Datasets:

ArmelR

/

the-pile-splitted

like 21

--- abstract: 'The rampant coronavirus disease 2019 (COVID-19) has brought global crisis with its deadly spread to more than 180 countries, and about 3,519,901 confirmed cases along with 247,630 deaths globally as on May 4, 2020. The absence of any active therapeutic agents and the lack of immunity against COVID-19 increases the vulnerability of the population. Since there are no vaccines available, social distancing is the only feasible approach to fight against this pandemic. Motivated by this notion, this article proposes a deep learning based framework for automating the task of monitoring social distancing using surveillance video. The proposed framework utilizes the YOLO v3 object detection model to segregate humans from the background and Deepsort approach to track the identified people with the help of bounding boxes and assigned IDs. The results of the YOLO v3 model are further compared with other popular state-of-the-art models, e.g. faster region-based CNN (convolution neural network) and single shot detector (SSD) in terms of mean average precision (mAP), frames per second (FPS) and loss values defined by object classification and localization. Later, the pairwise vectorized *L2* norm is computed based on the three-dimensional feature space obtained by using the centroid coordinates and dimensions of the bounding box. The violation index term is proposed to quantize the non adoption of social distancing protocol. From the experimental analysis, it is observed that the YOLO v3 with Deepsort tracking scheme displayed best results with balanced mAP and FPS score to monitor the social distancing in real-time.' author: - 'Narinder Singh Punn, Sanjay Kumar Sonbhadra and Sonali Agarwal [^1]' bibliography: - 'distance.bib' title: 'Monitoring COVID-19 social distancing with person detection and tracking via fine-tuned YOLO v3 and Deepsort techniques' --- COVID-19, Video surveillance, Social distancing, Object detection, Object tracking. Introduction ============ belongs to the family of coronavirus caused diseases, initially reported at Wuhan, China, during late December 2020. On March 11, it spread over 114 countries with 118,000 active cases and 4000 deaths, WHO declared this a pandemic [@bworld; @world2020director]. On May 4, 2020, over 3,519,901 cases and 247,630 deaths had been reported worldwide. Several healthcare organizations, medical experts and scientists are trying to develop proper medicines and vaccines for this deadly virus, but till date, no success is reported. This situation forces the global community to look for alternate ways to stop the spread of this infectious virus. Social distancing is claimed as the best spread stopper in the present scenario, and all affected countries are locked-down to implement social distancing. This research is aimed to support and mitigate the coronavirus pandemic along with minimum loss of economic endeavours, and propose a solution to detect the social distancing among people gathered at any public place. The word “social distancing” is best practice in the direction of efforts through a variety of means, aiming to minimize or interrupt the transmission of COVID-19. It aims at reducing the physical contact between possibly infected individuals and healthy persons. As per the WHO norms [@hensley2020social] it is prescribed that people should maintain at least 6 feet of distance among each other in order to follow social distancing. A recent study indicates that social distancing is an important containment measure and essential to prevent SARS-CoV-2, because people with mild or no symptoms may fortuitously carry corona infection and can infect others [@sa4]. Fig. \[fig2\] indicates that proper social distancing is the best way to reduce infectious physical contact, hence reduces the infection rate [@fong2020nonpharmaceutical; @ahmed2018effectiveness]. This reduced peak may surely match with the available healthcare infrastructure and help to offer better facilities to the patients battling against the coronavirus pandemic. ![An outcome of social distancing as the reduced peak of the epidemic and matching with available healthcare capacity.[]{data-label="fig2"}](fig2) Epidemiology is the study of factors and reasons for the spread of infectious diseases. To study epidemiological phenomena, mathematical models are always the most preferred choice. Almost all models descend from the classical SIR model of Kermack and McKendrick established in 1927 [@kermack1991contributions]. Various research works have been done on the SIR model and its extensions by the deterministic system [@eksin2019systematic], and consequently, many researchers studied stochastic biological systems and epidemic models [@zhao2016asymptotic]. Respiratory diseases are infectious where the rate and mode of transmission of the causing virus are the most critical factors to be considered for the treatment or ways to stop the spread of the virus in the community. Several medicine organizations and pandemic researchers are trying to develop vaccines for COVID-19, but still, there is no well-known medicine available for treatment. Hence, precautionary steps are taken by the whole world to restrict the spread of infection. Recently, Eksin et al. [@eksin2019systematic] proposed a modified SIR model with the inclusion of a social distancing parameter, $a(I, R)$ which can be determined with the help of the number of infected and recovered persons represented as $I$ and $R$, respectively. $$\begin{aligned} \frac{dS}{dt} &= -\beta S \frac{I}{N}a(I,N) \\ \frac{dI}{dt} &= -\delta I+\beta I \frac{I}{N}a(I,N) \\ \frac{dR}{dt} &= \delta I \end{aligned} \label{eq1}$$ where $\beta$ represents the infection rate and $\delta$ represents recovery rate. The population size is computed as $N = S + I + R$. Here the social distancing term ($a(I,R):{\mathbb{R}}^2 \ \epsilon \ [0,1]$) maps the transition rate from a susceptible state ($S$) to an infected state ($I$), which is calculated by $\frac{a\beta SI}{N}$. The social distancing models are of two types, where the first model is known as “long-term awareness” in which the occurrence of interaction of an individual with other is reduced proportionally with the cumulative percentage of affected (infectious and recovered) individuals (Eq. \[eq2\]), $$a = {\left (1- \frac{I+R}{N} \right)}^k \label{eq2}$$ Meanwhile, the second model is known as “short-term awareness”, where the reduction in interaction is directly proportional to the proportion of infectious individuals at a given instance (Eq. \[eq3\]), $$a = {\left (1- \frac{I}{N} \right)}^k \label{eq3}$$ where $k$ is behavior parameter defined as, $k \geq 0$. Higher value of $k$ implies that individuals are becoming sensitive to the disease prevalence. In the similar background, on April 16, 2020, a company Landing AI [@sA21] under the leadership of most recognizable names in AI, Dr. Andrew Ng [@sA22] announced the creation of an AI tool to monitor social distancing at the workplace. In a brief article, the company claimed that the upcoming tool could detect if people are maintaining the safe physical distance from each other by analyzing real-time video streams from the camera. It is also claimed that this tool can easily get integrated with existing security cameras available at different workplaces to maintain a safe distance among all workers. A brief demo was released that shows three steps: calibration, detection and measurement to monitor the social distancing. On April 21, 2020, Gartner, Inc. identified Landing AI as Cool Vendors in AI Core Technologies to appreciate their timely initiative in this revolutionary area to support the fight against the COVID -19 [@sA29]. Motivated by this, in this present work authors are attempting to check and compare the performance of popular object detection and tracking schemes in monitoring the social distancing. Rest of the paper structure is organized as follows: Section II presents the recent work proposed in this field of study, followed by the state-of-the-art object detection and tracking models in Section III. Later, in Section IV the deep learning based framework is proposed to monitor social distancing. In Section V experimentation and the corresponding results are discussed, accompanied by the outcome in Section VI. In Section VII the future scope and challenges are discussed and lastly Section VIII presents the conclusion of the present research work. Background study and related work ================================== Social distancing is surely the most trustworthy technique to stop the spreading of infectious disease, with this belief, in the background of December 2019, when COVID-19 emerged in Wuhan, China, it was opted as an unprecedented measure on January 23, 2020 [@sA8]. Within one month, the outbreak in China gained a peak in the first week of February with 2,000 to 4,000 new confirmed cases per day. Later, for the first time after this outbreak, there have been a sign of relief with no new confirmed cases for five consecutive days up to 23 March 2020 [@sA10]. This is evident that social distancing measures enacted in China initially, adopted worldwide later to control COVID-19. Prem et al. [@prem2020effect] aimed to study the effects of social distancing measures on the spread of the COVID-19 epidemic. Authors used synthetic location-specific contact patterns to simulate the ongoing trajectory of the outbreak using susceptible-exposed-infected-removed (SEIR) models. It was also suggested that premature and sudden lifting of social distancing could lead to an earlier secondary peak, which could be flattened by relaxing the interventions gradually [@prem2020effect]. As we all understand, social distancing though essential but economically painful measures to flatten the infection curve. Adolph et al. [@adolph2020pandemic] highlighted the situation of the United States of America, where due to lack of common consent among all policymakers it could not be adopted at an early stage, which is resulting into on-going harm to public health. Although social distancing impacted economic productivity, many researchers are trying hard to overcome the loss. Following from this context, Kylie et al. [@ainslie2020evidence] studied the correlation between the strictness of social distancing and the economic status of the region. The study indicated that intermediate levels of activities could be permitted while avoiding a massive outbreak. Since the novel coronavirus pandemic began, many countries have been taking the help of technology based solutions in different capacities to contain the outbreak [@sonbhadra2020target; @punn2020automated; @Punn2020.04.08.20057679]. Many developed countries, including India and South Korea, for instance, utilising GPS to track the movements of the suspected or infected persons to monitor any possibility of their exposure among healthy people. In India, the government is using the Arogya Setu App, which worked with the help of GPS and bluetooth to locate the presence of COVID-19 patients in the vicinity area. It also helps others to keep a safe distance from the infected person [@sA17]. On the other hand, some law enforcement departments have been using drones and other surveillance cameras to detect mass gatherings of people, and taking regulatory actions to disperse the crowd [@robakowska2017use; @8844927]. Such manual intervention in these critical situations might help flatten the curve, but it also brings a unique set of threats to the public and is challenging to the workforce. Human detection using visual surveillance system is an established area of research which is relying upon manual methods of identifying unusual activities, however, it has limited capabilities [@sulman2008effective]. In this direction, recent advancements advocate the need for intelligent systems to detect and capture human activities. Although human detection is an ambitious goal, due to a variety of constraints such as low-resolution video, varying articulated pose, clothing, lighting and background complexities and limited machine vision capabilities, wherein prior knowledge on these challenges can improve the detection performance [@wang2013intelligent]. Detecting an object which is in motion, incorporates two stages: object detection [@joshi2012survey] and object classification [@javed2002tracking]. The primary stage of object detection could be achieved by using background subtraction [@brutzer2011evaluation], optical flow [@aslani2013optical] and spatio-temporal filtering techniques [@dollar2005behavior]. In the background subtraction method [@piccardi2004background], the difference between the current frame and a background frame (first frame), at pixel or block level is computed. Adaptive Gaussian mixture, temporal differencing, hierarchical background models, warping background and non-parametric background are the most popular approaches of background subtraction [@xu2016background]. In optical flow-based object detection technique [@aslani2013optical], flow vectors associated with the object’s motion are characterised over a time span in order to identify regions in motion for a given sequence of images [@tsutsui2001optical]. Researchers reported that optical flow based techniques consist of computational overheads and are sensitive to various motion related outliers such as noise, colour and lighting, etc. [@agarwal2016review]. In another method of motion detection Aslani et al. [@dollar2005behavior] proposed spatio-temporal filter based approach in which the motion parameters are identified by using three-dimensional (3D) spatio-temporal features of the person in motion in the image sequence. These methods are advantageous due to its simplicity and less computational complexity, however shows limited performance because of noise and uncertainties on moving patterns [@niyogi1994analyzing]. Object detection problems have been efficiently addressed by recently developed advanced techniques. In the last decade, convolutional neural networks (CNN), region-based CNN [@zhao2019object] and faster region-based CNN [@krizhevsky2012imagenet] used region proposal techniques to generate the objectness score prior to its classification and later generates the bounding boxes around the object of interest for visualization and other statistical analysis [@ren2015faster]. Although these methods are efficient but suffer in terms of larger training time requirements [@chen2017implementation]. Since all these CNN based approaches utilize classification, another approach YOLO considers a regression based method to dimensionally separate the bounding boxes and interpret their class probabilities [@redmon2016you]. In this method, the designed framework efficiently divides the image into several portions representing bounding boxes along with the class probability scores for each portion to consider as an object. This approach offers excellent improvements in terms of speed while trading the gained speed with the efficiency. The detector module exhibits powerful generalization capabilities of representing an entire image [@putra2018convolutional]. Based on the above concepts, many research findings have been reported in the last few years. Crowd counting emerged as a promising area of research, with many societal applications. Eshel et al. [@eshel2008homography], focused on crowd detection and person count by proposing multiple height homographies for head top detection and solved the occlusions problem associated with video surveillance related applications. Chen et al. [@chen2009online] developed an electronic advertising application based on the concept of crowd counting. In similar application, Chih-Wen et al. [@su2009vision] proposed a vision-based people counting model. Following this, Yao et al. [@yao2011fast] generated inputs from stationary cameras to perform background subtraction to train the model for the appearance and the foreground shape of the crowd in videos. Once an object is detected, classification techniques can be applied to identify a human on the basis of shape, texture or motion-based features. In shape-based methods, the shape related information of moving regions such as points, boxes and blobs are determined to identify the human. This method performs poorly due to certain limitations in standard template-matching schemes [@wu2007detection; @eishita2012occlusion], which is further enhanced by applying part-based template matching [@singh2008human] approach. In another research, Dalal et al. [@dalal2005histograms] proposed texture-based schemes such as histograms of oriented gradient (HOG), which utilises high dimensional features based on edges along with the support vector machine (SVM) to detect humans. According to recent research, further identification of a person through video surveillance can be done by using face [@huang2010shape; @samal1992automatic] and gait recognition [@cunado1997using] techniques. However, detection and tracking of people under crowd are difficult sometimes due to partial or full occlusion problems. Leibe et al. [@leibe2005pedestrian] proposed trajectory estimation based solution while Andriluka et al. [@andriluka2008people] proposed a solution to detect partially occluded people using tracklet-based detectors. Many other tracking techniques, including a variety of object and motion representations, are reviewed by Yilmaz et al. [@yilmaz2006object]. A large number of studies are available in the area of video surveillance. Among many publically available datasets, KTH human motion dataset [@schuldt2004recognizing] shows six categories of activities, whereas INRIA XMAS multi-view dataset [@weinland2006free] and Weizmann human action dataset [@blank2005actions] contain 11 and 10 categories of actions, respectively. Another dataset named as performance evaluation of tracking and surveillance (PETS) is proposed by a group of researchers at university of Oxford [@parkhi2012oxford]. This dataset is available for vision based research comprising a large number of datasets for varying tasks in the field of computer vision. In the present research, in order to fine-tune the object detection and tracking models for identifying the person, open images datasets [@kuznetsova2020open] are considered. It is a collection of 19,957 classes out of which the models are trained for the identification of a person. The images are annotated with image-level labels and corresponding coordinates of the bounding boxes representing the person. Furthermore, the fine tuned proposed framework is simulated on the Oxford town center surveillance footage [@8844927] to monitor social distancing. We believe that having a single dataset with unified annotations for image classification, object detection, visual relationship detection, instance segmentation, and multimodal image descriptions will enable us to study and perform object detection tasks efficiently and stimulate progress towards genuine understanding of the scene. All explored literature and related research work clearly establishes a picture that the application of human detection can easily get extended to many applications to cater the situation that arises presently such as to check prescribed standards for hygiene, social distancing, work practices, etc. Object detection and tracking models ==================================== As observed from Fig. \[fig3\], the successful object detection models like RCNN [@girshick2014rich], fast RCNN [@girshick2015fast], faster RCNN [@ren2015faster], SSD [@liu2016ssd], YOLO v1 [@redmon2016you], YOLO v2 [@redmon2017yolo9000] and YOLO v3 [@redmon2018yolov3] tested on PASCAL-VOC [@everingham2010pascal] and MS-COCO [@lin2014microsoft] datasets, undergo trade-off between speed and accuracy of the detection which is dependent on various factors like backbone architecture (feature extraction network e.g. VGG-16 [@simonyan2014very], ResNet-101 [@he2016deep], Inception v2 [@szegedy2016rethinking], etc.), input sizes, model depth, varying software and hardware environment. A feature extractor tends to encode the model’s input into certain feature representation which aids in learning and discovering the patterns associated with the desired objects. In order to identify multiple objects of varying scale or size, it also uses predefined boxes covering an entire image termed as anchor boxes. Table \[tab1\] describes the performance in terms of accuracy for each of these popular and powerful feature extraction networks on ILSVRC ImageNet challenge [@russakovsky2015imagenet], along with the number of trainable parameters, which have a direct impact on the training speed and time. As highlighted in Table \[tab1\], the ratio of accuracy to the number of parameters is highest for Inception v2 model indicating that Inception v2 achieved adequate classification accuracy with minimal trainable parameters in contrast to other models, and hence is utilized as a backbone architecture for faster and efficient computations in the faster RCNN and SSD object detection models, whereas YOLO v3 uses different architecture Darknet-53 as proposed by Redmon et al. [@redmon2018yolov3]. ![Performance overview of the most popular object detection models on PASCAL-VOC and MS-COCO datasets. []{data-label="fig3"}](fig3) Backbone model Accuracy (a) Parameters (p) Ratio (a\*100/p) ------------------------------------------- -------------- ---------------- ------------------ VGG-16 [@simonyan2014very] 0.71 15 M 4.73 ResNet-101 [@he2016deep] 0.76 42.5 M 1.78 **Inception v2** [@szegedy2016rethinking] **0.74** **10 M** **7.40** Inception v3 [@szegedy2017inception] 0.78 22 M 3.58 Resnet v2 [@szegedy2017inception] 0.80 54 M 1.48 : Performance of the feature extraction network on ImageNet challenge. \[tab1\] Anchor boxes ------------ With the exhaustive literature survey, it is observed that every popular object detection model utilizes the concept of anchor boxes to detect multiple objects in the scene [@zhao2019object]. These boxes are overlaid on the input image over various spatial locations (per filter) with varying sizes and aspect ratio. In this article for an image of dimension breadth ($b$) $\times$ height ($h$) the anchor boxes are generated in the following manner. Consider the parameters, size as $p \ \epsilon \ (0,1]$ and aspect ratio as $r > 0$, then the anchor boxes for a certain location in an image can be constructed with dimensions as $bp\sqrt{r} \times hp\sqrt{r}$. Table \[tab2\] shows the values of $p$ and $r$ configured for each model. Later the object detection model is trained to predict for each generated anchor box to belong to a certain class, and an offset to adjust the dimensions of the anchor box to better fit the ground-truth of the object while using the classification and regression loss. Since there are many anchor boxes for a spatial location, the object can get associated with more than one anchor box. This problem is dealt with non-max suppression (NMS) by computing intersection over union (IoU) parameter that limits the anchor boxes association with the object of interest by calculating the score as the ratio of overlapping regions between the assigned anchor box and the ground-truth to the union of regions of the anchor box and the ground-truth. The score value is then compared with the set threshold hyperparameter to return the best bounding box for an object. [|l|l|l|l|l|]{} ----------- Detection model ----------- : Hyperparameters for generating the anchor boxes.[]{data-label="tab2"} & ------------- Size vector (p) ------------- : Hyperparameters for generating the anchor boxes.[]{data-label="tab2"} & -------------- Aspect ratio (r) -------------- : Hyperparameters for generating the anchor boxes.[]{data-label="tab2"} & -------- Anchor boxes -------- : Hyperparameters for generating the anchor boxes.[]{data-label="tab2"} & --------- IoU th. for NMS --------- : Hyperparameters for generating the anchor boxes.[]{data-label="tab2"} \ -------- Faster RCNN -------- : Hyperparameters for generating the anchor boxes.[]{data-label="tab2"} & [\[]{}0.25, 0.5, 1.0[\]]{} & [\[]{}0.5, 1.0, 2.0[\]]{} & 9 & 0.7\ SSD & [\[]{}0.2, 0.57, 0.95[\]]{} & [\[]{}0.3, 0.5, 1.0[\]]{} & 9 & 0.6\ YOLO v3 & [\[]{}0.25, 0.5, 1.0[\]]{} & [\[]{}0.5, 1.0, 2.0[\]]{} & 9 & 0.7\ ### Loss Function With each step of model training, predicted anchor box “a” is assigned a label as positive (1) or negative (0), based on its associativity with the object of interest having ground-truth box “g”. The positive anchor box is then assigned a class label $y_o \ \epsilon \ \{c_1, c_2,...., c_n\}$, here $c_n$ indicates the category of the $n^{th}$ object, while also generating the encoding vector for box “g” with respect to “a” as $f(g_a|a)$, where $y_o = 0$ for negative anchor boxes. Consider an image $I$, for some anchor “a”, model with trained parameters $\omega$, predicted the object class as $Y_{cls}(I|a;\omega)$ and the corresponding box offset as $Y_{reg}(I|a;\omega)$, then the loss for a single anchor prediction can be computed ($L_{cls}$) and bounding box regression loss ($L_{reg}$), as given by the Eq \[eq4\]. $$\begin{aligned} L(a|I; \omega) =\alpha. 1_{a}^{obj} L_{reg}(f(g_a|a) - Y_{reg}(I|a;\omega)) + \\ \beta. L_{cls} (y_a, Y_{cls}(I|a;\omega)) \end{aligned} \label{eq4}$$ where $1_{a}^{obj}$ is 1 if “a” is a positive anchor, $\alpha$ and $\beta$ are the weights associated with the regression and classification loss. Later, the overall loss of the model can be computed as the average of the $L(a|I;w)$ over the predictions for all the anchors. Faster RCNN ----------- Proposed by Ren et al. [@ren2015faster], the faster RCNN is derived from its predecessors RCNN [@girshick2014rich] and fast RCNN [@girshick2015fast], which rely on external region proposal approach based on selective search (SS) [@google]. Many researchers [@punn2020inception; @vaswani2018tensor2tensor; @amodei2016deep], observed that instead of using the SS, it is recommended to utilize the advantages of convolution layers for better and faster localization of the objects. Hence, Ren et al. proposed the Region Proposal Network (RPN) which uses CNN models, e.g. VGGNet, ResNet, etc. to generate the region proposals that made faster RCNN 10 times faster than fast RCNN. Fig. \[fig4\] shows the schematic representation of faster RCNN architecture, where RPN module performs binary classification of an object or not an object (background) while classification module assigns categories for each detected object (multi-class classification) by using the region of interest (RoI) pooling [@ren2015faster] on the extracted feature maps with projected regions. ![Schematic representation of faster RCNN architecture[]{data-label="fig4"}](fig4) ### Loss function The faster RCNN is the combination of two modules RPN and fast RCNN detector. The overall multi-task loss function is composed of classification loss and bounding box regression loss as defined in Eq. \[eq4\] with $L_{cls}$ and $L_{reg}$ functions defined in Eq. \[eq5\] $$\begin{aligned} L_{cls} (p_i, p_{i}^{*}) &= -p_{i}^{*} \log (p_i) - (1-p_{i}^{*}) \log (1- p_i) \\ L_{reg} (t^u, v) & = \sum_{x \epsilon {x, y, w, h} } L_{1}^{smooth}(t_{i}^{u} -v)\\ L_{1}^{smooth}(q)&=\begin{cases} 0.5 q^2, & if \mid q \mid < 1 .\\ \mid q \mid - 0.5, & \text{otherwise}. \end{cases} \end{aligned} \label{eq5}$$ where $t^u$ is the predicted corrections of the bounding box $t^u = \{t_{x}^{u}, t_{y}^{u}, t_{w}^{u}, t_{h}^{u}\}$. Here $u$ is a true class label, ($x$, $y$) corresponds to the top-left coordinates of the bounding box with height $h$ and width $w$, $v$ is a ground-truth bounding box, $p_{i}^{*}$ is the predicted class and $p_i$ is the actual class, Single Shot Detector (SSD) -------------------------- In this research, single shot detector (SSD) [@liu2016ssd] is also used as another object identification method to detect people in real-time video surveillance system. As discussed earlier, faster R-CNN works on region proposals to create boundary boxes to indicate objects, shows better accuracy, but has slow processing of frames per second (FPS). For real-time processing, SSD further improves the accuracy and FPS by using multi-scale features and default boxes in a single process. It follows the principle of the feed-forward convolution network which generates bounding boxes of fixed sizes along with a score based on the presence of object class instances in those boxes, followed by NMS step to produce the final detections. Thus, it consists of two steps: extracting feature maps and applying convolution filters to detect objects by using an architecture having three main parts. First part is a base pretrained network to extract feature maps, whereas, in the second part, multi-scale feature layers are used in which series of convolution filters are cascaded after the base network. The last part is a non-maximum suppression unit for eliminating overlapping boxes and one object only per box. The architecture of SSD is shown in Fig. \[fig5\]. ![Schematic representation of SSD architecture[]{data-label="fig5"}](fig5) ### Loss function Similar to the above discussed faster RCNN model, the overall loss function of the SSD model is equal to the sum of multi-class classification loss ($L_{cls}$) and bounding box regression loss (localization loss, $L_{reg}$), as shown in Eq. \[eq4\], where $L_{reg}$ and $L_{cls}$ is defined by Eq. \[eq6\] and \[eq7\]: $$\begin{aligned} L_{reg}(x,l,g) &=\sum_{i \epsilon pos}^{N} \sum_{m \epsilon c_x, c_y, w, h} x_{ij}^{k} {smooth}_{L_1}(l_{i}^{m} -{\hat{g}}_{j}^{m}) ,\\ {\hat{g}}_{j}^{c_x} &= \frac{ (g_{j}^{c_x}-a_{i}^{c_x})}{a_i^w} , \ {\hat{g}}_{j}^{c_y} = \frac{ (g_{j}^{c_y}-a_{i}^{c_y})}{a_i^h} ,\\ {\hat{g}}_{j}^{w} &= \log \left ( \frac{g_{j}^{w}}{a_{i}^{w}} \right ) ,\ {\hat{g}}_{j}^{h} = \log \left ( \frac{g_{j}^{h}}{a_{i}^{h}} \right ),\\ x_{ij}^{p} & =\begin{cases} 1, & \text{if IoU} > {0.5}\\ 0, & \text{otherwise}. \end{cases} \end{aligned} \label{eq6}$$ where $l$ is the predicted box, $g$ is the ground truth box, $x_{ij}^{p}$ is an indicator that matches the $i^{th}$ anchor box to the $j^{th}$ ground truth box, $c_x$ and $c_y$ are offsets to the anchor box $a$. $$\begin{aligned} L_{cls}(x,c) &= -\sum_{i \epsilon Pos}^{N} x_{ij}^{p} \log ({\hat{c}}_{i}^{p}) -\sum_{i \epsilon Neg} \log ({\hat{c}}_{i}^{o}) \end{aligned} \label{eq7}$$ where ${\hat{c}}_{i}^{p} = \frac{\exp{c_{i}^{p}}}{\sum_{p} \exp{c_{i}^{p}}}$ and $N$ is the number of default matched boxes. YOLO ---- For object detection, another competitor of SSD is YOLO [@redmon2016you]. This method can predict the type and location of an object by looking only once at the image. YOLO considers the object detection problem as a regression task instead of classification to assign class probabilities to the anchor boxes. A single convolutional network simultaneously predicts multiple bounding boxes and class probabilities. Majorly, there are three versions of YOLO: v1, v2 and v3. YOLO v1 is inspired by GoogleNet (Inception network) which is designed for object classification in an image. This network consists of 24 convolutional layers and 2 fully connected layers. Instead of the Inception modules used by GoogLeNet, YOLO v1 simply uses a reduction layer followed by convolutional layers. Later, YOLO v2 [@redmon2017yolo9000] is proposed with the objective of improving the accuracy significantly while making it faster. YOLO v2 uses Darknet-19 as a backbone network consisting of 19 convolution layers along with 5 max pooling layers and an output softmax layer for object classification. YOLO v2 outperformed its predecessor (YOLO v1) with significant improvements in mAP, FPS and object classification score. In contrast, YOLO v3 performs multi-label classification with the help of logistic classifiers instead of using softmax as in case of YOLO v1 and v2. In YOLO v3 Redmon et al. proposed Darknet-53 as a backbone architecture that extracts features maps for classification. In contrast to Darknet-19, Darknet-53 consists of residual blocks (short connections) along with the upsampling layers for concatenation and added depth to the network. YOLO v3 generates three predictions for each spatial location at different scales in an image, which eliminates the problem of not being able to detect small objects efficiently [@SK4]. Each prediction is monitored by computing objectness, boundary box regressor and classification scores. In Fig. \[fig6\] a schematic description of the YOLOv3 architecture is presented. ![Schematic representation of YOLO v3 architecture[]{data-label="fig6"}](fig6) ### Loss function The overall loss function of YOLO v3 consists of localization loss (bounding box regressor), cross entropy and confidence loss for classification score, defined as follows: $$\begin{aligned} {\lambda}_{coord} \sum_{i=0}^{S^2} \sum_{j=0}^{B} {{1}}_{i,j}^{obj} ({(t_x - {\hat{t}}_x)}^2 + {(t_y - {\hat{t}}_y)}^2 +{(t_w - {\hat{t}}_w)}^2 + \\ {(t_h - {\hat{t}}_h)}^2) \\ + \sum_{i=0}^{S^2} \sum_{j=0}^{B} {{1}}_{i,j}^{obj} (-\log(\sigma(t_o)) + \sum_{k=1}^{C} BCE({\hat{y}}_k, \sigma (s_k))) \\ + {\lambda}_{noobj}\sum_{i=0}^{S^2} \sum_{j=0}^{B} {{1}}_{i,j}^{noobj} (-\log(1- \sigma(t_o)) \end{aligned} \label{eq9}$$ where ${\lambda}_{coord}$ indicates the weight of the coordinate error, ${S^2}$ indicates the number of grids in the image, and $B$ is the number of generated bounding boxes per grid. ${1}_{i,j}^{obj} = 1$ describes that object confines in the $j^{th}$ bounding box in grid $i$, otherwise it is $0$. Deepsort -------- Deepsort is a deep learning based approach to track custom objects in a video [@wojke2017simple]. In the present research, Deepsort is utilized to track individuals present in the surveillance footage. It makes use of patterns learned via detected objects in the images which is later combined with the temporal information for predicting associated trajectories of the objects of interest. It keeps track of each object under consideration by mapping unique identifiers for further statistical analysis. Deepsort is also useful to handle associated challenges such as occlusion, multiple viewpoints, non-stationary cameras and annotating training data. For effective tracking, the Kalman filter and the Hungarian algorithm are used. Kalman filter is recursively used for better association, and it can predict future positions based on the current position. Hungarian algorithm is used for association and id attribution that identifies if an object in the current frame is the same as the one in the previous frame. Initially, a Faster RCNN is trained for person identification and for tracking, a linear constant velocity model [@wojke2018deep] is utilized to describe each target with eight dimensional space as follows: $$x = {[u, v, \lambda, h, x^{,}, y^{,}, {\lambda}^{,} , h^{,} ]}^T$$ where ($u,v$) is the centroid of the bounding box, $a$ is the aspect ratio and $h$ is the height of the image. The other variables are the respective velocities of the variables. Later, the standard Kalman filter is used with constant velocity motion and linear observation model, where the bounding coordinates ($u, v, \lambda, h$) are taken as direct observations of the object state. For each track $k$, starting from the last successful measurement association $a_k$, the total number of frames are calculated. With positive prediction from the Kalman filter, the counter is incremented and later when the track gets associated with a measurement it resets its value to $0$. Furthermore, if the identified tracks exceed a predefined maximum age, then those objects are considered to have left the scene and the corresponding track gets removed from the track set. And if there are no tracks available for some detected objects then new track hypotheses are initiated for each unidentified track of novel detected objects that cannot be mapped to the existing tracks. For the first three frames the new tracks are classified as indefinite until a successful measurement mapping is computed. If the tracks are not successfully mapped with measurement then it gets deleted from the track set. Hungarian algorithm is then utilized in order to solve the mapping problem between the newly arrived measurements and the predicted Kalman states by considering the motion and appearance information with the help of Mahalanobis distance computed between them as defined in Eq. \[eq10\]. $$d^{(1)} (i,j) = {(d_j -y_i)}^T S_{i}^{-1}(d_j -y_i) \label{eq10}$$ where the projection of the $i^{th}$ track distribution into measurement space is represented by ($y_i ,S_i$) and the $j^{th}$ bounding box detection by $d_j$. The Mahalanobis distance considers this uncertainty by estimating the count of standard deviations, the detection is away from the mean track location. Further, using this metric, it is possible to exclude unlikely associations by thresholding the Mahalanobis distance. This decision is denoted with an indicator that evaluates to 1 if the association between the $i^{th}$ track and $j^{th}$ detection is admissible (Eq. \[eq11\]). $$b_{i,j}^{(1)} = 1 [d^{(1)} (i,j) < t^{(1)} ] \label{eq11}$$ Though Mahalanobis distance performs efficiently but fails in the environment where camera motion is possible, thereby another metric is introduced for the assignment problem. This second metric measures the smallest cosine distance between the $i^{th}$ track and $j^{th}$ detection in appearance space as follows: $$d^{(2)} (i, j)= min\{1- {r_j}^T {r_k}^{(i)} \mid {r_k}^{(i)} \ \epsilon \ {\mathbb{R}}^2 \} \label{eq12}$$ Again, a binary variable is introduced to indicate if an association is admissible according to the following metric: $$b_{i,j}^{(1)} = 1 [d^{(2)} (i,j) < t^{(2)} ] \label{eq13}$$ and a suitable threshold is measured for this indicator on a separate training dataset. To build the association problem, both metrics are combined using a weighted sum: $$c_{i,j} = \lambda d^{(1)} (i,j) + (1 - \lambda) d^{(2)} (i, j)$$ where an association is admissible if it is within the gating region of both metrics: $$b_{i,j} = \prod_{m=1}{2} b_{i,j}^{(m)}.$$ The influence of each metric on the combined association cost can be controlled through hyperparameter $\lambda$. Proposed approach ================= The emergence of deep learning has brought the best performing techniques for a wide variety of tasks and challenges including medical diagnosis [@punn2020inception], machine translation [@vaswani2018tensor2tensor], speech recognition [@amodei2016deep], and a lot more [@pouyanfar2018survey]. Most of these tasks are centred around object classification, detection, segmentation, tracking, and recognition [@brunetti2018computer; @punn2019crowd]. In recent years, the convolution neural network (CNN) based architectures have shown significant performance improvements that are leading towards the high quality of object detection, as shown in Fig. \[fig3\], which presents the performance of such models in terms of mAP and FPS on standard benchmark datasets, PASCAL-VOC [@everingham2010pascal] and MS-COCO [@lin2014microsoft], and similar hardware resources. In the present article, a deep learning based framework is proposed that utilizes object detection and tracking models to aid in the social distancing remedy for dealing with the escalation of COVID-19 cases. In order to maintain the balance of speed and accuracy, YOLO v3 [@redmon2018yolov3] alongside the Deepsort [@wojke2017simple] are utilized as object detection and tracking approaches while surrounding each detected object with the bounding boxes. Later, these bounding boxes are utilized to compute the pairwise *L2* norm with computationally efficient vectorized representation for identifying the clusters of people not obeying the order of social distancing. Furthermore, to visualize the clusters in the live stream, each bounding box is color-coded based on its association with the group where people belonging to the same group are represented with the same color. Each surveillance frame is also accompanied with the streamline plot depicting the statistical count of the number of social groups and an index term (violation index) representing the ratio of the number of people to the number of groups. Furthermore, estimated violations can be computed by multiplying the violation index with the total number of social groups. Workflow --------- This section includes the necessary steps undertaken to compose a framework for monitoring social distancing. - Fine-tune the trained object detection model to identify and track the person in a footage. - The trained model is feeded with the surveillance footage. The model generates a set of bounding boxes and an ID for each identified person. - Each individual is associated with three-dimensional feature space ($x, y, d$), where ($x$, $y$) corresponds to the centroid coordinates of the bounding box and $d$ defines the depth of the individual as observed from the camera. $$d = ((2 * 3.14 * 180) / (w + h * 360) * 1000 + 3)$$ where $w$ is the width of the bounding box and $h$ is the height of the bounding box [@kaggle]. - For the set of bounding boxes, pairwise *L2* norm is computed as given by the following equation. $$||D||_2=\sqrt{\sum_{i=1}^{n} {(q_i -p_i)}^2}$$ where in this work $n = 3$. - The dense matrix of *L2* norm is then utilized to assign the neighbors for each individual that satisfies the closeness sensitivity. With extensive trials the closeness threshold is updated dynamically based on the spatial location of the person in a given frame ranging between ($90, 170$) pixels. - Any individual that meets the closeness property is assigned a neighbour or neighbours forming a group represented in a different color coding in contrast to other people. - The formation of groups indicates the violation of the practice of social distancing which is quantified with help of the following: - Consider $n_g$ as number of groups or clusters identified, and $n_p$ as total number of people found in close proximity. - $v_i = n_p/n_g$, where $v_i$ is the violation index. Experiments and results ======================= The above discussed object detection models are fine tuned for binary classification (person or not a person) with Inception v2 as a backbone network on the Nvidia GTX 1060 GPU, using the dataset acquired from the open image dataset (OID) repository [@google] maintained by the Google open source community. The diverse images with a class label as “Person” are downloaded via OIDv4 toolkit [@megapixels] along with the annotations. Fig. \[fig7\] shows the sample images of the obtained dataset consisting of 800 images which is obtained by manually filtering to only contain the true samples. The dataset is then divided into training and testing sets, in 8:2 ratio. In order to make the testing robust, the testing set is also accompanied by the frames of surveillance footage of the Oxford town center [@8844927]. Later this footage is also utilized to simulate the overall approach for monitoring the social distancing. ![Data samples showing (a) true samples and (b) false samples of a “Person” class from the open image dataset. []{data-label="fig7"}](fig7) In case of faster RCNN, the images are resized to $P$ pixels on the shorter edge with 600 and 1024 for low and high resolution, while in SSD and YOLO the images are scaled to the fixed dimension $P\times P$ with $P$ value as 416. During the training phase, the performance of the models is continuously monitored using the mAP along with the localization, classification and overall loss in the detection of the person as indicated in Fig. \[fig8\]. Table \[tab3\] summarizes the results of each model obtained at the end of the training phase with the training time (TT), number of iterations (NoI), mAP, and total loss (TL) value. It is observed that the faster RCNN model achieved minimal loss with maximum mAP, however, has the lowest FPS, which makes it not suitable for real-time applications. Furthermore, as compared to SSD, YOLO v3 achieved better results with balanced mAP, training time, and FPS score. The trained YOLO v3 model is then utilized for monitoring the social distancing on the surveillance video. ![Losses per iteration of the object detection models during the training phase on the OID validation set for detecting the person in an image. []{data-label="fig8"}](fig8) Output ====== The proposed framework outputs (as shown in Fig. \[fig9\]) the processed frame with the identified people confined in the bounding boxes while also simulating the statistical analysis showing the total number of social groups displayed by same color encoding and a violation index term computed as the ratio of the number of people to the number of groups. The frames shown in Fig. \[fig9\] displays violation index as 3, 2, 2, and 2.33. The frames with detected violations are recorded with the timestamp for future analysis. Model TT (in sec.) NoI mAP TL FPS ------------- -------------- ---------- ----------- ---------- -------- Faster RCNN 9651 12135 0.969 0.02 3 SSD 2124 1200 0.691 0.22 10 **YOLO v3** **5659** **7560** **0.846** **0.87** **23** : Performance comparison of the object detection models.[]{data-label="tab3"} ![Sample output of the proposed framework for monitoring social distancing on surveillance footage of Oxford Town Center. []{data-label="fig9"}](fig9) Future scope and challenges =========================== Since this application is intended to be used in any working environment; accuracy and precision are highly desired to serve the purpose. Higher number of false positive may raise discomfort and panic situation among people being observed. There may also be genuinely raised concerns about privacy and individual rights which can be addressed with some additional measures such as prior consents for such working environments, hiding a person’s identity in general, and maintaining transparency about its fair uses within limited stakeholders. Conclusion ========== The article proposes an efficient real-time deep learning based framework to automate the process of monitoring the social distancing via object detection and tracking approaches, where each individual is identified in the real-time with the help of bounding boxes. The generated bounding boxes aid in identifying the clusters or groups of people satisfying the closeness property computed with the help of pairwise vectorized approach. The number of violations are confirmed by computing the number of groups formed and violation index term computed as the ratio of the number of people to the number of groups. The extensive trials were conducted with popular state-of-the-art object detection models: Faster RCNN, SSD, and YOLO v3, where YOLO v3 illustrated the efficient performance with balanced FPS and mAP score. Since this approach is highly sensitive to the spatial location of the camera, the same approach can be fine tuned to better adjust with the corresponding field of view. Acknowledgment {#acknowledgment .unnumbered} ============== The authors gratefully acknowledge the helpful comments and suggestions of colleagues. Authors are also indebted to Interdisciplinary Cyber Physical Systems (ICPS) Programme, Department of Science and Technology (DST), Government of India (GoI) vide Reference No.244 for their financial support to carry out the background research which helped significantly for the implementation of present research work. [^1]: N. S. Punn, S. K. Sonbhadra, S. Agarwal, Indian Institute of Information Technology Allahabad, Jhalwa, Prayagraj, Uttar Pradesh, India; emails: {pse2017002, rsi2017502, sonali}@iiita.ac.in.

ArXiv

--- author: - 'Miguel Pereira-Santaella$^{1}$, Luigi Spinoglio$^{1}$, Paul P. van der Werf$^{2}$, Javier Piqueras López$^{3}$' title: 'Warm molecular gas temperature distribution in six local infrared bright Seyfert galaxies[^1]' --- ------------------ ------------ ------------- ------- ------------- --------------- ------------------- -------------------- ------------------------ -- -- -- Galaxy Name R.A. Decl. $cz$ $D_{\rm L}$ Nuclear $\log L_{\rm IR}$ $\alpha_{\rm AGN}$ Obs. ID (J2000.0) (J2000.0) () (Mpc) spect. class. () (%) NGC 34 00 11 06.5 $-$12 06 26 5881 77.0 Sy2 11.4 2 1342199253 IRASF 05189–2524 05 21 01.4 $-$25 21 45 12760 181 Sy2 12.2 71 1342192832, 1342192833 UGC 05101 09 35 51.6 +61 21 11 11802 168 Sy2 12.0 35 1342209278 NGC 5135 13 25 44.0 $-$29 50 01 4105 60.9 Sy2 11.3 20 1342212344 NGC 7130 21 48 19.5 $-$34 57 04 4842 63.6 Sy2 11.3 25 1342219565 NGC 7469 23 03 15.6 +08 52 26 4892 56.6 Sy1 11.5 20 1342199252 ------------------ ------------ ------------- ------- ------------- --------------- ------------------- -------------------- ------------------------ -- -- -- Introduction ============ Molecular gas is an important phase of the interstellar medium (ISM). This phase contains a significant fraction of the total mass, and stars form in it. But the study of molecular gas presents some complications. First, the lower energy levels of [H$_2$]{}, the main component of the ISM phase, have energies $>500$K, thus in cold molecular gas ($T< 100$K) most of the [H$_2$]{} is in the fundamental state and no [H$_2$]{} emission lines are produced. And second, only the near infrared (IR) ro-vibrational [H$_2$]{} transitions, with $E_{\rm up}>6000$K, are observable from ground telescopes, so only very high-temperature ($T > 1500$K) molecular gas can be detected. To overcome the first caveat, other abundant molecules with observable transitions in the millimeter range (like CO, HCN, etc.) are used as tracers of molecular gas. In particular, the lowest rotational transitions of CO, the second most abundant molecule, are commonly used to study the molecular gas content of galaxies. However, these low-$J$ CO transitions mainly originate in the coldest molecular gas. Thus, ground observations are limited to the study of either the warmest or the coldest molecular gas. Just recently, thanks to IR and sub-millimeter space observatories like the *Infrared Space Observatory* (*ISO*; @Kessler1996), the *Spitzer Space Telescope* [@Werner2004], and the [*Herschel*]{} Space Observatory [@Pilbratt2010Herschel], the rotational [H$_2$]{} transitions as well as the intermediate-$J$ CO transitions became accessible for a large number of local galaxies (e.g., @Rigopoulou02 [@Roussel07; @vanderWerf2010; @Pereira2013]). Therefore, now for the first time, it is possible to obtain a complete snapshot of molecular gas emission and study its physical properties (temperature, density, column density, etc.) and the excitation mechanisms (ultraviolet (UV) radiation, shocks, and X-ray and cosmic rays). In this work, we present new data obtained by the Fourier transform spectrometer (FTS) module of the Spectral and Photometric Imaging Receiver (SPIRE) instrument on-board [*Herschel*]{} [@Griffin2010SPIRE; @Naylor2010; @Swinyard2010] for six local active luminous IR galaxies. These [[SPIRE/FTS]{}]{} data cover the 210–670 (450–1440GHz) spectral range, so the mid-$J$ CO lines ($J=4-3$ to $J=13-12$) are observed. We completed the CO spectral line energy distributions (SLEDs) with ground-based observations of the three lowest $J$ CO transitions. In addition, we complemented the CO SLEDs with the [H$_2$]{} SLEDs obtained from near- and mid-IR observations of these galaxies. We used the available mid-IR spectroscopy obtained by the [*Spitzer*]{} IR spectrograph (IRS; @HouckIRS) to measure the lowest rotational [H$_2$]{} transitions (e.g., @Wu2009 [@Pereira2010]), and near-IR integral field spectroscopy obtained by the Spectrograph for INtegral Field Observations in the Near-Infrared (SINFONI; @Eisenhauer2003) on the Very Large Telescope (VLT) for the ro-vibrational [H$_2$]{} transitions. For the first time, we have performed a radiation transfer analysis of the whole set of molecular lines together (i.e., CO rotational and [H$_2$]{} rotational and ro-vibrational) in local IR bright galaxies. In total, the compiled CO and [H$_2$]{} SLEDs contain information for 26 transitions with upper level energies between 5 and 15000K, thus the emission from most of the molecular gas is included. The paper is organized as follows. In Section \[s:observations\], we present the sample and the data reduction. Sections \[s:rad\_models\] and \[s:sled\_fit\] describe the radiative transfer models used, and the fitting of the SLEDs. The cold-to-warm molecular gas ratio and the heating mechanisms are discussed in Sections \[s:cold-to-warm\_ratio\] and \[s:heating\], respectively. We summarize the main results in Section \[s:conclusions\]. Observations and data reduction {#s:observations} =============================== Sample ------ Our sample contains six local (50–180Mpc) IR bright Seyfert galaxies observed by [[SPIRE/FTS]{}]{} through two programs (PIs: P. van der Werf and L. Spinoglio). Their $\log L_{\rm IR}\slash L_{\rm \odot}$ ranges between 11.4 and 12.2 (see Table \[tab:sample\]). All of them host an active nucleus (AGN), although the AGN dominates the energy output of the galaxy only for IRASF 05189–2524. For the remaining galaxies, the AGN contributes between 2 and 35% of the total luminosity. Three of the galaxies (NGC 34, IRASF 05189–2524, and UGC 05101) are advanced major mergers. The other three galaxies (NGC 5135, NGC 7130, and NGC 7469) are spirals, although NGC 7130 and NGC 7469 have peculiar morphologies suggesting recent minor interactions [@Genzel1995; @Bellocchi2012]. [*Herschel*]{} [[SPIRE/FTS]{}]{} spectroscopy {#ss:spectroscopy} --------------------------------------------- We obtained [[SPIRE/FTS]{}]{} high-resolution (1.45GHz) spectroscopic observations of six nearby Seyfert luminous IR galaxies. Integration times varied between 5 and 34 ks corresponding to a 3$\sigma$ line detection limit $\sim$0.5–1$\times10^{-15}$ergcm$^{-2}$s$^{-1}$. The [[SPIRE/FTS]{}]{} consists of two bolometer arrays, the spectrometer short wavelength (SSW; 925–1545GHz) and the spectrometer long wavelength (SLW; 446–948GHz). They sparsely cover a field of view (FoV) of $\sim$2 with the central bolometer centered at the nuclei of the galaxies. The full-width half-maximum (FWHM) of the SSW beam is 18 almost constant with frequency. For the SLW bolometers the beam FWHM varies between $\sim$30 and $\sim$42 with a complicated dependence on frequency [@Swinyard2010]. The SSW beams size correspond to 5–15kpc at the distance of our galaxies, therefore these galaxies are almost point like sources for both the SSW and SLW bolometers (see also Figure 1 of @Pereira2013). The data was reduced as described by @Pereira2013, but using the more recent [*Herschel*]{} interactive pipeline environment software (HIPE) version 11 [@Ott2010HIPE]. In brief, the pipeline first creates the interferograms from the bolometer timelines. After the interferogram phase errors are corrected and the baselines removed a Fourier transform is applied to obtain the spectra. These are dominated by the thermal emission of the telescope that is later removed. Residual background emission is subtracted by averaging the off-source bolometers. Finally, the point-source calibration is applied to the spectra extracted from the central bolometers (SLWC3 and SSWD4). The final spectra are plotted in Figure \[fig:spectra\]. To measure the line fluxes, we fitted a sinc function to the line profiles[^2]. The local continuum was estimated using a linear fit around the line frequency. When two lines were close in frequency we fitted them simultaneously. The fluxes of the $^{12}$CO, HF, H$_2$O, and fine structure atomic transitions are listed in Table \[tab:lines\]. Transitions of CH$^+$, H$_2$O$^+$, and OH$^+$ detected in some galaxies are reported in Table \[tab:lines\_uncommon\]. The [[SPIRE/FTS]{}]{} data of two of these galaxies, UGC 05101 and NGC 7130, were already presented by @Pereira2013. To take advantage of the newest pipeline and calibration they were reprocessed with the rest of the sample. For these two galaxies, the line fluxes in Tables \[tab:lines\] and \[tab:lines\_uncommon\] agree with those in @Pereira2013 within the 1$\sigma$ uncertainties. {width="1.8\columnwidth"} {width="1.8\columnwidth"} {width="1.8\columnwidth"} {width="1.8\columnwidth"} {width="1.8\columnwidth"} {width="1.8\columnwidth"} [lrcccccccccc]{} Transition & & &\ \ & & & NGC 34 & IRASF 05189–2524 & UGC 05101 & NGC 5135 & NGC 7130 & NGC 7469\ $^{12}$CO J = 4–3 & 461.041 & & $<$10.6 & [  $\cdots$  ]{} & [  $\cdots$  ]{} & 33.2 $\pm$ 3.7 & 33.0 $\pm$ 3.8 & 28.8 $\pm$ 3.3\ $^{12}$CO J = 5–4 & 576.268 & & 14.3 $\pm$ 1.9 & 5.7 $\pm$ 1.3 & 11.6 $\pm$ 3.3 & 35.9 $\pm$ 3.9 & 34.2 $\pm$ 3.8 & 38.6 $\pm$ 4.2\ $^{12}$CO J = 6–5 & 691.473 & & 17.7 $\pm$ 2.1 & 9.1 $\pm$ 1.3 & 14.0 $\pm$ 2.0 & 36.1 $\pm$ 3.7 & 25.6 $\pm$ 2.9 & 37.5 $\pm$ 3.9\ $^{12}$CO J = 7–6 & 806.652 & & 21.3 $\pm$ 2.4 & 7.9 $\pm$ 1.1 & 10.5 $\pm$ 1.6 & 29.3 $\pm$ 3.1 & 25.5 $\pm$ 2.8 & 31.1 $\pm$ 3.3\ $^{12}$CO J = 8–7 & 921.800 & & 18.2 $\pm$ 2.3 & 10.5 $\pm$ 1.6 & 10.6 $\pm$ 2.3 & 23.4 $\pm$ 2.7 & 20.8 $\pm$ 2.5 & 26.9 $\pm$ 3.1\ $^{12}$CO J = 9–8 & 1036.912 & & 22.6 $\pm$ 2.6 & 8.9 $\pm$ 1.4 & $<$10.1 & 17.5 $\pm$ 2.1 & 24.2 $\pm$ 2.8 & 21.4 $\pm$ 2.4\ $^{12}$CO J = 10–9 & 1151.985 & & 18.3 $\pm$ 2.2 & 11.2 $\pm$ 1.5 & 12.1 $\pm$ 2.2 & 12.2 $\pm$ 1.8 & 16.9 $\pm$ 2.6 & 15.4 $\pm$ 2.1\ $^{12}$CO J = 11–10 & 1267.014 & & 17.1 $\pm$ 2.1 & 9.3 $\pm$ 1.4 & $<$10.0 & 7.1 $\pm$ 1.6 & 12.8 $\pm$ 2.1 & 13.0 $\pm$ 1.9\ $^{12}$CO J = 12–11 & 1381.995 & & 12.5 $\pm$ 1.8 & 8.0 $\pm$ 1.2 & $<$8.2 & 8.5 $\pm$ 1.4 & 10.5 $\pm$ 1.9 & 9.9 $\pm$ 1.7\ $^{12}$CO J = 13–12 & 1496.923 & & 9.3 $\pm$ 1.7 & 8.0 $\pm$ 1.3 & $<$10.5 & $<$11.6 & $<$12.1 & 9.1 $\pm$ 1.9\ \ o-H$_2$O 1$_{10}$–1$_{01}$ & 556.936 & & $<$6.2 & $<$5.3 & $<$12.9 & $<$7.6 & $<$14.7 & $<$12.7\ p-H$_2$O 2$_{11}$–2$_{02}$ & 752.033 & & 3.6 $\pm$ 1.0 & 4.5 $\pm$ 1.0 & 7.0 $\pm$ 1.6 & 5.4 $\pm$ 1.3 & 5.9 $\pm$ 1.6 & 5.4 $\pm$ 1.4\ p-H$_2$O 2$_{02}$–1$_{11}$ & 987.927 & & 9.6 $\pm$ 1.7 & $<$6.5 & $<$10.1 & 11.2 $\pm$ 2.5 & 18.8 $\pm$ 3.0 & $<$8.5\ o-H$_2$O 3$_{12}$–3$_{03}$ & 1097.365 & & $<$5.3 & 5.5 $\pm$ 1.3 & $<$9.1 & 6.3 $\pm$ 1.8 & 10.5 $\pm$ 2.3 & $<$6.0\ p-H$_2$O 1$_{11}$–0$_{00}$ & 1113.343 & & $<$7.4 & $<$5.7 & $<$9.1 & 8.9 $\pm$ 1.9 & $<$10.0 & 12.0 $\pm$ 1.6\ o-H$_2$O 3$_{21}$–3$_{12}$ & 1162.912 & & 7.7 $\pm$ 1.7 & 9.1 $\pm$ 1.2 & 9.0 $\pm$ 2.1 & $<$6.4 & $<$12.4 & 6.8 $\pm$ 2.0\ p-H$_2$O 4$_{22}$-4$_{13}$ & 1207.639 & & $<$5.6 & 4.6 $\pm$ 1.4 & $<$9.8 & $<$8.0 & $<$11.3 & $<$7.6\ p-H$_2$O 2$_{20}$–2$_{11}$ & 1228.789 & & 4.1 $\pm$ 1.1 & 4.0 $\pm$ 1.0 & $<$9.6 & 8.9 $\pm$ 1.5 & 12.8 $\pm$ 2.1 & 6.8 $\pm$ 1.7\ \ \ $[$$]$ $^{3}$P$_{1}$-$^{3}$P$_{0}$ & 492.161 & & 7.7 $\pm$ 2.1 & $<$6.0 & $<$12.6 & 28.6 $\pm$ 2.4 & 16.4 $\pm$ 3.5 & 19.9 $\pm$ 2.7\ $[$$]$ $^{3}$P$_{2}$-$^{3}$P$_{1}$ & 809.342 & & 9.9 $\pm$ 1.2 & 5.5 $\pm$ 0.7 & 12.8 $\pm$ 1.6 & 59.5 $\pm$ 1.0 & 26.7 $\pm$ 1.3 & 37.7 $\pm$ 1.3\ $[$$]$ $^{3}$P$_{1}$-$^{3}$P$_{0}$ & 1461.132 & & 20.7 $\pm$ 2.5 & 11.0 $\pm$ 1.4 & 31.3 $\pm$ 3.8 & 147.0 $\pm$ 3.3 & 118.3 $\pm$ 3.0 & 99.6 $\pm$ 3.7\ [llrccccccccc]{} Galaxy & Transition & & Flux\ & & & (10$^{-15}$ergcm$^{-2}$s$^{-1}$)\ UGC 05101 & HF $J = 1-0$ & 1232.476 & –10.4 $\pm$ 2.3$^a$\ \ NGC 5135 & CH$^+$ $J = 1-0$ & 835.079 & 5.1 $\pm$ 1.6\ & HF $J = 1-0$ & 1232.476 & 10.5 $\pm$ 1.5\ & o-H$_2$O$^+$ $1_{11}-0_{00}$ & 1115.204 & 8.8 $\pm$ 1.8\ & o-H$_2$O$^+$ $1_{11}-0_{00}$ & 1139.561 & 8.9 $\pm$ 1.7\ & OH$^+$ $1_{0}-0_{1}$ & 909.159 & 5.8 $\pm$ 1.6\ & OH$^+$ $1_{2}-0_{1}$ & 971.805 & 23.4 $\pm$ 2.8\ & OH$^+$ $1_{1}-0_{1}$ & 1032.998 & 14.9 $\pm$ 1.4\ \ NGC 7130 & HF $J = 1-0$ & 1232.476 & 8.3 $\pm$ 2.0\ & OH$^+$ $1_{0}-0_{1}$ & 909.159 & 6.5 $\pm$ 1.9\ & OH$^+$ $1_{2}-0_{1}$ & 971.805 & 26.1 $\pm$ 3.6\ & OH$^+$ $1_{1}-0_{1}$ & 1032.998 & 10.7 $\pm$ 2.2\ \ NGC 7469 & HF $J = 1-0$ & 1232.476 & 7.7 $\pm$ 1.6\ & o-H$_2$O$^+$ $1_{11}-0_{00}$ & 1115.204 & 12.7 $\pm$ 1.7\ & o-H$_2$O$^+$ $1_{11}-0_{00}$ & 1139.561 & $<$6.8\ & OH$^+$ $1_{0}-0_{1}$ & 909.159 & 8.4 $\pm$ 2.1\ & OH$^+$ $1_{2}-0_{1}$ & 971.805 & $<$9.8\ & OH$^+$ $1_{1}-0_{1}$ & 1032.998 & 12.7 $\pm$ 1.3\ Spitzer/IRS spectroscopy {#ss:spitzer_irs} ------------------------ Mid-IR Spitzer/IRS spectroscopic observations of these galaxies were available on the [*Spitzer*]{} data archive. Most of the data is already published in several papers (e.g., @Armus2004 [@Wu2009; @Pereira2010; @Tommasin2010; @Esquej2012]), but the fluxes (and upper limits) of the [H$_2$]{} transitions are not always reported. For this reason, we retrieved the data from the archive and reprocessed it in a uniform way. These observations include low- and high-resolution spectroscopy ($R\sim$60–120 and 600, respectively) covering the spectral range 5.5–14 (short-low resolution modules) and 10–36(short- and long-high-resolution modules). The observations obtained in the staring mode were processed using the standard pipeline version S18.18 included in the Spitzer IRS Custom Extraction software (SPICE). We assumed the point source flux calibration for these galaxies. To process the spectral mapping observations, we used CUBISM [@SmithCUBISM]. This software combines the individual slit observations to create the spectral data cubes. From the cubes, we extracted the spectra using a 15$\times$15 square aperture and then we applied an aperture correction (see @AAH2012a). The spectra are shown in Figure \[fig:spectra\_irs\]. We measured the fluxes of the [H$_2$]{} rotational transitions S(0), S(1), and S(2) in the high-resolution spectra[^3] fitting a Gaussian profile and a linear function for the local continuum (Table \[tab:h2fluxes\]). The [H$_2$]{} S(3) and S(5) transitions were measured in low-resolution spectra. Because of the large number of spectral features in the mid-IR range, it is not trivial to determine the continuum for the low-resolution spectra. Therefore, we used PAHFIT [@Smith07] to fit the complete spectrum and estimate the continuum and line fluxes. It is not possible to obtain a reliable measurement of the [H$_2$]{} S(4) transition at 8.03 because it is blended with the 7.7 PAH broad feature. Similarly, the [H$_2$]{} S(5) transition at 6.91 is blended with the \[\]6.99 line, which in our galaxies is always stronger than the [H$_2$]{} line. Therefore, the [H$_2$]{} S(5) fluxes should be considered with caution. The strength of the silicate feature at 9.7 ($S_{\rm Si}$) was measured in the low-resolution spectra as described by @Pereira2010. It is defined as $S_{\rm Si}=\ln \left(f_{\rm obs}\slash f_{\rm cont} \right)$, thus negative $S_{\rm Si}$ implies that the feature is seen in absorption (Table \[tab:extinction\]). We used the strength of the silicate feature to estimate extinction in our galaxies (see Section \[ss:extinction\]). ![Rest-frame [*Spitzer*]{}/IRS spectra. The short-low (5.5–14), short-high (10–19), and long-high (19–36) spectra are plotted for each galaxy. For clarity, the spectra are multiplied by the following factors 1, 2.5, 14, 100, 220, and 2000, from bottom to top. The wavelengths of the [H$_2$]{} rotational transitions are indicated. \[fig:spectra\_irs\]](fig_irs_spec.ps){width="0.95\columnwidth"} SINFONI integral field spectroscopy {#ss:sinfoni} ----------------------------------- We made use of K-band (1.95–2.45) seeing-limited and adaptive optics assisted near-IR SINFONI observations for several objects of the sample. The observations where carried out during different periods, i.e., 60A (NGC 7469), 77B (NGC 5135 and NGC 7130), 80B (IRASF 05189–2524), and 82B (NGC 34), with a spectral resolution of R$\sim$4000, using different plate scales that yield different FoVs. For further details of the NGC 7469 data, see [@Hicks2009], and for NGC 5135 and NGC 7130, please refer to [@Piqueras2012]. We made use of archive data for IRASF 05189–2524 and NGC 34, observed on December 2007 and October 2008, respectively. The observations were made using the 005$\times$010pixel$^{-1}$ setup, which yields a FoV of $\sim$3$\times$3, and split into individual exposures of 900s following a jittering O-S-O pattern for sky and on-source frames. The total on-source integration times for each galaxy were 5400s for IRASF 05189–2524 and 1800s for NGC 34. To perform the flux calibration and to correct for the instrument response and atmospheric absorption, we used three spectrophotometric standard stars, Hip032193 and Hip052202 for IRASF 05189–2524, and Hip001115 for NGC 34, which were observed with their respective sky frames. The reduction and calibration of the raw data were performed following the same procedure outlined in [@Piqueras2012], i.e., we used the standard ESO pipeline ESOREX (version 2.0.5) to perform the standard corrections of dark subtraction, flat fielding, detector linearity, geometrical distortion, and wavelength calibration. After these corrections were applied to each frame, the sky emission was subtracted from each individual on-source exposure. We constructed individual cubes from each sky-corrected frame that were then flux-calibrated separately, and combined into a single final cube taking the relative shifts in the jittering pattern into account. The absolute flux calibration to translate the counts into physical units is based on the K-band magnitudes from the 2MASS catalog [@Skrutskie2006]. To obtain the integrated K-band spectra, first we fitted the brightest [H$_2$]{} line (1–0 S(1) at 2.12) in every spaxel to produce the intensity and velocity map of the [H$_2$]{} emission. Then we combined the spectra of all the spaxels with signal to noise ratio $>$ 3 in the 1–0 S(1) line, correcting for the relative velocity of each spaxel. The integrated spectra are shown in Figure \[fig:spectra\_sinfoni\]. Finally, we measured the line fluxes and upper limits of the [H$_2$]{} transitions in the integrated spectra by fitting a Gaussian profile with fixed FWHM and position that were first determined from the 1–0 S(1) fit. Fluxes and upper limits are presented in Table \[tab:h2fluxes\]. [lrcccccccccc]{} Transition & & $A_{\rm \lambda}\slash A_{\rm K}$ & &\ \ & & & & NGC 34 & IRASF 05189–2524 & UGC 05101 & NGC 5135 & NGC 7130 & NGC 7469\ [H$_2$]{} 0–0 S(0) & 28.23 & 0.42 & & 32.2 $\pm$ 8.0 & $<$136 & 12.4 $\pm$ 4.1 & 18.0 $\pm$ 2.9 & 21.9 $\pm$ 2.9 & $<$69\ [H$_2$]{} 0–0 S(1) & 17.04 & 0.64 & & 150 $\pm$ 12 & 25.2 $\pm$ 3.6 & 47.0 $\pm$ 7.0 & 220 $\pm$ 11 & 116.1 $\pm$ 9.8 & 182.0 $\pm$ 9.5\ [H$_2$]{} 0–0 S(2) & 12.28 & 0.58 & & 61.7 $\pm$ 6.1 & 13.3 $\pm$ 2.5 & 22.7 $\pm$ 2.8 & 88 $\pm$ 15 & 37.7 $\pm$ 4.4 & 91.7 $\pm$ 9.2\ [H$_2$]{} 0–0 S(3) & 9.67 & 0.99 & & 194.3 $\pm$ 8.7 & 26.3 $\pm$ 3.5 & 20.4 $\pm$ 2.3 & 233.8 $\pm$ 6.1 & 109 $\pm$ 11 & 228.5 $\pm$ 8.2\ [H$_2$]{} 0–0 S(5) & 6.91 & 0.38 & & 87 $\pm$ 21 & $<$100 & 37.6 $\pm$ 3.4 & 66.4 $\pm$ 10.0 & 67 $\pm$ 10 & 91 $\pm$ 10\ \ [H$_2$]{} 1–0 S(0) & 2.22 & 0.95 & & 3.1 $\pm$ 0.2 & $<$1.3 & [  $\cdots$  ]{}& 6.4 $\pm$ 0.2 & 4.3 $\pm$ 0.1 & 1.5 $\pm$ 0.2\ [H$_2$]{} 1–0 S(1) & 2.12 & 1.02 & & 12.4 $\pm$ 1.0 & 5.0 $\pm$ 0.4 & [  $\cdots$  ]{}& 22.2 $\pm$ 1.2 & 14.5 $\pm$ 0.6 & 5.3 $\pm$ 0.5\ [H$_2$]{} 1–0 S(2) & 2.03 & 1.09 & & $<$7.0 & 2.0 $\pm$ 0.2 & [  $\cdots$  ]{}& 8.9 $\pm$ 0.4 & 5.6 $\pm$ 0.2 & 1.4 $\pm$ 0.2\ [H$_2$]{} 1–0 S(3) & 1.96 & 1.16 & & 14.9 $\pm$ 0.5 & 6.9 $\pm$ 0.4 & [  $\cdots$  ]{}& 22.4 $\pm$ 0.7 & 15.8 $\pm$ 0.6 & 5.3 $\pm$ 0.2\ \ [H$_2$]{} 2–1 S(1) & 2.25 & 0.93 & & 1.7 $\pm$ 0.2 & $<$1.3 & [  $\cdots$  ]{}& 3.2 $\pm$ 0.2 & 1.7 $\pm$ 0.1 & $<$0.9\ [H$_2$]{} 2–1 S(2) & 2.15 & 0.99 & & $<$1.9 & $<$0.9 & [  $\cdots$  ]{}& $<$1.6 & $<$1.2 & $<$0.8\ [H$_2$]{} 2–1 S(3) & 2.07 & 1.06 & & $<$2.0 & $<$1.2 & [  $\cdots$  ]{}& 2.7 $\pm$ 0.4 & 1.2 $\pm$ 0.2 & $<$0.8\ [H$_2$]{} 2–1 S(4) & 2.00 & 1.11 & & $<$2.8 & $<$1.1 & [  $\cdots$  ]{}& $<$3.2 & $<$1.7 & $<$1.5\ ------------------ -------------- ------------------ -- -- -- -- -- -- -- -- -- Galaxy $S_{\rm Si}$ $A_{\rm K}$ (mag) NGC34 $-$1.10 [  $\cdots$  ]{} IRASF 05189-2524 $-$0.36 0.79 UGC05101 $-$1.50 3.30 NGC5135 $-$0.39 0.86 NGC7130 $-$0.33 0.73 NGC7469 $-$0.13 0.29 ------------------ -------------- ------------------ -- -- -- -- -- -- -- -- -- : Extinction[]{data-label="tab:extinction"} ![Rest-frame SINFONI integrated spectra. For clarity the spectra are shifted by the following factors 0, 0, 3.5, 4.0, and 4.0 $\times$10$^{-11}$ergcm$^{-2}$s$^{-1}$ from bottom to top. The wavelengths of the H and [H$_2$]{} transitions are indicated. \[fig:spectra\_sinfoni\]](fig_sinfoni_spec.ps){width="0.95\columnwidth"} ------------------ ---------------- ------ ----------------- ------ ------------------ ------------------ -- -- -- -- -- Galaxy 1 Ref. 2 Ref. 3 Ref. 115.271GHz 230.538GHz 345.796GHz NGC34 5.6 $\pm$ 1.1 1,S 10.1 $\pm$ 2.0 1,S [  $\cdots$  ]{} [  $\cdots$  ]{} IRASF 05189-2524 1.8 $\pm$ 0.5 3,A 10.0 $\pm$ 2.6 3,I 29.4 $\pm$ 7.2 3,J UGC05101 1.8 $\pm$ 0.4 4,I 9.7 $\pm$ 2.0 4,I 22.5 $\pm$ 6.4 5,J NGC5135 14.6 $\pm$ 3.5 3,I 95 $\pm$ 21 3,I 225 $\pm$ 56 3,J NGC7130 12.4 $\pm$ 2.5 1,S 44.8 $\pm$ 9.0 1,S [  $\cdots$  ]{} [  $\cdots$  ]{} NGC7469 10.8 $\pm$ 2.6 2,I 70.0 $\pm$ 16.4 2,I 273.8 $\pm$ 65.2 2,J ------------------ ---------------- ------ ----------------- ------ ------------------ ------------------ -- -- -- -- -- Aperture matching {#ss:apertures} ----------------- We combined data from instruments with different FoVs and angular resolutions. The [[SPIRE/FTS]{}]{} beam size varies between 20 and 40, which correspond to 5–35kpc at the distance of our galaxies, so they appear as point-like sources at this resolution. By contrast, the [*Spitzer*]{}/IRS slit width varies between 3.6 and 11.1 (1–10kpc), depending on the IRS module. For the most distant galaxies in our sample, IRASF 05189-2524 and UGC 05101, these apertures encompass their mid-IR emission. Likewise, most of the mid-IR emission of NGC 34 and NGC 7469 is produced in the compact nuclear region (few kpc, @Esquej2012 [@Diaz-Santos2010b]), thus no aperture matching corrections are needed. For the more nearby NGC 5135 and NGC 7130, [*Spitzer*]{}/IRS spectral mapping observations are available and it is possible to extract their mid-IR spectra from a 15$\times$15 aperture (5$\times$5kpc) comparable to the [[SPIRE/FTS]{}]{} beam size and large enough to include most of their bright nuclear mid-IR emission. The SINFONI FoV for the IRASF 05189-2524 observations is 3$\times$3 (2.5$\times$2.5kpc). This source is unresolved in the SINFONI data (FWHM$\sim$0.3), and therefore we expect that most of the flux is included in our measurements. For NGC 7469, the FoV of SINFONI data is 0.8$\times$0.8 (0.2$\times$0.2kpc), so only the nuclear emission is included. NGC 7469 has a kpc scale star-forming ring, thus to account for the missing near-IR H$_2$ emission we added the emission measured in the star-forming ring by @DiazSantos2007 to the nuclear fluxes (Table \[tab:h2fluxes\]). This correction increases the fluxes by a factor of $\sim$2. Finally, the SINFONI FoVs are 8$\times$8 and 8$\times$16, respectively, for NGC 5135 and NGC 7130. This corresponds to 2.5–5kpc which is similar to the source sizes [@Arribas2012], so no correction is applied to the measured integrated SINFONI fluxes. Extinction correction {#ss:extinction} --------------------- Near- and mid-IR emission in luminous and ultra-luminous IR galaxies (LIRGs and ULIRGs) is affected by extinction due to the large amount of dust present in these galaxies. To estimate the extinction, we measured the $S_{\rm Si}$ (see Section \[ss:spitzer\_irs\]), $S_{\rm Si}$ and $A_{\rm K}$ are related according to the extinction law. In particular, using the @Chiar2006 local ISM extinction law we obtained that $A_{\rm K}\slash S_{\rm Si}=-2.20$. The $S_{\rm Si}$ values in our galaxies range from $-1.50$ to $-0.13$ (Table \[tab:extinction\]), which correspond to $A_{\rm K}$ between 0.3 and 3.3mag. Regions producing [H$_2$]{} and $\sim$10 continuum emissions (where the $S_{\rm Si}$ is measured) may have different extinction levels. We cannot use the near-IR lines to test if we can use the estimated $A_{\rm K}$ values to correct for the obscuration effects in the [H$_2$]{} transitions because the differential extinction between the different transitions is small ($A_{\rm \lambda}\slash A_{\rm K}=$0.93–1.16, see Table \[tab:h2fluxes\]). On the contrary, the [H$_2$]{} 0–0 S(3) transition at 9.67 lies close to the peak of the silicate absorption, so its flux is more affected by extinction than the rest of the mid-IR [H$_2$]{} lines. This is clear for UGC 05101, whose 0–0 S(1)/0–0 S(3) ratios are 2.3$\pm$0.4 and 0.8$\pm$0.1 before and after the extinction correction. Actually, all the galaxies (except NGC 34, see below) have very similar 0–0 S(1)/0–0 S(3) ratios, 0.76$\pm$0.06, after the extinction correction based on $S_{\rm Si}$. NGC 34 has a strong silicate absorption ($S_{\rm Si}=-1.10$) and an observed 0–0 S(1)/0–0 S(3) $=0.77\pm0.07$, which is compatible with the average extinction corrected ratio of the other galaxies. Therefore, it is possible that the regions producing the [H$_2$]{} and $\sim$10 continuum emissions in NGC 34 suffer from very different extinction levels. Moreover, the NGC 34 uncorrected [H$_2$]{} SLED does not have any indication of attenuation of the 0–0 S(3) transition (see Figure \[fig:powerlaw\_fit\]). This is at odds with the UGC 05101 ($S_{\rm Si}$=–1.50) SLED that shows a clear deficit of 0–0 S(3) emission. As a consequence, we do not apply any correction to NGC 34 since we do not have an accurate measurement of extinction toward the [H$_2$]{} emitting regions. Radiative transfer models {#s:rad_models} ========================= Previous works have analyzed the [*Herschel*]{} mid-$J$ CO SLEDs of galaxies by fitting non-LTE radiative transfer models (e.g., @Rangwala2011 [@Spinoglio2012; @Pereira2013; @Rigopoulou2013]). In general, a model with two components at different temperatures is used. In these models, the “cold” component (usually $T_{\rm kin}<100$K) reproduces the lowest $J$ CO lines, whereas the “warm” component ($T_{\rm kin}$ several hundred K) accounts for the higher $J$ emission. Similarly, when the mid-IR rotational [H$_2$]{} transitions are analyzed, models with two components are used (e.g., @Rigopoulou02 [@Roussel07]). In these studies, the derived $T_{\rm kin}$ are always higher than $\sim$100K since at lower temperatures the rotational [H$_2$]{} emission is negligible. In addition, it is assumed that [H$_2$]{} is in LTE because the critical density of these lines is low ($n_{\rm H_2}< 10^4$cm$^{-3}$, see Figure 3 of @Roussel07). For the near-IR ro-vibrational [H$_2$]{} lines, the molecular gas is usually assumed LTE too, although the critical densities of these transitions are higher ($n_{\rm crit}>$10$^{7}$cm$^{-3}$). The derived temperatures from these lines are around 2000K. In general, it is concluded that they are collisionally excited and UV excitation has relatively low importance for the excitation of the [H$_2$]{} $\nu=1-0$ transitions (e.g., @Davies2003 [@Bedregal2009]). In this work, we analyze the SLED of low- and mid-$J$ CO transitions as well as pure rotational and ro-vibrational [H$_2$]{} transitions. Therefore, to explain the emission of all these lines, our model would need at least four components with different temperatures ranging from $\sim$50K to 2000K. Instead, we use an alternative approach assuming a continuous distribution of temperatures. This is advantageous since we avoid using a somewhat arbitrary number of model components to fit the data. It is common to find that the observed SLEDs have a positive curvature, that is, the rotational temperature derived from a pair of consecutive transitions increases with the energy of their upper levels (e.g., @Neufeld2012). Therefore, emission could come from gas with a continuous distribution of temperatures instead of several discrete temperatures. To model this distribution of temperatures we assumed that the column density follows a power-law (d$N\sim T^{-\beta}$d$T$). A similar method has been used to model the ro-vibrational and rotational [H$_2$]{} emission in shocked gas (e.g., @Brand1988 [@Neufeld2008; @Shinn2009]), or the CO emission (e.g., @Goicoechea2013). In the following, we describe our model in detail. First, we explain how the models for single temperature and density are obtained, and then how they are combined to reproduce a power-law temperature distribution. Single temperature non-LTE model {#ss:single_model} -------------------------------- To solve the radiative transfer equations for the molecular emission we used the code RADEX [@vanderTak2007]. This code uses the escape probability approximation to obtain the molecular level populations and the intensities of the emission lines. We constructed two sets of models, one for CO and other for [H$_2$]{}, covering a wide range of kinetic temperature ($T_{\rm kin}=10-2800$K) and [H$_2$]{} density ($n_{\rm H_2} = 10^2-10^9$cm$^{-3}$). For the [H$_2$]{} grid, we also consider collisions with atomic H with $n_{\rm H}/n_{\rm H_2}$ ratios between 1 and 10$^{-5}$. In these grids, we assume the LTE [H$_2$]{} ortho-to-para ratio, which varies between 0 for low temperatures and 3 for $T_{\rm kin}>200$K (see Figure 4 of @Burton1992). The [H$_2$]{} transitions are optically thin for $N_{\rm H_2}\slash \Delta v$ lower than $\sim$10$^{24}$cm$^{-2}$()$^{-1}$, thus in our models we assume optically thin [H$_2$]{} emission[^4]. Similarly, we assume optically thin mid-$J$ CO emission, this condition is fulfilled when $N_{\rm CO}\slash \Delta v$ is lower than $\sim$10$^{15}$cm$^{-2}$()$^{-1}$. In general, this is the case for the mid-$J$ CO observations of galaxies (e.g., @Kamenetzky2012 [@Pereira2013]). For the CO grid, we used the collisional rate coefficients of @Yang2010 expanded by @Neufeld2012 for high-$J$ levels ($J=41$ to 80). This is important for the high temperature models as a non negligible fraction of the CO molecules are in energy levels with $J>40$ that otherwise are ignored. We created two sub-grids of [H$_2$]{} models. One for high-temperature ($T_{\rm kin}> 100$K) and another for low temperature. For the high-temperature sub-grid we used the [H$_2$]{}-[H$_2$]{} collisional rate coefficients of @LeBourlot1999 and the [H$_2$]{}-H coefficients of @Wrathmall2007 that cover kinetic temperatures between 100 and 6000K. For the lower temperature sub-grid ($T_{\rm kin}< 100$K) we used the [H$_2$]{}-[H$_2$]{} collisional coefficients of @Lee2008. They only include the lowest nine rotational energy levels of [H$_2$]{} ($E_{\rm upper}<5900$K), but at $T_{\rm kin}<$100K, no molecules are expected at higher rotational energy levels. Collisions with atomic H are mostly important for vibrational excited [H$_2$]{} levels, so they are neglected for the low-temperature sub-grid. The grids for ortho-[H$_2$]{} and para-[H$_2$]{} were constructed separately, although they were later combined assuming the LTE [H$_2$]{} ortho-to-para ratio. Power-law temperature distribution {#ss:power_law} ---------------------------------- We want to obtain the level populations for gas following a power-law temperature distribution, d$N\sim T^{-\beta}$d$T$. From the grids described in Section \[ss:single\_model\], we can obtain the fractional population of the molecular level $i$, $n_i\left(T, n_{\rm H_2}\right)$, for a given kinetic temperature and [H$_2$]{} density. Therefore, the total column density of molecules in the energy level $i$ for a kinetic temperature distribution following a power-law between $T_0$ and $T_1$ is $$N_i\left(\beta, n_{\rm H_2}\right) = A \int_{T_ 0}^{T_1}{T^{-\beta}n_i\left(T, n_{\rm H_2}\right){\rm d}T},$$ where $A=N (\beta - 1)\slash (T_0^{1 - \beta} - T_1^{1 - \beta})$ is a normalization constant and $N$ is total molecular gas column density[^5]. In our models, we use $T_0=20$K and $T_1=3500$K. This temperature range is required to model the lower-$J$ CO and the near-IR [H$_2$]{} transitions, respectively. In Section \[s:sled\_fit\], we show that it is not possible to fit simultaneously the SLEDs of both CO and [H$_2$]{} using a single power-law temperature distribution. We find that the temperature distribution is much steeper for [H$_2$]{} ($\beta\sim 5$) than for CO ($\beta\sim 3$). Therefore, for the SLED fitting we used a broken power-law model with two exponents ($\beta_1$ and $\beta_2$). This can be modeled with the following equation: $$\begin{aligned} N_i\left(\beta, n_{\rm H_2}\right) & = B \left( \int_{T_0}^{T_{\rm b}}{T^{-\beta_1}n_i\left(T, n_{\rm H_2}\right){\rm d}T} \right. \nonumber \\ & + \left. T_{\rm b}^{\beta_2 - \beta_1}\int_{T_{\rm b}}^{T_1}{T^{-\beta_2}n_i\left(T, n_{\rm H_2}\right){\rm d}T} \right)\end{aligned}$$ where $T_{\rm b}$ is the break temperature, and $B= \scriptstyle{N\left((T_{\rm b}^{1-\beta_1} - T_0^{1-\beta_1})\slash (1 - \beta_1) + T_{\rm b}^{\beta_2 - \beta_1}(T_1^{1-\beta_2} - T_{\rm b}^{1-\beta_2})\slash (1 - \beta_2)\right)^{-1}}$. CO and [H$_2$]{} SLED fitting {#s:sled_fit} ============================= To fit the SLEDs, we converted the observed fluxes into beam averaged column densities using the following relation $$\langle N_i \rangle = \frac{4\pi}{\Omega} \frac{F_{ij}}{h \nu_{ij} A_{ij}}$$ where $F_{ij}$ is the flux of the transition between levels $i$ and $j$, $\nu_{ij}$ and $A_{ij}$ are its frequency and Einstein coefficient, respectively, $h$ is the Planck constant, and $\Omega$ is the beam solid angle. Since our sources are barely resolved by [*Herschel*]{} we assume a beam FWHM of 18, which is the SSW beam size. This beam size is also used for the rest of measurements as they should include most of the emission in this area (see Section \[ss:apertures\] for details). Using this relation, we assume that the emission is optically thin, which is reasonable for the [H$_2$]{} transitions as well as for the mid-$J$ CO lines (see Section \[s:rad\_models\]). First, we attempted to fit the SLEDs using a power-law temperature distribution. The free parameters of this model are the CO and [H$_2$]{} column densities, $n_{\rm H_2}$ and $n_{\rm H}$ densities, and $\beta$. However, it was not possible to obtain a good fit to both the CO and [H$_2$]{} emissions. Nevertheless, the CO and [H$_2$]{} SLEDs individually are well reproduced by a power-law model, although each SLED has different $\beta$ values. In these models, there is a strong degeneracy between $\beta$ and $n_{\rm H_2}$, $\beta$ decreases for decreasing $n_{\rm H_2}$ when densities are below the critical density of the transitions ($n_{\rm H_2}<10^6$cm$^{-3}$ for CO and $n_{\rm H_2}<10^4$cm$^{-3}$ for [H$_2$]{}). This is because changes in $n_{\rm H_2}$ and $\beta$ produce similar variations in the SLED shape when the gas is not thermalized. In Figure \[fig:chi\_powerlaw\], we plot the $\chi^2$ values for one of the galaxies where this degeneracy is evident. In addition, it is clear that the CO and [H$_2$]{} SLEDs are not simultaneously reproduced by any pair of $\beta$ and $n_{\rm H_2}$ parameters. For our galaxies, we find that $\beta$ is always higher for [H$_2$]{} (steeper temperature distribution) for any given $n_{\rm H_2}$. The $\beta$ values range between 2.5 and 3.5 for CO and between 4.5 and 5.5 for [H$_2$]{} for $n_{\rm H_2}$ higher than the critical densities of the transitions. ![$\chi^2$ values for the power-law temperature distribution model as a function of $\beta$ and $n_{\rm H_2}$ for NGC 7130. The red (green) contours are the 1, 2, and 3$\sigma$ confidence levels for the CO ([H$_2$]{}) SLED.\[fig:chi\_powerlaw\]](chi_NGC7130.ps){width="0.85\columnwidth"} As anticipated in Section \[ss:power\_law\], to account for the difference on the $\beta$ values for the CO and [H$_2$]{} SLEDs, we used an alternative model consisting of a broken power-law, which has two exponent values, $\beta_1$ and $\beta_2$, for temperatures below and above the threshold temperature $T_{\rm b}$, respectively. With this model, it is possible to fit all the available CO and [H$_2$]{} transitions for our galaxies, with the exception of the 1 transition, which is always underpredicted by the model. The later can be caused by the presence of low temperature molecular gas that would not follow the power-law distribution. Consequently, we included a cold gas component at 15K assuming optically thin emission for the 1 transition. In general, 1 emission is not optically thin, so the obtained column densities are lower limits. The value of the breaking temperature ($T_{\rm b}$) is not well constrained because the upper level energies of the CO and [H$_2$]{} transitions do not overlap. When $T_{\rm b}$ is left as a free parameter it varies between 100 and 500K with 1$\sigma$ uncertainties $\sim$100K, and its variation affects the $\beta$ values slightly. The average $T_{\rm b}$ is 200$\pm$130K. It is not possible to accurately determine $T_{\rm b}$ with our data, therefore, we fixed it to the average value. As discussed above, $\beta_1$ and $n_{\rm H_2}$ are strongly correlated, specially for CO, when $n_{\rm H_2}$ is below the critical densities. The CO column density ($N_{\rm CO}$) is relatively constant when $\beta_1$ and $n_{\rm H_2}$ vary. But, on the contrary, the [H$_2$]{} column density variation is large. This is because we only measure the warm [H$_2$]{} column density directly, which represents a small fraction of the total [H$_2$]{}, and the total [H$_2$]{} column density is extrapolated using the $\beta_1$ exponent. Therefore, in our model, lower $n_{\rm H_2}$ also implies lower [H$_2$]{} column densities (and higher CO abundances) because of the smaller value of $\beta_1$. We used this to estimate a lower limit for $n_{\rm H_2}$ given that the CO abundance should be lower or equal to that of C. Assuming the solar atomic C abundance (2.4$\times$10$^{-4}$; @Lodders2003), the lower limit for $n_{\rm H_2}$ is $\sim$10$^{4.0 \pm 0.5}$cm$^{-3}$, depending on the galaxy. In the case of sub-solar abundances in these galaxies, the $n_{\rm H_2}$ upper limit would be slightly higher. Similarly, we can obtain a lower limit for the CO abundance by setting $n_{\rm H_2}$ to a value above the CO critical density (e.g., 10$^{6}$cm$^{-3}$). The minimum CO abundance ranges between 10$^{-5}$ and 10$^{-6}$. In Table \[tab:mol\_gas\], we list the best-fit parameters for $n_{\rm H_2}=10^{4.5}$cm$^{-3}$ and $n_{\rm H_2}=10^{6}$cm$^{-3}$. We plot the best-fit models in Figure \[fig:powerlaw\_fit\]. The $n_{\rm H}$ value depends mostly on the ratio between the intensities of the rotational and near-IR ro-vibrational [H$_2$]{} lines. For our galaxies we obtain $n_{\rm H}$ values between 10$^{3}$–10$^{4}$cm$^{-3}$. The critical densities for collisions with H of the near-IR [H$_2$]{} transitions are $n_{\rm H}\sim 10^{4}-10^{5}$cm$^{-3}$ at 1500K, much lower than the critical densities for collisions with [H$_2$]{}. Therefore, a relatively small abundance of atomic H can enhance the luminosity of the near-IR transitions. For the assumed $n_{\rm H_2}=10^{4.5}$ and 10$^6$cm$^{-3}$ the ro-vibrational near-IR emission would be one or two orders of magnitude lower than the observed if we did not include collisions with H in our model. Cold-to-warm molecular gas ratio {#s:cold-to-warm_ratio} ================================ To estimate the cold molecular gas mass, we used the CO to [H$_2$]{} conversion factor from @Leroy2011, $N_{\rm c}({\rm H_2})$(cm$^{-2}$) = $2\times 10^{20} I_{{\rm CO} J =1-0}$(K). We compare this cold molecular gas column density with the warm molecular gas column density obtained from the fit of the CO and [H$_2$]{} SLEDs (Table \[tab:mol\_gas\]). The warm molecular gas fraction ranges between 1 and 20% for the models with $n_{\rm H_2}=10^{4.5}$cm$^{-3}$ and between 15 and $>100$% for the models with $n_{\rm H_2}=10^{6}$cm$^{-3}$. This suggests that for the two galaxies where the warm column density exceeds the cold column density (UGC 05101 and NGC 5135) the model with $n_{\rm H_2}=10^{6}$cm$^{-3}$ overestimates the $N_{\rm w}({\rm H_2})$. Thus, in the following we use the $n_{\rm H_2}=10^{4.5}$cm$^{-3}$ model for these two galaxies. Moreover, these are the only two sources where the $\chi^2$ is lower for the $n_{\rm H_2}=10^{4.5}$cm$^{-3}$ model. For our sample of IR bright galaxies it might be adequate to use the lower CO to [H$_2$]{} conversion factor $N_{\rm H_2}$(cm$^{-2}$) = $0.5\times 10^{20} I_{{\rm CO} J =1-0}$(K) measured by @Downes1998 in active galaxies. Therefore, the cold-to-warm ratio could be lower by a factor of 4. Molecular gas heating {#s:heating} ===================== In the previous sections, we showed that the CO and [H$_2$]{} SLEDs are compatible with a broken power-law temperature distribution, however, we did not discuss how the molecular gas is heated. Three main mechanisms are invoked to explain the heating of the molecular gas: UV radiation, shocks, and X-ray and cosmic rays. In this section, we compare the molecular gas temperature distribution expected in photodissociation region (PDRs; UV radiation) and shocks regions with the observed temperature distribution. X-ray and cosmic ray heating can be important for AGNs, however, in this work we analyze active galaxies with strong star-formation (see Table \[tab:sample\]). The AGN contribution to the mid-$J$ CO emission seems to be small compared to that of SF (e.g., @Pereira2013). Likewise, SF dominates the [H$_2$]{} emission in LIRGs (e.g., @Roussel07 [@Pereira2010]). For these reasons, we consider that X-ray heating should not contribute significantly in the integrated spectra of these galaxies. {width="1.8\columnwidth"} UV radiation: PDRs {#ss:cloudy_pdr} ------------------ We used the radiative transfer code <span style="font-variant:small-caps;">Cloudy</span> version 13.02 [@Ferland2013] to estimate the gas temperature distribution of PDRs. We used a constant density slab model illuminated by the interstellar radiation field of @Black1987b. We created a grid of models with $n_{\rm H}$ between $10^4$ and $10^6$cm$^{-3}$ and radiation field intensity between $10^1$ and $10^4$$G_{\rm 0}$, where $G_{\rm 0}=1.6\times 10^{-3}$ergcm$^{-2}$s$^{-1}$ is the intensity integrated between 6 and 13.6eV. Since we are interested only in the PDR region we extinguished the input spectra to remove the ionizing radiation. We extracted the CO and [H$_2$]{} abundances and kinetic temperatures as a function of the column density from the output of the models, which we used to compute the temperature distribution of CO and [H$_2$]{}. Then we fitted these distributions with a power-law function. In the fit we only included gas with $T>20$K and $T>150$K, for CO and [H$_2$]{}, respectively, because lower temperature gas does not produce significant emission of the observed transitions. In Figure \[fig:pdr\_cloudy\], we show an example <span style="font-variant:small-caps;">Cloudy</span> PDR model and the fit to the temperature distribution. We find that, in general, [H$_2$]{} distributions are steeper than those of CO. The power-law index[^6] for CO is $\sim$3–10 for $n_{\rm H_2}=10^{4.5}$cm$^{-3}$, and $\sim$5–7 for $n_{\rm H_2}=10^{6}$cm$^{-3}$, and for [H$_2$]{} $\beta$ ranges between $\sim$10 and 15. That is, these $\beta$ values are higher than the values derived from our fits to the observed SLEDs (Section \[s:sled\_fit\]). Shocks {#ss:shocks} ------ To calculate the molecular gas temperature distribution in shocks we used a simple model. We assumed that molecular gas is heated at a constant rate up to a certain temperature, which depends on the shock velocity and is high enough to produce [H$_2$]{} emission, and after that it is cooled radiatively by the dominant cooling molecule. The main coolant of warm molecular gas is [H$_2$]{}, CO, or H$_2$O depending on the density and temperature of the gas [@Neufeld1993]. For our galaxies, the H$_2$O transitions are weaker than the CO transitions for those with similar $E_{\rm up}$ values. That is, H$_2$O cooling does not seem to dominate the cooling for the average conditions in these galaxies, and therefore, we do not consider H$_2$O in our model. First, we calculated the non-LTE cooling functions of CO and [H$_2$]{} as a function of the gas density and kinetic temperature using the grid of models described in Section \[ss:single\_model\]. For the [H$_2$]{} cooling function we do not include the effect of collisional excitation by H since it would only affect the cooling of high-temperature gas through ro-vibrational emission of [H$_2$]{}. The temperature dependence of these cooling functions can be approximated with a power-law (e.g., @Draine1993). Thus we fitted the calculated cooling functions using a power-law in the temperature range between 30 and 150K for CO, and 150 and 1000K for [H$_2$]{}. These are the temperature ranges where CO and [H$_2$]{} dominate the gas cooling [@Neufeld1993]. ![Power-law indices of the temperature dependence of the CO (top) and [H$_2$]{} (bottom) non-LTE cooling functions (solid line; see Section \[ss:shocks\]). The fits are valid for a temperature range between 30 and 150K for CO and 150 and 1000K for [H$_2$]{}. The fitted $\beta$ values form the models with $n_{\rm H_2}=10^{4.5}$cm$^{-3}$ and $n_{\rm H_2}=10^6$cm$^{-3}$ are also plotted for each galaxy. For visualization the abscissa of the data points are slightly offset. \[fig:shocks\_cooling\]](cooling_gamma.ps){width="0.9\columnwidth"} Then, assuming that the molecular gas cools radiatively after being heated according to these fitted cooling functions (i.e, d$U\slash$ d$t \sim -T^\gamma$, where $U=5\slash 2\,kT$ is the internal kinetic energy), the observed temperature distribution of the molecular gas is d$N\slash$d$T \sim T^{-\gamma}$. That is, the $\beta$ we fitted for the SLEDs should be compared with $\gamma$. The power-law indices, $\gamma$, for CO and [H$_2$]{} as a function of the density are shown in Figure \[fig:shocks\_cooling\]. Comparison with the observations -------------------------------- In our fits, we obtained that the power-law index varies between 1.2 and 3.5 for CO and between 4.3 and 4.9 for [H$_2$]{}. In general, these values are lower than those expected in PDR, but higher than those expected from shocks (see Table \[tab:mol\_gas\] and Figure \[fig:shocks\_cooling\]). Therefore, the combination of PDRs and shocks are needed to explain the intermediate power-law indices. The CO emission is compatible with that predicted by shock models only for NGC 5135[^7], whose best-fit model has $n_{\rm H_2}=10^{4.5}$cm$^{-3}$. This result, however, strongly depends on the adopted density. For instance, for the model with $n_{\rm H_2}=10^{6}$cm$^{-3}$, which provides only a slightly worse fit to the CO SLED of NGC 5135, the CO emission would need a PDR component. If we combine these PDR and shock models, the PDR SLEDs would dominate the lower-$J$ CO transitions because of their steeper power-law index, while the contribution from shocks would be dominant for higher-$J$ CO transitions. Moreover, only gas with $T>100$K produces [H$_2$]{} emission so, depending on the $G_0$ values of the PDRs and the shock velocities, the contributions of PDRs and shocks may differ in the CO and [H$_2$]{} SLEDs. Our sample includes three spiral galaxies (NGC 5135, NGC 7130, and NGC 7469) and three major mergers (NGC 34, IRASF 05189–2524, and UGC 05101). The latter have lower $\beta_1$ values for their CO SLEDs (Figure \[fig:shocks\_cooling\]) suggesting that the relative contribution from shocks to the CO emission in these mergers is larger than in the spirals. For the [H$_2$]{} emission only IRASF 05189–2524 has a $\beta_2$ value lower than the rest of galaxies, so in this source, shocks able to heat the molecular gas up to $T>100$K could be important. Also, this galaxy is the most luminous in our sample and the galaxy with the higher AGN contribution (Table \[tab:sample\]), so the bright AGN might affect its CO and [H$_2$]{} emissions. Comparison with two-component models ------------------------------------ The CO SLEDs of two of our galaxies (UGC 05101 and NGC 7130) were presented by @Pereira2013. In that paper, a model consisting of two components was used to fit the data: a cold component that contributed mainly to the 1 emission, and a warm component with variable [H$_2$]{} density and kinetic temperature. Once corrected for the different beam sizes used, the beam-averaged CO column densities derived by both methods agree within the uncertainties, although, this could be a particular case due to the specific SLEDs of these two galaxies. The interpretation of the [$n_{\rm H_2}$]{} and [$T_{\rm kin}$]{} of the two-component model (TCM) is not so straightforward. The high kinetic temperature needed by the TCM reveals the presence of warm molecular gas, but the exact temperature value is not necessarily a representative temperature of the molecular gas. The obtained [$n_{\rm H_2}$]{} and [$T_{\rm kin}$]{} are biased toward the conditions in the most luminous regions. Actually, the luminosity of each CO transition, for a given molecular mass, depends on [$n_{\rm H_2}$]{} and [$T_{\rm kin}$]{}. So, the derived [$n_{\rm H_2}$]{} and [$T_{\rm kin}$]{} will depend on the unknown distribution of densities and temperatures of the molecular gas. This problem is partially solved assuming a functional form for the d$N\slash$d$T$, as we adopted in our models. In addition, [$n_{\rm H_2}$]{} and [$T_{\rm kin}$]{} are correlated (lower [$n_{\rm H_2}$]{} and higher [$T_{\rm kin}$]{}, and higher [$n_{\rm H_2}$]{} and lower [$T_{\rm kin}$]{} produce similar CO SLEDs) and this adds an extra uncertainty to the derived values. This is equivalent to the uncertainty due to the $\beta$-[$n_{\rm H_2}$]{} correlation for our models (see Section \[s:sled\_fit\]). Although the TCM provides useful information, when the analyzed molecular transitions comprise a wide range of excitation temperatures, the need for additional components for the SLED fitting can make the modeling contrived. Instead, the power-law models used here could better represent the integrated temperature distribution of the molecular gas in a galaxy. Conclusions {#s:conclusions} =========== We have studied the integrated CO and [H$_2$]{} emission of six local IR bright galaxies using non-LTE models. Assuming a broken power-law distribution for the molecular gas temperatures, our model reproduces both the CO SLED (from $J_{\rm up}=1$ to $J_{\rm up}=13$) and the [H$_2$]{} SLED ($J_{\rm up}\leq 7$ for the lowest three vibrational levels) in our sample of galaxies. The main findings of this work are summarized in the following: 1. With a single power-law temperature distribution it is not possible to fit simultaneously the CO and [H$_2$]{} SLEDs. The [H$_2$]{} SLEDs have a much steeper power-law index ($\beta_2\sim 4-5$) than the CO SLED ($\beta_1\sim 1-3$). This is the expected behavior for the temperature distributions in PDRs and shocks. 2. We found that for most of the galaxies, the models with $n_{\rm H_2}=10^6$cm$^{-3}$ provide the best fit to the observed data, thus the majority of the transitions are close to LTE. The minimum acceptable density for the warm gas is $\sim n_{\rm H_2}=10^{4 \pm 0.5}$cm$^{-3}$, lower densities would imply CO abundances higher than the atomic C abundance. Likewise, we obtained that the minimum CO abundance in the warm gas is $x_{\rm CO}\sim 10^{-6}-10^{-5}$ assuming LTE conditions. 3. The column densities of the warm molecular gas ($T>20$K) represents between 10 and 100% of the molecular gas traced by the 1 transition. 4. We used PDR and shock models to determine the excitation mechanism of the molecular gas. Our models show that the temperature distributions are steeper for PDRs than for shocks. We also found that the temperature distribution of the warmest gas ($T>100$K) emitting in [H$_2$]{} is steeper than that of coldest gas ($T>30$K), which produces the mid-$J$ CO emission for both PDR and shocks models. This is because of the different main coolant of the warm and cold molecular gas ([H$_2$]{}, and CO, respectively). 5. Neither PDR nor shocks alone can explain the derived temperature distributions. A combination of both is needed to reproduce the observed SLEDs. In this case, the lower-$J$ CO transitions would be dominated by PDR emission, whereas the higher-$J$ CO transitions would be dominated by shocks. 6. The three major mergers among our targets (NGC 34, IRASF 05189–2524, and UGC 05101) have shallower temperature distributions for CO than the other three spirals. This suggests that the relative contribution of shocks to the heating of warm molecular gas ($T<100$K) in these major mergers is higher than in the other three spirals. 7. For only one of the mergers, IRASF 05189–2524, the shallower [H$_2$]{} temperature distribution (hot molecular gas) suggests that the relative importance of shocks is high. This galaxy also has a bright AGN that dominates the bolometric luminosity, which can contribute to the molecular gas heating. For the other two mergers, the [H$_2$]{} temperature distribution is similar to that of the spiral galaxies. Therefore the shocks producing the extra contribution to the CO emission in these mergers are not able to heat the molecular gas to temperatures higher than 100K, which would be necessary to see the [H$_2$]{} emission. We thank anonymous referee for comments that improved the paper. This work was funded by the Agenzia Spaziale Italiana (ASI) under contract I/005/11/0. JPL acknowledge support from the Spanish Plan Nacional de Astronomía y Astrofísica AYA2010-21161-C02-01 and AYA2012-32295. SPIRE has been developed by a consortium of institutes led by Cardiff Univ. (UK) and including: Univ. Lethbridge (Canada); NAOC (China); CEA, LAM (France); IFSI, Univ. Padua (Italy); IAC (Spain); Stockholm Observatory (Sweden); Imperial College London, RAL, UCL-MSSL, UKATC, Univ. Sussex (UK); and Caltech, JPL, NHSC, Univ. Colorado (USA). This development has been supported by national funding agencies: CSA (Canada); NAOC (China); CEA, CNES, CNRS (France); ASI (Italy); MCINN (Spain); SNSB (Sweden); STFC, UKSA (UK); and NASA (USA). This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. [63]{} natexlab\#1[\#1]{} , M., [Kr[ü]{}gel]{}, E., & [Chini]{}, R. 2007, , 462, 575 , A., [Garc[í]{}a-Mar[í]{}n]{}, M., [Monreal-Ibero]{}, A., [et al.]{} 2009, , 506, 1541 , A., [Pereira-Santaella]{}, M., [Rieke]{}, G. H., & [Rigopoulou]{}, D. 2012, , 744, 2 , L., [Charmandaris]{}, V., [Spoon]{}, H. W. W., [et al.]{} 2004, , 154, 178 , S., [Colina]{}, L., [Alonso-Herrero]{}, A., [et al.]{} 2012, , 541, A20 , A. G., [Colina]{}, L., [Alonso-Herrero]{}, A., & [Arribas]{}, S. 2009, , 698, 1852 , E., [Arribas]{}, S., & [Colina]{}, L. 2012, , 542, A54 , J. H. 1987, in Astrophysics and Space Science Library, Vol. 134, Interstellar Processes, ed. D. J. [Hollenbach]{} & H. A. [Thronson]{}, Jr., 731–744 , P. W. J. L., [Moorhouse]{}, A., [Burton]{}, M. G., [et al.]{} 1988, , 334, L103 , M. G., [Hollenbach]{}, D. J., & [Tielens]{}, A. G. G. 1992, , 399, 563 , J. E. & [Tielens]{}, A. G. G. M. 2006, , 637, 774 , R. I., [Sternberg]{}, A., [Lehnert]{}, M., & [Tacconi-Garman]{}, L. E. 2003, , 597, 907 , T., [Alonso-Herrero]{}, A., [Colina]{}, L., [Ryder]{}, S. D., & [Knapen]{}, J. H. 2007, , 661, 149 , T., [Charmandaris]{}, V., [Armus]{}, L., [et al.]{} 2010, , 723, 993 , D. & [Solomon]{}, P. M. 1998, , 507, 615 , B. T. & [McKee]{}, C. F. 1993, , 31, 373 , F., [Abuter]{}, R., [Bickert]{}, K., [et al.]{} 2003, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4841, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. M. [Iye]{} & A. F. M. [Moorwood]{}, 1548–1561 , P., [Alonso-Herrero]{}, A., [P[é]{}rez-Garc[í]{}a]{}, A. M., [et al.]{} 2012, , 423, 185 , G. J., [Porter]{}, R. L., [van Hoof]{}, P. A. M., [et al.]{} 2013, , 49, 137 , R., [Weitzel]{}, L., [Tacconi-Garman]{}, L. E., [et al.]{} 1995, , 444, 129 , J. R., [Etxaluze]{}, M., [Cernicharo]{}, J., [et al.]{} 2013, , 769, L13 , M. J., [Abergel]{}, A., [Abreu]{}, A., [et al.]{} 2010, , 518, L3 , E. K. S., [Davies]{}, R. I., [Malkan]{}, M. A., [et al.]{} 2009, , 696, 448 , J. R., [Roellig]{}, T. L., [van Cleve]{}, J., [et al.]{} 2004, , 154, 18 , F. P. 2009, , 493, 525 , J., [Glenn]{}, J., [Rangwala]{}, N., [et al.]{} 2012, , 753, 70 , M. F., [Steinz]{}, J. A., [Anderegg]{}, M. E., [et al.]{} 1996, , 315, L27 , J., [Pineau des For[ê]{}ts]{}, G., & [Flower]{}, D. R. 1999, , 305, 802 , T.-G., [Balakrishnan]{}, N., [Forrey]{}, R. C., [et al.]{} 2008, , 689, 1105 , A. K., [Bolatto]{}, A., [Gordon]{}, K., [et al.]{} 2011, , 737, 12 , K. 2003, , 591, 1220 , D. A., [Baluteau]{}, J.-P., [Barlow]{}, M. J., [et al.]{} 2010, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 7731, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series , D. A. 2012, , 749, 125 , D. A. & [Kaufman]{}, M. J. 1993, , 418, 263 , D. A. & [Yuan]{}, Y. 2008, , 678, 974 , S. 2010, in Astronomical Society of the Pacific Conference Series, Vol. 434, Astronomical Data Analysis Software and Systems XIX, ed. Y. [Mizumoto]{}, K.-I. [Morita]{}, & M. [Ohishi]{}, 139 , P. P., [van der Werf]{}, P. P., [Xilouris]{}, E. M., [et al.]{} 2012, , 426, 2601 , M., [Alonso-Herrero]{}, A., [Rieke]{}, G. H., [et al.]{} 2010, , 188, 447 , M., [Alonso-Herrero]{}, A., [Santos-Lleo]{}, M., [et al.]{} 2011, , 535, A93 , M., [Spinoglio]{}, L., [Busquet]{}, G., [et al.]{} 2013, , 768, 55 , G. L., [Riedinger]{}, J. R., [Passvogel]{}, T., [et al.]{} 2010, , 518, L1 , J., [Colina]{}, L., [Arribas]{}, S., [Alonso-Herrero]{}, A., & [Bedregal]{}, A. G. 2012, , 546, A64 , N., [Maloney]{}, P. R., [Glenn]{}, J., [et al.]{} 2011, , 743, 94 , D., [Hurley]{}, P. D., [Swinyard]{}, B. M., [et al.]{} 2013, , D., [Kunze]{}, D., [Lutz]{}, D., [Genzel]{}, R., & [Moorwood]{}, A. F. M. 2002, , 389, 374 , H., [Helou]{}, G., [Hollenbach]{}, D. J., [et al.]{} 2007, , 669, 959 , D. B., [Mazzarella]{}, J. M., [Kim]{}, D.-C., [Surace]{}, J. A., & [Soifer]{}, B. T. 2003, , 126, 1607 , J.-H., [Koo]{}, B.-C., [Burton]{}, M. G., [Lee]{}, H.-G., & [Moon]{}, D.-S. 2009, , 693, 1883 , M. F., [Cutri]{}, R. M., [Stiening]{}, R., [et al.]{} 2006, , 131, 1163 , J. D. T., [Armus]{}, L., [Dale]{}, D. A., [et al.]{} 2007, , 119, 1133 , J. D. T., [Draine]{}, B. T., [Dale]{}, D. A., [et al.]{} 2007, , 656, 770 , L., [Pereira-Santaella]{}, M., [Busquet]{}, G., [et al.]{} 2012, , 758, 108 , B. M., [Ade]{}, P., [Baluteau]{}, J.-P., [et al.]{} 2010, , 518, L4 , S., [Spinoglio]{}, L., [Malkan]{}, M. A., & [Fazio]{}, G. 2010, , 709, 1257 , F. F. S., [Black]{}, J. H., [Sch[ö]{}ier]{}, F. L., [Jansen]{}, D. J., & [van Dishoeck]{}, E. F. 2007, , 468, 627 , P. P., [Isaak]{}, K. G., [Meijerink]{}, R., [et al.]{} 2010, , 518, L42 , S., [Rupke]{}, D. S. N., [Kim]{}, D.-C., [et al.]{} 2009, , 182, 628 , M. W., [Roellig]{}, T. L., [Low]{}, F. J., [et al.]{} 2004, , 154, 1 , C. D., [Petitpas]{}, G. R., [Iono]{}, D., [et al.]{} 2008, , 178, 189 , S. A., [Gusdorf]{}, A., & [Flower]{}, D. R. 2007, , 382, 133 , Y., [Charmandaris]{}, V., [Huang]{}, J., [Spinoglio]{}, L., & [Tommasin]{}, S. 2009, , 701, 658 , B., [Stancil]{}, P. C., [Balakrishnan]{}, N., & [Forrey]{}, R. C. 2010, , 718, 1062 , T.-T., [Kewley]{}, L. J., & [Sanders]{}, D. B. 2010, , 709, 884 [^1]: [*Herschel*]{} is an ESA space observatory with science instruments provided by European-led Principal Investigator consortia and with important participation from NASA. [^2]: The sinc function accurately models the FTS instrumental line shape. [^3]: The [H$_2$]{} S(3) transition at 9.67 also lies in the high-resolution range for the higher redshift galaxies UGC 05101 and IRASF 05189-2524. [^4]: $\Delta v\sim 100$in our galaxies. [^5]: By definition $\sum_i{n_i\left(T, n_{\rm H_2}\right)} = 1$, and $\sum_i{N_i\left(\beta, n_{\rm H_2}\right)}=N$, then $N = A\int_{T_0}^{T_1}{T^{-\beta}{\rm d}T}$ and $A$ can be solved. [^6]: This index refers to the $\beta$ of the relation d$N\slash$d$T\sim T^{-\beta}$. [^7]: We do not consider UGC 05101 because its $\beta_1$ uncertainty, 2.4, is very large.

ArXiv

--- abstract: | We report on Cyg X-1 observations performed by the SPI telescope onboard the INTEGRAL mission and distributed over more than 6 years. We investigate the variability of the intensity and spectral shape of this peculiar source in the hard X-rays domain, and more particularly up to the MeV region. We first study the total averaged spectrum which presents the best signal to noise ratio (4 Ms of data). Then, we refine our results by building mean spectra by periods and gathering those of similar hardness. Several spectral shapes are observed with important changes in the curvature between 20 and 200 keV, even at the same luminosity level. In all cases, the emission decreases sharply above 700 keV, with flux values above 1 MeV (or upper limits) well below the recently reported polarised flux (Laurent et al. 2011), while compatible with the MeV emission detected some years ago by CGRO/COMPTEL (McConnell et al., 2002). Finally, we take advantage of the spectroscopic capability of the instrument to seek for spectral features in the 500 keV region with negative results for any significant annihilation emission on 2 ks and days timescales, as well as in the total dataset.\ author: - 'E. Jourdain, J. P. Roques and J. Malzac' - '*Received ; accepted*' title: 'The emission of Cygnus X-1: observations with *INTEGRAL* SPI [^1] from 20 keV to 2 Mev' --- Introduction ============ Cyg X-1 is an unavoidable target for any high energy instrument. Being one of the most luminous sources (up to the MeV range), it represents an ideal lab to study the mechanism at work in the direct environment of a black hole. There, the accretion flow is thought to form an optically thick disk and/or an optically thin corona, while the observed radio jets could originate from the same area (see e.g.the review by Done, Gierli[ń]{}ski & Kubota 2007). The high energy radiation provides insights on the physical processes at work in this region. Thanks to its persistent high flux ($\sim$ 1 Crab) and usually hard spectrum, Cygnus X-1 is easy to observe and there are numerous measurements of its the spectrum and variability in the X-rays. At higher energies however, the results are very scarce. Very few instruments were able to explore the properties of the emission above 200 keV. The high energy emission of Cyg X-1 is relatively well known from soft X-rays up to a few hundreds keV. The shape of the spectrum is variable and can change dramatically on time scales as short as a day. There are however two main, relatively stable, spectral states: the Hard State (HS, corona dominated) and the Soft State (SS, disk dominated). Their description can be found in various papers (see for example Liang & Nolan 1984; Gierlinski et al., 1997, 1999). So far, the MeV region of the spectrum was best explored by the Compton Gamma-ray Observatory (CGRO). McConnell et al (2002) have shown that, even though the HS hard X-ray emission is dominated by a thermal Comptonization component, both HS and SS spectra present a non thermal powerlaw component extending above 1 MeV with a slope of 2.6 in the SS and $>\sim$ 3 in the HS. Moreover, broad features around 1 MeV have been observed in several occasions in the past (see for example, Bassani et al. 1989 and references therein). Since CGRO observatory however, only the *INTEGRAL* mission contains instrument exploring the same energy domain. Recently, Laurent et al. (2011), stacked all the *INTEGRAL/IBIS* data available for Cygnus X-1 and detected the presence of a non-thermal power law component between 400 keV and 2 MeV. Interestingly they found that contrary to the thermal Comptonization component present below 400 keV, the non-thermal power-law emission appears to be strongly polarized. This non-thermal component appears to have a flux that is stronger than that measured by CGRO by a factor 5-10 and a much harder spectral slope $\Gamma \simeq 1.6$. Here, we use another *INTEGRAL* instrument operating in this energy range, the spectrometer SPI, to investigate the high energy spectral shape of Cyg X-1 and test the presence of non-thermal high energy excess.\ Our first goal is to take advantage of the sensitivity achieved with the SPI detector and the large amount of data and perform a detailed analysis of the energy extent of the hard X-ray emission together with its spectral variability.\ Moreover, the spectroscopic capability of the Germanium crystals allow us to seek for the presence of spectral features linked to the annihilation process. Hereafter, we present briefly the instrument, data set and the method followed for the analysis. Then, we report on our results and start by examining the total mean spectrum to determine the emission above a few hundreds of keV, where scarcity of photons imposes exposure as long as possible. In a second step, we analyse the source behavior during individual revolutions, and build several averaged spectra, following some hardness criteria which can be considered as characteristic of the spectral state of the source. We conclude with a comparison with previous results. Instrument, observations and data analysis ========================================== SPI is a spectrometer aboard the INTEGRAL observatory operating between 20 kev and 8 MeV. The description of the instrument and its performance can be found in Vedrenne et al. (2003) and Roques et al. (2003). The main features of interest for our study are a large FoV ($30^\circ$) with an angular resolution of $2.6^\circ$ (FWHM) based on a coded aperture mask. The Germanium camera, beyond an excellent spectroscopic capability, offers a good sensitivity over more than 2 decades in energy with a unique instrument. It is surrounded by a 5-cm thick BGO shield (ACS, Anti-Coincidence Shield) which measures the particle flux. This latter can be used as a good tracers of the background level.\ During a 3-day orbit, the usual dithering strategy (Jensen et al., 2003) consists of a hundred of 30-40 min exposures (also called scw for ’science window’), with a given region scanned by $2^\circ$ steps following pre-determined patterns. The recommended pattern for SPI observations is a grid of 5X5 around the chosen target. Unfortunately, in order to content more proposers, except a few exceptions, most of the Cyg X-1 data has been obtained through ’amalgamated’ observations, ie with the pattern center somewhere between Cyg X-1 and Cyg A region. As a consequence, Cyg X-1 appears only in one side of the FoV, reducing the mean efficient area (the source is partially coded), with some interruptions in the observation sequences, when the source goes out from the field of view. For our analysis, we have selected in the whole INTEGRAL observation plan, those revolutions in which Cyg X-1 is included in the $\pm 13 ^\circ$ FoV during more than 20 scw (50 ks). This gives a total of 42 revolutions. These observations encompass 4 Ms of effective time, from 2003, June to 2009 December. They were grouped together according to temporal proximity into 13 periods. Exposures with high background level (entry/exit of radiation belts, solar activity) or large source off-angle (source/pointing axis angle beyond 13$^\circ$) have been removed from our dataset. Table 1 gives some details about the observation pattern, beginning, end and useful duration, for each of the defined periods. We follow the analysis method described in Jourdain & Roques (2009), based on a sky model fitting, through a $\chi^2$ minimization. This methods makes use of the Pulse Shape Discrimination (PSD) system, a second electronic chain operating in parallel, to eliminate spurious events occurring in the MeV region. In the 650 keV-2.2 MeV energy range, we use only the PSD tagged events, while below 650 keV, all the single events are analysed. This procedure has been validated with the Crab Nebula observations.\ In this work, we consider a common sky model for all the observations, which consists in 4 sources, namely: Cyg X-1, Cyg X-3, EXO2023+375 and GRS1915+105. To keep the number of degrees of freedom to a minimum, we consider that the three latter have a constant flux over the revolution timescale. For Cyg X-1, we consider a variability timescale of 1 scw ($\sim$ 2 ks) for the lightcurve and hardness ratio studies, while the spectra have been built assuming a constant source flux during each revolution.\ The background determination relies on empty field observations. Because of detector failures, it is important to use proper empty fields for each configuration. Once the uniformity map of the detector plan is fixed, we allow its normalisation to vary on a timescale of $\sim$ 10-20 ks.\ The incident photon spectra have been deconvolved and fitted through the spectral fitting procedure from [xspec]{} 11 (Arnaud 1996). Due to some uncertainties on the energy response in the lowest channels and threshold effects, we exclude the first channels ($E <$ 23.5 keV) from the fit process. The spectra presented contain thus 41 (more or less logarithmic) channels between 23.5 keV and 2.2 MeV. Note that there are two strong narrow background lines (10 to 40 times higher than the continuum emission) at 139 and 198 keV. We use very narrow channels around these energies to isolate the line emission and prevent contamination of the neighbouring channels. They may appear as offset data points with large error bars in the spectra. Study of the spectral shape =========================== The total mean spectrum ------------------------ The total mean spectrum gathers the whole set of available clean data (4 Ms of observation distributed over more than 6 years). It has been built by extracting one spectrum per revolution then summing them. Due to the impressive signal to noise ratio at low energy, we are unable to ensure that the response matrices are known with a sufficient precision. We have thus added 0.5% of systematic errors to the data during the fit procedure.\ The data have been first fitted with a simple analytical cutoff power law and a Comptonization + reflection model (reflect\*comptt in [xspec]{}). Residuals clearly show deviation from these models above 200-300 keV where an excess of emission appears above the model prediction. To go further, we keep a Comptonization law (+ its reflected component) in the low energy part and focus on the high energy emission. Following Laurent et al. (2011), we first try to model this additional component by a single power law. The fit converges toward a photon index of $\sim1.8$ (close to the 1.6 $\pm$ 0.2 value reported by these authors) but the $\chi^2$ value remains unacceptable, with a huge contribution of the last channels. Indeed, the high energy part doesnot follow such a power-law : instead, the emission presents a rather sharp decrease after 700 keV and the source is not detected above 1 MeV. A better result is obtained when modeling this high energy component by a second Comptonization region with a temperature around 100 keV. As several couples (kT, $\tau$) can reproduce the data, we fixed $\tau$ to 0.5 and obtained a best fit value of kT=123 keV (see Table 2). Fig. \[fig:tot.comptt1et2\] presents the observed spectrum with one and two Comptonization models. Even though the set of parameters proposed in Table 2 is only a possible solution, the two Comptonization model provides a good description of the data and more specifically of the curvature observed up to 1 MeV. Flux, hardness and hardness versus flux evolution ------------------------------------------------- The MeV emission may depend on the source spectral state or intensity. To follow potential changes in the source emission, we have analysed in details each revolution and used the information to group observations corresponding to similar states.\ Fig. \[fig:clrev\] displays the temporal evolution of Cyg X-1 in 30-70 keV energy range, with the fluxes averaged by revolution (100-200 ks timescale) and Fig. \[fig:hrrev\] the evolution of the hardness (defined in the 30-120 keV range by $F_{70-120 keV}/F_{30-70 keV}$). In Fig. \[fig:clscw\] and Fig. \[fig:hrscw\], we present the same flux and hardness but detailed on time scales of a few days (with a resolution of one scw ie $\sim$ 2000 s). These graphs illustrate clearly the variability of the source. The hypothesis of constant flux has been tested by $\chi^2$ tests and is strongly rejected for all periods (reduced $\chi^2$ always greater than 5.9 for a number of degrees of freedom ranging from 100 to 300). The variability appears to be chaotic but in both cases over a limited amplitude: the source is always detected and varies by a factor not greater than 3. We can recognize the usual temporal behavior of Cyg X-1 (see e.g. Ling et al. 1987 and Zdziarski et al. 2002). Note that the first period, which corresponds the lowest and softest state in our sample, has been analysed by Malzac et al. (2006). They identified the (rather unusal) source behavior as a mini or failed transition between soft and hard state, in a so called “intermediate state”. However, no robust (that is, lasting more than a few hours) incursion in the soft state can be reported during our observations and we conclude that Cyg X-1 was always in a hard (LH) or intermediate states. Fig. \[fig:hrflux\] displays the hardness as a function of the 30-70 keV source flux (revolution averaged values). The hardness intensity plane can be divided in three regions corresponding to the main clusters of data points. Those three regions are outlined in Fig. \[fig:hrflux\]. The first two regions gather the points with hardnesses respectively below 0.24 and between 0.24 and 0.28, and, incidentally, correspond to the first part of the INTEGRAL mission (revolutions 79 to 261 ie June 2003 - November 2004).\ In the third group, the flux levels span a broader range, extending toward higher values, but the mean hardnesses never decrease below 0.29. During this period, which covers more than 3 years (mid 2006 up to end of 2009), the source evolution consists of an intensity variation within a factor of $\sim$ 2 without any visible change in the emission hardness (or spectral shape), so without notable change in the underlying processes. This behavior has already been observed in this source and is well known in transient sources. Conversely, it is worth noting that different spectral shapes are observed for an unchanged intensity level in the 30-70 keV band. To illustrate both of these effects, we have superimposed in Fig. \[fig:5spec.var\], five spectra representative of the global course followed by the source in Fig. \[fig:hrflux\] (the corresponding revolutions are identified in this figure by squares, numbered from 1 to 5). We recognize the two modes of evolution already identified for this source on the ks timescale (Malzac et al 2006): a pivoting of the spectral shape (spectra 2-4) and a change of luminosity at constant hardness (spectra 1 & 2 and 4 & 5). Note however, that the pivot energy is here at $\sim$ 45 keV (in the middle of the studied energy band, by construction). It could be attributed to an increase of both temperature and optical depth of the Comptonising medium. When combined, the two variability modes give rise to a global complex behavior in the hard X-ray domain, although the source remains in the HS (and intermediate states). To study in more details the spectral shape and its evolution in the high energy part, we have to accumulate data corresponding to the same hardness. Comparison of different averaged spectra ----------------------------------------- Based on the three hardness levels repered in Fig. \[fig:hrflux\], we have built the corresponding averaged spectra, hereafter refered to as ’low hardness’, ’mid hardness’ and ’high hardness’ samples, respectively. They are displayed on Fig. \[fig:3HR\], while best fit parameters are given in Table 3. Similarly to the total spectrum, the data are modeled by a Comptonization law plus its reflection component (required when looking at residuals) and a second hotter Comptonization. The evolution of the slope in the low energy part (in direct relation with the hardness) is clear, with an increase of the peak energy from $\sim$ 50 keV to $\sim$ 150 keV. This behavior is not related to the reflection component but appears as an evolution of the macroscopic parameters (kT and $\tau$) of the Comptonizing medium, which both increase from the soft to the hard sample (see Table 3).\ Looking now to the high energy part, no significant emission is detected above 700 keV. A degeneracy between parameters being unavoidable, we choose to fix the second Comptonizing medium temperature and optical depth to 130 keV and 0.6 respectively for all spectra, but free normalisation factors.\ Even though it represents only a possible description of our data, this model, involving a second (hotter and thiner) Comptonization medium with constant parameters $kT_{e}$ and $\tau$, provides an acceptable interpretation of the Cyg X-1 behavior in the hard X-ray domain. Even if other models could reproduce the data, such two Comptonizing regions (or more generally, variable $kT_{e}$ and/or $\tau$) scenario resembles those already applied in previous works to Cyg X-1 and several BHBs (see for example the Suzaku observations analysed by Makishima et al, 2008, and references therein). Beyond the specific sets of best fit parameter formally obtained, the inadequacy of the single comptonization region suggests spatial and/or rapid temporal variation of the temperature or optical depth of the Comptonization region. Annihilation feature ==================== Cygnus X-1 is one of the brightest Galactic sources of hard X-rays up to several hundreds of keV and is therefore a good candidate for positron production. The excellent energy resolution of the SPI germanium detector makes it the best instrument to seek for potential annihilation signatures in the observed emission.\ We have thus looked for any emission feature, narrow (10 keV FWHM) or broad (80 keV FWHM), transient (scw ie $\sim$ 2 ks timescale) or more persistent (revolution ie 1-2 day timescale). On the scw timescale, no significant emission is reported with 2 $\sigma$ upper limits of 2-3 $\times 10^{-3}$  and 0.5-1 $\times 10^{-2}$  for a narrow and broad line, respectively. On the revolution timescale, no significant excess above the continuum contribution are found, with upper limits ranging between 3-6 $\times 10^{-4}$  and 0.7-1.1 $\times 10^{-3}$ according to the revolution duration. Finally, considering the whole set of data (4 Ms), persistent emission features, if any, should be below 6 $\times 10^{-5}$  and 1.3 $\times 10^{-4}$ . Comparison with other instruments ================================= We now compare our observations with the available data in the MeV region. From Fig. \[fig:SPIGRO\], we can see that our “high hardness sample” observations has a spectral shape that is very close to that of the HS mean spectrum reported by McConnell et al. (2002) from GRO/ COMPTEL and OSSE instruments. To compare the whole set of data (INTEGRAL/SPI + CGRO/Comptel+OSSE), we use a model consisting in a cutoff power law plus a broken power-law with the first index fixed to -1 ( to avoid any contribution in the low energy part).\ Imposing the same photon index for the cutoff power-laws, the fit procedure converges toward slightly different cutoff energies ($\sim$ 140 and 160 keV, for SPI and OSSE data respectively, no contribution in COMPTEL range). We also impose a common second photon index in the broken power law and obtained 3.4 $\pm$ 0.5, while the energy break reflects the true difference between the spectra: close to 900 ($\pm$ 100) keV in the CGRO data, it means that this component is based essentially on the COMPTEL points. Around 420 ($\pm$ 25) keV in the SPI data, it allows us to recover the additional emission above the low energy component, the very soft index limiting its extension toward high energies.\ In conclusion, even if the SPI data do not show significant emission above 1 MeV, the upper limits are marginally consistent with the non-thermal tail observed at several MeV by CGRO/COMPTEL. We note however that this emission is not taken into account by the two-zone thermal Comptonization model proposed above which would require an additional component to fit the COMPTEL data. The overall good agreement between the SPI and OSSE spectra indicate that the average spectral properties in the hard state have remained constant between the two epochs. Recently, Laurent et al. (2011) presented a stacked spectrum of Cygnus X-1 obtained using the IBIS data during nearly the same observation period as ours. IBIS is composed of two position sensitive detector layers ISGRI (CdTe, 15-1000 keV; Lebrun et al. 2003), and PICsIT (CsI, 200 keV-10 MeV; Labanti et al. 2003). These two detectors are usually used independently to produce spectra and light curves. Laurent et al. (2011) used the IBIS/ISGRI data up to $\sim$ 400 keV and IBIS/PICsiT at higher energies. Fig. \[fig:laurent\] compares our averaged SPI spectra with the results of Laurent et al. 2011. We note that most of the published results from IBIS actually use only the ISGRI detector. With the recent versions of the public software, the results from IBIS/ISGRI are generally compatible with those of SPI. This good agreement is confirmed by our comparison of the stacked SPI and IBIS/ISGRI spectra. At higher energies however, the results are clearly different: the SPI fluxes are lower than the IBIS/PICsiT points by a factor of about 5 at least. The origin of this disagreement is unclear. We were not able to go further in the analysis of the discrepancy as there are very few published results from IBIS/PICsiT. Summary and Conclusions ======================== We used 4 Ms of *INTEGRAL* SPI data to study the emission of Cyg X-1 from 20 keV to the MeV region. While the source has been essentially detected in the HS, its presents complex variability on all timescales. We have studied the luminosity and hardness evolution of the source above 20 keV, on the scw (2000 s) and revolution (1-2 days) timescales. The revolution averaged data give a nice picture of the long term source behavior. A change by a factor of 2-3 in luminosity can be accompanied either by a significant softening in the 20-150 keV domain (see Fig. \[fig:3HR\]) or by the same spectral shape just moved up and down. Then, in order to improve the photon statistics at high energies we combined observations to produce long exposure time average spectra. The analysis of the stacked spectra reveals that the emission of Cyg X-1 extends up to $\sim$ 700 keV but presents a sharp cutoff around that energy. Whatever the criterion we used to built averaged spectra (low/mid/high hardness, low/high intensity, all data), no emission can be detected above 1 MeV.\ Nevertheless, a single Comptonization model does not provide a good description of the overall spectral shape. We have shown that a two temperature model provides a good fit of the SPI data, with a minimum of free parameters ($\tau$ and kT of the second Compton component can be considered as constant along the time). In a final step to investigate the high energy emission, we compare all the data available above 300 keV for Cyg X-1 in the last 2 decades and conclude that while our SPI upper limits are marginally compatible with the soft powerlaw reported by COMPTEL in the 90’ (McConnell et al, 2002), they clearly disagree (lower by a factor $\sim$ 5) with the recently reported IBIS/PICsiT emission (Laurent et al, 2011).\ In other words, the presence of a non-thermal mechanism participating to the power supply can not be excluded, but our results contradict the presence of a hard (polarised or not) power law emission. A last point concerns the potential emission linked to the positron production: No positive detection has been found, in a narrow (10 keV) and broad (80 keV) channel, on 2ks and day timescales, as well as in the total dataset. The question of the origin and the nature of the hard state high energy emission remains thus of prime interest. Cyg X-1 is clearly the most adequate target to investigate it. Since no mission operating above 300 keV is expected for a while, the only hope is that the INTEGRAL instruments will be able to accumulate a few more Ms of data on this important target to resolve the matter once and for all. Acknowledgments {#acknowledgments .unnumbered} =============== The *INTEGRAL* SPI project has been completed under the responsibility and leadership of CNES. We are grateful to ASI, CEA, CNES, DLR, ESA, INTA, NASA and OSTC for support. We thank A. A. Zdziarski for providing us with GRO/COMPTEL and OSSE data. [17]{} natexlab\#1[\#1]{} Bassani, L., Dean, A. J., Di Cocco, G., Perotti, F. and Stephen, J. B.,  1989, , 343, 313 Done, C., Gierli[ń]{}ski, M., & Kubota, A.,  2007, , 15, 1 Gierlinski, M., Zdziarski, A.A., Done C. et al., 1997, , 288, 958 Gierlinski, M., Zdziarski, A.A., Poutanen, J.. et al.,  1999, , 309, 496 Jensen, P.L., Clausen, K., Cassi, C. et al.,  2003, , 411, L7 Jourdain, E. & Roques, J. P.,  2009, , 704, 17 Laurent, P., Rodriguez, J., Wilms, J et al.,  2011, Science, 332, 438 Labanti, C., Di Cocco, G., Ferro, G., et al.,  2003, , 411, L149 Lebrun, F., Leray, J. P., Lavocat, P., et al.,  2003, , 411, L141 Liang, E. P. and Nolan, P. L.,  1984, SSRv, 38, 353 Ling J.C., Mahoney, W. A., Wheaton W. A. & Jacobson A. S.,  1987, , 321, L117 Makishima, K., Takahashi, H., Yamada, S. et al.,  2008, PASJ, 60, 585 Malzac, J., Petrucci, P.O., Jourdain, E. et al.,  2006, , 448, 1125 McConnell, M. L., Zdziarski, A. A., Bennett, K. et al.,  2002, , 572, 984 Roques J.P., Schanne S., Von Kienlin A., et al,  2003, , 411, L91 Vedrenne, G., Roques, J.P., Schonfelder, V. et al,  2003, , 411, L63 Zdziarski, A. A., Poutanen, J., Paciesas, W. S. and Wen L.,  2002, , 578, 357 [lccccccc]{} \[tab:revol\] 79-80 (5x5)& 2003-06-07 00:59 &2003-06-12 03:35 & 293\ 210-214 (A)& 2004-07-03 00:01 &2004-07-17 00:25 & 709\ 251-252 (A)& 2004-11-03 14:23 & 2004-11-07 16:26 & 176\ 259 & 261 (H) & 2004-11-26 12:28 & 2004-12-03 15:43 & 143\ 470 (EXO, H) & 2006-08-19 09:19 &2006-08-21 16:02 & 159\ 486 (EXO, H) & 2006-10-06 00:11 & 2006-10-08 07:55 & 160\ 498-505 (GP) & 2006-11-11 19:31 &2006-12-04 06:20 & 535\ 628-631 (A) & 2007-12-04 19:05 & 2007-12-15 21:08 & 388\ 673 (A) & 2008-04-18 17:41 &2008-04-19 22:09 & 54\ 682-684 (A) & 2008-05-14 08:13 & 2008-05-22 19:54 & 304\ 739-746 (A) & 2008-11-01 02:14 &2008-11-24 05:25 & 551\ 803-806 (A) & 2009-05-11 08:27 &2009-05-22 11:32 & 371\ 875(H\*) & 877(H) & 2009-12-12 16:18 & 2009-12-19 20:57 & 160\ [lcccccccc]{} \[tab:fittot\] Refl\*Comptt& 0.90 $\pm$ 0.3 & 75.0 $\pm$ 3 & 0.91 $\pm$ 0.03 & & & 6.0 (37)\ Refl\*Comptt + power law& 0.88 $\pm$ 0.3 & 56 $\pm$ 3 & 1.2 $\pm$ 0.1 & 1.8 $\pm$ 0.2 & & 5.1 (35)\ Refl\*Comptt+Comptt & 0.8 (fixed) & 38 $\pm$ 3 & 1.6 $\pm$ 0.15 & 123$\pm$ 10& 0.5 (fixed) & 1.7 (36)\ [lcccccccccccccc]{} \[tab:fit3hr\] Low hardness sample & 0.55 $\pm$ 0.25 & 33 $\pm$ 3 & 1.2 $\pm$ 0.1 & 130 (fix) & 0.6 (fix) & 0.99 (36) &\ Mid hardness sample & 0.73 $\pm$ 0.15 &36 $\pm$ 3 & 1.4 $\pm$ 0.2 & 130 (fix) & 0.6 (fix) &2.13 (36)\ High hardness sample & 0.68 $\pm$ 0.15 & 40 $\pm$ 3 & 1.6 $\pm$ 0.1 & 130 (fix) & 0.6 (fix) & 1.4 (36) &\ [^1]: Based on observations with INTEGRAL, an ESA project with instruments and science data centre funded by ESA member states (especially the PI countries: Denmark, France, Germany, Italy, Spain, and Switzerland), Czech Republic and Poland with participation of Russia and USA.

ArXiv

--- abstract: 'We consider a non-degenerate conic in ${{\rm{PG}}}(2,q^2)$, $q$ odd, that is tangent to $\ell_\infty$ and look at its structure in the Bruck-Bose representation in ${{\rm{PG}}}(4,q)$. We determine which combinatorial properties of this set of points in ${{\rm{PG}}}(4,q)$ are needed to reconstruct the conic in ${{\rm{PG}}}(2,q^2)$. That is, we define a set ${{\cal C}}$ in ${{\rm{PG}}}(4,q)$ with $q^2$ points that satisfies certain combinatorial properties. We then show that if $q\ge 7$, we can use ${{\cal C}}$ to construct a regular spread ${{\cal S}}$ in the hyperplane at infinity of ${{\rm{PG}}}(4,q)$, and that ${{\cal C}}$ corresponds to a conic in the Desarguesian plane ${\mathcal{P}}({{\cal S}})\cong{{\rm{PG}}}(2,q^2)$ constructed via the Bruck-Bose correspondence.' author: - | S.G. Barwick and Wen-Ai Jackson\ School of Mathematics, University of Adelaide\ Adelaide 5005, Australia title: 'Characterising pointsets in ${{\rm{PG}}}(4,q)$ that correspond to conics' --- AMS code: 51E20. Keywords: finite projective geometry, Bruck-Bose representation, conics, characterization Introduction ============ We begin by describing the Bruck-Bose representation of a translation plane of order $q^2$ with kernel containing ${{\rm{GF}}}(q)$ in ${{\rm{PG}}}(4,q)$, proved independently by André [@andr54] and Bruck and Bose [@bruc64; @bruc66]. Let ${{\Sigma_\infty}}$ be a hyperplane of ${{\rm{PG}}}(4,q)$ and let ${{\cal S}}$ be a spread of ${{\Sigma_\infty}}$. Consider the incidence structure whose [*points*]{} are the points of ${{\rm{PG}}}(4,q)\setminus{{\Sigma_\infty}}$, whose [*lines*]{} are the planes of PG$(4,q)$ which do not lie in $\Sigma_\infty$ but which meet $\Sigma_\infty$ in a line of ${{\cal S}}$ and where [*incidence*]{} is inclusion. This incidence structure is an affine translation plane and can be uniquely completed to a projective translation plane ${\mathcal{P}}({{\cal S}})$ of order $q^2$ by adjoining the line at infinity ${\ell_{\infty}}$ whose points are the elements of the spread ${{\cal S}}$. The line $\ell_\infty$ is a translation line for ${\mathcal{P}}({{\cal S}})$. The translation plane ${\mathcal{P}}({{\cal S}})$ is Desarguesian if and only if the spread ${{\cal S}}$ is regular ([@bruc64]). For more details on the Bruck-Bose representation, see [@barw08], in particular, note that Baer subplanes of ${{\rm{PG}}}(2,q^2)$ secant to ${\ell_{\infty}}$ are in one to one correspondence with affine planes of ${{\rm{PG}}}(4,q)$ that do not contain a line of the regular spread ${{\cal S}}$. In ${{\rm{PG}}}(4,q)$, with ${{\Sigma_\infty}}$ being the hyperplane at infinity, we call the points of ${{\rm{PG}}}(4,q)\setminus{{\Sigma_\infty}}$ [ *affine points*]{} and the points in ${{\Sigma_\infty}}$ [*infinite points*]{}. The lines and planes of ${{\rm{PG}}}(4,q)$ that are not contained in ${{\Sigma_\infty}}$ are called [ *affine lines*]{} and [*affine planes*]{} respectively. Now consider a non-degenerate conic $\overline{{\cal C}}$ in ${{\rm{PG}}}(2,q^2)$ that is tangent to ${\ell_{\infty}}$ in the point $P_\infty$. In the Bruck-Bose representation of ${{\rm{PG}}}(2,q^2)$ in ${{\rm{PG}}}(4,q)$, the affine points ${{\cal C}}=\overline{{\cal C}}\setminus\{P_\infty\}$ correspond to a set of $q^2$ affine points in ${{\rm{PG}}}(4,q)$ also denoted by ${{\cal C}}$. These points satisfy a variety of properties. In [@barwcaps], algebraic properties of these points were determined, and it was shown they were caps. In this article we are interested in combinatorial properties the points satisfy, in particular, combinatorial properties relating to the planes of ${{\rm{PG}}}(4,q)$. Our aim was to find the smallest set of these properties that would allow us to reconstruct the projective plane and the conic $\overline{{\cal C}}$. The properties of the conic in ${{\rm{PG}}}(2,q^2)$ that we are interested in are given in the next lemma, which is proved in Section \[section:lemma-proof\]. [conic-satisfies-props]{} Let $\overline{{\cal C}}$ be a non-degenerate conic of ${{\rm{PG}}}(2,q^2)$, $q$ odd, that is tangent to ${\ell_{\infty}}$ in a point $P_\infty$. Define a ${{\cal C}}$-plane to be a Baer subplane of ${{\rm{PG}}}(2,q^2)$ that is secant to ${\ell_{\infty}}$ and meets ${{\cal C}}=\overline{{\cal C}}\setminus\{P_\infty\}$ in $q$ points. Then 1. Each ${{\cal C}}$-plane meets ${{\cal C}}$ in a $q$-arc. Further, if a Baer subplane secant to ${\ell_{\infty}}$ meets ${{\cal C}}$ in more than four points, it is a ${{\cal C}}$-plane. 2. Every pair of points of ${{\cal C}}$ lie in exactly one ${{\cal C}}$-plane. 3. The affine points of ${{\rm{PG}}}(2,q^2)$ are of three types: points of ${{\cal C}}$; points on no ${{\cal C}}$-plane (the interior points of $\overline{{\cal C}}$); and points on exactly two ${{\cal C}}$-planes (the exterior points of $\overline{{\cal C}}$). We now consider what these properties correspond to in the Bruck-Bose representation in ${{\rm{PG}}}(4,q)$. The points of ${{\cal C}}$ correspond to affine points of ${{\rm{PG}}}(4,q)$, and the [${{\cal C}}$-plane]{}s correspond to affine planes of ${{\rm{PG}}}(4,q)$ that contain $q$ points of ${{\cal C}}$. We suppose that we have a set of affine points and ${{\cal C}}$-planes satisfying the ${{\rm{PG}}}(4,q)$ equivalence of the combinatorial properties of Lemma \[conic-satisfies-props\], and show that we can reconstruct the conic in the Bruck-Bose plane. The main result of this paper is the characterisation given in the next theorem. [mainthm]{} Let ${{\Sigma_\infty}}$ be the hyperplane at infinity in ${{\rm{PG}}}(4,q)$, $q\ge7$, $q$ odd. Let ${{\cal C}}$ be a set of $q^2$ affine points, called ${{\cal C}}$-points, and suppose there exists a set of affine planes called [${{\cal C}}$-plane]{}s satisfying the following properties: - Each [${{\cal C}}$-plane]{} meets ${{\cal C}}$ in a $q$-arc. Further, if a plane meets ${{\cal C}}$ in more than four points, it is a [${{\cal C}}$-plane]{}. - Every pair of points of ${{\cal C}}$ lie in exactly one [${{\cal C}}$-plane]{}. - The affine points of ${{\rm{PG}}}(4,q)$ are of three types: points of ${{\cal C}}$, points on no [${{\cal C}}$-plane]{}, and points on exactly two [${{\cal C}}$-plane]{}s. Then there exists a unique spread ${{\cal S}}$ in ${{\Sigma_\infty}}$ so that the [${{\cal C}}$-point]{}s in the Bruck-Bose plane ${\mathcal{P}}({{\cal S}})$ form a $q^2$-arc of ${\mathcal{P}}({{\cal S}})$. Moreover, the spread ${{\cal S}}$ is regular, and so ${\mathcal{P}}({{\cal S}})\cong{{\rm{PG}}}(2,q^2)$, and the $q^2$-arc can be completed to a conic of ${{\rm{PG}}}(2,q^2)$. We note that a similar characterisation when $q$ is even is given in [@conicqeven]. The rest of this paper is devoted to proving this theorem. The main structure of the proof is as follows. We need a number of preliminary results, leading to Theorem \[lines-form-a-spread\] where we show how to construct a spread ${{\cal S}}$ of ${{\Sigma_\infty}}$ from ${{\cal C}}$. In Corollary \[cor:regular-spread\], we show that ${{\cal S}}$ is a regular spread. In Theorem \[thm:isanarc\], we show that the set ${{\cal C}}$ corresponds to a $q^2$-arc in the Bruck-Bose plane ${\mathcal{P}}({{\cal S}})\cong{{\rm{PG}}}(2,q^2)$, and hence as $q$ is odd, ${{\cal C}}$ is contained in a unique conic. In Theorem \[thm:diffspread\], we show that if ${{\cal S}}'$ is any other spread of ${{\Sigma_\infty}}$, then the points of ${{\cal C}}$ will not correspond to an arc in the associated Bruck-Bose plane ${\mathcal{P}}({{\cal S}}')$. Proof of Lemma \[conic-satisfies-props\] ======================================== [section:lemma-proof]{} In this section we prove that a conic in ${{\rm{PG}}}(2,q^2)$ satisfies the combinatorial properties stated in Lemma \[conic-satisfies-props\]. [**of Lemma \[conic-satisfies-props\]**]{} Let $\overline{{\cal C}}$ be a non-degenerate conic in ${{\rm{PG}}}(2,q^2)$ tangent to ${\ell_{\infty}}$ in the point $P_\infty$, and let ${{\cal C}}=\overline{{\cal C}}\setminus\{ P_\infty\}$. We begin with a note about subconics of $\overline{{\cal C}}$ in Baer subplanes. Suppose $\pi$ is a Baer subplane that meets $\overline{{\cal C}}$ in a subconic ${{\cal O}}$. Note that for each point $P$ in ${{\cal O}}\cap\pi$, the tangent line of $\overline{{\cal C}}$ at $P$ is a line of $\pi$. Conversely, if a line $\ell$ of $\pi$ is a tangent line of $\overline{{\cal C}}$, then the point of contact $\ell\cap\overline{{\cal C}}$ lies in $\pi$. We now prove part 1. The $q$ points of ${{\cal C}}$ in a ${{\cal C}}$-plane lie on the conic $\overline{{\cal C}}$, so are clearly a $q$-arc. Suppose $\pi$ is a Baer subplane secant to ${\ell_{\infty}}$ which meets ${{\cal C}}$ in five points, then those five points define a unique conic in ${{\rm{PG}}}(2,q^2)$, namely $\overline{{\cal C}}$. Further, the five points define a unique conic ${{\cal O}}$ in $\pi$, which is necessarily a subconic of $\overline{{\cal C}}$. As $\pi$ is secant to ${\ell_{\infty}}$, it contains the tangent ${\ell_{\infty}}$ to ${\overline{{\cal C}}}$, hence by the above note, it contains the point $P_\infty$ of contact of the tangent ${\ell_{\infty}}$ to ${\overline{{\cal C}}}$. Hence $\pi$ meets ${{\cal C}}$ in $q$ points, and so is a ${{\cal C}}$-plane. For part 2, let $P,Q$ be two points of ${{\cal C}}$ with tangents to ${\overline{{\cal C}}}$ labeled $t_P$, $t_Q$ respectively. Suppose $\pi$ is any Baer subplane secant to ${\ell_{\infty}}$ containing $P,Q$. As $\pi$ contains ${\ell_{\infty}}$, it contains $P_\infty$ by the above note. Further, as $\pi$ contains $P,Q$, we have that $t_P,t_Q$ are lines of $\pi$. Thus $\pi$ contains the quadrangle $P,Q,t_P\cap {\ell_{\infty}},P_\infty$ and hence there is exactly one such subplane. However, the quadrangle $P,Q,t_P\cap {\ell_{\infty}},P_\infty$ defines a unique Baer subplane which contains five elements from the conic ${\overline{{\cal C}}}$: namely the points $P,Q,P_\infty$ and the tangents ${\ell_{\infty}},t_P$, and thus contains a subconic of ${\overline{{\cal C}}}$ containing $P_\infty$. Thus every pair of points of ${{\cal C}}$ lie on exactly one [${{\cal C}}$-plane]{}. Note that this allows us to count the number of ${{\cal C}}$-planes. Let ${{\cal A}}$ be the incidence structure with [*points*]{} the points of ${{\cal C}}$; [*lines*]{} the ${{\cal C}}$-planes; and inherited incidence. Then ${{\cal A}}$ is a $2$-$(q^2,q,1)$ design, and so is an affine plane of order $q$. Hence there are $q^2+q$ ${{\cal C}}$-planes. To prove part 3, let $\pi$ be a ${{\cal C}}$-plane, and let $X$ be an affine point of $\pi$ that is not in ${{\cal C}}$. Then $X$ lies on two tangents of $\pi\cap{\overline{{\cal C}}}$, which lie either in $\pi$ or in ${{\rm{PG}}}(2,q^2){\backslash}\pi$, depending on whether $X$ is an exterior point or an interior point of the conic $\pi\cap{\overline{{\cal C}}}$. In either case, $X$ lies on two tangents of ${\overline{{\cal C}}}$ in ${{\rm{PG}}}(2,q^2)$, and so is an exterior point of ${\overline{{\cal C}}}$. Hence the affine points in a ${{\cal C}}$-plane either lie in ${{\cal C}}$, or are exterior points of ${\overline{{\cal C}}}$. That is, the interior points of ${\overline{{\cal C}}}$ lie on zero ${{\cal C}}$-planes. It is straightforward to show that the group of homographies of ${{\rm{PG}}}(2,q^2)$ fixing $\overline{{\cal C}}$ and ${\ell_{\infty}}$ is transitive on the affine exterior points of $\overline{{\cal C}}$. Hence all the affine exterior points of $\overline{{\cal C}}$ lie on a common number $x$ of ${{\cal C}}$-planes. We now count incident pairs $(X,\pi)$ where $X$ is an affine exterior point of $\overline{{\cal C}}$ that lies on a ${{\cal C}}$-plane $\pi$. We have $q^2(q^2-1)/2\times x=(q^2+q)(q^2-q)$, hence $x=2$ as required. Proof of Theorem \[mainthm\] ============================ Properties of ${{\cal C}}$-planes --------------------------------- Let ${{\cal C}}$ be a set of $q^2$ affine points in ${{\rm{PG}}}(4,q)\setminus{{\Sigma_\infty}}$, $q\ge7$, $q$ odd, satisfying assumptions (A1), (A2) and (A3) of Theorem \[mainthm\]. We begin by noting that the ${{\cal C}}$-points and [${{\cal C}}$-plane]{}s form an affine plane. This gives us a natural set of parallel classes which will be useful in our proof. [affine-plane]{} Let ${{\cal A}}$ be the incidence structure with [*points*]{} the points of ${{\cal C}}\!$; [*lines*]{} the [${{\cal C}}$-plane]{}s; and inherited incidence. Then ${{\cal A}}$ is an affine plane of order $q$, and so there are $q^2+q$ [${{\cal C}}$-plane]{}s. Hence there are $q+1$ [*parallel classes*]{} of ${{\cal C}}$-planes, each containing $q$ [*parallel*]{} [${{\cal C}}$-plane]{}s. By (A1) and (A2), ${{\cal A}}$ is a 2-$(q^2,q,1)$ design and so is an affine plane of order $q$. Note that in ${{\rm{PG}}}(4,q)$, two ${{\cal C}}$-planes $\pi,\alpha$ meet in a point or a line. The parallel classes of the affine plane ${{\cal A}}$ tell us whether the intersection $\pi\cap\alpha$ contains a ${{\cal C}}$-point. So we have: if $\pi$ and $\alpha$ belong to the same parallel class, then they have no common ${{\cal C}}$-point; whereas if $\pi$ and $\alpha$ belong to different parallel classes, then they share exactly one ${{\cal C}}$-point. Let $\pi$ be a [${{\cal C}}$-plane]{}. By (A1), $\pi$ contains $q$ [${{\cal C}}$-point]{}s that form a $q$-arc. As $q$ is odd, this arc uniquely completes to a conic by the addition of a point which we denote $\pi_\infty$ (see [@hirs98 Theorem 10.28]). We call $\pi_\infty$ the [*$\infty$-point*]{} of $\pi$ since we will show in Lemma \[cplane-meet-theorem\] that $\pi_\infty$ is on the line $\pi\cap{{\Sigma_\infty}}$, and so is an infinite point of ${{\rm{PG}}}(4,q)$. That is: Let $\pi$ be a ${{\cal C}}$-plane, then the [*$\infty$-point*]{} $\pi_\infty$ of $\pi$ is the point that uniquely completes the $q$-arc $\pi\cap{{\cal C}}$ to a conic. To study the intersection of ${{\cal C}}$-planes in more detail, we will need the next lemma involving three ${{\cal C}}$-planes that all contain a fixed $\infty$-point. [one-parallel-class]{} Let $\pi$ be a [${{\cal C}}$-plane]{} with $\infty$-point $\pi_\infty$. Suppose two [${{\cal C}}$-plane]{}s $\alpha,\beta$ both meet $\pi$ in distinct lines through $\pi_\infty$. Then $\alpha$ and $\beta$ share no point other than $\pi_\infty$. Let $\pi,\alpha,\beta$ be ${{\cal C}}$-planes such that $\alpha\cap\pi$, $\beta\cap\pi$ are distinct lines through $\pi_\infty$. Note that if $\pi_\infty\not\in\Sigma_\infty$, then $\pi_\infty$ is on three [${{\cal C}}$-plane]{}s, namely $\pi$, $\alpha$ and $\beta$, contradicting (A3). So we have $\pi_\infty\in\Sigma_\infty$. Suppose $\alpha$ and $\beta$ meet in a line $\ell$ (so $\ell$ contains $\pi_\infty$). The line $\ell$ contains at most one [${{\cal C}}$-point]{} by (A2). Note that the $q+1$ lines of $\pi$ through $\pi_\infty$ consist of $q$ 1-secants to ${{\cal C}}$, and one 0-secant to ${{\cal C}}$. Hence at least one of the lines $\alpha\cap\pi$, $\beta\cap\pi$ is a 1-secant of ${{\cal C}}$. Without loss of generality, suppose $\pi\cap\beta$ meets ${{\cal C}}$ in the point $B$. As ${{\cal C}}$ meets $\pi$ in a $q$-arc, there are two 1-secants to ${{\cal C}}$ through $B$ in $\pi$, one is $B\pi_\infty$, let the other meet $\pi\cap\alpha$ in the point $R$. In the plane $\beta$, there are two 1-secants to ${{\cal C}}$ through $B$, one is $B\pi_\infty$, let the other meet $\beta\cap\alpha$ in the point $S$. We now show that there exists a 2-secant $m$ of $\alpha\cap{{\cal C}}$ which does not contain the points $S, R$ or $\pi_\infty$, and which does not meet $\ell$ or $\alpha\cap\pi$ in a [${{\cal C}}$-point]{}. As $q\geq7$, the $q$-arc $\alpha\cap{{\cal C}}$ has at least five points $X,Y,Z,V,W$ which do not lie on $\ell$ or on $\alpha\cap\pi$ (each of which contain at most one ${{\cal C}}$-point). A 2-secant through two of these points will meet $\ell$ and $\alpha\cap\pi$ in points that are not in ${{\cal C}}$. First consider the three 2-secants $XY,XZ,YZ$. If none of these is suitable, then each contains one of $R$, $S$, $\pi_\infty$. So next consider the three 2-secants $XW,YW,ZW$. If none of these is suitable, then each contains one of $R,S,\pi_\infty$. Without loss of generality suppose $R\in XW$, $S\in YW$ and $\pi_\infty\in ZW$. Hence we also have $R\in YZ$, $S\in XZ$, and $\pi_\infty\in XY$. Now consider the 2-secant $VX$. If $S\in VX$, then $X,Z,V$ would be collinear, contradicting $\alpha\cap{{\cal C}}$ being an arc. So $S\notin VX$. Similarly $R\not\in VX$ and $\pi_\infty\not\in VX$. So $m=VX$ is a 2-secant of ${{\cal C}}$ not containing $R,S$ or $\pi_\infty$, and meeting $\ell$ and $\alpha\cap\pi$ in points not in ${{\cal C}}$. So there exists a line $m$ in $\alpha$ that contains two ${{\cal C}}$-points, but does not contain the points $S, R$ or $\pi_\infty$, and does not meet $\ell$ or $\alpha\cap\pi$ in a [${{\cal C}}$-point]{}. Consider the plane $\tau=\langle m,B\rangle$. Now, $\tau$ meets $\beta$ in a line through $B$ that is not a 1-secant of ${{\cal C}}$ as $\pi_\infty,S\notin\tau$. Hence $\tau$ contains two points of $\beta\cap{{\cal C}}$ (one of which is $B$). Similarly, $\tau$ meets $\pi$ in two points of $\pi\cap{{\cal C}}$, one of which is $B$. Hence $\tau$ contains five distinct points of ${{\cal C}}$, and so is a [${{\cal C}}$-plane]{} by (A1). Hence $m$ contains two ${{\cal C}}$-points which lie on two [${{\cal C}}$-plane]{}s, namely $\alpha$ and $\tau$, contradicting (A2). Hence $\alpha$ and $\beta$ cannot meet in a line $\ell$, so they share no point other than $\pi_\infty$. The next lemma is key to our proof. We first show that for each [${{\cal C}}$-plane]{} $\pi$, the $\infty$-point $\pi_\infty$ of the $q$-arc $\pi\cap{{\cal C}}$ lies in ${{\Sigma_\infty}}$. Then we study how [${{\cal C}}$-plane]{}s can meet, and investigate properties of the parallel classes of the affine plane ${{\cal A}}$ defined in Lemma \[affine-plane\]. We use the following definition. If $\pi$ is a [${{\cal C}}$-plane]{}, then we call the line $\pi\cap{{\Sigma_\infty}}$ a [*[${{\cal C}}$-line]{}*]{}. [cplane-meet-theorem]{} 1. If $\pi$ is a ${{\cal C}}$-plane, then its $\infty$-point $\pi_\infty$ lies in ${{\Sigma_\infty}}$. 2. Let $\pi,\alpha$ be distinct ${{\cal C}}$-planes. 1. If $\pi,\alpha$ lie in the same parallel class, then they meet in exactly one point $\pi_\infty$ (which is equal to $\alpha_\infty$). 2. If $\pi,\alpha$ lie in different parallel classes, then they either meet in exactly one point of ${{\cal C}}$, or they meet in an affine line through $\pi_\infty$ (which is equal to $\alpha_\infty$) that contains one point of ${{\cal C}}$. 3. The [${{\cal C}}$-plane]{}s which are in a common parallel class all share the same [$\infty$-point]{}, and they pairwise intersect in only this [$\infty$-point]{}. Moreover, every [$\infty$-point]{} defines exactly two parallel classes. Let $\pi$ be a fixed [${{\cal C}}$-plane]{} and let $\pi_{{\cal C}}=\pi\cap{{\cal C}}$. We say a [${{\cal C}}$-plane]{} $\alpha$ (distinct from $\pi$) [*covers*]{} an affine point $P$ of $\pi{\backslash}\pi_{{\cal C}}$ if it contains $P$. Further an affine line of $\pi$ is called a [*cover line*]{} of $\pi$ if it is contained in a [${{\cal C}}$-plane]{} distinct from $\pi$. We work with the parallel classes of the affine plane ${{\cal A}}$ defined in Lemma \[affine-plane\]. The $q$ points of $\pi_{{\cal C}}$ each lie on $q$ further ${{\cal C}}$-planes (one in each of the parallel classes not containing $\pi$). The $q^2-q$ affine points of $\pi{\backslash}\pi_{{\cal C}}$ each lie in exactly one further ${{\cal C}}$-plane by (A3). That is, each of the $q^2-q$ affine points of $\pi{\backslash}\pi_{{\cal C}}$ is covered by exactly one ${{\cal C}}$-plane. We will investigate how the ${{\cal C}}$-planes cover these points. First we look at how another ${{\cal C}}$-plane $\alpha$ meets $\pi$. Now $\alpha\cap\pi$ is either a point or a line of ${{\rm{PG}}}(4,q)$. If $\alpha,\pi$ are in different parallel classes, then $\alpha,\pi$ contain exactly one common ${{\cal C}}$-point, so $\alpha\cap\pi$ is either a ${{\cal C}}$-point, or an affine 1-secant of ${{\cal C}}$. If $\alpha,\pi$ are in the same parallel class, then $\alpha,\pi$ contain no common ${{\cal C}}$-point. Let $${\mathscr P}=\{\beta_1,\ldots,\beta_{q-1}\}$$ be the ${{\cal C}}$-planes in the same parallel class as $\pi$. Then we have: - - If $\alpha\in{\mathscr P}$, then $\alpha$ meets $\pi$ in either a point of $\pi\setminus\pi_{{\cal C}}$ (possibly in ${{\Sigma_\infty}}$); or an affine 0-secant of $\pi_{{\cal C}}$; or an infinite line. - If $\alpha\not\in{\mathscr P}$, then $\alpha$ meets $\pi$ in either exactly one ${{\cal C}}$-point, or in an affine 1-secant of $\pi_{{\cal C}}$. We now focus on I(a) and show that no [${{\cal C}}$-plane]{} in $\mathscr P$ meets $\pi$ in an affine line. Suppose there is one such plane $\beta\in\mathscr P$ meeting $\pi$ in the affine line $b$, so $b$ is a 0-secant of $\pi_{{\cal C}}$ (we work towards a contradiction to show that $b$ cannot exist). No other cover line in $\pi$ can meet $b$ in an affine point, otherwise that point would be on three [${{\cal C}}$-plane]{}s, contradicting (A3). Hence every other cover line of $\pi$ (if any) meets $b$ in the point $X=b\cap{{\Sigma_\infty}}$. Let $\mathcal L$ be the set of $q$ affine lines of $\pi$ through $X$. There are four possibilities for these lines, depending on the location of the $\infty$-point $\pi_\infty$. Firstly, suppose $\pi_\infty\in{{\Sigma_\infty}}$. Note that $\pi_\infty$ lies on exactly one 0-secant of $\pi_{{\cal C}}$, which in this case is $\pi\cap{{\Sigma_\infty}}$. Hence $\pi_\infty\neq X$, as otherwise $X$ lies on two 0-secants of $\pi_{{\cal C}}$, namely $\pi\cap{{\Sigma_\infty}}$ and $b$. Thus the $q$ lines of $\mathcal L$ consist of one 1-secant, $\frac{q-1}2$ 2-secants, and $\frac{q-1}2$ 0-secants of $\pi_{{\cal C}}$. Secondly, if $\pi_\infty\not\in{{\Sigma_\infty}}$ and $X\pi_\infty$ is a 0-secant of $\pi_{{\cal C}}$, then $\mathcal L$ contains one 1-secant, $\frac{q-1}2$ 2-secants and $\frac{q-1}2$ 0-secants of $\pi_{{\cal C}}$. Thirdly, if $\pi_\infty\not\in{{\Sigma_\infty}}$ and $X\pi_\infty$ is a 1-secant of $\pi_{{\cal C}}$, then either $\mathcal L$ contains three 1-secants, $\frac{q-3}2$ 2-secants and $\frac{q-3}2$ 0-secants of $\pi_{{\cal C}}$, or $\mathcal L$ contains one 1-secant, $\frac{q-1}2$ 2-secants and $\frac{q-1}2$ 0-secants of $\pi_{{\cal C}}$. Let $\mathcal L'\subset\mathcal L$ be the set of 2-secants of $\pi_{{\cal C}}$ through $X$. In each of the four cases, $\mathcal L'$ contains either $\frac{q-1}2$ or $\frac{q-3}2$ lines. As $q\geq7$, $\mathcal L'$ contains at least two lines $\ell,m$. By I, the affine cover lines of $\pi$ are either 0-secants or 1-secants of $\pi_{{\cal C}}$, so $\ell$ and $m$ are not cover lines. Further, we noted above that all cover lines must contain $X$. Hence the affine points of $\ell,m$ not in ${{\cal C}}$ are not contained in any cover lines. Thus by I, the affine points of $\ell,m$ not in ${{\cal C}}$ must be covered by planes in ${\mathscr P}\setminus \{\beta\}$, with each plane covering at most one affine point. However, there are $2(q-2)$ such affine points and only $q-2$ [${{\cal C}}$-plane]{}s in ${\mathscr P}\setminus \{\beta\}$, a contradiction. Hence the line $b$ does not exist, so no [${{\cal C}}$-plane]{} in $\mathscr P$ meets $\pi$ in an affine line. Hence the following modification of I(a) holds: - If $\alpha\in{\mathscr P}$, then $\alpha$ meets $\pi$ in either a point of $\pi\setminus\pi_{{\cal C}}$, or in an infinite line. We now show that $\pi_\infty\in{{\Sigma_\infty}}$. By I(a)$'$, [${{\cal C}}$-plane]{}s in $\mathscr P$ meet $\pi$ in either 0 or 1 affine points. Hence the [${{\cal C}}$-plane]{}s in $\mathscr P$ cover at most $q-1$ of the $q^2-q$ affine points of $\pi\setminus\pi_{{\cal C}}$. Recall that each of these affine points is covered by exactly one ${{\cal C}}$-plane. By I(b), [${{\cal C}}$-plane]{}s not in ${\mathscr P}$ either meet $\pi$ in exactly a [${{\cal C}}$-point]{} or in a 1-secant of $\pi_{{\cal C}}$. So there are at least $((q^2-q)-(q-1))/(q-1)=q-1$ [${{\cal C}}$-plane]{}s not in $\mathscr P$ that meet $\pi$ in a cover line, denote them by $$\mathscr Q=\{\alpha_1,\ldots,\alpha_{q-1}\}.$$ For each $i$, $\alpha_i$ and $\pi$ have exactly one common ${{\cal C}}$-point, denoted $A_i$. As the cover line $\alpha_i\cap\pi$ is a 1-secant of the $q$-arc $\pi_{{\cal C}}$ in $\pi$, it is either the line $A_i\pi_\infty$, or it is the unique tangent line at $A_i$ to the conic $\pi_{{\cal C}}\cup\pi_\infty$, denote this tangent line by $\ell_i$. Suppose one of the cover lines $\alpha_1\cap\pi$ is the unique tangent line $\ell_1$ (we work towards a contradiction to show this cannot happen). Note that two cover lines cannot meet in an affine point of $\pi\setminus\pi_{{\cal C}}$ by (A3). As the cover lines $\alpha_2\cap\pi,\ldots,\alpha_{q-1}\cap\pi$ are all 1-secants, they either meet $\ell_1$ in $A_1$ (there is at most one other such 1-secant, namely $A_1\pi_\infty$) or in a 1-secant through $\ell_1\cap{{\Sigma_\infty}}$ (there are at most two other such 1-secants through $\ell_1\cap{{\Sigma_\infty}}$). This gives at most three other possibilities for cover lines, contradicting $|\mathscr Q\setminus\alpha_1|=(q-1)-1\geq 5$. Thus no tangent line $\ell_i$ is a cover line, so $A_i\pi_\infty$ is a cover line for $i=1,\ldots,q-1$. That is, $\pi_\infty$ lies on $q$ ${{\cal C}}$-planes, so by (A3), $\pi_\infty$ lies in $\pi\cap{{\Sigma_\infty}}$. This completes part 1. We now further improve I(a)$'$ and show that the ${{\cal C}}$-planes in $\mathscr P$ either meet $\pi$ in the infinite point $\pi_\infty$, or in an infinite line. First note that we have shown that the $q-1$ [${{\cal C}}$-plane]{}s in $\mathscr Q$ all contain $\pi_\infty$. Hence by Lemma \[one-parallel-class\], these planes pairwise meet only in $\pi_\infty$. In particular, they pairwise have no common ${{\cal C}}$-point. Hence the [*planes in ${\mathscr Q}$ all belong to the same parallel class*]{}, of which there is one more member, $\alpha_q$ say. Denote by $\mathscr X$ the set of $q-1$ affine points of $\pi{\backslash}\pi_{{\cal C}}$ that are not covered by the [${{\cal C}}$-plane]{}s in $\mathscr Q$. The points of $\mathscr X$ lie on a line through $\pi_\infty$. We will show that $\alpha_q$ meets $\pi$ in this line. Consider a plane $\alpha_1\in\mathscr Q$. A similar argument to that of part 1 shows that the $\infty$-point $(\alpha_1)_\infty\in\alpha_1\cap{{\Sigma_\infty}}$. If $\alpha_1$ and $\pi$ share [${{\cal C}}$-line]{}s, that is $\alpha_1\cap{{\Sigma_\infty}}=\pi\cap{{\Sigma_\infty}}$, then $\alpha_1,\pi$ must coincide since they share two lines, namely their [${{\cal C}}$-line]{} and an affine line. Hence their [${{\cal C}}$-line]{}s are distinct, and a similar argument to that of part 1 shows that $\pi$ meets $\alpha_1$ in a line through $(\alpha_1)_\infty$, hence $(\alpha_1)_\infty=\pi_\infty$. Next, similar to the set $\mathscr Q$ related to the [${{\cal C}}$-plane]{} $\pi$, consider the set $\mathscr T$ of at least $q-1$ [${{\cal C}}$-plane]{}s that each meet $\alpha_1$ in a line through $(\alpha_1)_\infty$. By the above argument, they lie in a common parallel class. As $\pi\in\mathscr T$, it follows that the parallel class containing $\mathscr T$ is the parallel class $\mathscr P\cup\pi$ (since every [${{\cal C}}$-plane]{} belongs to exactly one parallel class). Hence the (at least) $q-2$ planes $\beta_i\in\mathscr P\cap\mathscr T$ satisfy $(\beta_i)_\infty=(\alpha_1)_\infty=\pi_\infty$. Thus at least $q-2$ of the [${{\cal C}}$-plane]{}s in $\mathscr P$ meet $\pi$ exactly in the point $\pi_\infty$. In particular, at least $q-2$ of the $\beta_i$ contain no affine points of $\pi$. Hence there are at least $q-2$ affine points in the set $\mathscr X$ that are not covered by planes in $\mathscr P$. Hence by I, there is a ${{\cal C}}$-plane $\alpha$ not in $\mathscr P$ that meets $\pi$ in a 1-secant of $\pi_{{\cal C}}$ through $\pi_\infty$. This line covers $q-1$ affine points, so covers all the points of $\mathscr X$. Note that by Lemma \[one-parallel-class\], $\alpha$ belongs to the parallel class containing $\mathscr Q$, so $\alpha=\alpha_q$. In summary, we have: 1. - ${{\cal C}}$-planes in the parallel class ${\mathscr P}\cup {\pi}$ all contain $\pi_\infty$, and pairwise contain no common affine point; - Each ${{\cal C}}$-plane in the parallel class ${\mathscr Q}\cup\alpha_q$ meets $\pi$ in a 1-secant through $\pi_\infty$; - the remaining ${{\cal C}}$-planes each meet $\pi$ in exactly one ${{\cal C}}$-point. To complete the proof of part 2, we need to show that ${{\cal C}}$-planes in $\mathscr P$ meet $\pi$ in exactly $\pi_\infty$. Suppose not, that is, by II, suppose that there is a ${{\cal C}}$-plane $\beta_1$ in $\mathscr P$ that meets $\pi$ in an infinite line, so $\beta_1\cap\pi=\pi\cap{{\Sigma_\infty}}$. Let $\alpha\in\mathscr Q\cup\alpha_q$, so $\alpha$ meets $\pi$ in a 1-secant through $\pi_\infty$. As $\pi,\beta_1$ are in the same parallel class, $\beta_1$ and $\alpha$ lie in different parallel classes, so have a common ${{\cal C}}$-point, $A$ say. Then the line $A\pi_\infty$ lies in both $\beta$ and $\alpha$, contradicting Lemma \[one-parallel-class\]. Hence ${{\cal C}}$-planes in the parallel class $\mathscr P\cup\pi$ pairwise meet in exactly the point $\pi_\infty$. Hence we have shown that [${{\cal C}}$-plane]{}s in the same parallel class have the same $\infty$-point, and pairwise meet in exactly this point. To prove the second statement of part 3, we note that by II, the ${{\cal C}}$-planes in the parallel classes $\mathscr P\cup\pi$ and $\mathscr Q\cup\alpha_q$ all contain the point $\pi_\infty$. Further, the remaining ${{\cal C}}$-planes do not contain $\pi_\infty$. So $\pi_\infty$ lies on the ${{\cal C}}$-planes of precisely two parallel classes. This lemma shows that two ${{\cal C}}$-planes cannot meet in a ${{\cal C}}$-line, so we have: [count-cline]{} Each ${{\cal C}}$-plane has a unique ${{\cal C}}$-line, so there are $q^2+q$ distinct ${{\cal C}}$-lines in ${{\Sigma_\infty}}$. We now characterise the points of $\Sigma_\infty$ in relation to the [${{\cal C}}$-line]{}s. A point of $\Sigma_\infty$ that lies on no [${{\cal C}}$-line]{} is called a [*0-point*]{}. [lem:pttype]{} There are three types of points in $\Sigma_\infty$: ${q+1\over 2}$ [$\infty$-point]{}s, ${q+1\over 2}$ 0-points, and $q^3+q^2$ points which lie on exactly one [${{\cal C}}$-line]{} each. Let $\ell,m$ be two ${{\cal C}}$-lines in ${{\Sigma_\infty}}$, and let $\alpha,\beta$ be the unique ${{\cal C}}$-planes with ${{\cal C}}$-lines $\ell,m$ respectively. By Lemma \[cplane-meet-theorem\](2), $\ell,m$ meet in a exactly one point (namely $\alpha_\infty=\beta_\infty$), or not at all. So each point of ${{\Sigma_\infty}}$ is either a 0-point, an $\infty$-point, or lies on exactly one ${{\cal C}}$-line. By Corollary \[count-cline\], the number of points on exactly one ${{\cal C}}$-line is $(q^2+q)q=q^3+q^2$. By Lemma \[cplane-meet-theorem\](3), each [$\infty$-point]{} lies in exactly two parallel classes, hence there are ${q+1\over 2}$ [$\infty$-point]{}s. Thus the number of 0-points is $(q^3+q^2+q+1)-({q+1\over 2})-(q^3+q^2)={q+1\over 2}$. As a direct consequence of this proof, we have the following corollary which tells us how ${{\cal C}}$-lines are positioned in ${{\Sigma_\infty}}$. [cline-infty]{} If two ${{\cal C}}$-lines meet, then they do so in an $\infty$-point. Further, each $\infty$-point lies on exactly $2q$ ${{\cal C}}$-lines. Defining a spread in ${{\Sigma_\infty}}$ ---------------------------------------- In this section we show how we can use the points of ${{\cal C}}$ to construct a spread in ${{\Sigma_\infty}}$. We show that each ${{\cal C}}$-point $A$ defines a unique line $t_A$ in ${{\Sigma_\infty}}$. These resulting $q^2$ lines are mutually skew. The remaining $q+1$ points in ${{\Sigma_\infty}}$ are the 0-points and [$\infty$-point]{}s, we show they lie on a line ${t_\infty}$, and that the lines $t_A$ and ${t_\infty}$ form a spread in ${{\Sigma_\infty}}$. We begin by defining for each ${{\cal C}}$-point $A$, a set of $q+1$ points $t_A$ in ${{\Sigma_\infty}}$. [def:tA]{} Let $A$ be [${{\cal C}}$-point]{} and let $\pi$ be a [${{\cal C}}$-plane]{} through $A$. The unique tangent to the conic $(\pi\cap{{\cal C}})\cup\pi_\infty$ at the point $A$ meets ${{\Sigma_\infty}}$ in a point which we denote $A_\pi$. Note that $A_\pi$ is on the [${{\cal C}}$-line]{} of $\pi$. There are $q+1$ [${{\cal C}}$-plane]{}s $\pi_i$, $i=1,\ldots,q+1$, through $A$. Define $t_A$ to be the set of $q+1$ points $A_{\pi_i}$, $i=1,\ldots,q+1$. We show in Theorem \[cline-cover\] that the set $t_A$ is a line in ${{\Sigma_\infty}}$. First we show that the $\infty$-points and the $0$-points lie on a line of ${{\Sigma_\infty}}$, denoted $t_\infty$. [thm-tinfty]{} The set of [$\infty$-point]{}s and 0-points lie on a line ${t_\infty}$. Let $\ell$ denote the set of [$\infty$-point]{}s and 0-points, so by Lemma \[lem:pttype\], $|\ell|=q+1$. Let $\alpha$ be any plane of ${{\Sigma_\infty}}$. If $\alpha$ contains a [${{\cal C}}$-line]{}, then it contains an [$\infty$-point]{} and so meets $\ell$. If $\alpha$ does not contain a [${{\cal C}}$-line]{}, then $\alpha$ contains at most one point of each of the $q^2+q$ [${{\cal C}}$-line]{}s of ${{\Sigma_\infty}}$. As $\alpha$ has $q^2+q+1$ points, it follows that $\alpha$ meets $\ell$ in at least one point. Hence every plane of ${{\Sigma_\infty}}$ meets $\ell$, and so by [@hirs98 Theorem 3.5], $\ell$ is a line. The next two lemmas investigate the 3-space spanned by two ${{\cal C}}$-planes. [cor:old4]{} A $3$-space contains at most two [${{\cal C}}$-plane]{}s. Let $\alpha$, $\beta$ be two ${{\cal C}}$-planes that span a 3-space, so by Lemma \[cplane-meet-theorem\](2), they are in different parallel classes, and $\alpha_\infty=\beta_\infty$. Thus, for a 3-space to contain three [${{\cal C}}$-plane]{}s, they would have a common [$\infty$-point]{}, and the planes would be pairwise in different parallel class. This cannot occur as through any [$\infty$-point]{} there are at most two parallel classes by Lemma \[cplane-meet-theorem\](3). [two-cplanes-3-sp]{} Let $\Sigma$ be a $3$-space containing two distinct [${{\cal C}}$-plane]{}s $\alpha$ and $\beta$. Then the ${{\cal C}}$-points in $\Sigma$ are exactly the ${{\cal C}}$-points in $\alpha$ and $\beta$. If two [${{\cal C}}$-plane]{}s $\alpha$ and $\beta$ generate a 3-space $\Sigma$, then they meet in a line $\ell$, which by Lemma \[cplane-meet-theorem\](2) contains a unique [${{\cal C}}$-point]{} $A$. Further, $\alpha,\beta$ are in different parallel classes. Suppose $\Sigma$ contains a further [${{\cal C}}$-point]{} $B$ not in $\alpha$ or $\beta$. Of the $q+1$ [${{\cal C}}$-plane]{}s through $B$, one is parallel to $\alpha$, one is parallel to $\beta$ and one contains $A$. Let $\pi$ be any of the $q-2\geq 1$ remaining [${{\cal C}}$-plane]{}s through $B$. Then $\pi$ meets $\alpha$ in a [${{\cal C}}$-point]{} $C$ and $\beta$ in a [${{\cal C}}$-point]{} $D$. As $B,C,D$ are ${{\cal C}}$-points, they are not collinear, so $\pi=\langle B,C,D\rangle$. Hence $\pi,\alpha,\beta$ are three [${{\cal C}}$-plane]{}s contained in a 3-space $\Sigma$, contradicting Lemma \[cor:old4\]. We now show that the set of points $t_A$ defined in \[def:tA\] form a line of ${{\Sigma_\infty}}$. [cline-cover]{} For each [${{\cal C}}$-point]{} $A$, the set $t_A$ is a line of ${{\Sigma_\infty}}$. By Lemma \[lem:pttype\], there are $(q+1)/2$ $\infty$-points, label these $T_1,\ldots,T_{\frac{q+1}{2}}$. Let $A$ be a ${{\cal C}}$-point, each of the $q+1$ ${{\cal C}}$-planes through $A$ contains exactly one $\infty$-point. If three ${{\cal C}}$-planes $\alpha,\beta,\gamma$ through $A$ contained the same $\infty$-point $T_1$ say, then by Lemma \[cplane-meet-theorem\](2), they are all in different parallel classes. However, this contradicts Lemma \[cplane-meet-theorem\](3) as $T_1$ defines exactly two parallel classes. Hence each of these $q+1$ $\infty$-points lies in exactly two ${{\cal C}}$-planes that contain $A$. Hence the $q+1$ ${{\cal C}}$-planes $\pi_1,\ldots,\pi_{q+1}$ through $A$ can be ordered so that the pair $(\pi_1,\pi_2)$ both contain $T_1$, the pair $(\pi_3,\pi_4)$ both contain $T_2$, and so on. For each ${{\cal C}}$-plane $\pi_i$, $i=1,\ldots,q+1$, through $A$, define the point $A_{\pi_i}$ as in Definition \[def:tA\], and let $t_A=\{A_{\pi_1},\ldots,A_{\pi_{q+1}}\}$. Now $\pi_1,\pi_2$ both contain the line $AT_1$, so $\Sigma=\langle \pi_1,\pi_2\rangle$ is a 3-space. Consider the plane $\sigma_\infty=\Sigma\cap{{\Sigma_\infty}}$. By Lemma \[cor:old4\], $\pi_1$ and $\pi_2$ are the only ${{\cal C}}$-planes contained in $\Sigma$, so the remaining $q^2+q-2$ ${{\cal C}}$-planes each meet $\Sigma$ in a line. We now show that each of these $q^2+q-2$ lines meet the plane $\sigma_\infty$ in exactly one point. Suppose not, that is, suppose that some ${{\cal C}}$-plane $\pi$ meets $\sigma_\infty$ in a line $\ell$, so $\ell$ is the ${{\cal C}}$-line of $\pi$. Let $m,n$ be the ${{\cal C}}$-lines of $\pi_1,\pi_2$ respectively. So the plane $\sigma_\infty$ contains three ${{\cal C}}$-lines $\ell,m,n$ which pairwise meet. By Corollary \[cline-infty\], these ${{\cal C}}$-lines contain a common $\infty$-point, namely $T_1$. As $\pi_1,\pi_2$ meet in a line, by Lemma \[cplane-meet-theorem\](2), $\pi_1,\pi_2$ are in different parallel classes, and $\pi$ is in same parallel class as one of $\pi_1,\pi_2$. Suppose $\pi $ is in the same parallel class as $\pi_1$, then by Lemma \[cplane-meet-theorem\], $\pi$ meets $\pi_2$ in a line. So as $\ell$ is also in $\sigma_\infty$, we conclude that $\pi$ lies in $\Sigma$, contradicting Lemma \[cor:old4\]. Thus each of the $q^2+q-1$ [${{\cal C}}$-plane]{}s distinct from $\pi_1,\pi_2$ contain exactly one point of $\sigma_\infty$. In particular, as the [${{\cal C}}$-plane]{}s $\pi_3,\ldots,\pi_{q+1}$ all contain $A\in\Sigma$, they meet $\Sigma$ in affine lines labelled $\ell_3,\ldots,\ell_{q+1}$ respectively, so each line $\ell_i$ meets ${{\Sigma_\infty}}$ in a point. As $\pi_3,\ldots,\pi_{q+1}$ all contain the point $A$, they contain no further [${{\cal C}}$-point]{}s of $\pi_1$ and $\pi_2$. In particular, $\ell_i$, $i=3,\ldots,q+1$ does not contain any [${{\cal C}}$-point]{}s that lie in $\pi_1$ or $\pi_2$ (other than $A$). By Lemma \[two-cplanes-3-sp\], the only ${{\cal C}}$-points in $\Sigma$ are those in $\pi_1$ and $\pi_2$. Hence each $\ell_i$ contains exactly one ${{\cal C}}$-point, namely $A$. Recall that $\pi_1$ and $\pi_2$ have a common $\infty$-point $T_1$, and that $\pi_3,\ldots,\pi_{q+1}$ do not contain $T_1$. Thus in $\pi_i$, $\ell_i$ contains exactly one ${{\cal C}}$-point $A$, but does not meet ${t_\infty}$. Hence the line $\ell_i$ is the required tangent line to $A$ in $\pi_i$, and hence meets $\sigma_\infty$ in the point $A_{\pi_i}$. From this we conclude that if $\sigma_\infty$ is the plane defined by the [${{\cal C}}$-line]{}s of $\pi_1$ and $\pi_2$, then $\sigma_\infty$ contains all the points $A_{\pi_i}$ for $i=1,\ldots,q+1$. Next we repeat the argument with the pair $(\pi_3,\pi_4)$ and define $\sigma'_\infty$ to be the plane defined by the [${{\cal C}}$-line]{}s of $\pi_3$ and $\pi_4$. A similar argument shows that $\sigma'_\infty$ contains all the points $A_{\pi_i}$ for $i=1,\ldots,q+1$. As $\sigma_\infty$ and $\sigma'_\infty$ meet ${t_\infty}$ in distinct points, it follows that they meet in a line (disjoint from $t_\infty$), and this line $t_A$ contains the $q+1$ points $A_{\pi_i}$ for $i=1,\ldots,q+1$. The next two corollaries follow from the definition of $t_A$, and this result that $t_A$ is a line. [cor:tA]{} Any [${{\cal C}}$-plane]{} containing the point $A\in{{\cal C}}$ meets the line $t_A$, and conversely, any [${{\cal C}}$-plane]{} that meets $t_A$ contains the point $A$. [cor:tA2]{} Let $A$ be a ${{\cal C}}$-point, then the affine plane $\langle t_A,A\rangle$ contains exactly one ${{\cal C}}$-point. We now show that we can use these lines to construct a spread in ${{\Sigma_\infty}}$. [lines-form-a-spread]{} The lines $t_A$, $A\in{{\cal C}}$, together with ${t_\infty}$ form a spread ${{\cal S}}$ of $\Sigma_\infty$. Let $\pi$ be a ${{\cal C}}$-plane with ${{\cal C}}$-line $\ell=\pi\cap{{\Sigma_\infty}}$, and consider the conic ${{\cal D}}=({{\cal C}}\cap\pi)\cup\pi_\infty$ in $\pi$. The line $\ell$ is a tangent to ${{\cal D}}$, hence every point of $\ell\setminus\pi_\infty$ lies on one further tangent of ${{\cal D}}$. That is, each point of $\ell\setminus\pi_\infty$ lies in exactly one line $t_A$. By Lemma \[lem:pttype\], each point of ${{\Sigma_\infty}}{\backslash}t_\infty$ lies in exactly one ${{\cal C}}$-line, hence each point of ${{\Sigma_\infty}}{\backslash}t_\infty$ lies in exactly one of the $t_A$. Further, the number of points in ${{\Sigma_\infty}}{\backslash}t_\infty$ equals the number of points on the $q^2$ lines $t_A$. Hence the lines $t_A$, $A\in{{\cal C}}$ and $t_\infty$ are mutually disjoint, and form a spread of ${{\Sigma_\infty}}$. The spread ${{\cal S}}$ is regular ---------------------------------- In this section we will show that the spread ${{\cal S}}=\{t_A{:}A\in{{\cal C}}\}\cup\{{t_\infty}\}$ is regular. We begin with two lemmas showing how certain affine planes meet ${{\cal C}}$. [l-meets-Pinfty]{} Let $\ell$ be a line of ${{\Sigma_\infty}}$ that meets ${t_\infty}$, but is not a [${{\cal C}}$-line]{}. Then 1. Every affine plane containing $\ell$ meets ${{\cal C}}$ in at most two points. 2. If $\ell$ meets a spread line $t_A$ (with corresponding ${{\cal C}}$-point $A$), then the plane $\langle A,\ell\rangle$ contains exactly one [${{\cal C}}$-point]{}, namely $A$. Let $\ell$ be a line of ${{\Sigma_\infty}}$ that meets ${t_\infty}$ in a point, but is not a [${{\cal C}}$-line]{}. Suppose $\pi$ is an affine plane through $\ell$ that contains three ${{\cal C}}$-points $P,Q,R$. By (A2), the points $P,Q$ lie on a unique [${{\cal C}}$-plane]{}, denoted $\pi_{PQ}$. Similarly we have [${{\cal C}}$-plane]{}s $\pi_{PR},\pi_{QR}$. As $\ell$ is not a [${{\cal C}}$-line]{}, $\pi$ is not a ${{\cal C}}$-plane, hence the three [${{\cal C}}$-plane]{}s $\pi_{PQ},\pi_{PR},\pi_{QR}$ are distinct. Consider the 3-space $\Sigma=\langle t_\infty,\pi\rangle$. Now $P,Q,\pi_{PQ}\cap t_\infty$ are three non-collinear points in $\pi_{PQ}$ and in $\Sigma$, hence $\pi_{PQ}\subset\Sigma$. Similarly $\pi_{PR},\pi_{QR}\subset\Sigma$. This contradicts Lemma \[cor:old4\]. Hence every affine plane through $\ell$ meets ${{\cal C}}$ in at most two points, proving part 1. The line $\ell$ meets ${t_\infty}$ and $q$ other lines of the spread, let $t_A$ be one such line, and let $A$ be the ${{\cal C}}$-point corresponding to $t_A$. Consider the plane $\langle A,\ell\rangle$, and suppose it contains a second point $R$ of ${{\cal C}}$. By (A2), there is a unique [${{\cal C}}$-plane]{} $\pi_{AR}$ through $A,R$. The line $AR$ meets $\ell$ in a point $X$. By Corollary \[cor:tA2\], $X\notin t_A$, and the line through $A$ and $\ell\cap{t_\infty}$ is a 1-secant of ${{\cal C}}$, so $X$ lies on a unique spread line $t_B$, $B\neq A$. Hence the ${{\cal C}}$-plane $\pi_{AR}$ meets $\ell$ in this point $X$, that is, the [${{\cal C}}$-line]{} $m=\pi_{AR}\cap{{\Sigma_\infty}}$ meets $\ell$ in the point $X$ of $t_B$. Further $\ell\ne m$ as $m$ is a [${{\cal C}}$-line]{} but $\ell$ is not, so $\ell$ and $m$ meet ${t_\infty}$ in distinct points. Also note that as $A\in\pi_{AR}$, we have by Corollary \[cor:tA\] that $\pi_{AR}$ meets $t_A$, that is, $m$ meets $t_A$. Hence $\langle\ell,m\rangle$ is a plane that contains the spread lines ${t_\infty},t_A$, a contradiction as ${t_\infty},t_A$ are skew by Theorem \[lines-form-a-spread\]. Hence we have shown that $\langle A,\ell\rangle$ contains only one [${{\cal C}}$-point]{}, namely $A$. The next two lemmas examine planes that contain $t_\infty$. [lemma-plane-thru-tinfty]{} Each affine plane through ${t_\infty}$ meets ${{\cal C}}$ in exactly one point. Suppose $\alpha$ is a plane through ${t_\infty}$ that meets ${{\cal C}}$ in two points $A,B$. By (A2), there is a unique ${{\cal C}}$-plane $\pi_{AB}$ containing $A$ and $B$. Let $\pi_{AB}\cap {t_\infty}=Y$, then $Y$ is the $\infty$-point of $\pi_{AB}$, and so $Y$ can be added to $\pi_{AB}\cap{{\cal C}}$ to form a conic. This is a contradiction as the line $AB$ meets this conic in three points, namely $A,B,Y$. Hence each affine plane through ${t_\infty}$ meets ${{\cal C}}$ in at most one point. As there are $q^2$ affine planes through ${t_\infty}$ and $q^2$ points of ${{\cal C}}$, we have exactly one [${{\cal C}}$-point]{} in each affine plane about ${t_\infty}$. [cplane-parallel-classes]{} Every plane in ${{\Sigma_\infty}}$ containing ${t_\infty}$ contains $q$ [${{\cal C}}$-line]{}s through a common [$\infty$-point]{}, and the corresponding $q$ [${{\cal C}}$-plane]{}s are in one parallel class. Consider the $q+1$ planes of ${{\Sigma_\infty}}$ through ${t_\infty}$. By Corollary \[cline-infty\], each plane about ${t_\infty}$ contains at most $q$ [${{\cal C}}$-line]{}s through a common [$\infty$-point]{}. There are $q^2+q$ [${{\cal C}}$-line]{}s, hence each of the $q+1$ planes about ${t_\infty}$ contains exactly $q$ [${{\cal C}}$-line]{}s through a common $\infty$-point. Let $\ell,m$ be two ${{\cal C}}$-lines in the same plane of ${{\Sigma_\infty}}$ about ${t_\infty}$. Let $\alpha,\beta$ be the two [${{\cal C}}$-plane]{}s with [${{\cal C}}$-line]{}s $\ell,m$ respectively. Suppose $\alpha,\beta$ are in different parallel classes, then by Lemma \[cplane-meet-theorem\](2) they meet in a line containing a [${{\cal C}}$-point]{} $A$, and span a 3-space. Consider any other [${{\cal C}}$-line]{} $n$ in the plane $\langle {t_\infty},\ell,m\rangle$. Now if $\delta$ is the [${{\cal C}}$-plane]{} with [${{\cal C}}$-line]{} $n$, then $\delta$ cannot be parallel to both $\alpha$ and $\beta$ as $\alpha$ and $\beta$ are not in the same parallel class. So without loss of generality say $\delta$ meets $\alpha$ in the ${{\cal C}}$-point $B$. Consider the 3-space $\langle \alpha,\beta\rangle$. It contains $\ell$ and $m$, and so contains $n$. Further, it contains $B$, so $\delta\subset\langle \alpha,\beta\rangle$. This contradicts Lemma \[cor:old4\]. Hence $\alpha$ and $\beta$ are not in different parallel classes, that is, the [${{\cal C}}$-line]{}s in a plane of ${{\Sigma_\infty}}$ through ${t_\infty}$ pass through a common [$\infty$-point]{}, and are contained in [${{\cal C}}$-plane]{}s that belong to a common parallel class. We now consider a line of ${{\Sigma_\infty}}$ meeting $t_\infty$ and $q$ other spread lines, and show that the corresponding ${{\cal C}}$-points lie in a common ${{\cal C}}$-plane. [proj-result]{} If $\ell$ is a line of ${{\Sigma_\infty}}$ meeting spread lines ${t_\infty}, t_{A_1},\ldots t_{A_q}$, then the corresponding ${{\cal C}}$-points $A_1,\ldots,A_q$ lie on a common [${{\cal C}}$-plane]{}. Let $\ell$ be a line meeting spread lines ${t_\infty}, t_{A_1},\ldots t_{A_q}$. Let $V=\ell\cap t_\infty$. Let $\pi$ be a ${{\cal C}}$-plane not through $V$, let $A$ be a ${{\cal C}}$-point not in $\pi$ and let $\Sigma=\langle A,\pi\rangle$. We first show that $V\notin\Sigma$. By Lemma \[lemma-plane-thru-tinfty\], the plane $\langle {t_\infty},A\rangle$ contains exactly [${{\cal C}}$-point]{}, namely $A$. As each affine line of $\pi$ through $\pi_\infty=\pi\cap t_\infty$ meets ${{\cal C}}$ in a point, $\langle {t_\infty},A\rangle$ meets the plane $\pi$ in exactly the point $\pi_\infty$. Hence $\langle {t_\infty},A\rangle$ is not a plane of $\Sigma$, and so ${t_\infty}$ is not contained in $\Sigma$, thus $V\notin\Sigma$. Hence we can project the points of ${{\cal C}}$ from $V$ onto $\Sigma$ to obtain a set ${{\cal C}}'$. By Lemma \[lemma-plane-thru-tinfty\], no line through $V$ can contain 2 points of ${{\cal C}}$, hence ${{\cal C}}'$ contains $q^2$ distinct points. We consider this projection in the two cases: when $V$ is a $0$-point, and when $V$ is an $\infty $-point. Case 1: Suppose $V$ is a $0$-point, we will show that the set ${{\cal E}}={{\cal C}}'\cup\{\pi_\infty\}$ is an elliptic quadric in $\Sigma$. Further, each [${{\cal C}}$-plane]{} projects to a unique affine plane of $\Sigma$ through $\pi_\infty$, and conversely, each affine plane of $\Sigma$ through $\pi_\infty$ is the image of a unique [${{\cal C}}$-plane]{}. We first show that ${{\cal C}}'\cup\{ \pi_\infty\}$ is a cap of $\Sigma$. By Lemma \[l-meets-Pinfty\], a plane meeting ${t_\infty}$ in the 0-point $V$ meets ${{\cal C}}$ in at most 2 points. Hence no three points of ${{\cal C}}'$ are collinear (otherwise their preimages would be coplanar with $V$). Further, if the point $\pi_\infty$ were collinear with two points $B',C'$ of ${{\cal C}}'$, then their preimages $B,C\in{{\cal C}}$ would lie in a plane about ${t_\infty}$, which is not possible by Lemma \[lemma-plane-thru-tinfty\]. So ${{\cal E}}={{\cal C}}'\cup\{\pi_\infty\}$ is a set of $q^2+1$ points, no three collinear in a 3-space of order $q$, $q$ odd. Hence ${{\cal E}}$ is an elliptic quadric, see [@barl55] or [@pane55]. Now let $\alpha$ be a [${{\cal C}}$-plane]{}. As $V$ is a 0-point, $\alpha$ meets ${t_\infty}$ in a point distinct from $V$, and so $\alpha$ projects to an affine plane $\alpha'$ through $\pi_\infty$. Let $\alpha,\beta$ be two distinct ${{\cal C}}$-planes with images $\alpha',\beta'$ respectively. If $\alpha'=\beta'$, then as $\alpha,\beta$ together contain at least $2q-1$ ${{\cal C}}$-points, the plane $\alpha'$ contains $2q-1$ points of the elliptic quadric ${{\cal E}}$, a contradiction. As there are $q^2+q$ [${{\cal C}}$-plane]{}s and $q^2+q$ affine planes in $\Sigma$ through $\pi_\infty$ (that is, planes not contained in ${{\Sigma_\infty}}$), each [${{\cal C}}$-plane]{} projects to a unique affine plane of $\Sigma$ through $\pi_\infty$, and conversely. This completes the proof of the statement at Case 1. Recall that $\ell$ meets the spread lines $t_{A_1},\ldots,t_{A_q}$, we now consider the corresponding ${{\cal C}}$-points $A_i$, and their images $A_i'$ under the projection from $V$. By Lemma \[l-meets-Pinfty\], the plane $\pi_i=\langle\ell,A_i\rangle$ contains exactly one [${{\cal C}}$-point]{}, namely $A_i$. Let $L=\ell\cap\Sigma$, so $L\in{{\Sigma_\infty}}$ as $\ell\subseteq{{\Sigma_\infty}}$. The plane $\pi_i=\langle\ell,A_i\rangle$ projects to the line $LA_i'$ which contains exactly one point of ${{\cal C}}'$, namely $A_i'$, hence these lines are distinct for distinct $i$. As ${{\Sigma_\infty}}$ contains no [${{\cal C}}$-point]{}s, the projection $\Sigma\cap{{\Sigma_\infty}}$ of ${{\Sigma_\infty}}$ from $V$ onto $\Sigma$ is a tangent plane to ${{\cal E}}$, so $L\pi_\infty$ is a tangent line to ${{\cal E}}$. Thus from a point $L\in\Sigma$ we have a set $\{L\pi_\infty,LA_1',\ldots,LA_q'\}$ of $q+1$ tangent lines of ${{\cal E}}$. As ${{\cal E}}$ is an elliptic quadric, the points $\pi_\infty,A_1',\ldots,A_q'$ all lie on a plane $\beta'$, namely the polar plane of $L$, see [@hirs85 Theorem 15.3.10]. Hence by the above argument, $\beta'$ is the image of a [${{\cal C}}$-plane]{} $\beta$, and hence the points $A_i$ lie on a common [${{\cal C}}$-plane]{}, namely $\beta$. So the lemma holds in the case $V=\ell\cap t_\infty$ is a $0$-point. Case 2: Suppose $V$ is an [$\infty$-point]{}. We will show that the set ${{\cal C}}'$ can be completed to a hyperbolic quadric ${{\mathcal H}}$ with the addition of two lines through $\pi_\infty$ in $\Sigma\cap\Sigma_\infty$. Further, the [${{\cal C}}$-plane]{}s are of two types: the [${{\cal C}}$-plane]{}s through $V$ project to the generator lines of ${{\mathcal H}}$ not through $\pi_\infty$; and the [${{\cal C}}$-plane]{}s not through $V$ project to planes through $\pi_\infty$ which meet ${{\mathcal H}}$ in a conic. A [${{\cal C}}$-plane]{} through $V$ is projected onto a line of $\Sigma$ not through $\pi_\infty$. By Lemma \[cplane-meet-theorem\](2), there are two parallel classes ${\mathscr P}_1,{\mathscr P}_2$ of [${{\cal C}}$-plane]{}s through $V$, and [${{\cal C}}$-plane]{}s in the same parallel class pairwise meet in exactly $V$. Hence one parallel class ${\mathscr P}_1$ through $V$ is mapped to a set ${\mathcal T}_1$ of $q$ mutually skew lines (each line containing $q$ points of ${{\cal C}}'$), and the other parallel class ${\mathscr P}_2$ through $V$ is mapped to a set ${\mathcal T}_2$ of $q$ mutually skew lines (each line containing $q$ points of ${{\cal C}}'$), with each line from ${\mathcal T}_1$ meeting every line from ${\mathcal T}_2$ in a point of ${{\cal C}}'$. We complete these line sets into a regulus and its opposite regulus as follows. Consider the parallel class ${\mathscr P}_1$ of [${{\cal C}}$-plane]{}s through $V$. By Lemma \[cplane-parallel-classes\], the [${{\cal C}}$-line]{}s of these [${{\cal C}}$-plane]{}s lie in a common plane of ${{\Sigma_\infty}}$ through ${t_\infty}$. Hence they all meet $\Sigma\cap\Sigma_\infty$ in collinear points on a line $\ell_1$ through $\pi_\infty$. The line $\ell_1$ meets every [${{\cal C}}$-plane]{} in ${\mathscr P}_1$, and so meets every line in ${\mathcal T}_1$. Similarly we have a line $\ell_2$ through $\pi_\infty$ corresponding to the parallel class ${\mathscr P}_2$, and $\ell_2$ meets every line in ${\mathcal T}_2$. Thus the lines ${\mathcal T}_1\cup\ell_2$ form a regulus with opposite regulus ${\mathcal T}_2\cup\ell_1$. Hence ${{\cal C}}'\cup\ell_1\cup\ell_2$ is a hyperbolic quadric in $\Sigma$. So we have shown that the $2q$ [${{\cal C}}$-plane]{}s through $V$ project to $2q$ lines of the hyperbolic quadric ${{\mathcal H}}$. The remaining $q^2-q$ [${{\cal C}}$-plane]{}s not through $V$ project to planes through $\pi_\infty$ that contain $q$ points of ${{\cal C}}'$, hence meet ${{\mathcal H}}$ in a conic. Note that the remaining $2q+1$ planes through $\pi_\infty$ meet ${{\mathcal H}}$ in two lines of ${{\mathcal H}}$, with one of the lines necessarily a line through $\pi_\infty$. This proves the statement for Case 2. Recall that $\ell$ meets the spread lines $t_{A_1},\ldots,t_{A_q}$, we now consider the corresponding ${{\cal C}}$-points $A_i$, and their images $A_i'$ under the projection from $V$. If $\ell$ is a [${{\cal C}}$-line]{}, then the ${{\cal C}}$-points $A_1,\ldots,A_q$ lie on a common [${{\cal C}}$-plane]{} by Corollary \[cor:tA\]. So suppose $\ell$ is not a [${{\cal C}}$-line]{}. By Lemma \[l-meets-Pinfty\], the plane $\pi_i=\langle\ell,A_i\rangle$ contains exactly one [${{\cal C}}$-point]{}, namely $A_i$. So $\pi_i$ is mapped to a line through the points $A_i'$ and $L=\ell\cap(\Sigma\cap\Sigma_\infty)$. Note that $LA_i'$ meets ${{\cal C}}'$ in exactly one point, namely $A_i'$. The lines $LA_1',\ldots,LA_q'$ are distinct tangent lines to ${{\mathcal H}}$. Further, $L\pi_\infty$ is a tangent line to ${{\mathcal H}}$. Thus from a point $L\not\in{{\mathcal H}}$ we have a set of $q+1$ tangent lines of ${{\mathcal H}}$. Hence the points $\pi_\infty,A_1',\ldots,A_q'$ all lie on a plane through $\pi_\infty$ which meets ${{\mathcal H}}$ in a conic, namely the polar plane of $L$, see [@hirs85 Theorem 15.3.16]. Hence by the above argument, this plane is the image of a [${{\cal C}}$-plane]{}, and hence the points $A_i$ lie on a common [${{\cal C}}$-plane]{}. That is, the lemma also holds in the case when $V$ is an $\infty$-point. We now show that any regulus containing ${t_\infty}$ and two other lines of the spread ${{\cal S}}$ is contained in ${{\cal S}}$. Next we will use the Klein quadric to show that a spread with this property is regular. [partial-spread]{} Let $t_A,t_B$ be two elements of the spread ${{\cal S}}$. Then the unique regulus determined by the three lines ${t_\infty},t_A,t_B$ is contained in ${{\cal S}}$. Let $t_A,t_B$ be two elements of the spread ${{\cal S}}$. Through each point $V_i\in {t_\infty}$, $i=0,\ldots,q$, there is a unique line $\ell_i$ that meets both $t_A$ and $t_B$. Further, the lines $\ell_i$ form a regulus ${{\cal R}}'$. The opposite regulus ${{\cal R}}$ is the unique regulus containing ${t_\infty},t_A,t_B$. We want to show that ${{\cal R}}\subset {{\cal S}}$. Consider the ${{\cal C}}$-points $A,B$ corresponding to $t_A,t_B$ respectively. By (A2), they lie in a unique [${{\cal C}}$-plane]{} $\pi$. Now $\pi$ meets ${{\Sigma_\infty}}$ in a [${{\cal C}}$-line]{}  $\ell$, and $\ell$ meets ${t_\infty}$ by Lemmas \[cplane-meet-theorem\](1) and \[thm-tinfty\]. Further $\ell$ meets $t_A$ and $t_B$ by Corollary \[cor:tA\]. Hence $\ell$ is one of the lines $\ell_i\in{{\cal R}}'$. Now $\ell$ meets $q+1$ lines of the spread ${{\cal S}}$, denote them ${t_\infty},t_A,t_B,t_{C_3},\ldots,t_{C_q}$. We want to show that these are the lines of ${{\cal R}}$. Note that the corresponding ${{\cal C}}$-points $A,B,C_3,\ldots,C_q$ lie on the [${{\cal C}}$-plane]{} $\pi$ by Corollary \[cor:tA\]. Now consider the line $\ell_j\in{{\cal R}}'$, $\ell_j\neq\ell$, it meets $q+1$ lines of the spread ${{\cal S}}$, denote these by ${t_\infty},t_A,t_B,t_{D_3},\ldots,t_{D_q}$. By Lemma \[proj-result\], the corresponding ${{\cal C}}$-points $A,B,D_3,\ldots,D_q$ lie on a common [${{\cal C}}$-plane]{}, $\alpha$ say. As there is a unique [${{\cal C}}$-plane]{} containing $A,B$, we have $\alpha=\pi$, and so $\{C_3,\ldots,C_q\}=\{D_3,\ldots,D_q\}$. Hence each line $\ell_i$ in the regulus ${{\cal R}}'$ meets the spread lines ${t_\infty},t_A,t_B,t_{C_3},\ldots,t_{C_q}$, and so ${{\cal R}}=\{{t_\infty},t_A,t_B,t_{C_3},\ldots,t_{C_q}\}$, and so ${{\cal R}}\subset{{\cal S}}$. That is, the unique regulus containing ${t_\infty},t_A,t_B$ is contained in ${{\cal S}}$. To prove that ${{\cal S}}$ is a regular spread we now show that a spread satisfying the conditions of Lemma \[partial-spread\] is regular. We will use the Klein correspondence from the set of lines in ${{\rm{PG}}}(3,q)$ to the set of points of the Klein quadric ${{\mathcal H}}_5$, a hyperbolic quadric in ${{\rm{PG}}}(5,q)$. For details of this correspondence, see [@hirs85 Section 15.4]. We note that a regular spread of ${{\rm{PG}}}(3,q)$ corresponds under the Klein correspondence to a 3-dimensional elliptic quadric contained in ${{\mathcal H}}_5$, that is, the intersection of ${{\mathcal H}}_5$ and a 3-space that forms an elliptic quadric. Further, a regulus of ${{\rm{PG}}}(3,q)$ corresponds to a conic in ${{\mathcal H}}_5$. [Sisregular]{} Let ${{\cal S}}$ be a spread of ${{\rm{PG}}}(3,q)$ with a special line ${t_\infty}$ with the property that the regulus containing ${t_\infty}$ and any further two lines of ${{\cal S}}$ is contained in ${{\cal S}}$. Then ${{\cal S}}$ is a regular spread. We begin with some notation. Let ${{\cal S}}'$ be the points on the Klein quadric ${{\mathcal H}}_5$ corresponding to the lines of the spread ${{\cal S}}$. For each line $t_A$ of ${{\cal S}}$, let $t_A'$ denote the corresponding point of ${{\mathcal H}}_5$. Similarly, if ${{\cal R}}$ is a regulus of ${{\rm{PG}}}(3,q)$, let ${{\cal R}}'$ denote the set of points of ${{\mathcal H}}_5$ corresponding to the lines of ${{\cal R}}$, and note that the points in ${{\cal R}}'$ lie on a conic. For three mutually skew lines $\ell,m,n$ in ${{\rm{PG}}}(3,q)$, let ${\mathscr R}(\ell,m,n)$ denote the unique regulus containing them. We first show that all the points of ${{\cal S}}'$ are contained in a common 3-space. We will repeatedly use the assumption that the regulus determined by ${t_\infty}$ and two other lines of ${{\cal S}}$ is contained in ${{\cal S}}$. Fix a spread element $t_A\ne {t_\infty}$ and consider any two distinct reguli ${{\cal R}}_1,{{\cal R}}_2$ of ${{\cal S}}$ containing ${t_\infty}$ and $t_A$. Suppose that $$\begin{aligned} {{\cal R}}_1&={\mathscr R}({t_\infty},t_A,t_{B_1})=&\{{t_\infty},t_A,t_{B_1},\ldots,t_{B_{q-1}}\}\\ {{\cal R}}_2&={\mathscr R}({t_\infty},t_A,t_{C_1})=&\{{t_\infty},t_A,t_{C_1},\ldots,t_{C_{q-1}}\}.\end{aligned}$$ Two distinct reguli have at most two common lines, so ${{\cal R}}_1$ and ${{\cal R}}_2$ intersect in exactly ${t_\infty},t_A$. Hence in the Klein quadric, we have two conics ${{\cal R}}_1',{{\cal R}}_2'$, they lie in two distinct planes which meet in the line ${t_\infty}'t_A'$ and hence span a 3-space denoted $\Sigma$. Now consider the regulus ${{\mathcal T}}$ of ${{\cal S}}$ determined by ${t_\infty},t_{B_1},t_{C_1}$: $${{\mathcal T}}={\mathscr R}({t_\infty},t_{B_1},t_{C_1})=\{{t_\infty},t_{B_1},t_{C_1},t_{D_3},\ldots,t_{D_q}\}.$$ Note that the lines $t_{B_i},t_{C_j},t_{D_k}$ are all distinct. In ${{\rm{PG}}}(5,q)$, ${{\mathcal T}}'$ is a conic that contains three points ${t_\infty}',t_{B_1}',t_{C_1}'$ of $\Sigma$, hence ${{\mathcal T}}'\subset\Sigma$. We now use the lines $t_{D_3},\ldots,t_{D_q}$ of ${{\mathcal T}}$ to construct $q-3+1$ more reguli of ${{\cal S}}$ through ${t_\infty},t_A$: $$\begin{aligned} {{\cal R}}_i&=&{\mathscr R}({t_\infty},t_A,t_{D_i}),\quad i=3,\ldots,q.\end{aligned}$$ We have a set $\{{{\cal R}}_1,\ldots,{{\cal R}}_q\}$ of $q$ reguli of ${{\cal S}}$ that pairwise intersect in exactly the lines ${t_\infty},t_A$, so they cover $2+q(q-1)=q^2-q+2$ elements of ${{\cal S}}$. The remaining $q-1$ spread elements $t_{E_1},\ldots,t_{E_{q-1}}$ of ${{\cal S}}$ lie on a common regulus ${{\cal U}}$ through ${t_\infty},t_A$ (as every three elements determine a unique regulus), that is, $${{\cal U}}=\{{t_\infty},t_A,t_{E_1},\ldots,t_{E_{q-1}}\}.$$ Now each reguli ${{\cal R}}_i$, $i=3,\ldots,q$, is mapped to a conic ${{\cal R}}_i'$ of ${{\mathcal H}}_5$. Further each conic ${{\cal R}}_i'$, $i=3,\ldots,q$ contains three points ${t_\infty}',t_A',t_{D_i}'$ of $\Sigma$. Hence ${{\cal R}}_i'\subset \Sigma$, $i=1,\ldots,q$. To show that ${{\cal S}}'\subset\Sigma$, it remains to show that $t_{E_1}',\ldots,t_{E_{q-1}}'\in\Sigma$. Now consider the two reguli ${{\cal R}}_1$, ${{\cal U}}$ of ${{\cal S}}$. They map to two conics ${{\cal R}}_1',{{\cal U}}'$ of ${{\mathcal H}}_5$ that span a 3-space denoted by $\Sigma'$. Consider another regulus of ${{\cal S}}$: $${{\mathcal V}}={\mathscr R}({t_\infty},t_{B_1},t_{E_1})=\{{t_\infty},t_{B_1},t_{E_1},t_{F_3},\ldots,t_{F_q}\}.$$ Then ${{\mathcal V}}'$ is a conic of ${{\mathcal H}}_5$ with three points ${t_\infty}',t_{B_1}',t_{E_1}'$ in $\Sigma'$, and so ${{\mathcal V}}'$ is contained in $\Sigma'$. As ${{\cal U}},{{\mathcal V}}$ meet exactly in ${t_\infty},t_{E_1}$, the lines $t_{F_i}$, $i=3,\ldots,q$ are distinct from the lines $t_{E_i}$, $i=1,\ldots,q-1$ and so belong to the $q^2-q+2$ elements of ${{\cal S}}$ in $\Sigma$. Hence we have $t_{F_i}\subseteq\Sigma$ and so $t_{F_i}'\in\Sigma'\cap\Sigma$, $i=3,\ldots,q$. Thus $\Sigma\cap\Sigma'$ contains ${{\cal R}}_1'$ and ${{\mathcal V}}'$ which is more than a plane, and so $\Sigma=\Sigma'$. That is, $t_{E_1}',\ldots,t_{E_{q-1}}'\in\Sigma$. Thus the lines of ${{\cal S}}$ are mapped into points of ${{\mathcal H}}_5$ that lie in a 3-space $\Sigma$. The intersection ${{\cal Q}}$ of the 3-space $\Sigma$ with ${{\mathcal H}}_5$ is a quadric of $\Sigma$, hence is either an elliptic, hyperbolic or a degenerate quadric. Note that in all cases other than the elliptic quadric, ${{\cal Q}}$ does not correspond to a spread (or a set containing a spread) of ${{\rm{PG}}}(3,q)$. Hence ${{\cal Q}}$ is an elliptic quadric, and so ${{\cal S}}$ is a regular spread. As an immediate consequence of Lemma \[partial-spread\] and Theorem \[Sisregular\], we have that the spread ${{\cal S}}$ constructed in Theorem \[lines-form-a-spread\] is regular. [cor:regular-spread]{} The spread ${{\cal S}}=\{t_A{:}A\in{{\cal C}}\}\cup\{{t_\infty}\}$ is regular. ${{\cal C}}$ gives rise to an arc in ${{\rm{PG}}}(2,q^2)$ --------------------------------------------------------- By Corollary \[cor:regular-spread\], we have a regular spread ${{\cal S}}=\{t_A{:}A\in{{\cal C}}\}\cup\{{t_\infty}\}$ in ${{\Sigma_\infty}}$ from which we can construct a Desarguesian plane ${\mathcal{P}}({{\cal S}})\cong{{\rm{PG}}}(2,q^2)$ via the Bruck-Bose correspondence. Let ${{\mathcal O}}$ be the set of points in ${\mathcal{P}}({{\cal S}})$ corresponding to the affine points of ${{\cal C}}$ together with the point $T_\infty$ on ${\ell_{\infty}}$ corresponding to the spread line ${t_\infty}$. [thm:isanarc]{} ${{\mathcal O}}$ is a conic in ${\mathcal{P}}({{\cal S}})\cong{{\rm{PG}}}(2,q^2)$. We show that ${{\cal O}}={{\cal C}}\cup T_\infty$ is a $(q^2+1)$-arc in ${{\rm{PG}}}(2,q^2)$, and hence a conic. A line through $T_\infty$ in ${{\rm{PG}}}(2,q^2)$ corresponds to an affine plane of ${{\rm{PG}}}(4,q)$ that contains the spread line ${t_\infty}$. By Lemma \[lemma-plane-thru-tinfty\], the affine planes of ${{\rm{PG}}}(4,q)$ through ${t_\infty}$ meet ${{\cal C}}$ in exactly one point. Hence in ${{\rm{PG}}}(2,q^2)$, a line through $T_\infty$ meets ${{\mathcal O}}$ in at exactly one further point. Let $t_A$ be a line of the spread ${{\cal S}}$ in ${{\rm{PG}}}(4,q)$ with corresponding ${{\cal C}}$-point $A$. Let $\alpha$ be an affine plane of ${{\rm{PG}}}(4,q)$ through the spread line $t_A$ that contains three ${{\cal C}}$-points $P,Q,R$. We obtain a contradiction to show that this is not possible. Note that $\alpha\neq\langle A,t_A\rangle$ since by Corollary \[cor:tA2\], the plane $\langle A,t_A\rangle$ contains exactly one ${{\cal C}}$-point. By (A2), $P,Q$ lie in a unique ${{\cal C}}$-plane $\pi_{PQ}$ which meets $t_A$ in the point $PQ\cap t_A$. Hence by Corollary \[cor:tA\], $\pi_{PQ}$ contains the point $A$. Similarly, $P,R$ lie on a unique [${{\cal C}}$-plane]{} $\pi_{PR}$ that contains $A$. Hence we have two distinct [${{\cal C}}$-plane]{}s $\pi_{PQ},\pi_{PR}$ that both contain the two distinct points $A,P$ of ${{\cal C}}$, contradicting (A2). Hence any affine plane that contains a spread line $t_A$ contains at most two points of ${{\cal C}}$. Thus in the Bruck-Bose plane ${{\rm{PG}}}(2,q^2)$, a line through a point of ${\ell_{\infty}}\setminus T_\infty$ meets ${{\mathcal O}}$ in at most two points. Further, ${\ell_{\infty}}$ meets ${{\mathcal O}}$ in one point, so we have shown that ${{\mathcal O}}$ is a $(q^2+1)$-arc in ${{\rm{PG}}}(2,q^2)$. As $q$ is odd, by Segre [@segr55], ${{\mathcal O}}$ is a conic in ${{\rm{PG}}}(2,q^2)$. This almost completes the proof of Theorem \[mainthm\]. It remains to show that the spread ${{\cal S}}$ is unique, which we do in the next section. The spread ${{\cal S}}$ is unique --------------------------------- We now show that the spread ${{\cal S}}=\{t_A{:}A\in{{\cal C}}\}\cup\{{t_\infty}\}$ constructed in Theorem \[lines-form-a-spread\] is the only spread in ${{\Sigma_\infty}}$ for which the ${{\cal C}}$-points give rise to an arc in the Bruck-Bose plane ${\mathcal{P}}({{\cal S}})$. [thm:diffspread]{} Let ${{\cal S}}'$ be a spread of ${{\Sigma_\infty}}$ distinct from the spread ${{\cal S}}=\{t_A{:}A\in{{\cal C}}\}\cup\{{t_\infty}\}$. Then in the associated Bruck-Bose plane ${\mathcal{P}}({{\cal S}}')$, the set of points corresponding to ${{\cal C}}$ do not form an arc. Let $\ell$ be a line in ${{\Sigma_\infty}}$ disjoint from ${t_\infty}$ and assume that all the affine planes through $\ell$ meet ${{\cal C}}$ in at most two points. We will show that $\ell$ must be one of the spread lines $t_A$ for some point $A\in{{\cal C}}$. Let $X_1$ be a point on $\ell$, by Lemma \[lem:pttype\] and  \[thm-tinfty\], there exists a unique [${{\cal C}}$-line]{} $\ell_{1}$ through $X_1$. By Corollary \[count-cline\], $\ell_1$ is contained in a unique [${{\cal C}}$-plane]{} $\pi_{1}$. Let ${\mathcal E}_1$ be the conic in $\pi_{1}$, that is, ${\mathcal E}_1=({{\cal C}}\cap\pi_{1})\cup(\pi_{1}\cap {t_\infty})$. Now $X_1$ is on one tangent to ${\mathcal E}_1$, namely $\ell_{1}$, hence $X_1$ is on a second tangent $m_{1}$ to ${\mathcal E}_1$. There are $(q-1)/2$ 2-secants to ${\mathcal E}_1$ through $X_1$, each together with the line $\ell$ determines an affine plane, we label these planes $\alpha_1,\ldots,\alpha_{(q-1)/2}$. As each plane contains a 2-secant to ${\mathcal E}_1$ through $X_1$, each contains at least two [${{\cal C}}$-point]{}s, and hence by our assumption about $\ell$, contains exactly two [${{\cal C}}$-point]{}s. Similarly, let $X_2$ be a point on $\ell$, lying on a unique ${{\cal C}}$-line $\ell_2$, defining the unique [${{\cal C}}$-plane]{} $\pi_2$, and construct planes $\beta_1,\ldots,\beta_{(q-1)/2}$ which contain $\ell$ and exactly two [${{\cal C}}$-point]{}s. Suppose $\alpha=\alpha_i=\beta_j$ for some $i,j$. Then $\alpha$ contains at least three ${{\cal C}}$-points (as $\alpha_i$ contains a 2-secant of ${{\cal C}}$ through $X_1$, $\beta_j$ contains a 2-secant of ${{\cal C}}$ through $X_2$, and $X_1\neq X_2$). This contradicts our assumption that planes about $\ell$ contain at most two ${{\cal C}}$-points. Hence the planes $\alpha_1,\ldots,\alpha_{(q-1)/2}, \beta_1,\ldots,\beta_{(q-1)/2}$ are distinct, and cover $(q-1)\times 2$ distinct ${{\cal C}}$-points. Repeating this for the remaining points $X_3,\ldots,X_{q+1}$ of $\ell$, we obtain a set $ \mathcal K$ of distinct planes about $\ell$ (including the $\alpha_i,\beta_j$) of size $(q+1)\times (q-1)/2$. Each plane in $\mathcal K$ contains exactly two [${{\cal C}}$-point]{}s, accounting for $q^2-1$ points of ${{\cal C}}$ that lie on planes through $\ell$. Now for each $i$, consider the unique affine tangent line $m_{i}$ through $X_i$ (distinct from $\ell_i$) to the conic ${\mathcal E}_i$, in the plane $\pi_{i}$. We have a set of $q+1$ planes $\langle \ell,m_{i}\rangle$, not necessarily distinct, but distinct from the planes of $\mathcal K$, each meeting ${{\cal C}}$ in at least one point. However, there is only one [${{\cal C}}$-point]{} unaccounted for by the planes of $\mathcal K$, so the planes $\langle \ell,m_{i}\rangle$ are all the same plane $\alpha$ which contains exactly one point $Z$ of ${{\cal C}}$. The $q+1$ lines $ZX_1,\ldots,ZX_{q+1}$ through $Z$ in $\alpha$ are respectively the tangents $m_1,\ldots,m_{q+1}$ to the conic ${\mathcal E}_1,\ldots,{\mathcal E}_{q+1}$ in $\pi_{1},\ldots,\pi_{q+1}$ at $Z$. Further, the tangents $m_1,\ldots,m_{q+1}$ meet $\Sigma_\infty$ at points $X_1,\ldots,X_{q+1}$ of $\ell$, so by Definition \[def:tA\], $\ell=t_Z$. That is, $\ell$ is one of the spread lines of ${{\cal S}}$. We conclude that every line in ${{\Sigma_\infty}}$ which is disjoint from ${t_\infty}$ and not in ${{\cal S}}$ lies on an affine plane that contains at least three [${{\cal C}}$-point]{}s. Suppose ${{\cal S}}'$ is a spread of ${{\Sigma_\infty}}$ distinct from ${{\cal S}}$ such that in the Bruck-Bose plane ${\mathcal{P}}({{\cal S}}')$ the ${{\cal C}}$-points form an arc. Let ${{\cal S}}'\setminus{{\cal S}}$ denote the set of lines in ${{\cal S}}'$ that are not in ${{\cal S}}$. If ${{\cal S}}'\setminus{{\cal S}}$ contains a line $\ell$ disjoint from $t_\infty$, then by the above argument, $\ell$ lies on some affine plane $\alpha$ that contains at least three ${{\cal C}}$-points. In the Bruck-Bose plane ${\mathcal{P}}({{\cal S}}')$, $\alpha$ corresponds to a line that contains three ${{\cal C}}$-points, so ${{\cal C}}$ is not an arc, contradicting our assumption. Hence ${{\cal S}}'\setminus {{\cal S}}$ cannot contain a line $\ell$ disjoint from ${t_\infty}$, thus the lines in ${{\cal S}}'\setminus {{\cal S}}$ must meet ${t_\infty}$. So if ${{\cal S}}'\neq{{\cal S}}$, then ${{\cal S}}'\setminus{{\cal S}}$ contains $q+1$ lines that meet $t_\infty$ in a point. Let $\ell\in{{\cal S}}'\setminus{{\cal S}}$, so $\ell$ meets ${t_\infty}$ and another $q$ spread lines $t_{A_1},\ldots,t_{A_q}$ of ${{\cal S}}$. Let ${{\cal R}}=\{ {t_\infty},t_{A_1},\ldots,t_{A_q}\}$, then the lines of ${{\cal S}}'\setminus{{\cal S}}$ cover the same points as the lines of ${{\cal R}}$ do. Let ${{\cal S}}'\setminus{{\cal S}}=\{\ell,\ell_1,\ldots,\ell_q\}$, then each line $\ell_i$ meets each line in ${{\cal R}}$. Let $t_A$, $t_B$ be any two lines of ${{\cal R}}$ distinct from $t_\infty$. There are exactly $q+1$ lines meeting ${t_\infty},t_A,t_B$, and since $\ell,\ell_1,\ldots,\ell_q$ meet each of ${t_\infty},t_A,t_B$, they form a regulus ${{\cal R}}'$. A similar argument with three elements of ${{\cal R}}'$ shows that ${{\cal R}}$ is the opposite regulus of ${{\cal R}}'$. By Corollary \[cor:tA\], there exists a unique [${{\cal C}}$-line]{} $m$ meeting ${t_\infty},t_A,t_B$, namely the [${{\cal C}}$-line]{}  corresponding to the unique [${{\cal C}}$-plane]{} $\pi_{AB}$ through $A$ and $B$. The line $m$ meets three lines of ${{\cal R}}$ and hence it is a line of ${{\cal R}}'$. That is, $m$ is a line of ${{\cal S}}'$. However, the ${{\cal C}}$-plane $\pi_{AB}$ through $m$ corresponds to a line of ${\mathcal{P}}({{\cal S}}')$ that contains $q$ points of ${{\cal C}}$. Hence ${{\cal C}}$ is not an arc of ${\mathcal{P}}({{\cal S}}')$. Thus the only spread which gives rise to an arc in the corresponding Bruck-Bose plane is the regular spread ${{\cal S}}$ constructed in Theorem \[lines-form-a-spread\]. This completes the proof of Theorem \[mainthm\]. Conclusion ========== In this paper we characterised sets in ${{\rm{PG}}}(4,q)$, $q$ odd, $q\geq 7$ satisfying the combinatorial properties given in Theorem \[mainthm\] as corresponding via the Bruck-Bose correspondence to conics in ${{\rm{PG}}}(2,q^2)$. We note that a similar characterisation when $q$ is even is given in [@conicqeven]. The cases when $q=3$ or $5$ are still open. An interesting geometric question arises from the properties of a conic given in Lemma \[conic-satisfies-props\]. Let ${{\cal C}}$ be a conic in ${{\rm{PG}}}(2,q^2)$, $q$ odd, tangent to ${\ell_{\infty}}$, and let $\pi$ be a ${{\cal C}}$-plane. By property 3 of Lemma \[conic-satisfies-props\], the points of $\pi$ that are not in ${{\cal C}}$ lie on exactly one more ${{\cal C}}$-plane. Let $P$ be a point of $\pi\setminus{{\cal C}}$. If $P$ is an interior point of the subconic $\pi_{{\cal C}}=\pi\cap{{\cal C}}$, then we can use the polarity of $\pi_{{\cal C}}$ to construct the second ${{\cal C}}$-plane containing $P$. If $P$ is an exterior point of $\pi_{{\cal C}}$, then it would be interesting to have a geometric construction of the second ${{\cal C}}$-plane containing $P$. [99]{} J. André. Über nicht-Desarguessche Ebenen mit transitiver Translationsgruppe, [*Math. Z.*]{} [**60**]{} (1954), 156–186. A. Barlotti. [*Un’estensione del teorema di Segre-Kustaaheimo.*]{} Boll. U.M.I., [**10**]{} (1955) 498-506. S.G. Barwick and G.L. Ebert. Unitals in projective planes. Springer Monographs in Mathematics. Springer, New York, 2008. S.G. Barwick and W.A. Jackson. A characterisation of translation ovals in finite even order planes. http://arxiv.org/abs/1305.6673 S.G. Barwick, W.A. Jackson and C.T. Quinn. Conics and caps. [*J. Geom.*]{} 100 (2011) 15–28. R.H. Bruck and R.C. Bose. The construction of translation planes from projective spaces, [*J. Algebra*]{} [**1**]{} (1964) 85–102. R.H. Bruck and R.C. Bose. Linear representations of projective planes in projective spaces, [*J. Algebra*]{} [**4**]{} (1966) 117–172. J.W.P. Hirschfeld. [*Finite Projective Spaces of Three Dimensions.*]{} Oxford University Press, 1985. J.W.P. Hirschfeld. [*Projective Geometry over Finite Fields, Second Edition.*]{} Oxford University Press, 1998. G. Panella. Caratterizzazione delle quadriche di uno spazio (tri-dimensionale) lineare sopra un corpo finito. [*Boll. Un. Mat. Ital.*]{}, [**10**]{} (1955) 507–513. C.T. Quinn. The André/Bruck and Bose representation of conics in Baer subplanes of ${{\rm{PG}}}(2, q^2)$. [*J. Geom*]{} [**74**]{} (2002) 123-138. B. Segre. Ovals in a finite projective plane. [*Can. J. Math.*]{}, [**7**]{} (1955) 414–416. B. Segre. Introduction to Galois geometries. [*Atti Accad. Naz. Lincei Mem.*]{}, [**8**]{} (1967) 133–236. (edited by J.W.P. Hirschfeld.)

ArXiv

--- author: - | , Gunnar Bali, Andrea Nobile, Andreas Schäfer\ Institut für Theoretische Physik, Universität Regensburg,\ 93040 Regensburg, Germany\ E-mail:\ \ \ - | Yoshifumi Nakamura\ Center for Computational Sciences, University of Tsukuba,\ Tsukuba, Ibaraki 305-8577, Japan\ E-mail: - | James Zanotti\ School of Physics, University of Edinburgh,\ Edinburgh EH9 3JZ, UK\ E-mail: - '(QCDSF Collaboration)' title: Disconnected contributions to hadronic structure --- Introduction ============ In the last few years there has been an upsurge in interest in calculating the scalar matrix element on the lattice [@Ohki:2009mt; @Babich:2009rq; @Freeman:2009pu; @ramos; @jung] either directly, by calculating the corresponding connected and disconnected terms, or indirectly, via the Feynman-Hellman theorem. High statistics for dynamical simulations mean that reasonable signals can be obtained for disconnected terms and similarly the small statistical uncertainty on nucleon mass as a function of the quark masses enable reasonable fits to be made. Ideally, the results of both approaches should agree. Such calculations have also become particularly timely since the advent of the LHC because the scalar coupling $f_{T_q}=m_q\langle N|\bar{q}{q}|N\rangle/m_N$ determines the fraction of the proton mass $m_N$ that is carried by quarks of flavour $q$. The strength of the coupling of the Standard Model (SM) Higgs boson or of any similar scalar particle to the proton is mainly determined by $\sum_q f_{T_q}$ for $q\in\{u,d,s\}$. Therefore, an accurate calculation of these quantities will help to increase the precision of SM phenomenology and to shed light on non-SM processes. The spin of the nucleon can be decomposed into a quark spin contribution $\Delta\Sigma=\Delta u + \Delta d+ \Delta s+\ldots$, a quark angular momentum contribution $L_q$ and a gluonic contribution (spin and angular momentum) $\Delta G$: $$\frac12=\frac12 \Delta\Sigma+L_q+\Delta G\,.$$ Experimentally, $\Delta s$ is not well determined: HERMES obtained [@Airapetian:2007mh] $\Delta s=-0.085(13)(8)(9)$ in the $\overline{MS}$ scheme. However, the signal is dominated by contributions in the small $x$ region where models are used to extrapolate from the experimental results obtained at larger $x$. In these proceedings we present an update of an on-going project to calculate $f_{T_q}$ and $\Delta q$. In particular, the following improvements have been implemented since Lattice 2009 [@Bali:2009dz]: - Statistics on the $24^3\times 48$ and $32^3\times 64$ lattices have been significantly increased, by factors of roughly two and three, respectively, using the SFB/TR55 QPACE computers [@nobile; @yoshifumi] (details are given in the next section). - An additional volume of $40^3\times 64$ has been analyzed. The analysis is not yet finalized and we plan to further increase the statistics for the larger two volumes. The results presented here are preliminary. Simulation details ================== The simulations were performed on $n_{\mathrm f}=2$ configurations of nonperturbatively improved clover fermions with Wilson gauge action at $\beta=5.29$ and $\kappa_{\mathrm{sea}}=0.13632$. Details of the volumes and the number of trajectories analyzed are given in table \[configs\]. The pseudoscalar mass corresponding to this $\kappa_{\mathrm{sea}}$ value is around $270$ MeV, using an inverse lattice spacing of $2.59$ GeV determined from $r_0(\beta,\kappa)=0.467$ fm. ![The connected (left) and disconnected (right) diagrams associated with the scalar ($\Gamma=\mathbb{1}$) and axial vector matrix elements ($\Gamma=\gamma_5\gamma_i$).[]{data-label="diags"}](conn_discon.eps){height=".2\textwidth"} The scalar and axial-vector matrix elements we are interested in are extracted on the lattice from the three-point functions corresponding to the diagrams given in fig. \[diags\]. The axial-vector matrix element is related to $\Delta q$ through, $$\begin{aligned} \langle N,s|\bar{q}\gamma_{\mu}\gamma_5 q|N,s\rangle & = & 2m_N s_{\mu} \frac{\Delta q}{2},\end{aligned}$$ in Minkowski space notation, where $m_N$ is the nucleon mass and $s_\mu$ its spin ($s_\mu^2=-1$). The connected three-point functions were calculated separately by QCDSF, details of which can be found in refs. [@qcdsf; @Gockeler:2009pe]. The study presented here is only concerned with the disconnected terms. For $\Delta s$ and $\langle N|\bar{s}s|N\rangle$ only these terms contribute. We varied the quark mass of the current insertion (disconnected loop), as well as the mass of the valence quarks in the nucleon. In the following, we denote the $\kappa$ value corresponding to the quark loop and the valence quarks in the nucleon as $\kappa_{\mathrm{loop}}$ and $\kappa_{\mathrm{val}}$, respectively. All combinations $\kappa_{\mathrm{loop}}, \kappa_{\mathrm{val}}\in\{0.13550, 0.13609, 0.13632\}$ were used. These values correspond to the pseudoscalar masses, $m_{\mathrm{PS}}\approx 690$, $440$ and $270$ MeV, respectively. The heaviest $\kappa_{\mathrm{val}}=0.13550$ roughly corresponds to the strange quark mass. The disconnected contributions to $\Delta q$ and $\langle N|\bar{q}q|N\rangle$ were extracted from the ratios of three-point functions to two-point functions (at zero momentum), $$\label{eq:rati} R^{\mathrm{dis}}(t_{\mathrm f},t,t_{\mathrm i}) = - \frac{\mathrm{Re}\,\left\langle\Gamma_{\mathrm{2pt}}^{\alpha\beta}C^{\beta\alpha}_{\mathrm{2pt}}(t_{\mathrm f},t_{\mathrm i}) \sum_{\mathbf{x}}\mathrm{Tr}\,(M^{-1}(\mathbf{x},t;\mathbf{x},t)\Gamma_{\mathrm{loop}})\right\rangle}{\left\langle \Gamma_{\mathrm{unpol}}^{\alpha\beta} C^{\beta\alpha}_{\mathrm{2pt}}(t_{\mathrm f},t_{\mathrm i})\right\rangle}\,,$$ where the nucleon source and sink are at $t_{\mathrm i}$ and $t_{\mathrm f}$ respectively, and the current is inserted at $t$. The three-point function is simply the combination of the nucleon two-point function, $C_{\mathrm{2pt}}(t_{\mathrm f},t_{\mathrm i})$, and the disconnected loop, $ \sum_{\mathbf{x}}\mathrm{Tr}\,[M^{-1}(\mathbf{x},t;\mathbf{x},t)\Gamma_{\mathrm{loop}}]$. For the scalar matrix element we used, $\Gamma_{\mathrm{2pt}}=\Gamma_{\mathrm{unpol}}:=(\mathbb{1}+\gamma_4)/2$ and $\Gamma_{\mathrm{loop}}=\mathbb{1}$. For $\Delta q$ we calculated the difference between two polarizations: $\Gamma_{\mathrm{2pt}}= \gamma_j\gamma_5\Gamma_{\mathrm{unpol}}$ and $\Gamma_{\mathrm{loop}}=\gamma_j\gamma_5$, where we average over all three possible $j$-orientations. In the limit of large times, $t_{\mathrm f}\gg t\gg t_{\mathrm i}$, depending on the $\Gamma$-combination used, $R^{\mathrm{dis}}$ will either approach the disconnected axial matrix element $\Delta q^{\mathrm{dis}}$ or the disconnected scalar matrix element $\langle N|\bar{q}q|N\rangle^{\mathrm{dis}}$ (once the vacuum contribution is subtracted). However, the statistical noise increases rapidly with increasing $t-t_{\mathrm i}$ and this time difference needs to be minimized, using smeared sources and sinks for the nucleon, in order to obtain a reasonable signal. A smearing study on a limited number of configurations indicated that the nucleon plateaued around $t\ge 4a\approx 0.3\,$fm and we chose to insert the current at this timeslice ($t_{\mathrm i}=0$). However, the higher statistics now available mean the ground state dominates around $t\ge 6a$ only. At zero momentum, the excited state contribution to $R^{\mathrm{dis}}$ is governed by the time difference, $t_{\mathrm f}-t_{\mathrm i}$, and we must be careful to choose $t_{\mathrm f}$ large enough. In fig. \[rdis\] we show the results for $R^{\mathrm{dis}}$ as a function of $t_{\mathrm f}>t$ for all three volumes studied and the heaviest $\kappa_{\mathrm{val}}=\kappa_{\mathrm{loop}}=0.13550$ combination. So far we have chosen $t_{\mathrm f}=8a$ or $9a$ based on the quality of the plateau within given statistical errors. This depends on the observable and lattice volume. Once final statistics are reached we will fit the three- and two-point functions within $R^{\mathrm{dis}}$ of eq. (\[eq:rati\]) separately, as functions of $t_{\mathrm f}$, in order to extract the asymptotic values. The disconnected loop, $ \sum_{\mathbf{x}}\mathrm{Tr}\,[M^{-1}(\mathbf{x},t;\mathbf{x},t)\Gamma_{\mathrm{loop}}]$, was calculated using stochastic estimates, together with several noise reduction techniques: - Partitioning [@Bernardson:1993yg; @Wilcox:1999ab]: the stochastic source has support on eight timeslices. Additional two-point functions were generated for four time separated source points on each configuration. The forward and backward propagation from these four source points was combined with the loop to give us eight measurements of $R^{\mathrm{dis}}$ per configuration. - Hopping parameter expansion [@Thron:1997iy]: for the clover action the first two terms in the expansion of the disconnected loop, $$\mathrm{Tr}(M^{-1}\Gamma_{\mathrm{loop}})= 2\kappa\mathrm{Tr}[(\mathbb{1}-\kappa{\! \not \!\! D})^{-1}\Gamma_{\mathrm{loop}}]= \mathrm{Tr}[(2\kappa\mathbb{1}+2\kappa^2{\! \not \!\! D}+\kappa^2{\! \not \!\! D}^2M^{-1})\Gamma_{\mathrm{loop}}]\,,$$ vanish and hence only contribute to the noise. This means $\mathrm{Tr}\,[\kappa^2{\! \not \!\! D}^2M^{-1}(\mathbf{x},t;\mathbf{x},t)\Gamma_{\mathrm{loop}}]$ can be used as an improved estimate of the loop. (In the case of $\Gamma_{\mathrm{loop}}=\mathbb{1}$ the non-vanishing first term $\sum_{\mathbf{x}}2\kappa\mathrm{Tr}\,\mathbb{1}=24\kappa L^3$ can easily be corrected for.) - Truncated solver method [@Bali:2009hu]: $730$ conjugate gradient solves were used, where the solver was truncated after $40$ iterations. $50$ BiCGStab solves running to full convergence were generated to correct for the truncation error. The noise reduction techniques other than time partitioning are only necessary for determining $\Delta q$; for the scalar matrix element the gauge noise dominates. The scalar matrix element: $f_{T_s}$ and $m_q\langle N|\bar{q}q|N\rangle^{\mathrm{dis}}$ ======================================================================================== The results for $f_{T_s}$ are presented in fig. \[scalar\] as functions of $m_{\mathrm{PS}}^2$ corresponding to the mass of the (valence) quarks in the nucleon. No renormalization is required as the combination, $m_q\langle N|\bar{q}q|N\rangle$, is scale and scheme independent. There is consistency between the values obtained on the different volumes for the heaviest valence quark mass, however, the spread between the results increases when the quark mass is reduced. Whether this is an indication of finite size effects will be clarified once the statistics for the $40^3$ volume is increased. In fig. \[scalar\] we also display the results for the disconnected scalar matrix element, $m_q\langle N |\bar{q}q|N\rangle$, for the $32^3$ volume as a function of $m_{\mathrm{PS}}^2$ corresponding to the loop quark mass. This combination is relevant for extracting the sigma term, $$\sigma_N = m_q\langle N|\bar{u}u+\bar{d}d|N\rangle\,.$$ We found $2m_q\langle N|\bar{u}u|N\rangle^{\mathrm{dis}}$ for $\kappa_{\mathrm{val}}=\kappa_{\mathrm{loop}}=\kappa_{\mathrm{sea}}=0.13632$ to amount to roughly $40\%$ of the connected contribution to $\sigma_N$ for this volume. However, a more sophisticated method of extracting the matrix element from $R^{\mathrm{dis}}$ is required in order to make a firm comparison. The spin contribution: $\Delta s$ and $\Delta q$ ================================================ The results for $\Delta s$ on all three volumes are shown in fig. \[deltaq\]. No significant dependence on the valence quark mass nor on the lattice size is seen in the data. Neither is there any significant variation in the results if the loop quark mass is reduced. These numbers will have to be multiplied by a renormalization constant of approximately $0.8$ for the $\overline{MS}$ scheme [@Skouroupathis:2008mf]. In contrast to the scalar case, the disconnected contributions are much smaller than the connected terms, at around $10\%$ for $\Delta d$ and $5\%$ for $\Delta u$. Outlook ======= In the short term, the aim is to reach our target statistics of $2000$ trajectories for each volume. We then plan to begin an analysis close to the physical sea quark mass. The nonperturbative renormalization for $\Delta q$ needs to be calculated while for the scalar strangeness matrix element mixing with the light flavours needs to be considered. This work was supported by the EU ITN STRONGnet, the I3 HadronPhysics2 and the DFG SFB/Transregio 55. Sara Collins acknowledges support from the Claussen-Simon-Foundation (Stifterband für die Deutsche Wissenschaft). Computations were performed on the IBM BlueGene/L at EPCC (Edinburgh,UK), Regensburg’s Athene HPC cluster, the BlueGene/P (JuGene) and the Nehalem Cluster (JuRoPA) of the Jülich Supercomputer Center and the SFB/TR55 QPACE supercomputers. The Chroma software suite [@Edwards:2004sx] was used extensively in this work. [99]{} H. Ohki [*et al.*]{}, *Nucleon sigma term and strange quark content in 2+1-flavor QCD with dynamical overlap fermions*, *Pos* [**LAT2009**]{} 124 \[[arXiv:0910.3271 \[hep-lat\]](http://arxiv.org/abs/0910.3271)\]. R. Babich, R. Brower, M. Clark, G. Fleming, J. Osborn and C. Rebbi, *Strange quark content of the nucleon*, *PoS* [**LATTICE2008**]{} 160 \[[arXiv:0901.4569 \[hep-lat\]](http://arxiv.org/abs/0901.4569)\]. W. Freeman and D. Toussaint, *The strange quark content of the nucleon in 2+1 flavor lattice QCD*, *PoS* [**LAT2009**]{} 137 \[[arXiv:0912.1144 \[hep-lat\]](http://arxiv.org/abs/0912.1144)\]. A. Ramos et al. \[BMW Collaboration\], *Sigma term and strange content of the nucleon*, these proceedings. C. Jung and T. Izubuchi \[RBC/UKQCD Collaboration\], *$\langle N|\bar{s}s|N\rangle$ via reweighting on (2+1)-flavor DWF lattices*, these proceedings. A. Airapetian et al.  \[HERMES Collaboration\], *Precise determination of the spin structure function $g_1$ of the proton, deuteron and neutron*, *Phys. Rev. D* [**75**]{} (2007) 012007 \[[arXiv:hep-ex/0609039](http://arxiv.org/abs/hep-ex/0609039)\]. G. Bali, S. Collins and A. Schäfer \[QCDSF Collaboration\], *Strangeness and charm content of the nucleon*, *PoS* [**LAT2009**]{} 149 \[[arXiv:0911.2407 \[hep-lat\]](http://arxiv.org/abs/0911.2407)\]. A. Nobile, *Solving the Dirac equation on QPACE*, these proceedings. Y. Nakamura and H. Stüben \[QCDSF Collaboration\], *BQCD – Berlin quantum chromodynamics program*, these proceedings. G. Schierholz et al. \[QCDSF Collaboration\], *Low-energy parameters of pion and nucleon from two-flavor lattice QCD at physical quark masses*, these proceedings. M. Göckeler [*et al.*]{} \[QCDSF/UKQCD Collaboration\], *Lattice investigations of nucleon structure at light quark masses*, *PoS* [**LAT2009**]{} 125 \[[arXiv:0912.0167 \[hep-lat\]](http://arxiv.org/abs/0912.0167)\]. S. Bernardson, P. McCarty and C. Thron, *Monte Carlo methods for estimating linear combinations of inverse matrix entries in lattice QCD*, *Comput. Phys. Commun.*  [**78**]{} (1993) 256. W. Wilcox, *Noise methods for flavor singlet quantities*, [arXiv:hep-lat/9911013](http://arxiv.org/abs/hep-lat/9911013). C. Thron, S. J. Dong, K. F. Liu and H. P. Ying, *Pade-$Z_2$ estimator of determinants*, *Phys. Rev.  D* [**57**]{} (1998) 1642 \[[arXiv:hep-lat/9707001](http://arxiv.org/abs/hep-lat/9707001)\]. G. S. Bali, S. Collins and A. Schäfer, *Effective noise reduction techniques for disconnected loops in Lattice QCD*, *Comput. Phys. Commun.* [**181**]{} (2010) 1570 \[[arXiv:0910.3970 \[hep-lat\]](http://arxiv.org/abs/0910.3970)\]. A. Skouroupathis and H. Panagopoulos, *Two-loop renormalization of vector, axial-vector and tensor fermion bilinears on the lattice*, *Phys. Rev.  D* [**79**]{} (2009) 094508 \[[arXiv:0811.4264 \[hep-lat\]](http://arxiv.org/abs/0811.4264)\]. R. G. Edwards and B. Joo \[SciDAC, LHP and UKQCD Collaborations\], *The Chroma software system for lattice QCD, Nucl. Phys. Proc. Suppl.* [**140**]{} (2005) 832 \[[[ hep-lat/0409003]{}](http://arxiv.org/abs/hep-lat/0409003)\].

ArXiv

--- abstract: 'Within the resistive magnetohydrodynamic model, high-Lundquist number reconnection layers are unstable to the plasmoid instability, leading to a turbulent evolution where the reconnection rate can be independent of the underlying resistivity. However, the physical relevance of these results remains questionable for many applications. First, the reconnection electric field is often well above the runaway limit, implying that collisional resistivity is invalid. Furthermore, both theory and simulations suggest that plasmoid formation may rapidly induce a transition to kinetic scales, due to the formation of thin current sheets. Here, this problem is studied for the first time using a first-principles kinetic simulation with a Fokker-Planck collision operator in 3D. The low-$\beta$ reconnecting current layer thins rapidly due to Joule heating before onset of the oblique plasmoid instability. Linear growth rates for standard ($k_y = 0$) tearing modes agree with semi-collisional boundary layer theory, but the angular spectrum of oblique ($|k_y|>0$) modes is significantly narrower than predicted. In the non-linear regime, flux-ropes formed by the instability undergo complex interactions as they are advected and rotated by the reconnection outflow jets, leading to a turbulent state with stochastic magnetic field. In a manner similar to previous 2D results, super-Dreicer fields induce a transition to kinetic reconnection in thin current layers that form between flux-ropes. These results may be testable within new laboratory experiments.' author: - 'A. Stanier' - 'W. Daughton' - 'A. Le' - 'X. Li' - 'R. Bird' title: Influence of 3D plasmoid dynamics on the transition from collisional to kinetic reconnection --- mvpo \[sec:intro\]Introduction ========================= Magnetic reconnection is the change in topology of magnetic field-lines in a highly-conducting plasma. The reconnection associated release of stored magnetic energy into plasma kinetic energy is thought to be important in solar flares [@priest02; @su13], planetary magnetospheres [@dungey61; @burch16], and other astrophysical phenomena. In the laboratory, reconnection is usually associated with sawteeth that can lead to the fast collapse of core pressure profiles [@vongoeler74; @kadomtsev75; @chapman11], but it can also be utilized during tokamak start-up to obtain desired magnetohydrodynamic equilibrium states [@ebrahimi15; @stanier13]. ![\[fig:phasediagram\]Reconnection phase-diagram for Lundquist number $S$ and system-size $\lambda$ (see text for definitions). Different regimes of reconnection are delineated by the labelled (approximate) thresholds with $S_c=10^4$ and $\lambda_c=50$. The conditions for the solar corona and the FLARE experiment are marked. The tail of the arrow shows the initial conditions for the 3D kinetic simulation in this paper, and the head of the arrow gives the conditions just prior current-sheet break-up ($t\Omega_{ci}=88$).](phasediagram-eps-converted-to.pdf){width="50.00000%"} In these different plasma environments, the regimes of reconnection can vary, depending on the plasma size, collisionality, and the magnetic field configuration. Recent efforts [@jidaughton11; @baalrud11; @daughtonroytershteyn12; @cassak13; @huangbhattacharjee13; @le15; @loureirouzdensky16; @pucci17] have sought to classify the different regimes of reconnecting current sheets using a phase diagram in $S-\lambda$ space, for Lundquist number $S \equiv \mu_0\, v_{A}\, L_{CS}/\eta$ and normalized system-size $\lambda \equiv L/\delta_i$. Here, $v_{A}$ is the Alfvén velocity defined with the upstream (reconnecting) magnetic field, $L_{CS} = L/2$ is the current sheet half-length for a system of size $L$, $\eta$ is the Spitzer resistivity, and $\delta_i$ is the relevant ion kinetic scale. In a low-$\beta$ plasma, with $\beta$ the ratio of thermal to magnetic pressures, $\delta_i \equiv \rho_s = \sqrt{(T_i+T_e)m_i}/(q_iB)$ is the ion-sound radius defined with the ion/electron temperatures $T_{i/e}$, the magnetic field strength $B=|\boldsymbol{B}|$, and the ion charge $q_i$ and mass $m_i$. Figure \[fig:phasediagram\] shows an example phase-diagram that is similar to the one proposed in Ref. \[\]. Here, the value of $S_c \sim 10^4$ is assumed to be the critical threshold at which a collisional Sweet-Parker current layer breaks up due to the plasmoid instability in MHD, although this can depend in practice on the background fluctuation level in the system [@comisso16; @huang17]. A long standing problem in reconnection theory has been a viable explanation for the fast ($S$-independent) reconnection rates in solar flares. The initially promising Petschek model [@petschek64] invoked a microscopic value for the current sheet length, $L_{CS} \ll L$, with the primary energy conversion occuring at pairs of slow-mode shocks that bound the reconnection exhaust. However, an *ad-hoc* localized resistivity enhancement is necessary to access the solution within resistive-MHD [@ugai77; @forbes13], and it has not yet been validated with either first-principles numerical simulations or laboratory experiments (unlike the Sweet-Parker solution [@ji98; @daughton09b] with $L_{CS}\sim L$). An alternative idea invokes kinetic scales, following the now well established result from simulations [@birn01] and experiments [@yamada06; @egedal07] that reconnection becomes fast when the Sweet-Parker current sheet thickness $\delta_{SP}=S^{-1/2}L_{CS}$ falls below the ion kinetic scale $\delta_i$. This transition was historically considered using laminar Sweet-Parker layers, e.g. Ref. \[\], for which the threshold is the black line in Fig. \[fig:phasediagram\]. However, the Lundquist number in the corona [@jidaughton11] $S\sim 10^{13}$ is vastly above $S_c$, and it is now widely recognised that current sheets will become unstable to the plasmoid instability before the laminar Sweet-Parker layers have time to form [@puccivelli14; @uzdenskyloureiro16; @comisso16; @huang17; @pucci17]. Studies [@bhattacharjee09; @cassak09; @huang10; @loureiro12] have found that plasmoid-dominated reconnection can be fast in the “Multi X-line collisional” regime of Fig. \[fig:phasediagram\], which can be modelled with resistive MHD simulations without invoking kinetic scales. However, the applicability of these results to solar flares remains uncertain for several reasons. Firstly, the onset of the plasmoid instability may lead to kinetic scale reconnection more readily than by the thinning of a laminar Sweet-Parker layer. This idea was first suggested by Ref. \[\] who proposed that the plasmoid formation will lead to the formation of new secondary current sheets, which are also unstable to plasmoid formation. Applied recursively, this suggests a hierarchy of sheets and islands, which can form a cascade down to the ion kinetic scales where collisionless reconnection is triggered. This basic scenario has been confirmed in 2D using both Hall-MHD [@shepherd10; @huang11], as well as fully kinetic simulations [@daughton09a; @daughton09b], which give the theoretical basis for the blue line in Fig. \[fig:phasediagram\]. Within 3D reconnection layers, plasmoids are potentially unstable over a broader range of angles, and lead to the formation of flux ropes with considerably more freedom to interact. Large-scale 3D MHD simulations [@oishi15; @huang16; @beresnyak17; @kowal17] indicate that the reconnection layer becomes turbulent. While new thin current sheets are still produced, it is less clear how to estimate if this 3D dynamics leads to kinetic scale reconnection. Secondly, it is expected that the electric fields associated with solar flare reconnection should significantly exceed [@cassakshay10; @jidaughton11] the critical Dreicer [@dreicer59] threshold, $E_{\textrm{flare}} \gg E_D = (m_e T_e)^{1/2}\nu_{ei}/e$, at which fluid models break down [@daughton09a; @roytershteyn10]. These super-Dreicer electric fields may play a role in the generation of non-thermal distributions of particles that are often observed during solar flares [@lin06; @krucker10]. The Facility for Laboratory Reconnection Experiments (FLARE [@jiflare18]) has been designed to tackle these questions, amongst others. The maximum $S=10^4-10^5$ and $\lambda=10^2-10^3$ accessible are small compared with solar flare values, but should be large enough to study the phase transitions between the different reconnection regimes shown in Fig. \[fig:phasediagram\]. These more modest values are also becoming accessible for direct numerical simulation using first principles kinetic modelling, including the effects of Coulomb collisions [@daughton09a; @daughton09b; @roytershteyn10]. In particular, Refs. \[\] have studied these phase-transitions with $2D$ simulations using the Harris sheet equilibrium in the $\beta \approx 1$ regime. At these lower values of $S$, the Sweet-Parker layer is able to form initially (in contrast to coronal values) but thins due to Joule heating along with a temperature dependent resistivity. For small systems, reconnection transitions to the kinetic regime in laminar layers as $\delta_{SP}$ thins below $\delta_i$, but for larger layers this transition is triggered earlier by the onset of the plasmoid instability (as indicated by the blue line in Fig. \[fig:phasediagram\]). In the present paper, this transition is considered in 3D for an initially force-free current sheet in the low-$\beta$ regime, using a first-principles kinetic simulation with a Fokker-Planck collision operator. The low-$\beta$ regime is relevant for solar flares and magnetic reconnection experiments in FLARE. Compared with the $\beta \approx 1$ results of Ref. \[\], the low-$\beta$ current layer is found to thin much more rapidly from its initial thickness due to Joule heating and reach a significant Lundquist number $S\sim 10^4$ prior to plasmoid onset. At onset, the plasmoid instability in 3D results in multiple oblique modes that form at different rational surfaces [@daughton11; @baalrud12], and can be stretched [@huang17] and rotated by the reconnection outflow jets. It is found that the growth rates for the standard ($k_y \approx 0$) modes of the instability agree well with the semi-collisional predictions [@drakelee77] of boundary layer theory for the tearing instability, but the angular cut-off for the unstable oblique modes is significantly smaller than predicted. Although the initial conditions are force-free, temperature gradients develop self-consistently due to Joule heating in the initial phase and the possibility of diamagnetic stabilization due to these gradients is considered. The temperature gradient stabilization predicted for the semi-collisional drift-tearing mode [@connor12] is too small to explain this effect alone, but there may be additional stabilization due to the break-down of scale separation between the inner tearing layer thickness and the outer current sheet [@baalrud18]. In the non-linear regime, the oblique tearing modes grow to form flux-ropes that undergo a variety of kink and coalescence processes, while they continue to be rotated by the reconnection outflows. These interactions lead to a turbulent-like state with large regions of stochastic magnetic field. Despite these complications, this simulation suggests that the transition from collisional to kinetic reconnection can occur in a manner analogous to the 2D picture. Thin current layers form between the flux-ropes, where super-Dreicer electric fields are supported by collisionless terms in Ohm’s law. These thin current layers can become unstable to the generation of additional flux-ropes. The paper is organized as follows. In Section \[sec:setup\], the initial and boundary conditions, and the numerical parameters for the simulation are described. In Section \[sec:results\], an overview of the different stages of the simulation is given. These stages are then considered in more detail in the following sections. Section \[sec:thinning\] describes the thinning of the collisional Sweet-Parker current layer prior to onset. Section \[sec:obliqueplasmoid\] presents an account of the oblique plasmoid instability and compares with current theories, with focus on the role of outflow jets in the stretching and rotation of the oblique modes, the collisionality of the inner tearing layer, the angular spectrum of the oblique modes, the non-linear flux-rope processes, and the generation of stochastic magnetic field. Section \[sec:kinetictransition\] describes evidence for the transition to kinetic reconnection in thin current layers that form due to the plasmoid instability. Finally, a summary of results is given in Section \[sec:conclude\]. \[sec:setup\]Simulation set-up ============================== The primary simulation in this paper was performed with the VPIC particle in cell code [@bowers08]. Unless otherwise specified, velocities are normalized by the light speed $c$, frequencies by the electron plasma frequency $\omega_{pe0}=\sqrt{n_0 e^2/\epsilon_0 m_e}$ and distances by the electron skin-depth $d_{e0} = c/\omega_{pe0}$. The simulation described in this paper is initialised with a force-free current sheet in a uniform plasma of physical number density $n_{i0}=n_{e0}=n_0$, and temperature $T_{i0}=T_{e0}=T_{0}$, with electrons (ions) of mass $m_e$ ($m_i$) and charge $-e$ ($e$). The initial magnetic field profile is given by $$\label{initialfield}\boldsymbol{B} = B_{r0} \tanh{\left(z/\delta_0\right)} \boldsymbol{\hat{x}} + B_{r0} \sqrt{b_g^2 + 1-\tanh^2{\left(z/\delta_0\right)}}\boldsymbol{\hat{y}},$$ where $B_{r0} = 1/(\omega_{pe0}/\Omega_{ce0})$ is the asymptotic reconnecting magnetic field, $b_g = 0.6$ is the ratio of the guide field to the reconnecting field, and $\delta_0 = 2\,d_{i0}$ is the initial current sheet half-thickness in units of the ion inertial length $d_{i0}/d_{e0} = \sqrt{m_i/m_e}$. The initial ratio of the electron thermal to magnetic pressure based upon the reconnecting field is $\beta_{e0} = 2 \mu_0 n_0 T_0/B_{r0}^2 = 0.08$, and the ratio of the electron plasma frequency to the gyro-frequency $\omega_{pe0}/\Omega_{ce0} = 1$ (similar to a solar coronal value). In order to start in a collisional (Sweet-Parker) parameter regime, a large seperation of scales is needed between the current sheet length and the ion kinetic scales ($\lambda > \sqrt{4S}$ according to Fig. \[fig:phasediagram\]). A reduced ion-to-electron mass ratio of $m_i/m_e = 40$ is used to make such simulations tractable. The domain for the 3D simulation is a box of size $(L_x, L_y, L_z) =(164, 109.3, 54.7) \, d_{i0}$ that is periodic in $x$ and $y$, and has perfect conducting and particle reflecting boundaries in the $z$ direction. The spatial grid is $(n_x,n_y,n_z) = (3072, 2048, 1024)$ with $140$ particles per cell for each species (total $1.8 \times 10^{12}$ particles). The timestep is $\Delta t \omega_{pe} = 0.12$ (light wave CFL$=0.6$). Both ion and electron Coulomb collisions are studied. These are modelled using a Monte Carlo treatment of the Fokker-Planck collision operator [@takizuka77; @daughton09a]. The initial ratio of the electron-ion collision frequency to the cyclotron frequency is chosen to be $\nu_{ei0}/\Omega_{ce0} = 0.04$ such that the plasma is well magnetised. Since collisions are infrequent ($\omega_{pe0} = \Omega_{ce0} \gg \nu_{ei0}$), the collision operator is applied every $\Delta t_{\textrm{coll}} = 22\Delta t$ to reduce computational cost. This value was chosen based on numerical convergence to classical resistive friction [@braginskii65] within the current sheet at early time - when the plasma is cold and the requirement to resolve the collision frequency is the most restrictive. The initial conditions described above can be understood in the context of the reconnection phase diagram (Fig. \[fig:phasediagram\]). For a low-$\beta$ force-free current sheet, the key parameters are the system-size $\lambda = L/\rho_{s}$, and the Lundquist number based on the parallel resistivity $S_\parallel=L_{CS}v_A\mu_0/\eta_\parallel$. The latter can be written as $S_\parallel = (L_{CS}/d_{i0})/\hat{\eta}_\parallel$ for normalized resistivity $\hat{\eta}_{\parallel} = 0.51 (\nu_{ei0}/\Omega_{ce0}) (T_0/T_e)^{3/2}$. We follow the conventions of Ref. \[\] to define $L=L_x$ and $L_{CS} = L_x/4$. At $t=0$, the initial conditions above give $\lambda_0 = 676$ and $S_{\parallel 0} = 2010$. This position is marked in Fig. \[fig:phasediagram\](bottom) as the tail position of the blue arrow, which is within the ‘single X-line collisional’ regime and the operating regime of the FLARE magnetic reconnection experiment. An additional requirement for collisional reconnection is for the electric field to be less than the Dreicer field $E_y/E_D < 1$. At early time (see below) the electric field is given by the resistive friction, $E_y = \eta_\parallel j_\parallel$, where $j_\parallel$ is the current at the X-point. It can be shown that $$E_y/E_D = \frac{0.51}{\sqrt{\beta_e/2}(\delta/d_{i0})\sqrt{m_i/m_e}},$$ where $\beta_e$ is defined using the electron temperature and the upstream field. At $t=0$, $\beta_e=\beta_{e0}$ and $\delta=\delta_0$ to give $E_y/E_D=0.2$. A 2D perturbation is applied to the magnetic field to start the reconnection with $\boldsymbol{\delta B} = \boldsymbol{\nabla} \times \left(\delta A_y\boldsymbol{\hat{y}}\right)$, where $$\label{initialpert}\delta A_y = -\frac{0.0125 B_{r0} L_x}{\pi} \cos{\frac{2\pi (x-L_x/2)}{L_x}}\sin{\frac{\pi (z-L_z/2)}{L_z}}.$$ \[sec:results\]Simulation results ================================= ![\[fig:snapshots\]Volume rendering of the electron temperature $T_e$ with sample magnetic field lines (white). Movie (Multimedia view).](4panel-fig-labelled-halfsize-raster6-56dpi-axis.pdf){width="48.00000%"} Figure \[fig:snapshots\] shows several snapshots of the electron temperature $T_e$ over the course of the simulation. Since electron heat transport is primarily along the magnetic field, $T_e$ serves as a useful proxy to visualise the magnetic topology. In the first snapshot, the $T_e$ profile is due to Joule heating within a quasi-2D current sheet structure that is set-up from the initial magnetic field perturbation. As will be discussed below, the electron heating leads to current layer thinning until the layer becomes unstable to the primary plasmoid instability. The second panel shows this instability in the early non-linear phase. The formation of oblique flux-ropes breaks the initial symmetry, and they exhibit a range of kinking and coalescence processes. In the third panel, the magnetic flux-ropes produced by this instability are advected downstream, and further thin current layers form. These can also become unstable to secondary tearing-type instabilities to produce further flux-ropes as demonstrated in the fourth panel. The different stages of the simulation are discussed in further detail in the following sections. \[sec:thinning\]Single X-line collisional reconnection ====================================================== ![\[fig:ohmslaws\]Contributions to parallel force balance at $t\Omega_{ci}=60$, in the initial phase of the simulation prior to plasmoid instability onset. Contributions have been averaged in time and along the field-lines - see text for definitions. Quantities are expressed in ion units, after normalization by $n_0B_{r0}v_{A}$.](FLint-ohms-t60-L20de-w-labels-ionnorm-eps-converted-to.pdf){width="50.00000%"} \[sec:SPlayer\]Collisional current layer ---------------------------------------- In order to verify that the current layer is in the collisional regime in the initial phase of the simulation, and to quantify the applicability of classical transport theory [@braginskii65], the parallel component of the electron momentum balance across the current sheet is considered. The parallel electron momentum equation (Ohm’s law) is given by $$\label{ohmslaw}n_e e E_\parallel + \left[n_e m_e \frac{D \boldsymbol{u}_e}{D t}\right]_\parallel + \left[\boldsymbol{\nabla} \cdot \tensor{\boldsymbol{P}}_e\right]_\parallel = R_{e\parallel},$$ where $E_\parallel = \boldsymbol{E}\cdot \boldsymbol{\hat{b}}$, $\boldsymbol{\hat{b}}=\boldsymbol{B}/B$, $\boldsymbol{E}$ is the electric field, $\boldsymbol{u}_e$ is the electron bulk velocity, $D_t$ is the total derivative, $\tensor{\boldsymbol{P}}_e$ is the electron pressure tensor, and $\boldsymbol{R}_e$ is the collisional momentum exchange, which is identically zero in a collisionless plasma. In the strongly magnetized and collisional regime, $R_{e\parallel}$ can be computed from classical transport theory [@braginskii65] as $$\label{neEfric}R_{e\parallel} \approx n_e e \eta_\parallel j_\parallel - 0.71 n_e \nabla_\parallel T_e,$$ where the first term is due to the resistive friction, and the second term is due to the parallel thermal force. To test the closure, all of the terms in Eqs. (\[ohmslaw\],\[neEfric\]) were first averaged over a collision timescale $\nu_{ei}^{-1}$. Then, to further reduce statistical noise, the same terms were spatially averaged by integration along magnetic field-lines from an initial line of seed-points $\boldsymbol{x}=\boldsymbol{x}_0$ as $$<R_{e\parallel}> = \frac{1}{L_s}\int_{\boldsymbol{x}_0}^{\boldsymbol{x}_f} \boldsymbol{\hat{b}}[\boldsymbol{x}(s)]\cdot \boldsymbol{R}_e[\boldsymbol{x}(s)] ds,$$ where the final position $\boldsymbol{x}=\boldsymbol{x}_f$ is a distance $L_s$ along the field-line from $\boldsymbol{x}_0$. Ref. \[\] has shown that this method of spatial averaging gives less smearing out of the diffusion regions compared to averaging along the $y$-axis when structures do not align with the $y$-direction, which is the case after onset of the oblique plasmoid instability. Figure \[fig:ohmslaws\] shows the parallel force balance at $t\Omega_{ci} = 60$, when the reconnecting current layer has formed, but prior to plasmoid instability growth. Here, $\boldsymbol{x}_0 = (L_x/2,0,z)$ for $z\in [z_s-40,z_s+40]$, $z_s=L_z/2$, and $L_s=40$ ($d_e$). The term due to collisional momentum exchange, $<R_{e\parallel}>$ (grey), is calculated as the residual of the left hand side of Eq. (\[ohmslaw\]). It balances the term due to the parallel electric field, $<n_e e E_\parallel>$ (orange), and sets the thickness of the electron diffusion region at this time. The collisional transport closure (red) and the residual (grey) agree to within $3\%$ at the peak values, suggesting that classical transport is well founded in this early phase. Within the closure term, the friction term $<n_ee\eta_\parallel j_\parallel>$ is dominant over the thermal force $<-0.71 n_e \nabla_\parallel T_e>$. However, although the electron inertia (blue) and electron pressure tensor (green) are small, they are non-negligable in the center of the current sheet where they balance $20\%$ of the parallel electric field term at the X-point. A possible reason for this is partial runaway of electrons in the tail of the distribution function, which can occur even for sub-Dreicer electric fields [@dreicer60; @connorhastie75]. \[sec:thinningmodel\]Resistive thinning of current layer -------------------------------------------------------- ![\[fig:thinning\]Time traces of the electron temperature increase $T_e/T_{e0}$, the Lundquist number $S_\parallel$, the ratio of the system-size to the sound radius $\lambda=L/\rho_s$, the ratio of the layer thickness to the sound radius $\delta/\rho_s$, and the ratio of the electric field to the Dreicer runaway field $E_y/E_D$ in the early phase. Shown for the 3D first-principles kinetic simulation (blue) and a 2D single-fluid resistive MHD [@chacon08b] simulation (red) assuming $T_i=T_e$ with a temperature dependant Spitzer resistivity. The black dashed lines show the simple scaling model discussed in the text. The blue vertical lines show the start of plasmoid instability growth (“Linear” phase), and the time the magnetic islands are comparable to the current layer thickness (“Non-linear” phase).](figs-RMHD-vs-VPIC-eps-converted-to.pdf){width="50.00000%"} To characterize the initial current sheet thinning phase, prior to the plasmoid instability onset, Figure \[fig:thinning\] shows time traces of $T_e/T_{e0}$, $S_\parallel$, $\lambda=L/\rho_s$, $\delta/\rho_s$, and $E_y/E_D$ from the simulation (blue). Since the current sheet in the 3D simulation has symmetry along the $y$-direction (first panel in Fig. \[fig:snapshots\]), this 3D data is first reduced to 2D by averaging in $y$. The values of $T_e$, $T_i$, $|B|$, $n$ and $E_y = -\partial_t A_y$ are then measured at the dominant X-point of the mean-field magnetic flux profile, and $\delta$ is the half-thickness of the current layer at the thinnest point along its length (usually at the dominant X-point). The value of $\delta$ is estimated by fitting Eq. (\[initialfield\]), such that $\delta(t=0) = \delta_0$. For simplified fluid models with constant plasma resistivity, the current layer will thin due to the initial perturbation towards a constant Sweet-Parker thickness $\delta_{SP0} = S_0^{-1/2} L_{CS}$ for $S_0 \approx \textrm{const}$. Depending on the Lundquist number and the background noise level, the sheet may be either be stable, or break up before or after it is formed. In the present simulations, the plasma transport is determined self-consistently from the kinetic description of collisions and includes temperature dependent and anisotropic resistive and thermal friction, viscosity, heat conduction, and species thermal equilibration. Thus, $S_{\parallel} \neq \textrm{const.}$, and $\delta_{SP} = S_{\parallel}^{-1/2} L_{CS}$ can evolve in time. The precise evolution of the thickness $\delta(t)$ can, in principle, depend on all of the transport effects mentioned. To illustrate the most important physics, the same parameters are computed from a 2D resistive MHD simulation (red) with corresponding initial conditions, a temperature dependent Spitzer resistivity [@spitzer53], and which neglects heat conduction and assumes exact temperature equilibration $T_i=T_e$. Here, for the single-fluid model, $\rho_s$ and $E_D$ are not physically meaningful, but are computed to normalize quantities in the same manner as the kinetic simulation. The simplified MHD model reproduces reasonably well the overall profiles $T_e/T_{e0}$, $S_\parallel$, $\lambda$ and $\delta/\rho_s$. The kinetic model has a slightly larger $T_e$ (and therefore $S_\parallel$) than the MHD model, which is attributed to the preferential Joule heating of electrons while the equilibration timescale $\tau_{eq}\Omega_{ci} = (T_e/T_{e0})^{3/2}/(\nu_{ei0}/\Omega_{ce0}) = 25 (T_e/T_{e0})^{3/2}$ does not remain small compared to the timescale of current layer thinning. Despite this, the temperature ratio remains within a factor of $\tau \equiv T_e/T_i = 1.5$ at $t\Omega_{ci}=60$, and $\tau = 2$ at $t\Omega_{ci}=90$. Other noticeable differences include a slightly larger [^1] total temperature and thus $\rho_s$ in the MHD model, and a significantly weaker $E_y/E_D$ at late times. To verify that the thinning observed requires the temperature dependent resistivity, we performed a similar resistive-MHD simulation with uniform resistivity and found that $\delta$ is approximately $3$ times thicker (not shown) at $t\Omega_{ci}=90$. This result demonstrates that the temperature dependent (Spitzer) resistivity can play an important role in the evolution towards plasmoid unstable regimes. With the results described above, it is convenient to parameterize the thinning via a simplified analytic scaling model that can be used to plot the trajectory of the thinning phase onto the reconnection phase diagram. Firstly, based on the simulation data, it is assumed that $n$, $|B|$ and $L_{CS}$ are constant, and that $T_i \approx T_e$. With these assumptions, the phase-diagram co-ordinates vary only with $T_e/T_{e0}$ as $S_\parallel \propto \eta_\parallel^{-1} \propto (T_e/T_{e0})^{3/2}$ and $\lambda \propto \rho_s^{-1} \propto (T_{e}/T_{e0})^{-1/2}$. Then the temperature evolution can be estimated by neglecting heat conduction and viscous heating (which occurs primarily downstream of the X-point), such that the temperatures increase solely due to Ohmic heating within the layer $$\frac{3}{2}n_0 \partial_t (T_e + T_i) \approx \eta_{\parallel} j_{\parallel}^2.$$ Finally, it is assumed that the current at the X-point follows a Sweet-Parker scaling $j_{\parallel} \propto \delta_{SP}^{-1} \propto \eta_{\parallel}^{-1/2}$, such that $\partial_t T_e \propto \eta_{\parallel}^0$, i.e. an electron temperature that increases linearly in time $T_e \propto t$. The fractional heating rate can be estimated based upon the initial current density [@daughton09a], as $T_{e}/T_{e0} \approx 1 + Q_e t\Omega_{ci}$ where $$\label{heatingrate}Q_e \approx \frac{4}{6\, S_{\delta 0}\, \beta_{e0} (\delta_0/d_{i0})} \approx 0.0425,$$ with $S_{\delta 0} = S_{\parallel 0} \delta_0/L_{CS}$. It follows that $S_\parallel \approx S_{\parallel 0} (1 + Q_e t\Omega_{ci})^{3/2}$, $\lambda \approx \lambda_0 (1 + Q_e t\Omega_{ci})^{-1/2}$, $\delta \approx \delta_0 (1 + Q_e t\Omega_{ci})^{-3/4}$ ($\delta/\rho_s \approx \delta_0/\rho_0 (1 + Q_e t\Omega_{ci})^{-5/4}$), and $E_y/E_D \propto (1 + Q_e t\Omega_{ci})^{1/4}$. To compare the simple model against the simulation data, a linear profile (black dashed line) is fit to $T_e/T_{e0}$ for the kinetic simulation, which gives a measured value of $Q_e = 0.034$. The dashed lines in the other panels show predicted time profiles for each quantity using this measured value of $Q_e$, which give reasonable overall agreement with the data considering the number of assumptions made. Departures from these scalings are most noticeable in $\delta/\rho_s$ at early time, as it takes some time for Sweet-Parker reconnection to develop from the initial perturbation, and in $E_y/E_D$ at late time where $E_y$ deviates from $\eta_{\parallel} j_{\parallel}$ due to finite contributions from the pressure tensor and inertial terms in the momentum balance as discused above. These terms, which become significant during the early non-linear phase of the plasmoid instability (see Section \[sec:kinetictransition\]), are not present in the MHD model. The peak values of $T_e/T_{e0}$ and $S_\parallel$ are approximately $2.5$ and $6$ times larger respectively than simulations with similar parameters[^2] reported in Refs. \[\] and \[\] for the Harris sheet with $\beta \approx 1$. This follows from Eq. (\[heatingrate\]), where the fractional heating rate increases as $Q_e \propto \beta_{e0}^{-1}$ with other quantities equal. \[sec:obliqueplasmoid\]Oblique plasmoid instability =================================================== ![\[fig:obliquespectrum\]Reconnected component of the magnetic field $B_z(x,y,z=L_z/2)$ at $t\Omega_{ci0} = 88$. The angles of tearing fluctuations with respect to the $y$-axis are indicated by labelled red lines. Inset: 2D Fourier spectrum, $P(k_x,k_y,t\Omega_{ci}=88)$. The labelled white lines mark angles $\theta = \pm 30^\circ$. The time evolution of this Fourier spectrum is available in the multimedia view (Multimedia view). ](bz-with-spectrum-new.pdf){width="50.00000%"} Figure \[fig:obliquespectrum\] shows $B_z$, the reconnected component of the magnetic field, in a top down view of the $z=0$ plane at $t\Omega_{ci0}=88$. At this time, which is indicated by the second vertical blue line in Fig. \[fig:thinning\], tearing-type fluctuations in the current density become noticeable over the background current sheet structure. These fluctuations are visible in Fig. \[fig:obliquespectrum\] close to the center of the current layer, where they form at a range of oblique angles to the $y$-axis. To more clearly show the angular distribution of the fluctuations, Fig. \[fig:obliquespectrum\] inset shows $P(k_x,k_y,t\Omega_{ci}=88) = \log_{10}{\left(\int_0^{L_z} |\hat{B_z}(k_x, k_y,z)|^2\,dz\right)}$, the power spectrum of the magnetic energy density in $k_x-k_y$ space and integrated over the height of the simulation box $L_z$. Here, the peak values at $k_y = 0$ and $k_x d_e < 0.1$ are partly associated with the background reconnecting current sheet structure, but there is significant power across a range of oblique modes with $\theta = \arctan{(k_y/k_x)} \lesssim 30^\circ$. A full analysis of the plasmoid instability requires accounting for the detailed plasma physics of the inner tearing layer [@fkr63; @coppi79; @drakelee77; @cowley86], the evolution of the background current profiles during the current sheet thinning process [@uzdenskyloureiro16; @puccivelli14; @comisso16] ($\delta = \delta(t)$), and the role of outflow jets in the advection and stretching of weakly growing modes [@huang17]. The full analysis is not given here, but the relative importance of each of these is examined in this section from the simulation data in comparison with current theories. In particular, we quantify the importance of plasma collisions in the inner tearing layer and investigate the physics responsible for the maximum cutoff angle $\theta_{\textrm{cutoff}} \approx 30^\circ$ observed. Mode stretching and rotation by outflow jets -------------------------------------------- The multimedia view of Fig. \[fig:obliquespectrum\] shows the time evolution of the power spectrum $P(k_x,k_y,t)$, with frames every $2 \Omega_{ci}^{-1}$ from $t\Omega_{ci}=0$ to $t\Omega_{ci}=140$. As well as the growth of the oblique modes, there is notable advection of these modes towards $k_x = 0$ due to mode stretching by the reconnection outflow jets. Fig. \[fig:kxthetatime\] (top panel) shows a slice of the power spectrum in the $k_x-t$ plane for $t\Omega_{ci}\in[50,140]$ at constant $k_y L_y = 2\pi$ ($k_yd_e=0.0091$), where the background 2D current sheet profile with $k_y=0$ is not visible. The oblique modes are initially visible at $t\Omega_{ci}=70$ where they are slowly advected towards $k_x = 0$. Ref. \[\] has studied this effect in detail with 2D resistive MHD simulations (without oblique modes), and generalized a model of the plasmoid instability in time evolving current sheets by Ref. \[\] to account for this physics. In the model, the modes are assumed to be advected in the $k_x$-direction as $d_t k_x = -k_x v_x^\prime$ such that they follow characteristic trajectories $$\label{stretching}k_x = k_{x0} e^{-v_x^\prime t}.$$ Here, $k_{x0}$ is the initial component of the wavenumber in the $x$-direction, and $v_x^\prime$ is the gradient of the outflow jet velocity $v_x^\prime \approx v_{x,\textrm{max}}/L_{CS}$ for maximum outflow velocity $v_{x,\textrm{max}}$ and current sheet length $L_{CS}$. Two of these trajectories are plotted as the magenta and black curves in Fig. \[fig:kxthetatime\] (top panel), where $v_{x,\textrm{max}} \approx 0.5 v_A$ and $L_{CS} \approx L_x/4 = 41 d_i$ are assumed constant in time. The curves follow the visible mode stretching reasonably well. ![\[fig:kxthetatime\] Top panel: Power spectrum of modes $P(k_x, k_yL_y=2\pi, t)$ in the $k_x-t$ plane for $t\Omega_{ci} \in [50, 140]$. The black and magenta dotted lines show two sample trajectories $k_x(t)$ from Eq. (\[stretching\]). Middle panel: Growth rates $\gamma_{\textrm{eff}}/\Omega_{ci}$ along the two characteristic trajectories calculated as in Eq. (\[growthdef\]). Bottom panel: Power spectrum of modes $P(k_x,k_y,t)$ in the $\theta-t$ plane for constant $kd_e=0.133$. The white dash-dotted lines show sample trajectories from Eq. (\[rotation\]), and the vertical dashed red and green lines show sample times at which the growth rates are compared with boundary layer theory in Fig. \[fig:growthrates\].](threepanelfig-to-t140-eps-converted-to.pdf){width="50.00000%"} Since there are no outflow jets in the $y$-direction, the modes remain with approximately constant $k_y \approx k_{y0}$ (multimedia view of Fig. \[fig:obliquespectrum\]). An interesting consequence of this in 3D is that oblique modes rotate towards larger oblique angles due to the shear of the outflow jets. Using Eq. (\[stretching\]) for $k_x$ in the definition of the oblique angle $\theta = \arctan{(k_y/k_x)}$ gives the rotation as $$\label{rotation}\theta = \arctan{[\tan{(\theta_0)} \exp{(v_x^\prime t)}]},$$ where $\theta_0 = \arctan{(k_{y0}/k_{x0})}$. Fig. \[fig:kxthetatime\] (bottom panel) shows a slice of the mode spectrum in the $\theta-t$ plane for constant $kd_e = 0.133$. There is a slow but observable advection towards larger $|\theta|$, where the white dash-dotted lines show two sample trajectories in $\theta-t$ from Eq. (\[rotation\]). Collisionality of inner tearing layer ------------------------------------- The role of collisionality in the inner tearing layer depends on the relative magnitudes of the mode frequency $|\omega_r + i \gamma|$ and the collision frequency $\nu_{ei}$. The real frequency $\omega_r$ can be non-zero in the presence of temperature or density gradients across the rational surface, and we will discuss this further below. Following Ref. \[\], the power in each fourier mode can be modeled as $$\frac{d |\hat{B_z}(k_x,k_y,t)|^2}{dt} = \left[2\gamma(t) - v_x^\prime \right] |\hat{B_z}(k_x,k_y,t)|^2,$$ where $d_t = \partial_t - k_x v_x^\prime \partial_{k_x}$ is the derivative along the characteristics, and the growth rate $\gamma(t) = \gamma(\delta(t), k_x(t))$ depends upon the instantaneous current sheet thickness $\delta(t)$ [@comisso16]. Modes only grow when the growth-rate is large enough to overcome the mode stretching [@huang17], $\gamma(t) > v_x^\prime/2$. Rearranging this for the growth rate gives $$\label{growthdef}\gamma(t) = \frac{1}{2}\frac{d\left( \ln{|\hat{B_z}|^2}\right)}{dt} + v_x^\prime/2 \equiv \gamma_{\textrm{eff}}(t) + v_x^\prime/2.$$ Fig. \[fig:kxthetatime\] (middle panel) shows $\gamma_{\textrm{eff}}(t)/\Omega_{ci}$ calculated along the two curves in $k_x-t$ from the top panel using the data $P(k_x, k_yL_y=2\pi, t)$. Here, we have filtered the signal to remove high frequency waves while well preserving the time (zero phase delay) and peak magnitude of $\gamma_{\textrm{eff}}$. At $t\Omega_{ci} = 70$ both curves have $\gamma_{\textrm{eff}}/\Omega_{ci} \approx 0.05$, which is already significantly larger than $v_x^\prime/2\Omega_{ci} \approx 0.006$. At and after this time, the mode stretching is not a substantial effect and is neglected in the rest of the discussion on the linear growth with the assumption that $\gamma(t) \approx \gamma_{\textrm{eff}}(t)$. To estimate the importance of collisions, $\gamma/\Omega_{ci}$ can be compared with the electron-ion collision frequency $\nu_{ei}/\Omega_{ci} \approx (\nu_{ei}^0/\Omega_{ce}^0) (m_i/m_e) (T_e/T_{e0})^{-3/2} = 1.6 \,(T_e/T_{e0})^{-3/2}$. At $t\Omega_{ci} = 70$, $T_e/T_{e0} = 3.3$ (Fig. \[fig:thinning\]) such that $\nu_{ei}/\Omega_{ci} \approx 0.26$ is approximately $5$ times larger than $\gamma/\Omega_{ci}$ in Fig. \[fig:kxthetatime\] at this time. At a later time $t\Omega_{ci} = 82$, $T_e/T_{e0} = 3.8$ and $\nu_{ei}/\Omega_{ci} \approx 0.21$ is comparable to the instantaneous $\gamma/\Omega_{ci}$ of the two curves. We now proceed to compare the measured growth rates with those predicted from linear boundary layer theory in more detail. Comparison with semi-collisional theory --------------------------------------- Depending on the plasma collisionality, different asymptotic regimes of the tearing instability have been derived in the literature. In the collisionless (CLS) regime, electrons within a channel of thickness $\Delta_\textrm{CLS}$ from the rational surface $z_s$ ($|z -z_s| < \Delta_\textrm{CLS}$) are freely accelerated along the field-lines by the induced electric field of the mode. For $|z-z_s| \geq \Delta_\textrm{CLS}$ the Doppler frequency becomes larger than the mode frequency, $\omega_d \equiv k_{\parallel} v_{Te} \geq |\omega|$, and the electrons experience an alternating electric field that significantly reduces the current response. The thickness of the channel $\Delta_\textrm{CLS}$ is found at $|z-z_s|=\Delta_\textrm{CLS}$ where $|\omega|=\omega_d$. Using $k_\parallel \approx k(z-z_s)/L_s$, for a magnetic shear length $L_s$ (defined below), gives $$\Delta_{\textrm{CLS}} = \frac{|\omega| L_s}{k v_{Te}}.$$ Ref. \[\] derive a growth rate for this regime, under the assumption of cold ions, as $$\label{collisionlessnormal}\gamma_{CLS} = \frac{kv_{Te} d_e^2 \Delta^\prime}{2\sqrt{\pi} L_s}.$$ Here $\Delta^\prime$ is the parameter used to match asymptotic solutions from the outer ideal region $|z-z_s| \sim \delta \gg \Delta_\textrm{CLS}$ to the inner region $|z-z_s| \sim \Delta_\textrm{CLS}$. $\Delta^\prime$ is assumed small in the derivation of Eq. (\[collisionlessnormal\]). In this Section, it is assumed that the outer region is described by a 1D force-free profile. This is not strictly true for $t>0$, as reconnected ($B_z$) field develops within the current sheet during the initial Sweet-Parker phase giving a weakly 2D profile [@loureiro13], and the profile deviates from a force-free one due to Joule heating. Despite this, we find that profiles of the form of Eq. (\[initialfield\]) fit reasonably well the magnetic field data at $x=L_x/2$ for a fitting parameter $\delta(t)$. We thus consider below the role of temperature gradients only on the inner region. Eq. (\[initialfield\]) gives [@baalrud12; @liu13; @akcay16] $$\label{deltaprime}\Delta^\prime = \frac{2}{\delta}\left(\frac{1}{k\delta}\left(1+b_g^2 \tan^2{\theta}\right) - k\delta\right),$$ $$\label{shearlength}L_s = \frac{k}{k_\parallel^\prime(z_s)} = \frac{\delta \sqrt{1+b_g^2}}{\cos{\theta} \left(1-b_g^2 \tan^2{\theta}\right)},$$ and $$\label{rationalsurface}z_s = - \delta \operatorname{arctanh}{\left(\sqrt{1+b_g^2} \sin{(\theta)}\right)}.$$ As discussed above, $\nu_{ei} \geq \gamma$ for the early phase of the instability, and thus it is necessary to include the effects of collisions. In the semi-collisional regime ($\nu_{ei} \gg |\omega|$, $\Delta_{SC} \ll \rho_s$), the thickness of the current channel $|z-z_s| = \Delta_{SC}$ is found when the mode frequency is balanced by collisional diffusion of electrons along field-lines [@drakelee77], $|\omega| = k_\parallel^2 v_{Te}^2/\nu_{ei}$. The inner layer is thus broadened by collisions as $$\label{deltasc}\Delta_{SC} = \Delta_{\textrm{CLS}} (\nu_{ei}/|\omega|)^{1/2}.$$ Ref. \[\] has also derived closed form expressions for the growth rate in this semi-collisional regime under the assumptions of cold ions, small $\Delta^\prime$, and for weak density and temperature gradients. The growth rate is modified as[^3] $$\label{semicollisionalnormal}\gamma_{SC} = \left[\frac{3 \pi^{1/4}}{4 \Gamma(11/4)}\right]^{2/3} \gamma_{CLS}^{2/3} \,\nu_{ei}^{1/3}.$$ Figure \[fig:growthrates\] (top panel) shows the measured growth-rates $\gamma(\theta)/\Omega_{ci}$ ($\gamma = \gamma_{\textrm{eff}}$) for fixed $kd_e = 0.133$ at $t\Omega_{ci} = 70$ (red dots) and $t\Omega_{ci} = 82$ (green dots), where these times are indicated by vertical lines in Fig. \[fig:kxthetatime\] (bottom panel). These are not the fastest growing modes in the simulation, but for $kd_e=0.133$ the small-$\Delta^\prime$ theory is appropriate ($\Delta^\prime \rho_s^{1/2} \Delta_{SC}^{1/2} < 1$) [@zocco11]. In addition to the time filtering mentioned above, we have taken the mean of the positive and negative $\theta$ values to better compare with theory. The solid lines show the predicted growth rate $\gamma_{SC}$ from Eq. (\[semicollisionalnormal\]), where the collision term $\nu_{ei}$ is evaluated based on the local electron temperature $T_e(z)$ at the rational surface $z=z_s(\theta)$. Despite the assumptions that have been made in Eq. (\[deltaprime\]-\[semicollisionalnormal\]), namely that the profile remains a 1D force-free layer with cold ions, there is fairly good agreement for $\theta < 20^{\circ}$ between the measured growth rates and Eq. (\[semicollisionalnormal\]). It should also be noted that $\gamma/\nu_{ei} \sim 1$ at $t\Omega_{ci} = 82$ which is not strictly in the regime of validity for the semi-collsional mode ($\nu_{ei} \gg \gamma$). Refs. \[\] have argued that the onset of the plasmoid instability can occur earlier in the semi-collisional regime ($\Delta_{SC} \ll \rho_s$) than the resistive-MHD regime ($\Delta \gg \rho_s$), due to faster tearing mode growth rates. The precise threshold for onset in the semi-collisional regime is not considered here, but we note that onset occurs at a later time ($t\Omega_{ci} \approx 140$) in the 2D resistive-MHD simulation of Fig. \[fig:thinning\] than the kinetic simulation, despite the addition of a continuous random noise forcing term to the MHD velocity fields with amplitude larger than the PIC simulation noise level. ![\[fig:growthrates\] Top panel: Dotted lines show measured growth rates $\gamma_{\textrm{eff}}(\theta)/\Omega_{ci}$ at fixed $kd_e=0.133$ for $t\Omega_{ci}=70$ (red) and $t\Omega_{ci}=82$ (green). Solid lines show the predictions for the semi-collisional tearing mode from Eq. (\[semicollisionalnormal\]). Middle panel: The asymptotic matching parameter $\Delta^\prime$ from Eq. (\[deltaprime\]) divided by the critical value for marginal stability from boundary layer theory, $\Delta^\prime_{\textrm{crit}}$ in Eq. (\[deltaprimecrit\]), for the drift tearing mode on a logarithmic scale. Bottom panel: The ratio of the inner layer thickness $\Delta$ to the outer ideal region thickness $\delta$ on a logarithmic scale. Here $\Delta$ is calculated from Eqs. (\[deltasc\]) for modes with $\theta < 20^\circ$ (solid lines) and with Eq. (\[deltaT\]) for modes with $\theta>20^\circ$ (dotted lines).](growthrates-3panel-fig2-coldiontop-integralcalc-middle-innerlayer-with-data-newlabel-eps-converted-to.pdf){width="50.00000%"} Stabilization of oblique modes ------------------------------ Although there is good agreement for the modes with $\theta < 20^{\circ}$, there is clear disagreement between Eq. (\[semicollisionalnormal\]) and the measured growth rates for $\theta \gtrsim 20^{\circ}$. For $b_g = 0.6$, Eq. (\[semicollisionalnormal\]) predicts stabilization for $\theta = 59^{\circ}$ when $L_s \rightarrow \infty$ and $z_s\rightarrow \pm \infty$. The angle $\theta = 59^\circ$ is simply half the shear angle the magnetic field makes as it rotates across the current sheet, and is thus set by the background magnetic profile of the outer region. In contrast, the measured growth-rates are stabilized for $\theta \gtrsim 35^\circ$, at which the rational surface $z_s = \pm 0.81\delta$ is still within the current layer. This suggests there is some additional stabilization mechanism associated with the inner region. A possible explanation for the discrepancy, which will be presently considered, is the diamagnetic stabilization of oblique modes due to temperature and/or density gradients [@drake83; @cowley86; @baalrud18]. Such diamagnetic flows do not exist in the force-free initial conditions, but become finite over time. Here we consider only electron temperature gradients, which arise mainly due to the Joule heating, as we find density gradients and ion temperature gradients to be significantly smaller. The diamagnetic frequency due to gradients in $T_e$ is given by $\omega_T = k T_e/(eBL_T)$, where $L_T = 1/|\partial_z \ln{(T_e)}|$. For the standard tearing modes ($\theta = 0$), $\partial_z T_e \approx 0$ due to the symmetry of the current layer, but $\omega_T \neq 0$ for oblique modes. The marginal stability threshold for standard tearing modes ($\Delta^\prime \geq 0$) is then increased to $\Delta^\prime \geq \Delta^\prime_{\textrm{crit}}$ for both collisionless [@coppi79; @antonsen81; @cowley86] and semi-collisional [@drake83; @cowley86; @connor12] drift-tearing modes. For the semi-collisional case, Ref. \[\] found $\Delta^\prime_{\textrm{crit}} \sim \hat{\beta}_T/\rho_{se}$ for the cold ion limit where $\beta_T = (\beta_e/2) L_s^2/L_T^2$. Ref. \[\] then generalized this to include the effects of finite ion orbits in the regime with $\Delta_T/\rho_i \ll 1$, where $\rho_i$ is the ion Larmor radius, and $$\label{deltaT}\Delta_T = (\nu_{ei} \omega_T)^{1/2} L_s/(kv_{Te})$$ is the semi-collisional inner layer thickness (\[deltasc\]) with $\omega = \omega_T$. The critical value $\Delta^\prime_{\textrm{crit}} \sim (\hat{\beta}_T/\rho_i)\ln{(\rho_i/\Delta_T)}$. The full definition of $\Delta^\prime_\textrm{crit}$ that is used here to test for diamagnetic stabilization is given in Appendix \[apdx:drifttearing\], which is derived following Ref. \[\] for electron temperature gradients only [@zocco15]. Figure \[fig:growthrates\] (middle panel) shows the ratio of $\Delta^\prime$, from Eq. (\[deltaprime\]), to $\Delta^\prime_{\textrm{crit}}$, from Eq. (\[deltaprimecrit\]), on a logarithmic scale. This ratio is not plotted for $\theta < 20^\circ$, for which $\gamma > \omega_T$ and the strong drift assumption breaks down. The ratio of $\Delta^\prime/\Delta^\prime_\textrm{crit}$ decreases with increasing $\theta$. However, the precise threshold for stabilization ($\Delta^\prime=\Delta^\prime_\textrm{crit}$) only occurs for $\theta=55^{\circ}$ at $t\Omega_{ci} = 70$ and $\theta=57^{\circ}$ at $t\Omega_{ci} = 82$. At $\theta = 35^\circ$, where stabilization is observed, this predicted threshold from boundary layer theory is $13\times$ smaller for $t\Omega_{ci} = 70$ and $30\times$ smaller for $t\Omega_{ci} = 82$. A similar disagreement between the predictions of boundary layer theory and the measured growth rates of oblique modes has been found previously for the Harris current sheet [@daughton11]. In such an equilibrium, diamagnetic drifts occur only due to density gradients as the temperatures are uniform. Ref. \[\] studied this discrepancy in detail in the collisionless case, concluding that the stabilization is indeed due to electron diamagnetic drift. However, the stabilization was found to be enhanced with respect to the boundary layer theory predictions when the inner tearing layer thickness, $\Delta_\textrm{CLS}$, and the outer ideal region thickness, $\delta$, have insufficient scale separation such that the assumptions of boundary layer theory break down. Fig. \[fig:growthrates\] (bottom panel) shows the ratio of the inner ($\Delta$) to outer region thickness ($\delta$) for the two times on a logarithmic scale. For modes with $\theta < 20^\circ$, we use $\Delta_{SC}$ from Eq. (\[deltasc\]) for the inner layer thickness, with $\omega = i\gamma$ from the measured growth rates. For the oblique modes with $\theta \gtrsim 20^\circ$, we take it to be $\Delta_T$ ($\omega = \omega_T$) from Eq. (\[deltaT\]). The scale separation between the inner and outer regions is reduced for large oblique angles in a similar manner as seen for the Harris sheet in Fig. 8 of Ref. \[\]. At $\theta=35^{\circ}$, where stability is observed, $\Delta/\delta \approx 0.1$. Although this may seem sufficiently small, similar values in Fig. 8 of Ref. \[\] were large enough to significantly reduce the cut-off angle for oblique modes in the Harris sheet. The precise reason for the smaller cut-off angle observed here remains an open question. It is conceivable that the combination of electron temperature gradients and breakdown of boundary layer theory could account for this, but further study is required to confirm or reject this explanation. It is significant that two studies [@liu13; @akcay16] of collisionless oblique tearing modes in a 1D force-free equilibrium (without temperature gradients) do not find any additional stabilization of oblique modes, as the cut-off angle agrees with the predictions of Eq. (\[collisionlessnormal\]). Interestingly, Refs. \[,\] report the growth-rates of oblique modes to be larger than those predicted by boundary layer theory (and even the $\theta=0$ modes) for a range of strong guide fields. Non-linear phase ---------------- For the linear regime of the plasmoid instability, it is shown in Fig. \[fig:growthrates\] (top panel) that the fastest growing modes have small oblique angles ($\theta < 20^\circ$), and that that highly oblique modes with $\theta>35^\circ$ are stabilized. This reduction in the angular distribution of fluctuations may lead one to consider that 2D simulations, which include only the $\theta=0$ modes, may capture the main aspects of this 3D simulation. However, as described in this section, the angular range of fluctuations increases in the non-linear regime. Figure \[fig:kxthetatime\] (bottom panel) shows the angular distribution of fluctuations (at $kd_e=0.133$) also for the non-linear regime of the plasmoid instability, for $88 \lesssim t\Omega_{ci} \leq 140$, which is approximately between the first and second snapshots shown in Fig. \[fig:snapshots\]. Over this interval, the trajectories of the white dashed curves in $\theta-t$ from Eq. (\[rotation\]) continue to follow the peak values of the fluctuation spectrum, indicating that the mode rotation continues into the non-linear regime while the flux-ropes are not large enough to disrupt the mean properties of the outflow jets. However, beginning at $t\Omega_{ci} \approx 110$, significant power in $P(k_x,k_y,t)$ appears at larger oblique angles than can be expected from mode rotation alone. The flux-ropes shown in the second panel of Fig. \[fig:snapshots\] show signatures of secondary instabilities. Firstly, on the left side of the domain, at $x\approx L_x/3$, there is evidence of partial coalescence between neighbouring flux-ropes: two flux-ropes visible at the $y=0$ boundary merge into a single flux-rope at $y\approx L_y/2$. Secondly, the flux-ropes show signatures of the kink instability. This is most evident for the flux-rope that forms very close to the flow stagnation point at $x=L_x/2$ and is not monotonically advected downstream by the outflow jets (Fig. \[fig:snapshots\] Multimedia view). The safety factor $q(r)=2\pi rB_y/(L_yB_\theta)$ was checked for this flux-rope at $t\Omega_{ci} = 110$ (not shown), soon after it formed, where $r$ is the radial distance from the flux-rope center and $B_\theta(r)$ the poloidal field. For an isolated flux-rope with periodic boundaries, the condition for instability [@kruskal58] requires $q(a)<1$ where $r=a$ is the edge of the flux-rope. It is found that $q(a)=0.6$ at the flux-rope edge (which is taken to be the position of the maximum value of $B_\theta$). Moreover, the kinking of the flux-rope that is visible at $t\Omega_{ci} = 132$ interacts with the reconnection outflow jets and leads to further rotation of the flux-rope as can be seen at times $t\Omega_{ci}=240,320$. This flux-rope grows via reconnection at current sheets on either side to become a “monster” flux-rope [@loureirouzdensky16] with diameter as large as $\sim L_z/3$ ($18d_{i0}$) by the end of the simulation at $t\Omega_{ci}=400$. It should be noted that although secondary flux-ropes are observed at late times (e.g. $t\Omega_{ci}=320$), there are relatively few compared with those forming along thin separatrix current layers in the 3D collisionless simulations of Ref. \[\]. The separatrix current layers in the present simulation appear less intense than those in Ref. \[\], presumably due to collisional broadening. \[sec:heattransport\]Stochastic magnetic field and heat transport ----------------------------------------------------------------- {width="75.00000%"} In 3D, the formation of oblique plasmoids at multiple resonant surfaces can lead to the breakdown of magnetic surfaces. Since plasma transport is primarily along the magnetic field, the mixing of magnetic field-lines can lead to enhanced plasma mixing. In light of the above discussion on the stabilization of strongly oblique modes in the linear regime, and on the secondary instabilities in the non-linear regime, it is useful to briefly characterize the extent of any stochastic magnetic field regions and their role in plasma transport. Figure \[fig:poincare\] (top three panels) shows Poincaré surfaces of section with magnetic field-lines for different times during the simulation. Here, the field-lines are integrated a distance of $1000 L_x$ through the simulation domain, crossing through the periodic boundaries in the $x$ and $y$-directions, and the surface of section is the plane at $y=L_y/2$. To reliably integrate the field-lines over such a distance, we use a volume preserving method [@finnchacon05] that ensures $\boldsymbol{\nabla}\cdot \boldsymbol{B} = 0$ to numerical round-off and has been shown to well reproduce boundaries between domains of ordered and stochastic magnetic field [@ciaccio13]. Figure \[fig:poincare\] (top) shows a Poincaré section at $t\Omega_{ci} = 88$, which is at the start of the non-linear phase of the oblique plasmoid instability. As well as the upstream unreconnected flux, there are clearly visible magnetic flux-surfaces in the downstream region showing a magnetic island. This island is seeded in the single mode perturbation of Eq. (\[initialpert\]) and remains stable as it grows due to the quasi-2D nature of the Sweet-Parker reconnection. In between the upstream and downstream flux-surfaces is a thin region of stochastic magnetic field caused by the overlap of oblique magnetic flux-ropes. At the later times of $t\Omega_{ci} = 132$ (second panel) and $t\Omega_{ci}=396$ (third panel) the size of the stochastic region increases until it fills a significant part of the simulation volume at saturation. Within this middle volume there is no indication of any structure, suggesting that the “flux-ropes” that are visible in Fig. \[fig:snapshots\] do not confine magnetic field-lines over such large distances. To compare the regions of magnetic field mixing with plasma mixing, we consider the electron temperature $T_e$. The red contour in Fig. \[fig:poincare\] is $T_e/T_{e0} = 1.15$, just above the background value. Although this contour covers a significant part of the stochastic region, there are clear regions where the magnetic field is stochastic outside of this contour (choosing lower threshold values for the contour do not give better agreement). This result is similar to the test-particle study of Ref. \[\], where the electron mixing region was found to be somewhat smaller than the stochastic magnetic field region. In the present simulation, where the plasma and fields are self-consistently coupled, the finite electron velocity may limit the spread of electrons along the full volume of the stochastic region. To test this, we plot the magnetic field line exponentiation factor $\sigma$, which measures the exponential rate of separation of neighbouring magnetic field-lines [@boozer12; @daughton14; @le18]. It is defined as $$\sigma = \ln(\rho_{\textrm{max}}^{1/2}),$$ where $\rho_\textrm{max}$ is the maximum eigenvalue of the Cauchy-Green deformation tensor $(\nabla_{\boldsymbol{x}_0}\boldsymbol{x}_f)(\nabla_{\boldsymbol{x_0}}\boldsymbol{x}_f)^T$ of the field-line mapping $\boldsymbol{x}_0\rightarrow \boldsymbol{x}_f(\boldsymbol{x}_0)$. Here $\boldsymbol{x}_0$ are taken to be an array of seed points in the $y=L_y/2$ surface and $\boldsymbol{x}_f$ are the final positions after integrating a distance $L_s=L_y/2$ along the magnetic field-lines. The exponentiation factor is similar to the squashing degree $Q$, used to define quasi-separatrix layers [@titov02; @priest95], and the finite time Lyapunov exponent often used to characterize fluid flows. Fig. \[fig:poincare\] shows that the region of significant $\sigma$ agrees more closely with the electron temperature contour than the stochastic region shown in the Poincaré plot. The maximum $\sigma = 6.7$ at this time. We also find that the agreement between the $T_e$ contour and the region of significant $\sigma$ is fairly close for most of the simulation (not shown), apart from at early time where there is rapid change in $T_e$ due to Joule heating in the current layer. This suggests that the snapshots of $T_e$ in Fig. \[fig:snapshots\] trace out the magnetic topology to a reasonable degree, but due to the finite electron velocity they do not explore the whole stochastic region instantly. \[sec:kinetictransition\]Transition to kinetic reconnection =========================================================== {width="80.00000%"} The transition from collisional to kinetic reconnection has been previously studied using 2D first-principles simulations in Refs. \[\]. When the current layer thickness falls below the ion kinetic scale (either by thinning of laminar layers, or by new layers forming between magnetic islands due to the plasmoid instability), the reconnection electric field is observed to become larger than the critical Dreicer threshold $E_D$. This triggers rapid thinning of the current layer until it reaches electron kinetic scales ($\approx 2 d_e$) [@daughton09b]. As this occurs, resistive friction is no longer sufficient to balance the electric field and it is instead supported by gradients in the off-diagonal elements in the electron pressure tensor at the X-point location [@roytershteyn10]. The previous studies report these results for measurements taken at a single point in space - the primary X-point of the 2D reconnection layer. In this section, the physics of this transition is examined for the current simulation, with focus on the 3D spatial locations where signatures of kinetic reconnection occur. Figure \[fig:epar-agyro\] (top) shows an isovolume of $E_{\parallel}/E_D \geq 1$ (purple), the ratio of the parallel electric field to the critical Dreicer field at $t\Omega_{ci}=132$. The magnetic surfaces are indicated by a contour of $T_e$ (orange), which shows the flux-ropes have grown large enough to break-up the primary current layer. Intense current-layers that form between the flux-ropes are found to reach thicknesses on the electron kinetic scale $\delta \approx 1.5 d_e$ (not shown), in agreement with the findings of the previous 2D studies. The spatial locations of the super-Dreicer parallel electric fields are in good agreement with the locations of these thin current layers, and reach values as large as $E_\parallel/E_D = 5$. The bottom panel of Fig. \[fig:epar-agyro\] shows an isovolume of the electron pressure agyrotropy with $A\phi_e\geq 0.08$ (green). This agyrotropy is a scalar measure of the departure of the pressure tensor from cylindrical symmetry about the magnetic field [@scudder08], and significant values of $A\phi_e$ have been observed in both simulations and spacecraft data [@scudder12] at sites of collisionless magnetic reconnection. The isovolume of $A\phi_e \geq 0.08$ also appears to be spatially co-located with the isovolume of $E_\parallel/E_D\geq 1$, and the intense current layers that form between the flux-ropes. More quantitatively, there is a moderate positive correlation between $E_\parallel/E_D$ and $A\phi_e$ (Pearson coefficient $0.6$) in regions where the electric field is super-Dreicer, $E_\parallel/E_D\geq 1$. The primary mechanism for the generation of the agyrotropic electron distributions is presently unclear, although several possibilities have been suggested based upon tracking particles in simulations of collisionless reconnection with strong electric field gradients [@wendel16]. Since electron collisions act to isotropize the pressure tensor, significant $A\phi_e$ is taken here to be a signature of the transition to kinetic reconnection. ![\[fig:ohmslawsafter\]Contributions to parallel force balance at $t\Omega_{ci}=132$. Contributions have been averaged in time and along the field-lines - see text for definitions. Quantities are expressed in ion units, after normalization by $n_0B_{r0}v_{A}$. Inset: Decomposition of electron pressure tensor (green) into gyrotropic (cyan) and non-gyrotropic (magenta) components.](FLint-ohms-L20de-pvcols-w-decomp-INSET-ionunits-eps-converted-to.pdf){width="50.00000%"} Figure \[fig:ohmslawsafter\] shows the electron momentum balance at $t\Omega_{ci0} = 88$, from a line of seed points at $\boldsymbol{x}_0 = (428,0,z)$ for $z\in [z_s-12,z_s+12]$, $z_s=L_z/2$, and integrated a distance of $L_s=10$ ($d_e$). This small value of $L_s$ was chosen to prevent apparent broadening of the non-ideal electric field region when integrating along stochastic field-lines that exit the kinetic scale diffusion regions, but still have $E_\parallel \neq 0$ due to finite plasma resistivity. In contrast to Fig. \[fig:ohmslaws\], the region with finite $<n_e e E_\parallel>$ (orange) is significantly thinner, with half-thickness $\approx 1.5$ ($d_e$). The electron pressure tensor term (green) is now the largest balancing the electric field term at the center of the current layer. To examine this further, inset shows the break-down of the electron pressure term into the gyrotropic (cyan) and non-gyrotropic (magenta) components, where $\boldsymbol{P}_{eG} = P_{e\parallel}\boldsymbol{\hat{b}}\boldsymbol{\hat{b}} + P_{e\perp}\left(\mathbb{I}-\boldsymbol{\hat{b}}\boldsymbol{\hat{b}}\right)$, $\boldsymbol{P}_{eNG} = \boldsymbol{P}_{e}-\boldsymbol{P}_{eG}$, $P_{e\parallel} = \boldsymbol{P}_{e}:\boldsymbol{\hat{b}}\boldsymbol{\hat{b}}$ and $P_{e\perp} = \left(\textrm{Tr}[\boldsymbol{P}_{e}]-P_{e\parallel}\right)/2$. The non-gyrotropic part is significant in a thin region $\delta \approx \rho_e < d_e$ in the very center of the current layer, but on either side the gyrotropic part has a larger contribution to $<n_e e E_\parallel>$. This gyrotropic part has been observed in 3D collisionless reconnection simulations in Refs. \[\], \[\], although it is identically zero at the X-point in 2D simulations due to symmetry. As such, the role of this term in decoupling electrons from magnetic field-lines and permitting reconnection remains unclear. Strictly, $E_\parallel \neq 0$ is not a sufficient condition for reconnection in 3D, where the condition $\boldsymbol{\nabla}\times (\boldsymbol{E}+\boldsymbol{u}_e\times \boldsymbol{B})\neq 0$ is more appropriate [@hornig07]. Only the part of the gyrotropic term with non-zero curl is able break the electron frozen-in condition and determine the electron diffusion region but, unfortunately, the noise level in the components of the non-ideal electric field is too large to reliably compute the derivatives needed to examine this issue. Nevertheless, the $\approx 1.5 d_e$ half-thickness of the non-ideal region remains in good agreement with 2D collisionless simulations, as well as the significant role of both the electron inertia and non-gyrotropic pressure tensor terms in balancing the non-ideal electric field. These results are taken as confirmation of the transition to kinetic reconnection, which occurs within these thin current layers on the electron kinetic scale. \[sec:conclude\]Summary and discussion ====================================== The transition from collisional to kinetic magnetic reconnection was studied for the first time in 3D, using a first-principles kinetic approach with a Monte-Carlo treatment of the Fokker-Planck collision operator. Initial reconnection in the low-$\beta$ force free current sheet proceeded in a quasi-2D Sweet-Parker regime, keeping the symmetry of the initial perturbation, as expected for the “single X-line collisional” region of the reconnection phase diagram [@jidaughton11] shown in Fig. \[fig:phasediagram\]. In the low-$\beta$ sheet, intense Joule heating leads to more rapid thinning than reported for previous studies [@daughton09a; @daughton09b] with $\beta \approx 1$. While the current layer remains collisional, transport resulting from the kinetic description of collisions can include the classical effects of temperature dependent and anisotropic resistive and thermal friction, viscosity, heat conduction, and species thermal equilibration. However, a simplified resistive-MHD model that includes a Spitzer-type law for the resistivity, neglects heat conduction, and assumes equal ion and electron temperatures was found to reasonably well reproduce the current layer thinning profile for this simulation. Prior to disruption of the current layer, the resistive thinning causes the simulation to transition to the “Multiple X-line” hybrid region of Fig. \[fig:phasediagram\]. The 2D symmetry of the initial phase was broken by the oblique plasmoid instability, which occurred in the dynamically thinning Sweet-Parker sheet with well established reconnection outflow jets. In the early phase of the plasmoid instability, the tearing modes were found to be in the semi-collisional regime (with growth rates smaller than the collision frequency, $\nu_{ei}>\gamma$, and an inner layer thickness below the sound-radius, $\Delta_{SC} < \rho_s$). The growth rates for modes with small oblique angles ($\theta \lesssim 20^{\circ}$), which agreed well with linear semi-collisional theory predictions [@drakelee77], were found to be large compared to the rates of mode stretching and rotation by the outflow jets. However, strongly oblique modes were stabilized at a much lower angular cut-off ($\theta \approx 35^{\circ}$) than predicted for standard tearing modes. The presence of electron temperature gradients from the Joule heating was considered as a mechanism for this observed stabilization, but a theory accounting for this physics [@connor12] was also found to underpredict the amount of stabilization. The precise reason for the stabilization observed in the present simulation remains an open question, and it is possible that the validity of boundary layer theory is violated for the strongly oblique modes [@baalrud18]. Despite this narrow angular spectrum of oblique modes in the linear regime, magnetic energy is subsequently injected into oblique fluctuations by a combination of flux-rope rotation by the reconnection outflow jets, and secondary kink and coalescence instabilities. A region of stochastic magnetic field is formed by the plasmoid instability, which grows over time as more flux is reconnected, and agrees reasonably well with the observed extent of electron heat transport. Apart from long wavelength variations in the $y$-direction, the transition to kinetic reconnection proceeds in a manner similar to 2D simulations [@daughton09a; @daughton09b]. The parallel electric field becomes super-Dreicer ($1 \lesssim E_\parallel/E_D \lesssim 5$) at kinetic-scale current layers that form between the oblique flux-ropes, and a significant part of the parallel force is balanced by electron pressure tensor and inertia terms (although the former has a gyrotropic component [@liu13] not present in 2D). Secondary flux-ropes are observed to form in these thin current layers at late time in the simulation. The overall behavior described for this 3D simulation supports the picture of the plasmoid mediated transition to kinetic reconnection in the “Multiple X-line” hybrid regime, as indicated in Fig. \[fig:phasediagram\]. Solar flare reconnection, which occurs in low-$\beta$ force free current layers, is also argued to be in the “Multiple X-line hybrid” regime based on present understanding. With a flare Lundquist number of $S\sim 10^{13}$ and system-size [@jidaughton11] $\lambda = L/\delta_i \sim 4\times 10^7$, direct numerical simulation is unfeasible in the near future and we are left to extrapolate from smaller simulation and experimental studies. The $S\sim 10^{3-4}$ and $\lambda \sim 10^{2-3}$, as well as the low-$\beta$ initial conditions used for this paper are relevant to the newly constructed Facility for Laboratory Reconnection Experiments (FLARE [@jiflare18]). Recently, a number of laboratory magnetic reconnection experiments have observed the break up of current layers due to the plasmoid instability [@olsen16; @jaraalmonte16; @hare17], but the plasmoid mediated transition from collisional to kinetic reconnection has not yet been studied in detail. The full picture of the plasmoid instability in FLARE should take into account the resistive thinning of the Sweet-Parker layer that forms due to inductive current drive, the influence of the flux-core boundary conditions on the growth of oblique modes, the semi-collisional inner layer physics, and the role of outflow jets in the stretching and rotation of modes. It may also require the consideration of ion-neutral and neutral-neutral collisions [@jaraalmonte19]. Future work will extend the present study to experimentally realistic cylindrical geometry of the FLARE experiment, including the relevant physics, to better enable comparisons to be drawn. This work is supported by the Basic Plasma Science Program from the U.S. Department of Energy, Office of Fusion Energy Sciences. The large simulation was performed at the National Energy Research Scientific Computing Center (NERSC), a U.S. Department of Energy Office of Science User Facility operated under Contract No. DE-AC02-05CH11231. Supporting simulations used resources from the Los Alamos National Laboratory Institutional Computing Program, which is supported by the U.S. Department of Energy National Nuclear Security Administration under Contract No. DE-AC52-06NA25396. \[apdx:drifttearing\]$\Delta_{\textrm{crit}}^\prime$ due to temperature gradients ================================================================================= Ref. \[\] gives a theory for the semi-collisional drift-tearing and internal kink instabilities for arbitrary plasma-$\beta$ and $\Delta^{\prime}$, including ion-orbit effects via a gyrokinetic treatment. In general the dispersion relations need to be computed numerically, but closed form expressions can be found in certain asymptotic limits. Firstly, the semi-collisional theory of Eq. (\[semicollisionalnormal\]) can be found [@zocco15] in the limit of cold ions and small-$\Delta^\prime$. Secondly, a dispersion relation can be found for the strong drift regime (including finite ion orbits) by expanding $\omega = \omega_r + i\gamma$ in powers of $(\Delta/\rho_i)\ll 1$, where $\Delta$ is the semi-collisional layer thickness and $\rho_i$ is the ion gyroradius. Including only electron temperature gradients [@zocco15] so $\Delta=\Delta_T$ defined in Eq. (\[deltaT\]), and neglecting density and ion temperature gradients, the lowest order frequency $\omega_0$ is real: $$\frac{\omega_0}{\omega_T} = \frac{1.71 \sqrt{1+\tau}}{\sqrt{1+\tau} + \sqrt{2.13 \tau}},$$ where $\tau = T_e/T_i$, and $\omega_T = k T_e/(eBL_T)$. At the next order the growth rate scales as $$\frac{\gamma}{\omega_T} \sim \frac{\Delta_T}{\pi \hat{\beta}_T} \left[\Delta^\prime - \Delta^\prime_{\textrm{crit}}\right],$$ where $\beta_T = (\beta_e/2) L_s^2/L_T^2$. The critical threshold for instability, $\Delta^\prime_{\textrm{crit}}$, is given by the expression $$\label{deltaprimecrit}\Delta^\prime_{\textrm{crit}} = \frac{\sqrt{\pi}\hat{\beta}_T}{\rho_i} \frac{\omega_0^2}{\omega_T^2} \frac{\tau}{(1+\tau)^2}\ln\left[\frac{\rho_i}{\Delta_T} \sqrt{\frac{\omega_T}{2\omega_0}}\right] - \frac{\pi \hat{\beta}_T}{\rho_i} \frac{\omega_0^2}{\omega_T^2} \bar{I}(\tau).$$ Here, the integral $\bar{I}(\tau)$ is from the gyrokinetic ions [@connor12; @zocco15]. It is given by $$\bar{I}(\tau) = \int_0^{\infty} dk \left[\frac{F(k)}{G(k)} - \frac{\tau}{1+\tau} + \frac{\tau}{\sqrt{\pi} (1+\tau)^2 (1+k)}\right],$$ with $$F(k) = \tau \left[\exp{(-k^2/2)}I_0(k^2/2)-1\right],$$ $I_0$ is the modified Bessel function of the first kind, and $G(k) = F(k)-1$. To calculate $\Delta^\prime_{\textrm{crit}}$ that is used in Fig. \[fig:growthrates\], this integral is calculated numerically based upon the local $\tau$ at each rational surface. The integral is negative (it is stabilizing), and has typical value $\bar{I} \approx -0.5$ for the parameters used here. [92]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****,  ()](\doibase 10.1007/s001590100013) [****,  ()](\doibase 10.1038/nphys2675),  [****,  ()](\doibase 10.1103/PhysRevLett.6.47) [****,  ()](\doibase 10.1126/science.aaf2939) [****,  ()](\doibase 10.1103/PhysRevLett.33.1201) @noop [****,  ()]{} [****,  ()](\doibase 10.1088/0741-3335/53/1/013001) [****,  ()](\doibase 10.1103/PhysRevLett.114.205003) [****,  ()](\doibase 10.1063/1.4830104),  [****, ()](\doibase 10.1063/1.3647505),  [****,  ()](\doibase 10.1063/1.3633473),  [****,  ()](\doibase 10.1007/s11214-011-9766-z) [****, ()](\doibase 10.1063/1.4811120) [****,  ()](\doibase 10.1063/1.4802941),  [****,  ()](\doibase 10.1017/S0022377814000907) [****,  ()](\doibase 10.1088/0741-3335/58/1/014021),  [****,  ()](\doibase 10.3847/1538-4357/aa7b82),  [****,  ()](\doibase 10.1063/1.4964481),  [****,  ()](\doibase 10.3847/1538-4357/aa906d),  @noop [****,  ()]{} [****,  ()](\doibase 10.1017/S0022377800020663) [****, ()](\doibase 10.1063/1.4804337) [****,  ()](\doibase 10.1103/PhysRevLett.80.3256) [****, ()](\doibase 10.1103/PhysRevLett.103.065004) [****, ()](\doibase 10.1029/1999JA001001) [****,  ()](\doibase 10.1063/1.2203950) [****,  ()](\doibase 10.1103/PhysRevLett.98.015003) [****,  ()](\doibase 10.1086/505690),  [****,  ()](\doibase 10.1088/2041-8205/780/2/L19) [****,  ()](\doibase 10.1103/PhysRevLett.116.105003),  [****,  ()](\doibase 10.1063/1.3264103),  [****,  ()](\doibase 10.1063/1.3274462),  [****,  ()](\doibase 10.1063/1.3420208),  [****,  ()](\doibase 10.1063/1.3703318),  [****,  ()](\doibase 10.1186/BF03353258),  [****,  ()](\doibase 10.1103/PhysRevLett.105.015004),  [****,  ()](\doibase 10.1063/1.3606363),  [****,  ()](\doibase 10.1063/1.3191718) [****, ()](\doibase 10.1088/2041-8205/806/1/L12),  [****,  ()](\doibase 10.3847/0004-637X/818/1/20),  [****,  ()](\doibase 10.3847/1538-4357/834/1/47),  [****,  ()](\doibase 10.3847/1538-4357/aa6001),  [****,  ()](\doibase 10.1063/1.3274462),  [****,  ()](\doibase 10.1103/PhysRev.115.238) [****,  ()](\doibase 10.1063/1.3399787) [****, ()](\doibase 10.1007/s11214-006-9107-9) [****,  ()](\doibase 10.1088/0004-637X/714/2/1108) in @noop [**]{} () p.  [****,  ()](\doibase 10.1038/nphys1965) [****,  ()](\doibase 10.1063/1.3678211),  [****,  ()](\doibase 10.1063/1.862017) [****,  ()](\doibase 10.1088/0741-3335/54/3/035003),  [****, ()](\doibase 10.1063/1.5020777),  [****,  ()](\doibase 10.1063/1.2840133) [****,  ()](\doibase 10.1016/0021-9991(77)90099-7) @noop [****,  ()]{} [****,  ()](\doibase 10.1063/1.5027086),  [****,  ()](\doibase 10.1103/PhysRev.117.329) @noop [****,  ()]{} [****, ()](\doibase 10.1063/1.2838244) [****,  ()](\doibase 10.1103/PhysRev.89.977) [****, ()](\doibase 10.1063/1.1706761) [****,  ()](\doibase 10.1103/PhysRevLett.42.1058) [****, ()](\doibase 10.1063/1.865841) [****,  ()](\doibase 10.1103/PhysRevE.87.013102),  [****,  ()](\doibase 10.1103/PhysRevLett.110.265004) [****, ()](\doibase 10.1063/1.4940945),  [****, ()](\doibase 10.1063/1.3628639),  [****,  ()](\doibase 10.1017/S002237781800106X),  [****,  ()](\doibase 10.1063/1.864441) [****,  ()](\doibase 10.1016/0375-9601(81)90083-9) [****,  ()](\doibase 10.1088/0741-3335/57/6/065008),  [****,  ()](\doibase 10.1098/rspa.1958.0079),  [****, ()](\doibase 10.1063/1.1889156) [****,  ()](\doibase 10.1063/1.4811380) [****, ()](\doibase 10.1063/1.5004613) [****, ()](\doibase 10.1063/1.4765352) [****, ()](\doibase 10.1063/1.4875730) [****,  ()](\doibase 10.1029/2001JA000278) [****, ()](\doibase 10.1029/95JA02740) [****,  ()](\doibase 10.1029/2008JA013035) [****,  ()](\doibase 10.1103/PhysRevLett.108.225005) [****,  ()](\doibase 10.1063/1.4942031) [****, ()](\doibase 10.1063/1.5007860) @noop [**]{}, edited by  and  (, ) pp.  [****,  ()](\doibase 10.1103/PhysRevLett.116.255001) [****,  ()](\doibase 10.1103/PhysRevLett.117.095001) [****,  ()](\doibase 10.1103/PhysRevLett.118.085001),  [****,  ()](\doibase 10.1103/PhysRevLett.122.015101) [^1]: $(T_i+T_e)$ is $20\%$ larger in MHD at $t\Omega_{ci}=60$, presumably due to neglecting heat conduction [^2]: Here $T_e/T_{e0} \approx 4$ and $S_\parallel=1.6\times 10^4$ at $t\Omega_{ci} = 88$, compared with $T_e/T_{e0}\approx 1.6$ for a simulation with $\eta_\perp = 0.04$, $\delta_0/d_{i0}=1$, and $L_x=100d_i$ in Ref. \[\], and $S_{\textrm{max}} = 2500$ for a simulation with $\eta_\perp = 0.04$, $\delta_0/d_{i0}=2$, and $L_x=200d_i$ in Ref. \[\]. [^3]: In Alfvén units [@zocco11], this is the same small-$\Delta^\prime$ growth rate used for a recent model of the semi-collisional plasmoid instability [@bhat18]: $\gamma_{SC}/\omega_A \sim (kL_S)^{2/3} S^{-1/3} (\Delta^\prime \rho_s)^{2/3}$, where $\omega_A = L_s/v_A$, $S=L_s v_A/\eta$, and $\eta = d_e^2\nu_{ei}$.

ArXiv

--- abstract: 'For a characteristic $p > 0$ variety $X$ with controlled $F$-singularities, we state conditions which imply that a divisorial sheaf is Cohen-Macaulay or at least has depth $\geq 3$ at certain points. This mirrors results of Kollár for varieties in characteristic zero. As an application, we show that relative canonical sheaves are compatible with arbitrary base change for certain families with sharply $F$-pure fibers.' address: - | Department of Mathematics\ Princeton University\ Princeton, NJ, 08542, USA - | Department of Mathematics\ The Pennsylvania State University\ University Park, PA, 16802, USA author: - Zsolt Patakfalvi AND Karl Schwede title: 'Depth of $F$-singularities and base change of relative canonical sheaves' --- [^1] Introduction {#sec:introduction} ============ In the paper [@KollarALocalKawamataViehweg], Kollár proved that sheaves ${\mathcal{O}}_X(-D)$ satisfy strong depth conditions if $D$ is locally $\bQ$-linearly equivalent to a divisor $\Delta$ such that $(X, \Delta)$ is SLC or KLT. These results generalized [@AlexeevLimitsOfStablePairs Lemma 3.2], [@FujinoIntroductionToTheLMMP] and [@KollarMori Corollary 5.25]. Because depth conditions can be interpreted as vanishing of local cohomology, these results were described as a local version of the Kawamata-Viehweg vanishing theorem. In this paper, we obtain characteristic $p > 0$ analogs of the main results of [@KollarALocalKawamataViehweg]. This is particularly interesting because the (global) Kawamata-Viehweg vanishing theorem is false in positive characteristic [@raynaud_contre-exemple_1978]. We replace the KLT and SLC conditions by strongly $F$-regular and sharply $F$-pure singularities respectively (such characteristic $p > 0$ singularity classes are known as $F$-singularities). For the convenience of the reader, we recall that by [@HaraWatanabeFRegFPure; @MillerSchwedeSLCvFP] - KLT pairs correspond philosophically to strongly $F$-regular pairs, and - SLC pairs correspond philosophically to sharply $F$-pure pairs. Similar to [@KollarALocalKawamataViehweg], we can apply our results on depth to prove base change for relative canonical sheaves. [**(A special case of) .** ]{}[*[([[*cf.*]{}]{}[@KollarALocalKawamataViehweg 4.3])]{} Let $f : X {\xrightarrow{\ \ }}Y $ be a flat morphism of finite type with [$\textnormal{S}_2$]{}, [$\textnormal{G}_1$]{} equidimensional fibers to a smooth variety and let $\Delta \geq 0$ be a $\mathbb{Q}$-divisor on $X$ avoiding all the codimension zero and the singular codimension one points of the fibers. Further suppose that $K_X + \Delta$ is $\bQ$-Cartier, $p \notdivide \operatorname{ind}(K_X + \Delta)$ and $(X_y, \Delta_y)$ is sharply $F$-pure for every $y \in Y$. Then $\omega_{X/Y}$ is flat over $Y$ and compatible with arbitrary base change.* ]{} is hoped to be useful in constructing a moduli space for varieties of general type in positive characteristics. See [@PatakfalviBaseChange] for further explanation, and also for examples for which the above compatibility does not hold. We also remark here that $\omega_{X/Y}$ behaves surprisingly well with respect to base-change. It obeys base-change for example when the fibers are Cohen-Macaualay [@ConradGDualityAndBaseChange Theorem 3.6.1]. In particular, this pertains to families of normal surfaces. In contrast, the higher reflexive powers, $\omega_{X/Y}^{[m]}$ for $m >1$, are not compatible with base change in the surface case [@HaconKovacsClassificationOfHigherDimensional Section 14.A]. Similar differences between canonical and pluricanonical sheaves have been observed earlier [@KollarSubadditivityOfKodairaDimension page 2], [@KollarProjectivityOfCompleteModuli Remark 4.4]. The technical result on depth used to prove is as follows. It is a characteristic $p > 0$ version of [@KollarALocalKawamataViehweg Theorem 3(1)], also compare with [@AlexeevLimitsOfStablePairs Lemma 3.2], [@FujinoIntroductionToTheLMMP Theorem 4.21], [@AlexeevHaconNonRationalCenters Theorem 1.5] and [@KovacsIrrationalCenters Theorem 1.2, 1.5]. [**.** ]{}[ *[([[*cf.*]{}]{}[@KollarALocalKawamataViehweg Theorem 3(1)])]{} Suppose that $R$ is local, [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{} and that $0 \leq \Delta$ is an $\bR$-divisor on $X = \Spec R$ with no common components with the singular locus of $X$ and such that $(X, \Delta)$ is sharply $F$-pure. Set $x \in X$ to be the closed point and assume that $x$ is not an $F$-pure center of $(X, \Delta)$. Suppose that $0 \leq \Delta' \leq \Delta$ is another $\bR$-divisor and that $r\Delta'$ is integral for some $r > 0$ relatively prime to $p$. Further assume that $M$ is any rank-1 reflexive subsheaf of $K(X)$ such that $M^{(-r)} \cong {\mathcal{O}}_X( r\Delta')$ (here $\blank^{(\cdot)}$ denotes reflexive power). Then $$\depth_x M \geq \min\{3, \codim_X x\}.$$* ]{} Another interesting depth statement, again completely analogous to a theorem of Kollár is below. In the introduction we phrase it in the language of Frobenius splittings [@BrionKumarFrobeniusSplitting], but in the text it is phrased slightly more generally. [**, .**]{} [*[([[*cf.*]{}]{}[@KollarALocalKawamataViehweg Theorem 3(2)])]{} Suppose that $(R, \bm)$ is an [$\textnormal{S}_2$]{} local ring with Frobenius splitting ${\varphi}: F^e_* R {\xrightarrow{\ \ }}R$ which is not compatibly split with $\bm$. Additionally suppose that $Z$ is any union of compatibly ${\varphi}$-split subvarieties of $X = \Spec R$ such that no irreducible component of $Z$ coincides with an irreducible component of $X$. Suppose that $I_Z \subseteq R$ is the ideal defining $Z$, then $$\depth_{\bm} I_Z \geq \min\{3, 1+\codim_Z x\}.$$* ]{} The other main statement on depth we obtain, , asserts that if $(X, \Delta)$ is strongly $F$-regular and $(p^e - 1)D$ is an integral divisor linearly equivalent to $(p^e - 1)\Delta$, then ${\mathcal{O}}_X(-D)$ is Cohen-Macaulay. Compare with [@KollarALocalKawamataViehweg Theorem 2]. One should also compare the above results on depth, as well as the related characteristic zero results, to [@AberbachEnescuStructureOfFPure Theorem 4.8(vi)] where Aberbach and Enescu showed that the depth of an $F$-pure ring $R$ is always $\geq$ than the dimension of the minimal $F$-pure center (ie, of the dimension of $R$ modulo the splitting prime, which we know is equal to the $s$-dimension of [@AberbachEnescuStructureOfFPure] by [@BlickleSchwedeTuckerFSigPairs1]). [*Acknowledgements:*]{} The authors began working on this project at the workshop *ACC for minimal log discrepancies and termination of flips* held at the American Institute of Mathematics and organized by Tommaso de Fernex and Christopher Hacon. The authors would also like to thank Florian Enescu for valuable discussions as well as thank the referee, János Kollár and Sándor Kovács for many useful comments on a previous draft. Preliminaries on $F$-singularities ================================== Throughout this paper, all schemes are Noetherian, separated, of equal characteristic $p > 0$ and $F$-finite.[^2] Note that any such scheme $X$ is automatically locally excellent by [@KunzOnNoetherianRingsOfCharP] and also has a dualizing complex by [@Gabber.tStruc]. In particular, we are implicitly assuming all schemes are locally excellent and possess dualizing complexes. Little will be lost to the reader if he or she considers only schemes that are essentially of finite type over a perfect field. We remind the reader of some special divisors on non-normal schemes. We follow the notation of [@KollarFlipsAndAbundance Section 16]. For an [$\textnormal{S}_2$]{} reduced local ring $R$, set $X = \Spec R$. We define a *[W-divisor]{}* (or *Weil divisor*) to be a formal sum of codimension one subsets of $X$ *whose generic points are not singular points of $X$*. This has the same data as divisors on the regular locus of $X$ or as rank-1 [$\textnormal{S}_2$]{} submodules $M$ of $K(R)$ (the total ring of fractions of $R$) such that $M_{\eta} = R_{\eta}$ as a subset of $K(R)$, for every codimension 1 singular point $\eta$ of $X$. Later in the paper, we will need to instead work with the more general notion of Weil divisorial sheaves $\operatorname{{\bf WSh}}(X)$, rank one reflexive subsheaves of $K(X)$ that are invertible in codimension 1. In the non-local setting, such divisors are simply formal sums of irreducible subschemes that satisfy this definition locally. We now set $\operatorname{{\bf{WDiv}}_{\bQ}}(X) := \operatorname{{\bf{WDiv}}}(X) \tensor_{\bZ} \bQ$ and $\operatorname{{\bf{WDiv}}_{\bR}}(X) := \operatorname{{\bf{WDiv}}}(X) \tensor_{\bZ} \bR$. Note we have containments: $$\operatorname{{\bf{WDiv}}}(X) \subseteq \operatorname{{\bf{WDiv}}_{\bQ}}(X) \subseteq \operatorname{{\bf{WDiv}}_{\bR}}(X).$$ One can also form $\operatorname{{\bf WSh}_{\bQ}}(X) := \operatorname{{\bf WSh}}(X) \tensor_{\bZ} \bQ$ and $\operatorname{{\bf WSh}_{\bZ_{(p)}}}(X) := \operatorname{{\bf WSh}_{\bZ_{(p)}}}(X) \tensor_{\bZ} \bZ_{(p)}$ similarly, but the natural maps $\operatorname{{\bf WSh}}(X) {\xrightarrow{\ \ }}\operatorname{{\bf WSh}_{\bZ_{(p)}}}(X) {\xrightarrow{\ \ }}\operatorname{{\bf WSh}_{\bQ}}(X)$ are not necessarily injective [@KollarFlipsAndAbundance Section 16]. Given $\sum_i a_i D_i = \Delta \in \operatorname{{\bf{WDiv}}_{\bR}}(X)$, we use $\lceil \Delta \rceil$ to denote $\sum_i \lceil a_i \rceil D_i$ (such roundings are not necessarily well defined for $\operatorname{{\bf WSh}_{\bZ_{(p)}}}(X)$ or $\operatorname{{\bf WSh}_{\bQ}}(x)$). Finally, given $D \in \operatorname{{\bf{WDiv}}}(X)$, we use ${\mathcal{O}}_X(D)$ (or $R(D)$) to denote the corresponding subsheaf of $\sK(X)$ (or of $K(R)$) in the usual way. Note that $D$ is effective if and only if ${\mathcal{O}}_X(D) \supseteq {\mathcal{O}}_X$. Now we move away from divisors. Suppose that $R$ is a ring of characteristic $p > 0$. Following [@SchwedeTestIdealsInNonQGor; @BlickleTestIdealsViaAlgebras], we say a *Cartier subalgebra* $\sC$ is a graded subring of the graded ring $$\bigoplus_{e \geq 0} \Hom_R(F^e_* R, R) =: \sC^R$$ where multiplication is done by Frobenius twisted composition[^3] such that the zeroth graded piece $[\sC]_0 = \Hom_R(R, R) \cong R$. We note that even though we call $\sC$ a *Cartier subalgebra*, it is not an $R$-algebra because $R \cong [\sC]_0$ is not necessarily central. \[ex.CartierAlgebras\] Given an [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{} ring $R$, set $X = \Spec R$ and assume that $0 \leq \Delta \in \operatorname{{\bf{WDiv}}_{\bR}}(X)$ on $X$ (for example, if $R$ is normal, $\Delta$ is simply an $\bR$-divisor). We can form the Cartier subalgebra $\sC^{\Delta}$ where $$[\sC^{\Delta}]_e := \Hom_R(F^e_* R( \lceil (p^e - 1)\Delta \rceil), R) \subseteq \Hom_R(F^e_* R, R).$$ Suppose that $\sC^R$ is as above and ${\varphi}\in [\sC^R]_e$ for some $e > 0$. Then we can form the Cartier subalgebra $R\langle {\varphi}\rangle$ generated by $R = [\sC]_0$ and ${\varphi}$. Explicitly, this is the direct sum $R \oplus ({\varphi}\cdot (F^e_* R) ) \oplus ({\varphi}^2 \cdot (F^{2e}_* R)) \oplus \cdots$. Now we define sharply $F$-pure pairs and $F$-pure centers. If $\sC$ is a Cartier subalgebra on $R$, then we say that the pair $(R, \sC)$ is *sharply $F$-pure* if there exists some ${\varphi}\in [\sC]_e$ for some $e \geq 1$ such that ${\varphi}(F^e_* R) = R$. In particular, if $(R, \Delta)$ is a pair as in , then we say that $(R, \Delta)$ is *sharply $F$-pure* if the associated $(R, \sC^{\Delta})$ is sharply $F$-pure. If $(R, \sC^R)$ is sharply $F$-pure, then we simply say that $R$ is *$F$-pure*. If $(R, \sC)$ is a pair as above, then an ideal $I \subseteq R$ is called *$\sC$-compatible* if ${\varphi}(F^e_* I) \subseteq I$ for all ${\varphi}\in [\sC]_e$ and all $e \geq 0$. In the case that $\sC = R\langle {\varphi}\rangle$, we will sometimes simply say that $I$ is *${\varphi}$-compatible*. An irreducible closed set $W = V(Q) \subseteq \Spec R = X$, for some $Q \in \Spec R$, is called an *$F$-pure center* if the following two conditions hold: - The localization $(R_Q, \sC_Q)$ is sharply $F$-pure, and - For every for $e \geq 0$ and ${\varphi}\in [\sC]_e$, we have ${\varphi}(F^e_* Q) \subseteq Q$ (in other words, if $Q$ is $\sC$-compatible). Likewise we say that $W$ is an $F$-pure center of $(R, \Delta)$ if it is an $F$-pure center of $(R, \sC^{\Delta})$ where $\sC^{\Delta}$ is associated to $\Delta$ as in . We also define strongly $F$-regular pairs. If $R$ is a local ring, a pair $(R, \sC)$ is called *strongly $F$-regular* if the only proper $\sC$-compatible ideals of $R$ are $0$ and $R$ itself. If $R$ is not local, then we say $(R, \sC)$ is *strongly $F$-regular* if every localization is. A pair $(R, \Delta)$ is strongly $F$-regular if $(R, \sC^{\Delta})$ is strongly $F$-regular. Given a pair $(X, \Delta)$, all of the above definitions generalize to the non-affine setting by requiring them to hold at each stalk. The notion of Cartier subalgebras is somewhat more subtle in the non-affine setting however (but we will not need such generalities). We recall some facts about compatible ideals and $F$-pure centers. \[lem.FactsAboutSplittings\] Suppose that $(R, \sC)$ is a pair and $I \subseteq J \subseteq R$ are ideals. - The set of $\sC$-compatible ideals are closed under sum and intersection. - A prime ideal $Q$ is $\sC$-compatible if and only if $QR_Q$ is $\sC_Q$-compatible. - If $\sC' \subseteq \sC$ are Cartier subalgebras and $I$ is $\sC$-compatible, then $I$ is $\sC'$-compatible. - Given ${\varphi}\in \Hom_R(F^e_*R, R)$, we have ${\varphi}(F^e_* J) \subseteq J$ if and only if $J$ is $R\langle {\varphi}\rangle$-compatible. - Suppose that ${\varphi}: F^e_* R {\xrightarrow{\ \ }}R$ is surjective. Some $Q \in \Spec R$ is ${\varphi}$-compatible if and only if it is ${\varphi}^n$-compatible where $${\varphi}^n := \underbrace{{\varphi}\circ (F^e_* {\varphi}) \circ \cdots \circ (F^{(n-1)e}_* {\varphi}) }_{\text{$n$-times}}$$ - If $I$ is ${\varphi}$-compatible, then there exists a map ${\varphi}/I : F^e_* (R/I) {\xrightarrow{\ \ }}(R/I)$ such that the following diagram commutes: $$\xymatrix{ F^e_* R \ar[r]^{{\varphi}} \ar[d] & R \ar[d] \\ F^e_* (R/I) \ar[r]_{{\varphi}/I} & (R/I). }$$ Furthermore, $J \supseteq I$ is ${\varphi}$-compatible if and only if $J/I$ is ${\varphi}/I$-compatible. (This statement can also be done with Cartier subalgebras, but we will not need it). - $(R, \sC)$ is strongly $F$-regular if and only if for every $c \in R\setminus\{ \textnormal{minimal primes}\}$, there exists a ${\varphi}\in [\sC]_e$ for $e > 0$, in fact one may take $e$ to be any larger multiple, such that ${\varphi}(F^e_* c) = 1$. \(v) follows from the argument of [@SchwedeCentersOfFPurity Proposition 4.1]. (vii) can be found in this generality in [@SchwedeTestIdealsInNonQGor Proposition 3.23]. The rest are obvious. Our next goal is to give an example of a ${\varphi}$-compatible ideal that will be crucial in later sections. The main idea is that Frobenius maps and Frobenius splittings induce maps on local cohomology. Those induced maps can then be thought of as acting directly and explicitly on classes. For the convenience of the reader not already familiar with this construction, recall that if $X = \Spec R$ and $U = \Spec R \setminus \{\bm\}$, then for any coherent ${\mathcal{O}}_X$-module $M$, we have $H^i_{\bm}(M) = H^{i-1}(U, M)$ for $i > 1$ and also that $H^1_{\bm}(M) = H^{0}(U, M)/\operatorname{image}(H^0(X, M))$. We can then use the cohomology description of sheaf cohomology to define $H^i_{\bm}(M)$. For a more thorough description of local cohomology by the complex, see for example [@BrunsHerzog Section 3.5]. We now consider Frobenius action on local cohomology. The Frobenius map $R {\xrightarrow{\ \ }}F_* R$ yields $\Psi : H^i_{\bm}(R) {\xrightarrow{\ \ }}H^i_{\bm}(F_* R) \cong H^i_{\bm}(R)$. Given a class $[z] = [\dots, z_j, \dots] \in H^i_{\bm}(R)$, we have $\Psi([z]) \in H^i_{\bm}(F_* R)$. But certainly $\Psi([z]) = F^e_* [z]^{p^e} = F^e_* [\dots, z_j^{p^e}, \dots]$ is identified with raising the entries of $[z]$ to the $p^e$th power. Now we do the same computation with a Frobenius splitting. Suppose that ${\varphi}: F^e_* R {\xrightarrow{\ \ }}R$ is an $R$-linear Frobenius splitting, and so we have a map $H^i_{\bm}(F^e_* R) \xrightarrow{{\varphi}} H^i_{\bm}(R)$ induced by ${\varphi}$. Certainly ${\varphi}(F^e_* [\dots, y_j, \dots]) = [\dots, {\varphi}(F^e_* y_j), \dots]$. But now observe that for any $[z] \in H^i_{\bm}(R)$ and $r \in R$ we have that $${\varphi}(F^e_* (r \cdot [z]^{p^e})) = {\varphi}(F^e_* [\dots, r z_j^{p^e}, \dots]) = [\dots, {\varphi}(F^e_* r) z_j, \dots] = {\varphi}(F^e_* r) [z].$$ [([[*cf.*]{}]{}[@EnescuHochsterTheFrobeniusStructureOfLocalCohomology Theorem 4.1])]{} \[lem.LocalCohomologyCompatibleWithSplittings\] Suppose that $(R, \bm)$ is a local ring. Then $\Ann_R H^i_{\bm}(R)$ is compatible with every splitting[^4] ${\varphi}: F^e_* R {\xrightarrow{\ \ }}R$ of Frobenius $R {\xrightarrow{\ \ }}F^e_* R$. We have the following composition $$\xymatrix@R=1pt{ R \ar[r] & F^e_* R \ar[r]^{{\varphi}} & R\\ 1 \ar@{|->}[r] & F^e_* 1 \ar@{|->}[r] & 1. }$$ Now suppose that $r \in \Ann_R H^i_{\bm}(R)$. Then choose $[z] \in H^i_{\bm}(R)$. We want to show that ${\varphi}(F^e_* r).[z] = 0$. Now, it follows from the cohomology description of local cohomology, and ${\varphi}$’s action on it, that $$0 = {\varphi}(F^e_* 0) = {\varphi}(F^e_* ( r.[z]^{p^e})) = {\varphi}(F^e_* r).[z]$$ which completes the proof. We also recall the following fact. We include the proof because the method will be generalized later. \[lem.stronglyFRegularIsCohenMacaulay\] If $(R, \sC)$ is strongly $F$-regular, then $R$ is normal and Cohen-Macaulay. We first note that by (vii), the strong $F$-regularity hypothesis implies that there exists a Frobenius splitting $\psi$ such that $R {\xrightarrow{\ \ }}F^e_* R \xrightarrow{\psi} R$ is an isomorphism for some $e > 0$. It then easily follows that $R$ must be reduced since if not, the map $R {\xrightarrow{\ \ }}F^e_* R$ is not injective. Normality follows since the conductor ideal is compatible with every ${\varphi}\in \sC^R$ by the argument of [@BrionKumarFrobeniusSplitting Proposition 1.2.5]. For the Cohen-Macaulay condition, by working locally we assume that $(R,\bm)$ is a local domain. By local duality [@HartshorneResidues Chapter V, Theorem 6.2], each $H^i_{\bm}(R)$ is Matlis dual to $\myH^{i-\dim R} \omega_R^{\mydot}$ for some normalized dualizing complex $\omega_R^{\mydot}$. Since an element $c \in R$ annihilates a finitely generated $R$-module if and only if $c$ annihilates the Matlis dual of a module, it follows that there exists $0 \neq c \in R$ such that $c \cdot H^i_{\bm}(R) = 0$ for all $i < \dim R$. (vii) then implies that there exists ${\varphi}\in [\sC]_e$ such that the composition $$R {\xrightarrow{\ \ }}F^e_* R \xrightarrow{F^e_* (\cdot c)} F^e_* R {\xrightarrow{\ \ }}R$$ is an isomorphism. Taking local cohomology for $i < \dim R$ gives us an isomorphism: $$H^i_{\bm}(R) {\xrightarrow{\ \ }}H^i_{\bm}(F^e_* R) \xrightarrow{F^e_* (\cdot c)} H^i_{\bm}(F^e_* R) {\xrightarrow{\ \ }}H^i_{\bm}(R)$$ where the middle map is the zero map. Thus $H^i_{\bm}(R) = 0$ which completes the proof. We also state a generalization of [@SchwedeSmithLogFanoVsGloballyFRegular Theorem 4.3], similar computations were done in [@MillerSchwedeSLCvFP]. \[lem.S2G1ImpliesFrobeniusSplitYieldsDivisor\] Suppose that $X = \Spec R$ is [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{} and sharply $F$-pure. Then there exists an element $0 \leq \Delta \in \operatorname{{\bf{WDiv}}_{\bQ}}(X)$ such that $(p^e - 1)(K_X + \Delta)$ is $\bQ$-Cartier and $(X, \Delta)$ is sharply $F$-pure. A surjective map ${\varphi}\in \Hom_R(F^e_* R, R) \cong H^0(X, F^e_* {\mathcal{O}}_X( (1-p^e)K_X))$ induces an effective Weil divisorial sheaf[^5] $\Gamma_{{\varphi}}$ by [@HartshorneGeneralizedDivisorsOnGorensteinSchemes Proposition 2.9] such that $(p^e - 1)K_X + \Gamma_{{\varphi}} \sim 0$. We would like to show that $\Gamma_{{\varphi}}$ can be identified with an element of $\operatorname{{\bf{WDiv}}}(X)$. At the singular height one points $\eta$ of $X$, ${\mathcal{O}}_{X,\eta}$ is already Gorenstein. Thus we can consider the map $\Phi_{\eta}$ which generates $\Hom_{{\mathcal{O}}_{X,\eta}}(F^e_* {\mathcal{O}}_{X,\eta}, {\mathcal{O}}_{X,\eta})$ as an $F^e_* {\mathcal{O}}_{X,\eta}$-module. Set $\bm$ to be the maximal ideal of ${\mathcal{O}}_{X, \eta}$ and notice that $\bm$ is the conductor ideal since $F$-pure rings are seminormal [@HochsterRobertsFrobeniusLocalCohomology] and in particular the conductor is radical. But then $\Phi(F^e_* \bm) \subseteq \bm$ by the proof of [@BrionKumarFrobeniusSplitting Proposition 1.2.5]. Now, we know ${\varphi}_{\eta} : F^e_* {\mathcal{O}}_{X,\eta} {\xrightarrow{\ \ }}{\mathcal{O}}_{X,\eta}$ is equal to $\Phi(F^e_* (r \cdot \blank))$ for some $r \in {\mathcal{O}}_{X, \eta}$. We want to show that $r$ is a unit, which would prove that $\Gamma_{{\varphi}}$ is trivial at $\eta$. Since ${\varphi}_{\eta}$ is surjective, we see that $r \notin \bm$ and thus $r$ is a unit. This implies that the Weil divisorial sheaf $(\Gamma_{{\varphi}})_{\eta}$ coincides with ${\mathcal{O}}_{X,\eta}$ and thus $\Gamma_{{\varphi}} \in \operatorname{{\bf{WDiv}}}(X)$ as desired. Finally, set $\Delta = {1 \over p^e - 1} \Gamma_{{\varphi}}$. We conclude by recalling a well known lemma on the height of annihilators of local cohomology modules. However, because we lack a reference, we provide a proof. \[lem.DimensionsOfLocalCohomology\] Suppose that $(R, \bm)$ is a local ring and suppose that $M$ is a finitely generated $R$-module which is $\textnormal{S}_n$ in the sense[^6] of [@BrunsHerzog]. Set $Y_i = V\big(\Ann_R( H^i_{\bm}(M))\big) \subseteq \Spec R = X$. Suppose that $d$ is the minimum over the dimensions of the components of $\Supp M$. Then $\dim Y_i \leq i - n$ for $i < d$. Note $Y_i$ may not be the same as $\Supp H^i_{\bm}(M)$ since $H^i_{\bm}(M)$ is not finitely generated. Set $\omega_X^{\mydot}$ to be a normalized dualizing complex on $X$ (recall that all our rings are excellent and possess dualizing complexes). By local duality in the form of [@HartshorneResidues Chapter V, Theorem 6.2], it is equivalent to prove that $\dim \Supp \myH^{-i} \myR \Hom_R( M, \omega_R^{\mydot}) \leq i - n$. Suppose this is false, and thus that $W \subseteq \Supp \myH^{-i} \myR \Hom_R( M, \omega_R^{\mydot})$ is an irreducible component of dimension $t > i - n$ for some $i < d$. Set $\gamma$ to be the generic point of $W$ (which we also view as a prime ideal). By localizing at $\gamma$, we see that $$(\myH^{-i} \myR \Hom_R( M, \omega_R^{\mydot}))_{\gamma} = \myH^{-i} \myR \Hom_{R_{\gamma}} (M_{\gamma}, \omega_{R_{\gamma}}^{\mydot}[t])) = \myH^{-i + t} \myR \Hom_{R_{\gamma}} (M_{\gamma}, \omega_{R_{\gamma}}^{\mydot}))$$ is supported at a point. The shift by $[t]$ is necessary to keep the dualizing complexes normalized. Thus $H^{i-t}_{\gamma}(M_{\gamma}) \neq 0$ by local duality again. Now, $i - t < n$. Also observe that $\dim M_{\gamma} \geq d - t$ (this is why the $d$ is necessary since we do not know what component of $\Supp M$ we will be restricting to). Since $M_{\gamma}$ is still $\textnormal{S}_n$, we see that $H^{j}_{\gamma}(M_{\gamma}) = 0$ for $j < \min(n, \dim M_{\gamma})$. But then $$\min(n, \dim M_{\gamma}) \geq \min(n, d-t) > i-t$$ since $n > i - t$ and $d > i$. Setting $j = i-t$ we obtain a contradiction. Depth and $F$-singularities {#sec:depth} =========================== Our goal in this section is to prove several results on the depths of sheaves on schemes with controlled $F$-singularities. First we prove our result for pairs $(R, \Delta)$ which are strongly $F$-regular, this is the simplest case. [([[*cf.*]{}]{}[@KollarALocalKawamataViehweg Theorem 2])]{} \[thm.KollarTheorem2\] Suppose that $(R, \bm)$ is local and that $(X = \Spec R, \Delta)$ is strongly $F$-regular. Further suppose that $0 \leq \Delta' \leq \Delta$ is such that $rD \sim r \Delta'$ for some integral divisor $D$ and some integer $r > 0$ realtively prime to $p$. Then ${\mathcal{O}}_X(-D)$ is Cohen-Macaulay. By possibly multiplying $r$ with an integer, we may assume that $r= p^e -1$. Choose, using , $0 \neq c \in R \setminus\{ \textnormal{minimal primes}\}$ such that $c \cdot H^i_{\bm}( {\mathcal{O}}_X(-D) ) = 0$ for all $i < \dim R $. Note that since $\Delta' \leq \Delta$, $\sC^{\Delta} \subseteq \sC^{\Delta'}$ and then $(X,\Delta')$ is strongly $F$-regular as well by (iii). Therefore, by (vii) there exists an $e > 0$ and a splitting ${\varphi}$ such that the composition $${\mathcal{O}}_X {\xrightarrow{\ \ }}F^e_* {\mathcal{O}}_X {\xrightarrow{\ \ }}F^e_* {\mathcal{O}}_X( (p^e -1)\Delta') \xrightarrow{F^e_* (\cdot c)} F^e_* {\mathcal{O}}_X( (p^e -1)\Delta') \xrightarrow{{\varphi}} {\mathcal{O}}_X$$ is an isomorphism. By replacing $e$ by a multiple if necessary, we may assume that this $e > 0$ also satisfies the condition from the hypothesis. Twisting by ${\mathcal{O}}_X(-D)$, reflexifying, and applying $H^i_{\bm}(\blank)$ we obtain the following composition which is also an isomorphism. $$\begin{array}{rl} & H^i_{\bm}({\mathcal{O}}_X(-D)) \\ {\xrightarrow{\ \ }}& H^i_{\bm}(F^e_* {\mathcal{O}}_X( (p^e - 1)(\Delta' - D) - D )) \\ \xrightarrow{F^e_* (\cdot c)} & H^i_{\bm}(F^e_* {\mathcal{O}}_X( (p^e - 1)(\Delta' - D) - D )) \\ \xrightarrow{{\varphi}} & H^i_{\bm}({\mathcal{O}}_X(-D)) \end{array}$$ However, the map labeled $F^e_* (\cdot c)$ is the zero map for $i < \dim X$ since $$H^i_{\bm}(F^e_* {\mathcal{O}}_X( (p^e - 1)(\Delta' - D) - D )) = H^i_{\bm}(F^e_* {\mathcal{O}}_X(-D)).$$ Thus $H^i_{\bm}({\mathcal{O}}_X(-D)) = 0$ as desired. If one assumes that $(X, \Delta)$ is purely $F$-regular (an analog of purely log terminal [@TakagiPLTAdjoint]), the same result holds by the same proof. The point is that we may take $c$ annihilating $H^i_{\bm}(R)$ and which simultaneously doesn’t vanish along the support of any component of $\Delta'$. [([[*cf.*]{}]{}[@KollarMori Corollary 5.25])]{} If $(R, \bm)$ is local and $(X = \Spec R, \Delta)$ is strongly $F$-regular, then for every $\bQ$-Cartier integral divisor $D$, ${\mathcal{O}}_X(-D)$ is Cohen-Macaulay. If the index of $D$ is not divisible by $p$ then the statement is a special case of by setting $\Delta' :=0$. Hence assume that the index $m$ of $D$ is divisible by $p$. Choose then an effective divisor $E$ linearly equivalent to $D$ and set $r:= ms+1$, $\Delta':= \frac{1}{r} E$ for some integer $s \gg 0$. In this situation $r$ is relatively prime to $p$ and $$r D = (ms+1)D \sim D \sim E = r \left( \frac{1}{r} E \right) = r \Delta' .$$ Furthermore, for $s \gg 0$, $\left(X,\Delta+ \frac{1}{r} E \right)$ is strongly $F$-regular. Hence, we may apply for $\Delta$ replaced by $\Delta + \frac{1}{r} E$ and the above choices of $r$, $D$ and $\Delta'$. This concludes our proof. Before moving on to the sharply $F$-pure pairs, we need a Lemma on the existence of certain Frobenius splittings. \[lem.SplittingNotCompatibleWithM\] Suppose that $(R, \sC)$ is any pair where $\sC$ is a Cartier subalgebra on a local ring $(R, \bm)$. Suppose $\bm$ is not $\sC$-compatible. Then there exists some Frobenius splitting such that ${\varphi}(F^e_* \bm) = R \supsetneq \bm$. There obviously exists a map in $[\sC]_e$, $\psi : F^e_* R {\xrightarrow{\ \ }}R$, such that $\psi (F^e_* \bm) \not\subseteq \bm$. It follows that $\psi(F^e_* \bm) = R$. We have two cases: Suppose that there is a unit $d \in R \setminus \bm$ such that $\psi(F^e_* d) = u \notin \bm$. Thus $\psi(F^e_* (u^{-p^e} d)) = 1$. Consider the map ${\varphi}(F^e_* \blank) = \psi(F^e_*((u^{-p^e} d) \cdot \blank))$ and notice that ${\varphi}(F^e_* 1) = 1$ which shows that ${\varphi}$ is a splitting. Also notice that $\bm$ is not ${\varphi}$-compatible since ${\varphi}$ is a unit multiple of $\psi$. Thus we have found our ${\varphi}$. Since we have already handled Case 1, we may assume that $\psi(F^e_* d) \in \bm$ for all units $d \in R$. Choose $c \in \bm$ such that $\psi(F^e_* c) = 1$. Now then $\psi(F^e_* 1) \in \bm$ since $1$ is a unit. Thus $$\psi(F^e_* (c+1)) = \psi(F^e_* c) + \psi(F^e_* 1) \in 1 + \bm \not\subseteq \bm$$ is a unit. But this is a contradiction since $c$ is assumed not be a unit and so $c+1$ is a unit. \[rem.ThereExistsANiceDivisor\] If, in addition to the hypotheses of , $R$ is [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{}, then by applying the argument of to the splitting ${\varphi}$ constructed in the proof of , we obtain a $\bQ$-divisor $\Delta$ on $X = \Spec R$ such that - $(p^e - 1)(K_X + \Delta)$ is Cartier. - $(X, \Delta)$ is sharply $F$-pure. - $x = V(\bm)$ is not an $F$-pure center of $(X, \Delta)$. The second two statements follow since ${\varphi}\in [\sC^{\Delta}]_e$. [([[*cf.*]{}]{}[@KollarALocalKawamataViehweg Theorem 3(2)])]{} \[thm.KollarTheorem3(2)\] Suppose that $(R,\bm)$ is a local [$\textnormal{S}_2$]{}-ring, and $\sC$ is a Cartier-subalgebra on $R$ such that $(R, \sC)$ is sharply $F$-pure and $V(\bm)$ is not an $F$-pure center. Additionally suppose that $Z \subsetneq X = \Spec R$ is any union of $F$-pure centers of $(R, \sC)$. We also assume that $Z$ and $X$ have no common irreducible components. If $I_Z$ is the (radical) ideal defining $Z$, then $$\depth_{\bm} I_Z \geq \min\{3, 1+\codim_Z x\}.$$ Since shrinking $\sC$ is harmless, we set $\sC = R\langle {\varphi}\rangle$ for some splitting ${\varphi}: F^e_* R {\xrightarrow{\ \ }}R$ which by is not compatible with the origin $V(\bm)$. Indeed, by (iii) we can only increase the number of centers when restricting a Cartier subalgebra. We have the long exact sequence: $$\cdots H^1_{\bm}(I_Z) {\xrightarrow{\ \ }}H^1_{\bm}(R) {\xrightarrow{\ \ }}H^{1}_{\bm}(R/I_Z) {\xrightarrow{\ \ }}H^2_{\bm}(I_Z) {\xrightarrow{\ \ }}H^2_{\bm}(R) {\xrightarrow{\ \ }}H^2_{\bm}(R/I_Z) {\xrightarrow{\ \ }}H^{3}_{\bm}(I_Z) {\xrightarrow{\ \ }}\cdots$$ and recall we are trying to show that $H^i_{\bm}(I_Z) = 0$ for $i < \min\{3, 1+\codim_Z x\}$. Since $R$ is [$\textnormal{S}_2$]{}, $H^1_{\bm}(R) = H^0_{\bm}(R) = 0$. Since $Z \neq V(\bm)$ and $Z$ is reduced, $H^0_{\bm}({\mathcal{O}}_Z) = 0$ and so $H^1_{\bm}(I_Z) = 0$. Thus $\depth_{\bm} I_Z \geq 2$. It is now sufficient to prove the case when $\codim_Z x \geq 2$. This implies that $\dim X \geq 3$ (since $Z$ and $X$ have no common components). Furthermore, we can assume that every component of $Z$ has dimension at least 2. Indeed, suppose that $Z_1$ is an irreducible component of $Z$ such that $\dim Z_1 = 1$. If $Z_2$ is the union of the other components of $Z$ and $Z_2 \neq \emptyset$, then $Z_1 \cap Z_2 = x$ (for dimension reasons and since we working in a local ring). But this implies that $x$ is an $F$-pure center since intersections of $F$-pure centers are unions of $F$-pure centers by (i). Thus we can assume that $Z_1 = Z$ is $1$-dimensional. But then $\codim_Z x = 1$, which contradictions our assumption. By , we know that $\Ann_{R}(H^2_{\bm}(R))$ is compatible with $(R, \sC) = (R, R\langle {\varphi}\rangle)$. However, if $H^2_{\bm}(R) \neq 0$, then since $R$ is [$\textnormal{S}_2$]{} and of dimension $\geq 3$, $\sqrt{\Ann_{R}(H^2_{\bm}(R))} = \bm$ by . But $\Ann_{R}(H^2_{\bm}(R))$ is radical (since $\sC$ is sharply $F$-pure) so $\Ann_{R}(H^2_{\bm}(R)) = \bm$. But $V(\bm)$ is not an $F$-pure center, this is a contradiction. We conclude that $H^2_{\bm}(R) = 0$. Now we come to $H^1_{\bm}(R/I_Z)$. Again, since $R/I_Z$ is reduced, $R/I_Z$ is S1. Furthermore, since $Z$ has no 1-dimensional components we can apply to conclude that $\Ann_R H^1_{\bm}(R/I_Z)$ can either be $\bm$-primary or $R$. Suppose it is $\bm$-primary. Since ${\varphi}|_Z$ is still a splitting, it follows that $\Ann_{R/I_Z} H^1_{\bm}(R/I_Z)$ is ${\varphi}|_Z$-compatible and also radical and so equal to $\bm/I_Z$. But then $\bm$ is ${\varphi}$-compatible by basic facts about Frobenius splitting or by (vi). We conclude that $H^1_{\bm}(R/I_Z) = 0$. This forces $H^2_{\bm}(I_Z)$ to be zero and completes the proof. \[rem.KolThmInFSplitLanguage\] Another way to state a special case of using the language of Frobenius splittings is as follows: [*Suppose that $(R, \bm)$ is an [$\textnormal{S}_2$]{} local ring with Frobenius splitting ${\varphi}: F^e_* R {\xrightarrow{\ \ }}R$ which is not compatibly split with $\bm$. Additionally suppose that $Z$ is any union of compatibly split subvarieties of $X = \Spec R$ such that no irreducible component of $Z$ coincides with an irreducible component of $X$. Suppose that $I_Z \subseteq R$ is the ideal defining $Z$, then $$\depth_{\bm} I_Z \geq \min\{3, 1+\codim_Z x\}.$$*]{} Now we come to our main technical result on depth. It is a characteristic $p > 0$ version of [@KollarALocalKawamataViehweg Theorem 3(1)] but also compare with [@AlexeevLimitsOfStablePairs Lemma 3.2], [@FujinoIntroductionToTheLMMP Theorem 4.21], [@AlexeevHaconNonRationalCenters Theorem 1.5] and [@KovacsIrrationalCenters Theorem 1.2, 1.5]. [([[*cf.*]{}]{}[@KollarALocalKawamataViehweg Theorem 3(1)])]{} \[thm.KollarTheorem3(1)\] Suppose that $R$ is local, [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{}, $X = \Spec R$ and that $0 \leq \Delta \in \operatorname{{\bf{WDiv}}_{\bR}}(X)$ is such that $(R, \Delta)$ is sharply $F$-pure. Set $x \in X$ to be the closed point and assume that $x$ is not an $F$-pure center of $(R, \Delta)$. Suppose that $0 \leq \Delta' \leq \Delta$ is another element of $\operatorname{{\bf{WDiv}}_{\bR}}(X)$ and that $r \Delta'$ is integral for some $r > 0$ relatively prime to $p$. Further assume that $M$ is any rank-1 (along each component of $X$) reflexive coherent subsheaf of $K(X)$ such that $M^{(-r)} \cong {\mathcal{O}}_X( r\Delta')$ (here $\blank^{(\cdot)}$ denotes reflexive power).[^7] Then $$\depth_x M \geq \min\{3, \codim_X x\} = \min\{3, \dim R\}.$$ First observe that it is harmless to assume that $\dim R \geq 3$ since otherwise the statement is trivial since $M$ is reflexive and thus [$\textnormal{S}_2$]{} by [@HartshorneGeneralizedDivisorsOnGorensteinSchemes Theorem 1.9]. We may also assume that $M \subseteq {\mathcal{O}}_X$ is an ideal sheaf since we are working locally. We thus identify $M$ with an ideal of $R$ also denoted by $M$. Finally, by replacing $r$ by a power if necessary, we may assume that $r = p^e - 1$ for some $e > 0$. Using , we can find ${\varphi}$ a splitting, not compatible with $\bm$, making the following composition an isomorphism: $${\mathcal{O}}_X {\xrightarrow{\ \ }}F^e_* {\mathcal{O}}_X \hookrightarrow F^e_* {\mathcal{O}}_X( (p^e -1)\Delta')\hookrightarrow F^e_* {\mathcal{O}}_X( \lceil (p^e - 1)\Delta \rceil) \xrightarrow{{\varphi}} {\mathcal{O}}_X.$$ Twisting by $M$ and reflexifying (which we denote by $\blank^{**}$), we obtain $$\begin{array}{rcl} & & M\\ & {\xrightarrow{\ \ }}& F^e_* (M^{(p^e - 1)} \otimes M)^{**}\\ & \hookrightarrow & F^e_* ({\mathcal{O}}_X((p^e - 1)\Delta') \tensor M^{(p^e - 1)} \otimes M)^{**} \\ & \hookrightarrow & F^e_* ({\mathcal{O}}_X(\lceil(p^e - 1)\Delta\rceil) \tensor M^{(p^e - 1)} \otimes M)^{**}\\ & \xrightarrow{{\varphi}_M} & M. \end{array}$$ Using the fact that $({\mathcal{O}}_X((p^e - 1)\Delta') \tensor M^{(p^e - 1)})^{**} \cong {\mathcal{O}}_X$, we have a composition $$M {\xrightarrow{\ \ }}F^e_* M {\xrightarrow{\ \ }}M$$ that is an isomorphism (note the first map is not the usual inclusion of ideal sheaves via Frobenius). Certainly $H^1_{\bm}(M) = 0$ since $M$ is reflexive and thus [$\textnormal{S}_2$]{} by [@HartshorneGeneralizedDivisorsOnGorensteinSchemes Theorem 1.9]. We now study $H^2_{\bm}(M)$. Since $M$ is [$\textnormal{S}_2$]{}, it follows that either $\Ann_R H^2_{\bm}(M)$ is equal to $R$ or it is $\bm$-primary by . Since we have an injection $H^2_{\bm}(M) {\xrightarrow{\ \ }}H^2_{\bm}(F^e_* M)$, it follows that $\Ann_R H^2_{\bm}(M)$ is at the very least radical (since if $r^{p^e}$ kills $H^2_{\bm}(M)$, so does $r$). In particular, if $\Ann_R H^2_{\bm}(M) \neq R$, then it must be $\bm$. Fix $[z] \in H^2_{\bm}(M)$, and recall that we are considering $M$ as an ideal. If $z_j \in \Gamma(U, M)$[^8], then since $\Delta' \geq 0$ we have $$z_j^{p^e} \in \Gamma(U, M^{(p^e)}) \subseteq \Gamma\big(U, ({\mathcal{O}}_X((p^e - 1)\Delta') \tensor M^{(p^e - 1)} \otimes M)^{**}\big).$$ Thus we have a class $$[F^e_* z^{p^e}] \in H^2_{\bm}\big(F^e_* ({\mathcal{O}}_X((p^e - 1)\Delta') \tensor M^{(p^e - 1)} \otimes M)^{**}\big).$$ Now, when we apply ${\varphi}_M$ to this class, it is just applied component-wise. Thus ${\varphi}_M([F^e_* z^{p^e}]) = [z]$. For any $r \in \bm$, it follows that ${\varphi}_M((F^e_* r).[F^e_* z^{p^e}]) = {\varphi}(F^e_* r).[z]$. In particular, if an arbitrary $F^e_* r \in F^e_* \bm$ annihilates all classes $[y] \in H^2_{\bm}(F^e_* ({\mathcal{O}}_X((p^e - 1)\Delta') \tensor M^{(p^e - 1)} \otimes M)^{**}) \cong H^2_{\bm}(F^e_* M)$, then ${\varphi}(F^e_* r)$ also annihilates all such $[z] \in H^2_{\bm}(M)$. This proves that $\bm$ is ${\varphi}$-compatible, a contradiction. Applications {#sec:application} ============ Here we list the most important corollaries of the results of . The characteristic zero analog of many of them are already mentioned in [@KollarALocalKawamataViehweg]. We still state them here for the sake of completeness and we give a full proof of our main motivation, the compatibility of the relative canonical sheaf with base change. In Section \[sec:applications\_lemmas\] some lemmas are gathered while in Section \[sec:applications\_corollaries\] the promised corollaries are presented. Auxilliary results {#sec:applications_lemmas} ------------------ In this section, we prove a series of lemmas culminating with a base change statement for relative canonical sheaves for families with sharply $F$-pure fibers . ### Basic lemmas on depth and relative canonical sheaves We begin with a short section where we make note of some simple results on depth and relative canonical sheaves that we will use. \[fact:S\_d\_hyperplane\] [@BrunsHerzog Theorem 1.2.5] Let $\sF$ be a coherent sheaf on a Noetherian scheme $X$, $H$ a Cartier divisor on $X$ containing a point $P$, such that the local equation of $H$ at $P$ is not a zero divisor of $\sF_P$ (in other words, it is a regular element for $\sF_p$). Then 1. \[itm:S\_d\_hyperplane:first\] $\depth \sF_P \geq d \ \Leftrightarrow \ \depth (\sF|_H)_P \geq d-1$, 2. \[itm:S\_d\_hyperplane:second\] $\depth \sF_P \geq \min \{ d, \dim \sF_P \} \ \Leftrightarrow \ \depth (\sF|_H)_P \geq \min \{ d-1, \dim (\sF|_H)_P \}$. \[lem:S\_d\] If $f : X {\xrightarrow{\ \ }}Y$ is a morphism of Noetherian schemes, $\sF \neq 0$ is a coherent sheaf on $X$ flat over $Y$, such that $\sF|_{X_y}$ is [$\textnormal{S}_d$]{} for every $y \in Y$ (i.e., $\sF$ is relatively [$\textnormal{S}_d$]{} over $Y$) and $\sG \neq 0$ is an [$\textnormal{S}_d$]{} coherent sheaf on $Y$, then $\sF \otimes f^* \sG$ is [$\textnormal{S}_d$]{} as well. Fix an arbitrary $P \in X$ and set $Q:=f(P)$ and $F:=X_{Q}$. Then $$\begin{array}{rl} & \depth_{\sO_{X,P}} (\sF \otimes f^* \sG)_{P} \\ = & \underbrace{\depth_{\sO_{F,P}}\left( \sF|_F \right)_P \,\,+\,\, \depth_{\sO_{Y,Q}} \sG_{Q}}_{\textrm{$\sF$ is flat over $Y$ \& \cite[Proposition 6.3.1]{EGA_IV_II}}} \\ \geq & \underbrace{\min\big\{d, \dim_{\sO_{F,P}} \left(\sF|_F \right)_P\big\} + \depth_{\sO_{Y,Q}} \sG_{Q}}_{\textrm{$\sF|_F$ is {$\textnormal{S}_d$}}} \\ = & \min\big\{d + \depth_{\sO_{Y,Q}} \sG_Q, \dim_{\sO_{F,P}} \left(\sF|_F \right)_P + \depth_{\sO_{Y,Q}} \sG_Q\big\} \\ \geq & \underbrace{\min\big\{d + \depth_{\sO_{Y,Q}} \sG_Q, \dim_{\sO_{F,P}} \left(\sF|_F \right)_P + \min\{d, \dim_{\sO_{Y,Q}} \sG_Q\}\big\}}_{\textrm{$\sG$ is {$\textnormal{S}_d$}}} \\ = & \min\big\{d + \depth_{\sO_{Y,Q}} \sG_Q, \min\{\dim_{\sO_{F,P}} \left(\sF|_F \right)_P +d,\dim_{\sO_{F,P}} \left(\sF|_F \right)_P + \dim_{\sO_{Y,Q}} \sG_Q\}\big\} \\ = & \underbrace{\min\big\{d + \depth_{\sO_{Y,Q}} \sG_Q, \min\{\dim_{\sO_{F,P}} \left(\sF|_F \right)_P +d,\dim_{\sO_{X,P}} (\sF \otimes f^* \sG)_P \}\big\}}_{\textrm{$\sF$ is flat over $Y$ and \cite[Corollaire 6.1.2]{EGA_IV_II}}} \\ \geq & \min\big\{d + \depth_{\sO_{Y,Q}} \sG_Q, \min\{d,\dim_{\sO_{X,P}} (\sF \otimes f^* \sG)_P \}\big\} \\ \geq & \min\big\{d , \dim_{\sO_{X,P}} (\sF \otimes f^* \sG)_P \big\} . \end{array}$$ \[lem:Gorenstein\_base\] If $f : X {\xrightarrow{\ \ }}Y$ is a flat morphism of finite type to a Gorenstein scheme, then $\omega_{X/Y}^{\mydot}$ and $\omega_X^{\mydot}[-\dim Y]$ are locally isomorphic. In particular if $X$ is relatively Gorenstein over $Y$, then it is Gorenstein, and $\omega_{X/Y}$ and $\omega_X$ are locally isomorphic over $Y$. Locally on $Y$ the following isomorphisms hold. $$\omega_{X/Y}^{\mydot} = f^! \sO_Y \cong f^! \omega_Y \cong f^! \omega_Y^{\mydot}[- \dim Y] \cong \omega_{X}^{\mydot}[- \dim Y]$$ Before continuing, let us remind ourselves of how $F$-adjunction works and how it can allow us to restrict divisors. ### Restricting divisors by $F$-adjunction: the $F$-different Suppose that $\Delta \geq 0$ is a $\bQ$-divisor on an [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{} variety $X$ and that $(p^e - 1)(K_X + \Delta)$ is Cartier. In fact, everything we say even holds more generally if $\Delta \geq 0$ is a $\bZ_{(p)}$-Weil divisorial sheaf, which is intuitively something like a Weil divisor having components also in the singular locus[^9]. Further suppose that $D$ is a reduced Cartier divisor on $X$ that is itself [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{} and which has no common components with $\Delta$. We now explain how we can construct a canonical $\bZ_{(p)}$-Weil divisorial sheaf (not necessarily a $\bZ_{(p)}$-Weil divisor) which we call $\operatorname{Diff}_{F,D} \Delta$ on $D$. Here the subindex $F$ means that this is the $F$-singularity counterpart of the usual different known from minimal model program theory. However, contrary to the usual different, the construction of $\ Diff_{F,D} \Delta$ goes through without any further assumption requiring that $\Delta$ is $\bQ$-Cartier at certain points. Without loss of generality we can assume that $X = \Spec R$ and that $R$ is a local ring and that $D = V(f)$. The fact that the divisor is Cartier implies that is a free $F^e_* R$-module. Choose a generator of this module ${\varphi}$, which we can also view as an element of $\Hom_R(F^e_* R, R)$ since $\Delta$ is effective. Now define a new map $\psi : F^e_* R {\xrightarrow{\ \ }}R$ by the rule $\psi(F^e_* \blank) = {\varphi}(F^e_* (f^{p^e - 1} \cdot \blank))$. Certainly $\psi(F^e_* \langle f \rangle) = {\varphi}(F^e_* \langle f^{p^e} \rangle) \subseteq \langle f \rangle$ and thus $\psi$ induces a map $\overline{\psi} \in \Hom_{R/\langle f \rangle}(F^e_* R/\langle f \rangle, R/\langle f \rangle)$. Note that for every height one prime $\eta$ containing $f \in R$, [[*i.e.*]{}]{}, a minimal associated prime of $D$, we have that $\Delta_{\eta} = 0$ (since $\Delta$ and $D$ have no common components). Furthermore, $R_{\eta}$ is regular (since $R_{\eta}/\langle f \rangle$ is reduced and zero dimensional and hence regular). It follows from inspection that $\overline{\psi}$ is non-zero at every such $\eta$. By [@MillerSchwedeSLCvFP Theorem 2.4], it follows that $\overline{\psi}$ induces an effective $\bZ_{(p)}$-Weil divisorial sheaf on $D$. It is straightforward to verify that $\Delta|_D$ is independent of the choice of $e$ and ${\varphi}$ and so: We use $\operatorname{Diff}_{F,D} \Delta$ to denote the effective $\bZ_{(p)}$-Weil divisorial sheaf described above which coincides with $\overline{\psi}$. We also observe: \[obs.FAdjunction\] Notice now additionally that $\overline{\psi}$ (corresponding to $\operatorname{Diff}_{F,D} \Delta$) is surjective if and only if $\psi$ (corresponding to $\Delta + D$) is surjective. In other words, $(X, \Delta + D)$ is sharply $F$-pure near $D$ if and only if $(D, \operatorname{Diff}_{F,D} \Delta)$ is sharply $F$-pure. It will be useful for us to note that if $(D, \operatorname{Diff}_{F,D} \Delta)$ is sharply $F$-pure, then $\operatorname{Diff}_{F,D} \Delta$ is in fact an honest $\bZ_{(p)}$-Weil divisor and not just a $\bZ_{(p)}$-divisorial sheaf. Suppose not, then the Weil divisorial sheaf $(p^e-1)(\operatorname{Diff}_{F,D} \Delta)$ must properly contain ${\mathcal{O}}_D$, even at some generic point of the non-normal locus. A contradiction can then be obtained from the fact that the conductor is already compatible with every ${\varphi}\in \sC^R$ (this last fact follows from the argument of [@BrionKumarFrobeniusSplitting Proposition 1.2.5]). The above introduced $F$-different $\operatorname{Diff}_{F,D} \Delta$ is equal to $\Delta|_D$ in most cases when the latter is defined. \[lem:restriction\_equals\_different\] With the notation above, suppose additionally that $\Delta$ is $\bZ_{(p)}$-Cartier at all of the height-two primes primes of $R$ containing $f$ (the codimension 2 points of $X$ that are contained inside $D$). Then $\operatorname{Diff}_{F,D} \Delta$ coincides with the restriction $\Delta|_D$ of $\Delta$ to $D$. At each of those codimension 2 points $\bq \in \Spec R$, $R_{\bq}$ is already Gorenstein (since $R/\langle f \rangle$ is [$\textnormal{G}_1$]{} and so $R_{\bq}/\langle f \rangle$ is Gorenstein). It is enough to prove the result at each such $\bq$, so fix one such $\bq$. Further choose $e > 0$ as above and also sufficiently divisible such that $(p^e - 1)D$ is Cartier at $\bq$. We can thus write $(p^e - 1)\Delta = \Div_{\Spec R_{\bq}}(g)$ for some $g \in R_{\bq}$. Since $R_{\bq}$ is Gorenstein, we can choose $\Phi \in \Hom_{R_{\bq}}(F^e_* R_{\bq}, R_{\bq})$ generating the set as an $F^e_* R_{\bq}$-module. Consider the map $\Psi : F^e_* R_{\bq} {\xrightarrow{\ \ }}R_{\bq}$ defined by the rule $\Psi(F^e_* \blank) = \Phi(F^e_* (f^{p^e - 1} \cdot \blank))$. Certainly $\Psi$ restricts to a map $\overline{\Psi} \in \Hom_{R_{\bq}/\langle f \rangle}(F^e_* (R_{\bq}/\langle f \rangle), R_{\bq}/\langle f \rangle)$ as above. Furthermore, $\overline{\Psi}$ generates the $F^e_* (R_{\bq}/\langle f \rangle)$-module $\Hom_{R_{\bq}/\langle f \rangle}(F^e_* (R_{\bq}/\langle f \rangle), R_{\bq}/\langle f \rangle)$ by the diagrams in [@SchwedeFAdjunction Proof of Proposition 7.2]. It follows that $\psi = (F^e_* g)\cdot \Psi$ restricts to $\overline{{\varphi}} = (F^e_* g) \cdot \overline{\Psi}$ and hence corresponds to the naive restriction $\Delta|_W$. This proves the lemma. ### The relative canonical sheaf We apply the above ideas on $F$-different to the following. It is the inductional step in the proof of . \[lem:base\_change\] Let $ f : X {\xrightarrow{\ \ }}Y $ be a flat morphism of finite type with [$\textnormal{S}_2$]{}, [$\textnormal{G}_1$]{} equidimensional fibers to a smooth variety[^10] $Y$ and $\Delta \in \operatorname{{\bf{WDiv}}_{\bQ}}(X)$, such that $K_X + \Delta$ is $\bQ$-Cartier and $p \notdivide \operatorname{ind}(K_X + \Delta)$. Assume also that $Z \subseteq Y$ is a smooth Cartier divisor such that for $W:= X \times_Y Z$, $\Delta$ does not contain any component of $W$ and $(W,\operatorname{Diff}_{F, W} \Delta)$ is sharply $F$-pure.[^11] Then $\omega_{X/Y}|_{W} \cong \omega_{W/Z}$. By , both $X$ and $W$ are [$\textnormal{S}_2$]{}. Similarly, both are [$\textnormal{G}_1$]{} by . By $F$-adjunction , [[*cf.*]{}]{}[@SchwedeFAdjunction Main Theorem, Proposition 7.2, Remark 7.3], $(X, \Delta + W)$ is sharply $F$-pure in a neighborhood of $W$. Hence, so is $(X,\Delta)$. We now claim: No $F$-pure center of $(X, \Delta)$ is contained in $W$. Suppose that $Z \subseteq X$ was an $F$-pure center of $(X, \Delta)$ contained in $W$. Let $\eta$ denote the generic point of $Z$ and now we work in $R = {\mathcal{O}}_{X,\eta}$ with maximal ideal $\bm$ corresponding to $\eta$. For any element ${\varphi}: F^e_* {\mathcal{O}}_{X,\eta} \subseteq F^e_* {\mathcal{O}}_{X,\eta}(\lceil (p^e - 1)\Delta\rceil) {\xrightarrow{\ \ }}{\mathcal{O}}_{X,\eta}$ of $[\sC^{\Delta}]_e$, we notice that ${\varphi}(F^e_* \bm) \subseteq \bm$ since $Z$ is an $F$-pure center. Choose $f \in \bm$ to be the defining equation of the Cartier divisor $W$ in ${\mathcal{O}}_{X,\eta}$. It follows from construction that $(F^e_* f^{p^e-1}) \cdot [\sC^{\Delta}]_e = [\sC^{\Delta + W}]_e$. In other words, for any $\psi \in [\sC^{\Delta+W}]_e$, we can write $\psi(F^e_* \blank) = {\varphi}(F^e_* (f^{p^e-1} \cdot \blank))$ for some ${\varphi}\in [\sC^{\Delta}]_e$. With this notation, for any $r \in R$ we have $\psi(F^e_* r) = {\varphi}(F^e_* (f^{p^e-1}r)) \in {\varphi}(F^e_* \bm) \subseteq \bm$. This proves that $(X, \Delta+W)$ is not sharply $F$-pure at $\eta$, the generic point of $Z$. But we assumed that $(X, \ \Delta+W)$ was sharply $F$-pure, a contradiction which proves the claim. We return to the proof of . By then, for every $x \in W$, $$\depth_x \omega_{X} \geq \min\{3, \codim_X x\} = \min\{3, \dim \omega_{X,x}\}.$$ However, by , $\omega_X$ and $\omega_{X/Y}$ are isomorphic locally, and then in the above inequality $\omega_X$ can be replaced by $\omega_{X/Y}$. Then by , $\omega_{X/Y}|_W$ is [$\textnormal{S}_2$]{}. To be precise, to apply , one needs to prove a priori that the local equation of $W$ is not a zero-divisor of $\omega_{X/Y}$. For this it is enough to show that $\omega_{X/Y}$ is [$\textnormal{S}_1$]{}, which follows using again that locally $\omega_{X/Y}$ and $\omega_X$ are isomorphic and that $\omega_X$ is [$\textnormal{S}_2$]{} by [@KollarMori Corollary 5.69]. Therefore $\omega_{X/Y}|_W$ is [$\textnormal{S}_2$]{} indeed. However, so is $\omega_{W/Z}$ by using [@KollarMori Corollary 5.69] again. Furthermore, $\omega_{X/Y}|_W$ and $\omega_{W/Z}$ are isomorphic on the relative Gorenstein locus, since the relative canonical sheaf is compatible with base-change for Gorenstein morphisms [@ConradGDualityAndBaseChange Theorem 3.6.1]. Therefore [@HartshorneGeneralizedDivisorsOnGorensteinSchemes Theorems 1.9 and 1.12] yields the statement of the lemma. The next lemma is used in . It allows us to cite [@KollarHullsAndHusks]. \[lem:relative\_canonical\_reflexive\] Suppose that $Y$ is a scheme of finite type over an algebraically closed field. If $f : X {\xrightarrow{\ \ }}Y$ is a projective, flat, relatively [$\textnormal{S}_2$]{} and $G_1$, equidimensional morphism, then $\omega_{X/Y}$ is reflexive. According to [@HassettKovacsReflexivePullbacks Corollary 3.7], it is enough to exhibit an open set $U$ contained in the relative Gorenstein locus, such that 1. for $Z:= X \setminus U$, $\codim_{X_y} Z_y \geq 2$ for every $y \in Y$ and 2. for the inclusion of open set $j : U \hookrightarrow X$, the natural homomorphism $\omega_{X/Y} {\xrightarrow{\ \ }}j_* (\omega_{X/Y}|_U)$ is an isomorphism. Let $W$ be the non-relatively Gorenstein locus. Fix a finite surjective morphism $\pi : X {\xrightarrow{\ \ }}\bP^n_Y$ over $Y$, after possibly shrinking $Y$ ([[*cf.*]{}]{}[@KollarALocalKawamataViehweg proof of Corollary 24]). Set then $Z:= \pi^{-1} (\pi(W))$, $V:=\bP^n_Y \setminus \pi(W)$. Let $q : V {\xrightarrow{\ \ }}\bP^n_Y$ be the natural inclusion. With the above choices, $\codim_{X_y} Z_y \geq 2$ is satisfied for all $y \in Y$. For the other condition, notice that it is enough to prove that the natural homomorphism $\pi_* \omega_{X/Y} {\xrightarrow{\ \ }}\pi_* j_* (\omega_{X/Y}|_U)$ is an isomorphism. However $$\begin{gathered} \pi_* j_* (\omega_{X/Y}|_U) \cong q_* ((\pi_* \omega_{X/Y})|_V) \cong \underbrace{ q_* \sHom_{V}( (\pi_* \sO_X)|_V, \omega_{V/Y})}_{\textrm{Grothendieck duality}} \\ \cong \underbrace{ \sHom_{X}( \pi_* \sO_X, q_* \omega_{V/Y})}_{\textrm{adjoint functors}} \cong \underbrace{ \sHom_{X}( \pi_* \sO_X, \omega_{\bP^n_Y/Y})}_{\parbox{120pt}{\tiny \cite[Proposition 3.5]{HassettKovacsReflexivePullbacks} using that $\omega_{\bP^n_Y/Y}$ is flat and relatively $S_2$}} \cong \pi_* \omega_{X/Y},\end{gathered}$$ and the composition of the above isomorphisms is the natural homomorphism $\pi_* \omega_{X/Y} {\xrightarrow{\ \ }}\pi_* j_* (\omega_{X/Y}|_U)$. Consequences {#sec:applications_corollaries} ------------  \ We begin with a simple consequence on the depth of ${\mathcal{O}}_X$ and $\omega_X$. [([[*cf.*]{}]{}[@AlexeevLimitsOfStablePairs Lemma 3.2], [@FujinoIntroductionToTheLMMP Theorem 4.21], [@KollarALocalKawamataViehweg 4.1, 4.2, 4.3], [@AlexeevHaconNonRationalCenters Theorem 1.5], [@KovacsIrrationalCenters Theorem 1.2, 1.5])]{} Suppose that $X = \Spec R$ is [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{}. If $X$ is $F$-pure and $x \in X$ is not an $F$-pure center of $X$, then $$\depth_x {\mathcal{O}}_X \geq \{3, \codim_X x \} \text{ and } \depth_x \omega_X \geq \min\{3, \codim_X x\}.$$ We may assume that $X = \Spec R$ for a local ring $(R, \bm)$ with $x = V(\bm)$. Since $X$ is [$\textnormal{S}_2$]{} and [$\textnormal{G}_1$]{}, by using , we can assume that there exists some $\Delta \geq 0$ such that $(p^e - 1)(K_X + \Delta)$ is Cartier, such that $(X, \Delta)$ is sharply $F$-pure and such that $x$ is not an $F$-pure center of $(X, \Delta)$. Now the second statement follows from by setting $M = {\mathcal{O}}_X(K_X)$ and setting $\Delta' = \Delta$. The first statement also follows from by setting $M = {\mathcal{O}}_X$ and $\Delta' = 0$. Suppose that $(R, \bm)$ is $F$-injective. If $\bm$ is not an annihilator of any $F$-stable submodule of $H^i_{\bm}(R)$, does that imply any depth conditions on $R$ or $\omega_R$? To prove our main corollary, we need to introduce a generalization of $\operatorname{Diff}_{F, D} \Delta$ to the case when $D$ has higher codimension. We focus only on our case of interest, that is, when $D$ is the fiber over a smooth base. Let $f : X {\xrightarrow{\ \ }}Y $ be a flat morphism of finite type with [$\textnormal{S}_2$]{}, [$\textnormal{G}_1$]{} equidimensional fibers. Further suppose that $Y$ is a smooth variety, and $\Delta \in \operatorname{{\bf{WDiv}}_{\bQ}}(X)$ is such that $\Delta$ does not contain any component of any fiber. Fix a point $y \in Y$. Working locally, we may assume that $Y = \Spec A$ and $X = \Spec R$ for local rings $(A, \bn)$ and $(R, \bm)$. Further assume that $\bn = \langle f_1, \dots, f_n \rangle$ is regular system of generators with $Y_i = V(f_1, \dots, f_i)$ regular. Then, let $\Delta_i:= \operatorname{Diff}_{F,Y_i} \Delta_{i-1}$ and define then $\operatorname{Diff}_{F,X_y} \Delta:= \Delta_n$. The only question is whether this construction of $\operatorname{Diff}_{F,X_y} \Delta$ is independent of the choice of $f_i$. Following the method of $F$-adjunction, multiplying by each $f_i$ successively, we take a map corresponding to $\Delta$ and ${\varphi}: F^e_* R {\xrightarrow{\ \ }}R$ and obtain another map $\psi_f : F^e_* R {\xrightarrow{\ \ }}R$ defined by the rule $\psi_f(F^e_* \blank) = {\varphi}(F^e_* ( (f_1 \cdots f_n)^{p^e - 1} \cdot \blank))$. We then restrict this map to $X_y$ by modding out by $\bn$ and so obtain $\overline{\psi}_f$. Choosing different $Y_i$’s is simply choosing a different set of generators $\{g_1, \dots, g_n\}$ for $\bn$ which yields $\overline{\psi}_g$. To complete the proof of the claim, it is sufficient to show that these maps differ only by multiplication by a unit. We use $\bn^{[p^e]}$ to denote the ideal generated by the $p^e$th powers of the generators of $\bn$. Since $\bn^{[p^e]} : \bn = \langle (f_1 \cdots f_n)^{p^e - 1} \rangle + \bn^{[p^e]} = \langle (g_1 \cdots g_n)^{p^e - 1} \rangle + \bn^{[p^e]}$, see for example [@FedderFPureRat Proposition 2.1], it follows that $(f_1 \cdots f_n)^{p^e - 1} = u (g_1 \cdots g_n)^{p^e - 1} + \sum v_i h_i^{p^e}$ for some unit $u \in A$, elements $v_i \in A$ and $h_i \in \bn$. But now we see that $\overline{\psi}_f = (F^e_* u) \cdot \overline{\psi}_g$ since any multiple of $h_i^{p^e}$ will be sent into $\bn R$. This completes the proof. \[cor:relative\_canonical\_sheaf\_base\_change\] [([[*cf.*]{}]{}[@KollarALocalKawamataViehweg 4.3])]{} Let $f : X {\xrightarrow{\ \ }}Y $ be a flat morphism of finite type with [$\textnormal{S}_2$]{}, [$\textnormal{G}_1$]{}, equidimensional fibers to a smooth variety and let $\Delta \in \operatorname{{\bf{WDiv}}_{\bQ}}(X)$ be such that it does not contain any component of any fiber. Additionally assume that $K_X + \Delta$ is $\bQ$-Cartier, $p \notdivide \operatorname{ind}(K_X + \Delta)$ and $(X_y,\operatorname{Diff}_{F, X_y} \Delta)$ is sharply $F$-pure for every $y \in Y$. Then $\omega_{X/Y}$ is flat over $Y$ and compatible with arbitrary base change. We claim that $\omega_{X/Y}$ is flat over $Y$ and relatively [$\textnormal{S}_2$]{}. By [@BhattHoPatakfalviSchnell Lemma 2.13], flatness follows as soon as we prove that the restriction of $\omega_{X/Y}$ to every fiber is [$\textnormal{S}_1$]{}. On the other hand relatively [$\textnormal{S}_2$]{} means the stronger condition that the above restrictions are [$\textnormal{S}_2$]{}. Therefore to show the claim, it is enough to prove that $\omega_{X/Y}|_{X_y}$ is [$\textnormal{S}_2$]{} for every $y \in Y$. By [@KollarMori Corollary 5.69], $\omega_{X_y}$ is [$\textnormal{S}_2$]{} and hence, it is enough to show that $\omega_{X/Y} |_{X_y} \cong \omega_{X_y}$ locally around every point $x \in X_y$. We thus replace both $X$ and $Y$ by $\Spec {\mathcal{O}}_{X,x}$ and $\Spec {\mathcal{O}}_{Y,y}$, respectively. Therefore, we may assume that there is a sequence of smooth subvarieties: $Y=Y_0 \supseteq Y_1 \supseteq \dots \supseteq Y_{m-1} \supseteq Y_m = \{ y \} $, such that $Y_{i-1}$ is a Cartier divisor in $Y_i$. Set $X_i := X_{Y_i}$ and $\Delta_i:= \operatorname{Diff}_{F, X_i} \Delta_{i-1}$ with $\Delta_0 = \Delta$. Note that then $\ Delta_m = \operatorname{Diff}_{F,X_y} \Delta$. By applying (backwards) inductively and possibly further restricting $X$ around $x$, one obtains that $(X_i, \Delta_i + X_{i-1})$ and hence $(X_i, \Delta_i )$ is sharply $F$-pure for all $i$ (in fact, this also implies all the $\bZ_{(p)}$-Weil divisorial sheaves $\Delta_i$ are honest divisors). Finally, applying inductively again yields that $\omega_{X_i/Y_i}|_{X_{i-1}} \cong \omega_{X_{i-1}/Y_{i-1}}$ for all $i$, and consequently $\omega_{X/Y} |_{X_y} \cong \omega_{X_y}$. This finishes the proof of our claim. By our claim and [@HassettKovacsReflexivePullbacks Corollary 3.8] $\omega_{X/Y}$ and all its pullbacks are reflexive. Hence, by restricting to the relatively Cohen-Macaulay locus (whose complement has codimension $\geq 2$) and using [@HassettKovacsReflexivePullbacks Proposition 3.6], for any morphism $Z {\xrightarrow{\ \ }}Y$, $\omega_{X_Z/Z} \cong (\omega_{X/Y})_Z$. By , the appearance of $\operatorname{Diff}_{F,X_y} \Delta$ can be replaced by an actual “geometric” restriction, if we assume the following: $$\label{eq:index_at_relative_codimension_one} \parbox{350pt}{ for each $y \in Y$, there is some $r > 0$ relatively prime to $p$ such that $r\Delta$ is Cartier at the codimension 1 points of the fiber $X_y \subseteq X$.}$$ In particular, this is satisfied if $\Supp \Delta$ does not contain the singular codimension one points of the fibers. Indeed, let $\xi$ be a codimesnion 1 point of a fiber $X_y$. If $\xi$ is in the singular locus of $X_y$, then $\xi \not\in \Supp \Delta$ and hence $\Delta$ is Cartier at $\xi$. Otherwise, $X$ is smooth around $\xi$, and hence $K_X$ is Cartier at $\xi$. In particular then by $p \notdivide \operatorname{ind}(K_X + \Delta)$, we obtain that $\Delta$ is $\bZ_{(p)}$ Cartier at $\xi$. In either cases $\Delta$ satisfies , and therefore in the special case of stated in , the use of ordinary restriction of $\Delta$ was legitimate. When $f$ is projective, the compatibility of follows for arbitrary reduced base by an important result of Kollár [@KollarHullsAndHusks]. \[cor:relative\_canonical\_sheaf\_base\_change\_projective\] Let $f : X {\xrightarrow{\ \ }}Y $ be a flat projective morphism with [$\textnormal{S}_2$]{}, [$\textnormal{G}_1$]{} equidimensional fibers. Further suppose that $Y$ is a reduced, separated scheme of finite type over an algebraically closed field, and $\Delta$ a $\bQ$-Weil-divisor that avoids all the codimension zero and the singular codimension one points of the fibers. Additionally assume that there is a $p \notdivide N >0$, such that $N \Delta$ is Cartier in relative codimension one and $\omega_{X/Y}^{[N]}(N \Delta)$ [^12] is a line bundle and that $(X_y, \Delta_y)$ is sharply $F$-pure for every $y \in Y$. Then $\omega_{X/Y}$ is flat and compatible with arbitrary base change. First, we need some preparation about pulling back $\Delta$. Suppose $\tau : Y' {\xrightarrow{\ \ }}Y$ is a morphism and set $X' := X \times_Y Y'$, $\pi : X' {\xrightarrow{\ \ }}X$, and $f' : X' {\xrightarrow{\ \ }}Y'$ the induced morphisms. Then a natural pullback $\Delta'$ of $\Delta$ can be defined as follows. Let $U \subseteq X$ be the open set where $f$ is Gorenstein and $\Delta$ is $\bQ$-Cartier. Then, pull $\Delta|_U$ back to $\pi^{-1}U$, and finally extend it uniquely over $X'$. This extension is unique, since $\codim_{X'} X' \setminus \pi^{-1}U \geq 2$. We claim that $$\label{eq:relative_canonical_sheaf_base_change_projective:pullback} \pi^* \omega_{X/Y}^{[N]}(N \Delta) \cong \omega_{X'/Y'}^{[N]}(N \Delta').$$ Indeed, notice that by construction $\pi^* \omega_{X/Y}^{[N]}(N \Delta)$ and $\omega_{X'/Y'}^{[N]}(N \Delta')$ agree over $\pi^{-1}U$, that is, in relative codimension one. Notice also that since $\omega_{X/Y}^{[N]}(N \Delta)$ is assumed to be a line bundle, so is $\pi^* \omega_{X/Y}^{[N]}(N \Delta)$, and therefore $\pi^* \omega_{X/Y}^{[N]}(N \Delta)$ is reflexive. On the other hand, since $\omega_{X'/Y'}^{[N]}(N \Delta')$ is defined as a pushforward of a line bundle from relative codimension one, it is reflexive by [@HassettKovacsReflexivePullbacks Corollary 3.7]. Therefore by [@HassettKovacsReflexivePullbacks Proposition 3.6], holds. In particular, $\omega_{X'/Y'}^{[N]}(N \Delta')$ is a line bundle. The main consequence of the previous paragraph, is that the conditions of the corollary are invariant under pullback to another reduced, separated scheme $Y'$ of finite type over $k$. That is, $X'$, $f'$ and $\Delta'$ defined above satisfy all the assumptions of the corollary. Let us turn now to the actual proof of the corollary. First, we may assume that $Y$ is connected. Second, according to [@KollarHullsAndHusks Corollary 24] and , there is a locally closed decomposition $\amalg Y_i {\xrightarrow{\ \ }}Y$, such that if $T {\xrightarrow{\ \ }}Y$, then $\omega_{X_T/T}$ is flat and commutes with base-change if and only if $T {\xrightarrow{\ \ }}Y$ factors through some $Y_i {\xrightarrow{\ \ }}Y$. Now, for every irreducible component $Y'$ of $Y$, there is a regular alteration $S {\xrightarrow{\ \ }}Y'$ [@deJongAlterations Thorem 4.1]. By the above discussion $X_S {\xrightarrow{\ \ }}S$ satisfies the assumptions of the corollary, and hence also of . Therefore, $\omega_{X_S/S}$ is flat and compatible with arbitrary base-change. Hence $S {\xrightarrow{\ \ }}Y$ factors through one of the $Y_i$. In particular, since the image of $S {\xrightarrow{\ \ }}Y$ is the component $Y'$, $Y' \subseteq Y_i$. That is, every irreducible component of $Y$ is contained in one $Y_i$. However, $Y$ is connected, therefore all irreducible components of $Y$ are contained in the same $Y_i$, and hence by the reducedness of $Y$, $Y_i=Y$. In the case of $\dim Y=1$, if instead of assuming that $X_y$ is sharply $F$-pure, one assumes that $(X,X_y)$ is $F$-pure for all $y \in Y$, the $p \notdivide \operatorname{ind}(K_X + \Delta)$ assumption can be dropped from the above corollaries using the trick of . \[qtn:aribtrary\_base\] Does the compatibility of the relative canonical sheaf with base change stated in hold for singular $Y$ (with the adequate modification in the setup as in )? From the modular point of view, especially interesting would be the case of non-reduced $Y$. This case is open even in the projective case. It should be noted that the characteristic zero analogue of is known if $f$ is projective and $Y$ is arbitrary [@KollarKovacsLCImpliesDB Theorem 7.9]. That is, the answer to the characteristic zero analogue of is positive when $f$ is projective. \[qtn:log\_canonical\] Can one replace sharply $F$-pure by log-canonical (still assuming positive characteristic) in the statement of ? This would also be important from the modular point of view, since sharply $F$-pure varieties can be deformed to log-canonical but not sharply $F$-pure varieties. \[qtn:divisibility\] Can one remove the divisibility by $p$ condition from the statement of ? [@KollarALocalKawamataViehweg 4.10] The sheaf ${\mathcal{O}}_X(-D)$ in and cannot be replaced by ${\mathcal{O}}_X(D)$ as shown in [@KollarALocalKawamataViehweg]. We refer to [@KollarALocalKawamataViehweg] for the actual example. \#1[7 71000017 10000 -17 100007]{} \[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{} [BHPS12]{} : *The structure of [$F$]{}-pure rings*, Math. Z. **250** (2005), no. 4, 791–806. [MR2180375]{} : *Limits of stable pairs*, Pure Appl. Math. Q. **4** (2008), no. 3, Special Issue: In honor of Fedor Bogomolov. Part 2, 767–783. [2435844 (2009j:14020)]{} : *Non-rational centers of log canonical singularities*, arXiv:1109.4164. : *Moduli of products of stable varieties*, arXiv:1206.0438. : *Test ideals via algebras of $p^{-e}$-liner maps*, arXiv:0912.2255, to appear in J. Algebraic Geom. : *[$F$]{}-signature of pairs and the asymptotic behavior of [F]{}robenius splittings*, arXiv:1107.1082. : *Frobenius splitting methods in geometry and representation theory*, Progress in Mathematics, vol. 231, Birkhäuser Boston Inc., Boston, MA, 2005. [MR2107324 (2005k:14104)]{} : *Cohen-[M]{}acaulay rings*, Cambridge Studies in Advanced Mathematics, vol. 39, Cambridge University Press, Cambridge, 1993. [MR1251956 (95h:13020)]{} : *Grothendieck duality and base change*, Lecture Notes in Mathematics, vol. 1750, Springer-Verlag, Berlin, 2000. [ MR1804902 (2002d:14025)]{} : *Smoothness, semi-stability and alterations*, Inst. Hautes Études Sci. Publ. Math. (1996), no. 83, 51–93. [ 1423020 (98e:14011)]{} : *The [F]{}robenius structure of local cohomology*, Algebra Number Theory **2** (2008), no. 7, 721–754. [MR2460693 (2009i:13009)]{} : *[$F$]{}-purity and rational singularity*, Trans. Amer. Math. Soc. **278** (1983), no. 2, 461–480. [MR701505 (84h:13031)]{} : *Introduction to the log minimal model program for log canonical pairs*, arXiv:0907.1506. : *Notes on some [$t$]{}-structures*, Geometric aspects of Dwork theory. Vol. I, II, Walter de Gruyter GmbH & Co. KG, Berlin, 2004, pp. 711–734. : *Éléments de géométrie algébrique. [IV]{}. Étude locale des schémas et des morphismes de schémas. [II]{}*, Inst. Hautes Études Sci. Publ. Math. (1965), no. 24, 231. [ 0199181 (33 \#7330)]{} : *Classification of higher dimensional algebraic varieties*, Oberwolfach Seminars, vol. 41, Birkhäuser Verlag, Basel, 2010. [2675555 (2011f:14025)]{} : *F-regular and [F]{}-pure rings vs. log terminal and log canonical singularities*, J. Algebraic Geom. **11** (2002), no. 2, 363–392. [MR1874118 (2002k:13009)]{} : *Residues and duality*, Lecture notes of a seminar on the work of A. Grothendieck, given at Harvard 1963/64. With an appendix by P. Deligne. Lecture Notes in Mathematics, No. 20, Springer-Verlag, Berlin, 1966. [MR0222093 (36 \#5145)]{} : *Generalized divisors on [G]{}orenstein schemes*, Proceedings of Conference on Algebraic Geometry and Ring Theory in honor of Michael Artin, Part III (Antwerp, 1992), vol. 8, 1994, pp. 287–339. [MR1291023 (95k:14008)]{} : *Reflexive pull-backs and base extension*, J. Algebraic Geom. **13** (2004), no. 2, 233–247. [MR2047697 (2005b:14028)]{} : *The purity of the [F]{}robenius and local cohomology*, Advances in Math. **21** (1976), no. 2, 117–172. [MR0417172 (54 \#5230)]{} : *Subadditivity of the [K]{}odaira dimension: fibers of general type*, Algebraic geometry, [S]{}endai, 1985, Adv. Stud. Pure Math., vol. 10, North-Holland, Amsterdam, 1987, pp. 361–398. [ MR946244 (89i:14029)]{} : *Projectivity of complete moduli*, J. Differential Geom. **32** (1990), no. 1, 235–268. [1064874 (92e:14008)]{} : *Hulls and husks*, arXiv:0805.0576. : *A local version of the [K]{}awamata-[V]{}iehweg vanishing theorem*, Pure Appl. Math. Q. **7** (2011), no. 4, Special Issue: In memory of Eckart Viehweg, 1477–1494. [2918170]{} : *Flips and abundance for algebraic threefolds*, Société Mathématique de France, Paris, 1992, Papers from the Second Summer Seminar on Algebraic Geometry held at the University of Utah, Salt Lake City, Utah, August 1991, Astérisque No. 211 (1992). [MR1225842 (94f:14013)]{} : *Log canonical singularities are [D]{}u [B]{}ois*, J. Amer. Math. Soc. **23** (2010), no. 3, 791–813. [2629988]{} : *Birational geometry of algebraic varieties*, Cambridge Tracts in Mathematics, vol. 134, Cambridge University Press, Cambridge, 1998, With the collaboration of C. H. Clemens and A. Corti, Translated from the 1998 Japanese original. [MR1658959 (2000b:14018)]{} : *Irrational centers*, Pure Appl. Math. Q. **7** (2011), no. 4, Special Issue: In memory of Eckart Viehweg, 1495–1515. [2918171]{} : *On [N]{}oetherian rings of characteristic [$p$]{}*, Amer. J. Math. **98** (1976), no. 4, 999–1013. [MR0432625 (55 \#5612)]{} : *Semi-log canonical vs [$F$]{}-pure singularities*, J. Algebra **349** (2012), 150–164. [ 2853631]{} : *Base change behavior of the relative canonical sheaf related to higher dimensional moduli*, arXiv:1005.5207, to appear in [A]{}lgebra [N]{}umber [T]{}heory. : *Contre-exemple au “vanishing theorem” en caractéristique [$p>0$]{}*, C. [P]{}. [R]{}amanujam—a tribute, Tata Inst. Fund. Res. Studies in Math., vol. 8, Springer, Berlin, 1978, pp. 273–278. [541027 (81b:14011)]{} : *[$F$]{}-adjunction*, Algebra Number Theory **3** (2009), no. 8, 907–950. [2587408 (2011b:14006)]{} : *Centers of [$F$]{}-purity*, Math. Z. **265** (2010), no. 3, 687–714. [2644316 (2011e:13011)]{} : *Test ideals in non-[$\mathbb{{Q}}$]{}-[G]{}orenstein rings*, Trans. Amer. Math. Soc. **363** (2011), no. 11, 5925–5941. : *Globally [$F$]{}-regular and log [F]{}ano varieties*, Adv. Math. **224** (2010), no. 3, 863–894. [ 2628797 (2011e:14076)]{} : *A characteristic [$p$]{} analogue of plt singularities and adjoint ideals*, Math. Z. **259** (2008), no. 2, 321–341. [MR2390084 (2009b:13004)]{} [^1]: The second author was partially supported by the NSF grant DMS \#1064485 [^2]: Meaning the Frobenius morphism is a finite morphism. [^3]: If ${\varphi}\in [\sC^R]_e$ and $\psi \in [\sC^R]_d$, then ${\varphi}\cdot \psi = {\varphi}\circ (F^e_* \psi)$. See the aforementioned sources for more details. [^4]: A *splitting* is simply a map ${\varphi}: F^e_* R {\xrightarrow{\ \ }}R$ that sends $F^e_* 1$ to $1$. Splittings are necessarily surjective. [^5]: in the terminology of [@KollarFlipsAndAbundance Section 16] [^6]: In other words, $\depth_z M \geq \min(n, \dim M_z)$ for all $z \in \Spec R$. Note that here we use $\dim M_z$ not $\dim R_z$. [^7]: Note that it is also common to use the notation $\blank^{[\cdot]}$. We do not use that notation since it might be confused with Frobenius power. [^8]: Here $U= X \setminus \{x\}$, and local cohomology classes are treated as classes on $U$ as explained earlier. [^9]: See [@KollarFlipsAndAbundance] and [@MillerSchwedeSLCvFP] for definitions, in the latter source these are called $\bZ_{(p)}$-$AC$-divisors. [^10]: Variety here means a separated, integral scheme of finite type over an algebraically closed base-field. [^11]: The fact that $(W, \operatorname{Diff}_{F, W} \Delta)$ is sharply $F$-pure implies that $\operatorname{Diff}_{F, W} \Delta$ is a $\bZ_{(p)}$-divisor and not simply a $\bZ_{(p)}$-Weil divisorial sheaf. [^12]: Let $U$ be the intersection of the relative Gorenstein locus and the locus where $N \Delta$ is Cartier. Set $\iota : U {\xrightarrow{\ \ }}X$ for the natural inclusion. The sheaf $\omega_{X/Y}^{[N]}(N \Delta)$ is the reflexive hull of $\omega_{U/Y}^N(N \Delta|_U)$, i.e. the unique reflexive sheaf that restricts on $U$ to the above sheaf. It can be obtained as $\iota_* (\omega_{U/Y}^N(N \Delta|_U))$. Indeed, $\iota_* (\omega_{U/Y}^N(N \Delta|_U))$ is reflexive by [@HassettKovacsReflexivePullbacks Corollary 3.7] and it is unique by [@HassettKovacsReflexivePullbacks Proposition 3.6].

ArXiv

--- abstract: 'It is offered that $F(R)-$modified gravities can be considered as nonperturbative quantum effects arising from Einstein gravity. It is assumed that nonperturbative quantum effects gives rise to the fact that the connection becomes incompatible with the metric, the metric factors and the square of the connection in Einstein - Hilbert Lagrangian have nonperturbative additions. In the simplest approximation both additions can be considered as functions of one scalar field. The scalar field can be excluded from the Lagrangian obtaining $F(R)-$gravity. The essence of quantum correction to the affine connection as a torsion is discussed.' author: - Vladimir Dzhunushaliev title: ' Dynamical $F(R)$ gravities ' --- Introduction ============ Current theoretical cosmology has two fundamental problems: what are inflation and dark energy ? One possible resolution of these problems is the gravitational alternative (modified gravities or in other words: $F(R)-$gravities) for a unified description of the inflation and dark energy. From the author point of view at such resolution of these two problems erases another problem: how the nature chooses from even uncountable sets of all $F(R)-$theories one of which is realized in the nature ? Non-Einstein’s gravities have a long history. They arise: by the perturbative quantization of 4D Einstein gravity [@Buchbinder] and by the consideration of quantum fields on the background of curved spacetime (for the textbook, see [@birrell]). For modern review on $F(R)-$gravities one can see [@Nojiri:2010wj]-[@Capozziello]. In this paper we consider a model where $F(R)-$modified gravities appear as quantum effects in Einstein gravity. It will be what we call as dynamical $F(R)-$gravities. As we see below nonperturbative quantization is either an operator differential equations or an infinite equations set -. If we could resolve either or - then we would have the necessary information on the quantum behavior of gravitational field. But in the consequence of extremely complicated problem for solving either or - we can not do it. Therefore we have to use an approximate approach. The main idea for such approximation is following: (a) in quantum gravity we will consider a connection $\Gamma$ and metric $g$ as independent variables; (b) by a nonperturbative quantization we assume that approximately $\left\langle Q \left| \hat \Gamma^2 \right| Q \right\rangle \approx \{ \}^2 + B(\phi)$ [^1] and the metric factor $\left\langle Q \left| \widehat{\sqrt{-g}} \cdot \hat g^{\mu \nu} \right| Q \right\rangle$ in Einstein - Hilbert Lagrangian also has a quantum correction; (c) in the simplest approximation we assume that both quantum corrections are functions of one scalar function $\phi$; (d) considering the scalar field $\phi$ as an independent variable one can exclude the scalar field $\phi$ from the Einstein - Hilbert Lagrangian obtaining $F(R)-$modified theory. It means that $F(R)-$gravity appears after excluding nonperturbative quantum terms from averaged Einstein-Hilbert Lagrangian. In such approach there is a principal difference between standard and dynamical $F(R)-$gravities: For the first approach we should have either different functions $F(R)$ for an inflation era and a modern universe acceleration or one but very complicated function for the description both regimes. In the second approach these different regimes are different quantum effects in one theory (quantum Einstein gravity). Nonperturbative quantization technique for gravity ================================================== In 50th Heisenberg has offered the method for the nonperturbative quantization of a nonlinear spinor field [@heisenberg]. Following to Heisenberg, nonperturbative operators of a quantum field can be calculated using corresponding field equation(s) for this theory. According Heisenberg [@heisenberg] quantum operators of the metric $\hat g_{\mu \nu}$ obey to operator Einstein equations in the Palatini formalism $$\begin{aligned} \hat \Gamma^\rho_{\phantom{\rho} \mu \nu} - \frac{1}{2} \hat g^{\rho \sigma} \left( \frac{\partial \hat g_{\mu \sigma}}{\partial x^\nu} + \frac{\partial \hat g_{\nu \sigma}}{\partial x^\mu} - \frac{\partial \hat g_{\mu \nu}}{\partial x^\sigma} \right) &=& 0, \label{0-10}\\ \hat R_{\mu \nu} - \frac{1}{2} \hat g_{\mu \nu} \hat R &=& 0 \label{0-20}\end{aligned}$$ here for the simplicity we consider Einstein gravity without the matter; $\hat \Gamma^\rho_{\phantom{\rho} \mu \nu}$ is the operator of the connection; $\hat g_{\mu \nu}$ is the operator of the metric; $\hat R_{\mu \nu}$ is the operator of the Ricci tensor and defined in usual manner taking into that it is an operator $$\begin{aligned} \hat R_{\mu \nu} &=& \hat R^\rho_{\phantom{\rho} \mu \rho \nu}, \label{0-30}\\ \hat R^\rho_{\phantom{\rho} \sigma \mu \nu} &=& \frac{\partial \hat \Gamma^\rho_{\phantom{\rho} \sigma \nu}} {\partial x^\mu} - \frac{\partial \hat \Gamma^\rho_{\phantom{\rho} \sigma \mu}} {\partial x^\nu} + \hat \Gamma^\rho_{\phantom{\rho} \tau \mu} \hat \Gamma^\tau_{\phantom{\tau} \sigma \nu} - \hat \Gamma^\rho_{\phantom{\rho} \tau \nu} \hat \Gamma^\tau_{\phantom{\tau} \sigma \mu} . \label{0-40}\end{aligned}$$ Now we have obtained an extremely complicated problem for solving operator Einstein equations . Heisenberg’s approach for solving this problem is to write an infinite equations set for all Green functions $$\begin{aligned} \left\langle Q \left| \hat \Gamma(x_1) \cdot \text{ Eq. \eqref{0-10} } \right| Q \right\rangle &=& 0 , \label{0-50}\\ \left\langle Q \left| \hat g(x_1) \cdot \text{ Eq. \eqref{0-10} } \right| Q \right\rangle &=& 0 , \label{0-60}\\ \left\langle Q \left| \hat \Gamma(x_1) \cdot \text{ Eq. \eqref{0-20} } \right| Q \right\rangle &=& 0 , \label{0-70}\\ \left\langle Q \left| \hat g(x_1) \cdot \text{ Eq. \eqref{0-20} } \right| Q \right\rangle &=& 0 , \label{0-80}\\ \left\langle Q \left| \hat \Gamma(x_1) \hat \Gamma(x_2) \cdot \text{ Eq. \eqref{0-10} } \right| Q \right\rangle &=& 0 , \label{0-90}\\ \left\langle Q \left| \hat g(x_1) \hat \Gamma(x_2) \cdot \text{ Eq. \eqref{0-10} } \right| Q \right\rangle &=& 0 , \label{0-100}\\ \left\langle Q \left| \hat g(x_1) \hat g(x_2) \cdot \text{ Eq. \eqref{0-10} } \right| Q \right\rangle &=& 0 , \label{0-110}\\ \left\langle Q \left| \hat \Gamma(x_1) \hat \Gamma(x_2) \cdot \text{ Eq. \eqref{0-20} } \right| Q \right\rangle &=& 0 , \label{0-120}\\ \left\langle Q \left| \hat g(x_1) \hat \Gamma(x_2) \cdot \text{ Eq. \eqref{0-20} } \right| Q \right\rangle &=& 0 , \label{0-130}\\ \left\langle Q \left| \hat g(x_1) \hat g(x_2) \cdot \text{ Eq. \eqref{0-20} } \right| Q \right\rangle &=& 0 , \label{0-140}\\ \cdots &=& 0 , \label{0-150}\\ \left\langle Q \left| \text{ the product of $g$ and $\Gamma$ at different points $(x_1, \cdots , x_n)$} \cdot \text{ Eq. \eqref{0-10} or \eqref{0-20}} \right| Q \right\rangle &=& 0 \label{0-160}\end{aligned}$$ where $\left. \left|Q \right. \right\rangle$ is a quantum state, for example, it can be the quantum state for the inflation, the modern universe acceleration and so on. Below for the simplicity we will write $\left\langle \cdots \right\rangle$ instead of $\left\langle Q \left| \cdots \right| Q \right\rangle$. Schematically the first equation has $\left\langle \Gamma^2 \right\rangle$ and $\left\langle \Gamma \cdot g \right\rangle$ terms; the second equation has $g \cdot \Gamma$ and $g \cdot g^{\mu \nu} \cdot \partial g$; the third equation has $\left\langle \partial\Gamma \cdot \Gamma \right\rangle$ and $\left\langle \Gamma^3 \right\rangle$ terms and so on up to infinity. Thus all equations are linked and this is the main problem to solve such infinite equations set. Similar equations set one can find in statistical physics and turbulence theory. One can also remember the Heisenberg’s matrix mechanics which is working with an infinite equations set for matrix elements. On the perturbative language Eq’s - are Dyson - Schwinger equations but usually (for instance, in QED) they are written on the Feynman diagrams language. In Eq’s - there are products like $\left. \hat \Gamma^n \right|_{x_i=x}$, $\left. \hat g^n \right|_{x_i=x}$ or $\left. \hat \Gamma^n \right|_{x_i=x} \cdot \left. \hat g^m \right|_{x_i=x}$. By using the perturbative approach (Feynman diagram technique) these products leads to singularities because the product of field operators (for free fields) in one point is poorly defined. We have to point out on the difference between such products by perturbative and nonperturbative quantization: for the nonperturbative quantization such singularities may be much softer or be absent in general [@heisenberg]. The matter is that in the perturbative approach we have moving particles (quanta) only. The communication between two spacetime points is carried by such quanta. Consequently the correlation function between these points (Green function) is not zero only if the interaction between points with quanta is possible. It means that corresponding Green function is not zero inside of a light cone and is zero outside of them. Such kind of functions have to be singular on the light cone (it is well known). But for the nonperturbative case it is not the case: for self-interacting nonlinear fields may exist static configurations. For instance, it can be a nucleon, a glueball, a flux tube stretched between quark and antiquark and filled with a longitudinal chromoelectric field and so on. In this case the correlation function (Green function) is not zero outside of the light cone. This is the reason why the Green function for the nonperturbative case is not so singular in the comparison with the perturbative case. Mathematically it can be explained in such a way: for free and interacting operators of quantum fields we have two algebras. For free fields the algebra is well known and it is described by canonical commutations relationships where commutators and anticommutators are distributions. But for interacting fields the algebra is unknown and above we bring forward arguments that defining relationships will be ordinary functions (not distributions). It is high probable that the equations set neither nor - can not be solved analytically. One possible way to solve approximately these equations is following. We decompose $n-$th Green function $$G_n = G(x_1, x_2 \cdots , x_n) = \left\langle \Gamma(x_1) \cdots g(x_m) \cdots \right\rangle \label{0-170}$$ on the linear combination of the Green functions products of lower orders $$\begin{split} G_n(x_1, x_2 \cdots , x_n) \approx & G_{n-2}(x_3, x_4 \cdots , x_n) \left[ G_2(x_1, x_2) -C_2 \right] + \\ & \left( \text{permutations of $x_1,x_2$ with $x_3, \cdots , x_n$} \right) + \\ & G_{n-3} \left( G_3 -C_3 \right) + \cdots \end{split} \label{0-180}$$ where $C_{2,3 \cdots}$ are constants. In such a way one can cut off the infinite equation set - . Such technique is applied for solving similar equations set in statistical physics and turbulence theory. Another way to solve approximately the infinite equation set - is choose some functional (for instance, the action or something like to gluon condensates in quantum chromodynamics [^2]) and write down its average expression using corresponding Green functions. After that it is necessary to use some well-reasoned physical assumptions to express the highest order Green function through Green function of of lower order, see the decomposition . Finally we will have a functional which can be used to obtain Euler - Lagrange field equations for Green functions. Below we will use such method for obtaining Einstein - Hilbert + scalar field action. Let us note that equations - are similar to: (a) a Bogolyubov chain of equations (hierarchy) for the one-particle, two-particle, etc., distribution functions of a classical statistical system; (b) Dyson - Schwinger equations for n-point Green functions in quantum field theory; (c) equations connecting correlation functions for velocities, pressure and density in a turbulent fluid. In all cases the solution methods are similar to discussed above: the decomposition of n-point corresponding functions on the product of lower function order. For us more interesting is the case (b): Dyson - Schwinger equations. In a perturbative quantum field theory these equations usually written on the language of Feynman diagram technique. But for us this approach is not inapplicable because the gravity is strongly self-interacting system. Heisenberg has applied nonperturbative technique for the quantization of a nonlinear spinor field [@heisenberg]. In Ref. [@Dzhunushaliev:2010qs] Heisenberg’s technique was applied for the quantization in quantum chromodynamics and it was shown that such approach leads to the description of a gluon condensate in a glueball and in a flux tube filled with a longitudinal color electric field and stretched between quark - antiquark located at $\pm$infinities. The approaches from [@heisenberg; @Dzhunushaliev:2010qs] are not applicable for the quantization of general relativity because the nonlinearities in gravity are extremely strong. In usual field theory the nonlinearities are in potential term but in gravity they are in kinetic term. Therefore we can not apply the nonperturbative quantization methods from usual quantum field theory to the quantization of gravity. Nonperturbative calculations ============================ Excluding the hats $\widehat{(\cdots)}$ from equation we obtain the classical Riemann tensor for the affine connection $\Gamma^\rho_{\phantom{\rho} \mu \nu}$ as $$\tilde R^\rho_{\phantom{\rho} \sigma \mu \nu} = \frac{\partial \Gamma^\rho_{\phantom{\rho} \sigma \nu}}{\partial x^\mu} - \frac{\partial \Gamma^\rho_{\phantom{\rho} \sigma \mu}}{\partial x^\nu} + \Gamma^\rho_{\phantom{\rho} \tau \mu} \Gamma^\tau_{\phantom{\tau} \sigma \nu} - \Gamma^\rho_{\phantom{\rho} \tau \nu} \Gamma^\tau_{\phantom{\tau} \sigma \mu} \label{1-10}$$ where $\tilde{\phantom{R}}$ means that the scalar curvature $\tilde R$ is calculated for the affine connection $\Gamma$ which in general does not coincide with the Christoffel symbols $\{ \}$. Einstein - Hilbert Lagrangian is $$\mathcal L = \sqrt{-g} g^{\mu \nu} \tilde R^\rho_{\phantom{\rho} \mu \rho \nu} . \label{1-20}$$ In the classical approach varying with respect to independent variables the connection $\Gamma^\rho_{\phantom{\rho} \mu \nu}$ and the metric $g_{\mu \nu}$ leads to the fact that the connection becomes metric dependent $$\Gamma^\rho_{\phantom{\rho} \mu \nu} = \{^\rho_{\mu \nu}\} \label{1-30}$$ where $\{^\rho_{\mu \nu}\}$ are the Christoffel symbols. In general relativity we can use the metric and the connection as independent variables but using Palatini approach one can show that the connection is metric dependent. For quantum gravity again we work with the metric and the connection as with independent variables but now we assume that the Palatini relation is true for the operators only. Since the Palatini relation is nonlinear according to the metric it is not true for the averaged quantities $\left\langle \Gamma \right\rangle$ and $\left\langle g_{\mu \nu} \right\rangle$. Consequently the affine connection will have some additional contributions in addition to Christoffel symbols: torsion. Firstly we will consider the situation with the affine connection $\Gamma$. We suppose that the affine connection fluctuates about the Christoffel symbols. It means that the first natural assumption for the affine connection is following $$\left\langle \hat \Gamma^\rho_{\phantom{\rho} \mu \nu} \right\rangle \approx \{^\rho_{\mu \nu}\} \label{1-40}$$ but $$\left\langle \hat \Gamma^2 \right\rangle \neq \{\}^2 . \label{1-50}$$ The second assumption is about the correlation between the affine connections in two points. We suppose that such correlation (Green) function has a classical term (arising from the Christoffel symbols) and some quantum correction: $$\left\langle \hat \Gamma^{\rho_1}_{\phantom{\rho_1} \mu_1 \nu_1} (x_1) \hat \Gamma^{\rho_2}_{\phantom{\rho_2} \mu_2 \nu_2} (x_2) \right\rangle \approx \{^{\rho_1}_{\mu_1 \nu_1} \} (x_1) \{^{\rho_2}_{\mu_2 \nu_2} \} (x_2) + C^{\rho_1 \rho_2}_{\mu_1 \nu_1 \mu_2 \nu_2} \tilde B(x_1, x_2 ) \label{1-60}$$ where $C^{\rho_1 \rho_2}_{\mu_1 \nu_1 \mu_2 \nu_2}$ is some constant; $\tilde B(x_1, x_2 )$ is some function approximately describing the effects arising from the nonperturbative quantization. Secondly we will consider the situation with the quantization of the metric. The metric is a part of Lagrangian as the factor $\sqrt{-g} g^{\mu \nu}$. A distinguishing feature of this term is that it has not any derivatives (let us remember that we consider the metric and the connection as independent variables). Therefore we assume that nonperturbative quantum corrections for the factor $\sqrt{-g} g^{\mu \nu}$ have following multiplicative form $$\left\langle \widehat{\sqrt{-g}} \; \hat g^{\mu \nu} \right\rangle \approx \sqrt{-g} \; g^{\mu \nu} \left( 1 + A \right) \label{1-70}$$ where $\widehat{\sqrt{-g}}, \hat g^{\mu \nu}$ are the operators of the measure $\sqrt{-g}$ and the contravariant metric $g^{\mu \nu}$. The term $A \gtrsim 1$ for a strong quantum regime and $A \ll 1$ for a weak quantum region. Thirdly we assume that in the first approximation the factors $\widehat{\sqrt{-g}} \, \hat g^{\mu \nu}$ and $\hat{\tilde R}_{\mu \nu}$ do not correlate in the Lagrangian. Averaging Lagrangian with the operators $$\hat{\mathcal L} = \widehat{\sqrt{-g}} \, \hat g^{\mu \nu} \hat{\tilde R}_{\mu \nu} \label{1-75}$$ under a quantum state $\left.\left. \right| Q \right\rangle$ and taking into account we obtain following effective Lagrangian $$\mathcal L_{eff} = \left\langle \hat{\mathcal L} \right\rangle \approx \left\langle \widehat{\sqrt{-g}} \, \hat g^{\mu \nu} \right\rangle \left\langle \hat{\tilde R}_{\mu \nu} \right\rangle \approx \sqrt{-g} g^{\mu \nu} \left( 1 + A \right) \left[ R + B \right] \label{1-80}$$ where $R$ is the scalar curvature for the Christoffel symbols and $B$ is expressed through $C^{\rho_1 \rho_2}_{\mu_1 \nu_1 \mu_2 \nu_2}$ and $\tilde B$. The next assumption have to be done on the nonperturbative quantum corrections $A$ and $B$. We can not calculate their exactly. Most likely that by nonperturbative quantum calculations we have to do some well-founded physical assumptions of Green functions to cut off an infinite equations set for all Green functions. Bearing in mind this remark we assume that both functions $A$ and $B$ are function of one scalar field $\phi$: $A(\phi)$ and $B(\phi)$ (one scalar approximation). It leads to following effective Lagrangian $$\mathcal L_{eff} = \sqrt{-g} \left[ 1 + A(\phi) \right] \left[ R + B(\phi) \right] . \label{1-90}$$ Lagrangian is an intermediate Lagrangian on the way between scalar-tensor and $F(R)$ gravities (for details one can see the review [@Nojiri:2010wj]). Varying corresponding action with respect to $\phi$ we will obtain the equation for $\phi$ $$\left[ R + B(\phi) \right] A'(\phi) + \left[ 1 + A(\phi) \right] B'(\phi) = 0 \label{1-100}$$ where $d(\cdots)/d\phi = (\cdots)'$. This equation could be solved with respect to $\phi = \phi(R)$. After that we can exclude function $\phi$ from the Lagrangian $$\mathcal L_{eff} = \left\langle \mathcal L \right\rangle \approx \sqrt{-g} \left[ 1 + A(\phi(R)) \right] \left[ R + B(\phi(R)) \right] = \sqrt{-g} \, F(R) \label{1-110}$$ and obtain the $F(R)-$Lagrangian. Reconstructing of $\Lambda$CDM-type cosmology {#reconstructing} --------------------------------------------- In this subsection we would like to reconstruct $F(R)-$gravity having $\Lambda$CDM-type cosmology. In this case we will see what there should be the quantum corrections to have $\Lambda$CDM-type cosmology. Here we follow to Ref. [@Nojiri:2010wj], section III C 1. The $F(R)$ action is $$\label{2-10} S_{F(R)}= \int d^4 x \sqrt{-g} \left[ \frac{F(R)}{2\kappa^2} + \mathcal{L}_\mathrm{matter} \right] .$$ The action can be equivalently rewritten in the form as follows $$\label{2-20} S=\int d^4 x \sqrt{-g} \left\{ \frac{1}{2\kappa^2} \left[ P(\phi) R + Q(\phi) \right] + \mathcal{L}_\mathrm{matter} \right\}$$ where in our designation $1+A = P$ and $B=Q$ and the matter is added. To reconstruct the function $F(R)$ it is necessary to obtain the functions $P(R)$ and $Q(R)$. It can be done in the following way. The spatially flat FRW equations for $F(R)-$gravity are: $$\begin{aligned} \label{2-30} -6 H^2 P(\phi) - Q(\phi) - 6H\frac{dP(\phi(t))}{dt} + 2\kappa^2 \rho_\mathrm{matter} &=& 0 , \\ \label{2-40} \left(4\dot H + 6H^2\right)P(\phi) + Q(\phi) + 2\frac{d^2 P(\phi(t))}{dt^2} + 4H\frac{d P(\phi(t))}{dt} + 2\kappa^2 p_\mathrm{matter} &=& 0 .\end{aligned}$$ If we would like to reconstruct $F(R)-$modified gravity with the $\Lambda$CDM-type cosmology we have to consider as the equation for the function $Q(\phi)$ and the difference - as the equation for the function $P(\phi)$ where Hubble parameter $H = \dot a/a$ is given as $$\label{2-50} H = \dot g, \quad a = a_0 e^{g(t)} ,\quad g(t)=\frac{2}{3(1+w)}\ln \left\{ \alpha \sinh \left[ \frac{3(1+w)}{2l}\left( t - t_s \right) \right] \right\}$$ here, $t_s$ is a constant of the integration and $$\label{2-60} \alpha^2\equiv \frac{1}{3}\kappa^2 l^2 \rho_0 a_0^{-3(1+w)} ,$$ $w=p/\rho$ is the parameter of the equation of state; $l^2$ is the inverse cosmological constant and $\rho_\mathrm{matter}$ is given as $$\label{2-70} \rho_\mathrm{matter}=\rho_0 a^{-3(1+w_\mathrm{matter})}$$ One can exclude the function $Q$ from Eq’s obtaining inhomogeneous equation for the function $P$. Because this equation is a linear inhomogeneous equation, its general solution is given by the sum of the general solution that corresponds to the homogeneous equation $$\begin{aligned} \label{2-80} P = P_0 F(\tilde\alpha,\tilde\beta,\tilde\gamma; \phi) = P_0 \frac{\Gamma(\tilde\gamma)}{\Gamma(\tilde\alpha) \Gamma(\tilde\beta)} \sum_{n=0}^\infty \frac{\Gamma(\tilde\alpha + n) \Gamma(\beta + n)}{\Gamma(\tilde\gamma + n)} \frac{\phi^n}{n!} , \\ \tilde\gamma = 4 + \frac{1}{3(1+w)},\ \tilde\alpha + \tilde\beta + 1 = 6 + \frac{1}{3(1+w)},\ \tilde\alpha \tilde\beta = - \frac{1}{3(1+w)} \label{2-90}\end{aligned}$$ and of the special solution $$P=P_0(-\phi)^{-2/3} ,\quad P_0 = \frac{\eta}{\frac{10}{9} - \frac{2\left(\tilde\alpha + \tilde\beta + 1\right)}{3} + \tilde\alpha\tilde\beta} = - \frac{9\eta}{25} \label{2-100}$$ where $\Gamma$ is the $\Gamma-$function and $Q$ can be determined from equation . In the limit of $\phi \to + \infty$, one can arrive at $P(\phi)R + Q(\phi) \to P_0 R - 6P_0/l^2$. Identifying $P_0=1/2\kappa^2$ and $\Lambda = 6/l^2$, the Einstein theory with cosmological constant $\Lambda$ can be reproduced. Finally, having $P$ and $Q$ and recalling our designations $1+A = P$ and $B=Q$ we reconstruct what kind of quantum correlations should be to have $\Lambda$CDM-type cosmology. The case $A,B =$ const {#constant} ---------------------- Here we would like to discuss separately the simplest case when above mentioned nonperturbative quantum corrections $A,B$ have the simplest possible form: they are constants $A,B =$ const. In this case the Einstein constant $\kappa$ is redefined in the following way $$\frac{1}{2 \kappa^2_r} = \frac{1+A}{2 \kappa^2} \label{3-10}$$ and the cosmological constant $\Lambda$ is defined as $$- 2 \Lambda = B. \label{3-20}$$ Thus, if the quantum corrections for the connection and metric are constants then it leads to the $\Lambda$CDM-type cosmology directly. Now we would like to discuss in more details the essence of the quantum corrections of the affine connection. Let us remember that a torsion $Q^\mu_{\rho \sigma}$ is introduced in following way $$\Gamma^\mu_{\rho \sigma} = \left\{ ^\mu_{\rho \sigma}\right\} + Q^\mu_{\rho \sigma} \label{3-30}$$ where $\left\{ ^\mu_{\rho \sigma}\right\} = \Gamma^\mu_{\{ \rho \sigma \}}$ is the symmetrical part of the affine connection and the torsion $Q^\mu_{\rho \sigma} = \Gamma^\mu_{[ \rho \sigma ]}$ is the skew symmetrical part of the affine connection $\Gamma^\mu_{\rho \sigma}$. Then $$\Lambda \approx g^{\mu \nu} \left\langle \hat Q^\rho_{\tau \rho} \hat Q^\tau_{\mu \nu} - \hat Q^\rho_{\tau \nu} \hat Q^\tau_{\mu \rho} \right\rangle \equiv \left\langle \hat Q^2 \right\rangle \label{3-40}$$ where $\hat Q^\mu_{\rho \sigma}$ is the torsion operator. Taking into account we can estimate the variance of torsion (in the almost flat space $g^{\mu \nu} \approx 1$) as $$\left\langle \hat Q^2 \right\rangle \approx \left| \Lambda \right| \approx 10^{-57} \text{cm}^{-2} \quad \text{ with } \left\langle \hat Q \right\rangle = 0. \label{3-50}$$ Thus in the presented approach the modern acceleration epoch is caused by an extremely small torsion effect. The torsion appears as a nonperturbative quantum correction to the Christoffel symbols. In principle the torsion effects can be measured. The main problem for testing this torsion approach to $\Lambda$ cosmological constant is that $\left\langle \hat Q \right\rangle = 0$ and $\left\langle \hat Q^2 \right\rangle$ is extremely small. Conclusions =========== The main result of this paper is that $F(R)-$modified gravities can be considered as some nonperturbative quantum effects arising from Einstein gravity. In the subsections \[reconstructing\] and \[constant\] it was shown what kind should be quantum corrections to have $\Lambda$CDM-type cosmology. For the calculations we have used some assumptions about nonperturbative quantum effects because up to now we do not have any quantum gravity theory. These assumptions are: (a) the connection and the metric are independent dynamic variables; (b) in the first approximation these quantum effects give rise to an additional terms for the squared connection and the metric factors in Einstein-Hilbert action. As the result one can obtain an effective Lagrangian with the Ricci scalar + scalar field approximately describing nonperturbative quantum effects. It allows us to exclude the scalar fields from the Lagrangian obtaining $F(R)-$modified gravity. The result of this paper can be formulated as follows. For the nonperturbative gravity quantization we have the infinite equations set . If we could solve these equations precisely and then exclude all quantum effects then we will obtain some very complicated gravity (even may be nonlocal) with $F(R), R^{\mu \nu}R_{\mu \nu}, R^{\mu \nu \rho \sigma}R_{\mu \nu \rho \sigma}$ and so on. That we have made here it is very simple approximation of such exclusion process. We have shown that for the modern epoch $\Lambda$-term can be considered as a torsion effect where the torsion appears as nonperturbative correction for affine connection. The symmetrical part of the affine connection is a classical part and the skew symmetrical part are quantum corrections. In the approach presented here both inflation era and modern acceleration epoch are quantum effects. For the first case it is not surprisingly because the Universe has started from a Planck scale. But why the modern acceleration is made from quantum effects ? In our opinion it is similar to a hypothesized flux tube in quantum chromodynamics. In quantum chromodynamics the SU(3) gauge field between quark and antiquark is confined into a flux tube stretched between quark and antiquark even they are located at $+\infty$ and $-\infty$. Acknowledgments {#acknowledgments .unnumbered} =============== I am grateful to the Research Group Linkage Programme of the Alexander von Humboldt Foundation for the support of this research. [999]{} I.L. Buchbinder, S.D. Odintsov and I.L. Shapiro,, “Effective action in quantum gravity”, Institute of Physics Publishing;\ Mark H. Goroff, A. Sagnotti, “The ultraviolet behavior of Einstein gravity”, CALT-68-1289. N. D. Birrell and P. C. W. Davies, “Quantum fields in curved space”, Cambridge University Press, 1982. S. Nojiri and S. D. Odintsov, Phys. Rept.  [**505**]{}, 59 (2011) \[arXiv:1011.0544 \[gr-qc\]\]. S. Nojiri and S. D. Odintsov, eConf [**C0602061**]{}, 06 (2006) \[Int. J. Geom. Meth. Mod. Phys. [**4**]{}, 115 (2007)\] \[arXiv:hep-th/0601213\]; arXiv:0807.0685. S. Capozziello, M. De Laurentis and V. Faraoni, arXiv:0909.4672 \[gr-qc\];\ S. Capozziello and M. Francaviglia, Gen. Rel. Grav. [**40**]{}, 357 (2008) \[arXiv:0706.1146 \[astro-ph\]\];\ T. P. Sotiriou and V. Faraoni, Rev. Mod. Phys. [**82**]{}, 451 (2010) \[arXiv:0805.1726 \[gr-qc\]\];\ F. S. N. Lobo, arXiv:0807.1640 \[gr-qc\]. W. Heisenberg, *Introduction to the unified field theory of elementary particles.*, Max - Planck - Institut für Physik und Astrophysik, Interscience Publishers London, New York, Sydney, 1966. V. I. Zakharov, Int. J. Mod. Phys. A [**14**]{}, 4865 (1999) \[hep-ph/9906264\]. V. Dzhunushaliev, “SU(3) flux tube gluon condensate,” arXiv:1010.1621 \[hep-ph\];\ V. Dzhunushaliev, “SU(3) glueball gluon condensate,” \[arXiv:1110.1427 \[hep-ph\]\]. [^1]: where $\left.\left. \right| Q \right\rangle$ is a quantum state, $\{ \}$ are Christoffel symbols and the function $B(\phi)$ describes the effects of the nonperturbative quantization of the affine connection $\Gamma$ [^2]: in quantum chromodynamics there exist various gluon condensates: $\mathrm{tr}\left(F_{\mu \nu} F^{\mu \nu} \right)$, $\mathrm{tr} \left( F_\mu^\nu F_\nu^\rho F_\rho^\mu \right)$ and so on, for the review on the gluon condensate one can see Ref. [@Zakharov:1999jj]. Using these condensates and integrating their one can obtain various functionals.

ArXiv

--- abstract: 'Let $\varphi$ be a real-valued smooth function on ${\ensuremath{\mathbb{C}}}$ satisfying $0\leq\Delta\varphi\leq M$ for some $M \ge 0$. Denote by ${\ensuremath{\mathcal{H}L^{2}(\mathbb{C},e^{-\varphi})}}$ the space of all holomorphic functions which are square-integrable with respect to the measure $e^{-\varphi(z)}\,dz$. In this paper, we obtain a pointwise bound for any function in this space. We show that there exists a constant $K$ depending only on $M$ such that $$|f(z)|^2\leq Ke^{\varphi(z)}\|f\|^2_{L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})}$$ for any $f\in{\ensuremath{\mathcal{H}L^{2}(\mathbb{C},e^{-\varphi})}}$ and any $z\in{\ensuremath{\mathbb{C}}}$.' address: | Department of Mathematics\ Faculty of Science\ Chulalongkorn University\ Bangkok, Thailand 10330 author: - | Kamthorn Chailuek\ Wicharn Lewkeeratiyutkul\ title: 'A POINTWISE BOUND FOR A HOLOMORPHIC FUNCTION WHICH IS SQUARE-INTEGRABLE WITH RESPECT TO AN EXPONENTIAL DENSITY FUNCTION' --- Introduction ============ Let $U$ be a non-empty open subset of ${\ensuremath{\mathbb{C}}}$. Denote by ${\ensuremath{\mathcal{H}}}L^2(U,\alpha)$ the space of all holomorphic functions on $U$ which are square-integrable with respect to the measure $\alpha(\omega)\,d\omega$. For any $t>0$, consider the Gaussian measure $$d\mu_t(z)\;=\;\frac{1}{\pi t}e^{-|z|^2/t}\,dz.$$ Then the space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$ is called the *Segal-Bargmann space*. See \[GM\], \[H1\], \[H2\], \[F\] for detailed discussion about the importance of this space, and its relevance in quantum theory. It is well-known that a pointwise bound for any function $f \in {\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$ is given by $$\label{pwb:s:intro} |f(z)|^2 \; \leq \; e^{|z|^2/t}\|f\|^2_{L^2({\ensuremath{\mathbb{C}}},\mu_t)}.$$ This pointwise bound first appeared in Bargmann’s paper \[B\] and was revisited many times by other authors. More generally, for any space ${\ensuremath{\mathcal{H}}}L^2(U,\alpha)$, there exists a function $K(z,\omega)$ on $U\times U$, called the *reproducing kernel*, such that $$\label{pwb:k:intro} |f(z)|^2 \; \leq \; K(z,z)\|f\|^2_{L^2(U,\alpha)}$$ for any $f\in {\ensuremath{\mathcal{H}}}L^2(U,\alpha)$ and $z\in U$. The Bargmann’s pointwise bound (\[pwb:s:intro\]) for ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$ follows from the following formula of the reproducing kernel for the Segal-Bargmann space: $$\label{rep:sb:intro} K(z,\omega) \; = \; e^{z\overline{\omega}/t}.$$ In this work, we study a pointwise bound for a function in a more general holomorphic function space. First, we look at the space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$, where $\Delta\varphi$ is a positive constant. Note that $\Delta(|z|^2/t) = 4/t > 0$, so this is a generalization of the standard Segal-Bargmann space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$. The technique used here will be that of holomorphic equivalence \[H1\]. Two holomorphic function spaces ${\ensuremath{\mathcal{H}}}L^2(U,\alpha)$ and ${\ensuremath{\mathcal{H}}}L^2(U,\beta)$ are holomorphically equivalent if there exists a nowhere-zero holomorphic function $\phi$ on $U$ such that $$\beta(z)=\frac{\alpha(z)}{|\phi(z)|^{2}} \qquad \text{for all $z\in U$}.$$ If ${\ensuremath{\mathcal{H}}}L^2(U,\alpha)$ and ${\ensuremath{\mathcal{H}}}L^2(U,\beta)$ are holomorphically equivalent spaces, then their reproducing kernels are related by $$\label{h:equi:intro} \alpha(z)K_{\alpha}(z,z)\;=\;\beta(z)K_{\beta}(z,z).$$ We show that if $\Delta\varphi=c>0$, then ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ is holomorphically equivalent to the Segal-Barmann space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$ where $t=4/c$. It follows from (\[pwb:k:intro\]) and (\[h:equi:intro\]) that $$|f(z)|^2 \; \leq \; \frac{c}{4\pi}e^{\varphi(z)}\|f\|_{L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})}^2,$$ for any $f \in {\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ and any $z\in{\ensuremath{\mathbb{C}}}$.\ Next, we turn to the space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$, where $\Delta\varphi$ is positive and bounded, i.e. $0\leq\Delta\varphi\leq M$ for some $M \ge 0$. This space is not holomorphically equivalent to a Segal-Bargmann space, so we cannot apply the same technique here. Our proof relies on a technical lemma which can be stated as follows: For any $f\in{\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$, $$|f(0)|^2\;\leq\;Ce^{\varphi(0)}\int_{D(0,1)}|f(\omega)|^2e^{-\varphi(\omega)}\,d\omega$$ for some $C$ depending only on $M$. By translation to any point $z\in {\ensuremath{\mathbb{C}}}$, we obtain the following pointwise bound: $$|f(z)|^2 \; \leq \; Ce^{\varphi(z)}\|f\|^2_{L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})}.$$ Here is a brief summary of this work. In section 2, we study basic properties of holomorphic function spaces. We introduce the concept of holomorphic equivalence and establish a necessary and sufficient condition for two spaces to be holomorphically equivalent. In section 3, we establish a pointwise bound for functions in ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$. Holomorphic function spaces =========================== In this section, we review and prove some relevant facts about holomorphic function spaces that are needed in this paper. The main reference here is \[H1\]. Let $U$ be a non-empty open subset of ${\ensuremath{\mathbb{C}}}$. Denote by ${\ensuremath{\mathcal{H}}}(U)$ the space of all holomorphic functions on $U$. If $\alpha$ is a strictly positive function on $U$, let $L^2(U,\alpha)$ be the space of all functions on $U$ which are square-integrable with respect to the measure $\alpha(\omega)\,d\omega$. Then $L^2(U,\alpha)$ is a Hilbert space. Let ${\ensuremath{\mathcal{H}}}L^2(U,\alpha)={\ensuremath{\mathcal{H}}}(U)\cap L^2(U,\alpha)$. Then ${\ensuremath{\mathcal{H}}}L^2(U,\alpha)$ is a closed subspace of $L^2(U,\alpha)$ and hence a Hilbert space. Moreover, it is well-known that ${\ensuremath{\mathcal{H}}}L^2(U,\alpha)$ is separable. A *Segal-Bargmann space* is a space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$, where $$\mu_t(z)=\frac{1}{\pi t}e^{-|z|^2/t}$$ for some $t>0$. Let $K \colon U \times U \to {\ensuremath{\mathbb{C}}}$ be a reproducing kernel for the space ${\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$. We refer to \[H1\] for details of the discussion below. If $\{e_{i}\}_{i=0}^{\infty}$ is an orthonormal basis for ${\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$, then the reproducing kernel $K$ is given by $$\label{ker:ort:eqt} K(z,\omega) \; = \; \sum_{i=0}^{\infty}e_{i}(z) \overline{e_{i}(\omega)} \qquad (z,\,\omega\in U).$$ If we know the reproducing kernel of the space, the pointwise bound of any function $f$ in ${\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$ can be obtained by $$\label{est:k:f} |f(z)|^{2} \; \leq \; K(z,z)\|f\|_{L^2(U,\alpha)} ^{2}.$$ Moreover, for a fixed value of $z$, $K(z,z)$ is the smallest constant which makes the pointwise bound (\[est:k:f\]) holds for all $f \in {\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$. \[defn:holo-equiv\] Holomorphic function spaces ${\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$ and ${\ensuremath{\mathcal{H}}}L^{2}(U,\beta)$ are said to be *holomorphically equivalent* spaces if there exists a nowhere zero holomorphic function $\phi$ on $U$ such that $$\beta(z) = \frac{\alpha(z)}{|\phi(z)|^{2}} \qquad \text{for all $z\in U$.}$$ In this case, the map $f \mapsto \phi f$ is a unitary map from ${\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$ onto ${\ensuremath{\mathcal{H}}}L^2(U,\beta)$. \[inv:k\] Let ${\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$ and ${\ensuremath{\mathcal{H}}}L^{2}(U,\beta)$ be holomorphically equivalent spaces. Let $K_{\alpha}$ and $K_{\beta}$ be their respective reproducing kernels. Then for each $z\in U$, $$\alpha(z)K_{\alpha}(z,z)=\beta(z)K_{\beta}(z,z).$$ By formula \[ker:ort:eqt\] and the fact that a unitary map preserves orthonormal bases, we obtain $$K_{\beta}(z,\omega) \; = \; \phi(z)\overline{\phi(\omega)}K_{\alpha}(z,\omega).$$ It follows that $$K_{\beta}(z,z)\; = \; |\phi(z)|^{2}K_{\alpha}(z,z) = \;\frac{\alpha(z)}{\beta(z)}K_{\alpha}(z,z).$$ Thus, $\alpha(z)K_{\alpha}(z,z)=\beta(z)K_{\beta}(z,z)$. The next goal in this section is to establish a necessary and sufficient condition for two spaces to be holomorphically equivalent. \[Lemma-before-nec-suff\] Let $U$ be an open simply connected set in ${\ensuremath{\mathbb{C}}}$ and $\alpha$ a strictly positive smooth function on $U$. Then there exists a holomorphic function $\phi$ such that $|\phi|^{2}=\alpha$ if and only if $\log \alpha$ is harmonic. $(\Rightarrow)$ Since $\phi \in {\ensuremath{\mathcal{H}}}(U)$, by a standard result in complex analysis, there exists a function $\theta \in {\ensuremath{\mathcal{H}}}(U)$ such that $\phi = e^{\theta}$. Let $u=\text{Re}\,\theta$. Thus, $|\phi|=e^{u}$ and hence $\alpha = e^{2u}$. Then $\log\alpha = 2u$, which implies that $\Delta \log \alpha = \Delta 2u = 0$.\ $(\Leftarrow)$ Assume that $u = \log \alpha$ is harmonic. Then there exists a holomorphic function $f$ such that $u=\text{Re}f$. Hence, $e^{f}$ is also holomorphic. Let $\phi = e^{f/2}$. Then $\phi \in {\ensuremath{\mathcal{H}}}(U)$ and $e^f = \phi^2$. Hence, $\alpha = e^{u}=|e^{f}|=|\phi|^{2}.$ \[nec&suf\] Let $U$ be an open simply connected set in ${\ensuremath{\mathbb{C}}}$ and $\alpha$, $\beta$ strictly positive smooth functions on $U$. Then ${\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$ and ${\ensuremath{\mathcal{H}}}L^{2}(U,\beta)$ are holomorphically equivalent spaces if and only if $\Delta \log \alpha(z)=\Delta \log \beta(z)$. If ${\ensuremath{\mathcal{H}}}L^{2}(U,\alpha)$ and ${\ensuremath{\mathcal{H}}}L^{2}(U,\beta)$ are holomorphically equivalent, then there is a function $\phi \in {\ensuremath{\mathcal{H}}}(U)$ such that $\phi \neq 0$ and $|\phi(z)|^{2}= \frac{\alpha(z)}{\beta(z)}$. By Lemma \[Lemma-before-nec-suff\], $\log \frac{\alpha(z)}{\beta(z)}$ is harmonic. Hence, $\Delta(\log\alpha(z)-\log\beta(z))=0$, which shows that $\Delta\log\alpha(z)=\Delta\log\beta(z)$. It is easy to see that the reverse implication is true in each step. This immediately implies the following corollary: \[cor:nec&suf\] A holomorphic function space ${\ensuremath{\mathcal{H}}}L^{2}({\ensuremath{\mathbb{C}}},\alpha)$, where $\alpha$ is a strictly positive smooth function on ${\ensuremath{\mathbb{C}}}$, is holomorphically equivalent to one of the Segal-Bargmann spaces if and only if $\Delta \log \alpha =c < 0$. In particular, if $\varphi$ is a smooth function and $\Delta \varphi $ is a positive constant, then the space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ is holomorphically equivalent to a Segal-Bargmann space. Note that if $$\mu_t(z)\;=\;\frac{1}{\pi t}e^{-|z|^2/t},$$ then $$\Delta\log\mu_t(z)\,=\,-\Delta\frac{|z|^2}{t} \,=\,-\frac{4}{t}\frac{\partial^2}{\partial z \partial\overline{z}}(z\overline{z}) \,=\,-\frac{4}{t}\,<\,0.$$ Thus if ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\alpha)$ is holomorphically equivalent to the Segal-Bargmann space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$, then $\Delta\log\alpha = \Delta\log\mu_t < 0$. Conversely, if $\Delta\log\alpha = c < 0$, then $\Delta\log\alpha = \Delta\log\mu_t$ where $t=-4/c$. Therefore, ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\alpha)$ is holomorphically equivalent to the Segal-Bargmann space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$, where $t=-4/c$. A pointwise bound for a function in ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ =========================================================================================================== In this section, we obtain a pointwise bound for any function in the holomorphic function space ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$. First, we look at the case where $\Delta\varphi$ is a positive constant. Let $\varphi$ be a smooth function such that $\Delta\varphi=c$ where $c$ is a positive constant. Then, for any $f\in{\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ and any $z\in{\ensuremath{\mathbb{C}}}$, $$\label{pointwise-bound-const-case} |f(z)|^2 \; \leq \; \frac{c}{4\pi}e^{\varphi(z)}\|f\|_{L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})}^2.$$ By Corollary \[cor:nec&suf\], ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ is holomorphically equivalent to ${\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},\mu_t)$, where $t=4/c$. Then, by Lemma \[inv:k\], $$K_{e^{-\varphi}}(z,z) \;=\; \frac{1}{\pi t}e^{\varphi(z)} \;=\; \frac{c}{4\pi}e^{\varphi(z)}.$$ It follows that $$|f(z)|^2 \; \leq \; \frac{c}{4\pi}e^{\varphi(z)}\|f\|_{L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})}^2,$$ for any $f\in{\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ and any $z\in{\ensuremath{\mathbb{C}}}$. Note that when $\varphi = |z|^2/t$, we have $c = \Delta \varphi = 4/t$. Hence, in this case (\[pointwise-bound-const-case\]) reduces to the usual pointwise bound (\[pwb:s:intro\]) for the Segal-Bargmann space. Next, we turn to the situation in which $0 \leq \Delta\varphi \leq M$. The main result is contained in Theorem \[pwb:general-case\]. But first we need to establish a technical lemma. Recall that the function $\Gamma$ defined by $$\Gamma(z)=\frac{1}{2\pi}\log|z|$$ is the *fundamental solution* for the Laplace’s equation on ${\ensuremath{\mathbb{R}}}^2$. Thus if $\psi\in C_c^{\infty}({\ensuremath{\mathbb{C}}})$, then $$\Phi(z) \; = \; \Gamma\ast\psi(z) \; = \; \int_{{\ensuremath{\mathbb{C}}}}\Gamma(\zeta)\psi(z-\zeta)\,d\zeta$$ satisfies $\Delta\Phi=\psi$. \[z:0\] Let $\varphi\in C^{\infty}({\ensuremath{\mathbb{C}}})$ satisfying $0 \leq \Delta\varphi \leq M$. Then there exists a constant $C$ depending only on $M$ such that for any $f \in {\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$, $$|f(0)|^2 \leq C e^{\varphi(0)}\int_{D(0,1)}|f(\omega)|^2e^{-\varphi(\omega)}\,d\omega.$$ Choose a function $g\in C_{c}^{\infty}({\ensuremath{\mathbb{C}}})$ such that $0 \le g \le 1$, $g=1$ on $\overline{D(0,1)}$ and $g=0$ outside $D(0,2)$. Let $\psi = g\,\Delta\varphi$. Then $\psi\in C_{c}^{\infty}({\ensuremath{\mathbb{C}}})$, $0 \le \psi \le M$, $\psi=\Delta\varphi$ on $\overline{D(0,1)}$ and $\psi=0$ outside $D(0,2)$. Thus $\Phi=\Gamma\ast\psi$ satisfies $$\label{con:h:e} \Delta\Phi(z)=\psi(z)=\Delta\varphi(z)$$ for all $z\in D(0,1)$. First, we show that $\Phi$ is bounded above on $D(0,1)$. Note that $\Gamma(\zeta)\le 0$ if and only if $\zeta \in D(0,1)$. For any $\omega \in D(0,1)$, we have $$\begin{aligned} \Phi(\omega)\; &= \; \int_{{\ensuremath{\mathbb{C}}}}\Gamma(\zeta)\psi(\omega-\zeta)\,d\zeta \\ &= \; \int_{D(\omega,2)}\Gamma(\zeta)\psi(\omega-\zeta)\,d\zeta \\ &\leq \; \int_{D(\omega,2)\setminus D(0,1)}\Gamma(\zeta)\psi(\omega-\zeta)\,d\zeta \\ &\le \; \frac{M}{2\pi}\int_{D(\omega,2)\setminus D(0,1)}\log|\zeta|d\zeta.\end{aligned}$$ This shows that $\Phi(\omega) \le BM$ for all $\omega \in D(0,1)$, where $$B = \frac{1}{2\pi}\sup_{\omega\in D(0,1)}\int_{D(\omega,2)\setminus D(0,1)}\log|\zeta|d\zeta.$$ Write $\mathcal{U}= D(0,1)$ and let $h\in{\ensuremath{\mathcal{H}}}L^2(\mathcal{U},e^{-\Phi})$. Fix $0 < s < 1$. It is not hard to show that $$h(0) = \frac{1}{\pi s^2}\int_{D(0,s)}h(\omega)\,d\omega.$$ By the Cauchy-Schwarz inequality, it follows that $$|h(0)|^2 \; \leq \; (\pi s^2)^{-2} \left\|\chi_{D(0,s)}e^{\Phi}\right\|_{L^2(\mathcal{U},e^{-\Phi})}^2 \|h\|_{L^2(\mathcal{U},e^{-\Phi})}^2.$$ Hence, $$\begin{aligned} \left\|\chi_{D(0,s)}e^{\Phi}\right\|_{L^2(\mathcal{U},e^{-\Phi})}^2\; = \; \int_{D(0,s)}e^{\Phi(\omega)}\,d\omega \; \leq \; \int_{D(0,s)}e^{BM}\,d\omega \; = \; e^{BM}\pi s^2.\end{aligned}$$ Thus, for any $0 < s < 1$, $$|h(0)|^2 \; \leq \; \frac{e^{BM}}{\pi s^2}\,\|h\|^2_{L^2(\mathcal{U},e^{-\Phi})}.$$ It follows that $$|h(0)|^2 \; \leq \; \frac{e^{BM}}{\pi}\,\|h\|^2_{L^2(\mathcal{U},e^{-\Phi})}$$ for all $h \in {\ensuremath{\mathcal{H}}}L^2(\mathcal{U},e^{-\Phi})$. By a property of the reproducing kernel (see the paragraph preceding Definition \[defn:holo-equiv\]) we then have $$K_{e^{-\Phi}}(0,0) \; \leq \; \frac{e^{BM}}{\pi}$$ where $K_{e^{-\Phi}}$ is the reproducing kernel for ${\ensuremath{\mathcal{H}}}L^2(\mathcal{U},e^{-\Phi})$. Let $K_{e^{-\varphi}}$ be the reproducing kernel for ${\ensuremath{\mathcal{H}}}L^2(\mathcal{U},e^{-\varphi})$. Then, by equation (\[con:h:e\]) and Proposition \[nec&suf\], ${\ensuremath{\mathcal{H}}}L^2(\mathcal{U},e^{-\varphi})$ and ${\ensuremath{\mathcal{H}}}L^2(\mathcal{U},e^{-\Phi})$ are holomorphically equivalent and hence, by Lemma \[inv:k\], $$\begin{aligned} K_{e^{-\varphi}}(0,0) \; = \; \frac{e^{-\Phi(0)}}{e^{-\varphi(0)}}K_{e^{-\Phi}}(0,0) \; \leq \; C\,e^{\varphi(0)},\end{aligned}$$ where $C=e^{BM-\Phi(0)}/\pi$. Thus $$|h(0)|^2 \; \leq \; Ce^{\varphi(0)}\|h\|_{L^2(\mathcal{U},e^{-\varphi})}^2,$$ for any $h\in{\ensuremath{\mathcal{H}}}L^2(\mathcal{U},e^{-\varphi})$. Let $f \in {\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ and $h=f\big|_{\mathcal{U}}$. Then $h\in{\ensuremath{\mathcal{H}}}L^2(\mathcal{U},e^{-\varphi})$ and $$\begin{aligned} |f(0)|^2\;&= \; |h(0)|^2 \\ &\leq \; Ce^{\varphi(0)}\int_{D(0,1)}|h(\omega)|^2e^{-\varphi(\omega)}\,d\omega \\ &= \; Ce^{\varphi(0)}\int_{D(0,1)}|f(\omega)|^2e^{-\varphi(\omega)}\,d\omega.\end{aligned}$$ Finally, it remains to show that we can choose a constant $C$ to depend only on $M$. By straightforward calculations, we have $$\int_{D(0,1)}\Gamma(\zeta)\,d\zeta \; = \; -\frac{1}{4}.$$ Now, consider $$\begin{aligned} \Phi(0)\; = \; \int_{{\ensuremath{\mathbb{C}}}}\Gamma(\zeta)\psi(-\zeta)\,d\zeta \; \ge \;\int_{D(0,1)}\Gamma(\zeta)\psi(-\zeta)\,d\zeta \; \ge \; -\frac{M}{4}.\end{aligned}$$ Thus $e^{-\Phi(0)} \le e^{\frac{M}{4}}$, which shows that $C \le \frac{1}{\pi}e^{(B+\frac{1}{4})M}$. \[pwb:general-case\] Let $\varphi\in C^{\infty}({\ensuremath{\mathbb{C}}})$ with $0\leq\Delta\varphi\leq M$. Then there exists a constant $C$ depending only on $M$ such that for any $f \in {\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ and any $z \in {\ensuremath{\mathbb{C}}}$, $$|f(z)|^2 \; \leq \; Ce^{\varphi(z)}\|f\|_{L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})}^2.$$ Let $z\in{\ensuremath{\mathbb{C}}}$ and $g_z(\omega)=z+\omega$. Then $0\leq\Delta(\varphi\circ g_z)\leq M$. Let $f \in {\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})$ and $h = f\circ g_{z}$. Then $h \in {\ensuremath{\mathcal{H}}}L^2({\ensuremath{\mathbb{C}}},e^{-\varphi\circ g_z})$ and by Lemma \[z:0\], $$\begin{aligned} |f(z)|^2 \; &= \; |f\circ g_z(0)|^2 \; = \; |h(0)|^2 \\ &\leq\;Ce^{\varphi \circ g_z(0)}\int_{D(0,1)}|h(\omega)|^2e^{-\varphi\circ g_z(\omega)}\,d\omega\\ &=\;Ce^{\varphi(z)}\int_{D(0,1)}|f\circ g_z(\omega)|^2e^{-\varphi\circ g_z(\omega)}\,d\omega \\ &=\;Ce^{\varphi(z)}\int_{D(0,1)}|f(z+\omega)|^2e^{-\varphi(z+\omega)}\,d\omega\\ &\leq\;Ce^{\varphi(z)}\int_{{\ensuremath{\mathbb{C}}}}|f(\omega)|^2e^{-\varphi(\omega)}\,d\omega \\ &=\;Ce^{\varphi(z)}\|f\|_{L^2({\ensuremath{\mathbb{C}}},e^{-\varphi})}^2.\end{aligned}$$ Acknowledgments {#acknowledgments .unnumbered} =============== The authors are grateful to Brian Hall for helpful suggestions throughout the process of this work. We also thank Leonard Gross for useful comments. [999]{} V. Bargmann, On a Hilbert space of analytic functions and an associated integral transform, Part I, *Comm. Pure Appl. Math.* **14** (1961), 187–214. G. Folland, “Harmonic analysis on phase space,” Princeton Univ. Press, Princeton, N.J., 1989. L. Gross and P. Malliavin, Hall’s transform and the Segal-Bargmann map, in “Ito’s Stochastic Calculus and Probability Theory” (M. Fukushima, N. Ikeda, H. Kunita and S. Watanabe, Eds.), pp. 73–116. Springer-Verlag, Berlin/New York, 1996. B. Hall, Holomorphic methods in analysis and mathematical physics, in “First Summer School in Analysis and Mathematical Physics” (S. Pèrez Esteva and C. Villegas Blas, Eds.), pp. 1–59, Contemp. Math., Vol. 260, Amer. Math. Soc., Providence, RI, 2000. B. Hall, Harmonic Analysis with respect to heat kernel measure, *Bull. Amer. Math. Soc.* **38** (2001), 43–78.

ArXiv

--- abstract: 'We present a new method of measuring optical near-fields within ${\sim} 1\,\mathrm{nm}$ of a metal surface, based on rescattering of photoemitted electrons. With this method, we precisely measure the field enhancement factor for tungsten and gold nanotips as a function of tip radius. The agreement with Maxwell simulations is very good. Further simulations yield a field enhancement map for all materials, which shows that optical near-fields at nanotips are governed by a geometric effect under most conditions, while plasmon resonances play only a minor role. Last, we consider the implications of our results on quantum mechanical effects near the surface of nanostructures and discuss features of quantum plasmonics.' author: - Sebastian Thomas - Michael Krüger - Michael Förster - Markus Schenk - Peter Hommelhoff title: 'Probing of optical near-fields by electron rescattering on the 1nm scale' --- [^1] [^2] The excitation of enhanced optical near-fields at nanostructures allows the localization of electromagnetic energy on the nanoscale [@Novotny2006a; @Stockman2011]. At nanotips, this effect has enabled a variety of applications, most prominent amongst them are scanning near-field optical microscopy (SNOM) [@Wessel1985; @Inouye1994; @Kawata2009; @Schnell2011; @Hartschuh2008], which has reached a resolving power of $\mathrm{8\,nm}$ [@Raschke2005], and tip-enhanced Raman spectroscopy (TERS) [@Wessel1985; @Stockle2000]. Because of the intrinsic nanometric length scale, measuring and simulating the tips’ near-field has proven hard and led to considerably diverging results (see Refs. [@Novotny2006a; @Hartschuh2008] for overviews). Here we demonstrate a nanometric field sensor based on electron rescattering, a phenomenon well known from attosecond science [@Corkum2007]. It allows measurement of optical near-fields, integrating over only $1\,\mathrm{nm}$ right at the structure surface, close to the length scale where quantum mechanical effects become relevant [@Zuloaga2010; @Marinica2012; @Ciraci2012; @Wachter2012; @Teperik2013]. Hence, this method measures near-fields on a scale that is currently inaccessible to other techniques (such as SNOM or plasmonic methods in electron microscopy [@GarciadeAbajo2010; @Koh2011; @Willets2012]), and reaches down to the minimum length scale where one can meaningfully speak about a classical field enhancement factor. In the future, the method will allow tomographic reconstruction of the optical near-field and potentially the sensing of fields in more complex geometries such as bow-tie or split-ring antennas. ![(a) Time-averaged electric field (obtained from the simulation) near the apex of a tungsten tip ($R = 30\,\mathrm{nm}$, $\lambda = 800\,\mathrm{nm}$) in a plane spanned by the tip axis and the wave vector $\textbf{k}$ of the laser. The tip shape is indicated as a gray line. The white line at the bottom displays the near-field along $z = 0$ with the $1/e$ decay length $L$. The field rises from 1.2 to 3.4 over a distance of $29\,\mathrm{nm}$, where $1$ indicates the field strength in the bare laser focus without a tip. Note that the near-field is not symmetric with respect to the tip axis. This asymmetry is more prominent for larger tip radii [@Yanagisawa2010]. (b) Illustration of electron rescattering: electrons are emitted in the optical near-field of a metal nanotip. A fraction of the emitted electrons is driven back to the tip surface, where they can scatter elastically. The kinetic energy gained during the rescattering process depends sensitively on the electric field near the tip surface. Thus the strength of the optical near-field is mapped to the kinetic energy of the emitted electrons. (c) Typical energy distribution of electrons emitted in the forward direction. The high-energy plateau ($\sim 5$ to $10\,\mathrm{eV}$) arises due to rescattering. Its cut-off is related to the local electric field amplitude at the metal. We obtain it from the intersection of two exponential fit functions (red lines). (d) Decay length $L$ as function of tip radius $R$ for tungsten tips ([ $\filledmedsquare$]{}) and gold tips ([ $\medbullet$]{}), deduced from simulations. The lines show linear fits: $L = (0.90 \pm 0.03) R$ for tungsten and $L = (0.82 \pm 0.04) R$ for gold. As the shape of the near-field mainly depends on the tip geometry, other materials behave very similarly.[]{data-label="nfrescat"}](arfig1.pdf){width="\columnwidth"} In general, three effects contribute to the enhancement of optical electric fields at structures that are smaller than the driving wavelength [@Martin2001; @Hartschuh2008; @Zhang2009; @Martin1997]. The first effect is geometric in nature, similar to the electrostatic lightning rod effect: the discontinuity of the electric field at the material boundary and the corresponding accumulation of surface charges lead to an enhanced near-field at any sharp protrusion or edge. This effect causes singularities in the electric field at ideal edges of perfect conductors. For real materials at optical frequencies, the electric field is not as strongly enhanced and remains finite [@VanBladel1996]. The second effect occurs at structures whose size is an odd multiple of half the driving wavelength: optical antenna resonances can be observed there. The third effect concerns only plasmonic materials like gold and silver, where an enhanced electric field can arise due to a localized surface plasmon resonance. Antenna and plasmon resonances depend critically on the shape and material of the structure in question and occur only for specific wavelengths. In contrast, geometric effects are inherently broadband and result in a monotonically increasing field enhancement for increasing sharpness of the structure and for increasing discontinuity in the dielectric constant at the boundary. In spite of their different nature and properties, all three effects can be modeled in the framework of Maxwell’s equations with linear optical materials. However, field enhancement calculations remain challenging because they crucially depend on the shape of the illuminated object, while analytic solutions of Maxwell’s equations are known only for a few special cases like spheres and infinite cylinders. Accurate field enhancement measurements are equally challenging because of the nanometric length scale and the often unknown exact shape of the structure. In this letter, we present experimental measurements with a new technique, the results of which we compare to numerical simulations of optical field enhancement at nanometric metal tips. Illuminating such a tip with light polarized parallel to the tip axis leads to the excitation of an enhanced near-field, which is spatially confined in all directions on the length scale of the tip radius [@Martin2001; @Zhang2009; @Novotny2006a] (see Fig. 1(a)). The near-field drives a localized source of electrons at the tip apex [@Hommelhoff2006; @Hommelhoff2006a; @Ropers2007]. Such photoemission experiments have found applications in a variety of different contexts aside from nanotips [@Petek1997; @Aeschlimann2007; @Dombi2013]. Very recently, it has been observed that part of the electrons can be driven back to the parent tip within a single cycle of the optical field. There, the electrons can scatter elastically and gain more energy in the optical field [@Kruger2011; @Yalunin2011; @Herink2012; @Wachter2012]. This process, well known from atomic physics [@Corkum1993; @Corkum2007; @Paulus1994a], has been called rescattering and leads to pronounced spectral features that are sensitive to the local electric field. Here we exploit the rescattering effect to probe the near-field in the immediate vicinity of the tip surface, as illustrated in Fig. \[nfrescat\](b). Our experiment consists of an almost atomically smooth metal tip with a radius of curvature $R = 8$ to $50\,\mathrm{nm}$. Its apex lies in the focal spot of few-cycle laser pulses derived from a Ti:sapphire oscillator (wavelength $\lambda = 800\,\mathrm{nm}$, repetition rate $f_\mathrm{rep} = 80\,\mathrm{MHz}$, pulse duration $\tau \approx 6\,\mathrm{fs}$). The setup is described in more detail in Ref. [@Schenk2010]. While this laser system reaches intensities of up to ${\sim}10^{12}\,\mathrm{W/cm}^2$ in the focus, we do not observe any influence of possible optical non-linearities on the field enhancement factor, and all our results (e.g., the linear dependence of the rescattering cut-off on laser intensity [@Wachter2012]) are consistent with a linear model of the metal’s optical response. ![Experimental results for the field enhancement factor of tungsten tips ([ $\medbullet$]{}) and gold tips ([ $\filledmedsquare$]{}) as a function of the tip radius. The uncertainty in $\xi$ represents an estimated systematic error due to the uncertainty in laser intensity. The lines are simulation results for $\lambda = 800\,\mathrm{nm}$ (W: solid blue line, Au: dashed red line, Ag: dash-dotted black line). The dielectric functions of the metals are taken from experimental data [@Lide2004] (see Fig. \[epsilon\](b)). For technical reasons related to mock surface plasmon reflection, we simulate gold and silver tips with a smaller opening angle than tungsten tips (W: $5\degree$, Au, Ag: $0\degree$). Simulations of tips with different angles show that this should not alter the results by more than $5\%$.[]{data-label="radiusexp"}](arradiusexp.pdf){width="\columnwidth"} Optical field enhancement enables us to observe electron rescattering at moderate pulse energies of less than $1\,\mathrm{nJ}$. We measure the energy distribution of the electrons emitted in the forward direction using a retarding field spectrometer. The recorded spectra yield information about the dynamics of the emitted electrons. A typical spectrum is shown in Fig. \[nfrescat\](c). At small energies, such spectra display an exponential decrease in count rate, followed by a plateau towards larger energies. The latter is an indication of electron rescattering [@Paulus1994a; @Becker2002; @Kruger2011; @Kruger2012; @Wachter2012]. This process has found utmost attention as it is at the core of attosecond science [@Corkum2007]. It has been shown that rescattering is highly sensitive to the peak electric field strength $E$ via the ponderomotive potential $U_{\mathrm{p}}$ the electron experiences in the light field [@Paulus1994a; @Becker2002; @Kruger2012]: $U_{\mathrm{p}} = e^2 \lambda^2 E^2 / (16 \pi^2 m c^2).$ Here, $\lambda$ is the laser wavelength, $e$ and $m$ are the electron’s charge and mass, and $c$ is the speed of light. The cut-off kinetic energy (see Fig. \[nfrescat\](c)) after rescattering is given by $T_{\mathrm{cutoff}} = 10.007 \,U_{\mathrm{p}} + 0.538 \,\Phi$, where $\Phi$ denotes the tip’s work function [@Busuladzic2006]. Measuring $T_{\mathrm{cutoff}}$ hence yields $U_{\mathrm{p}}$ [@FESupp2012]. Series of spectra for both tungsten and gold tips [@Eisele2011] with various tip radii yield the dependence of the field enhancement factor on tip radius and material. We extract the cut-off position of the rescattering plateau and deduce, via the above expressions, the peak electric field $E$. We stress that $E$, the field acting on the electron, is the [*enhanced*]{} field present at the tip’s surface. We thus obtain the field enhancement factor $\xi = E / E_{\mathrm{in}}$, with the laser field $E_{\mathrm{in}}$ deduced from intensity measurements. Figure \[radiusexp\] shows the field enhancement factor $\xi$ as a function of the tip radius $R$. For tungsten tips, $\xi$ grows by around a factor of $2$ with decreasing $R$, from $2.6 \pm 0.6$ at $(51 \pm 5)\,\mathrm{nm}$ to $5.7 \pm 0.6$ at $(8 \pm 2)\,\mathrm{nm}$. For gold nanotips with radii between $(46 \pm 3)\,\mathrm{nm}$ and $(28 \pm 4)\,\mathrm{nm}$, we obtain field enhancement factors between $\xi = 3.3 \pm 0.6$ and $3.5 \pm 0.6$. We have been unable to produce sharper gold tips with a well-controlled surface. Tip radii are determined in situ with the ring counting method in field ion microscopy or, for radii $> 20\,\mathrm{nm}$, using a scanning electron microscope [@FESupp2012]. We compare our results to fully independent simulations of field enhancement at tungsten, gold, and silver tips. They were performed using Lumerical (7.0.1), a commercial Maxwell solver implementing the finite-difference time-domain (FDTD) algorithm [@Taflove2005]. From each simulation, we extract the field enhancement factor by fitting a quadratic decay to the near-field at the moment of greatest enhancement and extrapolating the result to the tip surface. This and other measures are essential to obtain meaningful results, as the finite mesh of the FDTD solver, together with the different length scales involved, makes this problem a tricky one. Further details and simulation results will be published elsewhere. The results for tungsten, gold, and silver are shown in Fig. \[radiusexp\]. Experimental and simulation results agree well within the error bars. Note that this agreement is obtained without any free parameters. Both experiment and simulation show that $\xi$ increases smoothly for sharper tips, an indication of field enhancement due to a geometric effect. ![Field enhancement factor as a function of the dielectric constant $\epsilon = \epsilon_\mathrm{r} + i \epsilon_\mathrm{i}$ obtained from simulations. The simulation parameters are $R = 10\,\mathrm{nm}$, $\lambda = 800\,\mathrm{nm}$, and an opening angle of $0\degree$. The dielectric constants of tungsten, gold, and silver at $800\,\mathrm{nm}$ are plotted in the complex plane for comparison (solid lines) [@Haynes2011]. The right color scale applies to the dots, the bottom one to the lines. Note that the absolute value $|\epsilon|$ of the three materials is similar around $800\,\mathrm{nm}$.[]{data-label="epsilon"}](arepsilon10nm.pdf){width="\columnwidth"} Comparing our results to literature values of $\xi$, we find good agreement for tungsten tips (experiment [@Neacsu2005; @Yanagisawa2010], theory [@Martin2001; @Yanagisawa2010]), while previous results for gold tips are inconsistent (experiment [@Neacsu2005; @Ropers2007], theory [@Martin2001; @Bouhelier2003]) with some authors reporting much higher enhancement [@Bouhelier2003; @Neacsu2005; @Ropers2007]. A possible explanation for this disagreement is that the near-field at plasmonic materials like gold is exceptionally sensitive to the geometry of the tip (the opening angle in particular [@Martin2001; @Zhang2009]) and its surface condition. This is supported by the large variance in enhancement factors at gold tips reported in Ref. [@Neacsu2005]. Note also that far higher field enhancement factors are observed for tips in close vicinity ($ \lesssim R$) to surfaces [@Yang2009]. In our experiments with gold tips, we do not observe a large variance of field enhancement factors even though not all the tips had the ideal conical shape assumed in the simulations. A possible reason for this is that, before any measurement, we use field ion microscopy in conjunction with field evaporation to clean the tip surface and to ensure that the tip is almost ideally spherical in the vicinity of the apex [@FESupp2012]. This is likely not the case in many other experiments. Evidently, more research is needed to fully understand the behavior of tips made of plasmonic materials. Such an investigation is beyond the scope of this letter. In the following analysis, we only consider perfectly smooth, conical tips (see Fig. \[nfrescat\](a)) with small opening angles between $0\degree$ and $5\degree$. It appears, at first glance, surprising that the field enhancement factor of such different materials as tungsten and gold is so similar (see Fig. \[radiusexp\]), considering that gold supports the excitation of surface plasmons [@Novotny2006a; @Stockman2011]. We therefore analyze the dependence of the field enhancement factor on tip material in a series of simulations where we vary the complex dielectric constant $\epsilon = \epsilon_\mathrm{r} + i \epsilon_\mathrm{i}$ of the tip. This allows us to give a field enhancement map for all materials with $|\epsilon| \lesssim 40$, shown in Fig. \[epsilon\]. The results demonstrate that field enhancement occurs for any material with $\epsilon \neq 1$, even for pure dielectrics. Furthermore, the enhancement factor increases monotonically with the absolute value of the dielectric constant $|\epsilon|$, confirming that we observe field enhancement due to a geometric effect. We also note a slightly larger enhancement for $\epsilon_\mathrm{r} > 0$ than for $\epsilon_\mathrm{r} < 0$, which corresponds to an $\epsilon$-dependent phase shift ($< \pi$) of the near-field with respect to the driving field: the field enhancement factor is proportional to the maximum of the total electric field, which is reduced if the driving field and the near-field are out of phase. The $\epsilon$-dependent simulations reveal why the field enhancement factor of tungsten, gold, and silver tips is similar: they have a similar value of $|\epsilon|$ at $800\,\mathrm{nm}$. In order to obtain higher enhancement factors, materials with larger values of $|\epsilon|$ are required. For example, we find $\xi = 7.6$ for $R = 10\,\mathrm{nm}$ aluminum tips ($\epsilon_\mathrm{Al} = -64 + 47i$ at $\lambda = 800\,\mathrm{nm}$, beyond the range of our simulations in Fig. \[epsilon\]). Alternatively, $\xi$ can be increased by using longer wavelengths, because both the tip sharpness relative to the wavelength and the absolute dielectric constant $|\epsilon|$ of many materials increase for longer wavelengths. We expect considerably higher field enhancement at sharp metal tips for mid- and far-infrared or terahertz radiation. An enhancement factor $\xi \approx 25$ has already been reported in SNOM experiments with terahertz radiation [@Huber2008]. In contrast to the increase with $|\epsilon|$, there is one point in Fig. \[epsilon\] close to $\epsilon = -2$ that shows a significantly higher enhancement than the points surrounding it. This can be interpreted as a localized plasmon resonance, similar to what is known from nanospheres [@Martin2001]. It can be observed with a wavelength of $\lambda \approx 360\,\mathrm{nm}$ at silver tips [@Zhang2009] or $\lambda \approx 520\,\mathrm{nm}$ at gold tips. In the analysis of our experimental results, we have neglected the spatial variation of the near-field on the rescattered (field-probing) electrons’ path, assuming instead a constant electric field. This is justified as the decay length of the near-field $L$ (see Fig. \[nfrescat\](d)) is much longer than the maximum extension of the electron’s path $M$: For our parameters, both classical [@Kruger2011; @Kruger2012] and quantum mechanical [@Wachter2012; @Yalunin2013] simulations indicate that the electrons’ path extends approximately $1\,\mathrm{nm}$ from the surface before rescattering. On this scale, the sharpest tips we investigate show a near-field variation of ${\sim} 20\%$. Including this spatial variation into classical calculations of rescattering changes the enhancement factor by $0.4$ only, even for the sharpest tip in our experiments. This is less than the measurement uncertainty. For longer wavelengths or higher field strengths, the maximum extension $M$ increases. In this case, the effect may be more significant and can even suppress rescattering completely [@Herink2012]. One intriguing application of our method is the investigation of quantum effects in nanoplasmonics, a new field that has recently been named quantum plasmonics. Self-consistent quantum mechanical calculations of small nanoparticles (radius of curvature $< 2\,\mathrm{nm}$) illuminated by laser pulses show that the excited surface charge density, the root cause of the optical near-field, extends over several [å]{}ngstr[" o]{}ms beyond the surface [@Zuloaga2010]. This “electron spill-out” reduces the strength of the near-field by up to ${\sim} 50\%$. For small nanoparticles, it was shown that these effects are noticeable only at a distance of $Q < 0.5\,\mathrm{nm}$ from the surface, while the near-field retains its classical shape for larger distances. As fully quantum mechanical calculations of larger nanoparticles remain difficult (although large steps are being made in this direction [@Teperik2013]), it is unclear if the length scale of nonclassical behavior $Q$ depends on the size of the nanoparticle. The authors of Ref. [@Zuloaga2010] suspect $Q$ to increase for larger nanoparticles. While a fully integrated quantum calculation of both field enhancement and electron rescattering is beyond the scope of this letter, we will discuss the effects of quantum plasmonics on rescattering qualitatively. They depend on three length scales: the extent of nonclassical field reduction $Q$, the near-field’s decay length $L$, and the rescattered electron’s maximum extension $M$. If $Q \approx L$ as in Ref. [@Zuloaga2010], the maximum of the near-field is significantly reduced, which implies a corresponding reduction of the cut-off energy. Extremely sharp nanostructures ($R \lesssim 3\,\mathrm{nm}$) will be required to reach this regime if $Q$ does not scale with structure size. As discussed earlier, rescattering may be suppressed in this case, depending on the relation of $M$ and $L$ [@Herink2012]. If $L \gg Q$, only a small fraction of the near-field’s extent is reduced in strength so that the maximum of the near-field is almost unchanged. In this case, quantum effects are only noticeable if $M \le Q$, because the rescattered electron would not be sensitive to the reduced field strength otherwise. The parameters in our experiments are $M \approx 1\,\mathrm{nm}$ (including a non-zero tunneling distance [@Busuladzic2006; @Kruger2012; @Hickstein2012]) and $L \ge 8\,\mathrm{nm}$, so quantum effects should only be visible if $Q$ becomes larger than $0.5\,\mathrm{nm}$ for larger nanostructures. The agreement between experimental results and classical theory seems to suggest that $M > Q$. Hence, Q does not seem to scale with structure size, as hypothesized in Ref. [@Zuloaga2010]. However, there is still the possibility of quantum plasmonic effects on a larger scale within the error bars of our results. An increase of $Q$ for larger tips might explain the steeper decrease of $\xi$ for larger radii we observe in the experiment as compared to the simulation (see Fig. \[radiusexp\]). In conclusion, we demonstrate a new method of probing optical near-fields within $1\,\mathrm{nm}$ distance from the surface of a nanoscale metal tip. The method is based on rescattering of electrons driven by short laser pulses. The length scale on which the near-field is measured reaches down to dimensions that are of utmost interest in the emerging field of quantum plasmonics. Experimental results for the field enhancement factors of tungsten and gold tips agree well with Maxwell simulations. Based on these results, we give a field enhancement map for a wide range of materials. Furthermore, the simulations reveal that geometric effects are the predominant mechanism of optical field enhancement at nanotips in most cases. Exceptions exist close to plasmon resonances. In the future, a tomographic reconstruction of the near-field, likely in three dimensions, will be possible by measuring the cut-off energy of the rescattered electrons while varying the laser power or wavelength. We would like to thank Peter Nordlander, Markus Raschke, and Hirofumi Yanagisawa for insightful discussions as well as Philipp Altpeter, Jakob Hammer, and Sebastian Stapfner for assistance with scanning electron microscope imaging. [49]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ** (, ). , ****, (). , ****, (). , ****, (). , , , ****, (). , , , , , , , ****, (). , ****, (). , , , , , , ****, (). , , , , ****, (). , ****, (). , , , ****, (). , , , , , ****, (). , , , , , , , , , ****, (). , , , , , , ****, (). , , , , ****, (). , ****, (). , , , , , ****, (). , ****, (), ISSN . , , , , , , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , **, IEEE Press Series on Electromagnetic Wave Theory (, ). , , , , ****, (). , , , ****, (). , , , , , ****, (). , ****, (). , , , , , , , , , ****, (). , , , , , , , ****, (). , , , ****, (). , , , ****, (). , , , , ****, (). , ****, (). , , , , , ****, (). , , , ****, (). , , , , , , in ** (, ), vol. , pp. . , , , , ****, (). , , , ****, (). , , , , , ****, (). , **, CRC Handbook of Chemistry and Physics, 85th Ed (, ). , ** (, ). , , , ****, (). , , , , ****, (). , , , ****, (), ISSN . , , , , , ****, (). , ** (, ). , , , , , , , , ****, (). , , , , , , , , , , , ****, (). [^1]: These authors contributed equally to this work. [^2]: These authors contributed equally to this work.

ArXiv

--- abstract: 'The progress in autonomous driving is also due to the increased availability of vast amounts of training data for the underlying machine learning approaches. Machine learning systems are generally known to lack robustness, e.g., if the training data did rarely or not at all cover critical situations. The challenging task of corner case detection in video, which is also somehow related to unusual event or anomaly detection, aims at detecting these unusual situations, which could become critical, and to communicate this to the autonomous driving system (online use case). Such a system, however, could be also used in offline mode to screen vast amounts of data and select only the relevant situations for storing and (re)training machine learning algorithms. So far, the approaches for corner case detection have been limited to videos recorded from a fixed camera, mostly for security surveillance. In this paper, we provide a formal definition of a corner case and propose a system framework for both the online and the offline use case that can handle video signals from front cameras of a naturally moving vehicle and can output a corner case score.' author: - 'Jan-Aike Bolte$^{\ast}$, Andreas Bär$^{\ast}$, Daniel Lipinski$^{\circ}$ and Tim Fingscheidt$^{\ast}$ [^1] [^2]' bibliography: - 'bolte\_ownbib.bib' title: '**Towards Corner Case Detection for Autonomous Driving** ' --- INTRODUCTION ============ The recent developments in machine learning also led to significant advancements in current autonomous driving systems. These systems more and more rely on deep learning techniques that use huge datasets for training, e.g., learning from this data how to behave in certain situations. The use of *black-box* deep learning systems poses a risk, which became apparent through the latest real-world accidents with autonomous cars being involved [@NTSB2018; @NTSB2018a]. Such accidents may occur, if the training data did rarely or not at all cover certain situations [@Tian2018], a typical case, when issues due to machine learning can be expected [@Koopman2016]. The goal of a corner case detection system is to detect unusual situations[^3] either in this training data, or, in a second step, online in an autonomous vehicle. The video signal from a monocular vehicle camera represents a part of the situation. For humans it is possible to distinguish normal from unusual events, even when they only have available the video from a mono vehicle camera. Therefore, it should also be possible to design a corner case detection system that can evaluate situations based on video data only.\ Corner case detection brings advantages both in offline as well as in online systems. In an online approach the corner case detection system can be employed as a redundant warning function accompanying an autonomous driving system, where it provides information about how unusual the current situation is (see Fig. 1 (a)). In an offline approach, the corner case detection system can be used to parse through vast amounts of collected video data and returns only a user-defined amount of unusual data. This can be used as a data selection procedure for large-scale data recordings, where it is undesirable to store too much redundant or irrelevant data. These selected corner cases can then be used for a more focused training of autonomous driving systems, tackling the problem of underrepresented critical training data, e.g., by oversampling the corner cases [@Chawla2002].\ [0.9]{}  [0.9]{}  A major challenge in the development of a corner case detection system is that the detection of corner cases is an ill-defined problem, since there is no universally accepted definition of what a corner case actually is. Corner case detection is closely related to anomaly detection [@Chandola2009] and novelty detection [@Pimentel2014], and these two disciplines are already very close to each other. Anomaly detection typically refers to the detection of samples during inference that do not conform with an expected normal behavior. To figure out what is abnormal, a world model can be trained on normal data so that any deviation from that learned behavior will be marked as an anomaly. Novelty detection is similar in the sense that a world model is trained on normal data. Any data that deviates from the already seen data is marked as a novelty. These novelties can be either anomalies or just samples that were not included in the training data. If enough diverse training data is available to train a world model that represents normality sufficiently well, any novelty will be an anomaly. If not, any anomaly is a novelty, but not any novelty is necessarily an anomaly.\ The core idea of our work is that the online corner case detection provides self-awareness and a criticality measure for a perception module by an advantageous combination of a semantic segmentation with a video-based prediction error, under the hypothesis that *non-predictive relevant objects in vicinity* to the car’s expected future trajectory is a corner case. Our contribution with this work is threefold, as we propose a formal definition of what a corner case in the context of perception in autonomous driving might be, secondly, we point out domain-specific challenges regarding corner case or anomaly detection in videos from car-mounted cameras that record scenes with a highly dynamic content and challenging ego motion. Thirdly, we propose a technical framework that will allow to develop corner case detection systems.\ The paper is structured as follows. In Section II we present related work in the field of anomaly detection, image prediction, and semantic segmentation. In Section III we propose a definition of a corner case. In Section IV we introduce the corner case detection system concept along with employed datasets and measures. Finally we will present the results of some first experiments in Section V, before we provide conclusions in Section VI. RELATED WORK ============ The corner case detection system that is proposed in this work has a modular structure, consisting of subsystems, which can be (partly) adopted from other research fields. As mentioned in Section 1, we employ an image prediction method in combination with a semantic segmentation to tackle the problem of corner case detection, being closely related to anomaly detection. Anomaly Detection ----------------- Literature offers a wide variety of approaches for the detection of unusual events. Most of them fall under the terms of anomaly detection or novelty detection [@Hodge2004; @Chandola2009; @Pimentel2014]. Driven by the rapid development in the field of deep learning, many recent methods follow the *reconstruction-based* approach by *unsupervised* training of neural networks, exploiting the important advantage that vast mounts of unlabeled data can be used in training [@Thompson2002; @Singh2004; @Mahadevan2010; @Hasan2016; @Carrera2017; @Chong2017; @Utkin2017; @Munawar2017; @Liu2018]. The unsupervised training is typically applied to datasets that contain a large number of normal samples and a negligible small amount of abnormal samples. We assume that we are given a sequence $\mathbf{x}_1^T = \left(\mathbf{x}_{t=1},\mathbf{x}_{t=2},...,\mathbf{x}_{t=T} \right) $, where $\mathbf{x}_t$ denotes the frame at the discrete frame index $t$, and $T$ is the length of the sequence. In this case the reconstruction-based method learns a model of normality $\mathcal{M}_{\text{normal}}$, given the training dataset $\mathcal{X}_{\text{train}}$, consisting of several sequences. The trained model is used to assign a novelty or abnormality score to the test dataset $\mathcal{X}_{\text{test}}$. Various forms of autoencoder (AE) networks can be used to train such a model.\ There are also some special approaches to anomaly detection in the context of autonomous driving or unmanned vehicles. Lin et al. [@Lin2010] used the Mahalanobis distance to measure the distance of mutliple sensor data vectors, thereby identifying the unusual events. However, their approach was not image-based and thus is not suitable for our work. Another approach is to employ particle filtering and maximum likelihood methods for anomaly detection [@Lampiri2017], but again this approach is not suitable for image-based anomaly detection.\ As a more general approach to anomaly detection, Munawar et al. [@Munawar2017] trained a spatio-temporal anomaly detection system for surveillance of industrial robots using a deep convolutional neural network (DCNN). They use a biologically plausible system for anomaly detection according to Egner et al. [@Egner2010], where an anomaly is determined on the basis of expectation and surprise. *Accordingly, an unusual video frame can be identified by its deviation from the predicted frame.* Liu et al. [@Liu2018] used a future frame prediction with spatial and motion constraints based on generative adversarial networks (GANs).\ Encouraged by these results, we will choose a simple reconstruction-based approach but will employ a predictively trained model of normality. Different to all image-based approaches mentioned above that were evaluated on datasets with a stationary camera, the ego motion of the vehicle makes the datasets in autonomous driving research much more challenging. We therefore will now consider various image prediction approaches for automotive datasets. Image Prediction ---------------- The task of predicting future frames in videos with parametric models that were trained in an unsupervised manner was rarely approached before the work of Ranzato et al. [@Ranzato2014]. Inspired by language modeling they introduced a baseline approach for extrapolating future frames that used features learned from video signals in an unsupervised fashion. It seemed to have been the first model that was able to generate realistic predictions of video sequences. Srivastava et al. [@Srivastava2015] introduced a long short-term memory autoencoder (LSTM-AE) that performed predictions on image patches, or so-called percepts, that are extracted by a preceding convolutional neural network (CNN). They showed that the predictive training even improved results in a classification task for action recognition. Mathieu et al. [@Mathieu2015] used a multi-scale architecture and an objective function that incorporated an adversarial loss and the differences in image gradients. Pătrăucean et al. [@Patraucean2015] used a convolutional LSTM-AE. They designed a spatio-temporal video autoencoder to emulate the human visual short-term memory in a basic form. The first real model for long-term prediction was introduced by Lu et al. [@Lu2017]. They designed a novel objective function and an autoencoder structure with LSTMs. However, again, all predictive approaches mentioned before have in common that they only evaluate their models on datasets with stationary cameras. In the context of automotive image prediction the well-known prediction network `PredNet` [@Lotter2016] can capture key aspects of both movement of the ego vehicle and movement of the objects in the visual sequences.\ In our experiments, we will evaluate `PredNet` and a predictive autoencoder that is based on the network proposed in [@Hasan2016], since the authors already used it for anomaly detection with a reconstruction-based approach. We will adopt the network and will propose an adversarially trained predictive approach for corner case detection. Semantic Segmentation --------------------- The aim of semantic segmentation is the semantic labeling of each pixel of an input image. Current state-of-the-art architectures for semantic segmentation rely on the concept of fully convolutional networks (FCNs), introduced by [@Long2015]. Here, a classifier, pre-trained on the ImageNet database [@Russakovsky2015], is modified for semantic segmentation. Typical classifiers in state-of-the-art models for semantic segmentation [@Zhao2016a; @Chen2017; @Wu2016; @Chen2018a; @RotaBulo2018] are residual networks (`ResNets`) [@He2016]. The appropriate amount of context is crucial for semantic segmentation. One common way of dealing with this problem is the use of dilated convolution [@Yu2016] to enlarge the receptive field in deeper layers [@Zhao2016a; @Chen2017; @Wu2016; @Chen2018a; @RotaBulo2018]. Further multi-scale context is addressed by some state-of-the-art architectures [@Zhao2016a; @Chen2017; @Chen2018a; @Chen2018; @RotaBulo2018] via combining the feature extractor with a pyramid pooling module. To restore the original resolution, all models are designed in a encoder-decoder fashion. One simple approach is to use bilinear interpolation of the network prediction as the decoder part [@Zhao2016a; @Chen2017; @Chen2018; @RotaBulo2018], while other approaches use more complex operations such as transposed convolution [@Wu2016; @Romera2018], or the additional use of low-level features through skip connections [@Chen2018a; @Ronneberger2015].\ Recently, along with a major interest in practical implementations, efficient semantic segmentation architectures with regard to computation and memory cost have been introduced. To address the memory problem, [@RotaBulo2018] proposes in-place activated batch normalization (`InPlace-ABN`), a memory-efficient approach in the training process through combining the leaky rectified linear unit (leaky ReLU) with batch normalization [@Ioffe2015]. In [@Chen2018a; @Sandler2018] depthwise separable convolutions are used to reduce the number of parameters and therefore computation cost and memory usage, while [@Romera2018] proposes factorized convolution in combination with residual connections to obtain a similar effect.\ We base our own segmentation network on the `DeepLabv3` [@Chen2017] with some improvements, since it is one of the best performing networks on the Cityscapes dataset. CORNER CASE DEFINITION ====================== A consequence of the so-far missing universally accepted definition of a corner case is that there is also no explicit metric existing. Motivated by [@Munawar2017], we will use a predictive approach for corner case detection. The idea is that if a novel or abnormal or critical suddenly occurring situation is technically predictable, it will not pose a major problem to an autonomous driving system, which will then naturally take care of adequate actions. Therefore, for us this would be a *don’t care* situation, although some might view it as a corner case. So our focus is on *technically unpredictable* situations. However, it is important to note that not each unpredictable situation in the field of autonomous driving is necessarily a corner case. An aircraft that suddenly enters the camera image in the sky may not be predictable, but luckily in most cases it will be irrelevant for the driving task. As opposed to that, pedestrians, cyclists and other moving objects on the ground are highly relevant classes. Beyond that, even a pedestrian acting in some highly unpredictable manner may be irrelevant for the driving task if it happens at a sufficiently large distance from the vehicle or its future trajectory.\ We therefore propose that a corner case is given if there is a *relevant* object (class) in *relevant* location that a modern autonomous driving system *cannot predict*. Therefore, the relevance of a corner case results from the three aspects noted in Figure 2. The relevance of objects and locations, as well as aspects of prediction, will be discussed in the next section. A corner case is given, if there is a $\underbrace{\text{\emph{non-predictable}}}_{1} \underbrace{\vphantom{p}\text{\emph{relevant object/class}}}_{2} \text{in} \underbrace{\vphantom{p}\text{\emph{relevant location}.}}_{3}$\ TOWARDS CORNER CASE DETECTION: FRAMEWORK AND MEASURES ===================================================== As a working hypothesis we assume that it is possible to detect corner cases with a camera-based system, because, as already mentioned, it is also possible for a human to distinguish normal from unusual events, even when only the video from a mono vehicle camera is available. We limit this work to the detection of non-predictive situations in the context of autonomous driving, meaning a camera in movement. Following our definition of a corner case we need a system that combines the three important aspects (Fig. 2): (1) First, we need an image prediction that gives us the prediction errors for each new frame. (2) Second, we need a semantic segmentation of the input frame that allows us to classify and localize the objects in the scene, with moving objects being considered as *relevant* (see Table 1), and (3) third, we need a detection system that processes the information from both image prediction and semantic segmentation by information fusion, comprising a check, whether the non-predictable (e.g., jumping) relevant class (e.g., pedestrian) is in a *relevant location* (will cross trajectory). The following subsections describe each part of the corner case detector that is illustrated in Figure 3, along with the datasets we employ for training, and related measures. Semantic Segmentation --------------------- Semantic segmentation of images aims at finding a transformation that partitions the input image into semantically related parts, by assigning each pixel to a specific class. As motivated in Section II, we adopt a segmentation network based on the `DeepLabv3` [@Chen2017] with some improvements from `WideResNet38` [@Wu2016], which is pre-trained on the ImageNet corpus. To be more specific, we replace `ResNet50` [@He2016] inside `DeepLabv3` by `WideResNet38`. This is similiar to the approach in [@RotaBulo2018], with the difference, that we don’t incorporate the proposed `InPlace-ABN`. Further, we do a few common modifications in semantic segmentation to `WideResNet38` [@Chen2017; @Wu2016]: First, we remove the classification layer of `WideResNet38` and connect the remaining network to the segmentation head of `DeepLabv3`. Second, to control the output stride (ratio between input resolution and output resolution) we decrease the stride of several convolutions from two to one in a bottom-up fashion and increase the dilation rate instead. Third, in contrast to [@Wu2016], we do not incorporate dropout in our segmentation framework as we observed lower performance results otherwise.\ The input image $\mathbf{x}_t \in \mathbb{G}^{H \times W \times C}$ with image pixel $x_t(i) \in \mathbb{G}$, where $\mathbb{G}$ is the set of gray values, $H$ and $W$ are the image height and width in pixels and $C=3$ is the number of color channels from set $\mathcal{C}=\{1,2,3\}$, is fed into a fully-convolutional neural network. It maps the input to output scores $\mathbf{P}_t \in \mathbb{I}^{H \times W \times \vert \mathcal{S} \vert }$, where $\mathcal{S}$ denotes the set of classes with cardinality $|\mathcal{S}|=19$ and $\mathbb{I}= [0,1]$. For each pixel position $i \in \mathcal{I}$, the third dimension of the output scores provides a posterior probability (score) $P_t(i,s)$ for each class $s \in \mathcal{S}$. Here, $\mathcal{I}$ is the set of pixel indices in the image, and $|\mathcal{I}| = H \cdot W$ the number of pixels. Taking the argmax over the output scores we obtain the $(H \times W)\text{-dimensional}$ mask $ \mathbf{m}_t = \operatorname*{argmax}_{s\in\mathcal{S}} \mathbf{P}_t $, which gives us a pixel-wise classification $m_t(i)\in\mathcal{S}$ of the frame for time $t$.\ \[tab:labels\] As a metric, the mean intersection over union (mIoU) is employed [@Everingham2015], which measures the accuracy of the segmentation mask and is defined as the mean of the frame-wise $$\text{IoU}_t = \frac{\text{TP}_t}{\text{TP}_t+\text{FP}_t+\text{FN}_t} \, ,$$ where $\text{TP}_t\text{, FP}_t\text{, and FN}_t$ are the numbers of true positive, false positive, and false negative pixels, respectively, in frame $t$.\ Modern neural networks have been shown to be overconfident in their classifications [@Gao2017]. This can pose a problem on system components used for semantic segmentation or object detection, since even unknown objects will be classified as one of the known classes. In a future corner case detection system this will be solved by a single-frame anomaly detection. Image Prediction ---------------- ![High-level block diagram of the corner case detector.[]{data-label="fig:short"}](figure2.pdf) As already mentioned, image prediction is an essential part of the corner case detection system. Modern autonomous driving systems already predict trajectories of other traffic participants. To identify corner cases in video streams, it is essential to understand the underlying states and dynamics within the given situations. This high-level abstraction can be learned by predictive models. For the image prediction approach we train a model that receives $n$ consecutive frames $\mathbf{x}_{t-n}^{t-1}= (\mathbf{x}_{t-n}, \mathbf{x}_{t-n+1},...,\mathbf{x}_{t-1})$ to compute a prediction $\hat{\mathbf{x}}_t$ of the current frame. As a metric for the corner case, we now calculate an error $$\mathbf{e}_t = \hat{\mathbf{x}}_{t} - \mathbf{x}_t,$$ between the predicted image and the actual image $\mathbf{x}_t$ (subtraction symbol in Figure 2), with elements $e_t(i), i \in \mathcal{I}$. Following Mathieu et al. [@Mathieu2015], we use the following metrics to evaluate the prediction models. The mean-squared error (MSE) distance is then given by $$\begin{aligned} D_\text{MSE}(\hat{\mathbf{x}}_{t}, \mathbf{x}_{t}) &=\frac{1}{|\mathcal{I}|} \sum_{i \in \mathcal{I}}e_t^2(i).$$ Additionally we employ the peak signal-to-noise ratio (PSNR), which is defined by $$\text{PSNR}(\hat{\mathbf{x}}_t, \mathbf{x}_t) = 10\log_{10} \frac{\rm{max}^2_{\hat{\mathbf{x}}}}{D_\text{MSE}(\hat{\mathbf{x}}_t, \mathbf{x}_t)},$$ where $\rm{max}^2_{\hat{\mathbf{x}}}$ shall denote the squared maximum possible value of the color channel image intensities, e.g., $255^2$ for 8-bit image formats with $\mathbb{G}=\{0,1,...,255\}$. Both metrics can also be applied to color images, where the metric is evaluated for each color channel separately, and then averaged. As a third metric, we use the structural similarity index measure (SSIM), which is a metric for perceived image quality being introduced by Wang et al. [@Wang2004]. The SSIM measures the perceptual difference between the original and the predicted image and is, unlike the other two metrics, based on visible structures in the image. Detection System ---------------- In the detection system, information from both of the two previous processing steps is combined. As its output the system generates a corner case score $\epsilon_t \in \left[0,1 \right] $ for each input frame $\mathbf{x}_t$, exploiting also past frames $\mathbf{x}_{t-n}^{t-1}$. Along with $\epsilon_t$, in principle also a localization of the corner case in the image can be performed. If we recall the definition of the corner case in Figure 1, we remember that we consider a corner case consisting of the logical AND combination of three aspects. The semantic segmentation provides the class and location information and the image prediction provides the *predictability*. To identify a *relevant location*, in a real system, typically one would use a perception approach based on a light detection and ranging (LIDAR) sensor to assign depth information to the image pixels [@Zhang2015; @Wei2018]. On the basis of that, a time to collision on a pixel basis can be estimated [@Bosnak2017]. For the purpose of this work, however, instead we adopt a rather simple approach for the reason of conciseness of presentation and evaluation. We simply assume that objects being further above the bottomline of the image are more distant to the ego vehicle.\ The error map (2) from the image prediction gives us a value of non-predictability for each pixel. Since we are for now only interested in the moving classes, we simply set the error of those pixels that do not belong to one of these moving classes to zero, given the following formula: $$\label{eq:comb_seg_pred} {e}_{t,\mathrm{rel}}(i) = \begin{cases} {e}_t(i), & m_t(i)\in\mathcal{S}_{\mathrm{rel}} \\ 0, & m_t(i)\notin\mathcal{S}_{\mathrm{rel}} \end{cases}$$ where $\mathcal{S}_{\mathrm{rel}}$ denotes the set of all relevant classes for the corner case detection. In our case, we use the eight classes that are printed in bold in Table I. The squared errors of the relevant classes ${e}_{t,\mathrm{rel}}^2(i)$ are then weighted depending on their distance from the bottom of the image and summed up resulting in an error score $$\epsilon_t^{\prime} = \sum_{i \in \mathcal{I}} {e}_{t,\mathrm{rel}}^2(i) \cdot (1-\frac{h_i}{H-1}),$$ with $h_i \in \{ 0,1,...,H\! - \!1 \}$ being the row index (bottom-up) of pixel $i$. Thereby, our simple definition of a relevant location weights the bottom row squared errors by a one, and the top row squared errors of the relevant classes by a zero. Finally, the corner case score is obtained by normalizing the error score $\epsilon_t^{\prime}$ to a value range from 0 to 1 using $$\epsilon_t = \frac{\epsilon_t^{\prime}-\min\limits_{\tau \in \mathcal{T}} \epsilon_{\tau}^{\prime}}{\max\limits_{\tau \in \mathcal{T}} \epsilon_{\tau}^{\prime}-\min\limits_{\tau \in \mathcal{T}} \epsilon_{\tau}^{\prime}} ,$$ where $\mathcal{T}$ denotes a set of time instants. For the online approach it may be $\mathcal{T}=\{1,2,...,t\}$, or all time instants of the validation data. For the offline approach $\mathcal{T}$ may contain all time instants of the video material currently being analyzed. If a *localization* of the corner case in the frame is needed, $\epsilon_t^{\prime}$ is obtained by (6) with summation over small patches $i \in \mathcal{I}_p \subset \mathcal{I}$ of the window. We therefore subdivide the image into patches of the same size, e.g., $32\! \times \! 32$ pixels. The patch size can be adjusted by the user, depending on how exactly he desires to localize the corner cases. The error scores of each of the patches can then also be normalized to a range between 0 and 1.\ Finally, the obtained corner case score is subject to thresholding. An appropriate threshold value $0<\theta_\epsilon<1$ has to be identified on validation data to tune the desired behavior of the detector regarding the false acceptance rate and the false rejection rate. In an offline system, $\theta_\epsilon$ can be chosen by the user in order to control the amount of detected corner cases in the data.\ DATASET, TRAINING, AND QUALITATIVE EVALUATION OF THE CORNER CASE DETECTION SYSTEM FRAMEWORK =========================================================================================== Dataset ------- We train both the segmentation and image prediction network on the Cityscapes dataset [@Cordts2016] that contains a diverse set of street scene images recorded in 50 different cities, being a widely used benchmark for semantic segmentation not only for autonomous driving research. The dataset is labeled with 19 classes that are used during training and inference (see Table 1). They denote the set $\mathcal{S}$. The Cityscapes dataset [@Cordts2016] offers a benchmark suite that serves as a baseline for future improvements in image segmentation. For the purpose of this conceptual paper, the image prediction is trained on the three demo videos provided by the dataset. Semantic Segmentation --------------------- We mainly adopted the training protocol from Chen et al. [@Chen2017]. For optimization, we used the stochastic gradient descent with momentum $\beta=0.9$ and a learning rate with polynomial decay: $$\label{eq:lr_schedule} \eta\left( k\right) = \eta_{\mathrm{start}} \cdot \left( 1-\dfrac{k}{k_{\mathrm{max}}}\right)^{\gamma}$$ where $\eta\left(k\right)$ is the current learning rate at iteration $k$, $\eta_{\mathrm{start}}$ is the initial learning rate, $k_{\mathrm{max}}$ the maximum number of iterations, and exponent $\gamma=0.9$. For data augmentation we perform random resizing in the range $\left[ 0.5, 2.0\right]$, left-right flipping, and cropping of the input image with a crop size of 700x700. With this configuration and our reimplementations we were able to fit a batch of size $b_1=4$ and $b_2=2$ using an output stride of $o_1=16$ and $o_2=8$, respectively, on an `Nvidia Geforce GTX 1080 Ti`. The training itself is organized in a two-stage fashion. In the first stage, we set the output stride to $o_1=16$ and train the network parameters, including the batch statistics, for 90,000 iterations with a batch of size $b_1=4$ and an initial learning rate $\eta_{\mathrm{start}}=0.001$. In the second stage, we set the output stride to $o_2=8$, freeze the batch statistics in the corresponding layers, and fine tune for an additional 120,000 iterations with a reduced batch of size $b_2=2$ and a reduced initial learning rate $\eta_{\mathrm{start}}=0.0005$. We evaluate our segmentation results by using the fine-tuned stage-two model and computing the mean intersection over union (mIoU, see (1)) between the network prediction and the ground truth with an output stride of $o=8$. We use multiple parallel fine-tuned models, which are fed by input images with three different scales $q \in \mathcal{Q}=\lbrace0.75, 1.0, 1.25 \rbrace$, resulting in score maps for each class $s \in \mathcal{S}$ and scale $q \in \mathcal{Q}$. These score maps are resized to the original input size and are fused by summation for each class $s \in \mathcal{S}$ separately across all scales $q \in \mathcal{Q}$. Our network was solely trained on the finely annotated training images from the Cityscapes dataset. It achieves a competitive mIoU of 78.4% on the Cityscapes evaluation set, being quite close to the best approaches known today. Image Prediction ---------------- For the image prediction task we tested two models to find out which one works best for the difficult conditions in automotive applications. The first model is the well-known `PredNet` [@Lotter2016] and the second is an autoencoder network that is based on the network proposed by Hasan et al. [@Hasan2016].\ To gain a quick first insight into the employment of image prediction the three demo videos provided by the Cityscapes dataset were used as training data for both prediction networks. ![Architecture of the **image prediction** by an autoencoder. A convolution, maxpooling, or upsampling block with parameters ($K,S$) uses a filter size of $K \times K$ and a stride of $S$. The annotations at the edges indicate the dimensions of the transferred tensors.[]{data-label="fig:short"}](figure3.pdf) Both networks were trained on the three demo videos provided by the Cityscapes dataset. The networks were trained in a leave-one-out scheme, where they were trained on two of the three videos and the test was performed on the third, unseen video. The training and testing was done for each combination of the three videos and the test results on the respective unseen video were averaged. For the exemplary test we limit ourself to the demo videos of the Cityscapes dataset, since our segmentation network was trained on this dataset and therefore we can definitely expect good performance on images from this domain.\ For the `PredNet`, we used the standard architecture and training protocol as described by Lotter et al. [@Lotter2016]. For the predictive AE we adopted the architecture proposed by Hasan et al. [@Hasan2016] with some improvements to refine the predictions. The architecture is shown in Figure 4. Typically a normal AE is trained using the MSE loss. It was shown that this leads to blurry predictions [@Mathieu2015; @Lotter2016a], which can be overcome by incorporating an adversarial loss. We added a discriminator network to the training procedure. The discriminator network uses the same architecture as the encoder network of the AE, extended with a patch-wise classification (also known as local adversarial loss) as proposed by [@Shrivastava2017], where the discriminator network is trained to classify the real images or predicted images of the image prediction network w.r.t. the respective classes, with $s^{\mathrm{(D)}} \in \mathcal{S}^{\mathrm{(D)}} = \left\lbrace s_{\mathrm{real}},s_{\mathrm{predicted}} \right\rbrace$ being the class upon which the discriminator decides. The architecture of the discriminator network is shown in Fig. 5. ![Architecture of the **discriminator** used for the adversarial loss of the **prediction training**. A convolution or maxpooling block with parameters ($K,S$) uses a filter size of $K \times K$ and a stride of $S$. The output class map $\mathbf{s}_t^{\mathrm{(D)}}$ of size $\frac{H}{8} \!\times\! \frac{W}{8}$ provides a class $s^{\mathrm{(D)}}$ for $8\!\times\!8$ pixel squares of the image.[]{data-label="fig:short"}](figure4.pdf)  The loss of the generator and of the discriminator are added in the loss function $$\label{eq:cost_func} J_{\mathrm{G}}(\mathcal{A}_{\mathrm{G}},\mathcal{A}_{\mathrm{D}})=J_{\mathrm{MSE}}(\mathcal{A}_{\mathrm{G}})+\lambda J_{\mathrm{ADV}}(\mathcal{A}_{\mathrm{D}},\mathcal{A}_{\mathrm{G}}),$$ with $J_{\mathrm{MSE}}$ being the standard MSE loss, $J_{\mathrm{ADV}}$ being the adversarial loss, and the weighting factor $\lambda = 0.25$ determining the influence of the adversarial loss on the overall loss. $\mathcal{A}_{\mathrm{G}}$ and $\mathcal{A}_{\mathrm{D}}$ denote the weights of the generator network, in this case the image prediction network, and the discriminator, respectively.\ We found a major improvement in stability of the training by using a cyclic learning rate as proposed by [@Smith2017]. We used the stochastic gradient descent and started the training with a learning rate $\eta_{\mathrm{start, G}}=1\cdot10^{-7}$ of the generator and then periodically increased the learning rate of the generator to $\eta_{\mathrm{max, G}}=7\cdot10^{-7}$ and decreased it back again. The period $\phi = 20 \text{ epochs}$ showed the best results. The learning rate of the discriminator did not follow the cyclic protocol and was set to $\eta_{\mathrm{D}}=1\cdot10^{-6}$. The maximum learning rate is also subject to an exponential decay after each epoch with an exponent of $\gamma=0.85$. The batch size for the training is set to $b=3$. For the training, the original Cityscapes images were downsampled to a resolution of $H = 256 \text{ and } W = 512$ pixels and converted to greyscale.\ The results in Table II show that the quality of the predictions from both networks are not far apart from each other for both PSNR and SSIM. In our experiment, however, the predictive autoencoder provides clearly a better MSE. We decide to use the predictive adversarial AE, since it is a simple feed-forward architecture that can be easily modified for further experiments and also provides the better MSE, which is closely related to the measure that we use for corner case detection.\ Detection System ---------------- As already mentioned, the task of corner case detection was an ill-defined problem lacking useful metrics so far. This is the reason why our approach to corner case detection relies on a clear definition of what a corner case is (Fig. 2), which is then conceptually implemented by a respective modular structure (Fig. 3), where two of the modules follow well-known quality metrics (see Sections IV.A and IV.B).\ Typical metrics for the third module, namely the detection system, such as the false acceptance rate (FAR), false rejection rate (FRR) or the area under curve (AUC) of the receiver operation characteristic (ROC) can only be applied if (human-)labeled test data is available. To the best of our knowledge, there is no dataset of labeled corner cases for autonomous driving available so far. Therefore, in this work, we show exemplary results that were achieved on demo video material from the Cityscapes dataset.\ Fig. 6 shows some exemplary results of $D_\text{MSE}(\hat{\mathbf{x}}_{t}, \mathbf{x}_{t})$ (3) (lower subfigure) and the final corner case score $\epsilon_t$ (7) (upper subfigure) on the demo video `stuttgart_00` of the Cityscapes dataset. A problem with the image prediction module is that the squared error $e_t^2(i)$ was considerably higher in image regions with high frequencies. \[tab:pred\_results\] We have therefore low-pass-filtered both the real image and the predicted image with a Gaussian kernel filter with kernel size $10\!\times\!10$ and calculated the squared error afterwards. In the lower subfigure the MSE (3) is depicted without such blurring (blue, dashed) and with blurring (green, solid). The effect of blurring both the predicted image and the real image with a Gaussian kernel filter before subtraction (2) is that it helps to achieve $D_\text{MSE} \approx 0$ in the many ordinary situations, while potential corner cases lead to fewer clearly observable periods of $D_\text{MSE}$ being larger than zero.\ The plot in the upper subfigure (red, solid) shows the corner case score $\epsilon_t$ that is calculated according to (7). The corner case score is based on the MSE of the blurred images (6). It can be seen, that the system focuses on the relevant classes and suppresses high MSE values for non-relevant objects, when there are either no relevant classes or they are all predictable. Marked are three exemplary situations yielding medium to high corner case scores. In the first situation, the preceding car is turning right and the driver of the ego vehicle is taking a slight left turn to overtake the other car. In the second situation, a pedestrian is suddenly crossing the street in a road curve, a potentially dangerous situation. The third situation is an oncoming car that is not predictable for our system and close to the ego vehicle, which in our case leads to a high corner case score. It can be seen that the woman immediately crossing in a road curve produces the highest corner case score in our system, a situation, which indeed can be considered a corner case. CONCLUSIONS =========== In this work we proposed a formal definition for a corner case that is applicable to autonomous driving. We consider that a corner case is given, if there is a *non-predictable* *relevant object/class* in *relevant location*. Each of the three aspects of our definition is covered by a module in the proposed corner case detector. The semantic segmentation to identify relevant objects and the image prediction are both subsystems that can be improved on their own. Their performance can be easily compared to knew methods due to widely accepted metrics in the respective research fields. For the third subsystem, the detection system, we presented a conceptual framework, which showed promising results in preliminary qualitative experiments. It also showed the urgent need for data that covers labeled corner cases. Accordingly, the next step of our work is to record videos of (arranged) corner cases along with labels in order to be able to also quantitatively evaluate the entire detection system.\ ACKNOWLEDGMENT ============== The authors gratefully acknowledge support of this work by Volkswagen Group Research, Wolfsburg, Germany. [^1]: $^{\ast}$Jan-Aike Bolte, Andreas Bär and Tim Fingscheidt are with the Institute for Communications Technology, Technische Universit[ä]{}t Braunschweig, Schleinitzstr. 22, 38106 Braunschweig, Germany [{j.bolte, a.baer, t.fingscheidt}@tu-bs.de]{} [^2]: $^{\circ}$Daniel Lipinski is with Volkswagen Group Research - Automated Driving, Berliner Ring 2, 38440 Wolfsburg, Germany [[email protected]]{} [^3]: In defining a “situation”, we follow the definition by Ulbrich et al. [@Ulbrich2015], “a situation is the entirety of circumstances, which are to be considered for the selection of an appropriate behaviour pattern at a particular point of time.”

ArXiv

--- author: - 'Nicole F. Bell,' - Yi Cai and - 'Rebecca K. Leane' bibliography: - 'darkhiggs\_bib.bib' title: Impact of Mass Generation for Simplified Dark Matter Models --- Introduction ============ The search for particle interactions of dark matter (DM) is currently being pursued across a great variety of experiments. Foremost amongst these are the searches for Weakly Interacting Massive Particles (WIMPs) [@Bergstrom:2000pn; @Bertone:2004pz]. The WIMP mass and coupling parameters are being probed with unprecedented sensitivity at in direct detection experiments such as LUX [@Akerib:2013tjd; @Akerib:2015rjg] and PandaX-II [@Tan:2016zwf], the mono-$X$ searches at the Large Hadron Collider [@Aaboud:2016tnv; @Aaboud:2016uro; @Aaboud:2016qgg; @ATLAS:2016tsc; @ATLAS:1437995; @CMS:2016flr; @CMS:2016fnh; @CMS:2016hmx; @CMS:2016pod; @CMS:2016mjh; @CMS:2016xok; @CMS:2016mxc; @CMS:2016uxr; @Carpenter:2012rg; @Carpenter:2013xra; @Petrov:2013nia; @Bell:2012rg; @Bell:2015rdw; @Birkedal:2004xn; @Gershtein:2008bf; @Goodman:2010ku; @Crivellin:2015wva; @Petriello:2008pu; @Berlin:2014cfa; @Lin:2013sca; @Fox:2011pm; @Bai:2015nfa; @Autran:2015mfa; @Gupta:2015lfa; @Ghorbani:2016edw], and in indirect detection analysis of astrophysical gamma-ray fluxes such as those measured by the Fermi-LAT satellite [@Ackermann:2015zua]. The physics reach of these searches is such that we can realistically expect to cover much of the WIMP parameter space in the near future. As such, it is imperative to have well-formulated models of DM interactions which span a comprehensive spectrum of possible interaction types, in a manner which is as model independent as possible. Simplified models address this aim by introducing a single DM candidate and a mediator which communicates between the dark and SM sectors  [@Abdallah:2014hon; @Buckley:2014fba; @Alves:2011wf; @Alwall:2008ag; @Abdallah:2015ter; @Abercrombie:2015wmb; @DeSimone:2016fbz; @Jacques:2016dqz]. The three most commonly considered benchmark simplified models involve the interaction of fermionic DM with Standard Model (SM) fermions via a spin-1 $s$-channel mediator, a spin-0 $s$-channel mediator, or a spin-0 $t$-channel mediator [@Abercrombie:2015wmb]. These simplified models are an improvement over the effective field theory approach [@Goodman:2010yf; @Goodman:2010ku; @Duch:2014xda] which was used for many recent collider and non-collider WIMP searches, yet suffers from unitarity issues when used outside the region of validity [@Shoemaker:2011vi; @Fox:2012ee; @Busoni:2013lha; @Buchmueller:2013dya; @Busoni:2014sya; @Endo:2014mja; @Busoni:2014haa; @Hedri:2014mua; @Bell:2015sza; @Baek:2015lna; @Bell:2016obu]. However, the simplified models are still far from ideal. Indeed, by their simplified nature, they are are not intrinsically capable of capturing the realistic phenomenology of many UV complete theories, which may have multiple dark-sector field content. More critically, separate consideration of the benchmark simplified models can lead to scenarios that are not physically viable. Indeed, the simplified models suffer some of the same issues that plague the effective field theory approach, such as violations of perturbative unitarity that arise because gauge invariance is not respected [@Shoemaker:2011vi; @Fox:2012ee; @Busoni:2013lha; @Buchmueller:2013dya; @Busoni:2014sya; @Endo:2014mja; @Busoni:2014haa; @Hedri:2014mua; @Bell:2015sza; @Baek:2015lna; @Kahlhoefer:2015bea; @Bell:2015rdw; @Haisch:2016usn; @Englert:2016joy; @Bell:2016obu]. As an example of such an issue, simplified models in which the DM has a non-zero axial-vector coupling to a spin-1 mediator will violate perturbative unitarity at high energies [@Cline:2014dwa; @Kahlhoefer:2015bea]. This can be remedied by introducing a dark Higgs field to unitarize the longitudinal component of the $Z'$ [@Cline:2014dwa; @Kahlhoefer:2015bea; @Bell:2016fqf; @Duerr:2016tmh]. The dark Higgs may also provide mass to the DM itself. The minimal self-consistent approach is then a multi-mediator model, featuring both spin-1 and spin-0 mediators[^1]. This of course can alter the phenomenology, even at low energies. In our recent work [@Bell:2016fqf], we considered indirect detection signals in a scenario with a Majorana DM candidate $\chi$, in which the couplings of a $Z'$ and scalar, $s$, are related by gauge invariance. In this scenario, the presence of both the $s$ and $Z'$ mediators opens a dominant $s$-wave annihilation channel, $\chi\chi\rightarrow sZ'$, that does not arise when a single-mediator is considered in isolation [@Bell:2016fqf]. This has a dramatic impact on the indirect detection phenomenology. An important consideration for DM models is the mass generation mechanism for the dark sector fields. Although commonly left unspecified in the simplified model approach, with mass terms simply added by hand, we shall argue that the mechanism of mass generation has significant consequences that cannot be ignored. For a spin-1 mediator with only vector couplings, a standard procedure is to appeal to the Stueckelberg mechanism to introduce a mass for the vector boson. However, this is valid only for a pure vector, with vanishing axial-vector couplings to fermions. This is a very specific scenario, and there is no reason to assume it is correct. In fact, the Higgs mechanism is the only mass generation mechanism we know is realized by nature, as confirmed by the recent experimental discovery of the SM Higgs boson. As such, it is well motivated to consider a variety of scenarios where different dark sector fields acquire their mass by various methods: the Stueckelberg mechanism, a dark Higgs mechanism, or in cases where it is allowed, simply with a bare mass term. We will show that the annihilation processes, and hence both the relic density and indirect detection constraints, are strongly dictated by the mass generation mechanisms. Interestingly, we will also show that depending on the choice of mass generation mechanism, only particular interactions types are allowed, as dictated by dark gauge invariance. In most cases, only pure vector couplings of the spin-1 mediator to fermionic DM are allowed. Conversely, if a single dark Higgs mechanism gives mass to all the dark sector fields, the axial-vector coupling of the spin-1 mediator to the DM is required to be non-zero. Such restrictions do not map to the single-mediator simplified models, despite being a compelling possibility (or in some cases, a requirement). Again, this phenomenology is not accurately captured by the single mediator simplified model framework. The purpose of this paper is to undertake a more complete study of simplified models that contain both a scalar and vector mediator. In all cases, we will be sure to enforce gauge invariance with respect to the dark $U(1)_\chi$ interaction (dark gauge invariance), which is important to ensure physically well behaved cross sections. We will consider Dirac DM, which allows for a wider combination of coupling types, each with their own distinct phenomenology. Results for Majorana DM can be obtained in the limit of one of the scenarios we investigate in this paper. We focus, in particular, on hidden-sector type models [@Pospelov:2007mp; @Pospelov:2008jd; @Pospelov:2008zw; @Feng:2008mu; @Feng:2008ya; @Rothstein:2009pm; @Mardon:2009gw; @Mardon:2009rc; @Meade:2009rb; @Cheung:2010gj; @Davoudiasl:2013jma; @Berlin:2014pya; @Liu:2014cma; @Hardy:2014dea; @Boehm:2014bia; @McDermott:2014rqa; @Chacko:2015noa; @Elor:2015tva; @Elor:2015bho; @Abdullah:2014lla; @Martin:2014sxa; @Ko:2015ioa; @Ko:2014gha; @Kim:2016csm; @Hooper:2012cw; @Berlin:2015wwa], where the DM annihilates directly to the mediators, which then decay to SM particles via small couplings between the dark and visible sectors. In section \[sec:models\] we outline mass generation for spin-1 simplified models, and in section \[sec:scenI\] we briefly discuss the standard assumption for mass generation in spin-1 models, before investigating three other compelling mass generation scenarios in sections \[sec:scenII\], \[sec:scenIII\] and \[sec:scenIV\], detailing models, annihilation processes and relic density constraints. We present indirect detection constraints in section \[sec:indirectdet\] and summarize our findings in section \[sec:summary\]. Mass Generation for Spin-1 Simplified Models {#sec:models} ============================================ The mass generation mechanism for fermionic DM in spin-1 simplified models is tightly correlated with the DM interaction type. In the case that DM is Majorana, the $Z'$ can have only axial-vector couplings to the DM, as vector couplings of Majorana particles vanish. In the case where DM is Dirac, both vector and axial-vector couplings to the $Z'$ can simultaneously be present. For both DM types, the presence of an axial-vector coupling is significant, as it implies that 1. The DM mass must arise after symmetry breaking, as the $U(1)_\chi$ gauge symmetry prevents a bare mass term for $\chi$, and 2. A $U(1)_\chi$ symmetry breaking mechanism is required to give the $Z'$ mass, in order to unitarize the longitudinal component of the $Z'$. A single dark Higgs field is an economical solution to these issues. In the following sections, we will show that the only scenario in which an axial-vector coupling is possible in a spin-1 mediator model is if there is a dark Higgs which interacts with both the DM and the dark gauge boson. Moreover, the axial coupling is not merely possible in this case, but in fact required to be non-zero by gauge invariance. We take the DM to be Dirac, as this permits the broadest range of possible coupling types. A related model involving Majorana fermions can be found in Ref. [@Bell:2016fqf] and is closely related to a specific realization of scenario II presented below. [|c|M[2.0cm]{}|M[2.1cm]{}|M[3.7cm]{}|M[2.8cm]{}|M[0.6cm]{}|]{} **Scenario** & **$\chi$ mass** & **$Z'$ mass** & **Required $\chi-Z'$ coupling type** & **Annihilation processes** & **$Z'$ pol**\ & Bare mass term & Stueckelberg mechanism & Vector & & $Z'_T$\ & Yukawa coupling to Dark Higgs & Dark Higgs mechanism & Vector $\&$ axial-vector\ or\ pure axial-vector. The $U(1)$ charge assignments of $\chi_L$ and $\chi_R$ determine the relative size of the V and A couplings. **The axial-vector coupling must be non-zero**. & & $Z'_T$ & $Z'_L$\ & Yukawa coupling to Dark Higgs & Stueckelberg mechanism & Vector & & $Z'_T$\ & Bare mass term & Dark Higgs mechanism & Vector & & $Z'_T$\ For Dirac DM, it is possible to have pure vectorlike couplings to the $Z'$ and so it is possible to include a bare mass term for DM, and use the Stueckelberg mechanism[^2] to provide a mass for the $Z'$, such that no dark Higgs is needed. Nonetheless, even in the case of pure vector couplings, a dark Higgs may still provide mass for one or both of the $Z'$ and DM. Furthermore, when the $Z'$ and DM masses arise from different mechanisms, the coupling of the DM to the scalar and vector mediators are no longer related to each other, and hence the phenomenology is less constrained. We are thus led to a spectrum of models in which both scalar and vector mediators would be present. We outline the phenomenologically distinct scenarios in Tab. (\[table:cases\]). Scenario I: Bare DM Mass and $Z'$ Mass from Stueckelberg Mechanism {#sec:scenI} ================================================================== *Interaction type required: Pure Vector* This is the most minimal gauge-invariant scenario, and is permitted only if there are pure vectors couplings between the DM and the $Z'$. Unlike the axial-vector scenario, a dark Higgs is not mandatory in the pure vector case because 1. The $Z'$ gauge boson can acquire a mass via the Stueckelberg mechanism. 2. As $\chi$ is vectorlike with respect to the $U(1)_\chi$, i.e. $Q_{\chi_R}=Q_{\chi_L}$, a bare $\chi$ mass term is permitted. Model ----- The Lagrangian for this case is simply $$\begin{aligned} \mathcal{L}= \mathcal{L}_{SM} + i\, \overline{\chi} (\partial_\mu +ig_\chi Q_V Z'_\mu)\gamma^\mu \chi -\frac{\sin\epsilon}{2} Z'^{\mu\nu}B_{\mu\nu}-m_\chi\overline{\chi}\chi +\frac{1}{2}m_{Z'}^2 Z'^\mu Z'_\mu.\end{aligned}$$ where $Q_V$ is the vectorlike $U(1)_\chi$ charge of the DM, which can be chosen freely, and the $\epsilon$ term describes kinetic mixing of the $U(1)_\chi$ gauge boson with the SM hypercharge gauge boson. This is the only spin-1 mediator scenario where it is possible to avoid the inclusion of a dark Higgs. This case has been thoroughly covered in the literature (for a review see, e.g., [@DeSimone:2016fbz]); we do not discuss it further. Scenario II: DM Mass and $Z'$ Mass both from Dark Higgs Mechanism {#sec:scenII} ================================================================= *Interaction type required: Non-Zero Axial-Vector* We now consider the case where both the DM and the dark gauge boson acquire mass from a single dark Higgs. We will show that this requires the axial-vector DM-$Z'$ interaction to be non-zero. The reason is simple: the dark Higgs field, $S$, must clearly carry $U(1)_\chi$ charge if its vacuum expectation value (vev) is to break that symmetry. A Yukawa coupling of the dark Higgs to the DM of the form $y_\chi \chi_R \chi_L S$ is then possible only if the DM is chiral, i.e. $\chi_L$ and $\chi_R$ carry different $U(1)_\chi$ charges. This guarantees that the axial couping is non-zero (while the vector couplings may be either zero or non-zero depending on the $U(1)_\chi$ charge assignments). Model ----- We investigate the phenomenology of the most minimal model containing a dark gauge boson and a dark Higgs field, by simply extending the Standard Model by an extra $U(1)$. The gauge group is thus: $SU(3)_c\otimes SU(2)_W\otimes U(1)_Y\otimes U(1)_\chi$. Here the covariant derivative is $D_\mu=D_\mu^{SM}+i Q g_{\chi}Z'_\mu$, where $Q$ denotes the $U(1)_\chi$ charge. The SM field content is augmented by a Dirac fermion DM candidate, $\chi$, a spin-1 dark gauge boson, $Z'$, and a dark Higgs field $S$. The vev of the dark Higgs field provides a mass generation mechanism for the dark sector fields $Z'$ and $\chi$. Before electroweak and $U(1)_\chi$ symmetry breaking, the most general Lagrangian is $$\begin{aligned} \mathcal{L} = \mathcal{L}_{\rm SM} &+ i \overline{\chi}_L\slashed{D}\chi_L + i \overline{\chi}_R\slashed{D}\chi_R - \left(y_\chi \overline{\chi}_R {\chi}_L S + h.c.\right) \nonumber\\ & + (D^\mu S)^\dagger (D_\mu S) - \mu_s^2 S^\dagger S - \lambda_s (S^\dagger S)^2 - \lambda_{hs}(S^\dagger S)(H^\dagger H) -\frac{\sin\epsilon}{2} Z'^{\mu\nu}B_{\mu\nu}.\end{aligned}$$ We assume that the SM fields are not charged under $U(1)_\chi$. There are thus only two possible terms that couple SM and dark-sector fields: the kinetic mixing of the $U(1)$ gauge boson with the hypercharge gauge boson, controlled by the kinetic mixing parameter $\epsilon$, and mixing of the dark Higgs, $S$, with the SM Higgs, $H$, controlled by the Higgs mixing parameter $\lambda_{hs}$. In order for the $\chi$-$S$ Yukawa term to be gauge invariant, the charges of the dark sector field must be chosen to satisfy[^3] $$\begin{aligned} Q_{\chi_R} - Q_{\chi_L} = Q_S \; . \end{aligned}$$ We can set the dark Higgs charge to be $Q_S=1$, without loss of generality, as any other choice can be absorbed into a rescaling of the dark gauge coupling. The $\chi$ charges therefore satisfy $$\begin{aligned} Q_A \equiv \frac{1}{2} (Q_{\chi_R} - Q_{\chi_L}) &= \frac{1}{2}, \\ Q_V \equiv \frac{1}{2} (Q_{\chi_R} + Q_{\chi_L}) &= \frac{1}{2} + Q_{\chi_L}. \end{aligned}$$ These charges determine the vector and axial-vector couplings of the $Z'$ to the $\chi$. We see that the axial-vector coupling is completely determined, while there is freedom to adjust the vector coupling by choosing $Q_{\chi_{L,R}}$ appropriately. For instance, $Q_{\chi_L} = 0$ would lead to equal vector and axial-vector couplings, while $Q_{\chi_L} \gg 1$ would lead to a vector coupling much larger than the axial-vector. Pure axial-vector is obtained with $Q_{\chi_L} = - 1/2$; this produces phenomenology similar to the Majorana model studied in [@Bell:2016fqf]. Pure vector, on the other hand, can only be approximately reached in the limit $Q_V \gg 1$, but never fully realized[^4], as dark gauge invariance prevents the axial-vector from being exactly zero. Both $S$ and $H$ obtain vevs, breaking $SU(3)_c\otimes SU(2)_L\otimes U(1)_Y\otimes U(1)_\chi$ down to $SU(3)_c\otimes U(1)_{em}$. In the broken phase, the terms of interest are $$\begin{aligned} \mathcal{L}\supset & -\frac{1}{2}m_{s}^2 s^2 +\frac{1}{2}m_{Z'}^2 Z'^\mu Z'_\mu -m_\chi \overline{\chi}\chi \nonumber \\ & + g_\chi^2 w Z'^\mu Z'_\mu s - \lambda_s w s^3 - 2\lambda_{hs} h \, s (v \, s+ w \, h) + g_f \sum_f Z'_\mu \overline{f}\Gamma^\mu_f f \\ & - g_\chi Q_V Z'_\mu \overline{\chi}\gamma^\mu \chi \nonumber - g_\chi Q_A Z'_\mu \overline{\chi}\gamma^\mu \gamma_5 \chi -\frac{y_\chi}{\sqrt{2}} s \overline{\chi}\chi \; , \label{eq:brokenlgn} \end{aligned}$$ where the component fields of $S$ and $H$ are defined in the broken phase as $S\equiv \frac{1}{\sqrt{2}}(w + s + ia)$ and $H=\left\{ G^+, \frac{1}{\sqrt{2}}(v + h + iG^0)\right\}$ with $G^+$, $G^0$ and $a$ being the Goldstone bosons of $W$, $Z$ and $Z^\prime$ respectively, while $s$ and $h$ are real scalars. The coupling $g_f$, which controls the interactions of the $Z'$ with SM fermions, is dictated by the kinetic mixing; the explicit form can be found, e.g., in Ref. [@Agashe:2014kda]. We assume that the scalar mixing parameter $\lambda_{hs}$ is small, which implies that the the SM Higgs is not significantly perturbed by the new physics. In this limit, the dark Higgs vev satisfies $w^2=-\mu_s^2/\lambda_s$ and the various masses are: $$\begin{aligned} m_{Z'}& =g_\chi w, \\ m_{\chi}& =\frac{1}{\sqrt{2}}y_\chi w, \\ m_{s}^2 & \simeq - 2\mu_s^2, \\ m_{h}^2 & \simeq -2 \mu_h^2 \; . \end{aligned}$$ \[eq:masses\] Importantly, because both the DM and $Z'$ masses are both proportional the to vev of the dark Higgs, their masses and couplings are not all independent parameters but instead are related as $$\label{eq:coupling} y_\chi/g_\chi = \sqrt{2} m_\chi/m_{Z'}.$$ Cross Sections -------------- The relevant annihilation process for this scenario are shown in Tab. (\[table:cases\]). The $\chi\overline{\chi} \rightarrow sZ'$ annihilation receives contributions from both $s$ and $t/u$ channel processes, while $s$-wave contributions to the $\chi\overline{\chi} \rightarrow Z'Z'$ process arise only from the $t/u$ channel diagrams. (Note, however, that the contribution of the $s$-channel scalar exchange diagram to the annihilation to $Z'Z'$ is necessary to unitarize the cross section at high energy. Without this contribution, longitudinal $Z_L'$ contributions would lead to unphysical high energy behavior of the $p$-wave term.) The $s$-wave contributions to the annihilation cross sections are given by $$\begin{aligned} \langle \sigma v \rangle_{\chi\overline{\chi}\to Z'Z'} & = \frac{g_{\chi }^4 \left(1-\eta _{Z'}\right){}^{3/2} \left( 16Q_V^4 \eta _{Z'} + 8 Q_V^2 \left(4-3 \eta _{Z'}\right)+ \eta _{Z'} \right)}{64 \pi m_{\chi }^2 \left(\eta _{Z'}-2\right){}^2 \eta _{Z'}},\end{aligned}$$ and $$\begin{aligned} \langle \sigma v \rangle_{\chi\overline{\chi}\to sZ'} & =\frac{g_{\chi }^4 \sqrt{ \left(\eta _s - \eta _{Z'} -4 \right)^2 - 16 \eta_{Z'}}}{1024 \pi m_{\chi }^2 \left(\eta _{Z'}-4\right)^2 \eta _{Z'}^2 \left(\eta _s +\eta _{Z'} - 4\right)^2} \times\nonumber\\ & \Big\{ \left(\eta _{Z'}-4\right)^2 (\eta_s + \eta_{Z'}-4)^2 \left( (\eta _s -\eta _{Z'} -4)^2 - 16 \eta_{Z'}\right) \\ &+4Q_V^2 \eta _{Z'} \Big[\eta _s^4 \eta _{Z'}-16 \eta _s^3 \eta _{Z'}-2 \eta _s^2 \left(\eta _{Z'}^3-44 \eta _{Z'}^2+80 \eta _{Z'}-64\right) \nonumber\\ &+ 64 \eta _s \left(\eta _{Z'}-4\right){}^2 \left(\eta _{Z'}-1\right)+ \left(\eta _{Z'}-4\right){}^4 \left(\eta _{Z'}+8\right)\Big] \Big\},\nonumber\end{aligned}$$ where $\eta_{s,Z'}=m_{s,Z'}^2/m_\chi^2$. As explained above, we have set $Q_S = 1 = 2Q_A$ without loss of generality, while $Q_V$ is left as a free parameter. Also note that we have used Eq. (\[eq:coupling\]) to replace the Yukawa coupling $y_\chi$ with the gauge coupling $g_\chi$. These annihilation cross sections are plotted in Fig. (\[fig:xsec\_scenII\]). We see that the $Z'Z'$ cross section becomes approximately independent of the DM mass when $m_\chi \gg m_{Z'}$, while the $sZ'$ cross section rises with $m_\chi$. (This is to be contrasted with the behavior in cases III and IV, where all cross sections decline as $m_\chi$ is increased.) This is an interesting consequence of having both vector and axial-vector interactions present: For the $Z'Z'$ process, there is a $V-A$ interference which gives rise to longitudinal $Z'$ domination in the $m_\chi^2 \gg m_{Z'}$ limit[^5]. The $sZ'$ process is also dominated by $Z'_L$ contributions in this limit. This can be understood by appealing to the Goldstone boson equivalence theorem which, in the high energy limit, relates the amplitude for emission of a longitudinally polarized gauge boson ($Z'_L$) with that for the emission of the corresponding Goldstone boson (the pseudoscalar $a$). For the $\chi\overline{\chi}\rightarrow sZ'$ process, in addition to the transverse contributions we have $\chi\overline{\chi}\rightarrow sZ'_L$, which in the high energy limit is equivalent to $\chi\overline{\chi}\rightarrow sa$. As this scalar plus pseudoscalar final state is odd under parity, this is an $s$-wave process. For the $\chi\overline{\chi}\rightarrow Z'Z'$ process, if both $Z'_L$ are replaced by their Goldstones we would have $\chi\overline{\chi}\rightarrow aa$, which is even under parity and thus $p$-wave. However, a combination of longitudinal and transverse modes are possible, $\chi\overline{\chi}\rightarrow Z'_LZ'_T$, which is equivalent to the $s$-wave process $\chi\overline{\chi}\rightarrow aZ'_T$ and thus dominates at high energy. Notice that the $sZ'$ process, in addition to external $Z'_L$ contributions, also receives contributions from the longitudinal $Z'$ mode in the $s$-channel $Z'$ propagator. This contribution leads to four powers of $m_{Z'}$ in the denominator of the $sZ'$ cross section. In contrast, the $Z'Z'$ cross section receives $Z'_L$ contributions only from a single final state $Z'$, and so has only two powers of $m_{Z'}$ in the denominator. The $Z'Z'$ process is thus is sub-dominant to the $sZ'$ process when both are kinematically allowed[^6]. Relic Density {#sec:relicdensII} ------------- An important requirement for a DM model is to produce the correct relic density. Note, however, that a full DM model is likely to have more dark sector fields than the simplified models considered here, which may impact the relic density determination. Nonetheless, we shall determine the relic density constraints for our simplified models, to serve as a guide to the viable regions of parameter space. We use [micrOMEGAs 3]{} [@Belanger:2013oya] to calculate the DM relic density, and compare with the recent determination by the Planck collaboration [@Ade:2013zuv], $$\begin{aligned} \Omega_\chi h^2 = 0.1196 \pm 0.0031.\end{aligned}$$ For different $Z'$ and dark Higgs mass, we scan the parameter space and find that the DM relic density can be saturated fairly easily. We show the relic contours in Fig. (\[fig:relicII\]) as a function of $m_\chi$ and $g_\chi$ for fixed $m_{Z'}$ and $m_s$, and $Q_V=1/2$. In each panel, the observed relic $\Omega_\chi h^2 = 0.1196$ is depicted by a black solid line, while red dot-dashed and blue dotted lines show contours for $\Omega_\chi h^2 = 0.01$ and $1.0$ respectively. The central panel clearly shows the resonant enhancement of the annihilation to $Z'Z'$ through the $s$-channel scalar exchange, as a spike near $m_\chi\sim 100$ GeV. Other features of the relic contours in the middle and right panels are associated with the $sZ'$ final state. The parameter regions shown in Fig. (\[fig:relicII\]) all satisfy $\lambda_s<\sqrt{4\pi}$. Scenario III: DM Mass from Dark Higgs Mechanism, $Z'$ Mass from Stueckelberg Mechanism {#sec:scenIII} ====================================================================================== *Interaction type required: Pure Vector* We now consider a scenario where the mass of the $\chi$ and $Z'$ arise from different mechanisms. Specifically, we assume the $\chi$ mass is due to a Higgs mechanism, while the $Z'$ mass arises from the Stueckelberg mechanism. As a result only pure vector interactions of the $\chi$ and $Z'$ are permitted. Here the dark $U(1)_\chi$ remains unbroken, and instead the dark Higgs must break some other symmetry under which the DM is charged. This scenario divorces the $Z'$ physics from the dark Higgs physics[^7]. Model ----- The most minimal Lagrangian for this setup is $$\begin{aligned} \mathcal{L} = \mathcal{L}_{SM} &+& i\, \overline{\chi}\left( \slashed{\partial} +ig_\chi Q_V \slashed{Z'} \right) \chi -\frac{y_\chi}{\sqrt{2}} \overline{\chi} {\chi} \phi \\ &+& \frac{1}{2}\partial_\mu \phi \partial^\mu \phi -\frac{1}{2} \mu_s^2 \phi^2 - \frac{1}{4}\lambda_s \phi^4 - \frac{1}{2}\lambda_{hs} \phi^2 (H^\dagger H) -\frac{\sin\epsilon}{2} Z'^{\mu\nu}B_{\mu\nu}, \nonumber\end{aligned}$$ with the real scalar $ \phi = w + s$, where $w$ is the vev of $\phi$ and $s$ is the dark Higgs. The vectorlike charge $Q_V$ can be chosen freely. Again, the dark sector interacts with the visible sector in two ways: via kinetic mixing or Higgs mass mixing. As the dark Higgs is responsible only for generating fermion masses, a real scalar is sufficient to accomplish this task. (The dark Higgs must break the $U(1)_\chi$ in all other scenarios we consider, which requires a complex scalar.) If we introduce a complex scalar instead, the extra degree of freedom will be a massless Goldstone boson and will contribute to the radiation energy density of the universe. If the Goldstones had the same temperature as the SM neutrinos, they would make a contribution equivalent to $N^\nu_{\rm eff}=4/7$, in marginal agreement with current experimental observations. However, their contribution to $N^\nu_{\rm eff}$ would be suppressed if they decoupled early enough to not be heated by the annihilations of some SM species [@Weinberg:2013kea]. Cross Sections -------------- As shown in Tab. (\[table:cases\]), both the $Z'Z'$ and $sZ'$ processes receive contributions only from the $t/u$ channel diagrams, as the absence of a $Z'$–$s$ interaction eliminates the $s$-channel diagrams of scenario II. The $s$-wave contributions to the annihilation cross sections are given by $$\langle\sigma v\rangle_{\chi\overline{\chi}\rightarrow Z' Z'} = \frac{g_{\chi}^4 Q_V^4 \left(1-\eta _{Z'}\right)^{3/2}}{4 \pi m_{\chi }^2 \left(\eta _{Z'}-2\right)^2}$$ and $$\langle\sigma v\rangle_{\chi\overline{\chi}\rightarrow s Z'} = \frac{g_{\chi}^2 Q_V^2 y_{\chi }^2 \sqrt{\left(\eta _{s} - \eta _{Z'} - 4\right)^2 -16 \eta _{Z'}} \left(\left(\eta_{s}-\eta _{Z'}-4\right)^2+8\eta _{Z'}\right)}{64 \pi m_{\chi }^2 \left(\eta _{s}+\eta _{Z'}-4\right)^2},$$ where $\eta_{s,Z'}=m_{s,Z'}^2/m_\chi^2$. \ The relative size and behavior of these cross sections can be seen in Fig. (\[fig:xsec\_scenIII\]). Given that the $Z'$ obtains mass from the Stueckelberg mechanism, there are no contributions to the cross sections from longitudinal $Z'$ modes. Therefore, all cross sections decrease with increasing DM mass. It is possible to dial the strength of one annihilation process relative to the other by adjusting the dark Yukawa coupling $y_\chi$ and the dark gauge coupling $g_\chi$, which are independent parameters. (This freedom was not available in scenario II, where the couplings were related.) This is shown in the top two panels of Fig. (\[fig:xsec\_scenIII\]). This also means that if $g_\chi \ll y_\chi$ then $p$-wave processes such as $\chi\overline{\chi}\rightarrow ss$ (which scale as $y_\chi^4$) may have an important effect on the relic density, as the otherwise dominant $Z'Z'$ and $sZ'$ processes (which scale as $g_\chi^4$ and $g_\chi^2y_\chi^2$ respectively) would be suppressed by the small gauge coupling. However, it is difficult to make the annihilation to $ss$ dominate in the universe today, where the $p$-wave modes are suppressed by $v_\chi^2\approx10^{-6}$. To do so would require $g_\chi^2 \sim 10^{-6} y_\chi^2 $ which, while possible, is a very tuned scenario that we shall not consider. The relevant diagrams for annihilation to $ss$ are shown in Fig. (\[fig:ss\_diagrams\]). Relic Density {#relic-density} ------------- In Fig. (\[fig:relic2\]), we show the relic density contours as a function of DM mass $m_\chi$ and the dark gauge coupling $g_\chi$ for various values of the $Z'$ mass, dark Higgs mass, $Q_V=1$ and fixed ratios of $y_\chi/g_\chi$. The color codes for the contours are the same as in the previous scenario. The different choices of $y_\chi/g_\chi$ are embodied in the thickness of the lines: thinner for $y_\chi/g_\chi=1$ and thicker for $y_\chi/g_\chi=5$. Obviously for the same $g_\chi$, a larger $y_\chi/g_\chi$ ratio results in a larger cross section for $\chi\overline{\chi}\to sZ'$ and thus a smaller relic density; a smaller $g_\chi$ is thus needed to obtain the same relic density, resulting in an overall downward shift of the contours. In this scenario, the quartic coupling $\lambda_s$ can be expressed as $\lambda_s\simeq y_\chi^2 m_s^2/(4 m_\chi^2)$. The perturbativity bounds for $\lambda_s$ are shown by the shaded gray regions, while the parameter space where $y_\chi > \sqrt{4\pi}$ ($y_\chi/g_\chi =5$ only) is shown as the light purple region. The dips in Fig. (\[fig:relic2\]) correspond to kinematic opening of various channels when $2m_\chi > m_1 + m_2$ with $m_{1,2}$ being the masses of the annihilation product. For $2m_\chi< m_1+m_2$, the annihilation of DM through these channels will be exponentially suppressed. So in the left panel, where $m_{Z'}=m_s=20$ GeV, all three channels, $sZ'$, $Z'Z'$ and $ss$, are open at the same time and no dips occur beyond this mass. For the middle and right panels of Fig. (\[fig:relic2\]), there is a dip at $m_\chi \simeq 110$ GeV corresponding to the $sZ'$ channel. In the middle panel, the $ss$ channel begins to contribute around $m_\chi\simeq 200$ GeV; the effect is more pronounced for larger $y_\chi$, leading to a prominent dip for $y_\chi/g_\chi=5$ but not for $y_\chi/g_\chi=1$. In the right panel, however, there is no dip around 200 GeV for the $Z'Z'$ channel, since the $Z'Z'$ cross section is always subdominant to the $sZ'$ cross section for the couplings chosen. Scenario IV: Bare DM Mass, $Z'$ Mass from Dark Higgs Mechanism {#sec:scenIV} ============================================================== *Interaction type required: Pure Vector* An alternative scenario in which the mass of the DM and $Z'$ arise from different mechanisms, is to have a bare mass for the $\chi$ and use a dark Higgs mechanism to provide mass for the $Z'$. In this scenario, again, only pure vector interactions of the $\chi$ and $Z'$ are permitted. Model ----- In this scenario, the most minimal gauge invariant Lagrangian is $$\begin{aligned} \mathcal{L}&= & \mathcal{L}_{SM} + i\, \overline{\chi}\left( \slashed{\partial} +ig_\chi Q_V \slashed{Z'}\right) \chi -\frac{\sin\epsilon}{2} Z'^{\mu\nu}B_{\mu\nu}-m_\chi\overline{\chi}\chi\\ &+& \left[ (\partial^\mu + ig_\chi Q_S Z'^\mu ) S \right]^\dagger \left[ (\partial_\mu + ig_\chi Q_S Z'_\mu ) S \right] - \mu_s^2 S^\dagger S - \lambda_s (S^\dagger S)^2 - \lambda_{hs}(S^\dagger S)(H^\dagger H). \nonumber\end{aligned}$$ The vectorlike charge $Q_V$ and dark Higgs charge $Q_S$ under the dark $U(1)_\chi$ can be chosen freely. Again the dark sector interacts with the visible sector in two ways: via kinetic mixing or Higgs mass mixing. Cross Sections -------------- As shown in Tab. (\[table:cases\]), the annihilation to $sZ'$ proceeds only via the $s$-channel diagram, as the DM does not interact directly with the dark Higgs. The annihilation to $Z'Z'$ proceeds via the $t/u$ and channel diagrams. The $s$-wave contributions to these annihilation cross sections are given by $$\langle\sigma v\rangle_{\chi \overline{\chi}\rightarrow Z' Z'} = \frac{g_\chi^4 Q_V^4 \left(1-\eta _{Z'}\right)^{3/2}}{4 \pi m_{\chi }^2 \left(\eta _{Z'}-2\right)^2}$$ and $$\langle\sigma v\rangle_{\chi \overline{\chi}\rightarrow s Z'} =\frac{g_\chi^4 Q_V^2 Q_S^2 \sqrt{ \left(\eta _s - \eta _{Z'} -4\right)^2 - 16 \eta_{Z'}} \left( \left(\eta _s - \eta _{Z'} - 4\right)^2+32\eta_{Z'}\right)}{256 \pi m_{\chi }^2 \left(\eta _{Z'}-4\right)^2},$$ where $\eta_{s,Z'}=m_{s,Z'}^2/m_\chi^2$. The behavior of these cross sections is depicted in Fig. (\[fig:xsec\_scenIV\]). We see that the shapes of the $sZ'$ and $Z'Z'$ cross sections are similar, as both fall off with DM mass as $1/m_\chi^2$. There is no production of longitudinal $Z'_L$ modes in the high energy limit, which is consistent with the fact that the DM does not interact with Goldstone modes, given the absence of a DM-Higgs coupling. Because $Q_{V}$ and $Q_S$ are independent, the relative size of the $Z'Z'$ and $sZ'$ processes can again be scaled relative to each other by appropriate choices of these charges. \ Relic Density {#relic-density-1} ------------- We plot the relic density contours for this scenario in Fig. (\[fig:relic3\]). As the dark Higgs is responsible for the $Z'$ mass, the quartic coupling may be expressed as $\lambda_s\simeq g_\chi^2 m_s^2/(2 m_{Z'}^2)$. Parameters excluded by the perturbativity bound on $\lambda_s$ are shaded gray; this bound is relevant only for the middle panel of Fig. (\[fig:relic3\]), where the ratio of $m_s/m_{Z'}$ is larger. Because there is no direct coupling of the scalar to the DM, there is no annihilation to $ss$. As a result, the features of the relic density contours are generally simpler than in the previous scenario. For the chosen values of $Q_{V,S}$, the annihilation to $sZ'$ is subdominant to the $Z'Z'$ process when both are kinematically allowed. This leads to a dip in the contours of the right panel at $m_\chi \simeq m_{Z'}$, where the $Z'Z'$ modes becomes allowed, but not in the left and center panels where the $Z'Z'$ mode always plays the dominant role. Indirect Detection Phenomenology {#sec:indirectdet} ================================ We now determine indirect detection constraints on the dominant annihilation modes for the scenarios discussed, $\chi \overline\chi \rightarrow Z'Z'$ and $\chi \overline\chi \rightarrow sZ'$. The $Z'$ and $s$ produced in these annihilations decay to SM particles, and subsequent hadronization/decay of these SM states leads to gamma-ray and other fluxes that we may compare with observational limits. We generate our gamma-ray spectra as per the method outlined in Ref. [@Bell:2016fqf], where a more detailed description can be found. The kinetic mixing of the $Z'$ with the SM hypercharge boson permits the decay $Z'\rightarrow f\overline{f}$, with a partial width given by $$\Gamma(Z'\rightarrow f\bar{f})=\frac{m_{Z'}N_c}{12\pi}\sqrt{1-\frac{4m_f^2}{m_{Z'}^2}}\left[g_{f,V}^2\left(1+\frac{2m_f^2}{m_{Z'}^2}\right)+g_{f,A}^2\left(1-\frac{4m_f^2}{m_{Z'}^2}\right)\right], \label{eq:zppartial}$$ where $N_c$ is a color factor, relevant for hadronic decays. The $g_{f,V}$ (vector) and $g_{f,A}$ (axial-vector) structure of the $Z'$–$f$ couplings are inherited from the kinetic mixing [@Agashe:2014kda]. The total decay width for the $Z'$ is then simply given by the sum over all the final state fermions, $\Gamma_Z'=\sum_f \Gamma(Z'\rightarrow f\bar{f})$. The dark Higgs decays to the SM due to mass mixing with the SM Higgs, and so it decays preferentially to heavier particles. The dark Higgs is also permitted to decay to pairs of $Z'$. In order to take into account loop decays and higher order corrections, we calculate the dark Higgs decay widths numerically with the [Fortran]{} package [ HDecay]{} [@Djouadi:1997yw]. The spectra generated are then compared to the strongest indirect detection limits available for our processes[^8]: the Fermi-LAT Pass 8 data on dwarf spheroidal galaxies (dSphs) of the Milky Way [@Ackermann:2015zua]. To find the limit on the cross section from dSphs, we use the maximal likelihood method to compare our spectra against those for the dSphs publicly provided by Fermi-LAT in the Pass 8 data, with the $J$ factor taken to be a nuisance parameter as per Ref. [@Ackermann:2015zua]. We take spectra from 15 dSphs: Bootes I, Canes Venatici II, Carina, Coma Berenices, Draco, Fornax, Hercules, Leo II, Leo IV, Sculptor, Segue 1, Sextans, Ursa Major II, Ursa Minor, and Willman 1. The 95$\%$ C.L. limits on the annihilation cross section are shown Fig. (\[fig:fermix1\]), for various dark Higgs and $Z'$ masses, for both $sZ'$ and $Z'Z'$ processes. \ The limits we show are independently set on either the $\chi\overline{\chi}\rightarrow sZ'$ process or the $\chi\overline{\chi}\rightarrow Z'Z'$ process. They can then be applied to any of the scenarios we study in this paper, assuming that one of the modes dominates. Indeed, they can also be applied to any model that features annihilations to a $sZ'$ or $Z'Z'$ final state, provided the $Z'$ and $s$ communicate with the SM via kinetic or Higgs mass mixing respectively, as the cross section limits depend only on the gamma-ray spectral shape that characterizes a given annihilation mode. From Fig. (\[fig:fermix1\]) it is clear that the limit on the cross section does not vary greatly with the mediator mass provided it is kinematically allowed; it is instead the DM mass with which the energy of final state photons and thus cross section limits is tightly correlated. The thermal relic cross section required to reproduce the correct relic density for conjugate DM is approximately $\langle\sigma v\rangle\approx 4.4 \times10^{-26}$ cm$^3$/s [@Steigman:2012nb], which excludes the low DM mass region where the Fermi limits surpass this sensitivity. However, this statement assumes that the $s$-wave contributions to the annihilation cross section dominate both at freeze-out and in the universe today. In fact in some cases, such as scenario III, the $p$-wave processes can make a non-negligible contribution at freeze-out. This means that the relic density constraint could be satisfied, yet the cross section in the universe today suppressed, escaping indirect detection bounds even for low DM mass. Discussion and Summary {#sec:summary} ====================== We have surveyed a spectrum of phenomenologically distinct two-mediator DM models, containing both a dark vector and dark scalar, where gauge invariance is respected and the mass terms for the dark sector fields are introduced in a self-consistent way. These two-mediator DM models correctly capture important phenomenology which is missing in the single mediator approach. Specifically, we modified the usual simplified model setup to incorporate mass generation for the DM candidate and vector mediator, by using combinations of bare mass terms, Higgs mechanisms and Stueckelberg mechansims. We found that the DM interaction types and annihilation processes, and hence both the relic density and indirect detection constraints, are strongly dictated by the mass generation mechanism we choose for the dark sector particles: - Unless the DM and $Z'$ masses both receive contributions from the vev of the same dark Higgs field, pure vector couplings of the spin-1 mediator and DM are required, as discussed in scenarios III and IV. In these scenarios DM annihilates to both $sZ'$ and $Z'Z'$, with the relative rates to these final states controlled by independent coupling constants. Moreover, in the high energy limit, only the $Z'_T$ polarization is produced by these annihilations. - However, if a dark Higgs mechanism gives mass to all the dark sector fields, as per scenario II, the axial-vector coupling between the spin-1 mediator and DM must be non-zero. In this scenario, the $sZ'$ and $Z'Z'$ DM annihilation channels are intrinsically linked. Furthermore, production of the $Z'_L$ polarization enhances the annihilation to $sZ'$. If both the vector and axial-vector couplings are non-zero, the annihilation to $Z'Z'$ is also enhanced by $Z'_L$ (via the $V-A$ interference) though it remains subdominant to the $sZ'$ mode when both are kinematically allowed. One may imagine generalizations of scenarios III and IV in which the $Z'$ and $\chi$ masses arise from [*two different*]{} Higgs mechanisms. Indeed, we would recover scenario III (Stueckelberg $Z'$ mass) in the limit that the Higgs responsible for the $Z'$ mass is taken to infinity. Likewise, we would recover scenario IV (bare $\chi$ mass) in the limit that the Higgs responsible for the $\chi$ mass is taken very large. In these generalizations, the $\chi$-$Z'$ coupling remains of pure vector form. Axial couplings always imply that a Higgs which Yukawa couples to the $\chi$ must carry $U(1)_\chi$ charge, and hence its vev also contributes to the $Z'$ mass, as in scenario II. Such two-scalar models would lead to additional complexity via mixing in the scalar sector, but would not introduce any qualitatively new $Z'$ physics. Our results are not captured by the single mediator approach, where the mass generation mechanism is left unspecified and constraints on the coupling types are not usually applied. This means that by continuing to use simplified models with a single spin-1 mediator, (i) we are at best only testing a very specific subset of the possibilities: Dirac DM with a bare mass and pure vector couplings to a $Z'$ with a Stueckelberg derived mass (i.e. scenario I) or (ii) at worst, experimental constraints may not be meaningful because the models have been oversimplified. Option (i) is not particularly appealing in that it does not cover well motivated possibilities such as Higgs mass generation (which, after all, is a mechanism we know is realized by nature) or Majorana DM. The remaining option, (ii), is far from desirable. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported in part by the Australian Research Council. RKL thanks John Beacom and the Center for Cosmology and AstroParticle Physics (CCAPP) at Ohio State University for their hospitality and support during her visit, where part of this work was completed. Feynman diagrams are drawn using [TikZ-Feynman]{} [@Ellis:2016jkw]. Appendix {#appendix .unnumbered} ======== Cross Sections ============== In the scenarios discussed in this paper, the charges were fixed to particular values either to satisfy gauge invariance, or to demonstrate the phenomenology. The full cross sections with explicit $Q_{A,V}$ and $Q_S$ dependence are listed in this appendix for reference. The full $s$-wave cross section for $\chi\overline{\chi}\rightarrow Z'Z'$ is $$\begin{aligned} \langle \sigma v \rangle_{\chi\overline{\chi}\to Z'Z'} & = \frac{g_{\chi }^4 \left(1-\eta _{Z'}\right)^{3/2} \left(2 Q_A^2 Q_V^2 \left(4-3 \eta _{Z'}\right)+Q_A^4 \eta _{Z'}+Q_V^4 \eta _{Z'}\right)}{4 \pi m_{\chi }^2 \left(\eta _{Z'}-2\right)^2 \eta _{Z'}}.\end{aligned}$$ This expression gives the $s$-wave contribution to the $\chi\overline{\chi}\rightarrow Z'Z'$ cross section for [*all*]{} cases, as only the $t/u$ channel diagrams contribute. (Scalar-mediator contributions only enter at the $p$-wave level.) We see that if either $Q_A$ or $Q_V$ is zero, the cross section scales as $1/m_\chi^2$ in the limit that $\eta_{Z'}= m^2_{Z'}/m^2_\chi \ll 1$, and is dominated by $Z_T'$ contributions only. In the case that both $Q_A$ and $Q_V$ are non-zero, the cross section instead scales as $1/m_{Z'}^2$ in the $\eta_{Z'} \ll 1$ limit, which arises due to the $Z_L'$ modes. Note however, that no violation of unitarity will occur – the $Z'$ mass cannot be made arbitrarily large while satisfying the constraint Eq. (\[eq:coupling\]) and restricting all couplings to perturbative values. This $Z'Z'$ cross section matches that in Refs. [@Alves:2015pea; @Alves:2015mua]. The full $s$-wave cross section for $\chi\overline{\chi}\rightarrow sZ'$ is $$\begin{aligned} \langle \sigma v \rangle_{\chi\overline{\chi}\to sZ'} & =\frac{g_{\chi }^2 \sqrt{ \left(\eta _s - \eta _{Z'} -4 \right)^2 - 16 \eta_{Z'} }}{256 \pi m_{\chi }^2 \left(\eta _{Z'}-4\right)^2 \eta _{Z'}^2 \left(\eta _s+\eta _{Z'}-4\right)^2}\times\\ &\Big\{g_{\chi}^2 Q_S^2 \left(\eta _s+\eta _{Z'}-4\right)^2 \Big[Q_A^2 \left(\eta _{Z'}-4\right){}^2 \left( \left(\eta _s - \eta _{Z'} -4 \right)^2 - 16 \eta_{Z'} \right)\nonumber\\ &+Q_V^2 \eta _{Z'}^2 \left( \left(\eta _s - \eta _{Z'} -4 \right)^2 +32 \eta_{Z'} \right)\Big]\nonumber\\ &+24 \sqrt{2} g_{\chi } Q_S Q_V^2 y_{\chi } \left(\eta _{Z'}-4\right) \eta _{Z'}^{5/2} \left(\eta _s^2-8 \eta _s-\eta _{Z'}^2+16\right)\nonumber\\ &+4 Q_V^2 y_{\chi }^2 \left(\eta _{Z'}-4\right)^2 \eta _{Z'}^2 \left( \left(\eta _s - \eta _{Z'} -4 \right)^2 +8 \eta_{Z'} \right)\Big\}\nonumber.\end{aligned}$$ Taking $2Q_A= Q_S=1$ and using the relation of the Yukawa and gauge coupling in Eq. (\[eq:coupling\]) recovers the cross sections for scenario II, $Q_A=Q_S= 0$ recovers the cross sections for scenario III, and $Q_A, y_\chi= 0$ gives the cross sections for scenario IV. It is still important to note however that the values for the charges cannot be chosen freely and should obey the constraints discussed in this paper. [^1]: Multi-mediator models have also been considered recently in Refs. [@Cline:2015qha; @Choudhury:2015lha; @Ghorbani:2015baa; @Bell:2016fqf; @Duerr:2016tmh]. [^2]: In Abelian gauge theories, the Stueckelberg mechanism can be taken as the limit of the Higgs mechanism where the mass of the real scalar is sent to infinity and only the pseudoscalar is present; however it is not always easily realized in more complicated scenarios. In particular, unitarity is already violated at tree-level in a non-Abelian theory with a Stueckelberg Lagrangian and thus the theory is not renormalizable [@Kunimasa:1967zza; @Slavnov:1972fg; @Veltman:1968ki; @Slavnov:1970tk; @Vainshtein:1971ip; @Shizuya:1976qf; @McKeon:1991hv; @Freedman:1980us; @Kafiev:1982hx]. In general the Stueckelberg mechanism should be treated as an alternative to the Higgs mechanism for mass generation. [^3]: For anomaly cancellation there must be additional fields charged under the dark $U(1)$. However, we include only the lightest of such fields as the DM candidate, as the others can be made heavier such that the phenomenology is not affected, as they make a subdominant contribution to the relic density, as can be seen from section \[sec:relicdensII\]. [^4]: It is also important to note that there exist relations between the axial-vector coupling size and the masses of the dark sector fields [@Kahlhoefer:2015bea]. Therefore, for almost all mass choices of the dark sector fields, it is not possible to make the axial-vector coupling vanishingly small relative to the vector coupling without the vector coupling becoming non-perturbative. Thus, the axial-vector coupling is effectively never negligible and cannot be neglected even in limiting cases. [^5]: If $Q_V$ were chosen to be zero, such that the $Z'$ coupling were pure axial, there would be no $V-A$ interference and the $s$-wave part of the $Z'Z'$ cross section would [*not*]{} be enhanced by longitudinal $Z'$ modes. This situation maps onto the Majorana DM case studied in Ref. [@Bell:2016fqf]. [^6]: Note that because the Yukawa and gauge coupling constants are related via Eq. (\[eq:coupling\]), it is not possible to change the relative size of the annihilation to $Z'Z'$ and $sZ'$ by adjusting these parameters. [^7]: This situation has a SM analogue where $\chi$ is replaced by the electron and $Z'$ is replaced by the photon: the electrons have vector couplings to the photon of an unbroken $U(1)_{\rm QED}$; the electron mass comes from breaking the electroweak symmetry with the SM Higgs; the SM Higgs does not couple to the photon or contribute to the photon mass. The annihilation $\chi\overline{\chi}\rightarrow sZ'$ is then the analogue of $e^+e^-\rightarrow h\gamma$. [^8]: We also include the approximate limit from AMS-02 at low DM masses, adapted from [@Elor:2015bho]. This approximate limit is only applicable if the sum of the final state mediators is less than about 70 GeV.

ArXiv

--- abstract: | We report initial performance results emerging from 600 hours of observations with the Automated Planet Finder (APF) telescope and Levy Spectrometer located at UCO/Lick Observatory. We have obtained multiple spectra of 80 G, K and M-type stars, which comprise 4,954 individual Doppler radial velocity (RV) measurements with a median internal uncertainty of 1.35 [${\rm m~s^{-1}}$]{}. We find a strong, expected correlation between the number of photons accumulated in the 5000-6200Å iodine region of the spectrum, and the resulting internal uncertainty estimates. Additionally, we find an offset between the population of G and K stars and the M stars within the data set when comparing these parameters. As a consequence of their increased spectral line densities, M-type stars permit the same level of internal uncertainty with 2x fewer photons than G-type and K-type stars. When observing M stars, we show that the APF/Levy has essentially the same speed-on-sky as Keck/HIRES for precision RVs. In the interest of using the APF for long-duration RV surveys, we have designed and implemented a dynamic scheduling algorithm. We discuss the operation of the scheduler, which monitors ambient conditions and combines on-sky information with a database of survey targets to make intelligent, real-time targeting decisions. author: - 'Jennifer Burt, Bradford Holden, Russell Hanson, Greg Laughlin, Steve Vogt, Paul Butler, Sandy Keiser, William Deich' title: The capabilities and performance of the Automated Planet Finder Telescope with the implementation of a dynamic scheduler --- [: Jennifer Burt, UCO/Lick Observatory, 1156 High St, Santa Cruz, USA, 95060; E-mail: ]{} Introduction {#sect:intro} ============ The Doppler Velocity technique now has two decades of success in enabling extrasolar planetary detections. It has produced candidates with masses approaching that of Earth, and has been especially successful in detecting long period planets. In recent years, Doppler Velocity confirmations have proven vital to gaining an understanding of the planetary candidates discovered by photometric space missions, especially Kepler. Doppler Velocity campaigns are responsible for 31% of the 1854 confirmed planetary discoveries in the past three decades, but they account for 87% of the 330 confirmed planets with periods longer than 1 year[@ExoplanetsEU]. Ground-based facilities, furthermore, are amenable to the operation of long-term surveys due to their relativity low construction and operational costs along with their ability to be upgraded as instrumentation improves. In order to find analogs of our own solar system, we need to extend the catalog of successful radial velocity (RV) planet detections to encompass longer-period planets (particularly true Jupiter analogs) and smaller mass, short-period planets. This means observing efforts must increase their temporal baselines and cadence of observation to more effectively populate each planet’s RV phase curve. The Automated Planet Finder, located at the Mt. Hamilton station of UCO/Lick Observatory, combines a 2.4-m telescope with a purpose-built, high-resolution echelle spectrograph, and is capable of 1[${\rm m~s^{-1}}$]{} Doppler Velocity precision[@Vogt2014]. Eighty percent of the telescope’s observing time is specifically dedicated to the detection of extrasolar planets. This time is shared evenly between two exoplanet research groups, one at UC Santa Cruz and one at UC Berkeley. Time is allocated in whole night segments, with a schedule developed quarterly by the telescope manager. Target lists and operational software are developed separately as the two exoplanet groups are focused on different types of planet detection/follow up. For a description of the UC Berkeley planet detection efforts, please see Fulton et. al, 2015[@Fulton2015]. The remaining 20% of telescope time is dedicated to at-large use by the University of California community. All users are allowed to request specific nights if it is beneficial to their science goals (for example: to obtain RV values while a planet transits its star), and such requests are taken into account by the telescope manager when setting the schedule. The APF leverages a number of inherent advantages to improve efficiency. For example, its Levy spectrometer, a high-resolution prism cross-dispersed echelle spectrograph with a maximum spectral resolving power of R$\sim150,000$, is optimized for high precision RV planet detection [@Radovan2010; @Vogt2014]. A full description of the design and the individual components of the APF is available in Vogt et al. (2014) [@Vogt2014]. To support long-running surveys, we have developed a dynamic scheduler capable of making real-time observing decisions and running the telescope without human interaction. Through automation and optimization, we increase observing efficiency, decrease operating costs and minimize the potential for human error. The scheduler’s target selection is driven by balancing scientific goals (what we [*want*]{} to observe based on scientific interest, required data quality, and desired cadences) and engineering constraints (what we [*can*]{} observe based on current atmospheric conditions and physical limitations of the telescope). To address these criteria, we need to know how the velocity precision extracted from a given stellar spectrum depends on inputs that can be monitored before and during each observation. We have assessed the influences of the various inputs by analyzing 16 months of data taken on the APF between June 2013 & October 2014. The plan of this paper is as follows: in §\[sect:data\] we describe the current APF radial velocity catalog, paying special attention to the variety of spectral types and the frequency of observations. In §\[sect:Inputs\] we evaluate the relations between velocity precision and parameters including stellar color and V band magnitude, airmass, seeing, date of observation and atmospheric transparency and we explain how these relations inform the nightly decision-making process executed by the observing software. In §\[sect:DismissedInputs\] we outline the parameters that were assumed to be important prior to on-sky observations, but that have since been determined to have little relevance. In §\[sect:Heimdallr\] we describe how the relevant relations are integrated into the scheduling software, and we discuss its structure, its dependencies, and its capabilities. In § \[sect:compare\] we discuss other automated and semi-automated observatories and highlight both the similarities and differences between those systems and that employed by the APF. Finally we conclude in §\[sect:campaigns\] by reviewing the application of the APF’s automated observing strategy to the telescope’s current and upcoming observing scientific campaigns. Data set description {#sect:data} ==================== Description of observing terminology ------------------------------------ This paper differs from most publications discussing precision radial velocity work in that all of the plots, equations and discussions presented are based on individual, un-binned exposures of stars. It is well known that pulsation modes (p-modes) in stars cause oscillations on the stellar surface, adding noise to the RV signal. It has also been well documented that the noise imparted by these p-modes in late-type stars can be averaged over by requiring total observation times longer than the $\sim 5-15$ minute periods typical of the pulsation cycles [@Santos2004; @Dumusque2011]. Thus the majority of radial velocity publications, especially those dealing with exoplanet detections, present [*binned*]{} velocities and error values. That is, they take multiple, individual [*exposures*]{} of a star during the night and then combine (bin) them to create one final [*observation*]{} with its own velocity and internal error estimate [@Dumusque2011; @Burt2014; @Fulton2015]. The binned observations therefore contain more photons than any individual exposure, but, more importantly, average over the pulsation modes on the star, and therefore exhibit a measurably smaller scatter (Figure \[fig:bin\_unbin\]). These terms, exposure and observation, will be used specifically throughout this paper to help make clear to the reader whether we are talking about an exposure - a single instance of the shutter being open and collecting photons from the star, or an observation - the combination of all exposures taken of a star over the course of a night. Similarly, exposure time will refer to the open shutter time during one exposure and observation time will refer to the total open shutter time spent on a target between all exposures. In the case of the star’s observation consisting of only one exposure, then the observation time is equivalent to the exposure time. ![ \[fig:bin\_unbin\] Our radial velocity dataset for the RV standard star HD 185144 showing the individual exposures (open circles in black) and the resulting binned velocities (solid circles in red). The binned velocities have a much smaller RMS value (0.99 [${\rm m~s^{-1}}$]{} for the first 120 days and 1.72 [${\rm m~s^{-1}}$]{} over the entire 400 day span, compared to the 1.72 [${\rm m~s^{-1}}$]{} and 2.06 [${\rm m~s^{-1}}$]{} exhibited by the un-binned data over the same respective time spans) due to their increased signal-to-noise. Additionally the individual exposures are each $\sim$60 seconds long while the binned velocities span at least $\sim$5min, so they average over the p-modes of the star. This paper uses the individual exposure data when carrying out all further analyses and calibrations. The change in the RMS values between the first 120 days and the full data set is largely due to a change in observing strategy. After the first 4 months we started observing HD 185144 less frequently and with fewer exposures in each observation to allow for more targets to be observed each night.](Figure1.pdf){height="8cm"} The APF data set ---------------- The APF’s star list is made up of legacy targets first observed as part of the Lick-Carnegie exoplanet campaign using the HIRES spectrograph on the Keck I telescope. Stars for the Lick-Carnegie survey were selected based on three main criteria: - Spectral type between G and M : stars with spectral types earlier than F5 experience oscillations that can produce pseudo-Keplerian signals, and are fast rotators with few spectral lines so they are avoided - Single stars : no stars with companions closer than 5", as this can lead to scattered light from the nearby companion making its way into the spectrograph slit and contaminating the target star’s signal - Quiet stars : Only stars without large amounts of emission in the Ca II H & K line cores are permitted, as core emission in these lines is an indicator of stellar chromospheric activity [@Noyes1984] which complicates the data reduction process and can produce pseudo-Keplerian signals. The resulting Lick-Carnegie target list is comprised of $\sim$1800 stars, which have been monitored using Keck/HIRES over the past two decades. When creating the initial target list for the APF, the Lick-Carnegie star list was culled for targets with V magnitudes brighter than 12 and declinations above -20$^{\circ}$. In order to efficiently prove the APF’s capabilities, we selected stars with suspected short-period planets (P &lt;100 days) that required only 1-2 more rounds of phase coverage to verify. The presence and the false-alarm probabilities of these short-period, Keplerian signals were determined by analyzing the existing Keck/HIRES RV datasets using the publicly available Systemic Console [@Meschiari2009]. Systemic allows users to fit planetary signals to RV data and derive the orbital properties, while also providing tools to handle error estimation and assess orbital stability. This selection process resulted in a list of 127 stars. The calibrations described in this paper are based on data taken with the APF between June 2013 and October 2014. The data set includes precision Doppler observations of 80 of the 127 stars selected from the Lick-Carnegie survey and this is before the development/inclusion of the dynamic scheduler, thus all data in this paper were obtained using fixed star lists. Our precision Doppler observations encompass spectra of 80 stars and incorporate 600 hours of open shutter time. The stars span spectral types from early G to mid M, have 3.5 &lt;V &lt;12, and are all located within 160 pc (Figure \[fig:APF\_CMD\]). Every APF star has a set of observations containing between one and seven hundred exposures. Individual exposures are restricted to a maximum length of fifteen minutes to avoid cosmic ray accumulation and to minimize uncertainty when calculating the photon-weighted midpoint times. Additionally, we enforce a total observing time limit of one hour per target per night to ensure that telescope time is not wasted on observing faint objects when conditions are poor. For each individual exposure, the $\rm \overline{FWHM}$ (the average FWHM of the star over the integration time) of the guide camera’s seeing disk is logged in the FITS image header, along with the total number of photons from the exposure meter and the total exposure time. Colors, magnitudes and distances for the stars are obtained from SIMBAD [@Wegner2000]. ![ \[fig:APF\_CMD\] Color-Magnitude diagram for stars observed with APF and used in the analysis described herein. Color coding represents the number of exposures that have been obtained for each star. Our dataset spans a wide range in color and magnitude and contains stars with spectral types from F6 to M4.](Figure2.pdf){height="8cm"} All spectra are analyzed using the data reduction techniques described in detail by Butler et al. (1996)[@Butler1996], which produces a measurement of the stellar RVs and associated internal uncertainties. The reduction pipeline analyzes each exposure’s extracted spectrum in 2Å chunks and determines the radial velocity shift for each chunk individually. The final reported velocity is the average over all the chunks, while the internal uncertainty is defined as the RMS of the individual chunk velocity values about the mean, divided by the square root of the number of chunks. Thus the internal uncertainty represents errors in the fitting process, which are dominated by the photon noise. The internal uncertainty does not include potential systematic errors associated with the instrument, nor does it account for astrophysical noise (or “jitter”) associated with the star and, therefore, represents only a lower limit to the accuracy of the data for finding companions. Determining additional systematic errors ---------------------------------------- We use the APF’s 737 individual exposures of HD 185144 (Sigma Draconis, HR 7462), a bright RV standard star, to estimate the precision with which we can measure radial velocities. It should be noted that the radial velocity values produced by our analysis pipeline are all relative velocities and thus have a mean value of zero. We examine the unbinned, or individual, exposures, finding that star’s mean internal uncertainty is $\mu_{int} = 1.0$ [${\rm m~s^{-1}}$]{}and that the mean of the absolute values of its velocity measurements is $\mu_{abs} = 1.8$ [${\rm m~s^{-1}}$]{}. Because $\mu_{abs}$ includes effects from the internal uncertainty [*in addition*]{} to other sources of error such as the stellar jitter and the instrument systematics, its value is always higher than the internal uncertainty value for a given exposure. The difference between these two parameters implies an additional quadrature offset of $1.5$ [${\rm m~s^{-1}}$]{}, which we then generalize as $\sigma_{s} = 1.5$ [${\rm m~s^{-1}}$]{}(Figure \[fig:185144\_unc\]). We take this offset value to represent the additional error contributions from all other systematics, including the known 5-15 minute pulsation modes of the star that the binned data averages over and the systematics from the instrument. If, instead, we compared the mean internal uncertainty to the standard deviation of the velocity measurements ($\sigma_{vel} = 2.3$ [${\rm m~s^{-1}}$]{}) we would require a larger additional offset term ($\sigma_{s} = 2.0$ [${\rm m~s^{-1}}$]{})[^1]. We note, however, that the data are strongly affected by the velocities of exposures that fall in the outlying, non-Gaussian tail. We thus choose to determine our error estimate using the mean of the absolute values of the velocity measurements (instead of the standard deviation of the velocities) as it mitigates the influence of the outliers. We use the relations of Wright et al. (2005)[@Wright2005], which presents radial velocity jitter estimates at the 20$^{th}$ percentile, median and 80$^{th}$ percentile levels to assess the expected stellar activity for HD 185144. We find an estimated median jitter of $\sigma'_{rv} = 3.5$ [${\rm m~s^{-1}}$]{}, and a 20$^{th}$ percentile value of $2.3$ [${\rm m~s^{-1}}$]{}. Noting that the $20^{th}$ percentile value matches the $\sigma_{vel}$ for our exposures, obtained over a timespan of 400 days, suggests that the star is in fact intrinsically quiet. That is to say, we expect only 20% of stars with the same evolution metric, activity metric (both described by Wright 2005) and B-V color as HD 185144 to have activity levels less than $2.3$ [${\rm m~s^{-1}}$]{}. Given that we measure $\mu_{abs} = 1.8$ [${\rm m~s^{-1}}$]{}, HD 185144 is, at a minimum, much quieter than expected based on its color, activity and evolution. Even so, it is reasonable to expect that some fraction of the observed uncertainty is due to the star itself (astrophysical noise) or currently unknown planets rather than instrumental effects, thus indicating that the instrument is performing very well. ![ \[fig:185144\_unc\] Radial velocities from APF exposures on the star HD 185144 \[G9V, $ \rm m_{v}$ = 4.68\], 737 points in total. The internal uncertainty estimates produced by the data reduction pipeline are noticeably smaller than the actual spread in the data. The internal error does not account for telescope systematic errors or sources of astrophysical noise in the star. Our average internal uncertainty is $\mu_{int}$ = 1.0 [${\rm m~s^{-1}}$]{}(blue line) but we find the mean absolute value of the HD185144 velocity measurements to be $\mu_{abs}$ = 1.8 [${\rm m~s^{-1}}$]{}(purple line). We adopt an estimate that the additional systematic uncertainty in our velocity precision is $\sigma$ = 1.5 [${\rm m~s^{-1}}$]{}. ](Figure3.pdf){height="8cm"} To verify that the size of the offset between $\mu_{int}$ and $\mu_{abs}$ is not specific to the HD 185144 data set, we compare the mean values of these parameters for every star observed by the APF during the time span of Figure \[fig:185144\_unc\]. That is, we apply the same procedure detailed above for HD 185144 to all stars observed during the same date range, and then compute and compare the mean values of $\mu_{int}$ and $\mu_{abs}$ for each star’s set of exposures (Figure \[fig:muabs\_meanintunc\]). As expected, the values for the mean of the absolute radial velocity values are always higher than the mean internal uncertainty values because $\mu_{abs}$ includes the effects of the internal uncertainty combined with additional sources of error such as the stellar jitter and instrument systematics. Additionally some of these stars are planet hosts, and thus display even higher $\mu_{abs}$ values because of the planet’s influence. However the quietest, non-planet hosting stars are able to reach $\mu_{abs}$ values of slightly less than 2 [${\rm m~s^{-1}}$]{}, suggesting that the additional error offset value determined using HD 185144 ($\sigma_{s} = 1.5$ [${\rm m~s^{-1}}$]{}) is appropriate. ![ \[fig:muabs\_meanintunc\] Mean values of $\mu_{int}$ and $\mu_{abs}$ for each star observed during the same time span as the HD 185144 analysis presented in Figure \[fig:185144\_unc\]. The $\mu_{abs}$ values are always larger because they include the internal uncertainty [*in addition*]{} to other effects such as stellar jitter and instrument systematics. For bright, quiet stars it is possible to reach $\mu_{abs}$ values of 1.8 [${\rm m~s^{-1}}$]{}.](Figure4.pdf){height="8cm"} Observing inputs {#sect:Inputs} ================ To create an efficient and scientifically informative exoplanet survey we must balance scientific interest in a range of target stars with the observing limitations presented by the weather and the telescope’s physical constraints. Parameters of scientific interest {#sect:SciParameters} --------------------------------- There are a number of criteria of astronomical importance. For each star, the following parameters are utilized by the scheduling software to determine whether the star is given a high rating for observation: - Observing priority - a numerical rating that reflects the observers’ assessment of potential Keplerian signals in the RV data. - Desired cadence - dictates the desired wait time between observations of a specific target. - Desired precision - the average allowable internal velocity uncertainty for measurements of a specific target. Both observing priority and cadence are determined via the observers’ examination of the star’s existing data set. As mentioned in Section \[sect:data\], we use the publicly available Systemic Console to analyze our RV data sets. The console produces a Lomb-Scargle periodogram from the selected RV data set, which displays peaks corresponding to periodic signals in the data. When a peak surpasses the analytic [*false alarm probability*]{} (FAP) threshold defined by this method, the observers note the period and the half-amplitude of the signal and use those values to decide upon an observing cadence and desired level of precision that will fill out the signal’s phase curve quickly and with data points of the appropriate SNR. The desired precision is further refined if the observers have some knowledge of the star’s stellar activity, as this sets a lower limit to the attainable precision. Details on the fitting procedures and the statistical capabilities of the Systemic Console are described in detail in Meschiari et. al, 2009. For each star, this information (along with other characteristics such as right ascension, declination, the V magnitude and the B-V color) is stored in an on-line Google spreadsheet accessible to team members and the telescope software. This database of target stars drives the survey design and target selection while also being easily accessed, understood and updated by observers. Relating iodine region photons to the internal uncertainty of RV values {#sect:relating_i2} ----------------------------------------------------------------------- In perfect conditions (no clouds, no loss of light due to seeing) all stars that are physically available and deemed in need of a new observation (based on their cadence) are simply ranked by observing priority and position in the night sky and then observed one after another, until dawn. Cloud cover and atmospheric turbulence, however, make such conditions rare. Furthermore, stars with low declinations spend only short periods at low air mass. Consequently there are a substantial number of constraints that affect the quality of an exposure. In order to maximize the scientific impact of each night’s exposures, the conditions must be evaluated dynamically, and data taken only when the desired precision listed in the database is likely to be attainable. To identify the targets that can be expected to attain their desired precision within the one hour observing time limit, the observing program must link the night-time conditions and the physical characteristics of each star with the resulting internal uncertainty over a given exposure time. To this end, we first relate the internal uncertainty of a given exposure to the number of photons that fall in the iodine (I$_{2}$) line-dense region of the spectrum (i.e. the $\lambda \sim 5000-6200$Å bandpass) where our radial velocity analysis is performed. We fit the relation between internal uncertainty and photons in the iodine region separately for the G and K star data set (comprised of 2790 G star exposures and 957 K star exposures) and the M star data set (comprised of 837 M star exposures). Because information on a star’s RV value comes from the location of its spectral lines, we expect the M stars - which contain many more spectral lines - to achieve better precision for the same number of photons. This expectation is validated by Figure \[fig:I2\_Uncert\] where a vertical offset between the G/K star best-fit line and the M star best-fit line is evident. Comparing the zero points for the best-fit lines (Table 1), we find a factor of 2 difference between the stellar groups. Thus the M stars reach the same level of internal uncertainty with half the number of photons required by the G and K star population. Additionally, the slopes of both best-fit lines are close to the [*m*]{}= -2 expected for shot-noise limited observations with [*m*]{}= -1.58 and -1.73 for the G/K and M stars, respectively. The fact that the slopes are shallower than [*m*]{}= -2, indicating performance better than shot-noise limited, is reasonable as the internal uncertainty only accounts for the errors resulting from the extraction of the spectrum from the original FITS files. We emphasize that this is just a piece of the error budget, and does not include other random or systematic errors. ![ \[fig:I2\_Uncert\] Observations of G (blue), K (green) and M (red) type stars during the 1.5 years of APF/Levy observations. We plot the individual 2,790 G star exposures, 957 K star exposures and 837 M star exposures. The gray diamonds represent M star data obtained using Keck/HIRES as part of the Lick-Carnegie planet survey where the Keck/HIRES pixels have been scaled to match the same Å /pixel scale as the APF/Levy ($\rm 0.0183\AA\ $ /pixel). We find that the G and K stars have the same zero point and the same slope, so we combine these two data sets for this analysis. The green and blue dashed line represents the best-fit to the APF/Levy’s combined G and K star dataset, while the red line shows the best-fit to the APF/Levy M star data set which is fit separately due to the increase of spectral lines in later spectral types. As expected, the APF/Levy M stars show higher data precision for the same number of photons in the $\rm I_{2}$ region of the spectrum. The percent errors quoted on the figure are calculated using the scatter in the difference between the observed $\rm I_{2}$ photons and the $\rm I_{2}$ photons from the best-fit lines, so they represent the scatter of the sample and not the error on the mean. The dark gray line is the best-fit to the Keck/HIRES M star data set. Comparing this to the red line reveals that the APF requires 5.75x fewer photons in the $\rm I_{2}$ region to achieve velocity precision comparable to Keck/HIRES on M stars down to at least M$\rm_{v}$=10.](Figure5.pdf){height="10cm"} The functional form of these best-fit lines is given by Equation \[eq:I2\_Unc\]. Note that the numeric values for each variable (in the case of Eq \[eq:I2\_Unc\], A and B) are listed in Table 1. This format will be used for all relations presented in §\[sect:Inputs\]. Functional form of the fits applied in Figure \[fig:I2\_Uncert\] : $$\label{eq:I2_Unc} \log_{10}(N_{med}) = A + B \cdot \log_{10}(\sigma_{int}) ,$$ The gray points in Figure \[fig:I2\_Uncert\] correspond to spectra of M stars obtained with Keck/HIRES since November 2002 as part of the Lick-Carnegie planet search. There are 168 stars represented, all with B-V $>$ 1.2, resulting in 8872 individual exposures. In order to compare these individual velocities to those obtained on the APF in a meaningful way, we rescale the Keck/HIRES pixels so that they represent the same range of Å per pixel as those on the the APF/Levy. This involves two different scaling factors as the HIRES instrument underwent a detector upgrade in 2004 that changed its pixel size from 24$\rm \mu m$ to 15$\rm \mu m$, resulting in different sampling values. Applying these factors means that all of the data shown in Figure \[fig:I2\_Uncert\] represents the median number of $\rm e^{-}$ per pixel, where each pixel covers $\rm 0.0183\AA\ $ - the native value for the APF/Levy in the $\rm I_{2}$ region. The figure shows that for M stars, the APF requires $\sim$5.75x fewer photons in the $I_{2}$ region to achieve velocity precision comparable to Keck[^2] Speed estimates for the APF/Levy, carried out last year[@Vogt2014], show that the telescope and instrument together are approximately 6x slower than Keck/HIRES. Combining these two effects indicates that the APF has essentially the same speed-on-sky as Keck/HIRES for precision RVs of M stars. This is not altogether unexpected, as HIRES was never specifically optimized for precision RV work. The APF’s Levy spectrograph was purpose-built for high-precision, RV science and therefore features much higher spectral resolution and finer wavelength sampling than HIRES. Both of these factors, as well as the significantly higher system efficiency of the APF/Levy optical train over that of Keck/HIRES[@Vogt2014], combine to make APF as fast as Keck/HIRES for precision RV work on M dwarfs, at least down to M$\rm_{v}$=10. Where N$_{med}$ is the median number of photons per pixel in the I$_{2}$ region for a given exposure, and $\sigma_{int}$ is the estimated internal uncertainty for the resulting radial velocity value. To assess whether the exposures of these stars represent a Gaussian distribution, we compare the standard RMS and the scatter calculated using a Tukey’s biweight method around each fit. The Tukey’s biweight scatter provides a more robust statistic for data drawn from a non-Gaussian distribution as it less heavily weights the outliers, which are assumed not to be part of a normal distribution [@Beers1990]. For the G and K star fit, the standard RMS is 0.182 while the biweight scatter is 0.186. Similarly, for the M stars, the standard RMS is 0.151 while the biweight scatter comes out to 0.155. In the limit of a true Gaussian distribution, these two metrics would produce the same result. Employing a bootstrap analysis of each method, we find the standard deviation of both the RMS and the biweight scatter of the G and K star sample to be 0.002. Similarly, the standard deviation of both the RMS and the biweight scatter of the M star sample is 0.003. Noting the similarity of these standard deviations with the actual offsets found between the RMS and biweight scatter, we determine that the observations for both sets of stars are drawn from a mostly Gaussian distribution. Real-time effects {#sec:realtime} ----------------- ### Data selection {#sec:dataselection} Knowing the median number of photons per pixel in the iodine region required to achieve a given level of RV precision enables us to determine the expected exposure time for a star [*if*]{} we know how quickly those photons accumulate. To determine this rate and the relation between the final exposure meter value and the number of photons in the $\rm I_{2}$ region (used to set upper bounds on observing time) we study a subset of the year of APF data described in Section \[sect:data\]. Cuts are applied to the main data set to select only those observations taken on clear nights and in photometric conditions, as non-photometric data will induce skew in the results. First, we select only exposures with seeing $\rm \overline{FWHM} < 2"$, which results in 935 individual observations. We then perform separate, multi-variate linear regressions on the photon accumulation rate in the $\rm I_{2}$ region of the spectrum and the photon accumulation rate for the exposure meter (Eq \[eq:LinRegress\_I2\], Eq \[eq:LinRegress\_Expmeter\]) on all remaining data points. In each case, we calculate the variance of deviations from the best-fitting relation for all of the data. We also calculate the variance for all of the points on a given night. We then reject nights using the F-ratio test. Namely, if the standard deviation of a given night is more than twice the standard deviation of the population as measured by the Tukey’s biweight then all observations from that night are rejected. This results in data sets of 865 exposures used in the $\rm I_{2}$ photon accumulation rate regression and 816 exposures in the exposure meter photon accumulation rate regression. Points that fall significantly below the regression line are most likely due to clouds while those falling above the line are likely due to erroneous readings from the exposure meter. Once the nights with large variance have been removed, we repeat the regressions on the remaining points to determine the actual fits described in Section \[sec:regressions\]. In Figures \[fig:LinRegress\_I2\] and \[fig:LinRegress\_Expmeter\] the points in color are those used to perform the linear regressions, while the points in gray are those we rejected after they were deemed non-photometric. Figure \[fig:I2\_Expmeter\] uses the same set of data as Figure \[fig:LinRegress\_Expmeter\] in order to keep only the normally distributed exposure meter readings. ### Linear regressions {#sec:regressions} To determine the predicted observation time of a star, we perform a multi-variate linear regression using the trimmed dataset resulting from the procedure described in the previous section. The regression estimates the rate at which photons accumulate in the pixels of the iodine region of the spectrum, and accounts for \[1\] the star’s V magnitude, \[2\] its B-V color, \[3\] slit loss due to the current seeing conditions, \[4\] the airmass based on the star’s location and \[5\] the modified date of the observation (Figure \[fig:LinRegress\_I2\]). The modified date is calculated by subtracting the maximum date from each observation following the selection process described above. This makes the zero point of the relationship the value at the time of the last photometric observation. We use the modified date parameter to address the degradation of the telescope’s mirror coatings over time. When the mirrors are recoated it will introduce a discontinuity in this parameter, and we will then adjust the zero point of all regression fits based on the new throughput estimates and watch for any changes that develop in the slope of the regressions. The multi-parameter fit over these five variables results in a best-fit plane, of which we present a projection in Figure \[fig:LinRegress\_I2\]. To help visualize the goodness of fit, we plot the data on one axis, the linear regression combination on the other and place a 1:1 line on top. This approach is also used when plotting the linear regression in Figure \[fig:LinRegress\_Expmeter\]. ![ \[fig:LinRegress\_I2\] Multi-variate linear regression of the iodine pixel photon accumulation rate which incorporates stellar color, stellar magnitude, atmospheric seeing and airmass. Colored points are used in calculating the regression, while gray points have been rejected as non-photometric data as described in Section \[sec:dataselection\]. The black line is a 1:1 relationship, and the grey dashed line shows the relation offset by one standard deviation, which is the limit we use operationally. The strong correlation between the data and the regression line enables prediction of the rate of photon accumulation in the spectrum’s iodine region (a value not calculated until the data reduction process) using the stellar properties and ambient conditions. We can thus estimate the observation time required to meet a specific median I2 photon value, and, in conjunction with Figure \[fig:I2\_Uncert\], a radial velocity precision.](Figure6.pdf){height="8cm"} The regression gives : $$\label{eq:LinRegress_I2} \log_{10}(r_{I_{2}}) = -\frac{1}{2.5}\bigg(V_{c}+\alpha_{I}(B-V)+\beta_{I}(\sec(z))+\gamma_{I}(\mathrm{MJD})+C_{I}\bigg) ,$$ with $$\label{eq:V_corr} V_{c}=V_{m} - 2.5\cdot \log_{10}(f_{t}) ,$$ where r$_{I_{2}}$ is the photon accumulation rate in the iodine region, [*z*]{} is the angle of the star relative to zenith, MJD is a modified Julian date and $f_{t}$ is the fraction of the starlight that traverses the spectrograph slit. By dividing the number of $I_{2}$ region photons necessary to meet our desired RV precision (derived from Eqn \[eq:I2\_Unc\]) with the photon accumulation rate in the $I_{2}$ region calculated using Eqn \[eq:LinRegress\_I2\], we can determine the predicted observation time for a star for a specified internal uncertainty in a given set of conditions. These predicted observation times account for atmospheric conditions such as seeing and airmass, but do not address the issue of atmospheric transparency. The APF lacks an all-sky camera with sufficient sensitivity to assess the brightness of individual stars, meaning that we cannot evaluate the relative instantaneous transparency of different regions of the sky. Instead we must determine the transparency during each individual observation by comparing the rate of photon accumulation we observe with what is expected for ideal transparency. Although the $\rm I_{2}$ region photons provide a straightforward way to determine the predicted exposure times, the number of iodine region photons is available only after the final FITS file for an observation has been reduced to yield a radial velocity measurement. Thus we cannot monitor in real time the rate at which they are registered by the detector to assess the cloud cover. Instead, we compute a transparency estimate during each exposure using the telescope’s exposure meter. The exposure meter is created by using series of 2D images from the guider camera that are updated every 1-30 seconds depending on the brightness of the target. Rather than guiding on light reflected off a mirrored slit aperture, as is traditionally done, the APF uses a beamsplitter to provide 4% of incoming light to the guider camera as a fully symmetric, unvignetted seeing disk. This allows a straightforward way to monitor how well the telescope is tracking its target and provide realtime corrections to both under and over guiding - both of which smear out the telescope’s point spread function on the CCD and result in broader full width at half maximum values for spectral lines. Guide cameras that utilize the reflected light off of a mirrored slit aperture are significantly more sensitive to these problems in good seeing, as the majority of the light falls through the spectrograph slit. In our experience, the loss of 4% of the star’s light is acceptable if it ensures that the telescope’s guiding is steady throughout the night and across the different regions of the sky. After each guiding exposure is completed, the guide camera then passes the 2D FITS images it creates to the SourceExtractor software [@Bertin1996] which analyzes each image and provides statistics on parameters such as the flux and full-width at half-maximum (FWHM), which are in turn used to evaluate the current atmospheric seeing. These guide camera images are also used to meter the exposures. Each image is integrated over the rectangular aperture corresponding to the utilized spectrograph slit, with background photons (determined using adjacent, background-estimating rectangles) subtracted off to determine the number of star-generated photons accumulated by the guide camera. Analysis of the existing APF data suggests that the exposure meter rate (much like the iodine photon accumulation rate) depends on the star’s color, its V magnitude, the atmospheric seeing (in the form of slit losses), the airmass and the date of observation. A multi-variate linear regression to the exposure meter rate over these five terms results in the correlation displayed in Figure \[fig:LinRegress\_Expmeter\]. Because the exposure meter is rapidly updated, we can monitor photon accumulation in real-time during an observation. Atmospheric seeing is already incorporated into the $V_{c}$ term in the linear regression (Eqns \[eq:V\_corr\], \[eq:LinRegress\_Expmeter\]), so any decrease from the expected exposure meter rate likely stems from an increase in clouds and corresponding decrease in atmospheric transparency. The ratio of expected exposure meter rate to observed exposure meter rate provides a “slowdown" factor that the scheduler tracks throughout the night and multiplies by the predicted ‘clear night’ observation times calculated using Eqn \[eq:LinRegress\_I2\] to determine a best guess exposure duration. ![ \[fig:LinRegress\_Expmeter\] Multi-variate linear regression to the exposure meter photon accumulation rate, as measured on the APF guider, which incorporates stellar color, stellar magnitude, atmospheric seeing and airmass. Colored points are used in calculating the regression, while gray points have been rejected as described in Section \[sec:dataselection\]. The black line is a 1:1 relation, and the grey dashed line shows the relation offset by one standard deviation, which is the limit we use operationally. The strong correlation permits prediction of the expected exposure photon accumulation rate for a given star in photometric conditions, and thus provides a measure of the transparency. Any decrease in exposure photon accumulation rate from what is predicted is presumed to arise from decreases in atmospheric transparency brought about by cloud cover.](Figure7.pdf){height="8cm"} The regression in Figure \[fig:LinRegress\_Expmeter\] gives : $$\label{eq:LinRegress_Expmeter} \log_{10}(r_{E}) = -\frac{1}{2.5}\bigg(V_{c}+\alpha_{E}(B-V)+\beta_{E}(\sec(z))+\gamma_{E}(\mathrm{MJD})+C_{E}\bigg) ,$$ where r$_{E}$ is the photon accumulation rate on the exposure meter, $\alpha_{E}$, $\beta_{E}$, $\gamma_{E}$ and C$_{E}$ have the same meanings as in Eqn \[eq:LinRegress\_I2\] but for the exposure meter photon accumulation rate instead of the Iodine photon accumulation rate, [*z*]{} is the star’s zenith angle, and V$_{c}$ is defined in Eqn \[eq:V\_corr\]. Setting upper bounds on exposure time {#sect:upperbounds} ------------------------------------- Finally, because of the scatter seen in Figures \[fig:LinRegress\_I2\] and \[fig:LinRegress\_Expmeter\], we must ensure that exposures end when photons sufficient to achieve the desired RV precision have accumulated instead of continuing on for extra seconds or minutes. Photons in the $\rm I_{2}$ region can’t be monitored during the observation. However the APF data set shows a strong relationship between the number of photons obtained in the $\rm I_{2}$ region of the spectrum and the photons registered by the exposure meter, which is monitored in realtime. The telescope’s guide camera, which is used for the exposure meter, has a broad bandpass and is unfiltered. This generates a strong color-dependent bias when comparing the guider photons to those that fall in the much narrower $\rm I_{2}$ region. We apply a quadratic B-V color correction term to produce the relation shown in Figure \[fig:I2\_Expmeter\]. Combining this with the equations identified in Figure \[fig:I2\_Uncert\], we obtain relations that allow us to relate the desired internal precision to the corresponding number of photons in the iodine region of the spectrum, and then to the number of photons required on the exposure meter. The resulting exposure meter threshold is used to place an upper limit on the exposure. This is particularly useful on nights with patchy clouds, where the cloud cover estimate calculated during the previous observation can be significantly higher than the cloud cover in other parts of the sky - resulting in artificially high predicted observation times. In this case, the exposure meter can be used to stop an observation if the desired photon count is reached early, improving efficiency. ![ \[fig:I2\_Expmeter\] Color-corrected relationship between the photons in the $\rm I_{2}$ region of the spectrum and the photons registered by the exposure meter. The black line shows the best-fit, and the grey dashed line shows the relation offset by one standard deviation, which is the limit we use operationally. To increase telescope efficiency, we require a way to ensure observations don’t continue when the number of iodine region photons necessary to achieve the desired internal uncertainty has already been achieved. As described in Section \[sec:regressions\] there is no way to measure the $\rm I_{2}$ region photon accumulation in real time. However, the tight correlation between $\rm I_{2}$ photons and photons on the exposure meter (which does update in real time) displayed here allows us to set a maximum exposure meter value based on our desired precision level. Thus the observation software will end the exposure when the specified exposure meter value is met, even if the open shutter time falls short of the predicted observation time. This is particularly useful for cases where the cloud cover used in calculating the predicted observation time is actually more than the cloud cover on the target, which would result in an erroneously long observation.](Figure8.pdf){height="8cm"} Fit applied in Figure \[fig:I2\_Expmeter\] : $$\label{eq:I2_Expmeter} \log_{10}(R) = \delta + \epsilon \rm (B-V) + \zeta \rm (B-V)^{2} ,$$ where R is the ratio of photons on the exposure meter to photons in the iodine region of the spectrum. Combining the fits {#sect:combining} ------------------ The above relations are combined to enable the scheduling algorithm to select scientifically optimal targets. The following list summarizes the combination scheme.\ [**Steps to determining an object’s predicted observation time and exposure meter cut-off**]{} 1. Query spreadsheet for stellar attributes (V, B-V, required precision, RA, Dec, observing priority). 2. Use Equation \[eq:I2\_Unc\] to determine the desired number of photons in the $I_{2}$ region of the spectrum. 3. Use Equation \[eq:LinRegress\_I2\] to calculate the observation time required to obtain the desired total number of $\rm I_{2}$ region photons in ideal transparency conditions. 4. Multiply the slowdown factor, calculated during the previous observation using Eqn \[eq:LinRegress\_Expmeter\], with the ideal transparency observation time estimates to get the predicted observation time in current conditions. 5. Calculate the exposure meter threshold based on the required number of $I_{2}$ region photons using Eqn \[eq:I2\_Expmeter\]. We find scatters of 11.5% and 11.1% in Figures \[fig:LinRegress\_I2\] and \[fig:I2\_Expmeter\] respectively (Table 1). This means that, even on a photometric night, we may not accumulate the number of photons in the $\rm I_{2}$ region necessary for the desired precision as the photon arrival rate could be too low. To increase the likelihood of getting enough photons to reach the desired number of photons in the $\rm I_{2}$ region, we increase the observation time estimate and the exposure meter threshold by 11.5% and 11.1%, or one standard deviation. By implementing this padding factor we ensure that 84% of the time we observe a target, we will obtain the desired number of $I_{2}$ photons. However, this does not necessarily guarantee that we will achieve the desired internal uncertainty, due to the scatter in the relation between the $\rm I_{2}$ region photons and the uncertainty estimates seen in Figure \[fig:I2\_Uncert\] and quantified in Table 1. Using these predicted observation times, the scheduler can evaluate whether any potential target can be observed at its desired precision within the one-hour observation time limit. Combining the predicted observation times with the targets’ coordinates determines whether it will remain within the allowed $20-85^{\circ}$ elevation range during the exposure. Stars that satisfy all these criteria are then ranked based on their observing priority, time past cadence requirement, and distance from the moon, with the highest scoring star being selected for observation. The scheduler then transmits the necessary information for the selected star, including its expected observation time and exposure meter threshold, to the observing software, breaking up the total observation time into a number of individual exposures if necessary. Dismissed factors {#sect:DismissedInputs} ================= Exposures obtained during 2013 and 2014 indicate that some factors initially suspected to be important need not impact target selection considerations. For example, the original observing protocol avoided targets within 45$^{\circ}$ of the direction of any wind above 5 mph to avoid wind shake in the telescope. We find, however, no discernible increase in the internal uncertainty (indicated by the color scale in Figure \[fig:Wind\_plot\]) as a function of wind speed or direction. This resilience likely stems from an effective wind shielding mode for the dome shutters, which opens them just enough to ensure that there is no vignetting of the target star [@Vogt2014]. In addition, substantial effort has been put into tuning the telescope’s servo motors in order to “stiffen" the telescope and thus mitigate the effect of wind gusts that do manage to enter the dome [@Lanclos2014]. ![ \[fig:Wind\_plot\] Windspeed (point size) and direction (azimuthal position) plotted for 3155 individual exposures reveals no strong correlation between pointing near/into the wind and the estimated internal uncertainty displayed in the color bar. The exposures represented in this figure were obtained before we had the means to determine condition-based exposure times. Thus all exposures were run until the reached their exposure meter threshold, or up to a static maximum exposure time of 900s and then terminated, regardless of the number of photons collected by the exposure meter. This means the wind based effects are not being mitigated by longer exposure times, and wind direction can thus be ignored when deciding which stars are considered viable targets for the next observation.](Figure9.pdf){height="8cm"} We also previously assigned higher priority to targets with elevation in the 60-70$^{\circ}$ range, removing scientifically interesting targets that were closer to the horizon from consideration. As shown in Figure \[fig:El\_error\], however, there is no significant loss in velocity precision as a function of elevation from 90$^{\circ}$ down to 20$^{\circ}$ thanks to the telescope’s atmospheric dispersion corrector (ADC) which works down to 15$^{\circ}$. The telescope has a hard observing limit of 15$^{\circ}$ because of the ADC’s range and because working at lower elevations leads to vignetting by the dome shutters. We still enforce an elevation range of $20-85^{\circ}$ to avoid mechanical problems in telescope tracking at the high elevations.. Working at elevations approaching our lower limit does result in longer predicted observation times (due to the airmass term) which can result in stars being skipped over in favor of other, higher elevation targets. ![ \[fig:El\_error\] The radial velocity precision as a function of the elevation shows no strong correlation, once we compare observations with a fixed exposure meter value. As our linear regressions in Figure 4 and 5 account for the decline in the photon accumulation rate with decreasing elevation, we do not need to add an additional term to account for other elevation effects such as increased seeing. Similar to the Figure \[fig:Wind\_plot\], the observations presented here were taken before the adaptive exposure time software was in use. Thus every exposure has a static maximum exposure time of 900s and the low elevation effects are not being mitigated by allowing for longer exposures. The telescope’s ADC only functions down to 15$^{\circ}$, the same elevation at which the telescope begins to vignette on the dome shutter, thus providing the lower limit for our observations.](Figure10.pdf){height="8cm"} Finally, we no longer assign a weighting value to the slew time necessary to move between targets. The APF is capable of moving at 3[${\rm ^{\circ}~s^{-1}}$]{} in azimuth and 2[${\rm ^{\circ}~s^{-1}}$]{} in elevation, which means that a direct slew to a target across the sky would take only one minute. Because the CCD takes approximately 40 seconds to read out each observation, this slew time factor is small enough to be considered unimportant. Furthermore, with the introduction of the wind shielding mode, the telescope’s movement was altered so that it must first drop to a “safe elevation" of $15^{\circ}$ before rotating to the target azimuth and then moving upwards to the appropriate elevation. This is done to protect the primary mirror from falling debris while the dome shutters are reconfigured to minimize wind-effects for the next observation during the slew. An additional result is that all slews take approximately the same amount of time, which provides justification for discounting slew time as an input to target selection. Dynamic scheduler overview {#sect:Heimdallr} ========================== In Section \[sect:Inputs\], we described a method to predict observation times for targets given their precision requirements and the current atmospheric conditions. To automate the determination of these observation times and the selection of the optimal target at any time throughout the night we have implemented a dynamic scheduler (written in Python) called [Heimdallr]{}. [Heimdallr]{} runs all of the APF’s target selection efforts and interfaces with pre-existing telescope control software so that, once it submits an observation request, it waits until that set of exposures is completed before reactivating. Telescope safety, system integrity, and alerts about current weather conditions are monitored by two services called [apfmon]{} and [checkapf]{}. Each of these software sets has the ability to override [Heimdallr]{} if they detect conditions that pose a threat to, or represent a problem with, the telescope. This ensures that the facility’s safety is always given priority. Additionally, while directing the night’s operations, [Heimdallr]{} uses other pre-existing utilities including [openatnight]{}, [prep-obs]{} and [closeup]{} which, as their names suggest, open the facility prior to nightfall (or when night time conditions warrant), prepare the instrument and optical train for observing, and close the facility, securing the telescope when conditions warrant. The setting of the guider camera, the configuration of dome shutters and the control of telescope movement to avoid interference with the cables wrapped around the telescope base is handled by yet another utility called [scriptobs]{}. Observing description {#sect:observing} --------------------- Typically, [Heimdallr]{} initiates in the afternoon and prompts the instrument control software to focus the instrument by obtaining a dewar focus cube (series of exposures of the quartz halogen lamp taken through the Iodine cell) using the standard observation slit. Once the software determines a satisfactory value for the instrument’s dewar position (via a simple linear least squares parabolic fit to the focus values), it proceeds to take all of the calibration exposures that are required by the data reduction pipeline. Upon completing these tasks, [Heimdallr]{} then waits until dusk, at which point it consults [checkapf]{} to ensure that there are no problematic conditions in the weather and then [apfmon]{} to ensure facility readiness. If both systems report safe conditions, it then opens the dome, allowing the telescope to thermalize with the outside air. [Heimdallr]{} runs a main loop that continuously monitors a variety of keywords supplied by the telescope. At 6$^{\circ}$ twilight, the telescope is prepped for observing and begins by choosing a rapidly rotating B star from a pre-determined list. The B star has no significant spectral lines in the $\rm I_{2}$ region and serves as a focus source for the telescope’s secondary mirror while also allowing the software to determine the current atmospheric conditions (as described in Section \[sec:regressions\]). At 9$^{\circ}$ twilight, [Heimdallr]{} accesses the online database of potential targets and parses it to obtain all the static parameters described in Section 3.1. It then checks the current date and time and eliminates from consideration those stars not physically available. The scheduler then employs the stars’ coordinates, B-V colors and V magnitudes, combined with the seeing and atmospheric transparency determined during the previous observation, to calculate the predicted observation time for each target given their desired precision levels. Stars unable to reach the desired precision within the one hour maximum observation time are eliminated from the potential target list. The remaining stars are ranked based on scientific priority and time past observing cadence. The star with the highest score is passed to [scriptobs]{} to initiate the exposure(s). When the exposures finish, [Heimdallr]{} updates the star’s date of last observation in the database with the photon-weighted midpoint of the last exposure and begins the selection process anew. When the time to 9$^{\circ}$ morning twilight becomes short enough that no star will achieve its desired level of precision before the telescope must close, [Heimdallr]{} shuts the telescope using [closeup]{} and initiates a series of post-observing calibration exposures. Once the calibrations finish, [Heimdallr]{} exits. Other operational modes {#sect:modes} ----------------------- In addition to making dynamic selections from the target database during the night, [Heimdallr]{} can also be initiated in a fixed list or ranked fixed list operational mode. The fixed list mode allows observers to design a traditional starlist that [Heimdallr]{} will move through, sending one line at a time to [scriptobs]{}. Any observations that are not possible (due to elevation constraints) will be skipped and the scheduler will simply move to the next line. The ranked fixed list option allows users to provide a target list that [Heimdallr]{} parses to determine the optimal order of observations. That is, after finishing one observation from the list, the scheduler will then perform a weighting algorithm similar to what is employed by the dynamic use mode to determine which line of the target list is best observed next. [Heimdallr]{} keeps track of all the lines it has already selected, so that they do not get initiated twice, and will re-analyze the remaining lines after each observation to select the optimal target. This option is especially useful for observing programs that have a large number of possible targets but don’t place a strong emphasis on which ones are observed during a given night. Comparing with other facilities {#sect:compare} =============================== There are a number of automated and semi-automated facilities that perform similar observations. Examples include HARPS-N, CARMENES, the Robo-AO facility at Palomar and the Las Cumbres Observatory Global Telescope (LCOGT) network. In order to place the APF’s operations in context relative to these other observatories, we will briefly discuss the approaches of these other dedicated radial velocity facilities (CARMENES, HARPS-N, and the NRES addition to LCOGT) and of the more general facility with queue scheduling (Robo-AO). We will then highlight the common approaches along with discussing what is simpler for our facility as it is dedicated to a single-use. RV surveys {#sect:auto_rv} ---------- Radial Velocity surveys generally require tens or hundreds of observations of the same star to detect planetary companions. Traditionally, high-precision velocities were obtained on shared-use facilities and observing time was limited. More recently, however, purpose built systems (such as HARPS) have presented the opportunity to obtain weeks or months of contiguous nights. The HARPS-N instrument, installed on the Telescopio Nazionale Galileo at Roque de los Muchachos Observatory in the Canary Islands, is a premier system for generating high-precision velocities. At present, it is primarily devoted to Kepler planet candidate follow up and confirmation. The system allows users to access XML standard format files that define target objects using the Short-Term-Scheduler GUI (STS) to guide the process and then assemble the objects into an observing block. When an observer initiates an observation, these blocks and their associated observing preferences are passed to the HARPS-N Sequencer software, which places all of the telescope subsystems into the appropriate states, performs the observation, and then triggers the data reduction process [@Cosentino2012]. Thus, while the building of target lists has been streamlined, nightly observation planning still requires attention from an astronomer or telescope technician. The forth-coming Calar Alto CARMENES instrument is expecting first light in 2015.\ CARMENES will also be used to obtain high-precision stellar radial velocity measurements on low-mass stars. Its automated scheduling mechanism relies on a two-pronged approach: the off-line scheduler which plans observations on a weekly to nightly time scale based on target constraints that can be known in advance, and the on-line scheduler which is called during the evening if unexpected weather or mechanical situations arise and require adapting the previously calculated nightly target list [@Garcia-Piquer2014]. Finally, the LCOGT network will soon implement the Network of Robotic Echelle Spectrographs (NRES), six high resolution optical echelle spectrographs slated for operation in late 2015. Like all instruments deployed on the LCOGT network, the scheduling and observing of NRES will be autonomous. Users submit observing requests via a web interface which are then passed to an adaptive scheduler which works to balance the requests’ observing windows with the hard constraints of day and night, target visibility, and any other user specified constraints (e.g. exposure time, filter, airmass). If the observation is selected by the adaptive scheduler, it creates a “block" observation tied to a specific telescope and time. This schedule is constructed 7 days out, but rescheduling can occur during the night if one or more telescopes become unavailable due to clouds, or if new observing requests arrive or existing requests are canceled. In this event, the schedule is recalculated, and observations are reassigned among the remaining available sites [@Eastman2014]. Automated queue scheduling {#sect:queue} -------------------------- Robo-AO is the first fully automated laser guide star adaptive optics instrument. It employs a fully automated queue scheduling system that selects among thousands of potential targets at a time with an observation rate of $\sim20$ objects $\rm hr^{-1}$. Its queue scheduling system employs a set of XML format files which use keywords to determine the required settings and parameters for an observation. When requested, the queue system runs each of the targets through a selection process, which first eliminates those objects that cannot be observed, and then assigns a weight to the remaining targets to determine their priority in the queue at that time. The optimal target is chosen, and the scheduler passes all observation information to the robotic system and waits for a response that the observations were successfully executed. Once the response is received, the observations are marked as completed and the relevant XML files are updated [@Riddle2014]. Comparing our approach to other efforts {#sect:compare_sub} --------------------------------------- We have designed this system based on our RV observing program carried out over the past 20 years at Keck and other facilities. This experience, coupled with the preexisting software infrastructure, has guided the development of both the dynamic scheduler software itself and our observing strategy. Comparison to these other observing facilities and the strategies they employ emphasizes some shared design decisions. For example, the APF has a similar target selection approach to that of the Robo-AO system. Both utilize a variety of user specified criteria to eliminate those targets unable to meet the requirements and then rank the remaining targets, passing the object with the highest score to the observing software. Additionally, as is common with all of the observing efforts mentioned above, our long term strategy is driven by our science goals and is in the hands of the astronomers involved with the project. Several differences are also notable. The first is that we lack an explicit long term scheduling component in our software. Our observing decisions are made in real time in order to address changes in the weather and observing conditions that occur on minute to hour time scales, and to maximize the science output of nights impacted by clouds or bad seeing. However, for a successful survey there must also be a longterm observing strategy for each individual target. We address this need via a desired cadence and required precision for every potential target. By incorporating the knowledge of how often each star needs to be observed and a way to assess whether the evening’s conditions are amenable to achieving the desired precision level, [Heimdallr]{} adheres to the longterm observing strategy outlined by our observing requirements and doesn’t need to generate separate multi-week observing lists. A second difference is that the final output of the scheduler is a standard star list text file, one line in length. This format has been in use in UCO-supported facilities such as Lick and Keck Observatories for more than 20 years, and thus is familiar to the user community. The file is a simple ASCII text file with key value pairs for parameters and a set of fixed fields for the object name and coordinates. This permits ready by-eye verification of the next observation if desired and allows the user to quickly construct a custom observing line that can be inserted into the night’s operations if needed. Observers can furthermore easily make a separate target list in this format for observations when not using the dynamic scheduler (see Sec \[sect:modes\]) Finally, we use a Google spreadsheet for storing target information and observing requirements, as opposed to a more machine-friendly format such as XML. Although Lick Observatory employs a firewall to sharply limit access to the APF hosts, it only operates on the incoming direction. Thus it is straightforward and non-compromising from a security standpoint for our internal computers to send a request to Google and pull the relevant data back onto the mountain machines. This approach provides team members with an accessible, easy-to-read structure that is familiar and easily exportable to a number of other formats. Google’s version control allows for careful monitoring of changes made to the spreadsheet and ensures that any accidental alterations can be quickly and easily undone. The use of the Google software also allows interaction with a browser, so no custom GUI development is required. Therefore, we are taking advantage of existing software to both minimize our development effort, and make it as easy as possible to have the scientists update and maintain the core data files that control the observations. Observing campaigns on the automated planet finder {#sect:campaigns} ================================================== The APF has operated at high precision for over a year, and has demonstrated precision levels of $\rm \sigma \sim 1$ [${\rm m~s^{-1}}$]{}on bright, quiet stars such as Sigma Draconis[@Vogt2014]. The telescope’s slew rate permits readout-limited cycling and 80-90% open shutter time, allowing for 50-100 Doppler measurements on clear nights. The Lick-Carnegie survey {#sect:LCsurvey} ------------------------ [Heimdallr]{}’s design dovetails with the need to automate the continued selective monitoring of more than 1,000 stars observed at high Doppler precision at Keck over the past 20 years [@Butler2015]. The APF achieves a superior level of RV precision and much-improved per-photon efficiency in comparison to Keck/HIRES for target stars with V &lt;10. As a primary user, we can employ the APF on 100+ nights per year in the service of an exoplanet detection survey. At present, 127 stars have been prioritized for survey-mode observation with the APF. This list emphasizes stars that benefit from the telescope’s more northern location, and gives preference to stars that display prior evidence of planetary signals. We adopt a default cadence of 0.5 observations per night. When a star is selected, it is observed in a set of 5-15 minute exposures with the additional constraint that the total of a night’s sequential exposures (the observation time) on the star is less than an hour. Additionally, information obtained at Keck has, in some cases, permitted estimates for the stellar activity of specific target stars. In these cases, observation times can be adjusted to conform to a less stringent desired precision. With its ability to predict observation times, [Heimdallr]{} readily achieves efficiencies that surpass the use of fixed lists, and indeed, its performance is comparable to that of a human observer monitoring conditions throughout the night. TESS pre-covery survey {#sect:TESSsurvey} ---------------------- [Heimdallr]{} is readily adopted to oversee a range of observational programs, and a particularly attractive usage mode for APF arises in connection with NASA’s Transiting Exoplanet Survey Satellite (TESS) mission. TESS is scheduled for launch in 2017, and is the next transit photometry planet detection mission in NASA’s pipeline [@Ricker2014]. Transit photometry observations of potential planet signatures generally require follow-up confirmation, with RV being the most common. Currently there is a dearth of high-precision RV facilities in the northern hemisphere, and HARPS-N is heavily committed to Kepler planet candidate follow up. NASA recently announced plans to develop an instrument for the 3.5m WIYN telescope at Kitt Peak Observatory capable of extreme precision Doppler spectrography to be used for follow up of TESS planet candidates [@WIYN]. However such an instrument will require time for development and commissioning and thus is not a viable candidate for pre-launch observations of TESS target stars. The APF’s year-round access to the bright stars near the north ecliptic pole, which will obtain the most observation time from TESS, makes it an optimal facility to conduct surveys in support of the satellite’s planet detection mission. At present, comparatively little is known about the majority of TESS’s target stars. We have little advance knowledge of which stars host properly inclined, short-period transiting planets observable by the satellite. We will thus start with a default value for the observing cadence and be poised to adapt quickly should hints of planetary signatures start to emerge. Additionally, initial desired precisions (and the corresponding observation times) must rest on the suspected stellar jitter of the targets. The availability of on-line databases to track all of the science-based criteria will be crucial for moving back and forth quickly between this project and further Lick-Carnegie follow up. Evaluations of the value of APF coverage are governed by three scientific criteria (priority, cadence and required precision), along with supporting physical characteristics (RA, Dec, V and B-V) for each target. Thus when beginning RV support for TESS, [Heimdallr]{} can easily be instructed to reference the TESS database when determining the next stellar target (instead of the Lick-Carnegie List). TESS observations also provide an excellent test bed for experimenting with alternate observational strategies. For example, Sinukoff et al. (2014)[@Sinukoff2014] stated that obtaining three 5-minute exposures of a star spaced approximately two hours apart from one another during the night results in a 10% increase in precision over taking a singe 15 minute exposure. The TESS stars that will be monitored by the APF are all located in the north ecliptic pole region, meaning that the slew times will be almost negligible. It is thus likely that an observing mode that subdivides exposures to improve precision could be very valuable. In short, the APF is extremely well matched to the TESS Mission. We are pleased to acknowledge support from the NASA TESS Mission through MIT sub award \#5710003702. We are also grateful for partial support of this work from NSF grant AST-0908870 to SSV, and NASA grant NNX13AF60G to RPB. This material is based upon work supported by the National Aeronautics and Space Administration through the NASA Astrobiology Institute under Cooperative Agreement Notice NNH13ZDA017C issued through the Science Mission Directorate. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. [10]{} I. [Zolotuhkin]{}, “[The Extrasolar Planets Encyclopaedia]{},” (1995). . S. S. [Vogt]{}, M. [Radovan]{}, R. [Kibrick]{}, R. P. [Butler]{}, B. [Alcott]{}, S. [Allen]{}, P. [Arriagada]{}, M. [Bolte]{}, J. [Burt]{}, J. [Cabak]{}, K. [Chloros]{}, D. [Cowley]{}, W. [Deich]{}, B. [Dupraw]{}, W. [Earthman]{}, H. [Epps]{}, S. [Faber]{}, D. [Fischer]{}, E. [Gates]{}, D. [Hilyard]{}, B. [Holden]{}, K. [Johnston]{}, S. [Keiser]{}, D. [Kanto]{}, M. [Katsuki]{}, L. [Laiterman]{}, K. [Lanclos]{}, G. [Laughlin]{}, J. [Lewis]{}, C. [Lockwood]{}, P. [Lynam]{}, G. [Marcy]{}, M. [McLean]{}, J. [Miller]{}, T. [Misch]{}, M. [Peck]{}, T. [Pfister]{}, A. [Phillips]{}, E. [Rivera]{}, D. [Sandford]{}, M. [Saylor]{}, R. [Stover]{}, M. [Thompson]{}, B. [Walp]{}, J. [Ward]{}, J. [Wareham]{}, M. [Wei]{}, and C. [Wright]{}, “[APF-The Lick Observatory Automated Planet Finder]{},” [*Publications of the ASP*]{} [**126**]{}, 359–379 (2014). B. J. [Fulton]{}, L. M. [Weiss]{}, E. [Sinukoff]{}, H. [Isaacson]{}, A. W. [Howard]{}, G. W. [Marcy]{}, G. W. [Henry]{}, B. P. [Holden]{}, and R. I. [Kibrick]{}, “[Three Super-Earths Orbiting HD 7924]{},” [*Astrophysical Journal*]{} [**805**]{}, 175 (2015). M. V. [Radovan]{}, G. F. [Cabak]{}, L. H. [Laiterman]{}, C. T. [Lockwood]{}, and S. S. [Vogt]{}, “[A radial velocity spectrometer for the Automated Planet Finder Telescope at Lick Observatory]{},” in [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{}, [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{} [**7735**]{}, 4 (2010). N. C. [Santos]{}, F. [Bouchy]{}, M. [Mayor]{}, F. [Pepe]{}, D. [Queloz]{}, S. [Udry]{}, C. [Lovis]{}, M. [Bazot]{}, W. [Benz]{}, J.-L. [Bertaux]{}, G. [Lo Curto]{}, X. [Delfosse]{}, C. [Mordasini]{}, D. [Naef]{}, J.-P. [Sivan]{}, and S. [Vauclair]{}, “[The HARPS survey for southern extra-solar planets. II. A 14 Earth-masses exoplanet around [$\mu$]{} Arae]{},” [*Astronomy & Astrophysics*]{} [**426**]{}, L19–L23 (2004). X. [Dumusque]{}, N. C. [Santos]{}, S. [Udry]{}, C. [Lovis]{}, and X. [Bonfils]{}, “[Planetary detection limits taking into account stellar noise. II. Effect of stellar spot groups on radial-velocities]{},” [*Astronomy & Astrophysics*]{} [**527**]{}, A82 (2011). J. [Burt]{}, S. S. [Vogt]{}, R. P. [Butler]{}, R. [Hanson]{}, S. [Meschiari]{}, E. J. [Rivera]{}, G. W. [Henry]{}, and G. [Laughlin]{}, “[The Lick-Carnegie Exoplanet Survey: Gliese 687 - A Neptune-mass Planet Orbiting a Nearby Red Dwarf]{},” [*The Astrophysical Journal*]{} [**789**]{}, 114 (2014). R. W. [Noyes]{}, L. W. [Hartmann]{}, S. L. [Baliunas]{}, D. K. [Duncan]{}, and A. H. [Vaughan]{}, “[Rotation, convection, and magnetic activity in lower main-sequence stars]{},” [*Astrophysical Journal*]{} [**279**]{}, 763–777 (1984). S. [Meschiari]{}, A. S. [Wolf]{}, E. [Rivera]{}, G. [Laughlin]{}, S. [Vogt]{}, and P. [Butler]{}, “[Systemic: A Testbed for Characterizing the Detection of Extrasolar Planets. I. The Systemic Console Package]{},” [*Publications of the ASP*]{} [**121**]{}, 1016–1027 (2009). M. [Wenger]{}, F. [Ochsenbein]{}, D. [Egret]{}, P. [Dubois]{}, F. [Bonnarel]{}, S. [Borde]{}, F. [Genova]{}, G. [Jasniewicz]{}, S. [Lalo[ë]{}]{}, S. [Lesteven]{}, and R. [Monier]{}, “[The SIMBAD astronomical database. The CDS reference database for astronomical objects]{},” [*Astronomy and Astrophysics*]{} [**143**]{}, 9–22 (2000). R. P. [Butler]{}, G. W. [Marcy]{}, E. [Williams]{}, C. [McCarthy]{}, P. [Dosanjh]{}, and S. S. [Vogt]{}, “[Attaining Doppler Precision of 3 m $\rm s^{-1}$]{},” [ *Publications of the ASP*]{} [**108**]{}, 500 (1996). J. T. [Wright]{}, “[Radial Velocity Jitter in Stars from the California and Carnegie Planet Search at Keck Observatory]{},” [*Publications of the ASP*]{} [**117**]{}, 657–664 (2005). P. [Connes]{}, “[Absolute astronomical accelerometry]{},” [*Astrophysics and Space Science*]{} [**110**]{}, 211–255 (1985). F. [Bouchy]{}, F. [Pepe]{}, and D. [Queloz]{}, “[Fundamental photon noise limit to radial velocity measurements]{},” [*Astronomy & Astrophysics*]{} [**374**]{}, 733–739 (2001). S. S. [Vogt]{}, S. L. [Allen]{}, B. C. [Bigelow]{}, L. [Bresee]{}, B. [Brown]{}, T. [Cantrall]{}, A. [Conrad]{}, M. [Couture]{}, C. [Delaney]{}, H. W. [Epps]{}, D. [Hilyard]{}, D. F. [Hilyard]{}, E. [Horn]{}, N. [Jern]{}, D. [Kanto]{}, M. J. [Keane]{}, R. I. [Kibrick]{}, J. W. [Lewis]{}, J. [Osborne]{}, G. H. [Pardeilhan]{}, T. [Pfister]{}, T. [Ricketts]{}, L. B. [Robinson]{}, R. J. [Stover]{}, D. [Tucker]{}, J. [Ward]{}, and M. Z. [Wei]{}, “[HIRES: the high-resolution echelle spectrometer on the Keck 10-m Telescope]{},” in [*Instrumentation in Astronomy VIII*]{}, D. L. [Crawford]{} and E. R. [Craine]{}, Eds., [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{} [**2198**]{}, 362 (1994). T. C. [Beers]{}, K. [Flynn]{}, and K. [Gebhardt]{}, “[Measures of location and scale for velocities in clusters of galaxies - A robust approach]{},” [ *Astronomical Journal*]{} [**100**]{}, 32–46 (1990). E. [Bertin]{} and S. [Arnouts]{}, “[SExtractor: Software for source extraction.]{},” [*Astronomy and Astrophysics*]{} [**117**]{}, 393–404 (1996). K. [Lanclos]{}, M. [Peck]{}, M. [Saylor]{}, R. I. [Kibrick]{}, and S. L. [Allen]{}, “[Tuning a 2.4-meter telescope... blindfolded]{},” in [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{}, [ *Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{} [**9145**]{}, 4 (2014). R. [Cosentino]{}, C. [Lovis]{}, F. [Pepe]{}, A. [Collier Cameron]{}, D. W. [Latham]{}, E. [Molinari]{}, S. [Udry]{}, N. [Bezawada]{}, M. [Black]{}, A. [Born]{}, N. [Buchschacher]{}, D. [Charbonneau]{}, P. [Figueira]{}, M. [Fleury]{}, A. [Galli]{}, A. [Gallie]{}, X. [Gao]{}, A. [Ghedina]{}, C. [Gonzalez]{}, M. [Gonzalez]{}, J. [Guerra]{}, D. [Henry]{}, K. [Horne]{}, I. [Hughes]{}, D. [Kelly]{}, M. [Lodi]{}, D. [Lunney]{}, C. [Maire]{}, M. [Mayor]{}, G. [Micela]{}, M. P. [Ordway]{}, J. [Peacock]{}, D. [Phillips]{}, G. [Piotto]{}, D. [Pollacco]{}, D. [Queloz]{}, K. [Rice]{}, C. [Riverol]{}, L. [Riverol]{}, J. [San Juan]{}, D. [Sasselov]{}, D. [Segransan]{}, A. [Sozzetti]{}, D. [Sosnowska]{}, B. [Stobie]{}, A. [Szentgyorgyi]{}, A. [Vick]{}, and L. [Weber]{}, “[Harps-N: the new planet hunter at TNG]{},” in [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{}, [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{} [**8446**]{}, 1 (2012). A. [Garcia-Piquer]{}, J. [Gu[à]{}rdia]{}, J. [Colom[é]{}]{}, I. [Ribas]{}, L. [Gesa]{}, J. C. [Morales]{}, A. [P[é]{}rez-Calpena]{}, W. [Seifert]{}, A. [Quirrenbach]{}, P. J. [Amado]{}, J. A. [Caballero]{}, and A. [Reiners]{}, “[CARMENES instrument control system and operational scheduler]{},” in [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{}, [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{} [**9152**]{}, 21 (2014). J. D. [Eastman]{}, T. M. [Brown]{}, J. [Hygelund]{}, J. [van Eyken]{}, J. R. [Tufts]{}, and S. [Barnes]{}, “[NRES: the network of robotic Echelle spectrographs]{},” in [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{}, [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{} [**9147**]{}, 16 (2014). R. L. [Riddle]{}, K. [Hogstrom]{}, A. [Papadopoulos]{}, C. [Baranec]{}, and N. M. [Law]{}, “[The Robo-AO automated intelligent queue system]{},” in [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{}, [ *Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{} [**9152**]{}, 1 (2014). R. P. [Butler]{}, “[The HIRES/Keck Precision Radial Velocity Exoplanet Survey (in prep)]{},” [*Astrophysical Journal*]{} (2016). G. R. [Ricker]{}, J. N. [Winn]{}, R. [Vanderspek]{}, D. W. [Latham]{}, G. [Á]{}. [Bakos]{}, J. L. [Bean]{}, Z. K. [Berta-Thompson]{}, T. M. [Brown]{}, L. [Buchhave]{}, N. R. [Butler]{}, R. P. [Butler]{}, W. J. [Chaplin]{}, D. [Charbonneau]{}, J. [Christensen-Dalsgaard]{}, M. [Clampin]{}, D. [Deming]{}, J. [Doty]{}, N. [De Lee]{}, C. [Dressing]{}, E. W. [Dunham]{}, M. [Endl]{}, F. [Fressin]{}, J. [Ge]{}, T. [Henning]{}, M. J. [Holman]{}, A. W. [Howard]{}, S. [Ida]{}, J. [Jenkins]{}, G. [Jernigan]{}, J. A. [Johnson]{}, L. [Kaltenegger]{}, N. [Kawai]{}, H. [Kjeldsen]{}, G. [Laughlin]{}, A. M. [Levine]{}, D. [Lin]{}, J. J. [Lissauer]{}, P. [MacQueen]{}, G. [Marcy]{}, P. R. [McCullough]{}, T. D. [Morton]{}, N. [Narita]{}, M. [Paegert]{}, E. [Palle]{}, F. [Pepe]{}, J. [Pepper]{}, A. [Quirrenbach]{}, S. A. [Rinehart]{}, D. [Sasselov]{}, B. [Sato]{}, S. [Seager]{}, A. [Sozzetti]{}, K. G. [Stassun]{}, P. [Sullivan]{}, A. [Szentgyorgyi]{}, G. [Torres]{}, S. [Udry]{}, and J. [Villasenor]{}, “[Transiting Exoplanet Survey Satellite (TESS)]{},” in [ *Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{}, [*Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*]{} [**9143**]{}, 20 (2014). T. [Beal]{}, “[NASA needs Kitt Peak telescope for exoplanet duty]{},” (2014). . E. [Sinukoff]{}, “[Optimization of Planet Finder Observing Strategy]{},” in [ *Search for Life Beyond the Solar System. Exoplanets, Biosignatures & Instruments*]{}, D. [Apai]{} and P. [Gabor]{}, Eds., 4P (2014). [^1]: As stated, the internal uncertainties and velocities used in this analysis are extracted from the individual spectra obtained by the telescope. Our normal operational mode determines the internal uncertainties from data binned on a two hour time scale, which is the approach used by Vogt et al. (2014)[@Vogt2014] and yields a standard deviation of $\sigma_{bin} = 1.05$ [${\rm m~s^{-1}}$]{} for HD 185144. This value is notably smaller than the standard deviation of the individual velocities, $\sigma_{vel} = 2.3$ [${\rm m~s^{-1}}$]{}, because we deliberately acquired 6 observations of HD 185144 in order to both average over the pulsation modes and achieve a high precision for the final binned observations. [^2]: Based on the work of Bouchy et al. (2001)[@Bouchy2001], we expect that for K dwarfs the relative speed should scale as the “information content” $Q$[@Connes1985], which is proportional to the ratio of the resolutions, $Q \propto (R_{APF}/R_{HIRES})$. For HIRES, the “throughput” (the resolving power times the angular size of the slit) is 39,000“ [@Vogt1994] and for the APF it is 114,000”[@Vogt2014]. Normally HIRES was used with the 0.861“ slit giving a filled aperture resolution of 45k while for the Levy a 1” slit is used, so the ratio of the resolutions is 2.5. The Levy demonstrates a larger than expected improvement over HIRES which could be explained by the increased number of lines in the iodine region for M dwarfs. Further investigation is beyond the scope of the current paper but we plan to include such analysis in a future publication. We note that the excellent seeing at MK means that some data were observed with a much higher effective R, up to 90k, which may explain the large scatter we see in Figure 5.

ArXiv

--- abstract: 'Let ${\mathscr{I}}$ be an ideal sheaf on ${\mathbb{P}}^n$ defining a subscheme $X$. Associated to ${\mathscr{I}}$ there are two elementary invariants: the invariant $s$ which measures the positivity of ${\mathscr{I}}$, and the minimal number $d$ such that ${\mathscr{I}}(d)$ is generated by its global sections. It is now clear that the asymptotic behavior of $\operatorname{reg}{\mathscr{I}}^t$ is governed by $s$ but usually not linear. In this paper, we first describe the linear behavior of the asymptotic regularity by showing that if $s=d$, i.e., $s$ reaches its maximal value, then for $t$ large enough $\operatorname{reg}{\mathscr{I}}^t=dt+e$ for some positive constant $e$. We then turn to concrete geometric settings to study the asymptotic regularity of ${\mathscr{I}}$ in the case that $X$ is a nonsingular variety embedded by a very ample adjoint line bundle. Our approach also gives regularity bounds for ${\mathscr{I}}^t$ once we know $\operatorname{reg}{\mathscr{I}}$ and assume that $X$ is a local complete intersection.' address: 'Department of Mathematics, Statistics, and Computer Science, University of Illinois at Chicago, 851 South Morgan Street, Chicago, IL 60607-7045, USA' author: - Wenbo Niu title: Some Results on Asymptotic Regularity of Ideal Sheaves --- Introduction ============ The motivation of this paper is to understand several interesting phenomena arose in recent research on the asymptotic Castelnuovo-Mumford regularity of an ideal sheaf. Throughout the paper, we work over the field of complex number ${\mathbb{C}}$ and a variety is always reduced and irreducible. Let us first mention an interesting result on the asymptotic regularity of a homogeneous ideal, for which the picture seems rather clear now thanks to the effort of many people. Suppose that $I$ is a homogeneous ideal in a polynomial ring. Swason [@Swanson:PowOfIdelas] first observed that for $t$ sufficiently large $\operatorname{reg}I^t$ is bounded by a linear function $dt+e$ for some constant $d$ and $e$. An effective result was soon established independently by Kodiyalam [@Kodiyalam:AymReg] and Cutkosky, Herzog and Trung [@Cutkosky:AsymReg], which says that for $t$ sufficiently large one actually can get $$\operatorname{reg}I^t=dt+e.$$ The slope $d$ in the formula has a concrete algebraic meaning in the result. However much less is known about the constant $e$ in the equality even recently. Turning to the geometric case, suppose that ${\mathscr{I}}$ is an ideal sheaf on the projective space ${\mathbb{P}}^n$. The shape of the asymptotic regularity of ${\mathscr{I}}$ was first described by Cutkosky, Ein and Lazarsfeld [@Ein:PosiComlIdeal] in the formula $$\lim_{t\rightarrow \infty}\frac{\operatorname{reg}{\mathscr{I}}^t}{t}=s.$$ The number $s$ is called the $s$-invariant of ${\mathscr{I}}$ and measures, roughly speaking, the positivity of ${\mathscr{I}}$. Thus this formula gives us an impression that the positivity of an ideal sheaf controls its asymptotic regularity. Several examples in [@Cutkosky:AsymReg] and [@Ein:PosiComlIdeal] have shown that $s$ could be an irrational number which prevents the asymptotic regularity from being linear. In general, the best one can hope for, is that the asymptotic regularity is bounded by linear functions [@Niu:AVanishing]. Since usually the Rees algebra $\oplus {\mathscr{I}}^t$ is not finitely generated it seems very hard that the asymptotic regularity is linear. However a recent work of Chardin [@Chardin:Powersofideals] showed a surprising result that $\operatorname{reg}{\mathscr{I}}^t$ could be linear for $t$ sufficiently large. In order to understand Chardin’s result and put those linear and nonlinear phenomena together in a clear picture, we give the following linearity theorem of asymptotic regularity. \[intr:01\] Let $d$ be the minimal number such that ${\mathscr{I}}(d)$ is generated by its global sections and $s$ be the $s$-invariant of ${\mathscr{I}}$. If $s=d$ then for $t$ large enough, one has $\operatorname{reg}{\mathscr{I}}^t=dt+e$ for some constant $e\geq 0$. The meaning of the constant $e$ can be explained in terms of the relative regularity and the biregularity of the blowing-up of ${\mathbb{P}}^n$ along ${\mathscr{I}}$ (for more details see Section 2). Thus this theorem strengthens Chardin’s result and justifies our intuition that positivity controls asymptotic regularity. As a quick corollary of the theorem, we give a typical variety having linear asymptotic regularity as follows. It is not so obvious and also tells us that varieties having linear asymptotic regularity are in fact not so rare. Suppose that ${\mathscr{I}}$ defines a (nonlinear) variety cut out by quadrics, then its asymptotic regularity is linear, i.e., for $t\gg 0$, one has $\operatorname{reg}{\mathscr{I}}^t=2t+e,$ for some positive integer $e\geq 0$. As another application of our linearity theorem, we can slightly strengthen a theorem due to Vermiere [@Vermeire:RegPowers] which gives the asymptotic regularity of a curve embedded by a line bundle of large degree. In this situation, the constant $e$ can be determined explicitly as either $0$ or $1$ depending on the surjectivity of the corresponding Gauss map (see Proposition \[pro:05\]). Let us further assume that ${\mathscr{I}}$ defines a projective subvariety $X$ in ${\mathbb{P}}^n$. The embedding of $X$ is determined by a very ample line bundle $L$ on $X$. The syzygies of $X$ therefore heavily depend on the positivity of $L$. An efficient way to describe such embedding is using Green’s condition $(N_p)$. We hope that knowing the syzygy of $X$ through $(N_p)$ condition could give information about the asymptotic regularity of $X$. The first case when $X$ is a nonsingular projective curve has been investigated by Vermiere [@Vermeire:RegPowers]. He showed that if $X$ is embedded by $L$ of degree $\geq 2g+3$, where $g$ is the genus of $X$, then ${\mathscr{I}}^t$ is $(2t+1)$-regular for $t\geq 1$. According to the theorem of Green and Lazarsfeld [@Lazarsfeld:SyzFiniteSet] such $L$ satisfies at least Property $(N_2)$. For higher dimensional nonsingular projective variety $X$, we consider the case that the embedding is determined by the following adjoint line bundle $$L_d=K_X+dA+B,$$ where $K_X$ is the canonical bundle of $X$, $A$ is a very ample line bundle and $B$ is a nef line bundle. A theorem of Ein and Lazarsfeld [@Ein:SyzygyKoszul] shows that if $d\geq \dim X+1+p$, then $L_d$ satisfies Property $(N_p)$. Thus an interesting question is up to which extend of the positivity of $L_d$ one could know the asymptotic regularity of $X$. The crucial point to get such asymptotic regularity bounds is to establish a vanishing theorem of the tensor algebra of the conormal bundle of $X$ (see Section 3). This idea is quite straightforward and has been used by Vermeire in the aforementioned work. As a general result we show the following theorem by assuming that $\dim X\geq 2$, that for the sufficient positive line bundle $L_d$ we can get regularity bounds for powers of an ideal sheaf. Assume that $d\geq 2(\dim X+1)$ in the adjoint line bundle $L_d$. Then for any $t\geq 1$, the ideal sheaf ${\mathscr{I}}^t$ is $(2t+2\dim X-2)$-regular. By the result of Ein and Lazarsfeld, such $L_d$ in the theorem satisfies Property $(N_{\dim X+1})$ and therefore $X$ is cut out by quadrics. Applying the linearity theorem \[intr:01\] above, we see that the constant $e$ satisfies $0\leq e\leq 2\dim X-2$. In particular if $X$ is a surface, we have $0\leq e\leq 2$. In the same way, we can also have an interesting result on regularity bounds for powers of ${\mathscr{I}}$ if $X$ is a locally complete intersection (equidimensional) and ${\mathscr{I}}$ is $r$-regular. Assume that $X\subset {\mathbb{P}}^n$ is a local complete intersection defined by ${\mathscr{I}}$ and assume that ${\mathscr{I}}$ is $r$-regular. 1. If $\dim X=1$, then for any $t\geq 1$, ${\mathscr{I}}^t$ is $rt$-regular. 2. If $\dim X=2$, then for any $t\geq 1$, ${\mathscr{I}}^t$ is $rt+r-2$-regular. 3. If $\dim X\geq 3$, then for any $t\geq 1$, ${\mathscr{I}}^t$ is $rt+\max(r,(\dim X -1)r-\dim X)$ regular. This paper is organized as follows. We first prove the linearity theorem of asymptotic regularity in Section 2. Then we build a vanishing theorem for tensor products of the conormal bundle of a variety in Section 3. In Section 4 and 5 we apply our vanishing theorem to get regularity bounds for powers of ideal sheaves. [*Acknowledgement.*]{} Special thanks are due to the author’s advisor Lawrence Ein who offers a lot of help and suggestions. The author also thanks Marc Chardin and Pete Vermeire for their explanations on their results and useful discussions and suggestions. Linearity of asymptotic regularity ================================== In this section, we study the linearity of the asymptotic regularity of an ideal sheaf on the projective space. Throughout, we fix our notation as follows. Suppose that ${\mathscr{I}}$ is an ideal sheaf on ${\mathbb{P}}^n$. Let $d$ be the minimal number such that ${\mathscr{I}}(d)$ is generated by its global sections. We denote by $H$ the hyperplane divisor of ${\mathbb{P}}^n$. We shall use line bundle and divisor interchangeable if there is confuse likely. Consider the blowing-up $$\mu: W=\operatorname{Bl}_{{\mathscr{I}}}{\mathbb{P}}^n\longrightarrow {\mathbb{P}}^n,$$ with an exceptional divisor $E$ such that ${\mathscr{I}}\cdot {\mathscr{O}}_W={\mathscr{O}}_W(-E)$. We denote by $V=H^0({\mathscr{I}}(d))$ the vector space of global sections of ${\mathscr{I}}(d)$. Those sections determine a surjective morphism $$V\otimes{\mathscr{O}}_{{\mathbb{P}}^n}(-d)\longrightarrow {\mathscr{I}}\longrightarrow 0.$$ Via the surjective morphism from the symmetric algebra of ${\mathscr{I}}$ to its Rees algebra, the morphism above gives the embedding of the blowing-up $W$ in the biprojective space $$W\hookrightarrow {\mathbb{P}}(V\otimes{\mathscr{O}}_{{\mathbb{P}}^n}(-d))\cong{\mathbb{P}}^n\times{\mathbb{P}}(V).$$ Note that ${\mathscr{O}}_W(d\mu^*H-E)$ is generated by its global sections. Let $p$ and $q$ be projections of ${\mathbb{P}}^n\times {\mathbb{P}}(V)$ to its components, we then have the diagram $$\label{eq:9} \xymatrix{ W=\operatorname{Bl}_{{\mathscr{I}}}{\mathbb{P}}^n \ar[dr]^{\mu} \ar@/^1.5pc/[rr]|\pi \ar@{^{(}->}[r] & {\mathbb{P}}^n\times {\mathbb{P}}(V) \ar[d]^p \ar[r]^q & {\mathbb{P}}(V) \\ & {\mathbb{P}}^n.}$$ On the biprojective space $Y={\mathbb{P}}^n\times {\mathbb{P}}(V)$, for any coherent sheaf ${\mathscr{F}}$, we denote by $${\mathscr{F}}(a,b)={\mathscr{F}}\otimes p^*{\mathscr{O}}_{{\mathbb{P}}^n}(a)\otimes q^*{\mathscr{O}}_{{\mathbb{P}}(V)}(b).$$ Under such notation, we see that $$\label{eq:10} {\mathscr{O}}_Y(0,1)|_W={\mathscr{O}}_W(0,1)=\pi^*{\mathscr{O}}_{{\mathbb{P}}(V)}(1)={\mathscr{O}}_W(d\mu^*H-E).$$ Now we recall some basic definitions. The first one is the $s$-invariant of ${\mathscr{I}}$, which measures its positivity. See [@Lazarsfeld:PosAG1 Section 5.4] for details. The [*$s$-invariant*]{} of ${\mathscr{I}}$ (with respect to the divisor $H$) is the positive real number $s({\mathscr{I}})=\min\{\ s\ |\ s\mu^*H-E \ \mbox{ is nef }\}$. Here $ s\mu^*H-E$ is considered as an ${\mathbb{R}}$-divisor on $W$. It is easy to see that $s\leq d$ since ${\mathscr{I}}(d)$ is generated by its global section and then the divisor ${\mathscr{O}}_W(d\mu^*H-E)$ is nef. There are two generalized notions of regularity: biregularity on a biprojective space and relative regularity on a projective bundle. \[def:1\] Let ${\mathscr{F}}$ be a coherent sheaf on a biprojective space $Y={\mathbb{P}}^a\times{\mathbb{P}}^b$. We say that ${\mathscr{F}}$ is $\textbf{m}=(m_1,m_2)$-regular if $$H^i(Y,{\mathscr{F}}(m_1-u,m_2-v))=0$$ for all $i>0$ and $u+v=i$ where $(u,v)\in {\mathbb{N}}^2$ (we assume $0\in {\mathbb{N}}$). Denote by $\operatorname{bireg}{\mathscr{F}}$ the set of the pair $\textbf{m}$ such that ${\mathscr{F}}$ is $\textbf{m}$-regular. We also denote by $b_1({\mathscr{F}})=\min\{m_1|\ \textbf{m}=(m_1,m_2)\in \operatorname{bireg}{\mathscr{F}}\}$ and $b_2({\mathscr{F}})=\min\{m_1|\ \textbf{m}=(m_1,m_2)\in \operatorname{bireg}{\mathscr{F}}\}$. Note that $b_1$ and $b_2$ could be $-\infty$. The basic property of biregualrity we will use is that if ${\mathscr{F}}$ is $\textbf{m}$-regular, then it is $\textbf{m}+{\mathbb{N}}^2$-regular. Let $X$ be a variety and ${\mathscr{E}}$ be a vector bundle on $X$, with the projectivization $\pi: {\mathbb{P}}({\mathscr{E}})\rightarrow X$. A coherent sheaf ${\mathscr{F}}$ on ${\mathbb{P}}({\mathscr{E}})$ is $m$-regular with respect to $\pi$ if $$R^i\pi_*({\mathscr{F}}\otimes{\mathscr{O}}_{{\mathbb{P}}({\mathscr{E}})}(m-i))=0$$ for $i>0$. Denote by $\operatorname{reg}_{\pi}{\mathscr{F}}$ the minimal number $m$ such that ${\mathscr{F}}$ is relative $m$-regular with respect to $\pi$. Note that $\operatorname{reg}_{\pi}{\mathscr{F}}$ could be $-\infty$. If ${\mathscr{F}}$ is relative $m$-regular, then it is relative $(m+1)$-regular. Let $X$ be a variety and ${\mathscr{E}}$ be a locally free sheaf with the projectivization $\pi:{\mathbb{P}}({\mathscr{E}})\rightarrow X$. Suppose that ${\mathscr{F}}$ is a coherent sheaf on ${\mathbb{P}}({\mathscr{E}})$ and let $c$ be a nonnegative integer. Then $R^i\pi_*{\mathscr{F}}(k)=0\mbox{ for all }i>c, k\in {\mathbb{Z}}$ if and only if for any $x\in X$, the restriction ${\mathscr{F}}_x$ of ${\mathscr{F}}$ on the fiber over $x$ satisfies $\dim \operatorname{Supp}{\mathscr{F}}_x\leq c$. The sufficient part is clear. We prove the necessary part, i.e., assume that $R^i\pi_*{\mathscr{F}}(k)=0\mbox{ for all }i>0, k\in {\mathbb{Z}}$. The plan is to use the formal function theorem and prove by contradiction. Suppose that there is a point $x\in X$ such that $P_x=\pi^{-1}(x)$ is the fiber and $d=\dim \operatorname{Supp}{\mathscr{F}}_x>c$. We claim that for $k\gg 0$, $$\label{eq:11} H^d(P_x,{\mathscr{F}}_x(-k))\neq 0.$$ For this, let $V=\operatorname{Supp}{\mathscr{F}}_x\subset P_x$ and we give a reduced scheme structure to $V$. Then it is enough to show that for $k\gg 0$, $$H^d(V,{\mathscr{F}}_x|_V(-k))\neq 0.$$ Because from the exact sequence $0\rightarrow \ker\rightarrow {\mathscr{F}}_x\rightarrow {\mathscr{F}}_x|_V\rightarrow 0$, we see that the sections of $H^d(V,{\mathscr{F}}_x|_V(-k))$ will be lifted to $H^d(P_x,{\mathscr{F}}_x(-k))$ since $\dim \operatorname{Supp}\ker\leq d$. Let $\omega_V$ be the dualizing sheaf of $V$. By duality, one has $$H^d(V,{\mathscr{F}}_x|_V(-k))=\operatorname{Hom}({\mathscr{F}}_x|_V,\omega_V(k))=H^0(V,\operatorname{\mathscr{H}om}({\mathscr{F}}_x|_V,\omega_V)(k)).$$ Thus it is enough to show that the sheaf $\operatorname{\mathscr{H}om}({\mathscr{F}}_x|_V,\omega_V)\neq 0$. This is then a local question. We may shrink $V$ if necessary to assume that $V$ is nonsingular and $\omega_V={\mathscr{O}}_V$. Thus locally we have a surjective morphism $\oplus {\mathscr{O}}_V\rightarrow ({\mathscr{F}}_x|_V)^{\vee}\rightarrow 0$ which gives an injection $({\mathscr{F}}_x|_V)^{\vee}\hookrightarrow \oplus{\mathscr{O}}_V$. Also note that ${\mathscr{F}}_x|_V$ is not torsion sheaf on $V$. By the generic flatness, the natural morphism ${\mathscr{F}}_x|_V\rightarrow ({\mathscr{F}}_x|_V)^{\vee}$ is nonzero. Composing it with the injection above, we have a nonzero morphism ${\mathscr{F}}_x|_V\rightarrow \oplus{\mathscr{O}}_V$. Thus by choosing to project to ${\mathscr{O}}_V$, we obtain a nonzero morphism ${\mathscr{F}}_x|_V\rightarrow {\mathscr{O}}_V$. Thus $\operatorname{\mathscr{H}om}({\mathscr{F}}_x|_V,{\mathscr{O}}_V)\neq 0$ and thus (\[eq:11\]) is true. Now let $nP_x$ is the scheme defined by ${\mathscr{I}}^n_{P_x}$, which is the $n$-th thickening of $P_x$, and let ${\mathscr{F}}_{nx}$ be the restriction of ${\mathscr{F}}$ to $nP_x$. Now from the sequence $$0\rightarrow \ker\rightarrow {\mathscr{F}}_{(n+1)x}\rightarrow {\mathscr{F}}_{nx}\rightarrow 0$$ and the fact that $\dim\operatorname{Supp}{\mathscr{F}}_x=d$, we see that any nonzero section of $H^d(P_x,{\mathscr{F}}_x(-k))$ in (\[eq:11\]) will be lifted to $H^d(P_x,{\mathscr{F}}_{(n+1)x}(-k))$. Thus by the formal function theorem, we have that $\hat{R}^d\pi_*{\mathscr{F}}(-k)\neq 0$ which is contradict to the assumption. Taking $c=0$ in the proposition, we get the following corollary. \[pro:10\] Keep notation as in Proposition above. $R^i\pi_*{\mathscr{F}}(k)=0\mbox{ for all }i>0, k\in {\mathbb{Z}}$ if and only if for any $x\in X$, the restriction ${\mathscr{F}}_x$ of ${\mathscr{F}}$ on the fiber over $x$ satisfies $\dim \operatorname{Supp}{\mathscr{F}}_x\leq 0$. Consider a coherent sheaf ${\mathscr{F}}$ on a biprojective space $Y={\mathbb{P}}^a\times {\mathbb{P}}^b$. We denote by $p_1$ and $p_2$ the projections of $Y$ to its components. Then ${\mathscr{F}}$ has biregularity and also has relative regularity with respect to $p_1$ and $p_2$. In the following theorem, we relate these two types of regularity together. \[pro:13\] Let ${\mathscr{F}}$ be a coherent sheaf on a biprojective space ${\mathbb{P}}^a\times {\mathbb{P}}^b$ with $p_1$ and $p_2$ the projection morphisms to its components. Let $r$ be an integer. Then ${\mathscr{F}}$ is relative $r$-regular with respect to $p_2$ if and only if ${\mathscr{F}}$ is $(r,r')$-biregular for some integer $r'$. We view $Y={\mathbb{P}}^a\times {\mathbb{P}}^b$ as a projectivized vector bundle over ${\mathbb{P}}^b$ with the tautological line bundle ${\mathscr{O}}_Y(1)=p^*_1{\mathscr{O}}_{{\mathbb{P}}^a}(1)$. Denote by $\delta=\dim Y$. Suppose that ${\mathscr{F}}$ is relative $r$-regular with respect to $p_2$, i.e., $R^i{p_2}_*{\mathscr{F}}(r-i)=0$ for all $i>0$. For any integers $k_1$ and $k_2$ we have a spectral sequence $$E^{p,q}_2=H^p({\mathbb{P}}^b,R^qp_{2*}{\mathscr{F}}(k_1)\otimes {\mathscr{O}}_{{\mathbb{P}}^b}(k_2)))\Rightarrow H^{p+q}(Y, {\mathscr{F}}(k_1)\otimes p^*_2{\mathscr{O}}_{{\mathbb{P}}^b}(k_2)).$$ By Serre’s vanishing theorem, there exists a number $n_0$ such that for all $k_2>n_0$ and for all $k_1=r,r-1,\cdots, r-\delta$, $$H^p({\mathbb{P}}^b,R^q_{p_2*}{\mathscr{F}}(k_1)\otimes {\mathscr{O}}_{{\mathbb{P}}^b}(k_2))=0,\quad\mbox{for all } p>0,q\geq 0.$$ We then deduce from the spectral sequence that for $k_2>n_0$ and $k_1=r,r-1,\cdots, r-\delta$, $$\label{eq:12} H^0({\mathbb{P}}^b,R^i_{p_2*}{\mathscr{F}}(k_1)\otimes {\mathscr{O}}_{{\mathbb{P}}^b}(k_2))=H^i(Y,{\mathscr{F}}(k_1)\otimes p^*_2{\mathscr{O}}_{{\mathbb{P}}^b}(k_2)),$$ Now for each $0<i\leq \delta$, since ${\mathscr{F}}$ is relative $r$-regular with respect to $p_2$, we see that $$R^ip_{2*}({\mathscr{F}}(r-u))=0,\quad \mbox{for }0\leq u\leq i.$$ Thus using (\[eq:12\]) with $k_1=r-u$, we see $$H^i(Y,{\mathscr{F}}(r-u)\otimes p^*_2{\mathscr{O}}_{{\mathbb{P}}^b}(k_2))=0,\quad\mbox{for all } k_2\geq n_0.$$ Then taking $r'>n_0+\delta$, we see that $$H^i(Y,{\mathscr{F}}(r-u,r'-v))=0,$$ where $(u,v)\in {\mathbb{N}}^2$ and $u+v=i$. This shows that ${\mathscr{F}}$ is $(r,r')$-biregular. Now suppose that ${\mathscr{F}}$ is $(r,r')$-biregular. Then ${\mathscr{F}}$ is $(r,k_2)$-biregular for all $k_2\geq r'$. Using spectral sequence above, for each $0\leq i\leq \delta$ and $k_2\gg 0$, we have $$H^0({\mathbb{P}}^b,R^ip_{2*}{\mathscr{F}}(r-i)\otimes {\mathscr{O}}_{{\mathbb{P}}^b}(k_2))=H^i(Y,{\mathscr{F}}(r-i)\otimes p^*_2{\mathscr{O}}_{{\mathbb{P}}^b}(k_2)).$$ Since by the biregularity assumption that $H^i(Y,{\mathscr{F}}(r-i)\otimes p^*_2{\mathscr{O}}_{{\mathbb{P}}^b}(k_2))=H^i(Y,{\mathscr{F}}(r-i,k_2))=0$, we immediately have $R^ip_{2*}{\mathscr{F}}(r-i)=0$. Thus ${\mathscr{F}}$ is relative $r$-regular with respect to $p_2$. \[pro:14\] Keep notation as in Proposition \[pro:13\]. Denote by $b_1({\mathscr{F}})=\min\{m_1|\ \textbf{m}=(m_1,m_2)\in \operatorname{bireg}{\mathscr{F}}\}$. Then one has $$b_1({\mathscr{F}})=\operatorname{reg}_{p_2}{\mathscr{F}}.$$ Note that in the corollary, we do allow $b_1({\mathscr{F}})$ and $\operatorname{reg}_{p_2}{\mathscr{F}}$ to be $-\infty$. In fact, using Corollary \[pro:10\], we can easily summarize in the following corollary the case when these numbers are $-\infty$. \[pro:12\] Keep notation as above. The following are equivalent 1. $b_1({\mathscr{F}})=-\infty$; 2. $\operatorname{reg}_{p_2}{\mathscr{F}}=-\infty$; 3. $\dim\operatorname{Supp}{\mathscr{F}}_x\leq 0$ for any $x\in {\mathbb{P}}^b$. Going back to the picture at the beginning. Think of ${\mathscr{O}}_W$ as a coherent sheaf on the biprojective space $Y={\mathbb{P}}^n\times {\mathbb{P}}(V)$, then it has biregularity and therefore the number $b_1({\mathscr{O}}_W)$ as in the definition (\[def:1\]) is defined. The sheaf ${\mathscr{O}}_W$ also has relative regularity $\operatorname{reg}_{q}{\mathscr{O}}_W$ with respect to $q$. The relation between these notions of regularity and the positivity of ${\mathscr{I}}$ is described in the following proposition. \[pro:15\] The following are equivalent 1. $b_1({\mathscr{O}}_W)=-\infty$; 2. $\operatorname{reg}_{q}{\mathscr{O}}_W=-\infty$; 3. $s<d$; 4. $\pi:W\rightarrow {\mathbb{P}}(V)$ is finite. \(1) $\Leftrightarrow$ (2) $\Leftrightarrow$ (4) is from Corollary \[pro:12\]. (3) $\Leftrightarrow$ (4) is from the definition of $s$-invariant. Now we come to our main theorem of this section. We show that if $s=d$ then the asymptotic regularity of ${\mathscr{I}}$ will be linear. We also give the way to compute the constant part of the asymptotic regularity; it is the relative regularity of the blowing-up ${\mathscr{O}}_W$. \[thm:02\] Suppose that $s=d$ and denote by $e=b_1({\mathscr{O}}_W)=\operatorname{reg}_{q}{\mathscr{O}}_W$ (the equality is guaranteed by Corollary \[pro:14\]). For $t$ large enough, one has $$\operatorname{reg}{\mathscr{I}}^t=dt+e,$$ and $e\geq 0$. Thus we see that for the ideal sheaf ${\mathscr{I}}$ with maximal positivity, it has linear asymptotic regularity. We first show that if $s=d$ then for any $t\geq 0$, $\operatorname{reg}{\mathscr{I}}^t\geq dt$. This is because otherwise suppose there is a number $t$ such that $\operatorname{reg}{\mathscr{I}}^t=dt-a$ for some $a>0$. Then ${\mathscr{I}}^t(dt-a)$ is generated by its global sections. Thus the line bundle ${\mathscr{O}}_W((dt-a)\mu^* H-tE)$ is nef. That means $s\leq (dt-a)/t<d$ which is contradict to $s=d$. Now let $\operatorname{reg}{\mathscr{I}}^t=dt+e_t$ for $t\geq 0$. Note that from the above argument $e_t\geq 0$ and therefore $\liminf \{e_t\}\geq 0$. Also notice that $e\neq-\infty$ since $s=d$ and by Proposition \[pro:15\]. Our plan is to prove the following inequalities $$e\leq \liminf \{e_t\}\leq \limsup \{e_t\}\leq e.$$ First we show $\limsup \{e_t\}\leq e$. Since $e=b_1({\mathscr{O}}_W)$ there is a number $l$ such that $\textbf{m}=(e,l)$ and ${\mathscr{O}}_W$ is $\textbf{m}$-regular. Then ${\mathscr{O}}_W$ is $\textbf{m}+{\mathbb{N}}^2$-regular. This implies that $$H^i(Y,{\mathscr{O}}_W(e-i,t))=0\quad\mbox{ for } t\gg 0.$$ Recall that for $t$ large enough, one has $H^i(Y,{\mathscr{O}}_W(e-i,t))=H^i({\mathbb{P}}^n,{\mathscr{I}}^t(dt+e-i))$. Thus we have the vanishing of $H^i({\mathbb{P}}^n,{\mathscr{I}}^t(dt+e-i))=0$ for $t$ sufficiently large and therefore $\operatorname{reg}{\mathscr{I}}^t\leq dt+e$. Thus $\limsup \{e_t\}\leq e$. Next we show that $e\leq \liminf \{e_t\}$. Let $r=\liminf \{e_t\}$. Then there is a sequence $\{e_{t_j}\}^{\infty}_{j=0}$ such that $e_{t_j}=r$ for $t_j\rightarrow +\infty$. Thus by $H^i(Y,{\mathscr{O}}_W(e_{t_j}-i,t_j))=H^i({\mathbb{P}}^n,{\mathscr{I}}^{t_j}(dt_j+e_{t_j}-i))$ again, one has $$H^i(Y,{\mathscr{O}}_W(r-i,t_j))=0\quad\mbox{ for } j\gg 0,$$ which implies that $R^iq_*{\mathscr{O}}_W(r-i)=0$. Thus ${\mathscr{O}}_W$ is relative $r$-regular with respect to $q$. Since $e=\operatorname{reg}_q{\mathscr{O}}_W$, we see that $e\leq r$ and therefore $e\leq \liminf \{e_t\}$. Hence we get $e=\lim e_t$ and therefore $\operatorname{reg}{\mathscr{I}}^t=dt +e$ for $t$ large enough. Immediately, we have the following corollary describing a typical type of variety having linear asymptotic regularity, which is not so obvious. Suppose that ${\mathscr{I}}$ defines a variety cut out by quadrics, then its asymptotic regularity is linear, i.e., for $t\gg 0$, one has $$\operatorname{reg}{\mathscr{I}}^t=2t+e,$$ for some positive integer $e$. Let $X$ be the variety defined by ${\mathscr{I}}$. Since $X$ is cut out by quadrics, we see $s\leq 2$, where $s$ is the $s$-invariant of ${\mathscr{I}}$. Also a generic secant line will cut $X$ by two distinct points. This shows that $s\geq 2$. Thus $s=2$ and then the result follows from Theorem \[thm:02\]. \(1) Note that the constant $e$ can be computed as the relative regularity of ${\mathscr{O}}_W$. But it still hard to obtain in practise. \(2) From the corollary above, we see that these varieties having linear asymptotic regularity in fact are not so rare as we expected, although their Rees algebra $\oplus {\mathscr{I}}^t$ are not finitely generated in general. \(3) In Chardin’s work [@Chardin:Powersofideals], he shows that if $I$ is a homogeneous ideal generated in a single degree $d$, then the limit $$\lim_{t\rightarrow \infty} (\operatorname{reg}(I^t)^{sat}-dt)$$ exist but could be $-\infty$. Thus by using the $s$-invariant $s$, we give a clear picture for his result. In the light of Theorem \[thm:02\], it would be convenient to give a name for the constant $e$. We suggest the following definition. Suppose that an ideal sheaf ${\mathscr{I}}$ satisfies $s({\mathscr{I}})=d({\mathscr{I}})$. The [*asymptotic regularity constant*]{} $e({\mathscr{I}})$ is defined as the constant part of $\operatorname{reg}{\mathscr{I}}^t=dt+e$ for $t$ large enough. We make a convention that if $s({\mathscr{I}})<d({\mathscr{I}})$, then $e({\mathscr{I}})=-\infty$. According to Theorem \[thm:02\], the asymptotic regularity constant $e$ is always nonnegative if it is not $-\infty$. We hope that this number would carry some geometric or algebraic information of ${\mathscr{I}}$. For example, we will see in Remark \[rmk:01\], it is determined by the surjectivity of a Guass map. We conclude this section by giving an example to compute a bound for $e$. Still we assume that $s({\mathscr{I}})=d({\mathscr{I}})$. If we know the degrees of the generators of $W$ in $Y={\mathbb{P}}^n\times {\mathbb{P}}(V)$, then we could give a bound for the asymptotic regularity constant $e$. Not surprisingly, such bound would be every large in general. Specifically, assume that ${\mathscr{I}}_W$ is cut out by equations of bidegree $(d_i,d'_i)$ for $i=1,\cdots, l$. Let $D=\max_i\{d_i\}$. Then for any point $y\in {\mathbb{P}}(V)$, in the fiber $Y_y$, $W_y$ is cut out by equations of degree no more than $D$. Assume that $(A,m)=({\mathscr{O}}_{{\mathbb{P}}(V),y},m_y)$ the local ring of $y$. Then for integer $k\geq 1$, in the thickening fiber $Y_{ky}=\operatorname{Proj}A/m^k[x_0,\cdots, x_n]$, the thickening $W_{ky}$ is still cut out by equations of degrees no more than $D$. According to the result of Chardin, Fall and Nagel [@Chardin:BoundRegModules Example 3.6], the regularity of $W_{ny}$ is bounded by $(n+1)(D-1)+1$ if $n\leq 2$, and by $(3D^3)^{2^{n-3}}$ if $n\geq 3$. Then an easy application of formal function theorem will give us that $$e \leq \left\{ \begin{array}{ll} (n+1)(D-1) & \textrm{if $n\leq 2$,}\\ (3D^3)^{2^{n-3}}-1 & \textrm{if $n\geq 3$.} \end{array} \right.$$ However, it is very difficult in general to find the defining equations of $W$ in $Y$. Tensor algebra of conormal bundles ================================== In the rest of the paper, we study the regularity bounds for powers of an ideal sheaf under its geometric conditions. We always assume in the sequel that $X$ is a locally complete intersection of equidimension $n$ in the projective space ${\mathbb{P}}$ (we omit the dimension of the projective space ${\mathbb{P}}$, which does not come into our results), defined by an ideal sheaf ${\mathscr{I}}$. Denote by $N^*$ the conormal sheaf of $X$ which is ${\mathscr{I}}/{\mathscr{I}}^2$ by definition. Since $X$ is a locally complete intersection the conormal sheaf $N^*$ is then a locally free sheaf on $X$, i.e., the conormal bundle of $X$. We fix a minimal free resolution of ${\mathscr{I}}$ as follows $$\cdots\rightarrow\oplus {\mathscr{O}}_{{\mathbb{P}}}(-d_{3,j})\rightarrow\oplus {\mathscr{O}}_{{\mathbb{P}}}(-d_{2,j})\rightarrow\oplus {\mathscr{O}}_{{\mathbb{P}}}(-d_{1,j})\rightarrow\oplus {\mathscr{O}}_{{\mathbb{P}}}(-d_{0,j})\rightarrow{\mathscr{I}}\rightarrow 0.$$ Tensoring the minimal resolution with the structure sheaf ${\mathscr{O}}_X$, we obtain a complex $$\label{eq:01} \cdots\rightarrow\oplus {\mathscr{O}}_X(-d_{3,j})\rightarrow\oplus {\mathscr{O}}_X(-d_{2,j})\rightarrow\oplus {\mathscr{O}}_X(-d_{1,j})\rightarrow\oplus {\mathscr{O}}_X(-d_{0,j})\rightarrow N^*\rightarrow 0,$$ which has homology sheaves $$\cdots,\quad {\mathscr{H}}_3=\wedge^4N^*,\quad {\mathscr{H}}_2=\wedge^3N^*,\quad {\mathscr{H}}_1=\wedge^2N^*,\quad {\mathscr{H}}_0=0,$$ i.e., $${\mathscr{H}}_i=\wedge^{i+1}N^*\quad\mbox{ for }i\geq 1.$$ The complex (\[eq:01\]) can be viewed as a non-exact resolution of the conormal bundle $N^*$. Our strategy is to deduce a vanishing theorem of the tensor algebra of $N^*$ from this complex. For any number $p\geq 0$, we say that ${\mathscr{I}}$ is $p$-th partial $m$-regular if $$d_{i,j}-i\leq m,\quad\mbox{ for }\ 0\leq i\leq p.$$ Obviously if ${\mathscr{I}}$ is $(p+1)$-th partial $m$-regular then it is $p$-th partial $m$-regular. We also assume that the structure sheaf ${\mathscr{O}}_X$ is $r_X$-regular. In order to avoid some trivial situation, we always assume that the regularity and the partial regularities of ${\mathscr{I}}$ are at least $2$ (in fact this just means that $X$ is not a linear space). The following lemma is easy to prove and can be found in [@Ein:SyzygyKoszul Lemma 2.10]. \[pro:31\] Let $$Q_\bullet: \cdots \rightarrow Q_2\rightarrow Q_1\stackrel{\epsilon}{\rightarrow} Q_0\rightarrow 0$$ be a complex of coherent sheaves on $X$, with $\epsilon$ surjective. Assume that 1. $H^k(X,Q_1)=H^{k+1}(X,Q_2)=\cdots=H^n(X,Q_{n-k+1})=0$; 2. $H^{k+1}(X,{\mathscr{H}}_1(Q_\bullet))=H^{k+2}(X,{\mathscr{H}}_2(Q_\bullet))=\cdots=H^n(X,{\mathscr{H}}_{n-k}(Q_\bullet))=0$. Then $H^k(X,Q_0)=0$. The theorem we established in this section is the following vanishing theorem of the tensor products of $N^*$. \[thm:01\] Assume that ${\mathscr{I}}$ is $p$-th partial $r_p$-regular for $p\leq n-1$ and ${\mathscr{O}}_X$ is $r_X$-regular. Then one has $$\begin{aligned} &H^n(X,T^aN^*(k))=0 &&\mbox{ for }k\geq ar_p+r_X-n,\\ &H^{n-1}(X,T^aN^*(k))=0 &&\mbox{ for }k\geq ar_p+1+r_X-n,\\ &H^{n-2}(X,T^aN^*(k))=0 &&\mbox{ for }k\geq (a+1)r_p+r_X-n,\\ &H^{n-3}(X,T^aN^*(k))=0 &&\mbox{ for }k\geq (a+2)r_p+r_X-n,\\ &\cdots && \cdots,\\ &H^{n-p}(X,T^aN^*(k))=0 &&\mbox{ for }k\geq (a+p-1)r_p+r_X-n.\end{aligned}$$ We first establish the vanishing of $H^n$ groups. From the complex (\[eq:01\]) and by Lemma \[pro:31\], we see that $$H^n(X,N^*(k))=0,\quad \mbox{for } k\geq r_p+r_X-n.$$ Tensoring $N^*$ to the the complex (\[eq:01\]), we obtain a complex $$\cdots\rightarrow \oplus N^*(-d_{0,j})\rightarrow T^2N^*\rightarrow 0.$$ Thus immediately by Lemma \[pro:31\], we have $$H^n(X,T^2N^*(k))=0,\quad \mbox{for } k\geq 2r_p+r_X-n.$$ Repeating such procedure we then get $H^n(X,T^aN^*(k))=0$ for $k\geq ar_p+r_X-n$. Now we establish the vanishing of $H^{n-1}$ groups. In the complex (\[eq:01\]) notice that $$H^{n-1}(X,{\mathscr{O}}_X(k-d_{0,j}))=H^n(X,{\mathscr{O}}_X(k-d_{1,j}))=0,\quad\mbox{for } k\geq r_p+1+r_X-n.$$ Then by Lemma \[pro:31\], we obtain $H^{n-1}(X,N^*(k))=0$, for $k\geq r_p+1+r_X-n$. Tensoring $N^*$ to the complex (\[eq:01\]), we have a complex $$\cdots\rightarrow \oplus N^*(-d_{1,j})\rightarrow \oplus N^*(-d_{0,j})\rightarrow T^2N^*\rightarrow 0,$$ which has homology sheaf ${\mathscr{H}}_0=0$. Notice that we have established $$H^{n-1}(X,N^*(k-d_{0,j}))=H^n(X,N^*(k-d_{1,j})),\quad \mbox{for }k\geq 2r_p+1+r_X-n.$$ Then by Lemma \[pro:31\], we obtain the vanishing of $H^{n-1}(X,T^2N^*(k))$. Repeatedly we then obtain $H^{n-1}(X,T^aN^*(k))=0$ for $k\geq ar_p+1+r_X-n$. Recall that for any nonnegative number $a$ and $b$, $\wedge^aN^*\otimes T^bN^*$ is a direct summand of $T^{a+b}N^*$ since $N^*$ is locally free and we work over ${\mathbb{C}}$. Thus the vanishing of the cohomology groups of $T^{a+b}N^*$ will automatically give the vanishing of the cohomology groups of $\wedge^aN^*\otimes T^bN^*$. We will use this fact repeatedly. In order to establish the vanishing of $H^{n-2}(X,T^aN^*(k))$ for $a\geq 1$, we tensor $T^{a-1}N^*$ to the complex (\[eq:01\]) to get a complex $$\cdots\rightarrow \oplus T^{a-1}N^*(-d_{2,j})\rightarrow \oplus T^{a-1}N^*(-d_{1,j})\rightarrow \oplus T^{a-1}N^*(-d_{0,j})\rightarrow T^aN^*\rightarrow 0,$$ which has homology sheaves ${\mathscr{H}}_0=0$ and $${\mathscr{H}}_i=\wedge^{i+1}N^*\otimes T^{a-1}N^*,\quad \mbox{for }i\geq 1.$$ Since inductively, we have established the vanishing of $$H^{n-2}(X,T^{a-1}N(k-d_{0,j})), H^{n-1}(X,T^{a-1}N^*(k-d_{1,j})), H^n(X,T^{a-1}N^*(k-d_{2,j}))$$ and the vanishing of $H^n(X,\wedge^2N^*\otimes T^{a-1}N^*(k))$, which from the vanishing of $H^n(X,T^{a+1}N^*(k))$. Thus by Lemma \[pro:31\] again, we have $H^{n-2}(X,T^aN^*(k))=0$ for $k\geq (a+1)r_p+r_X-n$. Inductively, we can finally prove the theorem. In order to give regularity bounds for powers of ${\mathscr{I}}$, we inductively consider the exact sequence for any $a\geq 1$, $$0\longrightarrow {\mathscr{I}}^{a+1}\longrightarrow {\mathscr{I}}^a\longrightarrow S^aN^*\longrightarrow 0.$$ For any $k\in {\mathbb{Z}}$, we denote the morphism $$\label{eq:02} \phi_a:H^0({\mathscr{I}}^a(k))\longrightarrow H^0(S^aN^*(k)).$$ The morphism $\phi_a$ actually depends on the twist $k$ but we omit it by abuse of notation. The crucial point is to prove the surjectivity of $\phi_a$ for specific twisting $k$. For this we will put $\phi_a$ in a commutative diagram and to analyze each morphisms in it. We tensor ${\mathscr{O}}_X$ to the morphism $\oplus {\mathscr{O}}_{{\mathbb{P}}}(-d_{0,j})\rightarrow {\mathscr{I}}\rightarrow 0$ of the first piece in the complex (\[eq:01\]) to get the following diagram $$\label{eq:03} \begin{CD} \oplus {\mathscr{O}}_{{\mathbb{P}}}(-d_{0,j}) @>>> {\mathscr{I}}@>>> 0\\ @VVV @VVV\\ \oplus{\mathscr{O}}_X(-d_{0,j})@>>> N^*@>>> 0 \end{CD}$$ Now for any $k,s\in {\mathbb{Z}}$, we have the following morphisms on global sections $$\begin{CD} \oplus H^0({\mathscr{O}}_{{\mathbb{P}}}(k-d_{0,j}) @>>> H^0({\mathscr{I}}(k))\\ @VVV @VVV\\ \oplus H^0({\mathscr{O}}_X(k-d_{0,j}))@>>> H^0(N^*(k)) \end{CD},\mbox{and} \quad\quad\quad \begin{CD} H^0({\mathscr{I}}^a(s))\\ @VV\phi_aV \\ H^0(S^aN^*(s)) \end{CD}$$ We tensor the left-hand-side diagram to the right-hand-side one above to deduce the following diagram and mark morphisms in it, $$\xymatrix{ \oplus H^0({\mathscr{O}}_{{\mathbb{P}}}(k-d_{0,j})\otimes H^0({\mathscr{I}}^a(s)) \ar[d]^{1\otimes\phi_a} \ar[r]& H^0({\mathscr{I}}(k))\otimes H^0({\mathscr{I}}^a(s)) \ar[dd]\ar[r] & H^0({\mathscr{I}}^{a+1}(k+s)) \ar[dd]^{\phi_{a+1}} \\ \oplus H^0({\mathscr{O}}_{{\mathbb{P}}}(k-d_{0,j}))\otimes H^0(S^aN^*(s)) \ar[d]^{c_a}&\\ \oplus H^0(S^aN^*(k-d_{0,j}+s))\ar[r]^{w_a} & H^0(N^*\otimes S^aN^*(k+s)) \ar[r]^{u_{a+1}} & H^0(S^{a+1}N^*(k+s))}$$ We simply write it as follows $$\xymatrix{ \oplus H^0({\mathscr{O}}_{{\mathbb{P}}}(k-d_{0,j})\otimes H^0({\mathscr{I}}^a(s)) \ar[d]^{1\otimes\phi_a} \ar[rr]&& H^0({\mathscr{I}}^{a+1}(k+s)) \ar[dd]^{\phi_{a+1}} \\ \oplus H^0({\mathscr{O}}_{{\mathbb{P}}}(k-d_{0,j}))\otimes H^0(S^aN^*(s)) \ar[d]^{c_a}&\\ \oplus H^0(S^aN^*(k-d_{0,j}+s))\ar[r]^{w_a} & H^0(N^*\otimes S^aN^*(k+s)) \ar[r]^{u_{a+1}} & H^0(S^{a+1}N^*(k+s))}$$ Thus inductively, in order to build the surjectivity of $\phi_{a+1}$ it is enough to build the surjectivity of $\phi_a$, $c_a$, $w_a$ and $u_{a+1}$. Since the formulas in Theorem \[thm:01\] depend on the dimension of $X$ we consider lower and higher dimension cases for $X$ in the following two sections. Higher dimensional varieties ============================ Keep notation as in Section 3 and assume that $n=\dim X\geq 3$. Assume also that ${\mathscr{I}}$ is $n$-th partial $r_n$-regular (recall that we always require that $r_n\geq 2$), the structure sheaf ${\mathscr{O}}_X$ is $r_X$-regular and $X$ is $n_0$-normal, i.e., $H^1({\mathbb{P}},{\mathscr{I}}(k))=0$ for $k\geq n_0$. We start by giving the regularity bounds for $T^aN^*$ which follows easily from Theorem \[thm:01\]. \[pro:01\]For $a\geq 1$ one has $$T^aN^* \mbox{ is } (a+n-2)r_n+r_X-n+1\mbox{ regular}.$$ It is an immediate consequence of Theorem \[thm:01\]. \[pro:03\] For any $a\geq 1$ the morphism $$u_a:H^0(N^*\otimes S^{a-1}N^*(k))\longrightarrow H^0(S^aN^*(k))$$ is surjective if $k\geq (a+n-2)r_n+r_X-n$. In the exact sequence $$\cdots\longrightarrow S^{a-2}N^*\otimes \wedge^2N^*\longrightarrow S^{a-1}N^*\otimes N^*\stackrel{\partial}{\longrightarrow} S^aN^*\longrightarrow 0.$$ each term is $(a+n-2)r_n+r_X-n+1$ regular by Proposition \[pro:01\]. Let ${\mathscr{K}}=\ker \partial$, then by chasing from this complex, it is easy to see that $H^1({\mathbb{P}},{\mathscr{K}}(a+n-2)r_n+r_X-n)=0$. Thus the surjectivity of $u_a$ follows immediately. \[pro:02\] For $a\geq 0$ the morphism $$w_a:\oplus H^0(S^aN^*(k-d_{0,j}))\longrightarrow H^0(N^*\otimes S^aN^*(k))$$ is surjective for $k\geq (n+a)r_n+r_X-n$. Tensor $S^aN^*$ to the complex (\[eq:01\]) to get the complex $$\cdots\rightarrow\oplus S^aN^*(-d_{2,j})\rightarrow\oplus S^aN^*(-d_{1,j})\rightarrow\oplus S^aN^*(-d_{0,j})\stackrel{\delta}{\rightarrow} N^*\otimes S^aN^*\rightarrow 0,$$ The homology sheaves of this complex are $${\mathscr{H}}_i=\wedge^{i+1}N^*\otimes S^aN^*\quad\quad \mbox{ for } i\geq 1$$ and ${\mathscr{H}}_0=0$. Let ${\mathscr{K}}$ be the kernel sheaf of the morphism $\delta$, then we have a complex $$\cdots\rightarrow\oplus S^aN^*(-d_{2,j})\rightarrow\oplus S^aN^*(-d_{1,j})\rightarrow {\mathscr{K}}\rightarrow 0$$ which is exact at ${\mathscr{K}}$. Thus it is enough to show that for $k\geq (n+a)r_n+r_X-n$ we have $H^1({\mathscr{K}}(k))=0$. Since from Proposition \[pro:01\], we see that for $k\geq (n+a)r_n+r_X-n$, $$H^1(S^aN^*(k-d_{1,j}))=H^2(S^aN^*(k-d_{2,j}))=\cdots=H^n(S^aN^*(k-d_{n,j}))=0,$$ and $$H^2(\wedge^2N^*\otimes S^aN^*(k))=H^3(\wedge^3N^*\otimes S^aN^*(k))=\cdots=H^n(\wedge^nN^*\otimes S^aN^*(k))=0.$$ Thus by Lemma \[pro:31\], our result follows. \[pro:04\] For $a\geq 1$, the morphism $$\phi_a:H^0({\mathscr{I}}^a(k))\longrightarrow H^0(S^aN^*(k))$$ is surjective for $k\geq ar_n+\max(n_0,(n-1)r_n+r_X-n)$. We prove by induction on $a$. We start by showing the surjectivity of the morphism $\phi_1: H^0({\mathscr{I}}(k))\rightarrow H^0(N^*(k))$. Twisting the diagram (\[eq:03\]) by ${\mathscr{O}}_{{\mathbb{P}}}(k)$ and then taking global sections to obtain the diagram $$\begin{CD} \oplus H^0({\mathscr{O}}_{{\mathbb{P}}}(k-d_{0,j})) @>>> H^0({\mathscr{I}}(k))\\ @Vc_0VV @VVV\\ \oplus H^0({\mathscr{O}}_X(k-d_{0,j}))@>w_0>> H^0(N^*(k)) \end{CD}$$ Then observe that 1. The surjectivity of $c_0$ is guaranteed by the normality of $X$ when $k\geq r_n+n_0$ 2. The surjectivity of $w_0$ has been built in Lemma \[pro:02\] when $k\geq nr_n+r_X-n$. Write $s_1=r_n+\max(n_0,(n-1)r_n+r_X-n)$ then $\phi_1$ is surjective if $k\geq s_1$. Note that $N^*$ is also $s_1$ regular by Proposition \[pro:01\]. Inductively suppose that $\phi_a$ is surjective. We write $s_a=ar_n+\max(n_0,(n-1)r_n+r_X-n)$ and note that $S^aN^*$ is $s_a$ regular by Proposition \[pro:01\]. we show the surjectivity of $\phi_{a+1}$. For this consider the diagram constructed in Section 3 $$\xymatrix{ \oplus H^0({\mathscr{O}}_{{\mathbb{P}}}(k'-d_{0,j})\otimes H^0({\mathscr{I}}^a(s_a)) \ar[d]^{1\otimes\phi_a} \ar[rr]&& H^0({\mathscr{I}}^{a+1}(k'+s_a)) \ar[dd]^{\phi_{a+1}} \\ \oplus H^0({\mathscr{O}}_{{\mathbb{P}}}(k'-d_{0,j}))\otimes H^0(S^aN^*(s_a)) \ar[d]^{c_a}&\\ \oplus H^0(S^aN^*(k'-d_{0,j}+s_a))\ar[r]^{w_a} & H^0(N^*\otimes S^aN^*(k'+s_a)) \ar[r]^{u_{a+1}} & H^0(S^{a+1}N^*(k'+s_a))}$$ Observe that 1. $1\otimes\phi_a$ is surjective since so is $\phi_a$. 2. $c_a$ is surjective if $k'\geq r_n$ since $S^aN^*$ is $s_a$-regular. 3. $w_2$ is surjective if $k'+s_a\geq (n+2)r_n+r_X-n$ by Lemma \[pro:02\]. 4. $u_{a+1}$ is surjective if $k'+s_a\geq (a+n-1)r_n+r_X-n$ by Lemma \[pro:03\]. Thus if write $s_{a+1}=(a+1)r_n+\max(n_0,(n-1)r_n+r_X-n)$ then $\phi_{a+1}$ is surjective if $k\geq s_{a+1}$. As an application of above results, we first consider the case that $X$ is a nonsingular projective variety embedded by an adjoint line bundle $$L_d=K_X+dA+B,$$ where $A$ is a very ample line bundle, $B$ is nef line bundle and $K_X$ is the canonical line bundle. According to the work of Ein and Lazarsfeld [@Ein:SyzygyKoszul], if $d\geq \dim X+1+p$ then $X$ satisfies Property $N_p$. For $L_d$ positive enough, we have a chance to get regularity bounds for powers of ${\mathscr{I}}$. As setting above. Assume that $d\geq 2(\dim X+1)$. Then for any $a\geq 1$, one has $${\mathscr{I}}^a \mbox{ is } 2a+2n-2 \mbox{ regular}.$$ Inductively using Proposition \[pro:04\], Proposition \[pro:01\] and exact sequences $$0\longrightarrow {\mathscr{I}}^{a+1}\longrightarrow {\mathscr{I}}^a\longrightarrow S^aN^*\longrightarrow 0,$$ the result follows immediately. As another application, we next consider the case that $X$ is a locally complete intersection and ${\mathscr{I}}$ is $r$-regular. Then we are able to give the regularity bounds for ${\mathscr{I}}^a$ in terms of $r$. As setting above, then for any $a\geq 1$ one has $${\mathscr{I}}^a \mbox{ is } ar+\max(r,(n-1)r-n) \mbox{ regular}.$$ Curves and Surfaces =================== In this section, we follow the same approach as previous section to study the case of curves and surfaces. Since the argument is exactly same as previous we shall be brief. Let us first start with the case of curves. Keep notation as in the beginning of Section 3 and assume that $\dim X=1$. 1. Assume that ${\mathscr{I}}$ is $0$-th partial $r_0$-regular and ${\mathscr{O}}_X$ is $r_X$-regular. Then $$T^aN^* \mbox{ is } (ar_0+r_X)\mbox{-regular}.$$ 2. Assume further that ${\mathscr{I}}$ is $1$-st partial $r_1$-regular and $X$ is $n_0$-normal, then for $a\geq 1$, the morphism $$\phi_a:H^0({\mathscr{I}}^a(k))\longrightarrow H^0(S^aN^*(k))$$ is surjective for $k\geq ar_1+\max(n_0,r_X)$. \(1) is an immediate corollary of Theorem \[thm:01\]. (2) is the same argument as Proposition \[pro:04\]. In fact for a nonsingular projective curve embedded by a large degree line bundle $L$, Vermeire [@Vermeire:RegPowers] has used the same idea to prove the following theorem. Assume that $X$ is a nonsingular projective curve of genus $g$ embedded by a line bundle $L$ with $\deg L\geq 2g+3$. Then for any $a\geq 1$, $${\mathscr{I}}^a \mbox{ is } (2a+1)\mbox{-regular}.$$ Just apply the proposition above. The syzygies of the curve embedded by such line bundle $L$ has been studied by Green and Lazarsfeld. And according to their theorem such curve satisfies at leas Property $(N_2)$. Also in his work, Vermeire has proved that ${\mathscr{I}}^a$ is $2a$-regular if and only if the Guass map $\Phi_L$ is surjective. Plus the linearity theorem we established in Section 2, we are able to give the effective value for $\operatorname{reg}{\mathscr{I}}^a$, which slightly strengthens Vermerie’s result. \[pro:05\] Assume that $X$ is a nonsingular projective curve of genus $g$ embedded by a line bundle $L$ with $\deg L\geq 2g+3$. Then for any $a\geq 1$, 1. $\operatorname{reg}{\mathscr{I}}^a=2a+1$ if and only if the Guass map $\Phi_L$ is not surjective 2. $\operatorname{reg}{\mathscr{I}}^a=2a$ if and only if the Guass map $\Phi_L$ is surjective. \[rmk:01\] Note that the proposition means that if for some $a$ $\operatorname{reg}{\mathscr{I}}^a=2a$ then for any $a$, $\operatorname{reg}{\mathscr{I}}^a=2a$. Thus the asymptotic behavior of the regularity of curves in this case is quite clear: the asymptotic regularity constant $e$ is either $0$ or $1$, which depends on the surjectivity of the Guass map $\Phi_L$. Next, we give regularity bounds for any powers of ${\mathscr{I}}$ by assuming that we know the regularity of ${\mathscr{I}}$. Assume that $X$ is a local complete intersection of dimension one in ${\mathbb{P}}$ defined by an ideal sheaf ${\mathscr{I}}$ which is $r$-regular, then for any $a\geq 1$, one has $${\mathscr{I}}^a \mbox{ is } ar\mbox{-regular}.$$ This formula is sharp. For example assume that $X$ is a rational normal curve, then $\operatorname{reg}{\mathscr{I}}=2$ and therefore $\operatorname{reg}{\mathscr{I}}^a=2a$, which is sharp. Now we turn to the case of surfaces. Keep notation as in the beginning of Section 3 and assume that $\dim X=2$. 1. Assume that ${\mathscr{I}}$ is $1$-th partial $r_1$-regular and ${\mathscr{O}}_X$ is $r_X$-regular. Then $$T^aN^* \mbox{ is } (ar_1+r_X)\mbox{-regular}.$$ 2. Assume further that ${\mathscr{I}}$ is $2$-nd partial $r_2$-regular and $X$ is $n_0$-normal, then for $a\geq 1$, the morphism $$\phi_a:H^0({\mathscr{I}}^a(k))\longrightarrow H^0(S^aN^*(k))$$ is surjective for $k\geq ar_2+\max(n_0,r_2+r_X-2)$. As in the higher dimensional case, here we consider a nonsingular surface embedded by an adjoint line bundle. This can be viewed as a generalization of Vermier’s result. Assume that $X$ is a nonsingular projetive surface embedded by an adjoint line bundle $L_d=K_X+dA+B$ where $A$ is a very ample line bundle and $B$ is nef line bundle and $K_X$ is the canonical line bundle. Assume that $d\geq \dim X+1+3$. Then for any $a\geq 1$, one has $${\mathscr{I}}^a \mbox{ is } (2a+2) \mbox{-regular}.$$ It would be very interesting to find the effective value of $\operatorname{reg}{\mathscr{I}}^a$. We hope that there is a similar result as in the case of curves in Proposition \[pro:05\]. Let us conclude this section by the following proposition, which gives a bound for the regularity of ${\mathscr{I}}^a$ in terms of the regularity of ${\mathscr{I}}$. Assume that $X$ is a local complete intersection of dimension $2$ in a projective space defined by an ideal sheaf ${\mathscr{I}}$ which is $r$-regular, then for any $a\geq 1$, one has $${\mathscr{I}}^a \mbox{ is } (ar+r-2) \mbox{-regular}.$$ Again this formula is sharp. For example if $X$ is a Veronese surface, then it has the minimal degree and therefore has the minimal regularity $2$. In this situation, we see that $\operatorname{reg}{\mathscr{I}}^a=2a$. [CHT99]{} Steven Dale Cutkosky, Lawrence Ein, and Robert Lazarsfeld. Positivity and complexity of ideal sheaves. , 321(2):213–234, 2001. Marc Chardin, Amadou Lamine Fall, and Uwe Nagel. Bounds for the [C]{}astelnuovo-[M]{}umford regularity of modules. , 258(1):69–80, 2008. Marc Chardin. Powers of ideals and the cohomology of stalks and fibers of morphisms, 2010. S. Dale Cutkosky, J[ü]{}rgen Herzog, and Ng[ô]{} Vi[ê]{}t Trung. Asymptotic behaviour of the [C]{}astelnuovo-[M]{}umford regularity. , 118(3):243–261, 1999. Lawrence Ein and Robert Lazarsfeld. Syzygies and [K]{}oszul cohomology of smooth projective varieties of arbitrary dimension. , 111(1):51–67, 1993. M. Green and R. Lazarsfeld. Some results on the syzygies of finite sets and algebraic curves. , 67(3):301–314, 1988. Vijay Kodiyalam. Asymptotic behaviour of [C]{}astelnuovo-[M]{}umford regularity. , 128(2):407–411, 2000. Robert Lazarsfeld. , volume 48 of [ *Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics \[Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics\]*]{}. Springer-Verlag, Berlin, 2004. Classical setting: line bundles and linear series. Wenbo Niu. A vanishing theorem and asymptotic regularity of powers of ideal sheaves, 2010. Irena Swanson. Powers of ideals. [P]{}rimary decompositions, [A]{}rtin-[R]{}ees lemma and regularity. , 307(2):299–313, 1997. Peter Vermeire. On the regularity of powers of ideal sheaves. , 131(2):161–172, 2002.

ArXiv

--- abstract: 'We report on the results from the analysis of our 114 ks *Chandra* HETGS observation of the Galactic core-collapse supernova remnant G292.0+1.8. To probe the 3D structure of the clumpy X-ray emitting ejecta material in this remnant, we measured Doppler shifts in emission lines from metal-rich ejecta knots projected at different radial distances from the expansion center. We estimate radial velocities of ejecta knots in the range of -2300 $\lesssim$ $v_r$ $\lesssim$ 1400 km s$^{-1}$. The distribution of ejecta knots in velocity vs. projected-radius space suggests an expanding ejecta shell with a projected angular thickness of $\sim$90$\arcsec$ (corresponding to $\sim$3 pc at d = 6 kpc). Based on this geometrical distribution of the ejecta knots, we estimate the location of the reverse shock approximately at the distance of $\sim$4 pc from the center of the supernova remnant, putting it in close proximity to the outer boundary of the radio pulsar wind nebula. Based on our observed remnant dynamics and the standard explosion energy of $10^{51}$ erg, we estimate the total ejecta mass to be $\lesssim$ 8 M$_{\astrosun}$, and we propose an upper limit of $\lesssim$ 35 M$_{\astrosun}$ on the progenitor’s mass.' author: - 'Jayant Bhalerao, Sangwook Park, Daniel Dewey, John P. Hughes , Koji Mori, and Jae-Joon Lee' title: 'X-Ray Ejecta Kinematics of the Galactic Core-Collapse Supernova Remnant G292.0+1.8' --- \[sec:intro\] INTRODUCTION ========================== G292.0+1.8 is a Galactic oxygen-rich (O-rich) core-collapse supernova remnant (CC SNR) that has been studied at different wavelengths over the past five decades. Previous studies have captured a complex portrait composed of typical elements for a CC SNR a pulsar (Camilo et al. 2002; Hughes et al. 2003), and its wind-blown nebula or pulsar wind nebula (PWN, Hughes et al. 2001; Gaensler & Wallace 2003 (GW03 hereafter); Park et al. 2007), the blast wave-shocked circumstellar medium (CSM, Park et al. 2002; Lee et al. 2009; Lee et al. 2010 (L10 hereafter)), and metal-rich ejecta knots strewn across the interior in intricate filamentary networks (Gonzalez & Safi-Harb 2003; Park et al. 2004; 2007; Ghavamian et al. 2005; 2009; 2012; Winkler & Long 2006; Winkler et al. 2009). Yet details about the progenitor star and how its explosion led to the complex patterns of shocked ejecta and CSM seen in the sky today, remain elusive. The mass of the progenitor star has not been tightly constrained ($\sim$20-40 M$_{\astrosun}$, Hughes & Singh 1994; Gonzales & Safi-Harb 2003; Park et al. 2004; L10; Kamitsukasa et al. 2014; Yang et al. 2014). It is unclear if the progenitor has gone through phases other than the red supergiant (RSG). The presence of the equatorial belt (a bright, belt-like emission feature of shocked dense CSM enhanced along the “equator” of the SNR, Park et al. 2002; Ghavamian et al. 2005; Lee et al. 2009) suggests that the progenitor was probably rapidly rotating and/or in a binary system, but extensive studies on the progenitor system have not been performed. The associated pulsar (PSR J1124-5916) is apparently off the geometric center of the SNR (e.g., Hughes et al. 2001) indicating a significant pulsar-kick which could be related to a non-symmetric SN explosion (Park et al. 2007). The details of the pulsar-kick and its relationships with the progenitor system and explosion mechanism in G292.0+1.8 are not known. In contrast to Cassiopeia A (Cas A, a $\sim$10 times younger cousin of G292.0+1.8), in which abundant Fe-group ejecta material is observed (e.g., Hwang & Laming 2012, HL12 hereafter), such explosive nucleosynthesis products had not been detected in G292.0+1.8. Recently, a *Suzaku* study detected faint Fe K-shell line emission in G292.0+1.8, probably originating from hot Fe-rich ejecta (Kamitsukasa et al. 2014). A supernova (SN) explosion releases elements synthesized by the life-long efforts of a star (somewhat modified during its explosion) as metal-rich ejecta gas that expands into the surrounding CSM. The interaction of the rapidly expanding ejecta with the surrounding CSM creates two powerful shock fronts: an outward-moving forward shock (FS) that heats the CSM, and an inward-moving reverse shock (RS) that propagates back heating the metal-rich ejecta near the SNR center (e.g., a recent review by Dewey 2010). The FS is clearly identified in G292.0+1.8 as the outermost boundary of the remnant in X-rays (L10), in radio (GW03), and in infrared (Lee et al. 2009; Ghavamian et al. 2009; 2012). The location of the RS is difficult to ascertain because the 3D ejecta distribution is projected on the plane of the sky. A useful method to probe the 3D structure of an SNR is to study the line-of-sight distribution of fast-moving ejecta knots by measuring their radial velocities ($v_r$). Mapping the 3D distribution of ejecta may help locate the RS front. The High Energy Transmision Grating Spectrometer (HETGS) on board *Chandra* provides a powerful high resolution spectroscopy to estimate Doppler shifts in the X-ray spectral lines of metal-rich ejecta knots, a measure of their $v_r$. The utility of this method has been successfully demonstrated with the bright ejecta-dominated SNR Cas A (e.g., Lazendic et al. 2006). Based on our *Chandra* HETGS observations, we apply a similar method to map the $v_r$ distribution of 33 bright knots and filaments in G292.0+1.8. Here we provide the first insight into the 3D internal architecture of this textbook-type CC SNR in X-rays. \[sec:obs\] OBSERVATIONS & DATA REDUCTION ========================================= We performed our *Chandra* HETGS observation of G292.0+1.8 between 2011 March 20 and 2011 March 27. The aim point was set at RA(J2000.0) = 11$^h$ 24$^m$ 39$^s$.5, Dec(J2000.0) = -5915$\arcmin$ 56$\arcsec$.40 to detect a majority of bright ejecta knots within $\sim$2$\arcmin$ off-axis. The observation was composed of three ObsIDs (12555, 13242, and 13243). In each observation all six ACIS-S CCDs were operated in the full-frame readout mode. We processed the raw event files using CIAO version 4.4 and CALDB version 4.4.3. We followed the standard data reduction methods involving grade and hot pixel filtering. We found no significant contamination from flaring background. We processed each ObsID individually, and all three ObsIDs were combined for data analysis to yield a total effective exposure of $\sim$114 ks. As supplementary data (see § 3), we also used the archival ACIS-I data of G292.0+1.8 (Park et al. 2007). We reprocessed all six ObsIDs of the ACIS-I data following standard data reduction procedures with CIAO version 4.3 and CALDB version 4.4.3, which resulted in a total effective exposure of 509 ks. \[sec:result\] ANALYSIS & RESULTS ================================= We extracted source spectra from numerous small regions in G292.0+1.8, and measured line center energies using methods similar to those applied for the study of ejecta knots in Cas A (Lazendic et al. 2006; Rutherford et al. 2013). We used a fixed order-sorting range of $\pm$10% to extract the first-order spectrum (Figure 1). We created the zeroth-order image of the ObsID with the longest exposure (ObsID 12555) in the 0.82.2 keV band in which bright K$\alpha$ lines from He- and H-like Ne, Mg and Si ions are present. Based on this image we identified the zeroth-order locations of bright, compact knots which would have small cross-dispersion widths ($\sim$2$\arcsec$ 9$\arcsec$ in angular sizes, and $\sim$4$\arcsec$ on average, for which the angular dispersion of these small source regions do not affect our Doppler line shift measurements). Using these line centers and cross-dispersion widths, we extracted the dispersed spectra from these small knots from all three ObsIDs applying standard CIAO tools  TGCat scripts.[^1] We show an example of an HETGS spectrum extracted from a small bright emission feature in Figure 1. We analyzed the first-order spectra, corresponding to orders MEG $\pm$1 and HEG $\pm$1, using custom scripts executed in the ISIS software package[^2] (Houck & Denicola 2000). For each knot, we combined the spectra extracted from all three ObsIDs. Five emission lines are useful for Doppler shift measurements of individual ejecta knots: atomic emission lines from the K-shell transitions in the He- and H-like ions of Ne and Mg, and in the He-like Si. The rest wavelengths for these lines are listed in Table 1. We detect and characterize these lines in the dispersed spectra of small knots in G292.0+1.8 using simple phenomenological model fits applied to a narrowband around each line. Our model consists of two Gaussians for the $\mathrm{Ly\alpha}$ lines (one for the line and the other to approximate the underlying continuum) and four Gaussians for the $\mathrm{He\alpha}$ triplets (three corresponding to the forbidden (*f*), intercombination (*i*) and resonance (*r*) lines, and one for the underlying continuum). We use a sum of broad Gaussians to approximate the underlying continua in the five line-regions that we fit: the Gaussians provide a computationally simple method that allows each local continuum level to be adjusted with reasonable independence, since the Gaussians decrease quickly outside of their wavelength ranges. Free parameters in our model are the line center, the line flux, the line width ($\sigma$), and the continuum flux. For the continuum Gaussian component, we fixed the center energy at the rest wavelength of the corresponding line while varying the area of the Gaussian. The model fits for the $\mathrm{He\alpha}$ triplets have the same degrees of freedom as those of the $\mathrm{Ly\alpha}$ lines, because the wavelength and fluxes of the *f* and *i* lines are set to be proportional to those of the *r* line. The *i/r* and *f/r* flux ratios we used were based on the observed values for Capella and SN 1987A (Canizares et al. 2000; Dewey et al. 2008) in which we assumed a low-density gas (which should also be the case for G292.0+1.8). We list these flux ratios in Table 1. We note that our primary goal of line shift measurements is not very sensitive to the exact ratios between these triplet lines. Also, the counting statistics dominate the observed line fluxes, and our Doppler velocity shift measurements are based on several emission line complexes (Table 2). For our Doppler velocity shift measurements, we first fitted each of the five lines listed above to detect a valid line feature. We scaled all the lines in the model to the model wavelength of the Ne IX line. The detected lines were then jointly fitted to estimate a common velocity shift. We fitted 65 small knots with these models, and used 33 knots that show statistically acceptable fits ( $\textless$ 2, for a combined fit of all detected lines) for our Doppler shift measurements. We excluded 32 knots from our $v_r$ measurements because their low signal-to-noise ratio did not allow us to detect valid line features. We show extracted spectra and best-fit models for three example regions in Figure 2. Based on the shifts in our measured line centers from the rest wavelengths, we estimate $v_r$ for these knots (Table 2). To identify the origins (shocked ejecta vs. CSM) of these 33 knots, we investigated their spectral properties using our deep 509 ks ACIS-I observation of G292.0+1.8 (Park et al. 2007). We used the ACIS data to utilize the significantly higher photon statistics (by more than an order of magnitude in the 0.3-5 keV band) than those in the HETG data . We performed spectral model fits for the observed ACIS spectra of these 33 regions to measure their metal abundances. For these spectral model fits we subtracted the background emission spectrum using spectra extracted from nearby source-free (dark, *ejecta-free*) regions within the SNR. We performed spectral model fits using the absorbed (*phabs* in XSPEC) non-equilibrium ionization (NEI) plane-parallel shock model (Borkowski et al. 2001) with variable abundances (*vpshock*, NEI version 2.0 with augmented ATOMDB, Smith et al. 2001; Badenes et al. 2006). We added a power law component for regions projected within or near the PWN. We varied O, Ne, Mg, Si, S and Fe abundances while fixing other elemental abundances at solar values (Anders & Grevesse 1989). Based on these abundance measurements we identified 24 ejecta knots (showing abundances typically $>$ several times solar for one or more elements). We identified 9 CSM-like features with sub-solar abundances for all fitted elements (Table 2). Most of the CSM features are positioned along the equatorial belt. We constructed a $v_r$$r_p$ distribution for these 33 knots (Figure 3), where $r_p$ is the projected distance from the expansion center. For a homologous expansion of ejecta knots in G292.0+1.8, the 3D spatial velocities ($v_{3D}$) of individual ejecta knots are proportional to their physical distances or 3D radii ($r_{3D}$) from the expansion center. The constant relating this proportionality is $r_{3D}/v_{3D}$ = 0.1055$\mathrm{\arcsec/km\,s^{-1}}$, assuming an expansion age of 3000 yr (Winkler et al. 2009) and a distance to the SNR of 6 kpc (GW03). Knots at the same $r_{3D}$ will differ in their $r_p$ and $v_r$ values depending on their projected locations. In Figure 3 we overlay four elliptical loci to relate $v_r$ and $r_p$ from the SNR’s expansion center, assuming this proportionality constant. The smallest elliptical locus corresponds to a physical distance ($r_{3D}$ from the expansion center) of $\sim$3.5 pc (at the projected angular distance $\sim$120) and roughly represents the angular size of the radio PWN, GW03). The next two loci at $\sim$3.8 pc (at $\sim$130) and $\sim$6.4 pc (at $\sim$220) have been qualitatively estimated by eyeball inspection to contain the majority of the ejecta knots within a shell. The outermost locus at $\sim$7.7 pc ($\sim$265) corresponds to the FS (L10). We roughly estimate (by eyes) the velocity centroid at +150 km s$^{-1}$ which is similar to that estimated in the optical band (Ghavamian et al. 2005). We show the projected positions for the 33 regional features and these elliptical loci in Figure 4. \[sec:disc\] DISCUSSION ======================= Our estimated radial velocity range of -2300 $\lesssim$ $v_r$ $\lesssim$ 1400 km s$^{-1}$ for X-ray ejecta knots is in plausible agreement with earlier optical measurements of ejecta velocities in G292.0+1.8. Ghavamian et al. (2005) reported ejecta radial velocities in the range of -1700 $\lesssim$ $v_r$ $\lesssim$ +1700 km s$^{-1}$ for O-rich optical ejecta knots in G292.0+1.8. Winkler et al. (2009) conducted proper motion studies of O-rich knots in the optical band, and measured east-west velocities of -1800 d$_6$ &lt; $v_x$ &lt; 1490 d$_6$ km s$^{-1}$ and north-south velocities in the range of -3570 d$_6$ &lt; $v_y$ &lt; 2340 d$_6$ km s$^{-1}$, where d$_6$ is the distance to G292.0+1.8 in units of 6 kpc. While X-ray and optical emissions originate from ejecta gas with different thermal conditions, and thus X-ray ejecta knots generally do not show optical counterparts, we detect some spatial correlations between X-ray and optical ejecta knots. The highly redshifted knot E7 shows positional coincidence with the largely redshifted optical spur in the southeast region of the SNR, and the blueshifted knots E5 and E11 are in similar positions to blueshifted optical knots in the northern parts of the SNR. Spatial correlation between X-ray and optical emission is also supported by the observation that several X-ray filamentary structures in the north coincide with optical knots located near their termini (Figure 13 in Ghavamian et al. 2005). Thus, in G292.0+1.8 the ejecta gas at various thermal states appears to share some bulk motion. We detect a significantly larger number of blueshifted knots than redshifted ones (17 of 24 ejecta knots are blueshifted). For the blueshifted ejecta knots, we also estimate generally higher velocity magnitudes than the redshifted ones: e.g., seven blueshifted knots show $v_r$ &gt; 1000 km s$^{-1}$ while only one redshifted ejecta knot shows such a high $v_r$. A similar non-symmetric $v_r$ distribution of ejecta in G292.0+1.8 was observed in the optical band, where a significantly larger number of blueshifted knots was detected, especially in the north (Ghavamian et al. 2005). Asymmetries in $v_r$ have also been seen in other O-rich SNRs, for which interpretations included asymmetric SN explosions and density variations in the CSM. SNR 1E 0102.27219, in the Small Magellanic Cloud, shows a larger number of blueshifted bright knots but the redshifted knots show generally higher $v_r$ (Vogt & Dopita 2010). SNR 0540-69.3, in the Large Magellanic Cloud (LMC), shows a generally redshifted spectrum of ejecta (Kirshner et al. 1989). SNR N132D, also in the LMC, shows higher $v_r$ in its blueshifted ejecta (Vogt & Dopita 2011). Cas A shows higher $v_r$ in its redshifted ejecta (Milisavljevic & Fesen 2013 and references therein). Possible origins for the observed $v_r$ asymmetry in G292.0+1.8 may include several scenarios such as an asymmetric SN explosion, CSM density variations (near vs. far sides of the SNR) along the line-of-sight, a clumpiness variation of the ejecta, and self-absorption of redshifted emission. We discuss each of these scenarios below. An asymmetric SN explosion may have channeled more kinetic energy towards the Earth along the line-of-sight. Observational evidence supporting an asymmetric SN explosion for G292.0+1.8 has been reported in previous works: e.g., higher X-ray ejecta temperatures in the northwest than in the southeast regions (Park et al. 2007), the absence of Si emission in the southeast (Park et al. 2002; Ghavamian et al. 2012), higher proper motions of optical ejecta knots along the north-south than in the east-west directions (Winkler et al. 2009), and the $\gtrsim$1 pc displacement (to southeast from the SNR’s expansion center) of the associated pulsar PSR J1124-5916 (e.g., Winkler et al. 2009). In such an asymmetric SN explosion, the energy output might have resulted in a larger amount of blueshifted fast-moving ejecta material as observed in X-rays (this work) and in optical (Ghavamian et al. 2005). Another tentative scenario for the observed $v_r$ asymmetry could be a non-uniform CSM. For instance, a significant CSM density variation between the near and far sides of the SNR might have created asymmetry in the RS structure, causing a greater inward migration of the RS on the near side (if the CSM density is higher there) than the far side, thus interacting with more ejecta material to produce more blueshifted material. A CSM density variation in G292.0+1.8 is suggested by large filamentary structures such as the equatorial belt (Park et al. 2004; Lee et al. 2009; L10; Ghavamian et al. 2012), and by a non-uniform circumstellar environment as seen in the mid-infrared (Park et al. 2007). Some azimuthal CSM density variation has been observed in G292.0+1.8 with regions in the southeast showing lower CSM densities than other regions (L10). However, it is not clear if this azimuthal CSM density structure orginated from variation in the progenitor’s wind density or from an asymmetric SN explosion. Also, a deeper migration of the RS on the near side of the SNR would create more blueshifted material with lower $v_r$ from the heating of slower moving central ejecta regions. One would therefore expect to see a larger number of low $v_r$ blueshifted ejecta regions projected near the SNR center, which is not clearly evident. Hence, the presence and contribution of a CSM density variation along the line-of-sight between the near and far sides of the SNR that would result in the observed blueshift-dominated ejecta in G292.0+1.8 is unclear, although it cannot be ruled out. A selection effect due to a clumpiness variation of the metal-rich ejecta in the SNR might also have contributed to the observed blueshift predominance in G292.0+1.8. Since we are more likely to select small bright knots (for our $v_r$ measurements) that would originate in clumpier regions than in smoother plasma, an SNR with a substantially larger number of clumpy ejecta features on the near side could result in the observed blueshift predominance. An asymmetric clumpy ejecta distribution has been proposed to explain observed optical emission line asymmetries in SNe 1993J (Spyromilio 1994) and 1990I (Elmhamdi et al. 2004), with further support in theoretical studies (e.g., Herrington et al. 2010). The observed $v_r$ asymmetry in G292.0+1.8 might have originated from self-absorption of redshifted emission by material within the SNR (e.g., ejecta dust, Ghavamian et al. 2009; 2012). However, there is no observational evidence for significant self-absorption of X-ray emission that could lead to the observed $v_r$ asymmetry in G292.0+1.8: e.g., we find that the column density ($N_{H}$) for the highly redshifted knot E7 is consistent with that for blueshifted regions. In general the existing observational evidence appears to favor an asymmetric explosion scenario for the observed $v_r$ asymmetry. However, the true origin remains elusive. Most of the ejecta knots in our study occupy a thick shell in $v_r$$r_p$ space, corresponding to the RS-heated hot ejecta gas (solid diamonds in Figure 3). On the other hand, most of the CSM filaments occupy the low $v_r$ region (open circles in Figure 3), and are positioned along the equatorial belt (Figure 4). Unless our 24 ejecta knots represent a heavily biased sample, the inner and outer radii of this ejecta shell may roughly correspond to the locations of the RS ($r_{in}$ $\sim$ 130$\arcsec$) and contact discontinuity (CD, $r_{out}$ $\sim$ 220$\arcsec$) respectively (the projected angular distances $r_{in}$ and $r_{out}$ are measured from the explosion site determined from the proper motions of optical ejecta knots (Winkler et al. 2009)). We find that nearly all of 62 fast-moving optical ejecta knots (Ghavamian et al. 2005) also lie within our X-ray-estimated ejecta shell (we estimate that only two of them are positioned at a slightly larger radius than our CD), further supporting our inferred location of the RS. We note that there are a few regions with large uncertainties in $v_r$, likely due to the relatively weak emission lines in these features. For example, $v_r$ measurements for regions C28, E29, C31 and E33 were based on only one, relatively faint line ($\mathrm{Si\:He\alpha}$, Table 2). Our estimate of the RS location gives a ratio between the radii of the RS and FS, R$_\text{RS}$/R$_\text{FS}$ $\sim$130$\arcsec$/265$\arcsec$ $\sim$ 0.5. This ratio is consistent with previous estimates at other wavelengths: $\sim$0.47 by Braun et al. (1986) based on radio and infrared data, and $\sim$0.5 by GW03, based on radio data. This R$_\text{RS}$/R$_\text{FS}$ ratio in G292.0+1.8 is similar to that seen in other young O-rich SNRs: e.g., $\sim$0.6-0.8 for Cas A (at age $\sim$ 330 yr, HL12) and $\sim$0.5-0.7 for 1E 0102.27219 (at age $\sim$ 1000 yr, Gaetz et al. 2000; Flanagan et al. 2004). A smaller ratio of $\sim$0.3 was estimated in N132D (age $\sim$ 2500 yr) suggesting a dynamically more evolved stage for this SNR, with its RS possibly accelerating towards the SNR center (Vogt & Dopita 2011). Our estimated R$_\text{RS}$/R$_\text{FS}$ ratio for G292.0+1.8 is significantly smaller than the values predicted by self-similar solutions (R$_\text{RS}$/R$_\text{FS}$ $>$ 0.7, Chevalier 1982), suggesting that this SNR has evolved beyond the early ejecta-dominated phase. Truelove & McKee (1999) (TM99 hereafter) developed a hydrodynamic framework that extends the model to later times when the RS reaches the ejecta core region and the FS approaches the late-time Sedov-Taylor phase. They presented explicit results for the SNR evolution in a uniform density medium. Laming & Hwang (2003) and HL12 (LH03-HL12 hereafter) extended the TM99 model for SNRs expanding into stellar winds with radial mass density profile $ \rho \propto r^{-2}$. Since G292.0+1.8 is expanding into an RSG wind (L10), we applied the LH03-HL12 model for G292.0+1.8 and successfully reproduce our estimated R$_\text{RS}$/R$_\text{FS}$ $\sim$ 0.5 at age = 3000 yr. The age of $\sim$3000 yr has been estimated for G292.0+1.8 based on kinematic studies (Ghavamian et al. 2005; Winkler et al. 2009), which is similar to the characteristic spin-down age of PSR J1124-5916 in G292.0+1.8 (2900 yr, Camilo et al. 2002). For these model calculations we assumed a canonical explosion energy E$_\text{0}$ = $1 \times 10^{51}$ erg, and a preshock CSM density 0.1-0.3 $\mathrm{ cm^{-3}}$ at R$_\text{FS}$ = 7.7 pc (L10). Based on the LH03-HL12 model, assuming the power-law index of 5 $\lesssim$ n $\lesssim$ 14 for the ejecta radial density profile in the outer layers (with an inner *constant* density core) and SNR age of 2890 $\lesssim$ t $\lesssim$ 3080 yr (Winkler et al. 2009), we calculate the total ejecta mass $\mathrm{M_{ej}}$ $\lesssim$ 8 M$_{\astrosun}$. Combining our upper limit for $M_{ej}$ with our previously determined averaged wind mass estimate of $M_w$ $\sim$ 25 M$_{\astrosun}$ (L10), we suggest an upper limit of $\sim$35 M$_{\astrosun}$ for the progenitor mass ($M_{prog}$) of G292.0+1.8. This upper limit provides a constraint on $M_{prog}$ based on the observed dynamics of the SNR, and is in plausible agreement with previous nucleosynthesis-based estimates for $M_{prog}$: e.g., 25 $M_{\astrosun}$ (Hughes & Singh 1994; Park et al. 2004), 30-40 $M_{\astrosun}$ (Gonzales & Safi-Harb 2003), 30-35 $M_{\astrosun}$ (Kamitsukasa et al. 2014) and 25-30 $M_{\astrosun}$ (Yang et al. 2014). Considering the sharp boundary between the PWN and the outer plateau in radio, GW03 suggested an RS-PWN interaction, probably in the very early stage where the RS has not compressed the PWN significantly. On the other hand, a large pressure difference between the PWN and the thermal gas in X-rays had suggested that the RS and PWN had not yet interacted (Park et al. 2004). Also, a large O/Ne mass ratio in the mid-infrared suggested that the inner explosive nucleosynthesis products might not have undergone significant mixing, and could still remain unshocked (Ghavamian et al. 2009), which generally supports a non-interaction between the RS and PWN. Our estimated RS location is overall close to the outer boundary of the radio PWN (Figure 4). The position of the X-ray PWN in G292.0+1.8 is generally consistent with its radio counterpart, and the projected angular extent of the X-ray PWN is smaller than that of the radio PWN (Figure 5). This X-ray-radio PWN size difference is consistent with the standard picture of more effective synchrotron loss of X-ray emission in the outer layers of PWNs (e.g., Gaensler & Slane 2006). Theoretical and observational studies suggest that late stages of PWN-RS interactions are characterized by irregular PWN morphologies. These studies also suggest that asymmetry in the RS structure of SNRs will result in displacement of PWNs relative to their PSRs, and inconsistencies in the sizes and positions of radio PWNs and their X-ray counterparts in late PWN-RS interaction stages (e.g., Gaensler & Slane 2006 and references therein). These signs of late-stage PWN-RS interactions are not clearly evident in G292.0+1.8. Therefore, a PWN-RS interaction, if it has started, should be in an early stage in this SNR (as suggested by GW03). We note that Park et al. (2004) estimated the thermal pressure of the SNR using only a small region on the equatorial belt. The RS front in G292.0+1.8 may not be smooth or spherically symmetric after interacting with a non-uniform CSM. Then we consider that the large pressure difference between the PWN and thermal gas estimated by Park et al. (2004) might have been a local effect. X-ray thermal pressure measurements from extensive areas of the SNR would be helpful to test this discrepancy (which is beyond the scope of this paper, and will be included in our follow-up work on the ACIS-I data analysis). If the RS front is close to the PWN (and probably interacting with it), it may be nearing the SNR’s central region where the Fe-rich ejecta might be expanding. A recent *Suzaku* detection of the Fe K-shell emission line in the hot ejecta gas of G292.0+1.8 (Kamitsukasa et al. 2014) may support this scenario, opening further avenues in the quest to decipher this complex remnant. \[sec:summ\] SUMMARY ==================== Based on our $\sim$114 ks *Chandra* HETGS observation of G292.0+1.8, we measure $v_r$ from Doppler line shifts for 33 bright knots in the SNR. Our measured $v_r$ is in the range of -2300 $\lesssim$ $v_r$ $\lesssim$ 1400 km s$^{-1}$. We detect a $v_r$ asymmetry with a larger number of blueshifted ejecta knots than redshifted ones. Our measured $v_r$ range and observed blueshifted ejecta knot predominance are generally consistent with results from optical observations (Ghavamian et al. 2005). We find that the blueshifted X-ray ejecta knots generally show higher velocity magnitudes than the redshifted ones. Other O-rich SNRs have also been found to show $v_r$ asymmetry. The cause for the $v_r$ asymmetry in G292.0+1.8 may have been an asymmetric SN explosion, although environmental effects such as CSM density variations along the line-of-sight cannot be ruled out. Based on the distribution of the ejecta knots in $v_r$$r_p$ space, we qualitatively locate the positions of the RS and CD. Our inferred RS position agrees with previous estimates based on radio and IR data, and with the hydrodynamic model for an SNR expanding into an RSG wind. Employing the SNR’s dynamics we calculate the total ejecta mass of $\lesssim$ 8 M$_{\astrosun}$, and propose an upper limit of $\sim$35 M$_{\astrosun}$ for the G292.0+1.8 progenitor mass. Our inferred location of the RS places it in close proximity to the outer boundary of the PWN, suggesting the possibility of early-stage PWN-RS interactions, and the possible onset of inner Fe-rich ejecta-heating by the RS. This work was supported in part by the SAO through *Chandra* grant GO1-12077X. JB acknowledges support from the NASA Texas Space Grant Consortium. DD was supported by NASA through SAO contract SV3-73016 to MIT for support of the *Chandra* X-ray Center and Science Instruments. Anders, E. & Grevesse, N. 1989, Geochim. Cosmochim. Acta, 53, 197 Badenes, C., Borkowski, K. J., Hughes, J. P., Hwang, U., & Bravo, E. 2006, ApJ, 645, 1373 Borkowski, K. J., Lyerly, W. J., & Reynolds, S. P. 2001, ApJ, 548, 820 Braun, R., Goss, W. M., Caswell, J. L., & Roger, R. S. 1986, , 162, 259 Camilo, F., Manchester, R. N., Gaensler, B. M., Lorimer, D. R., & Sarkissian, J. 2002, ApJ, 567, L71 Chevalier, R. A. 1982, ApJ, 258, 790 Chevalier, R. A. 2005, ApJ, 619, 839 Canizares. C. R., Huenemoerder, D. P., Davis, D.S., et al. 2000, ApJ, 539, L41 Dewey, D., Zhekov, S. A., McCray, R. & Canizares, C. R. 2008, ApJ, 676, L131 Dewey, D. 2010, SSRv, 157, 229 Drake, G.W. 1988, Can. J. Physics 66, 586 Elmhamdi, A., Danziger, I. J., Cappellaro, E., et al. 2004, A & A, 426, 963 Flanagan, K. A., Canizares, C. R., Dewey, D., et al. 2004, ApJ, 605, 230 Gaensler, B. M. & Wallace, B. J. 2003, ApJ, 594, 326 (GW03) Gaensler, B. M. & Slane, P. O. 2006, , 44, 17 Gaetz, T. J., Butt, Y. M., Edgar, R. J., et al. 2000, ApJ, 534, L47 Ghavamian, P., Hughes, J. P., & Williams, T. B. 2005, ApJ, 635, 365 Ghavamian, P., Raymond, J. C., Blair, W. P., et al. 2009, ApJ, 696, 1307 Ghavamian, P., Long, K. S., Blair, W. P., et al. 2012, ApJ, 750, 39 Gonzalez, M. & Safi-Harb, S. 2003, ApJ, 583, L91 Herrington, J., Ignace, R. & Hole, T. K., 2010, JSARA, 4, 15 Houck, J. C. & Denicola, L. A. 2000, in ASP Conf. Ser. 216, Astronomical Data Analysis Software and Systems IX, ed. N. Manset, C. Veillet, & D. Crabtree, (San Francisco, CA:ASP), 591 Hughes, J. P. & Singh, K. P. 1994, ApJ, 422, 126 Hughes, J. P., Slane, P. O., Burrows, D. N., et al. 2001, ApJ, 559, L153 Hughes, J. P., Slane, P. O., Park, S., Roming, P. W. A., & Burrows, D. N. 2003, ApJ, 591, L139 Hwang, U. & Laming, J. M. 2012, ApJ, 746, 130 (HL12) Johnson, W.R., & Soff, G. 1985, ADNDT, 33, 405 Kamitsukasa, F., Koyama, K., Tsunemi, H., et al. 2014, PASJ, 66, 648 Kirshner, R. P., Morse, J. A., Winkler, P. F., & Blair, W. P. 1989, ApJ, 342, 260 Laming, J. M. & Hwang, U. 2003, ApJ, 597, 347 (LH03) Lazendic, J. S., Dewey, D., Schulz, N. S., & Canizares, C. R. 2006, ApJ, 651, 250 Lee, H.-G., Koo, B.-C., Moon, D.-S., et al. 2009, ApJ, 706, 441 Lee, J.-J., Park, S., Hughes, J. P., et al. 2010, ApJ, 711, 861 (L10) Milisavljevic, D. & Fesen, R. A. 2013, ApJ, 772, 134 Park, S., Roming, P. W. A., Hughes, J. P. et al. 2002, ApJ, 564, L39 Park, S., Hughes, J. P., Slane, P. O., et al. 2004, ApJ, 602, L33 Park, S., Hughes, J. P., Slane, P. O., et al. 2007, ApJ, 670, L121 Rutherford, J., Dewey, D., Figueroa-Feliciano, E., et al. 2013, ApJ, 769, 64 Smith, R. K., Brickhouse, N. S., Liedahl, D. A. & Raymond, J. C. 2001, ApJ, 556, L91 Spyromilio, J. 1994, MNRAS, 266, L61 Truelove, J. K., & McKee, C. F. 1999, ApJS, 120, 299 Vogt, F. & Dopita, M. A., 2010, ApJ, 721, 597 Vogt, F. & Dopita, M. A., 2011, , 331, 521 Winkler, P. F. & Long, K. S. 2006, ApJ, 132, 360 Winkler, P. F., Twelker, K., Reith, C. N., & Long, K. S. 2009, ApJ, 692, 1489 Yang, X.-J., Liu, X.-Q., Li, S.-Y., & Lu, F.-J. 2014, RAA, 14, 1279 [cccc]{} $\mathrm{Ne\: IX\: \textit{r}\phm{a} (Ne\: He\alpha\: \textit{r})\phm{a}}$ & 13.447 & 0.26 & 0.59\ $\mathrm{Ne\: IX\: \textit{i}\phm{a} (Ne\: He\alpha\: \textit{i})\phm{a}}$ & 13.553 & &\ $\mathrm{Ne\: IX\: \textit{f}\phm{a} (Ne\: He\alpha\: \textit{f})\phm{a}}$ & 13.699 & &\ $\mathrm{Ne\: X\phm{a} (Ne\: Ly\alpha)~~~~~\phm{a}}$ & 12.134 & &\ $\mathrm{Mg\: XI\: \textit{r}\phm{a} (Mg\: He\alpha\: \textit{r})}$ & 9.169 & 0.21 & 0.43\ $\mathrm{Mg\: XI\: \textit{i}\phm{a} (Mg\: He\alpha\: \textit{i})}$ & 9.231 & &\ $\mathrm{Mg\: XI\: \textit{f}\phm{a} (Mg\: He\alpha\: \textit{f})}$ & 9.314 & &\ $\mathrm{Mg\: XII\: (Mg\: Ly\alpha)~~~}$ & 8.421 & &\ $\mathrm{Si\: XIII\: \textit{r}\phm{a} (Si\: He\alpha\: \textit{r})}$ & 6.648 & 0.23 & 0.43\ $\mathrm{Si\: XIII\: \textit{i}\phm{a} (Si\: He\alpha\: \textit{i})}$ & 6.688 & &\ $\mathrm{Si\: XIII\: \textit{f}\phm{a} (Si\: He\alpha\: \textit{f})}$ & 6.740 & &\ [cccrlccc]{} E1 & 125.8 & 102.4 & 41 & $\pm$297 & 1.59 & 1,2,3,4 & Ejecta\ E2 & 182.7 & 80.9 & -496 & $\pm$289 & 1.87 & 1,2,3,4 & Ejecta\ C3 & 221.9 & 252.1 & 677 & $\pm$751 & 1.30 & 2,3,5 & CSM\ E4 & 96.3 & 181.3 & -2221 & $\pm$407 & 1.48 & 2,3,4 & Ejecta\ E5 & 134.0 & 48.9 & -981 & $\pm$435 & 1.60 & 2,3,4 & Ejecta\ C6 & 62.5 & 90.0 & -503 & $\pm$2810 & 1.34 & 1,2 & CSM\ E7 & 54.3 & 117.7 & 1396 & $\pm$271 & 1.68 & 1,2,3,4 & Ejecta\ C8 & 15.4 & 97.5 & -792 & $\pm$655 & 1.94 & 2,3,5 & CSM\ C9 & 43.1 & 276.7 & -990 & $\pm$693 & 1.29 & 1,2,3,5 & CSM\ E10 & 92.1 & 248.9 & -1822 & $\pm$719 & 1.30 & 1,2,3,5 & Ejecta\ E11 & 75.0 & 330.5 & -1289 & $\pm$923 & 1.39 & 2,5 & Ejecta\ E12 & 69.7 & 306.2 & -1623 & $\pm$342 & 1.40 & 1,3,5 & Ejecta\ E13 & 156.0 & 324.8 & -479 & $\pm$235 & 1.40 & 1,2,3,4,5 & Ejecta\ E14 & 198.2 & 307.3 & -269 & $\pm$633 & 1.88 & 2,3,4 & Ejecta\ E15 & 172.1 & 301.2 & -78 & $\pm$332 & 1.59 & 1,2,3,4,5 & Ejecta\ C16 & 56.8 & 269.0 & 107 & $\pm$410 & 1.51 & 1,2,3,5 & CSM\ E17 & 203.2 & 333.2 & 320 & $\pm$413 & 1.57 & 2,3,4,5 & Ejecta\ E18 & 193.4 & 24.0 & -115 & $\pm$603 & 1.15 & 3,5 & Ejecta\ E19 & 57.3 & 210.8 & -453 & $\pm$712 & 1.57 & 2,3,4 & Ejecta\ E20 & 209.9 & 9.0 & 785 & $\pm$1048 & 1.01 & 1,2 & Ejecta\ E21 & 147.3 & 289.9 & 381 & $\pm$657 & 1.66 & 1,2,3,4 & Ejecta\ E22 & 203.0 & 355.8 & 839 & $\pm$452 & 1.33 & 2,3 & Ejecta\ E23 & 140.9 & 143.0 & -1346 & $\pm$313 & 1.05 & 2,3,4 & Ejecta\ E24 & 145.1 & 144.0 & -1074 & $\pm$271 & 1.14 & 1,2,3,4 & Ejecta\ E25 & 56.5 & 295.2 & -1007 & $\pm$528 & 1.85 & 2 & Ejecta\ E26 & 147.6 & 322.7 & -610 & $\pm$246 & 1.29 & 1,2,3,4 & Ejecta\ E27 & 163.1 & 328.9 & -230 & $\pm$389 & 1.19 & 2,3,4 & Ejecta\ C28 & 123.3 & 264.4 & -3087 & $\pm$1502 & 1.26 & 5 & CSM\ E29 & 172.2 & 245.2 & 394 & $\pm$1763 & 1.15 & 5 & Ejecta\ C30 & 85.6 & 265.3 & -420 & $\pm$778 & 1.04 & 2 & CSM\ C31 & 109.6 & 52.2 & -811 & $\pm$1528 & 1.74 & 5 & CSM\ C32 & 129.6 & 353.5 & -405 & $\pm$765 & 1.57 & 1,3 & CSM\ E33 & 76.9 & 336.0 & -2301 & $\pm$1799 & 1.39 & 5 & Ejecta\ \  \   [^1]: <http://tgcat.mit.edu/> [^2]: <http://space.mit.edu/CXC/ISIS/>

ArXiv

--- author: - | Ricardo López-Ruiz\ \ Department of Computer Science and BIFI,\ Facultad de Ciencias - Edificio B,\ Universidad de Zaragoza,\ 50009 - Zaragoza (Spain) date: title: '**An approach to the problem of generating irreducible polynomials over the finite field $GF(2)$ and its relationship with the problem of periodicity on the space of binary sequences**' --- -1 cm 2 cm [**Abstract**]{} A method for generating irreducible polynomials of degree $n$ over the finite field $GF(2)$ is proposed. The irreducible polynomials are found by solving a system of equations that brings the information on the internal properties of the splitting field $GF(2^n)$ . Also, the choice of a primitive normal basis allows us to build up a natural representation of $GF(2^n)$ in the space of $n$-binary sequences. Illustrative examples are given for the lowest orders. [**Keywords:**]{} Finite Fields; Generation of Irreducible Polynomials; Canonical Representations; Periodic Orbits; Computation on Binary Sequences; Solution’s Choice. [**AMS Subject Classifications:**]{} 11R32, 11R09, 11T71, 12E20, 12Y05. Introduction ============ The problem of determining the irreducible polynomials over a finite field of order $q$, $GF(q)$, is a task, from a qualitative mathematical point of view, similar to that of finding the prime numbers in the set of integer numbers [@hardy; @riesel]. Such polynomials are necessary in numerous applications. For instance, the generation of algebraic error-correcting codes and cryptography [@berlekamp] or the construction of fast transforms are based on finite field properties [@reed] that require algorithms to obtain the irreducible polynomials $f_n(x)$ of degree $n$ over a particular finite field. Tables of irreducible polynomials are found in the literature [@marsch]. These can be built up by different methods [@golomb; @shoup], including a close analog to the ’sieve’ method used in making tables of prime numbers. In this work a method to generate the irreducible polynomials of degree $n$ over the field $GF(2)$ is proposed. The strategy comprises the construction of a representation of the splitting field $GF(2^n)$ over the set of binary $n$-digits sequences. Each periodic orbit and its $m$ permutations, with $m$ a divisor of $n$, identify a irreducible polynomial of degree $m$ and its $m$ roots on $GF(2^n)$. In section $II$ the connection between the field $GF(2^n)$ and the space of $n$-binary sequences, ${\cal B}_n$, is identified. The two internal operations over ${\cal B}_n$ necessary in order to obtain a particular representation of $GF(2^n)$ are defined in section $III$. The method to find the irreducible polynomials $f_n(x)$ over $GF(2)$ is disclosed in section $IV$. Section $V$ summarizes the general scheme to be followed in order to obtain $f_n(x)$ for all $n$. Two illustrative examples are given in section $VI$. Last section is concerned with our conclusions. $GF(2^n)$ and ${\cal B}_n$ ========================== [**Notation**]{}: Let us call ${\cal B}_n$ the space of binary sequences of $n$ digits. ${\cal B}_n$ has $2^n$ elements. Each element $O\in{\cal B}_n$ is said to be an orbit. If $O$ is given by the binary sequence ($\alpha_{n-1}\alpha_{n-2}\cdots\alpha_1\alpha_0$), ($\alpha_i = 0\;\; or\;\; 1$), the left shifted orbit $O'$ defined by the sequence ($\alpha_{n-2}\alpha_{n-3}\cdots\alpha_1\alpha_0\alpha_{n-1}$) is represented by $shift(O)$. If $m$ is the smallest integer such that $O=shift^m(O)=shift~(~\stackrel{m}{\cdots}~(~shift~(~O)~)~)$ then the orbit $O$ is said to be of [*period $m$*]{} or $m$-periodic, and it is represented by $O_m$. It is obvious that if $O_m\in {\cal B}_n$ then $m$ is a divisor of $n$, $m\mid n$. The set $[O_m]$ formed by the $m$ different shifts of an $m$-periodic orbit, $O_m$, is called an orbital of period $m$: $[O_m]=\{O_m,shift(O_m),shift^2(O_m),\cdots,shift^{m-1}(O_m)\}$. As an example, $[O_3]=[100]=\{100,010,001\}$ or $[O_2]=[0101]=\{0101,1010\}\equiv [01]=\{01,10\}$. [**Fundamental Idea**]{}: Firstly, let us note that the number $N_n$ of irreducible polynomials (I.P.) of degree $n$ over $GF(2)$, is equal to the number of orbitals $[O_n]$ of period $n$ (table \[tabla1\],[@lopez; @mil]): $$N_n=\frac{2^n-\sum_{i=1}^{k}m_iN_{m_i}}{n},$$ where $\{m_1,m_2,\ldots,m_k\}$ are the integer divisors of $n$ less than $n$. The idea is to build up a representation of $GF(2^n)$ over ${\cal B}_n$, with the constraint that each orbital $[O_m]$ contained in ${\cal B}_n$ has an associated I.P. , $f_m(x)$, of degree $m$ over $GF(2)$. Moreover, the $m$ orbits $O_m$ that expand the orbital $[O_m]$ are the $m$ roots of $f_m(x)$ in $GF(2^n)$. Thus, two internal operations $(+,*)$ are defined over ${\cal B}_n$ in such a way that $({\cal B}_n,+,*)$ be the splitting field for all I.P. $f_m(x)$ over $GF(2)$ , with $m\mid n$. Field Construction in ${\cal B}_n$ ================================== [**Choice of a Basis**]{}: If we consider $GF(2^n)$ a vectorial space over $GF(2)$, a [*primitive normal basis*]{}, $B_{\sigma n}$={$\sigma_0$, $\sigma_1$,..., $\sigma_{n-1}$}, can be found in $GF(2^n)$ [@lidl; @lenstra1]. $B_{\sigma n}$ to be [*normal*]{} means that $\sigma_i^2=\sigma_{(i+1)\bmod n}$. $B_{\sigma n}$ to be [*primitive*]{} means that $\sigma_0$ generate $GF(2^n)^*=GF(2^n)-\{0\}$. Therefore, if we denote by $(+,*)$ the two internal operations in $GF(2^n)$, each element $a\in GF(2^n)^*$ has a unique representation as [*addition*]{} ($+$) or as [*product*]{} ($*$) of the $B_{\sigma n}$ elements (tables \[tabla2\], \[tabla3\], \[tabla4\], \[tabla5\], \[tabla6\]): (i) : each element $a\in GF(2^n)$ can be expanded in a linear combination of $B_{n\sigma}$, $a=\sum_{i=0}^{n-1}\alpha_i\sigma_i$ with $\alpha_i\in GF(2)\equiv\{0,1\}$. The binary sequence $a_+=(\alpha_{n-1}\alpha_{n-2}\cdots\alpha_1\alpha_0)\in {\cal B}_n$ is the $+$-representation of $a$. (ii) : each element $a\in GF(2^n)^*$ is generated by $\sigma_0$, consequently $a=\sigma_0^k$ with $k\in\{1,2,\cdots,2^n-1\}$. If we take the binary expression of the exponent, $k=\sum_{i=0}^{n-1}\beta_i2^i$, the binary sequence $a_*=(\beta_{n-1}\beta_{n-2}\cdots\beta_1\beta_0)\in {\cal B}_n$ is the $*$-representation of $a$. It can also be reformulated as $a=\Pi_{i=0}^{n-1}\sigma_i^{\beta_i}$ where $\sigma_i^0=1$. It is obvious that $(a^2)_*=shift(a_*)$. (table \[tabla5\]): If $B_{\sigma 4}$={$\sigma_0$, $\sigma_1$, $\sigma_2$, $\sigma_3$} is a primitive normal basis, the following expansions are obtained for some elements of $GF(2^4)$: $$\begin{aligned} a = \sigma_0^3 = \sigma_0+\sigma_1+\sigma_3 & \rightarrow & a_+ = 1011, \\ a = \sigma_0^3 = \sigma_0*\sigma_1 & \rightarrow & a_* = 0011, \\ & & \\ b = \sigma_0^4 = \sigma_2 & \rightarrow & b_+ = 0010, \\ b = \sigma_0^4 = \sigma_2 & \rightarrow & b_* = 0010, \\ & & \\ c = \sigma_0^7 = \sigma_2+\sigma_3 & \rightarrow & c_+ = 1100, \\ c = \sigma_0^7 = \sigma_0*\sigma_1*\sigma_2 & \rightarrow & c_* = 0111. \end{aligned}$$ The particular elements, $a=\sigma_0^{2^i}=\sigma_i$ and $a=\sigma_0^{2^n-1}=\sigma_{n-1}*\sigma_{n-2}*\cdots*\sigma_1*\sigma_0= \sigma_{n-1}+\sigma_{n-2}+\cdots+\sigma_1+\sigma_0=1$, verify $a_+=a_*$. In this case, $(00\cdots 010\cdots 00)$ and $(11\cdots 11\cdots 11)$, respectively. [**Internal operations in ${\cal B}_n$**]{}: we explain now how to perform the two internal operations $(+,*)$ in ${\cal B}_n$. : the property $\sigma_i+\sigma_i=0$ (the characteristic of $GF(2^n)$ is $2$) imposes the algorithm for performing this operation. Given $a,b\in GF(2^n)$ the element $a+b$ is obtained by summing bit to bit the binary sequences of $a$ and $b$ ($\bmod 2$) in the $+$-representation. If we have $$\begin{aligned} a_+ & = & \alpha_{n-1}\alpha_{n-2}\cdots\alpha_1\alpha_0, \\ b_+ & = & \beta_{n-1}\beta_{n-2}\cdots\beta_1\beta_0, \\ (a+b)_+ & = & \gamma_{n-1}\gamma_{n-2}\cdots\gamma_1\gamma_0,\end{aligned}$$ the addition $+$ presents the form: $$\begin{aligned} (a+b)_+ & = & a_+ + b_+ \bmod 2, \\ \gamma_i & = & (\alpha_i + \beta_i) \bmod 2. \end{aligned}$$ $({\cal B}_n, +)$ is a representation of the abelian additive group defined over $GF(2^n)$, where the neutral element is $0\equiv (00\cdots 00)$ and the inverse element $\overline a$ of $a$ is the element itself, $a+a=0$, then $\overline{a}=a$. (table \[tabla5\]): $$\begin{aligned} (\sigma_0^3)_+ & \rightarrow & 1011, \\ (\sigma_0^4)_+ & \rightarrow & \underline{0010}, \\ (\sigma_0^3+\sigma_0^4)_+ & \rightarrow & 1001 =\sigma_0^{14}.\end{aligned}$$ : the property $\sigma_i^2=\sigma_{(i+1) \bmod n}$ gives the clue to perform this operation. Given $a,b\in GF(2^n)$ the element $a*b$ is obtained after performing the addition with carrier ($\bmod 2$) of the binary sequences of $a$ and $b$ in the $*$-representation. If we have $$\begin{aligned} a_* & = & \alpha_{n-1}\alpha_{n-2}\cdots\alpha_1\alpha_0, \\ b_* & = & \beta_{n-1}\beta_{n-2}\cdots\beta_1\beta_0, \\ (a+b)_* & = & \gamma_{n-1}\gamma_{n-2}\cdots\gamma_1\gamma_0, \end{aligned}$$ then the multiplication $*$ presents the form: $$\begin{aligned} (a*b)_* \equiv a_* * b_* & = & (a_* + b_*\; with\;\;carrier)\bmod 2,\;\; \\ \gamma_i & = & (\alpha_i + \beta_i + carrier_{i})\bmod 2.\end{aligned}$$ To start we suppose $carrier_0=0$. If $\alpha_{n-1}+\beta_{n-1}+carrier_{n-1}>1$ then the operation is restarted with $carrier_0=1$. $({\cal B}_n, *)$ is a representation of the abelian multiplicative group defined over $GF(2^n)^*$, where the neutral element is $1\equiv (11\cdots 11)$ and the inverse element $a^{-1}$ of $a$ is represented by the complementary binary sequence, $(a^{-1})_*=cp(a_*)$ given by $digit_i((a^{-1})_*)=(digit_i(a_*)+1\;\;\bmod 2$). This is a consequence of the fact that if $a=\sigma_0^k$ then $a^{-1}=\sigma_0^{(2^n-1)-k}$. The multiplication by zero is $(a*0)_*=a_**(00\cdots 00)=(00\cdots 00)\equiv 0$ by definition. It can also be seen that if $(a*b)_*=0$, then only one of the two sequences are null, $a=0$ or $b=0$. : $$\begin{aligned} carrier_0=0, & & \\ (\sigma_0^{11})_* & \rightarrow & 1011, \\ (\sigma_0^4)_* & \rightarrow & \underline{0010}, \\ (\sigma_0^{11}*\sigma_0^4)_*& \rightarrow & 1101 =\sigma_0^{13}.\end{aligned}$$ $$\begin{aligned} carrier_0=1, & & \\ (\sigma_0^{11})_* & \rightarrow & 1011, \\ (1)_* & \rightarrow & \underline{1111}, \\ (\sigma_0^{11}*1)_*& \rightarrow & 1011 =\sigma_0^{11}.\end{aligned}$$ $$\begin{aligned} a=\sigma_0^{11} & \rightarrow & a_* = 1011, \\ a^{-1}=\sigma_0^{4} & \rightarrow & (a^{-1})_*=0100.\end{aligned}$$ : the central question in order to build up a representation of the field $GF(2^n)$ over ${\cal B}_n$ is to find the relationship between $a_+$ and $a_*$. This construction can be performed if that relation is known for the elements $\sigma_i*\sigma_j$ because all the other elements can be reduced to products of $\sigma_i*\sigma_j$ type. Thus, if we write $$\begin{aligned} \sigma_i*\sigma_j & = & \sum_{l=0}^{n-1} e_{ijl}\sigma_l,\end{aligned}$$ there are $n^3$ constants $e_{ijl}$ ($=0,1$) defining the field $GF(2^n)$ completely. By construction $\sigma_i*\sigma_i=\sigma_{i+1}$, then $e_{iil}=\delta_{(i+1)l}$. By the commutative property, $e_{ijl}=e_{jil}$. Also, $(\sigma_i*\sigma_j)^2=\sigma_{i+1}*\sigma_{j+1}$, then $shift((\sigma_i*\sigma_j)_*)=(\sigma_{i+1}*\sigma_{j+1})_*$. If we impose this property for the $+$-representation, that is, $shift((\sigma_i*\sigma_j)_+)=(\sigma_{i+1}*\sigma_{j+1})_+$, then $e_{(i+1)(j+1)(l+1)}=e_{ijl}$. Thus, only $(n-1)/2$ products $\sigma_i*\sigma_j$ when $n$ is odd, and $n/2$ products $\sigma_i*\sigma_j$ when $n$ is even, must be known in order to determine the field completely. That is, a total number of $n^2/2$ or $(n^2-n)/2$ constants $e_{ijl}$, depending on the parity of $n$, defines the field. These constants cannot be freely chosen as the distributive property must be verified: $$\begin{aligned} a*(b+c) & = & a*b + a*c, \\ \left( a_* * (b_+ + c_+)_*\right)_+ & = & (a_* * b_*)_+ + (a_* * c_*)_+.\end{aligned}$$ All the constants $e_{ijl}$ stay fixed by these conditions in some cases and the natural construction $({\cal B}_n,+,*)$ is unique. In general, this is not the case and there exist different constructions of $GF(2^n)$ in ${\cal B}_n$. Construction of the Irreducible Polynomials =========================================== Let us recall that $GF(2^n)$ is the splitting field of all irreducible polynomials $f_m(x)$ over $GF(2)$ of degree $m\leq n$ and $m\mid n$. If $a\in GF(2^n)$ is a root of $f_m(x)$, the $m$ roots of this irreducible polynomial are $\{a,a^2,a^{2^2},\cdots,a^{2^{(m-1)}}\}$. Therefore, the $m$ orbits that integrate an orbital $[O_m]\in {\cal B}_n$ are the $m$ roots of the irreducible polynomial $f_m(x)$ in the representation $({\cal B}_n,+,*)$ of $GF(2^n)$, as explained in the last section. For instance, in the case $n=4$ (table \[tabla5\]), the elements $\{\sigma_0\sigma_1,\sigma_1\sigma_2,\sigma_2\sigma_3,\sigma_3\sigma_0\}$ are the roots of the irreducible polynomial $x^4+x^3+x^2+x+1$. The objetive of this section is to establish a method to find the irreducible polynomials over $GF(2)$ working out this property. [**Trace of an orbital**]{}: The trace of an orbital $[O_m]$, $Tr(O_m)$, is defined as the trace of one of its elements: $\{a,a^2,a^{2^2},\cdots,a^{2^{(m-1)}}\}$ with $a\in GF(2^n)$, that is, $$Tr(O_m) = a+a^2+a^{2^2}+\cdots +a^{2^{(m-1)}}.$$ The trace is an element of $GF(2)$, in this case $0$ or $1$. The number of traces $N_T^{(n)}$ in $GF(2^n)$ is equal to the number of orbitals $[O_m]$ in ${\cal B}_n$, $$N_T^{(n)}=\sum_m N_m \;\;\; with \;\; m\mid n. \label{eqNT}$$ Let us call $N_{T1}^{(n)}$ the number of traces whose value is $1$ and $N_{T0}^{(n)}$ the number of those whose value is $0$. Hence, $N_T^{(n)} = N_{T0}^{(n)} + N_{T1}^{(n)}$.The important point for our proposal is that the coefficients of an I.P. over $GF(2)$ can be obtained by adding ($\bmod 2$) the traces of different orbitals of the field $GF(2^n)$. Therefore the I.P. can be identified if the traces are previously calculated. (table \[tabla5\]): (i) the I.P. associated to the orbital $\{\sigma_0\sigma_1,\sigma_1\sigma_2,\sigma_2\sigma_3,\sigma_3\sigma_0\}$ is a polynomial of degree $4$, $x^4 + d x^3 + e x^2 + f x + 1$. Its coefficients are functions of the traces: $$\begin{aligned} d = & \sigma_0\sigma_1+\sigma_1\sigma_2+\sigma_2\sigma_3+\sigma_3\sigma_0 & = Tr(\sigma_0\sigma_1), \\ e = & \sigma_0\sigma_1+\sigma_1\sigma_2+\sigma_2\sigma_3+\sigma_3\sigma_0 +\sigma_0\sigma_1\sigma_2\sigma_3+\sigma_0\sigma_1\sigma_2\sigma_3 & = Tr(\sigma_0\sigma_1), \\ f = & \sigma_0\sigma_1+\sigma_1\sigma_2+\sigma_2\sigma_3+\sigma_3\sigma_0 & = Tr(\sigma_0\sigma_1).\end{aligned}$$ (ii) the I.P. associated to the orbital $\{\sigma_0, \sigma_1, \sigma_2, \sigma_3\}$ is also of degree $4$, $x^4 + a x^3 + b x^2 + c x + 1$, where: $$\begin{aligned} a = & \sigma_0+\sigma_1+\sigma_2+\sigma_3 & = Tr(\sigma_0), \\ b = & \sigma_0\sigma_1+\sigma_1\sigma_2+\sigma_2\sigma_3+\sigma_3\sigma_0+ \sigma_0\sigma_2+\sigma_1\sigma_3 & = Tr(\sigma_0\sigma_1) + Tr(\sigma_0\sigma_2), \\ f = & \sigma_0\sigma_1\sigma_2+\sigma_1\sigma_2\sigma_3+ \sigma_2\sigma_3\sigma_0+\sigma_3\sigma_0\sigma_1 & = Tr(\sigma_0\sigma_1\sigma_2).\end{aligned}$$ [**Newton’s Formulas**]{}: the [*fundamental theorem on symmetrical polynomials*]{} can guide us in finding the different traces in the field $GF(2^n)$. This theorem can be expressed as follows [@edwards]: Let $r_1$, $r_2$, $\cdots$, $r_n$ be different variables and let us call $\theta_1$, $\cdots$, $\theta_n$ the following symmetrical polynomials in those variables: $$\begin{aligned} \theta_1 & = & r_1 + r_2 + \cdots + r_n, \\ \theta_2 & = & r_1 r_2 + r_1r_3 + \cdots + r_{n-1}r_n, \\ \theta_3 & = & r_1 r_2r_3 + r_1r_2r_4 + \cdots + r_{n-2}r_{n-1}r_n, \\ & \vdots & \\ \theta_n & = & r_1 r_2 r_3 \cdots r_n .\end{aligned}$$ Then every symmetrical polynomial in the variables {$r_1$, $r_2$, $\cdots$, $r_n$} and with coefficients over a field $F$ can be expressed as a polynomial in the new variables $\{\theta_1,\theta_2,\cdots,\theta_n\}$ with coefficients in $F$. In the particular case of the following symmetrical polynomials $s_k$: $$s_k = r_1^k + r_2^k +\cdots + r_n^k \;\;\;\; k = 1,2,3,\cdots$$ the expansion in the variables $\theta_k$ is given by the [*Newton’s formulas*]{} [@newton]: $$\begin{aligned} s_1 - \theta_1 & = & 0, \nonumber \\ s_2 - s_1\theta_1 + 2\theta_2 & = & 0, \nonumber \\ s_3 - s_2\theta_1 + s_1\theta_2 + 3\theta_3 & = & 0, \label{eqsk} \\ \vdots & & \nonumber \\ s_k - s_{k-1}\theta_1 + s_{k-2}\theta_2 - \cdots + (-1)^{k-1}s_1\theta_{k-1} + (-1)^k(k)\theta_k & = & 0, \nonumber \end{aligned}$$ where $\theta_j=0$ for $j>n$ and the symbol $(k)$ is interpreted as $(k)=1+1+\stackrel{k}{\cdots}+1$. In our problem, for instance, if we take $r_i=\sigma_{i-1}$ the expressions of $s_k$ and $\theta_k$ are functions of the $GF(2^n)$ traces. Then the Newton’s formulas give us a system of equations for the different traces of $GF(2^n)$. By solving this system it is possible build up the I.P. of degree $n$ over $GF(2)$. [**Finding the Traces**]{}: If the elements $\{\sigma_0,\sigma_1,\cdots,\sigma_{n-1}\} \in GF(2^n)$ are taken as the variables {$r_1$, $r_2$, $\cdots$, $r_n$} in the equations (\[eqsk\]), a system of equations for the different traces is obtained. The total number of unknowns (traces) is equal to the total number of orbitals, $N_T^{(n)}$, in $GF(2^n)$ (eq. \[eqNT\]). By construction: $$\begin{aligned} Tr(0) & = & 0, \\ Tr(\sigma_0) & = & 1, \\ Tr(\sigma_0\sigma_1\cdots\sigma_{n-1}) & = & 1.\end{aligned}$$ For large $n$ the number of orbitals is of the same order as $$N_T^{(n)}\sim \frac{2^n}{n}.$$ Noting that $r_i^{2^n}=\sigma_i^{2^n}=\sigma_i=r_i$ we will have $2^n$ equations. Therefore, a system of $2^n$ equations with $2^n/n$ unknowns (many equations are interdependent) should be solved. Only those solutions verifying that the number of traces with value $1$ is equal to $N_{T1}^{(n)}$ are accepted. If one of these solutions is chosen, then the I.P. are directly calculated. [**Note**]{}: Let us recall a property that allows us to generate the I.P. by pairs. Two polynomials, $f(x)$ and $g(x)$, are reciprocal if they can be expressed in the form: $$\begin{aligned} f(x) & = & a_nx^n + a_{n-1} x^{n-1} + \cdots + a_1 x + a_0, \\ g(x) & = & a_0x^n + a_1 x^{n-1} + \cdots + a_{n-1} x + a_n, \end{aligned}$$ or equivalently, $g(x)=x^nf(x^{-1})$. This means that the roots of $g(x)$ are the inverse roots of $f(x)$. Consequently, if $f_m(x)$ is an I.P. of degree $m$, the reciprocal polynomial $g_m(x)$ is also an I.P. of degree $m$. This property can be used as another useful tool to facilitate the finding of all irreducible polynomials. It is necessary to remark at this point that both polynomials can be the same, that is, $f_m(x)=g_m(x)$. General Methodology =================== Summarizing, the steps to follow for building up a natural representation of the field $GF(2^n)$ in ${\cal B}_n$ and for finding all I.P. of degree $n$ over $GF(2)$ are the following: [**(1)**]{} All the independent $n$-binary sequences (orbitals $[O_m]$ with $m\mid n$) in ${\cal B}_n$ are found (see appendix). Its total number, $N_{T}^{(n)}$, is the total number of I.P. over $GF(2)$. The splitting field is $GF(2^n)$. Let us observe that each orbital $[O_m]$ has associated an I.P. $f_m(x)$ of degree $m$. The roots of $f_m(x)$ in $({\cal B}_n,+,*)$ are the $m$ roots given by the $m$ orbits of the orbital $[O_m]$. It also is possible to evaluate $N_{T1}^{(n)}$ by calculating directly the trace for each orbit $O_m$, $$Tr(O_m) = O_m + shift(O_m) + shift^2(O_m) + \cdots + shift^{m-1}(O_m),$$ where $shift(O_m)$ is the left shift of the orbit $O_m$. If $\#_1(O_m)$ represents the number of $1$ in the orbit $O_m$ multiplied by $m/n$ then it can be deduced that $$\begin{aligned} Tr(O_m) = 1 & if & \#_1(O_m)\;\; is\;\; odd, \\ Tr(O_m) = 0 & if & \#_1(O_m)\;\; is\;\; even.\end{aligned}$$ For instance, $$\begin{aligned} \#_1(0011)=2\cdot\frac{4}{4}=2 & \Rightarrow & Tr(0011) = 0, \\ \#_1(0111)=3\cdot\frac{4}{4}=3 & \Rightarrow & Tr(0111) = 1, \\ \#_1(0101)=2\cdot\frac{2}{4}=1 & \Rightarrow & Tr(0101) = 1, \\ \#_1(1111)=4\cdot\frac{1}{4}=1 & \Rightarrow & Tr(1111) = 1. \end{aligned}$$ [**(2)**]{} To calculate the trace associated to each element in $GF(2^n)$, the $2^n$ non-independent equations deriving from the Newton’s formulas are established. There are $N_{T}^{(n)}\sim 2^n/n$ unknowns. [**(3)**]{} A solution of the former equation system is worked out. This solution must be compatible with the constraints: (i) $Tr(0)=0$, $Tr(\sigma_0)=1$ and $Tr(\sigma_0\sigma_1\cdots\sigma_{n-1})=1$; (ii) the number of traces with value $1$ must be equal to $N_{T1}^{(n)}$ (calculated in point (1)). The different possible solutions are related with the different possible constructions of $GF(2^n)$ in ${\cal B}_n$. [**(4)**]{} The irreducible polynomials are explicitely found in a deterministic way from a particular solution obtained at point (3). [**(5)**]{} The construction of $({\cal B}_n, +, *)$ is performed from the I.P. calculated in (4). This is a natural representation of $GF(2^n)$. [**(6)**]{} : As it can be seen in the tables below, an only representation of $GF(2^n)$ in ${\cal B}_n$ is possible when $\{n=1,2,3,4\}$. Examples ======== In this section the process formerly explained is applied for the cases $n=4,5$. [**n=4**]{}: Let us suppose $GF(2^4)$ expanded by the primitive normal basis $B_{\sigma 4}=\{\sigma_0,\sigma_1,\sigma_2,\sigma_3\}$. Then $\sigma_{i}^2=\sigma_{(i+1)\bmod 2}$. The field’s elements $\{0,\sigma_0,\sigma_0^2,\sigma_0^3,\cdots,\sigma_0^{14}, \sigma_0^{15}=1\}$ are generated by $\sigma_0$. If an element $a\in GF(2^4)$ verifies $a^{2^m}=a$, the $m$ powers of $a$, $\{a,a^2,a^{2^2},\cdots,a^{2^{m-1}}\}$, are the $m$ roots of an I.P. of degree $m$ over $GF(2)$. Obviously, $m$ is a divisor of $4$. The different representative elements of this kind in $GF(2^4)$ are: $a\in GF(2^4)$ $m$ ------------------------------------ ----- $0 $ $1$ $\sigma_0$ $4$ $\sigma_0\sigma_1$ $4$ $\sigma_0\sigma_2$ $2$ $\sigma_0\sigma_1\sigma_2$ $4$ $\sigma_0\sigma_1\sigma_2\sigma_3$ $1$ The different orbitals $[O_m]$ of period $m$, with $m\mid 4$, in ${\cal B}_4$ are given by the representative orbits: $[O_m]$ $Tr(O_m)$ m --------- ----------- ----- $0000$ 0 $1$ $0001$ 1 $4$ $0011$ 0 $4$ $0101$ 1 $2$ $0111$ 1 $4$ $1111$ 1 $1$ Thus, $N_T^{(4)}=N_1+N_2+N_3=6$ and $N_{T1}^{(4)}=4$ (see Table 1). It can also be observed that the orbit $\{\sigma_0\sigma_2,\sigma_1\sigma_3\}$ is of period $2$, then its binary $+$-representation is $\{0101,1010\}$ and by direct calculation $Tr(\sigma_0\sigma_2)=1$. If $\{\sigma_0,\sigma_1,\sigma_2,\sigma_3\}$ are taken as the roots of an I.P. of degree 4, the equations derived from Newton’s formulas (where $\theta_k=0$ for $k>4$) for the different traces can be established. The variables $\theta_k$ and $s_k$ are expressed as follows: $$\begin{aligned} \theta_1 & = & \sigma_0 + \sigma_1 + \sigma_2 + \sigma_3 \\ & = & Tr(\sigma_0) = 1, \\ \theta_2 & = & \sigma_0\sigma_1 + \sigma_1\sigma_2 + \sigma_2\sigma_3 + \sigma_3\sigma_0 + \sigma_0\sigma_2 + \sigma_1\sigma_3 \\ & = & Tr(\sigma_0\sigma_1) + Tr(\sigma_0\sigma_2), \\ \theta_3 & = & \sigma_0\sigma_1\sigma_2 + \sigma_1\sigma_2\sigma_3 + \sigma_2\sigma_3\sigma_0 + \sigma_3\sigma_0\sigma_1 \\ & = & Tr(\sigma_0\sigma_1\sigma_2), \\ \theta_4 & = & \sigma_0\sigma_1\sigma_2\sigma_3 \\ & = & Tr(\sigma_0\sigma_1\sigma_2\sigma_3) = 1,\\ s_k & = & \sigma_0^k + \sigma_1^k + \sigma_2^k + \sigma_3^k \\ & = & \sum_i Tr(\sigma^i) \;\; with\;\; k=1,2,\cdots,15.\end{aligned}$$ The independent Newton’s formulas from the system (\[eqsk\]) are: $$\begin{aligned} Tr(\sigma_0\sigma_2) & = & 1, \\ Tr(\sigma_0\sigma_2) + Tr(\sigma_0\sigma_1\sigma_2) & = & 1, \\ Tr(\sigma_0\sigma_1) & = & 1. \end{aligned}$$ The solution is unique: $$\begin{aligned} Tr(\sigma_0) & = & 1, \\ Tr(\sigma_0\sigma_1) & = & 1, \\ Tr(\sigma_0\sigma_2) & = & 1,\\ Tr(\sigma_0\sigma_1\sigma_2) & = & 0, \\ Tr(\sigma_0\sigma_1\sigma_2\sigma_3) & = & 1. \\\end{aligned}$$ Finally, the I.P. are: $$\begin{aligned} \; [O_4]\equiv\{\sigma_0\} \rightarrow & x^4 + a x^3 + b x^2 + c x + 1 & \rightarrow x^4 + x^3 + 1, \\ \; [O_4]\equiv\{\sigma_0^{-1}=\sigma_0\sigma_1\sigma_2\} \rightarrow & x^4 + c x^3 + b x^2 + a x + 1 & \rightarrow x^4 + x + 1, \\ & a = Tr(\sigma_0) = 1, & \\ & b = Tr(\sigma_0\sigma_1) + Tr(\sigma_0\sigma_2) = 0, & \\ & c = Tr(\sigma_0\sigma_1\sigma_2)=0, & \\ \\ \; [O_4]\equiv\{\sigma_0\sigma_1\} \rightarrow & x^4 + d x^3 + e x^2 + f x + 1 & \rightarrow x^4 + x^3 + x^2 + x + 1, \\ & d = Tr(\sigma_0\sigma_1) = 1, & \\ & e = Tr(\sigma_0\sigma_1) = 1, & \\ & f = Tr(\sigma_0\sigma_1) = 1, & \\ \\ \; [O_2]\equiv\{\sigma_0\sigma_2\} \rightarrow & x^2 + g x + 1 & \rightarrow x^2 + x + 1, \\ & g = Tr(\sigma_0\sigma_2) = 1, & \\ \\ \; [O_1]\equiv\{\sigma_0\sigma_1\sigma_2\sigma_3\} \rightarrow & x + 1 & \rightarrow x + 1, \\ \\ \; [O_1]\equiv\{0\} \rightarrow & x & \rightarrow x, \end{aligned}$$ where $\{\sigma_i\cdots\sigma_k\}$ is representing all the $GF(2^4)$ elements that correspond to the orbital \[$O_m$\]. The step to find the $+$-representation of every $a\in GF(2^4)$ is the following. Let us take $\sigma_0$ with associated I.P.: $x^4 + x^3 + 1$, then: $$\begin{aligned} \sigma_0^4 + \sigma_0^3 + 1 & = & 0, \\ \sigma_0^4 = \sigma_2 & \rightarrow & (\sigma_0^3)_+ = 1 + (\sigma_2)_+ = 1011.\end{aligned}$$ Taking the shifts of $1011$, we find the $+$-representation for all the elements of the orbital $\{\sigma_0^3\}\equiv\{\sigma_0\sigma_1\}$. The distributive property determines the $+$-representation of the orbit $\{\sigma_0\sigma_1\sigma_2\}$, with the final result shown in table \[tabla5\]. Let us observe that this construction of $GF(2^4)$ is unique. [**n=5**]{}: We repeat the same process as one of the case $n=4$. Let us suppose $GF(2^5)$ generated by the primitive normal basis $B_{\sigma 5}=\{\sigma_0,\sigma_1,\sigma_2,\sigma_3,\sigma_4\}$. Hence $\sigma_{i}^2=\sigma_{(i+1)\bmod 2}$. The field’s elements $\{0,\sigma_0,\sigma_0^2,\sigma_0^3,\cdots,\sigma_0^{30},\sigma_0^{31}=1\}$ are generated by $\sigma_0$. If an element $a\in GF(2^5)$ verifies $a^{2^m}=a$, then $m\mid 5$ and the $m$ powers of $a$, $\{a,a^2,a^{2^2},\cdots,a^{2^{m-1}}\}$, are the $m$ roots of an I.P. of degree $m$ over $GF(2)$. The representative different subsets of this kind in $GF(2^5)$ are: $a\in GF(2^5)$ $m$ -------------------------------------------- ----- $0 $ $1$ $\sigma_0$ $5$ $\sigma_0\sigma_1$ $5$ $\sigma_0\sigma_2$ $5$ $\sigma_0\sigma_1\sigma_2$ $5$ $\sigma_0\sigma_1\sigma_3$ $5$ $\sigma_0\sigma_1\sigma_2\sigma_3$ $5$ $\sigma_0\sigma_1\sigma_2\sigma_3\sigma_4$ $1$ The different orbitals $[O_m]$ of period $m$, with $m\mid 5$, in ${\cal B}_5$ are given by the representative orbits: $[O_m]$ $Tr(O_m)$ m --------- ----------- --- $00000$ 0 1 $00001$ 1 5 $00011$ 0 5 $00101$ 0 5 $00111$ 1 5 $01011$ 1 5 $01111$ 0 5 $11111$ 1 1 Thus, $N_T^{(5)}=N_1+N_5=8$ and $N_{T1}^{(5)}=4$.If $\{\sigma_0,\sigma_1,\sigma_2,\sigma_3,\sigma_4\}$ are taken as the roots of an I.P. of degree 5 the equations derived from Newton’s formulas (where $\theta_k=0$ for $k>5$) for the different traces can be established. The variables $\theta_k$ and $s_k$ are expressed as follows: $$\begin{aligned} \theta_1 & = & \sigma_0 + \sigma_1 + \sigma_2 + \sigma_3 \\ & = & Tr(\sigma_0) = 1,\\ \theta_2 & = & \sigma_0\sigma_1 + \sigma_1\sigma_2 + \sigma_2\sigma_3 + \sigma_3\sigma_4 +\sigma_4\sigma_0 \\ & & +\sigma_0\sigma_2 + \sigma_1\sigma_3 + \sigma_2\sigma_4 + \sigma_3\sigma_0 + \sigma_4\sigma_1 \\ & = & Tr(\sigma_0\sigma_1) + Tr(\sigma_0\sigma_2), \\ \theta_3 & = & \sigma_0\sigma_1\sigma_2 + \sigma_1\sigma_2\sigma_3 + \sigma_2\sigma_3\sigma_4 + \sigma_3\sigma_4\sigma_0 + \sigma_4\sigma_0\sigma_1 \\ & & +\sigma_0\sigma_1\sigma_3 + \sigma_1\sigma_2\sigma_4 + \sigma_2\sigma_3\sigma_0 + \sigma_3\sigma_4\sigma_1 + \sigma_4\sigma_0\sigma_2 \\ & = & Tr(\sigma_0\sigma_1\sigma_2) + Tr(\sigma_0\sigma_1\sigma_3), \\ \theta_4 & = & \sigma_0\sigma_1\sigma_2\sigma_3 +\sigma_1\sigma_2\sigma_3\sigma_4 +\sigma_2\sigma_3\sigma_4\sigma_0 \\ & & +\sigma_3\sigma_4\sigma_0\sigma_1+\sigma_4\sigma_0\sigma_1\sigma_2 \\ & = & Tr(\sigma_0\sigma_1\sigma_2\sigma_3), \\ \theta_5 & = & Tr(\sigma_0\sigma_1\sigma_2\sigma_3\sigma_4) = 1,\\ s_k & = & \sigma_0^k + \sigma_1^k + \sigma_2^k + \sigma_3^k + \sigma_4^k \\ & = & \sum_i Tr(\sigma^i) \;\; with\;\; k=1,2,\cdots,31.\end{aligned}$$ The independent Newton’s formulas from the system (\[eqsk\]) are: $$\begin{aligned} Tr(\sigma_0\sigma_2) + Tr(\sigma_0\sigma_1\sigma_2) + Tr(\sigma_0\sigma_1\sigma_3) & = & 1, \\ Tr(\sigma_0\sigma_1\sigma_3) + Tr(\sigma_0\sigma_1\sigma_2\sigma_3) & = & 0, \\ Tr(\sigma_0\sigma_1) + Tr(\sigma_0\sigma_1\sigma_2) & = & 0, \\ Tr(\sigma_0\sigma_2)(1+Tr(\sigma_0\sigma_1\sigma_3)) & = & 0, \\ Tr(\sigma_0\sigma_2)(1+Tr(\sigma_0\sigma_1)) & = & 0. \\\end{aligned}$$ One solution compatible with the constraints is: $$\begin{aligned} Tr(\sigma_0) & = & 1, \\ Tr(\sigma_0\sigma_1) & = & 0, \\ Tr(\sigma_0\sigma_2) & = & 0, \\ Tr(\sigma_0\sigma_1\sigma_2) & = & 0, \\ Tr(\sigma_0\sigma_1\sigma_3) & = & 1, \\ Tr(\sigma_0\sigma_1\sigma_2\sigma_3) & = & 1, \\ Tr(\sigma_0\sigma_1\sigma_2\sigma_3\sigma_4) & = & 1. \\\end{aligned}$$ Finally, the I.P. are calculated from this solution: $$\begin{aligned} \; [O_5]\equiv\{\sigma_0\} \rightarrow & x^5 + a x^4 + b x^3 + c x^2 + d x + 1 & \rightarrow x^5 + x^4 + x^2 \\ & & \;\;\; + x + 1, \\ \; [O_5]\equiv\{\sigma_0^{-1}\equiv \sigma_0\sigma_1\sigma_2\sigma_3\} \rightarrow & x^5 + d x^4 + c x^3 + b x^2 + a x + 1 & \rightarrow x^5 + x^4 + x^3 \\ & & \;\;\; + x + 1, \\ & a = Tr(\sigma_0) = 1, & \\ & b = Tr(\sigma_0\sigma_1) + Tr(\sigma_0\sigma_2) = 0, & \\ & c = Tr(\sigma_0\sigma_1\sigma_2) + Tr(\sigma_0\sigma_1\sigma_3) = 1, & \\ & d = Tr(\sigma_0\sigma_1\sigma_2\sigma_3) = 1, & \\ \\ \; [O_5]\equiv\{\sigma_0\sigma_1\} \rightarrow & x^5 + e x^4 + f x^3 + g x^2 + h x + 1 & \rightarrow x^5 + x^3 + 1, \\ \; [O_5]\equiv\{(\sigma_0\sigma_1)^{-1}\equiv \sigma_0\sigma_1\sigma_2\} \rightarrow & x^5 + h x^4 + g x^3 + f x^2 + e x + 1 & \rightarrow x^5 + x^2 + 1, \\ & e = Tr(\sigma_0\sigma_1) = 0, & \\ & f = Tr(\sigma_0\sigma_2) + Tr(\sigma_0\sigma_1\sigma_2\sigma_3) = 1, & \\ & g = Tr(\sigma_0\sigma_1\sigma_3) + Tr(\sigma_0) = 1, & \\ & h = Tr(\sigma_0\sigma_1\sigma_2) = 0, & \\ \\ \; [O_5]\equiv\{\sigma_0\sigma_2\} \rightarrow & x^5 + j x^4 + k x^3 + l x^2 + m x + 1 & \rightarrow x^5 + x^3 + x^2 \\ & & \;\;\; + x + 1, \\ \; [O_5]\equiv\{(\sigma_0\sigma_2)^{-1}\equiv \sigma_0\sigma_1\sigma_3\} \rightarrow & x^5 + m x^4 + l x^3 + k x^2 + j x + 1 & \rightarrow x^5 + x^4 + x^3 \\ & & \;\;\; + x^2 + 1, \\ & j = Tr(\sigma_0\sigma_2) = 0, & \\ & k = Tr(\sigma_0\sigma_1\sigma_2\sigma_3) + Tr(\sigma_0\sigma_1\sigma_2) = 1, & \\ & l = Tr(\sigma_0) + Tr(\sigma_0\sigma_1) = 1, & \\ & m = Tr(\sigma_0\sigma_1\sigma_3) = 1, & \\ \\ \; [O_1]\equiv\{\sigma_0\sigma_1\sigma_2\sigma_3\sigma_4\} \rightarrow & x + 1 & \rightarrow x + 1, \\ \\ \; [O_1]\equiv\{0\} \rightarrow & x & \rightarrow x, \end{aligned}$$ where $\{\sigma_i\cdots\sigma_k\}$ is representing all the $GF(2^5)$ elements that correspond to the orbital \[$O_m$\]. The steps to find the $+$-representation of every $a\in GF(2^5)$ are the following. Let us take $\sigma_0$ with associated I.P.: $x^5 + x^4 + x^2 + x + 1$, then: $$\begin{aligned} \sigma_0^5 + \sigma_0^4 + \sigma_0^2 + \sigma_0 + 1 & = & 0, \\ \sigma_0^2 = \sigma_1\; ;\;\sigma_0^4 = \sigma_2 & \rightarrow & (\sigma_0^5)_+ = 1 + (\sigma_0)_+ +(\sigma_1)_+ +(\sigma_2)_+ = 11000.\end{aligned}$$ The shifts of $11000$ are the $+$-representation for all the elements of the orbital $\{\sigma_0^5\}\equiv\{\sigma_0\sigma_2\}$. If we now take $\sigma_0^3$ with associated I.P.: $x^5 + x^3 + 1$, then: $$\begin{aligned} (\sigma_0^3)^5 + (\sigma_0^3)^3 + 1 & = & 0, \\ (\sigma_0^9)_+ = 00110 & \rightarrow & (\sigma_0^{15})_+ = 1 + (\sigma_0^9)_+ = 11001.\end{aligned}$$ The shifts of $11001$ are the $+$-representation for all the elements of the orbital $\{\sigma_0^{15}\}\equiv\{\sigma_0\sigma_1\sigma_2\sigma_3\}$. Proceeding now with $\sigma_0^7$ that has an associated I.P.: $x^5 + x^2 + 1$, then: $$\begin{aligned} (\sigma_0^7)^5 + (\sigma_0^7)^2 + 1 & = & 0, \\ \sigma_0^{35}= \sigma_0^4 & \rightarrow & (\sigma_0^{14})_+ = 1 + (\sigma_0^4)_+ = 11011.\end{aligned}$$ The shifts of $11011$ are the $+$-representation for all the elements of the orbital $\{\sigma_0^{7}\}\equiv\{\sigma_0\sigma_1\sigma_2\}$. Taking $\sigma_0^{15}$ with associated I.P.: $x^5 + x^4 + x^3 + x + 1$, then: $$\begin{aligned} (\sigma_0^{15})^5 + (\sigma_0^{15})^4 + (\sigma_0^{15})^3 + \sigma_0^{15} + 1 & = & 0, \\ \sigma_0^{60}=00111\; ;\;\sigma_0^{45}=11011\; & \rightarrow & (\sigma_0^{75})_+ = (\sigma_0^{13})_+ = \\ & & 1 + (\sigma_0^{15})_+ + (\sigma_0^{45})_+ + (\sigma_0^{60})_+ = 11010.\end{aligned}$$ The shifts of $11010$ are the $+$-representation for all the elements of the orbital $\{\sigma_0^{11}\}\equiv\{\sigma_0\sigma_1\sigma_3\}$. Finally, the $+$-representation for the elements of the orbital $\{\sigma_0^{3}\}\equiv\{\sigma_0\sigma_1\}$ are the shifts of $01001$. Other solutions for the Newton’s formulas can allow different natural representations of the field $GF(2^5)$ in ${\cal B}_n$. Table \[tabla6\] is one of them. Conclusion ========== A possible method to build up the irreducible polynomials of degree $n$ over the field $GF(2)$ has been exposed in this paper. The strategy exploits the properties and the connection between the splitting field $GF(2^n)$ and the space of $n$-binary sequences ${\cal B}_n$. The choice of a primitive normal basis allows us to perform a natural construction of $GF(2^n)$ in ${\cal B}_n$: (i) Two natural operations $(+,*)$ are defined on ${\cal B}_n$ in order to obtain a representation of the splitting field $GF(2^n)$; (ii) A system of equations for the different traces of the field is derived from the fundamental theorem on symmetrical polynomials; (iii) Different constraints, that have been established, limit the number of possible solutions of these equations; (iv) Any of these solutions gives us the irreducible polynomials of degree $m$, with $m\mid n$; (v) The field $({\cal B}_n,+,*)$ is finally built up with the help of the irreducible polynomials. We want to remark the direct relationship that seems to exist between the irreducible polynomials over $GF(2)$ and the periodic binary orbits, in such a way that this is a one-to-one application. Following this ideas and under a similar scheme, this method could be applied to find the irreducible polynomials of degree $n$ over a general finite field $GF(q)$ by using the connection between the properties of the splitting field $GF(q^n)$ and the space of $n$-q-ary sequences. [**Acknowledgements:**]{} This work was performed during a postdoctoral stage at the Ecole Normale Supérieure of Paris in 1995. I thank J-M. Couveignes (Paris) and J. Cabeza (Zaragoza) for useful discussions and references on this subject of finite fields, and the Gouvernement Français for a research grant.: This manuscript should be understood as the approach of a ’newbie’, plenty of questions, to a subject very far from its common working topics. [**Appendix: Calculation of $N_{T}^{(n)}$ and $N_{T1}^{(n)}$**]{} An algorithm to calculate $N_{T}^{(n)}$ and $N_{T1}^{(n)}$ is proposed in this appendix. The process to find the different orbitals in ${\cal B}_n$ involves the following steps: \(1) The decimal value of the different $n$-digit orbitals is stored in an indexed integer variable $O(i)$, then $0<i<N_{T}^{(n)}$. The initial conditions are: $O(0)=0$, $O(1)=1$. Another integer variable $O_p$ is defined and initialized as $O_p=1$. \(2) The operational loop to find $O(k+1)$ when all orbitals $O(m)$ are known for $m\leq k$ is the following: $\;\hspace{3cm} (**)\; O_p = O_p+1$ $\hspace{4cm} do\;\; i=0,n-1$ $\hspace{4.5cm} \overline{O_p} = O_p\cdot 2^i \bmod (2^n-1)$ $\;\hspace{4.5cm} do\;\; \;\;j=1,k$ $\;\hspace{5cm} if\;(\overline{O_p}.eq.O(j))\;\; goto\; (**)$ $\;\hspace{3cm} \hspace{1.5cm} end\; do$ $\;\hspace{4cm} end\; do$ $\;\hspace{4cm} O(k+1) = O_p$ $\;\hspace{4cm} k = k+1 $ $\;\hspace{4cm} goto\; (**)$ \(3) This is repeated for all sequences of ${\cal B}_n$ while $O_p\leq (2^n-1)$. The maximum index $k$ from the step (2) is $k_{max} = N_T^{(n)}-1$. \(4) The period ($m$) of each orbital $O(i)$ is calculated: $$\begin{aligned} per(O(i)) = m & \Leftrightarrow & m\; is\;the\; smallest\; integer\;\;1\leq m\leq n-1 \\ & & \rightarrow O(i) = O(i)\cdot 2^m \bmod (2^n-1) \end{aligned}$$ \(5) Each orbital is converted to its binary expression and its trace is calculated: $$\begin{aligned} O_b(i) & = & binary(O(i)) \\ Tr(O(i))\equiv Tr(O_b(i)) & = & O_b(i) + shift(O_b(i)) + shift^2(O_b(i))+\cdots + shift^m(O_b(i)) \end{aligned}$$ \(6) The number of traces equal to $1$ when $i$ ranges from $1$ to $N_{T}^{(n)}$ is $N_{T1}^{(n)}$. [16]{} G.H. Hardy and E.A. Wright, [*An Introduction to the Theory of Numbers*]{}, Oxford (1968). H. Riesel, [*Prime Numbers and Computer Methods for Factorization*]{}, Birkhauser, Boston (1985). E.R. Berlekamp, [*Algebraic Coding Theory*]{}, McGraw-Hill, New York (1968). I.S. Reed and T.K. Truong, [*The use of finite fields to compute convolutions*]{}, IEEE Trans. Information Theory, [**IT-21**]{} 208-213 (1975). R.W. Marsch, [*Table of irreducible polynomials over $GF(2)$ through degree $19$*]{}, Office of Thecnical Services, Commerce Dept., Washintong DC, Oct. 24 (1957). S.W. Golomb, [*Obtaining specified irreducible polynomials over finite fields*]{}, SIAM J. Alg. Dis. Meth. [**1**]{} 411-418 (1980). V. Shoup, [*New algorithms for finding irreducible polynomials over finite fields*]{}, Math. Comp. [**54**]{} 435-447 (1990). R. López-Ruiz, [*A binary approach to the Lorenz system*]{}, Chaos, Solitons and Fractals [**8**]{}, 1-6 (1997). R.L. Miller, [*Necklages, symmetries and self-reciprocal polynomials*]{}, Discrete Math., [**22**]{} 25-33 (1978). R. Lidl and H. Niederreiter, [*Finite Fields*]{}, Addison-Wesley, London (1983). H.W. Lenstra, Jr. and R.J. Schoof, [*Primitive normal basis for finite fields*]{}, Math. Comp., [**48**]{} 217-231 (1987).\ H.W. Lenstra, Jr., [*New algorithms in algebraic number theory*]{}, Bull. Am. Math. Soc., [**26**]{} 211-241 (1992). H.M. Edwards, [*Galois Theory*]{}, Springer-Verlag, New York (1984). I. Newton, [*Aritmetica Universalis*]{} (1707); in english, [*The Mathematical Works of Isaac Newton*]{}, vol. 2, Johnson Reprint Corp. (1967). Table Captions [**Table 1.**]{} Functions over the natural numbers $\cal N$, that appear in this work, are calculated for the first $n\in\cal N$. $N_n$ is the number of irreducible polynomials of degree $n$ over $GF(2)$. $N_T^{(n)}=\sum_m N_m$ with $m|n$ represents the total number of orbitals in ${\cal B}_n$. $N_{T1}^{(n)}$ and $N_{T0}^{(n)}$ are the number of orbitals in ${\cal B}_n$ with trace $1$ or $0$, respectively. [**Table 2,3,4,5,6.**]{} Natural representation of the finite field $GF(2^n)$ for the cases $n=1,2,3,4,5$ in the space of binary sequences ${\cal B}_n$, where the two internal operations are performed in the following way: (i) [*addition*]{}: $(a_++b_+)_+=a_++b_+$ ; (ii)[*multiplication*]{}: $(a*b)_*=a_**b_*$. The set (${\cal B}_n,+,*$) is a representation of the finite field $GF(2^n)$. The $m$ orbits that compose each orbital $[O_m]$, with $m\mid n$, are the $m$ roots of an irreducible polynomial (I.P.) of degree $m$ over $GF(2)$. $n$ $N_n$ $N_T^{(n)}$ $N_{T1}^{(n)}$ $N_{T0}^{(n)}$ ----- ------- ------------- ---------------- ---------------- $1$ $2$ $2$ $1$ $1$ $2$ $1$ $3$ $2$ $1$ $3$ $2$ $4$ $2$ $2$ $4$ $3$ $6$ $4$ $2$ $5$ $6$ $8$ $4$ $4$ $6$ $9$ $14$ $8$ $6$ $7$ $18$ $20$ $10$ $10$ : []{data-label="tabla1"} [|l|l|c|c|c|c|]{}\ & $a_*$ & $a_+$ & $Trace$ & $I.P.$\ $\sigma_0$ & $\sigma_0$ & $1$ & $1$ & $1$ & $x+1$\ $0$ & $0$ & $0$ & $0$ & $0$ & $x$\ [|l|l|l|c|c|c|c|]{}\ & $a_*$ & $a_+$ & $Trace$ & $I.P.$\ $\sigma_0$ & $\sigma_0$ & $\sigma_0$ & $01$ & $01$ & $1$ & $x^2+x+1$\ $\sigma_0^2$ & $\sigma_1$ & $\sigma_1$ & $10$ & $10$ & &\ $\sigma_0^3$ & $\sigma_0+\sigma_1$ & $\sigma_0\sigma_1$ & $11$ & $11$ & $1$ & $x+1$\ $0$ & $0$ & $0$ & $00$ & $00$ & $0$ & $x$\ [|l|l|l|c|c|c|c|]{}\ & $a_*$ & $a_+$ & $Trace$ & $I.P.$\ $\sigma_0$ & $\sigma_0$ & $\sigma_1$ & $001$ & $001$ & &\ $\sigma_0^2$ & $\sigma_1$ & $\sigma_2$ & $010$ & $010$ & $1$ & $x^3+x^2+1$\ $\sigma_0^4$ & $\sigma_2$ & $\sigma_3$ & $100$ & $100$ & &\ $\sigma_0^3$ & $\sigma_0+\sigma_2$ & $\sigma_0\sigma_1$ & $011$ & $101$ & &\ $\sigma_0^6$ & $\sigma_1+\sigma_0$ & $\sigma_1\sigma_2$ & $110$ & $011$ & $0$ & $x^3+x+1$\ $\sigma_0^5$ & $\sigma_2+\sigma_1$ & $\sigma_2\sigma_0$ & $101$ & $110$ & &\ $\sigma_0^7$ & $\sigma_0+\sigma_1+\sigma_2$ & $\sigma_0\sigma_1\sigma_2$ & $111$ & $111$ & $1$ & $x+1$\ $0$ & $0$ & $0$ & $000$ & $000$ & $0$ & $x$\ [|l|l|l|c|c|c|c|]{}\ & $a_*$ & $a_+$ & $Trace$ & $I.P.$\ $\sigma_0$ & $\sigma_0$ & $\sigma_0$ & $0001$ & $0001$ & &\ $\sigma_0^2$ & $\sigma_1$ & $\sigma_1$ & $0010$ & $0010$ & $1$ & $x^4+x^3+1$\ $\sigma_0^4$ & $\sigma_2$ & $\sigma_2$ & $0100$ & $0100$ & &\ $\sigma_0^8$ & $\sigma_3$ & $\sigma_3$ & $1000$ & $1000$ & &\ $\sigma_0^3$ & $\sigma_0+\sigma_1+\sigma_3$ & $\sigma_0\sigma_1$ & $0011$ & $1011$ & &\ $\sigma_0^6$ & $\sigma_1+\sigma_2+\sigma_0$ & $\sigma_1\sigma_2$ & $0110$ & $0111$ & $1$ & $x^4+x^3+x^2+x+1$\ $\sigma_0^{12}$ & $\sigma_2+\sigma_3+\sigma_1$ & $\sigma_2\sigma_3$ & $1100$ & $1110$ & &\ $\sigma_0^9$ & $\sigma_3+\sigma_0+\sigma_2$ & $\sigma_3\sigma_0$ & $1001$ & $1101$ & &\ $\sigma_0^5$ & $\sigma_0+\sigma_2$ & $\sigma_0\sigma_2$ & $0101$ & $0101$ & $1$ & $x^2+x+1$\ $\sigma_0^{10}$ & $\sigma_1+\sigma_3$ & $\sigma_1\sigma_3$ & $1010$ & $1010$ & &\ $\sigma_0^{7}$ & $\sigma_2+\sigma_3$ & $\sigma_0\sigma_1\sigma_2$ & $0111$ & $1100$ & &\ $\sigma_0^{14}$ & $\sigma_3+\sigma_0$ & $\sigma_1\sigma_2\sigma_3$ & $1110$ & $1001$ & $0$ & $x^4+x+1$\ $\sigma_0^{13}$ & $\sigma_0+\sigma_1$ & $\sigma_2\sigma_3\sigma_0$ & $1101$ & $0011$ & &\ $\sigma_0^{11}$ & $\sigma_1+\sigma_2$ & $\sigma_3\sigma_0\sigma_1$ & $1011$ & $0110$ & &\ $\sigma_0^{15}$ & $\sigma_0+\sigma_1+\sigma_2+\sigma_3$ & $\sigma_0\sigma_1\sigma_2\sigma_3$ & $1111$ & $1111$ & $1$ & $x+1$\ $0$ & $0$ & $0$ & $0000$ & $0000$ & $0$ & $x$\ [|l|l|l|c|c|c|c|]{}\ & $a_*$ & $a_+$ & $Trace$ & $I.P.$\ $\sigma_0$ & $\sigma_0$ & $\sigma_0$ & $00001$ & $00001$ & &\ $\sigma_0^2$ & $\sigma_1$ & $\sigma_0$ & $00010$ & $00010$ & & $x^5+x^4+x^2$\ $\sigma_0^4$ & $\sigma_2$ & $\sigma_2$ & $00100$ & $00100$ & $1$ & $+ x+1$\ $\sigma_0^8$ & $\sigma_3$ & $\sigma_3$ & $01000$ & $01000$ & &\ $\sigma_0^{16}$ & $\sigma_4$ & $\sigma_4$ & $10000$ & $10000$ & &\ $\sigma_0^3$ & $\sigma_0+\sigma_3$ & $\sigma_0\sigma_1$ & $00011$ & $01001$ & &\ $\sigma_0^6$ & $\sigma_1+\sigma_4$ & $\sigma_1\sigma_2$ & $00110$ & $10010$ & &\ $\sigma_0^{12}$ & $\sigma_2+\sigma_0$ & $\sigma_2\sigma_3$ & $01100$ & $00101$ & $0$ & $x^5+x^3+1$\ $\sigma_0^{24}$ & $\sigma_3+\sigma_1$ & $\sigma_3\sigma_4$ & $11000$ & $01010$ & &\ $\sigma_0^{17}$ & $\sigma_4+\sigma_2$ & $\sigma_4\sigma_0$ & $10001$ & $10100$ & &\ $\sigma_0^5$ & $\sigma_3+\sigma_4$ & $\sigma_0\sigma_2$ & $00101$ & $11000$ & &\ $\sigma_0^{10}$ & $\sigma_4+\sigma_0$ & $\sigma_1\sigma_3$ & $01010$ & $10001$ & & $x^5+x^3+x^2$\ $\sigma_0^{20}$ & $\sigma_0+\sigma_1$ & $\sigma_2\sigma_4$ & $10100$ & $00011$ & $0$ & $+ x+1$\ $\sigma_0^{9}$ & $\sigma_1+\sigma_2$ & $\sigma_3\sigma_0$ & $01001$ & $00110$ & &\ $\sigma_0^{18}$ & $\sigma_2+\sigma_3$ & $\sigma_4\sigma_1$ & $10010$ & $01100$ & &\ $\sigma_0^{7}$ & $\sigma_0+\sigma_2+\sigma_3+\sigma_4$ & $\sigma_0\sigma_1\sigma_2$ & $00111$ & $11101$ & &\ $\sigma_0^{14}$ & $\sigma_1+\sigma_3+\sigma_4+\sigma_0$ & $\sigma_1\sigma_2\sigma_3$ & $01110$ & $11011$ & &\ $\sigma_0^{28}$ & $\sigma_2+\sigma_4+\sigma_0+\sigma_1$ & $\sigma_2\sigma_3\sigma_4$ & $11100$ & $10111$ & $0$ & $x^5+x^2+1$\ $\sigma_0^{25}$ & $\sigma_3+\sigma_0+\sigma_1+\sigma_2$ & $ \sigma_3\sigma_4\sigma_0$ & $11001$ & $01111$ & &\ $\sigma_0^{19}$ & $\sigma_4+\sigma_1+\sigma_2+\sigma_3$ & $\sigma_4\sigma_0\sigma_1$ & $10011$ & $11110$ & &\ $\sigma_0^{11}$ & $\sigma_1+\sigma_2+\sigma_4$ & $\sigma_0\sigma_1\sigma_3$ & $01011$ & $10110$ & &\ $\sigma_0^{22}$ & $\sigma_2+\sigma_3+\sigma_0$ & $\sigma_1\sigma_2\sigma_4$ & $10110$ & $01101$ & & $x^5+x^4+x^3$\ $\sigma_0^{13}$ & $\sigma_3+\sigma_4+\sigma_1$ & $\sigma_2\sigma_3\sigma_0$ & $01101$ & $11010$ & $1$ & $+ x^2+1$\ $\sigma_0^{26}$ & $\sigma_4+\sigma_0+\sigma_2$ & $\sigma_3\sigma_4\sigma_1$ & $11010$ & $10101$ & &\ $\sigma_0^{21}$ & $\sigma_0+\sigma_1+\sigma_3$ & $\sigma_4\sigma_0\sigma_2$ & $10101$ & $01011$ & &\ $\sigma_0^{15}$ & $\sigma_0+\sigma_3+\sigma_4$ & $\sigma_0\sigma_1\sigma_2\sigma_3$ & $01111$ & $11001$ & &\ $\sigma_0^{30}$ & $\sigma_1+\sigma_4+\sigma_0$ & $\sigma_1\sigma_2\sigma_3\sigma_4$ & $11110$ & $10011$ & & $x^5+x^4+x^3$\ $\sigma_0^{29}$ & $\sigma_2+\sigma_0+\sigma_1$ & $\sigma_2\sigma_3\sigma_4\sigma_0$ & $11101$ & $00111$ & $1$ & $+ x+1$\ $\sigma_0^{27}$ & $\sigma_3+\sigma_1+\sigma_2$ & $\sigma_3\sigma_4\sigma_0\sigma_1$ & $11011$ & $01110$ & &\ $\sigma_0^{23}$ & $\sigma_4+\sigma_2+\sigma_3$ & $\sigma_4\sigma_0\sigma_1\sigma_2$ & $10111$ & $11100$ & &\ $\sigma_0^{31}$ & $\sigma_0+\sigma_1+\sigma_2+\sigma_3+\sigma_4$ & $\sigma_0\sigma_1\sigma_2\sigma_3\sigma_4$ & $11111$ & $11111$ & $1$ & $x+1$\ $0$ & $0$ & $0$ & $00000$ & $00000$ & $0$ & $x$\

ArXiv

--- abstract: 'Using a method of local transmission matrix, we generalize the well-known WKB formula for a barrier penetrability to multi-channel systems. We compare the WKB penetrability with a solution of the coupled-channels equations, and show that the WKB formula works well at energies well below the lowest adiabatic barrier. We also discuss the eigen-channel approach to a multi-channel tunneling, which may improve the performance of the WKB formula near and above the barrier.' address: - '$^1$Yukawa Institute for Theoretical Physics, Kyoto University, Kyoto 606-8502, Japan ' - '$^2$Department of Physics, Tohoku University, Sendai 980-8578, Japan' - '$^3$Department of Physics, University of Wisconsin, Madison, WI 53706' author: - 'K. Hagino$^{1,2}$ and A.B. Balantekin$^{3}$' title: 'WKB approximation for multi-channel barrier penetrability' --- Introduction ============ The coupled-channels approach has been a standard method in describing atomic, molecular, and nuclear reactions involving internal degrees of freedom [@BT98; @DHRS98; @HRK99; @BBGR93; @YK03]. In nuclear physics, for instance, the coupled-channels method has been successfully applied to heavy-ion fusion reactions at energies around the Coulomb barrier in order to discuss the effect of couplings between the relative motion of the colliding nuclei and inelastic excitations in the target nucleus [@BT98; @DHRS98; @HRK99]. At these energies, the fusion reaction takes place by a quantum tunneling, and the coupled-channels calculations well account for the enhancement of the barrier penetrability due to the channel couplings. A difficulty in the coupled-channels calculations, however, is that it is sometimes not so easy to obtain a numerically stable solution with a controlled accuracy. This is particularly the case in the presence of closed channels and/or when the coupling strength is strong in the classically forbidden region. Several methods have been proposed in order to stabilize the numerical solution [@BBGR93; @J78; @L84; @RO82]. In this paper, instead of directly integrating the coupled-channels equations with the stabilization techniques, we solve them using the WKB approximation. To this end, we employ the method of local transmission matrix, which was originally developed by Brenig and Russ in order to stabilize numerical solutions of the coupled-channels equations [@BBGR93]. We solve the equation for the local transmission matrix under the semi-classical assumption, and generalize the well-known WKB formula for a barrier penetrability for a single channel to coupled-channel systems. Since the penetrability is expressed in a compact form, the resultant WKB formula is entirely free from the problem of numerical instability. Moreover, the WKB method can be easily applied to systems with a large number of degrees of freedom, while obtaining a direct solution of the coupled-channels equations can be computationally demanding. Also, the WKB method is useful in gaining a physical intuition for the dynamics of multi-channel tunneling. The paper is organized as follows. In Sec. II, we set up the coupled-channels equations and introduce the local transmission matrix. We derive the semi-classical expression for the local transmission matrix, and obtain the WKB formula for a multi-channel penetrability. We apply the WKB formula to a three-channel problem, and compare the penetrability with the numerical solutions of the coupled-channel equations. In Sec. III. we discuss the penetrability at energies near and above the barrier height. The WKB formula which we derive works when the multiple reflection of the classical path under the barrier can be neglected, that is, at energies well below the barrier. We show that the eigen-channel approach can provide good prescriptions at higher energies, where the primitive WKB formula breaks down. We summarize the paper in Sec. IV. Multi-channel WKB formula ========================= Our aim in this paper is to derive the WKB formula for penetrability for a one dimensional potential barrier in the presence of channel couplings. We consider the following coupled-channels equations: $$-\frac{\hbar^2}{2m}\,\frac{d^2}{dx^2}\,u_{nn_0}(x)+\sum_m \left[V_{nm}(x)+\epsilon_m\delta_{n,m}-E\right]u_{mn_0}(x)=0.$$ Here, $m$ is the mass of a particle, $\epsilon_n$ is the excitation energy for the $n$-th channel, and $E$ is the total energy of the system. $u_{nn_0}(x)$ is the wave function matrix, where $n$ refers to the channel while $n_0$ specifies the incident channel. Notice that we express the wave functions in a matrix form by combining $N$ linearly independent solutions of the coupled-channels equations, $N$ being the dimension of the coupled equations. For the situation where the particle is incident on the barrier from the right hand side, the boundary conditions for $u_{nn_0}(x)$ are given by, $$\begin{aligned} u_{nn_0}(x)&\to& T_{nn_0}e^{-ik_nx}~~~~~~~~~~~~~~~~~~~(x\to -\infty), \label{boundary1}\\ &\to& \delta_{nn_0}e^{-ik_nx}+R_{nn_0}e^{ik_nx}~~~~(x\to \infty), \label{boundary2}\end{aligned}$$ where $k_n=\sqrt{2m(E-\epsilon_n)/\hbar^2}$ is the wave number for the $n$-th channel. The inclusive penetrability is then obtained as, $$P=\sum_n\frac{k_n}{k_{n_0}}\,|T_{nn_0}|^2. \label{P}$$ Let us now introduce the local transmission matrix defined by [@BBGR93] $${\mbox{\boldmath $\tau$}}(x)\equiv \frac{1}{2}\,\left[i{\mbox{\boldmath $q$}}(x)^{-1}{\mbox{\boldmath $u$}}'(x)+{\mbox{\boldmath $u$}}(x)\right],$$ where ${\mbox{\boldmath $q$}}(x)=[2m(E-{\mbox{\boldmath $W$}}(x))/\hbar^2]^{1/2}$ with $W_{nm}(x)=V_{nm}(x)+\epsilon_n\delta_{n,m}$. From Eqs. (\[boundary1\]) and (\[boundary2\]), the asymptotic form of ${\mbox{\boldmath $\tau$}}(x)$ reads, $$\begin{aligned} \tau_{nm}(x)&\to& T_{nm}e^{-ik_nx}~~~~~~~~~~(x\to -\infty), \label{boundary3} \\ &\to& \delta_{n,m}e^{-ik_nx}~~~~~~~~~~(x\to \infty). \end{aligned}$$ Here, we have used the fact $q_{nm}(x)\to k_n\delta_{n,m}$ as $x\to \pm\infty$. It is easy to show that the local transmission matrix ${\mbox{\boldmath $\tau$}}(x)$ obeys the equation[@BBGR93], $${\mbox{\boldmath $\tau$}}'(x)=-\frac{1}{2}{\mbox{\boldmath $q$}}^{-1}(x){\mbox{\boldmath $q$}}'(x)[1+{\mbox{\boldmath $\rho$}}(x)] {\mbox{\boldmath $\tau$}}(x)-i{\mbox{\boldmath $q$}}(x){\mbox{\boldmath $\tau$}}(x), \label{tau}$$ where $${\mbox{\boldmath $\rho$}}(x)\equiv [i{\mbox{\boldmath $q$}}^{-1}(x){\mbox{\boldmath $u$}}'(x)-{\mbox{\boldmath $u$}}(x)] \cdot[i{\mbox{\boldmath $q$}}^{-1}(x){\mbox{\boldmath $u$}}'(x)+{\mbox{\boldmath $u$}}(x)]^{-1},$$ is the local reflection matrix[@BBGR93]. The WKB approximation may be obtained by neglecting ${\mbox{\boldmath $\rho$}}(x)$ in Eq. (\[tau\]) [@DR79; @R03], that is, $${\mbox{\boldmath $\tau$}}'(x)= -\frac{1}{2}{\mbox{\boldmath $q$}}^{-1}(x){\mbox{\boldmath $q$}}'(x) {\mbox{\boldmath $\tau$}}(x)-i{\mbox{\boldmath $q$}}(x){\mbox{\boldmath $\tau$}}(x). \label{tauWKB}$$ A similar equation has been derived by Van Dijk and Razavy [@DR79; @R03], but by using the method of variable reflection amplitude (see also Ref. [@BK58]). Notice that the asymptotic form of the local reflection matrix ${\mbox{\boldmath $\rho$}}(x)$ is [@BBGR93] $$\begin{aligned} \rho_{nm}(x)&\to& 0~~~~~~~~~~~~~~~~~~~~~(x\to -\infty), \\ &\to& -R_{nm}e^{2ik_nx}~~~~~~~(x\to \infty). \end{aligned}$$ Neglecting ${\mbox{\boldmath $\rho$}}(x)$ in Eq. (\[tau\]) is thus equivalent to ignoring the reflection, that is reasonable in the semi-classical limit. This, in fact, corresponds to the lowest order of the Bremmer expansion [@B49; @L51; @B82; @MH96; @G74], where the WKB formula is obtained by approximating a smooth potential with a series of sharp potential steps [^1]. For a single channel problem, Eq. (\[tauWKB\]) can be easily integrated to yield $$\tau(x)=\sqrt{\frac{q({\infty})}{q(x)}}\,e^{-i\int^xq(x')dx'}.$$ For a coupled-channels problem, however, care must be taken in the integration, since ${\mbox{\boldmath $q$}}(x)$ and ${\mbox{\boldmath $q$}}'(x)$ do not commute to each other in general. We attempt to solve Eq. (\[tauWKB\]) by discretizing the coordinate with a mesh spacing of $\Delta x$. Replacing the derivative term by a simple point difference formula, we obtain $$\begin{aligned} {\mbox{\boldmath $\tau$}}(x_{n-1})&\sim& {\mbox{\boldmath $\tau$}}(x_{n})+\Delta x\left[i {\mbox{\boldmath $q$}}(x_n){\mbox{\boldmath $\tau$}}(x_n) +\frac{1}{2}{\mbox{\boldmath $q$}}^{-1}(x_n){\mbox{\boldmath $q$}}'(x_n) {\mbox{\boldmath $\tau$}}(x_n)\right], \\ &\sim& \left[1+\frac{\Delta x}{2}{\mbox{\boldmath $q$}}^{-1}(x_n){\mbox{\boldmath $q$}}'(x_n)\right] \cdot \left[1+i{\mbox{\boldmath $q$}}(x_n)\Delta x\right]{\mbox{\boldmath $\tau$}}(x_n), \\ &\sim& \left[1-\Delta x\,{\mbox{\boldmath $q$}}^{-1}(x_n){\mbox{\boldmath $q$}}'(x_n)\right]^{-1/2} \cdot e^{i{\mbox{\boldmath $q$}}(x_n)\Delta x}{\mbox{\boldmath $\tau$}}(x_n), \label{tauWKB2}\end{aligned}$$ to the lowest order of $\Delta x$. Using ${\mbox{\boldmath $q$}}(x_{n-1})\sim {\mbox{\boldmath $q$}}(x_{n})[1-\Delta x\,{\mbox{\boldmath $q$}}^{-1}(x_n){\mbox{\boldmath $q$}}'(x_n)]$, the first factor in Eq. (\[tauWKB2\]) is transformed to be $$[1-\Delta x\,{\mbox{\boldmath $q$}}^{-1}(x_n){\mbox{\boldmath $q$}}'(x_n)]^{-1/2} \sim\frac{1}{\sqrt{{\mbox{\boldmath $q$}}^{-1}(x_n){\mbox{\boldmath $q$}}(x_{n-1})}}\cdot \frac{1}{\sqrt{{\mbox{\boldmath $q$}}(x_n)}}\cdot \sqrt{{\mbox{\boldmath $q$}}(x_n)}.$$ Approximating $[{\mbox{\boldmath $q$}}^{-1}(x_n){\mbox{\boldmath $q$}}(x_{n-1})]^{-1/2} [{\mbox{\boldmath $q$}}(x_n)]^{-1/2} \sim [{\mbox{\boldmath $q$}}(x_n) {\mbox{\boldmath $q$}}^{-1}(x_n){\mbox{\boldmath $q$}}(x_{n-1})]^{-1/2} =1/\sqrt{{\mbox{\boldmath $q$}}(x_{n-1})}$, we finally obtain $${\mbox{\boldmath $\tau$}}(x_{n-1})= \frac{1}{\sqrt{{\mbox{\boldmath $q$}}(x_{n-1})}} \cdot e^{i{\mbox{\boldmath $q$}}(x_n)\Delta x} \sqrt{{\mbox{\boldmath $q$}}(x_n)} {\mbox{\boldmath $\tau$}}(x_n).$$ Iterating this equation backward from $x=\infty$, we obtain $${\mbox{\boldmath $\tau$}}(-\infty)= \frac{1}{\sqrt{{\mbox{\boldmath $q$}}(-\infty)}} \cdot \left(\prod_i e^{i{\mbox{\boldmath $q$}}(x_i)\Delta x}\right) \sqrt{{\mbox{\boldmath $q$}}(\infty)}. \label{taufin}$$ Substituting this expression into Eq. (\[P\]) together with Eq. (\[boundary3\]), the WKB approximation to the multi-channel penetrability reads $$P=\sum_n\left|\left\langle n\left| \prod_i e^{i{\mbox{\boldmath $q$}}(x_i)\Delta x}\right| n_0\right\rangle\right|^2. \label{PWKB}$$ This is the main result in this paper. Notice that the factor $k_n/k_{n_0}$ does not appear in the WKB formula, Eq. (\[PWKB\]). In practice, one can evaluate Eq. (\[PWKB\]) by diagonalizing ${\mbox{\boldmath $W$}}(x)$ at each $x_i$. This yields $$P=\sum_n\left|\left\langle n\left| \prod_i \left[\sum_m\left|m(x_i)\left\rangle e^{iq_m(x_i)\Delta x} \right\langle m(x_i)\right|\right] \right| n_0\right\rangle\right|^2, \label{PWKB2}$$ where $|m(x)\rangle$ is the eigen-vector of the matrix ${\mbox{\boldmath $W$}}(x)$ with the eigen-value of $\lambda_m(x)$, and $q_m(x)\equiv \sqrt{2m(E-\lambda_m(x))/\hbar^2}$. For a one dimensional problem, Eq. (\[PWKB\]) is reduced to the familiar WKB formula, $$P(E)= \exp\left[-2\int^{x_1}_{x_0}dx'\sqrt{\frac{2m}{\hbar^2} (V(x')-E)}\right],$$ where $x_0$ and $x_1$ are the inner and the outer turning points, respectively. Let us now apply the WKB formula (\[PWKB\]) to a three-level problem. We consider the following coupling potential: $${\mbox{\boldmath $W$}}(x)=\left(\matrix{V(x) & F(x) & 0 \cr F(x)& V(x)+\epsilon & F(x) \cr 0 & F(x) & V(x)+2\epsilon \cr } \right), \label{cc}$$ with $$\begin{aligned} V(x)&=&V_0e^{-x^2/2s^2}, \\ F(x)&=&F_0e^{-x^2/2s_f^2}. \end{aligned}$$ The parameters are chosen following Ref. [@DLW83] to be $V_0$=100 MeV, $F_0$=3 MeV, and $s=s_f=$3 fm, which mimic the fusion reaction between two $^{58}$Ni nuclei. The excitation energy $\epsilon$ and the mass $m$ are taken to be 2 MeV and 29$m_N$, respectively, where $m_N$=938 MeV is the nucleon mass. With these parameters, the three eigen-barriers $\lambda_i(x)$, which are obtained by diagonalizing ${\mbox{\boldmath $W$}}(x)$ at each $x$, have the barrier height of 97.31, 102.0, and 106.7 MeV, respectively, while the barrier height for the uncoupled barrier $V(x)$ is 100 MeV (See Fig.1). Before we study the three channel problem, let us first examine the validity of the WKB approximation for a single channel case to see whether the semi-classical approximation works in principle for the parameters which we choose. Figure 2 shows a comparison between the WKB penetrability for the uncoupled barrier $V(x)$ (the dashed line) obtained with Eq. (\[PWKB\]) and the exact solution. It is plotted in the linear and logarithmic scales in the upper and the lower panels, respectively. One clearly sees that the WKB approximation indeed works well at energies about 2 MeV below the barrier height and lower. As it is well known, the naive WKB approximation breaks down around the barrier. In fact, the WKB penetrability is unity at the barrier height, while the exact result is about a half. Around these energies, one needs to improve the WKB formula by using the uniform approximation in order to take into account the multiple reflection of the classical path under the barrier [@BS83; @B85]. We will discuss this problem more in the next section in connection to the penetrability in a coupled-channels system. Let us now solve the coupled-channels problem, Eq. (\[cc\]). We integrate Eq. (\[PWKB2\]) from $x=-15$ fm to $x$=15 fm with $\Delta x$=0.05 fm. The dashed line in Fig. 3 shows the penetrability in the WKB approximation for this problem, which is compared to the exact solution of the coupled-channels equations (the solid line). The figure also shows the penetrability for the uncoupled barrier as a comparison (the dotted line). Remarkably, the WKB formula (\[PWKB\]) reproduces almost perfectly the exact solution at energies well below the lowest adiabatic barrier, i.e., 97.31 MeV, as in the single channel problem. We notice that the WKB penetrability increases much more rapidly than the exact penetrability at energies corresponding to the height of each eigenbarrier. This is in close analogue to the single channel problem shown in Fig. 2. This behaviour may be expected in the eigen-channel approach discussed in Refs. [@DLW83; @HTB97; @DL87]. We will discuss this point in the next section. Penetrability near and above the barrier ======================================== In the previous section, we have shown that the multi-channel WKB formula works remarkably well at energies well below the lowest eigenbarrier. Therefore, it can be expected that the WKB formula provides a useful framework in discussing, for instance, the role of inelastic excitations in the colliding nuclei in nuclear reactions at extremely low energies, such as astrophysically relevant reactions, where the standard coupled-channels calculations may be difficult to carry out. At higher energies, however, we found that the agreement between the primitive WKB formula and the exact solution of the coupled-channels equations becomes poor. For a single-channel problem, one can cure this problem by using the uniform approximation [@BS83; @B85]. The WKB formula which is valid at all energies is given by, $$P(E)=\frac{1}{1+ \exp\left[2\int^{x_1}_{x_0}dx'\sqrt{\frac{2m}{\hbar^2} (V(x')-E)}\right]},$$ where the turning points $x_0$ and $x_1$ become complex numbers when the energy $E$ is above the barrier. It is not straightforward at all, however, to extend the uniform approximation to the coupled-channels problem. In this section, we instead present two prescriptions to deal with the coupled-channels penetrability at energies near and above the barrier. Dynamical Norm Method --------------------- The first prescription is closely related to the dynamical norm method developed in Ref. [@THA95]. It was argued in Ref. [@THA95] that the penetrability may be expressed as a product of the penetrability in the adiabatic limit and a multiplicative factor to it which accounts for the non-adiabatic effect. The latter factor, which was called the dynamical norm factor, was evaluated through the imaginary time evolution for an intrinsic degree of freedom with a classical path obtained with the adiabatic potential. We follow here the same idea as in the dynamical norm method, and re-express Eq. (\[taufin\]) as $${\mbox{\boldmath $\tau$}}(-\infty)= \frac{1}{\sqrt{{\mbox{\boldmath $q$}}(-\infty)}} \left(\prod_i e^{iq_0(x_i)\Delta x}e^{i[{\mbox{\boldmath $q$}}(x_i)-q_0(x_i)]\Delta x} \right) \sqrt{{\mbox{\boldmath $q$}}(\infty)},$$ where $q_0(x)=\sqrt{2m(E-\lambda_0(x))/\hbar^2}$ is the local wave number for the lowest eigen-barrier (i.e., the adiabatic barrier), $\lambda_0(x)$. The penetrability $P(E)$ is then given by $$P=\sum_n\left|\left\langle n\left| \prod_i e^{i[{\mbox{\boldmath $q$}}(x_i)-q_0(x_i)]\Delta x}\right|n_0\right\rangle\right|^2 \cdot \exp\left[-2\int^{x_1}_{x_0}dx'\sqrt{\frac{2m}{\hbar^2} (\lambda_0(x')-E)}\right]. \label{DNM}$$ The second factor on the right hand side (r.h.s.) of this equation is nothing more than the WKB penetrability for the adiabatic potential $\lambda_0(x)$. At this stage, one may replace it by the exact penetrability for the adiabatic potential, $P_{\rm ad}(E)$. The first factor on the r.h.s. of Eq. (\[DNM\]) expresses the non-adiabatic effect, as in the dynamical norm factor introduced in Ref. [@THA95]. Notice that our formula, (\[DNM\]), is in fact an improvement of the dynamical norm method in Ref. [@THA95], since the classical path is not assumed from the beginning in evaluating the dynamical norm factor. The result of the modified dynamical norm method is shown in Fig. 4 by the dashed line. It is evident that the agreement between the WKB approximation and the exact solution improves significantly at energies near and slightly above the adiabatic barrier, although the method provides essentially the same result as the original WKB formula, (\[PWKB\]), at higher energies. Eigen-channel approach ---------------------- The second prescription which we discuss is based on the eigen-channel approximation. In this approach, the penetrability is expressed as a weighted sum of the penetrability for the eigenbarriers [@DLW83; @HTB97; @DL87], that is, $$P(E)=\sum_nw_nP_n(E), \label{eigen}$$ where $P_n(E)$ is the penetrability for the eigen-potential $\lambda_n(x)$. The weight factors $w_n$ are usually estimated by assuming that the matrix ${\mbox{\boldmath $W$}}(x)$ is independent of $x$ through the interaction range [@DLW83; @DL87], while the coordinate dependence of ${\mbox{\boldmath $W$}}(x)$ is properly taken into account in calculating the penetrability, $P_n(E)$. One often takes the barrier position of the uncoupled barrier, $X_b$, in order to estimate the weight factors [@DL87]. This leads to (see Eq.(\[PWKB2\])) $$w_n=w_n(X_b)=|\langle n(X_b)|n_0\rangle|^2.$$ This procedure is indeed exact when the intrinsic degree of freedom has a degenerate spectrum [@HTB97; @HTBB95; @NBT86; @E81], since the matrix ${\mbox{\boldmath $W$}}(x)$ can be diagonalized independent of $x$. When the intrinsic states have a finite excitation energy, however, the unitary matrix which diagonalizes ${\mbox{\boldmath $W$}}(x)$ explicitly depends on $x$, and the results may depend strongly on the position where the weight factors are evaluated. Also the weight factors possess some energy dependence in general. In Ref. [@HTB97], we have explicitly shown for a two channel problem that the optimum weight factors are considerably different from those estimated at the barrier position, although their energy dependence appears to be weak. A satisfactory procedure to determine the weight factors has not yet been found so far when the excitation energy is finite. In Fig.3, we have shown that the WKB penetrability approaches to a constant value at the barrier height of each eigen barrier. Assuming that the weight factors are independent of energy, one can exploit this fact to determine a consistent value of the weight factors in the eigen-channel approach. For instance, at the barrier height of the lowest eigenbarrier, $E=B_0$, assuming that the contribution from the higher barriers is negligible, Eq. (\[eigen\]) suggests $$P(B_0)\sim w_0P_0(B_0) \sim w_0,$$ in the primitive WKB approximation (i.e., without the uniform approximation). Therefore, if one evaluates Eq. (\[PWKB\]) at $E=B_0$, it directly provides the weight factor for the lowest eigenbarrier. One can repeat this procedure $N-1$ times to determine the weight factors $w_0, w_1, \cdots, w_{N-2}$: suppose that the weight factors for the $k$ lowest eigen-barriers have been determined. The weight factor for the $(k+1)$-th eigenbarrier is then estimated as $$w_k=P_{\rm WKB}(B_k)-\sum_{i=0}^{k-1} w_i,$$ where $B_k$ is the barrier height of the $(k+1)$-th eigenbarrier, and $P_{\rm WKB}$ is the penetrability in the WKB approximation, (\[PWKB\]). Here, we have used the fact $P_i(B_k)=1$ for $i\leq k$ in the primitive WKB approximation, and assumed $P_i(B_k)=0$ for $i>k$. The weight factor for the highest eigenbarrier $\lambda_{N-1}(x)$ is evaluated as $$w_{N-1}=1-\sum_{i=0}^{N-2}w_i,$$ in order to ensure the unitarity. We apply this prescription to the three channel problem discussed in the previous section. The weight factors are evaluated to be 0.5914, 0.3543, and 0.0543 for the lowest, the second lowest, and the highest eigen barriers, respectively. The result of the eigen channel approximation with the weight factors thus estimated is denoted by the filled circles in Fig. 4. The result is indistinguishable from the exact solution of the coupled-channels equations for all the energy region shown in the figure. We thus conclude that the multi-channel WKB formula which we derive in this paper provides a consistent way to determine the weight factors in the eigen-channel approach, and it provides a useful and simple prescription to compute the penetrability in the presence of channel couplings at energies from well below to well above the potential barrier, as long as the weight factors are slowly varying functions of energy. Summary ======= We have extended the well known WKB formula for barrier penetrability to systems with intrinsic degrees of freedom. Applying the formula to a three channel problem, we have explicitly demonstrated that the WKB formula reproduces very nicely the exact solution of coupled-channels equations at energies well below the lowest eigenbarrier, i.e., the adiabatic barrier. The WKB formula which we derived is applicable even to systems with a large number of degrees of freedom, where the standard coupled-channels calculation cannot be performed due to a computational limitation. Our method may therefore provide a useful framework to discuss, e.g., a quantum scattering problem in the presence of coupling to a heat bath [@BA03; @CC95]. The WKB formula is also useful when one discusses the channel coupling effect on the tunneling rate at deep subbarrier energies, especially in the presence of closed channels, since the direct integration of the coupled-channels equations may suffer from a numerical instability. Such interesting problems include heavy-ion fusion reactions at extremely low energies [@HRD03], electron screening effects in nuclear astrophysical reactions [@SRSC01; @L93], and nuclear structure effects in astrophysical fusion reactions [@HHB03; @NCT97]. The WKB formula which we derived neglects the effect of multiple reflection of a classical path under the potential barrier. Such primitive formula breaks down at energies near and above the adiabatic barrier, as is well known. We discussed two prescription to cure this problem. One is the dynamical norm method, where the WKB formula is re-expressed as a product of the penetrability for the adiabatic barrier and a multiplicative factor which accounts for the non-adiabatic effect. By replacing the adiabatic penatrability in the WKB approximation by the exact one, we have shown that this prescription improves the result at energies near and slightly above the adiabatic barrier. The second prescription is the eigen-channel approximation, where the penetrability is given as a weighted sum of the eigen penetrability. By applying the WKB formula at energies corresponding to the barrier height of each eigen barrier, we have presented a consistent procedure to determine the weight factors. We have shown that this prescription works well at energies from well below to well above the barrier, as long as the energy dependence of the weight factors is negligible. This method would provide a useful way to simplify the coupled-channels calculations in realistic systems. For a single channel problem, the primitive WKB formula can be improved by using the uniform approximation [@BS83; @B85], where the resultant WKB formula is applicable at energies both below and above the barrier. It will be an interesting, but very challenging, problem to extend it to a multi-channel problem. For this purpose, it will be extremely helpful to construct a solvable coupled-channels model. A work towards this direction is now in progress [@HB04]. Acknowledgments {#acknowledgments .unnumbered} =============== The authors thank H. Horiuchi and Y. Fujiwara for useful discussions. A.B.B. gratefully acknowledges the 21st Century for Center of Excellence Program “Center for Diversity and Universality” at Kyoto University for financial support and thanks the Yukawa Institute for Theoretical Physics for its hospitality. This work was supported in part by the U.S. National Science Foundation Grant No. PHY-0244384 and in part by the University of Wisconsin Research Committee with funds granted by the Wisconsin Alumni Research Foundation. A.B. Balantekin and N. Takigawa, Rev. Mod. Phys. [**70**]{}, 77 (1998). M. Dasgupta, D.J. Hinde, N. Rowley, and A.M. Stefanini, Annu. Rev. Nucl. Part. Sci. [**48**]{}, 401 (1998). K. Hagino, N. Rowley, and A.T. Kruppa, Comp. Phys. Comm. [**123**]{}, 143 (1999). W. Brenig, T. Brunner, A. Gross, and R. Russ, Z. Phys. [**B93**]{}, 91 (1993); W. Brenig and R. Russ, Surf. Sci. [**315**]{}, 195 (1994). N. Yamanaka and Y. Kino, Phys. Rev. A[**68**]{}, 052715 (2003). B.R. Johnson, J. Chem. Phys. [**69**]{}, 4678 (1978). Z.H. Levine, Phys. Rev. A[**30**]{}, 1120 (1984). T.N. Rescigno and A.E. Orel, Phys. Rev. A[**25**]{}, 2402 (1982). W. Van Dijk and M. Razavy, Can. J. Phys. [**57**]{}, 1952 (1979). M. Razavy, [*Quantum Theory of Tunneling*]{}, (World Scientific Publishing Co., Singapore, 2003). R. Bellman and R. Kalaba, Proc. Nat. Acad. Sci. U.S.A. [**44**]{}, 317 (1958). H. Bremmer, Physics [**15**]{}, 593 (1949). R. Landauer, Phys. Rev. [**82**]{}, 80 (1951). M.V. Berry, J. Phys. A[**15**]{}, 3693 (1982). N.T. Maitra and E.J. Heller, Phys. Rev. A[**54**]{}, 4763 (1996). S. Gasiorowicz, [*Quantum Physics*]{} (Wiley, New York, 1974). C.H. Dasso, S. Landowne, and A. Winther, Nucl. Phys. [**A405**]{}, 381 (1983); [**A407**]{}, 221 (1983). D.M. Brink and U. Smilansky, Nucl. Phys. [**A405**]{}, 301 (1983). D.M. Brink, [*Semi-Classical Methods for Nucleus-Nucleus Scattering*]{} (Cambridge University Press, Cambridge, England, 1985). K. Hagino, N. Takigawa, and A.B. Balantekin, Phys. Rev. C[**56**]{}, 2104 (1997). C.H. Dasso and S. Landowne, Phys. Lett. [**B183**]{}, 141 (1987). N. Takigawa, K. Hagino, and M. Abe, Phys. Rev. C[**51**]{}, 187 (1995). K. Hagino, N. Takigawa, J.R. Bennett, and D.M. Brink, Phys. Rev. C[**51**]{}, 3190(1995). M.A. Nagarajan, A.B. Balantekin, and N. Takigawa, Phys. Rev. C[**34**]{}, 894(1986). H. Esbensen, Nucl. Phys. [**A352**]{}, 147(1981). B. Barkakaty and S. Adhikari, J. of Chem. Phys. [**118**]{}, 5302 (2003). A.H. Castro Neto and A.O. Caldeira, Phys. Lett. [**A209**]{}, 290 (1995). K. Hagino, N. Rowley, and M. Dasgupta, Phys. Rev. C[**67**]{}, 054603 (2003). F. Strieder, C. Rolfs, C. Spitaleri, and P. Corvisiero, Naturwissenschaften [**88**]{}, 461 (2001). K. Langanke, in [*Advances in Nuclear Physics*]{}, edited by J.W. Negele and E. Vogt (Plenum, New York, 1993). K. Hagino, M.S. Hussein, and A.B. Balantekin, Phys. Rev. C[**68**]{}, 048801 (2003). F.M. Nunes, R. Crespo, and I.J. Thompson, Nucl. Phys. [**A615**]{}, 69 (1997). K. Hagino and A.B. Balantekin, to be published. [^1]: The local reflection matrix ${\mbox{\boldmath $\rho$}}(x)$ satisfies the equation ${\mbox{\boldmath $\rho$}}'=i({\mbox{\boldmath $q$}}{\mbox{\boldmath $\rho$}}+{\mbox{\boldmath $\rho$}}{\mbox{\boldmath $q$}}) -(1-{\mbox{\boldmath $\rho$}}){\mbox{\boldmath $q$}}^{-1}{\mbox{\boldmath $q$}}'(1+{\mbox{\boldmath $\rho$}})/2$ [@BBGR93]. One may solve this equation perturbatively assuming that ${\mbox{\boldmath $\rho$}}(x)$ remains small and finding the correction to Eq. (\[tauWKB\]).

ArXiv

--- abstract: 'I describe an approach to fitting and comparison of radio spectra based on Bayesian analysis and realised using a new implementation of the nested sampling algorithm. Such an approach improves on the commonly used maximum-likelihood fitting of radio spectra by allowing objective model selection, calculation of the full probability distributions of the model parameters and provides a natural mechanism for including information other than the measured spectra through *priors*. In this paper I cover the theoretical background, the algorithms used and the implementation details of the computer code. I also briefly illustrate the method with some previously published data for three near-by galaxies. In forthcoming papers we will present the results of applying this analysis larger data sets, including some new observations, and the physical conclusions that can be made. The computer code as well as the overall approach described here may also be useful for analysis of other multi-chromatic broad-band observations and possibly also photometric redshift estimation. All of the code is publicly available, licensed under the GNU General Public License, at <http://www.mrao.cam.ac.uk/~bn204/galevol/speca/index.html>.' author: - | B. Nikolic\ Astrophysics Group, Cavendish Laboratory, Cambridge CB3 0HE, UK\ <email:[email protected]>\ <http://www.mrao.cam.ac.uk/~bn204/> bibliography: - '90ghzgals.bib' date: 'r(date: )' title: Fitting and Comparison of Models of Radio Spectra --- \[firstpage\] Introduction ============ The spectrum of radio emission from the majority of astronomical sources consists of a smooth continuum with (in some but not all sources) atomic and molecular lines superimposed, most notably the H[i]{} line at 1.42GHz and the carbon monoxide rotational line ladder starting at 115GHz. Measurements with significant fractional bandwidths are almost always dominated by the continuum emission. The shape of the continuum emission can be measured by making observations at widely spaced frequencies. The available measurements for some local objects span a very wide range of frequencies, from about 50MHz to above 1THz and into the far-infrared portion of the spectrum. The measured shape of the continuum emission naturally provides information about the physical properties of the sources and mechanisms for emission. There are many examples of physical properties which one can try to extract: - The slope of the synchrotron emission gives the energy spectrum of relativistic electrons in the source, providing constraints on the mechanism by which they are created - The change in slope in the synchrotron spectrum gives an estimate of the age of the source - The low frequency turn-over, i.e., where source becomes opaque due to absorption due to electrons, places constraints on the geometry of the source - The slope of the Raleigh-Jeans part of the dust spectrum places constraints on the properties of the dust - The frequency of the peak of the dust spectrum determines the temperature of the dust and the total power output of the source Additionally one can try to make an estimate of redshift of a distant object from the observed spectral shape. Most of the physics underlying the emission processes at these wavelengths is well understood and it is possible to calculate the expected spectrum given a model and its parameters. Therefore, analysis of radio continuum spectra often consists of ‘fitting’ a selection of known models to the observations. Some of the desirable outputs of such an analysis are: 1. An objective measure of how well the model explains the data 2. Unbiased estimates of parameters of the model, estimates of errors with which these estimates are made, and the correlations between the estimate errors 3. Full probability distributions of parameters in cases that they are significantly non-Gaussian 4. An objective way of comparing how well different models explain the same data All of these can be obtained simultaneously using Bayesian analysis and the nested sampling algorithm. In this paper I describe a computer code to perform this analysis and illustrate it with examples using previously published data for three near-by galaxies. This code was developed to support an ongoing observational programme to measure the radio spectral energy distributions of near-by star-forming galaxies. We are also currently applying he code to analyse the spectra of Ultra-Luminous InfraRed Galaxies (ULIRGs) with the goal of better understanding the their radio emission and the physical conditions within them. These results will be published separately in forthcoming papers. Method ====== The analysis proceeds in the usual fashion, starting with the Bayes equation [see for example: @Jaynes:PTLS; @SiviaD06]: $$p(\theta | D, H) = \frac{ p(D | \theta, H) p(\theta|H)}{p(D|H)}= \frac{ L(\theta) \pi(\theta)}{Z} \label{eq:BayesTheorem}$$ where the symbols have following meaning: $H$ : is the hypothesis under which the data are analysed. In this case, the hypothesis consists of the model we assume for the radio emission and the priors for each of the model parameters $\theta$ : is a vector with elements that are the parameters of the model (e.g., the spectral index $\alpha$, the frequency of the spectral break $\nu_{\rm br}$) $D$ : are the observed data (in this case, the observed flux densities at various frequencies) $p(\theta|H) = \pi(\theta) $ : is the probability that the model parameters take a particular value, i.e., the prior information associated with the hypothesis that we are using $p(D|\theta,H)=L(\theta)$ : is the likelihood, i.e., the probability of observing the data we have given some model parameters $\theta$ $p(D|H)=Z$ : is the so-called Bayesian *evidence*, that is a measure of how well our hypothesis (i.e., the model for the emission spectrum and the priors) predict the data actually observed $p(\theta | D,H)$ : is the posterior joint distribution of the model parameters The two *inputs* to the calculation are: 1. $H$, that is a model for the emission spectrum and the prior on its parameters (see \[sec:synchrotron-models\]) 2. $L(\theta)$, a function which uses observed data, the model and the parameters $\theta$ to calculate the likelihood of a predicted spectrum (see \[sec:likelihood\]) The computation is done using the nested sampling algorithm [@Skilling2006] as described in Section \[sec:nested-sampling\]. The two outputs are: 1. $Z$, the evidence for the model and the prior 2. $p(\theta | D,H)$, the posterior distribution of the model parameters Models of synchrotron radiation spectra {#sec:synchrotron-models} --------------------------------------- In general, the models that it makes sense to try when analysing a particular set of observations are determined by the type of object that has been observed and the region of the spectrum which is being analysed. The current version of the computer code described here already has implementations of models which combine absorption, synchrotron and thermal radiation. In this paper however, I restrict the description to models of non-thermal synchrotron emission. The reason is that in the examples shown later I use only measurements below 5GHz while the synchrotron mechanism is the dominant component of emission from the majority of extragalactic sources at frequencies below about 50GHz, so this is a reasonable approximation. The remaining models present in the software implementation will be used for forthcoming science papers and they can easily be extended further still with thermal dust or spinning dust emission for example. The simplest model of synchrotron emission spectrum is a simple power-law model: $$F_{\nu}(\nu)= F_{\nu}^{0} \cdot \left(\frac{\nu}{1\,{\mathrm{GHz}}}\right)^\alpha$$ with two parameters: $F_{\nu}^{0}$ : The flux density at the frequency of 1GHz $\alpha$ : The spectral index It is not however convenient to make a parametrisation directly in terms of $F_{\nu}^{0}$ since it can take large range of values for typical sources, and we do not *a-prior* know even the order magnitude of $F_{\nu}^{0}$ to expect. For example, if we assumed that $F_{\nu}^{0}$ could be any value between 0.01 and 1Jy with uniform probability then there would be far more values with *magnitude* of order of 1 than of order 0.01. Instead, I parametrise the model in terms of the logarithm of $F_{\nu}^{0}$, i.e., $\log_{10} (F_{\nu}^{0})$ and assume that this is uniformly distributed over a range of values (-2 to 0 in the example above). There is further discussion of this topic of so-called *scale* parameters by [@Jaynes:PTLS]. A more complicated and physically more accurate model is the so called continuous-injection model [e.g., @1962SvA.....6..317K]. In this model it is assumed that the energetic electrons have a power-law energy distribution when they are created, and that these freshly-created electrons are continuously added to the plasma at a constant rate. As the electrons emit synchrotron radiation they naturally loose energy. The rate of energy losses is however higher for the higher-energy electrons, and this leads to a ‘break’ in the emission spectrum of the plasma. The resulting spectrum can be *approximately* described by: $$\begin{aligned} F_{\nu}(\nu)= \left\{ \begin{array}{lr} \displaystyle F_{\nu}^{0} \cdot \left(\frac{\nu}{1\,{\mathrm{GHz}}}\right)^\alpha & \nu \leq \nu_{\rm br} \\ \displaystyle F_{\nu}^{0} \cdot \left(\frac{\nu}{1\,{\mathrm{GHz}}}\right)^{\alpha} \left(\frac{\nu}{\nu_{\rm br}}\right)^{-1/2} & \nu > \nu_{\rm br}. \end{array} \right.\end{aligned}$$ The one new parameter in this model is the frequency of break in the spectrum, $\nu_{\rm br}$. If this break frequency is known, it can be used to estimate the age of the source given the magnetic field strength within it and vice-versa. Like the flux density scaling parameter $F_{\nu}^{0}$, the frequency of the spectral break is a scale parameter and is best parametrised as the logarithm of the actual frequency: $\log_{10}(\nu_{\rm br})$. The models just described here are continuous functions of frequency while actual measurements are integrated over some finite bandwidth. The effect of averaging over a bandwidth can be taken into account analytically for the simple models described here or more generally by numerically integrating the average flux over the band. In this work however, I however make the approximation that the bandwidths are small and that it can be assumed without too great an error that each measurement is monochromatic. The above two models describe the intrinsic emission spectrum from relativistic electrons. Depending on the geometry of the source and total density of electrons, the electrons may absorb as well as emit radiation. This self-absorption process usually becomes significant at a low enough frequencies, leading to a turn-over in the spectrum. The self absorption factor, $A_{\rm s}$ [e.g., @1970ranp.book.....P] can be parametrised by the frequency at which emission peaks, $\nu_{\rm pk}$: $$\begin{aligned} x&= \nu/\nu_{\rm pk}\\ A_{\rm s}&= x^{-\alpha+5/2} \left[1- \exp\left(1-x^{\alpha-5/2}\right)\right]\end{aligned}$$ Again, this quantity is an additional parameter of models and should be parametrised in terms of its logarithm: $\log_{10}( \nu_{\rm pk})$. The possibility of Synchrotron-Self Absorption (SSA) leads to a total of four possible models in this simple example: power-law, power-law that also exhibits self-absorption, continuous-injection and continuous-injection with self-absorption. As described above, these models need to be combined with priors on their parameters for them to be useful. For this illustration I use simple priors with following two properties which ease analysis: 1. The prior factors into independent functions of one parameter only; e.g., for the most complex model: $$\begin{gathered} \pi(\{\log_{10}\left(F_{\nu}^{0}\right), \alpha, \log_{10}(\nu_{\rm br}), \log_{10}(\nu_{\rm pk})\})= \\ \pi\left[\log_{10}\left(F_{\nu}^{0}\right)\right]\cdot\pi(\alpha)\cdot \pi\left[\log_{10}(\nu_{\rm br})\right]\cdot \pi\left[\log_{10}(\nu_{\rm pk})\right]\end{gathered}$$ 2. Each prior is constant within a given range and zero outside: $$\pi(x)=\begin{cases} \frac{1}{x_{\rm high}-x_{\rm low}} & x_{\rm low}<x<x_{\rm high}\\ 0 & \text{otherwise} \end{cases}$$ For all of the examples presented below I used the same set of priors as follows: $$\begin{aligned} \log_{10}\left(F_{\nu}^{0}\right)_{\rm low}&=-1 & \log_{10}\left(F_{\nu}^{0}\right)_{\rm high}&=1 \label{eq:priors-bg} \\ \alpha_{\rm low}&=-3/2 & \alpha_{\rm high}=1/2\\ \log_{10}(\nu_{\rm br})_{\rm low} &= 8.5& \log_{10}(\nu_{\rm br})_{\rm high} &= 10.0\\ \log_{10}(\nu_{\rm pk})_{\rm low} &=7.5 & \log_{10}(\nu_{\rm pk})_{\rm low} &=8.5 . \label{eq:priors-end}\end{aligned}$$ Likelihood {#sec:likelihood} ---------- In this section I describe the calculation of the likelihood of the observed data given a model and its parameters. This part of the calculation depends on a good understanding of the observations and how they were processed so that realistic error estimates on measured flux densities can be assigned and any selection effects taken into account. In the simplest cases, it is possible to make two simplifying assumptions: 1. Measurements at each frequency are independent, and the likelihood therefore factors into a product of functions of flux densities at one frequency only 2. Errors on each measurement are normally distributed, and the likelihood therefore takes the standard Gaussian form If these assumptions are made, than the joint likelihood is simply: $$\log{L(\theta)} = - \frac{1}{2}\sum_{i} \left\{ \left[ \frac{D_{i}-F_{\nu}(\nu_i)} {\sigma_{i}} \right]^{2} + \log\left(2\pi \sigma_i^2\right) \right\}$$ where $D_i$ is the flux density observed at a frequency of $\nu_i$ and $\sigma_i$ is the estimate of the standard error of this measurement. The second term in the sum is the normalisation constant (i.e., it does not depend on the observed data or the model) which must be included for correct calculation of the evidence value. The above assumptions are the same as often made in analysis of radio spectra using more conventional techniques. One of the advantages of the present approach however is that these assumptions do *not* need to be made. For example, in multi-frequency surveys of relatively faint sources it is often the case that sources are selected at the lower frequency (where the survey speed is typically higher) and only the detected sources are followed up at the higher frequency. This leads to a cross-dependence in the likelihood function between the measurements at the lower and higher frequencies. Nested sampling {#sec:nested-sampling} --------------- The method I use for the calculation of the evidence, $Z$ and the posterior distribution of the model parameters, $p(\theta| D, H)$, is the *nested sampling* algorithm described by [@Skilling2006]. The key advantage of this algorithm for this application is that it allows efficient and accurate calculation of the evidence value even in the presence of relatively complex likelihood distributions. The starting set of points used by the sampler is initialised by randomly and uniformly distributing the points in the space allowed by the priors. Since all priors are flat this is sufficient to ensure a representative starting distribution. The sampling then proceeds by finding the point in the current set with the smallest likelihood and replacing it. The replacement needs to be selected uniformly from the prior space with the constraint that the likelihood of the replacement point is greater than the likelihood of the point it replaces. This is implemented by: 1. Selecting a point at random in the current set 2. Using a Markov chain with to find a new point subject to the likelihood constraint The step size and directions used in the Markov chain are determined by spread of the points in the current set. Specifically, a principal component analysis is carried out on the current set and the steps in the Markov chain alternate between each of the eigenvectors, which are scaled by 0.1 before being used. Normally 100 steps are made with the Markov chain before the new point is added to the current set. The nested sampling procedure is terminated when the requested number of samples has been made or when the Markov chain procedure fails to find a point with better likelihood than the worst point in the set. Because of this latter mechanism for termination of the sampler, the number of samples made should be inspected before further analysis to ensure the sampling proceeded far enough to provide accurate results. This procedure generally appears to work well but it should be noted that for multi-modal distributions with widely separated peaks it will generate step sizes which are too large and with too high a likelihood of leading to a lower-likelihood region. This would manifest itself as early termination of the nested sampling algorithm because the Markov chain constrained sampling does not produce a sample with a higher likelihood than the worst point in the live set. I note again that the models I have considered so far *are* (generally) multi-modal, but the modes are sufficiently close that the present scheme works well. If the models of radio spectra and the likelihood function are extended further to more complex problems, this potential problem presented by multi-modal distributions should be kept in mind. Significantly more advanced implementations of nested sampling algorithm are described by [@2008MNRAS.384..449F] and [@2009MNRAS.398.1601F]. I believe however that the present implementation is the only that is publicly available under the GPL and callable from C++ and Python. Presenting the results ---------------------- The evidence values calculated are simple numbers and can be tabulated for various combinations of models and priors. In the examples below only one set are priors is used, so only the one evidence value is given for each model. The relative magnitudes of the evidence allow objective model selection, with the higher evidence value implying the preferred model. For each model, the nested-sampling algorithm also provides the full joint posterior distribution of the parameters. These are conventionally visualised by marginalising to get to the marginal distributions of single parameters and of pairs of parameters. These can be then plotted as one-dimensional histograms and two-dimensional colour plots respectively. It is clearly also useful to visualise how well each combination of model and priors fits the observed data. For example, this often provides crucial information about how the models might need to be modified to explain the observations. Since the result of the nested-sampling analysis is a distribution of model parameters, there are various choices as to how the resulting models may be visualised. The simplest choice is to plot the model with the maximum-likelihood parameter set for each model. Good maximum likelihood estimate can obtained simply by taking the highest-likelihood point in the live set of the sampler at the completion of the algorithm. The approach of plotting the model with maximum-likelihood parameters however fails to capture the variation away from the maximum-likelihood solution that is the main reason for the Bayesian analysis approach in the first place. An alternative is to compute the probability distribution of the flux density as a function of frequency: $$\begin{aligned} p(F_\nu| H,D)= \int {\rm d}\theta \cdot F(\nu; \theta, H) p(\theta| H,D)\end{aligned}$$ where as before $H$ is the hypothesis, i.e., the model for emission and the priors on the model parameters; $p(\theta| H,D)$ is the posterior distribution which is an output of the nested sampling; and the integration over $\theta$ is of course over all its dimensions. This probability distribution can then be plotted by assigning frequency to horizontal position, flux density to vertical position and the probability to colour-scale on the plot. Such diagrams are sometimes called fan-diagrams in economics, and I adopt that terminology here. This visualisation approach is shown in Figures \[fig:ngc628-fan\]–\[fig:ngc7331-fan\]. Implementation ============== ------------------------------------------------------------ --------------------------------------------------------------- Power-law (Evidence: 185.874) Power-law + SSA (Evidence: 76.3034) {width="0.49\linewidth"} {width="0.49\linewidth"} CI (Evidence: 83.4132) CI + SSA (Evidence: 27.65) {width="0.49\linewidth"} {width="0.49\linewidth"} ------------------------------------------------------------ --------------------------------------------------------------- ------------------------------------------------------------ --------------------------------------------------------------- Power-law (Evidence: $9.7\times 10^{-15}$) Power-law + SSA (Evidence: $3.9\times10^{-15}$) {width="0.49\linewidth"} {width="0.49\linewidth"} CI (Evidence: $1.3\times10^{-09}$) CI + SSA (Evidence: $2.6\times10^{-10}$) {width="0.49\linewidth"} {width="0.49\linewidth"} ------------------------------------------------------------ --------------------------------------------------------------- ------------------------------------------------------------ --------------------------------------------------------------- Power-law (Evidence: $2.9\times10^{-17}$) Power-law + SSA (Evidence: 0.5) {width="0.49\linewidth"} {width="0.49\linewidth"} CI (Evidence: 70) CI + SSA (Evidence: $5.0\times10^{5}$) {width="0.49\linewidth"} {width="0.49\linewidth"} ------------------------------------------------------------ --------------------------------------------------------------- ------------------------------------------------------------------- ------------------------------------------------------------------ Overall flux Spectral index (at frequencies below the break) {width="0.49\linewidth"} {width="0.49\linewidth"} Frequency of the spectral break Frequency at which self-absorption becomes important {width="0.49\linewidth"} {width="0.49\linewidth"} ------------------------------------------------------------------- ------------------------------------------------------------------ The major computational parts of the algorithms described here are implemented in the C++ programming language in two separate libraries: - [bnmin1]{} This is a minimisation and inference library that contains the nested-sampling algorithms and the supporting functions - [radiospec]{} This is the library specialised for this application, and contains the models of radio spectra and descriptions of observations and their errors The first of these, [bnmin1]{}, is a general purpose minimisation/inference library that may be used in a variety of application. So far I have used this library for phase retrieval holography (, ), which is the application for which I first started to develop the library; and for phase correction algorithms for ALMA (@2009arXiv0903.1179N and @2009arXiv0907.4472N). The library has been available to the public under the GNU General Public License for a number of years and although it has been downloaded occasionally I am not aware of other public work using it. The majority of the functionality described in this paper has only just been added in the release of the library that accompanies this paper. The remaining functionality available in the library includes Levenberg-Marquardt fitting and Markov Chain Monte Carlo (MCMC). The C++ parts of both of the libraries are relatively self-contained with only two external dependencies: the Boost C++ libraries [@BoostLibs], and the GNU Scientific Library [@GSL]. The build system of [bnmin1]{} is based on the standard AutoTools chain, while [radiospec]{} is built using the SCons system. The top level commands, such as which algorithms to use, to enter the observed data, to control adjustable parameters etc are implemented in the Python programming language. The interface between the C++ libraries and Python is generated automatically using the standard SWIG[^1] package described by [@Beazley2003]. This architecture allows easy *interactive* use of the library. The supporting Python script for this application is available as part of [radiospec]{}. All of the code is available for public download under the GNU General Public License at <http://www.mrao.cam.ac.uk/~bn204/galevol/speca/index.html>. I should however point out some caveats: - The packages need to be compiled from source and this typically requires some experience. The instructions posted on the above web-pages will be updated over time to explain how to tackle common problems with compilation that are reported to me - In order to analyse new observations using existing models, you will need to program in Python. There is no graphical or command shell user interface to this software - In order to create new models, you will need to program in C++ Included with the code is a script which reproduces the illustrative examples given later in this paper. Implementation of radio spectra models -------------------------------------- The models of radio spectra are implemented as a polymorphic class hierarchy in C++ in the [radiospec]{} package. The base class defining the interface is [RadioModel]{} which in turn inherits from the [Model]{} class from the BNMin1 package. There are however only two relevant “virtual” functions in the interface: 1. [double fnu(double nu) const]{} which computes the flux density at specified frequency [nu]{} 2. [void AddParams(std::vector&lt; Minim::DParamCtr &gt; &pars)]{} which defines the parameters of the model by adding them to the vector of parameter definitions [pars]{} Any new models with which [radiospec]{} is extend must properly define these two functions to be useful. The reason for adopting the polymorphic inheritance approach rather than the more run-time efficient template approach is that the polymorphic inheritance can be easily accessed and manipulated from Python, allowing for example dynamic composition of several models into a new, more complex, single model. Implementation of the likelihood function ----------------------------------------- The likelihood function for the examples shown here is implemented in the [NormalLkl]{} class. This function combines an user specified model (as a pointer to a type [RadioModel]{} object), an user specified set of observations (as an object of [RadioObs]{} class) and the usual Gaussian probability formula to compute the likelihood. If a non-Gaussian likelihood function is required, it should be implemented in a similar way to [NormalLkl]{} class but it should *not* be derived from it. Implementation of the priors ---------------------------- In the current version of the code, only *flat* priors which are separable functions of single model parameter are supported. Additionally, every model parameter must have *some* prior defined for it because this information is used to initialise the nested sampling algorithm. Consequently, the user effectively has to define a parameter-space prior ‘box’ for each problem. The prior ‘box’ is specified in the Python layer as a dictionary of parameter names that map to a tuple specifying the lower and upper bound of the parameter. An example is given in the file [ methodex.py]{}. Plotting -------- The plotting of output is implemented using the PyX framework for Python (@PyXWeb). This library is able to directly write PostScript (PS) and Portable Document Format (PDF) files and to run LaTeX to generate properly type-set labels for the graphs. The main benefits of this library for this application are that it can be used directly from Python and that the output is of high-quality both visually and in terms of the efficiency and readability of the output PostScript code. The routines that build on top of PyX to make the plots shown in this paper are largely contained in the package [PyHLP]{} which is distributed separately from the other packages and can be downloaded at <http://www.mrao.cam.ac.uk/~bn204/technotes/pyhlp.html>. Results ======= The data I used to illustrate this approach are taken from . They are global spectra of three late-type galaxies with measurements at 5 to 7 frequencies in the range 50MHz–5GHz. I have taken the measurements and measurement errors listed by without any further edits and without referring to the original sources for the archival data that they used. The spectra of all three galaxies were processed as described above, including the use of the priors listed in Equations \[eq:priors-bg\]–\[eq:priors-end\]. The results are presented as the fan-plot for each model, together with the evidence value in Figures \[fig:ngc628-fan\]– \[fig:ngc7331-fan\]. For NGC 7331, in Figure \[fig:ngc7331-margin\] I also plotted the marginalised distributions of the model parameters. NGC 628 ------- The radio spectrum of the galaxy is relatively featureless and can be by-eye seen to be reasonably close to a pure power-law. The analysis presented here also reaches this expected conclusion. The fan-diagrams of all four models for this object are shown in Figure \[fig:ngc628-fan\]. The combination of model and prior with the highest evidence value is the power-law case, shown in the upper-left panel of this figure. This means that this simple model is the best model for the true underlying process given the observations out of the four models considered here. The continuous injection model fan-diagram in the lower-left panel shows that it can explain the data in two ways: either the break is at a frequency higher than the highest observations, or, the break is at around 300MHz and the lowest-frequency data point is predicted with a significant error. Therefore there in principle remains a possibility that this galaxy has a highly aged electron population with an intrinsic injection index of about $\alpha \sim -0.2$. This is however unlikely given the lower evidence value of this model compared to the power law. Finally it can be seen that the absorbed models on the right hand side of Figure \[fig:ngc628-fan\], i.e., the upper-right and lower-right plots, do not describe the data as well as the non-absorbed models. This is part because the minimum frequency of the turnover was set at 30MHz by priors. NGC 3627 -------- The fan-diagrams for this galaxy are shown in Figure \[fig:ngc3627-fan\]. The model with the highest evidence is the continuous-injection model without absorption at low frequencies. The lower evidence of the CI+SSA model indicates that with these data, there is no evidence for absorption in this source. NGC 7331 -------- The fan-diagrams for this galaxy are shown in Figure \[fig:ngc7331-fan\]. What is noticeable for this galaxy is that the most complex model, the continuous-injection with synchrotron self-absorption model (CI+SSA), has the highest evidence value by several orders of magnitude. This high evidence value implies that this complex model must be preferred given the available data. The fan-diagram of the CI+SSA model, in the lower-right panel of the figure, shows that it reproduces well the observed features of the spectrum while all of the others fail to reproduce one or more features. The marginalised distributions of the parameters of the CI+SSA model are shown in Figure \[fig:ngc7331-margin\]. It shows that the distributions of all of the parameters are well constrained, although at least the spectral index and the break frequency show non-Gaussian distributions. This should be interpreted to mean that care must be taken when using a simple single-value error estimate for these parameters in further calculations. Summary ======= Fitting of radio spectra of galaxies is a topic that is computationally relatively simple, since most models are either analytic or contain simple one-dimensional integrals. A proper statistical analysis is however not entirely straightforward for a combination of reasons: 1. There are few measurement points (typically 3–10) 2. Errors are often non-Gaussian (e.g., because they are dominated by calibration errors) and are sometimes not well quantified 3. There are many different models that could be tried An attractive way to tackle this problem is using Bayesian analysis because it provides: 1. A rigorous theoretical framework 2. Objective model selection 3. A natural way to introduce physical constraints on model parameters through priors 4. Full probability distributions for each model parameter 5. A complete picture of any degeneracies in the model parameters In this paper I have described a publicly available computer code which implements such Bayesian analysis of spectra using the nested sampling algorithm developed by [@Skilling2006]. This algorithm allows efficient calculation of all of the outputs of Bayesian analysis including the evidence value and the full joint distribution of all parameters. This means analysis is computed quickly and without the need to guide it ‘by hand’, by for example carefully choosing starting positions. The described code also allows visualisation of how well the each model explains the data using fan-diagrams. I believe this little-used approach to visualisation allows a good and intuitive understanding of implications of a particular distribution of model parameters. The code described is already being used for several of projects in radio astronomy but I expect it could useful for quite a broad range of applications. Such application to new areas will no doubt lead to discoveries of errors and shortcoming in the code and I would very much appreciate to be notified these at [mailto:[email protected]](mailto:[email protected]). If you obtain useful results from the code without finding any errors or shortcomings, than of course I would be even more happy to hear from you, at the same email address, and can place a link to your paper on the web-pages. Publication information {#publication-information .unnumbered} ======================= As an experiment, I will be publishing this paper on arXiv only. This is in part because future revision of this paper is likely to be necessary, once the code is used more extensively both by us and hopefully by the general community. In lieu of the normal referring process, I would be happy to hear from readers on any aspect of the paper and incorporate all corrections and (at least constructive) suggestions in any future versions. All of these will be credited unless requested otherwise. Alternatively if you have more extensive comments I suggest you use the arXiv trackback mechanism. \[lastpage\] [^1]: <http://www.swig.org/>

ArXiv

--- abstract: 'We investigate the existence of Taub-NUT/bolt solutions in Gauss-Bonnet gravity and obtain the general form of these solutions in $d$ dimensions. We find that for all non-extremal NUT solutions of Einstein gravity having no curvature singularity at $r=N$, there exist NUT solutions in Gauss-Bonnet gravity that contain these solutions in the limit that the Gauss-Bonnet parameter $\alpha$ goes to zero. Furthermore there are no NUT solutions in Gauss-Bonnet gravity that yield non-extremal NUT solutions to Einstein gravity having a curvature singularity at $r=N$ in the limit $\alpha \to 0$. Indeed, we have non-extreme NUT solutions in $2+2k$ dimensions with non-trivial fibration only when the $2k$-dimensional base space is chosen to be $\mathbb{CP}^{2k}$. We also find that the Gauss-Bonnet gravity has extremal NUT solutions whenever the base space is a product of 2-torii with at most a $2$-dimensional factor space of positive curvature. Indeed, when the base space has at most one positively curved two dimensional space as one of its factor spaces, then Gauss-Bonnet gravity admits extreme NUT solutions, even though there a curvature singularity exists at $r=N$. We also find that one can have bolt solutions in Gauss-Bonnet gravity with any base space with factor spaces of zero or positive constant curvature. The only case for which one does not have bolt solutions is in the absence of a cosmological term with zero curvature base space.' --- **NUT-Charged Black Holes in Gauss-Bonnet Gravity**\ M. H. Dehghani$^{1,2,3}$ and R. B. Mann$^{2,3}$ *1. Physics Department and Biruni Observatory, College of Sciences, Shiraz University, Shiraz 71454, Iran\ 2. Department of Physics, University of Waterloo, 200 University Avenue West, Waterloo, Ontario, Canada, N2L 3G1\ 3. Perimeter Institute for Theoretical Physics, 35 Caroline St. N., Waterloo, Ont. Canada* Introduction ============ Four-dimensional Taub-NUT and Taub-bolt-AdS solutions of Einstein gravity play a central role in the construction of diverse and interesting M-theory configurations. Indeed, the 4-dimensional Taub-NUT-AdS solution provided the first test for the AdS/CFT correspondence in spacetimes that are only locally asymptotically AdS [@Chamblin; @Hawking; @MannMisner], and the Taub-NUT metric is central to the supergravity realization of the $D6$ brane of type IIA string theory [@mbrane]. It is therefore natural to suppose that the generalization of these solutions to the case of Lovelock gravity, which is the low energy limit of supergravity, might provide us with a window on some interesting new corners of M-theory moduli space. The original four-dimensional solution [@Taub; @NUT] is only locally asymptotic flat. The spacetime has as a boundary at infinity a twisted $S^{1} $ bundle over $S^{2}$, instead of simply being $S^{1}\times S^{2}$. There are known extensions of the Taub-NUT solutions to the case when a cosmological constant is present. In this case the asymptotic structure is only locally de Sitter (for positive cosmological constant) or anti-de Sitter (for negative cosmological constant) and the solutions are referred to as Taub-NUT-(A)dS metrics. In general, the Killing vector that corresponds to the coordinate that parameterizes the fibre $S^{1}$ can have a zero-dimensional fixed point set (called a NUT solution) or a two-dimensional fixed point set (referred to as a ‘bolt’ solution). Generalizations to higher dimensions follow closely the four-dimensional case [@Bais; @Page; @Akbar; @Robinson; @Awad; @Lorenzo; @Mann1; @Mann2]. In this paper we consider Taub-NUT metrics in second order Lovelock gravity (referred to as Gauss-Bonnet gravity), which is a higher-dimensional generalization of Einstein gravity. In higher dimensions it is possible to use other consistent theories of gravity with actions more general than that of the Einstein-Hilbert action. Such an action may be written, for example, through the use of string theory. The effect of string theory on classical gravitational physics is usually that of modifying the low energy effective action that describes gravity at the classical level [@Wit1]. This effective action consists of the Einstein-Hilbert action plus curvature-squared terms and higher powers as well, and in general gives rise to fourth-order field equations containing ghosts. However if the effective action contains the higher powers of curvature in particular combinations, then only second-order field equations are produced and consequently no ghosts arise [@Zw]. The effective action obtained by this argument is precisely of the form proposed by Lovelock [@Lov]. Until now whether or not higher derivative gravity admits solutions of the Taub-NUT/bolt form has been an open question. Due to the nonlinearity of the field equations, it is very difficult to find nontrivial exact analytic solutions of Einstein’s equations modified with higher curvature terms. In most cases, one has to adopt some approximation methods or find solutions numerically. However, exact static spherically symmetric black hole solutions of second and third order Lovelock gravity have been obtained in Refs. [@Des; @Whe; @Deh1], and of Einstein-Maxwell-Gauss-Bonnet model in Ref. [@Wil]. Black hole solutions with nontrivial topology and their thermodynamics in this theory have been also studied [@Cai]. All of these solutions in Gauss-Bonnet gravity are static. Recently two new classes of rotating solutions of second order Lovelock gravity have been introduced and their thermodynamics have been investigated [@Deh2]. Here we investigate the existence of Taub-NUT and Taub-bolt solutions of Gauss-Bonnet gravity. We find that exact solutions exist, but that Lovelock gravity introduces some features not present in higher-dimensional Einstein-Hilbert gravity. The form of the metric function is sensitive to the base space over which the circle is fibred. Furthermore, we find that pure non-extreme NUT solutions only exist if the base space has a single factor of maximal dimensionality. We conjecture that this is a general property of Gauss-Bonnet gravity in any dimension. The outline of our paper is as follows. We give a brief review of the field equations of second order Lovelock gravity in Sec. \[Fiel\]. In Sec. \[6d\], we obtain all possible Taub-NUT/bolt solutions of Gauss-Bonnet gravity in six dimensions. The structure of these solutions suggests two conjectures, which we posit and then check in the remainder of the paper. Then, in sections \[8d\] and \[10d\], we present all kind Taub-NUT/bolt solutions of Gauss-Bonnet gravity in eight and ten dimensions and check the conjectures for them. In Sec \[dd\], we extend our study to the $d$-dimensional case, and find the general expressions for the Taub-NUT/bolt solutions. We finish our paper with some concluding remarks. Field Equations\[Fiel\] ======================= The most fundamental assumption in standard general relativity is the requirement that the field equations be generally covariant and contain at most second order derivatives of the metric. Based on this principle, the most general classical theory of gravitation in $d$ dimensions is Lovelock gravity. The gravitational action of this theory can be written as $$I_{G}=\int d^{d}x\sqrt{-g}\sum_{k=0}^{[d/2]}\alpha _{k}\mathcal{L}_{k} \label{Lov1}$$ where $[z]$ denotes the integer part of $z$, $\alpha _{k}$ is an arbitrary constant and $\mathcal{L}_{k}$ is the Euler density of a $2k$-dimensional manifold, $$\mathcal{L}_{k}=\frac{1}{2^{k}}\delta _{\rho _{1}\sigma _{1}\cdots \rho _{k}\sigma _{k}}^{\mu _{1}\nu _{1}\cdots \mu _{k}\nu _{k}}R_{\mu _{1}\nu _{1}}^{\phantom{\mu_1\nu_1}{\rho_1\sigma_1}}\cdots R_{\mu _{k}\nu _{k}}^{\phantom{\mu_k \nu_k}{\rho_k \sigma_k}} \label{Lov2}$$ In Eq. (\[Lov2\]) $\delta _{\rho _{1}\sigma _{1}\cdots \rho _{k}\sigma _{k}}^{\mu _{1}\nu _{1}\cdots \mu _{k}\nu _{k}}$ is the generalized totally anti-symmetric Kronecker delta and $R_{\mu \nu }^{\phantom{\mu\nu}{\rho\sigma}}$ is the Riemann tensor. We note that in $d$ dimensions, all terms for which $k>[d/2]$ are identically equal to zero, and the term $k=d/2$ is a topological term. Consequently only terms for which $k<d/2$ contribute to the field equations. Here we restrict ourselves to the first three terms of Lovelock gravity, which is known as Gauss-Bonnet gravity. In this case the action is $$I_{G}=\frac{1}{2}\int_{\mathcal{M}}dx^{d}\sqrt{-g}[-2\Lambda +R+\alpha (R_{\mu \nu \gamma \delta }R^{\mu \nu \gamma \delta }-4R_{\mu \nu }R^{\mu \nu }+R^{2})] \label{Ig}$$ where $\Lambda $ is the cosmological constant, $R$, $R_{\mu \nu \rho \sigma } $, and $R_{\mu \nu }$ are the Ricci scalar and Riemann and Ricci tensors of the spacetime, and $\alpha $ is the Gauss-Bonnet coefficient with dimension $(\mathrm{length})^{2}$. Since it is positive in heterotic string theory [@Des] we shall restrict ourselves to the case $\alpha >0$. The first term is the cosmological term, the second term is just the Einstein term, and the third term is the second order Lovelock (Gauss-Bonnet) term. From a geometric point of view, the combination of these terms in five-dimensional spacetimes is the most general Lagrangian producing second order field equations, analogous to the situation in four-dimensional gravity for which the Einstein-Hilbert action is the most general Lagrangian producing second order field equations.** ** Varying the action with respect to the metric tensor $g_{\mu \nu }$, the vacuum field equations are $$\begin{aligned} R_{\mu \nu }&-&\frac{1}{2}Rg_{\mu \nu }+\Lambda g_{\mu \nu }-\alpha \{\frac{1}{2}g_{\mu \nu }(R_{\kappa \lambda \rho \sigma }R^{\kappa \lambda \rho \sigma }-4R_{\rho \sigma }R^{\rho \sigma }+R^{2}) \notag \\ &-&2RR_{\mu \nu }+4R_{\mu \lambda }R_{\text{ \ }\nu }^{\lambda }+4R^{\rho \sigma }R_{\mu \rho \nu \sigma }-2R_{\mu }^{\ \rho \sigma \lambda }R_{\nu \rho \sigma \lambda }\}=0 \label{Geq}\end{aligned}$$ Equation (\[Geq\]) does not contain derivatives of the curvature, and therefore derivatives of the metric higher than two do not appear. We seek Taub-NUT solutions of the field equations (\[Geq\]). In constructing these metrics the idea is to regard the Taub-NUT space-time as a $U(1)$ fibration over a $2k$-dimensional base space endowed with an Einstein-K$\ddot{a}$hler metric $g_{\mathcal{B}}$. Then the Euclidean section of the $(2k+2)$-dimensional Taub-NUT spacetime can be written as: $$ds^{2}=F(r)(d\tau +NA)^{2}+F^{-1}(r)dr^{2}+(r^{2}-N^{2})g_{\mathcal{B}} \label{TN}$$ where $\tau $ is the coordinate on the fibre $S^{1}$ and $A$ has a curvature $F=dA$, which is proportional to some covariantly constant 2-form. Here $N$ is the NUT charge and $F(r)$ is a function of $r$. The solution will describe a ‘NUT’ if the fixed point set of the U(1) isometry $\partial/\partial\tau$ (i.e. the points where $F(r)=0$) is less than $2k$-dimensional and a ‘bolt’ if the fixed point set is $2k$-dimensional. Six-dimensional Solutions\[6d\] =============================== In this section we study the six-dimensional Taub-NUT/bolt solutions (\[TN\]) of Gauss-Bonnet gravity. We find that the function $F(r)$ for all the possible choices of the base space $\mathcal{B}$ can be written in the form $$\begin{aligned} F(r) &=&\frac{(r^{2}-N^{2})^{2}}{12\alpha (r^{2}+N^{2})}\left( 1+\frac{p\alpha }{(r^{2}-N^{2})}-\sqrt{B(r)+C(r)}\right) \notag \\ B(r) &=&1+\frac{4p\alpha N^{2}(r^{4}+6r^{2}N^{2}+N^{4})+12\alpha mr(r^{2}+N^{2})}{(r^{2}-N^{2})^{4}} \notag \\ &&+\frac{12\alpha \Lambda (r^{2}+N^{2})}{5(r^{2}-N^{2})^{4}}(r^{6}-5N^{2}r^{4}+15N^{4}r^{2}+5N^{6}) \label{F6}\end{aligned}$$ where $p$ is the sum of the dimensions of the curved factor spaces of $\mathcal{B}$, and the function $C(r)$ depends on the choice of the base space $\mathcal{B}$. We first study the solutions for which all factor spaces of $\mathcal{B}$ are curved. The first possibility is that the base space is $\mathcal{B}=\mathbb{CP}^{2}$, where $A$ and the $\mathbb{CP}^{2}$ metric are: $$\begin{aligned} A_{2} &=&6\sin ^{2}\xi _{2}(d\psi _{2}+\sin ^{2}\xi _{1}d\psi _{1}) \label{A2} \\ d{\Sigma _{2}}^{2} &=&6\{d{\xi _{2}}^{2}+\sin ^{2}\xi _{2}\cos ^{2}\xi _{2}(d\psi _{2}+\sin ^{2}\xi _{1}d\psi _{1})^{2} \notag \\ &&\ \ \ \ +sin^{2}\xi _{2}({d\xi _{1}}^{2}+\sin ^{2}\xi _{1}\cos ^{2}\xi _{1}{d\psi _{1}}^{2})\} \label{CP2}\end{aligned}$$ The function $C(r)$ for $\mathcal{B}=\mathbb{CP}^{2}$ is: $$C_{\mathbb{CP}^{2}}=-\frac{16\alpha ^{2}(r^{4}+6r^{2}N^{2}+N^{4})}{(r^{2}-N^{2})^{4}} \label{ac6CP}$$ The second possibility is $\mathcal{B}=S^{2}\times S^{2}$, where $S^{2}$ is the 2-sphere with $d\Omega ^{2}=d\theta ^{2}+\sin ^{2}\theta d\phi ^{2}$, and the one-form $A$ is $$A=2\cos \theta _{1}d\phi _{1}+2\cos \theta _{2}d\phi _{2} \label{6SS}$$ In this case $C(r)$ is $$C_{S^{2}\times S^{2}}=-\frac{32\alpha ^{2}(r^{4}+4r^{2}N^{2}+N^{4})}{(r^{2}-N^{2})^{4}} \label{c6SS}$$ Note that the asymptotic behavior of these solutions for positive $\alpha $is locally flat when $\Lambda $ vanishes, locally dS for $\Lambda >0$ and locally AdS for $\Lambda <0$ provided $\left| \Lambda \right| <5/(12\alpha ) $. We note that the function $F(r)$ given in (\[F6\]) has the same form in the limit $\alpha \rightarrow 0$, and reduces to $F(r)$ given in [@Lorenzo]. Factor spaces of zero curvature will be considered at the end of this section. If the base space has all factor spaces curved, then the remaining possibilities are $\mathcal{B}=S^{2}\times H^{2}$ or $\mathcal{B}=H^{2}\times H^{2}$, where $H^{2}$ is the two-dimensional hyperboloid with metric $d\Xi ^{2}=$ $d\theta ^{2}+\sinh ^{2}\theta d\phi ^{2}$. However we find that the function $F(r)$ is not real for all values of $r$ in the range $0\leq r\leq \infty $ for positive values of $\alpha $. The only case for which a black hole solution can be obtained is when $\alpha $ and $\Lambda $ are negative. As we mentioned earlier, we are interested in positive values of $\alpha $, and so we don’t consider them here. Taub-NUT Solutions ------------------ The solutions given in Eq. (\[F6\]) describe NUT solutions, if (i) $F(r=N)=0$ and (ii) $F^{\prime }(r=N)=1/(3N)$. The first condition comes from the fact that all the extra dimensions should collapse to zero at the fixed point set of $\partial/\partial\tau $, and the second one ensures that there is no conical singularity with a smoothly closed fiber at $r=N$. Using these conditions, one finds that Gauss-Bonnet gravity in six dimensions admits NUT solutions with a $\mathbb{CP}^{2}$ base space when the mass parameter is fixed to be $$m_{n}=-\frac{16}{15}N(3\Lambda N^{4}+5N^{2}-5\alpha ) \label{mn6cp}$$ Computation of the Kretschmann scalar at $r=N$ for the solutions given in the last section shows that the spacetime with $\mathcal{B}=S^{2}\times S^{2} $ has a curvature singularity at $r=N$ in Einstein gravity, while the spacetime with $\mathcal{B}=\mathbb{CP}^{2}$ has no curvature singularity at $r=N$. Thus, we conjecture that “*For all non-extremal NUT solutions of Einstein gravity having no curvature singularity at* $r=N$*, there exist NUT solutions in Gauss-Bonnet gravity that contain these solutions in the limit that the parameter* $\alpha $* vanishes. Furthermore there are no NUT solutions in Gauss-Bonnet gravity that yield non-extremal NUT solutions to Einstein gravity having a curvature singularity at* $r=N$* in the limit* $\alpha \rightarrow 0$* .*” Indeed, we have non-extreme NUT solutions in $2+2k$ dimensions with non-trivial fibration when the $2k$-dimensional base space is chosen to be $\mathbb{CP}^{2}$. We will test this conjecture throughout the rest of the paper, and consider the case of extremal NUT solutions at the end of this section. Taub-Bolt Solutions ------------------- The conditions for having a regular bolt solution are (i) $F(r=r_{b})=0$ and (ii)$\ F^{\prime }(r_{b})=1/(3N)$ with $r_{b}>N$. Condition (ii), which again follows from the fact that we want to avoid a conical singularity at the bolt, together with the fact that the period of $\tau $ will still be $12\pi N$ , gives the following equation for $r_{b}>N$ with the base space $\mathbb{CP}^{2}$ $$3N\Lambda {r_{b}}^{3}+(2+3\Lambda N^{2}){r_{b}}^{2}-N(4+3\Lambda N^{2})r_{b}-3\Lambda N^{4}-6N^{2}+8\alpha =0 \label{bmat6cp}$$ which has at least one real solution. This real solution for $N<\sqrt{\alpha }$ yields $r_{b}>N$, which is a bolt solution, while for $N\geq \sqrt{\alpha }$, there is no bolt solution – only the NUT solution can satisfy the regularity conditions. As $\Lambda $ goes to zero this real solution for $r_{b}$ diverges. However setting $\Lambda =0$ we obtain the asymptotically locally flat case, and Eq. (\[bmat6cp\]) becomes a quadratic equation with the following two roots: $$r_{b}=N\pm 2\sqrt{(N^{2}-\alpha )} \label{rb6}$$ Note that we have only one possible bolt solution (provided $N>\sqrt{\alpha}$), since the smaller root is less than $N$. For the case of $\mathcal{B}=S^{2}\times S^{2}$, $r_{b}$ can be found by solving the following equation: $$3N\Lambda {r_{b}}^{4}+2{r_{b}}^{3}-6N(\Lambda N^{2}+1){r_{b}}^{2}-2(N^{2}-4\alpha )r_{b}+3\Lambda N^{5}+6N^{3}-12\alpha N=0 \label{bmat6s}$$ where the condition for having bolt solution(s) is that $N\leq N_{\mathrm{\max }}$, where $N_{\mathrm{\max }}$ is the smaller root of the following equation: $$\begin{aligned} &&-144\Lambda ^{2}(1+3\Lambda \alpha )(4+15\Lambda \alpha )N^{10}-144\Lambda (72\Lambda ^{2}\alpha ^{2}+49\Lambda \alpha +8)N^{8} \notag \\ &&+8(2916\Lambda ^{3}\alpha ^{3}+2331\Lambda ^{2}\alpha ^{2}+396\Lambda \alpha -8)N^{6}-48\alpha (216\Lambda ^{2}\alpha ^{2}+51\Lambda \alpha -4)N^{4} \notag \\ &&+3\alpha ^{2}(1296\Lambda ^{2}\alpha ^{2}+360\Lambda \alpha -55)N^{2}+64\alpha ^{3}=0 \label{R6s}\end{aligned}$$ Note that the above equation (\[R6s\]) reduces to the equation for $N_{\mathrm{\max }}$ of Einstein gravity when $\alpha $ vanishes. For $N<N_{\mathrm{\max }}$ with negative $\Lambda $, we have two bolt solutions, while for $N=N_{\mathrm{\max }}$ we have only one bolt solution. It is straightforward to show that the value of $N_{\mathrm{\max }}$ is larger in Gauss-Bonnet gravity with respect to its value in Einstein gravity. As $\Lambda \rightarrow 0$, the larger bolt solution goes to infinity, so for the asymptotically flat solution, there is always one and only one bolt solution that is the real root of the following equation: $${r_{b}}^{3}-3N{r_{b}}^{2}-(N^{2}-4\alpha )r_{b}+3N^{3}-6\alpha N=0 \label{6dbolts2s2}$$ Eq. (\[6dbolts2s2\]) has either one or three real roots larger than $N$ since its left hand side is negative at $r_{b}=N$, and positive as $r_{b}$ goes to infinity. However it cannot have three real roots larger than $N$. This is because the product of the roots is equal to $-3N(N^{2}-2\alpha )< 0$, whereas the condition for having three real roots guarantees that $N^{2}>2\alpha $. Thus there exists one and only one real root larger than $N$, and so there is always one bolt solution. This kind of argument applies for all of the solutions we obtain in any dimension with any base space. Finally, we note an additional regularity condition not present in Einstein gravity. The metric function $F(r)$ is real for all values of $r$ in the range $0\leq r\leq \infty $ provided $\alpha <\alpha _{\mathrm{\max }}$, where $\alpha _{\mathrm{\max }}$ depends on the parameters of the metric. This is a general feature of solutions in Gauss-Bonnet gravity and also occurs for static solutions [@Des]. We also note that for $\Lambda>0$ the bolt radius increases with increasing $\alpha$, while for the case of $\Lambda=0$ it decreases as $\alpha$ increases. For the case of $\Lambda<0$, there are two bolt solutions provided $N<N_{\mathrm{\max }}$. As $\alpha$ increases the radius of smaller one decreases, while that of the larger solution increases. For $N=N_{\mathrm{\max }}$, only the smaller bolt solution exists, and its radius $r_b$ decreases as $\alpha$ increases. These features happen for all the bolt solutions in any dimension. Taub-NUT/Bolt Solutions with $T^{2}$ in the base ------------------------------------------------ We now consider the Taub-NUT/bolt solutions of Gauss-Bonnet gravity when $\mathcal{B}$ contains a 2-torus $T^{2}$ with metric $d\Gamma =d\eta ^{2}+d\zeta ^{2}$. There are two possibilities. The first possibility is to choose the base space $\mathcal{B}$ to be $S^{2}\times T^{2}$ with 1-form $$A=2\cos \theta d\phi +2\eta d\zeta \label{TS6}$$ and the second is $T^{2}\times T^{2}$ with 1-form $$A=2\eta _{1}d\zeta _{1}+2\eta _{2}d\zeta _{2} \label{TT6}$$ The function $F(r)$ for the above two cases is given by Eq. (\[F6\]), where $C(r)$ is zero for $\mathcal{B}=T^{2}\times T^{2}$ and is $$\text{\ }C_{ST}=\frac{4\alpha ^{2}}{(r^{2}-N^{2})^{2}} \label{cst6}$$ for $\mathcal{B}=S^{2}\times T^{2}$. The topology $\mathcal{B}=T^{2}\times T^{2}$ at $r\rightarrow \infty $ is $\mathbb{R}^{5}$. Although the boundary is topologically a direct product of the Euclidean time line and the spatial hypersurface ($\eta _{1},\eta _{2},\zeta _{1},\zeta _{2}$), the product is twisted and the boundary is not flat. Unlike the $B=CP^{2}$ case, an immediate consequence of the $T^{2}\times T^{2}$ topology is that the Euclidean time period** **$\beta =4\pi /F^{\prime }$** **will not be fixed by the value of the nut parameter** **$N$**.** Now we consider the NUT solutions. For these two cases the conditions of existence of NUT solutions are satisfied provided the mass parameter is $$\begin{aligned} m_{N} &=&-\frac{16}{5}\Lambda N^{5}, \label{mtt6} \\ m_{N} &=&-\frac{8}{15}N^{3}(6\Lambda N^{2}+5) \label{mst6}\end{aligned}$$ for $\mathcal{B}=T^{2}\times T^{2}$ and $\mathcal{B}=T^{2}\times S^{2}$ respectively. Indeed for these two cases $F^{\prime }(r=N)=0$, and therefore the NUT solutions should be regarded as extremal solutions. Computing the Kretschmann scalar, we find that there is a curvature singularity at $r=N$ for the spacetime with $\mathcal{B}=T^{2}\times S^{2}$, while the spacetime with $\mathcal{B}=T^{2}\times T^{2}$ has no curvature singularity at $r=N$. This leads us to our second conjecture: “*Gauss-Bonnet gravity has extremal NUT solutions whenever the base space is a product of 2-torii with at most one* $2$*-dimensional space of positive curvature*”. Indeed, when the base space has at most one two dimensional curved space as one of its factor spaces, then Gauss-Bonnet gravity admits an extreme NUT solution even though there exists a curvature singularity at $r=N$. Next we consider the Taub-bolt solutions. Euclidean regularity at the bolt requires the period of $\tau $ to be $$\beta =\frac{8\pi r_{b}(r_{b}^{2}-N^{2}+2\alpha )}{(r_{b}^{2}-N^{2})[1-\Lambda (r_{b}^{2}-N^{2})]} \label{betst6}$$ for $\mathcal{B}=T^{2}\times S^{2}$, and $$\beta =-\frac{8\pi r_{b}}{\Lambda (r{_{b}}^{2}-N^{2})} \label{bettt6}$$ for $\mathcal{B}=T^{2}\times T^{2}$. As $r_{b}$ varies from $N$ to infinity, one covers the whole temperature range from $0$ to $\infty $, and therefore we have non-extreme bolt solutions. Note that for the case of $\Lambda =0$, there is no black hole solution with $\mathcal{B}=T^{2}\times T^{2}$. This is also true for spherically symmetric solutions of Gauss-Bonnet gravity [@Cai]. Eight-dimensional Solutions\[8d\] ================================= In eight dimensions there are more possibilities for the base space $\mathcal{B}$. It can be a 6-dimensional space, a product of three 2-dimensional spaces, or the product of a 4-dimensional space with a 2-dimensional one. In all of these cases the form of the function $F(r)$ is $$\begin{aligned} F(r) &=&\frac{(r^{2}-N^{2})^{2}}{8\alpha (5r^{2}+3N^{2})}\left( 1+\frac{4p\alpha }{3(r^{2}-N^{2})}-\sqrt{B(r)+C(r)}\right) \notag \\ B(r) &=&1-\frac{16\alpha mr\left( 5r^{2}+3N^{2}\right) }{3(r^{2}-N^{2})^{5}}+\frac{16p\alpha N^{2}}{15(r^{2}-N^{2})^{5}}(r^{6}-15N^{2}r^{4}-45N^{4}r^{2}-5N^{6}) \notag \\ &&+\frac{16\alpha \Lambda \left( 5r^{2}+3N^{2}\right) }{105(r^{2}-N^{2})^{5}}(5r^{8}-28N^{2}r^{6}+70N^{4}r^{4}-140N^{6}r^{2}-35N^{8}) \label{F8}\end{aligned}$$ where $p$ is again the dimension of the curved factor spaces of $\mathcal{B} $, and the function $C(r)$ depends on the choice of the base space. The 1-form and the metric for the factor spaces $S^{2}$, $T^{2}$, and $\mathbb{CP}^{2}$ have been introduced in the last section. Here, for completeness we write down the metric and 1-form for the factor space $\mathbb{CP}^{3}$, and then we bring the function $F(r)$ for the various base spaces in a table. The metric and the 1-form $A$ for $\mathbb{CP}^{3}$ may be written as: $$\begin{aligned} d{\Sigma _{3}}^{2} &=&8\{d{\xi _{3}}^{2}+\sin ^{2}\xi _{3}\cos ^{2}\xi _{3}(d\psi _{3}+\frac{1}{6}A_{2})^{2}+\frac{1}{6}\sin ^{2}\xi _{3}d{\Sigma _{2}}^{2}\} \label{CP3} \\ A_{3} &=&8\sin ^{2}\xi _{3}\left\{ d\psi _{3}+\sin ^{2}\xi _{2}(d\psi _{2}+\sin ^{2}\xi _{1}d\psi _{1})\right\} \label{A3}\end{aligned}$$ The function $C(r)$ for various base spaces are: $\mathcal{B}$ $(r^{2}-N^{2})^{5}C(r)/\alpha ^{2}$ --------------------------------- --------------------------------------------------------------- ${\,{\mathbb{CP}^{3}}}$ $-16(r^{6}-15N^{2}r^{4}-45N^{4}r^{2}-5N^{6})$ $T^{2}\times T^{2}\times T^{2}$ $0$ $T^{2}\times T^{2}\times S^{2}$ $\frac{64}{9}(r^{2}-N^{2})^{3}$ $T^{2}\times S^{2}\times S^{2}$ $-\frac{64}{9}(r^{6}-15N^{2}r^{4}-45N^{4}r^{2}-5N^{6})$ $S^{2}\times S^{2}\times S^{2}$ $-\frac{128}{3}(r^{6}-9N^{2}r^{4}-21N^{4}r^{2}-3N^{6})$ $T^{2}\times \mathbb{CP}^{2}$ $\frac{128}{27}(r^{6}+9N^{2}r^{4}+51N^{4}r^{2}+3N^{6})$ $S^{2}\times \mathbb{CP}^{2}$ $-\frac{64}{27}(13r^{6}-135N^{2}r^{4}-345N^{4}r^{2}-45N^{6})$ One may note that the asymptotic behavior of all of these solutions is locally AdS for $\Lambda <0$ provided $\left| \Lambda \right| <21/(80\alpha ) $, locally dS for $\Lambda >0$ and locally flat for $\Lambda =0$. Note that all the different $F(r)$’s given in this section have the same form as $\alpha $ goes to zero, reducing to the solutions of Einstein gravity. Taub-NUT Solutions ------------------ Using the conditions for NUT solutions, (i) $F(r=N)=0$ and (ii) $F^{\prime }(r=N)=1/(4N)$, we find that Gauss-Bonnet gravity in eight dimensions admits non-extreme NUT solutions only when the base space is chosen to be $\mathbb{CP}^{3}$. The conditions for a nonsingular NUT solution are satisfied provided the mass parameter is fixed to be $$m_{N}=-\frac{8N^{3}}{105}(16\Lambda N^{4}+42N^{2}-105\alpha ) \label{mcp3}$$ On the other hand, the solutions with $\mathcal{B}=T^{2}\times T^{2}\times T^{2}=\mathcal{B}_{A}$ and $\mathcal{B}=T^{2}\times T^{2}\times S^{2}=\mathcal{B}_{B}$ are extermal NUT solutions provided the mass parameter is $$\begin{aligned} m_{n}^{A} &=&-\frac{128\Lambda N^{7}}{105}, \label{mttt8} \\ m_{n}^{B} &=&-\frac{16N^{5}}{105}(8\Lambda N^{2}+7) \label{mstt8}\end{aligned}$$ These results for eight-dimensional Gauss-Bonnet gravity are consistent with our conjectures. Again, one may note that the former extremal NUT solution does not have a curvature singularity at $r=N$ whereas the latter does. Taub-Bolt Solutions ------------------- The conditions for having a regular bolt solution are (i) $F(r=r_{b})=0$ and $F^{\prime }(r_{b})=1/(4N)$ with $r_{b}>N$. Condition (ii) again follows from the fact that we want to avoid a conical singularity at the bolt, together with the fact that the period of $\tau $ will still be $16\pi N$. Now applying these conditions for $\mathcal{B}=\mathbb{CP}^{3}$ gives the following equation for $r_{b}$: $$4N\Lambda {r_{b}}^{4}+3{r_{b}}^{3}-4(3N+2\Lambda N^{3}){r_{b}}^{2}+3(8\alpha -N^{2})r_{b}+4N(\Lambda N^{4}+3N^{2}-9\alpha )=0 \label{bmatcp8}$$ where again one has two bolt solutions for negative $\Lambda $ provided $N<N_{\text{max}}$ where $N_{\text{max}}$ is the smaller root of the following equation $$\begin{aligned} &&-320\Lambda ^{2}(27+64\Lambda \alpha )(15+64\Lambda \alpha )N^{10}-2880\Lambda (9+16\Lambda \alpha )(15+64\Lambda \alpha )N^{8} \notag \\ &&+9(1179648\Lambda ^{3}\alpha ^{3}+110182\Lambda ^{2}\alpha ^{2}+185760\Lambda \alpha -2025)N^{6} \notag \\ &&-216\alpha (43008\Lambda ^{2}\alpha ^{2}+8992\Lambda \alpha -375)N^{4} \notag \\ &&+324\alpha ^{2}(12288\Lambda ^{2}\alpha ^{2}+2688\Lambda \alpha -301)N^{2}+41472\alpha ^{3}=0 \label{R8cp3}\end{aligned}$$ As $\Lambda $ goes to zero Eq. (\[bmatcp8\]) becomes $${r_{b}}^{3}-4N{r_{b}}^{2}+(8\alpha -N^{2})r_{b}+4N(N^{2}-3\alpha )=0 \label{E8cp3}$$ which holds for locally asymptotically flat solutions. For all N, there is only one solution for which $r_b>N$. Taub-bolt solutions for the case of $\mathcal{B}=S^{2}\times \mathbb{CP}^{2}$ exist provided $N\leq N_{\mathrm{\max }}$ where $N_{\mathrm{\max }}$ is now given by the smallest real root of the following equation $$\begin{aligned} &&-4800\Lambda ^{2}(243+896\Lambda \alpha )(27+128\Lambda \alpha )N^{10}-2880\Lambda (499712\Lambda ^{2}\alpha ^{2}+1260496\Lambda \alpha +32805)N^{8} \notag \\ &&+(5532286976\Lambda ^{3}\alpha ^{3}+3243829248\Lambda ^{2}\alpha ^{2}+422703360\Lambda \alpha -4428675)N^{6} \notag \\ &&-72\alpha (3381248\Lambda ^{2}\alpha ^{2}+623232\Lambda \alpha -32265)N^{4} \notag \\ &&+139968\alpha ^{2}(6912\Lambda ^{2}\alpha ^{2}+1472\Lambda \alpha -159)N^{2}+10077696\alpha ^{3}=0\end{aligned}$$ In this case there exist two $r_{b}$’s which are the real roots of $$4N\Lambda {r_{b}}^{4}+3{r_{b}}^{3}-4(3N+2\Lambda N^{3}){r_{b}}^{2}+3(8\alpha -N^{2})r_{b}+4N(\Lambda N^{4}+3N^{2}-\frac{32}{3}\alpha )=0 \label{s2cp2reg}$$ Next we consider the Taub-bolt solutions for $\mathcal{B}=S^{2}\times S^{2}\times S^{2}$. One finds that $r_{b}$ is given by the solution of the following equation $$4N\Lambda {r_{b}}^{4}+3{r_{b}}^{3}-4(3N+2\Lambda N^{3}){r_{b}}^{2}+3(8\alpha -N^{2})r_{b}+4N(\Lambda N^{4}+3N^{2}-12\alpha )=0$$ One has two bolt solutions for $N<N_{\mathrm{\max }}$ and one solution for $N=N_{\mathrm{\max }}$ where $N_{\mathrm{\max }}$ is now given by the smallest real root of the following equation $$\begin{aligned} &&-64\Lambda ^{2}(27+128\Lambda \alpha )(25+128\Lambda \alpha )N^{10}-44\Lambda (45056\Lambda ^{2}\alpha ^{2}+19152\Lambda \alpha +2025)N^{8} \notag \\ &&+(12058624\Lambda ^{3}\alpha ^{3}+5548032\Lambda ^{2}\alpha ^{2}+597888\Lambda \alpha -6075)N^{6} \notag \\ &&-72\alpha (38912\Lambda ^{2}\alpha ^{2}+6144\Lambda \alpha -417)N^{4} \notag \\ &&+27648\alpha ^{2}(48\Lambda ^{2}\alpha ^{2}+10\Lambda \alpha -1)N^{2}+13824\alpha ^{3}=0\end{aligned}$$ For the locally asymptotic flat case, $r_{b}$ is the solution to the equation $${r_{b}}^{3}-4N{r_{b}}^{2}+(8\alpha -N^{2})r_{b}+12N(N^{2}-4\alpha )=0 \label{8dbolts2cubed}$$ and using reasoning similar to the 6-dimensional case we have only one bolt solution. For the case of $\mathcal{B}=T^{2}\times T^{2}\times T^{2}$ and $\mathcal{B}=T^{2}\times T^{2}\times S^{2}$, Euclidean regularity at the bolt requires the period of $\tau $ to be $$\beta =-\frac{12\pi r_{b}}{\Lambda ({r_{b}}^{2}-N^{2})}$$ and $$\beta =\frac{4\pi r_{b}(3{r_{b}}^{2}-3N^{2}+8\alpha )}{({r_{b}}^{2}-N^{2})[1-\Lambda ({r_{b}}^{2}-N^{2})]}$$ respectively. As $r_{b}$ varies from $N$ to infinity, one covers the whole temperature range from $0$ to $\infty $, and therefore one can have bolt solutions. Again, one may note that for the case of asymptotic locally flat solution with base space $\mathcal{B}=T^{2}\times T^{2}\times T^{2}$, there is no black hole solution. As with the six-dimensional case, there is a maximum value for $\alpha $ for all base spaces that ensures all bolt and NUT solutions are regular in eight dimensions (as in the case of static solutions). Ten-dimensional Solutions\[10d\] ================================ In ten dimensions there are more possibilities for the base space $\mathcal{B}$. It can be an $8$-dimensional space, the product of a $6$-dimensional space with a $2$-dimensional one, a product of two $4$-dimensional spaces, a product of a $4$-dimensional space with two $2$-dimensional spaces, or the product of four $2$-dimensional spaces. In all of these cases the form of the function $F(r)$ is $$\begin{aligned} F(r) &=&\frac{(r^{2}-N^{2})^{2}}{12\alpha (7r^{2}+3N^{2})}\left( 1+\frac{3p\alpha }{2(r^{2}-N^{2})}-\sqrt{B(r)+C(r)}\right) , \notag \\ B(r) &=&1+\frac{36\alpha mr\left( 7r^{2}+3N^{2}\right) }{(r^{2}-N^{2})^{6}}+\frac{6p\alpha N^{2}}{35(r^{2}-N^{2})^{6}}(3r^{8}-28N^{2}r^{6}+210N^{4}r^{4}+420N^{6}r^{2}+35N^{8}) \notag \\ &&+\frac{2\alpha \Lambda \left( 7r^{2}+3N^{2}\right) }{21(r^{2}-N^{2})^{6}}(7r^{10}-45N^{2}r^{8}-126N^{4}r^{6}-210N^{6}r^{4}+315N^{8}r^{2}+63N^{10}) \label{F10}\end{aligned}$$ where $p$ is the dimensionality of the curved portion of the base space, and the function $C(r)$ depends on the choice of the base space $\mathcal{B}$. The first case is when one chooses the $8$-dimensional base space to be $\mathbb{CP}^{4}$. In this case the one-form $A$ and the metric of $\mathbb{CP}^{4}$ may be written as: $$\begin{aligned} A_4&=&10\sin^2 \xi_4 \left(d\psi_4+ \frac{1}{8}A_3\right) \label{A4} \\ d{\Sigma _{3}}^{2} &=&10\{d{\xi_4}^2+\sin^2 \xi_4 \cos^2 \xi_4 (d\psi_4+\frac{1}{8}A_3)^2 +\frac{1}{8}\sin^2\xi_4 d{\Sigma _{3}}^{2} \} \label{CP4}\end{aligned}$$ where $A_3$ and $d{\Sigma _{3}}^{2}$ are given by Eqs. (\[A3\]) and (\[CP3\]) The 1-forms for all the other cases have been introduced in the previous sections, so we give only the function $C(r)$ for different base spaces in the following table: $\mathcal{B}$ $(r^{2}-N^{2})^{6}C(r)/\alpha ^{2}$ --------------------------------------------- ------------------------------------------------------------------------------ $\,{\mathbb{CP}^{4}}$ $-\frac{144}{25}(3r^{8}-28N^{2}r^{6}+210N^{4}r^{4}+420N^{6}r^{2}+35N^{8})$ $T^{2}\times T^{2}\times T^{2}\times T^{2}$ $0$ $T^{2}\times T^{2}\times T^{2}\times S^{2}$ $9(r^{2}-N^{2})^{4}$ $T^{2}\times T^{2}\times S^{2}\times S^{2}$ $\frac{12}{5}(r^{8}+4N^{2}r^{6}-90N^{4}r^{4}-220N^{6}r^{2}-15N^{8})$ $T^{2}\times S^{2}\times S^{2}\times S^{2}$ $-\frac{9}{5}(11r^{8}-76N^{2}r^{6}+450N^{4}r^{4}+820N^{6}r^{2}+75N^{8})$ $S^{2}\times S^{2}\times S^{2}\times S^{2}$ -$\frac{285}{5}(r^{8}-6N^{2}r^{6}+300N^{4}r^{4}+50N^{6}r^{2}+5N^{8})$ $T^{2}\times \mathbb{CP}^{3}$ $\frac{27}{5}(r^{8}+4N^{2}r^{6}-90N^{4}r^{4}-220N^{6}r^{2}-15N^{8})$ $S^{2}\times \mathbb{CP}^{3}$ $-\frac{18}{5}(9r^{8}-64N^{2}r^{6}+390N^{4}r^{4}+720N^{6}r^{2}+65N^{8})$ $T^{2}\times T^{2}\times \mathbb{CP}^{2}$ $\frac{4}{5}(17r^{8}-52N^{2}r^{6}-90N^{4}r^{4}-500N^{6}r^{2}-15N^{8})$ $T^{2}\times S^{2}\times \mathbb{CP}^{2}$ $-\frac{1}{5}(43r^{8}-428N^{2}r^{6}+3330N^{4}r^{4}+6740N^{6}r^{2}+555N^{8})$ $S^{2}\times S^{2}\times \mathbb{CP}^{2}$ $-\frac{8}{5}(29r^{8}-184N^{2}r^{6}+990N^{4}r^{4}+1720N^{6}r^{2}+165N^{8})$ $\mathbb{CP}^{2}\times \mathbb{CP}^{2}$ $-\frac{16}{5}(11r^{8}-76N^{2}r^{6}+450N^{4}r^{4}+820N^{6}r^{2}+75N^{8})$ Note that the asymptotic behavior of all of these solutions is locally AdS for $\Lambda <0$ provided $\left| \Lambda \right| <9/(42\alpha )$, locally dS for $\Lambda >0$ and locally flat for $\Lambda =0$. As with the 6 and 8 dimensional cases, all the different $F(r)$’s have the same form as $\alpha $ goes to zero and reduce to the solutions of Einstein gravity. Taub-NUT Solutions ------------------ Using the conditions for NUT solutions, (i) $F(r=N)=0$ and (ii) $F^{\prime }(r=N)=1/(5N)$, we find that Gauss-Bonnet gravity in ten dimensions admits non-extreme NUT solutions only when the base space is chosen to be $\mathbb{CP}^{4}$. There is no curvature singularity at $r=N$ for this solution provided the mass parameter is fixed to be $$m_{n}=-\frac{128N^{5}}{4725}(25\Lambda N^{4}+90N^{2}-378\alpha ) \label{mcp4}$$ On the other hand, the solutions with $\mathcal{B}=T^{2}\times T^{2}\times T^{2}\times T^{2}=\mathcal{B}_{A}$ and $\mathcal{B}=T^{2}\times T^{2}\times T^{2}\times S^{2}=\mathcal{B}_{B}$ are extermal NUT solution provided the mass parameter is $$\begin{aligned} m_{n}^{A} &=&-\frac{128\Lambda N^{9}}{189}, \label{mttt10} \\ m_{n}^{B} &=&-\frac{64N^{7}}{945}(10\Lambda N^{2}+9) \label{mstt10}\end{aligned}$$ These results for ten-dimensional Gauss-Bonnet gravity are consistent with our conjectures – we have found that all other spaces with 2-dimensional factors in the base yield singular solutions. It is also straightforward to show that the former extremal NUT solution has no curvature singularity at $r=N$, whereas the latter has**.** Taub-Bolt Solutions ------------------- The conditions for having a regular bolt solution are (i) $F(r=r_{b})=0$ and $F^{\prime }(r_{b})=1/(5N)$ with $r_{b}>N$. Now applying these conditions for $\mathcal{B}=\mathbb{CP}^{4}$ gives the following equation for $r_{b}$: $$5N\Lambda {r_{b}}^{4}+4{r_{b}}^{3}-10N(2+2\Lambda N^{2}){r_{b}}^{2}+4(12\alpha -N^{2})r_{b}+N(5\Lambda N^{4}+20N^{2}-96\alpha )=0$$ where there exists a maximum value for the NUT parameter given by the smallest real root of the following equation $$\begin{aligned} &&-25\Lambda ^{2}(4+15\Lambda \alpha )(1+15\Lambda \alpha )N^{10}-50\Lambda (8+27\Lambda \alpha )(1+15\Lambda \alpha )N^{8} \notag \\ &&+(34500\Lambda ^{3}\alpha ^{3}+18875\Lambda ^{2}\alpha ^{2}+2320\Lambda \alpha -16)N^{6} \notag \\ &&-2\alpha (6975\Lambda ^{2}\alpha ^{2}+1205\Lambda \alpha -48)N^{4} \notag \\ &&+\alpha ^{2}(6075\Lambda ^{2}\alpha ^{2}+1170\Lambda \alpha -109)N^{2}+48\alpha ^{3}=0\end{aligned}$$ in order for bolt solutions to exist. For the locally asymptotic flat case, $r_{b}$ is the solution of $${r_{b}}^{3}-5N{r_{b}}^{2}+(12\alpha -N^{2})r_{b}+N(5N^{2}-24\alpha )=0 \label{10dbolt}$$ which yields only a single bolt solution. For the case of $\mathcal{B}=T^{2}\times T^{2}\times T^{2}\times T^{2}$ and $\mathcal{B}=S^{2}\times T^{2}\times T^{2}\times T^{2}$, Euclidean regularity at the bolt requires the period of $\tau $ to be $$\beta =-\frac{16\pi r_{b}}{\Lambda ({r_{b}}^{2}-N^{2})}$$ and $$\beta =\frac{16\pi r_{b}({r_{b}}^{2}-N^{2}+3\alpha )}{({r_{b}}^{2}-N^{2})[1-\Lambda ({r_{b}}^{2}-N^{2})]}$$ respectively. As $r_{b}$ varies from $N$ to infinity, one covers the whole temperature range from $0$ to $\infty $, and therefore one can have bolt solutions. Again, one may note that for the case of asymptotic locally flat solution with base space $\mathcal{B}=T^{2}\times T^{2}\times T^{2}$, there is no black hole solution. Gauss-Bonnet gravity has other Taub-bolt solutions in ten dimensions corresponding to the various base space factors given in the above table. We shall omit the details of these solutions here, and simply note that it is sufficient to apply the regularity and reality conditions for bolt solutions mentioned at the beginning of this subsection to find the equation for** **$r_{b}$. The General Form of the Solutions in $d$ Dimensions\[dd\] ========================================================= In this section we write down the solutions of field equation (\[Geq\]) for the metric (\[TN\]). We find that the form of $F(r)$ in any dimension and for any choice of the base space is $$F(r)=\frac{(r^{2}-N^{2})\{r^{2}-N^{2}+\frac{2(d-4)}{d-2}p\alpha \}}{2(d-4)\alpha \{(d-3)r^{2}+3N^{2}\}}-\sqrt{F_{2}(r)} \label{Fd}$$ where $p$ is the sum of the dimensions of the curved factor spaces of $\mathcal{B}$. Here, we consider only the NUT/bolt solutions of those cases for which Gauss-Bonnet gravity admits NUT solutions, leaving out the other cases which have only bolt solution for reasons of economy**.** Our conjecture implies that there are three cases that have NUT solutions in $2k+2$ dimensions. Taub-NUT/bolt solutions for the base space $\mathcal{B}=\mathbb{CP}^{k}$ ------------------------------------------------------------------------ The only case for which Gauss-Bonnet gravity admits non-extreme NUT solutions in $2k+2$ dimensions is when the base space is $\mathcal{B}=\mathbb{CP}^{k}$. In this case the metric may be written as [@Pop] $$d\Sigma _{k}^{2}=(2k+2)\left\{ d\xi _{k}^{2}+\sin ^{2}\xi _{k}\cos ^{2}\xi _{k}(d\psi _{k}+\frac{1}{2k}A_{k-1})^{2}+\frac{1}{2k}\sin ^{2}\xi _{k}d\Sigma _{k-1}^{2}\right\} \label{CPk}$$ where $A_{k-1}$ is the Kähler potential of $\mathbb{CP}^{k-1}$. Here $\xi _{k}$ and $\psi _{k}$ are the extra coordinates corresponding to $\mathbb{CP}^{k}$ with respect to $\mathbb{CP}^{k-1}$. Also, the metric is normalized such that, Ricci tensor is equal to the metric, $R_{\mu \nu }=g_{\mu \nu }$. The 1-form $A_{k}$, which is the Kähler potential of $\mathbb{CP}^{k}$, is $$A_{k}=(2k+2)\sin ^{2}\xi _{k}(d\psi _{k}+\frac{1}{2k}A_{k-1}) \label{Ak}$$ Now the $tt$-component of the field equation can be written as: $$\Gamma _{1}rF^{\prime }(r)+\Gamma _{2}F^{2}(r)+\Gamma _{3}F(r)+\Gamma _{4}=0, \label{ECP}$$ where $\Gamma _{1}$, $\Gamma _{2}$, $\Gamma _{3}$ and $\Gamma _{4}$ are $$\begin{aligned} \Gamma _{1} &=&-\alpha \left( (d-3)r^{2}+3N^{2}\right) F(r)+(r^{2}-N^{2})\left( \alpha +\frac{r^{2}-N^{2}}{2(d-4)}\right) , \notag \\ \Gamma _{2} &=&-\frac{\alpha }{2(r^{2}-N^{2})}\left\{ (d-3)(d-5)r^{4}+2(d-9)N^{2}r^{2}+3N^{4}\right\} , \notag \\ \Gamma _{3} &=&-\frac{(r^{2}-N^{2})[(d-3)r^{2}+N^{2}]}{2(d-4)}+\alpha \left( (d-5)r^{2}+N^{2}\right) , \notag \\ \Gamma _{4} &=&(r^{2}-N^{2})\left\{ \Lambda \frac{(r^{2}-N^{2})^{2}}{(d-2)(d-4)}-\frac{(r^{2}-N^{2})}{2(d-4)}-\frac{(d-2)\alpha }{2d}\right\} \label{GamCP}\end{aligned}$$ We can write the general form of $F_{2}(r)$ in $d$ dimensions. Inserting the general form of $F(r)$ given in Eq. (\[Fd\]) into the differential equation (\[ECP\]), we find $F_{2}(r)$ for $\mathcal{B}=\mathbb{CP}^{k}$as: $$F_{2}(r)=\frac{r\left( (-1)^{(d-2)/2}m+\frac{1}{d(d-2)(d-4)^{2}\alpha }\int \Upsilon (r)dr\right) }{4\alpha \{(d-3)r^{2}+3N^{2}\}(r^{2}-N^{2})^{(d-6)/2}} \label{F2cp}$$ where $\Upsilon (r)$ is $$\begin{aligned} \Upsilon (r) &=&\frac{(r^{2}-N^{2})^{d/2}}{r^{2}[(d-3)r^{2}+3N^{2}]^{2}}\{-24(d-2)(d-3)(d-4)^{2}\alpha ^{2}+8d(d-2)\alpha \Lambda \lbrack (d-3)r^{2}+3N^{2}]^{2} \notag \\ &&+24d(d-2)(d-4)\alpha N^{2}+d(d-2)[(d-1)(d-3)r^{4}+6(d-1)r^{2}N^{2}+3N^{4}]\}\end{aligned}$$ The solutions given by Eqs. (\[Fd\]) and (\[F2cp\]) yield a NUT solution for any given (even) dimension $d>4$ provided the mass parameter $m$ is fixed to be $$m_{\mathrm{nut}}=-\frac{(\frac{d}{2}-3)!2^{d/2}N^{d-5}}{2d(d-1)!!}\left\{ 2d^{2}\Lambda N^{4}+d(d-1)(d-2)N^{2}-(d-1)(d-2)(d-3)(d-4)\alpha \right\} \label{mncp}$$ This solution has no curvature singularity at $r=N$. Solutions given by Eqs. (\[Fd\]) and (\[F2cp\]) for $m\neq m_{\mathrm{nut }}$ in any dimension can be regarded as bolt solutions. The value of the bolt radius $r_{b}>N$ may be found from the regularity conditions (i) $F(r=r_{b})=0$ and $F^{\prime }(r_{b})=2/(dN)$. Applying these for $\mathcal{B}=\mathbb{CP}^{k}$ gives the following equation for $r_{b}$: $$\begin{aligned} 0 &=&2d\Lambda N{r_{b}}^{4}+2(d-2){r_{b}}^{3}-dN\left[ (d-2)+4\Lambda N^{2}\right] {r_{b}}^{2}+2(d-2)\left[ 2(d-4)\alpha -N^{2}\right] r_{b} \notag \\ &&+N\left[ 2d\Lambda N^{4}+d(d-2)N^{2}-(d-4)(d-2)^{2}\alpha \right]\end{aligned}$$ which has two real roots larger than $N$, provided $N<N_{\mathrm{\max }}$ and one for $N=N_{\mathrm{\max }}$, where $N_{\mathrm{\max }}$ is the smaller real root of the following equation $$\begin{aligned} &&-4(d+2)d^{2}[8d(d-4)\alpha \Lambda +d^{2}-4][8d(d-4)(d-6)\alpha \Lambda +d^{3}-6d^{2}+12d-8]\Lambda ^{2}N^{10} \notag \\ &&-2d^{2}(d-2)(d+2)[8d(d-4)\alpha \Lambda +d^{2}-4][8(d-4)(d^{2}-6d-4)\alpha \Lambda +d^{3}-6d^{2}+12d-8]\Lambda N^{8} \notag \\ &&+(d-2)^{2}\{128d^{3}(d-4)^{3}(d^{2}-4d-14)\alpha ^{3}\Lambda ^{3}+32d^{2}(d-4)^{2}(d^{4}-4d^{3}-4d^{2}+40d+40)\Lambda ^{2}\alpha ^{2} \notag \\ &&+2d(d-4)(d^{2}-4)(d^{4}-4d^{3}+6d^{2}+80d+24)\alpha \Lambda -d^{6}+4d^{5}+4d^{4}-32d^{3}+16d^{2}+64d-64\}N^{6} \notag \\ &&-(d-4)(d-2)^{2}\{192d^{2}(d-4)^{2}(d^{2}+2d+4)\alpha ^{2}\Lambda ^{2}+16d(d-2)(d-4)(d^{3}+8d^{2}+14d-12)\alpha \Lambda \notag \\ &&-d^{6}+6d^{5}-16d^{4}-48d^{3}+176d^{2}+96d-384\}\alpha N^{4}+(d-4)^{2}(d-2)^{2}\{1728d^{2}(d-4)^{2}\alpha ^{2}\Lambda ^{2} \notag \\ &&+48d(d-2)(5d^{2}-18d-8)\alpha \Lambda -13d^{4}+16d^{3}+8d^{2}+192d-336\}\alpha ^{2}N^{2}\notag \\ && +128(d-2)^{4}(d-4)^{3}\alpha ^{3}=0\end{aligned}$$ For the locally asymptotically flat case, $r_{b}$ is the solution of $$2{r_{b}}^{3}-dN{r_{b}}^{2}+2\left[ 2(d-4)\alpha -N^{2}\right] r_{b}+N\left[ dN^{2}-(d-4)(d-2)\alpha \right] =0 \label{dbolt}$$ where there is no condition in order for a bolt solution to exist, except in six dimensions for which there exists a lower limit of the NUT charge. Next we write down the general form of the solutions with the base space $\mathcal{B}=T^{2}\times ...\times T^{2}$. The field equation is given by (\[ECP\]), where now $$\begin{aligned} \Gamma _{1} &=&-\alpha \{(d-3)r^{2}+3N^{2}\}F+\frac{(r^{2}-N^{2})^{2}}{2(d-4)}, \notag \\ \Gamma _{2} &=&-\frac{\alpha }{2(r^{2}-N^{2})}\left\{ (d-3)(d-5)r^{4}+2(d-9)N^{2}r^{2}+3N^{4}\right\} , \notag \\ \Gamma _{3} &=&-\frac{(r^{2}-N^{2})[(d-3)r^{2}+N^{2}]}{2(d-4)}, \notag \\ \Gamma _{4} &=&\Lambda \frac{(r^{2}-N^{2})^{3}}{(d-2)(d-4)} \label{GamTT}\end{aligned}$$ Inserting the general form of $F(r)$ given in Eq. (\[Fd\]) into the differential equation (\[ECP\]) with $\Gamma _{i}$’s of Eqs. (\[GamTT\]), we find that $F_{2}(r)$ for $\mathcal{B}=T^{2}\times T^{2}\times ...\times T^{2}$ is $$F_{2}(r)=\frac{r\left( (-1)^{(d-2)/2}m+\frac{1}{(d-2)(d-4)^{2}\alpha }\int \Upsilon (r)dr\right) }{4\alpha \{(d-3)r^{2}+3N^{2}\}(r^{2}-N^{2})^{(d-6)/2}} \label{F2TT}$$ where $\Upsilon (r)$ is $$\begin{aligned} \Upsilon (r) &=&\frac{(r^{2}-N^{2})^{d/2}}{r^{2}\{(d-3)r^{2}+3N^{2}\}^{2}}\{8(d-4)\alpha \Lambda \lbrack (d-3)r^{2}+3N^{2}]^{2} \\ &&+(d-2)[(d-1)(d-3)r^{4}+6(d-1)r^{2}N^{2}+3N^{4}]\end{aligned}$$ The solutions given by Eqs. (\[Fd\]) and (\[F2TT\]) yield a NUT solution for any given (even) dimension $d>4$ provided the mass parameter $m$ is fixed to be $$m_{\mathrm{nut }}=-\frac{d(\frac{d}{2}-3)!2^{d/2}}{(d-1)!!}\Lambda N^{d-1} \label{mnTT}$$ where in this case the spacetime has no curvature singularity at $r=N$. Also one may find that Euclidean regularity at the bolt requires the period of $\tau $ to be $$\beta =-\frac{2(d-2)\pi r_{b}}{\Lambda ({r_{b}}^{2}-N^{2})} \label{betTT}$$ and can have any value from zero to infinity as $r_{b}$ varies from $N$ to infinity, and therefore one can have bolt solution. Finally, we consider the solution when $\mathcal{B}=S^{2}\times T^{2}\times ...\times T^{2}$. In this case the field has the same form as Eq. (\[ECP\]) with $$\begin{aligned} \Gamma _{1} &=&-\alpha \{(d-3)r^{2}+3N^{2}\}F+(r^{2}-N^{2})\{\frac{2}{d-2}\alpha +\frac{r^{2}-N^{2}}{2(d-4)}\}, \notag \\ \Gamma _{2} &=&-\frac{\alpha }{2(r^{2}-N^{2})}\left\{ (d-3)(d-5)r^{4}+2(d-9)N^{2}r^{2}+3N^{4}\right\} , \notag \\ \Gamma _{3} &=&-\frac{(r^{2}-N^{2})[(d-3)r^{2}+N^{2}]}{2(d-4)}+\frac{4(d-4)}{d-2}\alpha \{(d-5)r^{2}+N^{2}\}, \notag \\ \Gamma _{4} &=&\frac{(r^{2}-N^{2})^{2}[\Lambda (r^{2}-N^{2})-1]}{(d-2)(d-4)} \label{GamST}\end{aligned}$$ Inserting the general form of $F(r)$ given in Eq. (\[Fd\]) into the differential equation (\[ECP\]) with $\Gamma _{i}$’s of Eqs. (\[GamST\]), we find that $F_{2}(r)$ for $\mathcal{B}=T^{2}\times T^{2}\times ...\times T^{2}$ is $$F_{2}(r)=\frac{r\left( (-1)^{(d-2)/2}m+\frac{1}{(d-2)^{2}(d-4)^{2}\alpha }\int \Upsilon (r)dr\right) }{4\alpha \{(d-3)r^{2}+3N^{2}\}(r^{2}-N^{2})^{(d-6)/2}} \label{F2ST}$$ where $\Upsilon (r)$ is $$\begin{aligned} \Upsilon (r) &=&\frac{(r^{2}-N^{2})^{d/2}}{r^{2}\{(d-3)r^{2}+3N^{2}\}^{2}}\{16(d-4)^{2}\alpha ^{2}\frac{(d-3)(d-5)r^{4}+6(d-3)N^{2}r^{2}+3N^{4}}{(r^{2}-N^{2})^{2}} \notag \\ &&+8(d-2)(d-4)\alpha \Lambda \lbrack (d-3)r^{2}+3N^{2}]^{2}+48(d-2)(d-4)\alpha N^{2} \notag \\ &&-(d-2)^{2}[(d-1)(d-3)r^{4}+6(d-1)r^{2}N^{2}+3N^{4}]\}\end{aligned}$$ The solutions given by Eqs. (\[Fd\]) and (\[F2ST\]) yield a NUT solution for any given (even) dimension $d\geq 4$ with curvature singularity at $r=N$, provided the mass parameter $m$ is fixed to be $$m_{\mathrm{nut}}=-\frac{(\frac{d}{2}-3)!2^{d/2}}{(d-1)!!}N^{d-3}\left\{ d\Lambda N^{2}+(d-1)\right\} \label{mnST}$$ Also one may find that the Euclidean regularity at the bolt requires the period of $\tau $ to be $$\beta =\frac{2(d-2)\pi r_{b}({r_{b}}^{2}-N^{2}+\frac{4(d-4)}{d-2}\alpha )}{({r_{b}}^{2}-N^{2})[1-\Lambda ({r_{b}}^{2}-N^{2})]} \label{betST}$$ Again, $\beta $ of Eq. (\[betST\]) can have any value from zero to infinity as $r_{b}$ varies from $N$ to infinity, and therefore one can have bolt solution. The asymptotic behavior of all of these solutions is locally AdS for $\Lambda <0$ provided $\left| \Lambda \right| <(d-1)(d-2)/[(d-3)(d-4)\alpha ]$ locally dS for $\Lambda >0$ and locally flat for $\Lambda =0$ . All the different $F(r)$’s for differing base spaces have the same form as $\alpha$ goes to zero and reduce to the solutions of Einstein gravity. Concluding Remarks \[con\] ========================== We have considered the existence of Taub-NUT/bolt solutions in Gauss-Bonnet gravity with and without cosmological term. These solutions are constructed as circle fibrations over even dimensional spaces that in general are products of Einstein-Kähler spaces. We found that the function $F(r)$ of the metric depends on the specific form of the base factors on which one constructs the circle fibration. In other words we found that the solutions are sensitive to the geometry of the base space, in contrast to Einstein gravity where the metric in any dimension is independent of the specific form of the base factors. We restricted ourselves to the cases of base spaces with zero or positive curvature factor spaces, since when the base space has $2$-hyperboloids with negative curvature, then the function $F(r)$ is not real for the whole range of $r$ when $\alpha $ is positive. We found that when Einstein gravity admits non-extremal NUT solutions with no curvature singularity at $r=N$, then there exists a non-extremal NUT solution in Gauss-Bonnet gravity. In $(2k+2)$-dimensional spacetime, this happens when the metric of the base space is chosen to be $\mathbb{CP}^{k}$. Indeed, Gauss-Bonnet gravity does not admit non-extreme NUT solutions with any other base space. We also found that when the base space has at most a 2-dimensional curved factor space with positive curvature, then Gauss-Bonnet gravity admits extremal NUT solutions where the temperature of the horizon at $r=N$ vanishes. We have extended these observations to two conjectures about the existence of NUT solutions in Gauss-Bonnet gravity. We also found the bolt solutions of Gauss-Bonnet gravity in various dimensions and different base spaces, and gave the equations which can be solved for the horizon radius of the bolt solution. For $\Lambda \neq 0$, there is a maximal value for the NUT charge in terms of the parameters of the metric that $N$ must be less than or equal to that in order to have bolt solutions. In this case for $\Lambda <0$ we have two bolt solutions in Gauss-Bonnet gravity if $N<N_{\mathrm{\max }}$ and one bolt solution if $N=N_{\mathrm{\max }}$. The value of $N_{\mathrm{max}}$ is larger in Gauss-Bonnet gravity than in Einstein gravity. For the case of $\Lambda =0$ we always have only one bolt solution. In the 6-dimensional case with the base space $\mathbb{CP}^{2}$ the bolt solution exists provided the NUT charge is larger than its minimal value of $\sqrt{\alpha }$. This situation is quite unlike that in Einstein gravity, where the NUT charge can smoothly approach zero. We note that Gauss-Bonnet gravity in six dimensions gives the most general second order differential equation in classical gravity, while in higher dimensions in order to have the most general second order differential equation, one should turn on other higher terms of Lovelock gravity. Our solutions have been generalized in an obvious way for even dimensions higher than ten. We gave the differential equation and the general form of the function $F(r)$ in any arbitrary even dimensions. For instance, in twelve dimensions, the base space is ten-dimensional and in general can be factorized as a $10$-dimensional space, the product of an $8$-dimensional space with a $2$-dimensional one, a product of $6$-dimensional space with a $4$-dimensional space, a product of a $6$-dimensional space with two $2$-dimensional spaces, a product of two $4$-dimensional spaces with a $2$-dimensional space, a product of a $4$-dimensional spaces with three $2$-dimensional spaces, or the product of five $2$-dimensional spaces. In any dimension $2k+2$, we have only one non-extremal NUT solution with $\mathbb{CP}^{k}$ as the base space, and two extremal NUT solutions with the base spaces $T^{2}\times T^{2}\times .....\times T^{2}$ and $S^{2}\times T^{2}\times T^{2}\times .....\times T^{2}$. There is no curvature singularity for the first two case, while for the latter case, the spacetime has curvature singularity at $r=N$. Insofar as Gauss-Bonnet gravity is expected to model leading-order quantum corrections to Einstein gravity, we see that quantum effects can be expected to single out a preferred non-singular base space and to yield a minimal value for the NUT charge in six dimensions. Thermodynamically this corresponds to a maximal temperature. The study of thermodynamic properties of these solutions, the investigation of the existence of NUT solutions in continued Lovelock gravity, or Lovelock gravity with higher order terms remain to be carried out in future. [**Acknowledgements**]{} We are grateful to C. Stelea for helpful discussions about $\mathbb{CP}^{k}$. M. H. D. would like to thank the University of Waterloo and the Perimeter Institute for Theoretical Physics for support and hospitality during the course of this work. This work was supported in part by the Natural Sciences and Engineering Council of Canada. [99]{} A. Chamblin, R. Emparan, C. V. Johnson and R. C. Myers,“Large N phases, gravitational instantons and the NUTs and bolts of AdS holography”, Phys. Rev. D **59**, 064010 (1999) \[arXiv:hep-th/9808177\]. S. W. Hawking, C. J. Hunter and D. N. Page, “Nut Charge, Anti-de Sitter Space and Entropy”, Phys. Rev. D **59**, 044033 (1999) \[hep-th/9809035\]. R. B. Mann, “Misner String Entropy", Phys. Rev. D **60**, 104047 (1999) \[hep-th/9903229\]. S. Cherkis and A. Hashimoto, “Supergravity Solution of Intersecting Branes and AdS/CFT with Flavour", JHEP **0211**, 036 (2002) \[hep-th/0210105\]; R. Clarkson, A. M. Ghezelbash and R. B. Mann, “New Reducible M-brane Solutions in D=11 Supergravity", JHEP **0404**, 063 (2004) \[hep-th/0404071\]; R. Clarkson, A. M. Ghezelbash and R. B. Mann, “New Reducible 5-brane Solutions in M-Theory", JHEP **0408**, [025]{} (2004) \[hep-th/0405148\]. A. H. Taub, “Empty Space-Times Admitting a Three Parameter Group of Motions”, Annal. Math. **53**, 472 (1951). E. Newman, L. Tamburino, and T. Unti, “Empty-space generalization of the Schwarzschild metric”, J. Math. Phys. **4**, 915 (1963). F. A. Bais and P. Batenburg, “A New Class Of Higher Dimensional Kaluza-Klein Monopole And Instanton Solutions”, Nucl. Phys. B **253**, 162 (1985). D. N. Page and C. N. Pope, “Inhomogeneous Einstein Metrics On Complex Line Bundles”, Class. Quant. Grav. **4**, 213 (1987). M. M. Akbar and G. W. Gibbons, “Ricci-flat metrics with U(1) action and the Dirichlet boundary-value problem in Riemannian quantum gravity and isoperimetric inequalities”, Class. Quant. Grav. **20**, 1787 (2003) \[arXiv:hep-th/0301026\]. M. M. Taylor-Robinson, “Higher dimensional Taub-Bolt solutions and the entropy of non compact manifolds” \[hep-th/9809041\]. A. Awad and A. Chamblin, “A bestiary of higher dimensional Taub-NUT-AdS spacetimes”, Class. Quant. Grav. **19**, 2051 (2002) \[hep-th/0012240\]. R. Clarkson, L. Fatibene and R. B. Mann, “Thermodynamics of ($d+1$)-dimensional NUT-charged AdS Spacetimes”, Nucl. Phys. **B652** 348 (2003) \[hep-th/0210280\]. Robert Mann and Cristian Stelea, “Nuttier (a)dS black holes in higher dimensions", Class. Quant. Grav. **21**, 2937 (2004). Robert Mann and Cristian Stelea, New multiply Nutty spacetimes" \[hep-th/0508203\]. M. B. Greens, J. H. Schwarz and E. Witten, *Superstring Theory*, (Cambridge University Press, Cambridge, England, 1987); D. Lust and S. Theusen, *Lectures on String Theory*, (Springer, Berlin, 1989); J. Polchinski, *String Theory*, (Cambridge University Press, Cambridge, England, 1998). B. Zwiebach, “Curvature squared terms and string theories", Phys. Lett. **B156**, 315 (1985); B. Zumino, Phys. Rep. “Gravity theories in more than four dimensions", **137**, 109 (1986). D. Lovelock, “The Einstein Tensor and Its Generalizations", J. Math. Phys. **12**, 498 (1971); N. Deruelle and L. Farina-Busto, “Lovelock gravitational field equations in cosmology", Phys. Rev. D **41**, 3696 (1990); G. A. Mena Marugan, “Perturbative formalism of Lovelock gravity", *ibid*. **46**, 4320 (1992); 4340 (1992). D. G. Boulware and S. Deser, Phys. Rev. Lett., “String-Generated Gravity Models", **55**, 2656 (1985); J. T. Wheeler, “Symmetric solutions to the Gauss-Bonnet extended Einstein equations ", Nucl. Phys. **B268**, 737 (1986). J. T. Wheeler, “Symmetric solutions to the maximally Gauss-Bonnet extended Einstein equations ", Nucl. Phys. **B273**, 732 (1986). M. H. Dehghani and M. Shahmirzaie, “Thermodynamics of Asymptotic Flat Charged Black Holes in Third Order Lovelock Gravity", Phys. Rev. D, *accepted*, \[hep-th/0506227\]. D. L. Wiltshire, “Spherically symmetric solutions of Einstein-Maxwell theory with a Gauss-Bonnet term", Phys. Lett. B **169**, 36 (1986). R. G. Cai, “Gauss-Bonnet black holes in AdS spaces", Phys. Rev. D **65**, 084014 (2002) \[hep-th/0109133\]; M. Cvetic, S. Nojiri, S.D. Odintsov, “Black hole thermodynamics and negative entropy in de Sitter and anti-de Sitter Einstein Gauss-Bonnet gravity", Nucl. Phys. **B628**, 295 (2002) \[hep-th/0112045\]. M. H. Dehghani, “Charged rotating black branes in anti–de Sitter Einstein-Gauss-Bonnet gravity"’ Phys. Rev. D **67**, 064017 (2003) \[hep-th/0211191\]; “Magnetic branes in Gauss-Bonnet gravity ", *ibid*. **69**, 064024 (2004) \[hep-th/0312030\]; “Asymptotically (anti)-de Sitter solutions in Gauss-Bonnet gravity without a cosmological constant, *ibid*. **70**, 064019 (2004) \[hep-th/0405206\]. P. Hoxha, R. R. Martinez-Acosta, C. N. Pope, “Kaluza-Klein Consistency, Killing Vectors, and Kahler Spaces", Class. Quant. Grav. **17**, 4207 (2000) \[hep-th/0005172\].

ArXiv

--- abstract: 'The effect of quantum fluctuations on a nearly flat, nonrelativistic two-dimensional membrane with extrinsic curvature stiffness and tension is investigated. The renormalization group analysis is carried out in first-order perturbative theory. In contrast to thermal fluctuations, which soften the membrane at large scales and turn it into a crumpled surface, quantum fluctuations are found to [*stiffen*]{} the membrane, so that it exhibits a Hausdorff dimension equal to two. The large-scale behavior of the membrane is further studied at finite temperature, where a nontrivial fixed point is found, signaling a crumpling transition.' address: | Institut für Theoretische Physik\ Freie Universität Berlin\ Arnimallee 14, 14195 Berlin. author: - 'M. E. S. Borelli[^1], H. Kleinert[^2], and Adriaan M. J. Schakel[^3]' title: | Quantum Statistical Mechanics of Nonrelativistic Membranes:\ Crumpling Transition at Finite Temperature --- Introduction {#intro} ============ Quantum oscillations, analogous to the superconducting Josephson effect, have recently been detected in samples of superfluid $^3$He [@qm1; @qm2]. The apparatus involved in these experiments consists of an inner cell, filled with $^3$He, contained in an outer cell, also filled with $^3$He. The two containers are separated by a stiff membrane glued to the bottom of the inner cell, and by a softer one attached to its top. The lower membrane contains an array of small apertures allowing for exchange of atoms between the cells, equivalent to the superconducting weak link. By manipulating this membrane, the pressure between the two systems can be kept at a fixed value, and the resulting mass current is determined by the displacement of the upper membrane from its original position. Josephson current oscillations at the weak link are then observed as oscillations of this membrane. It is thus of interest to investigate whether the membranes themselves display quantum fluctuations, and to what extent such effects may be observable. Thermal fluctuations are known to soften fluid membranes, and to increase their surface tension [@helf3; @peliti; @klein1; @forster; @helf2; @forster2; @fluid; @cai]. At a temperature $T$, the membrane’s bare surface tension $r_0$ and bending rigidity $1/\alpha_0$ are renormalized at large scales as follows: $$\begin{aligned} \label{rold} r &=& r_0 \left[ 1 + \frac{1}{4 \pi} u_{{\rm th},0} \ln (\Lambda L ) \right], \\ \label{kappa} \frac{1}{\alpha} &=& \frac{1}{\alpha_0} \left[ 1 - \frac{3}{4 \pi} u_{{\rm th},0} \ln (\Lambda L) \right] ,\end{aligned}$$ where $$\label{uth} u_{{\rm th},0} = k_{\rm B} T \alpha_0$$ is the (dimensionless) expansion parameter in the regime dominated by thermal fluctuations, while $\Lambda$ is an ultraviolet wavenumber cutoff set by the inverse size of the molecules in the membrane, and $L$ is an infrared cutoff determined by its finite size. The renormalization group equation extracted from Eq. (\[kappa\]) yields only the trivial fixed point $\alpha=0$, implying the absence of a crumpling transition. The fixed point turns out to be unstable in the infrared. As a result, the renormalized coupling constant $\alpha$ flows away from it, so that the bending rigidity $1/\alpha$ tends to zero with increasing membrane size. The length scale where the right-hand side of Eq. (\[kappa\]) vanishes gives the so-called de Gennes-Taupin persistence length $\xi_{{\rm th},0} = \Lambda^{-1} \exp (4 \pi / 3 u_{{\rm th},0})$ [@DGT], above which the bending rigidity $1/\alpha$ vanishes. As a consequence, such membranes are crumpled at these scales, and fill the space in which they are embedded completely. Such objects are said to have infinite Hausdorff dimension $d_{\rm H} = \infty$. The full SO($d$) rotational symmetry of the $d$-dimensional embedding space is recovered in this way. For a smooth surface, this symmetry would be spontaneously broken to its SO($d-2$) $\times$ SO(2) subgroup. The absence of this symmetry breaking is in accord with the Mermin-Wagner theorem [@MW], stating that in a two-dimensional system long-range order is destroyed at any finite temperature. The results (\[rold\]) and (\[kappa\]) are derived from the Canham-Helfrich model [@canham; @helf1], describing stiff membranes subject to thermal undulations. The Hamiltonian reads: $${\cal H}_0 = \int {\rm d}S \left(r_0 \ + \frac{1}{2\alpha_0} H^2 \right), \label{helfmodel}$$ where ${\rm d}S$ is the surface element, and $H$ corresponds to (twice) the mean curvature of the surface at each point. We omit a term proportional to the Gaussian curvature since it is only important if fluctuations change the topology of the membrane, characterized by the Euler characteristic, i.e., by the number of handles—a possibility which we shall neglect. Each point on the membrane is represented by a vector ${\bf X}(\sigma_1,\sigma_2)$ in the $d$-dimensional bulk space, depending on two coordinates $\sigma_1$ and $\sigma_2$. Explicitly, the Hamiltonian (\[helfmodel\]) reads (for reviews, see Ref. [@Jerusalem]) $${\cal H}_0 =\int {\rm d}^2 \sigma \sqrt{g} \left[r_0 + \frac{1}{2 \alpha_0} (\Delta {\bf X})^2 \right] \label{helfmodel2},$$ where $$\label{gmunu} g_{a b}= \partial_{a}{\bf X} \cdot \partial_{b}{\bf X}$$ is the metric induced by the embedding and $g \equiv \det[g_{a b}]$. The symbol $\partial_a \, (a=1,2)$ denotes the derivative with respect to the coordinates $\sigma = (\sigma_1, \sigma_2)$, and $\Delta = g^{a b}D_a D_b$ is the scalar Laplacian, where $D_a$ is the covariant derivative associated with the metric. By lowering the temperature, the effect of thermal fluctuations decreases, and the behavior of the membrane becomes dominated by quantum effects. To account for these, we have to include time-dependence and add a kinetic term to the classical theory. For an incompressible surface, the kinetic term reads, in euclidean spacetime, $${\cal T} = \frac{1}{2 \nu_0} \int {\rm d}^2 \sigma \sqrt{g} \, \dot{\bf X}^2, \label{dynterm}$$ where $\bf X$ is now time-dependent, $1/ \nu_0$ is the bare mass density, and the dot indicates a time derivative. The partition function $Z$ can be represented as a functional integral over all possible surface configurations ${\bf X}(\sigma, \tau)$: $$\label{partfunc2} Z = \int {\cal D}{\bf X} \exp(-S_0[{\bf X}]/\hbar),$$ with the euclidean action $$\label{action} S_0 = \int {\rm d} \tau \, {\rm d}^2 \sigma \sqrt{g} \left[ \frac{1}{2\nu_0} \dot{\bf X}^2 + r_0 + \frac{1}{2 \alpha_0} (\Delta {\bf X})^2 \right]. \label{euclaction}$$ We shall in this paper study the quantum statistical mechanics of a membrane described by this action. Specifically, in Sec. [\[1loop\]]{}, we give a one-loop analysis of the effect of quantum fluctuations on such a two-dimensional surface embedded in $d=3$ space dimensions. We show that, contrary to thermal fluctuations, quantum fluctuations lead to a stiffening of the membrane at large scales. In the one-loop approximation, the flow equations are again found to yield only the trivial fixed point $\nu = r = \alpha =0$, implying the absence of a crumpling transition. However, this fixed point is now, in contrast to the one generated by thermal fluctuations, stable in the infrared. This means that $\alpha$ flows towards this fixed point and the bending rigidity tends to infinity with increasing membrane size, so that the surface remains two-dimensional. We further investigate the behavior of the system at finite temperature. We find at the one-loop order a nontrivial fixed point arising here. Being unstable in the infrared, it gives rise to a crumpling transition. Below a critical temperature $T_{\rm c}$, $\alpha$ flows to zero, implying a flat membrane, while above $T_{\rm c}$, the flow is towards infinity, implying a crumpled membrane. Perturbative Expansion {#1loop} ====================== We shall work in the Monge parametrization, where a point on the surface embedded in three-dimensional space is described by a displacement field $\phi({\bf \sigma},\tau)$ with respect to a reference plane ${\bf \sigma}=(\sigma_1,\sigma_2)$, such that $${\bf X}({\bf \sigma},\tau) = \left({\bf \sigma},\phi({\bf \sigma},\tau)\right) . \label{monge}$$ Inserting this parametrization in the action (\[action\]), and expanding the resulting expression up to fourth order in powers of the displacement field $\phi$, we obtain $$\begin{aligned} \label{pertmodel} S_0 &=& \int {\rm d} \tau {\rm d}^2\sigma \left\{ \frac{1}{2}\left[ \frac{1}{\nu_0} {\dot \phi}^2 + r_0 (\partial_a \phi)^2 + \frac{1}{\alpha_0} (\partial^2\phi)^2 \right] + \frac{1}{4 \nu_0} {\dot \phi}^2 (\partial_a \phi)^2 - \frac{r_0}{8} (\partial_a \phi )^2( \partial_b \phi)^2 \right. \\&& \nonumber \hspace{1.8cm} \left. - \frac{1}{4 \alpha_0} (\partial_a \phi)^2(\partial^2\phi)^2 - \frac{1}{\alpha_0} (\partial_a \phi )(\partial_b \phi)( \partial_a \partial_b \phi)(\partial^2\phi) \right \}.\end{aligned}$$ The displacement field describes the undulations of the surface. Its spectrum, $\omega^2 = c_{\rm s}^2 q^2$, is gapless with $c_{\rm s}^2 = r_0 \nu_0$ the velocity of the transversal waves. In the one-loop approximation, the exponent in (\[partfunc2\]) may be expanded up to second order around a background configuration $\Phi({\bf \sigma},\tau)$ extremizing $S_0$. The resulting integral is Gaussian and yields the effective action $$\label{effaction} S_{\rm eff}[\Phi] = S_0 [\Phi] + S_1 [\Phi] = S_0 [\Phi] + \frac{\hbar}{2} \mbox{Tr} \ln \left[\left. \frac{\delta^2 S_0}{\delta \phi({\bf \sigma},\tau) \delta \phi({\bf \sigma'},\tau')} \right|_{\Phi} \right],$$ where the expression in square brackets is a functional matrix given by the second functional derivative of $S_0$, and Tr denotes the functional trace, i.e., the integral $\int {\rm d} \tau \, {\rm d}^2 \sigma$, as well as the integral $\int {\rm d} \omega \, {\rm d}^2 q/(2 \pi)^3$ over the (angular) frequency $\omega$ and the wavevector ${\bf q}$. Using a derivative-expansion method due to Fraser [@caroline], we expand the one-loop correction $S_1$ in Eq. (\[effaction\]) in powers of the derivatives of the field $\phi({\bf \sigma},\tau)$, as in Ref.[@memb]. To obtain the renormalization of the parameters $r_0$, $\alpha_0$ and $\nu_0$, it suffices to keep only the first three terms of the expansion: $$\label{seff} S_1 = \frac{\hbar}{2} \int {\rm d}\tau \, {\rm d}^2\sigma \left[ I_1 {\dot \phi}^2 + I_2 (\partial_a \phi)^2 + I_3 (\partial^2\phi)^2 + \ldots \right],$$ with $$\begin{aligned} I_1 &=& \frac{1}{2\nu_0} \int \frac{{\rm d}\omega}{2\pi} \frac{{\rm d}^2q}{(2\pi)^2} \frac{q^2}{\omega^2/\nu_0 + r_0 q^2 + q^4/\alpha_0} \label{I1} \\ I_2 &=& \int \frac{{\rm d}\omega}{2\pi} \frac{{\rm d}^2q}{(2\pi)^2} \frac{\frac{1}{2} \omega^2/\nu_0 -r_0 q^2 - \frac{3}{2} q^4/\alpha_0}{\omega^2/\nu_0 + r_0 q^2 + q^4/\alpha_0} \label{I2} \\ I_3 &=& -\frac{3}{2} \frac{1}{\alpha_0} \int \frac{{\rm d}\omega}{2\pi} \frac{{\rm d}^2q}{(2\pi)^2} \frac{q^2}{\omega^2/\nu_0 + r_0 q^2 + q^4/\alpha_0}. \label{I3}\end{aligned}$$ After the integrals over the loop energy $\omega$ have been carried out, the resulting momentum integrals in Eqs. (\[I1\])–(\[I3\]) diverge in the ultraviolet. To regularize them, we introduce a wavenumber cutoff $\Lambda$. Contributions proportional to positive powers of $\Lambda$ are irrelevant, and will be ignored. In dimensional regularization, where the number $D$ of space dimensions of the membrane is analytically continued to be less than two, $D=2-\epsilon$, these powerlike divergences never appear in the first place. Only logarithmic divergences arise as poles in $1/\epsilon$. Substituting the results of the integration into Eq.(\[seff\]), we obtain the effective action $$\label{seff2} S_{\rm eff} = \frac{1}{2} \int {\rm d}\tau \, {\rm d}^2 \sigma \left[ \frac{1}{\nu} {\dot \phi}^2 + r (\partial_a \phi)^2 + \frac{1}{\alpha} (\partial^2\phi)^2 + \ldots \right],$$ with the renormalized inverse mass density $\nu$, surface tension $r$, and inverse rigidity $\alpha$ $$\begin{aligned} \frac{1}{\nu} &=& \frac{1}{\nu_0} \left[ 1 - \frac{1}{16 \pi}u_{{\rm qm},0} \ln\left(\Lambda \xi_{{\rm qm},0} \right) \right], \label{zetaeff} \\ r &=& r_0 \\ \frac{1}{\alpha} &=& \frac{1}{\alpha_0} \left[ 1 + \frac{3}{16\pi}u_{{\rm qm},0} \ln\left(\Lambda \xi_{{\rm qm},0} \right) \right], \label{kappaeff}\end{aligned}$$ where the dimensionless parameter $u_{{\rm qm},0} = \hbar r_0 \alpha_0^{3/2} \nu_0^{1/2}$ is the (bare) expansion parameter in the quantum regime. Note that the surface tension is not renormalized by quantum fluctuations at this order. The parameter $\xi_{{\rm qm},0} = 2/\sqrt{r_0 \alpha_0}$ in the argument of the logarithm in Eqs.(\[zetaeff\]) and (\[kappaeff\]) defines a characteristic length scale of the problem. It sets the scale at which the tension and stiffness terms in the action (\[seff2\]) become equally important. At larger scales, the second term in the expansion (\[seff2\]) becomes more important and the undulations are dominated by tension, while at smaller scales, the third term dominates and the undulations are controlled by stiffness. To obtain the flow equations, we apply Wilson’s procedure [@Wilson]. Integrating out a momentum shell $\Lambda/s < q < \Lambda$, rescaling the coupling constants $\hat{\nu} = s^{\epsilon-3} \nu, \;\; \hat{r} \to s^{-(\epsilon-3)} r, \;\; \hat{\alpha} \to s^{\epsilon-1} \alpha$, we arrive at $$\begin{aligned} \label{bnu} \beta_\nu({\nu},{r},{\alpha}) &=& s \frac{\partial \hat{\nu}}{\partial s} \Biggl|_{s=1} = (\epsilon-3) {\nu} + \frac{1}{16\pi} {u}_{\rm qm} {\nu}, \\ \beta_r({\nu},{r},{\alpha}) &=& s \frac{\partial \hat{r}}{\partial s} \Biggl|_{s=1} = -(\epsilon-3) {r}, \\ \beta_\alpha({\nu},{r},{\alpha}) &=& s \frac{\partial \hat{\alpha}}{\partial s} \Biggl|_{s=1} = (\epsilon -1) {\alpha} -\frac{3}{16\pi} {u}_{\rm qm} {\alpha}, \label{betakappa}\end{aligned}$$ where $\epsilon = 2-D$ is assumed to be small and will be set to zero at the end of the calculation. The coefficients of the first terms at the right-hand sides denote the scaling dimension of the scaling fields. For small $\epsilon$, the scaling fields ${\nu}$ and ${\alpha}$ are irrelevant, while ${r}$ is relevant. Criticality is obtained by setting the relevant fields to zero, i.e., ${r}=0$ in this case. Starting somewhere on the critical surface ${r}=0$, the system flows towards the trivial fixed point ${\nu}={\alpha}=0$. This guarantees the stiffness of the membrane at large scales. There is no crumpling transition at the absolute zero of temperature; the membrane is always flat, where the normal vectors to its surface are strongly correlated. More specifically, the correlation function between the normal vectors to the surface behaves at large scales as $$\label{algebraic} \langle \partial_a {\bf X}(\sigma,\tau) \cdot \partial_b {\bf X}(\sigma',\tau) \rangle \sim \frac{\delta_{a b}}{|\sigma - \sigma'|^3}.$$ This algebraic fall-off implies the absence of a persistence length which would define the length scale above which the normals become uncorrelated and the surface becomes crumpled. To investigate this further, let us calculate the Hausdorff dimension $d_{\rm H}$ of the membrane. It can be defined by the relation between its mean surface area $\langle A \rangle$, where $$A = \int {\rm d}^2 \sigma \sqrt{g},$$ and the frame, or projected area $A_0 = \int {\rm d}^2 \sigma$. This relation is $$\langle A \rangle \sim A_0^{d_{\rm H}/2}, \label{defhausd}$$ so that the Hausdorff dimension is given by $$d_{\rm H} = 2 \frac{\partial \ln \langle A \rangle}{\partial \ln A_0}. \label{hausdorff}$$ Since the frame area $A_0$ scales with the cutoff $\Lambda$ as $$A_0 = \int {\rm d}^2 \sigma \sim \Lambda^{-2},$$ Eq. (\[hausdorff\]) can also be written in the form $$\label{alternative} d_{\rm H} = - \frac{\partial \ln \langle A \rangle}{\partial \ln \Lambda}.$$ At the one-loop level, the mean surface area is $$\begin{aligned} \langle A \rangle &=& \left\langle \int {\rm d}^2 \sigma \left[ 1 + \frac{1}{2} (\partial_a \phi)^2 + \ldots \right] \right\rangle \nonumber \\ &=& A_0 \left( 1 + \frac{\hbar}{2} \int \frac{{\rm d}\omega}{2\pi} \frac{{\rm d}^2q}{(2\pi)^2} \frac{ q^2}{ \omega^2/\nu_0 + r_0 q^2 + q^4/\alpha_0} \right), \label{area}\end{aligned}$$ so that we obtain from relation (\[alternative\]) the Hausdorff dimension $$d_{\rm H} = 2 + \frac{1}{16 \pi} {u}_{\rm qm}.$$ For large membranes, ${u}_{\rm qm} \to 0$, implying a Hausdorff dimension $d_{\rm H} =2$. Expressed in group theoretic terms, the SO($d$) rotational symmetry of $d$-dimensional space is spontaneously broken to its SO($d-2$) $\times$ SO(2) subgroup. Since we are now at zero temperature, where the membrane has an extra (time) dimension, this spontaneous symmetry breaking does not violate the Mermin-Wagner theorem. The algebraic decay found in Eq. (\[algebraic\]) is an immediate consequence of this symmetry breaking and identifies the two resulting Goldstone modes. We next investigate how the high-temperature regime dominated by thermal fluctuations, where the membrane is found to be always crumpled, goes over into the low-temperature regime dominated by quantum fluctuations, where the membrane remains flat. To investigate this temperature dependence, we adopt the imaginary-time approach to thermal field theory [@Rivers; @Kapusta]. It can be derived from the corresponding (euclidean) quantum theory at zero temperature by restricting the euclidean time to the finite interval $0 \leq \tau \leq \hbar/ k_{\rm B}T$, and substituting $$\label{fun:sub} \int \frac{{\rm d} \omega}{2\pi}\,g(\omega)\rightarrow \frac{k_{\rm B}T}{\hbar} \sum_{n} \, g(\omega_{n}),$$ where $g$ is an arbitrary function, and $\omega_n$ denote the Matsubara frequencies, $$\label{fun:matb} \omega_n = 2 \pi n k_{\rm B}T/\hbar , \;\;\;\;\; n = 0, \pm 1, \pm 2 \cdots.$$ Using this substitution in the zero-temperature integrals (\[I1\])–(\[I3\]), as well as the formula [@gradstein] $$\frac{k_{\rm B} T}{\hbar} \sum_n \frac{1}{\omega_n^2 + a^2} = \frac{1}{2 a} \coth\left(\frac{\hbar a}{2 k_{\rm B} T}\right)$$ to carry out the sum over the Matsubara frequencies, we arrive at $$\begin{aligned} \label{nu} \frac{1}{\nu}&=& \frac{1}{\nu_0}\left [1 + \frac{1}{8 \pi} u_{{\rm qm},0} F_1(T,\Lambda') \right], \\ \label{r} r &=& r_0 \left[ 1 - \frac{1}{2 \pi} u_{{\rm qm},0} F_2(T,\Lambda') \right], \\ \label{alpha} \frac{1}{\alpha} &=& \frac{1}{\alpha_0}\left [1 - \frac{3}{8 \pi} u_{{\rm qm},0} F_1(T,\Lambda') \right],\end{aligned}$$ with $$\begin{aligned} F_1(T,\Lambda')&=& \int^{\Lambda'} {\rm d}q' \frac{q'^2 \coth ({{\textstyle\frac{1}{2}}} \gamma_0 q' \sqrt{1 + q'^2} )}{\sqrt{1 + q'^2}} \label{F2} , \\ F_2(T,\Lambda')&=& \int^{\Lambda'} {\rm d}q' \frac{ ({{\textstyle\frac{3}{4}}} q'^2 + q'^4) \coth({{\textstyle\frac{1}{2}}} \gamma_0 q' \sqrt{1 + q'^2})}{\sqrt{1 + q'^2}}. \label{F1}\end{aligned}$$ Here, we rescaled the integration variables: $q'=q \xi_{{\rm qm},0}/2$, $\Lambda' = \Lambda \xi_{{\rm qm},0}/2$, and $\gamma_0$ stands for $\gamma_0 = \hbar r_0 \alpha_0^{1/2} \nu_0^{1/2} /k_{\rm B} T $. This dimensionless parameter can be expressed as the ratio of the expansion parameters in the quantum and classical regime, $$\gamma_0 = u_{{\rm qm},0}/u_{{\rm th},0}.$$ The integrals in Eqs. (\[F1\]) and (\[F2\]) diverge in the ultraviolet as $\Lambda' \to \infty$. As before, we disregard powerlike divergences, and consider only the logarithmically diverging terms. We thus arrive at: $$\begin{aligned} \label{nu'} \frac{1}{\nu} &=& \frac{1}{\nu_0}\left [1 + \frac{1}{4 \pi} \left(u_{{\rm th},0} - \frac{1}{4} u_{{\rm qm},0} \right)\ln (\Lambda') \right], \\ \label{r'} r &=& r_0 \left[ 1 + \frac{1}{4 \pi} u_{{\rm th},0} \ln(\Lambda') \right], \\ \label{alpha'} \frac{1}{\alpha} &=& \frac{1}{\alpha_0}\left [1 - \frac{3}{4 \pi} \left(u_{{\rm th},0} - \frac{1}{4} u_{{\rm qm},0} \right) \ln (\Lambda') \right].\end{aligned}$$ One may check that as the temperature $T$ and, consequently, $u_{{\rm th},0} \propto T$ tend to infinity, Eqs. (\[r’\]) and (\[alpha’\]) reproduce, up to finite terms, the high-temperature results (\[rold\]) and (\[kappa\]). On the other hand, in the limit $u_{{\rm th},0} \propto T \to 0$, we recover the zero-temperature results (\[zetaeff\])–(\[kappaeff\]). In order to explore the behavior of the membrane at large length scales, we compute the flow equations corresponding to the three parameters of the theory, as we did above. They are given by: $$\begin{aligned} \label{betanu} \beta_\nu({\nu},{r},{\alpha}) &=& (\epsilon -3) {\nu} -\frac{1}{4 \pi}\left({u}_{\rm th} - \frac{1}{4} {u}_{\rm qm} \right) {\nu}, \\ \label{betar} \beta_r({\nu},{r},{\alpha}) &=& -(\epsilon -3){r} + \frac{1}{4 \pi} {u}_{\rm th} {r} , \\ \label{betaalpha} \beta_\alpha({\nu},{r},{\alpha}) &=& (\epsilon -1) {\alpha} + \frac{3}{4 \pi}\left({u}_{\rm th} - \frac{1}{4} {u}_{\rm qm} \right) {\alpha}.\end{aligned}$$ This system of equations admits two possible fixed points (see Fig.\[fig:flows\]), viz. the trivial one at ${\nu}={r}={\alpha}=0$ which we already found above at $T=0$, and a new one at ${\nu}={r}=0, {\alpha}={\alpha}^*$, with $$\label{fp} {\alpha}^*=\frac{4\pi}{3 k_{\rm B} T}(1-\epsilon).$$ =20.cm Note that this fixed point exists even for a two-dimensional membrane ($\epsilon=0$). The scaling field ${r}$ is found to be relevant, while ${\nu}$ is found to be irrelevant with respect to both fixed points. Criticality is obtained by setting ${r}=0$. The scaling field ${\alpha}$ behaves differently: it is irrelevant with respect to the trivial fixed point, but relevant with respect to the new one. The presence of the unstable fixed point implies the existence of a crumpling transition for a two-dimensional membrane at a critical temperature $$\label{Tc} T_{\rm c}= \frac{4\pi}{3 k_{\rm B}} \frac{1}{{\alpha}^*}.$$ For $T<T_{\rm c}$, $\alpha$ flows away from $\alpha^*$ and towards the trivial fixed point $\alpha=0$. The correlation between the normals to the surface is long-ranged, and the membrane remains flat. For $T>T_{\rm c}$, on the other hand, ${\alpha}$ flows away from ${\alpha}^*$ in the other direction, that is ${\alpha} \to \infty$. In this case, thermal fluctuations dominate. The correlation between the normals to the surface is short-ranged, and the membrane is found to be crumpled. As the temperature tends to infinity, the time dimension shrinks to a point, making the integration $\int {\rm d} \tau$ disappear from the action (\[action\]). This implies that the parameter $\epsilon$ in the flow equations (\[betanu\])–(\[betaalpha\]) must be set equal to $\epsilon=\epsilon' + 1$ [@schmeltzer; @stephens], with $\epsilon'=0$ for $D=2$. We see that the fixed point (\[fp\]) reduces in the limit $T \to \infty$ to the trivial one for a two-dimensional membrane. A similar nontrivial fixed point to the one in (\[fp\]) was found in Refs. [@peliti; @leibler] for a $(2+|\epsilon'|)$-dimensional membrane ($\epsilon' <0$) described by the classical Canham-Helfrich model embedded in $3+|\epsilon'|$ dimensions. In our case, due to the presence of the extra time dimension, the nontrivial fixed point (\[fp\]) exists also in two dimensions. Conclusions =========== We have analyzed the large-scale behavior of a membrane subject to thermal and quantum fluctuations. At the absolute zero of temperature, the membrane is found to be flat, and the vectors normal to the surface are found to be strongly correlated. This behavior, being qualitatively different from that at high temperatures, led us to investigate how these two regimes are related, by studying the system at finite temperature. We found that it exhibits a crumpling transition: above a critical temperature $T_{\rm c}$, thermal fluctuations turn the membrane into a crumpled surface, while for temperatures below $T_{\rm c}$, quantum fluctuations render the membrane flat. We thank M. C. Diamantini, V. Schulte-Frohlinde, and C. A. Trugenberger for useful discussions. [99]{} S. V. Pereverzev, A. Loshak, S. Backhaus, J. C. Davis, and R. E. Packard, Nature [**388**]{}, 449 (1997). R. W. Simmonds, A. Loshak, A. Marchenkov, S. Backhaus, S. Pereversev, S. Vitale, J. C. Davis, and R. E. Packard, Phys. Rev. Lett. [**81**]{}, 1247 (1998). , Proceedings of the Fifth Jerusalem Winter School for Theoretical Physics, 1987/1988, edited by D. Nelson, T. Piran, and S. Weinberg (World Scientific, Singapore, 1989). W. Helfrich, J. Phys. (Paris) [**46**]{}, 1263 (1985). L. Peliti and S. Leibler, Phys. Rev. Lett. [**54**]{}, 1690 (1985). H. Kleinert, Phys. Lett. A [**114**]{}, 263 (1986). D. Förster, Phys. Lett. A [**114**]{}, 115 (1986). W. Helfrich, J. Phys. (Paris) [**48**]{}, 285 (1987). D. Förster, Europhys. Lett. [**4**]{}, 65 (1987). H. Kleinert, J. Stat. Phys. [**56**]{}, 227 (1989). W. Cai, T. C. Lubensky, P. Nelson, and T. Powers, J. Phys. II France [**4**]{}, 931 (1994). P. G. de Gennes and C. Taupin, J. Phys. Chem. [**86**]{}, 2294 (1982). N. D. Mermin and H. Wagner, Phys. Rev. Lett. [**17**]{}, 1133 (1969). P. B. Canham, J. Theor. Biol. [**26**]{}, 61 (1970). W. Helfrich, Z. Naturforsch. [**28c**]{}, 693 (1973). C. M. Fraser, Z. Phys. C [**28**]{}, 101 (1985). M. E. S. Borelli, H. Kleinert, and A. M. J. Schakel, Phys. Lett. A [**253**]{}, 239 (1999). K. G. Wilson, Rev. Mod. Phys. [**47**]{}, 773 (1975). R. J. Rivers, [*Path Integrals in Quantum Field Theory*]{} (Cambridge University Press, Cambridge, 1987). J. I. Kapusta, [*Finite-Temperature Field Theory*]{} (Cambridge University Press, Cambridge, 1989). I. Gradstheyn and I. M. Ryzhik, [*Table of Integrals, Series and Products*]{} (Academic Press, Boston, 1980). S. Leibler, in Ref. [@Jerusalem]. D. Schmeltzer, Phys. Rev. B [**32**]{}, 7512 (1985). D. O’Connor and C. R. Stephens, Nucl. Phys. B [**360**]{}, 297 (1991). [^1]: E-mail: [email protected] [^2]: E-mail: [email protected]; web site: http://www.physik.fu-berlin/$\sim$kleinert [^3]: E-mail: [email protected]

ArXiv

--- abstract: | In this study, we propose methods for the automatic detection of photospheric features (bright points and granules) from ultra-violet (UV) radiation, using a feature-based classifier. The methods use quiet-Sun observations at 214 nm and 525 nm images taken by *Sunrise* on 9 June 2009. The function of region growing and mean shift procedure are applied to segment the bright points (BPs) and granules, respectively. Zernike moments of each region are computed. The Zernike moments of BPs, granules, and other features are distinctive enough to be separated using a support vector machine (SVM) classifier. The size distribution of BPs can be fitted with a power-law slope -1.5. The peak value of granule sizes is found to be about 0.5 arcsec$^2$. The mean value of the filling factor of BPs is 0.01, and for granules it is 0.51. There is a critical scale for granules so that small granules with sizes smaller than 2.5 arcsec$^2$ cover a wide range of brightness, while the brightness of large granules approaches unity. The mean value of BP brightness fluctuations is estimated to be 1.2, while for granules it is 0.22. Mean values of the horizontal velocities of an individual BP and an individual BP within the network were found to be 1.6 km s$^{-1}$ and 0.9 km s$^{-1}$, respectively. We conclude that the effect of individual BPs in releasing energy to the photosphere and maybe the upper layers is stronger than what the individual BPs release into the network. author: - 'M.  $^1$, H. $^1$, A. $^{2}$ S. $^{1}$' title: 'Automatic Method for Identifying Photospheric Bright Points and Granules Observed by *Sunrise*' --- Introduction ============ The solar surface is stochastically covered by small-scale features such as bright points (BPs) and granules (Sánchez Almeida *et al.,* 2004; de Wijn *et al.,* 2008; Riethmüller *et al.,* 2010) at times of both peak and minimum of the solar cycles (Berger *et al.,* 1995; Zhang *et al.,* 1998). Statistical properties of these features are important subjects of solar physics. BPs, first reported by Dunn and Zirker (1973), are linked to the emergence of small magnetic elements in the solar surface [@Rieth]. Statistically speaking, they cover approximately five percent of the solar surface in TiO 705.68 nm images [@Andic] and are well observed in the G band and 214-515 nm wavelengths (de Wijn *et al.,* 2008; Utz *et al.,* 2009; Riethmüller *et al.,* 2010). Following Berger *et al.* (1995), visible-light signatures of photospheric magnetic concentrations have been classified into four main groups: faculae, filigree, facular BPs, and network BPs. BPs highlight the emergence and evolution of magnetic structures on the solar surface and granulation pattern (Frazier and Stenflo, 1978; de Wijn *et al.,* 2008; Nordlund, Stein, and Asplund, 2009). Solar granulation is widely considered to be resulting from a convective turbulent process (Cloutman, 1979; Roudier and Muller, 1987). The granule characteristic scale is $\sim$ 1.35 arcsec$^2$, ranging widely from 0.3 to 4 arcsec$^2$ (Rusu, 2005; Robitaille, 2011). The average center-to-center granule spacing is about 1300 km with a wide range. A granular cell has a lifetime of about 8–15 minutes, and the typical vertical and horizontal velocities are approximately 1 km s$^{-1}$ and 2 km s$^{-1}$, respectively [@Rusu]. An extended review on solar surface magneto-convection is given by Stein (2012). Solar space missions (*e.g.*, TRACE, *Hinode, and Sunrise*) and ground-based telescopes for high-resolution solar observations (*e.g.*, SST, NST, and GREGOR) provide images of the solar surface using various filters and cadences. To optimize the statistical analysis of various kinds of solar phenomena, it is necessary to develop automatic detection techniques. One of the significant methods for granules segmentation and finding BPs is multiple-level tracking (MLT4) based on pattern recognition methods [@Bovelet1]. Kobel *et al.* (2010) employed another classification method to derive distinguishing properties of both BPs and faculae. Aschwanden (2010) has given an extensive overview of image processing-techniques and feature-recognition algorithms in solar data. In the present study, we propose an automatic detection technique for BPs and granules at 214 and 525.02 nm in *Sunrise* images, taken of the quiet Sun, using Zernike moments, region growing, mean shift segmentation, and SVM. The paper is organized as follows: first, the data reduction is discussed in Section \[data\]. Identification methods of BPs and granules by using region growing, extracting Zernike moments, SVM, and applying the mean shift procedure are explained in Section \[identification\]. The results are given in Section \[Res\]. Finally, the conclusions are presented in Section \[Conc\]. Data Analysis {#data} ============= The *Sunrise* balloon-borne solar observatory was launched on 8 June 2009 (Barthol *et al.,* 2011; Solanki *et al.,* 2010). After reaching the required height in the Earth’s atmosphere, the high-resolution images in the UV began to be recorded [@Solanki]. *Sunrise* has a resolution of about 100 km. This observatory is equipped with the *Sunrise* filter imager [@Gandorfer], an imaging magnetograph experiment [@Pillet], an image stabilization and light distribution unit, and a correlating wavefront sensor [@Berkefeld]. The *Sunrise Filter Imager* (SuFI) is a diffraction-limited filter imager whose focal effective length is 121 m and which works in five different wavelength bands. The *Imaging Magnetograph eXperiment* (IMaX) data produce 936$\times$936 pixels with a resolution of 0.15–0.18 arcsec for studying solar magnetic and velocity fields. IMaX uses a Zeeman triplet with a Landé factor g = 3 (in the FeI 525.02 nm line). The *Sunrise* together with its instruments provides four levels of data. Level-0 are raw data, level-1 data are fully reduced, including phase-diversity reconstruction using individual and averaged wavefront exerts on both level-2 and level-3 data, respectively. We used a sequence of 32 images (level-2 data) recorded on 9 June 2009 (14:00-15:00 UT) at 214 nm (Figure \[fig1\]A) and reconstructed by deconvolution using a modified Wiener filter and the point-spread function of the optical system that was derived from through calibration applying phase diversity [@Solanki]. The images were recorded using an array of 711$\times$1972 pixels, with a pixel size of 0.02 arcsec (both $x$ and $y$ directions). The exposure time and cadence at 214 nm were 30 s and 42 s, respectively. In these images, both BPs and granulation were clearly observable [@Rieth]. Since intergranular lanes were not clearly visible at 214 nm, we used a sequence of 40 co-temporal and co-spatial IMaX (level-2) data (Figure \[fig2\]A) to recognize the granules. For our data cubes (SuFI and IMaX), a subsonic filter was used to suppress the global solar oscillations (*e.g.* five-minute p-modes) by attenuating modulations with horizontal speeds above 5 km s$^{-1}$. Identification of Photospheric Features {#identification} ======================================= Bright Points {#BP} ------------- It has been found that the photospheric BPs have the highest brightness contrast in 214 nm images (Riethmüller *et al.,* 2010). Riethmüller *et al.* (2010) have shown that sometimes the BP intensity reaches up to five times of the quiet-Sun mean intensity at this wavelength. Here, we present an automatic method for identifying BPs in images using Zernike moments and an SVM classifier. The following steps are employed: Pixels with an intensity slightly lower than twice the mean intensity of the images are marked. These marked pixels are candidates for a final seed point. The function of region growing is exerted on the seed points. Region growing is an approach that on the basis of the brightness-threshold criterion makes it possible to group and segment the image pixels in large regions. Adjacent pixels of each seed that share certain characteristics with the seed join the initial seed points, and subsequently, the region grows ($R_i$). The growth process terminates when all of the marked pixels from the previous step survive. In the final segmented images, the pixels of each region are separated (*i.e.*, $R_i\cap R_j=\emptyset$, $i\neq j$) [@Gonzalez]. The intensity-weighted centroid, ${\bf r}_{CI}=\sum{{\bf r}_{i}I({\bf r}_i)}/ \sum I({\bf r}_i)$, or the region’s pixel nearest to the intensity-weighted centroid is selected as the final seed point. In the output of region growing, each region is labeled by a different integer number based on the Bwlabel function [@Haralick]. The output of the region-growing function on 214 nm image is shown in Figure \[fig1\]B. These identified regions are grouped into the three main classes (individual BP, network BP, and non-BP). In each training step, we probed (investigated) the data to select individual and network BPs (manually). The criteria for selecting the individual BPs (class 1) are their shapes, which would not be elongated and crescent-like structures located in the intergranular lanes compared with co-spatial and co-temporal 525 nm images. The size of these selected BPs often range from $8\times8$ pixels to $20\times20$ pixels. The mean value of their equivalent diameter is $\approx 0.3\arcsec$. The regions with the smaller or larger size are classified into two different classes: network BPs elongated in intergranular lanes, class 2 (Figure \[fig1\]B, red rectangular boxes), and non-BPs, class 3 (Figure \[fig1\]B, yellow circles). The Zernike moments of these three classes of image tiles are calculated, see Section \[ZERM\]. The normalized magnitudes of moments are given to the SVM classifier. This classifier, the robust supervised machine-learning algorithm, was designed based on the structural-risk minimization [@Vapnik]. The purpose of this basically binary classifier is to find a decision boundary with a margin as large as feasible to reduce the classification error (Qu *et al.,* 2003; Theodoridis and Koutroumbas, 2009). Then, the classifier was applied to all labeled regions in all images. The code picks up a label 1 for an individual BP class, 2 for a network BP class, and 3 for a non-BP. When the individual BP or the network BP are classified, the size (number of pixels) and locations of the pixels are saved. Figure \[fig1\]C shows the individual BPs and network BPs. Granules -------- To detect granules, we used the 525.02 nm images. These images are segmented based on the mean shift procedure. This procedure is an old pattern-recognition approach in non-parametric feature-space analysis technique (Duda and Hart, 1973; Fukunaga and Hostetler, 1975). This feature-space analysis is the most useful and applicable tool for discontinuity-preserving smoothing, image filtering, and image segmentation. Discontinuity-preserving smoothing techniques were adapted to reduce both the smoothing amount near the abrupt changes at the edges and to remove the noise after a sufficiently many iterations. The mean shift image segmentation is a straightforward extension of the discontinuity-preserving smoothing algorithm [@Comaniciu]. In Figure \[fig2\]B, the segmented image is delineated. Similar to the BP training step (Section \[BP\]), the segmented regions are divided into two classes (granules and non-granular regions). The Zernike moments of these two classes (class 1, 2) of the image are computed and fed to the SVM classifier. The code is run for all images and picks up a label 1 for a granular class and 2 for a non-granular class. If it is a granule, the size and locations of pixels are saved. In Figure \[fig2\]C, the identified regions are presented. Zernike moments {#ZERM} --------------- During the past decades, various moment functions (Hu, Legendre, Zernike, etc) have been proposed to describe images according to their ability to represent the image features based on orthogonal polynomials (Hu, 1962; Teague, 1980). The Zernike polynomials form a complete set of orthogonal polynomials over the interior of the unit circle, $x^2+y^2\le1$. The Zernike polynomial is given by $$\begin{aligned} &&V_n^m=R_n^m(r)\exp(im\theta),\label{zernikeeq} \\ &&R_n^m(r)=\sum_{s=0}^{(n-m)/2}(-1)^s\frac{(n-s)!}{s!\left(\frac{n+|m|}{2}-s\right)!\left(\frac{n-|m|}{2}-s\right)!}r^{n-2s},\nonumber\end{aligned}$$ where the positive integer $n$ denotes the order number, and the repetition integer number $m$ that satisfies the constraint $n - |m|$ is even, and $|m| \leqslant n $ (Abandah and Anssari, 2009; Hosney, 2010). Using the Zernike polynomials, the image intensity function $I(x,y)$ is expanded as $$\begin{aligned} I(x,y)=\sum_{n=0}^{n_{up}}\sum_{m=0}^{n}Z_{nm}V_n^m(x,y).\label{zm}\end{aligned}$$ The Zernike moments $Z_{nm}$ are obtained as $$\begin{aligned} Z_{nm}=\frac{n+1}{2}\int\int_{\rm unit~ circle}V_n^{m*}(x,y)I(x,y)dxdy.\end{aligned}$$ The orthogonal properties of Zernike polynomials enable the contribution of each Zernike moment $Z_{nm}$ to be unique and independent of the information of an image. Thus, the Zernike moments for each feature are unique [@Hosney]. We calculated the Zernike moments of a maximum certain order up to $n_{up}=31$. This gives a $(n,m)$ set, which is labeled from 1 to 528. By testing different orders of Zernike moments, it was empirically realized that $n_{up}=31$ maintained the quality of the reconstructed image (BP, granule, etc) with both polygonal and simple shapes when compared with the original one based on the normalized reconstruction error [@Teh]. Figure \[fig3\] displays one BP (upper panel) and one granule (lower panel) with their reconstructed images and plots of the normalized reconstruction errors as a function of order number. For both the reconstructed BP and granule, the normalized reconstruction error decreases with increasing order number and approaches 0.2 after $n\approx9$. Empirically, the order number $n=31$ is chosen for the fast computations of the Zernike moments to support the convergence and robustness of the SVM classifier. The main question is how Zernike moments extract the information of the features with various $n$ and $m$ numbers. For a unit circle ($x^2+y^2 \le 1$) and $R_{nm}$ (Equation (\[zernikeeq\])), the weighted function becomes $r|R_{nm}(r)| \le r$. The square of the image pixels that were mapped into the unit circle are closer to the circle perimeter and have more weight than the pixels located near the origin. The higher the order and repetition numbers, the more the Zernike functions oscillate. This gives us much confidence to describe the image based on corresponding moments. To obtain scale and translation invariance, two methods are proposed: normalizing the image with respect to the regular geometrical moments before running the Zernike program [@Khotanzad], and making radial Zernike moments invariant to scale and translation without the need to use the regular moments (Belkasim, Hassan, and Obeidi, 2004). We adopted the second method. The magnitude of the Zernike moments is invariant under rotation by nature. The pixels of image tiles are mapped into the unit circle. The Zernike moments of two individual BPs, two network BPs, and two non-BPs are shown in Figure \[fig4\]. The Zernike moments of two granules and two non-granular regions are presented in Figure \[fig5\]. As shown in these figures, the magnitude values of the Zernike moments are clearly different for BPs and non-BPs. The BPs have a well-defined block structure. For non-BP features, we see some block structures, but they are different from those of the BPs. These differences give us much confidence in applying the SVM classifier to identify the photospheric features from in the 214 nm images. We created some artificial data (mimicking BPs) to describe the behavior of the Zernike moments. Artificial BPs are generated by using a two-dimensional Gaussian function, $$\begin{aligned} &&B(x,y)=A\exp\left(-(a(x-x_0)^2 +2b(x-x_0)(y-y_0) + c(y-y_0)^2)\right),\\ &&~~~~~~~~~~~~~~~~~~~~~a=\cos^2\theta/2\sigma_x^2 + \sin^2\theta/2\sigma_y^2,\nonumber\\&&~~~~~~~~~~~~~~~~~~~~~b=-\sin2\theta/4\sigma_x^2 + \sin2\theta/4\sigma_y^2,\nonumber\\&& ~~~~~~~~~~~~~~~~~~~~~c=\sin^2\theta/2\sigma_x^2 + \cos^2\theta/2\sigma_y^2, \nonumber\end{aligned}$$ where $A$ is the amplitude, $x_0$, $y_0$ are the positions of the center of the peak, $\sigma_x$, $\sigma_y$ are controlled width (FWHM) of the BP, and $\theta$ is the angle measured clockwise. In Figure \[fig6\], the Zernike moments of artificial BPs (for different amplitudes and widths) are computed. The structures of the Zernike moments for faint artificial BP ($A=10$) and bright artificial BP ($A=100$) are the same. The structures of moments with different widths ($\sigma_x=\sigma_y=2.5, 3, 3.5, 4$) do not change significantly. As expected, our calculations show that for $\sigma_x\neq\sigma_y$, with changing rotation angles, the structure of the moments are the same. Furthermore, the Zernike moments of the overlapping artificial BPs and mimicking network BPs with different arrangements appeared as similar block structures (Figure \[fig7\]). Additionally, the Zernike moments with higher order numbers are sensitive to noise (*e.g.*, Teh and Chin, 1988). Both our training and test (for BPs and granules) sets that include noisy image tiles change the structure of the moments so slightly that SVM classifier is able to distinguish them. Results {#Res} ======= By applying our automatic detection to SuFI images, about 1000 BPs were recognized. There are two groups of BPs that have been attained by our code: about 75% are individual BPs, most of which can be mapped on a square with an equivalent diameter of $\approx 0.3\arcsec$ ; the remaining 25% are network BPs. The size distribution of BPs and their filling factor (coverage area in each image) are shown in Figure \[fig8\]. The power-law distribution ($N\propto A^{-\alpha}$) is fitted with a power exponent $\alpha\approx1.5$ (Figure \[fig8\], upper panel). The filling factor of BPs per image with mean values of 0.01 is presented (Figure \[fig8\], lower panel). The brightness fluctuations and the scatter plot of BPs are shown in Figure \[fig9\]. The brightness fluctuations relative to the intensity of the photosphere are given by $$\label{fluct} I_{\rm fluc}=\sqrt{\frac{\sum(I-\bar{I})^2}{N\bar{I}^2}},$$ where $\bar{I}$ is the mean intensity of the photosphere and $N$ is the number of pixels (Roudier and Muller, 1987; Yu *et al.,* 2011). As we see, the brightness fluctuations of BPs (Figure \[fig9\], left panel) are larger than 1.1. As expected, the BP regions are brighter than their surrounding areas. The average of the brightness fluctuations in SuFI images is about 0.1. Increasing the size of BPs will subsequently increase the scatter in their average brightness (Figure \[fig9\], right panel). For network BPs ($\geq$ 0.2 arcsec$^2$), the brightness is more widely scattered. The latter could be related to the magnetic pressure inside flux tubes, combined with the lateral extension, inclination angle, and opacity. An individual BP and an individual BP within the network were tracked in seven time steps (each 234 second long). They were included in the image tiles and the function of region growing was exerted on them. We used the function align\_cube\_correl.pro, which is available in the SSW/IDL package, to coalign the images in the sequence. The intensity-weighted centroid was then computed for a sequence of the segmented image tiles (Figure \[fig10\]). The horizontal velocity of the centroid, $v_{CI}=\Delta{\bf r}_{CI}/\Delta t$, for both consecutive frames (with $\Delta t\simeq39$ seconds) was calculated and the mean value of velocity was derived (see, *e.g.*, Jafarzadeh *et al.,* 2013). The mean horizontal velocities of the individual BP and the individual one integrated within the network BP are $v_{_{\rm\small{IBP}}}$=1.6$\pm$ 0.3 km s$^{-1}$ and $v_{_{\rm\small{IBPN}}}$=0.9$\pm$ 0.5 km s$^{-1}$, respectively. These results are consistent with previous studies (1 - 2 km s$^{-1}$) obtained by Möstl *et el.* (2006) and Jafarzadeh *et al.* (2013). The excitation of the kink waves in the flux tubes of BPs might be related to horizontal motions (Choudhuri, Auffret, and Priest, 1993; Muller *et al.,* 1994; Wellstein, Kneer, and von Uexküll, 1998; Jafarzadeh *et al.,* 2013). The energy flux, carried by the excitation of the kink waves, is proportional to the square of the horizontal velocity, lifetime, filling factor, scale height and density of BPs. The proportion of the energy flux of the individual BP ($E_{_{\rm\small{IBP}}}$) to the network BP ($E_{_{\rm\small{IBPN}}}$) with almost the same filling factors and lifetimes is given by $$\label{Energy} \frac{E_{_{\rm\small{IBP}}}}{E_{_{\rm\small{IBPN}}}}=\frac{v_{_{\rm\small{IBP}}}^2}{v_{_{\rm\small{IBPN}}}^2}=3.16\pm2.32.$$ It shows that the contribution of an individual BP to the transfer of the energy to the upper layers is more than that of an individual BP into the network. Figure \[fig11\] illustrates the size distribution of the granules and their filling factor (coverage area in each image). The peak value of the size distribution is held on 0.5 arcsec$^2$ with an equivalent diameter of 0.8 arcsec (Figure \[fig11\], left panel). The two-parameter lognormal distribution with $\sigma=0.7$ and $\mu=5.9$ is fitted. The equivalent diameter is often concentrated on an interval of 0.6–2.5$\arcsec$. The mean value of the filling factors of the granules equals 0.51 (Figure \[fig11\], right panel). The difference between the filling factor in image number 21 and the total average was found to be 7%. Our computations show that the average brightness of these two images is about 8% lower than the overall average (average brightness of 40 images). The brightness fluctuations and the scatter plot of granules are shown in Figure \[fig12\]. The mean value of the brightness fluctuations of the granules (Figure \[fig12\], left panel) is about 0.22. The granules with areas $\le$2.5 arcsec$^2$ with an equivalent diameter shorter than 1.7 arcsec have excessive scatter in brightness (with large brightness fluctuations) and this scatter in brightness decreases with increasing granule size (Figure \[fig12\], right panel). For large granules, the brightness approaches unity. To estimate the reliability of the algorithm, we visually inspected 200 of the detected BPs. We found that 18 of them were not clear BPs. Using a constant threshold for all of the images seemed to be the cause of erroneous detections. In fact, background brightness, noises, etc, are different from image to image. The invariant Zernike moments can detect BPs even if they are small or have variable brightness (Alipour, Safari, and Innes, 2012). Because of the above-mentioned reasons, false-positive detections appeared when BPs are recognized. As we know, BPs are ubiquitously observed in intergranular lanes and their locations are correctly identified. We compared recognized BPs in SuFI images with co-temporal and co-spatial IMaX magnetograms (Figure \[fig13\]). As shown in Figure \[fig13\]B, the identified BPs show positive and/or negative polarities in the magnetogram (Figure \[fig13\]A, red box). The recognition of the granules has an efficiency slightly better than 85%, but regarding the intergranular regions, we obtain almost 80% of the answers correctly. The error in the latter result arises because of the heterogeneous shapes of the intergranular lanes; consequently, as a result of an oversegmentation in these areas, distinguishing between the regions would be somewhat difficult for the classifier. Conclusions {#Conc} =========== We presented an automated detection and characterization of photospheric BPs and granules. Image-processing methods and a machine-learning algorithm were applied to recognize and determine the physical properties of these features. The processes of two kinds of image segmentation and the invariant Zernike moments were employed. The structure of the Zernike moments for both faint and bright features with different sizes and various orientations are distinctive enough (Figures 6 and 7) to be identified using the SVM. The entire computing time has taken less than one hour for each image. BPs identified by our algorithm were located in the intergranular regions, as a comparison with IMaX data revealed. If the emergence and existence of BPs is taken as a criterion for the level of magnetic activity, it can be suggested that these images represented the quiet-Sun because the filling factor of BPs in such data is much lower than 5% (Andic *et al.,* 2011). With respect to the set of horizontal velocities of the photospheric BPs, it seems that energy transfer to the upper layers frequently returns to the individual BPs rather than the individual BPs within the network. Apparently, two regimes govern the relationship between the granules brightness and their sizes. Granules with sizes smaller than 2.5 arcsec$^2$, the critical scale, are more scattered in brightness, and for larger granules, the brightness approaches unity. This behavior of small granules can be attributed to the nature of their turbulent eddies, while large granules are considered as convective elements (Roudier and Muller, 1987; Yu *et al.,* 2011). Our code is extendable for the detection of both surface and atmospheric solar features (*e.g.*, sunspots, flares, and CMEs) observed by the ground-based telescopes (*e.g.*, SST, NST, GREGOR; and with future instruments ATST, EST, Indian NLST). Finally, the algorithm will be improved to track BPs and the granules in a sequence of images. The German contribution to *Sunrise* is funded by the Bundesministerium für Wirtschaft und Technologie through the Deutsches Zentrum für Luft- und Raumfahrt e.V. (DLR), Grant No.50 OU 0401, and by the Innovationsfonds of the President of the Max Planck Society (MPG). The Spanish contribution has been funded by the Spanish MICINN under projects ESP2006-13030-C06 and AYA2009-14105-C06 (including European FEDER funds). HAO/NCAR is sponsored by the National Science Foundation, and the HAO Contribution to *Sunrise* was partly funded through NASA grant number NNX08AH38G. The authors thank Julian Blanco and Valentin Martínez Pillet for providing access to IMaX data and Sami K. Solanki for helpful comments and discussions. The authors thank the anonymous referee for meticulous comments and suggestions. Abandah, G., Anssari, N.: 2009, [*Journal of Computrer Science 5*]{}, **3**, 226. Aipour, N., Safari, H., Innes, D. E.: 2012, [ [*Astrophys. J.* ]{}]{}**746**, 12. Andic, A., Chae, J., Goode, P. R., Cao, W., Ahn, K., Yurchyshyn, V., Abramenko, V.: 2011, [ [*Astrophys. J.* ]{}]{}**731**, 29. Aschwanden M. J.: 2010, [[*Solar Phys.* ]{}]{}**262**, 235. Barthol, P., Gandorfer, A., Solanki, S. K., Schüssler, M., Chares, B., Curdt, W., *et al.*: 2011, [[*Solar Phys.* ]{}]{}**268**, 1. Belkasim, S., Hassan, E., Obeidi, T.: 2004, Computer and Information Technology, **790**, 4. Berger, T. E., Scherijver, C. J., Shine, R. A., Tarbell, T. D., Title, A. M., Scharmer, G.: 1995,[ [*Astrophys. J.* ]{}]{}**454**, 531. Berkefeld, T., Schmidt, W., Soltau, D., Bell, A., Doerr, H. P., Feger, B., *et al.*: 2011, [[*Solar Phys.* ]{}]{}**268**, 1. Bovelet, B., Wiehr, E.: 2007, [[*Solar Phys.* ]{}]{}**243**, 121. Choudhuri, A. R., Auffret, H., Priest, E. R.: 1993, [[*Solar Phys.* ]{}]{}**143**, 49. Cloutman, L. D.: 1979, [ [*Astrophys. J.* ]{}]{}**227**, 614. Comaniciu, D., Meer, P.: 2002, [*IEEE Trans. Pattern Anal. Machine Intell.*]{} **24**, 603. de Wijn, A. G., Stenflo J. O., Solanki, S. K., Tsuneta, S.: 2008, [ [*Space Sci. Rev.* ]{}]{}**144**, 275. Duda, R. O., Hart, P. E.: 1973, [*Pattern Classification and Scene Analysis.*]{}, Wiley. Dunn, R. B., Zirker, J. B.:1973, [[*Solar Phys.* ]{}]{}**33**, 281. Frazier, E. N., Stenflo, J.O.: 1978, [ [*Astron. Astrophys.* ]{}]{}**70**, 789. Fukunaga, K., Hostetler, L. D.: 1975, [*IEEE Trans. Information Theory.*]{} **21**, 32. Gandorfer, A., Grauf, B., Barthol, P., Riethmüller, T. L., Solanki, S. K., Chares, B., *et al.*: 2011, [[*Solar Phys.* ]{}]{}**268**, 1. Gonzalez, R. C., Woods, R. E.: 2002, [*Digital Image Processing*]{}, 2nd edn., Pearson Prentice Hall, Upper Saddle River. Haralick, R. M., Linda, G. S.: 1992, Computer and Robot Vision, Volume I, Addison-Wesley, pp. 28-48. Hosney, K. M.: 2010, [*Information Sciences*]{}, **180**, 2299. Hu, M. K.: 1962, [*IRE Trans. Inf. Theory*]{}, **IT-8**, 179. Jafarzadeh, S., Solanki, S. K., Feller, A., Lagg, A., Pietarila, A., Danilovic, S., *et al.*: 2013, [ [*Astron. Astrophys.* ]{}]{}**549**, 116. Khotanzad, A., Hong, Y. H.: 1990, [*IEEE Trans. Pattern Anal. Machine Intell.*]{} **12**, 5. Kobel, P., Hirzberger, J., Solanki, S. K., Gandorfer, A., Zakharov, V.: 2010, [ [*Astron. Astrophys.* ]{}]{}**502**, 303. Martínez Pillet, V., del Toro Iniesta, J. C., Álvarez-Herrero, A., Domingo, V., Bonet, J. A., González Fernández, L., *et al.*: 2011, [[*Solar Phys.* ]{}]{}**268**, 1. Möstl, C., Hanslmeier, A., Sobotka, M., Puschmann, K., Muthsam, H. J.: 2006, [[*Solar Phys.* ]{}]{}**237**, 13. Muller, R., Roudier, Th., Vigneau, J., Auffret, H.: 1994, [ [*Astron. Astrophys.* ]{}]{}**283**, 232. Nordlund, Å., Stein, R. F., Asplund, M.: 2009, [*Living Rev.*]{} [[*Solar Phys.* ]{}]{}**6**, 2. Qu, M., Shih, F. Y., Jing, J., Wang, H.: 2003, [[*Solar Phys.* ]{}]{}**217**, 157. Riethmüller, T. L., Solanki, S. K., Martínez Pillet, V., Hirzberger, J., Feller, A., Bonet, J. A.,*et al.*: 2010, [ [*Astrophys. J. Lett.* ]{}]{}**723**, 169. Robitaille, P. M.: 2011, [*Progress in Physics*]{}, **3**, 79. Roudier, T., Muller, R.: 1987, [[*Solar Phys.* ]{}]{}**107**, 11. Rusu, M.: 2005, [*PADEU,*]{} **15**, 59. Sánchez Almeida, J., Márquez, I., Bonet, J. A., Domínguez Cerdeña, I., Muller, R.: 2004, [ [*Astrophys. J.* ]{}]{}**609**, 91. Solanki, S. K., Barthol, P., Danilovic, S., Feller, A., Gandorfer, A., Hirzberger, J., *et al.*: 2010, [ [*Astrophys. J.* ]{}]{}**723**, 127. Stein, R. F.: 2012, [*Living Rev.*]{} [[*Solar Phys.* ]{}]{}**9**, 4. Teague, M. R.: 1980, [*J. Opt. Soc. Am.*]{}, **70**, 920. Teh, C., Chin, R. T.: 1988, [*IEEE Trans. Pattern Anal. Machine Intell.*]{} **10**, 4. Theodoridis, S., Koutroumbas, K. : 2009, [*Pattern Recognition*]{}, 4th edn., Academic Press, Elsevier Inc. Utz, D., Hanslmeier, A., Möstl, C., Muller, R., Veronig, A., Muthsam, H.: 2009, [ [*Astron. Astrophys.* ]{}]{}**498**, 289. Vapnik, V. N.: 1995, [*The nature of statistical learning theory.*]{} Berlin Heidelberg, New York: Springer-Verlag. Wellstein, S., Kneer, F., von Uexküll, M.: 1998, [ [*Astron. Astrophys.* ]{}]{}**335**, 323. Yu, D., Xie, Z., Hu, Q., Yang, S., Zhang, J., Wang, J.: 2011, [ [*Astrophys. J.* ]{}]{}**743**, 58. Zhang, H., Scharmer, G., Lofdahl, M., Yi, Z.: 1998, [[*Solar Phys.* ]{}]{}**183**, 283.

ArXiv

--- abstract: 'We determined the orbits and stellar parameters of three eccentric eclipsing binaries by light curve solutions of their *Kepler* data. KIC 10992733 and KIC 5632781 undergo total eclipses while KIC 10026136 reveals partial eclipses. The components of the targets are G and K stars. KIC 10992733 exhibited variations which were attributed to variable visibility of spot(s) on asynchronously rotating component. KIC 5632781 and KIC 1002613 reveal tidally-induced features at periastron, i.e. they might be considered as eclipsing heartbeat stars. The characteristics of the periastron features (shape, width and amplitude) confirm the theoretical predictions.' address: 'Department of Physics and Astronomy, Shumen University, 115 Universitetska, 9700 Shumen, Bulgaria' author: - Diana Kjurkchieva and Doroteya Vasileva title: 'LIGHT CURVE SOLUTIONS OF THE ECCENTRIC BINARIES KIC 10992733, KIC 5632781, KIC 10026136 AND THEIR OUT-OF-ECLIPSE VARIABILITY' --- binaries: eclipsing – methods: data analysis – stars: fundamental parameters – stars: individual (KIC 10992733, KIC 5632781, KIC 10026136) INTRODUCTION ============ The eccentric binaries are important objects for the modern astrophysics because they present probes for study of tidal interaction. The gradual loss of energy due to tidal forces leads to circularization of the orbit and the synchronization of the rotation of these stars with their orbital motion (Regos et al. 2005). Recently we are witnesses of the next stage of development of the tidal interaction theory and testing of its predictions. The unprecedented high-accuracy *Kepler* observations (Borucki et al. 2010, Koch et al. 2010) allowed to discover and investigate new, fine, tidally-induced effects predicted by Kumar et al. (1995): light feature at the periastron and tidally-excited oscillations (Welsh et al. 2011, Fuller $\&$ Lai 2011, Burkart et al. 2012, Thompson et al. 2012, Nicholls $\&$ Wood 2012, Hambleton et al. 2013a). The newly discovered objects were called ”heartbeat” (HB) stars. They provide important tests for the stellar astrophysics and information about the stellar interiors (Hambleton et al. 2013b). The number of HB stars found in the *Kepler* data gradually increases (Kirk et al. 2016). Some of them have been objects of follow-up spectroscopy which shows a good agreement between the spectroscopic and photometric orbital elements (Smullen $\&$ Kobulnicky 2015, Shporer et al. 2016, Dimitrov et al. 2017). Last years we also found 12 HB stars between several samples of eclipsing *Kepler* systems (Kjurkchieva $\&$ Vasileva 2015, 2016, Kjurkchieva et al. 2016a, b). The goal of this study is determination of the orbits and physical parameters of three eccentric binaries, KIC 10992733, KIC 5632781 and KIC 10026136, as well as searching for tidally-excited effects. Table 1 presents available information for the targets from the EB catalog (Prsa et al. 2011, Slawson et al. 2011, Kirk et al. 2016). Kepler ID $P$ \[d\] $T_{m}$ \[K\] $w_1$ $w_2$ $d_1$ $d_2$ $\varphi_{2}$ ----------- ----------- --------------- -------- -------- -------- -------- --------------- 10992733 18.525 5274 0.0113 0.0153 0.405 0.268 0.72 5632781 11.025 5786 0.04 0.0457 0.395 0.376 0.66 10026136 9.08 6292 0.0361 0.0532 0.3124 0.2529 0.65 {width="13.5cm"} {width="13.5cm"} {width="13.5cm"} Light curve solutions ===================== The modeling of *Kepler* data (Archive MAST) was carried out by the package <span style="font-variant:small-caps;">PHOEBE</span> (Prsa $\&$ Zwitter 2005). We used its mode ”Detached binaries” because our targets are detached systems, which is clearly true from their light curves (Fig. 1). Long cadence (LC) data from many quarters are available in the *Kepler* archive for these binaries. We worked with the live version of the *Kepler* EB catalog (http://keplerebs.villanova.edu/). To ignore the effect of accidental light fluctuations in the procedure of the light curve solutions and to obtain adequate configuration parameters we modelled all available LC photometric data (above 50000 points for each target) after appropriate phase binning. For this aim we applied the PHOEBE option for binning of data. We used 1000 bins in phase (the smaller value leads to loss of information from the eclipses). The modeling procedure is described in Kjurkchieva $\&$ Vasileva (2015). Shortly it consists of several steps: fixing $T_1$ to target temperature $T_m$; preliminary calculation of the eccentricity *e* and periastron angle $\omega$; varying of *e* and $\omega$ to fit the eclipse phases; varying of the inclination *i*, mass ratio *q*, secondary temperature $T_{2}$ and potentials $\Omega_{1,2}$ to reproduce the whole curve; adjusting the real stellar temperatures $T_{1}^f$ and $T_{2}^f$ around the target value $T_{m}$ by the formulae (Ivanov et al. 2010): $$T_{1}^f = T_m + s \frac{\Delta T}{s+1},$$ $$T_{2}^f = T_1 - \Delta T,$$ where $s=l_2/l_1$ (ratio of relative luminosities) and $\Delta T = T_m-T_2$ were taken from the last <span style="font-variant:small-caps;">PHOEBE</span> fitting. Kepler ID 10992733 5632781 10026136 ------------------ ----------- ----------- ---------- *e* 0.380 (1) 0.273(2) 0.296(1) $\omega$ \[deg\] 25.34 (1) 22.5(1) 35.95(1) *i* \[deg\] 89.92 (6) 89.2(2) 87.33(2) *q* 0.509(3) 0.97(1) 0.42(1) $T_2$ \[K\] 4890(24) 5739(20) 6193(20) $\Omega_1$ 32.8(2) 11.41(4) 11.1(2) $\Omega_2$ 23.3(1) 13.01(5) 8.7(1) $r_1$ 0.031(1) 0.0993(4) 0.094(1) $r_2$ 0.023(1) 0.0834(4) 0.061(1) $\varphi_{per}$ 0.915 0.886 0.921 $l_{2}/l_{1}$ 0.471 0.685 0.392 $T_1$ \[K\] 5386(35) 5805(16) 6319(14) $T_2$ \[K\] 4995(23) 5758(20) 6220(20) : Fitted and calculated parameters of the light curve solutions: eccentricity $e$; periastron angle $\omega$; inclination *i*; mass ratio *q*; temperature of the secondary star $T_{2}$; surface potentials $\Omega_{1, 2}$; relative radius of the primary and secondary stars $r_{1, 2}$; periastron phase $\varphi_{per}$; ratio of the relative luminosities of the stellar components $l_2/l_1$; calculated temperatures of the primary and secondary stars $T_{1, 2}^f$[]{data-label="tab:3"} Although <span style="font-variant:small-caps;">PHOEBE</span> works with potentials, it gives a possibility to calculate directly all values (polar, point, side, and back) of the relative radius $r_i=R_i/a$ of each component ($R_i$ is linear radius and *a* is orbital separation). Moreover, <span style="font-variant:small-caps;">PHOEBE</span> yields as output parameters bolometric magnitudes $M_{bol}^i$ of the two components in conditional units (when radial velocity data are not available). But their difference $M_{bol}^2-M_{bol}^1$ determines the true luminosity ratio $s=L_2/L_1=l_2/l_1$. The method of differential corrections was used as a fitting algorithm. The parameters of the best light curve solutions are given in Table 2 while the corresponding synthetic curves are shown in Figs. 1–3 as continuous lines. The residual curves show bigger scatters during the eclipse phases (Figs. 1–3). Similar behavior could be seen also for other *Kepler* binaries (Hambleton et al. 2013a, Hambleton et al. 2013b, Lehmann et al. 2013, Maceroni et al. 2014). It was attributed to the effects of finite integration time (Kipping 2010). The reasons for this effect may be also numerical imperfectness of the physical model (Prsa et al. 2016) as well as contribution of pulsations and spots. The formal PHOEBE errors of the fitted parameters were unreasonably small. That is why we estimated the parameter errors manually based on the following rule (Dimitrov et al. 2017). The error of parameter $b$ corresponded to that deviation $\Delta b$ from its final value $b^{f}$ for which the mean residuals of the HB features increase by 3$\bar{\sigma}$ ($\bar{\sigma}$ is the mean photometric error of the target). The light curve solutions show that: (i) KIC 10992733 and KIC 5632781 undergo total eclipses and their mass ratios should be considered with a big confidence (Terrell $\&$ Wilson 2005); (ii) the targets have moderate eccentricity (0.27–0.38); (iii) the temperatures of target components are close (within 400 K, Table 2); (iv) mass ratios are in the range 0.42–0.97. Out-of-eclipse variability of the targets ========================================= The out-of-eclipse light of KIC 10992733 contains semi-regular oscillations (Fig. 4) with periods 5–7 d which undergo long-term modulation with amplitudes up to 0.005 mag. Their shape as well as lack of strong periodicity seem to exclude pulsation explanation. This is confirmed by the periodogram analysis of the out-of-eclipse data that reveals bad-defined peak (Fig. 5). Its period $P_{{out}}$ = 5.137 days is not harmonic of the orbital period (18.525 d). The ratio $P_{orb}/P_{out} \sim$ 3.6 of KIC 10992733 differs almost twice from the ratio of the pseudo-synchronous angular velocity and mean motion $\Omega_{ps}/n \sim$ 1.96 for $e$ = 0.38 (Hut 1981). Probably, $P_{{out}}$ = 5.137 days presents rotational period of some of the target components. Then the observed out-of-eclipse variability may be explained by variable spot visibility, the bad-defined peak of the Fourier transform may be result of differential rotation while the amplitude modulation may be attributed to spot activity cycle. The components of KIC 10992733 are K stars for which it is inherent characteristic. The asynchronism of the binary is expected considering its high eccentricity. Cool spot with angular size of order of 10$^{\circ}$ may reproduce the observed amplitude of out-of-eclipse oscillations of KIC 10992733. But there is a problem with the shape of oscillations: if the system is coplanar one spot of highly-inclined configuration would cause light variation whose shape is flat during almost half a cycle, different from the observed one. This problem may be overcame if one supposes that KIC 10992733 is not coplanar binary (expected for asynchronous system), i.e. far away from equilibrium state. Another alternative is KIC 10992733 to be coplanar binary with two diametrically opposed spots on some of its components but with a period of 2$P_{{out}}$ = 10.274 d. Then the ratio $P_{orb}/2P_{out} \sim$ 1.8 almost corresponds to pseudo-synchronous state of KIC 10992733. Precise spectral observations in the future may solve this ambiguity. ![The small-amplitude oscillations of KIC 10992733 probably are due to photospheric spot.[]{data-label="Fig4"}](fig4.eps){width="11cm"} ![Periodogram of out-of-eclipse variations of KIC 10992733[]{data-label="Fig5"}](periodogram.eps){width="7cm"} The out-of-eclipse light curves of KIC 5632781 and KIC 10026136 reveal light features around the periastron phases (Figs. 2–3, top panels) with amplitudes of 0.0035 mag and 0.0025 mag, i.e. these targets are eclipsing heartbeat stars (EB+HB type, Dimitrov et al. 2017). None of them reveals tidally-induced pulsations (Figs. 2–4). The synthetic <span style="font-variant:small-caps;">PHOEBE</span> features at the periastron reproduce well (but not perfectly) the observed ones. The possibility to reproduce such mmag details is a compliment of the physical model of binary stars, particularly of <span style="font-variant:small-caps;">PHOEBE</span>. The mean temperatures of KIC 5632781 and KIC 10026136 (Table 1) correspond to the maximum of temperature distribution of the MS heartbeat stars (figure 2 in Hambleton et al. 2013b). According to the model of Kumar et al. (1995) the amplitude of the periastron feature depends on the separation of the objects and their masses while its shape depends on the orbit parameters (Hambleton et al. 2013b): the argument of periastron $\omega$ determines the symmetry, the eccentricity *e* determines its relative width and the inclination *i* dictates the peak to trough ratio of the tidal pulse (see fig. 5 of Thompson et al. 2012). The contributions of the different parameters on the shape, width and amplitude of the periastron feature are apparent in figs. 3–4 of Dimitrov et al. (2017). It should be pointed out that both model sets (Thompson et al. 2012 and Dimitrov et al. 2017) describe result of clean tidal distortions without contribution of reflection and other faint effects. The last ones may be considerable for some cases (as KOI-54). The tidally-induced periastron features of KIC 5632781 and KIC 10026136 confirm the theoretical predictions. \(a) Their widths (around 0.40–0.45 in phase units) correspond to eccentricities of around 0.3. \(b) The tidally-induced periastron features have a ”hook” shape consisting of two parts with almost the same widths: light increasing followed by light trough (that coincides with the primary eclipse). This shape is expected for systems with $0 \leq \omega \leq 90^0$ (see fig. 3 of Dimitrov et al. 2017). The arguments of periastron $\omega$ of our two HB targets (Table 2) fulfil this condition. \(c) The ratios between the amplitudes of the two sections of the periastron features, the hump-shape part and the trough-shape part, of our targets are almost 1:1 (Figs. 2–3). This corresponds to binaries with $\omega \sim 30^0$ (fig. 3 of Dimitrov et al. 2017). The noneclipsing stars KIC 3547874, KIC 7622059 and KIC 11494130 from the sample of Thompson et al. (2012) exhibit tidally-induced hook-shape periastron features similar to ours. But their amplitudes of 1690, 1020 and 600 ppm are considerably smaller than those of our two targets. This result may be explained by almost 2 times shorter periods of our KIC 5632781 and KIC 10026136 than those of KIC 3547874, KIC 7622059 and KIC 11494130, because the shorter period means smaller separation, the stronger tidal interaction and correspondingly the bigger amplitude of the tidally-induced periastron feature. Conclusions =========== This paper presents the results of determination of the orbits and parameters of stellar configurations of the eclipsing eccentric binaries KIC 10992733, KIC 5632781 and KIC 10026136 on the basis of their *Kepler* data. The out-of-eclipse light variations of KIC 10992733 probably due to spot(s). KIC 5632781 and KIC 10026136 are eclipsing HB stars. The characteristics of their tidally-induced features are consistent with the theoretical predictions. The presented study adds new two members to the family of HB binaries and provides new data to search for dependencies of the tidally-induced effects of eccentric binaries on their orbital and stellar parameters. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported partly by project DN08/20 of the Fund for Scientific Research of the Bulgarian Ministry of Education and Science and by project RD 08-102/03.02.17 of Shumen University. It used the SIMBAD database and NASA Astrophysics Data System Abstract Service. We used data from *Kepler* EB catalog (http://keplerebs.villanova.edu/). The authors are very grateful to the anonymous referee for the valuable recommendations and notes. [00]{} Borucki W. et al., 2010, Sci, 327, 977 Burkart, J., Quataert, E., Arras, P., Weinberg, N.N., 2012, MNRAS, 421, 983 Claret, A., Gimenez, A., 1993, A & A, 277, 487 Dimitrov D., Kjurkchieva D., Iliev I., 2017, MNRAS, 469, 2089 Fuller, J., Lai, D., 2011, MNRAS, 420, 3126 Hambleton, K. et al., 2013a, MNRAS, 434, 925 Hambleton, K., Degroote, P., Conroy, K., et al. 2013b, EAS Publications Series, 64, 285 1, 4.6 Hut, P., 1981, A & A, 99, 126 Ivanov V.P., Kjurkchieva D.P., Srinivasa Rao M., 2010, BASI, 38, 83 Kipping, M., 2010, MNRAS, 408, 1758 Kirk, B. et al., 2016, AJ, 151, 68 Kjurkchieva, D., Vasileva, D., 2015, PASA, 32, 23 Kjurkchieva, D., Vasileva, D., 2016, NewA, 48, 30 Kjurkchieva, D., Vasileva, D., Dimitrov, D., 2016a, ApSS, 361, 132 Kjurkchieva, D., Vasileva, D., Dimitrov, D., 2016b, AJ, 152. 189 Koch, D.G. et al., 2010, ApJ, 713, L79 Kopal, Z., 1978, ASSL, 68, Dordrecht, D. Reidel Publishing Co. Kumar, P., Ao, C., Quataert, E., 1995, ApJ 449, 294 Lajoie, C., Sills, A., 2011, ApJ, 726, 67 Lehmann, H. et al., 2013, A & A, 557A, 79 Maceroni, C. et al., 2014, A & A, 563A, 59 Nicholls, C.P., Wood, P.R., 2012, MNRAS, 421, 2616 Prsa, A., Zwitter, T., 2005, ApJ, 628, 426 Prsa, A. et al., 2011, AJ, 141, 83 Regos, E., Bailey, V., Mardling, R., 2005, MNRAS, 358, 544 Sepinsky, J.F. et al., 2007, ApJ, 660, 1624 Sepinsky, J.F. et al., 2009, ApJ, 702, 1387 Slawson, R. et al., 2011, AJ, 142, 160 Smullen, R., Kobulnicky, H., 2015, ApJ, 808, 166S Shporer A. et al., 2016, ApJ, 829, 34 Terrell D., Wilson R., 2005, ApSpSci, 296, 221 Thompson, S.E. et al., 2012, ApJ, 753, 86 Welsh, W. et al., 2011, ApJS, 197, 4 Willems, B., Claret, A., 2005, SPC, 333, 52

ArXiv

harvmac =1600 Eva Silverstein Lyman Laboratory of Physics Harvard University Cambridge, MA 02138, USA .3in The simplest toroidally compactified string theories exhibit a duality between small and large radii: compactification on a circle, for example, is invariant under $R\rightarrow 1/(2R)$. Compactification on more general Lorentzian lattices (e.g. toroidal compactification in the presence of background metric, antisymmetric tensor, and gauge fields) yields theories for which a large-small equivalence is not so simple. Here an equivalence is demonstrated between large and small geometries for all toroidal compactifications. By repeatedly transforming the momentum mode corresponding to the smallest winding dimension to another mode on the lattice, it is possible to increase the volume to exceed a finite lower bound. It has been known for some time that the simplest toroidal string compactification exhibits an equivalence between small and large background geometries. That is, as pointed out in \[\] and later expanded upon in \[\] and \[4\], the presence of winding states leads to an $R\rightarrow 1/(2R)$ invariance of a theory compactified on a circle without any background fields; this generalizes easily to a G to $1/4G^{-1}$ invariance for a theory compactified on a torus given by the metric G. ($G_{ij}=e_i\cdot e_j$ where the $e_i$ form a basis for the winding vectors.) This raises the question of how generally this phenomenon occurs in arbitrary toroidal compactifications of string theory. To answer this question we must consider the moduli space of Lorentzian lattice compactifications spanned by the background metric, antisymmetric tensor, and gauge field (Wilson lines). The worldsheet action is $$S=\int{G_{ij}\partial_{\alpha}X^{i}\partial^{\alpha}X^{j} + \epsilon^{\alpha\beta}B_{ij} \partial_{\alpha}X^{i}\partial_{\beta}X^{j} + \epsilon^{\alpha\beta}A_i^J \partial_{\alpha}X^i\partial_{\beta}X^J}$$ As described in \[\], the left- and right-moving momenta are $$p_L=\bigl(P+A^Tn, G^{-1}\left(m/2-Bn-(1/4)AA^Tn-(1/2)AP\right)+n\bigr)$$ and $$p_R=\bigl(G^{-1}\left(m/2-Bn-(1/4)AA^Tn-(1/2)AP\right)-n\bigr)$$ where n,m $\epsilon Z^d$ are the winding and momentum vectors and P is a vector in the 16-dimensional root lattice of the gauge group. With general B and A values, interchanging $G$ and $1/4G^{-1}$ does not always yield a new state on the same lattice or even preserve the spectrum of $p_L^2+p_R^2$ values, as would be necessary for an invariance of this simple form. As discussed in \[\], the moduli space of Lorentzian lattice compactifications contains discrete equivalences given by O(16+d,d,Z) transformations on ($p_L$,$p_R$), which leave the lattice invariant but which can be obtained equivalently by changing the background fields. (Here d is the number of compactified spatial dimensions.) That is, the moduli space of compactifications is O(16+d,d)$\big/$O(16+d)xO(d)xO(16+d,d,Z). One example is the transformation taking $B_{ij}$ to $B_{ij}+1$; another is the well-known so-called duality transformation taking 2(B + G) to its inverse (discussed in \[4\] and \[\]). Thus we must investigate whether there exists some O(16+d,d,Z) transformation taking small G to large G. First we must specify precisely what we mean by the size of the compactified space. One condition for small-large duality requires the volume det G to have some finite effective lower bound. A stronger condition is that in addition the winding lengths $n^TGn$ should themselves all exceed some lower bound. As discussed in section 3, it can be shown that the latter implies the former: in any dimension, if all winding lengths exceed some minimum, the volume must also have a lower bound. In this paper we show that the compactified volume $\sqrt{\rm{det} G}$ can be raised to exceed a finite volume; we can make the further statement that if the lengths of the windings cannot all be made to exceed a finite lower bound, then we can transform the volume to $\infty$. In two dimensions with A=0 we will see explicitly that the winding lengths can be raised to exceed a lower bound. These results suggest that winding lengths can be raised in general; further work is necessary to show this for all cases. We will begin by discussing the d=2 case without Wilson lines in a way that we will then generalize to higher dimensions. Finally, we will demonstrate the result with Wilson lines included. As discussed in \[4\], the generalization of the interchange of momentum and winding vectors ($n\leftrightarrow m$) is $2(G+B)\rightarrow 1/2(G+B)^{-1}$. This transformation fails to transform all small metrics to large, as the following example demonstrates. When $G=RI$, we have $$2\left(\matrix{R^2&-B\cr B&R^2\cr}\right)\rightarrow {1\over{2(R^4+B^2)}}\left(\matrix{R^2&B\cr -B&R^2\cr}\right)$$ under this transformation (here $B=-B_{12}$). Thus as $R^2\rightarrow 0$ for $B\ne 0$ it is transformed to another nearly vanishing radius. It is possible, taking into account other available O(2,2,Z) transformations, to satisfy both conditions for small-large duality mentioned in section 1. As described in \[3\], we can raise the volume det G for $d=2$ using $SL(2,Z){\rm x} SL(2,Z) {\it=} SO(2,2,Z)$ transformations on the two complex parameters $$\rho = 2\left(B + i\sqrt{\rm det G}\right)$$ and $$\tau = {G_{12}\over{G_{11}}} + {i\sqrt{\rm det G}\over{G_{11}}}$$ We simply need to transform $\rho$ to its fundamental domain for which $2{\rm det}G \ge \sqrt{3}/2$ (see fig. 1). It is not too hard to see, however, that $SL(2,Z){\rm x} SL(2,Z)$ transformations do not suffice to raise the winding lengths of the two-dimensional torus: transforming $\tau$ by an SL(2,Z) transformation merely reparameterizes the torus (i.e. changes its basis); transforming $\rho$ does not affect the shape of the torus, just its volume. Consequently, since the fundamental domain contains only one point in the orbit of each $\rho$, a torus with a small winding situated in $\rho$’s fundamental domain cannot be transformed to one with larger windings by SL(2,Z)xSL(2,Z). To illustrate this, consider a diagonal torus with $R_1=\epsilon << 1$ and $R_2=1/\epsilon$. For this torus, det G = 1, putting it in $\rho$’s fundamental domain. Here Im $\tau = R_1/R_2$; transforming $\rho$ leaves this ratio fixed. Since $\rho$ is already in its fundamental domain, the volume cannot be further increased. Thus $R_1$ will stay small as long as we confine ourselves to SL(2,Z)xSL(2,Z). If, however, we interchange $\rho$ and $\tau$ (this is equivalent to exchanging the first component the momentum and winding modes, as will be seen in section 3), we find $$G\rightarrow\left(\matrix{1/(4\epsilon^2)&B/(2\epsilon^2)\cr B/(2\epsilon^2)&(1+B^2)/\epsilon^2\cr} \right)$$ and $$n^TGn\rightarrow (n_1 + (2B)n_2)^2/(4\epsilon^2) + n^2_2/\epsilon^2 \ge 1/(4\epsilon^2)$$ Thus we find that to raise winding lengths in $d=2$ we must make use of the rest of O(2,2,Z), namely the $Z_2$ interchange of $\rho$ and $\tau$. First we show that this transformation provides a different way to raise the volume from the SL(2,Z)xSL(2,Z) transformations. Given a $d=2$ torus G, reparameterize G if necessary to form a basis out of its two smallest winding lengths. The SL(2,Z) generators $\tau\rightarrow{\tau + 1}$, which takes $e_1 \rightarrow {e_1+e_2}$, and $\tau\rightarrow{-1/{\tau}}$, which interchanges the 1 and 2 components, provide the necessary transformations for this reparameterization. Suppose the smallest squared winding, $G_{11}$, is less than 1/2. Then the interchange of $\rho$ and $\tau$, taking det G to det G/$4G_{11}^2$, increases the volume. We may repeat this procedure as long as $G_{11}<1/2$. Then we have either ${\rm det} G\rightarrow\infty$ or at least $G_{11},G_{22}\ge 1/2$ (i.e. all squared winding lengths $\ge 1/2$, which is sufficient to put a lower bound on the volume, as discussed in section 3). This procedure will generalize to higher d. In fact, this procedure in $d=2$ gives us a two-step method for putting $\rho$ in its fundamental domain. In the basis consisting of the two smallest windings $e_1\le{e_2}$, the angle between $e_1$ and $e_2$ ($\theta_{12}$) must be between $\pi /3$ and $2\pi /3$; otherwise $\vert e_2-e_1 \vert < \vert e_2 \vert$, contradicting the fact that $e_1$ and $e_2$ are the two smallest windings. Thus once $\tau$ has been transformed to this basis, it falls in the shaded region in figure 2. Then $\tau$ can be placed in its fundamental region by $\tau \rightarrow \tau \pm 1$. Finally, $\rho \leftrightarrow \tau$ puts $\rho$ in its fundamental region, ensuring that $2\sqrt{{\rm det}G} > {\sqrt 3 /2}$. Next we show that a $d=2$ torus can be transformed via $\rho \leftrightarrow \tau$ into one with all windings bounded below. To see this we consider all possible windings $n^TG^{\prime}n$ of the new metric. We find $$G^\prime=\left(\matrix{1/(4G_{11})&B/(2G_{11})\cr B/(2G_{11})&({\rm det}G+B^2)/G_{11} \cr}\right)$$ This gives $$n^TG^{\prime}n={n_2^2{\rm detG}\over{G_{11}}} +{(n_1+2n_2B)^2\over{4G_{11}}}$$ We see here that with ${\rm det} G > 3/16$ (i.e. with $\rho$ in its fundamental domain so that $2\sqrt{{\rm det} G}>\sqrt{3}/2$), interchanging $\tau$ and $\rho$ yields new windings whose squared lengths all exceed $3/8$: If $n_2 \ne 0$, the squared winding $e^2 \ge 3/(16G_{11})\ge 3/8$ (for the smallest squared winding $G_{11}\le 1/2$). If $n_2 = 0$, we have $e^2 > 1/(4G_{11}) \ge 1/2$. Thus it takes at most two interchanges of $\rho$ and $\tau$ to transform any torus to one with all windings bounded below. Unfortunately this argument for putting a lower bound on the winding lengths does not generalize to all higher d, although the above argument for the weaker form of small-large duality does survive in higher dimensions. As discussed in \[3\], $$(1/2)(p_L^2+p_R^2)= \left(\matrix{n^T&m^T\cr}\right) \left(\matrix{G-BG^{-1}B&1/2BG^{-1}\cr -1/2G^{-1}B&1/4G^{-1}\cr}\right) \left(\matrix{n\cr m\cr}\right)$$ Thus to obtain the effect of an O(d,d,Z) transformation on the background fields, we simply transform the above matrix accordingly. That is, $$\left(\matrix{n \cr m \cr}\right) \rightarrow S \left(\matrix{n \cr m \cr}\right)$$ where $S\epsilon O(d,d,Z)$ is equivalent to $$M(G,B) \rightarrow S^TM(G,B)S$$ where M(G,B) is the matrix of background fields in the above expression for $p_L^2+p_R^2$. In particular, the lower right $d{\rm x}d$ block gives the transformed dual torus, which determines the new momentum modes $G^{-1}m$. The momentum modes characterize the volume much more directly than do the winding modes, since, as can be seen from the formulas for $p_L$ and $p_R$, turning on a winding mode generates nontrivial contributions to the momenta and thus to the energy which depend on B as well as G. This makes it difficult to extract G from the spectrum of winding modes. Under the O(d,d,Z) transformation which exchanges $n_1$ and $m_1$ taking the first momentum mode to the first winding mode (leaving all other momentum modes fixed), $G^{-1}$ becomes (generalizing $\rho\leftrightarrow\tau$ from 2 dimensions) $$1/4G^{\prime -1}=\pmatrix{ {G_{11}-{\left(BG^{-1}B\right)}_{11}} &1/2{\left(BG^{-1}\right)}_{12} &\cdots&1/2{\left(BG^{-1}\right)}_{1d}\cr 1/2{\left(BG^{-1}\right)}_{12}&1/4G^{-1}_{22} &\cdots&1/4G^{-1}_{2d}\cr \vdots&\vdots&\ddots&\vdots\cr 1/2\left(BG^{-1}\right)_{1d}&1/4G^{-1}_{2d} &\cdots&1/4G^{-1}_{dd}\cr}$$ Note that the squared volume det $G^{\prime -1}$ is independent of B: without the $G_{11}$ term, the first column is a linear combination of the other columns so that the contributions of $B_{ij}$ to the determinant cancel. That is, $$(BG^{-1}B)_{11}=B_{1k}G^{-1}_{kl}B_{l1}=(1/2B_{1k}G^{-1}_{kl})(2B_{l1})$$ and $$1/2(BG^{-1})_{1j}=1/2B_{1l}G^{-1}_{lj} =(1/4G^{-1}_{jl})(2B_{l1})$$ so that $${\rm column}1 = \left({\rm column} l\right)\left(2B_{l1}\right)$$ We are left with $${\rm det} G^{\prime-1}=4G_{11}V^*_{2d}$$ where $V^*_{2d}$ is the squared dual volume restriced to the 2,3,...,d directions (i.e. the determinant of $G^{-1}$ restricted to these components). Now $G_{11}=e_{1}^2$ and $e_{1}^2=1/cos^2{\theta_1}e^{*2}_1$ where $\theta_1$is the angle between $e_1$ and its dual $e^*_1$. So $${\rm det}G^{\prime-1}={\rm det}G^{-1}\left(4e_1^4 \right) \left({{V^*_{2d}e^{*2}_1{\rm cos}^2\theta_1} \over{V^*_{1d}}}\right)$$ But the last factor is 1, since the volume of the dual lattice is just the volume of the sublattice spanned by $e^*_2,...,e^*_d$ times the component of the remaining dual basis vector $e^*_1$ orthogonal to this sublattice. Thus exchanging $n_1$ and $m_1$ decreases ${\rm det}G^{-1}$ if $e_1^2 < 1/4$. By repeating this procedure, reparameterizing if necessary to render $e_1$ the smallest winding, we can continue increasing ${\rm det} G=1/{\rm detG}^{-1}$ as long as the smallest squared winding is less than 1/2. Then either det G increases ad infinitim or eventually all windings exceed 1/$\sqrt2$. In the latter case, it follows from a theorem in the geometry of numbers (a generalization of Blichfeldt’s theorem \[\]) that $$\sqrt{{\rm det}G} > V_d$$ where $$V_d={{\pi^{d\over 2}{\left(1/2\sqrt 2\right)^d}} \over{\Gamma(1+d/2)}}$$ is the volume of a sphere of diameter $1/\sqrt2$ in $R^d$. According to this theorem, any region in $R^d$ of volume V greater than $\sqrt{{\rm det} G}$ must contain two points whose difference is in the lattice determined by G. Consider a (d-1)-sphere whose diameter is 1/$\sqrt2$. After the above procedure we must have $\sqrt{{\rm det}G}$ greater than the volume of the sphere since otherwise the sphere would contain a winding shorter than 1/$\sqrt2$. Thus for any d, every theory is equivalent to one compactified on a volume exceeding $V_d$. With $A\ne 0$, the zero-mode mass spectrum is given by $$(1/2)(p_L^2+p_R^2)= \left(\matrix{n^T&m^T&P^T\cr}\right) M(G,B,A)\left(\matrix{n\cr m\cr P\cr}\right)$$ where M can be read off the expressions for $p_L$ and $p_R$. Since, as discussed above, the momentum modes characterize $G^{-1}$, we would like to transform the momentum modes orthogonally (i.e. via O(16+d,d,Z)) to other modes on the lattice with small values of $p_L^2+p_R^2$. If we transform only one of the momentum modes, that corresponding to the smallest winding $e_1$, leaving all other momentum modes fixed, the new dual volume will be the old dual subvolume for the 2 . . . d directions multiplied by the component of the transformed first momentum mode orthogonal to this subvolume. If this new orthogonal component is smaller than the original orthogonal component, the transformation will reduce the dual volume, as occurred for the $n_1\leftrightarrow m_1$ transformation with $A=0$. Knowing the transformed mode $(n,m,P)$ (with $2n\cdot m+P^2=0$ so that the new mode is still 0-norm) suffices to compute the new inverse metric $G^{\prime -1}$, as long as such an O(16+d,d,Z) transformation exists. We choose $$n=\left(\matrix{n_1\cr 0\cr \vdots\cr 0\cr}\right)$$ so that our new mode will still be orthogonal to the other momentum modes. As pointed out in \[\], it is always possible to find a dual $x^*$ to x=(n,m,P) such that the Lorentzian inner products $x^{*.} x=1$ and $x^{*2}=0$ hold and that $x^*$ is still orthogonal to the winding modes corresponding to the other momentum modes (2...d). (Here we are taking x to be indivisible since we are interested in small values of $p_L^2+p_R^2$; we can always divide out any common factor if necessary.) The 2...d momentum modes still have the corresponding winding modes as their duals. The space orthogonal to the new momentum and winding modes is a 16-dimensional even self-dual positive definite lattice, either $\Gamma_8 + \Gamma_8$, the root lattice for $E_8\times E_8$, or $\Gamma_{16}$, the root lattice for Spin(32)/$Z_2$. If upon transforming the first momentum mode, the orthogonal space switches from one root lattice to the other, our transformation is not quite Lorentzian and therefore cannot be absorbed into a background field transformation. It is, however, an O(16+d,d) transformation from a theory with the same G,B, and A and with our original momentum modes but with the orthogonal space the switched root lattice. This Lorentz transformation $\it{can}$ be absorbed into the background fields, giving $G^{\prime -1}$ as described above, since this is determined by the momentum-mode part of the transformation. This is simply a restatement of the isomorphism between $\Gamma_8+\Gamma_8$ with one compactified dimension and $\Gamma_{16}$ with one compactified dimension noted in \[3\] and \[7\]. Thus we can always transform the first momentum mode to any other 0-norm vector on the lattice in a way that can be obtained by an O(16+d,d) transformation on the background fields. We find $$G^{\prime -1}=\left(\matrix{ (1/2)(P+A^Tn)^2+n_1^2G_{11}+(1/4)z_iG^{-1}_{ij}z_j& (1/4)G^{-1}_{2k}z_k&\cdots&(1/4)G^{-1}_{dk}z_k\cr (1/4)G^{-1}_{2k}z_k&(1/4)G^{-1}_{22}&\cdots&(1/4)G^{-1}_{2d}\cr \vdots&\vdots&\ddots&\ddots\cr (1/4)G^{-1}_{dk}z_k&(1/4)G^{-1}_{d2}&\cdots&(1/4)G^{-1}_{dd}\cr} \right)$$ where $$z=m-(2B+(1/2)AA^T)n-AP$$ The (1,1) component of the new dual metric is of course the $(1/2)(p_L^2+p_R^2)$ value for our new momentum mode; the 2 . . . d components are those from the original dual metric, since we left the corresponding momentum modes alone. The others were obtained by transforming $M(G,B,A)\rightarrow O^TMO$, where O transforms the first momentum mode to the mode (n,m,P), leaving the others fixed. As in the A=0 case, we find that the first column contains a linear combination of the others (multiply column j by $z_j$). When we subtract this from the first column (since we are taking the determinant), the first row of the resulting matrix contains the same linear combination of the other rows as we had above for columns. Subtracting this out, the resulting dual volume is $$\rm{det}(1/4)G^{\prime -1}= [(1/2)(P_I+A^1_In_1)^2+n_1^2G_{11}](1/4)V^*_{2...d}(1/4)^{d-1} +z_1^2\rm{det}(1/4)G^{-1}$$ So $${\rm{det}(1/4)G^{\prime -1}\over\rm{det}(1/4)G^{-1}} = e_1^2[(1/2)(P+A_I^1n_1)^2+n_1^2e_1^2] +(m_1-(1/2)A_1^JA_J^1n_1-A^J_1P_J)^2$$ We need to find $n_1$, $m_1$, and P to render this ratio smaller than 1 as long as $e_1$ remains smaller than some lower bound. We can enforce the 0-(Lorentzian) norm condition by taking $n_1=2y^2$, $P_I=2S_Iy$ and $m_1=-S^2$ where y is an integer and S is a vector in the root lattice. This ensures that $2n_1m_1+P^2=0$, with $m_1$ integral. With this choice we have $$(P+A^1n_1)^2=(2Sy+A^1(2y^2))^2=4y^2(S+A^1y)^2$$ $$=4y^2(D_I+A^{\prime 1}_Iy)\Gamma^{IJ}(D_J+A^{\prime 1}_Jy)$$ and $$(m_1-(1/2)A_1^JA_J^1n_1-A_1P)^2=\bigl((S+A^1y)^2\bigr)^2$$ $$=\bigl((D_I+A^{\prime 1}_Iy)\Gamma^{IJ}(D_J+A^{\prime 1}_Jy) \bigr)^2$$ where the $D_I\epsilon Z$ give S in the integer basis: $S=D_IE_I$ where the $E_I$ are the basis vectors for the root lattice. Here $\Gamma$ is the metric for the root lattice: $E_I\cdot E_J=\Gamma_{IJ}$. Similarly $A^1=A^{\prime 1}_IE_I$. By a result of diophantine approximation \[\], it is possible to choose integers $D_I$ and y such that $$\vert D_I+A^{\prime 1}_Iy\vert < (\sqrt{e_1})^{1/16}$$ with $$y<(1/\sqrt{e_1})$$ Then $$e_1y\vert D_I+A^{\prime 1}_Iy\vert < (\sqrt{e_1})^{17/16}$$ and $$n_1^2e_1^4 < 4e_1^2$$ So $${\rm{det}G^{\prime -1}\over \rm{det} G^{-1}} < f(e_1)$$ where $f(e_1)$is a positive polynomial in $e_1^{1/16}$ since the root lattice metric $\Gamma$ is positive definite. Setting this polynomial equal to 1 gives a bound on $e_1$ below which the transformation lowers the dual volume, and thus raises the volume. Once again, if the smallest winding $e_1$ never exceeds this finite lower bound, the volume diverges by repeated transformations of the corresponding momentum mode. We have seen that by repeatedly transforming the momentum mode on the smallest circle, we can prevent the volume of the compactified dimensions from getting too small. In two compactified dimensions without Wilson lines, this transformation is sufficient in fact to raise all winding lengths. For any d, if it turns out that the winding lengths cannot be raised by O(16+d,d,Z), then we can at least say that the corresponding volume becomes unbounded. If this is the whole story, the fact that $\sqrt{{\rm det} G} > V_d $ will imply that some sort of uncertainty principle is operating on the winding lengths, requiring at least one of them to blow up if any approach zero. Since the theory has infinitely many chances to escape this volume divergence as we iteratively perform our transformation on the first momentum mode, and since in two dimensions without Wilson lines windings are bounded below, we expect windings to have a lower bound in all cases. This equivalence between large and small spaces is often taken to account for the necessary imprecision in measurements made by fundamental strings of nonzero (Planckian) size, since it keeps strings insensitive to arbitrarily small distances. This does not completely resolve the problem, however, since duality does not prevent one from measuring [*exactly*]{} the volumes of the large and small spaces. With a Planck-sized ruler one could certainly distinguish a tiny space from a huge one; what one could not do is measure precisely the size of either. The fact that they cannot be distinguished results directly from the presence of the winding modes. Further work is necessary to determine whether windings can be raised in all cases and the relation of small-large duality to the modular invariance of the worldsheet. I would like to thank C. Vafa for many helpful discussions and for introducing me to the subject of small-large duality. I would like to thank B. Gross for a helpful discussion on the automorphisms of the lattice. I am grateful to the Rowland Fund and NSF grant PHY-87-14654 for support.

ArXiv

--- abstract: 'In the framework of the most-studied doubly special relativity models the use of the naive formula $v=dE/dp$ has been argued to lead to inconsistencies connected to different rules of transformation, under boosts, of particles with the same energy but with different masses. In this paper we show that, at least in 1+1 dimensions, doubly special relativity can be formulated in such a way that the formula $v=dE/dp$ is fully consistent with the invariance of the relative rest, easily fitting to the relativity principle. It is also argued that, always in 1+1d, is not necessary to renounce to the usual (commutative) Minkowski space-time endowed with energy-independent boost transformations. The compatibility of the approach with superluminal propagation, with linear addition rule for energy, and possible extensions to 3+1 dimensions are also discussed.' author: - Gianluca Mandanici title: On the invariance of the relative rest in doubly special relativity --- Introduction {#S1} ============ In the quest for a quantum theory of gravity notable interest has been attracted over the last decade and more, from scenarios in which the Lorentz/Poincaré symmetry is deformed (but not broken) at the Planck scale. A much studied class of these scenarios is often referred in literature as doubly special relativity$\,$[@AmelinoCamelia:2000ge; @AmelinoCamelia:2000mn; @Magueijo:2001cr; @KowalskiGlikman:2002we; @AmelinoCamelia:2003ex] and typically consists in a two-scale extension of Einstein’s special relativity, being the second invariant scale eventually connected with the Planck energy/length. At first, doubly special relativity has been formulated in the energy-momentum sector, however, over the years, many attempts have been made to consistently extend the theory from the energy-momentum sector to the space-time sector (for a recent review see [@AmelinoCamelia:2010pd]). A number of these attempts are based on the assumption of a underling non-commutative space-time as a consequence of the fuzziness of space-time at length scale of the order of the Planck length (see e.g.$\,$[@Lukierski199590; @KowalskiGlikman:2002jr; @Mignemi:2003dm; @Galan:2007ns; @Ghosh:2007ai]). Other attempts instead assume a underlying commutative space-time (see e.g.$\,$[@Magueijo:2002am; @Kimberly:2003hp]). What is common within most of these approaches is that one obtains energy-dependent space-time transformations with the consequence that particles with the same speed, but with different masses, acquire relative motion when considered from different inertial observers. This unclassical feature of space-time transformations could appear in contrast with the relativity principle especially for the apparent loss of locality$\,$[@Hossenfelder:2010tm], that is recovered only in a form that depends on the choice of the inertial frame$\,$[@AmelinoCamelia:2011bm; @AmelinoCamelia:2010qv; @AmelinoCamelia:2011gy]. A central role in the delicate transition from the energy-momentum sector to the space-time sector of doubly special relativity is played by the formula that expresses the speed of a free particle as a function of its energy and its momentum. While the first proposals$\,$[@AmelinoCamelia:2000ge; @AmelinoCamelia:2000mn; @Magueijo:2001cr] have assumed the validity of the usual formula $v=dE/dp,$ in$\,$[@Kosinski:2002gu] this formula been argued to be incompatible with the standard hamiltonian formalism. In$\,$[@Mignemi:2003ab] starting from the request that the formula of the particle speed should not depend on the mass of the particles, the author was led to rule out the formula $v=dE/dp.$ Similar results have been obtained in$\,$[@Aloisio:2004rd] where has also been claimed that the dependence of the velocity on the mass of the boosted particle appear unavoidable using the mentioned formula. A study of velocity of particle in non-commutative space-time still compatible with the formula $v=dE/dp$ is instead that of$\,$[@AmelinoCamelia:2002tc]. The main aim of this paper is just to address the issue of the relative rest in doubly special relativity and its compatibility with the standard formula $v=dE/dp.$ The structure of the paper is the following. In section$\,$\[S2\] we discuss the implications of the request of covariance of the velocity in the form $dE/dp$ for the energy-momentum sector of the most-studied doubly special relativity models, providing explicit examples of covariance. In section$\,$\[S3\] we propose a construction of the space-time sector from given energy-momentum sectors. In section$\,$\[S9\] we address the issues of the compatibility of the approach with deformations of the boost action on the particle speed, with linear addition rule for energy and we also discuss a possible extension to 3+1 space-time dimensions. Finally, in section$\,$\[S10\], we present our conclusions. Momentum space formulation {#S2} ========================== The classical relativistic picture of relative rest --------------------------------------------------- The first working assumption of this paper is that *particles that do not possess relative motion in a given reference frame do not acquire relative motion in any other (inertial) reference frame*. In the usual formulation, for infinitesimal transformations, the speed of a boosted particle $v^{\prime}$ is obtained from the speed of the particle $v$ by means of the formula $$v^{\prime}(\xi)\simeq v+\xi\left\{ N,v\right\} ,\label{speedtransformationrule}$$ where $\xi$ is the boosting parameter and $\left\{ N,v\right\} $ is the infinitesimal boost action. In this notation, in order to obtain the invariance of the relative rest*,* the boost operator must be function only of the particle speed $v$ and of (possible) fundamental constants $c_{1},...,c_{n}$ of the model $$\left\{ N,v\right\} =K(v,c_{1},...,c_{n}).\label{boostv}$$ Notice that this request is satisfied in Galilei relativity where no invariant velocity scale is present. In fact in Galilei relativity, being $E=p^{2}/(2m)$, $v=dE/dp=p/m$ and $\left\{ N,p\right\} =m$, one finds $$\left\{ N,v\right\} =\left\{ N,p\right\} \frac{dv}{dp}=1,$$ so that $K(v)=1.$ Also in Einstein’s special relativity, where the invariant velocity scale $c$ is present, being $E^{2}=c^{2}p^{2}+c^{4}m^{2}$, $v=dE/dp=c^{2}p/E$ and $\left\{ N,p\right\} =E/c^{2},$ one finds $$\left\{ N,v\right\} =\left\{ N,p\right\} \frac{dv}{dp}=1-c^{2}p^{2}/E^{2},$$ i.e. $K(v,c)=1-v^{2}/c^{2}.$ Again the function $K(v)$ depends only on the particle speed and on the invariant scale $c$. In this latter case one also gets $K(c,c)=0$ consistently with the fact that $c$ is not only an invariant parameter of the theory but, rather, it is the invariant speed of (massless) particles. In both these known cases $K(v)$ depends only on the particle speed, so that two particles with the same speed in a given reference frame $v_{1}=v_{2}=v$ do not acquire relative motion in any other inertial reference frame $$v_{2}^{\prime}-v_{1}^{\prime}\simeq\xi\lbrack K(v,c)-K(v,c)]=0,$$ independently on their masses. Relative rest in the most studied doubly special relativity models ------------------------------------------------------------------ In most approaches to Planck-scale relativity there are strong departure from the classical picture as outlined above. In fact if one adopts the formula[^1] $$v=\frac{dE}{dp},\label{speed}$$ one finds that in general $K=K(v,c,\lambda cm,\lambda p)$, where $\lambda$ is the Planck-scale parameter, eventually identified with the inverse of the Planck momentum. This means that the action of the boosts on the speed of a particle depends not only on the particle speed itself but also on the particle mass. The main consequence of this dependency is that particles having the same speeds $\overline{v}$ in a given reference frame could acquire relative motion in a different reference frame, if they have different masses. In fact, being in general $K(\overline{v},c,\lambda cm_{1},\lambda p_{1})\neq K(\overline{v},c,\lambda cm_{2},\lambda p_{2}),$ one finds $$\overline{v}_{1}^{\prime}-\overline{v}_{2}^{\prime}=\xi\lbrack K(\overline{v},c,\lambda cm_{1},\lambda p_{1})-K(\overline{v},c,\lambda cm_{2},\lambda p_{2})]\neq0.$$ This effect has been analyzed in$\,$[@Aloisio:2004rd] in the framework of the so-called DSR1 model proposed in$\,$[@AmelinoCamelia:2000ge; @AmelinoCamelia:2000mn] that admits energy-momentum dispersion relation of the type $$E=c\left|\lambda\right|^{-1}\ln\left[\frac{\cosh\left(\lambda mc\right)+\sqrt{\cosh^{2}\left(\lambda mc\right)-(1-\lambda^{2}p^{2})}}{(1-\lambda^{2}p^{2})}\right],\label{DSR1}$$ but the same effect is also present in the other most-studied doubly special relativity model, the so-called DSR2 model, proposed in$\,$[@Magueijo:2001cr]. Again can be easily shown that from energy-momentum dispersion relation of the form $$\frac{E^{2}-p^{2}c^{2}}{(1-\lambda E/c)^{2}}=\frac{m^{2}c^{4}}{(1-\lambda mc)^{2}},\label{MSDDDR}$$ using (\[boostv\]) and (\[speed\]) one finds $$K(v,c,\lambda m)=\frac{1-v^{2}/c^{2}}{1+\lambda cm/\sqrt{(1-\lambda cm)^{2}-v^{2}(1-2\lambda cm)/c^{2}}}.\label{KDSR2}$$ From eq.(\[KDSR2\]) follows that two particles at rest in a given reference frame ($v_{1}=v_{2}=0$), will acquire a relative motion: $$v_{1}^{\prime}-v_{2}^{\prime}=\xi\lbrack K(0,c,\lambda m_{1})-K(0,c,\lambda m_{2})]\simeq\xi\lambda c(m_{1}-m_{2})\neq0.\label{relmotion}$$ The dependency of the boosts on the mass of the particle does not necessarily imply a violation of the relativity principle but, at least, suggests that locality should become a concept relative to the particular choice of the reference frame (see e.g.$\,$[@AmelinoCamelia:2011bm; @AmelinoCamelia:2010qv; @AmelinoCamelia:2011gy]). In the rest of this paper we wont address the point of whether or not effects of the type implied by formula (\[relmotion\]) can be incorporated in a relativistic scheme. Rather we intend to focus here on what could be the features of doubly special relativity models in which effects implied by formula (\[relmotion\]) are absent at all, like they are in the known classical cases. Approaching relative rest invariance in doubly special relativity ----------------------------------------------------------------- ### General theoretical framework Once we have discussed the way in which the problem manifests in doubly special relativity we can move forward looking for a solution. We start requiring the covariance under boosts of the energy-momentum dispersion relation $E=E(p).$ From this request it follows that it must be $$\left\{ N,E\right\} =\left\{ N,p\right\} \dfrac{dE}{dp},\label{CovRelDisp}$$ that fixes the relation between the action of the boosts on the energy and the action of the boosts on the momentum. Our next request is that $\left\{ N,v\right\} $ be a function of only the particle speed, expressed as in eq.(\[speed\]), from which one gets the following formula $$\left\{ N,v\right\} =\left\{ N,p\right\} \dfrac{d^{2}E}{dp^{2}}=K\left(\frac{dE}{dp}\right).\label{eq:RR}$$ The validity of eq.(\[eq:RR\]) directly implies that particles with the same speed (i.e. at rest in a given reference frame) wont acquire any relative motion in any other reference frame. Thus the request that the boosted velocity depends only on the particle speed fixes the infinitesimal action of the boost on the momentum in the form: $$\left\{ N,p\right\} =K\left(\frac{dE}{dp}\right)/\dfrac{d^{2}E}{dp^{2}}.\label{CovSpeedExpress}$$ Finally substituting (\[CovSpeedExpress\]) in (\[CovRelDisp\]) one obtains the infinitesimal action of the boost on the particle energy $$\left\{ N,E\right\} =K\left(\frac{dE}{dp}\right)\dfrac{dE}{dp}/\dfrac{d^{2}E}{dp^{2}}.\label{BoostEnergyAction}$$ Formulas (\[CovSpeedExpress\])-(\[BoostEnergyAction\]) together with (\[speed\]) are the main working tools of the analysis presented in the rest of the paper. ### A first-order example in the Planck scale We are now ready to apply the procedure outlined above to a general Planck-scale deformed dispersion relation. As a warm up we start the analysis with a first order deformation. The dispersion relation that we consider here is of the type of those studied in [@AmelinoCamelia:1997gz; @Gambini:1998it], that is a most studied energy-momentum dispersion relation: $$E^{2}-c^{2}p^{2}+\lambda p^{3}c^{2}=m^{2}c^{4}.\label{fodr}$$ Particle speed can be calculated from the dispersion relation using eq.(\[speed\]) obtaining $$v=\dfrac{dE}{dp}=\frac{c^{2}p(1-3/2\lambda p)}{\sqrt{c^{4}m^{2}+p^{2}c^{2}-\lambda c^{2}p^{3}}},$$ that, for massless particles, assumes the form $v\simeq c(1-\lambda p)$, at the first order in $\lambda.$ To derive the action of the boosts on the energy and on the momentum we use formulas (\[CovSpeedExpress\])-(\[BoostEnergyAction\]). The form of $K(v)$ is in principle arbitrary unless it provides the correct Minkowski limit. However, in this section, we will assume an undeformed expression for $K(v)$. Thus fixing $K(v)=1-v^{2}/c^{2}$, one finds for the action of the boost: $$\begin{aligned} \left\{ N,p\right\} & =\frac{\left[4c^{2}m^{2}+p^{3}\lambda\left(8-9p\lambda\right)\right]\sqrt{c^{2}m^{2}+p^{2}-\lambda p^{3}}}{cp^{3}\lambda(-4+3p\lambda)+4c^{3}m^{2}\left(1-3\lambda p\right)},\label{Np}\\ \left\{ N,E\right\} & =\left\{ N,p\right\} \frac{c^{2}p(1-3/2\lambda p)}{\sqrt{c^{4}m^{2}+p^{2}c^{2}-\lambda c^{2}p^{3}}},\label{NE}\end{aligned}$$ whose behavior, as a function of the particle momentum $p$, is reported in Fig.$\,$\[Fig1\]. ![(a) $v/c$ vs $\lambda p$ and (b) $\left\{ N,p\right\} $ vs $\lambda p$ for the dispersion relation of eq.(\[fodr\]) with $mc\lambda=0.001.$ The black-dashed lines ($\lambda=0$) are the predictions of special relativity. []{data-label="Fig1"}](Figura1 "fig:")![(a) $v/c$ vs $\lambda p$ and (b) $\left\{ N,p\right\} $ vs $\lambda p$ for the dispersion relation of eq.(\[fodr\]) with $mc\lambda=0.001.$ The black-dashed lines ($\lambda=0$) are the predictions of special relativity. []{data-label="Fig1"}](Figura2 "fig:") It is easy to check that $i)$ the dispersion relation (\[fodr\]) is invariant under boost action and that $ii)$ we have $v_{1}^{\prime}-v_{2}^{\prime}\simeq\xi\left[K(0,c,\lambda cm_{1})-K(0,c,\lambda cm_{2})\right]=0$ for all the possible values of the masses $m_{1}$ and $m_{2}$ of the particles involved. Thus the state of relative quite is preserved for all the particles in all the inertial frames. The entire analysis here reported is performed in the energy-momentum space. The only connection with space-time is provided by formula (\[speed\]) that defines the particle speed. In the case of $\lambda=0$ one recovers the usual special relativistic formulas. The case $\lambda<0$, that in principle should correspond to superluminal propagation (i.e. to $v>c$), in our approach returns $\left\{ N,p\right\} \simeq\left\{ N,E\right\} \simeq0$ as soon as $p\simeq p_{I}=\sqrt[3]{m^{2}c^{3}\lambda^{-1}/2}$, as it is clear both from the analytical expressions (\[Np\])-(\[NE\]) and from Fig.$\,$\[Fig1\]. This means that $p_{I}$ is an invariant momentum ($p_{I}^{\prime}\simeq p_{I}$), and the corresponding energy an invariant energy ($E_{I}^{\prime}\simeq E_{I}$). The reasons why one gets invariant energy and momentum from eqs.$\,$(\[Np\])-(\[NE\]) is that as soon as the momentum approaches $p_{I}$ the speed of the particle involved approaches the invariant speed $c,$ being in fact $\left\{ N,v(p_{I})\right\} \simeq0$. Notice that the invariant momentum $p_{I}$ depends on the mass of the particle involved and that it is different from the Planck momentum, as far as the mass of the particle involved is different from the Planck mass. The case $\lambda>0$ corresponds to subluminal propagation (i.e. to $v<c$). In this case the particle never approaches $v\simeq c$ and thus there is no invariant momentum. However there is a maximum momentum since, as soon as $p\simeq p_{max}=\sqrt[3]{m^{2}c^{3}\lambda^{-1}}$ , one gets $\left\{ N,p_{max}\right\} \simeq\left\{ N,E(p_{max})\right\} \simeq\infty$. ### Two all-order examples: the Magueijo-Smolin dispersion relation and the $\kappa$-Poincaré dispersion relation. We are now ready to apply our procedure to all-order (in the Planck scale) dispersion relations. The first type of dispersion relation that we consider is the Magueijo-Smolin dispersion relation (\[MSDDDR\]). Again using formula (\[speed\]) one gets the particle speed: $$v=\dfrac{dE}{dp}=\frac{c^{2}p(1-mc\lambda)^{2}}{\left(1-2\lambda cm\right)E+m^{2}c^{3}\lambda}.$$ From eqs.(\[CovSpeedExpress\])-(\[BoostEnergyAction\]) we obtain that the transformation rules between inertial observers are given by the actions $$\begin{aligned} \left\{ N,p\right\} & =\frac{E(1-2mc\lambda)+m^{2}c^{3}\lambda}{c^{2}(1-mc\lambda)^{2}}(1-\lambda^{2}p^{2}),\\ \left\{ N,E\right\} & =p(1-\lambda^{2}p^{2}),\end{aligned}$$ whose behavior is reported in Fig.$\,$\[Figura2\]. ![(a) $v/c$ vs $\lambda p$ and (b) $\left\{ N,p\right\} $ vs $\lambda p$ for the dispersion relation of eq.(\[MSDDDR\]) with $mc\lambda=0.001.$ The black-dashed line ($\lambda=0$) is the prediction of special relativity.[]{data-label="Figura2"}](Figura3 "fig:") ![(a) $v/c$ vs $\lambda p$ and (b) $\left\{ N,p\right\} $ vs $\lambda p$ for the dispersion relation of eq.(\[MSDDDR\]) with $mc\lambda=0.001.$ The black-dashed line ($\lambda=0$) is the prediction of special relativity.[]{data-label="Figura2"}](Figura4 "fig:") As expected as $v$ approaches the invariant speed $c$, the energy and the momentum approaches the invariant energy $c\lambda^{-1}$ and the invariant momentum $\lambda^{-1}$ respectively, as in the case of the original boost actions discussed in [@Magueijo:2001cr] but now, differently from [@Magueijo:2001cr], it is easy to check that also $$v_{1}=v_{2}\Rightarrow v_{1}^{\prime}=v_{2}^{\prime},$$ independently on the particle masses. Notice that it is also possible to express the energy and the momentum in terms of the mass and the speed of the particle $$\begin{aligned} p\left(v\right) & =\frac{mv}{\sqrt{\left(1-\lambda mc\right)^{2}-v^{2}(1-2mc\lambda)/c^{2}}},\label{eq:pv_ev1}\\ E(v) & =\frac{1}{1-2\lambda mc}\dfrac{m\left(1-\lambda cm\right)^{2}c^{2}}{\sqrt{\left(1-\lambda mc\right)^{2}-v^{2}(1-2mc\lambda)/c^{2}}}-\frac{m^{2}c^{3}\lambda}{1-2\lambda cm}.\label{eq:pv_ev2}\end{aligned}$$ The second example of all-order dispersion relation that we want to consider here is the $\kappa$-Poincaré motivated dispersion relation of eq.(\[DSR1\]). Following the same procedure we can easily find $v,$ $\left\{ N,p\right\} $ and $\left\{ N,E\right\} $ (again involving eqs.(\[speed\]), (\[CovSpeedExpress\]) and (\[BoostEnergyAction\])). Being the analytical expressions rather involved we do not report them. Instead we represent graphically the results in Fig.$\,$\[Figura3\] only for case $\lambda>0,$ since the case $\lambda<0$ does not possess, in our approach, the proper classical limit (see also the discussion in [@AmelinoCamelia:2010pd]). ![(a) $v/c$ vs $\lambda p$ and (b) $\left\{ N,p\right\} $ vs $\lambda p$ for the dispersion relation of eq.(\[DSR1\]) with $mc\lambda=0.001.$ The black-dashed lines ($\lambda=0$) are the predictions of special relativity. []{data-label="Figura3"}](Figura5 "fig:") ![(a) $v/c$ vs $\lambda p$ and (b) $\left\{ N,p\right\} $ vs $\lambda p$ for the dispersion relation of eq.(\[DSR1\]) with $mc\lambda=0.001.$ The black-dashed lines ($\lambda=0$) are the predictions of special relativity. []{data-label="Figura3"}](Figura6 "fig:") The space-time sector {#S3} ===================== In the preceding sections we have analyzed the invariance of the relative rest in the energy-momentum sector. In this section we try to address the same issue in the space-time sector. The link between the energy-momentum sector and the space-time sector is played by the definition of the particle speed given that, according to our assumptions, the following formula must hold $$v=\dfrac{dE}{dp}=\dfrac{dx}{dt}.\label{spacetimespeed}$$ The transformation rules in the energy-momentum sector have already been obtained therefore it is necessary that space-time transforms accordingly. We request that, as usual, different inertial systems are connected by uniform motion. Since we are looking for space-time boost transformations of the form $$x'(\xi)\simeq x+\xi\left\{ N,x\right\} ,$$ in order to have inertial systems moving at a uniform relative motion it must be $$\left\{ N,x\right\} =t.$$ Then the covariance in the energy-momentum sector, together with the request of covariance of (\[spacetimespeed\]), fixes the transformation rule $$\left\{ N,dt\right\} =\left\{ N,\dfrac{dx}{v}\right\} =\dfrac{1-K(v)}{v}dt.\label{timeboost}$$ Notice that the transformation of the time under boost does not depend on the mass of the particles involved, it depends only on their speeds. Notice also that the action of boost on time can be a non-linear function of the particle speed $v$, depending on the particular form of $K(v)$. Since we do not want to introduce new invariant velocity scales, we will assume that $K$ has not to be modified with respect to the standard special relativistic case.[^2] Adopting the usual form $K(v)=1-c^{-2}v^{2}$ and substituting in (\[timeboost\]) we get the action $$\left\{ N,t\right\} =c^{-2}x,$$ that is the usual infinitesimal action in the Minkowski space-time. It follows that all the elements regarding the space-time sector are not affected by the introduction of the Planck scale-parameter. In particular we get the usual dependence of the boost parameter $\xi$ on the relative velocity $v_{S}$ of the inertial systems, and the usual velocity composition rule $$v_{S}=c\tanh\left(\frac{\xi}{c}\right),\hspace{2cm}v^{\prime}=\frac{v+v_{S}}{1+\dfrac{vv_{S}}{c^{2}}}.$$ Notice however that being in general $E(p)\neq E_{SR}(p)$ it follows that $p=\left(dE/dp\right)^{-1}(v)\neq\left(dE_{SR}/dp\right)^{-1}(v)$ so that the momentum and the energy depend on the speed of the particle in a deformed way with respect to the special relativistic case, as already evident from eqs.(\[eq:pv\_ev1\])-(\[eq:pv\_ev2\]). Further issues {#S9} ============== In this section we consider other issues related to the construction developed. The issues that we want to address are *i)* the possibility of adopting different forms of the function $K(v)$, *ii)* the compatibility of the formalism with a linear addition law for energies, *iii)* the possibility to extend the formalism from 1+1 dimensions to 3+1 dimensions. On the possibility of different $K(v)$ -------------------------------------- The possibility of modifying the usual Minkowski space-time means, in the framework of this paper, considering deviations from the special relativistic formula $K(v)=1-v^{2}/c^{2}$. Since our basic assumption is that $K(v)$ be a function only of $v=dE/dp,$ we consider deformations of the type $$K(v)={\displaystyle \sum\limits _{k=0}^{n}}a_{2k}\left(\dfrac{v}{c}\right)^{2k},\label{Kmod}$$ where $k=0$, $a_{0}=1$ is the usual Galilean term and $k=1,$ $a_{2}=-1$ is the special-relativistic correction. Notice that according to eq.(\[Kmod\]) $c$ is not necessarily an invariant velocity anymore, being in general $K(c)\neq0$. However if we restrict our attention to function $K(v)$ such that $${\displaystyle \sum\limits _{k=0}^{n}}a_{2k}=0,$$ then $K(c)=0$, so that $c$ is still an invariant speed. If now we substitute (\[Kmod\]) in eqs.(\[CovSpeedExpress\])-(\[BoostEnergyAction\]) and consider dispersion relations of the type of eq.(\[fodr\]), being $$\begin{aligned} v & =\dfrac{dE}{dp}\simeq c\left(1-\frac{m^{2}c^{2}}{2p^{2}}-\lambda p\right),\\ \dfrac{d^{2}E}{dp^{2}} & \simeq c\left(\frac{m^{2}c^{2}}{p^{3}}-\lambda\right),\end{aligned}$$ we get in the large momentum limit ( $p\gg mc$ and $m\ll M_{Planck}=\lambda^{-1}c^{-1}$) $$\left\{ N,p\right\} \simeq\frac{\sum_{k=0}^{n}a_{2k}\left(1-\frac{m^{2}c^{2}}{2p^{2}}-\lambda p\right)^{2k}}{c\left(\frac{m^{2}c^{2}}{p^{3}}-\lambda\right)},\text{ \ \ \ \ \ \ \ \ \ \ }\left\{ N,E\right\} \simeq\frac{\sum_{k=0}^{n}a_{2k}\left(1-\frac{m^{2}c^{2}}{2p^{2}}-\lambda p\right)^{2k+1}}{c\left(\frac{m^{2}c^{2}}{p^{3}}-\lambda\right)}.\label{boostsvgec}$$ When $\lambda>0$, that here means subluminal propagation, it is not possible for $\left\{ N,p\right\} $ and $\left\{ N,E\right\} $ to vanish. Thus we cannot have in this case neither an invariant momentum/energy nor an invariant speed of particle, not even for massless particles. Moreover as soon as the momentum reaches $p\sim\sqrt[3]{m^{2}c^{3}\lambda^{-1}}$ the denominators of eqs.(\[boostsvgec\]) diverge. There is again a maximum allowed momentum $p_{\max}\sim\sqrt[3]{m^{2}c^{3}\lambda^{-1}}$ that depends on the particle mass and that can be many orders of magnitude smaller than the Planck momentum $\left\vert \lambda\right\vert ^{-1}$. Instead in the case $\lambda<0$ we get that the invariant speed $c$ is reached as soon as the momentum becomes of the order of $p_{I}\sim\sqrt[3]{m^{2}c^{3}\left|\lambda\right|^{-1}/2}$.[^3] For the same value of the momentum, the numerators of (\[boostsvgec\]) vanish so that the momentum $p_{I}$ is an invariant momentum and the corresponding energy an invariant energy. Thus the modifications introduced in $K(v)$ do not lead to significant differences in the predictions of the model. Always maintaining the same deformed form of $K(v)$ it is worth considering different dispersion relations. If we focus on the lowest orders, in the Planck parameter, of the dispersion relations of eq.(\[MSDDDR\]) we get $$\begin{aligned} v & =\dfrac{dE}{dp}\simeq c\left(1-\frac{m^{2}c^{2}}{2p^{2}}-\frac{m^{3}c^{4}}{p^{2}}\lambda+\frac{c^{4}m^{2}}{2}\lambda^{2}\right),\\ \dfrac{d^{2}E}{dp^{2}} & \simeq c\left(\frac{m^{2}c^{2}}{p^{3}}+\frac{2m^{3}c^{4}}{p^{3}}\lambda\right),\end{aligned}$$ that suggest that $v=c$ is still the invariant speed for massless particles, independently on their momenta, and that $p\sim\lambda^{-1}$ remains an invariant momentum (also) for massive particles. Even considering a dispersion relation that provides the following $$\begin{aligned} v & =\dfrac{dE}{dp}\simeq c\left(1-\frac{m^{2}c^{2}}{2p^{2}}-\alpha p^{2}m\lambda^{3}\right),\\ \dfrac{d^{2}E}{dp^{2}} & \simeq c\left(\frac{m^{2}c^{2}}{p^{3}}-2\alpha pm\lambda^{3}\right),\end{aligned}$$ where $\alpha$ is a further deformation parameter, we obtain that, for $\alpha>0$, the invariant speed would be reached at the invariant momentum $p_{I}\sim\sqrt[4]{mc^{2}/\left|\lambda\right|^{3}}$. Therefore, in all the analyzed cases, we find only minor departures from the behavior that the models exhibit with undeformed $K(v)$. On the compatibility of the approach with linear energy addition rules ---------------------------------------------------------------------- The second point that we want to address is the compatibility of our approach with the conservation of the energy for composite systems as analyzed in [@Mandanici:2007eb]. We have found so far that the covariance of the dispersion relation implies eq.(\[CovRelDisp\]) and that the covariance of the evolution equation $v=\left\{ x,E\right\} $, together with $\left\{ x,p\right\} =1$, implies eq.(\[CovSpeedExpress\]). In order to guarantee covariance to the rule of additivity of the energy (the momentum remaining not additive) according to [@Mandanici:2007eb] it is enough to choose $$\left\{ N,E\right\} =\pi(p),\hspace{1cm}\text{ }\left\{ N,p\right\} =E\left(\frac{d\pi}{dp}\right)^{-1},\label{laeaction}$$ where $\pi(p)$ is a general function of the physical momentum $p$ transforming according the usual special relativistic rules. From eq.(\[laeaction\]) (see [@Mandanici:2007eb]) follows $E=\sqrt{m^{2}+\pi^{2}}.$ Thus the point reduces to find a dispersion relation compatible with eqs.(\[laeaction\]) and with eqs.(\[CovRelDisp\])-(\[CovSpeedExpress\]). In order to achieve this, one has to solve the following system $$\begin{cases} E\left(\frac{d\pi}{dp}\right)^{-1}\frac{d^{2}E}{dp^{2}} & =K\left(\frac{dE}{dp}\right)\\ \frac{dE}{dp}E\left(\frac{d\pi}{dp}\right)^{-1} & =\pi \end{cases}\label{Sistema}$$ always with $E=\sqrt{m^{2}+\pi^{2}}$. Assuming again $K(v)$ to be undeformed, the second equation of (\[Sistema\]) is trivially satisfied for every $\pi(p)$. Instead the first equation implies $$\frac{d\pi}{dp}=1\Rightarrow\pi(p)=p.$$ Thus the only additive model compatible with our hypotheses, already in the 1+1-dimensional case, is the usual special-relativistic one. On the possibility to extend the analysis to 3+1 dimensions ----------------------------------------------------------- Finally let us came to the possibility to extended our analysis to 3+1 space-time dimensions. Particle speed in a boosted reference frame transforms according to the formula $$v_{i}^{\prime}=v_{i}+\xi_{j}\left\{ N_{j},p_{k}\right\} \dfrac{\partial^{2}E}{\partial p_{i}\partial p_{k}}.$$ The request of covariance of the energy-momentum dispersion relation in 3+1 dimensions becomes $$\left\{ N_{i},E\right\} =\dfrac{\partial E}{\partial p_{k}}\left\{ N_{i},p_{k}\right\} .\label{E3d}$$ Then the request that the boosted velocity depends only on the particle speed implies, in the 3+1-dimensional case, the following expression for the infinitesimal action of the boost on the spatial momentum: $$\left\{ N_{j},p_{k}\right\} =\left(\mathbf{H}^{-1}\mathbf{K}\right)_{jk},\label{p3d}$$ where we have defined $$\mathbf{H}_{ik}\mathbf{=}\dfrac{\partial^{2}E}{\partial p_{i}\partial p_{k}},\hspace{1cm}\text{ }\mathbf{K}_{ij}=K_{ij}\left(\frac{\partial E}{\partial p_{i}},\frac{\partial E}{\partial p_{j}}\right).$$ For the space-time sector in 3+1 space-time dimensions, we get by means of the formula $$v_{i}=\dfrac{\partial E}{\partial p_{i}}=\dfrac{dx_{i}}{dt},$$ and following the procedure of the 1+1-dimensional case, the boost actions on the space-time coordinates: $$\left\{ N_{j},x_{j}\right\} =\delta_{ij}t,\hspace{1cm}\left\{ N_{i},dt\right\} =\left\{ N_{i},\dfrac{dx_{j}}{v_{j}}\right\} =\dfrac{\delta_{ij}-K_{ij}}{v_{j}}dt.$$ Again we can maintain the usual Minkowski commutative space-time adopting the standard 3+1-dimensional expression $\mathbf{H}_{ik}=c^{2}E^{-1}\left(\delta_{ij}-p_{i}p_{j}c^{2}E^{-2}\right)$ and $\mathbf{K}_{ij}=\delta_{ij}-1/c^{2}\partial E/\partial p_{i}\partial E/\partial p_{j}$. However, in this 3+1-dimensional case, it can be shown by explicit calculations that the transformation rules constructed using formulas (\[E3d\])-(\[p3d\]) are not compatible with the group structure of the boost actions in the momentum space. Final remarks and conclusions {#S10} ============================= Under the hypothesis that the action of the boosts on the speed of a particle be a function of the particle speed alone, i.e. that $\left\{ N,v\right\} =K(v)$, we have addressed the issue of the relative rest in doubly special relativity. Our analysis has shown a way to construct, at least in 1+1 dimensions, double special relativity models in which the relative rest is not an inertial-observer dependent notion. The key point of our work is that our results can be obtained without renouncing to the usual $v=dE/dp$ formula for the particle speed. We have also argued that, in order our scheme to be fulfilled, there is no need to renounce to the usual (commutative) Minkowski space-time endowed with standard energy-independent boosts. From a conservative viewpoint we have mainly focused on the undeformed function $K(v)=1-v^{2}/c^{2}$. Within these assumptions we have found that those models in which particle speed approaches $v\simeq c$ admit invariant momentum $p_{I}$ and invariant energy $E_{I}$ that depend on the mass of the boosted particle. For instance, in the case of dispersion relation of the type $E^{2}-c^{2}p^{2}+\lambda p^{3}c^{2}=m^{2}c^{4}$ and for dispersion relation of the type DSR1, the invariant momentum has the form $p_{I}\sim\sqrt[3]{m^{2}c^{2}\lambda^{-1}/2}$. Considering the available bounds on photon mass $m_{\gamma}\lesssim10^{-18}eV/c^{2}$ and assuming $\lambda^{-1}\sim10^{19}GeV/c$ we would get for photons an invariant momentum $p_{I}$ $\lesssim10^{-3}eV/c$, that would be ruled out by available data. Instead, in the case of the DSR2 dispersion relation, we find that remains $p_{I}\sim\lambda^{-1}$ independently on the mass of the particle considered. Maintaining a Minkowski structure for space-time and a standard form of $K(v)$ also implies that the transformation laws of angular frequency $\omega$ and wavenumber $k$ must differ, in general, from those of $E$ and $p$, with influences on the phenomenology. In fact one would expect no deviations from special relativity when considering effects such as the Doppler effect (in space-time coordinates) and/or the time dilatation between inertial observers. Instead, as discussed, deviations from special relativity manifest when considering quantities involving the energy-momentum sector. Different form of $K(v)$ can modify the phenomenology but, within the class of functions that we have analyzed, no qualitative changes have emerged with respect to the described picture. The action of boosts on the energy-momentum space appears however incompatible with linear addition rules for energies in the sense of$\,$[@Mandanici:2007eb]. We also have found troublesome to extend the analysis to 3+1-spatial dimensions especially because of difficulties in maintaining the group structure of the boost actions. [10]{} G. Amelino-Camelia, “[Testable scenario for relativity with minimum-length]{},” [*Phys. Lett.*]{}, vol. B510, pp. 255–263, 2001. G. Amelino-Camelia, “[Relativity in space-times with short-distance structure governed by an observer-independent (Planckian) length scale]{},” [*Int. J. Mod. Phys.*]{}, vol. D11, pp. 35–60, 2002. J. Magueijo and L. Smolin, “[Lorentz invariance with an invariant energy scale]{},” [*Phys. Rev. Lett.*]{}, vol. 88, p. 190403, 2002. J. Kowalski-Glikman and S. Nowak, “[Doubly special relativity theories as different bases of kappa Poincare algebra]{},” [*Phys.Lett.*]{}, vol. B539, pp. 126–132, 2002. G. Amelino-Camelia, J. Kowalski-Glikman, G. Mandanici, and A. Procaccini, “[Phenomenology of doubly special relativity]{},” [*Int.J.Mod.Phys.*]{}, vol. A20, pp. 6007–6038, 2005. G. Amelino-Camelia, “[Doubly-Special Relativity: Facts, Myths and Some Key Open Issues]{},” [*Symmetry*]{}, vol. 2, pp. 230–271, 2010. J. Lukierski, H. Ruegg, and W. Zakrzewski, “[Classical and Quantum Mechanics of Free k-Relativistic Systems]{},” [*Annals of Physics*]{}, vol. 243, no. 1, pp. 90 – 116, 1995. J. Kowalski-Glikman and S. Nowak, “[Noncommutative space-time of doubly special relativity theories]{},” [*Int.J.Mod.Phys.*]{}, vol. D12, pp. 299–316, 2003. S. Mignemi, “[Transformations of coordinates and Hamiltonian formalism in deformed special relativity]{},” [*Phys.Rev.*]{}, vol. D68, p. 065029, 2003. P. Galan and G. A. Mena Marugan, “[Canonical realizations of doubly special relativity]{},” [*Int. J. Mod. Phys.*]{}, vol. D16, pp. 1133–1147, 2007. S. Ghosh and P. Pal, “[Deformed Special Relativity and Deformed Symmetries in a Canonical Framework]{},” [*Phys.Rev.*]{}, vol. D75, p. 105021, 2007. J. Magueijo and L. Smolin, “[Generalized Lorentz invariance with an invariant energy scale]{},” [*Phys.Rev.*]{}, vol. D67, p. 044017, 2003. D. Kimberly, J. Magueijo, and J. Medeiros, “[Nonlinear relativity in position space]{},” [*Phys.Rev.*]{}, vol. D70, p. 084007, 2004. S. Hossenfelder, “[Bounds on an energy-dependent and observer-independent speed of light from violations of locality]{},” [*Phys.Rev.Lett.*]{}, vol. 104, p. 140402, 2010. G. Amelino-Camelia, L. Freidel, J. Kowalski-Glikman, and L. Smolin, “[The principle of relative locality]{},” [*Phys.Rev.*]{}, vol. D84, p. 084010, 2011. G. Amelino-Camelia, M. Matassa, F. Mercati, and G. Rosati, “[Taming Nonlocality in Theories with Planck-Scale Deformed Lorentz Symmetry]{},” [ *Phys.Rev.Lett.*]{}, vol. 106, p. 071301, 2011. G. Amelino-Camelia, L. Barcaroli, and N. Loret, “[Modeling transverse relative locality]{},” [*Int.J.Theor.Phys.*]{}, vol. 51, pp. 3359–3375, 2012. P. Kosinski and P. Maslanka, “[On the definition of velocity in doubly special relativity theories]{},” [*Phys. Rev.*]{}, vol. D68, p. 067702, 2003. S. Mignemi, “[On the definition of velocity in theories with two observer-independent scales]{},” [*Phys. Lett.*]{}, vol. A316, pp. 173–176, 2003. R. Aloisio, A. Galante, A. F. Grillo, E. Luzio, and F. Mendez, “[Approaching space time through velocity in doubly special relativity]{},” [*Phys. Rev.*]{}, vol. D70, p. 125012, 2004. G. Amelino-Camelia, F. D’Andrea, and G. Mandanici, “[Group velocity in noncommutative spacetime]{},” [*JCAP*]{}, vol. 0309, p. 006, 2003. G. Amelino-Camelia, J. R. Ellis, N. Mavromatos, D. V. Nanopoulos, and S. Sarkar, “[Tests of quantum gravity from observations of gamma-ray bursts]{},” [*Nature*]{}, vol. 393, pp. 763–765, 1998. R. Gambini and J. Pullin, “[Nonstandard optics from quantum space-time]{},” [*Phys.Rev.*]{}, vol. D59, p. 124021, 1999. G. Mandanici, “[Undeformed (additive) energy conservation law in Doubly Special Relativity]{},” [*Mod. Phys. Lett.*]{}, vol. A24, pp. 739–745, 2009. [^1]: We emphasize here that this formula that works well in classical physics and quantum mechanics can also be derived in any Hamiltonian approach $v=\dot{x}=$ $\left\{ x,E(p)\right\} $ also compatible with $\left\{ x,p\right\} =1$, being $v=\left\{ x,E\right\} =\left\{ x,p\right\} dE/dp=dE/dp.$ [^2]: The possibility of introducing a second velocity scale in this framework is not void of interest and can deserve further investigations. [^3]: Notice that is this case the denominator of $\left\{ N,p\right\} $ does not vanish so that $\left\{ N,p\right\} $ does not diverge.

ArXiv

--- abstract: 'We discuss recent observations of stars located close to the symmetry plane of the Milky Way, and examine them in the context of theories of Galaxy formation and evolution. The kinematics, ages, and compositions of thin disk stars in the solar neighborhood display complex patterns, and interesting correlations. The Galactic disk does not seem to pose any unsurmountable obstacles to hierarchical galaxy formation theories, but a model of the Milky Way able to reproduce the complexity found in the data will likely require a meticulous study of a significant fraction of the stars in the Galaxy. Making such an observational effort seems necessary in order to make a physics laboratory out of our own galaxy, and ultimately ensure that the most relevant processes are properly understood.' --- Introduction ============ Large galaxies naturally produce disks. Radiative cooling of the gas and angular momentum conservation lead the early evolution of galaxies through dissipative collapse and disk formation. Disks are frequently observed in galaxies, even at high redshift (Förster Schreiber et al. 2006), and the Milky Way does not seem unique at all in showing a [*dual*]{} disk, with a distinct thin and thick components (Dalcanton & Bernstein 2002). Early attempts to place the Galactic thin disk in the context of a $\Lambda-$CDM universe exposed a number of problems. The number of observed surviving satellites appeared far too small compared to simulations. This problem is somewhat alleviated after the Sloan Digital Sky Survey (SDSS) has identified many new low-surface brightness galaxies in the immediate Galactic neighborhood (but see Koposov et al. 2008). It was also deemed hard for the disk to survive for as long as the observations suggested, 8–10 Gyr, and in particular to stay thin (Tóth & Ostriker 1992, Kauffmann & White 1993). More recent appraisals, however, indicate that as many as 85% of disk galaxies have not been involved in a merger with a mass ratio larger than 0.5 since redshift $\sim 1-1.5$, or approximately in the last 8 Gyr (Koda et al. 2009). A higher gas fraction in the accreted building blocks seems to favor disk survival (Hopkins et al. 2009). The structure of galaxy disks is usually studied measuring surface brightness distributions. Anomalies such as spiral arms and HII regions are smoothed out, taking de-projected azimuthal averages in nearly face-on galaxies, modeling the radial dependence of the light distribution with exponential profiles (see, e.g., Aguerri et al. 2000; Prieto et al. 2001). Edge-on galaxies are, in turn, used to study the light distribution perpendicular to the plane. In the Milky Way, the spatial distribution of stars is studied using deep imaging surveys – counting stars and exploiting photometric calibrations to estimate the luminosity of the main-sequence as a function of color (e.g. Jurić et al. 2008). Being [*insiders*]{} to the Galaxy provides some advantages; for example, we can measure in detail the properties of individual stars using spectroscopy. Modern surveys employ cameras with a very broad dynamical range, and massive multiplexing capabilities for spectroscopy (see, e.g., Gunn et al. 1998, 2005; Onaka et al. 2008), making it feasible to obtain large data sets fast. Main structural components of the Milky Way =========================================== There are multitude of studies of the main Galactic components using star counts. Two recent studies by Cabrera-Lavers et al. (2007) and Jurić et al. (2008) based on 2MASS (Skrutskie et al. 2006) and the SDSS (Abazajian et al. 2009), respectively, sample quite well the Galactic thin disk. These surveys are dominated by late-type (mainly K and M dwarf) stars, and their scale height is found to be about 200–300 pc. There is no consensus on the (radial) scale length of the thin disk, and estimates range between 2.5 to 3.5 kpc, but the larger distances involved make this measurement harder. The thin disk is thought to contribute about 85% of the stars in the Galactic plane. It is important to emphasize that the disk scale heights are expected to vary depending on where we look (Bilir et al. 2008), as the potential is far from perfectly smooth and axisymmetric. More dramatic is the variation of the scale height of the thin disk with age. Thus, Maíz-Apellaniz (2001) finds $h\sim 35$ pc from OB type stars – a value nearly 10 times smaller than determinations from late-type stars. This strong age dependence is also imprinted in the distribution of M-dwarfs observed spectroscopically in the SDSS, and West et al. (2004, 2006, 2008) find a decreasing fraction of active stars (the youngest) as they sample farther away from the Galactic plane. Further evidence is also seen in the stellar kinematics in the solar vicinity, which are discussed in the next section. Stellar kinematics in the Solar Neighborhood ============================================ The combination of Hipparcos (and Tycho) astrometry with radial velocities from high-resolution spectroscopy has provided detailed coordinates in phase space for stars in the solar neighborhood ($< 100$ pc from us). Observations with the CORAVEL spectrographs (Baranne et al. 1979) have been used to study large samples of giants (Famaey et al. 2005) and especially F- and G-type dwarfs (the Geneva-Copenhagen survey, described in Nordström et al. 2004 and Holmberg et al. 2007, 2009). The velocity distributions of nearby stars show plenty of structure. We discuss structure in more detail in the following section, and here we focus on other characteristics. Each of the velocity components of the thin disk ($U$, $V$ and $W$ for the radial, azimuthal, and vertical components in a cylindrical coordinate system; see Fig. \[fe\]) shows distributions that increase in width with age. The scatter in the vertical velocities ($W$) shows the sharpest and smoothest rise with age of the three velocity components from $\sigma \sim 8$ km s$^{-1}$ for stars that are $\sim 1$ Gyr old to roughly 30 km s$^{-1}$ for stars $\sim 10$ Gyr old. This is another way of looking at the increase in scale height with time. For a Mestel disk (surface density $\Sigma \propto R^{-1}$, where $R$ is the galactocentric distance), assuming an isothermal sheet $\rho \propto \cosh^{-2}(z)$, embedded on a spherical halo ($\rho \propto r^{-2}$, where $r$ is the radial coordinate), the disk rotational velocity $V_{\rm rot}$ is constant, and the vertical velocity dispersion can be written $$\sigma^2 = \frac{1}{2} V_{\rm rot}^2 \frac{h}{R_{\odot}} \left( \mu + \lambda(1-\mu) \frac{h}{R_{\odot}} \right),$$ (Tóth & Ostriker 1992), where $\lambda \equiv \int x^2 \cosh^{-2}(x) dx \simeq$ 1.645, $\mu(R)$ is the enclosed mass ratio (disk/total, about 0.35 for the Milky Way), and adopting $V_{\rm rot}\simeq 220$ km s$^{-1}$, $$\sigma^2 \simeq 24200 \frac{h}{R_{\odot}} \left(0.35 + 1.10 \frac{h}{R_{\odot}}\right).$$ This indicates that a range in $\sigma$ between 8 and 30 km s$^{-1}$, as observed for stars between 1 and 10 Gyr should be matched by an excursion in the scale height between 60 and 600 pc. Measuring directly scale heights from astrometry will have to wait for Gaia (Lindegren & de Bruijne 2005; Lindegren et al. 2009), as the Hipparcos parallaxes for late-type dwarfs are limited to $\sim 100$ pc. Structure in the Galactic disk ============================== Eggen is often singled out as the pioneer of the study of comoving groups of stars (superclusters and moving groups) in the solar vicinity (see, e.g., Eggen 1992). Recent data sets have shown these structures with sharper contrast, and revealed new ones (see, e.g. Famaey et al. 2005; Arifyanto & Fuchs 2006). Large surveys such as RAVE (e.g. Klement et al. 2008), and ultimately Gaia, are to provide improved statistics that will bring light on the important topic of the origin of superclusters and their connection to field stars and proper clusters. Previous work has demonstrated that some superclusters are indeed dissolving stellar clusters (e.g., the HR 1614 moving group, which has been found to exhibit a single age and metallicity by De Silva et al. 2007). But many others that have been scrutinized appear to exhibit broad age spans (e.g. the Pleiades, Hyades, or Hercules superclusters; Famaey et al 2008), or chemical abundance distributions (e.g. Hercules; Bensby et al. 2007), which suggests they have a dynamical origin (e.g. De Simone et al. 2004; Quillen & Minchev 2005; Chakrabarty 2007). Note that the accretion of an external stellar system may offer, in some cases, a plausible formation scenario, with stars directly brought in with common kinematics, or simply ’linked’ as a result of an accretion event (Minchev et al. 2009; Quillen et al. 2009). Ages and metallicities of disk stars ==================================== Being able to date individual stars is extremely valuable. The recovery of the star formation history of the stars in the solar neighborhood from the inversion of observed HR diagrams or chromospheric age estimates has been attempted in many studies (Rocha-Pinto et al. 2000; Hernández et al. 2000; Bertelli & Nasi 2001; Aumer & Binney 2009). Unfortunately, an examination of these and other works does not provide a coherent picture. Isochrone dating is mostly limited to subgiants, as it is at the turn-off where the basic fundamental parameters, mainly the luminosity, change quickly, with minimal degeneracy; i.e. isochrones spread nicely. Surface gravity determinations from spectroscopy are hardly useful, since spectra are only weakly sensitive to pressure, and fundamental measurements such as trigonometric parallaxes and angular diameters are best to constrain stellar luminosities. Gaia will dramatically change this field with parallaxes accurate to $\sim 20$ $\mu$m at 15 magnitude. Paying attention to details, in particular applying a rigorous statistical analysis, is important, and in some extreme cases critical. The last few years have seen a change in the methodologies for determining stellar ages, from the crude method of assigning the nearest isochrone to sophisticated statistical analyses (see, e.g., Reddy et al. 2003; Pont & Eyer 2004; J[ø]{}rgensen & Lindegren 2005). It has been emphasized by Lachaume et al. (1999) that different dating techniques are complementary. For example, isochrones are most useful for turn-off stars, and hence more likely applicable to intermediate mass stars, while activity and rotation can provide ages for low mass stars that stay on the main sequence longer than a Hubble time. Thin disk stars show a wide range of ages. Reddy et al. (2006) estimated ages between 1 and 9 Gyr, although predominantly $<5$ Gyr. Holmberg et al. (2009) and Haywood (2008), using larger samples, found wider ranges reaching up to 13–14 Gyr, although most concentrated again at $<4-5$ Gyr. Very old ages for thin-disk stars may be at odd with the upper limit to the age of the disk derived from the analysis of the white dwarf cooling sequence. For example, Leggett al. (1998) estimated $8 \pm 1.5$ Gyr, although a critical assessment of the literature by Fontaine, Brassard & Bergeron (2001) led them to propose a plausible range between 8.5 to 11 Gyr. ![Velocity and metallicity histograms for 2427 stars with metallicities from the catalog of Cayrel de Strobel, Soubiran & Ralite (2001), Hipparcos astrometry, and radial velocities from the compilations by Malaroda, Levato & Galliani (2001) and/or Barbier-Brossat & Figon (2000), showing velocities $V> -40$ and $-80 < W < +50$ km s$^{-1}$. The expected contamination by thick disk and halo populations is very small, and therefore we identify the observed distributions with the thin disk. The smooth solid lines are Gaussian curves fitted to the histograms. Note the peak associated with the Hyades at ($U,V,W$) $\simeq$ ($-43, -18, -2$) km s$^{-1}$. Adopted from Allende Prieto et al. (2004).[]{data-label="fe"}](carlos_f1.eps){width="3.4in"} An inspection of the literature in the last decade shows that there is now consensus regarding the metallicity distribution of the thin disk. Most authors find it is reasonably Gaussian, with a standard deviation of about 0.2 dex (see, e.g. Allende Prieto et al. 2004; Holmberg et al. 2007). More polemic is the exact mean of the distribution, which some argue could be as low as \[Fe/H\]=$-0.10$ dex, while others push for a value much close to solar (hence around $0.00$; see Luck & Heiter 2007; Haywood 2001; Taylor & Croxall 2005; Fuhrmann 2008). As noted some years ago, it is interesting that blindly adopting the metallicities compiled in the Cayrel de Strobel et al. catalog, and just by simply cleaning thick disk stars with [*slow*]{} Galactic rotation ($V<$ 50 km s$^{-1}$[^1]), one recovers again a \[Fe/H\] distribution centered at $-0.1$ dex with $\sigma\simeq 0.20$ dex – remarkably close to those found from the analysis of much more homogeneous data sets (see Fig. \[fe\]). The implication is that despite this compilation includes high-resolution determinations from many studies, the systematics across samples and analysis protocols are not large enough to widen the derived distribution. An additional complication should be noted. Although the metallicity distribution of the thin disk may be well determined, this is likely not an ideal quantity to compare with chemical evolution models. Selection effects such as mass biases due to different lifetimes need to be considered, and so does the role of radial migration, which could be bringing significant numbers of stars formed at different galactocentric distances, and hence with different compositions even if formed at exactly the same time (see, e.g., Haywood et al. 2008). Abundance ratios ================ There have been a multitude of studies performing high-resolution spectroscopy of nearby GFK stars (e.g., Edvardsson et al. 1993; Feltzing & Gustafsson 1998; Chen et al. 2000; Nissen et al. 2000; Fulbright 2002; Reddy et al. 2003, 2006; Takeda 2007; Ecuvillon et al. 2004; Gilli et al. 2006; Bensby et al. 2003; Ramírez et al. 2007; Fuhrmann 1998, 2004, 2008). Most of these studies found a remarkable uniformity in the abundance ratios for thin disk stars at any given \[Fe/H\]. Reddy et al. (2003) looked for and failed to find a cosmic scatter in the abundance ratios. Assuming \[Fe/H\] is a reliable clock, the interstellar medium where these samples formed was very well mixed. Many works encounter non-solar ratios at solar \[Fe/H\] for some elements. This puzzling result, which might fuel the idea of the Sun being somehow special, was later traced to systematic errors in the abundances associated with using the Sun as a reference for non-solar type stars (Allende Prieto 2008). There is no doubt that highly homogeneous samples, in particular those restricted to a narrow range in effective temperature ([*isothermal*]{} samples, if you will), can dramatically reduce systematic errors still present in the analyses. Meléndez et al., in these proceedings, show an extreme example of exploiting such a trick. A dichotomy between the thin and thick disks? ============================================= The thin and thick disks stars in the solar neighborhood can be easily separated, at least statistically. Although the distributions of the velocity components and metallicities overlap somewhat, combining all the data for $UVW$ as well as \[Fe/H\], makes their separation fairly straightforward. The age distributions have probably very little overlap, if any at all (see, e.g., Fig. 24 in Reddy et al. 2006). Star formation in the Milky Way has likely proceed in phases, with limited overlap: halo, thick disk, and thin disk. Yet, the connection between these three components, and in particular the thick and thin disks, is far from understood. Looking closely at the chemical compositions, a sharp distinction between the two disks has become evident in the abundances of many elements, such as the $\alpha$-capture nuclei (O, Mg, Si, S, Ca, Ti). This is illustrated in Fig. \[alfe\], borrowed from Reddy et al. (2006). Some argue there exists genuine transition objects, but not all studies find them in their samples. Galactic disks can become thinner by dissipative collapse while they are still rich in gas. Such straightforward connection between the thick and thin disks does not seem viable in the light of the distinct chemical patterns that separate the two Milky Way disks. Stellar disks can also become thicker with time, due to internal (scattering and other dynamical interactions in the disk) or external (satellite accretion) mechanisms. But again, such simple path does not match the distribution of ages. ![Average of the abundance ratios of Mg, Si, Ca and Ti to Fe for stars kinematically assigned to the thin (crosses) and thick disk (filled circles). The filled circles surrounded with an additional circumference are the so-called [*TKTA*]{} stars; they show thin-disk abundances but thick-disk kinematics. Adopted from Reddy et al. (2006).[]{data-label="alfe"}](carlos_f2.eps){width="3.4in"} Two scenarios recently proposed in the literature appear feasible. Modelers have argued for some time that mergers could have been responsible to produce the thick disk, perhaps disrupting a previously existing disk, but allowing the thin disk to form and evolve independently afterwards. While the accretion of dry (gasless) system(s) may lead to features that clash with existing observations (think rings of stars and thick-disk characteristics that vary with galactocentric distance), the acquisition of gas rich systems and on-the-fly formation of stars seems to work well (Brook et al. 2005, 2007). The second scenario is based on new, relatively simple, models which indicate that the observed distributions of metallicity, age, and $\alpha$/Fe ratios could emerge as a result of the natural density gradient and the associated variation of the star formation rate with galactocentric distance, when coupled to radial mixing of stellar orbits (Schönrich & Binney 2008, 2009). The disk beyond the solar neighborhood ====================================== The work in the solar neighborhood can now be complemented with observations of more distant stars and stellar systems. These large-range data sets will be most useful to discriminate among the proposed formation scenarios for the thick disk and its connection to thin disk and halo. Data on individual stars from the SDSS, which now accumulates close to 0.5 million stellar spectra, can tell us about abundance and kinematics of stars over a wide range of distances, from tens to hundreds of pc using low-mass stars, to tens of kpc for bright giants. An interesting hint from SDSS is that the median of the metallicity distribution of the thick disk, about \[Fe/H\]$= -0.7$ dex, does not seem to vary between 4 $<R<$ 14 kpc, in distinct contrast with observations for thin disk stars, where significant gradients are found for multiple elements using different tracers (see Fig. 13 in Allende Prieto et al. 2006). Both Cepheids (Andrievsky et al. 2002, 2004) and giants in open clusters (Yong et al. 2006) allow tracing the abundances of many elements at large galactocentric distances. A remarkable outcome of these studies is that the well-known thin disk abundance gradient (see the review by Maciel in these proceedings) may flatten out at $R> 12$ kpc (but see Sale et al. 2009 a for discrepant voice). This has been suggested to indicate a flat density profile in the inner stellar halo (Cescutti et al. 2007). Interestingly enough, these studies also show that the $\alpha/Fe$ ratios increase with galactocentric distance – and so do the ratios of lanthanum (an $s-$process tracer) as well as europium (an $r-$process tracer) to iron. The disk is by no means flat, and a better understanding of its structure is needed, in particular the flare and warp traced by stars (López-Corredoira et al. 2002; Momany et al. 2006) gas (Kalberla et al. 2007; Levine et al. 2006), and dust (Drimmel & Spergel 2001). Infrared spectroscopic observations of vast numbers of red giants across the disk should provide much insight. APOGEE, part of SDSS-III, plans to obtain high-resolution $H$-band spectra for 10$^5$ stars with a signal-to-noise ratio approaching 100 between 2011 and 2014 (Allende Prieto et al. 2008). Preliminary studies suggest that more than 15 chemical elements can be sampled within the $H$ band, where dust obscuration is 5 times less than in $V$. Closing remarks =============== The thin disk of the Galaxy likely fits in the overall bottom-up galaxy formation scenario in a $\Lambda$CDM universe, but a detailed picture of its formation is still missing. The stellar population of the thin disk is rich in kinematic structure, but appears chemically well-mixed. The two statements in the previous sentence need not be in contradiction, as fine structure is likely missed due to limited abundance precision (currently $\sim$0.05 dex), and especially if most of the structure has a dynamical origin, excited by resonances and/or (modest) accretion. The solar neighborhood needs to be placed in the context of the whole Galactic disk. Massive surveys of faint stars will do that, and they will happen over the next 5–10 years. Among the most pressing questions in this field, we could single out: What is has been the star formation history of the solar neighborhood? (Must consider radial mixing!). Which process(es) are mainly responsible for the stellar ÔclusteringÕ in phase space and the disk heating? What is the connection between the thin and thick disks? The tools to address these questions are already in place: global astrometry methods, accurate spectroscopy from efficient instruments, detailed chemical analysis techniques (ÔisothermalÕ samples, larger samples, refined analyses), chemo-dynamical modeling, improved statistical techniques, and last but not least, data, to be provided by SDSS, RAVE, Gaia, and other supporting facilities. Abazajian, K. N., et al. 2009, *ApJS*, 182, 543 Aguerri, J. A. L., Varela, A. M., Prieto, M., & Mu[ñ]{}oz-Tu[ñ]{}[ó]{}n, C. 2000, *AJ*, 119, 1638 Allende Prieto, C.  2008, The Metal-Rich Universe, 30 Allende Prieto, C., Beers, T. C., Wilhelm, R., Newberg, H. J., Rockosi, C. M., Yanny, B., & Lee, Y. S. 2006, *ApJ*, 636, 804 Allende Prieto, C., Barklem, P. S., Lambert, D. L., & Cunha, K. 2004, *A&A*, 420, 183 Allende Prieto, C., et al. 2008, *Astronomische Nachrichten*, 329, 1018 Andrievsky, S. M., Luck, R. E., Martin, P., & L[é]{}pine, J. R. D. 2004, *A&A*, 413, 159 Andrievsky, S. M., Kovtyukh, V. V., Luck, R. E., L[é]{}pine, J. R. D., Maciel, W. J., & Beletsky, Y. V. 2002, *A&A*, 392, 491 Arifyanto, M. I., & Fuchs, B. 2006, *A&A*, 449, 533 Aumer, M., & Binney, J. J. 2009, *MNRAS*, 397, 1286 Baranne, A., Mayor, M., & Poncet, J. L. 1979, *Vistas in Astronomy*, 23, 279 Barbier-Brossat, M., & Figon, P. 2000, *A&AS*, 142, 217 Bensby, T., Feltzing, S., & Lundstr[ö]{}m, I. 2003, *A&A*, 410, 527 Bensby, T., Oey, M. S., Feltzing, S., & Gustafsson, B. 2007, *ApJ*, 655, L89 Bertelli, G., & Nasi, E. 2001, *AJ*, 121, 1013 Bilir, S., Cabrera-Lavers, A., Karaali, S., Ak, S., Yaz, E., & L[ó]{}pez-Corredoira, M. 2008, *Publications of the Astronomical Society of Australia*, 25, 69 Brook, C., Richard, S., Kawata, D., Martel, H., & Gibson, B. K. 2007, *ApJ*, 658, 60 Brook, C. B., Gibson, B. K., Martel, H., & Kawata, D. 2005, *ApJ*, 630, 298 Cabrera-Lavers, A., Bilir, S., Ak, S., Yaz, E., & L[ó]{}pez-Corredoira, M. 2007,*A&A*, 464, 565 Cayrel de Strobel, G., Soubiran, C., & Ralite, N. 2001, *A&A*, 373, 159 Cescutti, G., Matteucci, F., Fran[ç]{}ois, P., & Chiappini, C. 2007, *A&A*, 462, 943 Chakrabarty, D. 2007, *A&A*, 467, 145 Chen, Y. Q., Nissen, P. E., Zhao, G., Zhang, H. W., & Benoni, T. 2000, *A&AS*, 141, 491 Dalcanton, J. J., & Bernstein, R. A. 2002, *AJ*, 124, 1328 De Silva, G. M., Freeman, K. C., Bland-Hawthorn, J., Asplund, M., & Bessell, M. S. 2007, *AJ*, 133, 694 De Simone, R., Wu, X., & Tremaine, S. 2004, *MNRAS*, 350, 627 Drimmel, R., & Spergel, D. N. 2001, *ApJ*, 556, 181 Ecuvillon, A., Israelian, G., Santos, N. C., Mayor, M., Villar, V., & Bihain, G. 2004, *A&A*, 426, 619 Edvardsson, B., Andersen, J., Gustafsson, B., Lambert, D. L., Nissen, P. E., & Tomkin, J. 1993, *A&A*, 275, 101 Eggen, O. J. 1992, *AJ*, 104, 2141 Famaey, B., Jorissen, A., Luri, X., Mayor, M., Udry, S., Dejonghe, H., & Turon, C. 2005, *A&A*, 430, 165 Famaey, B., Siebert, A., & Jorissen, A. 2008, *A&A*, 483, 453 Feltzing, S., & Gustafsson, B. 1998, *A&AS*, 129, 237 Fontaine, G., Brassard, P., & Bergeron, P. 2001, *PASP*, 113, 409 F[ö]{}rster Schreiber, N. M., et al. 2006,*ApJ*, 645, 1062 Fuhrmann, K. 2004, *Astronomische Nachrichten*, 325, 3 Fuhrmann, K. 1998, *A&A*, 338, 161 Fuhrmann, K. 2008, *MNRAS*, 384, 173 Fulbright, J. P. 2002, *AJ*, 123, 404 Gilli, G., Israelian, G., Ecuvillon, A., Santos, N. C., & Mayor, M. 2006, *A&A*, 449, 723 Gunn, J. E., et al. 1998, *AJ*, 116, 3040 Gunn, J. E., et al. 2006, *AJ*, 131, 2332 Haywood, M. 2001, *MNRAS*, 325, 1365 Haywood, M. 2008, *MNRAS*, 388, 1175 Hernandez, X., Valls-Gabaud, D., & Gilmore, G. 2000, *MNRAS*, 316, 605 Holmberg, J., Nordstr[ö]{}m, B., & Andersen, J. 2009, *A&A*, 501, 941 Holmberg, J., Nordstr[ö]{}m, B., & Andersen, J. 2007,*A&A*, 475, 519 Hopkins, P. F., Cox, T. J., Younger, J. D., & Hernquist, L. 2009, *ApJ*, 691, 1168 Juri[ć]{}, M., et al. 2008, *ApJ*, 673, 864 bibitem\[J[ø]{}rgensen & Lindegren(2005)\][2005AA...436..127J]{} J[ø]{}rgensen, B. R., & Lindegren, L. 2005, *A&A*, 436, 127 Kalberla, P. M. W., Dedes, L., Kerp, J., & Haud, U. 2007, *A&A*, 469, 511 Kauffmann, G., & White, S. D. M. 1993, *MNRAS*, 261, 921 Klement, R., Fuchs, B., & Rix, H.-W. 2008, *ApJ*, 685, 261 Koda, J., Milosavljevi[ć]{}, M., & Shapiro, P. R. 2009, *ApJ*, 696, 254 Koposov, S., et al.  2008, *ApJ*, 686, 279 Ma[í]{}z-Apell[á]{}niz, J. 2001, *AJ*, 121, 2737 Lachaume, R., Dominik, C., Lanz, T., & Habing, H. J. 1999, *A&A*, 348, 897 Leggett, S. K., Ruiz, M. T., & Bergeron, P. 1998, *ApJ*, 497, 294 Levine, E. S., Blitz, L., & Heiles, C. 2006, *ApJ*, 643, 881 Lindegren, L., et al. 2008, IAU Symposium, 248, 217 Lindegren, L., & de Bruijne, J. H. J. 2005, Astrometry in the Age of the Next Generation of Large Telescopes, 338, 25 L[ó]{}pez-Corredoira, M., Cabrera-Lavers, A., Garz[ó]{}n, F., & Hammersley, P. L. 2002, *A&A*, 394, 883 Luck, R. E., & Heiter, U. 2007, *AJ*, 133, 2464 Malaroda, S., Levato, H., & Galliani, S. 2001, VizieR Online Data Catalog, 3216, 0 (see Malaroda et al. 2000, *A&AS*, 144, 1 for more information) Minchev, I., Quillen, A. C., Williams, M., Freeman, K. C., Nordhaus, J., Siebert, A., & Bienaym[é]{}, O. 2009, *MNRAS*, 396, L56 Momany, Y., Zaggia, S., Gilmore, G., Piotto, G., Carraro, G., Bedin, L. R., & de Angeli, F. 2006, *A&A*, 451, 515 Nordstr[ö]{}m, B., et al. 2004, *A&A*, 418, 989 Onaka, P., Tonry, J. L., Isani, S., Lee, A., Uyeshiro, R., Rae, C., Robertson, L., & Ching, G. 2008, *SPIE Proc.*, 7014, Pont, F., & Eyer, L. 2004, *MNRAS*, 351, 487 Prieto, M., Aguerri, J. A. L., Varela, A. M., & Mu[ñ]{}oz-Tu[ñ]{}[ó]{}n, C. 2001, *A&A*, 367, 405 Quillen, A. C., & Minchev, I. 2005, *AJ*, 130, 576 Quillen, A. C., Minchev, I., Bland-Hawthorn, J., & Haywood, M. 2009, *MNRAS*, 397, 1599 Ram[í]{}rez, I., Allende Prieto, C., & Lambert, D. L. 2007, *A&A*, 465, 271 Reddy, B. E., Lambert, D. L., & Allende Prieto, C. 2006, *MNRAS*, 367, 1329 Reddy, B. E., Tomkin, J., Lambert, D. L., & Allende Prieto, C. 2003, *MNRAS*, 340, 304 Rocha-Pinto, H. J., Scalo, J., Maciel, W. J., & Flynn, C. 2000, *A&A*, 358, 869 Sale, S. E., et al. 2009, arXiv:0909.3857 (to appear in MNRAS) Sch[ö]{}nrich, R., & Binney, J. 2009, *MNRAS*, 399, 1145 Sch[ö]{}nrich, R., & Binney, J. 2009, *MNRAS*, 396, 203 Skrutskie, M. F., et al. 2006, *AJ*, 131, 1163 Takeda, Y. 2007, *PASJ*, 59, 335 Taylor, B. J., & Croxall, K. 2005, *MNRAS*, 357, 967 Toth, G., & Ostriker, J. P. 1992, *ApJ*, 389, 5 West, A. A., Hawley, S. L., Bochanski, J. J., Covey, K. R., Reid, I. N., Dhital, S., Hilton, E. J., & Masuda, M. 2008, *AJ*, 135, 785 West, A. A., Bochanski, J. J., Hawley, S. L., Cruz, K. L., Covey, K. R., Silvestri, N. M., Reid, I. N., & Liebert, J. 2006, *AJ*, 132, 2507 West, A. A., et al. 2004, *AJ*, 128, 426 Yong, D., Carney, B. W., & Teixera de Almeida, M. L. 2005, *AJ*, 130, 597 [^1]: Thick disk stars lag behind the thin disk rotation by roughly that much, although this depends on the distance from the plane.

ArXiv

--- abstract: 'Kernel methods have revolutionized the fields of pattern recognition and machine learning. Their success, however, critically depends on the choice of kernel parameters. Using Gaussian process (GP) classification as a working example, this paper focuses on Bayesian inference of covariance (kernel) parameters using Markov chain Monte Carlo (MCMC) methods. The motivation is that, compared to standard optimization of kernel parameters, they have been systematically demonstrated to be superior in quantifying uncertainty in predictions. Recently, the Pseudo-Marginal MCMC approach has been proposed as a practical inference tool for GP models. In particular, it amounts in replacing the analytically intractable marginal likelihood by an unbiased estimate obtainable by approximate methods and importance sampling. After discussing the potential drawbacks in employing importance sampling, this paper proposes the application of annealed importance sampling. The results empirically demonstrate that compared to importance sampling, annealed importance sampling can reduce the variance of the estimate of the marginal likelihood exponentially in the number of data at a computational cost that scales only polynomially. The results on real data demonstrate that employing annealed importance sampling in the Pseudo-Marginal MCMC approach represents a step forward in the development of fully automated exact inference engines for GP models.' author: - title: 'Bayesian Inference for Gaussian Process Classifiers with Annealing and Pseudo-Marginal MCMC' --- Introduction ============ Kernel methods have revolutionized the fields of pattern recognition and machine learning due to their nonlinear and nonparametric modeling capabilities [@Shawe-Taylor04]. Their success, however, critically depends on the choice of kernel parameters. In applications where accurate quantification of uncertainty in predictions is of primary interest, it has been argued that optimization of kernel parameters may not be desirable, and that inference using Bayesian techniques represents a much more reliable alternative [@FilipponeAOAS12; @FilipponeIEEETPAMI14; @Neal99; @Rue09; @Taylor12]. This paper focuses in particular on the problem of inferring covariance (kernel) parameters of Gaussian Process classification models using Markov chain Monte Carlo (MCMC) techniques. The choice of using GP classification as a working example is that they are formulated in probabilistic terms and are therefore particularly suitable candidates for carrying out Bayesian inference of their kernel parameters. The choice of employing MCMC based inference techniques is that for general GP models and for general kernels they offer an all-purpose solution to do so up to a given precision [@Flegal08], as discussed in [@Neal99; @FilipponeML13; @Neal93]. The formulation of GP classifiers, and that of GP models in general, makes use of a set of latent variables ${\mathbf{f}}$ that are assumed to be distributed according to a GP prior with covariance parameterized by a set of parameters ${\boldsymbol{\theta}}$. The application of MCMC to directly draw samples form the posterior distribution over covariance parameters would require the evaluation of the so called marginal likelihood, namely the likelihood where latent variables are integrated out of the model, which is analytically intractable. Recently, the Pseudo-Marginal (PM) MCMC approach has been proposed as a practical way to efficiently infer covariance parameters in Gaussian process classifiers exactly [@FilipponeIEEETPAMI14]. In this approach, computations do not rely on the actual marginal likelihood, but on an unbiased estimate obtained by approximate methods and Importance Sampling (IS). While the sampling of covariance parameters using PM MCMC improves on previous approaches for inferring covariance parameters, a large variance in the estimate of the marginal likelihood can negatively impact the efficiency of the PM MCMC approach, making convergence slow and efficiency low. In [@FilipponeIEEETPAMI14], IS was based on an importance distribution obtained by Gaussian approximations to the posterior over latent variables [@Rasmussen06; @Kuss05; @Nickisch08]. For certain values of the covariance parameters, the posterior over latent variables can be strongly non-Gaussian and the approximation can be poor, thus leading to a large variance in the IS estimate of the marginal likelihood [@Kuss05]. This effect is exacerbated by the dimensionality of the problem that makes the variance of IS grow exponentially large [@Neal01]. In the case of GP classification, estimating the marginal likelihood entails an integration in as many dimensions as the number of data, so this effect might be problematic in the case of large data sets. This paper presents the application of Annealed Importance Sampling (AIS) [@Neal01] to obtain a low-variance unbiased estimate of the marginal likelihood[^1]. This paper empirically demonstrate that compared to IS, AIS can reduce the variance of the estimate of the marginal likelihood exponentially in the number of data at a computational cost that scales only polynomially. Finally, two versions of PM MCMC approaches, employing AIS and IS respectively, are compared on five real data sets. The results on these data demonstrate that employing AIS in the PM MCMC approach represents a step forward in the development of fully automated exact Bayesian inference engines for GP classifiers. The remainder of this paper is organized as follows. Sections \[sec:gp\] and \[sec:ea:gp\] review GP models and their fully Bayesian treatment using the PM MCMC approach. Section \[sec:ais\] presents AIS to obtain an unbiased estimate of the marginal likelihood in GP models that can be used in the PM MCMC approach. Section \[sec:results\] reports results on synthetic and real data, and section \[sec:conclusions\] reports the conclusions. Bayesian Inference for GP Classification {#sec:gp} ======================================== Let $X = \{{\mathbf{x}}_1, \ldots, {\mathbf{x}}_n\}$ be a set of $n$ input data where ${\mathbf{x}}_i \in R^{d}$, and let ${\mathbf{y}}= \{y_1, \ldots, y_n\}$ be a set of associated observed binary responses $y_i \in \{-1, +1\}$. GP classification models are a class of hierarchical models where labels ${\mathbf{y}}$ are modeled as being independently distributed according to a Bernoulli distribution. The probability of class $+1$ for an input ${\mathbf{x}}_i$ is based on a latent variable $f_i$ and is defined as $p(y_i=+1 | f_i) = \Phi(f_i)$, where $\Phi$ is the cumulative distribution function of the standard normal distribution, so that $p({\mathbf{y}}| {\mathbf{f}}) = \prod_{i = 1}^n \Phi(y_i f_i)$. Latent variables ${\mathbf{f}}= \{f_1, \ldots, f_n\}$ are assumed to be distributed according to a GP prior, where a GP is a set of random variables characterized by the fact that any finite subset of them is jointly Gaussian. GPs are specified by a mean function and a covariance function; for the sake of simplicity, in the remainder of this paper we will employ zero mean GPs. The covariance function $k({\mathbf{x}}, {\mathbf{x}}^{\prime} | {\boldsymbol{\theta}})$ gives the covariance between latent variables at inputs ${\mathbf{x}}$ and ${\mathbf{x}}^{\prime}$ and it is assumed to be parameterized by a set of parameters ${\boldsymbol{\theta}}$. This specification results in a multivariate Gaussian prior over the latent variables $p({\mathbf{f}}| {\boldsymbol{\theta}}) = {\mathcal{N}}({\mathbf{f}}| {\mathbf{0}}, K)$ with $K$ defined as an $n \times n$ matrix with entries $k_{ij} = k({\mathbf{x}}_i, {\mathbf{x}}_j | {\boldsymbol{\theta}})$. A GP can be viewed as a prior over functions and it is appealing in situations where it is difficult to specify a parametric form for the function mapping $X$ into the probabilities of class labels. The covariance plays the role of the kernel in kernel machines, and in the remainder of this paper it will be assumed to be the Radial Basis Function (RBF) covariance $$\label{eq:rbf:covariance} k({\mathbf{x}}_i, {\mathbf{x}}_j | {\boldsymbol{\theta}}) = \sigma \exp\left[-\frac{1}{2} \sum_{r=1}^d \frac{ (x_{ir} - x_{jr})^2}{\tau_r^2} \right].$$ There can be one length-scale parameters $\tau_r$ for each feature, which is a suitable modelling assumption for Automatic Relevance Determination (ARD) [@Mackay94], or there can be one global length-scale parameter $\tau$ such that $\tau_1 = \ldots = \tau_d = \tau$. The parameter $\sigma$ represents the variance of the marginal distribution of each latent variable. A complete specification of a fully Bayesian GP classifier requires a prior $p({\boldsymbol{\theta}})$ over ${\boldsymbol{\theta}}$. When predicting the label $y_*$ for a new input data ${\mathbf{x}}_*$, it is necessary to estimate or infer all unobserved quantities in the model, namely ${\mathbf{f}}$ and ${\boldsymbol{\theta}}$. An appealing way of calculating predictive distributions is as follows: $$\label{eq:predictive} p(y_* | {\mathbf{y}}) = \int p(y_* | f_*) p(f_* | {\mathbf{f}}, {\boldsymbol{\theta}}) p({\mathbf{f}}, {\boldsymbol{\theta}}| {\mathbf{y}}) df_* d{\mathbf{f}}d{\boldsymbol{\theta}}.$$ In the last expression predictions are no longer conditioned on latent variables and covariance parameters, as they are integrated out of the model. Crucially, such an integration accounts for the uncertainty in latent variables and covariance parameters based on their posterior distribution $p({\mathbf{f}}, {\boldsymbol{\theta}}| {\mathbf{y}})$. In order to compute the predictive distribution in eq. \[eq:predictive\], a standard way to proceed is to approximate it using a Monte Carlo estimate: $$\label{eq:monte:carlo:integration} p(y_* | {\mathbf{y}}) \simeq \frac{1}{N} \sum_{i=1}^N \int p(y_* | f_*) p(f_* | {\mathbf{f}}^{(i)}, {\boldsymbol{\theta}}^{(i)}) df_* ,$$ provided that samples from the posterior $p({\mathbf{f}}, {\boldsymbol{\theta}}| {\mathbf{y}})$ are available. Note that in the case of GP classification, the remaining integral has a closed form solution [@Rasmussen06]. As it is not possible to directly draw samples from $p({\mathbf{f}}, {\boldsymbol{\theta}}| {\mathbf{y}})$, alternative ways to characterize it have been proposed. A popular way to do so employs deterministic approximations to integrate out latent variables [@Kuss05; @Nickisch08], but there is no way to quantify the error introduced by these approximation. Also, quadrature is usually employed to integrate out covariance parameters, thus limiting the applicability of GP models to problems with few covariance parameters [@Rue09]. Such limitations might not be acceptable in some pattern recognition applications, so we propose Markov chain Monte Carlo (MCMC) based inference as a general framework for tackling inference problems exactly in GP models. The idea underpinning MCMC methods for GP models is to set up a Markov chain with $p({\mathbf{f}}, {\boldsymbol{\theta}}| {\mathbf{y}})$ as invariant distribution. To date, most MCMC approaches applied to GP models alternate updates of latent variables and covariance parameters. All these approaches, however, face the complexity of having to decouple latent variables and covariance parameters, whose posterior dependence makes convergence to the posterior distribution slow. Reparameterization techniques are a popular way to attempt to decouple the two groups of variables [@Murray10; @Papaspiliopoulos07; @Yu11]. Also, jointly sampling ${\mathbf{f}}$ and ${\boldsymbol{\theta}}$ has been attempted in [@KnorrHeld02; @Rue04], and it is based on approximations to the posterior over latent variables. Despite these efforts, a satisfactory way of sampling the parameters ${\boldsymbol{\theta}}$ for general GP models is still missing, as demonstrated in a recent comparative study [@FilipponeML13]. At this point it is useful to notice that samples from the posterior distribution of latent variables and covariance parameters can be obtained by alternating the sampling from $p({\mathbf{f}}| {\boldsymbol{\theta}}, {\mathbf{y}})$ and $p({\boldsymbol{\theta}}| {\mathbf{y}})$. Obtaining samples from $p({\boldsymbol{\theta}}| {\mathbf{y}})$ is obviously difficult, as it requires the marginal likelihood $p({\mathbf{y}}| {\boldsymbol{\theta}})$; except for the case of a Gaussian likelihood, evaluating the marginal likelihood entails an integration which cannot be computed analytically [@Rasmussen06]. In the next section we will focus on the PM MCMC approach as a practical way of dealing with this problem. Obtaining samples from $p({\mathbf{f}}| {\mathbf{y}}, {\boldsymbol{\theta}})$, instead, can be done efficiently using Elliptical Slice Sampling (Ell-SS) [@Murray10b]. Ell-SS defines a transition operator $T({\mathbf{f}}^{\prime} | {\mathbf{f}})$, and is a variant of Slice Sampling [@Neal03] adapted to the sampling of latent variables in GP models. Ell-SS begins by randomly choosing a threshold $\eta$ for $\log[p({\mathbf{y}}| {\mathbf{f}})]$ $$u \sim U[0,1] \qquad \eta = \log[p({\mathbf{y}}| {\mathbf{f}})] + \log[u]$$ and by drawing a set of latent variables ${\mathbf{z}}$ from the prior ${\mathcal{N}}(\mathbf{0}, K)$. Then, a combination of ${\mathbf{f}}$ and ${\mathbf{z}}$ is sought, such that the log-likelihood of the resulting combination is larger than the threshold $\eta$. Such a combination is defined by means of sine and cosine of an auxiliary variable $\alpha$, which makes the resulting combination spanning a domain of points that is an ellipse in the latent variable space. The search procedure is based on slice sampling on $\alpha$ starting from the interval $[0, 2\pi]$. Due to the fact that Ell-SS does not require any tuning and it has been shown to be very efficient for several GP models [@FilipponeML13], it is the operator that will be used in the remainder of this paper to sample latent variables. However, note that latent variables can be also efficiently sampled by means of a variant of Hybrid Monte Carlo [@FilipponeML13]. Pseudo-Marginal Inference for GP models {#sec:ea:gp} ======================================= For the sake of simplicity, this work will focus on the Metropolis-Hastings (MH) algorithm [@Neal93; @Hastings70] to obtain samples from the posterior distribution over covariance parameters. The MH algorithm is based on the iteration of the following two steps: (i) proposing a new set of parameters ${\boldsymbol{\theta}}^{\prime}$ drawing from a user defined proposal distribution $\pi({\boldsymbol{\theta}}^{\prime} | {\boldsymbol{\theta}})$ and (ii) evaluating the Hastings ratio $$\tilde{z} = \frac{p({\mathbf{y}}| {\boldsymbol{\theta}}^{\prime}) p({\boldsymbol{\theta}}^{\prime})}{p({\mathbf{y}}| {\boldsymbol{\theta}}) p({\boldsymbol{\theta}})} \frac{\pi({\boldsymbol{\theta}}| {\boldsymbol{\theta}}^{\prime})}{\pi({\boldsymbol{\theta}}^{\prime} | {\boldsymbol{\theta}})}$$ to accept or reject ${\boldsymbol{\theta}}^{\prime}$. As previously discussed, the marginal likelihood $p({\mathbf{y}}| {\boldsymbol{\theta}}) = \int p({\mathbf{y}}| {\mathbf{f}}) p({\mathbf{f}}| {\boldsymbol{\theta}}) d{\mathbf{f}}$ cannot be computed analytically, except for the case of a Gaussian likelihood. The PM approach in [@FilipponeIEEETPAMI14] builds upon a remarkable theoretical result [@Beaumont03; @Andrieu09] stating that it is possible to plug an unbiased estimate of the marginal likelihood $\tilde{p}({\mathbf{y}}| {\boldsymbol{\theta}})$ in the Hastings ratio $$\tilde{z} = \frac{\tilde{p}({\mathbf{y}}| {\boldsymbol{\theta}}^{\prime}) p({\boldsymbol{\theta}}^{\prime})}{\tilde{p}({\mathbf{y}}| {\boldsymbol{\theta}}) p({\boldsymbol{\theta}})} \frac{\pi({\boldsymbol{\theta}}| {\boldsymbol{\theta}}^{\prime})}{\pi({\boldsymbol{\theta}}^{\prime} | {\boldsymbol{\theta}})}$$ and still obtain an MCMC algorithm sampling from the correct posterior distribution $p({\boldsymbol{\theta}}| {\mathbf{y}})$. In [@FilipponeIEEETPAMI14] an unbiased estimate of the marginal likelihood was obtained as follows. First, an approximation of the posterior over latent variables $p({\mathbf{f}}| {\mathbf{y}}, {\boldsymbol{\theta}})$, say $q({\mathbf{f}}| {\boldsymbol{\theta}}, {\mathbf{y}})$, was obtained by means of approximate methods, such as for example the Laplace Approximation (LA) or Expectation Propagation. Second, based on $q({\mathbf{f}}| {\boldsymbol{\theta}}, {\mathbf{y}})$, it was proposed to get an unbiased estimate of the marginal likelihood $p({\mathbf{y}}| {\boldsymbol{\theta}})$ using IS. In particular, this was achieved by drawing $N_{\mathrm{imp}}$ samples ${\mathbf{f}}^{(i)}$ from the approximating distribution $q({\mathbf{f}}| {\boldsymbol{\theta}}, {\mathbf{y}})$. Defining $$\label{eq:importance:weights} w_{\mathrm{IS}}^{(i)} = \frac{p({\mathbf{y}}| {\mathbf{f}}^{(i)}) p({\mathbf{f}}^{(i)} | {\boldsymbol{\theta}})}{q({\mathbf{f}}^{(i)} | {\boldsymbol{\theta}}, {\mathbf{y}})},$$ the marginal likelihood $p({\mathbf{y}}| {\boldsymbol{\theta}})$ was approximated by $$\label{eq:importance:pseudo} \tilde{p}({\mathbf{y}}| {\boldsymbol{\theta}}) \simeq \frac{1}{N_{\mathrm{imp}}} \sum_{i=1}^{N_{\mathrm{imp}}} w_{\mathrm{IS}}^{(i)}.$$ Such an estimate is unbiased and the closer $q({\mathbf{f}}| {\boldsymbol{\theta}}, {\mathbf{y}})$ is to $p({\mathbf{y}}| {\mathbf{f}}) p({\mathbf{f}}| {\boldsymbol{\theta}})$ the lower the variance of the estimate [@Neal01]. In the experiments shown in [@FilipponeIEEETPAMI14] this estimate was adequate for the problems that were analyzed, especially when accurate approximations based on Expectation Propagation were used. However, the variance of the IS estimate grows exponentially with the dimensionality of the integral [@Neal01], and this might represent a limitation when applying PM MCMC to large data sets. In particular, a large variance in the estimate of $p({\mathbf{y}}| {\boldsymbol{\theta}})$ can eventually lead to the acceptance of a ${\boldsymbol{\theta}}$ because the corresponding marginal likelihood is overestimated. If the overestimation is severe, it is unlikely that any new proposal will be accepted, resulting in slow convergence and low efficiency. The aim of this paper is to present a methodology based on AIS [@Neal01] which is capable of mitigating this effect. Marginal Likelihood estimation with Annealed Importance Sampling {#sec:ais} ================================================================ \ {width="1\linewidth"}\ \ {width="1\linewidth"} AIS is an extension of IS where the weights in eq. \[eq:importance:weights\] are computed based on a sequence of distributions going from one that is easy to sample from to the posterior distribution of interest. Following the derivation in [@Neal01], define $g_s({\mathbf{f}})$ as the unnormalized density of a distribution which is easy to sample from; in the next section we will study two of such distributions. Also, define $$g_0({\mathbf{f}}) = p({\mathbf{y}}| {\mathbf{f}}) p({\mathbf{f}}| {\boldsymbol{\theta}}) \propto p({\mathbf{f}}| {\boldsymbol{\theta}}, {\mathbf{y}}).$$ AIS defines a sequence of intermediate unnormalized distributions $$g_j({\mathbf{f}}) = g_0({\mathbf{f}})^{\beta_j} g_s({\mathbf{f}})^{1 - \beta_j}$$ with $1 = \beta_0 > \ldots > \beta_s = 0$. The AIS sampling procedure begins by drawing one sample ${\mathbf{f}}_{s-1}$ from $g_s({\mathbf{f}})$. After that, for $i = s-1, \ldots, 1$, a new ${\mathbf{f}}_{i-1}$ is obtained from ${\mathbf{f}}_{i}$ by iterating a transition operator $T_{i}({\mathbf{f}}^{\prime} | {\mathbf{f}})$ that leaves the normalized version of $g_i({\mathbf{f}})$ invariant. Finally, computing the average of the following weights $$w_{\mathrm{AIS}}^{(i)} = \frac{g_{s-1}({\mathbf{f}}_{s-1})}{g_{s}({\mathbf{f}}_{s-1})} \frac{g_{s-2}({\mathbf{f}}_{s-2})}{g_{s-1}({\mathbf{f}}_{s-2})} \cdots \frac{g_{1}({\mathbf{f}}_{1})}{g_{2}({\mathbf{f}}_{1})} \frac{g_{0}({\mathbf{f}}_{0})}{g_{1}({\mathbf{f}}_{0})}$$ yields an unbiased estimate of the ratio of the normalizing constants of $g_0({\mathbf{f}})$ and $g_s({\mathbf{f}})$, which immediately yields an unbiased estimate of $p({\mathbf{y}}| {\boldsymbol{\theta}})$. For numerical reasons, it is safe to implement the calculations using logarithm transformations. Also, note that although the annealing strategy is inherently serial, the computations with respect to multiple importance samples can be parallelized. We now analyze two ways of implementing AIS for GP models, which are visually illustrated in fig. \[fig:explain:annealing\]. Annealing from the prior ------------------------ When annealing from the prior, the intermediate distributions are between $g_s({\mathbf{f}}) = {\mathcal{N}}({\mathbf{f}}| 0, K)$ and $g_0({\mathbf{f}}) = {\mathcal{N}}({\mathbf{f}}| 0, K) p({\mathbf{y}}| {\mathbf{f}})$, namely $$g_j({\mathbf{f}}) = {\mathcal{N}}({\mathbf{f}}| 0, K) \left[p({\mathbf{y}}| {\mathbf{f}})\right]^{\beta_j}.$$ Employing Ell-SS as a transition operator for ${\mathbf{f}}$ for the intermediate unnormalized distributions $g_j({\mathbf{f}})$ is straightforward, as the log-likelihood is simply scaled by $\beta_j$. Annealing from the prior was proposed in [@Kuss05] where it was reported that a sequence of $8000$ annealed distributions was employed. This is because the prior and the posterior look very much different (see fig. \[fig:explain:annealing\]) and the only way to ensure a smooth transition from the prior to the posterior is by using several intermediate distributions. This is problematic from a computational perspective, as the calculation of the marginal likelihood has to be done at each iteration of the PM approach to sample from the posterior distribution over ${\boldsymbol{\theta}}$. We therefore propose an alternative starting distribution $g_s({\mathbf{f}})$ that leads to a reduction in the number of intermediate distributions while obtaining estimates of the marginal likelihood that are accurate enough to ensure good sampling efficiency when used in the PM MCMC approach. Annealing from an approximating distribution -------------------------------------------- Several Gaussian-based approximation schemes to integrate out latent variables have been proposed for GP models [@Minka01; @Opper00]. When an approximation to the posterior over latent variables is available, it might be reasonable to construct the sequence of intermediate distributions in AIS starting from it rather than the prior. When annealing from an approximating Gaussian distribution, the intermediate distributions are between $g_s({\mathbf{f}}) = q({\mathbf{f}}| {\boldsymbol{\theta}}, {\mathbf{y}}) = {\mathcal{N}}({\mathbf{f}}| {\boldsymbol{\mu}}, \Sigma)$ and $g_0({\mathbf{f}}) = {\mathcal{N}}({\mathbf{f}}| 0, K) p({\mathbf{y}}| {\mathbf{f}})$. In order to employ Ell-SS as a transition operator $T_i({\mathbf{f}}^{\prime} | {\mathbf{f}})$, it is useful to write the unnormalized intermediate distributions as $$g_j({\mathbf{f}}) = {\mathcal{N}}({\mathbf{f}}| {\boldsymbol{\mu}}, \Sigma) \left[\frac{{\mathcal{N}}({\mathbf{f}}| 0, K) p({\mathbf{y}}| {\mathbf{f}})}{{\mathcal{N}}({\mathbf{f}}| {\boldsymbol{\mu}}, \Sigma)}\right]^{\beta_j}.$$ In this way, the model can be interpreted as having a prior ${\mathcal{N}}({\mathbf{f}}| {\boldsymbol{\mu}}, \Sigma)$ and a likelihood given by the term in square brackets; applying Ell-SS to this formulation is straightforward. Experimental results {#sec:results} ==================== The first part of this section, compares the behavior of IS and AIS in the case of synthetic data. The second part of this section, reports an analysis of IS and AIS when employed in the PM MCMC approach applied to real data. In all experiments, the approximation was based on the Laplace Approximation (LA) algorithm. Also, we imposed Gamma priors on the parameters $\mathrm{Ga}(\sigma | a=1.1, b=0.1)$ and $\mathrm{Ga}(\tau_i | a=1, b=1)$ for the ARD covariance and $\mathrm{Ga}(\tau | a=1, b=1/\sqrt{d})$ for the isotropic covariance, where $a$ and $b$ are shape and rate parameters respectively. Following the recommendations in [@Neal01; @Neal96b], $s = \sqrt{n}$ intermediate distributions were defined based on a geometric spacing of the $\beta$’s. In particular, this was implemented by setting $s/2-1$ uniformly spaced values of $\log[\beta]$ between $\log[1]$ and $\log[0.2]$, $s/2$ uniformly spaced values between $\log[0.2]$ and $\log\left[10^{-6}\right]$, and finally $\beta_s = 0$. In AIS, the transitions $T_{i}({\mathbf{f}}^{\prime} | {\mathbf{f}})$ involved one iteration of Ell-SS. Synthetic data -------------- The aim of this section is to highlight the potential inefficiency in employing IS to obtain an unbiased estimate of the marginal likelihood and to demonstrate the effectiveness of AIS in dealing with this problem. In particular, this can be problematic in large dimensions, namely when analyzing large amounts of data. In order to show this effect, we generated data sets with an increasing number of data $n=10, 50, 100, 500, 1000$ in two dimensions with a balanced class distribution. Data were generated drawing input vectors uniformly in the unit square and a latent function from a GP with covariance in eq. \[eq:rbf:covariance\] with $\sigma=20$ and a global $\tau=0.255$. This combination of covariance parameters leads to a strongly non-Gaussian posterior distribution over the latent variables making IS perform poorly when $n$ is large. In order to obtain a measure of variability of the IS and AIS estimators of the marginal likelihood, we analyze the standard deviation of the estimator of $\log[p({\mathbf{y}}| {\boldsymbol{\theta}})]$ $$\label{eq:r:score} r = \mathrm{st\ dev} \left\{ \log_{10} \left[ \tilde{p}({\mathbf{y}}| {\boldsymbol{\theta}}) \right] \right\}.$$ In the experiments, $r$ was estimated based on $50$ repetitions; fig. \[fig:res:synth\] shows the distribution of $r$ based on $50$ draws of ${\boldsymbol{\theta}}$ from the posterior $p({\boldsymbol{\theta}}| {\mathbf{y}})$ obtained from a preliminary run of an MCMC algorithm. Ideally, a perfect estimator of the marginal likelihood would yield a degenerate distribution of $r$ over posterior samples of ${\boldsymbol{\theta}}$ at zero. In practice, the distribution of $r$ indicates the variability (across posterior samples of ${\boldsymbol{\theta}}$) around an average value of the standard deviation of the estimator of the logarithm of the marginal likelihood. The representation in $\log_{10}$ is helpful to get an idea of the order of magnitude of such a variability. For instance, a distribution of $r$ across posterior samples of ${\boldsymbol{\theta}}$ concentrated around $2$ would mean that, on average, the estimates of the marginal likelihood span roughly two orders of magnitude. ![This figure shows a measure of the quality of the IS and AIS (annealing from the prior and from an approximating distribution obtained by the LA algorithm) estimators of the marginal likelihood. The boxplot summarizes the distribution of $r$ in eq. \[eq:r:score\] for $50$ values of ${\boldsymbol{\theta}}$ drawn from $p({\boldsymbol{\theta}}| {\mathbf{y}})$. []{data-label="fig:res:synth"}](figs/PLOT_COMPARE_AIS_IS.eps "fig:"){width="0.98\linewidth"}\ Fig. \[fig:res:synth\] shows the distribution of $r$ for AIS when annealing from the prior and from an approximating distribution, along with the distribution of $r$ for IS as in [@FilipponeIEEETPAMI14]. In all methods we set $N_{\mathrm{imp}} = 4$. The results confirm that annealing from the prior offers much poorer estimates of the marginal likelihood compared to annealing from an approximating distribution and will not be considered further. The analysis of the results in fig. \[fig:res:synth\] reveal that when annealing from an approximating distribution, the reduction in variance of the estimate of the marginal likelihood compared to IS is exponential in $n$. When comparing the computational cost of running IS and AIS, instead, we notice that AIS increases it by a factor which scales only polynomially with $n$. This is because, after approximating the posterior over ${\mathbf{f}}$ (that typically costs $O(n^3)$ operations), in AIS drawing the initial importance samples, iterating Ell-SS, and computing the weights $w_{\mathrm{AIS}}$ costs $O(n^2)$ operations; this needs to be done as many times as the number of intermediate distributions $s$, which in our case means $O(\sqrt{n})$ times. In IS, drawing the importance samples and computing the weights $w_{\mathrm{IS}}$ requires $O(n^2)$ operations. Real data --------- [|c|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|]{}\ & & & & &\ & & & & &\ $N_{\mathrm{imp}}$ & IS & AIS & IS & AIS & IS & AIS & IS & AIS & IS & AIS\ 1 & $2.8 (1.6)$ & $5.2 (1.9)$ & $1.1 (1.0)$ & $3.2 (2.3)$ & $17.9 (2.4)$ & $28.0 (2.7)$ & $24.8 (1.4)$ & $29.3 (2.6)$ & $1.1 (0.6)$ & $3.2 (3.9)$\ 10 & $10.4 (3.1)$ & $11.4 (5.3)$ & $4.1 (3.8)$ & $6.4 (3.9)$ & $30.5 (4.1)$ & $36.4 (3.5)$ & $30.8 (2.6)$ & $30.8 (1.7)$ & $4.7 (1.0)$ & $9.2 (5.6)$\ \ [|c|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|C[1.25cm]{}|]{}\ & & & & &\ & & & & &\ $N_{\mathrm{imp}}$ & IS & AIS & IS & AIS & IS & AIS & IS & AIS & IS & AIS\ 1 & $1.3 (1.3)$ & $3.6 (2.3)$ & $0.4 (0.3)$ & $2.9 (1.8)$ & $1.8 (1.7)$ & $5.0 (2.5)$ & $17.1 (2.7)$ & $22.5 (3.3)$ & $1.3 (1.4)$ & $4.7 (2.1)$\ 10 & $2.5 (1.6)$ & $4.9 (3.2)$ & $6.9 (2.4)$ & $6.4 (2.0)$ & $7.7 (2.6)$ & $4.5 (1.8)$ & $22.8 (4.0)$ & $24.1 (3.9)$ & $5.8 (3.3)$ & $9.2 (3.1)$\ This section reports an analysis of the PM MCMC approach applied to five UCI data sets [@Asuncion07] when the marginal likelihood is estimated using AIS and IS. The Glass data set is multi-class, and we turned it into a two class data set by considering the data labelled as “window glass” as one class and data labelled as “non-window glass” as the other class. In all data sets, features were normalized to have zero mean and unit standard deviation. All experiments were repeated varying the number of importance samples $N_{\mathrm{imp}} = 1, 10$, and employing isotropic and ARD RBF covariance functions as in eq. \[eq:rbf:covariance\]. In order to tune the MH proposal, we ran a preliminary MCMC algorithm for $2000$ iterations. This was initialized from the prior and the marginal likelihood in the Hastings ratio was obtained by the LA algorithm. The proposal was then adapted to obtain an acceptance rate between $20\%$ and $30\%$. This set up was useful in order to avoid problems in tuning the proposal mechanism when a noisy version of the marginal likelihood is used, which may lead to a poor acceptance rate independently of the proposal mechanism. Tab. \[tab:res:real\] reports the average acceptance rate when switching to an unbiased version of the marginal likelihood obtained by IS or AIS for different values of $N_{\mathrm{imp}}$ after the adaptive phase. The average acceptance rate was computed based on $1500$ iterations, collected after discarding $500$ iterations, and over $5$ parallel chains. The results are variable across data sets and the type of covariance, but the general trend is that employing AIS in the PM MCMC approach improves on the acceptance rate compared to IS. In a few cases, it is striking to see how replacing an approximate marginal likelihood with an unbiased estimate in the Hastings ratio does not affect the acceptance rate, thus confirming the merits of the PM MCMC approach. In general, however, PM MCMC is affected by the use of an estimate of the marginal likelihood. In cases where this happens, AIS consistently offers a way to reduce the variance of the estimate of the marginal likelihood compared to IS, and this improves on the acceptance rate. Conclusions {#sec:conclusions} =========== This paper presented the application of annealed importance sampling to obtain an unbiased estimate of the marginal likelihood in GP classifiers. Annealed importance sampling for GP classifiers was previously proposed in [@Kuss05] where the sequence of distributions was constructed from the prior to the posterior over latent variables. Given the difference between these two distributions, the annealing strategy requires the use of several intermediate distributions, thus making this methodology impractical. This paper studied the possibility to construct a sequence of distributions from an approximating distribution rather than the prior, and empirically demonstrated that, compared to importance sampling, this reduces the variance of the estimator of the marginal likelihood exponentially in the number of data. Crucially, this reduction comes at a cost that is only polynomial in the number of data. Also, annealed importance sampling can be easily parallelized. The motivation for studying this problem was to plug the unbiased estimate of the marginal likelihood in the Hastings ratio in order to obtain an MCMC approach sampling from the correct posterior distribution over covariance parameters. The results on real data show that employing importance sampling within the pseudo-marginal MCMC approach can be satisfactory in many cases. However, in general, annealed importance sampling leads to a lower variance estimator of the marginal likelihood, and the resulting pseudo-marginal MCMC approach significantly improves on the average acceptance rate. These results suggest a promising direction of research towards the development of MCMC methods where the likelihood is estimated in an unbiased fashion, but the acceptance rate is as if the likelihood were known exactly. Given that the computational overhead scales with less than the third power of the number of data, the results indicate that this can be achieved with an acceptable computational cost. This paper considered GP classification as a working example, and the Laplace approximation algorithm to obtain the importance distribution. A matter of current investigation is the application of the proposed methodology to other GP models and other approximation schemes. Furthermore, this paper focused on the case of full covariance matrices. These results can be extended to deal with sparse inverse covariance matrices, which are popular when modeling spatio-temporal data, thus leading to the possibility to process massive amounts of data due to the use of sparse algebra routines. Finally, this paper did not attempt to optimize the annealing scheme, but it would be sensible to do so in order to minimize the variance of the annealed importance sampling estimator of the marginal likelihood [@Behrens10]. [10]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{} J. Shawe-Taylor and N. Cristianini, *[Kernel Methods for Pattern Analysis]{}*.1em plus 0.5em minus 0.4emNew York, NY, USA: Cambridge University Press, 2004. M. Filippone, A. F. Marquand, C. R. V. Blain, S. C. R. Williams, J. [Mourão-Miranda]{}, and M. Girolami, “[Probabilistic Prediction of Neurological Disorders with a Statistical Assessment of Neuroimaging Data Modalities]{},” *Annals of Applied Statistics*, vol. 6, no. 4, pp. 1883–1905, 2012. <http://dx.doi.org/10.1214/12-AOAS562> M. Filippone and M. Girolami, “Pseudo-marginal [B]{}ayesian inference for [G]{}aussian processes,” *IEEE Transactions on Pattern Analysis and Machine Intelligence*, 2014. <http://dx.doi.org/10.1109/TPAMI.2014.2316530> R. M. Neal, “[Regression and classification using Gaussian process priors (with discussion)]{},” *Bayesian Statistics*, vol. 6, pp. 475–501, 1999. H. Rue, S. Martino, and N. Chopin, “[Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations]{},” *Journal of the Royal Statistical Society: Series B (Statistical Methodology)*, vol. 71, no. 2, pp. 319–392, 2009. <http://dx.doi.org/10.1111/j.1467-9868.2008.00700.x> M. B. Taylor and J. P. Diggle, “[INLA or MCMC? A Tutorial and Comparative Evaluation for Spatial Prediction in log-Gaussian Cox Processes]{},” 2012, arXiv:1202.1738. J. M. Flegal, M. Haran, and G. L. Jones, “[Markov Chain Monte Carlo: Can We Trust the Third Significant Figure?]{}” *Statistical Science*, vol. 23, no. 2, pp. 250–260, 2007. <http://arxiv.org/abs/math.ST/0703746> M. Filippone, M. Zhong, and M. Girolami, “A comparative evaluation of stochastic-based inference methods for [G]{}aussian process models,” *Machine Learning*, vol. 93, no. 1, pp. 93–114, 2013. <http://dx.doi.org/10.1007/s10994-013-5388-x> R. M. Neal, “Probabilistic inference using [M]{}arkov chain [M]{}onte [C]{}arlo methods,” Dept. of Computer Science, University of Toronto, Tech. Rep. CRG-TR-93-1, 1993. C. E. Rasmussen and C. Williams, *[Gaussian Processes for Machine Learning]{}*.1em plus 0.5em minus 0.4emMIT Press, 2006. <http://www.gaussianprocess.org/gpml/> M. Kuss and C. E. Rasmussen, “[Assessing Approximate Inference for Binary Gaussian Process Classification]{},” *Journal of Machine Learning Research*, vol. 6, pp. 1679–1704, 2005. H. Nickisch and C. E. Rasmussen, “[Approximations for Binary Gaussian Process Classification]{},” *Journal of Machine Learning Research*, vol. 9, pp. 2035–2078, 2008. R. M. Neal, “[Annealed importance sampling]{},” *Statistics and Computing*, vol. 11, no. 2, pp. 125–139, 2001. <http://dx.doi.org/10.1023/A:1008923215028> D. J. C. Mac[k]{}ay, “[B]{}ayesian methods for backpropagation networks,” in *Models of Neural Networks [III]{}*, E. Domany, J. L. van Hemmen, and K. Schulten, Eds.1em plus 0.5em minus 0.4emSpringer, 1994, ch. 6, pp. 211–254. I. Murray and R. P. Adams, “[Slice sampling covariance hyperparameters of latent Gaussian models]{},” in *NIPS*, J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta, Eds.1em plus 0.5em minus 0.4emCurran Associates, 2010, pp. 1732–1740. O. Papaspiliopoulos, G. O. Roberts, and M. Sköld, “[A general framework for the parametrization of hierarchical models]{},” *Statistical Science*, vol. 22, no. 1, pp. 59–73, 2007. <http://www.jstor.org/stable/27645805> Y. Yu and X.-L. Meng, “[To Center or Not to Center: That Is Not the Question–An Ancillarity-Sufficiency Interweaving Strategy (ASIS) for Boosting MCMC Efficiency]{},” *Journal of Computational and Graphical Statistics*, vol. 20, no. 3, pp. 531–570, 2011. <http://amstat.tandfonline.com/doi/abs/10.1198/jcgs.2011.203main> L. Knorr-Held and H. Rue, “[On Block Updating in Markov Random Field Models for Disease Mapping]{},” *Scandinavian Journal of Statistics*, vol. 29, no. 4, pp. 597–614, 2002. <http://www.ingentaconnect.com/content/bpl/sjos/2002/00000029/00000004/art00308> H. Rue, I. Steinsland, and S. Erland, “[Approximating hidden Gaussian Markov random fields]{},” *Journal of the Royal Statistical Society: Series B (Statistical Methodology)*, vol. 66, no. 4, pp. 877–892, 2004. <http://dx.doi.org/10.1111/j.1467-9868.2004.B5590.x> I. Murray, R. P. Adams, and D. J. C. MacKay, “[Elliptical slice sampling]{},” *Journal of Machine Learning Research - Proceedings Track*, vol. 9, pp. 541–548, 2010. R. M. Neal, “[Slice Sampling]{},” *Annals of Statistics*, vol. 31, pp. 705–767, 2003. W. K. Hastings, “Monte [C]{}arlo sampling methods using [M]{}arkov chains and their applications,” *Biometrika*, vol. 57, no. 1, pp. 97–109, 1970. <http://dx.doi.org/10.1093/biomet/57.1.97> M. A. Beaumont, “[Estimation of Population Growth or Decline in Genetically Monitored Populations]{},” *Genetics*, vol. 164, no. 3, pp. 1139–1160, 2003. <http://www.genetics.org/content/164/3/1139.abstract> C. Andrieu and G. O. Roberts, “[The pseudo-marginal approach for efficient Monte Carlo computations]{},” *The Annals of Statistics*, vol. 37, no. 2, pp. 697–725, 2009. <http://dx.doi.org/10.1214/07-aos574> T. P. Minka, “[Expectation Propagation for approximate Bayesian inference]{},” in *Proceedings of the 17th Conference in Uncertainty in Artificial Intelligence*, ser. UAI ’01.1em plus 0.5em minus 0.4emSan Francisco, CA, USA: Morgan Kaufmann Publishers Inc., 2001, pp. 362–369. <http://dl.acm.org/citation.cfm?id=647235.720257> M. Opper and O. Winther, “[Gaussian processes for classification: Mean-field algorithms]{},” *Neural Computation*, vol. 12, no. 11, pp. 2655–2684, 2000. R. M. Neal, “[Sampling from multimodal distributions using tempered transitions]{},” *Statistics and Computing*, vol. 6, pp. 353–366, 1996. <http://www.springerlink.com.cyber.usask.ca/content/x6417mm4h7845735/> A. Asuncion and D. J. Newman, “[UCI]{} machine learning repository,” 2007. [http://www.ics.uci.edu/\$\\sim\$mlearn/{MLR}epository.html](http://www.ics.uci.edu/$\sim$mlearn/{MLR}epository.html) G. Behrens, N. Friel, and M. Hurn, “[Tuning Tempered Transitions]{},” *Statistics and Computing*, vol. 22, no. 1, pp. 65–78, 2010. <http://dx.doi.org/10.1007/s11222-010-9206-z> [^1]: The code to reproduce all the results in this paper can be found here:\ www.dcs.gla.ac.uk/$\sim$maurizio/pages/code\_ea\_mcmc\_ais/

ArXiv

--- abstract: 'Let $(M,g)$ be compact Riemannian manifold of dimension $n \geq 3$. A well-known conjecture states that the set of constant scalar curvature metrics in the conformal class of $g$ is compact unless $(M,g)$ is conformally equivalent to the round sphere. In this paper, we construct counterexamples to this conjecture in dimensions $n \geq 52$.' author: - Simon Brendle date: 'May 5, 2007' title: 'Blow-up phenomena for the Yamabe equation' --- Introduction ============ Let $(M,g)$ be a compact Riemannian manifold of dimension $n \geq 3$. The Yamabe problem is concerned with finding metrics of constant scalar curvature in the conformal class of $g$. This problem can be reduced to a semi-linear elliptic PDE. Indeed, the metric $u^{\frac{4}{n-2}} \, g$ has constant scalar curvature $c$ if and only if $$\label{yamabe.pde} \frac{4(n-1)}{n-2} \, \Delta_g u - R_g \, u + c \, u^{\frac{n+2}{n-2}} = 0,$$ where $\Delta_g$ is the Laplace operator with respect to $g$ and $R_g$ denotes the scalar curvature of $g$. Clearly, every solution of (\[yamabe.pde\]) is a critical point of the functional $$\label{yamabe.functional} E_g(u) = \frac{\int_M \big ( \frac{4(n-1)}{n-2} \, |du|_g^2 + R_g \, u^2 \big ) \, dvol_g}{\big ( \int_M u^{\frac{2n}{n-2}} \, dvol_g \big )^{\frac{n-2}{n}}}.$$ It is well-known that the PDE (\[yamabe.pde\]) has at least one positive solution for any choice of $(M,g)$. If $n \geq 6$ and $(M,g)$ is not locally conformally flat, this follows from results of T. Aubin [@Aubin1]. The remaining cases were solved by R. Schoen using the positive mass theorem [@Schoen1]. Solutions to (\[yamabe.pde\]) are not usually unique. As an example, consider the product metric on $S^1(L) \times S^{n-1}(1)$. If $L$ is sufficiently small, then the Yamabe PDE has a unique solution. On the other hand, there are many non-minimizing solutions if $L$ is large. D. Pollack [@Pollack] has used gluing techniques to construct high energy solutions on more general background manifolds: given any conformal class with positive Yamabe constant and any positive integer $N$, there exists a new conformal class which is close to the original one in the $C^0$-norm, and contains at least $N$ metrics of constant scalar curvature (see [@Pollack], Theorem 0.1). It is an interesting question whether the set of all solutions to the Yamabe PDE is compact (in the $C^2$-topology, say). A well-known conjecture states that this should be true unless $(M,g)$ is conformally equivalent to the round sphere (see [@Schoen2],[@Schoen3],[@Schoen4]). This conjecture has been verified in low dimensions and in the locally conformally flat case: if $(M,g)$ is locally conformally flat, compactness follows from work of R. Schoen [@Schoen2],[@Schoen3]. Moreover, Schoen proposed a strategy for proving the conjecture in the non-locally conformally flat case based on the Pohozaev identity. In [@Li-Zhu], Y.Y. Li and M. Zhu [@Li-Zhu] followed this strategy to prove compactness in dimension $3$. O. Druet [@Druet] proved the conjecture in dimensions $4$ and $5$. Recently, F. Marques [@Marques] showed that compactness holds up to dimension $7$. The same result was obtained independently by Y.Y. Li and L. Zhang [@Li-Zhang]. Moreover, Li and Zhang showed that compactness holds in all dimensions provided that $|W_g(p)| + |\nabla W_g(p)| > 0$ for all $p \in M$. M. Khuri, F. Marques, and R. Schoen [@Khuri-Marques-Schoen] proved compactness up to dimension $24$, assuming that the positive mass theorem holds.[^1] In this paper, we address the opposite question: is it possible to construct Riemannian manifolds $(M,g)$ such that the set of constant scalar curvature metrics in the conformal class of $g$ is non-compact? So far, the only known examples where compactness fails involve non-smooth background metrics. The first result in this direction was established by A. Ambrosetti and A. Malchiodi [@Ambrosetti-Malchiodi]. This result was subsequently improved by M. Berti and A. Malchiodi [@Berti-Malchiodi]. Given positive integers $n$ and $k$ such that $k \geq 2$ and $n \geq 4k+3$, Berti and Malchiodi showed that there exists a Riemannian metric $g$ on $S^n$ (of class $C^k$) for which the set of solutions to the Yamabe PDE (\[yamabe.pde\]) fails to be compact (see [@Berti-Malchiodi], Theorem 1.1). A survey of these results can be found in [@Ambrosetti]. Recently, O. Druet and E. Hebey [@Druet-Hebey1] showed that blow-up can occur for problems of the form $Lu + c \, u^{\frac{n+2}{n-2}} = 0$, where $L$ is a lower order perturbation of the conformal Laplacian on $S^n$. We improve the results of Berti and Malchiodi by showing that the set of solutions to the Yamabe PDE (\[yamabe.pde\]) can fail to be compact even if the background metric $g$ is $C^\infty$ smooth. In the examples we construct, the blowing-up sequence develops a singularity consisting of exactly one bubble. \[main.theorem\] Assume that $n \geq 52$. Then there exists a Riemannian metric $g$ on $S^n$ (of class $C^\infty$) and a sequence of positive functions $v_\nu \in C^\infty(S^n)$ ($\nu \in \mathbb{N}$) with the following properties: - $g$ is not conformally flat - $v_\nu$ is a solution of the Yamabe PDE (\[yamabe.pde\]) for all $\nu \in \mathbb{N}$ - $E_g(v_\nu) < Y(S^n)$ for all $\nu \in \mathbb{N}$, and $E_g(v_\nu) \to Y(S^n)$ as $\nu \to \infty$ - $\sup_{S^n} v_\nu \to \infty$ as $\nu \to \infty$ (Here, $Y(S^n)$ denotes the Yamabe energy of the round metric on $S^n$.) Let us sketch the main steps involved in the proof of Theorem \[main.theorem\]. For convenience, we will work on $\mathbb{R}^n$ instead of $S^n$. Let $g$ be a smooth metric on $\mathbb{R}^n$ which agrees with the Euclidean metric outside a ball of radius $1$. We will assume throughout the paper that $\det g(x) = 1$ for all $x \in \mathbb{R}^n$, so that the volume form associated with $g$ agrees with the Euclidean volume form. Our goal is to construct solutions to the Yamabe PDE on $(\mathbb{R}^n,g)$. In Section 2, we show that this problem can be reduced to finding critical points of a certain function $\mathcal{F}_g(\xi,\varepsilon)$, where $\xi$ is a vector in $\mathbb{R}^n$ and $\varepsilon$ is a positive real number. This idea has been used by many authors (see, e.g., [@Ambrosetti-Malchiodi] or [@Berti-Malchiodi]). In Section 3, we show that the function $\mathcal{F}_g(\xi,\varepsilon)$ can be approximated by an auxiliary function $F(\xi,\varepsilon)$. In Section 4, we prove that the function $F(\xi,\varepsilon)$ has a critical point, which is a strict local minimum. Finally, in Section 5, we use a perturbation argument to construct critical points of the function $\mathcal{F}_g(\xi,\varepsilon)$. From this the main result follows. It is a pleasure to thank Professor Richard Schoen and Professor Fernando Marques for numerous discussions on this subject. This project was supported by the Alfred P. Sloan Foundation and by the National Science Foundation under grant DMS-0605223.\ Lyapunov-Schmidt reduction ========================== Let $$\mathcal{E} = \bigg \{ w \in L^{\frac{2n}{n-2}}(\mathbb{R}^n) \cap W_{loc}^{1,2}(\mathbb{R}^n): \int_{\mathbb{R}^n} |dw|^2 < \infty \bigg \}.$$ By Sobolev’s inequality, there exists a constant $K$, depending only on $n$, such that $$\bigg ( \int_{\mathbb{R}^n} |w|^{\frac{2n}{n-2}} \bigg )^{\frac{n-2}{n}} \leq K \, \int_{\mathbb{R}^n} |dw|^2$$ for all $w \in \mathcal{E}$. We define a norm on $\mathcal{E}$ by $\|w\|_{\mathcal{E}}^2 = \int_{\mathbb{R}^n} |dw|^2$. It is easy to see that $\mathcal{E}$, equipped with this norm, is complete.\ Given any pair $(\xi,\varepsilon) \in \mathbb{R}^n \times (0,\infty)$, we define a function $u_{(\xi,\varepsilon)}: \mathbb{R}^n \to \mathbb{R}$ by $$u_{(\xi,\varepsilon)}(x) = \Big ( \frac{\varepsilon}{\varepsilon^2 + |x - \xi|^2} \Big )^{\frac{n-2}{2}}.$$ The function $u_{(\xi,\varepsilon)}$ satisfies the elliptic PDE $$\Delta u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} = 0.$$ It is well known that $$\int_{\mathbb{R}^n} u_{(\xi,\varepsilon)}^{\frac{2n}{n-2}} = \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}}$$ for all $(\xi,\varepsilon) \in \mathbb{R}^n \times (0,\infty)$. We next define $$\varphi_{(\xi,\varepsilon,0)}(x) = \Big ( \frac{\varepsilon}{\varepsilon^2 + |x - \xi|^2} \Big )^{\frac{n+2}{2}} \, \frac{\varepsilon^2 - |x - \xi|^2}{\varepsilon^2 + |x - \xi|^2}$$ and $$\varphi_{(\xi,\varepsilon,k)}(x) = \Big ( \frac{\varepsilon}{\varepsilon^2 + |x - \xi|^2} \Big )^{\frac{n+2}{2}} \, \frac{2\varepsilon \, (x_k - \xi_k)}{\varepsilon^2 + |x - \xi|^2}$$ for $k = 1,\hdots,n$. It is easy to see that the norm $\|\varphi_{(\xi,\varepsilon,k)}\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)}$ is constant in $\xi$ and $\varepsilon$. Finally, we define a closed subspace $\mathcal{E}_{(\xi,\varepsilon)} \subset \mathcal{E}$ by $$\mathcal{E}_{(\xi,\varepsilon)} = \bigg \{ w \in \mathcal{E}: \int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, w = 0 \quad \text{for $k = 0,1,\hdots,n$} \bigg \}.$$ Clearly, $u_{(\xi,\varepsilon)} \in \mathcal{E}_{(\xi,\varepsilon)}$. \[bound.for.error.term\] Consider a Riemannian metric on $\mathbb{R}^n$ of the form $g(x) = \exp(h(x))$, where $h(x)$ is a trace-free symmetric two-tensor on $\mathbb{R}^n$ satisfying $|h(x)| + |\partial h(x)| + |\partial^2 h(x)| \leq \alpha \leq 1$ for all $x \in \mathbb{R}^n$ and $h(x) = 0$ for $|x| \geq 1$. There exists a constant $C$, depending only on $n$, such that $$\Big \| \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \leq C \, \alpha$$ for all pairs $(\xi,\varepsilon) \in \mathbb{R}^n \times (0,\infty)$. **Proof.** Using the pointwise estimate $$\begin{aligned} &\Big | \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big | \\ &\leq C \, |h| \, |\partial^2 u_{(\xi,\varepsilon)}| + C \, |\partial h| \, |\partial u_{(\xi,\varepsilon)}| + C \, (|\partial^2 h| + |\partial h|^2) \, u_{(\xi,\varepsilon)}, \end{aligned}$$ we obtain $$\begin{aligned} &\Big \| \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \\ &\leq C \, \|h\|_{L^\infty(\mathbb{R}^n)} \, \|\partial^2 u_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} + C \, \|\partial h\|_{L^n(\mathbb{R}^n)} \, \|\partial u_{(\xi,\varepsilon)}\|_{L^2(\mathbb{R}^n)} \\ &+ C \, (\|\partial^2 h\|_{L^{\frac{n}{2}}(\mathbb{R}^n)} + \|\partial h\|_{L^n(\mathbb{R}^n)}^2) \, \|u_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \\ &\leq C \, \alpha. \end{aligned}$$ This proves the assertion.\ \[eigenvalue.estimate.1\] There exists a positive constant $\theta$, depending only on $n$, such that $$\begin{aligned} &\int_{\mathbb{R}^n} \Big ( |dw|^2 - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w^2 \Big ) \\ &\geq 2\theta \, \|w\|_{\mathcal{E}}^2 - \frac{16n^2}{\theta} \, \bigg ( \int_{\mathbb{R}^n} u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \, w \bigg )^2 \end{aligned}$$ for all $w \in \mathcal{E}_{(\xi,\varepsilon)}$. Proposition \[eigenvalue.estimate.1\] follows from an analysis of the eigenvalues of the Laplace operator on $S^n$. The details can be found in [@Rey].\ \[eigenvalue.estimate.2\] Consider a Riemannian metric on $\mathbb{R}^n$ of the form $g(x) = \exp(h(x))$, where $h(x)$ is a trace-free symmetric two-tensor on $\mathbb{R}^n$ satisfying $h(x) = 0$ for $|x| \geq 1$. There exists a positive constant $\alpha_0 \leq 1$, depending only on $n$, with the following property: if $|h(x)| + |\partial h(x)| + |\partial^2 h(x)| \leq \alpha_0$ for all $x \in \mathbb{R}^n$, then we have $$\bigg ( \int_{\mathbb{R}^n} |w|^{\frac{2n}{n-2}} \bigg )^{\frac{n-2}{n}} \leq 2K \, \int_{\mathbb{R}^n} \Big ( |dw|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, w^2 \Big )$$ for all $w \in \mathcal{E}$ and $$\begin{aligned} &\int_{\mathbb{R}^n} \Big ( |dw|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, w^2 - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w^2 \Big ) \\ &\geq \theta \, \|w\|_{\mathcal{E}}^2 - \frac{1}{\theta} \, \bigg ( \int_{\mathbb{R}^n} \Big ( \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big ) \, w \bigg )^2 \end{aligned}$$ for all $w \in \mathcal{E}_{(\xi,\varepsilon)}$. **Proof.** Using Proposition \[bound.for.error.term\] and Hölder’s inequality, we obtain $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} \Big ( \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big ) \, w \bigg | \\ &\geq 4n \, \bigg | \int_{\mathbb{R}^n} u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \, w \bigg | - C \, \alpha_0 \, \|w\|_{\mathcal{E}}. \end{aligned}$$ This implies $$\begin{aligned} &\bigg ( \int_{\mathbb{R}^n} \Big ( \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big ) \, w \bigg )^2 \\ &\geq 16n^2 \, \bigg ( \int_{\mathbb{R}^n} u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \, w \bigg )^2 - \theta^2 \, \|w\|_{\mathcal{E}}^2 \end{aligned}$$ if $\alpha_0$ is sufficiently small. Hence, the assertion follows from Proposition \[eigenvalue.estimate.1\].\ \[linearized.operator\] Consider a Riemannian metric on $\mathbb{R}^n$ of the form $g(x) = \exp(h(x))$, where $h(x)$ is a trace-free symmetric two-tensor on $\mathbb{R}^n$ satisfying $|h(x)| + |\partial h(x)| + |\partial^2 h(x)| \leq \alpha_0$ for all $x \in \mathbb{R}^n$ and $h(x) = 0$ for $|x| \geq 1$. Given any pair $(\xi,\varepsilon) \in \mathbb{R}^n \times (0,\infty)$ and any function $f \in L^{\frac{2n}{n+2}}(\mathbb{R}^n)$, there exists a unique function $w \in \mathcal{E}_{(\xi,\varepsilon)}$ such that $$\int_{\mathbb{R}^n} \Big ( \langle dw,d\psi \rangle_g + \frac{n-2}{4(n-1)} \, R_g \, w \, \psi - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w \, \psi \Big ) = \int_{\mathbb{R}^n} f \, \psi$$ for all test functions $\psi \in \mathcal{E}_{(\xi,\varepsilon)}$. Moreover, we have $\|w\|_{\mathcal{E}} \leq C \, \|f\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)}$, where $C$ is a constant that depends only on $n$. **Proof.** Suppose that $w \in \mathcal{E}_{(\xi,\varepsilon)}$ and $$\int_{\mathbb{R}^n} \Big ( \langle dw,d\psi \rangle_g + \frac{n-2}{4(n-1)} \, R_g \, w \, \psi - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w \, \psi \Big ) = \int_{\mathbb{R}^n} f \, \psi$$ for all test functions $\psi \in \mathcal{E}_{(\xi,\varepsilon)}$. This implies $$\int_{\mathbb{R}^n} \Big ( |dw|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, w^2 - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w^2 \Big ) = \int_{\mathbb{R}^n} f \, w$$ and $$\int_{\mathbb{R}^n} \Big ( \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big ) \, w = -\int_{\mathbb{R}^n} u_{(\xi,\varepsilon)} \, f.$$ Using Corollary \[eigenvalue.estimate.2\], we obtain $$\begin{aligned} \theta \, \|w\|_{\mathcal{E}}^2 &\leq \int_{\mathbb{R}^n} \Big ( |dw|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, w^2 - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w^2 \Big ) \\ &+ \frac{1}{\theta} \, \bigg ( \int_{\mathbb{R}^n} \Big ( \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big ) \, w \bigg )^2 \\ &\leq \bigg ( \int_{\mathbb{R}^n} |f|^{\frac{2n}{n+2}} \bigg )^{\frac{n+2}{2n}} \, \bigg ( \int_{\mathbb{R}^n} |w|^{\frac{2n}{n-2}} \bigg )^{\frac{n-2}{2n}} \\ &+ \frac{1}{\theta} \, \bigg ( \int_{\mathbb{R}^n} u_{(\xi,\varepsilon)}^{\frac{2n}{n-2}} \bigg )^{\frac{n-2}{n}} \, \bigg ( \int_{\mathbb{R}^n} |f|^{\frac{2n}{n+2}} \bigg )^{\frac{n+2}{n}} \\ &\leq K^{\frac{1}{2}} \, \|f\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \, \|w\|_{\mathcal{E}} + \frac{1}{\theta} \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n-2} {2}} \, \|f\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)}^2.\end{aligned}$$ Hence, it follows from Young’s inequality that $$\frac{\theta}{2} \, \|w\|_{\mathcal{E}}^2 \leq \frac{K}{2\theta} \, \|f\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)}^2 + \frac{1}{\theta} \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n-2}{2}} \, \|f\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)}^2.$$ From this the uniqueness statement follows easily. In order to prove the existence part, it suffices to minimize the functional $$\begin{aligned} &\int_{\mathbb{R}^n} \Big ( |dw|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, w^2 - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w^2 - 2fw \Big ) \\ &+ \frac{1}{\theta} \, \bigg ( \int_{\mathbb{R}^n} \Big ( \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big ) \, w \bigg )^2 \end{aligned}$$ over all functions $w \in \mathcal{E}_{(\xi,\varepsilon)}$.\ \[fixed.point.argument\] Consider a Riemannian metric on $\mathbb{R}^n$ of the form $g(x) = \exp(h(x))$, where $h(x)$ is a trace-free symmetric two-tensor on $\mathbb{R}^n$ satisfying $h(x) = 0$ for $|x| \geq 1$. Moreover, let $(\xi,\varepsilon) \in \mathbb{R}^n \times (0,\infty)$. There exists a positive constant $\alpha_1 \leq \alpha_0$, depending only on $n$, with the following property: if $|h(x)| + |\partial h(x)| + |\partial^2 h(x)| \leq \alpha_1$ for all $x \in \mathbb{R}^n$, then there exists a function $v_{(\xi,\varepsilon)} \in \mathcal{E}$ such that $v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} \in \mathcal{E}_{(\xi,\varepsilon)}$ and $$\int_{\mathbb{R}^n} \Big ( \langle dv_{(\xi,\varepsilon)},d\psi \rangle_g + \frac{n-2}{4(n-1)} \, R_g \, v_{(\xi,\varepsilon)} \, \psi - n(n-2) \, |v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} \, v_{(\xi,\varepsilon)} \, \psi \Big ) = 0$$ for all test functions $\psi \in \mathcal{E}_{(\xi,\varepsilon)}$. Moreover, we have the estimate $$\begin{aligned} &\|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}\|_{\mathcal{E}} \\ &\leq C \, \Big \| \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)}, \end{aligned}$$ where $C$ is a constant that depends only on $n$. **Proof.** Let $G_{(\xi,\varepsilon)}: L^{\frac{2n}{n+2}}(\mathbb{R}^n) \to \mathcal{E}_{(\xi,\varepsilon)}$ be the solution operator constructed in Proposition \[linearized.operator\]. We define a nonlinear mapping $\Phi_{(\xi,\varepsilon)}: \mathcal{E}_{(\xi,\varepsilon)} \to \mathcal{E}_{(\xi,\varepsilon)}$ by $$\begin{aligned} &\Phi_{(\xi,\varepsilon)}(w) \\ &= G_{(\xi,\varepsilon)} \Big ( \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big ) \\ &+ n(n-2) \, G_{(\xi,\varepsilon)} \Big ( |u_{(\xi,\varepsilon)} + w|^{\frac{4}{n-2}} \, (u_{(\xi,\varepsilon)} + w) - u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} - \frac{n+2}{n-2} \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w \Big ). \end{aligned}$$ It follows from Proposition \[bound.for.error.term\] that $\|\Phi_{(\xi,\varepsilon)}(0)\|_{\mathcal{E}} \leq C \, \alpha_1$. Using the pointwise estimate $$\begin{aligned} &\Big | |u_{(\xi,\varepsilon)} + w|^{\frac{4}{n-2}} \, (u_{(\xi,\varepsilon)} + w) - |u_{(\xi,\varepsilon)} + \tilde{w}|^{\frac{4}{n-2}} \, (u_{(\xi,\varepsilon)} + \tilde{w}) \\ &\hspace{20mm} - \frac{n+2}{n-2} \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, (w - \tilde{w}) \Big | \\ &\leq C \, (|w|^{\frac{4}{n-2}} + |\tilde{w}|^{\frac{4}{n-2}}) \, |w - \tilde{w}|, \end{aligned}$$ we obtain $$\begin{aligned} &\|\Phi_{(\xi,\varepsilon)}(w) - \Phi_{(\xi,\varepsilon)}(\tilde{w})\|_{\mathcal{E}} \\ &\leq C \, \Big \| |u_{(\xi,\varepsilon)} + w|^{\frac{4}{n-2}} \, (u_{(\xi,\varepsilon)} + w) - |u_{(\xi,\varepsilon)} + \tilde{w}|^{\frac{4}{n-2}} \, (u_{(\xi,\varepsilon)} + \tilde{w}) \\ &\hspace{20mm} - \frac{n+2}{n-2} \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, (w - \tilde{w}) \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \\ &\leq C \, (\|w\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)}^{\frac{4}{n-2}} + \|\tilde{w}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)}^{\frac{4}{n-2}}) \, \|w - \tilde{w}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \end{aligned}$$ for all functions $w,\tilde{w} \in \mathcal{E}_{(\xi,\varepsilon)}$. This implies $$\|\Phi_{(\xi,\varepsilon)}(w) - \Phi_{(\xi,\varepsilon)}(\tilde{w})\|_{\mathcal{E}} \leq C \, (\|w\|_{\mathcal{E}}^{\frac{4}{n-2}} + \|\tilde{w}\|_{\mathcal{E}}^{\frac{4}{n-2}}) \, \|w - \tilde{w}\|_{\mathcal{E}}$$ for $w,\tilde{w} \in \mathcal{E}_{(\xi,\varepsilon)}$. Hence, if $\alpha_1$ is sufficiently small, then the contraction mapping principle implies that the mapping $\Phi_{(\xi,\varepsilon)}$ has a unique fixed point. From this the assertion follows easily.\ We next define a function $\mathcal{F}_g: \mathbb{R}^n \times (0,\infty) \to \mathbb{R}$ by $$\begin{aligned} \mathcal{F}_g(\xi,\varepsilon) &= \int_{\mathbb{R}^n} \Big ( |dv_{(\xi,\varepsilon)}|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, v_{(\xi,\varepsilon)}^2 - (n-2)^2 \, |v_{(\xi,\varepsilon)}|^{\frac{2n}{n-2}} \Big ) \\ &- 2(n-2) \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}}. \end{aligned}$$ If we choose $\alpha_1$ small enough, then we obtain the following result:\ \[reduction.to.a.finite.dimensional.problem\] The function $\mathcal{F}_g$ is continuously differentiable. Moreover, if $(\bar{\xi},\bar{\varepsilon})$ is a critical point of the function $\mathcal{F}_g$, then the function $v_{(\bar{\xi},\bar{\varepsilon})}$ is a non-negative weak solution of the equation $$\Delta_g v_{(\bar{\xi},\bar{\varepsilon})} - \frac{n-2}{4(n-1)} \, R_g \, v_{(\bar{\xi},\bar{\varepsilon})} + n(n-2) \, v_{(\bar{\xi},\bar{\varepsilon})}^{\frac{n+2}{n-2}} = 0.$$ **Proof.** By definition of $v_{(\xi,\varepsilon)}$, we can find real numbers $a_k(\xi,\varepsilon)$, $k = 0,1, \hdots, n$, such that $$\begin{aligned} &\int_{\mathbb{R}^n} \Big ( \langle dv_{(\xi,\varepsilon)},d\psi \rangle_g + \frac{n-2}{4(n-1)} \, R_g \, v_{(\xi,\varepsilon)} \, v_{(\xi,\varepsilon)} \, \psi - n(n-2) \, |v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} \, v_{(\xi,\varepsilon)} \, \psi \Big ) \\ &= \sum_{k=0}^n a_k(\xi,\varepsilon) \, \int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, \psi \end{aligned}$$ for all test functions $\psi \in \mathcal{E}$. This implies $$\frac{\partial}{\partial \varepsilon} \mathcal{F}_g(\varepsilon,\xi) = 2 \, \sum_{k=0}^n a_k(\xi,\varepsilon) \, \int_{\mathbb{R}^n} \varphi_{\xi,\varepsilon,k)} \, \frac{\partial}{\partial \varepsilon} v_{(\xi,\varepsilon)}$$ and $$\frac{\partial}{\partial \xi_j} \mathcal{F}_g(\varepsilon,\xi) = 2 \, \sum_{k=0}^n a_k(\xi,\varepsilon) \, \int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, \frac{\partial}{\partial \xi_j} v_{(\xi,\varepsilon)}$$ for $j = 1, \hdots, n$. On the other hand, we have $$\int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) = 0$$ since $v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} \in \mathcal{E}_{(\xi,\varepsilon)}$. This implies $$\begin{aligned} 0 &= \int_{\mathbb{R}^n} \frac{\partial}{\partial \varepsilon} \varphi_{(\xi,\varepsilon,k)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) + \int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, \frac{\partial}{\partial \varepsilon} (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) \\ &= \int_{\mathbb{R}^n} \frac{\partial}{\partial \varepsilon} \varphi_{(\xi,\varepsilon,k)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) + \int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, \frac{\partial}{\partial \varepsilon} v_{(\xi,\varepsilon)} \\ &+ \frac{n-2}{2(n+1)} \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}} \, \varepsilon^{-1} \, \delta_{0k} \end{aligned}$$ and $$\begin{aligned} 0 &= \int_{\mathbb{R}^n} \frac{\partial}{\partial \xi_j} \varphi_{(\xi,\varepsilon,k)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) + \int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, \frac{\partial}{\partial \xi_j} (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) \\ &= \int_{\mathbb{R}^n} \frac{\partial}{\partial \xi_j} \varphi_{(\xi,\varepsilon,k)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) + \int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, \frac{\partial}{\partial \xi_j} v_{(\xi,\varepsilon)} \\ &- \frac{n-2}{2(n+1)} \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}} \, \varepsilon^{-1} \, \delta_{jk} \end{aligned}$$ for $j = 1,\hdots,n$. Putting these facts together, we obtain $$\begin{aligned} &- \frac{n-2}{n+1} \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}} \, a_0(\xi,\varepsilon) \\ &= \varepsilon \, \frac{\partial}{\partial \varepsilon} \mathcal{F}_g(\xi,\varepsilon) + 2\varepsilon \, \sum_{k=0}^n a_k(\xi,\varepsilon) \, \int_{\mathbb{R}^n} \frac{\partial}{\partial \varepsilon} \varphi_{(\xi,\varepsilon,k)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) \end{aligned}$$ and $$\begin{aligned} &\frac{n-2}{n+1} \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}} \, a_j(\xi,\varepsilon) \\ &= \varepsilon \, \frac{\partial}{\partial \xi_j} \mathcal{F}_g(\xi,\varepsilon) + 2\varepsilon \, \sum_{k=0}^n a_k(\xi,\varepsilon) \, \int_{\mathbb{R}^n} \frac{\partial}{\partial \xi_j} \varphi_{(\xi,\varepsilon,k)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) \end{aligned}$$ for $j = 1, \hdots, n$. Hence, if $(\bar{\xi},\bar{\varepsilon})$ is a critical point of $\mathcal{F}_g$, then we have $$\sum_{k=0}^n |a_k(\bar{\xi},\bar{\varepsilon})| \leq C \, \|v_{(\bar{\xi},\bar{\varepsilon})} - u_{(\bar{\xi},\bar{\varepsilon})}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \, \sum_{k=0}^n |a_k(\bar{\xi},\bar{\varepsilon})|,$$ where $C$ is a constant that depends only on $n$. On the other hand, we have $\|v_{(\bar{\xi},\bar{\varepsilon})} - u_{(\bar{\xi},\bar{\varepsilon})}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \leq C \, \alpha_1$. Hence, if we choose $\alpha_1$ sufficiently small, then we must have $a_k(\bar{\xi},\bar{\varepsilon}) = 0$ for $k = 0,1, \hdots, n$. Thus, we conclude that $$\int_{\mathbb{R}^n} \Big ( \langle dv_{(\bar{\xi},\bar{\varepsilon})},d\psi \rangle_g + \frac{n-2}{4(n-1)} \, R_g \, v_{(\bar{\xi},\bar{\varepsilon})} \, \psi - n(n-2) \, |v_{(\bar{\xi},\bar{\varepsilon})}|^{\frac{4}{n-2}} \, v_{(\bar{\xi},\bar{\varepsilon})} \, \psi \Big ) = 0$$ for all test functions $\psi \in \mathcal{E}$. It remains to show that the function $v_{(\bar{\xi},\bar{\varepsilon})}$ is non-negative. To that end, we put $\psi = \min \{v_{(\bar{\xi},\bar{\varepsilon})},0\}$. Since $v_{(\bar{\xi},\bar{\varepsilon})} \in \mathcal{E}$, we conclude that $\psi \in \mathcal{E}$. This implies $$\begin{aligned} &\int_{\{v_{(\bar{\xi},\bar{\varepsilon})} < 0\}} \Big ( |dv_{(\bar{\xi},\bar{\varepsilon})}|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, v_{(\bar{\xi},\bar{\varepsilon})}^2 \Big ) \\ &= n(n-2) \, \int_{\{v_{(\bar{\xi},\bar{\varepsilon})} < 0\}} |v_{(\bar{\xi},\bar{\varepsilon})}|^{\frac{2n}{n-2}}. \end{aligned}$$ Moreover, we have $$\begin{aligned} &\bigg ( \int_{\{v_{(\bar{\xi},\bar{\varepsilon})} < 0\}} |v_{(\bar{\xi},\bar{\varepsilon})}|^{\frac{2n}{n-2}} \bigg )^{\frac{n-2}{n}} \\ &\leq 2K \int_{\{v_{(\bar{\xi},\bar{\varepsilon})} < 0\}} \Big ( |dv_{(\bar{\xi},\bar{\varepsilon})}|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, v_{(\bar{\xi},\bar{\varepsilon})}^2 \Big ) \end{aligned}$$ by Corollary \[eigenvalue.estimate.2\]. From this we deduce that either $v_{(\bar{\xi},\bar{\varepsilon})} \geq 0$ almost everywhere, or $$\bigg ( \int_{\{v_{(\bar{\xi},\bar{\varepsilon})} < 0\}} |v_{(\bar{\xi},\bar{\varepsilon})}|^{\frac{2n}{n-2}} \bigg )^{\frac{2}{n}} \geq \frac{1}{2n(n-2) \, K}.$$ On the other hand, we have $$\bigg ( \int_{\{v_{(\bar{\xi},\bar{\varepsilon})} < 0\}} |v_{(\bar{\xi},\bar{\varepsilon})}|^{\frac{2n}{n-2}} \bigg )^{\frac{n-2}{2n}} \leq \bigg ( \int_{\mathbb{R}^n} |v_{(\bar{\xi},\bar{\varepsilon})} - u_{(\bar{\xi},\bar{\varepsilon})}|^{\frac{2n}{n-2}} \bigg )^{\frac{n-2}{2n}} \leq C \, \alpha_1.$$ Hence, if $\alpha_1$ is sufficiently small, then we have $v_{(\bar{\xi},\bar{\varepsilon})} \geq 0$ almost everywhere.\ An estimate for the energy of a “bubble” ======================================== Throughout this paper, we fix a multi-linear form $W: \mathbb{R}^n \times \mathbb{R}^n \times \mathbb{R}^n \times \mathbb{R}^n \to \mathbb{R}$. We assume that $W_{ijkl}$ satisfy all the algebraic properties of the Weyl tensor. Moreover, we assume that some components of $W$ are non-zero, so that $$\sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 > 0.$$ For abbreviation, we put $$H_{ik}(x) = \sum_{p,q=1}^n W_{ipkq} \, x_p \, x_q$$ and $$\overline{H}_{ik}(x) = (1 - |x|^2) \, H_{ik}(x).$$ It is easy to see that $H_{ik}(x)$ is trace-free, $\sum_{i=1}^n x_i \, H_{ik}(x) = 0$, and $\sum_{i=1}^n \partial_i H_{ik}(x) = 0$ for all $x \in \mathbb{R}^n$.\ We consider a Riemannian metric of the form $g(x) = \exp(h(x))$, where $h(x)$ is a trace-free symmetric two-tensor on $\mathbb{R}^n$ satisfying $h(x) = 0$ for $|x| \geq 1$, $$|h(x)| + |\partial h(x)| + |\partial^2 h(x)| \leq \alpha_1$$ for all $x \in \mathbb{R}^n$, and $$h_{ik}(x) = \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x)$$ for $|x| \leq \rho$. We assume that the parameters $\lambda$, $\mu$, and $\rho$ are chosen such that $\mu \leq 1$ and $\lambda \leq \rho \leq 1$. Note that $\sum_{i=1}^n x_i \, h_{ik}(x) = 0$ and $\sum_{i=1}^n \partial_i h_{ik}(x) = 0$ for $|x| \leq \rho$. Given any pair $(\xi,\varepsilon) \in \mathbb{R}^n \times (0,\infty)$, there exists a unique function $v_{(\xi,\varepsilon)}$ such that $v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} \in \mathcal{E}_{(\xi,\varepsilon)}$ and $$\int_{\mathbb{R}^n} \Big ( \langle dv_{(\xi,\varepsilon)},d\psi \rangle_g + \frac{n-2}{4(n-1)} \, R_g \, v_{(\xi,\varepsilon)} \, \psi - n(n-2) \, |v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} \, v_{(\xi,\varepsilon)} \, \psi \Big ) = 0$$ for all test functions $\psi \in \mathcal{E}_{(\xi,\varepsilon)}$ (see Proposition \[fixed.point.argument\]). For abbreviation, let $$\Omega = \bigg \{ (\xi,\varepsilon) \in \mathbb{R}^n \times \mathbb{R}: |\xi| < 1, \, \frac{n-8}{3(n+4)} < \varepsilon^2 < \frac{2(n-8)}{3(n+4)} \bigg \}.$$ \[estimate.for.error.term\] For every pair $(\xi,\varepsilon) \in \lambda \, \Omega$, we have $$\begin{aligned} &\Big \| \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \\ &\leq C \, \lambda^4 \, \mu + C \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}} \end{aligned}$$ and $$\begin{aligned} &\Big \| \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \\ &\hspace{10mm} + \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \\ &\leq C \, \lambda^8 \, \mu^2 + C \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}}. \end{aligned}$$ **Proof.** For abbreviation, we define two functions $A_1$ and $A_2$ by $$A_1 = \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}}$$ and $$A_2 = \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)}.$$ Using Proposition \[Taylor.expansion.of.scalar.curvature\] and the identity $\sum_{i=1}^n \partial_i h_{ik}(x) = 0$, we obtain $$|R_g(x)| \leq C \, |h(x)|^2 \, |\partial^2 h(x)| + C \, |\partial h(x)|^2 \leq C \, \mu^2 \, (\lambda + |x|)^6$$ for $|x| \leq \rho$. This implies $$\begin{aligned} |A_1| &= \Big | \sum_{i,k=1}^n \partial_i \big [ (g^{ik} - \delta_{ik}) \, \partial_k u_{(\xi,\varepsilon)} \big ] - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} \Big | \\ &\leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, (\lambda + |x|)^{4-n}\end{aligned}$$ and $$\begin{aligned} |A_1 + A_2| &= \Big | \sum_{i,k=1}^n \partial_i \big [ (g^{ik} - \delta_{ik} + h_{ik}) \, \partial_k u_{(\xi,\varepsilon)} \big ] - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} \Big | \\ &\leq C \, \lambda^{\frac{n-2}{2}} \, \mu^2 \, (\lambda + |x|)^{8-n}\end{aligned}$$ for $|x| \leq \rho$. Hence, we obtain $$\|A_1\|_{L^{\frac{2n}{n+2}}(B_\rho(0))} \leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, \bigg ( \int_{\mathbb{R}^n} (\lambda + |x|)^{-\frac{2n(n-4)}{n+2}} \bigg )^{\frac{n+2}{2n}} \leq C \, \lambda^4 \, \mu$$ and $$\|A_1 + A_2\|_{L^{\frac{2n}{n+2}}(B_\rho(0))} \leq C \, \lambda^{\frac{n-2}{2}} \, \mu^2 \, \bigg ( \int_{\mathbb{R}^n} (\lambda + |x|)^{-\frac{2n(n-8)}{n+2}} \bigg )^{\frac{n+2}{2n}} \leq C \, \lambda^8 \, \mu^2.$$ On the other hand, we have $$|A_1(x)| \leq C \, \lambda^{\frac{n-2}{2}} \, |x|^{-n}$$ for $\rho \leq |x| \leq 1$ and $$|A_2(x)| \leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, |x|^{4-n}$$ for $|x| \geq \rho$. Since the function $A_1(x)$ vanishes for $|x| \geq 1$, we conclude that $$\|A_1\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n \setminus B_\rho(0))} \\ \leq C \, \lambda^{\frac{n-2}{2}} \, \bigg ( \int_{\mathbb{R}^n \setminus B_\rho(0)} |x|^{-\frac{2n^2}{n+2}} \bigg )^{\frac{n+2}{2n}} \leq C \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}}$$ and $$\|A_2\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n \setminus B_\rho(0))} \leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, \bigg ( \int_{\mathbb{R}^n \setminus B_\rho(0)} |x|^{-\frac{2n(n-4)}{n+2}} \bigg )^{\frac{n+2}{2n}} \leq C \, \rho^4 \, \mu \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}}.$$ Putting these facts together, the assertion follows.\ \[estimate.for.v.1\] The function $v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}$ satisfies the estimate $$\|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \leq C \, \lambda^4 \, \mu + C \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}}$$ for $(\xi,\varepsilon) \in \lambda \, \Omega$. **Proof.** It follows from Proposition \[fixed.point.argument\] that $$\begin{aligned} &\|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \\ &\leq C \, \Big \| \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)}, \end{aligned}$$ where $C$ is a constant that depends only on $n$. Hence, the assertion follows from Proposition \[estimate.for.error.term\].\ We now prove a more refined estimate for the difference $v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}$. Using Proposition \[linearized.operator\] with $h = 0$, we conclude that there exists a unique function $w_{(\xi,\varepsilon)} \in \mathcal{E}_{(\xi,\varepsilon)}$ such that $$\begin{aligned} \label{def.w} &\int_{\mathbb{R}^n} \Big ( \langle dw_{(\xi,\varepsilon)},d\psi \rangle - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w_{(\xi,\varepsilon)} \, \psi \Big ) \notag \\ &= -\int_{\mathbb{R}^n} \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} \, \psi \end{aligned}$$ for all test functions $\psi \in \mathcal{E}_{(\xi,\varepsilon)}$. \[properties.of.w\] The function $w_{(\xi,\varepsilon)}$ is smooth. Moreover, if $(\xi,\varepsilon) \in \lambda \, \Omega$, then we have $$\begin{aligned} &|w_{(\xi,\varepsilon)}(x)| \leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, (\lambda + |x|)^{6-n} \\ &|\partial w_{(\xi,\varepsilon)}(x)| \leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, (\lambda + |x|)^{5-n} \\ &|\partial^2 w_{(\xi,\varepsilon)}(x)| \leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, (\lambda + |x|)^{4-n} \end{aligned}$$ for all $x \in \mathbb{R}^n$. **Proof.** Let $\varphi_{(\xi,\varepsilon,k)}$ be the functions defined in Section 2. We can find real numbers $b_k(\xi,\varepsilon)$, $k = 0,1,\hdots,n$, such that $$\begin{aligned} &\int_{\mathbb{R}^n} \Big ( \langle dw_{(\xi,\varepsilon)},d\psi \rangle - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w_{(\xi,\varepsilon)} \, \psi \Big ) \notag \\ &= -\int_{\mathbb{R}^n} \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} \, \psi + \sum_{k=0}^n b_k(\xi,\varepsilon) \, \int_{\mathbb{R}^n} \varphi_{(\xi,\varepsilon,k)} \, \psi \end{aligned}$$ for all test functions $\psi \in \mathcal{E}$. It follows from standard elliptic regularity theory that $w_{(\xi,\varepsilon)}$ is smooth. In the next step, we establish quantitative estimates for $w_{(\xi,\varepsilon)}$. To that end, we consider a pair $(\xi,\varepsilon) \in \lambda \, \Omega$. A straightforward calculation yields $$\label{estimate} \Big \| \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \leq C \, \lambda^4 \, \mu.$$ From this we deduce that $\|w_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \leq C \, \lambda^4 \, \mu$ and $\sum_{k=0}^n |b_k(\xi,\varepsilon)| \leq C \, \lambda^4 \, \mu$. This implies $$\begin{aligned} &\big | \Delta w_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w_{(\xi,\varepsilon)} \big | \\ &= \bigg | \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} - \sum_{k=0}^n b_k(\xi,\varepsilon) \, \varphi_{(\xi,\varepsilon,k)} \bigg | \\ &\leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, (\lambda + |x|)^{4-n}\end{aligned}$$ for all $x \in \mathbb{R}^n$. We claim that $$\sup_{x \in \mathbb{R}^n} (\lambda + |x|)^{\frac{n-2}{2}} \, |w_{(\xi,\varepsilon)}(x)| \leq C \, \lambda^4 \, \mu.$$ To show this, we fix a point $x_0 \in \mathbb{R}^n$ and put $r = \frac{1}{2} \, (\lambda + |x_0|)$. Clearly, $\lambda + |x| \geq r$ for all $x \in B_r(x_0)$. This implies $$u_{(\xi,\varepsilon)}(x)^{\frac{4}{n-2}} \leq C \, r^{-2}$$ and $$\big | \Delta w_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w_{(\xi,\varepsilon)} \big | \leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, r^{4-n}$$ for all $x \in B_r(x_0)$. Using standard interior estimates, we obtain $$\begin{aligned} r^{\frac{n-2}{2}} \, |w_{(\xi,\varepsilon)}(x_0)| &\leq C \, \|w_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(B_r(x_0))} \\ &+ C \, r^{\frac{n+2}{2}} \, \big \| \Delta w_{(\xi,\varepsilon)} + n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w_{(\xi,\varepsilon)} \big \|_{L^\infty(B_r(x_0))} \\ &\leq C \, \lambda^4 \, \mu + C \, \lambda^{\frac{n-2}{2}} \, \mu \, r^{-\frac{n-10}{2}} \\ &\leq C \, \lambda^4 \, \mu. \end{aligned}$$ Thus, we conclude that $$\sup_{x \in \mathbb{R}^n} (\lambda + |x|)^{\frac{n-2}{2}} \, |w_{(\xi,\varepsilon)}(x)| \leq C \, \lambda^4 \, \mu,$$ as claimed. Since $\sup_{x \in \mathbb{R}^n} |x|^{\frac{n-2}{2}} \, |w_{(\xi,\varepsilon)}(x)| < \infty$, we can express the function $w_{(\xi,\varepsilon)}$ in the form $$\label{convolution.formula} w_{(\xi,\varepsilon)}(x) = -\frac{1}{(n-2) \, |S^{n-1}|} \int_{\mathbb{R}^n} |x - y|^{2-n} \, \Delta w_{(\xi,\varepsilon)}(y) \, dy$$ for all $x \in \mathbb{R}^n$. We can now use a bootstrap argument to prove the desired estimate for $w_{(\xi,\varepsilon)}$. It follows from (\[convolution.formula\]) that $$\sup_{x \in \mathbb{R}^n} (\lambda + |x|)^\beta \, |w_{(\xi,\varepsilon)}(x)| \leq C \, \sup_{x \in \mathbb{R}^n} (\lambda + |x|)^{\beta+2} \, |\Delta w_{(\xi,\varepsilon)}(x)|$$ for all $0 < \beta < n-2$. Since $$\begin{aligned} |\Delta w_{(\xi,\varepsilon)}(x)| &\leq n(n+2) \, u_{(\xi,\varepsilon)}(x)^{\frac{4}{n-2}} \, |w_{(\xi,\varepsilon)}(x)| \\ &+ C \, \lambda^{\frac{n-2}{2}} \, \mu \, (\lambda + |x|)^{4-n} \end{aligned}$$ for all $x \in \mathbb{R}^n$, we conclude that $$\begin{aligned} \sup_{x \in \mathbb{R}^n} (\lambda + |x|)^\beta \, |w_{(\xi,\varepsilon)}(x)| &\leq C \, \lambda^2 \, \sup_{x \in \mathbb{R}^n} (\lambda + |x|)^{\beta-2} \, |w_{(\xi,\varepsilon)}(x)| \\ &+ C \, \lambda^{\beta-\frac{n-10}{2}} \, \mu \end{aligned}$$ for all $0 < \beta \leq n-6$. Iterating this inequality, we obtain $$\sup_{x \in \mathbb{R}^n} (\lambda + |x|)^{n-6} \, |w_{(\xi,\varepsilon)}(x)| \leq C \, \lambda^{\frac{n-2}{2}} \, \mu.$$ The estimates for the first and second derivatives of $w_{(\xi,\varepsilon)}$ follow now from standard interior estimates.\ \[estimate.for.v.2\] The function $v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} - w_{(\xi,\varepsilon)}$ satisfies the estimate $$\|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} - w_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \leq C \, \lambda^{\frac{4(n+2)}{n-2}} \, \mu^{\frac{n+2}{n-2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}}$$ for $(\xi,\varepsilon) \in \lambda \, \Omega$. **Proof.** Consider the functions $$B_1 = \sum_{i,k=1}^n \partial_i \big [ (g^{ik} - \delta_{ik}) \, \partial_k w_{(\xi,\varepsilon)} \big ] - \frac{n-2}{4(n-1)} \, R_g \, w_{(\xi,\varepsilon)}$$ and $$B_2 = \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)}.$$ Using (\[def.w\]), we obtain $$\begin{aligned} &\int_{\mathbb{R}^n} \Big ( \langle dw_{(\xi,\varepsilon)},d\psi \rangle_g + \frac{n-2}{4(n-1)} \, R_g \, w_{(\xi,\varepsilon)} \, \psi - n(n+2) \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, w_{(\xi,\varepsilon)} \, \psi \Big ) \\ &= -\int_{\mathbb{R}^n} (B_1 + B_2) \, \psi \end{aligned}$$ for all functions $\psi \in \mathcal{E}_{(\xi,\varepsilon)}$. Since $w_{(\xi,\varepsilon)} \in \mathcal{E}_{(\xi,\varepsilon)}$, it follows that $$w_{(\xi,\varepsilon)} = -G_{(\xi,\varepsilon)}(B_1 + B_2).$$ Moreover, we have $$v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} = G_{(\xi,\varepsilon)} \big ( B_3 + n(n-2) \, B_4 \big ),$$ where $$B_3 = \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}}$$ and $$B_4 = |v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} \, v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} - \frac{n+2}{n-2} \, u_{(\xi,\varepsilon)}^{\frac{4}{n-2}} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}).$$ Thus, we conclude that $$v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} - w_{(\xi,\varepsilon)} = G_{(\xi,\varepsilon)} \big ( B_1 + B_2 + B_3 + n(n-2) \, B_4 \big ),$$ where $G_{(\xi,\varepsilon)}: L^{\frac{2n}{n+2}}(\mathbb{R}^n) \to \mathcal{E}_{(\xi,\varepsilon)}$ denotes the solution operator constructed in Proposition \[linearized.operator\]. In particular, we have $$\|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} - w_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \leq C \, \big \| B_1 + B_2 + B_3 + n(n-2) \, B_4 \big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)}$$ by Proposition \[linearized.operator\]. Using Proposition \[properties.of.w\], we obtain $$|B_1(x)| \leq C \, \lambda^{\frac{n-2}{2}} \, \mu^2 \, (\lambda + |x|)^{8-n}$$ for $|x| \leq \rho$ and $$|B_1(x)| \leq C \, \lambda^{\frac{n-2}{2}} \, \mu \, |x|^{4-n}$$ for $\rho \leq |x| \leq 1$. Since the function $B_1(x)$ vanishes for $|x| \geq 1$, we conclude that $$\|B_1\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \leq C \, \lambda^8 \, \mu^2 + C \, \rho^4 \, \mu \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}}.$$ Moreover, we have $$\|B_2 + B_3\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \leq C \, \lambda^8 \, \mu^2 + C \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}}$$ by Proposition \[estimate.for.error.term\]. Finally, the function $B_4$ satisfies a pointwise estimate of the form $$|B_4| \leq C \, |v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}|^{\frac{n+2}{n-2}},$$ where $C$ is a constant that depends only on $n$. Hence, it follows from Corollary \[estimate.for.v.1\] that $$\begin{aligned} \|B_4\|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} &\leq C \, \|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)}^{\frac{n+2}{n-2}} \\ &\leq C \, \lambda^{\frac{4(n+2)}{n-2}} \, \mu^{\frac{n+2}{n-2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n+2}{2}}. \end{aligned}$$ Putting these facts together, we obtain $$\|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} - w_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \leq C \, \lambda^{\frac{4(n+2)}{n-2}} \, \mu^{\frac{n+2}{n-2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}},$$ as claimed.\ \[term.1\] We have $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} \Big ( |dv_{(\xi,\varepsilon)}|_g^2 - |du_{(\xi,\varepsilon)}|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, (v_{(\xi,\varepsilon)}^2 - u_{(\xi,\varepsilon)}^2) \Big ) \\ &\hspace{10mm} + \int_{\mathbb{R}^n} n(n-2) \, (|v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} - u_{(\xi,\varepsilon)}^{\frac{4}{n-2}}) \, u_{(\xi,\varepsilon)} \, v_{(\xi,\varepsilon)} \\ &\hspace{10mm} - \int_{\mathbb{R}^n} n(n-2) \, (|v_{(\xi,\varepsilon)}|^{\frac{2n}{n-2}} - u_{(\xi,\varepsilon)}^{\frac{2n}{n-2}}) \\ &\hspace{10mm} - \int_{\mathbb{R}^n} \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} \, w_{(\xi,\varepsilon)} \bigg | \\ &\leq C \, \lambda^{\frac{8n}{n-2}} \, \mu^{\frac{2n}{n-2}} + C \, \lambda^4 \, \mu \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^{n-2} \end{aligned}$$ for $(\xi,\varepsilon) \in \lambda \, \Omega$. **Proof.** Using Proposition \[fixed.point.argument\] with $\psi = v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}$, we obtain $$\begin{aligned} &\int_{\mathbb{R}^n} \Big ( |dv_{(\xi,\varepsilon)}|_g^2 - \langle du_{(\xi,\varepsilon)},dv_{(\xi,\varepsilon)} \rangle_g + \frac{n-2}{4(n-1)} \, R_g \, v_{(\xi,\varepsilon)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) \Big ) \\ &\hspace{10mm} - \int_{\mathbb{R}^n} n(n-2) \, |v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} \, v_{(\xi,\varepsilon)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) = 0. \end{aligned}$$ Moreover, it follows from Proposition \[estimate.for.error.term\] and Corollary \[estimate.for.v.1\] that $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} \Big ( \langle du_{(\xi,\varepsilon)},dv_{(\xi,\varepsilon)} \rangle_g - |du_{(\xi,\varepsilon)}|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) \Big ) \\ &\hspace{10mm} - \int_{\mathbb{R}^n} n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) \\ &\hspace{10mm} - \int_{\mathbb{R}^n} \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}) \bigg | \\ &\leq \Big \| \Delta_g u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)} + n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{n+2}{n-2}} \\ &\hspace{10mm} + \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} \Big \|_{L^{\frac{2n}{n+2}}(\mathbb{R}^n)} \\ &\hspace{5mm} \cdot \|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \\ &\leq C \, \lambda^{12} \, \mu^3 + C \, \lambda^4 \, \mu \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^{n-2}. \end{aligned}$$ Finally, we have $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)} \, (v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} - w_{(\xi,\varepsilon)}) \bigg | \\ &\leq C \, \lambda^4 \, \mu \, \|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)} - w_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \\ &\leq C \, \lambda^{\frac{8n}{n-2}} \, \mu^{\frac{2n}{n-2}} + C \, \lambda^4 \, \mu \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}} \end{aligned}$$ by (\[estimate\]) and Corollary \[estimate.for.v.2\]. Putting these facts together, the assertion follows.\ \[term.2\] We have $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} (|v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} - u_{(\xi,\varepsilon)}^{\frac{4}{n-2}}) \, u_{(\xi,\varepsilon)} \, v_{(\xi,\varepsilon)} - \frac{2}{n} \int_{\mathbb{R}^n} (|v_{(\xi,\varepsilon)}|^{\frac{2n}{n-2}} - u_{(\xi,\varepsilon)}^{\frac{2n}{n-2}}) \bigg | \\ &\leq C \, \lambda^{\frac{8n}{n-2}} \, \mu^{\frac{2n}{n-2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^n\end{aligned}$$ for $(\xi,\varepsilon) \in \lambda \, \Omega$. **Proof.** We have the pointwise estimate $$\begin{aligned} &\Big | (|v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} - u_{(\xi,\varepsilon)}^{\frac{4}{n-2}}) \, u_{(\xi,\varepsilon)} \, v_{(\xi,\varepsilon)} - \frac{2}{n} \, (|v_{(\xi,\varepsilon)}|^{\frac{2n}{n-2}} - u_{(\xi,\varepsilon)}^{\frac{2n}{n-2}}) \Big | \\ &\leq C \, |v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}|^{\frac{2n}{n-2}}, \end{aligned}$$ where $C$ is a constant that depends only on $n$. This implies $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} (|v_{(\xi,\varepsilon)}|^{\frac{4}{n-2}} - u_{(\xi,\varepsilon)}^{\frac{4}{n-2}}) \, u_{(\xi,\varepsilon)} \, v_{(\xi,\varepsilon)} - \frac{2}{n} \int_{\mathbb{R}^n} (|v_{(\xi,\varepsilon)}|^{\frac{2n}{n-2}} - u_{(\xi,\varepsilon)}^{\frac{2n}{n-2}}) \bigg | \\ &\leq C \, \|v_{(\xi,\varepsilon)} - u_{(\xi,\varepsilon)}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)}^{\frac{2n}{n-2}} \\ &\leq C \, \lambda^{\frac{8n}{n-2}} \, \mu^{\frac{2n}{n-2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^n \end{aligned}$$ by Corollary \[estimate.for.v.1\].\ \[term.3\] We have $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} \Big ( |du_{(\xi,\varepsilon)}|_g^2 + \frac{n-2}{4(n-1)} \, R_g \, u_{(\xi,\varepsilon)}^2 - n(n-2) \, u_{(\xi,\varepsilon)}^{\frac{2n}{n-2}} \Big ) \\ &\hspace{10mm} - \int_{B_\rho(0)} \frac{1}{2} \, \sum_{i,k,l=1}^n h_{il} \, h_{kl} \, \partial_i u_{(\xi,\varepsilon)} \, \partial_k u_{(\xi,\varepsilon)} \\ &\hspace{10mm} + \int_{B_\rho(0)} \frac{n-2}{16(n-1)} \, \sum_{i,k,l=1}^n (\partial_l h_{ik})^2 \, u_{(\xi,\varepsilon)}^2 \bigg | \\ &\leq C \, \lambda^{12} \, \mu^3 + C \, \Big ( \frac{\lambda}{\rho} \Big )^{n-2} \end{aligned}$$ for all $(\xi,\varepsilon) \in \lambda \, \Omega$. **Proof.** Note that $$\begin{aligned} &\Big | g^{ik}(x) - \delta_{ik} + h_{ik}(x) - \frac{1}{2} \, \sum_{l=1}^n h_{il}(x) \, h_{kl}(x) \Big | \\ &\leq C \, |h(x)|^3 \leq C \, \mu^3 \, (\lambda + |x|)^{12} \end{aligned}$$ for $|x| \leq \rho$. This implies $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} \big ( |du_{(\xi,\varepsilon)}|_g^2 - |du_{(\xi,\varepsilon)}|^2 \big ) + \int_{\mathbb{R}^n} \sum_{i,k=1}^n h_{ik} \, \partial_i u_{(\xi,\varepsilon)} \, \partial_k u_{(\xi,\varepsilon)} \\ &\hspace{10mm} - \int_{B_\rho(0)} \frac{1}{2} \sum_{i,k,l=1}^n h_{il} \, h_{kl} \, \partial_i u_{(\xi,\varepsilon)} \, \partial_k u_{(\xi,\varepsilon)} \bigg | \\ &\leq C \, \lambda^{n-2} \, \mu^3 \, \int_{B_\rho(0)} (\lambda + |x|)^{14-2n} + C \, \lambda^{n-2} \, \int_{\mathbb{R}^n \setminus B_\rho(0)} (\lambda + |x|)^{2-2n} \\ &\leq C \, \lambda^{12} \, \mu^3 + C \, \Big ( \frac{\lambda}{\rho} \Big )^{n-2}. \end{aligned}$$ By Proposition \[Taylor.expansion.of.scalar.curvature\], the scalar curvature of $g$ satisfies the estimate $$\begin{aligned} &\Big | R_g(x) + \frac{1}{4} \sum_{i,k,l=1}^n (\partial_l h_{ik}(x))^2 \Big | \\ &\leq C \, |h(x)|^2 \, |\partial^2 h(x)| + C \, |h(x)| \, |\partial h(x)|^2 \\ &\leq C \, \mu^3 \, (\lambda + |x|)^{10} \end{aligned}$$ for $|x| \leq \rho$. This implies $$\begin{aligned} &\bigg | \int_{\mathbb{R}^n} R_g \, u_{(\xi,\varepsilon)}^2 + \int_{B_\rho(0)} \frac{1}{4} \sum_{i,k,l=1}^n (\partial_l h_{ik})^2 \, u_{(\xi,\varepsilon)}^2 \bigg | \\ &\leq C \, \lambda^{12} \, \mu^3 \, \int_{B_\rho(0)} (\lambda + |x|)^{14-2n} + C \, \lambda^{n-2} \, \int_{\mathbb{R}^n \setminus B_\rho(0)} (\lambda + |x|)^{4-2n} \\ &\leq C \, \lambda^{12} \, \mu^3 + C \, \rho^2 \, \Big ( \frac{\lambda}{\rho} \Big )^{n-2}. \end{aligned}$$ At this point, we use the formula $$\begin{aligned} &\partial_i u_{(\xi,\varepsilon)} \, \partial_k u_{(\xi,\varepsilon)} - \frac{n-2}{4(n-1)} \, \partial_i \partial_k (u_{(\xi,\varepsilon)}^2) \\ &= \frac{1}{n} \, \Big ( |du_{(\xi,\varepsilon)}|^2 - \frac{n-2}{4(n-1)} \, \Delta (u_{(\xi,\varepsilon)}^2) \Big ) \, \delta_{ik}. \end{aligned}$$ Since $h_{ik}$ is trace-free, we obtain $$\sum_{i,k=1}^n h_{ik} \, \partial_i u_{(\xi,\varepsilon)} \, \partial_k u_{(\xi,\varepsilon)} = \frac{n-2}{4(n-1)} \sum_{i,k=1}^n h_{ik} \, \partial_i \partial_k (u_{(\xi,\varepsilon)}^2),$$ hence $$\int_{\mathbb{R}^n} \sum_{i,k=1}^n h_{ik} \, \partial_i u_{(\xi,\varepsilon)} \, \partial_k u_{(\xi,\varepsilon)} = \int_{\mathbb{R}^n} \frac{n-2}{4(n-1)} \sum_{i,k=1}^n \partial_i \partial_k h_{ik} \, u_{(\xi,\varepsilon)}^2.$$ Since $\sum_{i=1}^n \partial_i h_{ik}(x) = 0$ for $|x| \leq \rho$, it follows that $$\bigg | \int_{\mathbb{R}^n} \sum_{i,k=1}^n h_{ik} \, \partial_i u_{(\xi,\varepsilon)} \, \partial_k u_{(\xi,\varepsilon)} \bigg | \leq C \int_{\mathbb{R}^n \setminus B_\rho(0)} u_{(\xi,\varepsilon)}^2 \leq C \, \rho^2 \, \Big ( \frac{\lambda}{\rho} \Big )^{n-2}.$$ Putting these facts together, the assertion follows.\ \[key.estimate\] The function $\mathcal{F}_g(\xi,\varepsilon)$ satisfies the estimate $$\begin{aligned} &\bigg | \mathcal{F}_g(\xi,\varepsilon) - \int_{B_\rho(0)} \frac{1}{2} \, \sum_{i,k,l=1}^n h_{il} \, h_{kl} \, \partial_i u_{(\xi,\varepsilon)} \, \partial_k u_{(\xi,\varepsilon)} \\ &\hspace{10mm} + \int_{B_\rho(0)} \frac{n-2}{16(n-1)} \, \sum_{i,k,l=1}^n (\partial_l h_{ik})^2 \, u_{(\xi,\varepsilon)}^2 \\ &\hspace{10mm} - \int_{\mathbb{R}^n} \sum_{i,k=1}^n \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x) \, \partial_i\partial_k u_{(\xi,\varepsilon)} \, w_{(\xi,\varepsilon)} \bigg | \\ &\leq C \, \lambda^{\frac{8n}{n-2}} \, \mu^{\frac{2n}{n-2}} + C \, \lambda^4 \, \mu \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^{n-2} \end{aligned}$$ for $(\xi,\varepsilon) \in \lambda \, \Omega$. **Proof.** This follows by combining Proposition \[term.1\], Proposition \[term.2\], and Proposition \[term.3\].\ Finding a critical point of an auxiliary function ================================================= We define a function $F: \mathbb{R}^n \times (0,\infty) \to \mathbb{R}$ as follows: given any pair $(\xi,\varepsilon) \in \mathbb{R}^n \times (0,\infty)$, we define $$\begin{aligned} F(\xi,\varepsilon) &= \int_{\mathbb{R}^n} \frac{1}{2} \sum_{i,k,l=1}^n \overline{H}_{il}(x) \, \overline{H}_{kl}(x) \, \partial_i u_{(\xi,\varepsilon)}(x) \, \partial_k u_{(\xi,\varepsilon)}(x) \\ &- \int_{\mathbb{R}^n} \frac{n-2}{16(n-1)} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, u_{(\xi,\varepsilon)}(x)^2 \\ &+ \int_{\mathbb{R}^n} \sum_{i,k=1}^n \overline{H}_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)}(x) \, z_{(\xi,\varepsilon)}(x),\end{aligned}$$ where $z_{(\xi,\varepsilon)} \in \mathcal{E}_{(\xi,\varepsilon)}$ satisfies the relation $$\begin{aligned} &\int_{\mathbb{R}^n} \Big ( \langle dz_{(\xi,\varepsilon)},d\psi \rangle - n(n+2) \, u_{(\xi,\varepsilon)}(x)^{\frac{4}{n-2}} \, z_{(\xi,\varepsilon)} \, \psi \Big ) \\ &= -\int_{\mathbb{R}^n} \sum_{i,k=1}^n \overline{H}_{ik} \, \partial_i \partial_k u_{(\xi,\varepsilon)} \, \psi \end{aligned}$$ for all test functions $\psi \in \mathcal{E}_{(\xi,\varepsilon)}$. Our goal in this section is to show that the function $F(\xi,\varepsilon)$ has a critical point. \[symmetry\] The function $F(\xi,\varepsilon)$ satisfies $F(\xi,\varepsilon) = F(-\xi,\varepsilon)$ for all $(\xi,\varepsilon) \in \mathbb{R}^n \times (0,\infty)$. Consequently, we have $\frac{\partial}{\partial \xi_p} F(0,\varepsilon) = 0$ and $\frac{\partial^2}{\partial \varepsilon \, \partial \xi_p} F(0,\varepsilon) = 0$ for all $\varepsilon > 0$ and $p = 1, \hdots, n$. **Proof.** This follows immediately from the relation $\overline{H}_{ik}(-x) = \overline{H}_{ik}(x)$.\ \[integral.identity.1\] We have $$\begin{aligned} &\int_{\partial B_r(0)} \sum_{i,k,l=1}^n (\partial_l H_{ik}(x))^2 \, x_p \, x_q \\ &= \frac{2}{n(n+2)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \, r^{n+3} \\ &+ \frac{1}{n(n+2)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \, r^{n+3} \end{aligned}$$ and $$\begin{aligned} &\int_{\partial B_r(0)} \sum_{i,k=1}^n H_{ik}(x)^2 \, x_p \, x_q \\ &= \frac{2}{n(n+2)(n+4)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \, r^{n+5} \\ &+ \frac{1}{2n(n+2)(n+4)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \, r^{n+5}. \end{aligned}$$ **Proof.** By definition of $H_{ik}(x)$, we have $$\begin{aligned} &\int_{\partial B_r(0)} \sum_{i,k,l=1}^n (\partial_l H_{ik}(x))^2 \, x_p \, x_q \\ &= \int_{\partial B_r(0)} \sum_{i,j,k,l,m=1}^n (W_{ijkl} + W_{ilkj}) \, (W_{imkl} + W_{ilkm}) \, x_j \, x_m \, x_p \, x_q \\ &= \frac{2}{n(n+2)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \, r^{n+3} \\ &+ \frac{1}{n(n+2)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \, r^{n+3}. \end{aligned}$$ Moreover, it follows from Corollary \[useful.identities\] that $$\begin{aligned} &\int_{\partial B_r(0)} \sum_{i,k=1}^n H_{ik}(x)^2 \, x_p \, x_q \\ &= \int_{\partial B_r(0)} \sum_{i,j,k,l,m,s=1}^n W_{ijkl} \, W_{imks} \, x_j \, x_l \, x_m \, x_s \, x_p \, x_q \\ &= \frac{2}{n(n+2)(n+4)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \, r^{n+5} \\ &+ \frac{1}{2n(n+2)(n+4)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \, r^{n+5}. \end{aligned}$$ This completes the proof.\ \[integral.identity.2\] We have $$\begin{aligned} &\int_{\partial B_r(0)} \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, x_p \, x_q \\ &= \frac{2}{n(n+2)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \\ &\hspace{10mm} \cdot \Big [ r^{n+3} - \frac{2(n+8)}{n+4} \, r^{n+5} + \frac{n+16}{n+4} \, r^{n+7} \Big ] \\ &+ \frac{1}{n(n+2)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \\ &\hspace{10mm} \cdot \Big [ r^{n+3} - \frac{2(n+6)}{n+4} \, r^{n+5} + \frac{n+10}{n+4} \, r^{n+7} \Big ]. \end{aligned}$$ **Proof.** Using the identity $$\partial_l \overline{H}_{ik}(x) = (1 - |x|^2) \, \partial_l H_{ik}(x) - 2 \, H_{ik}(x) \, x_l$$ and Euler’s theorem, we obtain $$\begin{aligned} &\sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \\ &= (1 - |x|^2)^2 \, \sum_{i,k,l=1}^n (\partial_l H_{ik}(x))^2 \\ &- 4 \, (1 - |x|^2) \, \sum_{i,k,l=1}^n H_{ik}(x) \, x_l \, \partial_l H_{ik}(x) + 4 \, |x|^2 \, \sum_{i,k=1}^n H_{ik}(x)^2 \\ &= (1 - |x|^2)^2 \, \sum_{i,k,l=1}^n (\partial_l H_{ik}(x))^2 - 4 \, (2 - 3 \, |x|^2) \, \sum_{i,k=1}^n H_{ik}(x)^2. \end{aligned}$$ Hence, the assertion follows from the previous proposition.\ \[integral.identity.3\] We have $$\begin{aligned} \int_{\partial B_r(0)} \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 &= \frac{1}{n} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \\ &\hspace{10mm} \cdot \Big [ r^{n+1} - \frac{2(n+4)}{n+2} \, r^{n+3} + \frac{n+8}{n+2} \, r^{n+5} \Big ]. \end{aligned}$$ \[formula.for.F\] We have $$\begin{aligned} F(0,\varepsilon) &= -\frac{(n-2)(n+4)}{16n(n-1)(n+2)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \\ &\hspace{10mm} \cdot \Big [ \frac{n-8}{n+4} \, \varepsilon^4 - 2 \, \varepsilon^6 + \frac{n+8}{n-10} \, \varepsilon^8 \Big ] \, \int_0^\infty (1+r^2)^{2-n} \, r^{n+3} \, dr. \end{aligned}$$ **Proof.** Note that $z_{(0,\varepsilon)}(x) = 0$ for all $x \in \mathbb{R}^n$. This implies $$F(0,\varepsilon) = -\int_{\mathbb{R}^n} \frac{n-2}{16(n-1)} \, \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{2-n} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2.$$ Using Corollary \[integral.identity.3\], we obtain $$\begin{aligned} &\int_{\mathbb{R}^n} \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{2-n} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \\ &= \frac{1}{n} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \\ &\hspace{10mm} \cdot \int_0^\infty (1+r^2)^{2-n} \, \Big [ \varepsilon^4 \, r^{n+1} - \frac{2(n+4)}{n+2} \, \varepsilon^6 \, r^{n+3} + \frac{n+8}{n+2} \, \varepsilon^8 \, r^{n+5} \Big ] \, dr. \end{aligned}$$ Moreover, we have $$\int_0^\infty (1+r^2)^{2-n} \, r^{n+1} \, dr = \frac{n-8}{n+2} \int_0^\infty (1+r^2)^{2-n} \, r^{n+3} \, dr$$ and $$\int_0^\infty (1+r^2)^{2-n} \, r^{n+5} \, dr = \frac{n+4}{n-10} \int_0^\infty (1+r^2)^{2-n} \, r^{n+3} \, dr$$ by Proposition \[integration.by.parts\]. From this the assertion follows.\ \[critical.point.of.F\] Assume that $n \geq 52$. Moreover, suppose that $\varepsilon_* > 0$ is defined by $$\label{eps.star} \bigg ( 3 + \sqrt{9 - \frac{8(n+8)(n-8)}{(n+4)(n-10)}} \bigg ) \, \varepsilon_*^2 = \frac{2(n-8)}{n+4}.$$ Then $(0,\varepsilon_*)$ is a critical point of the function $F(\xi,\varepsilon)$. Moreover, we have $\frac{\partial^2}{\partial \varepsilon^2} F(0,\varepsilon_*) > 0$. In the next step, we show that $(0,\varepsilon_*)$ is a strict local minimum of the function $F$. To that end, we compute the Hessian of $F$ at a point $(0,\varepsilon)$.\ \[Hessian.of.F.1\] The second order partial derivatives of the function $F(\xi,\varepsilon)$ are given by $$\begin{aligned} \frac{\partial^2}{\partial \xi_p \, \partial \xi_q} F(0,\varepsilon) &= \int_{\mathbb{R}^n} (n-2)^2 \, \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{-n} \, \sum_{l=1}^n \overline{H}_{pl}(x) \, \overline{H}_{ql}(x) \\ &- \int_{\mathbb{R}^n} \frac{(n-2)^2}{4} \, \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{-n} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, x_p \, x_q \\ &+ \int_{\mathbb{R}^n} \frac{(n-2)^2}{8(n-1)} \, \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{1-n} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, \delta_{pq}. \end{aligned}$$ **Proof.** Using the identity $$\begin{aligned} &\sum_{i,k,l=1}^n \overline{H}_{il}(x) \, \overline{H}_{kl}(x) \, \partial_i u_{(\xi,\varepsilon)}(x) \, \partial_k u_{(\xi,\varepsilon)}(x) \\ &= (n-2)^2 \, \varepsilon^{n-2} \, (\varepsilon^2 + |x - \xi|^2)^{-n} \, \sum_{i,k,l=1}^n \overline{H}_{il}(x) \, \overline{H}_{kl}(x) \, (x_i - \xi_i) \, (x_k - \xi_k) \\ &= (n-2)^2 \, \varepsilon^{n-2} \, (\varepsilon^2 + |x - \xi|^2)^{-n} \, \sum_{i,k,l=1}^n \overline{H}_{il}(x) \, \overline{H}_{kl}(x) \, \xi_i \, \xi_k, \end{aligned}$$ we obtain $$\begin{aligned} &\frac{\partial^2}{\partial \xi_p \, \partial \xi_q} \bigg ( \frac{1}{2} \sum_{i,k,l=1}^n \overline{H}_{il}(x) \, \overline{H}_{kl}(x) \, \partial_i u_{(\xi,\varepsilon)}(x) \, \partial_k u_{(\xi,\varepsilon)}(x) \bigg ) \bigg |_{\xi=0} \\ &= (n-2)^2 \, \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{-n} \, \sum_{l=1}^n \overline{H}_{pl}(x) \, \overline{H}_{ql}(x). \end{aligned}$$ Moreover, we have $$\begin{aligned} &\frac{\partial^2}{\partial \xi_p \, \partial \xi_q} \bigg ( \frac{n-2}{16(n-1)} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, u_{(\xi,\varepsilon)}(x)^2 \bigg ) \bigg |_{\xi=0} \\ &= \frac{(n-2)^2}{4} \, \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{-n} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, x_p \, x_q \\ &- \frac{(n-2)^2}{8(n-1)} \, \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{1-n} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, \delta_{pq}. \end{aligned}$$ Finally, we have $$\begin{aligned} &\sum_{i,k=1}^n \overline{H}_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)}(x) \\ &= n(n-2) \, \varepsilon^{\frac{n-2}{2}} \, (\varepsilon^2 + |x - \xi|^2)^{-\frac{n+2}{2}} \, \sum_{i,k=1}^n \overline{H}_{ik}(x) \, (x_i - \xi_i) \, (x_k - \xi_k) \\ &= n(n-2) \, \varepsilon^{\frac{n-2}{2}} \, (\varepsilon^2 + |x - \xi|^2)^{-\frac{n+2}{2}} \, \sum_{i,k=1}^n \overline{H}_{ik}(x) \, \xi_i \, \xi_k \end{aligned}$$ since $\overline{H}_{ik}(x)$ is trace-free. Thus, we conclude that $$\begin{aligned} &\frac{\partial^2}{\partial \xi_p \, \partial \xi_q} \bigg ( \sum_{i,k=1}^n \overline{H}_{ik}(x) \, \partial_i \partial_k u_{(\xi,\varepsilon)}(x) \, z_{(\xi,\varepsilon)}(x) \bigg ) \bigg |_{\xi=0} \\ &= 2n(n-2) \, \varepsilon^{\frac{n-2}{2}} \, (\varepsilon^2 + |x|^2)^{-\frac{n+2}{2}} \, \sum_{i,k=1}^n \overline{H}_{pq}(x) \, z_{(0,\varepsilon)}(x) = 0. \end{aligned}$$ From this the assertion follows.\ \[Hessian.of.F.2\] The second order partial derivatives of the function $F(\xi,\varepsilon)$ are given by $$\begin{aligned} &\frac{\partial^2}{\partial \xi_p \, \partial \xi_q} F(0,\varepsilon) \\ &= \frac{4 (n-2)^2}{n(n+2)(n+4)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \\ &\hspace{10mm} \cdot \Big [ \varepsilon^4 - \frac{3(n+6)}{2(n-8)} \, \varepsilon^6 \Big ] \, \int_0^\infty (1 + r^2)^{-n} \, r^{n+5} \, dr \\ &+ \frac{(n-2)^2}{n(n+2)(n+4)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \\ &\hspace{10mm} \cdot \Big [ \varepsilon^4 - \frac{n+7}{n-8} \, \varepsilon^6 \Big ] \, \int_0^\infty (1 + r^2)^{-n} \, r^{n+5} \, dr. \end{aligned}$$ **Proof.** Using the identity $$\begin{aligned} &\int_{\partial B_r(0)} \sum_{l=1}^n \overline{H}_{pl}(x) \, \overline{H}_{ql}(x) \\ &= \int_{\partial B_r(0)} \sum_{i,j,k,l,m=1}^n W_{ipkl} \, W_{jqml} \, x_i \, x_j \, x_k \, x_m \, (1 - |x|^2)^2 \\ &= \frac{1}{n(n+2)} \, |S^{n-1}| \\ &\hspace{10mm} \cdot \sum_{i,j,k,l,m=1}^n W_{ipkl} \, W_{jqml} \, (\delta_{ij} \, \delta_{km} + \delta_{ik} \, \delta_{jm} + \delta_{im} \, \delta_{jk}) \, r^{n+3} \, (1 - r^2)^2 \\ &= \frac{1}{2n(n+2)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \, r^{n+3} \, (1 - r^2)^2, \end{aligned}$$ we obtain $$\begin{aligned} &\int_{\mathbb{R}^n} \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{-n} \, \sum_{i,k,l=1}^n \overline{H}_{pl}(x) \, \overline{H}_{ql}(x) \\ &= \frac{1}{2n(n+2)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \\ &\hspace{10mm} \cdot \int_0^\infty (1 + r^2)^{-n} \, \Big [ \varepsilon^2 \, r^{n+3} - 2 \, \varepsilon^4 \, r^{n+5} + \varepsilon^6 \, r^{n+7} \Big ] \, dr. \end{aligned}$$ Similarly, it follows from Proposition \[integral.identity.2\] that $$\begin{aligned} &\int_{\mathbb{R}^n} \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{-n} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, x_p \, x_q \\ &= \frac{2}{n(n+2)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \\ &\hspace{10mm} \cdot \int_0^\infty (1 + r^2)^{-n} \, \Big [ \varepsilon^2 \, r^{n+3} - \frac{2(n+8)}{n+4} \, \varepsilon^4 \, r^{n+5} + \frac{n+16}{n+4} \, \varepsilon^6 \, r^{n+7} \Big ] \, dr \\ &+ \frac{1}{n(n+2)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \\ &\hspace{10mm} \cdot \int_0^\infty (1 + r^2)^{-n} \, \Big [ \varepsilon^2 \, r^{n+3} - \frac{2(n+6)}{n+4} \, \varepsilon^4 \, r^{n+5} + \frac{n+10}{n+4} \, \varepsilon^6 \, r^{n+7} \Big ] \, dr. \end{aligned}$$ Moreover, we have $$\begin{aligned} &\int_{\mathbb{R}^n} \varepsilon^{n-2} \, (\varepsilon^2 + |x|^2)^{1-n} \, \sum_{i,k,l=1}^n (\partial_l \overline{H}_{ik}(x))^2 \, \delta_{pq} \\ &= \frac{1}{n} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \\ &\hspace{10mm} \cdot \int_0^\infty (1 + r^2)^{1-n} \, \Big [ \varepsilon^2 \, r^{n+1} - \frac{2(n+4)}{n+2} \, \varepsilon^4 \, r^{n+3} + \frac{n+8}{n+2} \, \varepsilon^6 \, r^{n+5} \Big ] \, dr. \end{aligned}$$ by Corollary \[integral.identity.3\]. Using Proposition \[Hessian.of.F.1\] and the identity $$\int_0^\infty (1 + r^2)^{1-n} \, r^{n+1} \, dr = \frac{2(n-1)}{n+2} \int_0^\infty (1 + r^2)^{-n} \, r^{n+3} \, dr,$$ we obtain $$\begin{aligned} &\frac{\partial^2}{\partial \xi_p \, \partial \xi_q} F(0,\varepsilon) \\ &= \frac{4 (n-2)^2}{n(n+2)(n+4)} \, |S^{n-1}| \, \sum_{i,k,l=1}^n (W_{ipkl} + W_{ilkp}) \, (W_{iqkl} + W_{ilkq}) \\ &\hspace{10mm} \cdot \int_0^\infty (1 + r^2)^{-n} \, \Big [ \varepsilon^4 \, r^{n+5} - \frac{3}{2} \, \varepsilon^6 \, r^{n+7} \Big ] \, dr \\ &+ \frac{(n-2)^2}{4n(n+2)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \\ &\hspace{10mm} \cdot \int_0^\infty (1 + r^2)^{-n} \, \Big [ \frac{2(n+6)}{n+4} \, \varepsilon^4 \, r^{n+5} - \frac{n+10}{n+4} \, \varepsilon^6 \, r^{n+7} \Big ] \, dr \\ &- \frac{(n-2)^2}{8n(n-1)} \, |S^{n-1}| \, \sum_{i,j,k,l=1}^n (W_{ijkl} + W_{ilkj})^2 \, \delta_{pq} \\ &\hspace{10mm} \cdot \int_0^\infty (1 + r^2)^{1-n} \, \Big [ \frac{2(n+4)}{n+2} \, \varepsilon^4 \, r^{n+3} - \frac{n+8}{n+2} \, \varepsilon^6 \, r^{n+5} \Big ] \, dr.\end{aligned}$$ Hence, the assertion follows from the identities $$\begin{aligned} &\int_0^\infty (1 + r^2)^{-n} \, r^{n+7} \, dr = \frac{n+6}{n-8} \int_0^\infty (1 + r^2)^{-n} \, r^{n+5} \, dr \\ &\int_0^\infty (1 + r^2)^{1-n} \, r^{n+3} \, dr = \frac{2(n-1)}{n+4} \int_0^\infty (1 + r^2)^{-n} \, r^{n+5} \, dr \\ &\int_0^\infty (1 + r^2)^{1-n} \, r^{n+5} \, dr = \frac{2(n-1)}{n-8} \int_0^\infty (1 + r^2)^{-n} \, r^{n+5} \, dr. \end{aligned}$$ \[strict.local.minimum\] Assume that $n \geq 52$ and $\varepsilon_* > 0$ is defined by (\[eps.star\]). Then the function $F(\xi,\varepsilon)$ has a strict local minimum at the point $(0,\varepsilon_*)$. **Proof.** It follows from Corollary \[critical.point.of.F\] that $(0,\varepsilon_*)$ is a critical point of the function $F(\xi,\varepsilon)$. Moreover, we have $\frac{\partial^2}{\partial \varepsilon^2} F(0,\varepsilon_*) > 0$. Since $n \geq 52$, we have $$\frac{6}{n+4} < \sqrt{9 - \frac{8(n+8)(n-8)}{(n+4)(n-10)}}.$$ This implies $$\frac{3(n+6)}{n+4} \, \varepsilon_*^2 < \bigg ( 3 + \sqrt{9 - \frac{8(n+8)(n-8)}{(n+4)(n-10)}} \bigg ) \, \varepsilon_*^2 = \frac{2(n-8)}{n+4}.$$ Thus, we conclude that $$\frac{n+7}{n-8} \, \varepsilon_*^2 < \frac{3(n+6)}{2(n-8)} \, \varepsilon_*^2 < 1.$$ Hence, it follows from Proposition \[Hessian.of.F.2\] that the matrix $\frac{\partial^2}{\partial \xi_p \, \partial \xi_q} F(0,\varepsilon_*)$ is positive definite. This proves the assertion.\ Proof of the main theorem ========================= \[perturbation.argument\] Assume that $n \geq 52$. Moreover, let $g$ be a smooth metric on $\mathbb{R}^n$ of the form $g(x) = \exp(h(x))$, where $h(x)$ is a trace-free symmetric two-tensor on $\mathbb{R}^n$ such that $|h(x)| + |\partial h(x)| + |\partial^2 h(x)| \leq \alpha \leq \alpha_1$ for all $x \in \mathbb{R}^n$, $h(x) = 0$ for $|x| \geq 1$, and $h_{ik}(x) = \mu \, (\lambda^2 - |x|^2) \, H_{ik}(x)$ for $|x| \leq \rho$. As above, we assume that $\lambda \leq \rho \leq 1$ and $\mu \leq 1$. If $\alpha$ and $\rho^{2-n} \, \mu^{-2} \, \lambda^{n-10}$ are sufficiently small, then there exists a positive function $v$ such that $$\Delta_g v - \frac{n-2}{4(n-1)} \, R_g \, v + n(n-2) \, v^{\frac{n+2}{n-2}} = 0,$$ $$\int_{\mathbb{R}^n} v^{\frac{2n}{n-2}} < \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}},$$ and $\sup_{|x| \leq \lambda} v(x) \geq c \, \lambda^{\frac{2-n}{2}}$. Here, $c$ is a positive constant that depends only on $n$. **Proof.** By Corollary \[strict.local.minimum\], the function $F(\xi,\varepsilon)$ has a strict local minimum at $(0,\varepsilon_*)$. Hence, we can find an open set $\Omega' \subset \Omega$ such that $(0,\varepsilon_*) \in \Omega'$ and $$F(0,\varepsilon_*) < \inf_{(\xi,\varepsilon) \in \partial \Omega'} F(\xi,\varepsilon) < 0.$$ Using Corollary \[key.estimate\], we obtain $$\begin{aligned} &|\mathcal{F}_g(\lambda\xi,\lambda\varepsilon) - \lambda^8 \, \mu^2 \, F(\xi,\varepsilon)| \\ &\leq C \, \lambda^{\frac{8n}{n-2}} \, \mu^{\frac{2n}{n-2}} + C \, \lambda^4 \, \mu \, \Big ( \frac{\lambda}{\rho} \Big )^{\frac{n-2}{2}} + C \, \Big ( \frac{\lambda}{\rho} \Big )^{n-2} \end{aligned}$$ for all $(\xi,\varepsilon) \in \Omega$. This implies $$\begin{aligned} &|\lambda^{-8} \, \mu^{-2} \, \mathcal{F}_g(\lambda\xi,\lambda\varepsilon) - F(\xi,\varepsilon)| \\ &\leq C \, \lambda^{\frac{16}{n-2}} \, \mu^{\frac{4}{n-2}} + C \, \rho^{\frac{2-n}{2}} \, \mu^{-1} \, \lambda^{\frac{n-10}{2}} + C \, \rho^{2-n} \, \mu^{-2} \, \lambda^{n-10} \end{aligned}$$ for all $(\xi,\varepsilon) \in \Omega$. Hence, if $\rho^{2-n} \, \mu^{-2} \, \lambda^{n-10}$ is sufficiently small, then we have $$\mathcal{F}_g(0,\lambda\varepsilon_*) < \inf_{(\xi,\varepsilon) \in \partial \Omega'} \mathcal{F}_g(\lambda\xi,\lambda\varepsilon) < 0.$$ Consequently, there exists a point $(\bar{\xi},\bar{\varepsilon}) \in \Omega'$ such that $$\mathcal{F}_g(\lambda\bar{\xi},\lambda\bar{\varepsilon}) = \inf_{(\xi,\varepsilon) \in \Omega'} \mathcal{F}_g(\lambda\xi,\lambda\varepsilon) < 0.$$ By Proposition \[reduction.to.a.finite.dimensional.problem\], the function $v = v_{(\lambda\bar{\xi},\lambda\bar{\varepsilon})}$ is a non-negative weak solution of the partial differential equation $$\Delta_g v - \frac{n-2}{4(n-1)} \, R_g \, v + n(n-2) \, v^{\frac{n+2}{n-2}} = 0.$$ Using a result of N. Trudinger, we conclude that $v$ is smooth (see [@Trudinger], Theorem 3 on p. 271). Moreover, we have $$\begin{aligned} 2(n-2) \int_{\mathbb{R}^n} v^{\frac{2n}{n-2}} &= 2(n-2) \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}} + \mathcal{F}_g(\lambda\bar{\xi},\lambda\bar{\varepsilon}) \\ &< 2(n-2) \, \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}}.\end{aligned}$$ Finally, it follows from Proposition \[fixed.point.argument\] that $\|v - u_{(\lambda\bar{\xi},\lambda\bar{\varepsilon})}\|_{L^{\frac{2n}{n-2}}(\mathbb{R}^n)} \leq C \, \alpha$. This implies $$|B_\lambda(0)|^{\frac{n-2}{2n}} \, \sup_{|x| \leq \lambda} v(x) \geq \|v\|_{L^{\frac{2n}{n-2}}(B_\lambda(0))} \geq \|u_{(\lambda\bar{\xi},\lambda\bar{\varepsilon})}\|_{L^{\frac{2n}{n-2}}(B_\lambda(0))} - C \, \alpha.$$ Hence, if $\alpha$ is sufficiently small, then we obtain $\lambda^{\frac{n-2}{2}} \, \sup_{|x| \leq \lambda} v(x) \geq c$.\ Let $n \geq 52$. Then there exists a smooth metric $g$ on $\mathbb{R}^n$ with the following properties: - $g_{ik}(x) = \delta_{ik}$ for $|x| \geq \frac{1}{2}$ - $g$ is not conformally flat - There exists a sequence of non-negative smooth functions $v_\nu$ ($\nu \in \mathbb{N}$) such that $$\Delta_g v_\nu - \frac{n-2}{4(n-1)} \, R_g \, v_\nu + n(n-2) \, v_\nu^{\frac{n+2}{n-2}} = 0$$ for all $\nu \in \mathbb{N}$, $$\int_{\mathbb{R}^n} v_\nu^{\frac{2n}{n-2}} < \Big ( \frac{Y(S^n)}{4n(n-1)} \Big )^{\frac{n}{2}}$$ for all $\nu \in \mathbb{N}$, and $\sup_{|x| \leq 1} v_\nu(x) \to \infty$ as $\nu \to \infty$. **Proof.** Choose a smooth cutoff function $\eta: \mathbb{R} \to \mathbb{R}$ such that $\eta(t) = 1$ for $t \leq 1$ and $\eta(t) = 0$ for $t \geq 2$. We define a trace-free symmetric two-tensor on $\mathbb{R}^n$ by $$h_{ik}(x) = \sum_{N=N_0}^\infty \eta(4N^2 \, |x - y_N|) \, 2^{-N} \, (2^{-N} - |x - y_N|^2) \, H_{ik}(x - y_N),$$ where $y_N = (\frac{1}{N},0,\hdots,0) \in \mathbb{R}^n$. It is straightforward to verify that $h(x)$ is $C^\infty$ smooth. Let $\alpha$ be the constant appearing in Proposition \[perturbation.argument\]. If $N_0$ is sufficiently large, then we have $|h(x)| + |\partial h(x)| + |\partial^2 h(x)| \leq \alpha$ for all $x \in \mathbb{R}^n$ and $h(x) = 0$ for $|x| \geq \frac{1}{2}$. Moreover, we have $h_{ik}(x) = 2^{-N} \, (2^{-N} - |x - y_N|^2) \, H_{ik}(x - y_N)$ provided that $N \geq N_0$ and $|x - y_N| \leq \frac{1}{4N^2}$. Hence, we can apply Proposition \[perturbation.argument\] with $\lambda = 2^{-N/2}$, $\mu = 2^{-N}$, and $\rho = \frac{1}{4N^2}$. From this the assertion follows.\ An asymptotic expansion for the scalar curvature ================================================ Suppose that $h(x)$ is a trace-free symmetric two-tensor defined on $\mathbb{R}^n$ satisfying $|h(x)| \leq 1$ for all $x \in \mathbb{R}^n$. We define a Riemannian metric $g$ on $\mathbb{R}^n$ by $g(x) = \exp(h(x))$. In this section, we derive an approximate formula for the scalar curvature of this metric. A similar formula is derived in [@Ambrosetti-Malchiodi].\ \[Taylor.expansion.of.scalar.curvature\] Let $R_g$ be the scalar curvature of $g$. There exists a constant $C$, depending only on $n$, such that $$\begin{aligned} &\Big | R_g - \partial_i \partial_k h_{ik} + \partial_i(h_{il} \, \partial_k h_{kl}) - \frac{1}{2} \, \partial_i h_{il} \, \partial_k h_{kl} + \frac{1}{4} \, \partial_l h_{ik} \, \partial_l h_{ik} \Big | \\ &\leq C \, |h|^2 \, |\partial^2 h| + C \, |h| \, |\partial h|^2.\end{aligned}$$ **Proof.** The Riemann curvature tensor is defined as $$\partial_i \Gamma_{jk}^m - \partial_j \Gamma_{ik}^m + \Gamma_{jk}^l \, \Gamma_{il}^m - \Gamma_{ik}^l \, \Gamma_{jl}^m.$$ Hence, the scalar curvature of $g$ is given by $$R_g = g^{jk} \, (\partial_i \Gamma_{jk}^i - \partial_j \Gamma_{ik}^i + \Gamma_{jk}^l \, \Gamma_{il}^i - \Gamma_{ik}^l \, \Gamma_{jl}^i).$$ Since $h$ is trace-free, we have $\det g(x) = 1$ for all $x \in \mathbb{R}^n$. This implies $\Gamma_{ik}^i = \frac{1}{2} \, g^{il} \, \partial_k g_{il} = \frac{1}{2} \, \partial_k \log \det g= 0$. Therefore, we obtain $$\begin{aligned} R_g &= g^{jk} \, \partial_i \Gamma_{jk}^i - g^{jk} \, \Gamma_{ik}^l \, \Gamma_{jl}^i \\ &= \partial_i(g^{jk} \, \Gamma_{jk}^i) + g^{jk} \, \Gamma_{ik}^l \, \Gamma_{jl}^i. \end{aligned}$$ Note that $$g^{jk} \, \Gamma_{jk}^i = g^{il} \, g^{jk} \, \partial_k g_{jl}.$$ From this it follows that $$\begin{aligned} &\Big | \partial_i(g^{jk} \, \Gamma_{jk}^i) - \partial_i \partial_k h_{ik} + \frac{1}{2} \, \partial_i(h_{il} \, \partial_k h_{kl}) + \frac{1}{2} \, \partial_i(h_{kl} \, \partial_k h_{il}) \Big | \\ &\leq C \, |h|^2 \, |\partial^2 h| + C \, |h| \, |\partial h|^2, \end{aligned}$$ hence $$\begin{aligned} &\Big | \partial_i(g^{jk} \, \Gamma_{jk}^i) - \partial_i \partial_k h_{ik} + \partial_i(h_{il} \, \partial_k h_{kl}) - \frac{1}{2} \, \partial_i h_{il} \, \partial_k h_{kl} + \frac{1}{2} \, \partial_i h_{kl} \, \partial_k h_{il} \Big | \\ &\leq C \, |h|^2 \, |\partial^2 h| + C \, |h| \, |\partial h|^2, \end{aligned}$$ Moreover, we have $$\Big | g^{jk} \, \Gamma_{ik}^l \, \Gamma_{jl}^i + \frac{1}{4} \, \partial_l h_{ik} \, \partial_l h_{ik} - \frac{1}{2} \, \partial_i h_{kl} \, \partial_k h_{il} \Big | \leq C \, |h| \, |\partial h|^2.$$ Putting these facts together, we obtain $$\begin{aligned} &\Big | R_g - \partial_i \partial_k h_{ik} + \partial_i(h_{il} \, \partial_k h_{kl}) - \frac{1}{2} \, \partial_i h_{il} \, \partial_k h_{kl} + \frac{1}{4} \, \partial_l h_{ik} \, \partial_l h_{ik} \Big | \\ &\leq C \, |h|^2 \, |\partial^2 h| + C \, |h| \, |\partial h|^2.\end{aligned}$$ This completes the proof.\ Some useful identities ====================== \[integration.by.parts\] Suppose that $\alpha$ and $\beta$ are real numbers satisfying $2\alpha - 2 > \beta + 1 > 0$. Then $$\int_0^\infty (1+r^2)^{1-\alpha} \, r^\beta \, dr = \frac{2\alpha-2}{2\alpha-\beta-3} \int_0^\infty (1+r^2)^{-\alpha} \, r^\beta \, dr$$ and $$\int_0^\infty (1+r^2)^{-\alpha} \, r^{\beta+2} \, dr = \frac{\beta+1}{2\alpha-\beta-3} \int_0^\infty (1+r^2)^{-\alpha} \, r^\beta \, dr.$$ **Proof.** Using the fundamental theorem of calculus, we obtain $$\begin{aligned} 0 &= \int_0^\infty \frac{d}{dr} \big [ (1+r^2)^{1-\alpha} \, r^{\beta+1} \big ] \, dr \\ &= (\beta+1) \int_0^\infty (1+r^2)^{1-\alpha} \, r^\beta \, dr - (2\alpha-2) \int_0^\infty (1+r^2)^{-\alpha} \, r^{\beta+2} \, dr. \end{aligned}$$ From this the assertion follows.\ Suppose that $p(x)$ is a homogenous polynomial of degree $d$. Then $$\int_{\partial B_1(0)} p(x) = \frac{1}{d(n+d-2)} \int_{\partial B_1(0)} \Delta p(x).$$ **Proof.** Using the divergence theorem, we obtain $$\begin{aligned} \int_{\partial B_1(0)} \Delta p(x) &= (n+d-2) \int_{B_1(0)} \Delta p(x) \\ &= (n+d-2) \int_{\partial B_1(0)} \sum_{k=1}^n x_k \, \partial_k p(x) \\ &= d(n+d-2) \int_{\partial B_1(0)} p(x). \end{aligned}$$ \[useful.identities\] We have $$\int_{\partial B_1(0)} x_i \, x_j = \frac{1}{n} \, |S^{n-1}| \, \delta_{ij},$$ $$\int_{\partial B_1(0)} x_i \, x_j \, x_k \, x_l = \frac{1}{n(n+2)} \, |S^{n-1}| \, (\delta_{ij} \, \delta_{kl} + \delta_{ik} \, \delta_{jl} + \delta_{il} \, \delta_{jk}),$$ and $$\begin{aligned} &\int_{\partial B_1(0)} x_i \, x_j \, x_k \, x_l \, x_p \, x_q \\ &= \frac{1}{n(n+2)(n+4)} \, |S^{n-1}| \, (\delta_{ij} \, \delta_{kl} \, \delta_{pq} + \delta_{ij} \, \delta_{kp} \, \delta_{lq} + \delta_{ij} \, \delta_{kq} \, \delta_{lp} \\ &\hspace{45mm} + \delta_{ik} \, \delta_{jl} \, \delta_{pq} + \delta_{ik} \, \delta_{jp} \, \delta_{lq} + \delta_{ik} \, \delta_{jq} \, \delta_{lp} \\ &\hspace{45mm} + \delta_{il} \, \delta_{jk} \, \delta_{pq} + \delta_{il} \, \delta_{jp} \, \delta_{kq} + \delta_{il} \, \delta_{jq} \, \delta_{kp} \\ &\hspace{45mm} + \delta_{ip} \, \delta_{jk} \, \delta_{lq} + \delta_{ip} \, \delta_{jl} \, \delta_{kq} + \delta_{ip} \, \delta_{jq} \, \delta_{kl} \\ &\hspace{45mm} + \delta_{iq} \, \delta_{jk} \, \delta_{lp} + \delta_{iq} \, \delta_{jl} \, \delta_{kp} + \delta_{iq} \, \delta_{jp} \, \delta_{kl}).\end{aligned}$$ [99]{} A. Ambrosetti, *Multiplicity results for the Yamabe problem on $S^n$,* Proc. Natl. Acad. Sci. USA 99 (2002), 15252–15256 A. Ambrosetti and A. Malchiodi, *A multiplicity result for the Yamabe problem on $S^n$,* J. Funct. Anal. 168, 529–561 (1999) T. Aubin, *Équations différentielles non linéaires et problème de Yamabe concernant la courbure scalaire,* J. Math. Pures Appl. 55, 269–296 (1976) T. Aubin, *Sur quelques problèmes de courbure scalaire,* J. Funct. Anal. 240, 269–289 (2006) T. Aubin, *Solution complète de la $C^0$ compacité de l’ensemble des solutions de l’équation de Yamabe,* J. Funct. Anal. 244, 579–589 (2007) M. Berti and A. Malchiodi, *Non-compactness and multiplicity results for the Yamabe problem on $S^n$,* J. Funct. Anal. 180, 210–241 (2001) O. Druet, *Compactness for Yamabe metrics in low dimensions,* Internat. Math. Res. Notices 23, 1143–1191 (2004) O. Druet and E. Hebey, *Blow-up examples for second order elliptic PDEs of critical Sobolev growth,* Trans. Amer. Math. Soc. 357, 1915–1929 (2004) O. Druet and E. Hebey, *Elliptic equations of Yamabe type,* International Mathematics Research Surveys 1, 1–113 (2005) M. Khuri, F. Marques, and R. Schoen, *A compactness theorem for the Yamabe problem,* preprint (2007) Y.Y. Li and L. Zhang, *Compactness of solutions to the Yamabe problem II,* Calc. Var. PDE 24, 185–237 (2005) Y.Y. Li and L. Zhu, *Yamabe type equations on three-dimensional Riemannian manifolds,* Commun. Contemp. Math. 1, 1–50 (1999) F.C. Marques, *A-priori estimates for the Yamabe problem in the non-locally conformally flat case,* J. Diff. Geom. 71, 315–346 (2005) D. Pollack, *Nonuniqueness and high energy solutions for a conformally invariant scalar equation,* Comm. Anal. Geom. 1, 347–414 (1993) O. Rey, *The role of the Green’s function in a non-linear elliptic equation involving the critical Sobolev exponent,* J. Funct. Anal. 89, 1–52 (1990) R.M. Schoen, *Conformal deformation of a Riemannian metric to constant scalar curvature,* J. Diff. Geom. 20, 479–495 (1984) R.M. Schoen, *Variational theory for the total scalar curvature functional for Riemannian metrics and related topics,* Topics in the calculus of variations (ed. by Mariano Giaquinta), Lecture Notes in Mathematics, vol. 1365, Springer Verlag 1989, 120–154 R.M. Schoen, *On the number of constant scalar curvature metrics in a conformal class,* Differential geometry (ed. by H. Blaine Lawson, Jr., and Keti Tenenblat), Pitman Monographs and Surveys in Pure and Applied Mathematics, vol. 52, Longman Scientific & Technical 1991, 311–320 R.M. Schoen, *A report on some recent progress on nonlinear problems in geometry,* In: Surveys in differential geometry, Lehigh University, Bethlehem, PA, 1991, 201–241 N. Trudinger, *Remarks concerning the conformal deformation of Riemannian structures on compact manifolds,* Annali Scuola Norm. Sup. Pisa 22, 265–274 (1968) [^1]: T. Aubin has recently claimed a general compactness theorem in all dimensions [@Aubin2],[@Aubin3]. We have, however, been unable to verify some of the arguments in [@Aubin2].

ArXiv

--- abstract: 'The kinetic theory for a fluid of hard spheres which undergo endothermic and/or exothermic reactions with mass transfer is developed. The exact balance equations for concentration, density, velocity and temperature are derived. The Enskog approximation is discussed and used as the basis for the derivation, via the Chapman-Enskog procedure, of the Navier-Stokes-reaction equations under various assumptions about the speed of the chemical reactions. It is shown that the phenomenological description consisting of a reaction-diffusion equation with a convective coupling to the Navier-Stokes equations is of limited applicability.' author: - 'James F. Lutsko' bibliography: - 'physics.bib' title: 'Kinetic Theory and Hydrodynamics of Dense, Reacting Fluids far from Equilibrium' --- Introduction {#introduction .unnumbered} ============ An understanding of chemically reactive flows is necessary in a wide range of disciplines including astrophysics, plasma physics and the chemical industry. Recently, applications in aerospace engineering have lead to a number of studies aimed at deriving phenomenological equations, the Navier-Stokes equations coupled to reactions, from the Boltzmann equation for increasingly complex systems including internal degrees of freedom and three-body interactions[@Chikhaoui94; @Chikhaoui00]. However, these investigations do not exhaust the range of interesting applications. A number of important applications arise in the physio-chemistry of cavitating bubbles[@Bubbles]. Aside from obvious examples such as the study of flames and explosions, a recent area of interest is sonochemistry in which ultrasound is used to induce conditions of extreme temperature and pressure inside bubbles with the effect of dramatically increasing the rates of chemical reactions[@SomnoChem]. Closely related is the phenomena of somnoluminescence - in which a fluid irradiated with ultrasound emits light - is believed to be caused by pressure waves acting on small bubbles of gas in the fluid(see, e.g. [@SomnoReview]). The bubbles are subjected to such rapid compression that shockwaves may develop and the concentrated energy drives many chemical reactions particularly when the shocks reach the center of the bubbles giving rise to high temperatures and densities. In fact, it has been suggested[@Didenko2002] that some (endothermic) reactions may play an important role in limiting the temperatures reached in the center of the bubble. It is therefore of interest for these applications, as well as some of the others mentioned above, to understand the phenomenological equations governing a reacting gas under extreme conditions and far from equilibrium. The Boltzmann equation cannot be considered an adequate basis for such a study due to the fact that it is only applicable at asymptotically low densities. In fact, the only simple fluid that is amenable to analytic investigation at finite densities is one composed of hard spheres. The purpose of this paper is therefore to review the kinetic theory of reacting hard-sphere systems and to use this as a basis for a hydrodynamic description of a reacting fluid far from equilibrium. In particular, the kinetic theory will be used to derive the exact balance equations describing the local concentration, density, velocity and temperature fields from which the extension of the Navier-Stokes equations to include reactions is developed based on the Chapman-Enskog procedure applied to the Enskog approximation to the kinetic theory. A primary result will be to show that the usual phenomenological description consisting of the Navier-Stokes equations coupled to a reaction-diffusion-advection equation is only applicable if the chemical reactions take place on a time scale which is comparable to the dissipative time scale $\lambda k^{2}$, where $\lambda $ is a transport coefficient and $% k$ a typical wavevector. If the reactions are slower, then all hydrodynamic relaxation takes place before the chemistry gets started and chemistry and hydrodynamics are effectively decoupled. Faster reactions, with time-scale comparable to $ck$, where $c$ is the sound velocity, leads to additional couplings of the reactions to the hydrodynamic fields. Even faster reactions lead to the chemistry taking place so fast that hydrodynamics is irrelevant. The hard-sphere interaction model has proven remarkably useful as models of single- and multi-component simple fluids since, in many respects, the phenomenology of the hard-sphere systems and atoms interacting via more realistic pair potentials is qualitatively identical. For example, hard-sphere systems exhibit the full range of transport coefficients found in all simple fluids[@McLennan], possess a freezing transition[Evans]{} and the structure of hard-sphere fluids in equilibrium is not much different from that of any other fluid[@HansenMcdonald]. On the theoretical side, the equation of state of hard-sphere fluids is easily modeled[@HansenMcdonald], kinetic theory is simplified by the fact that only binary collisions are important and it is possible to formulate an extension of the Boltzmann equation - the so-called Revised Enskog Theory or RET[@RET]- which not only describes the transport properties of multi-component hard-sphere fluids at finite densities but which also describes transport in the solid state[@KDEP]. More recently, inelastic hard spheres have been used as a model for driven granular fluids with similar success. The hard-sphere interaction is therefore an ideal model for understanding the extreme conditions occuring in sonochemical experiments. The kinetic theory of chemically reacting hard spheres has in fact been discussed in the literature[@Kapral78; @Bose; @Cukier; @KapralAdvChem]. The principal aim of these studies was to investigate contributions to the reaction rates coming from dense fluid effects (e.g., ring kinetic theory leading to mode-coupling models) at equilibrium. In these studies, the atoms carry labels indicating their species (sometimes called their color) and all intrinsic properties like the atomic radius and mass is specific to the species. When the atoms collide, there is a probability that a reaction takes place in which the species labels, and hence atomic properties, change. The probability typically depends on the rest-frame energy of the colliding atoms: if the rest frame energy is greater than some specified activation energy, the reaction can take place with a probability that is, in general, a function of the relative energy. Energy may be gained or lost (exo- or endothermic reactions) but the sizes of the atoms are generally invariant since, were they to also vary, a collision could result in one of the atoms overlapping with third atom. (Technically, there is no reason that atoms could not get *smaller* and most results would apply to such a model). Besides being restricted to a chemistry consisting of color labels (and so excluding, e.g., the exchange of mass upon collision), a common assumption in earlier work is that the chemical reactions are *slow* compared to other transport processes. Since the rate of chemical reactions is generally determined by the ratio of the temperature to the activation energies and the difference in the concentrations, this implies that the results are only applicable near equilibrium or for low temperatures. One of the primary goals of the present work is to indicate how the phenomenological description (Navier-Stokes equations coupled to a set of advective-reaction-diffusion equations) must be modified to account for large deviations from equilibrium. The organization of this paper is as follows. The second Section develops the formal statistical description of the system. The discussion of possible collision rules is followed by the development of the Liouville equation and the exact balance laws describing the evolution of the local mass, energy, momentum and partial densities. The Enskog approximation is also introduced. The third Section discusses the Chapman-Enskog solution of the Enskog equation and, particularly, the difference between the assumptions of fast and slow chemical reactions. It is shown that, if the chemical reactions are sufficiently slow, the fluid may be described by the Navier-Stokes equations for the total mass, energy and momentum densities and a reaction-diffusion-convection equation for the concentrations with the only coupling between the two being the convective term occurring in the latter (i.e., the ”usual” description). However, under less restrictive assumptions, the reactions are shown to depend in a much more complicated way on the hydrodynamic fields. The paper concludes with a discussion of the physical meaning of the different assumptions. Statistical Mechanics of Reacting Hard-Spheres ============================================== Consider a system of $N$ hard spheres of various species confined to a volume $V$ with positions $\left\{ \overrightarrow{q}_{i}\right\} _{i=1}^{N}$ and momenta $\left\{ \overrightarrow{p}_{i}\right\} _{i=1}^{N}$. Each atom will also be described by a set of discrete labels, $l_{i}$ for the $i$-th atom, which fix its ”chemical” properties (mass, hard-sphere diameter and reaction parameters). When two atoms collide, both the mechanical variables and the species labels of the atoms change. The dynamical variables are altered according to some deterministic collision rule so that for collisions between the $i$-th and $j-$th atoms, $$\begin{aligned} x_{i} &\rightarrow &x_{i}^{\prime }=\widehat{b}_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}x_{i} \\ x_{j} &\rightarrow &x_{j}^{\prime }=\widehat{b}_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}x_{j}, \notag\end{aligned}$$where the collision operator $\widehat{b}_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}$ describes a collision involving the reaction $% l_{i}+l_{j}\rightarrow l_{i}^{\prime }+l_{j}^{\prime }$ (i.e., the $i$-th atom changes from species $l_{i}$ to $l_{i}^{\prime }$, etc.) and clearly, one expects that $\widehat{b}_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}=% \widehat{b}_{l_{j}l_{i}}^{l_{j}^{\prime }l_{i}^{\prime }}$. Because the species change instantaneously and randomly upon collision, the species labels must be viewed as discrete random variables. Attention here will be restricted to the model in which the probability of making a particular transition is given by some function of the relative phases of the two colliding atoms $K_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}\left( x_{ij}\right) $ where the notation indicates that this probability depends on the relative velocity, $\overrightarrow{v}_{ij}=\overrightarrow{v}_{i}-% \overrightarrow{v}_{j}$, and position $\overrightarrow{q}_{ij}$ (e.g., through the combination $\left( \overrightarrow{v}_{ij}\cdot \overrightarrow{% q}_{ij}\right) ^{2}$). Collision Rules --------------- The coupling between chemistry and hydrodynamics can be captured simply by considering atoms that carry a label (e.g., color) that can change during collisions. Allowing for the non-conservation of energy gives a relatively broad model that includes endo- and exothermic reactions. However, in the interests of generality, the problem of modeling reactions that not only violate energy conservation, but that also allow for the exchange, or even loss, of mass will be considered. The modelling of the collision rules in the case that mass is either exchanged or lost upon collision is somewhat problematic. To understand why, consider the usual arguments leading to specular collision rules in the case that mass is invariant. Defining the total and relative momenta respectively as$$\begin{aligned} \overrightarrow{P} &=&\overrightarrow{p}_{1}+\overrightarrow{p}_{2} \\ \overrightarrow{p} &=&\overrightarrow{p}_{1}-\overrightarrow{p}_{2}, \notag\end{aligned}$$the conservation of total momentum means that$$\begin{aligned} \overrightarrow{P}^{\prime } &=&\overrightarrow{P} \label{rule} \\ \overrightarrow{p}^{\prime } &=&\overrightarrow{p}+\overrightarrow{\gamma }% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} \notag\end{aligned}$$where $\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} $ $(x_{1,}x_{2})=-\overrightarrow{\gamma }_{l_{2}l_{1}}^{l_{2}^{\prime }l_{1}^{\prime }}\left( x_{2},x_{1}\right) $ is to be determined. Second, the energy balance equation can be written as$$E(x_{1}^{\prime },x_{2}^{\prime })+\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}(x_{1},x_{2})=E(x_{1}^{\prime },x_{2}^{\prime })$$where $\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}$ $% (x_{1},x_{2}) $ is the energy lost during the collision. Substitution of Eq.(\[rule\]) gives$$\left( \gamma _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\right) ^{2}+2\left( \overrightarrow{p}+\frac{m_{l_{2}}-m_{l_{1}}}{% m_{l_{1}}+m_{l_{2}}}\overrightarrow{P}\right) \cdot \overrightarrow{\gamma }% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}+8\mu _{l_{1}l_{2}}\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=0 \label{coll1}$$where the reduced mass is $\mu _{l_{1}l_{2}}=\frac{m_{l_{1}}m_{l_{2}}}{% m_{l_{1}}+m_{l_{2}}}.$In $D$-dimensions, this gives one constraint on the $D$ independent components of $\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{% \prime }l_{2}^{\prime }}$ so, for example, it fixes the magnitude of $% \overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}$ if its direction is known: In one dimension, the problem is therefore solved. In higher dimensions, the conservation of angular momentum gives the needed additional constraint. This reads$$\overrightarrow{P}^{\prime }\times \frac{\overrightarrow{q}_{1}+% \overrightarrow{q}_{2}}{2}+\frac{1}{2}\overrightarrow{p}_{12}^{\prime }\times \overrightarrow{q}_{12}=\overrightarrow{P}\times \frac{% \overrightarrow{q}_{1}+\overrightarrow{q}_{2}}{2}+\frac{1}{2}\overrightarrow{% p}_{12}\times \overrightarrow{q}_{12}$$or, using the conservation of total momentum,$$\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\times \overrightarrow{q}_{12}=0$$thus fixing the direction of $\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{% \prime }l_{2}^{\prime }}$ as being along the line joining the centers of the atoms. In this case, eq.(\[coll1\]) gives$$\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=2\mu _{l_{1}l_{2}}\left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}-% \sqrt{\left( \overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) ^{2}-\frac{2}{\mu _{l_{1}l_{2}}}\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}\right) \widehat{q}_{12}.$$Taking $\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=0$ gives the usual result for elastic hard spheres whereas setting the loss to be a fixed fraction of the contribution to the rest-frame kinetic energy due to the velocity along the line joining the atoms, $\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\frac{1}{2}\mu _{l_{1}l_{2}}\left( \overrightarrow{v}_{12}\cdot \overrightarrow{q}% _{12}\right) ^{2}$, is the model used for inelastic hard spheres (i.e., granular fluids) and gives a coefficient of restitution $\alpha _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=\sqrt{1-\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}$. When mass can be exchanged upon collision, it is useful to introduce the total mass $M_{l_{1}l_{2}}=m_{l_{1}}+m_{l_{2}}$, the center of mass $% \overrightarrow{Q}=\frac{m_{l_{1}}\overrightarrow{q}_{1}+m_{l_{2}}% \overrightarrow{q}_{2}}{M_{l_{1}l_{2}}}$ and the center of mass velocity $% \overrightarrow{V}=\overrightarrow{P}/M_{l_{1}l_{2}}$. Notice that even with the conservation of total mass, the center of mass is not generally invariant if mass is exchanged and the positions are kept fixed. So, the conservation of angular momentum gives$$M_{l_{1}^{\prime }l_{2}^{\prime }}\overrightarrow{V}^{\prime }\times \overrightarrow{Q}^{\prime }+\mu _{l_{1}^{\prime }l_{2}^{\prime }}v^{\prime }\times \overrightarrow{q}_{12}=M_{l_{1}l_{2}}\overrightarrow{V}\times \overrightarrow{Q}+\mu _{l_{1}l_{2}}\overrightarrow{v}\times \overrightarrow{% q}_{12}$$or, using the conservation of total momentum,$$M_{l_{1}l_{2}}\overrightarrow{V}\times \left( \overrightarrow{Q}^{\prime }-% \overrightarrow{Q}\right) +\left( \mu _{l_{1}^{\prime }l_{2}^{\prime }}% \overrightarrow{v}^{\prime }-\mu _{l_{1}l_{2}}\overrightarrow{v}\right) \times \overrightarrow{q}_{12}=0. \label{angular}$$In general, this equation cannot be satisfied since it implies$$\left( \overrightarrow{V}\times \left( \overrightarrow{Q}^{\prime }-% \overrightarrow{Q}\right) \right) \cdot \overrightarrow{q}_{12}=0$$which is not generally true. The conclusion is that any collision rule which does not leave invariant the center of mass will necessarily result in a violation of the conservation of angular momentum at the microscopic level. It is not possible to compensate by allowing the positions of the atoms to shift during the collision since this could lead to overlapping configurations involving a third atom. In fact, one would expect that the inclusion of internal degrees of freedom, in particular of rotation of the spherical atoms, would allow for a galilean-invariant collision law and this will be explored at a later date. For present purposes, given that the collision rule cannot be uniquely fixed by appealing to general principles, the only recourse is to try to construct reasonable models. One possibility is to conserve the angular momentum in the center of mass (CM) rest frame since then $\overrightarrow{V}=0$ and angular momentum can indeed be conserved. Another is to work in analogy to the case of invariant masses and to require that either all momentum transfer be along $\overrightarrow{q}% _{12}$ (so $\overrightarrow{p}^{\prime }=\overrightarrow{p}+2\mu _{l_{1}l_{2}}\gamma _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{q}% _{12}$) However, since the former is not Galilean invariant when mass is transferred (since in a frame moving at velocity $\overrightarrow{u}$ the relative momentum is $\overrightarrow{p}_{boosted}=\overrightarrow{p}+\left( m_{l_{1}}-m_{l_{2}}\right) \overrightarrow{u}$), it is not clear how to uniquely apply it. In fact, if one tries to enforce this constraint in the CM frame, it gives the same result as fixing the angular momentum in the CM frame. A second possibility is to demand that all velocity change be along $% \overrightarrow{q}_{12}$ (so $\overrightarrow{v}^{\prime }=\overrightarrow{v}% +\overline{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{q}% _{12}$). For illustrative purposes, both options will be considered. For the sake of generality, it is also interesting to consider the consequences when mass is not only exchanged, but is lost. It is clear that the previous considerations concerning the specification of the collision rule under mass exchange apply to this case as well so that, again, a model must be introduced in order to specify the relation between the lost mass and the total momentum and energy. Without considering specific applications, it is not clear that any unique conclusions can be drawn, so by way of illustration, I will assume that the mass is carried away in such a way that the total momentum in the CM frame is conserved. This means in general that the law of conservation of momentum becomes$$\overrightarrow{P}^{\prime }=\overrightarrow{P}-\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\overrightarrow{V}$$where $\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=m_{l_{1}}+m_{l_{2}}-m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}$. The model adopted here is that the mass is carried away by $n_{l_{1}l_{2}}^{l_{1}^{% \prime }l_{2}^{\prime }}$ particles with masses $m_{i}^{0}$ so that $\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=\sum_{i=1}^{n_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}m_{i}^{0}$ and with rest-frame velocities $\overrightarrow{v}_{i}^{0}$ satisfying $% \sum_{i=1}^{n_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}\overrightarrow{v% }_{i}^{0}=0$. For the case in which the angular momentum in the rest CM frame is held constant, it is natural to also require that these particles carry no net angular momentum. Finally, some model must be specified for the energy lost (or gained), $% \delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}$, which might include contributions due to kinetic energy that is carried away by the lost mass and energy lost (or gained) through other mechanisms (excitation of internal degrees of freedom, radiation, exothermic and endothermic chemical reactions...). Suppressing the species indices for a moment, the energy differential can be written as the sum of two contributions, $\delta E=\delta E_{m}+\delta E_{0}$, where the first is the energy carried away with the lost mass and the second is due to any other inelastic processes. In the CM rest frame, $\delta E_{m}=\sum_{i=1}^{n}\frac{1}{2}m_{i}\left( \overrightarrow{v}_{i}^{0}\right) ^{2}\equiv \overline{\delta E_{m}}$ so in the lab frame, $\delta E_{m}=\sum_{i=1}^{n}\frac{1}{2}m_{i}\left( \overrightarrow{v}_{i}^{0}+\overrightarrow{V}\right) ^{2}=\overline{\delta E_{m}}+\frac{1}{2}\delta mV^{2}$. Further, I assume, as is commonly done, that the remaining energy loss (or gain) is frame independent (which means in particular that it can only be a function of the relative velocity of the colliding atoms). The energy balance equation therefore reads$$\frac{1}{2m_{l_{1}^{\prime }}}p_{1}^{\prime 2}+\frac{1}{2m_{l_{2}^{\prime }}}% p_{2}^{\prime 2}+\overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}+\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}V^{2}=\frac{1}{2m_{l_{1}}}p_{1}^{2}+\frac{1}{2m_{l_{2}}}% p_{2}^{2} \label{Ebal}$$where it is understood that $\overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}=\overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}(x_{1},x_{2})$ is a galilean-invariant function of the phases. This expression depends on the model for the lost mass (if any) but is independent of any other assumptions concerning the collision rule. From Eq.(\[angular\]), conservation of angular momentum in the rest frame then gives$$\overrightarrow{v}_{12}^{\prime }=\frac{\mu _{l_{1}l_{2}}}{\mu _{l_{1}^{\prime }l_{2}^{\prime }}}\left( \overrightarrow{v}_{12}+\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{q}\right)$$for some scalar $\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}(x_{1},x_{2})$. Substituting into Eq.(\[Ebal\]) gives $$\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=-\overrightarrow{v}% \cdot \widehat{q}+\sqrt{\left( \overrightarrow{v}\cdot \widehat{q}\right) ^{2}-\left( 1-\mu _{l_{1}^{\prime }l_{2}^{\prime }}/\mu _{l_{1}l_{2}}\right) v^{2}-\frac{2\mu _{l_{1}^{\prime }l_{2}^{\prime }}}{\mu _{l_{1}l_{2}}^{2}}% \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}-\frac{\mu _{l_{1}^{\prime }l_{2}^{\prime }}}{\mu _{l_{1}l_{2}}^{2}}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}V^{2}}$$ and one then finds that$$\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=2\mu _{l_{1}l_{2}}\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{q% }_{12}+\left( \frac{m_{l_{1}^{\prime }}-m_{l_{1}}}{m_{l_{1}}+m_{l_{2}}}-% \frac{m_{l_{2}^{\prime }}-m_{l_{2}}}{m_{l_{1}}+m_{l_{2}}}\right) \overrightarrow{P}.$$Demanding that the change in the relative velocity be along the line joining the atoms gives a very similar result$$\overrightarrow{v}_{12}^{\prime }=\overrightarrow{v}_{12}+\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{q}_{12}$$with$$\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=-\overrightarrow{v}% _{12}\cdot \widehat{q}_{12}+\sqrt{\left( \overrightarrow{v}_{12}\cdot \widehat{q}_{12}\right) ^{2}-\left( 1-\mu _{l_{1}l_{2}}/\mu _{l_{1}^{\prime }l_{2}^{\prime }}\right) v_{12}^{2}-\frac{2}{\mu _{l_{1}^{\prime }l_{2}^{\prime }}}\overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}-\frac{\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}{\mu _{l_{1}^{\prime }l_{2}^{\prime }}}V^{2}}$$and$$\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=\left( \frac{\mu _{l_{1}^{\prime }l_{2}^{\prime }}}{\mu _{l_{1}l_{2}}}% -1\right) \overrightarrow{p}_{12}+2\mu _{l_{1}^{\prime }l_{2}^{\prime }}\lambda _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{q}% _{12}+\left( \frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{% m_{l_{1}}+m_{l_{2}}}-\frac{\mu _{l_{1}^{\prime }l_{2}^{\prime }}}{\mu _{l_{1}l_{2}}}\left( \frac{m_{l_{1}}-m_{l_{2}}}{m_{l_{1}}+m_{l_{2}}}\right) \right) \overrightarrow{P}$$Note that these two models coincide in the special case that the reduced mass is invariant which obtains in one of two circumstances: the atomic masses are invariant or if the atoms just exchange masses so that $% m_{l_{1}^{\prime }}=m_{l_{2}}$ and vice-versa. In both cases, mass is necessarily conserved, $\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=0$, and all other conclusions are model-independent consequences of galilean-invariance. The evolution of phase functions -------------------------------- The dynamics of any hard-sphere model consists of free streaming interrupted by binary collisions. In non-reacting fluids, the collisions lead to an instantaneous change of the velocities of the colliding atoms. The generalization to the reacting fluid only requires that the chemical species labels, and hence the masses and any other species-specific properties, to be viewed as dynamical variables as well and, so, as part of an enlarged phase space. Two atoms, say atoms $i$ and $j$, collide at time $\tau _{ij}$ when their centers are separated by their relative hard-sphere diameter $\sigma _{l_{i}l_{j}}$$$\left| \overrightarrow{q}_{i}(\tau _{ij})-\overrightarrow{q}_{j}(\tau _{ij})\right| =q_{ij}^{2}(\tau _{ij})=\sigma _{l_{i}l_{j}}^{2} \label{cq}$$where, e.g., for additive models, the relative hard sphere diameter is simply the sum of the atoms’ radii $\sigma _{l_{i}l_{j}}=\frac{1}{2}\sigma _{l_{j}}+\frac{1}{2}\sigma _{l_{j}}$. The atoms do not have to all have the same size (e.g., an acceptable possibility is that different species have different sizes but chemical reactions always transform atoms from a species of a given size to other species of the same size). An exceptional possibility in which size could change is one in which atoms only get smaller upon collision: this might be useful to model certain granular materials that fragment upon collision (e.g., the ice composing the rings of Saturn) and could be handled within the present formalism as long as the position of the center of mass of each atom is invariant. From Eq.(\[cq\]), one has that the time of collision is $$\tau _{ij}\left( \Gamma \right) =-\frac{1}{v_{ij}^{2}}\overrightarrow{v}% _{ij}\cdot \overrightarrow{q}_{ij}-\frac{1}{v_{ij}^{2}}\sqrt{\left( \overrightarrow{v}_{ij}\cdot \overrightarrow{q}_{ij}\right) ^{2}-v_{ij}^{2}\left( q_{ij}^{2}-\sigma _{l_{i}l_{j}}^{2}\right) }$$where the sign has been chosen according to give the physical solution. If the right hand side is imaginary, then no collision takes place for the given initial conditions. This aspect of the dynamics is independent of what actually happens after the collision and is the reason that the structure of the pseudo-Liouville equation is independent of the collision rule. The pseudo-Liouville equation describing the time evolution of an arbitrary phase function, $A\left( \Gamma ;t\right) =A\left( x_{1},l_{1},...x_{N},l_{N};t\right) $ then follows immediately by analogy with the non-reacting fluid and is $$\begin{aligned} \frac{d}{dt}A &=&\frac{\partial }{\partial t}A+\widehat{\mathcal{L}}_{+}A \\ \widehat{\mathcal{L}}_{+} &=&\sum_{i}\overset{\cdot }{x}_{i}\frac{\partial }{% \partial x_{i}}+\sum_{i<j}\widehat{\mathcal{T}}_{+}\left( ij\right) \notag\end{aligned}$$where the binary collision operators are $$\widehat{\mathcal{T}}_{+}\left( ij\right) =-\overrightarrow{q}_{ij}\cdot \overrightarrow{v}_{ij}\delta \left( q_{ij}-\sigma _{l_{i}l_{j}}\right) \Theta \left( -\widehat{q}_{ij}\cdot \overrightarrow{v}_{ij}\right) \left( \sum_{l_{i}^{\prime }l_{j}^{\prime }}\mathcal{M}_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}\left( x_{ij}\right) \widehat{b}_{l_{i}l_{j}}^{l_{i}^{% \prime }l_{j}^{\prime }}-1\right) .$$As discussed in Appendix \[AppT\], this can be derived directly for a system of two atoms by writing the exact solution to the two-body problem and differentiating; the generalization to $N$-atoms follows immediately due to the fact that only binary collisions occur. Here, $\mathcal{M}% _{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}\left( x_{i}{}_{j}\right) $ is a random matrix which, in any realization, takes on the value $1$ for some single combination of $l_{i}^{\prime },l_{j}^{\prime }$ and is zero otherwise and which is distributed according to $$\left\langle \mathcal{M}_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}\left( x_{ij}\right) \right\rangle _{react}=K_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}\left( x_{ij}\right) . \label{stoch}$$(The notation used here indicates stochastic quantities by means of Calligraphic type and uses carets to denote operators and averages over the stochastic process are denoted as $\left\langle ...\right\rangle _{react}$.) For a non-reacting system, it becomes $\mathcal{M}_{l_{i}l_{j}}^{l_{i}^{% \prime }l_{j}^{\prime }}\left( x_{ij}\right) =\delta _{l_{i}^{\prime }l_{i}}\delta _{l_{j}^{\prime }l_{j}}$. The only other formal difference from the non-reacting case is that the momentum transfer operator, $\widehat{% b}_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}$, has the effect of altering both the mechanical variables and the species labels. So, just as this operator instantaneously changes the position in phase space of the $i$th atom from $x_{i}(t_{-})$ before a collision at time $t$ to $% x_{i}(t_{+})=x_{i}^{\prime }(t_{-})$ it also instantaneously alters the species labels from $l_{i}\left( t_{-}\right) $ to $l_{i}\left( t_{+}\right) =l_{i}^{\prime }\left( t_{-}\right) $ the difference being that $% x_{i}^{\prime }(t_{-})$ is a deterministic function of $x_{i}(t_{-})$ and $% x_{j}(t_{-})$ whereas the evolution of $l_{i}^{\prime }(t_{-})$ is stochastic. For phase functions which have no explicit time dependence, the Liouville equation can be formally solved to get$$\mathcal{A}(\Gamma ,t)=\exp \left( \widehat{\mathcal{L}}_{+}t\right) A(\Gamma )$$which has the meaning that the system evolves from the initial phase $\Gamma $. The most important difference from the non-reacting system appears in the evaluation of statistical averages. In the presence of reactions there are two statistical processes that must be considered: the distribution of initial conditions and the stochastic process that alters species labels at the collisions. For a given distribution of initial conditions $\rho ^{(0)}\left( \Gamma \right) =\rho _{l_{1}l_{2}...l_{N}}^{(0)}\left( x_{1},x_{2},...x_{N}\right) $ (giving the probability that the first atom begins with species $l_{1}$ and phase $x_{1}$, and so), one has$$\left\langle A;t\right\rangle =\int d\Gamma \;\rho ^{(0)}\left( \Gamma \right) \left\langle \mathcal{A}(t)\right\rangle _{react}=\int d\Gamma \;\rho ^{(0)}\left( \Gamma \right) \left\langle \exp \left( \widehat{% \mathcal{L}}_{+}t\right) A(\Gamma )\right\rangle _{react}$$and the notation should be understood as implying a sum over the initial species labels$$\begin{aligned} &&\int d\Gamma \;\rho ^{(0)}\left( \Gamma \right) \left\langle \exp \left( \widehat{\mathcal{L}}_{+}t\right) A(\Gamma )\right\rangle _{react} \\ &\equiv &\sum_{l_{1}...l_{N}}\int dx_{1}...dx_{N}\;\rho _{l_{1}...l_{N}}^{(0)}\left( x_{1},...x_{N}\right) \left\langle \exp \left( \widehat{\mathcal{L}}_{+}t\right) \right\rangle _{react}A(\Gamma ) \notag\end{aligned}$$where $A(\Gamma )$ can be taken outside of the average over reactions since it depends only on the initial conditions. Now, since each collision involves an independent stochastic process, it follows that $\left\langle \exp \left( \widehat{\mathcal{L}}_{+}t\right) \right\rangle _{react}=\exp \left( \left\langle \widehat{\mathcal{L}}_{+}\right\rangle _{react}t\right) $ which is evaluated using Eq.(\[stoch\]). Thus, the time averages become$$\left\langle A;t\right\rangle =\sum_{l_{1}...l_{N}}\int dx_{1}...dx_{N}\;\rho _{l_{1}...l_{N}}^{(0)}\left( x_{1},...x_{N}\right) \exp \left( \widehat{L}_{+}t\right) A(\Gamma ) \label{av}$$with the deterministic operator$$\widehat{L}_{+}=\sum_{i}\overset{\cdot }{x}_{i}\frac{\partial }{\partial x_{i}}+\sum_{i<j}\widehat{T}_{+}\left( ij\right)$$and the (reaction-averaged) collision operators are$$\widehat{T}_{+}\left( ij\right) =-\overrightarrow{q}_{ij}\cdot \overrightarrow{v}_{ij}\delta \left( q_{ij}-\sigma _{l_{i}l_{j}}\right) \Theta \left( -\widehat{q}_{ij}\cdot \overrightarrow{v}_{ij}\right) \left( \sum_{l_{i}^{\prime }l_{j}^{\prime }}K_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}\left( x_{ij}\right) \widehat{b}_{l_{i}l_{j}}^{l_{i}^{% \prime }l_{j}^{\prime }}-1\right) .$$This shows that, from the point of view of evaluating the statistical averages, it suffices to work with the deterministic dynamics defined by $% \widehat{L}_{+}$ which no longer treats the species labels as discrete stochastic variables. Instead, the phase functions are at all times averaged over the reactions and so do not explicitly represent dynamical quantities as might be realized in a computer simulation. In fact, they correspond to the average result of an ensemble of simulations, all beginning with identical initial conditions, but differing in the realization of the reaction process $\mathcal{M}_{l_{i}l_{j}}^{l_{i}^{\prime }l_{j}^{\prime }}\left( x_{i}{}_{j}\right) $. The evolution of the distribution function ------------------------------------------ The adjoint $\widehat{L}_{+}^{A}$ of the Liouville operator $\widehat{L}_{+}$ is defined as$$\int d\Gamma \;B\left( \Gamma \right) \widehat{L}_{+}A\left( \Gamma \right) =\int d\Gamma \;\left( \widehat{L}_{+}^{A}B\left( \Gamma \right) \right) A\left( \Gamma \right) \label{adjoint_def}$$from which one finds (see appendix \[AppA\])$$\widehat{L}_{+}^{A}=-\sum_{i}\overset{\cdot }{x}_{i}\frac{\partial }{% \partial x_{i}}+\sum_{i<j}\widehat{T}_{+}^{A}\left( ij\right)$$with the adjoint collision operator$$\widehat{T}_{+}^{A}\left( ij\right) =-\left[ \sum_{a,b}J_{ab}^{l_{i}l_{j}}% \left( x_{i},x_{j}\right) \left( \widehat{b}_{ab}^{l_{i}l_{j}}\right) ^{-1}K_{ab}^{l_{i}l_{j}}\left( x_{ij}\right) -1\right] \Theta \left( -% \overrightarrow{v}_{ij}\cdot \overrightarrow{q}_{ij}\right) \delta \left( q_{ij}-\sigma _{l_{i}l_{j}}\right) \overrightarrow{v}_{ij}\cdot \overrightarrow{q}_{ij}$$with$$J_{ab}^{l_{1}l_{2}}\left( x_{i},x_{j}\right) =\left| \frac{\partial \left( \left( \widehat{b}_{ab}^{l_{i}l_{j}}\right) ^{-1}x_{i},\left( \widehat{b}% _{ab}^{l_{i}l_{j}}\right) ^{-1}x_{j}\right) }{\partial \left( x_{i},x_{j}\right) }\right| ^{-1}.$$Here, the operator $\left( \widehat{b}_{ab}^{l_{i}l_{j}}\right) ^{-1}$ is the inverse of $\widehat{b}_{ab}^{l_{i}l_{j}}$ both in terms of the change of the mechanical variables as well as the species labels so that for an arbitrary function $\left( \widehat{b}_{ab}^{l_{i}l_{j}}\right) ^{-1}B\left( x_{i},l_{i};x_{j},l_{j}\right) =B\left( \left( \widehat{b}% _{ab}^{l_{i}l_{j}}\right) ^{-1}x_{i},a;\left( \widehat{b}_{ab}^{l_{i}l_{j}}% \right) ^{-1}x_{j},b\right) $. To illustrate the structure of this operator, consider the case of inelastic hard spheres used to model granular fluids. Specializing to a single species, one has$$\overrightarrow{v}_{ij}^{\prime }=\widehat{b}\overrightarrow{v}_{ij}=% \overrightarrow{v}_{ij}-\left( 1+\alpha \right) \overrightarrow{v}_{ij}\cdot \widehat{q}_{ij},$$where $\alpha $ is a constant, from which it follows that $$\overrightarrow{v}_{ij}=\widehat{b}^{-1}\overrightarrow{v}_{ij}^{\prime }=% \overrightarrow{v}_{ij}^{\prime }-\left( \frac{1+\alpha }{\alpha }\right) \overrightarrow{v}_{ij}^{\prime }\cdot \widehat{q}_{ij}$$giving$$\left| \frac{\partial \left( \widehat{b}^{-1}x_{i},\widehat{b}% ^{-1}x_{j}\right) }{\partial \left( x_{i},x_{j}\right) }\right| =\left| 1-\left( \frac{1+\alpha }{\alpha }\right) \right| =\frac{1}{\alpha }$$so that $$\widehat{T}_{+}^{A}\left( ij\right) B\left( x_{i},l_{i};x_{j},l_{j}\right) =-% \left[ \frac{1}{\alpha }\widehat{b}^{-1}-1\right] \Theta \left( -% \overrightarrow{v}_{ij}\cdot \overrightarrow{q}_{ij}\right) \delta \left( q_{ij}-\sigma _{l_{i}l_{j}}\right) \overrightarrow{v}_{ij}\cdot \overrightarrow{q}_{ij}B\left( x_{i},l_{i};x_{j},l_{j}\right) .$$which is the usual result[@HCSLiouville]. An important generalization of this result concerns the case that the inverse transformation $\widehat{b}_{ab}^{l_{i}l_{j}}x_{i}$ is not unique. This can happen even in the single species, inelastic case if the coefficient of restitution depends on the velocities. For example, if $% \alpha =\alpha \left( \overrightarrow{v}_{ij}\cdot \widehat{q}_{ij}\right) $ then the inverse collision rule is determined by solving$$\overrightarrow{v}_{ij}^{\prime }\cdot \widehat{q}_{ij}=-\alpha \left( \overrightarrow{v}_{ij}\cdot \widehat{q}_{ij}\right) \overrightarrow{v}% _{ij}\cdot \widehat{q}_{ij}$$which may or may not have a unique solution. In the latter case, $\widehat{T}% _{+}^{A}$ must be written in terms of a sum over the various branches and must include step functions which restrict the domain of integration in Eq.(\[adjoint\_def\]) to the appropriate domain for each branch. In practical calculations, it is usually most convenient to recast integrals over $% \widehat{T}_{+}^{A}\left( ij\right) $ into integrals involving $\widehat{T}% _{+}\left( ij\right) $ so as to avoid this complication. Given the adjoint operator $\widehat{L}_{+}^{A}$, Eq.(\[av\]) can be written as$$\left\langle A;t\right\rangle =\int d\Gamma \;\left( \exp \left( L^{A}t\right) \rho ^{(0)}\left( \Gamma \right) \right) A\left( \Gamma \right) \equiv \int d\Gamma \;\rho \left( \Gamma ;t\right) A\left( \Gamma \right)$$where the second equality defines the time-dependent distribution function. It’s time dependence is given by the pseudo-Liouville equation $$\left( \frac{\partial }{\partial t}+\sum_{i}\overset{\cdot }{x}_{i}\frac{% \partial }{\partial x_{i}}+\sum_{i<j}\overline{T}_{-}\left( ij\right) \right) \rho =0 \label{Liouville}$$where in the standard notation[@McLennan]$$\overline{T}_{-}\left( ij\right) =-T_{+}^{A}\left( ij\right) .$$ The Born-Bogoliubov-Green-Kirkwood-Yvon (BBGKY) hierarchy follows immediately from the Liouville equation. Defining the reduced distribution functions as $$f_{l_{1}...l_{m}}\left( x_{1}...x_{m}\right) =\frac{N!}{\left( N-m\right) !}% \sum_{l_{m+1}...l_{N}}\int dx_{m+1}...dx_{N}\;\rho \left( \Gamma \right) .$$and integrating the pseudo-Liouville equation over $x_{m+1}...x_{N}$ and summing over the corresponding species labels gives the $m$-th equation of the hierarchy$$\begin{aligned} &&\left( \frac{\partial }{\partial t}+\sum_{i=1}^{m}\overrightarrow{v}% _{i}\cdot \frac{\partial }{\partial \overrightarrow{q}_{i}}+\sum_{1\leq i<j\leq m}\overline{T}_{-}\left( ij\right) \right) f_{l_{1}...l_{m}}\left( x_{1}...x_{m}\right) \label{BBGKY} \\ &=&-\sum_{i=1}^{m}\sum_{l_{m+1}}\int dx_{m+1}\overline{T}_{-}\left( im+1\right) f_{l_{1}...l_{m+1}}\left( x_{1}...x_{m+1}\right) . \notag\end{aligned}$$The first equation of the hierarchy is the starting point for the Enskog kinetic theory as described below. Exact balance equations ======================= Now consider the phenomenology of the reacting fluid which is expressed in terms of the macroscopic hydrodynamic fields. The results presented here are derived using only the general form of the collision rule, eq.(\[rule\]), and the microscopic energy balance equation, eq.(\[Ebal\]) so that the only assumptions made with respect to the collision model are those concerning the energy transported by any lost mass. The local fields of interest are the number fractions$$n_{l}\left( \overrightarrow{r},t\right) =\left\langle \sum_{i}\delta \left( \overrightarrow{r}-\overrightarrow{q}_{i}\right) \delta _{ll_{i}};t\right\rangle =\int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1};t\right) ,$$and the mass, momentum and energy densities, defined respectively as$$\begin{aligned} \rho \left( \overrightarrow{r},t\right) &=&\sum_{l}m_{l}n_{l}\left( \overrightarrow{r},t\right) \\ \rho \left( \overrightarrow{r},t\right) \overrightarrow{u}(\overrightarrow{r}% ,t) &=&\left\langle \sum_{i}m_{l_{i}}\overrightarrow{v}_{i}\delta \left( \overrightarrow{r}-\overrightarrow{q}_{i}\right) ;t\right\rangle =\sum_{l}m_{l}\int d\overrightarrow{v}_{1}\;\overrightarrow{v}% _{1}f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1};t\right) \notag \\ \frac{D}{2}n\left( \overrightarrow{r},t\right) k_{B}T\left( \overrightarrow{r% },t\right) &=&\left\langle \sum_{i}\frac{1}{2}m_{l_{i}}V_{i}^{2}\delta \left( \overrightarrow{r}-\overrightarrow{q}_{i}\right) ;t\right\rangle =\sum_{l}\frac{1}{2}m_{l}\int d\overrightarrow{v}_{1}\;V_{1}^{2}f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1};t\right) \notag\end{aligned}$$where the excess velocity is $\overrightarrow{V}_{i}(t)=\overrightarrow{v}% _{i}(t)-\overrightarrow{u}\left( \overrightarrow{q}_{i},t\right) $and the total number density is $$n\left( \overrightarrow{r},t\right) =\sum_{l}n_{l}\left( \overrightarrow{r}% ,t\right) .$$It is also convenient to introduce the number fractions, or concentrations, $% x_{l}\left( \overrightarrow{r},t\right) =n_{l}\left( \overrightarrow{r}% ,t\right) /n\left( \overrightarrow{r},t\right) $. The balance equations for these quantities follow directly from their definitions and the first equation of the BBGKY hierarchy. The details of the derivation are given in appendix \[AppB\] and only the results summarized here. Number. mass and concentration ------------------------------ Integrating over the positions and velocities gives the balance equation for the local partial number density$$\frac{\partial }{\partial t}n_{l}+\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}n_{l}\right) +\overrightarrow{\nabla }\cdot \overrightarrow{j}_{l}=S_{l}^{(n)} \label{n1}$$with the source$$\begin{aligned} S_{l}^{(n)}\left( \overrightarrow{r},t\right) &=&-\frac{1}{2}% \sum_{abl_{1}l_{2}}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \label{n2} \\ &&\times f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) K_{l_{1}l_{2}}^{ab}\left( x_{12}\right) \left( \delta _{al}+\delta _{bl}-\delta _{ll_{1}}-\delta _{ll_{2}}\right) \notag\end{aligned}$$and the number current $\overrightarrow{j}_{l}=\overrightarrow{j}_{l}^{K}+% \overrightarrow{j}_{l}^{V}$ with$$\overrightarrow{j}_{l}^{K}\left( \overrightarrow{r},t\right) =\int d% \overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}% _{1},t\right) \overrightarrow{V}_{1} \label{n3}$$and$$\begin{aligned} \overrightarrow{j}_{l}^{V}\left( \overrightarrow{r},t\right) &=&-\frac{1}{2}% \sum_{abl_{1}l_{2}}\int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \label{n4} \\ &&\times f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) K_{l_{1}l_{2}}^{ab}\left( x_{12}\right) \left( \delta _{al}-\delta _{bl}-\delta _{ll_{1}}+\delta _{ll_{2}}\right) \notag \\ &&\times \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) . \notag\end{aligned}$$The source term represents the gain or loss of atoms of type $l$ due to chemical reactions. The kinetic part of the number current is familiar from the study of multi-component, non-reacting systems[@McLennan] where it takes the form $\overrightarrow{j}_{l}^{K}=\sum_{j}D_{lj}\overrightarrow{% \nabla }n_{j}+L_{l}\overrightarrow{\nabla }T+o\left( \nabla ^{2}\right) $ and, e.g., gives rise to Fick’s law when substituted into eq.(\[n1\]) . Here, it is seen that this diffusive current is enhanced by a second contribution, eq.(\[n4\]), that arises solely from the reactions (i.e., it vanishes if $K_{l_{1}l_{2}}^{ab}=\delta _{al_{1}}\delta _{bl_{2}}$). This is due to the transport of type-$l$ atoms due to the reaction process. The conservation of total number density immediately follows by summing over the species label$$\frac{\partial }{\partial t}n+\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}n\right) +\overrightarrow{\nabla }\cdot \sum_{l}% \overrightarrow{j}_{l}^{K}=0$$where the sum over the species of the collisional contributions to the number current vanishes. Similarly, multiplying by $m_{l}$ and then summing gives the balance equation for local mass density$$\frac{\partial }{\partial t}\rho +\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}\rho \right) +\overrightarrow{\nabla }\cdot \overrightarrow{Q}=S^{(\rho )},$$where, the mass flux is $$\begin{aligned} \overrightarrow{Q}\left( \overrightarrow{r},t\right) &=&-\frac{1}{2}% \sum_{abl_{1}l_{2}}\int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) K_{l_{1}l_{2}}^{ab}\left( x_{12}\right) \left( m_{a}-m_{b}-m_{l_{1}}+m_{l_{2}}\right) \notag \\ &&\times \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) , \notag\end{aligned}$$which vanishes if no mass is transported during collisions, and the mass source term is$$\begin{aligned} S_{l}^{(\rho )}\left( \overrightarrow{r},t\right) &=&\frac{1}{2}% \sum_{abl_{1}l_{2}}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) K_{l_{1}l_{2}}^{ab}\left( x_{12}\right) \delta m_{l_{1}l_{2}}^{ab}, \notag\end{aligned}$$which is only nonzero if the collisions do not conserve mass. Finally, using the definition of the concentrations, $x_{l}=n_{l}/n$, the reaction equation is found to be $$\frac{\partial }{\partial t}x_{l}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }x_{l}+n^{-1}\left[ \overrightarrow{\nabla }\cdot \overrightarrow{j}% _{l}-x_{l}\overrightarrow{\nabla }\cdot \sum_{l}\overrightarrow{j}_{l}^{K}% \right] =n^{-1}S_{l}^{(n)}$$where the term on the right is now identified as the reaction rate. Momentum and velocity fields ---------------------------- The balance equation for the local momentum, written in terms of the local velocity, is$$\frac{\partial }{\partial t}\rho \overrightarrow{u}+\overrightarrow{\nabla }% \cdot \left( \rho \overrightarrow{u}\overrightarrow{u}\right) +% \overrightarrow{\nabla }\cdot \left( \overleftrightarrow{P}+\overrightarrow{Q% }\overrightarrow{u}\right) =\overline{\overrightarrow{S}}^{(p)}+% \overrightarrow{u}S_{l}^{(\rho )}$$with the pressure tensor $\overleftrightarrow{P}=\overleftrightarrow{P}^{K}+% \overleftrightarrow{P}^{V}+\overleftrightarrow{P}^{M}$ where the kinetic contribution is $$\overleftrightarrow{P}^{K}\left( \overrightarrow{r},t\right) =\sum_{l}m_{l}\int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},% \overrightarrow{v}_{1},t\right) \overrightarrow{V}_{1}\overrightarrow{V}_{1},$$and the collisional contribution is$$\begin{aligned} \overleftrightarrow{P}^{V}\left( \overrightarrow{r},t\right) &=&-\frac{1}{2}% \sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;% \overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -% \widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \widetilde{\overrightarrow{\gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x% \overrightarrow{q}_{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) \notag\end{aligned}$$where $\widetilde{\overrightarrow{\gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}$ is the change of momentum in the rest frame$$\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=% \widetilde{\overrightarrow{\gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}+\left( m_{l_{1}^{\prime }}-m_{l_{1}}-m_{l_{2}^{\prime }}-m_{l_{2}}\right) \overrightarrow{V}_{12}.$$Finally, the contribution from the instantaneous exchange of mass is$$\begin{aligned} \overleftrightarrow{P}^{M} &=&-\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \left( \overrightarrow{V}_{12}-\overrightarrow{u}\right) \left( m_{l_{1}^{\prime }}-m_{l_{1}}-m_{l_{2}^{\prime }}-m_{l_{2}}\right) \notag \\ &&\times \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) \notag\end{aligned}$$The source terms arise due to momentum being carried away by the lost mass and the new term is given by$$\begin{aligned} \overline{\overrightarrow{S}}^{(p)}\left( \overrightarrow{r},t\right) &=&% \frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -% \widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times \left( \overrightarrow{V}_{12}-\overrightarrow{u}\right) f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\delta \left( \overrightarrow{r}-\overrightarrow{q}% _{1}\right) . \notag\end{aligned}$$By using the balance equation for total mass density the equation of motion for the velocity field is found to be $$\frac{\partial }{\partial t}\overrightarrow{u}+\overrightarrow{u}\cdot \overrightarrow{\nabla }\overrightarrow{u}+\rho ^{-1}\left( \overrightarrow{% \nabla }\cdot \overleftrightarrow{P}+\overrightarrow{Q}\cdot \overrightarrow{% \nabla }\overrightarrow{u}\right) =\rho ^{-1}\overline{\overrightarrow{S}}% ^{(p)}\text{.}$$ Energy density and temperature ------------------------------ The balance equation for total energy density is$$\frac{\partial }{\partial t}E+\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}E\right) +\overrightarrow{\nabla }\cdot \overrightarrow{q}+% \overrightarrow{\nabla }\cdot \left( \overrightarrow{u}\cdot \overleftrightarrow{P}\right) +\overrightarrow{\nabla }\cdot \left( \frac{1}{% 2}u^{2}\overrightarrow{Q}\right) =\xi +\overrightarrow{u}\cdot \overline{% \overrightarrow{S}}^{(p)}+\frac{1}{2}u^{2}\overline{S}^{(\rho )}.$$where the new source term, arising if energy is not conserved by the collisions, is$$\begin{aligned} \xi \left( \overrightarrow{r},t\right) &=&\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{% \prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}% _{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \\ &&\times \left[ \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}+\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( \overrightarrow{V}-\overrightarrow{u}\right) ^{2}% \right] f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) , \notag\end{aligned}$$which is recognized as the generalization of the source term studied in the context of granular fluids. The heat flux is written as a sum of several contributions$$\overrightarrow{q}=\overrightarrow{q}^{K}+\overrightarrow{q}^{V}+% \overrightarrow{q}^{m}+\overrightarrow{q}^{\delta E}$$where the kinetic part has the usual form$$\overrightarrow{q}^{K}\left( \overrightarrow{r},t\right) =\sum_{l}\frac{1}{2}% m_{l}\int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},% \overrightarrow{v}_{1},t\right) \overrightarrow{V}_{1}V_{1}^{2}$$as does the first part of the collisional contribution$$\begin{aligned} \overrightarrow{q}^{V}\left( \overrightarrow{r},t\right) &=&-\frac{1}{2}% \sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;% \overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -% \widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \left( \overrightarrow{V}-\overrightarrow{u}\right) \cdot \widetilde{% \overrightarrow{\gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) , \notag\end{aligned}$$which is a measure of energy displacement during the collision (i.e., one atom experiences a net gain of energy, the other a net loss and this represents an instantaneous movement of energy from the location of the second atom to the location of the first). Qualitatively new contributions arise from the instantaneous transfer of mass,$$\begin{aligned} \overrightarrow{q}^{m}\left( \overrightarrow{r},t\right) &=&-\frac{1}{2}% \sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\frac{m_{l_{2}^{\prime }}m_{l_{1}}-m_{l_{1}^{\prime }}m_{l_{2}}}{\left( m_{l_{2}^{\prime }}+m_{l_{1}^{\prime }}\right) \left( m_{l_{2}}+m_{l_{1}}\right) } \\ &&\times \int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}% _{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \notag \\ &&\times f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \mu _{l_{1}l_{2}}v^{2} \notag \\ &&\times \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) , \notag\end{aligned}$$and from the loss of energy,$$\begin{aligned} \overrightarrow{q}^{\delta E}\left( \overrightarrow{r},t\right) &=&-\frac{1}{% 2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}} \\ &&\times \int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}% _{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \notag \\ &&\times f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}} \notag \\ &&\times \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) . \notag\end{aligned}$$Alternatively, noting the relation between the total energy and the kinetic temperature$$E=\frac{D}{2}nk_{B}T+\frac{1}{2}\rho u^{2}$$the evolution of the kinetic temperature is found to be given by$$\left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) T-\frac{T}{n}\overrightarrow{\nabla }\cdot \sum_{l}% \overrightarrow{j}_{l}^{K}+\frac{2}{Dnk_{B}}\left[ \overleftrightarrow{P}:% \overrightarrow{\nabla }\overrightarrow{u}+\overrightarrow{\nabla }\cdot \overrightarrow{q}\right] =\frac{2}{Dnk_{B}}\xi .$$ Enskog Approximation -------------------- The expressions for the balance equations are exact. As a consequence, they depend on both the exact one-body and two-body distribution functions which are, in principle, determined by the BBGKY hierarchy. For example the equation for the one-body distribution is explicitly$$\begin{aligned} \left( \frac{\partial }{\partial t}+\overrightarrow{v}_{1}\cdot \frac{% \partial }{\partial \overrightarrow{q}_{1}}\right) f_{l_{1}}\left( x_{1};t\right) &=&-\sum_{a,b,l_{2}}\int d\overrightarrow{q}_{2}d% \overrightarrow{v}_{2} \label{BBGKY1} \\ &&\times \left[ \left| \frac{\partial \left( \widehat{b}% _{ab}^{l_{1}l_{2}}x_{1},\widehat{b}_{ab}^{l_{1}l_{2}}x_{2}\right) }{\partial \left( x_{1},x_{2}\right) }\right| ^{-1}\left( \widehat{b}% _{ab}^{l_{1}l_{2}}\right) ^{-1}K_{ab}^{l_{1}l_{2}}\left( x_{12}\right) -\delta _{l_{1}a}\delta _{l_{2}b}\right] \notag \\ &&\times \Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}% _{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \overrightarrow{v}_{12}\cdot \widehat{q}_{12} \notag \\ &&\times f_{l_{1}l_{2}}\left( \overrightarrow{q}_{1},\overrightarrow{v}_{1},% \overrightarrow{q}_{2},\overrightarrow{v}_{2};t\right) \notag\end{aligned}$$However, since the latter cannot be solved exactly, except in the special case of equilibrium, it is necessary to introduce an approximation. The most common approximation is to assume that the velocities of two colliding atoms are uncorrelated prior to the collision (they are of course correlated after the collision since the collision itself generates correlations). That this approximation is sufficient to decouple the BBGKY hierarchy is seen from the fact that the right hand side of eq. (\[BBGKY1\]) since the step function, $\Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) $, is non-zero only for atoms approaching one another and the delta function restricts the domain to the instant of contact. Thus, the assumption that atoms are uncorrelated just prior to a collision, Boltzmann’s ”assumption of molecular chaos”, is precisely the statement that$$\begin{aligned} &&\Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) f_{l_{1}l_{2}}\left( \overrightarrow{q}_{1},\overrightarrow{v}_{1},\overrightarrow{q}_{2},% \overrightarrow{v}_{2}\right) \label{mc} \\ &\simeq &\Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}% _{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) g\left( \overrightarrow{q}_{1},\overrightarrow{q}_{2};t\right) f_{l_{1}}\left( x_{1};t\right) f_{l_{2}}\left( x_{2};t\right) \notag\end{aligned}$$which, when substituted into Eq.(\[BBGKY1\]) gives the Enskog approximation to the one-body distribution function. The factor of $g\left( \overrightarrow{q}_{1},\overrightarrow{q}_{2};t\right) $, the spatial pair distribution function, allows for spatial correlations which always exist. In the Revised Enskog Theory, it is approximated by the equilibrium functional of the density evaluated for the local density field of the fluid[@RET]. The same approximation can be above to give the corresponding Enskog-approximation to the balance equations. A final consequence follows from the second equation of the BBGKY hierarchy which has the form$$\begin{aligned} &&\left( \frac{\partial }{\partial t}+\overrightarrow{v}_{1}\cdot \frac{% \partial }{\partial \overrightarrow{q}_{1}}+\overrightarrow{v}_{2}\cdot \frac{\partial }{\partial \overrightarrow{q}_{2}}+\overline{T}% _{-}(12)\right) f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) \label{BBGKY2} \\ &=&-n\int d3\;\left( \overline{T}_{-}(13)+\overline{T}_{-}(23)\right) f_{l_{1}l_{2}l_{3}}\left( x_{1},x_{2},x_{3};t\right) . \notag\end{aligned}$$Since atoms cannot interpenetrate the two-body distribution must have the form $f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) =\Theta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) y_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) $ for some function $y_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) $ which is continuous at $q_{12}=\sigma _{l_{1}l_{2}}$. Then, one expects that the singular terms in Eq.(\[BBGKY2\]), arising from the gradient acting on the step function and from the definition of $\overline{T}_{-}(12)$, must cancel, gives the constraint$$\overrightarrow{v}_{12}\cdot \widehat{q}_{12}\delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) =-\overline{T}% _{-}(12)f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right)$$and some rearrangement, together with the approximation of Eq.(\[BBGKY2\]) gives$$\begin{aligned} \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) f_{l_{1}l_{2}}\left( x_{1},x_{2};t\right) &\simeq &\delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) g\left( \overrightarrow{q}_{1},\overrightarrow{q}% _{2};t\right) f_{l_{1}}\left( x_{1};t\right) f_{l_{2}}\left( x_{2};t\right) \\ &&-\left( \overrightarrow{v}_{12}\cdot \widehat{q}_{12}\right) ^{-1}% \overline{T}_{-}(12)g\left( \overrightarrow{q}_{1},\overrightarrow{q}% _{2};t\right) f_{l_{1}}\left( x_{1};t\right) f_{l_{2}}\left( x_{2};t\right) \notag\end{aligned}$$which expresses the two-body distribution function at contact in terms of a completely uncorrelated piece, the first term on the right, and a correction that takes into account velocity correlations generated by the collision, the second term on the right. This can be used to evaluate two-body correlations at the Enskog level of approximation[@Lutsko96; @Lutsko2001; @LutskoHCS]. Chapman-Enskog Solution ======================= In the previous Section, the exact balance equations were developed and the Enskog approximation introduced. Next, this framework is used to derive the explicit equations governing the evolution of the hydrodynamic fields by means of the Chapman-Enskog approximation. As noted in the Introduction, previous studies of the kinetic theory for reacting systems have often made the assumption that the chemical reactions are *slow* relative to the hydrodynamic time-scales. The primary goal here is draw out, and make more precise, the meaning of this condition by outlining the Chapman-Enskog procedure under different assumptions about the speed of the chemical reactions. Before beginning, note that the phrase ”hydrodynamic fields” usually refers to those local fields which are conserved in the long-wavelength limit (which is to say that their sum over the entire system is conserved). For a non-reacting fluid of hard-spheres, this means the local partial number densities, and the momentum and energy densities. For a reacting fluid, the partial number densities are not conserved and for models of endo-/exothermic reactions, even the energy is not be conserved. Following the practice developed in the study of granular fluids (which are non-reactive but do not conserve energy) it seems natural to expand the definition of ”hydrodynamic” fields to include those fields which would be conserved in the limit of vanishing reaction probabilities. A partial justification for this is that all of these fields are necessary to develop a meaningful description of the non-reacting fluid, so one expects that they must also be included in any description of the reacting fluid (i.e., a minimal-coupling argument based on continuity of the description with respect to the control parameters). The Chapman-Enskog procedure attempts to construct a so-called normal solution of the Enskog equation which is to say, a solution which is a local functional of the (exact) hydrodynamic fields and for which all of the space and time dependence occurs implicitly through those fields[@McLennan]. This implies that the space and time derivatives of the distribution function can be written in terms of the corresponding derivatives of the fields and the functional derivative of the distribution with respect to the fields. In other words, one has$$f_{a}\left( \overrightarrow{q}_{1},\overrightarrow{v}_{1},t\right) =f_{a}% \left[ \overrightarrow{v}_{1}|x_{i}\left( \overrightarrow{q}_{1},t\right) ,n\left( \overrightarrow{q}_{1},t\right) ,\overrightarrow{u}\left( \overrightarrow{q}_{1},t\right) ,T\left( \overrightarrow{q}_{1},t\right) % \right] \label{norm1}$$so all of the dependence on $\overrightarrow{q}_{1}$ and $t$ occur through the hydrodynamic fields so that the time derivative can be expressed as$$\frac{\partial }{\partial t}f_{a}\left( \overrightarrow{q}_{1},% \overrightarrow{v}_{1},t\right) =\sum_{i}\frac{\partial x_{i}}{\partial t}% \frac{\partial f_{a}}{\partial x_{i}}+\frac{\partial n}{\partial t}\frac{% \partial f_{a}}{\partial n}+\frac{\partial \overrightarrow{u}}{\partial t}% \cdot \frac{\partial f_{a}}{\partial \overrightarrow{u}}+\frac{\partial T}{% \partial t}\frac{\partial f_{a}}{\partial T}. \label{norm2}$$Then, the kinetic equation determines the functional dependence of the distribution on the fields and their derivatives, while the fields are in turn fixed self-consistently by the balance equations (in the Enskog approximation). A further approximation which is made in practical calculations is to assume that spatial gradients are small so that the equations can be solved perturbatively via a gradient expansion. To order the terms, one introduces a uniformity parameter $\epsilon $ and replace $\overrightarrow{\nabla }$ with $\epsilon \overrightarrow{\nabla }$ and order terms in $\epsilon $. Since the space and time derivatives are related by the balance equations, one also introduces an expansion of the time derivative $\frac{\partial }{% \partial t}\equiv \partial _{t}=\partial _{t}^{(0)}+\epsilon \partial _{t}^{(1)}+...$ as well as of the distribution itself $$f_{a}\left( \overrightarrow{q}_{1},\overrightarrow{v}_{1},t\right) =f_{a}^{0}% \left[ \overrightarrow{v}_{1}|x_{i},n,\overrightarrow{u},T\right] +\epsilon f_{a}^{1}\left[ \overrightarrow{v}_{1}|x_{i},n,\overrightarrow{u},T\right] +...$$where the notation indicates that the distribution is a functional of the hydrodynamic fields. These expansions are substituted into both the Enskog equation and the balance equations and an order-by-order solution is sought. Writing the Enskog equation as$$\left( \frac{\partial }{\partial t}+\overrightarrow{v}_{1}\cdot \frac{% \partial }{\partial \overrightarrow{q}_{1}}\right) f_{a}\left( x_{1};t\right) =\sum_{bcd}J_{ab,cd}\left[ f_{c},f_{d}\right]$$so that$$\begin{aligned} J_{ab,cd}\left[ f_{c},f_{d}\right] &=&\int d\overrightarrow{q}_{2}d% \overrightarrow{v}_{2}\left[ \left| \frac{\partial \left( \widehat{b}% _{ab}^{cd}x_{1},\widehat{b}_{ab}^{cd}x_{2}\right) }{\partial \left( x_{1},x_{2}\right) }\right| ^{-1}\left( \widehat{b}_{ab}^{cd}\right) ^{-1}K_{ab}^{cd}\left( x_{12}\right) -\delta _{ac}\delta _{bd}\right] \\ &&\times \Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}% _{12}\right) \delta \left( q_{12}-\sigma _{cd}\right) \overrightarrow{v}% _{12}\cdot \widehat{q}_{12}f_{c}\left( \overrightarrow{q}_{1},% \overrightarrow{v}_{1};t\right) f_{d}\left( \overrightarrow{q}_{2},% \overrightarrow{v}_{2};t\right) \notag\end{aligned}$$it is also necessary to expand the non-locality of the collision operator which comes from the term $\delta \left( q_{12}-\sigma _{cd}\right) =\delta \left( q_{12}\right) +\epsilon \sigma _{cd}\delta ^{\prime }\left( q_{12}\right) +...$where the derivative of the delta-function, which will give rise to spatial gradients of the distribution, are scaled with an appropriate factor of $\epsilon $. In order to control the speed of the chemistry relative to the hydrodynamics, the non-diagonal part of the reaction probabilities is separated out as$$K_{ab}^{cd}\rightarrow \delta _{ac}\delta _{bd}+\epsilon ^{\alpha }\left( K_{ab}^{cd}-\delta _{ac}\delta _{bd}\right) \label{Kelastic}$$giving$$J_{ab,cd}\left[ f_{c},f_{d}\right] =\delta _{ac}\delta _{bd}J_{a,b}^{(invariat)}\left[ f_{a},f_{b}\right] +\epsilon ^{\alpha }J_{ab,cd}^{(reactive)}\left[ f_{c},f_{d}\right]$$where the non-reactive, or invariant, part is the usual collision operator for non-reactive (but possible energy non-conserving) multi-component fluids$$\begin{aligned} J_{a,b}^{(invariant)}\left[ f_{a},f_{b}\right] &=&\int d\overrightarrow{q}% _{2}d\overrightarrow{v}_{2}\left[ \left| \frac{\partial \left( \widehat{b}% _{ab}^{ab}x_{1},\widehat{b}_{ab}^{ab}x_{2}\right) }{\partial \left( x_{1},x_{2}\right) }\right| ^{-1}\left( \widehat{b}_{ab}^{ab}\right) ^{-1}-1% \right] \\ &&\times \Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}% _{12}\right) \delta \left( q_{12}-\sigma _{ab}\right) \overrightarrow{v}% _{12}\cdot \widehat{q}_{12}f_{a}\left( \overrightarrow{q}_{1},% \overrightarrow{v}_{1};t\right) f_{b}\left( \overrightarrow{q}_{2},% \overrightarrow{v}_{2};t\right) \notag\end{aligned}$$and, as indicated, the reactive part of the collision operator will be arbitrarily treated as being of order $\alpha $ in the gradient expansion. Thus, the full expansion of the collision operator will take the form$$\begin{aligned} J_{ab,cd}\left[ f_{c},f_{d}\right] &=&\delta _{ac}\delta _{bd}\left( J_{a,b}^{(invariant)0}\left[ f_{a},f_{b}\right] +\epsilon J_{a,b}^{(invariant)1}\left[ f_{a},f_{b}\right] +...\right) \\ &&+\epsilon ^{\alpha }\left( J_{ab,cd}^{(reactive)0}\left[ f_{c},f_{d}\right] +\epsilon J_{ab,cd}^{(reactive)1}\left[ f_{c},f_{d}\right] +...\right) . \notag\end{aligned}$$ Zeroth-order ------------ The zeroth order equation for the distribution is then$$\partial _{t}^{0}f_{a}^{(0)}=\sum_{b}J_{a,b}^{(elastic)0}\left[ f_{a}^{0},f_{b}^{0}\right] +\delta _{\alpha 0}\sum_{bcd}J_{ab,cd}^{(reactive)0}\left[ f_{c}^{0},f_{d}^{0}\right] \label{zeroth}$$which must be supplemented by the corresponding equations for the fields expanded to zeroth order $$\begin{aligned} \partial _{t}^{0}x_{l} &=&\delta _{\alpha 0}n^{-1}S_{l}^{(n)(reactive)0} \label{balance0} \\ \partial _{t}^{0}n &=&0 \notag \\ \partial _{t}^{0}\overrightarrow{u} &=&0 \notag \\ \partial _{t}^{0}T &=&\xi ^{(invariant)0}+\delta _{\alpha 0}\xi ^{(reactive)0}. \notag\end{aligned}$$These balance equations, together with the assumption of normality, eq. ([norm2]{}), serve to define the meaning of the term $\partial _{t}^{0}f_{a}^{(0)}$ in eq.(\[zeroth\]). Note that the fluxes do not enter, being of first order in the gradients, and that the sources are separated into a non-reactive and reactive part using eq.(\[Kelastic\]). For the concentration, mass and velocity fields, there are in general no non-reactive contributions to the sources whereas for the temperature, there is the possibility of such a contribution in which case, one recovers the inelastic hard-sphere system used to model granular fluids. Furthermore, use has been made of the fact that $f_{a}^{(0)}$ must be a function of $% \overrightarrow{v}_{1}-\overrightarrow{u}$ which implies that $\overline{% \overrightarrow{S}}^{(p)(reactive)0}=0$ (since there are no other zeroth-order vectors) so that no source can appear in the velocity equation, at this order. These zeroth-order equations illustrate a complication that occurs for fast reacting systems (e.g. $\alpha =0$) compared to non-reacting multi-component systems: namely that the sources in the balance equations at order $n$ require knowledge of the $n$-th order distribution. For non-reacting elastic systems, the $n$-th order balance equations generally require only the $n-1$ order distribution so that there is no coupling between the two. Non-reacting inelastic systems, i.e. granular fluids, share this complication as can be seen from the appearance of the source term $\xi ^{(non)0}$ in Eqs.(\[balance0\]). For $\alpha >0$, only the temperature can have a zeroth-order time dependence and so can contribute to the left side of eq.(\[zeroth\]). If this temperature source is zero then the left hand side of eq.(\[zeroth\]) is zero and the $f_{a}^{(0)}$ will simply be proportional to a Maxwellian. The solution of eqs.(\[zeroth\]) and (\[balance0\]) for the case $\alpha >0$ and the non-reactive source in the temperature equation being non-zero corresponds to the so-called homogeneous cooling state in granular fluids and has been discussed in detail in the literature for single-component[ErnstHCS]{} and multi-component systems[@HCS_Mix]. First-order ----------- At first order, one has$$\begin{aligned} &&\partial _{t}^{0}f_{a}^{(1)}+\left( \partial _{t}^{1}+\overrightarrow{v}% _{1}\cdot \overrightarrow{\nabla }\right) f_{a}^{0} \\ &=&\sum_{b}\left( J_{a,b}^{(invariant)0}\left[ f_{a}^{0},f_{b}^{1}\right] +J_{a,b}^{(invariant)0}\left[ f_{a}^{1},f_{b}^{0}\right] +J_{a,b}^{(invariant)1}\left[ f_{a}^{0},f_{b}^{0}\right] \right) \notag \\ &&+\delta _{\alpha 0}\sum_{bcd}\left( J_{ab,cd}^{(reactive)0}\left[ f_{c}^{0},f_{d}^{1}\right] +J_{ab,cd}^{(reactive)0}\left[ f_{c}^{1},f_{d}^{0}% \right] +J_{ab,cd}^{(reactive)1}\left[ f_{c}^{0},f_{d}^{0}\right] \right) \notag \\ &&+\delta _{\alpha 1}\sum_{bcd}J_{ab,cd}^{(reactive)0}\left[ f_{c}^{0},f_{d}^{0}\right] \notag\end{aligned}$$and for the fields$$\begin{aligned} \partial _{t}^{1}x_{l}+\overrightarrow{u}\cdot \overrightarrow{\nabla }x_{l} &=&\delta _{\alpha 0}n^{-1}S_{l}^{(n)(reactive)1} \label{balance1} \\ &&+\delta _{\alpha 1}n^{-1}S_{l}^{(n)(reactive)0} \notag \\ \partial _{t}^{1}n+\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}% n\right) &=&0 \notag \\ \partial _{t}^{1}\overrightarrow{u}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\overrightarrow{u}+\rho ^{-1}\overrightarrow{\nabla }p^{(0)} &=&\delta _{\alpha 0}\rho ^{-1}\overline{\overrightarrow{S}}^{(p)(0)} \notag \\ \partial _{t}^{1}T+\overrightarrow{u}\cdot \overrightarrow{\nabla }T+\frac{2% }{Dnk_{B}}\left[ p^{(0)}\overrightarrow{\nabla }\cdot \overrightarrow{u}+% \overrightarrow{\nabla }\cdot \left( \overrightarrow{u}w^{(0)}\right) \right] &=&\xi ^{(invariant)1} \notag \\ &&+\delta _{\alpha 0}\xi ^{(reactive)1}+\delta _{\alpha 1}\xi ^{(reactive)0} \notag\end{aligned}$$where we have used the fact that at zeroth order there are no velocity-independent vectors and only the unit tensor available so that we must have $\overleftrightarrow{P}^{(0)}=p^{(0)}\overleftrightarrow{1}$, $% \overleftrightarrow{W}^{(0)}=w^{(0)}\overleftrightarrow{1}$ and all vector fluxes must vanish. In general, the first order distribution must take the form$$f_{l}^{(1)}\left( \overrightarrow{r},\overrightarrow{V};t\right) =nx_{l}\phi _{l}\left( \overrightarrow{V}\right) \left[ \begin{array}{c} h_{l}(\overrightarrow{V})+\mathcal{A}_{l}(\overrightarrow{V})\overrightarrow{% V}\cdot \overrightarrow{\nabla }n+\mathcal{B}_{l}\overrightarrow{V}\cdot \overrightarrow{\nabla }T \\ +\overleftrightarrow{\mathcal{C}}_{l}:\left( \overrightarrow{\nabla }% \overrightarrow{V}-\frac{1}{D}\overleftrightarrow{1}\overrightarrow{\nabla }% \cdot \overrightarrow{V}\right) + \\ \mathcal{D}_{l}\overrightarrow{\nabla }\cdot \overrightarrow{V}+\sum_{k}% \mathcal{E}_{lk}\overrightarrow{V}\cdot \overrightarrow{\nabla }x_{k}% \end{array}% \right] \label{dist1}$$where I have written the zeroth order distribution in the form $% f_{l}^{(0)}=nx_{l}\phi _{l}\left( V\right) $. Here the coefficients $% \mathcal{A}_{l},\mathcal{B}_{l},...$are scalar functions of the velocity (and in general depend also on space and time through a dependence on the local hydrodynamic variables as does the zeroth-order distribution, although for the sake of conciseness this dependence has been suppressed). The function $h(V)$ represents the first-order correction to $\phi \left( V\right) $ due to the energy-dependent chemical reactions: for example, if the only allowed interaction were $A+A\rightarrow A+B$ and this only took place if the CM kinetic energy were greater than some threshold, $E_{AB}$, then starting with a system of all $A$-type atoms, one would expect to build up a preponderance of fast $B$ atoms and a corresponding deficit of fast $A$ atoms. It vanishes in the case that $\alpha >2$ and energy is conserved by the non-reactive dynamics. The consequences of different orderings of the reaction terms will be considered separately. ### Ultra-slow reactions: $\protect\alpha >2$ In this case, there are no reactive terms in the first order equations. The solution is therefore the same as for the equilibrium (or HCS)multi-component system. The second order balance equations will also have no reactive terms. Summing up to second order, the Navier-Stokes order balance equations are then$$\begin{gathered} \frac{\partial }{\partial t}x_{l}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }x_{l}+n^{-1}\left[ \overrightarrow{\nabla }\cdot \overrightarrow{j}% _{l}-x_{l}\overrightarrow{\nabla }\cdot \sum_{l}\overrightarrow{j}_{l}^{K}% \right] =0 \label{NS} \\ \frac{\partial }{\partial t}n+\overrightarrow{\nabla }\cdot \overrightarrow{u% }n+\overrightarrow{\nabla }\cdot \sum_{l}\overrightarrow{j}_{l}^{K}=0 \notag \\ \frac{\partial }{\partial t}\overrightarrow{u}+\overrightarrow{u}\cdot \overrightarrow{\nabla }\overrightarrow{u}+\rho ^{-1}\left( \overrightarrow{% \nabla }\cdot \overleftrightarrow{P}-\overrightarrow{u}\overrightarrow{% \nabla }\cdot \overrightarrow{Q}\right) =0 \notag \\ \frac{\partial }{\partial t}T+\overrightarrow{u}\cdot \overrightarrow{\nabla }T-\frac{T}{n}\overrightarrow{\nabla }\cdot \sum_{l}\overrightarrow{j}% _{l}^{K}+\frac{2}{Dnk_{B}}\left[ \overleftrightarrow{P}:\overrightarrow{% \nabla }\overrightarrow{u}+\overrightarrow{\nabla }\cdot \overrightarrow{q}+% \overrightarrow{\nabla }\cdot \left( \overrightarrow{u}\cdot \overleftrightarrow{W}\right) +\frac{1}{2}\overrightarrow{Q}\cdot \overrightarrow{\nabla }u^{2}\right] =\xi ^{(invariant)}\end{gathered}$$ where the fluxes are the sum of zeroth- and first-order contributions,$% \overleftrightarrow{P}=\overleftrightarrow{P}^{(0)}+\overleftrightarrow{P}% ^{(1)}$, and the source $\xi $ consists of (non-reactive) contributions summed through second order. This means that for the granular case, $\xi ^{(invariant)}\neq 0$, the Navier-Stokes order balance equations require knowledge of the second-order (or Burnett order) distribution function. There is, at this order, no coupling between the hydrodynamic equations and the reaction equations. Inserted into Eq.(\[NS\]), the result would be the Navier-Stokes equations for elastic hard-spheres, or their generalization for inelastic hard spheres. If this expansion is continued, the $\alpha -th$ order balance equation for the concentrations would be$$\partial ^{\left( \alpha \right) }x_{l}+n^{-1}\left[ \overrightarrow{\nabla }% \cdot \overrightarrow{j}_{l}^{(\alpha -1)}-x_{l}\overrightarrow{\nabla }% \cdot \sum_{l}\overrightarrow{j}_{l}^{K(\alpha -1)}\right] =n^{-1}S_{l}^{(n)(reactive)0}.$$Clearly, the reaction equation remains unknown in this case since one would need to consistently include the higher order hydrodynamic contributions that would come from the number current which, in turn, would bring in couplings to higher-order gradients of the hydrodynamic fields. Without knowledge of these higher order terms (and they are not known for even the one-component fluid) the reaction equation can only be consistently studied in the *absence of hydrodynamic gradients* when the reactive terms are treated as of order $\alpha >2$. ### Slow reactions:$\;\protect\alpha =2$ In this case, the first order solution is again the same as in the non-reacting case. However, the sources will have second-order contributions so that the Navier-Stokes equations take the form$$\begin{gathered} \frac{\partial }{\partial t}x_{l}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }x_{l}+n^{-1}\left[ \overrightarrow{\nabla }\cdot \overrightarrow{j}% _{l}-x_{l}\overrightarrow{\nabla }\cdot \sum_{l}\overrightarrow{j}_{l}^{K}% \right] =n^{-1}S_{l}^{(n)(reactive)0} \label{phenom} \\ \frac{\partial }{\partial t}n+\overrightarrow{\nabla }\cdot \overrightarrow{u% }n+\overrightarrow{\nabla }\cdot \sum_{l}\overrightarrow{j}_{l}^{K}=0 \notag \\ \frac{\partial }{\partial t}\overrightarrow{u}+\overrightarrow{u}\cdot \overrightarrow{\nabla }\overrightarrow{u}+\rho ^{-1}\left( \overrightarrow{% \nabla }\cdot \overleftrightarrow{P}-\overrightarrow{u}\overrightarrow{% \nabla }\cdot \overrightarrow{Q}\right) =\rho ^{-1}\overline{\overrightarrow{% S}}^{(p)(reactive)0} \notag \\ \frac{\partial }{\partial t}T+\overrightarrow{u}\cdot \overrightarrow{\nabla }T-\frac{T}{n}\overrightarrow{\nabla }\cdot \sum_{l}\overrightarrow{j}% _{l}^{K}+\frac{2}{Dnk_{B}}\left[ \overleftrightarrow{P}:\overrightarrow{% \nabla }\overrightarrow{u}+\overrightarrow{\nabla }\cdot \overrightarrow{q}+% \overrightarrow{\nabla }\cdot \left( \overrightarrow{u}\cdot \overleftrightarrow{W}\right) +\frac{1}{2}\overrightarrow{Q}\cdot \overrightarrow{\nabla }u^{2}\right] \notag \\ =\xi ^{(invariant)}+\xi ^{(reactive)0}.\end{gathered}$$For the simplest case that the reactions conserve energy and momentum, the reactions are governed by exactly the convective-reaction-diffusion equation that one might expect. The reaction rates are calculated using the local equilibrium distribution as in elementary treatments[@McQuarry]. Except for the usual modification of the transport properties arising from the use of the Enskog equation, as opposed to the Boltzmann equation, there are no new dense-fluid effects. ### Moderate reactions: $\protect\alpha =1$ For moderately fast reactions, the situation becomes more interesting. Considering here only the case that mass and energy are conserved by all collisions, the first order balance equations - the generalization of the Euler equations - are found to be $$\begin{aligned} \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) x_{l} &=&n^{-1}S_{l}^{(n)(reactive)0} \\ \frac{\partial }{\partial t}n+\overrightarrow{\nabla }\cdot \overrightarrow{u% }n &=&0 \notag \\ \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) \overrightarrow{u}+\rho ^{-1}\overrightarrow{\nabla }p^{(0)} &=&0 \notag \\ \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) T+\frac{2}{Dnk_{B}}p\overrightarrow{\nabla }\cdot \overrightarrow{u} &=&0. \notag\end{aligned}$$so that the reactions, with reaction rates calculated from the local-equilibrium distribution function, enter into the Euler equations. The Navier-Stokes equations will involve the reaction rates calculated up to first order in the distribution. In general, the only nonzero coupling in the reaction source will take the form $S_{l}^{(n)(reactive)1}=S_{l}^{(n)}% \overrightarrow{\nabla }\cdot \overrightarrow{u}$ where $S_{l}^{(n)}$ is a scalar function of the concentrations, density and temperature. The Navier-Stokes equations will therefore take the form $$\begin{aligned} \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) x_{l}+n^{-1}\left( \overrightarrow{\nabla }\cdot \overrightarrow{j}_{l}-x_{l}\overrightarrow{\nabla }\cdot \sum_{l}% \overrightarrow{j}_{l}^{K}\right) & =n^{-1}S_{l}^{(n)(reactive)0} \label{reaction1}\\ & +n^{-1}S_{l}^{(n)(reactive)1}+S_{l}^{(n)}\overrightarrow{\nabla }\cdot \overrightarrow{u} \notag \\ \frac{\partial }{\partial t}n+\overrightarrow{\nabla }\cdot \overrightarrow{u% }n+\overrightarrow{\nabla }\cdot \sum_{l}\overrightarrow{j}_{l}^{K}& =0 \notag \\ \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) \overrightarrow{u}+\rho ^{-1}\overrightarrow{\nabla }\cdot \overleftrightarrow{P}& =0 \notag \\ \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) T-\frac{T}{n}\overrightarrow{\nabla }\cdot \sum_{l}% \overrightarrow{j}_{l}^{K}+\frac{2}{Dnk_{B}}\left( \overleftrightarrow{P}:% \overrightarrow{\nabla }\overrightarrow{u}+\overrightarrow{\nabla }\cdot \overrightarrow{q}\right) & =0. \notag\end{aligned}$$The source term for the reactions has three contributions: the zeroth order reaction rate (calculated using the local equilibrium distribution function), the first order correction (due to deviations of the distribution from local equilibrium) and a new, dense-fluid effect which couples the reactions to the divergence of the velocity field with some field-dependent coefficient,$S_{l}^{(n)}$ . This coupling is a dense-fluid effect which does not exist in the Boltzmann approximation and not surprisingly, its origin is closely related to that of the bulk-viscosity which is also zero in the Boltzmann theory, but not the Enskog theory. (The calculation of these terms will be discussed in detail in a future publication but the fact that these are the only possible couplings is due to the fact that no other galilean-invariant scalars, linear in the gradients of the fields, can be constructed). ### Fast reactions: $\protect\alpha =0$ In the case of fast reactions, no a priori assumption is made about the speed of the reactions compared to the hydrodynamic time scales. The balance equations to first order, i.e. the Euler equations, are $$\begin{aligned} \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) x_{l} &=&n^{-1}S_{l}^{(n)(reactive)0}+S_{l}^{(n)}% \overrightarrow{\nabla }\cdot \overrightarrow{u} \\ \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) n+n\overrightarrow{\nabla }\cdot \overrightarrow{u} &=&0 \notag \\ \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) \overrightarrow{u}+\rho ^{-1}\overrightarrow{\nabla }p &=&0 \notag \\ \left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) T+\frac{2}{Dnk_{B}}\left[ p\overrightarrow{\nabla }\cdot \overrightarrow{u}+\overrightarrow{\nabla }\cdot \left( w\overrightarrow{u}% \right) \right] &=&\xi ^{(invariant)0}+\xi ^{(reactive)0}+\xi ^{(invariant)1}+\xi ^{(reactive)1} \notag\end{aligned}$$so that even the Euler equations show the dense-fluid correction to the reaction rates. The second-order, or Navier-Stokes, equations require the evaluation of the source terms to second order, which in turn requires knowledge of the distribution function to second order (also called Burnett order). One then expects that even for mass and energy conserving interactions, the reaction equation will contain couplings to gradients of all of the hydrodynamic fields. However, since the complete Burnett-order Chapman-Enskog solution of the Enskog equation is not even known for the case of a single-component fluid, there is no practical value in continuing the analysis for this case. Conclusion ========== In this paper, the kinetic theory of reactive hard-core systems has been extended to include the possibility of mass transfer and/or loss and energy gain/loss. When mass is not conserved, the collision rule becomes dependent on the model used to describe the lost mass. Nevertheless, quite general expressions for the dynamics of phase functions and the distribution function of the system can be given and used to derive equally general expressions for the exact balance laws for mass, momentum, energy and concentration. For example, when mass is conserved but a fixed fraction of the rest-frame kinetic energy is lost during collisions, the usual inelastic hard sphere kinetic theory, used as a model of granular fluids[HCSLiouville]{}, is recovered. The kinetic theory was used, within the Enskog approximation, to discuss the various phenomenological laws, extensions of the Navier-Stokes equations, that arise from different orderings of the reaction terms within the Chapman-Enskog procedure. It was noted that the intuitive model of the Navier-Stokes equations coupled to a reaction-diffusion equation through a convective term only arises when the ”speed” of the chemical reactions is comparable to some hydrodynamic time scale and even in this case, an additional coupling to the divergence of the velocity field can occur in dense fluids. How fast do we expect the chemistry to be relative to the hydrodynamics? In the model considered here, chemical reactions cannot be faster than the collision time. In fact, a typical reaction rate would be something like $% p\left( \delta x\right) e^{-E/k_{B}T}\nu _{col}$ where $p$ is the probability of a reaction occuring if the colliding atoms have energy greater than the reaction energy barrier, $E$, $\delta x$ is the difference between the concentration of the species and its equilibrium concentration and $\nu _{col}$ is the collision frequency. On the other hand, the Chapman-Enskog procedure is based on a gradient expansion, the small parameter $\epsilon $ will generally be a measure of the ratio of the typical microscopic length scale, the mean free path $l_{mfp},$ to a typical length scale for hydrodynamic gradients $L$. (In Fourier space, where gradients $\overrightarrow{\nabla }$ correspond to wavevectors $% \overrightarrow{k}$, this becomes $\epsilon \sim kl_{mfp}$.) So, setting $% pe^{-E/k_{B}T}\nu _{col}\sim \left( kl_{mfp}\right) ^{\alpha }\nu _{col}\sim \left( l_{mfp}/L\right) ^{\alpha }\nu _{col}$ gives $$\alpha \sim \frac{\ln p\left( \delta x\right) -\frac{E}{k_{B}T}}{\ln \left( l_{mfp}/L\right) }.$$For systems in which hydrodynamics is applicable, one has $l_{mfp}/L<<1$ so that $\alpha $ ranges from a minumum of $\frac{\ln p\left( \delta x\right) }{% \ln \left( l_{mfp}/L\right) }\geq 0$, for $k_{B}T\gg E$, to very large values for low temperatures. Far from chemical equilibrium, $\delta x\sim 1$, the lower limit could be arbitrarily close to zero depending on the reaction probability so that ”moderate” and ”fast” reactions are possible at high temperatures. Indeed, if all of these parameters are fixed, then fast reactions will always occur in the hydodrynamic regime limit $% l_{mfp}/L\rightarrow 0$. The conclusion is that, unless the concentrations are close to their equilibrium values, the reaction probabilities are very small or the temperature is extremely low, the concept of slow chemical reactions may be of limited applicability and, so, the correct phenomenological description, from the standpoint of kinetic theory, may be more complex than the reaction-diffusion-advection model. In summary,if chemical reactions are slow compared to the rate of dissipation in the fluid, then hydrodynamics and chemistry are not meaningfully coupled. If the reaction rate is comparable to the rate of dissipation in the fluid, i.e. $\lambda k^{2}$ for some transport coefficient $\lambda $ and wavevector $k$, then the usual reaction-diffusion-advection equation results. For faster reactions, additional couplings occur and the chemistry and hydrodynamics become more interdependent. The detailed solution of the Enskog, and the resulting phenomenological equations for particular reaction models will be the subject of a future publication where the importance for sonochemistry of additional terms, such as those occurring in Eq.(\[reaction1\]), will be investigated. The work presented here has benefited from discussions with Jim Dufty, Ray Kapral, Jean-Pierre Boon and Pierre Gaspard. Some financial support was received from the Université Libre de Bruxelles. The hard-core Liouville operator {#AppT} ================================ The goal in this appendix is to provide motivation for the statement in the text that the form of the pseudo-Liouville operator is independent of the collision rule. To start with, restrict attention to a system of 2 atoms. Let $X\left( \Gamma ;t\right) $ be the characteristic function for collisions after at time $t$ beginning with the phase $\Gamma $ at time $0$ so that if the two atoms do *not* collide between during the interval $% [0,t]$ then $X\left( \Gamma ;t\right) =0$ whereas if they *do* collide, $X\left( \Gamma ;t\right) =1$. Then the time evolution of the phase function $A_{\Gamma }\left( t\right) =A\left( \Gamma (t),t\right) $ is given by$$A_{\Gamma }\left( t\right) =\left( 1-X\left( \Gamma ;t\right) \right) A\left( \Gamma _{0}(t),t\right) +X\left( \Gamma ;t\right) A\left( \Gamma ^{\prime }(t),t\right)$$where $\Gamma _{0}(t)$ is just the phase of the system propagated a time $t$ into the future in the absence of interactions and is explicitly $\Gamma _{0}(t)=\left( \overrightarrow{q}_{1}+\overrightarrow{v}_{1}t,% \overrightarrow{v}_{1},\widehat{l}_{1},\overrightarrow{q}_{2}+% \overrightarrow{v}_{2}t,\overrightarrow{v}_{2},\widehat{l}_{2}\right) $. (Note that attention is restricted to the case that velocities are constant during free-streaming: generalization to include one-body forces is straightforward.) The phase point $\Gamma ^{\prime }(t)$ is the position the system would reach in phase space if a collision occurred at some time $\tau \left( \Gamma \right) $ $\in $ $\left[ 0,t\right] $. Explicit expressions can also be given for its components such as $\overrightarrow{q}_{1}^{\prime }(t)=\overrightarrow{q}_{1}+\overrightarrow{v}_{1}\tau +\overrightarrow{v}% _{1}^{\prime }\left( t-\tau \right) $, etc. Direct differentiation then gives$$\begin{aligned} \frac{dA_{\Gamma }\left( t\right) }{dt} &=&\frac{\partial A}{\partial t}% +\left( 1-X\left( \Gamma ;t\right) \right) \frac{\partial q(t)}{\partial t}% \cdot \frac{\partial }{\partial q(t)}A\left( \Gamma _{0}(t),t\right) \\ &&+X\left( \Gamma ;t\right) \frac{\partial q(t)}{\partial t}\cdot \frac{% \partial }{\partial q(t)}A\left( \Gamma ^{\prime }(t),t\right) \notag \\ &&+\frac{dX\left( \Gamma ;t\right) }{dt}\left( A\left( \Gamma ^{\prime }(t),t\right) -A\left( \Gamma _{0}(t),t\right) \right) . \notag\end{aligned}$$Now, from the definitions above$$\begin{aligned} &&\left( 1-X\left( \Gamma ;t\right) \right) \frac{\partial q(t)}{\partial t}% \cdot \frac{\partial }{\partial q(t)}A\left( \Gamma _{0}(t),t\right) +X\left( \Gamma ;t\right) \frac{\partial q(t)}{\partial t}\cdot \frac{% \partial }{\partial q(t)}A\left( \Gamma ^{\prime }(t),t\right) \\ &=&\left( 1-X\left( \Gamma ;t\right) \right) \sum_{i=1,2}\overrightarrow{v}% _{i}\cdot \frac{\partial }{\partial \overrightarrow{q}_{i}(t)}A\left( \Gamma _{0}(t),t\right) +X\left( \Gamma ;t\right) \sum_{i=1,2}\overrightarrow{v}% _{i}^{\prime }\cdot \frac{\partial }{\partial \overrightarrow{q}_{i}(t)}% A\left( \Gamma ^{\prime }(t),t\right) \notag \\ &=&\sum_{i=1,2}\overrightarrow{v}_{i}(t)\cdot \frac{\partial }{\partial \overrightarrow{q}_{i}(t)}A\left( \Gamma (t),t\right) \notag\end{aligned}$$giving$$\frac{dA_{\Gamma }\left( t\right) }{dt}=\frac{\partial A}{\partial t}% +\sum_{i=1,2}\overrightarrow{v}_{i}(t)\cdot \frac{\partial }{\partial \overrightarrow{q}_{i}(t)}A_{\Gamma }\left( t\right) +\frac{dX\left( \Gamma ;t\right) }{dt}\left( A\left( \Gamma ^{\prime }(t),t\right) -A\left( \Gamma _{0}(t),t\right) \right)$$Now, since $X\left( \Gamma ;t\right) $ has the form of a step function (it is zero if $t<\tau \left( \Gamma \right) $ and one otherwise), we must have$$\frac{dX\left( \Gamma ;t\right) }{dt}=\delta \left( t-\tau \left( \Gamma \right) \right) , \label{delta}$$and this also gives the correct result (zero) if $\tau \left( \Gamma \right) $ is imaginary (indicating that no collision ever occurs starting from the given state). Then, using$$\begin{aligned} \delta \left( t-\tau \left( \Gamma \right) \right) \left( A\left( \Gamma ^{\prime }(t),t\right) -A\left( \Gamma _{0}(t),t\right) \right) &=&\delta \left( t-\tau \left( \Gamma \right) \right) \left( A\left( \Gamma ^{\prime }(\tau \left( \Gamma \right) ),\tau \left( \Gamma \right) \right) -A\left( \Gamma _{0}(\tau \left( \Gamma \right) ),\tau \left( \Gamma \right) \right) \right) \\ &=&\delta \left( t-\tau \left( \Gamma \right) \right) \left( \sum_{l_{1}^{\prime }l_{2}^{\prime }}\widehat{M}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{b}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) A_{\Gamma }\left( \tau \left( \Gamma \right) \right) \notag \\ &=&\delta \left( t-\tau \left( \Gamma \right) \right) \left( \sum_{l_{1}^{\prime }l_{2}^{\prime }}\widehat{M}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{b}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) A_{\Gamma }\left( t\right) \notag\end{aligned}$$gives$$\frac{dA_{\Gamma }\left( t\right) }{dt}=\left[ \frac{\partial }{\partial t}% +\sum_{i=1,2}\overrightarrow{v}_{i}(t)\cdot \frac{\partial }{\partial \overrightarrow{q}_{i}(t)}+\delta \left( t-\tau \left( \Gamma \right) \right) \left( \sum_{l_{1}^{\prime }l_{2}^{\prime }}\widehat{M}% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{b}% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) \right] A_{\Gamma }\left( t\right)$$I order to express the right hand side entirely in terms of $\Gamma (t)$ rather than the initial condition $\Gamma $, the temporal delta-function is rewritten using $$\delta \left( q_{12}\left( t\right) -\sigma _{l_{1}l_{2}}\right) =\sum_{i}% \frac{\delta \left( t-\tau _{i}\left( \Gamma \right) \right) }{\left| \frac{% \partial }{\partial t}q_{12}\left( t\right) \right| _{t=\tau _{i}\left( \Gamma \right) }}$$where $\tau _{i}\left( \Gamma \right) $ are the roots of $q_{12}^{2}\left( \tau _{i}\right) -\sigma _{l_{1}l_{2}}^{2}=0$. They correspond to the time at which the two atoms are first in contact, i.e. the physical collision time which is denoted as $\tau \left( \Gamma \right) $, and the time at which they are last in contact if they are allowed to pass through one another (which is not physical). One picks out the correct root by noting that at the physical collision time $\overrightarrow{q}_{12}\left( \tau \right) \cdot \overrightarrow{v}_{12}(\tau )<0$ while the sign is reversed at the unphysical collision time so that$$\delta \left( q_{12}\left( t\right) -\sigma _{l_{1}l_{2}}\right) \Theta \left( -\overrightarrow{q}_{12}\left( t\right) \cdot \overrightarrow{v}% _{12}(t)\right) =\frac{\delta \left( t-\tau \left( \Gamma \right) \right) }{% \left| \frac{\partial }{\partial t}q_{12}\left( t\right) \right| _{t=\tau \left( \Gamma \right) }}=\frac{\delta \left( t-\tau \left( \Gamma \right) \right) }{\left| \widehat{q}_{12}\left( \tau \right) \cdot \overrightarrow{v}% _{12}(\tau )\right| }=\frac{\delta \left( t-\tau \left( \Gamma \right) \right) }{\left| \widehat{q}_{12}\left( t\right) \cdot \overrightarrow{v}% _{12}(t)\right| }$$or$$\delta \left( t-\tau \left( \Gamma \right) \right) =\delta \left( q_{12}\left( t\right) -\sigma _{l_{1}l_{2}}\right) \Theta \left( -% \overrightarrow{q}_{12}\left( t\right) \cdot \overrightarrow{v}% _{12}(t)\right) \left| \widehat{q}_{12}\left( t\right) \cdot \overrightarrow{% v}_{12}(t)\right|$$giving finally$$\frac{dA_{\Gamma }\left( t\right) }{dt}=\left[ \frac{\partial }{\partial t}+% \widehat{L}(t)\right] A_{\Gamma }\left( t\right) \label{L1}$$with$$\widehat{L}(t)=\sum_{1\leq i\leq 2}\overrightarrow{v}_{i}(t)\cdot \frac{% \partial }{\partial \overrightarrow{q}_{i}(t)}+\sum_{1\leq i<j\leq 2}% \widehat{T}_{+}(12;t) \label{L2}$$where, for arbitrary phase function $B\left( \Gamma ,t\right) $, $$\begin{aligned} \widehat{T}_{+}(12;t)B\left( \Gamma (t),t\right) &=&\delta \left( q_{12}(t)-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\overrightarrow{q}% _{12}(t)\cdot \overrightarrow{v}_{12}(t)\right) \left| \widehat{q}% _{12}(t)\cdot \overrightarrow{v}_{12}(t)\right| \\ &&\times \left( \sum_{l_{1}^{\prime }l_{2}^{\prime }}\widehat{M}% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\widehat{b}% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) B\left( \Gamma (t),t\right) . \notag\end{aligned}$$For more than two atoms, one simply extends the sums in eq.(\[L2\]) since, in a finite system, only binary collisions can occur. Starting with an initial condition $A_{\Gamma }\left( t\right) =A\left( \Gamma \right) $, iteration of eq.(\[L1\]) immediately gives$$\left. \frac{d^{n}A_{\Gamma }\left( t\right) }{dt^{n}}\right| _{t=0}=% \widehat{L}^{n}A\left( \Gamma \right)$$with $\widehat{L}=\widehat{L}(0)$ which implies that$$A_{\Gamma }\left( t\right) =\exp \left( \widehat{L}t\right) A\left( \Gamma \right)$$and$$\frac{d}{dt}A_{\Gamma }\left( t\right) =\widehat{L}\exp \left( \widehat{L}% t\right) A\left( \Gamma \right) =\widehat{L}A_{\Gamma }\left( t\right)$$as claimed in the text. The adjoint Liouville operator {#AppA} ============================== To derive the adjoint operator, begin with its definition$$\int d\Gamma \;B\left( \Gamma \right) L_{+}A\left( \Gamma \right) =\int d\Gamma \;\left( L_{+}^{A}B\left( \Gamma \right) \right) A\left( \Gamma \right)$$or, more explicitly,$$\sum_{l_{1},l_{2}...}\int dx_{1}dx_{2}...\;\;B\left( \Gamma \right) L_{+}A\left( \Gamma \right) =\sum_{l_{1},l_{2}}\int dx_{1}dx_{2}\;\;\left( L_{+}^{A}B\left( \Gamma \right) \right) A\left( \Gamma \right)$$Now, $$L_{+}=L_{+}^{(0)}+\sum_{i<j}T_{+}(ij)$$and it is obvious that, neglecting surface terms,$$\int d\Gamma \;B\left( \Gamma \right) L_{+}^{(0)}A\left( \Gamma \right) =\int d\Gamma \;\left( -L_{+}^{(0)}B\left( \Gamma \right) \right) A\left( \Gamma \right)$$so$$L_{+}^{(0)A}=-L_{+}^{(0)}.$$ Next, consider one of the collision operators and restrict attention to a system of two atoms. Then $$\begin{aligned} &&\sum_{l_{1},l_{2}}\int dx_{1}dx_{2}\;\;B\left( x_{1},l_{1};x_{2},l_{2}\right) \left[ T_{+}(12)A\left( x_{1},l_{1};x_{2},l_{2}\right) \right] \\ &=&-\sum_{l_{1},l_{2},a,b}\int dx_{1}dx_{2}\;B\left( x_{1},l_{1};x_{2},l_{2}\right) \Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) \delta \left( q_{12}-\sigma _{12}\right) \overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12} \notag \\ &&\times \left( K_{l_{1}l_{2}}^{ab}\left( x_{1},l_{1};x_{2},l_{2}\right) A(% \widehat{b}_{l_{1}l_{2}}^{ab}x_{1},a;\widehat{b}_{l_{1}l_{2}}^{ab}x_{2},b)-% \delta _{al_{1}}\delta _{bl_{2}}A\left( x_{1},l_{1};x_{2},l_{2}\right) \right) \notag\end{aligned}$$ Consider the first term. Relabeling the species in the sum gives$$\begin{aligned} &&-\sum_{l_{1},l_{2},a,b}\int dx_{1}dx_{2}\;B\left( x_{1},l_{1};x_{2},l_{2}\right) \Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12} \notag \\ &&\times K_{l_{1}l_{2}}^{ab}\left( x_{1},l_{1};x_{2},l_{2}\right) A(\widehat{% b}_{l_{1}l_{2}}^{ab}x_{1},a;\widehat{b}_{l_{1}l_{2}}^{ab}x_{2},b) \\ &=&-\sum_{l_{1},l_{2},a,b}\int dx_{1}dx_{2}\;B\left( x_{1},a;x_{2},b\right) \Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) \delta \left( q_{12}-\sigma _{ab}\right) \overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12} \notag \\ &&\times K_{ab}^{l_{1}l_{2}}\left( x_{1},a;x_{2},b\right) A(\widehat{b}% _{ab}^{l_{1}l_{2}}x_{1},l_{1};\widehat{b}_{ab}^{l_{1}l_{2}}x_{2},l_{2}). \notag\end{aligned}$$Assuming that the collision operator is invertible, then introducing new integration variables $y_{i}=\widehat{b}_{ab}^{l_{1}l_{2}}x_{i}$ and the corresponding Jacobian$$J_{ab}^{l_{1}l_{2}}\left( y_{1},y_{2}\right) =\left| \frac{\partial \left( \left( \widehat{b}_{ab}^{l_{1}l_{2}}\right) ^{-1}y_{1},\left( \widehat{b}% _{ab}^{l_{1}l_{2}}\right) ^{-1}y_{2}\right) }{\partial \left( y_{1},y_{2}\right) }\right|$$gives$$\begin{aligned} &&-\sum_{l_{1},l_{2},a,b}\int dx_{1}dx_{2}\;B\left( x_{1},l_{1};x_{2},l_{2}\right) \Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12} \\ &&\times K_{l_{1}l_{2}}^{ab}\left( x_{1},l_{1};x_{2},l_{2}\right) A(\widehat{% b}_{l_{1}l_{2}}^{ab}x_{1},a;\widehat{b}_{l_{1}l_{2}}^{ab}x_{2},b) \notag \\ &=&-\sum_{l_{1},l_{2},a,b}\int dy_{1}dy_{2}\;J_{ab}^{l_{1}l_{2}}\left( y_{1},y_{2}\right) B\left( \left( \widehat{b}_{ab}^{l_{1}l_{2}}\right) ^{-1}y_{1},a;\left( \widehat{b}_{ab}^{l_{1}l_{2}}\right) ^{-1}y_{2};b\right) \Theta \left( -\left( \widehat{b}_{ab}^{l_{1}l_{2}}\right) ^{-1}% \overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) \notag \\ &&\times \delta \left( q_{12}-\sigma _{ab}\right) \left( \left( \widehat{b}% _{ab}^{l_{1}l_{2}}\right) ^{-1}\overrightarrow{v}_{12}\right) \cdot \overrightarrow{q}_{12}K_{ab}^{l_{1}l_{2}}\left( \left( \widehat{b}% _{ab}^{l_{1}l_{2}}\right) ^{-1}y_{1},a;\left( \widehat{b}_{ab}^{l_{1}l_{2}}% \right) ^{-1}y_{2},b\right) A(y_{1},l_{1};y_{2},l_{2}) \notag \\ &=&-\sum_{l_{1},l_{2},a,b}\int dy_{1}dy_{2}\;A(y_{1},l_{1};y_{2},l_{2}) \notag \\ &&\times \left[ J_{ab}^{l_{1}l_{2}}\left( y_{1},y_{2}\right) \left( \widehat{% b}_{ab}^{l_{1}l_{2}}\right) ^{-1}\Theta \left( -\overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \overrightarrow{v}_{12}\cdot \overrightarrow{q}% _{12}K_{ab}^{l_{1}l_{2}}\left( y_{1},l_{1};y_{2},l_{1}\right) \right] \notag \\ &&\times B\left( x_{1},l_{1};x_{2},l_{2}\right) \notag\end{aligned}$$where the operator $\left( \widehat{b}_{ab}^{l_{1}l_{2}}\right) ^{-1}$ has the effect of changing the species from $l_{1},l_{2}$ to $a,b$. One can then write $$\begin{aligned} &&T_{+}^{A}\left( 12\right) B\left( x_{1},l_{1};x_{2},l_{2}\right) \\ &=&-\sum_{ab}\left[ J_{ab}^{l_{1}l_{2}}\left( x_{1},x_{2}\right) \left( \widehat{b}_{ab}^{l_{1}l_{2}}\right) ^{-1}K_{ab}^{l_{1}l_{2}}\left( x_{1},l_{1};x_{2},l_{2}\right) -1\right] \Theta \left( -\overrightarrow{v}% _{12}\cdot \overrightarrow{q}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \overrightarrow{v}_{12}\cdot \overrightarrow{q}_{12} \notag \\ &&\times B\left( x_{1},l_{1};x_{2},l_{2}\right) . \notag\end{aligned}$$In some cases of interest, the collision dynamics may not be invertible. For example, suppose that collisions with total rest frame energy less than some threshold, $E$, are elastic while those with energy greater than this are inelastic. Then, a pair of atoms with rest frame energy after collision of $% \frac{1}{2}\mu _{12}v_{12}^{\prime 2}<E$ might have resulted from either (a) a collision between two atoms with rest-frame energy below the threshold or (b) a collision between two atoms that had energy above the threshold but that lost part of this due to the inelastic process. In this case, it is necessary in the definition of the adjoint operator to include an additional sum over the various branches of the inverse collision dynamics. Even when it occurs, such a complication may not be of practical importance since it is often the case that expressions involving the adjoint operator $T_{+}^{A}$ can be rewritten in terms of the original operator $T_{+}$. Derivation of the balance equations {#AppB} =================================== In this section, the general form of the local balance equations is derive and then specialized to the partial density, momentum and energy fields. General form of the balance equations ------------------------------------- Consider any one-body phase function of the form$$\widehat{\Psi }_{l}\left( \overrightarrow{r}\right) =\sum_{i}\psi _{l_{i}}\left( \overrightarrow{v}_{i}\right) \delta \left( \overrightarrow{r}% -\overrightarrow{q}_{i}\right) \delta _{l_{i}l}$$and its average$$\begin{aligned} \Psi _{l}\left( \overrightarrow{r},t\right) &=&\left\langle \widehat{\Psi }% _{l}\left( \overrightarrow{r}\right) ;t\right\rangle =\frac{N}{V}% \sum_{l_{1}}\int dx_{1}\;f_{l_{1}}\left( x_{1},t\right) \psi _{l_{i}}\left( \overrightarrow{v}_{1}\right) \delta \left( r-\overrightarrow{q}_{1}\right) \delta _{l_{1}l} \\ &=&n\int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},% \overrightarrow{v}_{1},t\right) \psi _{l}\left( \overrightarrow{v}_{1}\right) \notag\end{aligned}$$The balance equation for this follows from the first BBGKY equation $$\left( \frac{d}{dt}+\overrightarrow{v}_{1}\cdot \frac{\partial }{\partial \overrightarrow{q}_{1}}\right) f_{l_{1}}\left( x_{1}\right) =\sum_{l_{2}}\int dx_{2}\overline{T}_{-}\left( 12\right) f_{l_{1}l_{2}}\left( x_{1}x_{2}\right)$$and is$$\begin{aligned} &&\frac{d}{dt}\Psi _{l}\left( \overrightarrow{r},t\right) +\overrightarrow{% \nabla }\cdot \int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},% \overrightarrow{v}_{1},t\right) \overrightarrow{v}_{1}\psi _{l}\left( \overrightarrow{v}_{1}\right) \\ &=&\sum_{l_{1}l_{2}}\int dx_{1}\;\delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}_{1}\right) \delta \left( r-\overrightarrow{q}_{1}\right) \int dx_{2}\overline{T}_{-}\left( 12\right) f_{ll_{2}}\left( x_{1}x_{2}\right) . \notag\end{aligned}$$Introducing the specific velocity, $\overrightarrow{V}_{1}\left( \overrightarrow{r},t\right) =\overrightarrow{v}_{1}-\overrightarrow{u}\left( \overrightarrow{r},t\right) $, the second term on the left becomes$$\begin{aligned} &&\int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},% \overrightarrow{v}_{1},t\right) \overrightarrow{v}_{1}\psi _{l}\left( \overrightarrow{v}_{1}\right) \\ &=&\overrightarrow{u}\left( \overrightarrow{r},t\right) \Psi _{l}\left( \overrightarrow{r},t\right) +\int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1},t\right) \overrightarrow{V}% _{1}\left( \overrightarrow{r},t\right) \psi _{l}\left( \overrightarrow{v}% _{1}\right) \notag\end{aligned}$$while it proves more convenient to rewrite the right hand side in terms of the $\widehat{T}_{+}$ collision operator $$\begin{aligned} &&\sum_{l_{1}l_{2}}\int dx_{1}\;\delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}_{1}\right) \delta \left( r-\overrightarrow{q}_{1}\right) \int dx_{2}\overline{T}_{-}\left( 12\right) f_{ll_{2}}\left( x_{1}x_{2}\right) \\ &=&\sum_{l_{1}l_{2}}\int dx_{1}dx_{2}f_{ll_{2}}\left( x_{1}x_{2}\right) T_{+}\left( 12\right) \delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}% _{1}\right) \delta \left( r-\overrightarrow{q}_{1}\right) \notag\end{aligned}$$so that the balance equation becomes$$\begin{aligned} &&\frac{d}{dt}\Psi _{l}\left( \overrightarrow{r},t\right) +\overrightarrow{% \nabla }\cdot \overrightarrow{u}\left( \overrightarrow{r},t\right) \Psi _{l}\left( \overrightarrow{r},t\right) +\overrightarrow{\nabla }\cdot \int d% \overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}% _{1},t\right) \overrightarrow{V}_{1}\psi _{l}\left( \overrightarrow{v}% _{1}\right) \\ &=&\sum_{l_{1}l_{2}}\int dx_{1}dx_{2}f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) \delta \left( r-\overrightarrow{q}_{1}\right) T_{+}\left( 12\right) \delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}_{1}\right) \notag\end{aligned}$$with$$\begin{aligned} &&T_{+}\left( 12\right) \delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v% }_{1}\right) \\ &=&-\overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \left( \sum_{l_{1}^{\prime }l_{2}^{\prime }}K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) \delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}_{1}\right) \notag \\ &=&-\overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \sum_{l_{1}^{\prime }l_{2}^{\prime }}\left( K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-\delta _{l_{1}l_{1}^{\prime }}\delta _{l_{2}l_{2}^{\prime }}\right) \delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}_{1}\right) . \notag\end{aligned}$$In general, the right hand side can be separated into a sum of a flux and a source term. Let$$B_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) =\left( K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-\delta _{l_{1}l_{1}^{\prime }}\delta _{l_{2}l_{2}^{\prime }}\right) \delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}_{1}\right)$$and define its even and odd components as$$\begin{aligned} F_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) &=&% \frac{1}{2}\left( B_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) -B_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{2},x_{1}\right) \right) \\ &=&\frac{1}{2}\left( K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-\delta _{l_{1}l_{1}^{\prime }}\delta _{l_{2}l_{2}^{\prime }}\right) \left( \delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}_{1}\right) -\delta _{ll_{2}}\psi _{l_{2}}\left( \overrightarrow{v}_{2}\right) \right) \notag \\ S_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) &=&% \frac{1}{2}\left( B_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) +B_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{2},x_{1}\right) \right) \notag \\ &=&\frac{1}{2}\left( K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-\delta _{l_{1}l_{1}^{\prime }}\delta _{l_{2}l_{2}^{\prime }}\right) \left( \delta _{ll_{1}}\psi _{l_{1}}\left( \overrightarrow{v}_{1}\right) +\delta _{ll_{2}}\psi _{l_{2}}\left( \overrightarrow{v}_{2}\right) \right) \notag\end{aligned}$$so that $$\begin{aligned} &&\frac{d}{dt}\Psi _{l}\left( \overrightarrow{r},t\right) +\overrightarrow{% \nabla }\cdot \overrightarrow{u}\left( \overrightarrow{r},t\right) \Psi _{l}\left( \overrightarrow{r},t\right) \\ &&+\overrightarrow{\nabla }\cdot \sum_{l}\int d\overrightarrow{v}% _{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1},t\right) \overrightarrow{V}_{1}\psi _{l}\left( \overrightarrow{v}_{1}\right) \notag \\ &=&-\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) \overrightarrow{q}% _{12}\cdot \overrightarrow{v}_{12}\delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \notag \\ &&\times \left( F_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) +S_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) \right) . \notag\end{aligned}$$Then, relabel the dummy variables to give$$\begin{aligned} &&\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) \overrightarrow{q}% _{12}\cdot \overrightarrow{v}_{12}\delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \\ &&\times \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) F_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) \notag \\ &=&\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}% _{12}\cdot \overrightarrow{v}_{12}\right) F_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) \notag \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) \left[ \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) -\delta \left( \overrightarrow{r}-\overrightarrow{q}_{2}\right) \right] \notag\end{aligned}$$where use has been made of the the asymmetry of $F_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) $ and of the symmetry of the distribution under an interchange of atoms. Finally, write$$\begin{aligned} \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) -\delta \left( \overrightarrow{r}-\overrightarrow{q}_{2}\right) &=&\int_{0}^{1}dx\;% \frac{d}{dx}\delta \left( \overrightarrow{r}-x\overrightarrow{q}_{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) \\ &=&-\overrightarrow{\nabla }\cdot \overrightarrow{q}_{12}\int_{0}^{1}dx\;% \delta \left( \overrightarrow{r}-x\overrightarrow{q}_{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) \notag\end{aligned}$$so that the balance equation becomes$$\frac{d}{dt}\Psi _{l}\left( \overrightarrow{r},t\right) +\overrightarrow{% \nabla }\cdot \overrightarrow{u}\left( \overrightarrow{r},t\right) \Psi _{l}\left( \overrightarrow{r},t\right) +\overrightarrow{\nabla }\cdot \overrightarrow{F}_{l}\left( \overrightarrow{r},t\right) =S_{l}\left( \overrightarrow{r},t\right) \label{Balance}$$with the flux written as $\overrightarrow{F}_{l}\left( \overrightarrow{r}% ,t\right) =\overrightarrow{F}_{l}^{K}\left( \overrightarrow{r},t\right) +% \overrightarrow{F}_{l}^{V}\left( \overrightarrow{r},t\right) $ where the kinetic contribution is $$\overrightarrow{F}^{K}\left( \overrightarrow{r},t\right) =\int d% \overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}% _{1},t\right) \overrightarrow{V}_{1}\psi _{l}\left( \overrightarrow{v}% _{1}\right)$$and the collisional contribution is$$\begin{aligned} \overrightarrow{F}^{V}\left( \overrightarrow{r},t\right) &=&-\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;% \overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -% \widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) F_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}_{1}-\left( 1-x\right) \overrightarrow{% q}_{2}\right) \notag\end{aligned}$$and the source is$$\begin{aligned} S_{l}\left( \overrightarrow{r},t\right) &=&-\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) S_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) \delta \left( \overrightarrow{r}% -\overrightarrow{q}_{1}\right) . \notag\end{aligned}$$ Local number density -------------------- Setting $\psi _{l}\left( \overrightarrow{v}_{1}\right) =1$ one has that$$\begin{aligned} F_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) &=&% \frac{1}{2}\left( K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \left( \delta _{ll_{1}^{\prime }}-\delta _{ll_{2}^{\prime }}\right) -\delta _{l_{1}l_{1}^{\prime }}\delta _{l_{2}l_{2}^{\prime }}\left( \delta _{ll_{1}}-\delta _{ll_{2}}\right) \right) \\ S_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) &=&% \frac{1}{2}\left( K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \left( \delta _{ll_{1}^{\prime }}+\delta _{ll_{2}^{\prime }}\right) -\delta _{l_{1}l_{1}^{\prime }}\delta _{l_{2}l_{2}^{\prime }}\left( \delta _{ll_{1}}+\delta _{ll_{2}}\right) \right) \notag\end{aligned}$$From the normalization condition $$1=\sum_{lb}K_{l_{1}l_{2}}^{lb}$$one has that$$\begin{aligned} \sum_{l_{1}^{\prime }l_{2}^{\prime }}F_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) &=&\frac{1}{2}% \sum_{l_{1}^{\prime }l_{2}^{\prime }}K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \left( \delta _{ll_{1}^{\prime }}-\delta _{ll_{2}^{\prime }}-\delta _{ll_{1}}+\delta _{ll_{2}}\right) \\ \sum_{l_{1}^{\prime }l_{2}^{\prime }}S_{l_{1}l_{2},l_{1}^{\prime }l_{2}^{\prime };l}\left( x_{1},x_{2}\right) &=&\frac{1}{2}% \sum_{l_{1}^{\prime }l_{2}^{\prime }}K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \left( \delta _{ll_{1}^{\prime }}-\delta _{ll_{2}^{\prime }}+\delta _{ll_{1}}-\delta _{ll_{2}}\right) \notag\end{aligned}$$so that the balance equation becomes$$\frac{d}{dt}n_{l}+\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}% n_{l}\right) +\overrightarrow{\nabla }\cdot \overrightarrow{j}% _{l}=S_{l}^{(n)}$$with the source$$\begin{aligned} S_{l}^{(n)} &=&-\frac{1}{2}\sum_{abl_{1}l_{2}}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) K_{l_{1}l_{2}}^{ab}\left( x_{12}\right) \left( \delta _{al}+\delta _{bl}-\delta _{ll_{1}}-\delta _{ll_{2}}\right) \notag\end{aligned}$$and the number current $\overrightarrow{j}_{l}=\overrightarrow{j}_{l}^{K}+% \overrightarrow{j}_{l}^{V}$ with$$\overrightarrow{j}_{l}^{K}=\int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1},t\right) \overrightarrow{V}_{1}.$$and$$\begin{aligned} \overrightarrow{j}_{l}^{V} &=&-\frac{1}{2}\sum_{abl_{1}l_{2}}\int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{ab}\left( x_{12}\right) \left( \delta _{al}-\delta _{bl}-\delta _{ll_{1}}+\delta _{ll_{2}}\right) \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x% \overrightarrow{q}_{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) . \notag\end{aligned}$$ The balance equations for total number and mass density follow immediately. Summing over $l$ gives$$\frac{d}{dt}n\left( \overrightarrow{r},t\right) +\overrightarrow{\nabla }% \cdot \overrightarrow{u}\left( \overrightarrow{r},t\right) n\left( \overrightarrow{r},t\right) +\overrightarrow{\nabla }\cdot \sum_{l}% \overrightarrow{j}_{l}^{K}=0$$since the sum of the collisional contributions to the number current vanishes. Similarly, multiplying by $m_{l}$ and then summing gives the balance equation for local mass density$$\frac{d}{dt}\rho \left( \overrightarrow{r},t\right) +\overrightarrow{\nabla }% \cdot \overrightarrow{u}\left( \overrightarrow{r},t\right) \rho \left( \overrightarrow{r},t\right) +\overrightarrow{\nabla }\cdot \overrightarrow{Q}% =S^{(\rho )}.$$Here, the anomalous mass flux is $$\begin{aligned} \overrightarrow{Q} &=&-\frac{1}{2}\sum_{abl_{1}l_{2}}\int dx_{1}dx_{2}\;% \overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -% \widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{ab}\left( x_{12}\right) \left( m_{a}-m_{b}-m_{l_{1}}+m_{l_{2}}\right) \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) \notag\end{aligned}$$and the mass source term is$$\begin{aligned} S_{l}^{(\rho )} &=&\frac{1}{2}\sum_{abl_{1}l_{2}}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) \delta \left( \overrightarrow{r}-\overrightarrow{q}_{1}\right) K_{l_{1}l_{2}}^{ab}\left( x_{12}\right) \delta m_{l_{1}l_{2}}^{ab}, \notag\end{aligned}$$which is only nonzero if the collisions do not conserve mass. Momentum density ---------------- Taking $\psi _{l}\left( \overrightarrow{v}_{1}\right) =m_{l}\overrightarrow{v% }_{1}=\overrightarrow{p}_{1}$ in Eq.(\[Balance\]) and summing over $l$ gives$$\frac{\partial }{\partial t}\rho \overrightarrow{u}+\overrightarrow{\nabla }% \cdot \left( \rho \overrightarrow{u}\overrightarrow{u}\right) +% \overrightarrow{\nabla }\cdot \left( \overleftrightarrow{P}+\overrightarrow{Q% }\overrightarrow{u}\right) =\overrightarrow{S}^{(p)}$$with $\overleftrightarrow{P}=\overleftrightarrow{P}^{K}+\overleftrightarrow{P% }^{V}+\overleftrightarrow{P}^{M}$ where the kinetic contribution is $$\overleftrightarrow{P}^{K}=\sum_{l}m_{l}\int d\overrightarrow{v}% _{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1},t\right) \overrightarrow{V}_{1}\overrightarrow{V}_{1}.$$To explicitly write the remaining flux and source terms, we need$$\left( b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\overrightarrow{p}_{1}-% \overrightarrow{p}_{1}\right) =\frac{1}{2}\left( \overrightarrow{\gamma }% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\overrightarrow{V}_{12}\right)$$giving$$\begin{aligned} S_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} &=&-\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( \overrightarrow{V}% _{12}-\overrightarrow{u}\right) -\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\overrightarrow{u} \\ F_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} &=&\frac{1}{2}% \overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}. \notag\end{aligned}$$At this point, it is useful to separate $\overrightarrow{\gamma }% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}$ into two parts: its value in the local rest-frame of the colliding atoms, $\widetilde{\overrightarrow{% \gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}$, and the part coming from the Galilean transformation to the lab frame$$\overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}=% \widetilde{\overrightarrow{\gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}+\left( m_{l_{1}^{\prime }}-m_{l_{1}}-m_{l_{2}^{\prime }}-m_{l_{2}}\right) \overrightarrow{V}_{12}$$so that we have that the collisional part of the flux is$$\begin{aligned} \overleftrightarrow{P}^{V} &=&-\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \widetilde{\overrightarrow{\gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x% \overrightarrow{q}_{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) \notag\end{aligned}$$and the contribution from the instantaneous exchange of mass is$$\begin{aligned} \overleftrightarrow{P}^{M} &=&-\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \left( \overrightarrow{V}_{12}-\overrightarrow{u}\right) \left( m_{l_{1}^{\prime }}-m_{l_{1}}-m_{l_{2}^{\prime }}-m_{l_{2}}\right) \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}_{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) \notag\end{aligned}$$and the source can be written as$$\begin{aligned} \overrightarrow{S}^{(p)} &=&\overrightarrow{u}S^{(\rho )}+\overline{% \overrightarrow{S}}^{(p)} \\ \overline{\overrightarrow{S}}^{(p)} &=&\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{% \prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}% _{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \notag \\ &&\times \left( \overrightarrow{V}-\overrightarrow{u}\right) f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\delta \left( \overrightarrow{r}-\overrightarrow{q}% _{1}\right) \notag\end{aligned}$$By using the balance equation for total mass density$$\frac{d}{dt}\rho \overrightarrow{u}+\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}\rho \overrightarrow{u}\right) =\rho \frac{d}{dt}% \overrightarrow{u}+\rho \overrightarrow{u}\cdot \overrightarrow{\nabla }% \overrightarrow{u}-\overrightarrow{u}\overrightarrow{\nabla }\cdot \overrightarrow{Q}+\overrightarrow{u}S^{(\rho )}$$we can write $$\frac{\partial }{\partial t}\overrightarrow{u}+\overrightarrow{u}\cdot \overrightarrow{\nabla }\overrightarrow{u}+\rho ^{-1}\left( \overrightarrow{% \nabla }\cdot \overleftrightarrow{P}+\overrightarrow{Q}\cdot \overrightarrow{% \nabla }\overrightarrow{u}\right) =\rho ^{-1}\overline{\overrightarrow{S}}% ^{(p)}\text{.}$$ Energy ------ Taking $\psi _{l}\left( \overrightarrow{v}_{1}\right) =\frac{1}{2}% m_{l}v_{1}^{2}$ in Eq.(\[Balance\]) and summing over $l$ gives$$\frac{d}{dt}E+\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}% E\right) +\overrightarrow{\nabla }\cdot \overrightarrow{F}=S^{(E)}$$where the kinetic part of the flux is$$\begin{aligned} \overrightarrow{F}^{K} &=&\sum_{l}\frac{1}{2}m_{l}\int d\overrightarrow{v}% _{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1},t\right) \overrightarrow{V}_{1}v_{1}^{2} \label{E2} \\ &=&\sum_{l}\frac{1}{2}m_{l}\int d\overrightarrow{v}_{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1},t\right) \overrightarrow{V}% _{1}\left( \overrightarrow{V}_{1}+\overrightarrow{u}\right) ^{2} \\ &=&\overrightarrow{q}^{K}+\overrightarrow{u}\cdot \overleftrightarrow{P}^{K} \notag\end{aligned}$$with the kinetic contribution to the heat flux being defined as$$\overrightarrow{q}^{K}\equiv \sum_{l}\frac{1}{2}m_{l}\int d\overrightarrow{v}% _{1}\;f_{l}\left( \overrightarrow{r},\overrightarrow{v}_{1},t\right) \overrightarrow{V}_{1}V_{1}^{2}.$$ The source term comes from the even part of the collision kernel$$\begin{aligned} S_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} &=&\frac{1}{2}\left( b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) \left( \frac{1}{% 2m_{l_{1}}}p_{1}^{2}+\frac{1}{2m_{l_{2}}}p_{2}^{2}\right) \\ &=&\frac{1}{2}\left[ \frac{1}{2m_{l_{1}^{\prime }}}p_{1}^{\prime 2}+\frac{1}{% 2m_{l_{2}^{\prime }}}p_{2}^{\prime 2}-\frac{1}{2m_{l_{1}}}p_{1}^{2}-\frac{1}{% 2m_{l_{2}}}p_{2}^{2}\right] \notag \\ &=&-\frac{1}{2}\left[ \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}+\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}V^{2}\right] \notag \\ &=&-\frac{1}{2}\left[ \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}+\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( \overrightarrow{V}-\overrightarrow{u}\right) ^{2}% \right] -\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( \overrightarrow{V}-\overrightarrow{u}\right) \cdot \overrightarrow{u% }-\frac{1}{4}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}u^{2} \notag\end{aligned}$$so that $$S^{(E)}=\xi +\overrightarrow{u}\cdot \overline{\overrightarrow{S}}^{(p)}+% \frac{1}{2}u^{2}\overline{S}^{(\rho )}$$with the (rest-frame) source term$$\begin{aligned} \xi &=&\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -% \widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times \left[ \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}+\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( \overrightarrow{V}-\overrightarrow{u}\right) ^{2}% \right] f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \delta \left( \overrightarrow{r}-% \overrightarrow{q}_{1}\right) . \notag\end{aligned}$$ The flux comes from the odd part of the collision kernel $$\begin{aligned} F_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} &=&\frac{1}{2}\left( b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) \left( \frac{1}{% 2m_{l_{1}}}p_{1}^{2}-\frac{1}{2m_{l_{2}}}p_{2}^{2}\right) \\ &=&\frac{1}{2}\left( b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) \left( \frac{1}{m_{l_{1}}+m_{l_{2}}}\left( p_{1}^{2}-p_{2}^{2}\right) -\frac{m_{l_{1}}-m_{l_{2}}}{m_{l_{1}}+m_{l_{2}}}% \left( \frac{1}{2m_{l_{1}}}p_{1}^{2}+\frac{1}{2m_{l_{2}}}p_{2}^{2}\right) \right) \notag\end{aligned}$$The first term gives$$\begin{aligned} &&\left( b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) \frac{1}{% m_{l_{1}}+m_{l_{2}}}\left( p_{1}^{2}-p_{2}^{2}\right) \\ &=&\frac{1}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\left[ \left( \overrightarrow{p}_{1}+\frac{1}{2}\overrightarrow{\gamma }% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\overrightarrow{V}\right) ^{2}-\left( \overrightarrow{p}_{2}-\frac{1}{2}\overrightarrow{\gamma }% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\overrightarrow{V}\right) ^{2}% \right] \notag \\ &&-\frac{1}{m_{l_{1}}+m_{l_{2}}}\left( p_{1}^{2}-p_{2}^{2}\right) \notag \\ &=&\overrightarrow{V}\cdot \overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{% \prime }l_{2}^{\prime }}+\frac{\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\overrightarrow{p}% _{12}\cdot \overrightarrow{V}+\left( \frac{1}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}-\frac{1}{m_{l_{1}}+m_{l_{2}}}\right) \left( p_{1}^{2}-p_{2}^{2}\right) \notag\end{aligned}$$while the second is $$\begin{aligned} &&\left( b_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}-1\right) \frac{% m_{l_{1}}-m_{l_{2}}}{m_{l_{1}}+m_{l_{2}}}\left( \frac{1}{2m_{l_{1}}}% p_{1}^{2}+\frac{1}{2m_{l_{2}}}p_{2}^{2}\right) \\ &=&-\frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\left( \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}+\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}V^{2}\right) +\left( \frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}-\frac{% m_{l_{1}}-m_{l_{2}}}{m_{l_{1}}+m_{l_{2}}}\right) \left( \frac{1}{2m_{l_{1}}}% p_{1}^{2}+\frac{1}{2m_{l_{2}}}p_{2}^{2}\right) \notag \\ &=&-\frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\left( \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}+\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}V^{2}\right) +\left( \frac{1}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\right) \left( \frac{m_{l_{1}^{\prime }}}{m_{l_{1}}}% p_{1}^{2}-\frac{m_{l_{2}^{\prime }}}{m_{l_{2}}}p_{2}^{2}\right) \notag \\ &&-\left( \frac{1}{m_{l_{1}}+m_{l_{2}}}\right) \left( p_{1}^{2}-p_{2}^{2}\right) \notag\end{aligned}$$so$$\begin{aligned} 2F_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} &=&\overrightarrow{V}% _{12}\cdot \overrightarrow{\gamma }_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}+\frac{\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\overrightarrow{p}_{12}\cdot \overrightarrow{V}_{12}+\frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{% m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\left( \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}+\frac{1}{2}\delta m_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}V_{12}^{2}\right) \\ &&+\left( \frac{m_{l_{1}}-m_{l_{1}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\right) \frac{1}{m_{l_{1}}}p_{1}^{2}+\left( \frac{% m_{l_{2}^{\prime }}-m_{l_{2}}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}% \right) \frac{1}{m_{l_{2}}}p_{2}^{2} \notag\end{aligned}$$which gives, after some algebra,$$\begin{aligned} 2F_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} &=&\overrightarrow{V}% _{12}\cdot \widetilde{\overrightarrow{\gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}+\frac{1}{m_{l_{1}}+m_{l_{2}}}\left( \frac{% m_{l_{1}}m_{l_{2}^{\prime }}-m_{l_{2}}m_{l_{1}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\right) \mu _{l_{1}l_{2}}v^{2} \\ &&+\frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}-\frac{1}{2}\left( -m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}-m_{l_{2}}+m_{l_{1}}\right) V_{12}^{2} \notag\end{aligned}$$This can also be written as $$\begin{aligned} 2F_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} &=&\left( \overrightarrow{V}% _{12}-\overrightarrow{u}\right) \cdot \widetilde{\overrightarrow{\gamma }}% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}+\frac{m_{l_{2}^{\prime }}m_{l_{1}}-m_{l_{1}^{\prime }}m_{l_{2}}}{\left( m_{l_{2}^{\prime }}+m_{l_{1}^{\prime }}\right) \left( m_{l_{2}}+m_{l_{1}}\right) }\mu _{l_{1}l_{2}}v^{2} \\ &&+\frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}-\frac{1}{2}\left( -m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}-m_{l_{2}}+m_{l_{1}}\right) \left( \overrightarrow{V}_{12}-\overrightarrow{% u}\right) ^{2} \notag \\ &&+\left( m_{l_{1}^{\prime }}-m_{l_{1}}-m_{l_{2}^{\prime }}+m_{l_{2}}\right) \left( \left( \overrightarrow{V}_{12}-\overrightarrow{u}\right) \cdot \overrightarrow{u}\right) +\frac{1}{2}\left( m_{l_{1}^{\prime }}-m_{l_{1}}-m_{l_{2}^{\prime }}+m_{l_{2}}\right) u^{2} \\ &&+\overrightarrow{u}\cdot \widetilde{\overrightarrow{\gamma }}% _{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }} \notag\end{aligned}$$so that$$\overrightarrow{F}^{(V)}=\overrightarrow{q}^{V}+\overrightarrow{q}^{m}+% \overrightarrow{q}^{\delta E}+\overrightarrow{u}.\left( \overleftrightarrow{P% }^{V}+\overleftrightarrow{P}^{M}\right) +\frac{1}{2}u^{2}\overrightarrow{Q}$$where the different pieces of the heat flux vector are the usual collisional contribution$$\begin{aligned} \overrightarrow{q}^{V} &=&-\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \left( \overrightarrow{V}_{12}-\overrightarrow{u}\right) \cdot \widetilde{% \overrightarrow{\gamma }}_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) , \notag\end{aligned}$$a part arising from the instantaneous transfer of mass$$\begin{aligned} \overrightarrow{q}^{m} &=&-\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\frac{m_{l_{2}^{\prime }}m_{l_{1}}-m_{l_{1}^{\prime }}m_{l_{2}}}{\left( m_{l_{2}^{\prime }}+m_{l_{1}^{\prime }}\right) \left( m_{l_{2}}+m_{l_{1}}\right) } \\ &&\times \int dx_{1}dx_{2}\;\overrightarrow{q}_{12}\left( \overrightarrow{q}% _{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}_{12}\cdot \overrightarrow{v}% _{12}\right) \notag \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \mu _{l_{1}l_{2}}v^{2} \notag \\ &&\times \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) , \notag\end{aligned}$$and a part arising from the loss of energy$$\begin{aligned} \overrightarrow{q}^{\delta E} &=&-\frac{1}{2}\sum_{l_{1}l_{2}l_{1}^{\prime }l_{2}^{\prime }}\frac{m_{l_{1}^{\prime }}-m_{l_{2}^{\prime }}}{% m_{l_{1}^{\prime }}+m_{l_{2}^{\prime }}}\int dx_{1}dx_{2}\;\overrightarrow{q}% _{12}\left( \overrightarrow{q}_{12}\cdot \overrightarrow{v}_{12}\right) \delta \left( q_{12}-\sigma _{l_{1}l_{2}}\right) \Theta \left( -\widehat{q}% _{12}\cdot \overrightarrow{v}_{12}\right) \\ &&\times f_{l_{1}l_{2}}\left( x_{1}x_{2}\right) K_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}\left( x_{12}\right) \overline{\delta E_{l_{1}l_{2}}^{l_{1}^{\prime }l_{2}^{\prime }}}% \int_{0}^{1}dx\;\delta \left( \overrightarrow{r}-x\overrightarrow{q}% _{1}-\left( 1-x\right) \overrightarrow{q}_{2}\right) . \notag\end{aligned}$$A little rearrangement allows us to write the energy balance equation as$$\frac{\partial }{\partial t}E+\overrightarrow{\nabla }\cdot \left( \overrightarrow{u}E\right) +\overrightarrow{\nabla }\cdot \overrightarrow{q}+% \overrightarrow{\nabla }\cdot \left( \overrightarrow{u}\cdot \overleftrightarrow{P}\right) +\overrightarrow{\nabla }\cdot \left( \frac{1}{% 2}u^{2}\overrightarrow{Q}\right) =\xi +\overrightarrow{u}\cdot \overline{% \overrightarrow{S}}^{(p)}+\frac{1}{2}u^{2}\overline{S}^{(\rho )}.$$Alternatively, noting the relation between the total energy and the kinetic temperature$$E=\frac{D}{2}nk_{B}T+\frac{1}{2}\rho u^{2}$$gives an equation for the evolution of the kinetic temperature$$\left( \frac{\partial }{\partial t}+\overrightarrow{u}\cdot \overrightarrow{% \nabla }\right) T-\frac{T}{n}\overrightarrow{\nabla }\cdot \sum_{l}% \overrightarrow{j}_{l}^{K}+\frac{2}{Dnk_{B}}\left[ \overleftrightarrow{P}:% \overrightarrow{\nabla }\overrightarrow{u}+\overrightarrow{\nabla }\cdot \overrightarrow{q}\right] =\xi .$$

ArXiv

--- abstract: 'We investigate the enrichment of the pre-solar cloud core with short lived radionuclides (SLRs), especially . The homogeneity and the surprisingly small spread in the ratio observed in the overwhelming majority of calcium-aluminium-rich inclusions (CAIs) in a vast variety of primitive chondritic meteorites places strong constraints on the formation of the the solar system. Freshly synthesized radioactive has to be included and well mixed within $20\kyr$. After discussing various scenarios including X-winds, AGB stars and Wolf-Rayet stars, we come to the conclusion that triggering the collapse of a cold cloud core by a nearby supernova is the most promising scenario. We then narrow down the vast parameter space by considering the pre-explosion survivability of such a clump as well as the cross-section necessary for sufficient enrichment. We employ numerical simulations to address the mixing of the radioactively enriched SN gas with the pre-existing gas and the forced collapse within $20\kyr$. We show that a cold clump of $10\Msun$ at a distance of $5\pc$ can be sufficiently enriched in and triggered into collapse fast enough - within $18\kyr$ after encountering the supernova shock - for a range of different metallicities and progenitor masses, even if the enriched material is assumed to be distributed homogeneously in the entire supernova bubble. In summary, we envision an environment for the birth place of the Solar System $4.567\Gyr$ ago similar to the situation of the pillars in M16 nowadays, where molecular cloud cores adjacent to an HII region will be hit by a supernova explosion in the future. We show that the triggered collapse and formation of the Solar System as well as the required enrichment with radioactive are possible in this scenario.' author: - 'M. Gritschneder$^{1}$,D.N.C. Lin$^{1,2}$, S.D. Murray$^{3}$, Q.-Z. Yin$^{4}$, M.-N. Gong$^{5}$' bibliography: - 'references.bib' title: The Supernova Triggered Formation and Enrichment of Our Solar System --- Introduction ============ The time-scale for the formation events of our Solar System can be derived from the decay products of radioactive elements found in meteorites. Short lived radionuclides (SLRs) such as , ${\rm^{41}Ca}$, ${\rm^{53}Mn}$ and ${\rm^{60}Fe}$ can be employed as high-precision and high-resolution chronometers due to their short half-lives. These SLRs are found in a wide variety of Solar System materials, including calcium-aluminium-rich inclusions (CAIs) in primitive chondrites [e.g. @Lee:1976fk; @Amelin:2002lr]. The decay of into  in particular, with a half-life of $\tau\approx0.7\Myr$, provides by far the highest resolution measurements that mark the initial formation of the proto-planetary disk [@Jacobsen:2008uq]. As decays, the ratio of at the time of condensation can be directly derived from the present day ${\rm^{26}Mg}/{\rm^{24}Mg}$ and ${\rm^{27}Al}/{\rm^{24}Mg}$ measurements. Furthermore, the spread of among different CAIs gives the precise time-span in which these CAIs formed. While there is considerable spread among the CAIs towards lower ratios of due to remelting and thermal metamorphism, the upper value, the so called ’canonical ratio’, [@MacPherson:1995lr] is now well established [e.g. @Jacobsen:2008uq; @Villeneuve:2009fj]. The general picture we adopt here is that a certain amount of is injected in the nascent solar nebula and then gets incorporated into the earliest formed CAIs as soon as the temperature drops below the condensation temperature of CAI minerals. These CAIs are most frequently found in CV-chondrites[^1]. Lower ratios of can then be explained by subsequent episodes of CAI remelting or thermal disturbance, thereby explaining the heterogeneity below the canonical value. Therefore, the CAIs found in chondrites represent the first known solid objects that crystalized within our Solar System and can be used as an anchor point to determine the formation time-scale of our Solar System. Various measurements of different CAIs by several research groups have not only confirmed the canonical ratio of $(5.23\pm0.13)\times10^{-5}$, but also established a very small spread (for a recent review see @Villeneuve:2009fj). This spread corresponds to an age range of less than $\simeq20\kyr$ [e.g. @Jacobsen:2008uq]. In addition, the Mg-isotope composition appears to be fairly uniform among bulk chondrites as well as Mars, Moon and Earth. This is a strong indication that initially Al- and Mg-isotopes are distributed quite homogeneously ($\pm 10\%$) in the proto-planetary disk [@Thrane:2006fk; @Villeneuve:2009fj; @Boss:2011lr]. The extremely small time-span together with the highly homogeneous mixing of isotopes poses a severe challenge for theoretical models on the formation of our Solar System Various theoretical scenarios for the formation of the Solar System have been discussed. Shortly after the discovery of SLRs, it was proposed that they were injected by a nearby massive star. This can happen either via a supernova explosion [@Cameron:1977yq] or by the strong winds of a Wolf-Rayet star [@Arnould:1997vn]. Another possibility would be the in situ enrichment inside the disk by the non-thermal activity of the Sun [@Shu:2001fk]. @Wasserburg:1995fj have proposed an asymptotic giant branch (AGB) star as an possible source of SLRs in the Solar System (see also @Trigo-Rodriguez:2009kx). Up to now, none of these models have been able to explain the extremely small range of CAI condensation ages (the spread of $\approx 20\kyr$). In this paper, we discuss several possible solutions to this problem. First, we show that most of these scenarios are not compatible with an almost coeval formation of CAIs. We then go on to perform numerical simulations on the promising scenario of a type IIa supernova as the trigger of our Solar System. Physical Processes and Feasibility of Solutions {#phys_proc} =============================================== In this section we discuss the physical processes in different scenarios and their viability. For a detailed discussion of these scenarios, including an assessment of their likelihood see the recent review by @Adams:2010lr. As mentioned before, the canonical value of is established to be $(5.23\pm0.13)\times10^{-5}$. Assuming an total mass for of ${M_{\rm^{27}Al}}=7.1\times10^{-5}\Msun$ [e.g. @Lodders:2003fk] we can convert the observed ratio into a total mass of ${M_{\rm^{26}Al}^{\rm canonical}}\approx3.71\times10^{-9}\Msun$ of radioactive , which has to be present in the early Solar System. In situ Enrichment ------------------ One solution to the incorporation of into CAIs was proposed by @Shu:2001fk. Young stars can undergo the so-called X-wind phase. In this phase the combination of magnetic fields, inflowing material and stellar outflows can form a wind between the proto-star and the proto-planetary disk [@Shu:1997qy]. SRLs are produced in this scenario by solar energetic particle irradiation of dust and gas near the proto-Sun [@Shu:2001fk] and incorporated into CAIs. This mechanism can explain the ratio of ${\rm^{10}Be/^{9}Be}=8.8\pm0.6\times10^{-4}$ inferred from measurements of ${\rm^{10}B}$, the decay product of ${\rm^{10}Be}$ [@McKeegan:2000lr]. Another way to explain this ratio was proposed by @Desch:2004fk. They calculate the contribution of ${\rm^{10}Be}$ trapped by galactic cosmic rays to the collapsing molecular cloud core and conclude that a large amount, if not all, ${\rm^{10}Be}$ can be attributed to cosmic rays. For the purpose of this study it is sufficient to note that this mechanism can not explain the observed amount of due to the following reasons. First of all, and ${\rm^{10}Be}$ do not correlate well in CAIs [e.g. @Marhas:2002qy]. In addition, SLRs produced by this mechanism are intrinsically expected to be highly heterogeneous. Variations in their relative abundances would reflect the local energetic particle environment and their episodic nature of production. This is in contradiction to the observed homogeneity ($\pm10\%$) of . Furthermore, the amount of which can be produced by this mechanism is at least an order of magnitude too low to explain the canonical value [@Duprat:2007fj; @Duprat:2008uq]. External Enrichment by AGB stars -------------------------------- Another source of are AGB stars. Every star of $\sim0.8-8\Msun$ [@Herwig:2005lr] undergoes this phase when helium fusion in the core is complete and helium shell burning begins. In this phase, the star loses most of its envelope in stellar winds, which could in term enrich the proto-solar cloud[e.g. @Wasserburg:1995fj]. This stage is at the end of a stellar life. However, AGB stars are not present in star forming regions and hence an encounter is very unlikely [@Kastner:1994lr]. In addition, the total amount of requires a massive AGB-star and the short enrichment-timescale requires a brief AGB-phase. External Enrichment by Wolf-Rayet stars {#WRstars} --------------------------------------- An interesting possibility was proposed by [@Tatischeff:2010uq] who followed up on an earlier idea by [@Arnould:1997vn]. Here, a run-away Wolf-Rayet star that gets ejected from its parental cluster of massive stars sweeps up a smaller shell while traveling into the ISM. This shell gets enriched by material from the strong stellar winds of the Wolf-Rayet star. As soon as this star explodes in a supernova it sweeps up an even denser shell which expands rapidly. After $\simeq10\kyr$ the shell starts to cool and a cold core can start to form due to dynamical and gravitational instabilities which subsequently undergoes collapse. During this collapse phase the the surface density of the core increases with time [@Whitworth:1994kx; @Heitsch:2008fj]. Since this theory assumes that the ratio increases as the surface density of the clump increases, the core becomes enriched sufficiently to explain the observed values in CAIs after $10^5-10^6\yr$. However, there is no physical reason why this ratio should increase as the core density rises. The initial ratio within the shell should be conserved and only modified by radioactive decay (the decay is taken into account by @Tatischeff:2010uq). A similar scenario was proposed by [@Gaidos:2009fk]. Here, an entire molecular cloud ($M>10^5\Msun$) is enriched in by a nearby star forming region. After $\sim4-5\Myr$ this cloud begins to form stars, including the Sun. During this stage, the Solar System is enriched with heavier SLRs. This scenario can explain the different abundances of SLRs (see §\[discussion\]). However, it requires a precise chain of events. In addition, this theory assumes continuous enrichment which is at odds with the small age spread of CAIs (see below). Constraints on continuous enrichment {#contin_enrich} ------------------------------------ All three scenarios described above are based on the assumption of a continuous enrichment. In that case, the characteristic time scale for the enrichment with radioactive isotopes is much longer than the brief duration ($<20\kyr$) of the CAI formation episode which is inferred from the meteoritic data. In external enrichment models, sufficient radioactive isotopes can only be intercepted by a tenuous and extended solar-system progenitor cloud prior to its collapse. But the free fall time scale for such a cloud is generally $\sim 100\kyr$ or longer. If the central star and disk accrete material from this parent envelope on such a time scale, dust and grains would be processed with different rates at different locations and in different stages. During this entire phase, the assembly and formation of the CAIs would be possible, leading to inhomogeneities and an age-spread of $\sim 100\kyr$. This is inconsistent with the observationally inferred homogeneity of Mg- and Al-isotopes during the formation epoch of CAIs. Here, we make a clear distinction and differentiate the initial formation epoch of CAIs from later epochs when some CAIs subsequently experienced a complex thermal history within the solar nebula and metamorphism on their host meteorite parent asteroids extending over $0.1-3\Myr$ [e.g. @MacPherson:1995lr; @Hsu:2000fk; @Young:2005kx; @MacPherson:2007uq]. One possible mechanism to homogenize the isotope’s distribution and reset the clock for grain condensation and CAI formation is to heat the infalling gas above the condensation temperature for CAIs of $\sim1500-1800$K [e.g. @Grossman:2002yq; @Lodders:2003fk]. During the infall, accretion and shocks can dissipate sufficient energy and heat up the disk to this temperature close to the central star (within $\sim 0.1$ AU). In order for the infalling gas to arrive at this location and form a compact disk, it must carry very little angular momentum. However, turbulent angular momentum transport, induced either by gravitational or magneto-rotational instabilities, would lead to mass diffusion. When the disk spreads beyond an AU, its mid-plane cools [@Ruden:1986lr; @Garaud:2007fk] well below the condensation temperature of the CAIs. If the infalling gas has a modest amount of angular momentum, it would form more extended disks with shock temperatures below $10^3$K at a distance outside 1 AU [@Walch:2009qy]. In either limit, condensation of the elements and the formation of CAI’s would continue until infall is terminated. A possible solution to reconcile with CAI’s brief formation epoch is to assume that all infalling gas was heated above their condensation temperature of $\sim1800$K and then cooled within $20\kyr$. Such a scenario would require a very compact disk and therefore a violent, external cause, i.e. a supernova, to trigger and drive the collapse of the entire progenitor cloud which formed the Solar System. As shown in Appendix A, the subsequent growth of CAIs would be possible on a short enough time-scale in such a disk. External Enrichment by a Type II Supernova ------------------------------------------ A frequently discussed possibility for SLRs is the external enrichment by a type IIa supernova (SN IIa). @Huss:2009lr estimate that a supernova with a progenitor of $20-60\Msun$ can produce the required SLRs. From the simple assumption that the cross-section of the enriched region has to be high enough to intercept enough we can derive a size-distance relation from geometrical considerations: $$\label{eq_crosssection} R=\sqrt{\frac{4\cdot M_{\rm Al^{26}_{\odot}}}{M_{\rm Al^{26}_{SN}}}} D$$ Where $R$ is the radius of the intercepting cloud or disk and $D$ is the distance from the supernova. $M_{\rm Al^{26}_{\odot}}$ and $M_{\rm Al^{26}_{SN}}$ are the amount of required in the Solar System and produced in the supernova, respectively. This equation is of course simplified, e.g. neglecting radioactive decay as the time-scales are very short, as well as assuming a perfect mixing efficiency, i.e. all intercepted is assumed to be mixed into the Solar System. ${M_{\rm Al^{26}_{SN}}}[\Msun]$ ---------------- --------------------------------- ---------------------- Mass $[\Msun]$ Z=0.02 Z=0.004 20 $1.5\times 10^{-5}$ $8.65\times 10^{-6}$ 40 $6.64\times 10^{-5}$ $2.23\times 10^{-5}$ : Typical yields in $\Msun$ for different supernova progenitor masses and metallicities taken from [@Nomoto:2006lr]. \[yields\_tab\] ### Enrichment in the Disk Stage From the values in Table \[yields\_tab\] one can calculate that at solar metallicity a $40\Msun$ progenitor supernova could deposit enough in a $1000\AU$ face-on disk at $D=0.3\pc$ or a $20\Msun$ progenitor enough at $D=0.15\pc$. That is very close to the hot, ionizing star. Although a proto-stellar disk could survive in this violent environment [@Adams:2010lr], there are additional timing issues. For example @Gounelle:2008fk estimate the probability that the proto-solar system is young enough ($<1\Myr$), able to survive photo-evaporation and close enough to get sufficiently enriched ($D<0.3\pc$) to be less than $10^{-3}$ in the most favorable case. Based on the required young age of the system, @Williams:2007yq argue for a very massive supernova progenitor, which in turn leads to a short life-time, i.e. an early supernova. As very massive stars are only found in massive clusters, they conclude that the likelihood is $<1\%$. [@Ouellette:2005kx; @Ouellette:2009fk; @Ouellette:2010qy] suggested an intriguing possibility where the disk is further away from the supernova ($\sim 2\pc$) and the enrichment is achieved via the injection of already formed grains into the pre-solar disk (’aerogel’ model). To achieve the high observed yield, the supernova gas has to be highly clumpy and the clumps have to be highly enriched in SLRs - to this possibility the authors estimate a probability of $\sim10^{-3}-10^{-2}$. There are, however, additional uncertainties. Recent models [e.g. @Ida:2004lr] suggest that the minimum mass solar nebula might be more massive than their assumed $0.01\Msun$, which would require an even more efficient enrichment. Future investigation on the clumpiness of SN-shock fronts are desirable (see §\[discussion\]). A remaining issue in this picture is the question why this scenario does not result in multiple enrichments of the early Solar System. After the first enrichment by a supernova, which leads to the canonical value of , it is likely that several more lighter and therefore more long-lived stars in the cluster undergo subsequent supernova. This would result in several enrichment episodes during the entire lifetime of the disk ($>10\Myr$). ### Enrichment in the Cloud Core Stage The enrichment of the pre-solar cloud core by a supernova has been suggested for a long time [@Cameron:1977yq; @Cameron:1995qy]. Recently, @Boss:2008uq [@Boss:2010qy] and @Boss:2010lr investigated the implosion and enrichment of a $1\Msun$ cloud core in detail. They assume the cloud core to be more than $10\pc$ away from the progenitor. Therefore, the supernova front has slowed down considerably by the time it reached the pre-solar cloud. In addition, it has a finite thickness. In their models, the shock speed ranges from $v_{\rm{S}}=1-100\kms$ [@Boss:2010qy], the shock thickness from $\delta_{\rm{S}}=3\times10^{-4}-3\times10^{-2}\pc$ and the shock density from $\rho_{\rm{S}}=3.6\times10^{-21}-2.4\times10^{-17}\dens$ [@Boss:2010lr]. They find that with certain initial conditions the enrichment is possible in their model. For example a thin ($\delta_{\rm{S}}=3\times10^{-4}$), dense( $\rho_{\rm{S}}=6\times10^{-18}\dens$) shock at a rather low velocity ($v_{\rm{S}}=20\kms$) and therefore far away from the source can enrich the clump sufficiently. This can be easily understood as they assume that the enriched supernova material is completely within the shock. Therefore, as the geometrical dilution rises, the shock has to be thinner to contain a sufficient amount of supernova material. In addition, the time-scale of the collapse is longer ($t>40\kyr$) than the value inferred from CAIs in this scenario. Again, this can be attributed to the slower supernova shock further away from the progenitor. Nevertheless, this is an encouraging development and we think this scenario is the most promising. Bearing the geometrical dilution in mind (cf Fig. \[param\_space\]) we revisit this scenario but place the pre-solar clump closer to the supernova. Therefore, the shock velocities are much higher (see §\[IC\]). We also take the more conservative assumption that the material is spread out in the hot phase or at least in a ring at the border within the supernova explosion bubble (see §\[sim5pc\] and \[sim10pc\]). Furthermore, we do not assume the shock already formed a density enhancement. Instead, we setup the supernova blast wave in the velocities under the assumption of a thin shock. As the evolution is expected to be much more violent, we investigate the collapse of a heavier ($10\Msun$) cloud core (see §\[IC\]). To assess the feasibility of this initial condition, we address the aspect of survival of the pre-solar cloud (§\[survival\]). A slightly different approach would be the enrichment in the cloud stage without triggering the collapse. Then, the cold cloud would collapse in an ’isolated’ manner at a later stage. However, there are clear short-comings of this scenario. To allow for the tight spread of only $\simeq20\kyr$, the collapse would have to lead to a stage, where the entire disk is heated above $\sim1800\K$ and then homogeneously cools to lower temperatures within $20\kyr$ (see also §\[contin\_enrich\]). To achieve the required heating from the gravitational energy the accretion disk would have to be $<0.5\AU$. At this scale, the disk could then cool coherently, the CAI could condense and later migrate outwards to explain the homogeneous in CAIs in different meteorites. However, isolated collapse leads to larger, cooler disks. Thus, a triggering event will be required to force the collapse to a small scale. It is straightforward to assume that this triggering event was the same event as the enrichment, i.e. the supernova shock. Thus, we investigate the triggered case in more detail. Model and Simulations {#simulations} ===================== In this section, we address the parameter space by accessing the enrichment, survivability and existence of the pre-solar cloud core in or adjacent to the HII region of the massive star. We investigate the mixing and the collapse probabilities by employing numerical simulations. In Fig. \[param\_space\] we plot the distance of the clump to the progenitor star versus the size of the clump. From the geometrical enrichment cross-section discussed before, we can derive a first constraint. Taking the yields in Table \[yields\_tab\] into account we plot Eq. \[eq\_crosssection\]. These diagonal lines denote the maximum enrichment efficiency. Even if all the material intersected by the initial cloud ends up in the Solar System ($100\%$ efficiency), the initial condition has to be above these lines to explain the ratio inferred from CAIs. Survival -------- The most obvious proof for the existence of cold cloud cores in the proximity of massive star comes from observations. In the pillars in M16 [e.g. @Hester:1996fk] [@McCaughrean:2002uq] find several cold cores in almost every region of the pillars. These pillars are about $2\pc$ away from the ionizing sources, the age of the region is estimated to be $1-2\Myr$ [@Hester:1996fk]. This agrees well with our assumption of structures being present at $5\pc$ after $3\Myr$ (see below). From a theoretical point of view, recent simulations on the formation and evolution of HII-regions show the ubiquity of these structures [e.g. @Mellema:2006yq; @Krumholz:2007kx; @Gritschneder:2009uq; @Arthur:2011qy]. These simulations focus mostly on the ionization of turbulent molecular clouds. As shown in [@Gritschneder:2010fj], the average time evolution of the front position is still well approximated by an analytic solution as given in [@Shu:1991vn]: $$\label{x_front} R(t)=R_\mathrm{s}\left(1+\frac{7}{4}\frac{c_\mathrm{s,hot}}{R_\mathrm{s}}(t-t_0)\right)^\frac{4}{7},$$ where $R_\mathrm{s}$ is the initial Stroemgren radius[^2] and $c_\mathrm{s,hot}$ is the sound speed of the hot, ionized gas. Using this equation, we can estimate the size of the HII-region. To investigate the most extreme cases, we assume a short-lived $40\Msun$-star with a very short life-time of $3\Myr$ as well as a $20\Msun$-star with a life-time of $10\Myr$ [e.g. @Hurley:2000qy] in different density regions. The resulting radii are depicted by the vertical lines in Fig. \[param\_space\]. Surviving clumps should be to the right of theses lines. Note that we assume a relatively high density in the neutral medium before the ignition of the star in order to reflect the birth environment of massive stars. @Myers:2009uq suggest that the birth of stellar groups is associated with hubs of column densities greater than $10^{22}\sdens$, for example in Taurus [@Goldsmith:2008qy], which corresponds to the assumed densities here. In addition, the high densities have to be present only in the direct surrounding of the massive stars, as the further evolution given in Eq. \[x\_front\] only depends on the initial Stroemgren radius (and the sound speed in the hot gas). Altogether, these values are chosen to reflect our current understanding of massive star formation. In addition, Fig. 1 gives the averaged front position, the pillars in the observations as well as the simulations can be a few pc further inside, pertruding into the ionized bubble. Collapse Timescale and Mixing in the Gas Phase ---------------------------------------------- The remaining issues are whether the gas from the supernova can be efficiently mixed with the gas from the pre-existing core, whether collapse occurs and if so, whether it is fast enough. To test these questions, we employ numerical simulations. ### Initial Conditions {#IC} Because the main focus lies on mixing as well as cooling, we employ the numerical code COSMOS. The non-relativistic hydrodynamics of COSMOS are discussed in [@Anninos:2003qy]. For the models described below, the code is run with the default hydrodynamics solver, which is a total variation diminishing (TVD) approximate Riemann solver and a third-order Runge-Kutta time-marching scheme [@Shu:1988ys]. Because of the importance of cooling, the code is run in internal energy mode. Shocks are handled with a zone-centered scalar artificial viscosity. Optically thin heating and cooling are included, using equilibrium cooling curves [@Dalgarno:1972rt; @Boehringer:1989vn]. Cooling by metals assumes solar metallicity. Mixing is followed by the use of a passive tracer, whose value is initialized to be proportional to the gas density behind the blast-wave. The problems are run in two dimensions. Cylindrical symmetry is imposed by taking advantage of the relativistic capability of COSMOS, and running the problems with a cylindrical metric. The problem region runs from -5 to 10 cloud radii in the direction of the shock, and from 0 to 4 cloud radii in the transverse direction. The large radial size is adopted so as to prevent any boundary affects from altering the cloud evolution. Due to the motion of the cloud after shock passage, we use a uniform grid spacing in the direction of the incoming shock, with 4096 zones along the axis, or 274 zones across the cloud radius. In order to enhance resolution of the collapsed cloud, geometrically increasing zone spacing is used in the radial direction. A total of 512 zones are included radially, with a ratio of 1.005 in their spacing, giving a minimum zone size of 0.0017 times the cloud radius. We set up an molecular cloud core in isolation, which is going to be hit by the supernova shock-wave. We approximate the core by an marginally stable Bonnor-Ebert-Sphere [@Bonnor:1956lr], the solution of the Lane-Emden equation for hydrostatic equilibrium $$\frac{1}{r^2}\frac{d}{dr}(\frac{r^2}{\rho}\frac{dp}{dr})=-4\pi G\rho,$$ where gravity and hydrodynamical pressure balance each other. We assume an initial temperature of $T_\mathrm{cold}=20\K$ and a central number density of $n_\mathrm{c}=5.7\times 10^4\ndens$, which gives a cloud core of mass $M=10\Msun$ with a radius $R_0=0.21\pc$. The sphere is embedded in a warm gas in pressure equilibrium with the sphere at a temperature $T_\mathrm{warm}=792\K$. This is realistic, as the core is embedded in the interstellar medium close to an HII-region. As this is an idealized case and the main focus of this study is the enrichment with SRLs, we don’t take any rotational or turbulent motion [e.g. @Walch:2010fk] inside the cloud core into account. While this initial rotation might be important in the later phases of the collapse - especially in determining the final disk size - its influence will be very small in the initial phase here, as the rotational velocities are orders of magnitude smaller than the shock speeds. In addition, the total angular momentum carried by a minimum mass nebula ($0.02\Msun$) out to $30\AU$ corresponds to that of $1\Msun$ of material at $0.6\AU$. Unless there is a substantial angular momentum loss since the initial collapse to the setup of a minimum mass nebula, the initial cloud must carry very little angular momentum. In such a low-angular momentum cloud, rotation is unlikely to inhibit the collapse until the size of the cloud is well below the resolution of our simulations and the dynamical time scale is much shorter than the $20\kyr$ of interest. The supernova shock-wave at the distances considered here ($5-10\pc$) has not yet reached the snow-plow phase. So the shock is still in the energy dominated phase and can be approximated by a Sedov-Taylor blast wave [@Taylor:1950qy; @Sedov:1959uq]. Following @Woosley:1995fj we assume an energy release of $E=10^{51}$erg for core-collapse supernovae with progenitor stars in the range of $20-40\Msun$. Considering the parameter space (§\[survival\]), we perform two simulations, placing the cloud core at a distance $D=5\pc$ (case I) and $D=10\pc$ (case II), respectively, from the supernova. We evaluate the Sedov solution for the given distances. In case I, the results are that the blast wave reaches the core after $\simeq8.5\kyr$ with a post-shock speed of $v_\mathrm{s}=276\kms$. The sound-speed in the post-shock gas is $c_\mathrm{s}=129\kms$. In case II the cold cloud core is reached after $\simeq48.2\kyr$ with $v_\mathrm{s}=97.5\kms$ and $c_\mathrm{s}=45.4\kms$. ### Case I {#sim5pc} In the first case, we place the cold pre-solar core at a distance $D=5\pc$ from the supernova. Fig. \[time\_evol\] shows the time evolution of the density in this simulation. The shock wave is propagating from the bottom to the top. As it can be clearly seen, the shock wave encompasses the cold core rapidly. Various hydrodynamical instabilities occur at the interface. After $t=4.15\kyr$ the front has already passed the core. The part of the front which hit the clump continues to interact with the pre-existing core, leading to two fronts wrapping around it. As these fronts collide, the material gets mixed with the pre-existing cold gas and is driven into collapse. After $t=8.33\kyr$ the central region is already at a very high density ($\rho_\mathrm{max}=3\times10^{-15}\dens$), the computation gets expensive and we terminate it. Fig. \[zoom5pc\] shows a zoom-in into the density at the final snapshot. In Fig. \[zoom5pcT\] we show the temperature distribution in this region. As one can already see, the central region is cold. We define the central region as the cold, dense material between $0.25$ and $0.75$ on the y-axis of Fig. \[zoom5pcT\], corresponding to an elongated filament of $\approx0.1\pc$ diameter. The mass in this region below a temperature of $20\K$ is ${M_{\rm core}}\simeq0.13\Msun$, the average density is $\bar{\rho}=1\times10^{-17}\dens$. This gives a Jeans-mass[^3] of $\simeq0.08\Msun$. The core is twice as heavy and therefore will definitely undergo collapse. This collapse will proceed at least with the free-fall time-scale, which is $t_\mathrm{ff}\approx 20\kyr$ at this average density. More likely, it will proceed faster as the central density is two orders of magnitude higher and since the material is already moving inwards. In summary, it is probably safe to say that the further collapse to a stage below $1800\K$, where the CAIs form, takes less than $10\kyr$. The mass is still lower than the solar value, but given the fact that cooling and accretion will still happen this value can increase. Although, given from the mass available in the surrounding we do not estimate it to reach more than $0.2\Msun$ in total. In addition, very strong inflows or colliding remnants of the shock are still present, flowing towards the centre with velocities of $\simeq10\kms$. These will lead to an enhanced, fast collapse likely force the material in a small central region, which favors the simultaneous crystallization of CAIs. The other main focus here is the mixing of the enriched hot supernova gas with the cold pre-existing clump. To follow this evolution, we include a tracer in the post-shock gas inside the shock-front. The ratio of to the total gas mass within the supernova bubble depends on several factors. First of all, different progenitor masses give different yields. Secondly, the exact distribution of the enriched gas within the supernova bubble remains an open question. We investigate two cases here. The worst case, where the supernova enrichment is fully mixed with the entire material in the hot bubble, i.e. the $20\Msun$ of total supernova yield are diluted with $\sim1000\Msun$ within the bubble[^4]. The second case is more optimistic, assuming that the enriched material stays close to the blast-wave. We assume it to be within the outer $10\%$ of the radius of the bubble, leading to a dilution by only $\sim350\Msun$ of pre-supernova material. However, even in the worst case of complete mixing, the enrichment is sufficient for a $40\Msun$ progenitor star at solar metallicity. This is shown in Fig. \[mixing\]. Color coded is the ratio of to total mass, normalized to the required value of $\simeq3.71\times10^{-9}$. Since in the collapsing regions the ratio is of order unity or higher, we conclude that the gas phase is sufficiently enriched to allow for the formation of the observed CAIs later on. In Table \[results\_tab\] we list the outcome of different scenarios. We give ${M_{\rm^{26}Al}^{\rm sim}}$, the total mass of inside the final collapsing core, as well as the dilution $f_{0}$ (i.e. the fraction of supernova material injected into the final core) and the injection efficiency $f_{\rm i}$ as defined in Eq. \[eq\_enrich\]. Finally, we list the resulting enrichment for a collapsing core scaled to the solar ratio $$f_{\odot} = \frac{{M_{\rm^{26}Al}^{\rm sim}}}{M_{\rm core}}\frac {M_{\odot}}{M_{\rm^{26}Al}^{\rm canonical}}.$$ $f_{\odot}$ can be interpreted as the success of the model, i.e. value of 1.0 corresponds to the observed in the Solar System. Our results differ depending on the assumed distribution of enriched material inside the supernova bubble and on the progenitor mass. As can be seen, in several cases the enrichment is sufficient, if the cloud core is placed at $D=5\pc$ from the supernova initially. Note that the dilution $f_0$ is systematically lower than the values of $f_0\approx (0.5-6)\times 10^{-4}$ as given by @Takigawa:2008lr. At the same time the mass in the cold core is only $0.13\Msun$ in our case, explaining that difference of one order in magnitude. In addition, we assume a high number density inside the supernova bubble ($n=100\ndens$), which leads to a very strong dilution. Thus, these estimates are very conservative and can be viewed as a worst case scenario enrichment. To further understand the reason for this very efficient mixing, it is useful to take a closer look at the earlier stage of the simulation. In Fig. \[mixingT\] we show the temperature distribution at the intermediate stage ($t=1.25\kyr$). As can be clearly seen the material is already cooling behind the shock front. This cold, enriched clumps are mixing very efficiently with the pre-existing core. Therefore, a precise prescription of cooling processes is vital to investigate enrichment scenarios. Distance Mass Metallicity SNe mixing ${M_{\rm^{26}Al}^{\rm sim}}[\Msun]$ $f_{0}$ $f_{\rm i}$ $f_{\odot}$ ---------- ----------- ------------- ------------ ------------------------------------- -------------------- -------------------- ------------- $5\pc$ $40\Msun$ $Z=0.02$ shell $1.6\times10^{-9}$ $2.3\times10^{-5}$ $5.3\times10^{-2}$ 3.34 complete $4.4\times10^{-10}$ $6.6\times10^{-6}$ $1.5\times10^{-2}$ 0.91 $Z=0.004$ shell $5.4\times10^{-10}$ $2.3\times10^{-5}$ $5.3\times10^{-2}$ 1.1 complete $1.6\times10^{-10}$ $7.0\times10^{-6}$ $1.6\times10^{-2}$ 0.32 $20\Msun$ $Z=0.02$ shell $3.6\times10^{-10}$ $2.5\times10^{-5}$ $5.6\times10^{-2}$ 0.75 complete $1.0\times10^{-10}$ $6.6\times10^{-6}$ $1.5\times10^{-2}$ 0.21 $Z=0.004$ shell $1.6\times10^{-10}$ $2.5\times10^{-5}$ $5.6\times10^{-2}$ 0.43 complete $5.7\times10^{-11}$ $6.5\times10^{-6}$ $1.5\times10^{-2}$ 0.11 $10\pc$ $40\Msun$ $Z=0.02$ shell $3.9\times10^{-12}$ $5.8\times10^{-8}$ $5.2\times10^{-4}$ 0.065 complete $1.1\times10^{-12}$ $1.6\times10^{-8}$ $1.5\times10^{-4}$ 0.017 $Z=0.004$ shell $1.3\times10^{-12}$ $5.8\times10^{-8}$ $5.2\times10^{-4}$ 0.022 complete $3.8\times10^{-13}$ $1.8\times10^{-8}$ $1.6\times10^{-4}$ 0.0064 $20\Msun$ $Z=0.02$ shell $8.7\times10^{-13}$ $5.8\times10^{-8}$ $5.2\times10^{-4}$ 0.014 complete $2.4\times10^{-13}$ $1.6\times10^{-8}$ $1.5\times10^{-4}$ 0.0041 $Z=0.004$ shell $5.1\times10^{-13}$ $5.9\times10^{-8}$ $5.4\times10^{-4}$ 0.0085 complete $1.4\times10^{-13}$ $1.6\times10^{-8}$ $1.4\times10^{-4}$ 0.0023 ### Case II {#sim10pc} In the second case, the cold cloud core is at a distance $D=10\pc$ and is embedded in an ambient medium which is a factor of ten less dense with a ten times higher temperature than case I. Therefore, the ambient pressure is the same. Due to the longer distance the angular cross-section is smaller, the enriched material is already more dispersed. In addition, the shock has already lost more energy, i.e. the front speed is slower. The effect of this can be seen in Fig. \[zoom10pc\]. After $t=14.6\kyr$ the cold core is still much less deformed by the supernova and not driven into collapse. Although the mixing seems to be very efficient, a closer look at Fig. \[mix10pc\] shows that the gravitational collapsing region, a remnant of the original cloud core is not strongly enriched (the blue region at $0.0-0.1$ cloud radii showing the highest density in Fig \[zoom10pc\]), whereas the enriched region further away (the light blue region between $0.7-1.3$ cloud radii) shows the effect of the two flanks of enriched supernova material colliding. In addition, the mass in the cold region is much smaller, only $0.01\Msun$, which is 30 times less than the Jeans mass at the average density of $\bar{\rho}=7\times10^{-19}\dens$ in this region. Therefore, the cloud will not collapse in this case. This is due to supernova shock wave already being substantially weakened so far away from the front. Of course it would be possible to start with a more condensed core, but then the enrichment, which is already difficult in this case (see below), becomes even more unlikely as the radius of the core would be smaller (cf Eq. \[eq\_crosssection\]). From Table \[results\_tab\] it can be seen that sufficient enrichment is very unlikely in this scenario, especially since the enriched material is now more diluted, as the volume of the supernova bubble is bigger, i.e. the material produced by the supernova is now mixed with $\sim10^4\Msun$ in the complete mixing case and with $\sim2800\Msun$ in the shell case. Altogether the clump will neither collapse, nor be sufficiently enriched in this case. Comparison of the enrichment in simulations with theoretical (geometrical) predictions {#enrichment} -------------------------------------------------------------------------------------- Here, we compare the predicted enrichment from the geometrical cross-section (Eq. \[eq\_crosssection\] and Fig. \[param\_space\]) with the finally achieved enrichment. From a geometrical point of view, the dilution $f_0$ is given as by the radius of the initial sphere $R_0$ and the distance $D$ to the source as: $$\label{eq_dilution} f_{0\rm,geom}=\frac{R_0^2\pi}{4D^2\pi}.$$ Thus, $f_{0\rm,geom}$ is the amount of intercepted geometrically by the initial Bonnor Ebert sphere. $f_{0\rm,geom}=4.4\times 10^{-4}$ and $f_{0\rm,geom}=1.1\times 10^{-4}$ for cases I and II, respectively. The values are systematically higher than $f_0$ in Table \[results\_tab\] as Eq. \[eq\_dilution\] assumes perfect mixing of the supernova material with the background gas. However, it can be used to determine an injection efficiency, which we can not define directly, as we do not assume the material to be in the front initially. Instead, we define the injection efficiency $f_{\rm i}$ as the ratio of the finally incorporated mass to expected intercepted mass from geometrical considerations $$\label{eq_enrich} f_{\rm i}=\frac{M_{\rm^{26}Al}^{\rm sim}}{f_{0\rm,geom}M_{\rm^{26}Al}^{\rm SNe}}=\frac{f_{0}}{f_{0\rm,geom}}.$$ $f_{\rm i}$ would be unity if all intercepted material is included in the collapsing core. We give the values in Table \[results\_tab\] as well. For case I $f_{\rm i}\approx 0.01-0.05$, i.e. a few percent of the material crossing the cold sphere initially are incorporated in the collapsing core finally. In case II, the values are much lower, $f_{\rm i}\approx0.0001-0.0006$. First, this is due to a larger amount of un-enriched material in the bigger bubble. Second, the shock front is much weaker in case II, and therefore the front does not mix as efficient with the cloud core. Discussion ========== The outcome of these simulations and the results given here are a first order proof of principle. From the current scenario it seems possible that every supernova with a progenitor of $20-40\Msun$ can produce a system with solar-like abundances by triggering a nearby ($\approx5\pc$) cloud core into collapse. However, we only tested a small range of the parameter space. Future simulations are desirable to further probe the likelihood. A first issue would be that we assume the (idealized) analytical Sedov-Solution for the blast wave. A more realistic treatment for the entire evolution of the shock from the progenitor is required in the future. This is of course computationally very expensive. Nevertheless, it is important to test e.g. the fragmentation and cooling of the shock front during the $8\kyr$ before it hits the cloud core. Another area to improve would be to include more detailed effects. For example, the core could be initially rotating. We assume here that the high shock velocities dominate over the rotation, but this should be further investigated. Furthermore, we neglect magnetic fields. They are of course present and should be involved in various processes, e.g. the details of the shock front and especially later on in the disk formation and the removement of angular momentum. In the early stages, however, the kinetic energy of the shock plays the dominant role. Still, it is worthwhile to investigate this regime of the parameter space in future simulations. A remaining issue is the production of heavy SLRs like ${\rm^{53}Mn}$ and ${\rm^{60}Fe}$. @Gounelle:2009uq suggest that the observed amount of ${\rm^{60}Fe}$ can be inherited from the ISM. Still, the amount of galactic is an order of magnitude lower than in the Solar System [@Diehl:2006fj]. Thus, a different enrichment mechanism for is still needed. In addition, in the supernova-enrichment scenario, ${\rm^{60}Fe}$ and especially ${\rm^{53}Mn}$ are over-produced by a factor of 10-100 relative to their abundance in meteorites when compared to the lighter SLRs ${\rm^{41}Ca}$ and [e.g. @Rauscher:2002lr; @Gaidos:2009fk]. This is often used as an argument against the supernova-enrichment scenario. However, there are several possible solutions. First of all, the precise determination of abundances, especially of radioactive isotopes in supernovae remains challenging. Second, there are alternative models, e.g. fallback-mixing within supernovae. Here, the innermost layers (commonly within the Si-burning layer) fall back onto the star. Therefore, there are much less SLRs heavier than Si within the supernova shock [e.g. @Takigawa:2008lr]. Third, as it is an overproduction, not a lack of these SLRs, there could be different incorporation efficiencies into CAIs and meteorites, which are not yet well understood. On a side note, the distribution of the supernova yield in the entire bubble versus the distribution of it in an outer rim gives different results (cf Table \[results\_tab\]). As the iron is produced in layers further inside it is valid to speculate that the different ratios might be due to an homogeneous distribution of ${\rm^{60}Fe}$ in the entire bubble, whereas might be concentrated in an outer layer. This scenario is of course only valid if the initial distribution of elements in the supernova is conserved while the bubble expands to $5\pc$. The different abundances of SLRs are often used to calibrate the delay between the supernova and the formation of CAIs. Mostly, these calculations give values of about a $\Myr$ (@Takigawa:2008lr: $0.7\Myr$, @Looney:2006lr: $1.8\Myr$). This contradicts the fast formation time required to produce the small spread observed in the CAIs. In our opinion, the supernova yields of SLRs are at present not understood well enough to derive meaningful delay time scales between supernova and CAI formation. Especially, since there might be different processes during the disk formation, which could also explain the different abundances of SLRs. In our simulations we find that a distance $D\approx5\pc$ is optimal to keep the balance of survivability for the pre-existing cold clump and the sufficient enrichment. This is larger than previous estimates - e.g. [@Looney:2006lr] derive a distance smaller than $4\pc$ from analytical estimates. The reason for this is the efficient mixing as discussed in §\[sim5pc\] (see also Fig. \[mixingT\]). For the survivability of the pre-solar clump we assume a quite high density and short lifetime for the HII-region. This leads to a small size, i.e. a small distance for the pre-existing clump from the O-star. The general picture would be the ionization front running into the paternal filament, from which the O-star originally formed. On a related note, this leads to the assumed high density of the ambient surrounding of $n=100\ndens$ (according to the analytical evolution of an HII-region, see e.g. @Shu:1991vn). This value is well within the observed range. @Lefloch:2002lr estimate the density in the Trifid Nebula to be $50\ndens$ whereas @Rubin:2011lr find values of $87-714\ndens$ and $74-1041\ndens$ for different lines in optical and infrared observations, respectively. In turn, the chosen value leads to a high density inside the supernova bubble, i.e. a very strong dilution of the enriched material. Thus, optimistic conditions for the survival lead to very conservative or pessimistic assumptions for the dilution. In addition, O/B-stars tend to form in bigger associations, therefore there would be more ionizing flux than from a single O-star. Furthermore, we neglect the effect of stellar winds. Altogether, the conditions for the survivability have been chosen quite optimistically to have a cold core close enough to the O-star to be sufficiently enriched ($D\approx5\pc$). Still, as the observations around various HII-regions show, pillar like-structures and therefore cold clumps are observed frequently at these distances. On a more technical side, these simulations are still only in 2D, so some physical processes might not be well resolved. This might apply to the gravitational collapse of two fronts colliding. However, colliding fronts of these speeds and densities have been shown to become gravitationally unstable with full 3D simulations [e.g. @Gritschneder:2009lr]. Another short-coming of 2D simulations might be the different behavior of instabilities and mixing in 2D and 3D [e.g. @Stone:2007fk]. This should be less of an issue here, as the mixing appears in the cold phase and is not solely due to hydrodynamical instabilities, which would be less well represented in 2D (see Fig. \[mixingT\]). The mixing is mainly influenced by the cooling, which can be well approximated in 2D. Nevertheless, further simulations in full 3D would be desirable to address this issues. Conclusions & Outlook ===================== In this study, we focus on the hydrodynamical interaction of a supernova with the pre-solar cloud core. A realistic supernova shock wave is assumed to interact with the cloud. This leads to the conclusion that it is possible to enrich a pre-existing clump sufficiently with to explain the observed ratio in the Solar System. In order to survive the previous HII-region and still be close enough to be sufficiently enriched and triggered into collapse, it turns out that $D\approx5\pc$ is an ideal distance. The previously stable core gets enriched and is triggered into gravitational collapse within $t\approx8\kyr$ after the first interaction with the supernova shock-wave. The further collapse until the formation of the CAIs, which takes places at temperatures below $1800\K$, can be estimated to be shorter than $\simeq10\kyr$. Whether the collapse finally goes all the way down to the condensation of the CAIs within $18\kyr$ or if there is an intermediate stage with a hot disk where the pre-meteorite gas is well mixed remains to be investigated. In either case, the CAIs will be able to grow fast enough to the observed cm-sizes within the background density of a disk, as shown in Appendix A. The hot disk, however, can as well only be achieved by a very violent (i.e. supernova triggered) collapse. In addition, the subsequent expansion and cooling of the disk will be rapid at these high densities so it is preferable if the material is well mixed before. On a side note, it is very interesting that the mixing occurs mainly by material from the supernova shock, which already cooled. This cooled portion interacts with the pre-existing cold gas in the surrounding of the progenitor star, whereas most of the supernova enriched material stays in the hot component. Therefore, it might be possible to determine the efficiency of enrichment by supernova feedback in the much larger context of galaxies by investigating the cooling rates within supernova shocks. Altogether, the precise time and location of the grain formation has to be determined by future investigations. In this work, we prove that it is possible to trigger a pre-existing stable core into collapse by a very nearby supernova ($D\approx5\pc$). In addition, we show for the first time that the collapsing core can be enriched sufficiently with SLRs to explain the observed abundances, even if the enriched material is not assumed to be in the shock initially, but instead distributed homogeneously within the entire supernova bubble. We would like to thank the referees for valuable comments on the manuscript. We thank Josh Wimpenny for carefully proof reading the final manuscript. M.G. acknowledges funding by the China National Postdoc Fund Grant No. 20100470108 and the National Science Foundation of China Grant No. 11003001. D.N.C.L. acknowledges support by NASA grant NNX07AI88G, NNX08AL41G and NNX08AM84G as well as the NSF grant AST-0908807. Q.Z.Y. acknowledges NASA Cosmochemistry grant NNX08AG57G and Origins of Solar Systems grant NNX09AC93G. The work by S.D.M. was performed under the auspices of the Lawrence Livermore National Security, LLC under contract No. DE-AC52-07NA27344. Assembly and Growth of CAIs =========================== Assuming that Aluminum sticks perfectly to the grains, the growth rate $\dot{M}$ for a particle of size $a$ at a certain radius within the disk with a relative velocity $v_\mathrm{rel}$ is $$\dot{M} = \pi a^2 \rho_m v_\mathrm{rel},$$ where $\rho_\mathrm{m}$ is the background density of the CAI-forming material, e.g. . If we assume that, as the particle grows the density of the particle $\rho_\mathrm{p}$ as well as $\rho_\mathrm{m}$ remain constant, we can write $$4\pi a^2 \dot{a} \rho_\mathrm{p} = \pi a^2 \rho_\mathrm{m} v_\mathrm{rel}$$ and solving for $\dot{a}$ $$\dot{a} = \frac{\rho_\mathrm{m}}{4\rho_\mathrm{p}}v_\mathrm{rel}.$$ If we ignore the migration of particle, that is, the particle stays in a certain radius $r$ as it grows, the radius $a$ simply grows linearly with time $$a = a_0 + \frac{\rho_\mathrm{m}}{4\rho_\mathrm{p}}v_\mathrm{rel}t = a_0 + kt,$$ where the growth rate $k = \frac{\rho_\mathrm{m}}{4\rho_\mathrm{p}}v_\mathrm{rel}$ is a function of radius. Since we investigate a longer period of time ($t\sim20\kyr$), the initial particle size is irrelevant, so we can assume $a_0 = 0$. Therefore, $$\frac{\rho_\mathrm{m}}{4\rho_\mathrm{p}}v_\mathrm{rel}=\frac{a}{t}.$$ To assemble cm-size objects within $20\kyr$, this gives: $$\frac{\rho_\mathrm{m}}{4\rho_\mathrm{p}}v_\mathrm{rel}=1.6\times10^{-17}\kms.$$ Assuming a density $\rho_\mathrm{p}=1\dens$ and a relative velocity of $v_\mathrm{rel}=1\kms$, the resulting Aluminum background density has to be $\rho_\mathrm{m}=6.4\times10^{-17}\dens.$ With the total mass of ${M_{\rm^{27}Al}}=7.1\times10^{-5}\Msun$ [e.g. @Lodders:2003fk] this leads to a required total background density of $$\rho_\mathrm{disk}=9.0\times10^{-13}\dens,$$ which is higher than the density in cloud cores, but can be easily reached in proto-stellar and proto-planetary disks. On the one hand this is an optimistic scenario for the growth, since we assume a perfect sticking efficiency. On the other hand, the injected supernova material most likely already contains $\sim0.1\mu$m-sized grains [e.g. @Nath:2008fr]. Furthermore, we only take into account here, whereas the CAIs are not formed by pure . [^1]: CV-chondrites are a class of CAI-rich carbonaceous chondrites named after the Vigarano meteorite [@Dauphas:2011fk] [^2]: The Stroemgren radius is the radius which can be immediately ionized by an ionizing source, without taking any hydrodynamic evolution into account [@Stromgren:1939lr]. [^3]: The Jeans-mass is calculated with a mean molecular weight of $\mu=2.5$, i.e. assuming that molecular hydrogen and helium are the dominant species. [^4]: To estimate the dilution we assume an average density of $n=100\ndens$ within the supernova bubble from here on. This is a high value, but still realistic (for a discussion see §\[discussion\]).

ArXiv

--- author: - | D.V.ANTONOV [^1]\ [*Institute of Theoretical and Experimental Physics*]{}\ [*B.Cheremushkinskaya 25, 117218, Moscow, Russia*]{}\ [*and*]{}\ [*Institut für Elementarteilchenphysik, Humboldt-Universität,*]{}\ [*Invalidenstrasse 110, D-10115, Berlin, Germany*]{}\ title: '**Towards the Realistic Gluodynamics String. Perturbative Gluons’ Contribution to the String Effective Action**' --- 175 mm =-2cm =-3cm **[Abstract]{}** Perturbation theory in the nonperturbative QCD vacuum and the non-Abelian Stokes theorem, representing a Wilson loop in the $SU(2)$ gluodynamics as an integral over all the orientations in colour space, are applied to derivation of the correction to the string effective action, obtained in Ref. 1. This correction is due to accounting in the lowest order of perturbation theory for the interaction of perturbative gluons with the string world sheet. It occurs that this interaction affects only the coupling constant of the rigidity term, while its contribution to the string tension of the Nambu-Goto term vanishes. The obtained correction to the rigid string coupling constant multiplicatively depends on the spin of the representation of the Wilson loop under consideration, the QCD coupling constant and a certain path integral, which includes the background Wilson average. Recently, a new approach to the gluodynamics string was suggested$^{1}$. Within this approach one considers the Wilson average written through the non-Abelian Stokes theorem$^{2,3}$ and the cumulant expansion$^{3,4}$ as a statistical weight in the partition function of some effective string theory. The action of this string theory may be then obtained via expansion of the Wilson average in powers of two small parameters, $\left(\alpha_sF_{\mu\nu}^a(0)F_{\mu\nu}^a(0)\right)^ {\frac{1}{2}}T_g^2$ and $\left(\frac{T_g}{r}\right)^2$, where $T_g\simeq 0,2 fm$ is the correlation length of the vacuum$^{5,6}$, and $r\simeq 1 fm$ is the size of the Wilson loop in the confining regime$^{7}$. This yields the so-called curvature expansion for the gluodynamics string effective action. In Ref. 1 only the first nonvanishing terms of this expansion, corresponding to the lowest, second, order in the former parameter were accounted for, which corresponds to the so-called bilocal approximation$^{3, 8-11}$, and the expansion up to the terms of the third order in the latter parameter was elaborated out. The first two terms of this expansion read as follows $$S_{biloc.}=\sigma\int d^2\xi\sqrt{g}+\frac{1}{\alpha_0}\int d^2\xi\sqrt{g}g^{ij}\left(\partial_it_{\mu\nu}\right)\left(\partial_j t_{\mu\nu}\right), \eqno (1)$$ where $$\sigma=4T_g^2\int d^2zD\left(z^2\right) \eqno (2)$$ is the string tension of the Nambu-Goto term, and $$\frac{1}{\alpha_0}=\frac{1}{4}T_g^4\int d^2zz^2\left(2D_1\left(z^2 \right)-D\left(z^2\right)\right) \eqno (3)$$ is an inverse bare coupling constant of the rigidity term, while the terms of the third order in $\left(\frac{T_g}{r}\right)^2$ contain higher derivatives of the induced metric $g_{ij}$ and/or of the extrinsic curvature tensor $t_{\mu\nu}$ w.r.t. world sheet coordinates, and we shall not quote them here (see Ref. 1 for the details)[^2]. In Eqs. (2) and (3) $D$ and $D_1$ stand for two renormalization group invariant coefficient functions, parametrizing the gauge-invariant bilocal correlator of gluonic field strength tensors$^{3, 8-10}$, and it is worth noting that since the nonperturbative parts of these functions are related to each other as $\mid D_1\mid\simeq\frac{1}{3}D$ according to lattice data$^{6}$, the inverse rigid string coupling constant (3) is negative, which according to Ref. 12 agrees with the mechanism of confinement, based on the dual Meissner effect$^{13}$ (see discussion in Ref. 1). This result confirms that the Method of Vacuum Correlators, developed in Refs. 3 and 8-11 (see also Refs. therein), provides us with the consistent description of the confining gluodynamics vacuum. The approach suggested in Ref. 1 enables one to express all the coupling constants of the terms emerging in the string action in higher orders of the curvature expansion through the gauge-invariant correlators of gluonic field strength tensor only. Notice also, that in Ref. 14 action (1) was applied to the derivation of the correction to the Hamiltonian of the QCD string with quarks, which was obtained in Ref. 15, arising due to the rigidity term, with the help of which a rigid string-induced term in the Hamiltonian of the relativistic quark model was then evaluated for the case of large masses of a quark and antiquark. However, it should be emphasized that the curvature expansion describes only the pure nonperturbative content of the gluodynamics string theory. As it was explained in Ref. 10, in order to get the correct spectrum of the open bosonic string and the exponential growth of the multiplicity of string states, which is necessary for ensuring the property of duality of amplitudes, contained in the Veneziano formula, one must account for the perturbative gluons interacting with the string, which can be done in the framework of the perturbation theory in the nonperturbative QCD vacuum$^{9}$ (see discussion in Ref. 1). In this letter, we shall take this interaction into account in the lowest order of perturbation theory and obtain the corresponding correction to the action (1). To this end, one needs to integrate over perturbative fluctuations in the expression for the Wilson average written through the non-Abelian Stokes theorem. This procedure is, however, looks rather difficult to be elaborated out in the case when one makes use of the version of the non-Abelian Stokes theorem, suggested in Refs. 2 and 3, due to the path-ordering, which is remained in the expression for the Wilson loop after rewriting it as a surface integral. In what follows, in order to get rid of it, we shall exploit another version of the non-Abelian Stokes theorem, which was proposed in Ref. 16, where the path-ordering was replaced by the integration over an auxiliary field from the $SU\left(N_c\right)/\left[U(1)\right]^{N_c-1}$ coset space. For simplicity we shall consider the $SU(2)$-case, when this field is a unit three-vector $\vec n$, which characterizes the instant orientation in the colour space, and the non-Abelian Stokes theorem takes a remarkably simple form. Integration over perturbative fluctuations then yields the interaction of the elements of the string world sheet via the nonperturbative gluonic exchanges. In other words, we arrive at a theory of the open nonperturbative strings between dynamical world sheet elements, which provides us with the quantitative description of the intuitive picture, described in the previous paragraph. Finally, due to the short rangeness of this nonperturbative interaction between world sheet elements, it occurs possible to perform the $\vec n$-averaging and extract all the remnant dependence on the background fields in the form of the Wilson average standing under a certain path integral, so that the dependence on the world sheet elements decouples and may be evaluated explicitly, which yields a correction to the rigidity term, while the string tension of the Nambu-Goto term does not acquire any corrections and keeps its pure nonperturbative value (2). All the points, mentioned above, will be worked out in the next Section. The main results of the letter are summarized in the Conclusion. The statistical weight of the effective string theory, we are going to derive, is the Wilson average in the $SU(2)$ gluodynamics $\left< W(C)\right>=\left<tr~P~\exp\left(ig\oint\limits_C^{} dx_\mu A_\mu^at^a\right)\right>$, which after rewriting it as a surface integral by virtue of the non-Abelian Stokes theorem, suggested in Ref. 16, splitting the total field $A_\mu^a$ into the background $B_\mu^a$ and the perturbative fluctuations $a_\mu^a, A_\mu^a=B_\mu^a+ a_\mu^a$, and making use of the background field formalism$^{9,17}$, takes the form $$\left<W(C)\right>=\int DB_\mu^a Da_\mu^a D\vec n \exp\Biggl[\int dx \Biggl[-\frac{1}{4}\left(F_{\mu\nu}^a\right)^2+\frac{1}{2}a_\mu^a D_\lambda^{ab}D_\lambda^{bc}a_\mu^c+a_\nu^aD_\mu^{ab}F_{\mu\nu}^b+$$ $$+\frac{iJ}{2}\int d\sigma_{\mu\nu}n^a\left(-g\left(F_{\mu\nu}^a+2D_\mu^{ab}a_\nu^b\right)+ \varepsilon^{abc}\left(D_\mu\vec n{\,}\right)^b\left(D_\nu\vec n{\,}\right)^c\right)\Biggr]\Biggr], \eqno (4)$$ where $F_{\mu\nu}^a=\partial_\mu B_\nu^a-\partial_\nu B_\mu^a+g\varepsilon^{abc}B_\mu^bB_\nu^c$ is a strength tensor of the background field, $D_\mu^{ab}=\delta^{ab}\partial_\mu-g\varepsilon^{abc}B_\mu^c$ is the corresponding covariant derivative, and $J=\frac{1}{2},1,\frac{3}{2},...$ is the spin of the representation of the Wilson loop under consideration. In what follows we shall be interested only in the effects of the lowest order of perturbation theory, so that on the R.H.S. of Eq. (4) we have replaced the full covariant derivative which should stand in the last term (the so-called Wess-Zumino term) by the background one and omitted the ghost term, the term, describing the interaction of two perturbative gluons with the background field strength tensor and with the string world sheet, and the terms which describe the interaction of three and four perturbative gluons. Integration over the perturbative fluctuations in Eq. (4) is Gaussian and yields $$\left<W(C)\right>=\int DB_\mu^a D\vec n \exp\Biggl[-\frac{1}{4}\int dx\left(F_{\mu\nu}^a\right)^2+\frac{iJ}{2}\int d\sigma_{\mu\nu}\varepsilon^{abc}n^a\left(D_\mu\vec n{\,}\right)^b \left(D_\nu\vec n{\,}\right)^c\Biggr]\cdot$$ $$\cdot\exp\Biggl(-\frac{iJg}{2}\int d\sigma_{\mu\nu}n^aF_{\mu\nu}^a\Biggr)\exp\Biggl[-\frac{1}{2}\int dxdy \Biggl[D_\mu^{ba}\left(iJgT_{\mu\nu}^a(x)+F_{\mu\nu}^a(x)\right)\Biggr] \int\limits_0^\infty ds\int (Dz)_{xy}e^{-\int\limits_0^s\frac{\dot z^2}{4}d\lambda}\cdot$$ $$\cdot\Biggl[P~\exp\Biggl(ig\int\limits_0^s d\lambda\dot z_\alpha B_\alpha\Biggr)\Biggr]^{bc}\Biggl[D_\rho^{cd}\left( iJgT_{\rho\nu}^d(y)+F_{\rho\nu}^d (y)\right)\Biggr]\Biggr], \eqno (5)$$ where $T_{\mu\nu}^a(x)\equiv\int d^2\xi\varepsilon^{ij}\left(\partial_ix_\mu(\xi)\right) \left(\partial_jx_\nu(\xi)\right)n^a\left(x(\xi)\right) \delta\left(x-x(\xi)\right)$ is the colour vorticity tensor current. Making use of the formula$^{4}$ $$\left<e^PQ\right>=\left(\exp\left(\sum\limits_{n=1}^\infty \frac{1}{n!}\left<\left< P^n\right>\right>\right)\right)\left(\left<Q\right>+ \sum\limits_{k=1}^\infty\frac{1}{k!} \left<\left< P^kQ\right>\right>\right),$$ where $P$ and $Q$ stand for two statistically dependent commuting quantities, and $\left<\left<...\right>\right>$ denotes the so-called cumulants, i.e. irreducible correlators, we get in the lowest order of the cumulant expansion the following correction to the string effective action (1) $$\Delta S=-\frac{J^2g^2}{2}\int dxdy\Biggl<\left(D_\mu^{ba}T_{\mu\nu}^a(x)\right)\int\limits_0^\infty ds\int\left(Dz\right)_{xy}e^{-\int\limits_0^s\frac{\dot z^2}{4}d\lambda}\cdot$$ $$\cdot\Biggl[P~\exp\Biggl(ig\int\limits_0^s d\lambda\dot z_\alpha B_\alpha\Biggr)\Biggr]^{bc}D_\rho^{cd}T_{\rho\nu}^d(y)\Biggr>_{\vec n{\,}, B_\mu^a}, \eqno (6)$$ where $$\left<...\right>_{\vec n{\,}}\equiv\int D\vec n{\,}\left(...\right)\exp\left(\frac{iJ}{2}\int d\sigma_{\mu\nu}\varepsilon^{abc}n^a\left(D_\mu\vec n{\,}\right)^b\left(D_\nu \vec n{\,}\right)^c\right),$$ $$\left<...\right>_{B_\mu^a}\equiv\int DB_\mu^a\left(...\right)\exp\left(-\frac{1}{4}\int dx\left(F_{\mu\nu}^a\right)^2\right),$$ and during the derivation of Eq. (6) we have omitted the term $$\exp\left[-\frac{1}{2}\int dxdy\left(D_\mu^{ba}F_{\mu\nu}^a(x)\right)\int\limits_0^\infty ds\int \left(Dz\right)_{xy}e^{-\int\limits_0^s\frac{\dot z^2}{4}d\lambda}\left[P~\exp\left(ig\int\limits_0^s d\lambda\dot z_\alpha B_\alpha\right)\right]^{bc}D_\rho^{cd}F_{\rho\nu}^d(y)\right]$$ on the R.H.S. of Eq. (5), which does not yield any contributions to the string action due to the lack of coupling with the world sheet and therefore may be absorbed into the measure $DB_\mu^a$. Integrating in Eq. (6) by parts, we arrive in the lowest order of perturbation theory at the following formula $$\Delta S=-\frac{J^2g^2}{2}\int d^2\xi\int d^2\xi^\prime\varepsilon^{ij}\varepsilon^{kl}\left(\partial_ix_\mu\right) \left(\partial_jx_\nu\right)(\partial_kx_\rho^\prime)\left( \partial_lx_\nu^\prime\right)\cdot$$ $$\cdot\left<\left<n^b\left(x\right)n^c\left( x^\prime \right)\right>_{\vec n{\,}}\frac{\partial^2}{\partial x_\mu \partial x_\rho^\prime}\int\limits_0^\infty ds\int\left(Dz\right)_{xx^\prime}e^{-\int\limits_0^s\frac{\dot z^2}{4}d\lambda}\left[P~\exp\left(ig\int\limits_0^s d\lambda\dot z_\alpha B_\alpha\right)\right]^{bc}\right>_{B_\mu^a}, \eqno (7)$$ where $x_\mu\equiv x_\mu\left(\xi\right)$, and $x_\mu^\prime\equiv x_\mu\left(\xi^\prime\right)$. Since $e^{-\int\limits_0^s\frac{\dot z^2}{4}d\lambda}P~\exp\left(ig\int\limits_0^s d\lambda\dot z_\alpha B_\alpha\right)$ is the statistical weight of a perturbative gluon, propagating from the point $x^\prime$ to the point $x$ along the trajectory $z_\alpha$ during the proper time $s$, it is the region where $s$ is small, which mainly contributes to the path integral on the R.H.S. of Eq. (7). This means that the dominant contribution to $\Delta S$ comes from those $x_\mu$’s and $x_\mu^\prime$’s, which are very close to each other, which is in the line with the curvature expansion, where $\mid x^\prime-x\mid\le T_g\ll r$. Within this approximation, one gets $$\left<n^b\left(x\right)n^c\left(x^\prime\right)\right>_{\vec n{\,}} \simeq\frac{\delta^{bc}}{3}\int D\vec n{\,}\exp\left(\frac{iJ}{2} \int d\sigma_{\mu\nu}\varepsilon^{def}n^d\left(D_\mu\vec n{\,}\right)^e\left(D_\nu\vec n{\,}\right)^f\right). \eqno (8)$$ It is worth noting, that the integral on the R.H.S. of Eq. (8) is a functional of the world sheet as a whole (it is independent of $x_\mu(\xi)$), and therefore may be also referred to the measure $DB_\mu^a$. Hence, as it was announced in the Introduction, we see that expression (7) for the correction to the string effective action (1) due to the perturbative gluons takes the form of the interaction of two elements of the world sheet, $d\sigma_{\mu\nu}(\xi)$ and $d\sigma_{\rho\nu}\left(\xi^\prime\right)$, via the nonperturbative gluonic string. In order to extract explicitly the dependence on the points $x$ and $x^\prime$ from the functional integral standing on the R.H.S. of Eq. (7), let us pass to the integration over the trajectories $u_\mu(\lambda)=z_\mu(\lambda)+\frac{\lambda}{s}\left(x^\prime- x\right)_\mu - x_\mu^\prime$, which yields $$\Delta S=J^2g^2\int d^2\xi\int d^2\xi^\prime\varepsilon^{ij}\varepsilon^{kl}\left(\partial_ix_\mu\right) \left(\partial_j x_\nu\right)(\partial_k x_\rho^\prime)\left(\partial_lx_\nu^\prime\right)\frac{\partial^2}{\partial x_\mu\partial x_\rho^\prime}\int\limits_0^\infty ds e^{-\frac{\left(x- x^\prime\right)^2}{4s}}\int \left(Du\right)_{00}e^{-\int\limits_0^s\frac {\dot u^2}{4}d\lambda}\cdot$$ $$\cdot\left<tr{\,}P{\,}\exp\left[ig\int\limits_0^s d\lambda \left(\frac{x-x^\prime}{s}+\dot u\right)_\alpha B_\alpha\left(u+x^\prime+ \frac{\lambda}{s}\left(x-x^\prime\right)\right)\right]\right>_{B_\mu^a}, \eqno (9)$$ and from now on we shall absorb the inessential constant factors into the measure $DB_\mu^a$. The Wilson loop standing on the R.H.S. of Eq. (9) may be expanded as follows $$tr{\,}P{\,}\exp\left[ig\int\limits_0^s d\lambda\left(\frac{x-x^\prime}{s} +\dot u\right)_\alpha B_\alpha\left(u+x^\prime+\frac{\lambda}{s}\left(x- x^\prime\right)\right)\right]=tr{\,}P{\,}\exp\left(ig\int\limits_0^s d \lambda\dot u_\alpha B_\alpha(u)\right)+$$ $$+ig{\,}tr\int\limits_0^s d\lambda\left[P{\,}\exp\left(ig\int\limits_0^ \lambda d\lambda^\prime\dot u_\alpha B_\alpha\left(u\left(\lambda^\prime \right)\right)\right)\right]\Biggl[\frac{\left(x-x^\prime\right)_\beta} {s}B_\beta\left(u\left(\lambda\right)\right)+$$ $$+\left(x^\prime+\frac{\lambda} {s}\left(x-x^\prime\right)\right)_\beta\dot u_\gamma\left(\lambda\right) \left(\partial_\beta B_\gamma\left(u\left(\lambda\right)\right)\right)\Biggr] \left[P{\,}\exp\left(ig\int\limits_\lambda^s d\lambda^{\prime\prime}\dot u_ \zeta B_\zeta\left(u\left(\lambda^{\prime\prime}\right)\right)\right)\right]+ O\left(g^2\right),$$ and since we are working in the lowest order of perturbation theory, all the terms of this expansion except for the first one will be omitted below, so that Eq. (9) yields $$\Delta S=J^2g^2\int d^2\xi\int d^2\xi^\prime\varepsilon^{ij}\varepsilon^ {kl}(\partial_ix_\mu)(\partial_jx_\nu)(\partial_kx_\rho^\prime)(\partial_l x_\nu^\prime)\cdot$$ $$\cdot\int\limits_0^\infty\frac{ds}{s}\left(\frac{\left(x-x^\prime \right)_\mu\left(x-x^\prime\right)_\rho}{2s}-\delta_{\mu\rho}\right) e^{-\frac{\left(x-x^\prime\right)^2}{4s}}\Phi (s), \eqno (10)$$ where $$\Phi (s)\equiv\int \left(Du\right)_{00}e^{-\int\limits_0^s\frac{\dot u^2}{4}d\lambda}\left<tr{\,}P{\,}\exp\left(ig\int\limits_0^s d\lambda \dot u_\alpha B_\alpha(u)\right)\right>_{B_\mu^a}. \eqno (11)$$ Finally, in order to get the desirable correction to the action (1), we shall expand the R.H.S. of Eq. (10) in powers of $\frac{s}{r^2}$ (according to the discussion in the paragraph before Eq. (8)), keeping in this expansion terms not higher in the derivatives w.r.t. world sheet coordinates than the rigidity, which corresponds to the expansion up to the second order in the parameter $\frac{s}{r^2}$. Omitting the full derivative terms of the form $\int d^2\xi\sqrt{g}R$, where $R$ is a scalar curvature of the world sheet, we get analogously to Ref. 1 the following values of the integrals standing on the R.H.S. of Eq. (10) $$\int d^2\xi\int d^2\xi^\prime \varepsilon^{ij}\varepsilon^{kl} \left(\partial_ix_\mu\right)\left(\partial_jx_\nu\right)(\partial_k x_\rho^\prime)\left(\partial_lx_\nu^\prime\right) \left(x-x^\prime\right)_\mu \left(x-x^\prime\right)_\rho\int\limits_0^\infty\frac{ds}{s^2} e^{-\frac{\left(x-x^\prime\right)^2}{4s}}\Phi (s)=$$ $$=4\pi\int\limits_0^\infty ds\Phi (s)\left(4\int d^2\xi\sqrt{g}-3s\int d^2\xi\sqrt{g}g^{ij}\left(\partial_i t_{\mu\nu}\right)\left(\partial_j t_{\mu\nu}\right)\right) \eqno (12)$$ and $$\int d^2\xi\int d^2\xi^\prime\varepsilon^{ij}\varepsilon^{kl}\left( \partial_ix_\mu\right)\left(\partial_jx_\nu\right)(\partial_k x_\mu^\prime)\left(\partial_lx_\nu^\prime\right)\int\limits_0^\infty \frac{ds}{s}e^{-\frac{\left(x-x^\prime\right)^2}{4s}}\Phi (s)=$$ $$=2\pi\int\limits_0^\infty ds\Phi (s)\left(4\int d^2\xi\sqrt{g}-s\int d^2\xi\sqrt{g}g^{ij}\left(\partial_i t_{\mu\nu}\right)\left(\partial_j t_{\mu\nu}\right)\right). \eqno (13)$$ Combining together Eqs. (12) and (13), we arrive at the following correction to the effective action (1) due to the accounting for the perturbative gluons in the lowest order of perturbation theory $$\Delta S=\left(\Delta\frac{1}{\alpha_0}\right)\int d^2\xi\sqrt{g}g^{ij}\left(\partial_i t_{\mu\nu}\right)\left(\partial_j t_{\mu\nu}\right), \eqno (14)$$ where $$\left(\Delta\frac{1}{\alpha_0}\right)=J^2g^2\int\limits_0^\infty dss\Phi (s). \eqno (15)$$ Notice, that as it was already pointed out in the Introduction, perturbative gluons do not change the value of the string tension (2) of the Nambu-Goto term and affect only the coupling constant of the rigidity term. Since this correction (15) to the nonperturbative rigid string coupling constant (3) is a pure perturbative effect, its sign is unimportant for the explanation of confinement in terms of the dual Meissner effect (see discussion in the Introduction), which allowed us to refer the constant factor in Eq. (15) to the measure $DB_\mu^a$ in Eq. (11). The nontrivial content of this correction emerges due to the path integral defined by Eq. (11), which includes the background Wilson average. In this letter we have applied perturbation theory in the nonperturbative QCD vacuum$^{9}$ and the non-Abelian Stokes theorem, which represents a Wilson loop in the $SU(2)$ gluodynamics as an integral over all the orientations in colour space$^{16}$ to the derivation of the correction to string effective action (1), found in Ref. 1, which emerges due to accounting for the interaction of perturbative gluons with the string world sheet in the lowest order of perturbation theory. This correction is given by formulae (11), (14) and (15) and affects only the rigidity term, while the string tension of the Nambu-Goto term keeps its pure nonperturbative value (2). Perturbative correction (15) to the inverse coupling constant of the rigidity term contains the dependence on the background fields in the form of the background Wilson average standing under a certain path integral (11). We have also demonstrated that perturbative fluctuations, when being taken into account, lead to the interaction defined by the R.H.S. of Eq. (7) between elements of the string world sheet by virtue of nonperturbative gluonic strings, which agrees with the qualitative scenario of excitation of the gluodynamics string by the perturbative gluons, suggested in Refs. 1 and 10. However, it is still remains unclear whether perturbative gluons may yield cancellation of the conformal anomaly in $D=4$ rather than in $D=26$, as it takes place for the ordinary bosonic string theory$^{18,19}$, and the solution of the problem of crumpling for the rigidity term$^{19,20}$. These problems will be treated in the next publications. I am deeply grateful to Professor Yu.A.Simonov for a lot of useful discussions on the problem studied in this paper. I would also like to thank the theory group of the Quantum Field Theory Department of the Institut für Physik of the Humboldt-Universität of Berlin for kind hospitality. 1. D.V.Antonov, D.Ebert and Yu.A.Simonov, [*Mod.Phys.Lett.*]{} [**A11**]{}, 1905 (1996) (preprint DESY 96-134).\ 2. M.B.Halpern, [*Phys.Rev.*]{} [**D19**]{}, 517 (1979); I.Ya.Aref’eva, [*Theor.Math.Phys.*]{} [**43**]{}, 111 (1980); N.Bralić, [*Phys.Rev.*]{} [**D22**]{}, 3090 (1980).\ 3. Yu.A.Simonov, [*Yad.Fiz.*]{} [**50**]{}, 213 (1989).\ 4. N.G. Van Kampen, [*Stochastic Processes in Physics and Chemistry*]{} (North-Holland Physics Publishing, 1984).\ 5. M.Campostrini, A. Di Giacomo and G.Mussardo, [*Z.Phys.*]{} [**C25**]{}, 173 (1984).\ 6. A. Di Giacomo and H.Panagopoulos, [*Phys.Lett.*]{} [**B285**]{}, 133 (1992).\ 7. I.-J.Ford et al., [*Phys.Lett.*]{} [**B208**]{}, 286 (1988); E.Laermann et al., [*Nucl.Phys.*]{} [**B26**]{} (Proc. Suppl.), 268 (1992).\ 8.  H.G.Dosch, [*Phys.Lett.*]{} [**B190**]{}, 177 (1987); Yu.A.Simonov, [*Nucl.Phys.*]{} [**B307**]{}, 512 (1988); H.G.Dosch and Yu.A.Simonov, [*Phys.Lett.*]{} [**B205**]{}, 339 (1988), [*Z.Phys.*]{} [**C45**]{}, 147 (1989); Yu.A.Simonov, [*Nucl.Phys.*]{} [**B324**]{}, 67 (1989), [*Phys.Lett.*]{} [**B226**]{}, 151 (1989), [*Phys.Lett.*]{} [**B228**]{}, 413 (1989), [*Yad.Fiz.*]{} [**54**]{}, 192 (1991); H.G.Dosch, A. Di Giacomo and Yu.A.Simonov, in preparation.\ 9. Yu.A.Simonov, [*Yad.Fiz.*]{} [**58**]{}, 113, 357 (1995), preprint ITEP 37-95; E.L.Gubankova and Yu.A.Simonov, [*Phys.Lett.*]{} [**B360**]{}, 93 (1995); Yu.A.Simonov, [*Lectures at the 35-th Internationale Universitätswochen für Kern- und Teilchenphysik, Schladming, March 2-9, 1996*]{} (Springer-Verlag, 1996); A.M.Badalian and Yu.A.Simonov, [*Yad.Fiz.*]{} [**60**]{}, 714 (1997); D.V.Antonov, [*Yad.Fiz.*]{} [**60**]{}, 365 (1997) ([*hep-th*]{}/9605044).\ 10. Yu.A.Simonov, [*Nuovo Cim.*]{} [**A107**]{}, 2629 (1984).\ 11. D.V.Antonov and Yu.A.Simonov, [*Int.J.Mod.Phys.*]{} [**A11**]{}, 4401 (1996); D.V.Antonov, [*JETP Lett.*]{} [**63**]{}, 398 (1996), [*Mod.Phys.Lett.*]{} [**A11**]{}, 3113 (1996) ([*hep-th*]{}/9612005), [*Int.J.Mod.Phys.*]{} [**A12**]{}, 2047 (1997), [*Yad.Fiz.*]{} [**60**]{}, 553 (1997) ([*hep-th*]{}/9605045).\ 12. P.Orland, [*Nucl.Phys.*]{} [**B428**]{}, 221 (1994).\ 13. S.Mandelstam, [*Phys.Lett.*]{} [**B53**]{}, 476 (1975); G.’t Hooft, in [*High Energy Physics*]{} (Ed. A.Zichichi) (Editrice Compositori, 1976).\ 14. D.V.Antonov, [*Pis’ma v ZhETF*]{} [**65**]{}, 673 (1997) ([*hep-th*]{}/9612109).\ 15. A.Yu.Dubin, A.B.Kaidalov and Yu.A.Simonov, [*Yad.Fiz.*]{} [**56**]{}, 213 (1993), [*Phys.Lett.*]{} [**B323**]{}, 41 (1994); E.L.Gubankova and A.Yu.Dubin, [*Phys.Lett.*]{} [**B334**]{}, 180 (1994), preprint ITEP 62-94.\ 16. D.I.Diakonov and V.Yu.Petrov, in [*Nonperturbative Approaches to QCD, Proceedings of the International Workshop at ECT\*, Trento, July 10-29, 1995*]{} (Ed. D.I.Diakonov) (PNPI, 1995), [*hep-th*]{}/9606104.\ 17. B.S. De Witt, [*Phys.Rev.*]{}, [**162**]{}, 1195, 1239 (1967); J.Honerkamp, [*Nucl.Phys.*]{} [**B48**]{}, 269 (1972); G.’t Hooft, [*Nucl.Phys.*]{} [**62**]{}, 444 (1973); L.F.Abbot, [*Nucl.Phys.*]{} [**B185**]{}, 189 (1981).\ 18. A.M.Polyakov, [*Phys.Lett.*]{} [**B103**]{}, 207 (1981).\ 19. A.M.Polyakov, [*Gauge Fields and Strings*]{} (Harwood Academic Publishers, 1987).\ 20. A.M.Polyakov, [*Nucl.Phys.*]{} [**B268**]{}, 406 (1986). [^1]: E-mail addresses: [email protected], [email protected], supported by Graduiertenkolleg [*Elementarteilchenphysik*]{}, Russian Fundamental Research Foundation, Grant No.96-02-19184, DFG-RFFI, Grant 436 RUS 113/309/0 and by the Intas, Grant No.94-2851. [^2]: From now on we shall use for the world sheet indices the letters from the middle of the Latin alphabet, $i, j, k,...,$ in order not to confuse them with the colour indices $a, b, c,...$. We shall also keep the standard notation “$g$” for the QCD coupling constant and would like to prevent the reader from confusing it with the determinant of the induced metric tensor, denoted by the same letter.

ArXiv

--- abstract: 'We report a measurement of exclusive [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} cross sections as a function of center-of-mass energy near [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} threshold with initial-state radiation. The analysis is based on a data sample collected with the Belle detector with an integrated luminosity of $967\,\mathrm{fb}^{-1}$.' author: - 'G. Pakhlova' - 'I. Adachi' - 'H. Aihara' - 'K. Arinstein' - 'T. Aushev' - 'T. Aziz' - 'A. M. Bakich' - 'V. Balagura' - 'E. Barberio' - 'A. Bay' - 'K. Belous' - 'V. Bhardwaj' - 'B. Bhuyan' - 'A. Bondar' - 'A. Bozek' - 'M. Bračko' - 'T. E. Browder' - 'A. Chen' - 'P. Chen' - 'B. G. Cheon' - 'R. Chistov' - 'I.-S. Cho' - 'K. Cho' - 'K.-S. Choi' - 'S.-K. Choi' - 'Y. Choi' - 'J. Dalseno' - 'M. Danilov' - 'Z. Doležal' - 'A. Drutskoy' - 'S. Eidelman' - 'D. Epifanov' - 'N. Gabyshev' - 'A. Garmash' - 'B. Golob' - 'H. Ha' - 'J. Haba' - 'K. Hayasaka' - 'H. Hayashii' - 'Y. Horii' - 'Y. Hoshi' - 'W.-S. Hou' - 'H. J. Hyun' - 'T. Iijima' - 'K. Inami' - 'R. Itoh' - 'M. Iwabuchi' - 'Y. Iwasaki' - 'N. J. Joshi' - 'T. Julius' - 'J. H. Kang' - 'P. Kapusta' - 'H. Kawai' - 'T. Kawasaki' - 'C. Kiesling' - 'H. J. Kim' - 'H. O. Kim' - 'M. J. Kim' - 'S. K. Kim' - 'Y. J. Kim' - 'K. Kinoshita' - 'B. R. Ko' - 'S. Korpar' - 'P. Križan' - 'T. Kumita' - 'A. Kuzmin' - 'Y.-J. Kwon' - 'S.-H. Kyeong' - 'J. S. Lange' - 'M. J. Lee' - 'S.-H. Lee' - 'C. Liu' - 'Y. Liu' - 'D. Liventsev' - 'R. Louvot' - 'A. Matyja' - 'S. McOnie' - 'K. Miyabayashi' - 'H. Miyata' - 'Y. Miyazaki' - 'R. Mizuk' - 'G. B. Mohanty' - 'T. Mori' - 'Y. Nagasaka' - 'E. Nakano' - 'M. Nakao' - 'H. Nakazawa' - 'S. Nishida' - 'K. Nishimura' - 'O. Nitoh' - 'T. Nozaki' - 'S. Ogawa' - 'T. Ohshima' - 'S. Okuno' - 'S. L. Olsen' - 'P. Pakhlov' - 'H. Palka' - 'C. W. Park' - 'H. Park' - 'H. K. Park' - 'R. Pestotnik' - 'M. Petrič' - 'L. E. Piilonen' - 'A. Poluektov' - 'M. Röhrken' - 'S. Ryu' - 'H. Sahoo' - 'K. Sakai' - 'Y. Sakai' - 'O. Schneider' - 'C. Schwanda' - 'K. Senyo' - 'M. E. Sevior' - 'M. Shapkin' - 'V. Shebalin' - 'C. P. Shen' - 'J.-G. Shiu' - 'B. Shwartz' - 'F. Simon' - 'P. Smerkol' - 'Y.-S. Sohn' - 'A. Sokolov' - 'E. Solovieva' - 'S. Stanič' - 'M. Starič' - 'T. Sumiyoshi' - 'Y. Teramoto' - 'I. Tikhomirov' - 'K. Trabelsi' - 'S. Uehara' - 'T. Uglov' - 'Y. Unno' - 'S. Uno' - 'G. Varner' - 'K. E. Varvell' - 'A. Vinokurova' - 'A. Vossen' - 'C. H. Wang' - 'M.-Z. Wang' - 'P. Wang' - 'M. Watanabe' - 'Y. Watanabe' - 'R. Wedd' - 'E. Won' - 'Y. Yamashita' - 'C. Z. Yuan' - 'Z. P. Zhang' - 'V. Zhilich' - 'P. Zhou' - 'V. Zhulanov' - 'T. Zivko' - 'A. Zupanc' - 'O. Zyukova' title: | \ Measurement of [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} cross sections near threshold using initial-state radiation --- Recently a number of measurements of exclusive cross sections for [$ e^+e^- $]{} annihilation into charmed hadron pairs above open-charm threshold were performed by the B-factory experiments using initial-state radiation (ISR). These include Belle measurements [@chcon] of [$ e^+e^- $]{} cross sections for the [$ D \overline D $]{} ($D={\ensuremath{ D^0 }}$ or [$ D^+ $]{}), [$ D^+ D^{*-} $]{}, [$ D^{*+} D^{*-} $]{}, [$ D^0 D^- \pi^+ $]{}, [$ {\ensuremath{ D^0 }}{\ensuremath{ D^{*-} }}\pi^+ $]{} and [$ \Lambda_c^+ \Lambda_c^- $]{} final states [@belle:dd; @belle:ddst; @belle:ddp; @belle:ddstp; @belle:ll] and BaBar measurements of the [$ D \overline D $]{}, [$ D \overline D{}^* $]{}, [$ D^* \overline D{}^* $]{} final states [@babar:dd; @babar:ddst], which are, in general, consistent with those of Belle. In addition, CLEO scanned the [$ e^+e^- $]{} center of mass energy range from 3.97 to 4.26[$\, {\mathrm{GeV}} $]{} and obtained exclusive cross sections for the [$ D \overline D $]{}, [$ D \overline D{}^* $]{}, [$ D^* \overline D{}^* $]{}, ${\ensuremath{ D \overline D }}\pi$ and ${\ensuremath{ D \overline D{}^* }}\pi$ final states [@cleo:cs]. The first measurements of the exclusive cross sections for [$ e^+e^- $]{} annihilation into [*charmed strange*]{} meson pairs [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} were performed by CLEO with high accuracy but with limited maximum energy (4.26[$\, {\mathrm{GeV}} $]{}) [@cleo:cs]. Recently BaBar presented [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} cross sections averaged over 100[$\, {\mathrm{MeV}} $]{} wide bins [@babar:dsds]. The observed cross sections were found to be an order of magnitude smaller than those for non-strange charmed meson production. Although the recent BES fit to the total cross section for hadron production in [$ e^+e^- $]{} provided new parameter values for the $\psi$ resonances [@bes:fit], the available [*exclusive*]{} [$ e^+e^- $]{} cross sections have not yet been qualitatively explained. One of the main problems is the numerous open charm thresholds in the region that influence the cross section behavior and, thus, complicate theoretical descriptions. The $Y$ states [@y_states] (with masses above open charm threshold and quantum numbers $J^{PC}= 1^{- -} $), which do not exhibit strong decays to any of the measured open charm final states and have remain unexplained since their discovery more than five years, provide additional motivation to pursue all possible experimental information about the decomposition of charmed particle production in the charm-threshold region. Here we report a measurement of exclusive [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} cross sections as a function of center-of-mass energy from the [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} thresholds to 5.0[$\, {\mathrm{GeV}} $]{}, continuing our studies of the exclusive open charm production in this mass range. The analysis is based on a study of events with ISR photons in a data sample collected with the Belle detector [@det] at the $\Upsilon(nS)$ ($n=1,..., 5$) resonances and nearby continuum with an integrated luminosity of $967\,\mathrm{fb}^{-1}$ at the KEKB [@kekb] asymmetric-energy [$ e^+e^- $]{} collider. We follow the full reconstruction method that was previously used for the measurements of the [$ e^+e^- $]{} cross sections to [$ D \overline D $]{}, [$ D^0 D^- \pi^+ $]{} and [$ {\ensuremath{ D^0 }}{\ensuremath{ D^{*-} }}\pi^+ $]{} final states [@belle:dd; @belle:ddp; @belle:ddstp]. We select [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} signal events in which the [$ D_s^{(*)+} $]{} and [$ D_s^{(*)-} $]{} mesons are fully reconstructed. ISR photon candidates are indicated by [$ \gamma_{\mathrm{ISR}} $]{}. In general, an [$ \gamma_{\mathrm{ISR}} $]{} is not required to be detected and its presence in the event is inferred from a peak at zero in the spectrum of recoil mass squared against the [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} system. The recoil mass squared is defined as: $$\begin{aligned} {\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}})=({\ensuremath{ E_{\mathrm{c.m.}} }}- E_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}})^2 - \nonumber \\ p^2_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}} ,\end{aligned}$$ where [$ E_{\mathrm{c.m.}} $]{} is the initial [$ e^+e^- $]{} center-of-mass ($\mathrm{c.m.}$) energy, $E_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}}$ and $p_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}}$ are energy and three-momentum of the [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} combination, respectively. To suppress backgrounds two cases are considered: (1) the [$ \gamma_{\mathrm{ISR}} $]{} is outside of the detector acceptance and the polar angle for the [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} combination in the $\mathrm{c.m.}$ frame is in the range $|\cos(\theta_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}})|>0.9$; (2) the [$ \gamma_{\mathrm{ISR}} $]{} is within the detector acceptance ($|\cos(\theta_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}})|<0.9$). In the latter case, the [$ \gamma_{\mathrm{ISR}} $]{} is required to be detected and the mass of the ${\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ combination should to be greater than (${\ensuremath{ E_{\mathrm{c.m.}} }}-0.5{\ensuremath{\, {\mathrm{GeV}} }}$). To suppress backgrounds from ${\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}(m)(\pi^+\pi^-){\ensuremath{ \gamma_{\mathrm{ISR}} }}, (m=1,2,...)$ processes, we exclude events that contain additional charged tracks that were not used in the [$ D_s^{(*)+} $]{} and the [$ D_s^{(*)-} $]{} reconstruction. All charged tracks are required to originate from the vicinity of the interaction point (IP); we impose the requirements $|dr|<1 \, {\mathrm{cm}}$ and $|dz|<4\,{\mathrm{cm}}$, where $dr$ and $dz$ are the impact parameters perpendicular to and along the beam direction with respect to the IP, respectively. Charged kaons are required to have a ratio of particle identification likelihood, $\mathcal{P}_K = \mathcal{L}_K / (\mathcal{L}_K + \mathcal{L}_\pi)>0.6$ [@nim]. No identification requirements are applied for pion candidates. $K^0_S$ candidates are reconstructed from $\pi^+ \pi^-$ pairs with an invariant mass within $10{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ of the $K^0_S$ mass. The distance between the two pion tracks at the $K^0_S$ vertex must be less than $1\,\mathrm{cm}$, the transverse flight distance from the interaction point is required to be greater than $0.1\,\mathrm{cm}$, and the angle between the $K^0_S$ momentum direction and the flight direction in the $x-y$ plane should be smaller than $0.1\,\mathrm{rad}$. Photons are reconstructed in the electromagnetic calorimeter as showers with energies greater than $50 {\ensuremath{\, {\mathrm{MeV}} }}$ that are not associated with charged tracks. Pairs of photons are combined to form $\pi^0$ candidates. If the mass of a $\gamma \gamma$ pair lies within $15{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ of the $\pi^0$ mass, the pair is fitted with a $\pi^0$ mass constraint and considered as a $\pi^0$ candidate. ISR photon candidates are required to have energies greater than $2.5 {\ensuremath{\, {\mathrm{GeV}} }}$. Photon candidates used in $\eta$, $\eta^{\prime}$ and [$ D_s^{*+} $]{} reconstruction are required to have energies greater than $100 {\ensuremath{\, {\mathrm{MeV}} }}$. $\eta$ candidates are reconstructed using $\pi^+ \pi^- \pi^0$ ($\pm 10{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ mass window) and $\gamma \gamma$ ($\pm 20{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ mass window) decay modes ($\sim 2.5\,\sigma$ in each case). $\eta^{\prime}$ candidates are reconstructed using $\eta \pi^+ \pi^-$ ($\pm 10{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ mass window) and $\gamma \pi^+ \pi^- $ ($\pm 15{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ mass window) decay modes ($\sim 2.0\,\sigma$ in each case). A mass- and vertex-constrained fit is applied to $\eta$ and $\eta^{\prime}$ candidates. [$ D_s^+ $]{} candidates are reconstructed using six decay modes: $K^0_S K^+$, $K^- K^+ \pi^+$, $K^- K^+ \pi^+ \pi^0$, $K^0_S K^- \pi^+ \pi^+$, $\eta \pi^+$ and $\eta^{\prime}\pi^+$. Before calculation of the [$ D_s^+ $]{} candidate mass, a vertex fit to a common vertex is performed for tracks that form the [$ D_s^+ $]{} candidate. A $\pm 15{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ mass signal window is used for all modes ($\sim 3\,\sigma$ in each case). To improve the momentum resolution of [$ D_s^+ $]{} meson candidates, the tracks from the [$ D_s^+ $]{} candidate are fitted to a common vertex with a mass [$ D_s^+ $]{} mass constraint. [$ D_s^{*+} $]{} candidates are reconstructed using the ${\ensuremath{ D_s^+ }}\gamma$ decay mode. A $\pm 15{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ mass window is used ($\sim 2.5\,\sigma$). A mass-constrained fit is also applied to [$ D_s^{*+} $]{} candidates. The [$ D_s^+ $]{} and [$ D_s^{*+} $]{} sidebands used for background studies are four times as large as the signal region and are divided into windows of the same width as that of the signal. To avoid signal over-subtraction, the selected sidebands are shifted by $30{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ from the signal region. The $D_s(D_s^*)$ candidates from these sidebands are refitted to the central mass value of each window. The ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}})$ distribution for ${\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}} }}<5.0{\ensuremath{\, {\mathrm{GeV}/c^2} }}$ after all the requirements described above is shown in Fig. \[rmx\_dsds\]a). A clear peak corresponding to the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} process is evident near zero recoil mass. The shoulder at positive values is mainly due to ${\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ background. We define the signal region with an asymmetric requirement $-0.7({\ensuremath{\, {\mathrm{GeV}/c^2} }})^2<{\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}})<0.4({\ensuremath{\, {\mathrm{GeV}/c^2} }})^2$ to suppress event this background. The polar angle distribution of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{} combinations and the mass spectrum of the ${\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ combinations (after subtraction of [$ E_{\mathrm{c.m.}} $]{}) in case (2), after ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}})$ requirement, are shown in Figs. \[rmx\_dsds\]b,c). These distributions are in agreement with a MC simulation and are typical of ISR production. The [$ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}} $]{} spectrum with all the requirements applied is shown in Fig. \[rmx\_dsds\]d). A clear peak is seen at threshold near the $\psi(4040)$ mass. ![a) The distribution of ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}})$ for ${\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}} }}<5.0{\ensuremath{\, {\mathrm{GeV}/c^2} }}$ after all the requirements are applied. b) The polar angle distribution of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{} combinations. c) The mass spectrum of the ${\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ combinations after subtraction of [$ E_{\mathrm{c.m.}} $]{} energy in case (2). d) The [$ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}} $]{} spectrum after all the requirements applied. Cross-hached histograms show the normalized [$ M_{{\ensuremath{ D_s^+ }}} $]{} and [$ M_{{\ensuremath{ D_s^- }}} $]{} sideband contributions. Feed down from the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} final state is shown by the open histograms. The signal windows are shown by vertical dashed lines.[]{data-label="rmx_dsds"}](prd_fig1.eps){width="49.00000%"} The ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}})$ distribution for ${\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}} }}<5.0{\ensuremath{\, {\mathrm{GeV}/c^2} }}$ after all the requirements are applied is shown in Fig. \[rmx\_dsdst\]a). A clear peak corresponding to [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} signal process is again evident around zero. We define the signal region for ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}})$ by a requirement $\pm 0.7({\ensuremath{\, {\mathrm{GeV}/c^2} }})^2$ around zero. The polar angle distribution of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} combinations and the mass spectrum of ${\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ combinations (after subtraction of the [$ E_{\mathrm{c.m.}} $]{} energy) in case (2) after the requirement on ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}})$ is applied are shown in Figs. \[rmx\_dsdst\]b,c). The [$ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}} $]{} spectrum after all the requirements are applied is shown in Fig. \[rmx\_dsdst\]d). Two clear peaks are seen at the $\psi(4160)$ and the $\psi(4415)$ masses. ![a) The distribution of the ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}})$ for ${\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}} }}<5.0{\ensuremath{\, {\mathrm{GeV}/c^2} }}$ after all the requirements applied. b) The polar angle distribution of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} combinations. c) The mass spectrum of the ${\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ combinations after subtraction of [$ E_{\mathrm{c.m.}} $]{} in case (2). d) The obtained [$ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}} $]{} spectrum. Cross-hached histograms show the normalized [$ M_{{\ensuremath{ D_s^+ }}} $]{} and [$ M_{{\ensuremath{ D_s^{*-} }}} $]{} sidebands contributions. The small contamination from the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} final state is shown by the open histograms. The signal windows are shown by vertical dashed lines.[]{data-label="rmx_dsdst"}](prd_fig2.eps){width="49.00000%"} The ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}})$ distribution for ${\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}} }}<5.0{\ensuremath{\, {\mathrm{GeV}/c^2} }}$ after all the requirements applied is shown in Fig. \[rmx\_dstdst\]a). A peak corresponding to [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} is again evident around zero recoil mass. We define the signal region for ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}})$ by a requirement $\pm 0.7({\ensuremath{\, {\mathrm{GeV}/c^2} }})^2$ around zero recoil mass. The polar angle distribution of the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} combinations and the mass spectrum of the ${\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ combinations (after subtraction of [$ E_{\mathrm{c.m.}} $]{}) that survive the ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}})$ requirement in case (2) are shown in Figs. \[rmx\_dstdst\]b,c). The full [$ M_{{\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}} $]{} spectrum after all the requirements are applied is shown in Fig. \[rmx\_dstdst\]d). With such limited statistics no structures are evident. ![a) The distributions of the ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}})$ for ${\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}} }}<5.0{\ensuremath{\, {\mathrm{GeV}/c^2} }}$ after all the requirements applied. b) The polar angle distribution of the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} combinations. c) The mass spectrum of the ${\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ combinations after subtraction of [$ E_{\mathrm{c.m.}} $]{} in case (2). d) The [$ M_{{\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}} $]{} spectrum after all the requirements applied. Cross-hached histograms show the normalized [$ M_{{\ensuremath{ D_s^{*+} }}} $]{} and [$ M_{{\ensuremath{ D_s^{*-} }}} $]{} sidebands contributions. The signal windows are shown by vertical dashed lines.[]{data-label="rmx_dstdst"}](prd_fig3.eps){width="49.00000%"} The contribution of multiple entries after all the requirements is found to be $\sim 6\%$, $\sim 22\%$, $\sim 23\%$, for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{}, [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} and the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} final states, respectively. In such cases, only one [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} combination per event, that with the minimum value of $\chi^2_{\mathrm{tot}} = \chi^2_{M({\ensuremath{ D_s^+ }})} + \chi^2_{M({\ensuremath{ D_s^- }})} + (\chi^2_{M({\ensuremath{ D_s^{*+} }})}) + (\chi^2_{M({\ensuremath{ D_s^{*-} }})}$), is used, where $\chi^2_{M({\ensuremath{ D_s^+ }})}$, $\chi^2_{M({\ensuremath{ D_s^- }})}$, $\chi^2_{M({\ensuremath{ D_s^{*+} }})}$ and $\chi^2_{M({\ensuremath{ D_s^{*-} }})}$ correspond to the mass fits for the [$ D_s^+ $]{}, [$ D_s^- $]{}, [$ D_s^{*+} $]{} and the [$ D_s^{*-} $]{} candidates. The following sources of background are considered: (1) combinatorial background under the [$ D_s^{(*)+} $]{}([$ D_s^{(*)-} $]{}) peak combined with a correctly reconstructed [$ D_s^{(*)-} $]{}([$ D_s^{(*)+} $]{}) from the signal or other processes; (2) both the [$ D_s^{(*)+} $]{} and the [$ D_s^{(*)-} $]{} are combinatorial; (3) for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{} final state: reflections from the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} and [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} processes followed by $D_s^* \to D_s \gamma$, where the low momentum $\gamma$ is not reconstructed; for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} final state: reflection from the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} process followed by $D_s^* \to D_s \gamma$, where the low momentum $\gamma$ is not reconstructed; (4) reflection from the ${\ensuremath{ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}}}\pi^0_{miss}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ processes, where the $\pi^0_{miss}$ is not reconstructed. (5) the contribution of ${\ensuremath{ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}}}\pi^0$, when an energetic $\pi^0$ is misidentified as a single [$ \gamma_{\mathrm{ISR}} $]{}. The contribution from background (1) is extracted using the [$ D_s^{(*)-} $]{} and [$ D_s^{(*)+} $]{} sidebands. Background (2) is present in both the [$ M_{{\ensuremath{ D_s^{(*)+} }}} $]{} and [$ M_{{\ensuremath{ D_s^{(*)-} }}} $]{} sidebands and is, thus, subtracted twice. To account for this over-subtraction we use a double sideband region, when events are selected from both the [$ M_{{\ensuremath{ D_s^{(*)+} }}} $]{} and the [$ M_{{\ensuremath{ D_s^{(*)-} }}} $]{} sidebands. The contributions from the combinatorial backgrounds (1–2) are shown in Figs. \[rmx\_dsds\]d) and \[rmx\_dsdst\]d) as cross-hatched histograms. Backgrounds (3–4) are suppressed by the tight requirement on ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}})$. The remaining background (3) for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} final state is estimated using a MC simulation of the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} process. To reproduce the shape of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} mass spectrum we use the initial measurement of the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} mass spectrum. The remainder of background (3) for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{} final state is estimated using a MC simulation of the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} and [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}{\ensuremath{ \gamma_{\mathrm{ISR}} }}$]{} processes. To reproduce the shape of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{} mass spectrum we use the initial measurement of the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} mass spectrum and the first iteration of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} mass spectrum. The contributions from background (3) for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{} and the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} final states are shown in Figs. \[rmx\_dsds\]a), \[rmx\_dsds\]d) and Figs. \[rmx\_dsdst\]a), \[rmx\_dsdst\]d) as open histograms. Uncertainties in these estimates are included in the systematic errors. To estimate the contribution from background (4), we study the ${\ensuremath{ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}}}\pi^0{\ensuremath{ \gamma_{\mathrm{ISR}} }}$ processes using fully reconstructed final states. From a MC study we estimate the fraction of reconstructed events for the cases where the $\pi^0$ is not detected. After the application of the requirement on ${\ensuremath{ M^2_{\mathrm{recoil}} }}({\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}})$ this contribution is found to be less than 0.5% and negligibly small; uncertainties in this estimate are included in the systematic errors. The contribution from background (5), in which an energetic $\pi^0$ is misidentified as the [$ \gamma_{\mathrm{ISR}} $]{} candidate, is determined from the data using fully reconstructed ${\ensuremath{ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}}}\pi^0$ events. Only three events with ${\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}} }}<5.0{\ensuremath{\, {\mathrm{GeV}/c^2} }}$ and $M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}\pi^0}-{\ensuremath{ E_{\mathrm{c.m.}} }}>0.5{\ensuremath{\, {\mathrm{GeV}} }}$ are found in the data. Assuming a uniform $\pi^0$ polar angle distribution, this background contribution in the $|\cos(\theta_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}})|>0.9$ signal sub-sample (case 1) is 3events/9$\epsilon_{\pi^0} \sim 0.6\,$ events in the whole [$ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}} $]{} mass range, where $\epsilon_{\pi^0}$ is the $\pi^0$ reconstruction efficiency. For the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} and the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} final states the expected backgrounds are $\sim 0.6\,$ events and 0 events in the whole [$ M_{{\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}} $]{} and [$ M_{{\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}} $]{} mass ranges. The probability of $\pi^0 \to \gamma$ misidentification due to asymmetric $\pi^0 \to \gamma \gamma$ decays is also estimated to be small. Thus the contribution from background (5) is found to be negligibly small; uncertainties in these estimates are included in the systematic error. The [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} cross sections are extracted from the background subtracted [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} mass distributions $$\begin{aligned} \sigma({\ensuremath{ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}}}) = \frac{ dN/dm }{ \eta_{\mathrm{tot}} dL/dm} \, ,\end{aligned}$$ where $m\equiv{\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}} }}$, $dN/dm$ is the obtained mass spectrum, $\eta_{\mathrm{tot}}$ is the total efficiency [@cs]. The factor $dL/dm$ is the differential ISR luminosity: $$\begin{aligned} dL/dm =\frac{\alpha}{\pi x}\Bigl((2-2x+x^2)\ln\frac{1+C}{1-C}-x^2C\Bigr) \frac{2m \mathcal{L}}{{{\ensuremath{ E_{\mathrm{c.m.}} }}}^2} \, ,\end{aligned}$$ where $x = 1 - m^2/{{\ensuremath{ E_{\mathrm{c.m.}} }}}^2$, $\mathcal{L}$ is the total integrated luminosity and $C = \cos\theta_0$, where $\theta_0$ denotes the polar angle range for [$ \gamma_{\mathrm{ISR}} $]{} in the [$ e^+e^- $]{} c.m. frame: $\theta_0<\theta_{{\ensuremath{ \gamma_{\mathrm{ISR}} }}}<180-\theta_0$. The total efficiencies determined by the MC simulation grow quadratically with energy from 0.015%, 0.010%, 0.005% near threshold to 0.045%, 0.025%, 0.011% at 5.0[$\, {\mathrm{GeV}/c^2} $]{} for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{}, [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} and the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} final states, respectively. The resulting [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} exclusive cross sections averaged over the bin width are shown in Fig. \[cs\]. Since the bin width is much larger than the [$ M_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}} $]{} resolution, which varies from $\sim2{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ around threshold to $\sim6{\ensuremath{\, {\mathrm{MeV}/c^2} }}$ at ${\ensuremath{ M_{{\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}} }}=5.0{\ensuremath{\, {\mathrm{GeV}/c^2} }}$, no correction for resolution is applied. The next-to-leading order radiative corrections are taken into account by the $dL/dm$ formula. The next-to-next-to-leading order corrections are included in the systematics. The contribution of final state radiation (FSR) is strongly suppressed [@quest] and is neglected in this study. ![The cross section averaged over the bin width for a) the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}$]{} process; b) the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}$]{} + c.c. process; c) the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}$]{} process. Error bars show statistical uncertainties only. There is a common systematic uncertainty for all measurements, 11% for [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{}, 17% for [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} and 31% for [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{}. This uncertainty is described in the text. The dotted lines show masses of the $\psi(4040)$, $\psi(4160)$ and $\psi(4415)$ states [@pdg].[]{data-label="cs"}](prd_fig4.eps){width="99.00000%"} The R ratio, defined as $R=\sigma({\ensuremath{ e^+e^- }}\to hadrons)/ \sigma({\ensuremath{ e^+e^- }}\to\mu^+\mu^-)$, where $\sigma({\ensuremath{ e^+e^- }}\to\mu^+\mu^-) = 4\pi\alpha^2/3s$, for the sum of the exclusive [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} cross sections is shown in Fig. \[r\_sum\]. --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![The R ratio for the sum of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{}, [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} and the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} final states. The vertical dotted lines show the masses of the $\psi(4040)$, $\psi(4160)$ and $\psi(4415)$ states [@pdg]. Error bars show statistical uncertainties only.[]{data-label="r_sum"}](prd_fig5.eps "fig:"){width="49.00000%"} --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- The systematic errors for the $\sigma({\ensuremath{ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}}})$ measurements are summarized in Table  \[tab1\]. Source [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{} [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} --------------------------- ----------------------------------------------------- -------------------------------------------------------- ----------------------------------------------------------- Background subtraction $\pm 4\%$ $\pm 7\%$ $\pm 24\%$ Cross section calculation $\pm 7\%$ $\pm 11\%$ $\pm 12\%$ $\mathcal{B}(D_s^{(*)})$ $\pm 5\%$ $\pm 5\%$ $\pm 5\%$ Reconstruction $\pm 6\%$ $\pm 10\%$ $\pm 16\%$ Kaon identification $\pm 2\%$ $\pm 2\%$ $\pm 2\%$ Total $\pm 11\%$ $\pm 17\%$ $\pm 31\%$ : Contributions to the systematic error on the cross sections.[]{data-label="tab1"} The systematic errors associated with the background (1–2) subtraction are estimated from the uncertainty in the scaling factors for the sideband subtractions. This is done using fits to the [$ M_{{\ensuremath{ D_s^{(*)+} }}} $]{} and [$ M_{{\ensuremath{ D_s^{(*)-} }}} $]{} distributions in the data with different signal and background parameterizations and are found to be 3%, 7% and 24% for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{}, [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} and the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} final states, respectively. Uncertainties in the contribution from background (3) are estimated to be 2% for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{} final state and smaller than 1% for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} final state. Uncertainties in the backgrounds (4–5) are estimated conservatively to be smaller than 1% of the signal for all final states. The systematic errors ascribed to the cross section calculation include an error on the differential ISR luminosity (2%) and statistical errors of the MC simulation for the total efficiency calculations. In the case of the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} state there is an additional uncertainty due to the unknown helicity distribution, estimated following a procedure similar to that used in Ref [@belle:ddst]. Another source of systematic error comes from the uncertainties in track and photon reconstruction efficiencies (1% per track, 1.5% per photon and 7% per soft photon). Other contributions include the uncertainty in the identification efficiency and the absolute [$ D_s^+ $]{} and [$ D_s^{*+} $]{} branching fractions [@pdg]. The total systematic uncertainties are 11%, 17% and 31% for the [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}$]{}, [$ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}$]{} and the [$ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}$]{} final states, respectively. In summary, we report the measurement of the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} exclusive cross sections over the center-of-mass energy range from the [$ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}$]{} thresholds to 5.0[$\, {\mathrm{GeV}} $]{}. A clear peak at threshold around the $\psi(4040)$ mass is seen in the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^- }}}}$]{} cross section. In the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}$]{} cross section two peaks are evident around the $\psi(4160)$ and the $\psi(4415)$ masses. The limited statistics do not reveal any structures in the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{*+} }}{\ensuremath{ D_s^{*-} }}}}$]{} cross section. The obtained R ratio for the sum of [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} cross sections has a rich structure including peaks around the $\psi(4040)$, $\psi(4160)$ and the $\psi(4415)$ masses. Both the [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^+ }}{\ensuremath{ D_s^{*-} }}}}$]{} cross section and the R ratio exhibit an obvious dip near the $Y(4260)$ mass, similar to what is seen in ${\ensuremath{ e^+e^- }}\to{\ensuremath{ D^* \overline D{}^* }}$ and in the total cross section for charm production. The obtained cross sections are consistent within errors with those from BaBar [@babar:dsds]. The CLEO exclusive cross sections [@cleo:cs] are not radiatively corrected and, therefore, cannot be directly compared to the results reported here. In this study we do not perform a fit to the obtained [$ {\ensuremath{ e^+e^- }}\to {\ensuremath{ {\ensuremath{ D_s^{(*)+} }}{\ensuremath{ D_s^{(*)-} }}}}$]{} cross sections. The numerous open charm thresholds in this region complicate the cross sections behavior and coupled-channel modifications to the description of any particular final state require one to take into account all other final states contributing to the total cross section for charm production. We thank the KEKB group for the excellent operation of the accelerator, the KEK cryogenics group for the efficient operation of the solenoid, and the KEK computer group and the National Institute of Informatics for valuable computing and SINET3 network support. We acknowledge support from the Ministry of Education, Culture, Sports, Science, and Technology (MEXT) of Japan, the Japan Society for the Promotion of Science (JSPS), and the Tau-Lepton Physics Research Center of Nagoya University; the Australian Research Council and the Australian Department of Industry, Innovation, Science and Research; the National Natural Science Foundation of China under contract No. 10575109, 10775142, 10875115 and 10825524; the Ministry of Education, Youth and Sports of the Czech Republic under contract No. LA10033 and MSM0021620859; the Department of Science and Technology of India; the BK21 and WCU program of the Ministry Education Science and Technology, National Research Foundation of Korea, and NSDC of the Korea Institute of Science and Technology Information; the Polish Ministry of Science and Higher Education; the Ministry of Education and Science of the Russian Federation and the Russian Federal Agency for Atomic Energy; the Slovenian Research Agency; the Swiss National Science Foundation; the National Science Council and the Ministry of Education of Taiwan; and the U.S. Department of Energy. This work is supported by a Grant-in-Aid from MEXT for Science Research in a Priority Area (“New Development of Flavor Physics”), and from JSPS for Creative Scientific Research (“Evolution of Tau-lepton Physics”). [99]{} Charge-conjugate modes are included throughout this paper. G. Pakhlova [*[et al.]{}*]{} (Belle Collab.), Phys. Rev. D [**77**]{}, 011103 (2008). G. Pakhlova [*et al.*]{} (Belle Collab.), Phys. Rev. Lett. [**98**]{}, 092001 (2007). G. Pakhlova [*et al.*]{} (Belle Collab.), Phys. Rev. Lett. [**100**]{}, 062001 (2008). G. Pakhlova [*et al.*]{} (Belle Collab.), Phys. Rev. D [**80**]{}, 091101 (2009). G. Pakhlova [*et al.*]{} (Belle Collab.), Phys. Rev. Lett. [**101**]{}, 172001 (2008). B. Aubert [*et al.*]{} (BaBar Collab.), Phys. Rev. D [**76**]{}, 111105 (2007). B. Aubert [*et al.*]{} (BaBar Collab.), Phys. Rev. D [**79**]{}, 092001 (2009). D. Cronin-Hennessy [*et al.*]{} (CLEO Collab.), Phys. Rev. D [**80**]{}, 072001 (2009). P. del Amo Sanchez [*et al.*]{} (BaBar Collab.), Phys. Rev. D [**82**]{}, 052004 (2010). M. Ablikim [*et al.*]{} (BES Collab.), Phys. Lett. B [**660**]{}, 315 (2008). B. Aubert [*et al.*]{} (BaBar Collab.), Phys. Rev. Lett. [**95**]{}, 142001 (2005); T. E. Coan [*et al.*]{} (CLEO Collab.), Phys. Rev. Lett. [**96**]{}, 162003 (2006); Q. He [ *et al.*]{} (CLEO Collab.), Phys. Rev. D [**74**]{}, 091104 (2006); C. Z. Yuan [*et al.*]{} (Belle Collab.), Phys. Rev. Lett. [**99**]{}, 182004 (2007); B. Aubert [*et al.*]{} (BaBar Collab.), arXiv:0808.1543 \[hep-ex\], (2008); B. Aubert [*et al.*]{} (BaBar Collab.), Phys. Rev. Lett. [**98**]{}, 212001 (2007); X. L. Wang [ *et al.*]{} (Belle Collab.), Phys. Rev. Lett. [**99**]{}, 142002 (2007); Z. Q. Liu, X. S. Qin and C. Z. Yuan, Phys. Rev. D [**78**]{}, 014032 (2008). A. Abashian [*et al.*]{} (Belle Collaboration), Nucl. Instr. and Meth. A [**479**]{}, 117 (2002). S. Kurokawa and E. Kikutani, Nucl. Instr. and Meth. A [**499**]{}, 1 (2003); and other papers included in this volume. E. Nakano, Nucl. Instrum. Meth., A [**494**]{}, 402 (2002). K. Nakamura [*et al.*]{} (Particle Data Group), J. Phys. G [**37**]{}, 075021 (2010). E. A. Kuraev and V. S. Fadin, Sov. J. Nucl. Phys. [**41**]{}, 466 (1985) \[Yad. Fiz [**41**]{}, 733 (1985)\]. S. Actis [*et al.*]{}, Eur. Phys. J. C [**66**]{}, 585 (2010).

ArXiv

--- abstract: 'We consider the initial boundary value problem for the inhomogeneous time-fractional diffusion equation with a homogeneous Dirichlet boundary condition and a nonsmooth right hand side data in a bounded convex polyhedral domain. We analyze two semidiscrete schemes based on the standard Galerkin and lumped mass finite element methods. Almost optimal error estimates are obtained for right hand side data $f(x,t)\in L^\infty(0,T;\dot H^q(\Omega))$, $-1< q \le 1$, for both semidiscrete schemes. For lumped mass method, the optimal $L^2(\Omega)$-norm error estimate requires symmetric meshes. Finally, numerical experiments for one- and two-dimensional examples are presented to verify our theoretical results.' address: 'Department of Mathematics and Institute for Applied Mathematics and Computational Science, Texas A&M University, College Station, TX 77843-3368 ([`btjin,lazarov,pasciak,[email protected]`]{})' author: - Bangti Jin - Raytcho Lazarov - Joseph Pasciak - Zhi Zhou bibliography: - 'frac.bib' date: 'started July, 2012; today is ' title: 'Error Analysis of Semidiscrete Finite Element Methods for Inhomogeneous Time-Fractional Diffusion' --- 1[[E\_[,1]{}]{}]{} Ł2o[[L\_2()]{}]{} \#1[H\^[\#1]{}()]{} \#1\#2[\#1 \_[H\^[\#2]{}()]{}]{} Introduction {#sec:intro} ============ We consider the model initial-boundary value problem for the fractional-order parabolic differential equation (FPDE) for $u(x,t)$: $$\begin{aligned} {3}\label{eq1} \Dal u-\Delta u&= f,&&\quad \text{in } \Omega&&\quad T \ge t > 0,\notag\\ u&=0,&&\quad\text{on}\ \partial\Omega&&\quad T \ge t > 0,\\ u(0)&=v,&&\quad\text{in }\Omega,&&\notag\end{aligned}$$ where $\Omega$ is a bounded convex polygonal domain in $\mathbb R^d\,(d=1,2,3)$ with a boundary $\partial\Omega$, $v$ and $f$ are given functions and $T>0$ is a fixed value. In the model , $\Dal u$ refers to the Caputo fractional derivative of order $\alpha$ ($0<\alpha<1$) of the function $u(t)$, and it is defined by [@KilbasSrivastavaTrujillo:2006 pp. 91, eq. (2.4.1)] $$\Dal u (t) = \frac{1}{\Gamma(1-\alpha)}\int_0^t \frac{1}{(t-s)^\alpha}u'(s)ds.$$ It is known that for the fractional order $\alpha=1$, the fractional derivative $\Dal u$ recovers the canonical first-order derivative $u'(t)$ [@KilbasSrivastavaTrujillo:2006 pp. 92, eq. (2.4.14)], and accordingly the model reduces to the classical time-dependent diffusion problem. Therefore, the model can be regarded as a fractional counterpart of the standard diffusion problem. The interest in is mainly motivated by anomalous diffusion processes, known as subdiffusion, in which the mean square variance grows slower than that in a Gaussian process. At a microscopic level, the diffusion process results from random motion of particles. In a subdiffusion process, the waiting time between consecutive random particle motion can be very large, and thus the mean waiting time diverges. Thus, the underlying stochastic process deviates significantly from the Brownian motion, and instead it can only be more adequately described by the continuous time random walk (CTRW) [@MontrollWeiss:1965]. This microscopic explanation has been frequently exploited in applications. For example, Adams and Gelhar [@AdamsGelhar:1992] observed that field data show anomalous diffusion in a highly heterogeneous aquifer, and Hatano and Hatano [@HatanoHatano:1998] applied the CTRW to model anomalous diffusion in an underground environmental problem. The macroscopic counterpart of the CTRW is a diffusion equation with a time fractional derivative $\Dal u(t)$, i.e. model . It has been successfully applied to many practical problems, including diffusion in media with fractal geometry [@Nigmatulin], the dynamics of viscoelastic materials [@GionaCerbeliRoman:1992], and contaminant transport within underground water flow [@Berkowitz:2002]. The modeling capabilities of FPDEs have generated considerable interest in deriving, analyzing, and testing numerical methods for such problems. As a result, a number of numerical techniques were developed, and their stability and convergence properties were investigated. Yuste and Acedo [@YusteAcedo:2005] presented a numerical scheme by combining the forward time centered space method for the ordinary diffusion equation and Grunwald-Letnikov discretization of the (Riemann-Liouville type) fractional-derivative and provided a von Neumann type stability analysis. Lin and Xu [@LinXu:2007] proposed a numerical method based a finite difference scheme in time and Legendre spectral method in space, showed its unconditional stability, and provided error estimates. Li and Xu [@LiXu:2009] developed a spectral method in both temporal and spatial variables and established various a priori error estimates. Mustapha [@Mustapha:2011] studied semidiscrete in time and fully discrete schemes and derived error bounds for smooth initial data [@Mustapha:2011 Theorem 4.3]. In all these useful studies, the error analysis was carried out under the assumption that the solution is sufficiently smooth. The optimality of the estimates with respect to the solution smoothness expressed through the problem data, i.e., the right hand side $f$ and the initial data $v$, was not considered. Thus, these useful studies do not cover the interesting case of solutions with limited regularity due to low regularity of the data, a typical case for related inverse problems; see, e.g., [@ChengNakagawaYamamoto:2009] and also [@KeungZou:1998] and [@XieZou:2006] for its parabolic counterpart. In our earlier work [@Bangti_LZ_2013], we have analyzed the semidiscrete Galerkin finite element method (FEM) and lumped mass method for problem with a zero right hand side. In particular, almost optimal error estimates were established for both smooth and nonsmooth initial data, i.e., $v\in \dH q$, $0\leq q\leq 1$. (See section 2.2 for the definitions of the space $\dH q$.) More recently in [@JinLazarovPasciakZhou:2013], the results were generalized to the case of very weak initial data, i.e., $v\in\dH q$, $-1<q<0$ . We also refer interested readers to [@McLean-Thomee-IMA; @McLeanThomee:2010] for related studies on FPDEs with a Riemann-Liouville fractional derivative and non-smooth initial data. In this work, we analyze the case of a non-smooth right hand side, i.e., $f\in L^r(0,T;\dH q)$, $-1<q\leq1$, $r>1$. Such problems occur in many practical applications, e.g., optimal control problems and inverse problems [@JinRundell:2012b]. Thus, it is of immense interest to develop and to analyze related numerical schemes. However, with such weak data it might be difficult to define a proper weak solution. In Remark \[weakest\_sol\] below, we note that a function $u(x,t)$ expressed by the representation satisfies the differential equation for $1< r$, but it satisfies (in generalized sense) the initial condition $u(x,0)=0$ for $r >1/{\alpha}$ only. Therefore, the natural class of weak data would be $f\in L^r(0,T;\dH q)$, $-1<q\leq1$, $r>1/{\alpha}$. In fact, all results (upon minor modifications) in this paper are valid for problem with such data. However, for the ease of exposition, we shall assume that $f \in L^\infty(0,T;\dH q)$, $-1<q\leq1$, which guarantees that the representation formula does give a legitimate solution weak solution $u(x,t)$ for all $0 < {\alpha}<1$. The goal of this paper is to develop an error analysis with optimal with respect to the regularity of the right hand side estimates for the semidiscrete Galerkin and the lumped mass Galerkin FEM for problem on convex polygonal domains. Now we describe the numerical schemes, using the standard notation from [@Thomee97]. Let $\{\mathcal{T}_h\}$, $0<h<1$, be a family of shape regular and quasi-uniform partitions of the domain $\Omega$ into $d$-simplices, called finite elements, with $h$ denoting the maximum diameter. The approximate solution $u_h$ will be sought in the finite element space $X_h=X_h(\Omega)$ of continuous piecewise linear functions over the triangulation $\mathcal{T}_h$ $$X_h = \{\chi\in H_0^1(\Omega): \chi \mbox{ is a linear function over } \tau \quad \forall \tau \in \mathcal{T}_h\}.$$ The semidiscrete Galerkin FEM for problem reads: find $u_h(t)\in X_h$ such that $$(\Dal u_h(t), \chi) + a(u_h, \chi) = (f,\chi),\quad \forall \chi\in X_h, \ \ t>0,$$ with $u_h(0)=0$, and $a(u,w) = (\nabla u, \nabla w)$ for $u,w\in H_0^1(\Omega)$. We shall study the convergence of the semidiscrete Galerkin FEM for the case of a right hand side $f\in L^\infty(0,T;\dH q)$, $-1<q\leq1$. The main difficulty in the analysis stems from limited smoothing properties of the solution operator, cf. Lemma \[lem:E\]. This difficulty is overcome by exploiting the mapping property of discrete solution operators established in Lemma \[lem:Eh\]. Our main results are as follows. First in Theorem \[thm:gal:l2\], we derive an optimal $L^2(0,T;\dot H^p(\Omega))$-norm, $p=0,1$, error bound: $$\| u_h - u \|_{L^2(0,T;L^2({\Omega}))} + h\|\nabla(u_h- u)\|_{L^2(0,T;L^2({\Omega}))} \le Ch^{2+\min(q,0)} \|f\|_{L^2(0,T;\dH q)},$$ for $-1<q\leq1$. Second, we derive an almost optimal $L^\infty(0,T;\dot H^p(\Omega))$-norm estimate of the error with an additional log-factor $\ell_h:=|\ln h|$ for $f\in L^\infty(0,T;\dH q)$, $-1<q\leq1$ (cf. Theorem \[thm:gal:linf\]): $$\| u_h(t) - u(t) \|_{L^2(\Omega)} + h \|\nabla(u_h(t) - u(t))\|_{L^2(\Omega)} \le Ch^{2+\min(q,0)} \, \ell_h^{2} \|f\|_{L^\infty(0,t;\dH q)}.$$ Further, we study the more practical lumped mass scheme, and show the same convergence rates for the gradient. For right hand side $f\in L^\infty(0,T;\dH q)$, $-1<q<1$, on general quasi-uniform meshes, we are only able to establish a suboptimal $L^2$-error bound of order $O(h^{1+q}\ell_h^2)$. For a class of special triangulations satisfying condition , (almost) optimal estimates of order $O(h^{2+\min(q,0)})$ and $O(h^{2+\min(q,0)}\ell_h^2)$ hold in $L^2(0,T;L^2(\Omega))$- and $L^\infty(0,T;L^2(\Omega))$-norm, respectively, cf. Theorems \[thm:lump:l2l2\] and \[thm:lump:linfl2\]. These results extend related results for nonsmooth initial data obtained in our earlier works [@JinLazarovPasciakZhou:2013; @Bangti_LZ_2013]. The rest of the paper is organized as follows. In Section \[sec:prelim\], we recall some preliminaries for the convergence analysis, including basic properties of the Mittag-Leffler function, the smoothing property of , and basic estimates for finite element projection operators. In Sections \[sec:galerkin\] and \[sec:lump\], we derive error estimates for the standard Galerkin FEM and lumped mass FEM, respectively. Finally, in Section \[sec:numerics\] we present numerical results for both one- and two-dimensional examples, including non-smooth and very weak data, which confirm our theoretical study. Throughout, we shall denote by $C$ a generic constant, which may differ at different occurrences and may depend on $T$, but it is always independent of the solution $u$ and the mesh size $h$. preliminaries {#sec:prelim} ============= In this part, we recall preliminaries for the convergence analysis, including the Mittag-Leffler function, solution representation, stability estimates, and basic properties of the finite element projection operators. Mittag-Leffler function ----------------------- The Mittag-Leffler function $E_{\alpha,\beta}(z)$ is defined by $$E_{\alpha,\beta}(z) = \sum_{k=0}^\infty \frac{z^k}{\Gamma(k\alpha+\beta)}\quad z\in \mathbb{C} \quad \mbox{with} \quad \Gamma(z) = \int_0^\infty t^{z-1}e^{-t} dt \,\quad ~~\Re(z) >0.$$ The function $E_{\alpha,\beta}(z)$ generalizes the exponential function in that $E_{1,1}(z)= e^z$. The variant $E_{\alpha,\alpha}(z)$, appears in the kernel of the solution representation of problem ; see and below. The following properties are essential in our analysis. \[lem:mlfbdd\] Let $0<\alpha<2$ and $\beta\in\mathbb{R}$ be arbitrary, and $\frac{\alpha\pi}{2} <\mu<\min(\pi,\alpha\pi)$. Then there exists a constant $C=C(\alpha,\beta,\mu)>0$ such that for $\mu\leq|\mathrm{arg}(z)|\leq \pi$ $$\label{M-L-bound} |E_{\alpha,\beta}(z)|\leq \left\{\begin{aligned} \frac{C}{1+|z|^2}, & \quad \beta-\alpha\in \mathbb{Z}^-\cup\{0\},\\ \frac{C}{1+|z|}, & \quad \mbox{otherwise}. \end{aligned}\right.$$ Moreover, for $\lambda>0$, ${\alpha}>0$, and $t>0$ we have $$\label{eq:mlfdiff} \begin{aligned} \Dal E_{{\alpha},1}(-{\lambda}t^{\alpha})&=-{\lambda}E_{{\alpha},1}(-{\lambda}t^{\alpha}) \quad \mbox{and} \quad \frac{d}{dt} E_{\alpha,1}(-{\lambda}t^{\alpha}) = -{\lambda}t^{\alpha-1}E_{{\alpha}.{\alpha}}(-{\lambda}t^{\alpha}). \end{aligned}$$ The estimate can be found in [@KilbasSrivastavaTrujillo:2006 pp. 43, equation (1.8.28)] or [@Podlubny_book Theorem 1.4], while is discussed in [@KilbasSrivastavaTrujillo:2006 Lemma 2.33, equations (2.4.58) and (1.10.7)]. Solution representation {#ssec:represt} ----------------------- For the solution representation to problem , we first introduce some notation. For $s\ge-1$, we denote by $\dH s\subset H^{-1}({\Omega})$ the Hilbert space induced by the norm: $$\|v\|_{\dH s}^2=\sum_{j=1}^{\infty}{\lambda}_j^s \langle v,{\varphi}_j \rangle^2$$ with $\{{\lambda}_j\}_{j=1}^\infty$ and $\{{\varphi}_j\}_{j=1}^\infty$ being respectively the eigenvalues and the $L^2(\Omega)$-orthonormal eigenfunctions of the Laplace operator $-\Delta$ on the domain $\Omega$ with a homogeneous Dirichlet boundary condition. As usual, we identify a function $f$ in $L^2(\Omega)$ with the functional $F$ in $H^{-1}(\Omega)$ defined by $\langle F,\phi\rangle = (f,\phi)$, for all $\phi\in H^1_0(\Omega)$. Then $\{{\varphi}_j\}_{j=1}^\infty$ and $\{{\lambda}_j^{1/2} {\varphi}_j\}_{j=1}^\infty$, form orthonormal basis in $L^2(\Omega)$ and $H^{-1}(\Omega)$, respectively. Further, $\|v\|_{\dH 0}=\|v\|_{L^2(\Omega)}=(v,v)^{1/2}$ is the norm in $L^2(\Omega)$ and $\|v\|_{\dH {-1}} = \|v\|_{H^{-1}(\Omega)}$ is the norm in $H^{-1}(\Omega)$. Besides, it is easy to verify that $\|v\|_{\dH 1}= \|\nabla v\|_{L^2(\Omega)}$ is also the norm in $H_0^1({\Omega})$ and $\|v\|_{\dH 2}=\|\Delta v\|_{L^2(\Omega)}$ is equivalent to the norm in $H^2({\Omega})\cap H^1_0({\Omega})$ (cf. the proof of Lemma 3.1 of [@Thomee97]). Note that $\dot H^s(\Omega)$, $s\ge -1$ form a Hilbert scale of interpolation spaces. Motivated by this, we denote $\|\cdot\|_{H^s(\Omega)}$ to be the norm on the interpolation scale between $H^1_0(\Omega)$ and $L^2(\Omega)$ when $s$ is in $[0,1]$ and $\|\cdot\|_{H^{s}(\Omega)}$ to be the norm on the interpolation scale between $L^2(\Omega)$ and $H^{-1}(\Omega)$ when $s$ is in $[-1,0]$. Then, $\| \cdot \|_{H^s(\Omega)}$ and $\|\cdot\|_{\dH s}$ are equivalent for $s\in [-1,1]$ by interpolation. For a Banach space $B$, we define the space $$L^r(0,T;B) = \{u(t)\in B \mbox{ for a.e. } t\in (0,T) \mbox{ and } \|u\|_{L^r(0,T;B)}<\infty\},$$ for any $r\geq 1$, and the norm $\|\cdot\|_{L^r(0,T;B)}$ is defined by $$\|u\|_{L^r(0,T;B)} = \left\{\begin{aligned}\left(\int_0^T\|u(t)\|_B^rdt\right)^{1/r}, &\quad r\in [1,\infty),\\ \mathrm{esssup}_{t\in(0,T)}\|u(t)\|_B, &\quad r= \infty. \end{aligned}\right.$$ Now we give a representation of the solution to problem using the eigenpairs $\{({\lambda}_j,{\varphi}_j)\}_{j=1}^\infty$. We define an operator $\bar E(t)$ for $\chi \in L^2(\Omega)$ by $$\label{Duhamel} {\bar E}(t) \chi = \sum_{j=1}^\infty t^{{\alpha}-1} \Mitagaa(-{\lambda}_j t^{\alpha})\,(\chi,{\varphi}_j)\, {\varphi}_j(x).$$ Then by separation of variables we get the following representation of the solution $u(x,t)$ to problem for initial data $v=0$: $$\label{Duhamel2} u(x,t)= \int_0^t {\bar E}(t-s) f(s) ds. $$ Our first task is to find the weakest class of right hand side data $f(x,t)$ so that is indeed a solution to the problem . As we see below, for any $f \in L^2(0,T;\dot H^s(\Omega))$, $-1 < s \le 1$ the function $u(x,t)$ from satisfies the differential equation as an element in the space $L^2(0,T;\dot H^{s+2}(\Omega))$. However, it may not satisfy the homogeneous initial condition $u(x,0)=0$. In Remark \[weakest\_sol\] we argue that the weakest class of source term that produces a legitimate weak solution of is $f \in L^r(0,T;\dot H^s(\Omega))$ with $r>1/{\alpha}$ and $-1 < s \le 1$. Obviously, for $1/2<{\alpha}<1$, the representation does give a solution $u(x,t) \in L^2(0,T;\dot H^{s+2}(\Omega))$. We begin with the following important smoothing property of the solution operator $\bar{E}$. \[lem:E\] For any $t>0$, we have for $0\le p-q \le 4$ $$\|\bar E(t) \chi \|_{\dH p} \le Ct^{-1+{\alpha}(1+(q-p)/2)}\|\chi\|_{\dH q}.$$ The definition of $\bar{E}$ in and Lemma \[lem:mlfbdd\] yield $$\begin{split} \|\bar E(t) \chi \|_{\dH p}^2 & =\sum_{j=1}^{\infty}{\lambda}_j^p |t^{{\alpha}-1}\Mitagaa(-{\lambda}_j t^{\alpha})|^2 |(\chi,{\varphi}_j)|^2\\ &=t^{-2+(2+q-p){\alpha}}\sum_{j=1}^{\infty} ({\lambda}_j t^{{\alpha}})^{p-q} |\Mitagaa(-{\lambda}_j t^{\alpha})|^2 {\lambda}_j^q |(\chi,{\varphi}_j)|^2\\ &\le C t^{-2+(2+q-p){\alpha}}\sum_{j=1}^{\infty} \frac{({\lambda}_j t^{{\alpha}})^{p-q}}{(1+({\lambda}_jt^{{\alpha}})^2)^2} {\lambda}_j^q |(\chi,{\varphi}_j)|^2\\ &\le C t^{-2+(2+q-p){\alpha}}\sum_{j=1}^{\infty} {\lambda}_j^q |(\chi,{\varphi}_j)|^2 \le C t^{-2+(2+q-p){\alpha}}\|\chi\|_{\dH q}, \end{split}$$ where in the last line we have used the inequality $\sup_{j}\frac{({\lambda}_j t^{{\alpha}})^{p-q}}{(1+({\lambda}_jt^{{\alpha}})^2)^2} \le C$ for $0\le p-q \le 4$. From this inequality the desired estimate follows immediately. Next we state some stability estimates for the solution $u$ to problem for $f\in L^\infty(0,T;\dH q)$, $-1<q\leq 1$. These estimates will be essential for the convergence analysis of the standard Galerkin FEM in Section \[sec:galerkin\]. The first estimate in Theorem \[thm:reg\] in the case $q=0$ was already established in [@Sakamoto_2011 Theorem 2.1, part (i)]. Below it is extended for the whole range of $q$. \[thm:reg\] Assume that $f\in L^2(0,T;\dH q)$, $-1<q \le 1$. Then the expression represents a function $u\in L^2(0,T;\dot H^{q+2}(\Omega))$ which satisfies the differential equation in and the estimate: $$\label{eq:reg} \|u\|_{L^2(0,T;\dH {q+2})}+\|\partial_t^\alpha u\|_{L^2(0,T;\dH {q})}\leq C\|f\|_{L^2(0,T;\dH q)}.$$ If $f\in L^\infty(0,T;\dH q)$, $-1<q \le 1$, then the function $u(x,t)$ belongs to $ L^{\infty}(0,T;\dot H^{q+2-\epsilon}(\Omega))$ and satisfies $$\label{eq:regeps} \|u(\cdot,t)\|_{\dH {q+2-\epsilon}} \leq C\epsilon^{-1}t^{{\epsilon}{\alpha}/2}\|f\|_{L^\infty(0,t;\dH q)} \quad \mbox{ for any} \quad \epsilon>0 .$$ Hence, is a solution to the initial value problem with a homogeneous initial data $v=0$. By the complete monotonicity of the function $E_{\alpha,1}(-t^\alpha)$ (with $\alpha\in (0,1)$) [@Sakamoto_2011 Lemma 3.3], i.e., $$(-1)^n\frac{d^n}{dt^n}E_{\alpha,1}(-t^\alpha)\geq 0 \quad\mbox{ for all } t>0,\ \ n=0,1,\ldots,$$ and the differentiation formula in Lemma \[lem:mlfbdd\], we deduce $E_{\alpha,\alpha}(-\eta)\geq0$, $\eta\geq0$. Therefore, for $t>0$ $$\label{ineq1} \begin{aligned} \int_0^t |t^{\alpha-1}E_{\alpha,\alpha}(-\lambda_n t^\alpha)|dt & = \int_0^t t^{\alpha-1} E_{\alpha,\alpha}(-\lambda_n t^\alpha)dt \\ & = -\frac{1}{\lambda_n}\int_0^t \frac{d}{dt}E_{\alpha,1}(-\lambda_n t^\alpha)dt\\ & = \frac{1}{\lambda_n}(1-E_{\alpha,1}(-\lambda_nt^\alpha)) \le \frac{1}{\lambda_n}. \end{aligned}$$ Meanwhile, by the differentiation formula [@KilbasSrivastavaTrujillo:2006 pp. 140-141], we get $$\begin{aligned} A_n:=& \partial_t^\alpha \int_0^t(f(\cdot,\tau),{\varphi}_n)(t-\tau)^{\alpha-1}E_{\alpha,\alpha}(-\lambda_n(t-\tau)^\alpha)d\tau\\ =& (f(\cdot,t),{\varphi}_n) -\lambda_n \int_0^t(f(\cdot,\tau),{\varphi}_n)(t-\tau)^{\alpha-1}E_{\alpha,\alpha}(-\lambda_n(t-\tau)^\alpha)d\tau . \end{aligned}$$ Now by means of Young’s inequality for convolution, we deduce $$\begin{aligned} \|A_n\|_{L^2(0,T)}^2 &\leq C_1\int_0^T|(f(\cdot,t),{\varphi}_n)|^2 dt + C_2\int_0^T |(f(\cdot,t),{\varphi}_n)|^2 dt\left(\int_0^T|\lambda_nt^{\alpha-1}E_{\alpha,\alpha}(-\lambda_nt^\alpha)|dt\right)^2\\ &\leq C\int_0^T|(f(\cdot,t),{\varphi}_n)|^2 \, dt. \end{aligned}$$ Thus there holds $$\begin{aligned} \|\partial_t^\alpha u \|_{L^2(0,T;\dH q)}^2 &= \sum_{n=1}^\infty \int_0^T\lambda_n^q|\partial_t^\alpha\int_0^t(f(\cdot,\tau),{\varphi}_n)(t-\tau)^{\alpha-1}E_{\alpha,\alpha}(-\lambda_n(t-\tau)^{\alpha})d\tau|^2 dt \\ &\leq C\sum_{n=1}^\infty \int_0^T\lambda_n^q|(f(\cdot,t),\phi_n)|^2 dt= C\|f\|_{L^2(0,T;\dH q)}^2. \end{aligned}$$ Now using equation and the triangle inequality, we also get $\|\Delta u\|_{L^2(0,T;\dH q)} \leq C\|f\|_{L^2(0,T;\dot{H}^q(\Omega))}$. This shows the first assertion. By Lemma \[lem:E\] we have $$\begin{aligned} \|u(\cdot,t)\|_{\dH {q+2-{\epsilon}}} & = \|\int_0^t \bar{E}(t-s)f(s)ds\|_{\dH {q+2-{\epsilon}}} \leq \int_0^t \|\bar{E}(t-s)f(s)\|_{\dH {q+2-{\epsilon}}} ds \\ & \leq C\int_0^t (t-s)^{{\epsilon}{\alpha}/2-1} \|f(s)\|_{\dH q}ds \leq C\epsilon^{-1}t^{{\epsilon}{\alpha}/2}\|f\|_{L^\infty(0,t;\dot H^q(\Omega))} \end{aligned}$$ which shows estimate . Finally, it is follows directly from this that the representation $u$ satisfies also the initial condition $u(0)=0$, i.e., for any ${\epsilon}>0$, $\lim_{t\to 0 }\|u(\cdot,t)\|_{\dH {q+2-{\epsilon}}}=0$, and thus it is indeed a solution of the initial value problem . The first estimate in Theorem \[thm:reg\] can be improved to $$\|u\|_{L^r(0,T;\dH {q+2})}+\|\partial_t^\alpha u\|_{L^r(0,T;\dot H^{q})}\leq C\|f\|_{L^r(0,T;\dH q)}$$ for any $r\in(1,\infty)$. The proof is essentially identical. The $\epsilon$ factor in the second estimate reflects the limited smoothing property of the fractional differential operator, resulting from the slow decay of the Mittag-Leffler function $E_{\alpha,\alpha}(-z)$. \[weakest\_sol\] The condition $f\in L^\infty(0,T;\dH q)$ can be weakened to $f\in L^r(0,T;\dH q)$ with $r>1/\alpha$. This follows from Lemma \[lem:E\] and the Cauchy-Schwarz inequality with $r'$, $1/r' + 1/r=1$ $$\begin{aligned} \|u(\cdot,t)\|_{\dH q} & \leq \int_0^t \|\bar{E}(t-s)f(s)\|_{\dH {q}} ds \leq \int_0^t (t-s)^{\alpha-1} \|f(s)\|_{\dH q}ds \\ & \leq \tfrac{C}{1+r'(\alpha-1)}t^{1+r'(\alpha-1)}\|f\|_{L^r(0,t;\dot H^q(\Omega))}, \end{aligned}$$ where $1+r'(\alpha-1)>0$ by the condition $r>1/\alpha$. It follows from this that the initial condition $u(x,0)=0$ holds in the following sense: $\lim_{t\to0^+}\|u(t)\|_{\dH q}=0$. Hence for any $\alpha\in (1/2,1)$ the representation formula remains a legitimate solution under the weaker condition $f\in L^2(0,T;\dH q)$. \[L-infinity\] For the ease of exposition, in the error analysis we shall restrict our discussion to the case $f\in L^\infty(0,T;\dH q)$. Nonetheless, we note that for $p=0,1$ the $L^2(0,T;\dH p)$-norm estimate of the error below remain valid under the weakened regularity condition on the source term $f$. Ritz and $L^2$-orthogonal projections ------------------------------------- In our analysis we shall also use the $L^2$-orthogonal projection $P_h:L^2(\Omega)\to X_h$ and the Ritz projection $R_h:H^1_0(\Omega)\to X_h$, respectively, defined by $$\begin{aligned} (P_h \psi,\chi) & =(\psi,\chi) \quad\forall \chi\in X_h,\\ (\nabla R_h \psi,\nabla\chi) & =(\nabla \psi,\nabla\chi) \quad \forall \chi\in X_h. \end{aligned}$$ We shall need some properties of the Ritz projection $R_h$ and the $L^2$-projection $P_h$ [@JinLazarovPasciakZhou:2013]. \[lem:prh-bound\] Let the mesh be quasi-uniform. Then the operator $R_h$ satisfies: $$\begin{aligned} \|R_h \psi-\psi\|_{L^2(\Omega)}+h\|\nabla(R_h \psi-\psi)\|_{L^2(\Omega)}\le Ch^q\| \psi\|_{\dH q}{\quad{\mbox{for}}\quad}\psi \in \dH q, \ q=1,2.\end{aligned}$$ Further, for $s\in [0,1]$ we have $$\begin{aligned} \|(I-P_h)\psi \|_{H^s(\Omega)} &\le Ch^{2-s} \|\psi\|_{\dH 2} {\quad{\mbox{for}}\quad}\psi\in H^2(\Omega)\cap H^1_0(\Omega),\\ \|(I-P_h)\psi \|_{H^s(\Omega)} &\le Ch^{1-s} \|\psi\|_{\dH 1} {\quad{\mbox{for}}\quad}\psi\in H^1_0(\Omega). \end{aligned}$$ In addition, by duality $P_h$ is stable on $\dH s$ for $s\in [-1,0]$. Semidiscrete Galerkin FEM {#sec:galerkin} ========================= Finite element method {#ssec:fem} --------------------- The semidiscrete Galerkin FEM for problem with $v=0$ is: find $u_h \in X_h$ such that $$\label{fem} \begin{split} {( \Dal u_{h},\chi)}+ a(u_h,\chi)&= {(f, \chi)}, \quad \forall \chi\in X_h,\ T \ge t >0 \quad \mbox{and} \quad u_h(0)=0. \end{split}$$ Upon introducing the discrete Laplace operator ${\Delta}_h: X_h\to X_h$ $$\label{eqn:Delh} -({\Delta}_h\psi,\chi)=(\nabla\psi,\nabla\chi)\quad\forall\psi,\,\chi\in X_h,$$ the spatial discrete problem can be written as $$\label{fem-operator} \Dal u_{h}(t)-{\Delta}_h u_h(t) =f_h(t) {\quad{\mbox{for}}\quad}t\ge0 \quad \mbox{with} \quad u_h(0)=0,$$ where the discrete source term $f_h=P_h f$. Then the solution of can be represented by the eigenpairs $\{({\lambda}^h_j,{\varphi}_j^h)\}_{j=1}^{\NN}$ of the discrete Laplacian $-{\Delta}_h$. Now we introduce the discrete analogue $\bar{E}_h $ of the solution operator $\bar{E}$ defined in for $t>0 $: $$\label{E-tilde} \Etilh(t) f_h= \sum_{j=1}^\NN t^{{\alpha}-1} \Mitagaa(-{\lambda}^h_jt^{\alpha})\,(f_h,{\varphi}^h_j) \, {\varphi}_j^h.$$ Then the solution $u_h(x,t)$ of the discrete problem can be expressed by: $$\label{Duhamel_o} u_h(x,t)= \int_0^t \Etilh(t-s) f_h(s)\,ds.$$ Now we define the discrete norm $\normh{\cdot}{p}$ on $X_h$ for any $p\in\mathbb{R}$ $$\label{eqn:normhp} \normh{\psi}{p}^2 = \sum_{j=1}^N({\lambda}_j^h)^p(\psi,{\varphi}_j^h)^2\quad \psi\in X_h.$$ Analogously, we introduce the associated spaces $\tribar \cdot\tribar_{L^r(0,T;\dH p)}$, $r\in[1,\infty]$, on the space $X_h$. We have the following norm equivalence and inverse inequality. \[lem:inverse\] For all $\psi \in X_h$ and any real $l>s$ $$ \normh{ \psi}{l}\le Ch^{s-l}\normh{\psi}{s}.$$ For any $s\in [-1,1]$, the norms $\normh{\cdot}{s}$ and $\|\cdot\|_{\dH s}$ are equivalent on $X_h$. The inverse estimate was shown in [@Bangti_LZ_2013 Lemma 3.3]. By the definition of the $\normh{\cdot}{s}$-norm and the discrete Laplace operator, it is easy to show $\normh{\cdot}{s}$ is equivalent to $\|\cdot\|_{\dH s}$ with $s=0,1$. Then the assertion for $0\le s\le 1$ follows by interpolation, and by duality it is also equivalent to $\|\cdot\|_{\dH s}$ for $-1\leq s\leq0$. Properties of the discrete solution {#ssec:sol} ----------------------------------- Next, analogous to Lemma \[lem:E\], we introduce some smoothing properties of $\Etilh(t)$. The proof is identical with that for Lemma \[lem:E\], and hence omitted. \[lem:Eh\] Let $\Etilh$ be defined by and $ \psi \in X_h$. Then we have for all $t >0$, $$\normh{ \Etilh(t) \psi }{p} \le \left \{ \begin{array}{ll} Ct^{ -1 + {\alpha}(1 + (q-p)/2)}\normh{\psi}{q}, & \quad p-4\leq q \le p, \\[1.3ex] Ct^{ -1 + \alpha }\normh{\psi}{q}, & \quad p< q. \end{array}\right .$$ The following estimate is the discrete analogue of Theorem \[thm:reg\]. It is essential for the analysis of the lumped mass method in Section 4. \[lem:reg-d\] Let $u_h$ be the solution of . Then for arbitrary $p>-1$ $$\label{Gal-L2-est-1} \begin{aligned} \int_0^T\normh{\Dal u_h(t)}{p}^2 +\normh{ u_h(t) }{p+2}^2 \, dt & \le \int_0^T \normh {f_h(t)}{p}^2 dt, \end{aligned}$$ and $$\label{Gal-L2-est-2} \begin{aligned} \normh{ u_h(t) }{p+2-\epsilon} & \leq C\epsilon^{-1}t^{{\epsilon}{\alpha}/2}\tribar f_h\tribar_{L^\infty(0,t;\dH {p})}. \end{aligned}$$ The solution $u_h(t)$ of can be represented by , and hence $$\begin{split} \normh {u_h(t)}{p+2}^2 =&\sum_{j=1}^{N} ({\lambda}_j^h)^{p+2} \Big |\int_0^t (t-\tau)^{{\alpha}-1} \Mitagaa(-{\lambda}^h_j(t-\tau)^{\alpha})(f_h(\cdot,\tau),{\varphi}_j^h)\,d\tau \Big |^2\\ =&\sum_{j=1}^{N} ({\lambda}_j^h)^p \Big |\int_0^t {\lambda}_j^h(t-\tau)^{{\alpha}-1} \Mitagaa(-{\lambda}^h_j(t-\tau)^{\alpha}) (f_h(\cdot,\tau),{\varphi}_j^h)\,d\tau \Big |^2. \end{split}$$ Then by Young’s inequality for convolution, we deduce $$\begin{split} \int_0^T\normh{ u_h(t) }{p+2}^2 \, dt & \le \sum_{j=1}^N ({\lambda}_j^h)^p \int_0^T |(f_h(\cdot,t),{\varphi}_j^h)|^2 \,dt \left( \int_0^T {\lambda}_j^h t^{{\alpha}-1} \Mitagaa(-{\lambda}^h_j t^{\alpha}) \, dt\right)^2\\ &\le C \int_0^T \normh {f_h(t)}{p}^2 dt, \end{split}$$ where the last line follows from the identity $\int_0^T {\lambda}_j^h t^{{\alpha}-1} \Mitagaa(-{\lambda}^h_j t^{\alpha})dt=1-E_{\alpha,1}(-{\lambda}^h_jt^\alpha)\in (0,1)$, cf. . Now the first estimate follows from this and the triangle inequality and equation . The second estimate follows from Lemma \[lem:Eh\] $$\begin{aligned} \normh {u_h(\cdot,t)}{p+2-\epsilon} & = \normh{\int_0^t \bar{E}_h(t-s)f_h(s)ds}{p} \leq \int_0^t \normh {\bar{E}_h(t-s)f_h(s)}{p+2-{\epsilon}} ds \\ & \leq \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1} \normh {f_h}{p} \leq C\epsilon^{-1}t^{{\epsilon}{\alpha}/2}\tribar f_h\tribar_{L^\infty(0,t; H^{p}(\Omega))}. \end{aligned}$$ This completes the proof of the lemma. The rest of this part is devoted to the error analysis for the semi-discrete Galerkin scheme for a nonsmooth source term $f\in L^\infty(0,T;\dH q) $, $-1<q\leq 1$. To this end, we employ the $L^2$-projection $P_hu$, and split the error $u_h-u$ into two terms as: $$\label{separate} u_h-u=(u_h-P_hu)+(P_hu-u):={{\widetilde \vartheta}}+ {{\widetilde \varrho}}.$$ Obviously, $ P_h \Dal {{\widetilde \varrho}}= \Dal P_h(P_hu-u)=0$ and using the identity $\Delta_hR_h=P_h\Delta$, we get the following equation for ${{\widetilde \vartheta}}$: $$\label{eq:thettil} \Dal {{\widetilde \vartheta}}(t) -\Delta_h {{\widetilde \vartheta}}(t) = - \Delta_h (R_h u - P_h u)(t), \quad t>0, \quad {{\widetilde \vartheta}}(0)=0.$$ Then in view of the representation formula , ${{\widetilde \vartheta}}(t)$ can be represented by $$\label{eqn:vtht} {{\widetilde \vartheta}}(t) = - \int_0^t\Etilh(t-s)\Delta_h(R_hu-P_hu)(s)\,ds.$$ Next we shall treat the $L^2$- and $L^\infty$- in time error estimates separately, due to the different stability estimates in the two cases, cf. Theorem \[thm:reg\]. Error estimates for solutions in $L^2(0,T;\dot H^p(\Omega))$ ------------------------------------------------------------ In this part, we establish error estimates in $L^2$-norm in time. The case $-1<q\leq 0$ is stated in the next theorem, while the case $0<q\leq 1$ follows directly and is commented in Remark \[rmk:gal:l2\] below. \[thm:gal:l2\] Let $f\in L^\infty(0,T;\dH q)$, $-1<q\leq 0$, and $u$ and $u_h$ be the solutions of and with $f_h=P_hf$, respectively. Then $$\| u_h - u \|_{L^2(0,T;L^2({\Omega}))} + h\|\nabla(u_h- u)\|_{L^2(0,T;L^2({\Omega}))} \le Ch^{2+q} \|f\|_{L^2(0,T;\dH q)}.$$ We use the splitting . By Theorem \[thm:reg\] and Lemma \[lem:prh-bound\] $$\begin{split} \| {{\widetilde \varrho}}\|_{L^2(0,T;L^2({\Omega}))} + h \|\nabla {{\widetilde \varrho}}\|_{L^2(0,T;L^2({\Omega}))} &\le Ch^{2+q} \|u\|_{L^2(0,T;\dH {2+q})} \le Ch^{2+q} \|f\|_{L^2(0,T;\dH q)}. \end{split}$$ By , and Lemmas \[lem:reg-d\] and \[lem:prh-bound\], we have for $p=0,\,1$: $$\begin{split} \int_0^T \|{{\widetilde \vartheta}}(t)\|^2_{p} dt & \le C\int_0^T \normh{\Delta_h (R_h u - P_h u)(t) }{p-2}^2 dt \\ & \le C\int_0^T \normh{(R_h u - P_h u)(t)}{p}^2 dt\\ & \le C h^{4+2q-2p} \| u(t) \|_{L^2(0,T;\dH {2+q})}^2 \le C h^{4+2q-2p}\| f(t) \|_{L^2(0,T;\dH q)}^2. \end{split}$$ Combing the preceding two estimates yields the desired assertion. \[rmk:gal:l2\] Theorem \[thm:reg\] implies that for $0<q\leq 1$, there holds $$ \| u_h - u \|_{L^2(0,T;L^2({\Omega}))} + h\|\nabla(u_h- u)\|_{L^2(0,T;L^2({\Omega}))} \le Ch^2\|f\|_{L^2(0,T;\dH q)}.$$ Error estimates for solutions in $L^\infty(0,T;\dot H^p(\Omega))$ ----------------------------------------------------------------- Now we turn to error estimates in $L^\infty$-norm in time. Like before, we first consider the case $-1<q\leq 0$. By Theorem \[thm:reg\] and Lemma \[lem:prh-bound\], the following estimate holds for ${{\widetilde \varrho}}$: $$\begin{split} \| {{\widetilde \varrho}}(t) \|_{L^2(\Omega)} + h \|\nabla{{\widetilde \varrho}}(t)\|_{L^2(\Omega)} &\le Ch^{2+q-\epsilon} \|u(t)\|_{\dH {2+q-\epsilon}} \le C\epsilon^{-1}h^{2+q-\epsilon}\|f\|_{L^\infty(0,t;\dH q)}. \end{split}$$ Now the choice $\ell_h= |\ln h|,\, \epsilon=1/\ell_h$ yields $$\label{Ph-bound} \| {{\widetilde \varrho}}(t) \|_{L^2(\Omega)} + h \|\nabla{{\widetilde \varrho}}(t)\|_{L^2(\Omega)}\le C\ell_hh^{2+q} \|f\|_{L^\infty(0,t;\dH q)}.$$ Thus, it suffices to bound the term ${{\widetilde \vartheta}}$, which is shown in the following lemma. \[lem:vtht\] Let ${{\widetilde \vartheta}}(t)$ be defined by . Then for $f\in L^\infty(0,T;\dH q)$, $-1<q\leq 0$, we have $$\|{{\widetilde \vartheta}}(t)\|_{L^2(\Omega)}+h\|\nabla{{\widetilde \vartheta}}(t)\|_{L^2(\Omega)}\leq Ch^{2+q}\ell_h^2\|f\|_{L^\infty(0,t;\dH q)} \quad \mbox{with } \quad \ell_h=|\ln h|.$$ By and Lemmas \[lem:inverse\] and \[lem:Eh\], we deduce that for $p=0,1$ $$\begin{aligned} \|{{\widetilde \vartheta}}(t)\|_{\dH p} &\leq \int_0^t \|\Etilh(t-s)\Delta_h(R_h u-P_hu)(s)\|_{\dH p} ds\\ &\leq C\int_0^t (t-s)^{\epsilon{\alpha}/2 - 1} \normh {\Delta_h(R_h u -P_hu)(s)}{p-2+\epsilon}ds\\ & \leq C\int_0^t(t-s)^{\epsilon{\alpha}/2-1} \normh{R_h u(s) -P_hu(s)}{p+\epsilon}ds := A. \end{aligned}$$ Further, we apply the inverse estimate from Lemma \[lem:inverse\] for $R_h u - P_h u$ and the bounds in Lemma \[lem:prh-bound\], for $P_h u -u$ and $R_hu -u$, respectively, to deduce $$\begin{aligned} A & \leq C h^{-\epsilon} \int_0^t(t-s)^{\epsilon{\alpha}/2-1} \| R_h u(s) -P_hu(s)\|_{\dot H^p(\Omega)}ds \\ & \leq C h^{2+q-p-2\epsilon} \int_0^t(t-s)^{\epsilon{\alpha}/2-1} \| u(s) \|_{\dH {2+q-\epsilon}}ds. \end{aligned}$$ Further, by applying estimate and choosing ${\epsilon}=1/\ell_h$ we get $$\begin{aligned} A & \leq C \epsilon^{-1} h^{2+q-p-2\epsilon} \|f\|_{L^\infty(0,t;\dH q)} \int_0^t(t-s)^{\epsilon{\alpha}/2-1}t^{\epsilon \alpha/2}ds \le Ch^{2+q-p}\ell_h^2\|f\|_{L^\infty(0,t;\dH q)}. \end{aligned}$$ This completes the proof of the lemma. Now we can state the main result of this part, namely, an almost optimal error estimate of the Galerkin approximation $u_h$ for solutions $u \in L^\infty(0,T;\dH p)$. \[thm:gal:linf\] Let $f\in L^\infty(0,T;\dH q)$, $-1<q\leq0$, and $u$ and $u_h$ be the solutions of and with $f_h=P_hf$, respectively. Then with $\ell_h =| \ln h|$, there holds $$\| u_h(t) - u(t) \|_{L^2({\Omega})} + h \|\nabla(u_h(t) - u(t))\|_{L^2(\Omega)} \le Ch^{2+q} \ell_h^{2} \|f\|_{L^\infty(0,t;\dH q)}.$$ In comparison with the $L^2(0,T;\dH p)$-norm estimates, the $L^\infty(0,T;\dH p)$-norm estimates suffer from the factor $\ell_h^2$. This is due to Lemma \[lem:Eh\] and the regularity estimate in Theorem \[thm:reg\] (in turn this goes back to Lemma \[lem:E\]), which is attributed to the limited smoothing property of the time-fractional differential operator. \[rmk:gallogh\] An inspection of the proof of Lemma \[lem:vtht\] indicates that for $0<q\leq1$, one can get rid of one factor $\ell_h$, and hence there holds $$\| u_h(t) - u(t) \|_{L^2({\Omega})} + h \|\nabla(u_h(t) - u(t))\|_{L^2({\Omega})} \le Ch^2 \ell_h\|f\|_{L^\infty(0,t;\dH q)}.$$ Problems with more general elliptic operators {#general_problems} --------------------------------------------- We can study problem with a more general bilinear form $a(\cdot, \cdot): V \times V\mapsto \mathbb{R}$ of the form: $$\label{general-form} a(u,\chi) = \int_{\Omega}(k(x) \nabla u \cdot \nabla \chi + q(x) u \chi)\, dx,$$ where $k(x) $ is a symmetric $d \times d$ matrix-valued measurable function on the domain $\Omega$ with smooth entries and $q(x)$ is an $L^\infty(\Omega)$-function. We assume that $$c_0|\xi|^2\leq \xi^T k(x) \xi \leq c_1 |\xi|^2,\ {\quad{\mbox{for}}\quad}\xi \in {\mathbb R}^d,\ x \in {\Omega},$$ where $c_0,c_1>0$ are constants, and the bilinear form $a(\cdot, \cdot)$ is coercive on $V \equiv H_0^1({\Omega})$. Further, we assume that the problem $a(u,\chi)=(f,\chi), \, \forall \chi \in V$ has a unique solution $u\in V$, with the fully regularity $\|u\|_{H^2(\Omega)} \le C\|f\|_{L^2(\Omega)}$. Under these conditions we can define a positive definite operator ${\mathcal A}: H^1_0(\Omega) \to H^{-1}(\Omega),$ which has a complete set of eigenfunctions ${\varphi}_j(x)$ and respective eigenvalues ${\lambda}_j({\mathcal A})>0$. Then we can define the spaces $\dH q$ as in Section \[ssec:represt\] and all the equivalent properties are satisfied. Further, we define the discrete operator ${\mathcal A}_h: X_h \to X_h$ by $$({\mathcal A}_h \psi,\chi) = a(\psi,\chi), ~~\forall \psi, \chi \in X_h .$$ Then all results for problem can be easily extended to fractional-order problems with an elliptic part of this more general form. Lumped mass method {#sec:lump} ================== In this section, we consider the semidiscrete scheme based on the more practical lumped mass FEM (see, e.g. [@Thomee97 Chapter 15, pp. 239–244]) and derive related error estimates. Lumped mass FEM --------------- For completeness we shall first introduce this approximation. Let $z_j^\tau$, $j=1,\dots,d+1$ be the vertices of the $d$-simplex $\tau \in \mathcal{T}_h$. Consider the quadrature formula $$\label{quadrature} Q_{\tau,h}(f) = \frac{|\tau|}{d+1} \sum_{j=1}^{d+1} f(z^\tau_j) \approx \int_\tau f dx.$$ We then define an approximation of the $L^2({\Omega})$-inner product in $X_h$ by $$\label{h-inner} (w, \chi)_h = \sum_{\tau \in \T_h} Q_{\tau,h}(w \chi).$$ Then the lumped mass Galerkin FEM is: find $ \luh (t)\in X_h$ such that $$\label{fem-lumped} {(\Dal \luh, \chi)_h}+ a(\luh,\chi)= (f, \chi) \quad \forall \chi\in X_h,\ t >0, \quad \luh(0)=0.$$ The lumped mass method leads to a diagonal mass matrix, and thus enhances the computational efficiency. We now introduce the discrete Laplace operator $-\bDelh:X_h\rightarrow X_h$, corresponding to the approximate $L^2(\Omega)$-inner product $(\cdot,\cdot)_h$, defined by $$\label{eqn:bDelh} -(\bDelh\psi,\chi)_h = (\nabla \psi,\nabla \chi)\quad \forall\psi,\chi\in X_h.$$ Also, we introduce a projection operator $\Pbarh: L^2({\Omega}) \rightarrow X_h$ by $$(\Pbarh f, \chi)_h = (f, \chi), \quad \forall \chi\in X_h.$$ Similarly as in Section \[sec:galerkin\], we introduce the discrete solution operator $\Ftilh$ by $$\label{eqn:Ftilh} \Ftilh f_h(t) = \sum_{j=1}^Nt^{{\alpha}-1}E_{{\alpha},{\alpha}}(-\bar{{\lambda}}_j^h t^{\alpha})(f_h,\bar{{\varphi}}_j^h)_h\bar{{\varphi}}_j^h,$$ where $\{\bar{{\lambda}}_j^h\}_{j=1}^N$ and $\{\bar{{\varphi}}_j^h\}_{j=1}^N$ are respectively the eigenvalues and the eigenfunctions of $-\bDelh$ with respect to $(\cdot,\cdot)_h$. Then with $f_h=\Pbarh f$, the solution $\bar{u}_h$ to problem can be represented by $$\label{eqn:Duhamel_oo} \bar{u}_h(t) = \int_0^t\Ftilh(t-s)f_h(s)ds.$$ We need the following modification of the discrete norm , still denoted by $\normh {\cdot}{p}$, on the space $X_h$ $$\label{eqn:normhbarp} \normh{\psi}{p}^2 = \sum_{j=1}^N (\bar{{\lambda}}_j^h)^p(\psi,\bar{{\varphi}}_j^h)_h^2\quad \forall p\in\mathbb{R}.$$ The following norm equivalence result and inverse estimate are useful for our analysis \[lem:normequivhbarp\] The norm $\normh {\cdot}{p}$ defined in is equivalent to the norm $\|\cdot\|_{\dH p}$ on the space $X_h$ for $-1\leq p \leq 1$. Further for all $\psi \in X_h$ we have for any real $l>s$ $$\label{inverse:hbarp} \normh{\psi}{l}\le Ch^{s-l} \normh{\psi}{s}. $$ The norm equivalence for $q=0,1$ is well known. The interpolation and duality arguments show $\normh{\cdot }{p}$ defined in is equivalent to $\|\cdot\|_{\dH p}$ on the space $X_h$ for $ -1\leq p \le 1$. \[lem:Ftilh\] Let $\Ftilh$ be defined by . Then we have for $\psi \in X_h$ and all $t>0$, $$\normh{\Ftilh(t) \psi}{p}\leq \left\{\begin{array}{ll} Ct^{-1+{\alpha}(1+(q-p)/2)}\normh{ \psi }{q}, & p-4\leq q\leq p,\\[1.3ex] Ct^{-1+{\alpha}}\normh{ \psi }{q},& p<q. \end{array}\right.$$ We also need the quadrature error operator $Q_h: X_h\rightarrow X_h$ defined by $$\label{eqn:Q} (\nabla Q_h\chi,\nabla \psi) = \epsilon_h(\chi,\psi) : = (\chi,\psi)_h-(\chi,\psi)\quad \forall \chi,\psi\in X_h.$$ The operator $Q_h$, introduced in [@chatzipa-l-thomee12], represents the quadrature error (due to mass lumping) in a special way. It satisfies the following error estimate [@chatzipa-l-thomee12 Lemma 2.4]. \[lem:Q\] Let $\bDelh$ and $Q_h$ be defined by and , respectively. Then $$\|\nabla Q_h\chi\|_{L^2(\Omega)}+h\|\bDelh Q_h\chi\|_{L^2(\Omega)} \leq Ch^{p+1}\|\nabla^p\chi\|_{L^2(\Omega)} \quad \forall \chi\in X_h, ~~~p=0,1.$$ The rest of this section is devoted to the error analysis of the lumped mass method for the case $f\in L^\infty(0,T;\dH q)$, $-1<q\leq1$. The error $\luh(t)-u(t)$ can be split as $\luh(t)-u(t)=u_h(t) -u(t) + \delta(t)$ with $\delta(t)=\luh(t)-u_h(t)$ and $u_h(t)$ being the standard Galerkin approximation, i.e., the solution of . Thus it suffices to establish proper error bounds for $\delta(t)$. By the definitions of $u_h(t)$, $\luh(t)$, and $Q_h$ the function $\delta(t)$ satisfies $$\Dal \delta(t) - \bDelh \delta(t) = \bDelh Q_h\Dal u_h(t) \quad \mbox{ for } \quad T \ge t>0 \quad \mbox{and} \quad \delta(0)=0.$$ By Duhamel’s principle , $\delta(t)$ can be expressed as $$\label{splitlump} \delta(t) = \int_0^t \Ftilh (t-s)\bDelh Q_h\Dal u_h(s)ds.$$ Like before, now we discuss the $L^2$- and $L^\infty$-norm in time error estimates separately. In the error analysis, the quadrature error operator $Q_h$ and the inverse estimate play essential role. Error estimates for solutions in $L^2(0,T;\dH p)$ ------------------------------------------------- In this part, we derive an $L^2(0,T;\dH p)$-error estimate, $p=0,1$, for the lumped mass method. The main result is stated in the following theorem. \[thm:lump:l2\] Let $f\in L^\infty(0,T;\dH q)$, $-1<q\leq 1$, and $u$ and $u_h$ be the solutions of and with $f_h=\bar P_hf$, respectively. Then there holds $$\begin{aligned} \|\nabla (\luh-u)\|_{L^2(0,T;L^2(\Omega))}&\leq Ch^{1+\min(q,0)}\|f\|_{L^2(0,T;\dH q)}, \\ \|\luh-u\|_{L^2(0,T;L^2(\Omega))}&\leq Ch^{1+q}\|f\|_{L^2(0,T;\dH q)}.\\ \end{aligned}$$ By repeating the proof of Lemma \[lem:reg-d\], we deduce from and Lemmas \[lem:normequivhbarp\] and \[lem:Q\] that $$\begin{split} \int_0^T \| \nabla \delta(t)\|^2_{L^2(\Omega)} dt & \le C\int_0^T \normh {\bDelh Q_h\Dal u_h(t)}{-1}^2 dt \le C\int_0^T \normh{ Q_h\Dal u_h(t) }{1}^2 dt\\ & \le C\int_0^T\| \nabla Q_h \Dal u_h(t)\|_{L^2(\Omega)}^2 dt \leq Ch^{2} \int_0^T \|\Dal u_h(t)\|_{L^2(\Omega)}^2dt. \end{split}$$ The desired assertion for the case $q \ge 0$ now follows immediately from Lemma \[lem:reg-d\]. For $-1 < q <0$ we use the inverse estimate of Lemma \[lem:inverse\], the stability of the Galerkin solution $u_h$ established in Lemma \[lem:reg-d\] and the stability of $P_h$ from Lemma \[lem:prh-bound\] to get $$\begin{aligned} \int_0^T \|\nabla \delta(t)\|^2_{L^2(\Omega)} dt & = Ch^{2+2q}\int_0^T\normh{\Dal u_h(t)}{q}^2 dt\\ & \leq Ch^{2+2q}\int_0^T\normh{f_h(t)}{q}^2 dt \leq Ch^{2+2q}\|f\|_{L^2(0,T;\dH q)}^2. \end{aligned}$$ Now we turn to the $L^2$-estimate. By repeating the preceding arguments, we arrive at $$\begin{split} \int_0^T \|\delta(t)\|^2_{L^2(\Omega)} dt & \le C\int_0^T \normh {\bDelh Q_h\Dal u_h(t)}{-2}^2 dt = C\int_0^T \tribar {Q_h\Dal u_h(t)} \tribar_{L^2(\Omega)}^2 dt\\ & \le C\int_0^T \|\nabla Q_h\Dal u_h(t) \|_{L^2(\Omega)}^2 dt \le Ch^{4}\int_0^T\|\Dal u_h(t)\|_{\dH 1}^2 dt \\ & \leq Ch^{2+2q} \int_0^T \normh{\Dal u_h(t)}{q}^2dt. \end{split}$$ where the second line follows from the trivial inequality $\|\chi\|_{L^2(\Omega)}\leq C\|\nabla \chi\|_{L^2(\Omega)}$ for $\chi\in X_h$ and the norm equivalence in Lemma \[lem:normequivhbarp\]. The rest of the proof is identical with that in the preceding part, and hence omitted. The estimate in $L^2(0,T;L^2(\Omega))$-norm of Theorem \[thm:lump:l2\] is suboptimal for any $q<1$. An optimal estimate can be obtained under an additional condition on the mesh. \[thm:lump:l2l2\] Let the assumptions in Theorem \[thm:lump:l2\] be fulfilled and the operator $Q_h$ satisfy $$\label{eqn:condQ} \|Q_h\chi\|_{L^2(\Omega)}\leq Ch^2\|\chi\|_{L^2(\Omega)}\quad \forall\chi \in X_h.$$ Then $$\|\luh-u\|_{L^2(0,T;L^2(\Omega))}\leq Ch^{2+\min(q,0)}\|f\|_{L^2(0,T;\dot H^q(\Omega))}.$$ It follows from the condition on the operator $Q_h$ that $$\begin{split} \int_0^T \|\delta(t)\|^2_{L^2(\Omega)} dt & \le C\int_0^T \normh{\bDelh Q_h\Dal u_h}{-2}^2 dt \le C\int_0^T \tribar Q_h\Dal u_h(t) \tribar^2_{L^2(\Omega)} dt\\ & \le Ch^4\int_0^T \|\Dal u_h(t) \|_{L^2(\Omega)}^2 dt \le Ch^{4+2\min(q,0)}\int_0^T\normh{ \Dal u_h(t)}{q}^2 dt. \end{split}$$ The rest of the proof is identical with that of Theorem \[thm:lump:l2\], and this completes the proof. The condition on the operator $Q_h$ is satisfied for symmetric meshes [@chatzipa-l-thomee12 section 5]. In one dimension, the symmetry requirement can be relaxed to almost symmetry [@chatzipa-l-thomee12 section 6]. In case does not hold, we were able to show only a suboptimal $O(h^{1+q})$-convergence rate for the $L^2$-norm of the error, which is reminiscent of that for the initial value problem for the classical parabolic equation [@chatzipa-l-thomee12 Theorem 4.4] and time-fractional diffusion problem [@Bangti_LZ_2013 Theorem 4.5]. Error estimates for solutions in $L^\infty(0,T;\dH p)$ ------------------------------------------------------ Now we derive an estimate in $L^\infty(0,T;\dH p)$-norm for the lumped mass approximation $\luh$. \[thm:lump:linf\] Let $f\in L^\infty(0,T;\dH q)$, $-1<q\leq 1$, and $u$ and $\luh$ be the solutions of and , respectively, with $\bar f_h=\bar P_hf$. Then with $\ell_h=|\ln h|$, the following estimates are valid for $t >0$: $$\label{H1-lumpled} \begin{aligned} \|\nabla(\luh(t)-u(t))\|_{L^2(\Omega)}&\leq Ch^{1+q} \ell_h^2\|f\|_{L^\infty(0,t;\dH q)} \quad \mbox{for} \quad -1< q \le 0. \end{aligned}$$ Moreover, for $-1< q \le 1$ we have $$\label{L2-lumpled} \begin{aligned} \|\luh(t)-u(t)\|_{L^2(\Omega)} &\leq Ch^{1+q}\ell_h^2 \|f\|_{L^\infty(0,t; \dH q)}. \end{aligned}$$ We first note that by the smoothing property of $\Ftilh$ in Lemma \[lem:Ftilh\] and the inverse inequality , we have for $\chi\in X_h $, $\epsilon>0$, and $p=0,1$ $$\label{eqn:lumpbasic1} \begin{aligned} \normh{ \Ftilh(t)\bDelh Q_h\chi}{p} &\leq Ct^{\epsilon{\alpha}/2-1} \normh{ \bDelh Q_h\chi}{p-2+\epsilon} = Ct^{\epsilon{\alpha}/2-1} \normh{ Q_h\chi}{p+\epsilon}\\ & \leq Ct^{\epsilon{\alpha}/2-1}h^{-\epsilon}\normh{ Q_h \chi}{p}. \end{aligned}$$ We first prove estimate . Setting $\chi=\Dal u_h(t)$ in and Lemma \[lem:Q\] yield $$\label{eqn:lumpbasicI1} \begin{aligned} \normh{ \Ftilh(t-s)\bDelh Q_h\Dal u_h(s)}{1} &\leq C(t-s)^{\epsilon{\alpha}/2-1}h^{-\epsilon} \|Q_h\Dal u_h(s)\|_{\dH 1}\\ & \le Ch^{1-{\epsilon}}(t-s)^{{\epsilon}{\alpha}/2-1}\|\Dal u_h(s)\|_{L^2(\Omega)}. \end{aligned}$$ Then it follows from relation of Galerkin approximation $u_h$ and the triangle and inverse inequalities that $$\begin{aligned} \|\nabla\delta(t)\|_{L^2(\Omega)} &\leq Ch^{1-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}\|\Dal u_h(s)\|_{L^2(\Omega)}ds \\ & \leq Ch^{1-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}(\|\Delta_h u_h(s)\|_{L^2(\Omega)}+\|f_h(s)\|_{L^2(\Omega)})ds \\ & \leq Ch^{1-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}(h^{-{\epsilon}}\normh{\Delta_h u_h(s)}{-{\epsilon}}+\|f_h(s)\|_{L^2(\Omega)})ds. \end{aligned}$$ Further, using the discrete stability estimate in Lemma \[lem:reg-d\] and the estimate of $P_h$ in Lemma \[lem:prh-bound\] we further get for $-1 <q \le 0$ $$\begin{aligned} \|\nabla\delta(t)\|_{L^2(\Omega)} & \leq Ch^{1+q-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}(h^{-{\epsilon}}\normh{\Delta_h u_h(s)}{q-{\epsilon}}+\normh{ f_h(s)}{q})ds \\ & \leq Ch^{1+q-2\epsilon} \int_0^t(t-s)^{{\epsilon}{\alpha}/2-1}({\epsilon}^{-1}s^{{\epsilon}{\alpha}/2}\tribar f_h\tribar_{L^\infty(0,s;\dH q)} +\normh{f_h(s)}{q})ds\\ & \leq C\epsilon^{-1}h^{1+q -2\epsilon} \tribar f_h\tribar_{L^\infty(0,t;\dH q)}\int_0^t(t-s)^{{\epsilon}{\alpha}/2-1}s^{{\epsilon}{\alpha}/2} ds \\ & \leq C\epsilon^{-2}h^{1+q -2\epsilon} \tribar f_h\tribar_{L^\infty(0,t;\dH q)} \leq Ch^{1+q} \ell_h^2 \|f\|_{L^\infty(0,t;\dH q)}, \end{aligned}$$ where in the last inequality we have chosen $\epsilon=1/\ell_h$. Now follows from this and Theorem \[thm:gal:linf\]. Next we derive the $L^2$-error estimate. Similar to the derivation of , we get $$\begin{aligned} \| \Ftilh(t-s)\bDelh Q_h\Dal u_h(s)\|_{L^2(\Omega)} &\leq C (t-s)^{{\alpha}/2-1} \|\nabla Q_h\Dal u_h(s)\|_{L^2(\Omega)}\\ &\le Ch^{2}(t-s)^{{\alpha}/2-1}\| \Dal u_h(s) \|_{\dH 1}. \end{aligned}$$ Consequently, by the triangle inequality, inverse estimate in Lemma \[lem:inverse\] and the discrete stability estimate in Lemma \[lem:reg-d\], there holds $$\begin{aligned} \|\delta(t)\|_{L^2(\Omega)} &\leq Ch^{2}\int_0^t(t-s)^{{\alpha}/2-1}(\|\Delta u_h(s)\|_{\dH 1}+\|P_hf(s)\|_{\dH 1}) ds \\ &\leq Ch^{2}\int_0^t(t-s)^{{\alpha}/2-1}(h^{-{\epsilon}}\normh{\Delta u_h(s)}{1-{\epsilon}}+\normh{ f_h(s)}{1}) ds \\ &\leq Ch^{1+q}\int_0^t(t-s)^{{\alpha}/2-1}(h^{-{\epsilon}}\normh{\Delta u_h(s)}{q-{\epsilon}}+\normh{ f_h(s)}{q}) d\tau \\ &\leq Ch^{1+q}\int_0^t(t-s)^{{\alpha}/2-1}({\epsilon}^{-1}h^{-{\epsilon}}s^{{\epsilon}{\alpha}/2}\tribar f_h\tribar_{L^\infty(0,s;\dH q)}+\normh{f_h(s)}{q})ds. \end{aligned}$$ The $L^2$-estimate follows directly from Theorem \[thm:gal:linf\] and setting $\epsilon=1/\ell_h$ in the above inequality. For $q>0$, we have $\|\nabla(\luh(t)-u(t))\|_{L^2(\Omega)}\leq Ch\ell_h^2$ and the error cannot be improved even if the function $f$ is smoother. In view of Remark \[rmk:gallogh\], for $0<q\leq1$, the following slightly improved estimate holds $$\|\luh(t)-u(t)\|_{L^2(\Omega)}\leq Ch^{1+q}\ell_h \|f\|_{L^\infty(0,t; \dH q)}.$$ In the case of $f\in L^\infty(0,T;\dH q)$, $0<q\leq1$, we can obtain an improved estimate of $\|\nabla \delta(t)\|_{L^2(\Omega)}$: $$\begin{aligned} \|\nabla\delta(t)\|_{L^2(\Omega)} &\leq Ch^{2-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}\|\Dal u_h(s)\|_{\dH 1}ds \\ & \leq Ch^{1+q-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}(h^{-{\epsilon}}\normh{\Delta_h u_h(s)}{q-{\epsilon}}+\tribar f_h(s)\tribar_{\dH q})ds \\ & \leq Ch^{1+q-2\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}({\epsilon}^{-1}s^{{\epsilon}{\alpha}/2}\tribar f_h\tribar_{L^\infty(0,s;\dH q)}+\tribar f_h(s)\tribar_{\dH q})ds\\ & \leq C\epsilon^{-2}h^{1+q-2\epsilon} \|f\|_{L^\infty(0,t;\dH q)} \leq Ch^{1+q} \ell_h^2 \|f\|_{L^\infty(0,t;\dH q)}. \end{aligned}$$ We record this observation in a remark. In the case of a right hand side of intermediate smoothness, i.e., $f\in L^\infty(0,T;\dH q)$, $0< q\leq1$, the gradient estimate $\|\nabla\delta(t)\|_{L^2(\Omega)}$ can be improved to $(1+q)$th-order at the expense of an extra factor $\ell_h$: $$\|\nabla\delta(t)\|_{L^2(\Omega)} \leq Ch^{1+q} \ell_h^2 \|f\|_{L^\infty(0,t;\dH q)}.$$ Like in the case of $L^2(0,T;L^2(\Omega))$-estimate, the $L^\infty(0,T;L^2(\Omega))$ estimate is suboptimal for any $q\in(-1,1)$, and can be improved to an almost optimal one by imposing condition . \[thm:lump:linfl2\] Let the conditions in Theorem \[thm:lump:linf\] be fulfilled and the operator $Q_h$ satisfy , i.e., $\|Q_h\chi\|_{L^2(\Omega)}\leq Ch^2\|\chi\|_{L^2(\Omega)}$ for all $\chi \in X_h$. Then there holds (with $\ell_h=|\ln h|$) $$\|\luh(t)-u(t)\|_{L^2(\Omega)}\leq Ch^{2+\min(q,0)} \ell_h^2 \|f\|_{L^\infty(0,t;\dH q)}, \quad -1 < q \le 1.$$ If holds, then applying with $p=0$ we get $$\begin{aligned} \|\Ftilh(t-s)\bDelh Q_h\Dal u_h(s)\|_{L^2(\Omega)} &\leq C(t-s)^{{\epsilon}{\alpha}/2-1}h^{-\epsilon} \|Q_h\Dal u_h(s)\|_{L^2(\Omega)}\\ & \le Ch^{2-{\epsilon}}(t-s)^{{\epsilon}{\alpha}/2-1}\|\Dal u_h(s)\|_{L^2(\Omega)}. \end{aligned}$$ Consequently, this together with the inverse estimate from Lemma \[lem:inverse\], the discrete stability estimate in Lemma \[lem:reg-d\], and the stability of $P_h$ in Lemma \[lem:prh-bound\], yields $$\begin{aligned} \|\delta(t)\|_{L^2(\Omega)} & \leq Ch^{2-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}(\|\Delta_h u_h(s)\|_{L^2(\Omega)}+\|f_h(s)\|_{L^2(\Omega)})ds\\ & \leq Ch^{2+\min(q,0)-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}(h^{-{\epsilon}}\normh{\Delta_h u_h(s)}{-{\epsilon}+q}+\normh{ f_h(s)}{q})ds \\ & \leq Ch^{2+\min(q,0)-\epsilon} \int_0^t (t-s)^{{\epsilon}{\alpha}/2-1}(h^{-{\epsilon}}{\epsilon}^{-1}s^{{\epsilon}{\alpha}/2}\tribar f_h\tribar_{L^\infty(0,s;\dH q)}+\normh {f_h(s)}{q})ds\\ & \leq C\epsilon^{-2}h^{2+\min(q,0)-2\epsilon} \|f\|_{L^\infty(0,t;\dH q)} \leq Ch^{2+\min(q,0)} \ell_h^2 \|f\|_{L^\infty(0,t;\dH q)}, \end{aligned}$$ where the last line follows from the choice $\epsilon = 1/\ell_h$. Now the desired assertion follows immediately from this and Theorem \[thm:gal:linf\]. Numerical results {#sec:numerics} ================= In this section, we present numerical results to verify the theoretical error estimates in Sections \[sec:galerkin\] and \[sec:lump\]. We present the errors $\|u-\luh\|_{L^\infty(0,T;L^2(\Omega))}$ and $\|\nabla(u-\luh)\|_{L^\infty(0,T; L^2(\Omega))}$ for the lumped mass method only, since the errors for the Galerkin FEM are almost identical. In the tables and figures below, we use the following notation convention: $\|u(t)-\luh(t) \|_{\dH p}$ for $p=0$ and $p=1$ is simply referred to as $L^2$-norm and $H^1$-norm error, respectively. 1-d examples {#ssect:1D} ------------ First, we consider for $d=1$ on the unit interval $\Omega=(0,1)$ and perform numerical tests on the following three data sets: 1. Nonsmooth data: $f(x,t)=(\chi_{[1/2,1]}(t)+1)\chi_{[0,{1/2}]}(x)$, where $\chi_S$ is the characteristic function of the set $S$. The jump at $x=1/2$ leads to $f(t,\cdot) \notin \dot H^1(\Omega)$; nonetheless, for any $\epsilon >0$, $f(x,t) \in L^\infty(0,T;H^{{1/2}-\epsilon}(\Omega))$. 2. Very weak data: $f(x,t)=(\chi_{[1/2,1]}(t)+1)\delta_{1/2}(x)$ where $f(x,t)$ involves a Dirac $\delta_{1/2}(x)$-function concentrated at $x=0.5$. 3. Variable coefficients: in we take $k(x)= 3 +\sin(2\pi x)$, $f(x,t)=(\chi_{[1/2,1]}(t)+1)\chi_{[0,1/2]}(x)$, and $q(x)=0$. The exact solution for examples (1a) and (1b) can be explicitly expressed by an infinite series involving the Mittag-Leffler function $E_{\alpha,\alpha}(z)$ as . To accurately evaluate the Mittag-Leffler functions, we employ the algorithm developed in [@Seybold:2008], which is based on three different approximations of the function, i.e., Taylor series, integral representations and exponential asymptotics, in different regions of the domain. In our computation, we divide the unit interval $(0,1)$ into $N+1$ equally spaced subintervals, with a mesh size $h=1/(N+1)$. The finite element space $X_h$ consists of continuous piecewise linear functions. The eigenpairs $({\lambda}^h_j, {\varphi}^h_j(x) )$ and $( \bar{\lambda}^h_j, \bar {\varphi}^h_j(x) )$ of the one-dimensional discrete Laplacian $-\Delta_h$ and $-\bar \Delta_h$, defined by and , respectively, satisfy $$(-\Delta_h {\varphi}^h_j,v)={\lambda}^h_j ({\varphi}^h_j,v) \quad\mbox{ and }\quad (-\bar\Delta_h \bar{\varphi}^h_j,v)_h=\bar{\lambda}^h_j (\bar{\varphi}^h_j,v)_h \quad \forall v \in X_h.$$ Here $(w,v)$ and $(w,v)_h$ refer to the standard $L^2$-inner product and the approximate $L^2$-inner product (cf. eq. ) on the space $X_h$, respectively. Then for $ j=1,\dots,N$, there hold $${\lambda}^h_j= \bar{{\lambda}}_j^h /(1 - \tfrac{h^2}{6} \bar{{\lambda}}_j^h), ~~ \bar{{\lambda}}_j^h=\frac{4}{h^2}\sin^2\frac{\pi j}{2(N+1)}, ~~\mbox{and} ~~ {\varphi}^h_j(x_k)=\bar{\varphi}^h_j(x_k) = \sqrt2 \sin(j\pi x_k)$$ for $x_k$ being a mesh point and $ {\varphi}^h_j$ and $ \bar{\varphi}^h_j$ linear over the finite elements. Then the solutions of the standard Galerkin method and lumped mass method can be computed by and , respectively. In the case of (1c), there is no convenient solution representation. To compute the semidiscrete solution, we have used a direct numerical technique by first discretizing the time interval, $t_n=n\tau$, $n=0,1,\dots$, with $\tau$ being the time step size, and then using a weighted finite difference approximation of the fractional derivative $\Dal u(x,t_n)$ developed in [@LinXu:2007]: $$\label{eqn:timedisc} \begin{aligned} \Dal u(x,t_n) &= \frac{1}{\Gamma(1-{\alpha})}\sum^{n-1}_{j=0}\int^{t_{j+1}}_{t_j} \frac{\partial u(x,s)}{\partial s} (t_n-s)^{-{\alpha}}\, ds \approx \frac{1}{\Gamma(2-\alpha)}\sum_{j=0}^{n-1}b_j\frac{u(x,t_{n-j})-u(x,t_{n-j-1})}{\tau^\alpha}, \end{aligned}$$ where the weights $\{b_j\}$ are given by $b_j=(j+1)^{1-\alpha}-j^{1-\alpha}$, $j=0,1,\ldots,n-1$. If the solution $u(x,t)$ is sufficiently smooth, then the local truncation error of the finite difference approximation is bounded by $C\tau^{2-{\alpha}}$ for some $C$ depending only on $u$ [@LinXu:2007 equation (3.3)]. Hence with a small $\tau$, the fully discrete solution can approximate the semidiscrete solution well. ### Numerical results for example (1a) In Tables \[tab:nonsmooth1DL2\] and \[tab:nonsmooth1DLinf\] we show the errors $\|u-\luh\|_{L^2(0,T;\dH p)}$ and $\|u-\luh\|_{L^\infty(0,T;\dH p)}$ for $p=0,1$. The convergence rates are almost identical for the $L^2$- and $L^\infty$- in time estimates since both the $L^2(\Omega)$-norm of the error $e=u-\luh$ and its gradient are bounded independent of the time. Thus in what follows, we only present the errors $\|u(t)-\luh(t)\|_{L^2(\Omega)}$ and $\|\nabla(u(t)-\luh(t))\|_{L^2(\Omega)}$. In Figure \[fig:nonsmooth\_Linf\], we show the plot of the results from Table \[tab:nonsmooth1DLinf\] in a log-log scale. The errors are almost independent of the fractional order $\alpha$, and hence the three curves nearly coincide with each other. The numerical results fully confirm our theoretical predictions, i.e., $O(h^2)$ and $O(h)$ convergence rates for the $L^2(\Omega)$- and $H^1(\Omega)$-norms of the error, respectively. ${\alpha}$ $k$ $3$ $4$ $5$ $6$ $7$ rate ----------------- ------------ --------- --------- --------- --------- --------- --------------- ${\alpha}=0.1$ $L^2$-norm 9.96e-4 2.49e-4 6.23e-5 1.55e-5 3.89e-6 $O(h^{2.00})$ $H^1$-norm 2.45e-2 1.23e-2 6.13e-3 3.07e-3 1.53e-3 $O(h^{1.00})$ ${\alpha}=0.5$ $L^2$-norm 9.97e-4 2.50e-4 6.24e-5 1.56e-5 3.90e-6 $O(h^{2.00})$ $H^1$-norm 2.46e-2 1.23e-2 6.14e-3 3.08e-3 1.54e-3 $O(h^{1.00})$ ${\alpha}=0.95$ $L^2$-norm 1.01e-3 2.51e-4 6.28e-5 1.57e-5 3.93e-6 $O(h^{2.00})$ $H^1$-norm 2.48e-2 1.24e-2 6.20e-3 3.10e-3 1.55e-3 $O(h^{1.00})$ : Numerical results, i.e., errors $\|u-\luh\|_{L^2(0,1;L^2(\Omega))}$ ($L^2$-norm) and $\|u-\luh\|_{L^2(0,1;\dot H^1(\Omega))}$ ($H^1$-norm), for example (1a) (nonsmooth data) with mesh sizes $h=1/2^k$.[]{data-label="tab:nonsmooth1DL2"} ${\alpha}$ $k$ $3$ $4$ $5$ $6$ $7$ rate ----------------- ------------ --------- --------- --------- --------- --------- --------------- ${\alpha}=0.1$ $L^2$-norm 9.98e-4 2.50e-4 6.24e-5 1.56e-5 3.90e-6 $O(h^{2.00})$ $H^1$-norm 2.47e-2 1.23e-2 6.17e-3 3.09e-3 1.54e-3 $O(h^{1.00})$ ${\alpha}=0.5$ $L^2$-norm 1.01e-3 2.51e-4 6.29e-5 1.57e-5 3.93e-6 $O(h^{2.00})$ $H^1$-norm 2.53e-2 1.27e-2 6.33e-3 3.17e-3 1.58e-3 $O(h^{1.00})$ ${\alpha}=0.95$ $L^2$-norm 1.04e-3 2.59e-4 6.49e-5 1.62e-5 4.06e-6 $O(h^{2.00})$ $H^1$-norm 2.58e-2 1.29e-2 6.46e-3 3.23e-3 1.62e-3 $O(h^{1.00})$ : Numerical results, i.e., errors $\| u(t)-\luh(t) \|_{\dH p}$, $p=0,1$ at $t=1$, for example (1a) (nonsmooth data) with mesh sizes $h=1/2^k$.[]{data-label="tab:nonsmooth1DLinf"} ### Numerical results for example (1b) In this example, we consider the case of very weak data, which involves the Dirac $\delta$-function. Since the Dirac $\delta$-function $\delta_{1/2}(x)$ is a bounded linear functional on $C_0(0,1)$. and $H_0^{1/2+\epsilon}(0,1)$ with $\epsilon>0$ embeds continuously into $C_0(0,1)$, we have $\delta_{1/2}(x) \in H^{-1/2-\epsilon}(\Omega)$. Thus, the source term $f(x,t)=(\chi_{[1/2,1]}(t)+1)\delta_{1/2}(x) \in L^{\infty}(0,T;H^{-1/2-\epsilon}(\Omega))$. In Table \[tab:weak1D\] we show the error and convergence rates for the lumped mass method for three different ${\alpha}$ values. Here the mesh size is chosen to be $h=1/(2^k+1)$, and thus the Dirac $\delta$ function is not aligned with the grid. The results indicate an $O(h^{1/2})$ and $O(h^{3/2})$ convergence rate for the $H^1(\Omega)$- and $L^2(\Omega)$-norm of the error, respectively, which agrees well with the theoretical prediction. In Table \[tab:weak1Dsuper\], we report the results for the case that the $\delta$-function is supported at a grid point. We observe that the method converges at the expected rate in $H^1(\Omega)$-norm, while the convergence rate in the $L^2(\Omega)$-norm is $O(h^2)$, i.e., the method exhibits a superconvergence of one half order, which theoretically remains to be established. The plots of the results for $\alpha=0.5$ in Tables \[tab:weak1D\] and \[tab:weak1Dsuper\] are shown respectively in Figures \[fig:weak1D\] and \[fig:weak1Dsuper\] in a log-log scale. ${\alpha}$ $k$ $3$ $4$ $5$ $6$ $7$ rate ----------------- ------------ --------- --------- --------- --------- --------- --------------- ${\alpha}=0.1$ $L^2$-norm 5.35e-3 2.07e-3 7.67e-4 2.78e-4 9.94e-5 $O(h^{1.44})$ $H^1$-norm 1.56e-1 1.14e-1 8.23e-2 5.88e-2 4.17e-2 $O(h^{0.48})$ ${\alpha}=0.5$ $L^2$-norm 5.38e-3 2.08e-3 7.68e-4 2.78e-4 9.95e-5 $O(h^{1.44})$ $H^1$-norm 1.57e-1 1.15e-1 8.25e-2 5.89e-2 4.17e-2 $O(h^{0.48})$ ${\alpha}=0.95$ $L^2$-norm 5.39e-3 2.08e-3 7.69e-4 2.79e-4 9.95e-5 $O(h^{1.44})$ $H^1$-norm 1.58e-1 1.15e-1 8.26e-2 5.89e-2 4.18e-2 $O(h^{0.48})$ : Numerical results, i.e., errors $\|u(t)-\luh(t) \|_{\dH p}$, $p=0,1$ at $t=1$, for example (1b) (Dirac $\delta$-function) with mesh sizes $h=1/(2^k+1)$.[]{data-label="tab:weak1D"} ${\alpha}$ $k$ $3$ $4$ $5$ $6$ $7$ rate ----------------- ------------ --------- --------- --------- --------- --------- --------------- ${\alpha}=0.1$ $L^2$-norm 1.59e-4 3.93e-5 9.80e-6 2.45e-6 6.15e-7 $O(h^{2.00})$ $H^1$-norm 5.46e-2 3.91e-2 2.78e-2 1.97e-3 1.39e-3 $O(h^{0.49})$ ${\alpha}=0.5$ $L^2$-norm 7.12e-5 1.76e-5 4.37e-6 1.10e-6 2.76e-7 $O(h^{2.00})$ $H^1$-norm 5.53e-2 3.93e-2 2.79e-2 1.97e-3 1.40e-3 $O(h^{0.50})$ ${\alpha}=0.95$ $L^2$-norm 4.91e-5 1.21e-5 3.03e-6 7.66e-7 1.96e-7 $O(h^{2.02})$ $H^1$-norm 5.60e-2 3.96e-2 2.80e-2 1.98e-3 1.40e-3 $O(h^{0.50})$ : Numerical results, errors $\|u(t)-\luh(t) \|_{\dH p}$, $p=0,1$ at $t=1$, for example (1b) (Dirac $\delta$ function) with mesh sizes $h=1/2^k$.[]{data-label="tab:weak1Dsuper"} ### Numerical results for example (1c) Here the coefficient is variable, and the discrete solution $\luh$ does not have a convenient explicit representation. Hence, we apply the fully discrete scheme with finite difference discretization on time by . The benchmark solution $u$ is computed on a very fine mesh, i.e., with a spacial mesh size $h=1/512$ and a time step size $\tau=1.0 \times 10^{-5}$. The $L^2(\Omega)$- and $H^1(\Omega)$-norms of the error are reported in Table \[tab:variable\] at $t=1$ for ${\alpha}=0.5$. The results confirm the discussions in Section \[general\_problems\]. $h$ $1/8$ $1/16$ $1/32$ $1/64$ $1/128$ rate ------------- --------- --------- --------- --------- --------- --------------- $L^2$-error 8.42e-4 2.14e-4 5.38e-5 1.34e-5 3.30e-6 $O(h^{2.00})$ $H^1$-error 1.76e-2 8.89e-3 4.45e-3 2.21e-3 1.08e-3 $O(h^{1.01})$ : Numerical results, i.e., errors $\|u(t)-\luh(t)\|_{\dH p}$, $p=0,1$, at $t=1$, for example (1c) (variable coefficients) with ${\alpha}=0.5$ and mesh sizes $1/2^k$.[]{data-label="tab:variable"} 2-d Examples {#ssect:2D} ------------ Now we consider for $d=2$ on the unit square ${\Omega}=[0,1]^2$ for the following data: 1. Nonsmooth data: $ f(x,t) = (\chi_{[1/2,1]}(t)+1)\chi_{[1/4,3/4]\times[1/4,3/4]}$. 2. Very weak data: $f(x,t)=(\chi_{[1/2,1]}(t)+1)\delta_\Gamma $ with $\Gamma$ being the boundary of the square $[1/4,3/4]\times[1/4,3/4]$ with $\langle \delta_\Gamma,\phi\rangle = \int_\Gamma \phi(s) ds$. One may view $(v,\chi)$ for $\chi \in X_h \subset\dot H^{1/2+\epsilon}({\Omega})$ as duality pairing between the spaces $H^{-1/2-\epsilon}({\Omega})$ and $\dot H^{1/2+\epsilon}({\Omega})$ for any $\epsilon >0$ so that $\delta_\Gamma \in H^{-1/2-\epsilon}(\Omega)$. Indeed, it follows from Hölder’s inequality and the continuity of the trace operator from $\dot H^{{1/2}+\epsilon}(\Omega)$ to $L^2(\Gamma)$ [@AdamsFournier:2003] that $$\begin{aligned} \|\delta_\Gamma\|_{H^{-1/2-\epsilon}(\Omega)}&= \sup_{\phi \in\dot H^{1/2+\epsilon}(\Omega)} \frac{|\int_\Gamma \phi(s)ds|}{\|\phi\|_{\dot H^{1/2+\epsilon}(\Omega)}} \le |\Gamma|^{1/2} \sup_{\phi \in\dot H^{1/2+\epsilon} (\Omega)} \frac{\|\phi\|_{L^2(\Gamma)}}{\|\phi\|_{\dot H^{1/2+\epsilon}(\Omega)}} \le C. \end{aligned}$$ To discretize the problem, we divide $(0,1)$ into $N=2^k$ equally spaced subintervals with a mesh size $h=1/N$ so that unit square $(0,1)^2$ is divided into $N^2$ small squares. We get a symmetric triangulation of the domain $(0,1)^2$ by connecting the diagonal of each small square. Therefore, the lumped mass method and standard Galerkin method have the same convergence rates. On these meshes, $\bar {\lambda}^h_{n,m}$ and $\bar {\varphi}^h_{n,m}$, $1 \le n,m \le N-1$, i.e., eigenvalues and eigenvectors of the discrete Laplacian $\bar \Delta_h$, are explicitly given by: $$\bar{\lambda}^h_{n,m}= \frac{4}{h^2} \left(\sin^2\frac{n\pi h}{2} + \sin^2\frac{m\pi h}{2} \right), \quad \bar{\varphi}^h_{n,m}(x_i,y_j)= 2\sin(n\pi x_i)\sin(m\pi y_j).$$ respectively, where $(x_i,y_j)$, $i,j=1,\ldots,N-1$, is a mesh point. Then the semidiscrete approximation can be computed via the explicit representation . ### Numerical results for example (2a) In this example the right hand side data $f(x,t)$ is in the space $L^{\infty}(0,1;\dot H^{1/2-{\epsilon}}({\Omega}))$ with ${\epsilon}> 0$, and the numerical results were computed at $t=1$ for ${\alpha}=0.1,0.5$ and $0.95$; see Table \[tab:nonsmooth2D\] and Figure \[fig:nonsmooth2D\]. The slopes of the error curves in a log-log scale are $2$ and $1$, respectively, for $L^2(\Omega)$- and $H^1(\Omega)$-norm of the errors, which agrees well with the theoretical results for the nonsmooth case. ${\alpha}$ $k$ $3$ $4$ $5$ $6$ $7$ rate ----------------- ------------ --------- --------- --------- --------- --------- --------------- ${\alpha}=0.1$ $L^2$-norm 9.66e-4 2.48-4 6.26e-5 1.57e-5 3.93e-6 $O(h^{1.99})$ $H^1$-norm 2.06e-2 1.04e-2 5.24e-3 2.63e-3 1.31e-3 $O(h^{0.99})$ ${\alpha}=0.5$ $L^2$-norm 9.82e-4 2.52-4 6.36e-5 1.59e-5 3.99e-6 $O(h^{1.99})$ $H^1$-norm 2.10e-2 1.07e-2 5.35e-3 2.68e-3 1.34e-3 $O(h^{0.99})$ ${\alpha}=0.95$ $L^2$-norm 9.82e-4 2.52-4 6.36e-5 1.61e-5 4.02e-6 $O(h^{1.99})$ $H^1$-norm 2.13e-2 1.08e-2 5.42e-3 2.71e-3 1.36e-3 $O(h^{0.99})$ : Numerical results, i.e., errors $\|u(t)-\luh(t) \|_{\dH p}$, $p=0,1$ at $t=1$, for example (2a) (nonsmooth data) with mesh sizes $h=1/2^k$ in either direction.[]{data-label="tab:nonsmooth2D"} ### Numerical results for example (2b) In Table \[tab:weak2D\], we present the $L^2(\Omega)$- and $H^1(\Omega)$-norms of the error for this example. The $H^1(\Omega)$-norm of the error decays at the theoretical rate, however the $L^2(\Omega)$-norm of the error exhibits superconvergence. This is attributed to the fact that the boundary $\Gamma$ is fully aligned with element edges. In contrast, if we choose $P_h f$ as the discrete right hand side for the lumped mass semidiscrete problem instead of $\bar P_h f$, then the $L^2(\Omega)$-norm of the error converges only at the standard order; see Table \[tab:weak2D2\]. ${\alpha}$ $k$ $3$ $4$ $5$ $6$ $7$ rate ----------------- ------------ --------- --------- --------- --------- --------- --------------- ${\alpha}=0.1$ $L^2$-norm 4.61e-3 1.25e-3 3.31e-4 8.68e-5 2.39e-5 $O(h^{1.90})$ $H^1$-norm 1.60e-1 9.92e-2 6.43e-2 4.43e-2 3.16e-2 $O(h^{0.58})$ ${\alpha}=0.5$ $L^2$-norm 4.67e-3 1.26e-3 3.34e-4 8.76e-5 2.40e-5 $O(h^{1.91})$ $H^1$-norm 1.60e-1 9.92e-2 6.44e-2 4.50e-2 3.17e-2 $O(h^{0.58})$ ${\alpha}=0.95$ $L^2$-norm 4.70e-3 1.27e-3 3.36e-4 8.81e-5 2.42e-5 $O(h^{1.91})$ $H^1$-norm 1.61e-1 9.98e-2 6.46e-2 4.50e-2 3.17e-2 $O(h^{0.58})$ : Numerical results, i.e., errors $\|u(t)-\luh(t)\|_{\dH p}$, $p=0,1$ at $t=1$, for example (2b) (Dirac $\delta$-function) with mesh sizes $h=1/2^k$ in either direction.[]{data-label="tab:weak2D"} ${\alpha}$ $k$ $3$ $4$ $5$ $6$ $7$ rate ---------------- ------------ --------- --------- --------- --------- --------- --------------- ${\alpha}=0.5$ $L^2$-norm 1.19e-2 4.55e-3 1.69e-3 6.15e-4 2.22e-4 $O(h^{1.44})$ $H^1$-norm 3.28e-1 2.32e-1 1.67e-1 1.13e-1 8.21e-2 $O(h^{0.50})$ : Numerical results, i.e., errors $\|u(t)-\luh(t)\|_{\dH p}$, $p=0,1$, at $t=1$, for example (2b) (Dirac $\delta$-function) with $f_h=P_h f$ and mesh sizes $h=1/2^k$ in either direction.[]{data-label="tab:weak2D2"} conclusion ========== In this paper, we have studied two semidiscrete finite element schemes, i.e., the semidiscrete Galerkin method and the lumped mass method, for the time-fractional diffusion problem with a nonsmooth right hand side $f\in L^\infty(0,T;\dH q)$, with $-1<q\leq 1$. We derived almost optimal estimates of the error and its gradient for the Galerkin method, and also almost optimal estimates of the gradient of the error for the lumped mass method. Optimal estimates for the $L^2(\Omega)$-norm of the error can only be shown under certain restrictions on the mesh such that condition is fulfilled. There are several possible extensions of the work. First, the error estimates are expected to be useful for analyzing relevant inverse problems, which will be studied in a future work. Second, it is natural to study the fully discrete scheme, e.g., with finite difference or discontinuous Galerkin method in time. Acknowledgments {#acknowledgments .unnumbered} =============== The research of B. Jin has been support by NSF Grant DMS-1319052, R. Lazarov and Z. Zhou was supported in parts by US NSF Grant DMS-1016525 and J. Pasciak has been supported by NSF Grant DMS-1216551. The work of all authors has been supported in parts also by Award No. KUS-C1-016-04, made by King Abdullah University of Science and Technology (KAUST).

ArXiv

--- address: - 'LTCI, Télécom Paristech, Université Paris Saclay, Paris, France' - 'Smart Impulse, Paris, France' author: - Simon Henriet - Umut Şimşekli - Benoit Fuentes - Gaël Richard bibliography: - 'nilm.bib' title: 'A Generative Model for Non-Intrusive Load Monitoring in Commercial Buildings' --- NILM, Commercial buildings, Synthetic data generation, Source separation, Matrix factorization, Statistical analysis References {#references .unnumbered} ==========

ArXiv

--- abstract: 'The charged lepton masses obey to high precision the so-called Koide relation. We propose a generalization of this relation to quarks. It includes up and down quarks of the three generations and is numerically reasonably close to the Koide limit.' author: - 'A. Kartavtsev$^{a}$' title: A remark on the Koide relation for quarks --- The pole masses of three charged leptons have been measured to an unprecedented precision [@PDG]: \[LeptonMassesLowScale\] $$\begin{aligned} m_e&=0.510998910 \pm 0.000000013\,\, {\rm MeV}\,,\\ m_\mu&=105.658367 \pm 0.000004\,\, {\rm MeV}\,,\\ m_\tau&=1776.82^{+0.16}_{-0.26}\,\, {\rm MeV}\,.\end{aligned}$$ To a surprising degree of accuracy they obey the empirical Koide mass relation [@Koide:1982si; @koide_new_1983]: $$\begin{aligned} \label{KoideLeptons} K_\ell\equiv \frac{m_e+m_ \mu+m_\tau}{(\sqrt{m_e}+\sqrt{m_\mu}+\sqrt{m_\tau})^2}=\frac23\,.\end{aligned}$$ Defining a deviation from the Koide limit, $\Delta K_\ell\equiv (3 K_\ell/2-1)$, and substituting the experimental charged lepton masses into we find $-2\cdot 10^{-5} < \Delta K_\ell < 9\cdot 10^{-7}$. It is also astonishing that the electron mass, being orders of magnitude smaller than the muon and tau-lepton masses, still plays an important role in this relation. Neglecting $m_e$ in we would find $\Delta K_\ell \sim 2 \cdot 10^{-2}$. There have been quite a few attempts to derive from flavor symmetries or grand-unified extensions of the Standard Model [@Koide:1989jq; @koide_analytical_1996; @koide_u3-family_1995]. Inspired by the idea of grand unification, one may wonder whether a similar mass relations also exist for quarks. Using current values of $d$, $s$ and $b$–quark masses [@PDG]: \[DownQuarkMasses\] $$\begin{aligned} m_d&=4.1 - 5.7\,\, {\rm MeV}\,,\\ m_s&=100^{+30}_{-20}\,\, {\rm MeV}\,,\\ m_b&=4.19^{+0.18}_{-0.06}\,\, {\rm GeV}\,,\end{aligned}$$ and defining, analogously to , $$\begin{aligned} \label{KoideDown} K_{d}\equiv \frac{m_d+m_s+m_b}{(\sqrt{m_d}+\sqrt{m_s}+\sqrt{m_b})^2}\,,\end{aligned}$$ we find $K_{d}\sim 0.72$, i.e. a considerable deviation from the Koide limit with $\Delta K_d \sim 8 \cdot 10^{-2}$. For up-type quarks the mass spectrum has a stronger hierarchy than for down-type quarks [@PDG]: \[UpQuarkMasses\] $$\begin{aligned} m_u&=1.7 - 3.1\,\, {\rm MeV}\,,\\ m_c&=1.29^{+0.05}_{-0.11}\,\, {\rm GeV}\,,\\ m_t&=172.9 \pm 0.6 \pm 0.9 \,\, {\rm GeV}\,,\end{aligned}$$ and therefore the analogue of for up-type quarks, $$\begin{aligned} \label{KoideUp} K_{u}\equiv \frac{m_u+m_c+m_t}{(\sqrt{m_u}+\sqrt{m_c}+\sqrt{m_t})^2}\,,\end{aligned}$$ deviates from the Koide limit even stronger, $K_{u}\sim 0.85$, which corresponds to $\Delta K_u \sim 27 \cdot 10^{-2}$. One could argue, that due to the non-perturbative nature of quantum chromodynamics at low energies, the pole masses of light quarks are not well defined and one should make use of the running quark masses at the scale $M_Z = 91.2$ GeV [@rodejohann_extended_2011]. The latter are given by [@xing_updated_2007]: \[QuarkMassesAtMZ\] $$\begin{aligned} m_u&=1.27^{+0.50}_{-0.42}\,\, {\rm MeV}\,,\\ m_d&=2.90^{+1.24}_{-1.19}\,\, {\rm MeV}\,,\\ m_s&=55^{+16}_{-15}\,\, {\rm MeV}\,,\\ m_b&=0.619\pm 0.084\,\, {\rm GeV}\,,\\ m_c&=2.89\pm 0.09\,\, {\rm GeV}\,,\\ m_t&=171.7\pm 3.0\,\, {\rm GeV}\,.\end{aligned}$$ Substituting into and we find $K_{d}\sim 0.74$ and $K_{u}\sim 0.89$ respectively. This corresponds to $\Delta K_d \sim 11 \cdot 10^{-2}$ and $\Delta K_u \sim 33 \cdot 10^{-2}$. That is, the running effects induce an even stronger deviation from the Koide limit. The same also applies to leptons. The running charged lepton masses at the scale $M_Z = 91.2$ GeV are given by [@xing_updated_2007]: $$\begin{aligned} m_e&=0.486570161 \pm 0.000000042\,\, {\rm MeV}\,,\\ m_\mu&=102.7181359 \pm 0.0000092\,\, {\rm MeV}\,,\\ m_\tau&=1746.24^{+0.20}_{-0.19}\,\, {\rm MeV}\,.\end{aligned}$$ Substituting these values into the Koide relation for leptons we find $1.87 \cdot 10^{-3} < \Delta K_\ell < 1.91\cdot 10^{-3}$. This implies in particular, that the running masses at high scales are not compatible with the exact Koide limit, $K_\ell=2/3$. This result is somewhat counterintuitive. If there is an underlying high-scale symmetry which ensures for leptons, one could expect that the deviation from the Koide limit would be smaller at high scales. It was noted in [@rodejohann_extended_2011] that one could divide quarks into light and heavy ones, instead of up- and down-like quarks, that is, according to their mass instead of the isospin. The corresponding Koide parameters are defined by: \[KoideLightHeavy\] $$\begin{aligned} K_{light}&\equiv \frac{m_u+m_d+m_s}{(\sqrt{m_u}+\sqrt{m_d}+\sqrt{m_s})^2}\,,\\ K_{heavy}&\equiv \frac{m_c+m_b+m_t}{(\sqrt{m_c}+\sqrt{m_b}+\sqrt{m_t})^2}\,.\\end{aligned}$$ Substituting and into we find a rather good agreement with the Koide limit for the heavy quarks, $-5 \cdot 10^{-3} < \Delta K_{heavy} < 1 \cdot 10^{-2}$, and a deviation from it for the light ones, $-20 \cdot 10^{-2} < \Delta K_{light} < - 6 \cdot 10^{-2}$ . If we take the running masses at the scale $M_Z = 91.2$ GeV instead, we will obtain $5 \cdot 10^{-2} < \Delta K_{heavy} < 9 \cdot 10^{-2}$ and $-20 \cdot 10^{-2} < \Delta K_{light} < - 2 \cdot 10^{-2}$ respectively. Although dividing quarks into the light and the heavy ones numerically gives a rather good agreement with the Koide relation, the fact that $s$ and $c$ quarks – components of the same $SU_L(2)$ doublet – enter the expressions for $K_{light}$ and $K_{heavy}$ separately is counterintuitive. A somewhat more natural generalization of the Koide relation to quarks is an expression which sums over up *and* down components of the three generations: $$\begin{aligned} \label{KoideAll} K_{q}\equiv \frac{\sum m_q}{\left(\sum\sqrt{m_q}\right)^2}\,.\end{aligned}$$ Since the active neutrino masses are very small, a similar generalization of the Koide relations for leptons would not affect the range for $\Delta K_\ell$ indicated above. Substituting and into we find for the deviation from the Koide limit $-5 \cdot 10^{-2} < \Delta K_q < - 4 \cdot 10^{-2}$. If we take the running quark masses at the scale $M_Z = 91.2$ GeV instead, we will obtain $2 \cdot 10^{-2} < \Delta K_q < 5 \cdot 10^{-2}$. It is interesting that due to the running effects $\Delta K_q$ crosses zero. That is, at some scale 2 GeV $< \mu < M_Z$ the quark masses satisfy (within the uncertainties) the Koide relation $K_q=2/3$. It was noted in [@foot_note_1994] that the Koide relation has a geometric interpretation. Let $(\sqrt{m_e},\sqrt{m_\mu},\sqrt{m_\tau})$ define a vector in a three-dimensional vector space and $\theta_\ell$ be the angle between this vector and the vector $(1,1,1)$, $$\begin{aligned} \label{theta_ell} \theta_\ell\equiv \arccos\frac{(\sqrt{m_e},\sqrt{m_\mu},\sqrt{m_\tau})\cdot(1,1,1)}{|(\sqrt{m_e},\sqrt{m_\mu},\sqrt{m_\tau})||(1,1,1)|}\,.\end{aligned}$$ Then the Koide relation is equivalent to the statement $\theta_\ell=\pi/4$. Any deviation of $K$ from the Koide limit, $K=2/3$, implies also a deviation from $\theta=\pi/4$. Thus, $\theta_d$ and $\theta_d$ substantially deviate from $\pi/4$, whereas $\theta_{light}$ and $\theta_{heavy}$ are close to this limit. It is interesting that if we define an analogue of for quarks in a six-dimensional vector space, $$\begin{aligned} \theta_q\equiv \arccos\frac{(\sqrt{m_d},\sqrt{m_u},\sqrt{m_s}, \ldots)\cdot(1,1,1,\ldots)}{|(\sqrt{m_d},\sqrt{m_u},\sqrt{m_s},\ldots)||(1,1,1,\ldots)|}\,,\end{aligned}$$ then we will find that this angle is again a rational multiple of $\pi$, $\theta_q\approx\pi/3$. To summarize, generalizations of the original Koide relation to quarks which include only the up-type or only the down-type quarks show a considerable deviation from the Koide limit $K=2/3$. In this letter we have proposed a generalization which includes up *and* down quarks of the three generations. Interestingly enough, although masses of the light ($u$, $d$ and $s$) quarks are much smaller than masses of the heavy ones ($c$, $b$ and $t$), their inclusion into the sum makes it substantially closer to the Koide limit. This resembles the role played by the electron mass in the Koide relation for charged leptons. Furthermore, similarly to the Koide relation for leptons, the proposed relation for quarks also allows for a geometric interpretation: the angle between vectors $(\sqrt{m_d},\sqrt{m_u},\sqrt{m_s}, \ldots)$ and $(1,1,1,\ldots)$ in a six-dimensional vector space is very close to $\pi/3$. Acknowledgements {#acknowledgements .unnumbered} ---------------- This work was supported by Deutsche Forschungsgemeinschaft under Grant KA-3274/1-1. The author would like to thank A. Hohenegger for careful reading of the manuscript. [9]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , ****, (). , ****, (). , ****, (). , (), . , (), . , ****, (), . , , , (), . , ().

ArXiv

--- abstract: 'We present a variation of the Autoencoder (AE) that explicitly maximizes the mutual information between the input data and the hidden representation. The proposed model, the InfoMax Autoencoder (IMAE), by construction is able to learn a robust representation and good prototypes of the data. IMAE is compared both theoretically and then computationally with the state of the art models: the Denoising and Contractive Autoencoders in the one-hidden layer setting and the Variational Autoencoder in the multi-layer case. Computational experiments are performed with the MNIST and Fashion-MNIST datasets and demonstrate particularly the strong clusterization performance of IMAE.' author: - Vincenzo Crescimanna - Bruce Graham title: An information theoretic approach to the autoencoder --- Introduction ============ Nowadays, Deep Neural Networks (DNNs) are considered the best machine learning structure for many different tasks. The principal property of this family is the ability to learn a representation of the input data in the hidden layers. Indeed, as underlined in [@ben_replen], in order to understand the world, an intelligent system needs to learn a compact representation of the high dimensional input data. In biological Neural Networks (NNs) processing sensory information, it is hypothesised that the first layers encode a task-independent representation that maximizes the mutual information with the input [@barlow]. Given that it is desirable to learn a representation that is able to identify and disentangle the underlying explanatory hidden factors in input data, and inspired by the behaviour of real NNs, we propose an unsupervised Infomax Autoencoder (IMAE) NN that maximizes the mutual information between the encoded representation and the input. The results of computational experiments show that IMAE is both robust to noise and able to identify the principal features of the data. The paper is structured as follows. In the next section we describe related work. Section 3 describes the derivation of the model, which is based on a classic approximation of differential entropy. Due to comparison with other AEs, we give also a geometric and a generative description of the model. The theoretical differences are confirmed in the fourth section by computational experiments, where the clusterization (representation) and the robustness of IMAE are evaluated. The work is concluded with a brief comment on the results and a description of future work. Related Work ============ The first neural network defined to explicitly maximize the information between the input and the hidden layer was proposed by Linsker [@linsker], giving to the objective function the name, *Infomax principle*. This model is a linear model that actually maximizes only the entropy of the representation, performing Principal Component Analysis (PCA) in the case of Gaussian distributed data. The same objective function was applied by Bell and Sejnowsky [@bell95info], showing that a 1-layer neural network with sigmoid activation performs Independent Component Analysis (ICA). Both these models are quite restrictive, indeed they work only under the assumption that the visible data is a linear combination of hidden features. A more general way to extract hidden features is given by the Autoencoder (AE), a NN that is a composition of an encoder and a decoder map, respectively $W$ and $V$. This model can be seen as a generalization of the PCA model, because in the assumption $W$ and $V$ are linear maps, the space spanned by $W$ is the same of the one spanned by principal components, see. e.g. [@Baldi]. An information theoretic description of an AE was given by Vincent et al. [@Vincent1] where, with restrictive assumptions, they observed that reducing the reconstruction loss of an AE is related to maximizing the mutual information between the visible and hidden variables. The IMAE, as we will see in the next section, is an AE that is able to learn a robust representation. In the literature there are many models in this family that are defined to learn a good representation, see e.g. [@karhunen2015] and references therein. For practical reasons, in the following section we compare the IMAE with the AEs that are recognized to learn the best features: Denoising, Contractive and Variational Autoencoders. Infomax autoencoder model ========================= Shannon [@Shannon] developed Information theory in order to define a measure of the efficiency and reliability of signal. In this theory a key role is assumed by Mutual Information $I$, i.e. a measure of the information shared between two variables, in the signal transmission case: the original message and the received one. Formally, given two random variables: $X$ and $Y$, the mutual information between these two variables is defined as $$\label{eq_IM} I(X,Y) = h(Y) - h(Y|X)$$ where $h(Y)$ is the (differential) entropy of $Y$, a measure of the average amount of information conveyed, and $h(Y|X)$ is the conditional entropy $Y$ with respect to $X$, a measure of the average of the information conveyed by $Y$ given $X$, i.e. the information about the processing noise, rather than about the system input $X$. Given the assumption the hidden representation $Y$ lives in a manifold embedded in a subspace of the visible space, by the definition above, the Infomax objective function finds the projection from the visible (higher dimension space) to the hidden manifold that preserves as much information as possible. Defining $W$ as the projection map, the Infomax objective can be state as: $\max_W I(X, W(X))$, with the representation, $Y = W(X)$, that maximizes the mutual information with the system input $X$. In general, to compute the mutual information between two variables is not trivial, indeed it is necessary to know the distribution of the input and of the hidden representation, and to compute an integral. For this reason, we propose an approximation of the mutual information $I(X,Y)$. In the following, capital letters $X, \hat{X}, Y$ denote the random variables and lowercase letters $x, \hat{x} \in \mathbb{R}^d, y \in \mathbb{R}^l $ the respective realisations. The map between the input space $\mathbb{R}^d$ to the hidden representation $\mathbb{R}^l$ is $W, Y = W(X) $, and the map from $\mathbb{R}^l$ to the input space is $V, \hat{X} = V(Y)$. In the first step we assume that the map $W$ is invertible, then both $x, y \in \mathbb{R}^d$. Our approximation starts by considering the conditional entropy in eq (\[eq\_IM\]). In order to estimate this quantity it is necessary to suppose there exits an unbiased efficient estimator $\hat{X} = V(Y) $ of the input $X$, in this way the conditional entropy $h(Y|X)$ can be approximated by $h(\hat{X}|X)$. Indeed, the conditional entropy is the information about the processing noise, and the noise added in the encoding process is proportional to the noise between the input and its reconstruction. Formally, see e.g. [@papa] p. 565, in the case of a bijiective function the following holds $$\label{_ineq} h(\hat{X}|X) \leq h(Y|X) + \int \log |J_V(y)| dp(y),$$ where the equality holds in the case that $V$ is an orthogonal transformation, with $J_V$ the Jacobian of the function $V$ and $|\cdot| = |\det(\cdot)| $ . Let us start by assuming the second term on the right-hand-side (RHS) is bounded (in the end we show that this term is bounded). In order to have the inequality as tight as possible, the idea is to assume the conditional variable $\hat{X}|X$ is Gaussian distributed with mean $X$ and variance $E[(X-\hat{X}) (X-\hat{X})^T]$. Indeed, the Gaussian distribution has maximum entropy. Since $\hat{X}|X$ is Gaussian distributed, the conditional entropy $h(\hat{X}|X)$ is an (increasing) function of the covariance matrix $\Sigma_{\hat{x}|x}$, i.e. $h(\hat{X}|X) \propto \mbox{const} +\log(|\Sigma_{\hat{x}|x}|)$. By the symmetric definition of the covariance matrix and the well known relationship $log(|\Sigma|) = trace(\log(\Sigma))$, it is sufficient to consider as the loss function to minimize the trace of $\Sigma_{\hat{x}|x}$, i.e. the $L_2$-norm of the input with its reconstruction. In this way, we obtain formally that $V$ should be near, in the $L_2$ space of functions, to the inverse of $W$, i.e. $V\circ W \approx Id$; then also the Jacobian $J_{V \circ W} \approx Id$. Now let us focus on the first term of the RHS of (\[eq\_IM\]), the differential entropy $h(Y)$. Defining the function $W$ as a composition of two functions: $W = \sigma \circ W_0$, the entropy function can be written as follows: $$h(Y) = \int\log |J_\sigma(Y_0)|p(y_0) + h(Y_0)$$ where $Y_0 = W_0(X)$. In our application, we consider the non-linearity $\sigma$ as the logistic function $\sigma(x) = (1+\exp(-x))^{-1}$; this allows to have the representation $Y_0$ distributed following the logistic distribution, see [@bell95info]. In order to compute the second term, $h(Y_0)$, following [@hyvrinen1998new; @hyvarinen2000independent], this term can be seen as a sparse penalty in the $Y_0$ output, that, as observed in [@lee_ica], encourages to have an independent distribution of the components of $Y_0$. Since the $Y_0$ components are almost independent, a good approximation for the determinant of the Jacobian of the sigmoid is the product of its diagonal elements. Then observing that the variance of $Y_0$ is controlled by the Jacobian of the sigmoid, the entropy of $Y$, can be approximated as: $$\label{ent} h(Y) = \sum_i \sigma_i(Y_0)(1-\sigma_i(Y_0)) - (\log(\cosh(Y_0)_i))^2,$$ where the first term in the RHS is associated to the $\log$ of the determinant sigmoid Jacobian, and the second term is the sparsity penalty, associated to $h(Y_0)$. The sparsity penalty in the equation above ensures $W$ to be a bounded operator and, by the open mapping theorem, $J_W$ and $J_V$ have bounded determinants different from zero, thus the inequality (\[\_ineq\]) is well defined. #### Generalization to under- and over-complete representations. The approximation of the conditional entropy as the $L_2$ reconstruction loss was derived under the condition that the input and the representation live in the same space. A way to generalize the derivation to the under- and over-complete setting is thinking of this structure as a noisy channel, where some units (components) in the hidden and input layers are masked. Indeed, since by definition $V$ behaves as the inverse of the (corrupted) map $W$, it is not adding more, or even different, noise than that already added in the map $W$. Thus, we can conclude that the approximation $h(Y|X) \approx h(\hat{X}|X)$ holds both in the over- and under- complete setting. By construction, the approximation of the entropy term does not have a particular requirement, so can be described for a generic number of hidden units. #### Properties. The difference of IMAE from the other models in the literature is given by the latent loss term i.e. the maximization of the latent entropy. Choosing the sigmoid activation in the latent layer, in accordance with the theory of Independent Component Analysis and processing signals [@bookhyva], we suppose that the latent distribution is logistic and concentrated around the mean. In particular, the first term of the entropy (\[ent\]), ensures that the distribution is peaked around the mean value, guaranteeing the learning of good prototypes of the input data. Instead, the sparseness penalty encourages the independence between the variables, guaranteeing a model robust to noise. ### Relationship with other approaches. In order to describe the properties of IMAE from different perspectives, we compare it with the most common AE variants able to learn good representations. In all the cases that we consider the encoder map has the form $\sigma \circ W_0$. #### Contractive Autoencoder (CAE). Starting from the idea that a low-derivative hidden representation is more robust, because it is less susceptible to small noise, Rifai et al. [@rifai] suggested a model that explicitly reduces (contracts) the Jacobian of the encoder function. The proposed Contractive Autoencoder (CAE) has a structure similar to IMAE, with sigmoid activation in the hidden layer, but the latent loss is the Froebenius norm of the encoder Jacobian $J_W$. In the case $W_0 $ is a linear matrix, the latent loss can be written as $$\label{rif_eq} \| J_W(x) \|_F^2 = \sum_i^l (\sigma_i(1-\sigma_i))^2\sum_j^d (W_0)_{ij}^2.$$ From eq (\[rif\_eq\]), it is clear that CAE encourages a flat hidden distribution, the opposite of IMAE, that encourages a big derivative around the mean value; the consequences of such differences will be underlined and clarified in the computational experiment section, below. From an information theory perspective, observing that the Froebenius norm is an approximation of the absolute value of the determinant, the CAE representation can be described as a low entropy one. Indeed, by changing variables in the formula, in the case of a complete representation, the entropy of $Y$ is a linear function of the log-determinant of the Jacobian of $W$. #### Denoising Autoencoder (DAE). Assuming that robust reconstruction of the input data implies robust representation, Vincent et al. [@Vincent1] proposed a variant of the AE trained with corrupted input, termed the Denoising Autoencoder (DAE). From a manifold learning perspective, Vincent et al. observed that the DAE is able to learn the latent manifold, because it learns the projection from the corrupted $\tilde{X}$ to the original $X$ input. The problem, as highlighted in section 4, is that there is no assumption on the distribution of the data around the hidden manifold i.e. DAE learns only a particular type of projection and then is robust to only one type of noise. Differently in IMAE, the preferred projection direction is suggested by the choice of the hidden activation. #### Variational Autoencoder (VAE). The Variational Autoencoder (VAE) [@kigma] is not a proper neural network, but actually a probabilistic model; differently from the other models described until now it does not discover a representation of the hidden layer, but rather the parameters $\lambda$ that characterize a certain distribution chosen a-priori $q_\lambda(y_0|x)$ as a model of the unknown probability $p(y_0|x)$. In the case that $q_\lambda$ is Gaussian, the parameters are only the mean and variance and VAE can be written as a classic AE with loss function: $$\begin{aligned} \mathcal{L}_{VAE} = - E_q(\log p(x|y_0)) + D_{KL}(q_\lambda(y_0|x)\| p(y_0)).\end{aligned}$$ The first term of $\mathcal{L}_{VAE}$ is approximately the reconstruction loss. The Kullback-Leibler (K-L) divergence term, $D_{KL}$, which is the divergence in the probability space between the a-priori and the real distributions, represents the latent loss of the neural network, approximated in our case as: $\sum_i \mu(y_0)_i^2 + \sigma(y_0)_i^2 - \log(\sigma(y_0)_i^2) -1. $ From the IMAE perspective, we observe that the first term of the latent loss, $-h(Y)$, encourages the $Y_0$-distribution to be logistic with small variance, and the second term penalizes the dependencies between the $Y_0$ components. Then we can conclude that, from a probabilistic perspective, the IMAE latent loss works like a K-L divergence, with the target function a product of independent logistic distributions. Computational experiments ========================= In this section we describe the robustness and the clusterization performance of IMAE, comparing it with the models described above and the classic AE. We carry out the tests with two different architectures: single- and multi-hidden layers. The first comparison is made with AE, CAE and DAE; in the second comparison, which is suitable for a small representation, we compare IMAE with VAE. 1-hidden layer -------------- The first tests are carried using the MNIST dataset, a collection of $28\times 28$ grey-scale handwritten digits, subdivided into ten classes, with 60 000 training and 10 000 test images. We consider two different AutoEncoder architectures, one under-complete with 200 hidden units and the other over-complete with 1000 hidden units. All the neural networks are trained using batch Gradient Descent with learning rate equal to 0.05 and batch size composed by 500 input patterns (images), trained for 2000 epochs. For all the models, the encoder and decoder functions have form: $W(x) = \sigma(W_0x)$ and $V(x) = Vx$. The listed results are obtained with tied weights, $W_0 = V^T$, but equivalent results are obtained with $W$ and $V$ not related to each other. The DAE is trained both with Gaussian (DAE-g) and mask (DAE-b) noise. In DAE-g, each pixel of the training images is corrupted with zero-mean Gaussian noise with standard deviation $\sigma = 0.3$; in DAE-b, each pixel is set to zero with probability $p = 0.3$ . In CAE and IMAE, the hyper-parameters associated with the latent loss are set equal to $0.1$ and $1$, respectively. #### Robust reconstruction. As suggested by [@Vincent1], a way to evaluate the robustness of the representation is by evaluating the robustness of the reconstruction. Under this assumption, for each model we compute the $L_2$-loss between the reconstructed image and the original (clean) one. Results for this measure are listed in Table \[l2\_MNIST\]. From these results we can deduce that IMAE is equally robust in all the considered settings and, excluding the respective DAEs trained with the test noise, has the best reconstruction performance. In contrast, DAE is robust only to the noise with respect to which it was trained and CAE is only robust in the over-complete setting with respect to Gaussian noise. The classic AE is best in the zero-noise case. \ #### Prototype identification. A good, task-independent representation should include the hidden prototypes for each class. Assuming that the hidden representation of each element lies close to the respective prototype, to evaluate representation quality we measure the clusterization performance. In particular, we compute the Rand-index between the ten clusters identified by the K-means algorithm on the hidden space and the ones defined by ground-truth in the MNIST data set. Formally we compute the Rand-index $R$ as follows: $$R = \max_m \frac{\sum_{i=1}^N \mathbf{1}(m(c_i) = l_i)}{N}$$ where $c_i\in \{1, \dots, k\}$ is the hidden representation cluster in which the current input, $x_i$, belongs, $l_i \in \{1, \dots, k \}$ is the ground truth label associated with $x_i \in \{ x_i\}_1^N$, $m$ is a 1-1 map from cluster to label, and $\mathbf{1}$ is the indicator function, that is 1 if $m(c_i) = l_i$, 0 otherwise. In our case $k = 10$, the number of different classes, and $N = 1000$; the results are shown in Table \[clust\]. IMAE has the best clusterization performance (high Rand index) and is also the more robust to noise. In particular, we observe that these performances are associated with the mean derivative of the non-linear hidden activation, where a high derivative is associated with good performance. This observation leads us to conclude that the derivative value can be considered as an empirical measure of the clusterization performance. Concerning both sets of results, we underline that reconstruction robustness is not associated with robust clusterization. For example, according to the dropout idea, DAE-b has a more robust representation than DAE-g also in the case of the Gaussian noise corruption on which it was not trained. Multi-hidden layers ------------------- In many applications a shallow network structure is not sufficient, because the considered data are more redundant than the MNIST data set and the desired representation is really small. For this reason it is important to consider a deeper structure. In this setting AE, CAE and DAE cannot naturally be extended, so we compare IMAE with VAE, the AE that, better than others, is able to learn a hidden representation. We now consider a network with hidden architecture of type: 1100-700-$n_h$-700-1100, with softplus activation in each hidden layer except for the $n_h$ units, where a sigmoid nonlinearity is applied. In the following, we use hidden units to refer to the $n_h$-variables, where $n_h \in \{5, 10, 20\}$. The tests are carried out with MNIST and Fashion-MNIST (F-MNIST), a MNIST-like dataset where the digits are replaced by images of clothes, [@fmnist]. This dataset is suitable for this network structure since it is more redundant than MNIST but still low-dimensional. The network is trained as in the 1-hidden layer architecture, with learning rate equal to $5\cdot 10^{-3}$. Since we are interested to learn a small representation of the data, we compare the clusterization performance of VAE and IMAE. We do not consider the robustness of reconstruction because, by construction, the models are robust to really small noise. #### Clusterization Performance. Table \[l2\_MNIST\_tab\] lists the Rand-indices obtained in the hidden space, with possibly Gaussian-noise-corrupted input data. In the noisy case the input data is corrupted with zero mean Gaussian noise with standard deviation, respectively, $\sigma = 0.01, 0.1$ for MNIST and F-MNIST. For MNIST, we choose this relatively small noise because, since it is not particularly redundant, the deep structure is not robust. ------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- -------- -------- ---------- ---------- $n_h$ noisy $n_h$ noisy Model 5 10 20 5 10 20 5 10 20 5 10 20 VAE 70.1 [57.2]{} 52.9 61 [51.7]{} 36 40.3 42.3 42.4 38.1 41.8 41.3 IMAE **76.8** **75.7** **54.8** **68.7** **60.5** **42.2** **55.5** **59.3** **56** **55** **55.3** **55.3** ------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- -------- -------- ---------- ---------- : Average over 50 iterations of Rand-indices of VAE and IMAE with networks having 5, 10 or 20 hidden units. []{data-label="l2_MNIST_tab"} The results highlight how IMAE is able to clusterize the data in a small hidden space and identify the prototypes for each class also in the case of noisy input data i.e. the hidden prototypes are robust. In order to clarify the robustness of the clusters, Figure \[tsne\_hidden\] visualises the hidden representations of the two models in the case of noisy input data. In the IMAE representation, despite the noise, it is possible to clearly distinguish the different clusters; instead in the VAE representation all the points are mixed, except largely the orange ones (bottom cluster), which represent the digit 1. Conclusions and future work =========================== In this paper, we derived formally an Autoencoder that explicitly maximizes the information between the input data and the hidden representation and introduces an empirical measure of the clusterization performance on the hidden space. The experiments show that IMAE is versatile, works both in shallow and deep structures, has good reconstruction performance in the case of noisy input data, and is able to learn good and robust prototypes in the hidden space. Finally, we observe that starting from its generative description, IMAE can be extended to a context-dependent model, a generative model able to learn a representation of the input associated to a context variable. The idea, as suggested in [@kay], is to consider an AE with lateral (contextual) hidden connections, that modify the shape of activation, allowing to learn a context-specific representation for the input. a gated NN able to learn a context-dependent representation, simply modifying the shape of the hidden activation, in the same fashion as done in [@kay]. By IMAE generative description, where the distribution of the hidden units is given by their activation We observed further that IMAE can be seen as a generative model, where the distribution of the hidden units is given by their activation. That observation allows to generalize the IMAE to a gated NN that, modulating the activation, is able to learn a context-dependent representation The latter point, that is also the principal innovation, is particularly useful for applications where it is necessary to integrate different representations. For example, in case-based reasoning approaches [@kim_cbm], where there are some contextual variables influencing the behaviour of an intelligent system, the ability of IMAE to recognize, with a small representation, the context, would be useful. The latter point, that is also the principal innovation, is particularly useful for applications, such as case-based reasoning approaches [@kim_cbm], where it is necessary to identify clear prototypes to represent each case. ### Acknowledgments: {#acknowledgments .unnumbered} This research is funded by the the University of Stirling CONTEXT research programme and by Bambu (B2B Robo advisor, Singapore). [8]{} Baldi, P., and Hornik, K.: Neural networks and principal component analysis: Learning from examples without local minima. Neural networks, **2** (1), 53–58, (1989). Barlow, H.: Possible principles underlying the transformation of sensory messages, Sensory Communication, MIT Press, (1961) Bell, A. J., and Sejnowski, T. J. An information-maximization approach to blind separation and blind deconvolution. Neural computation, **7** (6), 1129–1159, (1995). Bengio, Y., Courville, A. and Vincent, P.: Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, **35** (8), 1798–1828 (2013). Brunel, N. and Nadal, J. P.: Mutual information, Fisher information, and population coding. Neural computation, **10** (7), 1731–1757, (1998). Hyvärinen, A., Hurri, J. and Hoyer, P. O.: Natural image statistics: a probabilistic approach to early computational vision. Springer, (2009). Hyvrinen, A.: New approximations of differential entropy for independent component analysis and projection pursuit, Neural Information Processing Systems, **10**, 273–279, (1998) Hyvrinen, A. and Oja, E: Independent component analysis: algorithms and applications, Neural networks, Elsevier, **13** (4-5), 411–430, (2000). Kay, J.W. and Phillips, W.A.: Coherent Infomax as a computational goal for neural systems. Bulletin of mathematical biology, **73** (2), 344–372, (2011). Karhunen, J., Raiko, T. and Cho, K.: Unsupervised deep learning: A short review. In Advances in Independent Component Analysis and Learning Machines, 125–142 (2015). Kim, B., Rudin, C., and Shah, J. A.: The bayesian case model: A generative approach for case-based reasoning and prototype classification. In Advances in Neural Information Processing Systems, 1952–1960 (2014). Kingma, D. P. and Welling, M.: Auto-Encoding Variational Bayes. Proceedings of the 2nd International Conference on Learning Representations, (2014). Lee, T-W and Girolami, M. and Bell A.J. and Sejnowski, T. J.: A unifying information-theoretic framework for independent component analysis. Computers & Mathematics with Applications **39** (11), 1–21, (2000). Linsker, R.: Self-organization in a perceptual network. Computer, IEEE, **21** (3), 105–117, (1988) Maaten, L. V. D. and Hinton G.: Visualizing data using t-SNE. Journal of machine learning research **9** (Nov), 2579–2605 (2008). Papoulis, A. and Pillai, S. U.: Probability, random variables, and stochastic processes. Tata McGraw-Hill Education (2002). Rifai, S., Vincent, P., Muller X., Glorot, X. and Bengio, Y.: Contractive auto-encoders: Explicit invariance during feature extraction. In Proceedings of the 28th International Conference on International Conference on Machine Learning,Omnipress, 833–840, (2011). Shannon, C. E. :A Mathematical Theory of Communication, Bell System Technical Journal. **27** (3): 379–423, (1948). Tishby, N. and Slonim, N. : Data clustering by markovian relaxation and the information bottleneck method. In Advances in neural information processing systems. 640–646, (2001). Vincent, P., Larochelle, H., Bengio, Y., and Manzagol, P. A.: Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, July, 1096–1103, (2008). Xiao, H., Rasul, K. and Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, (2017).

ArXiv

--- abstract: 'We give a new treatment of the $\pi$-calculus based on the semantic theory of separation logic, continuing a research program begun by Hoare and O’Hearn. Using a novel resource model that distinguishes between public and private ownership, we refactor the operational semantics so that sending, receiving, and allocating are commands that influence owned resources. These ideas lead naturally to two denotational models: one for safety and one for liveness. Both models are fully abstract for the corresponding observables, but more importantly both are very simple. The close connections with the model theory of separation logic (in particular, with Brookes’s action trace model) give rise to a logic of processes and resources.' address: | College of Computer and Information Science\ Northeastern University\ Boston MA, USA author: - Aaron Turon - Mitchell Wand bibliography: - 'pi-semantics.bib' title: 'A resource analysis of the $\pi$-calculus' --- Names play a leading role in the $\pi$-calculus [@Milner1992]: they are both the means of communication, and the data communicated. This paper presents a study of the $\pi$-calculus based on a new mechanism for name management, which is in turn rooted in separation logic. The main benefit of this study is a very simple—but fully abstract—denotational semantics for the $\pi$-calculus. Traditionally, the use of names in the $\pi$-calculus is governed by lexical, but dynamically-expandable, scope. In the composite process $P | \new x. Q$ for example, the channel $x$ is by virtue of scope initially *private* to $Q$. The prefix $\new x$ is not an imperative allocation. It is a binder that remains fixed as $Q$ evolves—a constant reminder that $x$ is private—until $Q$ sends $x$ in a message. At that point, the binder is lifted to cover both $P$ and $Q$, dynamically “extruding” the scope of $x$. The $\pi$-calculus relies on $\alpha$-renaming and side conditions about freshness to ensure that its privacy narrative is borne out. In contrast, work on separation logic has led to models of dynamically-structured concurrency based on resources and ownership, rather than names and scoping [@Brookes2007; @Calcagno2007]. From this perspective, programs consist of imperative commands that use certain resources (their “footprint”) while leaving any additional resources unchanged. Concurrent processes must divide resources amongst themselves, with each process using only those resources it owns. Ownership makes it possible to constrain concurrent interference, and thereby to reason compositionally about process behavior. In this paper, we reanalyze the $\pi$-calculus in terms of resources and ownership, establishing a clear connection with models of separation logic. The analysis hinges on the use of resources to specify not just that a process can do something, but that other processes cannot.[^1] Concretely, channels are resources that can be owned either publicly or privately. Public ownership asserts only that a channel can be used by the owning process. Private ownership asserts moreover that a channel cannot be used by other processes. And the prefix $\new x$ becomes an imperative action, allocating an initially private channel. Armed with this simple resource model, we give a new operational semantics for the $\pi$-calculus (). The semantics is factored into two layers. The first layer generates the basic labeled transitions, without regard to their global plausibility. The second layer then uniformly interprets those labels as resource transformers, filtering out implausible steps. The two-layer setup is reminiscent of Brookes’s semantics for concurrent separation logic [@Brookes2007; @Brookes2002], and allows us to blend message-passing and imperative interpretations of actions. More importantly, the resource model also enables a very simple denotational treatment of the $\pi$-calculus. We give two denotational interpretations, both trace-theoretic. The first () captures safety properties only, while the second () is also sensitive to divergence and some branching behavior, along the lines of the failures/divergences model with infinite traces [@Roscoe1993]. We prove that each model is fully abstract with respect to appropriate observables. The semantic foundation reconciles the model theory of separation logic with the $\pi$-calculus; what about the proof theory? We sketch an integration of separation logic with refinement calculus for processes (). Refinement is justified by the denotational semantics, so the calculus is sound for contextual approximation. Resource reasoning allows us to derive an *interference-free expansion law* that uses privacy assertions to rule out interference on a channel. To provide an accurate model of the $\pi$-calculus, public/private resources must be *conservative* in a certain sense: once a resource has been made public, it is impossible to make it private again. Work in separation logic has shown the usefulness of more “aggressive” resource models that capture not just what can and cannot be done, but assert that certain things *may* not be done. We sketch a few such aggressive resource models (), including an interpretation of fractional permissions [@Boyland2003] and of session types [@Honda1998]. Hoare and O’Hearn initiated a study of a $\pi$-calculus-like language in terms of separation logic semantics [@Hoare2008]. That study provided the impetus for our work, which goes farther by (1) handling the full calculus, (2) handling liveness, (3) proving full abstraction and (4) building a logic on the semantics. There have also been several fully abstract models of the $\pi$-calculus [@Stark2002; @Hennessy2002; @Fiore2002] based on functor categories for modeling scope. Our models complement these by providing an elementary account of behavior, structured around resources and abstract separation logic. A full discussion of related work is in . A resource-driven operational semantics {#sec:operational} ======================================= There are many variants of the $\pi$-calculus; here’s ours: $$\begin{array}{c} P \ ::=\ \sum \pi_i. P_i \GA P \nc Q \GA \new x. P \GA P|Q \GA \rec X. P \GA X \\ \pi \ ::=\ \ov{e}e' \GA e(x) \qquad \qquad e \ ::=\ x \GA c \end{array}$$ We distinguish between external choice ($+$) and internal choice ($\oplus$), which simplifies the liveness semantics () but is not essential. We also distinguish between channels ($c, d$) and channel variables ($x, y, z$) and include a simple grammar of channel expressions ($e$) ranging over both. A *closed* process has no unbound channel or process variables. Closed processes may, however, refer to channel constants and thereby communicate with an environment. We write $0$ for an empty summation, which is an inert process. Generating actions ------------------ The operational semantics of closed processes is given in two layers, via two labelled transition systems. In both systems, the labels are (syntactic) *actions*, given by the following grammar: $$\begin{aligned} \alpha &::=& c!d \GA c?d \GA \nu c \GA \tau \GA \fault \qquad (\textsc{Action})\end{aligned}$$ Actions record the concrete channels involved in sending, receiving, and allocating, respectively. The action $\tau$, as usual, represents an internal (unobservable) step on the part of the process. The action $\fault$ represents a fault, caused by using an unowned channel (). Communication actions are dual: $\ov{c!d} = c?d$ and $\ov{c?d} = c!d$, while $\ov{\nu c}$, $\ov{\tau}$, and $\ov{\fault}$ are undefined. The first transition system generates all conceivable actions associated with a process, without considering whether those actions are globally plausible: [Operational semantics: action generation]{} $$\begin{array}{rcl} \cdots + \ov{c}d.P + \cdots &\step{c!d}& P {}\\ \cdots + c(x).P +\cdots &\step{c?d}& P \{d/x\} {} \\ P_1 \nc P_2 &\step{\tau}& P_i {} \\ \new x.P &\step{\nu c}& P\{c/x\} {} \\ \rec X. P &\step{\tau}& P\{\rec X.P/ X\} {} \end{array} \quad \begin{array}{c} \infer[{}] {P \step{\alpha} P'} {P|Q \step{\alpha} P'|Q} \qquad \infer[{}] {Q \step{\alpha} Q'} {P|Q \step{\alpha} P|Q'} \\ \\ \infer[{}] {P \step{\alpha} P' \\ Q \step{\ov{\alpha}} Q'} {P|Q \step{\tau} P'|Q'} \end{array}$$ According to this semantics, we will have transitions like $$\new x. \new y. \ov{x}y. 0 \ \step{\nu c}\ \new y. \ov{c}y. 0 \ \step{\nu c}\ \ov{c}c. 0 \ \step{c!c}\ 0$$ where $c$ is allocated twice, and used to communicate with an environment that cannot know it. To filter out such executions, we use resources. Resources and action semantics {#sec:action-sem} ------------------------------ The execution above is intuitively impossible because, after the first $\nu c$ action, the process *already owns* the channel $c$. Similarly, for the process $\new x. \ov{x}x . 0$ the trace $$\new x. \ov{x}x. 0 \ \step{\nu c}\ \ov{c}c. 0 \ \step{c!c}\ 0$$ is impossible because the channel $c$, having just been allocated, is unknown to the environment—so no parallel process could possibly be on the other side of the communication, receiving along $c$. Formally, resources are elements $\sigma$ of the domain $ \Sigma \eqdef \textsc{Chan} \rightharpoonup \{ \pub, \pri \} $, where $\pub$ and $\pri$ are distinct atoms. If a process is executing with resources $\sigma$, it owns the channels $\dom(\sigma)$, and $\sigma(c)$ tells, for each $c$, whether that ownership is exclusive. Therefore, if $c \in \dom(\sigma)$, the action $\nu c$ is impossible. Likewise, if $\sigma(c) = \pri$, the action $c!c$ is impossible. The resources owned at a particular point in time determine not only what is *possible*, but also what is *permissible*. For example, the process $\ov{c}d.0$ immediately attempts a communication along the channel $c$. If this channel is not allocated (*i.e.*, not owned, *i.e.*, not in $\dom(\sigma)$) then the process is *faulty*: it is attempting to use a dangling pointer. We interpret actions $\alpha$ as *resource transformers* of type $\Sigma \ra \Sigma^\top_\bot$.[^2] Since all nondeterminism is resolved during the generation of actions, these transformers are deterministic. A result of $\top$ or $\bot$ represents that an action is not permissible or not possible, respectively. Given the semantics $\ASem{\alpha} : \Sigma \ra \Sigma^\top_\bot$ of actions (defined below), we can define a transition system that *executes* actions according to the currently-owned resources: \[$P,\sigma \step{\alpha} P',\sigma'$\][Operational semantics: resource sensitivity]{} $$\infer[{}] {P \step{\alpha} P' \\ \ASem{\alpha}{\sigma} = \sigma'} {P,\sigma \rstep{\alpha} P', \sigma'} \qquad \infer[{}] {P \step{\alpha} P' \\ \ASem{\alpha}{\sigma} = \top} {P,\sigma \rstep{\fault} 0,\sigma}$$ Successful actions proceed normally, updating the owned resources—note that if $\ASem{\alpha}{\sigma} = \sigma'$ then in particular $\ASem{\alpha}{\sigma} \neq \top, \bot$. Impermissible actions noisily fail, generating the faulting label $\fault$. Impossible actions silently fail to occur. The semantics of actions is as follows: [Action semantics]{} $$\begin{array}[t]{rcl@{\quad}rcl} \ASem{c!d}{\sigma} &\eqdef& \begin{cases} \top & \{c,d\}\not\subseteq \dom(\sigma) \\ \sigma[d \ \pub] & \sigma(c) = \pub \\ \bot & \ow \end{cases} & \ASem{c?d}{\sigma} &\eqdef& \begin{cases} \top & c \notin \dom(\sigma) \\ \sigma[d \ \pub] & \! \begin{array}[t]{l} \sigma(c) = \pub,\\ \quad \sigma(d) \neq \pri \end{array} \\ \bot & \ow \end{cases} \\ \ASem{\nu c}{\sigma} &\eqdef& \begin{cases} \sigma[c \ \pri] & c \notin\dom(\sigma) \\ \bot & \ow \end{cases} & \ASem{\tau}{\sigma} &\eqdef& \sigma \qquad \ASem{\fault}{\sigma} \ \eqdef\ \top \end{array}$$ Allocation is always permitted, but is not possible if the channel is already allocated. Allocated channels are initially private. Sending a channel publicizes it, but the communication is only possible if performed over an already public channel, and only permitted over an allocated channel. A locally-unknown channel received from the environment is known to the environment, and hence public; a locally-known channel received from the environment cannot possibly have been private. #### Examples Consider the process $\new x. 0$. We have $$\new x. 0\quad \step{\nu c}\quad 0$$ for every channel $c$. It follows that $$\new x. 0,\ \emptyset\quad \rstep{\nu c}\quad 0,\ [c \mapsto \pri]$$ for every channel $c$, while executing with more resources $$\new x. 0,\ [c \mapsto \pri]\quad \rstep{\nu d}\quad 0,\ [c \mapsto \pri] \uplus [d \mapsto \pri]$$ results in constrained allocation: the $\uplus$ here denotes disjoint union, meaning that $c \neq d$. The fact that $c$ was already allocated pruned one trace (preventing it from taking an impossible step), but introduced no new traces. Similarly, $$\new x. \ov{x}x. 0\quad \step{\nu c} \ov{c}c.0\quad \step{c!c} 0$$ but, taking resources into account, we have $$\new x. \ov{x}x. 0,\ \emptyset\quad \rstep{\nu c}\quad \ov{c}c.0,\ [c \mapsto \pri]$$ at which point the process is stuck: the action $c!c$ is prevented from occurring, because $\ASem{c!c}[c\mapsto \pri] = \bot$. This deadlock is exactly what we expect to see when a process attempts to communicate along a private channel. Finally, we have $$\new x. (\ov{x}x.0 | x(y).\ov{y}x.0) \quad\step{\nu c}\quad \ov{c}c.0 | c(y).\ov{y}c.0 \quad\step{\tau}\quad 0 | \ov{c}c.0 \quad\step{c!d}\quad 0 | 0$$ which, with resources, yields $$\new x. (\ov{x}x.0 | x(y).\ov{y}x.0),\ \emptyset \ \ \rstep{\nu c}\ \ \ov{c}c.0 | c(y).\ov{y}c.0,\ [c \mapsto \pri] \ \ \rstep{\tau}\ \ 0 | \ov{c}c.0,\ [c \mapsto \pri]$$ Here we see that *internal* communication along a private channel is both possible and permitted: such internal steps appear as $\tau$ actions to the resource-sensitive stepping relation, and hence always pass through. On the other hand, the internal communication also leaves the ownership of $c$ unchanged. Because it remains private, the final communication $\ov{c}c$ is stuck, as it should be. Process safety -------------- With the simple public/private resource model, faulting occurs only when using an unallocated channel. Our semantic framework can accommodate deallocation, but doing so complicates the full abstraction result, and we wish to focus on the standard $\pi$-calculus. Avoiding deallocation allows us to easily characterize “safe” processes: we say $\sigma \proves P\ok$ iff $P$ is closed and all channel constants in $P$ are in $\dom(\sigma)$, and have: If $\sigma \gives P \ok$ then $P,\sigma \stackrel{\fault}{\not\rightarrow}$, and if $P,\sigma \step{\alpha} P',\sigma'$ then $\sigma' \gives P' \ok$. Denotational semantics: safety traces {#sec:safety} ===================================== Resources provide an intriguing refactoring of the operational semantics for $\pi$-calculus, but their real payoff comes in the elementary denotational model they support. We begin with a simple trace model capturing only (some) safety properties, which allows us to focus on the role of resources. Afterwards we incorporate liveness () and its interaction with resources. For the safety model, we have traces $t$, trace sets $T$ and behaviors $B$: $$\begin{array}{c} \textsc{Trace} \ \eqdef\ \textsc{Action}^* \qquad \textsc{Beh} \ \eqdef\ \Sigma \ra \textsc{TraceSet} \\ \textsc{TraceSet} \ \eqdef\ \{ T\ :\ \emptyset \subset T \subseteq \textsc{Trace},\ T\ \textrm{prefix-closed} \} \end{array}$$ Processes will denote behaviors: sets of action traces determined by the initially-available resources. Not every action is observable. We follow standard treatments of $\pi$-calculus [@Sangiorgi2001; @Hennessy2002] in considering $\tau$ steps unobservable, and eliding $\nu c$ steps until just before the allocated channel $c$ is sent over a public channel (a “bound send”). Our denotational semantics shows that the operators of the $\pi$-calculus are congruent for these observables, and the cited works prove that similar observables are fully abstract for yet coarser notions of observation. The observables of an action $\alpha$ are a (possibly empty) trace, depending on the available resources: \[$|\alpha|_\sigma : \textsc{Trace}$\] [Action observables]{} $$\begin{array}{rcl} |\tau|_\sigma &\eqdef& \epsilon \\ |\nu c|_\sigma &\eqdef& \epsilon \end{array} \qquad \begin{array}{rcl} |\fault|_\sigma &\eqdef& \fault \\ |c?d|_\sigma &\eqdef& c?d \end{array} \qquad \begin{array}{rcl} |c!d|_\sigma &\eqdef& \begin{cases} \nu d \cdot c!d & \sigma(d) = \pri \\ c!d & \textrm{otherwise} \end{cases} \end{array}$$ We write $t \cdot u$ or $tu$ for trace concatenation, and $\epsilon$ for the empty trace. Although $\nu c$ is not immediately observable, taking a $\nu c$ step affects the resources owned by the process, so exposing $c$ later will cause the $\nu c$ step to reemerge. The behavior of a process can be read from its operational semantics: [Safety observation]{} $$\infer { } {\epsilon \in \Ob{P}{\sigma}} {} \qquad \infer {P,\sigma \rstep{\alpha} P', \sigma' \\ t \in \Ob{P'}{\sigma'}} {|\alpha|_\sigma t \in \Ob{P}{\sigma}} {}$$ The goal of the denotational semantics is to calculate the same traces compositionally over process structure. $\textsc{TraceSet}$ is a complete lattice under the subset order, and behaviors inherit this order structure pointwise: we write $B \refines B'$ if $B(\sigma) \subseteq B'(\sigma)$ for all $\sigma$ and have $(B \lub B')(\sigma) = B(\sigma) \cup B'(\sigma)$. The semantic operators are monotonic (in fact, continuous), so we are justified in defining as a fixpoint. For the safety semantics, which is based on finite observation, it is the least fixpoint. The safety trace model is insensitive to branching behavior of processes [@Glabbeek1988], so internal and external choice are indistinguishable. We interpret both forms of choice using $\lub$, merging behaviors from all the alternatives. For empty summations, $\lub$ yields the smallest behavior: $\lambda \sigma. \{ \epsilon \}$. The denotation function is parameterized by an environment $\rho$, here taking channel variables $x$ to channels $c$, and process variables $X$ to behaviors $B$. It uses two additional operators, $\pref$ and $\parallel$, which we will define shortly. \[$\SemB{P} : \textsc{Env} \ra \textsc{Beh}$\] [Denotational semantics (safety)]{} $$\begin{array}{r@{\ \ \eqdef \ \ }r@{\ \pref\ }l} \Sem{\ov{e}e'.P}{\rho} & \rho e ! \rho e' & \Sem{P}{\rho} \\ \Sem{e(x).P}{\rho} & \biglub_c \rho e ? c & \Sem{P}{\rho[x \mapsto c]} \\ \Sem{\new x.P}{\rho} & \biglub_c \nu c & \Sem{P}{\rho[x \mapsto c]} \\ \Sem{\rec X.P}{\rho} & \multicolumn{2}{r}{\mu B. \Sem{P}{\rho[X \mapsto B]}} \end{array} \qquad \begin{array}{r@{\ \ \eqdef\ \ }l} \Sem{\sum \pi_i.P_i}{\rho} & \biglub_i \Sem{\pi_i.P_i}{\rho} \\ \Sem{P \nc Q}{\rho} & \Sem{P}{\rho} \lub \Sem{Q}{\rho} \\ \Sem{P | Q}{\rho} & \Sem{P}{\rho} \parallel \Sem{Q}{\rho} \\ \Sem{X}{\rho} & \rho(X) \end{array}$$ The interpretation of prefixed processes resembles the operational semantics: each clause of the denotational semantics generates all locally-reasonable actions, without immediately checking global plausibility. We use $\lub$ to join the behaviors arising from each action—once more reflecting nondeterminism—and we update the environment as necessary. The operator $\alpha \pref B$ prefixes an action $\alpha$ to a behavior $B$ in a resource-sensitive way, playing a role akin to the second layer of the operational semantics: \[$\alpha \pref B : \textsc{Beh}$\] [Semantic prefixing]{} $$\begin{array}{r@{\ \ \eqdef\ \ }l} (\alpha \pref B)(\sigma) & \{ \alpha t\ : \ASem{\alpha}{\sigma} = \sigma',\ t \in B(\sigma') \} \ \cup\ \{ \fault\ :\ \ASem{\alpha}{\sigma} = \top \} \ \cup\ \{ \epsilon \} \end{array}$$ To maintain prefix-closure, we include $\epsilon$ as a possible trace. A quick example: $$\Sem{\new x. \ov{x}x.0}{\emptyset} \ =\ \biglub_c \nu c \pref \Sem{\ov{x}x.0}{x \mapsto c} \ =\ \biglub_c \nu c \pref c!c \pref \Sem{0}{x \mapsto c} \ =\ \biglub_c \nu c \pref c!c \pref \lambda \sigma. \{ \epsilon \}$$ This expansion of the definition resembles the traces we see from the first layer of the operational semantics, without taking resources into account. The denotation, recall, is a *behavior*: to extract its set of traces, we must apply it to some particular resource $\sigma$. If we use the empty resource, we see that $$\begin{aligned} \left(\biglub_c \nu c \pref c!c \pref \lambda \sigma. \{ \epsilon \}\right)(\emptyset) &=& \{ \epsilon \} \cup \bigcup_c \left\{ \nu c \cdot t\ :\ t \in \left(c!c\pref \lambda \sigma. \{\epsilon\}\right)[c\mapsto \pri] \right\} \\ &=& \{ \epsilon \} \cup \bigcup_c \left\{ \nu c \cdot t\ :\ t \in \{ \epsilon \} \right\} $$ in other words, we have $\Sem{\new x. \ov{x}x.0}{\emptyset}(\emptyset) = \{ \epsilon \} \cup \bigcup_c \{ \nu c \}$. Just as in the operational semantics, the fact that $\ASem{c!c}[c \mapsto \pri] = \bot$ prevents the $c!c$ step from being recorded. Here, the prefix closure (in particular, the inclusion of $\epsilon$ in every application of $\pref$) ensures that we see the trace up to the point that we attempt an impossible action. Finally, we have parallel composition—the most interesting semantic operator. Here we must ask a crucial question for the denotational semantics: if $\sigma$ is the resource belonging to $P|Q$, what resources do we provide to $P$ and $Q$? The question does not come up in the operational semantics, which maintains a single, global resource state, but a compositional semantics must answer it. Consider the process $\new x.(\ov{x}c\ |\ x(z))$. When the process reaches the parallel composition, $x$ will still be private. The privacy of $x$ means that the subprocesses can only communicate with each other (yielding $\tau$), not with the external environment of the process. But the subprocesses *are* communicating with environments external to themselves—namely, each other. That is, $x$ is private to $\ov{x}c\ |\ x(z)$, which cannot communicate along it externally, but it is *public* to the *subprocesses* $\ov{x}c$ and $x(z)$, which can. Formally, we capture this narrative as follows: \[$B_1 \parallel B_2 : \textsc{Beh}$\] [Semantic parallel composition]{} $$\begin{array}{r@{\ \eqdef\ }l} (B_1 \parallel B_2)(\sigma) & \bigcup_{t_i \in B_i(\widehat{\sigma})} (t_1 \tparallel t_2)(\sigma) \end{array} \ \textrm{where}\ \widehat{\sigma}(c)\ \eqdef\ \begin{cases} \pub & c \in \dom(\sigma) \\ \textrm{undefined} & \textrm{otherwise} \end{cases}$$ The resource $\sigma$ given to a parallel composition of behaviors is fed in *public-lifted* form ($\widehat{\sigma}$) to the composed behaviors, yielding two sets of traces. For each pair of traces $t_1$ and $t_2$ from these sets, we calculate all interleavings $t_1 \mbox{$\tparallel$} t_2$: \[$t \tparallel u : \textsc{Beh}$\] [Trace interleavings]{} $$\begin{array}{rcll} t \tparallel u &\eqdef& \lambda \sigma.\{\epsilon\} &\IF t = \epsilon = u \\&\lub& \alpha \pref (t' \tparallel u) &\IF t = \alpha t' \\&\lub& \alpha \pref (t \tparallel u') &\IF u = \alpha u' \\&\lub& t' \tparallel u' &\IF t = \alpha t',\ u = \ov{\alpha} u' \end{array}$$ Interleaving at first glance appears standard, but note the use of semantic prefixing $\pref$: *the interleavings are not simply another set of traces, they are given as a *behavior* that must be evaluated*. We evaluate with the *original* resources $\sigma$. The effect is that each interleaving is checked with respect to the resources held by the *combined* process. This additional check is the key to making the “declare everything public” approach work, allowing us to take into account channels that are private from the point of view of the combined process, but public between the subprocesses. An example helps illuminate the definitions: take the process $\ov{d}c\ |\ d(z)$ with resources $\sigma = [c \mapsto \pub][d \mapsto \pri]$. It is easy to calculate that $$\begin{array}{l} \Sem{\ov{d}c}{\emptyset}\!(\widehat{\sigma})\ =\ \{ \epsilon, d!c \} \qquad \Sem{d(z)}{\emptyset}\!(\widehat{\sigma})\ =\ \{ \epsilon \} \cup \{ d?e\ :\ e \in \textsc{Chan} \} \\ d!c \parallel d?c \ =\ \left(d!c \pref d?c \pref \lambda \sigma. \{\epsilon\}\right) \lub \left(d?c \pref d!c \pref \lambda \sigma. \{\epsilon\}\right) \lub \left(\lambda \sigma. \{\epsilon\}\right) \end{array}$$ The interleaving $d!c \parallel d?c$ includes the case that $d!c$ and $d?c$ are two sides of the same communication (yielding $\lambda \sigma. \{\epsilon\}$) and the two possible orderings if they are not. From the point of view of $\widehat{\sigma}$, which has lost the information that $d$ is private to the combined process, this is the most we can say. However, the interleaving is built using the prefixing operation $\pref$, so when we evaluate it with respect to the original $\sigma$, some traces will be silently dropped: $$\begin{aligned} && (d!c \parallel d?c)(\sigma)\\ &\ =\ & (d!c \pref d?c \pref \lambda \sigma. \{\epsilon\})(\sigma) \cup (d?c \pref d!c \pref \lambda \sigma. \{\epsilon\})(\sigma) \cup (\lambda \sigma. \{\epsilon\}) (\sigma) \\ &\ =\ & \{ \epsilon \} \cup \{ \epsilon \} \cup \{ \epsilon \}\end{aligned}$$ In particular, for any $B$ we have $ (d!c \pref B)(\sigma) = (d?c \pref B)(\sigma) = \{ \epsilon \} $ because $\sigma(d) = \pri$. We are left only with traces that could arise from internal communication, as expected. That is, $\Sem{\new x.(\ov{x}c|x(y))}{\emptyset}[c \mapsto\pub] = \{ \epsilon \}$. More generally, we can show $\Sem{\new x.(\ov{x}c|x(y))}{\emptyset}\sigma = \Sem{0}{\emptyset}\sigma$ whenever $c \in \dom(\sigma)$. Because $\ASem{\fault}{\sigma} = \top$, we have $\fault \pref B = \lambda \sigma. \{ \fault, \epsilon \}$ for any $B$. Thus, when a $\fault$ action is interleaved, the interleaving is terminated with that action. In summary, we calculate the traces of $P|Q$ by calculating the traces of $P$ and $Q$ under conservatively public-lifted resources, then evaluating the interleavings with complete information about what resources $P|Q$ actually owns. #### Example calculations Before proving full abstraction, we briefly examine a few of the expected laws. For example, why does $\SemB{\new x.0} = \SemB{0}$? Expanding the former, we get $\biglub_c \nu c \pref \lambda \sigma . \{ \epsilon \}$. When applied to a particular $\sigma$, this behavior yields the simple set $\{ \epsilon \}$, because $|\nu c|_\sigma = \epsilon$. This simple example sheds light on the importance of action observation $|-|$: it is crucial for ignoring when, or in some cases whether, channels are allocated. A more complex example is the following: $$\begin{aligned} \Sem{\new x.\new y.P}{\rho} &=& \biglub_c \nu c \pref \Sem{\new y.P}{\rho[x \mapsto c]} \\ &=& \biglub_c \nu c \pref \biglub_d \nu d \pref \Sem{P}{\rho[x \mapsto c, y \mapsto d]} \\ &=& \biglub_{c,d} \nu c \pref \nu d \pref \Sem{P}{\rho[x \mapsto c, y \mapsto d]} \\ &=& \biglub_{c,d} \nu d \pref \nu c \pref \Sem{P}{\rho[x \mapsto c, y \mapsto d]} \\ &=& \biglub_d \nu d \pref \biglub_c \nu c \pref \Sem{P}{\rho[x \mapsto c, y \mapsto d]} \\ &=& \biglub_d \nu d \pref \Sem{\new x.P}{\rho[y \mapsto d]} \ =\ \Sem{\new y.\new x.P}{\rho}\end{aligned}$$ The key step is swapping $\nu c$ and $\nu d$, which relies on the lemma $\nu c \pref \nu d \pref B = \nu d \pref \nu c \pref B$. The validity of this lemma, again, relies on observability: $|\nu c|_\sigma = |\nu d|_\sigma = \epsilon$ for all $\sigma$. Congruence for the basic operators ---------------------------------- We prove full abstraction by proving a *congruence* result for each operator in the language. For the operators other than parallel composition, we show: All of the following equivalences on closed processes hold: 1. $\ObB{0} = \Sem{0}{\emptyset}$ 2. $\ObB{\ov{c}d.P} = c!d \pref \ObB{P}$ 3. $\ObB{c(x).P} = \biglub_d c?d \pref \ObB{P\{d/x\}}$ 4. $\ObB{\new x.P} = \biglub_c \nu c \pref \ObB{P\{c/x\}}$ 5. $\ObB{\sum_i P_i} = \biglub_i \ObB{P_i}$ 6. $\ObB{P \nc Q} = \Ob{P} \lub \ObB{Q}$ These equivalences are straightforward to show; we prove each by showing containment in both directions. For illustration, we give the proof that $\ObB{c(x).P} \subseteq \biglub_d c?d \pref \ObB{P\{d/x\}}$: Let $\sigma \in \Sigma$ and $t \in \Ob{c(x).P}{\sigma}$. We analyze cases on the derivation of $t \in \Ob{c(x).P}{\sigma}$: Let $d$ be a channel. Then $t = \epsilon \in c?d \pref \ObB{P\{d/x\}}$ by definition of $\pref$. The result follows by monotonicity of $\lub$. Reasoning by inversion, we see that there are two subcases: Then $t = \alpha t' \in \biglub_d c?d \pref \Ob{P\{d/x\}}$ trivially by the definition of $\pref$. Then $t = \alpha t' = \fault$ because $\Ob{0}{\sigma'} = \{ \epsilon \}$. That $\fault \in \biglub_d c?d \pref \Ob{P\{d/x\}}$ again follows easily by the definition of $\pref$. Congruence for parallel composition ----------------------------------- The justification of our treatment of parallel composition goes back to the intuitions from the beginning of the paper: concurrent process must divide resources amongst themselves, with each process using only those resources it owns. We say $\sigma$ separates into $\sigma_1$ and $\sigma_2$ if the following conditions hold: [Parallel separation]{} $$\sigma \in (\sigma_1 \parallel \sigma_2) \ \eqdef\ \left\{ \begin{array}{l} \dom(\sigma) = \dom(\sigma_1) \cup \dom(\sigma_2) \\ \sigma_1(c) = \pri \implies \sigma(c)=\pri,\ c\notin\dom(\sigma_2) \\ \sigma_2(c) = \pri \implies \sigma(c)=\pri,\ c\notin\dom(\sigma_1) \end{array} \right.$$ We understand this definition as saying: if $\sigma_1$ and $\sigma_2$ are resources separately held by $P$ and $Q$ respectively, then $\sigma$ is *possibly* the resource held by $P|Q$. The subresources $\sigma_i$ do not uniquely determine a combination $\sigma$ because resources public to the subprocess may, or may not, be private to the combined process.[^3] Separation crisply captures the desired meaning of public and private ownership: if one subprocess owns a resource privately ($\sigma_1(c) = \pri$), then the other subprocess does not own the resource at all ($c \notin\dom(\sigma_2)$), but both processes may own a resource publicly. To show that that $\ObB{P_1|P_2} = \ObB{P_1} \mbox{$\parallel$} \ObB{P_2}$, we must show that our strategy of interleaving traces from publicly-lifted resources agrees with the global operational semantics. A key idea is that $\sigma \in \sigma_1 \parallel \sigma_2$ constitutes an invariant relationship between the resources owned by subprocesses (in the denotational semantics) and those owned by the composite process (in the operational semantics). The invariant holds initially because $\sigma \in \widehat{\sigma} \parallel \widehat{\sigma}$. The unobservability of $\nu c$ steps complicates matters somewhat: it means there is an additional perspective on resources—call it $\sigma_{\rm den}$—owned by a composite process. Generally, $\sigma_{\rm den}$ underestimates the true resources $\sigma$ of the operational semantics. Consider the denotational interleaving of two traces $t_1$ and $t_2$ from subprocesses $P_1$ and $P_2$ respectively. If $P_1$ allocates a channel, that allocation does not appear immediately in $t_1$, and hence does not appear immediately in the resources $\sigma_{\rm den}$ of the interleaving, while it would appear in $\sigma$ operationally. During denotational interleaving, the same channel can even be owned privately in *both* $\sigma_1$ and $\sigma_2$. The key observation here is that either both subprocesses eventually reveal a given private channel—in which case the denotational interleaving is filtered out—or at least one subprocess does not—in which case its choice of channel is irrelevant. Altogether, the four resources—$\sigma_{\textrm{op}}$, $\sigma_{\textrm{den}}$, $\sigma_1$, and $\sigma_2$—are always related: $$\CI(\sigma_{\textrm{op}}, \sigma_{\textrm{den}}, \sigma_1, \sigma_2)\ \eqdef\ \sigma_{\textrm{op}} \in \sigma_1\parallel\sigma_2,\ \sigma_{\rm den} = \sigma_{\textrm{op}} \setminus \{ c\ :\ \sigma_1(c) = \pri \vee \sigma_2(c) = \pri \}$$ Validating parallel composition requires another important lemma, *locality* from abstract separation logic [@Calcagno2007].[^4] If $\sigma \in \sigma_1 \parallel \sigma_2$ then - if $\ASem{\alpha}{\sigma} = \top$ then $\ASem{\alpha}{\sigma_1} = \top$, and - if $\ASem{\alpha}{\sigma} = \sigma'$ then $\ASem{\alpha}{\sigma_1} = \top$ or $\ASem{\alpha}{\sigma_1} = \sigma'_1$ for some $\sigma'_1$ with $\sigma' \in \sigma'_1 \parallel \sigma_2$. The lemma characterizes the transformations an action can make given some composite resources $\sigma$ in terms of its behavior on subresources $\sigma_1$. Providing additional resources can never introduce new faults, and if the action does not fault given just $\sigma_1$ resources, then the changes it makes to $\sigma$ must only change the $\sigma_1$ portion (framing). Locality was introduced to characterize the frame rule of separation logic [@Calcagno2007], but we use it here to characterize interleaving steps in parallel composition. We have a related lemma for internal communication steps: If $\sigma \in \sigma_1 \parallel \sigma_2$, $\ASem{\alpha}{\sigma_1} = \sigma'_1$ and $\ASem{\ov{\alpha}}{\sigma_2} = \sigma'_2$ then $\sigma \in \sigma'_1 \parallel \sigma'_2$. We prove each direction of congruence separately: If $\CI(\sigma_{\textrm{op}}, \sigma_{\textrm{den}}, \sigma_1, \sigma_2)$, $\sigma_i \proves P_i \ok$ and $t \in \Ob{P_1|P_2}{\sigma_{\textrm{op}}}$ then\ $t \in (t_1\parallel t_2)(\sigma_{\textrm{den}})$ for some $t_i \in \Ob{P_i}{\sigma_i}$. If $\CI(\sigma_{\textrm{op}}, \sigma_{\textrm{den}}, \sigma_1, \sigma_2)$, $\sigma_i \proves P_i \ok$, $t_i \in \Ob{P_i}{\sigma_i}$, and\ $t \in (t_1\parallel t_2)(\sigma_{\textrm{den}})$ then $t \in \Ob{P_1|P_2}{\sigma_{\textrm{op}}}$. The first of these two lemmas is easier to prove, because we are given a trace $t$ derived from the operational semantics of the composite processes. This means that the subprocesses are guaranteed not to independently allocate the same channel. The second lemma requires more care, using the insights mentioned above about renaming unexposed channels. The assumptions $\sigma_i \proves P_i \ok$ are needed to ensure that the processes we are working with do not fault. The reason that faulting is problematic is seen in the following example: $$\begin{aligned} && \new x.\ov{c}x.0\ |\ c(y).\ov{c}y.\ov{d}y.0),\ [c \mapsto \pub] \\ &\rstep{\nu d}\ \ & \ov{c}d.0\ |\ c(y).\ov{c}y.\ov{d}y.0,\ [c \mapsto \pub, d \mapsto \pri] \\ &\rstep{\tau}\ \ & 0\ |\ \ov{c}d.\ov{d}c.0,\ [c \mapsto \pub, d\mapsto\pri] \\ &\rstep{c!d}\ \ & 0\ |\ \ov{d}c.0,\ [c \mapsto \pub, d\mapsto\pub] \\ &\rstep{d!c}\ \ & 0\ |\ 0,\ [c \mapsto \pub, d\mapsto\pub] \end{aligned}$$ The uncomfortable aspect of this derivation is that the channel $d$ occurred in the process initially, even though it was not owned. As a result, the process was able to *allocate* $d$, in a sense falsely capturing the constant $d$ that initially appeared. In cases where the process allocates a different channel than $d$, it will fault when it attempts to communicate along the constant channel $d$. But in this “lucky” case, the operational semantics allows communication along the constant channel. The denotational semantics, however, *always* generates a fault. It computes the traces compositionally, meaning that a channel $d$ allocated by one subprocess is not immediately available for use by a parallel subprocess. Our full abstraction result applies only to nonfaulty processes, which, fortunately, is a trivial syntactic check. However, this does limit its applicability to languages that include features like deallocation, which makes checking for safety more difficult. Full abstraction ---------------- To complete the proof of full abstraction, we must deal with recursion. We begin with the usual unwinding lemma, proved in the standard syntactic way: We have $\ObB{\rec X.P} = \biglub_n \Ob{\recB_n X.P}$, where $\recB_0 X.P \eqdef \rec X. X$ and $\recB_{n+1} X.P \eqdef P\{\recB_n X.P/X\}$. We also have the standard substitution lemmas: We have $\Sem{P[Q/X]}{\rho} = \Sem{P}{\rho[X \mapsto Q]}$ and\ $\Sem{P[c/x]}{\rho} = \Sem{P}{\rho[x \mapsto c]}$. Combined these lemmas with the previous congruence results, it is straightforward to show the following theorem relating the observed operational traces to those calculated denotationally: If $P$ closed and $\sigma \gives P \ok$ then $\Ob{P}{\sigma} = \Sem{P}{\emptyset}\sigma$. To prove this theorem, we must generalize it to deal with open terms. We do this by introducing a *syntactic environment* $\eta$ as a finite map taking channel variables to channels and process variables to closed processes. Given a syntactic environment $\eta$ the corresponding semantic environment $\widehat{\eta}$ is given by: $$(\widehat{\eta})(x)\ \eqdef\ \eta(x) \qquad (\widehat{\eta})(X)\ \eqdef\ \ObB{\eta(X)}$$ We write $\eta P$ for the application of $\eta$ as a syntactic substitution on $P$. The needed induction hypothesis for congruence is then if $\sigma \gives \eta P \ok$ then $\Ob{\eta P}{\sigma} = \Sem{P}{\widehat{\eta}}\sigma$. Define $P =_{\textsc{Den}} Q$ iff $\Sem{P}{\rho}\sigma = \Sem{Q}{\rho}\sigma$ for all $\sigma$ such that $\sigma \gives P\ok$ and $\sigma \gives Q\ok$. Likewise, let $P =_{\textsc{Op}} Q$ iff $\ObB{C[P]}\sigma = \ObB{C[Q]}\sigma$ for all contexts $C$ with $\sigma \gives C[P]\ok$ and $\sigma \gives C[Q]\ok$. Full abstraction follows by compositionality: $P =_{\textsc{Den}} Q$ iff $P =_{\textsc{Op}} Q$. Denotational semantics: adding liveness {#sec:liveness} ======================================= To round out our study of $\pi$-calculus, we must account for liveness properties. Liveness in process algebra appears under diverse guises, differing in sensitivity to branching behavior and divergence [@Glabbeek1988]. Each account of liveness corresponds to some choice of basic observable: given a process $P$ and a context $C$, what behavior of $C[P]$ matters? The standard observable for the $\pi$-calculus is barbed bisimilarity [@barbed], which sits quite far on the branching side of the linear-branching time spectrum [@Glabbeek1988]. Here, we choose a treatment more in the spirit of linear time: an adaptation of acceptance traces [@Hennessy2002]. This choice is partly a matter of taste, but it also allows us to stick with a purely trace-theoretic semantics, which keeps the domain theory to a minimum. We do not see any immediate obstacles to applying our resource-based handling of names to a branching-time semantics. Branching sensitivity and resource-sensitivity seem largely orthogonal, though of course branches may be pruned when deemed impossible given the owned resources. Liveness observables -------------------- We say that a process *diverges* if it *can* perform an infinite sequence of unobservable (*i.e.*, internal) steps without any intervening interactions with its environment—which is to say, the process can livelock. On the other hand, a process that can make *no* further unobservable steps is blocked (waiting for interaction from its environment). The basic observables in our model are: - A finite sequence of interactions, after which the process diverges or faults; - A finite sequence of interactions, after which the process is blocked, along with which channels it is blocked on; and - An infinite sequence of interactions. Notice that we have conflated divergence and faulting: we view both as erroneous behavior. In particular, we view any processes that are capable of immediately diverging or faulting as equivalent, regardless of their other potential behavior. This perspective is reasonable—meaning that it yields a congruence—because such behavior is effectively uncontrollable. For example, if $P$ can immediately diverge, so can $P|Q$ for any $Q$. Formally, we add a new action $\delta_\Delta$ which records that a process is blocked attempting communication along the finite set of *directions* $\Delta$: $$\alpha\ ::=\ \cdots \GA \delta_\Delta \qquad \Delta \subseteq_\textrm{fin} \textsc{Dir} \eqdef \{ c!\ :\ c\in\textsc{Chan} \} \cup \{ c?\ :\ c\in\textsc{Chan} \}$$ We then define $$\textsc{LTrace}\ \eqdef\ \textsc{NTAction}^*;\{ \fault, \delta_\Delta\}\ \cup\ \textsc{NTAction}^\omega \qquad \textsc{LBeh}\ \eqdef\ \Sigma \rightarrow 2^{\textsc{LTrace}}$$ where <span style="font-variant:small-caps;">NTAction</span> (for “non-terminating action”) refers to all actions except for $\fault$ or blocking actions $\delta_\Delta$. Thus finite liveness traces must end with either a $\delta_\Delta$ action or a $\fault$ action, whereas neither of these actions can appear in an infinite trace. Each liveness trace encompasses a *complete* behavior of the process: either the process continues interacting indefinitely, yielding an infinite trace, or diverges, faults or gets stuck after a finite sequence of interactions. Therefore, sets of liveness traces are not prefixed-closed. As with the safety traces, we can observe liveness traces from the operational semantics. However, we do so using the *greatest* fixpoint of the following rules: [Liveness observation]{} $$\begin{array}{c} \infer[{}] {P,\sigma \rstep{\alpha }P',\sigma' \\\\ \alpha \neq \fault \\ t \in \LOb{P'}{\sigma'}} {|\alpha|_\sigma t \in \LOb{P}{\sigma}} \textrm{gfp} \qquad \infer[{}] {P,\sigma \rstep{\fault}} {\fault \in \LOb{P}{\sigma}} \textrm{gfp} \qquad \infer[{}] {P,\sigma \blocked \Delta} {\delta_\Delta \in \LOb{P}{\sigma}} \textrm{gfp} \end{array}$$ where $P, \sigma \blocked \Delta$ means that $P, \sigma$ can only take communication steps, and $\Delta$ contains precisely the directions of available communication. Since the owned resources influence which communications are possible, they also influence the directions on which a process is blocked: $$\delta_{\{c!\}} \in \LOb{\ov{c}c.0}{[c \mapsto \pub]} \qquad \delta_\emptyset \in \LOb{\ov{c}c.0}{[c \mapsto \pri]}$$ The action $\delta_\emptyset$ reflects a completely deadlocked process, and is for example the sole trace of the inert process $0$. Defining the observations via a greatest fixpoint allows for infinite traces to be observed, but also means that if a process diverges after a trace $t$, its behavior will contain all traces $tu$, in particular $t\fault$. For example, suppose $P,\sigma \rstep{\tau} P,\sigma$. If $t$ is any liveness trace whatsoever, we can use the first inference rule to show, coinductively, that $t \in \LOb{P}{\sigma}$. We merely assume that $t \in \LOb{P}{\sigma}$, and derive that $|\tau|_\sigma t = t \in \LOb{P}{\sigma}$. Thus, divergence is “catastrophic” (as in failures/divergences [@Brookes1984]). An important step toward making these observables coherent is the notion of *refinement*. In general, saying that $P$ refines $Q$ (or $P$ “implements” $Q$) is to say that every behavior of $P$ is a possible behavior of $Q$. In other words, $P$ is a more deterministic version of $Q$. We define a refinement order on traces: $$t \refines t \qquad t\delta_\Delta \refines t\delta_{\Delta'}\ \textrm{if}\ \Delta' \subseteq \Delta \qquad tu \refines t\fault$$ which we lift to sets of traces as: $T \refines U$ iff $\forall t \in T.\ \exists u\in U.\ t \refines u$. This notion of refinement, which closely follows that of acceptance traces [@Hennessy2002], says that an implementation must allow at least the external choices that its specification does. It also treats faulting as the most permissive specification: if $Q$ faults, then any $P$ will refine $Q$. Moreover, any two immediately-faulting processes are equivalent. Since faulting and divergence are treated identically, the same holds for divergent processes. Thus, the simple refinement ordering on traces has an effect quite similar to the closure conditions imposed in failures/divergences semantics. The ordering on trace sets inherits the complete lattice structure of $2^{\textsc{LTrace}}$, as does the pointwise order on <span style="font-variant:small-caps;">LBeh</span>. We again exploit this fact when interpreting recursion. Liveness semantics ------------------ To complete the semantic story, we need to interpret blocking actions. We define $$\begin{aligned} \ASem{\delta_\Delta}{\sigma} &\eqdef& \begin{cases} \top & \exists c.\ (c! \in \Delta \vee c? \in \Delta) \wedge c \notin\dom(\sigma) \\ \sigma & \ow \end{cases} \\ |\delta_\Delta|_\sigma &\eqdef& \delta_{\Delta'} \ \textrm{where}\ \Delta' = \Delta\upharpoonright \{c\ :\ \sigma(c) = \pub \}\end{aligned}$$ which shows the interaction between resources and blocking: blocking on a private resource is possible, but unobservable (*cf.* projection on $\delta$ in [@Brookes2002]). For example, we have $$\begin{array}{c@{\qquad}c} \ASem{\delta_{\{c!\}}}{[c \mapsto \pub]} = [c \mapsto \pub] & |\delta_{\{c!\}}|_{[c \mapsto \pub]} = \delta_{\{c!\}} \\ \ASem{\delta_{\{c!\}}}{[c \mapsto \pri]} = [c \mapsto \pri] & |\delta_{\{c!\}}|_{[c \mapsto \pri]} = \delta_\emptyset \end{array}$$ The denotational semantics for liveness, $\LSemB{-}$, is largely the same as that for safety, except for the following clauses: $$\begin{aligned} \LSem{\rec X.P}{\rho} &\eqdef&\ \nu B. \LSem{P}{\rho[X \mapsto B]} \\ \LSem{\sum \pi_i. P_i}{\rho} &\eqdef& \left(\biglub \LSem{\pi_i. P_i}{\rho} \right) \lub \left(\delta_{\{\dir(\rho\pi_i)\}} \pref \lambda \sigma.\emptyset \right)\end{aligned}$$ Recursion is given by a greatest fixpoint, as expected. A summation of prefixed actions now generates a corresponding blocking set, recording the external choice (where dir extracts the direction of a prefix). The blocking action is “executed” using the prefixing operator $\pref$ so that the actual observed action corresponds to the available resources, as in the example above. Finally, we use the following definition of interleaving: $$\begin{array}{rcll} t \tparallel u &\eqdef_{\textrm{gfp}}& \alpha \pref (t' \tparallel u) &\IF t = \alpha t',\ \alpha \ \textrm{not blocking} \\&\lub& \alpha \pref (t \tparallel u') &\IF u = \alpha u',\ \alpha \ \textrm{not blocking} \\&\lub& \delta_{\Delta \cup \Delta'} &\IF t = \delta_\Delta,\ u = \delta_{\Delta'},\ \ov{\Delta} \pitchfork \Delta' \\&\lub& t' \tparallel u' &\IF t = \alpha t',\ u = \ov{\alpha} u' \end{array}$$ Liveness interleaving is given by a greatest fixpoint. An infinite sequence of internal communications (operationally, an infinite sequence of $\tau$ moves) therefore yields *all* possible traces, including faulting ones, as it should. An interleaved trace is blocked only when both underlying traces are, and only when they do not block in opposite directions ($\ov{\Delta}$ is $\Delta$ with directions reversed, and $\pitchfork$ denotes empty intersection). If two processes are blocked in opposite directions, then their parallel composition is in fact *not* blocked, since they are willing to communicate with each other (*cf* stability [@Brookes1984]). Full abstraction ---------------- The proof of full abstraction is structured similarly to the proof for the safety semantics. Congruence proofs must take into account blocking actions, which is straightforward in all cases except for parallel composition. There, we require a lemma: Suppose $\CI(\sigma_{\textrm{op}}, \sigma_{\textrm{den}}, \sigma_1, \sigma_2)$. Then - If $\delta_{\Delta_i} \in \LOb{P_i}{\sigma_i}$ and $\Delta_1 \pitchfork \ov{\Delta_2}$ then $|\delta_{\Delta_1 \cup \Delta_2}|_{\sigma_{\textrm{den}}} \in \LOb{P_1|P_2}{\sigma_{\textrm{op}}}$. - If $\delta_\Delta \in \LOb{P_1|P_2}{\sigma_{\textrm{op}}}$ then $\delta_{\Delta_i} \in \LOb{P_i}{\sigma_i}$ for some $\Delta_1$, $\Delta_2$ with $\Delta_1 \pitchfork \ov{\Delta_2}$ and $|\delta_{\Delta_1 \cup \Delta_2}|_{\sigma_{\textrm{den}}} = \delta_\Delta$. Defining $=_{\textsc{LDen}}$ and $=_{\textsc{LOp}}$ analogously to the safety semantics, we again have full abstraction: $P =_{\textsc{LDen}} Q$ iff $P =_{\textsc{LOp}} Q$. Logic {#sec:logic} ===== We now sketch a logic for reasoning about the safety semantics of processes. The logic proves *refinement* between open processes—denotationally, trace containment; operationally, contextual approximation. The refinements are qualified by assertions about owned resources, which is what makes the logic interesting. The basic judgment of the logic is $\Gamma \gives p \asm P \refines Q$, which says the traces of $P$ are traces of $Q$, as long as the initial resources and environment, respectively, satisfy assertions $p$ and $\Gamma$ (defined below). Resource assertions $p$ are as follows: $$p\ ::=\ \true \GA \false \GA p \wedge q \GA p \vee q \GA p * q \GA x\Pub \GA x\Pri \GA x=y \GA x \neq y$$ and we let $x\Known \eqdef x\Pub \vee x\Pri$. Satisfaction of assertions depends on both the environment and resources, as in these illustrative cases: $$\begin{array}{lcl} \rho, \sigma \models x \Pub &\eqdef& \sigma(\rho(x)) = \pub \\ \rho, \sigma \models p_1 * p_2 &\eqdef& \exists \sigma_1,\sigma_2. \sigma = \sigma_1 \uplus \sigma_2 \textrm{ and } \rho, \sigma_i \models p_i \end{array}$$ Resource assertions like $x\Pub$ are intuitionistic [@Reynolds2002]; without deallocation there is no reason to use the classical reading, which can assert nonownership. We are using the standard interpretation of separation logic’s $*$ as disjoint separation to enable *sequential* reasoning about resource transformers in our logic. Action interpretations $\ASemB{\alpha}$ are local with respect to $*$, just as they were for $\parallel$. Environment assertions $\Gamma$ constrain process variables: $$\begin{array}{c} \Gamma\ ::=\ \emptyset \GA \Gamma,\ (p \asm X \refines P) \\ \rho \models (p \asm X \refines P) \ \ \eqdef\ \ \forall \sigma.\ (\rho, \sigma \models p) \implies \rho(X)(\sigma) \subseteq \Sem{P}{\rho}{\sigma} \end{array}$$ The definition of entailment is thus: $$\Gamma \models p \asm P \refines Q\ \ \eqdef\ \ \forall \rho, \sigma.\ (\rho \models \Gamma\ \wedge\ \rho, \sigma \models p) \implies \Sem{P}{\rho}\sigma \subseteq \Sem{Q}{\rho}\sigma$$ By qualifying refinements by resource assertions we can incorporate Hoare logic-like reasoning. Take, for example, the rule $$\infer {\Gamma \gives p * (x\Pub \wedge y\Pub) \asm P \refines Q} {\Gamma \gives p * (x\Pub \wedge y\Known) \asm \ov{x}y.P \refines \ov{x}y.Q}$$ for sending over a public channel. It is a kind of congruence rule, but we shift resource assumptions for the subprocesses, corresponding to the Hoare triple $$\{ p * (x\Pub \wedge y\Known) \}\ \ov{x}y\ \{ p * (x\Pub \wedge y\Pub) \}$$ The syntactic structure of prefixes (rather than sequential composition) prevents a clean formulation of the logic using Hoare triples. This is why the frame $p$ is included, rather than added via a separate frame rule; we are using “large” rather than “small” axioms [@OHearn2001]. A better treatment is possible if we semantically interpret prefixing as sequential composition, which requires a variables-as-resources model [@Parkinson]. For sending over a private channel, we have an axiom: $\ov{x}y.P$ refines *any* process when $x$ is private, because $\ov{x}y.P$ is stuck. The corresponding Hoare triple is $\{ x\Pri \wedge y\Known\} \ \ov{x}y\ \{\false\}$. Here is a fragment of the logic, focusing on resource-sensitive rules: [A selection of logical rules for safety behavior]{} $$\infer {\Gamma \gives p * (x\Pub \wedge y\Pub) \asm P \refines Q} {\Gamma \gives p * (x\Pub \wedge y\Known) \asm \ov{x}y.P \refines \ov{x}y.Q} \quad \infer { } {\Gamma\gives x\Pri \wedge y\Known \asm \ov{x}y.P \refines Q}$$ $$\infer {\Gamma \gives (p * x\Pub) \wedge y\Pub \asm P \refines Q \\ y \notin \fv(p, \Gamma)} {\Gamma \gives p * x\Pub \asm x(y).P \refines x(y).Q} \quad \infer { } {\Gamma \gives x\Pri \asm x(y).P \refines Q}$$ $$\infer {\Gamma \gives p * x\Pri \asm P \refines Q \\ x \notin \fv(p, \Gamma)} {\Gamma \gives p \asm \new x.P \refines \new x.Q} \qquad $$ $$\infer {p \asm X \refines P \in \Gamma } {\Gamma \gives p \asm X \refines P} \qquad \infer {\Gamma, p \asm X \refines Q \gives p \asm P \refines Q} {\Gamma \gives p \asm \rec X. P \refines Q} \qquad \infer {p \models p' \\ \Gamma \gives p' \asm P \refines Q} {\Gamma \gives p \asm P \refines Q}$$ The congruence rule for parallel composition performs public-lifting $\widehat{p}$ on resource assertions (by replacing $\pri$ by $\pub$ in the assertion). Fixpoint induction is resource-qualified as well. We reason about the body $P$ of a recursive definition $\rec X.P$ using a hypothetical bound on $X$ as the induction hypothesis. That hypothesis, however, is only applicable under the *same* resource assumptions $p$ that were present when it was introduced—making $p$ the loop invariant. In addition to these resource-sensitive rules, we have the usual laws of process algebra, including the expansion law. Combining those laws with the ones we have shown, we can derive an *interference-free* expansion law, as in this simplified version: $ \Gamma \gives x\Pri \wedge y\Known \asm \ov{x}y.P | x(z).Q \equiv P | Q\{y/z\} $. Discussion ========== Future work: richer resources {#sec:resources} ----------------------------- Our resource model captures exactly the guarantees provided by the $\pi$-calculus: until a channel is exposed, it is unavailable to the environment; afterwards, all bets are off. This property is reflected in the fact that $\Sigma$ is not a separation algebra, since $c\ \pub \parallel c\ \pub$ can result in $c\ \pub$ or $c\ \pri$. No amount of public ownership adds up definitively to private ownership. Rather than using resources to model the guarantees of a language, we can instead use them to enforce guarantees we intend of programs, putting ownership “in the eye of the asserter” [@O'Hearn2007]. We can then recover privacy just as Boyland showed [@Boyland2003] how to recover write permissions from read permissions: via a fractional model of ownership: $ \Sigma_{\textsc{Frac}} \eqdef \textsc{Chan} \ra [0, 1] $. Unlike traditional fractional permissions, owning a proper fraction of a channel does not limit what can be done with the channel—instead, it means that the environment is *also* allowed to communicate on the channel. The fractional model yields a separation algebra, using (bounded) summation for resource addition. An easy extension is distinguishing send and receive permissions, so that interference can be ruled out in a direction-specific way. One can also imagine encoding a session-type discipline [@Honda1998] as a kind of resource: $\Sigma_\textsc{Sess} \eqdef \textsc{Chan} \rightharpoonup \textsc{Session}$ where $$s \in \textsc{Session}\ ::=\ \ell.s \oplus \ell.s \GA \ell.s\ \&\ \ell.s \GA !.s \GA ?.s \GA \textsf{end}$$ Separation of session resources corresponds to matching up dual sessions, and actions work by consuming the appropriate part of the session. Ultimately, such resource models could yield rely-guarantee reasoning for the $\pi$-calculus, borrowing ideas from deny-guarantee [@Dodds2009]. A challenge for using these models is managing the ownership protocol in a logic: how are resources consistently attached to channels, and how are resources split when reasoning about parallel composition? We are far from a complete story, but believe our semantics and logic can serve as a foundation for work in this direction. Related work {#sec:related} ------------ Hoare and O’Hearn’s work [@Hoare2008] introduced the idea of connecting the model theory of separation logic with the $\pi$-calculus, and provided the impetus for the work presented here. Their work stopped short of the full $\pi$-calculus, modelling only point-to-point communication and only safety properties. Our liveness semantics, full abstraction results, and refinement calculus fill out the rest of the story, and they all rely on our new resource model. In addition, our semantics has clearer connections to both Brookes’s action trace model [@Brookes2002] and abstract separation logic [@Calcagno2007]. Previous fully abstract models of the $\pi$-calculus are based on functor categories [@Stark2002; @Hennessy2002; @Fiore2002], faithfully capturing the traditional role of scope for privacy in the $\pi$-calculus. Those models exploit general, abstract accounts of recursion, nondeterminism, names and scoping in a category-theoretic setting. We have similarly sought connections with a general framework, but have chosen resources, separation and locality as our foundation. An immediate question is: why do we get away with so much less mathematical scaffolding? This question is particularly pertinent in the comparison with Hennessy’s work [@Hennessy2002], which uses a very similar notion of observation. Hennessy’s full abstraction result is proved by extracting, from his functor-categorical semantics, a set of acceptance traces, and showing that this extraction is injective and order preserving. The force of this “internal full abstraction” is that the functor-categorical meaning of processes is completely determined by the corresponding acceptance traces. But note, these traces are *not* given directly via a compositional semantics: they are extracted only after the compositional, functor-categorical semantics has been applied. What we have shown, in a sense, is that something like acceptance traces for a process can be calculated directly, and compositionally, from process syntax. Beyond providing a new perspective on the $\pi$-calculus, we believe the resource-oriented approach will yield new reasoning techniques, as argued above. We have also emphasized concreteness, giving an elementary model theory based on sets of traces. Finally, it is worth noting that substructural type systems have been used to derive strong properties (like confluence) in the $\pi$-calculus [@Kobayashi1999], just as we derived interference-free expansion. Here, we have used a resource theory to explain the $\pi$-calculus as it is, rather than to enforce additional discipline. But the ideas of  take us very much into the territory of discipline enforcement. More work is needed to see what that territory looks like for the resource-based approach. #### Acknowledgements We are grateful to Paul Stansifer and Tony Garnock-Jones for feedback on drafts of this paper, and to the anonymous reviewers who provided guidance on presentation. [^1]: Such a reading of resources has already appeared in *e.g.* deny-guarantee reasoning[@Dodds2009]. [^2]: The notation $\Sigma^\top_\bot$ denotes the set $\{\Sigma, \top, \bot\}$ and implies an ordering $\bot \leq \sigma \leq \top$ for all $\sigma \in \Sigma$. The order structure follows abstract separation logic [@Calcagno2007], and is related to locality (). [^3]: This means that $\Sigma$ with $\parallel$ does not form a separation algebra [@Calcagno2007]; see . [^4]: For simplicity we avoid the order-theoretic definition here, which requires lifting some of our constructions to $2^\Sigma$ in a way that is not otherwise useful.

ArXiv

The Standard Model (SM), despite its enormous success, is believed to be the low-energy limit of a more fundamental theory whose nature will be probed by the next generation of colliders. New physics effects have been studied in a variety of processes using model independent approaches [@leff], as well as within specific models [@spec.models]. All such investigations aim at providing a clear unambiguous signal for non SM physics effects. In this letter we propose one such signal obtained through simple $b$-jet counting. The approach is best suited to lepton colliders but may be extended to hadron colliders, e.g., it applies to the Fermilab Tevatron $p \bar p$ collider to the extent that the sea $b$-quark content of the protons can be ignored. Consider the inclusive multiple $b$-jet production in $\ell^+\ell^-$ collisions $$\begin{aligned} \ell^+ \ell^- \to n \cdot b +X \label{nbx}~,\end{aligned}$$ where $n$ denotes the number of $b$ and $\bar b$-jets in the final state (FS) and $X$ stands for non $b$-jets, leptons and/or missing energy; it is understood that this represents the state after top-quark decay. For the reactions (\[nbx\]) we introduce a useful approximate symmetry we call $b$-Parity ($b_P$), defined as $$\begin{aligned} b_P=(-1)^n ~.\end{aligned}$$ In the limit where the quarks mixing CKM matrix $V$ [@ckm] satisfies $ V_{3j}=V_{j3}=0$ for $ j \not=3 $, all SM processes are $b_P$-even since in this case the third generation quarks do not mix with the others, and this leads to the conservation of the corresponding flavor number. Given the fast top decay and since ${\rm Br}( t \to b W) \simeq 1 $, the experimentally-observed flavor number is in fact carried only by the $b$-quarks. Therefore, the measured quantum number reduces to the net number of detected $b$-quarks; we find it convenient to use instead the derived quantity $ b_P $. The only SM processes that violate this conserved number necessarily involve the charged current interactions and are, therefore, suppressed by the corresponding small off-diagonal CKM elements $|V_{cb}|^2$, $|V_{ub}|^2$, $|V_{ts}|^2$ or $|V_{td}|^2$. As a consequence the irreducible SM background to $b_P=-1$ processes (induced by new flavor physics beyond the SM) is strongly suppressed: [*the SM is essentially $b_P$-even*]{}. In the following we will look for experimental signatures of $ b_P$-odd physics within multi-jet events. We will assume that a sample with a definite number of jets has been selected (we will use 2 and 4 jet samples) and determine (within each sample) the experimental sensitivity needed to detect – or rule out – new flavor physics of this type up to a certain scale. In contrast with other observables the determination of $b_P$ within a sample with a fixed number of jets relies primarily on the $b$-tagging efficiency and purity of the sample used and not on the particular structure of a given FS, nor does it require the identification of any other particle but the $b$. Thus, the main obstacle in this use of $b_P$ is the reducible SM background, due to jet mis-identification. This results from having a $b$-tagging efficiency $ \epsilon_b$ below 1, and/or having non-zero probabilities $t_c$ and $ t_j $ of misidentifying a $c$ or light jets for a $b$-jet, respectively. This type of background would of course disappear as $ \epsilon_b \to 1 $ and $t_{c,j} \to 0$, but even for the small value $t_c =0.1$ and high $b$-tagging efficiency $ \epsilon_b= 0.7$, can produce a significant number of (miss-identified) events in the detector. Since for most experiments $ t_j $ is very small [@opal.example], the only relevant experimental parameters for this probe of new physics are $\epsilon_b$ and $t_c$. Consider now the inclusive $b$ and $\bar b$-jet production process in (\[nbx\]).[^1] Focusing only on multi-jet FS, let $\sigma_{n m \ell}$ be the cross-section for $$e^+e^- \to n \cdot b + m \cdot c + \ell \cdot j \label{nml} ~,$$ where $j$ is a light-quark or gluon jet and $c$ is a $c$-quark jet. Since our method does not require the detection of the charge of the $b$, $n$ is the number of $b$ + $\bar b$-quarks and similarly $m$ and $\ell$ are the number of the corresponding jets in (\[nml\]) irrespective of the parent quarks charges. We denote by $t_c$ the $c$-jet mis-tagging probability (i.e., that of mistaking a $c$-jet for a $b$-jet) and by $t_j$ the light-jet mis-tagging probability (i.e., that of mistaking a light-jet or gluon-jet for a $b$-jet). Using these, the probability (or cross-section) for detecting precisely $k$ $b$-jets in the reaction (\[nml\]) is given by $$\begin{aligned} \bar\sigma_k = \sum_{u,v,w} && P_u^n P_v^m P_w^\ell \left[ \epsilon_b^u (1-\epsilon_b)^{n-u} \right] \left[ t_c^v (1-t_c)^{m-v} \right] \nonumber \\ &&\left[ t_j^w (1-t_j)^{\ell - w} \right] \sigma_{ n m \ell} \delta_{u+v+w,k} \label{sigmabar} ~,\end{aligned}$$ where $P^i_j = i!/j!/(i-j)!$. To experimentally detect $b_P$-odd signals generated by new physics one should simply measure the number of events with an odd number of $b$-jets in the FS. In particular, for the reaction (\[nml\]), we define $N_{k,J}$ to be the number of events with $k$ (taken odd) $b$-jets in a FS with a total of $J$ jets. The sensitivity of $N_{k,J}$ to $b_P$-violating new physics is determined by comparing the theoretical shift due to the underlying $b_P=-1$ interactions with the expected error ($\Delta$) in measuring the given quantity. Thus, requiring a signal of at least $N_{SD}$ standard deviations, we have $$\begin{aligned} \left| N_{k,J} - N_{k,J}^{\rm (SM)} \right| \geq N_{SD} \Delta \label{nsd}~.\end{aligned}$$ We will include three contributions to $ \Delta $ which we combine in quadrature: $\Delta^2 = \Delta_{\rm stat}^2 + \Delta_{\rm sys}^2 + \Delta_{\rm theor}^2$, where $\Delta_{\rm stat}=\sqrt {N_{k,J}}$ is the statistical error, $\Delta_{\rm sys}=N_{k,J} \delta_s$ is a systematic error and $\Delta_{\rm theor}= N_{k,J} \delta_t$ is the theoretical error in the numerical integration of the corresponding cross sections. The quantities $ \delta_{s,t} $ denote the statistical and theoretical errors per event; $ \delta_s $ is estimated using experimental values from related processes (eg. $R_b$ measurements), $ \delta_t $ is derived from the errors in the Monte Carlo integration used in calculating the various cross sections. There are various types of specific models beyond the SM (e.g., multi-Higgs models, supersymmetry, etc.) that can alter the SM prediction for the cross-section of reaction (\[nml\]). In this letter we will take a model-independent approach in which we investigate the limits that can be placed on the scale $\Lambda$ of a new short-distance theory that can generate flavor violation, and which we parameterize using an effective Lagrangian [@aewpaper] $$\begin{aligned} {\cal L}_{eff} = \frac{1}{\Lambda^2} \sum_i f_{i} {\cal O}_{i} + O(1/\Lambda^3) \label{leff}~,\end{aligned}$$where ${\cal O}_{i}$ are mass-dimension 6 gauge-invariant effective operators (some of which may have new flavor dynamics) [^2], and $f_i$ are coefficients that can be estimated using naturality arguments[@aewpaper]. As a concrete example that clearly illustrates the significance of $b_P$, we consider the effects of the $b_P$-odd effective four-Fermi operator $$\begin{aligned} {\cal O} = \left( \bar \ell \gamma^\mu \ell \right) \left( \bar q_i \gamma_\mu q_j \right) \label{fourfermi}~,\end{aligned}$$ where $\ell$ and $q$ are the SM left-handed lepton and quark $SU(2)_L$ doublets and $i,j=1,2~{\rm or}~3$ label the generation. This operator gives rise to contact $e^+e^- t \bar c$ and $e^+e^- b \bar s$ vertices (and their charged conjugates). It can be generated, for example, by an exchange of a heavy boson in the underlying theory (see [@aewpaper]). Although our method applies to any $b_P=-1$ process, in what follows we will investigate the effects of (\[fourfermi\]) on the reaction (\[nml\]) as an illustration. In particular, on $N_{1,2}$ (i.e., 1 $b$-jet signal in a 2-jet sample, $J=n+m+\ell=2$) and on $N_{1,4}$ and $N_{3,4}$ (i.e., 1 and 3 $b$-jet signals in a 4-jet sample, $J=n+m+\ell=4$). Consider first the $2$-jet sample case: in the limit $m_q=0$ for all $q \neq t$, the only relevant cross-sections are $\sigma_d = \sigma(e^+ e^- \to d \bar d)$, $\sigma_u = \sigma(e^+e^- \to u \bar u)$, where $d=d,~s,~b$ and $u=u,~c$, that are generated by the SM, and $\sigma_{bs} = \sigma(e^+e^- \to b \bar s) = \sigma(e^+e^- \to \bar b s)$ generated by the $eebs$ contact term. These cross-sections are calculated by means of the CompHEP package [@comphep], in which we implemented the Feynman rules for the $e^+e^-b \bar s$ and $e^+e^- t \bar c$ vertices generated by the operator (\[fourfermi\]). Using (\[sigmabar\]), we get the following cross-section for the 2-jet events, one of which is identified as a $b$-jet ($\bar\sigma_1$ with $J=n+m+\ell=2$) $$\begin{aligned} \bar\sigma_1 &=& P^2_1 \left[ \epsilon_b (1-\epsilon_b) + 2 t_j (1-t_j) \right] \sigma_d \cr && + P^2_1 \left[ t_c (1-t_c) + t_j (1-t_j) \right] \sigma_u \cr && + 2 (P_1^1)^2 \left[ \epsilon_b (1-t_j) + t_j (1-\epsilon_b) \right] \sigma_{bs} \label{sig1} ~\end{aligned}$$ that is used to calculate $N_{1,2}$. In Table \[tab1\] we give the largest $\Lambda$ (the scale of the new $b_P=-1$ physics) that can be probed or excluded at the level of 3 standard deviations ($N_{SD} = 3$), derived using (\[nsd\]), for the three representative $b$-tagging efficiencies of $25\%,~40\%$ and $60\%$ and fixing the $c$-jet and light-jet purity factors to $10\%$ and $2\%$, respectively.[^3] Results are given for three collider scenarios: $\sqrt s = 200$ GeV with $L=2.5$ fb$^{-1}$, $\sqrt s = 500$ GeV with $L=100$ fb$^{-1}$ and $\sqrt s = 1$ TeV with $L=200$ fb$^{-1}$. Both the systematic error $\delta_{s}$ and the theoretical uncertainty $\delta_{t}$ are assumed to be $5\%$. Also, an angular cut on the c.m. scattering angle of $|\cos\theta| < 0.9$ is imposed on each of the 2-jet cross-sections in (\[sig1\]). As expected, we see from Table \[tab1\] that the sensitivity to the new flavor physics induced by the four-Fermi interaction increases with the $b$-tagging efficiency. In Figs. \[fig1\] we show the regions in the $\epsilon_b -t_c$ plane (enclosed in the dark areas) where the flavor physics parameterized by (\[fourfermi\]) can be probed or excluded at the 3 standard deviation level (or higher); as an illustration we chose $\Lambda = 4 \sqrt s$ ($\sqrt{s}$ denotes the collider CM energy) for the collider scenarios mentioned above. The calculation was done using (\[nsd\]) for $N_{1,2}$ with $\delta_s=\delta_t=0.05$, $|\cos\theta| < 0.9$, $t_j=0.02$. Evidently, $\Lambda$ as large as four times the c.m. energy of any of the three colliders may be probed or excluded even for rather small $b$-tagging efficiencies; typically $\epsilon_b \gsim 25\%$ will suffice as long as the purity factors (in particular $t_c$ being the more problematic one) are kept below the $10\%$ level. [|c|c||c|c|c|]{}\ $\sqrt s$ & $L$ & $\epsilon_b=0.25$ & $\epsilon_b=0.4$& $\epsilon_b=0.6$\ (GeV) & (fb$^{-1}$)& $$&$$ & $$\ 200 & 2.5 & 0.68 & 0.74 & 0.81\ 500 & 100 & 1.81 & 1.96 & 2.15\ 1000 & 200 & 3.61 & 3.91 & 4.36\ For the 4-jet sample there are numerous processes that can contribute to $N_{1,4}$ and $N_{3,4}$. At the parton level the 4-jet events may be categorized as follows: (1) events containing 2 quark-antiquark pairs or one quark-antiquark pair and two gluons, $(q \bar q)( q^\prime {\bar q}^\prime)$, $(q \bar q) gg$, where both $q$ and $q^\prime$ denote any light quark ($q,q^\prime \neq t$) including the case $q=q^\prime$. (2) events with two charged quark pairs: $(u \bar d) ({\bar u}^\prime d^\prime)$, where $u$, $u^\prime$ are either $u$ or $c$-quarks and $d$, $d^\prime$ are any of the down-type quarks, excluding the states $u=u^\prime$ and $d=d^\prime$ since these are induced in type (1) above. (3) the 4 combinations $(b \bar s) gg$, $(b \bar b) (b \bar s)$, $(d \bar d) (b \bar s)$ and $(s \bar s) (s \bar b)$ (and the corresponding charged conjugate states) generated by the presence of the four-Fermi operator. It is worth noting that the $eetc$ contact term also contributes through graphs containing a virtual top-quark exchange. In order to get a reliable jet separation within the 4-jet sample, we use the so-called Durham criterion [@durham], that requires the quantities $y_{ij}^D = 2 {\rm min}(E_i^2,E_j^2)(1-\cos\theta_{ij})/s$, where $E_i$ and $E_j$ are the energies of the particles $i$ and $j$ and $\theta_{ij}$ is their relative angle ($i \neq j = 1,\ldots,4$). We evaluate all 4-jet cross-sections using the CompHEP package with the cuts $y_{ij}^D \geq y_{\rm cut}$ on all possible parton pairs $ij$ – we present our numerical results for $y_{\rm cut}=0.01$. In addition we neglect all quark masses except $m_{\rm top}$, and the strong coupling constant $\alpha_s$ was evaluated to the next-to-next-to leading order at a scale $Q$ equal to half the CM energy for 5 or 6 active quark flavors depending on whether $Q<m_{\rm top}$ or not respectively. For 6 flavors we used $ \Lambda_{QCD} =118.5$MeV (see [@comphep] for details). The results for the 4-jet case, using $N_{1,4}$, are shown in Figs. \[fig2\] for a 200, 500 and 1000 GeV colliders, where, as in Fig. \[fig1\], any value in the $\epsilon_b -t_c$ plane inclosed by the dark area will suffice for probing or ruling out (at $3\sigma$) the new four-Fermi operator in (\[fourfermi\]) with a scale $\Lambda$ as indicated in the figure. As for the 2-jet sample, we take $t_j=0.02$ and a systematic error of $5\%$. Using the results of the CompHEP Monte-Carlo integration we estimate that our calculated 4-jet cross-sections are accurate up to the level of about $10\%$, accordingly we choose $\delta_{t}=0.1$ in Fig. \[fig2\]. The 4-jet case is less sensitive to a $b_P$-odd signal induced by the four-Fermi operator in (\[fourfermi\]). For example, we see that for a 500 GeV collider with $t_c \sim 0.1$ and $t_j=0.02$, a $b$-tagging efficiency of about $70\%$ will be needed in order to probe or exclude a value of $\Lambda \sim 1300$ GeV by measuring $N_{1,4}$, while a $40\%$ $b$-tagging efficiency will suffice for probing $\Lambda \sim 2000$ GeV using $N_{1,2}$ in the 2-jet sample. Equivalently, for a given value of $ t_{c,j}$, higher $\epsilon_b$ will be required in the 4-jet measurement compared to the 2-jet one in order to detect a $b_P$-odd signal generated by the operator in (\[fourfermi\]) for the same value of $ \Lambda $. (this may be somewhat improved by reducing the theoretical uncertainties). We also find that $N_{3,4}$ is less sensitive than $N_{1,4}$ to ${\cal O}_{\ell q}^{(1)}$ in (\[fourfermi\]). Though $N_{1,2}$ is more efficient for probing type of new flavor physics which generate the four-Fermi operator (\[fourfermi\]) at low energies, this is not necessarily a general feature: certain types of new physics will not contribute to the 2-jet FS and must be probed using the 4-jet sample. This is the case, for example, for an effective vertex generating a right-handed $Wbc$ coupling, that may alter the flavor structure of SM, and give rise to sizable $b_P=-1$ effects. Note, however, that this refers only to the [*exclusive*]{} 2 and 4 jet samples, since such a $Wcb$ right-handed coupling will give rise to a $b_P$-odd signal in [*inclusive*]{} 2-jet reactions such as $ e^+ e^- \to b +j + X $, where $j$ is a light jet. The analysis of these events is, however, considerably more complex. We note that our cross sections include terms of order $ 1/\Lambda^4$ that will be modified by dimension 8 effective operators, which are in general present in (\[leff\]). Note, however, that such dimension 8 operators, if generated by the underlying high energy theory, are expected to give an additional uncertainty of order $ (s/\Lambda^2)^2 $, below $3\%$ for the results presented. In addition we note that the above analysis assumes unbiased pure samples with a fixed jet number, the effects of contamination from events with different jet number have not been included. Before we summarize we wish to note that the following issues need further investigation: - Our $b$-jet counting method can be used to constrain specific models containing $b_P=-1$ interactions. For example, supersymmetry with R-parity violation or with explicit flavor violation in the squark sector and/or multi-Higgs models without natural flavor conservation can give rise to $t \to c,~ t \to u$ (or $b \to s,b \to d$) transitions, which may lead to sizable $b_P$-odd signals in leptonic colliders. - In leptonic colliders with c.m. energies $\gsim 1.5$ TeV, $t$-channel vector-boson fusion processes become important. At such energy scales, the SM $b_P=-1$ reducible background needs to be reevaluated. At the same time, the $V_1V_2$-fusion processes ($V_{1,2} = \gamma,Z$ or $W$) give rise to a variety of new possible $b_P=-1$ signals from new flavor physics (see e.g., [@mhdm]). To summarize, we have shown that $b$-jet counting that relies on $b$-tagging (with moderate efficiency in a relatively pure multi-jet sample) can be used to efficiently probe physics beyond the SM. Reactions with $n$ final $b$-jets can be characterized through the use of the quantum number $b_P=(-1)^n$ that we called $b$-Parity. Due to small off-diagonal CKM matrix elements, $b_P$ is conserved within the SM to very good accuracy; it follows that the SM contributions to the above reactions are $b_P$-even. Despite the presence of a (reducible) background, due to reduced $b$-tagging efficiency and sample purity, we showed that our method is sensitive enough to provide very useful limits on new flavor physics in a variety of scenarios (of which two examples are provided) using realistic values for $ \epsilon_b $. We would like to thank D. Atwood R. Cole W. Gary R. Hawkings and B. Shen for illuminating comments and insights. This research was supported in part by US DOE contract number DE-FG03-94ER40837(UCR). T. Barklow [*et al.*]{}, hep-ph/9611454; R.D. Peccei and X. Zhang, Nucl. Phys. [**B337**]{}, 269 (1990); W. Buchmuller and D. Wyler, Nucl. Phys. [**B268**]{}, 621 (1986); A. Dobado, A. [*et al.*]{}, [*Effective Lagrangians for the Standard Model*]{} (Springer-Verlag, 1997) See, for example, [*New directions for high-energy physics*]{}: proceedings of the 1996 DPF/DPB Summer Study on High-Energy Physics, June 25-July 12, 1996, Snowmass, Colorado, edited by D.G. Cassel, [*et al.*]{} (SLAC, Menlo Park, CA, c1997) and references therein. M. Kobayashi and T. Maskawa, Prog. Theor. Phys.  [**49**]{}, 652 (1973). G. Abbiendi [*et al.*]{} \[OPAL Collaboration\], Eur. Phys. J.  [**C8**]{} (1999) 217 \[hep-ex/9810002\]. W. Buchmuller and D. Wyler, Ref. [@leff]; C. Arzt, M.B. Einhorn and J. Wudka, Nucl. Phys. [**B433**]{}, 41 (1995). E.E. Boos [*et al.*]{}, hep-ph/9503280; P.A. Baikov [*et al.*]{}, hep-ph/9701412; also in http://theory.npi.msu.su/comphep. S. Catani [*et al.*]{}, Phys. Lett. [**B269**]{}, 432 (1991). S. Bar-Shalom, G. Eilam, A. Soni and J. Wudka, Phys. Rev. Lett. [**79**]{}, 1217 (1997); [*ibid.*]{} Phys. Rev. [**D57**]{}, 2957 (1998), and references therein. [^1]: To be specific, we will consider reactions in $e^+e^-$ colliders, but the method is clearly extendible to muon colliders. [^2]: We assume there are no significant lepton-number violation effects at scale $ \Lambda $ that would generate dimension 5 operators. [^3]: Note that the limits derived here and throughout the rest of the paper assume $|f|=1$ \[see (\[leff\])\]. Alternatively, they can be interpreted as limits on $\Lambda/\sqrt {|f|}$.

ArXiv

--- abstract: 'We present the controlled creation of single nitrogen-vacancy (NV) centers via ion implantation at the center of a photonic crystal cavity which is fabricated in an ultrapure, single crystal diamond membrane. High-resolution placement of NV centers is achieved using collimation of a -nitrogen ion beam through a pierced tip of an atomic force microscope (AFM). We demonstrate coupling of the implanted NV centers’ broad band fluorescence to a cavity mode and observe Purcell enhancement of the spontaneous emission. The results are in good agreement with a master equation model for the cavity coupling.' author: - 'Janine Riedrich-Möller' - Sébastien Pezzagna - Jan Meijer - Christoph Pauly - Frank Mücklich - Matthew Markham - 'Andrew M. Edmonds' - Christoph Becher title: Nanoimplantation and Purcell enhancement of single NV centers in photonic crystal cavities in diamond --- The nitrogen-vacancy (NV) center [@Doherty2013] in diamond has been successfully implemented as solid state quantum bit that meets all essential requirements for quantum information processing such as optical initialization, control and readout of the spin state. The challenge remains to extend the quantum system from a small number of qubits to large scale networks. Seminal experiments already demonstrated remote entanglement between individual NV centers via two-photon quantum interference [@Sipahigil2012; @Bernien2013]. The hitherto poor rate of entanglement events [@Sipahigil2012; @Bernien2013] could be strongly increased by coupling the NV centers to optical microcavities. The effects range from enhancement and spectral reshaping of the NV spectrum over cavity-enhanced spin state readout [@Young2009] to cavity mediated entanglement between two NV centers [@Wolters2014]. Photonic crystal (PhC) cavities directly fabricated in diamond are ideal for color center-cavity coupling experiments as they exhibit high Q-factors and extremely small mode volumes. For solid state systems, it is however challenging to precisely place the emitter in the maximum of the cavity electric field to achieve optimum coupling. Past experiments that demonstrated coupling of single NV centers to PhC cavities [@Faraon2012; @Hausmann2013] have largely relied on random positioning. Controlled lateral positioning and emitter-cavity coupling has recently been achieved via a tailored fabrication process of a PhC around a single silicon-vacancy center in diamond [@Riedrich-Moller2014b]. Here, we pursuit the complementary approach based on targeted implantation of NV centers into pre-defined cavities in diamond. In recent years, several techniques for spatially selective formation of single NV centers in bulk diamond have been developed involving focused nitrogen ion beam [@Lesik2013], implantation through pierced AFM-tips [@Meijer2008; @Pezzagna2010] and through small apertures in e-beam resist [@Toyli2010; @Spinicelli2011], mica foils [@Pezzagna2011a], and silicon masks [@Bayn2015]. Using the silicon mask simultaneously as an etch mask would allow for controlled emitter-cavity placement [@Schroder2014a]. In our experiment, we achieve high resolution implantation of NV centers within two-dimensional diamond-based PhC cavities using a combined system of a nitrogen ion beam and an atomic force microscope (Fig. \[fig1\](a)) that allows for collimation and lateral positioning of the ion beam [@Meijer2008; @Pezzagna2010]. We verify the successful formation of a small number of NV centers and demonstrate Purcell enhancement of the broad NV emission when coupled to the confined cavity field. The PhCs consist of a triangular lattice of air holes milled in a single crystal diamond membrane with a refractive index of $n=2.4$. The cavity is introduce by a one-, three- or seven-hole defect at the center, referred to as M1-, M3- or M7-cavity, respectively. For membrane preparation a high purity synthetic diamond ( nitrogen concentration) was synthesized using microwave assisted chemical vapor deposition. The as grown (001) single crystal diamond was processed using standard diamond lapidary for bulk material removal followed by scaife polishing to thin the diamond to . The membrane is bonded on a silicon substrate via a spin-on-glass adhesion layer (hydrogen silsesquioxane, Dow corning XR-1541). The silicon substrate has been partially removed in order to obtain a free-standing diamond membrane which is subsequently thinned to using reactive ion etching in an oxygen plasma and patterned with an array of air holes using focused ion beam milling with -Ga$^+$ ions [@Riedrich-Moller2012]. As a final step, the sample is annealed at 800$^\circ$C for in vacuum and is thoroughly cleaned in a boiling mixture of nitric, sulfuric, and perchloric acid. The lattice constant $a = $ and air hole radii $R = $ of the M1-, M3- and M7-cavities are chosen such that the resonant modes are close to the design wavelength of with experimental quality factors of $Q=150-1200$ and mode volumes of $V \approx 1(\lambda/n)^3$. The hole radii of the M3- and M7-cavity are uniform in size, whereas the next neighbor holes around the M1-defect have been displaced and reduced in size to optimize the cavity $Q$ [@Kreuzer2008] (c.f. Figs. \[fig1\](b,c)). The fabricated structures are investigated using a home-build confocal microscopy setup with a continuous-wave excitation laser where the sample is mounted in a continuous flow liquid-helium cryostat. Figure \[fig1\](d) shows the room temperature photoluminescence (PL) spectrum of the fabricated M1-cavity revealing three pronounced cavity modes $c_1$, $c_2$ and $c_3$ at 653, 670 and , respectively, but no signature of NV emission in the ultrapure diamond material. For deterministic creation of NV centers within the PhC cavities, we first use an AFM to image the PhC structures (c.f. Figs. \[fig1\](c)). A small hole (diameter ) drilled in the AFM tip serves as an aperture for the ion beam that allows for high resolution implantation [@Meijer2008; @Pezzagna2010] of $^{15}$N$^+$ ions with an energy of at the cavities’ center at different doses of . The low ion energy is chosen to achieve a high spatial resolution of [@Pezzagna2010]. According to Monte Carlo simulations (SRIM [@Srim2008]), the average implantation depth of the -nitrogen ions is with a small ion straggle of . After implantation, the diamond sample is annealed at $800^\circ$C for in vacuum such that lattice vacancies diffuse in the diamond host material towards the implanted nitrogen ions to form optically active NV centers. Finally, the sample is cleaned again in a boiling acid mixture for in order to oxidize any graphite-like residuals and to convert the NV centers to the negative charge state. Ensemble NV emission spectra taken at different reference spots implanted at high dose () aside the photonic structures reveal that up to 70% of the NV centers are converted to their negative charge state after all oxidation steps. In our analysis, we take into account the PL intensities integrated in a spectral range of around the NV$^0$ ($\lambda =$ ) and NV$^-$ zero-phonon line (ZPL) ($\lambda =$ ) as well as varying detection efficiencies of our spectrometer, different absorption [@Beha2012a] and quantum [@GattoMonticone2013] efficiencies of the two charge states. In the following, we refer to the most abundant NV$^-$ center simply as NV center. We verify the successful formation of NV centers within PhC cavities using confocal spectroscopy at . Figure \[fig2\] shows the PL scan and spectra of the M1-cavity after nitrogen ion implantation at the lowest dose of . Besides the three cavity modes $c_1$, $c_2$ and $c_3$, a clear signature of NV ZPLs around is visible in the spectrum collected at the M1-cavity center (black spectrum in Fig. \[fig2\](a)). A zoom into the spectral region around (Fig. \[fig2\](b)) reveals $3\pm1$ narrow Gaussian-shaped lines with linewidths of $\Delta \nu_0 \approx$ . At low temperature the linewidth is limited by spectral diffusion [@Wolters2013a]. Each line corresponds to the ZPL of a single NV center that has been created upon ion implantation and subsequent annealing. The implantation is solely restricted to the cavity center. Reference spectra collected off the cavity center (gray spectrum in Fig. \[fig2\](a)) do not show any signature of NV ZPLs, which verifies the high spatial resolution of the implantation process. We determine the number of NV centers created within each M1-, M3- and M7-cavity that were implanted at various ion doses by integrating the background corrected PL signal in the spectral range of and normalize it to the average ZPL intensity of a single NV center. As displayed in Fig. \[fig2\](c), the number of produced NV centers monotonically increases as a function of the applied implantation dose. By dividing the number of NV centers by the amount of implanted nitrogen ions, we obtain the NV creation yield shown in Fig. \[fig2\](d). For the ion energy of , we find a creation yield of $0.8\pm 0.2\%$ that is constant over the whole range of implantation dose. The small creation yield is within the range of experimental observations of yields $< 0.1 \%$ [@Pezzagna2010a] to $\sim 25\%$ [@Antonov2014] which strongly depend on annealing and surface conditions. The creation yield is limited by loss of vacancies to the surface upon shallow implantation and surface effects possibly shifting the NV center charge state to NV$^+$. From Fig. \[fig2\](c), we deduce an optimal dose of at an ion energy of for deterministic creation of one single optically active NV center. In the following, we analyze the intensity enhancement of the NV emission at the resonant wavelength of the M1-cavity mode $c_1$. Thereby, we take into account that the mode $c_1$ preferentially overlaps with the NV phonon side band (PSB) and not with the ZPL. As the resonance wavelengths and linewidths are known from the M1-cavity spectrum (Fig. \[fig1\]) prior implantation, we can estimate the bare NV emission (orange line in Fig. \[fig3\](a)) without cavity modes. By comparing the integrated intensity of the cavity-enhanced emission $I_{\text{on}}$ (orange + gray area in Fig. \[fig3\](b)) to the bare spectrum $I_{\text{off}}$ (orange area), we find an experimental enhancement factor of $I_{\text{on}}/I_{\text{off}} = 1.24$. Here, we solely focus on the intensity increase by the dominant $c_1$ mode and disregard other modes $c_2$ and $c_3$. In addition the emission efficiency into the cavity mode is $\beta = I^{c_1}_{\text{on}}/I^{\text{tot}}_{\text{on}} = 0.31$, where $I^{c_1}_{\text{on}}$ is the emission channeled into mode $c_1$ (pink area in Fig. \[fig3\](c)) and $I^{\text{tot}}_{\text{on}}$ is the overall intensity (pink + gray area). To relate the resonant intensity enhancement of the NV PSB to a generalized Purcell factor $F^*$, we adopt the master equation model [@Albrecht2013; @Auffeves2009] for broad-band emitter-cavity coupling. In analogy to [@Albrecht2013], the NV emission is modeled as a multi-level system (inset in Fig. \[fig3\](a)). The model input parameters are obtained from Lorentzian fits to the uncoupled NV spectrum (Fig. \[fig3\](a)). Solving the master equation model, we compute the generalized Purcell factor $F^*$ [@Albrecht2013; @Auffeves2009] (Fig. \[fig3\](d)) and the associated emission factor $\beta = F^*/(1+F^*)$ (Fig. \[fig3\](e)) into the cavity when the resonant mode $c_1$ with $Q_{c_1}=160$ and $V_{c_1} = 1.1\, (\lambda/n)^3$ is tuned across the modeled NV spectrum. The individual contributions of the ZPL and PSBs to the total emitter-cavity coupling are shown by the filled curves in Figs. \[fig3\](d,e). For simplicity, our analysis assumes unity quantum efficiency and perfect spatial and orientational overlap of the two NV dipoles with the cavity field. At the resonant wavelength $\lambda_{c_1} = $ of the M1-cavity mode (c.f. Fig. \[fig3\](d)), we find a theoretical Purcell enhancement of $1+F^* = 1.7$ and an emission efficiency $\beta = 0.42$ that result in a theoretical intensity increase $I_{\text{on}}/I_{\text{off}} = ((1+F^*)\,e_{1,2}\, \gamma+(1-e_{1,2})\gamma) / \gamma = 1.2$, considering that the $c_1$ mode preferentially overlaps with the first and second NV PSB with a relative contribution to the total NV emission of $e_{1,2} = 0.29$. These theoretical values are in excellent agreement with our experiment. In conclusion, we have demonstrated high resolution creation of a small number of NV centers at the center of diamond-based PhC cavities using collimated nanoimplantation of nitrogen ions through a pierced AFM-tip. For an ion energy of , we found a constant NV creation yield of 0.8%, independent of implantation dose and cavity size. The lowest ion dose of yielded $3\pm1 $ NV centers placed at the center of a M1-cavity. The emitter-cavity coupling leads to an intensity enhancement of $I_{\text{on}} / I_{\text{off}} = 1.24$ when the cavity mode is in resonance with the NV PSB and an emission efficiency into the cavity mode of $\beta = 0.31$ which is in very good agreement with theoretical predictions. From our experiment, we deduce an optimal dose of for the targeted creation of one single NV center within a PhC cavity. The here demonstrated high resolution implantation of single NV centers within PhC cavities is an essential step towards scalable solid-state quantum networks [@Nemoto2013] or quantum repeaters [@Childress2005] based on NV nanocavity systems. We cordially thank R. Albrecht and A. Bommer for helpful discussions, S. Wolff and C. Dautermann (Nano Structuring Center, University of Kaiserslautern) for assistance with the bonding process and A. Baur and M. Wandt (IMTEK, University of Freiburg) for deep reactive ion etching. This research has been partially funded by the European Community’s Seventh Framework Programme (FP7/2007-2013) under Grant Agreement No. 618078 (WASPS) and No. 611143 (DIADEMS). EU funding for the project AME-Lab (European Regional Development Fund C/4-EFRE 13/2009/Br) for the FIB/SEM is acknowledged. MM and AME additionally acknowledge finical support from the DARPA SPARQC. [29]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****,  ()](\doibase 10.1016/j.physrep.2013.02.001) [****,  ()](\doibase 10.1103/PhysRevLett.108.143601) @noop [****,  ()]{} @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevA.89.060303) [****,  ()](\doibase 10.1103/PhysRevLett.109.033604) @noop [****,  ()]{} [****,  ()](\doibase 10.1021/nl502327b) [****,  ()](\doibase 10.1002/pssa.201300102) @noop [****,  ()]{} @noop [****,  ()]{} [****,  ()](\doibase 10.1021/nl102066q) @noop [****, ()]{} [****,  ()](\doibase 10.1002/pssa.201100455) [ (), doi:10.1021/nl504441m](\doibase doi:10.1021/nl504441m) in [**](\doibase 10.1364/CLEO_QELS.2014.FW1B.6) (, ) p.  [****,  ()](\doibase 10.1038/NNANO.2011.190) @noop [****,  ()]{} [“,” ](http://www.srim.org/) () [****,  ()](\doibase 10.1103/PhysRevLett.109.097404) [****,  ()](\doibase 10.1103/PhysRevB.88.155201) @noop [****,  ()]{} @noop [****,  ()]{} [****,  ()](\doibase http://dx.doi.org/10.1063/1.4860997) [****,  ()](\doibase 10.1103/PhysRevLett.110.243602) @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevX.4.031022) @noop [****,  ()]{}

ArXiv

--- abstract: 'The interest in studying relativistic compact objects play an important role in modern astrophysics with an aim to understand several astrophysical issues. It is therefore natural to ask for internal structure and physical properties of specific classes of compact star for astrophysical observable, and we obtain a class of new relativistic solutions with anisotropic distribution of matter for compact stars. More specifically, stellar models, described by the anisotropic fluid, establish a relation between metric potentials and generating a specific form of mass function are explicitly constructed within the framework of General Relativity. New solutions that can be used to model compact objects which adequately describe compact strange star candidates like SMC X-1, Her X-1 and 4U 1538-52, with observational data taken from Gangopadhyay *et al.* [@Gangopadhyay]. As a possible astrophysical application the obtained solution could explain the physics of self-gravitating objects, might be useful for strong-field regimes where data are currently inadequate.' author: - 'S. K. Maurya' - Ayan Banerjee - 'Y. K. Gupta' title: 'Exact solution of anisotropic compact stars via. mass function' --- I.   Introduction {#i.introduction .unnumbered} ================= Over the past few years, there has been growing interest in static spherically symmetric compact object consistent with general relativity which are most often observed as pulsars, spinning stars with strong magnetic fields. However, in spite of the fact, it is important to know the exact composition and nature of particle interactions, which allow us to completely describe them in terms of their mass, spin angular momentum and charge. In particular, observations of compact stars are considered primary targets of the forthcoming field of gravitational wave astronomy. In the recent times we have experimental evidence that such objects do exist from observations with very high densities [@Lattimer], and some of the compact objects like X-ray pulsar Her X-1, X-ray burster 4U 1820-30, X-ray sources 4U 1728-34, PSR 0943+10, millisecond pulsar SAX J 1808.4-3658, and RX J185635-3754 strongly favour the possibility that they could actually be strange stars. In other words, there is no any strong evidence to make conclusion/ understood the mechanism about compact objects. In spite of this drawback, the internal composition and consequent geometry of such objects are still considered an open question in the scientific community. From a theoretical point of view, such compact objects are composed of a perfect fluid [@Ivanov]. Generally, polytropic equation of state (EOS) in the form P = k $\rho^{\gamma}$ and bag model [@Witten; @Glendenning] have widely used to describe a white dwarf or a less compact star [@Shapiro]. An example, Herrera and Barreto [@Barreto] carry out a general study on polytropic general relativistic stars with anisotropic pressure whereas Lai and Xu [@Lai] have studied polytropic quark star model. Though, interior of a star is an important astrophysical question and hence it is pertinent to construct relativistic models by assuming anisotropy fluid distribution. The theoretical study of the influence of anisotropic compact objects was first initiated by Bowers and Liang [@Bowers] and another investigation led be Ruderman [@Ruderman] showed that nuclear matter may have anisotropic features at least in certain very high density ranges $10^{15}$ g/$cm^3$, where the nuclear interaction must be treated relativistically. Later on, several models (see for example, [@Gleiser; @Dev2003; @Herrera2002]) have been proposed in this direction. The procedure has been found to be interesting and useful on the physical grounds that anisotropy affects the critical mass, mass-radius relation and stability of highly compact relativistic bodies. It is also well-established fact that a magnetic field acting on a Fermi gas produces pressure anisotropy was discussed recently in [@Chaichian; @Martinez; @Ferrer]. Recently, Maurya and Gupta in [@Maurya2015a; @Maurya2016a] have studied charged anisotropic stars whereas without charged solution has been analysed in [@Hossein; @Mehedi2012]. The role of anisotropy, with the linear equation of state was pursed by Mak and Harko [@Mak(2002)], and extending the work by Sharma and Maharaj [@Sharma] obtained an exact analytical solutions assuming a particular form of mass function. In yet another paper, Victor Varela et al [@Varela] studied Charged anisotropic matter with linear or nonlinear equation of state . However, it has recently been proposed that an arbitrary d dimensional (pseudo-) Riemannian space can always be locally embedded into any Riemann space of dimension N $\geq$ d(d + 1)/2. Riemann’s seminal work in 1868 [@Riemann] had inspired Schl$\ddot{a}$fli that how one can locally embed such manifolds in Euclidean space [@Schlafli]. He discussed the local form of the embedding problem and conjectured that maximum number of extra dimensions that can embedded as a submanifold of a Euclidean space $E_N$ with N= $\frac{1}{2}n(n+1)$ or, equivalently, when the codimension r = N - n it gives r = n(n - 1)/2. Furthermore, one motivation in the treatment of embedding has been devoted, namely, Randall- Sundrum braneworld model is based on the assumption that four dimensional space-time is a three-brane, embedded in a five-dimensional Einstein space [@Randall]. On the other hand Nash in 1956 [@nash] established the idea of global isometric embedding theorem of $V_n$ into Euclidean space $E_N$. In a sense one could say that all Riemannian manifold has a local and a global isometric embedding in an Euclidean space. This result opened new perspectives in embedding theorems with increasing degrees of generality and soon it became an powerful tool to construct and classify solutions of GR. Inspired by these advances, a popular approach has been emerged in embedding 4-dimensional space-time into 5-dimensional flat space-time by using the spherical coordinates transformation and known as embedding class one if it satisfies the Karmarkar condition [@Karmarkar]. One of the primary motivations of such embedding is to establish a relationship between metric potentials and obtained exact solutions of Einstein’s equation to a single-generating function. In connection to this an exact anisotropic solution of embedding class one, has been developed in [@Maurya(2016); @Maurya2017a; @Newton2016]. They showed that in seeking solution for relativistic static fluid spheres one can utilized this technique successfully. Motivated by the above facts, we will see how this theorem can be utilized for obtaining an exact solution of Einstein’s field equation for compact star models. In this article we consider the static spherically symmetric spacetime metric with embedding class one conditions, which can be altered to fit with a set of astrophysical objects. This paper is outlined in the following manner. In Sec. II, we present the structural equations for anisotropic fluid distributions of stellar models applying the embedding class one conditions. Specific models are then analyzed in a brief description by obtaining a particular form of mass function. In Section III, we match our interior solution to an exterior Schwarzschild vacuum solution at the boundary surface and then determine the values of constant parameters. In Section IV, we continue our discussion thorough geometrical analysis of the solution such as energy conditions, hydrostatic equilibrium and stability of the star by fixing of certain parameters. We end our discussion by concluding remarks in V. II.  Class one condition for Spherical symmetric metric and General relativistic equations {#ii.class-one-condition-for-spherical-symmetric-metric-and-general-relativistic-equations .unnumbered} ========================================================================================== The simplest configuration for a star is the static and spherically symmetric geometry has the usual form $$ds^{2}=e^{\nu(r)}dt^{2}-e^{\lambda(r)}dr^{2}-r^{2}\left(d\theta^{2}+\sin^{2}\theta d\phi^{2} \right),\label{eq1}$$ where the coordinates (t, r, $\theta$, $\phi$) are the spherical coordinates and the metric coefficient $\lambda$ and $\nu$ are the functions of the radial coordinate $r$, and yet to be determined by solving the Einstein equations. As the metric (1) is time independent and spherically symmetric, we restrict ourselves that the space time is of emending class one, if it satisfies the Karmarker condition (see Ref. [@Karmarkar; @Maurya(2016)] for more details discussion) and metric functions $e^{\lambda}$ and $e^{\nu}$ are dependent on each other as: $$e^{\lambda}=1+K\,F'^2, \label{eq2}$$ where  $F=e^{\nu/2}$, $F = F(r)$ and $K > 0$. However $F$ = constant, leads to flat space time metric. The energy momentum tensor $T_{\mu\nu}$ associated to a spherical distribution of matter bounded by gravitation is locally anisotropic, that is $T_{\mu\nu}$ = $\text{diag} \left(\rho, -p_r, -p_t, -p_t\right)$, where $p_r$ and $p_t$ are the radial and tangential pressures and $\rho$ is the energy density of the fluid, respectively. Thus, the Einstein field equation, $G_{\mu\nu}$ = 8$\pi T_{\mu\nu}$, where $G_{\mu\nu}$ is the Einstein tensor then reduce to the following ordinary differential equations for the metric \[eq1\] (since we use natural units where G = c = 1) as: $$8\pi\,\rho(r) = e^{-\lambda}\,\left[\frac{\lambda'}{r}+\frac{e^{\lambda}-1}{r^2}\right], \label{eq3}$$ where the primes ($\prime$) denote differentiation with respect to r. Then one may write the solution in a very familiar form of mass function of the compact object, $$m(r)=\int_{0}^{r}{4\,\pi\,r^2\,\rho(r)\,dr}. \label{eq4}$$ It is worth noting that the mass is the density inside a proper volume element within a radius r. Using the above Eq. (\[eq4\]) and Eq. (\[eq2\]), we can write the metric component $e^\lambda$ which is given by the equality $$e^{-\lambda}=1-\frac{2m}{r}=(1+K\,F'^2)^{-1}. \label{eq5}$$ Rewriting the above expression in terms of mass function as: $$m(r)= \frac{K\,r\,F'^2}{2\,(1+K\,F'^2)} \,\,\, ~~\text{or}~~\,\, F'^2=\frac{2m}{K\,(r-2m)}. \label{eq6}$$ Since $F'(0) = 0$ and $F^{\prime\prime}(0)> 0$ because $F=e^{\nu/2}$, $\nu^{\prime}(0)= 0$ and $\nu^{\prime\prime}(0)> 0$ which is already proved by Herrera *et al.* [@Herrera1] and Maurya *et al.* [@Maurya1]. From Eq. (\[eq6\]), we obtain $m(0)\,= \,m'(0)\,=\, m''(0)=0$ and $m'''(0)=3\,F''(0)/K > 0$. This solution clearly implies that $m(r)$ is positive and monotonic increasing function of r. It is possible to derive the mass function in terms of F(r). Aiming this the value of F(r) from Eq. (\[eq6\]), we have $$F(r)=\frac{1}{\sqrt{K}}\,\int{\sqrt{\frac{2m}{r-2m}}}. \label{eq7}$$ To solve this integral, we shall assume $2m(r) = r-f(r)$, that we claim $f(r)$ should be zero at centre i.e. $f(0)=0$ due to m$(0)=0$, which indicating that the mass function is regular at the origin. Equation (\[eq7\]) can be rewritten as $$F (r) = \frac{1}{\sqrt{K}}\,\int{\sqrt{\frac{r-f(r)}{f(r)}}}. \label{eq8}$$ To make the above integral equations tractable and construct a physically viable model, we suppose that $\frac{r-f(r)}{f(r)}=ar^4+br^2$ in a particular form, which gives $$f(r)=\frac{r}{ar^4+br^2+1}. \label{eq9}$$ where two constants a, b are positive and makes the Eq. (\[eq8\]) in a simplest form $$F (r) = \frac{1}{\sqrt{K}}\,\int{\sqrt{a\,r^4+b\,r^2}}. \label{eq10}$$ Solving the integral gives $$F=A+\frac{(a\,r^2+b)^{3/2}}{3\,a\,\sqrt{K}}, \label{eq11}$$ where A stands for integrating constant. Substituting Eq. (\[eq11\]) in Eq. (\[eq2\]), we obtain one of the metric coefficient in the form $$e^{\lambda}=1+b\,r^2+a\,r^4, \label{eq12}$$ and the mass function is given by $$m(r)=\frac{r^3\,(a\,r^2+b)}{2\,\Bigl(1+b\,r^2+a\,r^4 \Bigl)}. \label{eq13}$$ The motivation of choice for particular dimensionless function lies on the fact that the obtained mass function a monotonic decreasing energy density in the interior of the star. To construct a physically viable model this type of mass function is not new but similar works have been considered earlier by Matese & Whitman [@Matese] and Finch & Skea [@Finch] for isotropic fluid spheres, and Mak & Harko [@Mak] for anisotropic fluid spheres. The system of equations used to study for spherically symmetric configurations with anisotropic fluid distribution are $$\begin{aligned} \hspace{-2.6cm} 8\pi\,\rho (r)=\frac{b^2\,r^2 + a\,r^2\, (5 + a\,r^4) + b \Bigl(3 + 2\,a\,r^4\Bigl)}{\Bigl(1 + b\,r^2 + a\,r^4\Bigl)^2}, \label{eq14}\end{aligned}$$ $$8\pi\,p_r (r)= -\frac{1}{\Bigl(a\,r^4+b\,r^2+1\Bigl)}\left[{b+a\,\left(r^2-\frac{6\,B\,\sqrt{a\,r^2+b}} {A+B\,(a\,r^2+b)^{3/2}}\right)}\right], \label{eq15}$$ $$8\pi\,p_t(r) = \frac{a^2\,r^2\,B\,(a\,r^4+b\,r^2+9)-b\,(b^2\,B+A\,\sqrt{a\,r^2+b})+f_1(r)}{\sqrt{a r^2+b}~~ \Bigl(1 + b r^2 + a r^4\Bigl)^2~~ \Bigl[A + B (b + a r^2)^{3/2}\Bigl]}, \label{eq16}$$ where we use $f_1(r)=a\,\left[6\,b\,B-b^2\,B\,r^2-2\,A\,r^2\,\sqrt{ar^2+b}\right]$ for our notational convention. Firstly, we will present the anisotropic effect by a term $(p_t - p_r)/r$, by taking into account Eqs. (\[eq15\]-\[eq16\]), may be expressed in the following equivalent form $$\begin{aligned} \Delta=\frac{r^2\,\Bigl[B\,(b^2+a^2\,r^4)+A\,g - a\,B\,(3-2\,b\,r^2)\,(b^2+a^2\,r^4-a+2\,a\,b\,r^2)\,\Bigl]}{8\pi\,\sqrt{a\,r^2+b}~~ \Bigl(1 + b \,r^2 + a\,r^4\Bigl)^2~~ \Bigl[A + B (b + a r^2)^{3/2}\Bigl]}, \label{eq17}\end{aligned}$$ which representing a force that is due to the anisotropic nature of the fluid. The anisotropy will be repulsive or directed outwards if $p_t > p_r$, and attractive or directed inward when $p_t < p_r$. The effects of anisotropy forces maintain the stability and equilibrium configurations of a stellar stricture, as we discuss later. {width="5cm"} {width="5cm"} In more detail, it is important to emphasize the results. That’s why we extend our calculation for the first order differential equation, which are $$\begin{aligned} \frac{d\rho}{dr}= - \frac{2\,r\,\left[\,a^3 r^8 + 3 a^2 r^4 (4 + b r^2) + b^2 (5 + b r^2) + a (-5 + 13 b r^2 + 3 b^2 r^4)\,\right]}{8\pi\,\Bigl(1 + b r^2 + a r^4\Bigl)^3}, \label{eq18}\end{aligned}$$ $$\begin{aligned} \hspace{-0.6cm} \frac{dp_r}{dr}= {\frac{-2ar\,(g\,f^2+9\,a\,B^2 g^{3}-3\,a\,B\,f)}{8\,\pi\,g~f^2\,\Bigl(1 + b\,r^2 + a\,r^4\Bigl)}} + \frac{2\,r\,(2\,a\,r^2+b)\,(f~g^2-6\,B\,g)}{8\,\pi\,\Bigl(1 + b\,r^2 + a\,r^4 \Bigl)^2~f}, \label{eq19}\end{aligned}$$ $$\begin{aligned} \hspace{-1.7cm} \frac{dp_t}{dr}= \frac{2\,r\,\Bigl[\,\Psi_1(r)+\Psi_2(r)+\Psi_3(r)+\Psi_4(r)+\Psi_5(r)\,\Bigl]} {16\,\pi\,\,(b + a r^2)^{3/2} (1 + b r^2 + a r^4)^3 \Bigl[A + B (b + a r^2)^(3/2)\Bigl]^2}, \label{eq20}\end{aligned}$$ for our notational conventional we use $f=\Bigl[A + B (b + a r^2)^{3/2}\Bigl]$,    $g=\sqrt{b+ a r^2}$,     and\ $\Psi_1(r)=-6 a^6 B^2 r^{12} g + 4 b^3 \Bigl[2 A b^2 B + A^2 g + b^3 B^2 g ] + a^5 B r^8 [15 A r^2 - 2 B g (44 + 9 b r^2)\Bigl]$,\ $\Psi_2(r)=2 a b \Bigl[-11 b^3 B^2 g h - 2 A^2 g (4h-3) - A b^2 B (21 h - 8)\Bigl]$,\ $\Psi_3(r)=-a^4 B r^4\,\Bigl[-7 A x (-8 + 9 b r^2) + 2 B g (9 + 131 b r^2 + 4 b^2 r^4)\Bigl]$,\ $\Psi_4(r)=a^2 \Bigl[4 A^2 r^2 g (-1 + 6 b r^2) - 2 b^2 B^2 g (3 + 65 b r^2 - 21 b^2 r^4) + 3 A b B (4 - 31 b r^2 f)\Bigl]$,\ $\Psi_5(r)=a^3 r^2 \Bigl[12 A^2 r^4 g - 2 b B^2 g [12 + b r^2 (141- 14 b r^2)] + A B (9 - 123 b r^2 + 107 b^2 r^4)\Bigl]$. We exercised our results in various ways first by checking calculation and putting restriction on the physical parameter based on logarithmic principle and then by using graphical representation which are illustrated in Figs. (1-3), describing the metric functions, energy density, radial and transverse pressures and measure of anisotropic within the given radius. In the next section we will look for supplementary restrictions on the model to make it physically viable. {width="5cm"} {width="5cm"}\ {width="5cm"} {width="5cm"} III.   Boundary conditions {#iii.boundary-conditions .unnumbered} ========================== To do the matching properly, we start by joining an interior spacetime $\mathcal{M_{-}}$, to an exterior $\mathcal{M_{+}}$, Schwarzschild vacuum solution at the boundary surface r = R. At this boundary the metric should be continuous. The Schwarzschild solution is given by $$\begin{aligned} \label{21} &\qquad\qquad\hspace{-2.7cm}ds^{2} =\left(1-\frac{2M}{r} \right)\, dt^{2} -\left(1-\frac{2M}{r} \right)^{-1} dr^{2}-r^{2} (d\theta ^{2} +\sin ^{2} \theta \, d\phi ^{2} ),\nonumber \\\end{aligned}$$ where $M$ denote the total mass of the compact star. In deciding criterions for an anisotropic compact star the radial pressure $p_{r}$ must be finite and positive inside the stars, and it should vanishs at the boundary $r = R$ of the star [@Misner]. In determining the value of constrains parameters we use the boundary condition $p_r(R) =0$, which can express as $$\begin{aligned} \label{22} \frac{A}{B}=\frac{\sqrt{b + a\, R^2}~ \Bigl[\,6 a B - b^2 B - 2\,a\,b\, B\, R^2 -a^2\,B\, R^4 \,\Bigl]}{(b + a\,R^2)}. \label{16}\end{aligned}$$ In order to match smoothly on the boundary surface, we must require the continuity of the first and the second fundamental forms across that surface. Then it follows the condition $e^{\nu(R)}=e^{-\lambda(R)}$ which gives the value of constant parameter B as $$\begin{aligned} B=\frac{1}{\sqrt{1+bR^2+aR^4}~\left[\,\frac{A}{B}+(aR^2+b)^{3/2}\,\right]}, \label{23}\end{aligned}$$ and using the condition $e^{-\lambda(R)}=1-\frac{2M}{R}$, we have $$\begin{aligned} M=\frac{R^3\,\left(aR^2+b \right)}{1+bR^2+aR^4}. \label{24}\end{aligned}$$ This represents the total mass of the sphere as seen by an outside observer. Additionally, bounds on stellar structures is an important source of information and classification criterion for compact objects, to determine the mass-radius ratio which was proposed by Buchdahl [@Buchdahl]. This bound has been considered for thermodynamically stable perfect fluid compact star with ratio 2M/R, must be less than 8/9. We have carried out the analysis for our compact star candidates SMC X-1, Her X-1 and 4U 1538-52, that are used to calculate the values of constants A and B. For this purpose, we present these results (central density, surface density, central pressure, mass-radius ratio) in table I & II, respectively. Table I -------------- ---------- --------------------------- ------------------- ------------------- ------------- -------------- Compact Star   R (km)   M$\left(M_\odot\right)$ $~~~a\,(km^{-4})$    $b\,(km^{-2})$     A  $B(km^{3})$ SMC X-1 8.301 1.04 0.0000270 0.006648 0.439167064 452.08447 Her X-1 8.1 0.85 0.0000185 0.005620 0.481332147 618.82624 4U 1538-52 7.866 0.87 0.0000247 0.006300 0.481382726 490.02078 Let us now focus on surface gravitational redshift, which gives a wealth of information about compact objects, and defined by Z = $\Delta \lambda/\lambda_{e} $ = $\frac{\lambda_{0}-\lambda_{e}}{\lambda_{e}}$, where $\lambda_{e} $ is the emitted wavelength at the surface of a nonrotating star and the observed wavelength $\lambda_{e}$. Thus one can estimate the gravitational redshift, defined by $Z_s $, from the surface of the star as measured by a distant observer by the following relation $$Z_{s} = -1+\Bigl\rvert g_{tt}(R)\Bigl\rvert ^{-1/2} = -1+\left( 1-\frac{2M}{R}\right) ^{-1/2},$$ where $g_{tt}$ = $e^{\nu(R)}$ = $\left( 1-\frac{2M}{R}\right)$. Measurement of the gravitational redshift for a static perfect fluid sphere is not larger than $Z_{s}$ = 2 [@Buchdahl], whereas for an anisotropic fluid sphere this value may be increase up to $Z_{s}$ = 3.84, as given in Ref. [@Ivanov]. We are trying to estimate the surface redshift given table II for different compact star candidates. Table II -------------- ------------------------- -------------------------- ------------------------- ----------------- ---------- Compact star $\rho_{0} (gm/cm^{3})$   $\rho_{R} (gm/cm^{3})$   $p_{c} (dyne/cm^{2})$     $2M/R$    $~~~Z_s$ SMC X-1 1.0710$\times 10^{15} $ 7.3060$\times 10^{14} $ 1.0052$\times 10^{35} $ 0.36959 $<$ 8/9 0.25948 Her X-1 9.0538$\times 10^{14} $ 6.6539$\times 10^{14} $ 6.3783$\times 10^{34}$ 0.30957 $<$ 8/9 0.20348 4U 1538-52 1.0149$\times 10^{15} $ 7.3928$\times 10^{14} $ 7.9036$\times 10^{34}$ 0.32628 $<$ 8/9 0.218346 IV.   Physical features and Comparative study of the physical parameters for compact star model {#iv.physical-features-and-comparative-study-of-the-physical-parameters-for-compact-star-model .unnumbered} =============================================================================================== To proceed further discussion based on the obtained solution that must satisfy some general physical requirements. In order to simplify the analysis and make the solution more viable we explore some physical features of the compact star and carry out a comparative study between the data of the model parameters with a set of astrophysical objects in connection to direct comparison of some strange/compact star candidates. Energy Conditions ----------------- Let us first discuss a very simple but important features of a stellar model. The study of energy conditions within the framework of GR is an essential part for studying the compact objects. Here we examine the energy conditions, namely : (i) Null energy condition (NEC), (ii) Weak energy condition (WEC) and (iii) Strong energy condition (SEC), at all points in an interior of a star holds simultaneously, by the following inequalities $$\begin{aligned} \textbf{NEC:}~ \rho(r)-p_r \geq 0,\\ \textbf{WEC:}~ \rho(r)-p_r(r) \geq 0~~(WEC_r),~~~ \rho(r)-p_t(r) \geq 0~~(WEC_t),\\ \textbf{SEC:}~ \rho(r)-3p_r(r) \geq 0~~(SEC_1), ~~~\rho-p_r(r)-3p_t(r) \geq 0 ~~(SEC_2).\end{aligned}$$ From the above inequalities we provide a graphical representation that one can easily justify the nature of energy condition for three different compact objects in Fig. 3. Due to the complexity of the expression, we only able to write down the above inequalities and plotted the graphs against the above criterions. In fact, all the energy conditions are not violated and well behaved in the stellar interior. {width="4.5cm"} {width="4.5cm"} {width="4.5cm"} Generalized Tolman-Oppenheimer-Volkov Equation ---------------------------------------------- On the other hand, for a given compact star, it is possible to test for hydrostatic equilibrium under the different forces namely gravitational, hydrostatic and anisotropic forces. Nevertheless, in our investigations we need to apply generalized Tolman-Oppenheimer-Volkov (TOV) equation [@Oppenheimer1939; @Leon1993], which is commonly used for an anisotropic fluid distribution is given by $$\begin{aligned} & \qquad\hspace{-1cm}-\frac{M_G(r)(\rho+p_r)}{r^2}e^{\frac{\lambda-\nu}{2}}-\frac{dp_r}{dr}+\frac{2}{r}(p_t-p_r)=0, \label{eq25}\end{aligned}$$ where the effective gravitational mass $M_G(r)$ is defined by $$\begin{aligned} &\qquad\hspace{1cm} M_G(r)=\frac{1}{2}{{r}^{2}}e^{\frac{\nu-\lambda}{2}}\nu'. \label{eq26}\end{aligned}$$ Then equation (\[eq25\]) may be rewritten as $$\begin{aligned} &\qquad\hspace{-1cm}-\frac{\nu'}{2}(\rho+p_r)-\frac{dp_r}{dr}+\frac{2}{r}(p_t-p_r)=0, \label{eq27}\end{aligned}$$ Let us now attempt to explain the Eq. (\[eq27\]) from an equilibrium point of view, which was first shown by Tolman [@Tolman1939] and Oppenheimer and then Volkoff [@Oppenheimer1939], where they predicted the stable equilibrium condition for the compact star as a sum of three different forces, viz. gravitational force ($F_g$), hydrostatics force ($F_h$) and anisotropic force ($F_a$). Thus the above condition assumed the following form, namely $$F_g+F_h+F_a=0. \label{eq28}$$ The components of forces can be expressed in explicit form as: $$\begin{aligned} &\quad\hspace{-1.9cm} F_g= \frac{-\nu'(\rho+p_r)}{2} =-\frac{6ar\,B\,g}{8\,\pi}\,\left[\frac{A\,(b + 2 a r^2) + B\, g~ [b^2 + 5 a^2 r^4 + a\, (3 + 6 b r^2)]} {(1 + b r^2 + a r^4)^2\, [A + B\, (b + a r^2)^{3/2}]^2}\right],\label{eq29}\end{aligned}$$ $$\begin{aligned} &\quad\hspace{-0.4cm} F_h=-\frac{dp_r}{dr} =\frac{a\,r}{(1 + b\,r^2 + a\,r^4)}{\,\left[ \frac{gf^2+9\,a B^2 g^{3}-3\,a\,B\,f}{4\,\pi\,g~~f^2}\right]} -\frac{r\,(2\,a\,r^2+b)}{(1 + b\,r^2 + a\,r^4)^2} \left[\frac{g^2\,f-6\,B\,g}{4\,\pi\,f}\right], \label{eq30}\end{aligned}$$ $$\begin{aligned} &\quad\hspace{-0.5cm} F_a=\frac{2}{r}(p_t-p_r) =\frac{r\,[B\,(b^2+a^2\,r^4)+A\,\sqrt{a\,r^2+b} - a\,B\,(3-2\,b\,r^2)\,(b^2+a^2\,r^4-a+2a\,b\,r^2)\,]}{4\,\pi\,\sqrt{a r^2+b}~~ \left(1 + b r^2 + a r^4\right)^2~~ \left[A + B (b + a r^2)^{3/2}\right]}, \label{eq31}\end{aligned}$$ At this point in the derivation, a stable configuration of the system is counter balance by the components of different forces. To illustrate this in more detail let us introduce an graphical representation as evidenced in Fig. 4. Here, the combine forces of hydrostatic ($F_h$) and anisotropic ($F_a$), dominate the gravitational force ($F_g$), while the anisotropic stress has a less role to the action of equilibrium condition. Hence, following the reason outlined above and making use of Eq. (\[eq28\]), we established the stable configuration model as shown in Fig. 4. {width="4.5cm"} {width="4.5cm"}{width="4.5cm"} Stability Analysis ------------------ In this section we analyze the speed of sound propagation $v^2_s $, is given by the relation $v^2_s = dp/ d\rho$. For a physically interesting stellar geometry should require that the sound speed does not exceed the speed of light, i.e., the physically relevant region is always less than unity. Here we will investigate the speed of sound for anisotropic fluid distribution and propagating along radial as well as transverse direction, should satisfy the bounds $0 < v_{r}^{2}=\frac{dp_r}{d\rho} < 1$ and $0 < v_{t}^{2}=\frac{dp_t}{d\rho} < 1$, [@Herrera(2016)]. We first carry out an analysis of sound velocity with graphical representation. Our result for obtained velocity of sound for compact star with anisotropic matter is presented in Fig. 5, for strange star candidates SMC X-1, Her X-1 and 4U 1538-52. As the resulting expressions are very elaborated, so we only plot the results for different compact stars. It is interesting to note that both $v_r^2,\,v_t^2 < 1$ and monotonic decreasing function, which all show a behavior similar to that found for other compact objects. This is a sufficient condition for the solution to be causal. A notable characteristic of the sound velocity is the estimation of the potentially stable and unstable eras, by considering the expression $0< \Bigl\rvert v_{t}^{2}-v_{r}^{2}\Bigl\rvert \leq 1$, for stable potential [@Andreasson]. According to Fig. 5 (extreme right) our solution gives us the stable star configuration. Our investigation shows that obtained mass function by using the Karmarker condition for strange compact star matter, satisfies both energy and stability conditions {width="4.6cm"} {width="4.6cm"} {width="4.5cm"} Concluding Remarks ================== To summarize, we have studied compact objects supported by anisotropic fluid distribution, which plays an important role in preventing gravitationally collapse. The motivation behind such a construction is that, we established a relation between metric potentials by imposing an embedding theorem, known as embedding class one within the framework of GR. As a result of this approach, we generate a mass function which we have used to study the interior of stellar objects with density decreasing outwards. In next, we have started by deriving the basic equations of Einstein’s field equations that describing the structure of compact objects. In section (II), using the structural equations we derived the energy density, radial and transverse pressures and measure the nature of anisotropy. The complicated expressions given by equations (14-17) are plotted as a function of the radius for our compact stars candidates SMC X-1, Her X-1 and 4U 1538-52. An important feature of this model is the energy density might not vanish at the boundary r =R, though the radial pressure vanishes for all parameter values in according to the boundary condition. Additionally, at the boundary the interior spacetime have been matched by a Schwarzschild metric, and determine the values of arbitrary constants A and B for our compact stars candidates SMC X-1, Her X-1 and 4U 1538-52 in section (III). A comparative study of our results with that of the compact star candidates are provided in Table I and II. Once the mass function is specified, in order to close the system based on physical requirements, we further proceed by investigating the energy conditions, hydrostatic equilibrium under the different forces, and velocity of sound. All the physical properties are well behaved within the stellar radius. We also showed that upper bound of the mass-radius ratio must be less than 8/9 as proposed by Buchdahl [@Buchdahl], for different compact star candidates which we have used for our model. This indicates that the approach adapted in this paper is likely to produce other meaningful models with specific mass function that also greatly help in understanding the properties of other different static compact configurations. In our further study it would be interesting to investigate other forms of mass function that exhibit more general behaviour and thereby describe strange stars related with observational details. Acknowledgments {#acknowledgments .unnumbered} =============== AB is thankful to the authority of Inter-University Centre for Astronomy and Astrophysics, Pune, India for providing research facilities. [99]{} T. Gangopadhyay *et al.*: . J. Lattimer: (2010) http://stellarcollapse.org/nsmasses. B. V. Ivanov: . E. Witten: . N. K. Glendenning, Ch. Kettner & F. Weber: . S. L. Shapiro & S. A. Teukolosky: [*Black Holes, White Dwarfs and Neutron Stars: The Physics of Compact Objects*]{} (Wiley, New York,1983). L. Herrera & W. Barreto: . X.Y. Lai & R.X. Xu: . R.L. Bowers & E. P. T. Liang: . R. Ruderman: . K. Dev & M. Gleiser: . K. Dev & M. Gleiser: . L. Herrera, J. Martin & J. Ospino: . M. Chaichian *et al.*: . A. Perez Martinez, H. Perez Rojas & H. J. Mosquera Cuesta: . E. J. Ferrer *et al.*: . S.K. Maurya *et al.*: . S.K. Maurya *et al.*: . Sk. Monowar Hossein *et al.*: . Mehedi Kalam *et al.*: . M.K. Mak & T. Harko: . R. Sharma & S. D. Maharaj: . Victor Varela *et al.*: . B. Riemann & Abh. K$\ddot{o}$nigl: . L. Schl$\ddot{a}$: . L. Randall & R. Sundrum: ; :. J. Nash: . K. R. Karmarkar: . S. K. Maurya *et al.*: . S. K. Maurya *et al.*: . Ksh. Newton Singh *et al.*: . L. Herrera : . H. A. Buchdahl: ;\ H. A. Buchdahl: . B. V. Ivanov: . J. J. Matese & P. G. Whitman: . M. R. Finch & J. E. F. Skea: . M. K. Mak & T. Harko: . L. Herrera, J. Ospino & A. Di Parisco: . S. K. Maurya, Y.K. Gupta & S. Ray: arXiv: . C. W. Misner & D. H. Sharp: . R.C. Tolman: . J.R. Oppenheimer & G.M. Volkoff: . J. Ponce de Le[ó]{}n: . V. Varela: . J. Devitt & P.S. Florides: . H. Andreasson: .

ArXiv

--- author: - | ,$^{abc}$ Anna Williams$^{d}$ and Sarrvesh S. Sridhar$^{cb}$\ CSIRO Astronomy and Space Science, 26 Dick Perry Avenue, Kensington WA 6151, Australia\ ASTRON, the Netherlands Institute for Radio Astronomy, Postbus 2, 7990 AA, Dwingeloo, The Netherlands\ Kapteyn Astronomical Institute, University of Groningen, PO Box 800, 9700 AV, Groningen, The Netherlands\ Department of Astronomy, University of Wisconsin-Madison, 475 N. Charter Street, Madison, WI, USA\ E-mail: title: 'Magnetic fields on a wide range of scales in star-forming galaxies' --- Introduction {#section:intro} ============ Magnetic fields are a crucial element of galaxies, and play an important role in setting the stage for their structure and evolution. At the scale of active star formation, the role of magnetism is becoming clear through simulations [e.g., @birnboim_etal_2015; @kim_ostriker_2015] and novel observations [@planck_2016]. Simulations are also key to understanding the importance of magnetism in defining the multiphase structure of the interstellar medium [ISM; e.g., @hill_etal_2012]. On larger scales, it is still unclear to what degree magnetic forces play an important dynamical role [see, e.g., @elstner_etal_2014 and references therein], but magnetic field lines are certainly essential for guiding the propagation of cosmic rays [see @kotera_olinto_2011] and thus in understanding the isotropy of ultra-high energy cosmic ray flux as seen from the Earth [e.g., @alvarez_etal_2002]. Magnetic fields may contribute to setting the properties of galactic outflows [see @heesen_etal_2011] and the disk-halo interface [e.g., @benjamin_2002; @henriksen_irwin_2016], and on the largest scales to connecting galaxies to the intergalactic medium [IGM; e.g. @kronberg_etal_1999; @bernet_etal_2013]. A clear understanding of the properties (shape, strength, and energetic content) of galactic magnetic fields on all scales is plainly of great astrophysical interest. Much has been written about the “classical” techniques for detecting and characterising magnetic fields in galaxies. See for example @condon_1992 [@heiles_robishaw_2009; @andersson_2015; @heald_2015; @beck_2016] and references therein. In this contribution, we focus on the detailed study of synchrotron radiation and its polarization, although optical/infrared polarization and Zeeman splitting also provide extremely useful and unique information about galactic magnetic fields. Radio synchrotron measurements can be used as powerful and highly flexible probes of various aspects of the ordered and turbulent components of the magnetic field. It is important to distinguish the ordered/regular components of the field from the turbulent/tangled components. These have different observational signatures. For an excellent summary and illustration, see @jaffe_etal_2010. Crucially, total synchrotron intensity measurements probe the total magnetic field (ordered plus disordered components), while the corresponding polarized emission probes only the ordered (but not necessarily uniformly directed) fields in the plane of the sky; the Faraday rotation experienced by that polarized emission yields estimates of the uniformly directed (regular) field along the line of sight. Polarization is particularly powerful in distinguishing various subtly different configurations of gas, magnetic fields, and cosmic rays within the telescope beam as well as across extended sources. It is in fully understanding and exploiting these detailed measurements that defines the modern era of radio polarimetry. Depolarization, often viewed as a hinderance, is uniquely sensitive to the magnetic field structure through its strong $\lambda^2$ dependence, and provides a detailed probe of magnetism [see, e.g., @horellou_fletcher_2014]. A basic but powerful tool that is often used to try to make sense of broadband polarization measurements is the Rotation Measure (RM) Synthesis technique [@burn_1966; @brentjens_debruyn_2005; @heald_2009]. The concept has this year reached its $50^\mathrm{th}$ anniversary, but it has only been widely used in the last decade. The procedure relies on a Fourier-like relationship between the observed linear polarization quantities (Stokes [*Q*]{} and [*U*]{}), and the amount of polarized emission at each Faraday depth ($\phi$) — a quantity that describes how much Faraday rotation has occurred along the line of sight $l$ (specifically, $\phi\propto\int\,n_e\vec{B}\cdot d\vec{l}$). A deconvolution approach can be used to attempt to isolate the true signal from the instrumental sampling function [[RMCLEAN]{}; @heald_etal_2009]. However, it has been noted that systematic issues can prevent a reliable reconstruction of the input signal when using these techniques [e.g., @farnsworth_etal_2011]. As a consequence, a complementary technique is generally needed: directly fitting the primary broadband [*QU*]{} data to theoretical models [@farnsworth_etal_2011; @osullivan_etal_2012; @sun_etal_2015]. In practice, a combination of these techniques should be used because while RM Synthesis can be used arbitrarily with some important limitations, the appropriate physical model must first be correctly identified for [*QU*]{} fitting to be successful. Indeed, the output of RM Synthesis and [RMCLEAN]{} can be used as a prior for running [*QU*]{} fitting [e.g., @mao_etal_2015], highlighting the complementary nature of the two analytical approaches. The key distinguishing features that define various physical models are encoded in the detailed properties of the resulting broadband depolarization. - : This is a largely $\lambda$-independent effect. It can of course be addressed by observing with improved angular resolution. - : These effects are all dependent on $\lambda^2$, but each has distinct features provided sufficient frequency coverage is utilized. For example, here is a partial list of simple geometries of relevance for the discussion in §\[section:smallscale\]: - [*Differential Faraday rotation (DFR)*]{}: Caused by a regular field in a region that is both emitting and rotating. Emission originating from locations closer to the observer experience less Faraday rotation; thus polarized signal is present over a range of Faraday depth. - [*Internal Faraday dispersion (IFD)*]{}: Turbulent and regular field in a region that is both emitting and rotating. The turbulence also generates depolarization in addition to the DFR. - [*External Faraday dispersion or inhomogeneous Faraday screen (IFS)*]{}: The emitting and turbulent rotating regions are separated, and the rotating screen is turbulent. It may only cover part of the emitting source (partial IFS; PIFS). - [*Extended partial coverage model*]{}: Partial coverage of a depolarizing screen as above (i.e., partial IFS or PIFS), as well as a depolarizing screen associated with the emitting source. Modern radio telescopes are very well suited to distinguish and characterise the corresponding astrophysical situations [@heald_etal_2015]. This contribution will illustrate that depolarization provides a highly useful tool both to learn about detailed small-scale ISM physics (§\[section:smallscale\]), and to ‘dissect’ large-scale magnetic structure (§\[section:largescale\]). Based on the use of standard techniques, as well as the careful interpretation of earlier narrowband data, a consistent and detailed picture of the broad properties of galactic magnetic fields has developed over the course of the past several decades. The general structure is reviewed by @beck_2016. Typically, galaxies demonstrate axisymmetric spiral fields in the disk, and X-shaped vertical extensions when viewed from an edge-on perspective. Field strengths vary but are typically in the range of several to tens of $\mu\mathrm{G}$, with generally similar contributions from ordered and random field components. The magnetic energy density in galaxy ISMs can be comparable to, or even greater than, that of other contributors to the energy budget. One of the areas where new progress is now being made is in developing a better picture of the small scales, understanding the detailed internal structure of the magnetized ISM and the connection between magnetic fields, multiphase gas, and star formation. Meanwhile, various observational techniques are beginning to unveil the properties of magnetic fields and gas at the outskirts of galaxies. Still, many of the observations that have been employed so far probe two-dimensional (2D) projections of the magnetic field structure. By embracing the effects of depolarization, multifrequency observations now allow us to peel the onion of galactic magnetic fields and unveil their 3D structure. Small scale magnetic fields {#section:smallscale} =========================== ![Illustration of the depolarization modeling that is revealing the detailed magnetic field structure in NGC 6946 (Williams et al., in prep). Left: Observed fractional polarization values as a function of $\lambda^2$ in a particular region of the galaxy, along with a model fit to the data (in this case, a PIFS model; see §\[section:intro\]). The top-left panel shows the two linear Stokes parameters [*Q*]{} and [*U*]{}, while the bottom-left panel shows $P=\sqrt{Q^2+U^2}$. Right: Reconstruction of the underlying regular field in NGC 6946, recovered after modeling out the depolarization structure in the ISM.[]{data-label="figure:smallscales"}](mod_pifs_example_fit.png "fig:"){height="0.5\hsize"}![Illustration of the depolarization modeling that is revealing the detailed magnetic field structure in NGC 6946 (Williams et al., in prep). Left: Observed fractional polarization values as a function of $\lambda^2$ in a particular region of the galaxy, along with a model fit to the data (in this case, a PIFS model; see §\[section:intro\]). The top-left panel shows the two linear Stokes parameters [*Q*]{} and [*U*]{}, while the bottom-left panel shows $P=\sqrt{Q^2+U^2}$. Right: Reconstruction of the underlying regular field in NGC 6946, recovered after modeling out the depolarization structure in the ISM.[]{data-label="figure:smallscales"}](n6946bfieldzoom.pdf "fig:"){height="0.5\hsize"} It remains unclear how magnetic fields are organized on small scales ($\lesssim1\,\mathrm{kpc}$) within galaxies. The smallest scales can most readily be observed in the Milky Way, where a turbulent spectrum is observed with characteristic scales reminiscent of the size of [[Hii]{}]{} regions and supernova remnants, depending on location in the disk [e.g., @haverkorn_etal_2008 and references therein]. A number of observational techniques have been brought to bear in external galaxies, including analysis of RMs from background sources [@gaensler_etal_2005] and from the diffuse magnetized ISM itself [@mao_etal_2015], from dispersion of RMs [@fletcher_etal_2011] or polarization angles [@houde_etal_2013], and potentially in the future from fluctuation statistics of the synchrotron intensity [total intensity together with polarized intensity; @herron_etal_2016]. Still, a comprehensive observational picture of the small-scale structure and relevant spatial scales of interstellar magnetic fields has not yet been developed. What are the typical properties of magnetic field fluctuations across the galaxy population, and what are the relative contributions of ordered and random magnetic field components? To what degree are these related to the structure of the gaseous ISM and the star formation activity taking place in the disk? Modern broadband polarimetric observations are now allowing a more detailed study of the magnetized ISM in galaxies. An example is illustrated in Figure \[figure:smallscales\]. By drawing together polarization maps of NGC 6946 at 3, 6, 13, 18, and 22 cm, Williams et al. (in prep) are able to model the structure of the magnetized ISM using simple configurations of the type summarized in §\[section:intro\]. At lower frequencies (i.e., larger values of $\lambda^2$) the data are of sufficiently high spectral resolution and sensitivity that a combination of RM Synthesis and [*QU*]{} fitting is beginning to distinguish depolarization caused within the emitting medium from the effect of a foreground screen. Meanwhile, the spatial resolution ($\sim400$ pc) is approaching the scale of giant molecular clouds (GMCs). New 3D descriptions of the small-scale magnetic fields in external galaxies can now be made. These modeling efforts not only reveal how the magnetoionized ISM varies across the galaxy, but also allows a reconstruction of the underlying large-scale magnetic field shape, as seen in Figure \[figure:smallscales\]. Large scale magnetic fields {#section:largescale} =========================== An interesting development in the past few years has been the use of depolarization as a tool to permit ‘slicing’ galaxies into different depths to investigate the structure of the 3D field in a tomographic manner. For example, @heald_etal_2009 and @braun_etal_2010 have used this approach to first recognize and then interpret generally applicable, large-scale symmetric patterns in the distribution of polarized synchrotron emission from a large number of galaxy disks. Specifically, they noted that for moderately inclined galaxies, a minimum in polarization consistently occurs along the kinematically receding major axis. This feature has been subsequently recognized in other galaxy samples [e.g., @vollmer_etal_2013]. It is illustrated in Figure \[figure:asymmetry\]. Along with an azimuthal variation in polarized emissivity, the galaxies also demonstrate a corresponding azimuthal variation in the Faraday rotation measure. Finally, @braun_etal_2010 report ‘ghost’ polarized emission in some galaxies, at extreme values of RM [but cf. @mao_etal_2015]. ![Illustration of the magnetic field and depolarization model proposed by @braun_etal_2010. Left: Axisymmetric spiral field in the disk. Right: Quadrupolar poloidal field as viewed from an edge-on perspective. Due to the turbulent depolarization in the midplane (induced by active star formation injecting energy into the ISM; [*blue region*]{}), strong depolarization ([*red region*]{}) prevents the observer from detecting polarization at $\sim1\,\mathrm{GHz}$ within and behind the thin disk. Polarized synchrotron radiation ([*purple region*]{}) is visible to such an observer from the front surface of the disk only. That polarized signal may be Faraday rotated by thermal electrons and magnetic fields in the halo ([*green region*]{}).[]{data-label="figure:asymmetry"}](bfieldgeom-faceon.jpg "fig:"){width="0.5\hsize"}![Illustration of the magnetic field and depolarization model proposed by @braun_etal_2010. Left: Axisymmetric spiral field in the disk. Right: Quadrupolar poloidal field as viewed from an edge-on perspective. Due to the turbulent depolarization in the midplane (induced by active star formation injecting energy into the ISM; [*blue region*]{}), strong depolarization ([*red region*]{}) prevents the observer from detecting polarization at $\sim1\,\mathrm{GHz}$ within and behind the thin disk. Polarized synchrotron radiation ([*purple region*]{}) is visible to such an observer from the front surface of the disk only. That polarized signal may be Faraday rotated by thermal electrons and magnetic fields in the halo ([*green region*]{}).[]{data-label="figure:asymmetry"}](bfieldgeom-edgeon.jpg "fig:"){width="0.5\hsize"} The cause of these polarized features is interpreted in the following way. Turbulent depolarization in the midplane, induced by active star formation, serves to depolarize synchrotron radiation from within the disk and to a large degree from the back-side thick disk region as well. (Ghost emission at extreme RM, if present, is interpreted to originate from the back-side of the disk.) In this picture, at $\sim1~\mathrm{GHz}$ frequency, the bulk of the polarized emission originates from a layer just above the star forming thin disk [see also @horellou_etal_1992]. Due to a combination of azimuthal and vertical magnetic field components, the observed polarized asymmetry is expected to be missing from the region where the ordered magnetic field points predominantly toward the observer. @braun_etal_2010 model simple magnetic field configurations and conclude that the observations are consistent with a mixture of an axisymmetric spiral field and a quadrupolar poloidal field. This picture works if galaxies have trailing spiral arms, as indeed seems to be the case [@devaucouleurs_1958]. Galaxies with leading arms would be expected to have a minimum in polarization along the approaching major axis. A further prediction of the model is that the asymmetry should vanish at higher frequency due to the lower impact of turbulent depolarization, as is indeed seen in the case of NGC 6946 [@beck_2007] at wavelengths of 3 and 6 cm. A rather more detailed application of the same basic analysis has been performed for the individual case of M 51 [@fletcher_etal_2011]. By drawing together a large multifrequency data set spanning wavelengths from 3 to 20 cm, the authors were able to consider separately the thermal electron population and magnetic field properties in the disk and halo regions. Ultimately they found that different spiral modes were present in the magnetic fields of the disk and halo ($m=0+2$ in the disk, and $m=1$ in the halo; note that the qualitative appearance of the RM maps presented by @heald_etal_2009 are consistent both with this model and with the interpretation of the $\sim1\,\mathrm{GHz}$ emission originating from above the disk). Low-frequency ($\nu\lesssim300\,\mathrm{MHz}$) synchrotron radiation probes even farther into the outskirts of galaxies. While synchrotron and inverse Compton energy losses mean that cosmic rays no longer emit radiation at high frequencies once they have propagated far from regions of star formation, the low-energy cosmic ray population loses energy more slowly [$\dot{E}\propto\,E^2$; @condon_1992] and thus emits at low radio frequency for a longer span of time, i.e. $\tau\propto\,E^{-1}$. The new generation of low frequency radio telescopes is in this way opening the window to the outer parts of galactic magnetic fields. For example, total intensity LOFAR observations have been used to probe into the far outer disk [e.g., in M51; @mulcahy_etal_2014] and into the upper disk-halo interface region (e.g., NGC 5775; Heald et al. in prep, see Figure \[figure:lofar\]). One ultimate aim for this kind of research is to start to recover polarization at such large distances from the star forming disk, in order to further constrain the [*ordered*]{} component of the large-scale field (and possibly also the regular component from associated RMs). ![Examples of nearby galaxies viewed at low radio frequency (here, $150\,\mathrm{MHz}$) with high angular resolution ($\sim20^{\prime\prime}$) and excellent sensitivity ($<1\,\mathrm{mJy\,beam}^{-1}$) using LOFAR. Left: Wide field of view image surrounding the nearby galaxy M51 [@mulcahy_etal_2014]. The inset shows a DSS image of the galaxy overlaid with LOFAR contours, starting at $1\,\mathrm{mJy\,beam}^{-1}$ and increasing by factors of 1.5. Right: False-color SDSS composite of the nearby edge-on galaxy NGC 5775 (Heald et al., in prep) and its vertically extended radio halo (orange colors), visible with LOFAR extending up to 15 kpc from the star forming disk. The companion galaxy NGC 5774 is also detected, as is a bridge of radio emission corresponding to an [[Hi]{}]{} bridge [see @irwin_1994].[]{data-label="figure:lofar"}](Whirlpool_Galaxy.jpg "fig:"){height="0.425\hsize"}![Examples of nearby galaxies viewed at low radio frequency (here, $150\,\mathrm{MHz}$) with high angular resolution ($\sim20^{\prime\prime}$) and excellent sensitivity ($<1\,\mathrm{mJy\,beam}^{-1}$) using LOFAR. Left: Wide field of view image surrounding the nearby galaxy M51 [@mulcahy_etal_2014]. The inset shows a DSS image of the galaxy overlaid with LOFAR contours, starting at $1\,\mathrm{mJy\,beam}^{-1}$ and increasing by factors of 1.5. Right: False-color SDSS composite of the nearby edge-on galaxy NGC 5775 (Heald et al., in prep) and its vertically extended radio halo (orange colors), visible with LOFAR extending up to 15 kpc from the star forming disk. The companion galaxy NGC 5774 is also detected, as is a bridge of radio emission corresponding to an [[Hi]{}]{} bridge [see @irwin_1994].[]{data-label="figure:lofar"}](comboRGBimagebb.jpg "fig:"){height="0.425\hsize"} At the very largest distance from the star forming regions of galaxies, where cosmic rays have lost too much energy to be detectable in synchrotron radiation with current technology, different methods must be brought to bear to identify and study the extended magnetic field structure. The Universe provides excellent probes of galactic magnetic fields on the largest scales: background polarized radio galaxies [as have been used to study the magnetic extent of galaxy clusters; @clarke_etal_2001]. For galaxies, a similar approach has led @bernet_etal_2013 to find evidence for strong magnetic fields (tens of $\mu$G) extending up to 50 kpc from galaxies, and with properties suggestive of galactic outflows rather than large coherent magnetic structures. Follow-up studies with forthcoming radio surveys will be of particular interest in the next several years. Future prospects ================ While recent progress in our understanding of magnetism has been rapid and very encouraging, there is also still clearly much to learn using future radio telescopes and accompanying large-scale survey efforts. For example, key questions remain surrounding the buildup and evolution of magnetism in the Universe and in bound objects like galaxies; the detailed internal structure of magnetic fields on both small and large scales in galaxies; and the interplay between magnetic fields and the gaseous content and kinematics in galaxies. In this section, we provide a brief overview of some of the radio surveys that are currently underway and will be pursued with the next generation of observatories, leading to the SKA. One interesting survey that is currently enhancing our view of the magnetic content of nearby galaxies is called “Continuum HAlos in Nearby Galaxies – an EVLA Survey” [CHANG-ES; @irwin_etal_2012a]. This project makes use of sensitive broadband continuum observations of a large number (35) of edge-on nearby star-forming galaxies, in full polarization, at L- and C-bands ($1.2-1.9$ and $5-7\,\mathrm{GHz}$) using the Jansky Very Large Array (VLA), or EVLA as it was known at the time. From these observations a clear picture of the vertical extent of the magnetic component of nearby galaxies is starting to be revealed [e.g., @irwin_etal_2012b; @wiegert_etal_2015]. Surveys of this type are set to become commonplace as improvements in the sensitivity, field of view, and instantaneous bandwidth of radio telescopes are ubiquitous. Two aspects that will need continued care for future efforts are ([*i*]{}) instrumental polarization, not just at the pointing center but across the whole field of view (requiring control of, and excellent understanding of, the beamshape), and ([*ii*]{}) characterization of the ionospheric contribution to Faraday rotation measure. As we look toward the horizon, the community is greatly anticipating the next generation of radio telescope facilities, amongst these the Australian Square Kilometre Array Pathfinder [ASKAP; @johnston_etal_2008] and the South African MeerKAT [formerly the Karoo Array Telescope; @jonas_2009]. Both of these will incorporate polarimetric observations either as dedicated surveys or as fundamental aspects of larger survey activities. These new telescopes promise to deliver both excellent polarization performance as well as the broad bandwidth that is needed to characterize the internal structure of the astrophysical sources of interest. Particular surveys of note include the Polarization Sky Survey of the Universe’s Magnetism [POSSUM; @gaensler_2009] with the ASKAP telescope, which will develop a catalog containing an estimated $\sim3$ million polarized sources over a surveyed sky area of $30,000$ square degrees. By covering the frequency range $1150-1450\,\mathrm{MHz}$, accurate RMs will be available for a large number of those sources, leading to an “RM grid” of POSSUM sources with a density as high as 100 RMs sq deg$^{-1}$. The catalog will allow large statistical studies of polarized source populations, as well as enabling investigations of the extended magnetic structure of intervening objects along the line of sight (e.g. the Milky Way, nearby galaxies and clusters). In South Africa, the MeerKAT [[Hi]{}]{} Observations of Nearby Galactic Objects: Observing Southern Emitters (MHONGOOSE)[^1] survey will probe the magnetic fields in 30 nearby galaxies to exquisite depth (200 hr per target) and with broadband frequency coverage. Meanwhile the MeerKAT International GigaHertz Tiered Extragalactic Exploration (MIGHTEE)[^2] survey will observe broadband continuum emission in a tiered approach, covering 1000 square degrees to a depth of $\sim5\,\mu\mathrm{Jy\,beam}^{-1}$ along with two smaller and deeper sky areas, enabling great strides in understanding the cosmic evolution of magnetism in distant AGN and star-forming galaxies. On somewhat longer timescales, the community is working toward transformational breakthroughs in several facets of cosmic magnetism science with the SKA. Several headline science projects are being developed, with the highest priority embodied in a large-area, deep, broadband survey for Faraday rotation measures [@johnston-hollitt_etal_2015]. Additional science cases include those focused on nearby galaxies [@heald_etal_2015; @beck_etal_2015], the Milky Way galaxy [@haverkorn_etal_2015], detailed broadband studies of radio galaxies [@gaensler_etal_2015], and intermediate-redshift normal and starforming galaxies [@taylor_etal_2015]. There clearly remain a great number of outstanding questions of fundamental importance that will be pursued in the coming decades of radio surveys. [99]{} Alvarez-Mu[ñ]{}iz, J., Engel, R., & Stanev, T. *Ultrahigh-Energy Cosmic-Ray Propagation in the Galaxy: Clustering versus Isotropy*, *ApJ* [**572**]{} (2002) 185 \[arXiv:astro-ph/0112227\] Andersson, B.-G.  *Interstellar Grain Alignment: Observational Status*, in *Magnetic Fields in Diffuse Media*, Springer-Verlag Berlin, Heidelberg 2015, p. 59 Beck, R. *Magnetism in the spiral galaxy NGC 6946: magnetic arms, depolarization rings, dynamo modes, and helical fields*, *A&A* [**470**]{} (2007) 539 \[arXiv:0705.4163\] Beck, R., et al.  *Structure, dynamical impact and origin of magnetic fields in nearby galaxies in the SKA era*, in proceedings of *Advancing Astrophysics with the Square Kilometre Array* (2015) \[arXiv:1501.00385\] Beck, R. *Magnetic fields in spiral galaxies*, *A&ARv* [**24**]{} (2016) 4 \[arXiv:1509.04522\] Benjamin, R. A. *The Interstellar Disk-Halo Connection: The Rotation of Extra-planar Gas*, *ASPC* [**276**]{} (2002) 201 Bernet, M. L., Miniati, F., & Lilly, S. J. *The Extent of Magnetic Fields around Galaxies out to z \~ 1*, *ApJ* [**772**]{} (2013) L28 \[arXiv:1307.2250\] Birnboim, Y., Balberg, S., & Teyssier, R. *Galaxy evolution: modelling the role of non-thermal pressure in the interstellar medium*, *MNRAS* [**447**]{} (2015) 3678 \[arXiv:1311.1206\] Braun, R., Heald, G., & Beck, R. *The Westerbork SINGS survey. III. Global magnetic field topology*, *A&A* [**514**]{} (2010) A42 \[arXiv:1002.1776\] Brentjens, M. A., & de Bruyn, A. G. *Faraday rotation measure synthesis*, *A&A* [**441**]{} (2005) 1217 \[arXiv:astro-ph/0507349\] Burn, B. J. *On the depolarization of discrete radio sources by Faraday dispersion*, *MNRAS* [**133**]{} (1966) 67 Clarke, T. E., Kronberg, P. P., & B[ö]{}hringer, H. *A New Radio-X-Ray Probe of Galaxy Cluster Magnetic Fields*, *ApJ* [**547**]{} (2001) L111 \[arXiv:astro-ph/0011281\] Condon, J. J. *Radio emission from normal galaxies*, *ARA&A* [**30**]{} (1992) 575 de Vaucouleurs, G.  *Tilt Criteria and Direction of Rotation of Spiral Galaxies.*, *ApJ* [**127**]{} (1958) 487 Elstner, D., Beck, R., & Gressel, O. *Do magnetic fields influence gas rotation in galaxies?*, *A&A* [**568**]{} (2014) A104 \[arXiv:1407.6998\] Farnsworth, D., Rudnick, L., & Brown, S. *Integrated Polarization of Sources at [$\lambda$]{} \~ 1 m and New Rotation Measure Ambiguities*, *AJ* [**141**]{} (2011) 191 \[arXiv:1103.4149\] Fletcher, A., Beck, R., Shukurov, A., Berkhuijsen, E. M., & Horellou, C. *Magnetic fields and spiral arms in the galaxy M51*, *MNRAS* [**412**]{} (2011) 2396 \[arXiv:1001.5230\] Gaensler, B. M., Haverkorn, M., Staveley-Smith, L., Dickey, J. M., McClure-Griffiths, N. M., Dickel, J. R., & Wolleben, M. *The Magnetic Field of the Large Magellanic Cloud Revealed Through Faraday Rotation*, *Science* [**307**]{} (2005) 1610 \[arXiv:astro-ph/0503226\] Gaensler, B. M. *Cosmic magnetism with the Square Kilometre Array and its pathfinders*, *IAUS* [**259**]{} (2009) 645 \[arXiv:0901.2952\] Gaensler, B., et al.  *Broadband Polarimetry with the Square Kilometre Array: A Unique Astrophysical Probe*, in proceedings of *Advancing Astrophysics with the Square Kilometre Array* (2015) \[arXiv:1501.00626\] Haverkorn, M., Brown, J. C., Gaensler, B. M., & McClure-Griffiths, N. M. *The Outer Scale of Turbulence in the Magnetoionized Galactic Interstellar Medium*, *ApJ* [**680**]{} (2008) 362-370 \[arXiv:0802.2740\] Haverkorn, M., et al. *Measuring magnetism in the Milky Way with the Square Kilometre Array*, in proceedings of *Advancing Astrophysics with the Square Kilometre Array* (2015) \[arXiv:1501.00416\] Heald, G. *The Faraday rotation measure synthesis technique*, in proceedings of *Cosmic Magnetic Fields: From Planets, to Stars and Galaxies* [**259**]{} (2009) 591 Heald, G., Braun, R., & Edmonds, R. *The Westerbork SINGS survey. II Polarization, Faraday rotation, and magnetic fields*, *A&A* [**503**]{} (2009) 409 \[arXiv:0905.3995\] Heald, G. *Synchrotron Radiation and Faraday Rotation*, in *Magnetic Fields in Diffuse Media*, Springer-Verlag Berlin, Heidelberg 2015, p. 41 Heald, G., et al.  *Magnetic Field Tomography in Nearby Galaxies with the Square Kilometre Array*, in proceedings of *Advancing Astrophysics with the Square Kilometre Array* (2015) \[arXiv:1501.00408\] Heesen, V., Beck, R., Krause, M., & Dettmar, R.-J. *Cosmic rays and the magnetic field in the nearby starburst galaxy NGC 253 III. Helical magnetic fields in the nuclear outflow*, *A&A* [**535**]{} (2011) A79 \[arXiv:1109.0255\] Heiles, C., & Robishaw, T. *Zeeman splitting in the diffuse interstellar medium-The Milky Way and beyond*, *IAUS* [**259**]{} (2009) 579 Henriksen, R. N., & Irwin, J. A. *Magnetized galactic halos and velocity lags*, *MNRAS* in press (2016) \[arXiv:1603.02973\] Herron, C. A., Burkhart, B., Lazarian, A., Gaensler, B. M., & McClure-Griffiths, N. M. *Radio Synchrotron Fluctuation Statistics as a Probe of Magnetized Interstellar Turbulence*, *ApJ* in press (2016) \[arXiv:1603.02751\] Hill, A. S., Joung, M. R., Mac Low, M.-M., Benjamin, R. A., Haffner, L. M., Klingenberg, C., & Waagan, K. *Vertical Structure of a Supernova-driven Turbulent, Magnetized Interstellar Medium*, *ApJ* [**750**]{} (2012) 104 \[arXiv:1202.0552\] Horellou, C., Beck, R., Berkhuijsen, E. M., Krause, M., & Klein, U. *Faraday effects in the spiral galaxy M 51*, *A&A* [**265**]{} (1992) 417 Horellou, C., & Fletcher, A. *Magnetic field tomography, helical magnetic fields and Faraday depolarization*, *MNRAS* [**441**]{} (2014) 2049 \[arXiv:1401.4152\] Houde, M., Fletcher, A., Beck, R., Hildebrand, R. H., Vaillancourt, J. E., & Stil, J. M. *Characterizing Magnetized Turbulence in M51*, *ApJ* [**766**]{} (2013) 49 \[arXiv:1301.7508\] Irwin, J. A. *Arcs and bridges in the interacting galaxies NGC 5775/NGC 5774*, *ApJ* [ **429**]{} (1994) 618 Irwin, J., et al.  *Continuum Halos in Nearby Galaxies: An EVLA Survey (CHANG-ES). I. Introduction to the Survey*, *AJ* [**144**]{} (2012) 43 \[arXiv:1205.5694\] Irwin, J., et al.  *Continuum Halos in Nearby Galaxies: An EVLA Survey (CHANG-ES). II. First Results on NGC 4631*, *AJ* [**144**]{} (2012) 44 \[arXiv:1205.5771\] Jaffe, T. R., Leahy, J. P., Banday, A. J., Leach, S. M., Lowe, S. R., & Wilkinson, A. *Modelling the Galactic magnetic field on the plane in two dimensions*, *MNRAS* [**401**]{} (2010) 1013 \[arXiv:0907.3994\] Johnston, S., et al.  *Science with ASKAP. The Australian square-kilometre-array pathfinder*, *Experimental Astronomy* [**22**]{} (2008) 151 \[arXiv:0810.5187\] Johnston-Hollitt, M., et al. *Using SKA Rotation Measures to Reveal the Mysteries of the Magnetised Universe*, in proceedings of *Advancing Astrophysics with the Square Kilometre Array* (2015) \[arXiv:1506.00808\] Jonas, J. L. *MeerKAT - The South African Array With Composite Dishes and Wide-Band Single Pixel Feeds*, *IEEEP* [**97**]{} (2009) 1522 Kim, C.-G., & Ostriker, E. C. *Vertical Equilibrium, Energetics, and Star Formation Rates in Magnetized Galactic Disks Regulated by Momentum Feedback from Supernovae*, *ApJ* [**815**]{} (2015) 67 \[arXiv:1511.00010\] Kotera, K., & Olinto, A. V. *The Astrophysics of Ultrahigh-Energy Cosmic Rays*, *ARA&A* [**49**]{} (2011) 119 \[arXiv:1101.4256\] Kronberg, P. P., Lesch, H., & Hopp, U. *Magnetization of the Intergalactic Medium by Primeval Galaxies*, *ApJ* [**511**]{} (1999) 56 Mao, S. A., Zweibel, E., Fletcher, A., Ott, J., & Tabatabaei, F. *Properties of the Magneto-ionic Medium in the Halo of M51 Revealed by Wide-band Polarimetry*, *ApJ* [**800**]{} (2015) 92 \[arXiv:1412.8320\] Mulcahy, D. D., et al. *The nature of the low-frequency emission of M 51. First observations of a nearby galaxy with LOFAR*, *A&A* [**568**]{} (2014) A74 \[arXiv:1407.1312\] O’Sullivan, S. P., et al. *Complex Faraday depth structure of active galactic nuclei as revealed by broad-band radio polarimetry*, *MNRAS* [**421**]{} (2012) 3300 \[arXiv:1201.3161\] Planck Collaboration, et al. *Planck intermediate results. XXXV. Probing the role of the magnetic field in the formation of structure in molecular clouds*, *A&A* [**586**]{} (2016) A138 \[arXiv:1502.04123\] Sun, X. H., et al.  *Comparison of Algorithms for Determination of Rotation Measure and Faraday Structure. I. 1100-1400 MHz*, *AJ* [**149**]{} (2015) 60 \[arXiv:1409.4151\] Taylor, R., et al.  *SKA Deep Polarization and Cosmic Magnetism*, in proceedings of *Advancing Astrophysics with the Square Kilometre Array* (2015) \[arXiv:1501.02298\] Vollmer, B., Soida, M., Beck, R., Chung, A., Urbanik, M., Chy[ż]{}y, K. T., Otmianowska-Mazur, K., & Kenney, J. D. P. *Large-scale radio continuum properties of 19 Virgo cluster galaxies. The influence of tidal interactions, ram pressure stripping, and accreting gas envelopes*, *A&A* [**553**]{} (2013) A116 \[arXiv:1304.1279\] Wiegert, T., et al.  *CHANG-ES. IV. Radio Continuum Emission of 35 Edge-on Galaxies Observed with the Karl G. Jansky Very Large Array in D ConfigurationData Release 1*, *AJ* [**150**]{} (2015) 81 \[arXiv:1508.05153\] [^1]: See <http://mhongoose.astron.nl/> [^2]: See <http://public.ska.ac.za/meerkat/meerkat-large-survey-projects>

ArXiv

--- abstract: 'In this paper we consider completed coverings that are branched coverings in the sense of Fox. For completed coverings between PL manifolds we give a characterization of the existence of a monodromy representation and the existence of a locally compact monodromy representation. These results stem from a characterization for the discreteness of a completed normal covering. We also show that completed coverings admitting a monodromy representations are discrete and that the image of the branch set is closed.' address: 'Department of Mathematics and Statistics, P.O.Box 68, 00014 Univeristy of Helsinki' author: - Martina Aaltonen bibliography: - 'viite.bib' title: Monodromy representations of completed coverings --- Introduction ============ By the classical theory of covering spaces, a covering map $f \colon X\to Z$ between manifolds is a factor of a normal (or regular) covering; there exists a normal covering $p \colon Y {\rightarrow}X$ so that $q = f \circ p \colon Y \to Z$ is a normal covering and the deck-transformation group of $q$ is isomorphic to the monodromy group of $f,$ $$\label{rr} \xymatrix{ & Y \ar[ld]_p \ar[rd]^q &\\ X \ar[rr]^f & & Z. }$$ In this case, the monodromy group $G$ of $f$ acts on $Y$ and there exists a subgroup $H \subset G$ for which $Y/G \approx Z$ and $Y/H \approx X.$ The normal coverings $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ are orbit maps. In this article we are interested in ramifications of this construction for discrete open mappings $f \colon X {\rightarrow}Z$ between manifolds. By [Č]{}ernavski  [@CER] and Väisälä [@V] a discrete and open mapping $f \colon X {\rightarrow}Z$ between manifolds is almost a local homeomorphism in the following sense. Let $B_f \subset X$ be the branch set i.e.the set of points in $X,$ where $f$ is not a local homeomorphism. By the theorem of [Č]{}ernavski and Väisälä the topological codimension of $B_f$ is at least $2.$ In fact, if we set $Z':=Z \setminus f(B_f)$ and $X':=X \setminus f^{-1}(f(B_f)),$ then $X' \subset X$ is a dense and connected subset of $X$ and $Z' \subset Z$ is a dense and connected subset of $Z$ and $g:=f {|}X' \colon X' {\rightarrow}Z'$ is a local homeomorphism. Suppose now in addition that $Z' \subset Z$ is open and $g$ is a covering. By the classical argument above there exists an open manifold $Y'$ and a commutative diagram of discrete and open mappings $$\label{hei} \xymatrix{ & Y' \ar[ld]_{p'} \ar[rd]^{q'} &\\ X \ar[rr]^f & & Z }$$ where $p' \colon Y' {\rightarrow}X'$ and $q' \colon Y' {\rightarrow}Z'$ are normal coverings, the deck-transformation group of the covering $q' \colon Y' {\rightarrow}Z'$ is isomorphic to the monodromy group of $g$ and $q'=f \circ p'.$ It becomes a question, whether there exists a space $Y \supset Y'$ so that $p'$ and $q'$ extend to discrete orbit maps $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ satisfying $q=f \circ p.$ We formalize this question as follows. \[kysymys\]Suppose $f \colon X {\rightarrow}Z$ is an open discrete mapping between manifolds so that $Z \setminus f(B_f) \subset Z$ is open and $$f {|}\big(X \setminus f^{-1}(f(B_f))\big) \colon \big(X \setminus f^{-1}(f(B_f))\big) {\rightarrow}\big(Z \setminus f(B_f)\big)$$ is a covering. Does there exists a locally connected Hausdorff space $Y,$ an embedding $\iota \colon Y' {\rightarrow}Y$ and orbit maps $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ so that $$\label{ww} \xymatrix{ & Y \ar[ld]_p \ar[rd]^q &\\ X \ar[rr]^f & & Z}$$ is a commutative diagram of discrete and open mappings satisfying $p'= p \circ \iota$ and $q'=q \circ \iota,$ where $p'$ and $q'$ are as in . Further, we require that $\iota(Y')\subset Y$ is a dense subset and $Y \setminus \iota(Y')$ does not locally separate $Y.$ This question stems from an article of Berstein and Edmonds [@BE] where they show that for open and discrete mappings between compact manifolds the answer to Question \[kysymys\] is positive and the orbit maps $p$ and $q$ are induced by the action of the monodromy group of $f$ on $Y.$ They use this construction to give degree estimates for simplicial maps between compact manifolds; see also Pankka-Souto [@PS] for another application. In this article we extend the construction of Berstein and Edmonds [@BE 2.2Prop.] for completed coverings (Definition \[cc\]) that are branced coverings in the sense of Fox [@F]. Completed coverings form a subclass of open and surjective mappings from a Hausdorff space onto a manifold. Examples of completed coverings between manifolds are discrete and open mappings between compact manifolds (see [@CH]), proper quasiregular mappings (see [@R]) and surjective open and discrete simplicial mappings between PL manifolds (see [@I]). The class of completed coverings also include mappings outside these classes of mappings (see [@DP] and [@L]). The definition of a completed covering is technical and it comes from the theory of complete spreads in Fox [@F]. We only mention here, that the class of completed coverings is sufficiently large and natural in our setting. Edmonds’ results on orbit maps with finite multiplicity [@E] ensures that the answer to Question \[kysymys\] is known for completed coverings that have finite multiplicity: By Edmods’ result completed normal coverings that have finite multiplicity are discrete orbit maps; see [@E Thm.4.1]. Based on this the argument of Berstein and Edmonds [@BE] relies on the observation that a discrete and open mapping $f \colon X {\rightarrow}Z$ between compact manifolds is a completed covering that has a finite monodromy group. For completed coverings the finiteness of the monodromy group is a sufficient condition for the argument of Bernestein and Edmonds. Thus, by the finiteness of the monodromy group, the answer to Question \[kysymys\] is positive for every completed covering $f \colon X {\rightarrow}Z$ between manifolds that has finite multiplicity. On the other hand, for Question \[kysymys\] to have a positive answer for a mapping $f \colon X {\rightarrow}Z$ having infinite multiplicity, it is not enough to assume that $f$ is a completed covering. The underlying reason is that the monodromy group of $f$ is in this case infinite, and a completed normal covering $p \colon Y {\rightarrow}Z$ with infinite multiplicity is not necessary discrete; see Montesinos [@MA Ex.10.6]. To answer Question \[kysymys\], our first main Theorem is the following. \[hilu\]A completed normal covering $p \colon Y {\rightarrow}Z$ from a Hausdorff space $Y$ onto a manifold $Z$ is an orbit map if and only if $p$ is a discrete map. For this reason in the heart of Question \[kysymys\] is the characterization of discrete completed normal coverings. In this direction we obtain the following results. \[ilta\] A completed normal covering $p \colon Y {\rightarrow}Z$ from a Hausdorff space $Y$ onto a PL manifold $Z$ is discrete if and only if $p$ is a stabily completed normal covering. Stabily completedness is defined in Section \[dcc\]. By Montesinos a completed normal covering $p \colon Y {\rightarrow}Z$ from a Hausdorff space $Y$ onto a manifold $Z$ is discrete if it has locally finite multiplicity; see [@MA Thm. 9.14]. We prove the following. \[kuuskytkolme\]Let $p \colon Y {\rightarrow}Z$ be a discrete completed normal covering from a Hausdorff space $Y$ onto a PL manifold $Z.$ Then $Y$ is locally compact if and only if $p$ has locally finite multiplicity. [**Regularity and existence of monodromy representations.**]{} We begin with a result showing the naturality of completed coverings. We show that if the space $Y$ in diagram is locally compact and locally connected, then all the maps in the diagram are completed coverings. \[sade\] Suppose $f \colon X {\rightarrow}Z$ is a discrete and open mapping between manifolds, $X':= X \setminus f^{-1}(f(B_f))$ and $Z':= Z \setminus f(B_f)$ for the branch set $B_f$ of $f.$ Suppose $Z'\subset Z$ is open and $f {|}X' \colon X' {\rightarrow}Z'$ is a covering. Let $$\xymatrix{ & Y' \ar[ld]_{p'} \ar[rd]^{q'} &\\ X \ar[rr]^f & & Z }$$ be a commutative diagram of discrete and open mappings so that $p' \colon Y' {\rightarrow}X'$ and $q' \colon Y' {\rightarrow}Z'$ are normal coverings. Suppose $Y$ is a locally compact and locally connected Hausdorff space so that there exists an embedding $\iota \colon Y' {\rightarrow}Y$ and orbit maps $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ for which $$\xymatrix{ & Y \ar[ld]_p \ar[rd]^q &\\ X \ar[rr]^f & & Z }$$ is a commutative diagram of discrete and open mappings satisfying $p'= p \circ \iota$ and $q'=q \circ \iota,$ and so that $\iota(Y')\subset Y$ is dense and $Y \setminus \iota(Y')$ does not locally separate $Y.$ Then $f \colon X {\rightarrow}Z,$ $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ are completed coverings. Let $X$ and $Z$ be manifolds and $f \colon X {\rightarrow}Z$ a completed covering. We say that a triple $(Y,p,q)$ is *a monodromy representation* of $f$ if $Y$ is a locally connected Hausdorff space and the monodromy group $G$ of $f$ has an action on $Y$ and a subgroup $H \subset G$ so that $Y/G \approx Z,$ $Y/H \approx X$ and the associated orbit maps $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ are completed coverings satisfying $q=p \circ f.$ We call a monodromy representation $(Y,p,q)$ *locally compact* if $Y$ is locally compact. We present the following four regularity results for monodromy representations; \[ll\] Let $f \colon X {\rightarrow}Z$ be a completed covering between manifolds. Suppose $(Y,p,q)$ is a monodromy representation of $f.$ Then $f,$ $p$ and $q$ are discrete maps. \[yopo\] Let $f \colon X {\rightarrow}Z$ be a completed covering between manifolds. Suppose $(Y,p,q)$ is a monodromy representation of $f$ and $B_f \subset X,$ $B_p \subset Y$ and $B_q \subset Y$ the respective branch sets. Then $f(B_f) \subset Z,$ $p(B_p) \subset X$ and $q(B_q) \subset Z$ are closed sets. \[mm\] Let $f \colon X {\rightarrow}Z$ be a completed covering between PL manifolds. Suppose $(Y,p,q)$ is a monodromy representation of $f.$ Then $f,$ $p$ and $q$ are stabily completed coverings. \[kk\] Let $f \colon X {\rightarrow}Z$ be a completed covering between PL manifolds. Suppose $(Y,p,q)$ is a locally compact monodromy representation of $f.$ Then for every polyhedral path-metric $d_s$ on $Z$ there exists a path-metric $d_s^*$ on $Y,$ so that - the topology induced by $d_s^*$ coincides with the topology of $Y$, - $(Y,d_s^*)$ is a locally proper metric space, - $q \colon (Y,d_s^*) {\rightarrow}(Z,d_s)$ is a $1$-Lipschitz map, and - the action of the monodromy group of $f$ in the monodromy representation $(Y,p,q)$ is by isometries on $(Y,d_s^*).$ We note that the path metric $d_s^*$ in Theorem \[kk\] is a pullback of the path metric $d_s$ in $Z.$ Thus the map $q \colon (Y,d_s^*) {\rightarrow}(Z,d_s)$ is a $1$-BLD map. Further, if $f \colon (X,e_s) {\rightarrow}(Z,d_s)$ is a $L$-BLD map for a path metric $e_s$ on $X,$ then $p \colon (Y,d^*_s) {\rightarrow}(X,e_s)$ is a $L$-BLD map. A mapping $f \colon X {\rightarrow}Z$ between length manifolds is $L$-BLD for $L \geq 1$ if $f$ is discrete and open and satisfies $$\frac{1}{L} \ell(\gamma) \leq \ell(f \circ \gamma) \leq L \ell(\gamma)$$ for every path $\gamma$ in $X,$ where $\ell(\cdot)$ is the length of a path. We refer to [@HR] for a detailed discussion of BLD-maps. For the existence of a monodromy representations, we have the following characterization. Together with Theorem \[ll\], this answers to Question \[kysymys\] in the context of completed coverings between PL manifolds. \[relas\] A completed covering $f \colon X {\rightarrow}Z$ between PL manifolds has a monodromy representation $(Y,p,q)$ if and only if $f$ is *stabily completed*. As a corollary of Theorem \[relas\] open and surjective simplicial mappings between PL manifolds have monodromy representations. Another corollary is the following. Let $f \colon X {\rightarrow}Z$ be a $L$-BLD mapping between PL $2$-manifolds. Then $f$ has a monodromy representation if and only if $f(B_f) \subset Z$ is a discrete set. Indeed, if $f(B_f)\subset Z$ is a discrete set, then the BLD-mapping $f$ is a completed covering by Luisto [@L], and hence $f$ is a stabily completed covering. Then $f$ has a monodromy representation by Theorem \[relas\]. If $f(B_f) \subset Z$ is not a discrete set, then $f(B_f) \subset Z$ it not a closed set by Stoilow’s Theorem. Thus, by Theorem \[yopo\], $f$ has no monodromy representation. We characterize the existence of locally compact monodromy representations as follows. Local monodromy groups are defined in Definition \[br\]. \[hopo\] A completed covering $f \colon X {\rightarrow}Z$ between PL manifolds has a locally compact monodromy representation $(Y,p,q)$ if and only if $f$ is stabily completed and $f$ has a finite local monodromy group at each point of $Z.$ We show that not every monodromy representation is locally compact. However, our results combined with previous results by Fox, Edmonds and Montesinos have an easy corollary in the positive direction. For the following statement we say that a covering $g \colon X' {\rightarrow}Z'$ is *virtually normal* if $g_*(\pi(X',x_0))\subset \pi(Z',z_0)$ contains a finite index subgroup which is normal in $\pi(Z',z_0).$ \[ilo\] Let $f \colon X {\rightarrow}Z$ be a completed virtually normal covering between PL manifolds. Then $f$ has a locally compact monodromy representation $(Y,p,q),$ where $p$ has finite multiplicity. By Fox [@F] we find a completed normal covering $p \colon Y {\rightarrow}Z$ so that $p$ has finite multiplicity and $q:=f \circ p$ is a completed normal covering. Since $p$ has finite multiplicity, $p$ is an orbit map by Edmonds [@E Thm.4.1] and $q$ has locally finite multiplicity. Since $q$ has locally finite multiplicity, the map $q$ is a discrete map by Montesinos [@MA Thm. 9.14]. Thus $q$ is an orbit map by Theorem \[hilu\] and $(Y,p,q)$ a monodromy representation of $f$. Since $q \colon Y {\rightarrow}Z$ is a completed normal covering onto a PL manifold that is an orbit map and that has locally finite multiplicity, the space $Y$ is by Theorem \[kuuskytkolme\] locally compact. Thus the monodromy representation $(Y,p,q)$ is locally compact. This article is organized as follows. In Sections \[nc\] and \[mc\] we discuss monodromy of covering maps. In Sections \[sp\] and \[ccc\] we introduce completed coverings. In Section \[os\] we study the relation between orbit maps and completed normal coverings and prove Theorem \[sade\]. In Section \[ooo\] we prove Theorem \[ll\] by showing that competed normal coverings are orbit maps if and only if they are discrete maps. In Section \[ando\] we prove Theorem \[yopo\]. In Section \[dcc\] we prove Theorems \[mm\] and \[relas\] by showing that a completed normal covering onto a PL manifold is discrete if and only if it is stabily completed. In Sections \[brr\] and \[ccamm\] we study completed normal coverings having locally finite multiplicity and prove Theorems \[kuuskytkolme\], \[kk\] and \[hopo\]. [**Acknowledgements.**]{} I want to thank my PhD advisor Pekka Pankka for our numerous discussions on the topic and for reading the manuscript and giving excellent suggestions. Preliminaries ============= Normal coverings {#nc} ---------------- In this section we recall some facts on normal coverings and fix some notation related to coverings; we refer to Hatcher [@H Ch.1] for a detailed discussion. Let $Y$ be a topological space. A *path* is a continuous map ${\alpha}\colon [0,1] {\rightarrow}Y.$ The *inverse path* ${\alpha}^{-1} \colon [0,1] {\rightarrow}Y$ of a path ${\alpha}\colon [0,1] {\rightarrow}Y$ is the path $t \mapsto {\alpha}(1-t).$ The *path composition* ${\alpha}{\beta}\colon [0,1]{\rightarrow}Y$ of paths ${\alpha}\colon [0,1] {\rightarrow}Y$ and ${\beta}\colon [0,1] {\rightarrow}Y$ satisfying ${\alpha}(1)={\beta}(0)$ is the path $t \mapsto {\alpha}(2t)$ for $t \in [0,1/2]$ and $t \mapsto {\beta}(2(t-1/2))$ for $t \in [1/2,1].$ For $y_0, y_1 \in Y$ a path ${\alpha}\colon [0,1] {\rightarrow}Y$ satisfying ${\alpha}(0)=y_0$ and ${\alpha}(1)=y_1$ is denoted ${\alpha}\colon y_0 {\curvearrowright}y_1.$ Given $y_0 \in Y,$ we denote by $\pi(Y,y_0)$ the fundamental group of $Y$ (at $y_0$). A map $p \colon X {\rightarrow}Y$ between path-connected spaces $X$ and $Y$ is a covering map if for every $y \in Y$ there exists an open neighbourhood $U$ of $y$ such that the pre-image $p^{-1}(U) \subset X$ is a union of pairwise disjoint open sets homeomorphic to $U$. A space $X$ is called *a cover* of $Y$ if there is a covering map from $X$ to $Y.$ Let $p \colon X {\rightarrow}Y$ be a covering map, $y_0 \in Y$ and $x_0 \in p^{-1}\{y_0\}.$ A path ${\widetilde}{{\alpha}} \colon [0,1] {\rightarrow}X$ is called a *lift* of ${\alpha}\colon [0,1] {\rightarrow}Y$ in $p$ if ${\alpha}= p \circ {\widetilde}{{\alpha}}.$ For every path ${\alpha}\colon [0,1] {\rightarrow}Y$ satisfying ${\alpha}(0)={\alpha}(1)=y_0$ there exists a unique lift ${\widetilde}{{\alpha}}$ satisfying ${\widetilde}{{\alpha}}(0)=x_0.$ We denote this lift by ${\widetilde}{{\alpha}}_{x_0}.$ Given a loop ${\alpha}\colon (S^1,e_0) {\rightarrow}(Y,y_0)$ we denote by ${\widetilde}{{\alpha}}$ the lift of the path $[0,1] {\rightarrow}Y, t \mapsto {\alpha}(\text{cos}(2 \pi t),\text{sin}(2 \pi t)).$ The *deck-transformation group* ${\mathcal{T}}(p)$ of a map $p \colon X {\rightarrow}Y$ is the group of all homeomorphisms $\tau \colon X {\rightarrow}X$ satisfying $p(\tau(x))=p(x)$ for all $x \in X.$ Let $\tau_1 \colon X {\rightarrow}X$ and $\tau_2 \colon X {\rightarrow}X$ be deck-transformations of a covering $p \colon X {\rightarrow}Y.$ Then $\tau_1=\tau_2$ if and only if there exists $x \in X$ so that $\tau_1(x)=\tau_2(x).$ A covering $p \colon X {\rightarrow}Y$ is *normal* if the subgroup $p_*(\pi(X,x_0)) {\subset}\pi(Y,y_0)$ is normal for $y_0 \in Y$ and $x_0 \in p^{-1}\{y_0\}.$ The following statements are equivalent: - $p$ is a normal covering, - for every $y \in Y$ and pair of points $x_1, x_2 \in p^{-1}\{y\}$ there exists a unique deck transformation $\tau \colon X {\rightarrow}X$ satisfying $\tau(x_1)=x_2,$ - $X/{\mathcal{T}}(p) \approx Y,$ and - $p_*(\pi(X,x_0))=p_*(\pi(X,x_1))$ for all $x_1 \in p^{-1}\{p(x_0)\}.$ Moreover, ${\mathcal{T}}(p) \cong \pi(Y,y_0)/p_*(\pi(X,x_0))$ if and only if $p$ is a normal covering. For every manifold $Y,$ base point $y_0 \in Y$ and normal subgroup $N \subset \pi(Y,y_0),$ there exists a normal covering $p \colon X {\rightarrow}Y$ so that $p_*(\pi(X,x_0))=N$ for every $x_0 \in p^{-1}\{y_0\}.$ Moreover, the cover $X$ is a manifold. Thus ${\mathcal{T}}(p)$ is countable, since $p^{-1}\{y_0\}$ is countable. We also recall the following properties of covering maps. Let $p \colon (X,x_0) {\rightarrow}(Y,y_0)$ be a normal covering and ${\widetilde}{{\alpha}}_{x_0} \colon [0,1] {\rightarrow}X$ a lift of a loop ${\alpha}\colon (S^1,e_0) {\rightarrow}(Y,y_0).$ Then the point ${\widetilde}{{\alpha}}_{x_0}(1) \in p^{-1}\{y_0\}$ depends only on the class of ${\alpha}$ in $\pi(Y,y_0)/p_*(\pi(X,x_0)).$ For any pair of points $y_1 \in Y$ and $x_1 \in p^{-1}\{y_1\}$ and a path ${\beta}\colon y_1 {\curvearrowright}y_0,$ the lift ${\widetilde}{({\beta}{\alpha}{\beta}^{{\leftarrow}})}_{x_1}$ is a loop if and only if the lift ${\widetilde}{{\alpha}}_{{\widetilde}{{\beta}}_{x_1}(1)}$ is a loop. Further, ${\widetilde}{({\beta}{\alpha}{\beta}^{{\leftarrow}})}_{x_1}(1)={\widetilde}{({\beta}\gamma{\beta}^{{\leftarrow}})}_{x_1}(1)$ for a loop $\gamma \colon (S^1,e_0) {\rightarrow}(Y,y_0)$ if and only if ${\widetilde}{{\alpha}}_{{\widetilde}{{\beta}}_{x_1}}(1)={\widetilde}{\gamma}_{{\widetilde}{{\beta}}_{x_1}}(1).$ Monodromy {#mc} --------- In this section we recall some facts on the monodromy of maps. Let $f \colon X {\rightarrow}Y$ be a covering, $y_0 \in Y$ and $K: = f^{-1}\{y_0\}.$ For every $[{\alpha}] \in \pi(Y,y_0)$ we define a map $m_{[{\alpha}]} \colon K {\rightarrow}K$ by setting $m_{[{\alpha}]}(k)={\widetilde}{{\alpha}}_k(1)$ for every $k \in K.$ For every $[{\alpha}] \in \pi(Y,y_0),$ the map $m_{[{\alpha}]}$ is a bijection, since $m_{[{\alpha}]} \circ m_{[{\alpha}]^{-1}}=m_{[{\alpha}]^{-1}} \circ m_{[{\alpha}]}=\mathrm{id}.$ The *monodromy* of $f$ is the homomorphism $$\sigma_f \colon \pi(Y,y_0) {\rightarrow}\text{Sym}(K),\, [{\alpha}] \to m_{[{\alpha}]}.$$ We call the quotient $\pi(Y,y_0)/\mathrm{Ker}(\sigma_f)$ the monodromy group of $f.$ Let $p \colon (Y,y_0) {\rightarrow}(X,x_0)$ be a normal covering, $f \colon (X,x_0) {\rightarrow}(Z,z_0)$ a covering and $q:=f \circ p \colon (Y,y_0) {\rightarrow}(Z,z_0).$ Then $q$ is a covering and $q_{*}(\pi(Y,y_0))$ is a normal subgroup of $f_{*}(\pi(Z,z_0)).$ We note that, $q$ is a normal covering if and only if $q_{*}(\pi(Y,y_0))$ is a normal subgroup of $\pi(Z,z_0).$ The question weather the covering $q$ is a normal is related to the monodromy of the covering $f$ in the following way. \[spread\_3\] Let $X$ be a manifold, $f \colon (X,x_0) {\rightarrow}(Z,z_0)$ a covering and $N \subset \pi(Z,z_0)$ a normal subgroup. Then there exists a normal covering $p \colon (Y,y_0) {\rightarrow}(X,x_0)$ satisfying $(f \circ p)_{*}(\pi(Y,y_0))=N$ if and only if $N \subset \mathrm{Ker}(\sigma_f).$ In order to prove Proposition \[spread\_3\], we prove first Lemma \[nalkat\] and Lemma \[nalka\]. \[nalkat\]Let $f \colon (X,x_0) {\rightarrow}(Y,y_0)$ be a covering map. Then $\mathrm{Ker} (\sigma_f)$ is a normal subgroup of $\pi(Y,y_0)$ satisfying $\mathrm{Ker} (\sigma_f) \subset f_*(\pi(X,x_0)).$ Clearly, $\mathrm{Ker} (\sigma_f) \subset \pi(Y,y_0)$ is a normal subgroup. Let ${\alpha}\colon (S^1,e_0) {\rightarrow}(Y,y_0)$ be a loop such that $[{\alpha}] \in \mathrm{Ker} (\sigma_f)$ and let ${\widetilde}{{\alpha}}_{x_0} \colon [0,1] {\rightarrow}X$ be the lift of ${\alpha}$ in $f$ starting from $x_0.$ Since $${\widetilde}{{\alpha}}_{x_0}(1)=m_{[{\alpha}]}(x_0)=\text{id}(x_0)=x_0={\widetilde}{{\alpha}}_{x_0}(0),$$ ${\widetilde}{{\alpha}}_{x_0}$ is a loop. Since ${\alpha}=f \circ {\widetilde}{{\alpha}}_{x_0},$ we have $[{\alpha}] \in f_*(\pi(X,x_0)).$ Hence $\mathrm{Ker} (\sigma_f) \subset f_*(\pi(X,x_0)).$ \[nalka\]Let $f \colon (X,x_0) {\rightarrow}(Y,y_0)$ be a covering and $N \subset f_*(\pi(X,x_0))$ a normal subgroup of $\pi(Y,y_0).$ Then $N \subset \mathrm{Ker} (\sigma_f).$ We need to show that for every loop ${\alpha}\colon (S^1,e_0) {\rightarrow}(Y,y_0)$ for which $[{\alpha}] \in N$ the lift ${\widetilde}{{\alpha}}_x$ in $f$ is a loop for every $x \in f^{-1}\{y_0\}$. Let ${\beta}\colon x_0 {\curvearrowright}x$ be a path. Since $N$ is a normal subgroup of $\pi(Y,y_0),$ we have $$[\gamma]:=[f \circ {\beta}][{\alpha}][f \circ {\beta}]^{-1}=[(f \circ {\beta}){\alpha}(f \circ {\beta})^{{\leftarrow}}] \in N\subset f_*(\pi(X,x_0)).$$ Thus the lift ${\widetilde}{\gamma}_{x_0}$ is a loop. By the uniqueness of lifts, ${\widetilde}{{\alpha}}_x$ is a loop. This shows that $N \subset \mathrm{Ker}(\sigma_f).$ Let $p \colon (Y,y_0) {\rightarrow}(Z,z_0)$ be a normal covering. Suppose first that $(f \circ p)_{*}(\pi(Y,y_0))=N.$ Then $N \subset \mathrm{Ker}(\sigma_f)$ by Lemma \[nalka\]. Suppose now that $N \subset \mathrm{Ker}(\sigma_f).$ Then, by Lemma \[nalkat\], $N \subset f_*(\pi(X,x_0)).$ Since $f_*$ is a monomorphism, the pre-image $\tilde{N}$ of $N$ under $f_*$ is a normal subgroup of $\pi(X,x_0)$ and isomorphic to $N.$ Hence there exists a normal covering $p \colon (Y,y_0) {\rightarrow}(X,x_0)$ satisfying $p_{*}(\pi(Y,y_0))={\widetilde}{N} \cong N.$ In particular, $(f \circ p)_{*}(\pi(Y,y_0))=f_{*}(\tilde{N})=N.$ The following observation is due to Berstein-Edmonds [@BE]. \[he\] Let $$\xymatrix{&Y \ar[dl]_{p} \ar[dr]^{q:=f \circ p} \\X \ar[rr]_{f} && Z}$$ be a diagram of covering maps so that $q_{*}(\pi(Y,y_0))=\mathrm{Ker}(\sigma_f)$ for $z_0 \in Z$ and $y_0 \in q^{-1}\{z_0\}.$ Then the deck-transformation group ${\mathcal{T}}(q)$ of the covering $q \colon Y {\rightarrow}Z$ is finite if and only if the set $f^{-1}\{z_0\} \subset X$ is finite. Suppose that the group ${\mathcal{T}}(q)$ is finite and $\#{\mathcal{T}}(q) = m$. Since $q$ is a normal covering, $\#(q^{-1}\{z_0\})=m$. Hence $\#(f^{-1}\{z_0\})\leq m,$ since $q=f \circ p.$ Suppose that $f^{-1}\{z_0\} \subset X$ is finite. Then the finiteness of the symmetric group $\text{Sym}(f^{-1}\{z_0\})$ implies that $${\mathcal{T}}(q) \cong \pi(Z,z_0)/\mathrm{Ker}(\sigma_f) \cong \mathrm{Im}(\sigma_f) \subset \text{Sym}(f^{-1}\{y_0\})$$ is a finite group. Next we define monodromy for a broader class of maps including completed coverings defined in the following section. We say, that an open dense subset $Z' \subset Z$ is *large* if $Z \setminus Z'$ does not locally separate $Z.$ \[something\] Let $f \colon X {\rightarrow}Z$ be a continuous and open map. Suppose that for $i \in \{1,2\}$ there exist large subsets $X_i \subset X$ and $Z_i \subset Z$ so that $g_i:=f {|}X_i \colon X_i {\rightarrow}Z_i$ is a covering. Then $\mathrm{Im}(\sigma_{g_1})$ is isomorphic to $\mathrm{Im}(\sigma_{g_2}).$ Let $z_0 \in Z_1 \cap Z_2$ and $g:= f {|}X_1 \cap X_2 \colon X_1 \cap X_2 {\rightarrow}Z_1 \cap Z_2.$ Then $g$ is a covering. It is sufficient to show that $\mathrm{Im}(\sigma_{g_1})=\mathrm{Im}(\sigma_{g}).$ Let $\iota_* \colon \pi(Z_1 \cap Z_2,z_0) {\rightarrow}\pi(Z_1,z_0)$ be the homomorphism induced by the inclusion $\iota \colon Z_1 \cap Z_2 {\hookrightarrow}Z_1.$ Then $\iota_*$ is surjective, since $Z_1 \cap Z_2 \subset Z_1$ is dense and $Z_1 \setminus(Z_1 \cap Z_2)$ does not locally separate $Z_1.$ Thus $\mathrm{Im}(\sigma_{g_1})\cong\mathrm{Im}(\sigma_{g}).$ Let $f \colon X {\rightarrow}Z$ be an open and continuous map so that $g:=f {|}X' \colon X' {\rightarrow}Z'$ is a covering for large subsets $X' \subset X$ and $Z' \subset Z.$ Then we say that the monodromy $\sigma_g \colon \pi(Z',z_0) {\rightarrow}\mathrm{Sym}(f^{-1}\{z_0\})$ is a *monodromy* $\sigma_f$ of $f.$ We call the monodromy group $\pi(Z',z_0)/\mathrm{Ker}(\sigma_g)$ of $g$ the *monodromy group of* $f.$ We note that the isomorphism class of the monodromy group of $f$ is independent of the choice of $z_0 \in Z',$ and thus by Lemma \[something\], it is independent of the choice of monodromy. Spreads and completed coverings =============================== Spreads {#sp} ------- In this section we discuss the theory of spreads; see Fox [@F] and Montesinos [@MA] for a more detailed discussion. A mapping $g \colon Y {\rightarrow}Z$ between locally connected Hausdorff spaces is called a *spread* if the components of the inverse images of the open sets of $Z$ form a basis of $Y.$ Clearly spreads are continuous and a composition of spreads is a spread. A map $g \colon Y {\rightarrow}Z$ between topological manifolds is a spread if and only if $g$ is light; for every $z \in Z$ the set $g^{-1}\{z\}$ is totally disconnected, see [@MA Cor.4.7]. In particular, if $Z$ is a manifold and $g \colon Y {\rightarrow}Z$ is an open mapping which is a covering to its image, then $g$ is a spread. The definition of a complete spread is formulated by Fox [@F]: a spread $g \colon Y {\rightarrow}Z$ is *complete* if for every point $z$ of $Z$ and every open neighbourhood $W$ of $z$ there is selected a component $V_W$ of $g^{-1}(W)$ in such a way that $V_{W_1} \subset V_{W_2}$ whenever $W_1 \subset W_2,$ then $\cap_{W}V_{W}\neq {\varnothing}.$ This definition can equivalently be written as follows. For every $z \in Z,$ we denote by $\mathcal{N}_Z(z)$ the set of open connected neighbourhoods of $z.$ A function $\Theta \colon \mathcal{N}_Z(z) {\rightarrow}\text{Top}(Y)$ is a *selection function* of a spread $g\colon Y {\rightarrow}Z$ if $\Theta(W)$ is a component of $g^{-1}(W)$ for every $W \in \mathcal{N}_Z(z)$ and $\Theta(W_1)\subset \Theta(W_2)$ whenever $W_1 \subset W_2.$ A spread $g \colon Y {\rightarrow}Z$ is *complete* if for all $z \in Z$ every selection function $\Theta \colon \mathcal{N}_Z(z) {\rightarrow}\mathrm{Top}(Y)$ of $g$ satisfies $$\bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W)\neq {\varnothing}.$$ Let $g \colon Y {\rightarrow}Z$ be a complete spread. Since $Y$ is Hausdorff, the set $\bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W)\subset Y$ is a point for every $z \in Z$ and every selection function $\Theta \colon \mathcal{N}_Z(z) {\rightarrow}\text{Top}(Y).$ Moreover, for every $z \in Z$ and $y \in g^{-1}\{z\}$ there exists a selection function $\Theta \colon \mathcal{N}_Z(z) {\rightarrow}\text{Top}(Y)$ of $g$ for which $\bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W)=\{y\}.$ \[aurinko\]Let $g \colon Y {\rightarrow}Z$ be a spread. Suppose that for every $z \in Z$ the following conditions are satisfied: - $g(V)=W$ for every $W \in \mathcal{N}_Z(z)$ and every component $V$ of $g^{-1}(W),$ and - there exists $W_0 \in \mathcal{N}_Z(z)$ so that for every component $V$ of $g^{-1}(W_0)$ the set $V \cap g^{-1}\{y\}$ is finite. Then $g$ is a complete spread. Let $z \in Z$ and let $\Theta \colon \mathcal{N}_Z(z) {\rightarrow}\text{Top}(Y)$ be a selection function for $g$. We need to show that $\bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W) \neq {\varnothing}.$ Suppose $\bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W)={\varnothing}.$ By condition (2), the set $g^{-1}\{z\} \cap \Theta(W_0)$ is finite for some $W_0 \in \mathcal{N}_Z(z).$ Let $\{y_1, \ldots, y_k\}=g^{-1}\{z\} \cap \Theta(W_0).$ Since $\bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W)={\varnothing},$ we find for every $j=0, \ldots, k$ a set $W_j \in \mathcal{N}_Z(z)$ for which $y_j \notin \Theta(W_j).$ Let $W'=W_0 \cap \cdots \cap W_k.$ Then $\Theta(W') \cap g^{-1}\{z\}={\varnothing}.$ Thus $z \notin \bigcap_{W \in \mathcal{N}_Z(z)}g(\Theta(W))\subset g(\Theta(W')).$ On the other hand, $z \in \bigcap_{W \in \mathcal{N}_Z(z)}g(\Theta(W))$ as a consequense of condition (1). This is a contradiction and we conclude that $\bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W)\neq {\varnothing}.$ For completion of presentation we include the following proposition well known for the experts. Let $f \colon X {\rightarrow}Z$ be a discrete and open map between compact manifolds. Then $f$ is a complete spread. The map $f$ is a spread by [@MA Cor.4.7]; see also proof of Theorem \[pakkanen\]. We prove that $f$ is complete by showing that conditions (1) and (2) in Lemma \[aurinko\] hold. Since $f$ is an open and discrete map from a compact manifold $X,$ the map $f$ has finite multiplicity. Thus condition (2) holds since $f$ is a spread. We then suppose that condition (1) does not hold. Then there exists $x \in X$ and $V \in \mathcal{N}_Z(f(x))$ so that the $y$-component $U$ of $f^{-1}(V)$ satisfies $f(U)\subsetneq V.$ Since $V$ is connected there is a point $z \in \partial f(U) \cap V.$ Since $X$ is compact $\overline{U} \subset X$ is compact. Thus $z \in \overline{f(U)} \subset f(\overline{U})$ and there exists a point $x' \in \overline{U} \cap p^{-1}\{z\}.$ This is a contradiction since $x' \in (\overline{U} \setminus U) \cap p^{-1}(V)$ and $U$ is a component of $p^{-1}(V).$ We conclude that (1) holds and that $f$ is a complete spread. We recall that an open dense subset $Z' \subset Z$ is large if $Z \setminus Z'$ does not locally separate $Z.$ \[completion\]Let $X,$ $Y$ and $Z$ be Hausdorff-spaces. A complete spread $g \colon Y {\rightarrow}Z$ is a *completion* of a spread $f \colon X {\rightarrow}Z$ if there is an embedding $i \colon X {\rightarrow}Y$ such that $i(X)\subset Y$ is large and $$\xymatrix{X \ar[rr]^{i} \ar[dr]_{f} && Y \ar[dl]^g \\&Z.}$$ Two completions $g_1 \colon Y_1 {\rightarrow}Z$ and $g_2 \colon Y_2 {\rightarrow}Z$ of a spread $f \colon X {\rightarrow}Z$ are *equivalent* if there exists a homeomorphism $j \colon Y_1 {\rightarrow}Y_2$ such that $g_2 \circ j=g_1$ and $j(x)=x$ for all $x \in X.$ By the fundamental result of Fox, every spread has a completion [@F Sec.2] and it is unique up to equivalence [@F Sec.3] ; see also [@MA Cor. 2.8 and Cor. 7.4]. Completed coverings {#ccc} ------------------- Let $Y$ be a Hausdorff space and $Z$ a manifold. Suppose $f \colon Y {\rightarrow}Z$ is a complete spread so that there are large subsets $Y' \subset Y$ and $Z' \subset Z$ for which $f':=f {|}Y' \colon Y' {\rightarrow}Z'$ is a covering. Then $f$ is the completion of $f' \colon Y' {\rightarrow}Z,$ and since $Z' \subset Z$ is large, $f$ is an open surjection, see [@C Teo. 2.1]. \[cc\] Let $Y$ be a Hausdorff space and $Z$ a manifold. A map $f \colon Y {\rightarrow}Z$ is a *completed covering* if there are large subsets $Y' \subset Y$ and $Z' \subset Z$ so that $f{|}Y' \colon Y' {\rightarrow}Z'$ is a covering and $f$ is the completion of $f'{|}Y' \colon Y' {\rightarrow}Z.$ The results in [@I] cover the following example. Let $f \colon X {\rightarrow}Z$ be an open and discrete simplicial map from a PL $n$-manifold $X$ with triangulation $K$ onto a PL $n$-manifold $Z$ with triangulation $L.$ Let $K^{n-2}$ be the $(n-2)$-skeleton of $K,$ $L^{n-2}$ the $(n-2)$-skeleton of $L,$ $X'=|K| \setminus |K^{n-2}|$ and $Z'=|L| \setminus |L^{n-2}|.$ Then $f {|}X' \colon X' {\rightarrow}Z'$ is a covering and $f$ is the completion of $f {|}X' \colon X' {\rightarrow}Z.$ Next we prove some basic properties of completed coverings. \[biotin\]Let $g \colon Y' \to Z'$ be a normal covering onto a large subset $Z'\subset Z$ of a manifold $Z$ and $f \colon Y {\rightarrow}Z$ the completion of $g \colon Y' {\rightarrow}Z.$ Then $f^{-1}(Z')=Y'.$ Suppose there exists a point $y \in Y \setminus Y'$ so that $z:=f(y)\in Z'.$ Let $\Theta \colon \mathcal{N}_Z(z) {\rightarrow}\text{Top}(Y)$ be a selection function of $f$ satisfying $$\bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W)=\{y\}.$$ Since $Z' \subset Z$ is open, we have $\mathcal{N}_{Z'}(z) \subset \mathcal{N}_Z(z).$ Moreover, since $Y' \subset Y$ is large, the set $\Theta(W) \cap Y'$ is a component of $g^{-1}(W)$ for every $W \in \mathcal{N}_{Z'}(z).$ Define $\Theta' \colon \mathcal{N}_{Z'}(z) {\rightarrow}\text{Top}(Y')$ by $\Theta'(W)= \Theta(W) \cap Y'.$ Then $\Theta'$ is a selection function for $g \colon Y' {\rightarrow}Z'.$ Since $g$ is a covering, it is also a complete spread. Thus there exists $y' \in Y'$ so that $$\bigcap_{W \in \mathcal{N}_{Z'}(z)} \Theta'(W)=\{y'\}.$$ This is a contradiction, since $$y' \in \bigcap_{W \in \mathcal{N}_Z(z)} \Theta(W)=\{y\} \subset Y \setminus Y'.$$ \[brunate\] Let $g \colon Y' \to Z'$ be a covering onto a large subset $Z'\subset Z$ of a manifold $Z$ and $f \colon Y {\rightarrow}Z$ the completion of $g \colon Y' {\rightarrow}Z.$ Let $U \subset Z$ be a connected and open set, and let $V$ be a component of $f^{-1}(U).$ Then $f(V)=U.$ Since $Z'\subset Z$ is large, the set $U \cap Z'$ is path-connected. Furthermore, since $Y' \subset Y$ is dense, we may fix a component $W$ of $g^{-1}(U \cap Z')$ contained in $V \cap Y'.$ Every path into $U \cap Z'$ has a total lift into $W,$ since $g$ is a covering map. Thus $U \cap Z'= \mathrm{Im}(g {|}W).$ In particular, $g {|}W \colon W {\rightarrow}Z' \cap U$ is a covering onto a large subset of $U$. Since $W$ is a connected and open set and $f$ is a complete spread, there exists such a connected subset $V' \subset Y$ that $f {|}V'$ is the completion of the spread $g {|}W \colon W {\rightarrow}U.$ Since $U \cap Z'\subset U$ is large, the map $f {|}V' \colon V' {\rightarrow}U$ is a surjective spread by [@C Thm. 2.1]. Since $V'$ is connected and $W \subset V,$ we have $V' \subset V.$ Thus $U \subset f(V') \subset f(V)\subset U.$ \[elama\] Let $f_1 \colon X {\rightarrow}Y$ and $f_2 \colon Y {\rightarrow}Z$ be completed coverings. Then $f_2 \circ f_1 \colon X {\rightarrow}Z$ is a completed covering. Let $X'\subset X,$ $Y' \subset Y$ and $Z' \subset Z$ be large subsets so that $g_1:= f_1 {|}X' \colon X' {\rightarrow}Y'$ and $g_2:= f_1 {|}Y' \colon Y' {\rightarrow}Z'$ are coverings. Then $f_1$ is the completion of $g_1 \colon X' {\rightarrow}Y$ and $f_2$ is the completion of $g_2 \colon Y' {\rightarrow}Z$ since $f_1$ and $f_2$ are complete spreads. Since $Z'\subset Z$ is large $f_2\circ f_1 \colon X' {\rightarrow}Z$ is a spread. Moreover, since $X' \subset X$ is large, it is sufficient to show that $f_2 \circ f_1 \colon X {\rightarrow}Z$ is complete. If there are no selection functions for $f_2 \circ f_1,$ then $f_2 \circ f_1$ is trivially a complete spread. Then suppose $z \in Z$ and $\Theta \colon \mathcal{N}_{Z}(z) {\rightarrow}\text{Top}(X)$ is a selection function for $f_2 \circ f_1.$ We need to show that ${\bigcap}_{W \in \mathcal{N}_Z(z)} \Theta(W) \neq {\varnothing}.$ Since $\Theta$ is a selection function, there are for every $W \in \mathcal{N}_Z(z)$ components $U_{W}$ of $f_2^{-1}(W)$ and $V_{U_{W}}$ of $f_2^{-1}(U_{W})$ so that $\Theta(W)=V_{U_{W}}$ and $U_{W_1} \subset U_{W_2}$ whenever $W_1 \subset W_2.$ Define $\Theta_2 \colon \mathcal{N}_{Z}(z) {\rightarrow}\text{Top}(Y)$ by $W \mapsto U_{W}.$ Then $\Theta_2$ is a selection function of $f_2.$ Since $f_2$ is a complete spread, there exists a point $y \in f^{-1}_2\{z\}$ satisfying ${\bigcap}_{W \in \mathcal{N}_Z(z)} \Theta_2(W)=\{y\}.$ Thus there exists a selection function $\Theta_1 \colon \mathcal{N}_{Y}(y) {\rightarrow}\text{Top}(X)$ satisfying $\Theta_1(U_W)=V_{U_{W}}$ for every $W \in \mathcal{N}_Z(z).$ Since $f_1$ is complete, there exists a point $x \in f_1^{-1}\{y\}$ satisfying ${\bigcap}_{U \in \mathcal{N}_Y(y)}\Theta_1(U)=\{x\}.$ Now $$\{x\}={\bigcap}_{U \in \mathcal{N}_Y(y)} \Theta_1(U) \subseteq {\bigcap}_{W \in \mathcal{N}_Z(z)} V_{U_W}= {\bigcap}_{W \in \mathcal{N}_Z(z)} \Theta(W).$$ Thus $ {\bigcap}_{W \in \mathcal{N}_Z(z)} \Theta(W) \neq {\varnothing},$ and we conclude that $f_2\circ f_1$ is a complete spread. \[sina\] Let $f_1 \colon X {\rightarrow}Y$ be a completed covering and $f_2 \colon Y {\rightarrow}Z$ a spread between manifolds and suppose $Y'' \subset Y$ and $Z'' \subset Z$ are large subsets so that the map $f_2 {|}Y'' \colon Y'' {\rightarrow}Z''$ is a covering. Then $f_2$ is a completed covering if $f:=f_2 \circ f_1 \colon X {\rightarrow}Z$ is a completed covering. Let $X'\subset X,$ $Y' \subset Y$ and $Z' \subset Z$ be large subsets so that the maps $g_1:= f_1 {|}X' \colon X' {\rightarrow}Y'$ and $f_2 {|}Y' \colon Y' {\rightarrow}Z'$ are coverings. Then $f_1$ is the completion of $g_1 \colon X' {\rightarrow}Y$, $g:= f {|}X' \colon X' {\rightarrow}Z'$ is a covering and $f$ is the completion of $g \colon X' {\rightarrow}Z$. Since $f_2$ is a spread and $Y' \subset Y$ is large, it is sufficient to show that $f_2 \colon Y {\rightarrow}Z$ is complete. Suppose there are no selection functions for $f_2.$ Then $f_2$ is trivially complete. Suppose then that $z \in Z$ and $\Theta \colon \mathcal{N}_{Z}(z) {\rightarrow}\text{Top}(Y)$ is a selection function for $f_2.$ Since $f_1$ is a completed covering, there exists a selection function $\Theta' \colon \mathcal{N}_{Z}(z) {\rightarrow}\text{Top}(X)$ so that $\Theta'(W)$ is a component of $f_1^{-1}(\Theta(W))$ for every $W \in \mathcal{N}_Z(z)$ as a consequence of Lemma \[brunate\]. Since $f$ is complete, there exists a point $x \in {\bigcap}_{W \in \mathcal{N}_Z(z)} \Theta'(W).$ Now $$f_1(x) \in f\left({\bigcap}_{W \in \mathcal{N}_Z(z)} \Theta'(W)\right) \subseteq{\bigcap}_{W \in \mathcal{N}_Z(z)} f(\Theta'(W))\subseteq{\bigcap}_{W \in \mathcal{N}_Z(z)} \Theta(W).$$ Thus $f_2$ is complete and we conclude that $f_2$ is a completed covering. Orbit maps {#os} ---------- In this section we prove Theorem \[sade\] in the introduction. An open continuous map $p \colon Y {\rightarrow}Z$ is an *orbit map* if there exists a subgroup $G \subset \text{Homeo}(Y)$ and a homeomorphism $\phi \colon Y/G {\rightarrow}Z$ so that the diagram $$\xymatrix{Y \ar[rr]^{p} \ar[dr]_{\pi} && Z \\& Y/G \ar[ru]_{\phi}},$$ commutes for the canonical map $\pi \colon Y {\rightarrow}Y/G$ \[vindpark\]Let $f \colon Y {\rightarrow}Z$ be an orbit map. Let $U \subset Z$ be a connected and open set and let $V_1$ and $V_2$ be components of $f^{-1}(U)$ satisfying $f(V_1)=f(V_2)=U.$ Then there exists a deck-transformation $\tau \colon Y {\rightarrow}Y$ satisfying $\tau(V_1)=V_2.$ Let $z \in U$ and fix points $y_1 \in V_1 \cap f^{-1}\{z\}$ and $y_2 \in V_2 \cap f^{-1}\{z\}.$ Since $f$ is an orbit map there is a deck-transformation $\tau \colon Y {\rightarrow}Y$ for which $\tau(y_1)=y_2.$ Since $\tau(f^{-1}(U))=f^{-1}(U)$ and $V_1$ and $V_2$ are components of $f^{-1}(U)$, we have $\tau(V_1)\subset V_2$ and $\tau^{-1}(V_2)\subset V_1.$ Thus $\tau(V_1)=V_2.$ \[pakkanen\] Let $Y$ be a locally compact and locally connected Hausdorff space, $Z$ a manifold and $p \colon Y {\rightarrow}Z$ a discrete, continuous and open orbit map. Suppose there are large subsets $Y' \subset Y$ and $Z' \subset Z$ so that $g \colon Y' {\rightarrow}Z'$ is a covering. Then $p$ is a completed covering. Suppose $U\subset Y$ is open. To show that $p$ is a spread, it is sufficient to show that for every $y \in U$ there exists a neighbourhood $V \subset p(U)$ of $p(y)$ for which the $y$-component $D$ of $p^{-1}(V)$ is contained in $U.$ Fix $y \in U.$ Since $Y$ is a locally compact and $p$ discrete there exists a neighbourhood $W\subset U$ of $y,$ so that $\overline{W}$ is compact and $\partial W \cap p^{-1}\{p(y)\}={\varnothing}.$ Since $p$ is open and continuous the set $p(W) \setminus p(\partial W)\subset Z$ is open. Since $Z$ is locally connected the $p(y)$-component $V$ of $p(W) \setminus p(\partial W)$ is open and since $Y$ is locally connected the $y$-component $D$ of $f^{-1}(V)$ is open. Since $D \cap \partial W={\varnothing},$ we get $D \subset W \subset U.$ We conclude that $p$ is a spread. Since $Y' \subset Y$ and $Z' \subset Z$ are large and $p {|}Y' \colon Y' {\rightarrow}Z'$ is a covering, to show that $p$ is a completed covering it is sufficient to show that the spread $p$ is complete. For this let $z \in Z.$ By Lemma \[aurinko\] it is suffient to show that $p(U)=W$ for every $W \in \mathcal{N}_Z(z)$ and $U$ a component of $p^{-1}(W),$ and that there is a connected neighbourhood $W_0$ of $z$ so that for every component $V_0$ of $p^{-1}(W_0)$ the set $V_0 \cap p^{-1}\{z\}$ is finite. Towards this let $W \in \mathcal{N}_Z(z).$ Since $p$ is an orbit map $p$ is necessarily onto. Hence there is a component $U_1$ of $p^{-1}(W)$ for which $z \in p(U_1).$ Let $U_2$ be a component of $p^{-1}(W).$ Since $p$ is an orbit map the deck-transformation group ${\mathcal{T}}(p)$ acts transitively on fibers $p^{-1}\{z'\}$ for every $z' \in Z$. Thus Lemma \[vindpark\] implies that $p(U_1)=p(U_2)$ or $p(U_1) \cap p(U_2)={\varnothing}.$ This implies that both $p(U_1)$ and $W \setminus p(U_1)$ are open sets. Since $W$ is connected, this is only possible if $W \setminus p(U_1)={\varnothing}.$ Thus $p(U)=W$ for every component $U$ of $p^{-1}(W).$ Let then $y \in p^{-1}\{z\}$. Since $Y$ is locally compact and locally connected there is a neighbourhood $U_0$ of $y$ so that $p^{-1}\{z\}\cap U_0$ is finite and $U_0$ is the $y$-component of a $p^{-1}(p(U_0)).$ Further by Lemma \[vindpark\], $\#(p^{-1}\{z\}\cap V_0)=\#(p^{-1}\{z\}\cap U_0)$ for every component $V_0$ of $p^{-1}(p(U_0))$. Thus $p$ is a complete spread by Lemma \[aurinko\]. Let $f \colon X {\rightarrow}Z$ be a mapping and $X' \subset X$ and $Z' \subset Z$ large subsets as in the statement. Similarly, let $Y$ be a space and $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ be orbit maps as in the statement. By assumption, there exists a large subset $Y' \subset Y$ for which the maps $p {|}Y' \colon Y' {\rightarrow}X'$ and $q {|}Y' \colon Y' {\rightarrow}X'$ are normal coverings. Thus, by Theorem \[pakkanen\], the maps $p$ and $q$ are completed coverings. Since $q=f \circ p$ and $p$ and $q$ are completed coverings, we have by Lemma \[sina\], that the map $f$ is a completed covering. Existence of monodromy representations ====================================== Completed normal coverings and orbit maps {#ooo} ----------------------------------------- In this section we will prove Theorems \[hilu\] and \[ll\] in the introduction. \[vauva\]Let $g \colon Y' \to Z'$ be a normal covering onto a large subset $Z'\subset Z$ of a manifold $Z,$ $f \colon Y {\rightarrow}Z$ the completion of $g \colon Y' {\rightarrow}Z$ and ${\mathcal{T}}(f)$ the deck-transformation group of $f.$ Then for the canonical maps $\pi$ and $\phi$ $$\xymatrix{ Y \ar[dr]_{\pi} \ar[rr]^{f} && Z\\ &Y/{\mathcal{T}}(f) \ar[ur]_{\phi} &}$$ the map $\phi$ is a homeomorphism if and only if $f$ is discrete. In particular, $f$ is an orbit map if and only if $f$ is discrete. \[huu\]Let $g \colon Y' \to Z'$ be a normal covering onto a large subset $Z'\subset Z$ of a manifold $Z$ and $f \colon Y {\rightarrow}Z$ the completion of $g \colon Y' {\rightarrow}Z.$ Then the deck transformations of $g$ are restrictions of deck transformations of $f$ by Lemma \[biotin\], that is, ${\mathcal{T}}(g)=\{\tau {|}Y' \colon Y' {\rightarrow}Y' : \tau \in {\mathcal{T}}(f)\},$ and each $\tau \in {\mathcal{T}}(f)$ is the completion of $\tau {|}Y' \colon Y' {\rightarrow}Y$ since $Y' \subset Y$ is large. In particular, the homomorphism $$\varphi \colon {\mathcal{T}}(f) {\rightarrow}{\mathcal{T}}(g), \tau \mapsto \tau {|}Y',$$ is an isomorphism by the uniqueness of completions, see [@MA Cor. 7.8]. Suppose first that $f$ is discrete. Since $\phi$ is open and continuous, it is sufficient to show that it is bijective; we need to show that for every $z \in Z$ and $y_1, y_2 \in f^{-1}\{z\}$ there exists $\tau \in {\mathcal{T}}(f)$ so that $\tau(y_1)=y_2.$ Since $f$ is discrete, there exists $W \in \mathcal{N}_Z(z)$ so that the $y_i$-components $V_i$ of $f^{-1}(W)$ satisfy $f^{-1}\{z\} \cap V_i=\{y_i\}$ for $i \in \{1,2\}.$ Since $g$ is a normal covering and $Y' \subset Y$ is large, there exists $\tau' \colon Y' {\rightarrow}Y'$ in ${\mathcal{T}}(g)$ for which $\tau'(Y' \cap V_1)=Y' \cap V_2.$ The completion $\tau \in {\mathcal{T}}(f)$ of $\tau' \colon Y' {\rightarrow}Y$ satisfies $\tau(y_1)=y_2.$ Suppose now that $\phi$ is a homeomorphism. Thus ${\mathcal{T}}(f)$ acts transitively on $f^{-1}\{z\}$ for every $z \in Z.$ Arguing towards contradiction assume $f$ is not discrete. Then there is a point $y \in Y$ so that $f^{-1}\{f(y)\} \cap U$ is an infinite set for every neighbourhood $U$ of $y.$ Let $z:=f(y).$ Since $g \colon Y' {\rightarrow}Z'$ is a covering between manifolds, ${\mathcal{T}}(g)$ is countable. Since ${\mathcal{T}}(g)\cong{\mathcal{T}}(f),$ for a contradiction it is sufficient to show that $f^{-1}\{z\}$ is uncountable. Towards a contradiction suppose that the set $f^{-1}\{z\}$ is countable and let $f^{-1}\{z\}=\{y_0,y_1, \ldots \}$ be an enumeration of $f^{-1}\{z\}.$ Then for every $k \in {\mathbb{N}}$ and neighbourhood $U$ of $y_k$ the set $f^{-1}\{z\} \cap U$ is infinite, since $f$ is an orbit map. Fix $W_0 \in \mathcal{N}_Z(z)$ and let $U_0$ be the $y_0$-component of $f^{-1}(W_0).$ Then fix a point $y'_1:=y_k \in \{y_1, y_2 \ldots \} \cap U_0$ for which $y_i \notin U_0, i < k,$ and let $W_1 \in \mathcal{N}_Z(z)$ be such that the $y'_1$-component $U_1$ of $f^{-1}(W_1)$ satisfies $U_0 \supset U_1$ but $y_0 \notin U_1.$ Proceed in this fashion to construct sequences $W_0 \supset W_1 \supset \cdots$ and $U_0 \supset U_1 \supset \cdots $ so that ${\bigcap}_{n=0}^{\infty}W_n=\{z\}.$ Since $f$ is a complete spread, there exists a point $y_{\infty}$ so that ${\bigcap}_{n=0}^{\infty}U_n=\{y_{\infty}\}.$ However, by the construction, ${\bigcap}_{n=0}^{\infty}U_n \cap f^{-1}\{z\}$ is empty. This is a contradiction. Hence $f^{-1}\{z\}$ is uncountable and $\phi$ is not injective. Since $\phi$ is bijective, this is a contradiction and we conclude that $f$ is discrete. Theorems \[hilu\] and \[ll\] in the introduction are now easy corollaries of Theorem \[vauva\]. Let $p \colon Y {\rightarrow}Y$ be a completed normal covering. Then there are large subsets $Y'\subset Y$ and $Z' \subset Z$ so that $p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering and $p$ is the completion of $p {|}Y' \colon Y' {\rightarrow}Z.$ By Theorem \[vauva\] $p$ is an orbit map if and only if $p$ is a discrete map. Let $f \colon X {\rightarrow}Z$ be a completed covering. Suppose $f$ has a monodromy representation $(Y,p,q).$ We need to show that $f,$ $p$ and $q$ are discrete maps. Since $p$ and $q$ are completed coverings and orbit maps, $p$ and $q$ are completed normal coverings. Thus $p$ and $q$ are discrete maps by Theorem \[vauva\]. Since $p$ is an open surjection and $q=p \circ f,$ the discreteness of $f$ follows from the discreteness of $q.$ Image of the branch set of a completed covering {#ando} ----------------------------------------------- In this section we prove Theorem \[yopo\] in the introduction. We recall that the branch set $B_f$ of a completed covering $f \colon X {\rightarrow}Z$ is the set of points where $f$ fails to be a local homeomorphism. We say that a completed covering $f \colon X {\rightarrow}Z$ is *uniformly discrete* if every $z \in Z$ has a neighbourhood $U$ so that $f^{-1}\{z\} \cap D$ is a point for every component $D$ of $f^{-1}(U).$ \[lamppu\] Let $p \colon Y {\rightarrow}Z$ be a discrete completed normal covering. Then $p(B_p) \subset Z$ is closed for the branch set $B_p$ of $p.$ Let $U:=Y \setminus p^{-1}(p(B_p)).$ Since $p$ is an orbit map by Theorem \[vauva\], we have $p^{-1}(p(B_p))=B_p$ and $p(B_p)=Z \setminus p(U).$ Now $U \subset Y$ is open, since $B_p \subset Y$ is closed. Thus $p(B_p)=Z \setminus p(U)$ is closed, since $p$ is an open map. \[gingerale\] Let $f \colon X {\rightarrow}Z$ be a completed covering between manifolds. Suppose there are uniformly discrete completed normal coverings $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ satisfying $q:=f \circ p.$ Then $f(B_f) \subset Z$ is closed. For proving Proposition \[gingerale\] we need the following lemma. \[kylajuhlat\] Let $p \colon Y {\rightarrow}Z$ be a uniformly discrete completed normal covering. Let $z \in Z$ and $U, V \in \mathcal{N}_Z(z)$ be such that $U \subset V$ and $p^{-1}\{z\} \cap D$ is a point for every component $D$ of $p^{-1}(V).$ Let $D$ be a component of $p^{-1}(V)$ and $C$ a component of $p^{-1}(U)$ so that $C \subset D.$ Then the groups $$H_D:=\{ \tau \in {\mathcal{T}}(p) : \tau(D)=D\} \text{ and }H_C:=\{ \tau \in {\mathcal{T}}(p) : \tau(C)=C\}$$ satisfy $H_D=H_C.$ Let $Y' \subset Y$ and $Z' \subset Z$ be large subsets so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering. Then $D \cap Y'$ is a component of $V \cap Z'$ and $C \cap Y'$ is a component of $U \cap Z'.$ By Theorem \[vauva\] $p$ is an orbit map, and in particular, the map ${\mathcal{T}}(p) {\rightarrow}{\mathcal{T}}(g), \tau \mapsto \tau {|}Y'$ is an isomorphism [@MA Cor. 7.8]. Thus it is sufficient to show that $H'_D=H_C'$ for the groups $$H'_D:=\{ \tau \in {\mathcal{T}}(g) : \tau(D \cap Y')=D \cap Y'\}$$ and $$H'_C:=\{ \tau \in {\mathcal{T}}(g) : \tau(C \cap Y')=C \cap Y'\}.$$ Let $z' \in Z' \cap U.$ Since $p^{-1}\{z\} \cap D$ is a point, we have $p^{-1}(U) \cap D=C$ as a consequence of Lemma \[brunate\]. Thus $g^{-1}\{z'\} \cap (D \cap Y')=g^{-1}\{z'\} \cap (C \cap Y').$ Let $y' \in g^{-1}\{z'\} \cap (C \cap Y')$ and let $\tau_{y} \colon Y' {\rightarrow}Y'$ be for every $y \in p^{-1}\{z'\}$ the unique deck-transformation of $g$ that satisfies $\tau_y(y')=y.$ Then $$\begin{aligned} H'_{D}&=\{ \tau_y : y \in g^{-1}\{z'\} \cap (D \cap Y')\} \\&=\{ \tau_y : y \in g^{-1}\{z'\} \cap (C \cap Y')\}=H'_C.\end{aligned}$$ Thus $H_D=H_C.$ Let $z \in Z \setminus f(B_f)$ and $U \in \mathcal{N}_Z(z)$ be a neighbourhood of $z$ so that the set $q^{-1}\{z\} \cap D$ is a point for every component $D$ of $q^{-1}(U)$. We note that, since $p \colon Y {\rightarrow}X$ is onto $X$, $U\subset Z \setminus f(B_f)$ if $f {|}p(D)$ is an injection for every component $D$ of $q^{-1}(U).$ Let $D$ be a component of $q^{-1}(U)$ and $H_D:=\{\tau \in {\mathcal{T}}(q) : \tau(D)=D\}.$ By Theorem \[vauva\], $q$ is an orbit map. Thus $f {|}p(D)$ is an injection, if $H_D \subset {\mathcal{T}}(p).$ Let $y \in q^{-1}\{z\} \cap D$ and $x:=p(y).$ Since $x \in f^{-1}(Z \setminus B_f)$ there exists $U_0 \in \mathcal{N}_Z(z)$ so that $f {|}V$ is an injection for the $x$-component $V$ of $f^{-1}(U_0)$ and so that $U_0 \subset U.$ Let $C$ be the $y$-component of $q^{-1}(U_0).$ Then $f {|}p(C)=f {|}V$ by Lemma \[brunate\]. Thus $H_C:=\{\tau \in {\mathcal{T}}(q) : \tau(C)=C\} \subset {\mathcal{T}}(p),$ since $f {|}p(C)$ is an injection. By Lemma \[kylajuhlat\], $H_D=H_C,$ since $C \subset D.$ Thus $H_D=H_C \subset {\mathcal{T}}(p).$ Thus $f {|}p(D)$ is an injection. Thus $U \subset Z \setminus f(B_f)$ and $f(B_f)\subset Z$ is closed. Towards proving Theorem \[yopo\] we make the following observation. \[spread\_5\]Let $p \colon Y \to Z$ be a completed normal covering. Then $p$ is a discrete map if and only if $p$ is uniformly discrete. Suppose $p$ is discrete. Then, by Theorem \[vauva\], the map $p$ is an orbit map. Let $y \in Y$ and $z:=p(y).$ Then since $p$ is both a discrete map and a spread, there exists a neighbourhood $V$ of $y$ so that $V$ is the $y$-component of $p^{-1}(p(V))$ and $V \cap p^{-1}\{p(y)\}=\{y\}.$ Let $U$ be a component of $p^{-1}(p(V)).$ By Lemma \[vindpark\] there is a deck-transformation $\tau \colon Y {\rightarrow}Y$ of $p$ so that $\tau(V)=U.$ Thus $$p^{-1}\{p(y)\} \cap U=\{\tau(y)\}$$ and $p$ is uniformly discrete. Suppose now that every $z \in Z$ has a neighbourhood $W$ such that $p^{-1}\{z\} \cap V$ is a point for each component $V$ of $p^{-1}(W).$ Then $p$ is discrete, since $p$ is a spread. Let $f \colon X {\rightarrow}Z$ be a completed covering between manifolds and $(Y,p,q)$ a monodromy representation of $f.$ As a consequence of Theorem \[vauva\] and Lemma \[spread\_5\] $p \colon Y {\rightarrow}X$ and $q \colon Y {\rightarrow}Z$ are uniformly discrete completed normal coverings. By Proposition \[lamppu\], $p(B_p) \subset X$ and $q(B_q) \subset Z$ are closed. By Proposition \[gingerale\], $f(B_f) \subset Z$ is closed, since $q=f \circ p.$ Discrete completed normal coverings {#dcc} ----------------------------------- In this section we prove Theorems \[mm\] and \[relas\] in the introduction. Let $g \colon Y' \to Z'$ be a normal covering onto a large subset $Z'\subset Z$ of a PL manifold $Z,$ $p \colon Y {\rightarrow}Z$ the completion of $g \colon Y' {\rightarrow}Z,$ $z_0 \in Z'$ and $x_0 \in g^{-1}\{z_0\}.$ In general $p$ is not discrete, see [@MA Ex.10.6.], but if $p$ has locally finite multiplicity, then $p$ is discrete, see [@MA Thm.9.14.]. The discreteness of the map $p$ depends on the properties of the group $\pi(Z',z_0)/g_*(\pi(X',x_0))$ and the manifold $Z.$ To capture this relation we define stability of $g \colon Y' {\rightarrow}Z$ and show that $p \colon Y {\rightarrow}Z$ is discrete if and only if $g \colon Y' {\rightarrow}Z$ is stable. Let $\mathcal N_Z(z;z_0)$ be the set of open and connected neighbourhoods $U \subset Z$ of $z$ satisfying $z_0 \in U.$ We denote by $\iota_{U,Z'}$ the inclusion $U {\hookrightarrow}Z'.$ Let $Z$ be a PL manifold, $Z'\subset Z$ a large subset, $z_0 \in Z'$ and $N {\subset}\pi(Z',z_0)$ a normal subgroup. A subset $U \in \mathcal{N}_Z(z;z_0)$ is a $(Z',N)$*-stable* neighbourhood of $z$ if $$\mathrm{Im}(\pi \circ (\iota_{V \cap Z',Z'})_*)=\mathrm{Im}(\pi \circ (\iota_{U \cap Z',Z'})_*)$$ for every $V \in \mathcal{N}_Z(z;z_0)$ contained in $U;$ $$\xymatrix{\pi(V \cap Z',z_0) \ar[dr]_{(\iota_{V \cap Z',Z'})_*} && \pi(U \cap Z',z_0) \ar[dl]^{(\iota_{U \cap Z',Z'})_*}\\ &\pi(Z',z_0)\ar[d]_{\pi}&\\ &\pi(Z',z_0)/N,& }$$ where $\pi \colon \pi(Z',z_0) {\rightarrow}\pi(Z',z_0)/N$ is the canonical quotient map. \[stability\] Let $Z$ be a PL manifold and $Z' \subset Z$ a large subset, $z_0 \in Z'$ and $N {\subset}\pi(Z',z_0)$ a normal subgroup. The manifold $Z$ is $(Z',N)$-*stable* if every point $z \in Z$ has a $(Z',N)$-stable neighbourhood. \[stabhuomio\]Let $Z$ be a PL manifold, $Z' \subset Z$ a large subset, $z_0 \in Z'$ and $N'$ and $N$ normal subgroups of $\pi(Z',z_0)$ so that $N' \subset N.$ Then $Z$ is $(Z',N)$-stable if $Z$ is $(Z',N')$-stable. Indeed, if $U \in \mathcal{N}_Z(z,z_0)$ is a $(Z',N')$-stable neighbourhood of $z,$ then $U$ is a $(Z',N)$-stable neighbourhood of $z.$ A mapping $g \colon Y' {\rightarrow}Z$ is *stable* if $g \colon Y' {\rightarrow}g(Y')$ is a covering, $g(Y') \subset Z$ is large and $Z$ is $(g(Y'),\mathrm{Ker}(\sigma_g))$-stable for the monodromy $\sigma_g$ of the covering $g.$ We say that a completed covering $f \colon Y {\rightarrow}Z$ is a *stabily completed covering* if $f {|}Y' \colon Y' {\rightarrow}Z$ is stable for a large subset $Y' \subset Y.$ We say that a completed covering $f \colon Y {\rightarrow}Z$ is a *stabily completed normal covering*, if $g:=f {|}Y' \colon Y' {\rightarrow}Z$ is stable and $g \colon Y' {\rightarrow}g(Y')$ a normal covering for a large subset $Y' \subset Y.$ \[TFAE\] Let $p \colon Y {\rightarrow}Z$ be a completed normal covering, $Y' \subset Y$ and $Z' \subset Z$ large subsets so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering, $z_0 \in Z'$ and $y_0 \in g^{-1}\{z_0\}.$ Then the following conditions are equivalent: - $p\colon Y {\rightarrow}Z$ is discrete, - $p\colon Y {\rightarrow}Z$ is uniformly discrete and - $g \colon Y' {\rightarrow}Z$ is stable. The proof consists of three parts. Lemma \[spread\_5\] proves $(a) \Leftrightarrow (b),$ Proposition \[world\] proves $(c)\Rightarrow(b)$ and Proposition \[somac\] proves $(b)\Rightarrow(c).$ Next we are going to state two lemmas used in proving Proposition \[world\] and Proposition \[somac\]. \[D\_1\]Let $Z$ be a manifold, $Z' \subset Z$ large and $U$ and $V$ simply-connected open subsets of $Z$ so that $U \subset Z'$ and the set $U \cap V$ is a non-empty path-connected set. Let $z_0 \in U$ and $z_1\in U \cap V \cap Z'.$ Then for every path $\gamma \colon z_0 {\curvearrowright}z_1$ in $(U \cup V) \cap Z'$ and every loop ${\alpha}\colon (S_1,e_0) {\rightarrow}((U \cup V) \cap Z', z_0)$ there exists a loop ${\beta}\colon (S^1,e_0) {\rightarrow}(V \cap Z',z_1)$ so that $[{\alpha}]=[\gamma{\beta}\gamma^{{\leftarrow}}]$ in the group $\pi((U \cup V) \cap Z',z_0).$ We denote $W:=U \cup V.$ Let $\gamma \colon z_0 {\curvearrowright}z_1$ be a path in $W \cap Z'$ and ${\alpha}\colon (S^1,e_0) {\rightarrow}(W \cap Z',z_0)$ a loop. Since $U \subset Z',$ we have $W \cap Z'=U \cup (V \cap Z').$ Moreover, the sets $U$, $V \cap Z'$ and $U \cap (V \cap Z')$ are path-connected open subsets of $W \cap Z',$ since $Z' \subset Z$ is large. Since $U$ is simply-connected, the homomorphism $\iota_* \colon \pi(V \cap Z',z_1) {\rightarrow}\pi(W \cap Z',z_1)$ induced by the inclusion $\iota \colon V \cap Z' {\rightarrow}W \cap Z'$ is by the Seifert-van Kampen Theorem an epimorphism, see [@Mas Thm 4.1.]. Hence there exist a loop ${\beta}\colon (S^1,e_0) {\rightarrow}(V \cap Z',z_1)$ so that $[{\beta}]=[\gamma^{{\leftarrow}}{\alpha}\gamma]$ in $\pi(W \cap Z',z_1).$ Now $[{\alpha}]=[\gamma{\beta}\gamma^{{\leftarrow}}]$ in $\pi(W \cap Z',z_0).$ For the terminology of simplicial structures we refer to Hatcher [@H Ch.2]. We denote by $\sigma(v_0,\ldots, v_n)$ the simplex in ${\mathbb{R}}^k$ spanned by vertices $v_0, \ldots, v_n \in {\mathbb{R}}^k.$ The open simplex $\mathrm{Int}(\sigma):=\sigma \setminus \partial(\sigma)$ we call the simplicial interior of $\sigma.$ For a simplicial complex $K$ consisting of simplices in ${\mathbb{R}}^k$ the polyhedron is a subset $|K| \subset {\mathbb{R}}^k.$ For $z \in |K|$ we call $\mathrm{St}(z):={\bigcup}_{\sigma \in K; z \in \sigma} \mathrm{Int}(\sigma)$ the star neighbourhood $z$ (with respect to $K$). A *triangulation* of a manifold $Z$ is a simplicial complex $T$ satisfying $Z \approx |T|.$ We say that a manifold is a *PL manifold* if it has a triangulation. If $Z$ is a PL manifold and $T$ a triangulation of $Z$ with simplexes in ${\mathbb{R}}^k,$ then we assume from now on that $$Z=|T| \subset {\mathbb{R}}^k.$$ Let $K$ be a simplical complex. We will introduce for the polyhedron $|K|$ a topological basis related to the star neighbourhoods $\mathrm{St}(x), x \in |K|,$ and show a technical property for this particular basis. Let $x_0 \in |K|$ be a point and $t \in (0,1].$ For a simplex $\sigma=\sigma(v_0, \ldots, v_m) \in K$ for which $x_0 \in \sigma$ we denote by $\tau(\sigma,x_0,t)$ the simplex spanned by the vertices $v'_j=x_0+t(v_j-x_0), j \in \{0, \ldots, m\}.$ The $t$-*star* of $x_0$ is the set $$\mathrm{St}_t(x_0)= {\bigcup}_{\{\sigma \in K : x_0 \in \sigma\}} \mathrm{Int}(\tau(\sigma,x_0,t)),$$ where $\mathrm{Int}(\tau(\sigma,x_0,t))$ is the simplicial interior of $\tau(\sigma,x_0,t)$ for all $\sigma \in K$ satisfying $x_0 \in \sigma.$ We note that the collection $(\mathrm{St}_{1/n}(x_0))_{n \in {\mathbb{N}}}$ forms a neighbourhood basis of the point $x_0 \in |K|,$ and that there exists a contraction $$H \colon \mathrm{St}(x_0) \times [0,1] {\rightarrow}\mathrm{St}(x_0)$$ of the open star $\mathrm{St}(x_0)$ to $x_0$ for which $H(\mathrm{St}(x_0) \times \{1/n\})=\mathrm{St}_{1/n}(x_0)$ for every $n \in {\mathbb{N}}$ and $H(\{x\} \times [0,1])\subset [x,x_0]$ for all $x \in \mathrm{St}(x_0).$ \[D\_2\]Let $Z$ be a PL manifold, $Z' \subset Z$ large and $T$ a triangulation of $Z$ and $z_0 \in Z'$ a vertex in $T.$ Let $z \in Z$ be a point and $\mathrm{St}_{1/m}(z)$ the $1/m$-star of $z$ for $m>1.$ Suppose $X \subset Z$ is an open connected satisfying $z_0\in X$ and $\mathrm{St}_{1/m}(z) \subset X.$ Then there exists a simply-connected open set $U \subset X \cap Z'$ so that $z_0 \in U$ and that the set $U \cap \mathrm{St}_{1/m}(z)$ is non-empty and path-connected. Suppose $z_0 \in \overline{\mathrm{St}_{1/m}(z)}.$ Since $z_0$ is a vertex of $T$ and $m>1,$ we have $z=z_0.$ Thus $z=z_0 \in \mathrm{St}_{1/m}(z) \cap Z'.$ Since $\mathrm{St}_{1/m}(z)\cap Z' \subset Z$ is open, there exists a simply-connected neighbourhood $U\subset \mathrm{St}_{1/m}(z) \cap Z'$ of $z.$ This proves the claim in the case $z_0 \in \overline{\mathrm{St}_{1/m}(z)}.$ Suppose then that $z_0 \notin \overline{\mathrm{St}_{1/m}(z)}.$ Since $Z'\subset Z$ is large and $\mathrm{St}_{1/m}(z) \subsetneq X, $ there exists a point $z_1 \in \partial (\mathrm{St}_{1/m}(z)) \cap X \cap Z',$ that is a simplicial interior point of an $n$-simplex $\sigma \in T$ and a face $f$ of $\tau(\sigma,z,1/m).$ Since $|T|\subset {\mathbb{R}}^k$ for some $k \in {\mathbb{N}},$ the Euclidean metric of ${\mathbb{R}}^k$ induces a metric on $|T|$. Let $$r_1:=\text{min}\{\text{dist}(z_1, f) \mid f \text{ is a face of } \tau(\sigma,z,1/m) \text{ and } z_1 \notin f\}$$ and $$r_2=\text{dist}(z_1, Z \setminus (Z'\cap X )).$$ Then $r:=\text{min}\{r_1,r_2\}>0.$ We fix $k \in {\mathbb{N}}$ and a $1$-subcomplex $L$ in the $k^{th}$ barycentric subdivision $\mathrm{Bd}^k(T)$ of $T$ so that $|L|\subset |T|$ is a simple closed curve in $X \cap Z'\cap (Z \setminus \overline{\mathrm{St}_{1/m}(z)})$ from $z_0$ to $\mathrm{Int}(\sigma)$ satisfying $\text{dist}(z_1,|L| \cap \sigma) <r/4.$ We fix $l\geq k+2$ large enough so that for $$W:={\bigcup}_{z' \in |L|} \mathrm{St}^{l}(z'),$$ where $\mathrm{St}^{l}(z')$ is for every $z' \in |L|$ the open star of $z'$ in $\mathrm{Bd}^l(T),$ we have $$\overline{W} \subset X \cap Z' \cap (Z \setminus \overline{\mathrm{St}_{1/m}(z)}).$$ Then $|L|$ is a strong deformation retract of $W$ since $l\geq k+2;$ see regular neighbourhoods in [@RS Sec.2]. In particular, $W$ is simply-connected since $|L|$ is simply connected. Let $K < \mathrm{Bd}^l(T)$ be the $(n-1)$-subcomplex satisfying $\partial W=|K|$ and let $S$ be the collection of $(n-1)$-simplexes $\tau \in K$ for which $$\text{dist}(z_1, \overline{W} \cap \sigma)=\text{dist}(z_1, \tau).$$ Since $z_1$ is an interior point of $f$ and $\text{dist}(z_1,\overline{W} \cap \sigma)<r_1,$ there exists $\tau \in S$ and $z_2 \in \mathrm{Int}(\tau)$ for which the line segment $[z_1,z_2] \subset \mathrm{Int}(\sigma)$ has length at most $r/2$ and $$[z_1,z_2] \cap (\mathrm{St}_{1/m}(z) \cup W)=\{z_1,z_2\}.$$ Now, for all $t<r/2$ the $t$-neighbourhood $W_t \subset \mathrm{Int}(\sigma)$ of $[z_1,z_2]$ satisfies $W_t \subset X \cap Z'$ and the intersections $W \cap W_t$ and $\mathrm{St}_{1/m}(z) \cap W_t$ are non-empty. Since $z_2 \in \mathrm{Int}(\tau),$ the intersections $W \cap W_{t_0}$ and $\mathrm{St}_{1/m}(z) \cap W_{t_0}$ are non-empty intersections of two convex sets for some $t_0<r/2$. Thus $W \cap W_{t_0}$ and $\mathrm{St}_{1/m}(z) \cap W_{t_0}$ are path-connected. Since the sets $W_{t_0}$ and $W$ are simply-connected and the set $W \cap W_{t_0}$ path-connected, the set $U=W \cup W_{t_0} \subset Z'$ is simply-connected by the Seifert-van Kampen Theorem. Further, $z_0 \in W \subset U.$ Since $\mathrm{St}_{1/m}(z) \cap W_{t_0}$ is path-connected and $U \cap \mathrm{St}_{1/m}(z)=\mathrm{St}_{1/m}(z) \cap W_{t_0},$ this proves the claim. \[world\]Let $p \colon Y {\rightarrow}Z$ be a completed normal covering, $Y' \subset Y$ and $Z' \subset Z$ large subsets so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering. If $g \colon Y' {\rightarrow}Z$ is stable, then $p \colon Y \to Z$ is uniformly discrete. Let $z_0 \in Z'$ and $y_0 \in g^{-1}\{z_0\}.$ Let $z \in Z.$ Since $g$ is stable the manifold $Z$ is $(Z',g_*(\pi(Y',y_0))$-stable. Let $T$ be a triangulation of $Z$ having $z_0$ a vertex. Let $V$ be a $(Z',g_*(\pi(Y',y_0)))$-stable neighbourhood of $z$ and $D$ a component of $p^{-1}(V).$ It is now sufficient to show that $p^{-1}\{z\} \cap D$ is a point. By Lemma \[brunate\], we have $p(D)=V,$ and hence $p^{-1}\{z\} \cap D\neq {\varnothing}.$ We suppose towards a contradiction that there are distinct points $d_1$ and $d_2$ in $p^{-1}\{z\} \cap D.$ Using points $d_1$ and $d_2,$ we construct a loop ${\alpha}_1 \colon (S^1,e_0) {\rightarrow}(V \cap Z',z_0)$ and $V_1 \in \mathcal{N}_Z(z,z_0)$ so that $V_1 \subset V$ and $$[{\alpha}_1]g_*(\pi(Y',y_0)) \notin \mathrm{Im}(\pi \circ \iota_{Z' \cap V_1,Z'})).$$ This contradicts the $(Z',g_*(\pi(Y',y_0)))$-stability of $Z$ and proves the claim. We begin by choosing the neighbourhood $V_1 \subset V$ of $z.$ For all $n \in {\mathbb{N}}$ let $\mathrm{St}_{1/n}(z)$ be the $1/n$-star of $z$ in the triangulation $T.$ Then $(\mathrm{St}_{1/n}(z))_{n \in {\mathbb{N}}}$ is a neighbourhood basis for $z.$ Since $p \colon Y \to Z$ is a spread and $Y$ a Hausdorff space, there exists $n \geq 2$ so that $\mathrm{St}_{1/n}(z) \subset V$ and the $d_i$-components $C_i,$ $i \in \{1,2\},$ of $p^{-1}(\mathrm{St}_{1/n}(z))$ are disjoint. By Lemma \[D\_2\], there is such an open simply-connected set $W \subset V \cap Z'$ containing $z_0$ that the sets $W \cup \mathrm{St}_{1/n}(z)$ and $W \cap \mathrm{St}_{1/n}(z)$ are path-connected. Hence $$V_1=W \cup \mathrm{St}_{1/n}(z) \subset V$$ is an open connected neighbourhood of $z$ containing $z_0$. Configuration illustrated in Figure \[wer\]. ![[]{data-label="wer"}](Martina.1) We continue by choosing a loop ${\alpha}_1 \colon (S^1,e_0) {\rightarrow}(V \cap Z',z_0)$. Fix $z_1 \in W \cap \mathrm{St}_{1/n}(z) \cap Z'.$ Since $W$ is a connected set containing points $z_0$ and $z_1,$ there is a path $\gamma \colon z_0 {\curvearrowright}z_1$ in $W.$ Since $\mathrm{St}_{1/n}(z)$ is connected, Lemma \[brunate\] implies that there are points $c_i \in p^{-1}\{z_1\} \cap C_i$ for $i \in \{1,2\}.$ By Lemma \[biotin\], $p^{-1}(Z')=Y'.$ Thus $c_1, c_2 \in Y' \cap D.$ Since $Y' \subset Y$ is large, the set $Y \setminus Y'$ does not locally separate $Y$ and the set $D \cap Y'$ is path-connected. Let ${\beta}\colon [0,1] {\rightarrow}D \cap Y'$ be a path ${\beta}\colon c_1 {\curvearrowright}c_2.$ Then $g \circ {\beta}\colon [0,1] {\rightarrow}V \cap Z'$ is a loop at $z_1.$ We set ${\alpha}_1=\gamma(g \circ {\beta}) \gamma^{\leftarrow}.$ Suppose there is a loop ${\alpha}_2 \colon (S^1,e_0) {\rightarrow}(V_1 \cap Z',z_0)$ satisfying $$[{\alpha}_2] \in [\gamma(g \circ {\beta}) \gamma^{\leftarrow}]g_*(\pi(Y',y_0)).$$ Since $W \subset Z'$ and $\gamma$ is a path in $W \subset V_1 \cap Z',$ there is by Lemma \[D\_1\] a loop ${\alpha}_3 \colon (S^1,e_0) {\rightarrow}(\mathrm{St}_{1/n}(z) \cap Z',z_1)$ satisfying $$[{\alpha}_2]=[\gamma {\alpha}_3 \gamma^{\leftarrow}].$$ Next we change the base point of $Y$ as follows. Since $g \colon Y' {\rightarrow}Z'$ is a normal covering and $W \subset Z',$ the path $\gamma^{\leftarrow} \colon [0,1] {\rightarrow}W$ has a lift ${\widetilde}{\gamma^{\leftarrow}}_{c_1} \colon [0,1]{\rightarrow}Y'.$ We set $y_1:={\widetilde}{\gamma^{\leftarrow}}_{c_1}(1).$ Since $g \colon Y' {\rightarrow}Z'$ is a normal covering, we have $$g_*(\pi(Y',y_1))=g_*(\pi(Y',y_0)).$$ Hence $$[\gamma {\alpha}_3 \gamma^{\leftarrow}] \in [\gamma(g \circ {\beta}) \gamma^{\leftarrow}]g_*(\pi(Y',y_0))=[\gamma(g \circ {\beta}) \gamma^{\leftarrow}]g_*(\pi(Y',y_1)).$$ Thus the lifts ${\widetilde}{\gamma {\alpha}_3 \gamma^{\leftarrow}}_{y_1}$ and ${\widetilde}{\gamma(g \circ {\beta}) \gamma^{\leftarrow}}_{y_1}$ have a common endpoint. By the uniqueness of lifts, $${\widetilde}{{\alpha}_3}_{c_1}(1)={\widetilde}{(g \circ {\beta})}_{c_1}(1)={\beta}(1)=c_2.$$ On the other hand, since ${\alpha}_3$ is a path in $\mathrm{St}_{1/n}(z),$ the lift ${\widetilde}{{\alpha}_3}_{c_1}$ is a path in $p^{-1}(\mathrm{St}_{1/n}(z)).$ Since $C_1$ is a component of $p^{-1}(\mathrm{St}_{1/n}(z)),$ this implies that ${\widetilde}{{\alpha}_3}_{c_1}[0,1] \subset C_1.$ Thus $c_2={\widetilde}{{\alpha}_3}_{c_1}(1) \in C_1,$ which is a contradiction. Hence $p^{-1}\{p(y)\} \cap D$ is a point. \[somac\] Let $p \colon Y {\rightarrow}Z$ be a completed normal covering, $Y' \subset Y$ and $Z' \subset Z$ large subsets so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering. If $p$ is uniformly discrete, then $g \colon Y' {\rightarrow}Z$ is stable. Let $z_0 \in Z',$ $y_0 \in g^{-1}\{z_0\}$ and $T$ a triangulation of $Z$ so that $z_0$ is a vertex. We prove the statement by constructing a $(Z',g_*(\pi(Y',y_0))$-stable neighbourhood for every $z \in Z.$ Suppose that $z=z_0.$ Since $p$ is uniformly discrete and $z_0 \in Z',$ there exists a simply-connected neighbourhood $V \subset Z'$ of $z$ so that $p^{-1}\{z\} \cap D$ is a point for each component $D$ of $p^{-1}(V).$ Then $V$ is a $(Z',g_*(\pi(Y',y_0))$-stable neighbourhood of $z,$ since $V \subset Z'$ and $V$ is simply-connected. Let then $z \in Z \setminus \{z_0\}.$ For all $n \in {\mathbb{N}}$ let $\mathrm{St}_{1/n}(z)$ be the $1/n$-star of $z$ in the triangulation $T.$ Since $(\mathrm{St}_{1/n}(z))_{n \in {\mathbb{N}}}$ is a neighbourhood basis for $z$ and $p$ is uniformly discrete, there exists $n \geq 2$ so that $p^{-1}\{z\} \cap D$ is a point for each component $D$ of $p^{-1}(\mathrm{St}_{1/n}(z)).$ By Lemma \[D\_2\], there is an open simply-connected set $U \subset Z'$ containing $z_0$ so that $U \cup \mathrm{St}_{1/n}(z)$ and $U \cap \mathrm{St}_{1/n}(z)$ are path-connected. Let $V:=U \cup \mathrm{St}_{1/n}(z).$ Towards showing that $V$ is a $(Z',g_*(\pi(Y',y_0))$-stable neighbourhood of $z,$ let ${\alpha}\colon (S^1,e_0) {\rightarrow}(V \cap Z',z_0)$ be a loop and let $V_1 \in \mathcal{N}_Z(z;z_0)$ be an open set contained in $V.$ It is sufficient to show that there exists a loop ${\beta}\colon (S^1,e_0) {\rightarrow}(V_1 \cap Z',z_0)$ satisfying $[{\beta}]g_*(\pi(Y',y_0))=[{\alpha}]g_*(\pi(Y',y_0)).$ We choose first a suitable representative from the class $[{\alpha}]$ as follows. Let $W$ be the $z$-component of $\mathrm{St}_{1/n}(z) \cap V_1.$ Since $Z' \subset Z$ is large, there exists a point $z_1 \in W \cap Z'$ and a path $\gamma \colon z_0 {\curvearrowright}z_1$ in $V_1 \cap Z'.$ By Lemma \[D\_1\], there exists a loop ${\alpha}_1 \colon (S^1,e_0) {\rightarrow}(\mathrm{St}_{1/n}(z) \cap Z' ,z_1)$ for which $[{\alpha}]=[\gamma {\alpha}_1 \gamma^{\leftarrow}]$ in $\pi(V \cap Z',z_0).$ Denote the endpoint of the lift ${\widetilde}{\gamma_{y_0}}$ in $g$ by $y_1$ and the endpoint of the lift ${\widetilde}{{\alpha}_1}_{y_1}$ in $g$ by $y_2.$ Let $D_1$ be the $y_1$-component of $p^{-1}(\mathrm{St}_{1/n}(z)).$ We proceed by showing that there exists a loop ${\alpha}_2 \colon (S^1,e_0) {\rightarrow}(W \cap Z',z_1)$ so that the lift ${\widetilde}{{\alpha}_2}_{y_1}$ in $g$ that begins at $y_1$ ends at $y_2.$ Let $C_1\subset D_1$ be the $y_1$-component of $p^{-1}(W).$ Since $W$ is a connected neighbourhood of $z,$ every component of $p^{-1}(W)$ intersects the set $p^{-1}\{z\}$ by Lemma \[brunate\]. Since $W \subset \mathrm{St}_{1/n}(z)$ and the set $p^{-1}\{z\} \cap D_1$ is a point, this implies $$p^{-1}(W) \cap D_1=C_1.$$ Hence the points $y_1$ and $y_2$ belong to $C_1 \cap Y'.$ Since $Y \setminus Y'$ does not locally separate $Y,$ there is a path ${\beta}\colon y_1 {\curvearrowright}y_2$ in $C_1 \cap Y'.$ Let ${\alpha}_2:=g \circ {\beta}.$ Since the lifts ${\widetilde}{\gamma {\alpha}_1 \gamma^{\leftarrow}}_{y_0}$ and ${\widetilde}{\gamma {\alpha}_2 \gamma^{\leftarrow}}_{y_0}$ end at the common point, $$[\gamma {\alpha}_2 \gamma^{\leftarrow}]\in [\gamma{\alpha}_1\gamma^{\leftarrow}]g_*(\pi(Y',y_0))= [{\alpha}]g_*(\pi(Y',y_0)).$$ Since $\gamma {\alpha}_2 \gamma^{\leftarrow}$ is a loop in $V_1 \cap Z'$ at $z_0,$ $V$ is a $(Z',g_*(\pi(Y',y_0)))$-stable neighbourhood of the point $z.$ Since this holds for every $z \in Z$ the manifold $Z$ is $(Z',g_*(\pi(Y',y_0)))$-stable. This concludes the proof of Theorem \[TFAE\]. After the following two consequences of Theorem \[TFAE\] we are ready to prove Theorems \[ilta\], \[mm\] and \[relas\] in the introduction. \[viimeinen\] Let $f \colon X {\rightarrow}Z$ be a completed covering between PL manifolds, $X' \subset X$ and $Z' \subset Z$ large subsets and $$\xymatrix{ & Y' \ar[ld]_{p'} \ar[rd]^{q'} &\\ X' \ar[rr]^{f'} & & Z' }$$ a commutative diagram of coverings so that $p'$ and $q'$ are normal coverings. Let $z_0 \in Z'$ and $y_0 \in q'^{-1}\{z_0\}$, $N=q'_*(\pi(Y',y_0))$ and $$\xymatrix{ & Y \ar[ld]_{p} \ar[rd]^{q} &\\ X \ar[rr]^{f} & & Z }$$ be a commutative diagram of open continuous maps so that $p$ is the completion of $p' \colon Y' {\rightarrow}X$ and $q$ is the completion of $p' \colon Y' {\rightarrow}Z.$ Then, the maps $p$ and $q$ are orbit maps if and only if $Z$ is $(Z',N)$-stable. Let $f \colon X {\rightarrow}Z$ be a completed covering between manifolds and $q \colon Y {\rightarrow}Z$ a completed covering that is an orbit map. We say that $q$ is *natural of* $f$ if there are large subsets $X' \subset X,$ $Y' \subset Y$ and $Z' \subset Z$ so that $g:=f {|}X'\colon X' {\rightarrow}Z'$ and $h:=q {|}Y' \colon Y' {\rightarrow}Z'$ are coverings satisfying $\mathrm{Ker}(\sigma_g)=\mathrm{Ker}(\sigma_h).$ We recall that the monodromy group of $f$ is the monodromy group of $g.$ \[hh\] Let $f \colon X {\rightarrow}Z$ be a stabily completed covering between PL manifolds. Then $f$ has a monodromy representation $(Y,p,q),$ where $q$ is natural of $f.$ Let $X'\subset X$ and $Z' \subset Z$ be large subsets so that $g:=f {|}X' \colon X' {\rightarrow}Z'$ is a covering and $Z$ is $(Z',\mathrm{Ker}(\sigma_g))$-stable. By Theorem \[spread\_3\] there is a normal covering $p' \colon Y' {\rightarrow}X'$ so that $q':=g \circ p'$ is a normal covering satisfying $q_*'(\pi(Y',y_0))=\mathrm{Ker}(\sigma_g).$ Thus $\mathrm{Ker}(\sigma_{q'})=q_*'(\pi(Y',y_0))=\mathrm{Ker}(\sigma_g)$ and the deck-transformation group ${\mathcal{T}}(q')$ is isomorphic to the monodromy group of $g.$ Let $p \colon Y {\rightarrow}X$ be the completion of $p' \colon Y' {\rightarrow}X$ and $q:=f\circ p \colon Y {\rightarrow}Z.$ Then $q$ is the completion of $q' \colon Y' {\rightarrow}Z$ by Proposition \[elama\] and $q$ is natural of $f.$ Since ${\mathcal{T}}(q')$ is isomorphic to the monodromy group of $g,$ it is sufficient to show that $p$ and $q$ are orbit maps and ${\mathcal{T}}(q)\cong {\mathcal{T}}(q').$ Since $Z$ is $(Z',\mathrm{Ker}(\sigma_{q'}))$-stable, $q$ is discrete by Theorem \[TFAE\]. Thus $q$ is an orbit map by Theorem \[vauva\], and ${\mathcal{T}}(q)\cong {\mathcal{T}}(q')$ by Remark \[huu\]. Since $q=f \circ p$ is discrete, $p$ is discrete. Thus $p$ is an orbit map by Theorem \[vauva\]. Let $p \colon Y {\rightarrow}Z$ be a completed normal covering onto a PL manifold $Z.$ Let $Y' \subset Y$ and $Z' \subset Z$ be large subsets so that $g:= p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering. Since $p$ is discrete, $g \colon Y' {\rightarrow}Z$ is stable by Theorem \[TFAE\]. Thus $p$ is a stabily completed normal covering. Suppose then that $p$ is a stabily completed normal covering. Then there are such large subsets $Y'' \subset Y$ and $Z'' \subset Z$ that $h:=p {|}Y'' \colon Y'' {\rightarrow}Z''$ is a normal covering and the map $h \colon Y'' {\rightarrow}Z$ is stable. Thus the map $p$ is discrete by Theorem \[TFAE\]. Suppose $f \colon X {\rightarrow}Z$ is a completed covering between PL manifolds and $(Y,p,q)$ a monodromy representation of $f.$ By Corollary \[viimeinen\], the maps $p$ and $q$ are stabily completed normal coverings. Since $q$ is a stabily completed covering, as a consequence of Theorem \[spread\_3\] and Remark \[stabhuomio\], also $f$ is a stabily completed covering. By Theorem \[mm\] and Theorem \[hh\] a completed covering $f \colon X {\rightarrow}Z$ between PL manifolds is a stabily completed covering if and only if $f$ has a monodromy representation $(Y,p,q).$ We end this section with a corollary of Theorem \[viimeinen\]. We recall that for a simplical $n$-complex $K,$ $K^{n-2}$ is the codimension $2$ skeleton of $K.$ \[simplisiaalinen\] Let $f \colon X {\rightarrow}Z$ be a surjective, open and discrete simplicial map from a PL $n$-manifold $X$ with triangulation $K$ onto a PL $n$-manifold $Z$ with triangulation $L.$ Let also $z_0 \in |L| \setminus |L^{n-2}|$ and $x_0 \in f^{-1}\{z_0\}.$ Then $g:=f {|}|K| \setminus |K^{n-2}| \colon |K| \setminus |K^{n-2}| {\rightarrow}|L| \setminus |L^{n-2}|$ is a covering and for all normal subgroups $N {\subset}\pi(|L| \setminus |L^{n-2}|,z_0)$ is contained in $\mathrm{Ker}(\sigma_g)$ for the monodromy $\sigma_g$ of $g,$ there is a commutative diagram $$\xymatrix{ & Y \ar[ld]_{p} \ar[rd]^{q} &\\ X \ar[rr]^{f} & & Z }$$ of completed coverings so that $p$ and $q$ are orbit maps and the deck-transformation groups ${\mathcal{T}}(p)$ and ${\mathcal{T}}(q)$ satisfy $${\mathcal{T}}(p)\cong \pi(|K| \setminus |K^{n-2}|,x_0)/N \text{ and } {\mathcal{T}}(q) \cong \pi(|L| \setminus |L^{n-2}|,z_0)/N .$$ Since $f \colon X {\rightarrow}Z$ is an open discrete simplicial map the branch set of $f$ is contained in $|K^{n-2}|.$ In particular $g:=f {|}|K| \setminus |K^{n-2}| \colon |K| \setminus |K^{n-2}| {\rightarrow}|L| \setminus |L^{n-2}|$ is a covering and $f$ is the completion of $f \colon |K| \setminus |K^{n-2}| {\rightarrow}Z.$ Clearly, $Z$ is $(|L| \setminus |L^{n-2}|,\{e\})$-stable for the trivial group $\{e\} \subset \pi( |L| \setminus |L^{n-2}|,z_0).$ Further, $Z$ is $(|L| \setminus |L^{n-2}|, N )$-stable by Remark \[stabhuomio\]. We conclude the statement from Theorem \[spread\_3\] and Theorem \[viimeinen\]. If the map $p \colon Y {\rightarrow}X$ in Corollary \[simplisiaalinen\] has locally finite multiplicity, then the space $Y$ is a locally finite simplicial complex and $p$ and $q$ are simplicial maps, see [@F Sec.6]. However, in the next section we show that $p$ need not have locally finite multiplicity. Existence of locally compact monodromy representations ====================================================== Uniformly bounded local multiplicities -------------------------------------- In this section we show the existence of a completed covering that has a monodromy representation that is not locally compact. Let $f \colon X {\rightarrow}Z$ be a completed covering and $z \in Z.$ We say that $f$ has *uniformly bounded local multiplicities* in $f^{-1}\{z\} \subset X$ if there exist a neighbourhood $U \in \mathcal{N}_Z(z)$ so that every point $z' \in U$ satisfies $$\sup\{\#(f^{-1} \{z'\} \cap D) : D \text{ is a component of }f^{-1}(U)\}< \infty.$$ ![[]{data-label="qer"}](kaavio.1) \[kotiin\] Suppose $f \colon X {\rightarrow}Z$ is a completed covering and $(Y,p,q)$ is a locally compact monodromy representation of $f.$ Then $f$ has uniformly bounded local multiplicities in $f^{-1}\{z\} \subset X$ for every $z \in Z.$ Let $z \in Z$ and $y \in q^{-1}\{z\}.$ Since $q \colon Y {\rightarrow}Z$ is an open and discrete map, the map $q$ has locally finite multiplicity, since $Y$ is locally compact. Thus there exists such $U \in \mathcal{N}_Z(z)$ that $q {|}D_y$ has finite multiplicity for the $y$-component $D_y$ of $q^{-1}(U).$ Let $z' \in U$ and $k_{z'}:=\#(q^{-1}\{z'\} \cap D_y).$ Since $q$ is an orbit map, $$\sup\{\#(q^{-1} \{z'\} \cap D) : D \text{ is a component of }q^{-1}(U)\}= k_{z'}.$$ Hence $$\sup\{\#(f^{-1} \{z'\} \cap E) : E \text{ is a component of }f^{-1}(U)\}\leq k_{z'} < \infty,$$ since $q=f \circ p.$ Thus $f$ has uniformly bounded local multiplicities in $f^{-1}\{z\}.$ Suppose $f \colon X {\rightarrow}Z$ is a completed covering between manifolds so that there exists $z \in Z$ so that $f$ does not have uniformly bounded local multiplicities in $f^{-1}\{z\} \subset X.$ Then, as a consequence of Lemma \[kotiin\], $f$ is not an orbit map. There exists a BLD-mapping from ${\mathbb{R}}^2$ to the $2$-sphere $S^2$ having a monodromy representation that is not locally compact. Let $f \colon {\mathbb{R}}^2 {\rightarrow}S^2$ be a BLD-mapping for which $f(B_f) \subset S^2$ is a finite set, but for a point $z \in S^2$ the map $f$ does not have uniformly bounded local multiplicities in $f^{-1}\{z\} \subset {\mathbb{R}}^2$; see Figure \[qer\]. Since $f$ is BLD-mapping and $f(B_f)\subset S^2$ is closed, $f$ is a completed covering by [@L]. Since $f(B_f) \subset S^2$ is a finite set, $f$ is a stabily completed covering. Thus it has a monodromy representation $(Y,p,q)$ by Theorem \[hh\]. By Proposition \[kotiin\], $(Y,p,q)$ is not locally compact. Completed normal coverings with locally finite multiplicity {#brr} ----------------------------------------------------------- Let $Z$ be a PL manifold and $f \colon X {\rightarrow}Z$ a stabily completed covering. Then there are large subsets $Y' \subset Y$ and $Z' \subset Z$ so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a covering and $Z$ is $(Z', \mathrm{Ker}(\sigma_g))$-stable. Fix $z_0 \in Z'.$ Let $z \in Z$ and $U \in \mathcal{N}_Z(z,z_0)$ be a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $z.$ Let $\pi \colon \pi(Z',z_0) {\rightarrow}\pi(Z',z_0)/\mathrm{Ker}(\sigma_g)$ be the canonical map. \[br\] A group $H \subset \pi(Z',z_0)/\mathrm{Ker}(\sigma_g)$ is a *local monodromy group* of the completed covering $f \colon X {\rightarrow}Z$ at $z$ if there exists a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood $U \in \mathcal{N}_Z(z,z_0)$ of $z$ so that $H=\mathrm{Im}(\pi \circ (\iota_{Z' \cap U, Z'})_*)$ for the inclusion $\iota_{Z' \cap U,Z'}$. For a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood $U \in \mathcal{N}_Z(z,z_0)$ we denote $$\text{Mono}_f(U;Z'):=\mathrm{Im}(\pi \circ (\iota_{Z' \cap U, Z'})_*).$$ If $U, V \in \mathcal{N}_Z(z,z_0)$ are $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhoods of $z \in Z,$ then $$\text{Mono}_f(V;Z')=\text{Mono}_f(U;Z')$$ if $V \subset U.$ \[whynot\]Let $Z$ be a PL manifold and $p \colon Y {\rightarrow}Z$ a stabily completed normal covering. Let $Y' \subset Y$ and $Z' \subset Z$ be such large subsets that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering and $Z$ is $(Z', \mathrm{Ker}(\sigma_g))$-stable. Let $y \in Y,$ $z:=p(y)$ and $V$ be a $(Z', \mathrm{Ker}(\sigma_g)))$-stable neighbourhood of the point $z.$ Let $Z$ have a triangulation $T$ having a vertex $z_0 \in Z'$ and let $n \geq 2$ be such that $\mathrm{St}_{1/n}(z)\subset V.$ Let $D_y$ be the $y$-component of $p^{-1}(\mathrm{St}_{1/n}(z)).$ Then - $\#(p^{-1}\{z'\}\cap D_y) = \#\mathrm{Mono}_p(V;Z')$ for $z' \in \mathrm{St}_{1/n}(z) \cap Z'$ and - $\#(p^{-1}\{z'\}\cap D_y) \leq \#\mathrm{Mono}_p(V;Z')$ for $z' \in \mathrm{St}_{1/n}(z).$ We first prove claim (a). Let $z'$ and $z''$ be points in $\mathrm{St}_{1/n}(z) \cap Z'.$ Then $\#(p^{-1}\{z'\}\cap D_y)=\#(p^{-1}\{z''\}\cap D_y),$ since $\mathrm{St}_{1/n}(z) \cap Z'$ is connected and $D_y \cap Y'$ is a component of $g^{-1}(\mathrm{St}_{1/n}(z) \cap Z').$ Thus we only need to show that there is a point $z_1 \in \mathrm{St}_{1/n}(z) \cap Z'$ satisfying $\#(p^{-1}\{z_1\}\cap D_y)=\#\text{Mono}_p(V).$ By Lemma \[D\_2\], there is an open set $U \subset V \cap Z'$ so that $U \cap \mathrm{St}_{1/n}(z)$ is path-connected and $V_n:=U \cup \mathrm{St}_{1/n}(z) \in \mathcal{N}_Z(z,z_0).$ Since $V_n\subset V,$ the set $V_n$ is a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $z.$ Hence $$\#\mathrm{Im}(\pi \circ (\iota_{V_n \cap Z',Z'})_*)=\#\mathrm{Im}(\pi \circ (\iota_{V \cap Z',Z'})_*).$$ Let $z_1 \in V_n \cap Z'$ and let ${\beta}\colon [0,1]{\rightarrow}V_n \cap Z'$ be a path ${\beta}\colon z_0 {\curvearrowright}z_1.$ Let $y_0 \in p^{-1}\{z_0\}$ and $y_1:={\widetilde}{{\beta}}_{y_0}(1).$ Let $D_{y_1}$ be the $y_1$-component of $f^{-1}(\mathrm{St}_{1/n}(z)).$ Since $p$ is an orbit map, there is by Lemma \[vindpark\] a deck-transformation $\tau \in {\mathcal{T}}(p)$ satisfying $\tau(D_y)=D_{y_1}.$ Hence $$\#(p^{-1}\{z_1\}\cap D_y)=\#(p^{-1}\{z_1\}\cap D_{y_1}).$$ Since $g$ is a normal covering, we have $\mathrm{Ker}(\sigma_g)=\mathrm{Im}(g_*).$ By Lemma \[D\_1\] there is for every loop $\gamma \colon (S^1,e_0){\rightarrow}(V_n \cap Z',z_0)$ a loop ${\alpha}\colon (S^1,e_0){\rightarrow}(\mathrm{St}_{1/n}(z) \cap Z',z_1)$ satisfying $[\gamma]=[{\beta}{\alpha}{\beta}^{{\leftarrow}}]$ in $V_n \cap Z'.$ Hence $$\#(p^{-1}\{z_1\}\cap D_{y_1})=\#\mathrm{Im}(\pi \circ (\iota_{V_n \cap Z',Z'})_*),$$ since $g^{-1}(Z')=Y'$ by Lemma \[biotin\]. We conclude that $$\begin{aligned} \#(p^{-1}\{z_1\}\cap D_y)=&\#(p^{-1}\{z_1\}\cap D_{y_1})=\#\mathrm{Im}(\pi \circ (\iota_{V_n \cap Z',Z'})_*)\\=&\#\mathrm{Im}(\pi \circ (\iota_{V \cap Z',Z'})_*)=\#\text{Mono}_p(V).\end{aligned}$$ We then prove claim (b). Since $p^{-1}\{z'\}$ is a countable set for every $z' \in Z,$ the statements holds trivially if $\#\text{Mono}_p(V)=\infty.$ Suppose $\#\text{Mono}_p(V)=k$ for $k \in {\mathbb{N}}.$ Towards a contradiction suppose that $z'\in \mathrm{St}_{1/n}(z) \setminus Z'$ and there exists $k+1$ points $y_0, \ldots, y_k$ in $p^{-1}\{z'\}\cap D_y.$ Let $U\subset \mathrm{St}_{1/n}(z)$ be an open connected set so that the $y_i$-components $D'_{y_i}\subset D$ of $p^{-1}(U)$ are pairwise disjoint. By Lemma \[brunate\], $p(D'_{y_i})=U$ for every $i \in \{0, \ldots, k\}.$ Since $Z'\subset Z$ is large, there exists a point $z'_1 \in U \cap Z'$ for which $\#(p^{-1}\{z'\}\cap D_y)\geq k.$ This is a contradiction with (a). Hence $\#(p^{-1}\{z'\}\cap D_y)\leq k$ for all $z' \in \mathrm{St}_{1/n}(z).$ \[torstai\] Let $Z$ be a PL manifold and $p \colon Y {\rightarrow}Z$ a discrete completed normal covering. Then $p$ has locally finite multiplicity if and only if $p$ has a finite local monodromy group at each $z \in Z$. The map $p$ is an orbit map by Theorem \[vauva\]. Thus there is by Theorem \[TFAE\] large subsets $Y' \subset X$ and $Z' \subset Z$ so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering and $Z$ is $(Z', \mathrm{Ker}(\sigma_g))$-stable. Fix a triangulation $T$ of $Z$ having a vertex $z_0 \in Z'$. Suppose first that $p$ has locally finite multiplicity. Let $z \in Z$ and $y \in p^{-1}\{z\}.$ Let $D_n$ be the $y$-component of $p^{-1}(\mathrm{St}_{1/n}(z))$ for every $n \geq 1.$ Then $(D_n)_{n \in {\mathbb{N}}}$ is a neighbourhood basis of $y,$ since $p$ is a spread. Let $V \in \mathcal{N}_Z(z,z_0)$ be a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $z.$ Since $p$ has locally finite multiplicity, there exists $n \geq 2$ so that $\mathrm{St}_{1/n}(z)\subset V$ and $\#(p^{-1}\{z'\}\cap D_{n})<\infty$ for every $z' \in \mathrm{St}_{1/n}(z).$ Let $z_1 \in \mathrm{St}_{1/n}(z) \cap Z'.$ By Lemma \[whynot\], $$\#\text{Mono}_p(V)=\#(p^{-1}\{z_1\}\cap D_n)<\infty.$$ Suppose then that $p$ has a finite local monodromy group at each $z \in Z$. Let $y \in Y$ and $z:=p(y).$ Let $V\in \mathcal{N}_Z(z,z_0)$ be a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of the point $z$ so that $\#\text{Mono}_p(V)< \infty.$ Let $n \geq 2$ be such that $\mathrm{St}_{1/n}(z)\subset V$ and $D_y$ be the $y$-component of $p^{-1}(\mathrm{St}_{1/n}(z)).$ By Lemma \[whynot\], $p {|}D_y$ has finite multiplicity. Thus $p$ has locally finite multiplicity. Next we show regularity results for local monodromy. \[viimeista\] Let $f \colon X {\rightarrow}Z$ be a stabily completed covering between PL manifolds. Let $$\xymatrix{ & Y \ar[ld]_{p} \ar[rd]^{q} &\\ X \ar[rr]^{f} & & Z }$$ be a commutative diagram of completed coverings so that $p$ and $q$ are orbit maps and $q$ is natural of $f.$ Then $q$ has locally finite multiplicity if and only if every $z \in Z$ has a finite local monodromy group of $f$ at $z.$ By naturality of $q$ we may fix large subsets $X' \subset X,$ $Y' \subset Y$ and $Z' \subset Z$ for which $g:=f {|}X' \colon X' {\rightarrow}Z'$ and $h:=q {|}Y' \colon Y' {\rightarrow}Z'$ are coverings satisfying $\mathrm{Ker}(\sigma_{g})=\mathrm{Ker}(\sigma_{h}).$ Suppose first that $f$ has a finite local monodromy group of $f$ at each point of $Z$. Let $z \in Z.$ Then there is a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood $V \in \mathcal{N}_{Z}(z,z_0)$ of $z$ for which $\#\text{Mono}(V) <\infty.$ Since $\mathrm{Ker}(\sigma_{g})=\mathrm{Ker}(\sigma_{h}),$ $V$ is a $(Z', \mathrm{Ker}(\sigma_h))$-stable neighbourhood of $z.$ Since $\#\text{Mono}(V) <\infty,$ $q$ has a finite local monodromy group at $z.$ Thus $q$ has locally finite multiplicity by Theorem \[torstai\]. Suppose then that $q$ has locally finite multiplicity. Let $z \in Z.$ By Theorem \[torstai\] $q$ has a finite local monodromy group at $z$. Since $\mathrm{Ker}(\sigma_{g})=\mathrm{Ker}(\sigma_{h}),$ the finite local monodromy group of $q$ at $z$ is a finite local monodromy group of $f$ at $z.$ As a direct consequence of Theorem \[viimeista\] we obtain an analogy of Theorem \[relas\]. \[matka\] Let $f \colon X {\rightarrow}Z$ be a stabily completed covering between PL manifolds, so that $f$ has a finite local monodromy group at each $z \in Z.$ Then there is a monodromy representation $(Y,p,q)$ of $f,$ where $q$ has locally finite multiplicity. We conclude this section by defining for a PL manifold $Z,$ completed normal covering $p \colon Y {\rightarrow}Z$ and $y \in Y,$ the homotopical index ${\mathcal{H}}(y,p).$ By Lemmas \[whynot\] and \[torstai\] we may define as follows. \[prev\]Let $Z$ be a PL manifold, $p \colon Y {\rightarrow}Z$ a completed normal covering having locally finite multiplicity and $y \in Y.$ Let $Y' \subset Y$ and $Z' \subset Z$ be large subsets so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering. Let $T$ be a triangulation of $Z$ having a vertex $z_0 \in Z',$ $n\geq 2$ be such that $\mathrm{St}_{1/n}(p(y))$ is contained in a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $p(y),$ $D$ be the $y$-component of $\mathrm{St}_{1/n}(p(y))$ and $z' \in \mathrm{St}_{1/n}(p(y)) \cap Z'.$ Then $${\mathcal{H}}(p,y):=\#(p^{-1}\{z'\} \cap D)$$ is the *homotopical index* of $y$ in $p.$ When $p \colon Y {\rightarrow}Z,$ $y \in Y$ and $D \subset Y$ are as in definition \[prev\] we get the following upper bound for the homotopical indices of points in $D.$ \[maxindeksi\] Let $y' \in D$ and ${\mathcal{H}}(y',p)$ be the homotopical index of $y'$ in $p.$ Then ${\mathcal{H}}(y',p) \leq {\mathcal{H}}(y,p).$ Let $y' \in D.$ Let $m \geq 2$ be such that $\mathrm{St}_{1/m}(p(y')) \subset \mathrm{St}_{1/n}(p(y))$ and such that $\mathrm{St}_{1/m}(p(y'))$ is contained in a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $p(y').$ Let $C \subset D$ be the $y'$-component of $\mathrm{St}_{1/m}(p(y))$ and $z' \in \mathrm{St}_{1/m}(p(y')) \cap \mathrm{St}_{1/n}(p(y)) \cap Z'.$ Then by Lemma \[whynot\] $${\mathcal{H}}(p,y)=\#(p^{-1}\{z'\} \cap D) \leq \#(p^{-1}\{z'\} \cap C)={\mathcal{H}}(p,y'),$$ since $C \subset D.$ Metrization of monodromy representations {#ccamm} ---------------------------------------- In this section we consider PL manifolds as length manifolds and prove Theorem \[kk\] and Theorem \[hopo\] in the introduction. Let $|K|\subset {\mathbb{R}}^n$ be the polyhedron of a simplical complex $K$ and $$\label{tad}\ell(\gamma):=\sup\left\{\sum_{i=1}^{k-1}|\gamma(t_i)-\gamma(t_{i+1})| : t_1, \ldots, t_k \in [0,1], t_1 < \cdots < t_k, k \in {\mathbb{N}}\right\}$$ the *length* of $\gamma$ for every path $\gamma$ in $|K|.$ A path $\gamma$ in $|K|$ is called *rectifiable* if the length $\ell(\gamma)$ is finite. Since $|K| \subset {\mathbb{R}}^n$ is a polyhedron for $z_1, z_2 \in |K|$ there exists a rectifiable path $\gamma \colon z_1 {\curvearrowright}z_2.$ In particular, the formula $$d_s(z_1,z_2)=\text{inf}_{\gamma} \{\ell(\gamma) \mid \gamma \colon z_1 {\curvearrowright}z_2\}$$ defines a path-metric on $|K|.$ For a detailed study of path length structures see [@G Sec.1]. We note that the path-metric $d_s$ coincides with the metric of ${\mathbb{R}}^n$ when restricted to any simplex $\sigma \in K.$ By local finiteness of $K$ the metric topology induced on $|K|$ by $d_s$ coincides with the relative topology of $|K|$ as a subset of ${\mathbb{R}}^n.$ For a PL manifold $Z,$ there exists an embedding $\iota \colon Z {\hookrightarrow}{\mathbb{R}}^n$ satisfying $\iota(Z)=|K|$ for a simplical complex $K$ and for the next theorem we assume $Z=(|K|,d_s).$ For the next theorem we also assume the following, if $p \colon Y {\rightarrow}Z$ is a completed normal covering and there are fixed large subsets $Y'\subset Y$ and $Z'\subset Z$ so that $p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering, then $K$ has a vertex $z_0 \in Z'.$ \[kumpula\] Let $Z$ be a PL manifold, $p \colon Y {\rightarrow}Z$ a completed normal covering that has locally finite multiplicity. Then there exists a path metric $d_s^*$ on $Y$ so that - the topology induced by $d_s^*$ on $Y$ coincide with the original topology of $Y,$ - $(Y,d_s^*)$ is a locally proper metric space, - $p \colon (Y,d_s^*) {\rightarrow}(Z,d_s)$ is a $1$-Lipschitz map and - every deck-transformation $\tau \in {\mathcal{T}}(p)$ is an isometry with respect to $d_s^*.$ The proof consists of Proposition \[pitkäperjantai\] and Proposition \[lankalauantai\]. We obtain metric $d_s^*$ in Theorem \[kumpula\] by lifting paths. We show an analogy of the result [@R II.3.4] concerning local path-lifting for open discrete maps between manifolds. In Lemmas \[kiire\], \[kiireet\] and \[indeksit\] we assume that $Z$ is a PL manifold and $p \colon Y {\rightarrow}Z$ is a completed normal covering that has locally finite multiplicity. Since $p$ has locally finite multiplicity, it is discrete by [@MA Thm.9.14]. We recall also that $p$ is an orbit map by Theorem \[vauva\], and that there are by Theorem \[TFAE\] large subsets $Y' \subset Y$ and $Z' \subset Z$ so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering and $Z$ is $(Z', \mathrm{Ker}(\sigma_g))$-stable. \[kiire\] Let $z \in Z$ and $y \in p^{-1}\{z\}.$ Let $m \geq 2$ be such that $\mathrm{St}_{1/m}(z)$ is contained in a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $z$ and let $D$ be the $y$-component of $p^{-1}(\mathrm{St}_{1/m}(z)).$ Let $\gamma \colon [0,1] {\rightarrow}\mathrm{St}_{1/m}(z)$ be a path and let $\gamma' \colon [0,1] {\rightarrow}D$ a function satisfying $p \circ \gamma'=\gamma.$ Then $\gamma'$ is continuous at every $t \in [0,1]$ satisfying ${\mathcal{H}}(\gamma'(t);p)={\mathcal{H}}(y;p).$ Let $\mathrm{St}_{1/n}(\gamma(t))$ be the $1/n$-star of $\gamma(t)$ and $C_n$ the $\gamma'(t)$ component of $p^{-1}(\mathrm{St}_{1/n}(\gamma(t)))$ for every $n \geq 2.$ Let $n_0 \geq 2$ be such that $\mathrm{St}_{1/{n_0}}(\gamma(t))$ is contained in $\mathrm{St}_{1/m}(z)$ and in a $(Z', \mathrm{Ker}(\sigma_p))$-stable neighbourhood of $\gamma(t).$ Since ${\mathcal{H}}(\gamma'(t);p)={\mathcal{H}}(y;p),$ $p^{-1}(\mathrm{St}_{1/n}(\gamma(t)))\cap D=C_n$ for every $n\geq n_0$ by Lemma \[whynot\]. Since $\gamma$ is continuous, there exists for every $n\geq n_0$ such $\epsilon_n > 0$ that $\gamma(t-\epsilon_n, t +\epsilon_n)\subset \mathrm{St}_{1/n}(\gamma(t)).$ Now for every $n\geq n_0$ we have $\gamma'(t- \epsilon_n, t +\epsilon_n)\subset C_n.$ Since $(C_n)_{n\geq n_0}$ is a neighbourhood basis at $\gamma'(t)$ the function $\gamma'$ is continuous at $t.$ \[kiireet\] Let $h \colon (0,1) {\rightarrow}Z$ be a continuous map. Suppose there exists for every $t \in (0,1)$ such $\epsilon>0$ that for every $y \in p^{-1}\{h(t)\}$ there exists a continuous map $h' \colon (t-\epsilon, t+ \epsilon) {\rightarrow}Y$ satisfying $h'(t)=y$ and $p \circ h' =h {|}(t-\epsilon, t+ \epsilon).$ Then there exists for every $y \in p^{-1}\{h(1/2)\}$ a continuous map ${\widetilde}{h} \colon (0,1) {\rightarrow}Y$ satisfying ${\widetilde}{h}(1/2)=y$ and $p \circ {\widetilde}{h}=h.$ Let $y \in p^{-1}\{h(1/2)\}.$ By Zorn’s Lemma there exists a maximal connected set $I\subset (0,1), 1/2 \in I,$ for which there exists a continuous map $h' \colon I {\rightarrow}Y$ satisfying $p \circ h'(1/2)=y$ and $p \circ h'=h {|}I.$ By the existence of local lifts $I$ is an open interval $(a,b) \subset [0,1].$ We need to show that $a=0$ and $b=1.$ Fix a map $h' \colon I {\rightarrow}Y$ satisfying $h'(1/2)=y$ and $p \circ h'=h {|}I.$ Suppose $b\neq 1.$ Let $m \geq 2$ be such that $\mathrm{St}_{1/m}(h(b))$ is contained in a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $h(b),$ $\delta \in (0,(b-a)/2)$ be such that $h(b -2\delta ,b) \subset \mathrm{St}_{1/m}(h(b)),$ $D$ be the component of $p^{-1}(\mathrm{St}_{1/m}(h(b)))$ that contains $h'(b -2\delta ,b),$ $y'$ be the point in $D \cap \{h(b)\}$ and $h''\colon (a,b] {\rightarrow}Y$ be the extension of $h'$ defined by $h''(b)=y.$ By Lemma \[kiire\], $h'' {|}[b-\delta,b] \colon [b-\delta,b] {\rightarrow}D$ is continuous. Thus $h''$ is continuous. This is a contradiction with the maximality of $I=(a,b) \subset [0,1].$ Thus $b=1.$ By a similar argument $a=1.$ \[indeksit\] Let $z \in Z$ and $m \geq 2$ such that $\mathrm{St}_{1/m}(z)$ is contained in a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $z.$ Let $\gamma \colon [0,1] {\rightarrow}\mathrm{St}_{1/m}(z)$ be a path satisfying $\gamma(0)=z$ and $y \in p^{-1}\{z\}.$ Then $\gamma$ has a lift ${\widetilde}{\gamma} \colon [0,1] {\rightarrow}Y$ satisfying $\gamma(0)=y.$ Let $k:={\mathcal{H}}(y,p).$ We prove the existence of the lift ${\widetilde}{\gamma}_y$ by induction on $k$. Let $D$ be the $y$-component of $p^{-1}(\mathrm{St}_{1/m}(z)).$ By Lemma \[brunate\], $p(D)=\mathrm{St}_{1/m}(z).$ Suppose $k=1.$ Then $p {|}D \colon D {\rightarrow}\mathrm{St}_{1/m}(z)$ is a homeomorphism. Thus $\gamma$ has a lift ${\widetilde}{\gamma}$ satisfying ${\widetilde}{\gamma}(0)=y.$ Suppose the statement holds for all $j \leq k-1.$ Let $$F:=\{z' \in \mathrm{St}_{1/m}(z) : {\mathcal{H}}(y',p)={\mathcal{H}}(y,p) \text{ for every } y' \in p^{-1}\{z'\}\}$$ and $$U:= \{z' \in \mathrm{St}_{1/m}(z) : {\mathcal{H}}(y',p)<{\mathcal{H}}(y,p) \text{ for every } y' \in p^{-1}\{z'\}\}.$$ By Lemma \[whynot\] and Lemma \[maxindeksi\], $U=\mathrm{St}_{1/m}(z) \setminus F,$ $U \subset \mathrm{St}_{1/m}(z)$ is open and $p^{-1}\{z'\} \cap D$ is a point for every $z' \in F.$ Now $\gamma^{-1}(U) \subset [0,1]$ is a countable union ${\bigcup}_{n \in N}I_n$ of disjoint open intervals $I_n\subset [0,1].$ We use Lemma \[kiireet\] to show that there exists for every $n \in N$ a continuous map ${\widetilde}{\gamma {|}I_n} \colon I_n {\rightarrow}D$ satisfying $p \circ {\widetilde}{\gamma {|}I_n}= \gamma {|}I_n.$ Let $n \in N$ and $t \in I_n.$ Then there exists such $p \geq 2$ that $\mathrm{St}_{1/p}(\gamma(t))$ is contained in a $(Z', \mathrm{Ker}(\sigma_g))$-stable neighbourhood of $\gamma(t)$ and such $\epsilon >0$ that $(t- \epsilon, t+\epsilon) \subset I_n$ and $\gamma(t- \epsilon, t+\epsilon) \subset \mathrm{St}_{1/p}(\gamma(t)).$ Let $h:=I_n {|}(t- \epsilon, t + \epsilon).$ Let $y' \in p^{-1}(\gamma(t)).$ Since $\gamma(t) \in U,$ we have ${\mathcal{H}}(y',p) < k.$ Thus, by the induction hypotheses there are continuous maps $h_1' \colon (t- \epsilon ,t] {\rightarrow}Y$ and $h_2' \colon [t,t + \epsilon) {\rightarrow}Y$ satisfying $h_1'(t)=y'=h_2'(t),$ $p \circ h_1'=h {|}(t- \epsilon,t]$ and $p \circ h_2'=h {|}[t,t+ \epsilon).$ Hence there exists a continuous map $h' \colon (t - \epsilon, t + \epsilon) {\rightarrow}Y$ satisfying $h'(t)=y'$ and $p \circ h'=h.$ Let $t_0 \in I_n$ be the center of $I_n$ and $y'' \in p^{-1}\{\gamma(t_0)\} \cap D.$ By Lemma \[kiireet\], there exists a continuous map ${\widetilde}{\gamma {|}I_n} \colon I_n {\rightarrow}Y$ satisfying $p \circ {\widetilde}{\gamma {|}I_n}= \gamma {|}I_n$ and ${\widetilde}{\gamma {|}I_n}(t_0)=y'' \in D.$ Since $D$ is a component of $\mathrm{St}_{1/m}(z),$ ${\widetilde}{\gamma {|}I_n} \colon I_n {\rightarrow}D.$ Let then $J:=\gamma^{-1}(F).$ Since $p^{-1}\{z'\} \cap D$ is a point for every $z' \in F,$ there exists a unique function ${\widetilde}{\gamma {|}J} \colon J {\rightarrow}D$ satisfying $p \circ {\widetilde}{\gamma {|}J}=\gamma {|}J.$ Let ${\widetilde}{\gamma} \colon [0,1] {\rightarrow}D$ be the unique function satisfying ${\widetilde}{\gamma} {|}I_n={\widetilde}{\gamma {|}I_n}$ for every $n \in N$ and ${\widetilde}{\gamma} {|}J={\widetilde}{\gamma {|}J}.$ Then $p \circ {\widetilde}{\gamma}=\gamma.$ Since $[0,1] \setminus J$ is open, the function ${\widetilde}{\gamma}$ is continuous at every $t \in [0,1] \setminus J.$ By Lemma \[kiire\], ${\widetilde}{\gamma}$ is continuous at every $t \in J.$ Thus ${\widetilde}{\gamma}$ is continuous. Since $\gamma(0)=z \in F,$ ${\widetilde}{\gamma}(0)=y.$ \[LBAPp\]Let $Z$ be a PL manifold and $p \colon Y {\rightarrow}Z$ a completed normal covering that has locally finite multiplicity. Then for every pair of points $y_1$ and $y_2$ in $Y$ there exists a path $\gamma \colon y_1 {\curvearrowright}y_2$ so that $p \circ \gamma$ is rectifiable. Let $Y' \subset Y$ and $Z' \subset Z$ be large subsets so that $g:=p {|}Y' \colon Y' {\rightarrow}Z'$ is a normal covering. By Lemma \[indeksit\], there exists points $z_1, z_2 \in Z'$ and rectifiable paths ${\alpha}_1 \colon p(y_1) {\curvearrowright}z_1$ and ${\alpha}_2 \colon p(y_2) {\curvearrowright}z_2$ having lifts $({\widetilde}{{\alpha}_1})_{y_1}$ and $({\widetilde}{{\alpha}_2})_{y_2}$ by $p.$ We denote $y'_1=({\widetilde}{{\alpha}_i})_{y_1}(1)$ and $y'_2=({\widetilde}{{\alpha}_i})_{y_2}(1)$. Since $g \colon Y' {\rightarrow}Z'$ is a covering between open manifolds, there exists a path ${\beta}\colon y_1' {\curvearrowright}y_2'$ so that $\ell(p \circ {\beta})<\infty.$ Now $\gamma:=({\widetilde}{{\alpha}_1})_{y_1} {\beta}({\widetilde}{{\alpha}_2})_{y_1}^{{\leftarrow}} \colon y_1 {\curvearrowright}y_2$ satisfies $$\ell(p \circ \gamma)\leq\ell({\alpha}_1)+\ell(p \circ {\beta}) + \ell({\alpha}_2^{{\leftarrow}}) < \infty.$$ Let $Z$ be a PL manifold and $p \colon Y {\rightarrow}Z$ a completed normal covering that has locally finite multiplicity and $d_s$ the path-metric of $Z.$ We call $d_s^* \colon Y \times Y {\rightarrow}{\mathbb{R}}_+$ defined by $$d_s^*(y_1,y_2) = \inf \{\ell(p \circ \gamma) \mid p \circ \gamma,\, \gamma \colon y_1 {\curvearrowright}y_2\}$$ the *pullback* of the path-metric $d_s$ by $p.$ \[pitkäperjantai\]Let $Z$ be a PL manifold and $p \colon Y {\rightarrow}Z$ a completed normal covering that has locally finite multiplicity, $d_s$ the path-metric of $Z$ and $d_s^*$ the pullback of $d_s$ by $p.$ Then $d_s^*$ is a metric, every $\tau \in {\mathcal{T}}(p)$ is an isometry with respect to $d_s^*$ and $p \colon (Y,d_s^*) {\rightarrow}(Y,d_s)$ is a $1$-Lipschitz map. The map $p$ is discrete by [@MA Thm.9.14]. Thus $d_s^*$ separates points of $Y$ and $d_s^*$ is a metric by Lemma \[LBAPp\]; see [@G Sec.1]. By the definition of $d_s^*$ every $\tau \in {\mathcal{T}}(p)$ is an isometry with respect to $d_s^*$ and $p \colon (Y,d_s^*) {\rightarrow}(Y,d_s)$ is a $1$-Lipschitz map. In the following proposition we show that the topology induced by $d_s^*$ on $Y$ coincides with the original topology of $Y.$ \[kiirastorstai\]Let $Z$ be a PL manifold and $p \colon Y {\rightarrow}Z$ a completed normal covering that has locally finite multiplicity. Let $d_s$ be the path-metric on $Z$ and ${\mathcal{T}}$ the topology of $Y$. Let $d_s^*$ be the pullback of the path-metric $d_s$ of $Z$ by $p$ and ${\mathcal{T}}_{d_s^*}$ the topology induced on $Y$ by the metric $d_s^*.$ Then ${\mathcal{T}}={\mathcal{T}}_{d_s^*}.$ We first show that ${\mathcal{T}}\subset {\mathcal{T}}_{d_s^*}.$ Since $p$ is a spread, it is sufficient to show that for every open connected subset $V \subset Z$ and component $U$ of $p^{-1}(V)$ we have $U \in {\mathcal{T}}_{d_s^*}.$ Let $V \subset Z$ be an open and connected set, and let $U \subset Y$ be a component of $p^{-1}(V).$ Fix $y \in U.$ Since $p \colon (Y,d_s^*) {\rightarrow}(Y,d_s)$ is a $1$-Lipschitz map, there exists $r_y \in (0,1)$ so that $p(B(y,r_y))\subset B(y,r_y) \subset V.$ Now for every point $y' \in B(y,r_y)$ there exists by the definition of $d_s^*$ a path $\gamma \colon y {\curvearrowright}y'$ in $p^{-1}(B(y,r_y))\subset p^{-1}(V).$ Thus $B(y,r_y)\subset U,$ since $U$ is the $y$-component of $p^{-1}(V).$ We conclude $U \in {\mathcal{T}}_{d_s^*}.$ We then show that ${\mathcal{T}}_{d_s^*} \subset {\mathcal{T}}.$ Let $U \in {\mathcal{T}}_{d_s^*}.$ The map $p$ is discrete by [@MA Thm.9.14], uniformly discrete Theorem \[TFAE\] and an orbit map by Theorem \[vauva\]. Thus, as a consequence of Lemma \[indeksit\], there exists for every $y \in U$ a radius $r_y \in (0,1)$ that satisfies the following conditions: - for the $y$-component $U_y$ of $p^{-1}(B(p(y),r_y)),$ $U_y \cap p^{-1}\{p(y)\}=\{y\},$ - every path in $B(p(y),r_y)$ beginning at $p(y)$ has a total lift into $Y$ beginning at $y,$ - $[z,p(y)] \subset B(p(y),r_y)$ for every $z \in B(p(y),r_y)$ and - $B(y,r_y) \subset U.$ Since $p$ is a spread, $U_y \in {\mathcal{T}}$ for every $y \in U.$ It is suffices to show that $U_y \subset B(y,r_y)$ for every $y \in U,$ since then $$U={\bigcup}_{y \in U}B(y,r_y)={\bigcup}_{y \in U}U_y \in {\mathcal{T}}.$$ Let $y_1 \in U_y.$ Then $p(y_1) \in B(p(y),r_y)$ and there exists a path $\gamma \colon p(y) {\curvearrowright}p(y_1)$ in $B(p(y),r_y)$ satisfying $\ell(\gamma)<r_y.$ Let ${\widetilde}{\gamma}_y$ be a lift of $\gamma$ in $U_y$ beginning at $y.$ Then $$d_s^*(y, {\widetilde}{\gamma}_y(1)) \leq \ell(p \circ {\widetilde}{\gamma}_y)=\ell(\gamma)<r_y.$$ Hence ${\widetilde}{\gamma}_y(1) \in B(y,r_y).$ Since $p$ is an orbit map there is a deck-transformation $\tau \in {\mathcal{T}}(p)$ satisfying $\tau ({\widetilde}{\gamma}(1))=y_1.$ Since $y_1$ and ${\widetilde}{\gamma}_y(1)$ belong to $U_y,$ we have $\tau(U_y)=U_y$ by Lemma \[brunate\]. Hence $\tau(y)=y,$ since $U_y \cap p^{-1}\{p(y)\}=\{y\}.$ Thus $$d_s^*(y_1,y)=d_s^*(\tau({\widetilde}{\gamma}_y(1)), \tau(y))=d_s^*({\widetilde}{\gamma}_y(1),y)<r_y,$$ since $\tau$ is an isometry with respect to $d_s^*.$ Thus $y_1 \in B(y,r_y).$ Thus $U_y \subset B(y,r_y)$ and we conclude that ${\mathcal{T}}_{d_s^*} \subset {\mathcal{T}}.$ In the following proposition we show that the topology induced by the pullback metric is locally proper. \[lankalauantai\]Let $p \colon Y {\rightarrow}Z$ be a completed normal covering onto a PL manifold $Z$ that has locally finite multiplicity. Let $d_s^*$ be the pullback of the path-metric $d_s$ of $Z$ by $p.$ Then $Y$ is a locally proper metric space with respect to $d_s^*$. Fix $y \in Y.$ Since $p$ has locally finite multiplicity, Lemma \[indeksit\] implies that there exists $r \in (0,1)$ that satisfies $p(\overline{B(y,r)})=\overline{B(p(y),r)}$ and satisfies for the $y$-component $D$ of $p^{-1}(B(p(y),2r))$ the following conditions: - $p^{-1}\{p(y)\} \cap D=\{y\}$ - $p{|}D$ has finite multiplicity and - $\overline{B(y,r)} \subset D.$ We prove the claim by showing that $\overline{B(y,r)}$ is compact. Let $\mathcal{U}$ be an open cover of $\overline{B(y,r)}.$ Since $p$ is an open map and the set $p^{-1}\{z\} \cap \overline{B(y,r)}$ is finite, there exists for every $z \in p(\overline{B(y,r)})$ a radius $r_z \in (0,1)$ so that for every $y' \in p^{-1}\{z\} \cap \overline{B(y,r)}$ there exists $U \in \mathcal{U}$ satisfying $B(y',r_z)\subset U$ and $p(B(y',r_z))=B(z,r_z).$ Since $p(\overline{B(y,r)})=\overline{B(p(y),r)} \subset Z$ is compact, we may fix $z_1, \ldots, z_k \in p(\overline{B(y,r)})$ so that $\{B(z_i,r_{z_i}) : 1\leq i \leq k\}$ is an open cover of $p(\overline{B(y,r)}).$ The set $\{B(y',r_{z_i}) \mid y' \in p^{-1}\{z_i\} \cap \overline{B(y,r)}, i \in \{1, \ldots, k\}\}$ is now finite and for every $i \in \{1, \ldots, k\}$ and $y' \in p^{-1}\{z_i\} \cap \overline{B(y,r)}$ there exists $U \in \mathcal{U}$ so that $B(y',r_{z_i})\subset U.$ Thus it suffices to show that $$\overline{B(y,r)} \subset V:={\bigcup}\{B(y',r_{z_i}) : y' \in p^{-1}\{z_i\} \cap \overline{B(y,r)}, i \in \{1, \ldots, k\}\}.$$ Let $y' \in \overline{B(y,r)}.$ Fix $i \in \{1, \ldots, k\}$ satisfying $p(y') \subset B(z_i,r_{z_i})$ and $e \in p^{-1}\{z_i\} \cap \overline{B(y,r)}.$ Then there exists a point $y'' \in p^{-1}\{p(y')\} \cap B(e,r_{z_i}),$ since $p(B(e,r_{z_i}))=B(z_i,r_{z_i}).$ Let $\tau \colon (Y,d_s^*) {\rightarrow}(Y, d_{s}^*)$ be a deck-transformation isometry satisfying $\tau(y'')=y'.$ Then $y' \in B(\tau(e),r_{z_i})$ and $\tau(e) \in p^{-1}\{z_i\}.$ Since $y''$ and $y'$ belong to $D,$ $\tau(D)=D$ by Lemma \[vindpark\]. Thus $\tau(y)=y,$ since $p^{-1}\{p(y)\} \cap D=\{y\}.$ Hence $$d_s^*(\tau(e),y)=d_s^*(\tau(e),\tau(y))=d_s^*(e,y) \leq r.$$ Thus $\tau(e) \in p^{-1}\{z_i\} \cap \overline{B(y,r)}$ and $y' \in B(\tau(e),r_{z_i}) \subset V.$ We conclude that ${\mathcal{U}}$ has a finite subcover, $\overline{B(y,r)}\subset Y$ is compact and $(Y,d_s^*)$ is a locally proper metric space. This concludes the proof of Theorem \[kumpula\] and we are ready for the proofs of Theorems \[kuuskytkolme\], \[kk\] and \[hopo\] in the introduction. We say that a path-metric $d'_s$ on $Z$ is a *polyhedral path-metric* on $Z,$ if there exists such a simplicial complex $K,$ polyhedron $|K|\subset {\mathbb{R}}^n$ and embedding $\iota \colon Z {\rightarrow}{\mathbb{R}}^n$ satisfying $\iota(Z)=|K|,$ that $d'_s$ is the pullback of $d_s$ by $\iota$ for the path-metric $d_s$ of $|K|.$ We recall that a completed normal covering $p \colon Y {\rightarrow}Z$ is an open map and the domain $Y$ is by definition a Hausdorff space. In particular, $p$ has thus locally finite multiplicity, if $p$ is discrete and $Y$ is locally compact. Let $Z$ be a PL manifold and $p \colon Y {\rightarrow}Z$ be a discrete completed normal covering. If $Y$ is locally compact, then $p$ has locally finite multiplicity. Suppose that $p$ has locally finite multiplicity. By Theorem \[kumpula\], there is a metric $d_s^*$ on $Y$ so that the topology induced by $d_s^*$ coincides with the original topology of $Y$ and $(Y,d_s^*)$ is a locally proper metric space. Thus $Y$ is locally compact. Let $f \colon X {\rightarrow}Z$ be a completed covering between PL manifolds, $(Y,p,q)$ a locally compact monodromy representation of $f,$ $d_s$ a polyhedral path-metric of $Z.$ By Theorem \[vauva\], the orbit map $q$ is discrete. Thus $q$ has locally finite multiplicity, since $Y$ is locally compact. Since $q$ has locally finite multiplicity, by Theorem \[kumpula\], the pullback $d_s^*$ of $d_s$ is a path-metric on $Y$ satisfying conditions (a)–(d) in Theorem \[vauva\]. For the statement we need to show that a completed covering $f \colon X {\rightarrow}Z$ between PL manifolds has a locally compact monodromy representation $(Y,p,q)$ if and only if $f$ is stabily completed and $f$ has a finite local monodromy group at each $z \in Z.$ Suppose first that $f$ is stabily completed covering and $f$ has a finite local monodromy group at each $z \in Z.$ By Corollary \[matka\] $f$ has a monodromy representation $(Y,p,q),$ where $q$ has locally finite multiplicity. By Theorem \[kumpula\], there is a metric $d_s^*$ on $Y$ so that the topology induced by $d_s^*$ coincides with the original topology of $Y$ and $(Y,d_s^*)$ is a locally proper metric space. Thus $Y$ is locally compact and $(Y,p,q)$ a locally compact monodromy representation of $f.$ Suppose then that $f$ has a locally compact monodromy representation $(Y,p,q).$ Then the map $q$ is discrete by Theorem \[vauva\]. Thus $q$ has locally finite multiplicity. By Theorem \[TFAE\], $q$ is a stabily completed normal covering. Thus, by Theorem \[torstai\], $q$ has a finite local monodromy group at each $z \in Z,$ since $q$ has locally finite multiplicity. Fix $z \in Z$ and let $H$ be a finite local monodromy group of $q$ at $z.$ By Theorem \[spread\_3\] and Remark \[stabhuomio\] there is a quotient of $H$ that is a local monodromy group of $f$ at $z.$ Since every quotient of a finite group is finite, $f$ has a finite local monodromy group at $z.$

ArXiv

--- abstract: 'A notion of [*intermediate vertex subalgebras*]{} of lattice vertex operator algebras is introduced, as a generalization of the notion of [*principal subspaces*]{}. Bases and the graded dimensions of such subalgebras are given. As an application, it is shown that the characters of some modules of an intermediate vertex subalgebra between $E_7$ and $E_8$ lattice vertex operator algebras satisfy some modular differential equations. This result is an analogue of the result concerning the “hole" of the [*Deligne dimension formulas*]{} and the [*intermediate Lie algebra*]{} between the simple Lie algebras $E_7$ and $E_8$.' address: 'Department of Mathematical Sciences, University of Tokyo, Komaba, Tokyo, 153-8914, Japan.' author: - Kazuya KAWASETSU title: The intermediate vertex subalgebras of the lattice vertex operator algebras --- Introduction ============ The [*Deligne exceptional series of simple Lie algebras*]{} is the series $$A_1\subset A_2 \subset G_2 \subset D_4 \subset F_4 \subset E_6 \subset E_7 \subset E_8$$ of simple Lie algebras [@D]. For irreducible components of some tensor products of the adjoint representations of the simple Lie algebras in the above exceptional series, remarkable dimension formulas, called [*Deligne dimension formulas*]{}, were established [@CdM; @D; @LaM2]. They are expressed as rational functions in the dual Coxeter number $h^\vee$. For example, $$\dim \mathfrak{g} = \frac{2(5h^\vee -6)(h^\vee +1)}{h^\vee +6}$$ and $$\dim \mathfrak{g}^{(2)} = \frac{5h^{\vee 2}(2h^\vee+3)(5h^\vee-6)}{(h^\vee+6)(h^\vee+12)}.$$ When $h^\vee=24$, which intermediates between the dual Coxeter numbers $18$ of $E_7$ and $30$ of $E_8$, the formulas give integer values $\dim\mathfrak{g}=190$ and $\dim\mathfrak{g}^{(2)}=15504$. However there is no such a simple Lie algebra. Later, this “hole" of the exceptional series was filled in. In [@LaM1], a Lie algebra $E_{7+1/2}$, which is non-reductive and intermediates between $E_7$ and $E_8$, was constructed, and the dimension formulas for this algebra were proved. The Lie algebra $E_{7+1/2}$ is an [*intermediate Lie algebra*]{} [@LaM1; @S1; @GZ1]. The same exceptional series appeared in earlier studies of modular differential equations. In 1988, Mathur, Mukhi and Sen, in their work of classification of rational conformal field theories ($C_2$-cofinite rational vertex operator algebras (VOAs) of CFT-type) with two characters [@MMS], studied the modular differential equations of the form $$\label{eq:diff1} \left( q\frac{d}{dq}\right)^2 f(\tau) + 2 E_2(\tau)\left(q\frac{d}{dq}\right)f(\tau) + 180\mu\cdot E_4(\tau)f(\tau)=0.$$ Here $\mu$ is a numerical constant, $\tau$ a complex number in the complex upper half-plane $\mathbb{H}$ with $q=e^{2\pi i \tau}$, and $E_k(\tau) (k=2,4,6,\ldots)$ the Eisenstein series $$E_k(\tau)=-\frac{B_k}{k!}+\frac{2}{(k-1)!}\sum_{n\geq 1}\frac{n^{k-1}q^n}{1-q^n},$$ where $B_k$ is the $k$-th Bernoulli number. (Differential equations equivalent to (\[eq:diff1\]) were studied by Kaneko and Zagier [@Kan3] in number theory.) By studying (\[eq:diff1\]), they showed, roughly speaking, that the characters of the rational conformal field theories with two characters are that of the level one affine VOAs associated to the Deligne exceptional simple Lie algebras. The list obtained is shown in Table \[tb:table1\]. ($c$ denotes the central charge, $h$ the non-zero conformal weight, and $\dim V_1$ the dimension of the weight one subspace of such a theory.) When $\mu=11/900$ and $551/900$, there are solutions of the differential equations (\[eq:diff1\]) of the form $ f(\tau)= q^{-c/24} \sum_{n=0}^\infty a_n q^n $ with $a_0=1$, $a_n\in \mathbb{Z}_{> 0}$ ($n=1,2,3,\ldots$) and $c=2/5$ and $38/5$. However, according to the Verlinde formula, there are no rational conformal field theories of central charges $c=2/5$ and $38/5$ with two characters. (In [@MMS], the Virasoro minimal model at $c=-22/5$ was assigned to the case $\mu=11/900$. The characters agree with the famous [*Rogers-Ramanujan functions*]{}.) Note that the value $c=38/5$ intermediates between the central charges $c=7$ of $L_{1,0}(E_7)$ and $c=8$ of $L_{1,0}(E_8)$. Here, $L_{1,0}(\mathfrak{g})$ is the level one affine VOA associated to a simple Lie algebra $\mathfrak{g}$. Note also that actually $a_1$ is $190$ and agrees with $\dim E_{7+1/2}=190$. \[tb:table1\] $\mu$ $\dim V_1$ $c$ $h$ Identification ----------- ------------ -------- --------- ---------------- $11/900$ $1$ $2/5$ $1/5$ $5/144$ $ 3 $ $ 1 $ $1/4$ $L_{1,0}(A_1)$ $1/12$ $ 8 $ $2$ $1/3$ $L_{1,0}(A_2)$ $119/900$ $14$ $14/5$ $2/5$ $L_{1,0}(G_2)$ $2/9$ $28$ $4$ $1/2$ $L_{1,0}(D_4)$ $299/900$ $52$ $26/5$ $3/5$ $L_{1,0}(F_4)$ $5/12$ $78$ $6$ $2/3$ $L_{1,0}(E_6)$ $77/144$ $133$ $7$ $3/4$ $L_{1,0}(E_7)$ $551/900$ $190$ $38/5$ $4/5$ $2/3$ $248$ $8$ $(5/6)$ $L_{1,0}(E_8)$ : The characters of the RCFTs with “two" characters and $h\geq 0$ Motivated by the above works, we construct an $\mathbb{N}$-graded vertex algebra and a module with the “characters" satisfying the equation (\[eq:diff1\]) with $\mu=551/900$, by considering the analogy of the intermediate Lie algebras. We consider the vertex subalgebra $$\label{eq:e712affine} \langle E_{7+1/2} \rangle_{{\rm v.a.}} \subset L_{1,0}(E_8)$$ (and a module). Here, $\langle E_{7+1/2}\rangle_{{\rm v.a.}}$ denotes the smallest vertex subalgebra containing $E_{7+1/2}$. The vertex algebra $\langle E_{7+1/2} \rangle_{{\rm v.a.}}$ is isomorphic to an [*intermediate vertex subalgebra*]{} of a lattice VOA, which we introduce in this paper. For the purpose, we establish the formula to describe the graded dimensions of such subalgebras and modules. The intermediate vertex subalgebra $W(R,S)$ is by definition the vertex subalgebra generated by the subset $\{e^\rho, e^{\pm \sigma}, \sigma_{-1}\otimes 1 | \rho\in R,\sigma\in S\}$ of the lattice VOA $V_L$ associated with an integral lattice $L$. Here, $R$ and $S$ are disjoint subsets of a $\mathbb{Z}$-basis $B$ of $L$. (See DEFINITION \[sec:def1\] and \[sec:def2\], for more detail.) The notion of the intermediate vertex subalgebras is a generalization of the notion of the [*principal subspaces*]{} introduced by Feigin and Stoyanovsky [@FS; @SF]. A principal subspace is the subspace $$W(\Lambda)=U(\bar{\mathfrak{n}})\cdot v_\Lambda$$ of a standard $A_n^{(1)}$-module $L(\Lambda)$, where $\mathfrak{n}$ is the nilradical of a Borel subalgebra of $sl_{n+1}$. When $n=1$, the graded dimensions of $W(\Lambda_0)$ and $W(\Lambda_1)$ agree with the Rogers-Ramanujan functions. The notion clearly extends to an arbitrary highest weight module for an affine Lie algebra. The principal subspaces were studied in [@CLM2; @CalLM3; @G; @AKS; @CoLM; @P1; @FFJMM] and others. Recently, Milas and Penn considered the lattice VOA $V_L$ and the vertex subalgebra $W_L(B)=\langle e^{\beta_1},\ldots,e^{\beta_n} \rangle _{{\rm v.a.}}$, called the principal subalgebra [@MP]. This is a generalization of the principal subspaces of level one standard modules over simply-laced simple Lie algebras. Combinatorial bases and the graded dimensions of the subalgebra $W_L(B)$ and some modules were given in [@MP]. Note that the principal subalgebra $W_L(B)$ agrees with the intermediate vertex subalgebra $W(B,\emptyset)$. By using the bases of the principal subalgebras and modules, we construct combinatorial bases and give the formula to compute the graded dimensions of the intermediate vertex subalgebra $W(R,S)$ and [*intermediate modules*]{}. We use the formula to study the character of the vertex subalgebra (\[eq:e712affine\]). We consider the lattice VOA $V_{E_8}$ associated with the $E_8$ root lattice and the intermediate vertex subalgebra $V_{E_{7+1/2}}=W(\{\alpha_1\},\{\alpha_2,\ldots,\alpha_8\})$. This vertex algebra is isomorphic to the vertex algebra (\[eq:e712affine\]) via the isomorphism $V_{E_8}\cong L_{1,0}(E_8)$. Next, we consider the intermediate module $V_{E_{7+1/2}+\alpha_1}$. Then we will show that the characters of the subalgebra and module form a basis of the solutions of the modular differential equation (\[eq:diff1\]) with $\mu=551/900$. Note that by means of Tuite’s result in [@T2], which was motivated by the work of Matsuo [@Mat], our result can be thought of as a conformal field theory version of filling in the “hole" of the exceptional series. In section 2, we recall the definition of the lattice VOAs and introduce the notion of the intermediate vertex subalgebras and intermediate modules. We then describe the structures and give the formula to compute the graded dimensions. In section 3, we consider the intermediate vertex subalgebra $V_{E_{7+1/2}}$ and the module $V_{E_{7+1/2}+\alpha_1}$. We compute the characters and show that they form a basis of the solutions of (\[eq:diff1\]) with $\mu=551/900$. For the purpose, we decompose the characters into the form of polynomials in the characters of the modules of some VOAs. In section 4, we prove the structure theorem of the intermediate vertex subalgebras and modules. The intermediate vertex subalgebras =================================== The setting {#sec:setting} ----------- Let $L$ be a rank $n$ non-degenerate integral lattice with the $\mathbb{Z}$-bilinear form $\langle\cdot , \cdot \rangle: L\times L\rightarrow \mathbb{Z}$. Let $B=\{ \beta_1,\ldots,\beta_n\}$ be a $\mathbb{Z}$-basis of $L$. Put $\mathfrak{h}=\mathbb{C}\otimes_\mathbb{Z} L$. Consider the affine central extension $\widehat{\mathfrak{h}}=\mathfrak{h}\otimes \mathbb{C}[t,t^{-1}]\oplus \mathbb{C}{\bf k}$ and its irreducible induced module $$M(1)=U(\widehat{\mathfrak{h}}) \otimes_{U(\mathfrak{h} \otimes \mathbb{C}[t]\oplus \mathbb{C}{\bf k})} \mathbb{C},$$ where $\mathfrak{h}\otimes \mathbb{C}[t] $ acts trivially and ${\bf k}$ acts as $1$ on the one-dimensional module $\mathbb{C}$. The space $M(1)$ can be identified with the symmetric algebra ${\rm S}(\widehat{\mathfrak{h}}_-)$, where $$\widehat{\mathfrak{h}}_-=\mathfrak{h}\otimes t^{-1}\mathbb{C}[t^{-1}].$$ Consider the corresponding lattice vertex operator (super)algebra (lattice VOA) $$V_L\cong M(1) \otimes \mathbb{C}[L]$$ [@B]. Recall that the vertex operator is given by the following formula: $$Y(e^\alpha,x)=\sum_{m\in \mathbb{Z} } (e^\alpha)_m x^{-m-1} = E^- (-\alpha,x) E^+ (-\alpha,x)e_\alpha x^\alpha.$$ Here $$e_\alpha \cdot (h\otimes e^\beta )=\epsilon (\alpha , \beta )h\otimes e^{\alpha+\beta}, \ \ \ e^{\beta} \in \mathbb{C} [L], h\in M(1),$$ $$E^-(-\alpha,x)=\exp \left(-\sum_{j<0} \frac{x^{-j} } {j} \alpha_j \right),$$ and $$E^+(-\alpha,x)=\exp \left(-\sum_{j>0} \frac{x^{-j} } {j} \alpha_j \right).$$ For $B'\subset B$, we set $$L(B')=\bigoplus_{\beta \in B'} \mathbb{Z}\beta \ \ \ \ \mbox{and} \ \ \ \ L_+(B')=\bigoplus_{\beta \in B'} \mathbb{Z}_{\geq 0} \beta.$$ Then $L(B')$ is a sublattice, and $L_+(B')$ is a submonoid of $L$. Furthermore, we set $\mathfrak{h}(B')$, $\widehat{\mathfrak{h}}(B')$ and $\widehat{\mathfrak{h}}(B')_-$ and consider the lattice VOA $V_{L(B')}$ as a vertex operator subalgebra of $V_L$. Let $R$ and $S$ be disjoint subsets of $B$. Let $R\sqcup S$ denote the disjoint union. Put $r=|R|$ and $s=|S|$. We arrange the indices of the basis so that $ R=\{\beta_1,\ldots,\beta_r \}$ and $ S=\{ \beta_{r+1},\ldots, \beta_{r+s}\}. $ We put $\rho_i=\beta_i$ for $1\leq i \leq r$ and $\sigma_j=\beta_{r+j}$ for $1\leq j \leq s$. That is, $$R=\{ \rho_1,\ldots,\rho_r \} \ \ \ \mbox{and} \ \ \ S= \{ \sigma_1,\ldots,\sigma_s \}.$$ Then we set $$L(R,S)= L_+(R) \oplus L(S).$$ This is a submonoid of $L$. Let $\langle A \rangle_{{\rm v.a.}}$ denote the smallest vertex subalgebra containing the subset $A$ of $V_L$. \[sec:def1\] The (weak) [*intermediate vertex subalgebra*]{} $W(R,S)$ of $V_L$ associated with $(R,S)$ is the vertex subalgebra $$W(R,S) = \langle e^{\rho}, e^{\pm \sigma}, \sigma_{-1}\otimes 1 | \rho\in R, \sigma \in S \rangle_{{\rm v.a.}}.$$ We set $W(R)=W(R,\emptyset)$. Let $L^\circ$ denote the dual lattice of $L$. Consider a $V_L$-module $V_{L^\circ}=M(1)\otimes \mathbb{C}[L^\circ]$. Let $\lambda$ be an element of $L^\circ$. \[sec:def2\] The (weak) [*intermediate module*]{} $W(R,S;\lambda)$ over $W(R,S)$ is the cyclic $W(R,S)$-module $$W(R,S;\lambda) = W(R,S) \cdot e^\lambda \subset V_{L^\circ}.$$ We set $W(R;\lambda)=W(R,\emptyset;\lambda)$. Put $L'=L(R)$ and $L''=L(S)$. The intermediate vertex subalgebra $W(\emptyset,S)$ agrees with the lattice VOA $V_{L''}$ associated with the lattice $L''$. On the other hand, the intermediate vertex subalgebra $W(R)=W(R,\emptyset)$ is the principal subalgebra $W_{L'}(R)$, and the intermediate modules $W(R;\lambda)=W(R,\emptyset;\lambda)$ agree with the principal subspaces $W_{L'+\lambda}(R)$ [@MP]. Moreover, if $L'$ is an ADE root lattice, $W(R)$ and $W(R;\lambda_i)$ correspond to the level one principal subspaces $W (\Lambda_0)$ and $W(\lambda_i)$, studied in [@CalLM3]. Here, $\lambda_i$ $(i=1,\ldots,n)$ are the fundamental weights of the root system of $L'$. Now let us define some (bi-)gradings on the above vertex algebras and modules. Put $V=V_{L^\circ}$. First, take the conformal vector $\omega$ in $V_L$ and the stress-energy tensor $$Y(\omega,z)=T(z)=\sum_{i\in \mathbb{Z}}L(i) z^{-i-2}.$$ Then, the operator $L(0)$ on $V$ is diagonalizable, and the eigenvalues are rational numbers. Let $V_r$ denote the $r$-eigenspace. The grading $V=\bigoplus_r V_r$ is called the [*weight grading*]{}. For $v\in V_r$, we call $r$ the [*weight*]{} of $v$ and write $r={\rm wt}(v)$. For any vector subspace $X$ of $V$, we set $X_r=X\cap V_r$. Then the vector spaces $V_L$, $W(R,S)$ and $W(R,S;\lambda)$ are graded vector subspaces of $V$. We call the restricted grading the weight grading. Next, take an element $\tau$ of $L^\circ$. Consider the subspace $$V^\tau=M(1)\otimes e^\tau.$$ Then consider the subspaces $$\begin{aligned} X^\tau = X\cap V^\tau, \ \ \ \ (X=V_L, W(R,S)),\end{aligned}$$ and $$(W(R,S;\lambda))^\tau=W(R,S;\lambda)\cap V^{\tau+\lambda}.$$ Let $X$ be one of the vector spaces $V, V_L, W(R,S)$ and $W(R,S;\lambda)$. The grading $X=\bigoplus_\tau X^\tau$ is called the [*charge-grading*]{}. For any $v\in X^\tau$, we call $\tau$ the [*charge*]{} of $v$. If $X^\tau\neq 0$, we call $\tau$ a [*charge*]{} of $X$. Note that the set of the charges of $V_L$ agrees with $L$, and that of $W(R,S)$ and $W(R,S;\lambda)$ agree with $L(R,S)$. Note that for any charge $\tau$, the subspace $X^\tau$ is $L(0)$-invariant, and the weight grading and charge grading are compatible. Consider the bi-grading $X=\bigoplus_{r,\tau} X^\tau_r$, where $X^\tau_r$ denote the subspace $X^\tau \cap X_r$. We call this the [*charge and weight grading*]{}. Note that as the bi-graded vector spaces, $W(R,S;0)=W(R,S)$. Structure of the intermediate vertex subalgebras and modules ------------------------------------------------------------ Recall that $W(R)=W(R,\emptyset)$ is a principal subalgebra [@MP]. The following theorem is our main result. \[sec:str\] The module $W(R,S;\lambda)$ decomposes as an ${{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}\otimes W(R)$-module into the form $$\label{eqn:streqn1} W(R,S;\lambda) \cong {{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}\otimes \bigoplus_{\delta \in L(S)} W(R;\delta+\lambda).$$ We prove the theorem in §4. Now, we construct a basis of $W(R,S;\lambda)$ using the theorem and the result of [@MP]. Write $r=|R|$ and $s=|S|$. Put $\rho_i=\beta_i$ for $1\leq i \leq r$, and $\sigma_j=\beta_{r+j}$ for $1\leq j \leq s$. Let $k_1,\ldots,k_r$ be non-negative integers. For $1\leq i \leq r$, consider the set $$\begin{aligned} M_i(R;\lambda;k_1,\ldots,k_r) &=& \{ \left(m_{k_i},\ldots,m_1\right) \in \mathbb{Z}^{k_i} | \\ && \ \ \ \ \ \ m_1 \leq -1 - \sum_{l=1}^{i-1} k_l \langle \rho_i, \rho_l \rangle - \langle \rho_i, \lambda \rangle, \\ && \ \ \ \ \ \ m_{j+1}\leq m_j - \langle \rho_i, \rho_i \rangle \ (1\leq j \leq k_i-1) \}.\end{aligned}$$ For each sequence $\mu=(m_k,\ldots,m_1)$ of integers and $\beta\in B$, set $\varepsilon^\beta_\mu=(e^\beta)_{m_k}\ldots(e^\beta)_{m_1}$. Consider the set $$\begin{aligned} {{\mathcal}B}(R;\lambda;k_1,\ldots,k_r) &=& \{ \varepsilon^{\rho_r}_{\mu_r}\ldots \varepsilon^{\rho_1}_{\mu_1}. e^{\lambda} | \mu_i\in M_i(R;\lambda;k_1,\ldots,k_r) \ (1\leq i \leq r) \}.\end{aligned}$$ Note that the elements of the set ${{\mathcal}B}(R;\lambda;k_1,\ldots,k_r)$ have the charge $k_1\rho_1+\cdots+k_r\rho_r$. The following lemma is the result of [@MP]. ([@MP], Corollary 4.8.) \[sec:lemmp\] If $k_1,\ldots,k_r$ are non-negative integers, then the set ${{\mathcal}B}(R;\lambda;k_1,\ldots,k_r)$ is a $\mathbb{C}$-basis of the vector space $(W(R;\lambda))^\rho$, where $\rho={k_1\beta_1+\cdots+k_r\beta_r}$. Note that for each $\tau\in L(R,S)$, the subspace $(W(R,S;\lambda))^\tau$ is a free ${{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}$-module. By THEOREM \[sec:str\] and LEMMA \[sec:lemmp\], we obtain the following corollary. \[sec:cor1\] If $\delta$ is an element of $L(S)$, and $k_1,\ldots,k_r$ are non-negative integers, then the set ${{\mathcal}B}(R;\lambda+\delta;k_1,\ldots,k_r)$ is a basis of the free ${{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}$-module $(W(R,S;\lambda))^\tau$, where $\tau={k_1\rho_1+\cdots+k_r\rho_r+\delta}$. Consider the set $$\begin{aligned} {{\mathcal}B}({{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}) &=& \{ (\sigma_1)_{i_1^1}\ldots(\sigma_1)_{i_1^{l_1}}\ldots(\sigma_s)_{i_s^1}\ldots(\sigma_s)_{i_s^{l_s}} | \\ && \ \ \ \ \ l_1,\ldots,l_s \geq 0, i_j^1 \leq \cdots \leq i_j^{l_j} \leq -1 \ (1\leq j \leq s)\}.\end{aligned}$$ Then ${{\mathcal}B}({{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )})$ is a basis of ${{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}$. Consider the set $${{\mathcal}B}(R,S;\lambda)={{\mathcal}B}({{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )})\otimes \coprod_{\delta\in L(S), k_1, \ldots, k_r\geq 0} {{\mathcal}B}(R;\lambda+\delta;k_1,\ldots,k_r).$$ Here, for vector spaces $P$ and $Q$ and subsets $X\subset P$ and $Y\subset Q$, we denote by $X\otimes Y$ the set $ \{ x\otimes y \in P \otimes Q | x\in X, y\in Y\}. $ Then we obtain the following corollary. \[sec:cor11\] The following hold. (i) : If $\delta$ is an element of $L(S)$, and $k_1,\ldots,k_r$ are non-negative integers, then the set ${{\mathcal}B}({{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )})\otimes {{\mathcal}B}(R;\lambda+\delta;k_1,\ldots,k_r)$ is a $\mathbb{C}$-basis of the vector space $W(R,S;\lambda)^\tau$, where $\tau={k_1\rho_1+\cdots+k_r\rho_r+\delta}$. (ii) : The set ${{\mathcal}B}(R,S;\lambda)$ is a $\mathbb{C}$-basis of the vector space $W(R,S;\lambda)$. This is a generalization of the result of [@MP]. Graded dimensions of the intermediate vertex subalgebras and modules -------------------------------------------------------------------- Recall that $V_{L^\circ}$ and our subspaces carry the bi-gradings: the charge and weight gradings. The [*graded dimension*]{} $\chi _{W(R,S;\lambda) }$ of $W(R,S;\lambda)$ is $$\chi _{W(R,S;\lambda) } ({\bf x};q) = \sum_{\tau \in L^\circ, r\in \mathbb{Q}} \dim_{\mathbb{C}} \left( \left(W(R,S;\lambda) \right)^\tau_r \right) q^r x_1^{k_1} \cdots x_n^{k_n},$$ where $k_1\beta_1 +\cdots +k_n \beta_n=\tau$. To compute the graded dimension, consider the symbols $$(q)_k=(q; q)_k=(1-q)\cdots(1-q^k), \ \ \ (k\geq 1),$$ $(q)_0=1$, and $(q)_{\infty}=\prod_{i=1}^\infty (1-q^i)$. Here $ (a;q)_k=(1-a)(1-aq)\cdots (1-aq^{k-1}) $ is the $q$-Pochhammer symbol. Recall that $1/(q)_k$ agrees with the generating function of the partitions into parts not greater than $k$, therefore agrees with the generating function of the partitions into at most $k$ parts. Recall further that $$\frac{1}{(q)_{\infty}} = \frac{1}{\varphi(q)} = \sum_{k=0} ^{\infty} p(k) q^k.$$ Here, $\varphi(q)$ is the Euler function and $p(k)$ is the number of the un-restricted partitions of an integer $k$. Consider the Gram matrix $A=(\langle \beta_i,\beta_j \rangle)_{i,j}$. Put $r=|R|$ and $s=|S|$. Since $\lambda\in L^\circ$, $\lambda$ has the form $l_1\beta_1+\cdots +l_n\beta_n$ with $l_1,\ldots, l_n \in \mathbb{Q}$. Put ${\bf l}=(l_1,\ldots,l_n)$. \[sec:char\] The graded dimension of $W(R,S;\lambda)$ is given by $$\chi _{W(R,S;\lambda) } ({\bf x};q) = \sum_{k_1,\ldots,k_r \geq 0, k_{r+1},\ldots,k_{r+s} \in \mathbb{Z}} \frac{q^{ \frac{ ({\bf k}+{\bf l}) \cdot A \cdot ({\bf k}+{\bf l}) } {2} } } { (q)_{k_1} \cdots (q)_{k_{r}}\cdot (q)_{\infty}^s } x_1^{k_1} \cdots x_{r+s}^{k_{r+s}},$$ where ${\bf k}=(k_1,\ldots,k_{r+s},0,\ldots,0)$. The assertion follows from COROLLARY \[sec:cor11\] and the relation $${\rm wt}(e^{\tau+\lambda})=\frac{ \langle \tau+\lambda, \tau+\lambda \rangle} {2} = \frac{ ({\bf k}+{\bf l}) \cdot A \cdot ({\bf k}+{\bf l}) } {2},$$ where $\tau=k_1\beta_1+\cdots +k_n\beta_n$. For later use, we set $ \chi'_{W(R,S;\lambda)}({\bf x};q)=q^{-\langle \lambda,\lambda \rangle /2}\chi_{W(R,S;\lambda)}({\bf x};q). $ Applications (Filling in the “hole" of the characters of the “exceptional" series) =================================================================================== As an application of Theorem \[sec:char\], we will study the intermediate vertex subalgebra $V_{E_{7+1/2}}$ between the lattice VOAs $V_{E_7}$ and $V_{E_8}$. A well-known example: the intermediate vertex subalgebra $V_{A_{1/2}}$ ---------------------------------------------------------------------- First we consider the well-known result as an example of the intermediate vertex subalgebras. Let $A_1$ be a root lattice of type $A_1$ with the $\mathbb{Z}$-bilinear form $\langle, \rangle$, and let $A_1^\circ$ be the dual lattice. Consider the lattice VOA $V_{A_1}$ associated with $A_1$. Let $\alpha$ be a simple root and $\omega$ the fundamental weight. Set $\Delta_1=\{\alpha\}$. Consider the intermediate vertex subalgebra $V_{A_{1/2}}=W(\Delta_1,\emptyset)$ and the intermediate module $V_{A_{1/2}+\omega}=W(\Delta_1,\emptyset;\omega)$. Note that they agree with the principal subspaces of the basic representations of the affine Lie algebra $A_1^{(1)}$. The graded dimensions were described in [@FS; @SF]: $$\chi_{V_{A_{1/2}}}(x;q)=\sum_{k=0}^\infty \frac{q^{k^2}} {(q)_k}x^k$$ and $$\chi'_{V_{A_{1/2}+\omega} } (x;q) = \sum_{k=0}^\infty \frac{q^{k^2+k} } {(q)_k}x^k.$$ Set $$c=\frac{2}{5} \ \ \ \ \mbox{and} \ \ \ \ h=\frac{1}{5},$$ as in the row of $\mu=11/900$ in Table \[tb:table1\]. We define the characters of $V_{A_{1/2}}$ and $V_{A_{1/2}+\omega}$ to be $$Z(V_{A_{1/2}};\tau)= q^{-c/24} \chi_{V_{A_{1/2}}}(1;q)$$ and $$Z(V_{A_{1/2}+\omega};\tau)=q^{h-c/24} \chi'_{V_{A_{1/2}+\omega}}(1;q),$$ where $\tau \in \mathbb{H}$ and $q=e^{2\pi i \tau}$. Then the vector space spanned by the characters $Z(V_{A_{1/2}};\tau)$ and $Z(V_{A_{1/2}+\omega};\tau)$ is invariant under modular transformations and is the space of the solutions of the modular differential equation (\[eq:diff1\]) with $\mu=11/900$: $$\left( q\frac{d}{dq}\right)^2 f(\tau) + 2 E_2(\tau)\left(q\frac{d}{dq}\right)f(\tau)-\frac{11}{5}E_4(\tau)f(\tau)=0.$$ In fact, we have $$\label{eq:rr1} Z\left(V_{A_{1/2}};\tau\right)=q^{-1/60} \sum_{k=0}^\infty \frac{q^{k^2}} {(q)_k}$$ and $$\label{eq:rr2} Z\left(V_{A_{1/2}+\omega} ;\tau\right) = q^{11/60} \sum_{k=0}^\infty \frac{q^{k^2+k} } {(q)_k}.$$ The RHS of (\[eq:rr2\]) agrees with the character of the Virasoro minimal model $L(-22/5,0)$ at $c=-22/5$, and the RHS of (\[eq:rr1\]) agrees with that of the unique non-identity irreducible module $L(-22/5,-1/5)$. Hence, $$Z\left(V_{A_{1/2}};\tau\right)=Z\left(L\left(-22/5,-1/5\right);\tau\right),$$ and $$Z\left(V_{A_{1/2}+\omega} ;\tau\right)=Z\left(L\left(-22/5,0\right);\tau\right).$$ Therefore the characters of the intermediate subalgebra $V_{A_{1/2}}$ and the module $V_{A_{1/2}+\omega}$ form a basis of the space of the solutions of a modular differential equation. The differential equation can be standardly deduced using a singular vector of the Virasoro Verma module $V(-22/5,0)$. It coincides with the above differential equation. (cf: [@M] Theorem 6.1.) Thus the assertion holds. Functions (\[eq:rr1\]) and (\[eq:rr2\]) are the Rogers-Ramanujan functions. These functions arise naturally in solutions of the Regime I of Baxter’s Hard Hexagon model [@Bax; @A2]. For later use, put $\phi_1(\tau)=\eta(\tau)^{2/5} Z(V_{A_{1/2}})$ and $\phi_2(\tau)=\eta(\tau)^{2/5} Z(V_{A_{1/2}+\omega})$. Here, $\eta(\tau)$ is the Dedekind eta function. Note that the functions $\phi_1$ and $\phi_2$ are holomorphic modular forms of weight $1/5$ (with a suitable multiplier system) on the congruence subgroup $\Gamma(5)$ [@BKMS; @I; @Kan]. The intermediate vertex subalgebra $V_{E_{7+1/2}}$ -------------------------------------------------- $$\[email protected]@C40pt@R5pt{ &&&&\alpha_8\cr &&&&\maru\ar@{-}[ddd]\cr \vbox to 20pt{}\cr \alpha_1\cr \maru\ar@{-}[r]&\maru\ar@{-}[r]&\maru\ar@{-}[r]&\maru\ar@{-}[r]&\maru\ar@{-}[r]&\maru\ar@{-}[r]&\maru\cr &\alpha_2&\alpha_3&\alpha_4&\alpha_5&\alpha_6&\alpha_7\cr \vbox to 20pt{}\cr \maru\ar@{.}[uuu]\cr \widetilde{\alpha}\cr \vbox to 20pt{}\cr }$$ Let $E_8$ be a root lattice of type $E_8$ with the $\mathbb{Z}$-bilinear form $\langle , \rangle$. Let $\alpha_1,\ldots,\alpha_8$ be simple roots of $E_8$ and denote the highest root by $\tilde{\alpha}$, as illustrated in Figure \[fig:fig1\]. Note that a sublattice $E_7=\langle \alpha_2,\ldots,\alpha_8 \rangle \subset E_8$ is a root lattice of type $E_7$. Set $\Delta_8=\{\alpha_1,\ldots,\alpha_8\}$. Consider the lattice VOA $V_{E_8}$ associated with $E_8$ and the lattice subVOA $V_{E_7}\subset V_{E_8}$ associated with $E_7$. Set $R=\{\alpha_1\}$ and $S=\{\alpha_2,\ldots,\alpha_8\}$. Consider the intermediate vertex subalgebra $V_{E_{7+1/2}}=W(R,S)$ of $V_{E_8}$ associated with $(R,S)$. That is, $$V_{E_{7+1/2}}= \langle e^{\alpha_1}, e^{\pm \alpha_2},\ldots,e^{\pm \alpha_8}, \alpha_2,\ldots,\alpha_8 \rangle_{{\rm v.a.}} \subset V_{E_8}.$$ As in the previous subsection, we will fill in the “hole" of the “exceptional series" using the intermediate vertex subalgebra $V_{E_{7+1/2}}$. For the purpose, we need to take a “non-identity" intermediate module with the character satisfying the modular differential equation. If we consider the ways of the theories of the intermediate Lie algebras and modules, we should take it in the “non-identity irreducible module" of $V_{E_8}$. But such a module does not exist since $E_8$ is a unimodular lattice, that is, $E_8=E_8^\circ$. Therefore we try to take it in the identity module $V_{E_8}$. It will go well. Consider the element $\alpha_1\in E_8$ and the intermediate module $W(R,S;\alpha_1)$. We set $V_{E_{7+1/2}+\alpha_1}=W(R,S;\alpha_1)$. Set $$c=\frac{38}{5} \ \ \ \ \mbox{and} \ \ \ \ h=\frac{4}{5},$$ as in the row of $\mu=551/900$ in Table \[tb:table1\]. We define the characters of $V_{E_{7+1/2}}$ and $V_{E_{7+1/2}+\alpha_1}$ to be $$Z(V_{E_{7+1/2}};\tau)= q^{-c/24} \chi_{V_{E_{7+1/2}}}(1,\ldots,1;q)$$ and $$Z(V_{E_{7+1/2}+\alpha_1};\tau)=q^{h-c/24} \chi'_{V_{E_{7+1/2}+\alpha_1}}(1,\ldots,1;q),$$ where $\tau \in \mathbb{H}$ and $q=e^{2\pi i \tau}$. Then the vector space spanned by the characters $Z(V_{E_{7+1/2}};\tau)$ and\ $Z(V_{E_{7+1/2}+\alpha_1};\tau)$ is invariant under modular transformations and is the space of the solutions of the modular differential equation (\[eq:diff1\]) with $\mu=551/900$: $$\label{eq:diff712} \left( q\frac{d}{dq}\right)^2 f(\tau) + 2 E_2(\tau)\left(q\frac{d}{dq}\right)f(\tau)-\frac{551}{5}E_4(\tau)f(\tau)=0.$$ Let us prove the assertion. Denote by $(\omega_2,\ldots,\omega_8)$ ($\omega_i \in E_7^\circ$ ($i=2,\ldots,8$)) the dual basis of the basis $(\alpha_2,\ldots,\alpha_8)$ of the sublattice $E_7$. Consider the Virasoro minimal model $L(-3/5,0)$ at $c=-3/5$ and the modules $L(-3/5,h)$ with the conformal weights $h=0,3/4, 1/5$ and $-1/20$. First, we show that the characters satisfy the following equalities: $$\begin{aligned} \label{eq:zv712} Z\left(V_{E_{7+1/2}};\tau\right)&=&Z\left(V_{E_7};\tau\right)\cdot Z\left(L\left(-3/5,-1/20\right);\tau\right) + \nonumber \\ && Z\left(V_{E_7+\omega_2};\tau\right) \cdot Z\left(L\left(-3/5,1/5\right);\tau\right)\end{aligned}$$ and $$\begin{aligned} \label{eq:zl712} Z\left(V_{E_{7+1/2}+\alpha_1};\tau\right)&=&Z\left(V_{E_7};\tau\right)\cdot Z\left(L\left(-3/5,3/4\right);\tau\right) + \nonumber \\ && Z\left(V_{E_7+\omega_2};\tau\right) \cdot Z\left(L\left(-3/5,0\right);\tau\right).\end{aligned}$$ By Theorem \[sec:char\], the characters are $$\label{eq:zv712pre} Z\left(V_{E_{7+1/2}};\tau\right)=q^{-19/60} \cdot \sum_{ k_1\geq 0, k_2,\ldots,k_8 \in \mathbb{Z}} \frac{q^{ \frac{ {\bf k}\cdot M_8 \cdot {\bf k}^{\rm T} }{2} } }{(q)_{k_1} (q)_{\infty}^7 }$$ and $$\label{eq:zl712pre} Z\left(V_{E_{7+1/2}+\alpha_1};\tau\right)=q^{29/60}\cdot \sum_{ k_1\geq 1, k_2,\ldots,k_8 \in \mathbb{Z}} \frac{q^{ \frac{ {\bf k}\cdot M_8 \cdot {\bf k}^{\rm T} }{2} } }{(q)_{k_1-1} (q)_{\infty}^7},$$ where ${\bf k}=(k_1,...,k_8)$. Here $M_8$ is the Cartan matrix $M_8=(\langle \alpha_i, \alpha_j \rangle)_{(i,j=1,\ldots,8)}$ of $E_8$. In (\[eq:zv712pre\]), ${\bf k}=(k_1,\ldots,k_8)$ implies the charge $k_1\alpha_1+\cdots +k_8\alpha_8$ of $V$. Note that the set of the charges of $V_{E_{7+1/2}}$ agrees with the set $L(R,S)=\{k_1\alpha_1+\cdots +k_8\alpha_8 | k_1\geq 0, k_2,\ldots,k_8 \in \mathbb{Z}\}$. Recall (cf. [@K2]) that the highest root $\tilde{\alpha}$ of $E_8$ is $$\tilde{\alpha}= 2\alpha_1+3\alpha_2+4\alpha_3+5\alpha_4+6\alpha_5+4\alpha_6+2\alpha_7+3\alpha_8.$$ Set $$L(R,S)^{\rm even}=\{k_2\alpha_2+\cdots +k_8\alpha_8+ k\tilde{\alpha} | k_2,\ldots,k_8 \in \mathbb{Z}, k\geq 0\}$$ and $$L(R,S)^{\rm odd}=\{ k_2\alpha_2+\cdots+k_8\alpha_8+k\tilde{\alpha} +\alpha_1 | k_2,\ldots,k_8 \in \mathbb{Z}, k\geq 0\}.$$ Then $L(R,S)=L(R,S)^{\rm even} \sqcup L(R,S)^{\rm odd}$. Let $k_2,\ldots,k_8$ be integers and $k$ a non-negative integer. Set $\mu=k_2\alpha_2+\cdots+k_8\alpha_8+k\tilde{\alpha}$ and $\nu=k_2\alpha_2+\cdots+k_8\alpha_8+k\tilde{\alpha}+\alpha_1$. Then $\mu\in L(R,S)^{\rm even}$, and $\nu\in L(R,S)^{\rm odd}$. Set $\beta=k_2\alpha_2+\cdots+k_8\alpha_8$ and ${\bf k}'=(k_2,\ldots,k_8)$. Then $\beta$ belongs to $L(S)=E_7$. Considering the extended Dynkin diagram of $E_8$ (Figure \[fig:fig1\]), we obtain $$\begin{aligned} \langle \beta,\beta\rangle &=& {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}, \\ \langle \beta, \tilde{\alpha} \rangle &=&0, \\ \langle \beta, \alpha_1 \rangle &=& \langle k_2\alpha_2, \alpha_1 \rangle =-k_2\end{aligned}$$ and $$\langle \tilde{\alpha}, \alpha_1 \rangle = \langle 3\alpha_2+2\alpha_1, \alpha_1 \rangle = 1.$$ Here $M_7$ is the Cartan matrix $(\langle \alpha_i, \alpha_j \rangle)_{(i,j=2,\ldots,8)}$ of $E_7$. Then we have $$\begin{aligned} \frac{\langle \mu, \mu \rangle}{2} = \frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}} }{2} + k^2\end{aligned}$$ and $$\begin{aligned} \frac{\langle \nu, \nu \rangle}{2} = \frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}} }{2} - k_2+ k^2+ k + 1.\end{aligned}$$ Therefore we have $$\begin{aligned} && Z\left(V_{E_{7+1/2}};\tau\right) \nonumber \\ &=& q^{-19/60} \sum_{k\geq 0, k_2,\ldots, k_8 \in \mathbb{Z}} \frac {q^{\frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}}{2} + k^2}} {(q)^7_\infty (q)_{2k}} + \nonumber \\ &&q^{-19/60} \sum_{k\geq 0, k_2,\ldots, k_8 \in \mathbb{Z}} \frac {q^{\frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}} {2} - k_2+k^2+k+1}} {(q)^7_\infty (q)_{2k+1}} \nonumber \\ &=& \left( q^{-7/24}\sum_{k_2,\ldots, k_8\in \mathbb{Z}} \frac {q^{\frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}}{2}}} {(q)_\infty^7} \right) \cdot \left( q^{-1/40} \sum_{k\geq 0} \frac {q^{k^2}} {(q)_{2k}} \right) + \nonumber \\ && \left( q^{11/24} \sum_{k_2,\ldots, k_8 \in \mathbb{Z}} \frac {q^{\frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}}{2} -\langle \beta,\omega_2\rangle }} {(q)_\infty^7} \right) \cdot \left( q^{-9/40} \sum_{k\geq 0} \frac {q^{k^2+k}} {(q)_{2k+1}} \right), \nonumber \\ \label{eq:bhh1}\end{aligned}$$ where ${\bf k}'= (k_2,\ldots, k_8)$, and $\beta=k_2\alpha_2+\cdots +k_8\alpha_8$. Similarly, we have $$\begin{aligned} && Z\left(V_{E_{7+1/2}+\alpha_1};\tau\right) \nonumber \\ &=& \left( q^{-7/24}\sum_{k_2,\ldots,k_8 \in \mathbb{Z}} \frac {q^{\frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}}{2}}} {(q)_\infty^7} \right) \cdot \left( q^{-9/40} \sum_{k\geq 1} \frac {q^{k^2}} {(q)_{2k-1}} \right) + \nonumber \\ && \left( q^{11/24} \sum_{k_2,\ldots, k_8 \in \mathbb{Z}} \frac {q^{\frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}}{2} -\langle \beta,\omega_2\rangle }} {(q)_\infty^7} \right) \cdot \left( q^{1/40} \sum_{k\geq 0} \frac {q^{k^2+k}} {(q)_{2k}} \right), \nonumber \\ \label{eq:bhh2}\end{aligned}$$ where ${\bf k}'=(k_2,\ldots, k_8)$, and $\beta=k_2\alpha_2+\cdots +k_8\alpha_8$. It is known [@KKMM] that the four functions appearing in the above equalities agree with the characters of the Virasoro minimal model at $c=-3/5$: $$\begin{aligned} q^{-1/40}\sum_{k\geq 0} \frac {q^{k^2}} {(q)_{2k}} &=& Z\left(L\left(-3/5,-1/20\right);\tau\right), \\ q^{9/40}\sum_{k\geq 0} \frac {q^{k^2+k}} {(q)_{2k+1}} &=& Z\left(L\left(-3/5,1/5\right);\tau\right), \\ q^{-9/40}\sum_{k\geq 1} \frac {q^{k^2}} {(q)_{2k-1}} &=& Z\left(L\left(-3/5,3/4\right);\tau\right)\end{aligned}$$ and $$\begin{aligned} q^{1/40}\sum_{k\geq 0} \frac {q^{k^2+k}} {(q)_{2k}} &=& Z\left(L\left(-3/5,0\right);\tau\right).\end{aligned}$$ Furthermore, by the definition of the characters of the lattice VOAs, we have $$\begin{aligned} q^{-7/24}\sum_{k_2,\ldots,k_8 \in \mathbb{Z}} \frac {q^{\frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}}{2}}} {(q)_\infty^7} = Z\left(V_{E_7};\tau\right)\end{aligned}$$ and $$\begin{aligned} q^{11/24} \sum_{k_2,\ldots, k_8 \in \mathbb{Z}} \frac {q^{\frac{ {\bf k}' \cdot M_7 \cdot {\bf k}'^{{\rm T}}}{2} -\langle \beta,\omega_2\rangle }} {(q)_\infty^7} = Z\left(V_{E_7-\omega_2};\tau\right),\end{aligned}$$ where ${\bf k}'=(k_2,\ldots, k_8)$ and $\beta=k_2\alpha_2+\cdots+ k_8\alpha_8$. Since $$-\omega_2=-\frac{1}{2}(3\alpha_2+4\alpha_3+5\alpha_4+6\alpha_5+4\alpha_6+2\alpha_7+3\alpha_8)$$ is a representative of the unique non-zero element of $E_7^\circ/E_7\cong \mathbb{Z}/2\mathbb{Z}$, the module $V_{E_7-\omega_2}$ (also, $V_{E_7+\omega_2}$) is the unique non-identity irreducible module of the lattice VOA $V_{E_7}$ up to equivalence. Thus we have (\[eq:zv712\]) and (\[eq:zl712\]). The four functions $$\sum_{k\geq 0} \frac {q^{k^2}} {(q)_{2k}}, \ \ \ \sum_{k\geq 0} \frac {q^{k^2+k}} {(q)_{2k+1}}, \ \ \ \sum_{k\geq 1} \frac {q^{k^2}} {(q)_{2k-1}} \ \ \mbox{and} \ \ \sum_{k\geq 0} \frac {q^{k^2+k}} {(q)_{2k}}$$ appearing in equalities (\[eq:bhh1\]) and (\[eq:bhh2\]) arise naturally in solutions of the Regime IV of Baxter’s Hard Hexagon model [@Bax; @A2]. By the theories of VOAs (or theories of infinite dimensional Lie algebras [@K2; @W],) we obtain the rules of modular transformations for the characters of the VOAs $V_{E_7}$ and $L(-3/5,0)$: $$\begin{aligned} \begin{pmatrix} Z\left(V_{E_7}; -1/\tau\right) \\ Z\left(V_{E_7+\omega_2};-1/\tau\right) \end{pmatrix} =\frac{1}{\sqrt{2}} \begin{bmatrix} 1 & 1 \\ 1 & -1 \end{bmatrix} \begin{pmatrix} Z\left(V_{E_7};\tau\right) \\ Z\left(V_{E_7+\omega_2};\tau\right) \end{pmatrix},\label{eqn:mod1}\end{aligned}$$ $$\begin{aligned} \begin{pmatrix} Z\left(V_{E_7};\tau+1\right) \\ Z\left(V_{E_7+\omega_2}; \tau + 1 \right) \end{pmatrix} = \begin{bmatrix} e^{2\pi i(-7/24)} & 0 \\ 0& e^{2\pi i(11/24)} \end{bmatrix} \begin{pmatrix} Z\left(V_{E_7}; \tau\right) \\ Z\left(V_{E_7+\omega_2}; \tau\right) \end{pmatrix},\label{eqn:mod2}\end{aligned}$$ $$\begin{aligned} && \begin{pmatrix} Z(L(-3/5,0);-1/\tau) \\ Z(L(-3/5,3/4);-1/\tau) \\ Z(L(-3/5,1/5);-1/\tau) \\ Z(L(-3/5,-1/20);-1/\tau) \end{pmatrix} \nonumber \\ &=& \sqrt{\frac{2}{5}} \begin{bmatrix} \sin(2\pi/5) & -\sin(2\pi/5) & -\sin(\pi/5)& \sin(\pi/5) \\ -\sin(2\pi/5) & -\sin(2\pi/5) & \sin(\pi/5 & \sin(\pi/5) \\ -\sin(\pi/5)& \sin(\pi/5)& \sin(2\pi/5)& \sin(2\pi/5) \\ \sin(\pi/5)& \sin(\pi/5)& \sin(2\pi/5)& \sin(2\pi/5) \end{bmatrix} \cdot \nonumber \\ && \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \begin{pmatrix} Z(L(-3/5,0);\tau) \\ Z(L(-3/5,3/4);\tau) \\ Z(L(-3/5,1/5);\tau) \\ Z(L(-3/5,-1/20);\tau) \end{pmatrix} \label{eqn:mod3}\end{aligned}$$ and $$\begin{aligned} && \begin{pmatrix} Z(L(-3/5,0);\tau+1) \\ Z(L(-3/5,3/4);\tau+1) \\ Z(L(-3/5,1/5);\tau+1) \\ Z(L(-3/5,-1/20);\tau+1) \end{pmatrix} \nonumber \\ &=& \begin{bmatrix} e^{2\pi i(1/40)} & 0&0&0 \\ 0& e^{2\pi i(31/40)} & 0 & 0 \\ 0 & 0 & e^{2\pi i(9/40)} & 0 \\ 0& 0&0& e^{2\pi i(-1/40)} \end{bmatrix} \cdot \nonumber \\ && \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \begin{pmatrix} Z(L(-3/5,0);\tau) \\ Z(L(-3/5,3/4);\tau) \\ Z(L(-3/5,1/5);\tau) \\ Z(L(-3/5,-1/20);\tau) \end{pmatrix} .\label{eqn:mod4}\end{aligned}$$ By (\[eq:zv712\]),(\[eq:zl712\]) and (\[eqn:mod1\]) – (\[eqn:mod4\]), we have $$\begin{aligned} && \begin{pmatrix} Z\left(V_{E_{7+1/2}};-1/\tau\right) \\ Z\left(V_{E_{7+1/2}+\alpha_1};-1/\tau\right) \end{pmatrix} \nonumber \\ &=& \frac{2}{\sqrt{5}} \begin{bmatrix} \sin(2\pi/5) & \sin(\pi/5) \\ \sin(\pi/5) & -\sin(2\pi/5) \end{bmatrix} \cdot \begin{pmatrix} Z\left(V_{E_{7+1/2}};\tau\right) \\ Z\left(V_{E_{7+1/2}+\alpha_1};\tau\right) \end{pmatrix}\end{aligned}$$ and $$\begin{aligned} && \begin{pmatrix} Z\left(V_{E_{7+1/2}};\tau+1\right) \\ Z\left(V_{E_{7+1/2}+\alpha_1}; \tau + 1 \right) \end{pmatrix} \nonumber \\ &=& \begin{bmatrix} e^{2\pi i(-19/60)} & 0 \\ 0& e^{2\pi i(29/60)} \end{bmatrix} \cdot \begin{pmatrix} Z\left(V_{E_{7+1/2}};\tau\right) \\ Z\left(V_{E_{7+1/2}+\alpha_1};\tau\right) \end{pmatrix}.\end{aligned}$$ Hence, especially, the vector space spanned by the characters of the intermediate vertex subalgebra $V_{E_{7+1/2}}$ is invariant under modular transformations. Furthermore, the rules coincides with that of modular transformations of a system of the solutions of the modular differential equations (\[eq:diff1\]) with $\mu=551/900$ [@MMS2]: $$f_1(\tau)=\left( \frac{1}{16}\lambda(1-\lambda)\right)^{(1-x)/6} F\left(\frac{1}{2}-\frac{1}{6}x,\frac{1}{2}-\frac{1}{2}x,1-\frac{1}{3}x,\lambda\right)$$ and $$f_2(\tau)=N\cdot \left( \frac{1}{16}\lambda(1-\lambda)\right)^{(1+x)/6} F\left(\frac{1}{2}+\frac{1}{6}x,\frac{1}{2}+\frac{1}{2}x,1+\frac{1}{3}x,\lambda\right),$$ where $N$ is a normalization constant, $ x=\sqrt{1+36\mu}=1+\frac{1}{2}c $ and $ \lambda=\vartheta^4_2(\tau)/\vartheta^4_3(\tau). $ Here, $\vartheta_i$ $(i=1,\ldots,4)$ are the Jacobi theta functions and $F$ are the hypergeometric functions. If we multiply the solutions \[resp., the characters\] by $\eta(\tau)^{38/5}$, we obtain holomorphic modular forms $$g_1=\eta(\tau)^{38/5}f_1(\tau) \ \ \ \mbox{and} \ \ \ g_2=\eta(\tau)^{38/5}f_2(\tau)$$ $$[\mbox{resp.,} \ \ \psi_1=\eta(\tau)^{38/5}Z\left(V_{E_{7+1/2}};\tau\right) \ \ \ \mbox{and} \ \ \ \psi_2=\eta(\tau)^{38/5}Z\left(V_{E_{7+1/2}};\tau\right) ]$$ of weight $19/5$ on $\Gamma(5)$ (with a multiplier system that is $19$ times of that of $\phi_1=\eta(\tau)^{2/5}Z\left(V_{A_{1/2}}\right)$ and $\phi_2=\eta(\tau)^{2/5}Z\left(V_{A_{1/2}+\omega}\right))$. It is known that the ring of holomorphic modular forms of weight $\frac{1}{5}\mathbb{Z}$ on $\Gamma(5)$ with this multiplier system is the polynomial ring $\mathbb{C}[\phi_1,\phi_2]$ [@BKMS; @I; @Kan]. Therefore, $g_1,g_2, \psi_1$ and $\psi_2$ are homogeneous polynomials in $\phi_1$ and $\phi_2$ of degree $19$, hence $f_1,f_2,Z\left(V_{E_{7+1/2}}\right)$ and $Z\left(V_{E_{7+1/2}+\alpha_1}\right)$ are homogeneous polynomials in $p_1=Z\left(V_{A_{1/2}}\right)$ and $p_2=Z\left(V_{A_{1/2}+\omega}\right)$ of degree $19$. Then explicitly computing the fist few terms of the $q$-expansions of the homogeneous polynomials $p_1^ip_2^{19-i}$ ($i=0,\ldots,19$), the solutions, and the characters, we obtain $ f_1(\tau)=Z\left(V_{E_{7+1/2}}\right)$ and $f_2(\tau)=Z\left(V_{E_{7+1/2}+\alpha_1}\right). $ We show $f_1(\tau)=Z\left(V_{E_{7+1/2}}\right)$. We have $$f_1(\tau)=q^{-19/60}(1+190q+2831q^2+22306q^3+O(q^4))$$ and $$Z\left(V_{E_{7+1/2}};\tau\right)=q^{-19/60}(1+190q+2831q^2+22306q^3+O(q^4)).$$ Since both the functions $f_1$ and $Z(V_{E_{7+1/2}})$ belong to the ring $q^{-19/60}\mathbb{C}[[q]]$, the functions $f_1$ and $Z(V_{E_{7+1/2}})$ must be linear combinations of the polynomials $ p_1^{19}$, $p_1^{14}p_2^5$, $p_1^{9}p_2^{10}$ and $ p_1^4p_2^{15}. $ The first four terms of these polynomials are as follows: $$\begin{aligned} p_1^{19} &=& q^{-19/60}(1+19q+190q^2+1330q^3+O(q^4)), \\ p_1^{14}p_2^5 &=&q^{-19/60}(0+q+14q^2+110q^3+O(q^4)), \\ p_1^9p_2^{10} &=& q^{-19/60}(0+0+q^2+9q^3+O(q^4))\end{aligned}$$ and $$p_1^4p_2^{15}=q^{-19/60}(0+0+0+q^3+O(q^4)).$$ Therefore we obtain $$f_1(\tau)=p_1^{19}+171p_1^{14}p_2^5+247p_1^9p_2^{10}-57p_1^4p_2^{15}$$ and $$\label{eq:kanv712} Z\left(V_{E_{7+1/2}};\tau\right)=p_1^{19}+171p_1^{14}p_2^5+247p_1^9p_2^{10}-57p_1^4p_2^{15}.$$ Hence we obtain $ f_1(\tau)=Z\left(V_{E_{7+1/2}};\tau\right). $ Similarly, we obtain $$\label{eq:kanl712} f_2(\tau)=57p_1^{15}p_2^4+247p_1^{10}p_2^9-171p_1^5p_2^{14}+p_2^{19}=Z\left(V_{E_{7+1/2}+\alpha_1}\right).$$ Thus we have the assertion that the characters of the intermediate vertex subalgebra $V_{E_{7+1/2}}$ form a basis of the space of the solutions of the modular differential equation (\[eq:diff712\]). (Note that the polynomials appearing in (\[eq:kanv712\]) and (\[eq:kanl712\]) were studied in [@Kan] as a system of the solutions of the differential equation [@Kan; @Kan2; @Kan3] $$\begin{aligned} f''(\tau)-\frac{k+1}{6}E_2(\tau)f'(\tau) + \frac{k(k+1)}{12}E'_2(\tau)f(\tau)=0\end{aligned}$$ with $k=19/5$. The above differential equation is equivalent to the modular differential equation (\[eq:diff1\]) with $\mu=k(k+2)/36$. More precisely, $f(\tau)$ satisfies (\[eq:diff1\]) with $\mu=k(k+2)/36$ if and only if $\eta(\tau)^{2k}f(\tau)$ satisfies the above equation. Hence we can deduce from (\[eq:zv712\]) and (\[eq:zl712\]) the new descriptions of the system of the solutions of the above differential equation with $k=19/5$ using\ 1) the characters of the intermediate vertex subalgebra $V_{E_{7+1/2}}$ and module or\ 2) the characters of the lattice VOA associated with the $E_7$ root lattice and Virasoro minimal model at $c=-3/5$.) The lowest weight subspace of the module $V_{E_{7+1/2}+\alpha_1}$ is $57$-dimensional and contains the $56$-dimensional irreducible module of $E_{7}\subset E_{7+1/2}$. The module $V_{E_{7+1/2}+\alpha_1}$ is expected to give modules of $E_{7+1/2}$ not having been studied. Proof of THEOREM \[sec:str\] ============================ In this section, we prove the main theorem. Let us take over the setting and notations in §2. \[sec:lem1\] Let $\lambda$ be an element of $L^\circ$. Let $\tau_1,\ldots,\tau_l$ be elements of $L$. Let $i_1,\ldots,i_l$ be integers. Then, (i) : there exists a unique $g\in {{\rm S}( {{\widehat{\mathfrak{h}}}}{}_- )}$ such that $$(e^{\tau_1})_{(i_1+\langle \tau_1,\lambda+\tau_l+\cdots +\tau_2 \rangle )}\ldots(e^{\tau_l})_{(i_l+\langle \tau_l,\lambda \rangle)} \mbox{\boldmath $1$} =g\otimes e^{\tau_1+\cdots +\tau_l}. \label{eqn:lem21}$$ Furthermore, (ii) : there exists a non-zero constant $r\in \mathbb{C}$ such that $$(e^{\tau_1})_{i_1}\ldots(e^{\tau_l})_{i_l} e^\lambda =r \cdot g\otimes e^{\lambda+\tau_1+\cdots +\tau_l}. \label{eqn:lem2}$$ The assertions follow from the definition of vertex operators for the lattice VOA. Note that if one puts $$r=\epsilon(\tau_l,\lambda)\cdot \epsilon(\tau_{l-1},\lambda+\tau_l)\cdots \epsilon(\tau_1,\lambda+\tau_l+\cdots+\tau_2) \in \mathbb{C},$$ $r$ is non-zero, and (\[eqn:lem2\]) holds. If $\lambda\in (L(R))^\circ$, the assertion was proved in [@MP] (Corollary 4.8.) The rest of the assertions follow from Lemma \[sec:lem1\] and that of the case $\lambda=0\in (L(R))^\circ$. \[sec:cor1prop1\] Let $R'$ and $R''$ be disjoint subsets of $B$. Let $\delta'$ be an element of $L_+(R')$ and $\delta''$ an element of $L_+(R'')$. Then $$(W(R'\sqcup R'';\lambda))^{\delta'+\delta''} \subset {{\rm S}( {{\widehat{\mathfrak{h}}}(R')}{}_- )}\cdot (W(R'';\delta'+\lambda))^{\delta''}. \label{eqn:corprop1eq1}$$ Denote the RHS of (\[eqn:corprop1eq1\]) by X. We show the case $R'=\{\beta_1,\ldots,\beta_{n-1}\}$ and $R''=\{\beta_n\}$. Since $\delta'\in L_+(R')$, $\delta'$ has the form $k_1\beta_1+\cdots+k_{n-1}\beta_{n-1}$ with $k_1,\ldots,k_{n-1}\geq 0$. Since $\delta''\in L_+(R'')$, $\delta''$ has the form $k_n\beta_n$ with $k_n \geq 0$. Take $v\in {{\mathcal}B}(R'\sqcup R'';\lambda;k_1,\ldots,k_n)$. By Lemma \[sec:lemmp\], it suffices to show $v\in X$. The element $v$ has the form $ \varepsilon^{\beta_n}_{\mu_n} \ldots \varepsilon^{\beta_1}_{\mu_1} e^\lambda $ with $\mu_i \in M_i(R'\sqcup R'';\lambda;k_1,\ldots,k_n)$ for $1\leq i \leq n$, and $ \varepsilon^{\beta_{n-1}}_{\mu_{n-1}} \ldots \varepsilon^{\beta_1}_{\mu_1}. e^\lambda $ has the form $ g \cdot e^{k_1\beta_1+\cdots+k_{n-1}\beta_{n-1}+\lambda} $ with $g\in {{\rm S}( {{\widehat{\mathfrak{h}}}(R')}{}_- )}$. Then we have $ v=\varepsilon^{\beta_n}_{\mu_n}. g\cdot e^{\delta'+\lambda}. $ By the commutation relation $$[h_k, (e^\alpha)_l]=\langle h,\alpha \rangle (e^\alpha)_{k+l}, \ \ \ \alpha \in L , h\in \mathfrak{h}, k,l \in \mathbb{Z}, \label{eqn:opee}$$ of vertex operators of lattice VOAs, it follows that $\varepsilon^{\beta_n}_{\mu_n}. g$ is the sum of operators of the form $ h\cdot (e^{\beta_n})_{q_{k_n}} \ldots (e^{\beta_n})_{q_1} $ with $h\in {{\rm S}( {{\widehat{\mathfrak{h}}}(R')}{}_- )}$ and $q_1,\ldots,q_{k_n} \in \mathbb{Z}$. Since $\beta_n \in R''$, we have $ (e^{\beta_n})_{q_{k_n}} \ldots (e^{\beta_n})_{q_1}.e^{\delta'+\lambda} \in (W(R'';\delta'+\lambda))^{k_n\beta_n}. $ Therefore, $v\in X$. Thus we have proved the lemma. Let $\lambda$ be an element of $L^\circ$. Let $\gamma$ be an element of $L_+(R)$ and $\delta$ an element of $L(S)$. For $\mu \in L(R,S)$, consider the set $$\begin{aligned} T(\mu;\lambda) &=& \{ f\cdot (e^{\tau_l})_{i_l}\ldots(e^{\tau_1})_{i_1}.e^{\tau_0+\lambda} | f\in {{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}, l\geq 0, i_1,\ldots,i_l\in \mathbb{Z}, \\ && \ \ \ \ \ \tau_0\in L(S), \tau_1,\ldots,\tau_l \in R\sqcup S \sqcup (-S) \ \ \mbox{with} \\ && \ \ \ \ \ \tau_0+\tau_1+\cdots +\tau_l=\mu \}.\end{aligned}$$ Here, $- S=\{-\sigma \in L | \sigma \in S\}$. \[sec:lemmonomial\] The set $T(\delta+\gamma;\lambda)$ spans the vector space $W(R,S;\lambda)^{\delta+\gamma}$. For $\tau\in L(S)$, we have $ e^{\tau+\lambda}= \varepsilon (\tau,\lambda)(e^\tau)_{-1-\langle \tau,\lambda \rangle}e^\lambda. $ Then it follows that each element of $T(\delta+\gamma;\lambda)$ belongs to $W(R,S;\lambda)^{\delta+\gamma}$. Consider the subset $$\begin{aligned} && \{ f\cdot (e^{\tau_l})_{i_l}\ldots(e^{\tau_1})_{i_1}.e^{\lambda} | f\in {{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}, l\geq 0, i_1,\ldots,i_l\in \mathbb{Z}, \\ && \ \ \ \ \ \tau_1,\ldots,\tau_l \in R\sqcup S \sqcup (-S) \ \ \mbox{with} \ \ \tau_1+\cdots +\tau_l=\delta+\gamma \}\end{aligned}$$ of $T(\delta+\gamma;\lambda)$. By (\[eqn:opee\]), it follows that the subset spans $W(R,S;\lambda)^{\delta+\gamma}$. Hence the set $T(\delta+\gamma;\lambda)$ spans $W(R,S;\lambda)^{\delta+\gamma}$. \[sec:thm1\] $$W(R,S;\lambda)^{\delta+\gamma}={{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )} \cdot W(R;\delta+\lambda)^{\gamma}. \label{eqn:thm1}$$ We show the case $\lambda=0$. That is, we show $ W(R,S)^{\delta+\gamma}={{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}\cdot W(R;\delta)^{\gamma}. $ Denote the LHS by $X$ and the RHS by $Y$. Note that both $X$ and $Y$ are free ${{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}$-modules. We have $X \supset Y$, since $W(R;\delta)\subset W(R,S)$. Let us show $X \subset Y$. Take $v\in T(\delta+\gamma;0)$. By Lemma \[sec:lemmonomial\], it suffices to show $ v\in Y. $ The element $v$ has the form $ f \cdot (e^{\tau_l})_{i_l}\ldots(e^{\tau_1})_{i_1}. e^{\tau_0} $ with $f\in {{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}$, $l\geq 0$, $i_1,\ldots,i_l\in \mathbb{Z}$, $\tau_0 \in L(S)$ and $\tau_1,\ldots,\tau_l \in R\sqcup S \sqcup (-S)$ with $\tau_0+\tau_1+\cdots +\tau_l=\delta+\gamma$. If $\tau_1,\ldots,\tau_l\in R$, then $\tau_0=\delta$ and $\tau_1+\cdots+\tau_l=\gamma$, since $R\cap (S\sqcup (-S))=\emptyset$. Then it follows that $v\in Y$. So assume that $j \in \{1,\ldots,l\}$ is the minimum number satisfying $\tau_j\not \in R$. Then $\tau_j$ has the form $ \tau_j=t\cdot \sigma_k $ with $t\in \{\pm 1\}$ and $k\in \{1,\ldots,s\}$. Put $R'=t\cdot S$ and $R''=R$. Here, $t\cdot S=\{t\cdot \sigma_1,\ldots,t\cdot \sigma_s\}$. Then $R'$ and $R''$ are disjoint. Consider the vector $ v'= (e^{\tau_j})_{i_j} \cdots (e^{\tau_1})_{i_1} e^{\tau_0}. $ (Note that $v=f\cdot (e^{\tau_l})_{i_l}\ldots(e^{\tau_{j+1}})_{i_{j+1}}.v'$.) Then $v'\in (W(R'\sqcup R'';\tau_0))^{\tau_1+\cdots+\tau_j}$, since $\tau_j \in R'$ and $\tau_1,\ldots,\tau_{j-1} \in R''$. Moreover, $\tau_j\in L_+(R')$, and $\tau_1+\cdots+\tau_{j-1}\in L_+(R'')$. From Lemma \[sec:cor1prop1\] and the relation ${{\rm S}( {{\widehat{\mathfrak{h}}}(t\cdot S)}{}_- )}={{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}$, it follows that $v' \in {{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}\cdot (W(R;\tau_0+\tau_j))^{\tau_1+\cdots+\tau_{j-1}}. $ Therefore $v'$ is the sum of elements of the form $$g\cdot (e^{\tau'_{j-1}})_{i'_{j-1}}\ldots(e^{\tau'_1})_{i'_1}.e^{\tau_0+\tau_j}$$ with $g\in{{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}$, $i'_1,\ldots,i'_{j-1}\in \mathbb{Z}$ and $\tau'_1,\ldots,\tau'_{j-1}\in R$ with $\tau'_1+\cdots+\tau'_{j-1}=\tau_1+\cdots+\tau_{j-1}$. Therefore, to show $v\in Y$, it suffices to show that the vector $ f \cdot (e^{\tau_l})_{i_l}\ldots(e^{\tau_{j+1}})_{i_{j+1}}.g\cdot (e^{\tau'_{j-1}})_{i'_{j-1}}\ldots(e^{\tau'_1})_{i'_1}.e^{\tau_0+\tau_j} $ belongs to $Y$. By (\[eqn:opee\]), the operator $(e^{\tau_l})_{i_l}\ldots(e^{\tau_{j+1}})_{i_{j+1}}.g$ is the sum of operators of the form $ h\cdot (e^{\tau_l})_{p_l}\cdots (e^{\tau_{j+1}})_{p_{j+1}} $ with $h\in {{\rm S}( {{\widehat{\mathfrak{h}}}(S)}{}_- )}$ and $p_{j+1},\ldots,p_l \in \mathbb{Z}$. Therefore, to show $v\in Y$, it suffices to show that the vector $$w=f\cdot h\cdot (e^{\tau_l})_{p_l}\ldots(e^{\tau_{j+1}})_{p_{j+1}}.(e^{\tau'_{j-1}})_{i'_{j-1}}\ldots(e^{\tau'_1})_{i'_1}.e^{\tau_0+\tau_j}$$ belongs to $Y$. Since $\tau_j \in S\sqcup (-S)$, the vector $w$ belongs to $T(\delta+\gamma;0)$ as the vector $v$. Since $\tau_j \in S\sqcup (-S)$ and $\tau'_1,\ldots,\tau'_{j-1} \in R$, the number of elements of the set $$\{\tau| \tau\not \in R, \tau\in\{\tau'_1,\ldots,\tau'_{j-1},\tau_{j+1},\ldots,\tau_l\} \}$$ is fewer than the number of elements of the set $ \{\tau| \tau\not \in R, \tau \in \{\tau_1,\ldots,\tau_l\}\}. $ Thus, repeating this procedure, we have $v\in Y$, and the proof is complete. The assertion follows from Proposition \[sec:thm1\] and the fact that the set of the charges of $W(R,S;\lambda)$ agrees with $L(R,S)$. Acknowledgments {#acknowledgments .unnumbered} --------------- The author wishes to express his thanks to his advisor, Professor A. Matsuo for helpful advice and kind encouragement. He also wishes to express his thanks to H. Yamauchi, H. Shimakura and M. Okumura for helpful conversations. Note added: After finishing this work, we learned of a recent related work by Kaneko, Nagatomo and Sakai [@Kan4], where the Kaneko-Zagier equations and the Mathur-Mukhi-Sen classification are studied in detail. [100]{} Andrews, G. E.: The theory of partitions. Encyclopedia of Mathematics and Its Applications, Vol. 2. Addison-Wesley Publishing Co., Reading, Mass.-London-Amsterdam (1976) Andrews, G. E.: $q$-Series: Their development and application in analysis, number theory, combinatorics, physics, and computer algebra, Regional Conf. Ser. in Math., no. [**66**]{} (1986) Ardonne, E, Kedem, R., Stone, M.: Fermionic characters of arbitrary highest-weight integrable $sl_{r+1}$-modules, Comm.  Math. Phys. [**264**]{} 427-464 (2006) Bannai, E., Koike, M., Munemasa, A., Sekiguti, J.: Klein’s icosahedral equation and modular forms, preprint (1999) Baxter, R. J.: Exactly Solved Models in Statistical Mechanics, Academic Press, London and New York, (1982) Borcherds, R. E.: Vertex algebras, Kac-Moody algebras, and the Monster. Proc. Nat. Acad. Sci. U.S.A. [**83**]{}, no. 19, 3068-3071 (1986) Calinescu, C., Lepowsky, J., Milas, A.: Vertex-algebraic structure of the principal subspaces of level one modules for the untwisted affine Lie algebras of type A, D, E. J. Algebra [**323**]{}, no. 1, 167-192 (2010) Cohen, A. M., de Man, R.: Computational evidence for Deligne’s conjecture regarding exceptional Lie groups, C. R. Acad.  Sci. Paris Sér. I Math. [**322**]{} 427-432 (1996) Capparelli, S., Lepowsky, J., Milas A.: The Rogers-Selberg recursions, the Gordon-Andrews identities and intertwining operators, The Ramanujan Journal [**12**]{} 379-397 (2006) Cook, W. J., Li, H., Misra, K. C.: A recurrence relation for characters of highest weight integrable modules for affine Lie algebras, Comm. in Contemporary Math.  [**9**]{}, no. 2, 121-133 (2007) Deligne, P.: La série exceptionalle de groupes de Lie, C. R. Acad. Sci. Paris Sér. I Math. [**322**]{} 321-326 (1996) Feigin, B., Feigin, E., Jimbo, M., Miwa, T., Mukhin, E.: Principal $\widehat{\mathfrak{sl}_3}$ subspaces and quantum Toda Hamiltonian, arXiv:0707.1635. (2007) Feigin, B. L., Stoyanovsky, A. V.: Quasi-particles models for the representations of Lie algebras and geometry of flag manifold. arXiv:hep-th/9308079. (1993) Georgiev, G. N.: Combinatorial constructions of modules for infinite-dimensional Lie algebras, I. Principal subspace, J.  Pure Appl. Algebra [**112**]{}, 247-286 (1996) Gel’fand, I. M., Zelevinsky, A. V.: Models of representations of classical groups and their hidden symmetries, (Russian) Funkt. Anal. i Prilozhen. 18 (3) 14-31 (1984) Ibukiyama, T.: Modular forms of rational weights and modular varieties, Abh. Math. Sem. Univ. Hamburg, [**70**]{} 315-339 (2000) Kac, V. G.: Infinite dimensional Lie algebras. Third Edition. Cambridge University Press, Cambridge, UK (1990) Kaneko, M.: On modular forms of weight $(6n+1)/5$ satisfying a certain differential equation, Number Theory: Tradition and Modernization 97-102 (2006) Kaneko, M., Koike, M.: On modular forms arising from a differential equations of hypergeometric type. Ramanujan J. [**7**]{} 145-164 (2003) Kaneko, M., Nagatomo, K., Sakai, Y.: Modular forms and second order ordinary differential equations: applications to vertex operator algebras. Lett. Math. Phys. [**103**]{} 439-453 (2013) Kaneko, M., Zagier, D.: Supersingular $j$-invariants, hypergeometric series, and Atkin’s orthogonal polynomials. AMS/IP Stud. Adv. Math. [**7**]{} 97-126 (1998) Kedem, R., Klassen, T. R., McCoy, B. M., Melzer, E.: Fermionic sum representations for conformal field theory characters, Phys. Lett. B vol. [**307**]{} issue 1-2, 68-76 (1993) Landsberg, J. M., Manivel, L.: Triality, exceptional Lie algebras, and Deligne dimension formulas, Adv. Math. [**171**]{} 59-85 (2002) Landsberg, J. M., Manivel, L.: The sextonions and $E_{7\frac{1}{2}}$, Adv. Math. [**201**]{} (1) 143-179 (2006) Matsuo, A.: Norton’s trace formulae for the Griess algebra of a vertex operator algebra with larger symmetry, Commun. Math.  Phys. [**224**]{} 565-591 (2001) Milas, A.: Ramanujan’s “Lost Notebook" and the Virasoro algebra, Comm. Math. Phys. vol. [**251**]{}, no. 3, 567-588 (2004) Milas, A., Penn, M.: Lattice vertex algebras and combinatorial bases: general case and ${{\mathcal}W}$-algebras. New York J. Math. [**18**]{} 621-650 (2012) Mathur, S., Mukhi, S., Sen, A.: On the classification of rational conformal field theories, Phys. lett. B, Vol. [**213**]{}, Issue. 3, 303-308 (1988) Mathur, S., Mukhi, S., Sen, A.: Reconstruction of conformal field theories from modular geometry on the torus, Nucl. Phys. B, Vol. [**318**]{} 483-540 (1989) Primc, M.: Vertex operator construction of standard modules for $A_n^{(1)}$, Pacific J. Math. [**162**]{} 143-187 (1994) Shtepin, V. V.: Intermediate Lie algebras and their finite-dimensional representations. Russian Acad. Sci. Izv. Math. Vol. [**43**]{}, No.3 559-579 (1994) Stoyanovskii, A. V., Feigin, B. L., Functional models for representations of current algebras and semi-infinite Schubert cells. Funct. Anal. Appl. [**28**]{}, no. 1, 55-72 (1994) Tuite, M. P.: Exceptional vertex operator algebras and the Virasoro algebra, Contemp. Math. [**497**]{} 213-225 (2009) Wakimoto, M.: Lectures on infinite dimensional Lie algebras. World Scientific Publishing, Singapore (2001)

ArXiv

--- address: - | Department of Physics and Mathematical Physics and\ Special Research Centre for the Subatomic Structure of Matter\ University of Adelaide\ Adelaide SA 5005, AUSTRALIA\ E-mail: athomas,[email protected] - | Department of Physics, National Taiwan University,\ Taipei, TAIWAN\ E-mail: [email protected] author: - 'A. W. THOMAS, D. B. LEINWEBER' - 'D. H. LU' title: 'NON-PERTURBATIVE CHIRAL CORRECTIONS FOR LATTICE QCD' --- =cmr8 1.5pt \#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{} [Invited talk presented at NEWS 99]{}\ [University of Osaka, March 9-12, 1999]{}\ [ADP-99-18/T360 ]{} Introduction ============ At present the only known way to calculate the properties of QCD directly is through the formulation of lattice gauge theory on a discrete space-time lattice. While the lattice formulation of QCD is well established [@lattice98], there have recently been a number of exciting advances in lattice action improvement which are greatly facilitating the reduction of systematic uncertainties associated with the finite lattice volume and the finite lattice spacing. However, direct simulation of QCD for light current quark masses, near the chiral limit, remains computationally intensive. In particular, present lattice calculations of masses in full QCD are limited to equivalent pion masses of order 500 MeV or more – although there have been recent preliminary results from CP-PACS using improved actions at a mass as low as 300 MeV [@2]. For nucleon magnetic moments the situation is much worse, with the best calculations still being made within quenched QCD for pion masses greater than 600 MeV. This is clearly beyond the range of validity of chiral perturbation theory ($\chi$PT). In view of this situation, the present approach of calculating the properties of QCD using quark masses away from the chiral regime and extrapolating to the physical world is likely to persist for the foreseeable future. It is therefore vital to understand the quark mass dependence of hadronic observables calculated on the lattice and how to connect these calculations with the physical world. A major difficulty in this endeavour is the rapid rise of the pseudoscalar mass for small increases in the quark mass away from the chiral limit – with $m_\pi^2 \propto \bar{m}$, where $\bar{m}$ is the light quark mass. Historically, lattice results were often linearly extrapolated with respect to $m_\pi^2$, particularly in exploratory calculations. More recently the focus has turned to chiral perturbation theory ($\chi$PT), which provides predictions for the leading nonanalytic quark-mass dependence of observables in terms of phenomenological parameters [@golterman94; @labrenz94]. Indeed, it is now relatively standard to use $\chi$PT in the extrapolation of lattice simulation data for hadron masses and decay constants [@lattice98]. On the other hand, earlier attempts to apply $\chi$PT predictions for the quark-mass dependence of baryon magnetic moments failed, as the higher order terms of the chiral expansion quickly dominate the truncated expansion as one moves away from the chiral limit. To one meson loop, $\chi$PT expresses the nucleon magnetic moments as [@jenkins93] $$\mu_N = \mu_0 + c_1 \, m_\pi + c_2 \, m_\pi^2 \log m_\pi + c_3 \, m_\pi^2 + \cdots \, ,$$ where $\mu_0$ and $c_3$ are fitted phenomenologically and $c_1$ and $c_2$ are predicted by $\chi$PT. The $m_\pi^2 \log m_\pi$ term quickly dominates as $m_\pi$ moves away from the chiral limit, making contact with the lattice results impossible. As a result of these early difficulties, lattice QCD results for baryon magnetic moments[@dblOctet; @wilcox92; @dong97] remain predominantly based on linear quark mass (or $m_\pi^2$) extrapolations of the moments expressed in natural magnetons. This approach systematically underestimates the measured moments by 10 to 20%. Finite lattice volume and spacing errors are expected to be some source of systematic error. However, $\chi$PT clearly indicates the linear extrapolation of the simulation results is also suspect. It is therefore imperative to find a method which can bridge the void between the realm of $\chi$PT and lattice simulations. We report such a method, which provides predictions for the quark mass (or $m_\pi^2$) dependence of nucleon magnetic moments well beyond the chiral limit – for more details we refer to Ref. [@Us]. The method is motivated by studies based on the cloudy bag model (CBM), a chiral quark model which preserves the correct leading non-analytic behaviour of chiral perturbation theory while providing what should be a fairly reliable transition to the regime of large pion mass. These studies suggest a Padé approximant which incorporates both the leading nonanalytic structure of $\chi$PT and Dirac-moment mass dependence in the heavy quark-mass regime. We apply this method to the existing lattice data as an illustration of how important such an approach will be to the analysis of future lattice QCD calculations. The Cloudy Bag Model ==================== The linearized CBM Lagrangian, with pseudoscalar pion-quark coupling (to order $1/f_\pi$), is given by [@CBM; @thomas84] $$\begin{aligned} \protect{\cal L} &=& \left[ \overline q (i\gamma^\mu \partial_\mu-m_q)q - B\right]\theta_V - {1\over 2}\overline q q \delta_S \nonumber \\ && + {1\over 2} (\partial_\mu \pi)^2 - {1\over 2} m^2_\pi \pi^2 - {i\over 2f_\pi} \overline q \gamma_5 \tau \cdot \pi q \delta_S, \label{LAG}\end{aligned}$$ where $B$ is the bag constant, $f_\pi$ is the $\pi$ decay constant, $\theta_V$ is a step function (unity inside the bag volume and vanishing outside) and $\delta_S$ is a surface delta function. In a lowest order perturbative treatment of the pion field, the quark wave function is not affected by the pion field and is simply given by the MIT bag model. Our calculation is carried out in the Breit frame with the center-of-mass correction for the bag performed via Peierls-Thouless projection. The detailed formulas for calculating nucleon electromagnetic form factors in the CBM are given in Ref. [@lu98]. In the CBM, a baryon is viewed as a superposition of a bare quark core and bag plus meson states. Both the quark core and the meson cloud contribute to the baryon magnetic moments. These two sources are balanced around a bag radius, $R = 0.7 - 1.1$ fm [@MM]. A large bag radius suppresses the contributions from the pion cloud, and enhances the contribution from the quark core. The CBM reproduces the leading non-analytic behavior of $\chi$PT, which has as its origin the process shown in Fig. 1(c), along with the other contributions in the model. For the $\pi NN$ vertex, we take a phenomenological, monopole form, $u(k)=(\Lambda^2 - \mu^2)/(\Lambda^2 + k^2)$, where $k$ is the loop momentum and $\Lambda$ is a cut-off parameter. As current lattice simulations indicate that $m_\pi^2$ is approximately proportional to $m_q$ over a wide range of quark masses [@cppacs97], we scale the mass of the quark confined in the bag as $m_q= \left(m_\pi/\mu\right)^2 m_q^{(0)}$, with $m_q^{(0)}$ being the current quark mass corresponding to the physical pion mass ($\mu$). In order to obtain a first idea of the behaviour within the model between the lowest mass lattice point and the experimental data, the parameters $R_0$, $\Lambda$ and $m_q^{(0)}$ were tuned to reproduce the experimental moment while accommodating the lattice data. It turns out that the best fits are obtained with $m_q^{(0)}$ in the range 6 to 7 MeV for a bag radius of 0.8fm and $\Lambda$ of order 600-700 MeV, which are all quite satisfactory. Inspection of the results of this calculation, in Fig. \[proton.ps\], show clearly that the pion cloud contribution to the nucleon magnetic moments decreases very quickly, becoming quite small for large quark masses – especially in the range corresponding to the current lattice calculations. In particular, the total pionic correction at the first lattice data point (for $m_\pi$ around 600 MeV) is of the order of only 10-15% of the total. Since the present lattice data is based on a quenched calculation, which gives incorrect chiral contributions, this is quite good news. It suggests that in using this data as the basis for a correct chiral extrapolation to the physical mass quenching should not induce a major error. It is also clear from Fig. \[proton.ps\] that below $m_\pi$ = 600 MeV the behaviour of the magnetic moment is highly non-linear. One clearly needs to account for such behaviour in a reliable manner if we are to make believable extrapolations of the lattice data. Rather than rely on a model dependent extrapolation method, we chose to investigate whether it was possible to find a simple phenomenological form, [*with a sound physical basis*]{}, which could do the job. The successful conclusion to that search is described next. Encapsulating Formula ===================== After considerable effort we found that the following simple Padé approximant, which builds in both the linear behaviour in $m_\pi$ as $m_\pi \rightarrow 0$ (i.e., a square root branch point in $\bar{m}$) and the Dirac moment for large $m_\pi^2$ (i.e., $\mu_N \sim 1/\bar{m}$), was able to reproduce the behaviour found in the CBM calculations: $$\mu_N(m_\pi) = {\mu_N^{(0)} \over 1 + \frac{\alpha}{\mu_N^{(0)}} m_\pi + \beta m_\pi^2} \, . \label{fit}$$ The fit for the proton case is also shown in Fig. 2. Even more remarkable is that the fit parameter $\alpha$ turned out to be quite close to the value required by $\chi$PT (e.g., $\alpha = 4.54$ for the proton, compared with $\alpha = 4.41$ in $\chi$PT). This result encourages us to propose that Eq.(\[fit\]), with $\alpha$ taken from $\chi$PT, should be used as the method of extrapolation in future analyses of lattice data for nucleon magnetic moments – and suitably generalized, for all members of the baryon octet. As an example, we show in Fig.3 the result of two parameter fits to the proton and neutron data. The nucleon magnetic moments at the physical pion mass, obtained from this extrapolation, are $$\mu_p = 2.85(22)\ \mu_N \quad \mbox{and} \quad \mu_n = -1.90(15)\ \mu_N ,$$ which agree surprisingly well with the experimental measurements, 2.793 and $-$1.913 $\mu_N$ respectively. We note that the data required to do a fit of the lattice results in which covariances are taken into account is no longer available. As such, the uncertainties quoted here should be regarded as indicative only. Conclusion ========== We have explored the quark mass dependence of nucleon magnetic moments. Quark masses beyond the region appropriate to chiral perturbation theory have been explored using the cloudy bag model (CBM) which reproduces the leading nonanalytic behaviour of $\chi$PT while modeling the internal structure of the hadron under investigation. We find that the predictions of the CBM are succinctly described by a simple formula which reproduces the leading nonanalytic behavior of $\chi$PT in the limit $m_\pi \to 0$ [*and*]{} provides the anticipated Dirac moment behavior in the limit $m_\pi \to \infty$. As an example we applied this encapsulating formula to the existing lattice data for nucleon magnetic moments, leading to surprisingly accurate predictions in comparison with the observed values. It will be interesting to see how the fit parameters change as finite volume and lattice spacing artifacts are eliminated in future simulations and whether the level of agreement seen in this investigation is maintained or improved. We strongly advocate the use of the Padé approximant given in Eq. (\[fit\]) in future lattice QCD investigations of octet baryon magnetic moments. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported by the Australian Research Council. References {#references .unnumbered} ========== [99]{} , edited by T. DeGrand, C. DeTar, R. Sugar, and D. Toussaint, Boulder, CO, 1998, [Nucl. Phys. B]{} ([Proc. Suppl. ]{}), (1998). S. Aoki et al. (CP-PACS), hep-lat/9902018. D. B. Leinweber and T. D. Cohen, , 2147 (1993), hep-lat/9211058; A. W. Thomas, Australian J. Phys. [**44**]{}, 173 (1991). M. F. L. Golterman, , 1731 (1994). J. N. Labrenz and S. R. Sharpe, ([Proc. Suppl. ]{}) [**34**]{}, 335 (1994). E. Jenkins, M. Luke, A. V. Manohar, and M. J. Savage, B [**302**]{}, 482 (1993), B [**388**]{}, E866 (1996). D. B. Leinweber, R. M. Woloshyn, and T. Draper, , 1659 (1991). W. Wilcox, T. Draper, and K. F. Liu, , 1109 (1992). S. J. Dong, K. F. Liu, and A. G. Williams, 074504 (1998). D. B. Leinweber et al., hep-lat/9810005. S. Théberge, A. W. Thomas and G. A. Miller, D [**22**]{} (1980) 2838; D [**23**]{} (1981) 2106(E). A. W. Thomas, in [*Advances in Nuclear Physics, Vol. 13 (1984) 1*]{}, edited by J. Negele and E. Vogt, New York, 1984, Plenum Press; G. A. Miller, in [*Int. Rev. Nucl. Phys.*]{} [**2**]{} (1984) 90. D. H. Lu, A. W. Thomas, and A. G. Williams, , 2628 (1998). S. Théberge and A. W. Thomas, (1983) 252. S. Aoki et al., , in [*Lattice ’97, [Proc. of XVth International Symposium on Lattice Field Theory]{}*]{}, edited by C. T. H. Davies et al., Edinburgh, UK 1997, [Nucl. Phys. B]{} ([Proc. Suppl. ]{}), (1998), hep-lat/9710056.

ArXiv

--- abstract: 'Recently, Andrews, Chan, Kim and Osburn introduced the even strings and the odd strings in the overpartitions. We show that their conjecture $$A_k (n) \ge B_k (n)$$ holds for large enough positive integers $n$, where $A_k (n)$ (resp. $B_k (n)$) is the number of odd (resp. even) strings along the overpartitions of $n$. We introduce $m$-strings and show how this new combinatorial object is related with another positivity conjecture of Andrews, Chan, Kim, and Osburn. Finally, we confirm that the positivity conjecture is also true for large enough integers.' address: - | School of Liberal Arts\ Seoul National University of Science and Technology\ 232 Gongneung-ro, Nowongu, Seoul,139-743, Republic of Korea - | Center for Applications of Mathematical Principles, National Institute for Mathematical Sciences\ 70 Yuseong-daero 1689-gil, Yuseong-gu, Daejeon 305-811, Republic of Korea - | Department of Mathematics, Sogang University\ 35 Baekbeom-ro, Mapo-gu, Seoul 121-742, Republic of Korea author: - Byungchan Kim - Eunmi Kim - Jeehyeon Seo title: On the number of even and odd strings along the overpartitions of $n$ --- [^1] Introduction ============ A partition of a non-negative integer $n$ is a non-increasing sequence of positive integers whose sum is $n$. For example, The $5$ partitions of 4 are $4$, $3+1$, $2+2$, $2+1+1$, and $1+1+1+1$. To explain Ramanujan’s famous three partition congruences, the partition statistics, the rank (resp. the crank) were introduced by Dyson [@dyson] (resp. Andrews and Garvan [@ag]). Atkin and Garvan [@atg] initiated the study of rank and crank moments to examine the relations between the rank and the crank. For $k \geq 1$, the $k$-th rank moment $N_{k}(n)$ and the $k$-th crank moment $M_{k}(n)$ are given by $$\begin{aligned} N_{k}(n) &:= \sum_{m \in \mathbb{Z}} m^{k} N(m,n) \intertext{and} \label{kcrank} M_{k}(n)&:= \sum_{m \in \mathbb{Z}} m^{k} M(m,n),\end{aligned}$$ where $N(m,n)$ (resp. $M(m,n)$ denotes the number of partitions of $n$ whose rank (resp. crank) is $m$. By the symmetries $N(-m,n)=N(m,n)$ [@dyson] and $M(-m,n)=M(m,n)$ [@ag], we see $N_{k}(n)=M_{k}(n)=0$ for $k$ odd. To define nontrivial odd moments, Andrews, Chan and Kim [@ack] introduced the positive rank and crank moments $$\begin{aligned} {N}_{k}^{+}(n) &:= \sum_{m=1}^{\infty} m^{k} N(m,n) \intertext{and} {M}_{k}^{+}(n) &:= \sum_{m=1}^{\infty} m^{k} M(m,n).\end{aligned}$$ Andrews, Chan and Kim [@ack] showed that $$\label{oddrc} {M}_{k}^{+}(n) > {N}_{k}^{+}(n),$$ for all positive integers $k$ and $n$. Indeed, this extended Garvan’s result [@gar2]: $$\label{rcforpart} M_{2k}(n) > N_{2k}(n),$$ for all positive integers $k$ and $n$ (Note that $M_{2k} (n) = 2 M_{2k}^{+} (n)$ and $N_{2k} (n) = 2 N_{2k}^{+} (n)$). As ${M}_{k}^{+}(n) - {N}_{k}^{+}(n)$ is positive, it is natural to ask what they count. In this direction, Andrews [@Andrews:spt] showed that $$\label{sptmoment} {\operatorname{spt}}(n) = {M}^{+}_2(n) - {N}^{+}_2(n),$$ where ${\operatorname{spt}}(n)$ is the number of the smallest parts in the partitions of $n$. Moreover, Andrews, Chan and Kim introduced a new counting function ${\operatorname{ospt}}(n)$ satisfying $$\label{osptdefn} {\operatorname{ospt}}(n) ={M}^{+}_1(n) - {N}^{+}_1(n),$$ where ${\operatorname{ospt}}(n)$ counts the number of certain strings along the partitions of $n$ (For the precise definition of ${\operatorname{ospt}}(n)$, see [@ack]). More recently, Andrews, Chan, Kim and Osburn [@acko] considered an overpartition analog of positive moments. Recall that an overpartition [@lc] is a partition in which the first occurrence of each distinct number may be overlined. For example, the $14$ overpartitions of $4$ are $$\begin{gathered} 4, \overline{4}, 3+1, \overline{3} + 1, 3 + \overline{1}, \overline{3} + \overline{1}, 2+2, \overline{2} + 2, 2+1+1, \overline{2} + 1 + 1, 2+ \overline{1} + 1, \\ \overline{2} + \overline{1} + 1, 1+1+1+1, \;\text{and}\;\;\overline{1} + 1 + 1 +1. \end{gathered}$$ Let $\overline{N}(n,m)$ denote the number of overpartitions of $n$ whose rank is $m$ and $\overline{M}(n,m)$ denote the number of overpartitions of $n$ whose (first residual) crank is $m$ [@blo]. As an analog of ordinary positive moments, Andrews, Chan, Kim and Osburn [@acko] defined positive rank and crank moments for overpartitions: $$\begin{aligned} \overline{N}_{k}^{+}(n) &:= \sum_{m=1}^{\infty} m^{k} \overline{N}(m,n) \intertext{and} \overline{M}_{k}^{+}(n) &:= \sum_{m=1}^{\infty} m^{k} \overline{M}(m,n).\end{aligned}$$ As an analog of ${\operatorname{spt}}(n)$ and ${\operatorname{ospt}}(n)$, they defined $${\overline{\operatorname{ospt}}}(n) := \overline{M}_{1}^{+}(n) - \overline{N}_{1}^{+}(n).$$ To explain what ${\overline{\operatorname{ospt}}}(n)$ counts, they introduced even and odd strings, which are our main objects. We define an odd string starting from $2k-1$ in an overpartition as 1. $2k-1, 2k, \ldots, 2k+2\ell-3$ appears at least once, i.e. there are $2\ell -1$ consecutive part sizes starting from $2k-1$. 2. There is no other part of size $2\ell^2 - \ell$ and there is no part of size $4\ell + 2k -2$. Similarly, we define an even sting starting from $2k$ in an overpartition as 1. $2k-1, 2k, \ldots, 2k+2\ell-2$ appears at least once, i.e. there are $2\ell$ consecutive part sizes starting from $2k-1$. 2. There is no other part of size $2\ell^2 + \ell$ and there is no part of size $4\ell + 2k$. Let $A_{k} (n)$ be the number of odd strings starting from $2k-1$ along the overpartitions of $n$, and let $B_{k} (n)$ be the number of even strings starting from $2k-1$ along the overpartitions of $n$. Then, Andrews, Chan, Kim and Osburn showed that $${\overline{\operatorname{ospt}}}(n) = \sum_{k=1}^{\lfloor(n+1)/2\rfloor} ( A_{k}(n) - B_{k} (n) )$$ and conjectured that $A_k (n) \ge B_k (n)$. Our main result is that this conjecture is true for large enough integers. \[mainthm\] For all positive integers $k$, $$A_k (n) > B_k (n)$$ for all large enough integers $n$. We are going to prove Theorem \[mainthm\] by employing the circle method. Due to the linear sum of partial theta functions, the generating function for $A_k (n) - B_k (n)$ is not modular. Therefore, we cannot use modular transformation formula to apply the circle method. However, by analyzing the behavior of generating function near $q=1$ and away from $q=1$, we can obtain a desirable asymptotic formula for the difference $A_k (n) - B_k (n)$. As the difference grows exponentially, we deduce the main theorem. This approach has been very successful to get asymptotic formula for the coefficients of $q$-expansions involving partial theta functions, and a similar approach was used in [@bm2]. Before stating the second result, we need to introduce some notations. In [@acko Lemma 2.2], Andrews, Chan, Kim, and Osburn introduced the function $$\label{hqdefn} h(q) := \sum_{n=1}^{\infty} \frac{(-1)^{n+1} q^{n(n+1)/2}}{1-q^n} = \sum_{j=1}^{\infty} q^{j^2} (1+2q^{j} + 2 q^{2j} + \cdots + 2 q^{j^2 -j} + q^{j^2} ),$$ and conjectured that $$\frac{1}{(q)_{\infty}} ( h(q) - m h(q^m) )$$ has positive coefficients, where we use the standard $q$-series notation, $(a)_{\infty} = (a;q)_{\infty} = \prod_{n=1}^{\infty} (1-aq^{n-1} )$.. Here we define a new combinatorial object to give a combinatorial interpretation for the coefficients of $ \frac{1}{(q)_{\infty}} h( q^m )$. We define $m$-string in an ordinary partition as the parts consisting of $m(1+k)$, $m(3+k)$, $\ldots$, $m(2j-1+k)$ with a positive integer $j$ and a nonnegative integer $k \le j$. We also define a weight of $m$-string as 1 if $k=0$ or $j$, and 2, otherwise. Then, we define $C_m (n)$ as the weighted sum of the number of $m$-strings along the partitions of $n$, i.e. $$C_m (n) = \sum_{{\lambda}\vdash n} \sum_{\substack{\text{$\pi$ is a} \\ \text{$m$-string of ${\lambda}$}}} \operatorname{wt} (\pi).$$ It is straightforward to see that $$\frac{1}{(q)_{\infty}} h(q^m) = \sum_{n \ge 1} C_m (n) q^n.$$ Therefore, the second conjecture of Andrews, Chan, Kim, and Osburn [@acko] can be translated as there are more (weighted count of) $1$-strings than $m$ times of (weighted count of) $m$-strings along the partitions of $n$. By employing the similar argument before, we can prove that the above conjecture is also asymptotically true. \[second main\] For all integer $m\geq 2$, $$\frac{1}{(q)_{\infty}}(h(q)-mh(q^m))$$ has positive power series coefficients for all large enough integers $n$, i.e. $$C_1 (n) > m C_m (n)$$ holds for large enough integers $n$. \[rem2\] Despite the differences are positive for large integers $n$, the proof of Theorem \[second main\] reveals that, for any positive integer $m \ge 2$, $$C_1 (n) \sim m C_m (n)$$ as $n$ goes to the infinity. \[rem1\] For all integer $m\geq 2$, $$\frac{(-q)_{\infty}}{(q)_{\infty}}(h(q)-mh(q^m))$$ has also positive power series coefficients for all large enough integers $n$. Actually, by comparing the generating function of positive moments in [@acko], we see that $$\overline{C}_1 (n) - 2 \overline{C}_2 ( n) = \overline{M}_{1}^{+} (n) - \overline{N}_{1}^{+} (n) = {\overline{\operatorname{ospt}}}(n),$$ where $\overline{C}_{m}(n)$ be the weighted count of the number of $m$-strings along the overpartitions of $n$. Then, by the exactly same method for the proof of Theorem \[second main\], we find $${\overline{\operatorname{ospt}}}(n) \sim \frac{1}{64 n} e^{\pi \sqrt{n}} \sim \frac{1}{8} \overline{p} (n).$$ We also note that $\overline{C}_1 (n) \sim m \overline{C}_{m} (n)$ for positive integers $m \ge 2$. Proofs of Results ================= For a positive integer $a$ and an integer $b$ with $a+b> 0$, we define $$f_{j,a,b}(\tau ):= \sum_{n=1}^{\infty} (-1)^n n^{-j} q^{(an^2+bn)/2}.$$ Here and throughout the paper, we set $q=e^{2\pi i \tau}$ with $\tau = x+iy$. From [@acko], we write the generating function $F_k (q)$ as $$F_k(q) = \sum_{n=1}^{\infty} (A_k(n)-B_k(n))q^n = \frac{(-q)_{\infty}}{(q)_{\infty}}\left[2 f_{0,2,(4k-2)}( \tau )- f_{0,1,(4k-1)}(\tau)- f_{0,1,(4k-3)}(\tau)\right].$$ We begin with investigate asymptotic behavior of partial theta function. \[f\_0\] Assume $|x|\leq y$. As $y \rightarrow 0+$, $$f_{0,a,b}(\tau)=-\frac{1}{2}+\frac{b}{8}(-2 \pi i \tau) + \frac{ab}{16}(-2 \pi i \tau)^2 + \mathcal{O}(y^3)$$ If we separate even and odd terms, then we obtain $$f_{a,b}(\tau) = e^{-\frac{b^2}{4a}\pi i \tau} g_{a,b}(\tau),$$ where $$g_{a,b}(\tau)=\sum_{n=0}^{\infty} \left[ e^{4\pi a \left(n+1+\frac{b}{4a}\right)^2 i \tau} - e^{4\pi a \left(n+\frac{1}{2}+\frac{b}{4a}\right)^2 i \tau} \right].$$ The real part of $g_{a,b}(\tau)$ can be written as $$\text{Re} \left(g_{a,b}(\tau)\right) = \sum_{n=0}^{\infty} \left[u_{\frac{x}{y}}\left( \left(n+1+\frac{b}{4a}\right)\sqrt{y} \right)-u_{\frac{x}{y}}\left( \left(n+\frac{1}{2}+\frac{b}{4a}\right)\sqrt{y} \right)\right],$$ where $$u_s(t)=e^{-4 \pi a t^2} \cos(4\pi a s t^2)=1 - 4 \pi a t^2 +\mathcal{O}(t^6) ~~\text{as}~~ t \rightarrow 0+.$$ By Zagier’s result on asymptotic expansions for series (the first generalization of Proposition 3 in [@z] with a correction on the sign), for $\dfrac{b}{4a}+\dfrac{1}{2}>0$ (this is the case as $a>0$, $a+b>0$), $$\begin{aligned} &\text{Re} \left(g_{a,b}(\tau)\right)\\ &= \left[ \frac{I_u}{\sqrt{y}}-B_1\left(1+\frac{b}{4a}\right)-(- 4 \pi a)\frac{B_3\left(1+\frac{b}{4a}\right)}{3} y- 8 \pi^2 a^2 \frac{B_5\left(1+\frac{b}{4a}\right)}{5} (y^2-x^2)+\mathcal{O}_{\frac{x}{y}}(y^3) \right]\\ & \quad -\left[ \frac{I_u}{\sqrt{y}}-B_1\left(\frac{1}{2}+\frac{b}{4a}\right)-(- 4 \pi a)\frac{B_3\left(\frac{1}{2}+\frac{b}{4a}\right)}{3} y- 8 \pi^2 a^2 \frac{B_5\left(\frac{1}{2}+\frac{b}{4a}\right)}{5} (y^2-x^2) + \mathcal{O}_{\frac{x}{y}}(y^3) \right] \end{aligned}$$ where $|I_u|=\left| \int_{0}^{\infty} u_s(t) \, dx \right| < \infty$ from the assumption $|x|\leq y$. Similarly, $$\begin{aligned} \text{Im} \left(g_{a,b}(\tau)\right) =& \left[ \frac{I_v}{\sqrt{y}}-4 \pi a\frac{B_3\left(1+\frac{b}{4a}\right)}{3} x- (-16 \pi^2 a^2) \frac{B_5\left(1+\frac{b}{4a}\right)}{5} xy + \mathcal{O}_{\frac{x}{y}}(y^3) \right]\\ & -\left[ \frac{I_v}{\sqrt{y}}-4 \pi a\frac{B_3\left(\frac{1}{2}+\frac{b}{4a}\right)}{3} x - (-16 \pi^2 a^2) \frac{B_5\left(\frac{1}{2}+\frac{b}{4a}\right)}{5} xy + \mathcal{O}_{\frac{x}{y}}(y^3) \right] \end{aligned}$$ where $v_s(t)=e^{-4 \pi a t^2} \sin(4\pi a s t^2)$ and $|I_v|=\left| \int_{0}^{\infty} v_s(t) \, dx \right| < \infty$. The real and imaginary parts together will give us $$g_{a,b}(\tau)=-\frac{1}{2}-\frac{b(2a+b)}{8a} (\pi i \tau) +\frac{b(8a^3-4ab^2-b^3)}{64a^2}(\pi i \tau)^2 + \mathcal{O}(y^3).$$ By considering the Taylor expansion of $e^{-\frac{b^2}{4a}\pi i \tau}$, we get the claimed expansion. Now we estimate $F_k(q)$ near $q=1$. \[Sum\_near1\] Assume $y=\dfrac{1}{4 \sqrt{N}}$ and $|x|\leq y$. As $N \rightarrow \infty$, $$F_k(q)=\frac{2k-1}{8\sqrt{\pi}}(-2 \pi i \tau)^{5/2} e^{\frac{\pi i}{8 \tau}} + \mathcal{O}\left(N^{-7/4}e^{\frac{\pi}{2}\sqrt{N}}\right).$$ From Lemma \[f\_0\], we conclude that $$2 f_{0,2,(4k-2)}(\tau)- f_{0,1,(4k-1)}(\tau)- f_{0,1,(4k-3)}(\tau) = \frac{2k-1}{4}(-2 \pi i \tau)^2 + \mathcal{O}(y^3)$$ as $|x| \le y \rightarrow 0+$. By the modular inversion formula for Dedekind eta function, we derive $$\begin{aligned} \frac{(-q)_{\infty}}{(q)_{\infty}}=\sqrt{\frac{-i \tau}{2}}\frac{ \eta\left(-\frac{1}{2\tau}\right)}{\eta^2 \left(-\frac{1}{\tau}\right)} &= \sqrt{\frac{-i \tau}{2}} e^{\frac{\pi i}{8 \tau}}\left(1+ \mathcal{O}\left(e^{-2\pi\sqrt{N}}\right)\right). \end{aligned}$$ As $N \rightarrow \infty$ with $y=\dfrac{1}{4 \sqrt{N}}$ and $|x|\leq y$, we arrive at $$\begin{aligned} F_k(q)&=\frac{(-q)_{\infty}}{(q)_{\infty}}\left[2 f_{0,2,(4k-2)}(\tau)- f_{0,1,(4k-1)}(\tau)- f_{0,1,(4k-3)}(\tau)\right]\\ &=\sqrt{\frac{-i \tau}{2}} e^{\frac{\pi i}{8 \tau}}\left(1+ \mathcal{O}\left(e^{-2\pi\sqrt{N}}\right)\right)\left[\frac{2k-1}{4}(-2 \pi i \tau)^2 + \mathcal{O}(N^{-3/2})\right]\\ &=\left(\frac{2k-1}{4}\right)\frac{e^{\frac{\pi i}{8 \tau}}}{2\sqrt{\pi}}(-2 \pi i \tau)^{5/2} + \mathcal{O}\left(N^{-7/4}e^{\frac{\pi}{2}\sqrt{N}}\right). \end{aligned}$$ The following proposition describes the behavior of $F_k(q)$ away from $q=1$. \[sum away1\] For $y=\dfrac{1}{4\sqrt{N}}$ with $N>0$ and $y \leq |x| \leq 1/2$, $$|F_k(q)| \ll N^{1/2} e^{\frac{\pi}{4}\sqrt{N}}.$$ For $a>0$ and $a+b>0$, we observe $$\left|f_{0,a,b}(\tau)\right| \leq \sum_{n=1}^{\infty}|q|^{n/2} \ll N^{1/2}$$ by bounding each term. As Corollary 3.4 in [@bm2], the inversion formula gives the bound $$\left| \frac{(-q)_{\infty}}{(q)_{\infty}} \right| \ll |\tau|^{1/2} e^{\frac{\pi y}{8(x^2+y^2)}} \ll e^{\frac{\pi}{16 y}}.$$ Now we are ready to apply the circle method. By Cauchy’s integral formula, we see that $$A_k(N)-B_k(N)=\frac{1}{2 \pi i} \int_{\mathcal{C}} \frac{F_k(q)}{q^{N}} \, dq =\int_{-1/2}^{1/2} F_k\left( e^{-\frac{\pi}{2\sqrt{N}}+2\pi i x} \right)e^{\frac{\pi}{2}\sqrt{N}-2\pi i N x} \, dx,$$ where $\mathcal{C}=\left\{|q|=e^{-\frac{\pi}{2\sqrt{N}}}\right\}$. We divide the integral into the two parts: $$\begin{aligned} I'&=\int_{|x|\leq \frac{1}{4\sqrt{N}}} F_k\left( e^{-\frac{\pi}{2\sqrt{N}}+2\pi i x} \right)e^{\frac{\pi}{2}\sqrt{N}-2\pi i N x} \, dx\\ \intertext{and} I''&=\int_{\frac{1}{4\sqrt{N}} \leq |x|\leq \frac{1}{2}} F_k\left( e^{-\frac{\pi}{2\sqrt{N}}+2\pi i x} \right)e^{\frac{\pi}{2}\sqrt{N}-2\pi i N x} \, dx.\end{aligned}$$ The integral $I'$ will contribute the main term of the asymptotic formula and $I''$ will be absorbed into the error term. We first calculate the major contribution, i.e. $I'$. \[majorarc\] As $N \rightarrow \infty$, $$I'= (2k-1) \frac{\pi^3}{64 \sqrt{2}}N^{-7/4}I_{-7/2}\left(\pi\sqrt{N}\right)+\mathcal{O}\left(N^{-9/4}e^{\pi \sqrt{N}}\right),$$ where $I_{s}(z)$ is the modified Bessel function of the first kind. By replacing $u$ by $4\sqrt{N}x$ and Proposition \[Sum\_near1\], we deduce that $$\begin{aligned} I'=& \frac{1}{4\sqrt{N}}\int_{-1}^1 F_k \left( e^{\frac{\pi}{2\sqrt{N}}(-1+iu)} \right) e^{\frac{\pi}{2} \sqrt{N} (1-iu)} \, du\\ =& \frac{1}{4\sqrt{N}}\int_{-1}^1 \left[\frac{2k-1}{8 \sqrt{\pi}} \left( \frac{\pi(1-iu)}{2\sqrt{N}} \right)^{5/2} e^{\frac{\pi}{2}\sqrt{N}\left(\frac{1}{1-iu}+(1-iu)\right)} +\mathcal{O}\left(N^{-7/4}e^{\pi\sqrt{N}}\right)\right] \, du\\ =& (2k-1) \frac{\pi^3}{64 \sqrt{2}} N^{-7/4} P_{5/2} + \mathcal{O}\left(N^{-9/4}e^{\pi\sqrt{N}}\right), \end{aligned}$$ where $$P_s := \frac{1}{2 \pi i} \int_{1-i}^{1+i} v^s e^{\frac{\pi}{2} \sqrt{N} \left(v+\frac{1}{v}\right)} \, dv.$$ The same argument as Lemma 4.2 in [@bm2] implies $$P_s = I_{-s-1}(\pi \sqrt{N})+\mathcal{O}\left(e^{\frac{3\pi}{4}\sqrt{N}}\right)$$ as $N \rightarrow \infty$, where we adopt Lemma 17 in [@w] and a bound for incomplete Gamma function. The contribution from $I''$ is relative small, so that it can be absorbed in the error term in $I'$. \[minorarc\] As $N \rightarrow \infty$, $$|I''| \ll N^{1/2} e^{\frac{3\pi}{4}\sqrt{N}}.$$ By Proposition \[sum away1\], we have $$|I''| \leq \int_{\frac{1}{4\sqrt{N}} \leq |x|\leq \frac{1}{2}} \left| F_k\left( e^{-\frac{\pi}{2\sqrt{N}}+2\pi i x} \right)e^{\frac{\pi \sqrt{N}}{2}-2\pi i N x} \right| \, dx \ll e^{\frac{\pi}{2}\sqrt{N}}N^{1/2}e^{\frac{\pi}{4}\sqrt{N}}.$$ By combining Propositions \[majorarc\] and \[minorarc\], we obtain the asymptotic for $A_k (N) - B_k (N)$. \[asym\] For $k \geq 1$, as $N \rightarrow \infty$, $$A_k(N)-B_k(N)=(2k-1) \frac{\pi^3}{64 \sqrt{2}}N^{-7/4}I_{-7/2}\left(\pi\sqrt{N}\right)+\mathcal{O}\left(N^{-9/4}e^{\pi \sqrt{N}}\right).$$ By the famous asymptotic formulae for the $I$-Bessel function, $I_{s} (z) \sim \frac{e^z}{\sqrt{2\pi z}}$, Theorem \[mainthm\] follows immediately from Theorem \[asym\]. Theorem \[second main\] can be proved similarly, thus we will not give all details. For $m \geq 2$, let $$H_m (q) := \sum_{N=0}^{\infty} (C_1 (N) - m C_m (N))q^N =\frac{1}{(q)_{\infty}}\left( h(q)-m h(q^m) \right).$$ Next two propositions give the asymptotic behavior of $H_m(q)$ near $q=1$ and away from $q=1$. \[H\_near1\] Assume $y=\dfrac{1}{2 \sqrt{6N}}$ and $|x|\leq y$. As $N \rightarrow \infty$, $$H_m(q)=\frac{m-1}{8\sqrt{2 \pi}}(-2 \pi i \tau)^{1/2} e^{\frac{\pi i}{12 \tau}} + \mathcal{O}\left(N^{-3/4}e^{\pi \sqrt{\frac{N}{6}}} \right).$$ From the Mittag-Leffler decomposition [@bm2 eqn. (3.1)], for $w \in \mathbb{C}$, $$\frac{e^{\pi i w}}{1-e^{2 \pi i w}}=\frac{1}{-2\pi i w}+\frac{1}{-2 \pi i} \sum_{k=1}^{\infty} \left( \frac{1}{w-k} + \frac{1}{w+k} \right)$$ holds. Using this, $h(q)$ can be rewritten as $$\begin{aligned} h(q) &= \sum_{n=1}^{\infty} (-1)^{n+1} q^{n^2/2} \left[ \frac{1}{-2\pi i n \tau}+\frac{1}{-2 \pi i} \sum_{k=1}^{\infty} \left( \frac{1}{n \tau-k} + \frac{1}{n \tau+k} \right) \right]\\ &= -\frac{1}{2\pi i \tau}\sum_{n=1}^{\infty} (-1)^{n+1} n^{-1} q^{n^2/2} - \frac{1}{2 \pi i} \sum_{n=1}^{\infty} (-1)^{n+1} q^{n^2/2} \sum_{k=1}^{\infty} \left( \frac{1}{n \tau-k} + \frac{1}{n \tau+k} \right). \end{aligned}$$ Here, the first sum equals $\dfrac{1}{2 \pi i \tau}f_{1,1,0}(\tau)$. For the second sum $S_2$, we first note that $$|n^2\tau^2 -k^2| \geq | \textrm{Re} (n^2\tau^2 -k^2)| \geq k^2+n^2(y^2-x^2) \ge k^2 .$$ Therefore, we find that $$\sum_{k=1}^{ \infty } \left| \frac{1 }{n^2 {\tau}^2-k^2} + \frac{1}{k^2} \right| \le \sum_{k=1}^{\infty} \frac{2n^2 y^2}{k^4} \le 3 n^2 y^2 .$$ Hence, the second sum is $$\begin{aligned} S_2 &= \frac{1}{ 2 \pi i} \sum_{n=1}^{\infty} (-1)^{n+1} q^{n^2/2} \sum_{k=1}^{\infty} 2n\tau \left( \frac{1}{n^2 \tau^2 - k^2} + \frac{1}{k^2} - \frac{1}{k^2} \right) \\ &=\frac{1}{ \pi i} \sum_{n=1}^{\infty} (-1)^{n+1} n\tau q^{n^2/2} \left( - \frac{\pi^2}{6} + T_{n, \tau} \right), \end{aligned}$$ where $|T_{n,\tau}| \le 3 n^2 y^2$. By employing $$\sum_{n=1}^{\infty} n^{3} e^{-\pi n^2 y } = \mathcal{O} ( y^{-2} ),$$ and $f_{-1,1,0}(\tau) = -1/4 + \mathcal{O} (y)$ from the same argument as Lemma \[f\_0\], we can deduce that $$S_2 = \mathcal O (y)$$ as $|x| \le y \rightarrow 0+$. Similarly, we can find $$h(q^m)=\frac{1}{2 \pi i m \tau} f_{1,m,0}(\tau) + \mathcal{O}\left( y \right).$$ Now we need the asymptotes for $f_{1,m,0}(\tau)$ when $|x|\leq y$ and $y \rightarrow 0+$. By the same argument as Lemma \[f\_0\] on the second sum of $$f_{1,m,0}(\tau)= \sum_{n=1}^{\infty} \frac{(-1)^n}{n} + \sum_{n=1}^{\infty} (-1)^n \frac{q^{m n^2/2}-1}{n},$$ we derive $$f_{1,m,0}(\tau)=-\log 2 -\frac{m}{8}(2\pi i \tau)+\mathcal{O}(y^2)$$ as $y \rightarrow 0+$ with $|x|\leq y$. In summary, we have arrived at $$\begin{aligned} h(q)-m h(q^m)&= \left[-\log 2(2\pi i \tau)^{-1}-\frac{1}{8}+\mathcal{O}({y})\right]-\left[-\log 2(2\pi i \tau)^{-1}-\frac{m}{8}+\mathcal{O}({y})\right]\\ &= \frac{m-1}{8} + \mathcal{O}({y}), \end{aligned}$$ which completes the proof by combining with the expansion for $1 / (q)_{\infty} $ [@bm2 eqn. (3.8)]. \[H\_away1\] If $y=\dfrac{1}{2\sqrt{6N}}$ with $N>0$ and $y \leq |x| \leq 1/2$, $$\left|\frac{1}{(q)_{\infty}}\left( h(q)-m h(q) \right)\right| \ll N^{3/4} e^{\frac{\pi}{2}\sqrt{\frac{N}{6}}}.$$ By employing the Hardy-Ramanujan circle method on the integral representation for the coefficients of $H_m(q)$ with the contour $\mathcal{C}=\left\{ |q|=e^{-\frac{\pi}{\sqrt{6 N}}} \right\}$, we find the asymptotic formula with aids of Propositions \[H\_near1\] and \[H\_away1\]. Theorem \[second main\] follows immediately from this asymptotic formula. For $m \geq 2$, as $N \rightarrow \infty$, $$C_1 (N) - m C_m (N)=(m-1) \frac{\pi}{16 \sqrt[4]{54}} N^{-3/4} I_{-3/2} \left( \pi\sqrt{\frac{2N}{3}}\right)+\mathcal{O}\left(N^{-5/4}e^{\pi\sqrt{\frac{2N}{3}}} \right).$$ Concluding Remarks ================== Though we have obtained that $$A_k (n) > B_k (n)$$ holds for large enough $n$’s, it is still desirable to find a combinatorial object which is identical with $A_k (n) - B_k (n)$. The existence of such object will guarantee that $A_k (n) \ge B_k (n)$. Moreover, as ${\overline{\operatorname{ospt}}}(n)$ is the sum of the differences between $A_k (n)$ and $B_k (n)$, such a combinatorial model can be regarded as a refinement of Andrews, Chan, Kim, and Osburn’s result. Numerical data suggest that for all positive integers $n$, $${\overline{\operatorname{ospt}}}(n) > \frac{1}{8} \overline{p} (n)$$ though ${\overline{\operatorname{ospt}}}(n) \sim \frac{1}{8} \overline{p} (n)$. It would be very interesting if one can find a combinatorial or a $q$-theoretic proof for the above inequality. Acknowledgments {#acknowledgments .unnumbered} =============== The authors thank Robert Rhoades for valuable comments on an earlier version of this paper. [15]{} G.E. Andrews, *The number of smallest parts in the partitions of $n$*, J. Reine Angew. Math. **624** (2008), 133–142. G.E. Andrews, S.H. Chan, B. Kim, *The odd moments of ranks and cranks*, J. Combin. Theory Ser. A **120** (2013), no. 1, 77–91. G.E. Andrews, F.G. Garvan, *Dyson’s crank of a partition*, Bull. Amer. Math. Soc. (N.S.) **18** (1988), no. 2, 167–171. G.E. Andrws, S.H. Chan, B. Kim, R. Osburn, *The first positive rank and crank moments for overpartitions*, preprint. A.O.L. Atkin, F.G. Garvan, *Relations between the ranks and cranks of partitions*, Ramanujan J. **7** (2003), no. 1-3, 343–366. K. Bringmann, J. Lovejoy, R. Osburn, *Rank and crank moments for overpartitions*, J. Number Theory **129** (2009), no. 7, 1758–1772. K. Bringmann, K. Mahlburg, *Asymptotic inequalities for positive crank and rank moments*, Trans. Amer. Math. Soc., to appear. F.J. Dyson, *Some guesses in the theory of partitions*, Eureka **8** (1944), 10–15. F. G. Garvan, *Higher order spt-functions*, Adv. Math. **228** (2011), no. 1, 241–265. B. Kim, E. Kim, *On the subpartitions of the ordinary partitions, II*, preprint. J. Lovejoy, S. Corteel, *Overpartitions*, Trans. Amer. Math. Soc. **356** (2004), no. 4, 1623–1635. E. Wright, *Asymptotic Partition Formulae: (II) Weighted Partitions*, Proc. London Math. Soc. (2) **36** (1933), no. 1, 117–141 D. Zagier, *The Mellin transform and other useful analytic techniques*, Appendix to E. Zeidler, Quantum Field Theory I: Basics in Mathematics and Physics. A Bridge Between Mathematicians and Physicists Springer-Verlag, Berlin-Heidelberg-New York (2006), 305–323. [^1]: This research has been supported by TJ Park Science Fellowship of POSCO TJ Park Foundation.

ArXiv

--- abstract: 'We present CN and CH indices and [Ca[ii]{}]{} triplet metallicities for 34 giant stars and chemical abundances for 33 elements in 14 giants in the globular cluster M2. Assuming the program stars are cluster members, our analysis reveals ($i$) an extreme variation in CN and CH line strengths, ($ii$) a metallicity dispersion with a dominant peak at \[Fe/H\] $\approx$ $-$1.7 and smaller peaks at $-$1.5 and $-$1.0, ($iii$) star-to-star abundance variations and correlations for the light elements O, Na, Al and Si and ($iv$) a large (and possibly bimodal) distribution in the abundances of all elements produced mainly via the $s$-process in solar system material. Following @roederer11, we define two groups of stars, “$r+s$” and “$r$-only”, and subtract the average abundances of the latter from the former group to obtain a “$s$-process residual”. This $s$-process residual is remarkably similar to that found in M22 and in M4 despite the range in metallicity covered by these three systems. With recent studies identifying a double subgiant branch in M2 and a dispersion in Sr and Ba abundances, our spectroscopic analysis confirms that this globular cluster has experienced a complex formation history with similarities to M22, NGC 1851 and $\omega$ Centauri.' author: - | David Yong,$^{1}$[^1][^2] Ian U. Roederer,$^2$ Frank Grundahl,$^3$ Gary S. Da Costa,$^1$ Amanda I. Karakas,$^1$ John E. Norris,$^1$ Wako Aoki,$^4$ Cherie K. Fishlock,$^1$ A. F. Marino,$^1$ A. P. Milone$^1$ and Luke J. Shingles$^1$.\ $^{1}$Research School of Astronomy and Astrophysics, Australian National University, Canberra, ACT 2611, Australia\ $^{2}$Department of Astronomy, University of Michigan, 500 Church Street, Ann Arbor, MI 48109, USA\ $^{3}$Stellar Astrophysics Centre, Department of Physics and Astronomy, Aarhus University, Ny Munkegade 120, DK-8000 Aarhus C, Denmark\ $^{4}$National Astronomical Observatory, Mitaka, Tokyo 181-8588, Japan\ title: 'Iron and neutron-capture element abundance variations in the globular cluster M2 (NGC 7089)[^3]' --- \[firstpage\] Stars: abundances $-$ Galaxy: abundances $-$ globular clusters: individual: NGC 7089 INTRODUCTION ============ Photometric studies have revealed complex structure in the colour-magnitude diagrams (CMD) of Galactic globular clusters (e.g., see review by @piotto09). The subgiant branch region is of particular interest because differences in the luminosity of stars at this evolutionary stage require distinct ages and/or chemical compositions. Any globular cluster that exhibits a broadened or split subgiant branch must therefore have experienced a complex, and likely prolonged, chemical enrichment history when compared to globular clusters with a single subgiant branch population. $\omega$ Centauri and M22 (NGC 6656) are two Galactic globular clusters with multiple subgiant branches (e.g., @bedin04; @marino09). These two clusters are also notable for exhibiting a large star-to-star dispersion in the abundance of Fe-peak and neutron-capture elements (e.g., @norris95 [@smith00; @marino09; @marino11; @johnson10; @roederer11]). NGC 1851 is another globular cluster with multiple subgiant branches [@milone08]. Although the difference in metallicity between the two populations, $\Delta$\[Fe/H\] $\approx$ 0.07 dex [@carretta101851], is less pronounced in NGC 1851 compared to $\omega$ Cen and M22, a large star-to-star dispersion in the neutron-capture element abundances is also present (e.g., @yong081851 [@villanova09; @carretta11]). While theoretical studies indicate that multiple population globular clusters could be formed through mergers or that some may be the remnants of dwarf galaxies (e.g., @bekki03 [@carretta10; @bekki11; @bekki12]), understanding the sequence of events that produce multiple population globular clusters remains a major challenge (e.g., @marcolini07 [@dercole08; @dantona10; @conroy11; @herwig12; @vesperini13]). An important step in advancing our knowledge of the formation of multiple population globular clusters is to understand the full range of phenomena and relative frequency present in the Galactic globular cluster system. @piotto12 identified five new Galactic globular clusters with broadened or split subgiant branches based on [$Hubble~Space~Telescope$]{} photometry. Their sample included M2 (NGC 7089), a little-studied cluster. @smith90 measured the strengths of the CN and CH molecular features in a sample of 19 M2 red giants. In addition to the usual bimodal distribution of CN band strengths [@smith87], they noted that two objects are CH stars. CH stars are rare in globular clusters, and at the time of that paper, the only other clusters known to contain CH stars included the apparently normal cluster M55 as well as the peculiar systems M22 and $\omega$ Cen. @smolinski11 studied the CN and CH bands from Sloan Digital Sky Survey spectroscopy in a number of globular clusters including M2. They did not identify any stars with unusually strong CN or CH in this cluster, and all of their program stars lie on the canonical red giant branch (RGB). @lardo12 studied the CN and CH band strengths as well as the C and N abundances in a sample of 35 M2 red giants. They also noted the presence of an additional RGB in the $V$ versus $U-V$ CMD. Both CH stars identified by @smith90 are located on the anomalous RGB (see Figure 14 in @lardo12). Examination of the @grundahl99 [Str[" o]{}mgren]{}photometry also confirms the peculiar nature of the RGB. While @lardo12 did not observe any stars on the anomalous RGB, in a subsequent study they obtained spectra for such stars [@lardo13]. Stars belonging to the two RGBs had distinct C, N, Sr and Ba abundances and @lardo13 argued that M2 has experienced a complex star formation history with similarities to $\omega$ Cen, M22 and NGC 1851. High-resolution spectroscopy and chemical abundance measurements for a larger suite of elements for stars on the canonical and anomalous RGBs of M2 are essential to reveal the true nature of this multiple population globular cluster. The purpose of this paper is to measure CN and CH indices and chemical abundances for a sample of stars in M2 belonging to the canonical and anomalous RGBs. The sample selection and observations are described in Section 2. Section 3 contains the analysis. The results are presented in Section 4. Section 5 includes a discussion on the nature of this cluster. SAMPLE SELECTION AND OBSERVATIONS ================================= The program stars were selected from the $uvby$ [Str[" o]{}mgren]{} photometry by @grundahl99. In Figure \[fig:cmd100\], we present the $u-y$, $v-y$, and $b-y$ CMDs. As noted by @lardo12, we confirm the presence of an additional RGB sequence. Such stars are highlighted in red and aqua in Figure \[fig:cmd100\] and were selected from the $v$ versus $u-y$ CMD (upper panel in Figure \[fig:cmd100\]). We refer to these as anomalous RGB stars. (The reason for using two sets of colours for the anomalous RGB stars will become clear when we present the chemical abundances for these objects: the red symbols are stars with \[Fe/H\] $\approx$ $-$1.5 and the aqua symbols are stars with \[Fe/H\] $\approx$ $-$1.0.) While @lardo12 showed that the two CH stars identified by @smith90 were located on the anomalous giant branch, neither was included in the @grundahl99 photometry. ![Colour-magnitude diagrams for $v$ versus $u-y$ (upper), $y$ versus $v-y$ (middle) and $y$ versus $b-y$ (lower). The black symbols are stars that lie on the canonical RGB, AGB or HB as well as one UV bright object. The red (and aqua) symbols are stars that lie on the anomalous RGB, and were selected from the $v$ versus $u-y$ CMD, upper panel. (The aqua symbols are the unusually metal-rich objects as determined from high-resolution spectroscopy.) Crosses represent stars observed with the AAT. Square symbols (Subaru Telescope) and diamond symbols (Magellan Telescope) represent objects observed at high spectral resolution. \[fig:cmd100\] ](fig1.ps){width=".95\hsize"} Medium-resolution spectroscopic observations -------------------------------------------- We observed candidate M2 members using the AAOmega multi-object spectrograph [@saunders04] on the Anglo-Australian Telescope as part of two separate observing runs. The two CH stars from @smith90 were also observed. The first set of observations, obtained on 2010 September 30, used the 1700B blue and 2000R red gratings. These provide spectral coverage of 3750Å to 4440Å and 5800Å to 6300Å at resolutions $R$ = $\lambda$/$\Delta\lambda$ = 3500 and 8000 for the red and blue arms, respectively. The second set of observations, from 2011 November 1, employed the 580V blue and 1700D red gratings. These gratings provide a wavelength coverage from 3750Å to 5500Å at a resolution of 1300 in the blue arm, and 8350Å to 8825Å at $R$ = 10000 in the red arm. In each case the cluster observations were obtained together with flat-field and arc lamp calibration exposures. Data reduction to wavelength-calibrated sky-subtracted individual stellar spectra was accomplished using the standard [2dfdr]{}[^4] software. In generating the fibre configurations for the observations, high priority was given to the stars that lie on the anomalous giant branch. Although the same input catalogue was used for both sets of observations, the combination of different field plates and different available fibre numbers meant that the two sets of stars observed are not identical. The first set contains 22 stars, the second 23 with 11 in common for a total sample of 34 candidate M2 members. The stars observed are listed in Table \[tab:aao\]. We note that 14 stars belong to the anomalous RGB and the two CH stars from @smith90 may also be regarded as anomalous RGB objects [@lardo12; @lardo13]. Based on their CMD location, we offer some comments on a handful of the stars observed with AAOmega. Star NR 82, if a cluster member, would be classified as a asymptotic giant branch (AGB) star rather than a RGB star. Star NR 184, if a cluster member, would be a UV-bright object lying more than a magnitude above the cluster blue horizontal branch. Similarly, star NR 648 has the photometric colours and magnitude of a cluster blue horizontal branch star. However, the AAOmega spectra (from the second data set) are that of a cluster-like red giant star. We have no straightforward explanation for this anomaly. Star NR 707, again if it is a cluster member, has a magnitude and colour that would suggest it is a red horizontal branch (RHB) star. M2, however, is not normally considered to have a RHB population given its dominant blue HB morphology (e.g., @lardo12). We have not considered these stars any further in the analysis. [@llcccccrrrr@]{} Name1[^5] & Name2 & RA. 2000 & Dec. 2000 & P(%)[^6] & Flag[^7] & $V$ & RV ([km s$^{-1}$]{}) & $\sigma$RV ([km s$^{-1}$]{}) & $S(3839)$ & $m_{\rm CH}$\ (1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) & (11)\ \ HI-240 & AIII-43 & 21 33 10.70 & $-$00 51 09.67 & 99 & …& 14.25 & 1.6 & 2.0 & 1.183 & 0.082\ HI-451 & …& 21 33 39.11 & $-$00 49 30.18 & …& …& 15.86 & 2.1 & 4.6 & 0.574 & 0.150\ NR 38 & …& 21 33 28.91 & $-$00 50 00.94 & …& 1 & 13.60 & 1.1 & 2.2 & 0.520 & 0.045\ NR 76 & HI-104,AII-30 & 21 33 17.91 & $-$00 48 19.82 & 99 & …& 13.85 & $-$1.5 & 4.9 & 0.291 & 0.049\ NR 81 & …& 21 33 27.08 & $-$00 48 19.41 & …& 1 & 13.81 & $-$16.8 & 2.2 & 0.350 & 0.157\ NR 124 & …& 21 33 27.81 & $-$00 47 30.43 & …& …& 14.21 & 4.3 & 2.3 & 0.421 & 0.009\ NR 132 & …& 21 33 23.10 & $-$00 48 11.53 & …& 1 & 14.20 & 0.8 & 2.5 & 0.245 & 0.023\ NR 184[^8] & CR57 & 21 33 24.94 & $-$00 50 41.42 & 99 & …& 15.24 & 0.0 & 7.8 & 0.002 & $-$0.113\ NR 216 & CR19 & 21 33 31.48 & $-$00 49 06.33 & 99 & 1 & 14.82 & $-$12.1 & 2.7 & 0.652 & 0.038\ NR 225 & HI-586,AI-58 & 21 33 29.27 & $-$00 45 55.49 & 99 & …& 14.82 & $-$10.6 & 5.2 & 0.492 & 0.025\ NR 301 & AIV-37 & 21 33 32.80 & $-$00 50 27.06 & 99 & 1 & 14.93 & $-$6.3 & 3.2 & 0.693 & 0.098\ NR 358 & HI-521,AI-79 & 21 33 34.05 & $-$00 47 32.10 & 99 & 1 & 15.06 & $-$16.0 & 4.2 & 1.459 & 0.043\ NR 378 & AI-50 & 21 33 30.32 & $-$00 47 24.54 & 99 & 1 & 15.22 & 3.2 & 5.2 & 0.363 & 0.086\ NR 386 & …& 21 33 26.18 & $-$00 49 21.35 & …& 1 & 15.70 & $-$3.4 & 3.7 & 0.057 & $-$0.061\ NR 388 & …& 21 33 27.16 & $-$00 50 25.43 & …& …& 15.23 & $-$6.0 & 6.0 & 0.293 & $-$0.022\ NR 417 & CR76 & 21 33 23.48 & $-$00 48 46.57 & 99 & …& 15.39 & $-$2.9 & 3.3 & 0.212 & 0.017\ NR 721 & …& 21 33 24.32 & $-$00 49 41.46 & …& 1 & 15.79 & $-$8.8 & 1.8 & 0.309 & 0.002\ NR 811 & …& 21 33 22.85 & $-$00 50 34.00 & …& …& 15.74 & $-$9.3 & 5.2 & 0.388 & 0.013\ NR 847 & …& 21 33 23.45 & $-$00 46 24.34 & …& …& 15.79 & $-$16.1 & 2.8 & 0.430 & $-$0.008\ NR 915 & …& 21 33 35.36 & $-$00 49 57.45 & …& …& 15.83 & $-$5.2 & 6.9 & 0.319 & 0.037\ NR 1178 & …& 21 33 31.64 & $-$00 49 59.80 & …& 1 & 16.06 & 3.0 & 5.1 & 0.343 & 0.059\ NR 1204 & AIII-26 & 21 33 20.08 & $-$00 50 13.76 & 99 & 1 & 15.82 & $-$7.9 & 3.3 & 0.561 & 0.154\ \ HI-240 & AIII-43 & 21 33 10.70 & $-$00 51 09.67 & 99 & …& 14.25 & 2.9 & 2.5 & 1.069 & 0.085\ HI-451 & …& 21 33 39.11 & $-$00 49 30.18 & …& …& 15.86 & 4.3 & 2.7 & 0.609 & 0.199\ NR 47 & CR12 & 21 33 28.52 & $-$00 48 43.92 & 99 & 1 & 13.70 & 5.1 & 1.5 & 0.483 & 0.058\ NR 76 & HI-104,AII-30 & 21 33 17.91 & $-$00 48 19.82 & 99 & …& 13.85 & $-$1.5 & 1.9 & 0.283 & 0.045\ NR 82$^{\ref{note1}}$ & CR190 & 21 33 33.63 & $-$00 50 29.50 & 99 & …& 13.91 & $-$3.1 & 1.7 & 0.424 & $-$0.015\ NR 99 & AIII-86 & 21 33 23.59 & $-$00 50 41.07 & 99 & …& 13.69 & $-$1.3 & 2.1 & 0.250 & 0.018\ NR 124 & …& 21 33 27.81 & $-$00 47 30.43 & …& …& 14.21 & 4.4 & 1.9 & 0.382 & 0.017\ NR 132 & …& 21 33 23.10 & $-$00 48 11.53 & …& 1 & 14.20 & 1.2 & 2.2 & 0.337 & 0.053\ NR 207 & …& 21 33 27.48 & $-$00 49 51.35 & …& 1 & 14.89 & $-$1.1 & 2.0 & 0.219 & 0.004\ NR 225 & HI-586,AI-58 & 21 33 29.27 & $-$00 45 55.49 & 99 & …& 14.82 & $-$10.8 & 2.0 & 0.430 & 0.049\ NR 254 & …& 21 33 29.37 & $-$00 49 42.84 & …& 1 & 15.05 & 3.4 & 2.0 & 0.255 & 0.027\ NR 299 & AI-22 & 21 33 35.32 & $-$00 49 22.13 & 99 & …& 14.94 & 0.3 & 1.9 & 0.307 & 0.120\ NR 358 & HI-521,AI-79 & 21 33 34.05 & $-$00 47 32.10 & 99 & 1 & 15.06 & $-$16.7 & 1.2 & 1.328 & 0.114\ NR 375 & …& 21 33 30.62 & $-$00 50 08.33 & …& …& 15.20 & 2.1 & 1.7 & 0.322 & 0.022\ NR 378 & AI-50 & 21 33 30.32 & $-$00 47 24.54 & 99 & 1 & 15.22 & $-$2.3 & 2.0 & 0.432 & 0.081\ NR 403 & CR58 & 21 33 25.64 & $-$00 50 43.12 & 99 & …& 15.20 & $-$15.3 & 1.8 & 0.331 & 0.015\ NR 648$^{\ref{note1}}$ & …& 21 33 26.32 & $-$00 49 10.58 & …& …& 15.99 & $-$1.6 & 1.9 & 0.194 & $-$0.049\ NR 707$^{\ref{note1}}$ & …& 21 33 23.02 & $-$00 48 56.69 & …& …& 15.81 & 0.3 & 1.9 & 0.095 & $-$0.053\ NR 801 & …& 21 33 27.46 & $-$00 46 53.10 & …& …& 15.79 & $-$2.3 & 2.2 & 0.305 & $-$0.004\ NR 847 & …& 21 33 23.45 & $-$00 46 24.34 & …& …& 15.79 & $-$3.9 & 2.0 & 0.377 & 0.034\ NR 915 & …& 21 33 35.36 & $-$00 49 57.45 & …& …& 15.83 & $-$11.8 & 2.3 & 0.332 & 0.047\ NR 947 & AIII-10 & 21 33 18.75 & $-$00 49 44.09 & 99 & …& 15.79 & $-$6.2 & 1.9 & 0.190 & 0.059\ NR 1204 & AIII-26 & 21 33 20.08 & $-$00 50 13.76 & 99 & 1 & 15.82 & $-$10.6 & 1.7 & 0.449 & 0.123\ High-resolution spectroscopic observations ------------------------------------------ Three stars (NR 76, NR 81 and NR 132) were observed in service mode using the High Dispersion Spectrograph (HDS, @noguchi02) on the Subaru Telescope on 2011 August 3. Six additional stars (NR 47, NR 99, NR 124, NR 207, NR 254 and NR 378) were observed using HDS in classical mode on 2013 July 17. All nine stars were also observed with the AAOmega instrument. For both sets of observations, we used the StdYb setting and the 08 slit which resulted in a wavelength coverage from $\sim$4100Å to $\sim$6800Å at a spectral resolution of $R = 45000$. A telluric standard was also observed. The spectra were reduced using [iraf]{}[^9] adopting a similar approach as in @yong06. Five stars (NR 37, NR 38, NR 58, NR 60 and NR 77) were also observed using the Magellan Inamori Kyocera Echelle (MIKE) spectrograph [@bernstein03] at the Magellan Telescope on 2012 August 26. NR 60 is a likely AGB star based on CMD location. Full wavelength coverage was obtained ($\sim$3400Å to $\sim$9000Å), and we used the 07 slit which provided a spectral resolution of $R = 40000$ in the blue arm and $R = 35000$ in the red arm, as measured from the ThAr lines. The spectra were reduced using the [CarPy]{} pipeline[^10] and independently in [iraf]{} using the [mtools]{} package[^11]. One star, NR 77, had a faint, nearby companion. This object was reduced in two different ways using [iraf]{}. In the first approach, we adopted a conservative aperture placement to try to avoid flux from the faint companion. In the second approach, the flux from both stars was extracted. In the subsequent section, we analyze both sets of spectra independently in order to quantify the contamination from the nearby companion. The program stars are listed in Table \[tab:param\]. We note that of these 14 objects observed at high spectral resolution, six belong to the canonical RGB and eight are anomalous RGB stars. [@llccccccrcccccc@]{} Name1[^12] & Name2 & R.A. & Dec. & P[^13] & Flag[^14] & Run[^15] & $V$ & RV & $\sigma$RV & [$T_{\rm eff}$]{}& [$\log g$]{}& [$\xi_t$]{}& \[m/H\][^16] & \[Fe/H\]\ & & & & (%) & & & & & (K) & (cgs) & ([km s$^{-1}$]{}) & (dex) & (dex)\ (1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) & (11) & (12) & (13) & (14) & (15)\ \ NR 37 & CR78 & 21 33 25.44 & $-$00 48 53.73 & 99 & …& M12 & 13.56 & $-$15.3 & 1.0 & 4250 & 0.70 & 1.77 & $-$1.6 & $-$1.66\ NR 58 & CR30 & 21 33 32.17 & $-$00 50 01.17 & 99 & …& M12 & 13.55 & 11.8 & 1.0 & 4225 & 0.70 & 1.89 & $-$1.6 & $-$1.64\ NR 60[^17] & CR28 & 21 33 32.57 & $-$00 49 45.72 & 99 & …& M12 & 13.55 & $-$7.1 & 1.0 & 4325 & 0.30 & 2.19 & $-$1.7 & $-$1.75\ NR 76 & HI-104 & 21 33 17.91 & $-$00 48 19.82 & 99 & …& S11 & 13.85 & $-$1.3 & 0.6 & 4375 & 0.90 & 1.73 & $-$1.7 & $-$1.69\ NR 99 & AIII-86 & 21 33 23.59 & $-$00 50 41.07 & 99 & …& S13 & 13.69 & $-$1.5 & 0.6 & 4275 & 0.70 & 1.78 & $-$1.6 & $-$1.66\ NR 124 & …& 21 33 27.81 & $-$00 47 30.43 & …& …& S13 & 14.21 & 3.4 & 0.7 & 4425 & 0.85 & 1.81 & $-$1.6 & $-$1.64\ \ NR 38 & …& 21 33 28.91 & $-$00 50 00.94 & …& 1 & M12 & 13.60 & 3.7 & 1.3 & 4175 & 0.60 & 2.12 & $-$1.6 & $-$1.61\ NR 47 & CR12 & 21 33 28.52 & $-$00 48 43.92 & 99 & 1 & S13 & 13.70 & 3.3 & 0.5 & 4050 & 0.65 & 1.77 & $-$1.4 & $-$1.42\ NR 77 & …& 21 33 24.45 & $-$00 48 36.29 & …& 1 & M12 & 13.92 & 6.6 & 1.0 & 4350 & 1.00 & 2.25 & $-$1.5 & $-$1.46\ NR 81 & …& 21 33 27.08 & $-$00 48 19.41 & …& 1 & S11 & 13.81 & $-$22.0 & 0.5 & 4275 & 1.00 & 1.85 & $-$1.6 & $-$1.55\ \ NR 132 & …& 21 33 23.10 & $-$00 48 11.53 & …& 1 & S11 & 14.20 & 0.7 & 0.5 & 4325 & 1.30 & 1.88 & $-$1.0 & $-$0.97\ NR 207 & …& 21 33 27.48 & $-$00 49 51.35 & …& 1 & S13 & 14.89 & $-$2.1 & 0.4 & 4425 & 1.30 & 1.40 & $-$1.1 & $-$1.08\ NR 254 & …& 21 33 29.37 & $-$00 49 42.84 & …& 1 & S13 & 15.05 & 3.2 & 0.5 & 4525 & 1.60 & 1.61 & $-$1.0 & $-$0.97\ NR 378 & AI-50 & 21 33 30.32 & $-$00 47 24.54 & 99 & 1 & S13 & 15.22 & $-$2.9 & 0.5 & 4750 & 1.50 & 1.68 & $-$1.1 & $-$1.08\ In Figures \[fig:spec1s\] to \[fig:spec2m\], we plot regions of the high dispersion spectra for the program stars from both telescope+instrument combinations. These figures demonstrate that there are star-to-star variations in the strengths of Zr and La lines which could be caused by differences in stellar parameters and/or chemical abundance ratios. In Section 3.2, we shall seek to quantify the stellar parameters and chemical abundances. ![A portion of the Subaru HDS spectra for nine program stars. The yellow region highlights [Zr[i]{}]{} lines used in the analysis. There are clear star-to-star differences in the [Zr[i]{}]{} line strengths, and also for [V[i]{}]{} and [Ti[i]{}]{} lines. Black lines represent stars that lie on the canonical RGB. Red lines are stars on the anomalous RGB. The aqua lines are the unusually metal-rich objects on the anomalous RGB. (The colours are consistent with those used in Figure \[fig:cmd100\].) The positions of other atomic lines and the stellar parameters ([$T_{\rm eff}$]{}/[$\log g$]{}/\[Fe/H\]) are included. \[fig:spec1s\] ](fig2.ps){width=".99\hsize"} ![A portion of the Magellan MIKE spectra for five program stars. As in Figure \[fig:spec1s\], Zr lines are highlighted and there are significant star-to-star differences in line strengths. The black spectra denote that those stars lie on the canonical RGB while red spectra represent stars that lie on the anomalous RGB. The positions of other atomic lines and the stellar parameters ([$T_{\rm eff}$]{}/[$\log g$]{}/\[Fe/H\]) are included. \[fig:spec1m\] ](fig3.ps){width=".99\hsize"} ![Same as Figure \[fig:spec1s\] but for a region containing a La line used in the analysis. There are significant star-to-star differences in the line strength of La. The positions of CN lines are marked. \[fig:spec2s\] ](fig4.ps){width=".99\hsize"} ![Same as Figure \[fig:spec2s\] but for the Magellan MIKE spectra. \[fig:spec2m\] ](fig5.ps){width=".99\hsize"} ANALYSIS ======== Radial velocities and line indices from medium-resolution spectra ----------------------------------------------------------------- Radial velocities were measured from the medium-resolution spectra by cross-correlating each program star against HI-240. The radial velocity for HI-240 was determined by measuring the wavelengths of a small set of lines (sodium doublet 5889.951Å and 5895.924Å and the calcium triplet 8498.03Å, 8542.09Å and 8662.14Å). Given the superior spectral resolution in the red arm, we adopted those values as the radial velocities and corrected for the heliocentric motion. We measured the [$S(3839)$]{} and [$m_{\rm CH}$]{} indices in the AAOmega spectra using the definitions given in @smith90. Calcium triplet line strengths were measured via Gaussian line profile fits to the observed data for the two stronger [Ca[ii]{}]{} triplet lines at 8542Å and 8662Å using the technique first described in @AD91. The heliocentric radial velocities and [$S(3839)$]{} and [$m_{\rm CH}$]{} indices are presented in Table \[tab:aao\]. An assessment of the internal errors associated with these measurement can be obtained by consideration of the 11 objects observed on both runs. For the radial velocities, 10 of the 11 objects showed no evidence ($\le$1-$\sigma$, i.e., less than one standard deviation) for radial velocity variation between the two observing runs. One star, NR 847, exhibited evidence for radial velocity variability; $-$16.1 [km s$^{-1}$]{} $\pm$ 2.8 [km s$^{-1}$]{} (2010) versus $-$3.9 [km s$^{-1}$]{}$\pm$ 2.0 [km s$^{-1}$]{} (2011). Excluding NR 847, the average difference in radial velocity for stars observed on both runs is 1.2 [km s$^{-1}$]{} $\pm$ 0.9 [km s$^{-1}$]{} ($\sigma$ = 2.9 [km s$^{-1}$]{}). For the [$S(3839)$]{} and [$m_{\rm CH}$]{} indices, we find mean differences for stars observed on both runs of 0.028 $\pm$ 0.023 ($\sigma$ = 0.075) and 0.018 $\pm$ 0.009 ($\sigma$ = 0.029), respectively. Since for the [Ca[ii]{}]{} triplet spectra only single observations are available, we adopt the uncertainty in the pseudo-equivalent widths which results from the uncertainties in the Gaussian fit parameters for the observed line profiles. An assessment of the systematic errors can be obtained by comparison of our measurements with literature values. For the radial velocities, five of our program stars were also observed by @lardo12 [@lardo13], noting that on average their measurement errors ($\langle\sigma$RV$\rangle$ = 16.5 [km s$^{-1}$]{}) are larger than ours ($\langle\sigma$RV$\rangle$ = 3.0 [km s$^{-1}$]{}). For three of these five stars our radial velocity measurements are in agreement. The two stars with poor agreement are NR 132, $-$16.4 [km s$^{-1}$]{} $\pm$ 7 [km s$^{-1}$]{} [@lardo13] versus 0.9 [km s$^{-1}$]{} $\pm$ 3.4 [km s$^{-1}$]{} (this study), and NR 378, $-$60.3 [km s$^{-1}$]{} $\pm$ 5.8 [km s$^{-1}$]{} [@lardo13] versus 0.5 [km s$^{-1}$]{} $\pm$ 5.6 [km s$^{-1}$]{} (this study). These stars may be spectroscopic binaries. The [$S(3839)$]{} and [$m_{\rm CH}$]{} values are in good agreement with those of @smith90 for the two stars in common. For HI-240, our mean values are [$S(3839)$]{} = 1.126 and [$m_{\rm CH}$]{} = 0.084 and the @smith90 values are 1.111 and 0.067, respectively. For HI-451, our mean values are [$S(3839)$]{} = 0.592 and [$m_{\rm CH}$]{} = 0.175 and the @smith90 values are 0.571 and 0.165, respectively. We note that the differences for [$S(3839)$]{} and [$m_{\rm CH}$]{} between this study and @smith90 are comparable to mean differences for the 11 stars observed on both AAOmega runs. Stellar parameters, chemical abundances and radial velocities from high-resolution spectra ------------------------------------------------------------------------------------------ Equivalent widths (EW) were measured from the high-resolution spectra using routines in [iraf]{} and [daospec]{} [@stetson08]. When using [iraf]{} to measure an EW, every line in every star was visually inspected. In a given star, lines regarded to be blended or poorly fit were excluded, and weak (EW $<$ 5 mÅ) and strong (EW $>$ 130 mÅ) lines were also removed from the analysis. When using [daospec]{} to measure EWs, the continuum was the same as in the [iraf]{} analysis, i.e., [daospec]{} did not re-adjust the continuum level. Additionally, the set of lines measured using [daospec]{} was identical to those already measured, and visually inspected, using [iraf]{}. For the Subaru and Magellan spectra, there was good agreement ($\sigma$ = 1.5 mÅ) between the two sets of EW measurements, and we adopted the [daospec]{} values. For the Magellan spectra, EW measurements could be compared between the [CarPy]{} reduction and the [iraf]{} reduction. Again, there was excellent agreement between the two sets of measurements ($\sigma$ = 1.4 mÅ). The EW measurements and line list are presented in Table \[tab:ew\]. ------------ -------------- ------- ----------- ------- ------- ------- ------- ------- ------- ------- ------- ------------- -- -- -- -- -- -- Wavelength Species[^18] L.E.P $\log gf$ NR 37 NR 38 NR 47 NR 58 NR 60 NR 76 NR 77 NR 81 Source[^19] Å eV mÅ mÅ mÅ mÅ mÅ mÅ mÅ mÅ (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) 6300.31 8.0 0.00 $-$9.75 32.3 53.5 … 43.7 23.3 … … 49.3 B 6363.78 8.0 0.02 $-$10.25 … 26.3 … 16.1 12.0 … 14.2 22.3 A 4751.82 11.0 2.10 $-$2.11 … … … … … … 11.2 … B 4982.83 11.0 2.10 $-$0.91 … … … … … 20.8 48.2 39.0 A 5682.65 11.0 2.10 $-$0.67 51.6 49.7 112.7 37.8 45.5 24.9 … 57.6 B ------------ -------------- ------- ----------- ------- ------- ------- ------- ------- ------- ------- ------- ------------- -- -- -- -- -- -- To determine the stellar parameters, we adopted a traditional spectroscopic approach. We used one dimensional local thermodynamic equilibrium (LTE) model atmospheres with \[$\alpha$/Fe\] = +0.4 from the @castelli03 grid. To produce particular models we used the interpolation software tested in @allende04. Chemical abundances were computed using the LTE stellar line analysis program [moog]{} [@moog; @sobeck11]. The effective temperature, [$T_{\rm eff}$]{}, was adjusted until there was no trend between the abundance from [Fe[i]{}]{} lines and the lower excitation potential (L.E.P). The surface gravity, [$\log g$]{}, was adjusted until the abundance from [Fe[i]{}]{} lines was the same as from [Fe[ii]{}]{}lines. The microturbulent velocity, [$\xi_t$]{}, was established when there was no trend between the abundance from [Fe[i]{}]{} and the reduced equivalent width, EW$_r$ = $\log ($EW$/\lambda)$. Finally, we required that the derived metallicity was within 0.1 dex of the value adopted in the model atmosphere. The final stellar parameters (see Table \[tab:param\]) were obtained when these four conditions were simultaneously satisfied. We note that NR 60, whose CMD location is consistent with being an AGB star, has a surface gravity appropriate for that evolutionary phase. Uncertaintes in the stellar parameters were obtained in the following manner. For [$T_{\rm eff}$]{} and [$\xi_t$]{} we measured the uncertainty in the slope between the abundance from [Fe[i]{}]{} lines and L.E.P. and EW$_r$, respectively. We then adjusted [$T_{\rm eff}$]{} or [$\xi_t$]{} until the slope matched the relevant uncertainty. For [$\log g$]{}, we added the standard error of the mean for [Fe[i]{}]{} and [Fe[ii]{}]{} in quadrature, then adjusted [$\log g$]{} until the difference in abundances from [Fe[i]{}]{}and [Fe[ii]{}]{} was equal to this value. Adopting this approach, we estimate that the internal uncertainties in [$T_{\rm eff}$]{}, [$\log g$]{} and [$\xi_t$]{} are 50 K, 0.2 dex and 0.2 [km s$^{-1}$]{}, respectively, and these are slightly conservative estimates. For [$T_{\rm eff}$]{} and [$\log g$]{}, we can compare the spectroscopic values to photometric values. For [$T_{\rm eff}$]{}, we used the infrared flux method (IRFM) metallicity-dependent colour-temperature relations of @ramirez05 for giant stars. We assumed a reddening $E(B-V)$ = 0.06 as in the @harris96 catalogue[^20]. The values are the weighted mean from the $b-y$, $V-J$, $V-H$ and $V-K$ colours ($JHK$ photometry from 2MASS, @2mass). The surface gravity was determined assuming the photometric [$T_{\rm eff}$]{}, a distance modulus $(m-M)_V$ = 15.5 [@harris96], bolometric corrections from @alonso99 and a mass of 0.8 [M$_{\odot}$]{}. The mean differences (photometric $-$ spectroscopic) in [$T_{\rm eff}$]{} and [$\log g$]{} are $-$13 $\pm$ 26 K ($\sigma$ = 78 K) and +0.08 $\pm$ 0.07 dex ($\sigma$ = 0.20 dex), respectively. These differences are within the uncertainties estimated above. For Ni and lighter elements, chemical abundances were computed using the measured EWs, final model atmospheres and [moog]{}. For Cu, Zn and the neutron-capture elements, abundances were determined via spectrum synthesis (e.g., see Figure \[fig:synthpb\] for the Pb analysis). Lines affected by hyperfine splitting (hfs) and/or isotope shifts (IS) were treated appropriately using the hfs data from @kurucz95 or other sources as noted in Table \[tab:ew\]. We adopted the @asplund09 solar abundances. The chemical abundances are presented in Table \[tab:abun\]. ![A portion of the Magellan MIKE spectra near the 4057Å [Pb[i]{}]{}  line, highlighted in yellow. The black thick line represents the best-fitting synthetic spectra. The red and aqua lines show synthetic spectra with unsatisfactory ratios of \[Pb/Fe\], namely $\Delta$Pb $\pm$ 0.5 dex from the line of best fit. The dotted black line is a synthesis containing no Pb. The values written in the bottom of each panel are [$T_{\rm eff}$]{}/[$\log g$]{}/\[Fe/H\]/\[Pb/Fe\]. \[fig:synthpb\] ](fig6.ps){width=".99\hsize"} [@lcrcrc@]{} Name & A(X) & N$_{\rm lines}$ & s.e.$_{\log\epsilon}$ & \[X/Fe\] & Total Error\ \ NR 37 & 7.48 & 1 & …& 0.44 & 0.23\ NR 38 & 7.71 & 2 & 0.06 & 0.64 & 0.19\ NR 47 & …& …& …& …& …\ NR 58 & 7.62 & 2 & 0.00 & 0.57 & 0.19\ NR 60 & 7.20 & 2 & 0.10 & 0.26 & 0.18\ NR 76 & …& …& …& …& …\ NR 77 & …& …& …& …& …\ NR 81 & 7.83 & 2 & 0.03 & 0.69 & 0.18\ NR 99 & …& …& …& …& …\ NR 124 & …& …& …& …& …\ NR 132 & …& …& …& …& …\ NR 207 & …& …& …& …& …\ NR 254 & …& …& …& …& …\ NR 378 & 7.97 & 1 & …& 0.36 & 0.24\ \ NR 37 & 4.76 & 3 & 0.03 & 0.18 & 0.13\ NR 38 & 4.73 & 3 & 0.06 & 0.11 & 0.13\ NR 47 & 5.44 & 3 & 0.01 & 0.63 & 0.14\ NR 58 & 4.44 & 3 & 0.03 & $-$0.16 & 0.13\ NR 60 & 4.84 & 3 & 0.06 & 0.35 & 0.13\ NR 76 & 4.43 & 3 & 0.06 & $-$0.12 & 0.13\ NR 77 & 5.29 & 4 & 0.10 & 0.52 & 0.12\ NR 81 & 4.84 & 5 & 0.02 & 0.15 & 0.11\ NR 99 & 4.43 & 2 & 0.06 & $-$0.14 & 0.15\ NR 124 & 4.86 & 3 & 0.01 & 0.26 & 0.13\ NR 132 & 5.14 & 5 & 0.04 & $-$0.13 & 0.10\ NR 207 & 5.01 & 4 & 0.04 & $-$0.14 & 0.11\ NR 254 & 5.08 & 4 & 0.03 & $-$0.18 & 0.11\ NR 378 & 4.93 & 2 & 0.03 & $-$0.22 & 0.15\ \ NR 37 & 6.27 & 4 & 0.03 & 0.33 & 0.11\ NR 38 & 6.55 & 4 & 0.06 & 0.56 & 0.11\ NR 47 & …& …& …& …& …\ NR 58 & 6.42 & 5 & 0.04 & 0.47 & 0.10\ NR 60 & 6.27 & 3 & 0.01 & 0.42 & 0.13\ NR 76 & 6.23 & 3 & 0.02 & 0.32 & 0.12\ NR 77 & 6.59 & 4 & 0.05 & 0.46 & 0.11\ NR 81 & 6.40 & 3 & 0.05 & 0.35 & 0.12\ NR 99 & 6.38 & 3 & 0.10 & 0.45 & 0.13\ NR 124 & 6.25 & 2 & 0.03 & 0.29 & 0.15\ NR 132 & 6.86 & 4 & 0.04 & 0.22 & 0.11\ NR 207 & 6.79 & 1 & …& 0.28 & 0.22\ NR 254 & 6.89 & 1 & …& 0.26 & 0.22\ NR 378 & 6.75 & 2 & 0.02 & 0.23 & 0.15\ This table is published in its entirety in the electronic edition of the paper. A portion is shown here for guidance regarding its form and content. To determine the errors in chemical abundances, we repeated the analysis varying the stellar parameters, one at a time, by the relevant uncertainties noted above. Additionally, we also changed the metallicity in the model, \[m/H\], by 0.2 dex. We added these four error terms in quadrature to obtain the systematic uncertainty. We replaced the random error (s.e.$_{\log \epsilon}$) by max(s.e.$_{\log \epsilon}$, 0.20/$\sqrt{N_{\rm lines}}$) where the second term is what would be expected for a set of $N_{\rm lines}$ with a dispersion of 0.20 dex (a conservative value based on the abundance dispersion exhibited by [Fe[i]{}]{} lines). To obtain the total error (presented in Table \[tab:abun\]), we added the systematic and random errors in quadrature. As noted in Section 2.2, star NR 77 had a faint companion and we extracted the spectrum for this star in two different ways. In the first approach, we placed the apertures for each order in such a way as to avoid flux from the faint companion. In the second approach, we extracted the flux from both stars. The stellar parameters and chemical abundances were essentially identical between the two approaches. We present the values from the first approach and are confident that the results for this star are not affected by contamination from the faint companion. For the 14 stars observed at high spectral resolution, radial velocities were obtained from the observed wavelengths of the lines used in the equivalent-width analysis. Heliocentric corrections were applied and the radial velocities are presented in Table \[tab:param\]. For the ten stars observed at high- and medium-resolution, we find an average radial velocity difference (high-resolution $-$ medium-resolution) of $-$1.0 $\pm$ 0.7 [km s$^{-1}$]{} ($\sigma$ = 2.1 [km s$^{-1}$]{}). This agreement gives us additional confidence in our heliocentric radial velocity measurements. RESULTS ======= Cluster membership ------------------ Cluster membership for any given star can be established through a combination of the following criteria: ($i$) evolutionary status, ($ii$) location in CMD, ($iii$) radial velocity, ($iv$) distance from cluster center and ($v$) proper motions. Regarding point ($i$), all stars selected from [Str[" o]{}mgren]{} CMDs have colours and magnitudes consistent with being giant stars at the distance of M2. In particular, all 14 stars observed with the Subaru Telescope or Magellan Telescope are red giants with magnitudes consistent with the distance modulus of M2. Concerning point ($ii$), all stars occupy plausible locations in all CMDs (although we shall revisit this aspect in Section 4.4 taking into account the derived metallicities). Regarding point ($iii$), the heliocentric radial velocity of M2 is $-$5.3 [km s$^{-1}$]{} $\pm$ 2 [km s$^{-1}$]{} and the central velocity dispersion is 8.2 [km s$^{-1}$]{} $\pm$ 0.6 [km s$^{-1}$]{} [@harris96]. While all stars have a radial velocity consistent with cluster membership, the small value means that radial velocity alone cannot confirm cluster membership. Concerning point ($iv$), we note that all stars lie within the tidal radius (2145, @harris96). For point ($v$), proper motions, and membership probabilities based on those measurements, were published by @cudworth87. For the 16 stars with proper motion measurements, we note that all are high probability cluster members, $P$ = 99%. Whether or not the four anomalous RGB stars with \[Fe/H\] $\approx$ $-$1.0 are cluster members obviously affects our conclusions. We remain open to both possibilities, i.e., that these four stars may, or may not, be members. That said, in an upcoming study by Milone et al. (in preparation), recent [$Hubble~Space~Telescope$]{}photometry reveals that the four metal-rich stars appear to lie on a narrow well-defined RGB sequence that can be traced to the subgiant branch and main sequence regions supporting the case for cluster membership. Radial velocity and velocity dispersion --------------------------------------- To determine the radial velocity and velocity dispersion for M2, we took the following approach. We exclude NR 847 as this star exhibits radial velocity variation. For stars with multiple radial velocity measurements, we adopt the weighted mean for a given star. Assuming all stars are cluster members, we find that the heliocentric radial velocity for M2 is $-$3.9 $\pm$ 1.1 [km s$^{-1}$]{}  ($\sigma$ = 7.0 [km s$^{-1}$]{})[^21]. These values are in good agreement with those listed in the @harris96 catalogue. CN and CH indices ----------------- In the upper panel of Figure \[fig:cnch\], we plot the [$S(3839)$]{} index against $V$ mag. In this figure, we include the data from @smith90 and exclude the UV-bright (NR 184), HB (NR 648 and NR 707) and AGB (NR 82) stars. As discussed in Section 3.1, our measurements are on the same scale as @smith90. The middle panel shows the generalised histogram of the [$S(3839)$]{}residuals, $\delta$[$S(3839)$]{}, measured with respect to the same baseline as in @smith90, namely $S_0(3839) = -0.1V + 1.644$. The generalised histogram was produced using a Gaussian kernel with a FWHM of 0.03. We note that while @smith90 identified a particularly CN rich star (HI-240, [$S(3839)$]{} = 1.110), our sample includes an even more extreme example, NR 358 with [$S(3839)$]{} = 1.394. In the following subsection, however, we note that NR 358 (not observed at high resolution) has a CMD location inconsistent with cluster membership given the metallicity of this star assuming no significant age spread in the cluster. In the lower panel of Figure \[fig:cnch\], we plot the [$m_{\rm CH}$]{} index against $V$ mag. Consideration of the measurement errors would indicate a genuine spread in the [$m_{\rm CH}$]{} index within this cluster. In addition to the two CH stars identified by @smith90, there are three stars with [$m_{\rm CH}$]{} $>$ 0.1, NR 81, NR 299 and NR 1204. Given the metallicity of NR 1204, the CMD location is inconsistent with cluster membership (i.e., we use the same argument as for NR 358 above that will be described in the following subsection). We have no reason to suspect non-membership for the other two stars with strong [$m_{\rm CH}$]{}indices, NR 81 and NR 299. There is no obvious anti-correlation between the [$S(3839)$]{} and [$m_{\rm CH}$]{} indices. Indeed, the two CH stars from @smith90 also exhibit large [$S(3839)$]{} indices. [*The first key result is that we confirm the presence of unusually CN and/or CH strong stars in M2.*]{} ![The [$S(3839)$]{} CN index versus $V$ magnitude (upper), the distribution of CN excess $\delta$ [$S(3839)$]{} (middle) and the [$m_{\rm CH}$]{} CH index versus $V$ magnitude (lower). (We exclude stars 82, 184, 648 and 707 since they are not on the RGB.) The program stars are shown as black circles (canonical RGB), red triangles (metal-poor anomalous RGB) and aqua stars (metal-rich anomalous RGB). Filled symbols are proper-motion members according to @cudworth87. The crosses are stars from @smith90, and the two CH objects are indicated by large red crosses. A representative error bar is shown in the top and bottom panels. \[fig:cnch\] ](fig7.ps){width=".99\hsize"} Calcium triplet and high-resolution metallicities ------------------------------------------------- Based on the iron abundances derived from the high dispersion spectra, it is clear that the anomalous RGB stars have higher \[Fe/H\] values than those for the normal RGB stars. In particular, the six normal RGB stars in Table \[tab:param\] have a mean iron abundance of $\langle$\[Fe/H\]$\rangle$ = $-$1.67 $\pm$ 0.02 ($\sigma$ = 0.04). The eight anomalous stars separate into two metallicity groups (and in the following subsection we shall see that the two groups exhibit distinct \[X/Fe\] ratios). The more metal-poor group of anomalous RGB stars includes four objects (NR 38, NR 47, NR 77 and NR 81) and has $\langle$\[Fe/H\]$\rangle$ = $-$1.51 $\pm$ 0.04 ($\sigma$ = 0.09) dex. The more metal-rich group of anomalous RGB stars consists of four objects (NR 132, NR 207, NR 254 and NR 378) and has $\langle$\[Fe/H\]$\rangle$ = $-$1.03 $\pm$ 0.03 ($\sigma$ = 0.06) dex. When defined in this way, each of the three groups of stars (canonical RGB, metal-poor anomalous RGB and metal-rich anomalous RGB) likely have metallicities consistent with a single value, i.e., the dispersion in \[Fe/H\] for a given group can probably be explained entirely by the measurement uncertainties. We now turn to the [Ca[ii]{}]{} triplet spectra to investigate the presence of a metallicity dispersion in this cluster. In Figure \[fig\_CaT\] we plot the sum of the EWs of the two stronger [Ca[ii]{}]{} triplet lines against the magnitude difference from the horizontal branch, $V-V_{HB}$ for the M2 stars observed at this wavelength setting with AAOmega. Here the $y$ magnitudes were assumed to be equivalent to $V$ and the value of $V_{HB}$ was taken from Harris (1996). “Normal” RGB stars are plotted as black circles while the triangle and star symbols show the location of stars from the anomalous RGB. The two CH-stars identified by Smith & Mateo (1990) are shown as red crosses. ![The sum of the EWs of the [Ca[ii]{}]{} triplet lines at 8542Å and 8662Å are plotted against magnitude difference from the horizontal branch $V-V_{HB}$. M2 stars lying on the “normal” RGB are shown as black circles while stars from the “anomalous” RGB are plotted as red triangles or aqua star symbols. The two CH-stars identified by Smith & Mateo (1990) are shown as red crosses. The solid lines are the relations between summed EWs and $V-V_{HB}$ for the standard clusters. In order of increasing summed EW the standard clusters are NGC 7099 (\[Fe/H\] = $-$2.27), NGC 2298 ($-$1.96), NGC 1904 ($-$1.58), NGC 288 ($-$1.32) and 47 Tuc ($-$0.76). The dashed line is a fit of a line with slope $-$0.60 Å/mag to the M2 normal RGB stars. \[fig\_CaT\]](fig8.ps){width="0.99\hsize"} In order to calibrate the line strengths in terms of \[Fe/H\] we have made use of similar observations of red giants in “standard” globular clusters that have well established abundances. The clusters are NGC 7099 (M30), NGC 2298, NGC 1904, NGC 288 and 47 Tuc. The observations for these clusters were obtained with AAOmega during the same observing run as that for the M2 [Ca[ii]{}]{} triplet spectra, using an identical instrumental setup. The standard cluster stars observed were chosen using the photometry lists made publicly available by Peter Stetson at the Canadian Astronomy Data Centre[^22]. A similar analysis to that described here for M2 led to the identification of RGB cluster members from the observations. The [Ca[ii]{}]{} triplet line strengths of these stars were then measured using the same procedure as for the M2 stars described in Section 3.1. The numbers of confirmed RGB cluster members ranged from 8 and 10 in NGC 2298 and NGC 7099 to 33 and 46 in NGC 288 and 47 Tuc. In each cluster the RGB stars covered at least two magnitudes in $V-V_{HB}$ at luminosities exceeding $V-V_{HB}$ $\approx$ 0.0 mag, and we adopted $V_{HB}$ from @harris96. A slope of $-$0.60 $\pm$ 0.01 Å/mag was found to fit consistently each set of cluster data. This value is similar to that found in other [Ca[ii]{}]{} triplet studies: for example, @saviane12 find a value for the slope of $-$0.627 while the original study of @AD91 found a slope of $-$0.619 Å/mag. Adopting W$^\prime$ as the value of the summed EW at $V-V_{HB}$ = 0 with the adopted slope of $-$0.60 Å/mag, and \[Fe/H\] abundances from @Ca09 for the standard clusters, then yields a very well defined linear relationship between W$^\prime$ and \[Fe/H\]: \[Fe/H\] = 0.590 W$^\prime$ $-$ 3.253 dex. The RMS about the fitted relation is only 0.02 dex, indicating excellent consistency between the [Ca[ii]{}]{} triplet line strength measurements for these clusters and the @Ca09 \[Fe/H\] abundances. The relation is valid for the abundances encompassed by the standard clusters, i.e., from \[Fe/H\] $\approx$ $-$2.3 to \[Fe/H\] $\approx$ $-$0.7 dex. Returning now to Figure \[fig\_CaT\], we note that the normal M2 RGB stars cluster tightly around the fitted line of slope $-$0.60 Å/mag, shown as the dashed line. In particular, there is no evidence for any intrinsic dispersion in \[Fe/H\] values from the [Ca[ii]{}]{} triplet line strengths of these stars. The \[Fe/H\] abundance derived from the mean W$^\prime$ value is \[Fe/H\]$_{\rm CaT}$ = $-$1.58 $\pm$ 0.08 dex, where the error includes the RMS deviation about the fitted line for the 11 normal RGB stars and the (minor) calibration uncertainty. This value of \[Fe/H\] is somewhat higher than the value listed in @Ca09 for M2, \[Fe/H\] = $-$1.66 $\pm$ 0.07, and in the latest version of the Harris (1996) catalogue (\[Fe/H\] = $-$1.65). Both of these values stem from the measurement of [Ca[ii]{}]{} triplet line strengths in an integrated spectrum of M2 obtained by @AZ88. The value is also somewhat higher than the mean abundance, $-$1.67 $\pm$ 0.02 (std. error of mean), of the six normal RGB stars observed at high dispersion. Nevertheless, there is good agreement between the \[Fe/H\] values derived from the [Ca[ii]{}]{} line strengths and from high dispersion analysis for the three normal RGB stars in common (NR 76, 99 and 124). For these three stars, the mean difference in \[Fe/H\], in the sense of the high dispersion values minus the [Ca[ii]{}]{} values, is $-$0.03 $\pm$ 0.01 dex ($\sigma$ = 0.02). This consistency also applies to the 5 anomalous RGB stars (NR 47, 132, 207, 254 and 378) in common between the two datasets. Here the mean difference is 0.00 $\pm$ 0.05 dex ($\sigma$ = 0.11) suggesting we can combine the \[Fe/H\] determinations for the anomalous RGB stars into a single sample. There are then 10 anomalous RGB star \[Fe/H\] determinations, eight from the high dispersion analysis, seven from the [Ca[ii]{}]{} triplet spectroscopy with five stars in common. For the latter stars the \[Fe/H\] values have been averaged, weighted by the uncertainties. We assume for the present that all the stars are cluster members. The mean abundance of the anomalous RGB stars is the $\langle$\[Fe/H\]$\rangle$ = $-$1.29 $\pm$ 0.09, considerably more metal-rich than that for the normal RGB stars, and with a substantial dispersion of 0.28 dex. The \[Fe/H\] range shown by the anomalous RGB stars is $\sim$0.8 dex indicating that there is a substantial intrinsic iron abundance spread present. Moreover, the value of the mean abundance and the size of the intrinsic abundance spread do not change significantly even if the sample is restricted to the four anomalous RGB stars with 99% membership probabilities. Further, although the sample is not large, the anomalous RGB stars appear to fall into two distinct metallicity groups, each containing 5 objects. The first, consisting of stars NR 38, 47, 77, 81 and 1204, has a mean abundance of $\langle$\[Fe/H\]$_{\rm CaT}\rangle$ = $-$1.47 $\pm$ 0.05 ($\sigma$ = 0.11). For the four stars in this group with high dispersion spectra, the mean abundance is $\langle$\[Fe/H\]$\rangle$ = $-$1.51 $\pm$ 0.04 ($\sigma$ = 0.09). Similarly, for the second group of stars, NR 132, 207, 254, 358 and 378, the mean abundance is $\langle$\[Fe/H\]$_{\rm CaT}\rangle$ = $-$0.98 $\pm$ 0.06 ($\sigma$ = 0.13) dex and for the four stars with high dispersion spectra, the mean abundance is $\langle$\[Fe/H\]$\rangle$ = $-$1.03 $\pm$ 0.03 ($\sigma$ = 0.06). Within each group the intrinsic abundance dispersion is notably smaller than for the full sample, and these two groups mirror those identified by our high resolution spectroscopic analysis. We note in passing that we have not included the two CH stars in the above discussion. Nevertheless, the [Ca[ii]{}]{} triplet spectra of these two objects appear very similar to those of the other M2 stars observed. The measured line strengths imply abundances of \[Fe/H\]$_{\rm CaT}$ = $-$1.69 $\pm$ 0.11 for HI-240 and \[Fe/H\]$_{\rm CaT}$ = $-$1.29 $\pm$ 0.12 for HI-451. The former is consistent with that for the normal RGB stars as well that of anomalous RGB stars such as NR 38 (\[Fe/H\]$_{\rm CaT}$ = $-$1.61 $\pm$ 0.05). The latter is similar to those for the anomalous RGB stars NR 207 (\[Fe/H\]$_{\rm CaT}$ = $-$1.11 $\pm$ 0.07) and NR 1204 (\[Fe/H\]$_{\rm CaT}$ = $-$1.34 $\pm$ 0.09). In the above discussion we have implicitly assumed that the stars observed are all members of M2, deriving abundances under that assumption. There seems no reason to doubt the membership of any of the stars in the normal RGB samples. There is, however, a consistency check that we can apply to further investigate the membership status of the anomalous RGB stars. The check is as follows: given the reasonable assumption that the age range in M2 is small ($\lesssim$2 Gyr, @piotto12), stars that are M2 members with higher \[Fe/H\] abundances should lie to the red of normal RGB stars at the same magnitude in the CMD by an amount that depends on the excess in \[Fe/H\] above that for the normal RGB stars. Ideally such an investigation would use, for example, an ($I$, $V-I$) CMD to minimise the potential influence of molecular bands on the photometry at bluer wavelengths. However, such photometry is not available for most of the anomalous RGB stars. We have therefore used a ($V$, $B-V$) CMD based on the M2 photometry given in Stetson’s Photometric Standard Star fields available from the Canadian Astronomy Data Centre. The M2 normal RGB is well-defined in this data set. We then plotted the stars observed spectroscopically in the CMD using either Stetson’s photometry where available or by generating $V$ and $(B-V)$ values from the @grundahl99 $y$ and $(b-y)$ photometry. Here we have $y$ = $V$ and $(B-V)$ = 1.64 $(b-y)$ with the latter relation determined from 11 stars in common between Stetson’s photometry list and the stars observed at the [Ca[ii]{}]{} triplet. The RMS deviation about the relation is only 0.009 mag. We then use isochrones for metallicities of \[Fe/H\] = $-$1.65, $-$1.25 and $-$0.85 dex, \[$\alpha$/Fe\] = +0.4 and an age of 13 Gyr from the Dartmouth isochrone set [@Dotter08] to provide an indication of the colour shift expected for the metallicities of the anomalous RGB stars. We adopt values from the current on-line version of the Harris (1996) database for the reddening and distance modulus of M2 and with these parameters the \[Fe/H\] = $-$1.65 theoretical RGB is an acceptable representation of the normal RGB stars in the CMD. Specifically, for each anomalous RGB star, we have interpolated in the isochrones at the $V$ magnitude of the star to determine the $(B-V)$ that corresponds to the \[Fe/H\] value. This colour, and its uncertainty derived from the uncertainty in the \[Fe/H\] value, is then compared with the observed $(B-V)$ value. Stars NR 38, 47, 77 and 81 have predicted colours that agree well with the observed colours on the metal-poor anomalous RGB: the mean difference (observed – predicted) is 0.00 $\pm$ 0.03 with, in each case, the predicted colour lying within 2$\sigma$ of the observed colour. We conclude therefore that all four of these stars are likely to be members of the cluster: one (NR 47) has a 99% membership probability from Cudworth & Rauscher (1987) while the others are not classified. Conversely, with this approach it seems probable that stars NR 132, 358, 378 and 1204 are not members of the cluster. Here the colour differences on the metal-rich anomalous RGB are –0.18 $\pm$ 0.04, –0.29 $\pm$ 0.04, –0.15 $\pm$ 0.03, and –0.08 $\pm$ 0.02, respectively; i.e., in each case the location of the star in the CMD is at least 3.5$\sigma$ bluer than predicted for the star’s metallicity. The observed colours can only be reproduced if the age of the stars is at least 6 Gyr younger than the 13 Gyr assumed, which seems unlikely, although as we have already noted in passing, this argument only considers metallicity and that other elements (He, C, N, O and $\alpha$ elements) can also affect the $B-V$ colour. We note further that our classification contrasts with the fact that three of these stars (NR 358, 378 and 1204) have 99% membership probabilities in the Cudworth & Rauscher (1987) study.[^23] For stars NR 207 and NR 254 the comparison suggests that these stars may also be non-members: both lie in the CMD 0.10 $\pm$ 0.03 mag bluer than the predicted colour. Neither has a classification in the @cudworth87 study. For these stars we will need to rely on the similarity of their chemical abundance distributions with those of the cluster members, or with the non-members, for the membership classification. We conclude therefore that at least some of the anomalous RGB stars, in particular the stars NR 38, NR 47, NR 77 and NR 81 (and perhaps also NR 207 and NR 254) are likely bona-fide members of M2. [*If this is indeed the case then the second key result we find is that M2 joins other clusters like M22 [@dacosta09; @marino09; @marino11], M54 [@carretta10; @saviane12] and NGC 5824 [@dacosta14] in possessing a modest intrinsic \[Fe/H\] range: M2 has member stars with \[Fe/H\] values up to 0.25 dex above that for the majority of cluster members, perhaps up to 0.7 dex depending on the membership status of the metal-rich anomalous RGB stars.*]{} To investigate the likelihood of observing field stars in the vicinity of M2 with stellar parameters ([$T_{\rm eff}$]{}, [$\log g$]{} and \[Fe/H\]) similar to that of the four metal-rich anomalous RGB stars observed at high resolution, we make use of the Trilegal Galactic model [@girardi05]. First, we consider all stars within a one degree square field centered on M2. Secondly, we restricted the sample to lie in the same region in the $v$ versus $u-y$ CMD from which we selected the anomalous RGB stars. We find 17285 such stars in the Trilegal model. Thirdly, of these 17285 stars, we counted the number that satisfied the following constraints: ($i$) $-$25 $\le$ RV $\le$ $+$25 [km s$^{-1}$]{} and ($ii$) $-$1.2 $\le$ \[Fe/H\] $\le$ $-$0.8 dex. And finally, we counted the numbers of stars that lay in a particular region in the [$T_{\rm eff}$]{}-[$\log g$]{} plane, specifically, the area is bounded at the left edge by the line from ([$T_{\rm eff}$]{},[$\log g$]{}) = (5000,2.0) to ([$T_{\rm eff}$]{},[$\log g$]{}) = (4300,0.0), at the right edge by the line from ([$T_{\rm eff}$]{},[$\log g$]{}) = (4400,2.0) to ([$T_{\rm eff}$]{},[$\log g$]{}) = (3700,0.0) both with 0.0 $\le$ [$\log g$]{} $\le$ 2.0. We found 46 stars in the Trilegal model that satisfied all criteria and therefore estimate that given a sample of stars occupying similar locations in the $v$ versus $u-y$ CMD as the program stars, the probability of observing a field star with stellar parameters and a radial velocity consistent with the metal-rich population is roughly 0.3%. We reach similar conclusions when using the Besançon model [@robin03]. Accurate proper-motion and parallax measurements from GAIA will establish cluster membership, or otherwise, for the M2 stars. Given the strong bias towards anomalous RGB stars in the samples selected for observation here, we have little constraint on the form of the iron abundance distribution function other than noting that the normal RGB population is dominant and the anomalous RGB is not prominent (e.g., @lardo12). In this context, the anomalous fainter subgiant branch contains only a small fraction of stars, $\sim$4%, relative to the dominant brighter subgiant branch [@piotto12]. An unbiased sample of RGB stars is needed to constrain the iron abundance distribution and allow comparison with those of other clusters. We now examine the element-to-iron abundance ratios from the high dispersion spectra of the normal and anomalous RGB stars. Chemical Abundance Ratios ------------------------- In Figure \[fig:onamgalsi\], we plot combinations of the light elements (O, Na, Mg, Al and Si) against one another. M2 exhibits star-to-star abundance variations of the light elements along with the usual correlations and anti-correlations between these elements found in globular clusters (e.g., see reviews by @smith87; @kraft94; @gratton04 [@gratton12]). In particular, we note that the observed dispersions in \[X/Fe\] for Na, Al and Si are considerably larger than the average measurement uncertainties indicating genuine abundance spreads. The six canonical RGB stars (black circles in Figure \[fig:onamgalsi\]) clearly exhibit abundance dispersions for Na and Al as well as a correlation between these elements. The four metal-poor anomalous RGB stars (red triangles in Figure \[fig:onamgalsi\]) also exhibit these abundance patterns and this would suggest that they are cluster members. The four metal-rich anomalous RGB stars (aqua star symbols in Figure \[fig:onamgalsi\]) do not exhibit abundance variations for Na and Al. On the other hand, Si does not usually exhibit a star-to-star abundance variation within a given cluster, with a handful of exceptions including NGC 6752 [@yong05] and NGC 4833 [@carretta14]. For O and Mg, there is no compelling evidence for an abundance dispersion within our sample. ![Abundance ratios for combinations of the light elements (O, Na, Mg, Al and Si) for the stars observed at high spectral resolution. The black points are stars on the canonical RGB while the red and aqua points are stars on the anomalous RGB. The aqua points are the unusually metal-rich objects. Open symbols reflect upper limits. The dashed blue line is the linear fit to the data (slope and error are included). The average error ($<\sigma$\[X/Fe\]$>$) and dispersion ($\sigma$) in the x-direction and y-direction are included. \[fig:onamgalsi\] ](fig9.ps){width="0.99\hsize"} Next, in Figure \[fig:sna\], we plot \[X/Fe\] versus \[Na/Fe\] for six neutron-capture species (Y, Zr, La, Nd, Eu and Pb). While there is no evidence for any significant trend between \[X/Fe\] versus \[Na/Fe\], it is clear that the four stars on the anomalous RGB with \[Fe/H\] $\approx$ $-$1.5 exhibit large overabundances of the $s$-process elements with respect to the six stars on the canonical RGB. Such a result is not unexpected given the clear star-to-star line strength differences for neutron-capture elements seen in Figures \[fig:spec1s\] to \[fig:spec2m\]. Confirmation of the presence of a large spread in neutron-capture element abundances can be obtained by noting that the observed dispersion exceeds the average measurement uncertainty. [*The third key result is that we identify an intrinsic abundance dispersion for the neutron-capture elements in M2 thereby verifying and extending the results of @lardo13.*]{} M2 joins the small, but growing, group of globular clusters that exhibit abundance variations for the neutron-capture elements as well as iron abundance dispersions. These clusters include $\omega$ Cen, M22 and NGC 1851 [@norris95; @smith00; @yong081851; @marino09; @marino11; @villanova09; @johnson10; @carretta11; @dorazi11; @roederer11]. Additionally, there are other globular clusters with a dispersion in neutron-capture element abundances, but no obvious iron abundance dispersion including M15 [@sneden97; @sneden00; @otsuki06; @sobeck11; @worley13] and NGC 362 [@carretta13]. ![Same as Figure \[fig:onamgalsi\] but for neutron-capture elements versus \[Na/Fe\]. \[fig:sna\] ](fig10.ps){width="0.99\hsize"} In Figure \[fig:abundfield\] we plot \[X/Fe\] versus \[Fe/H\] for the program stars and field stars from @fulbright00. Here one sees that the six M2 giants on the canonical RGB appear to follow the trends exhibited by field halo stars (although we recognise that there may be systematic abundance differences between this analysis and that of @fulbright00). Similarly, in this figure the four $s$-process rich stars with \[Fe/H\] $\approx$ $-$1.5 have \[X/Fe\] ratios (excluding Y and Zr) consistent with field stars at the same metallicity. For both sets of stars, Na and Al may exhibit higher abundance ratios compared to field stars at the same metallicity. For the four metal-rich stars with \[Fe/H\] $\approx$ $-$1.0, the abundance ratios for all elements included in this figure are consistent with field stars of comparable metallicity. {width="0.65\hsize"} In Figure \[fig:err\], we compare the average measurement errors[^24] with the observed dispersion in \[X/Fe\] ratios for the three groups of stars: (1) the six canonical RGB objects (NR 37, 58, 60, 76, 99 and 124), (2) the four $s$-process rich anomalous RGB stars (NR 38, 47, 77 and 81) and (3) the four metal-rich anomalous RGB stars (NR 132, 207, 254, 378). For the second group, all are likely members based on our analysis in the previous subsection whereas for the third group, their membership is questionable based on the analysis presented in Section 4.4, although [$Hubble~Space~Telescope$]{} photometry suggests that these stars may indeed be members (Milone et al. in preparation). For reasons that will become clearer in the following section, we refer to the three groups as the $r$-process only group (“$r$-only”), the $r$- + $s$-process group (“$r+s$”) and the “metal-rich” groups, respectively. For the purposes of this exercise, we assumed that the \[Al/Fe\] limits are detections, and therefore the observed dispersion for \[Al/Fe\] in the $r$-only group is effectively a lower limit. In general, there is a suggestion that the abundance errors are overestimated as the majority of elements lie below the 1:1 relation. For the $r$-only group, only Na (and perhaps Al) exhibits an abundance dispersion that significantly exceeds the average measurement error. For the $r+s$ group, a handful of elements including Na, Al, Cr, Zn, Zr and Ba exhibit abundance dispersions that exceed the average measurement error. For the metal-rich group, all elements exhibit abundance dispersions that are consistent with the expected dispersion given the average measurement error. That said, it is important to emphasise that for most elements, there is no evidence for an intrinsic abundance dispersion within a given group of stars. That is, with the exception of a few elements in the “$r$-only” and “$r+s$” groups, the dispersion in \[X/Fe\] is consistent with the measurement error. {width="0.9\hsize"} DISCUSSION ========== The aim of this discussion is to examine the nature of M2 in light of the chemical abundance ratios with an emphasis on the neutron-capture elements (Sections 5.2 and 5.3). In Table \[tab:meanabun\], we present the average abundance ratios and dispersions for $\log\epsilon$ (X) and \[X/Fe\] for the $r$-only, $r+s$ and metal-rich groups of stars. ------------------ ------------------ --------------- ---------------- ---------- -- ------------------ ---------- ---------------- ---------- -- ------------------ ---------- ---------------- ---------- Species $<\log\epsilon>$ $\sigma$[^25] $<$\[X/Fe\]$>$ $\sigma$ $<\log\epsilon>$ $\sigma$ $<$\[X/Fe\]$>$ $\sigma$ $<\log\epsilon>$ $\sigma$ $<$\[X/Fe\]$>$ $\sigma$ [O[i]{}]{} 7.43 0.21 0.42 0.16 7.77 0.08 0.66 0.04 7.97 … 0.36 … [Na[i]{}]{} 4.63 0.21 0.06 0.23 5.07 0.34 0.35 0.26 5.04 0.09 $-$0.17 0.04 [Mg[i]{}]{} 6.30 0.08 0.38 0.08 6.51 0.10 0.46 0.11 6.82 0.06 0.25 0.03 [Al[i]{}]{} 5.22 0.20 0.45 0.18 5.55 0.32 0.61 0.24 5.56 0.13 0.12 0.08 [Si[i]{}]{} 6.23 0.05 0.40 0.01 6.61 0.19 0.61 0.12 6.74 0.03 0.26 0.03 [Ca[i]{}]{} 4.95 0.06 0.28 0.02 5.17 0.11 0.34 0.05 5.50 0.09 0.19 0.05 [Sc[ii]{}]{} 1.45 0.09 $-$0.03 0.06 1.58 0.10 $-$0.06 0.08 1.99 0.12 $-$0.13 0.06 [Ti[i]{}]{} 3.45 0.05 0.17 0.02 3.75 0.13 0.32 0.05 4.07 0.14 0.14 0.08 [Ti[ii]{}]{} 3.70 0.09 0.43 0.07 3.76 0.06 0.33 0.09 4.24 0.11 0.32 0.05 [Cr[i]{}]{} 3.91 0.06 $-$0.06 0.03 4.21 0.19 0.08 0.18 4.60 0.13 $-$0.02 0.08 [Cr[ii]{}]{} 4.11 0.10 0.14 0.08 4.22 0.07 0.12 0.15 4.74 0.15 0.11 0.09 [Mn[i]{}]{} 3.34 0.07 $-$0.41 0.05 3.51 0.03 $-$0.41 0.06 4.04 0.12 $-$0.36 0.07 [Fe[i]{}]{}[^26] 5.82 0.04 $-$1.68 0.04 5.99 0.08 $-$1.51 0.08 6.47 0.07 $-$1.03 0.07 [Fe[ii]{}]{}$^b$ 5.83 0.06 $-$1.66 0.06 5.99 0.10 $-$1.51 0.10 6.48 0.06 $-$1.02 0.06 [Co[i]{}]{} 3.19 0.02 $-$0.13 0.05 3.46 0.10 $-$0.02 0.07 3.87 0.11 $-$0.11 0.05 [Ni[i]{}]{} 4.49 0.05 $-$0.05 0.02 4.68 0.10 $-$0.03 0.03 5.14 0.08 $-$0.05 0.02 [Cu[i]{}]{} 1.84 0.07 $-$0.68 0.05 2.28 0.13 $-$0.40 0.09 2.63 0.19 $-$0.53 0.13 [Zn[i]{}]{} 2.93 0.13 0.04 0.11 3.13 0.29 0.08 0.27 3.56 0.10 0.02 0.14 [Sr[i]{}]{} 0.63 0.07 $-$0.56 0.10 1.31 0.20 $-$0.04 0.23 1.22 0.22 $-$0.62 0.18 [Y[ii]{}]{} 0.36 0.09 $-$0.18 0.06 1.07 0.11 0.38 0.12 1.06 0.16 $-$0.12 0.11 [Zr[i]{}]{} 0.83 0.15 $-$0.08 0.17 1.62 0.07 0.56 0.05 1.56 0.12 0.01 0.06 [Zr[ii]{}]{} 1.17 0.13 0.26 0.11 1.78 0.22 0.72 0.23 1.85 0.13 0.30 0.07 [Mo[i]{}]{} 0.10 … $-$0.03 … … … … … … … … … [Ba[ii]{}]{} 0.69 0.17 0.19 0.15 1.59 0.28 0.92 0.25 1.45 0.16 0.30 0.12 [La[ii]{}]{} $-$0.47 0.09 0.11 0.07 0.28 0.18 0.69 0.13 0.34 0.13 0.27 0.06 [Ce[ii]{}]{} $-$0.10 0.09 $-$0.01 0.07 0.57 0.14 0.50 0.12 0.72 0.09 0.16 0.03 [Pr[ii]{}]{} $-$0.88 0.07 0.08 0.04 $-$0.30 0.08 0.49 0.03 $-$0.12 0.20 0.19 0.14 [Nd[ii]{}]{} $-$0.10 0.08 0.15 0.05 0.51 0.17 0.61 0.14 0.74 0.14 0.35 0.08 [Sm[ii]{}]{} $-$0.45 0.10 0.26 0.08 $-$0.12 0.10 0.43 0.08 0.44 0.17 0.50 0.11 [Eu[ii]{}]{} $-$0.78 0.15 0.38 0.14 $-$0.74 0.14 0.25 0.08 0.06 0.17 0.57 0.12 [Gd[ii]{}]{} $-$0.28 0.09 0.33 0.08 $-$0.05 0.06 0.43 0.13 0.80 … 0.70 … [Tb[ii]{}]{} $-$1.13 0.07 0.27 0.14 … … … … … … … … [Dy,[ii]{}]{} $-$0.25 0.14 0.33 0.12 0.12 0.16 0.56 0.23 … … … … [Er[ii]{}]{} $-$0.68 … 0.05 … … … … … … … … … [Tm[ii]{}]{} $-$1.84 … $-$0.19 … … … … … … … … … [Yb[ii]{}]{} $-$0.94 0.14 $-$0.08 0.07 … … … … … … … … [Hf[ii]{}]{} $-$0.77 0.06 0.08 0.13 $-$0.27 0.11 0.42 0.00 … … … … [Pb[i]{}]{} 0.15 0.18 0.08 0.13 1.40 0.21 1.18 0.32 … … … … ------------------ ------------------ --------------- ---------------- ---------- -- ------------------ ---------- ---------------- ---------- -- ------------------ ---------- ---------------- ---------- M2 shares similar, and peculiar, characteristics found in the unusual globular clusters M22, NGC 1851 and $\omega$ Cen, namely, a dispersion in metallicity and neutron-capture abundance ratios. If a subset of the metal-rich group are genuine cluster members, then M2 would host stars that span a range in metallicity from \[Fe/H\] $\approx$ $-$1.6 to \[Fe/H\] $\approx$ $-$1.0, a factor of four. We note further that even if the most metal-rich stars are not members, there still remains a metallicity spread of order 0.25 dex among the stars for which we assert cluster membership. M2 appears to be different from M22 and NGC 1851; for the latter two clusters, the number of stars on the bright subgiant branch is similar to the number on the faint subgiant branch. In contrast, for M2 the canonical RGB stars represent the overwhelming majority of stars. As noted in Section 4.4, the relative numbers of canonical and anomalous RGB stars in M2 is probably comparable to the relative numbers of bright ($\sim$96%) and faint ($\sim$4%) subgiant branch stars [@piotto12]. Light-, $\alpha$- and Fe-peak elements -------------------------------------- Regarding the light elements, even with our limited sample it is apparent that the $r$-only and $r+s$ groups both exhibit star-to-star abundance variations and correlations between \[Na/Fe\] and \[Al/Fe\]. The two populations in M22 and NGC 1851 both exhibit a O-Na anticorrelation [@marino11; @carretta11], and for $\omega$ Cen, the O-Na anticorrelation is present across a broad metallicity range [@norris95b; @johnson10]. Indeed, every well studied Galactic globular cluster exhibits star-to-star abundance variations for the light elements C, N, O, F, Na, Mg and Al (e.g., see reviews by @kraft94; @gratton04; @gratton12). While these abundance variations are believed to be produced through hydrogen-burning, the specific site continues to be debated (e.g., @fenner04 [@ventura05; @decressin07; @demink09]). For the metal-rich group, the apparent absence of a star-to-star abundance variation for the light elements is intriguing, although the sample size is small. No such abundance spread would be expected if these were all field stars. On the other hand, a similar situation is present in the M54+Sagittarius (Sgr) system. While the O-Na anticorrelation is evident in M54, the more metal-rich Sgr stars do not exhibit this pattern [@carretta10b]. If the four stars in the metal-rich group are indeed cluster members, then M2 would share this peculiar feature with M54+Sgr. For the $\alpha$ and Fe-peak elements, there is no compelling evidence for a star-to-star abundance variation within a given group. Additionally, the abundance ratios \[X/Fe\] for a given star are compatible with field stars at the same metallicity. In other words, these elements appear to be well-behaved. The abundance of Cu offers an important tool to distinguish between field stars and “$\omega$ Cen-like” systems. For M2, the Cu abundance may help establish additional similarities with $\omega$ Cen and potentially cluster membership, or otherwise, for the four metal-rich objects for the following reasons. In the metallicity regime $-$2.0 $\lesssim$ \[Fe/H\] $\lesssim$ $-$0.5, field stars exhibit a systematic increase in \[Cu/Fe\] with increasing metallicity [e.g., @sneden88; @mishenina02; @primas08]. Mono-metallic globular clusters in the same metallicity range appear to follow the field star trend [@simmerer03]. $\omega$ Cen, however, displays a near constant Cu abundance, \[Cu/Fe\] $\approx$ $-$0.5, over the range $-$1.9 $\le$ \[Fe/H\] $\le$ $-$0.8 [@cunha02]. At higher metallicities, $-$1.2 $\le$ \[Fe/H\] $\le$ $-$0.4, there is evidence for an increase in the \[Cu/Fe\] ratio in $\omega$ Cen [@pancino02], although the rate of that increase is smaller than in field stars. Chemical evolution models of $\omega$ Cen and the Milky Way by @romano07 attribute the nucleosynthesis of Cu to massive stars and successfully reproduce the observed trends. {width="0.8\hsize"} In Figure \[fig:cu\], we plot \[Cu/Fe\] versus \[Fe/H\] for M2, field stars [@mishenina02], mono-metallic globular clusters [@simmerer03] and $\omega$ Cen [@cunha02; @pancino02]. The Cu abundances in M2 displayed in this figure have been adjusted in the following manner. Following @simmerer03, the abundances from the 5105Å and 5782Å lines are referenced to solar values of $\log \epsilon_\odot$ = 4.21 and $\log \epsilon_\odot$ = 4.06, respectively. Such an approach reflects the different solar abundances obtained from these lines, and we note that the abundances we derive for program stars from the 5782Å line are, on average, 0.21 dex $\pm$ 0.03 dex ($\sigma$ = 0.11 dex) higher than those from the 5105Åline. The gradient of the linear fit to M2 is not affected by these zero-point offsets. We also stress that although this figure includes data from numerous studies, the linear fit in each panel is performed upon data obtained from a single study (for the lower left panel, M2, M22 and NGC 1851 are from different studies but those data are not included in the linear fit). So long as each sample is analysed uniformly, the slopes should be robust and we can compare them in a quantitative manner. In Figure \[fig:cu\], the slopes for the field stars and mono-metallic globular clusters are in good agreement, and these slopes differ from that seen in $\omega$ Cen. The behaviour of the slope of \[Cu/Fe\] versus \[Fe/H\] in M2—whether or not the $r+s$ stars are considered—is different from the field stars, mono-metallic globular clusters and $\omega$ Cen over the metallicity range $-$1.7 $\leq$ \[Fe/H\] $\leq$ $-0.9$. The metal-rich stars, relative to the $r$-only and $r+s$ groups, do not follow the field star trend, and this may be the strongest abundance-based evidence that they are cluster members. Furthermore, if the metal-rich stars are indeed cluster members, then M2 does not share a similar chemical enrichment history to $\omega$ Cen, at least for Cu. Figure \[fig:cu\] also demonstrates that the mean \[Cu/Fe\] ratios in NGC 1851 [@carretta11] and the $r$-only groups in M2 and M22 [@roederer11] match the trends established by the mono-metallic globular clusters. The Cu in the $r+s$ group of stars may include small contributions from $s$-process nucleosynthesis, so we do not discuss Cu in the $r+s$ group here. Neutron-capture abundance patterns in M2 {#ncap} ---------------------------------------- We now turn our attention to the neutron-capture elements. In the abundance analysis described in Section 3.2, we examined up to 122 lines of elements with atomic numbers $Z \geq$ 38 in each of the program stars. All abundances were computed by matching synthetic spectra, generated using one dimensional plane-parallel model atmospheres, to the observed spectra under the assumption that LTE holds in the line-forming layers. The abundances of Sr and Pb were derived from neutral lines. For the program stars, Sr <span style="font-variant:small-caps;">i</span> and Pb <span style="font-variant:small-caps;">i</span> are minority species, and LTE calculations will tend to underestimate the populations of the lower levels of the Sr <span style="font-variant:small-caps;">i</span> 4607Å and Pb <span style="font-variant:small-caps;">i</span> 4057Å transitions. Abundances of strontium and lead derived in LTE from these lines are thus underestimated. Calculations allowing for departures from LTE in the line-forming layers by making reasonable assumptions for the photoionization cross sections have been made for stars with stellar parameters similar to those in our sample. These non-LTE calculations suggest that our LTE analysis may underestimate the strontium abundance by $\approx$ 0.3$-$0.5 dex [@bergemann12; @hansen13] and the lead abundance by $\approx$ 0.3$-$0.4 dex [@mashonkina12]. The values presented in our tables and figures reflect the LTE values. Neglecting the non-LTE corrections for these two elements should not significantly affect any of the abundance differences between the $r$-only and $r+s$ groups of stars that we shall discuss below. Figure \[starplot\] illustrates the heavy element abundance patterns found in each star of our sample. The six stars shown in the left-hand panels are those on the canonical RGB, the four stars in the middle panels are the neutron-capture rich anomalous RGB stars and the four stars in the right-hand panels are the metal-rich anomalous RGB stars. For comparison, in each panel of this figure we overplot the heavy element abundance pattern found in the rich standard star  (normalised to the Eu abundance), whose metallicity is only a factor of $\approx$ 2.5 lower than the majority of stars in M2. The stars on the canonical RGB have heavy element abundance patterns very similar to one another and to the  pattern in , and the overall amounts of heavy elements are constant within their mutual uncertainties. We refer to these six stars (NR 37, NR 58, NR 60, NR 76, NR 99 and NR 124) on the canonical RGB as the “$r$-only group.” The reasoning behind this name will be made clear shortly. {width="0.95\hsize"} As shown in the middle panels of Figure \[starplot\], the heavy elements in the neutron-capture rich anomalous RGB stars in M2 tell a different story. All heavy elements except europium in these four stars exhibit noticeable abundance enhancements relative to the stars on the canonical RGB and, therefore, enhancements relative to the  standard . The pattern changes little from one star to the next, and the overall abundances in this group of stars are also constant within their mutual uncertainties. We refer to these four stars (NR 38, NR 47, NR 77 and NR 81) as the “$r+s$ group.” The consistent patterns and levels of enhancement found within each of the $r$-only and $r+s$ groups suggests that we can average together their abundances to reduce the random uncertainties, which is especially helpful for elements whose abundances are derived from small numbers of lines. These mean abundance patterns are listed in Table \[tab:meanabun\] and illustrated in Figure \[meanplot\]. Subtle differences between the stars in the $r$-only group and , (e.g., small overabundances in M2 for strontium, yttrium, zirconium, barium, cerium and neodymium, as well as small underabundances in M2 for ytterbium) may simply reflect differing combinations of material produced by the so-called weak and main components of the  enriching M2 and . This is plausible because the overall level of $r$-process enhancement relative to iron is different in  and M2, with \[Eu/Fe\] = +0.9 and +0.4, respectively. Regardless, Figure \[starplot\] demonstrates that the heavy elements in the stars in the $r$-only group in M2 owe their origin to nucleosynthesis with little or no  contributions. ![Top panel: Mean logarithmic abundances for the six $r$-only stars (black circles) and the four $r+s$ stars (red triangles). (Only elements measured in more than one star are included, i.e., we exclude Mo, Er and Tm.) The zirconium abundance derived from Zr <span style="font-variant:small-caps;">ii</span> lines is shown. The gray line illustrates the abundances in the  standard star  normalized to the europium abundance. Bottom panel: Differences in these mean abundances. The dotted line indicates zero difference. \[meanplot\] ](fig15.ps){width="0.99\hsize"} In contrast, the middle panels of Figure \[starplot\] demonstrate that the $r+s$ stars have abundance patterns that are inconsistent with nucleosynthesis alone. Figure \[sfracplot\] demonstrates that the excess of heavy elements ($Z \geq$ 38) found in the $r+s$ group relative to the $r$-only group exhibits an unmistakable correlation with the fraction of each element attributed to an  origin in solar system material. Elements with a high  fraction in the solar system are most overabundant in the $r+s$ group in M2, and those with small  fractions in the solar system show little excess. Only a few percent of the europium in the solar system is attributed to  nucleosynthesis, and this element shows a constant abundance in both the $r$-only and $r+s$ groups of stars; the average $\log\epsilon$ (Eu) and \[Eu/Fe\] ratios are very similar between these two groups. This suggests that there is a common  abundance foundation in each star in the $r$-only and $r+s$ groups in M2. (We set aside, for now, the metal-rich group.) Following @roederer11, we speculate that the stars in the $r+s$ group formed from additional material enriched by products of  nucleosynthesis as well as iron. We emphasise that the abundance differences between these two groups of stars should be nearly insensitive to any non-LTE effects given the modest range in stellar parameters ([$T_{\rm eff}$]{}, [$\log g$]{}, \[Fe/H\]) spanned by the sample. {width="0.9\hsize"} We now consider the metal-rich group (stars NR 132, NR 207, NR 254 and NR 378) shown in the right-hand panels of Figure \[starplot\] noting that these stars may, or may not, be cluster members. The heavy element abundances in the metal-rich group closely resemble an  pattern, despite the fact that the overall metallicity of these objects is a factor of $\approx$ 4 higher than the other cluster members. Furthermore, the \[X/Fe\] ratios (where X denotes any element with $Z \geq$ 56) in the metal-rich group are on average 0.17 dex $\pm$ 0.02 dex ($\sigma$ = 0.05 dex) higher than in the $r$-only stars, i.e., a factor of $\approx$1.5. If we assume that the overall metal content in an isolated stellar system increases monotonically with time, the metal-rich group should have formed later than either the $r$-only or $r+s$ groups. To the best of our knowledge, no isolated self-enriched stellar system shows a return to  dominance after previous enrichment by a substantial amount of material. We conclude that the four metal-rich stars cannot be easily understood as members of a single self-enriched stellar system. That said, the data do not preclude a scenario in which M2 is composed of independent fragments that experienced different chemical enrichment histories [@searle77]. The origin of the $s$-process material {#sproorigin} -------------------------------------- If we assume that the  enrichment is common to both the $r$-only and $r+s$ groups in M2, we can subtract the average abundances found in the $r$-only group from those in the $r+s$ group to obtain the intrinsic abundance ratios of the  material added to the $r+s$ group. These differences are shown in Figure \[meanplot\] (as a function of atomic number) and in Figure \[sfracplot\] (as a function of the fraction of each element attributed to the $s$-process in solar system material). @roederer11 performed a similar calculation for the two stellar groups in M22 and the unrelated clusters M4 and M5 using data from @yong08m4m5a [@yong08m4m5b]. In Figure \[sfracplot\], we include the abundance differences for the M22 groups as well as the abundance differences when subtracting the mean values for M5 from those of M4 which we denote as “M4 $-$ M5”[^27]. In this figure we adopt the values from @bisterzo11 for the fraction of each element attributed to the $s$-process in solar system material. As noted, there is a clear trend between the abundance differences and the $s$-process fraction. As originally proposed by @roederer11, we argue that the abundance residual represents $s$-process material. When considering all elements with $Z$ $\ge$ 38 (i.e., Sr and heavier elements), the three “systems,” (1) $\langle r+s \rangle$ $-$ $\langle r$-only$\rangle$ in M2, (2) $\langle r+s \rangle$ $-$ $\langle r$-only$\rangle$ in M22 and (3) $\langle$M4$\rangle$ $-$ $\langle$M5$\rangle$, exhibit identical gradients within their mutual uncertainties. Such a result is surprising given that the yields for the $s$-process elements in AGB stars are mass and metallicity dependent (e.g., @busso01 [@cristallo11; @karakas12]; Fishlock et al., in preparation). If our interpretation that the abundance residuals in these systems represent $s$-process material is correct, then the implication is that these three systems, which span a range in metallicity from \[Fe/H\] $\approx$ $-$1.8 to \[Fe/H\] $\approx$ $-$1.2, experienced enrichment by $s$-process material of indistinguishable composition. Quantitative chemical evolution modelling is needed to test this intriguing hypothesis, and Shingles et al. (in preparation) are investigating M22 and M4 and comparing the predicted and observed enrichment patterns taking into account yields from AGB and massive stars. If we consider only elements with 38 $\le$ $Z$ $\le$ 45 (i.e., Sr to Rh), the gradients do not exhibit any consistent patterns. In sharp contrast, however, the elements from 56 $\le$ $Z$ $\le$ 72 (i.e., Ba to Hf) exhibit identical gradients within their mutual uncertainties. For these elements, the measured abundance differences in each system are consistent with a single relation. This implies that the enrichment in M2, M22 and M4 involved $s$-process material of remarkably similar composition despite the factor of $\sim$ 4 difference in metallicity. The intrinsic  ratios and indices[^28] are \[Pb/La\]$_{\rm s}$ $= +$0.53 in M2, $+$0.18 in M22 and $-$0.01 in M4 $-$ M5; \[$hs$/$ls$\]$_{\rm s}$ $=-$ 0.02 in M2, $-$0.01 in M22 and $-$0.50 in M4 $-$ M5; and \[Pb/$hs$\]$_{\rm s}$ $= +$0.72 in M2, $+$0.29 in M22 and $+$0.28 in M4 $-$ M5. Uncertainties on these ratios are typically 0.1$-$0.2 dex. These ratios and indices are largely insensitive to uncertainties in the atomic data and non-LTE effects. For \[Pb/La\], M2 exhibits a higher ratio than M22 and M4 $-$ M5. For \[$hs$/$ls$\], M4 $-$ M5 exhibits lower ratios than M22 and M2, although this may reflect the higher metallicity of M4 and M5 relative to the other two clusters. For \[Pb/$hs$\], M2 exhibits a higher ratio than the other two systems. A number of studies have investigated  nucleosynthesis in metal-poor stars on the AGB [@goriely00; @goriely01; @cristallo09; @cristallo11; @bisterzo10 Fishlock et al., in preparation]. While most of these models fail to offer an exact match for the metallicity of M2, we can use them to get a sense of  nucleosynthesis ratios predicted for metallicities higher and lower than M2. We find encouraging agreement when comparing our results to the \[$hs$/$ls$\] and \[Pb/$hs$\] indices presented in Figures C3 and C5 of @bisterzo10 for 3 and 5 [M$_{\odot}$]{} AGB stars at the appropriate metallicities for M2, M22 and M4 $-$ M5. Furthermore, we note that the yields of Fishlock et al. (in preparation) for their 3 [M$_{\odot}$]{} and 3.5 [M$_{\odot}$]{} models for \[Fe/H\] = $-$1.2 bracket the \[Pb/La\]$_{\rm s}$, \[$hs$/$ls$\]$_{\rm s}$ and \[Pb/$hs$\]$_{\rm s}$ ratios in M2, M22 and M4 $-$ M5. Quantitative chemical evolution models based on their yields would be of great interest. Overall, we reach the same conclusion drawn by @roederer11: AGB stars with masses less than 3 [M$_{\odot}$]{} cannot reproduce the observed ratios unless the standard $^{13}$C pocket efficiency in the models is reduced by factors of 30 or more. If we assume that the stars in the $r+s$ group in M2 formed later than the stars in the $r$-only group, and that the AGB stars responsible for distributing this  material in M2 formed simultaneously with the stars in the $r$-only group, this sets an upper limit on the amount of time that passed between the formation of the $r$-only group and the $r+s$ group. For a 3 [M$_{\odot}$]{} AGB star, adopting the approximate stellar lifetimes computed by @mowlavi12, this sets a limit of no more than 300 Myr or so between the two groups. Of course, this limit would be even smaller if higher-mass AGB stars were the source of the  material. Finally, although we have focused on the neutron-capture elements, the difference in \[Fe/H\] between the $r$-only and $r+s$ groups requires some source(s) that produces the elements from Si to Zn (and perhaps other elements) to increase the abundances of these elements between these groups. CONCLUSION ========== In this paper we present a spectroscopic analysis of giant stars in the multiple population globular cluster M2. Our principal and novel results include the following. First, we identify a star-to-star dispersion in iron abundance with the anomalous RGB stars (i.e., stars lying redward of the dominant RGB) being more metal-rich than the canonical RGB objects. The iron abundance distribution has a dominant peak at \[Fe/H\] $\approx$ $-$1.7 and smaller peaks at $-$1.5 and $-$1.0, although membership for the latter group remains to be established. Secondly, the neutron-capture element abundances exhibit a star-to-star dispersion with a possible bimodal distribution. In this regard, M2 is chemically similar to the globular clusters M22, NGC 1851 and $\omega$ Cen, whose subgiant branches exhibit multiple sequences. It is likely that M2 has therefore experienced a similarly complex formation history. Thirdly, when subtracting the average abundances in the $r$-only group from those of the $r+s$ group, the abundance residual exhibits a striking correlation with the fraction of each element attributed to the $s$-process in solar system material. This residual is remarkably similar to that found in M22 and in M4 $-$ M5. Such a similarity would indicate that M2, M22, and M4 were enriched by $s$-process material of identical composition and potentially offers important observational constraints on the nature of the $s$-process in low metallicity environments. A comparison with theoretical predictions reveals that AGB stars with masses less than 3 [M$_{\odot}$]{} are unlikely to have played a major role in the chemical enrichment of M2. In addition to the AGB star contribution, some source(s) is needed to increase the abundances of the elements from Si to Zn in the $r+s$ group relative to the $r$-only group. Additional studies are essential to understand the formation and evolution of this complex cluster. Acknowledgments {#acknowledgments .unnumbered} =============== We thank K. Cudworth for providing electronic data and the referee, Kim Venn, for helpful comments. Some of the AAT data were taken in time allocated to the AEGIS program (PI Keller). The authors are grateful to the AEGIS team for access to these data. D.Y, G.D.C, A.I.K, J.E.N, A.F.M and A.P.M gratefully acknowledge support from the Australian Research Council (grants DP0984924, FT110100475, DP120100475, DP120100991 and DP120101237). Funding for the Stellar Astrophysics Centre is provided by The Danish National Research Foundation. The research is supported by the ASTERISK project (ASTERoseismic Investigations with SONG and Kepler) funded by the European Research Council (Grant agreement no.: 267864). [139]{} natexlab\#1[\#1]{} , C., [Barklem]{}, P. S., [Lambert]{}, D. L., & [Cunha]{}, K. 2004, , 420, 183 , A., [Arribas]{}, S., & [Mart[í]{}nez-Roger]{}, C. 1999, , 140, 261 , T. E. & [Da Costa]{}, G. S. 1991, , 101, 1329 , T. E. & [Zinn]{}, R. 1988, , 96, 92 , H. C. 1955, , 60, 317 , M., [Grevesse]{}, N., [Sauval]{}, A. J., & [Scott]{}, P. 2009, , 47, 481 , L. R., [Piotto]{}, G., [Anderson]{}, J., [Cassisi]{}, S., [King]{}, I. R., [Momany]{}, Y., & [Carraro]{}, G. 2004, , 605, L125 , K. 2011, , 412, 2241 , K. & [Freeman]{}, K. C. 2003, , 346, L11 , K. & [Yong]{}, D. 2012, , 419, 2063 , M., [Hansen]{}, C. J., [Bautista]{}, M., & [Ruchti]{}, G. 2012, , 546, A90 , R., [Shectman]{}, S. A., [Gunnels]{}, S. M., [Mochnacki]{}, S., & [Athey]{}, A. E. 2003, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4841, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. M. [Iye]{} & A. F. M. [Moorwood]{}, 1694–1704 , E., [Baudoux]{}, M., [Kurucz]{}, R. L., [Ansbacher]{}, W., & [Pinnington]{}, E. H. 1991, , 249, 539 , [É]{}., [Blagoev]{}, K., [Engstr[ö]{}m]{}, L., [Hartman]{}, H., [Lundberg]{}, H., [Malcheva]{}, G., [Nilsson]{}, H., [Whitehead]{}, R. B., [Palmeri]{}, P., & [Quinet]{}, P. 2011, , 414, 3350 , E., [Garnir]{}, H. P., [Palmeri]{}, P., [Li]{}, Z. S., & [Svanberg]{}, S. 2000, , 312, 116 , E., [Grevesse]{}, N., [Hannaford]{}, P., & [Lowe]{}, R. M. 1981, , 248, 867 , S., [Gallino]{}, R., [Straniero]{}, O., [Cristallo]{}, S., & [K[ä]{}ppeler]{}, F. 2010, , 404, 1529 —. 2011, , 418, 284 , D. E., [Booth]{}, A. J., [Haddock]{}, D. J., [Petford]{}, A. D., & [Leggett]{}, S. K. 1986, , 220, 549 , D. E., [Ibbetson]{}, P. A., [Petford]{}, A. D., & [Shallis]{}, M. J. 1979, , 186, 633 , D. E., [Lynas-Gray]{}, A. E., & [Smith]{}, G. 1995, , 296, 217 , D. E., [Petford]{}, A. D., & [Shallis]{}, M. J. 1979, , 186, 657 , D. E., [Petford]{}, A. D., [Shallis]{}, M. J., & [Simmons]{}, G. J. 1980, , 191, 445 , M., [Gallino]{}, R., [Lambert]{}, D. L., [Travaglio]{}, C., & [Smith]{}, V. V. 2001, , 557, 802 , E., [Bragaglia]{}, A., [Gratton]{}, R., [D’Orazi]{}, V., & [Lucatello]{}, S. 2009, , 508, 695 , E., [Bragaglia]{}, A., [Gratton]{}, R. G., [D’Orazi]{}, V., [Lucatello]{}, S., [Momany]{}, Y., [Sollima]{}, A., [Bellazzini]{}, M., [Catanzaro]{}, G., & [Leone]{}, F. 2014,  in press (arXiv:1401.7325) , E., [Bragaglia]{}, A., [Gratton]{}, R. G., [Lucatello]{}, S., [Bellazzini]{}, M., [Catanzaro]{}, G., [Leone]{}, F., [Momany]{}, Y., [Piotto]{}, G., & [D’Orazi]{}, V. 2010, , 520, A95 —. 2010, , 714, L7 , E., [Bragaglia]{}, A., [Gratton]{}, R. G., [Lucatello]{}, S., [D’Orazi]{}, V., [Bellazzini]{}, M., [Catanzaro]{}, G., [Leone]{}, F., [Momany]{}, Y., & [Sollima]{}, A. 2013, , 557, A138 , E., [Gratton]{}, R. G., [Lucatello]{}, S., [Bragaglia]{}, A., [Catanzaro]{}, G., [Leone]{}, F., [Momany]{}, Y., [D’Orazi]{}, V., [Cassisi]{}, S., [D’Antona]{}, F., & [Ortolani]{}, S. 2010, , 722, L1 , E., [Lucatello]{}, S., [Gratton]{}, R. G., [Bragaglia]{}, A., & [D’Orazi]{}, V. 2011, , 533, A69 , F. & [Kurucz]{}, R. L. 2003, in IAU Symp. 210, Modelling of Stellar Atmospheres, ed. N. Piskunov, W. W. Weiss, & D. F. Gray (San Francisco, CA: ASP), A20 , C. & [Spergel]{}, D. N. 2011, , 726, 36 , J. J., [Sneden]{}, C., [Beers]{}, T. C., [Lawler]{}, J. E., [Simmerer]{}, J., [Truran]{}, J. W., [Primas]{}, F., [Collier]{}, J., & [Burles]{}, S. 2005, , 627, 238 , J. J., [Sneden]{}, C., [Burles]{}, S., [Ivans]{}, I. I., [Beers]{}, T. C., [Truran]{}, J. W., [Lawler]{}, J. E., [Primas]{}, F., [Fuller]{}, G. M., [Pfeiffer]{}, B., & [Kratz]{}, K.-L. 2002, , 572, 861 , S., [Piersanti]{}, L., [Straniero]{}, O., [Gallino]{}, R., [Dom[í]{}nguez]{}, I., [Abia]{}, C., [Di Rico]{}, G., [Quintini]{}, M., & [Bisterzo]{}, S. 2011, , 197, 17 , S., [Straniero]{}, O., [Gallino]{}, R., [Piersanti]{}, L., [Dom[í]{}nguez]{}, I., & [Lederer]{}, M. T. 2009, , 696, 797 , K. M. & [Rauscher]{}, B. J. 1987, , 93, 856 , K., [Smith]{}, V. V., [Suntzeff]{}, N. B., [Norris]{}, J. E., [Da Costa]{}, G. S., & [Plez]{}, B. 2002, , 124, 379 , G. S., [Held]{}, E. V., & [Saviane]{}, I. 2014, , 438, 3507 , G. S., [Held]{}, E. V., [Saviane]{}, I., & [Gullieuszik]{}, M. 2009, , 705, 1481 , F., [Ventura]{}, P., [Caloi]{}, V., [D’Ercole]{}, A., [Vesperini]{}, E., [Carini]{}, R., & [Di Criscienzo]{}, M. 2010, , 715, L63 , S. E., [Pols]{}, O. R., [Langer]{}, N., & [Izzard]{}, R. G. 2009, , 507, L1 , T., [Meynet]{}, G., [Charbonnel]{}, C., [Prantzos]{}, N., & [Ekstr[ö]{}m]{}, S. 2007, , 464, 1029 , E. A., [Lawler]{}, J. E., [Sneden]{}, C., & [Cowan]{}, J. J. 2003, , 148, 543 —. 2006, , 167, 292 , A., [Vesperini]{}, E., [D’Antona]{}, F., [McMillan]{}, S. L. W., & [Recchi]{}, S. 2008, , 391, 825 , V., [Gratton]{}, R. G., [Pancino]{}, E., [Bragaglia]{}, A., [Carretta]{}, E., [Lucatello]{}, S., & [Sneden]{}, C. 2011, , 534, A29 , A., [Chaboyer]{}, B., [Jevremovi[ć]{}]{}, D., [Kostov]{}, V., [Baron]{}, E., & [Ferguson]{}, J. W. 2008, , 178, 89 , Y., [Campbell]{}, S., [Karakas]{}, A. I., [Lattanzio]{}, J. C., & [Gibson]{}, B. K. 2004, , 353, 789 , J. R. & [Wiese]{}, W. L. 2009, in CRC Handbook of Chemistry and Physics, 90th edn. CRC Press, Boca Raton, p. 10 , J. P. 2000, , 120, 1841 , L., [Groenewegen]{}, M. A. T., [Hatziminaoglou]{}, E., & [da Costa]{}, L. 2005, , 436, 895 , S. & [Mowlavi]{}, N. 2000, , 362, 599 , S. & [Siess]{}, L. 2001, , 378, L25 , R., [Sneden]{}, C., & [Carretta]{}, E. 2004, , 42, 385 , R. G., [Carretta]{}, E., & [Bragaglia]{}, A. 2012, , 20, 50 , R. G., [Carretta]{}, E., [Claudi]{}, R., [Lucatello]{}, S., & [Barbieri]{}, M. 2003, , 404, 187 , F., [Catelan]{}, M., [Landsman]{}, W. B., [Stetson]{}, P. B., & [Andersen]{}, M. I. 1999, , 524, 242 , C. J., [Bergemann]{}, M., [Cescutti]{}, G., [Fran[ç]{}ois]{}, P., [Arcones]{}, A., [Karakas]{}, A. I., [Lind]{}, K., & [Chiappini]{}, C. 2013, , 551, A57 , W. E. 1975, , 29, 397 —. 1996, , 112, 1487 Herwig, F., VandenBerg, D. A., Navarro, J. F., Ferguson, J., & Paxton, B. 2012, , 757, 132 , I. I., [Kraft]{}, R. P., [Sneden]{}, C., [Smith]{}, G. H., [Rich]{}, R. M., & [Shetrone]{}, M. 2001, , 122, 1438 , I. I., [Simmerer]{}, J., [Sneden]{}, C., [Lawler]{}, J. E., [Cowan]{}, J. J., [Gallino]{}, R., & [Bisterzo]{}, S. 2006, , 645, 613 Ivans, I. I., Sneden, C., Kraft, R. P., et al. 1999, , 118, 1273 , S., [Litz[é]{}n]{}, U., & [Wahlgren]{}, G. M. 2001, , 64, 455 , C. I. & [Pilachowski]{}, C. A. 2010, , 722, 1373 , A. I., [Garc[í]{}a-Hern[á]{}ndez]{}, D. A., & [Lugaro]{}, M. 2012, , 751, 8 , R. P. 1994, , 106, 553 , R. & [Bell]{}, B. 1995, Atomic Line Data (R.L. Kurucz and B. Bell) Kurucz CD-ROM No. 23. Cambridge, Mass.: Smithsonian Astrophysical Observatory, 1995., 23 , C., [Pancino]{}, E., [Mucciarelli]{}, A., [Bellazzini]{}, M., [Rejkuba]{}, M., [Marinoni]{}, S., [Cocozza]{}, G., [Altavilla]{}, G., & [Ragaini]{}, S. 2013, , 433, 1941 , C., [Pancino]{}, E., [Mucciarelli]{}, A., & [Milone]{}, A. P. 2012, , 548, A107 , J. E., [Bonvallet]{}, G., & [Sneden]{}, C. 2001, , 556, 452 , J. E., [den Hartog]{}, E. A., [Labby]{}, Z. E., [Sneden]{}, C., [Cowan]{}, J. J., & [Ivans]{}, I. I. 2007, , 169, 120 , J. E., [Den Hartog]{}, E. A., [Sneden]{}, C., & [Cowan]{}, J. J. 2006, , 162, 227 , J. E., [Sneden]{}, C., [Cowan]{}, J. J., [Ivans]{}, I. I., & [Den Hartog]{}, E. A. 2009, , 182, 51 , J. E., [Sneden]{}, C., [Cowan]{}, J. J., [Wyart]{}, J.-F., [Ivans]{}, I. I., [Sobeck]{}, J. S., [Stockett]{}, M. H., & [Den Hartog]{}, E. A. 2008, , 178, 71 , J. E., [Wickliffe]{}, M. E., [Cowley]{}, C. R., & [Sneden]{}, C. 2001, , 137, 341 , J. E., [Wickliffe]{}, M. E., [den Hartog]{}, E. A., & [Sneden]{}, C. 2001, , 563, 1075 , J. E., [Wyart]{}, J.-F., & [Blaise]{}, J. 2001, , 137, 351 , R., [Chatelain]{}, R., [Holt]{}, R. A., [Rehse]{}, S. J., [Rosner]{}, S. D., & [Scholl]{}, T. J. 2007, , 76, 577 , G., [Nilsson]{}, H., [Asplund]{}, M., & [Johansson]{}, S. 2006, , 456, 1181 , A., [Sollima]{}, A., [D’Ercole]{}, A., [Gibson]{}, B. K., & [Ferraro]{}, F. R. 2007, , 382, 443 , A. F., [Milone]{}, A. P., [Piotto]{}, G., [Villanova]{}, S., [Bedin]{}, L. R., [Bellini]{}, A., & [Renzini]{}, A. 2009, , 505, 1099 , A. F., [Sneden]{}, C., [Kraft]{}, R. P., [Wallerstein]{}, G., [Norris]{}, J. E., [da Costa]{}, G., [Milone]{}, A. P., [Ivans]{}, I. I., [Gonzalez]{}, G., [Fulbright]{}, J. P., [Hilker]{}, M., [Piotto]{}, G., [Zoccali]{}, M., & [Stetson]{}, P. B. 2011, , 532, A8 , L., [Ryabtsev]{}, A., & [Frebel]{}, A. 2012, , 540, A98 , A. 1998, , 115, 1640 , A. P., [Bedin]{}, L. R., [Piotto]{}, G., [Anderson]{}, J., [King]{}, I. R., [Sarajedini]{}, A., [Dotter]{}, A., [Chaboyer]{}, B., [Mar[í]{}n-Franch]{}, A., [Majewski]{}, S., [Aparicio]{}, A., [Hempel]{}, M., [Paust]{}, N. E. Q., [Reid]{}, I. N., [Rosenberg]{}, A., & [Siegel]{}, M. 2008, , 673, 241 , T. V., [Kovtyukh]{}, V. V., [Soubiran]{}, C., [Travaglio]{}, C., & [Busso]{}, M. 2002, , 396, 189 , N., [Eggenberger]{}, P., [Meynet]{}, G., [Ekstr[ö]{}m]{}, S., [Georgy]{}, C., [Maeder]{}, A., [Charbonnel]{}, C., & [Eyer]{}, L. 2012, , 541, A41 , K., [Aoki]{}, W., [Kawanomoto]{}, S., [Ando]{}, H., [Honda]{}, S., [Izumiura]{}, H., [Kambe]{}, E., [Okita]{}, K., [Sadakane]{}, K., [Sato]{}, B., [Tajitsu]{}, A., [Takada-Hidai]{}, T., [Tanaka]{}, W., [Watanabe]{}, E., & [Yoshida]{}, M. 2002, , 54, 855 , J. E. & [Da Costa]{}, G. S. 1995, , 447, 680 —. 1995, , 441, L81 , K., [Honda]{}, S., [Aoki]{}, W., [Kajino]{}, T., & [Mathews]{}, G. J. 2006, , 641, L117 , E., [Pasquini]{}, L., [Hill]{}, V., [Ferraro]{}, F. R., & [Bellazzini]{}, M. 2002, , 568, L101 , G. 2009, in IAU Symposium, Vol. 258, IAU Symposium, ed. E. E. [Mamajek]{}, D. R. [Soderblom]{}, & R. F. G. [Wyse]{}, 233–244 , G., [Milone]{}, A. P., [Anderson]{}, J., [Bedin]{}, L. R., [Bellini]{}, A., [Cassisi]{}, S., [Marino]{}, A. F., [Aparicio]{}, A., & [Nascimbeni]{}, V. 2012, , 760, 39 , F. & [Sobeck]{}, J. 2008, in American Institute of Physics Conference Series, Vol. 1001, Evolution and Nucleosynthesis in AGB Stars, ed. R. [Guandalini]{}, S. [Palmerini]{}, & M. [Busso]{}, 230–234 , J. X., [Naumov]{}, S. O., [Carney]{}, B. W., [McWilliam]{}, A., & [Wolfe]{}, A. M. 2000, , 120, 2513 , I. & [Mel[é]{}ndez]{}, J. 2005, , 626, 465 , S. V. & [Cohen]{}, J. G. 2002, , 123, 3277 , A. C., [Reyl[é]{}]{}, C., [Derri[è]{}re]{}, S., & [Picaud]{}, S. 2003, , 409, 523 , I. U., [Cowan]{}, J. J., [Karakas]{}, A. I., [Kratz]{}, K.-L., [Lugaro]{}, M., [Simmerer]{}, J., [Farouqi]{}, K., & [Sneden]{}, C. 2010, , 724, 975 , I. U. & [Lawler]{}, J. E. 2012, , 750, 76 , I. U., [Lawler]{}, J. E., [Cowan]{}, J. J., [Beers]{}, T. C., [Frebel]{}, A., [Ivans]{}, I. I., [Schatz]{}, H., [Sobeck]{}, J. S., & [Sneden]{}, C. 2012, , 747, L8 , I. U., [Lawler]{}, J. E., [Sneden]{}, C., [Cowan]{}, J. J., [Sobeck]{}, J. S., & [Pilachowski]{}, C. A. 2008, , 675, 723 , I. U., [Lawler]{}, J. E., [Sobeck]{}, J. S., [Beers]{}, T. C., [Cowan]{}, J. J., [Frebel]{}, A., [Ivans]{}, I. I., [Schatz]{}, H., [Sneden]{}, C., & [Thompson]{}, I. B. 2012, , 203, 27 , I. U., [Marino]{}, A. F., & [Sneden]{}, C. 2011, , 742, 37 , I. U., [Sneden]{}, C., [Lawler]{}, J. E., & [Cowan]{}, J. J. 2010, , 714, L123 , D. & [Matteucci]{}, F. 2007, , 378, L59 , W., [Bridges]{}, T., [Gillingham]{}, P., [Haynes]{}, R., [Smith]{}, G. A., [Whittard]{}, J. D., [Churilov]{}, V., [Lankshear]{}, A., [Croom]{}, S., [Jones]{}, D., & [Boshuizen]{}, C. 2004, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 5492, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. A. F. M. [Moorwood]{} & M. [Iye]{}, 389–400 , I., [da Costa]{}, G. S., [Held]{}, E. V., [Sommariva]{}, V., [Gullieuszik]{}, M., [Barbuy]{}, B., & [Ortolani]{}, S. 2012, , 540, A27 , L. 1977, in Evolution of Galaxies and Stellar Populations, ed. B. M. [Tinsley]{} & R. B. G. [Larson]{}, D. Campbell, 219 , J., [Sneden]{}, C., [Ivans]{}, I. I., [Kraft]{}, R. P., [Shetrone]{}, M. D., & [Smith]{}, V. V. 2003, , 125, 2018 , M. F., [Cutri]{}, R. M., [Stiening]{}, R., [Weinberg]{}, M. D., [Schneider]{}, S., [Carpenter]{}, J. M., [Beichman]{}, C., [Capps]{}, R., [Chester]{}, T., [Elias]{}, J., [Huchra]{}, J., [Liebert]{}, J., [Lonsdale]{}, C., [Monet]{}, D. G., [Price]{}, S., [Seitzer]{}, P., [Jarrett]{}, T., [Kirkpatrick]{}, J. D., [Gizis]{}, J. E., [Howard]{}, E., [Evans]{}, T., [Fowler]{}, J., [Fullmer]{}, L., [Hurt]{}, R., [Light]{}, R., [Kopan]{}, E. L., [Marsh]{}, K. A., [McCallon]{}, H. L., [Tam]{}, R., [Van Dyk]{}, S., & [Wheelock]{}, S. 2006, , 131, 1163 , G. H. 1987, , 99, 67 , G. H. & [Mateo]{}, M. 1990, , 353, 533 , V. V., [Suntzeff]{}, N. B., [Cunha]{}, K., [Gallino]{}, R., [Busso]{}, M., [Lambert]{}, D. L., & [Straniero]{}, O. 2000, , 119, 1239 , J. P., [Martell]{}, S. L., [Beers]{}, T. C., & [Lee]{}, Y. S. 2011, , 142, 126 , C. 1973, , 184, 839 , C. & [Crocker]{}, D. A. 1988, , 335, 406 , C., [Johnson]{}, J., [Kraft]{}, R. P., [Smith]{}, G. H., [Cowan]{}, J. J., & [Bolte]{}, M. S. 2000, , 536, L85 , C., [Kraft]{}, R. P., [Shetrone]{}, M. D., [Smith]{}, G. H., [Langer]{}, G. E., & [Prosser]{}, C. F. 1997, , 114, 1964 , C., [Lawler]{}, J. E., [Cowan]{}, J. J., [Ivans]{}, I. I., & [Den Hartog]{}, E. A. 2009, , 182, 80 , J. S., [Kraft]{}, R. P., [Sneden]{}, C., [Preston]{}, G. W., [Cowan]{}, J. J., [Smith]{}, G. H., [Thompson]{}, I. B., [Shectman]{}, S. A., & [Burley]{}, G. S. 2011, , 141, 175 , P. B. & [Pancino]{}, E. 2008, , 120, 1332 , P. & [D’Antona]{}, F. 2005, , 635, L149 , E., [McMillan]{}, S. L. W., [D’Antona]{}, F., & [D’Ercole]{}, A. 2013, , 429, 1913 , S., [Piotto]{}, G., & [Gratton]{}, R. G. 2009, , 499, 755 , W. & [Brault]{}, J. W. 1988, , 38, 707 , M. E. & [Lawler]{}, J. E. 1997, , 110, 163 , M. E., [Lawler]{}, J. E., & [Nave]{}, G. 2000, , 66, 363 , C. C., [Hill]{}, V., [Sobeck]{}, J., & [Carretta]{}, E. 2013, , 553, A47 , D., [Aoki]{}, W., [Lambert]{}, D. L., & [Paulson]{}, D. B. 2006, , 639, 918 , D. & [Grundahl]{}, F. 2008, , 672, L29 , D., [Grundahl]{}, F., [Nissen]{}, P. E., [Jensen]{}, H. R., & [Lambert]{}, D. L. 2005, , 438, 875 , D., [Karakas]{}, A. I., [Lambert]{}, D. L., [Chieffi]{}, A., & [Limongi]{}, M. 2008, , 689, 1031 , D., [Lambert]{}, D. L., [Paulson]{}, D. B., & [Carney]{}, B. W. 2008, , 673, 854 \[lastpage\] [^1]: E-mail: [email protected] [^2]: Stromlo Fellow [^3]: Based in part on data collected at Subaru Telescope, which is operated by the National Astronomical Observatory of Japan. This paper includes data gathered with the 6.5 meter Magellan Telescopes located at Las Campanas Observatory, Chile. [^4]: http://www.aao.gov.au/2df/aaomega/aaomega\_2dfdr.html [^5]: AXXX names are from @arp55, CRXXX names are from @cudworth87, HXXX names are from @harris75, and NR XXX names are from the @grundahl99 photometry. [^6]: Probability of cluster membership from @cudworth87. [^7]: 1 = stars which lie on the anomalous giant branch selected from the $v$ versus $u-y$ CMD. All other stars lie on the canonical RGB. [^8]: \[note1\]NR 184 is a UV-bright star, NR 648 is a BHB star, NR 707 is a RHB (or AGB) star and NR 82 is an AGB star. [^9]: [iraf]{} is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation. [^10]: http://code.obs.carnegiescience.edu/mike [^11]: http://www.lco.cl/telescopes-information/magellan/instruments/mike/iraf-tools/iraf-mtools-package [^12]: AXXX names are from @arp55, CRXXX names are from @cudworth87, HXXX names are from @harris75, and NR XXX names are from the @grundahl99 photometry. [^13]: Probability of cluster membership from @cudworth87. [^14]: 1 = stars which lie on the anomalous giant branch selected from the $v$ versus $u-y$ CMD. All other stars lie on the canonical RGB. [^15]: M12 = Magellan Telescope 2012 08 26, S11 = Subaru Telescope 2011 08 03, S13 = Subaru Telescope 2013 07 17. [^16]: \[m/H\] refers to the metallicity used to generate the model atmosphere. [^17]: NR 60 is a likely AGB star. [^18]: The digits to the left of the decimal point are the atomic number. The digit to the right of the decimal point is the ionization state (“0” = neutral, “1” = singly ionised). [^19]: A = $\log gf$ values taken from @yong05 where the references include @denhartog03, @ivans01, @kurucz95, @prochaska00, @ramirez02; B = @gratton03; C = Oxford group including @blackwell79feb [@blackwell79fea; @blackwell80fea; @blackwell86fea; @blackwell95fea]; D = @biemont91; E1 = @fuhr09, using line component patterns for hfs/IS from @kurucz95; E2 = @roederer12c; E3 = @fuhr09; E4 = @biemont11; E5 = @biemont81; E6 = @ljung06; E7 = @whaling88; E8 = @fuhr09, using hfs/IS from @mcwilliam98; E9 = @lawler01a, using hfs from @ivans06; E10 = @lawler09; E11 = @li07, using hfs from @sneden09; E12 = @ivarsson01, using hfs from @sneden09; E13 = @denhartog03, using hfs/IS from @roederer08 when available; E14 = @lawler06, using hfs/IS from @roederer08 when available; E15 = @lawler01b, using hfs/IS from @ivans06; E16 = @roederer12b; E17 = @denhartog06; E18 = @lawler01c, using hfs from @lawler01d [@lawler09]; E19 = @wickliffe00; E20 = @lawler08; E21 = @wickliffe97; E22 = @sneden09 for  and hfs/IS; E23 = @lawler07; E24 = @biemont00, using hfs/IS from @roederer12b.\ This table is published in its entirety in the electronic edition of the paper. A portion is shown here for guidance regarding its form and content. [^20]: Here and throughout the paper, we use the values found in the 2010 version of the catalogue (available online) rather than the values in the original @harris96 paper. [^21]: This value is the observed dispersion and is not corrected for the contribution from velocity errors. [^22]: www3.cadc-ccda.hia-iha.nrc-cnrc.gc.ca/community/STETSON/standards/ [^23]: As discussed in @cudworth87, the proper motions membership probabilities are based on a relative system with the zero point set by the mean of all the measurements. Since the M2 sample is dominated by cluster members (see Table II of Cudworth & Rauscher 1987) whose absolute proper motions will be small given the large distance, any relatively distant field star, as distinct from nearby dwarfs, will likely also have a small proper motion and therefore potentially be assigned an erroneous high membership probability. [^24]: For a given element in a set of stars, the “average measurement error” is the average of the Total Error presented in Table \[tab:abun\]. [^25]: These values are the standard deviation. [^26]: This is \[[Fe[i]{}]{}/H\] or \[[Fe[ii]{}]{}/H\], not $<$\[X/Fe\]$>$. [^27]: These are two well-studied unrelated clusters of similar metallicity, \[Fe/H\] $\simeq$ $-$1.2, and M4 is known to exhibit a moderate enhancement in $s$-process element abundances compared to M5 [@ivans99; @ivans01]. As in @roederer11, subtracting the abundances for M5 from M4 attempts to quantify the $s$-process contribution to M4. [^28]: We adopt the indices as defined by @bisterzo10: the ratios of light ($ls$) and heavy ($hs$)  abundances are \[$ls$/Fe\] $\equiv \frac{1}{2}$(\[Y/Fe\] $+$ \[Zr/Fe\]) and \[$hs$/Fe\] $\equiv \frac{1}{3}$(\[La/Fe\] $+$ \[Nd/Fe\] $+$ \[Sm/Fe\]). These include elements at the first (Sr, Y, Zr) and second (Ba, La, Ce, Pr, Nd)  peaks; Pb is the sole representative of the third  peak. Similarly, \[$hs$/$ls$\] $\equiv$ \[$hs$/Fe\] $-$ \[$ls$/Fe\].

ArXiv

--- abstract: 'The tropical arithmetic operations on $\R$ are defined by $a\op b=\min\{a,b\}$ and $a\ot b=a+b$. Let $A$ be a tropical matrix and $k$ a positive integer, the problem of Tropical Matrix Factorization (TMF) asks whether there exist tropical matrices $B\in\R^{m\times k}$ and $C\in\R^{k\times n}$ satisfying $B\ot C=A$. We show that no algorithm for TMF is likely to work in polynomial time for every fixed $k$, thus resolving a problem proposed by Barvinok in 1993. TMF is also shown to be hard for matrices with bounded tropical rank. Proving that TMF can be solved by a polynomial-time algorithm if $k\leq3$, we answer a question posed by Develin, Santos, and Sturmfels. Another question they have posed asks whether every tropical matrix of factor rank $k$ has a rank-$k$ submatrix of size at most $N(k)\times N(k)$; we answer this question in the negative for every $k>4$.' author: - Yaroslav Shitov title: | The complexity of tropical\ matrix factorization --- Introduction ============ The *tropical semiring* is the set $\R$ of real numbers equipped with the operations of tropical addition and tropical multiplication, which are defined by $a\op b=\min\{a,b\}$, $a\ot b=a+b$. The tropical semiring is essentially the same structure as the *max-plus algebra*, which is the set $\R$ with the operations of maximum and sum, and is being studied since the 1960’s, when the applications in the optimization theory have been found [@Vor]. The tropical arithmetic operations on $\R$, which allow us to formulate a number of important non-linear problems in a linear-like way, arise indeed in a variety of topics in pure and applied mathematics. The study of tropical mathematics has applications in operations research [@CG], discrete event systems [@BCOQ], automata theory [@Sim], optimal control [@KM; @McE], algebraic geometry [@DSS; @EKL], and others; we refer to [@GP2] for a detailed survey of applications. A considerable number of important problems in tropical mathematics has a linear-algebraic nature. For instance, the concepts of eigenvalue and eigenvector, the theory of linear systems, and the algorithms for computing rank functions are useful for different applications [@AGG; @DSS; @HOW; @GP2]. Some applications also give rise to studying the multiplicative structure of tropical matrices [@Pin], and in this context, the Burnside-type problems are important [@Gau; @Sim]. Another interesting problem is to study the subgroup structure of the semigroup of tropical matrices under multiplication [@IJK; @JK2]. In our paper, we consider the problem of matrix factorization, which is also related to the concept of factor rank of matrices over semirings [@BKS]. The study of factor rank dates back to the 1980’s [@BP2], and has now numerous applications in different contexts of mathematics. Being considered on the semiring of nonnegative matrices, the factor rank is known as nonnegative rank and has applications in quantum mechanics, statistics, demography, and others [@CR]. The factor rank of matrices over the binary Boolean semiring is also called Boolean rank and has applications in combinatorics and graph theory [@Bea; @GP]. Finally, for matrices over a field, the factor rank coincides with the classical rank function. In the context of matrices over the tropical semiring, the factor rank is also known as *combinatorial rank* [@Barv2] and *Barvinok rank* [@DSS], and the study of this notion has arisen from combinatorial optimization [@Barv]. The factor rank appears in the formulation of a number of problems in optimization, for instance, in the Traveling Salesman problem with warehouses [@Barv3]. Also, the notion of factor rank is of interest in the study of tropical geometry [@Dev2; @DSS], where the factor rank can be thought of as the minimum number of points whose tropical convex hull contains the columns of a matrix. Let us define the factor rank function for tropical matrices, assuming that multiplication of tropical matrices is understood as ordinary matrix multiplication with $+$ and $\cdot$ replaced by the tropical operations $\op$ and $\ot$. \[factrank\] The *factor rank* of a tropical matrix $A\in\R^{m\times n}$ is the smallest integer $k$ for which there exist tropical matrices $B\in\R^{m\times k}$ and $C\in\R^{k\times n}$ satisfying $B\ot C=A$. The most straightforward way of computing the factor rank is based on the quantifier elimination algorithm for the theory of reals with addition and order [@FR]. Indeed, Definition \[factrank\] allows us to define the set of all $m$-by-$n$ matrices with factor rank $k$ by a first-order order formula. We then employ the decision procedure based on the quantifier elimination algorithm provided in [@FR] to check whether a given $m$-by-$n$ matrix indeed has factor rank $k$. However, the computational complexity of quantifier elimination makes the algorithm mentioned unacceptable for practical use. Another algorithm for computing the factor rank is given by Develin in the paper [@Dev2], where he develops the theory of tropical secant varieties. He characterizes the factor rank from a point of view of tropical geometry, and the characterization obtained provides an algorithm for computing the factor rank. Unfortunately, neither the algorithm by Develin nor any other algorithm is likely to compute the factor rank of a tropical matrix in polynomial time. Indeed, the problem of computing the factor rank is NP-hard even for tropical $01$-matrices, see [@DSS]. In other words, the general problem of *Tropical Matrix Factorization* (*TMF*), which asks whether a given matrix $A$ and a given integer $k$ are such that $A=B\ot C$ for some $B\in\R^{m\times k}$ and $C\in\R^{k\times n}$, turns out to be NP-hard. Besides the general problem of computing the factor rank, the problem of detecting matrices with fixed factor rank deserved attention. Certain hard problems of combinatorial optimization admit fast solutions if the input matrices are required to have factor rank bounded by a fixed number [@Barv; @Barv2]. The Traveling Salesman problem (TSP) also admits a fast solution if we require the distance matrix to have a fixed factor rank [@Barv3; @Barv4]; this special case of TSP is also known as TSP with warehouses [@Barv3]. Matrices with bounded factor rank arise naturally in the problems mentioned and in a number of other problems in combinatorial optimization [@Barv] and tropical geometry [@DSS]. These considerations led to the following interesting question on tropical matrix factorizations. [@Barv; @Barv2]\[que1\] Does there exist an algorithm that solves TMF for every fixed $k$ in polynomial time? In [@Barv], Barvinok expected that Question \[que1\] can be answered in the positive, this question has also been mentioned in [@Barv2]. Further investigations on the problem of determining matrices with fixed factor rank have been carried out in [@CRW; @Dev1; @Dev2; @DSS; @HJ]. However, the problem remained open even in the case $k=3$ and has been formulated again in [@DSS]. [@DSS Section 8, Question 2]\[quemain\] Is there a polynomial-time algorithm that decides whether a given tropical matrix has factor rank $3$? In [@DSS], a number of related problems has also been posed. The following question asks whether the set of matrices with fixed factor rank can be defined by ranks of the submatrices of bounded size. \[que2\][@DSS Section 8, Question 3a] Is there a number $N(k)$ such that if all minors of a tropical matrix $A$ of size at most $N(k)$ have factor rank at most $k$ then $A$ itself has factor rank at most $k$? Assume that Question \[que2\] is answered in the positive for some $k$. To solve Problem \[probmain\], it is then sufficient to check whether every minor of size at most $N(k)\times N(k)$ has factor rank at most $k$. The number of minors of fixed size is easily seen to be polynomial in the size of a matrix, and each of these minors can be treated in fixed time thanks to, say, quantifier elimination algorithm. Thus we see that, indeed, as noted in [@DSS], a positive answer to Question \[que2\] would allow to construct a fast algorithm for recognizing tropical matrices with factor rank $k$. Besides the factor rank, there are other important rank concepts for tropical matrices, see [@AGG; @DSS]. One of them is presented by the function of tropical rank, which arises naturally in tropical linear algebra and tropical geometry [@DSS]. First, we define the *tropical permanent* (which is also called the *tropical determinant*) of a tropical matrix $D\in\R^{n\times n}$ as $$\label{permanent} \textrm{perm}(D)=\min\limits_{\sigma\in\S_n} \left\{D_{1,\sigma(1)}+\ldots+D_{n,\sigma(n)}\right\},$$ where $\S_n$ stands for the symmetric group on $\{1,\ldots,n\}$. $D$ is said to be *tropically singular* if the minimum in (\[permanent\]) is attained at least twice. The *tropical rank* of a tropical matrix $A$ is the largest integer $r$ such that $A$ has a non-singular $r$-by-$r$ submatrix. Now we can formulate another problem related to the complexity of tropical matrix factorization. \[que3\][@DSS Section 8, Question 3b] Is there a polynomial-time algorithm for the factor rank of matrices with bounded tropical rank? The progress in solving the problems we have mentioned has mostly been based on studying matrices with factor rank at most $2$, and the set of such matrices is now indeed well studied. First, the TMF problem with $k\leq2$ can be solved by a linear-time algorithm, see [@CRW; @DSS]. Further, it has been proven in [@DSS] that the factor rank of a matrix is at most $2$ if and only if all its $3$-by-$3$ minors have factor ranks at most $2$. The set of $d$-by-$n$ matrices with factor rank $2$ has been studied as a simplicial complex in [@Dev1]. This set has also been studied from the topological point of view in [@HJ], and the space of $d$-by-$n$ matrices of factor rank two modulo translation and rescaling has been shown to form a manifold. For $k\geq3$, the question of fast algorithm for recognizing tropical matrices with factor rank $k$ remained open. In our paper, we answer Question \[que1\], showing that no algorithm is likely to solve TMF in polynomial time for every fixed $k$. More precisely, we show that the problem of recognizing tropical matrices with factor rank $8$ is NP-hard. However, it turns out that the answer for Question \[quemain\] is positive. We also answer Question \[que2\] in the negative for every $k\geq4$ and obtain a negative answer for Question \[que3\]. The following notation is used throughout our paper. By $A_{ij}$ we will denote the $(i,j)$th entry of a matrix $A$, by $A[r_1,\ldots,r_p|c_1,\ldots,c_q]$ the submatrix of $A$ formed by the rows with labels $r_1,\ldots,r_p$ and columns with $c_1,\ldots,c_q$. We will say that tropical matrices $B,C\in\R^{m\times n}$ coincide modulo *scaling* if there exist real numbers $\alpha_1,\ldots,\alpha_m$, $\beta_1,\ldots,\beta_n$ such that $B_{ij}=C_{ij}+\alpha_i+\beta_j$ for all $i$ and $j$. Straightforwardly, $B$ and $C$ have the same tropical and the same factor ranks in that case. Preliminaries {#prelim} ============= In this section, we give a more specific formulation for the TMF problem and prove the preliminary results on computational complexity of this problem. We describe the construction of the extended tropical semiring $\R\cup\{\i\}$, which will be helpful for further considerations, and we prove a number of auxiliary results on the factor rank of matrices over $\R\cup\{\i\}$. Tropical matrix $k$-factorization --------------------------------- We will consider a family of decision problems which specify a value of inner dimension $k$ in the general formulation of the TMF problem. We require an input matrix to consist of integers in these problems, since this assumption will be convenient in our further considerations. For a positive integer $k$, the problem we deal with is as follows. \[facrankprob\]\[probmain\] TROPICAL MATRIX $k$-FACTORIZATION ($k$-TMF). Given a tropical matrix $A\in\Z^{m\times n}$. Question: Do there exist tropical matrices $B\in\R^{m\times k}$ and $C\in\R^{k\times n}$ satisfying $B\ot C=A$? In the input of Problem \[probmain\], as well as throughout our paper, we assume that an integer matrix is presented in the usual way as a rectangular array of numbers written in the decimal system. Namely, matrices will be considered to be words over the alphabet that consists of the characters $"$0$"$, $"$1$"$, $\ldots$, $"$9$"$, $"$+$"$, $"$-$"$, a space, and an end-of-line symbol. In particular, the size of the input in Problem \[probmain\] equals $\sum\limits_{i,j}\left[\log_{10}\max\{|A_{ij}|,1\}\right]+3mn,$ where $[x]$ denotes the integer part of $x$. Now we will prove that the TMF problem belongs to the class NP if an input matrix is required to consist of integers. We need the following technical lemma. \[NP-tmf-lem\] Let a matrix $A\in\Z^{m\times n}$ have factor rank at most $k$. Let $g$ and $l$ denote, respectively, the greatest and the least elements of $A$, set also $h=|g|+|l|$. Then there exist matrices $B\in\Z^{m\times k}$ and $C\in\Z^{k\times n}$ such that $B\ot C=A$ and $|B_{i\tau}|\leq h$, $|C_{\tau j}|\leq h$ for any triple of indexes $(i, j, \tau)$. By Definition \[factrank\], there are matrices $B'\in\R^{m\times k}$ and $C'\in\R^{k\times n}$ satisfying $B'\ot C'=A$. Consider matrices $B''$ and $C''$ defined by $B''_{i\tau}=B'_{i\tau}-\{B'_{i\tau}\}$, $C''_{\tau j}=C'_{\tau j}+\{-C'_{\tau j}\}$, where $\{x\}=x-[x]$ is the fractional part of $x$. The matrices $B''$ and $C''$ are then integer, and we have $|(B''_{i\tau}+C''_{\tau j})-(B'_{i\tau}+C'_{\tau j})|<1$. Therefore, $B'_{i\tau}+C'_{\tau j}=A_{ij}$ implies $B''_{i\tau}+C''_{\tau j}=A_{ij}$, and from $B'_{i\tau}+C'_{\tau j}\geq A_{ij}$ it follows that $B''_{i\tau}+C''_{\tau j}\geq A_{ij}$. Thus we have that $B'\ot C'=A$ implies $B''\ot C''=A$. Further, we subtract $b_\tau$, the least element of the $\tau$th column of $B''$, from every entry of the $\tau$th column of $B''$ and add $b_\tau$ to every entry of the $\tau$th row of $C''$. The matrices $B$ and $C$ obtained satisfy $B\ot C=A$, and zero appears as the minimal element of every column of $B$. The definition of matrix multiplication then shows that every entry of $C$ is greater than or equal to $l$. Finally, those entries of $B$ and $C$ that are greater than $|g|+|l|$ can be then replaced by $|g|+|l|$ without changing the product $B\ot C$. Now we can prove our first results concerning the computational complexity of tropical matrix factorization. \[NP-tmf\] Given a tropical matrix $A\in\Z^{m\times n}$ and a positive integer $k$. The problem of deciding whether $A$ has factor rank at most $k$ belongs to the class NP. The factor rank of $A$ is at most $\min\{m,n\}$ (see [@DSS Proposition 2.1]), so the problem can be solved immediately if $k\geq\min\{m,n\}$. For $k<\min\{m,n\}$, we apply Lemma \[NP-tmf-lem\]. If $A$ has factor rank at most $k$, then there exist matrices $B\in\Z^{m\times k}$ and $C\in\Z^{k\times n}$ that satisfy $B\ot C=A$ and require at most $(m+n)k\left(\log_{10}h+3\right)$ symbols to be written ($h$ denotes here the sum of the absolute values of the greatest and the least elements of $A$). The size of the input data of the problem considered is at least $\log_{10}\frac{h}{2}+mn$, so the result follows. Theorem \[NP-tmf\] shows that TMF, the general problem of tropical matrix factorization, belongs to NP if the input matrices are assumed to consist of integers. The problem $k$-TMF, introduced in this section, is thus in NP as well. \[NP-k-tmf\] The $k$-TMF problem belongs to NP. Follows from Theorem \[NP-tmf\]. Extended tropical semiring -------------------------- It will be useful for our further considerations to extend the tropical semiring by an infinite positive element, which we denote by $\i$. We also write $\Ro$ for $\R\cup\{\i\}$ and assume $a\op\i=a$, $a\ot\i=\i$ for $a\in\Ro$. Defining the factor rank, one can now think of a tropical matrix $A\in\R^{m\times n}$ as a matrix over $\Ro$, and allow matrices $B$ and $C$ from Definition \[factrank\] to contain infinite elements. However, it turns out that the rank function defined in this way is the same as that defined with respect to Definition \[factrank\]. Indeed, if matrices $A\in\R^{m\times n}$, $B\in\Ro^{m\times k}$, $C\in\Ro^{k\times n}$ satisfy $A=B\ot C$, then we can replace infinite entries of $B$ and $C$ with a sufficiently large real without changing the product of the matrices. We can therefore extend Definition \[factrank\] to the case of matrices over $\Ro$ and define the *factor rank* of $A\in\Ro^{m\times n}$ to be the smallest integer $k$ for which there exist $B\in\Ro^{m\times k}$ and $C\in\Ro^{k\times n}$ satisfying $A=B\ot C$. The following technical lemma will play an important role in the considerations of our paper. \[Ro-&gt;R\] Let zero appear as a minimal element of every row and every column of a matrix $A\in\Ro^{m\times n}$. Consider the matrix $A'$ obtained from $A$ by replacing every infinite entry with the number $2g+1$, where $g$ stands for the maximal finite entry of $A$. Then the factor ranks of $A$ and $A'$ are the same. Consider matrices $B'\in\R^{m\times k}$ and $C'\in\R^{k\times n}$ satisfying $B'\ot C'=A'$. For $t\in\{1,\ldots,k\}$, we subtract $b_t$, the least element of the $t$th column of $B$, from every entry of that column and add $b_t$ to every entry of the $t$th row of $C$. The matrices $B''$ and $C''$ obtained satisfy $B''\ot C''=A'$, and zero appears as a minimal element of every row of $B''$ and of every column of $C''$. Further, we replace by $\i$ every entry of $B''$ and $C''$ that is greater than $g$, and we denote the matrices obtained by $B$ and $C$. It is then easy to check that $B\ot C=A$. So we have proven that the factor rank of $A$ is at most that of $A'$. It is thus sufficient to note that $A'=E\ot A$, where $E$ stands for an $m$-by-$m$ tropical matrix satisfying $E_{ij}=0$ if $i=j$ and $E_{ij}=2g+1$ otherwise. The following easy observation will also be useful. \[block\] The factor rank of a tropical matrix $A\in\R^{m\times n}$ is one less than that of the matrix $A'\in\R^{(m+1)\times (n+1)}$ defined by $A'_{ij}=A_{ij}$, $A'_{i,n+1}=A'_{m+1,j}=\i$ for $i\leq m$, $j\leq n$, and $A'_{m+1,n+1}=0$. The rest of our paper is organized as follows. In Section 3, we consider the $k$-TMF problem for large enough values of inner dimension $k$. We show that $k$-TMF is NP-complete for $k\geq8$, thus giving a negative answer for Question \[que1\]. As a corollary of this result, we show that the problem of deciding whether the factor rank of a matrix $A$ exceeds $8$ is NP-hard even if $A$ is assumed to have tropical rank at most $8$; in particular, this gives a negative answer for Question \[que3\]. Note that this situation is in contrast with that of the Gondran–Minoux rank, which can be computed in polynomial time if a given matrix has bounded tropical rank [@our]. In Section 4, we consider the problem of tropical matrix factorization for small values of $k$. We construct an algorithm that requires at most $O\left((m+n)^3mn\log(mn)+m^3n^3\right)$ arithmetic operations and decides whether a given matrix $A\in\R^{m\times n}$ has factor rank at most $3$. This answers Question \[quemain\] in the affirmative and shows that the $3$-TMF problem belongs to the class P. In Section 5, we consider the question of whether matrices with fixed factor rank are defined by ranks of their bounded-size minors, and we give a negative answer for Question \[que2\] in the case when $k\geq4$. $8$-TMF is NP-complete {#sec2} ====================== In this section, we consider the problem of tropical matrix factorization for sufficiently large but fixed values of $k$, and show that this problem remains hard if the dimension $k$ is assumed to be bounded. We show that no algorithm is likely to solve $k$-TMF efficiently if $k\geq8$. More precisely, we construct a polynomial-time reduction from an NP-complete problem to $k$-TMF, thus proving that $k$-TMF is NP-hard. In other words, we show that efficient solution of $k$-TMF would provide efficient solutions to many problems that are of interest for a long time and widely believed to be hard [@GJ]. Therefore, the results of this section give a negative answer for Question \[que1\]. As a corollary, we obtain a negative answer for Question \[que3\]. The NP-complete problem which we reduce to $k$-TMF is known as SET SPLITTING. This problem has been considered in [@GJ] and is formulated as follows. SET SPLITTING. Given a set $S$ and a family $\Ccal$ of subsets of $S$. Question: Are there sets $\Phi_1$ and $\Phi_2$ such that $\Phi_1\cap\Phi_2=\es$, $\Phi_1\cup\Phi_2=S$, and no subset from $\Ccal$ is entirely contained in either $\Phi_1$ or $\Phi_2$? Reformulation of SET SPLITTING and some specific conventions ------------------------------------------------------------ Our reduction to $k$-TMF, however, does not go from SET SPLITTING directly. Let us introduce the following reformulation of SET SPLITTING, which is useful for our considerations. \[ssref\] SET SPLITTING REFORMULATED. Given a nondecreasing integer sequence $\si_0=0,\si_1,\ldots,\si_m$ and a sequence of pairwise disjoint sets $\Rcal_1,\ldots,\Rcal_n$ such that $\Rcal_1\cup\ldots\cup\Rcal_n=\{1,\ldots,\si_m\}$. Question: Are there sets $\Phi_1$ and $\Phi_2$ such that $\Phi_1\cap\Phi_2=\es$, $\Phi_1\cup\Phi_2=\{1,\ldots,n\}$, and the sets $$\left(\bigcup_{i\in\Phi_1}\Rcal_i\right)\bigcap\left\{x\left|\right.\si_{\mu-1}<x\leq\si_\mu\right\} \mbox{ and } \left(\bigcup_{i\in\Phi_2}\Rcal_i\right)\bigcap\left\{x\left|\right.\si_{\mu-1}<x\leq\si_\mu\right\}$$ are both nonempty for every $\mu\in\{1,\ldots,m\}$? Problem \[ssref\], as well as SET SPLITTING, turns out to be NP-complete. \[ssrefnpc\] Problem \[ssref\] is NP-complete. Let $S=\{s_1,\ldots,s_n\}$ be a set and $\Ccal_1,\ldots,\Ccal_m$ a sequence of its subsets, by $\Psi$ we denote the set of all pairs $(s_i,\Ccal_j)$ satisfying $s_i\in\Ccal_j$. We set $\sigma_0=0$ and $\sigma_\mu=|\Ccal_1|+\ldots+|\Ccal_\mu|$ for every $\mu\in\{1,\ldots,m\}$. Then it is easy to construct a bijection $\nu:\Psi\rightarrow\{1,\ldots,\sigma_m\}$ mapping every $(s_i,\Ccal_j)\in\Psi$ to some integer $x$ satisfying $\sigma_{j-1}<x\leq\sigma_j$. Finally, for $i\in\{1,\ldots,n\}$, we define $\Rcal_i$ to consist of all numbers whose preimage under $\nu$ has $s_i$ as the first entry. These settings can be straightforwardly checked to give a polynomial reduction from SET SPLITTING (which is known to be NP-complete, see [@GJ]) to Problem \[ssref\]. On the other hand, the fact that Problem \[ssref\] is in NP is clear. Our way of proving that $8$-TMF is NP-hard is based on the polynomial-time reduction from Problem \[ssref\]. In order to describe this reduction explicitly, we need to specify some conventions that will be used throughout this section. We will denote the input data of Problem \[ssref\] by $\Sigma=(\sigma_0,\ldots,\sigma_m)$, $\Rcal=(\Rcal_1,\ldots,\Rcal_n)$. The conventions to be used are given by the following remarks. \[diff2\] Note that if $|\sigma_i-\sigma_j|<2$ for some distinct $i$ and $j$, then the answer for Problem \[ssref\] is trivially ’no’. Consequently, Problem \[ssref\] remains NP-complete under an additional assumption that $|\sigma_i-\sigma_j|<2$ implies $i=j$. In what follows, we suppose that $\Sigma$ satisfies this assumption. \[nu\] The formulation of Problem \[ssref\] requires that for every $u\in\{1,\ldots,\sigma_m\}$, there exists a unique $\nu=\nu(u)$ such that $u\in\Rcal_\nu$. \[rhogammaGsigma\] For any $u\in\{1,\ldots,\sigma\}$ and $v\in\{1,\ldots,\sigma-m\}$, we set $$\label{CuRv} \gamma_u=20u-10\lambda-9,\mbox{ $ $ $ $ $ $} \rho_v=20v+10\mu+1,$$ where $\lambda$ and $\mu$ stand for the maximal subscripts satisfying $u\geq\sigma_\lambda$ and $v>\sigma_\mu-\mu$, respectively. We will also write $\sigma$ for $\sigma_m$, and $G$ for the number $21\sigma$. Encoding the input of Problem \[ssref\] by a matrix --------------------------------------------------- Let us now begin to describe the reduction from Problem \[ssref\] to $8$-TMF. Namely, this subsection is devoted to the construction of a tropical matrix $A=A(\Sigma,\Rcal)$ which we assign to the input data $(\Sigma,\Rcal)$ of Problem \[ssref\]. By our setting, the matrix $A$ will have $\sigma-m+n+9$ rows and $\sigma+10$ columns. The following convention will be useful to present this matrix in a more concise way: The rows of $A$ will be labeled with the elements of the set $$\Ical=\{1\pi,\ldots,9\pi,1\bu,\ldots,n\bu,1\ch,\ldots,(\sigma-m)\ch\},$$ the columns with the elements of $$\Jcal=\{1\be,\ldots,10\be,1\di,\ldots,\sigma\di\}.$$ Now we can define the entries of $A(\Sigma,\Rcal)$. \[A\] It is easy enough to show that the function assigning the matrix $A(\Sigma,\Rcal)$ to a tuple $(\Sigma,\Rcal)$ has the polynomial complexity. \[polyred\] Given a tuple $(\Sigma,\Rcal)$. The number of operations needed to construct the matrix $A(\Sigma,\Rcal)$ is polynomial in the size of the data. The numbers of the rows and columns of $A(\Sigma,\Rcal)$ are given by linear functions of variables appearing in the data. On the other hand, every entry of $A(\Sigma,\Rcal)$ can clearly be computed in bounded number of operations. We will further consider tropical factorizations of the matrix $A=A(\Sigma,\Rcal)$. In order to adapt the notation of the row and column indexes of $A$ to the product of matrices, we need the following convention. \[BCnotat\] Throughout this section, for considering matrices $B$ and $C$ satisfying $A(\Sigma,\Rcal)=B\ot C$, we will use the following notation. The elements of $\Ical$ will label the rows of $B$, the elements of $\Jcal$ the columns of $C$, the numbers $1,\ldots,k$ the columns of $B$ as well as the rows of $C$. In our settings, $A=B\ot C$ means that $A_{ij}=\min_{\tau=1}^k \{B_{i\tau}+C_{\tau j}\}$ for all $i\in\Ical$ and $j\in\Jcal$. The following lemma shows that the factor rank of $A(\Sigma,\Rcal)$ does not exceed $8$ if the question of Problem \[ssref\] is answered in the affirmative on the query $(\Sigma,\Rcal)$. \[fromlemma\] Let an input $(\Sigma,\Rcal)$ be such that the answer in Problem \[ssref\] is ’yes’. Then there are matrices $B\in\Ro^{|\Ical|\times 8}$ and $C\in\Ro^{8\times |\Jcal|}$ such that $A(\Sigma,\Rcal)=B\ot C$. The assumption of the lemma implies that there exist $\Phi_1$ and $\Phi_2$ such that $\Phi_1\cap\Phi_2=\es$, $\Phi_1\cup\Phi_2=\{1,\ldots,n\}$, and the sets $$\left(\bigcup_{i\in\Phi_1}\Rcal_i\right)\bigcap\{\sigma_{\mu-1}+1,\ldots,\sigma_\mu\} \mbox{ and } \left(\bigcup_{i\in\Phi_2}\Rcal_i\right)\bigcap\{\sigma_{\mu-1}+1,\ldots,\sigma_\mu\}$$ are both nonempty for every $\mu\in\{1,\ldots,m\}$. Then for every $\mu$, the set $\{\nu(\sigma_{\mu-1}+1),\ldots,\nu(\sigma_\mu)\}$ has nonempty intersections with both $\Phi_1$ and $\Phi_2$ (recall that, with respect to Remark \[nu\], $\nu(u)$ denotes a unique number for which the condition $u\in\Rcal_{\nu(u)}$ is satisfied). By $\Hcal_\mu$ we will denote the greatest number of the set $\{\sigma_{\mu-1}+1,\ldots,\sigma_\mu\}$ for which the numbers $\nu(\Hcal_\mu)$ and $\nu(\sigma_\mu)$ are neither both in $\Phi_1$ nor both in $\Phi_2$. Now we can present the matrices $B$ and $C$. In the rest of our proof, we will denote an arbitrary number from $\{1,\ldots,\sigma\}$ by $u$, from $\{1,\ldots,\sigma-m\}$ by $v$, from $\{1,\ldots,n\}$ by $\eta$; also, $\chi$ will stand for either $1$ or $2$. As in Definition \[A\], $\rho_u$, $\gamma_v$, and $G$ stand for the numbers defined in Remark \[rhogammaGsigma\]. In items B4, B5 and C3, we assume that the index $\mu$ runs over the set $\{1,\ldots,m\}$. B1 : The matrix $B[1\pi,\ldots,9\pi|1,\ldots,8]$ is equal to $$\label{B[..|..]}\left( \begin{array}{cccccccc} 2&2&0&2&2&2&\i&\i\\ 0&2&2&2&2&2&\i&\i\\ 2&0&2&2&2&2&\i&\i\\ 2&2&\i&0&2&2&\i&\i\\ \i&\i&\i&\i&0&2&\i&\i\\ 0&0&\i&\i&0&0&\i&\i\\ \i&\i&\i&\i&\i&0&\i&\i\\ \i&\i&\i&\i&\i&\i&0&\i\\ \i&\i&\i&\i&\i&\i&\i&0 \end{array} \right).$$ B2 : We take $B_{\eta\bu,\chi}=0$ if $\eta\in\Phi_\chi$, and $B_{\eta\bu,\chi}=\i$ if $\eta\notin\Phi_\chi$; take also $B_{\eta\bu,3}=0$, $B_{\eta\bu,4}=B_{\eta\bu,5}=\i$, $B_{\eta\bu,6}=2$, $B_{\eta\bu,7}=2\eta G$, and $B_{\eta\bu,8}=2(n+1-\eta)G$. B3 : We set $B_{v\ch,3}=B_{v\ch,7}=B_{v\ch,8}=\i$, $B_{v\ch,4}=B_{v\ch,5}=\rho_{v}$, $B_{v\ch,6}=0$. B4 : If $g\in\{\sigma_{\mu-1}-\mu+2,\ldots,\sigma_\mu-\mu\}$ is less than $\Hcal_\mu-\mu+1$, then we set $B_{g\ch,\chi}=10g-7$; if $g$ is greater than $\Hcal_\mu-\mu+1$, then $B_{g\ch,\chi}=10g-8$. B5 : Finally, if $g=\Hcal_\mu-\mu+1$, then we take $B_{g\ch,\chi}=10g-8$ if $\nu(\Hcal_\mu)\in\Phi_\chi$, and $B_{g\ch,\chi}=10g-7$ if $\nu(\Hcal_\mu)\notin\Phi_\chi$. We proceed with the description of the matrix $C$. C1 : The matrix $C[1,\ldots,8|1\be,\ldots,10\be]$ is equal to $$\label{C[..|..]}\left( \begin{array}{cccccccccc} 2&0&0&2&2&2&2&\i&\i&\i\\ 2&2&0&2&0&2&2&\i&\i&\i\\ 2&0&2&0&0&2&2&\i&\i&\i\\ 0&2&2&2&2&2&2&0&\i&\i\\ 2&2&2&2&2&0&2&0&\i&\i\\ 2&2&2&2&2&2&0&\i&\i&\i\\ \i&\i&\i&\i&\i&\i&\i&\i&0&\i\\ \i&\i&\i&\i&\i&\i&\i&G'&\i&0\\ \end{array} \right)$$ with $G'$ standing for $-2(n+1)G$. C2 : We set $C_{3u\di}=C_{6,u\di}=\gamma_u$, $C_{4u\di}=0$, $C_{5,u\di}=A_{5\pi,u\di}$, $C_{7,u\di}=-2\nu(u)G+G$, $C_{8,u\di}=-2(n+1-\nu(u))G+G$. C3 : Assume $h\in\{\sigma_{\mu-1}+1,\ldots,\sigma_\mu\}$. We set $C_{\chi, h\di}=\gamma_h$ if $\nu(h)\in\Phi_\chi$, set also $C_{\chi, h\di}=10h-2$ if $\nu(h)\notin\Phi_\chi$ and $\Hcal_\mu<h<\sigma_\mu$, finally, set $C_{\chi, h\di}=10h-3$ if $\nu(h)\notin\Phi_\chi$ and either $h\leq\Hcal_\mu$ or $h=\sigma_\mu$. It is straightforward to check that the matrices $B\in\Ro^{|\Ical|\times 8}$ and $C\in\Ro^{8\times |\Jcal|}$ are well defined. In order to complete our proof, we need to verify that indeed $B\ot C=A(\Sigma,\Rcal)$, that is, $$\label{1111}A_{ij}=\min_{\tau=1}^8 \{B_{i\tau}+C_{\tau j}\}$$ for all $i\in\Ical$ and $j\in\Jcal$. Consider the six possible cases. 1.For $i\in\{1\pi,\ldots,9\pi\}$ and $j\in\{1\be,\ldots,10\be\}$, it suffices to check that the product of the matrices (\[B\[..|..\]\]) and (\[C\[..|..\]\]) equals the matrix (\[A\[..|..\]\]). 2\. Let $i=\eta\bu$ and $j\in\{1\be,\ldots,10\be\}$. We need to verify that the row $(A_{\eta\bu,1\be},\ldots,A_{\eta\bu,10\be})$ is a product of the row of $B$ labeled by $\eta\bu$ and the tropical matrix (\[C\[..|..\]\]). Namely, we need to check that the row $\left(B_{\eta\bu,1},\,B_{\eta\bu,2},\,0,\,\i,\,\i,\,2,\,2\eta G,\,2(n+1-\eta)G\right)$ multiplied by the tropical matrix (\[C\[..|..\]\]) equals $\left(2,\,0,\,0,\,0,\,0,\,2,\,2,\,-2\eta G,\,2\eta G,\,2(n+1-\eta)G\right)$. We note that also $\{B_{\eta\bu,1},\,B_{\eta\bu,2}\}=\{0,\i\}$ by the definition of $B$, checking is then straightforward. 3\. For $i=v\ch$ and $j\in\{1\be,\ldots,10\be\}$, checking (\[1111\]) means checking that the row $(A_{v\ch,1\be},\ldots,A_{v\ch,10\be})$ is a product of the row of $B$ labeled by $v\ch$ and the tropical matrix (\[C\[..|..\]\]). We note that $\min\{B_{v\ch,1},\,B_{v\ch,2}\}\geq2$ by the definition of $B$. Then the row $\left(B_{v\ch,1},\,B_{v\ch,2},\,\i,\,\rho_v,\,\rho_v,\,0,\,\i,\,\i\right)$ multiplied by the tropical matrix (\[C\[..|..\]\]) can be straightforwardly checked to equal the row $\left(2,\,2,\,2,\,2,\,2,\,2,\,0,\,\rho_v,\,\i,\,\i\right)$. 4\. Further, assume $i\in\{1\pi,\ldots,9\pi\}$ and $j=u\di$. Observe that we have $\min\{C_{1,u\di},C_{2,u\di}\}=10u-3$ for $u=\sigma_{\mu-1}+1$, and $A_{5\pi,u\di}=10u-3$ for $u\neq\sigma_{\mu-1}+1$. Also, by our settings, $C_{1,u\di},C_{2,u\di}\in\{10u-2,10u-3,\gamma_u\}$ and $A_{5\pi,u\di}\in\{10u-3, 10u+1\}$ for any $u$. So we can check that $$\left( \begin{array}{cccccccc} 2&2&0&2&2&2&\i&\i\\ 0&2&2&2&2&2&\i&\i\\ 2&0&2&2&2&2&\i&\i\\ 2&2&\i&0&2&2&\i&\i\\ \i&\i&\i&\i&0&2&\i&\i\\ 0&0&\i&\i&0&0&\i&\i\\ \i&\i&\i&\i&\i&0&\i&\i\\ \i&\i&\i&\i&\i&\i&0&\i\\ \i&\i&\i&\i&\i&\i&\i&0 \end{array} \right)\otimes \left(\begin{array}{c}C_{1,u\di}\\ C_{2,u\di}\\ \gamma_u\\0\\ A_{5\pi,u\di}\\ \gamma_u\\G_1\\G_2\end{array}\right)= \left(\begin{array}{c}2\\2\\2\\0\\A_{5\pi,u\di}\\10u-3\\ \gamma_u\\G_1\\ G_2\end{array}\right),$$ where $G_1$ and $G_2$ stand for $-2G\nu(u)+G$ and $-2G(n+1-\nu(u))+G$, respectively. This implies that (\[1111\]) holds. 5\. Now consider $i=\eta\bu$ and $j=u\di$. In this case, we have $\min\{B_{\eta\bu,1}+C_{1,u\di},B_{\eta\bu,2}+C_{2,u\di}\}>0$, and $\min\{B_{\eta\bu,1}+C_{1,u\di},B_{\eta\bu,2}+C_{2,u\di}\}=\gamma_u$ if $\eta=\nu(u)$. We also have $\min\{B_{\eta\bu,7}+C_{7,u\di},B_{\eta\bu,8}+C_{8,u\di}\}=G-2G\left|\eta-\nu(u)\right|$ and $B_{\eta\bu,\tau'}+C_{\tau',u\di}\geq\gamma_u$ for $\tau'\in\{3,4,5,6\}$. This implies $$\min_{\tau=1}^8 \{B_{\eta\bu,\tau}+C_{\tau, u\di}\}=\min\left\{\gamma_u,G-2G\left|\eta-\nu(u)\right|\right\},$$ thus proving (\[1111\]). 6\. Finally, we assume $i=v\ch$, $j=u\di$. In this case, the minimum in (\[1111\]) computed for $\tau$ running over $\{3,\ldots,8\}$ equals $\min\{\rho_v,\gamma_u\}$. It is therefore sufficient to prove that $$\label{2222}A_{v\ch,u\di}=\min\{B_{v\ch,1}+C_{1,u\di},B_{v\ch,2}+C_{2,u\di},\rho_v,\gamma_u\}$$ for all $u$ and $v$. Note that by the definitions of $B$ and $C$, it holds that $B_{v\ch,\chi}+C_{\chi,u\di}\geq 10u+10v-11$ for $\chi\in\{1,2\}$. By $\lambda$ and $\mu$ we will denote the maximal subscripts satisfying $u\geq\sigma_\lambda$ and $v>\sigma_\mu-\mu$, respectively. Let us consider possible cases. 6.1. If $10u+10v-11\geq\min\{\rho_v,\gamma_u\}$, then the minimum in (\[2222\]) equals $\min\{\rho_v,\gamma_u\}$. Item 1 of Definition \[A\] shows that also $A_{v\ch,u\di}=\min\{\rho_v,\gamma_u\}$, proving (\[2222\]). 6.2. If the assumption of item 6.1 fails to hold, then $10u+10v-11<20u-10\lambda-9$ and $10u+10v-11<20v+10\mu+1$, so that $10v+10\lambda<10u+2<10v+10\mu+14$, and we obtain $v+\mu+1\geq u\geq v+\lambda$. In particular, we see that the assumption of item 6.1 holds if $\lambda>\mu+1$, so (\[2222\]) holds in this case as well. 6.3. Now suppose $\lambda<\mu$. The definition of $\lambda$ and $\mu$ implies $u\leq\sigma_{\lambda+1}-1$ and $v>\sigma_\mu-\mu$. By Remark \[diff2\], $\sigma_\mu-\sigma_{\lambda+1}\geq 2(\mu-\lambda-1)$. Thus we have $$u\leq\sigma_{\lambda+1}-1\leq \sigma_{\lambda+1}+\mu-\lambda-2\leq\sigma_\mu-\mu+\lambda<v+\lambda,$$ and then $10v+10\lambda>10u+2$, so $10u+10v-11>\gamma_u$. We see that the assumption of item 6.1 holds again, proving (\[2222\]). 6.4. Further, suppose $\lambda=\mu+1$. By items 6.1 and 6.2, it can be assumed without a loss of generality that $u=v+\lambda$. In this case, the definition of $\lambda$ and $\mu$ shows that $u\geq\sigma_\lambda$ and $v\leq\sigma_\lambda-\lambda$, so we have $u=v+\lambda=\sigma_\lambda$. Note that $\nu(\Hcal_{\lambda})\in\Phi_\chi$ if and only if $\nu(\sigma_{\lambda})\notin\Phi_\chi$ by the definition of the index $\Hcal_{\lambda}$. From items B4, B5 and C3 it therefore follows that $\min\{B_{v\ch,1}+C_{1,u\di},B_{v\ch,2}+C_{2,u\di}\}=10u+10v-11$, from item 1 of Definition \[A\] that $A_{v\ch,u\di}=\min\{10(u+v)-11, \gamma_u, \rho_v\}$, so (\[2222\]) holds. 6.5. Finally, assume $\lambda=\mu$. We use items 6.1 and 6.2 and then assume without a loss of generality that either $u=v+\mu$ or $u=v+\mu+1$. In this case, the definition of $\lambda$ and $\mu$ shows that $u\in\{\sigma_\mu+1,\ldots,\sigma_{\mu+1}-1\}$. Now from items B4, B5 and C3 it follows that $\min\{B_{v\ch,1}+C_{1,u\di},B_{v\ch,2}+C_{2,u\di}\}=10u+10v-10$, from item 1 of Definition \[A\] that $A_{v\ch,u\di}=\min\{10(u+v)-10, \gamma_u, \rho_v\}$, so (\[2222\]) holds again. Items 6.1–6.5 consider all the possibilities, so the consideration of case 6 is complete. The cases 1–6 prove (\[1111\]) for every $(i,j)\in\Ical\times\Jcal$, so we can conclude that indeed $B\otimes C=A(\Sigma,\Rcal)$. The main results ---------------- In this subsection, we finalize the proof of the fact that $k$-TMF is NP-hard for $k\geq8$. What we need is the following lemma. \[tolemma\] Let an input $(\Sigma,\Rcal)$ be such that the answer in Problem \[ssref\] is ’no’, assume $B\in\Ro^{|\Ical|\times 8}$, $C\in\Ro^{8\times |\Jcal|}$. Then $B\ot C\neq A(\Sigma,\Rcal)$. The proof is by *reductio ad absurdum*. We assume that $B\ot C=A(\Sigma,\Rcal)$, in other words, that the equality $$\label{111}A_{ij}=\min_{\tau=1}^8 \{B_{i\tau}+C_{\tau j}\}$$ holds for all $i\in\Ical$ and $j\in\Jcal$. ***Step 1.*** 1.1. Let $\tau'\in\{1,\ldots,8\}$ satisfy $C_{\tau'\psi}\neq\i$ for some $\psi\in\{1\be,\ldots,7\be\}$. From Definition \[A\] it follows that $A_{8\pi, \psi}=A_{9\pi, \psi}=\i$, so the equality (\[111\]) implies $B_{8\pi,\tau'}+C_{\tau'\psi}=B_{9\pi,\tau'}+C_{\tau'\psi}=\i$, and thus we have $B_{8\pi,\tau'}=B_{9\pi,\tau'}=\i$. 1.2. By $\overline{\tau}=\{\overline{\tau}_1,\ldots,\overline{\tau}_g\}$ we denote the set of all $\tau\in\{1,\ldots,8\}$ satisfying both $\min\{C_{\tau, 9\be}, C_{\tau, 10\be}\}\neq\infty$ and $C_{\tau, 1\be}=\ldots=C_{\tau, 7\be}=\infty$. From item 1.1 it follows that either $B_{8\pi,\tau''}=B_{9\pi,\tau''}=\i$ or $C_{\tau'', 9\be}=C_{\tau'', 10\be}=\i$ for every $\tau''\notin\overline{\tau}$. Thus the equality (\[111\]) implies that $$A[8\pi,9\pi|9\be, 10\be]=B[8\pi,9\pi|\overline{\tau}_1,\ldots,\overline{\tau}_g]\otimes C[\overline{\tau}_1,\ldots,\overline{\tau}_g|9\be, 10\be],$$ and, since from Definition \[A\] it follows that $A[8\pi,9\pi|9\be, 10\be]=\left(\begin{smallmatrix}0 & \i \\\i & 0 \\\end{smallmatrix}\right)$, we see that $\overline{\tau}$ has cardinality at least $2$. 1.3. Using item 1.2, we assume without a loss of generality that $7,8\in\overline{\tau}$. Thus, $$\label{step1a}\min\{C_{7, 9\be}, C_{7, 10\be}\}\neq\infty,\,\,\,\,\,\min\{C_{8, 9\be}, C_{8, 10\be}\}\neq\infty,$$ $$\label{step1b}C_{7\psi}=C_{8\psi}=\infty\mbox{ $ $ $ $ for $ $ $ $} \psi\in\{1\be,\ldots,7\be\}.$$ ***Step 2.*** From the equality (\[111\]) it follows that the condition $$\label{step2a} \min\{A_{s',9\be},A_{s',10\be}\}=\min_{\tau=1}^8 \left\{B_{s',\tau}+\min\{C_{\tau,9\be},C_{\tau,10\be}\}\right\}$$ holds, in particular, for every $s'\in\{1\pi,\ldots,7\pi,1\ch,\ldots,(\sigma-m)\ch\}$. By Definition \[A\], we have $A_{s',9\be}=A_{s',10\be}=\i$, so that the left-hand side of (\[step2a\]) is indeed equal to $\i$. Now from the conditions (\[step1a\]) it follows that $$\label{step2b} B_{s'7}=B_{s'8}=\i\mbox{ $ $ $ $ for $ $ $ $} s'\in\{1\pi,\ldots,7\pi,1\ch,\ldots,(\sigma-m)\ch\}.$$ ***Step 3.*** Note that the product $B\otimes C$ will remain unchanged if we add some number $r\in\mathbb{R}$ to every entry of the $\tau$th column of $B$ and $-r$ to every entry of the $\tau$th row of $C$. So we can assume without a loss of generality that for $t\in\{1,\ldots,6\}$, the $t$th row of $C$ either consists of $\i$’s or has $0$ as its minimal element. 3.1. If the $t$th row of $C$ consists of $\i$’s, then, analogously, we can add a sufficiently large number to every entry of the $t$th column of $B$ without changing the product $B\otimes C$. Thus it can be assumed without a loss of generality that in this case, the $t$th column of $B$ consists of positive elements. 3.2. Now assume that the $t$th row of $C$ has $0$ as its minimal element (namely, $C_{tg'}=0$ for some $g'\in\Jcal$) and that $B_{s't}<0$ for some $s'\in\{1\pi,\ldots,7\pi,1\ch,\ldots,(\sigma-m)\ch\}$. Then from the equality (\[111\]) it follows that $A_{s'g'}\leq B_{s't}+C_{tg'}<0$, a contradiction with Definition \[A\]. As a conclusion of Step 3, we therefore obtain that $B_{s't}\geq0$, $C_{tj}\geq0$ for all $s'\in\{1\pi,\ldots,7\pi,1\ch,\ldots,(\sigma-m)\ch\}$, $t\in\{1,\ldots,6\}$, and $j\in\Jcal$. ***Step 4.*** Fix arbitrary $u\in\{1,\ldots,\sigma\}$ and $v\in\{1,\ldots,\sigma-m\}$. The condition (\[step2b\]) implies that $\min_{t=1}^6 \{B_{s't}+C_{tj}\}=\min_{\tau=1}^8 \{B_{s'\tau}+C_{\tau j}\}$ for $s'\in\{1\pi,\ldots,7\pi,1\ch,\ldots,(\sigma-m)\ch\}$. In particular, this means that the product of matrices $\widehat{B}=B[1\pi,\ldots,7\pi,v\ch|1,\ldots,6]$ and $\widehat{C}=C[1,\ldots,6|1\be,\ldots,8\be,u\di]$ is equal to the matrix $A[1\pi,\ldots,7\pi,v\ch|1\be,\ldots,8\be,u\di]$. By Definition \[A\], we have that $$\label{step4a} \widehat{B}\otimes\widehat{C}= \left(\begin{array}{ccccccccc} *&0&*&0&0&*&*&*&*\\ *&0&0&*&*&*&*&*&*\\ *&*&0&*&0&*&*&*&*\\ 0&*&*&*&*&*&*&0&0\\ *&*&*&*&*&0&*&0&*\\ *&0&0&*&0&0&0&0&*\\ *&*&*&*&*&*&0&*&*\\ *&*&*&*&*&*&0&*&*\\ \end{array}\right),$$ where the asterisks stand for certain (possibly different) positive elements from $\Ro$. By the conclusion of Step 3, the matrices $\widehat{B}$ and $\widehat{C}$ consist of nonnegative elements. Then, it turns out that the equality (\[step4a\]) determines the ’zero patterns’ of $\widehat{B}$ and $\widehat{C}$ uniquely up to a permutation of row and column indexes. Namely, one can check that $$\label{step4b} B[1\pi,\ldots,7\pi,v\ch|\omega(1),\ldots,\omega(6)]= \left(\begin{array}{cccccc} *&*&0&*&*&*\\ 0&*&*&*&*&*\\ *&0&*&*&*&*\\ *&*&*&0&*&*\\ *&*&*&*&0&*\\ 0&0&*&*&0&0\\ *&*&*&*&*&0\\ *&*&*&*&*&0\\ \end{array}\right),$$ $$\label{step4c} C[\omega(1),\ldots,\omega(6)|1\be,\ldots,8\be,u\di]= \left(\begin{array}{ccccccccc} *&0&0&*&*&*&*&*&*\\ *&*&0&*&0&*&*&*&*\\ *&0&*&0&0&*&*&*&*\\ 0&*&*&*&*&*&*&0&0\\ *&*&*&*&*&0&*&0&*\\ *&*&*&*&*&*&0&*&*\\ \end{array}\right),$$ where the asterisks stand again for certain positive elements from $\Ro$, and $\omega$ for some permutation on $\{1,\ldots,6\}$. Note that performing the same permutation of column indexes of $B$ and row indexes of $C$ will not change the product $B\otimes C$. Thus we can assume without a loss of generality that $\omega$ is the identity permutation. ***Step 5.*** Fix arbitrary $\eta\in\{1,\ldots,n\}$ and $h\in\{2,3,4,5\}$. 5.1. From Definition \[A\] it follows that $A_{\eta\bu,h\be}=0$. 5.2. The condition (\[step1b\]) shows that $C_{7,h\be}=C_{8,h\be}=\i$. Further, assume that $B_{\eta\bu,t'}+C_{t',h\be}\leq0$ for some $t'\in\{4,5,6\}$; then from (\[step4c\]) it follows that $B_{\eta\bu,t'}<0$. In this case, we use (\[step4c\]) again and obtain that $B_{\eta\bu,4}+C_{4,1\be}<0$, or $B_{\eta\bu,5}+C_{5,6\be}<0$, or else $B_{\eta\bu,6}+C_{6,7\be}<0$. On the other hand, Definition \[A\] implies $A_{\eta\bu,1\be}=A_{\eta\bu,6\be}=A_{\eta\bu,7\be}=2$, so we get a contradiction with the equality (\[111\]). Thus we have $\min_{t''=4}^8 \{B_{\eta\bu,t''}+C_{t'',h\be}\}>0$. 5.3. Items 5.1 and 5.2 and the equality (\[111\]) show that $\min_{t=1}^3 \{B_{\eta\bu,t}+C_{t,h\be}\}=0$. From the equality (\[step4c\]) it then follows that $$\label{step5a}B_{\eta\bu,3}=0\mbox{ $ $ $ $ for $ $ $ $} \eta\in\{1,\ldots,n\},$$ and $\min\{B_{\eta\bu,1},B_{\eta\bu,2}\}=0$. 5.4. By $\Psi_1$ (respectively, $\Psi_2$) we denote the set of all $\eta$ satisfying $B_{\eta\bu,1}=0$ (respectively, $B_{\eta\bu,2}=0$); from item 5.3 it follows that $\Psi_1\cup\Psi_2=\{1,\ldots,n\}$. The initial assumption of the lemma states that, given the input $(\Sigma,\Rcal)$, the answer in Problem \[ssref\] is ’no’. Therefore, for some $\mu\in\{0,\ldots,m-1\}$ and $\chi\in\{1,2\}$, we have $$\label{step5b}\{\sigma_{\mu}+1,\ldots,\sigma_{\mu+1}\}\subset\bigcup\limits_{i\in\Psi_\chi}\Rcal_i.$$ 5.5. The condition (\[step5b\]) means that for every $u'\in\{\sigma_{\mu}+1,\ldots,\sigma_{\mu+1}\}$, there exists $\eta(u')\in\{1,\ldots,n\}$ for which $u'\in\Rcal_{\eta(u')}$ and $\eta(u')\in\Psi_\chi$. (Here and in the rest of the proof, $\mu$ and $\chi$ denote the indexes defined in item 5.4.) Item 2 of Definition \[A\] then shows that $A_{\eta(u')\bu, u'\di}=\gamma_{u'}$, the definition of $\Psi_\chi$ that $B_{\eta(u')\bu,\chi}=0$. The number $\gamma_{u'}$ is to be understood here with respect to (\[CuRv\]), namely, for our case, we have that $$\label{step5bb}\gamma_{u'}=20u'-10\mu-9\,\,\mbox{if}\,\,u'\neq\sigma_{\mu+1}, \,\,\,\,\mbox{and}\,\,\gamma_{\sigma_{\mu+1}}=20\sigma_{\mu+1}-10\mu-19.$$ 5.6. The equality (\[111\]) shows that $B_{\eta(u')\bu, \chi}+C_{\chi,u'\di}\geq A_{\eta(u')\bu, u'\di}$; item 5.5 therefore implies $$\label{step5c}C_{\chi,u'\di}\geq\gamma_{u'}\mbox{ $ $ $ $ for every $ $ $ $} u'\in\{\sigma_{\mu}+1,\ldots,\sigma_{\mu+1}\}.$$ ***Step 6.*** 6.1. From Definition \[A\] it follows that $A_{6\pi,\sigma\di}=10\sigma-3$, so the equality (\[111\]) implies $B_{6\pi,4}+C_{4,\sigma\di}\geq10\sigma-3$. By (\[step4c\]), it holds that $C_{4,\sigma\di}=0$, so we obtain $$\label{step6a}B_{6\pi,4}\geq10\sigma-3.$$ 6.2. From Definition \[A\] it also follows that $A_{7\pi,u'\di}=\gamma_{u'}$ for $u'\in\{\sigma_{\mu}+1,\ldots,\sigma_{\mu+1}\}$. The equality (\[111\]) therefore implies $B_{7\pi,6}+C_{6,u'\di}\geq\gamma_{u'}$. By (\[step4b\]), it holds that $B_{7\pi,6}=0$, so we have $$\label{step6b}C_{6,u'\di}\geq\gamma_{u'}\mbox{ $ $ $ $ for every $ $ $ $} u'\in\{\sigma_{\mu}+1,\ldots,\sigma_{\mu+1}\}.$$ 6.3. Again, Definition \[A\] implies that for every $u'\in\{\sigma_{\mu}+1,\ldots,\sigma_{\mu+1}\}$, we have $A_{\nu(u')\bu,u'\di}=\gamma_{u'}$ (where the function $\nu$ is defined with respect to Remark \[nu\]). Therefore, by the equality (\[111\]), we obtain $B_{\nu(u')\bu,3}+C_{3,u'\di}\geq\gamma_{u'}$. By (\[step5a\]), it holds that $B_{\nu(u')\bu,3}=0$, thus, $$\label{step6c}C_{3,u'\di}\geq\gamma_{u'} \mbox{ $ $ $ $ for every $ $ $ $} u'\in\{\sigma_{\mu}+1,\ldots,\sigma_{\mu+1}\}.$$ 6.4. We use Definition \[A\] again and obtain $A_{v'\ch, 8\be}=\rho_{v'}$ for every $v'\in\{\sigma_\mu-\mu+1,\ldots,\sigma_{\mu+1}-\mu-1\}$; the number $\rho_{v'}$ is to be understood here with respect to (\[CuRv\]), namely, $$\label{step6cc}\rho_{v'}=20v'+10\mu+1.$$ Therefore, by the equality (\[111\]), we obtain $B_{v'\ch,4}+C_{4,8\be}\geq\rho_{v'}$, $B_{v'\ch,5}+C_{5,8\be}\geq\rho_{v'}$. By (\[step4c\]), it holds that $C_{4,8\be}=C_{5,8\be}=0$, thus, $$\label{step6d}B_{v'\ch,4}\geq\rho_{v'},\,\,B_{v'\ch,5}\geq\rho_{v'} \mbox{ $ $ $ $ for every $ $ $ $} v'\in\{\sigma_\mu-\mu+1,\ldots,\sigma_{\mu+1}-\mu-1\}.$$ 6.5. We have $A_{5\pi, (\sigma_\mu+1)\di}=10\sigma_\mu+11$ by Definition \[A\]. Therefore, by the equality (\[111\]), we obtain $B_{5\pi,5}+C_{5,(\sigma_\mu+1)\di}\geq10\sigma_\mu+11$. By the equality (\[step4b\]), it holds that $B_{5\pi,5}=0$, thus, $$\label{step6e}C_{5,(\sigma_\mu+1)\di}\geq10\sigma_\mu+11.$$ ***Step 7.*** 7.1. We have $A_{6\pi, (\sigma_\mu+1)\di}=10\sigma_\mu+7$ by Definition \[A\]. 7.2. The conditions (\[step5c\]), (\[step6b\]), (\[step6c\]), and (\[step6e\]) show that $C_{\widehat{t}, (\sigma_\mu+1)\di}>10\sigma_\mu+7$ for $\widehat{t}\in\{\chi,3,5,6\}$; the conclusion of Step 3 that $B_{6\pi,\widehat{t}}\geq0$, thus we have $$\label{step7a}B_{6\pi,\widehat{t}}+C_{\widehat{t}, (\sigma_\mu+1)\di}>A_{6\pi, (\sigma_\mu+1)\di}.$$ 7.3. By (\[step6a\]), $B_{6\pi,4}\geq10\sigma-3$. From Remark \[diff2\] it follows that $\sigma-\sigma_\mu\geq2$, so we obtain $B_{6\pi,4}>10\sigma_\mu+7$. The equality (\[step4c\]) implies $C_{4, (\sigma_\mu+1)\di}=0$, so that $$\label{step7b}B_{6\pi,4}+C_{4, (\sigma_\mu+1)\di}>A_{6\pi, (\sigma_\mu+1)\di}.$$ 7.4. From (\[step2b\]) it follows that, finally, $$\label{step7c}B_{6\pi,7}+C_{7, (\sigma_\mu+1)\di}=B_{6\pi,8}+C_{8, (\sigma_\mu+1)\di}=\i.$$ 7.5. The conditions (\[step7a\])–(\[step7c\]) show that the minimum in (\[111\]) for $i=6\pi$, $j=(\sigma_\mu+1)\di$ is provided by $\tau=\varphi$ (where $\varphi$ stands for the index from $\{1,2\}$ not equal to $\chi$, that is, $\{\varphi\}=\{1,2\}\setminus\{\chi\}$). So, using item 7.1, we see that $$B_{6\pi,\varphi}+C_{\varphi, (\sigma_\mu+1)\di}=10\sigma_\mu+7;$$ by (\[step4b\]), $B_{6\pi,\varphi}=0$, thus, $$\label{step7d}C_{\varphi, (\sigma_\mu+1)\di}=10\sigma_\mu+7.$$ ***Step 8.*** Now we consider indexes $\uo\in\{\sigma_\mu+1,\ldots,\sigma_{\mu+1}\}$ and $\vo\in\{\sigma_\mu-\mu+1,\ldots,\sigma_{\mu+1}-\mu-1\}$ that satisfy the condition either $\vo=\uo-\mu-1$ or $\vo=\uo-\mu$. 8.1. Definition \[A\] shows that $A_{\vo\ch,\uo\di}=10\uo+10\vo-10$ if $\uo\neq\sigma_{\mu+1}$, and $$A_{\left(\sigma_{\mu+1}-\mu-1\right)\ch,\sigma_{\mu+1}\di}=20\sigma_{\mu+1}-10\mu-21.$$ 8.2. Step 3 shows that $B_{\vo\ch,t}\geq0$ and $C_{t,\uo\di}\geq0$ for $t\in\{1,\ldots,6\}$, the condition (\[step2b\]) that $B_{\vo\ch,7}=B_{\vo\ch,8}=\i$. 8.3. The condition (\[step6d\]) shows that $B_{\vo\ch,4}\geq\rho_{\vo}$, $B_{\vo\ch,5}\geq\rho_{\vo}$. Using item 8.1 and the equality (\[step6cc\]), we obtain $B_{\vo\ch,4}>A_{\vo\ch,\uo\di}$, $B_{\vo\ch,5}>A_{\vo\ch,\uo\di}$. Further, the conditions (\[step5c\]), (\[step6b\]), and (\[step6c\]) imply that the elements $C_{\chi,\uo\di}$, $C_{3,\uo\di}$, and $C_{6,\uo\di}$ are all greater than or equal to $\gamma_{\uo}$. Item 8.1 and the equation (\[step5bb\]) show therefore that $C_{\chi,\uo\di}$, $C_{3,\uo\di}$, and $C_{6,\uo\di}$ are greater than $A_{\vo\ch,\uo\di}$. 8.4. Items 8.2 and 8.3 show that $B_{\vo\ch,t}+C_{t,\uo\di}>A_{\vo\ch,\uo\di}$ whenever $t\in\{\chi,3,4,5,6,7,8\}$. The equality (\[111\]) thus shows that $B_{\vo\ch,\varphi}+C_{\varphi,\uo\di}=A_{\vo\ch,\uo\di}$, and, therefore, $$\label{step8a}B_{\vo\ch,\varphi}+C_{\varphi,\uo\di}=10\uo+10\vo-10\mbox{$ $ $ $ for $ $ $ $} \uo\neq\sigma_{\mu+1}, \mbox{$ $ $ $ and $ $ $ $}$$ $$\label{step8b}C_{\varphi,\sigma_{\mu+1}\di}=20\sigma_{\mu+1}-10\mu-21-B_{\left(\sigma_{\mu+1}-\mu-1\right)\ch,\varphi}.$$ 8.5. By the initial setting of Step 8, $\vo$ can be either $\uo-\mu$ or $\uo-\mu-1$ if $\uo\notin\{\sigma_\mu+1,\sigma_{\mu+1}\}$, and $\vo=\sigma_\mu-\mu+1$ if $\uo=\sigma_\mu+1$ in (\[step8a\]). Therefore, the conditions (\[step8a\]) together with (\[step7d\]) and (\[step8b\]) form a classical linear system with $2(\sigma_{\mu+1}-\sigma_\mu)-1$ real unknowns and with the same number of equations; the system has a unique solution and it is given by $B_{\vo\ch,\varphi}=10\vo-7$, $C_{\varphi,\uo\di}=10\uo-3$ for $\uo\neq\sigma_{\mu+1}$, and $$\label{step8c}C_{\varphi,\sigma_{\mu+1}\di}=10\sigma_{\mu+1}-4.$$ 8.6. Finally, we note that $B_{6\pi,\varphi}=0$ by the equality (\[step4b\]), so the equality (\[step8c\]) implies $B_{6\pi,\varphi}+C_{\varphi,\sigma_{\mu+1}\di}=10\sigma_{\mu+1}-4$. Now we use the equality (\[111\]) to derive that $A_{6\pi,\sigma_{\mu+1}\di}\leq10\sigma_{\mu+1}-4$ and obtain a contradiction with Definition \[A\] with respect to which $A_{6\pi,\sigma_{\mu+1}\di}=10\sigma_{\mu+1}-3$. The contradiction obtained shows that the initial assumption was false. Therefore, matrices $B\in\Ro^{|\Ical|\times 8}$ and $C\in\Ro^{8\times |\Jcal|}$ can not satisfy $B\ot C=A(\Sigma,\Rcal)$. Now we are ready to prove the main results of the present section. \[8-tmf\] The $k$-TMF problem is NP-complete for any integer $k\geq8$. By Lemmas \[fromlemma\] and \[tolemma\], the matrix $A(\Sigma,\Rcal)$ has factor rank at most $8$ if and only if Problem \[ssref\] returns a positive answer on the query $(\Sigma,\Rcal)$. Using Proposition \[block\], we construct the matrix $\Acal(\Sigma,\Rcal,k)$, which has factor rank at most $k$ if and only if Problem \[ssref\] returns a positive answer. By Lemma \[polyred\], the number of operations needed to construct $\Acal(\Sigma,\Rcal,k)$ is polynomial in the size of the data. Finally, we perform an appropriate scaling and apply Lemma \[Ro-&gt;R\], thus giving a polynomial transformation from Problem \[ssref\] to $k$-TMF. That $k$-TMF is in NP follows from Theorem \[NP-k-tmf\]. Theorem \[8-tmf\] allows us to obtain a negative answer for Question \[que1\]. \[que1answ\] It is an NP-hard problem to decide whether the factor rank of a given tropical matrix equals $8$. In order to check whether the factor rank is at most $8$, we need to solve the $8$-TMF problem. So the result follows from Theorem \[8-tmf\]. Another result that can be obtained as a corollary of Theorem \[8-tmf\] gives an answer for Question \[que3\]. In contrast with the situation of the Gondran–Minoux rank, which can be computed in polynomial time if a given matrix has bounded tropical rank [@our], Question \[que3\] is answered in the negative. \[troprankhard\] Given a matrix $A$ of tropical rank at most $8$, it is NP-hard to decide whether the factor rank of $A$ is at most $8$. Using the definition of tropical rank explicitly, one can check in time $O(m^9n^9)$ whether the tropical rank of an $m$-by-$n$ matrix is greater than $8$. Further, Theorem 1.4 of [@DSS] shows that the factor rank of a matrix is greater than $8$ if the tropical rank is. Thus the present theorem follows from Theorem \[8-tmf\]. When TMF is easy: the case $k\leq3$ =================================== In this section, we discuss the computational complexity of tropical matrix factorization in the case when the dimension $k$ is small. It turns out that the TMF problem admits an easy solution if $k$ is required to be small enough, indeed, the case of $k=1$ is trivial. In [@DSS], the authors show that every tropical matrix whose factor rank is greater than $2$ has a $3$-by-$3$ submatrix with full factor rank, thus giving a polynomial-time algorithm for solving TMF in the case $k\leq2$. In fact, the TMF problem with $k\leq2$ can be solved by a linear-time algorithm, see [@CRW; @DSS]. However, the question of whether the TMF problem can be solved in polynomial time in the case $k=3$ remained open. The goal of this section is to give an affirmative answer to this question. The proof of the main result of this section is based on a reduction to a problem of linear programming. In the most general setting, the problem of *linear programming* asks whether a given system of $m$ linear inequalities with $n$ real unknowns has a solution. The polynomial-time algorithm for this problem has been constructed by Khachiyan in [@Kh]. However, the number of operations that Khachiyan’s algorithm needs to perform depends on the magnitude of coefficients, and this number can be unbounded even if $m$ and $n$ are fixed [@Meg]. On the other hand, it turns out that in our case, it is sufficient to consider the following special case of linear programming. \[lp2\] Given a system of $m$ linear inequalities involving $n$ real unknowns but no more than two unknowns per inequality. Question: Does there exist a solution of the given system? In contrast with the general problem of linear programming, a *strongly-polynomial* algorithm is known for solving Problem \[lp2\]. In other words, the number of operations this algorithm requires is bounded by a polynomial of $m$ and $n$. [@Meg]\[lp2p\] There is an algorithm that solves Problem \[lp2\] and requires at most $O\left(mn^3\log m\right)$ arithmetic operations. In order to prove our main result, we also need the following technical lemma. \[techmin\] Let $r_i$ be the greatest element of the $i$th row of a matrix $A\in\R^{p\times q}$, and $c_j$ the greatest element of the $j$th column. If numbers $u_i$ and $v_j$ satisfy $\min\{u_i,v_j\}=A_{ij}$ for any pair of indexes, then either $u_i=r_i$ for every $i$ or $v_j=c_j$ for every $j$. If $u_{i'}<r_{i'}$ for some $i'$, then we have $A_{i'j}<r_{i'}$ for every $j$, thus obtaining a contradiction with the maximality of $r_{i'}$. So we see that $u_i\geq r_i$ for every $i$. Analogously, we can show that $v_j\geq c_j$ for every $j$. Finally, if we have $u_i>r_i$ and $v_j>c_j$ for some $i$ and $j$, then $A_{ij}$ is greater than $\min\{r_i,c_j\}$, again contradicting the maximality of $r_i$ and $c_j$. Now we will prove the main result of this section. Throughout the proof, when we say that a certain computation can be made in time $\varphi(m,n)$, we mean that the computation requires at most $\varphi(m,n)$ arithmetic operations in the worst case. \[main2sec\] Given a matrix $A\in\R^{m\times n}$. There is an algorithm that decides whether the factor rank of $A$ is greater than $3$ and requires at most $O\left((m+n)^3mn\log(mn)+m^3n^3\right)$ arithmetic operations in the worst case. We need to decide whether the system of equations $$\label{sys}\min_{\tau=1}^3 \{B_{i\tau}+C_{\tau j}\}=A_{ij}$$ has a solution $B\in\R^{m\times 3}$, $C\in\R^{3\times n}$. 1\. Proposition 6.1 from [@DSS] states that $A$ has factor rank at least $3$ if and only if there is a full-rank $3$-by-$3$ submatrix of $A$. The $3$-by-$3$ submatrices of $A$ can be treated in time $O(m^3n^3)$, so we can assume without a loss of generality that $A[1,2,3|1,2,3]$ has factor rank $3$. Then we perform an appropriate scaling and relabeling, and can therefore assume that $0$ is the minimal element of every row and every column of $A$, and that $A[1,2,3|1,2,3]$ equals either $$\label{A[1,2,3]}\left( \begin{array}{ccc} 0 & * & * \\ * & 0 & * \\ * & * & 0 \\ \end{array} \right)\,\,\,\mbox{or}\,\,\, \left( \begin{array}{ccc} * & 0 & 0 \\ 0 & * & 0 \\ 0 & 0 & * \\ \end{array} \right),$$ where the asterisks stand for certain (possibly different) positive numbers. 2\. Assume that the $j$th column of $A$ consists of zero elements. In this case, we check whether $\min\{A_{i1},A_{i2},A_{i3}\}=0$ for every $i$. If so, the $j$th column of $A$ is a tropical sum of the first three columns, and we can remove the $j$th column without changing the factor rank. Otherwise, if $\min\{A_{i1},A_{i2},A_{i3}\}>0$ for some $i$, then the submatrix $A[1,2,3,i|1,2,3,j]$ equals either $$\left( \begin{array}{cccc} 0 & * & *&0 \\ * & 0 & *&0 \\ * & * & 0&0 \\ *&*&*&0\\ \end{array}\right)\,\,\,\mbox{or}\,\,\, \left( \begin{array}{cccc} * & 0 & 0&0 \\ 0 & * & 0&0 \\ 0 & 0 & *&0 \\ *&*&*&0\\ \end{array} \right),$$ so the factor rank of $A$ is at least $4$. Note that zero columns can be straightforwardly detected in time $O(mn)$. We can therefore assume that no column (as well as no row, analogously) of $A$ consists of zero entries. 3\. Suppose that matrices $B$ and $C$ indeed satisfy (\[sys\]). Then for every $\tau\in\{1,2,3\}$, we subtract $b_\tau$, the least element of the $\tau$th column of $B$, from every entry of that column and add $b_\tau$ to every entry of the $\tau$th row of $C$. The matrices $B'$ and $C'$ obtained satisfy $B'\ot C'=A$, and the minimal element of every column of $B'$ equals $0$. From item 1 it then follows that every element of $C'$ is nonnegative as well. Thus we can assume without a loss of generality that$$\label{BandCnonneg} B_{i\tau}\geq0,\,\,\,C_{\tau j}\geq0$$ for every triple of indexes $i$, $j$, $\tau$. It is thus sufficient to decide whether there exist matrices $B\in\R^{m\times 3}$ and $C\in\R^{3\times n}$ that satisfy (\[sys\]) and (\[BandCnonneg\]) simultaneously. Let us start to describe the algorithm that makes this decision. 4\. Items 1 and 3 allow us to determine the ’zero patterns’ of $B[1,2,3|1,2,3]$ and $C[1,2,3|1,2,3]$ up to relabeling the columns of $B$ and relabeling the rows of $C$. Namely, the problem reduces to the following three cases, each of which can be treated by the algorithm separately: $$\label{B[1,2,3]1}B[1,2,3|1,2,3]=C[1,2,3|1,2,3]=\left( \begin{array}{ccc} 0 & * & * \\ * & 0 & * \\ * & * & 0 \\ \end{array} \right),$$ $$\label{B[1,2,3]2}B[1,2,3|1,2,3]=\left( \begin{array}{ccc} 0 & * & * \\ * & 0 & * \\ * & * & 0 \\ \end{array} \right),\,\,\, C[1,2,3|1,2,3]=\left( \begin{array}{ccc} * & 0 & 0 \\ 0 & * & 0 \\ 0 & 0 & * \\ \end{array} \right),$$ $$\label{B[1,2,3]3}B[1,2,3|1,2,3]=\left( \begin{array}{ccc} * & 0 & 0 \\ 0 & * & 0 \\ 0 & 0 & * \\ \end{array} \right),\,\,\, C[1,2,3|1,2,3]=\left( \begin{array}{ccc} 0 & * & * \\ * & 0 & * \\ * & * & 0 \\ \end{array} \right).$$ If $B[1,2,3|1,2,3]$ has the form as in (\[B\[1,2,3\]1\]), then, by (\[sys\]), we necessarily have $C_{\tau j}=0$ if $A_{\tau j}=0$ and $C_{\tau j}>0$ if $A_{\tau j}>0$ for every $\tau\in\{1,2,3\}$ and $j\in\{1,\ldots,n\}$. Further, if $C[1,2,3|1,2,3]$ has the form as in (\[B\[1,2,3\]1\]), then (\[sys\]) implies that $B_{i\tau}=0$ if $A_{i\tau}=0$ and $B_{i\tau}>0$ if $A_{i\tau}>0$ for every $\tau\in\{1,2,3\}$ and $i\in\{1,\ldots,m\}$. Now consider the case (\[B\[1,2,3\]2\]), then the sign of every entry of $C$ is determined uniquely by the considerations of the previous paragraph. From item 2 it follows that for every $i$, there exists $j=j(i)$ satisfying $A_{i,j(i)}>0$. If $j(i)\in\{1,2,3\}$, then we use (\[sys\]) to obtain that $B_{i\tau}=0$ if $C_{\tau,j(i)}>0$ and $B_{i\tau}>0$ if $C_{\tau,j(i)}=0$. Otherwise, if $A_{i1}=A_{i2}=A_{i3}=0$, then we have $j(i)>3$, and from (\[sys\]) it can be deduced that again $B_{i\tau}=0$ if $C_{\tau,j(i)}>0$ and $B_{i\tau}>0$ if $C_{\tau,j(i)}=0$. Thus we have determined how the ’zero patterns’ of $B$ and $C$ look like in the cases (\[B\[1,2,3\]1\]) and (\[B\[1,2,3\]2\]), and we have performed $O(mn)$ operations to do this. The case (\[B\[1,2,3\]3\]) can be considered in the same way as (\[B\[1,2,3\]2\]). 5\. Let $A_{ij}$ equal $\beta_j$, the greatest element of the $j$th column of $A$. By item 4, we know a $\tau(j)\in\{1,2,3\}$ for which $B_{i,\tau(j)}=0$, the equations (\[sys\]) then imply that $C_{\tau(j),j}\geq\beta_j$. Moreover, if the equations (\[sys\]) and (\[BandCnonneg\]) are satisfied by some $B$ and $C$, then they can not be then broken by setting $C_{\tau(j),j}=\beta_j$. Note that computations of item 5 also require at most $O(mn)$ operations. 6\. Thus for every $j$, we can assume without a loss of generality that we have yet computed the values $C_{\tau_1(j),j}=0$, $C_{\tau_2(j),j}=\beta_j$, where $\tau_1(j),\tau_2(j)\in\{1,2,3\}$. Analogously, for every $i$, it be can assumed that $B_{i,\tau'_1(i)}=0$, $B_{i,\tau'_2(i)}=\gamma_i$ for $\tau'_1(i),\tau'_2(i)\in\{1,2,3\}$ ($\gamma_i$ stands here for the greatest element of the $i$th row of $A$). Note that the numbers $\tau_1(j)$ and $\tau'_1(i)$ are determined by the computations of item 4, the numbers $\tau_2(j)$ and $\tau'_2(i)$ by the computations of item 5. Note that by item 2, $\beta_j\neq0$ and $\gamma_i\neq0$ for every $i$ and $j$, so that $\tau_1(j)\neq\tau_2(j)$ and $\tau'_1(i)\neq\tau'_2(i)$. We also denote $\{\tau_3(j)\}=\{1,2,3\}\setminus\{\tau_1(j),\tau_2(j)\}$ and $\{\tau'_3(i)\}=\{1,2,3\}\setminus\{\tau'_1(i),\tau'_2(i)\}$. 7\. Let $t_1,t_2,t_3$ be different elements from $\{1,2,3\}$. By $\Ical_{t_1,t_2}$ we denote the set of all $i$ satisfying $\tau'_1(i)=t_1$, $\tau'_2(i)=t_2$; by $\Jcal_{t_1,t_2}$ denote the set of all $j$ satisfying $\tau_1(j)=t_1$, $\tau_2(j)=t_2$. Consider the subsystem of the system (\[sys\]) formed by the equations with $i\in\Ical_{1,2}$, $j\in\Jcal_{3,2}$. We can now apply Lemma \[techmin\] and reduce our problem to the two cases, each of which can be treated separately. In one of these cases, the value of $B_{i3}$ is determined for every $i\in\Ical_{1,2}$; in the other of them, $C_{1j}$ is determined for every $j\in\Jcal_{3,2}$. We repeat our considerations with arbitrary $\Ical_{t_1,t_2}$ and $\Jcal_{t_3,t_2}$ instead of $\Ical_{1,2}$ and $\Jcal_{3,2}$, the algorithm needs then consider $2^6=64$ cases separately. In what follows, we can therefore assume that, if $\tau'_2(i)=\tau_2(j)$ and $\tau'_1(i)\neq\tau_1(j)$, then either $B_{i,\tau'_3(i)}$ or $C_{\tau_3(j),j}$ is yet known. Finally, we note that the computations of item 7 can also be performed in time $O(mn)$. 8\. Now we will prove that for every $i$ and $j$, the equation (\[sys\]) either gives a contradiction or is equivalent to a linear inequality or equation with at most $2$ unknowns. This will reduce our problem to Problem \[lp2\], and then we will complete our proof by applying Theorem \[lp2p\]. Indeed, let $i\in\{1,\ldots,m\}$, $j\in\{1,\ldots,n\}$. If the entry $B_{i,\tau'_3(i)}$ is known, then (\[sys\]) is either a contradiction or can be rewritten as either $B_{i,\tau_3(j)}+C_{\tau_3(j),j}=A_{ij}$ or $B_{i,\tau_3(j)}+C_{\tau_3(j),j}\geq A_{ij}$. The case when the entry $C_{\tau_3(j),j}$ is known can be treated in the same way. There are the four possibilities in the case when both $B_{i,\tau'_3(i)}$ and $C_{\tau_3(j),j}$ are unknown. *Case A.* Suppose $\tau'_3(i)=\tau_3(j)$. In this case, if (\[sys\]) is not a contradiction, then it can be rewritten as either $B_{i,\tau'_3(i)}+C_{\tau_3(j),j}=A_{ij}$ or $B_{i,\tau'_3(i)}+C_{\tau_3(j),j}\geq A_{ij}$. *Case B.* Let $\tau'_3(i)=\tau_2(j)$. By item 4, we have $B_{i,\tau'_3(i)}>0$, thus $B_{i,\tau'_3(i)}+C_{\tau'_3(i),j}>\beta_j\geq A_{ij}$. If (\[sys\]) is not a contradiction, then it can be therefore rewritten as either $B_{i,\tau_3(j)}+C_{\tau_3(j),j}=A_{ij}$ or $B_{i,\tau_3(j)}+C_{\tau_3(j),j}\geq A_{ij}$. *Case C.* Let $\tau'_2(i)=\tau_3(j)$. By item 4, we have $B_{i,\tau_3(j)}>0$, thus $B_{i,\tau_3(j)}+C_{\tau_3(j),j}>\gamma_i\geq A_{ij}$. If (\[sys\]) is not a contradiction, then it can be therefore rewritten as either $B_{i,\tau'_3(i)}+C_{\tau'_3(i),j}=A_{ij}$ or $B_{i,\tau'_3(i)}+C_{\tau'_3(i),j}\geq A_{ij}$. *Case D.* The possibility $\tau'_2(i)=\tau_2(j)$, $\tau'_1(i)\neq\tau_1(j)$ is not indeed realizable thanks to item 7. The cases A–D cover all the possibilities, so that for every indexes $i$ and $j$, the equation (\[sys\]) either gives a contradiction or is equivalent to a linear inequality or equation with at most $2$ unknowns. What we need is to decide whether the systems (\[sys\]) and (\[BandCnonneg\]) have a simultaneous solution, and this can be now done in time $O\left((m+n)^3mn\log(mn)\right)$ by applying the algorithm mentioned in Theorem \[lp2p\]. Theorem \[main2sec\] shows that the TMF problem admits a polynomial-time solution in the case $k=3$. We can now obtain a positive answer for Question \[quemain\]. \[quemainansw\] There is a polynomial-time algorithm that decides whether a given tropical matrix has factor rank $3$. Checking whether the factor rank is at least $3$ requires at most $O(m^3n^3)$ operations because it is sufficient to check whether some $3$-by-$3$ submatrix has factor rank $3$, see Proposition 6.1 from [@DSS]. Checking whether the factor rank is at most $3$ can be done in polynomial time as Theorem \[main2sec\] shows. We also prove the result on the complexity of the $3$-TMF problem. \[P-3-tmf\] The $3$-TMF problem belongs to the class P. Solving $3$-TMF means deciding whether a given matrix has factor rank at most $3$. So the result follows from Theorem \[main2sec\]. The $k$-TMF problem, which belongs to NP by Theorem \[NP-k-tmf\], is still not known to be either in P or NP-complete if $k\in\{4,5,6,7\}$. For these values of $k$, the question of whether $k$-TMF admits a polynomial-time solution seems to be an interesting topic for future research. Factor rank and minors of fixed size {#sec3} ==================================== In this section, we discuss a problem that is related to determining computational complexity of tropical matrix factorization. The problem we are interested in, Question \[que2\] of the present paper, has been formulated by Develin, Santos, and Sturmfels in [@DSS]. The question asks to find some $N=N(k)$ for which every matrix with rank greater than $k$ necessarily contains a submatrix whose rank is greater than $k$ and whose size is at most $N\times N$. For instance, the basic result of classical linear algebra states that we can set $N_\F(k)=k+1$ for usual rank function of matrices over a field. Also, it follows directly from the definition that for tropical rank of a tropical matrix, one can set $N_{\mathrm{tr}}(k)=k+1$. The same question has also been studied for the Gondran–Minoux rank functions of tropical matrices, and it has been proved that setting $N_{\mathrm{GM}}(k)=2^{k+1}$ is sufficient in that case [@our]. For the factor rank, it is trivial that we can set $N(0)=1$, and it is easy to check that setting $N(1)=2$ suffices. The fact that $N(2)$ can be set to equal $3$ has been proved in [@DSS]. It was not known, however, whether there exists an $N(k)$ if $k\geq3$. If there were such $N(k)$, it would be bounded below by a function having an exponential growth as $k$ approaches infinity, see Proposition 2.2 of [@DSS]. In this section, we construct an example showing that indeed no such $N(k)$ exists for $k\geq4$. We fix an arbitrary integer $\nu>1$ and will use its designation throughout this section. Our goal is to construct a matrix $C$ with the following properties: (i) every minor of $C$ of size at most $\nu\times\nu$ has factor rank at most $4$, (ii) the factor rank of $C$ itself is greater than $4$. Now we present an example of such a matrix. \[defn\_H\]\[&gt;4exa\] *Denote an arbitrary number from $\{1,\ldots,\nu+1\}$ by $u$, from $\{1,\ldots,\nu+2\}$ by $v$. The entries of the tropical matrix $C=C(\nu)\in\R^{(\nu+6)\times(\nu+6)}$ are defined as follows:* *(1) $C_{uv}=\min\left\{10u+10v-10,20u+1,20v-9\right\}$ if $(u,v)$ equals neither $(\nu,\nu+1)$ nor $(\nu+1,\nu+1)$;* *(2) $C_{\nu,\nu+1}=20\nu-1$; $C_{\nu+1,\nu+1}=20\nu+11$;* *(3) $C_{u,\nu+3}=0$, $C_{u,\nu+4}=C_{u,\nu+5}=C_{u,\nu+6}=2$;* *(4) $C_{\nu+2,u}=10u-3$, $C_{\nu+3,u}=11$, $C_{\nu+4,u}=C_{\nu+5,u}=2$, $C_{\nu+6,u}=0$;* *(5) the matrix $C[\nu+2,\ldots,\nu+6|\nu+2,\ldots,\nu+6]$ equals $$\left( \begin{array}{ccccc} 0&0&0&0&2\\ 0&0&2&0&2\\ 2&2&0&2&2\\ 0&2&2&0&2\\ 0&2&2&2&0\\ \end{array} \right).$$* Let us prove that the factor rank of the matrix introduced is at least $5$. \[&gt;4\] The factor rank of the matrix $C(\nu)$ is greater than $4$. Assume the converse. Definition \[factrank\] then implies that $C=A\ot B$ for some $A\in\R^{(\nu+6)\times4}$ and $B\in\R^{4\times(\nu+6)}$, that is, $$\label{equ2} \min_{q=1}^4\{A_{sq}+B_{qt}\}=C_{st}$$ for every $s,t\in\{1,\ldots,\nu+6\}$. 1\. Note that the product $A\otimes B$ will remain unchanged if we add some number $r\in\mathbb{R}$ to every entry of the $q$th column of $A$ and $-r$ to every entry of the $q$th row of $B$. So we can assume without a loss of generality that for every $q\in\{1,2,3,4\}$, the $q$th row of $B$ has $0$ as its minimal element. In this case, the matrix $A$ also consists of nonnegative elements, because from Example \[&gt;4exa\] it follows that the elements of $C=A\ot B$ are nonnegative. 2\. Let us denote by $A'$ the $4$-by-$4$ matrix formed by the rows of $A$ with labels $\nu+1,\nu+4,\nu+5,\nu+6$, by $B'$ the $4$-by-$4$ matrix formed by the columns of $B$ with numbers $\nu+3,\nu+4,\nu+5,\nu+6$. From the definition of the matrix $C$, which is equal to $A\ot B$, it follows that the product $A'\ot B'$ has the form $$\label{diagstar} \left(\begin{array}{cccc} 0&*&*&*\\ *&0&*&*\\ *&*&0&*\\ *&*&*&0\\ \end{array} \right),$$ where the asterisks stand for certain (possibly different) positive elements. Item 1 shows that $A'$ and $B'$ are nonnegative, so that the matrices $A'$ and $B'$ have the form (\[diagstar\]) up to permutations of rows and columns. Note that performing the same permutation on column indexes of $A$ and row indexes of $B$ will not change the product $A\otimes B$. So we can assume without a loss of generality that the matrices $A'$ and $B'$ have exactly the form (\[diagstar\]). 3\. From item 1 it follows that the elements of $A$ and $B$ are nonnegative, from item 2 that $A_{\nu+5,q}=0$ if and only if $q=3$. So from (\[equ2\]) it follows that $B_{3t}=0$ if and only if $C_{\nu+5,t}=0$, where $t$ is an arbitrary index from $\{1,\ldots,\nu+6\}$. Analogously, item 2 implies that $A_{\nu+6,q}=0$ if and only if $q=4$, so that the equation (\[equ2\]) shows that $B_{4t}=0$ if and only if $C_{\nu+6,t}=0$. 4\. Item 2 also implies that $B_{q,\nu+3}=0$ if and only if $q=1$; $B_{q,\nu+4}=0$ if and only if $q=2$; $B_{q,\nu+5}=0$ if and only if $q=3$. Thus the equation (\[equ2\]) shows that $A_{s1}=0$ if and only if $C_{s,\nu+3}=0$; $A_{s2}=0$ if and only if $C_{s,\nu+4}=0$; $A_{s3}=0$ if and only if $C_{s,\nu+5}=0$; here $s$ runs over $\{1,\ldots,\nu+6\}$. 5\. In particular, items 3 and 4 imply that the entries $A_{\nu+1,1}$, $A_{\nu+3,1}$, $A_{\nu+2,2}$, $A_{\nu+3,3}$, $B_{3,\nu+2}$, $B_{4,\nu+1}$, $B_{4,\nu+2}$ are all equal to $0$. 6\. From (\[equ2\]) it also follows that $\min\{A_{s3}+B_{3,\nu+2},A_{s4}+B_{4,\nu+2}\}\geq C_{s,\nu+2}$ for every $s\in\{1,\ldots,\nu+6\}$. Item 5 implies that $B_{3,\nu+2}=B_{4,\nu+2}=0$, so we have $\min\{A_{s3},A_{s4}\}\geq C_{s,\nu+2}$. Applying the definition of the matrix $C$, we obtain $A_{i3}\geq 20i+1$, $A_{i4}\geq 20i+1$ for every $i\in\{1,\ldots,\nu\}$. 7\. Analogously, from item 5 it follows that $A_{\nu+1,1}=0$, so (\[equ2\]) implies that $B_{1t}\geq C_{\nu+1,t}$ for every $t\in\{1,\ldots,\nu+6\}$. For $j\in\{1,\ldots,\nu+1\}$, we therefore obtain $B_{1j}\geq 20j-9$. 8\. Again, item 5 implies $B_{4,\nu+1}=0$, so (\[equ2\]) shows that $A_{\nu+2,4}\geq C_{\nu+2,\nu+1}=10\nu+7$. Further, by item 5, $A_{\nu+3,1}=A_{\nu+3,3}=0$, so (\[equ2\]) implies $\min\{B_{11},B_{31}\}\geq C_{\nu+3,1}$, that is, $B_{11}\geq 11$, $B_{31}\geq 11$. 9\. Item 8 shows that the number $A_{\nu+2,q}+B_{q1}$ is greater than $7$ unless $q=2$. The definition of the matrix $C$ shows that $C_{\nu+2,1}=7$, so from (\[equ2\]) it follows that $A_{\nu+2,2}+B_{21}=7$. Item 5 implies that $A_{\nu+2,2}=0$, and we therefore obtain $B_{21}=7$. 10\. Now let $i\in\{1,\ldots,\nu\}$ and either $j=i$ or $j=i+1$. The definition of the matrix $C$ implies $C_{ij}\leq 10i+10j-10$. From items 6 and 7 it then follows that the numbers $A_{i3}$, $A_{i4}$, and $B_{1j}$ are always greater than $C_{ij}$. The equation (\[equ2\]) therefore implies $A_{i2}+B_{2j}=C_{ij}$. We thus obtain the following system of $2\nu+1$ linear (in the classical sense) equations with $2\nu+1$ real unknowns: $B_{21}=7$ and $$\label{equ3} A_{i2}+ B_{2i}=C_{ii},\,\,\,A_{i2}+ B_{2,i+1}=C_{i,i+1},\,\,\mbox{for every}\,\,i\in\{1,\ldots,\nu\}.$$ Solving the system, we obtain that, in particular, $B_{2,\nu+1}=10\nu+6$. From the equation (\[equ2\]) it thus follows that $C_{\nu+2,\nu+1}\leq A_{\nu+2,2}+B_{2,\nu+1}$. Item 5 shows that $A_{\nu+2,2}=0$, so that $C_{\nu+2,\nu+1}\leq 10\nu+6$. However, the definition of $C$ shows that $C_{\nu+2,\nu+1}=10\nu+7$. The contradiction obtained shows that our initial assumption is not true, so the result follows. Let us show that every $\nu$-by-$\nu$ submatrix of $C(\nu)$ has factor rank at most $4$. \[&lt;5\] Let $\mu\in\{1,\ldots,\nu+1\}$. The matrix $C'$ that is obtained from $C(\nu)$ by removing the $\mu$th column has factor rank at most $4$. We need to provide matrices $A\in\R^{(\nu+6)\times4}$ and $B\in\R^{4\times(\nu+5)}$ satisfying $A\ot B=C'$. Denoting an arbitrary number from $\{1,\ldots,\nu+1\}$ by $i$, from $\{1,\ldots,\nu+1\}\setminus\{\mu\}$ by $j$, we can construct these matrices as follows. We set $A_{i1}=0$ and $A_{i3}=A_{i4}=20i+1$ for every $i$, $A_{i2}=10i-7$ if $i<\mu$, $A_{i2}=10i-8$ if $\mu\leq i<\nu+1$, and $A_{\nu+1,2}=20\nu+21$. Set also $B_{1j}=20j-9$ and $B_{4j}=0$ for every $j$, $B_{3j}=10j-3$ if $j\neq1$, $B_{3j}=11$ if $j=1$. Define $B_{2j}=10j-3$ if either $j<\mu$ or $j=\nu+1$, set $B_{2j}=10j-2$ if $\mu<j<\nu+1$. Finally, we define the matrices $$\left( \begin{array}{cccc} 0&0&0&10\nu+7\\ 0&11&0&11\\ 2&0&2&2\\ 2&2&0&2\\ 2&2&2&0\\ \end{array} \right)\,\,\mbox{and}\,\, \left( \begin{array}{ccccc} 20\nu+20&0&2&2&2\\ 20\nu+20&2&0&2&2\\ 0&2&2&0&2\\ 0&2&2&2&0\\ \end{array} \right)$$ to be, respectively, the matrix formed by the rows of $A$ with numbers $\nu+2,\ldots,\nu+6$ and the matrix formed by the columns of $B$ with numbers $\nu+2,\ldots,\nu+6$. The product $A\ot B$ can now be straightforwardly checked to equal the matrix $C'$. Note that the factor rank of a submatrix is always less than or equal to the factor rank of whole matrix, Theorem \[&lt;5\] thus implies that every submatrix of $C(\nu)$ of size at most $\nu$-by-$\nu$ has factor rank at most $4$. Therefore, the matrix $C(\nu)$ gives a negative answer for Question \[que2\] if $k=4$. We are now ready to prove the main result of this section, and answer Question \[que2\] for every $k\geq4$. \[large&gt;3\] Let $k$ and $\nu$ be integers and $k\geq4$. There exists a tropical matrix $\Ccal(k,\nu)$ with the following properties: (i) the factor rank of $\Ccal(k,\nu)$ is greater than $k$; (ii) every minor of $\Ccal(k,\nu)$ of size at most $\nu$-by-$\nu$ has factor rank at most $k$. Apply Proposition \[block\] and then Lemma \[Ro-&gt;R\] to the matrix from Example \[&gt;4exa\]. [99]{} *Linear independence over tropical semirings and beyond*, Contemporary Mathematics, AMS, **495**(2009) 1–38. A.I. Barvinok. *Combinatorial Optimization and Computations in the Ring of Polynomials*, DIMACS Technical Report 93-13, 1993. A.I. Barvinok. *Two Algorithmic Results for the Traveling Salesman Problem*, Mathematics of Operations Research, **21(1)**(1996), 65–84. A. Barvinok, E. Kh. Gimadi, A. I. Serdyukov. *The maximum traveling salesman problem*, in: The Traveling Salesman problem and its variations (G. Gutin and A. Punnen, eds.), Kluwer, 2002. A. Barvinok, D.S. Johnson, G.J. Woeginger, and R. Woodroofe. *The maximum traveling salesman problem under polyhedral norms*, Integer programming and combinatorial optimization, Lecture Notes in Comput. Sci, **1412**, Springer, Berlin, 1998. Synchronization and Linearity, Wiley, 1992. *Isolation number versus Boolean rank*, Linear Algebra Appl., **436**(2012), 3469–3474. L. B. Beasley, S. J. Kirkland, B. L. Shader. *Rank comparisons*, Linear Algebra Appl., **221**(1995), 171–188. *Semiring rank versus column rank*, Linear Algebra Appl., **101**(1988), 33–48. *On the Barvinok rank of matrices*, presentation at the 2nd Aussois Workshop on Combinatorial Optimization, February 1998. *Nonnegative ranks, decompositions, and factorizations of nonnegative matrices*, Linear Algebra Appl., **190**(1993), 149–168. Minimax algebra, volume **166** of Lecture Notes in Economics and Mathematical Systems, Springer-Verlag, Berlin, 1979. *The moduli space of $n$ tropically collinear points in $\R^d$*, Collect. Math., **56(1)**(2005), 1–19. *Tropical Secant Varieties of Linear Spaces*, Discr. Comp. Geom., **35(1)**(2006), 117–129. *On the rank of a tropical matrix*, in: Discrete and Computational Geometry (E. Goodman, J. Pach and E. Welzl, eds.), MSRI Publications, Cambridge Univ., Press, 2005. *Non-Archimedean amoebas and tropical varieties*, J. Reine Angew. Math., [**601**]{}(2006), 139–157. J. Ferrante, C. Rackoff. *A decision procedure for the first order theory of real addition with order*, SIAM J. Computing, **4**(1975), 69–76. *Tropical patterns of matrices and the Gondran–Minoux rank function*, Linear Algebra Appl., to appear. Max Plus at Work: Modeling and Analysis of Synchronized Systems: A Course on Max- Plus Algebra and Its Applications, Princeton Univ. Press, 2006. The topology of the space of matrices of Barvinok rank two, arXiv:0902.3380v1. Computers and Intractability: A Guide to the Theory of NP-Completeness, W. H. Freeman and Company, San Francisco, 1979. *On the Burnside problem for semigroups of matrices in the $(\max,+)$ algebra*, Semigroup Forum **52**(1996), 271–292. S. Gaubert, M. Plus. *Methods and applications of (max,+) linear algebra*, in Lecture Notes in Computer Science, **1200**, Springer Verlag, New York, 1997. *Biclique Coverings of Regular Bigraphs and Minimum Semiring Ranks of Regular Matrices*, J. Combin. Theory Ser. B, **51**(1991), 73–89. Z. Izhakian, M. Johnson, M. Kambites, *Tropical matrix groups*, arXiv:1203.2449. Multiplicative structure of $2\times2$ tropical matrices, Linear Algebra Appl. 435 (2011) 1612–1625. L. G. Khachiyan. *Polynomial algorithms in linear programming*, USSR Comp. Math. and Math. Phys., **20(1)**(1980), 53–72. Idempotent analysis and applications. Kluwer Acad. Publishers, 1997. Max-plus methods for nonlinear control and estimation, Systems $\verb"&"$ Control: Foundations $\verb"&"$ Applications, Birkhäuser Boston Inc., Boston, MA, 2006. N. Megiddo. *Towards a Genuinely Polynomial Algorithm for Linear Programming*, SIAM J. Comput. **12**(1983), 347–353. J.-E. Pin. *Tropical semirings*, In Idempotency (Bristol, 1994), Publ. Newton Inst., **11**, Cambridge Univ. Press, Cambridge, 1998. *Limited Subsets of a Free Monoid*, in Proc. 19th Annual Symposium on Foundations of Computer Science, Piscataway, N.J., Institute of Electrical and Electronics Engineers, 1978. *Extremal algebra of positive matrices*, Elektron. Informationsverarbeitung und Kybernetik, **3**(1967), 39–71.

ArXiv

--- abstract: 'Weak gravitational lensing provides a unique method to map directly the distribution of dark matter in the universe and to measure cosmological parameters. This cosmic-shear technique is based on the measurement of the weak distortions that lensing induces in the shape of background galaxies as photons travel through large-scale structures. This technique is now widely used to measure the mass distribution of galaxy clusters and has recently been detected in random regions of the sky. In this review, we present the theory and observational status of cosmic shear. We describe the principles of weak lensing and the predictions for the shear statistics in favored cosmological models. Next, we review the current measurements of cosmic shear and show how they constrain cosmological parameters. We then describe the prospects offered by upcoming and future cosmic-shear surveys as well as the technical challenges that have to be met for the promises of cosmic shear to be fully realized.' author: - '[Alexandre Refregier]{}' title: | Weak Gravitational Lensing by\ Large-Scale Structure\ ------------------------------------------------------------------------ --- psfig.sty Cosmology, Dark Matter, Cosmic Shear, Structure Formation, Statistics INTRODUCTION ============ Gravitational lensing provides a unique method to directly map the distribution of dark matter in the universe. It relies on the measurement of the distortions that lensing induces in the images of background galaxies (for reviews, see Bartelmann & Schneider 1999, Bernardeau 1999, Kaiser 1999, Mellier 1999, Narayan & Bartelmann 1996, Schneider 1995, Wittman 2002). This method is now widely used to map the mass of clusters of galaxies (see Fort & Mellier 1994 for a review) and has been extended to the study of superclusters (Gray et al. 2002, Kaiser et al. 1998) and groups (Hoekstra et al. 2001). Recently, weak lensing was statistically detected for the first time in random patches of the sky (Brown et al. 2003; Bacon et al. 2002; Bacon, Refregier & Ellis 2000; Hamana et al. 2003; Hämmerle et al. 2002; Hoekstra et al. 2002a; Hoekstra, Yee & Gladders 2002a; Jarvis et al. 2002; Kaiser, Wilson & Luppino 2000; Maoli et al. 2001; Refregier, Rhodes, & Groth 2002; Rhodes, Refregier & Groth 2001; van Waerbeke et al. 2000; van Waerbeke et al. 2001a; van Waerbeke et al. 2002; Wittman et al. 2000). These cosmic shear surveys provide direct measurements of large-scale structure in the universe and, therefore, of the distribution of dark matter. Unlike other methods that probe the distribution of light, weak lensing measures the mass and can thus be directly compared to reliable theoretical models of structure formation. Cosmic shear can therefore be used to measure cosmological parameters in the context of these models, thereby opening wide prospects for cosmology (Bernardeau, van Waerbeke & Mellier 1997; Hu & Tegmark 1999; Jain & Seljak 1997; Kaiser 1998). In the present review, we describe the theoretical and observational status of cosmic shear. Earlier reviews of this fast-evolving field can be found in Hoekstra et al. (2002b), Mellier et al. (2001), van Waerbeke et al. (2002), Wittman (2002). Here, we first describe the principles of weak lensing (Section 2). We then summarize the different statistics used to measure cosmic shear and describe how they are used to constrain cosmological parameters (Section 3). In Section 4, we survey the different methods used to derive the lensing shear from the shapes of background galaxies. We present, in Section 5, the current observations and their cosmological significance. Future cosmic-shear surveys and the prospects they offer for cosmology are described in Section 6. In Section 7, we outline how systematic effects present challenges that must be met for the potential of these future surveys to be fully realized. We summarize our conclusions in [Section 8]{}. THEORY ====== The idea of cosmic shear can be traced back to a lecture given by Richard Feynman at Caltech in 1964 (J.E. Gunn, personal communication). Several theorists (e.g., Gunn 1967, Jaroszyńsky et al. 1990, Kristian & Sachs 1966, Lee & Paczyńsky 1990, Schneider & Weiss 1988) then studied the propagation of light in an inhomogeneous universe. Predictions for the statistics of the weak-lensing distortions were then computed in a modern cosmological context by several groups (Babul & Lee 1991, Blandford et al. 1991, Kaiser 1992, Miralda-Escudé 1991, Villumsen 1996). More recently, the power of cosmic shear to measure cosmological parameters was the object of many theoretical studies (Bernardeau, van Waerbeke & Mellier 1997; Jain & Seljak 1997; Hu & Tegmark 1999; Kaiser 1998; Kamionkowski et al. 1997; van Waerbeke, Bernardeau & Mellier 1999). In this section, we briefly describe the principles of weak lensing and show how this technique can be used to map the dark matter in the universe. As they travel from a background galaxy to the observer, photons get deflected by mass fluctuations along the line of sight (see Figure \[fig:lensing\]). As a result, the apparent images of background galaxies are subject to a distortion that is characterized by the distortion matrix: $$\label{eq:psi_def_theory} \Psi_{ij} \equiv \frac{\partial (\delta \theta_{i})}{\partial \theta_{j}} \equiv \left( \begin{array}{cc} \kappa +\gamma_{1} & \gamma_{2}\\ \gamma_{2} & \kappa - \gamma_{1} \\ \end{array} \right),$$ where $\delta \theta_{i}({\mathbf \theta})$ is the deflection vector produced by lensing on the sky. The convergence $\kappa$ is proportional to the projected mass along the line of sight and describes overall dilations and contractions. The shear $\gamma_{1}$ ($\gamma_{2}$) describes stretches and compressions along (at $45^{\circ}$ from) the x-axis. Figure \[fig:ellip\] illustrates the geometrical meaning of the two shear components. The distortion matrix is directly related to the matter density fluctuations along the line of sight by $$\label{eq:psi_dchi} \Psi_{ij} = \int_{0}^{\chi_{h}} d\chi ~g(\chi) \partial_{i} \partial_{j} \Phi,$$ where $\Phi$ is the Newtonian potential, $\chi$ is the comoving distance, $\chi_{h}$ is the comoving distance to the horizon, and $\partial_{i}$ is the comoving derivative perpendicular to the line of sight. The radial weight function $g(\chi)$ is given by $$g(\chi) = 2 \int_{\chi}^{\chi_{h}} d\chi'~n(\chi') \frac{r(\chi)r(\chi'-\chi)}{r(\chi')},$$ where $r=a^{-1}D_{A}$, and $D_{A}$ is the angular-diameter distance. The function $n(\chi)$ is the distribution of the galaxies as a function of the comoving distance $\chi$ from the observer and is assumed to be normalized as $\int d\chi n(\chi)=1$. As we discuss in Section 4, galaxy shapes can be averaged over a patch of the sky to measure the shear, which is thus an observable. The shear pattern expected in a standard Cold Dark Matter (SCDM) model is shown in Figure \[fig:jsw\] for a $1\times1$ deg$^{2}$ region. Jain, Seljak, & White (2000) derived this map from ray tracing through N-body simulations. Tangential patterns around the overdensities corresponding to clusters and groups of galaxies, along with a more complicated network of shear fluctuations, are apparent. By inverting the lensing equation (Equation \[eq:psi\_dchi\]), the shear map can be converted into a map of the projected mass $\kappa$ and, therefore, of the dark matter distribution. COSMIC-SHEAR STATISTICS AND COSMOLOGY {#statistics} ===================================== The statistical characteristics of the cosmic-shear field can be quantified using a variety of measures, which can then be used to constrain cosmological models. First, we consider the most basic two-point statistic of the shear field, namely the two-dimensional power spectrum (Jain & Seljak 1997, Kaiser 1998, Kamionkowski et al. 1997, Schneider et al. 1998a). The shear power spectrum $C_{{\mathbf l}}$ is defined as a function of multipole moment $l$ (or inverse angular scale) by $\sum_{i=1}^{2} \langle \widetilde{\gamma_{i}}({\mathbf l}) \widetilde{\gamma_{i}}({\mathbf l'}) \rangle = (2\pi)^{2} \delta({\mathbf l} - {\mathbf l'}) C_{{\mathbf l}}$, where tildes denote Fourier transforms (with the conventions of Bacon, Refregier, & Ellis 2000), $\delta$ is the two-dimensional Dirac-delta function, and the brackets denote an ensemble average. Applying Limber’s equation in Fourier space (e.g., Kaiser 1998) to Equation \[eq:psi\_dchi\] and using the Poisson equation, one can easily express the shear power spectrum $C_{l}$ in terms of the three-dimensional power spectrum $P(k,\chi)$ of the mass fluctuations $\delta \rho/\rho$ and obtain $$C_{l} = \frac{9}{16} \left( \frac{H_{0}}{c} \right)^{4} \Omega_{m}^{2} \int_{0}^{\chi_h} d\chi~\left[ \frac{g(\chi)}{a r(\chi)} \right]^{2} P\left(\frac{l}{r}, \chi\right),$$ where $a$ is the expansion parameter, and $H_{0}$ and $\Omega_{m}$ are the present value of the Hubble constant and matter density parameter, respectively. The lensing power spectra for four CDM models are shown in Figure \[fig:cl\] (see color insert) (see Bacon, Refregier & Ellis 2000 for the exact cosmological parameter values of each model). They were derived using the fitting formula for the nonlinear matter power spectrum $P(k,\chi)$ of Peacock & Dodds (1996). In Figure 4, the galaxies were assumed to lie on a sheet at a redshift $z_{s}=1$. A more realistic redshift distribution $n(z)$ would require corrections of only approximately 10% on these power spectra, as long as the median redshift of the galaxies were kept at $z_{s}=1$. The three cluster-normalized models ($\Lambda$CDM, OCDM and $\tau$CDM) yield power spectra of similar amplitudes but with different shapes. The COBE-normalized SCDM model has more power on a small scale and thus yields a larger normalization. For the $\Lambda$CDM model, the power spectrum corresponding to a linear evolution of structures is also shown in Figure 4 for comparison. For $l {\raisebox{-0.5ex}{$\,\stackrel{>}{\scriptstyle \sim}\,$}}1000$ (corresponding to angular scales smaller than approximately 10’), nonlinear corrections dominate the power spectrum (Jain & Seljak 1997), making cosmic-shear sensitive to gravitational instability processes. The measurement of the lensing power spectrum can thus be used to measure cosmological parameters, such as $\Omega_{m}$, $\Omega_{\Lambda}$, $\sigma_{8}$, and $\Gamma$. (Bernardeau, van Waerbeke & Mellier 1997; Hu & Tegmark 1999; Jain & Seljak 1997; Kaiser 1998; van Waerbeke, Bernardeau & Mellier 1999). A full-sky cosmic-shear survey would yield a precision of these parameters comparable to that for future cosmic microwave background (CMB) missions. More realistically for the short term, a precision on the order of 10% can be achieved with surveys of approximately 10 square degrees. Such cosmic-shear surveys can also be combined with CMB anisotropy measurements to break degeneracies present when the CMB is considered alone (Hu & Tegmark 1999; Contaldi, Hoekstra & Lewis 2003). This would yield improvements in the precision of cosmological parameters by approximately one order of magnitude. In addition, the use of photometric redshifts can provide a tomographic measurement of matter fluctuations and improve the precision of cosmological parameters by up to an order of magnitude (Hu 1999, Hu & Keeton 2002, Taylor 2001). In practice, it often is more convenient to measure other two-point statistics. In particular, the variance of the shear $\sigma_{\gamma}^{2} \equiv \langle \overline{\gamma}^{2} \rangle$ in randomly placed cells is widely used. It is related to the shear power spectrum by $$\sigma_{\gamma}^{2} = \frac{1}{2\pi} \int_{0}^{\infty} dl~ l C_{l} \left| \widetilde{W}_{l} \right|^{2}, \label{eq:sigma_gamma}$$ where $\widetilde{W}_{l}$ is the Fourier transform of the cell aperture. The shear two-point correlation functions (eg. Kaiser 1998) are Fourier transforms of the power spectra and have also been measured by various groups (eg. Bacon et al. 2002; van Waerbeke et al. 2001a; Hoekstra et al. 2002a). The $M_{ap}$ statistics (Schneider et al. 1998a) is another convenient statistic based on the average of the shear within a compensated filter. Its advantages are that its window function in $l$-space is narrow, that adjacent cells are effectively uncorrelated, and that it can be easily related to the statistics of the projected mass $\kappa$. In analogy with electromagnetism, a tensor field like the shear field $\gamma_{i}$ can be decomposed into an electric (E, or gradient) component and a magnetic (B, or curl) component (Crittenden et al. 2001b, Kaiser 1992, Kamionkowski et al. 1997, Stebbins 1996). Because the distortions it induces arise from a scalar field (the gravitational potential), weak lensing only produces E-type fluctuations. On the other hand, systematic effects and intrinsic galaxy alignments are likely to produce both E-type and B-type fluctuations (See Heavens 2001 for a review; also see Section 7.6). The presence of B-modes can thus be used as a measure of contaminants to the cosmic-shear signal. In practice, this decomposition can be performed using the $M_{ap}$ statistic (Schneider et al. 1998a, van Waerbeke et al. 2001a) or the power spectrum (Hu & White 2000, Padmanabhan, Seljak & Pen 2002, Pen et al. 2001). \[eb\] As is apparent in Figure \[fig:jsw\], the shear field is not Gaussian but, instead, has apparent coherent structures. Bernardeau, van Waerbeke, & Mellier (1997) have shown that the skewness $$S_{3} = \frac{\langle \kappa^{3} \rangle}{\langle \kappa^{2} \rangle^2}$$ of the convergence field $\kappa$ can be used to break the degeneracy between $\sigma_{8}$ and $\Omega_{m}$, which is present when the shear variance alone is considered. The measurement of higher-point statistics is thus of great cosmological interest, but their computation is made difficult by the fact that the angular scales accessible to obervation ($\theta {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle \sim}\,$}}10'$) are in the nonlinear regime. Predictions for higher-point correlation functions have been calculated using perturbation theory, the hierarchical ansatz, (Bernardeau & Valageas 2000; Bernardeau, van Waerbeke & Mellier 1997; Hui 1999; Munshi & Coles 2000; Munshi & Jain 2001; van Waerbeke et al. 2001b) and halo models (Cooray & Hu 2001a; Cooray, Hu, & Miralda-Escudé 2000). These techniques have also been used to compute the full probability distribution function of the convergence field (Bernardeau & Valageas 2000, Munshi & Jain 2000, Valageas 2000) and the errors in two-point statistics (Cooray & Hu 2001b; Munshi & Coles 2002; Schneider et al. 2002). Another possible way to compute the full nonlinear field is to use ray tracing through N-body simulations (Blandford et al. 1991; Hamana, Colombi & Mellier 2000; Jain, Seljak & White 2000; Premadi et al. 2001; Wambsganss, Cen, & Ostriker 1998; White & Hu 2000) to produce simulated shear maps such as those by Jain, Seljak & White (2000) (shown in Figure \[fig:jsw\]). These can be used to compute and study other proposed measures of non-Gaussianity such as peak statistics (Jain & van Waerbeke 2000), a generalized maximum likelihood (Taylor & Watts 2000), and cluster counts (Bartelmann, King & Schneider 2001). In practice, the complex geometry of surveys makes it difficult to infer a convergence $\kappa$ map from the observed shear $\gamma_{i}$. For this reason, a number of researchers have recently proposed the use of high-order statistics of the shear field rather than the skewness $S_{3}$ of the convergence (Bernardeau, van Waerbeke, & Mellier 2003; Schneider & Lombardi 2002; Takada & Jain 2002; Zaldarriaga & Scoccimarro 2002). SHEAR MEASUREMENT METHODS {#method} ========================= Because the sought-after lensing signal is of only a few percent in amplitude, the data acquisition and analysis must be performed carefully, and systematic effects must be tightly controlled. All the current measurements of cosmic shear were derived from deep optical images taken with charged-coupled devices (CCD). It is advantageous for the exposures to be homogeneous in depth and for the ground to be subject to as small a seeing as possible. The fields are generally chosen to lie far away from each other to ensure that they are statistically independent and to minimize cosmic variance. The first step in the data analysis is image processing. After flat fielding, the different exposures are co-added to produce the final reduced images. If necessary, any instrumental distortion induced by the telescope optics is corrected for at this stage. This can be done very accurately by measuring the astrometric offsets from several dithered exposures. For a detailed description of the different image-processing steps, see Kaiser et al. (1999). An example of a processed deep image from the cosmic-shear survey of Bacon, Refregier & Ellis (2000) is shown in Figure \[fig:wht\]. The next step consists of deriving an estimator for the shear from the shapes of the galaxies in the co-added images. The point-spread function (PSF), which smears the images of galaxies and is generally not circular, complicates this task. In general, the PSF varies spatially and in time, and it must be measured and modelled for each image individually. This can be done by measuring the shape of the stars in the field, whose number can be optimized by tuning the galactic latitudes of the observations. Figure \[fig:star\] shows the ellipticity pattern of the PSF from one of the William Hershel telescope (WHT) fields of Bacon, Refregier & Ellis 2000. Several methods have been developed to tackle this difficult and crucial task. The more rigorous method of Kaiser, Squires & Broadhurst (KSB; 1995), further developed by Luppino & Kaiser (1997) and Hoekstra et al. (1998), replaced the earlier method by Bonnet & Mellier (1995). The KSB method is now widely used for cluster studies, and it has been used by the majority of the groups involved in measuring cosmic shear (see Bartelmann & Schneider 1999 for a detailed review of the KSB method). It is based on the measurement of the quadrupole moment of the galaxy surface brightness $I({\mathbf x})$, $$Q_{ij} \equiv \int d^{2}x x_{i} x_{j} w(x) I({\mathbf x}),$$ where $w(x)$ is a weight function conveniently taken to be Gaussian. These moments capture the lowest-order shape information of the galaxy and can be combined to form the ellipticity of the galaxy $$\epsilon_{1}= \frac{Q_{11}-Q_{22}}{Q_{11}+Q_{22}},~~~ \epsilon_{2}= \frac{2Q_{12}}{Q_{11}+Q_{22}}.$$ The first (second) component of the ellipticity describes compressions and elongations along (at 45$^{\circ}$ from) the x and y axes (see Figure \[fig:ellip\] for an illustration). The ellipticity vanishes for circular galaxies. . \[fig:star\] The first step in the KSB method is to correct the observed galaxy ellipticity $\epsilon_{i}^{g \prime}$ for the anisotropy of the PSF. The corrected galaxy ellipticity $\epsilon^{g}$ is given by $$\epsilon^{g} = \epsilon^{g \prime} - P_{\rm sm}^{g} (P_{\rm sm}^{*})^{-1} \epsilon^{*},$$ where $\epsilon^{*}$ is the PSF ellipticity derived from the stars, and $P_{\rm sm}^{g}$ and $P_{\rm sm}^{*}$ are the smear susceptibility tensors for the galaxy and star, respectively, and can be derived from higher moments of the images. The shear in a patch of the sky can then be measured by averaging over the (corrected) ellipticities in the patch of the sky using $$\gamma = (P_{\gamma})^{-1} \langle \epsilon^{g} \rangle,$$ where the tensor $P_{\gamma}$ quantifies the susceptibility to shear acting before isotropic PSF smearing and is given by $$P_{\gamma} = P_{\rm sh}^{g} - P_{\rm sh}^{*} (P_{\rm sm}^{*})^{-1} P_{\rm sm}^{g},$$ where the shear susceptibility tensors $P_{\rm sh}^{g}$ and $P_{\rm sh}^{*}$ for the galaxies and the stars can also be measured from higher moments of their respective light distribution. The KSB method was thoroughly tested using realistic simulated images by Erben et al. (2001) and Bacon et al. (2001a). These studies showed that it is accurate to within a few tenths of percent in reconstructing an input shear and is thus sufficient for the current cosmic-shear surveys (see Section 5). However, this accuracy is insufficient for future, more sensitive surveys (see Section 6). Moreover, Kuijken (1999) and Kaiser (2000) have shown that the KSB method is ill-defined mathematically and unstable for PSF’s found in practice. This inadequacy has led a number of researchers to develop alternative methods. Rhodes, Refregier & Groth (2000) have modified the KSB method to be better suited for HST images. Kuijken (1999) considered a different approach that consisted of fitting the observed galaxy shape with a smeared and sheared circular model. Kaiser (2000) introduced a new method based on a finite resolution shear operator. Refregier & Bacon (2003) and, independently, Bernstein & Jarvis (2001) developed a new method based on the decomposition of the galaxies into shape components or “shapelets”. The gauss-hermite orthogonal basis functions used in this approach allow shears and PSF convolutions to be described as simple matrix operations, using the formalism developped for the quantum harmonic oscillator (Refregier 2003). These new methods are promising but require extensive testing to establish whether they will achieve the required precision. OBSERVATIONS ============ Because the expected distortions are only of a few percent, the measurement of cosmic shear requires large survey areas and excellent image quality. Early searches for cosmic shear signals with photographic plates were unsuccesful (Kristian 1967; Valdes, Jarvis & Tyson 1983). Mould et al. (1994), performed the first attempt to detect a cosmic-shear signal with CCDs, but only derived an upper limit. Using the same data, Villumsen (1995) reported a $4.5\sigma$ detection. Schneider et al. (1998b) then reported a detection of cosmic shear in one of three QSO fields, an area too small to draw any constraints on cosmology. Within a few weeks, four independent groups (Bacon, Refregier & Ellis 2000; Kaiser, Wilson & Luppino 2000; van Waerbeke et al. 2000; Wittman et al. 2000) reported the first firm statistical detections of cosmic shear using three different 4m-class telescopes: the Cerro Tololo Inter-American Observatory (CTIO), the Canada-France-Hawaii Telescope (CFHT), and the William Herschel Telescope (WHT). These were later confirmed by more precise measurements of the cosmic-shear amplitude from the ground (Bacon et al. 2002; Brown et al. 2003; Maoli et al. 2001; Hamana et al. 2003; Hoekstra et al. 2002a; Hoekstra, Yee & Gladders 2002a; Jarvis et al. 2002; van Waerbeke et al. 2001; van Waerbeke et al. 2002a) using 2-m – (MPG/ESO), 4-m– (WHT, CFHT, CTIO) and 8-m– (Very Large Telescope, Keck, Subaru) class ground-based telescopes. Meanwhile, cosmic shear was also detected (Hämmerle et al. 2002; Rhodes, Refregier & Groth 2001) and measured (Refregier, Rhodes & Groth 2002) from space using HST. Space-based surveys are currently limited by the small field of view of HST, but they are deeper and less prone to systematics thanks to the absence of atmospheric seeing. Table \[tab:surveys\] summarizes the existing cosmic-shear surveys and highlights the wide range of telescopes, survey areas, and depths. The shear variance $\sigma_{\gamma}^{2}$ (Equation \[eq:sigma\_gamma\]) measured recently by several groups is shown in Figure \[fig:var\] (see color insert) as a function of the radius a circular cell. The results by Hämmerle et al. (2002) and Hamana et al. (2002a) are not displayed. Note that, for the shear variance, the data points at different angular scales are not independent. For comparison, the shear variance predicted for a $\Lambda$CDM model with $\sigma_{8}=1$ is also shown in Figure \[fig:var\]. The model is displayed for median galaxy redshifts of $z_{m}$ from 0.8 to 1.0, corresponding approximately to the range of depths of the top five surveys displayed (van Waerbeke et al. 2002a; Brown et al. 2003; Bacon et al. 2002, WHT and Keck; Refregier et al. 2002). These observations are approximately consistent with each other and with the $\Lambda$CDM model on angular scales from 0.7 to 20 arcmins. This is compelling given that these were performed with different telescopes (and therefore different instrumental systematics) and independent data-analysis pipelines. The bottom two surveys (Hoekstra et al. 2002b; Jarvis et al. 2002) have a median redshift in the range $z_{m} \simeq 0.6$–$0.7$ and yield lower shear variances. As indicated by the theoretical curves in figure \[fig:var\], this redshift dependence is expected in CDM models and thus confirms the detection of a cosmic shear signal. ------------------------------------------------------------------------------------------------------------------------------------------------------------ Reference Telescope Area (deg$^{2}$) Mag. limit $\sigma_{8}$[^1] --------------------------- ------------ ------------------ ------------------------------------------------------------------- ---------------------------- Wittman et al. 2000 CTIO 1.0 $R {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle \sim}\,$}}26$ van Waerbeke et al. 2000 CFHT 1.7 Kaiser et al. 2000 CFHT 0.96 $I{\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle \sim}\,$}}24$, $V {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle \sim}\,$}}25$ Bacon et al. 2000 WHT 0.5 $R {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $1.50^{+0.50}_{-0.50}$ \sim}\,$}}26$ Maoli et al. 2001 VLT 0.65 $I {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $1.03^{+0.03}_{-0.03}$[^2] \sim}\,$}}24.5$ Rhodes et al. 2001 HST/WFPC2 0.05 $I {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.91^{+0.25}_{-0.30}$ \sim}\,$}}26$ van Waerbeke et al. 2001a CFHT 6.5 $I {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.88^{+0.02}_{-0.02}$[^3] \sim}\,$}}24.5$ Hämmerle et al. 2002 HST/STIS 0.02 Hoekstra et al. 2002a CFHT, CTIO 24 $R {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.81^{+0.07}_{-0.09}$ \sim}\,$}}24$ van Waerbeke et al. 2002a CFHT 8.5 $I {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.98^{+0.06}_{-0.06}$ \sim}\,$}}24.5$ Refregier et al. 2002 HST/WFPC2 0.36 $\langle I \rangle \simeq 23.5$ $0.94^{+0.14}_{-0.14}$ Bacon et al. 2002 WHT, Keck 1.6 $R {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.97^{+0.13}_{-0.13}$ \sim}\,$}}26$ Hoekstra et al. 2002b CFHT, CTIO 53 $R {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.86^{+0.04}_{-0.05}$ \sim}\,$}}24$ Brown et al. 2003 MPG/ESO 1.25 $R {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.72^{+0.09}_{-0.09}$ \sim}\,$}}25$ Hamana et al. 2003 Subaru 2.1 $R {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.69^{+0.18}_{-0.13}$ \sim}\,$}}26$ Jarvis et al. 2002 CTIO 75 $R {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle $0.71^{+0.06}_{-0.08}$ \sim}\,$}}23$ ------------------------------------------------------------------------------------------------------------------------------------------------------------ Although the shear variance was displayed for the purpose of Figure \[fig:var\], some of the groups have used instead the shear correlation function or $M_{ap}$ statistic to quantify their lensing signal. Recently, several groups used the latter statistic to separate their signal into $E$ and $B$ components (see Section 3) and thus to estimate and subtract systematic effects (Jarvis et al. 2002; Hamana et al. 2003; Hoekstra, Yee & Gladders 2002; van Waerbeke et al. 2001, 2002a). Pen, van Waerbeke & Mellier (2001) and Brown et al. (2003) chose instead to measure directly the shear power spectrum for each $E$ and $B$ mode. Figure \[fig:hoek\]a shows the measurement of the $E$ and $B$ signals using the $M_{ap}$ statistic by Hoekstra, Yee & Gladders (2002a), for several $R_{c}$-magnitude ranges in the Red-Sequence Cluster (RCS) survey. They measure a clear lensing signal apparent as $E$ modes. However, the significant $B$ modes reveals the presence of residual systematics on small scales ($\theta {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle \sim}\,$}}10'$), especially for bright galaxies. In their analysis, these authors use the amplitude of the $B$ modes as a measure systematic uncertainties. Existing cosmic-shear measurements already yield interesting constraints on the amplitude of the matter power spectrum $\sigma_{8}$ on which the lensing signal strongly depends. For instance, Figure \[fig:hoek\]b shows cosmological constraints for a $\Lambda$CDM model derived by Hoekstra, Yee & Gladders (2002a) from the measurement of the $M_{ap}$ statistic for their $22<R_{c}<24$ galaxy sample. The degeneracy between $\sigma_{8}$ and $\Omega_{m}$ apparent in the figure is typical of cosmic-shear measurements involving only two-point statistics. Hoekstra, Yee & Gladders (2002a) found that the constraints are well described by $\sigma_{8} \Omega_{m}^{0.52}=0.46^{+0.05}_{-0.07}$ (95% CL), where priors from CMB and galaxy survey data have been used to marginalize over $\Gamma$ and $\Omega_{\Lambda}$. Other surveys find constraints of the same form with similar exponents for $\Omega_{m}$. Table \[tab:surveys\] lists the values of $\sigma_{8}$ found by the different groups. The errors have been converted to 68% confidence level (CL) when necessary, and values of $\Omega_{m}=0.3$, $\Omega_{\Lambda}=0.7$, and $\Gamma=0.21$ have been assumed, when possible. The most recent values are displayed in Figure \[fig:sigma8\] (see color insert), along with their average $\sigma_{8} \simeq 0.83 \pm 0.04$ ($1\sigma$). The values derived by the different cosmic shear groups are in the range $\sigma_{8}\simeq 0.7$–$1.0$, with the most recent measurements (Brown et al. 2003; Hamana et al. 2003; Jarvis et al. 2002) yielding lower values. The $2$–$3\sigma$ mutual inconsistencies between some of the measurements may be symptomatic of a small level of residual systematics. In particular, calibration errors in the shear measurement method would not be detected via the $E$-$B$ decomposition and are a likely explanation (see Hirata & Seljak 2003). Another source of discrepancy arises from the different fitting functions for the non-linear evolution of the power spectrum (see section \[theory\]). Apart from Brown et al. (2003; see also Contaldi et al. 2003) who used the more accurate results of Smith et al. (2003), the other groups used the Peacock & Dodds (1997) fitting function which yields an underprediction of $\sigma_{8}$ by roughly 5–10%. The impact of these systematics will be discussed further in section \[challenges\]. Interestingly, an independent measurement of $\sigma_{8}$ is provided by the abundance of X-ray clusters and can be directly compared to this value. Initially, this method yielded normalisations of $\sigma_{8}\sim 0.9$ (Eke et al. 1998; Pierpaoli, Scott & White 2001; Viana & Liddle 1999), in agreement with the early cosmic shear results (see table \[tab:surveys\] and figure \[fig:sigma8\]). This value was subsequently revised downward to $\sigma_{8}\sim 0.7$–$0.8$ by the use of the observed, rather than simulated, mass-temperature relation in clusters (Borgani et al. 2001; Reiprich & Böhringer 2001; Seljak 2001; Viana, Nichol & Liddle 2001; Pierpaoli, Scott & White 2002). Recently, Spergel et al. (2003) derived a value of $\sigma_{8}=0.84\pm0.04$ (68%CL) from a joint analysis of CMB anisotropy measurements from the Wilkinson Microwave Anistropy Probe (WMAP) and other experiments, galaxy clustering and the Lyman $\alpha$ forest. For comparison, these results along with a representative value of the revised cluster-abundance normalization (Pierpaoli, Scott & White 2002) are also shown in figure \[fig:sigma8\]. The average of the cosmic shear results is formally in good agreement with the determination of $\sigma_{8}$ using the other techniques. Future surveys are however needed to confirm this, by resolving the discrepancies between the current cosmic shear measurements. Also, a full likelihood analysis would be required to establish the significance of the agreement between the different techniques (see Contaldi et al. 2003). This comparison is important as it provides a strong test of the $\Lambda$CDM model, the gravitational instability paradigm, the physics of clusters, and of the biased formation of galaxies. As discussed in Section 3, the degeneracy between $\sigma_{8}$ and $\Omega_{m}$ can be broken by measuring higher-order correlation functions of the lensing field. Bernardeau, Mellier, & van Waerbeke (2002) recently reported the first detection of a non-Gaussian shear signal using the 3-point shear correlation function formalism of Bernardeau, van Waerbeke & Mellier (2003). Another measure of non-gaussianity was recently performed by Miyazaki et al. (2002) using peak statistics in their Subaru survey. Although these results are consistent with that expected from structure formation models, larger survey areas are needed to infer cosmological constraints. Another approach to probe the dark matter is to measure the bias between the mass and galaxies by cross-correlating the shear map with that of the light from foreground galaxies in the same region of the sky (Cooray 2002, Schneider 1998, van Waerbeke 1998). Hoekstra, Yee & Gladders (2001a), Hoekstra et al. (2002b), and Wilson, Kaiser & Luppino (2001) have recently measured this cross-correlation on large scales. FUTURE SURVEYS AND PROSPECTS {#future} ============================ The existing measurements described above are primarily limited by statistics. They will therefore be improved upon by ongoing surveys on existing telescopes, such as the Legacy Survey on CFHT (CFHTLS; Mellier et al. 2000), the Deep Lens Survey (Wittman et al. 2002), surveys with the Subaru telescope (Hiroyasu et al. 2001) and the Sloan Digital Sky Survey (Stebbins, McKay & Frieman 1995). Future instruments dedicated to surveys and for which cosmic shear is a primary science driver are being planned, such as Megacam on CFHT (Boulade et al. 2000), the Visible and Infrared Survey Telescope for Astronomy (VISTA; Taylor et al. 2003), the Large aperture Synoptic Survey Telescope (LSST; Tyson et al. 2002a,b), or the novel Panoramic Survey Telescope and Rapid Response System (Pan-STARRS; Kaiser, Tonry & Luppino 2000). From space, the new Advanced Camera for Surveys (ACS) on HST and, much more ambitiously, the future Supernova Acceleration Probe satellite (SNAP; Perlmutter et al. 2003; Rhodes et al. 2003; Massey et al. 2003; Refregier et al. 2003) also offer exciting prospects. Table \[tab:future\] lists the characteristics of some of these future surveys. Broadly speaking, ground-based measurements will cover large areas, whereas space-based surveys will yield higher-resolution maps and reduced systematics thanks to the absence of atmospheric seeing. ------------------ -------------- -------------- -------------- -------------- ------------- ------------ Telescope/Survey Ground/Space Diameter FOV Area Start$^{a}$ Ref.$^{b}$ (m) (deg$^{2}$) (deg$^{2}$) DLS ground $2\times4$ $2\times0.3$ 28 1999$^{c}$ 1 CFHTLS ground 3.6 1 172 2003 2 VST ground 2.6 1 $x$100$^{d}$ 2004 3 VISTA$^{e}$ ground 4 2 10000 2007 4 Pan-STARRS ground $4\times1.8$ $4\times4$ 31000 2008 5 LSST ground 8.4 7 30000 2012 6 SNAP space 2 0.7 300 2011 7 ------------------ -------------- -------------- -------------- -------------- ------------- ------------ [$^{a}$ planned start date of the cosmic shear surveys\ $^{b}$ references: 1: Wittman et al. 2002; 2: Mellier et al. 2000; 3: Kuijken et al. 2002; 4: Taylor et al. 2003; 5: Kaiser et al. 2000; 6: Tyson et al. 2002a,b; 7: Perlmutter et al. 2003, Rhodes et al. 2003, Massey et al. 2003, Refregier et al. 2003\ $^{c}$ the survey will be complete in 2003\ $^{d}$ a survey of several 100 deg$^{2}$ is planned\ $^{e}$ assuming the availability of the optical camera\ ]{} Radio surveys offer another interesting prospect. Ongoing efforts are aimed at detecting cosmic shear with the FIRST radio survey (Chang & Refregier 2002, Kamionkowski et al. 1997, Refregier et al. 1998). The future radio telescopes LOFAR (Low Frequency Array) and SKA (Square Kilometer Array) will yield cosmic-shear measurements of comparable sensitivity to the most ambitious optical surveys (Schneider 1999). The advantages of radio surveys are that they cover large solid angles, that the bright radio sources are at a higher redshift, and that the PSF is fully predictable and reproducible. These future surveys will provide very accurate measurements of cosmological parameters through the measurement of the lensing power spectrum and higher-order statistics. Figure \[fig:cl\_omw\] (see color insert) shows, for instance, the excellent accuracy with which the lensing power spectrum will be measured with the SNAP wide survey. They will also allow us to test some of the foundations of the standard cosmological model. For example, the measurement of the power spectrum on nonlinear scales at different redshift slices, and of the hierarchy of high-order correlation functions, will yield a direct test of the gravitational instability paradigm. The lensing power spectrum can also be used to measure the equation of the state of the dark energy $w$ and thus complement supernovae measurements in the constraining of quintessence models (Benabed & Bernardeau 2001; Hui 1999; Huterer 2001; Hu 2001, 2002; Munshi & Wang 2002; Weinberg & Kamionkowski 2002). Figure \[fig:cl\_omw\] shows, for instance, that a change of 40% in $w$ can easily be measured by SNAP (see Refregier et al. 2003). Cosmic-shear measurements can also be used to test general relativity (Uzan & Bernardeau 2001). Another promising approach to measure weak lensing is to use the fluctuations of the CMB temperature as the background sources (Bernardeau 1997, 1998; Cooray & Kesden 2002; Hirata & Seljak 2002; Hu 2000; Seljak 1996; Seljak & Zaldarriaga 1999; Zaldarriaga & Seljak 1998). Because these fluctuations are produced at a redshift of approximately 1100, CMB lensing provides a probe of the evolution of mass fluctuations at redshifts larger than those probed by optical galaxies. Lensing indeed produces distinct non-Gaussian signatures that can be used to reconstruct the foreground mass distribution and probe the growth of structures. This approach will become feasible with the advent of future CMB missions such as Planck Surveyor or ground-based instruments with high angular resolutions. CHALLENGE: SYSTEMATIC EFFECTS {#challenges} ============================= For the future surveys described above to yield their full potential, a number of challenges must be met. Indeed, these surveys require the measurement of shears on the order of 1% with an accuracy better than 0.1%; thus, they require a very tight control of systematic effects. In the following, we review the main sources of systematics. PSF Anisotropy -------------- The most serious systematic for ground-based surveys is that produced by the rather large PSF ellipticities ($\sim 10\%$) observed by the different groups. The KSB method (see Section 4) provides a correction of the PSF anisotropy by, at most, a factor of approximately 10, which is insufficient for a precision of 0.1% in shear. Although new shear measurement methods may improve upon this (see Section 4 and below), the correction is fundamentally limited by the finite surface density and signal-to-noise ratio of the stars used to measure the PSF shape. It is, therefore, necessary for the PSF ellipticity to be reduced in hardware as well as in software. This can be done by using tighter constraints in the tracking and optical systems of the new telescopes and instruments as well as with interactions between telescope designers, engineers, and scientists. Shear Measurement Method ------------------------ Even if the PSF ellipticity were guaranteed to be isotropic, future measurements will be limited by the precision of the shear measurement methods. The KSB method is accurate to measure shears of approximately 1% with only a 10% accuracy (Bacon et al. 2001, Erben et al. 2001). This will soon become an important limiting factor (see Hirata & Seljak 2003). Although new methods (see Section 4) promise to improve upon this, more extensive simulations are required to establish that the same accuracy can be reached for shears of 0.1%. Redshift Distribution --------------------- To convert cosmic-shear measurements into constraints on cosmological parameters, the redshift distribution of the background galaxies must be known (see Section 4). The uncertainty in the median galaxy redshift is already one of the dominant contributions to the uncertainty in the amplitude of the matter power spectrum from cosmic shear. The determination of the galaxy redshift distribution is made difficult by the depth of the cosmic-shear surveys. In addition, the sample of galaxies is not simply magnitude limited, but it is subject to complex selection cuts throughout the shear measurement pipeline. Dedicated spectroscopic surveys and photometric redshift studies (such as that by Brown et al. 2003) are thus required to overcome this limitation. CCD Nonlinearities ------------------ All the shear measurement methods rely on the assumption that the instrumental response is linear. It is therefore important to test whether CCD cameras do not have subtle pixel-to-pixel nonlinearities that would induce biases in the shear measurements. The mean shear offset and the shear gradient across the chip found by van Waerbeke et al. (2001) could be due to this effect. Overlapping Isophotes --------------------- Two neighboring galaxies in an image yield “peanut-shaped” isophotal contours and thus appear to have aligned ellipticities. This overlapping-ellipticities effect may produce a spurious ellipticity correlation signal on small scales. van Waerbeke et al. (2000) suggested that this effect may explain the excess shear signal, which they measured on small scale and which disappeared when close pairs were discarded. Evidence for this effect was also found in the image simulation by Bacon et al. (2001). More extensive and detailed simulations would need to be performed to ascertain and calibrate this effect. Intrinsic Correlations ---------------------- The measurement of the lensing shear relies on the assumptions that, in the absence of lensing, the ellipticities of the galaxies are uncorrelated. However, an intrinsic correlation of galaxy shapes could exist owing to the coupling of the galaxy angular momentum or shape to the tidal field or to galaxy interactions (see Heavens 2001 for a review). Theoretical estimation of the size of this effect has been performed using numerical simulations (Croft & Metzler 2000; Heavens, Refregier & Heymans 2000; Jing 2002) and analytical methods (Catalan, Kamionkowski & Blandford 2001; Crittenden et al. 2001a,b; Lee & Pen 2001; Mackey, White & Kamionkowski 2001). Measurements of intrinsic correlations have also been performed (Brown et al. 2000; Pen, Lee & Seljak 2000). Although considerable uncertainty remains regarding the amplitude of this effect, a consensus is arising that intrinsic correlations are likely to be small for the deep current surveys with $z_{m} \sim 1$, but may be dominant for shallower surveys with $z_{m} {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle \sim}\,$}}0.2$ (see however the conflicting results of Jing 2002). Although lensing distortions are coherent over a large redshift range, intrinsic alignments are only significant for small physical separations. Photometric redshifts can thus be used to separate and reduce intrinsic correlations from cosmic-shear signals (Heymans & Heavens 2002; King & Schneider 2002a,b). Another approach is to search for the $B$-type correlation signal produced by intrinsic correlations (Crittenden et al. 2001b). \[intrinsic\] Theoretical Uncertainties ------------------------- Most of the signal in cosmic-shear surveys arises from small scales ($\theta {\raisebox{-0.5ex}{$\,\stackrel{<}{\scriptstyle \sim}\,$}}10'$) and thus from nonlinear structures (Jain & Seljak 1997). The existing prescriptions for computing the nonlinear corrections (Peacock & Dodds 1997; Ma 1998) to the matter power spectrum are only accurate to approximately 10% and disagree at that level with one another (see discussion in Huterer 2001). This theoretical uncertainty will soon become one of the dominating errors in the determination of cosmological parameters from cosmic-shear surveys (see discussion in van Waerbeke et al. 2002). New prescriptions based on more recent N-body simulations such as those by Smith et al. (2003) will help improve the accuracy of the predictions. The problem of the prediction of higher-order statistics is even more difficult, but it is not as pressing given the large uncertainties in the the measurements within current surveys (see Bernardeau, Mellier, van Waerbeke 2002). The inclusion of second-order terms in the weak-lensing approximation will eventually also be required (Cooray 2002, Cooray & Hu 2002). These systematic effects must be controlled to match the statistical accuracy of future surveys. Along with the studies suggested above, further measurements with different colors and comparison of various surveys in overlapping regions would help to control systematics and to test data analysis pipelines. On the theoretical side, larger simulations coupled with advanced analytical techniques will be required. CONCLUSIONS {#conclusion} =========== Cosmic shear has emerged as a powerful method to measure the large-scale structure in the universe. It can be thought of as the measurement of background fluctuations in the space-time metric. Although other methods rely on assumptions relating the distribution of light to that of the mass, weak lensing is based on “clean” physics and can be directly compared to theory. The past three years have yielded impressive observational progress, as the first statistical detections and measurements of cosmic shear have been achieved. In analogy with the CMB, cosmic shear has moved from the COBE era to that of the first generation of anisotropy experiments. However, the measurement of cosmic shear differs from that of CMB anisotropies in several respects. First, the fluctuations are on the order of $10^{-2}$ as opposed to $10^{-5}$, making them easier to measure while retaining the validity of linear calculations. Second, the cosmic-shear field is non-Gaussian and therefore contains more information than that quantified by the power spectrum. Existing cosmic-shear measurements have started to yield significant constraints on cosmological parameters. The measurement of the amplitude of the matter power spectrum $\sigma_{8} \Omega_{m}^{0.5}$ from cosmic shear should soon replace that derived from the local abundance of clusters. This latter technique is indeed limited by the finite number of bright clusters and by systematic uncertainties in the physics of clusters. With better statistics, the angular and redshift dependence of the shear signal as well as with higher-order moments of the convergence field will break the degeneracy between $\sigma_{8}$ and $\Omega_{m}$ and yield constraints on further parameters such as $\Omega_{\Lambda}$ and $\Gamma$. The current measurements of cosmic shear are now primarily limited by statistics. They will therefore be improved upon by a number of upcoming and future instruments such as Megacam, VST, VISTA, LSST, and Pan-STARRS from the ground, and HST/ACS and SNAP in space. For several of these instruments, a weak-lensing survey has been listed as one of the primary science drivers. These surveys will potentially yield measurements of cosmological parameters that are comparable in precision and complementary to those derived from the CMB. They will also be able to address more far-reaching questions in cosmology by measuring parameters beyond the standard model. For instance, they can be used to provide a test of the gravitational instability paradigm, a measure of the equation of state of the dark energy, and a test of general relativity. For these instruments to yield the full promise of cosmic shear, a number of challenges have to be met. First, observationally, systematic effects, such as the PSF anisotropy and CCD nonlinearities, must be controlled and corrected for. From the theoretical point of view, calculations of the nonlinear power spectrum, of high-order statistics, and of the associated errors must be improved to meet the precision of future measurements. The observational and theoretical efforts required to overcome these difficulties are worthwhile given the remarkable promise that cosmic shear offers to cosmology. [Acknowledgments]{} The author thanks his collaborators David Bacon, Richard Massey, Tzu-Ching Chang, Jason Rhodes, and Richard Ellis for numerous fruitful discussions. The author thanks Bhuvnesh Jain, David Bacon, and Henk Hoekstra for their permission to reproduce their figure. He also thanks Tony Tyson, Hervé Aussel and Andy Taylor for precisions regarding the parameters of future cosmic shear surveys. He is also grateful to Richard Ellis, Henk Hoekstra, Gary Bernstein, Yannick Mellier, Richard Massey, and Ivan Valtchanov for useful comments on the manuscript. .25in The Annual Review of Astronomy and Astrophysics is online at http://astro.annualreviews.org Babul, A Lee MH. 1991. [*MNRAS*]{} 250:407 Bartelmann M, King LJ, Schneider P. 2001. [*Astron. Astrophys.*]{} 378:361 Bartelmann M, Schneider P. 1999. astro-ph/9912508 Bacon DJ, Refregier A, Clowe D, Ellis R. 2001. [*MNRAS*]{} 325:1065 Bacon DJ, Refregier A, Ellis R. 2000. [*MNRAS*]{} 318:625 Bacon DJ, Massey R, Refregier A, Ellis R. 2002. astro-ph/0203134 Benabed K, Bernardeau F. 2001. [*Phys. Rev. D*]{} 64:083501 Bennet DP, Rhie SH. 2000. [*GEST Home Page*]{}.\ [http://bustard.phys.nd.edu/GEST. astro-ph/0011466]{} Bernardeau F. 1997. [*Astron. Astrophys.*]{} 324:15 Bernardeau F. 1998. [*Astron. Astrophys.*]{} 338:767 Bernardeau F. 1999. Theoretical and Observational Cosmology. [*Proc. Cargese Summer Sch.*]{}, [*Cargese*]{}, [*France*]{}, ed. M Lachieze-Rey. astro-ph/9901117 Bernardeau F, Mellier Y, van Waerbeke L. 2002. [*Astron. Astrophys.*]{} 389:L28 Bernardeau F, Valageas P. 2000. [*Astron. Astrophys.*]{} 364:1 Bernardeau F, van Waerbeke L, Mellier Y. 1997. [*Astron. Astrophys.*]{} 322:1 Bernardeau F, van Waerbeke L, Mellier Y. 2003. [*Astron. Astrophys.*]{} 397:405 Berstein GM, Jarvis M. 2001. astro-ph/0107431 Blandford RD, Saust AB, Brainerd TG, Villumsen JV. 1991. [*MNRAS*]{} 241:600 Bonnet H, Mellier Y. 1995. [*Astron. Astrophys.*]{} 303:331 Borgani S, Rosati P, Tozzi P, Stanford SA, Eisenhardt PR, et al. 2001. [*Ap. J.*]{} 561:13 Boulade O, Charlot X, Abbon P, Aune S. Borgeaud P. et al. 2000. [*Proc. SPIE*]{} 4008:657 [*Megacam Home Page.*]{}\ [ http://www-dapnia.cea.fr/Phys/Sap/Activites/Projets/Megacam/page.shtml]{} Brown M, Taylor AN, Hambly N, Dye S. 2000. astro-ph/0009499 Brown M, Taylor AN, Bacon, D.J., Gray, M.E., Dye S., Meisenheimer, K., Wolf, C. 2003. [*MNRAS*]{} 341:100 Catalan P, Kamionkowksi M, Blandford R. 2001. [*MNRAS*]{} 320:7 Chang T-C, Refregier A. 2002. [*Ap. J.*]{} 570:447 Contaldi, C, Hoekstra, H., Lewis, A. 2003. submitted to [*Phys. Rev. Letters*]{}. astro-ph/0302435 Cooray A. 2002. astro-ph/0206068 Cooray A, Hu W. 2001a. [*Ap. J.*]{} 548:7 Cooray A, Hu W. 2001b. [*Ap. J.*]{} 554:56 Cooray A, Hu W. 2002. [*Ap. J.*]{} 574:19 Cooray A, Hu W, Miralda-Escudé J. 2000. [*Ap. J.*]{} 535:9 Cooray A, Kesden M. 2002. astro-ph/0204068 Crittenden R, Natarajan P, Pen U, Theuns, T. 2001a. [*Ap. J.*]{} 559:552 Crittenden R, Natarajan P, Pen U, Theuns T. 2001b. astro-ph/0012336 Croft RAC, Metzler CA. 2000. [*Ap. J.*]{} 545:561 Eke V, Cole S, Frenk C, Patrick Henry J. 1998. [*MNRAS*]{} 298:1145 Erben T, van Waerbeke L, Bertin E, Mellier Y, Schneider P. 2001. [*Astron. Astrophys.*]{} 3667:17 Fort B, Mellier Y. 1994. [*Astron. Astrophys. Rev.*]{} 5:239 Gray ME, Taylor AN, Meisenheimer K, Dye S, Wolf C, Thommes E. 2002. [*Ap. J.*]{} 568:141 Gunn JE. 1967. [*Ap. J.*]{} 150:737 Hamana T, Colombi S, Mellier Y. 2000. Cosmological Physics with Gravitational Lensing. [*Proc. XXth Moriond Astrophys. Meet.*]{}, [*Les Arcs*]{}, [*France*]{}, ed. J-P Kneib, Y Mellier, M Mon, J Tran Thanh Van. astro-ph/0009459 Hamana T, Miyazaki S, Shimasaku K, Furusawa H, Doi M, et al. 2002, submitted to [*Ap. J.*]{}, preprint astro-ph/0210450 Hämmerle H, Miralles JM, Schneider P, Erben T, Fosbury RA. 2002. [*Astron. Astrophys.*]{} 385:743 Heavens AF. 2001. [*Intrinsic Galaxy Alignments and Weak Gravitational Lensing*]{}. Yale Worksh. Shapes Galaxies Haloes, May. astro-ph/0109063 Heavens A, Refregier A, Heymans C. 2000. [*MNRAS*]{} 319:649 Heymans C, Heavens A. 2002. astro-ph/0208220 Hirata C, Seljak U. 2002. astro-ph/0209489 Hirata C, Seljak U. 2003. to appear in [*MNRAS*]{}. astro-ph/0301054 Hiroyasu A, et al. 2001. [*Subaru Home Page*]{}. [http://www.subaru.naoj.org/]{} Hoekstra H, Franx M, Kuijken K, Squires G. 1998. [*Ap. J.*]{} 504:636 Hoekstra H, Franx M, Kuijken K, Carlberg RG, Yee HKC, et al. 2001. [*Ap. J.*]{} 548:5 Hoekstra H, Yee HKC, Gladders, M. 2001. [*Ap. J.*]{} 558:11 Hoekstra H, Yee HKC, Gladders M. 2002a. [*Ap. J.*]{} 577:595 Hoekstra H, Yee HKC, Gladders M. 2002b. astro-ph/0205205 Hoekstra H, Yee HKC, Gladders M, Felipe Barrientos L, Hall PB, Infante L. 2002a. [*Ap. J.*]{} 572:55 Hoekstra H, van Waerbeke L, Gladders MD, Mellier Y, Yee HKC. 2002b. [*Ap. J.*]{} 577:604 Hu W. 1999. [*Ap. J.*]{} 522L:21 Hu W. 2000. [*Phys. Rev. D*]{} 63:3504 Hu W. 2001. astro-ph/010890 Hu W. 2002. astro-ph/0208093 Hu W, Keeton CR. 2002. astro-ph/0205412 Hu W, Tegmark M. 1999. [*Ap. J.*]{} 514L:65 Hu W, White M. 2000. [*Ap. J.*]{} 554:67 Hui L. 1999. [*Ap. J.*]{} 519:9 Huterer D. 2001. astro-ph/0106399 Jain B, Seljak U. 1997 [*Ap. J.*]{} 484:560. Jain B, Seljak U, White S. 2000. [*Ap. J.*]{} 530:547 Jain B, van Waerbeke L. 2000. [*Ap. J.*]{} 530:L1 Jaroszyńsky M, Park C, Paczyńsky B, Gott JR. 1990. [*Ap. J.*]{} 365:22 Jarvis, M., Bernstein, G.M., fisher, P., Smith, D., Jain, B., Tyson, J.A., Wittman, D. 2002,[*Ap. J.*]{} 125:1014 Jing YP. 2002. [*MNRAS*]{} 335:L89 Kaiser N. 1992. [*Ap. J.*]{} 388:272. Kaiser N. 1998. [*Ap. J.*]{} 498:26. Kaiser N. 1999. [*Weak Lensing by Galaxy Clusters*]{}. Boston Lensing Meet. astro-ph/9912569 Kaiser N. 2000. [*Ap. J.*]{} 537:555 Kaiser N, Squires G, Broadhurst T. 1995. [*Ap. J.*]{} 449:460 Kaiser N, Tonry JL, Luppino GA. 2000. [*PASP*]{} 112:768. [*Pan-STARRS homepage*]{} [http://pan-starrs.ifa.hawaii.edu/]{} Kaiser N, Wilson G, Luppino GA. 2000. astro-ph/0003338 Kaiser N, Wilson G, Luppino GA, Dahle H. 1999. astro-ph/99077229 Kaiser N, Wilson G, Luppino G, Kofman L, Gioia I, et al. 1998. astro-ph/9809268 Kamionkoski M, Babul A, Cress CM, Refregier A. 1997. [*MNRAS*]{} 301:1064. astro-ph/9712030 King L, Schneider P. 2002a. astro-ph/0208256 King L, Schneider P. 2002b. astro-ph/0209474 Kristian J, Sachs RK. 1966. [*Ap. J.*]{} 143:379 Kristian J. 1967. [*Ap. J.*]{} 147:864 Kuijken K. 1999. [*Astron. Astrophys.*]{} 352:355 Kuijken K, Bender R, Cappellaro E, Muschielok B, Baruffolo A et al. 2002. [*ESO Messenger*]{} 110: 15; VST homepage [http://www.na.astro.it/vst/]{} Lee MH, Paczyńsky B. 1990. [*Ap. J.*]{} 357:32 Lee J, Pen U. 2001. [*Ap. J.*]{} 532:5 Luppino GA, Kaiser N. 1997. [*Ap. J.*]{} 475:20 Ma C-P. 1998. [*Ap. J.*]{} 508:5 Mackey J, White M, Kamionkowksi M. 2001. astro-ph/0106364 Massey R., Rhodes J., Refregier A., Albert J., Bacon D. et al. 2003. astro-ph/0304418 Maoli R, van Waerbeke L, Mellier Y, Schneider P, Jain B, et al. 2001. [*Astron. Astrophys.*]{}, 368:766 Mellier Y. 1999. [*Annu. Rev. Astron. Astrophys.*]{} 37:127 Mellier Y, van Waerbeke L, Bertin E, Tereno I, Bernardeau F. 2002. astro-ph/0210091 Mellier Y, van Waerbeke L, Maoli R, Schneider P, Jain B, et al. 2001. Cosmic shear surveys. [*Deep Fields, Proc. Eur. South. Obs.*]{}, [Oct.]{}, [*Garching*]{}, [Ger.]{} astro-ph/0101130 Mellier Y, van Waerbeke L, Radovich M, Bertin E, Dantel-Fort M. 2000. [*ESO Proceedings, Mining the Sky, Garching, July 2000, A.J. Banday et al eds. astro-ph/0012059*]{}. [*CFHTLS homepage [http://cdsweb.u-strasbg.fr:2001/Science/CFHLS/]{}*]{} Mould J, Blandford R, Villumsen J, Brainerd T, Smail I. 1994. [*MNRAS*]{} 271:31 Miyazaki S, Hamana T, Shimasaku K, Furusawa H, Doi M, et al. 2002. [*Ap. J.*]{} 580:97 Munshi D, Coles P. 2000. astro-ph/0003354 Munshi D, Coles P. 2002. astro-ph/0003481 Munshi D, Jain B. 2000. [*MNRAS*]{} 318:109 Munshi D. Jain B. 2001. [*MNRAS*]{} 322:107 Munshi D, Wang Y. 2002. astro-ph/0206483 Narayan R, Bartelmann M. 1999. In [*Formation of Structure in the Universe*]{}, ed. A Dekel, JP Ostriker, p. 360. astro-ph/9606001 Padmanabhan N, Seljak U, Pen UL. 2002. [*New Astronomy*]{} 8:581 Peacock J, Dodds SJ. 1997. [*MNRAS*]{} 280:L19 Pen U, Lee J, Seljak U. 2000. [*Ap. J.*]{} 543:L107 Pen U, van Waerbeke L, Mellier Y. 2001. astro-ph/0109182 Perlmutter, et al. 2003. [*SNAP Home Page*]{}. [http://snap.lbl.gov]{} Pierpaoli E, Scott D, White M. 2001. [*MNRAS*]{} 325:77 Pierpaoli E, Scott D, White M. 2002. submitted to [*MNRAS*]{}. astro-ph/0210567 Premadi P, Martel H, Matzner R, Futamase T. 2001. [*Ap. J. Suppl.*]{} 135:7 Refregier A. 2003. [*MNRAS*]{} 338:35 Refregier A, Bacon DJ. 2003. [*MNRAS*]{} 338:48 Refregier A, Rhodes J, Groth E. 2002. [*Ap. J.*]{} 572:L131 Refregier A, Brown ST, Kamionkowski M, Helfand DJ, Cress CM, et al. 1998. Wide Field Surveys in Cosmology. [*Proc. XIVth IAP Meet.*]{}, ed. Y Mellier, S Colombi. Paris. astro-ph/9810025 Refregier A, Massey M, Rhodes J, Ellis R, Albert J, et al. 2003. astro-ph/0304419 Reiprich TH, Böhringer H. 2001. astro-ph/0111285 Rhodes J, Refregier A. Groth E. 2000. [*Ap. J.*]{} 536:79 Rhodes J, Refregier A, Groth E. 2001. [*Ap. J.*]{} 552:L85 Rhodes J., Refregier A., Massey R., Albert, J., Bacon D., et al. 2003, astro-ph/0304417 Schneider P. 1995. [*Proc. Laredo Adv. Summer Sch.*]{}, Sept. astro-ph/9512047 Schneider P. 1998. [*Ap. J.*]{} 498:43 Schneider P. 1999. [*Proc. Perspec. Radio Astron.*]{}, [*April*]{}, [*Amsterdam*]{}. astro-ph/9907146 Schneider P, Lombardi M. 2002. astro-ph/0207454 Schneider P, van Waerbeke L, Jain B, Kruse G. 1998a. [*MNRAS*]{} 296:873 Schneider P, van Waerbeke L, Kilbinger M, Mellier Y. 2002. [*Astron. Astrophys*]{} 396:1 Schneider P, Weiss A. 1988. [*Ap. J.*]{} 327:526 Schneider P, van Waerbeke L, Mellier Y, Jain B, Seitz S, Fort B. 1998b. [*Astron. Astrophys.*]{} 333:767 Seljak U. 1996. [*Ap. J.*]{} 463:1 Seljak U. 2001. astro-ph/0111362 Seljak U, Zaldarriaga M. 1999. [*Phys. Rev. D*]{} 60:43504 Smith RE, Peacock JA, Jenkins A, White SDM, Frenk CS, et al. 2003. [*MNRAS*]{} 341:1311 Spergel, D., Verde, L., Peiris, H.V., Komatsu, E., Nolta, M.R., et al. 2003, submitted to [*Ap. J.*]{}. astro-ph/0302209 Stebbins A. 1996. astro-ph/9609149 Stebbins A, McKay T, Frieman JA. 1995. [*Proc. IAU Symposium 173*]{}. astro-ph/9510012 Takada M. Jain B. 2002. astro-ph/0205055 Taylor A. 2001. astro-ph/0111605 Taylor A, Watts P. 2000. astro-ph/0010014 Taylor A, et al. 2003. in preparation [*VISTA Home Page.*]{} [http://www.vista.ac.uk]{} Tyson JA, Wittman D, Hennawi JF, Spergel DN. 2002a. [*Proc. 5th Int. UCLA Symp. Sources Detect. Dark Matter*]{}, [Feb.]{}, [*Marina del Rey*]{}, ed. D Cline. astro-ph/0209632 Tyson JA, & the LSST collaboration 2002b. [*Proc. SPIE Int.Soc.Opt.Eng.*]{} 4836, 10-20. astro-ph/0302102. [*LSST Home Page*]{} [http://lsst.org]{} Uzan J-P, Bernardeau F. 2001. [*Phys. Rev. D*]{} 64:083004 Valageas P. 2000. [*Astron. Astrophys.*]{} 356:771 Valdes F, Jarvis JF, Tyson JA. 1983. [*Ap. J.*]{} 271:431 van Waerbeke L. 1998. [*Astron. Astrophys.*]{} 334:1 van Waerbeke L, Bernardeau F, Mellier Y. 1999. [*Astron. Astrophys.*]{} 342:15 van Waerbeke L, Mellier Y., Erben T., Cuillandre JC, Bernardeau F, et al. 2000. [*Astron. Astrophys.*]{} 358:30 van Waerbeke L, Mellier Y., Radovich M., Bertin E., Dantel-Fort M. et al. 2001a. [*Astron. Astrophys.*]{} 374:757 van Waerbeke L, Hamana T, Scoccimarro R, Colombi S, Bernardeau F. 2001b. [*MNRAS*]{} 322:918 van Waerbeke L, Mellier Y, Pelló R, Pen U-L, McCracken HJ, Jain B. 2002. [*Astron. Astrophys.*]{} 393:369 Viana P, Liddle A. 1999. [*MNRAS*]{} 303:535 Viana P, Nichol RC, Liddle A. 2001. astro-ph/0111394 Villumsen J. 1995. astro-ph/9507007 Villumsen J. 1996. [*MNRAS*]{} 281:369 van Waerbeke L, Mellier Y, Tereno I. 2002. astro-ph/0206245 Wambsganss J, Cen R, Ostriker JP. 1998. [*Ap. J.*]{} 494:29 Weinberg NN, Kamionkowski M. 2002. astro-ph/0210134 White M, Hu W. 2000. [*Ap. J.*]{} 537:1 Wilson G, Kaiser N, Luppino GA. 2001. [*Ap. J.*]{} 556:601 Wittman DM, Tyson J, Kirkman D, Dell’Antonio I, Bernstein G. 2000. [*Nature*]{} 405:143 Wittman DM. 2002. [*Dark Matter and Gravitational Lensing*]{}, [*LNP Top. Vol.*]{}, ed. F Courbin, D Minniti. Springer-Verlag. astro-ph/0208063 Wittman DM, Tyson JA, Dell’Antonio IP, Becker AC, Margoniner VE, et al. 2002. [*Proc. SPIE*]{} 4836 v.2. astro-ph/0210118. [*Deep Lens Survey web page*]{} [http://dls.bell-labs.com/]{} Zaldarriaga M, Scoccimarro R. 2002. astro-ph/0208075 Zaldarriaga M, Seljak U. 1998. astro-ph/9810257 [^1]: for $\Lambda$CDM model with $\Omega_{m}=0.3$, $\Omega_{\Lambda}=0.7$; $\Gamma$ is marginalised over or set to $0.21$ when possible; errors correspond to 68%CL [^2]: for combination of Maoli et al. (2001), van Waerbeke et al. (2000), Bacon et al. (2000), and Wittman et al. (2000); cosmic variance not included. [^3]: cosmic variance not included.

ArXiv

--- abstract: 'In this paper, we propose a method of improving temporal Convolutional Neural Networks (CNN) by determining the optimal alignment of weights and inputs using dynamic programming. Conventional CNN convolutions linearly match the shared weights to a window of the input. However, it is possible that there exists a more optimal alignment of weights. Thus, we propose the use of Dynamic Time Warping (DTW) to dynamically align the weights to the input of the convolutional layer. Specifically, the dynamic alignment overcomes issues such as temporal distortion by finding the minimal distance matching of the weights and the inputs under constraints. We demonstrate the effectiveness of the proposed architecture on the Unipen online handwritten digit and character datasets, the UCI Spoken Arabic Digit dataset, and the UCI Activities of Daily Life dataset.' address: 'Department of Advanced Information Technology, Kyushu University, Fukuoka, Japan' bibliography: - 'master.bib' - 'deep.bib' - 'dtwnn.bib' title: DYNAMIC WEIGHT ALIGNMENT FOR TEMPORAL CONVOLUTIONAL NEURAL NETWORKS --- Time series classification, convolutional neural network, dynamic programming, dynamic time warping Introduction {#sec:intro} ============ Neural networks and perceptron learning models have become a powerful tool in machine learning and pattern recognition. Early models were introduced in the 1970s, but recently have achieved state-of-the-art results due to improvements in data availability and computational power [@schmidhuber2015deep]. Convolutional Neural Networks (CNN) [@lecun1998gradient] in particular have achieved the state-of-the-art results in many areas of image recognition, such as offline handwritten digit recognition [@wan2013regularization], text digit recognition [@lee2016generalizing; @uchida2016further], and object recognition [@clevert2015fast; @he2016deep]. Most recent successes in time series recognition have been through the use of Recurrent Neural Networks (RNN) [@jaeger2002tutorial] and in particular, Long Short-Term Memory (LSTM) networks [@Hochreiter_1997]. Typically, CNN-based models have been used in the image domain, however, they have also been used for time series patterns. A predecessor to CNNs, Time Delay Neural Networks (TDNN) [@waibel1989phoneme; @lang1990time] used time-delay windows similar to the filters of CNNs. CNNs were also used to classify time series by embedding the sequences into vectors [@Zheng_2014] and matrices [@razavian2015temporal; @yang2015deep]. CNNs use sparsely connected shared weights that act as a feature extractor and maintain the structural aspects from the input. In particular, these shared weights are linearly aligned to each corresponding window value of the input. However, the linear alignment assumes that each element of the input window correspond directly to each weight of the filter in a one-to-one fashion. It is possible that there is a more optimal alignment of the shared weights and the input values. We propose a method of finding that alignment using dynamic programming, namely Dynamic Time Warping (DTW) [@sakoe1978dynamic]. DTW estimates the globally minimal distance between two time series patterns by elastically matching elements using dynamic programming along a constrained path on a cost matrix. While DTW is traditionally used just as a distance measure, we exploit the elastic matching byproduct of DTW to align the weights of the filter to the elements of the corresponding receptive field to create more efficient feature extractors for CNNs. The contribution of this paper is twofold. First, we propose a novel method of aligning weights within the convolutional filters of CNNs by dynamically matching the weights to similar input values. Using the discovered dynamic weight alignment, we create a nonlinear matching to create more effective convolutions. Second, we demonstrate the effectiveness of the proposed method on multiple time series datasets including: Unipen online handwritten character datasets, the UCI Spoken Arabic Digit dataset, and the UCI Activities of Daily Life dataset and perform a comparative study to reveal the benefits of the proposed weight alignment. Relation to prior work {#sec:related} ====================== Dynamic neural networks is an emerging field in neural model learning Dynamic Filter Networks (DFN) [@de2016dynamic] use filter-generating networks to produce filters that are used depending on the input. Dynamic Convolutional Neural Networks (DCNN) [@kalchbrenner2014convolutional] use dynamic $k$-Max Pooling to simplify CNNs for sentence modeling. Deformable Convolutional Networks [@dai2017deformable] use deformable convolutions to relax the constraints of a traditional convolutional window. DTW-NNs [@iwana2016robust] similarly use DTW as a nonlinear inner product for regular feed forward neural networks. The distinction between these models and the proposed method is that we use dynamic programming to estimate the optimal weight alignment within convolutions. Dynamic Weight Alignment for CNNs {#sec:nonlinear} ================================= The goal of the proposed method is to exploit dynamic programming to determine the optimal alignment of weights for convolutional layers in CNNs. In this case, we define “optimal” as the globally minimal warping path determined by DTW. In other words, instead of the conventional linear inner product of a convolution, the convolutional filter weights and the input window values are dynamically matched to minimize the difference between similar features of the weights and the input values. Figure \[fig:cnncompare\] demonstrates the difference between a conventional convolutional layer with linear weight alignment and the proposed CNN with dynamic weight alignment. Convolutional Neural Networks ----------------------------- A CNN is an artificial neural network which contains one or more convolutional layers. The key features of convolutional layers is that they have sparse connectivity and use parameter sharing. Specifically, the weights of a convolutional layer are shared for each corresponding output element’s local receptive field. In this way, a forward calculation of a convolutional layer is identical to a convolution operation where the shared weights are the filter and the output is a feature map. Formally, the feature map $z^{(l)}_{j}$ of a convolutional layer is defined as: $$\label{eq:conv} z^{(l)}_{j}=\sum^{I-1}_{i=0} w^{(l)}_{i} a^{(l-1)}_{i+j}+b^{(l)}$$ for each element $j$, where $l$ is the convolutional layer, $l-1$ is the previous layer, $i$ is the index of the filter, and $I$ is the window size. We denote $w^{(l)}_{i}$, $a^{(l-1)}_{i+j}$, and $b^{(l)}$ as the shared weights, the previous layer activations, and the bias respectively. In other words, $z^{(l)}_{j}$ is the inner product of the shared weights $\mathbf{w}^{l}$ and each window of the previous layer $a^{(l-1)}_{j},\dots,a^{(l-1)}_{j+(I-1)}$. This inner product linearly matches the weights to the inputs within the window. However, it is plausible that there exist instances where particular weights should be matched with more optimal inputs, for example noisy elements or feature translation and scale variance within the filter. Dynamic Weight Alignment {#sec:weight} ------------------------ The conventional inner product of a convolution acts much like a similarity function. Thus, the general idea is to align the weights so that there is a stronger activation to input windows that are similar but only slightly misaligned. To optimize the alignment of weights, we adopt a dynamic programming solution, specifically DTW. ### Dynamic Time Warping DTW is an asymmetric positive semi-definite similarity function that is traditionally used as a distance measure between sequences. It is calculated using dynamic programming to determine the optimal match of elements between two sequences. By matching elements, the sequences are *warped* in the time dimension to align similar features of the time series. DTW finds the total cost over an optimal warping path of a local cost matrix using dynamic programming. Given two discrete time series, sequence $\mathbf{p}=p_1,\dots,p_i,\dots,p_I$ of length $I$ and sequence $\mathbf{s}=s_1,\dots,s_j,\dots,s_J$, where $i$ and $j$ are the index of each time step and $p_i$ and $s_j$ are elements at each time step, the DTW-distance is the global summation of local distances between pairwise element matches. Namely, the DTW-distance is denoted as: $$\label{eq:dtwdef} \mathrm{DTW}(\mathbf{p},\mathbf{s})=\sum_{(i',j')\in\mathcal{M}}\left|\left| p_{i'} - s_{j'} \right|\right|,$$ where $(i',j')$ is a pair of matched indices $i'$ and $j'$ corresponding to the original indices $i$ of $\mathbf{p}$ and $j$ of $\mathbf{s}$, respectively. The set $\mathcal{M}$ contains all matched pairs of $i'$ and $j'$. Additionally, the set of matched pairs $\mathcal{M}$ can contain repeated and skipped indices of $i$ and $j$ from the original sequences, therefore, $\mathcal{M}$ has a nonlinear correspondence to $1,\dots,i,\dots,I$ and $1,\dots,j,\dots,J$. $||\cdot||$ is a local distance function between elements. ### Dynamic Weight Alignment with Shared Weights The forward pass calculation is done in two steps. First, DTW is calculated between the shared weights of each convolution and the receptive field window of the input. This is possible if we consider the weights of the convolution as the time series $\mathbf{p}$ and the window of the input as $\mathbf{s}$. The result is a mapping of the shared weights to the input values based on minimizing the $L^2$ distance between sequence elements. Second, the convolution is calculated using the stored mapping. Namely, we propose using DTW to determine $\mathcal{M}_j$ and then calculate the result of the convolution $z^{(l)}_{j}$: $$\begin{aligned} z^{(l)}_{j}&=\sum_{(i', j')\in \mathcal{M}_j} w^{(l)}_{i'} a^{(l-1)}_{j'}+b^{(l)}, \label{eq:weightalignment}\end{aligned}$$ where $\mathcal{M}_j$ is the set of matched indices $i'$ and $j'$ corresponding to the index $i$ of $\mathbf{w}^{(l)}$ and the index $j$ in $\mathbf{a}^{(l-1)}$ , respectively. When used in this manner, we create a nonlinear convolutional filter that acts as a feature extractor similar to using shapelets with DTW [@ye2009time]. In addition, it is important to note that unlike a conventional CNN, the set of matched indicies $\mathcal{M}_j$ allows for duplicate and skipped values of $w^{(l)}_{i'}$ and $a^{(l-1)}_{j'}$. The idea is that DTW will match similar features from the filter to the input and skip elements with a very high distance to the weights and perform small translations. Therefore, the process of aligning the weight using DTW is repeated for every stride of the convolution during all forward passes including during training and testing. Consequently, the alignment is only kept for the immediate forward and backward round and recalculated on the fly for subsequent iterations. Backpropagation of Convolutions with Dynamic Weight Alignment ------------------------------------------------------------- In order to train the network, Stochastic Gradient Decent (SGD) is used to determine the gradients of the weights with respect to the error. This is done to update the weights in order to minimize the loss. For a CNN, the gradient of the error with respect to the shared weights is the partial derivative: $$\label{eq:cnnback} \frac{\partial C}{\partial {w^{(l)}_{i} }}=\sum_i \frac{\partial C}{\partial {z^{(l)}_{j} }} \frac{\partial {z^{(l)}_{j} }}{\partial {w^{(l)}_{i} }},$$ where $C$ is the loss function. In a conventional CNN, $w^{(l)}_{i}$ has a linear relationship to $z^{(l)}_{j}$, thus $\frac{\partial {z^{(l)}_{j} }}{\partial {w^{(l)}_{i} }}$ can be calculated simply. However, given the nonlinearity of the weight alignment, the calculation of the gradient is reliant on the matched elements determined by the forward pass in: $$\begin{aligned} \label{eq:nonlinearback} \frac{\partial C}{\partial {w^{(l)}_{i} }}&=\sum_i \frac{\partial C}{\partial {z^{(l)}_{j} }} \frac{\partial {\left( \sum_{(i', j')\in \mathcal{M}_j} w^{(l)}_{i'} a^{(l-1)}_{j'}+b^{(l)}\right)}}{\partial {w^{(l)}_{i} }} \\&= \delta^{(l+1)} \sum_{(i', j')\in \mathcal{M}_j} a^{(l-1)}_{j'}.\end{aligned}$$ where $\delta^{(l+1)}$ is the backpropagated error from the previous layer as determined by the chain rule. Experiments and Results {#sec:results} ======================= Datasets and Evaluation ----------------------- We demonstrate the effectiveness of the proposed method by quantitatively evaluating the architecture and compare it to baseline methods for three diverse datasets. The Unipen multi-writer 1a, 1b, and 1c datasets [@guyon1994unipen] are constructed from pen tip trajectories of isolated numerical digits, uppercase alphabet characters, and lowercase alphabet characters respectively. The UCI Spoken Arabic Digit Data Set [@Ganchev05comparativeevaluation] contains spoken Arabic digit patterns encoded using 13-frequency Mel-Frequency Cepstrum Coefficients (MFCC) in 10 classes. The UCI Activities of Daily Life (ADL) Recognition with Wrist-worn Accelerometer Data Set [@bruno2013analysis] is made of patterns from 7 classes of ADL actions. The Unipen and the UCI ADL datasets were divided into three sets for training, a test of 10% of the data, a training set of 90% of the data, and 50 patterns set aside from the training set for a validation set. The UCI Arabic data has a pre-defined division of the data with a speaker-independent training set and test set. --------------------------------------------- ---------- ---------- ----------- ---------- ---------- (r)[2-4]{} (r)[5-6]{} Method 1a 1b 1c Arabic ADL Proposed 98.54 96.08 **95.92** 96.95 **90.0** CNN 98.08 94.67 95.33 95.50 87.1 LSTM 96.84 92.31 89.79 96.09 81.4 SVM GDTW [@bahlmann2002online] 96.2 92.4 87.9 – – HMM CSDTW [@bahlmann2004writer] 97.1 92.8 90.7 – – DTW-NN [@iwana2016robust] 96.8 – – – – Google [@keysers2017multi] **99.2** **96.9** 94.9 – – Tree Dist [@hammami2010improved] – – – 93.1 – CHMM – – – **98.4** – $\Delta$($\Delta$MFCC) [@hammami2012second] WNN [@hu2011spoken] – – – 96.7 – GMM + GMR [@bruno2013analysis] – – – – 63.1 Decision Tree [@kannaactivities] – – – – 80.9 --------------------------------------------- ---------- ---------- ----------- ---------- ---------- : Accuracy (%) on the evaluated datasets. The highest accuracy for each dataset is in bold.[]{data-label="tab:results"} Architecture Settings --------------------- For the experiment, we implement a five-layer CNN. The first two hidden layers are convolutional layers with 50 nodes of the proposed dynamically aligned filters. In addition, we use batch normalization [@ioffe2015batch] on the results of the convolutional layers. The third and fourth layers are fully-connected layers with a hyperbolic tangent $\tanh$ activation and have 400 and 100 nodes respectively. The final output layer uses softmax with the number of outputs corresponding to the number of classes. The learning rate $\eta_t$ at iteration $t$ is defined as $\eta_t=\frac{\eta_0}{1+\alpha t}$, where $\eta_0$ is the initial learning rate and $\alpha$ is the decay parameter. For all of the experiments, we use the $1/t$ progressive learning rate with a $\eta_0=0.001$ and $\alpha=0.001$ for the convolutional layers and a static learning rate of 0.0001 between the fully-connected layers. Given that the experimental datasets are made of sequences of different dimensions, the filters should correspond accordingly. The convolutional filters were of size $8 \times 2$ at stride 2, $6 \times 13$ at stride 2, and $12 \times 3$ at stride 4 for the Unipen datasets, the UCI Arabic dataset, and the UCI ADL dataset, respectively. A stride was used to reduce redundant information and decrease computation time. The experiment uses batch gradient decent with a batch size of 100, 50, and 5 for the three datasets respectively and for 60,000 iterations. The batch sizes were selected based on the size of the training sets and were chosen to iterate through epochs at generally the same rates. This is the reason for the very small batch of 5 used for the ADL dataset. In the DTW implementation, we used the asymmetric slope constraint proposed by Itakura [@itakura1975minimum] and Euclidean distance as the local distance function $||\cdot||$ of Eq. . Comparison Methods ------------------ We report classification results literature as well as evaluate the datasets on established state-of-the-art neural network methods. To evaluate the proposed method, we compare the accuracy to current methods from literature. For the online handwritten character evaluations, we compare results from two classical methods, SVM GDTW [@bahlmann2002online] and HMM CSDTW [@bahlmann2004writer], and two state-of-the-art neural network methods, DTW-NN [@iwana2016robust] and Google [@keysers2017multi]. For the spoken Arabic digits, there is one reported neural network solution using a WNN [@hu2011spoken] as well as other models using a Tree Distribution model [@hammami2010improved] and a Continuous HMM of the second-order derivative MFCC (CHMM $\Delta$($\Delta$MFCC)) [@hammami2012second]. For the ADL dataset, we compare our results to the original dataset proposal [@bruno2013analysis] using a Gaussian Mixture Modeling and Gaussian Mixture Regression (GMM + GMR) and the best results of Kanna et al. [@kannaactivities] using a Decision Tree. The evaluated baselines were designed to be direct comparisons for the proposed method. The LSTM is used as the established state-of-the-art neural network method for sequence and time series recognition and a traditional CNN is used as a direct comparison using standard convolutional layers. Both comparative models are provided with the same exact training, test, and validation sets as the proposed method. Furthermore, the evaluated methods use the same batch size and number of iterations as the proposed method for the respective trials. For the LSTM evaluation, an LSTM with two recursive hidden layers, two fully-connected layers, and a softmax output layer was used. The second comparative evaluation was using a CNN with the same exact hyperparameters as the proposed method, but with standard convolutional nodes. Results and Discussion ---------------------- The results of the experiments are shown in Table \[tab:results\]. The results show that the proposed method surpassed all of the results of a conventional CNN as well as the LSTM. Furthermore, the results are competitive with the state-of-the-art methods despite many of them being tailored to the respective datasets and data types. In the online handwriting and ADL experiments, the LSTM performed poorly compared to both CNNs. One reason for the limited performance of the LSTM is that each individual element of those datasets do not contain a significant amount of information and the model needs to know how all elements work together to form spacial structures. For example, for large tri-axial accelerometer data, individual long-term dependencies are not as important as the local and global structures whereas CNNs excels. Another reason for the poor performance of the ADL dataset could be the low amount of training data (600 training samples), high amounts of noise, and a high variation of patterns within each class. However, the LSTM did comparatively well on the spoken Arabic digits. The most important comparison is the conventional CNN with linearly aligned weights against the proposed method with dynamically aligned weights. In addition to the increased accuracy, we observed from Fig. \[fig:results\] that compared to the conventional CNN, the proposed method achieves a higher accuracy during all parts of training but especially during the early stages. This indicates that the nonlinear alignment is able to optimize the weights efficiently. One explanation of the improved accuracy is that aligning the weights to their similar corresponding inputs is more efficient than conventional linear matching. The weights of a convolutional layer learned by a CNN act like filter for feature extraction [@lecun1998gradient]. The purpose of using dynamically aligned weights is to warp the assignment of weights to their most similar corresponding inputs. In this way, noisy input values can be skipped and normally muted but relevant features are enhanced. This provides a more robust convolution. Computational Complexity ------------------------ In the case of the proposed method, the number of elements in the aligned sequences is equal to $I$ and $J$, where $I$ and $J$ is the width of the filter and the input, respectively. Furthermore, the complexity of each DTW calculation is $O(IJ)$, which is required for every application of a convolutional filter. Thus, the computational complexity of the convolutional layer with dynamic weight alignment becomes $O(\frac{N I J^2}{S})$, where $N$ is the number of convolutional nodes and $S$ is the stride. Compared with the standard convolution of a temporal CNN with a complexity of $O(\frac{N I J}{S})$, this a relatively small increase in complexity compared to the overall network. The per classification runtime for the traditional CNN was 0.036s, 0.092s, and 0.029s for the Unipen, ADL, and Spoken Arabic datasets, respectively. The proposed method had runtimes of 0.114s, 0.403s, and 0.078s, respectively. The networks were constructed in Python using Numpy with no GPU on a desktop computer with an Intel Xeon 2.6 GHz CPU. However, these speeds can be further optimized with the use of GPUs and deep learning libraries. Conclusion {#sec:conc} ========== In this paper, we proposed a novel method of optimizing the weights within a convolutional filter of a CNN through the use of dynamic programming. We implemented DTW as a method of sequence element alignment between the weights of a filter and the inputs of the corresponding receptive field. In this way, the weights of the convolutional layer are aligned to maximize their relationship to the data from the previous layer. Furthermore, we show that the proposed model is able to tackle time series pattern recognition. We evaluated the proposed model on a variety of datasets to reach state-of-the-art results. This shows that the proposed method a viable feedforward neural network model for time series recognition and an effective method of optimizing the convolutional filter in CNNs. There is potential for this work to be extended to any CNN-based model.

ArXiv

--- abstract: 'In this paper, we propose the use of spatial and harmonic features in combination with long short term memory (LSTM) recurrent neural network (RNN) for automatic sound event detection (SED) task. Real life sound recordings typically have many overlapping sound events, making it hard to recognize with just mono channel audio. Human listeners have been successfully recognizing the mixture of overlapping sound events using pitch cues and exploiting the stereo (multichannel) audio signal available at their ears to spatially localize these events. Traditionally SED systems have only been using mono channel audio, motivated by the human listener we propose to extend them to use multichannel audio. The proposed SED system is compared against the state of the art mono channel method on the development subset of TUT sound events detection 2016 database [@dcase2016Data]. The usage of spatial and harmonic features are shown to improve the performance of SED.' address: 'Department of Signal Processing, Tampere University of Technology' bibliography: - 'refs.bib' title: SOUND EVENT DETECTION IN MULTICHANNEL AUDIO USING SPATIAL AND HARMONIC FEATURES --- Sound event detection, multichannel, time difference of arrival, pitch, recurrent neural networks, long short term memory Introduction {#sec:intro} ============ A sound event is a segment of audio that a human listener can consistently label and distinguish in an acoustic environment. The applications of such automatic sound event detection (SED) are numerous; embedded systems with listening capability can become more aware of its environment [@environmentalSED][@robot]. Industrial and environmental surveillance systems and smart homes can start automatically detecting events of interest [@surveillance]. Automatic annotation of multimedia can enable better retrieval for content based query methods [@contentRetrieval]. The task of automatic SED is to recognize the sound events in a continuous audio signal. Sound event detection systems built so far can be broadly classified to monophonic and polyphonic. Monophonic systems are trained to recognize the most dominant of the sound events in the audio signal [@monoPhonic2013]. While polyphonic systems go beyond the most dominant sound event and recognize all the overlapping sound events in a segment [@monoPhonic2013][@polyHough2013][@emre2015][@giam2016]. We propose to tackle such polyphonic soundscape which replicates real life scenario in this paper. Some SED systems have tackled polyphonic detection using mel-frequency cepstral coefficients (MFCC) and hidden Markov models (HMMs) as classifiers with consecutive passes of the Viterbi algorithm [@Mesaros2010_EUSIPCO]. In [@nmf], a non-negative matrix factorization was used as a pre-processing step, and the most prominent event in each of the stream was detected. However, it still had a hard constraint of estimating the number of overlapping events. This was overcome by using coupled NMF in [@coupledNmf]. Dennis et al [@polyHough2013] took an entirely different path from the traditional frame-based features by combining generalized Hough transform (GHT) with local spectral features. More recently, the state of the art SED systems have used log mel-band energy features in DNN [@emre2015], and RNN-LSTM [@giam2016] networks trained for multi-label classification. Motivated by the good performance of RNN-LSTM over DNN as shown in [@giam2016], we continue to use the multi-label RNN-LSTM network. The present state of the art polyphonic SED systems have been using a single channel of audio for sound event detection. Polyphonic events can potentially be tackled better if we had multichannel data. Just like humans use their two ears (two channels) to recognize and localize the sound events around them [@iid_itd], we can also potentially train machines to learn sound events from multichannel of audio. Recently, Xiao et al [@multiChannelASR] have successfully used spatial features from multichannel audio for far field automatic speech recognition (ASR) and shown considerable improvements over just using mono channel audio. This further motivates us to use spatial features for SED tasks. In this paper, we propose a spatial feature along with harmonic feature and prove its superiority over mono channel feature even with a small dataset of around 60 minutes. The remaining of the paper is structured as follows. We describe in Section 2 the features used and the proposed approach. Section 3 presents a short introduction to RNNs and long short-term memory (LSTM) blocks. Section 4 presents the experimental set-up and results on a database of real life recordings. Finally, we present our conclusions in Section 5. SOUND EVENT DETECTION {#sec:sed} ===================== The sound event detection task involves identifying temporally the locations of sound event and assigning them to one among the known set of labels. Sound events in real life have no fixed pattern. Different contexts, for example, forest, city, and home have a different variety of sound events. They can be of different sparsity based on the context, and can occur in isolation or be completely overlapped with other sound events. While recognizing isolated sounds have been done with an appreciable accuracy [@isolatedSED], detecting the mixture of labels in an overlapped sound event is a challenging task, where still a considerable amount of improvements can be made. Figure \[fig:multiLabel\] shows a snippet of sound event annotation, where three sound events - speech, car, and dog bark happen to occur. At time frame *t*, two events - speech and car are overlapping. An ideal SED system should be able to handle such overlapping events. The human auditory system has been successfully exploiting the stereo (multichannel) audio information it receives at its ears to isolate, localize and classify the sound events. A similar set up is envisioned and implemented, where the sound event detection system gets a stereo input and suitable spatial features are implemented to localize and classify sound events. The proposed sound event detection system, shown in Figure \[fig:framework\], works on real life multichannel audio recordings and aims at detecting and classifying isolated and overlapping sound events. Three sets of features -log mel-band energies, pitch frequency, and its periodicity, and time difference of arrival (TDOA) in sub-bands, are extracted from the stereo audio. All features are extracted at a hop length of 20 ms to have consistency across features. {width="\columnwidth"} \[fig:framework\] Log mel-band energy {#ssec:mel} ------------------- Log mel-band energies have been used for mono channel sound event detection extensively [@emre2015][@giam2016][@sainath2015] and have proven to be good features. In the proposed system we continue to use log mel-band energies, and extract it for both the stereo channels. This is motivated from the idea that human auditory system exploits the interaural intensity difference (IID) for spatial localization of sound source [@iid_itd]. Neural networks are capable of performing linear operations, which includes the difference. Therefore, when trained on the stereo log mel-band energy data, it will learn to obtain information similar to IID. Each channel of the audio is divided into 40 ms frames with 50% overlap using hamming window. Log mel-band energies are then extracted for each of the frames ($mel$ in Table \[table:1\]). We use 40 mel-bands spread across the entire spectrum. Harmonic features {#ssec:pitch} ----------------- The pitch is an important perceptual feature of sound. Human listeners have evolved to identify different sounds using the pitch cues, and can make efficient use of pitch to acoustically separate each of the mixture in an overlapping sound event [@pitchSED1990]. Uzkent et al [@uzkent2012] have shown improvement in accuracy of non speech environmental sound detection used pitch range along with MFCC’s. Here we propose using the absolute pitch and its periodicity as the features ($pitch$ in Table \[table:1\]). The librosa implementation of pitch tracking [@librosa] on thresholded parabolically-interpolated STFT [@jos] was used to estimate the pitch and periodicity. Since we are handling multi-label classification it is intuitive to identify as many dominant fundamental frequencies as possible and use them to identify the sound events. The periodicity feature gives the confidence measure for the extracted pitch value and helps the classifier to make better decisions based on pitch. The overlapping sound events in the training data (Section \[ssec:dataset\]) did not have more than three events overlapping at a time, hence we have limited ourselves to using the top three dominant pitch values per frame. So, for each of the channels, top three pitch values, and its respective periodicity values are extracted at every frame in 100-4000 Hz frequency range ($pitch3$ in Table \[table:1\]). {width="\columnwidth"} \[fig:multiLabel\] Feature Name Length Description -------------- -------- ------------------------------------------------------------------------------- $mel$ 40 Log mel-band energy extracted on a single channel of audio $pitch$ 2 Most dominant pitch value and periodicity extracted on a single channel $pitch3$ 6 Top three dominant pitch and periodicity values extracted on a single channel $tdoa$ 5 Median of multi-window TDOA’s extracted from stereo audio $tdoa3$ 15 Concatenated multi-window TDOA’s extracted from stereo audio \[table:1\] Time difference of arrival (TDOA) features {#ssec:spatialFeat} ------------------------------------------ Overlapping sound events have forever troubled classification systems. This is mainly because the feature vector for the overlapped frame is a combination of different sound events. But, human listeners have been able to successfully identify each of the overlapping sound events by isolating and localizing the source spatially. This has been possible due to the interaural time delay (ITD) [@iid_itd] Each sound event has its own frequency band, some occur in low frequencies, some in high, and some occur all across the frequency band. If we can divide the frequency spectrum into different bands, and identify the spatial location of the sound source in each of these bands, then this is an extra dimension of the feature, which the classifier can learn to estimate the number of possible sources in each frame, and their orientation in the space. We implement this by dividing the spectral frame into five mel-bands and calculating the time difference of arrival (TDOA) at each of these bands. For example, if a non-overlapping isolated sound event is spread across the entire frequency range, and we are calculating the TDOA in five mel-bands. We should have the same TDOA values for each of the bands. However, if we have two overlapping sounds $S_1$ and $S_2$, where $S_1$ is spread in the first two bands and $S_2$ is spread in the last two bands. The feature vector will have different TDOA values for each of the sounds, which the classifier can learn to isolate and identify them as separate sound events. The TDOA can be estimated using the generalized cross-correlation with phase-based weighting (GCC-PHAT) [@Knapp_Carter-1976]. Here, we extract the correlation for each mel-band separately: $$R_b(\Delta_{12},t) = \sum_{k=0}^{N-1} H_b(k) \frac{ X_1(k,t) \cdot X_2^\ast(k,t)}{\arrowvert X_1(k,t)\arrowvert \arrowvert X_2(k,t) \arrowvert} e^{ i 2\pi k \Delta_{12} / N },$$ where $N$ is the number of frequency bands, $X(k,t)$ is the FFT coefficient of the $k$th frequency band at time frame $t$ and the subscript specifies the channel number, ${H}_b(k)$ is the magnitude response of the $b$th mel-band of total of $B$ bands and $\Delta_{12}$ is the sample delay value between channels. The TDOA is extracted as the location of correlation peak magnitude for each mel-band and time frame. $$\tau(b,t)= \underset{\Delta_{12}}{\textrm{argmax}}\left\{ R_b(\Delta_{12},t)\right\}$$ The maximum and minimum TDOA values are truncated between values $-2\tau_\textrm{max},2\tau_\textrm{max}$, where $\tau_\textrm{max}$ is the maximum sample delay between a sound wave traveling between microphones. The sound events in the training set were seen to be varying from 50 ms to a few seconds. In order to accommodate such variable length sound events, TDOA was calculated in three different window lengths — 120, 240 and 480 ms, with a constant hop length of 20 ms. The TDOA values of these three windows were concatenated for each mel-band to form one set of TDOA features. So, TDOA values extracted in five mel-band, and for three window lengths, on concatenation gives 15 TDOA values per frame ($tdoa3$ in Table \[table:1\]). TDOA values in small windows are generally very noisy and unreliable. To overcome this, the median of the TDOA values from the above three different window lengths for each sub-band of the frame was used as the second set of TDOA features ($tdoa$ in Table \[table:1\]). Post filtering across window lengths, the TDOA values in each mel-band were also median filtered temporally using a kernel of length three to remove outliers. MULTI-LABEL RECURRENT NEURAL NETWORK BASED SOUND EVENT DETECTION {#sec:rnn} ================================================================ Deep neural networks have shown to perform well on complex pattern recognition tasks, such as speech recognition [@graves2013speech], image recognition [@krizhevsky2012imagenet] and machine translation [@bahdanau2014neural]. A deep neural network typically computes a map from an input to an output space through several subsequent matrix multiplications and non-linear activation functions. The parameters of the model, i.e. its weights and biases, are iteratively adjusted using a form of optimization such as gradient descent. When the network is a directed acyclic graph, i.e. information is only propagated forward, it is known as a feedforward neural network (FNN). When there are feedback connections the model is called a recurrent neural network (RNN). An RNN can incorporate information from previous timesteps in its hidden layers, thus providing context information for tasks based on sequential data, such as temporal context in audio tasks. Complex RNN architectures — such as long short-term memory (LSTM) [@hochreiter1997long] — have been proposed in recent years in order to attenuate the vanishing gradient problem [@bengio1994learning]. LSTM is currently the most widely used form of RNN, and the one used in this work as well. In SED, RNNs can be used to predict probabilities for each class to be active in a given frame at timestep $t$. The input to the network is a sequence of feature vectors $\bf{x}(\textmd{t})$; the network computes hidden activations for each hidden layer, and at the output layer a vector of predictions for each class $\bf{y}(\textmd{t})$. A sigmoid activation function is used at the output layer in order to allow several classes to be predicted as active simultaneously. By thresholding the predictions at the output layer it is possible to obtain a binary activity matrix. Neural network configurations {#ssec:nn} ----------------------------- For each recording, we obtain a sequence of feature vectors, which is normalized to zero mean and unit variance, and the scaling parameters are saved for normalizing the test feature vectors. The sequences are further split into non-overlapping sequences of length 25 frames. Each of these frames has a target binary vector, indicating which classes are present in the feature vector. We use a multi-label RNN-LSTM with two hidden layers each having 32 LSTM units. The number of units in the input layer depends on the length of the feature being used. The output layer has one neuron for each class. The network is trained by back propagation through time (BPTT) [@bptt1990] using binary cross-entropy as loss function, Adam optimizer [@adamKeras] and block mixing [@giam2016] data augmentation. Early stopping is used to reduce over-fitting, the training is halted if the segment based error rate (ER) (see Section \[ssec:metrics\]) on the validation set does not decrease for 100 epochs. At test time we use scaling parameters estimated on training data to scale the feature vectors and present them in non-overlapping sequences of 25 frames, and threshold the outputs with a fixed threshold of 0.5, i.e., we mark an event is active if the posterior in the output layer of network is greater than 0.5 and otherwise inactive. [| p[5cm]{} | p[3cm]{} | c | c | c | c | c | c |]{} & Feature combination & & &\ & & ER & F (%) & ER & F (%) & ER & F (%)\ Baseline system using GMM classifier in DCASE 2016 [@dcase2016Data][@dcase2016task3web] & $mfcc;delta;acc$ & 0.96 & 15.9 & 0.86 & 31.5 & 0.91 &23.7\ Mono channel feature With RNN-LSTM network & $mel_1$ & **0.94 & **27.4 & 0.88 & 38.3 & 0.91 & 32.9\ **** & $mel_1;pitch_1$ & 0.97 & 25.4 & 0.85 & 43.4 & 0.91 & 34.4\ & $mel_1;pitch3_1$ & 0.96 & 27.6 & 0.88 & 43.9 & 0.92 & 35.7\ & $mel_1;tdoa$ & 1.02 & 19.4 & 0.89 & 40.2 & 0.96 & 29.8\ & $mel_1;tdoa3$ & 0.98 & 25.9 & 0.87 & 40.5 & 0.92 & 33.2\ & $mel_2$ & 1.03 & 25.4 & 0.84 & 45.9 & 0.93 & 35.6\ & $mel_2;pitch_2$ & 1.03 & 24.9 & 0.93 & 40.9 & 0.98 & 32.9\ & $mel_2;pitch3_2$ & 0.97 & 26.6 & 0.88 & 41.7 & 0.92 & 34.2\ & $mel_2;tdoa$ & 1.01 & 24.4 & **0.82 & **46.4 & **0.91 & **35.4\ & $mel_2;tdoa3$ & 0.96 & 24.9 & 0.86 & 38.5 & 0.91 & 31.7\ & $mel_2;tdoa3;pitch_2$ & 0.97 & 25.7 & 0.85 & 43.1 & 0.91 & 34.4\ & $mel_2;tdoa3;pitch3_2$ & 0.99 & 26.5 & 0.91 & 35.2 & 0.95 & 30.9\ & $mel_2;tdoa;pitch_2$ & 0.98 & 24.7 & 0.87 & 43.8 & 0.92 & 34.2\ & $mel_2;tdoa;pitch3_2$ & 0.94 & 26.3 & 0.89 & 40.5 & 0.91 & 33.4\ ******** \[table:2\] EVALUATION AND RESULTS {#sec:eval} ====================== Dataset {#ssec:dataset} ------- We evaluate the proposed SED system on the development subset of TUT sound events detection 2016 database [@dcase2016Data]. This database has stereo recordings which were collected using binaural Soundman OKM II Klassik/studio A3 electret in-ear microphones and Roland Edirol R09 wave recorder using 44.1 kHz sampling rate and 24-bit resolution. It contains two contexts - home and residential area. Home context has 10 recordings with 11 sound event classes and the residential area context has 12 recordings with 7 classes. The length of these recordings is between 3-5 minutes. In the development subset provided, each of the context data is already partitioned into four folds of training and test data. The test data was collected such that each recording is used exactly once as the test, and the classes in it are always a subset of the classes in the training data. Also, 20% of the training data recordings in each fold were selected randomly to be used as validation data. The same validation data was used across all our evaluations. Metrics {#ssec:metrics} ------- We perform the evaluation of our system in a similar fashion as [@dcase2016Data] which uses the established metrics for sound event detection defined in [@dcase2016Metrics]. The error rate (ER) and F-scores are calculated on one second long segments. The results from all the folds are combined to produce a single evaluation. This is done to avoid biases caused due to data imbalance between folds as discussed in [@apple]. Results {#ssec:results} ------- The baseline system for the dataset [@dcase2016Data] uses 20 static (excluding the 0th coefficient), 20 delta and 20 acceleration MFCC coefficients extracted on mono audio with 40 ms frames and 20 ms hop length. A Gaussian mixture model (GMM) consisting of 16 Gaussians is then trained for each of the positive and negative values of the class. This baseline system gives a context average ER of 0.91 and F-score of 23.%. An ideal system should have an ER of 0 and an F-score of 100%. In Table \[table:2\] we compare the segment based ER and F-score for different combinations of proposed spatial and harmonic features. In all these evaluations, only the size of the input layer changes based on the feature set, with the rest of the configurations in the RNN-LSTM network remaining unchanged. Mono channel audio was created by averaging the stereo channels in order to compare the performance of the proposed spatial and harmonic features for multichannel audio. One of the present state of the art SED system for mono channel is proposed in [@giam2016]. An RNN-LSTM network is trained in a similar fashion with log mel-band energy feature (Section \[ssec:mel\]) and evaluated. Across contexts, the F-score was seen to be better than the GMM baseline system with comparable ER. Here onwards we use this mono-channel log mel-band feature and RNN-LSTM network configuration result as a baseline for comparisons. A set of hybrid combinations were tried as shown in Table \[table:2\]. All combinations other than $mel_1;tdoa$ performed better than the baseline across contexts in F-score. Finally, the full spectrum of proposed spatial and harmonic features were evaluated in different combinations with RNN-LSTM network. With a couple of exceptions - $mel_2;pitch_2$ and $mel_2;tdoa3;pitch3_2$, all the combinations of features performed equal to or better than the baseline in average F-scores, with marginally similar average ER as baseline. Given the dataset size of around 60 minutes, it is difficult to conclusively say that the binaural features are far superior to monaural features; but they surely look promising. Binaural features - $mel_2$ and $mel_2;tdoa;pitch_2$ in Table \[table:3\] were submitted to the DCASE 2016 challenge [@dcase2016task3web], where they were evaluated as the top performing systems. Monaural feature $mel_1$ was submitted unofficially to compare the performance with binaural features. The hyper-parameters of the network were tuned before the submission, and hence the development set results in Table \[table:3\] are different from Table \[table:2\]. Three hidden layers with 16 LSTM units each were used for $mel_2$, while $mel_1$ and $mel_2;tdoa;pitch_2$ were trained with two layers each having 16 LSTM units. ---------------------- ------ ------- ------ ------- ER F (%) ER F (%) $mel_1$ 0.79 46.6 0.90 35.3 $mel_2$ 0.80 47.8 0.88 34.7 $mel_2;tdoa;pitch_2$ 0.88 37.9 0.87 34.8 ---------------------- ------ ------- ------ ------- : Comparison of segment based error rate (ER) and F-score for development and evaluation dataset. The evaluation dataset scores are the result of DCASE 2016 challenge [@dcase2016task3web]. \[table:3\] CONCLUSION {#sec:conclusion} ========== In this paper, we proposed to use spatial and harmonic features for multi-label sound event detection along with RNN-LSTM networks. The evaluation was done on a limited dataset size of 60 mins, which included four cross validation data for two contexts — home and residential area. The proposed multi-channel features were seen to be performing substantially better than the baseline system using mono-channel features. Future work will concentrate on finding novel data augmentation techniques. Augmenting spatial features is an unexplored space, and will be a challenge worth looking into. Concerning the model, further studies can be done on different configurations of RNN like extending them to bidirectional RNN’s and coupling with convolutional neural networks.

ArXiv

--- abstract: 'We propose a semantic segmentation model that exploits rotation and reflection symmetries. We demonstrate significant gains in sample efficiency due to increased weight sharing, as well as improvements in robustness to symmetry transformations. The group equivariant CNN framework is extended for segmentation by introducing a new equivariant $(G\rightarrow\mathbb{Z}^2)$-convolution that transforms feature maps on a group to planar feature maps. Also, equivariant transposed convolution is formulated for up-sampling in an encoder-decoder network. To demonstrate improvements in sample efficiency we evaluate on multiple data regimes of a rotation-equivariant segmentation task: cancer metastases detection in histopathology images. We further show the effectiveness of exploiting more symmetries by varying the size of the group.' --- Introduction ============ Convolutional networks are able to model local patterns in data efficiently by sharing parameters in the convolutional layers. In doing so, the translational symmetry of images is preserved: shifting a layer’s input produces a proportionate shift $T$ in the layer’s output, i.e. $f(T\mathbf{x})=Tf(\mathbf{x})$. This property known as translation equivariance exploits the spatial structure inherent to many sensory data like images, video and audio. It has been shown that the effectiveness of exploiting symmetries extends beyond mere translations. Recent work by [@Cohen2016-do] on group equivariant convolutions has explored a higher degree of weight sharing to encode discrete rotation and reflection symmetries. The resulting rotation equivariant CNN exhibits an increased expressive power and in turn a higher sample efficiency, without an increase in the number of parameters. These networks have shown considerable gains in classification performance, but have not been extended to function in a segmentation setting. {width="\textwidth"} We leverage the group convolutions of [@Cohen2016-do] to encode rotation and reflection symmetries in a pixel-wise segmentation model. To enable segmentation for this framework, we extend it with a convolution operation that maps from a group representation to the $\mathbb{Z}^2$ grid and we use group convolution in a transposed convolution layer, which we show to be equivariant. Domains that benefit most from exploiting rotation symmetries are those that lack a canonical orientation, such as medical imaging and in particular histopathology data. This domain is especially well-suited for the use of equivariant convolutional networks; the limited availability of pixel-level annotations requires a high sample efficiency. As such, to demonstrate the potential of the proposed model we evaluate on a histopathology dataset derived from the Camelyon16 challenge [@bejnordi2017diagnostic]. We show that the increased weight sharing by explicitly encoding rotation and reflection symmetries leads to consistent performance gains, especially under limited training dataset size. We further establish that conventional CNNs trained on histopathology data demonstrate erratic predictions under $\pi/2$ rotations and reflections. Such behavior is alleviated by the proposed group equivariant model which helps accommodate the requirements of model predictability in a clinical setting. Related work {#related-work .unnumbered} ------------ A common approach to improve robustness to rotations and reflections for CNNs is to use extensive data augmentation during training [@Liu2017-jq; @litjens2017survey]. Although this potentially leads to better generalization, it also increases the amount of training samples. Furthermore, the additional rotated copies of training data induce the model to learn permuted copies of the same filter, adding redundancy to the network’s weights and increasing the risk of overfitting. Our proposed model reduces these effects with built-in equivariance, and hence could exhibit an improved sample efficiency. Additionally even if rotation equivariance is achieved on the training data, there is no guarantee that this generalizes to a test set. To approximate test time equivariance, [@Liu2017-jq; @Ciresan2013-wv] propose a test-time augmentation strategy that averages the predictions of $90\degree$-rotated and mirrored versions. This however increases the computational cost of inference by a factor of eight and does not guarantee equivariance [@DBLP:journals/corr/LencV14]. To achieve equivariance, we focus on the straight-forward G-CNN framework from [@Cohen2016-do] applied on discrete rotation and reflection groups. Work by [@Worrall2017-ji] further exploits rotational symmetries allowing for full $360\degree$-equivariance, such Harmonic Networks constrain the set of filters to circular harmonics. [@Weiler2017-oz] employs steerable filters and uniformly samples a small number of rotations. Although the groups in the G-CNN framework from [@Cohen2016-do] do not cover the full continuous rotational symmetries, the empirical evidence gathered so far shows that $90\degree$ rotation equivariance improves performance significantly [@Weiler2017-oz; @veeling2018]. Similar work by [@veeling2018] leverages the G-CNN framework on histopathology data in a classification setting, demonstrating consistent performance gains and improvements in sample efficiency when comparing an equivariant model with an equivalent standard CNN. Methodology =========== {width="\textwidth"} Mathematical framework ---------------------- Here, we further develop the mathematical framework of G-CNNs [@Cohen2016-do] such that it can be used in a segmentation setting, by formalizing transposed group convolutions and introducing a new convolutional operation transforming group feature maps to planar feature maps. The implementation of the equivariant layers is available on: `https://github.com/nom/gcnn-seg`. G-CNNs utilize group convolutions, which enjoy increased weight sharing and better statistical efficiency than regular convolutions. Specifically, it is implemented for the $p4$ group, consisting of translations and rotations by multiples of $\pi/2$, and the $p4m$ group, which additionally includes reflections. G-CNNs are generalizations of CNNs where feature maps are considered as functions on these groups, e.g. for the $p4m$ group each feature map contains eight orientation channels, corresponding to the number of roto-reflections in the group. In Fig. \[equiv\] we visualize such G-CNN feature maps and demonstrate how equivariance is preserved throughout the different group convolutions. The first layer transforms an input image $f : \mathbb{Z}^2 \rightarrow \mathbb{R}^K$, with $K$ the number of channels using a filter $\psi$, which is defined as the ($\mathbb{Z}^2 \rightarrow G$)-convolution: $$[f * \psi](g) = \sum_{y \in \mathbb{Z}^2} \sum_{k=1}^K f_k(y) \psi_k(g^{-1} y),$$ where $g = (r, t)$ is a roto-translation (in case $G=p4$) or a roto-reflection-translation (in case $G=p4m$). In the next layers, the feature maps are functions on $G$, hence the filters must also be functions on $G$, for which the $(G\rightarrow G)$-convolution is used: $$[f * \psi](g) = \sum_{h\in G}\sum^K_{k=1} f_k(h)\psi_k(g^{-1} h).$$ The transpose of this linear operation, the *transposed* convolution, is used to increase the spatial size of the feature maps. Contrary to strided group convolution [@Cohen2016-do], the choice of stride for transposed group convolution does not affect equivariance. To allow for the equivariant transformation of a feature map in $G$ to a two-dimensional segmentation mask $m : \mathbb{Z}^2 \rightarrow \mathbb{R}^C$, with $C$ the number of classes, we propose the definition of a $(G \rightarrow \mathbb{Z}^2 )$-convolution: $$[f * \psi](y) = \sum_{h\in G}\sum^K_{k=1} f_k(h)\psi_k(z(y)^{-1}h), \label{eq:gtoz2}$$ where $z(y)$ is the translation by $y$ in $\mathbb{Z}^2$. Contrary to the $(G\rightarrow G)$-convolution, the filter $\psi$ in Eq. \[eq:gtoz2\] is a function on $\mathbb{Z}^2$. This single planar filter is shared across the orientation channels following the transformations of the group, similar to the ($\mathbb{Z}^2 \rightarrow G$)-convolution (see Fig. \[equiv\]). This introduces a learnable transformation from a feature map in $G$ to a segmentation mask which is naturally more expressive then pooling the orientation channels [@Cohen2016-do; @Weiler2017-oz]. GU-Net architecture ------------------- To obtain pixel-wise segmentation maps, we use the conventional U-Net architecture [@ronneberger2015u] as a baseline for our rotation equivariant model. The *GU-Net* architecture is constructed by replacing all (transposed) convolution layers with their group equivariant counterparts. Two-layer convolution blocks are followed by a $3 \times 3$ max-pool that incrementally reduces the spatial size, up to a depth level of four. Then pooling is replaced by $3 \times 3$ transposed convolutions with zero padding to recover the spatial size of the input image and enable per-pixel classification. Batch normalization is applied after every convolution operation, including transposed convolutions. For the *GU-Net* version, these batch normalization operations are made group equivariant by aggregating moments per group feature map rather than spatial feature map. The proposed $(G \rightarrow \mathbb{Z}^2 )$-convolution is finally used to transform the orientation channels to the two-dimensional grid of the output mask. To ensure a fair comparison between the proposed equivariant model and the baseline model, the number of parameters is kept constant [@Cohen2016-do]. To this end the amount of filters used by the G-CNN model is divided by the square root of the group size. This way the additional parameters introduced by an increase in orientation channels, increasing each layers input size, are compensated for. Experiments =========== Dataset ------- The proposed model is evaluated on the PatchCamelyon dataset [@veeling2018] derived from the Camelyon16 challenge [@bejnordi2017diagnostic] with the task of tumor localization. The original challenge data contains 400 H&E stained whole slide images (WSIs) of sentinel lymph node sections with pixel-level annotations. The PatchCamelyon dataset consists of 300.000 patches of $320\times320$ pixels at 10$\times$ magnification, with a 8:1:1 split into training, validation and test set. To prevent selecting background and non-tissue patches, patches are converted to HSV, blurred and selected if max pixel saturation lies above 0.07 (range $[0,1]$), and value above 0.1. This was empirically verified to not drop tissue patches. The patches were extracted by uniformly sampling WSIs and drawing tumor/non-tumor patches with equal probability. Training details ---------------- Models are optimized using Adam [@KingmaB14] with initial learning rate $1\mathrm{e}{-3}$ (halved after 20 epochs of no improvement in validation loss). Epochs consist of 468 batches with a batch size of 64. Weights with lowest validation loss are selected for test evaluation. Model stability --------------- To assess the predictive stability of the P4M U-Net model, we examine the standard deviation of predictive probabilities under roto-reflection transformations of the input as compared to the baseline model. Fig. \[model\_stability\] shows the analysis for an example patch. We observe that the standard U-Net is prone to erratic prediction behavior under roto-reflections, especially for uncommon patterns that occur only in a limited set of orientations in the training data such as tumor boundaries. Segmentation performance ------------------------ The segmentation performance of the proposed model is evaluated using the Dice similarity coefficient (DSC), Table \[data-regimes\] reports the results. To compare with a data augmentation strategy, the training data is augmented with roto-reflection transformations for the baseline U-Net. To compare the sample efficiency, we experimented with multiple data regimes, where the number of samples in the training set is incrementally reduced by a factor of two. Model Data DSC ------- ---------- ------ $100\%$ 83.7 $50\%$ 82.1 $25\%$ 81.6 $12.5\%$ 79.6 $100\%$ 79.8 $50\%$ 78.5 $25\%$ 77.6 $12.5\%$ 74.2 : Performance on the Camelyon16 derived dataset. \[data-regimes\] These results show that the proposed model performs consistently better than the baseline method in terms of the DSC metric, and when trained on the 25% data regime it even outperforms the baseline model trained on the entire dataset. We see that the superiority of our proposed architecture is predominantly due to the increased weight sharing in the $p4m$-equivariant model, which frees up model capacity and reduces the redundancy of detecting the same local patterns in different orientations. We also observe that the performance gap between our model and the baseline increases when we limit the dataset size by removing training samples. For example, our method outperforms the baseline on 1/8th of the data by 7.2%. This seems to indicate that the performance in the small-data regime benefits significantly from the sample efficiency of P4M U-Net, with diminishing returns when the amount of data is sufficient for the baseline network to achieve (approximate) rotation equivariance. This performance gap remains for the full data set. We further vary the group size to only roto-transformations to study whether adding symmetries contributes to a better performance. For the $p4$ group, we measure a Dice coefficient on the full test set of 82.5 compared to 83.7 for the $p4m$ group, which indicates that additionally exploiting the reflection symmetry corresponds to better generalization. Conclusion ========== By extending the group equivariant CNN framework and using it in an encoder-decoder architecture for pixel-wise predictions, we show a consistent performance increase compared to an equivalent conventional CNN. We further demonstrate that the performance gap between the G-CNN and the baseline model is most significant for smaller data regimes, highlighting a substantial enhancement of sample efficiency by G-CNNs. We also confirm experimentally, by increasing the size of the group, that exploiting more symmetries leads to better generalization. Notably, we further demonstrate that conventional CNNs show erratic predictions for rotations and reflections which is alleviated by the use of group convolutions.

ArXiv

--- abstract: 'We calculate the screening of the ion-ion potential due to electrons in the presence of a large background magnetic field, at densities of relevance to neutron star crusts. Using the standard approach to incorporate electron screening through the one-loop polarization function, we show that the magnetic field produces important corrections both at short and long distances. In extreme fields, realized in highly magnetized neutron stars called magnetars, electrons occupy only the lowest Landau levels in the relatively low density region of the crust. Here our results show that the screening length for Coulomb interactions between ions can be smaller than the inter-ion spacing. More interestingly, we find that the screening is anisotropic and the screened potential between two static charges exhibits long range Friedel oscillations parallel to the magnetic field. This long-range oscillatory behavior is likely to affect the lattice structure of ions, and can possibly create rod-like structures in the magnetar crusts. We also calculate the imaginary part of the electron polarization function which determines the spectrum of electron-hole excitations and plays a role in damping lattice phonon excitations. We demonstrate that even for modest magnetic fields this damping is highly anisotropic and will likely lead to anisotropic phonon heat transport in the outer neutron star crust.' author: - 'Rishi Sharma, Sanjay Reddy' bibliography: - 'screen.bib' title: Electronic screening and damping in magnetars --- Introduction ============ Highly magnetized neutron stars called magnetars that feature extreme magnetic fields ($B$) — as large as $10^{15}$G on the surface and perhaps even larger fields inside — have been detected in recent years. Currently there are 13 magnetar candidates which are classified as anomalous x-ray pulsars (AXPs) or soft gamma repeaters (SGRs) (http://www.physics.mcgill.ca/ pulsar/magnetar/main.html). Such large magnetic fields can strongly influence the state of matter in neutron stars, especially in the lower density region were the characteristic energy scales associated with the matter fields is small compared to the magnetic “perturbation”. Several decades ago, it was realized that even modest magnetic fields could influence the structure of atoms in the atmospheres of magnetized neutron stars and affect the spectral features of the emitted radiation [@Ruderman:1974] . At the surface, where electrons are localized around nuclei and the system consists of neutral atoms, large magnetic fields distort the atomic structure resulting in rod-like or cigar shapes elongated parallel to $B$. (For a recent review see [@Ruder:1994].) It was realized quite early that these cigar shaped atoms would bind with each other along the magnetic field direction forming polymer chains, and that these chains would interact in the perpendicular plane, resulting in a condensed phase [@Ruderman:1971]. More recently the physics of these condensed phases for hydrogen has been analyzed and its implications for neutron atmospheres is discussed in Ref. [@Lai:1997]. Here we will consider the effects of the magnetic fields at higher density, in the crusts of magnetars where the electrons are not localized in atomic states but rather form a degenerate Fermi system. This will be important in determining the structural and transport properties of the magnetar crust. Our finding will have implication for crust oscillations (excited in explosive phenomena such as giant flares) and for heat and electrical transport which are important ingredients in the study of magnetic and thermal evolution of magnetars [@Geppert:2004; @Aguilera:2007]. In this work we will typically restrict the study to strong fields where only a few Landau levels are occupied and this will necessarily restrict us to relatively low density region in the outer crust of the neutron star. ($\rho {\, \, \raisebox{-0.8ex}{$\stackrel{\textstyle <}{\sim}$ }}10^{10}$ g/cm$^3$ for $B$ ${\, \, \raisebox{-0.8ex}{$\stackrel{\textstyle <}{\sim}$ }}$ 10$^{15}$ G.) The key microscopic quantity required to characterize the response in our approach is the polarization function, which is the Fourier transform of the density-density correlation function. The expression for the polarization function for a magnetized electron gas is calculated and is found in agreement with previous work [@Danielsson:1995]. Our new findings in this work are related to identifying and understanding the implications of the electronic response in strong magnetic fields for neutron star structure and transport. The real part of the electron polarization function is related to the screening of the ion-ion potential in the crust. In an unmagnetized crust, screening is dominated by the usual Debye screening. In neutron stars where the typical magnetic field is small ($B$ ${\, \, \raisebox{-0.8ex}{$\stackrel{\textstyle <}{\sim}$ }}10^{11}$ G), the electron screening length is larger than the inter-ion distance and this implies that the effects of electron screening are usually unimportant in determining the crystal structure and thermodynamics. In contrast, we find that for large fields when only one or a few Landau levels are filled, the screening length can be either much smaller or much larger than in the unmagnetized case depending on the density and magnetic field. At relatively low density the screening is enhanced resulting in a screening length that is generically smaller than the inter-ion distance. Our finding of enhanced screening for large $B$ is not new, this was already noted in early work which explored some its implications for nuclear fusion reactions [@Shalybkov:1987]. The qualitatively new finding of our work is the realization that the sharp Fermi surface in the electron momentum distribution along the $B$ field will give rise to long-range oscillatory behavior called Friedel oscillations in the ion-ion potential — a well studied feature in the condensed matter context for $B=0$. The period of this spatial oscillation $\lambda_F = \pi/(k_{f}^z)$, where $k_{f}^z$ the Fermi momentum of the electrons in the $z$-direction, where we have assumed that $B$ is parallel to $z$. This potential falls off slowly as $1/z$ on a scale of several inter-ion distances and consequently can be very important in determining the phase structure of the crust. We speculate that this will lead to the formation of an anisotropic crystal with rod-like structures akin to the polymer chains expected in the atomic regime. The imaginary part of the polarization is related to dissipation and gives rise to the usual Landau damping observed in Fermi systems. In the crust of a neutron star, excitations such as lattice phonons couple to the electrons, and their mean free paths are largely determined by the efficiency of the electronic Landau damping where the phonon decays by producing an electron-hole pair. We show that this decay rate has a non-trivial dependence on the angle between their propagation direction and the magnetic field because the electron-hole excitation for large $B$ is highly anisotropic. This anisotropy leads to an anisotropy in heat transport properties of lattice phonons. We find that the lattice phonons can be scattered more easily if they are moving perpendicular to the field, compared to when they are moving parallel to it. The paper is organized as follows. In Section \[Section:PolarizationMagnetic\] we derive the polarization function of electrons in the presence of a magnetic field. In the Sections \[Section:Screening\] and \[Section:HeatTransport\], we apply the result to two physical problems. In Section \[Section:Screening\] we present numerical results for Friedel oscillations. In Section \[Section:HeatTransport\] we discuss the effect of anisotropic polarization on the decay rates of the lattice phonons as a function of the angle between the propagation direction and the magnetic field. We discuss the implications for the structure of the magnetar crust in Section \[Section:Conclusions\]. Details of calculations of the expression for the screened potential are given in Appendix \[Section:LLL dominates\] and Appendix \[Section:Friedel derivation\]. Polarization of electrons in a magnetic field.\[Section:PolarizationMagnetic\] ============================================================================== The Lagrangian for the electron gas in an external electromagnetic field and at finite chemical potential $\mu$ is given by, $$L = \int d^3x \bar{\psi}(i{D\!\!\!\!/}-m_e + \mu\gamma^0)\psi\;,$$ where $D_\mu = \partial_\mu - ie A_\mu$ and $m_e$ is the electron mass. We consider a time independent, uniform magnetic field $B$ in the ${z}$ direction and choose a gauge such that the only non-zero component of the gauge field $A$ is $A^x = -A_x = By$. The spectrum of electron energy levels in an external magnetic field is well known and electrons occupy Landau levels with energies given by $$E_{m} = \sqrt{(k^z_{m})^2+2meB+m_e^2}\,,$$ where $m$ is the quantum number associated with the Landau level and $k^z_m$ is the momentum in the $z$ direction. The number density of electrons is $$n_e =\frac{eB}{(2\pi)^2}(\int_{-\infty}^{\infty} dk^z f_0 + 2\sum_{m>0}\int_{-\infty}^{\infty} dk^z f_m) \label{eq:ne}$$ where $f_{m} = (\exp(\frac{E_{m}-\mu}{T})+1)^{-1}$ is the Fermi distribution function. At $T=0$, we obtain $$n_e =\frac{eB}{2\pi^2}(\sqrt{\mu^2-m_e^2} + 2\sum_{m>0}\theta(\mu^2-m_e^2-2meB)\sqrt{\mu^2-m_e^2-2meB})\;. \label{eq:neT0}$$ We will often restrict to the low-density or large magnetic field limit, where only the lowest Landau level is occupied. This occurs for $2eB>\mu^2-m_e^2$. The density-density correlation function is defined as $$\chi_B(x^\mu,y^\mu) = \frac{1}{i}\langle T\{n(x)n(y)\}\rangle\label{eq:chiB}\,,$$ where $n(x)= \psi^\dagger(x) \psi(x)$ is the density operator. The Lagrangian is time independent for a static external field and even though the Lagrangian looks position dependent in a particular gauge, the system itself is translationally invariant. (The change in the Hamiltonian due to a translation in ${x}$ can be undone by a gauge transformation.) Therefore we can write the correlation function in momentum space as, $$\chi_B(x^\mu,y^\mu) = \int \frac{d\omega}{2\pi} \frac{1}{V} \sum_q e^{iq_\mu (x-y)^\mu} \Pi_B(\omega,{\bf{q}})\label{eq:piB}\;,$$ where $q^\mu=(\omega,{{\bf{q}}})=(\omega,{q_{\perp}},{q^{z}})$. To evaluate $\Pi_B$, we insert a complete set of many-body eigenstates between the two density operators in Eq. \[eq:chiB\]. We ignore the anti-particle contribution to the polarization function because they are well below the Fermi level and retain only particle-hole contributions. In this approximation we can write the polarization as, $$\begin{split} \Pi_B(q^\mu) =& \frac{eB}{2\pi} \sum_{m,n} \int_{-\infty}^\infty \frac{dk^z_m}{2\pi}f_m (1-f_n) W(E_m,E_n,m,n,k^z_m,k^z_n,{q_{\perp}})\\ &\bigl[\frac{1}{\omega-E_n+E_m+i\epsilon} -\frac{1}{\omega+E_n-E_m-i\epsilon} \bigr]|_{k^z_n=k^z_m+{q^{z}}}\label{eq:PiB simplified}\;, \end{split}$$ where, $m$ refers to the (initial) energy eigenstate with energy $E_{m} = \sqrt{(k^z_{m})^2+2meB+m_e^2}$ and $n$ to the (final) state with energy $E_{n} = \sqrt{(k^z_{m}+q^z)^2+2neB+m_e^2}$. The function $W$ contains the matrix elements of the density operator between different electron eigenstates and is given by , $$\begin{split} W = W(E_m,E_n,m,n,k^z_m,k^z_n,{q_{\perp}}) =& \frac{1}{2E_nE_m}\bigl[(E_mE_n+k^z_mk^z_n+m_e^2)H_{m-1,n-1}({q_{\perp}})H^*_{m-1,n-1}({q_{\perp}})\\ &+(\sqrt{2meB}\sqrt{2neB})H_{m-1,n-1}({q_{\perp}})H^*_{m,n}({q_{\perp}})\\ &+(\sqrt{2meB}\sqrt{2neB})H_{m,n}({q_{\perp}})H^*_{m-1,n-1}({q_{\perp}})\\ &+(E_mE_n+k^z_mk^z_n+m_e^2)H_{m,n}({q_{\perp}})H^*_{m,n}({q_{\perp}}) \bigr]\label{eq:W defined}\;, \end{split}$$ where, $$H({q_{\perp}})=\left\{\begin{array}{cc} m\geq n &\bigl[ e^{-{q_{\perp}}^2l^2/4}\sqrt{\frac{n!}{m!}} \Bigl(\frac{(-q^x-iq^y)l}{\sqrt{2}}\Bigr)^{m-n} {{L}}_n^{m-n}({q_{\perp}}^2l^2/2)\bigr]\\ m\leq n &\bigl[ e^{-{q_{\perp}}^2l^2/4}\sqrt{\frac{m!}{n!}} \Bigl(\frac{(q^x+iq^y)l}{\sqrt{2}}\Bigr)^{n-m} {{L}}_m^{n-m}({q_{\perp}}^2l^2/2) \bigr] \end{array}\right.\label{eq:H defined}\;\;.$$ Here, ${{L}}_a^b(x)$ are the associated Laguerre polynomials, and $l=1/\sqrt{eB}$ is the magnetic length. For $m=0$ or $n=0$, $H_{m-1,n-1}(X)=0$. The imaginary and real parts of the response function can be separated by following the pole prescription, $$\frac{1}{\omega - \Delta E \pm i \epsilon} \ = {\rm{P}}\bigl(\frac{1}{\omega - \Delta E }\bigr) \mp i\pi \delta(\omega - \Delta E )\;.$$ This gives the imaginary part, $$\begin{split} \Im m[\Pi_B(q^\mu)] =& \frac{eB}{2\pi} (-\pi)\sum_{m,n} \int_{-\infty}^\infty \frac{dk^z_m}{2\pi} f_m (1-f_n)W(E_m,E_n,m,n,k^z_m,k^z_n,{q_{\perp}})\\ & \bigl[\delta({\omega-E_n+E_m}) -\delta({\omega+E_n-E_m}) \bigr]|_{k^z_n=k^z_m+{q^{z}}}\label{eq:ImPi}\;, \end{split}$$ We want to consider situations where the temperature is much less than the chemical potential $\mu$. In this case, it is appropriate to approximate $f_m$ by $\theta(\mu-E_m)$ and $1-f_n$ by $\theta(E_n-\mu)$. The real part of $\Pi_B$ is given by, $$\begin{split} \Re e[\Pi_B(q^\mu)] =& \frac{eB}{2\pi} \sum_{m,n}\int_{-\infty}^\infty \frac{dk^z_m}{2\pi} f_m W(E_m,E_n,m,n,k^z_m,k^z_n,{q_{\perp}})\\ &{\rm{P}}\bigl[\frac{1}{\omega-E_n+E_m} -\frac{1}{\omega+E_n-E_m} \bigr]|_{k^z_n=k^z_m+{q^{z}}}\label{eq:RePi}\;, \end{split}$$ In particular, to calculate screening, we will be interested in the static response, $\omega=0$, which is given by, $$\begin{split} \Re e[\Pi_B(0,{{\bf{q}}})] = & \frac{eB}{\pi} \sum_{m,n}\int_{-\infty}^\infty \frac{dk^z_m}{2\pi} f_m W(E_m,E_n,m,n,k^z_m,k^z_n,{q_{\perp}})\\ &{\rm{P}}\bigl[\frac{1}{E_m-E_n} \bigr]|_{k^z_n=k^z_m+{q^{z}}}\label{eq:RePiw0}\;. \end{split}$$ In the limit $T=0$, $$\begin{split} \Re e[\Pi_B(0,{{\bf{q}}})] =& \frac{eB}{2\pi} 2 \sum_{m,n}\int_{-\infty}^\infty \frac{dk^z_m}{2\pi} W(E_m,E_n,m,n,k^z_m,k^z_n,{q_{\perp}})\\ &\theta(\mu-\sqrt{(k^z)^2+m_e^2+2meB}){\rm{P}}\bigl[\frac{1}{E_m-E_n} \bigr]|_{k^z_n=k^z_m+{q^{z}}}\label{eq:RePiw0T0}\;. \end{split}$$ We will now use these results for calculating two important physical processes of relevance to magnetars: (i) the screening of electric charge by the electron gas and the resulting modification of the ion-ion potential in the crust (Section \[Section:Screening\]), and (ii) the anisotropic spectrum of electron-hole excitations (Landau damping) and its consequences for the phonon mean free paths in the outer crust (Section \[Section:HeatTransport\]). Screening in an electron gas\[Section:Screening\] ================================================= The screened potential between two ions with charges $Z_1$ and $Z_2$ separated by a displacement ${{\bf{r}}}$ in an electron gas is given by $$\begin{aligned} V_{\rm mod}({{\bf{r}}}) &=& \frac{Z_1 Z_2~e^2}{4 \pi r}g({{\bf{r}}})\;, \nonumber \\ {\rm where} \quad g({{\bf{r}}}) &=& {4\pi r}\int\;\frac{d^3q}{(2\pi)^3} \;\frac{e^{i{{\bf{q}}}\cdot{{\bf{r}}}}}{q^2+F({{\bf{q}}})} \label{eq:vmod}\end{aligned}$$ gives the modification of the bare interaction due to the polarizability of the electron gas. The function $F({{\bf{q}}})=-e^2\Re e[\Pi_{00}(\omega=0,{{\bf{q}}})]$ is related to the static polarization function of the electron gas, $\Pi_{00}(0,{{\bf{q}}})$. In evaluating the integral in Eq. \[eq:vmod\], typically it is assumed that for distances $r {\, \raisebox{-0.8ex}{$\stackrel{\textstyle >}{\sim}$ }}1/k_f$, where $k_f$ is the Fermi momentum of the electrons, we can replace $\Pi_{00}(0,{{\bf{q}}})$ by $\Pi_{00}(0,{\bf{0}})$. In doing so one obtains the usual Debye screened potential in coordinate space $$V_{\rm mod}(r) = \frac{Z_1 Z_2~e^2}{4 \pi r}~ \exp(-r/\lambda_D)\,,$$ or $g(r) = \exp(-r/\lambda_D)$ where $ \lambda_D=1/\sqrt{F({\bf{0}})}$ is the Debye screening length, and is simply related to the Debye mass $m_D$, by $\lambda_D=1/m_D$. F([**[0]{}**]{}) can be calculated using the well known compressibility sum-rule [@FetterWalecka:1971] and is given by $$F({\bf{0}}) = e^2~\frac{\partial n_e}{\partial \mu_e}\,, \label{eq:fzero}$$ where $n_e$ is the equilibrium density and $\mu_e$ is the chemical potential and $e^2=4\pi/137$. For a free electron gas $F({\bf{0}}) = e^2k_f \mu/\pi^2$ with $\mu=(k_f^2+m_e^2)^{1/2}$. The ion density $n_I=3/(4\pi a^3)$ — where $a$ is the average inter-particle separation (for a simple cubic crystal $a$ is related to the lattice spacing $L$ by $(4\pi/3)^{1/3}a=L$) — and the electron density are related by electric charge neutrality which requires that $n_e=Zn_I$, where $Z$ is the charge of each ion. Consequently, the electron Fermi momentum $k_f$ and the average inter-ion distance $a$ are related by $k_f=(3\pi^2 Z n_I)^{1/3} =5.66~(Z_{26})^{1/3}~a^{-1}$ where $Z_{26}=Z/26$. Further, from Eq. \[eq:fzero\] we have $\lambda_D= 10.37\sqrt{v_f}/k_f$ where $v_f=k_f/\mu$. The ratio of the screening length to the average inter-ion distance is therefore given by $$\frac{\lambda_D}{a} \simeq \frac{1.83}{(Z_{26})^{1/3}}\sqrt{v_f}\label{eq:lDvsa} \,,$$ implying that in the relativistic regime, where $k_f \gg m_e$ and $v_f \simeq 1$, $\lambda_D\simeq 1.83~a$. Since the interaction between the nearest ionic neighbors is essentially unscreened in the relativistic electron gas, screening is not considered to be important for determining the structure of the neutron star crust. This situation is changed if the electrons are non-relativistic. For $Z_{26}=1$, $\lambda_D<a$ for $v_f<0.3$. However at lower densities, where $v_f \ll 1$, orbital effects from the localization of electrons may be important and a simple screening approximation is not valid. ![(color online) The screening length $\lambda_D$ versus the inter-particle separation $a$ for $Z=26$ and $eB=10eB_c$. The dotted line corresponds to $\lambda_D=a$. For $a\gtrsim512$(fm), the lowest Landau level is occupied, and for large regions $\lambda_D<a$ . As $a$ decreases, more levels are occupied.[]{data-label="fig:lambdavsa"}](lambdavsa_eB2.62.eps){width="4in"} Extending the Debye approximation to an electron gas in a magnetic field, setting $F_B({{\bf{q}}})$ to its long-wavelength limit $F_B({{\bf{q}}}={\bf{0}})$, we can deduce how the Debye screening is modified by the magnetic field. When only one Landau level is occupied, we find $$\begin{aligned} \lambda_D(2eB>{\mu_e^2-m_e^2}) &=& \frac{\pi \sqrt{2 }}{e\sqrt{eB}}~\sqrt{v_{f}^z} \label{eq:lDvsaB}\,, \end{aligned}$$ which can be smaller or larger than $a$. Hence the screening length between ions relative to their separation can tuned arbitrarily by introducing a large magnetic field. To illustrate this we plot $\lambda_D$ versus $a$ for $eB=10eB_c$ in Fig. \[fig:lambdavsa\]. In the low density regime where only one Landau level is filled, $\lambda_D < a $ for $a>1164$(fm). With increasing density (smaller $a$), $\lambda_D$ becomes larger than $a$ and continues to oscillate around the $B=0$ value as more Landau levels are occupied. The amplitude of these oscillations rapidly decreases with $n$ where $n$ is the number of Landau levels filled. As noted earlier, for $B=0$ the screening length is always larger than $a$ when electrons are relativistic. At low temperatures the Fermi surface of the electrons is sharp and this gives rise to non-analyticities in $F({{\bf{q}}})$. Therefore it is not possible to approximate $F({{\bf{q}}})$ in Eq. \[eq:vmod\] by $F({\bf{0}})$ even for $r\gg 1/k_f$. The screened potential in position space depends on the polarization function for values of ${{\bf{q}}}$ other than ${{\bf{q}}}={\bf{0}}$, and has qualitative new features compared to the Debye screened result. This phenomenon is well known for the free electron gas ($B=0$) [@FetterWalecka:1971]. Kapusta and Toimela [@KapustaToimela:1988] employed the fully relativistic polarization function for electrons at one-loop order to calculate the most general form of the screened potential in a free electron gas. The central result of their calculation is that $F({{\bf{q}}})$ has branch cuts along $q=|{{\bf{q}}}|=\pm2k_f+i\eta$ for $\eta>0$. Consequently at large distances, where $r \gg \lambda_D$ the screened potential exhibits characteristic oscillations called Friedel oscillations. It is well established [@FetterWalecka:1971] that in the non-relativistic limit the screening function is given by $$g_{\rm Friedel(NR)}(r\gg \lambda_D)= \frac{4\xi^2}{(4+\xi)^2}\frac{\cos{(2 k_f r)}}{( r/\lambda_D)^2} \,,$$ where $\xi=m_{D}^2/2 k_f^2= (e^2 \mu)/(\pi^2 k_f)$. At very low density, where $\xi {\, \raisebox{-0.8ex}{$\stackrel{\textstyle >}{\sim}$ }}1$ Friedel oscillations are non-negligible and have been well studied in condensed matter physics context [@FetterWalecka:1971]. The potential between two ions for $r\lesssim \lambda_D$ is dominated by Debye screening. In the region $r \simeq \lambda_D$ a simple analytic form does not exist and the screening function needs to be calculated numerically. It is also shown in Ref. [@KapustaToimela:1988] that in the extreme relativistic limit, when $m_e\rightarrow 0$, Friedel oscillations are modified. At large distances the screening function is of the form $$g_{\rm Friedel(R)}= (2\xi)^{3/2}\frac{\sin{(2 k_f r)}}{( r/\lambda_D)^3} \,.$$ Unlike in the non-relativistic regime, here $ \xi = 0.0047 \ll 1$, and amplitude of these oscillations is greatly suppressed. Furthermore they decay as $1/r^3$ instead of $1/r^2$ in the non-relativistic limit. Consequently, Friedel oscillations can be ignored in the relativistic regime. In the presence of a magnetic field, the screening function $F_B({{\bf{q}}})$ is no longer rotationally symmetric and we expect the screened potential to be anisotropic. We now proceed to calculate $V({{\bf{r}}})$ in this case. The calculation of the polarization tensor when the densities are high and several Landau levels are occupied, is technically complicated. On the other hand, in this limit one expects the results to be very similar to the free electron gas. Therefore we restrict ourselves to the large field limit, when only the lowest Landau level is filled. In cylindrical variables we write $F_B({{\bf{q}}})$ as a function of ${q^{z}}=q \cos{\theta}$ and ${q_{\perp}}=q \sin{\theta}$, where $\theta$ is the angle between the ${\bf{q}}$ and the magnetic field ${\bf{B}}$. Note that Debye screening — which depends only on $F_B({\bf{0}})$ — is completely isotropic. In cylindrical spatial coordinates the potential between two ions $$\begin{aligned} V_{\rm mod}(\rho,z) &=& \frac{Z_1Z_2e^2}{4\pi r}g(\rho,z) \;, \nonumber \\ {\rm where} \quad g(\rho,z) &=& \frac{r}{\pi}\int_0^\infty d{q_{\perp}}{q_{\perp}}J_0({q_{\perp}}\rho)\int_{-\infty}^\infty d{q^{z}}\frac{e^{i{q^{z}}z}}{q^2+F_B({q_{\perp}},{q^{z}})} \,, \label{eq:vmodmag}\end{aligned}$$ with $r=\sqrt{\rho^2 + z^2}$, $F_B(q_\perp,q^z)=-e^2 \Pi_B(0,{\bf{q}})$, and $\Pi_B(0,{\bf{q}})$ is the static polarizability of the electron gas in the presence of a magnetic field. Restricting to $m=0$ in the sum in Eq. \[eq:RePiw0T0\], we have, $$\begin{split} F_{B}(q_\perp,q^z) &= e^2\frac{eB}{2\pi}\sum_{n}\int_{-\infty}^\infty\frac{dk^z}{2\pi} 2\theta(\mu-\sqrt{(k^z)^2+m_e^2})\frac{W(E_0,E_n,m,n,k^z_0,k^z_n,q^{z})}{E_n-E_0}\,. \\ \label{eq:Fb} \end{split}$$ The sum over $n$ in Eq. \[eq:Fb\] runs over all non-negative integers, but the most important contribution for $eB\gg\mu^2$ and $eB\gg q^2$ is the $n=0$ term. (This is shown in Appendix \[Section:LLL dominates\].) Therefore for simplicity, we drop the $n>0$ terms, which allows us to calculate analytic expressions for $F_B({{\bf{q}}})$, and the screened Coulomb interaction in certain limits. With these approximations, $$\begin{split} F_{B}({q_{\perp}},{q^{z}}) \sim (\frac{e}{\pi})^2&\bigl(\frac{eB}{2}\bigr)e^{-{q_{\perp}}^2/(2eB)} \int_{-k_f^z}^{k_f^z} d k^z\frac{1}{\sqrt{(k^z+{q^{z}})^2+m_e^2}-\sqrt{(k^z)^2+m_e^2}}\\ &\frac{((k^z+{q^{z}})k^z+m_e^2+\sqrt{(k^z)^2+m_e^2}\sqrt{(k^z+{q^{z}})^2+m_e^2})} {2\sqrt{(k^z)^2+m_e^2}\sqrt{(k^z+{q^{z}})^2+m_e^2}}\,, \end{split}$$ where $k_f^z=\sqrt{\mu^2-m_e^2}$ is the Fermi momentum of the electrons in the $z$ direction for the $m=0$ level. If higher levels are occupied, each will have a different Fermi momentum. Now we consider two limiting cases. First, in the non-relativistic regime where $m_e\gg k_f^z$, we obtain for large $B$, $$\begin{split} F_B({q_{\perp}},{q^{z}})&\sim (\frac{e}{\pi})^2\bigl(\frac{eB}{2}\bigr)e^{-{q_{\perp}}^2/2eB} \int_{-k_f^z}^{k_f^z} d k^z \bigl(\frac{m_e}{(k^z+{q^{z}})^2-(k^z)^2}\bigr)\\ &= \bigl(\frac{e}{\pi}\bigr)^2\bigl(\frac{eB}{2}\bigr)e^{-{q_{\perp}}^2l^2/2} \log\bigl(\frac{2k_f^z+{q^{z}}}{2k_f^z-{q^{z}}}\bigr)\frac{m_e}{{q^{z}}}\label{eq:FB nonrel}\;. \end{split}$$ The Debye mass is given by $$m_D^2 = \lim_{\bf{q}\rightarrow 0} F_B({q_{\perp}},{q^{z}}) = (\frac{e}{\pi})^2(\frac{eB}{2})(\frac{m_e}{k_f^z}) \label{eq:mDnrel}\,,$$ and since $n_e=eB/(2\pi^2)\sqrt{\mu^2-m_e^2}$, the above expression (Eq. \[eq:mDnrel\]) is consistent with Eq. \[eq:fzero\]. The key feature of the expression for $F_B({q_{\perp}},{q^{z}})$ in Eq. \[eq:FB nonrel\] is its non-analytic behavior as a function of $q^z$. In the complex plane, Eq. \[eq:FB nonrel\] has branch cuts along $q^z=\pm 2k_f^z + i\eta$, $\eta>0$. Restricting $q^z$ to the real axis, one sees a kink at $q^z=\pm 2k_f^z$ (Fig. \[fig:FBqz\]). On the other hand the expression is analytic in $q_{\perp}$. Based on this we can expect long range oscillations in $g(z)$ with wavelength $\pi/k_f^z$ in the $z$ direction, but no long range features in the $x-y$ plane. In the relativistic regime, where $m_e\ll\mu$ we find that $$\begin{split} F_{B}({q_{\perp}},{q^{z}}) &\sim (\frac{e}{\pi})^2\bigl(\frac{eB}{2}\bigr)e^{-{q_{\perp}}^2/2eB} \int_{-k_f^z}^{k_f^z} d k^z \frac{(k^z+{q^{z}})k^z+|k^z||k^z+{q^{z}}|} {2|k^z||k^z+{q^{z}}|}\bigl(\frac{1}{-|k^z|+|k^z+{q^{z}}|}\bigr)\\ &= \left\{ \begin{array}{cc} (\frac{e}{\pi})^2 \bigl(\frac{eB}{2}\bigr) e^{-{q_{\perp}}^2/2eB} & {q_{\perp}}< k_f^z \\ (\frac{e}{\pi})^2 \bigl(\frac{eB}{2}\bigr) e^{-{q_{\perp}}^2/2eB}~\frac{k_f^z}{|{q^{z}}|} & {q_{\perp}}> k_f^z \end{array} \right.\;.\label{eq:FB rel} \end{split}$$ The debye mass square is $m_D^2 = \lim_{\bf{q}\rightarrow 0} F_B({q_{\perp}},{q^{z}}) = (\frac{e}{\pi})^2(\frac{eB}{2})$. The number of electrons in these limits is $n_e=eB/(2\pi^2)\sqrt{\mu^2-m_e^2}\sim (eB/(2\pi^2))\mu$ and therefore $m_D^2=e^2 dn_e/d\mu_e=e^2 (eB/(2\pi^2))$ as expected. ![(color online) $F_B$ as a function of $q^z$ for $q_\perp=0$, $eB=100eB_c$, and three values of $\mu$. As the electrons become more relativistic ($\mu$ increases), the kink at $2k_f^z$ becomes less sharp. The shoulder at $k_f^z$ becomes non analytic in the ultra-relativistic limit.[]{data-label="fig:FBqz"}](FB_eB26.eps){width="5in"} From Eq. \[eq:FB rel\] we note that in the ultra-relativistic limit $F_{B}$ does not have any non-analytic behavior at $q^z=2k_f^z$. On the other hand $F_B$ is non-differentiable at $q^z=k_f^z$. This behavior is different from that of a free electron gas, where the non-analyticity occurs at $\pm 2k_f^z$ in both the relativistic and non-relativistic limits. However, this is an artifact of the ultra-relativistic limit, and for any finite $m_e$, $F_B$ is differentiable at $q^z=k_f^z$. We plot $F_B(q^z)$ for three values of $\mu$ in Fig. \[fig:FBqz\]. We notice that for fixed $\mu$ there are two distinct features in the plot versus $q^z$. First, there is shoulder at $k_f^z$ which becomes sharper as $m_e/\mu$ decreases. Second, there is a kink at $2k_f^z$ which becomes less prominent as $m_e/\mu$ decreases. In the ultra-relativistic limit the kink disappears, and the shoulder becomes a non-differentiable point. However, for any finite $m_e$, the fact that the kink becomes less prominent suggests that the Friedel oscillations will become weaker as $m_e/\mu$ decreases. Hence, we will now focus on the non-relativistic limit. In the non-relativistic limit, we can obtain an analytic expression for $g(\rho, z)$ valid for $z\gg1/k_f^z$. We find $$\begin{split} g(\rho,z) &= g_D(\rho, z) + g_F(\rho, z) \label{eq:g large z}\;, \end{split}$$ where, $$g_D(\rho, z) = e^{-m_D\sqrt{\rho^2+z^2}}$$ is the Debye screening formula which comes from the pole in the integrand in Eq. \[eq:vmodmag\] at $q = i m_D$, and $$\begin{split} g_F(\rho,z) =& -\frac{\sqrt{\rho^2+z^2}}{\pi}\frac{\cos(2k_f^zz)}{z} \frac{m_D^2(\pi/4)\rho}{\sqrt{(2k_f^z)^2+m_D^2(1/2)\ln(4k_f^zz)}}\\ &\times K_1(\rho\sqrt{(2k_f^z)^2+m_D^2(1/2)\ln(4k_f^zz)}) \end{split}$$ arises from the branch cuts at $q^z=\pm 2k_f^z+i\eta$. $K$ is a modified Bessel function of the second kind. The derivation of Eq. $\ref{eq:g large z}$ is given in Appendix \[Section:Friedel derivation\]. For $\rho=0$, we see that $g_F(0,z)$ exhibits long-range Friedel oscillations in the $z$ direction given by $$g_F(0,z) \sim -\cos(2k_f^z z)\;.$$ Eq. \[eq:g large z\] is a good approximation for $z\gtrsim 1/k_f^z$. For $\rho\rightarrow 0 $ and $z\rightarrow 0$, we expect that $g(\rho,z)$ should $\rightarrow 1$ since the short range behavior of $V({\bf{r}})$ can not be modified by screening. For $z\rightarrow 0$ the expression Eq. \[eq:g large z\] breaks down because $(2k_f^z)^2+m_D^2(1/2)\ln(4k_f^zz)<0$ but in our derivation we used the fact that $z\gtrsim 1/k_f^z$ and we can not trust Eq. \[eq:g large z\] for $z<1/k_f^z$ anyway. For $\rho=0$ our derivation remains valid but Eq. \[eq:g large z\] can not be used directly since $K_1(a\rho)$ goes as $1/(a\rho)$ as $\rho\rightarrow 0$. The limit of $g$ as $\rho\rightarrow 0$, however, is finite because $a\rho K_1(a\rho)$ tends to $1$ as $\rho\rightarrow 0$. Therefore we define, $$\begin{split} g(\rho=0,z)&= e^{-m_Dz} -\frac{\cos(2k_f^zz)}{\pi} \frac{m_D^2(\pi/4)}{{(2k_f^z)^2+m_D^2(1/2)\ln(4k_f^zz)}} \label{eq:g rho 0}\;. \end{split}$$ For $z=0$ the value of $g(z)$ is well approximated by the Debye screened value $$\begin{split} g(\rho,z=0)&= e^{-m_D\rho} \label{eq:g z 0}\;. \end{split}$$ ![(color online) Plots of $g$ as a function of $\rho$ for $z=0$, and as a function of $z$, for $\rho=0$ for $eB=100~eB_c$, $\mu=0.7$ MeV. $g$ shows an exponential decay as a function of $\rho$, while along the $z$ axis it shows long range oscillations $\propto -\cos(2 k_f^z z)$. Also shown is the Debye screening length $\lambda_D$ (vertical line, brown online) and $a_I=(3/(4\pi n_I))^{1/3}$ (vertical dot dashed line, orange online, assuming $Z=26$). The solid lines correspond to the analytic approximations given in Eq. \[eq:g rho 0\] ($\rho=0$) and Eq. \[eq:g z 0\] ($z=0$), while the points correspond to a numerical computation of the Fourier transform in Eq. \[eq:vmod\]. The right panel shows the 3-d plot of $g(\rho,z)$ in the $\rho$, $z$ plane with $z$ and $\rho$. []{data-label="fig:g_separate"}](rhozg_separate_eB26v2.eps "fig:"){width="3.5in"} ![(color online) Plots of $g$ as a function of $\rho$ for $z=0$, and as a function of $z$, for $\rho=0$ for $eB=100~eB_c$, $\mu=0.7$ MeV. $g$ shows an exponential decay as a function of $\rho$, while along the $z$ axis it shows long range oscillations $\propto -\cos(2 k_f^z z)$. Also shown is the Debye screening length $\lambda_D$ (vertical line, brown online) and $a_I=(3/(4\pi n_I))^{1/3}$ (vertical dot dashed line, orange online, assuming $Z=26$). The solid lines correspond to the analytic approximations given in Eq. \[eq:g rho 0\] ($\rho=0$) and Eq. \[eq:g z 0\] ($z=0$), while the points correspond to a numerical computation of the Fourier transform in Eq. \[eq:vmod\]. The right panel shows the 3-d plot of $g(\rho,z)$ in the $\rho$, $z$ plane with $z$ and $\rho$. []{data-label="fig:g_separate"}](rhozg_3d_eB26.ps "fig:"){width="3in"} We have calculated $g(\rho,z)$ for $eB=100 ~eB_c$ and $\mu=0.7$ MeV numerically. For these values, only the lowest Landau level is occupied and the Fermi momentum of the lowest Landau level is $k_f^z=0.48$ MeV. So the system is mildly relativistic and the electron Fermi energy is large compared to atomic binding energy. The two relevant length scales for the screened potential are the Debye screening length $\lambda_D=468$ (fm), and the wavelength of Friedel oscillations $\lambda_F=\pi/k_f^z=eB/(2\pi n_e)=1296$ (fm). In the left panel (Fig. \[fig:g\_separate\]) we plot $g(\rho, z)$ along the $z$ and the $\rho$ directions separately. This allows us to see the Friedel oscillations in the $z$ direction clearly, and also compare the numerical results with the approximate analytic expressions. Along the $z$ axis, Eq. \[eq:g rho 0\] is a good approximation to the numerical result, and in the $x-y$ plane, the exponential Debye formula describes the result well. A 3-dimensional plot of $g(\rho,z)$ in the $\rho$, $z$ plane is shown in the right panel and depicts that the Friedel oscillations are restricted to a narrow cylinder of radius $r \simeq 1/\sqrt{eB}$ in the x-y direction. For large chosen field of $4.4\times10^{15}$ Gauss, which is near the upper end of what can be found in magnetar surfaces, the amplitude of Friedel oscillations can be very large as shown in Fig. \[fig:g\_separate\]. Nonetheless, it is important to note that even small amplitude oscillations can affect the lattice structure due to their long-range character. In our zero temperature treatment the Friedel oscillations are undamped in the $z$ direction due to our assumption of a sharp Fermi surface. A finite temperature will smear the Fermi surface and this will result in exponential damping of Friedel oscillations by the factor $\exp{(-z/\xi_T)}$, where $\xi_T=2 \pi v_{Fe}/T$ is the thermal correlation length [@FetterWalecka:1971] and $v_{Fe}=k_f^z /\mu$ is the Fermi velocity. In neutron stars, the temperature is small and the thermal correlation length $\xi_T \gg a$ where $a$ is the inter-ion distance. Consequently even for small amplitudes, Friedel oscillations can have important effects due to their long-range nature when the screening potential from different ions add coherently. Heat transport by longitudinal lattice phonons\[Section:HeatTransport\] ======================================================================= Lattice phonons (lPhs) are space and time dependent vibrations of the ions, which can transport heat from one region of a solid to another. Typically both longitudinal and transverse modes contribute to heat conduction. In this analysis we restrict ourselves to only longitudinal modes since they primarily couple to electron-hole excitations. At low temperatures, when Umklapp process are suppressed [@Ziman:1972], and the occupation numbers of lattice modes is small enough that the non-linear terms in the lPh lagrangian are small, the dominant scattering of lPhs is with the electrons. We first re-derive the relation between the specific heat conductivity of lattice phonons and the imaginary part of the electronic polarization tensor, and recall results for scattering of lattice phonons in a free electron gas. Then we consider scattering in the presence of a magnetic field. At low temperatures the form of the interaction term between lattice phonons and electrons is well approximated by $${\cal{L}}_{el} = -\frac{1}{f_{el}}\int d^3 r \psi^\dagger\psi \partial_i\xi^i\;,$$ where $1/f_{el} = (Ze^2 n_I)/(m_D^2\sqrt{m_I n_I})$, with $n_I$, the number density of ions, $m_I$, the mass of ions, and $m_D$, the debye screening mass [@Ziman:1972]. The rate for phonon decay into an electron-hole ($e-h$) pair can be calculated from the thermal width of the phonon due to absorption by electrons, and is given by the imaginary part of the self energy correction due to the electron loop. $$\omega \Gamma (q^\mu) = -\frac{q^2}{f^2_{el}} \Im m \Pi(q^\mu) \;,~\label{eq:lPh width}$$ where $\Gamma(q^\mu)$ is the inverse lifetime $1/\tau(q^\mu)$ of the lPh, and $q^\mu=(\omega, {\bf{q}})$. For an on-shell photon $\omega=c_l q$. From Eq. \[eq:lPh width\], we obtain the mean free path $l_l(q^\mu)=\tau(q^\mu)c_l$, $$\begin{split} l_{l}(\omega, {{\bf{q}}}) &= -\bigl(\frac{c_l^3 f_{el}^2}{\omega \Im m \Pi(\omega, {{\bf{q}}})}\bigr)\;, \end{split}$$ where $c_l$ is the lPh speed. ![(color online) A plot of $\Im m[\Pi(\omega=3T,q_{z}=\frac{3T}{c_s}\cos\theta,q_{\perp}=\frac{3T}{c_s}\sin\theta)]$ as a function of the angle $\theta$ between the magnetic field and the lPh propagation direction, for three different densities $\rho_{12}$(10$^{12}$gm/cc). For this plot $eB=10eB_c$, $T=5$keV, and $c_s=0.05$. $\cos\theta=0$ corresponds to a lPh traveling perpendicular to the magnetic field, and $\cos\theta=1$ corresponds to one traveling parallel to it. At the lowest density, only the lowest Landau level is occupied and scattering is kinematically allowed only at one specific angle. As more levels are occupied, more angles are kinematically allowed.[]{data-label="fig:imPivscth"}](imPivsth_eB2.62.eps){width="5in"} The thermal conductivity contribution of the lattice phonons from kinetic theory is $\kappa = \frac{1}{3}C_v c_l l_{l}$. $C_v$ for lattice phonons is simply $$C_v = \bigl(\frac{T}{c_l}\bigr)^3\bigl(\frac{2\pi^2}{5}\bigr)~\label{eq:C_v lPh}\;.$$ The final expression for the specific heat conductivity for a typical phonon with $\omega=3T$ is then, $$\begin{split} \kappa = -\frac{2\pi^2}{15} T^3 \bigl(\frac{ c_l f_{el}^2}{\omega \Im m \Pi(\omega, {{\bf{q}}})}\bigr) =- \bigl(\frac{c_l 2\pi^2 T^2 f_{el}^2}{45\Im m \Pi(\omega, {{\bf{q}}})}\bigr)~\label{eq:kappa}\;. \end{split}$$ For $B=0$ the polarization function (ignoring the antiparticle contribution) of an electron gas is well known [@Jancovici:1962; @FetterWalecka:1971]. Here, for convenience, we will approximate $\Im m \Pi(q^\mu)$ by its value at $T=0$. The corrections to this approximation are small if $T/\mu$ is small, except at the kinematic boundaries. At $T=0$ the imaginary part of $\Pi(q^\mu)$ is, $$\Im m [\Pi(q^\mu)] = -\frac{\mu^2\omega}{2\pi q} \theta(q v_f-|\omega|)\label{eq:ImPiB0T0}\,.$$ For phonon decay and conductivity in Eq. \[eq:kappa\] we are interested in the imaginary part for $\omega/q = c_l < v_f$, which is given by $\Im m[\Pi(q^\mu)]= -\mu^2c_l/(2\pi)$. For $B\neq0$, $\Im m[\Pi_B(q^\mu)]$ is given in Eq. \[eq:ImPi\]. For a representative large value of the field in a magnetar we choose $eB=10eB_c$, and show results for the imaginary part of the polarization tensor, and the specific heat conductivity $\kappa$. The lattice phonons have a typical energy $\omega\sim 3T$, and we take $T=5$keV for our calculation. The magnitude of the momentum is given by $q=\omega/c_s$, and the $z$ and the $\perp$ components are $q\cos\theta$ and $q\sin\theta$ respectively. The speed of lattice phonons depends on the depth, but we take a representative value $c_s\sim0.05$. We show the results as a function of the mass density, in commonly used units $\rho_{12}$ corresponding to $10^{12}$gm/cc. To convert the mass density, which is dominated by ions, to the electron number density and hence the electron chemical potential, we take the atomic number of ions to be $Z=26$ and their mass number $A=56$ (Fe). In Fig. \[fig:imPivscth\], we show $\Im m [\Pi_B(q^\mu)]$ as a function of $\cos\theta$ for a three different values of $\rho_{12}$. In Fig. \[fig:kappavsrho\] we show the specific heat conductivity parallel ($\kappa_{z}$) and perpendicular ($\kappa_\perp$) to the magnetic field, as a function of $\rho_{12}$. From Fig. \[fig:imPivscth\] we see that at the lowest densities, the electrons occupy only the lowest Landau levels, and the response is highly anisotropic, and peaked around very specific values of $\theta$, where energy conservation and momentum conservation along the $z$ direction can be simultaneously satisfied. As the density of electrons increases, they occupy higher landau levels and overall response is obtained by summing contributions from various levels, and is non-zero for several values of $\theta$. ![(color online) Plot of $\kappa$ as a function of density. $\kappa_z$ is the specific heat conductivity parallel to the magnetic field, shown by the dashed (red) line. This should be considered as an upper bound for $\kappa_z$ because our calculation ignores Umklapp processes. $\kappa_\perp$ is the conductivity perpendicular the field shown by the solid line (green). The dotted line (black) is the conductivity for $B=0$ at the same density.[]{data-label="fig:kappavsrho"}](kappavsrho_eB2.62.eps){width="5in"} To calculate $\kappa_\perp$ and $\kappa_{z}$ in Fig. \[fig:kappavsrho\], we approximate the effective $\Im m[\Pi_B(\omega, {{\bf{q}}})]$ for scattering parallel to the magnetic field by the average between $\theta=[0,\pi/4]$, and the value perpendicular the magnetic field by the average between $\theta=[\pi/4,\pi/2]$. We see that the lPh conductivity is suppressed perpendicular to the magnetic field but not parallel to it. This result is easy to see in the ultra-relativistic limit. Energy-momentum conservation in this limit requires $|k^z+q\cos\theta|-|k^z|=c_s q$, which can occur only at $\cos\theta=c_s\ll\frac{1}{\sqrt{2}}$. Only at large densities where inter-level transitions become kinematically allowed, does the net response function become relatively isotropic. Formally, $\kappa_z$ is very large in our calculation, but we have ignored the Umklapp processes which will be important in the $z$ direction, because electron scattering is kinematically suppressed. One can include them by using $\kappa_z^{net}= \kappa_z^e\kappa^U/(\kappa_z^e+\kappa^U)$, where $\kappa_z^e$ can be read from Fig. \[fig:kappavsrho\], and $\kappa^U$ is the Umklapp contribution. In the perpendicular direction we expect electron scattering to be dominant, and it is alright to ignore the Umklapp contribution. We note that averaging the response function over angles is a rough method to include the effect of anisotropic scattering. A rigorous procedure will involve solving the transport equation using the anisotropic collision term, that can be deduced from the expression for $\Im m[\Pi_B(q^\mu)]$. We leave such a calculation for future work. Conclusions\[Section:Conclusions\] ================================== By calculating the one-loop electron-hole polarization function in a strong magnetic field we have found that the screening of the ion-ion potential is significantly different from those obtained in earlier studies where the Debye approximation was assumed. We show for the first time that the Fermi surface of electrons in the direction parallel to the magnetic field leads to Friedel oscillations in the ion-ion potential. These oscillations are large and could fundamentally change the structure of the ionic solid at large field and relatively low density when only the lowest Landau levels are occupied. For typical magnetar field strengths of order $10^{15}$ G, matter up to densities of the order $10^{10}$gm$/$cc will be affected by our finding here. To evaluate how Friedel oscillations and modifications to the average screening length affects the structure and melting properties of the solid we will need to perform either Path Integral Monte Carlo or classical Monte Carlo simulations because the ion-ion interaction is highly non-perturbative. This is beyond the scope of this work and will be investigated separately. Here, we present plausible implications based on qualitative arguments to explore how our findings will modify the ionic structure. Considering the parameters used in Fig. \[fig:g\_separate\], $eB=100eB_c$ and $\mu=0.7$(MeV) and assuming $Z=26$, the average number density of ions in the system is $n_I=3.2\times10^{-9}$ (fm)$^{-3}$. If one assumes that the ions form a regular bcc lattice, then the separation between nearest ions is $(\sqrt{3}/2)\times (2/n_I)^{1/3}=743$(fm). This is comparable to the screening length $\lambda_D$, and smaller than the wavelength for Friedel oscillations $\lambda_F$. If significant screening in the $x-y$ plane prevents the formation of a regular bcc structure, it may be favorable to form an anisotropic crystal structure. One possibility is that the ions fit into the troughs in the potential formed by superposition of the $V(z)$ separated in the $z$ axis by $\lambda_F$. To maintain neutrality, the ions should arrange themselves in the $x-y$ plane (in a regular or irregular structure) with an average separation $a_\perp=(\pi n_I\lambda_F)^{-1/2} = 278$(fm). If indeed the ions in the $z$ direction arrange themselves with separation $\lambda_F$, then the following picture arises for the structure of the ion-electron system with changing density, or as we move deeper into the neutron star crust. A single chain of ions arranged along the $z$ axis will look like a rod. $\lambda_F\propto eB/n_I$, and for fixed $eB$ decreases linearly with increasing density. The transverse separation between the rods, $a_\perp\propto (eB)^{1/2}$, on the other hand remains a constant. At the surface, it is known that the elongated atoms form chains parallel to the $B$ field. At low densities (near the surface of the crust), ions will form a plasma of rods with a small charge per unit length (since $\lambda_F$ is large). These rods will interact with each other by a two dimensional Debye screened Coulomb interaction. As we move deeper into the star, the charge per unit length will keep on increasing, and once $\lambda_F$ becomes comparable to $a_\perp$, the system will look more isotropic. This picture will break down when higher Landau levels start being occupied. It is interesting to note that for $Z=26$, $a_\perp\sim \lambda_F$ implies that $\mu^2 \gtrsim 2 eB$, meaning that the next Landau level is occupied. Hence, the point where the screened potential becomes more isotropic coincides with the point where the lattice structure becomes more isotropic. Hence, in our picture, the region where the electrons occupy only the lowest Landau level interpolates smoothly between the surface and the inner crust of the neutron star. We have also showed that the heat conduction due lattice phonons can become anisotropic because their damping rate due to electron-hole excitations is anisotropic. It is well known that the electronic heat conductivity perpendicular to the magnetic field is suppressed [@Urpin:1980], and this in turn results in a temperature anisotropy at the surface of the neutron star [@Geppert:2004]. Our finding that the heat conductivity due to phonons in the direction perpendicular to the magnetic field is also suppressed, is new. This effect may modify the temperature anisotropy in the outer regions of the magnetars during their early thermal evolution. Acknowledgements\[Section:Acknowledgements\] ============================================ We acknowledge discussions with Tanmoy Bhattacharya, Joe Carlson, Stefano Gandolfi and Charles Horowitz. We thank Dima Yakovlev for notes on screening in a magnetic field. RS thanks Huaiyu Duan on discussions about the accurate implementation of Laguerre polynomials. Dominance of the Lowest Landau level for $eB\gg\mu^2, q^2$ \[Section:LLL dominates\] ==================================================================================== To see that the $n=0$ term dominates in the sum in Eq. \[eq:Fb\], we look at the integrand, $$\frac{W(E_0,E_n,0,n,k^z_m,k^z_n,q^{z})}{E_m-E_n} =\frac{(E_0E_n+k^z_mk^z_n+m_e^2)e^{-q^2l^2/2}(q^2l^2/2)^n/n!}{2E_0E_n}\frac{1}{E_n-E_m}\;,$$ where we have used $L_0^n(x)=1$. Now, $E_m=\sqrt{(k^z_m)^2+m_e^2}\sim\mu$. For $n=0$, $E_n=\sqrt{(k^z+{q_{\perp}})^2+m_e^2}\sim \mu$ while for $n>0$, $E_n \sim \sqrt{2neB}$. Therefore, for $n=0$, $$\frac{W(E_0,E_n,0,n,k^z_m,k^z_n,q^{z})}{E_n-E_m} \sim e^{-q^2l^2/2}\frac{1}{2\mu}\;,$$ while for $n>0$, $$\frac{W(E_0,E_n,0,n,k^z_m,k^z_n,q^{z})}{E_n-E_m} \sim \frac{e^{-q^2l^2/2}(q^2l^2/2)^n/n!}{2(E_n-E_m)}\sim \frac{e^{-q^2l^2/2}(q^2l^2/2)^n/n!}{2\sqrt{2eB}}\;.$$ Thus, we can conclude that the processes involving excitation to the level $n$ is suppressed compared to the $m=0$, $n=0$ transition by a factor $(\mu/(\sqrt{2enB}))(q^2/(2eB))^n(1/n!)$ which is small if $eB\gg\mu^2$, $eB\gg q^2$ . Derivation of $g(r)$ in the non-relativistic limit \[Section:Friedel derivation\] ================================================================================= To proceed with the derivation in the non-relativistic limit, we start form the expression, $$\begin{split} g(\rho,z) &= \frac{\sqrt{\rho^2+z^2}}{\pi}\int_0^\infty d{q_{\perp}}{q_{\perp}}J_0({q_{\perp}}\rho)~\int_{-\infty}^\infty d{q^{z}}\frac{\exp{(i {q^{z}}z)}}{({q^{z}})^2+{q_{\perp}}^2+F_B({q_{\perp}},{q^{z}})} \\ &= \frac{\sqrt{\rho^2+z^2}}{\pi}\int_0^\infty d{q_{\perp}}{q_{\perp}}J_0({q_{\perp}}\rho) I({q_{\perp}},z)\;, \end{split}$$ where $$I({q_{\perp}},z) = \int_{-\infty}^\infty d{q^{z}}\frac{\exp{(i{q^{z}}z)}}{({q^{z}})^2+({q_{\perp}})^2+F_B({q_{\perp}},{q^{z}})} \;.$$ We calculate the integral $I({q_{\perp}},z)$ by using contour integration. We assume $z>0$ and close the contour in the upper half of the complex plane. We have to deform the contour to go around the branch cuts $\pm 2k_f+i\eta$, $\eta>0$. We call these two parts of the closed contour, $C_1$ (${q_{\perp}}=2k_f+i\eta$) and $C_2$ (${q_{\perp}}=-2k_f+i\eta$). This gives $I({q_{\perp}},z)+\int_{C_1}d{q^{z}}\frac{\exp{(i{q^{z}}z)}}{({q^{z}})^2+{q_{\perp}}^2+F_B({q_{\perp}},{q^{z}})}+\int_{C_2}d{q^{z}}\frac{\exp{(i{q^{z}}z)}}{({q^{z}})^2+{q_{\perp}}^2+F_B({q_{\perp}},{q^{z}})} =2\pi i {\rm{Res}} (\frac{\exp{(i{q^{z}}z)}}{({q^{z}})^2+{q_{\perp}}^2+F_B({q_{\perp}},{q^{z}})}) |_{{q^{z}}\ni ({q^{z}})^2+{q_{\perp}}^2+F_B({q_{\perp}},{q^{z}})=0} $. We treat these contributions one by one. First, we look at the pole contribution which we will call the Debye part $g_D$. The pole is where $$\begin{split} 0&={q_{\perp}}^2+({q^{z}})^2+F_B({q_{\perp}},{q^{z}})\\ &={q_{\perp}}^2+({q^{z}})^2+\bigl(\frac{e}{\pi}\bigr)^2\bigl(\frac{eB}{2}\bigr)e^{-{q_{\perp}}^2l^2/2} \log\bigl(\frac{2k_f+{q^{z}}}{2k_f-{q^{z}}}\bigr)\frac{m_e}{{q^{z}}} \end{split}$$ If we assume that the dominant contributions comes from ${q^{z}}\ll 2k_f$, then we want $$\begin{split} 0&={q_{\perp}}^2+({q^{z}})^2+\bigl(\frac{e}{\pi}\bigr)^2\bigl(\frac{eB}{2}\bigr)e^{-{q_{\perp}}^2l^2/2} \frac{m_e}{k_f}\\ &={q_{\perp}}^2+({q^{z}})^2+m_D^2e^{-{q_{\perp}}^2l^2/2} \end{split}$$ This gives, ${q^{z}}=i\sqrt{{q_{\perp}}^2+m_D^2\exp{(-{q_{\perp}}^2/(2eB))}}$, where we consider the pole in the upper half complex plane. Furthermore if the values of ${q_{\perp}}$ that contribute strongly to the $d{q_{\perp}}$ integral are ${q_{\perp}}\ll \sqrt{2eB}$, we can approximate $\exp{(-{q_{\perp}}^2/(2eB))}$ by $1$ and $$\begin{split} I_D({q_{\perp}},z) &= 2\pi i {\rm{Res}} (\frac{\exp{(i{q^{z}}z)}}{({q^{z}})^2+{q_{\perp}}^2+m_D^2})\\ & = \frac{2\pi}{2}\frac{e^{-z\sqrt{{q_{\perp}}^2+m_D^2}}}{\sqrt{{q_{\perp}}^2+m_D^2}} \end{split}$$ This is exactly what we obtain if we replace $F({q_{\perp}},{q^{z}})$ by $m_D^2$ in the first place. Hence, this integral is the same as what we obtain if we assume that there is a simple Debye like screening, and therefore gives, $$g_D(\rho,z) = e^{-m_D\sqrt{\rho^2+z^2}}$$ This is the answer if ${q_{\perp}}\ll k_f,\sqrt{eB}$ dominates the ${q_{\perp}}$ integral, and if $(\frac{e}{\pi})^2(\frac{eB}{2})(m_e/k_f)\ll k_f^2$. Now we consider the integrals over the contours $C_1$ and $C_2$, $$\begin{split} I_F&({q_{\perp}},z) = \int_{C_1}d{q^{z}}\frac{e^{(i{q^{z}}z)}}{({q^{z}})^2+{q_{\perp}}^2+F_B({q_{\perp}},{q^{z}})}+ \int_{C_2}d{q^{z}}\frac{e^{(i{q^{z}}z)}}{({q^{z}})^2+{q_{\perp}}^2+F_B({q_{\perp}},{q^{z}})}\\ &=\Re e\Bigl[\int_{C_1}d{q^{z}}\frac{2e^{(i{q^{z}}z)}}{({q^{z}})^2+{q_{\perp}}^2+F_B({q_{\perp}},{q^{z}})}\Bigr]\\ &=\Re e\Bigl[ \int_{\infty}^{0}d\eta\;\;2ie^{(2ik_f-\eta)z}\Bigl(\\ & \phantom{+}\frac{1} {{q_{\perp}}^2+(2k_f+i\eta)^2+(e/\pi)^2(eB/2)(m/(2k_f+i\eta))\exp(-{q_{\perp}}^2/(2eB))(1/2)(\ln((4k_f+i\eta)^2/\eta^2)-i\pi)}\\ & - \frac{1} {{q_{\perp}}^2+(2k_f+i\eta)^2+(e/\pi)^2(eB/2)(m/(2k_f+i\eta))\exp(-{q_{\perp}}^2/(2eB))(1/2)(\ln((4k_f+i\eta)^2/\eta^2)+i\pi)}\\ & \Bigr)\Bigr]\\ &=\Re e\Bigl[ \int_{0}^{\infty} d\eta\;\;2ie^{(2ik_f-\eta)z}\Bigl(\\ & \phantom{+} \frac{-1} {{q_{\perp}}^2+(2k_f+i\eta)^2+m_D^2(k_f/(2k_f+i\eta))\exp(-{q_{\perp}}^2/(2eB))(1/2)(\ln((4k_f+i\eta)^2/\eta^2)-i\pi)}\\ & + \frac{1} {{q_{\perp}}^2+(2k_f+i\eta)^2+m_D^2(k_f/(2k_f+i\eta))\exp(-{q_{\perp}}^2/(2eB))(1/2)(\ln((4k_f+i\eta)^2/\eta^2)+i\pi)} \Bigr)\Bigr]\;. \end{split}$$ Because of the $\exp(-\eta z)$ in the integration over $\eta$, for $z\gg 1/k_f$ the integral is dominated by $\eta\sim 1/z \ll k_f$. Therefore we can ignore $\eta$ whenever it is added to $k_f$. This simplifies the integral somewhat, $$\begin{split} I_F&({q_{\perp}},z) \sim \Re e\Bigl[ \int_{0}^{\infty} d\eta\\ &\frac{2e^{(2ik_f-\eta) z}m_D^2\exp(-{q_{\perp}}^2/(2eB))(\pi/4)} {({q_{\perp}}^2+(2k_f)^2+m_D^2\exp(-{q_{\perp}}^2/(2eB))(1/2)\ln(4k_f/\eta))^2+(m_D^2\exp(-{q_{\perp}}^2/(2eB))(\pi/4))^2} \Bigr]\;. \end{split}$$ To obtain an analytic form, we need one more simplification. We replace $\eta$ in $\ln(4k_f/\eta)$ by the value where we expect the integral to dominate, namely $\eta\sim 1/z$. The integral over $\eta$ can then be evaluated simply, and $$I_F({q_{\perp}},z) \sim\Bigl[ \frac{1}{z} \frac{2\cos{(2k_f z)}m_D^2\exp(-{q_{\perp}}^2/(2eB))(\pi/4)} {({q_{\perp}}^2+(2k_f)^2+m_D^2\exp(-{q_{\perp}}^2/(2eB))(1/2)\ln(4k_fz))^2+(m_D^2\exp(-{q_{\perp}}^2/(2eB))(\pi/4))^2} \Bigr]\;. $$ The Friedel contribution to $g$ is then, $$\begin{split} g_F(\rho,z) &= -\sqrt{\rho^2+z^2}\frac{1}{\pi} \int_{0}^{\infty} d{q_{\perp}}{q_{\perp}}J({q_{\perp}}\rho) I_F({q_{\perp}},z)\\ & = -\sqrt{\rho^2+z^2}\frac{1}{\pi}\frac{2\cos(2k_fz)}{z} \int_{0}^{\infty} d{q_{\perp}}{q_{\perp}}J({q_{\perp}}\rho)\\ & \frac{m_D^2\exp(-{q_{\perp}}^2/(2eB))(\pi/4)} {({q_{\perp}}^2+(2k_f)^2+m_D^2\exp(-{q_{\perp}}^2/(2eB))(1/2)\ln(4k_fz))^2+(m_D^2\exp(-{q_{\perp}}^2/(2eB))(\pi/4))^2}\\ \end{split}$$ Once again assuming ${q_{\perp}}\ll\sqrt{eB}$ we ignore the $\exp(-{q_{\perp}}^2/(2eB))$ term as before. Finally, assuming $k_f\gg m_D$ we can drop the $(m_D^2\pi/4)$ term in the denominator, and $$\begin{split} g_F(\rho,z)&\sim -\sqrt{\rho^2+z^2}\frac{1}{\pi}\frac{2\cos(2k_fz)}{z} \int_{0}^{\infty} d{q_{\perp}}{q_{\perp}}J({q_{\perp}}\rho) \frac{m_D^2(\pi/4)} {({q_{\perp}}^2+(2k_f)^2+m_D^2(1/2)\ln(4k_fz))^2}\\ &= -\frac{\sqrt{\rho^2+z^2}}{\pi}\frac{\cos(2k_fz)}{z} \frac{m_D^2(\pi/4)\rho}{\sqrt{(2k_f)^2+m_D^2(1/2)\ln(4k_fz)}} K_1(\rho\sqrt{(2k_f)^2+m_D^2(1/2)\ln(4k_fz)}) \label{eq:gFrhoz}\;. \end{split}$$ These manipulations work only if $z\gg 1/k_f$, (long distances in $z$) and $\rho\ll 1/k_f$. The assumption $\rho\ll 1/k_f$ is not a very essential one because for $\rho\gtrsim 1/k_f$, $g_F(\rho,z)$ decreases very rapidly and this is captured by the expression in Eq. \[eq:gFrhoz\], because of the modified Bessel function, $K_1$. Thus, the final answer for large $z$ is, $$\begin{split} g(\rho,z) &= g_D(\rho,z) + g_F(\rho,z)\\ &= e^{-m_D\sqrt{\rho^2+z^2}}\\ &-\frac{\sqrt{\rho^2+z^2}}{\pi}\frac{\cos(2k_fz)}{z} \frac{m_D^2(\pi/4)\rho}{\sqrt{(2k_f)^2+m_D^2(1/2)\ln(4k_fz)}} K_1(\rho\sqrt{(2k_f)^2+m_D^2(1/2)\ln(4k_fz)})\;. \end{split}$$

ArXiv

--- abstract: 'We present a theory of quantum circuits based on logical qubits encoded in chirality of electron spin complexes in lateral gated semiconductor triple quantum dot molecules with one electron spin in each dot. Using microscopic Hamiltonian we show how to initialize, coherently control and measure the quantum state of a chirality based coded qubit using static in-plane magnetic field and voltage tuning of individual dots. The microscopic model of two interacting coded qubits is established and mapped to an Ising Hamiltonian, resulting in conditional two-qubit phase gate.' author: - 'Chang-Yu Hsieh' - Pawel Hawrylak title: Quantum circuits based on coded qubits encoded in chirality of electron spin complexes in triple quantum dots --- Introduction {#sec:INTRO} ============ There is currently interest in exploiting electron spin for nano-spintronic[@sachrajda_hawrylak_book2003] and quantum information processing[@brum_hawrylak_sm1997; @korkusinski_hawrylak_book_2008; @loss_divincenzo_pra1998; @hanson_kouwenhoven_2007]. This is partly motivated by electron spin long coherence times [@hanson_witkamp_prl2003] and availability of scalable semiconductor technology. In the simplest approach, a physical qubit is identified with the two states of an electron spin, which can be manipulated by applying local magnetic fields. Much progress has been achieved using micro-magnet[@pioro-ladriere_tokura_apl2007; @pioro_obata_nat2008] technology with electron spin qubits. An alternative approach is to encode a logical qubit in a two level system constructed with spin complexes[@korkusinski_hawrylak_book_2008]. This includes singlet-triplet two electron qubit[@levy_prl2002] and a logical qubit encoded in a degenerate ground state of a three spin complex[@divincenzo_bacon_nature2000]. As suggested by DiVincenzo et al.[@divincenzo_bacon_nature2000], in the framework of a Heisenberg model for a logical qubit encoded in a linear chain of three spins, such a logical qubit can be manipulated by the control of the exchange interaction between pairs of spins. The ability to manipulate spin state with a voltage is related to the relation between the orbital and spin part of the many-electron wavefunction. A detailed microscopic model of a three electron complex in a single semiconductor quantum dot (QD), exploring the orbital and spin relation, has been investigated and compared with experiment by one of us[@hawrylak_prl1993]. A similar microscopic model of three electrons localized in a lateral gated triangular triple quantum dot molecule in the plane of a GaAs/GaAlAs heterojunction was proposed and explored by one of us in Ref. . The coded qubit was identified with chirality of three electron spin complex, or equivalently, two possible directions of a minority spin motion. Other proposals to implement coded qubits with both triple quantum dots[@weinstein_hellberg_pra2005_2] (TQDs) and atom traps[@georgeot_mila_arxiv2009; @trif_troiani_prl2008]also exist. The advantages of working with a TQD-based coded qubit are two-fold. First, every quantum gate can be implemented electrically. In such a scheme, magnetic field will only be used for initialization and (or) measurement of a coded qubit. Second, the coded qubit involves Decoherence Free Subspace[@bacon_kempe_prl2000; @viola_fortunato_sci2001; @weinstein_hellberg_pra2005_2; @lidar_whaley_book2008] (DFS) and is immune to channels of collective decoherence. This reduces decoherence of a TQD-based coded qubit due to charge fluctuations as discussed in Ref. . Other interesting phenomena involving TQD molecules include non-Fermi liquid behavior when coupled to leads[@delgado_hawrylak_prb2008; @ingersent_ludwig_prl2005], potential for generating maximally entangled three-partite GHZ and W states[@eibl_kiesel_prl2004; @ghosh_kar_njp2002]and manipulation of total spin[@shim_delgado_prb2009]. In this work we develop a microscopic theory of quantum circuits based on coded qubits encoded in chirality of electron spin complexes in TQD. We use a combination of Linear Combination of Harmonic Orbitals - Configuration Interactions (LCHO-CI)[@gimenez_korkusinski_prb2007], Hubbard and Heisenberg models to determine a set of optimal conditions for single qubit operations and describe the two qubit gate. We show that there exists an in-plane magnetic field direction and magnitude optimal for single qubit operations. However, the magnetic field will rotate qubit states in an undesirable way; and we show how this rotation can be controlled by tuning voltages on the gates. Once the optimal sets of magnetic fields and voltages have been obtained, the exact diagonalization techniques are used to establish and verify an effective two-qubit Hamiltonian. It is shown that the effective two qubit interaction is Ising-like, leading to a two-qubit phase gate. The present work establishes both single qubit and two-qubit operations necessary for performing quantum computation using TQD-based coded qubits. The paper is organized as follows. Sec. \[sec:INTRO\] contains introduction. Sec. \[sec:SYS\] describes the quantum circuit based on chirality of 3 electron spin complexes and computational methodology. In Sec. \[sec:CQBT\], we present a definition of a TQD-based coded qubit, discuss initialization, single qubit operations using voltages, and measurement of the state of the coded qubit. In Sec. \[sec:DQOP\], we present an effective Hamiltonian of 2 coupled coded qubits and show that it can be canonically transformed to an Ising interaction. Sec. \[sec:CON\] contains summary. The Model {#sec:SYS} ========= Fig.(\[fig:layout\]a) shows schematically layout of quantum circuits based on coded qubits encoded in chirality of electron spin complexes in triangular TQDs. The circles denote individual lateral quantum dots formed in the 2D electron gas (2DEG) at the heterojunctions of AlGaAs/GaAs by metallic gates on the AlGaAs surface. The gates are set to confine a single electron in each dot denoted by an arrow. Additional gates (not shown here) are used to control tunneling between dots in the same TQD molecule, shown schematically as solid lines. Dashed lines indicate tunneling between neighboring TQDs. The tunneling between any pair of QDs is responsible for exchange interaction of electron spins localized on each QD. The brackets indicate two TQDs isolated from the rest of the circuit. It is assumed that any number of TQDs can be isolated from the rest of the circuit by turning off the boundary exchange interactions with gate voltages. For comparison, Fig.(\[fig:layout\]b) shows the circuit composed of linear, instead of triangular, TQDs, similar to the linear chain of spins proposed by DiVincenzo et al., in which the coded qubits in the chain are implemented with only two exchange interactions. Fig.(\[fig:layout\]c) is another possible architecture studied by Weinstein[@weinstein_hellberg_pra2005_2] et. al. for spin system. However, in this design, bringing TQDs close together induces interactions between quantum dots beyond the ones indicated by dashed line. In this work we focus mainly on triangular TQDs-based quantum circuit shown in Fig.(\[fig:layout\]a). Since electrons are well localized in each QD, a system of two TQDs in a chain can be very well described by an extended Hubbard model[@korkusinski_gimenez_prb2007]. With $c^{\dag}_{i\sigma} (c_{i\sigma}) $ electron creation (annihilation) operator for the electron with spin $\sigma=\pm1$ on the $i-th$ QD, the Hubbard Hamiltonian reads: $$\begin{aligned} \label{eq:NHubbard} \hat{H}_{2} &=& \sum_{n=1}^{2}\sum_{i=3n-2}^{3n}\sum_\sigma E_{i\sigma} \hat{n}_{i\sigma} + \sum_{n=1}^{2} \sum_{\substack{i, j = 3n-2\\ i \neq j }}^{3n} \sum_ \sigma t_{ij} \hat{c}^{\dag}_{i\sigma}\hat{c}_{j\sigma} \nonumber \\ & & + \frac{1}{2}\sum_{n=1}^{2} \sum_{\substack{i, j = 3n-2\\ i \neq j }}^{3n}V \hat{\rho}_{i} \hat{\rho}_{j} + \sum_{n=1}^{2} \sum_{i=3n-2}^{3n}U \hat{n}_{i\uparrow}\hat{n}_{j\downarrow} \nonumber \\ & & + \sum_{\sigma} t' \left(\hat{c}^{\dag}_{3\sigma}\hat{c}_{4\sigma}+\hat{c}^{\dag}_{4\sigma}\hat{c}_{3\sigma}\right) + V' \hat{\rho}_{3} \hat{\rho}_{4},\end{aligned}$$ where $n=1,2$ labels the two TQD molecules from left to right, indices $i,j$ range from $3n-2$ to $3n$ label each QD from left to right in the $n-th$ TQD molecule. The intra-TQD Hubbard parameters $t_{ij}$, $V$, $U$ and $E_{i\sigma}$ are the tunneling matrix element between the $i-th$ and the $j-th$ QDs, the off-site Coulomb interaction parameter between any two QDs in the same TQD molecule, the on-site Coulomb interaction strength for any QD, and the on-site energy for the $i-th$ QD respectively. The on-site energy depends on spin and magnetic field which is applied in the plane of a TQD molecule. The Hubbard parameters $t'$ and $V'$ represent the inter-molecular tunneling matrix element and inter-molecular Coulomb interactions. $\hat{n}_{i\sigma}=\hat{c}^{\dag}_{i\sigma}\hat{c}_{i\sigma} $ is the number operator for the $\sigma$ electron on the $i-th$ QD. $\hat{\rho}_i = \hat{n}_{i\uparrow} + \hat{n}_{i\downarrow}$ is the electron charge density operator on the $i-th$ QD. When we discuss triangular resonant TQDs, we drop the subscripts on all the parameters. In our model, we consider the parameters corresponding to regime of strong correlations: $t'_{ij} < t_{ij} << V' < V < U$. The intra-TQD Hubbard parameters, $E_i$, $ t_{ij}$, $V$, and $U$, in Eq.(\[eq:NHubbard\]) are obtained from a microscopic calculation for single TQD based on LCHO-CI method as explained in Ref. . The inter-TQD Hubbard parameter $V'$ is taken to be the direct Coulomb interaction between 2 charges localized on adjacent edge dots of two neighboring TQD molecules. The energy spectrum and eigenstates of the Hubbard Hamiltonian for one and two TQD molecules are obtained using Configuration Interaction technique. For a given number of electons $N_e$ we construct all possible configurations $\vert k={i_{N_e}\sigma_{N_e},..,i_2\sigma_2,i_1\sigma_1}\rangle= \hat{c}^{\dag}_{i_{N_e} \sigma_{N_e}}... \hat{c}^{\dag}_{i_2\sigma_2}\hat{c}^{\dag}_{i_1\sigma_1}\vert 0 \rangle$, build Hamiltonian matrix in the space of configurations, and diagonalize it numerically. At half-filling, the low-energy spectrum of the Hubbard model can be approximated by a spectrum of a Heisenberg model[@korkusinski_gimenez_prb2007; @scarola_park_prl2004; @scarola_dassarma_pra2005] describing electron spins localized in each dot: $$\label{eq:Heis1} \hat{H}_{heis} = \sum_{ i < j } J_{ij} \mathbf{S}_i \cdot \mathbf{S}_j.$$ The exchange interactions, $J_{ij}$, for the TQDs can be expressed in terms of tunneling matrix elements and quantum dot energies: $$\label{eq:SingleExchange} J_{ij} = 2\vert t_{ij} \vert ^2\left(\frac{1}{U-V+(E_i-E_j)}+\frac{1}{U-V-(E_i-E_j)}\right).$$ The exchange interaction can be controlled by either tuning the tunneling matrix element $t_{ij}$ by, for example, additional gates controlling the height of the tunneling barrier, or by biasing the dots and changing their on-site energy $E_{i(j)}$. Coded qubit encoded in chirality of electron spin complex in a triple quantum dot molecule {#sec:CQBT} ========================================================================================== In this section, we discuss a single coded qubit shown in Fig.(\[fig:bfield\]a): its preparation, initialization, operation and measurement in the presence of a lateral magnetic field in the $y$ direction. The qubit is encoded in quantum states of a three electron spin complex in a fully symmetric and half-filled TQD. An example of one of the three possible configurations $\left| \uparrow \downarrow \downarrow \right\rangle= \hat{c}^{\dag}_{1 \uparrow}\hat{c}^{\dag}_{2 \downarrow} \hat{c}^{\dag}_{3\downarrow}|0>$ with $S_y=-1/2$ is shown schematically in Fig.(\[fig:bfield\]a). The three configurations with the minority spin on QD 1, 2 or 3 form a doubly degenerate ground state with total spin $S=1/2$ and an excited state with total spin $S=3/2$ separated from the ground state by $J_{eff}$, as shown in Fig.(\[fig:bfield\]b). The doubly degenerate ground state with fixed $S_y=-1/2$ forms an effective two level system. We identify the coded qubit levels $\vert q_{+} \rangle$ and $\vert q_{-} \rangle$ with $$\begin{aligned} \label{eq:ChiralStates} \left| q_s \right\rangle & = & \frac{1}{\sqrt{3}}\sum_{n=1}^{3} e^{i 2\pi n s / 3} \left| n \right\rangle,\end{aligned}$$ where $s=\pm$ and $\left| n \right\rangle = S_n^+ \left| \downarrow \downarrow \downarrow \right\rangle$. The three spin complex is characterized by chirality $\chi = \left( \mathbf{S}_1 \times \mathbf{S}_2 \right) \cdot \mathbf{S}_3$ which measures the degree of collinearity of the three spins. The two coded qubit levels, Eq.(\[eq:ChiralStates\]), are the eigenstates of chirality operator with eigenvalues $\chi= \pm \sqrt{3}/4 $. These two states also can be more intuitively characterized by minority spin moving either to the left or to the right as in Resonant Valence Bond (RVB) plaquette [@wen_wilczek_prb1989; @fradkin_book; @tatara_garcia_prl2003; @bulaevskii_batista_prb2008]. We note that in the absence of the magnetic field the coded qubit states in $S_y=1/2$ subspace are degenerate with coded qubit states in $S_y=-1/2$ subspace. If we are to work in the computational space corresponding to $S_y=-1/2$, any process which flips the electron spin will remove the coded qubit from its computational space. In order to separate the computational Hilbert space $S_y=-1/2$ from $S_y=+1/2$ subspace, a magnetic field $B_y$ is applied along the $y$ direction as shown in Fig(\[fig:bfield\]a). We avoid applying magnetic field applied in the $z$ direction, because it activates undesirable, higher order spin-spin interactions[@scarola_park_prl2004; @scarola_dassarma_pra2005], whereas the magnetic field applied in the plane only modifies the on-site energies of QDs: for $B_y$ field, QDs 1 and 3 energy level as $E_{1(3)}(B)= E_{1(3)}(0)+1/2 \omega_c (R/2)^2+ g\mu B S_y$, where $R$ is the spatial separation of QDs 1 and 3 and $\omega_c = eB_y/m^*_ec$ is the cyclotron frequency, while the QD 2 energy level only acquires the Zeeman term. Hence magnetic field in the $y$ direction effectively lowers the energy of the QD 2 by $\Delta E_2= - 1/2 \omega_c (R/2)^2$ with respect to QDs 1 and 3. Fig.(\[fig:bfield\]b) shows the energy spectrum of the coded qubit as a function of $\omega_c $ obtained in LCHO-CI method. In our LCHO-CI calculation, we use the following parameters for a specific resonant TQD. The inter-dot distance is 10 $a_B^*$, where $a_B^*=9.79 nm$ is the effective Bohr radius for $GaAs$. The confining Gaussian potential on the $i-th$ QD is of the form, $V_i(x,y) = -V_i^0 \exp\left(-((x-x_i)^2+(y-y_i)^2)/d_i^2\right)$, where $V_i^0=4.0 Ry^*$ and $d_i=2.5 a_B^*$. $ Ry^*=5.93 meV$ is the effective Rydberg for $GaAs$. In Fig.(\[fig:bfield\]b), the optimal cyclotron frequency $\omega_c^*$ corresponds to $B^* = 0.69 T$. At $B=0$, the four-fold degenerate ground state is separated from the four-fold degenerate spin polarized excited state by $J_{eff}$. The two computational Hilbert spaces corresponding to $S_y = \pm 1/2$ separate energetically with increasing magnetic field while the spin polarized states decrease in energy. At magnetic field $B^*$, such that $ J_{eff}= 2 g \mu B^*$, the energy gap is maximized and should correspond to the working point that can maintain the longest coherence time of the coded qubit. The coded qubit should operate at this value of magnetic field. However, as discussed above, the magnetic field effectively biases the QD 2. This removes the degeneracy of the two qubit levels and rotates them from their zero magnetic field states. The energy splitting of the two coded qubit levels as a function of applied magnetic field is shown in Fig.(\[fig:gate\]a). The splitting is a fraction of the large energy scale $J_{eff}$. For the largest gap, i.e. $B=B^*$, the splitting is $\Delta$. In order to restore the degeneracy of the two coded qubit levels, one can apply voltage to the QD 2. In Fig.(\[fig:gate\]b), microscopic calculations done in LCHO-CI show that a positive voltage bias on the QD 2 can indeed bring the TQD back on resonance in the presence of $B$ field. We have now established the coded qubit and the best conditions for its operation. In order to operate the coded qubit, we need to be able to initialize it. We propose to initialize the coded qubit by turning off both interaction between QDs 1 and 3 and interaction between QDs 1 and 2. The only remaining interaction is between QDs 2 and 3. The ground state of a TQD, $\left| L_0 \right\rangle = \vert \downarrow_1 \rangle \vert S_{23} \rangle $, becomes a product of a spin down state of an electron on the QD 1 and a singlet state of electrons across QDs 2 and 3. This intuitively is a ground state in magnetic field. The singlet state of two electrons in a pair of QDs 2 and 3 can be generated in real time starting from two electrons in a biased QD 3. This procedure does not generate directly the coded qubit levels $\vert q_ {+} \rangle$ and $\vert q_{-} \rangle$, but the state $\left| L_0 \right\rangle $ is a linear superposition of the two qubit levels: $\left| L_0 \right\rangle = {-i \over{\sqrt{2}}} ( e^{i2\pi/3} \vert q_- \rangle -e^{-i2\pi/3}\vert q_+ \rangle ) $. Once interactions are turned on, any state can be obtained from the initial state. This can be seen by writing the TQD Heisenberg Hamiltonian in the basis of the two coded qubit levels $\left\{ \vert q_{\pm} \rangle \right\}$ , $$\label{eq:singleCodedQubit} \hat{H}_{1q} = \frac{1}{2}\left(J_{12} - \frac{1}{2}J_{13}- \frac{1}{2}J_{23}\right) \hat{\sigma}_x + \frac{\sqrt{3}}{4}\left( J_{13}-J_{23}\right) \hat{\sigma}_y.$$ If we take $J_{13}=J_{23}$, and let $2J_{12} > J_{13}+J_{23}$, then the Heisenberg Hamiltonian corresponds to $\hat{\sigma}_x$ operation. If we take $J_{23} > J_{13}$ and $2J_{12} = J_{13} + J_{23}$, then the coded qubit Hamiltonian Eq.(\[eq:singleCodedQubit\]) corresponds to $\hat{\sigma}_y$ operation. The capability to rotate a qubit with respect to two different axes on a Bloch sphere allowes us to generate arbitrary single qubit operations. In practice, we tune the exchange interaction $J_{ij}$ through biasing QDs and changing their energies. As long as the biasing $|\Delta E| << U-V$ is satisfied, we expect the quantum state to remain in the qubit subspace during the process of tuning the exchange interactions as discussed in Ref. . Next, we discuss the measurement of coded qubits. Several proposals[@bulaevskii_batista_prb2008; @cao_hu_pla2008; @georgeot_mila_arxiv2009] discusses methods of detecting chirality of a triangular (three-body) antiferromagnetic cluster. For our specific TQD-based coded qubit, we propose to apply a $B_z$ field to split the two coded qubit levels with different chirality. As discussed earlier, the application of $B_z$ field will give rise to additional terms proportional to chirality operator on top of the Heisenberg model, Eq.(\[eq:Heis1\]), used for qubit modeling. Since the additional term to the Heisenberg model is strictly proportional to the chirality, it splits the two coded qubit levels with the gap given by $$\label{eq:chiralgap} \frac{12\sqrt{3}|t|^3}{(U-V)^2}\sin\left( \frac{2\pi\Phi}{\Phi_0}\right),$$ where $\Phi$ is the magnetic flux through the device and $\Phi_0 = eh/c$ is the flux quanta. This energy gap can be interpreted as the energy difference between two magnetic dipole moments[@bulaevskii_batista_prb2008] oriented in opposite directions under $B_z$. In Ref. , Scarola et. al. also showed that the unintended qubit rotation due to the additional chirality term can be significantly reduced for a range of specific weak $B_z$ fields that satisfy $R/l_h \geq 1$ and $\omega_c/\omega_o \leq 1$, where $l_h = \sqrt{\hbar c/eB}(1+4\omega_o^2/\omega_c^2)^{-1/4}$ is the modified magnetic length, and $\omega_o$ is the confining frequency used to approximate the potential of a QD. Thus, initializing the system under $B_z$ might not be a good idea as the field has to stay on for the entire period of quantum computation and eventually modify the quantum state of the coded qubit due to magnetic moment coupling with $B_z$. However, a measurement done with advanced spectroscopy only requires $B_z$ to be turned on for a relatively short time and significantly limit the extent to which the coded qubit will be modified. Double qubit operations {#sec:DQOP} ======================= We now turn to discuss double coded qubit operations using both Hubbard and Heisenberg models. Starting with the Hubbard Hamiltonian, Eq.(\[eq:NHubbard\]), we derive perturbatively a Heisenberg Hamiltonian for a complex of six electron spins with a set of exchange interactions $J_{ij}$ $$\begin{aligned} \label{eq:DoubleExchange} J_{12} = J_{56} = J_a & = & \frac{4t^2}{U-V} ,\nonumber \\ J_{13} = J_{23} = J_{45} = J_{46} = J_b & = & \frac{2t^2}{(U-V+V')}+\frac{2t^2}{U-V-V'} , \nonumber \\ J_{34} = J_c & = & \frac{4t'^2}{U-V'},\end{aligned}$$ with all other exchange interactions set to zero. For all the Hubbard parameters considered in this study (i.e. weak inter-TQD tunneling and strong Coulomb interaction), $J_c$ is usually about two order of magnitudes smaller than $J_a$ and $J_b$, mainly due to the fact that $t' < t / 2$ in our study. We treat the inter-TQD exchange interaction, scaled by $J_c$, perturbatively to derive an effective interaction in the coupled coded qubits subspace, $\left\{ \vert q_+^L\rangle \vert q_+^R\rangle,\vert q_+^L\rangle \vert q_-^R\rangle, \vert q_-^L\rangle \vert q_+^R\rangle,\vert q_-^L\rangle \vert q_-^R\rangle \right\}$, where the superscript $L$ and $R$ stand for left and right TQD respectively. The effective qubit-qubit interacting Hamiltonian reads $$\label{eq:codedqubit4} \hat{H}_{2q} = 4|\alpha'|(\hat{n} \cdot \vec{S}_1)(\hat{n} \cdot \vec{S}_2)+2|\beta'|\hat{m} \cdot \vec{S}_1+2|\beta'|\hat{m} \cdot \vec{S}_2,$$ where $ \alpha' = e^{-\frac{i2\pi}{3}}J_c/9,$ and $ \beta' = e^{-\frac{i\pi}{3}}\left(-J_a/2+J_b/2+J_c/27\right),$ and the direction vectors in Eq.(\[eq:codedqubit4\]) are defined as follows: $ \hat{n} = \left(\sqrt{|\alpha'|+Re(\alpha')}/\sqrt{2|\alpha'|}, \sqrt{|\alpha'|-Re(\alpha')|}\sqrt{2|\alpha'|},0\right),% \nonumber \\ \hat{m} = \left( Re(\beta')/|\beta'|, Im(\beta')/|\beta'|, 0 \right)$. Thus, the effective interaction for 2 TQD-based code qubits is equivalent to an $XY$ Hamiltonian under a uniform in-plane magnetic field for 2 spins. If we rotate the effective qubit-qubit interacting Hamiltonian, Eq.(\[eq:codedqubit4\]), from the coded qubit basis to the Jacobian basis, $\left\{ \vert L_0^L\rangle \vert L_0^R\rangle,\vert L_0^L\rangle \vert L_1^R\rangle, \vert L_1^L\rangle \vert L_0^R\rangle,\vert L_1^L\rangle \vert L_1^R\rangle \right\}$, the rotated Hamiltonian reads $$\label{eq:2qubitL} \hat{H}_{2Lq} = \alpha S_1^zS_2^z + \beta S_1^z + \beta S_2^z ,$$ where $\alpha =J_c / 9 $, and $\beta = -J_a/2 + J_b/2 + J_c/18$ . The Jacobian state $\vert L_0 \rangle$ is defined in Sec.\[sec:CQBT\] when we discuss the initialization of the coded qubit. The Jacobian state $\vert L_1\rangle = \sqrt{1/3}\vert \downarrow_1 \rangle \vert T_{23}^0 \rangle - \sqrt{2/3} \vert \uparrow_1 \rangle \vert T_{23}^- \rangle$, where $\vert T_{23}^0\rangle$ is a $S_z=0$ triplet state of electrons on QDs 2 and 3 and $\vert T_{23}^-\rangle$ is a $S_z=-1$ triplet state of electrons on QDs 2 and 3. Similarly, $\vert L_1 \rangle = -1/\sqrt{2} ( e^{i2\pi/3} \vert q_- \rangle + e^{-i2\pi/3}\vert q_+ \rangle )$ can also be written as a linear combination of the coded qubit levels. Therefore this rotation of basis only requires single qubit operations applied to each coded qubit; the two Hamiltonians, Eq.(\[eq:codedqubit4\]) and Eq.(\[eq:2qubitL\]), are locally equivalent[@makhlin_qip2002]. As Ising interaction can be used to generate CNOT gate[@nielsen_chuang_book], we show how to generate non-local two qubit interactions with 2 TQD-based coded qubits. Fig.(\[fig:ising\]) shows results of exact diagonalization of the Hubbard Hamiltonian for two coupled TQDs as a function of increasing coupling $t'$ between the two molecules. The inset shows the entire energy spectrum at $t'/t=0.2$. The low energy spectrum consists of four levels, characterized mostly by the two lowest levels of each TQD. These 4 low-lying energy levels are well separated from the rest of the spectrum and constitute the two coupled coded qubit subspace. The low-lying spectrum contains a doubly degenerate level, which is a signature of the Ising model in an external field. In practice, the tuning of $J_c$, the inter-TQD exchange interaction, should be done via tuning the tunneling parameter $t'$. Fig. (\[fig:ising\]) shows that the Ising model features of the energy spectrum are maintained over a wide range of values of $t'$, and hence the Ising model Hamiltonian, derived above, describes the coupling of two coded qubits very well. Conclusion {#sec:CON} ========== In summary, we present a theory of quantum circuits based on coded qubits encoded in chirality of electron spin complexes in lateral gated semiconductor triple quantum dot molecules with one electron spin in each dot. Using microscopic Hamiltonian and exact diagonalization techniques we show how to initialize, coherently control and measure the quantum state of a chirality based coded qubit using static in-plane magnetic field and voltage tuning of individual QDs. The microscopic model of two interacting coded qubits is established and mapped to an Ising Hamiltonian. Hence both conditional two-qubit phase gate and voltage controlled single qubit operations are demonstrated. Acknowledgment {#sec:Ack} ============== The authors thank NSERC, QUANTUMWORKS, CIFAR, NRC-CNRC CRP and OGS for support. C.-Y. Hsieh would like to thank Y.-P. Shim and A. Sharma for useful discussions. [36]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , ** (, ). , ****, (). , ** (, ). , ****, (). , , , , , **** (). , , , , , ****, (). , , , , , **** (). , , , , , , , , ****, (). , ****, (). , , , , , ****, (). , ****, (). , ****, (). , , ****, (). , ****, (). , . , , , , ****, (). , , , , ****, (). , , , , , , ****, (). , ** (, ). , , , , ****, (). , . , , , ****, (). , , , , , ****, (). , , , , , ****, () , , , ****, (). , , , , , , ****, (). , ****, (). , , , ****, (). , ****, (). , , , ****, (). , ** (, ). , ****, (). , , , , ****, (). , ****, (). , ****, (). , ** (, ). ![ Three possible layouts for a chain of TQD-based coded qubits. Each solid and dash line represents an exchange interaction between a pair of QDs. The layouts (a) and (b) differ in that the constituent coded qubits are implemented with triangular TQD in (a) and with linear TQD in (b). The layouts (a) and (c) differ in the way logical qubits are connected in a chain. If the inter-qubit distance is reduced beyond a threshold, an additional set of exchange interactions will arise and dominate the behavior of the system in layout (c). []{data-label="fig:layout"}](Fig1.eps){width="90.00000%"} ![(a). A schematic of a single TQD initialized in $S_y = -1/2$ subspace under a $B_y$ field. The $B_y$ field breaks the discrete rotational symmetry, and it affects QD 2 differently from QDs 1 and 3. (b). The energy spectrum of a triangular TQD as a function of $\omega_c / \omega_c^*$, where $\omega_c$ is the cyclotron frequency used in LCHO-CI calculation. $\omega_c^*$ corresponds to a magnetic field $B^*$ such that the gap between computational subspace (the lowest energy level in the plot) and the rest of spectrum is maximized as indicated by the black arrow. []{data-label="fig:bfield"}](Fig2.eps){width="90.00000%"} ![ (a). The energy spectrum of qubit levels (the zoom in of the lowest level in Fig. (2b)) as a function of $\omega_c / \omega_c^*$. The $B_y$ field splits the doubly degenerate ground state, and $\Delta$ corresponds to the gap between the two levels at $\omega_c^*$ or equivalently at $B_y^*$. (b). The energy spectrum of the qubit levels under $B_y^*$ as a function of voltage $V_2$ on QD2, and $V_2^0$ is the unbiased voltage at resonant condition when there is no external $B$ field. The splitting of levels under $B$ field can be restored via gate voltage tuning.[]{data-label="fig:gate"}](Fig3.eps){width="70.00000%"} ![The four lowest energy levels of 2 coupled triangular TQDs as a function of inter-TQD tunneling $t'$. The energy spectrum resembles that of the Ising model with an external field: a doubly degenerate levels corresponding to state $\vert \downarrow \uparrow \rangle$ and $\vert \uparrow \downarrow \rangle$ and two unique levels $\vert \downarrow \downarrow \rangle$ and $\vert \uparrow \uparrow \rangle$. “$X2$” denotes a doubly degenerate level. Inset: The entire energy spectrum at $t' / t = 0.2$ calculated with LCHO-CI method.[]{data-label="fig:ising"}](Fig4.eps){width="90.00000%"}

ArXiv

-1cm -1.5cm [**Stringy Sphalerons and Gauss–Bonnet Term**]{}\ \ [*Laboratory of High Energies,\ Joint Institute for Nuclear Research, 141980 Dubna, Russia,\ e–mail: [email protected]* ]{}\ and\ [**Dmitri V. Gal’tsov,**]{}\ [*Department of Theoretical Physics, Physics Faculty,\ Moscow State University, 119899 Moscow, Russia,\ e–mail: [email protected]* ]{}\ [**Abstract**]{} The effect of the Gauss–Bonnet term on the $SU(2)$ non–Abelian regular stringy sphaleron solutions is studied within the non–perturbative treatment. It is found that the existence of regular solutions depends crucially on the value of the numerical factor $\beta$ in front of the Gauss–Bonnet term in the four–dimensional effective action. Numerical solutions are constructed in the $N=1, 2, 3$ cases for different $\beta$ below certain critical values $\beta_N$ which decrease with growing $N$ ($N$ being the number of nodes of the Yang–Mills function). It is proved that for any static spherically symmetric asymptotically flat regular solution the ADM mass is exactly equal to the dilaton charge. No solutions were found for $\beta$ above critical values, in particular, for $\beta=1$. Since the Bartnik and McKinnon’s discovery [@bm] of the regular particle–like solutions to the coupled system of the Einstein–Yang–Mills (EYM) equations there was a growing interest in revealing their possible physical significance. It was shown that these solution could play at the ultramicroscopic distances a role analogous to that of electroweak sphalerons [@gv]. Sphaleron interpretation is supported by the existence of the odd–parity YM negative modes [@onm] (apart from the previously known even–parity ones [@sz]), as well as the fermion zero modes and the level–crossing phenomenon [@fer]. Natural question arises whether the EYM sphalerons survive in more sophysticated field models suggested by the theory of superstrings. It was shown recently that regular sphaleron solutions exist within the context of the Einstein–Yang–Mills–Dilaton (EYMD) theory [@dg1], [@lm], [@tm], [@dg2], [@oneill]. Remarkably, they have a dilaton charge exactly equal to the ADM mass. This property is similar to that of the extremal dilaton black holes which are likely (at least some) to represent exact solutions of the string theory. To further investigate possible relevance of the EYMD sphalerons to the string theory we study here the EYMD system with the Gauss–Bonnet (GB) term which is typically present in stringy gravity as the lowest order curvature correction. Similar problem for Abelian dilatonic black holes was studied recently within the perturbative approach [@msm]. However, for regular solutions, one needs a more precise treatment. In order to see in a continuous way how EYMD solutions are modified by the GB term we introduce into the lagrangian a numerical factor $\beta$ so that $\beta=0$ corresponds to the pure EYMD system. It turns out that series expansion of the regular solution near the origin is essentially $\beta$–dependent. Also, computing the GB contribution into the effective energy density on the background EYMD solutions, one can observe that the GB effect becomes non–small for $\beta$ of the order of unity. For this reason we avoid any perturbative treatment of the GB term and attack the problem numerically. Starting with $\beta=0$ we increase gradually the value of this parameter and search (using the shooting strategy) for solutions interpolating smoothly between the regular asymptotic expansion near the origin and an asymptotically flat expansion at infinity. Although the leading terms of expansions near infinity are not modified by the GB corrections, those near the origin [*are affected*]{} substantially. We construct numerical solutions for $N=1, 2, 3$ and some $\beta \neq 0$ and show that regular solution cease to exist above certain critical values $\beta_N$ depending of the number of nodes $N$ of the YM function. For all solutions found within the domains of existence, modifications due to GB term are relatively small, and all characteristic functions still preserve the typical behaviour they have in the pure EYMD case. We also prove analytically that the dilaton charge of any regular solution (with an exact account for the GB term) [*is equal*]{} to its ADM mass independently on the value of $\beta$. For $\beta=0$ a stronger relation holds between $g_{00}$ and the dilaton factor everywhere. We start with the following bosonic part of the heterotic string effective action in four dimensions in the Einstein frame : $$S = \frac{1}{16\pi} \, \int \;\left\{(-{\it R} + \, 2\partial_{\mu} \Phi \partial^{\mu} \Phi ) - \alpha'\exp (-2 \Phi) (F_{a\mu\nu} \, F_a^{\mu\nu} - \beta {G})\right\}\, \sqrt{-g} d^4x \; ,$$ where $\Phi$ is the dilaton, $F$ is the Yang-Mills field strength and $ G$ is the Gauss–Bonnet term which can be presented as the divergence of the topological current $$G = R_{\mu\nu\lambda\tau} R^{\mu\nu\lambda\tau} - 4 R_{\mu\nu} R^{\mu\nu} + R^2 = \nabla_{\mu} K^{\mu} \; .$$ Integrating by parts the GB term in (1) one can rewrite the action in somewhat simpler form (both in (1) and (2) we ignore surface terms which are not relevant for the present analysis) $$S = \frac{1}{16\pi} \, \int \, \,\left\{((-{\it R} + \, 2\partial_{\mu} \Phi \partial^{\mu} \Phi ) - \alpha' e^{-2 \Phi} (F_{a\mu\nu} \, F_a^{\mu\nu} - 2 \beta (\partial_{\mu} \Phi) { K^{\mu}})\right\}\,\sqrt{-g} d^4x \; .$$ We parametrize the metric of the static spherically symmetric spacetime as $$ds^2 = W dt^2 - \frac{dr^2}{w} - R^2 (d\theta^2 + \sin^2 \theta d\phi^2)\; ,$$ where $ W = w \sigma^2$ and all functions depend on the single variable $r$. In this case only the radial component of the topological current is relevant $$K^r = \frac{4 (w \sigma^2)' (wR'^2 - 1)}{R^2\sigma^2} \;.$$ Here and below primes mean derivatives with respect to $r$. A magnetic part of the static spherically symmetric $SU(2)$ Yang–Mills connection can be expressed in terms of the single function of the radial variable $f(r)$ $$A^a_{\mu} T_a dx^\mu = (f-1)(L_\phi d\theta - L_\theta\sin \theta d\phi)\, ,$$ where $L_r =T_a n^a$, $L_\theta=\partial_\theta L_r$, $L_\phi= (\sin \theta)^{-1}\partial_\phi L_r\;,$ $n^a = (\sin \theta \cos \phi, \sin \theta \sin \phi, \cos \theta)$ is the unit vector and $T_a$ are normalized Hermitean generators of the $SU(2)$ group. Integrating out the angular variables in (2) and eliminating some total derivatives one obtains the following reduced effective action $$S = \int dt dr \left(L_{g} + L_{m} + L_{ GB} \right)\; ,$$ where $$L_{g} = \frac{\sigma}{2} \left( R'(wR)' + 1\right) + w \sigma'RR'\; ,$$ is the gravitational part, $$L_{m} = - \frac{1}{2} wR^2 \Phi'^2 - \frac{\alpha'}{2} { F} e^{-2\Phi}\; ,$$ is the matter part, $$L_{GB} = 2 \alpha' \beta \sigma^{-1} \Phi' W' (wR'^2 - 1) e^{-2\Phi}\; ,$$ is the Gauss–Bonnet contribution, and $${ F} = 2wf'^2 + \frac{(1 - f^2)^2}{R^2}\; .$$ Note that an arbitrary rescaling of the slope parameter $\alpha' \rightarrow k \alpha'$ together with the corresponding rescaling of the radial variable $r \rightarrow \sqrt{k} r$ is a symmetry transformation of the effective action (7). Choosing Planck units $\alpha'=1$ we are left with the only dimensionless parameter $\beta$. The equations of motion (including an Einstein constraint) can be obtained by direct variation of (7) over $\sigma, w, R, f, \Phi$. Then fixing the gravitational gauge as $R=r$ one finds the following set of equations $$\frac{\sigma'}{\sigma} = r\Phi'^2 + \frac{2f'^2 e^{-2\Phi}}{r} + \frac{4\beta}{r}\Big((\frac{\Phi'(w-1) e^{-2\Phi}}{\sigma})' \sigma - \frac{W' \Phi'}{\sigma^2} e^{-2\Phi} \Big)\; ,$$ $$w'\left(1 - \frac{4 \beta \Phi' (1 - 3w) e^{-2\Phi}}{r}\right) + \frac{{ F}}{r}e^{-2\Phi} + rw \Phi'^2 = \frac{(1 - w)}{r} \left(1 - 4 \beta w (e^{-2\Phi})''\right)\; ,$$ $$\frac{1}{2}\left(\frac{W'}{\sigma}\right)' r + \left(\frac{W}{\sigma}\right)' + \sigma \left(wr\Phi'^2 -\frac{(1 - f^2)^2}{r^3} e^{-2\Phi}\right) + 4 \beta \left(\frac{W'w \Phi'e^{-2\Phi}}{\sigma}\right)' = 0 \; ,$$ $$\left(w \sigma f' e^{-2\Phi}\right)' + \frac{\sigma f (1 - f^2) e^{-2\Phi}}{r^2} = 0\; ,$$ $$\left(\sigma r^2 w \Phi'\right)' + \sigma { F} e^{-2\Phi} + 2 \beta \left(\frac{W'(1 - w)}{\sigma}\right)' e^{-2\Phi} = 0\; .$$ It is useful also to compute an effective energy density as it enters the standard Einstein equations with account for the GB term $$2 T^0_0 = w \Phi'^2 + \frac{{ F}}{r^2} e^{-2\Phi} + \frac{4 \beta}{r^2}\left\{2w(w -1)(\Phi'' - 2\Phi'^2) + \Phi' w'(3w - 1)\right\} e^{-2\Phi}\; .$$For $\beta=0$ the system reduces to that of [@dg1] and the corresponding solutions exibit typical BK structure of the YM function: solutions start from $f=\pm 1$ and goes asymptotically to $\mp 1$ either monotonically $(N=1)$ or after $N-1$ oscillations around zero. As a first step of the analysis we calculate the GB term and the corresponding density $\sqrt{-g}G$ substituting the sphaleron solutions found without an account for the GB term. Numerical results are shown on the Fig. 1. One can see that the value of GB term increases with growing number of nodes of the YM function. It can be anticipated that its influence on the sphaleron solutions will increase for higher $N$. We have also calculated the effective energy density (17) for the background EYMD solutions. Fig. 2 clearly shows that relative contribution of the GB term for $\beta=1$ is not small. This presumably invalidate any attempt to treat the GB term perturbatively, so we are faced with the problem of constructing numerical solutions to the system (12)–(16). To define the ADM mass $M$ and the dilaton charge $D$ one writes asymptotic expansions for $W$ $$W = 1 - \frac{2M}{r} - \frac{2D^2 M}{r^3} + O(\frac{1}{r^4})\; ,$$and the dilaton $$\Phi = \Phi_{\infty} + \frac{D}{r} + \frac{DM}{r^2} + \frac{8M^2 D - D^3}{6r^3} + O(\frac{1}{r^4})\; .$$The corresponding expansion of $\sigma$ reads $$\sigma = 1 - \frac{D^2}{2r^2} - \frac{4D^2 M}{3r^3} + O(\frac{1}{r^4})\; .$$ To ensure asymptotic flatness it is sufficient (as it is for $\beta=0$) to have for the Yang–Mills function $f$ $$f = \pm 1 + O(\frac{1}{r})\; .$$Clearly, GB–induced terms do not influence the leading behaviour of solutions near infinity. In contrary, an expansion of regular solutions near the origin [*is*]{} affected by the curvature terms. From the system (12)–(16) one finds $$\begin{aligned} f&=& -1 + br^2 + O(r^4)\,,\\ \Phi&=& \Phi_0 + \Phi_2 r^2 + O(r^4)\,,\\ \sigma&=&\sigma_0 + \sigma_2 r^2 + O(r^4)\,,\\ W&=&W_0 + W_2 r^2 + O(r^4)\,, \end{aligned}$$ or in terms of $w$: $$w = 1 + w_2 r^2 + O(r^4)\,,$$ where the following relations hold $$W_0 = \sigma_0^2, \quad W_2 = 2 \sigma_0 \sigma_2 + w_2 \sigma_0^2 .$$ Let us prove that for any regular solution to the system (12)–(16) (if exists), the ADM mass $M$ is exactly equal to the dilaton charge $D$. Combining Eqs. (12), (14) and (16), after some rearrangment one can find the following identity $$\left(2 \sigma r^2 w \Phi' + \frac{W' r^2}{\sigma}\right)' = 4 \beta Q'\;,$$ where $$Q = \sigma^{-1} \left\{(w - 1)(W' + 2W \Phi') - 2r \Phi w W\right\}\; .$$ Integrating this relation over the semiaxis with account for (18)–(20) on gets $$\left.\left(\frac{W}{\sigma}r^2 \Phi' + \frac{W' r^2}{2\sigma}\right)\right|^{\infty}_{0}\equiv M - D = 2 \beta \left[Q(\infty) - Q(0)\right]\; .$$ Now from the expansions (18)–(21) and (22)–(26) it can be found that both above boundary values of $Q$ are equal to zero, what proves the exact equality $M=D$. Remarkably, this property of regular EYMD solutions observed first in [@dg1], remains true with account for the GB term for any value of $\beta$. There is an important difference, however. In the case $\beta=0$ a stronger identity $$W = \exp (-2\Phi)$$holds, which is similar to the well–known relation for the extremal magnetic dilatonic black holes, where it ensures regularity of the metric in the string frame. When GB term is taken into account this is no longer true while the relation $M=D$ exibiting the validity of (31) in the asymptotic region still holds. Similarly to the system of Einstein–Yang–Mills–Dilaton equations [@dg1], [@lm], [@tm], [@oneill] without Gauss–Bonnet term there are three independent parameters in the series solutions of the system (12–16) near the origin: $b$, $\Phi_0$ and $\sigma_0$. From them the quantity $\Phi_0$ is somewhat trivial because of the symmetry of the system under a dilaton shift accompanied by suitable rescaling of the radial coordinate (if desired, $\exp (-2\Phi_0)$ may be absorbed into redefinition of parameters in (22)–(26)). However, there is a substantial complication as compared with the pure EYMD theory. In order to fulfil the system (12–16) in the first leading order, the coefficient $\Phi_2$ has to be one of the real roots of the following algebraic equation of the forth order $$\left(\Phi_2 + 2 b^2 e^{-2\Phi_0}\right) \left(1 + 16 \beta \Phi_2 e^{-2\Phi_0}\right)^3 + 32 \beta b^4 e^{-6\Phi_0} \left(1 + 8 \beta \Phi_2 e^{-2\Phi_0}\right) = 0 .$$ Once $\Phi_2$ is found, two other coefficients $w_2$ and $\sigma_2$ can be obtained as $$\begin{aligned} w_2 &=& - \frac{4b^2 e^{-2\Phi_0}}{1 + 16 \beta \Phi_2 e^{-2\Phi_0}}\,,\\ \sigma_2 &=& \frac{\sigma_0 e^{-2\Phi_0} (4b^2 + 4 \beta w_2 \Phi_2)}{1 + 16 \beta \Phi_2 e^{-2\Phi_0}}\; . \end{aligned}$$ It is convenient to regard the Eq. (32) as giving the value of $\Phi_2$ as a function of $b$, while $\Phi_0$ is fixed. In fact, a dilaton shift $$\Phi_0 \rightarrow \Phi_0 + \delta \Phi_0$$ leads to a solution related with the initial one by a radial rescaling. Physically the normalization $\Phi_{\infty}=0$ is preferable since it ensures a unique mass scale for all solutions. But technically is is convenient to solve the system first by fixing $\Phi_0$ arbitrarily, say, $\Phi_0=0$. Then the rescaled solution will result from $$b \rightarrow b \exp (2\delta \Phi_0) ,\quad \Phi_2 \rightarrow \Phi_2 \exp (2\delta \Phi_0) ,\quad \sigma_0 \rightarrow \sigma_0 .$$ At the final stage of the calculation we rescaled solutions imposing the condition $\Phi_{\infty}=0$ in order to fix a unique mass scale for all of them. The numerical strategy consists in solving the system (12)–(16) starting from the series solution (22)–(26) near the origin. The crucial role is played by the parameter $b$ which shoul take a discrete sequence of values. For $\beta=0$ the solution of Eq. (32) reads $\Phi_2=-2b^2\exp (-2\Phi_0)$, and clearly this does not impose any restriction on this parameter. But for $\beta\neq 0$ it turns out that real solutions for $\Phi_2$ do not exist in some region of $b$. Hence, in addition to the problem of “quantization” of $b$ one has to ensure that $b$ belongs to region where the real roots of the Eq. (32) exist. It happens that if $\beta$ is greater than some ($N$–dependent) critical value $\beta_N$, the allowed region of $b$ does not contain those quantized values for which regular solutions exist. Only for $\beta < \beta_N$ regular solutions exist and exibit behaviour similar to that of the EYMD solutions. Real roots of the algebraic equation (32) form two branches as shown on the Fig. 3a,b in terms of the quantities ${\tilde b}=b\exp (-2\Phi_0),\, {\tilde \Phi_2}=\Phi_2 \exp (-2\Phi_0)$. For roots from the second branch (Fig. 3b) we didn’t find any solution, they seem to correspond to $b$ outside the above quantization domain. Note that this branch does not contain the EYMD root corresponding to $\beta=0$. The first branch 3a [*has*]{} a solution for $\beta=0$, while for any $\beta\neq 0$ there are two negative solutions with absolute values $\Phi_2^{max}(b)$ and $\Phi_2^{min}(b)$. From these two, it is just the second one, $\Phi_2^{min}(b)$, which has the limiting value $\Phi_2=-2b^2\exp (-2\Phi_0)$ when $\beta \rightarrow 0$. No regular solutions to the system (12)–(16) corresponding to $\Phi^{max}(b)$ were found neither. Starting with the known $\beta=0$, $N=1$ EYMD solution [@dg1] we increased gradually the value of $\beta$ searching for the desired quantized $b$ related to $\Phi_2^{min}(b)$. Numerical integration of the system (12)–(16) was done using the Runge–Kutta fourth order scheme. The values of the parameters for $N=1$ case, found numerically for some $\beta$ together with the corresponding ADM mass $M=D$ are given in the Table 1. The solutions were rescaled to ensure $\Phi_{\infty}=0$. **Table 1. N=1.** $\beta$ $b$ $\Phi_0$ $\sigma _0$ $\Phi_2^{min}$ $\Phi_2^{max}$ $M=D$ $w_2 $ --------- -------- ---------- ------------- ---------------- ---------------- ------- --------- 0. 1.073 0.9311 0.3936 -0.3576 — 0.578 -0.7153 0.1000 1.026 0.9199 0.3840 -0.3523 -3.390 0.573 -0.7344 0.2000 0.9866 0.9122 0.3744 -0.3566 -1.475 0.568 -0.7697 0.3000 0.9619 0.9120 0.3597 -0.3833 -0.8376 0.563 -0.8496 0.3700 0.9657 0.9231 0.3421 -0.4938 -0.5198 0.560 -1.0933 One can observe that with increasing $\beta$ two real roots $\Phi_2^{min}(b)$ and $\Phi_2^{max}(b)$ converge and merge together for a limiting value $\beta_1$ approximately equal to $0.37$. For $\beta>\beta_1$ there are no such $b$ which could generate asymptotically flat solutions with $N=1$ compatible with the existence of the real root $\Phi_2(b)$ of the Eq. 32. Similar situation was encountered for higher–$N$ solutions. Numerical results for $N=2$ and $N=3$ are presented in the Tables 2, 3. Figures 4–8 depict the corresponding numerical curves for some values of $\beta$ and $N$. **Table 2. N=2.** $\beta$ $b$ $\Phi_0$ $\sigma _0$ $\Phi_2^{min}$ $\Phi_2^{max}$ $M=D$ $w_2$ --------- -------- ---------- ------------- ---------------- ---------------- ------- -------- 0. 8.3612 1.7923 0.1665 -3.8796 — 0.685 -7.760 0.1000 7.1902 1.7481 0.1529 -3.5165 -15.982 0.673 -7.558 0.2000 6.4017 1.7297 0.1370 -3.6597 -5.7461 0.660 -8.161 0.2208 6.3344 1.7343 0.1320 -4.2904 -4.3127 0.657 -9.478 **Table 3. N=3.** $\beta$ $b$ $\Phi_0$ $\sigma _0$ $\Phi_2^{min}$ $\Phi_2^{max}$ $M=D$ $w_2$ --------- --------- ---------- ------------- ---------------- ---------------- -------- --------- 0. 53.8351 2.6920 0.0678 -26.600 — 0.7042 -53.202 0.2000 36.5453 2.5930 0.0504 -22.348 -30.543 0.6744 -49.817 0.21117 36.2683 2.5956 0.0492 -25.151 -25.212 0.6726 -55.558 Note, that the numerical values of ADM mass/dilaton charge monotonically decrease with growing $\beta$ for each fixed $N$. Also, it can be observed that the limiting values of $\beta_{N}$ decrease with the increasing $N$: ($\beta_{1}=0.37, \beta_{2}=0.22, \beta_{3}=0.21, ...$). It can be anticipated that $\beta_{N}$ has a limiting value $\beta_{\infty} $ as $N \rightarrow \infty$, which would presumably give an absolut bound of the existence of static spherically symmetric regular EYMD–Gauss–Bonnet solutions. It is also interesting to note, that although the contribution of the GB terms to the energy density for $\beta$ close to $\beta_N$ is not small (as it is shown on Fig. 5), the behaviour of $f$ and metric functions is very similar to that of pure EYMD solutions. It has also to be noted that, when the limiting value $\beta_{N}$ is approached, neither singularities no other numerical problems arise; so the only reason for the absence solutions when $\beta$ exceeds the above critaical value is an intrinsic incompatibility of the series expansion near the origin. We conclude with the following remarks. When Gauss–Bonnet term is included, the total number of derivatives in the system of equations increases, as well as the dergee of its non–linearity. However, in a limited region of the numerical factor $\beta$ the behaviour of solutions remains qualitatively the same as in the pure EYMD case. Moreover, the remarkable equality of the ADM mass to the dilaton charge remains unaffected by the GB term for any $\beta$. However it is likely that EYMD sphalerons are destroyed by the Gauss–Bonnet term for sufficiently large values of $\beta$. The most persistent is the $N=1$ solution, which exists up to $\beta=0.37$. Higher $N$ solutions cease to exist for lower $\beta$, the limiting value is likely to be of the order of 0.2. This work was supported in part by the ISF Grant M79000 and by the Russian Foundation for Fundamental Research Grant 93–02–16977. [91]{} R. Bartnik and J. McKinnon, Phys. Rev. Lett. [**61**]{} (1988) 141. D.V. Gal’tsov, M.S. Volkov, Phys. Lett. [**B 273**]{} (1991) 255. M.S. Volkov and D.V. Gal’tsov, Phys. Lett. [**B 341**]{} (1995) 279. N. Straumann and Z. Zhou, Phys. Lett. [**B 237**]{} (1990) 35; [**B 243**]{} (1990), 33. G.W. Gibbons, A.R. Steif, Phys. Lett. [**B 314**]{} (1993) 13; M.S. Volkov, Phys. Lett. [**B 328**]{} (1994) 89, [**B 334**]{} (1994) 40. E.E. Donets and D.V. Gal’tsov, Phys. Lett. [**B 302**]{} (1993) 411. G. Lavrelashvili and D. Maison, Nucl.Phys. [**B 410**]{} (1993), 407. K. Torii, K. Maeda, Phys. Rev. [**D 48**]{} (1993), 1643. C.M. O’Neill, Phys.Rev. [**D 50**]{} (1994) 865. E.E. Donets and D.V. Gal’tsov, Phys. Lett. [**B 312**]{} (1993) 392. S. Mignemi, N.R. Stewart, Charged Black Holes in Effective String Theory, Preprint CNRS/URA 769 (1992); Makoto Natsuume, Higher Order Corrections to the GHS String Black Holes, Preprint NSF–ITP–94–66, hep-th/9406079. Fig. 1. GB term (B) and GB density (A), calculated for pure EYMD $N=1$ solutions [@dg1], curves (C) and (D) – GB density for $N=2,3$ EYMD solutions. Fig. 2. Contributions to the energy density $r^2 * T^0_0$, from YMD ($\beta=0$) and GB parts ($\beta=1$) calculated using EYMD solutions : (A): $N=1$, YMD; (B): $N=1$, GB; (C): $N=2$, YMD; (D): $N=2$, GB. Fig. 3a,b. Real roots of Eq. 32 (two different branches), $\beta=0.1, 0.2, 0.37, 0.5,1$ in terms of ${\tilde b}=b\exp (-2\Phi_0),\, {\tilde \Phi_2}=\Phi_2 \exp (-2\Phi_0)$. Fig. 4. “Gauss–Bonnet” mass distribution (contribution to ADM mass from $\beta$ –dependent terms) for solutions with $\beta=0.2$, $N=1,2,3$. Fig. 5. Energy density for $N=3$, $\beta=0.2$. (A): total energy density; (B): contribution from $\beta$ -independent terms; (C): GB contribution. Fig. 6. Yang-Mills function $f$ for $N=1,2,3$. Solid lines: solutions with GB term ($\beta=0.2$), dashed lines: purely EYMD solutions Fig. 7. Metric function $W=g_{00}$ (dashed lines) and $\exp (-2\Phi)$ (solid lines) for $N=1,2,3$, $\beta=0.2$. Fig. 8. Metric function $\sigma$ for $N=1,2,3$. Solid lines: solutions with GB term ($\beta=0.2$); dashed lines: purely EYMD solutions ($\beta=0$).

ArXiv

--- abstract: 'We propose a set of scanning tunneling microscopy experiments in which the surface of superconductor is scanned by a *superconducting* tip. Potential capabilities of such experimental setup are discussed. Most important anticipated results of such an experiment include the position-resolved measurement of the superconducting order parameter and the possibility to determine the nature of the secondary component of the order parameter at the surface. The theoretical description based on the tunneling Hamiltonian formalism is presented.' address: | $^1$Theoretical Physics, Department of Physics, Royal Institute of Technology, SE-10044 Stockholm, Sweden\ $^2$Theoretical Division, Los Alamos National Laboratory, Los Alamos, NM 87545 author: - 'Jurij Šmakov$^{1,2}$[@smakov], Ivar Martin$^2$ and Alexander V. Balatsky$^2$' date: Printed title: 'Josephson scanning tunneling microscopy.' --- 0.4pt 0.4pt 0.4pt 9.75in -25mm [2]{} Introduction ============ Recent technological advances allowed Pan *et al.* to conduct very low temperature scanning tunneling microscopy (STM) experiments reaching temperatures as low as 220 mK[@davis1] and making possible the direct imaging of the surface of high-$T_c$ superconductors (SC) in a superconducting state with high spatial and energy resolution.[@davis2] The Pt/Ir tips used in these experiments were not superconducting, but the same group has reported earlier,[@davis3] that they were able to obtain the atomic resolution images of superconducting NbSe$_2$ with the *superconducting* atomically sharp Nb tip in the quasiparticle tunneling regime. The idea of STM experiment in which the Josephson effect would be observed between the tip and the surface was also suggested.[@davis3] If both the tip and the surface are superconducting and the temperature is low enough to have a well-defined phase difference between the superconductors, the tunneling contact may be considered a Josephson junction with tunable parameters (hence the term “Josephson STM” or JSTM). The phase-dependent supercurrent will not be washed out by thermal fluctuations if the temperature is sufficiently low, $$\label{cond} I_c\Phi_0 \gtrsim 2\pi k_BT,$$ where $I_c$ is the critical Josephson current and $\Phi_0=h/2e$ is the unit magnetic flux. Assuming that the Josephson current can be roughly estimated as $I=\sigma\Delta/e$,[@baratoff] where $\Delta$ is the magnitude of the superconducting order parameter (OP) and $\sigma$ is the normal state conductance of the junction, Eq. (\[cond\]) translates into the condition imposed on the normal state resistance: $$\label{res} R \lesssim \frac{\Delta}{2 k_BT}R_0.$$ Here $R_0=\hbar/e^2\approx 4\,\textrm{k}\Omega$ is the resistance quantum. For conventional superconductors ($\Delta\sim 2\,\textrm{meV}$) and temperature $T\sim 0.1\,\textrm{K}$ the junction resistance should not exceed $\sim 500\,\textrm{K}\Omega$ for the Josephson current to be observable. For higher gap values, such as in high-$T_c$ materials, we estimate $R\lesssim 7.5\,\textrm{M}\Omega$. Since the temperature enters into criteria (\[res\]) only via ratio $\Delta/kT$, it might be easier to observe Josephson current using high-$T_c$ material with a larger value of $\Delta$ as a tip rather than to lower temperature by an order of magnitude. Typical resistances of STM experiments are at least by an order of magnitude larger, than the estimated upper bounds for Josephson effect. Thus, Josephson supercurrent is difficult to observe in a setup with an atomically sharp superconducting tip. Therefore we propose the use of a tip with a finite tunneling area. While it will be impossible to achieve the atomic-scale spatial resolution, to capture the microscopic variations of superconducting OP only the resolution comparable to the coherence length of SC is needed. This characteristic length is about 20 Å for the high-$T_c$ materials and a few thousand Å for conventional superconductors. Also, the finite tunneling area will lead to relatively big capacitance of the junction, minimizing the charging energy and making the effects of Coulomb blockade negligible. Since the supercurrent depends on the magnitudes of the superconducting OP on both sides of the junction,[@baratoff] it should be possible to measure the microscopic variations in OP depending on the position of the tip. Such measurement can provide a direct insight into the microscopic nature of the superconducting state. Another potential application of the JSTM is the probing of the *symmetry* of superconducting OP. It is known that a perturbation, such as an applied magnetic field or a surface, can induce a secondary component of the superconducting OP. The type of the secondary component can be determined from the modification of the Josephson current. In case when both the tip and the surface are unconventional superconductors, like high-$T_c$ cuprates, the OP is angle-dependent and the variations in the tunneling supercurrent may be caused by changing the mutual orientation, providing the direct information about the symmetry of OP. Finally, the JSTM could provide information about the nature of the gap in high-$T_c$ superconductors. It is believed, that the normal state of the underdoped high-$T_c$ cuprates is a pseudogap (PG) state with partially gapped Fermi surface, from which superconductivity emerges as the temperature is lowered through $T_c$.[@timusk] Currently there is no general agreement on the origin of PG state. Some models attribute PG to superconducting phase fluctuations above $T_c$,[@emery] others - to a competing non-superconducting order parameter.[@chak] Since JSTM is sensitive only to the superconducting OP, the measurements of the critical Josephson current on differently doped samples can shed some light on the origin of PG. The rest of the paper is organized as follows: in Section \[descr\] we describe the tunneling Hamiltonian formalism used in the calculations and two practically important cases of JSTM, followed by proposed experimental setup (Section \[setup\]) and summary (Section \[summary\]). Theoretical description {#descr} ======================= We base our theoretical description of JSTM on the tunneling Hamiltonian formalism.[@baratoff; @cohen] In this formalism, time-dependent Josephson current through the tunneling contact is expressed as[@mahan] $$\label{current} I_J(t)=2e\,\textrm{Im}[\exp(-2ieVt/\hbar)\Phi_{\textrm{ret}}(eV)],$$ where $V$ is the applied voltage and $\Phi_{\textrm{ret}}(eV)$ is the voltage-dependent retarded correlation function, which in the limit of the zero temperature has the form $$\label{rcf} \Phi_{\textrm{ret}}(eV)=\frac{1}{2}\sum_{\mathbf{k}\mathbf{p}} T_{\mathbf{k}\mathbf{p}}T_{-\mathbf{k},-\mathbf{p}} \frac{\Delta_{\mathbf{k}}^{*}\Delta_{\mathbf{p}}}{E_{\mathbf{k}}E_{\mathbf{p}}} \times$$ $$\times\left(\frac{1}{eV+E_{\mathbf{k}}+E_{\mathbf{p}}+i\delta}- \frac{1}{eV-E_{\mathbf{k}}-E_{\mathbf{p}}+i\delta} \right).$$ Here indices $\mathbf{k}$ and $\mathbf{p}$ refer to the quasiparticle momenta of the tip and surface, respectively, $T_{\mathbf{k}\mathbf{p}}$ is the tunneling matrix element, $\Delta_{\mathbf{k}}$ and $\Delta_{\mathbf{p}}$ are the superconducting order parameters, $E_{\mathbf{k}}$ and $E_{\mathbf{p}}$ are the excitation energies, and the infinitesimal quantity $i\delta$ ensures the convergence of the integral. Using Eqs. (\[current\]) and (\[rcf\]) we can calculate the Josephson current for a tunneling contact of two superconductors. Unique feature of the Josephson effect is that below a certain threshold value, the dc supercurrent can flow without any voltage drop on the contact. This threshold value – critical Josephson current – is easily measured, because when the value of the current reaches the critical value, the finite voltage drop abruptly appears at the contact. If the analytical expression for the critical current is known, its measurements can provide the information about the magnitudes of OPs, its symmetry and possibly other quantities of interest. $s$-$d$ junction {#s-d} ---------------- First we consider the case when the STM tip is a conventional $s$-wave SC (superconducting metal, like Nb) with momentum-independent real order parameter $\Delta_{\mathbf{k}}\equiv\Delta_1$ and the surface under study is the $a$-$b$ plane of a $d$-wave unconventional SC (e.g., YBCO or BSCCO). It is generally agreed, that perturbation of the unconventional $d$-wave superconductors, like magnetic field or the distortion of crystal lattice caused by presence of the surface, induces a secondary component of the superconducting OP.[@balatsky; @fogel; @kleiner; @kouz] Tunneling between pure $s$- and $d$-wave superconductors is expected to be zero by symmetry arguments, at least if one does not take into account the higher order processes.[@tanaka] Experimental evidence exists,[@kleiner; @kouz; @greene] that the secondary component in high-$T_c$ cuprates possesses $s$-wave symmetry. In order to determine the dependence of the critical Josephson current both on the magnitude and the phase of the secondary OP component, we have chosen the following form for the superconducting OP on the surface: $$\Delta_{\mathbf{p}}\equiv \Delta_2 =\Delta^{(d)}_2\cos 2\varphi+ e^{i\alpha}\Delta^{(s)}_2.$$ Here $\Delta^{(d)}_2$ and $\Delta^{(s)}_2$ are the magnitudes of the primary and secondary components (respectively) and $\varphi$ is the azimuthal angle corresponding to momentum $\mathbf{p}$ in the coordinate system associated with the surface plane. By varying angle $\alpha$ the secondary component can be assigned an arbitrary phase: In particular $\alpha=0$ corresponds to pure $d+s$ symmetry and $\alpha=\pi/2$, to $d+is$. The excitation energies $E_{\mathbf{k}}$ and $E_{\mathbf{p}}$ are given by familiar relations $$E_{\mathbf{k}}=\sqrt{\xi_{\mathbf{k}}^2+\Delta_1^2}$$ and $$E_{\mathbf{p}}=\sqrt{\xi_{\mathbf{p}}^2+|\Delta_2|^2},$$ where $\xi_{\mathbf{k}}$ and $\xi_{\mathbf{p}}$ are the single-particle energies, measured with respect to corresponding chemical potentials. We assume here that the tunneling matrix element is independent of momenta, i.e. $$T_{\mathbf{k}\mathbf{p}}T_{-\mathbf{k},-\mathbf{p}}=|T|^2e^{i\phi}.$$ Here $|T|^2$ is a constant, depending only on the geometry of the tunneling contact (it drops exponentially with increasing distance between tip and the surface) and $\phi$ is the phase difference, giving rise to Josephson supercurrent. In order to calculate the sums in Eq. (\[rcf\]), we convert them into integrals over the energies $\xi_{\mathbf{k}}$, $\xi_{\mathbf{p}}$ and the azimuthal angle $\varphi$. In the case of zero applied voltage, the integral with respect to energies can be calculated using the result due to Ambegaokar and Baratoff[@baratoff] to yield $$\label{rcfv0} \Phi_{\textrm{ret}}(0)=\frac{\sigma_0}{2\pi e^2}e^{i\phi}\int_0^{2\pi} \frac{\Delta_1\Delta_2}{\Delta_1+|\Delta_2|}K\left( \frac{|\Delta_1-|\Delta_2||}{\Delta_1+|\Delta_2|}\right)\,d\varphi$$ Here $K(x)$ is the complete elliptic integral of the first kind and $\sigma_0$ is the conductance of the tunneling contact in the normal state[@mahan] $$\sigma_0=4\pi e^2 N_1N_2|T|^2$$ with $N_1$ and $N_2$ being the densities of states at the Fermi levels of tip and surface respectively. The critical Josephson current in this case is $$I_c=I_c^0|\chi(\gamma,\kappa,\alpha)|,$$ where $\gamma=\Delta^{(d)}_2/\Delta_1$ and $\kappa=\Delta^{(s)}_2/\Delta_1$ are dimensionless parameters, $I_c^0=\Delta_1\sigma_0/\pi e$ and $\chi(\gamma,\kappa,\alpha)$ is the value of integral in Eq. (\[rcfv0\]). Assuming, that $\Delta_1$ is fixed, we have studied numerically the dependence of the critical current on $\gamma$, $\kappa$ and $\alpha$ for realistic parameter ranges. The dependence of the critical Josephson current for $\alpha=0$ (corresponds to pure $d+s$ symmetry) on $\gamma$ and $\kappa$ is shown on Fig. \[s-d-alpha0\]. =3.0in Another important case is $\alpha=\pi/2$, which corresponds to $d+is$ symmetry of superconducting OP. The results for this case are presented in Fig. \[s-d-alphapi2\]. =3.0in Useful information can also be extracted from the observation of quasiparticle current between superconductors at a finite temperature. The current can be written as[@mahan; @tinkham] ------------------------------------------------------------------------ ------------------------------------------------------------------------ $$I=\frac{\sigma_0}{2\pi e}\int_0^{2\pi}\,d\varphi\int_{-\infty}^{+\infty} \frac{|E|}{\sqrt{E^2-\Delta_1^2}}\frac{|E+eV|}{\sqrt{(E+eV)^2-|\Delta_2|^2}} [f(E)-f(E+eV)]\,dE,$$ [ ]{} [2]{}where $f(E)$ is a Fermi-Dirac distribution function and the regions $|E|<\Delta_1$ and $|E+eV|<|\Delta_2|$ are excluded from the integration. We have numerically calculated current-voltage characteristic for a realistic set of parameters: $\Delta_1/kT=2$, $\Delta_2^{(d)}/\Delta_1=10$ and $\Delta_2^{(s)}/\Delta_2^{(d)}=0.025$. It is known,[@tinkham] that when the tunneling contact of two $s$-wave superconductors with different constant real order parameters $\Delta_1$ and $\Delta_2$ is considered, the quasiparticle current-voltage characteristic possesses two significant features: a peak at the voltage $|\Delta_1-\Delta_2|$ and an abrupt increase in current at voltage $\Delta_1+\Delta_2$, corresponding to transition to Ohmic regime. We find that the position of the peak and the Ohmic transition depends on the symmetry of the superconducting OP. In particular (as seen in Fig. \[quasi\]) the current is sharply peaked at the voltage $\Delta_2^{(d)}-\Delta_1$ for $\alpha=\pi/2$ ($d$+$is$ symmetry), while for $\alpha=0$ the peak is at the voltage $\Delta_2^{(d)}+\Delta_2^{(s)}- \Delta_1$. The transition to Ohmic regime is also shifted to higher voltages by the value of $\Delta_2^{(s)}$. Thus, such a measurement can complement the measurements of Josephson critical current to provide additional information about the nature and magnitude of superconducting OP on the surface. =3.0in $d$-$d$ junction {#d-d} ---------------- The $d$-wave nature of a pair requires finite size tunneling area. If the tunneling occurs from only one site on the tip it would imply effective momentum averaging of pair wavefunction and would yield zero Josephson current for $d$-wave pairs. Thus it is desirable to have a finite region of the tip where tunneling can occur. In this case, one looses spatial resolution, compared to conventional tips. However the gain is the phase coherent current between the tip and the surface. While the idea of using the high-$T_c$ cuprates as a material of STM tip has not yet been reported in the literature, the fabrication of relatively small (of the order 100 Å) flat high-$T_c$ probes, suitable for use as tips, is feasible. Theory based on tunneling Hamiltonian predicts, that if the tunneling matrix element is momentum-independent, $c$-axis Josephson current at zero temperature is identically zero for superconductors with pure $d$-wave symmetry of OP. It is possible to argue that, as discussed in previous section, the presence of the surface will induce the secondary OP component, which can lead to a nonzero current. However, since the magnitude of secondary component is usually much smaller than the magnitude of primary one, this induction will produce a second order effect and we assume that it can be neglected. =3.0in Here we discuss a different scenario, supported by recent theoretical and experimental work of Latyshev *et al.*[@latysh] In their work, the *intrinsic* $c$-axis quasiparticle transport between layers of BSCCO was studied. It was found that interlayer transport is predominantly momentum-conserving, i.e. the in-plane components of quasiparticle momentum are conserved in the tunneling process. It seems natural to make an assumption, that in-plane momentum is also conserved in case of Josephson supercurrent. As we show below, adoption of this hypothesis leads to non-zero current even in the case of pure $d$-wave symmetry. We consider an $a$-$b$ plane tunneling contact of two $d$-wave superconductors. We allow the magnitudes of the superconducting OP to be different. Because of the angular dependence of the superconducting OP, the Josephson current will depend on the mutual orientation of two superconductors and we introduce angle $\theta$ (angle between $a$-axes of superconductors) to describe this orientation. The problem is essentially two-dimensional and the order parameters of the tip $\Delta_1$ and surface $\Delta_2$ may be written as $$\Delta_1=\Delta^{(0)}_1\cos 2\varphi$$ and $$\Delta_2=\Delta^{(0)}_2\cos 2(\varphi+\theta).$$ Tunneling matrix element has the form $$T_{\mathbf{k}\mathbf{p}}T_{-\mathbf{k},-\mathbf{p}}=|T|^2e^{i\phi} \delta(\mathbf{k}-\mathbf{p}),$$ where $\mathbf{k}$ and $\mathbf{p}$ are two-dimensional vectors. Again, we change the sums in Eq. (\[rcf\]) into integrals, which can be calculated numerically. In the case of zero applied voltage and zero rotation angle $\theta$ it is possible to calculate the integral analytically and find that the critical Josephson current in this limit is $$I_c=I_c^0\frac{\gamma\ln \gamma^2}{\gamma^2-1},$$ where $\gamma=\Delta^{(0)}_1/\Delta^{(0)}_2$, assuming $\Delta^{(0)}_1<\Delta^{(0)}_2$, and $I_c^0$ is constant current independent of $\gamma$. This dependence is presented in Fig. \[gaps\]. $\gamma=0$ corresponds to absence of the superconductivity in the tip, which leads to dissapearance of the Josephson current. We have also studied the dependence of the critical current on the rotation angle $\theta$ at zero applied voltage (Fig. \[rot\]). As expected, it has a maximum at $\theta=0$, monotonically decreases with increasing $\theta$ and becomes zero at $\theta=\pi/4$. For the angles bigger than $\pi/4$, the critical current can be obtained as $$I_c\left(\frac{\pi}{4}+\theta\right)=-I_c\left(\frac{\pi}{4}-\theta\right),$$ i.e. for $\theta>\pi/4$ the current changes direction, but the curve remains symmetric with respect to the point $\theta=\pi/4$. It is notable, that since the curve is symmetric with respect to $\theta=0$, there is a kink in the angle dependence at $\theta=0$. It can be seen from Fig. \[rot\], that the decrease is almost linear with $\theta$ and linear approximation becomes better as $\gamma$ is reduced. For practical purposes an approximate formula $$\label{linear} I_c=I_c^0\frac{\gamma\ln\gamma^2}{\gamma^2-1}\left(1-\frac{4}{\pi}\theta \right)$$ may be used. =3.0in It is worth mentioning that when the angle $\theta$ is non-zero, the denominator in Eq. (\[rcf\]) has a positive lower bound. This means that for voltages much smaller than this lower bound, the integrand may be expanded in powers of voltage. It turns out that all first order terms in voltage exactly cancel out and dependence on voltage enters only as higher order corrections (proportional to the square of the voltage), so the zero-voltage behavior is not altered significantly. The range of applicability of this approximation depends on actual values of $\gamma$ and $\theta$. Experimental setup {#setup} ================== The sketches of experimental setup for $s$-$d$ and $d$-$d$ tunneling are shown in Fig. \[s-dexp\] and Fig. \[d-dexp\] respectively. In both cases, superconducting tip and surface form an electric circuit together with current (voltage) source and measuring equipment. =3.0in In $s$-$d$ case, measurements can be performed in two stages. First, by driving a large current through the tunneling contact, the superconductivity is suppressed and the normal state conductance of the tunneling junction at the specific position on the surface is measured. Next, the measurement of the critical Josephson current (by starting with zero current and increasing it until a voltage drop appears on the contact) is performed. Using this information, a conclusion can be made about the magnitudes of the primary and/or secondary OP components on the surface. By repeating the measurement, a surface map of variations in order parameter may be obtained. Both the surface, formed by $a$-$b$ plane and the one, parallel to $c$-axis can be studied. =3.0in In $d$-$d$ case, the apparatus must be designed in such a way that the tip can be rotated around vertical axis. Measuring the Josephson current as a function of rotation angle, the information about the symmetry of OP can be extracted. If Josephson current has nodes, separated by $\Delta\theta=\pi/2$, a conclusion can be made that the non-zero Josephson current is due to coherent tunneling and superconductors possess pure $d$-wave symmetry. If, on the other hand, current never becomes zero or becomes zero with different periodicity, then the secondary component of the OP must be a primary reason for the existence of the current. Finally, the experiments aimed at establishing the nature of the pseudogap should be conducted with differently doped tip and surface. For example, measuring critical Josephson current in a setup with underdoped tip and overdoped surface and comparing it with measurements on identically underdoped (or overdoped) contacts will provide information about the magnitude of the pseudogap and whether or not it is of superconducting origin. Detailed discussion of such an experiment will be given elsewhere. Summary ======= In summary, we have a proposed a series of STM experiments of a superconducting surface with a superconducting tip. Treating the tunneling contact as a Josephson junction with tunable parameters, we have developed a theoretical description of these experiments, calculating the dependence of critical Josephson current on various relevant parameters. Since Josephson current is sensitive to the magnitudes of superconducting order parameters in the tip and in the surface, this new technique allows to measure the microscopic spatial variations of the OP on the surface and determine the magnitude and phase of the secondary OP component, induced by the surface. It is also possible to probe the symmetry of superconducting OP directly, exploiting the dependence of the Josephson critical current on the mutual orientation of $d$-wave superconductors. Finally, since the Josephson current is sensitive only to the superconducting gap, the new technique can provide information about the nature of the pseudogap, in particular, confirm or falsify its superconducting origin. Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful to J. C. Davis, M. J. Graf and A. Rosengren for many valuable comments. One of us (J. Š.) would like to thank Los Alamos National Laboratory for hospitality and acknowledge partial financial support from Swedish Natural Science Research Council. [99]{} Corresponding author.\ Electronic address: [[email protected]]{}. S. H. Pan, E. W. Hudson and J. C. Davis, Rev. Sci. Instrum. [**70**]{} (2), 1459 (1999). S. H. Pan *et al.*, Nature [**403**]{}, 746 (2000); S. H. Pan *et al.*, Phys. Rev. Lett. [**85**]{}, 1536 (2000). S. H. Pan, E.W. Hudson and J. C. Davis, Appl. Phys. Lett. [**73**]{}, 2992 (1998). V. Ambegaokar and A. Baratoff, Phys. Rev. Lett. [**10**]{}, 486 (1963); *ibid.* [**11**]{}, 104 (E) (1963). T. Timusk and B. Statt, Rep. Prog. Phys. [**62**]{}, 61 (1999). V. J. Emery and S. A. Kivelson, Nature [**374**]{}, 434 (1995). P. A. Lee and X. G. Wen, Phys. Rev. Lett. [**76**]{}, 503 (1996); I. Martin, G. Ortiz, A. V. Balatsky and A. R. Bishop, cond-mat/0003316; S. Chakravarty, R. B. Laughlin, D. K. Morr and C. Nayak, cond-mat/0005443. M. H. Cohen, L. M. Falicov and J. C. Phillips, Phys. Rev. Lett. [**8**]{}, 316 (1962). G. D. Mahan, *Many-Particle Physics* (Plenum Press, New York, 1990). K. Krishana, et al., Science [**277**]{}, 83 (1997); R. B. Laughlin, Phys. Rev. Lett. [**80**]{}, 5188 (1998); A. V. Balatsky, P. Kumar and J. R. Schrieffer, Phys. Rev. Lett. [**84**]{}, 4445 (2000) and references therein. M. Fogelström, D. Rainer and J. A. Sauls, Phys. Rev. Lett. [**79**]{}, 281 (1997) R. Kleiner *et al.*, Phys. Rev. Lett. [**76**]{}, 2161 (1996). K. A. Kouznetsov *et al.*, Phys. Rev. Lett. [**79**]{}, 3050 (1997). Y. Tanaka, Phys. Rev. Lett., [**72**]{}, 3871 (1994). M. Covington *et al.*, Phys. Rev. Lett. [**79**]{}, 277 (1997). M. Tinkham, *Introduction to Superconductivity* (McGraw-Hill, New York, 1996). Yu. I. Latyshev *et al.*, Phys. Rev. Lett. [**82**]{}, 5345 (1999).

ArXiv

--- bibliography: - 'DIPTBTP.bib' --- Adjunct Professor of Finance, Stuart School of Business, Illinois Institute of Technology and Principal, Market Pattern Research, Inc. Alumni Professor of Financial Modeling and Stochastic Optimization (Emeritus), University of British Columbia, Vancouver, BC, and Distinguished Visiting Research Associate, Systemic Risk Centre, London School of Economics, UK Introduction ============ [@moffitt:ziemba:2017a] show that expected returns of $10\%$-$25\%$ can be achieved under certain conditions from betting all the tickets in a lottery that pays its entire jackpot in equal shares to winning ticket holders. For many large government lotteries, this strategy of “buying the pot” is not feasible because the logistical problems are insurmountable. In the California Powerball Lottery, for example, the number of ticket combinations is over $175,000,000$ and the rules do not allow betting large numbers of combinations on single paper tickets. The Canadian 6/49 Lotto, however, has a large but manageable number of ticket combinations ($13,983,816$) and allows paper tickets that have multiple combinations. The 6/49 is played in other countries, including the UK. Here we focus on the Canadian version. The purpose of this paper is threefold: (1) to modify the *pure jackpot model* in [@moffitt:ziemba:2017a] to accomodate the irregular payout features of the Canadian 6/49 Lotto, (2) to derive conditions under which the expected return from buying the pot is positive, and (3) to discuss the implications of our findings for lottery design. Previous Work and Instances of Buying the Pot {#S:PreviousWorkAndInstancesOfBuyingThePot} ============================================= Each lottery has the following rules — players buy tickets and the winning ticket is selected using an equiprobable drawing. Those who hold the winning ticket share equally in a jackpot that consists of a carryover pot from the previous lottery plus an after tax portion the monies wagered. If there is no winner, the jackpot pool carries over to the next drawing. There can be multiple carryovers. [@moffitt:ziemba:2017a] use the following assumptions and notation to analyze the *pure jackpot model*: - Each lottery has $t$ tickets costing $\$1$ apiece. - A single winning ticket $w$, $1 \le w \le t$ is drawn from $i=1,\ldots,t$ using probabilities $p_i = 1/t$. - The syndicate buys one of each ticket for a total of $t$ tickets, and $c$ individuals (the “crowd”) independently buy one ticket apiece using probabilities $q_i$, $1 \le i \le t$. - A cash jackpot $v = a + (t + c)(1 - x)$ is awarded in equal shares to all holders of the winning ticket $w$, where $a \ge 0$ is the current carryover from the previous lottery draws, $c$ is the number of tickets bet by the crowd, and $x$ is the the (fractional) take. [@moffitt:ziemba:2017a] show the following for the pure jackpot model: 1. [*Recursion:*]{} When $t$ and $c$ are large, $q_i = 1/t$ for each $i$, and $X$ is the random number of winning tickets held by the crowd, the expected value $E\left[ \frac{n}{n+X} \right]$, $n$ an integer $\ge 1$, is to close approximation equal to $$\label{E:EVrecursion} E\left[ \frac{n}{n+X} \right] = \begin{cases} \frac{1}{\lambda(c)} \left( 1 - e^{-\lambda(c)} \right) & n=1 \\ \frac{n}{\lambda(c)} \left\{ 1 - E\left[ \frac{n-1}{n-1+X} \right] \right\} & n > 1, \end{cases}$$ where $\lambda(c) = c/t$. \[Enum:Recursion\] 2. [*Condition under which Buying the Pot has Positive Expected Return:*]{} The expected gain for a syndicate that bets one of each ticket is positive $$(a + (t + c)(1 - x)) \, E\left[ \frac{1}{1+X} \right] \, - \, t \, > 0 \label{E:FairSplitCondition}$$ provided that $a/(t + c) - x \ge 0$. Since $a/(t + c) - x$ is the after tax value of a ticket assuming the pot $a$ is fairly split, this condition implies that a syndicate earns more than a fair split of the jackpot. In a lottery with no take, the returns to the syndicate typically range between 10% and 25%. \[Enum:BTPEdge\] 3. [*Optimal Strategies:*]{} 1. The best returning strategy for the crowd consists of using $q_i = 1/t$ for each $i$. 2. Let $E_q[X/(1+X)]$ be the expectation for a crowd that bets with probability vector $q = (q_1, \ldots, q_t)'$, and let $1_t/t$ be the probability t-vector that has $1/t$ for each entry. Then if $q \ne 1_t/t$ $$E_q[X/(1+X)] < E_{1_t/t}[X/(1+X)]. \label{E:OptimalityOfCrowdProportionalBetting}$$ \[Enum:OptimalStrategy\] Several studies of lottery strategy and design have appeared in the economic literature. [@chernoff1980analysis; @Chernoff1981] studied the Massachusetts Numbers Game, proposing that playing unpopular numbers might be a winning strategy. However, the results from a test were disappointing because of learning (unpopular numbers became less unpopular) and gambler’s ruin (betting funds were exhausted). [@ziemba1986dr] carry this further and study various Canadian lotto games, their unpopular numbers and the uniformity of betting. [@10.2307/2290349] has additional discussion of this latter point and [@10.2307/3314913], [@10.2307/2290073] and [@Ziemba2008183] further analyze unpopular numbers. [@citeulike:1337256] review the behavioral evidence in efficient markets for a persistence of betting at unfavorable odds. [@RePEc:inm:ormnsc:v:38:y:1992:i:11:p:1562-1585] investigate the use of Kelly optimal wagering on unpopular tickets and find that this strategy has positive expectation, but the waiting time to achieve reliable gains with high probability is millions of years! [@10.2307/1942724] discuss behavioral bases of betting and along with [@Walker2008459], discuss design considerations for lotteries. None of these studies consider the strategy that in a short time achieves reliable gains — buying the pot. There are anecdotal accounts of successful buyings of the pot. One putative attempt involved a syndicate that tried but failed to buy all tickets. But they were lucky, having had time to bet only about $70\%$ of all tickets according to one source and $85\%$ according to another ([@NYTimes:US:BuyThePot1992]). The syndicate ostensibly bet about \$5 million and won about \$27 million. There are examples of when it was optimal to buy the pot or betting was advantageous. In June 1984 four western Canadian provinces jointly ran the *Lotto West 6/8/56*, in which players choose six numbers from a field of 56, but eight winning numbers and a bonus number are drawn. The jackpot is shared among all tickets that select six of the eight drawn, second price among all that had five of six, and other payouts to those having five of six plus the bonus, four of six or three of six. These rules make the jackpot about twelve times easier to hit than the 6/49 Lotto: 1 in $1,159,587$ (See [@ziemba1986dr]). In 1987, the provinces went their own ways, at which time the BC Lotto Corporation had about \$10 million in unclaimed prize money. Rather than donate it, they created a version of Lotto 6/8/56 to give it back on March 27, 1987. As before eight numbers were drawn from 56, but players could now choose 1, 2, 3, 4, 5 or 6 numbers on a ticket. A schedule of payouts was published for 1/1, 2/2, /3/3, 4/4, 5/5 and for 3/6, 4/6, 5/6 and 6/6. With these payouts, the expected payback on a \$1 ticket was $\$0.385$. To promote the game, the Corporation offered six tickets for the price of one, for an expected return of \$0.385 times 6, or \$2.31, a $131\%$ edge ([@drz:ColumnsOnRacing]). Ziemba and colleagues at the University of British Colombia knew that individual tickets had a positive expected return, and in a makeshift effort, they bought about 13,000 of the combinations. They made a nice return, but spent hours buying and then locating the winning tickets. Some $\$3.5$ million was paid out of the unclaimed prize fund. A game where it was optimal to buy the pot was the 5/40 Lotto played in British Colombia and Rhode Island. Ninety-one percent of the net pool went to 5/5 with a minimum shared pool of $\$150,000$ and maximum of $\$300,000$. There were small prizes for 1+, 2+, 3, 4 and 4+ where “+” means getting the sixth bonus number correct. There were $658,008$ combinations. But the jackpot that had built up slowly fell because the public viewed it as unwinnable, so it became a prime target for buying the pot; see [@drz:ColumnsOnRacing]. Rules of the 6/49 Lotto {#S:RulesOfThe649Lotto} ======================= A *ticket* in the 6/49 Lotto is a unique choice of $6$ different numbers from integers $1$ to $49$. Thus the total number of tickets is the number of combinations of $49$ things taken $6$ at a time: $$t = \binom{49}{6} = \frac{49!}{43! 6!} = 13, 983, 816. \label{Nbr-6/49-Tickets}$$ The Canadian 6/49 Lotto holds drawings twice a week and lumps together the monies wagered for purposes of awarding prizes, whose allocation is described below. On the drawing day, 6 numbers (the “*winning numbers*”) are selected equiprobably and without replacement from 1, 2, …49. Following that, a $7^{th}$ “*bonus number*” is selected. We introduce notation to describe types of prize-wining tickets. A $x$/6- ticket is one that contains exactly $x$ of the six winning numbers but does not contain the bonus number and a $x$/6+ ticket is one that contains exactly $x$ of the 6 numbers plus the bonus number. A x/6 ticket contains x of the 6 numbers, irrespective of the status of the bonus number; it is therefore a union of types x/6- and x/6+. A 5/6-, for example, contains exactly 5 of the 6 winning numbers with the other not being the bonus number, and a 5/6+ ticket contains exactly 5 of the 6 winning numbers plus the bonus number. For example, if the six numbers drawn were 46, 13, 4, 21, 38, 25 and the bonus number was 43 then ticket 1-4-20-21-32-43 would be a 2/6+ ticket because it contains 4 and 21 from the six plus the bonus number. Similarly, ticket 4-13-21-25-43-46 would be a 5/6+ ticket. Rules for the Original Lottery: 1982-2004 ----------------------------------------- Table \[Ta:PrizeAllocIn649:1982-2004\] has the initial 6/49 payout scheme (1982-2004) for 3/6, 4/6, 5/6-, 5/6+ and the *Jackpot* 6/6. The cost of a single ticket was \$1, with the lottery sponsors taking 55% of each daily *betting pool* and committing the remaining 45% (the “*prize pool*”) for player payouts. The 45% *prize pool* was allocated as follows: all 3/6 tickets were paid \$10, and the remainder was paid to holders of 4/6, 5/6-, 5/6+ and 6/6 using percentage allocation rules in Table \[Ta:PrizeAllocIn649:1982-2004\]. That game is analyzed thoroughly in [@ziemba1986dr]. See also [@10.2307/2290073]. For other analyses of such games, see [@citeulike:1337256] and [@Haigh2008481]. [lrlll]{} ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ Prize & Combinations & Probability & Allocation Rule & Type\ ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ $6/6$ & $1$ & $p_1 \sim 7.1e$-$08$ & 50% of the Pools Fund & Share\ $5/6+$ & $6$ & $p_2 \sim 4.3e$-$07$ & 15% of the Pools Fund & Share\ $5/6-$ & $252$ & $p_3 \sim 1.8e$-$05$ & 12% of the Pools Fund & Share\ $4/6$ & $13,545$ & $p_4 = 0.000969$ & 23% of the Pools Fund & Share\ $3/6$ & $246,820$ & $p_5 = 0.017650$ & $\$10$ per ticket & Fixed\ No Win & $13,723,192$ & $p_8 = 0.981362$ & Non-winner = \$0 & Returns 0\ ------------------------------------------------------------------------ \ Figure \[G:PoolShare\] shows the leveraging effect of the fixed \$10 3/6 prize when there are average numbers, popular numbers, and unpopular numbers. The impact of popular vs. unpopular numbers selected in the drawing is significant, producing a 17% versus a 36% jackpot share. The large prizes 5/6-, 5/6+ and 6/6 for unpopular numbers in the drawing are typically seven times larger than for popular ones. See examples in [@ziemba1986dr]. ![[]{data-label="G:PoolShare"}](lottery649-popularity.pdf) Rules for the Current Lottery: 9/18/2013 - ------------------------------------------- In the 6/49 Lotto, new rules were introduced in June, 2004 and again on September 18, 2013. We discuss only the latter rules. These included (1) a single ticket cost of \$3, (2) three fixed prizes, the same 3/6 paying \$10, a 2/6+ paying \$5 and a 2/6- that earns a free play at the next drawing, and (3) altered payout percentages for 4/6, 5/6-, 5/6+ and 6/6 (Table \[Ta:PrizeAllocIn6492013-\]), (4) an increase in the take from 55% to 60%, and (5) a greater allocation to 6/6 winners. The intention of these changes was to increase sales by growing jackpots faster, and creating of many small consolation prices (2/6-, 2/6+ and 3/6). This is a typical convex prize structure where most of the daily payout goes to the smallest (to make them feel that the lottery is winnable) and to the largest (to show that a huge gain can be made). Ziemba has used this in lottery consulting over the years. [@RePEc:eee:jfinec:v:13:y:1984:i:2:p:253-282] call this a “silver lining” for non-winners. We call the number of tickets bet at a drawing (twice a week in the 6/49), the *ticket pool*, contributors to which are the crowd in amount $c$ and the syndicate in amount $t$. Thus the total number of tickets bet is $c + t$. The *betting pool* $d_{\!_{BP}}$ is the total number of dollars contributed by the bettors. The *betting pool* is divided among the lottery sponsors and the bettors as follows: [*Sponsors.*]{} Sponsors (the state, the lottery organization) receive $0.60*d_{\!_{BP}}$, with the remaining $0.40*d_{\!_{BP}}$, the *prize pool*, awarded as prizes or added to the carryover pool as indicated below. The “lottery take” $0.60*d_{\!_{BP}}$ is used to cover expenses of running the lottery and to provide funds for community and government services and for donations. The lottery itself, however, is run by a non-governmental company. [*Prize Distribution.*]{} The *prize pool* has eight classes ($i = 1, 2, \ldots, 8$) of payouts grouped into four types: (A) fixed dollar (2/6+ and 3/6), (B) free play in the next lottery (2/6-), (C) payouts that split among 4/6, 5/6-, 5/6+ and 6/6 tickets the remaining *prize pool* after deductions for type (A) and (B) payouts, and (D) non-winner tickets that receive no payout. Table \[Ta:PrizeAllocIn6492013-\] details these payouts by showing in the first column the type of ticket, in the second column a notation for the number of each class determined after the random, equiprobable drawing of 6 numbers and a bonus, the third column showing the notation for the class, the fourth column the number of tickets matching a randomly drawn ticket, the fifth column having the probability that a randomly chosen ticket is in the class, the sixth column having the allocation rule and the last, whether the ticket payout is fixed, shared as part of a pool, or returns nothing. The 2/6+ and 3/6 tickets receive \$5 and \$10, respectively, and 2/6- tickets receive a free play in the next lottery, but a charge of \$1.41 is applied to the *prize pool*. See Example \[Exmp:ExampleOfPrizePayouts\] for details. The payouts are shown in the first four lines of the table for 4/6, 5/6-, 5/6+ and 6/6 tickets. These type (C) tickets share the remainder of the $0.40*d_{\!_{BP}}$ after deductions for tickets of types (A) and (B). The amount $0.40*d_{\!_{BP}} - (\text{payouts to 2/6+, 3/6 and charges for 2/6-})$ is called the *Pools Fund*. Type (C) tickets share in a pool whose percentage of the total bets varies greatly, depending on the winning numbers of 2/6+, 3/6 and free plays. The lottery also guarantees a \$5,000,000 pool to holders of 6/6 tickets. [lclrlll]{} ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ & \# Crowd & & \# Combinations & & Allocatio & Share\ Type & Tickets & Class & for any ticket & Probability & Rule & Status\ ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ (C) & $N_1$ & $6/6$ & $1$ & $p_1 \sim 7.1e$-$08$ & 79.5% of the Pools Fund & Share\ (C) & $N_2$ & $5/6+$ & $6$ & $p_2 \sim 4.3e$-$07$ & 6% of the Pools Fund & Share\ (C) & $N_3$ & $5/6-$ & $252$ & $p_3 \sim 1.8e$-$05$ & 5% of the Pools Fund & Share\ (C) & $N_4$ & $4/6$ & $13,545$ & $p_4 = 0.000969$ & 9.5% of the Pools Fund & Share\ (A) & $N_5$ & $3/6$ & $246,820$ & $p_5 = 0.017650$ & $\$10$ per ticket & Fixed\ (A) & $N_6$ & $2/6+$ & $172,200$ & $p_6 = 0.012314$ & $\$5$ per ticket & Fixed\ (B) & $N_7$ & $2/6-$ & $1,678,950$ & $p_7 = 0.120064$ & free play ($\$1.41$ deduction) & Fixed\ (D) & $N_8$ & No Win & $11,872,042$ & $p_8 = 0.848984$ & Non-winner = \$0 & Returns 0\ ------------------------------------------------------------------------ \ Any unclaimed monies in the *Pools Fund* are added to the current jackpot and carried over to the next drawing. From Table \[Ta:PrizeAllocIn6492013-\], it is clear that the majority contribution to the carryover is the $79.5\%$ that occurs when there is no 6/6 winner. But 5/6+ and 5/6- tickets also have low probabilities of occurring and when there are no 5/6+ or 5/6- winners, those shares of 6% and 5%, respectively, are added to the carrover pool for the next lottery. The probabilities of these tickets occurring in an equiprobable lottery are denoted by $p_1$, $p_2$, etc. This notation is useful in the analytical expressions developed below. \[Exmp:ExampleOfPrizePayouts\] The carryover is \$30,000,000 and the crowd bets $10,000,000$ tickets, of which $1,000,000$ are assumed to be free plays, yielding a net cash contribution of $\$27,000,000$. Assuming the crowd chooses quick picks with probabilities of $1/t$, numbers for each ticket will have a binomial distribution. Random selections under the binomial are displayed in column 3 of Table \[Ta:ExamplePrizeAllocIn6492013-\], which gives the probability of this class of ticket winning. The first column is the winning ticket type, the second, the number of combinations, the fourth the total payouts to the crowd, the fifth the number of tickets held by the syndicate and the sixth, the total payouts to the syndicate. Using these numbers, we calculate the *prize pool*, the fixed payouts to crowd and syndicate, and the pools fund as follows: - Prize Pool: $\$27,580,579 = 0.40 * (13,983,816 * 3 \, + \, 0.90*30,000,000)$. - Crowd Fixed: $\$4,077,490 = 176,933*\$10 + 123,569*\$5 + 1,198,805*\$1.41$. - Syndicate Fixed: $\$5,696,520 = 246,820*\$10 + 172,200*\$5 + 1,678,950*\$1.41$. - Pools Fund: $\$17,806,569 = \$27,580,579 - \$4,077,490 - \$5,696,520$. [lrrrrr]{} ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ & & \# Crowd & Crowd & \# Syndicate & Syndicate\ Type & Combinations & Tickets & Payout & Tickets & Payout\ ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ $6/6$ & $1$ & 0 & \$0 & 1 & \$44,156,222\ $5/6+$ & $6$ & 6 & \$534,135 & 6 & \$534,135\ $5/6-$ & $252$ & 185 & \$375,960 & 252 & \$514,534\ $4/6$ & $13,545$ & 9,773 & \$708,909 & 13,545 & \$982,518\ $3/6$ & $246,820$ & 176,933 & \$1,769,330 & 246,820 & \$2,468,200\ $2/6+$ & $172,200$ & 123,569 & \$617,845 & 172,200 & \$861,000\ $2/6-$ & $1,678,950$ & 1,198,805 & \$0 & 1,678,950 & \$0\ ------------------------------------------------------------------------ \ Summing all payouts in the syndicate payout column gives \$49,516,609, for a gain of $$\$7,565,161 = \$49,516,609 - \$3*13,983,816.$$ plus 1,678,200 free plays in the next lottery. The cash payout from non-6/6 tickets is \$5,360,387, despite a crowd and syndicate bet of \$68,951,454. Clearly, the jackpot must be large in order for buying the pot to be justifiable. Expected Return from Buying the Pot {#S:ExpectedReturnFromBuyingThePot} =================================== Notation and Terminology ------------------------ Table \[Ta:FixedParametersForThe649\] gives the fixed parameters of the lottery, namely those that do not involve betting strategies of the syndicate or crowd. [ll]{} ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ Notation & Description\ ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ $t$ & Number of tickets in the lottery = 13,983,816.\ $a$ & Carryover pool in dollars, $a \ge 0$.\ $p_i$ & Probability that a ticket is of class $i$ assuming that\ & the winning ticket is drawn equiprobably (see Table \[Ta:PrizeAllocIn6492013-\]).\ $f$ & Fraction of tickets that are “free plays.”\ $c$ & Number of tickets bet by the crowd.\ ------------------------------------------------------------------------ \ Table \[Ta:RandomVariablesThatInvolveBettorStrategies\] has the notation for the random variables that account for stochasticity and strategy in playing the lottery. [ll]{} ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ Notation & Description\ ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ $N_i$ & Random variable for the number of tickets of class $i$ bet by\ & the crowd (see Table \[Ta:PrizeAllocIn6492013-\]).\ $N$ & Vector $N = (N_1, N_2, \ldots, N_8)'$.\ $d_{\!_{AB}}$ & Dollars awarded or deducted for tickets of types (A) and (B).\ $d_{\!_{BP}}$ & Dollars in the *betting pool*.\ $d_{\!_{PP}}$ & Dollars in the *prize pool*.\ $d_{\!_{PF}}$ & Dollars in the *Pools Fund*.\ ------------------------------------------------------------------------ \ Using the notation in Tables \[Ta:FixedParametersForThe649\] and \[Ta:RandomVariablesThatInvolveBettorStrategies\], the number of dollars in each fund is $$\begin{aligned} d_{\!_{AB}} &= 10 (N_5 + t p_5) \, + \, 5 (N_6 + t p_6) \, + \, 1.41 (N_7 + t p_7) \label{dAB} \\ d_{\!_{BP}} &= 3 ( t + c (1 - f)) \label{dBP} \\ d_{\!_{PP}} &= d_{\!_{BP}} 0.4 \label{dPP} \\ d_{\!_{PF}} &= d_{\!_{PP}} - d_{\!_{AB}} \label{dPF}\end{aligned}$$ Since $f$ is non-random, the second entry of the above table is the only non-stochastic entry. Equiprobable Betting by the Crowd {#S:EquiprobableBettingByTheCrowd} --------------------------------- We calculate first the expected return to a syndicate that buys the pot when the crowd chooses tickets independently and equiprobably. As we discuss in Section \[S:Non-equiprobableBettingByTheCrowd\], this is the crowd’s optimal strategy, although they do not employ it in practice — and the cost of this “mistake” is considerable. ### Syndicate’s Expected Value for Equiprobable Crowd Betting In Appendix \[S:TheSyndicate’sExpectedValueWhenTheCrowdBetsEquiprobably\], we develop a formula for the syndicate’s expected gain $G(c)$ from the wagering of $\$41,951,448 = \$3 * 13,983,816$ on $13,983,816$ tickets: $$\begin{aligned} E[ \, G(c) \, ] & = (a + 0.795*\mu(c)) \, \lambda(c)^{-1} (1 - exp(-\lambda(c))) \label{E:SyndicateExpectedReturn} \\ & + \left( 0.06 \nu(c) \, + \, 0.145 \frac{1}{1 + c/t} \right) \mu(c) \nonumber \\ & + \$3,329,200 \, - \, \$3 t, \nonumber\end{aligned}$$ where $$\begin{aligned} \lambda(c) &= c/t, \\ \mu(c) &= 0.40 \cdot 3 \left( ( t + c \cdot (1 - f)) \right) - \, (t + c) \cdot (p_5 \$10 \, + \, p_6 \$5 \, + \, p_7 \$1.41), \\ \nu(c) &= E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right], \qquad \text{for } X \sim Bin(c,6/t).\end{aligned}$$ The term $\nu(c)$ is calculated using the recursive formula in Appendix \[S:TheSyndicate’sExpectedValueWhenTheCrowdBetsEquiprobably\] and appears as the last column of Table \[Ta:BettingThresholds\]. ### Parameters that Lead to Positive Expected Return {#SSS:ParametersThatLeadToPositiveExpectedReturn} Consider the implications of the 6/49 rules and of Formula . Because the lottery sponsors take such a high percentage of the betting pool (60%), a large jackpot is needed for a syndicate to have a positive expected return. When a syndicate bets one of each ticket, previous analysis showed the syndicate’s numbers of winning tickets are known exactly, irrespective of the winning numbers from the drawing. There will always be exactly 1 winning ticket, exactly 6 5/6+ tickets, exactly 252 5/6- tickets, and so on. The RHS of first line of formula dominates the others when a jackpot $a$ is large. Table \[Ta:BettingThresholds\] shows the results of applying formula for 10 levels of total crowd betting (*c*) to solve for the sizes of carryover pools (*a*) that produce expected returns of 0%, 10% and 20% for the syndicate. Since the cost of buying the pot is $\$3 * 13,983,816 =\$41,951,448$, a return of 10% is $\$4,195,145$. When the crowd bets \$30 million, for example, any carryover larger than \$36.92 million is a potential play for the syndicate, and carryovers of \$42.80 and \$48.67 million have expected returns of 10% and 20%, respectively. The last three columns of Table \[Ta:BettingThresholds\] provide insight into the payout structure. The sixth column shows the expected amount in the *Pools Fund* and the next column is its percentage in the *prize pool*. Thus when the crowd bets \$40 million, the expected *Pools Fund* is \$19.97 million, which is 49.68% of the prize pool. Thus, the charges for fixed payout tickets amount to \$50.32% of the *prize pool*. The final column is the expected value for the 5/6+ factor: $$\text{EV56+} = E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right]. \label{E:ExpValOf5/6+Factor}$$ The expected value declines when the crowd bets more, as one expects since $X_{\text{5/6+}}$ is generally larger. [rccccccc]{} ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ & Crowd & & Expected & (Pools Fund) &\ & \$ Bet & Breakeven & +10% & +20% & Pools Fund & (Prize Pool) & Eqn.\ & (millions) & (millions) & (millions) & (millions) & (millions) & (%) & (%)\ ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ 3.3 & 9 & 30.33 & 35.05 & 39.76 & 13.28 & 58.39 & 82.71\ 6.7 & 18 & 33.46 & 38.74 & 44.01 & 15.51 & 54.02 & 70.15\ 10.0 & 27 & 36.92 & 42.80 & 48.67 & 17.74 & 51.32 & 60.71\ 13.3 & 36 & 40.71 & 47.22 & 53.73 & 19.97 & 49.68 & 53.40\ 16.7 & 45 & 44.81 & 51.99 & 59.17 & 22.20 & 48.73 & 47.60\ 20.0 & 54 & 49.21 & 57.10 & 64.99 & 24.44 & 48.27 & 42.90\ 23.3 & 63 & 53.90 & 62.52 & 71.15 & 26.67 & 48.14 & 39.02\ 26.7 & 72 & 58.84 & 68.24 & 77.63 & 28.90 & 48.25 & 35.77\ 30.0 & 81 & 64.03 & 74.23 & 84.42 & 31.13 & 48.53 & 33.01\ 33.3 & 90 & 69.45 & 80.46 & 91.48 & 33.37 & 48.92 & 30.64\ ------------------------------------------------------------------------ \ Recall from Example \[Exmp:ExampleOfPrizePayouts\] that the crowd bet a net \$27,000,000 on 10 million tickets and the carryover was \$30,000,000 — yet the syndicate won over \$6 million. According to Table \[Ta:BettingThresholds\], the syndicate should not bet under these conditions, since a minimum carryover of \$36.92 million is necessary. There is no problem here, since the numbers in the table are expected values and it is quite possible for a syndicate to win despite making an unfavorable bet. The syndicate in that example just got lucky. Non-equiprobable Betting by the Crowd {#S:Non-equiprobableBettingByTheCrowd} ------------------------------------- The calculations in Section \[S:EquiprobableBettingByTheCrowd\] assumed that the crowd bets independently using $q = \frac{1}{t} 1_t$, where $1_t$ is a t-vector of all ones. What happens when the crowd bets using $q \ne \frac{1}{t} 1_t$? In part \[Enum:OptimalStrategy\], we stated a result from [@moffitt:ziemba:2017a], that for *pure jackpot* lotteries (ones having a single prize, a non-stochastic jackpot[^1] $v$) the expected payoff is $$E_q \left[ v \frac{1}{1 + N_1} \right] \, = v E_q \left[ \frac{1}{1 + N_1} \right] \, > \, v E_{1_t/t} \left[ \frac{1}{1 + N_1} \right] = \, E_{1_t/t} \left[ v \frac{1}{1 + N_1} \right], \label{E:CrowOptimalqi}$$ where $q \ne 1/t 1_t$, $N_1$ is the random number of 6/6 tickets held by the crowd. However, formula does not apply in the present case because $v$ is stochastic, depending on the size of the *Pools Fund*. Consider a non-stochastic configuration of single ticket bets $n_j = (n_{j1}, n_{j2}, \ldots, n_{jt})'$ for individuals $j = 1, \ldots, c$, each having zeroes except for a single $1$ in some position. Define $z_k = \sum_{j=1}^{j=c} n_{jk}$ and t-vector $z = (z_1, \ldots, z_t)'$. Clearly, $\sum_{k=1}^{k=t} z_k = c$. To compute the expected values of ticket types 1, 2, …7 with respect to an equiprobable drawing, observe that as $i$ ranges over all ticket drawings $i = 1, \ldots t$, for **any** $n_j$, the number of 6/6 is 1, the number of 5/6+ is 6, the number of 5/6 is 252, and so on as indicated in Table \[Ta:PrizeAllocIn6492013-\]. Since the drawing is equiprobable, dividing each of these by $t$ gives the probability that **any** non-stochastic ticket will be of the indicated type under an equiprobable drawing. Define indicator functions on single ticket t-vectors $n$ as: $$I_{x/6}(n) = \begin{cases} 1 & \text{if ticket } n \text{ is a x/6 ticket}, \\ 0 & \text{otherwise.} \end{cases}$$ Applying this to fixed payout types 3/6, 2/6+ and 2/6, we obtain for $d_{\!_{AB}}$ in formula $$\begin{aligned} E_e[d_{\!_{AB}}] &= E_e \left[ \sum_{j=1}^{j=c} \left\{ \$10 I_{3/6}(n_j) \, + \, 5 I_{2/6+}(n_j) \, + \, 1.41 I_{2/6} \right\} \right] + \$5,696,520 \nonumber \\ &= \left( \$10 p_{3/6} \, + \$5 \, p_{2/6+} \, + \, \$1.41 p_{2/6} \right) c \, + \, \$5,696,520, \nonumber \\ &= \$0.4073651 \cdot c * \, + \, \$5,696,520. \label{E:ExpectedFixedTicketPayout}\end{aligned}$$ where the notation $E_e$ emphasizes that the expectation is taken over equiprobable drawings and $\$5,696,520$ is the fixed payout/deduction for the syndicate.[^2] The (stochastic) jackpot is $v = a \, + \, 0.795 d_{\!_{PF}}$ and the random 6/6 payout to the syndicate is $$\begin{aligned} v \frac{1}{1 + N_1} &= \left( a \, + \, 0.795 d_{\!_{PF}} \right) \frac{1}{1 + N_1} \nonumber \\ &= \left( a \, + \, 0.795 (0.4 (3( t \, + \, c(1 - f))) - d_{\!_{AB}}) \right) \frac{1}{1 + N_1} \nonumber \\ &= \left( a \, + \, 0.954( t \, + \, c(1 - f)) \right) \frac{1}{1 + N_1} - \frac{d_{\!_{AB}}}{1 + N_1} \nonumber \\ &= \left( a \, + \, 0.954( t \, + \, c(1 - f)) - \$5,696,520 \right) \frac{1}{1 + N_1} - \label{E:SyndicateTermOfJackpotPayout} \\ & \qquad \frac{\$10 N_5 \, + \, 5 N_6 \, + \, 1.41 N_7}{1 + N_1} \label{E:CrowdTermOfJackpotPayout}\end{aligned}$$ In , the factor multiplying $1/(1 + N_1)$ is fixed. Its expectation using is $$\begin{aligned} & E_q \left[ \left( a \, + \, 0.954( t \, + \, c(1 - f) - \$5,696,520) \right) \frac{1}{1 + N_1} \right] \nonumber \\ & \qquad > \left( a \, + \, 0.954( t \, + \, c(1 - f) - \$5,696,520) \right) \frac{1}{\lambda} ( 1 - \exp(-\lambda) ), \label{E:SyndicateCostBound}\end{aligned}$$ where $\lambda = c/t$. Thus for this term at least, the syndicate gets more than a fair split of the jackpot since $$\frac{1}{\lambda} ( 1 - \exp(-\lambda) ) > \frac{t}{t + c}.$$ The second term depends on $N_1$, $N_5$, $N_6$ and $N_7$, which respectively, are the numbers of 6/6, 3/6, 2/6+ and 2/6 tickets held by the crowd, and these are dependent on the crowd betting probabilities $q = (q_1, \ldots, q_t)'$. But we do not have the data to model the joint distribution of $(N_1, N_5, N_6, N_7)$ which is needed to evaluate . However, we have circumstantial evidence that $N_5$, $N_6$ and $N_7$ are positively correlated with $N_1$. Therefore we make a crude assumption that the joint crowd payouts for 3/6, 2/6+ and 2/6 tickets are increased linearly with the winning ticket, that is, the payout for ticket $i$ is proportional to $q_i$: $$\frac{\$10 N_5 \, + \, \$5 N_6 \, + \, \$1.41 N_7}{1 + N_1} \cdot q_i / (1/t).$$ Thus if the winning ticket $i$ is bet with twice the frequency of an equiprobable bet, so that $t q_i = 2$, then the fixed payouts/deductions will be twice that expected in the equiprobable case (see the discussion leading to equation ). Using $H = \$10 p_5 \, + \, \$5 p_6 \, + \, \$1.41 p_7$, we calculate $$\begin{aligned} E_{1_t/t} \left[ \min \left( c H t q_i, d_{PF} \right) \frac{1}{1 + N_1} \right] &\le E_{1_t/t} \left[ c H t q_i \frac{1}{1 + N_1} \right] \nonumber \\ &= H t E_{1_t/t} \left[ c q_i \frac{1}{c q_i} (1 - e^{-cq_i}) \right] \nonumber \\ &= H t \sum_{i=1}^{i=t} \frac{1}{t} (1 - e^{-cq_i}) \nonumber \\ &\le H t (1 - e^{-c/t}) \label{E:CrowdCost} \\ &= \frac{c H}{\lambda} (1 - e^{-\lambda}), \label{E:CrowdCostBound}\end{aligned}$$ where $\lambda = c/t$ and the step follows from Jensen’s inequality since $1 - e^{-cq}$ is a concave function of $q$. Jensen’s inequality can be stated as follows. A function $f: [a,b] \rightarrow \mathbb{R}$ that satisfies $f(ta + (1-t)b) \le t f(a) + (1-t) f(b)$ for all $t \in (0,1)$ is called *convex*, and if the inequality is strict, *strictly convex*. For a random variable $X$ and *convex* function $f$, Jensen’s inequality asserts that $f(E[X]) \le E[f(X)]$. Further, if $X$ is not degenerate and $f$ is strictly convex, then $f(E[X]) < E[f(X)]$. A function $f$ is (*strictly*) *concave* if $-f$ is (*strictly*) *convex*i, so Jensen’s inequality is reversed for *concave* functions. Putting together with yields $$\begin{aligned} E \left[v \frac{1}{1 + N_1} \right] &\ge \left( a \, + \, 0.954( t \, + \, c(1 - f)) - \$5,696,520 \, - \, c H \right) \nonumber \\ &\qquad \cdot \frac{1}{\lambda} (1 - e^{-\lambda}) \label{E:SyndicateExpectedReturn}\end{aligned}$$ where $\lambda = c/t$ and $H = \$10 p_5 \, + \, \$5 p_6 \, + \, \$1.41 p_7$. This calculation shows that the syndicate obtains a better result than when the crowd bets proportionally, as in the corresponding result for *pure jackpot* lotteries. Design Considerations for Lotteries {#SS:DesignConsiderationsForLotteries} =================================== Lottery design includes the goal to maximize the the sponsors’ earnings. Assuming fairly constant fixed costs of running the lottery, sponsors should strive to make the lottery popular, thereby increasing profitability. The most recent changes to payouts were made with that goal in mind — these changes increased the “convexity” of payouts, meaning many little prizes and greater jackpot growth. Ziemba recommended these designs in his work in the 1980’s and [@Walker2008459] later also recommended them. Convex designs encourage players because more “get something back,” while at the same time growing large jackpots quickly. This design is supported by research in behavioral finance. Lopes’ SP/A (Security-Potential/Aspiration) model ([@Lopes1987255]), an improved version of the classic Friedman/Savage (1948) utility curves, argues that many unsophisticated gamblers prefer strategies of buying safe prospects with a few longshots (the “Cautiously Hopeful” pattern of SP/A). Regarding large *jackpots*, Daniel Kahneman has written > “For emotionally significant events, the size of the probability simply doesn’t matter. What matters is the possibility of winning. People are excited by the image in their mind. The excitement grows with the size of the prize, but it doesn’t diminish with the size of the probability.” Source: [@NYTimes:YourMoney:KahnemanQuote:Online]. There is another aspect of lottery design, namely, discouraging syndicates from buying the pot. There are two ways to accomplish this: (1) creating a large number of tickets making it logistically difficult to buy the pot, and (2) using convex designs, which reduces the likelihood that pot buying situations will occur. Method (1) is not feasible except for large lotteries like the California Powerball lottery. The reason is that if the number of tickets sold are too small relative to the total number of tickets, the jackpot may build slowly and seldom be won. On the other hand, method (2) can be effective regardless of the size of the lottery. To illustrate, consider a *pure jackpot* lottery with the same carryover, take and crowd betting as in Table \[Ta:BettingThresholds\]. The results are shown in Table \[Ta:BettingThresholdsForPureJockpotAnd649\]. The first column has the number of tickets, which after a 10% deduction for free plays, equals the crowd contribution to the *betting pool* shown in the second column. Then assuming a take of 60%, breakeven thresholds of 0%, 10% and 20% for the pure jackpot lottery are shown in columns 3-5 and for the 6/49 in columns 6-8. The results show that buying the pot thresholds are lower in the pure lottery, but not as much as might be expected. But one can see the reason by a simple argument. When the sponsors takes 60%, only 40 cents is returned as prizes for each dollar wagered. Therefore, a syndicate needs to recover 60% of the covering bet, or $0.6*\$3*13,983,816 = \$25,170,869$, regardless of the lottery’s rules. As we have shown, the syndicate earns its fair share of consolation prizes, but the free plays it earns are not worth too much since after the lottery is hit the next lottery when those tickets will be used will have a small purse. Assuming the the crowd bets $\$1,000,000$ on the next lottery the expected value of these $1,678,950$ tickets will under optimal wagering be worth about $\$150,000$. [rccccccc]{} ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ & Crowd & &\ & \$ Bet & Breakeven & +10% & +20% & Breakeven & +10% & +20%\ & (millions) & (millions) & (millions) & (millions) & (millions) & (millions) & (millions)\ ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ 3.3 & 9 & 26.77 & 31.48 & 36.20 & 30.33 & 35.05 & 39.76\ 6.7 & 18 & 28.76 & 34.04 & 39.31 & 33.46 & 38.74 & 44.01\ 10.0 & 27 & 31.14 & 37.02 & 42.89 & 36.92 & 42.80 & 48.67\ 13.3 & 36 & 33.90 & 40.41 & 46.92 & 40.71 & 47.22 & 53.73\ 16.7 & 45 & 37.02 & 44.20 & 51.38 & 44.81 & 51.99 & 59.17\ 20.0 & 54 & 40.49 & 48.38 & 56.26 & 49.21 & 57.10 & 64.99\ 23.3 & 63 & 44.28 & 52.91 & 61.53 & 53.90 & 62.52 & 71.15\ 26.7 & 72 & 48.37 & 57.77 & 67.17 & 58.84 & 68.24 & 77.63\ 30.0 & 81 & 52.75 & 62.94 & 73.13 & 64.03 & 74.23 & 84.42\ 33.3 & 90 & 57.38 & 68.39 & 79.41 & 69.45 & 80.46 & 91.48\ ------------------------------------------------------------------------ \ We conclude the discussion by examining the impacts of design choices in the 6/49 Lotto. The 6/49 Lotto’s convex design according to Table \[Ta:BettingThresholdsForPureJockpotAnd649\] raised the bar for buying-the-pot strategies, making carryover thresholds roughly 12% to 20% higher. We now compare the impacts of the 6/49’s design features toward increasing the threshold for buying the pot. We identify four factors: (1) the take, (2) the payouts for small prizes, (3) the payouts for large, non 6/6 prizes, and (4) free plays. Then we compare by 1. Changing the take only, using alternatives 55%, 60% (current) and 65%. 2. Eliminating fixed payouts 3/6 and 2/6+ only. 3. Eliminating 4/6, 5/6 and 5/6+ payouts only. 4. Eliminating free plays only. Table \[Ta:BettingThresholdsForDesignFactors\] shows breakeven carryover thresholds for these design factors. The factor is indicated in the first column and the other 5 columns are carryover thresholds when the crowd bets the indicated millions of dollars, 20, 40, etc. In the second column (corresponding to a crowd bet of \$20 million), the numbers in parenthesis are differences of threshold carryovers from the current 6/49 values (second row, second column). Since the relative impacts of these factors are the same for the five crowd betting amounts, their impacts on the buying the pot strategy can be assessed using this column. The greatest factor impact is due to free plays; removing them drops the threshold by $\$3.39$ million ($\sim 10\%$). The largest inhibitor is clearly the take — increasing it by $0.05\%$ from to $65\%$ has a large impact on breakeven carryovers. [lcccccccccc]{} ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ &\ & 20 & 40 & 60 & 80 & 100\ & million & million & million & million & million\ ------------------------------------------------------------------------ \ ------------------------------------------------------------------------ \ TAKE=0.55 & 30.56 (-2.90) & 36.98 & 44.64 & 53.41 & 63.15\ CURRENT 6/49 & 33.46 ( 0.00) & 40.71 & 49.21 & 58.84 & 69.45\ TAKE=0.65 & 36.37 ( 2.91) & 44.44 & 53.79 & 64.27 & 75.74\ NO 2/6+, 3/6 & 32.88 (-0.58) & 39.65 & 47.76 & 57.08 & 67.44\ NO 4/6, 5/6 & 32.99 (-0.47) & 39.87 & 48.07 & 57.48 & 67.91\ NO FREE PLAY & 30.07 (-3.39) & 36.28 & 43.73 & 52.29 & 61.81\ ------------------------------------------------------------------------ \ Based on these statistics, we make recommendations for state lotteries using ratings of the form $(+=-\pm\mp, +=-\pm\mp)$. The first sign is for popularity, the second for inhibiting buyers of the pot. For example, $(+,-)$ indicates that a factor increases the lottery’s popularity, but encourages buying the pot. 1. $(\mp,+)$ If possible, add combinations to the lottery by increasing the numbers. 2. $(+,+)$ Initiate a free play feature. 3. $(\mp,+)$ Increase the take. 4. $(+,=)$ Offer many small prizes. 5. $(+,=)$ Increase the allocation of the *Pools Fund* to 6/6 winners. 6. $(=,+)$ Decrease the awards to hard-to-win non 6/6 tickets. Increasing the allocation to 6/6 allows quicker build-up of jackpots, which encourages greater crowd participation. However, we did not address the question of build-up speed of the jackpot, nor the acceleration of betting on larger jackpots. These need to be studied in order to design prizes and allocations to optimize betting flows. Conclusions =========== In this paper, we have shown conditions under which buying the pot in the 6/49 Lotto has positive expected return when the crowd bets equiprobably. We also indicated that equiprobable betting is optimal for the crowd, that is, expected return is lower when it does not bet equiprobably. We illustrated the advantages of lotteries with convex designs by calculating 6/49 carryover thresholds and comparable *pure jackpot* carryover thresholds. We then rated various design features for their likelihood of increasing a lottery’s popular, and decreasing the likelihood of buyers of the pot. The Syndicate’s Expected Value when the Crowd bets Equiprobably {#S:TheSyndicate'sExpectedValueWhenTheCrowdBetsEquiprobably} =============================================================== Assuming that the lottery’s tickets are equiprobable, $(N_1, \ldots N_8)'$ has a multinomial distribution $$(N_1, \ldots N_8)' \sim Multin(c + t, p),$$ where $p = (p_1, p_2, \ldots, p_8)'$. The distribution of deductions $d_{\!_{AB}}$ from the *prize pool* is given by $$d_{\!_{AB}} = \beta_{\!_{AB}}' N,$$ where $\beta_{\!_{AB}} = (0,0,0,0,10,5,1.41,0)'$ and $N \sim Multin(c + t, p)$. Substituting and into equation gives the *Pools Fund* as $$d_{\!_{PF}} = 0.40 \cdot 3 \cdot ( t + c \cdot (1 - f)) \, - \, d_{\!_{AB}}.$$ and in RHS of this expression, only $d_{\!_{AB}}$ is random. Using results from [@moffitt:ziemba:2017a], the expected value $G$ of the syndicate’s net gain, given $d_{\!_{PF}}$, as $$\begin{aligned} E[ G(c) \, | \, d_{\!_{PF}}] &= 0.795 \cdot (a + d_{\!_{PF}}) E \left[ \frac{1}{1 + X_{\text{6/6}}} \right] \label{E:6/6Expectation} \\ & + 0.06 \cdot d_{\!_{PF}} E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right] \label{E:5/6+Expectation} \\ & + 0.05 \cdot d_{\!_{PF}} E \left[ \frac{252}{252 + X_{\text{5/6-}}} \right] \label{E:5/6-Expectation} \\ & + 0.095 \cdot d_{\!_{PF}} E \left[ \frac{13545}{13545 + X_{\text{4/6}}} \right] \label{E:4/6Expectation} \\ & + \$2,468,200 \, + \, \$86,100 \, - \, \$3 t \label{E:2-3/6Expectation}\end{aligned}$$ where $$\begin{aligned} (X_{\text{6/6}}, X_{\text{5/6+}}, X_{\text{5/6-}}, X_{\text{4/6}})' \sim Multin(c,(1, 6, 252, 13545)'/t).\end{aligned}$$ Since buying one of each ticket gives the same exact payout regardless of the winning ticket (numbers of tickets shown in Table \[Ta:PrizeAllocIn6492013-\]), we know that a covering strategy pays \$2,468,200 and \$86,100, respectively, for 3/6 and and 2/6+ tickets. This explains the term . Using the formulas from we obtain for $\lambda(c) = c/t$ $$E \left[ \frac{1}{1 + X_{\text{6/6}}} \right] = \lambda(c)^{-1} (1 - exp(-\lambda(c)))$$ and values $\nu(c) = E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right]$ using recursion. These calculations take care of terms and . Using the Law of Large Numbers, the expectation in the term can be approximated by $$\frac{252}{252 \, + \, 252 c/t} = \frac{1}{1 + c/t}. \label{E:FairSplitExpectation3}$$ and in term by $$\frac{13545}{13545 \, + \, 13545 c/t} = \frac{1}{1 + c/t}. \label{E:FairSplitExpectation4}$$ Basically, these two approximations amount to fair split of the corresponding share of the *Funds Pool*. Thus $$\begin{aligned} E[ G(c) \, | \, d_{\!_{PF}}] & = (a + 0.795 \cdot d_{\!_{PF}}) \lambda(c)^{-1} (1 - exp(-\lambda(c))) \label{E:6/6ExpectationC} \\ & + 0.06 \cdot d_{\!_{PF}} E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right] \label{E:5/6+ExpectationC} \\ & + 0.145 \cdot d_{\!_{PF}} \frac{1}{1 + c/t} \label{E:5/6-ExpectationC} \\ & + 2,468,200 \, + \, 86,100 \, - \, \$3 t \label{E:2-3/6ExpectationC} \end{aligned}$$ where $\lambda(c) = c/t$. To complete the calculation, we need to eliminate the dependence of $E[ G(c) \, | \, d_{\!_{PF}}]$ on $d_{\!_{PF}}$ by determining its distribution and performing an integration. But this is straightforward: the first three terms , and are linear in $d_{\!_{PF}}$ so that the expectation $\mu(c) = E[d_{\!_{PF}}]$ should be substituted for $d_{\!_{PF}}$. The expectation $E[d_{\!_{PF}}]$can be calculated by substituting into and into taking expectations: $$\mu(c) = E[d_{\!_{PF}}] = 0.40 \cdot 3 \cdot ( t + c \cdot (1 - f)) \, - \, E[d_{\!_{AB}}]$$ We calculate $E[d_{\!_{AB}}]$ as follows. For any ticket $i$, the number of tickets that are 3/6, 2/6+ and 2/6- are respectively $248,820$, $172,200$, and $1,678,950$, respectively. Therefore, the probability that ticket $i$ is a 3/6, 2/6+ or 2/6- ticket, *given that a winning ticket is drawn equiprobably*, is $p_5 = 248,820/t$, $p_6 = 172,200/t$ and $p_7 = 1,678,950/t$, respectively. Now consider any choice of $c$ tickets. By linearity of expectations, the expected number of 3/6 tickets is $c * p_5$, of 2/6+ tickets is $c * p_6$ and of 2/6- tickets, $c * p_7$. Therefore, $$\nu(c) = E[d_{\!_{AB}}] = (t + c) (p_5*\$10 + p_6*\$5 + p_7*\$1.41).$$ Summarizing, the expected gain $G(c)$ to a syndicate that covers the pool is $$\begin{aligned} E[ \, G(c) \, ] & = (a + 0.795 \cdot \mu(c)) \lambda(c)^{-1} (1 - exp(-\lambda(c))) \label{E:SyndicateExpectedReturnAppendix} \\ & + \, \left( 0.06 \nu(c) \, + \, 0.145 \frac{1}{1 + c/t} \right) \mu(c) \nonumber \\ & + \, \$2,553,300 \, - \, \$3 t, \nonumber\end{aligned}$$ where $$\begin{aligned} \lambda(c) &= c/t, \\ \mu(c) &= 0.40 \cdot 3 \left( ( t + c \cdot (1 - f)) \right) - \, (t + c) \cdot (p_5 \$10 \, + \, p_6 \$5 \, + \, p_7 \$1.41), \\ \nu(c) &= E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right], \qquad \text{for } X \sim Bin(c,6/t).\end{aligned}$$ [^1]: We are assuming that the crowd’s number of tickets, $c$, is known. [^2]: $\$5,696,520 = \$10*246,820 + \$5*172,200 + \$1.41*1,678,950$.

ArXiv

--- abstract: 'We study effects of possible tachyonic perturbations of dark energy on the CMB temperature anisotropy. Motivated by some models of phantom energy, we consider both Lorentz-invariant and Lorentz-violating dispersion relations for tachyonic perturbations. We show that in the Lorentz-violating case, the shape of the CMB anisotropy spectrum generated by the tachyonic perturbations is very different from that due to adiabatic scalar perturbations and, if sizeable, it would be straightforwardly distinguished from the latter. The tachyonic contribution improves slightly the agreement between the theory and data; however, this improvement is not statistically significant, so our analysis results in limits on the time scale of the tachyonic instability. In the Lorentz-invariant case, tachyonic contribution is a rapidly decaying function of the multipole number $l$, so that the entire observed dipole may be generated without conflicting the data at higher multipoles. On the conservative side, our comparison with the data places limit on the absolute value of the (imaginary) tachyon mass in the Lorentz-invariant case.' author: - | M.V. Libanov$^{\ddagger}$, V.A. Rubakov$^{\ddagger}$, O.S. Sazhina$^{\star}$ and M.V.Sazhin$^{\star}$\ \ \ $^{\dagger}$ Institute for Nuclear Research of the Russian Academy of Sciences,\ 60th October Anniversary Prospect, 7a, Moscow, Russia;\ $^{\star}$ P.K. Sternberg Astronomical Institute of the Moscow State University,\ Universitetsky Prospect, 13, Moscow, Russia. title: | CMB anisotropy induced by\ tachyonic perturbations of dark energy --- [**PACS:**]{} 98.80.-k, 95.36.+x Introduction ============ Recently, a number of suggestions have been put forward for explaining the observed accelerated expansion of the Universe. Among them there are the presence of the cosmological constant, modification of gravity at ultra-large scales, existence of new light fields (for reviews, see, e.g., Refs. [@Padmanabhan:2002ji; @Sahni:2004ai; @Copeland:2006wr; @Sahni:2006pa; @Frieman:2008sn]). In the latter case, dark energy can be characterized by equation of state $p=w\rho$, where the parameter $w$ is different from $-1$ and generically depends on time. In a simple version, dark energy is due to a scalar field (quintessence), and the parameter $w$ obeys the constraint $w>-1$, while $w=-1$ for the cosmological constant. However, the value of $w$ may be strongly negative, $w<-1$; this is the case for phantom dark energy. Existing data do not exclude also a possibility  [@Komatsu:2008hk; @Sahni:2008xx; @Xia:2008ex] that at relatively large redshift the dark energy is of quintessence type $w>-1$, and later it becomes phantom with $w<-1$. Phantom energy violates null energy condition, which is usually a signal for instabilities. As an example, the simplest model of scalar field with the wrong sign of kinetic term [@Caldwell:2002] suffers from the presence of ghost (negative energy state) at arbitrarily high spatial momenta. This implies catastrophic vacuum instability. However, in phantom models that break Lorentz-invariance, violation of the null energy condition at cosmological scales (related to the property that $w<-1$ for spatially homogeneous phantom field) does not necessarily imply unacceptable instabilities at shorter scales. This suggests that Lorentz-violating phantom theories may be viable. Indeed, models of this sort have been recently constructed [@Senatore; @Crem:2006; @Rubakov:2006; @Libanov:2007]. A property of one class of these models [@Rubakov:2006; @Libanov:2007] is that there is a tachyonic mode in the perturbation spectrum about the homogeneous phantom background. This mode occurs at sufficiently small spatial momenta only, so that the time scale of the tachyonic instability may be roughly comparable to the age of the Universe. This is not particularly dangerous. It is conceivable that the existence of tachyonic modes at low spatial momenta is a fairly generic property of a class of phantom models: the violation of the null energy condition may show up precisely in this way. Therefore, it is of interest to study observable consequences of such models. In this paper we consider one of these consequences, namely, the effect of tachyonic modes on the anisotropy of CMB temperature. We adopt the phenomenological approach, and instead of using results obtained within a concrete model, we parametrize the tachyonic instability by a few parameters. We will consider models with a Lorentz-violating dispersion relation and Lorentz-invariant one. In the Lorentz-violating case, we parametrize the dispersion relation as follows: $$\omega^2 = \alpha |{\bf p}| ( M - |{\bf p}|), \label{dispers}$$ where $\alpha$ and $M$ are constant parameters, and ${\bf p}$ is the physical spatial momentum. Our convention is that positive values of $\omega ^{2}$ correspond to exponential growth of perturbations, while the usual oscillatory behaviour occurs at negative $\omega ^{2}$. Thus, the parameter $M$ equals to the momentum below which the mode is tachyonic. For given $\alpha$, the parameter $M$ determines also the time scale of instability. The parametrization (\[dispers\]) is chosen in accord with Refs. [@Rubakov:2006; @Libanov:2007] where similar dispersion relation has been found in a concrete model of phantom energy. The analysis presented below can be straightforwardly generalized to other forms of dispersion relation. In the Lorentz-invariant case the dispersion relation has the form $$\omega ^{2}=M^{2}-\mathbf{p}^{2} \label{Eq/Pg3/1:PhanTach}$$ In this case too, the tachyon mass $M$ equals to the spatial momentum below which the mode is unstable, and $M^{-1}$ is the time scale of instability. From our perspective, the important difference between the dispersion relations (\[dispers\]) and (\[Eq/Pg3/1:PhanTach\]) is that in the latter case the “frequency” is nonzero at ${\bf p}=0$ and monotonously decreases as $|{\bf p}|$ increases, whereas in the former, the “frequency” vanishes as ${\bf p}=0$ and has a maximum at finite $|{\bf p}|$. This will lead to qualitatively different shapes of the contributions to the CMB anisotropy spectra. It is worth noting in this regard that the Lorentz-invariant model may be viewed as a representative of a class of theories with tachyonic perturbations: analogous results would hold for Lorentz-violating models with dispersion relations similar to (\[Eq/Pg3/1:PhanTach\]). In the cosmological context, the dispersion relations (\[dispers\]) and (\[Eq/Pg3/1:PhanTach\]) are written as follows: $$\begin{aligned} \omega^2(t) &=& \alpha \dfrac{k}{a(t)}\left( M - \dfrac{k}{a(t)}\right) \; , \nonumber \\ \omega^2 (t) &=& M^2 - \frac{k^2}{a^2(t)} \; , \nonumber\end{aligned}$$ respectively, where $k$ is the time-independent conformal momentum and $a(t)$ is the scale factor. In the expanding Universe, a mode of given $k$ is first normal, and after the physical momentum gets reshifted down to $k/a=M$, it becomes unstable. For $ M > H_{0}$, where $H_{0}$ is the present value of the Hubble parameter, the tachyonic modes in both models start to grow exponentially at times preceding the present cosmological epoch. The growth of the tachyonic perturbations gives rise to the growth of the gravitational potential $\Phi$ generated by these perturbations[^1] [@rub08] $$\begin{aligned} \Phi (t,\mathbf{x})&=&\frac{1}{(2\pi )^{3/2}}\int \limits_{}^{}\!d\mathbf{k}\Phi (t,\mathbf{k})\mathrm{e}^{i\mathbf{kx}} + \mbox{h.c.}, \nonumber\\ \Phi(t, {\bf k}) &=&{A({\bf k} )} \exp\left(\int\limits^t_{t_k}\omega(t^{\prime})\;dt^{\prime}\right), \label{gravpot}\end{aligned}$$ where $A( {\bf k} )$ is the amplitude of primordial fluctuations in the mode with conformal momentum ${\bf k}$ at the time $t_k$ when this mode becomes unstable. We will discuss the range of the amplitudes $A(\mathbf{k})$ later on. We stress that the gravitational potential (\[gravpot\]) is generated by tachyonic perturbations rather than inhomogeneities in the ordinary matter. In this paper we calculate CMB multipoles generated by the gravitational potential (\[gravpot\]). In the Lorentz-violating case, the mechanism we discuss gives rise to the contribution to the CMB anisotropy spectrum which is quite different from the standard spectrum coming from adiabatic scalar perturbations generated, e.g., at inflationary stage (for the latter see, e.g., Refs. [@book1; @book2; @book3]). As we will see below, the tachyonic contribution leads to potentially observable features in the CMB spectrum at relatively low multipoles. On the other hand, there are hints towards the existence of deviations in the observed spectrum [@Komatsu:2008hk; @Nolta:2008ih] from the predictions based on flat (Harrison-Zeldovich) or almost flat spectrum of primordial adiabatic perturbations. So, one is tempted to employ tachyonic instabilities to explain these deviations. We add the contribution due to the tachyonic perturbations to the standard contribution of the adiabatic modes and compare the result with the observed spectrum. We show that the tachyonic contributions slightly improve the agreement between the theory and data, but this improvement is statistically insignificant. So, our analysis only leads to limits on the parameters of the tachyonic perturbations. In the Lorentz-invariant case, tachyonic perturbations contribute to the lowest multipoles only. An interesting possibility here is that they may generate the entire observed dipole without getting in conflict with measurements of higher multipoles. This paper is organized as follows. We begin with preliminaries on the cosmological model in section \[sec:background\], and then discuss in detail the growth of the gravitational potential due to the tachyonic modes in section \[sec:growth\], first in the Lorentz-violating model and then in the Lorentz-invariant one. In section \[sec:multipoles\] we calculate the contributions to the CMB multipoles, again distinguishing Lorentz-violating and Lorentz-invariant cases. We compare our results with the data in section \[sec:data\], and conclude in section \[sec:conclusion\]. The cosmological model {#sec:background} ====================== The background space-time we consider in this paper corresponds to the “almost” standard cosmological model with the only special feature being that the accelerated expansion of the Universe is driven by phantom energy instead of the cosmological constant. The background metric is that of the spatially flat expanding Universe, $$ds^2 = dt^2 - a^2(t)d\mathbf{x}^2.$$ The scale factor $a(t)$ is determined by the Friedmann equation, which can be written as follows: $$\left( \frac{\dot a(t)}{a(t)} \right)^{2}=H_{0}^{2} \left[ \Omega_m \left( \frac{a(t_0)}{a(t)} \right)^3 + \Omega_p \left( \frac{a(t)}{a(t_0)} \right)^{-3(1+w_{p})} \right], \label{fried1}$$ where $H_{0}$ is the present value of the Hubble parameter, $a(t_0) = a_0 = 1$ is the present value of the scale factor, the dot denotes the derivative with respect to cosmic time $t$, $\Omega_p$ and $w_p$ refer to phantom energy. We assume for simplicity that $w_p \equiv p_p/\rho_p$ is independent of time; we will see in what follows that the effects we discuss are largely independent of $w_p$, so this assumption is not restrictive. The values we use in this paper are $\Omega_m=0.27$, $\Omega_p=0.73$. According to observational data [@Komatsu:2008hk] the parameter $w_p$ belongs to the interval $-1.38 < w_{p} < -0.86$. When calculating the CMB multipoles, we work with conformal time $\eta$ instead of cosmic time $t$, $$\eta(t) = \int\limits_0^t \dfrac{d\hat t}{a(\hat t)}.$$ Equation (\[fried1\]), written in terms of conformal time, has the form $$H_0 d\eta = \dfrac{ da}{\sqrt{a}\sqrt{\Omega_m + \Omega_p a^{-3w_p}}}.$$ Growth of perturbations {#sec:growth} ======================= The gravitational potential of a mode in the tachyonic regime is given by eq. (\[gravpot\]). The exponent $$N(t, k) = \int\limits_{t_k}^t \; d\hat t \; \omega(\hat t) \label{Eq/Pg5/1:PhanTach}$$ determines the growth of the potential in both Lorentz-violating model (\[dispers\]) and Lorentz-invariant one (\[Eq/Pg3/1:PhanTach\]). Properties of this function are different in the two cases, however. Lorentz-violating model. {#sub:lv-growth} ------------------------ In the Lorentz-violating model, the function (\[Eq/Pg5/1:PhanTach\]) can be written as an integral over the scale factor, $$N(a, k) = \sqrt{\alpha }\frac{M}{H_{0}}\sqrt{\nu }\int\limits_{a_k}^a d\hat a\dfrac{\sqrt{\hat a-\nu}}{\sqrt{\hat a}\sqrt{\Omega _{m} + \Omega _{p} \hat a^{-3w_p}}}, \label{fold}$$ where we have introduced dimensionless wave number $${\bf \nu } = \dfrac{k}{M}.$$ A few comments are in order. First, for a given mode the tachyonic regime begins when the integrand in (\[fold\]) becomes real, i.e., at $a = \nu$. Since $a \leq a_0 =1$, the maximum value of $\nu$ for tachyonic modes is $\nu =1$, that corresponds to the modes that are entering the tachyonic regime today. In fact, because of the Hubble friction, the exponential growth of the tachyonic mode starts not quite at the time when $a=\nu$, namely, it starts at the time $t_k$ when the tachyonic “frequency” becomes comparable to the Hubble parameter, $$\omega (t_{k}) \simeq H(t_{k}). \label{Eq/Pg6/1A:Phantom}$$ The lower limit of integration in (\[fold\]) is the value of the scale factor at that time, $a_k = a(t_k)$. Second, the integral in (\[fold\]) is a dimensionless smooth function of its arguments $\{a,\nu \}$, both of which do not exceed 1. Therefore, this integral is not parametrically large or small. On the other hand, the amplitude $A(\mathbf{k})$ is small (see below), so the effect of the tachyonic instability may be considerable only if $N(a,k)$ is sufficiently large. This can happen if $$\sqrt{\alpha }M>H_{0}. \label{Eq/Pg6/1:Phantom}$$ We assume in what follows that this inequality is indeed valid. Third, the integral in (\[fold\]) is saturated near its upper limit, and because of (\[Eq/Pg6/1:Phantom\]) this integral is practically insensitive to the lower limit of integration. Therefore, it is an excellent approximation to set $a_{k}=\nu $, that is, approximate the time $t_k$ by the time at which the dispersion relation becomes tachyonic. In fact, by solving eq. (\[Eq/Pg6/1A:Phantom\]) numerically, we have found that $a_{k} = \nu $ with precision of order $10^{-5}$ in a wide range of values of $\nu $ ($1\geq \nu \geq 0.05$). By setting $a_{k}=\nu $ in (\[fold\]) one finds that the dependence of $N$ on the parameters $M$ and $\alpha $ factors out $$\begin{aligned} N(a,\nu )&=&\sqrt{\alpha }\frac{M}{H_{0}}\mathcal{ N}(a,\nu )\; , \nonumber\\ \mathcal{ N}(a,\nu )&=&\sqrt{\nu }\int\limits_{\nu }^a d\hat a\dfrac{\sqrt{\hat a-\nu}}{\sqrt{\hat a}\sqrt{\Omega _{m} + \Omega _{p} \hat a^{-3w_p}}}\; . \nonumber\end{aligned}$$ Fourth, the function $\mathcal{ N}(a, \nu)$ is obviously a growing function of the scale factor $a$, ![: The growth function $\mathcal{ N}(a,\nu )$ as a function of $\nu$ at the values of the scale factor $a=0.8$, $a=0.9$ and $a=1.0$ (dashed, dotted and solid lines, respectively). Clearly, $\mathcal{ N}$ has a maximum at $\nu \simeq 0.2$ and grows considerably as the scale factor approaches its present value $a=1$. [**Right**]{}: The growth factor $\exp[N(a,\nu )]$ as a function of the scale factor $a$ at $\nu = \nu_{\mathrm{max}} \simeq 0.2$. The function $\exp[N(a,\nu)]$ becomes comparable to its present value only at $a \gtrsim 0.95$. Calculations are done for $w_{p} = -1$, $\sqrt{\alpha }M/H_{0}=100$. \[fig1\] ](fig1) as shown in the left panel of Fig. \[fig1\]. This is especially relevant since the gravitational potential (\[gravpot\]) depends on $N$ exponentially. We present in the right panel of Fig. \[fig1\] the dependence of $\exp[N(a, \nu)]$ on the scale factor at $\nu =0.2$; the plot is given for the values of parameters[^2] such that $\sqrt{\alpha }M/H_0 =100$. It is clear that the major effect of tachyonic modes occurs at late times. Finally, let us discuss the dependence of $\mathcal{ N}(a, \nu)$ on the parameter $w_{p}$. We show in Fig. \[fig3\] the function $\mathcal{ N}(a,\nu)$ at $a=1$ for different values of this parameter. Clearly, the dependence on $w_{p}$ is weak; in what follows we take $w_{p}=-1$ for definiteness. ![Function $\mathcal{ N} (a,\nu) $ at $a=1$ for different values of the equation of state parameter of phantom energy, $w_p = -1.0$, $w_p= -1.17$ and $w_p=-1.33$ (solid, dotted and dashed lines, respectively). []{data-label="fig3"}](fig2) The growth function $\mathcal{ N}(a, \nu)$ as a function of $\nu $ has a pronounced maximum at $\nu _{\mathrm{max}}=0.2$. The existence of a maximum of $\mathcal{ N}(a, \nu)$ is a fairly generic property independent of the particular form of the parametrization (\[dispers\]), provided that $\omega ({\bf p})$ vanishes at ${\bf p}=0$. Indeed, the tachyonic regime begins at $k/a(t)=M$. This implies, in particular, that $\mathcal{ N}(a, \nu)$ vanishes at $\nu =1$. Now, under the assumption that $\omega ({\bf p}=0) = 0$ one observes that $\mathcal{ N}(a, \nu)$ vanishes at $\nu =0$ as well. Hence, $\mathcal{ N}(a, \nu)$ has a maximum at some intermediate $\nu$. In other words, modes of high conformal momentum have not entered the tachyonic regime yet. On the other hand, $\mathcal{ N}(a, \nu)$ is small also at low conformal momenta, since the tachyonic regime begins too early for the corresponding modes. At that time the scale factor increases too rapidly; the function $\mathcal{ N}(a, \nu)$ does not have enough time to grow by the epoch when the physical momentum $k/a(t)$ gets close to zero and the growth terminates. The maximum growth occurs for intermediate momenta. We note in passing that the situation is different in the Lorentz-invariant model (\[Eq/Pg3/1:PhanTach\]), since in that case $\omega ({\bf p}=0) = M \neq 0$, and the maximum growth takes place at the lowest spatial momenta. Besides the growth function $N$, the gravitational potential (\[gravpot\]) is determined by the amplitude $A(\mathbf{k})$ at the time the perturbations enter the tachyonic regime. So, we have to estimate this amplitude at $k/a = M$. We assume that the tachyonic perturbations are generated from vacuum fluctuations. Hence, the amplitude is the Gaussian random field with zero expectation value, $\langle A({\bf k}) \rangle = 0$. This field is completely determined by its two-point correlation function, which we parametrize as follows: $$\langle A({\bf k}) A^*({\bf k}^{\prime} ) \rangle = \dfrac{f(k)}{k^3} \frac{H_{0}^2 \delta X^{2}}{M_{PL}^{2}M^{2}} \delta \left( {\bf k} - {\bf k^{\prime}} \right), \label{Eq/Pg7/1:PhantomP}$$ where $\delta X$ is the primordial perturbations of the phantom field. Let us clarify the form chosen for the overall factor. We recall that the gravitational potential is generated by the perturbations of the phantom field, which in turn is assumed to cause the accelerated cosmological expansion. This implies, in particular, that the energy density of the homogeneous phantom field is of the order of the present critical density $$\rho _{p}\sim M_{PL}^{2}H_{0}^{2}. \label{Eq/Pg8/1:Phantom}$$ On the other hand, this energy density can be estimated as $$\rho _{p}\sim M_{X}^{2}X^{2}, \label{Eq/Pg8/2:Phantom}$$ where $M_{X}$ is the mass scale characteristic of the phantom field and $X$ estimates the value of this field. By comparing eq. (\[Eq/Pg8/1:Phantom\]) to eq. (\[Eq/Pg8/2:Phantom\]) we obtain an estimate for this value, $$X\sim \frac{M_{PL}H_{0}}{M_{X}}. \label{Eq/Pg13/1:tachyon}$$ Let us assume that all dimensionful parameters in the phantom Lagrangian are of order of the mass parameter $M$ entering the dispersion relation (\[dispers\]), so that $M_{X}\sim M$. Therefore, $$\delta \rho _{p}\sim M_{X} ^{2}X\delta X\sim M M_{PL}H_{0}\delta X.$$ Now, we make use of the Poisson equation for the gravitational potential, $$\triangle\Phi \sim \frac{\delta \rho _{p}}{M_{PL}^{2}},$$ and find that at physical momenta of order $M$ the gravitational potential is of order $$\Phi \sim \frac{H_{0}\cdot\delta X}{M_{PL}\cdot M} \; .$$ This gives rise to the overall factor in (\[Eq/Pg7/1:PhantomP\]). The dimensionless function $f(k)$ in (\[Eq/Pg7/1:PhantomP\]) parametrizes possible deviations from the flat spectrum; it is likely that this function is model dependent. However, we have seen that the growth factor $\exp[N(a, \nu)]$ is peaked at $\nu = \nu_{max}$. This implies that the integrals over conformal momenta, determining CMB anisotropies, are saturated in a narrow region of $k$. Hence, if $f(k)$ is sufficiently smooth, our final results are insensitive to its shape, so we can set $f(k)=C$ where $C$ is a constant. This constant may be somewhat different from 1, as it may contain, e.g., a power of $H_{0}/M$; it cannot, however, contain extra factors involving the Planck mass. We will see that our final results are determined by the interplay between the large growth factor $\mbox{exp}(N)$ and small factor $\propto M_{PL}^{-1}$ in $A({\bf k})$, so a possible deviation of $C$ from unity is unimportant in the end. Furthermore, we will see that the shape of the tachyonic contribution to the CMB anisotropy spectrum (the position of the maximum and width) is almost independent of $M$; the parameter $M$, as well as $\delta X$, determine the overall magnitude only. Therefore, the constant $C$ can be set equal to 1 by redefinition of the parameter $M$. In view of these observations we write for the two-point correlator, without loss of generality, $$\langle A({\bf k}) A^*({\bf k}^{\prime} ) \rangle = \frac{H_{0}^2 \delta X^{2} }{M_{PL}^{2}M^{2}} \dfrac{1}{k^3} \delta \left( {\bf k} - {\bf k^{\prime}} \right) \; . \label{Eq/Pg7/1:Phantom}$$ It is this expression that will be used in what follows. Note that the amplitude $A(\mathbf{k})$ is small, so the linearized treatment of the problem is legitimate even for exponentially growing gravitational potential. Let us now discuss possible range of the amplitudes $\delta X$. The lowest value is obtained by assuming that the perturbations of the phantom field are in the vacuum state just before the tachyonic regime sets in. Since we are interested in momenta of order $M$ at that time, their amplitude is estimated as $\delta X\sim M$. The largest conceivable primordial perturbations are those generated at inflationary epoch: their amplitude is then comparable to the amplitudes of other (nonconformal) light fields, $\delta X\sim H_{\mathit{infl}}/(2\pi )$, where $H_{\mathit{infl}}$ is the Hubble parameter some 60 $e$-foldings before the end of inflation [@book1; @book2; @book3]. In view of the observational constraint $H_{\mathit{infl}}<1\cdot 10^{-5}M_{PL}$, we estimate the maximum value of $\delta X$ as $\delta X\sim 10^{-5}M_{PL}$. We will see that the interesting range of the parameter $M_{X}\sim M$ is roughly $M\lesssim 10^{3}H_{0}$, so the latter amplitude is small compared to the background value (\[Eq/Pg13/1:tachyon\]) of the phantom field. In what follows we present the results in the two extreme cases, $$\frac{H_{0}\delta X}{MM_{PL}}\sim \frac{M}{M_{PL}} \label{a}$$ and $$\frac{H_{0}\delta X}{MM_{PL}}\sim 10^{-5}\frac{H_{0}}{M}. \label{b}$$ The first case is obtained for vacuum amplitudes of fluctuations $\delta X$ before they get tachyonic, and we neglected the relatively mild difference between $M$ and $H_{0}$. The second case corresponds to the generation of primordial phantom perturbations at inflation with the maximum possible $H_{\mathit{infl}}$. We will see that the fact that these values differ by many orders of magnitude, the properties of the contributions to the CMB spectrum are qualitatively similar in the two cases. Lorentz-invariant model. {#sub:li-growth} ------------------------ The analysis performed in section \[sub:lv-growth\] is straightforwardly repeated in the case of Lorentz-invariant dispersion relation (\[Eq/Pg3/1:PhanTach\]). There are two properties, however, that make this case different from the Lorentz-violating one. The first property has to do with the behaviour of the function $\mathcal{ N}(a,\nu )$ that enters the growth factor, $$N(a,\nu )=\frac{M}{H_{0}}\mathcal{ N}(a,\nu ) \; . \label{Eqn/Pg10/1A:PhanTach}$$ In the Lorentz-invariant case one has $$\mathcal{ N}(a,\nu )=\int\limits_{\nu }^a d\hat a\dfrac{\sqrt{\hat a^{2}-\nu^{2}}}{\sqrt{\hat a}\sqrt{\Omega _{m} + \Omega _{p} \hat a^{-3w_p}}} \; . \label{Eqn/Pg10/1:PhanTach}$$ Unlike in the Lorentz-violating model, $\mathcal{ N}(a, \nu )$ as a function of $\nu$ monotonously decreases as $\nu$ increases, so the maximum of $\mathcal{ N}(a, \nu )$ is at $\nu=0$. It is therefore instructive to find $\mathcal{ N}(a, \nu )$ at small $\nu$. To this end, we write for the integral (\[Eqn/Pg10/1:PhanTach\]) $$\mathcal{ N}=\mathcal{ N}_{0}+\mathcal{ N}_{1}+\mathcal{ N}_{2},\nonumber$$ where $$\begin{aligned} \mathcal{ N}_{0}&=&\int\limits_{\nu }^a d\hat a\dfrac{\hat{a}}{\sqrt{\hat{a}}\sqrt{\Omega _{m} + \Omega _{p} \hat a^{-3w_p}}}, \label{Eqn/Pg10/2:PhanTach}\\ \mathcal{ N}_{1}&=&\int\limits_{\nu }^\infty d\hat a\dfrac{\sqrt{\hat{a}^{2}-\nu ^{2}}-\hat{a}}{\sqrt{\hat{a}}\sqrt{\Omega _{m} + \Omega _{p} \hat a^{-3w_p}}}, \label{Eqn/Pg10/3:PhanTach}\\ \mathcal{ N}_{2}&=&-\int\limits_{a }^\infty d\hat a\dfrac{\sqrt{\hat{a}^{2}-\nu ^{2}}-\hat{a}}{\sqrt{\hat{a}}\sqrt{\Omega _{m} + \Omega _{p} \hat a^{-3w_p}}}. \label{Eqn/Pg10/4:PhanTach}\end{aligned}$$ The integrals (\[Eqn/Pg10/3:PhanTach\]), (\[Eqn/Pg10/4:PhanTach\]) are convergent, since at large $\hat{a}$ one has $$\sqrt{\hat{a}^{2}-\nu ^{2}}-\hat{a}\simeq-\frac{\nu ^{2}}{2\hat{a}}.$$ Furthermore, the latter expression shows that at small $\nu$ the integral (\[Eqn/Pg10/4:PhanTach\]) is of order $\nu ^{2}$. The integral (\[Eqn/Pg10/3:PhanTach\]) can be written as $$\mathcal{ N}_{1}=\nu ^{3/2}\int \limits_{1}^{\infty }d y \frac{\sqrt{y^{2}-1}-y}{\sqrt{y}\sqrt{\Omega _{m}+\Omega _{p}\nu ^{-3w_{p}}y^{-3w_{p}}}}.$$ Since $w_{p}<0$, the term with $\Omega _{p}$ in the integrand can be neglected at small $\nu $, and then the remaining integral is straightforwardly evaluated. Finally, the integral $\mathcal{ N}_{0}$ is readily calculated at[^3] $w_{p}=-1$, $$\begin{aligned} \mathcal{ N}_{0}&=&\int \limits_{\nu }^{a}d\hat{a}\frac{\sqrt{\hat{a}}}{\sqrt{\Omega _{m}+\Omega _{p}\hat{a}^{3}}}=\frac{2}{3\sqrt{\Omega _{p}}}\left(\mbox{Arcsinh}\sqrt{a^{3}\frac{\Omega _{p}}{\Omega _{m}}} - \mbox{Arcsinh}\sqrt{\nu ^{3}\frac{\Omega _{p}}{\Omega _{m}}} \right) \nonumber\\ &=&\frac{2}{3\sqrt{\Omega _{p}}}\,\mbox{Arcsinh}\sqrt{a^{3}\frac{\Omega _{p}}{\Omega _{m}}} - \frac{2}{3\sqrt{\Omega _{m}}}\nu ^{3/2} +\mathcal{ O}(\nu ^{9/2})\; . \nonumber\end{aligned}$$ With all contributions included, we obtain finally that at small $\nu$ $$\begin{aligned} \mathcal{ N}(a,\nu )=\frac{2}{3\sqrt{\Omega _{p}}}\,\mbox{Arcsinh}\sqrt{a^{3}\frac{\Omega _{p}}{\Omega _{m}}} -\frac{\mathcal{ C}}{\sqrt{\Omega _{m}}}\nu ^{3/2}+\mathcal{ O}(\nu ^{2}), \label{Eqn/Pg11/1:PhanTach}\\ \mathcal{ C}=\frac{2}{3}-\int \limits_{1}^{\infty }dy\frac{\sqrt{y^{2}-1}-y}{\sqrt{y}} =\sqrt{\pi }\frac{\Gamma (-3/4)}{\Gamma (-1/4)}\simeq 1.75.\end{aligned}$$ Note that the expressions (\[Eqn/Pg10/1A:PhanTach\]) and (\[Eqn/Pg11/1:PhanTach\]) imply that the tachyonic contribution is non-negligible only for $M>H_{0}$. This is the analog of the inequality (\[Eq/Pg6/1:Phantom\]). It is clear from eq. (\[Eqn/Pg11/1:PhanTach\]) that unlike in the Lorentz-violating case, the growth factor $\mbox{exp}[N(a,\nu)]$ is peaked at $\nu = 0$. This is due to the fact that low-momentum modes are tachyonic already at early times, and their “frequency” $\omega=M$ is not small. However, the modes of very low spatial momenta $k =M\nu$ contribute to the [*monopole*]{} CMB harmonic only. This contribution strongly depends on the primordial spectrum of the tachyonic perturbations, i.e., on the shape of the function $f(k)$ in (\[Eq/Pg7/1:PhantomP\]); this is the second special property of the Lorentz-invariant model. Now, the monopole contribution merely renormalizes the average CMB temperature; it is not directly measurable and will not be discussed in this paper. Multipoles with $l\neq 0$ are less model-dependent: we will see in section \[sub:li-multi\] that the integrals over momenta are saturated in a relatively narrow region where $N(a,\nu)$ is not damped, i.e., in the region where $$\nu \sim \left(\frac{H_{0}}{M} \right)^{2/3} \; . \label{Eq/Pg11/1:PhanTach}$$ We assume that $f(k)$ does not change much in an interval $\Delta k \sim k$ for momenta belonging to the region (\[Eq/Pg11/1:PhanTach\]). Then for calculating the CMB multipoles with $l\neq 0$, we can still use the spectrum (\[Eq/Pg7/1:Phantom\]). The above discussion of the primordial amplitude $\delta X$ holds in the Lorentz-invariant case as well. So, we will again concentrate on the two extreme cases (\[a\]) and (\[b\]). CMB multipoles {#sec:multipoles} ============== The contribution to CMB anisotropy we are interested in is generated fairly recently, when the tachyon-induced gravitational potential becomes sizeable. Therefore, the only phenomenon responsible for this contribution is the integrated Sachs-Wolfe effect. It is clear from Fig. \[fig1\] that in the Lorentz-violating model this effect operates at the late cosmological epoch beginning at $z\sim 0.05$ ($a \sim 0.95$). A similar picture holds in the Lorentz-invariant case. Let us use the standard notation for the temperature anisotropy $$\Theta (\mathbf{n})=\frac{T(\mathbf{n})-T_{0}}{T_0} \; ,$$ where ${\bf n}$ is the direction of observation and $T_0$ is the average CMB temperature at present. The integrated Sachs-Wolfe effect then reads (see, e.g., Ref. [@Giov]) $$\Theta (\mathbf{n})=2\int \limits_{0}^{\eta _{0}}\!d\eta \left.\frac{\partial \Phi (\eta ,\mathbf{x})}{\partial \eta }\right|_{\mathbf{x}=\mathbf{n}(\eta _{0}-\eta )}, \label{Eq/Pg9/2:Phantom}$$ where $\eta_0$ is the present time, and in view of the above discussion we set the lower limit of integration equal to zero instead of the time of last scattering. The integrand in (\[Eq/Pg9/2:Phantom\]) exponentially grows toward the present epoch. CMB anisotropy is characterized by the multipoles $$C_l = \dfrac{1}{2l+1} \sum\limits_{m = -l}^{m = l} \langle |a_{lm}|^2\rangle , \label{Eq/Pg9/3:Phantom}$$ where $a_{lm}$ are the coefficients of the decomposition of the anisotropy over spherical harmonics, $$a_{lm} = \int \; d{\bf n} \; \Theta ({\bf n}) \; Y_{lm}({\bf n}). \label{multip}$$ Making use of eqs. (\[gravpot\]), (\[Eq/Pg7/1:Phantom\]) and (\[Eq/Pg9/2:Phantom\]) and performing the angular integration one obtains the following expression for the multipoles in terms of integrals over conformal momenta: $$C_l = \dfrac{8 H_{0}^2\delta X^{2}}{\pi M^{2}M^2_{PL}} \; \; \int\limits^1_{0} \; \dfrac{d \nu}{\nu} \Delta^2_l(\nu) \label{multip2}$$ The quantity $\Delta^2_l(\nu)$ is the analog of the power spectrum. It is expressed in terms of the integral over conformal time, $$\Delta_{l}(\nu )=\int \limits_{\eta _{k}}^{\eta _{0}}\!d\tau \omega (\tau )a(\tau )\exp\left\{N[a(\tau ), \nu] \right\}j_{l}[M\nu (\eta _{0}-\tau )], \label{Eq/Pg1/1:appendix}$$ where $$j_l(x) = \sqrt{\dfrac{\pi}{2 x}}J_{l + \frac{1}{2}}(x)$$ is the spherical Bessel function of the first kind. The integrals (\[Eq/Pg1/1:appendix\]) and (\[multip2\]) have different properties in Lorentz-violating and Lorentz-invariant models. Lorentz-violating model. {#lorentz-violating-model.} ------------------------ We plot in the left panel of Fig. \[fig4\] the tachyonic contribution to the angular spectrum of CMB temperature in the Lorentz-violating model (\[a\]) with $\alpha = 1 \cdot 10^{-3}$ and $M/H_{0} = 9770$, using the standard quantity $$D_l = \dfrac{l(l+1)}{2\pi}C_l \; .$$ A very similar plot is obtained for the Lorentz-violating model (\[b\]) with $\alpha =1\cdot 10^{-3}$, $M/H_{0}=726$. This is a very general situation: the tachyonic contributions to the CMB spectrum in models (\[a\]) and (\[b\]) at the same value of $\alpha $ are virtually indistinguishable if the value of $M$ appropriately scaled down by about an order of magnitude. The reason for that will become clear later, see eqs. (\[Eq/Pg3/1:appendix\]), (\[Eq/Pg14/1:Phantom\]). The analysis of this section applies to both models (\[a\]) and (\[b\]); we will illustrate our results using the model (\[a\]) for definiteness. It is clear that the spectrum shown in Fig. \[fig4\] has a rather narrow maximum at $l=l_{\mathrm{max}}$ ($l_{\mathrm{max}} \approx 7$ in this example). We will see below that the position of the maximum is determined solely by the parameter $\alpha$, and that $l_{\mathrm{max}}$ grows as $\alpha$ decreases, see eq. (\[Eq/Pg14/1:Phantom\]). ![: The tachyonic contribution to the angular spectrum of CMB temperature in the Lorentz-violating model (\[a\]) with $\alpha = 1 \cdot 10^{-3}$ and $M/H_{0} = 9770$. [**Right**]{}: Same, but for $\alpha = 1.0$ and $M/H_{0} = 315$. []{data-label="fig4"}](fig3) For $\alpha = 1$, the spectrum is shown in the right panel of Fig. \[fig4\]. It is clear from this figure that at large enough $\alpha$, the tachyonic perturbations contribute to the lowest multipoles only. It is also clear that the dipole and quadrupole components differ by one order of magnitude only. Hence, we disregard in what follows the dipole component, as for realistic quadrupole it is negligible compared to the observed dipole, which is supposedly due to the motion of the Earth in the CMB reference frame. We note in passing that the situation is different in the Lorentz-invariant model: it is meaningful to discuss the dipole component in that case, see sections \[sub:li-multi\] and \[sub:li-exp\]. The multipoles $C_l$ can be calculated analytically in the regime $\alpha (M/H_{0})^{2} \gg 1$, see (\[Eq/Pg6/1:Phantom\]). As we discussed in section \[sub:lv-growth\], the function $\mbox{exp}[N(a,\nu )]$ rapidly grows with the scale factor, and as a function of $\nu$ it has a peak at $\nu _{\mathrm{max}}\simeq 0.2$. Thus, the main contribution into the integral (\[Eq/Pg1/1:appendix\]) comes from late times, while the integral (\[multip2\]) is saturated at $\nu \approx \nu _{\mathrm{max}}$ (it is important at this point that $\nu _{\mathrm{max}}$ is different from both 0 and 1). This means, in the first place, that the lower limit of integration in (\[Eq/Pg1/1:appendix\]) may be set equal to zero. Second, one can make use of expansion of the function $N(a, \nu)$ near $a=1$, $$N[a(\eta ), \nu]=N(1, \nu)-\omega(\nu) \cdot(\eta _{0}-\eta ), \label{Eq/Pg1/2:appendix}$$ and for $\nu=\nu _\mathrm{max}\simeq 0.2$ we have $$N(1,\nu _{\mathrm{max}})=0.39 \sqrt{\alpha }\frac{M}{H_{0}}.$$ Third, one can set the pre-exponential factor $\omega(\eta ) a(\eta )$ in (\[Eq/Pg1/1:appendix\]) equal to its value at $a=1$. In this way we obtain $$\Delta_{l}(\nu )=\sqrt{\alpha }\sqrt{\frac{1-\nu}{\nu } }\exp\left[ N(1,\nu )\right]\int \limits_{0}^{\nu M\eta_{0} }\!dx\exp\left(-\sqrt{\alpha }\sqrt{\frac{1-\nu }{\nu }}x \right)j_{l}(x), \label{Eq/Pg1/3:appendix}$$ where we introduced the integration variable $x=\nu M(\eta _{0}-\eta )$. Finally, let us recall that $\eta _{0}\sim 1/H_{0}$. Then the upper limit of integration in (\[Eq/Pg1/3:appendix\]) is $\nu M\eta _{0}\sim \nu _{\mathrm{max}}M/H_{0}\gg 1$, so the integration may be extended to infinity. Then the integral in (\[Eq/Pg1/3:appendix\]) is calculated by making use of the formula [@GR-Abr] $$\begin{aligned} \int \limits_{0}^{\infty }\!dx\frac{J_{l+\frac{1}{2}}(x)}{\sqrt{x}}\exp\left(-\gamma x \right)&=&\frac{1}{(1+\gamma ^{2})^{\frac{1}{4}}}\Gamma (l+1)P_{-\frac{1}{2}}^{-l-\frac{1}{2}}\left[\frac{\gamma }{\sqrt{1+\gamma ^{2}}} \right]=\nonumber\\ &=&\frac{1}{(1+\gamma^{2})^{\frac{1}{4}}}\frac{\Gamma (l+1)}{\Gamma (l+\frac{3}{2})}\left[\frac{1-z}{1+z} \right]^{\frac{l}{2}+\frac{1}{4}} F\left(\frac{1}{2},\frac{1}{2},l+\frac{3}{2},\frac{1-z}{2}\right), \label{Eqn/Pg15/1:PhanTach}\end{aligned}$$ where $P_{-\frac{1}{2}}^{-l-\frac{1}{2}}$ is the Legendre function, $F$ is the hypergeometric function[^4], and $$z=\frac{\gamma}{ \sqrt{1+\gamma ^{2}}} =\sqrt{\frac{\alpha (1-\nu )}{\nu +\alpha (1-\nu )}}.$$ Because of the exponential dependence on $\nu $ of the factor $\exp [N(1,\nu )]$ in (\[Eq/Pg1/3:appendix\]), the integral in (\[multip2\]) can be evaluated in the saddle-point approximation, and we obtain $$\begin{aligned} D_l &=&\frac{2H_{0}^{2}\delta X^{2} }{M^{2}M_{PL}^{2}\nu_{\mathrm{max}}}\sqrt{\frac{1}{\pi |N''(1,\nu _{\mathrm{max}})|}}\exp(2N(1,\nu _{\mathrm{max}})) \cdot l(l+1)\cdot\left[ \frac{\Gamma (l+1)}{\Gamma (l+3/2)}\right]^{2}\times\nonumber\\ &&\times \frac{z^{2}}{\sqrt{1-z^{2}}}\left(\frac{1-z}{1+z} \right)^{l+1/2}F^2 \left(\frac{1}{2},\frac{1}{2},l+\frac{3}{2},\frac{1-z}{2} \right) \; , \label{exact}\end{aligned}$$ where $z=z(\nu _{\mathrm{max}})$, $N''(1,\nu )$ is the second derivative with respect to $\nu $. At $\nu _{\mathrm{max}} \simeq 0.2$ we have $$\begin{aligned} D_l & = & \mathcal{A}_0 \cdot \left[\dfrac{\Gamma(l+1)}{\Gamma(l+3/2)}\right]^2 \dfrac{l(l+1)}{\sqrt{1 + 4\alpha}} \cdot \left(\dfrac{1-z}{1+z}\right)^{l+1/2} F^2\left(\frac{1}{2}, \frac{1}{2}; l+\frac{3}{2}, \frac{1-z}{2}\right), \label{exact1}\\ z&=&\dfrac{2\sqrt{\alpha}}{\sqrt{1 + 4\alpha}},\nonumber\\ \mathcal{A}_0 &=& 9.4 \cdot \alpha ^{3/4}\left(\frac{H_{0}}{M} \right)^{5/2}\frac{\delta X^{2}}{M_{PL}^{2}} \cdot \exp\left[0.78\sqrt{\alpha}\frac{M}{H_0}\right]. \label{C0value}\end{aligned}$$ This is the desired analytical expression for the multipoles; it works with the precision which is certainly sufficient for our purposes.. At $\alpha \ll 1$ and $l>1$, the expression (\[exact\]) simplifies to $$D_l=\frac{2\alpha H_{0}^{2}\delta X^{2} }{M^{2}M_{PL}^{2}}\frac{1-\nu _{\mathrm{max}}}{\nu _{\mathrm{max}}^{2}}\sqrt{\frac{1}{\pi |N''(1,\nu _{\mathrm{max}})|}}\exp(2N(1,\nu _{\mathrm{max}})) \cdot (l+1)\cdot\exp\left(\!-2l\sqrt{\alpha }\sqrt{\frac{1-\nu_{\mathrm{max}} }{\nu_{\mathrm{max}} }} \right). \label{Eq/Pg3/1:appendix}$$ Corrections to the latter formula are $\mathcal{ O}(1/l)$ and they are numerically small even for $l=2$. A few comments are in order. First, it is seen from (\[exact1\]) that the dependence on $l$ and on $M$ has factorized. Therefore, the position of the maximum $l_{\mathrm{max}}$ and the width of the peak in the spectrum depend on the parameter $\alpha $ and do not depend on the parameters $M$ and $\delta X$. In particular, for $\alpha \ll 1$ the maximum of the function (\[Eq/Pg3/1:appendix\]) is at $$l_{\mathrm{max}}=\frac{1}{2\sqrt{\alpha }}\sqrt{\frac{\nu _{\mathrm{max}}}{1-\nu _{\mathrm{max}}}}-1\simeq \frac{1}{4\sqrt{\alpha }}-1. \label{Eq/Pg14/1:Phantom}$$ The overall magnitude of the spectrum depends on $\alpha $, $M$ and $\delta X$. The dependence on $M$ is exponential, since $N(1,\nu _{\mathrm{max}}) \propto M$. This justifies the use of (\[Eq/Pg7/1:Phantom\]) for the primordial spectrum of perturbations of the gravitational potential. Second, the exponential dependence of $C_{l}$ on $l$ has the following interpretation. The problem has the characteristic time scale $\tau (k)\sim \omega^{-1} =(\sqrt{\alpha }M\sqrt{\nu (1-\nu )})^{-1} \ll H^{-1}_{0}$. This scale determines the time of the development of the tachyonic instability in a mode with momentum $k$. Since the growth function $N(\nu )$ has a maximum, the relevant modes have momenta near $k_{\mathrm{max}}=\nu _{\mathrm{max}}M$. Therefore, the gravitational potential is small at distances $r\gg \tau (k_{\mathrm{max}})$ along the light cone emanating from the observer (the tachyonic instability has not developed yet). On the other hand, at distances $r<\tau (k_{\mathrm{max}})$ the gravitational potential is almost constant in time (according to (\[Eq/Pg6/1:Phantom\]), the expansion of the Universe has a negligible effect, while the tachyonic instability gives rise to mild growth of the potential). In other words, the gravitational potential at $r< \tau (k_{\mathrm{max}})$ is the superposition of random, time-independent waves with almost constant amplitude and almost constant wavelength $2\pi/k_{\mathrm{max}}$. At $r> \tau (k_{\mathrm{max}})$ the amplitude of these waves decays as $\exp[-r/\tau(k _{\mathrm{max}})]$ as $r$ increases. The period of a wave located at distance $r$ is seen at an angle $\triangle\theta _{r}\simeq 2\pi/(rk_{\mathrm{max}})$. Hence, this wave contributes to the multipoles with $l\simeq rk_{\mathrm{max}}$. The multipoles $a_{lm}$ are not exponentially suppressed for $r<\tau (k_{\mathrm{max}})$, i.e., $l < 1/[k_{\mathrm{max}}\tau (k_{\mathrm{max}})]$, and are exponentially small in the opposite case. Recalling (\[Eq/Pg9/3:Phantom\]), one finds that this behaviour of $a_{lm}$ leads to the following dependence of $C_l$ on $l$, $$C_{l}\propto \exp\left(-\frac{2l}{k_{\mathrm{max}}\tau (k_{\mathrm{max}})} \right)= \exp\left(-2l\sqrt{\alpha }\sqrt{\frac{1-\nu _{\mathrm{max}}}{\nu _{\mathrm{max}}}} \right) \; ,$$ in complete agreement with (\[Eq/Pg14/1:Phantom\]). The fact that for relatively large $\alpha $ sizeable contributions are obtained by the lowest multipoles only (see Fig. \[fig4\]) can be seen directly from (\[Eq/Pg9/2:Phantom\]). Indeed, inserting (\[gravpot\]) into (\[Eq/Pg9/2:Phantom\]), making use of (\[Eq/Pg1/2:appendix\]), and integrating over time, we obtain $$\Theta (\mathbf{n})\sim\int \limits_{}^{}\!d^{3}k F(k,\omega ) A(\mathbf{k}) \frac{1}{\sqrt{\alpha }\sqrt{\frac{1-\nu }{\nu }} -i\frac{\mathbf{k \cdot n}}{k}}+\mbox{h.c.}, \label{nov29-1}$$ where $F(k,\omega )$ is a smooth function independent of the direction of $\mathbf{k}$. Recalling that $\nu \simeq \nu _{\mathrm{max}}=0.2$, we have $\sqrt{\alpha }\sqrt{(1-\nu)/\nu } >1$ for sufficiently large $\alpha $. In that case the denominator in (\[nov29-1\]) can be expanded in a series in ($\mathbf{k \cdot n}$), which just corresponds to the expansion in spherical harmonics. The $l$-th harmonic is thus suppressed as $ (\sqrt{\alpha }\sqrt{(1-\nu)/\nu })^{-l} $. This is relatively mild suppression, in accord with the right panel of Fig. \[fig4\]. At $\sqrt{\alpha }\sqrt{(1-\nu)/\nu } <1$ the expansion of the denominator in (\[nov29-1\]) is not legitimate, and one has to perform more sophisticated analysis leading to (\[exact1\]). Lorentz-invariant model. {#sub:li-multi} ------------------------ In the Lorentz-invariant case we make use of the expression (\[Eqn/Pg11/1:PhanTach\]) for the function $\mathcal{ N}(a,\nu )$ to calculate the integral (\[Eq/Pg1/1:appendix\]). We find $$\Delta_{l}(\nu )=\frac{\sqrt{\pi }}{2}\frac{\Gamma (l+1)}{\Gamma (l+3/2)}\left(\frac{\nu }{2} \right)^{l}\cdot\exp\left(\frac{2}{3}\frac{M}{H_{0}\sqrt{\Omega _{p}}}\,\mbox{Arcsinh}\sqrt{\frac{\Omega _{p}}{\Omega _{m}}}-\frac{\mathcal{ C}M}{H_{0}\sqrt{\Omega _{m}}}\nu ^{3/2} \right). \label{Eq/Pg17/1:PhanTach}$$ When obtaining this expression we used the fact that $\nu$ is small (see (\[Eq/Pg11/1:PhanTach\])), again extended the time integration to infinity in the same way as we have done after eq. (\[Eq/Pg1/3:appendix\]) and made use of in eq. (\[Eqn/Pg15/1:PhanTach\]). We also kept the leading terms in $\nu$ in the expression (\[Eq/Pg17/1:PhanTach\]). Inserting (\[Eq/Pg17/1:PhanTach\]) into the integral in the expression (\[multip2\]) for multipoles, and changing the integration variable, we arrive at the following integral: $$\left(\frac{H_{0}\sqrt{\Omega _{m}}}{2\mathcal{ C}M} \right)^{4l/3}\int \limits_{0}^{\frac{\mathcal{ C}M}{H_{0}\sqrt{\Omega _{m}}}}dx x^{4l/3-1}\mathrm{e}^{-x}=\left(\frac{H_{0}\sqrt{\Omega _{m}}}{\mathcal{ C}M} \right)^{4l/3} \gamma \left(\frac{4l}{3}, \frac{\mathcal{ C}M}{H_{0}\sqrt{\Omega _{m}}} \right), \label{Eq/Pg18/1:PhanTach}$$ where $\gamma(\beta ,x) $ is an incomplete $\Gamma $ function. We notice that the second argument of this function $$x = \frac{\mathcal{ C}M}{H_{0}\sqrt{\Omega _{m}}} \; ,$$ is large, $x \gg 1$, otherwise the overall factor in (\[multip2\]) makes the effect we discuss negligibly small. Hence, for $\beta \equiv 4l/3 \ll x$ we use the approximation $\gamma(\beta ,x) = \Gamma (\beta)$, and obtain the final formula for the multipoles $$D_l = \frac{H_{0}^2\delta X^{2}}{3\pi M^{2}M^2_{PL}}\frac{l(l+1)}{2^{2l-1}} \exp \left(\frac{4M}{3H_0}\frac{1}{\sqrt{\Omega_p}}\mbox{Arcsinh} \sqrt{\frac{\Omega_p}{\Omega_m}}\right) \left[ \dfrac{\Gamma(l+1)}{\Gamma(l+3 / 2)}\right]^2\! \left( \dfrac{H_0 \sqrt{\Omega_m}}{2\mathcal{ C}M}\right)^{4l \over 3} \!\!\!\!\Gamma\!\!\left (\frac{4l}{3}\right). \label{power1}$$ This formula is valid for relatively low multipoles, $l<\mathcal{ C}M/(H_{0}\sqrt{\Omega _{m}})$. In the opposite case the behaviour of the incomplete $\Gamma$ function is $$\gamma (\beta ,x)\simeq \frac{x^{\beta }}{\beta }\mathrm{e}^{-x}\,$$ Hence, the multipoles at large $l$ are negligibly small in the interesting case of large $x$. Because of the exponential factor in (\[power1\]), the low multipoles may be fairly large at large enough $M/H_0$. On the other hand, a large value of this parameter implies that the multipoles rapidly decay with $l$. Consider first the case (\[a\]) of small primordial amplitude $\delta X$. As we discuss in section \[sub:li-exp\], the interesting range is $M/H_0 \sim 100$, in which case the multipole $D_{l+1}$ is suppressed by a factor of about $10^{-3}$ as compared to $D_l$. So, it is sufficient to consider the dipole and quadrupole anisotropies only. With our parametrization, these are, respectively, $$\begin{aligned} D_1 &=& 6.4\cdot 10^{-125} \left(\dfrac{M}{H_0}\right)^{2/3} \exp \left(1.98\cdot \frac{M}{H_0}\right) \label{Eqn/Pg19/1:PhanTach}\\ D_2 &=& 4.1\cdot 10^{-126} \left(\dfrac{H_0}{M}\right)^{2/3} \exp \left(1.98\cdot \frac{M}{H_0}\right) \label{Eqn/Pg19/2:PhanTach}\\ \dfrac{D_1}{D_2} &=& 15.7 \left(\dfrac{M}{H_0}\right)^{4/3} \label{Eqn/Pg19/3:PhanTach}\end{aligned}$$ These expressions are in agreement, within $10\%$, with the values that we obtained numerically[^5]. Let us now turn to high primordial amplitude $\delta X$, case (\[b\]). In that case, the interesting range is $M/H_{0}\sim 10$, and $$D_2 = 2.7\cdot 10^{-14} \left(\dfrac{H_0}{M}\right)^{14/3} \exp \left(1.98\cdot \frac{M}{H_0}\right),$$ while eq. (\[Eqn/Pg19/3:PhanTach\]) remains valid. Comparison with the data {#sec:data} ======================== Overall, the data on the anisotropy of CMB temperature are in good agreement with the standard picture of adiabatic scalar perturbations whose primordial spectrum is close to the Harrison-Zeldovich one. Still, the observed angular spectrum may possibly show deviations whose nature is unclear. Our analysis was partially motivated by the desire to understand whether these deviations may be due to the contributions coming from the tachyonic perturbations of phantom energy. As we discuss in this section, the deviations [*cannot*]{} be explained in this way. So, our analysis enables us only to place limits on the parameters of the tachyonic perturbations. Lorentz-violating case. ----------------------- We begin with the Lorentz-violating model, and consider a wide range of the parameter $\alpha$, $$2.5\cdot 10^{-7} < \alpha < 1.0. \label{Eq/Pg14/1:Phantom15_09}$$ This range is representative: at $\alpha \gtrsim 1$ the main effect is in the lowest multipoles, whereas at $\alpha =2.5\cdot 10^{-7}$ the contribution is peaked at $l\sim l_{\mathrm{max}}\simeq 500$, see (\[Eq/Pg14/1:Phantom\]). As we pointed out above, we do not consider the dipole anisotropy in the Lorentz-violating case, as it is much smaller than the observed dipole anisotropy supposedly originating from the motion of the Earth. In our study we used the data on multipoles given in Ref. [@wmapcl] and organized as a table “$C_l$ vs. $l$.” These are not combined in bins, unlike the data usually presented. In the model we study, the multipoles are the sums of two terms, one due to adiabatic scalar perturbations and another due to tachyonic modes, $$C_l = C_l^{(ad)} + C_l^{(t)} \; . \label{dec2-1}$$ Making use of the code CMBFast [@cmbfast], we calculated the angular spectrum $C_l^{(ad)}$ generated by adiabatic scalar perturbations for various values of the spectral index $n_s$ in the range $0.8 \le n_s \le 1.5$, in the standard cosmological model with the following values of parameters: the Hubble constant $H_0 = 72~\mbox{km} \cdot \mbox{s}^{-1} \cdot \mbox{Mpc}^{-1}$, baryon plus CDM contribution to the present energy density $\Omega_{m} = 0.27$, contribution of hot dark matter $\Omega_{hdm} = 0$, dark energy contribution $\Omega_p \equiv \Omega_{\Lambda} = 0.73$, $^4\mbox{He}$ abundance $Y=0.24$, number of massless neutrino species $N_\nu=3$. We assumed that the tensor perturbations are absent. The second term $C_l^{(t)}$ in (\[dec2-1\]) was calculated by making use of the analytical expression (\[exact1\]). To compare the model with the data, we analyzed the difference between the measured and calculated multipoles, $$\epsilon_l = C_l^{(exp)} - C_l^{(ad)} - C_l^{(t)}.$$ The study of the moments and correlation properties of $\epsilon_l$ has shown that they are independent and their average is zero within statistical error. Furthermore, an $\chi^2$ estimate has shown that with $95\%$ probability the distribution of $\epsilon_l$ is Gaussian. To obtain the parameters of the theoretical spectrum, we used maximum likelihood method with the likelihood function $$\begin{aligned} F( \epsilon | \theta) &=& \prod\limits_{l=2, 600} f(\epsilon_l |\theta) \; , \nonumber \\ f(\epsilon_l |\theta) &=& \exp\left(-\dfrac{\epsilon_l^2(\theta)}{2 \sigma_l^2}\right) \; . \nonumber\end{aligned}$$ Here $\theta$ is the set of four parameters: the spectral index and amplitude of adiabatic perturbations, the amplitude $\mathcal{A}_0$ and parameter $\alpha$ of the tachyonic perturbations. In view of (\[Eq/Pg14/1:Phantom\]) and (\[Eq/Pg14/1:Phantom15\_09\]), we included in our analysis multipoles with $l\leq 600$ only. For each $\alpha$ from the range (\[Eq/Pg14/1:Phantom15\_09\]) we obtained the best fit values of the three other parameters. The variations of individual multipoles have been calculated with the use of the error estimations given in the third column of the table in Ref. [@wmapcl]. We show in Fig. \[fig6\] the best fit value of the tachyonic contribution as a function of the parameter $\alpha$. As a measure of this contribution we use the maximum in the anisotropy spectrum generated by the tachyonic perturbations, $$D_{\mathrm{max}} = \max_{l} \left[ \frac{l(l+1)}{2\pi} C^{(t)}_l \right] \; . \label{nov30-1}$$ This maximum is at $l=l_{\mathrm{max}}$ (as an example, $l_{\mathrm{max}} \approx 7$ in the left panel of Fig. \[fig4\]). It is clear from Fig. \[fig6\] that at $\alpha > 10^{-4}$ the best fit value is equal to zero, whereas at $\alpha < 10^{-4}$ it is considerably different from zero. This means that the tachyonic contribution improves the agreement between the theory and data. It is worth noting that the addition of this contribution moves the best fit value of the spectral index up from $n_{s}=0.96$ obtained in Ref. [@wmapcos]; in particular, for some values of $\alpha$ the best fit values of $n_s$ are larger than 1. ![ The best fit value of the tachyonic contribution to CDM temperature anisotropy as a function of $\alpha$. The parameter $D_{\mathrm{max}}$ is defined in (\[nov30-1\]). The maximum of $D_{\mathrm{max}}$ is at $\alpha= 3.0 \cdot 10^{-6}$, which corresponds to $l_{\mathrm{max}}\simeq 143$. \[fig6\] ](fig4) This improvement, however, is not statistically significant. We show in Fig. \[fig7\] the maximum likelihood function as a function of $D_{\mathrm{max}}$ at $\alpha=1.8 \cdot 10^{-6}$. It is clear from Fig. \[fig7\], that even though the best fit value of $D_{\mathrm{max}}$ is nonzero, the difference of the likelihood function at the best fit value and at $D_{\mathrm{max}}=0$ is small. The same is true for all values of $\alpha$ in the range considered. So, the data is consistent with the absence of the tachyonic contribution to the CMB temperature anisotropy. ![Likelihood function at the best fit values of the amplitude and spectral index of the adiabatic perturbations as a function of $D_{\mathrm{max}}$ at $\alpha = 1.8 \cdot 10^{-6}$. The dashed line shows the limit on $D_{\mathrm{max}}$ at 95 % confidence level. The dotted line corresponds to the best fit value of $D_{\mathrm{max}}$. The maximum value of the likelihood function is $F_{\mathrm{max}} = 0.577$. []{data-label="fig7"}](fig5) Thus, we can only place limits on the overall magnitude of the tachyonic contribution, $\mathcal{A}_0$, at various values of $\alpha$, which can then be translated into the limits on the physical parameter $M/H_0$. These limits, at $95~\% $ confidence level, are shown in Fig. \[fig8\] for both cases of low primordial amplitude (\[a\]) and high amplitude (\[b\]). ![: Upper limit on the amplitude $\mathcal{A}_0$ for $\alpha \in\left[1.0 \div 2.5\cdot 10^{-7}\right]$ at 95 % confidence level. [**Right**]{}: Upper limit on the parameter $M$ of the tachyonic perturbations in units of the Hubble constant, at 95 % confidence level, in the model (\[a\]) (solid line) and model (\[b\]) (dashed line). []{data-label="fig8"}](fig6.eps) Lorentz-invariant case. {#sub:li-exp} ----------------------- As we have seen in section \[sub:li-multi\], in the model with Lorentz-invariant spectrum the CMB multipoles generated by the tachyonic perturbations rapidly decrease as $l$ increases. Therefore, only two multipoles — the dipole and quadrupole — are relevant for comparison with the data. The measured dipole component of CMB temperature is [@wmap03_1] $d = 3.358 \pm 0.017$ mK, and the direction in the Galactic polar coordinate frame is $l = 263.86 \pm 0.04^{\circ}$, $b=48.24 \pm 0.10^{\circ}$. In the standard parametrization, the dipole anisotropy is $$D^{exp}_1 = \frac{1}{3\pi}\sum\limits_{m = -1}^{m = 1} |a_{1 m}|^2= 1.6\cdot 10^{-7}, \label{dip}$$ while the quadrupole component is given by $$D^{exp}_2 = \frac{3}{5\pi}\sum\limits_{m = -2}^{m = 2} |a_{2 m}|^2 = 2.9\cdot 10^{-11}. \label{quad}$$ To obtain conservative limits on the parameter $M$ of Lorentz-invariant tachyonic perturbations, we do not impose any priors on the contribution to the dipole anisotropy due to the motion of the Earth and on the quadrupole anisotropy generated by adiabatic perturbations. In the model with small primordial spectrum (\[a\]) we make use of the expression (\[Eqn/Pg19/1:PhanTach\]) and take into account the fact that the dipole has 3 degrees of freedom. In this way we find that the analysis of the dipole anisotropy leads to the limit $$\frac{M}{H_{0}}\le 135.9 \;\;\;\;\; \mbox{at}~~95~\%~~\mbox{c.l.}\;. \label{Eq/Pg24/1:PhanTach}$$ The limit coming from the quadrupole anisotropy is obtained by making use of eq. (\[Eqn/Pg19/2:PhanTach\]). It reads $$\frac{M}{H_{0}}\le 136.4\;\;\;\;\; \mbox{at}~~95~\%~~\mbox{c.l.}\;. \label{Eq/Pg24/2:PhanTach}$$ Interestingly, the limit coming from the dipole anisotropy is similar to that obtained from the quadrupole. One can turn this result around and speculate that the large observed dipole may be due to the tachyonic perturbations, with no contradiction to the data at higher multipoles. In our model the latter property is natural in the sense that the quadrupole and higher angular harmonics are small automatically. Another way to phrase this is to pretend that the observed dipole anisotropy is due to the tachyonic perturbations, i.e., equate (\[Eqn/Pg19/1:PhanTach\]) and (\[dip\]), and then calculate the contribution to the quadrupole from (\[Eqn/Pg19/2:PhanTach\]). This gives for the tachyonic contribution $D_2 = 1.5\cdot 10^{-11}$, which is safely below the observed value (\[quad\]). The octupole is suppressed by another three orders of magnitude. So, our model would serve as an alternative to other explanations [@GZ; @Turner; @PP; @JP; @lan96; @lan96a; @lan97] of the large dipole component of the CMB anisotropy, if such an explanation were needed (for observational aspects of this issue see Refs. [@Kamionkowski:2002nd; @Gordon:2007sk] and references therein). In the model with large primordial perturbations (\[b\]), the situation is different. In that case, the strongest limit on $M/H_{0}$ is obtained from the quadrupole $$\frac{M}{H_{0}}<8.6\;\;\;\;\; \mbox{at}~~95~\%~~\mbox{c.l.}\;.$$ As seen from eq. (\[Eqn/Pg19/3:PhanTach\]), the dipole is not so much enhanced as compared to quadrupole; at $M/H_{0}=8.6$ its value is $D_{1}\simeq 1\cdot 10^{-8}$, which is safely below the observed value (\[dip\]). On the other hand, the octupole and higher harmonics are still suppressed compared to the quadrupole by more than an order of magnitude. Discussion {#sec:conclusion} ========== In this paper we have considered the effects on the anisotropy of CMB temperature due to possible tachyonic perturbations of dark energy. Because of the exponential growth, these perturbations may generate large gravitational potential $\Phi$ at the recent cosmological epoch, and only at that epoch. This results in a sizeable Sachs-Wolfe effect. Note that the tachyonic perturbations we have discussed are unrelated to perturbations in baryons or dark matter, so their contribution to the CMB anisotropy does not correlate with the distribution of structure in the Universe. Our analysis was mostly motivated by the Lorentz-violating models of phantom energy. Hence, we have studied in detail the tachyonic perturbations with the dispersion relation (\[dispers\]). We have seen that their effect on the CMB angular spectrum has a pronounced maximum whose position depends on one of the parameters, $\alpha$, and is practically insensitive to other parameters. It is expected that similar shape of the angular spectrum is characteristic to a wide class of models with Lorentz-violating tachyonic perturbations, as it is closely related to the fact that these perturbations become sizeable at late times only. We have also considered tachyonic perturbations with Lorentz-invariant dispersion relation (\[Eq/Pg3/1:PhanTach\]). In that case, an interesting possibility is that the largest contribution to the CMB ansotropy is received by the dipole component, and the angular spectrum rapidly decays with the increase of $l$. We have seen that even if the entire observed dipole anisotropy is attributed to the tachyonic perturbations, the quadrupole component generated by them is still consistent with the observational data. It is worth noting that this result should be inherent not only in the Lorentz-invariant model, but also in Lorentz-violating models with tachyonic dispersion relations, provided that the “frequency” does not vanish at zero momentum and decreases as momentum increases. Our main conclusion is that even if perturbations of the tachyonic type exist in the Universe, their contribution to the CMB anisotropy is small. Nevertheless, we do not exclude a possibility that growing precision of observations, and especially elaborate analysis of correlations between CMB anisotropy and structures in the Universe, may lead to hints toward the possible exotic property of dark energy, the tachyonic behaviour of its perturbations. Acknowledgemets {#acknowledgemets .unnumbered} =============== This work has been partially supported by Russian Foundation for Basic Research grant Nos. 07-02-01034a (O.S. and M.S) and 08-02-00473 (M.L. and V.R), grant of the President of RF for leading scientific schools under Contract No. NS-1616.2008.2 (M.L and V.R.), grant of the President of RF under Contract No. MK-2503.2008.2 (O.S.) and grant of Dynasty Foundation (M.L.) M.L. is indebted to Université Libre de Bruxelles, where part of this work has been done under partial support by the Belspo:IAP-VI/11 and IISN grants, for hospitality. [99]{} T. Padmanabhan, Phys. Rept.  [**380**]{} (2003) 235 \[arXiv:hep-th/0212290\]. V. Sahni, Lect. Notes Phys.  [**653**]{} (2004) 141 \[arXiv:astro-ph/0403324\]. E. J. Copeland, M. Sami and S. Tsujikawa, Int. J. Mod. Phys.  D [**15**]{} (2006) 1753 \[arXiv:hep-th/0603057\]. V. Sahni and A. Starobinsky, Int. J. Mod. Phys.  D [**15**]{} (2006) 2105 \[arXiv:astro-ph/0610026\]. J. Frieman, M. Turner and D. Huterer, *Dark Energy and the Accelerating Universe*, arXiv:0803.0982 \[astro-ph\]. E. Komatsu [*et al.*]{} \[WMAP Collaboration\], *Five-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Cosmological Interpretation*, arXiv:0803.0547 \[astro-ph\]. V. Sahni, A. Shafieloo and A. A. Starobinsky, *Two new diagnostics of dark energy*, arXiv:0807.3548 \[astro-ph\]. J. Q. Xia, H. Li, G. B. Zhao and X. Zhang, *Determining Cosmological Parameters with Latest Observational Data*, arXiv:0807.3878 \[astro-ph\]. R. R. Caldwell, Phys. Lett. B [**545**]{}, 23 (2002) L. Senatore, Phys. Rev.  D [**71**]{}, 043512 (2005) \[arXiv;astro-ph/0406187\]. P. Creminelli, M. A. Luty, A. Nicolis and L. Senatore, JHEP [**0612**]{}, 080 (2006) \[arXiv:hep-th/0606090\]. V. A. Rubakov, Theor. Math. Phys. [**149**]{}, 1651 (2006) \[Teor. Mat. Fiz. **149**, 409 (2006)\] \[arXiv:hep-th/0604153\]. M. Libanov, V. Rubakov, E. Papantonopoulos, M. Sami and S. Tsujikawa, JCAP **0708**, 010 (2007) \[arXiv:0704.1848 \[hep-th\]\]. A. Sergienko, V. Rubakov, *Phantom dark energy with tachyonic instability: metric perturbations*, 2008 \[arXiv:0803.3163 \[hep-th\]\]. A.D. Linde, *Elementary Particle Physics and Inflationary Cosmology*, Moscow, “Nauka”, 1981. A.D. Dolgov, Ya.B. Zeldovich, V.V. Sazhin, *Cosmology of the Early Universe*, Moscow State University Press, 1988. V. Mukhanov *Physical Foundations of Cosmology*, (Cambridge University Press, Cambridge, England, 2005). M. R. Nolta [*et al.*]{} \[WMAP Collaboration\], *Five-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Angular Power Spectra,* arXiv:0803.0593 \[astro-ph\]. M. Giovannini, Int. J. Mod. Phys.  D [**14**]{} (2005) 363 \[arXiv:astro-ph/0412601\]. I.S. Gradshteyn, I. M. Ryzhik, *Table of Integrals, Series, and Products*, Academic Press, New York 2000;\ M. Abramowitz and I.A. Stegun, *Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables*, Dover, New York, 1964. http://lambda.gsfc.nasa.gov/product/map/dr3/pow\_tt\_spec\_get.cfm http://lambda.gsfc.nasa.gov/toolbox/tb\_cmbfast\_form.cfm J. Dunkley [*et al.*]{} \[WMAP Collaboration\], *Five-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Likelihoods and Parameters from the WMAP data*, arXiv:0803.0586 \[astro-ph\]. G. Hinshaw [*et al.*]{} \[WMAP Collaboration\], Astrophys. J. Suppl.  [**170**]{} (2007) 288 \[arXiv:astro-ph/0603451\]. L.P. Grishchuk and Ya.B. Zeldovich, Sov. Astron. [**22**]{} (1978) 125. B. Paczynski and T. Piran, Astroph. J. [**364**]{} (1990) 341. M. S. Turner, Phys. Rev.  D [**44**]{} (1991) 3737. M. Jaroszynski and B. Paczynski, Astroph. J. [**448**]{} (1995) 488. D. Langlois and T. Piran, Phys. Rev.  D [**53**]{} (1996) 2908 \[arXiv:astro-ph/9507094\]. D. Langlois, Phys. Rev.  D [**54**]{} (1996) 2447 \[arXiv:gr-qc/9606066\]. D. Langlois, Phys. Rev.  D [**55**]{} (1997) 7389. M. Kamionkowski and L. Knox, Phys. Rev.  D [**67**]{} (2003) 063001 \[arXiv:astro-ph/0210165\]. C. Gordon, K. Land and A. Slosar, *Determining the motion of the solar system relative to the cosmic microwave background using type Ia supernovae*, arXiv:0711.4196 \[astro-ph\]. [^1]: Refs. [@Libanov:2007; @rub08] consider the Lorentz-violating case, but that analysis is straightforwardly repeated in the Lorentz-invariant situation. [^2]: The effect of the tachyonic instability on the CMB anisotropy is sizeable if $\sqrt{\alpha }M/H_0 \sim 100$ (see section \[sub:li-exp\]), hence our choice here. [^3]: Like in the Lorentz-violating case, the dependence on $w_p$ is weak. [^4]: The second representation of the integral (\[Eqn/Pg15/1:PhanTach\]) is convenient because the hypergeometric function is only slightly different from unity at $0 \leq z \leq 1$, which corresponds to $0\leq \alpha <\infty $. [^5]: The reason for the $10\%$ discrepancy is the omission of the term of order $\nu^{2}$ in the expression (\[Eqn/Pg11/1:PhanTach\]) for $\mathcal{N}(a,\nu)$; note that $\mathcal{N}(a,\nu)$ enters the final result exponentially.

ArXiv

--- abstract: | In this paper, we look at the effect of volatility clustering on the risk indifference price of options described by Sircar and Sturm in the paper ‘From smile asymptotics to market risk measures’ [@SS]. The indifference price in [@SS] is obtained by using dynamic convex risk measures given by backward stochastic differential equations (BSDEs). Volatility clustering is modeled by a fast mean-reverting volatility in a stochastic volatility model for stock price. Asymptotics of the indifference price of options and their corresponding implied volatility are obtained in this paper, as the mean-reversion time approaches zero. Correction terms to the asymptotic option price and implied volatility are also obtained.\ [: risk measures, indifference price, implied volatility, volatility clustering, comparison principle.]{} author: - 'Rohini Kumar [^1] [^2]' title: Effect of volatility clustering on indifference pricing of options by convex risk measures --- Introduction ============ In an incomplete market, there are several ways of pricing options. The arbitrage-free method of option pricing, where the option price is given by the expected value of the discounted payoff under a risk-neutral equivalent martingale measure, is widely known and studied (see for example [@FPS00]). Another method is the indifference pricing method. Indifference pricing of options using utility functions have been studied extensively in [@HN89], [@MZ04], [@RKar00], etc. Later it was observed that the same idea of indifference pricing could be extended from utility functions to dynamic convex risk measures, see [@BK09]. Sircar and Sturm in [@SS] derived a nonlinear partial differential equation (PDE) characterizing the indifference price of put options given by dynamic convex risk measures. In their paper, they used the residual risk measure after hedging for pricing options; the residual risk measure was given in terms of a BSDE. They also obtained the implied volatility, corresponding to this indifference option price, as the viscosity solution of a nonlinear PDE. The significance of the results in [@SS] is that, via the indifference pricing scheme, the market risk, reflected in the implied volatility skew, can be related to convex risk measure theory. Typically, convex risk measures are defined abstractly via BSDEs (see [@BK09]). So the right driver in the BSDE which gives a good risk measure is hard to determine. Using the indifference pricing of Sircar and Sturm, we can calibrate the driver from the market implied volatility data. In their paper, [@SS], small-maturity asymptotics of the implied volatility yielded simple formulas. However, in general, a closed form solution to the non linear PDE in[@SS] is hard, if not impossible, to find. In this paper, an effort is made to give some meaningful, simplified formulas for the implied volatility surface, by considering the effect of volatility clustering. It is believed that market volatility fluctuates frequently between high and low periods. While volatility cannot be observed directly, this “clustering" behavior is estimated from observed stock prices. To model stock prices subject to this clustering behavior of volatility we use stochastic volatility models where the volatility is a fast mean-reverting ergodic process. Such stochastic volatility models with fast mean-reverting volatility were found to be a good fit for stock price data (see chapter 4 of [@FPS00]). The question of interest in this paper is, how does fast mean-reversion in volatility affect option prices? As the rate of mean-reversion increases, the long-run behavior of the ergodic volatility process manifests. Consequently, the effect of volatility gets averaged with respect to the invariant distribution of the ergodic volatility process. The analysis of fast mean-reversion of volatility on option prices has been studied in the case of no-arbitrage pricing, see chapter 5 of [@FPS00]. In this paper, we will look at the effect of fast mean-reversion of volatility on the indifference prices of options given in [@SS]. In the no-arbitrage pricing case, option prices are given as the solution of linear PDEs. For small mean-reversion time, denoted by the parameter $\ep$, this leads to a singular perturbation problem. In [@FPS00], assuming the volatility process is an Ornstein-Uhlenbeck process, the option price is expanded in powers of $\ep$ and the asymptotic option price, when $\ep\to0$, is obtained together with a correction term of order $\sqrt{\ep}$. The corresponding corrected implied volatility is also obtained. In this paper, unlike the no-arbitrage case, the option price is not the solution of a linear PDE. In fact, this indifference price is given in terms of the solutions to BSDEs. Initially, finding the asymptotic indifference price appears to involve averaging of BSDEs. However, there is a difficulty: Sircar and Sturm use quadratic drivers in their BSDEs which are not Lipschitz and hence we cannot use the established stability results for BSDEs (for example in [@HPeng97]). We instead avail of the nonlinear Feynman-Kac formula derived by Pardoux and Peng in [@PP92], and express the option price in terms of solutions to nonlinear PDEs; thus converting this into a problem of averaging nonlinear PDEs. The formal derivation of the corrected option price and implied volatility, in sections \[sec:heuristics\] and \[sec:IV\], follows the same line of reasoning seen in [@FPS00]. The difference in method lies in the rigorous proof of accuracy of the corrected asymptotic formula, which is in section \[sec:rigorous\]. For the rigorous proof, we use the maximum principle to bound the solution of the nonlinear PDE, which gives the option price, within $o(\sqrt{\ep})$ distance of the corrected asymptotic option price formula. The purpose of this paper is two-fold: on one hand, we look at the effect of an observed phenomenon, viz., volatility clustering, on option pricing; secondly, the averaging of the volatility leads to a simpler formula for implied volatility, which makes calibration easier. The dependence of the corrected implied volatility formula on the risk parameters can potentially be used to calibrate risk measures from the market implied volatility data. It should be interesting to see if this work extends to other risk measures besides those given by BSDEs. This however falls outside the scope of this paper and will be considered for future work. In general, regardless of risk measure used, if the security price can be expressed as a solution of a PDE, we can potentially use the method in this paper for averaging out the effect of fast mean-reverting volatility. The paper is organized as follows. In section \[sec:model\] we recall the indifference pricing of options from [@SS] and introduce the nonlinear PDE that gives the indifference price of put options. Section \[sec:heuristics\] has heuristic calculations for the corrected asymptotic option price. The main result and the rigorous proof of accuracy of the obtained corrected asymptotic option price formula is in section \[sec:rigorous\]. The corrected implied volatility formula is obtained in section \[sec:IV\]. Preliminaries {#sec:model} ============= We begin by introducing the stochastic volatility model for stock price. Stochastic volatility model. ---------------------------- Let $S_{\ep,t}$ denote stock price at time $t$, where the parameter $\ep$ refers to the mean-reversion time scale for volatility. As $\ep$ approaches $0$, the speed of mean-reversion, $1/\ep$, increases. Let $(\Omega, \mathcal F, P)$ denote the probability space on which $S_{\ep,t}$ satisfies the following stochastic volatility model. \[FSDEs\] $$\begin{aligned} dS_{\ep,t}=&b(\yept)S_{\ep,t}dt+\s_1(\yept)S_{\ep,t}dW^{(1)}_t, \quad 0\leq t\leq T,\\ d\yept=&\frac{m-\yept}{\ep}dt+\frac{\s_2(\yept)}{\sqrt{\ep}}(\rho dW_t^{(1)}+\sqrt{1-\rho^2}dW^{(2)}_t), \quad 0\leq t\leq T,\\ (S_{\ep,0},Y_{\ep,0})&=(s,y),\end{aligned}$$ where $|\rho|< 1$, $W^{(1)}$ and $W^{(2)}$ are independent Brownian motions on $(\Omega,\mathcal F, P)$. We make the same assumptions on the stochastic volatility model as in [@SS]. For the reader’s convenience we recall these assumptions. \[assmpn\] We assume that 1. $\s_1, \s_2\in C^{1+\beta}_{loc}(\R)$, where $C^{1+\beta}_{loc}(\R)$ is the space of differentiable functions with locally Hölder-continuous derivatives with Hölder-exponent $\beta>0$.\ 2. Both $\s_1$ and $\s_2$ are bounded and bounded away from zero: $$0<\underline{c_1}<\s_1<\overline{c_1}<\infty,\qquad \text{ and }0<\underline{c_2}<\s_2<\overline{c_2}<\infty,$$ 3. $b\in C^{0+\beta}_{loc} $, and $b$ is bounded. Let $\B$ denote the infinitesimal generator of the $Y$ process when $\ep=1$. Then, for $f\in C^2(\R)$, $$\label{generator-B} \B f(y):= (m-y)\p_yf(y)+\frac1 2 \s_2^2(y)\p^2_{yy}f(y).$$ By the general theory of 1-D diffusions (see Karlin and Taylor [@KT81], page 221) it is easy to see that there exists a unique probability measure $$\label{inv-meas} \pi(dy)=Z^{-1}\frac{\exp\l\{\int_0^y\frac{2(m-z)}{\s_2^2(z)}dz\r\}}{\s_2^2(y)}dy,$$ such that $\int \B f(y) \pi(dy)=0$ for all $f\in C_c^2(\R)$; $Z$ is the normalizing constant, so that $\int \pi(dy)=1$. The invariant distribution of $Y$ given by plays an important role in the following analysis. Indifference option price. -------------------------- We consider a European put option with maturity time $T$ and strike price $K$. The indifference price of this European put option at time $t$, $P_{\ep}(t,x,y)$, is given in [@SS] in terms of risk measures as follows. $$P_\ep(t,x,y)=\Rt-\rt,$$ where $(\tilde{R}_\cdot, \tilde{Z}_\cdot)$ and $(R_\cdot, Z_\cdot)$ are respectively solutions of the following BSDEs $$\label{BSDE1} \Rt=-\int_t^Tf(\Zs, \ZZs)ds-\int_t^T\Zs dW_s^{(1)}-\int_t^T\ZZs dW_s^{(2)};$$ and $$\label{BSDE2} \rt=-(K-S_{\ep,T})^+-\int_t^Tf(\zs, \zzs)ds-\int_t^T\zs dW_s^{(1)}-\int_t^T\zzs dW_s^{(2)}.$$ The function $f$ in the above BSDEs satisfies the criteria for admissible drivers (see Definition 2.3 in [@SS]) to ensure the solvability of the BSDEs. In this paper, we will consider a specific family of admissible drivers called [*distorted entropic risk measures*]{} which were introduced in [@SS] (see section 3.1 of [@SS]). This class of drivers has the following form: $$g^{\eta,\gamma}(z_1,z_2):=\frac{\gamma}{2}\l((z_1+\eta z_2)^2+z_2^2\r),$$ and is parametrized by two parameters: the risk aversion parameter $\gamma>0$ and the volatility risk premium $\eta$. When $\eta=0$ the driver reduces to the classical entropic risk measure whose level curves are circles with radius depending on the risk aversion parameter, $\gamma$. By introducing $\eta$, we distort this circle into an ellipse (if $|\eta|<1$). Under hedging, the risk measure gets adjusted and is now given by a BSDE where the driver $g^{\eta,\gamma}$ is transformed to $$\begin{aligned} \tilde{g}^{\eta,\gamma}(z_1,z_2)&=\inf_{\nu\in \R}\l(g^{\eta,\gamma}(z_1+\s_1(y)\nu,z_2)+b(y)\nu\r)\\ &=z_1\frac{b(y)}{\s_1(y)}+\frac{b^2(y)}{2\gamma \s^2_1(y)}-\frac{\gamma}{2}z_2^2+\frac{\eta b(y)z_2}{\s_1(y)}.\end{aligned}$$ Henceforth we will take this to be our driver, i.e. define $$f(z_1,z_2):=z_1\frac{b(y)}{\s_1(y)}+\frac{b^2(y)}{2\gamma \s^2_1(y)}-\frac{\gamma}{2}z_2^2+\frac{\eta b(y)z_2}{\s_1(y)}.$$ It is important to note that this driver is not Lipschitz, as it is quadratic in $z_2$ Instead of stock price $S_{\ep,t}$, we will work with the logarithm of stock price normalized by the strike price: $X_{\ep,t}:=\ln \l(\frac{S_{\ep,t}}{K}\r)$. With this change of variable, becomes \[FSDEs-alt\] $$\begin{aligned} d\xept=&\l(b(\yept)-\frac1 2\s_1^2(\yept)\r)dt+\s_1(\yept)dW^{(1)}_t, \quad 0\leq t\leq T,\\ d\yept=&\frac{m-\yept}{\ep}dt+\frac{\s_2(\yept)}{\sqrt{\ep}}(\rho dW_t^{(1)}+\sqrt{1-\rho^2}dW^{(2)}_t), \quad 0\leq t\leq T,\\ (X_{\ep,0},Y_{\ep,0})&=(x,y).\end{aligned}$$ Sircar and Sturm in [@SS] use the generalized Feynman-Kac formula, given by Pardoux and Peng in [@PP92], to describe the solutions of the forward-backward SDEs, $\Rt$ and $\rt$, in terms of solutions of the nonlinear PDE below. Before introducing this PDE, we will first make another change of variable $\tau:=T-t$, which gives the time to maturity. Define the differential operator $\L_\ep$ by $$\label{L_ep} \begin{split} \L_\ep g:=&\frac{1}{2}\s_1^2(y)\p^2_{xx}g+\frac{1}{2\ep}\s_2^2(y)\p^2_{yy}g+\frac{\rho}{\sqrt{\ep}}\s_1(y)\s_2(y)\p^2_{xy}g\\ &+\l(\frac{1}{\ep}(m-y)-\frac{\rho}{\sqrt{\ep}}\frac{b(y)\s_2(y)}{\s_1(y)}\r)\p_y g-\frac1 2 \s_1^2(y)\p_xg\\ &-\frac{b^2(y)}{2\gamma\s^2_1(y)}+\frac{\gamma(1-\rho^2)}{2\ep}\s_2^2(y)(\p_yg)^2-\frac{\eta\sqrt{1-\rho^2 }}{\sqrt{\ep}}\frac{b(y)\s_2(y)\p_y g}{\s_1(y)}, \end{split}$$ for $g\in C^{2}(\R^2)$. Let $\tilde{u}_\ep$ and $u_\ep$ denote solutions to the PDE $$\label{PDE} \p_\tau u=\L_\ep u,$$ with the initial conditions $\tilde{u}_\ep(0,x,y)=0$ and $u_\ep(0,x,y)=-[K-Ke^x]^+$, respectively. By Theorem 2.9 in [@SS], we get the put option price $$\label{Pep}P_\ep(\tau,x,y)=\tilde{u}_\ep(\tau,y)-u_\ep(\tau,x,y).$$ As mentioned in [@SS], by Ladyshenskaya et al. [@Lad67 Theorem V.8.1], $u_\ep$ and $\tilde{u}_\ep$ are unique bounded classical solutions to the semilinear parabolic equation , with bounded derivatives in $[0,T]\times \mathbb R\times \R$. Since the coefficients of the PDE in and the initial condition of $\tilde{u}$ are x-independent, we get $\tilde{u}$ to be x-independent. While we only consider European put options, the results in the paper extend to any other option with bounded and continuous payoff. Asymptotic option price {#sec:asymptotics_option_price} ======================= We will begin with heuristic arguments for obtaining the asymptotic option price and correction terms to the asymptotic price. The heuristic calculations follow along the same lines as the no-arbitrage option pricing case, seen in [@FPS00] Heuristics {#sec:heuristics} ---------- Assume the following expansions of $u_\ep, \tilde{u}_\ep$ and $P_\ep$ in powers of $\sqrt{\ep}$: $$\begin{aligned} u_\ep &=u_0+\sqrt{\ep}u_1+\ep u_2+\ep^{3/2}u_3+\hdots, \label{u-exp}\\\tilde{u}_\ep&=\tilde{u}_0+\sqrt{\ep}\tilde{u}_1+\ep \tilde{u}_2+\ep^{3/2}\tilde{u}_3+\hdots, \label{tildeu-exp}\\P_\ep&=P_0+\sqrt{\ep}P_1+\ep P_2+\hdots.\label{P-exp}\end{aligned}$$ We define the following differential operators: for $g\in C^{2}( \mathbb R\times \mathbb R)$, $$\begin{aligned} \L_0g(x,y):&=&\frac1 2 \s_1^2(y)\p^2_{xx}g(x,y)-\frac1 2 \s_1^2(y)\p_xg(x,y)-\frac{b^2(y)}{2\gamma\s_1^2(y)},\\\L_1g(x,y):&=&\rho\s_1(y)\s_2(y)\p^2_{xy}g(x,y)\\&&-\l(\rho+\eta\sqrt{1-\rho^2 }\r)\frac{b(y)\s_2(y)}{\s_1(y)}\p_y g(x,y),\\\Lnl g(x,y):&=&\B g(x,y)+\frac{\gamma(1-\rho^2)}{2}\s_2^2(y)\l(\p_yg(x,y)\r)^2.$$ Note that the operator $\Lnl$ is nonlinear. Equation can be rewritten as $$\label{PDE-op} \p_\tau u_\ep=\L_\ep u_\ep= \L_0u_\ep+\frac{1}{\sqrt{\ep}}\L_1u_\ep+\frac{1}{\ep}\Lnl u_\ep.$$ ### Leading order term Using the expansion in and collecting terms of order $1/\ep$ we get $$\B u_0= -\frac{\gamma(1-\rho^2)}{2}\s_2^2(y)(\p_yu_0)^2,$$ which is satisfied if $u_0$ is $y$ independent. We will thus assume $u_0(\tau,x)$ is independent of $y$. Using this and collecting terms of order $1/\sqrt{\ep}$ we get $$\B u_1=0,$$ which is satisfied by taking $u_1$ independent of $y$. As $u_0(\tau,x)$ and $u_1(\tau,x)$ are both $y$ independent, terms of $O(1)$ in equation satisfy $$\p_\tau u_0(\tau,x)=\L_0 u_0(\tau,x)+\B u_2.$$ Thus, $u_2$ must satisfy the Poisson equation $$\label{Poisson-u2}\B u_2=\p_\tau u_0(\tau,x)-\L_0 u_0,\,$$ which has a solution provided the following centering condition holds: $$\label{centering} \p_\tau u_0=\frac1 2 \overline{\s_1}^2\p^2_{xx}u_0- \frac1 2\overline{\s_1}^2(y)\p_xu_0 - \frac{1}{2\gamma}\overline{\frac{b^2(y)}{\s_1^2(y)}}.$$ Here $\overline{\s_1}^2$ and $\overline{\frac{b^2(y)}{\s_1^2(y)}}$ denote the average of the terms $\s_1^2(y)$ and $\frac{b^2(y)}{\s_1^2(y)}$, respectively, with respect to the invariant distribution, $\pi$, of the $Y$ process. Similarly, $\tilde{u}_0$ satisfies equation . The initial conditions for $u_\ep$ and $\tilde{u}_\ep$ give the initial conditions for $u_0$ and $\tilde{u}_0$ respectively, i.e. $\tilde{u}_0(0)=0$ ($\tilde{u}_0$ is independent of both $x$ and $y$, so only a function of $\tau$) and $u_0(0,x)=-[K-Ke^x]^+$. Observe that the first order approximation term to the option price, $P_0(\tau,x)=\tilde{u}_0(\tau)-u_0(\tau,x)$, satisfies the equation $$\p_tP_0=\frac1 2 \overline{\s_1}^2\p^2_{xx}P_0- \frac1 2\overline{\s_1}^2(y)\p_xP_0,\qquad P_0(0,x)=[K-Ke^x]^+,$$ which is simply the equation for the Black-Scholes put option price, $P_{BS}(\tau,x;\overline{\s_1})$, with volatility parameter $\overline{\s_1}.$ Therefore, $$\label{P0}P_0(\tau,x)=P_{BS}(\tau,x;\overline{\s_1})=KN(-d_2)-Ke^xN(-d_1),$$where $N(z)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^z e^{-y^2/2}dy$, $d_1=\frac{x+\frac1 2 \overline{\s_1}^2\tau}{\overline{\s_1}\sqrt{\tau}}$ and $d_2=\frac{x-\frac1 2 \overline{\s_1}^2\tau}{\overline{\s_1}\sqrt{\tau}}$. Observe that $$\label{u0}u_0(\tau,x)=P_0(\tau,x)- \frac{1}{2\gamma}\overline{\frac{b^2(y)}{\s_1^2(y)}}\tau.$$ ### Higher order terms Define $\hat{\L}_0:=\L_0+\frac{b^2}{2\gamma\s^2_1}=\frac1 2 \s_1^2\p^2_{xx}-\frac1 2 \s_1^2\p_x$. So $\hat{\L}_0$ is a linear operator. Equating terms of order $\sqrt{\ep}$ in gives the equation $$\p_\tau u_1=\hat{\L}_0u_1+\L_1u_2+\B u_3.$$ Thus $u_3$ is the solution of the Poisson equation $$\label{u3-equation} \B u_3= \p_\tau u_1-\hat{\L}_0u_1-\L_1u_2.$$ provided the following centering condition is satisfied: $$\p_\tau u_1=\overline{\hat{\L}_0u_1+\L_1u_2}.$$ Henceforth, a line over any term indicates averaging of the term with respect to the invariant distribution of the $Y$ process given in . Since $u_1$ is independent of $y$, the above equation becomes $$\label{u1-1} \p_\tau u_1=\frac1 2 \overline{\s_1}^2\p^2_{xx}u_1-\frac1 2 \overline{\s_1}^2\p_xu_1+\overline{\L_1u_2}.$$ To simplify the right hand side of the above, recall that $u_2$ is the solution of the Poisson equation in . Together with , we see that $u_2$ is the solution of the following equation, $$\B u_2(\tau,x,y)=\frac{b^2(y)}{2\gamma\s_1^2(y)}-\overline{\frac{b^2(y)}{2\gamma\s_1^2(y)}}+\frac1 2\l(\s_1^2(y)-\overline{\s_1}^2\r)\l(\p_{x}u_0(\tau,x)- \p^2_{xx}u_0(\tau,x)\r).$$ Let $\phi_1(y)$ and $\phi_2(y)$ denote the solutions of $$\B \phi_1(y)=\frac{b^2(y)}{2\gamma\s_1^2(y)}-\overline{\frac{b^2(y)}{2\gamma\s_1^2(y)}}$$ and $$\B \phi_2(y)=\frac1 2 \l(\s_1^2(y)-\overline{\s_1}^2\r)$$ respectively. Define $$\label{U2} U_2(\tau,x,y):= \phi_1(y)+\phi_2(y)\l(\p_x u_0(\tau,x)-\p^2_{xx}u_0(\tau,x)\r).$$ Then $$\label{u2} u_2(\tau,x,y)=U_2(\tau,x,y) +F(\tau, x)$$ where $F(\tau,x)$ will be determined later, see . \[Rem-U2\] Note that the function $U_2(\tau,x,y)$ satisfies equations and . We compute $$\begin{aligned} &\overline{\L_1u_2}\\ &=\overline{\rho\s_1(y)\s_2(y)\phi^\prime_2(y)}\l(\p^2_{xx}u_0(\tau,x)-\p^3_{xxx}u_0(\tau,x)\r)\\ &-\l(\rho+\eta\sqrt{1-\rho^2 }\r)\overline{\frac{b(y)\s_2(y)}{\s_1(y)}\phi^\prime_2(y)}\l(\p_x u_0(\tau,x)-\p^2_{xx}u_0(\tau,x)\r)\\ &\quad-\l(\rho+\eta\sqrt{1-\rho^2 }\r)\overline{\frac{b(y)\s_2(y)}{\s_1(y)}\phi^\prime_1(y)}.\end{aligned}$$ Substituting this in we get $$\begin{aligned} \p_\tau u_1=&\frac1 2 \overline{\s_1}^2\p^2_{xx}u_1-\frac1 2\overline{\s}_1^2(y)\p_xu_1-A\p^3_{xxx}u_0(\tau,x)\\ &+(A+B)\p^2_{xx}u_0(\tau,x)-B\p_xu_0(\tau,x)-\tilde{A}, $$ where $$\label{A}A=\overline{\rho\s_1(y)\s_2(y)\phi^\prime_2(y)},$$ $$\label{tildeA}\tilde{A}=\l(\rho+\eta\sqrt{1-\rho^2 }\r)\overline{\frac{b(y)\s_2(y)}{\s_1(y)}\phi^\prime_1(y)}$$ and $$\label{B}B=\l(\rho+\eta\sqrt{1-\rho^2 }\r)\overline{\frac{b(y)\s_2(y)}{\s_1(y)}\phi^\prime_2(y)}.$$ It is easy to verify that the solution to the above equation is $$\label{u1-formula} u_1(\tau,x)=\tau\l[ -A\p^3_{xxx}u_0(\tau,x)+(A+B)\p^2_{xx}u_0(\tau,x)-B\p_xu_0(\tau,x)-\tilde{A} \r].$$ By a similar argument, we get $$\tilde{u}_1(\tau)=\tau\l[ -A\p^3_{xxx}\tilde{u}_0(\tau)+(A+B)\p^2_{xx}\tilde{u}_0(\tau)-B\p_x\tilde{u}_0(\tau)-\tilde{A} \r] = -\tilde{A}\tau .$$ Thus, $$\label{P1}P_1(\tau,x)=\tau\l[ -A\p^3_{xxx}P_0(\tau,x)+(A+B)\p^2_{xx}P_0(\tau,x)-B\p_xP_0(\tau,x) \r].$$ ### Terms of order $\ep$ {#sec:u2} Using the asymptotic expansion of $u_\ep$ in and collecting terms of $O(\ep)$, we get $$\p_\tau u_2=\hat{\L}_0u_2+\L_1 u_3+\B u_4+\frac{\gamma(1-\rho^2)}{2}\s_2^2(\p_y u_2)^2.$$ Thus $u_4$ is the solution of the Poisson equation $$\label{u4} \B u_4=\p_\tau u_2-\hat{\L}_0u_2-\L_1 u_3-\frac{\gamma(1-\rho^2)}{2}\s_2^2(\p_y u_2)^2,$$ provided the following centering condition holds: $$\label{F} \overline{\p_\tau u_2-\hat{\L}_0u_2-\L_1 u_3-\frac{\gamma(1-\rho^2)}{2}\s_2^2(\p_y u_2)^2}=0.$$ The unknown function $F(\tau,x)$ in is determined by the above equation . Accuracy of corrected asymptotic price formula {#sec:rigorous} ---------------------------------------------- \[asymp-price\] Let $P_\ep$ denote the put option price given by , and let $P_0$ and $P_1$ be as defined in and , respectively. Then, under assumptions \[assmpn\], $$|P_\ep(\tau,x,y)-(P_0(\tau,x)+\sqrt{\ep}P_1(\tau,x))|\leq O(-\ep\log\ep).$$ We use a maximum/comparison principle argument to prove this result. Define $$\label{psi}\psi_\ep(y)=\ep (y-m)^2+D_\ep,$$ where $D_\ep>0$ is a constant chosen such that $$\label{psi-ineq} \frac{\psi_\ep(y)}{2}>|u_2(\tau,x,y)|+\sqrt{\ep}\ |u_3(\tau,x,y)|$$ for all $(\tau,x,y)\in [0,T]\times\R\times\R.$ It is possible to find such a $D_\ep$, as $U_2$ and $u_3$ have at most logarithmic growth in $y$ (see , , , and Lemma \[bounded-phi-prime\]) and are bounded functions of $\tau$ and $x$ (see and ). It is easy to check that $D_\ep=O(-\log \ep)$, see Lemma \[D\_ep\] in the Appendix. Choose a constant $C>0$ large enough so that $$\label{const-C} C+2(y-m)^2>\p_\tau U_2-\hat{\L}_0U_2-\L_1 u_3.$$ It is possible to choose such a $C$ since $U_2$ and $u_3$ are bounded functions of $\tau$ and $x$ and have at most logarithmic growth in $y$ (see , , , and Lemma \[bounded-phi-prime\]). Define $$\ul{u}_\ep(\tau,x,y):=u_0(\tau,x)+\sqrt{\ep} u_1(\tau, x)+\ep U_2(\tau,x,y)+\ep^{3/2}u_3(\tau,x,y)-\ep \psi_\ep(y)-\ep (C+\bar{c}_2^2)\tau,$$ and $$\ol{u}_\ep(\tau,x,y):= u_0(\tau,x)+\sqrt{\ep} u_1(\tau, x)+\ep U_2(\tau,x,y)+\ep^{3/2}u_3(\tau,x,y)+\ep \psi_\ep(y)+\ep (C+\bar{c}_2^2)\tau,$$ where $\bar{c}_2$ is the upper bound on $\s_2$ in Assumption \[assmpn\], $u_0, u_1, U_2$ and $\psi_\ep$ are defined in , , and respectively. We compute $$\begin{aligned} \p_\tau \ul{u}_\ep-\L_\ep\ul{u}_\ep=&\p_\tau u_0+\sqrt{\ep}\p_\tau u_1+\ep\p_\tau U_2+\ep^{3/2}\p_\tau u_3-\ep(C+\bar{c}_2^2)\\ &-\L_0u_0-\sqrt{\ep}\hat{\L}_0 u_1-\ep\hat{\L}_0U_2-\ep^{3/2}\hat{\L}_0u_3\\ &-\sqrt{\ep}\L_1 U_2+\sqrt{\ep}\L_1\psi_\ep-\ep\L_1 u_3\\ &-\B U_2-\sqrt{\ep}\B u_3+\B\psi_\ep-\ep\frac{\gamma(1-\rho^2)\s^2_1(y)}{2}\l(\p_y U_2+\sqrt{\ep}\p_y u_3-\p_y\psi_\ep\r)^2,\end{aligned}$$ where we have used the independence of $y$ in $u_0$ and $u_1$ to eliminate some terms. Using Remark \[Rem-U2\], , and $\B \psi_\ep< -\ep 2(y-m)^2+\ep\bar{c}_2^2$, we get $$\begin{aligned} \p_\tau \ul{u}_\ep-\L_\ep\ul{u}_\ep <& -\ep(C+2(y-m)^2)\\ &+\ep^{3/2}\l[-2(\rho+\eta\sqrt{1-\rho^2})\frac{b(y)\s_2(y)}{\s_1(y)}(y-m)+\p_\tau u_3-\hat{\L}_0u_3\r]\\ &+\ep\Bigl[\p_\tau U_2-\hat{\L}_0U_2-\L_1u_3-\frac{\gamma(1-\rho^2)\s^2_1(y)}{2}\l(\p_y U_2+\sqrt{\ep}\p_y u_3-\ep 2(y-m)\r)^2\Bigr].\end{aligned}$$ The term $-\ep(C+2(y-m)^2)$ dominates the other terms on the right hand side of the above equation, uniformly, for small enough $\ep$. To see this, first look at the terms of $O(\ep^{3/2})$. The first term grows linearly in $y$ as $b$ and $\s_2$ are bounded and $\s_1$ is bounded away from $0$. The terms that depend on $u_3$ grow at most logarithmically in $y$ and are bounded in $\tau$ and $x$. This can be seen from , , , and Lemma \[bounded-phi-prime\]. Thus the $O(\ep^{3/2})$ terms are dominated by $\ep(C+2(y-m)^2)$ uniformly for all $\tau,x,y$, for small enough $\ep$. We now turn to the remaining term of $O(\ep)$ in the third line of the above inequality. By choice of $C$ in , this term is dominated by $\ep(C+2(y-m)^2)$. Therefore the term $-\ep(C+2(y-m)^2)$ dominates the other terms for small enough $\ep$ making the right hand side of the above inequality negative. Hence, $$\label{max-principle-ineq1} \p_\tau \ul{u}_\ep(\tau,x,y)-\L_\ep\ul{u}_\ep(\tau,x,y)<0, \quad \forall (\tau,x,y)\in (0,T)\times \mathbb R\times \mathbb R,$$ for small enough $\ep$. Similarly, it can be shown that $$\label{max-principle-ineq2} \p_\tau \ol{u}_\ep(\tau,x,y)-\L_\ep\ol{u}_\ep(\tau,x,y)>0, \quad \forall (\tau,x,y)\in (0,T)\times \mathbb R\times \mathbb R,$$ for small enough $\ep$. By the maximum principle, we can show that $\ul{u}_\ep\leq u_\ep\leq \ol{u}_\ep$, as follows. The goal is to prove that at the point of maximum of $\ul{u}_\ep-u_\ep$ we have $\ul{u}_\ep\leq u_\ep$. For this we will first need to ensure the maximum of $\ul{u}_\ep-u_\ep$ is attained and to this end we first perturb our function $\ul{u}_\ep$ slightly. Define $$\ul{u}_\ep^\delta=\ul{u}_\ep-\delta\l(\sqrt{1+x^2}+\frac{1}{T-\tau}\r),$$ where $0<\delta\ll1$. Note that the first and second order derivatives of $\sqrt{1+x^2}$ are bounded. Therefore the strict inequalities and still hold if $\underline{u}_\ep$ is replaced with $\underline{u}_\ep^\delta$, for small enough $\delta$. Because of the term $-\psi_\ep(y)$ in $\ul{u}_\ep$ which dominates the growth in $y$, $\ul{u}_\ep\to-\infty$ when $|y|\to\infty$. Also observe that $\ul{u}_\ep$ is bounded in $\tau$ and $x$, so the perturbation of $\ul{u}_\ep$ by $-\delta\l(\sqrt{1+x^2}+\frac{1}{T-\tau}\r)$ ensures $\ul{u}^\delta_\ep\to -\infty$ when either $|x|\to\infty$ or $\tau\to T$. Since $u_\ep$ is bounded, we see that $\ul{u}^\delta_\ep-u_\ep\to -\infty$ when either $|y|\to\infty$, $|x|\to\infty$ or $\tau\to T$. Therefore, $\ul{u}^\delta_\ep-u_\ep$ must attain its maximum at some finite point, say $(\tau_0,x_0,y_0)$, in the domain $[0,T)\times\R\times\R$. For $(\tau_0,x_0,y_0)$ in the interior of $(0,T)\times \mathbb R\times \mathbb R$, we have $$\label{max-principle}0=\p_\tau u_\ep(\tau_0,x_0,y_0)-\L_\ep u_\ep(\tau_0,x_0,y_0)\leq \p_\tau\ul{u}^\delta_\ep(\tau_0,x_0,y_0)- \L_\ep\ul{u}^\delta_\ep(\tau_0,x_0,y_0)<0;$$ which is a contradiction. Therefore the point of maximum must occur at the boundary $\tau=0$. Recall that at $\tau=0$, $u_0(0,x)=u_\ep(0,x,y)=\ -[K-Ke^x]^+$ and $u_1(0,x)=0$ from . Then, by the choice of $D_\ep$ in the definition of $\psi_\ep$ i.e. , we get $\ul{u}^\delta_\ep(0,x,y) <u_\ep(0,x,y)$, for all $(x,y)\in\R\times\R$ and thus $\ul{u}^\delta_\ep<u_\ep$ everywhere, for all $\delta\ll1$. Taking the limit at $\delta\to 0$, we have the desired comparison $$\label{ineq1}\ul{u}_\ep(\tau,x,y)\leq u_\ep(\tau,x,y), \quad \forall (\tau,x,y)\in [0,T]\times \mathbb R\times \mathbb R.$$ By a similar argument, we get $$\label{ineq2} \ol{u}_\ep\geq u_\ep(\tau,x,y), \quad \forall (\tau,x,y)\in [0,T]\times \mathbb R\times \mathbb R.$$ Putting inequalities and together and using , we get $$|u_\ep(t,x,y)-(u_0(\tau,x)+\sqrt{\ep}u_1(\tau,x))|\leq \frac{\ep\psi_\ep(y)}{2}+\ep(C+\bar{c}_2^2)\tau,$$ for all $(\tau,x,y)\in [0,T]\times \mathbb R\times \mathbb R$. Recall that the constant $D_\ep$ in the definition of $\psi_\ep$ is of order $-\log(\ep)$ and that $\psi_\ep$ has a term with quadratic growth in $y$, which gives us $$|u_\ep(t,x,y)-(u_0(\tau,x)+\sqrt{\ep}u_1(\tau,x))|\leq O(-\ep \log\ep)+O(\ep^2)y^2.$$ We can repeat the same argument for $\tilde{u}_\ep$, which leads to the desired result $$|P_\ep(t,x,y)-(P_0(\tau,x)+\sqrt{\ep}P_1(\tau,x))|\leq O(-\ep \log\ep)+O(\ep^2)y^2,$$ for all $(\tau,x,y)\in [0,T]\times \mathbb R\times \mathbb R$. Implied Volatility {#sec:IV} ================== In theory, we could use the corrected option price formula for calibration. However, option prices are typically quoted in terms of their implied volatility and so it is useful to derive a corrected implied volatility formula corresponding to the corrected option price. Let $I_\ep$ denote the implied volatility corresponding to the put option price $P_\ep$. Recall that $x=\ln(S/K)$ and the Black-Scholes formula for put option prices with volatility $\s$ is given by the formula $$P_{BS}(\tau,x;\s)=KN(-d_2)-Ke^xN(-d_1),$$ where $N(z)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^z e^{-y^2/2}dy$, $d_1=\frac{x+\frac1 2 \s^2\tau}{\s\sqrt{\tau}}$ and $d_2=\frac{x-\frac1 2 \s^2\tau}{\s\sqrt{\tau}}$. By definition, the implied volatility $I_\ep$ is obtained by setting $$\label{IV-definition}P_{BS}(\tau,x;I_\ep(\tau,x,y))=P_\ep(\tau,x,y).$$ From , we know that the leading order term of the put option price $P_0(\tau,x)=P_{BS}(\tau,x;\overline{\s_1})$. Thus, the leading order term of the implied volatility is simply $\overline{\s_1}$. To obtain a correction term for the asymptotic implied volatility we will expand $I_\ep$ about $\overline{\s_1}$ in powers of $\sqrt{\ep}$ as follows $$\label{imp-vol-expansion} I_\ep(\tau,x,y)=\overline{\s_1}+\sqrt{\ep}I_1(\tau,x,y)+\ep I_2(\tau,x,y)+\cdots.$$ Using the expansion of $I_\ep$ in , we get $$\label{equating} \begin{split} &P_{BS}(\tau,x;\overline{\s_1})+\sqrt{\ep}I_1\frac{\p P_{BS}}{\p\s}(\tau,x;\overline{\s_1})+\cdots\\ &=P_0(\tau,x)+\sqrt{\ep}P_1(\tau,x,y)+\cdots. \end{split}$$ Thus the correction term to the implied volatility is $$\label{I1} I_1=P_1(\tau,x,y)\l[\frac{\p P_{BS}}{\p\s}(\tau,x;\overline{\s_1})\r]^{-1}.$$ Differentiating the Black-Scholes formula, we get $$\frac{\p P_{BS}}{\p\s}=\frac{Ke^{-d_2^2/2}\sqrt{\tau}}{\sqrt{2\pi}}.$$Substituting the formula for $P_1$ in , we can rewrite as $$\label{Iep}\begin{split}I_\ep=&\overline{\s_1}-\sqrt{\ep}\frac{\sqrt{2\pi}}{K\sqrt{\tau}e^{-d_2^2/2}}\bigl[\tau\bigl( -A\p^3_{xxx}P_0(\tau,x)+(A+B)\p^2_{xx}P_0(\tau,x)\\ &-B\p_xP_0(\tau,x)\bigr) \bigr]+o(\sqrt{\ep}),\end{split}$$ where $A$ and $B$ are defined in and respectively. The constants $A$ and $B$ can be simplified to: $$\label{A-alt} A=\int_{-\infty}^\infty\rho\frac{\s_1(y)}{\s_2(y)}\l(\int_{-\infty}^y (\s_1^2-\overline{\s_1}^2)\pi(dy)\r)dy,$$ and $$\label{B-alt} B=\int_{-\infty}^\infty(\rho+\eta\sqrt{1-\rho^2})\frac{b(y)}{\s_1(y)\s_2(y)}\l(\int_{-\infty}^y (\s_1^2-\overline{\s_1}^2)\pi(dy)\r)dy.$$ On substituting the derivatives of $P_0$ in , we get $$\begin{split} I_\ep&=\overline{\s_1}+\sqrt{\ep}\l[ A\frac{d_2}{\overline{\s_1}^2\sqrt{\tau}}+\frac{B}{\overline{\s_1}}\r]+o(\sqrt{\ep})\\ &=\overline{\s_1}+\sqrt{\ep}\l[A\frac{x}{\overline{\s_1}^3\tau}+\frac{B-A/2}{\overline{\s_1}}\r]+o(\sqrt{\ep}). \end{split}$$ We see that, as in the no-arbitrage pricing case (see [@FPS00]), the corrected implied volatility is an affine function of log-moneyness-to-maturity ratio (LMMR), where $$LMMR=\frac{\ln(strike\ price/asset\ price)}{time\ to\ maturity}=\frac{-x}{\tau}.$$ For calibration purposes, it is convenient to rewrite the formula for implied volatility, $I_\ep$, as $$\label{IV-surface} I_\ep=a(LMMR)+d+o(\sqrt{\ep}),$$ where $a=\frac{-\sqrt{\ep}A}{\overline{\s_1}^3}$, and $d=\overline{\s_1}+\frac{\sqrt{\ep}}{\overline{\s_1}}\l(B-\frac{A}{2}\r)$. On calibration of $a$ and $d$ from market implied volatility data, we can determine $A$ and $B$ from the formulas $A=\frac{-\overline{\s_1}^3 a}{\sqrt{\ep}}$ and $B=\frac{1}{\sqrt{\ep}}\l((d-\overline{\s_1})\overline{\s_1}-\frac{\overline{\s_1}^3 a}{2} \r)$. In [@FPS00] an affine function of LMMR is fitted to S&P 500 European call option implied volatility data and $a$ and $d$ are estimated to be -0.154 and 0.149 respectively. Using these estimates we graph the implied volatility surface given by , see Figure \[fig1\]. ![Implied volatility surface given by with $a=-0.154$ and $d=0.149$.[]{data-label="fig1"}](Imp_vol_surface){width="4.5in"} Taking $Y_t$ to be an OU-process and $\sigma_1(y)$ an arctangent volatility function, we compute the implied volatility from the corrected asymptotic option price formula in Theorem 1. To be precise, we take $\sigma_1(y)=0.3+\frac{0.5}{\pi}\arctan(y)$, $\sigma_2(y)=0.2$, $m=0$, $b(y)=1$ and $\rho=-0.2$ in our stochastic volatility model . We fix $\tau=0.25$ and assume $\ep=0.004$. Figure \[fig2\] gives the implied volatility as a function of log moneyness i.e. $-x=\log K/S$, for three different values of the risk parameter $\eta$. The skew of the implied volatility function is clearly seen. Observe that as the value of $\eta$ increases, the implied volatility gets shifted down. This was also observed in [@SS] (see Figure 3 in [@SS]). In Figure \[fig2\] implied volatility appears to be an almost linearly decreasing function of log moneyness. This agrees with the formally derived formula for corrected asymptotic implied volatility in . The corrected implied volatility formula only depends on the volatility risk parameter $\eta$ and the correlation term $\rho$, and has no dependence on the risk aversion parameter $\gamma$. By calibration, the correlation coefficient $\rho$ and the risk parameter $\eta$ can be obtained from and . ![Implied volatility as a function of log moneyness for the arctangent stochastic volatility model, when $\eta=-0.25,0,0.25$.[]{data-label="fig2"}](IV){width="4in"} [10]{} Buckdahn, R., Hu, Y., Peng, S. , Probabilistic Approach to homogenization of viscosity solutions of parabolic PDEs, NoDEA, Nonlinear Differ. Equ. Appl. 6, N.4, pp. 395 Ð 411, 1999 Barrieu, P. and El Karoui, N., Pricing, hedging, and designing derivatives with risk measures. In R. Carmona, editor, Indifference pricing. Theory and applications, pages 75Ð146. Princeton University Press, 2009. O. A. Ladyzhenskaya, V. A. Solonnikov, and N. N. Ural’ceva. Linear and quasi-linear equations of parabolic type, Translated from the Russian by S. Smith. Translations of Mathematical Monographs, Vol. 23. American Mathematical Society, Providence, R.I., 1967 Fouque, J.-P., Papanicolaou, G. and Sircar, R.K., Derivatives in financial markets with stochastic volatility, Cambridge University Press, Cambridge, 2000. Fouque, J.-P., Papanicolaou, G., Sircar, R. K. and Solna,K., Multi-scale stochastic volatility asymptotics, multi-scale Model. Simul. 2 (2003), no. 1, 22–42. Hodges, S.D., Neuberger, A., Optimal Replication of Contingent Claims under Transaction Costs, Review of Futures Markets 8, 222-239 (1989). Hu, Y. and Peng, S., A stability theorem of backward stochastic differential equations and its application, C. R. Acad. Sci. Paris Sér. I Math. 324 (1997), pp. 1059Ð1064 S. Karlin and H. M. Taylor, [*A second course in stochastic processes*]{}, Academic Press Inc. \[Harcourt Brace Jovanovich Publishers\], 1981, New York Musiela, M., Zariphopoulou, T.: A Valuation Algorithm for Indifference Prices in Incomplete Market, Finance and Stochastics 8, 399-414 (2004). Pardoux, E. and Peng, S., Backward stochastic differential equations and quasilinear parabolic partial differential equations, Lecture Notes in Control and Information Sciences, Springer Berlin / Heidelberg (1992), volume 176, 200-217. Rouge, R., El Karoui, N.: Pricing via Utility Maximization and Entropy, Mathematical Finance 10, 259-276 (2000) Sircar, R.K. and Sturm,S., From Smile Asymptotics to Market Risk Measures, arXiv:1107.4632v1 \[q-fin.GN\] Appendix ======== \[bounded-phi-prime\] Suppose $f\in C_b(\R)$ is a bounded continuous function which is centered with respect to the invariant distribution $\pi$, i.e. $\int fd\pi=0$. Then, the Poisson equation $$\B v=f,$$ has a solution that grows at most logarithmically and has bounded derivative $v^\prime$. We construct a solution that satisfies the required growth condition. Suppose $f$ is bounded and $v$ is a solution of $\B v=f$. Recall the definition of the differential operator $\B$ in . Multiplying the equation, $\B v=f$, by the integrating factor $\exp\{\int_m^y\frac{2(m-z)}{\s_2^2(z)}dz\}$, we get $$\l(e^{\int_m^y\frac{2(m-z)}{\s_2^2(z)}dz}v^\prime(y)\r)^\prime=\frac{2f(y)}{\s_2^2(y)}e^{\int_m^y\frac{2(m-z)}{\s_2^2(z)}dz}$$ Without loss of generality, we can assume $m=0$, so $$\l(e^{\int_0^y\frac{-2z}{\s_2^2(z)}dz}v^\prime(y)\r)^\prime=\frac{2f(y)}{\s_2^2(y)}e^{\int_0^y\frac{-2z}{\s_2^2(z)}dz}$$ Integrating, we get $$\label{soln} v^\prime(y)=e^{\int_0^y\frac{2z}{\s_2^2(z)}dz}\int_{-\infty}^y \frac{2f(u)}{\s_2^2(u)}e^{\int_0^u\frac{-2z}{\s_2^2(z)}dz}du,$$ under the assumption that $\lim_{y\to-\infty}e^{\int_0^y\frac{-2z}{\s_2^2(z)}dz}v^\prime(y)$ is $0$. Let us denote the right hand side of by the function $G(y)$ i.e. $$G(y):=e^{\int_0^y\frac{2z}{\s_2^2(z)}dz}\int_{-\infty}^y \frac{2f(u)}{\s_2^2(u)}e^{\int_0^u\frac{-2z}{\s_2^2(z)}dz}du.$$ Using l’hôpital’s rule as $|y|\to\infty$ and the boundedness of $f$ and $\s_2$, we see that $G$ is a bounded function. Then $v(y):=\int_0^yG(u)du$ gives us a solution to the Poisson equation $\B v=f$. We will show that this solution $v$ to the Poisson equation has at most logarithmic growth. Recall the invariant measure $\pi$ given in . Using the centering condition $\int fd\pi$, we get $$v^\prime(y)=e^{\int_0^y\frac{2z}{\s_2^2(z)}dz}\int_y^\infty \frac{2f(u)}{\s_2^2(u)}e^{\int_0^u\frac{-2z}{\s_2^2(z)}dz}du.$$ Let $y>1$. By the boundedness of $f$ and $\s_2$, we can bound $$\label{bd1}\begin{split} |v^\prime(y)|&\leq ce^{\int_0^y\frac{2z}{\s_2^2(z)}dz}\int_y^\infty \frac{2}{\s_2^2(u)}e^{\int_0^u\frac{-2z}{\s_2^2(z)}dz}du\intertext{(for some constant $c>0$)} &=ce^{\int_0^y\frac{2z}{\s_2^2(z)}dz}\int_y^\infty \frac{1}{-u}\frac{-2u}{\s_2^2(u)}e^{\int_0^u\frac{-2z}{\s_2^2(z)}dz}du,\intertext{which on integrating by parts gives} &=ce^{\int_0^y\frac{2z}{\s_2^2(z)}dz}\l[\frac{1}{y}e^{-\int_0^y\frac{2z}{\s_2^2(z)}dz}-\int_y^\infty \frac{1}{u^2}e^{\int_0^u\frac{-2z}{\s_2^2(z)}dz}du \r]\\ &\leq \frac{c_1}{y} \end{split}$$ for some $c_1>0$. Let $y<-1$. We repeat the same argument using the bound $$\label{bd2}\begin{split} |v^\prime(y)|&\leq e^{\int_0^y\frac{2z}{\s_2^2(z)}dz}\int_{-\infty}^y \frac{2|f(u)|}{\s_2^2(u)}e^{\int_0^u\frac{-2z}{\s_2^2(z)}dz}du\\ &=ce^{\int_0^y\frac{2z}{\s_2^2(z)}dz}\l[\frac{1}{-y}e^{-\int_0^y\frac{2z}{\s_2^2(z)}dz}-\int_{-\infty}^y \frac{1}{u^2}e^{\int_0^u\frac{-2z}{\s_2^2(z)}dz}du \r]\\ &\leq \frac{c_2}{-y} \end{split}$$ for some $c_2>0$. The boundedness of $|v^\prime(y)|$ for all $y$ together with the bounds and when $|y|>1$ gives us $$|v(y)|\leq C_1\log(1+|y|)+C_2, \quad \forall y\in\R,$$ for some $C_1,C_2>0$. \[D\_ep\] $D_\ep=O(-\log \ep)$. Recall that $D_\ep$ is chosen to be a positive number such that $$D_\ep>|u_2(\tau,x,y)|+\sqrt{\ep}\ |u_3(\tau,x,y)|-\ep(y-m)^2,$$ for all $\tau,x,y$. Since $u_2$ and $u_3$ are bounded in $\tau$ and $x$ and have at most logarithmic growth in $y$, we can write $$|u_2(\tau,x,y)|+\sqrt{\ep}\ |u_3(\tau,x,y)|\leq c_1\log(1+(y-m)^2)+c_2\quad \text{ for all }\tau,x,y,$$ for some positive constants $c_1$ and $c_2$. So it suffices to choose $D_\ep$ such that $$\label{ineq}D_\ep>c_1\log(1+(y-m)^2)+c_2-\ep(y-m)^2,$$ for all $\tau,x,y$. The quadratic term $(y-m)^2$ grows faster than $\log(1+(y-m)^2)$ for large $|y|$, therefore the maximum of the right hand side in is attained at finite values of $y$. We will determine the maximum value of the right hand side of and choose $D_\ep$ to be larger than that. Let $y_0$ denote a point of maximum of the right hand side of , then $$\frac{d}{dy}\l[ c_1\log(1+(y-m)^2)+c_2-\ep(y-m)^2 \r]|_{y=y_0}=0,$$ i.e.$$\frac{c_12(y_0-m)}{1+(y_0-m)^2}-\ep 2(y_0-m)=0.$$ Solving for $y_0$ we get, either $y_0=m$, in which case choose $D_\ep>c_2$ or $$(y_0-m)^2=\frac{c_1}{\ep}-1,$$ and the right hand side of becomes $$c_1\log(\frac{c_1}{\ep})+c_2-c_1+\ep.$$ Therefore, for small enough $\ep$, it suffices to choose $D_\ep= -2c_1\log \ep$. [^1]: Department of Mathematics, Wayne State University, Detroit, MI 48202 ([[email protected]]{}) [^2]: Work supported in part by National Science Foundation grant DMS 1209363.

ArXiv

--- abstract: 'We construct higher order rogue wave solutions and breather profiles for the quasi-one-dimensional Gross-Pitaevskii equation with a time-dependent interatomic interaction and external trap through the similarity transformation technique. We consider three different forms of traps, namely (i) time-independent expulsive trap, (ii) time-dependent monotonous trap and (iii) time-dependent periodic trap. Our results show that when we change a parameter appearing in the time-independent or time-dependent trap the second and third-order rogue waves transform into the first-order like rogue waves. We also analyze the density profiles of breather solutions. Here also we show that the shapes of the breathers change when we tune the strength of trap parameter. Our results may help to manage rogue waves experimentally in a BEC system.' author: - 'K. Manikandan' - 'P. Muruganandam' - 'M. Senthilvelan' - 'M. Lakshmanan' title: 'Manipulating matter-rogue waves and breathers in Bose-Einstein condensates' --- Introduction ============ During the past several years considerable interest has been shown in exploring localized nonlinear waves in the variable coefficient nonlinear Schrödinger (NLS) equation and its generalizations [@pono; @serki; @cent]. The motivation comes from the fact that NLS equation and its variants appear in several branches/topics of physics, including nonlinear optics [@hase; @solli] and Bose-Einstein condensates (BECs) [@peth:smi; @dalf; @carr], etc. Focusing our attention on BECs alone, it is well known that the Gross-Pitaevskii (GP) equation governs the evolution of macroscopic wave function at ultra low temperatures [@peth:smi; @pit:str]. In particular, for cigar shaped BECs, it has been shown that the GP equation can be reduced to the 1D variable coefficient NLS equation [@perez; @raj:mur; @atre:pani] $$\begin{aligned} i\psi_t+\frac{1}{2}\psi_{xx}+R(t)\vert \psi\vert ^2\psi+\frac{1}{2} \beta(t)^2 x^2 \psi=0, \label{eq:1d-gp}\end{aligned}$$ where $\psi(x,t)$ is the condensate wave function, $t$ is the dimensionless normalized time, $x$ is the dimensionless normalized coordinate in the axial direction, $R(t)$ represents the effective scattering length and $\beta(t)$ is the axial trap frequency. A simple and straightforward way of exploring the localized/periodic structures of (\[eq:1d-gp\]) is by transforming it into a constant coefficient NLS equation through a suitable transformation. From the known solutions of the latter equation the solutions of the former equation can be identified. Using this procedure a class of solutions, in particular various localized solutions, have been identified for the model (\[eq:1d-gp\]), including soliton, breather, and rogue wave (RW) solutions. In order to appreciate the relevance of the above type of localized structures for Eq. (\[eq:1d-gp\]), we may first consider their existence in the case of constant coefficient NLS equation, $$\begin{aligned} iU_{T}+\frac{1}{2}U_{XX}+ \vert U\vert ^2U=0. \label{nls}\end{aligned}$$ It is well known that the standard NLS equation (\[nls\]) admits the following basic localized profiles and their higher-order versions [@mlsr]. #### (i) Envelope soliton: {#i-envelope-soliton .unnumbered} It is a solitary wave (localized envelope along with a carrier wave) that retains its characteristics (amplitude/shape and velocity) under collision with another soliton, except for a change in phase. The intensity profile of the soliton is shown in Fig. \[figa\](a). The typical form of the envelope soliton is $$\begin{aligned} U(X,T)= P_{1R} \exp^{i \eta_1} {\normalfont\mbox{sech}\,}{\eta_2}, \label{nls1}\end{aligned}$$ where $P_1(=P_{1R}+iP_{1I})$ is a complex constant, $\eta_1=P_{1I} X+[(P_{1R}^2-P_{1I}^2)T]/2+\eta_{1I}^{(0)}$ and $\eta_2=P_{1R}(X- P_{1I} T)+\eta_{1R}^{(0)}+\log{(1/2P_{1R})}$. Here $\eta_{1R}^{(0)}$ and $\eta_{1I}^{(0)}$ are constant parameters. {width="0.8\linewidth"} #### (ii) Breather: {#ii-breather .unnumbered} It is a localized solution with temporally and/or spatially periodic structures having constant background exhibiting internal oscillations and bound states of nonlinear wave packets [@mand] which is represented in Fig. \[figa\](b). Its typical form reads [@tajri] $$\begin{aligned} U(X,T) &= \rho_0 \cos{(2\phi_R)}\exp{i(\theta+2\phi_R)} \left[1+\frac{1}{\sqrt{a}\cosh{(\eta_R+\sigma)}+\cos{\eta_I}} \right. \notag \\ & \left. \times\left\{\left(\frac{\cosh{2\phi_I}}{\cos{2\phi_R}}-1\right)\cos{\eta_I}+i\left(\tan{2\phi_R}\sinh{(\eta_R+\sigma)}-\frac{\sinh{2\phi_I}}{\cos{2\phi_R}}\sin{\eta_I}\right)\right\}\right] \label{nls2}\end{aligned}$$ where $\rho_0$ is the amplitude of the plane wave, $\theta=kX-\omega T$, $\sigma$ , $k$, $\omega$ are constant parameters, $\eta_R=P_R X-(\Omega_R T)/2+\eta_R^0$, $\eta_I=P_I X-(\Omega_I T)/2+\eta_I^0$, $\Omega_R=2kP_R-\frac{P_R^2-P_I^2\sin{2\phi_R}+2P_RP_I\sinh{2\phi_I}}{\cosh{2\phi_I}-\cos{2\phi_R}}$, $\Omega_I=2kP_I+\frac{P_R^2-P_I^2\sin{2\phi_I}+2P_RP_I\sinh{2\phi_R}}{\cosh{2\phi_I}-\cos{2\phi_R}}$, $a=\cosh^2{\phi_I}/\cos^2{\phi_R}$, $P_R=-2\rho_0\cos{\phi_R}\sinh{\phi_I}$, $P_I=2\rho_0\sin{\phi_R}\cosh{\phi_I}$, and $\phi=\phi_R+i\phi_I$ is a complex constant. This solution is also called a general breather (GB) since it is periodic both in space and in time. Two important special cases are the following: (i) When $\phi_R\neq 0$ and $\phi_I=0$, the GB solution corresponds to an Akhmediev breather (AB) which is periodic in space and localized in time. (ii) If we take $\phi_R=0$ and $\phi_I\neq 0$, (\[nls2\]) becomes a Ma breather (MB) which is periodic in time only and localized in space. #### (iii) Rogue wave: {#iii-rogue-wave .unnumbered} A further specialized structure, which is localized both in space and in time with a constant plane wave background, is the so called Peregrine soliton or RW [@pere]. It can be obtained by taking the limits $\phi_R=\epsilon\gamma$ and $\phi_I=\epsilon\delta$, and $\epsilon\to 0$, where $\gamma$, $\delta$ and $\epsilon$ are constants. Then Eq. (\[nls2\]) reduces to the following form, $$\begin{aligned} U(X,T)=\rho_0 \,\text{e}^{ i \theta }\left(1-\frac{4+8i\rho_0^2 T}{1+4\rho_0^2(X-kT)^2+4\rho_0^4T^2}\right). \label{nls3}\end{aligned}$$ Very often the above localized nonlinear wave is described as a wave that “appears from nowhere and disappears without a trace" [@osbrn:rato] as shown in Fig. \[figa\](c). It was first observed in the area of oceanography [@osbrn] and is traditionally defined as a wave whose wave-height (the distance from trough to crest) is more than twice the significant wave height (SWH). The latter is generally defined as the average wave height among one third of highest waves in a given time series [@osbrn; @khar:pelin]. Further, it has also been explained that the above structure (\[nls3\]) arises due to a modulation instability (MI) [@pere; @benj:feir] of the plane wave solution of the constant coefficient NLS equation. Very recently higher-order RWs (HRWs) which correspond to the higher-order rational solutions of the NLS equation (\[nls\]) have been deduced [@akmv:anki]. The explicit expressions for the second and third-order RWs of the NLS equation are given in appendix A. These HRWs have higher amplitudes than the first-order RW. The RWs have also been observed experimentally in physical systems such as water wave tank [@chab], capillary waves [@shatz] and nonlinear optics [@solli; @kibler]. Several theoretical studies on the dynamics of RWs in nonlinear fiber optics [@porse; @pors], plasma physics [@mose], laser-plasma interactions [@veldas], and even econophysics [@tan], described by scalar NLS equation, have been made in recent times. Now, as mentioned in the beginning, the dynamics of a cigar shaped BEC at absolute zero temperature is usually described by the mean-field GP equation (\[eq:1d-gp\]), which is a generalized form of the ubiquitous constant coefficient NLS equation (\[nls\]), for the wave function of the condensate. Since the NLS equation (\[nls\]) admits breather and RW solutions, it is natural to expect that RWs and breathers may also be found in BEC systems as well. In this context, the RWs can correspond to a sudden increase of peaks in the condensate clouds similar to the nature of high peaks in the open sea, while breathers are generalizations of the RWs. The formative mechanism for the matter RWs in BECs is the accumulation of energy and atoms towards its central part and their spreading out to a constant density background. The formation of matter breather is the periodic exchange of atoms between the profile and the plane wave background. From an experimental point of view, the existence of RW and breather structures in a BEC system can be effectively controlled by tuning the nonlinear interaction between atoms by Feshbach resonance technique [@dalf; @stali; @abdul] and modulating the trapping frequency of the external potential. It will therefore be of great interest to study the characteristics and/or controlling of structures of RWs and breathers due to their localization both in space and in time in BEC experiments. Past explorations of GP equation in BECs have paved the way for important developments in manipulating coherent matter waves for application, including atom interferometry [@inter], coherent atom transport [@mandel] and quantum information processing or quantum computation [@cirac]. Therefore, it is of high significance to study the dynamics of RWs and breather profiles of the GP equation (\[eq:1d-gp\]). However only few attempts have been made to identify and analyze the RWs and breather solutions of (\[eq:1d-gp\]) [@blud:kono; @yan; @wen:li; @zhao; @wu; @he]. To the best of our knowledge neither higher-order RW solutions (with certain free parameters) nor higher order breather solutions of (\[eq:1d-gp\]) have been taken up for study. Motivated by these observations, in this work, we construct the aforementioned localized and periodic solutions of (\[eq:1d-gp\]). Besides constructing these two families of solutions we also investigate how to manipulate the RWs and breathers through the effective scattering length and the strength of trap parameter. Having stated the motivation we now proceed to construct a transformation that transforms Eq. (\[eq:1d-gp\]) to the standard NLS equation (\[nls\]). Following the standard procedure [@raj:mur; @raj:ml] we find that the required similarity transformation should be of the form $$\begin{aligned} \label{a7} \psi(x,t)& = r_0\sqrt{R(t)}[U(X,T)] \notag \\ & \times \exp \left[{i\left(c_1r_0^2 Rx-\frac{R_t}{R}x^2-\frac{1}{2}c_1^2r_0^4\int{R^2(t)}dt\right)}\right],\end{aligned}$$ where \[a4\] $$\begin{aligned} X(x,t)= & \, r_0 R(t)x-c_1r_0^3\int{R^2(t)}dt, \\ T(t)= & \, r_0^2\int{R^2(t)}dt,\end{aligned}$$ and $U(X,T)$ is the solution of the standard NLS equation (\[nls\]). In the above $b,r_0$ are arbitrary constants and the modulational functions $R(t)$ and $\beta(t)^2$ should satisfy the following condition $$\begin{aligned} \label{a5} \frac{d}{dt}\left(\frac{R_t}{R}\right)-\left(\frac{R_t}{R}\right)^2+\beta(t)^2=0,\end{aligned}$$ which is a Riccati type equation with dependent variable $(R_t/R)$ and independent variable $t$. Regardless of what $R(t)$ is, as long as the condition (\[a5\]) is satisfied, the GP equation is integrable [@raj:mur; @raj:ml]. We also note here that the Painlevé singularity structure analysis performed on Eq. (\[eq:1d-gp\]) confirms the same restriction (\[a5\]) on the system coefficients [@pain]. We further note that the solution (\[a7\]) provides us some flexibility to generate new structures related to the RWs which may be useful for the BEC experiments. Even though one can arbitrarily choose the functions $R(t)$ and $\beta(t)$ that satisfy the constraint (\[a5\]) and generate the required solutions, in this paper we consider the trap frequency to be of the following three forms: (i) $\beta(t)^2=\beta_0^2$, (ii) $\beta^2(t)=\left({\beta_0^2}/{2} \right)\left[1-\tanh\left({\beta_0 t/}{2}\right)\right]$ and (iii) $\beta(t)^2=2\beta_0^2[1+3\tan^2(\beta_0 t)]$, where $\beta_0$ is a constant. As shown in reference [@raj:mur] the effective scattering length for these three cases turn out to be (i) $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$, (ii) $R(t)=1+\tanh \left({\beta_0 t}/{2}\right)$ and (iii) $R(t)=1+\cos{(2 \beta_0 t)}$. We consider the trap frequency to be in the above forms since it has been shown that they are valid forms in BEC experiments [@raj:mur]. We consider each one of the cases separately and substitute them in (\[a7\]) along with the RW and breather solutions of (\[nls\]). We then analyze in detail how the nature of the RW and breather structures get modified by the above two functions $\beta(t)$ and $R(t)$. Our analysis shows that the amplitude parameter $(r_0)$ plays a vital role in the formation of RW, and the trap frequency and the effective scattering length modify the structure of the RW and breather profiles, allowing one the possibility of manipulating the RWs and breathers in specific ways. The paper is organized as follows. In the following section, we construct RW solutions for time-independent and time-dependent traps and study their characteristics in detail. To the best of our knowledge, for the first time in the literature, we observe that the second- and third-order RWs transform to first-order RW-like structures when we tune a parameter which appears in the harmonic traps (time-independent and time-dependent traps). In Sec. III, we also construct RW solutions with free parameters which allow us to split the symmetric form solution into a multi-peaked solution for (\[eq:1d-gp\]) and investigate how these RW structures get modified in the plane wave background by increasing the strength of the trap. In Sec. IV, we construct one-breather and two-breather solutions of (\[eq:1d-gp\]) and investigate their characteristics when we alter the trap parameter. Finally, in Sec. V, we present a summary of the results and our conclusions. Characteristics of rogue waves ============================== To begin with, we consider the case in which the trap frequency is a constant, that is $\beta(t)^2$ = constant = $\beta_0^2$. A time-independent trap frequency implies that the frequency does not change with time and space. We then consider the trap frequency to be time-dependent and investigate the associated RW solutions. Time-independent trap --------------------- Substituting $\beta(t)^2=\beta_0^2$ in the integrability condition $(\ref{a5})$, we find that the time-dependent interaction term should be of the form $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$, where $\delta$ is an integration constant. {width="0.8\linewidth"} ![(Color online) (a) First-order RW in BEC and (b) the corresponding contour plot obtained by numerically solving Eq. (\[eq:1d-gp\]) through split-step Crank-Nicolson method for the time-dependent nonlinearity coefficient $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$ and time-independent trap frequency $\beta(t)^2=\beta_0^2$. The initial condition chosen corresponds to the analytic solution of Fig. \[fig8\](d).[]{data-label="num-rog"}](fig-num3){width="\linewidth"} Plugging this expression in (\[a7\]), we find $$\begin{aligned} \label{a15} \psi_j(x,t) = r_0\sqrt{{\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}}\, U_j(X,T) \eta(x,t),\end{aligned}$$ where $$\begin{aligned} \eta(x,t) = & \exp\bigg\{i\bigg[c_1 r_0^2 {\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}x \notag \\ & \left.\left. +\frac{(\beta_0^2 x^2-c_1^2r_0^4)\tanh{(\beta_0 t+\delta)}}{2\beta_0}\right]\right\}, \notag\end{aligned}$$ and $U_j(X,T)$’s, $j=1, 2, 3$, are the first, second and third-order RW solutions of the NLS equation (\[nls\]) whose explicit expressions are given in the Appendix (vide Eqs. (\[a8\]), (\[a11\]) and (\[a12\])). Also $X(x,t)$ and $T(t)$ have the forms as given in Eqs. (\[a4\]). {width="0.8\linewidth"} {width="0.8\linewidth"} {width="0.8\linewidth"} In Fig. \[fig8\], the first, second, and third columns represent the density profiles of the first, second, and third-order RWs obtained from (\[a15\]). In this figure we present the formation of RWs in cigar-shaped BECs. In the context of BECs, it is the fluctuation in the density of atoms, localized in space and time, which is what we observe as RWs. For instance consider the first-order RWs depicted in Figs. \[fig8\](a)-(d). These may be interpreted as follows: Atoms in the condensate, initially at a constant density background, suddenly accumulate to form a hump towards the center of the condensate at finite time while leaving voids in the density which appear as troughs in the RWs, depending on the initial state. The crest and troughs reach their extrema as time progresses, and then the condensate atoms spread out so as to recover the constant density background in finite time, thus revealing the unstable nature of RWs. Further, by tuning the amplitude parameter $r_0$ in (\[a15\]), we can visualize the formation and manipulation of RWs. For example, when we increase the amplitude parameter $r_0$ smoothly from $0.5$ to $1.2$, we can observe in Figs. \[fig8\](a)-(d) the formation of more and more localized first-order RWs with increasing amplitude. At $r_0=1.2$ one can visualize a large amplitude wave which is sufficiently localized both in space and in time which in turn confirms the formation of the first-order RW in BEC. The formation of second and third-order RWs is also demonstrated in Figs. \[fig8\](c)-(h) and \[fig8\](i)-(l), respectively, for the same set of parameter values of $r_0$. From these figures we infer that the time-dependent nonlinear interaction between the atoms induces density fluctuations over the condensate which gets more and more localized both in space and in time as we increase the order of the RW. In order to confirm the existence of the RWs further, we have also performed a direct numerical simulation of (\[eq:1d-gp\]) with the aid of the split-step Crank-Nicolson method using an initial wave function which is the same as the function (\[a15\]) and with space step $dx=0.01$ and time step $dt=0.001$ [@Muruganandam+]. In Fig. \[num-rog\], we have presented the computer generated density profile of the first-order RW and the corresponding contour plot with the parameters chosen as $r_0=1.2$, $c_1=0.01$, $\beta_0=0.1$, and $\delta=0.01$ which are same as that of Fig. \[fig8\](d). One can easily observe a very good agreement between the numerical results and the analytical predictions for the emergence of the RWs. We have also verified numerically the presence of second- and third-order RWs of (\[eq:1d-gp\]) as well, replicating Figs. \[fig8\]. Next we demonstrate how these localized structures vary with respect to the trap parameter $\beta_0$. Fig. \[fig1\] displays the first-order RW for the same nonlinearity management parameter $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$ and the external trap frequency $\beta(t)^2=\beta_0^2$. The nature of the first-order RW for $\beta_0=0.1$ is depicted in Fig. \[fig1\](a). When we increase the strength of the trap parameter the density profiles corresponding to the first-order RW become more and more localized in time as shown in Figs. \[fig1\](b) and \[fig1\](c), respectively. The corresponding contour plots are given in Figs. \[fig1\](d)-(f). Fig. \[fig1a\] displays the density profiles of the second-order RWs for the same nonlinearity management parameter as a function of $\beta_0$. In Fig. \[fig1a\](a) we display the second-order RW for $\beta_0=0.1$. When the strength of the parameter $\beta_0$ is increased to $1.2$, the wave subcrests start to stretch as shown in Fig. \[fig1a\](b). The wave subcrests become more and more localized in time when we increase the value of $\beta_0$ and finally the second-order RW attains a new structure as given in Fig. \[fig1a\](c) for $\beta_0 = 5.0$. The resultant structure looks almost like a first-order RW, see Fig. \[fig1\](c). Figs. \[fig1a\](d)-(f) are the corresponding contour plots. We also observe similar effects in the third-order RW case as well when we increase the value of $\beta_0$. Fig. \[fig3a\] demonstrates the changes in the third-order RW when we vary the interaction strength. At $\beta_0=1.5$, the third-order RW transforms into the second-order RW-like structure as shown in Fig. \[fig3a\](b). When we increase the value of the parameter $\beta_0$ to $5.0$, one obtains a first-order like RW which is displayed in Fig. \[fig3a\](c). These facts are also confirmed by the corresponding contour plots, Figs. \[fig3a\](d)-(f). The aforementioned results reveal that when we increase the strength of the trap parameter $\beta_0$, the second and third-order RWs become more localized in time and delocalized in space, approaching the structure of a first-order RW. Thus the robustness of the density profiles can be controlled by varying the strength of trap frequencies. Time-dependent monotonous trap ------------------------------ Next we consider the time-dependent trap frequency in the form $\beta^2(t)=\left({\beta_0^2}/{2} \right)\left[1-\tanh\left({\beta_0 t/}{2}\right)\right]$. {width="0.8\linewidth"} {width="0.9\linewidth"} {width="0.9\linewidth"} For this choice, the relation (\[a5\]) fixes the interatomic interaction term to be of the form $R(t)=1+\tanh \left({\beta_0 t}/{2}\right)$. The first, second, and third-order RW solutions for this trap frequency and strength of interatomic interaction are found to be $$\begin{aligned} \label{a22} \psi_j(x,t) = r_0\sqrt{1+\tanh\Big(\frac{\beta_0}{2}t\Big)} \, U_j(X,T) \, \eta(x,t),\end{aligned}$$where $j = 1, 2, 3$ and $$\begin{aligned} \eta(x,t) = \exp\left\{i \left[\frac{\beta_0{\normalfont\mbox{sech}\,}^2\left(\frac{\beta_0 t}{2}\right)x^2}{4\left[1+\tanh(\frac{\beta_0 t}{2})\right]}-c_1r_0^2\left[1+\tanh{\left(\frac{\beta_0 t}{2}\right)}\right]x +\frac{c_1^2r_0^4\left(\beta_0 t+2\log\left[\cosh(\frac{\beta_0 t}{2})-\tanh(\frac{\beta_0 t}{2})\right]\right)}{\beta_0}\right]\right\}. \notag\end{aligned}$$ The qualitative nature of the first, second, and third-order RWs for $R(t)=1+\tanh\big({\beta_0 t}/{2}\big)$ and $\beta(t)^2={\beta_0^2}/{2}\left[1-\tanh\left({\beta_0 t}/{2}\right)\right]$ turns out to be the same as in the previous case (Fig. \[fig8\](d), (h) and (l)) when the amplitude parameter $r_0$ is varied and so we do not display the outcome here separately. On the other hand, we identify interesting structures while varying the parameter $\beta_0$ which is discussed in the following. In Fig. \[fig4\], we depict the first-order RW for these choices of $R(t)$ and $\beta(t)$. When $\beta_0 = 0.1$ the first-order RW is as shown in Fig. \[fig4\](a) (see also Fig. \[fig4\](d) for the corresponding contour plot). By altering the value of the trap parameter $\beta_0$ to $1.0$, the structure of the first-order RW gets modified as shown in Fig. \[fig4\](b)/Fig. \[fig4\](e). One can also see from Fig. \[fig4\] that, as the trap parameter $\beta_0$ is increased, the RW gradually becomes more localized in time and the condensate atoms settle down to a slightly higher density background due to the attractive nature of the potential. When $\beta_0 = 5.0$ the modified structure of first-order RW is given in Fig. \[fig4\](c)/Fig. \[fig4\](f), where this feature is even more prominent. The density profiles of the corresponding second-order RWs are presented in Fig. \[fig5\]. When we tune the parameter $\beta_0$ from $0.1$ upwards the wave subcrests start to stretch. From the contour plots we can observe that the stretches occur on one side of the RW only. When we increase the value of $\beta_0$ further the second-order RW gets modified to a first-order RW like structure which is demonstrated in Fig. \[fig5\](c). A similar transition has also been observed in the third-order RW case as well which is illustrated in Fig. \[fig6\]. The third-order RW acquires a new structure as shown in Fig. \[fig6\](b)/Fig. \[fig6\](e) when we increase the value of the parameter $\beta_0$ from $0.1$ to $1.0$. At $\beta_0=5.0$, we observe that the third-order RW acquires a further modified structure which is displayed in Fig. \[fig6\](c) and Fig. \[fig6\](f). Note the similarity in the central part with that of the first-order RW as given in Fig. \[fig4\](c) and Fig. \[fig4\](f). Time-dependent periodic trap ---------------------------- In the case of the third choice we consider the time-dependent periodic trap frequency to be of the form $\beta(t)^2=2\beta_0^2[1+3\tan^2(\beta_0 t)]$ so that the strength of the time-dependent periodic interatomic interaction turns out to be $R(t)=1+\cos{(2 \beta_0 t)}$. Substituting these two expressions in (\[a7\]), we find $$\begin{aligned} \label{a24} \psi_j(x,t)=r_0\sqrt{1+\cos{(2\beta_0 t)}} \, U_j(X,T) \, \eta(x,t),\end{aligned}$$ where $$\begin{aligned} \eta(x,t)= & \exp\bigg\{i\bigg[\beta_0 \tan{(\beta_0 t)}x^2+2c_1 r_0^2 \cos{(\beta_0 t)^2} x \notag \\ & \left.\left.-\frac{c_1^2 r_0^4(12\beta_0 t+8\sin{(2\beta_0 t)}+\sin{(4\beta_0 t)})}{16\beta_0}\right]\right\}.\notag\end{aligned}$$ Here $U_j(X,T)$, $ j = 1, 2, 3$, are again the first, second and third-order RW solutions of the NLS equation (vide Eqs. (\[a8\]), (\[a11\]), and (\[a12\])). {width="0.8\linewidth"} In Fig. \[fig7\], we present the density profiles of the first, second, and third-order RWs (first row) and their corresponding contour plots (second row) for the strength of time-dependent interatomic interaction $R(t)=1+\cos{(2 \beta_0 t)}$ and time-dependent periodic trap frequency $\beta(t)^2=2\beta_0^2[1+3\tan^2(\beta_0 t)]$ . Fixing the value of $\beta_0$ at $2.5$, from Fig. \[fig7\] we see that the RW exists on a periodic background for the above forms of $R(t)$ and $\beta^2(t)$. Further we have also verified that the behaviour of the RWs as a function of $\beta_0$ follows the same qualitative picture discussed above for the other forms of $R(t)$ and $\beta^2(t)$. Characteristics of triplet RWs ============================== Very recently it has been shown that one can generalize the expressions for the higher order RW solutions of the scalar NLS equation (\[nls\]) given in Appendix A further by introducing certain free parameters which allow one to split a symmetric form RW solution into a multi-peaked solution and that by varying these free parameters one can extract certain novel patterns of RWs [@ankie]. The introduction of free parameters decompose the higher-order RW solutions into $n(n+1)/2$ first-order forms, where $n$ is the order of the RW, however maintaining the symmetry of the higher-order solutions even in their decomposed forms. Further the free parameters are shown to determine the size and orientation of the first-order solutions [@ankie]. Triplets are symmetry preserving first-order structures revealing the fact that the (higher-) second-order RW solution is a family of three first-order rational solutions. It is evident that the existence of triplets in an ocean corresponds to three big waves on the water surface in a row or “ three sister waves" [@juli]. Inspired by such a possibility, in the following, we consider the second- and third-order RW solutions of (\[eq:1d-gp\]) with suitable free parameters and analyze the symmetrical structures that arise due to these free parameters when we vary the strength of nonlinearity and the trap parameter. To begin with, we confine our attention to the second-order RW solution. In this case, we have the following modified expressions [@ankie] for $G_2$, $H_2$ and $D_2$ in Eq. (\[a9\]), that is $$\begin{aligned} G_2 = &\, 12 \big[3-16 X^4-24 X^2(4 T^2+1)-48 l X-80 T^4 \notag \\ &\, -72 T^2-48 m T\big], \notag \\ H_2 = &\, 24 \big\{ T\big[ 15-16 X^4+24 X^2-48 l X-8(1-4 X^2)T^2 \notag \\ &\, -16 T^4 \big]+6m(1-4 T^2+4X^2)\big\}, \notag \\ D_2 = &\, 64 X^6+48 X^4 (4T^2+1)+12X^2(3-4T^2)^2+64 T^6\notag \\ &\, +432 T^4+396 T^2+9 +48m\big[18m+T(9+4 T^2 \notag \\ &\, -12X^2)\big]+48l\big[(18l+X(3+12T^2 -4X^2)\big]. \end{aligned}$$ Note that this RW solution now contains two free parameters, namely $l$ and $m$. The parameters $l$ and $m$ describe the relative positions of the first-order RWs in the triplet. Substituting the above expressions in (\[a15\]), (\[a22\]) and (\[a24\]), for $j=2$, we obtain the corresponding second-order RW solutions to the GP equation (\[eq:1d-gp\]) with the free parameters $l$ and $m$ included. {width="0.8\linewidth"} When $l= m=0$, this solution coincides with the one given earlier (vide Eq. (\[a11\])) which contains one largest crest and four subcrests with two deepest troughs (Fig. \[fig8\](h)). When $l$ and $m$ are not equal to $0$, the second-order RW splits into three first-order RWs. These waves emerge in a triangular fashion (a triplet pattern). The three first-order RWs form a triangular pattern with $120$ degrees of angular separation between them [@ankie]. In Figs. \[fig8b\](a)-(c) we display the triplet pattern for $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$ and $\beta(t)^2=\beta_0^2$ when $l=15$ and $m=25$. The triplet RW pattern for $\beta_0=0.1$ is shown in Fig. \[fig8b\](a). When we increase the parameter $\beta_0$ to $0.3$ we observe that the triplet pattern has started to collapse as shown in Fig. \[fig8b\](b) and a complete collapse is observed as in Fig. \[fig8b\](c) when we increase the $\beta_0$ value further to $1.0$. Figs. \[fig8b\](d)-(g) represent the triplet pattern for $R(t)=1+\tanh\big({\beta_0 t}/{2}\big)$ and $\beta(t)^2={\beta_0^2}/{2}\left[1-\tanh\left({\beta_0 t}/{2}\right)\right]$. The formation of triplet RWs is shown in Fig. \[fig8b\](d) when $\beta_0=0.1$. When we increase the value $\beta_0$ to $0.3$ one of the single RWs in the triplet pattern vanishes which is illustrated in Fig. \[fig8b\](e). By increasing the parameter $\beta_0$ further we observe that two first-order RWs are more localized in time as shown in Fig. \[fig8b\](f). Figs. \[fig8b\](g)-(i) represent the triplet pattern for $R(t)=1+\cos{(2 \beta_0 t)}$ and $\beta(t)^2=2\beta_0^2[1+3\tan^2(\beta_0 t)]$. The form of the triplet pattern for $\beta_0=0.1$ is displayed in Fig. \[fig8b\](g). Here also when we increase the value of $\beta_0$ we observe the collapse of the triplet pattern in the periodic wave background (vide Fig. \[fig8b\](i)). {width="0.9\linewidth"} We then move on to investigate the structure of the third-order RW solution with four free parameters, namely $l,m,g$ and $h$. The third-order RW solution with four free parameters is much lengthier than the one without free parameters and so we do not give the explicit expression here and analyze the results only graphically. Here also we analyze the solution with respect to the free parameters. When $l= m= g= h=0$, we have the classical third-order RW solution which is shown in Fig. \[fig8\](l). It has one largest crest and six subcrests with two deepest troughs. For non-zero values of $l$, $m$, $g$ and $h$, the third-order RW splits into six separated first-order RWs. When we increase the value of the free parameters, the six first-order RWs take new positions. The sextet pattern is displayed in Figs. \[fig8h\](a)-(c) for $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$ and $\beta(t)^2=\beta_0^2$ when $l=10$, $m=20$, $g=500$ and $h=500$. For $\beta_0=0.1$ the set of six first-order RWs is shown in Fig. \[fig8h\](a). When we increase the value of $\beta_0$ to $0.25$, three peaks disappear and the remaining peaks start to bend as shown in Fig. \[fig8h\](b). When we increase the value of $\beta_0$ further the RWs bend in the plane wave background as given in Fig. \[fig8h\](c). Figs. \[fig8h\](d)-(f) represent six first-order RWs for the time-dependent nonlinearity strength $R(t)=1+\tanh\big({\beta_0 t}/{2}\big)$ and the time-dependent external trap frequency $\beta(t)^2={\beta_0^2}/{2}\left[1-\tanh\left({\beta_0 t}/{2}\right)\right]$. The formation of six first-order RWs at $\beta_0=0.1$ is shown in Fig. \[fig8h\](d). When we increase the strength of the parameter $\beta_0$ to $0.25$, one of six first-order RWs vanishes as seen in Fig. \[fig8h\](e). If we increase the value $\beta_0$ further three out of six peaks bend in the plane wave background and eventually collapse which is shown in Fig. \[fig8h\](f). Figs. \[fig8h\](g)-(i) represent the sextet pattern for $R(t)=1+\cos{(2 \beta_0 t)}$ and $\beta(t)^2=2\beta_0^2[1+3\tan^2(\beta_0 t)]$. The form of the six first-order RWs for $\beta_0=0.05$ is displayed in Fig. \[fig8h\](g) and for further increase in $\beta_0$, the modified structures in the periodic wave background are as shown in Figs. \[fig8h\](h) and \[fig8h\](i). Characteristics of Breathers ============================ In the previous two sections we have analyzed how the RW profiles get modified by the variations of the distributed coefficients present in the variable coefficient NLS Eq. (\[eq:1d-gp\]). In this section we analyze how the breather structures get modified in the condensates when we vary the strength of the external trap parameter. To begin with we consider the first-order breather solution of the NLS equation (\[nls\]), which is given in [@eleon] and is a special case of the GB solution (\[nls2\]), $$\begin{aligned} \tilde{U_1}(X,T) = & \,\biggr[\frac{f^2 \cosh[\alpha (T - T_1)] + 2 i \,f v \sinh[\alpha (T - T_1)]}{2 \cosh[\alpha (T - T_1)] - 2v \cos[ f (X - X_1) ]} \notag \\ & \, -1\biggr] \times \exp{(iT)}, \label{b1}\end{aligned}$$where the parameters $f$ and $v$ are expressed in terms of a complex eigenvalue (say $\lambda$), that is $f=2\sqrt{1+\lambda^2}$ and $v = \mbox{Im}(\lambda)$, and $X_1$ and $T_1$ serve as coordinate shifts from the origin. The real part of the eigenvalue represents the angle that the one-dimensionally localized solutions form with the $T$ axis, and the imaginary part characterizes frequency of periodic modulation. The parameter $\alpha$ $(=f v)$ in (\[b1\]) is the growth rate of modulation instability. Substituting this breather solution of the NLS equation into (\[a15\]), (\[a22\]) and (\[a24\]), we study the underlying dynamics of (\[eq:1d-gp\]). ![(Color online) (a) AB for $\lambda=0.6 i$, (b) MB for $\lambda=1.4 i$ when $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$ and $\beta(t)^2=\beta_0^2$. (c) and (d) are their corresponding contour plots. The other parameters are $r_0=1.0$, $\beta_0=0.01$, $c_1=0.01$, and $\delta=0.01$. []{data-label="fig9"}](1D-sech-bre){width="0.95\linewidth"} For illustration, we consider the case $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$ and $\beta(t)^2=\beta_0^2$, and plot the outcome in Fig. \[fig9\]. The first row in Fig. \[fig9\] represents an Akhmediev breather (AB) solution for the eigenvalue $\lambda=0.6 i$ and Ma breather (MB) solution for $\lambda=1.4 i$ and the second row represents their contour plots. The AB and MB solutions are localized in time and space, respectively, as discussed in Sec. I. This is clearly demonstrated in the two columns of Fig. \[fig9\]. When we tune the parameter $\beta_0$, a new structure against a breather background is obtained. To visualize this we fix the value of $\beta_0$ to be $0.8$. For this value a stretching occurs in space in the case of AB (Fig. \[fig9a\](a)) whereas in the case of MB (Fig. \[fig9a\](b)) the stretching occurs in time. Figs. \[fig9a\](c)-(d) illustrate the AB and MB profiles for the time-dependent nonlinearity coefficient $R(t)=1+\tanh\big({\beta_0 t}/{2}\big)$ and time-dependent trap frequency $\beta(t)^2=({\beta_0^2}/{2})\left[1-\tanh\left({\beta_0 t}/{2}\right)\right]$. Here also we tune the strength of the trap parameter $\beta_0$ to 0.8 and observe that stretching occurs over space in AB which is depicted in Fig. \[fig9a\](c) and the MB gets more localized and when $t\leq 0$ the breather profile completely disappears as shown in Fig. \[fig9a\](d). Our results reveal the fact that when we tune the parameter $\beta_0$ in the obtained breather solution, the breather gets a modified structure corresponding to a distortion of the breather profile. ![(Color online) Stretching of breathers. (a), (c) and (e) AB with an eigenvalue $\lambda=0.6 i$, (b), (d) and (f) MB with $\lambda=1.4 i$ for the three different forms of $R(t)$ discussed in the text. (a)-(d) $\beta_0=0.8$, (e)-(f) $\beta_0=1.5$. The other parameters are $r_0=1.0$, $c_1=0.01$ and $\delta=0.01$.[]{data-label="fig9a"}](1D-comb-bre){width="0.95\linewidth"} Figs. \[fig9a\](e) and (f) respectively represent the AB and MB in the periodic background for $R(t)=1+\cos{(2 \beta_0 t)}$ and $\beta(t)^2=2\beta_0^2[1+3\tan^2(\beta_0 t)]$ when $\beta_0=1.5$. {width="0.8\linewidth"} Next, we proceed to construct the two-breather solutions of (\[eq:1d-gp\]) and analyze how these solutions are distorted by the variations of modulation parameters. The two-breather solution of the NLS equation is given by [@kadz] $$\begin{aligned} \label{2b1} \tilde{U_2}(X,T)=\left[1+\frac{\tilde{G_2}(X,T)+i \tilde{H_2}(X,T)}{\tilde{D_2}(X,T)}\right]\exp{(iT)},\end{aligned}$$ where $\tilde{G_2}$, $\tilde{H_2}$, and $\tilde{D_2}$ are given by $$\begin{aligned} \tilde{G_2} = &\, -(k_1^2-k_2^2) \biggr[\frac{k_1^2\delta_2}{k_2}\cosh(\delta_1T_{s1})\cos(k_2X_{s2}) \notag \\ &\, - (k_1^2-k_2^2)\cosh(\delta_1T_{s1})\cosh(\delta_2T_{s2}) \notag \\ &\, -\frac{k_2^2\delta_1}{k_1}\cosh(\delta_2T_{s2})\cos(k_1 X_{s1})\biggr], \end{aligned}$$ $$\begin{aligned} \tilde{H_2} = &\, -2(k_1^2-k_2^2)\biggr[\frac{\delta_1\delta_2}{k_2}\sinh(\delta_1 T_{s1})\cos(k_2 X_{s2})\notag \\ &\, - \frac{\delta_1 \delta_2}{k_1}\sinh(\delta_2T_{s2}) \cos(k_1 X_{s1}) \notag \\ &\, -\delta_1 \sinh(\delta_1 T_{s1})\cosh(\delta_2 T_{s2}) \notag \\ &\, + \delta_2 \sinh(\delta_2 T_{s2})\cosh(\delta_1 T_{s1})\biggr], \end{aligned}$$ $$\begin{aligned} \tilde{D_2} = &\, 2 (k_1^2 + k_2^2) \frac{\delta_1 \delta_2}{k_1 k_2} \cos(k_1X_{s1})\cos(k_2 X_{s2}) \notag \\ &\, + 4 \delta_1 \delta_2 (\sin(k_1 X_{s1}) \sin(k_2 X_{s2}) \notag \\ &\, + \sinh(\delta_1 T_{s1})\sinh(\delta_2 T_{s2})) \notag \\ &\,- (2 k_1^2 - k_1^2 k_2^2 + 2 k_2^2) \cosh(\delta_1 T_{s1})\cosh(\delta_2 T_{s2}) \notag \\ &\, - 2 (k_1^2 - k_2^2) \biggr[\frac{\delta_1}{k_1}\cos(k_1 X_{s1})\cosh(\delta_2 T_{s2}) \notag \\ &\, - \frac{\delta_2}{k_2}\cos(k_2 X_{s2}) \cosh(\delta_1 T_{s1})\biggr], \end{aligned}$$ where the modulation frequencies, $k_j=2\sqrt{1+\lambda_j^2}$, $j=1,2$, are described by the (imaginary) eigenvalues $\lambda_j$. In the above expressions, $X_j$, $T_j$, $j=1,2$, represent the shifted point of origin, $\delta_j$ $(=k_j\sqrt{4-k_j^2}/2)$ is the instability growth rate of each component and $X_{sj}=X-X_j$ and $T_{sj}=T-T_j$ are shifted variables. {width="0.8\linewidth"} With two purely imaginary eigenvalues, $\lambda_j$, $j=1,2$, the solution (\[2b1\]) is capable of describing a variety of possible second-order breather structures. The solution includes ABs, MBs and the intersection of AB and MB solutions for certain combinations of eigenvalues. For example, when the imaginary parts of both the eigenvalues $\mbox{Im}(\lambda_j)$, $j=1,2$, lie between $0$ and $1$, we obtain the ABs. On the other hand when both of them are greater than one ($\mbox{Im}(\lambda_j)>1$) we obtain the MBs and in the mixed possibility case, that is one of the eigenvalues is less than one ($\mbox{Im}(\lambda_1)<1$) and the other eigenvalue ($\mbox{Im}(\lambda_2>1$) is greater than one, we obtain the intersection of AB and MB solutions. Substituting the two-breather NLS solution (\[2b1\]) in (\[a15\]), (\[a22\]) and (\[a24\]) we obtain the general two-breather solution of (\[eq:1d-gp\]). Fig. \[br2\] displays the evolution of the two-breather solution of (\[eq:1d-gp\]) for $R(t)={\normalfont\mbox{sech}\,}{(\beta_0 t+\delta)}$ and $\beta(t)^2 =\beta_0^2 = (0.01)^2 $ with imaginary eigenvalues. To obtain the ABs from (\[2b1\]) we consider the situation where the magnitudes of both the eigenvalues $\lambda_1$ and $\lambda_2$ are less than $1$ ($\lambda_1=0.55i$ and $\lambda_2=0.75i$). One AB developing with a time delay after another is shown in Fig. \[br2\](a), while in Fig. \[br2\](b) we present the case where there is no such time delay. In Fig.  \[br2\](c) we depict the case when one AB along with a RW coexist for the choice $\lambda_1=0.55i$ and $\lambda_2=0.99i$. When we take the eigenvalues $\lambda_1$ and $\lambda_2$ to be $1.3i$ and $1.4i$, respectively, we obtain two MB solutions. Similarly the evolution of the two MBs with and without spatial delay is shown in Figs. \[br2\](d) and \[br2\](e), respectively. We also observe that the distance between the MBs increases when we set both the eigenvalues to be nearly equal, say for example $\lambda_1=1.3i$ and $\lambda_2=1.31i$ which is not displayed here. When we take the eigenvalues as $\lambda_1=0.5i$ and $\lambda_2=1.3i$, the AB intersects with the MB which is displayed in Fig. \[br2\](f). When we tune the strength of the trap parameter $\beta_0$ to $0.15$, both the ABs get stretched in the plane wave background which is demonstrated in Fig. \[br2a\](a). When $\beta_0=0.5$, in the intersection of AB and MB solutions we observe that the AB gets a bending structure while the MB fully disappears in the plane wave background which is shown in Fig. \[br2a\](b). In Fig. \[br2a\](c) we note that both the MBs develop a bending structure in the plane wave background when $\beta_0=0.2$. Figs. \[br2a\](d)-(f) display the evolution of the two-breather solution of (\[eq:1d-gp\]) for $R(t)=1+\tanh\big({\beta_0 t}/{2}\big)$ and $\beta(t)^2={\beta_0^2}/{2}\left[1-\tanh\left({\beta_0 t}/{2}\right)\right]$ with the imaginary eigenvalues. When $\beta_0 = 0.15$ one of the ABs gets stretched which is not shown here and on further increase of the value of $\beta_0$ to $1.5$, one of the ABs gets annihilated and the other AB bends in the plane wave background which is demonstrated in Fig. \[br2a\](d). When $\beta_0 = 0.5$ the intersection of AB-MB structures is as shown in Fig. \[br2a\](e). In Fig.  \[br2a\](f) we observe that both the MBs get a bending structure in the plane wave background. Figs. \[br2a\](g)-(i) show the evolution of the two-breather solution of (\[eq:1d-gp\]) for $R(t)=1+\cos{(2 \beta_0 t)}$ and $\beta(t)^2=2\beta_0^2[1+3\tan^2(\beta_0 t)]$ with the imaginary eigenvalues. Here also when we increase the value of $\beta_0$ we observe the collapse of the two-breather solution in the periodic wave background. Conclusion ========== In this work, we have constructed higher order RW solutions with and without free parameters for the quasi one-dimensional GP equation with time-dependent interatomic interation and external trap through the similarity transformation technique. By mapping the variable coefficient NLS equation onto the constant coefficient NLS equation we have derived these solutions. We have shown that the mapping can be done when the external trap and the nonlinearly interatomic interaction of atoms satisfy a constraint. From the known higher order RW and breather solutions of the constant coefficient NLS equation, we have derived the solutions of (\[eq:1d-gp\]). In our analysis, we have considered the harmonic trap frequency in three different forms, namely (i) time-independent expulsive trap, (ii) time-dependent monotonous trap, and (iii) time-dependent periodic trap, and correspondingly fixed the effective scattering length. We then studied the characteristics of the constructed RW solutions in detail. We have observed that the second- and third-order RWs transform to first-order RW-like structures when a parameter appearing in the harmonic trap (time-independent and time-dependent traps) is varied. We have then analyzed the characteristics of triplet and sextet patterns of matter RWs for (\[eq:1d-gp\]). We have also constructed one-breather and two breather solutions of (\[eq:1d-gp\]). We have investigated how these periodic localized waves change in the plane wave background when we tune the trap parameter in the obtained breather solutions. Our results may provide possibilities to manipulate RWs experimentally in a BEC system. KM thanks the University Grants Commission (UGC-RFSMS), Government of India, for providing a research fellowship. The work of PM forms part of Department of Science and Technology (Ref. No. SR/S2/HEP-03/2009) and Council of Scientific and Industrial Research (Ref. No. 03(1186)/10/EMR-II), Government of India funded research projects. The work of MS forms part of a research project sponsored by NBHM, Government of India. The work forms part of an IRHPA project and a Ramanna Fellowship project of ML, sponsored by the Department of Science and Technology (DST), Government of India, who is also supported by a DAE Raja Ramanna Fellowship. In the absence of the trap, and the nonlinearity strength $R(t)$ is equal to one, Eq. (\[eq:1d-gp\]) reduces to the standard NLS equation $(\ref{nls})$. Several localized and periodic structures of standard NLS are documented in the literature [@akmv:anki; @eleon; @kadz]. Equation $(\ref{nls})$ admits $N^{th}$ order RW solution. We present the RW solution of the NLS equation in the following form [@akmv:anki], $$\begin{aligned} \label{a6} U_j(X,T)=\left[(-1)^j+\frac{G_j(X,T)+iTH_j(X,T)}{D_j(X,T)}\right]\exp{(iT)},\end{aligned}$$ where $j=1,2,..,G_j,H_j$ and $D_j$ are polynomials in the variables $T$ and $X$. For the first-order $(j=1)$ RW solution $G_1=4$, $H_1=8$ and $D_1=1+4X^2+4T^2$. From (\[a6\]) we get $U_1=\left(4\frac{1+2iT}{1+4X^2+4T^2}-1\right)\exp{(iT)}$. For covenience we multiply this expression by $-1 = \exp[i\pi]$ and consider the solution in the form $$\begin{aligned} \label{a8} U_1(X,T)=\left(1-4\frac{1+2iT}{1+4X^2+4T^2}\right)\exp{[iT]}.\end{aligned}$$ We use only this form of expression in our analysis. This solution is same as the one given in Eq. (\[nls3\]) when $k=0$, $\omega=1$ and $\rho_0=1$. For the second-order $(j=2)$ RW solution, the function $G_2$, $H_2$, and $D_2$ turn out to be [@akmv:anki] $$\begin{aligned} \label{a9} G_2 = & \frac{3}{8}-3X^2-2X^4-9T^2-10T^4-12X^2T^2, \notag \\ H_2 = & \frac{15}{4}+6X^2-4X^4-2T^2-4T^4-8X^2T^2\end{aligned}$$ and $$\begin{aligned} D_2 = & \frac{1}{8}\left(\frac{3}{4}+9X^2+4X^4+\frac{16}{3}X^6+33T^2+36T^4 \right. \notag \\ & \left.+\frac{16}{3}T^6-24X^2T^2+16X^4T^2+16X^2T^4\right). \notag\end{aligned}$$ Then we get $$\begin{aligned} \label{a11} U_2(X,T)=\left[1+\frac{G_2+iTH_2}{D_2}\right]\exp{(iT)}.\end{aligned}$$ The profiles of second-order RW and third-order RW of the constant coefficient NLS equation are shown in Fig. \[nls-rog\]. ![(Color online) Profiles. (a) Second-order RW and (b) Third-order RW of the standard NLS equation.[]{data-label="nls-rog"}](nls-rogue){width="0.99\linewidth"} For the third-order $(j=3)$ RW solution, we have $$\begin{aligned} \label{a12} U_3(X,T)=\left[-1+\frac{G_3+iTH_3}{D_3}\right]\exp{(iT)},\end{aligned}$$ where $$\begin{aligned} G_3(X,T)= & \, g_0 + (2 T)^2 g_2 + (2 T)^4 g_4 + (2 T)^6 g_6 \notag \\ & + (2 T)^8 g_8 + (2 T)^{10} g_{10},\end{aligned}$$ with $$\begin{aligned} \label{solg} g_0 = & 1 - (2 X)^2 - \frac{2}{3} (2 X)^4 + \frac{14}{45} (2 X)^6 + \frac{(2 X)^8}{45} + \frac{(2 X)^{10}}{675},\notag \\ g_2 = & -3 - 20 (2 X)^2 + \frac{2}{3}(2 X)^4 - \frac{4}{45} (2 X)^6 + \frac{(2 X)^8}{45},\notag \\ g_4 = & 2 \left[-\frac{17}{3} + 5 (2 X)^2 - \frac{(2 X)^4}{3^2} + \frac{(2 X)^6}{3^3}\right]\notag \\ g_6 = & \frac{2}{45}\left[73 +14 (2 X)^2 + \frac{7}{3}(2 X)^4\right],\notag \\ g_8 = & \frac{1}{15} (11 + (2 X)^2), \;\;\;\; g_{10} = \frac{11}{675}\end{aligned}$$ and $$\begin{aligned} H_3(X,T)= & \, h_0 + (2 T)^2 h_2 + (2 T)^4 h_4 + (2 T)^6 h_6 \notag \\ & + (2 T)^8 h_8 + (2 T)^{10} h_{10}, \end{aligned}$$ with $$\begin{aligned} h_0 = & 2 \left[7 + 7 (2 X)^2 - 2 (2 X)^4 - \frac{2}{3^2} (2 X)^6 - \frac{(2 X)^8}{45} \right. \notag \\ & \left. + \frac{(2 X)^{10}}{675} \right], \notag \\ h_2 = & \frac{2}{3} \left[-11 - 28 (2 X)^2 - 2 (2 X)^4 - \frac{28}{45} (2 X)^6 + \frac{(2 X)^8}{45}\right],\notag \\ h_4 = & \frac{4}{15} \left[-107 + 19 (2 X)^2 - \frac{7}{3} (2 X)^4 + \frac{(2 X)^6}{3^2}\right],\notag \\ h_6 = & \frac{4}{45} \left[-29 - 2 (2 X)^2 +\frac{(2 X)^4}{3}\right],\notag \end{aligned}$$ $$\begin{aligned} h_8 = & \frac{2}{3^3} \left[1 + \frac{(2 X)^2}{5}\right], \;\;\; h_{10} =\frac{2}{675}.\end{aligned}$$ The denominator is represented by $$\begin{aligned} D_3(X,T)=& \, d_0 + (2 T)^2 d_2 + (2 T)^4 d_4 + (2 T)^6 d_6 + (2 T)^8 d_8 \notag \\ & + (2 T)^{10} d_{10}+ (2 T)^{12} d_{12},\notag \\\end{aligned}$$ where $$\begin{aligned} d_0 = & \frac{1}{2^3} \left[1 + 6 (2 X)^2 + \frac{5}{3} (2 X)^4 + \frac{52}{45} (2 X)^6 + \frac{(2 X)^8}{15} \right.\notag \\ & \left. + \frac{2}{675} (2 X)^{10} + \frac{(2 X)^{12}}{2025}\right],\notag \end{aligned}$$ $$\begin{aligned} d_2 = & 23 - 9 (2 X)^2 + \frac{10}{3} (2 X)^4 + \frac{2}{15} (2 X)^6 - \frac{(2 X)^8}{45} \notag \\ & + \frac{(2 X)^{10}}{675},\notag \\ d_4 = & 2 \left[71 + \frac{116}{3} (2 X)^2 - \frac{2}{3} (2 X)^4 - \frac{4}{45} (2 X)^6 + \frac{(2 X)^8}{135}\right],\notag \end{aligned}$$ $$\begin{aligned} d_6 = & \frac{32}{3} \left[\frac{17}{3} + 5 (2 X)^2 +\frac{(2 X)^4}{45} + \frac{(2 X)^6}{135}\right],\notag \\ d_8 = & \frac{32}{15} \left[\frac{83}{3} + 2 (2 X)^2 +\frac{(2 X)^4}{3^2}\right),\notag \\ d_{10} = & \frac{2^8}{225} \left[7 + \frac{(2 X)^2}{3}\right], \;\;\; d_{12} = \frac{2^9}{2025}.\end{aligned}$$ [90]{} S. Ponomarenko and G.P. Agrawal, Phys. Rev. Lett. [**97**]{}, 13901 (2006). V.N. Serkin, A. Hasegawa, and T.L. Belyaeva, Phys. Rev. Lett. [**98**]{}, 074102 (2007). M. Centurion, M.A. Porter, P.G. Kevrekidis, and D. Psaltis, Phys. Rev. Lett. [**97**]{}, 033903 (2006). A. Hasegawa and M. Matsumoto, *Optical Solitons in Fibers* (Springer, Berlin, 2003). D.R. Solli, C. Ropers, P. Koonath, and B. Jalali *Nature* [**450**]{}, 1054 (2007). C.J. Pethick and H. Smith, *Bose-Einstein condensation in Dilute Gases* (Cambridge University Press, Cambridge, 2008). F. Dalfovo, S. Giorgini, L.P. Pitaevskii, and S. Stringari, Rev. Mod. Phys. [**71**]{} 463 (1999). R. Carretero-Gonzalez, D.J. Frantzeskakis, and P.G. Kevrekidis, Nonlinearity, [**21**]{} R139 (2008). L. Pitaevskii and S. Stringari, *Bose-Einstein Condensation* (Oxford University Press, New York, 2003). V.M. Perez-Garcia, H. Michinel, and H. Herrero, Phys. Rev. A [**57**]{} 3837 (1998). S. Rajendran, P. Muruganandam, and M. Lakshmanan, Physica D [**239**]{} 366 (2010). R. Atre, P.K. Panigrahi, and G.S. Agarwal, Phys. Rev. E [**73**]{} 056611 (2006). M. Lakshmanan and S. Rajasekar, *Nonlinear Dynamics: Integrablity, Chaos and Patterns* (Springer-Verlag, Heidelberg, 2003 ) D. Mandelik, H.S. Eisenberg, Y. Silberberg, R. Morandotti, and J.S. Aitchison, Phys. Rev. Lett. [**90**]{} 253902 (2003). M. Tajri and Y. Watanabe, Phys. Rev. E [**57**]{} 3 (1998). D.H. Pergrine, J. Austral. Math. Soc. Ser. B, [**25**]{} 16 (1983). A.R. Osborne, M. Onorato, and M. Serio, Phys. Lett. A [**275**]{} 386 (2000). A.R. Osborne, *Nonlinear Ocean waves* (Academic Press, New York, 2009). C. Kharif, E. Pelinovsky, and A. Slunyaev, *Rogue waves in the Ocean* (Springer-Verlag, Berlin, 2009). T.B. Benjamin and J.E. Feir, J. Fluid Mech. [**27**]{} 417 (1967). N. Akhmediev, A. Ankiewicz, and J.M. Soto-Crespo, Phys. Rev. E [**80**]{} 026601 (2009). A. Chabchoub, N.P. Hoffman, and N. Akhmediev, Phys. Rev. Lett. [**106**]{} 204502 (2011). M. Shatz, H. Punzmann, and H. Xia, Phys. Rev. Lett. [**104**]{} 104503 (2010). B. Kibler, J. Fatome, C. Finot, F. Dias, G. Genty. N. Akhmediev, and J.M. Dudley, Nat. Phys. [**6**]{} 790 (2010). J.S. He, S.W. Hu, and K. Porsezian, Phys. Rev. E [**86**]{} 066603 (2012). J.S. He, H.R. Zhang, L.H. Wang, K. Posezian, and A.S. Fokas, Phys. Rev. E [**87**]{} 052914 (2013). W.M. Moselem, Phys. Plasmas [**18**]{}, 032301 (2011). G.P. Veldas, J. Borhanian, M. Mckerr, V. Saxena, D.J. Frantzeskakis, and I. Kourakis, J. Optics. [**15**]{}, 064003 (2013). Z.Y. Yan, Commun. Theor. Phys. [**54**]{}, 947 (2010). K. Staliunas, S. Longhi, and G.J. deValcarcel, Phys. Rev. Lett. [**89**]{}, 210406 (2002). F.Kh. Abdullaev, J.G. Caputo, R.A. Kraenkel, and B.A. Malomed, Phys. Rev. A [**67**]{} 013605 (2003). A.D. Cronin, J. Schmiedmayer, and D.E. Pritchard, Rev. Mod. Phys. [**81**]{}, 3 (2009). O. Mandel, M. Greiner, A. Widera, T. Rom, T. H. Hänsea, and I. Bloch, Phys. Rev. Lett. [**91**]{} 1 (2003). J.I. Cirac and P. Zadar, Phys. Rev. Lett. [**74**]{} 4091 (1995). Y.V. Bludov, V.V. Konotop, and N. Akhmediev, Phys. Rev. A [**80**]{}, 033610 (2009). Z.Y. Yan, Phys. Lett. A [**374**]{} 672 (2010). Lin Wen, Lu Li, Zai-Dang Li, Shu-Wei Song, Xiao-Fei Zhang, and W.M. Liu, Eur. Phys. J. D [**64**]{} 473 (2011). L.C. Zhao, Annals of Physics [**329**]{} 73 (2013) X.F. Wu, G.S. Hua, and Z.Y. Ma, Commun. Nonlinear Sci. Num. Simulat. [**18**]{} 3325 (2013). J.S. He, E.G. Charalampidis, P.G. Kevrekidis, and D.J. Frantzeskakis, Phys. Lett. A [**378**]{} 577 (2014). S. Rajendran, P. Muruganandam, and M. Lakshmanan, J. Phys. B: At. Mol. Opt. Phys, [**42**]{} 145307 (2009). C. Özemír and F. Güngor, Rev. Math. Phys. [**24**]{} 1250015 (2012). P. Muruganandam and S.K. Adhikari, Comput. Phys. Commun. [**180**]{}, 1888 (2009); D. Vudragović, I. Vidanović, A. Balaz, P. Muruganandam, and S. K. Adhikari, *ibid.* 183, 2021 (2012). A. Ankiewicz, D.J. Kedziora, and N. Akhmediev, Phy. Lett. A [**375**]{} 2782 (2011). J.F. Wolff, Lake Superior Shipwrecks, Lake Superior Marine Museum Association, Inc., Duluth, MN, USA, 1979, ISBN 0932212-18-8, p. 28. N. Akhmediev, V.M. Eleonskii, and N.E. Kulagin, Zh. Eksperimentalnoii i Teoreticheskoii Fiziki [**89**]{}, 1542 (1985) \[Sov. Phys. JETP [**62**]{}, 894 (1985)\]. D.J. Kedziora, A. Ankiewicz, and N. Akhmediev, Phys. Rev. E [**85**]{}, 066601 (2012).

ArXiv

--- abstract: 'We study the effect of one-loop vacuum polarization on photon propagation in Siklos spacetimes in the geometric optics limit. We show that for photons with a general polarization in the transverse plane, the quantum correction vanishes in spacetimes with $H_{xy}=0$. For photons polarized along a transverse axis, subluminal and superluminal solutions are admitted for certain subclasses of Siklos spacetimes. We investigate the results in the Kaigorodov and Defrise spacetimes and obtain explicit expressions for the phase velocities. In Kaigorodov spacetime with $H\sim x^3$, photons polarized along $x$-axis are subluminal in regions where $H$ is positive, and superluminal in regions where $H$ is negative, while photons polarized along $y$-axis are superluminal in $H>0$ regions and subluminal in $H<0$ regions. In Defrise spacetime, $H\sim x^{-2}$, $x$-polarized and $y$-polarized photons are superluminal for $H<0$ and subluminal for $H>0$. We comment on motion in other Siklos spacetimes.' author: - Morteza Mohseni title: Vacuum polarization in Siklos spacetimes --- Introduction ============ The interplay between gravitational and quantum effects has been a topic of huge interest in recent decades, but in spite of massive efforts, there is no theory fully reconciling these effects yet. Still, one can study quantum effects in gravitational systems in certain contexts. Within this framework, [@1980PhRvD..22..343D] have investigated the QED contribution to the photon effective action from one-loop vacuum polarization in curved backgrounds. The calculation showed that the photon propagation get altered due to the quantum corrections and superluminal or subluminal motion could be possible, at least in principle. They also applied the formulation to gravitational wave spacetime (in the weak-field approximation) and also to the Schwarzschild spacetime and the superluminal-subluminal motion and also birefringence were shown explicitly. Their theory has then been examined in several spacetimes; namely in Reissner-Nordstrum spacetime in [@1994NuPhB.425..634D], in Kerr spacetime in [@1996PhLB..367...75D], in dilaton black hole spacetimes in [@1997PhRvD..56.6416C], in the static and rotating topological back hole backgrounds in [@1998NuPhB.524..639C], and more recently in [@2015EPJC...75..247B], in which by considering the Kerr- de Sitter and static de Sitter cosmic string spacetimes, the effect of a positive cosmological canstant was studied. Other aspects of this theory have been discussed in the literature, including generalization to high frequency limit and discussion of various kinds of velocity in [@2002NuPhB.633..271S], the issue of causality violation in [@1996NuPhB.460..379S; @2016JHEP...03..129H], and the problem of superluminality in [@2017JHEP...02..134G]. A review of the subject may be found in [@2003ConPh..44..503S]. Also, it has been shown in [@1988PhRvD..37.2743T] that including the quantum terms (but with arbitrary coefficients) in the electromagnetic Lagrangian breaks the conformal invariance of the action and this could be responsible in producing sizable magnetic fields during inflation. In the present work, we investigate photon propagation in Siklos spacetimes. These spacetimes, first introduced in [@1985gasr.book..247S], may be considered as exact gravitational waves propagating in anti-de Sitter universe [@1986PhLB..171..390G; @1998CQGra..15..719P]. A particular Siklos spacetime, the Kaigorodov spacetime [@1963SPhD....7..893K], has been of interest particularly in the context of AdS/CFT correspondence [@1999NuPhB.545..309C; @2001NuPhB.607..155B]. Some other aspects of the Kaigorodov spacetime have been investigated in the literature, see; e.g. [@2003CQGra..20.2087P] and the references therein. Although the current observations favour a positive cosmological constant, models with negative cosmological constant are still of interest and appears in different contexts including BTZ black holes [@1992PhRvL..69.1849B], and string theory and supersymmetry [@2014IJMPA..2930010B]. Plane gravitational and electromagnetic fields in spaces with cosmological constant have also been studied in [@1985JMP....26.1755O]. The paper is organized in the following order. We begin with a brief review of the electromagnetic field equations in a general curved spacetime in the limit of geometric optics and also collect the relevant equations for the one-loop vacuum polarization. Then, in section \[sec2\], we solve the equations of vacuum polarization for photons propagating in a general Siklos spacetime. We obtain expressions for $k^2$ and show that solutions with positive, zero, or negative values are admitted. We also compute the phase velocities. In section \[sec3\] we consider the particular case of propagation in a background Kaigorodov spacetime and show that depending on the photon polarization and the spacetime region on which photons are moving, both subluminal or superluminal photon propagation are possible. In section \[sec4\], we take the contribution of photons to the background spacetime into account, which is achieved by considering the Defrise spacetime. In section \[sec5\], we consider propagation in other Siklos spacetimes. We conclude the paper with a summary of the results. We use the natural units with the Lorentz-Heaviside units for electromagnetic fields, the metric signature $(-+++)$, and the following convention for the Riemann tensor ${R^\mu}_{\nu\rho\sigma}=\partial_\rho\Gamma^\mu_{\nu\sigma}+\Gamma^\mu_{\rho\kappa}\Gamma^\kappa_{\nu\sigma}- \{\rho\leftrightarrow\sigma\}$. The vacuum polarization ======================= The electromagnetic action $$\label{el1} S_0=\frac{-1}{4}\int d^4x\sqrt{-g}F_{\mu\nu}F^{\mu\nu}$$ gives the Maxwell equations for free fields $$\label{el2} D_\mu F^{\mu\nu}=0,$$ or, in terms of the field $A^\mu$ (subject to the Lorentz condition $D_\mu A^{\mu}=0$), $$\label{el3} D_\nu D^\nu A^\mu=R^\mu_\nu A^\nu,$$ $R_{\mu\nu}$ being the Ricci tensor. In the limit of geometric optics, the solution to eq. (\[el3\]) is given by $$\label{e14c} A^\mu=(a^\mu+i\varepsilon b^\mu+\cdots)\exp(\frac{i\varphi}{\varepsilon}).$$ Here, it is assumed that the wavelength $\lambda$ is small compared with the length scale of the spacetime curvature, $L_0$. In fact, one may take $\varepsilon=O(\frac{\lambda}{L_0})$. See e.g., [@padmanabhan2010gravitation] for more details and a pedagogical review of the geometric optics limit in curved spacetime. To the leading order, Eqs. (\[el3\]) and (\[e14c\]) result in $k_\mu k^\mu=0$ where $k_\mu=\partial_\mu\varphi$. Thus, the integral curve of $k^\mu$ is null. Also, the Lorentz condition implies that the polarizations are transverse $$\label{e14f} k_\mu a^\mu=0.$$ The next-to-leading order terms give $$\label{el4a} k^\nu D_\nu a^\mu=-\frac{1}{2}a^\mu D_\nu k^\nu$$ and the Lorentz condition leads to $$\label{el4b} D_\mu a^\mu=k_\mu b^\mu.$$ The one-loop corrected action is given by $S=S_0+S_1$, where $$\begin{aligned} \label{el5} S_1&=&\frac{1}{m^2_e}\int d^4x\sqrt{-g}(aRF_{\mu\nu}F^{\mu\nu}+bR_{\mu\nu}{F^\mu}_\kappa F^{\nu\kappa} \nonumber\\&&+cR_{\mu\nu\kappa\lambda} F^{\mu\nu}F^{\kappa\lambda}+dD_\mu F^{\mu\nu}D_\kappa {F^\kappa}_\nu) \end{aligned}$$ in which $a=\frac{\alpha}{144\pi}, b=-\frac{13\alpha}{360\pi}, c=\frac{\alpha}{360\pi}, d=\frac{e^2}{120\pi^2}$, $\alpha$ being the fine structure constant, and $m_e$ is the electron mass. It should be noted that the above action is in fact a truncated form the QED effective field theory action. The full action contains other terms including some curvature-independent and some UV divergent ones. Omitting such terms from the action, restricts the range of energies over which it is valid. A discussion of this may be found in [@2017JHEP...02..134G]. Including $S_1$ (with the last term being neglected) into the action results in the following field equation [^1] $$\label{e16} D_\mu F^{\mu\nu}=\frac{2}{m^2_e}D_\mu Q^{\mu\nu}$$ where $Q^{\mu\nu}=2aRF^{\mu\nu}+b(R^\mu_\rho F^{\rho\nu}-R^\nu_\rho F^{\rho\mu})+2cR^{\mu\nu\kappa\lambda}F_{\kappa\lambda}$. In the special case of a maximally symmetric spacetime, where $R_{\mu\nu\kappa\lambda}=P(g_{\mu\kappa}g_{\nu\lambda}- g_{\mu\lambda}g_{\nu\kappa})$, this reduces to $$\label{s1} \left(1+\frac{7P\alpha}{90\pi m^2_e}\right)D_\mu F^{\mu\nu}=0$$ and the vacuum polarization does not affect the propagation. For Ricci-flat spacetimes, we have from eq. (\[e16\]) $$\label{s2} D_\mu F^{\mu\nu}-\frac{\alpha}{90\pi m^2_e}{R^{\mu\nu}}_{\kappa\lambda}D_\mu F^{\kappa\lambda}=0.$$ Eq. (\[e16\]) can be expanded into the following form $$\begin{aligned} \label{e14d} &&\left(1-\frac{4aR}{m^2_e}\right)D_\mu F^{\mu\nu}-\frac{2b}{m^2_e}(R^\mu_\sigma D_\mu F^{\sigma\nu}-R^\nu_\sigma D_\mu F^{\sigma\mu})\nonumber\\&& -\frac{4c}{m^2_e}{R^{\mu\nu}}_{\sigma\tau}D_\mu F^{\sigma\tau}+\frac{2b+8c}{m^2_e}F^{\sigma\mu}D_\mu R^\nu_\sigma\nonumber\\&& -\frac{4a+b}{m^2_e}F^{\mu\nu}D_\mu R=0,\end{aligned}$$ which by using Eqs. (\[e14c\]) and (\[e14f\]), to $O(\frac{1}{\varepsilon^2})$ gives $$\begin{aligned} \label{e16d} &&\left(-\left(1-\frac{4aR}{m^2_e}\right)k_\mu k^\mu+\frac{2b}{m^2_e}R^\mu_\sigma k_\mu k^\sigma\right)a^\nu-\frac{2b}{m^2_e}R^\mu_\sigma k_\mu a^\sigma k^\nu \nonumber\\&&+\frac{2b}{m^2_e}k_\mu k^\mu a^\sigma R^\nu_\sigma-\frac{8c}{m^2_e}{R^{\mu\nu}}_{\sigma\tau}k_\mu k^\tau a^\sigma=0.\end{aligned}$$ The general equation, in which higher order terms in the expansion are also included, involve derivative of $a^\mu$ and $k^\mu$. To the leading order which Eq. (\[e16d\]) is based on, such terms are absent. The Siklos spacetimes {#sec2} ===================== In the chart $(u,x,y,v)$ with $u, v$ being the light-cone coordinates, the Siklos spacetimes metric is described by $$\label{e14} ds^2=\frac{-3}{\Lambda x^2}(H(u,x,y)du^2-dudv+dx^2+dy^2)$$ in which $\Lambda<0$ is the cosmological constant. Inserting this into the Einstein equation with cosmological constant, $G_{\mu\nu}+\Lambda g_{\mu\nu}=0$, results in $$\label{e13} H_{xx}+H_{yy}=\frac{2}{x}H_{x}$$ where subscripts represent differentiation. For photons propagating along the $z$ axis with $k^\mu=(A,0,0,B)$, we take $a^\mu=(0,C,F,0)$ in which $C,F$ are constants. This is consistent with the condition $$\label{o1} k_\mu a^\mu=0.$$ Inserting the above data into Eq. (\[e16d\]) results in (for $\mu=x,y$, respectively) the following two equations $$\begin{aligned} \frac{A}{m^2_e\Lambda x^2}[A\Lambda(CK+Fcx^2H_{xy})+C(AH-B)N]&=&0\nonumber\\ \label{eq1}\\ \frac{A}{m^2_e\Lambda x^2}[A\Lambda(FL+Ccx^2H_{xy})+F(AH-B)N]&=&0\nonumber\\ \label{eq3}\end{aligned}$$ where $$\begin{aligned} K&\equiv& 4c(x^2H_{xx}-xH_x)+E,\\ L&\equiv& 4c(x^2H_{yy}-xH_x)+E,\\ N&\equiv& (48a+12b+8c)\Lambda-3m^2_e,\\ E&\equiv& bx^2\left(H_{xx}+H_{yy}-\frac{2}{x}H_x\right)\end{aligned}$$ For $H$ satisfying Eq. (\[e13\]), we have $E=0$. The above equations admit the trivial solution $A=0$ which corresponds to $k^2=0$. Now if $H_{xy}=0$, $C\neq 0$, and $F\neq 0$, Eqs. (\[eq1\]) and (\[eq3\]) are inconsistent unless we take $A=0$ (except for the particular case where $K=L$ for which $A$ can be nonzero. This particular condition is satisfied for Siklos spacetimes with $H_{xx}=H_{yy}$). However, if we further take either $F=0, C\neq 0$ or $C=0, F\neq 0$, then the above system of equations can be satisfied with $A\neq 0$. If we take $H_{xy}=0$, with $F=0, C\neq 0$, Eq. (\[eq3\]) is automatically satisfied and from Eq. (\[eq1\]) we get $$\begin{aligned} \label{kk2} \frac{3A^2 K}{Nx^2}=-\frac{3A}{\Lambda x^2}(AH-B)\end{aligned}$$ where $K=c(x^2H_{xx}-xH_x)+E$. From Eq. (\[kk2\]) one can easily read off the value of $k^2=\frac{-3A}{\Lambda x^2}(AH-B)$. Thus, $$\label{vp2} k^2=\frac{3A^2 K}{Nx^2}.$$ The phase velocity of photons can be obtained from $v_{p}=\frac{\omega}{|\vec{k}|}$. To compute this, we first perform the coordinates transformation $$\begin{aligned} du&=&\frac{1}{\sqrt{h(h+H)}}(hdt-dz),\label{tr1}\\ dv&=&\frac{1}{\sqrt{h(h+H)}}(dt+hdz),\label{tr2}\end{aligned}$$ in which $h\equiv {\sqrt{H^2+1}}-H$. This brings the metric to the following form $$\label{met1} ds^2=\frac{-3}{\Lambda x^2}(-hdt^2+\frac{1}{h}dz^2+dx^2+dy^2).$$ Thus, we obtain $$\begin{aligned} \label{p1} \omega&=&\sqrt{\frac{-3}{4\Lambda x^2(h+H)}}|hA+B|\nonumber\\ &=&\sqrt{\frac{-3}{4\Lambda x^2(h+H)}}|A|\left|h+H+\frac{\Lambda K}{N}\right|,\end{aligned}$$ and $$\begin{aligned} \label{p2} |\vec{k}|&=&\sqrt{\frac{-3}{4\Lambda x^2h^2(h+H)}}|A-Bh|\nonumber\\ &=&\sqrt{\frac{-3}{4\Lambda x^2(h+H)}}|A|\left|h+H-\frac{\Lambda K}{N}\right|.\end{aligned}$$ We therefore obtain $$\label{vp1} v_p=\left|\frac{N(h+H)+\Lambda K}{N(h+H)-\Lambda K}\right|.$$ Similarly, if we take $C=0, F\neq 0$, then Eq. (\[eq1\]) is automatically satisfied and from Eq. (\[eq3\]) we get the same expressions as above but with $K$ replaced by $L=c(x^2H_{yy}-xH_x)+E$. In the case where $H_{xy}\neq 0$, Eqs. (\[eq1\]) and (\[eq3\]) are satisfied with $A\neq 0$ by choosing particular $\frac{F}{C}$ ratios. Here the solutions are given by the roots of $$\label{rev1} \mbox{det} \begin{pmatrix} A\Lambda K+N(AH-B) & A\Lambda cx^2H_{xy} \\ A\Lambda cx^2H_{xy} & A\Lambda L+N(AH-B) \end{pmatrix} =0.$$ These are given by $B-AH=\frac{\Lambda A}{N}W_{\pm}$ where $W_{\pm}=\left(1+\frac{2c}{b}\right)E\pm cx^2\sqrt{4(H_{xx}-H_{yy})^2+H_{xy}^2}$ from which we obtain $$\label{hh1} k^2=\frac{3A^2}{Nx^2}W_{\pm}$$ corresponding to $\frac{F}{C}=\frac{W_{\pm}-K}{cx^2H_{xy}}$ respectively. Motion in Kaigorodov Spacetime {#sec3} ============================== A particular solution of Eq. (\[e13\]), which in addition satisfies $H_{xy}=0$, is given by $H=\sigma x^3$ in which $\sigma$ is a constant. With this choice, the metric (\[e14\]) reduces to $$\label{e12} ds^2=\frac{-3}{\Lambda x^2}(\sigma x^3du^2-dudv+dx^2+dy^2)$$ This describes the Kaigorodov space-time [@1963SPhD....7..893K] in Siklos horospherical-type coordinates (or more formally, Fefferman-Graham coordinates, see; e.g. [@2012PhRvD..85d6007T]). This metric can be obtained from the following one $$\label{hor1} ds^2=\pm e^{-lr}dX^2+e^{2lr}(-dXdT+dY^2)+dr^2$$ by imposing the coordinate transformation $x=\pm e^{-lr}, u=lX, v=lT, y=lY$ where $l=\sqrt{\frac{-\Lambda}{3}}$ [@2003CQGra..20.2087P]. The minus and plus signs correspond to $x>0$ and $x<0$ regions respectively, and $\sigma$ is regarded as unity for simplicity. The positive and negative $x$ regions are disjointed and $x=0$ represents the null infinity. Now, noting that for this metric we have $E=0$, Eq. (\[vp2\]) gives $$\label{kk1} k^2=-\frac{9\beta A^2\sigma}{7\Lambda(3+\beta)}x,$$ in which $\beta=\frac{7\alpha\Lambda}{90\pi m^2_e}$. Also, Eq. (\[vp1\]) results in $$\label{pv1} v_p=\left|\frac{(3\sigma x^3-7\sqrt{\sigma^2x^6+1})\beta-21\sqrt{\sigma^2x^6+1}}{(3\sigma x^3+7\sqrt{\sigma^2x^6+1})\beta+21\sqrt{\sigma^2x^6+1}}\right|$$ or, for $\beta\ll 1$, $$\label{pv11} v_p=1+\beta\frac{6\sigma x^3}{21\sqrt{\sigma^2x^6+1}}$$ Similarly, for $C=0, F\neq 0$ we obtain $$\label{kk1a} k^2=\frac{9\beta A^2\sigma}{7\Lambda(3+\beta)}x,$$ and $$\label{pv1a} v_p=\left|\frac{(3\sigma x^3+7\sqrt{\sigma^2 x^6+1})\beta+21\sqrt{\sigma^2x^6+1}}{(3\sigma x^3-7\sqrt{\sigma^2x^6+1})\beta-21\sqrt{\sigma^2x^6+1}}\right|$$ or $$\label{pv12} v_p=1-\beta\frac{6\sigma x^3}{21\sqrt{\sigma^2x^6+1}}$$ respectively. In both cases we have $v_p\rightarrow 1$ as $x\rightarrow 0$. Now, noting that $m_e=5.1\times 10^5 eV$, and $|\Lambda|=4.6\times 10^{-66} eV^2$ (corresponding to the experimental value $+1.19\times 10^{-52} m^{-2}$), we have $\beta=-3.2\times 10^{-81}$. Thus, for the case $F=0$, Eq. (\[kk1\]) gives $sign(k^2)=sign(-\sigma x)$ which, if we assume $\sigma>0$, corresponds to subluminal photons in $x>0$ region and superluminal photons in $x<0$ region. Similarly, for the case $C=0$, Eq. (\[kk1a\]) gives $sign(k^2)=sign(\sigma x)$ which shows superluminal photons in $x>0$ region and subluminal photons in $x<0$ region. These are also confirmed explicitly by Eqs. (\[pv11\]) and (\[pv12\]). One can obtain the reverse situation by choosing $\sigma<0$. For the spacetime described by metric (\[e12\]), the curvature length-scale is of order of $L_0\sim (-\Lambda)^{-1/2}\sim 10^{33} eV^{-1}$ which is very large compared to the Compton wavelength $\lambda_c\sim 10^{-5} eV^{-1}$. Thus, the requirement $L_0\gg\lambda_c$ ([@1980PhRvD..22..343D]) for the validity of the one-loop calculations is well satisfied. On the other hand, we should also have $-k^2<4m^2_e$ (see e.g., [@weinberg1995quantum]). Thus, from Eqs. (\[kk1\]) or (\[kk1a\]) we obtain $A<\sqrt{\left|\frac{28\Lambda}{3\beta\sigma x}\right|}m_e$ or, equivalently, $A<\sqrt{\frac{120\pi}{\alpha|\sigma x|}}m^2_e$. This puts an upper bound on the value of $A$. Motion in Defrise spacetime {#sec4} =========================== In the previous section, we investigated the photon propagation in a particular Siklos spacetime, the Kaigorodov spacetime, in which the source of spacetime curvature is only the cosmological constant. This implies that the contribution of photons to the energy-momentum tensor is neglected. There are other subclasses of Siklos solutions in which such contributions can be accounted for. In particular, the Defrise spacetime [@defrise1969] is obtained when in addition to the cosmological constant, there is a pure radiation field with the the energy-momentum tensor $T_{\mu\nu}=\rho k_\mu k_\nu$, with $\rho$ being a constant [@2001GReGr..33.1093P]. The Defrise metric is described by Eq. (\[e14\]) by setting $H=-\delta x^{-2}$ in which $\delta$ is constant. It can be obtained by taking $\rho=\frac{5\Lambda^2\delta}{18\pi G}$ and $k^{\mu}=(0,0,0,1)$, corresponding to massless photons propagating in the $z$ direction. In this spacetime we have $E\neq 0$, and with the choice $F=0, C\neq 0$, we have from Eq. (\[vp2\]) $$\label{j1} k^2=\frac{54\beta A^2\delta}{7(3+\beta)\Lambda x^4},$$ and Eq. (\[vp1\]) results in $$\label{j1a} v_p=\left|\frac{(18\delta+7\sqrt{x^4+\delta^2})\beta+21\sqrt{x^4+\delta^2}}{(18\delta-7\sqrt{x^4+\delta^2})\beta-21\sqrt{x^4+\delta^2}}\right|$$ or $$\label{j1aw} v_p=1-\beta\frac{12\delta}{7\sqrt{x^4+\delta^2}}$$ showing subluminal motion for $\delta<0$ and superluminal motion for $\delta>0$. Similarly, with the choice $C=0, F\neq 0$, we obtain $$\label{j2} k^2=\frac{36\beta A^2\delta}{7(3+\beta)\Lambda x^4},$$ and $$\label{j2a} v_p=\left|\frac{(12\delta+7\sqrt{x^4+\delta^2})\beta+21\sqrt{x^4+\delta^2}}{(12\delta-7\sqrt{x^4+\delta^2})\beta-21\sqrt{x^4+\delta^2}}\right|$$ or $$\label{j2aw} v_p=1-\beta\frac{8\delta}{7\sqrt{x^4+\delta^2}}$$ which shows again subluminal motion for $\delta<0$ and superluminal motion for $\delta>0$. In both cases we have $v_p\rightarrow 1$ as $x\rightarrow\infty$. Motion in other Siklos spacetimes {#sec5} ================================= It has been shown in [@1985gasr.book..247S] that Eq. (\[e13\]) admits a general solution of the following form $$\label{pt1} H(u,x,y)=x^2\frac{\partial}{\partial x}\left(\frac{f(\zeta,u)+\bar f(\bar\zeta,u)}{x}\right)$$ where $f$ is an arbitrary function and $\zeta=x+iy$. Thus, for example, choosing $f(\zeta,u)=\frac{1}{4}\zeta^3$ reproduces the Kaigorodov metric. It is possible to apply the formulation given in section \[sec2\] to various subsets of the above general wave profile. The procedure is straightforward, but the results depend on the explicit form of the wave profile. As an interesting example, for a generalized Kaigorodov metric with $H(u,x,y)=w(u)x^3$ in which $w(u)$ is arbitrary, Eq. (\[vp2\]) reduces to $$\label{vp2u} k^2=\pm\frac{9A^2\beta}{7\Lambda(3+\beta)}xw(u)$$ and the behavior depends on $w(u)$. On the other hand, there are other solutions to Eq. (\[e13\]), such as $H(u,x,y)=w(u)x^3+s(u)y$ in which $s(u)$ is arbitrary, which give the same results as Eq. (\[vp2u\]). Another profile of this kind is $H(u,x,y)=q(u)(x^2+y^2)$ where $q(u)$ is also arbitrary. Interestingly, the later (up to a conformal transformation) is also the wave profile of an exact gravitational wave produced by a light wave in otherwise empty spacetime [@doi:10.1063/1.522612], (see also [@2011ForPh..59..284V]). It is also possible to generalize the Defrise metric by $H(u,x,u)=j(u)x^{-2}$, $j(u)$ being arbitrary, which would results in different behavior compared to the ones discussed in section \[sec4\]. An example of Siklos spacetimes with $H_{xy}\neq 0$ is $H(u,x,y)=x^3+\frac{1}{3}y^3+x^2y$. For this spacetime, Eq. (\[hh1\]) gives $$\label{hh1a} k^2=\frac{\mp 3\sqrt{37}\beta A^2}{14\Lambda(\beta+3)}x$$ representing both superluminal or sublumninal propagation. Conclusions =========== We studied the effect of one-loop correction of photon vacuum polarization on photon propagation in Siklos spacetimes in the geometric optics limit. In Siklos spacetimes with $H_{xy}=0$, for photons with nonzero polarization in both $x, y$ directions, the quantum correction vanishes. In Kaigorodov spacetime, $H\sim x^3$, we showed that in addition to usual massless photons, there exists a solution for which photons polarized along $x$ axis are superluminal in the $H<0$ regions and subluminal in the $H>0$ regions, while photons polarized along the $y$ axis are subluminal in the $H<0$ regions and superluminal in the $H>0$ regions. Thus, phenomenon of birefringence is shown to exhibit in Kaigorodov, and some other subclasses of Siklos spacetimes. The deviation from the standard speed of light are tiny, of the order of $\frac{\alpha\Lambda}{m^2_e}$. In Defrise spacetime, $H\sim x^{-2}$, photons polarized either along the $x$-axis or the $y$-axis are superluminal in $H<0$ regions and subluminal in $H>0$ regions. For the class of Siklos spacetimes with off-diagonal terms in the wave profile, $H_{xy}\neq0$, spuperluminal/sublumibal propagation is possible with arbitrary polarization in the transverse plane. Acknowledgements {#acknowledgements .unnumbered} ================ I would like to thank an anonymous referee of PRD for several comments. [30]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase 10.1103/PhysRevD.22.343) [****,  ()](\doibase 10.1016/0550-3213(94)90291-7),  [****,  ()](\doibase 10.1016/0370-2693(95)01468-3),  [****,  ()](\doibase 10.1103/PhysRevD.56.6416),  [****,  ()](\doibase 10.1016/S0550-3213(98)00274-0),  [****,  ()](\doibase 10.1140/epjc/s10052-015-3477-7),  [****,  ()](\doibase 10.1016/S0550-3213(02)00240-7),  [****,  ()](\doibase 10.1016/0550-3213(95)00646-X),  [****,  ()](\doibase 10.1007/JHEP03(2016)129),  [****,  ()](\doibase 10.1007/JHEP02(2017)134), [****,  ()](\doibase 10.1080/00107510310001617106),  [****, ()](\doibase 10.1103/PhysRevD.37.2743) “,” in @noop [**]{},  (, , ) pp.  [****,  ()](\doibase 10.1016/0370-2693(86)91426-7) [****,  ()](\doibase 10.1088/0264-9381/15/3/019), @noop [****,  ()]{} [****,  ()](\doibase 10.1016/S0550-3213(99)00002-4),  [****,  ()](\doibase 10.1016/S0550-3213(01)00170-5),  [****,  ()](\doibase 10.1088/0264-9381/20/11/310),  [****,  ()](\doibase 10.1103/PhysRevLett.69.1849),  [****,  ()](\doibase 10.1142/S0217751X14300105),  [****,  ()](\doibase 10.1063/1.526887) @noop [**]{} (, ) pp.  [****, ()](\doibase 10.1103/PhysRevD.85.046007),  @noop [**]{}, Vol.  (, ) pp.  **, @noop ,  () [****,  ()](\doibase 10.1023/A:1010284400184),  [****, ()](\doibase 10.1063/1.522612),  [****,  ()](\doibase 10.1002/prop.201000088),  [^1]: As elaborated in [@2017JHEP...02..134G], there is some confusion in the literature regarding the conventions and signs. This causes some references; e.g. to use (mistakenly) an extra minus sign before $\frac{2}{m^2_e}$ in Eq. (\[e16\]).

ArXiv

--- abstract: 'We consider the Cauchy problem for systems of nonlinear wave equations with multiple propagation speeds in three space dimensions. Under the null condition for such systems, the global existence of small amplitude solutions is known. In this paper, we will show that the global solution is asymptotically free in the energy sense, by obtaining the asymptotic pointwise behavior of the derivatives of the solution. Nonetheless we can also show that the pointwise behavior of the solution itself may be quite different from that of the free solution. In connection with the above results, a theorem is also developed to characterize asymptotically free solutions for wave equations in arbitrary space dimensions.' address: | Department of Mathematics, Wakayama University, 930 Sakaedani, Wakayama 640-8510, Japan\ Tel: +81-73-457-7343 author: - Soichiro Katayama title: Asymptotic behavior for systems of nonlinear wave equations with multiple propagation speeds in three space dimensions --- Introduction ============ We consider the Cauchy problem for a system of nonlinear wave equations of the following type with small initial data: $$\begin{aligned} \label{OurSystem} & {\square}_{c_j} u_j(t,x)=F_j\bigl({\partial}u (t,x), {\partial}^2 u(t,x)\bigr), \quad (t,x)\in(0,\infty)\times {{\mathbb R}}^3,\\ \label{OurData} & u_j(0,x)={\varepsilon}f_j(x), \ ({\partial}_t u_j)(0,x)={\varepsilon}g_j(x), \quad x\in {{\mathbb R}}^3\end{aligned}$$ for $j=1, \ldots, N$, where ${\square}_c={\partial}_t^2-c^2\Delta_x={\partial}_t^2-c^2\sum_{k=1}^3{\partial}_{x_k}^2$ for $c>0$, and the propagation speeds $c_1,\ldots, c_N$ are positive constants, while ${\partial}u$ and ${\partial}^2 u$ denote the first and second derivatives of $u=(u_l)_{1\le l\le N}$, respectively. More specifically, we write $${\partial}u=({\partial}_a u_l)_{1\le l\le N,\, 0\le a\le 3},\ {\partial}^2 u=({\partial}_a{\partial}_b u_l)_{ 1\le l\le N,\, 0\le a,b\le 3}$$ with the notation $${\partial}_0:={\partial}_t=\frac{{\partial}}{{\partial}t}, \text{ and }{\partial}_k:={\partial}_{x_k}=\frac{{\partial}}{{\partial}x_k} \text{ for $k=1,2,3$.}$$ We assume that $f=(f_j)_{1\le j\le N}, g=(g_j)_{1\le j\le N}\in C^\infty_0({{\mathbb R}}^3;{{\mathbb R}}^N)$. ${\varepsilon}$ in is a small and positive parameter. $F_j$ can include nonlinear terms of higher order, but for simplicity we assume that $F_j$ can be written as $$\label{Non01} F_j=\sum_{k, l=1}^N \left(\sum_{a,b,b'=0}^3 p_{jkl}^{abb'} ({\partial}_a u_k)({\partial}_b{\partial}_{b'} u_l)+\sum_{a,b=0}^3 q_{jkl}^{ab}({\partial}_a u_k)({\partial}_b u_l)\right)$$ for $1\le j\le N$ with appropriate real constants $p_{jkl}^{abb'}$ and $q_{jkl}^{ab}$, where $p_{jkl}^{a00}=0$ for $1\le j,k, l\le N$ and $0\le a\le 3$. To assure the hyperbolicity, we assume the symmetry condition $$\label{symmetry} p_{jkl}^{abb'}=p_{lkj}^{abb'}=p_{jkl}^{ab'b}$$ for $1\le j, k, l \le N$, and $0\le a,b,b' \le 3$. For general quadratic nonlinearity, the classical solution to – may blow up in finite time for some $(f,g)$ no matter how small ${\varepsilon}$ is. Klainerman [@Kla86] introduced a sufficient condition for small data global existence for the single speed case where $c_1=\cdots=c_N(=1)$ (see also Christodoulou [@Chr86]). This sufficient condition, called the [*null condition*]{}, was extended to the multiple speed case: To simplify the description, we assume that the speeds are distinct, and that $$\label{distinct speeds} 0<c_1<c_2<\cdots<c_N.$$ Let the constants $p_{jkl}^{abb'}$ and $q_{jkl}^{ab}$ be from . We say that the null condition (associated with the speeds $c_1,\ldots, c_N$) is satisfied if we have $$\label{NullC} \sum_{a,b,b'=0}^3 p_{jjj}^{abb'} X_aX_bX_{b'}=\sum_{a,b=0}^3 q_{jjj}^{ab}X_aX_b =0,\quad X\in{\mathcal N}_j,\ 1\le j\le N,$$ where ${\mathcal N}_j:=\{X=(X_a)_{0\le a\le 3}\in {{\mathbb R}}^4;\, X_0^2-c_j^2\sum_{k=1}^3X_k^2=0\}$. Small data global existence under the null condition for the multiple speed case was obtained by Yokoyama [@Yok00] (see also Sideris-Tu [@SidTu01], Sogge [@Sog03], and Hidano [@Hid04]; see Kubota-Yokoyama [@Kub-Yok01], the author [@Kat04:02], and Metcalfe-Nakamura-Sogge [@MetNaSo05b] for the case where nonlinearity of higher order depends not only on $({\partial}u, {\partial}^2 u)$, but also on $u$). We introduce the [*null forms*]{} $$\begin{aligned} \label{NullForm01} Q_0(\varphi,\psi;c)=&({\partial}_t\varphi)({\partial}_t\psi)-c^2\sum_{k=1}^3 ({\partial}_k\varphi)({\partial}_k\psi),\\ \label{NullForm02} Q_{ab}(\varphi,\psi)=&({\partial}_a\varphi)({\partial}_b \psi)-({\partial}_b\varphi)({\partial}_a\psi),\quad 0\le a, b\le 3.\end{aligned}$$ Then it is shown in [@Yok00] that the quadratic nonlinearity satisfying the null condition can be written as $$\label{Non02} F_j({\partial}u,{\partial}^2 u)=N_j({\partial}u, {\partial}^2 u)+R_j^{\rm I}({\partial}u, {\partial}^2 u) {}+R_j^{\rm II}({\partial}u, {\partial}^2 u),$$ where $$\begin{aligned} N_j=& \sum_{0\le |\alpha|\le 1} A_j^{\alpha} Q_0(u_j,{\partial}^\alpha u_j; c_j) {}+\sum_{a,b,b'=0}^3 B_j^{abb'}Q_{ab}(u_j, {\partial}_{b'}u_j), \label{DefNullT}\\ R_j^{\rm I}=& \sum_{\{(k,l); k\ne l\}}\left(\sum_{a,b,b'=0}^3 p_{jkl}^{abb'} ({\partial}_a u_k)({\partial}_b{\partial}_{b'} u_l)+\sum_{a,b=0}^3 q_{jkl}^{ab}({\partial}_a u_k)({\partial}_b u_l)\right), \label{DefNonResI}\\ R_j^{\rm II}=& \sum_{\{k; k\ne j\}}\left(\sum_{a,b,b'=0}^3 p_{jkk}^{abb'} ({\partial}_a u_k)({\partial}_b{\partial}_{b'} u_k)+\sum_{a,b=0}^3 q_{jkk}^{ab}({\partial}_a u_k)({\partial}_b u_k)\right) \label{DefNonResII}\end{aligned}$$ with some constants $A_j^{\alpha}$ and $B_j^{abb'}$ ($p_{jkl}^{abb'}$ and $q_{jkl}^{ab}$ are from ). Here ${\partial}=({\partial}_a)_{0\le a\le 3}$, and $\alpha=(\alpha_a)_{0\le a\le 3}$ is a multi-index. We refer to terms involved in $R^{\rm I}=(R_j^{\rm I})_{1\le j\le N}$ and $R^{\rm II}=(R_j^{\rm II})_{1\le j\le N}$ as the [*non-resonant terms of type I*]{} and [*type II*]{}, respectively. Now we turn our attention to the asymptotic behavior of the global solutions. Let $c>0$. It is known that if $G\in L^1\bigl((0,\infty);L^2({{\mathbb R}}^3)\bigr)$, then the solution $v$ to ${\square}_c v(t,x)=G(t,x)$ is [*asymptotically free in the energy sense*]{}, that is to say, there exists a solution $v^+$ to the free wave equation ${\square}_c v^+=0$ such that $$\lim_{t\to \infty} \|(v-v^+)(t)\|_{E,c}=0,$$ where the energy norm $\|\varphi(t)\|_{E,c}$ is defined by $$\|\varphi(t)\|_{E,c}^2=\frac{1}{2}\int_{{{\mathbb R}}^3} \left(\frac{1}{c^2}|{\partial}_t\varphi(t,x)|^2+|\nabla_x \varphi(t,x)|^2\right)dx$$ with $\nabla_x=({\partial}_1,{\partial}_2,{\partial}_3)$. For the single speed case, investigating the proof in [@Kla86], we have $$F_j({\partial}u, {\partial}^2 u)\in L^1\bigl((0, \infty); L^2({{\mathbb R}}^3)\bigr)$$ under the null condition because of the extra decay for the null forms, and as an immediate consequence we see that the global solution $u$ is asymptotically free in the energy sense. As for the multiple speed case, by the estimates obtained in [@Kub-Yok01] (cf.  and below), it is easy to see that $$\label{AsympCond} N_j({\partial}u,{\partial}^2 u)+R_j^{\rm I}({\partial}u,{\partial}^2 u)\in L^1\bigl((0,\infty);L^2({{\mathbb R}}^3)\bigr)$$ for the global solution $u$ to –, because we can expect some gain in the decay rate for the null forms and the non-resonant terms of type I, compared to general quadratic nonlinearity (see and below). Therefore, if $R^{\rm II}\equiv 0$ in , then implies that the global solution $u$ is asymptotically free in the energy sense; namely for $j=1,\ldots, N$, there exists a solution $u_j^+$ of the free wave equation ${\square}_{c_j} u_j^+=0$ such that $\lim_{t\to\infty}\|(u_j-u_j^+)(t)\|_{E,c_j}=0$. By contrast, there is no explicit gain in the decay rate for the non-resonant terms of type II unless they can be written in terms of the null forms, and we cannot expect $R_j^{\rm II}({\partial}u,{\partial}^2 u)\in L^1\bigl((0,\infty);L^2({{\mathbb R}}^3)\bigr)$ in general (see below), although its influence is weak enough for the solution to exist globally. Hence it is not clear whether the global solution $u$ is asymptotically free or not when $R^{\rm II}$ is not written in terms of the null forms. Our aim in this paper is to determine the asymptotic behavior in the presence of the non-resonant terms of type II by modifying the method developed in [@Kat11]. Main Results {#MainResults} ============ Asymptotically free functions in the energy sense. -------------------------------------------------- To begin with, we will characterize the asymptotically free functions in the energy sense. We only need the three space dimensional result for an application in this paper, but we consider the general space dimensional case here for future applications. Let $n$ be a positive integer. Let $\dot{H}^1({{\mathbb R}}^n)$ be the completion of $C^\infty_0({{\mathbb R}}^n)$ with respect to the norm $\|\varphi\|_{\dot{H}^1({{\mathbb R}}^n)}=\|\nabla_x\varphi\|_{L^2({{\mathbb R}}^n)}$. We put $$X_n:=C\bigl([0,\infty); \dot{H}^1({{\mathbb R}}^n)\bigr) \cap C^1\bigl([0,\infty); L^2({{\mathbb R}}^n)\bigr). \label{FunctionSpace}$$ We say that a function $v=v(t,x)\in X_n$ is [*asymptotically free in the energy sense associated with the speed $c$*]{}, if there is $(v_0^+, v_1^+)\in \dot{H}^1({{\mathbb R}}^n)\times L^2({{\mathbb R}}^n)$ such that $$\lim_{t\to\infty} \|v(t,\cdot)-v^+(t,\cdot)\|_{E,c}=0,$$ where $v^+\in X_n$ is a unique solution to $${\square}_c v^+(t,x)=0,\quad (t,x)\in(0,\infty)\times {{\mathbb R}}^n$$ with initial data $\bigl(v^+(0),{\partial}_t v^+(0)\bigr)=(v_0^+, v_1^+)$, and the energy norm (associated with the speed $c$) is given by $$\|\varphi(t,\cdot)\|_{E,c}^2:=\frac{1}{2}\int_{{{\mathbb R}}^n} \left(\frac{1}{c^2}|{\partial}_t \varphi(t,x)|^2+|\nabla_x \varphi(t,x)|^2\right)dx.$$ Here ${\square}_c={\partial}_t^2-c^2\sum_{k=1}^n {\partial}_k^2$ with the notation ${\partial}_k={\partial}/{\partial}x_k$ for $1\le k\le n$. Note that we do not suppose that $v$ is a solution to some wave equation here. For $c>0$ and $x\in {{\mathbb R}}^n\setminus\{0\}$, we define $$\label{speedconst} \vec{\omega}_c(x)=\bigl({\omega}_{c,a}(x)\bigr)_{0\le a\le n}:=(-c, |x|^{-1}x).$$ \[AFSE\] Let $n\ge 2$ and $c>0$. A function $v\in X_n$ is asymptotically free in the energy sense associated with the speed $c$, if and only if there is a function $V=V(\sigma, \omega)\in L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})$ such that $$\lim_{t\to\infty} \bigl\|{\partial}v(t,\cdot)-\vec{\omega}_c(\cdot){V}^\sharp(t,\cdot) \bigr\|_{L^2({{\mathbb R}}^n)}=0,$$ where ${\partial}=({\partial}_0, {\partial}_1, \ldots, {\partial}_n)$, and $${V}^\sharp(t,x):=|x|^{-(n-1)/2}V(|x|-ct, |x|^{-1}x),\quad (t,x)\in (0,\infty)\times({{\mathbb R}}^n\setminus\{0\}).$$ This theorem for $n=3$ was implicitly proved in [@Kat11]. The author believes that this result for general space dimensions is a new observation, but it possibly has already appeared in some literature. We will give the proof in Section \[ProofAFSE\]. See [@KatMurSun11] for an application of Theorem \[AFSE\] to semilinear wave equations in two space dimensions. Another application in three space dimensions can be found in [@KatMatSun12]. Asymptotic behavior of the global solutions ------------------------------------------- Next we examine the asymptotic behavior of the global solutions to –. For $h\in C^\infty({{\mathbb R}}^3)$, we define its Radon transform ${\mathcal R}[h]$ by $${\mathcal R}[h](\sigma, \omega)=\int_{y\cdot\omega=\sigma} h(y) dS(y),$$ where $dS(y)$ denotes the surface element on the plane $\{y\cdot\omega=\sigma\}$. Now, restricting our attention to the three space dimensional case, we introduce the Friedlander radiation field (its definition in general space dimensions will be given in below): For $(\varphi, \psi)\in C^\infty_0({{\mathbb R}}^3)\times C^\infty_0({{\mathbb R}}^3)$, we define the [*Friedlander radiation field*]{} $$\label{Friedlander3D} {\mathcal F}_0[\varphi,\psi](\sigma, \omega)=\frac{1}{4\pi} \bigl({\mathcal R}[\psi](\sigma, \omega)-({\partial}_\sigma {\mathcal R}[\varphi])(\sigma,\omega)\bigr)$$ for $(\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^2$. For $z\in {{\mathbb R}}^d$ with a positive integer $d$, we put ${\left\langle z \right\rangle}:=\sqrt{1+|z|^2}$. \[Main01\] Suppose that , , , and the null condition are fulfilled. Let ${\partial}=({\partial}_0, {\partial}_1, {\partial}_2, {\partial}_3)$, and $\vec{\omega}_c(x)$ be given by .\ [(1)]{} We fix small $\delta>0$. Then for any $f,g\in C^\infty_0({{\mathbb R}}^3; {{\mathbb R}}^N)$ and sufficiently small ${\varepsilon}>0$, there is a function $P=(P_j)_{1\le j\le N}$ of $(\sigma,\omega)\in {{\mathbb R}}\times {{\mathbb S}}^2$ such that $$\begin{aligned} |x| {\partial}u_j(t, x)=& {\varepsilon}\vec{\omega}_{c_j}(x) P_j(|x|-c_jt, |x|^{-1}x) \nonumber\\ &{}+O\bigl({\varepsilon}{\left\langle t+|x| \right\rangle}^{-1+\delta}{\left\langle c_jt-|x| \right\rangle}^{-\delta}\bigr) \label{Main11}\end{aligned}$$ for $(t,x)\in [0,\infty)\times({{\mathbb R}}^3\setminus\{0\})$ and $1\le j\le N$, where $u=(u_j)_{1\le j\le N}$ is the global solution to the Cauchy problem –. Moreover we have $$\label{Main12} P_j(\sigma, \omega)={\partial}_\sigma{\mathcal F}_0[f_j,c_j^{-1}g_j](\sigma, \omega) +O\bigl({\varepsilon}{\left\langle \sigma \right\rangle}^{-1}\bigr),\quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^2.$$ [(2)]{} We further assume that $R^{\rm II}=(R_j^{\rm II})_{1\le j\le N}$ has the null structure, that is to say $$R_j^{\rm II}=\sum_{\{k;k\ne j\}} \left(\sum_{0\le |\alpha|\le 1} A_{jk}^{\alpha} Q_0(u_k, {\partial}^\alpha u_k; c_k) +\sum_{a, b, b'=0}^3 B_{jk}^{abb'}Q_{ab}(u_k, {\partial}_{b'}u_k)\right)$$ for $1\le j\le N$ with some constants $A_{jk}^\alpha$ and $B_{jk}^{abb'}$. Then for any $f,g\in C^\infty_0({{\mathbb R}}^3,{{\mathbb R}}^N)$ and sufficiently small ${\varepsilon}>0$, there is a function $U=(U_j)_{1\le j\le N}$ of $(\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^2$ such that $$\begin{aligned} \label{Main31} |x| u_j(t, x)= & {\varepsilon}U_j(|x|-c_jt, |x|^{-1}x)+O\bigl({\varepsilon}{\left\langle t+|x| \right\rangle}^{-1}\log(2+t)\bigr), \\ \label{Main32} |x| {\partial}u_j(t, x)= & {\varepsilon}\vec{\omega}_{c_j}(x) ({\partial}_\sigma U_j)(|x|-c_jt, |x|^{-1}x) \nonumber\\ & {}+O\bigl({\varepsilon}{\left\langle t+|x| \right\rangle}^{-1}{\left\langle c_jt-|x| \right\rangle}^{-1}\bigr)\end{aligned}$$ for $(t,x)\in [0,\infty)\times({{\mathbb R}}^3\setminus\{0\})$ and $1\le j\le N$, where $u=(u_j)_{1\le j\le N}$ is the global solution to the Cauchy problem –. Moreover we have $$\label{Main33} {\partial}_\sigma^m U_j(\sigma, \omega)={\partial}_\sigma^m{\mathcal F}_0[f_j,c_j^{-1}g_j](\sigma, \omega)+O\bigl({\varepsilon}{\left\langle \sigma \right\rangle}^{-1-m}\bigr),\ (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^2$$ for $m=0,1$ and $1\le j\le N$. By we see that the asymptotic pointwise behavior of ${\partial}u_j$ is similar to that of the derivatives of the free solution (see Lemma \[PointwiseFriedlander\] below). Combining and with Theorem \[AFSE\], we see that the solution $u$ is asymptotically free in the energy sense. \[AsympFreeEnergy\] Suppose that , , , and the null condition are fulfilled. Then, for any $f,g\in C^\infty_0({{\mathbb R}}^3; {{\mathbb R}}^N)$ and sufficiently small ${\varepsilon}>0$, there exist $f^+=(f_j^+)_{1\le j\le N}\in \dot{H}^1({{\mathbb R}}^3;{{\mathbb R}}^N)$ and $g^+=(g_j^+)_{1\le j\le N} \in L^2({{\mathbb R}}^3;{{\mathbb R}}^N)$ such that $$\lim_{t\to\infty}\|(u_j-u_j^+)(t)\|_{E,c_j}=0,\quad 1\le j\le N,$$ where $u=(u_j)_{1\le j\le N}$ is the global solution to the Cauchy problem –, and $u_j^+$ is the solution to ${\square}_{c_j}u_j^+=0$ with initial data $(u_j^+, {\partial}_t u_j^+)=(f_j^+, g_j^+)$ at $t=0$. Theorem \[Main01\] and Corollary \[AsympFreeEnergy\] will be proved in Section \[P01\], after giving some preliminaries in Section \[P00\]. When $R^{\rm II}$ has the null structure, we see from that the solution $u_j$ itself also behaves similarly to the free solution. The next result shows that the lack of the estimate corresponding to in the presence of $R^{\rm II}$ without the null structure is inevitable. \[NotFree\] Let $0<c_1<c_2$, and let $A_1, A_2$ be real constants. We consider the Cauchy problem for $$\label{OurExample} \begin{cases} {\square}_{c_1}u_1=A_1({\partial}_t u_2)^2,\\ {\square}_{c_2}u_2=A_2({\partial}_t u_1)^2 \end{cases} \text{ in $(0,\infty)\times {{\mathbb R}}^3$}.$$ If $A_1\ne 0$, then there exist $(f,g)\in C^\infty_0({{\mathbb R}}^3;{{\mathbb R}}^2)\times C^\infty_0({{\mathbb R}}^3;{{\mathbb R}}^2)$, $M>0$, $T_0\ge 2$, and $C>0$ such that $$\label{Main41} C^{-1}{\varepsilon}\bigl(1+{\varepsilon}\log(2+t)\bigr) \le |r u_1(t,x)|\le C{\varepsilon}\bigl(1+{\varepsilon}\log(2+t)\bigr)$$ for any $(t,x)$ satisfying $T_0\le c_1t\le r(=|x|)\le c_1t+M$, where $u=(u_1, u_2)$ is the global solution to with initial data $u={\varepsilon}f$ and ${\partial}_t u={\varepsilon}g$ at $t=0$ for sufficiently small ${\varepsilon}(>0)$. This theorem will be proved in Section \[P02\]. From this result, we see that there is some loss in the decay rate of $u_1$: says that $u_1(t,x)$ decays like $(1+t)^{-1}\log(2+t)$ along the ray $r=c_1t+\sigma$ for $0\le \sigma \le M$, while implies that $u_j$ decays like $(1+t)^{-1}$ and has the same decay property as the free solution along the ray $r=c_jt+\sigma$ if $R^{\rm II}$ has the null structure. Accordingly we find that estimates like cannot hold in the presence of the non-resonant terms of type II without the null structure. To sum up the results, an interesting character of the non-resonant terms of type II is revealed: Their effect is weak enough for the solution to exist globally (Theorem 1.1 in [@Yok00]) and to be asymptotically free in the energy sense (Corollary \[AsympFreeEnergy\]); however it is strong enough to affect the decay rate of the solution $u$ itself (Theorem \[NotFree\]), though that of ${\partial}u$ is not affected (Theorem \[Main01\]). Throughout this paper, various positive constants are denoted by the same letter $C$. Thus the actual value of $C$ may change line by line. Asymptotics for Homogeneous Wave Equations {#ProofAFSE} ========================================== Our aim in this section is to prove Theorem \[AFSE\]. Most of the necessary materials for this purpose are rather standard, but we give the details and proofs to make this section self-contained. In what follows, we use the formal expression of writing distributions as if they are functions. Solutions to the free wave equation ----------------------------------- We put $H_0({{\mathbb R}}^n)=\dot{H}^1({{\mathbb R}}^n)\times L^2({{\mathbb R}}^n)$, and we define $$\label{DefNormH0} \|(\varphi,\psi)\|_{H_0({{\mathbb R}}^n)}^2:=\frac{1}{2}\left( \|\nabla_x \varphi\|_{L^2({{\mathbb R}}^n)}^2+\|\psi\|_{L^2({{\mathbb R}}^n)}^2\right).$$ $H_0({{\mathbb R}}^n)$ can also be understood as the completion of $C^\infty_0({{\mathbb R}}^n)\times C^\infty_0({{\mathbb R}}^n)$ with respect to the norm $\|\cdot\|_{H_0({{\mathbb R}}^n)}$. Let $c>0$, and consider the Cauchy problem for the free wave equation $$\begin{aligned} & {\square}_c w(t,x)=0, \quad (t,x)\in (0,\infty)\times {{\mathbb R}}^n, \label{FreeWaveEq}\\ & \bigl(w(0,x), ({\partial}_t w)(0,x)\bigr)=\bigl(w_0(x), w_1(x)\bigr),\quad x\in {{\mathbb R}}^n. \label{FreeData}\end{aligned}$$ Let $X_n$ be defined by . It is known that for $(w_0, w_1)\in H_0({{\mathbb R}}^n)$, – admits a unique solution $w\in X_n$, and we have the conservation of the energy $$\|w(t,\cdot)\|_{E,c}=\|w(0,\cdot)\|_{E,c}=\|(w_0, c^{-1}w_1)\|_{H_0({{\mathbb R}}^n)}.$$ For $\operatorname{\rm Re}a>-1$, let $\chi_+^a$ be defined by $$\chi_+^a(s):=\begin{cases} \displaystyle \frac{s^a}{\Gamma(a+1)}, & s> 0,\\ 0, & s\le 0, \end{cases}$$ where $\Gamma$ denotes the Gamma function. Then we have $\chi_+^{a}(s)=(\chi_+^{a+1})'(s)$ for $\operatorname{\rm Re}a>-1$. We can extend the definition of the distribution $\chi_+^a$ to all $a\in {{\mathbb C}}$ so that we have $\chi_+^{a}(s)=(\chi_+^{a+1})'(s)$ for any $a\in {{\mathbb C}}$. Note that $\chi_+^a$ can have its singularity only at $s=0$. Especially we have $$\chi_+^{-k}(s)=\delta^{(k-1)}(s),\quad k\in {{\mathbb N}},$$ where $\delta$ is the Dirac function and $\delta^{(j)}$ denotes its $j$-th derivative. For a positive integer $m$, we define $$E_m(t,x)=\frac{1}{2\pi^{(m-1)/2}}\chi_+^{(1-m)/2}(t^2-|x|^2).$$ Then the solution $w$ to – with $(w_0, w_1)\in \bigl(C^\infty_0({{\mathbb R}}^n)\bigr)^2$ can be written as $$\label{FreeSolExp} w(t,x)=c^{-1}{\partial}_t \bigl(E_n(ct,\cdot)*w_0\bigr)(x)+c^{-1}\bigl(E_n(ct,\cdot)*w_1)(x),$$ where the convolution $*$ is taken with respect to $x$-variable (see Hörmander [@Hoe90-01 Section 6.2] for instance). From now on, we suppose that $(w_0, w_1)\in \bigl(C^\infty_0({{\mathbb R}}^n)\bigr)^2$, and that $w_0(x)=w_1(x)=0$ for $|x|\ge M$ with some positive constant $M$. Since $\operatorname{\rm supp}E_n(t,\cdot)\subset \{x; |x|\le t\}$, it follows from that $$\label{Huygens01} w(t,x)=0,\quad |x|\ge ct+M.$$ If $n(\ge 3)$ is odd, then $\operatorname{\rm supp}E_n(t,\cdot)=\{x; |x|=t\}$ and we also get $$\label{Huygens02} w(t,x)=0,\quad |x|\le ct-M,$$ which is called the [*(strong) Huygens principle*]{}. If $n$ is even, is not valid in general, but we have a faster decay away from the light cone $ct=|x|$: Let $t/2\ge |x|$ and $t\ge 4M$, say. Then $t^2-|x-y|^2\ge 7t^2/16\ge C {\left\langle t+|x| \right\rangle}^2>0$ for $|y|\le M$. Hence, observing that $\chi^{(1-n)/2}(s)=A_ns^{(1-n)/2}$ for $s>0$ with an appropriate constant $A_n$, we get $$|{\partial}_{t,x}^\alpha E_n(t,x-y)|\le C_\alpha {\left\langle t+|x| \right\rangle}^{-|\alpha|+(1-n)},\quad t/2\ge \max\{|x|, 2M\},\ |y|\le M$$ with a positive constant $C_\alpha$. Therefore leads to $$\label{Huygens03} |{\partial}^\alpha w(t,x)|\le C_\alpha{\left\langle t+|x| \right\rangle}^{-|\alpha|+(1-n)}, \quad ct/2\ge |x|,$$ because for $2M\ge ct/2\ge |x|$ is easily shown. Following the arguments in Hörmander [@Hoe97 Section 6.2], we will obtain a useful expression of $E_m(t)*\varphi$ for $\varphi\in C^\infty_0({{\mathbb R}}^n)$ with $\varphi(x)=0$ for $|x|\ge M$. Note that we have $$\bigl(E_m(t)*\varphi\bigr)(x)=0, \quad |x|\ge t+M, \label{Huygens01'}$$ and $$\bigl(E_m(t)*\varphi\bigr)(x)=0, \quad |x|\le t-M\text{ when $m(\ge 3)$ is odd} \label{Huygens02'}$$ as in and . Let $x=r\omega$ with $r=|x|$ and $\omega\in {{\mathbb S}}^{n-1}$. We assume $$2M\le \frac{t}{2}\le r\le t+M.$$ We put $\sigma=r-t$. Then we get $-r\le \sigma\le M$. Since we have $$t^2-|x-y|^2=2r(\omega\cdot y-\sigma)+\sigma^2-|y|^2,$$ the homogeneity of $\chi_+^{(1-m)/2}$ implies that $$\begin{aligned} (2\pi r)^{(m-1)/2}\bigl(E_m(t)*\varphi\bigr)(x) =& \frac{1}{2}\int_{{{\mathbb R}}^n} \chi_+^{(1-m)/2}\left(\omega\cdot y-\sigma+ \frac{\sigma^2-|y|^2}{2r}\right)\varphi(y)dy \\ =& \frac{1}{2}\int_{{{\mathbb R}}} \chi_+^{(1-m)/2}\left(s-\sigma+ \frac{\sigma^2}{2r}\right){\mathcal G}[\varphi](s, \omega, r^{-1})ds,\end{aligned}$$ where ${\mathcal G}[\varphi](s, \omega, z)$ is given by $${\mathcal G}[\varphi](s, \omega, z)=\int_{{{\mathbb R}}^n}\delta\left(s-\omega\cdot y+\frac{|y|^2}{2}z\right)\varphi(y) dy$$ for $(s, \omega, z)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}\times [0, (2M)^{-1}]$. If we put $\rho=s-\omega\cdot y+|y|^2z/2$, then $\nabla_y\rho=-\omega+zy$. Since $|\nabla_y\rho|\ge 1-z|y|\ge 1/2$ for $z\in [0, (2M)^{-1}]$ and $|y|\le M$, ${\mathcal G}[\varphi]$ can be written as an integral of a compactly supported function $\varphi$ over a hyper-surface $\{y\in {{\mathbb R}}^n; s-\omega\cdot y+|y|^2z/2=0\}$ which smoothly depends on $(s,\omega, z)\in {{\mathbb R}}\times{{\mathbb S}}^{n-1}\times [0, (2M)^{-1}]$, and we see that $${\mathcal G}[\varphi]\in C^\infty\left({{\mathbb R}}\times {{\mathbb S}}^{n-1}\times \left[0, (2M)^{-1}\right]\right).$$ We also see that if ${\mathcal G}[\varphi](s,\omega,z)\ne 0$ for some $(\omega,z)\in {{\mathbb S}}^{n-1}\times \left[0, (2M)^{-1}\right]$, then we have $-5M/4\le s\le M$. Indeed, since the assumption implies $s=\omega\cdot y-(|y|^2/2)z$ for some $y$ and $(\omega, z)$ with $|y|\le M$ and $(\omega, z)\in {{\mathbb S}}^{n-1}\times \left[0, (2M)^{-1}\right]$, we get $$-\frac{5M}{4}\le -|y|-\frac{|y|^2}{2}\frac{1}{2M}\le \omega\cdot y-\frac{|y|^2}{2}z(=s) \le |y|\le M.$$ For a multi-index $\alpha$ with $|\alpha|=k\le 1$, we have $${\mathcal G}[{\partial}_x^\alpha \varphi](s, \omega, z)= {\partial}_s^k \int_{{{\mathbb R}}^n} \delta\left(s-\omega\cdot y+\frac{|y|^2}{2}z\right)(\omega-zy)^\alpha \varphi(y) dy. \label{DeriG}$$ We define $${\mathcal H}_m[\varphi](\sigma, \omega, z):=\frac{1}{2(2\pi)^{(m-1)/2}}\int_{{{\mathbb R}}} \chi_+^{(1-m)/2}\left(s-\sigma+ \frac{\sigma^2}{2}z\right){\mathcal G}[\varphi](s, \omega, z)ds.$$ Then we obtain $$\label{FreeSolExp02} r^{(m-1)/2} \bigl(E_m(t)*\varphi\bigr)(x)={\mathcal H}_m[\varphi](\sigma, \omega, r^{-1}).$$ Since we have $$\begin{aligned} & {\mathcal H}_m[\varphi](\sigma, \omega, z)\\ & \quad =\frac{(-1)^k}{2(2\pi)^{(m-1)/2}}\int_{{{\mathbb R}}} \chi_+^{k+(1-m)/2}\left(s-\sigma+ \frac{\sigma^2}{2}z\right)\bigl({\partial}_s^k{\mathcal G}[\varphi]\bigr)(s, \omega, z)ds\end{aligned}$$ for any nonnegative integer $k$, and since we have $\chi_+^a\in C^1({{\mathbb R}})$ for $a>1$, we can easily see that ${\mathcal H}_m[\varphi]\in C^\infty({{\mathbb R}}\times {{\mathbb S}}^{n-1}\times [0, (2M)^{-1}])$. Moreover we have $$\label{SupportHE} \sigma \le M \text{ in $\operatorname{\rm supp}{\mathcal H}_m[\varphi]$ when $m$ is even,}$$ and $$\label{SupportHO} |\sigma| \le M\text{ in $\operatorname{\rm supp}{\mathcal H}_m[\varphi]$ when $m(\ge 3)$ is odd.}$$ Indeed and for $z\ne 0$ follow from , , and , while they follow immediately from the definition of ${\mathcal H}_m$ when $z=0$. The Radon transform and the Friedlander radiation field ------------------------------------------------------- Let ${\mathcal S}({{\mathbb R}}^n)$ denote the set of rapidly decreasing functions on ${{\mathbb R}}^n$. For $\varphi\in {\mathcal S}({{\mathbb R}}^n)$ we define the [*Radon transform*]{} ${\mathcal R}[\varphi]$ of $\varphi$ by $$\label{DefRadonT} {\mathcal R}[\varphi](\sigma, \omega):=\int_{y\cdot \omega=\sigma} \varphi(y) dS(y), \quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1},$$ where $dS(y)$ denotes the surface element on the hyperplane $\{y\in {{\mathbb R}}^n; y\cdot\omega=\sigma\}$. It is easy to see that ${\mathcal R}[\varphi]\in{\mathcal S}({{\mathbb R}}\times {{\mathbb S}}^{n-1})$. Since we have ${\mathcal R}[\varphi](\sigma, \omega)={\mathcal G}[\varphi](\sigma, \omega, 0)$, it follows from that $$\label{RadonDeriExp01} {\mathcal R}[{\partial}_x^\alpha \varphi](\sigma, \omega)=\omega^\alpha {\partial}_\sigma^k{\mathcal R}[\varphi](\sigma, \omega)$$ for any multi-index $\alpha$ with $|\alpha|=k\le 1$. For a positive integer $m$ and $\varphi\in {\mathcal S}({{\mathbb R}}^n)$, we put $$\begin{aligned} {\mathcal R}_m[\varphi](\sigma, \omega):= & \frac{1}{2(2\pi)^{(m-1)/2}}\int_{{{\mathbb R}}}\chi_+^{(1-m)/2}(s-\sigma) {\mathcal R}[\varphi](s,\omega) ds\\ =& \frac{1}{2(2\pi)^{(m-1)/2}} \bigl(\chi_-^{(1-m)/2}*{\mathcal R}[\varphi](\cdot, \omega)\bigr)(\sigma), \quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1},\end{aligned}$$ where $\chi_-^a(\sigma):=\chi_+^a(-\sigma)$ for $a\in {{\mathbb C}}$, and $*$ is the convolution with respect to $\sigma$-variable. Note that we have ${\mathcal R}_m[\varphi]={\mathcal H}_m(\cdot,\cdot,0)\in C^\infty({{\mathbb R}}\times {{\mathbb S}}^{n-1})$. For any multi-index $\alpha$ with $|\alpha|=k\le 1$, we obtain from that $$\begin{aligned} {\mathcal R}_m[{\partial}_x^\alpha\varphi](\sigma,\omega) =& \frac{1}{2(2\pi)^{(m-1)/2}}\int_{{{\mathbb R}}}\chi_+^{(1-m)/2}(s-\sigma)\omega^\alpha {\partial}_s^k{\mathcal R}[\varphi](s,\omega)ds \nonumber\\ =& \frac{(-1)^k\omega^\alpha}{2(2\pi)^{(m-1)/2}}\int_{{{\mathbb R}}}\chi_+^{(1-m)/2-k}(s-\sigma){\mathcal R}[\varphi](s,\omega)ds\nonumber\\ =& \omega^\alpha {\partial}_\sigma^k{\mathcal R}_m[\varphi](\sigma, \omega). \label{RadonDeriExp02}\end{aligned}$$ If $\varphi(x)=0$ for $|x|\ge M$, then we immediately see by that $$\label{SupportRadon} {\mathcal R}[\varphi](\sigma,\omega)=0, \quad |\sigma|\ge M, \ \omega\in {{\mathbb S}}^{n-1}.$$ Consequently we get $$\label{SupportRNG} {\mathcal R}_m[\varphi](\sigma,\omega)=0, \quad \sigma\ge M,\ \omega\in {{\mathbb S}}^{n-1}.$$ When $m(\ge 3)$ is odd, since $\operatorname{\rm supp}\chi_+^{(1-m)/2}=\{0\}$, we obtain $$\label{SupportRNGO} {\mathcal R_m}[\varphi](\sigma,\omega)=0, \quad |\sigma|\ge M,\ \omega\in {{\mathbb S}}^{n-1}.$$ \[DecayPropRN\] Let $m$ be a positive integer with $m\ge 2$. For $\varphi\in C^\infty_0({{\mathbb R}}^n)$, a nonnegative integer $j$, and a multi-index $\alpha$ with $|\alpha|=k\le 1$, there is a positive constant $C$ such that $$\left|{\partial}_\sigma^j{\mathcal R}_m[{\partial}_x^\alpha \varphi](\sigma, \omega)\right|\le C {\left\langle \sigma \right\rangle}^{-j-k+(1-m)/2},\quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}.$$ Suppose that $\varphi(x)=0$ for $|x|\ge M$ with a positive constant $M$. Since ${\partial}_\sigma^j{\mathcal R}_m[{\partial}_x^\alpha \varphi]\in C^\infty({{\mathbb R}}\times {{\mathbb S}}^{n-1})$, we get $$\label{DP01} |{\partial}_\sigma^j{\mathcal R}_m[{\partial}_x^\alpha \varphi](\sigma, \omega)|\le C\le C{\left\langle 2M \right\rangle}^{j+k+(m-1)/2}{\left\langle \sigma \right\rangle}^{-j-k+(1-m)/2}$$ for $(\sigma,\omega)\in [-2M,M]\times {{\mathbb S}}^{n-1}$, which implies the desired result for odd $m$ because of . Let $m$ be even. In view of and , it suffices to consider the case where $\sigma\le -2M$. Then we have $s-\sigma\ge |\sigma|/2\ge C{\left\langle \sigma \right\rangle}>0$ for $|s|\le M$. Hence we obtain from that $$\begin{aligned} \left|{\partial}_\sigma^j{\mathcal R}_m[{\partial}_x^\alpha \varphi](\sigma, \omega)\right|=&\left|\frac{(-1)^{j+k}\omega^\alpha}{2(2\pi)^{(m-1)/2}}\int_{-M}^M \chi_+^{-j-k+(1-m)/2}(s-\sigma){\mathcal R}[\varphi](s, \omega) ds\right|\\ \le & C \int_{-M}^M (s-\sigma)^{-j-k+(1-m)/2} ds\le C{\left\langle \sigma \right\rangle}^{-j-k+(1-m)/2},\end{aligned}$$ because of . This completes the proof. \[FriedAsymp00\] Let $n\ge 2$, and $m$ be a positive integer with $m\ge 2$. Suppose that $\varphi\in C^\infty_0({{\mathbb R}}^n)$ and $\varphi(x)=0$ for $|x|\ge M$ with a positive constant $M$. Then, for any integer $j$ with $0\le j\le 2$, and a multi-index $\alpha$ with $|\alpha|=k\le 1$, there is a positive constant $C$ such that $$\begin{aligned} & \left|r^{(m-1)/2}{\partial}_t^j\bigl(E_m(t)*{\partial}_x^\alpha\varphi\bigr)(x)-\bigl((-{\partial}_\sigma)^j{\mathcal R}_m[{\partial}_x^\alpha \varphi]\bigr)(r-t, \omega) \right| \nonumber\\ & \qquad\qquad\qquad\qquad\qquad \le Cr^{-1}{\left\langle t-r \right\rangle}^{-j-k+(3-m)/2},\quad r\ge \frac{t}{2}\ge 2M, \label{BFE01}\end{aligned}$$ where $r=|x|$ and $\omega=|x|^{-1}x$. Recall the definitions of ${\mathcal G}$ and ${\mathcal H}_m$ in the previous subsection, and that we have $$\begin{aligned} {\mathcal G}[{\partial}_x^\alpha \varphi](s, \omega,0)=&{\mathcal R}[{\partial}_x^\alpha \varphi](s,\omega),\ {\mathcal H}_m[{\partial}_x^\alpha \varphi](\sigma, \omega, 0)={\mathcal R}_m[{\partial}_x^\alpha \varphi](\sigma,\omega).\end{aligned}$$ We suppose that $r\ge t/2\ge 2M$. By and , we may also assume $r\le t+M$. First we assume that $j=0$. We put $\sigma=r-t$ as before. By we obtain $$\begin{aligned} & \left|r^{(m-1)/2}\bigl(E_m(t)*{\partial}_x^\alpha\varphi\bigr)(x)-({\mathcal R}_m[{\partial}_x^\alpha \varphi])(r-t, \omega) \right|\\ & \quad =|{\mathcal H}_m[{\partial}_x^\alpha\varphi](\sigma,\omega, r^{-1})-{\mathcal H}_m[{\partial}_x^\alpha\varphi](\sigma, \omega, 0)| \\ & \quad \le r^{-1}\int_0^1\left|({\partial}_z {\mathcal H}_m[{\partial}_x^\alpha \varphi])(\sigma, \omega, \theta r^{-1})\right|d\theta,\end{aligned}$$ which leads to with $j=0$ if we can show $$\label{STAR} |{\partial}_z {\mathcal H}_m[{\partial}_x^\alpha \varphi](\sigma, \omega, z)| \le C{\left\langle \sigma \right\rangle}^{-k+(3-m)/2}$$ for $(\sigma, \omega, z)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}\times [0,(2M)^{-1}]$. Since ${\mathcal H}_m[{\partial}_x^\alpha \varphi]\in C^\infty\bigl({{\mathbb R}}\times {{\mathbb S}}^{n-1}\times [0, (2M)^{-1}]\bigr)$, we have $$\label{BHE01} |{\partial}_z {\mathcal H}_m[{\partial}_x^\alpha \varphi](\sigma, \omega, z)|\le C\le C{\left\langle \sigma \right\rangle}^{-k+(3-m)/2}$$ for $(\sigma,\omega, z)\in [-2M, M]\times {{\mathbb S}}^{n-1}\times [0, (2M)^{-1}]$, which leads to for odd $m$ because of . Let $m$ be even. In view of and , it suffices to show for $(\sigma,\omega,z)\in (-\infty, -2M]\times {{\mathbb S}}^{n-1}\times[0,(2M)^{-1}]$. Suppose $\sigma\le -2M$, $\omega\in {{\mathbb S}}^{n-1}$, and $0\le z \le(2M)^{-1}$. We compute $$\begin{aligned} & 2(2\pi)^{(m-1)/2}{\partial}_z{\mathcal H}_m[{\partial}_x^\alpha \varphi](\sigma, \omega, z) \\ & \quad = \int_{{{\mathbb R}}} \frac{\sigma^2}{2}\chi_+^{-(1+m)/2}\left(s-\sigma+\frac{\sigma^2}{2}z\right){\mathcal G}[{\partial}_x^\alpha \varphi](s,\omega, z)ds\\ &\qquad {}+\int_{{{\mathbb R}}} \chi_+^{(1-m)/2}\left(s-\sigma+\frac{\sigma^2}{2}z\right)({\partial}_z{\mathcal G}[{\partial}_x^\alpha \varphi])(s,\omega, z)ds=:I_1+I_2.\end{aligned}$$ Since $s-\sigma+\sigma^2z/2\ge -5M/4+|\sigma|\ge 3|\sigma|/8\ge C{\left\langle \sigma \right\rangle}>0$ for $\sigma\le -2M$, $z\ge 0$, and $s\ge -5M/4$, recalling we obtain $$\begin{aligned} |I_1|\le & \left|\int_{{{\mathbb R}}} \frac{\sigma^2}{2}\chi_+^{-k-(1+m)/2}\left(s-\sigma+\frac{\sigma^2}{2}z\right)\Phi_\alpha(s,\omega, z)ds\right|\\ \le & C|\sigma|^2{\left\langle \sigma \right\rangle}^{-k-(1+m)/2}\le C{\left\langle \sigma \right\rangle}^{-k+(3-m)/2},\end{aligned}$$ where $$\Phi_{\alpha}(s, \omega, z)=\int_{{{\mathbb R}}^n} \delta\left(s-\omega\cdot y+\frac{|y|^2}{2}z\right)(\omega-zy)^\alpha\varphi(y)dy$$ which is a $C^\infty$-function in ${{\mathbb R}}\times {{\mathbb S}}^{n-1}\times [0, (2M)^{-1}]$ with $-5M/4\le s\le M$ in $\operatorname{\rm supp}\Phi_\alpha$. Similarly we get $|I_2| \le C{\left\langle \sigma \right\rangle}^{-k+(3-m)/2}$. This completes the proof of for even $m$, and with $j=0$ is established. Note that by with $j=0$ and Lemma \[DecayPropRN\] we get $$\label{DecaySol11} \bigl|r^{(m-1)/2}\bigl(E_m(t)*{\partial}_x^\alpha \varphi\bigr)(x) \bigr|\le C{\left\langle \sigma \right\rangle}^{-k+(1-m)/2}.$$ Direct calculations lead to $${\partial}_t E_m(t,x)=2\pi t E_{m+2}(t,x),\ -{\partial}_\sigma{\mathcal R}_m[{\partial}_x^\alpha \varphi](\sigma, \omega) =2\pi {\mathcal R}_{m+2}[{\partial}_x^\alpha\varphi](\sigma,\omega).$$ Hence it follows from and with $j=0$ that $$\begin{aligned} & \bigl|r^{(m-1)/2}{\partial}_t \bigl(E_m(t)*{\partial}_x^\alpha\varphi\bigr)(x)-(-{\partial}_\sigma){\mathcal R}_m[{\partial}_x^\alpha\varphi](\sigma,\omega)\bigr|\\ & \quad \le 2\pi \bigl|r^{(m+1)/2} \bigl(E_{m+2}(t)*{\partial}_x^\alpha\varphi\bigr)(x)-{\mathcal R}_{m+2}[{\partial}_x^\alpha\varphi](\sigma,\omega)\bigr|\\ & \qquad {}+2\pi \bigl|r^{(m-1)/2}(t-r)\bigl(E_{m+2}(t)*{\partial}_x^\alpha \varphi\bigr)(x)\bigr| \le Cr^{-1}{\left\langle \sigma \right\rangle}^{-k+(1-m)/2},\end{aligned}$$ which is for $j=1$. Observing that $${\partial}_t^2E_m(t,x)=2\pi E_{m+2}(t,x)+(2\pi t)^2E_{m+4}(t)$$ and $$(-{\partial}_\sigma)^2{\mathcal R}_m[{\partial}_x^\alpha \varphi](\sigma, \omega) =(2\pi)^2 {\mathcal R}_{m+4}[{\partial}_x^\alpha \varphi](\sigma,\omega),$$ we can show for $j=2$ in a similar way. For $(\varphi, \psi)\in \bigl({\mathcal S}({{\mathbb R}}^n)\bigr)^2$, we define the [*Friedlander radiation field*]{} ${\mathcal F}_0[\varphi,\psi]$ by $$\label{DefGeneralFried} {\mathcal F}_0[\varphi,\psi](\sigma, \omega)=-{\partial}_\sigma{\mathcal R}_n[\varphi](\sigma, \omega)+{\mathcal R}_n[\psi](\sigma, \omega), \quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}.$$ Observe that this definition is a generalization of the previous definition that was given only for $n=3$ and $(\varphi,\psi)\in \bigl(C^\infty_0({{\mathbb R}}^3)\bigr)^2$, because ${\mathcal R}_3[h]=(4\pi)^{-1}{\mathcal R}[h]$. The next lemma is a slight refinement of [@Hoe97 Theorem 6.2.1] (see also Friedlander [@Fri62] and Katayama-Kubo [@Kat-Kub09a]). \[PointwiseFriedlander\] Let $n\ge 2$ and $c>0$. Let $w$ be the solution to the Cauchy problem –. If $(w_0, w_1)\in \bigl(C^\infty_0({{\mathbb R}}^n)\bigr)^2$, then there is a positive constant $C$ such that $$\bigl|r^{(n-1)/2} w(t, x)-W(r-ct,\omega)\bigr| \le C{\left\langle t+r \right\rangle}^{-1}{\left\langle ct-r \right\rangle}^{(3-n)/2} \label{FA01}$$ and $$\bigl|r^{(n-1)/2}{\partial}w(t,x)-\vec{\omega}_c(x) ({\partial}_\sigma W)(r-ct,\omega)\bigr| \le C{\left\langle t+r \right\rangle}^{-1}{\left\langle ct-r \right\rangle}^{(1-n)/2} \label{FA02}$$ for all $(t, x)\in [0,\infty)\times ({{\mathbb R}}^n\setminus\{0\})$, where $W(\sigma,\omega)={\mathcal F}_0[w_0, c^{-1}w_1](\sigma,\omega)$, $\vec{\omega}_c(x)$ is given by , $r=|x|$, $\omega=|x|^{-1}x$, and ${\partial}=({\partial}_0,{\partial}_1,\ldots, {\partial}_n)$. We may assume that $c=1$, because the general result is easily obtained by a change of variables. We assume that $w_0(x)=w_1(x)=0$ for $|x|\ge M$. Firstly we suppose that $r\ge t/2\ge 2M$. Then we have $r^{-1}\le C{\left\langle t+r \right\rangle}^{-1}$. Let $|\alpha|=k$, and suppose that $j$ and $k$ are nonnegative integers with $0\le j+k\le 1$. From we get $${\partial}_t^j{\partial}_x^\alpha w(t,x)={\partial}_t^{j+1}\bigl(E_n(t)*{\partial}_x^\alpha w_0\bigr)(x)+ {\partial}_t^j\bigl(E_n(t)*{\partial}_x^\alpha w_1\bigr)(x).$$ Then Lemma \[FriedAsymp00\] (with $m=n$) and lead to $$\begin{aligned} r^{(n-1)/2}{\partial}_t^j{\partial}_x^\alpha w(t,x)=& (-1)^j\omega^{\alpha} \bigl({\partial}_\sigma^{j+k}W\bigr)(r-t,\omega) \\ &{}+O\left({\left\langle t+r \right\rangle}^{-1}{\left\langle t-r \right\rangle}^{-j-k+(3-n)/2}\right),\end{aligned}$$ which implies when $j+k=0$, and when $j+k=1$. Secondly we suppose that either $r\le t/2$ or $t\le 4M$ holds. Then we have ${\left\langle t-r \right\rangle}^{-1}\le C{\left\langle t+r \right\rangle}^{-1}$. Hence by Lemma \[DecayPropRN\] (with $m=n$) we get $$\label{FaLast01} |({\partial}_\sigma^j W)(r-t, \omega)|\le C{\left\langle t-r \right\rangle}^{-j+(1-n)/2}\le C {\left\langle t+r \right\rangle}^{-j+(1-n)/2}.$$ Now we are going to prove $$\label{FaLast} |r^{(n-1)/2}{\partial}^\alpha w(t,x)|\le C {\left\langle t+r \right\rangle}^{-|\alpha|+(1-n)/2}$$ for $(t,x)$ with either $r\le t/2$ or $t\le 4M$, which implies and with the help of . If we assume $-2M\le r-t\le M$ in addition, then we get $t\le 4M$ and $r\le 5M$. Now, from we can easily obtain $$|r^{(n-1)/2}{\partial}^\alpha w(t,x)|\le C \le C{\left\langle t+r \right\rangle}^{-|\alpha|+(1-n)/2}, \quad t\le 4M,\ r\le 5M.$$ When $n$ is odd, this shows because of and . Thus we assume $n$ is even and $r-t\le -2M$. Accordingly we have $t/2\ge r$, and immediately implies the desired result. This completes the proof. The translation representation ------------------------------ For $\varphi\in {\mathcal S}({{\mathbb R}}^n)$, we write $\widehat{\varphi}(={\mathcal F}[\varphi])$ for its Fourier transform. To be more precise, we put $$\widehat{\varphi}(\xi)={\mathcal F}[\varphi](\xi)=\frac{1}{(2\pi)^{n/2}}\int_{{{\mathbb R}}^n} e^{-i x\cdot\xi} \varphi(x) dx,\quad \xi\in {{\mathbb R}}^n,$$ where $i=\sqrt{-1}$. For a function $\psi=\psi(\sigma, \omega)\in {\mathcal S}({{\mathbb R}}\times {{\mathbb S}}^{n-1})$, we define $$\widetilde{\psi}(\rho,\omega)=\widetilde{\mathcal F}[\psi](\rho,\omega):=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty e^{-i\rho\sigma}\psi(\sigma, \omega) d\sigma,\quad (\rho,\omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1},$$ which is the one-dimensional Fourier transform of $\psi(\cdot,\omega)$ with a parameter $\omega\in {{\mathbb S}}^{n-1}$. Let $\varphi\in {\mathcal S}({{\mathbb R}}^n)$. For $(\rho,\omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}$, we have $$\begin{aligned} \widehat{\varphi}(\rho\omega)=& \frac{1}{(2\pi)^{n/2}} \int_{{{\mathbb R}}^n} e^{-i\rho(y\cdot\omega)} \varphi(y) dy =\frac{1}{(2\pi)^{n/2}}\int_{-\infty}^\infty e^{-i \rho \sigma} {\mathcal R}[\varphi](\sigma, \omega)d\sigma\nonumber\\ = &\frac{1}{(2\pi)^{(n-1)/2}} \widetilde{{\mathcal R}[\varphi]}(\rho,\omega). \label{RadonFourier}\end{aligned}$$ Equation implies $$\label{Symm02} \widetilde{{\mathcal R}[\varphi]}( -\rho, -\omega)=(2\pi)^{(n-1)/2}\widehat{\varphi}(\rho\omega)=\widetilde{{\mathcal R}[\varphi]}(\rho, \omega),\quad (\rho,\omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}.$$ In other words $\widetilde{{\mathcal R}[\varphi]}$ is an even function in $(\rho,\omega)$. \[isometry\] For $(\varphi,\psi)\in \bigl({\mathcal S}({{\mathbb R}}^n)\bigr)^2$ we have $$\|{\partial}_\sigma{\mathcal F}_0[\varphi,\psi]\|_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}=\|(\varphi,\psi)\|_{H_0({{\mathbb R}}^n)}.$$ Since we have ${\mathcal R}_n[h](\sigma, \omega)=B_n\bigl(\chi_-^{(1-n)/2}*{\mathcal R}[h](\cdot,\omega)\bigr)(\sigma)$ for $h\in {\mathcal S}({{\mathbb R}}^n)$ with $B_n=1/\left(2(2\pi)^{(n-1)/2}\right)$, we get $$\label{RadonFried} \widetilde{{\mathcal R}_n[h]}(\rho,\omega) ={\mathcal F}_1\bigl[{\mathcal R}_n[h](\cdot,\omega)\bigr](\rho) =\sqrt{2\pi}B_n \mathcal{F}_1\bigl[\chi_{-}^{(1-n)/2}\bigr](\rho) \widetilde{{\mathcal R}[h]}(\rho,\omega),$$ where ${\mathcal F}_1$ denotes the one-dimensional Fourier transformation (of a tempered distribution). It is known that $$\mathcal{F}_1[\chi_-^a](\rho)=\frac{1}{\sqrt{2\pi}} e^{i\pi(a+1)/2}(\rho+i0)^{-a-1}$$ for every $a\in {{\mathbb C}}$, where $z^b$ for $z\in {{\mathbb C}}\setminus (-\infty,0]$ and $b\in {{\mathbb C}}$ is given by $z^b=\exp(b\log z)$ if we write $\log z=\log |z|+i \arg z$ with $-\pi<\arg z<\pi$, and $(\rho+i0)^b$ is defined by $(\rho+i0)^{b}=\lim_{{\varepsilon}\to +0} (\rho+i{\varepsilon})^{b}$ (see Hörmander [@Hoe90-01 Example 7.1.17] for instance). Especially we have $$\bigl|\mathcal{F}_1\bigl[\chi_-^{(1-n)/2}\bigr](\rho)\bigr|^2=\frac{1}{2\pi} |\rho|^{n-3},\quad \rho\in {{\mathbb R}}.$$ Therefore, for $\varphi, \psi \in {\mathcal S}({{\mathbb R}}^n)$ and nonnegative integers $j, k$, it follows from the Plancherel formula for the one-dimensional Fourier transform and that $$\begin{aligned} & \left\langle {\partial}_\sigma^j {\mathcal R}_n[\varphi], {\partial}_\sigma^k {\mathcal R}_n[\psi]\right\rangle_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})} \nonumber\\ & \ ={(-i)^ki^{j}B_n^2}\int_{{{\mathbb S}}^{n-1}}\left(\int_{-\infty}^\infty \widetilde{{\mathcal R}[\varphi]}(\rho,\omega) \overline{\widetilde{{\mathcal R}[\psi]}(\rho,\omega)} |\rho|^{n-3}\rho^{j+k} d\rho\right) dS_\omega, \label{Pla02}\end{aligned}$$ where $dS_\omega$ denotes the surface element on ${{\mathbb S}}^{n-1}$. By , we see that the integrand on the right-hand side of is an odd (resp. even) function in $(\rho,\omega)$ if $j+k$ is odd (resp. even). Hence we have $\left\langle {\partial}_\sigma^2{\mathcal R}_n[\varphi], {\partial}_\sigma {\mathcal R}_n[\psi]\right\rangle_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}=0$. Now we find $$\|{\partial}_\sigma{\mathcal F}_0[\varphi,\psi]\|_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}^2=\|{\partial}_\sigma^2{\mathcal R}_n[\varphi]\|_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}^2+\|{\partial}_\sigma{\mathcal R}_n[\psi]\|_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}^2.$$ We obtain from , , and that $$\begin{aligned} \|{\partial}_\sigma{\mathcal R}_n[\psi]\|_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}^2 =& 2{B_n^2}\int_{{{\mathbb S}}^{n-1}}\left(\int_{0}^\infty \left|\widetilde{{\mathcal R}[\psi]}(\rho,\omega)\right|^2 \rho^{n-1} d\rho\right) dS_\omega \\ =& \frac{1}{2}\bigl\|\widehat{\psi}\bigr\|_{L^2({{\mathbb R}}^n)}^2=\frac{1}{2}\|\psi\|_{L^2({{\mathbb R}}^n)}^2.\end{aligned}$$ Just in the same manner, we obtain $$\|{\partial}_\sigma^2{\mathcal R}_n[\varphi]\|_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}^2 =\frac{1}{2}\bigl\|\,|\cdot|\,\widehat{\varphi}\bigr\|_{L^2({{\mathbb R}}^n)}^2=\frac{1}{2}\|\nabla_x \varphi\|_{L^2({{\mathbb R}}^n)}^2.$$ This completes the proof. By Lemma \[isometry\], we can uniquely extend the linear mapping $$\bigl({\mathcal S}({{\mathbb R}}^n)\bigr)^2\ni (\varphi, \psi) \mapsto {\partial}_\sigma{\mathcal F}_0[\varphi,\psi] \in L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})$$ to the linear mapping ${\mathcal T}$ from $H_0({{\mathbb R}}^n)$ to $L^2({{\mathbb R}}^n)$ with $$\|{\mathcal T}[\varphi,\psi]\|_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}=\|(\varphi, \psi)\|_{H_0({{\mathbb R}}^n)},\quad (\varphi,\psi)\in H_0({{\mathbb R}}^n). \label{Isos}$$ This mapping ${\mathcal T}$ is called the [*translation representation*]{} in Lax-Phillips [@LaxPhi89]. The following lemma was essentially proved in [@LaxPhi89] for odd $n$, and in [@LaxPhi73] for even $n$ (see also Melrose [@Mel79]). The mapping $\mathcal T$ defined above is an isometric isomorphism from $H_0({{\mathbb R}}^n)$ to $L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})$. What is left to show is that ${\mathcal T}$ is surjective. Because of , we only have to prove the following: For any function $v$ in some dense subset of $L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})$, there is $(\varphi,\psi)\in H_0({{\mathbb R}}^n)$ such that ${\mathcal T}[\varphi,\psi]=v$. We write $h\in {\mathcal S}_0({{\mathbb R}}\times {{\mathbb S}}^{n-1})$ if $h \in{\mathcal S}({{\mathbb R}}\times {{\mathbb S}}^{n-1})$ and there is a positive constant $\delta$ such that $h(\rho, \omega)=0$ for all $(\rho, \omega)\in (-\delta,\delta)\times {{\mathbb S}}^{n-1}$. We put $${\mathcal S}_1({{\mathbb R}}\times {{\mathbb S}}^{n-1})=\left\{ \widetilde{\mathcal F}^{-1}[h]; h\in {\mathcal S}_0({{\mathbb R}}\times {{\mathbb S}}^{n-1}) \right\},$$ where $$\widetilde{{\mathcal F}}^{-1}[h](\sigma,\omega)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty e^{i\rho\sigma} h(\rho, \omega) d\rho,$$ which is the one-dimensional inverse Fourier transform of $h(\cdot,\omega)$ with a parameter $\omega\in {{\mathbb S}}^{n-1}$. It is easy to see that ${\mathcal S}_0({{\mathbb R}}\times {{\mathbb S}}^{n-1})$ is dense in $L^2({{\mathbb R}}\times{{\mathbb S}}^{n-1})$, and hence ${\mathcal S}_1({{\mathbb R}}\times {{\mathbb S}}^{n-1})$ is also dense in $L^2({{\mathbb R}}\times{{\mathbb S}}^{n-1})$. Let $v\in {\mathcal S}_1({{\mathbb R}}\times {{\mathbb S}}^{n-1})$. We want to find $(\varphi, \psi)\in H_0({{\mathbb R}}^n)$ such that ${\mathcal T}[\varphi,\psi]=v$. We put $$v_0(\rho, \omega)=\frac{1}{\sqrt{2\pi}B_n{\mathcal F}_1\bigl[\chi_-^{(1-n)/2}\bigr](\rho)} \widetilde{v}(\rho,\omega),\quad (\rho,\omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}.$$ Note that we have $v_0\in {\mathcal S_0}({{\mathbb R}}\times {{\mathbb S}}^{n-1})$, because $\widetilde{v}\in \mathcal{S}_0({{\mathbb R}}\times {{\mathbb S}}^{n-1})$, and the singularity of $1/{\mathcal F}_1\bigl[\chi_-^{(1-n)/2}\bigr]$ lies only at $\rho=0$. For $\xi\in {{\mathbb R}}^n\setminus\{0\}$ we put $$\begin{aligned} v_1(\xi)=& \frac{v_0(|\xi|, |\xi|^{-1}\xi)+v_0(-|\xi|, -|\xi|^{-1}\xi)}{2(2\pi)^{(n-1)/2}|\xi|^2}, \\ v_2(\xi)=& \frac{v_0(|\xi|, |\xi|^{-1}\xi)-v_0(-|\xi|, -|\xi|^{-1}\xi)}{2i(2\pi)^{(n-1)/2}|\xi|}.\end{aligned}$$ We also set $v_1(0)=v_2(0)=0$. Then, since $v_0\in{\mathcal S}_0({{\mathbb R}}\times {{\mathbb S}}^{n-1})$, we find that $v_1, v_2\in {\mathcal S}({{\mathbb R}}^n)$. Hence if we set $\varphi={\mathcal F}^{-1}[v_1]$ and $\psi={\mathcal F}^{-1}[v_2]$, then we get $(\varphi,\psi)\in \bigl({\mathcal S}({{\mathbb R}}^n)\bigr)^2\subset H_0({{\mathbb R}}^n)$. Using and , we get $$\widetilde{\mathcal F}\left[ {\partial}_\sigma{\mathcal F}_0[\varphi,\psi]\right](\rho,\omega)=\rho^2 \widetilde{{\mathcal R}_n[\varphi]}(\rho,\omega)+i\rho\widetilde{\mathcal{R}_n[\psi]}(\rho,\omega)=\widetilde{v}(\rho,\omega)$$ for $(\rho, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}$, which shows ${\mathcal T}[\varphi,\psi]={\partial}_\sigma{\mathcal F}_0[\varphi, \psi]=v$. This completes the proof. Theorem \[AFSE\] is an immediate consequence of the following lemma. \[Final01\] Let $c>0$, and $(w_0, w_1)\in H_0({{\mathbb R}}^n)$. Let $w$ be a solution to the Cauchy problem –, and $${\mathcal W}(t,x)=|x|^{-(n-1)/2}{\mathcal T}[w_0, c^{-1}w_1](|x|-ct,|x|^{-1}x)$$ for $(t,x)\in [0,\infty)\times ({{\mathbb R}}^n\setminus\{0\})$. Then we have $$\lim_{t\to \infty} \bigl\|{\partial}^{(c)} w(t,\cdot)-\vec{\omega}(\cdot){\mathcal W}(t,\cdot)\bigr\|_{L^2({{\mathbb R}}^n)} =0,$$ where ${\partial}^{(c)}:=(c^{-1}{\partial}_t, {\partial}_1, \ldots , {\partial}_n)$ and $\vec{\omega}(x):=\vec{\omega}_1(x)=(-1, |x|^{-1}x)$. Let ${\varepsilon}>0$. There is $(w_0^*,w_1^*)\in \bigl(C^\infty_0({{\mathbb R}}^n)\bigr)^2$ such that $$\left\|(w_0, c^{-1}w_1)-(w_0^*, c^{-1} w_1^*)\right\|_{H_0({{\mathbb R}}^n)}^2<\left(\frac{{\varepsilon}}{3}\right)^2.$$ Let $w^*$ and ${\mathcal W}^*$ be defined similarly to $w$ and ${\mathcal W}$, respectively, by replacing $(w_0, w_1)$ with $(w_0^*, w_1^*)$. Recalling , we obtain from the energy identity that $$\begin{aligned} & \frac{1}{2}\|{\partial}^{(c)} w(t,\cdot)-{\partial}^{(c)} w^*(t,\cdot)\|_{L^2({{\mathbb R}}^n)}^2 \\ & \quad = \frac{1}{2}\left(\|c^{-1}(w_1-w_1^*)\|_{L^2({{\mathbb R}}^n)}^2+\|\nabla_x(w_0-w_0^*)\|_{L^2({{\mathbb R}}^n)}^2\right)\\ & \quad = \|(w_0, c^{-1}w_1)-(w_0^*, c^{-1}w_1^*)\|_{H_0({{\mathbb R}}^n)}^2 < \left(\frac{{\varepsilon}}{3}\right)^2.\end{aligned}$$ Since $|\vec{\omega}(x)|=2$, using we get $$\begin{aligned} & \frac{1}{2}\bigl\| \vec{\omega}(\cdot)\bigl({\mathcal W}(t,\cdot)-{\mathcal W}^*(t,\cdot)\bigr) \bigr\|_{L^2({{\mathbb R}}^n)}^2 \nonumber\\ & \quad = \int_0^\infty \left(\int_{{{\mathbb S}}^{n-1}} |{\mathcal T}[w_0-w_0^*, c^{-1}(w_1-w_1^*)](r-ct,\omega)|^2 dS_\omega\right) dr \nonumber\\ & \quad \le \|{\mathcal T}[w_0-w_0^*, c^{-1}(w_1-w_1^*)]\|_{L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})}^2 \nonumber\\ & \quad =\|(w_0, c^{-1}w_1)-(w_0^*, c^{-1}w_1^*)\|_{H_0({{\mathbb R}}^n)}^2< \left(\frac{{\varepsilon}}{3}\right)^2. \nonumber\end{aligned}$$ By Lemma \[PointwiseFriedlander\] we obtain $$\begin{aligned} & \frac{1}{2}\|{\partial}^{(c)} w^*(t,\cdot)-\vec{\omega}(\cdot){\mathcal W}^*(t,\cdot)\|_{L^2({{\mathbb R}}^n)}^2 \\ & \quad \le \frac{1}{2} \int_0^\infty\left(\int_{{{\mathbb S}}^{n-1}} |r^{(n-1)/2} {\partial}^{(c)} w^{*}(t,r\omega)- \vec{\omega}(r\omega) ({\partial}_\sigma {\mathcal F}_0^*) (r-ct,\omega)|^2dS_\omega\right)dr \\ & \quad \le C \int_0^\infty (1+t+r)^{-2}(1+|t-r|)^{-(n-1)} dr\le C(1+t)^{-1},\end{aligned}$$ where ${\mathcal F}_0^*(\sigma, \omega)={\mathcal F}_0[w_0^*, c^{-1}w_1^*](\sigma, \omega)$. Hence there is a positive constant $t_0>0$ such that $t\ge t_0$ implies $$\frac{1}{2} \|{\partial}^{(c)} w^*(t,\cdot)-\vec{\omega}(\cdot){\mathcal W}^*(t,\cdot)\|_{L^2({{\mathbb R}}^n)}^2 <\left(\frac{{\varepsilon}}{3}\right)^2.$$ To sum up, for any ${\varepsilon}>0$ there is a positive constant $t_0$ such that we have $$\frac{1}{\sqrt{2}} \bigl\|{\partial}^{(c)} w(t,\cdot)-\vec{\omega}(\cdot){\mathcal W}(t,\cdot)\bigr\|_{L^2({{\mathbb R}}^n)} <{\varepsilon},\quad t\ge t_0.$$ This completes the proof. Now we are in a position to prove Theorem \[AFSE\]. Suppose that there is $V\in L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})$ such that $$\label{Fin03} \lim_{t\to \infty} \bigl\|{\partial}v(t,\cdot)-\vec{\omega}_c(\cdot){V}^\sharp(t,\cdot)\bigr\|_{L^2({{\mathbb R}}^n)} =0,$$ where $$\label{Fin01a} {V}^\sharp(t,x)=|x|^{-(n-1)/2} V(|x|-ct, |x|^{-1}x).$$ Then we get $$\label{Fin03a} \lim_{t\to \infty} \bigl\|{\partial}^{(c)} v(t,\cdot)-\vec{\omega}(\cdot){V}^\sharp(t,\cdot)\bigr\|_{L^2({{\mathbb R}}^n)} =0,$$ where ${\partial}^{(c)}$ and $\vec{\omega}(x)$ are defined as in Lemma \[Final01\]. We define $(w_0, c^{-1}w_1)={\mathcal T}^{-1} [V]\bigl(\in H_0({{\mathbb R}}^n)\bigr)$, and let $w$ be the solution to the Cauchy problem – for this $(w_0, w_1)$. Then it follows from Lemma \[Final01\] that $$\label{Fin02} \lim_{t\to\infty} \bigl\|{\partial}^{(c)} w(t,\cdot)-\vec{\omega}(\cdot){V}^\sharp(t,\cdot)\bigr\|_{L^2({{\mathbb R}}^n)} =0,$$ which, together with implies $$\label{Fin01} \lim_{t\to\infty} \|v(t,\cdot)-w(t,\cdot)\|_{E,c} =\lim_{t\to\infty}\frac{1}{\sqrt{2}}\|{\partial}^{(c)}v(t,\cdot)-{\partial}^{(c)}w(t,\cdot)\|_{L^2({{\mathbb R}}^n)} =0.$$ Conversely, suppose that there is $(w_0, w_1)\in H_0({{\mathbb R}}^n)$ such that holds for the solution $w$ to the Cauchy problem –. If we put $$V(\sigma, \omega)={\mathcal T}[w_0, c^{-1}w_1](\sigma, \omega)\bigl(\in L^2({{\mathbb R}}\times {{\mathbb S}}^{n-1})\bigr)$$ and define ${V}^\sharp$ by , then Lemma \[Final01\] implies . Now and yield , which is equivalent to . This completes the proof. Vector Fields Associated with the Wave Equations {#P00} ================================================ We restrict our attention to the three space dimensional case from now on. We introduce vector fields $$\begin{aligned} S: = & t{\partial}_t+x\cdot\nabla_x,\\ \Omega= & (\Omega_1,\Omega_2,\Omega_3):=x\times \nabla_x=(x_2{\partial}_3-x_3{\partial}_2, x_3{\partial}_1-x_1{\partial}_3, x_1{\partial}_2-x_2{\partial}_1),\end{aligned}$$ where the symbols $\cdot$ and $\times$ denote the inner and exterior products in ${{\mathbb R}}^3$, respectively. We also use ${\partial}=({\partial}_t,\nabla_x)=({\partial}_a)_{0\le a\le 3}$. We define $$Z=(Z_0, Z_1,\ldots, Z_7)=(S,\Omega,{\partial})=(S, \Omega_1,\Omega_2, \Omega_3,{\partial}_0,{\partial}_1,{\partial}_2,{\partial}_3).$$ These vector fields are compatible with the system of wave equations having multiple propagation speeds, since we have $[{\square}_c, \Omega_j]=[{\square}_c, {\partial}_a]=0$ for $1\le j\le 3$ and $0\le a\le 3$, and $[{\square}_c, S]=2{\square}_c$, where $c>0$, and $[P,Q]=PQ-QP$ for operators $P$ and $Q$. Using a multi-index $\alpha=(\alpha_0,\alpha_1,\ldots, \alpha_7)$, we write $Z^\alpha=Z_0^{\alpha_0}Z_1^{\alpha_1}\cdots Z_7^{\alpha_7}$. For a nonnegative integer $s$ and a (scalar- or vector-valued) smooth function $\varphi=\varphi(t,x)$, we define $$|\varphi(t,x)|_s=\sum_{|\alpha|\le s}| Z^\alpha\varphi(t,x)|.$$ We can check that we have $[Z_a, Z_b]=\sum_{d=0}^7 C^{ab}_d Z_d$ and $[Z_a, {\partial}_b]=\sum_{d=0}^3 D^{ab}_d {\partial}_d$ with appropriate constants $C^{ab}_d$ and $D^{ab}_d$. Hence, for any multi-indices $\alpha$ and $\beta$, and any nonnegative integer $s$, there exist some positive constants $C_{\alpha,\beta}$ and $C_s$ such that we have $$\begin{aligned} & |Z^\alpha Z^\beta \varphi(t,x)|\le C_{\alpha,\beta}|\varphi(t,x)|_{|\alpha|+|\beta|},\\ & C_s^{-1}|{\partial}\varphi(t,x)|_s \le \sum_{|\alpha|\le s}\sum_{a=0}^3 |{\partial}_a Z^\alpha\varphi(t,x)| \le C_s|{\partial}\varphi(t,x)|_s\end{aligned}$$ for any smooth function $\varphi$. We write $r=|x|$, $\omega=(\omega_1,\omega_2,\omega_3)=|x|^{-1}x$, and ${\partial}_r=\sum_{j=1}^3\omega_j {\partial}_j$. Then we can write $S=t{\partial}_t+r{\partial}_r$. For $c> 0$, we define $${\partial}_{\pm}^{(c)}={\partial}_t\pm c{\partial}_r,\text{ and } D_\pm^{(c)}=\pm\frac{1}{2c}{\partial}_\pm^{(c)}=\frac{1}{2}({\partial}_r\pm c^{-1}{\partial}_t).$$ Since we have ${\partial}_t=-c\bigl(D_-^{(c)}-D_+^{(c)}\bigr)$, ${\partial}_r=D_-^{(c)}+D_{+}^{(c)}$, and $$(1+r){\partial}_{+}^{(c)}=cS-(ct-r){\partial}_t+({\partial}_t+c{\partial}_r),$$ there exists a positive constant $C$ such that $$\begin{aligned} & \bigl|\bigl({\partial}_t-(-c)D_{-}^{(c)}\bigr)\varphi(t,x)\bigr|+\bigl|\bigl({\partial}_r-D_-^{(c)}\bigr) \varphi(t,x)\bigr| \nonumber\\ & \qquad\qquad\quad \le \frac{C}{1+r}\left(|Z\varphi(t,x)|+|ct-r|\,|{\partial}\varphi(t,x)|\right) \label{frame01}\end{aligned}$$ for any smooth function $\varphi$. Since we have $\nabla_x=\omega{\partial}_r-r^{-1}\omega\times \Omega$, we get $$\label{frame02} |({\partial}_k-\omega_k{\partial}_r)\varphi(t,x)|\le C(1+r)^{-1}|Z\varphi(t,x)|,\quad k=1,2,3.$$ From and we obtain $$\label{frame03} \left|{\partial}\varphi(t,x)-\vec{\omega}_c(x)D_-^{(c)}\varphi(t,x)\right| \le \frac{C}{1+r}\left(|Z\varphi(t,x)|+|ct-r|\,|{\partial}\varphi(t,x)|\right)$$ for any smooth function $\varphi$, where $\vec{\omega}_c(x)$ is given by . As an immediate consequence we also get $$\label{frame04} \left|r{\partial}\varphi(t,x)-\vec{\omega}_c(x)D_{-}^{(c)}\bigl(r\varphi(t,x)\bigr)\right| \le C\left(|\varphi(t,x)|_1+|ct-r|\,|{\partial}\varphi(t,x)|\right).$$ We remark that the term $|ct-r|\,|{\partial}\varphi|$ was not needed in the estimates used in [@Kat11] instead of and ; however the vector field $L=t\nabla_x+x{\partial}_t$, which is not compatible with the multiple speed case, was involved. We need the term $|ct-r|\,|{\partial}\varphi|$ here to compensate the lack of the vector field $L$. These kinds of identities and estimates without the vector field $L$ were developed and used in [@KlaSid96], [@Age00], [@Sid00], [@SidTu01], [@Yok00], and so on (see also [@Kat12], [@KatKub08], [@SidThomases05], [@SidThomases06], and [@SidThomases07] for the related topics). Using , we can easily show the following estimate for the null forms $Q_0$ and $Q_{ab}$ given by and (see [@Sid00], [@SidTu01] and [@Yok00] for the details of the proof): \[NullEst\] Let $c>0$. Then we have $$\begin{aligned} |Q_0(\varphi, \psi;c)|+\sum_{0\le a<b\le 3} |Q_{ab}(\varphi,\psi)|\le & C{\left\langle r \right\rangle}^{-1}\left(|Z\varphi|\,|{\partial}\psi| {}+|{\partial}\varphi|\,|Z\psi|\right)\\ &{}+C{\left\langle r \right\rangle}^{-1}{\left\langle ct-r \right\rangle}|{\partial}\varphi|\,|{\partial}\psi|\end{aligned}$$ at $(t,x)\in(0,\infty)\times {{\mathbb R}}^3$ with $r=|x|$. Fix $c>0$, and we define $R=(R_a)_{0\le a\le 3}={\partial}-\vec{\omega}_c(x)D_-^{(c)}$. Substituting ${\partial}=\vec{\omega}_c(x)D_{-}^{(c)}+R$, we get $$\begin{aligned} |Q_0(\varphi,\psi;c)|+\sum_{a,b}|Q_{ab}(\varphi,\psi)|\le & C \bigl(\bigl|D_-^{(c)}\varphi\bigr|\,|R\psi|+|R\varphi|\,\bigl|D_-^{(c)}\psi\bigr|+|R\varphi|\,|R\psi|\bigr)\\ \le & C \left(|{\partial}\varphi|\,|R\psi|+|R\varphi|\,|{\partial}\psi|\right),\end{aligned}$$ where we have used $|R\psi|\le C|{\partial}\psi|$ to obtain the last line. The point here is that terms including $\bigl(D_-^{(c)}\varphi\bigr)\bigl(D_-^{(c)}\psi\bigr)$ are canceled out because of the structure of the null forms. Now, using to evaluate $|R\varphi|$ and $|R\psi|$, we obtain the desired result. Let $c_1, \ldots, c_N$ satisfy . We define $$\label{DefLambda0} \Lambda_0=\left\{(t,r)\in {{\mathbb R}}_+ \times {{\mathbb R}}_+; r\ge \frac{c_1t}{2}\ge 1\right\},$$ where ${{\mathbb R}}_+=[0,\infty)$. For $1\le j\le N$, we define $$t_{0, j}(\sigma):=\max\{-2(2c_j-c_1)^{-1}\sigma, 2c_1^{-1}\} \label{Deft0i}$$ so that we have $\bigcup_{\sigma\in {{\mathbb R}}} \{(t, c_j t+\sigma); t\ge t_{0,j}(\sigma)\}=\Lambda_0$. We also define $r_{0, j}(\sigma):=c_jt_{0, j}(\sigma)+\sigma$ for $\sigma\in {{\mathbb R}}$. The following lemma is a modification of a key lemma in [@Kat11] to obtain the asymptotic behavior. \[KeyLemma\] Assume , and let $j\in\{1,\ldots, N\}$. Suppose the following: - $\mu_j>1$ and $\kappa_j\ge 0$. - $\mu_k\ge 0$ and $\kappa_k>1$ for any $k=1,\ldots, N$ with $k\ne j$. If $v=v(t,r,\omega)\in C^1(\Lambda_0\times {{\mathbb S}}^2)$ satisfies $$\label{Kanon} {\partial}_+^{(c_j)}v(t,r,\omega)=G(t,r,\omega),\quad (t,r)\in \Lambda_0,\ \omega\in {{\mathbb S}}^2,$$ and $G$ satisfies $$\label{Air} |G(t, r, \omega)|\le \sum_{k=1}^N B_k{\left\langle t+r \right\rangle}^{-\mu_k}{\left\langle c_kt-r \right\rangle}^{-\kappa_k},\quad (t,r)\in \Lambda_0,\ \omega\in {{\mathbb S}}^2$$ with some positive constants $B_1,\ldots, B_N$, then there exists a positive constant $C$ such that $$\begin{aligned} |v(t,r,\omega)-V(r-c_jt,\omega)|\le & {CB_j}{\left\langle t+r \right\rangle}^{-\mu_j+1}{\left\langle c_jt-r \right\rangle}^{-\kappa_j} \nonumber\\ & {} +C\sum_{\substack{1\le k\le N \\ k\ne j}} B_k {\left\langle t+r \right\rangle}^{-\mu_k} \label{Clannad}\end{aligned}$$ for any $(t,r)\in \Lambda_0$ and $\omega\in {{\mathbb S}}^2$, where $V$ is defined by $$\label{LittleBusters} V(\sigma, \omega)=v\bigl(t_{0, j}(\sigma), r_{0, j}(\sigma),\omega \bigr) {}+\int_{t_{0, j}(\sigma)}^\infty G(s, c_j s+\sigma, \omega) ds.$$ The constant $C$ above is determined only by $c_k$, $\mu_k$, and $\kappa_k$ with $1\le k\le N$. First we note that $V$ is well-defined because of . By we find $$v(t,c_jt+\sigma,\omega)=v\bigl(t_{0,j}(\sigma), r_{0,j}(\sigma),\omega\bigr) {}+\int_{t_{0,j}(\sigma)}^t G(s, c_js+\sigma, \omega) ds,\quad t\ge t_{0,j}(\sigma)$$ for $(\sigma,\omega) \in {{\mathbb R}}\times {{\mathbb S}}^2$. From and we get $$|v(t,c_jt+\sigma,\omega)-V(\sigma, \omega)| \le \int_{t}^\infty |G(s, c_js+\sigma, \omega)|ds \le C\sum_{k=1}^N B_k I_k(t,\sigma), \label{TomoyoAfter}$$ where $$I_k(t,\sigma)=\int_t^\infty \left. (1+\tau+\rho)^{-\mu_k}(1+|c_k\tau-\rho|)^{-\kappa_k}\right|_{(\tau, \rho)=(s,c_js+\sigma)}ds$$ for $1\le k\le N$. By direct calculations, we obtain $$\label{Planetarian} I_j(t,\sigma)=\frac{1}{(c_j+1)(\mu_j-1)}(1+(c_j+1)t+\sigma)^{-\mu_j+1}(1+|\sigma|)^{-\kappa_j},$$ since $\mu_j>1$. If $k\ne j$, then we have $$\begin{aligned} I_k(t,\sigma)\le & (1+(c_j+1)t+\sigma)^{-\mu_k}\int_t^\infty (1+|(c_k-c_j)s-\sigma|)^{-\kappa_k}ds \nonumber\\ \le & (1+(c_j+1)t+\sigma)^{-\mu_k}\frac{1}{|c_k-c_j|}\int_{-\infty}^\infty(1+|\tau|)^{-\kappa_k}d\tau \nonumber\\ \le & \frac{2}{(\kappa_k-1)|c_k-c_j|}(1+(c_j+1)t+\sigma)^{-\mu_k} \label{KudoWafter}\end{aligned}$$ for $t\ge t_{0,j}(\sigma)$ and $\sigma\in {{\mathbb R}}$, because we have $\mu_k\ge 0$, $\kappa_k>1$, and $c_k\ne c_j$. Now we obtain immediately by putting $\sigma=r-c_jt$ in , , and . Proof of Theorem \[Main01\] and Corollary \[AsympFreeEnergy\] {#P01} ============================================================= In the following, we write $r=|x|$ and $\omega=|x|^{-1}x$. Let the assumptions in Theorem \[Main01\] be fulfilled. Then we have the global solution $u$ to the Cauchy problem – for sufficiently small ${\varepsilon}$ by the global existence theorems in [@SidTu01], [@Sog03], or [@Yok00]. For $j=1, \ldots, N$, we put $${u}_j^1(t,x)=u_j(t,x)-{\varepsilon}u_j^0(t,x),$$ where $u_j^0$ is the solution to ${\square}_{c_j}u_j^0=0$ with initial data $u_j^0=f_j$ and ${\partial}_t u_j^0=g_j$ at $t=0$. From and , we have $$\begin{aligned} \label{OurSysM} & {\square}_{c_j} u_j^1=F_j({\partial}u, {\partial}^2 u)\quad\text{ in $(0,\infty)\times {{\mathbb R}}^3$}, \\ \label{OurDataM} & u_j^1(0,x)=({\partial}_t {u}_j^1)(0,x)=0,\quad x\in {{\mathbb R}}^3.\end{aligned}$$ We put $$U_j^0(\sigma, \omega)={\mathcal F}_0[f_j, c_j^{-1}g_j](\sigma, \omega).$$ Suppose that we have $f(x)=g(x)=0$ for $|x|\ge M$. Then, as in , , and , we get $$\begin{aligned} \label{Supportui0} u_j^0(t,x)=&0,\quad |r-c_jt|\ge M,\\ U_j^0(\sigma,\omega)=&0,\quad |\sigma|\ge M \label{SupportUi0}\end{aligned}$$ for $1\le j\le N$. Hence Lemma \[PointwiseFriedlander\] implies $$\begin{aligned} & \left| r u_j^0(t,x)-U_j^0(r-c_jt,x) \right| {}+\left| r {\partial}u_j^0(t,x)-\vec{\omega}_{c_j}(x)({\partial}_\sigma U_j^0)(r-c_jt,x) \right| \nonumber\\ & \qquad\qquad\qquad\qquad\qquad\qquad \le C{\left\langle t+r \right\rangle}^{-1}{\left\langle c_jt-r \right\rangle}^{-1},\quad 1\le j\le N \label{FreeApprox}\end{aligned}$$ for $(t,x)\in [0,\infty)\times({{\mathbb R}}^3\setminus\{0\})$, because and imply $1\le {\left\langle c_jt-r \right\rangle}\le {\left\langle M \right\rangle}$ on the support of the functions on the left-hand side of . Let $\Lambda$ be the set of $(t,x)\in [0,\infty)\times {{\mathbb R}}^3$ with $(t,r)\in \Lambda_0$, where $\Lambda_0$ is given by ; namely we put $$\Lambda=\left\{(t,x)\in[0,\infty)\times ({{\mathbb R}}^3\setminus\{0\}); |x|\ge \frac{c_1t}{2}\ge 1\right\}.$$ We set $\Lambda^{\rm c}=\bigl([0,\infty)\times ({{\mathbb R}}^3\setminus\{0\})\bigr) \setminus \Lambda$. Recall the definitions of ${\partial}_\pm^{(c)}$ and $D_\pm^{(c)}$ in the previous section. Following the proof of the global existence theorem in [@Kub-Yok01] (see also [@Kat04:02 Proposition 4.2] and its proof) we obtain $$\begin{aligned} |u_k(t,x)|_2+{\varepsilon}^{-1}|{u}_k^1(t,x)|_2\le & C{\varepsilon}{\left\langle t+r \right\rangle}^{-1}\log \left(1+\frac{1+c_k t+r}{1+|c_k t-r|}\right), \label{Basic01a}\\ |{\partial}{u}_k(t,x)|_1+{\varepsilon}^{-1}|{\partial}{u}_k^1(t,x)|_1\le & C{\varepsilon}{\left\langle r \right\rangle}^{-1}{\left\langle c_kt-r \right\rangle}^{-1} \label{Basic01b}\end{aligned}$$ for $1\le k\le N$. We choose small $\delta>0$. Then, by we get $$\label{Basic01a'} |{u}_k(t,x)|_2+{\varepsilon}^{-1}|{u}_k^1(t,x)|_2\le C{\varepsilon}{\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_kt-r \right\rangle}^{-\delta}$$ for $1\le k\le N$, because there is a positive constant $C_\delta$ such that we have $\log(1+z)\le C_\delta z^\delta$ for $z\ge 1$. Fix $j=1,2,\ldots, N$. Switching to the polar coordinates, we get $$\label{Polar} r({\square}_{c_j} {u}_j^1)(t, r\omega) ={\partial}_{+}^{(c_j)}{\partial}_{-}^{(c_j)}\bigl(r{u}_j^1(t, r\omega)\bigr) {}-c_j^2 r^{-1}\Delta_\omega {u}_j^1(t,r\omega)$$ for $(t,r,\omega)\in (0,\infty)\times (0,\infty)\times {{\mathbb S}}^2$, where $\Delta_\omega=\sum_{k=1}^3 \Omega_k^2$ is the Laplace-Beltrami operator on ${{\mathbb S}}^2$. We put $$\label{DefVj} {v}_j(t,r,\omega):=D_{-}^{(c_j)}\bigl(r{u}_j^1(t, r\omega)\bigr).$$ Then and lead to $$\label{ReducedEq} {\partial}_{+}^{(c_j)}{v}_j(t,r,\omega)=G_j(t,r,\omega),$$ where $$G_j(t,r,\omega):=-\frac{1}{2c_j}\left\{c_j^2 r^{-1}\Delta_\omega {u}_j^1(t, r\omega)+r F_j\bigl({\partial}u(t,r\omega), {\partial}^2 u(t,r\omega)\bigr)\right\}.$$ We suppose $(t, r)\in \Lambda_0$ and $\omega\in {{\mathbb S}}^2$ for a while. Note that we have $${\left\langle 1+2c_1^{-1} \right\rangle}^{-1} {\left\langle t+r \right\rangle}\le r\le {\left\langle r \right\rangle}\le {\left\langle t+r \right\rangle},\quad (t, r)\in \Lambda_0.$$ From we obtain $$\label{Est01} |r^{-1}\Delta_\omega {u}_j^1(t,r\omega)|\le C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-2+\delta} {\left\langle c_jt-r \right\rangle}^{-\delta}.$$ Recall the definitions of $N_j$, $R_j^{\rm I}$, and $R_j^{\rm II}$ given in , , and , respectively. By Lemma \[NullEst\], , and , we get $$\label{Est02} r|N_j({\partial}u, {\partial}^2 u)|\le C{\varepsilon}^2{\left\langle t+r \right\rangle}^{-2+\delta}{\left\langle c_jt-r \right\rangle}^{-1-\delta}.$$ Noting that we have $${\left\langle c_kt-r \right\rangle}{\left\langle c_lt-r \right\rangle}\ge C {\left\langle t+r \right\rangle}\min\{{\left\langle c_k t-r \right\rangle}, {\left\langle c_l t-r \right\rangle}\}$$ for $c_k\ne c_l$, we obtain from that $$\begin{aligned} r|R_j^{\rm I}({\partial}u, {\partial}^2 u)|\le & C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-1} \sum_{k\ne l} {\left\langle c_kt-r \right\rangle}^{-1}{\left\langle c_lt-r \right\rangle}^{-1} \nonumber\\ \le & C{\varepsilon}^{2}{\left\langle t+r \right\rangle}^{-2} \sum_{k=1}^N {\left\langle c_kt-r \right\rangle}^{-1}. \label{Est03}\end{aligned}$$ By we also have $$\label{Est04} r|R_j^{\rm II}({\partial}u, {\partial}^2 u)|\le C{\varepsilon}^2 \sum_{k\ne j} {\left\langle t+r \right\rangle}^{-1} {\left\langle c_kt-r \right\rangle}^{-2}.$$ Now , , , and lead to $$\begin{aligned} |G_j(t,r,\omega)|\le & C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-2+\delta} {\left\langle c_jt-r \right\rangle}^{-\delta} \nonumber\\ & {}+C{\varepsilon}^2\sum_{k\ne j} {\left\langle t+r \right\rangle}^{-1} {\left\langle c_k t-r \right\rangle}^{-2}. \label{Est05}\end{aligned}$$ We define $$\label{Construct01} {V}_j(\sigma,\omega)={v}_j\bigl( t_{0,j}(\sigma),r_{0,j}(\sigma),\omega \bigr)+\int_{t_{0,j}(\sigma)}^\infty G_j(s, c_js+\sigma, \omega) ds,$$ where $t_{0,j}(\sigma)$ is defined by , and $r_{0,j}(\sigma)=c_jt_{0,j}(\sigma)+\sigma$. Since we have and , Lemma \[KeyLemma\] implies $$\begin{aligned} |{v}_j(t, r, \omega)-{V}_j(r-c_jt,\omega)|\le & C{\varepsilon}^2 \left({\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_jt-r \right\rangle}^{-\delta} +{\left\langle t+r \right\rangle}^{-1}\right) \nonumber\\ \le & C{\varepsilon}^2{\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_jt-r \right\rangle}^{-\delta} \label{Est06}\end{aligned}$$ for $(t,r)\in \Lambda_0$ and $\omega\in {{\mathbb S}}^2$. By , , and , we get $$\begin{aligned} |{v}_j(t, r, \omega)| \le & C\left(r|{\partial}u_j^1(t,r\omega)|+|u_j^1(t,r\omega)|\right) \nonumber\\ \le & C{\varepsilon}^2 \left({\left\langle c_jt-r \right\rangle}^{-1}+{\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_jt-r \right\rangle}^{-\delta}\right) \nonumber\\ \le & C{\varepsilon}^2 {\left\langle c_j t-r \right\rangle}^{-1}.\end{aligned}$$ Hence, putting $r=c_jt+\sigma$ in , we get $$\begin{aligned} |V_j(\sigma, \omega)|\le & |v_j(t, c_jt+\sigma, \omega)|+C{\varepsilon}^2{\left\langle (c_j+1)t+\sigma \right\rangle}^{-1+\delta} {\left\langle \sigma \right\rangle}^{-\delta}\nonumber\\ \le & C{\varepsilon}^2\bigl({\left\langle \sigma \right\rangle}^{-1}+{\left\langle (c_j+1)t+\sigma \right\rangle}^{-1+\delta} {\left\langle \sigma \right\rangle}^{-\delta}\bigr) \label{BlackCat}\end{aligned}$$ for all $(\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^{n-1}$ and $t\ge t_{0,j}(\sigma)$. Taking the limit in as $t\to\infty$, we obtain $$\label{Est07} |{V}_j(\sigma, \omega)|\le C{\varepsilon}^2 {\left\langle \sigma \right\rangle}^{-1},\quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^2.$$ Now we define ${P}_j^1(\sigma,\omega)={\varepsilon}^{-1}{V}_j(\sigma, \omega)$. Then, recalling the definition of ${v}_j$, we obtain from , , and that $$\left|r{\partial}{u}_j^1(t,x)-\vec{\omega}_{c_j}(x){v}_j(t, r, \omega)\right| \le C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_jt-r \right\rangle}^{-\delta},\quad (t,x)\in \Lambda.$$ Therefore leads to $$\label{NonlinApprox01} \left|r{\partial}{u}_j^1(t,x)-{\varepsilon}\vec{\omega}_{c_j}(x){P}_j^1(r-c_jt,\omega)\right| \le C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_jt-r \right\rangle}^{-\delta}$$ for $(t,x)\in \Lambda$. Finally we define $$P_j(\sigma, \omega)={\partial}_\sigma U_j^0(\sigma, \omega) {}+{P}_j^1(\sigma,\omega). \label{FiveTwentyFive}$$ Then and yield for $(t,x)\in \Lambda$. Noting that and the definition of $P_j^1$ lead to $$\label{PseudoTwoFive} |P_j^1(\sigma, \omega)|\le C{\varepsilon}{\left\langle \sigma \right\rangle}^{-1},\quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^{2},$$ and recalling , we immediately obtain by . Now what is left to prove is for $(t,x)\in \Lambda^{\rm c}$. Since we have either $t\le 2c_1^{-1}$ or $r\le c_1t/2(\le c_jt/2)$ in $\Lambda^{\rm c}$, we get ${\left\langle c_jt-r \right\rangle}^{-1}\le C{\left\langle t+r \right\rangle}^{-1}$. Therefore, by we get $$\label{KL01} |r{\partial}u_j(t,x)|\le C{\varepsilon}{\left\langle c_jt-r \right\rangle}^{-1}\le C{\varepsilon}{\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_jt-r \right\rangle}^{-\delta}$$ for $(t,x)\in \Lambda^{\rm c}$. It follows from , , and that $$|P_j(\sigma, \omega)|\le C{\left\langle \sigma \right\rangle}^{-1},\quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^2,$$ which immediately yields $$\label{KL02} |P_j(r-c_jt,\omega)|\le C{\left\langle c_jt-r \right\rangle}^{-1}\le C{\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_jt-r \right\rangle}^{-\delta}$$ for $(t,x)\in \Lambda^{\rm c}$. From and we obtain $$|r{\partial}u_j(t, x)-{\varepsilon}\vec{\omega}_{c_j}(x)P_j(r-c_jt,\omega)|\le C{\varepsilon}{\left\langle t+r \right\rangle}^{-1+\delta}{\left\langle c_jt-r \right\rangle}^{-\delta}, \quad (t, x)\in \Lambda^{\rm c},$$ which is for $(t,x)\in \Lambda^{\rm c}$. This completes the proof. Suppose that $R^{\rm II}$ has the null structure, and fix $\rho\in (1/2, 1)$. Then following the proof of the global existence theorem in [@Kat05] (see Proposition 5.2 and its proof in [@Kat05] specifically), we have $$\begin{aligned} |{u}_k(t,x)|_2+{\varepsilon}^{-1}|{u}_k^1(t,x)|_2\le & C{\varepsilon}{\left\langle t+r \right\rangle}^{-1}{\left\langle c_kt-r \right\rangle}^{-1}, \label{Basic02a} \\ |{\partial}{u}_k(t,x)|_2+{\varepsilon}^{-1}|{\partial}{u}_k^1(t,x)|_2\le & C{\varepsilon}{\left\langle r \right\rangle}^{-1}{\left\langle c_kt-r \right\rangle}^{-1-\rho}, \label{Basic02b00}\end{aligned}$$ which are better than and . Using decay estimates in [@KatYok06] (see Lemmas 3.2 and 6.1 in [@KatYok06]), we can further improve in the decay rate and we get $$|{\partial}{u}_k(t,x)|_1+{\varepsilon}^{-1}|{\partial}{u}_k^1(t,x)|_1\le C{\varepsilon}{\left\langle r \right\rangle}^{-1}{\left\langle c_kt-r \right\rangle}^{-2}. \label{Basic02b}$$ Indeed, similarly to and , we obtain from and that $$\begin{aligned} r |N_j|_1 \le & C{\varepsilon}^2\left({\left\langle t+r \right\rangle}^{-1}{\left\langle r \right\rangle}^{-1}{\left\langle c_jt-r \right\rangle}^{-2-\rho}+{\left\langle r \right\rangle}^{-2}{\left\langle c_jt-r \right\rangle}^{-1-2\rho}\right)\\ \le & C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-2}w_-(t,r)^{-1-2\rho},\\ r|R_j^{\rm I}|_1\le & C{\varepsilon}^2 {\left\langle r \right\rangle}^{-1}\sum_{k\ne l} {\left\langle c_kt-r \right\rangle}^{-1-\rho}{\left\langle c_lt-r \right\rangle}^{-1-\rho}\\ \le & C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-2} w_-(t,r)^{-1-2\rho}\end{aligned}$$ at $(t,x)\in (0,\infty)\times {{\mathbb R}}^3$, where $w_-(t,r)=\min_{0\le k\le N} {\left\langle c_kt-r \right\rangle}$ with $c_0=0$. Here we have used the estimates ${\left\langle r \right\rangle}{\left\langle c_jt-r \right\rangle}\ge C {\left\langle t+r \right\rangle}w_-(t,r)$ and $${\left\langle r \right\rangle}{\left\langle c_kt-r \right\rangle}{\left\langle c_lt-r \right\rangle}\ge C{\left\langle t+r \right\rangle}^2w_-(t,r),\quad k\ne l.$$ Since $R^{\rm II}$ has the null structure, we see that $R^{\rm II}$ enjoys the same estimate as $N_j$. Now we have proved that $$r|F({\partial}u, {\partial}^2 u)|_1\le C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-2}w_-(t,r)^{-1-2\rho}, \label{Kiririn}$$ and we obtain from the following estimate, which comes from Lemmas 3.2 and 6.1 in [@KatYok06] (especially see the estimates (3.7), (6.1) and (6.2) in [@KatYok06]): For $c, \kappa, \mu>0$, it holds that $$\begin{aligned} & {\left\langle r \right\rangle}{\left\langle ct-r \right\rangle}^{1+\kappa}|{\partial}\phi(t,x)|\\ & \qquad \le C\sup_{y\in {{\mathbb R}}^3} {\left\langle y \right\rangle}^{2+\kappa} |{\partial}\phi(\tau,y)|_1\bigr|_{\tau=0}\\ & \qquad\quad {}+C\sup_{(\tau,y)\in[0,t]\times {{\mathbb R}}^3} |y|{\left\langle \tau+|y| \right\rangle}^{1+\kappa}w_-(\tau, |y|)^{1+\mu} |({\square}_c\phi)(\tau,y)|_1.\end{aligned}$$ Let $(t,r)\in \Lambda_0$ and $\omega\in {{\mathbb S}}^2$ for a while. Recall the proof of (1). We use and , instead of (or ) and . Then , , and are replaced by $$\begin{aligned} |r^{-1}\Delta_\omega{u}_j^1(t, r\omega)|\le & C{\varepsilon}^2{\left\langle t+r \right\rangle}^{-2}{\left\langle c_jt-r \right\rangle}^{-1},\\ r|N_j({\partial}u, {\partial}^2 u)|\le & C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-2}{\left\langle c_jt-r \right\rangle}^{-3},\\ r|R_j^{\rm I}({\partial}u, {\partial}^2 u)|\le & C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-3}\sum_{k=1}^N{\left\langle c_kt-r \right\rangle}^{-2},\end{aligned}$$ respectively. Since $R^{\rm II}$ has the null structure, we can use Lemma \[NullEst\] to get $$r|R_j^{\rm II}({\partial}u, {\partial}^2 u)|\le C{\varepsilon}^2\sum_{k\ne j}{\left\langle t+r \right\rangle}^{-2}{\left\langle c_kt-r \right\rangle}^{-3},$$ instead of . These estimates lead to $$|G_j(t,r,\omega)|\le C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-2}\left( {\left\langle c_jt-r \right\rangle}^{-1} {}+\sum_{k\ne j}{\left\langle c_kt-r \right\rangle}^{-3}\right)$$ for $(t,r)\in \Lambda_0$ and $\omega\in {{\mathbb S}}^2$. Now going similar lines to through , we can construct ${P}_j^1$ satisfying $$\label{NonlinApprox02} \left|r{\partial}{u}_j^1(t, x)-{\varepsilon}\vec{\omega}_{c_j}(x) {P}_j^1(r-c_jt, \omega)\right|\le C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-1}{\left\langle c_jt-r \right\rangle}^{-1}$$ for $(t,x)\in \Lambda$, and $$\label{Est31} |{P}_j^1(\sigma, \omega)|\le C{\varepsilon}{\left\langle \sigma \right\rangle}^{-2},\quad (\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^2.$$ We define $$\label{DefUi} {U}_j^1(\sigma, \omega):=-\int_\sigma^\infty {P}_j^1(\lambda,\omega) d\lambda$$ so that we have ${\partial}_\sigma {U}_j^1(\sigma, \omega)={P}_j^1(\sigma, \omega)$. Note that the right-hand side of is finite because of , and that $U_j^1(\sigma, \omega)\to 0$ as $\sigma\to\infty$. It follows from and that $$\label{NonlinApprox03} \left|{\partial}_r\bigl(r {u}_j^1(t, r\omega)\bigr)-{\varepsilon}({\partial}_\sigma {U}_j^1)(r-c_jt, \omega)\right| \le C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-1}{\left\langle c_jt-r \right\rangle}^{-1}$$ for $(t,x)\in \Lambda$. Since implies $ru_j^1(t,r\omega)\to 0$ as $r\to \infty$, from we get $$\begin{aligned} |r{u}_j^1(t,r\omega)-{\varepsilon}{U}_j^1(r-c_jt, \omega)|\le & \left|-\int_r^\infty {\partial}_\lambda\bigl(\lambda {u}_j^1(t,\lambda\omega)-{\varepsilon}{U}_j^1 (\lambda-c_jt, \omega)\bigr)d\lambda \right| \nonumber\\ \le & C{\varepsilon}^2\int_r^\infty (1+t+\lambda)^{-1}(1+|c_jt-\lambda|)^{-1}d\lambda. \label{Lost}\end{aligned}$$ If $r\ge 3c_jt/2$, then we have $$\int_r^\infty (1+t+\lambda)^{-1}(1+|c_jt-\lambda|)^{-1}d\lambda\le C\int_r^\infty(1+t+\lambda)^{-2}d\lambda \le C(1+t+r)^{-1}.$$ For $r$ satisfying $1\le c_1t/2\le r\le 3c_jt/2$, we get $$\begin{aligned} \int_r^\infty (1+t+\lambda)^{-1}(1+|c_jt-\lambda|)^{-1}d\lambda \le & C(1+t)^{-1}\int_{c_1t/2}^{3c_jt/2}(1+|c_jt-\lambda|)^{-1}d\lambda\\ & +C\int_{3c_jt/2}^\infty(1+t+\lambda)^{-2}d\lambda \nonumber\\ \le & C(1+t+r)^{-1}\log(2+t). \nonumber\end{aligned}$$ Now it follows from that $$\label{NonlinApprox04} \left|r {u}_j^1(t, x)-{\varepsilon}{U}_j^1(r-c_jt, \omega)\right| \le C{\varepsilon}^2 {\left\langle t+r \right\rangle}^{-1}\log(2+t),\quad (t,x)\in \Lambda.$$ For $(\sigma, \omega)\in {{\mathbb R}}\times {{\mathbb S}}^2$, by and we get $$|{U}_j^1(\sigma, \omega)|\le C{\varepsilon}{\left\langle (c_j+1)t+\sigma \right\rangle}^{-1}\log(2+t)+C{\varepsilon}{\left\langle \sigma \right\rangle}^{-1}$$ for sufficiently large $t$. Now taking the limit as $t\to\infty$, we obtain $$\label{Est32} |{U}_j^1(\sigma, \omega)|\le C{\varepsilon}{\left\langle \sigma \right\rangle}^{-1}.$$ Finally, putting $$U_j(\sigma, \omega)=U_j^0(\sigma,\omega)+{U}_j^1(\sigma,\omega),$$ we obtain and for $(t,x)\in \Lambda$ from , , and . is an immediate consequence of and . As in the proof of (1), using , , , and , as well as , we can show that and hold also for $(t,x)\in \Lambda^{\rm c}$. This completes the proof. Now we are in a position to prove Corollary \[AsympFreeEnergy\]. Let the assumptions of Corollary $\ref{AsympFreeEnergy}$ be fulfilled. Let $P_j$ for $1\le j\le N$ be from Theorem \[Main01\] (1). Recalling , we see from that $|P_j(\sigma, \omega)|\le C {\left\langle \sigma \right\rangle}^{-1}$, which implies $P_j\in L^2({{\mathbb R}}\times {{\mathbb S}}^2)$. If we put $P_j^\sharp(t,x)=|x|^{-1}P_j(|x|-c_jt, |x|^{-1}x)$, it follows from that $$\begin{aligned} & \bigl\|{\partial}u_j(t,\cdot)-{\varepsilon}\vec{\omega}_{c_j}(\cdot)P_j^\sharp(t,\cdot) \bigr\|_{L^2({{\mathbb R}}^3)}^2\\ & \quad = \int_0^\infty \int_{{{\mathbb S}}^2} |r{\partial}u_j(t,r\omega)-{\varepsilon}\vec{\omega}_{c_j}(r\omega)P_j(r-c_jt,\omega)|^2 dS_\omega dr\\ & \quad \le C{\varepsilon}^2 \int_0^\infty {\left\langle t+r \right\rangle}^{-2+2\delta}{\left\langle c_jt-r \right\rangle}^{-2\delta} dr \le C{\varepsilon}^2(1+t)^{-1} \to 0,\quad t\to\infty.\end{aligned}$$ Now Theorem \[AFSE\] (with $n=3$, $c=c_j$, $v=u_j$, and $V={\varepsilon}P_j$) implies that each $u_j$ is asymptotically free in the energy sense. This completes the proof. Proof of Theorem \[NotFree\] {#P02} ============================ Suppose that ${\varepsilon}$ is sufficiently small, and let $u$ be the global solution to with initial data $u={\varepsilon}f$ and ${\partial}_t u={\varepsilon}g$ at $t=0$. Since the last half of follows from $\eqref{Basic01a}$ for any $f, g\in C^\infty_0({{\mathbb R}}^3)$, we will prove the existence of $(f,g)$ for which the first half holds. Without loss of generality, we may assume $A_1>0$ and $A_2\ge 0$. We may also assume $c_1=1<c=c_2$. We suppose that $f\equiv 0$, and that $g=(g_1,g_2)$ is radially symmetric, i.e., $g_j(x)={g}_j^*(|x|)$ with some function ${g}_j^*$ for $j=1,2$. Then $u=(u_1,u_2)$ is also radially symmetric in $x$-variable, i.e., $u_j(t,x)={u}_j^*(t, |x|)$ with some ${u}_j^*(t,r)$ for $j=1,2$. For $r\in {{\mathbb R}}$, we put $$h_j(r):=\frac{r}{2}{g}_j^*(|r|),\ v_j(t,r):=r {u}_j^*(t,|r|),\quad j=1,2.$$ From we obtain $$\label{Exp01} v_j(t,r)=\frac{{\varepsilon}}{c_j}\int_{r-c_jt}^{r+c_jt} h_j(\lambda)d\lambda {}+\frac{1}{c_j}\int_0^t\left(\int_{r-c_j(t-\tau)}^{r+c_j(t-\tau)}G_j(\tau, \lambda)d\lambda\right)d\tau$$ for $j=1,2$, where $$\begin{aligned} G_1(t,r)=& \frac{A_1r}{2} \bigl({\partial}_t {u}_2^*(t, |r|)\bigr)^2 =\frac{A_1}{2r}\bigl({\partial}_t v_2(t,r)\bigr)^2,\\ G_2(t,r)=& \frac{A_2r}{2} \bigl({\partial}_t {u}_1^*(t, |r|)\bigr)^2 =\frac{A_2}{2r}\bigl({\partial}_t v_1(t,r)\bigr)^2.\end{aligned}$$ Assume that $g_2^*$ is a nonnegative function. Since we have $$\begin{aligned} {\partial}_tv_2(t,r)=& {\varepsilon}\left\{h_2(r+c t)+h_2(r-c t)\right\} \\ & {}+\int_0^t\left\{G_2\bigl(\tau,r+c(t-\tau)\bigr)+G_2\bigl(\tau,r-c(t-\tau)\bigr)\right\}d\tau,\end{aligned}$$ we get $$\label{Exp02} {\partial}_t v_2(t, r)\ge {\varepsilon}h_2(r-ct)(\ge 0),$$ provided that $r-ct\ge 0$. We fix $M>0$, and assume that $$C_0:=\int_{0}^M |h_2(\lambda)|^2d\lambda>0.$$ We suppose that $0\le \sigma=r-t\le M$ in what follows. We put $$\tau_0=\frac{r-t}{c-1}=\frac{\sigma}{c-1},\quad \tau_1=\frac{r+t-M}{c+1}=\frac{2t+\sigma-M}{c+1},$$ so that we have $c\tau_0=r-t+\tau_0$ and $c\tau_1+M=r+t-\tau_1$. Let $2t\ge (c-1)^{-1}(c+1)M$ hold, so that we have $\tau_1\ge \tau_0$. Then it is easy to see that ${\mathcal E}\subset {\mathcal D}$, where $$\begin{aligned} {\mathcal D}=&\{(\tau,\lambda);\, 0\le \tau\le t,\, r-t+\tau\le \lambda\le r+t-\tau\},\\ {\mathcal E}=&\{(\tau,\lambda);\, \tau_0\le \tau\le \tau_1,\, c\tau\le \lambda\le c\tau+M\}.\end{aligned}$$ Therefore, noting that we have $G_1(t,r)\ge 0$ for $r>0$, we get $$\begin{aligned} \iint_{{\mathcal D}} G_1(\tau, \lambda) d\tau d\lambda \ge & \iint_{{\mathcal E}} G_1(\tau, \lambda) d\tau d\lambda.\end{aligned}$$ From the definition of $G_1$ and , we obtain $$\begin{aligned} \iint_{{\mathcal E}} G_1(\tau, \lambda) d\tau d\lambda \ge & \frac{A_1{\varepsilon}^2}{2}\int_{\tau_0}^{\tau_1} \left(\int_{c\tau}^{c\tau+M} \frac{|h_2(\lambda-c\tau)|^2}{\lambda}d\lambda\right)d\tau\nonumber\\ \ge & \frac{A_1{\varepsilon}^2}{2}\int_{\tau_0}^{\tau_1} \frac{1}{c\tau+M} \left(\int_0^M |h_2(\lambda)|^2d\lambda\right)d\tau \nonumber\\ =& \frac{C_0A_1{\varepsilon}^2}{2c}\log \frac{c\tau_1+M}{c\tau_0+M} \ge \frac{C_0A_1{\varepsilon}^2}{2c}\log \frac{(c-1)(2ct+M)}{(c+1)(2c-1)M}. \label{Exp05}\end{aligned}$$ Now we choose a nonnegative function $g_1^*$ satisfying $\int_{M}^{M+1} h_1(\lambda)d\lambda\ge 1$. Then we get $$\label{Exp06} \int_{r-t}^{r+t} h_1(\lambda)d\lambda\ge 1$$ for $(t,r)$ with $0\le r-t\le M$ and $t \ge M+1$ From with $j=1$, , and , we obtain $$v_1(t,r)\ge C\bigl({\varepsilon}+{\varepsilon}^2\log(2+t)\bigr)$$ for $0\le r-t\le M$ and $t\gg 1$. This completes the proof. Acknowledgments {#acknowledgments .unnumbered} =============== This work is partially supported by Grant-in-Aid for Scientific Research (C) (No. 20540211 and No. 23540241), JSPS. [99]{} R. Agemi, [*Global existence of nonlinear elastic waves*]{}, Invent. Math. [**142**]{} (2000), 225–250. D. Christodoulou, [*Global solutions of nonlinear hyperbolic equations for small initial data*]{}, [Comm. Pure Appl. Math.]{} [**39**]{} (1986), [267–282]{}. F. G. Friedlander, [*On the radiation field of pulse solutions of the wave equation*]{}, Proc. Roy. Soc. A. [**269**]{} (1962), 53–65. K. Hidano, [*The global existence theorem for quasi-linear wave equations with multiple speeds*]{}, Hokkaido Math. J. [**33**]{} (2004), 607 – 636. L. Hörmander, [The Analysis of Linear Partial Differential Operators I, Distribution Theory and Fourier Analysis, Second Edition]{}, Springer-Verlag, Berlin, 1990. L. Hörmander, [Lectures on Nonlinear Hyperbolic Differential Equations]{}, Springer-Verlag, Berlin, 1997. S. Katayama, [*Global and almost–global existence for systems of nonlinear wave equations with different propagation speeds*]{}, Diff. Integral Eqs. [**17**]{} (2004), 1043–1078. S. Katayama, [*Global existence for systems of wave equations with nonresonant nonlinearities and null forms*]{}, J. Differential Equations [**209**]{} (2005), 140–171. S. Katayama, [*Global existence for coupled systems of nonlinear wave and Klein-Gordon equations in three space dimensions*]{}, Math. Z. [**270**]{} (2012), 487–513. S. Katayama, [*Asymptotic pointwise behavior for systems of semilinear wave equations in three space dimensions*]{}, J. Hyperbolic Differ. Equ. [**9**]{} (2012), 263–323. S. Katayama and H. Kubo, [*Decay estimates of a tangential derivative to the light cone for the wave equation and their application*]{}, SIAM J. Math. Anal. [**39**]{} (2008), 1851–1862. S. Katayama and H. Kubo, [*The rate of convergence to the asymptotics for the wave equation in an exterior domain*]{}, Funkcial. Ekvac. [**53**]{} (2010), 331–358. S. Katayama, T. Matoba and H. Sunagawa, [*Semilinear hyperbolic systems violating the null condition*]{}, preprint (arXiv: 1206.0066 \[math.AP\]). S. Katayama, D. Murotani and H. Sunagawa, [*The energy decay and asymptotics for a class of semilinear wave equations in two space dimensions*]{}, J. Evol. Equ. [**12**]{} (2012), 891–916. S. Katayama and K. Yokoyama, [*Global small amplitude solutions to systems of nonlinear wave equations with multiple speeds*]{}, Osaka J. Math. [**43**]{} (2006), 283–326. S. Klainerman, [*The null condition and global existence to nonlinear wave equations*]{}, in: [Nonlinear Systems of Partial Differential Equations in Applied Mathematics, Part 1]{}, Lectures in Appl. Math. [**23**]{}, Amer. Math. Soc., Providence, 1986, pp. 293–326. S. Klainerman and T. C. Sideris, [*On almost global existence for nonrelativistic wave equations in $3$D*]{}, Comm. Pure Appl. Math. [**49**]{} (1996), 307–321. K. Kubota and K. Yokoyama, [*Global existence of classical solutions to systems of nonlinear wave equations with different speeds of propagation*]{}, Japan. J. Math. (N.S.) [**27**]{} (2001), 113–202. P. D. Lax and R. S. Phillips, [*Scattering theory for the acoustic equation in an even number of space dimensions*]{}, Indiana Univ. Math. J. [**22**]{} (1972/73), 101–134. P. D. Lax and R. S. Phillips, Scattering Theory, Revised Edition, Academic Press, New York, 1989. R. B. Melrose, [ *Singularities and energy decay in acoustical scattering*]{}, Duke Math. J. [**46**]{} (1979), 43–59. J. Metcalfe, M. Nakamura and C. D. Sogge, [*Global existence of quasilinear, nonrelativistic wave equations satisfying the null condition*]{}, Japan. J. Math. (N.S.) [**31**]{} (2005), 391–472. T. C. Sideris, [*Nonresonance and global existence of prestressed nonlinear elastic waves*]{}, Ann. of Math. (2) [**151**]{} (2000), 849–874. T. C. Sideris and B. Thomases, [*Global existence for three-dimensional incompressible isotropic elastodynamics via the incompressible limit*]{}, Comm. Pure Appl. Math. [**58**]{} (2005), 750–788. T. C. Sideris and B. Thomases, [*Local energy decay for solutions of multi-dimensional isotropic symmetric hyperbolic systems*]{}, J. Hyperbolic Differ. Equ. [**3**]{} (2006), 673–690. T. C. Sideris and B. Thomases, [*Global existence for three-dimensional incompressible isotropic elastodynamics*]{}, Comm. Pure Appl. Math. [**60**]{} (2007), 1707–1730. T. C. Sideris and Shun-Yi Tu, [*Global existence for systems of nonlinear wave equations in 3D with multiple speeds*]{}, SIAM J. Math. Anal. [**33**]{} (2001), 477 –488. C. D. Sogge, [*Global existence for nonlinear wave equations with multiple speeds*]{}, in: W. Beckner [*et. al.*]{} (Eds.), Harmonic Analysis at Mount Holyoke, Contemp. Math. [**320**]{}, Amer. Math. Soc., Providence, RI, 2003, pp. 353 – 366. K. Yokoyama, [*Global existence of classical solutions to systems of wave equations with critical nonlinearity in three space dimensions*]{}, J. Math. Soc. Japan [**52**]{} (2000), 609–632.

ArXiv

--- abstract: 'We show how to construct covariant amplitudes for processes involving higher spins in this paper. First we give the explicit expressions of Rarita-Schwinger wave functions and propagators for bosons with spins, then kinematic singularity free 3-leg effective vertexes are derived and given in a list. Equivalence relations are worked out to get these independent vertexes. Constraints of space reflection symmetry and boson symmetry are considered and shown in a explicit way. Some helicity amplitudes for two-body decays in center of frame are calculated. Finally the covariant helicity amplitudes for the process $a_1\rightarrow \pi^+\pi^+\pi^-$ are constructed to illustrate how to include background (1PI) amplitudes. Both background amplitudes and resonance amplitudes are needed to give reliable descriptions to high energy reactions.' address: | CCAST(World Lab), P. O. Box 8730, Beijing, 100080, P. R. China\ and Center for Fundamental Physics, USTC, Hefei, Anhui, 230026, P. R. China[^1] author: - 'Jie-Jie Zhu[^2], Mu-Lin Yan' title: Covariant amplitudes for mesons --- Introduction ============ Model independent amplitudes are needed to analyze high energy experimental data. Such amplitudes are usually written in terms of helicity formalism first proposed by Jacob and Wick [@JacobH59; @ChouH58; @BermanH65; @WickH62], or tensor formalism by Zemach [@ZemachTA64; @ZemachTA65], in a non-covariant form. Recently Chung [@ChungTA93] and Filippini [*et al.*]{} [@FilippiniTA95] emphasized on the importance of covariance to get reliable results. Chung gave some examples of two-body decays with their amplitudes calculated in center of mass frame in his works [@ChungTA93; @ChungSF71] on covariant tensor formalism. The case of spin-$J$ $\rightarrow$ spin-$j$ + spin-$0$ with $J, j \leq 2$ has been discussed in detail in Ref. [@FilippiniTA95]. We construct covariant amplitudes in a view of S-matrix approach. We call a S-matrix (maybe off shell) after stripping off external lines (wave functions) effective vertexes. Effective vertexes are related to Green functions by LSZ reduction formulations [@LSZ55; @BogoliubovAFT], which can be divide into one-particle irreducible (1PI) parts and one-particle reducible parts. The former ones are often called backgrounds and the later resonances. One should repeat the process to divide those sub-vertexes connected by (full) propagators in resonance parts, until arrives at 3-leg vertexes that can not be separated. These effective vertexes (with three or more legs) should be constructed from the four-momenta of outer legs and isotropic tensors of Lorentz group. We use wave functions satisfying Rarita-Schwinger conditions [@RaritaWF41] in this paper. This specific choice of wave functions will not introduce any model dependence, since we can change our results into any other representations via basis transformations. Information on the structure of particles is contained in effective vertexes. The general form of an effective vertex for bosons are tensors to be constructed from $p_i^\mu$ ($i=1,2,\cdots.)$ and $g^{\mu\nu}, \varepsilon^{\alpha\beta\gamma\delta}$. Following the assumption of maximal analyticity [@StappAF62; @HaraKS64; @StappKS67], such vertex functions are free of kinematic singularities (K. S.) [@HaraKS64; @WangKS66; @TannoudjiKS68; @BardeenKS68; @BrownKS71]. However, one should be careful when writing down independent effective vertexes. Since there are redundant components to be removed by Rarita-Schwinger conditions, some seemingly independent terms are related by equivalence relations [@ScadronVF68]. We work out equivalence relations and give a general list of 3-leg effective vertexes. Some special cases of them have already been given by Scadron [@ScadronVF68]. Additional symmetries give constraints on the form factors in effective vertexes. Space reflection symmetry demands effective vertexes being tensors or pseudo-tensors, depending on spin-parities of their outer legs. The ratio of form factors in tensor parts and those in pseudo-tensor parts can be taken as a parameter of parity violation. For 3-leg effective vertexes involving two spin-$0$ particles, some processes are forbidden. If an effective vertex is connected to two identical bosons, it must satisfy boson symmetry. There are kinematic zeros in form factors that should be factored out. When both of the two bosons are on shell, the “anti-symmetric” parts of the 3-leg effective vertex vanish. For the case of one particle is on shell while another is off shell, the contribution from “anti-symmetric” parts is not independent from background terms, since the form factors in these parts contribute a factor which eliminates the denominator of the propagator for the off shell particle. We give a list of 3-leg effective vertexes, with all these cases considered. Covariant amplitudes in different reference frame are related by Wigner rotations [@WignerLG39]. We calculate the amplitudes for some two-body decays in center of mass (CM) frame. The relation between amplitudes in laboratory frame and those in CM frame is derived. Background amplitudes should not be neglected in order to give a full description to a reaction, and to get reliable information from data [@Zhua23pi]. 1PI amplitudes will not give flat distributions if there are particles with non-zero spins. The covariant helicity amplitudes for $a_1\rightarrow \pi^+\pi^+\pi^-$ are constructed as a demonstration. Our present work is partly based on Ref.[@ZhuPHD]. In Sec. \[sec:wf\] we give a brief introduction to wave functions and propagators. 3-leg effective vertexes are listed in Sec. \[sec:3v\]. Sec. \[sec:sr\] and Sec. \[sec:bs\] are devoted to the constraint of space reflection symmetry and boson symmetry. Two-body decays are considered in Sec. \[sec:2bd\]. In Sec. \[sec:bk\] we discuss the process $a_1\rightarrow \pi^+\pi^+\pi^-$. Wave functions {#sec:wf} ============== Let $L(p)$ be a Lorentz transformation, $$\label{eq:lrtz} p^\mu={L^\mu}_\nu(p)k^\nu.$$ The standard momentum for a mass-$W$ particles is $(k^\mu)=(W;\ \roarrow{0})$. The space-time metric we use is $(g^{\mu\nu})={\rm diag}\{1,-1,-1,-1\}.$ Now define one-particle states as [@WeinbergQFT1] $$\label{eq:qsdef} |p,\sigma\rangle =U(L(p))|k,\sigma\rangle$$ with $U(L(p))$ the unitary representation of $L(p)$ in Hilbert space. It satisfies $$\hat{p}^\mu|p,\sigma\rangle=p^\mu|p,\sigma\rangle .$$ We can choose the orthonormal condition to be $$\langle p^{'}, \sigma^{'} | p, \sigma \rangle = (2 \pi)^3(2 p^0) \delta(\roarrow{p}^{'} - \roarrow{p}) \delta_{\sigma^{'}\sigma}.$$ Under a Lorentz transformation $\Lambda$, $$\label{eq:qswr} U(\Lambda)| p,\sigma \rangle = \sum\limits_{\sigma^{'}} D_{\sigma^{'}\sigma}(W(\Lambda,p))| \Lambda p, \sigma^{'} \rangle,$$ where $$W(\Lambda,p) \equiv L^{-1}(\Lambda p)\Lambda L(p)$$ is the Wigner rotation [@WignerLG39] and $\{D_{\sigma^{'}\sigma}\}$ furnishes a representation of $SO(3).$ If we define the Lorentz transformation in Eq. (\[eq:qsdef\]) to be a pure Lorentz boost $$\label{eq:lc} \begin{array}{rcl} L(p)&=&L(\roarrow{p})\\ &\equiv&R(\varphi,\theta,0)L_{z} (|\roarrow{p}|)R^{-1}(\varphi,\theta,0), \end{array}$$ we obtain canonical states. Here $R(\varphi,\theta,0)$ is the rotation that takes $z$-axis to the direction of $\roarrow{p}$, and the boost $L_{z}(|\roarrow{p}|)$ takes the four-momentum $(k^\mu)=(W;\ \roarrow{0})$ to $\left( \sqrt{W^2+\roarrow{p}^2};\ 0,\ 0,\ |\roarrow{p}|\right)$. For a particle of spin-$j$, $\sigma\sim(j,m)$. For canonical states, Eq. (\[eq:qswr\]) becomes $$U(\Lambda)|\roarrow{p},j,m\rangle=\sum\limits_{m^{'}} D_{m^{'} m}^j(L^{-1}(\roarrow{\Lambda}p)\Lambda L(\roarrow{p})) |\roarrow{\Lambda}p,j,m^{'}\rangle.$$ $ D_{m^{'} m}^j $ is the ordinary $D$-function. Especially, under a rotation $R$, $$\label{eq:rot1} U(R)|\roarrow{p},j,m\rangle=\sum\limits_{m^{'}} D_{m^{'} m}^j(R) |\roarrow{R}p,j,m^{'}\rangle.$$ Defining the Lorentz transformation in another way leads to helicity states [@JacobH59]: $$\label{eq:lh} \begin{array}{rcl} L(p)&=&L(\roarrow{p})R^{-1}(\varphi,\theta,0)\\ &\equiv&R(\varphi,\theta,0)L_{z}(|\roarrow{p}|). \end{array}$$ We have $$\label{eq:lrt} U(\Lambda)|\roarrow{p},j,\lambda\rangle=\sum\limits_{\lambda^{'}} D_{\lambda^{'}\lambda}^j (L^{-1}(\roarrow{\Lambda}p)\Lambda L(\roarrow{p})) |\roarrow{\Lambda}p,j,\lambda^{'}\rangle$$ and $$\label{eq:rot2} U(R)|\roarrow{p},j,\lambda\rangle= |\roarrow{R}p,j,\lambda\rangle.$$ The two kinds of definitions are related to each other, $$|\roarrow{p},j,\lambda\rangle_{helicity} =\sum\limits_m D^j_{m\lambda}(\varphi,\theta,0) |\roarrow{p},j,m\rangle_{canonical}.$$ Quantum states in terms of creation and annihilation operators read: $$|\roarrow{p}, j, \sigma\rangle= \sqrt{(2 \pi)^3 2 p^0}a^\dagger(\roarrow{p}, j, \sigma)|0\rangle,$$ where $|0\rangle$ is the vacuum state. We use “$\sigma$” to mean that the relation holds for both helicity states and canonical states. Quantum fields for spin-$j$ bosons are constructed as [@WeinbergQFT1] $$\phi_{\mu_1\mu_2...\mu_j}=\\ \int\frac{d^3 p}{\sqrt{\left(2\pi\right)^3 2 p^0}} \{ \sum\limits_\sigma e_{\mu_1\mu_2...\mu_j}(\roarrow{p}, j, \sigma) a(\roarrow{p}, j, \sigma) e^{-ip\cdot x} + \sum\limits_\sigma e^{\ast}_{\mu_1\mu_2...\mu_j}(\roarrow{p}, j, \sigma) {a^c}^\dagger (\roarrow{p}, j, \sigma)e^{ip\cdot x} \},$$ with ${a^c}^\dagger$ the annihilation operator of the antiparticle, and $$U(\Lambda,a)\phi_{\mu_1\mu_2...\mu_j}U^{-1}(\Lambda,a)= {\Lambda_{\mu_1}}^{\nu_1} {\Lambda_{\mu_2}}^{\nu_2}... {\Lambda_{\mu_j}}^{\nu_j} \phi_{\nu_1\nu_2...\nu_j}(\Lambda x +a).$$ $e_{\mu_1\mu_2...\mu_j}$ is the wave function in momentum space satisfying [@WeinbergQFT1] $$\label{eq:wltr} {\Lambda_{\mu_1}}^{\nu_1} {\Lambda_{\mu_2}}^{\nu_2}... {\Lambda_{\mu_j}}^{\nu_j} e_{\nu_1\nu_2...\nu_j}(\roarrow{p}, j, \sigma) =\sum\limits_{\sigma^{'}}D_{\sigma^{'} \sigma}(W(\Lambda,p)) e_{\mu_1\mu_2...\mu_j}(\roarrow{p}, j, \sigma^{'}),$$ so its definition is $$e_{\mu_1\mu_2...\mu_j}(\roarrow{p}, j, \sigma) = {\Lambda_{\mu_1}}^{\nu_1} {\Lambda_{\mu_2}}^{\nu_2} ... {\Lambda_{\mu_j}}^{\nu_j} e_{\mu_1\mu_2...\mu_j}(\roarrow{0}, j, \sigma).$$ From the following infinitesimal generators of the Lorentz group $$\begin{array}{ccc} \left({{J_1}^\mu}_\nu\right)= \left( \begin{array}{cccc} 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0\\ 0 & 0 & 0 & -i\\ 0 & 0 & i & 0 \end{array} \right), & \left({{J_2}^\mu}_\nu\right)= \left( \begin{array}{cccc} 0 & 0 & 0 & 0\\ 0 & 0 & 0 & i\\ 0 & 0 & 0 & 0\\ 0 & -i & 0 & 0 \end{array} \right), & \left({{J_3}^\mu}_\nu\right)= \left( \begin{array}{cccc} 0 & 0 & 0 & 0\\ 0 & 0 & -i & 0\\ 0 & i & 0 & 0\\ 0 & 0 & 0 & 0 \end{array} \right), \\ \left({{K_1}^\mu}_\nu\right)= \left( \begin{array}{cccc} 0 & i & 0 & 0\\ i & 0 & 0 & 0\\ 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 \end{array} \right), & \left({{K_2}^\mu}_\nu\right)= \left( \begin{array}{cccc} 0 & 0 & i & 0\\ 0 & 0 & 0 & 0\\ i & 0 & 0 & 0\\ 0 & 0 & 0 & 0 \end{array} \right), & \left({{K_3}^\mu}_\nu\right)= \left( \begin{array}{cccc} 0 & 0 & 0 & i\\ 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0\\ i & 0 & 0 & 0 \end{array} \right), \end{array}$$ one can obtain $$\left({L_z\left(|\roarrow{p}|\right)^\mu}_\nu\right) \equiv e^{-i \alpha K_3} = \left( \begin{array}{cccc} \cosh \alpha & 0 & 0 & \sinh \alpha \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ \sinh \alpha & 0 & 0 & \cosh \alpha \end{array} \right) = \left( \begin{array}{cccc} E/W & 0 & 0 & |\roarrow{p}|/W \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ |\roarrow{p}|/W & 0 & 0 & E/W \end{array} \right),$$ $$\left(R\left(\varphi, \theta, 0\right){^\mu}_\nu \right) \equiv \left({ \left( e^{-i \varphi J_3}e^{-i \theta J_2} \right)^\mu}_\nu \right) = \left( \begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & \cos \theta \cos \varphi & -\sin \varphi & \sin \theta \cos \varphi \\ 0 & \cos \theta \sin \varphi & \cos \varphi & \sin \theta \sin \varphi \\ 0 & - \sin \theta & 0 & \cos \theta \end{array} \right).$$ Choose wave functions at rest frame to be $$\begin{aligned} \left(e^\mu(\roarrow{0},0)\right) & = & \left( \begin{array}{c} 0\\0\\0\\1 \end{array} \right), \\ \left(e^\mu(\roarrow{0},\pm 1)\right) & = & \mp \frac{1}{\sqrt{2}} \left( \begin{array}{c} 0\\1\\ \pm i \\ 0 \end{array} \right), \end{aligned}$$ we get the familiar explicit expressions of spin-$1$ canonical wave functions($E$ is the energy of the particle) $$\label{eq:wfc} \begin{array}{rcl} \left(e_c^\mu(\roarrow{p},0)\right) & = & \left( \begin{array}{c} \frac{|\roarrow{p}|}{W}\cos\theta\\ \frac{1}{2}\left(\frac{E}{W}-1\right) \sin 2\theta\cos\varphi\\ \frac{1}{2}\left(\frac{E}{W}-1\right) \sin 2\theta\sin\varphi\\ \frac{1}{2}\left(\frac{E}{W}-1\right)(1+\cos 2\theta)+1 \end{array} \right),\\ \left(e_c^\mu(\roarrow{p},\pm 1)\right) & = & \mp\frac{1}{\sqrt{2}}\left( \begin{array}{c} \frac{|\roarrow{p}|}{W}\sin\theta e^{\pm i\varphi}\\ \left(\frac{E}{W}-1\right)\sin^2\theta\cos\varphi e^{\pm i\varphi}+1\\ \left(\frac{E}{W}-1\right)\sin^2\theta\sin\varphi e^{\pm i\varphi}\pm 1\\ \left(\frac{E}{W}-1\right)\cos\theta\sin\theta e^{\pm i\varphi} \end{array} \right) \end{array}$$ and spin-$1$ helicity wave functions $$\label{eq:wfh} \begin{array}{rcl} \left(e_h^\mu(\roarrow{p},0)\right) & = & \left( \begin{array}{c} \frac{|\roarrow{p}|}{W}\\ \frac{E}{W}\sin \theta\cos\varphi\\ \frac{E}{W}\sin \theta\sin\varphi\\ \frac{E}{W}\cos \theta \end{array} \right),\\ \left(e_h^\mu(\roarrow{p},\pm 1)\right) & = & \frac{1}{\sqrt{2}}\left( \begin{array}{c} 0\\ \mp\cos\theta\cos\varphi +i\sin\varphi\\ \mp\cos\theta\sin\varphi -i\cos\varphi\\ \pm\sin\theta \end{array} \right). \end{array}$$ Wave functions for higher integral spin particles can be defined recurrently by $$\label{eq:wfj} e_{\mu_1\mu_2\cdots\mu_j}(\roarrow{p},j,\sigma) =\sum\limits_{\sigma^{'}_{j-1},\sigma_j} (j-1,\sigma^{'}_{j-1};1,\sigma_j|j,\sigma) e_{\mu_1\mu_2\cdots\mu_{j-1}}(\roarrow{p},j-1,\sigma_{j-1}^{'}) e_{\mu_j}(\roarrow{p},\sigma_j).$$ Using the following C-G coefficient relation $$\begin{array}{rl} & \sum\limits_{\sigma^{'}_3, \sigma^{'}_4, \cdots, \sigma^{'}_n} (j_1,\sigma_1;j_2,\sigma_2|j_1+j_2,\sigma^{'}_3) (j_1+j_2,\sigma^{'}_3;k_3,\sigma_3|j_1+j_2+j_3,\sigma^{'}_4)\cdots\\ & \ \ \ \times (j_1+j_2+\cdots+j_{n-1},\sigma^{'}_n; j_n,\sigma_n|j_1+j_2+\cdots+j_n, \sigma^{'}_n+\sigma_n)\\ =&\left[\prod\limits_{i=1}^{n} \frac{(2j_i)!}{(j_i+\sigma_i)! (j_i-\sigma_i)!}\right]^\frac{1}{2} \left\{\frac{\left[\sum\limits^n_{i=1}(j_i+\sigma_i)\right]! \left[\sum\limits^n_{i=1}(j_i-\sigma_i)\right]!} {\left(2\sum\limits_{i=1}^n j_i\right)!} \right\}^\frac{1}{2}, \end{array}$$ we find $$\begin{array}{rl} & e_{\mu_1\mu_2\cdots\mu_j}(\roarrow{p},j,\sigma) \\ =& \sum\limits_{\sigma_1, \sigma_2,\cdots, \sigma_j} \left\{ \frac{2^j(j+\sigma)!(j-\sigma)!} {(2j)!\prod\limits^j_{i=1}[(1+\sigma_i)!(1-\sigma_i)!]} \right\}^{\frac{1}{2}} \delta_{\sigma_1+\sigma_2+\cdots+\sigma_j,\sigma} e_{\mu_1}(\roarrow{p},\sigma_1) e_{\mu_2}(\roarrow{p},\sigma_2)\cdots e_{\mu_j}(\roarrow{p},\sigma_j). \end{array}$$ The above expression is equivalent to that given by Scadron [@ScadronVF68] and Chung [@ChungHA95], since they come from the same definition of Eq. (\[eq:wfj\]). $e_{\mu_1\mu_2\cdots\mu_j}(\roarrow{p},j,\sigma)$ satisfies Rarita-Schwinger conditions: space-like, symmetric and traceless: $$\begin{aligned} \label{eq:rs1} p^{\mu_1} e_{\mu_1\mu_2...\mu_j}(\roarrow{p}, j, \sigma) & = & 0,\\ \label{eq:rs2} e_{\mu_1...\mu_k...\mu_l...\mu_j}(\roarrow{p}, j, \sigma) & = & e_{\mu_1...\mu_l...\mu_k...\mu_j}(\roarrow{p}, j, \sigma),\\ \label{eq:rs3} g^{\mu_1\mu_2} e_{\mu_1\mu_2...\mu_j}(\roarrow{p}, j, \sigma) & = & 0.\end{aligned}$$ Spin projection operator is [@FronsdalSP58] $$\label{eq:prj} \begin{array}{rl} & {\cal P}^{(j)}_{\mu_1 \mu_2 ... \mu_j ; \nu_1 \nu_2 ... \nu_j} \\ \equiv & \sum\limits_{\sigma} e_{\mu_1\mu_2...\mu_j}(\roarrow{p},j,\sigma) e^\ast_{\nu_1\nu_2...\nu_j}(\roarrow{p},j,\sigma) \\ = & \frac{1}{(j!)^2} \sum\limits_{\begin{array}{l} P\left\{ \mu_1 \mu_2 ... \mu_j \right\} \\ P\left\{ \nu_1 \nu_2 ... \nu_j \right\} \end{array} } \left( \prod\limits^{j}_{i=1}\tilde{g}_{\mu_i \nu_i}+ a_{1}^{(j)} \tilde{g}_{\mu_1 \mu_2} \tilde{g}_{\nu_1 \nu_2} \prod\limits^{j}_{i=3}\tilde{g}_{\mu_i \nu_i} + ... \right . \\ & + \left\{ \begin{array}{ll} \left . a_{j/2}^{(j)}\prod\limits_{i=1}^{j/2}\left( \tilde{g}_{\mu_{2i-1}\mu_{2i}}\tilde{g}_{\nu_{2i-1}\nu_{2i}} \right) \right), & {\rm for\; even}\; j, \\ \left. a_{(j-1)/2}^{(j)} \tilde{g}_{\mu_j\nu_j} \prod\limits_{i=1}^{(j-1)/2}\left( \tilde{g}_{\mu_{2i-1}\mu_{2i}}\tilde{g}_{\nu_{2i-1}\nu_{2i}} \right) \right), & {\rm for\; odd}\; j; \end{array} \right . \end{array}$$ where $$a_K^{(j)} = \frac{(-1)^K j! } {2^K K! (j-2K)!} \frac{1} {(2j-1)(2j-3)...(2j-2K+1)},$$ and $$\tilde{g}_{\mu\nu} = -g_{\mu \nu} + \frac{{p_\mu} {p_\nu}} {W^2}$$ is the spin-$1$ projection operator. Spin projection operators are useful when we sum over final (or initial) particles’ spins. They also serve as numerators of propagators [@WeinbergQFT1]. From Eq. [(\[eq:prj\])]{}, the first five projection operators read $$\begin{aligned} {\cal P}^{(1)}_{\mu \nu} & = & -g_{\mu \nu} + \frac{{p_\mu} {p_\nu}} {W^2} ,\\ {\cal P}^{(2)}_{\mu_1 \mu_2 ;\nu_1 \nu_2} & = & \frac{1}{2}(\tilde{g}_{\mu_1 \nu_1} \tilde{g}_{\mu_2 \nu_2} + \tilde{g}_{\mu_2 \nu_1} \tilde{g}_{\mu_1 \nu_2}) - \frac{1}{3}\tilde{g}_{\mu_1 \mu_2}\tilde{g}_{\nu_1 \nu_2}, \end{aligned}$$ $$\begin{array}{rl} {\cal P}^{(3)}_{\mu_1 \mu_2 \mu_3 ;\nu_1 \nu_2 \nu_3} & = \\ +\frac{1}{6}\sum\limits_{P\{\nu_1,\nu_2,\nu_3\}} & \tilde{g}_{\mu_1 \nu_1} \tilde{g}_{\mu_2 \nu_2} \tilde{g}_{\mu_3 \nu_3}\\ -\frac{1}{30}\sum\limits_{P\{\nu_1,\nu_2,\nu_3\}} & (\tilde{g}_{\mu_1 \mu_2}\tilde{g}_{\nu_1 \nu_2}\tilde{g}_{\mu_3\nu_3} +\tilde{g}_{\mu_1 \nu_1}\tilde{g}_{\mu_2 \mu_3}\tilde{g}_{\nu_2\nu_3} +\tilde{g}_{\mu_1 \mu_3}\tilde{g}_{\nu_1 \nu_3}\tilde{g}_{\mu_2\nu_2}), \end{array}$$ $$\begin{array}{rl} {\cal P}^{(4)}_{\mu_1 \mu_2 \mu_3 \mu_4; \nu_1 \nu_2 \nu_3 \nu_4} & = \\ +\frac{1}{24}\sum\limits_{P\{\nu_1,\nu_2,\nu_3,\nu_4\}} & \tilde{g}_{\mu_1\nu_1} \tilde{g}_{\mu_2\nu_2} \tilde{g}_{\mu_3\nu_3} \tilde{g}_{\mu_4\nu_4} \\ -\frac{1}{168}\sum\limits_{P\{\nu_1,\nu_2,\nu_3,\nu_4\}} & (\tilde{g}_{\mu_1\mu_2}\tilde{g}_{\mu_3\nu_1}\tilde{g}_{\mu_4\nu_2}+ \tilde{g}_{\mu_2\mu_3}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_4\nu_2}+ \tilde{g}_{\mu_1\mu_3}\tilde{g}_{\mu_2\nu_1}\tilde{g}_{\mu_4\nu_2}+\\ & \tilde{g}_{\mu_1\mu_4}\tilde{g}_{\mu_2\nu_1}\tilde{g}_{\mu_3\nu_2}+ \tilde{g}_{\mu_2\mu_4}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_3\nu_2}+ \tilde{g}_{\mu_3\mu_4}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_2\nu_2} ) \tilde{g}_{\nu_3\nu_4} \\ +\frac{1}{840}\sum\limits_{P\{\nu_1,\nu_2,\nu_3,\nu_4\}} & ( \tilde{g}_{\mu_1\mu_2}\tilde{g}_{\mu_3\mu_4}+ \tilde{g}_{\mu_1\mu_3}\tilde{g}_{\mu_2\mu_4}+ \tilde{g}_{\mu_1\mu_4}\tilde{g}_{\mu_2\mu_3} ) \tilde{g}_{\nu_1\nu_2}\tilde{g}_{\nu_3\nu_4}, \end{array}$$ $$\begin{array}{rl} {\cal P}^{(5)}_{\mu_1\mu_2\mu_3\mu_4\mu_5; \nu_1\nu_2\nu_3\nu_4\nu_5} & = \\ +\frac{1}{120}\sum\limits_{P\{\nu_1,\nu_2,\nu_3,\nu_4,\nu_5\}} & \tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_2\nu_2}\tilde{g}_{\mu_3\nu_3} \tilde{g}_{\mu_4\nu_4}\tilde{g}_{\mu_5\nu_5} \\ -\frac{1}{1080}\sum\limits_{P\{\nu_1,\nu_2,\nu_3,\nu_4,\nu_5\}} & (\tilde{g}_{\mu_1\mu_2}\tilde{g}_{\mu_3\nu_1}\tilde{g}_{\mu_4\nu_2} \tilde{g}_{\mu_5\nu_3} +\tilde{g}_{\mu_1\mu_3}\tilde{g}_{\mu_2\nu_1}\tilde{g}_{\mu_4\nu_2} \tilde{g}_{\mu_5\nu_3}\\ & +\tilde{g}_{\mu_1\mu_4}\tilde{g}_{\mu_2\nu_1}\tilde{g}_{\mu_3\nu_2} \tilde{g}_{\mu_5\nu_3} +\tilde{g}_{\mu_1\mu_5}\tilde{g}_{\mu_2\nu_1}\tilde{g}_{\mu_3\nu_2} \tilde{g}_{\mu_4\nu_3}\\ & +\tilde{g}_{\mu_2\mu_3}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_4\nu_2} \tilde{g}_{\mu_5\nu_3} +\tilde{g}_{\mu_2\mu_4}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_3\nu_2} \tilde{g}_{\mu_5\nu_3}\\ & +\tilde{g}_{\mu_2\mu_5}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_3\nu_2} \tilde{g}_{\mu_4\nu_3} +\tilde{g}_{\mu_3\mu_4}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_2\nu_2} \tilde{g}_{\mu_5\nu_3}\\ & +\tilde{g}_{\mu_3\mu_5}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_2\nu_2} \tilde{g}_{\mu_4\nu_3} +\tilde{g}_{\mu_4\mu_5}\tilde{g}_{\mu_1\nu_1}\tilde{g}_{\mu_2\nu_2} \tilde{g}_{\mu_3\nu_3} )\tilde{g}_{\nu_4\nu_5}\\ +\frac{1}{7560}\sum\limits_{P\{\nu_1,\nu_2,\nu_3,\nu_4,\nu_5\}} & (\tilde{g}_{\mu_1\mu_2}\tilde{g}_{\mu_3\mu_4}\tilde{g}_{\mu_5\nu_1} +\tilde{g}_{\mu_1\mu_3}\tilde{g}_{\mu_2\mu_4}\tilde{g}_{\mu_5\nu_1} +\tilde{g}_{\mu_1\mu_4}\tilde{g}_{\mu_2\mu_3}\tilde{g}_{\mu_5\nu_1}\\ & +\tilde{g}_{\mu_1\mu_2}\tilde{g}_{\mu_3\mu_5}\tilde{g}_{\mu_4\nu_1} +\tilde{g}_{\mu_1\mu_3}\tilde{g}_{\mu_2\mu_5}\tilde{g}_{\mu_4\nu_1} +\tilde{g}_{\mu_1\mu_5}\tilde{g}_{\mu_2\mu_3}\tilde{g}_{\mu_4\nu_1}\\ & +\tilde{g}_{\mu_1\mu_2}\tilde{g}_{\mu_4\mu_5}\tilde{g}_{\mu_3\nu_1} +\tilde{g}_{\mu_1\mu_4}\tilde{g}_{\mu_2\mu_5}\tilde{g}_{\mu_3\nu_1} +\tilde{g}_{\mu_1\mu_5}\tilde{g}_{\mu_2\mu_4}\tilde{g}_{\mu_3\nu_1}\\ & +\tilde{g}_{\mu_1\mu_3}\tilde{g}_{\mu_4\mu_5}\tilde{g}_{\mu_2\nu_1} +\tilde{g}_{\mu_1\mu_4}\tilde{g}_{\mu_3\mu_5}\tilde{g}_{\mu_2\nu_1} +\tilde{g}_{\mu_1\mu_5}\tilde{g}_{\mu_3\mu_4}\tilde{g}_{\mu_2\nu_1}\\ & +\tilde{g}_{\mu_2\mu_3}\tilde{g}_{\mu_4\mu_5}\tilde{g}_{\mu_1\nu_1} +\tilde{g}_{\mu_2\mu_4}\tilde{g}_{\mu_3\mu_5}\tilde{g}_{\mu_1\nu_1} +\tilde{g}_{\mu_2\mu_5}\tilde{g}_{\mu_3\mu_4}\tilde{g}_{\mu_1\nu_1} )\tilde{g}_{\nu_2\nu_3}\tilde{g}_{\nu_4\nu_5}. \end{array}$$ Now we can state Feynman rules for bosons as: (1) Every incoming particle or incoming antiparticle contributes a factor of $e_{\mu_1 \mu_2 ... \mu_j}(\roarrow{p}, j , \sigma)$. (2) Every outgoing particle or outgoing antiparticle contributes a factor of $e^{\ast}_{\mu_1 \mu_2 ... \mu_j}(\roarrow{p}, j , \sigma)$. (3) For each spin-$j$ internal line, include a factor [@WeinbergQFT1] $$\label{eq:prop} i \frac{ {\cal P}^{(j)}_{\mu_1 \mu_2 ... \mu_j ; \nu_1 \nu_2 ... \nu_j} } {p^2 -W^2 + i \epsilon}.$$ Note that we have dropped constants unnecessary for amplitude analysis. The denominator of Eq. [(\[eq:prop\])]{} is often changed to Breit-Wigner form as an approximation to the full propagator: $$\label{eq:bw} i \frac{ {\cal P}^{(j)}_{\mu_1 \mu_2 ... \mu_j ; \nu_1 \nu_2 ... \nu_j} } {p^2 -W^2 + i \Gamma W},$$ where $\Gamma$ is the width of the particle. The width is either determined from experiments or from the chain approximation of theoretical models. The problem left now is how to write down effective vertexes. We will give a list of 3-leg effective vertexes in next section. Three-leg effective vertexes {#sec:3v} ============================ The effective vertex $\Gamma$ should be constructed from momentums and isotropic tensors of the Lorentz group. Free indexes left after contraction of $p_i^\mu , g^{\mu \nu} , \varepsilon^{\alpha \beta \gamma \delta}$ are the Lorentz indexes of $\Gamma$. Here $p_i^\mu$ ($i = 1, 2, ... , n$) is the four-momenta for particle $i$ and $\varepsilon^{\alpha \beta \gamma \delta}$ the antisymmetric tensor. Not all possible constructions are independent since we have conservation of energy and momentum $$p_1^\mu + p_2^\mu + ... + p_n^\mu = 0$$ and Rarita-Schwinger conditions. The antisymmetric tensor has the property $$\label{eq:2eps} \varepsilon_{\mu \nu \alpha \beta} \varepsilon_{\mu^{'} \nu{'} \alpha^{'} \beta^{'}} = -\det \left\{ \begin{array}{cccc} g_{\mu \mu^{'}} & g_{\nu \mu^{'}} & g_{\alpha \mu^{'}} & g_{\beta \mu^{'}} \\ g_{\mu \nu^{'}} & g_{\nu \nu^{'}} & g_{\alpha \nu^{'}} & g_{\beta \nu^{'}} \\ g_{\mu \alpha^{'}} & g_{\nu \alpha^{'}} & g_{\alpha \alpha^{'}} & g_{\beta \alpha^{'}} \\ g_{\mu \beta^{'}} & g_{\nu \beta^{'}} & g_{\alpha \beta^{'}} & g_{\beta \beta^{'}} \end{array} \right\},$$ so products of $\varepsilon_{\mu \nu \alpha \beta}$ can be absorbed into other terms K. S. freely. Let’s move on to the case of effective vertexes with only three legs. As shown in Fig. \[fig:3leg\], the four-momenta of the three particles are $p_1$, $p_2$ and $p_3$. The spins are $J_1$, $J_2$ and $J_3$, and helicities being $\lambda_1$, $\lambda_2$ and $\lambda_3$. The Lorentz indexes for these particles are $(\alpha_1, \alpha_2, ..., \alpha_{J_1})$, $(\beta_1, \beta_2, ..., \beta_{J_2})$ and $(\gamma_1, \gamma_2, ..., \gamma_{J_3})$, so the effective vertex $\Gamma$ has $J_1 + J_2 + J_3$ indexes. The antisymmetric tensor $\varepsilon^{\alpha \beta \gamma \delta}$ in a three-leg vertex should contract with at least one four-momenta since wave functions are symmetric. We define $$\begin{aligned} A_1^{\alpha \beta \gamma} & \equiv & p_{1 \mu} \varepsilon^{\mu \alpha \beta \gamma}, \\ A_2^{\alpha \beta \gamma} & \equiv & p_{2 \mu} \varepsilon^{\mu \alpha \beta \gamma}, \\ Q^{\mu \nu} & \equiv & p_{1 \alpha} p_{2 \beta} \varepsilon^{\alpha \beta \mu \nu}.\end{aligned}$$ In effective vertexes involving higher spin particles, some seemingly independent terms we write down are in fact not independent. There are equivalence relations among them. These equivalence relations come from Eq. [(\[eq:2eps\])]{} and the following identity [@ScadronVF68] $$\begin{aligned} \varepsilon^{\alpha\beta\gamma\delta}g^{\mu\nu} & = & \varepsilon^{\mu\beta\gamma\delta}g^{\alpha\nu} +\varepsilon^{\alpha\mu\gamma\delta}g^{\beta\nu} +\varepsilon^{\alpha\beta\mu\delta}g^{\gamma\nu} +\varepsilon^{\alpha\beta\gamma\mu}g^{\delta\nu}.\end{aligned}$$ We can find[^3] $$\begin{aligned} \label{eq:scadron1} Q^{\alpha_1 \gamma_1} p_1^{\beta_1} & \simeq & p_1^2 A_2^{\alpha_1 \beta_1 \gamma_1} -(p_1 \cdot p_2)A_1^{\alpha_1 \beta_1 \gamma_1} +Q^{\alpha_1 \beta_1} p_1^{\gamma_1} , \\ \label{eq:scadron2} Q^{\beta_1 \gamma_1} p_2^{\alpha_1} & \simeq & - (p_1 \cdot p_2)A_2^{\alpha_1 \beta_1 \gamma_1} +p_2^2 A_1^{\alpha_1 \beta_1 \gamma_1} +Q^{\alpha_1 \beta_1} p_1^{\gamma_1},\end{aligned}$$ $$\begin{aligned} Q^{\alpha_1 \beta_1} g^{\alpha_2 \gamma_1} & \simeq & A_1^{\alpha_1 \beta_1 \gamma_1} p_2^{\alpha_2} +Q^{\alpha_1 \gamma_1} g^{\alpha_2 \beta_1},\\ Q^{\alpha_1 \beta_1} g^{\beta_2 \gamma_1} & \simeq & -A_2^{\alpha_1 \beta_1 \gamma_1} p_1^{\beta_2} -Q^{\beta_1 \gamma_1} g^{\alpha_1 \beta_2} ,\end{aligned}$$ $$\label{eq:scadron3} Q^{\alpha_1 \gamma_1} g^{\beta_1 \gamma_2} -Q^{\beta_1 \gamma_1} g^{\alpha_1 \gamma_2} \simeq A_1^{\alpha_1 \beta_1 \gamma_1} p_1^{\gamma_2} + A_2^{\alpha_1 \beta_1 \gamma_1} p_1^{\gamma_2},$$ $$\begin{array}{rl} & g^{\alpha_1 \gamma_1}p_1^{\beta_1}p_1^{\beta_2}p_1^{\gamma_2}p_2^{\alpha_2} - g^{\beta_1 \gamma_1}p_1^{\beta_2}p_1^{\gamma_2}p_2^{\alpha_1}p_2^{\alpha_2}\\ \simeq & g^{\alpha_1\beta_1}p_1^{\beta_2}p_1^{\gamma_1}p_1^{\gamma_2}p_2^{\alpha_2} +[(p_1\cdot p_2)^2 - p_1^2 p_2^2] g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1}g^{\beta_2\gamma_2}\\ & -\frac{1}{2}p_2^2 g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} p_1^{\beta_1}p_1^{\beta_2} -\frac{1}{2}p_1^2 g^{\beta_1\gamma_1}g^{\beta_2\gamma_2} p_2^{\alpha_1}p_2^{\alpha_2}\\ & -\frac{1}{2}(p_1+p_2)^2 g^{\alpha_1\beta_1}g^{\alpha_2\beta_2} p_1^{\gamma_1}p_1^{\gamma_2} -(p_1\cdot p_2)g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2} p_1^{\beta_2}p_2^{\alpha_2}\\ & +[(p_1\cdot p_2)+p_2^2]g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} p_1^{\beta_2}p_1^{\gamma_2} -[(p_1\cdot p_2)+p_1^2]g^{\alpha_1\beta_1}g^{\beta_2\gamma_1} p_1^{\gamma_2}p_2^{\alpha_2}, \end{array}$$ $$A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2}g^{\beta_2\gamma_2}+ A_2^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2}p_1^{\beta_2} \simeq - g^{\alpha_1\beta_1}Q^{\alpha_2\gamma_1}g^{\beta_2\gamma_2} - g^{\alpha_1\beta_1}Q^{\beta_2\gamma_1}g^{\alpha_2\gamma_2},$$ $$\begin{array}{rl} & A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2}p_1^{\beta_2}p_1^{\gamma_2} + A_2^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2}p_1^{\beta_2}p_1^{\gamma_2}\\ \simeq & \frac{1}{2}p_1^2 A_2^{\alpha_1\beta_1\gamma_1} p_2^{\alpha_2}g^{\beta_2\gamma_2} - \frac{1}{2}(p_1\cdot p_2)A_1^{\alpha_1\beta_1\gamma_1} p_2^{\alpha_2}g^{\beta_2\gamma_2} \\ & + \frac{1}{2}(p_1\cdot p_2)A_2^{\alpha_1\beta_1\gamma_1} g^{\alpha_2\gamma_2}p_1^{\beta_2} - \frac{1}{2}p_2^2 A_1^{\alpha_1\beta_1\gamma_1} g^{\alpha_2\gamma_2}p_1^{\beta_2} \\ & + \frac{1}{2}[(p_1\cdot p_2)-p_1^2]g^{\alpha_1\beta_1} A_2^{\alpha_2\beta_2\gamma_1}p_1^{\gamma_2} \\ & + \frac{1}{2}[(p_1\cdot p_2)-p_2^2]g^{\alpha_1\beta_1} A_1^{\alpha_2\beta_2\gamma_1}p_1^{\gamma_2} \\ & - g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2} p_1^{\gamma_1}p_1^{\gamma_2}. \end{array}$$ The symbol “$\simeq$” means that the left hand side and the right hand side of the equation are equal only when they are contracted with wave functions satisfying Eqs. [(\[eq:rs1\]-\[eq:rs3\])]{}. The left hand side of these equations can be absorbed into those terms on the right hand side without introducing kinematic singularities. A K. S. free effective vertex is written as $$\Gamma = \Gamma_+ + \Gamma_-,$$ where $\Gamma_+$($\Gamma_-$) is the tensor(pseudo-tensor) part of the vertex. We sort the three particles in ascending order of their spins, i.e., $J_1 \leq J_2 \leq J_3$. After considering these equivalence relations, one finds: - $(J_1, J_2, J_3) = (0, 0, j)$, with $j \leq 0$ $$\Gamma_+ = c_1 \left( p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right),$$ $$\label{eq:00j-} \Gamma_- = 0;$$ - $(J_1, J_2, J_3) = (0, j, j^{'})$, with $1 \leq j \leq j^{'}$ $$\begin{array}{rcl} \Gamma_+ & = & \; c_1 \left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & + c_2 \left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & +\cdots \\ & & + c_j \left( g^{\beta_1\gamma_1}p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j^{'}} \right)\\ & & + c_{j+1} \left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j^{'}} \right), \end{array}$$ $$\begin{array}{rcrl} \Gamma_- & = & Q^{\beta_1\gamma_1} & \left\{ c_{j+2}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_j^{'}} \right) \right. \\ & & & + c_{j+3} \left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_{j}}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_j^{'}} \right)\\ & & & + \cdots \\ & & & + c_{2j}\left( g^{\beta_2\gamma_2}p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & \left. + c_{2j+1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_{j}} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j^{'}}} \right) \right\}; \end{array}$$ - $(J_1, J_2, J_3) = (1, j, j)$, with $j \leq 1$ $$\begin{array}{rcrl} \Gamma_+ & = & g^{\alpha_1\gamma_1} & \left\{ c_1\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j} \right)\right.\\ & & & + c_2\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_{j-2}\gamma_{j-1}} p_1^{\beta_{j-1}}p_1^{\beta_j}p_1^{\gamma_j} \right)\\ & & & + \cdots \\ & & & \left. + c_j \left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right) \right\}\\ & & + p_2^{\alpha_1} & \left\{ c_{j+1}\left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_j\gamma_j} \right)\right. \\ & & & + c_{j+2}\left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j} \right)\\ & & & +\cdots \\ & & & \left. + c_{2j+1}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right) \right\} \\ & & + g^{\alpha_1\beta_1} & \left\{ c_{2j+2}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j} \right) \right.\\ & & & + c_{2j+3}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j} \right)\\ & & & +\cdots \\ & & & \left. + c_{3j+1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right)\right\}, \end{array}$$ $$\begin{array}{rcrl} \Gamma_- & = & A_1^{\alpha_1\beta_1\gamma_1} & \left\{ c_{3j+2}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} \right)\right.\\ & & & + c_{3j+3}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} \right)\\ & & & + \cdots \\ & & & \left. + c_{4j+1} \left( p_1^{\beta_2}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right)\right\}\\ & & + A_2^{\alpha_1\beta_1\gamma_1} & \left\{ c_{4j+2}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} \right)\right.\\ & & & + c_{4j+3}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} \right)\\ & & & + \cdots \\ & & & \left. + c_{5j+1} \left( p_1^{\beta_2}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right)\right\}\\ & & + Q^{\alpha_1\beta_1} & \left\{ c_{5j+2}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j} \right)\right. \\ & & & + c_{5j+3}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j} \right)\\ & & & + \cdots \\ & & & \left. + c_{6j+1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right)\right\}; \end{array}$$ - $(J_1, J_2, J_3) = (1, j, j^{'})$, with $1 \leq j < j^{'}$ $$\begin{array}{rcrl} \Gamma_+ & = & g^{\alpha_1\gamma_1} & \left\{ c_1 \left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_j\gamma_{j+1}} p_1^{\gamma_{j+2}}p_1^{\gamma_{j+3}}\cdots p_1^{\gamma_{j^{'}}} \right) \right. \\ & & & +c_2 \left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j}p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & + \cdots \\ & & & + \left. c_{j+1}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j^{'}} \right) \right\} \\ & & + p_2^{\alpha_1} & \left\{ c_{j+1}\left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right) \right. \\ & & & + c_{j+2} \left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_{j}}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & + \left. c_{2j+2}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j^{'}} \right) \right\}\\ & & + g^{\alpha_1\beta_1} & \left\{ c_{2j+3}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right) \right. \\ & & & + c_{2j+4}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. + c_{3j+2}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j^{'}}} \right) \right\} , \end{array}$$ $$\begin{array}{rcrl} \Gamma_- & = & A_1^{\alpha_1\beta_1\gamma_1} & \left\{ c_{3j+3}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right) \right.\\ & & & + c_{3j+4}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. + c_{4j+2}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j^{'}} \right) \right\}\\ & & + A_2^{\alpha_1\beta_1\gamma_1} & \left\{ c_{4j+3}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right) \right.\\ & & & + c_{4j+4}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. + c_{5j+2}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j^{'}} \right) \right\}\\ & & + Q^{\alpha_1\gamma_1} & c_{5j+3}\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_j\gamma_{j+1}} p_1^{\gamma_{j+2}}p_1^{\gamma_{j+3}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & + Q^{\alpha_1\beta_1} & \left\{ c_{5j+4}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right) \right. \\ & & & + c_{5j+5}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j}\cdots p_1^{\gamma_{j^{'}}} \right) \\ & & & \left. + c_{6j+3}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j^{'}}} \right) \right\}; \end{array}$$ - $(J_1, J_2, J_3) = (2, 2, 2)$ $$\begin{array}{rcl} \Gamma_+ & = &\;\;\; c_1 g^{\alpha_1\beta_1}g^{\alpha_2\beta_2} p_1^{\gamma_1}p_1^{\gamma_2} \\ & & +c_2 g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} g^{\beta_2\gamma_2}\\ & & +c_3 g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} p_1^{\beta_2}p_1^{\gamma_2} \\ & & +c_4 g^{\alpha_1\beta_1}p_2^{\alpha_2} g^{\beta_2\gamma_1}p_1^{\gamma_2}\\ & & +c_5 g^{\alpha_1\beta_1}p_2^{\alpha_2} p_1^{\beta_2}p_1^{\gamma_1}p_1^{\gamma_2} \\ & & +c_6 g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} p_1^{\beta_1}p_1^{\beta_2}\\ & & +c_7 g^{\alpha_1\gamma_1}p_2^{\alpha_2} g^{\beta_1\gamma_2}p_1^{\beta_2}\\ & & +c_8 g^{\alpha_1\gamma_1}p_2^{\alpha_2} p_1^{\beta_1}p_1^{\beta_2}p_1^{\gamma_2}\\ & & + c_9 p_2^{\alpha_1}p_2^{\alpha_2} g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\\ & & + c_{10} p_2^{\alpha_1}p_2^{\alpha_2} p_1^{\beta_1}p_1^{\beta_2}p_1^{\gamma_1}p_1^{\gamma_2}, \end{array}$$ $$\begin{array}{rcl} \Gamma_- & = & \;\;\; c_{11} g^{\alpha_1\beta_1}A_1^{\alpha_2\beta_2\gamma_1} p_1^{\gamma_2} \\ & & +c_{12} g^{\alpha_1\beta_1}A_2^{\alpha_2\beta_2\gamma_1} p_1^{\gamma_2}\\ & & + c_{13} g^{\alpha_1\beta_1}Q^{\alpha_2\gamma_1} g^{\beta_2\gamma_2}\\ & & +c_{14} g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2} p_1^{\gamma_1}p_1^{\gamma_2}\\ & & +c_{15} A_1^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} p_1^{\beta_2}\\ & & +c_{16} A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} g^{\beta_2\gamma_2}\\ & & +c_{17} A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} p_1^{\beta_2}p_1^{\gamma_2}\\ & & + c_{18} A_2^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} g^{\beta_2\gamma_2}\\ & & + c_{19} Q^{\alpha_1\beta_1}p_2^{\alpha_2} p_1^{\beta_2}p_1^{\gamma_1}p_1^{\gamma_2}; \end{array}$$ - $(J_1, J_2, J_3) = (2, j, j)$, with $j \geq 3$ $$\begin{array}{rcrl} \Gamma_+ & = & g^{\alpha_1\beta_1}g^{\alpha_2\beta_2} & \left\{ c_1\left( g^{\beta_3\gamma_1}g^{\beta_4\gamma_2}\cdots g^{\beta_j\gamma_{j-2}} p_1^{\gamma_{j-1}}p_1^{\gamma_j} \right)\right. \\ & & & +c_2\left( g^{\beta_3\gamma_1}g^{\beta_4\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-3}} p_1^{\beta_j}p_1^{\gamma_{j-2}}p_1^{\gamma_{j-1}} p_1^{\gamma_j} \right) \\ & & & +\cdots \\ & & & \left. + c_{j-1} \left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right) \right\} \\ & & + g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} & \left\{ c_j\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} \right) \right. \\ & & & +c_{j+1}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j} \right) \\ & & & +\cdots \\ & & & \left. +c_{2j-1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right) \right\} \\ & & +g^{\alpha_1\beta_1}p_2^{\alpha_2} & \left\{ c_{2j}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j} \right)\right. \\ & & & +c_{2j+1}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{3j-1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right) \right\} \\ & & +g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} & \left\{ c_{3j}\left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4}\cdots g^{\beta_{j-2}\gamma_j} p_1^{\beta_{j-1}}p_1^{\beta_j} \right) \right. \\ & & & +c_{3j+1}\left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4}\cdots g^{\beta_{j-3}\gamma_{j-1}} p_1^{\beta_{j-2}}p_1^{\beta_{j-1}}p_1^{\beta_j}p_1^{\gamma_j} \right) \\ & & & +\cdots \\ & & & \left. +c_{4j-2}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right)\right\} \\ & & +g^{\alpha_1\gamma_1}p_2^{\alpha_2} & \left\{ c_{4j-1}\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j} \right) \right. \\ & & & +c_{4j}\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j} \right) \\ & & & +\cdots \\ & & & \left. +c_{5j-2}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right)\right\} \\ & & +p_2^{\alpha_1}p_2^{\alpha_2} & c_{5j-1}\left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_j\gamma_j} \right)\\ & & +p_2^{\alpha_1}p_2^{\alpha_2} & c_{5j}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right), \end{array}$$ $$\begin{array}{rcrl} \Gamma_- & = & g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}Q^{\beta_3\gamma_1} & \left\{ c_{5j+1}\left( g^{\beta_4\gamma_2}g^{\beta_5\gamma_3}\cdots g^{\beta_j\gamma_{j-2}} p_1^{\gamma_{j-1}}p_1^{\gamma_j} \right)\right. \\ & & & +c_{5j+2}\left( g^{\beta_4\gamma_2}g^{\beta_5\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-3}} p_1^{\beta_j}p_1^{\gamma_{j-2}}p_1^{\gamma_{j-1}}p_1^{\gamma_j} \right) \\ & & & +\cdots \\ & & & \left. +c_{6j-2}\left( p_1^{\beta_4}p_1^{\beta_5}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right)\right\}\\ & & +g^{\alpha_1\beta_1}A_1^{\alpha_2\beta_2\gamma_1} & \left\{ c_{6j-1}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j} \right)\right.\\ & & & +c_{6j}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{7j-3}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right)\right\}\\ & & +g^{\alpha_1\beta_1}A_2^{\alpha_2\beta_2\gamma_1} & \left\{ c_{7j-2}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j} \right)\right.\\ & & & +c_{7j-1}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{8j-4}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right)\right\} \\ & & +g^{\alpha_1\beta_1}Q^{\alpha_2\gamma_1} & c_{8j-3}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} \right)\\ & & +g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2} & c_{8j-2}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right)\\ & & +A_1^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} & \left\{ c_{8j-1}\left( g^{\beta_2\gamma_3}g^{\beta_3\gamma_4}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j} \right)\right.\\ & & & +c_{8j}\left( g^{\beta_2\gamma_3}g^{\beta_3\gamma_4}\cdots g^{\beta_{j-2}\gamma_{j-1}} p_1^{\beta_{j-1}}p_1^{\beta_j}p_1^{\gamma_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{9j-3}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right)\right\}\\ & & +A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} & \left\{ c_{9j-2}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} \right)\right. \\ & & & +c_{9j-1}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{10j-3}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right)\right\}\\ & & +A_2^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} & c_{10j-2}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} \right)\\ & & +Q^{\alpha_1\beta_1}p_2^{\alpha_2} & c_{10j-1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right); \end{array}$$ - $(J_1, J_2, J_3) = (2, 2, 3)$ $$\begin{array}{rcl} \Gamma_+ & = &\;\;\; c_1 g^{\alpha_1\beta_1}g^{\alpha_2\beta_2} p_1^{\gamma_1}p_1^{\gamma_2}p_1^{\gamma_3} \\ & & +c_2 g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} g^{\beta_2\gamma_2}p_1^{\gamma_3} \\ & & +c_3 g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} p_1^{\beta_2}p_1^{\gamma_2}p_1^{\gamma_3}\\ & & +c_4 g^{\alpha_1\beta_1}p_2^{\alpha_2} g^{\beta_2\gamma_1}p_1^{\gamma_2}p_1^{\gamma_3}\\ & & +c_5 g^{\alpha_1\beta_1}p_2^{\alpha_2} p_1^{\beta_2}p_1^{\gamma_1}p_1^{\gamma_2}p_1^{\gamma_3}\\ & & +c_6 g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} g^{\beta_1\gamma_3}p_1^{\beta_2}\\ & & +c_7 g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} p_1^{\beta_1}p_1^{\beta_2}p_1^{\gamma_3}\\ & & +c_8 g^{\alpha_1\gamma_1}p_2^{\alpha_2} g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\\ & & +c_9 g^{\alpha_1\gamma_1}p_2^{\alpha_2} g^{\beta_1\gamma_2}p_1^{\beta_2}p_1^{\gamma_3}\\ & & +c_{10} g^{\alpha_1\gamma_1}p_2^{\alpha_2} p_1^{\beta_1}p_1^{\beta_2}p_1^{\gamma_2}p_1^{\gamma_3}\\ & & + c_{11} p_2^{\alpha_1}p_2^{\alpha_2} g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p_1^{\gamma_3}\\ & & + c_{12} p_2^{\alpha_1}p_2^{\alpha_2} p_1^{\beta_1}p_1^{\beta_2} p_1^{\gamma_1}p_1^{\gamma_2}p_1^{\gamma_3}, \end{array}$$ $$\begin{array}{rcl} \Gamma_- & = & \;\;\; c_{13} g^{\alpha_1\beta_1}A_1^{\alpha_2\beta_2\gamma_1} p_1^{\gamma_2}p_1^{\gamma_3}\\ & & +c_{14} g^{\alpha_1\beta_1}A_2^{\alpha_2\beta_2\gamma_1} p_1^{\gamma_2}p_1^{\gamma_3}\\ & & + c_{15} g^{\alpha_1\beta_1}Q^{\alpha_2\gamma_1} g^{\beta_2\gamma_2}p^{\gamma_3}\\ & & +c_{16} g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2} p_1^{\gamma_1}p_1^{\gamma_2}p_1^{\gamma_3}\\ & & +c_{17} A_1^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} g^{\beta_2\gamma_3}\\ & & +c_{18} A_1^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} p_1^{\beta_2}p_1^{\gamma_3}\\ & & +c_{19} A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} g^{\beta_2\gamma_2}p_1^{\gamma_3}\\ & & +c_{20} A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} p_1^{\beta_2}p_1^{\gamma_2}p_1^{\gamma_3}\\ & & + c_{21} A_2^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} g^{\beta_2\gamma_3}\\ & & + c_{22} A_2^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} g^{\beta_2\gamma_2}p_1^{\gamma_3}\\ & & + c_{23} Q^{\alpha_1\beta_1}p_2^{\alpha_2} p_1^{\beta_2}p_1^{\gamma_1}p_1^{\gamma_2}p_1^{\gamma_3}; \end{array}$$ - $(J_1, J_2, J_3) = (2, j, j+1)$, with $j \geq 3$ $$\begin{array}{rcrl} \Gamma_+ & = & g^{\alpha_1\beta_1}g^{\alpha_2\beta_2} & \left\{ c_1\left( g^{\beta_3\gamma_1}g^{\beta_4\gamma_2}\cdots g^{\beta_j\gamma_{j-2}} p_1^{\gamma_{j-1}}p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\right. \\ & & & +c_2\left( g^{\beta_3\gamma_1}g^{\beta_4\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-3}} p_1^{\beta_j}p_1^{\gamma_{j-2}}p_1^{\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right) \\ & & & +\cdots \\ & & & \left. + c_{j-1} \left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j+1}} \right) \right\} \\ & & + g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} & \left\{ c_j\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}} \right) \right. \\ & & & +c_{j+1}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right) \\ & & & +\cdots \\ & & & \left. +c_{2j-1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right) \right\} \\ & & +g^{\alpha_1\beta_1}p_2^{\alpha_2} & \left\{ c_{2j}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\right. \\ & & & +c_{2j+1}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\\ & & & +\cdots \\ & & & \left. +c_{3j-1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j+1}} \right) \right\} \\ & & +g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} & \left\{ c_{3j}\left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4}\cdots g^{\beta_{j-1}\gamma_{j+1}} p_1^{\beta_j} \right) \right. \\ & & & +c_{3j+1}\left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4}\cdots g^{\beta_{j-2}\gamma_j} p_1^{\beta_{j-1}}p_1^{\beta_j}p_1^{\gamma_{j+1}} \right) \\ & & & +\cdots \\ & & & \left. +c_{4j-1}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_{j+1}} \right)\right\} \\ & & +g^{\alpha_1\gamma_1}p_2^{\alpha_2} & \left\{ c_{4j}\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_j\gamma_{j+1}} \right) \right. \\ & & & +c_{4j+1}\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j}p_1^{\gamma_{j+1}} \right) \\ & & & +\cdots \\ & & & \left. +c_{5j}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j+1}} \right)\right\} \\ & & +p_2^{\alpha_1}p_2^{\alpha_2} & c_{5j+1}\left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}} \right)\\ & & +p_2^{\alpha_1}p_2^{\alpha_2} & c_{5j+2}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j+1}} \right), \end{array}$$ $$\begin{array}{rcrl} \Gamma_- & = & g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}Q^{\beta_3\gamma_1} & \left\{ c_{5j+3}\left( g^{\beta_4\gamma_2}g^{\beta_5\gamma_3}\cdots g^{\beta_j\gamma_{j-2}} p_1^{\gamma_{j-1}}p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\right. \\ & & & +c_{5j+4}\left( g^{\beta_4\gamma_2}g^{\beta_5\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-3}} p_1^{\beta_j}p_1^{\gamma_{j-2}}p_1^{\gamma_{j-1}}p_1^{\gamma_j} p_1^{\gamma_{j+1}} \right) \\ & & & +\cdots \\ & & & \left. +c_{6j}\left( p_1^{\beta_4}p_1^{\beta_5}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j+1}} \right)\right\}\\ & & +g^{\alpha_1\beta_1}A_1^{\alpha_2\beta_2\gamma_1} & \left\{ c_{6j+1}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\right.\\ & & & +c_{6j+2}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\\ & & & +\cdots \\ & & & \left. +c_{7j-1}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j+1}} \right)\right\}\\ & & +g^{\alpha_1\beta_1}A_2^{\alpha_2\beta_2\gamma_1} & \left\{ c_{7j}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\right.\\ & & & +c_{7j+1}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\\ & & & +\cdots \\ & & & \left. +c_{8j-2}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j+1}} \right)\right\} \\ & & +g^{\alpha_1\beta_1}Q^{\alpha_2\gamma_1} & c_{8j-1}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p^{\gamma_{j+1}} \right)\\ & & +g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2} & c_{8j}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j+1}} \right)\\ & & +A_1^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} & \left\{ c_{8j+1}\left( g^{\beta_2\gamma_3}g^{\beta_3\gamma_4}\cdots g^{\beta_j\gamma_{j+1}} \right)\right.\\ & & & +c_{8j+2}\left( g^{\beta_2\gamma_3}g^{\beta_3\gamma_4}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j}p_1^{\gamma_{j+1}} \right)\\ & & & +\cdots \\ & & & \left. +c_{9j}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_{j+1}} \right)\right\}\\ & & +A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} & \left\{ c_{9j+1}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}} \right)\right. \\ & & & +c_{9j+2}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j}p_1^{\gamma_{j+1}} \right)\\ & & & +\cdots \\ & & & \left. +c_{10j}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j+1}} \right)\right\}\\ & & +A_2^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} & c_{10j+1}\left( g^{\beta_2\gamma_3}g^{\beta_3\gamma_4}\cdots g^{\beta_j\gamma_{j+1}} \right)\\ & & +A_2^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} & c_{10j+2}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}} \right)\\ & & +Q^{\alpha_1\beta_1}p_2^{\alpha_2} & c_{10j+3}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j+1}} \right); \end{array}$$ - $(J_1, J_2, J_3) = (2, 2, j)$, with $j\geq 4$ $$\begin{array}{rcl} \Gamma_+ & = &\;\;\; c_1 g^{\alpha_1\beta_1}g^{\alpha_2\beta_2} \left( p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right) \\ & & +c_2 g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} \left( g^{\beta_2\gamma_2} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right) \\ & & +c_3 g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} \left( p_1^{\beta_2}p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right) \\ & & +c_4 g^{\alpha_1\beta_1}p_2^{\alpha_2} \left( g^{\beta_2\gamma_1} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right) \\ & & +c_5 g^{\alpha_1\beta_1}p_2^{\alpha_2} \left( p_1^{\beta_2} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right) \\ & & +c_6 g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} \left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4} p_1^{\gamma_5}p_1^{\gamma_6}\cdots p_1^{\gamma_j} \right) \\ & & +c_7 g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} \left( g^{\beta_1\gamma_3} p_1^{\beta_2}p_1^{\gamma_4}p_1^{\gamma_5}\cdots p_1^{\gamma_j} \right) \\ & & +c_8 g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} \left( p_1^{\beta_1}p_1^{\beta_2} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right) \\ & & +c_9 g^{\alpha_1\gamma_1}p_2^{\alpha_2} \left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3} p_1^{\gamma_4}p_1^{\gamma_5}\cdots p_1^{\gamma_j} \right) \\ & & +c_{10} g^{\alpha_1\gamma_1}p_2^{\alpha_2} \left( g^{\beta_1\gamma_2} p_1^{\beta_2}p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right) \\ & & +c_{11} g^{\alpha_1\gamma_1}p_2^{\alpha_2} \left( p_1^{\beta_1}p_1^{\beta_2} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right) \\ & & + c_{12} p_2^{\alpha_1}p_2^{\alpha_2} \left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right)\\ & & + c_{13} p_2^{\alpha_1}p_2^{\alpha_2} \left( p_1^{\beta_1}p_1^{\beta_2} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right), \end{array}$$ $$\begin{array}{rcl} \Gamma_- & = & \;\;\; c_{14} g^{\alpha_1\beta_1}A_1^{\alpha_2\beta_2\gamma_1} \left( p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right) \\ & & +c_{15} g^{\alpha_1\beta_1}A_2^{\alpha_2\beta_2\gamma_1} \left( p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right) \\ & & + c_{16} g^{\alpha_1\beta_1}Q^{\alpha_2\gamma_1} \left( g^{\beta_2\gamma_2} p^{\gamma_3}p^{\gamma_4}\cdots p^{\gamma_j} \right)\\ & & +c_{17} g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2} \left( p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right)\\ & & +c_{18} A_1^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} \left( g^{\beta_2\gamma_3} p_1^{\gamma_4}p_1^{\gamma_5}\cdots p_1^{\gamma_j} \right) \\ & & +c_{19} A_1^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} \left( p_1^{\beta_2} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right)\\ & & +c_{20} A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} \left( g^{\beta_2\gamma_2} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right) \\ & & +c_{21} A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} \left( p_1^{\beta_2} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_j} \right)\\ & & + c_{22} A_2^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} \left( g^{\beta_2\gamma_3} p_1^{\gamma_4}p_1^{\gamma_5}\cdots p_1^{\gamma_j} \right)\\ & & + c_{23} A_2^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} \left( g^{\beta_2\gamma_2} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_j} \right)\\ & & +c_{24} Q^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} \left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4} p_1^{\gamma_5}p_1^{\gamma_6}\cdots p_1^{\gamma_j} \right)\\ & & + c_{25} Q^{\alpha_1\beta_1}p_2^{\alpha_2} \left( p_1^{\beta_2} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_j} \right); \end{array}$$ - $(J_1, J_2, J_3) = (2, j, j^{'})$, with $j \geq 3$ and $j^{'}\geq j+2$ $$\begin{array}{rcrl} \Gamma_+ & = & g^{\alpha_1\beta_1}g^{\alpha_2\beta_2} & \left\{ c_1\left( g^{\beta_3\gamma_1}g^{\beta_4\gamma_2}\cdots g^{\beta_j\gamma_{j-2}} p_1^{\gamma_{j-1}}p_1^{\gamma_j}\cdots p_1^{\gamma_{j^{'}}} \right)\right. \\ & & & +c_2\left( g^{\beta_3\gamma_1}g^{\beta_4\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-3}} p_1^{\beta_j}p_1^{\gamma_{j-2}}p_1^{\gamma_{j-1}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. + c_{j-1} \left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j^{'}}} \right) \right\} \\ & & + g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1} & \left\{ c_j\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right) \right. \\ & & & +c_{j+1}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right) \\ & & & +\cdots \\ & & & \left. +c_{2j-1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j^{'}}} \right) \right\} \\ & & +g^{\alpha_1\beta_1}p_2^{\alpha_2} & \left\{ c_{2j}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right)\right. \\ & & & +c_{2j+1}\left( g^{\beta_2\gamma_1}g^{\beta_3\gamma_2}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. +c_{3j-1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j^{'}}} \right) \right\} \\ & & +g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} & \left\{ c_{3j}\left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4}\cdots g^{\beta_j\gamma_{j+2}} p_1^{\gamma_{j+3}}p_1^{\gamma_{j+4}}\cdots p_1^{\gamma_{j^{'}}} \right) \right. \\ & & & +c_{3j+1}\left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4}\cdots g^{\beta_{j-1}\gamma_{j+1}} p_1^{\beta_j}p_1^{\gamma_{j+2}}p_1^{\gamma_{j+3}}\cdots p_1^{\gamma_{j^{'}}} \right) \\ & & & +\cdots \\ & & & \left. +c_{4j}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_{j^{'}}} \right)\right\} \\ & & +g^{\alpha_1\gamma_1}p_2^{\alpha_2} & \left\{ c_{4j+1}\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_j\gamma_{j+1}} p_1^{\gamma_{j+2}}p_1^{\gamma_{j+3}}\cdots p_1^{\gamma_{j^{'}}} \right) \right. \\ & & & +c_{4j+2}\left( g^{\beta_1\gamma_2}g^{\beta_2\gamma_3}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j}p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right) \\ & & & +\cdots \\ & & & \left. +c_{5j+1}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j^{'}}} \right)\right\} \\ & & +p_2^{\alpha_1}p_2^{\alpha_2} & c_{5j+2}\left( g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & +p_2^{\alpha_1}p_2^{\alpha_2} & c_{5j+3}\left( p_1^{\beta_1}p_1^{\beta_2}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j^{'}}} \right), \end{array}$$ $$\begin{array}{rcrl} \Gamma_- & = & g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}Q^{\beta_3\gamma_1} & \left\{ c_{5j+4}\left( g^{\beta_4\gamma_2}g^{\beta_5\gamma_3}\cdots g^{\beta_j\gamma_{j-2}} p_1^{\gamma_{j-1}}p_1^{\gamma_j}\cdots p_1^{\gamma_{j^{'}}} \right)\right. \\ & & & +c_{5j+5}\left( g^{\beta_4\gamma_2}g^{\beta_5\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-3}} p_1^{\beta_j}p_1^{\gamma_{j-2}}p_1^{\gamma_{j-1}}\cdots p_1^{\gamma_{j^{'}}} \right) \\ & & & +\cdots \\ & & & \left. +c_{6j+1}\left( p_1^{\beta_4}p_1^{\beta_5}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j^{'}}} \right)\right\}\\ & & +g^{\alpha_1\beta_1}A_1^{\alpha_2\beta_2\gamma_1} & \left\{ c_{6j+2}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right)\right.\\ & & & +c_{6j+3}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. +c_{7j}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j^{'}}} \right)\right\}\\ & & +g^{\alpha_1\beta_1}A_2^{\alpha_2\beta_2\gamma_1} & \left\{ c_{7j+1}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_j\gamma_{j-1}} p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right)\right.\\ & & & +c_{7j+2}\left( g^{\beta_3\gamma_2}g^{\beta_4\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-2}} p_1^{\beta_j}p_1^{\gamma_{j-1}}p_1^{\gamma_j}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. +c_{8j-1}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j^{'}}} \right)\right\} \\ & & +g^{\alpha_1\beta_1}Q^{\alpha_2\gamma_1} & c_{8j}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p^{\gamma_{j+1}}p^{\gamma_{j+2}}\cdots p^{\gamma_{j^{'}}} \right)\\ & & +g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2} & c_{8j+1}\left( p_1^{\beta_3}p_1^{\beta_4}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & +A_1^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} & \left\{ c_{8j+2}\left( g^{\beta_2\gamma_3}g^{\beta_3\gamma_4}\cdots g^{\beta_j\gamma_{j+1}} p_1^{\gamma_{j+2}}p_1^{\gamma_{j+3}}\cdots p_1^{\gamma_{j^{'}}} \right)\right.\\ & & & +c_{8j+3}\left( g^{\beta_2\gamma_3}g^{\beta_3\gamma_4}\cdots g^{\beta_{j-1}\gamma_j} p_1^{\beta_j}p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. +c_{9j+1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_3}p_1^{\gamma_4}\cdots p_1^{\gamma_{j^{'}}} \right)\right\}\\ & & +A_1^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} & \left\{ c_{9j+2}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right)\right. \\ & & & +c_{9j+3}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_{j-1}\gamma_{j-1}} p_1^{\beta_j}p_1^{\gamma_j}p_1^{\gamma_{j+1}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & & +\cdots \\ & & & \left. +c_{10j+1}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_2}p_1^{\gamma_3}\cdots p_1^{\gamma_{j^{'}}} \right)\right\}\\ & & +A_2^{\alpha_1\beta_1\gamma_1}g^{\alpha_2\gamma_2} & c_{10j+2}\left( g^{\beta_2\gamma_3}g^{\beta_3\gamma_4}\cdots g^{\beta_j\gamma_{j+1}} p_1^{\gamma_{j+2}}p_1^{\gamma_{j+3}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & +A_2^{\alpha_1\beta_1\gamma_1}p_2^{\alpha_2} & c_{10j+3}\left( g^{\beta_2\gamma_2}g^{\beta_3\gamma_3}\cdots g^{\beta_j\gamma_j} p_1^{\gamma_{j+1}}p_1^{\gamma_{j+2}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & +Q^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2} & c_{10j+4}\left( g^{\beta_1\gamma_3}g^{\beta_2\gamma_4}\cdots g^{\beta_j\gamma_{j+2}} p_1^{\gamma_{j+3}}p_1^{\gamma_{j+4}}\cdots p_1^{\gamma_{j^{'}}} \right)\\ & & +Q^{\alpha_1\beta_1}p_2^{\alpha_2} & c_{10j+5}\left( p_1^{\beta_2}p_1^{\beta_3}\cdots p_1^{\beta_j} p_1^{\gamma_1}p_1^{\gamma_2}\cdots p_1^{\gamma_{j^{'}}} \right); \end{array}$$ etc. One can give more effective vertexes for those cases in which all three spins $\geq 3$, but this will take too much space and these vertexes are seldom used. We drop the factor $\delta (p_1+p_2+p_3)$ in effective vertexes and amplitudes through out this paper. Scalar coefficients $c_i$ are complex functions of the form $$c_i = c_i(p_1^2, p_2^2, p_1\cdot p_2).$$ They are just constants when all of the particles are on shell. These coefficients are also called invariant amplitudes, coupling coefficients, structure functions or form factors in different references. The propagator of an off shell particle does not satisfy Eq. [(\[eq:rs1\])]{} and Eq. [(\[eq:rs3\])]{}. In this case, those terms proportional to $$p_1^{\alpha_i},\; p_2^{\beta_i},\; (p_1+p_2)^{\gamma_i},\; g^{\alpha_i\alpha_j},\; g^{\beta_i\beta_j},\; g^{\gamma_i\gamma_j}$$ will not vanish. However, such an off shell term always contributes a factor of $p_1^2-W_1^2$, $p_2^2-W_2^2$ or $p_3^2-W_3^2$ after contraction with the propagator. This factor will eliminate the denominator of the propagator, which makes the amplitude free of pole at the mass of the intermediate particle[^4]. These contributions are not independent from background terms. They can be absorbed into back ground amplitudes without introducing kinematic singularities. Off shell terms are not needed when back ground amplitudes are included. One can refer to Sec. \[sec:bk\] or Ref. [@Zhua23pi] for examples. Symmetry under space reflection {#sec:sr} =============================== Under space reflection operation ${\bf P}$, canonical quantum states transform as $${\bf P}|\roarrow{p}, j, m\rangle = \eta |-\roarrow{p}, j, m\rangle.$$ If parity is conserved $${\bf P}^\dagger {\bf S} {\bf P} = {\bf S},$$ $$\label{eq:pp} \begin{array}{rl} & \langle \roarrow{p}_1^{'},J_1^{'},m_1^{'};\roarrow{p}_2^{'},J_2^{'},m_2^{'}; \cdots|{\bf S}|\roarrow{p}_1,J_1,m_1;\roarrow{p}_2,J_2,m_2;\cdots\rangle \\ = & {\eta_1^{'}}^{\ast}{\eta_2^{'}}^{\ast}\cdots \eta_1\eta_2\cdots\langle -\roarrow{p}_1^{'},J_1^{'},m_1^{'}; -\roarrow{p}_2^{'},J_2^{'},m_2^{'};\cdots |{\bf S}|-\roarrow{p}_1,J_1,m_1;-\roarrow{p}_2,J_2,m_2;\cdots\rangle, \end{array}$$ where ${\bf S}$ is the S-matrix operator. The space reflection matrix is defined as [@ChungSF71] $$\left( {P^\mu}_{\nu} \right) = {\rm diag}\{1, -1, -1, -1\},$$ we have $$\begin{aligned} \bar{p}_i^\mu & \equiv & {P^\mu}_{\nu}p_i^\nu = \left(E_i, -\roarrow{p}_i\right),\\ \bar{p}_i^{'\mu} & \equiv & {P^\mu}_{\nu}p_i^{'\nu} = \left(E^{'}_i, -\roarrow{p}^{'}_i\right),\\ \bar{g}^{\mu\nu} & \equiv & {P^\mu}_{\alpha} {P^\nu}_{\beta} g^{\alpha\beta} = g^{\mu\nu},\\ \bar{\varepsilon}^{\alpha\beta\gamma\delta} & \equiv & {P^\alpha}_{\alpha^{'}}{P^\beta}_{\beta^{'}} {P^\gamma}_{\gamma^{'}}{P^\delta}_{\delta^{'}} \varepsilon^{\alpha^{'}\beta^{'}\gamma^{'}\delta^{'}} = - \varepsilon^{\alpha\beta\gamma\delta},\\ \bar{e}^{\mu_1\mu_2\cdots\mu_j} & \equiv & {P^{\mu_1}}_{\nu_1}{P^{\mu_2}}_{\nu_2}\cdots{P^{\mu_j}}_{\nu_j} e^{\nu_1\nu_2\cdots \nu_j}.\end{aligned}$$ Eq. [(\[eq:pp\])]{} reads $$\begin{array}{rl} & e^{\ast}_c(p_1^{'},J_1^{'},m_1^{'})e^{\ast}_c(p_2^{'},J_2^{'},m_2^{'})\cdots e_c(p_1,J_1,m_1)e_c(p_2,J_2,m_2)\cdots\times\\ & \times\Gamma(p_1^{'},p_2^{'},\cdots,p_1,p_2,\cdots,g^{\mu\nu}, \varepsilon^{\alpha\beta\gamma\delta})\\ = & {\eta_1^{'}}^{\ast}{\eta_2^{'}}^{\ast}\cdots\eta_1\eta_2\cdots e^{\ast}_c(\bar{p}_1^{'},J_1^{'},m_1^{'}) e^{\ast}_c(\bar{p}_2^{'},J_2^{'},m_2^{'})\cdots e_c(\bar{p}_1,J_1,m_1)e_c(\bar{p}_2,J_2,m_2)\cdots\times\\ & \times\Gamma(\bar{p}_1^{'},\bar{p}_2^{'},\cdots, \bar{p}_1,\bar{p}_2,\cdots,g^{\mu\nu}, \varepsilon^{\alpha\beta\gamma\delta}). \end{array}$$ From Eq. [(\[eq:wfc\])]{} one can see $$e_c^\mu (\bar{p}, m) = - \bar{e}_c^\mu (p, m),$$ and for spin-$j$ canonical wave functions, $$e_c^{\mu_1\mu_2\cdots\mu_j} (\bar{p},j,m) = (-1)^j \bar{e}_c^{\mu_1\mu_2\cdots\mu_j} (p,j,m).$$ The requirement of parity conservation becomes $$\begin{array}{rl} & e^{\ast}_c(p_1^{'},J_1^{'},m_1^{'})e^{\ast}_c(p_2^{'},J_2^{'},m_2^{'})\cdots e_c(p_1,J_1,m_1)e_c(p_2,J_2,m_2)\cdots\times\\ & \times\Gamma(p_1^{'},p_2^{'},\cdots,p_1,p_2,\cdots,g^{\mu\nu}, \varepsilon^{\alpha\beta\gamma\delta})\\ = & {\eta_1^{'}}^{\ast}{\eta_2^{'}}^{\ast}\cdots\eta_1\eta_2\cdots (-1)^{J_1+J_2+\cdots-J_1^{'}-J_2^{'}-\cdots} e^{\ast}_c(p_1^{'},J_1^{'},m_1^{'}) e^{\ast}_c(p_2^{'},J_2^{'},m_2^{'})\cdots e_c(p_1,J_1,m_1)e_c(p_2,J_2,m_2)\cdots\times\\ & \times\Gamma(p_1^{'},p_2^{'},\cdots, p_1,p_2,\cdots,\bar{g}^{\mu\nu}, \bar{\varepsilon}^{\alpha\beta\gamma\delta}) \end{array}$$ or $$\begin{array}{rl} &\Gamma(p_1^{'},p_2^{'},\cdots,p_1,p_2,\cdots,g^{\mu\nu}, \varepsilon^{\alpha\beta\gamma\delta})\\ = & {\eta_1^{'}}^{\ast}{\eta_2^{'}}^{\ast}\cdots\eta_1\eta_2\cdots (-1)^{J_1+J_2+\cdots-J_1^{'}-J_2^{'}-\cdots} \Gamma(p_1^{'},p_2^{'},\cdots,p_1,p_2,\cdots,g^{\mu\nu}, -\varepsilon^{\alpha\beta\gamma\delta}). \end{array}$$ That is, if space reflection parity is conserved, the effective vertexes will consist of only tensors for the case $${\eta_1^{'}}^{\ast}{\eta_2^{'}}^{\ast}\cdots\eta_1\eta_2\cdots (-1)^{J_1+J_2+\cdots-J_1^{'}-J_2^{'}-\cdots} = 1,$$ and only pseudo-tensors if $${\eta_1^{'}}^{\ast}{\eta_2^{'}}^{\ast}\cdots\eta_1\eta_2\cdots (-1)^{J_1+J_2+\cdots-J_1^{'}-J_2^{'}-\cdots} = - 1.$$ [*Mixing of tensor and pseudo-tensor vertexes always means violation of space reflection symmetry.*]{} One can also prove the same result using helicity wave functions [@Zhua23pi]. We can see from Eq. [(\[eq:00j-\])]{} that, if space reflection parity is conserved, a spin-$j$ particle decaying into two scalar(or pseudo-scalar) particles must has the parity of $(-1)^j$, while a particle decaying into a scalar and a pseudo-scalar particle has the parity of $-(-1)^j$. Similarly, for spin-$0$ $\rightarrow$ spin-$j$ + spin-$0$, the parity of the spin-$j$ particle will be $(-1)^j$ (or $-(-1)^j$) if the two spin-$0$ particles have the same ( or opposite) parities. Boson symmetry {#sec:bs} ============== For two identical bosons, we have $$\left[ a^\dagger( \roarrow{p}, j, \sigma ), a^\dagger( \roarrow{p}^{'}, j, \sigma^{'} ) \right] =0,$$ $$\left|\cdots ;\roarrow{p},j,\sigma; \cdots; \roarrow{p}^{'},j,\sigma^{'}; \cdots \right> = \left|\cdots ;\roarrow{p}^{'},j,\sigma^{'}; \cdots; \roarrow{p},j,\sigma; \cdots \right>.$$ This demands[^5] $$\label{eq:bs} \Gamma^{\cdots \mu_1\mu_2\cdots\mu_j\cdots \nu_1\nu_2\cdots\nu_j\cdots} = \Gamma^{\cdots \nu_1\nu_2\cdots\nu_j\cdots \mu_1\mu_2\cdots\mu_j\cdots}.$$ Suppose particle 1 and particle 2 are identical particles. We define $$p=p_1+p_2,\;k=p_1-p_2,$$ $$\begin{aligned} A_+^{\alpha\beta\gamma} & = & k_\mu \varepsilon^{\mu\alpha\beta\gamma},\\ A_-^{\alpha\beta\gamma} & = & p_\mu \varepsilon^{\mu\alpha\beta\gamma}.\end{aligned}$$ Let’s first separate the 3-leg effective vertexes given in Sec. \[sec:3v\] into 1-2 symmetric parts $\Gamma_\pm^{(S)}$ and 1-2 anti-symmetric parts $\Gamma_\pm^{(A)}$: - $(J_1, J_2, J_3) = (0, 0, j)$, with $j$ an even number $$\begin{aligned} \Gamma_+^{(S)} & = & c_1 (k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j}),\\ \Gamma_+^{(A)} & = & 0, \\ \Gamma_-^{(S,A)} & = & 0. \end{aligned}$$ - $(J_1, J_2, J_3) = (0, 0, j)$, with $j$ an odd number Similar to the previous case, with $\Gamma_\pm^{(S)}\longleftrightarrow \Gamma_\pm^{(A)}$: $$\begin{aligned} \Gamma_+^{(S)} & = & 0, \label{eq:00j.1}\\ \Gamma_+^{(A)} & = & c_1 (k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j}),\\ \Gamma_-^{(S,A)} & = & 0. \label{eq:00j.2}\end{aligned}$$ - $(J_1, J_2, J_3) = (j, j, 0)$, with $j \geq 1$ $$\begin{array}{rcl} \Gamma_+^{(S)} & = & \;\;\; c_1 \left( g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}\cdots g^{\alpha_j\beta_j} \right)\\ & & + c_2 \left( g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j\beta_j} \right)\\ & & + \cdots \\ & & + c_{j+1} \left( p^{\alpha_1}p^{\alpha_2}\cdots p^{\alpha_j} p^{\beta_1}p^{\beta_2}\cdots p^{\beta_j} \right), \end{array}$$ $$\Gamma_+^{(A)} = 0,$$ $$\begin{array}{rcrl} \Gamma_-^{(S)} & = & Q^{\alpha_1\beta_1} & \left\{ \; c_{j+2} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right. \\ & & & +c_{j+3} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{2j+1} \left( p^{\alpha_1}p^{\alpha_2}\cdots p^{\alpha_j} p^{\beta_1}p^{\beta_2}\cdots p^{\beta_j} \right)\right\}, \end{array}$$ $$\Gamma_-^{(A)} = 0.$$ - $(J_1, J_2, J_3) = (1, 1, j)$, with $j$ an even number and $j \geq 2$ $$\begin{array}{rcl} \Gamma_+^{(S)} & = & \;\;\;c_1 g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2} \left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_2 \left( g^{\alpha_1\gamma_1}p^{\beta_1} -g^{\beta_1\gamma_1}p^{\alpha_1} \right) \left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right) \\ & & +c_3 g^{\alpha_1\beta_1} \left( k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j} \right)\\ & & +c_4 p^{\alpha_1}p^{\beta_1} \left( k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j} \right), \end{array}$$ $$\Gamma_+^{(A)} = c_5 \left( g^{\alpha_1\gamma_1}p^{\beta_1} +g^{\beta_1\gamma_1}p^{\alpha_1} \right) \left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right),$$ $$\begin{array}{rcl} \Gamma_-^{(S)} & = & \;\;\; c_6 A_-^{\alpha_1\beta_1\gamma_1}\left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right)\\ & & +c_7 Q^{\alpha_1\beta_1}\left( k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j} \right), \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(A)} & = & \;\;\; c_8 \left( Q^{\alpha_1\gamma_1}g^{\beta_1\gamma_2} + Q^{\beta_1\gamma_1}g^{\alpha_1\gamma_2} \right) \left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_9 A_+^{\alpha_1\beta_1\gamma_1}\left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right). \end{array}$$ - $(J_1, J_2, J_3) = (1, 1, j)$, with $j$ an odd number and $j \geq 3$ Similar to the previous case. Just interchange the 1-2 symmetric vertexes with the 1-2 anti-symmetric ones. - $(J_1, J_2, J_3) = (j, j, 1)$, with $j \geq 1$ $$\begin{array}{rcrl} \Gamma_+^{(S)} & = & \left( g^{\alpha_1\gamma_1}p^{\beta_1} + g^{\beta_1\gamma_1}p^{\alpha_1} \right) & \left\{ \; c_1 \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right. \\ & & & +c_2 \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right) \\ & & & +\cdots \\ & & & \left. + c_j \left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}, \end{array}$$ $$\begin{array}{rcrl} \Gamma_+^{(A)} & = & \left( g^{\alpha_1\gamma_1}p^{\beta_1} - g^{\beta_1\gamma_1}p^{\alpha_1} \right) & \left\{ \; c_{j+1} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right. \\ & & & +c_{j+2} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right) \\ & & & +\cdots \\ & & & \left. + c_{2j} \left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}\\ & & +k^{\gamma_1} & \left\{ c_{2j+1} \left( g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}\cdots g^{\alpha_j\beta_j} \right)\right. \\ & & & +c_{2j+2}\left( g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right) \\ & & & +\cdots \\ & & & \left. +c_{3j+1}\left( p^{\alpha_1}p^{\alpha_2}\cdots p^{\alpha_j} p^{\beta_1}p^{\beta_2}\cdots p^{\beta_j} \right)\right\}, \end{array}$$ $$\begin{array}{rcrl} \Gamma_-^{(S)} & = & A_+^{\alpha_1\beta_1\gamma_1} & \left\{ \; c_{3j+2} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right. \\ & & & +c_{3j+3} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{4j+1} \left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}, \end{array}$$ $$\begin{array}{rcrl} \Gamma_-^{(A)} & = & A_-^{\alpha_1\beta_1\gamma_1} & \left\{ \; c_{4j+2} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right. \\ & & & +c_{4j+3} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{5j+1} \left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}\\ & & +Q^{\alpha_1\beta_1}k^{\gamma_1} & \left\{ \; c_{5j+2} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & & +c_{5j+3} \left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & & +\cdots \\ & & & \left. +c_{6j+1} \left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}. \end{array}$$ - $(J_1, J_2, J_3) = (2, 2, 2)$ $$\begin{array}{rcl} \Gamma_+^{(S)} & = & \;\;\; c_1 g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}k^{\gamma_1}k^{\gamma_2} \\ & & + c_2 g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1}g^{\beta_2\gamma_2} \\ & & + c_3 g^{\alpha_1\beta_1} \left( g^{\alpha_2\gamma_1}p^{\beta_2} - g^{\beta_2\gamma_1}p^{\alpha_2} \right) k^{\gamma_2} \\ & & + c_4 g^{\alpha_1\beta_1}p^{\alpha_2}p^{\beta_2}k^{\gamma_1}k^{\gamma_2} \\ & & + c_5 \left( g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}p^{\beta_1}p^{\beta_2} + g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p^{\alpha_1}p^{\alpha_2} \right)\\ & & + c_6 g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2}p^{\alpha_2}p^{\beta_2} \\ & & + c_7 p^{\alpha_1}p^{\alpha_2}p^{\beta_1}p^{\beta_2} k^{\gamma_1}k^{\gamma_2}, \end{array}$$ $$\begin{array}{rcl} \Gamma_+^{(A)} & = & \;\;\; c_8 g^{\alpha_1\beta_1} \left( g^{\alpha_2\gamma_1}p^{\beta_2} + g^{\beta_2\gamma_1}p^{\alpha_2} \right) k^{\gamma_2} \\ & & +c_9 \left( g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}p^{\beta_1}p^{\beta_2} - g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p^{\alpha_1}p^{\alpha_2} \right) \\ & & + c_{10} \left( g^{\alpha_1\gamma_1}p^{\beta_1} + g^{\beta_1\gamma_1}p^{\alpha_1} \right) p^{\alpha_2}p^{\beta_2}k^{\gamma_2}, \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(S)} & = & \;\;\; c_{11} g^{\alpha_1\beta_1}A_-^{\alpha_2\beta_2\gamma_1}k^{\gamma_2}\\ & & +c_{12} g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2}k^{\gamma_1}k^{\gamma_2}\\ & & +c_{13}A_-^{\alpha_1\beta_1\gamma_1}\left( p^{\alpha_2}g^{\beta_2\gamma_2} -p^{\beta_2}g^{\alpha_2\gamma_2} \right) \\ & & +c_{14} A_+^{\alpha_1\beta_1\gamma_1}\left( p^{\alpha_2}g^{\beta_2\gamma_2} +p^{\beta_2}g^{\alpha_2\gamma_2} \right)\\ & & +c_{15} Q^{\alpha_1\beta_1} p^{\alpha_2}p^{\beta_2}k^{\gamma_1}k^{\gamma_2}, \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(A)} & = & \;\;\; c_{16} g^{\alpha_1\beta_1}\left( Q^{\alpha_2\gamma_1}g^{\beta_2\gamma_2} + Q^{\beta_2\gamma_1}g^{\alpha_2\gamma_2} \right)\\ & & +c_{17} g^{\alpha_1\beta_1}A_+^{\alpha_2\beta_2\gamma_1}k^{\gamma_2}\\ & & +c_{18} A_-^{\alpha_1\beta_1\gamma_1}\left( p^{\alpha_2}g^{\beta_2\gamma_2} +p^{\beta_2}g^{\alpha_2\gamma_2} \right)\\ & & +c_{19} A_+^{\alpha_1\beta_1\gamma_1}p^{\alpha_2}p^{\beta_2}k^{\gamma_2}. \end{array}$$ - $(J_1, J_2, J_3) = (2, 2, 3)$ $$\begin{array}{rcl} \Gamma_+^{(S)} & = & \;\;\; c_1 g^{\alpha_1\beta_1}\left( g^{\alpha_2\gamma_1}p^{\beta_2} +g^{\beta_2\gamma_1}p^{\alpha_2} \right)k^{\gamma_2}k^{\gamma_3}\\ & & +c_2 g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2}\left( g^{\alpha_2\gamma_3}p^{\beta_2} +g^{\beta_2\gamma_2}p^{\alpha_2} \right)\\ & & +c_3 \left( g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}p^{\beta_1}p^{\beta_2} -g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p^{\alpha_1}p^{\alpha_2} \right) k^{\gamma_3}\\ & & +c_4 \left( g^{\alpha_1\gamma_1}p^{\beta_1} +g^{\beta_1\gamma_1}p^{\alpha_1} \right)p^{\alpha_2}p^{\beta_2}k^{\gamma_2}k^{\gamma_3}, \end{array}$$ $$\begin{array}{rcl} \Gamma_+^{(A)} & = & \;\;\; c_5 g^{\alpha_1\beta_1}g^{\alpha_2\beta_2} k^{\gamma_1}k^{\gamma_2}k^{\gamma_3} \\ & & +c_6 g^{\alpha_1\beta_1}g^{\alpha_2\gamma_1}g^{\beta_2\gamma_2} k^{\gamma_3} \\ & & +c_7 g^{\alpha_1\beta_1}\left( g^{\alpha_2\gamma_1}p^{\beta_2} -g^{\beta_2\gamma_1}p^{\alpha_2} \right)k^{\gamma_2}k^{\gamma_3}\\ & & +c_8 g^{\alpha_1\beta_1}p^{\alpha_2}p^{\beta_2} k^{\gamma_1} k^{\gamma_2} k^{\gamma_3}\\ & & +c_9 g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2}\left( g^{\alpha_2\gamma_3}p^{\beta_2} -g^{\beta_2\gamma_2}p^{\alpha_2} \right)\\ & & +c_{10} \left( g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}p^{\beta_1}p^{\beta_2} +g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p^{\alpha_1}p^{\alpha_2} \right) k^{\gamma_3}\\ & & +c_{11} g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2} p^{\alpha_2}p^{\beta_2}k^{\gamma_3}\\ & & +c_{12} p^{\alpha_1}p^{\alpha_2}p^{\beta_1}p^{\beta_2} k^{\gamma_1}k^{\gamma_2}k^{\gamma_3}, \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(S)} & = & \;\;\; c_{13} g^{\alpha_1\beta_1}A_+^{\alpha_1\beta_2\gamma_1} k^{\gamma_2}k^{\gamma_3}\\ & & +c_{14} g^{\alpha_1\beta_1}\left( Q^{\alpha_2\gamma_1}g^{\beta_2\gamma_2} +Q^{\beta_2\gamma_1}g^{\alpha_2\gamma_2} \right)k^{\gamma_3} \\ & & +c_{15} A_+^{\alpha_1\beta_1\gamma_1} g^{\alpha_2\gamma_2}g^{\beta_2\gamma_3}\\ & & +c_{16} A_-^{\alpha_1\beta_1\gamma_1}\left( g^{\alpha_2\gamma_2}p^{\beta_2} +g^{\beta_2\gamma_2}p^{\alpha_2} \right)k^{\gamma_3} \\ & & +c_{17} A_+^{\alpha_1\beta_1\gamma_1} p^{\alpha_2}p^{\beta_2}k^{\gamma_2}k^{\gamma_3}, \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(A)} & = & \;\;\; c_{18} g^{\alpha_1\beta_1}A_-^{\alpha_2\beta_2\gamma_1} k^{\gamma_2}k^{\gamma_3}\\ & & +c_{19} g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2} k^{\gamma_1}k^{\gamma_2}k^{\gamma_3} \\ & & +c_{20} A_-^{\alpha_1\beta_1\gamma_1} g^{\alpha_2\gamma_2}g^{\beta_2\gamma_3}\\ & & +c_{21} A_-^{\alpha_1\beta_1\gamma_1}\left( g^{\alpha_2\gamma_2}p^{\beta_2} -g^{\beta_2\gamma_2}p^{\alpha_2} \right)k^{\gamma_3}\\ & & +c_{22} A_+^{\alpha_1\beta_1\gamma_1}\left( g^{\alpha_2\gamma_2}p^{\beta_2} +g^{\beta_2\gamma_2}p^{\alpha_2} \right)k^{\gamma_3}\\ & & +c_{23} Q^{\alpha_1\beta_1}p^{\alpha_2}p^{\beta_2} k^{\gamma_1}k^{\gamma_2}k^{\gamma_3}. \end{array}$$ - $(J_1, J_2, J_3) = (2, 2, j)$, with $j$ an even number and $j \geq 4$ $$\begin{array}{rcl} \Gamma_+^{(S)} & = & \;\;\; c_1 g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}\left( k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j} \right)\\ & & +c_2 g^{\alpha_1\beta_1} g^{\alpha_2\gamma_1}g^{\beta_2\gamma_2}\left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_3 g^{\alpha_1\beta_1}\left( g^{\alpha_2\gamma_1}p^{\beta_2} -g^{\beta_2\gamma_1}p^{\alpha_2} \right)\left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right)\\ & & +c_4 g^{\alpha_1\beta_1}p^{\alpha_2}p^{\beta_2}\left( k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j} \right)\\ & & +c_5 g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}g^{\beta_1\gamma_3}g^{\beta_2\gamma_4} \left( k^{\gamma_5}k^{\gamma_6}\cdots k^{\gamma_j} \right)\\ & & +c_6 g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2}\left( g^{\alpha_2\gamma_3}p^{\beta_2} -g^{\beta_2\gamma_3}p^{\alpha_2} \right)\left( k^{\gamma_4}k^{\gamma_5}\cdots k^{\gamma_j} \right)\\ & & +c_7 g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2} p^{\alpha_2}p^{\beta_2}\left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_8 \left( g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}p^{\beta_1}p^{\beta_2} +g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p^{\alpha_1}p^{\alpha_2} \right)\left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_9 p^{\alpha_1}p^{\alpha_2}p^{\beta_1}p^{\beta_2}\left( k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j} \right), \end{array}$$ $$\begin{array}{rcl} \Gamma_+^{(A)} & = & \;\;\; c_{10} g^{\alpha_1\beta_1}\left( g^{\alpha_2\gamma_1}p^{\beta_2} +g^{\beta_2\gamma_1}p^{\alpha_2} \right)\left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right)\\ & & +c_{11} g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2}\left( g^{\alpha_2\gamma_3}p^{\beta_2} +g^{\beta_2\gamma_3}p^{\alpha_2} \right)\left( k^{\gamma_4}k^{\gamma_5}\cdots k^{\gamma_j} \right)\\ & & +c_{12} \left( g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}p^{\beta_1}p^{\beta_2} -g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p^{\alpha_1}p^{\alpha_2} \right)\left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_{13} \left( g^{\alpha_1\gamma_1}p^{\beta_1} +g^{\beta_1\gamma_1}p^{\alpha_1} \right)p^{\alpha_2}p^{\beta_2}\left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right), \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(S)} & = & \;\;\; c_{14} g^{\alpha_1\beta_1}A_-^{\alpha_2\beta_2\gamma_1}\left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right)\\ & & +c_{15} g^{\alpha_1\beta_1}Q^{\alpha_2\beta_2}\left( k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j} \right)\\ & & +c_{16} A_-^{\alpha_1\beta_1\gamma_1} g^{\alpha_2\gamma_2}g^{\beta_2\gamma_3}\left( k^{\gamma_4}k^{\gamma_5}\cdots k^{\gamma_j} \right)\\ & & +c_{17} A_+^{\alpha_1\beta_1\gamma_1}\left( g^{\alpha_2\gamma_2}p^{\beta_2} +g^{\beta_2\gamma_2}p^{\alpha_2} \right)\left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_{18} A_-^{\alpha_1\beta_1\gamma_1}\left( g^{\alpha_2\gamma_2}p^{\beta_2} -g^{\beta_2\gamma_2}p^{\alpha_2} \right)\left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_{19} Q^{\alpha_1\beta_1}p^{\alpha_2}p^{\beta_2}\left( k^{\gamma_1}k^{\gamma_2}\cdots k^{\gamma_j} \right), \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(A)} & = & \;\;\; c_{20} g^{\alpha_1\beta_1}A_+^{\alpha_2\beta_2\gamma_1}\left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right)\\ & & +c_{21} g^{\alpha_1\beta_1}\left( Q^{\alpha_2\gamma_1}g^{\beta_2\gamma_2} +Q^{\beta_2\gamma_1}g^{\alpha_2\gamma_2} \right)\left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_{22} A_+^{\alpha_1\beta_1\gamma_1} g^{\alpha_1\gamma_2}g^{\beta_2\gamma_3}\left( k^{\gamma_4}k^{\gamma_5}\cdots k^{\gamma_j} \right)\\ & & +c_{23} A_-^{\alpha_1\beta_1\gamma_1}\left( g^{\alpha_2\gamma_2}p^{\beta_2} +g^{\beta_2\gamma_2}p^{\alpha_2} \right)\left( k^{\gamma_3}k^{\gamma_4}\cdots k^{\gamma_j} \right)\\ & & +c_{24} A_+^{\alpha_1\beta_1\gamma_1} p^{\alpha_2}p^{\beta_2}\left( k^{\gamma_2}k^{\gamma_3}\cdots k^{\gamma_j} \right)\\ & & +c_{25} \left( Q^{\alpha_1\gamma_1}g^{\beta_1\gamma_2} +Q^{\beta_1\gamma_1}g^{\alpha_1\gamma_2} \right)g^{\alpha_2\gamma_3}g^{\beta_2\gamma_4}\left( k^{\gamma_5}k^{\gamma_6}\cdots k^{\gamma_j} \right). \end{array}$$ - $(J_1, J_2, J_3) = (2, 2, j)$, with $j$ an odd number and $j \geq 5$ Similar to the previous case, with $\Gamma_\pm^{(S)}\longleftrightarrow \Gamma_\pm^{(A)}$. - $(J_1, J_2, J_3) = (j, j, 2)$, with $j \geq 3$ $$\begin{array}{rcl} \Gamma_+^{(S)} & = & \left(g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}p^{\beta_1}p^{\beta_2} +g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p^{\alpha_1}p^{\alpha_2} \right) \times \\ & & \left\{c_1 \left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_2 \left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & +\cdots \\ & & \left. +c_{j-1}\left( p^{\alpha_3}p^{\alpha_4}\cdots p^{\alpha_j} p^{\beta_3}p^{\beta_4}\cdots p^{\beta_j} \right)\right\}\\ & & +g^{\alpha_1\gamma_1}g^{\beta_1\gamma_2}\times\\ & & \left\{c_{j}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{j+1}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & +\cdots \\ & & \left. +c_{2j-1}\left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}\\ & & \left(g^{\alpha_1\gamma_1}p^{\beta_1} +g^{\beta_1\gamma_1}p^{\alpha_1}\right)k^{\gamma_2}\times\\ & & \left\{c_{2j}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{2j+1}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & \cdots \\ & & \left. +c_{3j-1}\left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}\\ & & + c_{3j}k^{\gamma_1}k^{\gamma_2}\left( g^{\alpha_1\beta_1}g^{\alpha_2\beta_2}\cdots g^{\alpha_j\beta_j} \right)\\ & & + c_{3j+1}k^{\gamma_1}k^{\gamma_2}\left( p^{\alpha_1}p^{\alpha_2}\cdots p^{\alpha_j} p^{\beta_1}p^{\beta_2}\cdots p^{\beta_j} \right), \end{array}$$ $$\begin{array}{rcl} \Gamma_+^{(A)} & = & \left(g^{\alpha_1\gamma_1}g^{\alpha_2\gamma_2}p^{\beta_1}p^{\beta_2} -g^{\beta_1\gamma_1}g^{\beta_2\gamma_2}p^{\alpha_1}p^{\alpha_2} \right) \times \\ & & \left\{c_{3j+2} \left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{3j+3} \left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & +\cdots \\ & & \left. +c_{4j}\left( p^{\alpha_3}p^{\alpha_4}\cdots p^{\alpha_j} p^{\beta_3}p^{\beta_4}\cdots p^{\beta_j} \right)\right\}\\ & & \left(g^{\alpha_1\gamma_1}p^{\beta_1} -g^{\beta_1\gamma_1}p^{\alpha_1}\right)k^{\gamma_2}\times\\ & & \left\{c_{4j+1}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{4j+2}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & \cdots \\ & & \left. +c_{5j}\left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}, \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(S)} & = & A_-^{\alpha_1\beta_1\gamma_1}\left( p^{\alpha_2}g^{\beta_2\gamma_2} -p^{\beta_2}g^{\alpha_2\gamma_2} \right)\times\\ & & \left\{c_{5j+1}\left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{5j+2}\left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & +\cdots \\ & & \left. +c_{6j-1}\left( p^{\alpha_3}p^{\alpha_4}\cdots p^{\alpha_j} p^{\beta_3}p^{\beta_4}\cdots p^{\beta_j} \right)\right\}\\ & & +A_+^{\alpha_1\beta_1\gamma_1}\left( p^{\alpha_2}g^{\beta_2\gamma_2} +p^{\beta_2}g^{\alpha_2\gamma_2} \right)\times\\ & & \left\{c_{6j}\left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{6j+1}\left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & +\cdots \\ & & \left. +c_{7j-2}\left( p^{\alpha_3}p^{\alpha_4}\cdots p^{\alpha_j} p^{\beta_3}p^{\beta_4}\cdots p^{\beta_j} \right)\right\}\\ & & +c_{7j-1}A_-^{\alpha_1\beta_1\gamma_1}k^{\gamma_2}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\\ & & +Q^{\alpha_1\beta_1}k^{\gamma_1}k^{\gamma_2}\times\\ & & \left\{c_{7j}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{7j+1}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & +\cdots \\ & & \left. +c_{8j-1}\left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}, \end{array}$$ $$\begin{array}{rcl} \Gamma_-^{(A)} & = & A_-^{\alpha_1\beta_1\gamma_1}\left( p^{\alpha_2}g^{\beta_2\gamma_2} +p^{\beta_2}g^{\alpha_2\gamma_2} \right)\times\\ & & \left\{c_{8j}\left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{8j+1}\left( g^{\alpha_3\beta_3}g^{\alpha_4\beta_4}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & +\cdots \\ & & \left. +c_{9j-2}\left( p^{\alpha_3}p^{\alpha_4}\cdots p^{\alpha_j} p^{\beta_3}p^{\beta_4}\cdots p^{\beta_j} \right)\right\}\\ & & +A_+^{\alpha_1\beta_1\gamma_1}k^{\gamma_2}\times\\ & & \left\{c_{9j-1}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j} \right)\right.\\ & & +c_{9j}\left( g^{\alpha_2\beta_2}g^{\alpha_3\beta_3}\cdots g^{\alpha_{j-1}\beta_{j-1}} p^{\alpha_j}p^{\beta_j} \right)\\ & & +\cdots \\ & & \left. +c_{10j-2}\left( p^{\alpha_2}p^{\alpha_3}\cdots p^{\alpha_j} p^{\beta_2}p^{\beta_3}\cdots p^{\beta_j} \right)\right\}\\ & & +c_{10j-1} \left(Q^{\alpha_1\gamma_1}g^{\beta_1\gamma_2} +Q^{\beta_1\gamma_1}g^{\alpha_1\gamma_2}\right) \left(g^{\alpha_2\beta2}g^{\alpha_3\beta_3}\cdots g^{\alpha_j\beta_j}\right) \end{array}$$ and so on. We need to consider three cases: 1. Both of the two identical particles are internal lines According to Eq. [(\[eq:bs\])]{}, one can obtain $$\Gamma_\pm = \Gamma_\pm^{(S)} + (p_1^2-p_2^2)\Gamma_\pm^{(A)}$$ with the coefficients satisfying $$c_i = c_i\left(p_1^2+p_2^2, (p_1^2-p_2^2)^2, p_1\cdot p_2 \right).$$ 2. One is on shell, another an internal line The factor $(p_1^2-W^2)$ or $(W^2-p_2^2)$ will eliminate the denominator of the propagator for the off shell particle. This make the contribution from $\Gamma_\pm^{(A)}$ has no pole at the particle’s mass, so that it can be absorbed K. S. freely into background amplitudes. $$\Gamma_\pm = \Gamma_\pm^{(S)},$$ $$c_i = c_i\left(p_1^2, p_1\cdot p_2 \right).$$ 3. Both of the two identical particles are on shell Since $p_1^2-p_2^2 = 0$, we have $$\Gamma_\pm = \Gamma_\pm^{(S)}$$ with the coefficients $$c_i = c_i\left(p_1\cdot p_2 \right).$$ From Eqs. [(\[eq:00j.1\],\[eq:00j.2\])]{} one can infer that a particle decaying into two identical spin-$0$ particles must has a even spin. If parity is conserved, the parity of such a particle must be +1. For example, $\rho^0,\; \eta,\; \eta\prime,\; \omega,\; \phi,\; a_1,\; f_1$ etc. will not decay into two neutral pions[^6]. Examples for incorporating boson symmetry in 4-leg vertexes can be found in Sec. \[sec:bk\] and Ref. [@Zhua23pi]. Covariant helicity amplitudes for two-body decays {#sec:2bd} ================================================= Helicity amplitudes for two body-decays can be write down directly using the wave functions and vertexes given in previous sections. Such amplitudes can be calculated in arbitrary reference frame. Especially they can be calculated in laboratory frame so that no Lorentz transformation is needed. However, one might still favor amplitudes in center of mass frame(CM frame). We will give some explicit results calculated in the rest frame of parent particles in this section. Suppose a spin-$J$ particle with momentum $p$ decays into a spin-$s$ and a spin-$\sigma$ particle with momentum $q$ and $k$. The helicity amplitude of such a process is $$\begin{array}{rcl} {\cal M}_{\delta\lambda\nu}(p,q,k) & \equiv & \langle \roarrow{q},s,\lambda;\roarrow{k},\sigma,\nu|{\bf S} |\roarrow{p},J,\delta\rangle\\ & = & \sum\limits_{\lambda^{'}\nu^{'}} D^{s\ast}_{\lambda^{'}\lambda}\left( W(L^{-1}(p),q) \right) D^{\sigma\ast}_{\nu^{'}\nu}\left( W(L^{-1}(p),k) \right) \langle \roarrow{q}^{'},s,\lambda^{'};\roarrow{k}^{'},\sigma^{'},\nu|{\bf S} |\roarrow{0},J,\delta\rangle \end{array}$$ with $$\begin{aligned} q^{'\alpha} = {L^{-1 \alpha}}_\beta(p)q^\beta,\\ k^{'\alpha} = {L^{-1 \alpha}}_\beta(p)k^\beta.\end{aligned}$$ Here Eq. [(\[eq:lrt\])]{} has been used. Notice that we should use $L(p)$ defined in Eq. [(\[eq:lh\])]{}. From Eqs. [(\[eq:rot1\],\[eq:rot2\])]{} one can rotate $\roarrow{q}^{'}$ to the direction of $z$-axis: $$\label{eq:rel} {\cal M}_{\delta\lambda\nu}(p,q,k) = \sum\limits_{\lambda^{'}\nu^{'}} D^{s\ast}_{\lambda^{'}\lambda}\left( W(L^{-1}(p),q) \right) D^{\sigma\ast}_{\nu^{'}\nu}\left( W(L^{-1}(p),k) \right) D^{J\ast}_{\lambda^{'}-\nu{'},\delta}(\varphi,\vartheta,0) F_{\lambda^{'}\nu^{'}},$$ where $(\vartheta, \varphi)$ is the direction of $\roarrow{q}^{'}$. $F_{\lambda\nu}$ is the helicity amplitude in the rest frame of the parent particle, $$F_{\lambda\nu} = \langle q,s,\lambda; k,\sigma,\nu|{\bf S}| p,J,\lambda-\nu \rangle _{CM} .$$ $p,q,k$ have been redefined to their CM frame values in the above equation. We follow the convention of Chung [@ChungHA95] in this section: $$\begin{array}{cclccr} (p^\alpha) & = & ( W; & 0, & 0, & 0 ),\\ (q^\alpha) & = & ( q_0; & 0, & 0, & r/2 ),\\ (k^\alpha) & = & ( k_0; & 0, & 0, & -r/2 ). \end{array}$$ Eq. [(\[eq:rel\])]{} is the relation between helicity amplitudes in laboratory frame and those in CM frame. It can be derived in a alternative way by writing down the explicit expressions for these amplitudes and use Lorentz transformation properties of wave functions in Eq. [(\[eq:wltr\])]{}. The masses of the daughter particles are $m$ and $\mu$, $$\begin{aligned} W & = & q_0 + k_0 , \\ q_0 & = & \sqrt{m^2 + \frac{r^2}{4}},\\ k_0 & = & \sqrt{\mu^2 + \frac{r^2}{4}}.\end{aligned}$$ The corresponding space reflection parity of the three particles are $\eta_J$, $\eta_s$ and $\eta_\sigma$. Parity conserving helicity amplitudes in CM frame satisfy [@ChungTA93; @ChungSF71] $$F_{\lambda\nu} = \eta_J \eta_s \eta_\sigma (-1)^{J-s-\sigma} F_{-\lambda,-\nu}.$$ If the two daughter particles are identical, one has [@JacobH59; @ChungTA93; @ChungSF71] $$F_{\lambda\nu} = (-1)^J F_{\nu\lambda}.$$ Some explicit results for $F_{\lambda\nu}$ are listed below. We assume space reflection symmetry in all processes. - Spin-1 $\longrightarrow$ spin-0 + spin-1, $\eta_J \eta_s \eta_\sigma = -1$ We should choose pseudo-tensor effective vertexes. $$F_{\lambda\nu} = - F_{-\lambda,-\nu} ;$$ $$F_{01} = \frac{i}{2} c W r.$$ Here $c$ is a scalar. - Spin-1 $\longrightarrow$ spin-0 + spin-1, $\eta_J \eta_s \eta_\sigma = +1$ The effective vertex should be a tensor. $$F_{\lambda\nu} = + F_{-\lambda,-\nu};$$ $$\begin{aligned} F_{01} & = & -\frac{k_0}{\mu}c_2 + \frac{W}{4\mu}c_1 r^2,\\ F_{00} & = & -c_2.\end{aligned}$$ It can be applied to the process $a_1(1260)\longrightarrow \pi\rho$. - Spin-1 $\longrightarrow$ spin-2 + spin-1, $\eta_J \eta_s \eta_\sigma = +1$ The effective vertex is a tensor. $$F_{\lambda\nu} = + F_{-\lambda,-\nu};$$ $$\begin{aligned} F_{00} & = & \sqrt{\frac{2}{3}}\frac{k_0 q_0^2}{\mu m^2}c_5 +\frac{1}{2\sqrt{6}\mu m^2}\left(c_5 q_0 + c_3 k_0 q_0 W -c_2 k_0 W^2 + c_4 q_0 W^2 \right)r^2 +\frac{W}{8\sqrt{6}\mu m^2}\left(c_3 + c_1 W^2 \right)r^4,\\ F_{10} & = & \frac{k_0 q_0}{\sqrt{2}\mu m}c_5 + \frac{1}{4\sqrt{2}\mu m}\left(c_5 + c_4 W^2 \right)r^2,\\ F_{01} & = & -\frac{1}{\sqrt{6}}c_5 -\frac{W^2}{2\sqrt{6}m^2}c_2 r^2,\\ F_{11} & = & - \frac{q_0}{\sqrt{2} m} c_5 -\frac{W}{2\sqrt{2} m} c_3 r^2,\\ F_{21} & = & -c_5.\end{aligned}$$ These amplitudes are applicable to the case $J/\psi\longrightarrow a_2(1320) \rho$. - Spin-1 $\longrightarrow$ spin-4 + spin-1, $\eta_J \eta_s \eta_\sigma = +1$ Use the tensor vertexes given in Sec. [\[sec:3v\]]{}. $$F_{\lambda\nu} = + F_{-\lambda,-\nu};$$ $$\begin{aligned} F_{00} & = & \frac{k_0 q_0 W^2}{\sqrt{70} m^4}c_5 r^2 +\frac{W^2}{\sqrt{1120} m^4}\left( c_5 q_0 + c_3 k_0 q_0 W - c_2 k_0 W^2 + c_4 q_0 W^2 \right) r^4 + \frac{W^3}{\sqrt{17920} m^4} \left(c_3 + c_1 W^2 \right) r^6,\\ F_{01} & = & -\frac{W^2}{\sqrt{280}m^2}c_5 r^2 -\frac{W^4}{\sqrt{1120}m^4}c_4 r^4,\\ F_{10} & = & \frac{k_0 q_0 W^2}{4\sqrt{7}m^3}c_5 r^2 +\frac{W^2}{16\sqrt{7}m^3}\left(c_5 +c_4 W^2 \right) r^4,\\ F_{11} & = & -\frac{q_0 W^2}{4\sqrt{7}m^3}c_5 r^2 -\frac{W^3}{16\sqrt{7}m^3}c_3 r^4,\\ F_{21} & = & -\frac{W^2}{4\sqrt{7}m^2}c_5 r^2.\end{aligned}$$ - Spin-0 $\longrightarrow$ spin-1 + spin-1, $\eta_J \eta_s \eta_\sigma = -1$ The effective vertex is a pseudo scalar. $$F_{\lambda\nu} = - F_{-\lambda,-\nu};$$ $$F_{11} = i g W r.$$ These amplitudes automatically satisfy $$F_{\lambda\nu} = F_{\nu\lambda}.$$ - Spin-0 $\longrightarrow$ spin-1 + spin-1, $\eta_J \eta_s \eta_\sigma = +1$ The vertex is a tensor. $$F_{\lambda\nu} = + F_{-\lambda,-\nu};$$ $$\begin{aligned} F_{00} & = & -\frac{k_0 q_0}{m \mu}c_2 -\frac{1}{4m\mu}\left(c_2 + c_1 W^2 \right)r^2,\\ F_{11} & = & c_2.\end{aligned}$$ - Spin-0 $\longrightarrow$ spin-0 + spin-1, $\eta_J \eta_s \eta_\sigma = -1$ The vertex should be a tensor. $$F_{00} = -\frac{W}{2\mu}c .$$ - Spin-2 $\longrightarrow$ spin-0 + spin-0, $\eta_J \eta_s \eta_\sigma = +1$ The vertex should be a tensor, $$F_{00} = -\frac{1}{\sqrt{24}}c r^2 .$$ It can be applied to the decay $f_2(1270)\longrightarrow \pi^+ \pi^-$. - Spin-2 $\longrightarrow$ two identical spin-1 particles, $\eta_J = +1$ The vertex should satisfy boson symmetry. $$\begin{aligned} F_{\lambda\nu} & = & + F_{-\lambda,-\nu},\\ F_{\lambda\nu} & = & F_{\nu\lambda}; \end{aligned}$$ $$\begin{aligned} F_{00} & = & \sqrt{\frac{2}{3}}\frac{q_0^2}{m^2}c_4 +\frac{1}{\sqrt{6}m^2}\left( \frac{c_2}{2}q_0^2 - c_3 q_0 W \right) r^2 +\frac{1}{\sqrt{384}m^2}\left( c_2 + c_1 W^2 \right) r^4,\\ F_{01} & = & \frac{q_0}{\sqrt{2}m}c_4 -\frac{W}{4\sqrt{2}m}c_3 r^2,\\ F_{1,-1} & = & c_4 ,\\ F_{11} & = & \frac{1}{\sqrt{6}}c_4 -\frac{1}{2\sqrt{6}}c_2 r^2 .\end{aligned}$$ - Spin-2 $\longrightarrow$ spin-2 + spin-0, $\eta_J \eta_s \eta_\sigma = +1$ The effective vertex should be a tensor. $$F_{\lambda\nu} = + F_{-\lambda,-\nu};$$ $$\begin{aligned} F_{00} & = & \frac{1}{3}\left(1+\frac{2q_0^2}{m^2}\right)c_3 -\frac{q_0 W}{6 m^2}c_2 r^2 +\frac{W^2}{24 m^2}c_1 r^4 ,\\ F_{10} & = & \frac{q_0}{m}c_3 - \frac{W}{8 m}c_2 r^2,\\ F_{20} & = & c_3 .\end{aligned}$$ - Spin-4 $\longrightarrow$ two identical spin-1 particles, $\eta_J = +1$ The vertex should satisfy boson symmetry. $$\begin{aligned} F_{\lambda\nu} & = & + F_{-\lambda,-\nu},\\ F_{\lambda\nu} & = & F_{\nu\lambda}; \end{aligned}$$ $$\begin{aligned} F_{00} & = & -\frac{q_0^2}{\sqrt{70}m^2}c_4 r^2 -\frac{1}{\sqrt{1120}m^2}\left( c_2 q_0^2 + 2 c_3 q_0 W \right)r^4 -\frac{1}{\sqrt{17920}m^2}\left(c_2 + c_1 W^2 \right) r^6 ,\\ F_{01} & = & -\frac{q_0}{4\sqrt{7}m}c_4 r^2 -\frac{W}{16\sqrt{7}m}c_3 r^4 ,\\ F_{1,-1} & = & -\frac{1}{4\sqrt{7}}c_4 r^2,\\ F_{11} &= & -\frac{1}{\sqrt{280}}c_4 r^2 +\frac{1}{\sqrt{1120}}c_2 r^4 .\end{aligned}$$ - Spin-4 $\longrightarrow$ spin-2 + spin-0, $\eta_J \eta_s \eta_\sigma = +1$ The effective vertex is a tensor. $$F_{\lambda\nu} = + F_{-\lambda,-\nu};$$ $$\begin{aligned} F_{00} & = & \frac{1}{2\sqrt{105}}\left(1+ \frac{2q_0^2}{m^2} \right)c_3 r^2 -\frac{W q_0}{4\sqrt{105}m^2}c_2 r^4 +\frac{W^2}{16\sqrt{105}m^2}c_1 r^6 ,\\ F_{10} & = & \frac{q_0}{\sqrt{56}m}c_3 r^2 -\frac{W}{\sqrt{3584}m}c_2 r^4 ,\\ F_{20} & = & \frac{1}{4\sqrt{7}}c_3 r^2 .\end{aligned}$$ - Spin-1 $\longrightarrow$ spin-0 + spin-0, $\eta_J \eta_s \eta_\sigma = -1$ The effective vertex should be a vector. $$F_{00} = -\frac{c}{2}r .$$ Resonances and backgrounds {#sec:bk} ========================== For a process involving more than three particles, we must separate the vertexes into one-particle irreducible (1PI) parts and one-particle reducible (1PR) parts. Usually 1PI parts are called backgrounds, while 1PR parts are called resonances. An example with 4-leg vertex are shown in Fig. [\[fig:4leg\]]{}. Feynman graphs for the process $a_1\rightarrow \pi^+\pi^+\pi^-$ are illustrated in Fig. \[fig:a13pi\]. We consider only $\rho$ resonance here. The four-momenta of $a_1$ and the three pions are $p$, $p_1$, $p_2$ and $p_3$, with $$p = p_1 + p_2 + p_3.$$ The corresponding spin-parities of $a_1, \rho$ and $\pi$ are $1^-$, $1^-$ and $0^-$[@pdg98]. The background amplitude can be found after an analysis similar to that in Ref. [@Zhua23pi][^7]: $$\begin{aligned} {\cal M}_\lambda^{(b)} & = & e_\mu(p,\lambda) \{ b_1 (p_1+p_2)^\mu + [(p_1-p_2)\cdot p_3] b_2 (p_1-p_2)^\mu \}, \\ b_i & = & b_i\left( (p_1+p_2)\cdot p_3, [(p_1-p_2)\cdot p_3]^2 \right).\end{aligned}$$ We use Breit-Wigner factors as approximation to the full propagators in Fig. \[fig:a13pi\], i.e., write the propagators of $\rho^0$ as $$\frac{g^{\alpha\beta}-(p_2+p_3)^\alpha(p_2+p_3)^\beta/m_\rho^2} {(p_2+p_3)^2- m_\rho^2 +i \Gamma_\rho m_\rho}$$ and $$\frac{g^{\alpha\beta}-(p_1+p_3)^\alpha(p_1+p_3)^\beta/m_\rho^2} {(p_1+p_3)^2- m_\rho^2 +i \Gamma_\rho m_\rho},$$ where $m_\rho$ is the mass of $\rho^0$ and $\Gamma_\rho$ its width. Alternatively, one can choose propagators in other forms to get better approximations. The $\rho^0\pi^+\pi^-$ vertexes must be vectors if parity is conserved, which can be read out from the list of Sec. \[sec:3v\]: $$\begin{aligned} \Gamma_{\rho\pi^+\pi^-}^\beta (p_2,p_3) & = & c(p_2\cdot p_3) \left(p_2-p_3 \right)^\beta,\\ \Gamma_{\rho\pi^+\pi^-}^\beta (p_1,p_3) & = & c(p_1\cdot p_3) \left(p_1-p_3 \right)^\beta.\end{aligned}$$ The $a_1\pi^+\rho^0$ vertexes must be tensors, $$\begin{aligned} \Gamma_{a_1\pi^+\rho}^{\mu\alpha}(p_1, p_2+p_3) & = & c^{'}_1(p_2\cdot p_3) g^{\mu\alpha} + c^{'}_2(p_2\cdot p_3) p_1^\mu p_1^\alpha, \\ \Gamma_{a_1\pi^+\rho}^{\mu\alpha}(p_2, p_1+p_3) & = & c^{'}_1(p_1\cdot p_3) g^{\mu\alpha} + c^{'}_2(p_1\cdot p_3) p_2^\mu p_2^\alpha.\end{aligned}$$ Combining all of these together, we find the resonance part of the amplitude to be $$\begin{array}{rcrl} {\cal M}_\lambda^{(res)} & = & e_\mu(p,\lambda)\{ & \;\;\;D_{23}[\frac{3}{2}c_1(s_{23}) +\frac{1}{2}p_1\cdot (p_2-p_3)c_2(s_{23})] (p_1+p_2)^\mu\\ & & & +D_{23}[-\frac{1}{2}c_1(s_{23}) +\frac{1}{2}p_1\cdot (p_2-p_3)c_2(s_{23})] (p_1-p_2)^\mu\\ & & & +D_{13}[\frac{3}{2}c_1(s_{13}) +\frac{1}{2}p_2\cdot (p_1-p_3)c_2(s_{13})] (p_1+p_2)^\mu\\ & & & +D_{13}[-\frac{1}{2}c_1(s_{13}) +\frac{1}{2}p_2\cdot (p_1-p_3)c_2(s_{13})] (p_2-p_1)^\mu \}. \end{array}$$ Here $p_3^\mu \simeq - (p_1+p_2)^\mu$ have been used, and $$\begin{aligned} s_{23} & = & (p_2 + p_3)^2,\\ s_{13} & = & (p_1 + p_3)^2,\end{aligned}$$ $$\begin{aligned} D_{23} & = & \frac{1}{s_{23}-m_\rho^2+i \Gamma_\rho m_\rho},\\ D_{13} & = & \frac{1}{s_{13}-m_\rho^2+i \Gamma_\rho m_\rho}.\end{aligned}$$ It is easy to see $$\begin{aligned} p_1\cdot (p_2-p_3) & = & \frac{m_{a_1}^2 + 9 m_\pi^2 - 2 (2 s_{13} + s_{23})}{2},\\ p_2\cdot (p_1-p_3) & = & \frac{m_{a_1}^2 + 9 m_\pi^2 - 2 (s_{13} + 2 s_{23})}{2}.\end{aligned}$$ After redefinition of $c_1$, $c_2$ and $b_2$, the covariant helicity amplitude becomes $$\begin{array}{rcl} {\cal M}_\lambda & \equiv & {\cal M}_\lambda^{(b)} + {\cal M}_\lambda^{(res)}\\ & = & e_\mu(p,\lambda)(p_1+p_2)^\mu \times \\ & & \{ \;\;b_1\left(s_{13}+s_{23}, (s_{13}-s_{23})^2\right)\\ & & \;+ 3c_1(s_{23}) D_{23} + 3c_1(s_{13}) D_{13} \\ & & \;+ [m_{a_1}^2 + 9 m_\pi^2 - 2 (2 s_{13} + s_{23})] c_2(s_{23}) D_{23} \\ & & \;+ [m_{a_1}^2 + 9 m_\pi^2 - 2 (s_{13} + 2 s_{23})] c_2(s_{13}) D_{13} \} \\ & & +e_\mu(p,\lambda)(p_1-p_2)^\mu \times \\ & & \{ \;\; [s_{13}-s_{23}] b_2\left(s_{13}+s_{23}, (s_{13}-s_{23})^2\right) \\ & & \;-c_1(s_{23})D_{23} + c_1(s_{13})D_{13}\\ & & \;+[m_{a_1}^2 + 9 m_\pi^2 - 2 (2 s_{13} + s_{23})] c_2(s_{23})D_{23} \\ & & \;-[m_{a_1}^2 + 9 m_\pi^2 - 2 (s_{13} + 2 s_{23})] c_2(s_{13})D_{13} \}. \end{array}$$ The background amplitude will not give a flat distribution in the Dalitz plot of the three pions: $$\begin{array}{rcl} \sum\limits_{\lambda}|{\cal M}_\lambda^{(b)}|^2 & = & -|b_1|^2 (p_1+p_2)^2 - |b_2|^2 (p_1-p_2)^2 [(p_1-p_2)\cdot p_3]^2\\ & & +\frac{1}{m_{a_1}^2} \{ \;\;|b_1|^2[p\cdot (p_1+p_2)]^2 + |b_2|^2[p\cdot (p_1-p_2)]^2 [(p_1-p_2)\cdot p_3]^2\\ & & \;\;\;\;\;\;\;\;\;\; +(b_1^*b_2+b_1b_2^*) [p\cdot (p_1+p_2)][p\cdot (p_1-p_2)][(p_1-p_2)\cdot p_3] \}. \end{array}$$ In fact for any process involving particles with non-zero spins, the background distributions are not flat. If we do not include such background(1PI) terms, those resonances terms we have considered might just simulating the background distributions. Let’s take the process $a_1\rightarrow \pi^+\pi^+\pi^-$ as an example. We can not say that we have seen $\rho\prime$ or some other resonances, if the background term ${\cal M}_\lambda^{(b)}$ is not considered when we fit data. Particularly, this [*is*]{} the case for resonances far off shell or with large widths. [**One must include background terms**]{}. A resonance in a process can be taken as well established under two conditions: (1) We must include the background terms in amplitudes with and without such a resonance. (2) The amplitudes including the resonance significantly improve the best fit to experimental data, comparing to those ignoring it. Summary ======= The main results of this paper are summarized bellow: A list of general 3-leg effective vertexes for bosons is given, with kinematic singularities carefully avoided. Space reflection symmetry demand effect vertexes to be tensors or pseudo-tensors depending on spin-parities of external lines. Mixing of tensor and pseudo-tensor vertexes always means violation of parity conservation. Boson symmetry require that effective vertexes take the special form given in Sec.\[sec:bs\]. The requirement of parity conservation and boson symmetry leads to selection rules. These results are needed when we construct phenomenological models or write amplitudes to fit data of high energy experiments. Helicity amplitudes in laboratory frame are related to those in center of mass frame by Wigner rotations. For two-body decays, it is possible to write the explicit expressions of covariant helicity amplitudes in a concise form. S-matrixes for processes involving more than three particles can be divide into 1PI parts and one-particle reducible parts, or in another words, backgrounds and resonances. We emphasize that such background terms are important when one try to extract meaningful information on resonances. This is especially the case if the width of the resonances are large, or the resonances are far off shell. Constraints of gauge invariance on effective vertexes will be the content of another paper. Acknowledgement {#acknowledgement .unnumbered} =============== Jie-Jie Zhu wishes to thank Prof. Yu-Can Zhu and Dr. Liao-Yuan Dong for discussions. This work is partially supported by National Founds of China through C. N. Yang, Funds of IHEP of China and the Grant LWTZ-1298 of Chinese Academy of Science. M. Jacob and G. C. Wick, Ann. Phys. [**7**]{}, 404 (1959). Chou Kuang-Chao and M. I. Shirokov, J. Exp. Theor. Phys. [**34**]{}, 1230 (1958). S. M. Berman and M. Jacob, Phys. Rev. [**139**]{}, B1023 (1965). G. C. Wick, Ann. Phys. (N.Y.) [**18**]{}, 65 (1962). C. Zemach, Phys. Rev. [**133**]{}, B1201 (1964). C. Zemach, Phys. Rev. [**140**]{}, B97 (1965). S. U. Chung, Phys. Rev. D [**48**]{}, 1225 (1993). V. Filippini, A. Fontana and A. Rotondi, Phys. Rev. D [**51**]{}, 2247 (1995). S. U. Chung, [*Spin Formalisms*]{}, CERN 71-8 (1971). H. Lehmann, K. Symanzik and W. Zimmermann, Nuovo Cimento, [**10**]{}, 205 (1955). N. N. Bogoliubov, A. A. Logunov and I. T. Todorov, [*Introduction to Axiomatic Field Theory*]{}, (Benjamin/Cummings Reading MA, 1975). W. Rarita and J. Schwinger, Phys. Rev. [**60**]{}, 61 (1941). H. P. Stapp, Phys. Rev. [**125**]{}, 2139 (1962). Y. Hara, Phys. Rev. [**136**]{}, B507 (1964). H. P. Stapp, Phys. Rev. [**160**]{}, 1251 (1967). L. L. C. Wang, Phys. Rev. [**142**]{}, 1187 (1966). G. Cohen-Tannoudji, A. Morel and H. Navelet, Ann. Phys., [**46**]{}, 239 (1968). W. A. Bardeen and Wu-Ki Tung, Phys. Rev. [**173**]{}, 1423 (1968). R. W. Brown and I. J. Muzinich, Phys. Rev. D [**4**]{}, 1496 (1971). M. D. Scadron, Phys. Rev. [**165**]{}, 1640 (1968). E. P. Wigner, Ann. Math. [**40**]{}, 149 (1939). Jie-Jie Zhu and Tu-Nan Ruan, Phys. Rev. D [**57**]{}, 5468 (1998). Jie-Jie Zhu, [*thesis*]{}, (University of Science and Technology of China, 1997), unpublished. Steven Weinberg, [*The Quantum Theory of Fields, Vol. I, Foundations*]{}, (Cambridge University Press, Cambridge, 1995). S. U. Chung, BNL Report No. BNL-QGS94-22 (1995). C. Fronsdal, Nuovo Cimento, [**Suppl. 9**]{}, 416 (1958). Particle Data Group, C. Caso [*et al.*]{}, Eur. Phys. J. C [**3**]{}, 1 (1998). [^1]: Permanent address. [^2]: Electronic address: [email protected] [^3]: Three of these equivalence relations, Eq. [(\[eq:scadron1\])]{}, Eq. [(\[eq:scadron2\])]{} and Eq. [(\[eq:scadron3\])]{}, have been listed in Ref. [@ScadronVF68]. [^4]: See Ref.[@WeinbergQFT1] for arguments on polology. [^5]: Strictly speaking, we should use Green functions to derive properties of off shell effective vertexes. [^6]: $\rho^0\rightarrow 2 \pi^0$ also violates (approximate) isospin symmetry. Since boson symmetry is an exact symmetry, this process is absolutely forbidden. [^7]: The $c_3$ term in Eq.(10) of Ref. [@Zhua23pi] can be dropped without introducing K. S..

ArXiv

--- abstract: 'Brain computer interfaces (BCIs) offer individuals suffering from major disabilities an alternative method to interact with their environment. Sensorimotor rhythm (SMRs) based BCIs can successfully perform control tasks; however, the traditional SMR paradigms intuitively disconnect the control and real task, making them non-ideal for complex control scenarios. In this study we design a new, intuitively connected motor imagery (MI) paradigm using hierarchical common spatial patterns (HCSP) and context information to effectively predict intended hand grasps from electroencephalogram (EEG) data. Experiments with 5 participants yielded an aggregate classification accuracy–intended grasp prediction probability–of 64.5% for 8 different hand gestures, more than 5 times the chance level.' author: - | Seyed Sadegh Mohseni Salehi$^{2}$, Mohammad Moghadamfalahi, Fernando Quivira\ Alexander Piers, Hooman Nezamfar, and Deniz Erdogmus$^{1}$[^1][^2][^3] bibliography: - 'refs.bib' title: '**Decoding Complex Imagery Hand Gestures\*** ' --- INTRODUCTION ============ Brain Computer Interfaces have shown promises in providing individuals with alternative interaction methods via brain activities. Major electroencephalogram (EEG) based BCI systems can be categorized into three different categories based on the type of brain activity they utilize, event-related potential (ERP) based, steady state visually evoked potential (SSVEP) based, and motor imagery (MI) based. ERP and SSVEP are more popular in BCI systems where the short response time and the availability of several options are of concern. Controlling a wheelchair [@nezamfar2013brain] and typing with characters [@nezamfar2016flashtype] are two examples that rely on the above features, respectively. In control and navigation applications SSVEP-based BCI systems are widely used due to their fast response. SSVEP systems are fast, but they require the visual stimulation and vision capabilities [@nezamfar2015stimuli] not suitable for individuals with severely impaired vision. MI, the process of imagining a physical movement without execution, can be categorized as imagining moving different limbs (simple MI) and imagining different movements of the same limb (compound MI)[@yi2013eeg]. MI tasks desynchronize the stable, resting alpha/mu ($7-12$ Hz) and beta ($12-30$ Hz) rhythms, that can be measured through EEG potentials on the scalp[@yi2013eeg; @pfurtscheller1999event]. The spatial distribution of brain activity measured by EEG signals in response to these imaginations depend on the complexity of the MI task; simple MI (imagination of left hand vs. right leg) produces activity in different spatial locations while compound MI (open palm vs. closed fist) produces activities in similar spatial cortical areas [@vuvckovic2012two]. The high spatial correlation of compound MI coupled with the volume conduction effect, the spatial distortion of EEG signals on the scalp caused by travelling through tissue and fluid, make different MI tasks difficult to distinguish. One particular application of BCIs is the control of prosthetics or exoskeletons for individuals with motor control loss. From the BCI perspective, a brain controlled hand requires rapidly (real-time) and accurately classified MI that maps directly to the motion of the target limb. Simple MI is not sufficiently intuitive, and compound MIs are difficult to differentiate. Hence classifying EEG potentials suitable to control a robotic limb is a challenge. Several studies have classified gestures with moderate success, but these gestures are not sufficiently intuitive to be used in a BCI prosthetic. These studies investigated classifying simple MI by simply distinguishing imagination of right and left hand movements with $80\%-93\%$ accuracy[@vuvckovic2012two; @pfurtscheller1997eeg; @hamedi2014neural]. Others have completed binary classifications on different same-limb movements, such as distinguishing between wrist flexion, extension, supination, pronation, and finger movement[@vuvckovic2012two; @vuckovic2008delta; @mohamed2011single] resulting in accuracies between $58-82\%$. However, the higher accuracies are only achieved when wrist extension has been one of the binary options; without wrist extension classification accuracies tend to run closer to the lower side of the reported range. Binary classification does not provide enough degrees of freedom to control a prosthetic hand, and the imaginary movements used for classification are not necessarily intuitive hand motions. A recent study [@edelman2016eeg] classified four hand states (wrist flexion, extension, supination, and pronation) using source imaging analysis [@edelman2014discriminating] to improve signals’ separability. Even with this combination, the imagined wrist motions do not map directly or intuitively enough to natural hand movements. In this paper, we propose classification of four grasps on each hand totaling to 8 options. The grasps mimic four of the required grasps to complete everyday tasks. The goal is to improve classification accuracy and time to control a prosthetic hand, utilizing the natural choice of grasps. Method ====== Experiment design ----------------- Classifier parameters are estimated using data collected during a supervised calibration session. The classifier is meant to discriminate EEG signals corresponding to imagination of four gestures shown in Fig. 1.a. Prior to data collection, participants get familiarized with gestures through executing each gesture two times, for 5 seconds. During the calibration session, participants are presented with the *target* gesture for 2 seconds during which they are asked to focus on that particular gesture. Then, upon the presentation of the word *start* on the display participants imagine the target gesture for 5 seconds (Fig. 1.b). Here, each MI task for a gesture is called a *trial*. Inter trial interval in the calibration task is arbitrary and is decided by the user based on their required time to rest. Each participant performs 20 trials of each gesture, summing to a total of 160 MI trials. {width="0.8\columnwidth"} \[fig:4gestures\] EEG feature extraction ---------------------- Prior to feature extraction, EEG signals are filtered with a $3-30$ Hz FIR band-pass filter to retain the alpha/mu and beta frequency bands, as they show the most activity in response to imagination and execution of hand gestures [@pfurtscheller1997eeg]. Assuming the band-passed EEG signal follows a Gaussian distribution within the time window, and source activity constellations between different class pairs are independent, we can design a spatial filter (spatial transform) that maximizes the signal’s variance for one class and minimizes for the other. We used the common spatial pattern (CSP) algorithm [@ramoser2000optimal] to calculate the spatial transforms. CSP algorithm is used to find the transformation matrix between class pairs. CSP finds the best coordinates between binary classes. In the multiclass case, we propose a Hierarchical Common Spatial Pattern (HCSP) to utilize a series of binary transformations found by CSP. In HCSP, classes with common properties are placed in one category. For example, first level contains two categories, motions on the right hand and motions on the left hand. Similarly, extension or flexion of fingers and abduction or adduction of thumbs are considered different categories in the second and the third level respectively. Fig \[fig:Hierarchy\] outlines the HCSP for classification of different hand gestures. The goal of each classifier is to classify the categories not classes. By cascading the category classifiers, all 8 classes can be detected. \[fig:Hierarchy\] The normalized spatial covariance matrix is estimated as $$C_j = \frac{1}{N_j}\sum_{i=1}^{N_j}\frac{E_i^{(j)}{E_i^{(j)}}'}{trace(E_i^{(j)}{E_i^{(j)}}')}$$ where $E_i^{(j)}$ is an $m\times t$ matrix representing filtered windowed EEG evidence corresponding to the $i^{th}$ trial of the $j^{th}$ class. Additionally, $m$ is the number of channels, $t$ is the number of samples in a time window, $'$ is the transpose operator, $trace(.)$ is the sum of the diagonal elements of the matrix, and $N_j$ is the number of trials for the $j^{th}$ class. In order to find the spatial transformation matrix at each level,$V_L$, the optimization problem in equation \[eqe:CSP\] should be solved. $$V_L^* = \underset{V_L}{\operatorname{argmax}}\frac{V_L^T \sigma_{L_{l=-1}} V_L}{{V_L}' (\sigma_{L_{l=-1}}+\sigma_{L_{l=+1}}) V_L} \label{eqe:CSP}$$ where $\sigma_{L_{l=-1}}$ and $\sigma_{L_{l=+1}}$ are the average covariance matrices - where $\pm1$ represent one of the categories - in level $L$ and are calculated by $$\sigma_{L_{l}} = \frac{\sum_{j\in category({L_{l}})}N_j C_j}{\sum_{j\in category({L_{l}})}N_j} \label{eqe:sigma}$$ $\sum_{j\in category({L_{l}})}N_j$ is the number of trials for each category at each level. Solving the optimization problem represented in equation \[eqe:CSP\] is identical to solving generalized eigenvalue problem $$V_{L}'\sigma_{L_{l=-1}}V_L = D \wedge V_{L}'(\sigma_{L_{l=-1}}+\sigma_{L_{l=+1}})V_L = I \label{eqe:GEP}$$ where $D$ is a diagonal matrix. According to each spatial filter $V_L$, computed for level $L$, the projected data $P_{L_l}$ corresponding to the $i^{th}$ trial of the $j^{th}$ class, $E_i^{(j)}$, can be calculated as $$P_{L_l} = {V_L}' E^{(j\in category(L_l))} \label{eqe:projected data}$$ In equation \[eqe:GEP\] the $k$ smallest and largest eigenvalues in $D$ correspond to $k$ leftmost/rightmost columns in spatial filter $V$ respectively. These values yield to the smallest variance in category $L_{l=-1}$ and simultaneously largest variance in category $L_{l=+1}$ and vice versa. To extract the features, the first and the last $k$ rows of $P_{L_l}$ are considered. Using $P_{L_l}^{(K)}$ ($K=1 \hdots 2k)$ which maximizes the difference of variances between two categories, $f_{L_l}^{(K)}$, features for each trial of the corresponding category at each level are calculated as $$f_{L_l}^{(K)} = log \left (\frac{var(P_{L_l}^{(K)})}{\sum_{q=1}^{2k}var(P_{L_l}^{(q)})}\right ) \label{eqe:feature}$$ where, $i$ is the trial index in each the category. EEG Likelihood Probability Extraction {#sec:method_classification} ------------------------------------- For proper estimation of the classifier performance a leave-one-out cross validation approach is used. At each level, the corresponding feature vector, $f_{L_{l_i}}^{(K)}$, is calculated for each trial. The classifier is trained in two steps. First, Fisher LDA algorithm is used to extract fisher scores by $$F_{L_{l_i}}^{(K)} = \left (w_L^{(K)}\right )' f_{L_{l_i}}^{(K)}$$ where $$w_L^{(K)} \propto \left (\Sigma_{L_{l={-1}}}^{(K)} + \Sigma_{L_{l={+1}}}^{(K)}\right )^{-1} \left (\mu_{L_{l={+1}}}^{(K)}-\mu_{L_{l={-1}}}^{(K)}\right )$$ and $\Sigma_{L_{l={-1}}}^{(K)}, \mu_{L_{l={-1}}}^{(K)}$ and $\Sigma_{L_{l={+1}}}^{(K)}, \mu_{L_{l={+1}}}^{(K)}$ are the category’s covariance and mean. Then, the likelihood probability densities are calculated for the fisher scores in each category $P(F_{L_{l}}^{(K)}|L=l)$. Multi-variate Gaussian model is considered at each level for density estimation due to the nature of EEG signals. Classification accuracy is defined as the probability of the correct answer. Different integer window lengths from 1 to 5 seconds are considered for feature extraction. Graphical model {#sec:graphical_model} --------------- The graphical model in Fig \[fig:GraphicalModel\] is proposed to probabilistically merge the category classifiers and employ the inter-level information to calculate the posterior probability of each gesture. \[fig:GraphicalModel\] This model represents the generative model of the collected EEG data in epoch $e$ and the time required for the model to make a decision ($1-5$ s). The goal of this graphical model is to estimate the next state $s_e$ with incorporation of inter-level prior information. The inter-level information is the prior probability of a gesture at a level given the state at the higher level. This information is gathered based on the combination of gestures used over time for different tasks. For instance, the probability of the thumb being closed would be increased if closed fingers is detected at the higher level. Specifically in this graphical model:$s_e$ is the decision between all classes at epoch $e$, $L_e^j$ represents the decision from the level j classifier, and $\epsilon_{e_i}^j$ is the extracted CSP feature from EEG evidence at time $i$ for decision $L_e^j$. Using the graphical model above, the posterior probability of each state given the observed random variables ($\{\epsilon_{e_i}^1\}_{i=1}^{t},\{\epsilon_{e_i}^2\}_{i=1}^{t},\{\epsilon_{e_i}^3\}_{i=1}^{t}$) can be calculated to construct the PMF over all gestures, required to make a decision. The posterior probabilities are calculated as follows: $$\begin{split} P(s_e&\ |\{\epsilon_{e_i}^1\}_{i=1}^{t},\{\epsilon_{e_i}^2\}_{i=1}^{t},\{\epsilon_{e_i}^3\}_{i=1}^{t}) \\ &\ = P(L_e^1,L_e^2,L_e^3|\{\epsilon_{e_i}^1\}_{i=1}^{t},\{\epsilon_{e_i}^2\}_{i=1}^{t},\{\epsilon_{e_i}^3\}_{i=1}^{t}) \\ &\ \propto P(\{\epsilon_{e_i}^1\}_{i=1}^{t},\{\epsilon_{e_i}^2\}_{i=1}^{t},\{\epsilon_{e_i}^3\}_{i=1}^{t}|L_e^1,L_e^2,L_e^3)\\ &\ \times P(L_e^1,L_e^2,L_e^3) \\ &\ \propto P(\{\epsilon_{e_i}^1\}_{i=1}^{t}|L_e^1)P(\{\epsilon_{e_i}^2\}_{i=1}^{t}|L_e^2)P(\{\epsilon_{e_i}^3\}_{i=1}^{t}|L_e^3) \\ &\ P(L_e^3|L_e^2)P(L_e^2|L_e^1)P(L_e^1) \end{split}$$ In this set of equations: $P(\{\epsilon_{e_i}^M\}_{i=1}^{t}|L_e^M)$, represents EEG observation likelihood for a particular gesture at level $M$, estimated using the calibration data; $P(L_e^M|L_e^{M-1})$, is the probability of each command at level $M$ given the inter-level information; and $P(L_e^1)$, represents the prior probability of each hand being target. Decision Criteria {#sysDec} ----------------- One decision is made in every epoch. Decisions are made using a Maximum-a-Posteriori (MAP) estimate. $$\hat{s}_e=\arg\max_{s_e}{P(s_e|\{\epsilon_{e_i}^1\}_{i=1}^{t},\{\epsilon_{e_i}^2\}_{i=1}^{t},\{\epsilon_{e_i}^3\}_{i=1}^{t})} \label{eq:decision}$$ At each time interval $t$, if the posterior probability of $\hat{s}_e$ is greater than the system *Confidence Threshold*, it will be chosen as the final decision of the epoch. If after $N$ time interval (here $N=5$) the posterior probability is not greater than the system *Confidence Threshold*, the maximum posterior probability of $\hat{s}_e$ across all time intervals will be chosen as the final decision of the epoch. Analysis and results Analysis and Results ==================== Five healthy participants (3 males and 2 females) in the range of 20-30 years old consented and participated in a data collection session following an approved protocol by Northeastern University’s IRB office. Participants were not under the influence of any chemicals, such as caffeine. Before starting the experiment, different gestures were described to the participants; participants executed these gestures to get more familiar with the motions to be imagined during the imagination task. Participants were asked to imagine the entire progress of the gestures, not only the final position. Between each trial (2 seconds preparation, 5 second gesture imagination), participants rested until they felt comfortable to continue. On average, the experiment consisting of 160 trials, 4 different gestures on each hand took less than 45 minuets. All possible number of features ($K$) and window lengths ($t\geqslant3s$) combinations were evaluated offline to find the parameter values leading to the highest classification accuracy. Figure \[fig:accuracy8\] shows the total system accuracy as a function of the window length and number of features. As expected, more EEG evidence (larger $t$) results in higher accuracy. The average maximum accuracy of 64.5% (71% the highest) is achieved using $K=6$ and the window length of $t=5$ seconds using uniform inter-level prior probability. ![Average system accuracy using different number of features (x-axis) and different time-window lengths of EEG evidence.[]{data-label="fig:accuracy8"}](fig/All_acc.jpg){width="0.9\columnwidth"} The average confusion matrix is shown in figure \[table:8 classes\]. The confusion matrix shows the system accurately classifies between left and right hand (level 1), hence the two approximately zero blocks on the top-right and bottom-left of the matrix (Green squares). The highest confusion is between classes that differ in thumb position (level 3), as shown by the highlighted red blocks. ![Average confusion matrix for 5 participants with using $K=6$ features and the window length of $t=5$ seconds[]{data-label="table:8 classes"}](fig/Confusion_MI.jpg){width="0.9\columnwidth"} The confusion matrix represented in figure \[table:8 classes\] shows gestures with differences in details are not as separable as the more general levels like left and right hand. DISCUSSION ========== With the goal of utilizing intuitive and natural gestures towards building motor imagery based classifiers. The natural gestures are more complex they reflect on the same region and require more sophisticated classifiers as opposed to the tasks performed or imagined on the similar limbs on the different sides. We proposed a hierarchical model combining binary classifiers on different levels taking advantage of the similarities in different gestures. Binary CSP is used to extract features at each level. The results from each classifier is probabilistically merged, and inter-level probabilities are employed to achieve a multiclass gesture classifier. In a study with 5 healthy participants our method resulted in an average accuracy of $64.5\%$ among 8 complex hand gestures, more than 5 times the chance level. [^1]: \*Our work is supported by NSF (IIS-1149570, CNS-1544895), NIDLRR (90RE5017-02-01), and NIH (R01DC009834). Relevant code and data will be disseminated via Northeastern University Digital Repository Service, in collection NEU/COE/ECE/CSL (permalink: http://hdl.handle.net/2047/D20199232). [^2]: $^{1}$ [Electrical and Computer Engineering Department, Northeastern University, Boston]{} [^3]: $^{2}$ [[email protected]]{}

ArXiv

[**Universal Power-law Decay in Hamiltonian Systems?**]{} The understanding of the asymptotic decay of correlations and of the distribution of Poincaré recurrence times $P(t)$ has been a major challenge in the field of Hamiltonian chaos for more than two decades. In a recent Letter, Chirikov and Shepelyansky [@CS99] claimed the universal decay $P(t) \sim t^{-3}$ for Hamiltonian systems. Their reasoning is based on renormalization arguments and numerical findings for the sticking of chaotic trajectories near a critical golden torus in the standard map. We performed extensive numerics and find clear deviations from the predicted asymptotic exponent of the decay of $P(t)$. We thereby demonstrate that even in the supposedly simple case, when a critical golden torus is present, the fundamental question of asymptotic statistics in Hamiltonian systems remains unsolved. As in Ref. [@CS99] we study the standard map $$\label{standardmap} q_{n+1}=q_n+p_n \,{\rm mod}\,2\pi\qquad p_{n+1}=p_n+K\sin q_{n+1} \,\, ,$$ at $K=K_c=0.97163540631$, where the golden torus is critical (Fig. 1, inset). We determine the Poincaré recurrence time distributions $P(t)$ for trajectories starting below and above the critical golden torus by using the same numerical approach as in Ref. [@CS99]. By considerably increasing the statistics we are able to extend the distribution by almost two orders of magnitude in recurrence times. We verify that our statistical data are not affected by the unavoidable finite numerical precision by comparing data for double ($\approx$16 significant digits) and quadruple ($\approx$32 digits) precision. The data for approaching the critical golden torus from above and below are presented in Fig. \[fig:tcube\]. For times $t < 10^8$ our data agree with the results presented in Fig. 2 of Ref. [@CS99]. For larger times, however, we find strong deviations from the predicted universal power law $P(t) \sim t^{-3}$ (dashed lines in Fig. \[fig:tcube\]). The deviations might be explained in two ways: The onset of the claimed asymptotic decay might occur for larger times, which is in contradiction to the prefactors determined in Ref. [@CS99]. On the other hand, the long-time trapping of chaotic trajectories might be dominated by islands of stability (non-principal resonances) that are neglected by the renormalization arguments. In fact, the latter possibility is supported by a detailed investigation [@WHK]. If even in the supposedly simple case of a critical golden torus the decay $P(t) \sim t^{-3}$ is not observed, the claim for a universal existence of this decay cannot be maintained. It thus remains a fundamental challenge in the field of Hamiltonian chaos whether the asymptotic behavior of $P(t)$ follows a universal power law and what the value of its exponent would be. =8.4cm We thank Dima Shepelyansky for candid discussions. M. Weiss$^{1,2}$, L. Hufnagel$^1$, and R. Ketzmerick$^1$ [$^1$ Max-Planck-Institut für Strömungsforschung and Institut für Nichtlineare Dynamik der Universität Göttingen, Germany]{} [$^2$ EMBL, Meyerhofstr. 1, 69117 Heidelberg, Germany]{} PACS numbers: 05.45.Mt B. V. Chirikov and D. L. Shepelyansky, Phys. Rev. Lett. [**82**]{}, 528 (1999). M. Weiss, L. Hufnagel, and R. Ketzmerick, nlin.CD/0106021.

ArXiv

--- abstract: 'Massive theories of abelian $p$-forms are quantized in a generalized path-representation that leads to a description of the phase space in terms of a pair of dual non-local operators analogous to the Wilson Loop and the ’t Hooft disorder operators. Special atention is devoted to the study of the duality between the Topologically Massive and the Self-Dual models in $2+1$ dimensions. It is shown that these models share a geometric representation in which just one non local operator suffices to describe the observables.' --- hep-th/0206082\ UCVFC-DF-16-2001 1truein \ [**[Pío J. Arias[${}^{a, b, }$]{}[^1], Lorenzo Leal[${}^{a,c, }$]{}[^2] and J.C. Pérez-Mosquera[${}^{a, }$]{}[^3]]{}**]{}\ ${}^a$[*Grupo de Campos y Partículas, Departamento de Física, Facultad de Ciencias, Universidad Central de Venezuela, AP 47270, Caracas 1041-A, Venezuela*]{}\ ${}^b$[*Centro de Astrofísica Teórica, Facultad de Ciencias, Universidad de Los Andes, La Hechicera, Mérida 5101, Venezuela*]{}\ ${}^c$ [*Departamento de Física Teórica, Universidad Autónoma de Madrid, Cantoblanco 28049, Madrid, Spain*]{} Introduction ============ As it is known, the electric-magnetic duality of free Maxwell theory may be seen as a particular case of a duality tansformation relating different abelian gauge theories of arbitrary rank in the appropriate space time dimensions[@Townsend:1981nu; @DFNS]. In $D$ dimensions, the ranks $p_1$, $p_2$ of the generalized potentials describing the dual abelian theories must obey the equality $p_1 + p_2 = D-2$. For example, in four dimensions, Maxwell theory is self-dual, while the second rank gauge theory is dual to the massles scalar field. In reference[@LL] it was shown that the “electric-magnetic” duality of abelian gauge theories allows to describe their physical phase space in terms of a pair of non-local observables that are dual in the Kramers-Wannier sense[@Kramers:1941kn]. The algebra that they obey results to be invariant under spatial diffeomorphisms. This topological algebra, the Dual Algebra (DA), admits a realization in terms of operators acting on functionals that depend on extended objects, inasmuch as the dual operators themselves. For instance, in the case of Maxwell theory in four space time dimensions, the dual operators are the Wilson Loop and the ’t Hooft disorder operator[@tHooft]. Both operators depend on closed spatial loops, and may be realized on a loop-dependent Hilbert space (see section \[sec2\]). The DA of the three and four dimensional Maxwell theory had been previously analized[@AlgNL-Maxwell], due to their close relation with the Yang-Mills field. Furthermore, non local operators that obey commutation relations of the DA type have been used to quantize topological excitations in interacting field theories[@Marino:1992rq]. In this paper we discuss how the ideas of reference[@LL] can be extended to the conventional (i.e., non topological) Abelian massive theories in arbitrary dimension, and to the Self- Dual [@TPN] and Topologically Massive theories[@DJT] in 2+1 dimensions, which are known to be dual to each other[@DJ]. In all the cases, the program that we develop is as follows: one starts from a first order master lagrangian that encodes the dual theories simultaneously. We take this master lagrangian to be of the Stüeckelberg form [@St], in order to maintain gauge invariance even in the massive case. The master theory is then quantized within the Dirac scheme[@Dirac], and the phase space is taken into account by choosing non local operators that encode all the gauge invariant content of the original canonical operators. The algebra obeyed by these dual operators is then studied and realized onto an appropriate set of functionals. We shall see that the DA of massive Abelian theories is also characterized by a topological quantity, namely, the intersection number between the extended objects that support the non-local dual operators. This contrasts with the masless case, where the DA is governed by the linking number of the closed extended objects that enter in the construction of the dual operators[@LL; @AlgNL-Maxwell]. This and other differences between both cases are studied. The case of the Self-Dual and Topologically Massive theories presents several interesting peculiarities, regarding the DA study. Perhaps the more relevant one is that instead of a pair of Wilson Loop operators, as in both the massless and the conventional massive theories, it suffices with only one non-local operator to describe the gauge invariant content of the theory. Consecuently, this operator has to play both the “coordinate” and “momentum” roles. As we shall see, this feature has an interesting geometrical counterpart when the non-local operator is realized in a path-dependent Hilbert space. In other direction the Proca model in 2+1 dimensions is equivalent to two non-interacting Self-Dual models with opposite spins. The paper is organized as follows. In section \[sec2\] we review the massless case, following reference[@LL], focusing mainly in the study of Maxwell theory in four dimensions. In section \[sec3\] the DA of the Proca model in three dimensions is considered. Section \[sec4\] is dedicated to the study of the Self-Dual and Topologically Massive theories. Some concluding remarks are given in the last section. In the Appendix we summarize the generalization of the study presented in section \[sec3\], to the case of forms of arbitrary rank in arbitrary dimension. Maxwell theory {#sec2} ============== Let us summarize the results of reference[@LL] regarding Maxwell theory. The starting point is the first order Lagrangian density $${\cal L}=\frac 12\epsilon^{\mu\nu\lambda\rho}\partial_\mu A_\nu B_{\lambda\rho}-\frac 14\left(B_{\lambda\rho}+\partial_\lambda C_\rho-\partial_\rho C_\lambda\right)\left(B^{\lambda\rho}+\partial^\lambda C^\rho-\partial^\rho C^\lambda\right),$$ which is invariant under the simultaneous gauge transformations $$\begin{aligned} \delta A_\mu&=&\partial_\mu\Lambda,\\ \delta B_{\lambda\rho}&=&\partial_\rho\xi_\lambda-\partial_\lambda\xi_\rho, \\ \delta C_\rho&=&\xi_\rho + \partial_\rho\xi.\label{deltaC}\end{aligned}$$ Equation [(\[deltaC\])]{} shows that the field $C_\rho$ is pure gauge. Its presence just serves to enforce gauge invariance. When this field is gauged away in (1), the equations of motion become $$\begin{aligned} \epsilon^{\mu\nu\lambda\rho}\partial_\mu A_\nu&=&B^{\lambda\rho},\label{EdMA}\\ \epsilon^{\mu\nu\lambda\rho}\partial_\mu B_{\lambda\rho}&=&0.\label{EdMB}\end{aligned}$$ Substituting $B_{\lambda\rho}$ from [(\[EdMA\])]{} into the master lagrangian (with $C_\rho=0$), one finds the standard Maxwell lagrangian. If, instead, one solves equation [(\[EdMB\])]{} locally $$B_{\lambda\rho}=\partial_\lambda\tilde A_\rho-\partial_\rho\tilde A_\lambda,$$ and substitutes the above expression into the equation (1) (again with $C_\rho=0$) we obtain (after an integration by parts) the “dual” lagrangian density $$\tilde L=-\frac 14 F_{\mu\nu}(\tilde A) F^{\mu\nu}(\tilde A),$$ in correspondence with the fact that in $D=4$ Maxwell theory is self-dual. The canonical analysis may be summarized as follows. There are three secondary first class constraints $$\begin{aligned} \psi&=&-\frac 12 \epsilon^{ijk}\partial_i B_{jk}\approx0,\label{ligadura1}\\ \theta^i&=&-(\pi_C^i-\epsilon^{ijk}\partial_j A_k)\approx0\label{ligadura2}\\ \theta&=&-\partial_i\pi_C^i\approx0\label{ligadura3}\end{aligned}$$ where $\pi_c^i$ is the momentum canonically conjugate to $C_i$. We are taking $g_{\mu\nu} = \hbox{diag}(1,-1,-1,-1)$. These constraints are reducible ($\partial_i\theta^i-\theta=0$) and appear associate, respectively, to $A_0$, $B_{0i}$ and $C_0$ as their Lagrange multipliers. The fields $A_i$ and $\frac 12\epsilon^{ijk}B_{jk}$ are mutually conjugate $$\left[A_k(\vec x),\frac 12\epsilon^{lij}B_{ij}(\vec y)\right]=i\delta_k^l{\delta{^{(3)}}(\vec x-\vec y)},$$ as can be seen from the first order $BF$ term in the master lagrangian (see [@Faddeev-Jackiw]). $\psi$, $\theta^i$ and $\theta$ generate the gauge transformations for $A_i$, $B_{ij}$, $C_i$ and $\pi_C^i$. The gauge transformations for the remaining fields are obtained imposing the gauge invariance on the extended action, taking into account the reducibility of the first class constraints. On the physical sector, the Hamiltonian reduces to $$H=\int d^3\vec x\frac 12\left({\cal B}^i{\cal B}^i+{\cal E}^i{\cal E}^i\right),$$ with the magnetic and electric fields given, respectivelly, by $$\begin{aligned} {\cal B}^k&\equiv&\epsilon^{ijk}\partial_j A_i,\\ {\cal E}^i&\equiv&\frac 12\epsilon^{ijk}\left(B_{jk}+F_{jk}(C)\right).\end{aligned}$$ The gauge invariant combinations of the operators appearing in the above expressions indicate which are the non-local dual operators we are interested in. They are the Wilson loop $$W(\gamma)=\exp\left(i\oint_\gamma dy^iA_i(\vec y)\right),\label{WL-Maxwell}$$ with $\gamma$ a closed spatial path, and the operator $$\Omega(\Sigma,\Gamma)=\exp{\left(i\oint_\Gamma dy^iC_i(\vec y)\right)} \exp{\left(i\int_\Sigma d\Sigma_k\epsilon^{kij}B_{ij}\right)}, \label{WD-Maxwell}.$$ wich depends on the spatial open surface $\Sigma$ whose boundary is $\Gamma$. In virtue of the constraint [(\[ligadura1\])]{}, one has $$\Omega\left(\Sigma_{\hbox{closed}}\right){\left|\psi_{\hbox{physical}}\right\rangle}= {\left|\psi_{\hbox{physical}}\right\rangle},$$ i.e. $\Omega$ does not depend on the surface $\Sigma$, but only on its boundary $\Gamma$. The algebra obeyed by the dual operators (the DA) is given by $$W(\gamma)\Omega(\Gamma)=e^{i{\cal L}(\gamma,\Gamma)} \Omega(\Gamma)W(\gamma).\label{AlgDual}$$ where the quantity $${\cal L}(\gamma,\Gamma)=\frac 1{4\pi}\oint_\gamma dx^i\oint_\Gamma dy^j\epsilon_{ijk}\frac{(\vec x-\vec y)^k}{|\vec x-\vec y|^3},$$ measures the Gauss linking number between $\gamma$ and $\Gamma$, which are closed curves in $R^3$, and is a topological object, since it does not depend on the metric properties of the space. The operator $\Omega (\Gamma)$ results to be the “dual” Wilson loop, i.e. the contour integral of the dual potential $\tilde A$ along $\Gamma$[@AlgNL-Maxwell]. It must be noticed, however, that these results are obtained from a formulation that does not include this potential as a lagrangian variable, which would be redundant. The DA [(\[AlgDual\])]{} is fulfilled if the operators are defined to act onto loop dependent functionals $\Psi(\gamma)$ as $$\begin{aligned} W(\gamma)\Psi(\gamma_1)&=&\Psi({\gamma}\circ{\gamma_1}), \label{RealW} \\ \Omega(\Gamma)\Psi(\gamma_1)&=&e^{-i{\cal L}(\Gamma,\gamma_1)}\Psi(\gamma_1). \label{RealO}\end{aligned}$$ Here ${\gamma}\circ{\gamma}'$ denotes the Abelian group of loops product[@GT-em; @NC:Max]. It is worth recalling that an Abelian loop is an equivalence class of closed curves, defined as follows. The curves $\gamma_1$ and $\gamma_2$ are equivalent if their form factors $T^i(\vec x,\gamma_1)$ and $T^i(\vec x,\gamma_2)$, with $$T^i(\vec x,\gamma)\equiv\int_\gamma dy^i{\delta{^{(3)}}(\vec x-\vec y)},$$ are equal. With this definition it is easy to see that the usual composition of curves is lifted to a group product. The electric and magnetic fields may be obtained from $W(\gamma)$ and $\Omega(\Gamma)$ through the expressions $$\begin{aligned} {\cal B}^i(\vec x)&=&\left.-i\epsilon^{ijk}\Delta_{jk}(\vec x)W(\gamma) \right|_{\gamma=0},\label{DefB}\\ {\cal E}^i(\vec x)&=&\left.-i\epsilon^{ijk}\Delta_{jk}(\vec x) \Omega(\Gamma)\right|_{\Gamma=0},\label{DefE}\end{aligned}$$ where we have made use of the loop derivative $\Delta_{ij}(x)$ of Gambini-Trias[@GT-em], $$\delta\sigma^{ij}\Delta_{ij}(\vec x)f(\gamma)\equiv f(\delta\gamma\circ\gamma)-f(\gamma)\label{LD},$$ that measures the change experimented by a loop dependent object $f(\gamma)$ when its argument $\gamma$ is modified by attaching a small plaquette $\delta\gamma$ of area $\delta\sigma^{ij}$ at the point $\vec x$. In view of equations [(\[DefB\])]{},[(\[DefE\])]{}, the Hamiltonian and the other observables of the theory may be expressed in terms of the basic operators $W$ and $\Omega$. Equations [(\[AlgDual\])]{}, [(\[RealW\])]{} and [(\[RealO\])]{} are the basic results of the geometric formulation of massless theories that we are going to extend to massive cases, with and without topological terms, in the following sections. Proca theory in three dimensions {#sec3} ================================ In order to preserve gauge invariance, we start from lagrangian of the Proca model in the Stüeckelberg form $${\cal L}=-\frac 14F_{\mu\nu}F^{\mu\nu}+\frac 12 m^2(A_\mu+\partial_\mu f) (A^\mu+\partial^\mu f).\label{LagProca}$$ It is a trivial matter to see that the equation of motion associated to the auxiliar field $f$ is nothing but a consistence requisite for the other equation, which is the relevant one. This reflects the invariance of the lagrangian density [(\[LagProca\])]{} under the gauge transformations $$\begin{aligned} \delta A_\mu&=&\partial_\mu\Lambda,\\ \delta f&=&-\Lambda.\end{aligned}$$ As in the Maxwell case, $f$ may be eliminated by choosing $f=0$. To incorporate the dual formulation of the theory [(\[LagProca\])]{}, we take the master lagrangian $${\cal L'}=m\epsilon^{\mu\nu\lambda}\partial_\mu A_\nu B_\lambda+ \frac{m^2}2\left(B_\mu+\partial_\mu\omega\right)\left(B^\mu+\partial^\mu\omega\right)+ \frac{m^2}2\left(A_\mu+\partial_\mu f\right)\left(A^\mu+\partial^\mu f\right), \label{proca1}$$ which is first order in the Proca field $A_\mu$ and the dual field $B_\mu$. Besides $f$ we have introduced the Stüeckelberg field $\omega$, associated to $B_\mu$, to promote gauge invariance. It can be seen that [(\[proca1\])]{} corresponds to two self-dual models [@TPN] with opposite spins. In fact if we do the change $$\begin{aligned} A_\mu&=&\frac1{\sqrt{2}}\left(a_{\mu}^1+a_{\mu}^2\right),\nonumber \\ B_\mu&=&\frac1{\sqrt{2}}\left(a_{\mu}^1-a_{\mu}^2\right),\nonumber \\ f&=&\frac1{\sqrt{2}}\left(f_1+f_2\right),\nonumber \\ \omega&=&\frac1{\sqrt{2}}\left(f_1-f_2\right).\end{aligned}$$ we will get two decoupled self-dual lagrangians (see equation [(\[eq:1.2\])]{} further) in Stückelberg form. Each of them describe a massive mode with spin +1 for one mode and spin -1 for the other [@DJ]. The invariance under P and T transformations is accomplished if we exchange the fields $a_{\mu}^1$ and $a_{\mu}^2$ (and so with the fields $f_1$ and $f_2$). In this sense we see that the field $B_{\mu}$ behaves as a pseudovector. The equations of motion that results after eliminating the Stüeckelberg fields are $$\begin{aligned} \epsilon^{\mu\nu\lambda}\partial_\nu B_\lambda+mA^\mu&=&0,\label{EdM1}\\ \epsilon^{\mu\nu\lambda}\partial_\nu A_\lambda+mB^\mu&=&0.\label{EdM2}\end{aligned}$$ By substitution of $B^\lambda$ from [(\[EdM2\])]{} into [(\[proca1\])]{} we obtain the Proca lagrangian [(\[LagProca\])]{} (with $f=0$). Doing an analogous procedure with $A^\mu$ from [(\[EdM1\])]{} we obtain the same Proca lagrangian, but this time in terms of the dual field $B_\mu$. In this sense, one says that the theory is self dual, and the master lagrangian ${\cal L'}$ is a good starting point to explore the geometrical consequences of this duality. Let us now quantize the theory. From the “BF” term we directly read the commutator[@Faddeev-Jackiw] $$\left[A_i(\vec x),\epsilon^{kj}B_j(\vec y)\right]=i\frac 1{m}\delta_i^k{\delta{^{(2)}}(\vec x-\vec y)}.$$ The Hamiltonian results to be $$\begin{aligned} H=\int d^2\vec x\left(\frac 1{2m^2}\pi_\omega^2+\frac 1{2m^2}\pi_f2+ \frac {m^2}2(B_i+\partial_i\omega)(B_i+\partial_i\omega)\right. \nonumber\\ +\frac{m^2}2(A_i+\partial_i f)(A_i+\partial_i f) -A_0\left(\pi_f+m\epsilon^{ij}\partial_i B_j\right)\nonumber\\ \left.-B_0\left(\pi_\omega+m\epsilon^{ij} \partial_i A_j\right)\right),\label{Ham-Pro}\end{aligned}$$ where $\pi_w$ and $\pi_f$ are the momenta conjugate to $\omega$ and $f$ respectivelly. We did not consider $A_0$ and $B_0$ as canonical variables since their role as lagrange multipiers is clear. They are associate to the first class constraints $$\begin{aligned} \pi_f+m\epsilon^{ij}\partial_i B_j&\approx&0,\label{v1}\\ \pi_\omega+m\epsilon^{ij}\partial_i A_j&\approx&0,\label{v2}\end{aligned}$$ that generate the time independent gauge transformations of the theory. At this point we can compare equation [(\[Ham-Pro\])]{} with the hamiltonian of the Proca theory obtained from the standard action. Starting from equation [(\[LagProca\])]{} with $f=0$, and following the canonical quantization procedure one obtains the hamiltonian $$\begin{aligned} H_{\hbox{Proca}}&=&\int d^2\vec x\left(\frac 14F_{ij}F_{ij}+ \frac{m^2}2A_iA_i+\frac 12\pi_i\pi_i\right.\nonumber\\ &&\qquad\qquad\qquad\left.+{\frac1{2m^2}} \left(\partial_i\pi^i\right)^2\right)\label{HProca},\end{aligned}$$ after solving the second class constraint to eliminate the time component of the vector field. In equation [(\[HProca\])]{}, $A_i$ and $\pi^i$ are canonically conjugate. On the other hand, in the gauge invariant model the Hamiltonian $H$ may be expressed as $$\begin{aligned} H=\int d^2\vec x\left(\frac 12\left(\epsilon^{ij}\partial_iA_j\right)^2+ \frac 12\left(\epsilon^{ij}\partial_iB_j\right)^2\right.\nonumber\\ \left.+\frac {m^2}2\left(B_i+\partial_i\omega\right) \left(B_i+\partial_i\omega\right)+ \frac{m^2}2\left(A_i+\partial_i f\right)\left(A_i+\partial_i f\right)\right).\end{aligned}$$ The equivalence of the two formulations is clear after fixing $f=0$, $\omega=0$ and identifying $\pi^i$ with $m\epsilon^{ij}B_j$. Examining the first class constraints one realizes that the gauge invariant combinations that can be formed from the canonical fields $A_i$, $B_i$, $f$ and $\omega$ are $A_i+\partial_if$ and $B_i+\partial_i\omega$. Hence, it is natural to introduce the non local Wilson like operators $$\begin{aligned} W(\gamma_x^{x'})&=&\exp{\left(ie\int_{\gamma_x^{x'}}dx^i \left(A_i+\partial_i f\right)\right)},\label{3.18}\\ \Omega(\Gamma_y^{y'})&=&\exp{\left(ie\int_{\Gamma_y^{y'}}dx^i \left(B_i+\partial_i\omega\right)\right),}\label{3.19}\end{aligned}$$ where $\gamma_x^{x'}$ ($\Gamma_y^{y'}$) is an open curve in $R^2$, starting at $x$ ($y$) and ending at $x'$ ($y'$) and e is a constant with units $L^{-\frac 12}$. These operators play the role of the Wilson Loop and its dual (equations [(\[WL-Maxwell\])]{},[(\[WD-Maxwell\])]{}) in the four dimensional Maxwell theory. In virtue of the constraints [(\[v1\])]{}, [(\[v2\])]{} the introduction of non-local operators associated with $\pi_f$ and $\pi_{\omega}$ would be redundant. In fact, the exponential of $i$ times the integral of $\pi_{\omega}$ over the region of ${\cal R}^2$ bounded by a closed contour $\cal C$ is equivalen to $W(\cal C)$. A similar argument holds for $\pi_f$ and $\Omega$. It is simple to show that the operators [(\[3.18\])]{},[(\[3.19\])]{} obey $$W(\gamma)\Omega(\Gamma)=e^{-i{\frac {e^2}m}N(\gamma,\Gamma)}\Omega(\Gamma)W(\gamma), \label{DA-Proca}$$ where $$\begin{aligned} N(\gamma,\Gamma)&=&\int_\gamma dy^i\int_\Gamma dx^j\epsilon^{ij}{\delta{^{(2)}}(\vec x-\vec y)},\nonumber \\ &=&\int_{\Gamma} dx^j\epsilon^{ij}T^i(\vec x,\gamma), \nonumber \\ &=&\int_{\gamma} dx^i\epsilon^{ij}T^j(\vec x,\Gamma), \label{NumCort}\end{aligned}$$ is the oriented number of intersections between the curves $\gamma$ and $\Gamma$. This topological quantity obeys the relations $$\begin{aligned} N(\gamma,\Gamma)&=&-N(\Gamma,\gamma),\\ N(\gamma_1,\gamma_2\cdot\gamma_3)&=&N(\gamma_1,\gamma_2)+ N(\gamma_1,\gamma_3).\end{aligned}$$ Ec.[(\[DA-Proca\])]{} is the Dual Algebra for the $D=2+1$ Proca theory. As it was pointed out in the introducction, its topological “structure constant” involves the intersection index of the curves that enter in the definition of the non-local operators, instead of linking numbers, which are the objects that appear in the DA of masless theories[@LL]. Our next step will be to realize the DA [(\[DA-Proca\])]{} in an appropriate geometrical representation. To this end we employ the Abelian open path representation, which has been discussed in references[@LRC; @Rolando]. The main features may be summarized as follows. One groups the piecewise continue (and not necessarily closed) curves of $R^2$ in equivalence classes characterized by the equality of their form factors $T(\vec x,\gamma)$. Then the usual composition of curves turns into a group product. It is a trivial matter to show that the Abelian group of loops is a subgroup of the Abelian group of open paths. Besides the loop derivative eq. [(\[LD\])]{}, one can define the path derivative[@LRC; @LO] $$h^i\delta_i(\vec x)\Psi(\gamma)=\Psi(\delta\gamma\circ\gamma)- \Psi(\gamma),\label{PD}$$ which computes the variation of a path-dependent function when an infinitesimal open path $\delta\gamma_x^{x+h}$ going from $x$ to $x+h$ ($h\rightarrow 0$) is appended to $\gamma$. It is related to the loop derivative through the expression $$\Delta_{ij}(\vec x)=\partial_i\delta_j(\vec x)-\partial_j\delta_i(\vec x).$$ The DA [(\[DA-Proca\])]{} may be realized onto open-path dependent wave functionals in the form $$\begin{aligned} W(\gamma)\Psi(\gamma_1)&=&\Psi(\gamma\circ\gamma_1),\label{48}\\ \Omega(\Gamma)\Psi(\gamma_1)&=&e^{i{\frac {e^2}m}N(\Gamma,\gamma_1)} \psi(\gamma_1).\label{49}\end{aligned}$$ As in the Maxwell case, we have chosen a geometric representation in which the non local operator associated to the “direct” field, i.e., the Wilson path, produces a “translation” in path-space, while that associated to the dual field is diagonal. One could also interchange these roles. Since the theory is self-dual, the dual geometric representation results to be a path representation too. With the use of the derivative [(\[PD\])]{}, the basic local observables of the theory may be obtained from the non local dual operators $$\begin{aligned} A_i+\partial_if& = &\left.-i{\frac {\delta_i(\vec x)}e}W(\gamma) \right|_{\gamma=0},\\ B_i+\partial_i\omega& = &\left.-i{\frac {\delta_i(\vec x)}e}\Omega(\Gamma) \right|_{\Gamma=0}.\end{aligned}$$ As we show in the Appendix, the program developed in this section can also be carried out for massive $p$-forms in arbitrary dimensions. In $D$ dimensions, Abelian massive theories of $p_1$ and $p_2$-forms are dual for $p_1 + p_2 = D-1$. For instance, the four dimensional Proca theory is dual to the massive Kalb-Ramond model. On the other hand, massive $p$-form theories are associated to generalized Wilson Surfaces $W(\Sigma_p)$, where $\Sigma_p$ is an open $p$-surface. Then the Dual Algebra generalizes to $$W(\Sigma_{p_1})\Omega(\Sigma_{p_2})=e^{-iN(\Sigma_{p_1},\Sigma_{p_2})}\Omega(\Sigma_{p_2})W(\Sigma_{p_1}),$$ where $N(\Sigma_{p_1},\Sigma_{p_2})$ is the intersection index of the open surfaces $\Sigma_{p_1}$ and $\Sigma_{p_2}$. Self-Dual and Topologically Massive theories {#sec4} ============================================ Master Lagrangian and Canonical Quantization -------------------------------------------- It is well known that the Topologically Massive [@DJT] $$\label{eq-1.1} S_{TM}=\int d^3x\left(-\frac 14F_{\mu\nu}F^{\mu\nu}+ \frac m{4}{\varepsilon^{\mu\nu\lambda}}F_{\mu\nu}(A)A_\lambda\right),$$ and Self-Dual theories [@TPN] $$\label{eq:1.2} S_{AD}=\int d^3x\left(-\frac m{4}{\varepsilon^{\mu\nu\lambda}}F_{\mu\nu}A_\lambda+ \frac{m^2}2A_\mu A^\mu\right),$$ provide locally equivalent descriptions of spin $1$ massive particles in $2+1$ dimensions [@DJ], although they exhibit different global behaviors depending on the topological properties of the space time where they are defined[@PR; @ST]. The local equivalence between these models may be viewed by noticing that they are dual, in the sense that both may be obtained from the master action [@DJ] $$\label{eq:1-1} S_M=\frac m{2}{\int d^3x\left({\varepsilon^{\mu\nu\lambda}}F_{\mu\nu}(A)\left(C_\lambda+ \frac 12A_\lambda\right)+ mC_\mu C^\mu\right)},$$ where $F_{\mu\nu}=\partial_{\mu}A_{\nu}-\partial_{\nu}A_{\mu}$. We shall take $g_{\mu\nu}= \hbox{diag}(1, -1,-1)$. The equations of motions, obtained by varying the independent fields $A_\mu,C_\mu$, are $$\begin{aligned} \label{eq:1-2} &&{\varepsilon^{\mu\nu\lambda}}\partial_\nu\left(C_\lambda+A_\lambda\right)=0,\\ &&{\varepsilon^{\mu\nu\lambda}}\partial_{\nu}A_{\lambda}+ mC^\mu=0\label{eq:1-3}.\end{aligned}$$ Using equation [(\[eq:1-2\])]{} to eliminate $A_\lambda$ in [(\[eq:1-3\])]{} we obtain the equations of motion for the SD model. In other direction, from equation [(\[eq:1-3\])]{} we can eliminate $C_\mu$ in [(\[eq:1-2\])]{} to obtain the equations of of motion of the TM model. This proves the local classical equivalence. As in the previous sections, the “$C^2$” term spoils gauge invariance. We remedy this fact by introducing an auxiliary Stüeckelberg field $\omega$ $$\begin{aligned} S_M'=\frac m{2}\int d^3x\left({\varepsilon^{\mu\nu\lambda}}F_{\mu\nu}(A)\left(C_\lambda+\frac 12A_\lambda\right)\right.\nonumber\\ +m(C_\mu+\partial_\mu\omega)(C^\mu+\partial^\mu\omega)\Big)\label{eq:1-4}.\end{aligned}$$ This action is invariant under the simultaneous gauge transformations $$\begin{aligned} \label{eq:1-5} \delta A_\mu&=&\partial_\mu\xi,\\ \delta C_\mu&=&\partial_\mu\zeta,\label{eq:1-6}\\ \delta\omega&=&-\zeta\label{eq:domega}.\end{aligned}$$ From equation (\[eq:domega\]) we see that the field $\omega$ is pure gauge, as corresponds to the Stüeckelberg formulation. Now we apply the canonical procedure of quantization to the master model. First, we decompose the action [(\[eq:1-4\])]{} into spatial and temporal parts $$\begin{aligned} S_M'&=&\int d^3x\left(m\varepsilon^{ij}F_{0i}\left(C_j+ \frac 12A_j\right)\nonumber\right.\\ &&+\frac m{2}\varepsilon^{ij}F_{ij}\left(C_0+\frac 12 A_0\right)+ \frac m{2}(C_0+\dot\omega)^2\nonumber\\ &&\left.-\frac m{2}(C_i+\partial_i\omega) (C_i+\partial_i\omega)\right)\label{eq:1-7}.\end{aligned}$$ The canonical momenta conjugate to the dinamical variables $A_i$, $C_i$ and $\omega$ are $$\begin{aligned} \label{eq:1-8} \pi^i_A&=&m\varepsilon^{ij}\left(C_j+\frac 12 A_j\right),\\ \pi^i_C&=&0,\label{eq:1-9}\\ \pi_\omega&=&m(C_0+\dot\omega)\label{eq:1-10}.\end{aligned}$$ We consider the fields $A_0$ and $C_0$ as non dynamical. They will appear in the next step as Lagrange multipliers. Equations [(\[eq:1-8\])]{} and [(\[eq:1-9\])]{} are just primary constraints among the phase space variables $$\begin{aligned} \label{eq:1-11} \psi_i^A&\equiv& \pi^i_A-m\varepsilon^{ij}\left(C_j+\frac 12 A_j\right)\approx 0,\\ \psi_i^C&\equiv&\pi^i_C\approx 0\end{aligned}$$ while equation [(\[eq:1-10\])]{} allows us to obtain the velocities associated to $\omega$. Thus, the Hamiltonian on the manifold defined by the primary constraints is given by $$\label{eq:1-13} H=\int d^2\vec x\left(\frac1{2m^2}\pi_{\omega}^2+\frac{m^2}{2}(C_i+\partial_i\omega)^2 +A_0\theta_1+C_0\theta_2\right),$$ where $$\label{eq:1-14} \theta_1\equiv-m\varepsilon^{ij}\partial_i(A_j+C_j); \quad\theta_2\equiv-m\varepsilon^{ij}\partial_iA_j -\pi_\omega.$$ Following the scheme of quantization of Dirac, we extend the Hamiltonian to the whole phase space: $$\begin{aligned} \tilde H&=&\int d^2\vec x\left(\frac1{2m^2}\pi_{\omega}^2+ \frac{m^2}{2}(C_i+\partial_i\omega)(C_i+\partial_i\omega)\right.\nonumber\\ &&\left.+A_0\theta_1+C_0\theta_2+\lambda_A^i\psi_i^A+\lambda_C^i\psi_i^C\right).\label{eq:1-15}\end{aligned}$$ At this point we observe that the variables $A_0$ and $C_0$ are the Lagrange multipliers associated to the “secondary” constraints $\theta_1$ and $\theta_2$, respectively. Now, we define the Poisson Brackets among the canonical variables by $$\begin{aligned} \label{eq:1-16} \left\{A_i(\vec x),\pi^j_A(\vec y)\right\}&=&\delta_i^j{\delta{^{(2)}}(\vec x-\vec y)}\\ \left\{C_i(\vec x),\pi^j_C(\vec y)\right\}&=&\delta_i^j{\delta{^{(2)}}(\vec x-\vec y)}\\ \left\{\omega(\vec x)\pi_\omega(\vec y)'\right\}&=&{\delta{^{(2)}}(\vec x-\vec y)}\end{aligned}$$ (the remaining Poisson brackets vanish) and proceed to require the preservation in time of the contraints, taking $\tilde H$ as the generator of time translations. This leads to determine the Lagrange multipliers associated to the primary constraints [(\[eq:1-8\])]{} and [(\[eq:1-9\])]{} $$\begin{aligned} \label{eq:1-17} \lambda_A^i=\partial_iA_0+m\varepsilon^{ij}(C_j+\partial_j\omega),\\ \lambda_C^i=\partial_iC_0-m\varepsilon^{ij}(C_j+\partial_j\omega),\end{aligned}$$ and it is seen that no further secondary constraints arise. Substituting the multipliers into the Hamiltonian yields $$\begin{aligned} \label{eq:1-19} \tilde H =\int d^2\vec x\left(\frac1{2m^2}{\pi_{\omega}^2}+ \frac{m^2}{2}(C_i+\partial_i\omega)(C_i+\partial_i\omega)+ A_0{\varphi}_1\right.\nonumber\\ +C_0{\varphi}_2+\mu\varepsilon^{ij}(C_j+\partial_j\omega)\psi_i^A\nonumber\\ \left.-\mu\varepsilon^{ij}(C_j+\partial_j\omega)\psi_i^C\right),\end{aligned}$$ where $$\begin{aligned} \label{eq:1-20} {\varphi}_1&\equiv&-\partial_i\psi_i^A+\theta_1= -\partial_i\pi^i_A-\frac m{2}\varepsilon^{ij}\partial_iA_j,\nonumber\\ {\varphi}_2&\equiv&-\partial_i\psi_i^C+\theta_2= -\partial_i\pi^i_C-\pi_{\omega}-m\varepsilon^{ij}\partial_iA_j,\end{aligned}$$ result to be the first class constraints of the theory. The matrix associated to the second class constraints $\Psi_a\equiv(\psi_i^A,\psi_i^C)$ may be written down as $$\label{eq:1-21} C_{ab}(\vec x,\vec y)\equiv\left\{\Psi_a(\vec x),\Psi_b(\vec y)\right\}= -m\varepsilon^{ij}\left( \begin{array}{ll} 1&1\\ 1&0 \end{array}\right){\delta{^{(2)}}(\vec x-\vec y)}.$$ Its inverse is given by $$\begin{aligned} C^{-1}_{ab}(\vec x,\vec y)&\equiv&\left\{\Psi_a(\vec x), \Psi_b(\vec y)\right\}^{-1}\nonumber\\ &=&-\frac{1}m\varepsilon^{ij}\left( \begin{array}{cc} 0&1\\ 1&-1 \end{array}\right){\delta{^{(2)}}(\vec x-\vec y)}\label{eq:1-22},\end{aligned}$$ and allows us to define the Dirac brackets $${\left\{F,G\right\}^*}\equiv\left\{F,G\right\} -\int d^2\vec xd^2\vec y\left\{F,\Psi_a(\vec x)\right\} C^{-1}_{ab}(\vec x,\vec y)\left\{\Psi_b(\vec y),G\right\}.\nonumber\\$$ Recalling that Dirac brackets are consistent with second class constraints, we can eliminate $\pi^i_A$ and $\pi^i_C$ from now on, employing the constraints $\Psi_a$. The brackets between the reduced phase space variables are then $$\begin{aligned} {\left\{A_i(\vec x),A_j(\vec y)\right\}^*}&=&0\label{DBA-1}\\ {\left\{A_i(\vec x),C_j(\vec y)\right\}^*}&=&\frac1{m}\epsilon^{ij}{\delta{^{(2)}}(\vec x-\vec y)}\\ {\left\{C_i(\vec x),C_j(\vec y)\right\}^*}&=&-\frac1{m}\epsilon^{ij}{\delta{^{(2)}}(\vec x-\vec y)}\\ {\left\{\omega(\vec x),\pi_\omega(\vec y)\right\}^*}&=&{\delta{^{(2)}}(\vec x-\vec y)}\label{DBA-Ult}.\end{aligned}$$ Once the phase space has been reduced, the first class constraints become $$\theta_1\approx 0\qquad,\qquad\theta_2\approx 0.$$ and it can be seen that the time independent gauge transformations generated by these constraints on the reduced phase space variables are given by $$\begin{aligned} \delta A_i(\vec x)&=&\partial_i\xi,\\ \delta C_i(\vec x)&=&\partial_i\zeta,\\ \delta\omega(\vec x)&=&-\zeta,\\ \delta\pi_\omega&=&0,\end{aligned}$$ as expected. The next step in the quantization procedure is to promote the fields to operators acting on a Hilbert space $\cal H$, obeying commutation relations given by $i{\left\{\cdot,\cdot\right\}^*}$, and ask the physical vectors ${\left|\psi\right\rangle}$ to belong to the kernel of both first class constraint operators: $\theta_1{\left|\psi\right\rangle}=0$ and $\theta_2{\left|\psi\right\rangle}=0$ . The basic observables (in the sense of Dirac), from which all relevant gauge invariant information of the theory can be recovered, are the operators $-\epsilon^{ij}\partial_i A_j$, $C_i+\partial_i \omega$ and $\pi_\omega$. It is then natural, within the spirit of the previous sections, to introduce the non-local operators $$\begin{aligned} W(\cal C)&\equiv&\exp{\left(ie\oint_{\cal C}A_idx^i\right)},\label{WL}\\ \Omega(\gamma_x^{x'})&\equiv&\exp{\left(ie\int_{\gamma_x^{x'}} \left(C_i+\partial_i\omega\right)dx^i\right)}.\label{Wpath}\end{aligned}$$ In this expression, $e$ is a constant with dimensions $L^{-\frac12}$. The Wilson Loop ($W$) and Wilson Path ($\Omega$) operators, depend on the closed and open paths $\cal C$ and $\gamma$ respectivelly. It is easy to see from equations [(\[DBA-1\])]{}-[(\[DBA-Ult\])]{} that these operators obey $$\begin{aligned} W({\cal C})W({\cal C'})&=&W({\cal C'})W({\cal C})\label{dual-al1}\\ W({\cal C})\Omega(\gamma)&=&e^{-i\frac{e^2}m N({\cal C},\gamma)} \Omega(\gamma)W({\cal C})\label{dual-al2}\\ \Omega(\gamma)\Omega(\Gamma)&=&e^{i\frac{e^2}m N(\gamma,\Gamma)} \Omega(\Gamma)\Omega(\gamma)\label{dual-al3},\end{aligned}$$ where $N(\gamma,\Gamma)$, the oriented number of intersections between $\gamma$ and $\Gamma$, was defined in equation [(\[NumCort\])]{}. One could also introduce a non local operator associated to $\pi_\omega$: the exponential of $i$ times the integral of $\pi_\omega$ over the region of ${\cal R}^2$ bounded by a closed contour $\cal C$. However, in virtue of the first class constraint $\theta_2\approx 0$, this operator would be just another representation for the Wilson Loop [(\[WL\])]{} when restricted to the physical space of states, so we do not gain anything with its introduction. It should be remarked that, as in the previous cases, the local gauge invariant operators may be obtained from the non-local ones. In fact, the local operator $C_i+\partial_i\omega$ can be recovered from the Wilson path by considering an infinitesimal open path $\delta\gamma$, i.e. $$\Omega(\delta\gamma)=1+ie\delta\gamma^i\left(C_i+\partial_i\omega\right)+ O(\delta\gamma^2).\label{C-O}$$ On the other hand, if $\delta\gamma$ is closed, we have $$\Omega(\delta\gamma)=1+ie\delta\sigma^{ij}\tilde F_{ij}+O(\delta\sigma^2),$$ where $\tilde F_{ij}\equiv\partial_iC_j-\partial_jC_i$. In a similar way, the local gauge invariant operator $F_{ij}=\partial_iA_j-\partial_jA_i\equiv\epsilon_{ij}B$ and the Wilson Loop are related through $$W(\delta\gamma)=1+ie\delta\sigma^{ij}\epsilon_{ij}B+O(\delta\sigma^2),\label{B-W}$$ where $\delta\gamma$ is an infinitesimal loop. From the latter expansions it is straight forward to see that $$\begin{aligned} C_i+\partial_i\omega&=&\left.-\frac{i}{e}\delta_i(\vec x)\Omega(\gamma) \right|_{\gamma=0},\\ B&=&\left.-\frac{i}{2e}\epsilon^{ij}\Delta_{ij}(\vec x)W(C) \right|_{C=0}.\end{aligned}$$ Finally, the evolution of the physical states is governed by the Schrödinger equation $$i\frac d{dt}{\left|\psi(t)\right\rangle}=\int d^2\vec x\left(\frac12B^2 + m^2(C_i+\partial_i\omega)^2\right){\left|\psi(t)\right\rangle}.$$ Geometrical Representation and the Dual Algebra ----------------------------------------------- The algebra of non local operators [(\[dual-al1\])]{}- [(\[dual-al3\])]{} may be realized on the space of open-path dependent functionals $\psi(\gamma)$ of section \[sec3\], if we prescribe $$\begin{aligned} W(\cal C)\psi(\gamma)&\equiv&\exp{\left(i\frac{e^2}m N({\cal C,\gamma})\right)}\psi(\gamma)\label{realization1}\\ \Omega(\Gamma)\psi(\gamma)&\equiv& \exp{\left(-i\frac{e^2}{2m}N(\Gamma,\gamma)\right)} \psi(\Gamma\circ\gamma).\label{realization2}\end{aligned}$$ For instance, using eq. [(\[realization2\])]{} one has $$\begin{aligned} \Omega(\Gamma)\Omega(\Gamma')\psi(\gamma)&=&\Omega(\Gamma) \left[e^{-i\frac{e^2}{2m}N(\Gamma',\gamma)}\psi(\Gamma'\circ\gamma)) \right]\nonumber\\ &=&e^{-i\frac{e^2}{2m} N(\Gamma,\gamma)}e^{-i\frac{e^2}{2m} N(\Gamma',\Gamma\circ\gamma)}\psi(\Gamma\circ\Gamma'\circ\gamma)\nonumber\\ &=&e^{i\frac{e^2}m N(\Gamma,\Gamma')}\Omega(\Gamma')\Omega(\Gamma)\psi(\gamma),\end{aligned}$$ in agreement with eq. [(\[dual-al3\])]{}. Besides realizing the non-local algebra, we have to consider the restrictions that the first class constraints impose onto the path dependent states. The constraint $\theta_2$ is automatically satisfied if the non local operator associated to $\pi_\omega$ (see comment after [(\[dual-al3\])]{}) is realized as (essentially) the Wilson Loop $W$. So it remains to study the constraint $\theta_1$ : $$\epsilon^{ij}\partial_i(A_j+C_j)\approx 0,$$ which may be imposed onto the states as $$\begin{aligned} \exp{\left(ie\oint_{{\cal C}}(\hat A_i+\hat C_i)dx^i\right)}{\left|\psi\right\rangle} &=&\exp\left(-\frac12\oint_{\cal C}\oint_{\cal C}dx^idy^j{\left[\hat A_i(\vec x),\hat C_j(\vec y)\right]}\right)W({\cal C})\Omega({\cal C}){\left|\psi\right\rangle},\nonumber \\ &=&W({\cal C})\Omega({\cal C}){\left|\psi\right\rangle},\nonumber \\ &=&{\left|\psi\right\rangle}, \label{Vin-NL}\end{aligned}$$ or, in other words, $$W({\cal C}) \approx \Omega(\overline{\cal C}), \label{phys}$$ within the physical sector of the Hilbert Space. To obtain this result we used that $N({\cal C}_1 , {\cal C}_2)= 0$ for closed paths ${\cal C}_1 , {\cal C}_2$. Equation [(\[phys\])]{} states that instead of a pair of “Wilson” operators it suffices with just one of them, namely, $\Omega(\gamma)$ , that simultaneously plays the role of “coordinate” and “momentum”. That $N({\cal C}_1 , {\cal C}_2)$ vanishes when the curves are closed also matters to see that when eq. [(\[phys\])]{} holds (i.e., on the physical sector), equation [(\[dual-al3\])]{} already implies eqs. [(\[dual-al1\])]{} and [(\[dual-al2\])]{}. Therefore, it is just eq. [(\[dual-al3\])]{} which corresponds to the Dual Algebrae of the previous sections (equations [(\[AlgDual\])]{} and [(\[DA-Proca\])]{}), with which it should be compared. So far, it remains to study how eq. [(\[phys\])]{} restricts the space of states. Combining equation [(\[phys\])]{} with [(\[realization1\])]{} and [(\[realization2\])]{} one obtains $$\psi({\cal C}\circ\gamma)=e^{-i\frac{e^2}{2m}N({\cal C},\gamma)} \psi(\gamma)\label{vin-int}.$$ Taking ${\cal C}$ to be an infinitesimal closed path we can see that this equation is just the non-local version of the differential constraint recently obtained in a study of the path-space quantization of the Maxwell-Chern-Simons theory [@LO] $$\left(\rho(\vec x,\gamma)-i\frac m {e^2}\epsilon^{ij}\Delta_{ij}(\vec x) \right)\psi(\gamma)=0, \label{ppp}$$ where $$\begin{aligned} \rho(\vec x,\gamma) & \equiv & -\partial_i T^i(\vec x,\gamma)\nonumber\\ &=& -\sum_s (\delta^2(\vec x - \vec {\beta}_s) -\delta^2(\vec x - \vec {\alpha}_s)),\label{ro}\end{aligned}$$ is a functional that depends on the boundary $(\vec\alpha,\vec\beta\,)$ of the path $\gamma$. Since the path may comprises several pieces $s$, both $\vec\alpha$ and $\vec\beta$ denote the set of starting and ending points, respectivelly. It can be said then that $\rho(\vec x,\gamma)$ is the “density of extremes” of the path $\gamma$. The solution to [(\[ppp\])]{} was found to be[@LO] $$\psi(\gamma)=e^{i\chi(\gamma)}\Phi(\vec\alpha,\vec\beta),\label{oswaldo}$$ where $\Phi(\vec\alpha,\vec\beta\,)$ is an arbitrary functional of the boundary $(\vec\alpha,\vec\beta\,)$ and $$\begin{aligned} \chi(\gamma)&=&-i\frac{e^2}{4\pi m}\sum_s\int_\gamma dx^i\epsilon^{ij}\left(\frac{(x^j-\beta^j_s)}{|\vec x-\vec\beta_s|^2} -\frac{(x^j-\alpha^j_s)}{|\vec x-\vec\alpha_s|^2}\right)\nonumber\\ &=&-i\frac{e^2}{4\pi m}\Delta\Theta(\gamma),\label{chi}\end{aligned}$$ with $\Delta\Theta(\gamma)$ being the algebraic sum of the angles subtended by the pieces of $\gamma$, measured from their ending points $\vec\beta$, minus that measured from their starting points $\vec\alpha$. The path-dependent function $\chi(\gamma)$ is ill defined due to the ambiguous definition of the angle subtended by a path when it is measured from their own ending points. In fact, when the point from which the angle is measured coincides with one of the extremes of the path, one loses the straight line connecting that point with the extreme, which would serve as a reference to compute the desired angle. We can replace that fidutial straight line by the tangent to the path at the problematic point. For instance, if we want to compute the angle subtended by the path $\gamma$, given as a map from the interval $[0,1]$ to $R3$, measured from its starting point $\vec y(0)=\vec\alpha$, we can take the prescription $$\Theta(\gamma,\vec\alpha)\equiv\lim_{a\rightarrow0^+}\int_a^1dt\frac{dy^i(t)}{dt} \epsilon^{ij}\frac{(y^j(t)-\alpha^j)}{|\vec y-\vec\alpha|^2}.$$ It may be seen that this prescription is consistent with the fact that $\chi(\gamma)$ must be a path-dependent function, and not merely a curve-dependent one. It is worth noticing, from eq. [(\[oswaldo\])]{}, that the path dependence of the wave functionals is realized through the boundary points of the paths, and through the way they wind around these points. Hence, we see that in this case not only the DA shows a topological character, but also the geometrical representation of the algebra carries a topological content. Equations [(\[C-O\])]{}-[(\[B-W\])]{} together with the realization [(\[realization1\])]{} and [(\[realization2\])]{} for the DA allow us to see that the gauge invariant operators $B(\vec x)$ and $C_i(\vec x)+\partial_i\omega(\vec x)$ are realized as $$\begin{aligned} B(\vec x)&\rightarrow&-i\frac{e^2}m\rho(\vec x,\gamma)\\ C_i(\vec x)+\partial_i\omega(\vec x)&\rightarrow&-i{\cal D}_i(\vec x) \equiv -i\delta_i(\vec x)-\frac{e^2}{2m}\epsilon^{ij}T^j(\vec x,\gamma),\end{aligned}$$ whose action on gauge invariant functionals can be seen to respect the form given in eq.[(\[oswaldo\])]{} [@LO]. The same is then true for the non-local operator $\Omega({\gamma})$, in view of its definition [(\[Wpath\])]{}. Now, let us quote the path-representation expressions for the Poincare generators of the theory: $$\begin{aligned} H&=&\frac{m^2}{e^2}\int d^2\vec x\left[-\Delta_{ij}(\vec x)\Delta_{ij}(\vec x) -\frac 12{\cal D}_i(\vec x){\cal D}_i(\vec x)\right]\label{Real-H}\\ P^i&=&\frac{m}{2e^2}\int d^2\vec x\epsilon^{jk}\Delta_{jk}(\vec x)i{\cal D}_i(\vec x),\\ J&=&\frac{m}{e^2}\int d^2\vec x\epsilon^{ij}x^i\epsilon^{kl}\Delta_{kl}(\vec x){\cal D}_j(\vec x)\end{aligned}$$ It can be shown that the operator $P^i$ ($J$) generates rigid translations (rotations) of the path $\gamma$ appearing in the argument of the wave functional $\psi(\gamma)$, as it should be expected. Since $\chi(\gamma)$ is invariant under both translations and rotations, the above results does not contradict the fact that $P^i$ and $J$ are gauge invariant operators. In other words, one has, for an infinitesimal translation along $u^i$ $$\left(1+u^iP^i\right)\psi(\gamma)=e^{i\chi(\gamma)} \left(1+u^iP^i\right)\Phi(\vec\alpha,\vec\beta).$$ Thus, $P^i$ translates the boundary $(\vec\alpha,\vec\beta)$ of the path while maintaining the form of the wave functional given by [(\[oswaldo\])]{}, which is dictated by gauge invariance. A similar argument holds for infinitesimal rotations. Discussion ========== We have studied the duality symmetry between massive Abelian $p-$forms, with and without topological terms, from the point of view of the geometrical representations that, in each case, generalize the loop representation of Maxwell theory. We found that in the cases without topological terms, and within the physical sector of the Hilbert space, the canonical algebra of local operators can be translated into a non local albegra of a pair of gauge invariant operators, that exhibit an interesting geometrical content, and that is characterized by a topological quantity, namely, the intersection index between the geometrical objects that constitute the argument of the gauge invariant operators. This algebra, the Dual Algebra, may be realized in a basis of wave functionals depending on open paths, or $p-$surfaces, according to the rank of the forms involved. In general, for any pair of dual theories, there is also a pair of dual geometric representations. This situation degenerates in the case of self-duality, since then the “direct” and the “dual” theories are equivalent. Regarding the study of the TM and SD case theories in $2+1$ dimensions, we found that, as in the Proca model in $2+1$ dimensions, the topological quantity that characterizes the DA is the number of intersections of two paths. However, unlike the Proca’s case, these paths are different arguments of the same operator. Another important difference is that in the TM and SD models, the open paths involved fall into equivalence classes labeled by their boundary $\partial\gamma$ and their winding properties described by $\Delta\Theta(\gamma)$. One could say that in this case the geometric representation in one of “rubber bands with fixed ends”, rather than a path representation. The results of this study could contribute to put both massless and massive Abelian gauge theories under a common scope, regarding their geometrical properties. It remains to explore whether or not these ideas find a suitable extension to the non-Abelian case. Also, it would be interesting to study how the equivalence between the Proca model and two selfdual models with opposite spins is manifested in the geometrical representation. Conventional Massive Theories {#Ap-A} ============================= In this appendix we discuss briefly how to extend the results of section \[sec3\] to the general case of duality between massive Abelian forms $A_{\mu_1\cdots\mu_p}$ and $B_{\mu_1\cdots\mu_{D-p-1}}$, for $p=0,1,\cdots,D-1$, in $D$ spacetime dimensions. We start from the Stückelberg form of the master action, i.e. $$\begin{aligned} I_M{^{(p,D)}}&=&\int d^Dx\left(\frac{g(-1)^{pD}}{(p+1)!(D-p-1)!}\epsilon^{\mu_1\cdots\mu_{p+1}\nu_1 \cdots\nu_{D-p-1}}F_{\mu_1\cdots\mu_{p+1}}(A)B_{\nu_1\cdots\nu_{D-p-1}} -\right.\nonumber\\ &&\qquad\qquad-\frac{g(-1)^p}{2(D-p-1)!}(B_{\mu_1\cdots\mu_{D-p-1}}+ F_{\mu_1\cdots\mu_{D-p-1}}(C))^2 -\nonumber\\ &&\qquad\qquad\qquad\left.-\frac{g(-1)^p\mu^2}{2p!}(A_{\mu_1\cdots\mu_p}+ F_{\mu_1\cdots\mu_p}(\omega))^2\right),\label{general}\end{aligned}$$ with $F(f) \equiv df$ for any form $f$. Here, $\omega$ and $C$ are auxiliary Stückelberg $p-1$ and $D-p-2$ forms respectivelly. From the space-time decomposition of the master action [(\[general\])]{}, which is given by $$\begin{aligned} I_M{^{(p,D)}}&=&\int d^Dx\left({\frac 1{2(D-p-2)!}} (B_{0i_1\cdots i_{D-p-2}}+ F_{0i_1\cdots i_{D-p-2}}(C))^2\right.\nonumber\\ &&-{\frac 1{2(D-p-1)!}}(B_{i_1\cdots i_{D-p-1}}+ F_{i_1\cdots i_{D-p-1}}(C))^2\nonumber\\ &&+{\frac{\mu^2}{2(p-1)!}} (A_{0i_1\cdots i_{p-1}}+ F_{0i_1\cdots i_{p-1}}(\omega))^2-{\frac{\mu^2}{2p!}} (A_{i_1\cdots i_p}+F_{i_1\cdots i_p}(\omega))^2\nonumber\\ &&+{\frac{g(-1)^{pD}}{p!(D-p-1)!}} \epsilon^{i_1\cdots i_pj_1\cdots j_{D-p-1}} \dot A_{i_1\cdots i_p}B_{j_1\cdots j_{D-p-1}}\nonumber\\ &&-{\frac{g(-1)^{pD}}{(p-1)!(D-p-1)!}} \epsilon^{i_1\cdots i_pj_1\cdots j_{D-p-1}} \partial_{i_2} A_{0i_2\cdots i_{p-1}}B_{j_1\cdots j_{D-p-1}}\nonumber\\ &&\left.-{\frac{g(-1)^{pD}}{p!(D-p-2)!}}(-1)^p \epsilon^{i_1\cdots i_{p+1}j_1\cdots j_{D-p-2}}\partial_{i_1} A_{i_2\cdots i_{p+1}}B_{0j_1\cdots j_{D-p-2}}\right),\nonumber\\\end{aligned}$$ we can read the fundamental Poisson bracket $$\left\{A_{i_1\cdots i_p}(\vec x),B_{j_1\cdots j_{D-p-1}} (\vec y)\right\}= g\epsilon^{i_1\cdots i_pj_1\cdots j_{D-p-1}} \delta{^{(d)}}(\vec x-\vec y),\label{Alg-MasLoc}$$ and obtain the Hamiltonian as $$\begin{aligned} H_M{^{(p,D)}}&=&\int d^d\vec x\left\{\frac 1{2(D-p-2)!}\left(\pi_C^{i_1\cdots i_{D-p-2}}\right)^2+ \frac 1{2(p-1)!} \left(\pi_\omega^{i_1\cdots i_{p-1}}\right)^2\right.\nonumber\\ &&+\frac 1{2(D-p-1)!} \left(B_{i_1\cdots i_{D-p-1}}+F_{i_1\cdots i_{D-p-1}}(C)\right)^2+ \frac 1{2p!}\left(A_{i_1\cdots i_p}+ F_{i_1\cdots i_p}(\omega)\right)^2\nonumber\\ &&+B_{0i_1\cdots i_{D-p-2}} \Theta_1^{i_1\cdots i_{D-p-2}}+ C_{0i_1\cdots i_{D-p-3}} \Theta_2^{i_1\cdots i_{D-p-3}}+ \omega_{0i_1\cdots i_{p-2}}\Theta_3^{i_1\cdots i_{p-1}}\nonumber\\ &&\left.+A_{0i_1\cdots i_{p-1}} \Theta^{i_1\cdots i_{p-1}}\right\}.\end{aligned}$$ In this equation, the quantities $$\begin{aligned} \Theta_1^{i_1\cdots i_{D-p-2}}&=&-\frac 1{(D-p-2)!} \left[\pi_C^{i_1\cdots i_{D-p-2}}+\frac{g(-1)^{p(D+1)}}{p!} \epsilon^{j_1\cdots j_{p+1}i_1\cdots i_{D-p-2}} \partial_{j_1}A_{j_2\cdots j_{p+1}}\right]\nonumber\\ \Theta_2^{i_1\cdots i_{D-p-3}}&=&-\frac 1{(D-p-3)!}\partial_i\pi_C^{ii_1\cdots i_{D-p-3}},\nonumber\\ \Theta_3^{i_1\cdots i_{p-2}}&=& -\frac 1{(p-2)!}\partial_i\pi_\omega^{ii_1\cdots i_{p-2}},\\ \Theta_4^{i_1\cdots i_{p-1}}&=& \frac 1{(p-1)!}\left[\pi_\omega^{i_1\cdots i_{p-1}}+ \frac{g(-1)^{p(D+1)}}{(D-p-1)!} \epsilon^{i_1\cdots i_{p-1}j_1\cdots j_{D-p}} \partial_{j_1}B_{j_2\cdots j_{D-p}}\right],\end{aligned}$$ are the secondary first class constraints associated to the Lagrange multipliers $B_0,C_0,\omega_0$ y $A_0$, respectively. The non-local and gauge-invariant “Wilson operators” of this theory are $$\begin{aligned} W(\Sigma_p)&=&\exp{\left(i\int_{\Sigma_p}(A+d\omega)\right)},\\ \Omega(\Sigma_{D-p-1})&=&\exp{\left(i\int_{\Sigma_{D-p-1}}\!\!\!\!\!\!\!\!(B+dC)\right)}.\end{aligned}$$ They obey the Dual Algebra $$W(\Sigma_p)\Omega(\Sigma_{D-p-1})=\exp{\left(-iN(\Sigma_p,\Sigma_{D-p-1})\right)}\Omega(\Sigma_{D-p-1})W(\Sigma_p),$$ where $N(\Sigma_p,\Sigma_{D-p-1})$ is the oriented number of intersection between the hypersurfaces $\Sigma_p$ and $\Sigma_{D-p-1}$. This model admits two dual geometric representations. In one of them, $W(\Sigma_p)$ appends a $p-$surface $\Sigma_p$ to the argument of the surface-dependent functional $\Psi({\Sigma_p '})$ on which it acts, while $\Omega(\Sigma_{D-p-1})$ counts ($i$ times the exponential of) how many times $\Sigma_p '$ and $\Sigma_{D-p-1}$ intersect each other. In the “dual” geometric representation, on the other hand, these roles are interchanged: $\Omega$ appends $ \Sigma_{D-p-1}$ surfaces while $W$ counts intersection numbers. This result should be compared with the $2+1$ case discussed in section \[sec3\]. [**Acknowlegment**]{}\ This work is supported by Project G-2001000712 of FONACIT. [25]{} P. K. Townsend, in [*C81-02-18.8*]{} CERN-TH-3067 [*Lecture given at 18th Winter School of Theoretical Physics, Karpacz, Poland, Feb 18 - Mar 3, 1981*]{}. S. E. Hjelmeland and U. Lindstrom, hep-th/9705122. L. Leal, Mod. Phys. Lett. A [**11**]{}, 1107 (1996) \[hep-th/9603006\]. H. A. Kramers and G. H. Wannier, Phys. Rev.  [**60**]{}, 252 (1941). G. ’t Hooft, Nucl. Phys. B [**153**]{}, 141 (1979). M. B. Halpern, Phys. Rev. D [**19**]{}, 517 (1979). E. C. Marino and J. E. Stephany, Int. J. Mod. Phys. A [**7**]{}, 171 (1992). P. K. Townsend, K. Pilch and P. van Nieuwenhuizen, Phys. Lett.  [**136B**]{}, 38 (1984) \[ [**137B**]{}, 443 (1984)\]. S. Deser, R. Jackiw and S. Templeton, Ann. Phys.  [**140**]{}, 372 (1982). S. Deser and R. Jackiw, Phys. Lett. B [**139**]{}, 371 (1984). E. C. Stüeckelberg, Helv. Phys. Acta [**15**]{}, 23 (1942). P.A.M. Dirac, [*Lectures on Quantum Mechanics*]{}, Belfer Graduate School, Yeshiva University, New York,1964. L. Faddeev and R. Jackiw, Phys. Rev. Lett.  [**60**]{}, 1692 (1988). R. Gambini and A. Trias, Phys. Rev. D [**22**]{}, 1380 (1980); D [**23**]{}, 553 (1981); D [**27**]{}, 2935 (1983); X. Fustero, R. Gambini and A. Trias, Phys. Rev. D [**31**]{}, 3144 (1985). C. di Bartolo, F.Nori, R. Gambini and A. Trias, Lett. Nuovo Cim.  [**38**]{}, 497 (1983). J. Camacaro, R. Gaitan and L. Leal, Mod. Phys. Lett. A [**12**]{}, 3081 (1997) \[hep-th/9606121\]. R. Gaitan and L. Leal, Int. J. Mod. Phys. A [**11**]{}, 1413 (1996). L. Leal and O. Zapata, Phys. Rev. D [**63**]{}, 065010 (2001) \[hep-th/0008049\]. P. J. Arias and A. Restuccia, Phys. Lett. B [**347**]{}, 241 (1995) \[hep-th/9410134\]. J. Stephany, Phys. Lett. B [**390**]{}, 128 (1997) \[hep-th/9605074\]. [^1]: [email protected] [^2]: [email protected] [^3]: [email protected]

ArXiv

--- abstract: 'We construct firstly the complete list of five quantum deformations of $D=4$ complex homogeneous orthogonal Lie algebra $\mathfrak{o}(4;\mathbb{C})\cong \mathfrak{o}(3;\mathbb{C})\oplus \mathfrak{o}(3;\mathbb{C})$, describing quantum rotational symmetry of four-dimensional complex space-time, in particular we provide the corresponding universal quantum $R$-matrices. Further applying four possible reality conditions we obtain all sixteen Hopf-algebraic quantum deformations for the real forms of $\mathfrak{o}(4;\mathbb{C})$: Euclidean $\mathfrak{o}(4)$, Lorentz $\mathfrak{o}(3,1)$, Kleinian $\mathfrak{o}(2,2)$ and quaternionic $\mathfrak{o}^{\star}(4)$. For $\mathfrak{o}(3,1)$ we only recall well-known results obtained previously by the authors, but for other real Lie algebras (Euclidean, Kleinian, quaternionic) as well as for the complex Lie algebra $\mathfrak{o}(4;\mathbb{C})$ we present new results.' author: - | A. Borowiec$^{1}$, J. Lukierski$^{1}$ and V.N. Tolstoy$^{2}$\ \ $^{1}$Institute for Theoretical Physics,\ University of Wroc[ł]{}aw, pl. Maxa Borna 9,\ 50–205 Wroc[ł]{}aw, Poland\ \ $^{2}$Lomonosov Moscow State University,\ Skobeltsyn Institute of Nuclear Physics,\ Moscow 119991, Russian Federation title: 'Basic quantizations of $D=4$ Euclidean, Lorentz, Kleinian and quaternionic $\mathfrak{o}^{\star}(4)$ symmetries ' --- Introduction ============ In recent years due to the efforts to construct models of quantum gravity characterized by non-commutative spacetime structures at Planckian distances [@Maj88]–[@Gar95], the ways in which one can deform the algebras of spacetime coordinates and space-time symmetries became important. In noncommutative description of spacetime the numerical coordinates are replaced by noncommutative algebra, which is consistent with new type of uncertainty relations between the pairs of operator-valued coordinates [@DFR95] called further DFR uncertainty relation relation. This extension of noncommutativity into the spacetime sector describes the limitations on spcetime localization measurements if the quantum gravitational background is present. It appears that during such procedure the high density of energy added by measurement leads to the creation of mini black holes, and one can show that below Planck distance $\lambda_P=10^{-33}m$ operationally the classical spacetime is not longer applicable. The quantum spacetime is effectively atomized, with lattice structure, and following the derivation in QM of Heisenberg algebra from Heisenberg uncertainty relations, one can deduce from DFR uncertainty relation the noncommutativity of quantum spacetime. Such noncommutative and/or discrete nature of quantum spacetime follows as well from loop quantum gravity (LQG) approach [@AshLew; @Diana; @Thi], where the discretization is dynamical[^1], based on the existence in LQG framework of minimal lengths, minimal area surfaces or minimal volume quanta. In particular recently by applying LQG techniques to the quantum deformation of $D=3$ gravity with positive cosmological constant $\Lambda$ it has been shown [@JKG] that one gets the quantum symmetry of $U_q(\mathfrak{o}(4))$, where $\ln q\sim\Lambda$, in analogy with earlier results of [@LuRuNoTo] and [@Smolin] for Lorentz signature. In this paper we shall consider the noncommutative structures as linked with the quantum groups, which are described as non-cocommutative Hopf algebras [@Dr2; @EtKazh; @Wor87; @ChPr94; @Maj95; @Klimyk; @EtSch02]. The deformed spacetime algebra is described as the irreducible representation (noncommutative Hopf algebra module) of quantum rotations algebra, with semidirect (smash) product structure and build-in covariant action of quantum-deformed symmetry algebra on quantum noncommutative spacetime. The aim of this paper is to provide the quantum Hopf algebras and universal $R$-matrices which are obtained by quantization of classical $r-$matrices classified recently in [@BLTnov15; @LT17; @BLT17]. In this paper we shall describe the Hopf-algebraic deformations of any real four-dimensional rotational algebra given by the real form of $\mathfrak{o}(4;\mathbb{C})$. Fortunately, all classical $r-$matrices presented in [@BLTnov15; @BLT17] can be quantized by providing explicit formulae for coproducts, antipodes and universal $R$-matrices. [^2] We see therefore that the present paper provides the completion of the research program which we started in ref. [@BLTnov15; @LT17; @BLT17]. The plan of our paper is the following: In Sect.2 we shall present some generalities on quantization of ’infinitesimal’ versions of quantum deformations described by complex and real classical $r$-matrices, which provides the triangular and nontriangular cases and further we present reality conditions for the universal $R$-matrices. Further, in Sect. 3, we shall illustrate the quantization of classical $r$-matrices by the explicit presentation, for $\mathfrak{sl}(2;\mathbb{C})$ case, of Jordanian and standard deformations. In Sect. 4 we shall recall $D=4$ complex Lie algebra $\mathfrak{o}(4;\mathbb{C})$ and its all four real forms. In particular, besides three real forms $\mathfrak{o}(4-k,k)$ ($k=0,1,2$) differing by the choice of signature, it should be added fourth quaternionic real form $\mathfrak{o}(2;\mathbb{H})\equiv\mathfrak{o}(2,1)\oplus \mathfrak{o}(3)\equiv\mathfrak{o}^*(4)$. We shall work mostly with the generators of four-dimensional complex rotations in Cartan-Weyl bases. In Sect. 5 we quantize the full list of five classical $r$-matrices from [@BLTnov15; @BLT17], i.e. provide the complete list of all Hopf-algebraic deformations of $U_q(\mathfrak{o}(4;\mathbb{C}))$: two of them triangular, and remaining three quasitriangular.[^3] In order to present the results in detail we shall calculate the coproducts, antipods and universal quantum $R$-matrices. Further we specify all real forms of the quantum deformations of $\mathfrak{o}(4;\mathbb{C})$ described by $\star$-Hopf algebras. Following the standard recipe (see e.g. [@KhTo1; @LNR91; @Wor91]) we assume that the $\star$-operation, defining respective real form, acts on tensor product (coproduct) in unflipped way $(a\otimes b)^\star= a^\star\otimes b^\star$. We add that all quantum deformations of Lorentz algebra were already obtained earlier by the present authors [@BLT06; @BLT08] and five out of eight Kleinian $D=4$ real deformations can be obtained from the complex $\mathfrak{o}(4;\mathbb{C})$ deformations listed in Sect.4 simply by replacing the complex $\mathfrak{sl}(2;\mathbb{C})$ generators by the real ones describing $\mathfrak{sl}(2;\mathbb{R})$ algebra. The Hopf-algebraic deformation of Euclidean $\mathfrak{o}(4)$ algebra some $\mathfrak{o}(2,2)$ deformations and quaternionic $\mathfrak{o}(2;\mathbb{H})$ case are the most important because the obtained results are new. In Sect. 6 we shall present a brief outlook; the paper contains also two appendices. The quantum deformations of four-dimensional rotational symmetries presented in this paper can be applied at least in the following contexts: i\) The deformed $D=4$ rotation groups with the signature $(+,+,-,-)$ (Kleinian case) describe the deformed $D=3$ AdS symmetry and for Lorentz signature $(+,-,-,-)$ the $D=3$ dS quantum symmetries. If we introduce (A)dS radius $\Lambda$ and the re-scaling of three rotations $M_{1k}\rightarrow \tilde M_{1k}=\Lambda P_k$ ($P_k$ describes curved (A)dS momenta, $k=1,2,3$), by suitable quantum Wigner-Inönü contraction [@CeGiSoTa; @LuRuNoTo] one can get various $\kappa$-deformed $D=3$ Poincaré algebras. ii\) The knowledge of classical $r-$matrices permits to introduce explicitly the action of deformed (super)string models, described by so-called YB (Yang-Baxter) sigma models [@Klim02]–[@BKLSY15]. The quantum deformations presented in this paper can be applied to the description of the YB deformation of principal $\mathfrak{o}(4-k,k)$ $\sigma-$models ($k=0,1,2$) as well as to the coset sigma models with noncommutative target space, described by the deformed cosets $\frac{\mathfrak{o}(4-k,k)}{\mathfrak{o}(3-k,k)}$. It appears that such deformed $\mathfrak{o}(4-k,k)$ group or their coset manifolds can appear as parts of internal symmetry target spaces obtained by the reduction to $D=4$ of deformed $D=10$ Green–Schwarz superstrings. iii\) The classical $r$-matrices and their quantizations provide a powerful algebraic tool in description of integrable models and provide effective methods for studying their multihamiltonian systems [@VDr; @MBl; @StS]. In particular, the methods of noncommutative geometry permits to consider as well the Hamiltonian theories over the noncommutative rings [@Dorfman; @Fokas] and their integrability conditions. iv\) Eight quantum deformations of $\mathfrak{o}(2,2)$ presented in the paper provide the set of finite $D=2$ quantum conformal algebras, with six generators, which in general case cannot be factorized into a sum of “left” and “right” ($X_\pm=X_1\pm X_0$) $D=1$ quantum deformed conformal algebras. It is interesting to study which $\mathfrak{o}(2,2)$ deformations presented in the paper can be consistently extended to infinite-dimensional quantum groups, describing new classes of deformed $D=2$ infinite-dimensional conformal Virasoro algebras. v\) For various real forms of quantum-deformed $\mathfrak{o}(4;\mathbb{C})$ groups one can obtain corresponding four-dimensional spacetime with different signatures (see e.g. [@Rita17]). With all Hopf-algebraic deformations of $\mathfrak{o}(4;\mathbb{C})$ which will be presented in this paper one can obtain the complete list of quantum spacetimes with signatures $(4,0)$, $(3,1)$ and $(2,2)$. Further remarks related with the applications of quantum deformations considered in this paper we shall present also in Sect. 6. Quantizations of complex and real Lie algebras: general remarks =============================================================== From classical $r$-matrices to quantum universal $R$-matrices ------------------------------------------------------------- It is known that formulated by Drinfeld [@Dr2] the quantization problem of Lie bialgebras has been answered by Etingof and Kazhdan [@EtKazh]: to each Lie bialgebra one can associate a quantized enveloping algebra supplemented with Hopf algebra structure. Unfortunately, their proof is not constructive and the methods of explicit quantizations are known only in specific situations, as e.g. Drinfeld-Jimbo quantization of semi-simple Lie algebras and twist quantization in the triangular case (when twist tensor can be constructed explicitly). We shall show however that the known quantization techniques are sufficient for finding all explicit non-isomorphic quantizations of the enveloping algebra $\mathfrak{o}(4;\mathbb{C})$ and their real form. Principal tool for the classification of quantum deformations is provided by the classical $r$-matrices [@BD82; @BD84; @Sem-T-Sh83] which determine coboundary Lie bialgebra [^4] structures. Quantization procedure od bialgebras leads to the construction of quantum-deformed associative and coassociative Hopf-algebras [@Dr1] and determine the corresponding universal (quantum) $R-$matrices [@ChPr94; @Maj95; @KhTo1]. For semi-simple Lie algebras, due to the classical Whithead lemma, all bialgebras are coboundary. In such a case there is one-to-one correspondence between the Lie bialgebra structure and the corresponding classical $r$-matrix given as the skewsymmetric element $r\in\mathfrak{g}\wedge\mathfrak{g}$ satisfying the classical (homogenous or inhomogenous) YB equation: $$\begin{aligned} \label{crm1} [[r,\,r]]\!\!&=\!\!& t\,\Omega , \quad t\,\in\mathbb{C}~.\end{aligned}$$ with $[[\cdot,\cdot]]$ denoting Schouten bracket $$\label{i4} [[r,r]]\equiv [r_{12},r_{13}+r_{23}]+[r_{13},r_{23}].$$ where $r_{12}=r^{(1)}\otimes r^{(2)}\otimes 1\in \mathfrak{g}\otimes\mathfrak{g}\otimes\mathfrak{g}$ etc. For skew-symmetric 2-tensor monomials $x\wedge y=x\otimes y-y\otimes x$ and $u\wedge v$ ($x,y,u,v\in\mathfrak{g}$) the explicit formula for Schouten brackets reads[^5] $$\begin{aligned} \label{schouten} \begin{array}{rcl} [[x\wedge y,\,u\wedge v]]\!\!&:=\!\!&x\wedge\bigl([y,u]\wedge{v}+u\wedge[y,v]\bigr) \\[4pt] &&-y\wedge\bigl([x,u]\wedge{v}+u\wedge[x,v]\bigr) \\[5pt] \!\!&\phantom{:}=\!\!&[[u\wedge v,\,x\wedge y]] \end{array}\end{aligned}$$ where the three-form $\Omega$ is the $\mathfrak{g}$-invariant element in $\mathfrak{g}\wedge\mathfrak{g}\wedge\mathfrak{g}$, i.e. $$\label{i8a} ad_x\Omega\equiv[x\otimes 1\otimes 1+1\otimes x\otimes 1+1\otimes 1\otimes x, \Omega]=0 ,\quad x\in\mathfrak{g}.$$ The complex Lie bialgebra is described by a pair $\mathsf{g}\equiv(\mathfrak{g}, r)$ consisting of complex Lie algebra $\mathfrak{g}$ and skew-symmetric classical $r$-matrix $r$ satisfying the equation (\[crm1\]). One can distinguish two cases (cf. e.g. [@ChPr94; @Maj95; @EtSch02]): A\) If in (\[crm1\]) $t= 0$, one gets the so-called triangular or non-standard case with vanishing Schouten brackets describing homogenous classical Yang-Baxter equation (denoted as CYBE). In such a case the $r$-matrix can be rescaled arbitrarily without changing the corresponding bialgebra structure. The triangularity is preserved by Lie algebra homomorphisms and can be reduced to non-degenerate case on the Borel subalgebra. B\) If $t\neq 0$, eq. (\[crm1\]) describes so-called non-triangular (quasitriangular) classical $r$-matrix, satisfying inhomogenous or modified classical Yang-Baxter equation (mCYBE). In such case one can introduce $ r_{BD}\in \mathfrak{g}\otimes \mathfrak{g}$ called Belavin-Drinfeld form of the $r$-matrix satisfying CYBE, such that $r= r_{BD}- r_{BD}^\tau$ ($(x\otimes y)^\tau=y\otimes x$ is the flip operation) and the symmetric element $r_{BD}^s\equiv r_{BD} +r_{BD}^\tau$ which is ad-invariant. [^6] In general, the initial skew-symmetric $r$-matrix is not scale invariant nor preserved by a Lie algebra homomorphisms. It is remarkable that Belavin-Drinfeld $r$-matrices for simple Lie algebras has been fully classified by means of so-called Belavin-Drinfel triples in [@BD84]. Quantization of (complex) Lie bialgebra leads to quantum groups in Drinfeld sense [@Dr2] with the Hopf algebra structure supplementing the complex deformed universal enveloping algebra $U(\mathfrak{g})$ (in general one needs its topological $\xi$-adic extension $U_\xi(\mathfrak{g}) \equiv U(\mathfrak{g})[[\xi]]$ formulated also for multiparameter deformation, i.e. $\xi\rightarrow (\xi_1,\ldots,\xi_k)$, see e.g. [@Dr2; @ChPr94; @Maj95; @Klimyk]). According to the cases A), B) indicated above, there are two ways of introducing quantum-enveloping algebra what will be described shortly below. Before we would like to focus our attention on the quantum universal $R$-matrix as an important byproduct of the quantization procedure. [^7] The universal $R$-matrix is an invertible element of $U_\xi(\mathfrak{g})\otimes U_\xi(\mathfrak{g})$ which provides the flip $\tau$ of the noncocommutative coproduct $\tau:\Delta_\xi\rightarrow \Delta^\tau_\xi$ given by the following similarity transformation $$\label{i2} \Delta^\tau_\xi(\cdot)= R \,\Delta_\xi(\cdot) \,R^{-1}$$ The universal $R-$matrix describes quantum group (see [@Dr1]) if it satisfies quasitriangularity conditions $$\label{i3a} (\Delta^{}_\xi\otimes id)R =R_{12} R_{23} ,\qquad (id\otimes\Delta^{}_\xi)R =R_{13} R_{12}$$ where $R =R^{(1)} \otimes R^{(2)} $ and $R_{12} =R^{(1)} \otimes R^{(2)} \otimes 1$, etc.. The properties (\[i2\])–(\[i3a\]) imply in $U_\xi(\mathfrak{g})\otimes U_\xi(\mathfrak{g})\otimes U_\xi(\mathfrak{g})$ quantum Yang-Baxter equation (qYBE) in the form $$\label{i3b} R_{12} R_{13} R_{23} =R_{23} R_{13} R_{12}$$ as well as the following normalization conditions $$\label{i3c} (\epsilon\otimes id)R =(id \otimes\epsilon)R =1,$$ where $\epsilon$ denotes a counit. In fact, the same properties (\[i2\])–(\[i3c\]) are satisfied by another universal $R$-matrix, which is $(R^\tau)^{-1}$. Therefore one can distinguish two case: i\) the element $Q_R=R R^\tau=1$ is trivial ii\) $Q_R\neq 1$ is non-trivial\ It turns out that the first case corresponds to the triangular or twist quantization case while the second characterizes the non-triangular case. In order to describe their difference let us expand the $R$-matrix (\[i2\]) in the powers of the deformation parameter $\xi$, entering linearly in the definition of classical $r$-matrix [^8] $$\label{i1} R(\xi)= 1\otimes 1 + \tilde r + O(\xi^2)$$ From (\[i1\]) and (\[i3b\]) it follows that the element $\tilde r\in \mathfrak{g}\otimes\mathfrak{g}$ satisfies classical Yang–Baxter equation (CYBE) [@ChPr94; @Maj95]. In triangular case one has $\tilde r+\tilde r^\tau=0$, i.e. $\tilde r$ is skew-symmetric and can be identified with the classical $r$-matrix satisfying (\[crm1\]) with $t=0$. In the second (non-triangular, see ii)) case, $\tilde r$ is not skew-symmetric, satisfies CYBE and takes the Belavin-Drinfeld form of $r$-matrix, i.e. $\tilde r=r_{BD}$. The classical $r$-matrices describe the infinitesimal version of quantum deformed Lie-algebraic symmetries; the quantum deformation parameterized by an arbitrary (formal) deformation parameter $\xi$ determines Hopf-algebraic quantization and universal $R-$matrix . In general case it is not known how to obtain the universal $R$-matrix from the solutions of (\[i4\]); however for canonical Belavin–Drinfeld nontriangular $r-$matrices [@BD82] the explicit formula for universal $R-$matrices is well-known (see e.g. [@KhTo1]). It is worth noticing that in contrast to the triangular case, the non-triangular one provides two different quantum $R$-matrices: $R(\xi)$ and $R^\tau(\xi)^{-1}$. The element $Q(\xi)\equiv R^\tau(\xi)R(\xi)=1+(r+r^\tau) + O(\xi^2)$ is called a quantum Killing form since its first order term, if not degenerate, defines a classical Cartan-Killing form on $\mathfrak{g}$. We would like to add that the skew-symmetric classical $r$-matrices are sufficient for classification [^9] as well as for the description of correspondence with classical Lie-Poisson groups. Reality conditions providing the quantizations of real bialgebras ----------------------------------------------------------------- We remind that a real Lie algebra structure $(\mathfrak{g}, \divideontimes)$ can be introduced by adding an antilinear involutive (Lie algebra) anti-automorphism $\divideontimes : \mathfrak{g}\rightarrow\mathfrak{g}$ ($\divideontimes$-operation, conjugation) acting on the complex Lie algebra $\mathfrak{g}$. It relies on finding Lie algebra basis with real structure constants for which $\divideontimes$-operation is anti-Hermitian (i.e. $x^\divideontimes=-x$) [^10]. Subsequently, the real coboundary Lie bialgebra can be considered as a triple $(\mathsf{g},\divideontimes)\equiv(\mathfrak{g}, \divideontimes, r)$, where the skew-symmetric element $r$ is assumed to be anti-Hermitian, i.e. $$\label{rc_1} r^{\divideontimes\otimes\divideontimes}=-r =r^\tau.$$ Such conditions lead to the suitable reality conditions for parameters, what is particularly important in the inhomogenous case ($t\neq 0$), due to the lack of scale invariance. The $\divideontimes$-operation extends, by the property $(ab)^\divideontimes=b^\divideontimes a^\divideontimes$ (i.e. as an antilinear antiautomorphism), to the enveloping algebra $U_{}(\mathfrak{g})$, as well as to quantized enveloping algebra, making both of them associative $\divideontimes$-algebras. The real Hopf-algebraic structure represented on quantized enveloping algebra $U_{q_{}}(\mathfrak{g})$ by $\divideontimes$-involution is defined by the following conditions for coproducts and antipodes (see also [@LNR91; @Wor91]) $$\label{d-1} \Delta_{q_{}}(a^\divideontimes)\;=\;(\Delta_{q_{}}(a))^\divideontimes,\quad\;\; S_{q_{}}((S_{q_{}} (a^\divideontimes))^{\divideontimes})\;=\;a,\quad\;\epsilon(a^\divideontimes)=\epsilon(a)^*\quad(\forall a\in U_{q_{}}(\mathfrak{g})~.$$ where the $\divideontimes$-involution on the tensor product (\[d-1\]) acts as follows [^11] $$\label{d-1b} (a\otimes b)^\divideontimes = a^\divideontimes\otimes b^\divideontimes$$ One can get (\[i2\]) and (\[d-1\]) compatible and consistently defined quasitriangular $\divideontimes$-Hopf algebras by imposing two distinct reality constraints on the universal $R$-matrix (see e.g. [@Maj95]): a\) $R^{\divideontimes\otimes\divideontimes}=R^\tau$ ($R$ is called real); b\) $R^{\divideontimes\otimes\divideontimes}=R^{-1}$ and the corresponding quantum $R$-matrix is $\divideontimes$-unitary ($R$ is called antireal). Particularly, in the triangular case, due to the identity $R^\tau=R^{-1}$, the conditions a) and b) are the same. In non-triangular case ($R^\tau\neq R^{-1}$), the second universal R-matrix $(R^\tau)^{-1}$ satisfies the same reality constraints [^12]. It should be noted that for any element $\tilde r\in\mathfrak{g}\otimes\mathfrak{g}$ satisfying CYBE, $\tilde r^{\divideontimes\otimes\divideontimes} $ satisfies again CYBE. Therefore, one can distinguish two cases: i\) the classical $r$-matrix $r=\tilde r$ corresponding to the universal $R$-matrix (cf. \[i1\]) is skew-symmetric; then $r$ should be anti-Hermitian and satisfy the relation (\[rc\_1\]). ii) if the element $\tilde r$ is not skew-symmetric this corresponds to the non-triangular case; then $\tilde r^{\divideontimes\otimes\divideontimes}=\tilde r^\tau$ [^13] for $R$ real and $\tilde r^{\divideontimes\otimes\divideontimes}=-\tilde r$ for $R$ antireal. It is easy to check that in any case the skew-symmetric part of $\tilde r$ remains anti-Hermitian, i.e. $(\tilde r-\tilde r^\tau)^{\divideontimes\otimes\divideontimes}=-(\tilde r-\tilde r^\tau)$.\ It is easy to show that twisting of real form of quasitriangular Hopf algebra by unitary twist leads again to real quasitriangular Hopf algebra. More precisely, if $(\mathcal{H}, \Delta, S, \epsilon, R, \star)$ is a quasitriangular $\star$-Hopf algebra with $R$ being real (resp. antireal) universal $R$-matrix, then for any unitary, normalized 2-cocycle twist $F=(F^{-1})^{\star\otimes\star}\in \mathcal{H}\otimes\mathcal{H}$ the quantized algebra $(\mathcal{H}, \Delta_F, S_F, \epsilon, R_F, \star)$ is a quasitriangular $\star$-Hopf algebra such that $R_F=F^\tau R F^{-1}$ is real (resp. antireal). This property will be used in the consideration of some cases of chain quantization of $\mathfrak{o}(4;\mathbb{C})$ (see Sect. 5). The basic $\mathfrak{sl}(2;\mathbb{C})$ example: complex and real Lie bialgebra and their quantizations ======================================================================================================= Complex and real bialgebras ---------------------------- We recall that classical $r$-matrices, providing Lie bialgebra structure of a given Lie algebra as well as quantum deformations of the corresponding enveloping algebra, are classified up to the isomorphisms; in particular for real Lie algebras one should use the isomorphisms preserving reality condition. [^14] Fixing the basis (structure constant) one deals with Lie algebra automorphisms. For simple Lie algebras these are (modulo discrete automorphisms) the internal automorphisms generated by the adjoint actions of the Lie algebra upon itself. The important example of such classification for real forms of $\mathfrak{o}(4;\mathbb{C})$ has been investigated in [@BLTnov15; @LT17; @BLT17]. It is well known that for the complex Lie algebra $\mathfrak{sl}(2;\mathbb{C})\cong\mathfrak{o}(3;\mathbb{C})$ there exists up to $\mathfrak{sl}(2,\mathbb{C})$ automorphisms two solutions of mCYBE [^15], namely Jordanian $r_{J}^{}$ (triangular, called also non-standard) and the standard one $r_{st}^{}$ (non-triangular): $$\begin{aligned} \label{rm1a} r_{J}^{}\!\!&=\!\!&\xi\,E_{+}\wedge H,\quad[[r_{J}^{},r_{J}^{}]]\;=\;0, \\[3pt]\label{rm2} r_{st}^{}(\gamma)\!\!&=\!\!&\gamma E_{+}^{}\wedge E_{-}^{},\quad[[r_{st}^{},r_{st}^{}]]\;=\;{\gamma}^{2}\Omega,\end{aligned}$$ where we use the Cartan–Weyl (CW) basis $$\label{rm3} [H,E_{\pm}]\,=\,\pm E_{\pm},\qquad [E_{+},E_{-}]\,=\,2H .$$ In (\[rm1a\]) the parmeter $\xi$ can be replaced by $\xi=1$ due to the scale invariance of CYBE. In the standard case (\[rm2\]) the non skew-symmetric counterpart of $r_{st}$ satisfies CYBE if it takes the Belavin-Drinfeld form $$\begin{aligned} \label{rm5} r_{BD}^{}(\gamma)\!\!&=\!\!&\gamma\bigl(E_{+}\otimes E_{-}+H\otimes H\bigr) $$ Its symmetric part, described by $\mathfrak{sl}(2;\mathbb{C})$ bilinear split Casimir $$E_{+}\otimes E_{-}+ E_-\otimes E_+ +2H\otimes H$$ is an invariant element in $\mathfrak{sl}(2,\mathbb{C})\otimes \mathfrak{sl}(2,\mathbb{C})$ and determines the Cartan-Killing form. We recall that the (complex) simple Lie algebra $\mathfrak{o}(3; \mathbb{C})\cong\mathfrak{sl}(2; \mathbb{C})$ has up to $\mathfrak{sl}(2; \mathbb{C})$-isomorphisms two real forms: compact $\mathfrak{o}(3)\cong\mathfrak{su}(2)$ and noncompact $\mathfrak{o}(2,1)\cong\mathfrak{su}(1,1)\cong \mathfrak{sl}(2)$. It is known, see e.g. [@LT17], that with these two real forms there are linked four real Lie bialgebras, one compact and three noncompact ones, which can be expressed in $\mathfrak{su}(1,1)\cong \mathfrak{sl}(2)$ or $\mathfrak{o}(2,1)$ bases (see also Appendix A). The unique compact real bialgebra one can write in $\mathfrak{su}(2)$ basis (cf. [@Maj95; @Klimyk]) [^16], [^17] satisfying reality conditions ($\divideontimes=\dag$) $$\begin{aligned} \label{rm6-1} H^{\dagger}\!\!&=\!\!&H,\qquad E_{\pm}^{\dagger}\;=\;E_{\mp},\quad r_{st}(\gamma)\, ,\,\gamma\in \mathbb{R} \quad \;\;(\mathsf{su}_\gamma(2)\;\; {\rm{standard\, bialgebra}})\end{aligned}$$ From three noncompact inequivalent real bialgebras we choose to write one in $\mathfrak{su}(1,1)$ basis ($\divideontimes=\#$) $$\begin{aligned} \label{rm6-2} H^{\#}\!\!&=\!\!&H,\quad\;E_{\pm}^{\#}\;=\;-E_{\mp},\quad\; r_{st}(\gamma)\, ,\,\gamma\in \mathbb{R} \quad\;\;(\mathsf{su}_\gamma(1,1) \;\; {\rm{standard\, bialgebra}})\end{aligned}$$ and remaining two in $\mathfrak{sl}(2)$ basis ($\divideontimes=\star$) $$\begin{aligned} \label{rm6-3} H^{\star}\!\!&=\!\!&-H,\quad\;E_{\pm}^{\star}\;=\;-E_{\pm},\quad\;\; r_{st}(\gamma)\, ,\,\gamma\in \imath\mathbb{R} \quad\;\;(\mathsf{sl}_\gamma(2)\;\; {\rm{standard\, bialgebra}})\end{aligned}$$ $$\begin{aligned} \label{rm6-4} H^{\star}\!\!&=\!\!&-H,\quad\;E_{\pm}^{\star}\;=\;-E_{\pm},\quad\;\; r_{J}\quad \qquad\quad (\mathsf{sl}_J(2)\;\qquad {\rm{nonstandard\,\, bialgebra}}) \end{aligned}$$ First three $r$-matrices (\[rm6-1\])-(\[rm6-3\]) are standard (non-triangular) while the last (\[rm6-4\]) is Jordanian (triangular) without multiplicative parameter because it has been rescaled to $1$ by suitable $\mathfrak{sl}(2)$-automorphism. The first and second bialgebra depends on real parameter, and the third one is multiplied by purely imaginary parameter (it is antireal). We stress that however $\mathsf{su}(1,1)$ and $\mathsf{sl}(2)$ are isomorphic with real Lie algebra $\mathfrak{o}(2,1)$, they are not isomorphic as real Lie bialgebras (cf. [@LT17]). In formulae (\[rm6-1\])-(\[rm6-3\]) there are used for the complex CW basis (\[rm3\]) three different reality conditions defining $\mathfrak{su}(2)$, $\mathfrak{su}(1,1)$ and $\mathfrak{sl}(2)$ real algebras. Because $\mathfrak{o}(2,1)\cong\mathfrak{su}(1,1)\cong \mathfrak{sl}(2)$, the involutions $\#$ and $\star$ (see (\[rm6-2\])-(\[rm6-3\])) can be identified and related with the reality condition defining $\mathsf{o}(2,1)$ as real form of $\mathsf{o}(3;\mathbb{C})$. Indeed, the $\mathfrak{su}(1,1)$ real basis $(H', E_\pm')$ and $\mathfrak{sl}(2)$ real bases $(H, E_\pm)$ can be related by the following linear complex $\mathsf{sl}(2;\mathbb{C})\cong\mathfrak{o}(3; \mathbb{C})$ automorphism $$\begin{aligned} \label{jl1} \begin{array}{rcl} H' = \displaystyle-\frac{\imath}{2}\big(E_{+}-E_{-}\big) ,\quad E_{\pm}'= \displaystyle\mp\imath H+\frac{1}{2}\big(E_{+}+E_{-}\big). \end{array}\end{aligned}$$ One can use the complex Cartesian basis $I_k\in \mathfrak{o}(3; \mathbb{C})$ ($k=1,2,3$) $$\begin{aligned} \label{jl2} \begin{array}{rcl} [I_{i},\,I_{j}]\!\!&=\!\!&\varepsilon_{ijk}I_{k} \end{array}$$ which is antireal for the real compact form $\mathfrak{o}(3)\cong\mathfrak{su}(2)$ $$\begin{aligned} \label{jl3b} I_{i}^{\dag}\!\!&=\!\!&-\,I_{i}^{}\quad(i=1,2,3)\quad{\rm{for}}\;\;\mathfrak{o}(3).\end{aligned}$$ For both cases $\mathfrak{su}(1,1)$ and $\mathfrak{sl}(2)$ the reality condition in Cartesian basis takes the same $\mathfrak{o}(2,1)$ form $$\begin{aligned} \label{jl3} I_{i}^{\star}\!\!&=\!\!&(-1)^{i-1}I_{i}^{}\quad(i=1,2,3)\quad{\rm{for}}\;\;\mathfrak{o}(2,1).\end{aligned}$$ One can relate the $\mathfrak{su}(1,1)$ and $\mathfrak{sl}(2)$ bases with the Cartesian $\mathsf{o}(2,1)$ generators satisfying the same reality condition (\[jl3\]) by the following formulae $$\begin{aligned} &&\begin{array}{rcl}\label{jl4} &&H'\,:=\,\imath I_{2},\qquad\qquad E_{\pm}'^{}\,:=\,\imath I_{1}\pm I_{3}, \qquad{\rm{for}}\;\;\mathfrak{su}(1,1),\\[4pt] &&H\,:=\,\imath I_{3},\qquad\qquad E_{\pm}\,:=\,\imath I_{1}\mp I_{2}, \qquad{\rm{for}}\;\;\mathfrak{sl}(2,\mathbb{R}).\end{array}$$ Both CW bases $\{E_{\pm}',H'\}$ and $\{E_{\pm},H\}$ have different reality properties which follow from the same reality condition (\[jl3\]) $$\begin{aligned} \begin{array}{rcl}\label{jl5} &&{H'}^{\star}\;=\;H',\qquad\;\; {E_{\pm}'}^{\star}=-E_{\mp}'^{} \qquad{\rm{for}}\;\;\mathfrak{su}(1,1), \\[3pt] &&{H}^{\star}\;=\;-H,\qquad{E_{\pm}}^{\star}=-E_{\pm} \qquad{\rm{for}}\;\;\mathfrak{sl}(2;\mathbb{R}),\end{array}\end{aligned}$$ It should be noted that in the case of $\mathfrak{su}(1,1)$ the Cartan generator $H'$ is compact while for the case $\mathfrak{sl}(2)$ the generator $H$ is noncompact, what also explains the difference between $\mathfrak{su}(1,1)$ and $\mathfrak{sl}(2)$ CW basis. In this way the involutions (\[jl3\]) and (\[rm6-2\]) – (\[rm6-3\]) are identified (it can be checked that the relations (\[jl1\]) and (\[jl4\]) are consistent). Concluding, it is sufficient for $\mathfrak{sl}(2; \mathbb{C})$ to introduce only two involutions: defining $\mathfrak{o}(3)\cong\mathfrak{su}(2)$ and $\mathfrak{o}(2,1)\cong\mathfrak{su}(1,1)\cong \mathfrak{sl}(2)$. In fact the formulae (\[jl4\]) can be used for the introduction of Cartesian basis in all classical $\mathfrak{o}(2, 2)$ and $\mathfrak{o}^*(4)$ $r$-matrices containing the $\mathfrak{su}(1,1)$ and $\mathfrak{sl}(2)$ sectors. In the next subsection we shall describe explicitly the quantization of the complex bialgebras (\[rm1a\]) and (\[rm2\]). In order to obtain the quantization of the bialgebras listed in (\[rm6-1\]) - (\[rm6-4\]) one should insert the generators $(H, E_\pm)$ satisfying the respective reality condition and impose the suitable restriction on the parameter $\gamma$. Standard deformation of simple Lie algebra is given by the explicit algorithm introduced firstly by Drinfeld and Jimbo. Non-standard quantum deformation of $\mathsf{g}\equiv(\mathfrak{g}, r)$, where $r$ a skew symmetric solution of CYBE, is obtained by employing the 2-cocycle Drinfeld twist element $ F \in U(\mathfrak{g}) \otimes U(\mathfrak{g})$ which remains unchanged the algebra and modifies the coproduct $\Delta$ and antipode $S$ as follows (see e.g. [@ChPr94; @Maj95]): $$\label{blte7} \Delta \longrightarrow \Delta_{F} = F \, \Delta^{} \, F^{-1} \, ,\qquad S^{} \longrightarrow S_{F} = u \, S^{} \, u^{-1} \, ,$$ where $$\begin{aligned} \label{blte8} & F = \sum\limits_{i} f^{(1)}_{i} \otimes f^{(2)}_{i} \, , \qquad u = \sum\limits_{i} f^{(1)}_{i}\, S\, ( f^{(2)}_{i})\, .\end{aligned}$$ If classical enveloping Lie algebra $U(\mathfrak{g})$ is considered as a Hopf algebra $H^{(0)} = (U (\mathfrak{g}), m, \Delta^{(0)}, S^{(0)}, \epsilon)$ then $$\begin{aligned} &\Delta^{(0)} (x) = x \otimes 1 + 1 \otimes x,\quad S^{(0)}(x)=-x, \ \quad \forall x\in \mathfrak{g}\,.\nonumber $$ In order to get the coassociative coproduct one should postulate the normalized 2-cocycle condition for the invertible twist element ${\mathcal F}$ (see [@Dr2]) $$\label{coc} F^{12}(\Delta \otimes id)\left( F\right) = F^{23}(id\otimes \Delta )\left( F\right), \quad \left( \epsilon \otimes id\right)( F)=1=(id\otimes \epsilon) (F) .$$In $H^{(0)}$ one can introduce the universal (quantum) $R$-matrix by the formula $$\label{Rtwist} R_F= F^\tau\, F^{-1}\,=\, (R_F^\tau)^{-1} \sim 1\otimes 1+r+o(\chi^2)$$ which under the reality conditions (\[rc\_1\]) becomes unitary (at the same time real and antireal). More generally, twist deformation of quasitriangular Hopf algebra $(H,R)$ give rise to quasitriangular Hopf algebra with new universal $R$-matrix $R\longrightarrow R_F=F^\tau\,R\, F^{-1}$. The simplest case one can deal is an Abelian twist $$\label{Atwist} F_{A,1/2}=\exp{(-{\chi\over 2}X\wedge Y)}, \qquad R_A=\exp{(\chi X\wedge Y)}$$ where two primitive commuting elements $[X,Y]=0$ determines $r_A=\chi X\wedge Y$ the skew-symmetric solutions of CYBE. In fact, the same Abelian quantum $R$-matrix $R_A$ can be implemented by the one-parameter family of Abelian twists $$\label{Ab} F_{A,s}=\exp{\xi(s\,X\otimes Y-(1-s)\,Y\otimes X)},\quad s\in [0,1]$$ which are related with each other by a trivial (coboundary) twists . For example $$\label{Atwist2} F_{A,1}\equiv \exp{\chi\,X\otimes Y}= (W^{-1}\otimes W^{-1})\,F_{A,1/2}\, \Delta^{}(W), \qquad$$ where $W=\exp({\xi\over 2}XY)$. [^18] Assuming $X,Y$ real (antireal), i.e. $X^\divideontimes=\pm X, Y^\divideontimes=\pm Y$, the formal parameter $\chi$ has to be imaginary and all twist $F_{A,s}(\chi)$ are unitary. Consequently, anyone can be used to deform equivalently $\divideontimes$-Hopf algebras. However, there is an advantage of using (\[Atwist\]). In this case the element $u$ (see (\[blte8\])) reduces to the unit and the antipodes map remains unchanged. If $X^\divideontimes=\pm Y, Y^\divideontimes=\pm X$ then only (\[Atwist\]) is unitary for $\chi$ real. Thus the inverse transformation $F_{A,1/2}=(W^{}\otimes W^{})\,F_{A,1}\,\Delta^{}(W^{-1})$ can be treated as unitarizing the non-unitary twist $F_{A,1}$ by the coboundary twist $(W^{}\otimes W^{})\Delta^{}(W^{-1})$ (cf \[Atwist2\]). Alternatively, by introducing the (non-standard) flipped conjugation on the tensor product (see [@LNR91]) $$\label{d-1f} (a\otimes b)^\divideontimes = b^\divideontimes\otimes a^\divideontimes$$ and in the formulae (\[d-1\]) one can regain all the twist $F_{A,s}$ unitary as well for the imaginary parameter $\chi$. This property will be used later in Sect. 5.4 for the case of quantized Abelian twist in a quantized Lorentz algebra. Two basic $\mathfrak{sl}(2; \mathbb{C})$ quantizations and their real versions ------------------------------------------------------------------------------ ### Quantization of Jordanian $r$-matrix $r_J$ The quantum twist $F_{J_{}^{}}$ corresponding to the classical Jordanian $r$-matrix $r_J$ is well known since a long time [@Og93] [^19] $$\label{j1} F_{J^{}}^{}(\chi)\,=\,\exp{(H\otimes\sigma}),\qquad\sigma\;=\;\ln(1+\chi E_{+})~.$$ The twisted coproducts and antipodes are easy to derive $$\begin{aligned} \label{j2} \Delta_{J}(E_{+})&=&{\mathcal F}(\chi)\Delta^{(0)}(E_+) {\mathcal F}^{-1}(\chi)=E_{+}\otimes e^{\sigma}+1\otimes E_{+} \nonumber \\ && \nonumber\\ \Delta_{J}(H) &= &H\otimes e^{-\sigma}+1\otimes H =H\otimes 1+1\otimes H-\chi H\otimes E_{+} e^{-\sigma} \\ && \nonumber\\ \Delta_{J}(E_{-})&=&E_{-}\otimes e^{-\sigma}+1\otimes E_{-}+2\chi H\otimes H e^{-\sigma} \nonumber \\[8pt] && -\chi^2 H (H-1)\otimes E_{+}e^{-2\sigma}\nonumber \end{aligned}$$ and $$\begin{aligned} \label{j3} S_{J}(E_{+})&=&-E_{+}\,e^{-\sigma} ,\qquad\qquad S_{J}(H)=-H\,e^{-\sigma}\nonumber\\[8pt] S_{J}(E_{-})&=&-E_{-}\,e^{\sigma} +2\chi H^2 e^{\sigma} +\chi^2 H (H-1)E_{+}e^{\sigma} $$ The quantum $R$-matrix takes the form ($R_J=F_J^{21}F_J^{-1}$) $$\label{j4} R_{J^{}}^{}(\chi)\,=\,F_J^{21}(\chi)F_J^{-1}(\chi )=\exp{(\sigma\otimes H)} \exp{(-H\otimes\sigma)} .$$ The only compatible reality condition for the $\mathfrak{sl}(2;\mathbb{C})$ Jordanian deformation is of non-compact $\mathsf{sl}_J(2)$ type (see \[rm6-4\]) obtained if the parameter $\chi\in\imath\mathbb{R}$. In such case the Jordanian twist $F_J= \exp{(H\otimes\ln(1+\chi E_+))}$ is unitary, provides deformed coproducts and antipodes satisfying automatically the conditions (\[d-1\]). Therefore, it provides (see \[Rtwist\]) the (real=antireal) universal R-matrix $R_J=\exp{(\ln(1+\chi E_+)\otimes H)}\exp{(-H\otimes\ln(1+\chi E_+))}$.\ ### Standard quantization of $\mathfrak{sl}(2;\mathbb{C})$ The standard (non-triangular) quantum deformation is corresponding to the solution of CYBE given by (\[rm5\]). It is described by $q$-analog or Drinfeld-Jimbo quantum deformation, with algebraic and coalgebraic sectors given by the following formulae $$\begin{aligned} \label{ex1} q_{}^{\H}\E_{\pm}\!\!&=\!\!&q_{}^{\pm 1}\E_{\pm}\,q_{}^{\H}~,\quad q_{}^{\H_{}}q_{}^{-\H_{}}=q_{}^{-\H_{}}q_{}^{\H_{}}=1~,\quad [\E_{+},\,\E_{-}]\;=\;\frac{q_{}^{2\H}-q_{}^{-2\H}}{q_{}^{}-q_{}^{-1}} ~, \\[7pt]\label{ex2} \Delta_{q_{}}^{}(q_{}^{\pm \H_{}})\!\!&=\!\!&q_{}^{\pm \H_{}}\otimes q_{}^{\pm \H_{}}~, \qquad \Delta_{q_{}}^{}(\E_{\pm})\;=\;\E_{\pm}\otimes q_{}^{\H_{}}+q_{}^{-\H_{}}\otimes \E_{\pm}~, \\[7pt]\label{ex3} S_{q_{}}^{}(q_{}^{\pm \H_{}})\!\!&=\!\!&q_{}^{\mp \H_{}}~,\qquad S_{q_{}}^{}(\E_{\pm})\;=\;-q_{}^{\pm 1}\E_{\pm}~, \\[7pt]\label{ex4} \epsilon_q(q_{}^{\pm\H})\!\!&=\!\!& 1~,\qquad \epsilon_q(\E_{\pm})=0~.\end{aligned}$$ where we denote by $(q^\H, \E_\pm)$ the $q$-deformed or quantum CW basis. [^20] The quantum universal $R$-matrix satisfying QYBE (\[i3b\]) as well as the conditions (\[i2\])-(\[i3a\]) is given by the formula: $$\label{ex5b} R_{q}=\exp_{q_{}^{-2}}\Bigl((q_{}^{}-q_{}^{-1})\E_{+}\, q_{}^{-\H_{}}\otimes q_{}^{\H_{}}\E_{-}\Bigr) q_{}^{2\H_{}\otimes \H_{}}= q_{}^{2\H_{}\otimes \H_{}}\exp_{q_{}^{-2}}\Bigl((q_{}^{}-q_{}^{-1})\E_{+}\, q_{}^{\H_{}}\otimes q_{}^{-\H_{}}\E_{-}\Bigr) ~.$$ where we use the standard definition of $q$-exponential $\exp_{q_{}^{-2}}$ (cf. Appendix B) $$\begin{aligned} \label{ct16} \exp_{q}(x)\!\!&:=\!\!&\sum_{n\geq0}\,\frac{x^n}{(n)_{q}^{}!}~, \quad\;(n)_{q}^{}!:=(1)_{q}^{}(2)_{q}^{}\cdots (n)_{q}^{},\quad(n)_{q}^{}=\frac{1-q^n}{1-q}~.\end{aligned}$$ The alternative second version of the universal $R$-matrix has the form: $$\label{ex5b2} R_{q}^{\tau-1}=\exp_{q_{}^{2}}\Bigl((q_{}^{-1}-q_{}^{})\E_{-}\, q_{}^{-\H_{}}\otimes q_{}^{\H_{}}\E_{+}\Bigr) q_{}^{-2\H_{}\otimes \H_{}}= q_{}^{-2\H_{}\otimes \H_{}}\exp_{q_{}^{2}}\Bigl((q_{}^{-1}-q_{}^{})\E_{-}\, q_{}^{\H_{}}\otimes q_{}^{-\H_{}}\E_{+}\Bigr) ~.$$ and provides nontrivial element $Q_q=R_qR_q^\tau$. This quantum $R$-matrices describe by their linear terms, in the limit $\gamma\mapsto 0, q\mapsto 1$, non-skewsymmetric classical $r$-matrices in the Belavin-Drinfeld form (\[rm5\]). Three standard real forms (\[rm6-1\])-(\[rm6-3\]) impose the following reality conditions on $q$-deformed generators $(q^{\pm\H},\E_\pm)$: $$\begin{aligned} \label{ex6-1} (q^\H)^{\dagger}\!\!&=\!\!&q^\H,\qquad \E_{\pm}^{\dagger}\;=\;\E_{\mp},\quad q\in\mathbb{R}\Leftrightarrow\gamma\in \mathbb{R} \quad{\rm{for}}\;\;\mathsf{su}_\gamma(2),\\[5pt]\label{ex6-2} (q^\H)^{\#}\!\!&=\!\!&q^\H,\quad\;\E_{\pm}^{\#}\;=\;-\E_{\mp},\quad\; q\in\mathbb{R}\Leftrightarrow\gamma\in \mathbb{R} \quad\;{\rm{for}}\;\;\mathsf{su}_\gamma(1,1), \\[5pt]\label{ex6-3} (q^\H)^{\star}\!\!&=\!\!&q^\H,\quad\;\E_{\pm}^{\star}\;=\;-\E_{\pm},\quad\;\; |q| =1\Leftrightarrow\gamma\in \imath\mathbb{R} \quad{\rm{for}}\;\;\mathsf{sl}_\gamma(2), \end{aligned}$$ which turn, in each case, the Hopf algebra (\[ex1\])–(\[ex4\]) into the real Hopf algebra satisfying the reality conditions (\[d-1\]). Taking into consideration the restriction on the values of $q$ one can see that reality conditions (\[rm6-1\])-(\[rm6-3\]) for $(H, E_\pm)$ have the same form as for $(\H, \E_\pm)$ (see (\[ex6-1\])-(\[ex6-3\])). The last two (non-compact) real forms coincide in the classical limit $\gamma\mapsto 0$. In the classical limit $\gamma\mapsto 0$ deformed and undeformed generators can be identified, i.e. $\H\mapsto H,\ \E_\pm\mapsto E_\pm,\ q^\H\mapsto 1$. For the first two real forms the corresponding universal R-matrix (\[ex4\]) is real, the last case (\[rm6-3\]) is antireal. We recall that however if the Jordanian Lie bialgebra has no effective deformation parameter its quantization requires the introduction of a (formal) parameter $\chi$, which permits to construct the twist and the quantum $R$-matrix as a formal power series, elements of $U(\mathfrak{sl}(2; \mathbb{C}))\otimes U(\mathfrak{sl}(2; \mathbb{C}))[[\chi]]$. In contrast, Lie bialgebras corresponding to standard deformations are parametrized by numerical (complex or real) factor $\gamma$, describing effective deformation parameter. Lie bialgebras of complex $D=4$ rotations and their real forms ============================================================== In this section we describe Lie bialgebra of $D=4$ complex rotations $\mathfrak{o}(4;\mathbb{C})$ [^21] and its real forms: Euclidian, Lorentz, Kleinian and quaternionic orthogonal Lie algebras in terms of chiral left $(H,E_\pm)$ and right $(\bar H, \bar E_\pm)$ CW bases: [^22] $$\begin{aligned} \label{3o} [H,\,E_{\pm}]\!\!&=\!\!&E_{\pm}~,\quad[E_{+},\,E_{-}]\,=\,2H~,\quad[\bar{H},\,\bar{E}_{\pm}]\,=\,\bar{E}_{\pm}~,\quad [\bar{E}_{+},\,\bar{E}_{-}]\,=\,2\bar{H}~.\end{aligned}$$ Due to the fact that each $\mathfrak{sl}(2;\mathbb{C})$ sector has two bialgebra structures (single Jordanian and standard one-parameter family) one can easily to identify three (up to the flip) types of bialgebra structures on $\mathfrak{o}(4;\mathbb{C})$, namely the direct sums $$\begin{aligned} \label{3o1} \mathsf{o}_{\gamma,\bar\gamma}(4;\mathbb{C})\!\!&=\!\!&\mathsf{sl}_\gamma(2;\mathbb{C})\oplus\bar{\mathsf{sl}}_{\bar\gamma}(2;\mathbb{C}), \\ \label{3o2} \mathsf{o}_{\gamma,\bar J}(4;\mathbb{C})\!\!&=\!\!&\mathsf{sl}_\gamma(2;\mathbb{C})\oplus\bar{\mathsf{sl}}_{J}(2;\mathbb{C}). \\ \label{3o3} \mathsf{o}_{J,\bar J}(4;\mathbb{C})\!\!&=\!\!&\mathsf{sl}_J(2;\mathbb{C})\oplus\bar{\mathsf{sl}}_J(2;\mathbb{C}).\end{aligned}$$ with the classical $r$-matrices obtained by summing up the pair of chiral and antichiral contributions, e.g. $r_{J,\bar J}=r_J+\bar r_J=H\wedge E_+ +\bar H\wedge \bar E_+$ in (\[3o3\]), etc. The list (\[3o1\])–(\[3o3\]) does not exhaust all possible bialgebra structures because it does not take into account the mixed terms belonging to $\mathfrak{sl}(2;\mathbb{C})\wedge\bar{\mathfrak{sl}}(2;\mathbb{C})$, which can also contribute to the classical $r$-matrices. In [@BLTnov15] using purely algebraic methods we classified all $\mathfrak{o}(4;\mathbb{C})$ bialgebras. We found five families of complex skewsymmetric $r$-matrices: three, each with three-parameters, one two-parameter and one with one parameter.The list of $\mathfrak{o}(4;\mathbb{C})$ $r$-matrices looks as follows [@BLTnov15]: [^23] $$\begin{aligned} r_{I}^{}(\chi)\!\!&=\!\!&\chi(E_{+}+\bar{E}_{+})\wedge(H+\bar{H})~, \label{r1} \\[7pt] r_{II}^{}(\chi,\bar{\chi},\varsigma)\!\!&=\!\!&\chi\,E_{+}\wedge H+\bar{\chi}\,\bar{E}_{+}\wedge\bar{H}+\varsigma E_{+}\wedge\bar{E}_{+}~, \label{r2}\\[7pt] r_{III}^{}(\gamma,\bar{\gamma},\eta)\!\!&=\!\!&\gamma\,E_{+}\wedge E_{-} +\bar{\gamma}\,\bar{E}_{+}\wedge\bar{E}_{-} + \eta\,H\wedge\bar{H}~, \label{r3}\\[7pt] r_{IV}^{}(\gamma,\varsigma)\!\!&=\!\!&\gamma\left(E_{+}\wedge E_{-} -\bar{E}_{+}\wedge\bar{E}_{-} -2H\wedge\bar{H}\right)+\varsigma E_{+}\wedge\bar{E}_{+} \label{r4}\\[7pt] r_{V}^{}(\gamma,\bar{\chi},\rho)\!\!&=\!\!&\gamma\,E_{+}\wedge E_{-} +\bar{\chi}\,\bar{E}_{+}\wedge\bar{H}+\rho H\wedge\bar{E}_{+}~. \label{r5} $$ Here all parameters $\gamma$, $\bar{\gamma}$, $\eta$, $\chi$, $\bar{\chi}$, $\varsigma$, $\rho$, $\bar{\rho}$ are arbitrary complex numbers and they are independent in different $r$-matrices. The first two $r$-matrices $r_{I}^{}(\chi)$ and $r_{II}^{}(\chi,\bar{\chi},\varsigma)$, generate twist and they satisfy the homogeneous CYBE (\[i4\]). Moreover the first $r$-matrix $r_{I}^{}(\chi)$ is pure Jordanian type and the second $r$-matrix $r_{II}^{}(\chi,\bar{\chi},\varsigma)$ is the sum of two Jordanian ones with third one describing Abelian twist: $r_{II}^{}(\chi,\bar{\chi},\varsigma)= r_{II}^{}(\chi,0,0)+r_{II}^{}(0,\bar{\chi},0)+r_{II}^{}(0,0,\varsigma)$. The third $r$-matrix $r_{III}^{}(\gamma,\bar{\gamma},\eta)$ is the sum of two standard $r$-matrices and one Abelian: $r_{III}^{}(\gamma,\bar{\gamma},\eta)= r_{III}^{}(\gamma,0,0)+r_{III}^{}(0,\bar{\gamma},0)+r_{III}^{}(0,0,\eta)$. The fourth $r$-matrix $r_{IV}^{}(\gamma,\chi')$ is the sum of special choice of the third $r$-matrix and the Abelian $r$-matrix: $r_{IV}^{}(\gamma,\varsigma):=r_{III}^{}(\gamma,-\gamma,-2\gamma)+\varsigma E_{+}\wedge\bar{E}_{+}$. The last $r$-matrices $r_{V}^{}(\gamma,\bar{\chi},\rho)$ is the sum of standard, Jordanian and Abelian $r$-matrices: $r_{V}^{}(\gamma,\bar{\chi},\rho)=r_{V}^{}(\gamma,0,0)+r_{V}^{}(0,\bar{\chi},0)+r_{V}^{}(0,0,\rho)$. The formulae for ($r_{II}, r_{III}, r_{V}$) are obtained by supplementing (\[3o1\]) - (\[3o3\]) with particular additional Abelian contributions belonging to $\mathfrak{sl}(2;\mathbb{C})\wedge\bar{\mathfrak{sl}}(2;\mathbb{C})$. We shall calculate as well in next Section for all five quantizations generated by (\[r1\])–(\[r5\]) the universal $R$-matrices. Using formula (\[i1\]) one obtains in third, fourth and fifth cases the following Belavin-Drinfeld type of matrices which appear in the expansion (\[i1\]): $$\begin{aligned} \begin{array}{rcl} \tilde r_{III}^{}(\gamma,\bar{\gamma},\eta)\!\!&=\!\!&\gamma\,\left(E_{+}\otimes E_{-}+ H\otimes H \right)+\bar{\gamma}\,\left( \bar{E}_{+}\otimes\bar{E}_{-}+\bar H\otimes\bar H\right)+\eta\,H\wedge\bar{H}~, \label{r3t}\\[7pt] \tilde r_{IV}^{}(\gamma,\varsigma)\!\!&=\!\!&\gamma\,\left(E_{+}\otimes E_{-}+H\otimes H -\bar{E}_{+}\otimes\bar{E}_{-}-\bar H\otimes\bar H -2H\wedge\bar{H}\right)+\varsigma E_{+}\wedge\bar{E}_{+} \label{r4t}\\[7pt] \tilde r_{V}^{}(\gamma,\bar{\chi},\rho)\!\!&=\!\!&\gamma\,\left(E_{+}\otimes E_{-}+ H\otimes H \right)+\bar{\chi}\,\bar{E}_{+}\wedge\bar{H}+\rho H\wedge\bar{E}_{+}~. \label{r5t} \end{array}\end{aligned}$$ There is unique compact real form $\mathfrak{o}(4)$ and three real non-compact forms of $\mathfrak{o}(4)$: [*the Lorentz algebra $\mathfrak{o}(3,1):=\mathfrak{o}(3,1;\mathbb{R})\cong\mathfrak{sl}(2;\mathbb{C})^{\mathbb{R}}$*]{}, [*the Kleinian algebra $\mathfrak{o}(2,2):=\mathfrak{o}(2,2;\mathbb{R})\cong\mathfrak{o}(2,1)\oplus\mathfrak{o}(2,1)$*]{} and [*the quaternionic Lie algebra $\mathfrak{o}^{\star}(4):=\mathfrak{o}(2;\mathbb{H})\cong\mathfrak{o}(2,1)\oplus\mathfrak{o}(3)$*]{}. These real forms can be expressed as the following six direct sums of $\mathfrak{sl}(2;\mathbb{C})$- real forms listed in (\[rm6-1\])–(\[rm6-3\]) $$\begin{aligned} \label{rf1} &H^{\dagger}=H,\quad E_{\pm}^{\dagger}=E_{\mp},\quad\bar{H}^{\dagger}=\bar{H},\quad\bar{E}_{\pm}^{\dagger}=\bar{E}_{\mp}\quad{\rm{for}}\;\mathfrak{o}(4), \\[5pt] &\begin{array}{l}\label{rf2} {H}^{\dagger}=H,\quad E_{\pm}^{\dagger}=E_{\mp},\quad(\bar{H})^{\#}=\bar{H},\quad(\bar{E}_{\pm})^{\#}=-\bar{E}_{\mp}, \\[2pt] H^{\dag}=H,\quad E_{\pm}^{\dag}=E_{\mp},\quad(\bar{H})^{\star}=-\bar{H},\quad(\bar{E}_{\pm})^{\star}=-\bar{E}_{\pm} \end{array}\;\;{\rm{for}}\;\mathfrak{o}^{\star}(4), \\[5pt] &\begin{array}{l}\label{rf3} {H}^{\#}=H,\quad {E_{\pm}}^{\#}=-E_{\mp},\quad(\bar{H})^{\#}=\bar{H},\quad(\bar{E}_{\pm})^{\#}=-\bar{E}_{\mp}, \\[2pt] {H}^{\#}=H,\quad {E_{\pm}}^{\#}=-E_{\mp},\quad(\bar{H})^{\star}=-\bar{H},\quad(\bar{E}_{\pm})^{\star}=-\bar{E}_{\pm}, \\[2pt] {H}^{\star}=-H,\quad{E_{\pm}}^{\star}=-E_{\pm},\quad(\bar{H})^{\star}=-\bar{H},\quad(\bar{E}_{\pm})^{\star}=-\bar{E}_{\pm} \end{array}{\rm{for}}\;\mathfrak{o}(2,2), \\[5pt]\label{rf4} &H^{\ddag}=-\bar{H},\quad E_{\pm}^{\ddag}=-\bar{E}_{\pm},\quad (\bar{H}^{\ddag}=-H,\quad\bar{E}_{\pm}^{\ddag}=-E_{\pm})\quad{\rm{for}}\;\mathfrak{o}(3,1).\end{aligned}$$ The last real form (\[rf4\]) characterizing the Lorentz $\mathfrak{o}(3,1)$-algebra, does not preserve the chiral decomposition. By imposing all real involutions in the list of classical complex $r$-matrices (\[r1\])–(\[r5\]) we get complete set of real bialgebra structures on the Lie algebra $\mathfrak{o}(4;\mathbb{C})$. The list of all real bialgebras for $\mathfrak{o}(4;\mathbb{C})$, together with specified values for the corresponding parameters, is presented in the table below, where $\mathfrak{o}^\star(4)$, $\mathfrak{o'}^{\star}(4)$ denotes the bialgebras after imposing the reality conditions (\[rf2\]), and $\mathfrak{o}''(2,2)$, $\mathfrak{o}'(2,2)$, $\mathfrak{o}''(2,2)$ denotes three bialgebras obtained by applying three reality conditions (\[rf3\]). -------------------------------------- --------------------------- --------------------------------------------------------------- -------------------------------------------------------------- --------------------------------------- ------------------------------------------------------------ \[-11pt\] • $ r_I(\chi) $ $ r_{II}(\chi, \bar\chi,\varsigma)$ $ r_{III}(\gamma, \bar\gamma,\eta)$ $ r_{IV}(\gamma, \varsigma)$ $ r_V(\gamma, \bar\chi,\rho)$ \[-11pt\] $\mathfrak{o}(4)$ • • $\gamma,\bar\gamma\in\mathbb{R}$ ; $\eta\in\imath\mathbb{R}$ • • \[-11pt\] $\mathfrak{o}^\star(4)$ • • $\gamma,\bar\gamma\in\mathbb{R}$ ; $\eta\in\imath\mathbb{R}$ • • \[-11pt\] $\mathfrak{o'}^{\star}(4)$ • • $\gamma,\eta\in\mathbb{R}$ ; $\bar\gamma\in\imath\mathbb{R}$ • $\gamma,\rho\in\mathbb{R}$ ; $\bar\chi\in\imath\mathbb{R}$ \[-11pt\] $\mathfrak{o}(2,2)$ • • $\gamma,\bar\gamma\in\mathbb{R}$ ; $\eta\in\imath\mathbb{R}$ • • \[-11pt\] $\mathfrak{o}'(2,2)$ • • $\gamma,\eta\in\mathbb{R}$ ; $\bar\gamma\in\imath\mathbb{R}$ • $\gamma,\rho\in\mathbb{R}$ ; $\bar\chi\in\imath\mathbb{R}$ \[-11pt\] $\mathfrak{o}''(2,2)$ $\chi\in\imath\mathbb{R}$ $\chi,\bar\chi, \varsigma\in\imath\mathbb{R}$ $\gamma,\bar\gamma, \eta\in\imath\mathbb{R}$ $\gamma,\varsigma\in\imath\mathbb{R}$ $\gamma,\bar\chi, \rho\in\imath\mathbb{R}$ \[-11pt\] $\mathfrak{o}(3,1)$ $\chi\in\imath\mathbb{R}$ $\chi=\bar\chi\in\imath\mathbb{R}$ ; $\varsigma\in\mathbb{R}$ $\bar\gamma=-\gamma^*\in\mathbb{C}$ ; $\eta\in\mathbb{R}$ $\gamma,\varsigma\in\mathbb{R}$ • -------------------------------------- --------------------------- --------------------------------------------------------------- -------------------------------------------------------------- --------------------------------------- ------------------------------------------------------------ : All real Lie bialgebras for $\mathfrak{o}(4;\mathbb{C})$[]{data-label="tab:Real"} In the following Section we shall describe the Hopf-algebraic quantization of five complex $\mathfrak{o}(4;\mathbb{C})$ $r$-matrices (\[r1\])–(\[r5\]). Out of these five complex quantizations after imposing seven reality conditions we obtain sixteen real $\mathfrak{o}(4;\mathbb{C})$ Hopf algebra structures: $r_{III}$ provides seven real forms, $r_V$ – three, and each of remaining three leads to two real quantizations. Explicit quantizations of $\mathfrak{o}(4;\mathbb{C})$ and their real forms =========================================================================== Jordanian quantization of $\mathfrak{o}(4;\mathbb{C})$ ($r$-matrix $r_I$) ------------------------------------------------------------------------- Following the previous considerations (Subsect. 2.2) the quantum twist $F_{1_{}^{}}$ corresponding to the classical Jordanian $r$-matrix (\[r1\]) can be written as $$\label{J1} F_{1^{}}^{}(\chi)\,=\,\exp{((H+\bar H)\otimes\sigma}),\qquad\sigma\;=\;\ln(1+\chi (E_{+}+\bar E_+))~,$$ Coproducts and antipodes are easy to derive (cf. (\[j1\])–(\[j4\])) $$\begin{aligned} \label{J2}\Delta_{1}(E_{k+}) &=&{\mathcal F}(\chi)\Delta^{(0)}(E_k) {\mathcal F}^{-1}(\chi)=\Delta_1(E_{k+})=E_{k+}\otimes e^{\sigma}+1\otimes E_{k+} \nonumber \\ && \nonumber\\ \Delta_{1}(H_k) &= & H _k\otimes 1+1\otimes H_k-\chi(H+\bar H)\otimes E_{k+} e^{-\sigma} \nonumber\\ && \\ \Delta_{1}(E_{k-})&=&E_{k-}\otimes e^{-\sigma}+1\otimes E_{k-} +2\chi (H+\bar H)\otimes H_ke^{-\sigma} \nonumber \\[8pt] && -\chi^2 (H+\bar H)(H+\bar H-1)\otimes E_{k+}e^{-2\sigma}\nonumber \end{aligned}$$ where $k\in \{0, 1\}\equiv \mathbb{Z}_2$ and in order to reduce the number of formulae we denoted $X_0= \{H=H_0, E_\pm=E_{0\pm}\}$ and $X_1= \{\bar H=H_1, \bar E_\pm=E_{1\pm}\}$. [^24] Similarly, the formulae for the antipodes look as follows $$\begin{aligned} \label{J3} S_{1}(E_{k+})&=&-E_{k+}\,e^{-\sigma} ,\qquad\qquad S_{1}(H_k)=-H_k -\chi (H+\bar H) E_{k+}\nonumber\\[8pt] S_{1}(E_{k-})&=&-E_{k-}\,e^{\sigma} +2\chi (H+\bar H)H_k e^{\sigma} +\chi^2 (H+\bar H)(H+\bar H-1)E_{k+}e^{\sigma} $$ The universal quantum $R$-matrix takes the form ($R=F^{21}F^{-1}$) $$\label{J4} R_{1^{}}^{}(\chi)\,=\,\exp{(\sigma\otimes (H+\bar H)}) \exp{(-(H+\bar H)\otimes\sigma}) .$$ This simple one-parameter deformation admits two real quantum group structures (cf. Table \[tab:rr1\]) as indicated below. Since the twist is Jordanian, the reality conditions (\[d-1\]) are valid if the deformation parameter $\chi$ is imaginary. The Lorentzian case requiring as well imaginary $\chi$ has been already studied in [@BLT08] with more details. -------------------------------- ----------------------------------------------- ------------------------------------------------------------- ---------------------------------------------------------------- \[-8pt\] $\mathfrak{o}''(2,2)$ $\ \ \ \ \ \chi\in\imath\mathbb{R}\ \ \ \ \ $ $H^{\star}=-H, E_{\pm}^{\star}=-E_{\pm}$ $\bar H^{\star}= -\bar H , \bar E_{\pm}^{\star}=-\bar E_{\pm}$ \[-8pt\] $\mathfrak{o}(3,1)$ $\ \ \ \ \ \chi\in\imath\mathbb{R}\ \ \ \ \ $ $H^{\ddagger} =-\bar{H} , E_{\pm}^{\ddagger}=-\bar E_{\pm}$ $\bar H^{\ddagger}= -H , \bar E_{\pm}^{\ddagger}=- E_{\pm}$ -------------------------------- ----------------------------------------------- ------------------------------------------------------------- ---------------------------------------------------------------- : Real quantizations of $r_{I}(\chi)=\chi(E_{+}+\bar{E}_{+})\wedge(H+\bar{H})$[]{data-label="tab:rr1"} Left and right Jordanian quantizations intertwined by Abelian twist ($r$-matrix $r_{II}$) ------------------------------------------------------------------------------------------ We see that for $\varsigma=0$ the $r$-matrix (\[r2\]) describes two complex Jordanian $r$-matrices, each one for chiral sectors $\mathfrak{sl}(2,\mathbb{C})$ and $\overline{\mathfrak{sl}(2;\mathbb{C})}$. They do commute with each other and can be quantized as the product of two Ogievetsky twists ($k=1,2$) ([@Og93] see also (\[j1\])) $$\label{e10} F_{J,0}(\chi) = \exp{(H\otimes \Sigma)} \,\quad F_{J,1}(\bar\chi) = \exp{(\bar H\otimes \bar\Sigma)}$$ where $\Sigma=ln{(1+\chi E_+)}$, $\bar\Sigma=ln{(1+\bar\chi \bar E_+)}$. The next step is to consider the Abelian part of the classical $r$-matrix $r_{II}$ belonging to $\mathfrak{sl}(2;\mathbb{C})\wedge\bar{\mathfrak{sl}}(2;\mathbb{C})$ intertwining two chiral coalgebra sectors which ceases to be independent. Because the generators $(H,\,E_{\pm})$ and $(\bar H,\, \bar E_\pm)$ do commute the twist function corresponding to (\[r2\]) is given by the following formula: $$\label{e11} F_2(\chi,\bar\chi, \varsigma)= F_A (\chi,\bar\chi, \varsigma) F_{J,1}(\bar\chi) F_{J,0}(\chi) = F_A (\chi,\bar\chi, \varsigma) F_{J,0}(\chi) F_{J,1}(\bar\chi)\,.$$ where the Abelian twist $F_A$ takes the form[^25]$$\label{e12} F_A(\chi,\bar\chi, \varsigma)=\exp{(\frac{\varsigma}{\chi\bar\chi}\Sigma\wedge\bar\Sigma)}$$ which follows from the property that elements $\Sigma,\ \bar\Sigma$ are primitive after performing Jordanian deformation. We would like to mention here that the form of the twist function given above by formula (\[e11\]) was proposed firstly by Kulish and Mudrov [@KM99]. If we use (\[blte7\]–\[blte8\]), and (\[e11\]) we obtain the following formulae for the coproducts of $sl(2;\mathbb{C})\oplus \bar{sl}(2;\mathbb{C})$ generators $(H_k,\, E_{k+},\,E_{k-})$, $k=0,1\in \mathbb{Z}_2$ $$\begin{aligned} \label{e21} \Delta_{2}(E_{k+}) &=&{\mathcal F}(\chi,\bar\chi,\varsigma)\Delta^{(0)}(E_k) {\mathcal F}^{-1}(\chi,\bar\chi,\varsigma)=E_{k+}\otimes e^{\Sigma_k}+1\otimes E_{k+}\nonumber \\\nonumber && \\\nonumber \Delta_{2}(H_k) &= & H _k\otimes e^{-\Sigma_k}+1\otimes H_k\, + \nonumber \\[8pt]&& (-1)^k \frac{\varsigma}{\chi_{k+1}}\,\left(\Sigma_{k+1}\otimes E_{k+} e^{-\Sigma_k}-\, E_{k+} e^{-\Sigma_k}\otimes\Sigma_{k+1} e^{-\Sigma_k}\right)\nonumber \\\nonumber && \\\Delta_{2}(E_{k-})&=&E_{k-}\otimes e^{-\Sigma_k}+1\otimes E_{k-} +2\chi_k H_k\otimes H_ke^{-\Sigma_k} +\chi_k H_k(H_k-1)\otimes\Lambda_k+\\[8pt]&& \,(-)^k\frac{2\varsigma}{\chi_{k+1}}\,\left( H_ke^{-\Sigma_k}\otimes\Sigma_{k+1}e^{-\Sigma_k} - H_k\Sigma_{k+1}\otimes\Lambda_k -\Sigma_{k+1}\otimes H_ke^{-\Sigma_k}\right) \nonumber \\[8pt]&& \,(-)^k\frac{2\varsigma}{\chi_{k+1}}\, \left(\, \Lambda_k e^{\Sigma_k}\otimes H_k\Sigma_{k+1}e^{-\Sigma_k}\,+ H_k\Lambda_ke^{\Sigma_k}\otimes\Sigma_{k+1}\Lambda_k\, \right) \nonumber \\[8pt]&& +\,(-)^k\,\frac{\varsigma}{\chi_{k+1}}\,\left( \left(1-e^{-2\Sigma_k}\right)\otimes\Sigma_{k+1}\Lambda_k\, +\Sigma_{k+1}\otimes\Lambda_k\,-\Lambda_k\otimes \Sigma_{k+1}e^{-\Sigma_k}\right) \nonumber\\[8pt] && {1\over\chi_k}\left(\frac{\varsigma}{\chi_{k+1}}\right)^2\,\left(\Lambda^2_ke^{2\Sigma_{k}}\otimes\Sigma^2_{k+1}\Lambda_k +\Lambda_k\otimes\Sigma^2_{k+1}e^{-\Sigma_k} +\Sigma^2_{k+1}\otimes\Lambda_k \right) \nonumber\\[8pt]&& - {2\over\chi_k}\left(\frac{\varsigma}{\chi_{k+1}} \right)^2\,\Lambda_k \Sigma_{k+1} e^{\Sigma_k}\otimes\Sigma_{k+1}\Lambda_k\, \nonumber\end{aligned}$$ Here $\Sigma_{k+1}$ is denoted with index mod 2, i.e. $\Sigma_{k+1}$ is equal to $\Sigma_0$ for $k=1$; further $\Lambda_k=e^{-2\Sigma_k}-e^{-\Sigma_k}= -\chi_kE_{k+}e^{-2\Sigma_k}$ [^26]. Therefore, $\Lambda_k$ is proportional to $\chi_k$. Using further the relations (\[blte7\]-\[blte8\]) one obtains the following formulae for the antipodes $$\begin{aligned} \label{blte22} S_{2}(E_{k+})&=&-E_{k+}\,e^{-\Sigma_k} ,\qquad\qquad S_{2}(H_k)=-H_ke^{\Sigma_k}\nonumber\\[8pt] S_{2}(E_{k-})&=&-E_{k-}\,e^{\Sigma_k} +\chi_k H^2_ke^{\Sigma_k}(e^{\Sigma_k}+1) -\chi_k^2 H_kE_{k+}e^{\Sigma_k} $$ We notice that the Abelian twist (\[e12\]) does not contribute to the antipodes (\[blte22\]). The quantum universal $R$-matrix $R_2\equiv R_{2^{}}^{}(\chi,\bar\chi,\varsigma)$ takes the form $$\label{qr2} R_{2^{}}^{}=\exp{(\frac{-\varsigma}{\chi\bar\chi}\Sigma\wedge\bar\Sigma)}\exp{(\Sigma\otimes H)} \exp{(-H\otimes \Sigma)} \exp{(\bar\Sigma \otimes \bar H)} \exp{(-\bar H\otimes \bar\Sigma)} \exp{(\frac{-\varsigma}{\chi\bar\chi}\Sigma\wedge\bar\Sigma)}.$$ The formulae (\[e21\])–(\[qr2\]) present the general three-parameter deformation which can be studied in various two-parameter limits. For example, if $\chi\mapsto 0$ one should take into account that $\lim_{\chi\mapsto 0}\frac{\Sigma}{\chi}=E_+$, $\lim_{\chi\mapsto 0}\Lambda=0$ and $\lim_{\chi\mapsto 0}\frac{\Lambda}{\chi}=-E_+$. In this case the left chiral sector will be deformed only by Abelian twist. The case $\varsigma=0$ provides obviously the product of two independent Jordanian deformations. In real cases the independence of parameters may be not valid. Only for the real $\mathfrak{o}(2,2)$ deformation all three parameters are imaginary and independent. In the Lorentzian case [^27] two Jordanian parameters $(\chi, \bar\chi)$ are replace by one as follows from the condition $\chi=(\bar\chi)^*$ in the table below. -------------------------------- ----------------------------------------------------------------- ------------------------------------------------------------ --------------------------------------------------------------- \[-8pt\] $\mathfrak{o}''(2,2)$ $\chi,\bar\chi, \varsigma\in\imath\mathbb{R}$ $H^{\star}=-H, E_{\pm}^{\star}=-E_{\pm}$ $\bar H^{\star}=-\bar H , \bar E_{\pm}^{\star}=-\bar E_{\pm}$ \[-8pt\] $\mathfrak{o}(3,1)$ $\chi=\bar\chi^*\in\imath\mathbb{R}$ ; $\varsigma\in\mathbb{R}$ $H^{\ddagger} =-\bar H , E_{\pm}^{\ddagger}=-\bar E_{\pm}$ $\bar H^{\ddagger}= - H , \bar E_{\pm}^{\ddagger}=-E_{\pm}$ -------------------------------- ----------------------------------------------------------------- ------------------------------------------------------------ --------------------------------------------------------------- : Real quantizations of $r_{II}(\chi, \bar\chi, \varsigma)=\chi\,E_{+}\wedge H+\bar{\chi}\,\bar{E}_{+}\wedge\bar{H}+\varsigma E_{+}\wedge\bar{E}_{+}$[]{data-label="tab:rr2"} All the twists present in the formula (\[e11\]) are unitary (if the corresponding parameters are as indicated in the Table \[tab:rr2\]) and the reality conditions (\[d-1\]) are satisfied. Twisted pair of $q$-analogs ($r$-matrix $r_{III}$) --------------------------------------------------- From the structure of the classical $r$-matrix $r_{3}$ (see (\[r3\])) for $\eta=0$ follows that a quantum deformation $U_{r'_{3}} (\mathfrak{o}(3,1))$ is a combination of two independent $q$-analogs (stanadard deformations) of $U(\mathfrak{sl}(2;\mathbb{C}))$ with the parameter $q=\exp{{1\over 2}\gamma}=q_{0}$ and $\bar q=\exp{{1\over 2} \bar{\gamma}}=q_{1}$. Moreover one has the splitting $U_{(q,\bar q)}(\mathfrak{o}(4;\mathbb{C}))\cong U_{q_{}}(\mathfrak{sl}(2;\mathbb{C}))\otimes U_{\bar q_{}}(\mathfrak{sl}(2;\mathbb{C}))$. This implies that the starting point for further considerations is a pair of standard (Drinfeld-Jimbo) deformations in each chiral sector. They are described by nonlinear (quantum) generators $q_{k^{}}^{\pm \H_{k}}$, $\E_{k\pm}\ (k=0,1)$ which satisfy the following defining relations $$\begin{aligned} \label{ct1} q_{k}^{\H_k}\E_{k\pm}\!\!&=\!\!&q_{k}^{\pm 1}\E_{k\pm}\,q_{k}^{\H_k}~,\qquad [\E_{k+},\,\E_{k-}]\;=\;\frac{q_{k}^{2\H_k}-q_{k}^{-2\H_k}} {q_{k}^{}-q_{k}^{-1}}~, $$ The co-products $\Delta_{3'}$ and antipodes $S_{3'_{}}$ are given by the formulas : $$\begin{aligned} \label{ct3} \Delta_{3'_{}}^{}(q_{k}^{\pm \H_{k}})\!\!&=\!\!&q_{k}^{\pm \H_{k}}\otimes q_{k}^{\pm \H_{k}}~, \qquad\Delta_{3'_{}}^{}(\E_{k\pm})\;=\;\E_{k\pm}\otimes q_{k}^{\H_{k}}+q_{k}^{-\H_{k}}\otimes \E_{k\pm}~, $$ $$\begin{aligned} \label{ct5} S_{3'_{}}^{}(q_{k}^{\pm \H_{k}})\!\!&=\!\!&q_{k}^{\mp \H_{k}}~,\qquad S_{3'_{}}^{}(\E_{k\pm})\;=\;-q_{k}^{\pm1}\E_{k\pm}~, $$ The universal $R$-matrices $R_{3'k}$ for each chiral sector are well-known and using deformed CW generators (\[ct1\]) take the form ($q_k=\exp{{1\over 2}\gamma_k}$): $$\begin{aligned} \label{ct13} R_{3'k}(\gamma_k)\!\!&=\!\!&\exp_{q_{k}^{-2}}\Bigl((q_{k}^{}-q_{k}^{-1})\E_{k+}\, q_{k}^{-\H_{k}}\otimes q_{k}^{\H_{k}}\E_{k-}\Bigr) q_{k}^{2\H_{k}\otimes \H_{k}}~, $$ Following the discussion of nontriangular case in Sect. 2.1, there exists alternative universal $R$-matrix in the form $$\begin{aligned} \label{ct13bis} (R_{3'k}^{\tau})^{ -1}\!\!&=\!\!&q_{k}^{-2\H_{k}\otimes \H_{k}}\,\exp_{q_{k}^{2}}\Bigl((q_{k}^{-1}-q_{k}^{})q_{k}^{\H_{k}}\E_{k-}\otimes \E_{k+}\,q_{k}^{-\H_{k}}\Bigr) ~. $$ Therefore, the universal $R$-matrix $R_{3'_{}}$, which connects the coproducts $\Delta_{3'_{}}^{12}:=\Delta_{3'_{}}^{}$ and the flipped one $\Delta_{3'_{}}^{21}$ can be written in two equivalent forms: [^28] $$\begin{aligned} \label{ct12} R_{3'_{}}(\gamma,\bar\gamma)\!\!&=\!\!&R_{3'0}(\gamma)R_{3'1}(\bar\gamma)\,=\,R_{3'1}(\bar\gamma)R_{3'0}(\gamma)~,\end{aligned}$$ Expanding (\[ct12\]) up to first order in deformation parameters $(\gamma,\bar\gamma)$ one gets $$\begin{aligned} \label{ct18} R_{3'_{}}(\gamma,\bar\gamma)\!\!&=\!\!&1+r_{3'BD}^{}+O(\gamma^2,\gamma\bar\gamma,\bar\gamma^2)~,\end{aligned}$$ where $r_{3'BD}^{}$ is in Belavin-Drinfeld form [^29] $$\begin{aligned} \label{ct19} r_{3'BD}^{}\!\!&=\!\!&\gamma\bigl(E_{+}\otimes E_{-}+H\otimes H\bigr) +\bar\gamma\bigl(\bar E_{+}\otimes \bar E_{-}+\bar H\otimes \bar H\bigr)\end{aligned}$$ This $r$-matrix is not skew-symmetric and satisfies the condition $$\begin{aligned} \label{ct21} r_{BD}^{12}+r_{BD}^{21}\!\!&=\!\!&\omega\end{aligned}$$ where $\omega$ is the quadratic split Casimir of $\mathfrak{o}(4; \mathbb{C})$ $$\begin{array}{rcl}\label{ct22} \omega\!\!&=\!\!&\gamma \bigl(E_{+}\otimes E_{-}+E_{-}\otimes E_{+}+2H\otimes H\bigr) \\[7pt] &&\!\!+\bar\gamma \bigl(\bar E_{+}\otimes \bar E_{-}+\bar E_{-}\otimes \bar E_{+}+2\bar H\otimes \bar H\bigr) \end{array}$$ We recall that the Belavin-Drinfeld $r$-matrix $r_{BD}^{}$ satisfies CYBE and the $r$-matrix $r_3'$ is a skew-symmetric part of it. Now we consider deformation of the quantum algebra $U_{(\gamma, \bar\gamma)}(\mathfrak{o}(4;\mathbb{C}))\cong U_{\gamma}(\mathfrak{sl}(2;\mathbb{C}))\otimes U_{\bar\gamma}(\mathfrak{sl}(2;\mathbb{C}))$ generated by the $r$-matrix $r''_{3}=\eta H\otimes\bar H$, (see (\[r3\])). Since the generators $\H$ and $\bar \H$ have the primitive coproduct $$\begin{aligned} \label{ct24} \Delta_{3'_{}}(\H_{k})\!\!&=\!\!&\H_{k}\otimes 1+1\otimes \H_{k}\quad(k=0, 1)~,\end{aligned}$$ the Abelian two-tensor ($\tilde q=\exp{{1\over 4}\eta}$) $$\begin{aligned} \label{ct25} F_{3''}(\eta)\!\!:=\!\!&\tilde q_{}^{\H\wedge\bar \H}\qquad $$ satisfies the 2-cocycle condition (\[coc\]). Thus the complete deformation generated by the $r$-matrix $r_{3}^{}$ is the twist deformation of $U_{(\gamma,\bar\gamma)}(\mathfrak{o} (4;\mathbb{C}))$; the resulting coproduct $\Delta_{3_{}}^{}$ is given as follows $$\begin{aligned} \label{ct28} \Delta_{3_{}^{}}^{}(a)\!\!&=\!\!&F_{3''}^{}\Delta_{3'}^{}(a) F_{3''}^{-1}\quad(\forall a\in U_{r'_{3}}(\mathfrak{o}(4;\mathbb{C}))~,\end{aligned}$$ and the antipode $S_{3^{}}^{}$ is not changed ($S_{3^{}}^{}=S_{3'_{}}^{}$). Applying the twist (\[ct25\]) to the formulas (\[ct3\]) we obtain $$\begin{aligned} \label{ct29} \Delta_{3_{}^{}}(q_{k}^{\pm \H_{k}})\!\!&=\!\!&q_{k}^{\pm \H_{k}}\otimes q_{k}^{\pm \H_{k}},\\[7pt]\label{ct30} \Delta_{3_{}^{}}(\E_{k\pm})\!\!&=\!\!&\E_{k\pm}\otimes q_{k}^{\H_{k}} \tilde q_{}^{\pm (-)^k\H_{k+1}}+q_{k}^{-\H_{k}}\tilde q_{}^{\mp (-)^k \H_{k+1}}\otimes \E_{k\pm}~. $$ The universal $R$-matrix, $R_{3_{}^{}}(\gamma,\bar\gamma,\eta)$, corresponding to the complete $r$-matrix $r_{3}^{}$, has the form $$\begin{aligned} R_{3_{}^{}}(\gamma,\bar\gamma,\eta)\!\!&=\!\!&\tilde q_{}^{\bar \H_{}\wedge \H_{}}R_{3'_{}}(\gamma,\bar\gamma) \tilde q_{}^{\bar \H_{}\wedge \H_{}}\;=\;R_{30}^{}(\gamma,\eta)R_{31}^{}(\bar\gamma,\eta) \tilde q_{}^{2\bar \H_{}\wedge \H_{}}\;= \;R_{31}^{}(\bar\gamma,\eta)R_{30}^{}(\gamma,\eta) \tilde q_{}^{2\bar \H_{}\wedge \H_{}}~,\nonumber \label{ct32}\end{aligned}$$ where $$\begin{aligned} \label{ct33} R_{3k}^{}(\gamma_k,\eta)\!\!&=\!\!&\exp_{q_{k}^{-2}}\Bigl((q_{k}^{}-q_{k}^{-1})\E_{k+} q_{k}^{-\H_{k}} \tilde q_{}^{(-)^{k+1}\H_{k+1}}\otimes q_{k}^{\H_{k}} \tilde q_{}^{(-)^{k+1}\H_{k+1}} \E_{k-}\Bigr)\,q_{k}^{2\H_{k}\otimes \H_{k}}\end{aligned}$$ In the linear limit we obtain (cf. (\[ct12\]), (\[ct18\])) $$\begin{aligned} \label{ct37} R_{3_{}^{}}\!\!&\sim\!\!&1+ r_3 $$ This deformation admits seven real forms which employ all four conjugations (cf. (\[rf1\]) – (\[rf4\]) ) . The list of real forms with corresponding restricted values of the deformation parameters $\gamma,\bar\gamma,\eta$ is presented in the Table \[tab:rr3\], with real bialgebras denoted in the first column (cf. Table \[tab:Real\]). ------------------------------------- -------------------------------------------------------------- --------------------------------------------------------------------------- ---------------------------------------------------------------------------------- --- \[-8pt\] $\mathfrak{o}(4)$ $\gamma,\bar\gamma\in\mathbb{R}$ ; $\eta\in\imath\mathbb{R}$ $(q^\H)^{\dag}=q^\H, \E_{\pm}^{\dag}=\E_{\mp}$ $(\bar q^{\bar\H})^{\dag}=\bar q^{\bar{\H} }, \bar\E_{\pm}^{\dag}=\bar\E_{\mp}$ R \[-8pt\] $\mathfrak{o}^\star(4)$ $\gamma,\bar\gamma\in\mathbb{R}$ ; $\eta\in\imath\mathbb{R}$ $(q^\H)^{\dag}=q^\H, \E_{\pm}^{\dag}=\E_{\mp}$ $(\bar q^{\bar\H})^{\#}=q^{\bar\H} , \bar\E_{\pm}^{\#}=-\bar\E_{\mp}$ R \[-8pt\] $\mathfrak{o'}^{\star}(4)$ $\gamma,\eta\in\mathbb{R}$ ; $\bar\gamma\in\imath\mathbb{R}$ $(q^\H)^{\dag}=q^\H, \E_{\pm}^{\dag}=\E_{\mp}$ $(\bar q^{\bar\H})^{\star}=\bar q^{\bar\H} , \bar\E_{\pm}^{\star}=-\bar\E_{\pm}$ H \[-8pt\] $\mathfrak{o}(2,2)$ $\gamma,\bar\gamma\in\mathbb{R}$ ; $\eta\in\imath\mathbb{R}$ $(q^\H)^{\#}=q^\H, \E_{\pm}^{\#}=-\E_{\mp}$ $(\bar q^{\bar\H})^{\#}=\bar q^{\bar\H} , \bar\E_{\pm}^{\#}=-\bar\E_{\mp}$ R \[-8pt\] $\mathfrak{o}'(2,2)$ $\gamma,\eta\in\mathbb{R}$ ; $\bar\gamma\in\imath\mathbb{R}$ $(q^\H)^{\#}=q^\H, \E_{\pm}^{\#}=-\E_{\mp}$ $(\bar q^{\bar\H})^{\star}=\bar q^{\bar\H} , \bar\E_{\pm}^{\star}=-\bar\E_{\pm}$ H \[-8pt\] $\mathfrak{o}''(2,2)$ $\gamma,\bar\gamma, \eta\in\imath\mathbb{R}$ $(q^\H)^{\star}=q^\H, \E_{\pm}^{\star}=-\E_{\pm}$ $(\bar q^{\bar\H})^{\star}=\bar q^{\bar\H} , \bar\E_{\pm}^{\star}=-\bar\E_{\pm}$ A \[-8pt\] $\mathfrak{o}(3,1)$ $\bar\gamma=-\gamma^*\in\mathbb{C}$ ; $\eta\in\mathbb{R}$ $(q^\H)^{\ddagger} =\bar{q}^{\bar\H} , \E_{\pm}^{\ddagger}=-\bar\E_{\pm}$ $(\bar q^{\bar\H})^{\ddagger}=q^{\H} , \bar\E_{\pm}^{\ddagger}=-\E_{\pm}$ A ------------------------------------- -------------------------------------------------------------- --------------------------------------------------------------------------- ---------------------------------------------------------------------------------- --- : Real quantizations of $r_{III}(\gamma, \bar\gamma, \eta)= \gamma\,E_{+}\wedge E_{-} +\bar{\gamma}\,\bar{E}_{+}\wedge\bar{E}_{-} + \eta\,H\wedge\bar{H} $[]{data-label="tab:rr3"} The letters in the last column (R=real, A=antireal, H=hybrid) indicate the properties of the $R$-matrix under respective conjugation: $R^\star=R^\tau$ for real, $R^\star=R^{-1}$ for antireal cases. In the hybrid case the $R$-matrix decomposes into a product of three factors, with first real, second antireal and third is given by twist which satisfies both reality conditions. It shoul be mentioned that only the classical $r$-matrix $r_{III}$ provides the quantum deformations of real $\mathfrak{o}(4)= \mathfrak{su}(2)\oplus \mathfrak{su)(2}$ algebra (see first line in Tab \[tab:rr3\]). Particular case, with $\eta=0$ was derived as describing quantum symmetries of $D=3$ LQG [@JKG]. Twisting of $\mathfrak{o}(4;\mathbb{C})$ Belavin–Drinfeld triple ($r$-matrix $r_{IV}$) --------------------------------------------------------------------------------------- Next, we describe quantum deformation corresponding to the classical $r$-matrix $r_{IV}$ (\[r4\]). Since the $r$-matrix $r_{IV}(\gamma,0):=r_{4'}$ is a particular case of $r_{III}^{}(\gamma,\bar\gamma,\eta)$, namely $r_{IV'}(\gamma)=r_{III}^{}(\gamma,-\gamma,-2\gamma)$, $\gamma\in\mathbb{C}$, the quantum deformation corresponding to the $r$-matrix $r_{IV'}$ is obtained from the formulae in Sect. 5.3 by setting $\bar q=\tilde q= q^{-1}$. The quantum deformation corresponding to $r_{IV}$ is generated by the elements $q_{}^{\pm \H_{k}}$, $\E_{k\pm}$ (k=0,1) with the following defining relations (cf. (\[ct1\])) $$\begin{aligned} \label{BD1} q_{}^{\H_k}\E_{k\pm}\!\!&=\!\!&q_{}^{\pm1}\E_{k\pm}\,q_{}^{\H_k}~,\qquad [\E_{k+},\,\E_{k-}]\;=\;\frac{q_{}^{2\H_k}-q_{}^{-2\H_k}} {q_{}^{}-q_{}^{-1}}$$ constituting the algebra $U_{q_{}}(\mathfrak{sl}(2;\mathbb{C}))\otimes U_{ q^{-1}}(\mathfrak{sl}(2;\mathbb{C}))$. The co-products $\Delta_{4'_{}}$ and antipodes $S_{4'_{}}$ generated by $r_{IV'}$ are given by the formulas (cf. (\[ct29\])-(\[ct30\])): $$\begin{aligned} \label{BD2} \Delta_{4'_{}}(q_{}^{\pm \H_{k}})\!\!&=\!\!&q_{}^{\pm \H_{k}}\otimes q_{}^{\pm \H_{k}} ~,\nonumber \\[7pt]\label{BD3} \Delta_{4'_{}}(\E_{k\pm})\!\!&=\!\!&\E_{k\pm}\otimes q_{}^{(-)^k(\H_{k}\pm \H_{k+1})}+ q_{}^{(-)^{k+1}(\H_{k}\pm \H_{k+1})}\otimes \E_{k\pm}~, \\[7pt]\label{BD4} S_{4'_{}}^{}(q_{}^{\pm \H_{k}})\!\!&=\!\!&q_{}^{\mp \H_{k}} ,\qquad S_{4'_{}}^{}(\E_{k\pm})\!\!=\!\!-q_{}^{\pm(-)^k}\E_{k\pm}~,\nonumber $$ The full deformation of the quantum algebra (\[BD1\])–(\[BD2\]) is obtained after performing the twist quantization generated by the remaining part of the $r$-matrix $r_{IV}$ namely $r_{IV''}=\varsigma E_+\wedge\bar E_+$, described by the following quantum Abelian twist factor [@IO01]: $$\begin{aligned} \label{BD7} F_{4_{}''}(\gamma,\varsigma)\!\!&:=\!\!&\exp_{q_{}^{2}}^{}\big(\varsigma \E_{+}q_{}^{\H_{}+\bar \H_{}} \otimes q_{}^{\H_{}+\bar \H_{}}\bar \E_{+}\big)~.\end{aligned}$$ It can be shown that the two-tensor (\[BD7\]) satisfies the 2-cocycle equation (\[coc\]). Explicit form of the co-products $\Delta_{4_{}^{}}^{}(\cdot)= F_{4_{}''}^{} \Delta_{4_{}'}^{}(\cdot)F_{4_{}''}^{-1}$ in the complex Cartan-Weyl bases of $U_{r_{4}'}(\mathfrak{o}(4;\mathbb{C}))$ can be calculated using $q$-analog of Hadamard formula (Appendix B) $$\begin{aligned} \label{BD8} \Delta_{4_{}}(q_{}^{\pm(\H_{}-\bar \H)})\!\!&=\!\!&q_{}^{\pm(\H_{}-\bar \H)}\otimes q_{}^{\pm(\H_{}-\bar \H)}~,\nonumber \\[5pt]\label{BD9} \Delta_{4_{}}(\,q_{}^{\;\,\H_{}+\bar \H})\!\!&=\!\!&\mathbb{X}^{-1}\, q_{}^{\H_{}+\bar \H}\otimes q_{}^{\H_{}+\bar \H}~,\nonumber \\[7pt]\label{BD10} \Delta_{4_{}}(q_{}^{\;-\H_{}-\bar \H})\!\!&=\!\!&q_{}^{\,-\H_{}-\bar \H}\otimes q_{}^{-\H_{}-\bar \H}\,\mathbb{X}~,\nonumber \\[7pt]\label{BD11} \Delta_{4_{}}(\E_{+})\!\!&=\!\!&\E_{+}\otimes q_{}^{\H_{}+\bar \H_{}}+ q_{}^{-\H_{}-\bar \H_{}} \otimes \E_{+}\,\mathbb{X}~,\nonumber \\[7pt]\label{BD12} \Delta_{4_{}}(\bar \E_{+})\!\!&=\!\!&\bar \E_{+}\otimes q_{}^{-\H_{}-\bar \H_{}}\mathbb{X}+ q_{}^{\H_{}+\bar \H_{}}\otimes \bar \E_{+}~,\end{aligned}$$ $$\begin{aligned} \label{BD13} \begin{array}{rcl} \Delta_{4_{}}(\E_{-})\!\!&=\!\!&\E_{-}\otimes q_{}^{\H_{}-\bar \H_{}}+ q_{}^{\bar \H_{}-\H_{}}\otimes \E_{-}\,- \\[12pt] &&-\;\displaystyle\frac{\varsigma}{q_{}^{}-q_{}^{-1}}\,\bigl(q_{}^{-4\H_{}}\otimes1- \mathbb{X}^{-1}\bigr)\bigl(q_{}^{3\H_{}+\bar \H_{}}\otimes \bar \E_{+}q_{}^{2\H}\bigr)~, \end{array}\nonumber\end{aligned}$$ $$\begin{aligned} \label{BD14} \begin{array}{rcl} \Delta_{4_{}}(\bar \E_{-})\!\!&=\!\!&\bar \E_{-}\otimes q_{}^{\H_{}-\bar \H_{}}+ q_{}^{\bar \H_{}-\H_{}}\otimes \bar \E_{-}\,- \\[12pt] &&-\;\displaystyle\frac{\varsigma}{q_{}^{}-q_{}^{-1}}\,\bigl(1\otimes q_{}^{-4\bar \H_{}}- \mathbb{X}^{-1}\bigr)\bigr(\E_{+}q_{}^{2\bar \H}\otimes q_{}^{\H_{}+3\bar \H_{}}\bigr)~, \end{array}\nonumber\end{aligned}$$ where $$\begin{aligned} \label{BD15} \mathbb{X}\!\!&:=\!\!&1+\varsigma (q_{}^{2}-1)\E_{+}q_{}^{\H_{}+\bar \H_{}}\otimes q_{}^{\H_{}+\bar \H_{}} \bar \E_{+}~.\end{aligned}$$ Explicit formulas for antipodes $S_{4}(\cdot)=uS_{4_{}'}(\cdot)u^{-1}$ where $$\begin{aligned} \label{BD16} u^{-1}\!\!&=\!\!&m\circ (S_{4_{}'}\otimes\mathop{\rm id})\exp_{q_{}^{2}}^{}\big(\varsigma\,\E_{+}q_{}^{\H_{}+\bar \H_{}}\otimes q_{}^{\H_{}+\bar \H_{}}\bar \E_{+}\big)\,=\,\exp_{q_{}^{2}}^{}\big(\!\varsigma\, \E_{+}\bar \E_{+}\big)~,\end{aligned}$$ are given (as results from $q$-Hadamard formula) below $$\begin{aligned} \label{BD17} S_{4_{}}(q_{}^{\pm(\H_{}-\bar \H)})\!\!&=\!\!&q_{}^{\mp(\H_{}-\bar\H)}~,\qquad S_{4_{}}(\E_{k+})\!\!=\!\!-q_{}^{(-)^k}\E_{k+} \,,\nonumber \\[7pt]\label{BD18} S_{4_{}}(\,q_{}^{\H_{}+\bar \H})\!\!&=\!\!&q_{}^{-\H_{}-\bar \H}X^{-1}~,\quad S_{4_{}}(q_{}^{\;-\H_{}-\bar \H})\;=\;X\,q_{}^{\H_{}+\bar \H}~, \\[7pt]\label{BD19} S_{4_{}}(\E_{k-})\!\!&=\!\!&-q_{}^{(-1)^{k+1}}\E_{k-}+\frac{(-)^k\varsigma}{q_{}^{2(-)^k}-1}\, \E_{(k+1)+}\bigl(q_{}^{2\H_k}-q_{}^{-2\H_k}X^{-1}\bigr)~,\nonumber $$ where $$\begin{aligned} \label{BD22} X\!\!&:=\!\!&1+\varsigma(q_{}^{2}-1)\E_{+}\bar \E_{+}~.\end{aligned}$$ Therefore a total universal $R$-matrix for this case is the following product (now $\bar\gamma=-\gamma \Leftrightarrow\bar q=q^{-1}$) $$\begin{aligned} \label{ct12b} R_{4_{}}(\gamma,\varsigma)\!\!&=\!\!&F^\tau_{4_{}''}(\gamma,\varsigma)R_{3'0}(\gamma)R_{3'1}(-\gamma) F^{-1}_{4_{}''}(\gamma,\varsigma)~. $$ Two real qunatizations are described in the Table \[tab:rr4\]. -------------------------------- --------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------------------- --- \[-8pt\] $\mathfrak{o}''(2,2)$     $\gamma,\,\varsigma\in\imath\mathbb{R}\ \ \ $ $(q^\H)^{\star}=q^\H, \E_{\pm}^{\star}=-\E_{\pm}$ $(\bar q^{\bar\H})^{\star}=\bar q^{\bar\H} , \bar\E_{\pm}^{\star}=-\bar\E_{\pm}$ A \[-8pt\] $\mathfrak{o}(3,1)$ $\ \ \ \ \gamma\in\mathbb{R}\ \ ,\ \varsigma=0 $ $(q^\H)^{\ddagger} =q^{-\bar\H} , \E_{\pm}^{\ddagger}=-\bar\E_{\pm}$ $(q^{\bar\H})^{\ddagger}=q^{-\H} , \bar\E_{\pm}^{\ddagger}=-\E_{\pm}$ A -------------------------------- --------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------------------- --- : Real quantizations of $r_{IV}(\gamma,\varsigma)= \gamma\left(E_{+}\wedge E_{-} -\bar{E}_{+}\wedge\bar{E}_{-} -2H\wedge\bar{H}\right)+\varsigma E_{+}\wedge\bar{E}_{+} $[]{data-label="tab:rr4"} It should be noted that the value $\varsigma=0$ in the Lorentzian case is due to the fact that twist (\[BD7\]), in contrast to the $\mathfrak{o}''(2,2)$ case where $|q|=1$, is not unitary for real $q$. In order to have formulae (\[BD8\])– (\[BD17\]) compatible with the Lorentzian conjugation (\[rf4\]) it is helpful to introduce flipped conjugation (\[d-1f\]) on the tensor product of quantized algebras (see [@BLT08]). Alternatively, one can keep the standard (non-flipped) conjugation and seek for the unitarizing coboundary twist - the quantum analog of (\[Atwist2\]) [^30]. Examples of quantum coboundary twists can be found e.g. in [@Sam06]. The realization of this task is postponed to our future work. Yet another method relying on quantum deformation of the real involution ($\star$-involution) has been studded in [@Lyub90; @Maj95; @Osei]. Assuming $q$ real, the quantum twist (\[BD7\]) in the real Hopf algebra $(U_{q_{}}(\mathfrak{sl}(2;\mathbb{C}))\otimes U_{ q^{-1}}(\mathfrak{sl}(2;\mathbb{C}), \Delta_{4'}, S_{4'}, \ddag)$ satisfies the condition (see [@Maj95], Prop. 2.3.7, p.59) $$(S_{4'}^{}\otimes S_{4'})(F_{4''}^{\ddag\otimes\ddag})=F_{4''}^\tau \nonumber$$ for $\varsigma$ real. This permits to introduce new conjugation quantum-deformed by the similarity transformation $$()^{\ddag^\prime} = u\,()^\ddag\,u^{-1} \nonumber$$ where $u^{-1}$ is given by the formula (\[BD16\]) (in our case $S^{-1}(u)=u^{}$). Explicit calculations with the help of $q$-Hadamard formula leads to the following results $$\begin{aligned} \label{BDstar} (q_{}^{\pm(\H_{}-\bar \H)})^{\ddag'}\!\!&=\!\!&(q_{}^{\pm(\H_{}-\bar \H)})^{\ddag}=q_{}^{\pm(\H_{}-\bar\H)}~,\qquad (\E_{k+})^{\ddag'}=(\E_{k+})^{\ddag}=-\E_{(k+1)+} \,,\nonumber \\[7pt] (\,q_{}^{\H_{}+\bar \H})^{\ddag'}\!\!&=\!\!&q_{}^{-\H_{}-\bar \H}X^{-1}~,\quad (q_{}^{\;-\H_{}-\bar \H})^{\ddag'}\;=\;X^{}\,q_{}^{\H_{}+\bar \H}~, \\[7pt] (\E_{k-})^{\ddag'}\!\!&=\!\!&-\E_{(k+1)-}+\frac{\varsigma}{q^{}-q^{-1}}\, \E_{k+}\bigl(q_{}^{2\H_{k+1}}-q_{}^{-2\H_{k+1}}X^{-1}\bigr)~,\nonumber $$ where $X$ is given by (\[BD22\]). In this way Belavin-Drinfeld type quantum deformation of the Lorentz algebra is described by the real Hopf algebra $(U_{q_{}}(\mathfrak{sl}(2;\mathbb{C}))\otimes U_{ q^{-1}}(\mathfrak{sl}(2;\mathbb{C})[[\varsigma]], \Delta_{4}, S_{4}, \ddag')$. Left $q$-analog and right Jordanian deformation intertwined by Abelian twist ($r$-matrix $r_{V}$) -------------------------------------------------------------------------------------------------- In this case we start with the left sector as q-deformed with $q=\exp{{1\over 2}\gamma}$ $$\begin{aligned} \label{BD} q_{}^{\H}\E_{\pm}\!\!&=\!\!&q_{}^{\pm1}\E_{\pm}\,q_{}^{\H}~,\qquad [\E_{+},\,\E_{-}]\;=\;\frac{q_{}^{2\H}-q_{}^{-2\H}} {q_{}^{}-q_{}^{-1}}\quad\;\; ~,\end{aligned}$$ the right sector is deformed by Jordanian twist $F_J$ expressed in undeformed CW basis (cf. Sect. 3.1.1) $$\begin{aligned} \label{pr16} [\bar{H},\,\bar{E}_{\pm}]\!\!&=\!\!&\bar{E}_{\pm}~,\quad [\bar{E}_{+},\,\bar{E}_{-}]\,=\,2\bar{H}~.\end{aligned}$$ Further we perform the subsequent quantization by using the quantized Abelian twist $$F_{5''}(\bar\chi,\rho)=\tilde q^{\H\wedge\bar\Sigma} ,\qquad \tilde q=\exp{{\rho\over 4\bar\chi}}$$ The explicit coproduct formuale are the following $$\begin{aligned} \label{ct29b} \Delta_{5_{}^{}}(q_{}^{\pm \H_{}})\!\!&=\!\!&q_{}^{\pm \H_{}}\otimes q_{}^{\pm \H_{}},\\[7pt]\label{ct30b} \Delta_{5_{}^{}}(\E_{\pm})\!\!&=\!\!&\E_{\pm}\otimes q_{}^{\H_{}} \tilde q_{}^{\pm \bar\Sigma}+q_{}^{-\H_{}}\tilde q_{}^{\mp \bar\Sigma }\otimes \E_{\pm}~.\end{aligned}$$ $$\begin{aligned} \label{blte21} \Delta_{5}(\bar E_{+}) &=&\bar E_{+}\otimes e^{\bar\Sigma}+1\otimes \bar E_{+} \\\nonumber && \\\nonumber \Delta_{5}(\bar H) &= & \bar H\otimes e^{-\bar\Sigma}+1\otimes \bar H\, - \frac{\rho}{4}\,\left(\H\otimes \bar E_{+} e^{-\bar\Sigma}-\, \bar E_{+} e^{-\bar\Sigma}\otimes \H e^{-\bar\Sigma}\right) \\\nonumber && \\\nonumber \Delta_{5}(\bar E_{-})&=&\bar E_{-}\otimes e^{-\bar\Sigma}+1\otimes \bar E_{-} +2\bar\chi\bar H\otimes \bar H e^{-\bar\Sigma} +\bar\chi\bar H(\bar H-1)\otimes\bar\Lambda+ \nonumber \\[8pt]&& \,-\frac{\rho}{2}\,\left( \bar H e^{-\bar\Sigma}\otimes\H e^{-\bar\Sigma} - \bar H\H \otimes\bar\Lambda -\H \otimes \bar H e^{-\bar\Sigma}\right) \nonumber \\[8pt]&& \,-\frac{\rho}{2}\, \left(\, \bar\Lambda e^{\bar\Sigma}\otimes \bar H\H e^{-\bar\Sigma}\,+ \bar H\bar\Lambda e^{\bar\Sigma}\otimes\H \bar\Lambda\, \right)\nonumber \\[8pt]&& \,-\,\frac{\rho}{4}\,\left( \left(1-e^{-2\bar\Sigma}\right)\otimes\H \bar\Lambda\, +\H \otimes\bar\Lambda\,-\bar\Lambda\otimes \H e^{-\bar\Sigma}\right) \nonumber\\[8pt] && {1\over\bar\chi}\left(\frac{\rho}{4}\right)^2\,\left(\bar{\Lambda}^2 e^{2\bar\Sigma}\otimes\H^2 \bar\Lambda +\bar\Lambda\otimes\H^2 e^{-\bar\Sigma} +\H^2\otimes\bar\Lambda \right) \nonumber\\[8pt]&& - {2\over\bar\chi}\left(\frac{\rho}{4} \right)^2\,\bar\Lambda \H e^{\bar\Sigma}\otimes\H\bar\Lambda\,\end{aligned}$$ The antipodes do not depend on the Abelian twist and look as follows: $$\begin{aligned} \label{blte22b } S_{5_{}}^{}(q_{}^{\pm \H_{}})&=&q_{}^{\mp \H_{}}~,\qquad S_{5_{}}^{}(\E_{\pm})\;=\;-q_{}^{\pm1}\E_{\pm}~,\nonumber\\[8pt] S_{5}(\bar E_{+})&=&-\bar E_{+}\,e^{-\bar\Sigma} ,\qquad\qquad S_{5}(\bar H)=-\bar He^{\bar\Sigma}\\[8pt] S_{5}(\bar E_{-})&=&-\bar E_{-}\,e^{\bar\Sigma} +\bar\chi \bar{H}^2 e^{\bar\Sigma}(e^{\bar\Sigma}+1) -\bar{\chi}^2 \bar H \bar E_{+}e^{\bar\Sigma} .\nonumber \end{aligned}$$ Quantum universal $R$-matrix generated from $r_V$ takes the following form $$\label{R5} R_5(\gamma,\bar\chi,\rho)= \tilde q^{\bar\Sigma\wedge \H} R_{3'0}(\gamma)F^\tau_{J1}(\bar\chi)F^{-1 }_{J1}(\bar\chi) \tilde q^{\bar\Sigma\wedge \H}.$$ Three real quantizations we describe in the Table \[tab:rr5\] below. ------------------------------------- ------------------------------------------------------------ --------------------------------------------------- ---------------------------------------------------------------- --- \[-8pt\] $\mathfrak{o'}^{\star}(4)$ $\gamma,\rho\in\mathbb{R}$ ; $\bar\chi\in\imath\mathbb{R}$ $(q^\H)^{\dag}=q^\H, \E_{\pm}^{\dag}=\E_{\mp}$ $\bar H^{\star}= -\bar H , \bar E_{\pm}^{\star}=-\bar E_{\pm}$ R \[-8pt\] $\mathfrak{o}'(2,2)$ $\gamma,\rho\in\mathbb{R}$ ; $\bar\chi\in\imath\mathbb{R}$ $(q^\H)^{\#}=q^\H, \E_{\pm}^{\#}=-\E_{\mp}$ $\bar H^{\star}= -\bar H , \bar E_{\pm}^{\star}=-\bar E_{\pm}$ R \[-8pt\] $\mathfrak{o}''(2,2)$ $\gamma,\rho, \bar\chi\in\imath\mathbb{R}$ $(q^\H)^{\star}=q^\H, \E_{\pm}^{\star}=-\E_{\pm}$ $\bar{H}^{\star}=-\bar H , \bar E_{\pm}^{\star}=-\bar E_{\pm}$ A ------------------------------------- ------------------------------------------------------------ --------------------------------------------------- ---------------------------------------------------------------- --- : Real quantizations of $r_{V}(\gamma, \rho, \bar\chi)= \gamma\,E_{+}\wedge E_{-} +\bar{\chi}\,\bar{E}_{+}\wedge\bar{H}+\rho H\wedge\bar{E}_{+}$[]{data-label="tab:rr5"} Concluding remarks and outlook ============================== In this paper we presented the complete set of Hopf-algebraic quantum deformations generated by classical $r$-matrices for $\mathfrak{o}(4;\mathbb{C})$ and its real forms given in [@BLTnov15; @BLT17]. The explicit formulae describing algebraic and coalgebraic sectors are provided as well as there are given the universal $R$-matrices which permits the tensoring of quantum modules (representations of quantum-deformed Hopf algebras). We recall that the universal $R$-matrices describe the braided structure of quantum-covariant tensor products of modules [@Pod92; @Maj95] what has been used in quantum-covariant NC field theory [@FW07; @JLMW]. For quantum twist deformations of enveloping Lie algebras $U(\mathfrak{g})$ ($\mathfrak{g}=\mathfrak{o}(4;\mathbb{C}), \mathfrak{o}(4-k,k)$ ($k=0,1,2$), and $\mathfrak{o}^*(4)=\mathfrak{0}(2;\mathbb{H})$); for $\mathfrak{o}(4;\mathbb{C})$ (see Sect.5.1, 5.2) the algebra of quantum modules, describing e.g. NC quantum fields, can be represented by the functions of classical (commutative) fields with twist-dependent nonlocal star product multiplication rule [@Bloh03; @Kul05]. Basic role plays in relativistic physics $D=4$ Minkowski space $\mathbb{R}^{3,1}$, with signature $(+,+,+,-)$, and its Lorentz rotations $\mathfrak{o}(3,1)$. In order to describe the Lie algebra generating relativistic group of motion, one adds four generators $P_\mu\in \mathbb{T}^{3,1}$ of translations, i.e. one extends Lorentz algebra $\mathfrak{o}(3,1)$ to $D=4$ Poincaré algebra $\mathfrak{o}(3,1)\ltimes \mathbb{T}^{3,1}$. It is known that only two out of four quantum deformations of $\mathfrak{o}(3,1)$ can be extended to quantum deformations of $D=4$ Poincaré algebra (see [@Zak94; @Zak97; @Tol07; @BLT08]). The studies providing the complete list of possible quantum deformations of inhomogeneous $D=4$ Euclidean $\mathfrak{o}(4)\ltimes \mathbb{T}^{4}$ algebra and of inhomogeneous $D=4$ Kleinian $\mathfrak{o}(2,2)\ltimes \mathbb{T}^{2,2}$ algebra has not been presented [^31].We add that inhomogeneous extension of quaternionic real form $\mathfrak{o}^*(4)\equiv \mathfrak{o}(2;\mathbb{H})$ of $\mathfrak{o}(4; \mathbb{C})$ contains four complex or two quaternionic translations and its applications to the description of physical symmetries are, according to our knowledge, not known. The real forms of considered quantum groups describe the quantum symmetries of $D=3$ compact Euclidean ($S^3$), de Sitter ($dS_3$) or anti-de-Sitter ($AdS_3$) spacetimes, with finite nonvanishing constant curvature and curved $D=3$ Euclidean, dS$_3$ or AdS$_3$ translations. In $D=4$ rotation algebras $\mathfrak{o}(3,1)$ ($\mathfrak{o}(2,2)$), the dS (AdS) radius $\mathcal{R}$ is introduced by suitable rescaling of the generators in the coset $\frac{\mathfrak{o}(4-k,k)}{\mathfrak{o}(3-k,k)}$ ($k=1,2$), with $\Lambda=\mathcal{R}^{-1}$ which can be treated as a deformation parameter. The quantum deformations of $ \mathfrak{o}(4-k,k)$ ($k=1,2$) have been extensively studied as describing the NC geometry of $2+1$-dimensional QG with cosmological constant $\Lambda\neq 0$ [@Witten88; @Ball14]. The classical action of $D=3$ gravity can be introduced geometrically as gauge theory described by $D=3$ Chern-Simons (CS) model. Following Fock-Rosly construction [@Fock99; @Alek95], in such framework we describe gravitational degrees of freedom as parameterizing the Poisson-Lie group manifold. If we search for quantum deformations of Fock-Rosly construction, it appears that only classical $r$-matrices obtained from Drinfeld double (DD) structures [@Dr2] are allowed [@BS09; @BS10]. The DD structures and corresponding classical $r$-matrices for $\mathfrak{o}(3,1)$ and $\mathfrak{o}(2,2)$ algebras were recently constructed and classified [@Ball14]. We see that such quantum deformation which are well adjusted to the description of quantum-deformed $D=3$ gravity are generated by a subclass of classical $r$-matrices, listed in [@BLTnov15; @BLT17] and quantized in this paper. The next step in our program is to construct the complete list of classical $r$-matrices for the $D=4$ complex inhomogeneous Euclidean algebra $\mathcal{E}(4;\mathbb{C}):=\mathfrak{io}(4;\mathbb{C}):=\mathfrak{o}(4;\mathbb{C})\ltimes\mathbf{T}(4;\mathbb{C})$ (orthogonal rotations together with translations) and for its real forms, in particular $\mathfrak{o}(4-k,k)\ltimes\mathbf{T}(4-k,k;\mathbb{R})$ ($k=0,1,2$). Until present time the most complete results were obtained for $\mathfrak{o}(3,1)\ltimes\mathbf{T}(3,1)$ by Zakrzewski [@Zak97], who provided almost complete list of 21 different, not related by Poincaré automorphism real $D=4$ Poincaré $r$-matrices (see also [@Tol07; @BP14]). It should be noticed that the complete classifications of r-matrices for both inhomogenous $D=3$ Poincaré and $D=3$ Euclidean algebras have been given by Stachura [@Stach98]. Recently in [@BLMT12; @BLMT11] the present authors complexified Zakrzewski results and then imposed $D=4$ Euclidean reality constraints. It appeared that 8 out of 21 complexified Zakrzewski $r$-matrices are consistent with the Euclidean conjugation in $\mathfrak{o}(4;\mathbb{C})$ (see (\[rf1\])). It can be shown, however, that the complexified Zakrzewski $r$-matrices do not describe all $r$-matrices for $\mathcal{E}(4;\mathbb{C})$[^32]. Using the constructive method analogous to the one proposed in this paper we intend to describe the complete classification of classical $r$-matrices for $D=4$ complex inhomogeneous Lie algebra $\mathcal{E}(4;\mathbb{C})$ and for its all real forms. We add that in [@BLMT12; @BLMT11] we considered also the $N=1$ superextension of Poincaré and Euclidean classical $r$-matrices. Recently we derived in analogous way as well new class of $N=2$ Poincaré and Euclidean supersymmetric $r$-matrices (see [@BLT15]). We hope that our constructive method of providing the complete list of classical $r$-matrices for the complex $\mathcal{E}(4;\mathbb{C})$ case can be applied as well to $N$-extended Euclidean superalgebras $\mathcal{E}(4|N;\mathbb{C})$ for $N=1,2,4$ and further classify and quantize the supersymmetric $r$-matrices for the corresponding real forms. Acknowledgments {#acknowledgments .unnumbered} --------------- This work has been supported by Polish National Science Center (NCN), project 2014/13/B/ST2/04043 (A.B. and J.L.) and by COST (European Cooperation in Science and Technology) Action MP1405 QSPACE. V.N.T. was supported by RFBR grant No.14-01-00474-a. All $\mathfrak{o}(3)$ and $\mathfrak{o}(2,1)$ Lie bialgebras ============================================================ Classification of r-matrices is the same task as classification of coboundary Lie bialgebras up to isomorphisms (an isomorphism which preserves the structure constants is called an automorphism). In geometric terms they can be seen as orbits of an action of the Lie algebra automorphism group in the space of skew-symmetric solution of mCYBE. For simple algebras all bialgebra structures are coboundary due to the Whitehead lemma. Let us consider, for completeness as well as for pedagogical reason, geometric classification scheme for classical r-matrices of simple 3-dimensional real rotational Lie algebras (for purely algebraic approach see [@LT17]). Up to an isomorphism there are only two non-isomorphic real simple Lie algebras: compact $\mathfrak{o}(3)$ and non-compact $\mathfrak{o}(2,1)$, both are real form of $\mathfrak{o}(3;\mathbb{C})$. Consider firstly the compact $\mathfrak{o}(3)$ case with the canonical vectorial basis ($I_k^\dag=-I_k$ cf. (\[jl2\]) – (\[jl3b\])) $$\label{a1} [I_1, I_2]=I_3, \quad [I_1, I_3]= - I_2, \quad [I_2, I_3]=I_1$$ We notice that any element $r(a,b,c)= a I_2\wedge I_3+ b I_3\wedge I_1+c I_1\wedge I_2 \in \mathfrak{o}(3)\wedge\mathfrak{o}(3)$ is a classical r-matrix since it satisfies $$\label{a2} [[r(a,b,c), r(a,b,c)]]= (a^2+b^2+c^2)\Omega$$ where $\Omega=I_1\wedge I_2\wedge I_3\in \mathfrak{o}(3)\wedge\mathfrak{o}(3)\wedge\mathfrak{o}(3)$ is a unique up to the constant invariant element. The non-isomorphic Lie bialgebra structures for $\mathfrak{o}(3)$ case can be identify with orbits of the automorphism group in the space of free parameters $(a,b,c)\in \mathbb{R}^3$ with the Euclidean metric. The group of automorphisms contain $SO(3)$ subgroup. Moreover, bivector and vector representations are equivalent in dimension 3. Due to this property we look only for $SO(3)$- orbits in $\mathbb{R}^3$. These are the 2-dimensional spheres represented by a radius $\xi>0$ or by the vector $\xi(1,0,0)$. Thus as a result of final classification one gets the following family of non-trivial $\mathfrak{o}(3)$ r-matrices (the trivial $r=0$ $r$-matrix corresponds to singular one-point orbit at $(0,0,0)$) $$\label{a3} r_\xi= \xi I_1\wedge I_2$$ Notice that the values of the real parameter $\xi>0$ are effective and lead to nonequivalent Lie bialgebra structures. Similar analysis applied to the non-compact real form $\mathfrak{o}(2,1)$ provides qualitatively different results. Arbitrary $\mathfrak{o}(2,1)$ r-matrix satisfies the following YB equation (a,b,c real) $$\label{a4} [[r(a,b,c), r(a,b,c)]]= (a^2-b^2 +c^2) J_1\wedge J_2\wedge J_3$$ where $r(a,b,c)= a J_2\wedge J_3+ b J_3\wedge J_1+c J_1\wedge J_2 \in \mathfrak{o}(2,1)\wedge\mathfrak{o}(2,1)$ is written in the canonical $\mathfrak{o}(2,1)$ basis. We choose noncompact vectorial generators $J_1, J_2, J_3$ following our choice of $\mathfrak{o}(2,1)$ reality conditions $J_k^\star=-J_k$ $$\label{a5} [J_1, J_2]=J_3, \quad [J_1, J_3]= J_2, \quad [J_2, J_3]= J_1$$ The automorphisms group $SO(2,1)$ of the Lie algebra $\mathfrak{o}(2,1)$ acts in three-dimensional Minkowski space $\mathbb{R}^{2,1}$. There are three types of non-trivial orbits in $\mathbb{R}^{2,1}=\{(a,b,c): a,b,c\in\mathbb{R}\}$, characterizing three independent $\mathfrak{o}(2,1)$ r-matrices. 1\. single light-cone orbit represented by the light-like vector $\xi(1,1,0)$ which provides the solution of homogeneous classical YB equation (CYBE) 2\. one-parameter family of space-like orbits represented by space-like vectors\ $\xi (1,0,0), \xi\neq 0$ (solution of modified CYBE) 3\. one-parameter families of time-like orbits represented by time-like vectors\ $\xi (0,1,0), \xi\neq 0 $ (solution of modified CYBE) Three canonical $\mathfrak{o}(2,1)$ r-matrices corresponding to three types of orbits take the form $$\label{a6} r^{0}_\xi= \xi (J_1\wedge J_3+ J_1\wedge J_2); \quad r^{-}_\xi= \xi J_3\wedge J_2; \quad r^{+}_\xi= \xi J_1\wedge J_3$$ where in the first case one gets the same deformation for any value of the parameter $\xi\neq 0$, while in the remaining two cases different values of $\xi$ lead to different Lie bialgebras. In this setting we get one reality condition and three different types of $r$-matrices representing nonequivalent bialgebra structures. $q$-exponent and $q$-Hadamard formula ===================================== Our aim here is to introduce some formulas (mainly concerning a $q$-deformed Hadamard lemma), which were main tools for calculations presented in Sect. 5.4. Let $A$ and $B$ be two arbitrary elements of some quantum algebra and let $\exp_{q}(A)$ be a formal $q$-exponential $$\begin{aligned} \label{q-exp} \exp_{q}(A)\!\!&:=\!\!&\sum_{n\geq0}\,\frac{A^n}{(n)_{q}^{}!}~, \quad\;(n)_{q}^{}!:=(1)_{q}^{}(2)_{q}^{}\cdots (n)_{q}^{},\quad(n)_{q}^{}=\frac{1-q^n}{1-q}~.\end{aligned}$$ of the element $A$. As the $q$-exponential $\exp_{q^{-1}}(-A)$ is inverse to $\exp_{q}(A)$, i.e. $\bigl(\exp_{q}(A)\bigr)^{-1}= \exp_{q^{-1}}(-A)$ thus the $q$-analog of Hadamard formula can be obtained as follows (see [@KhTo1]) $$\begin{aligned} \label{A1} \begin{array}{rcl} \exp_{q}(A)\,B\bigl(\exp_{q}(A)\bigr)^{-1}\!\!&=\!\!&\exp_{q}(A)\,B\exp_{q^{-1}}(-A)\; \equiv\;\bigl(\mathop{\rm Ad}\exp_{q}(A)\bigr)(B)\;=\; \\[10pt] \!\!&=\!\!&\displaystyle\Bigl(\sum_{n\geq 0}\frac{1}{(n)_{q}!}(\mathop{{\rm ad}_q}A)^n\Bigr)(B)\;=\;\bigl(\exp_{q}(\mathop{{\rm ad}_q}A)\bigr)(B)~, \end{array}\end{aligned}$$ where the $q$-adjoint action is defined by means of $q$-brackets ($[C,\,D]_{q'}\;\equiv\;CD-q'DC$ ): $$\begin{aligned} \label{B2} \begin{array}{rcl} (\mathop{{\rm ad}_q}A)^0(B)\!\!&\equiv\!\!&B~,\quad(\mathop{{\rm ad}_q}(A))^1(B)\;\equiv \;[A,\,B]~,\quad(\mathop{{\rm ad}_q}(A))^2(B)\;\equiv\;[A,\,[A,\,B]]_{q}~, \\[10pt] (\mathop{{\rm ad}_q}(A))^3(B)\!\!&\equiv\!\!&[A,\,[A,\,[A,\,B]]_{q}]_{q^2}~,\ldots, (\mathop{{\rm ad}_q}(A))^{n+1}(B)\;=\;[A,\,(\mathop{{\rm ad}_q}(A))^{n}(B)]_{q^n}~. \end{array}\end{aligned}$$ Consider the spacial case ($q'\neq q$ in general) $$\begin{aligned} \label{A3} [A,\,B]_{q'}\!\!&=\!\!&0~,\end{aligned}$$ one gets $$\begin{aligned} \label{A4a} \begin{array}{rcl} (\mathop{{\rm ad}_q}(A))^{n+1}(B)\!\!&=\!\!&\displaystyle(1-q'^{-1}q^{n})A\bigl(\mathop{{\rm ad}_q}(A)\bigr)^{n}(B)\;=\;\prod_{k=0}^{n}(1-q'^{-1}q^{k})A^{n}B \\[10pt] \!\!&=\!\!&(q'^{-1};q)_{n}\,A^{n}B~. \end{array}\end{aligned}$$ using the standard notation $(a;q)_n$ from the theory of basic hypergeometric series (see e.g. [@Klimyk; @GaRa]). Substituting (\[A4a\]) in (\[A1\]) we obtain $$\begin{aligned} \label{A5} \begin{array} {rcl} \exp_{q}(A)\,B\bigl(\exp_{q}(A)\bigr)^{-1}\!\!&=\!\!&\displaystyle\Bigl( \sum_{n=0}^{\infty} \frac{(q'^{-1};q)_{n}}{(q;q)_n}\,(1-q)^nA^{n}\Bigr)B\;=\; \\[10pt] \!\!&=\!\!&\displaystyle{}_1\phi_{0}(q'^{-1};-;q;(1-q)A)\,B\;=\;\frac{\bigl(q'^{-1} (1-q)A;q\Bigr)_{\infty}}{((1-q)A;q)_{\infty}}\,B~, \end{array}\end{aligned}$$ as the result of the $q$-binomial theorem (see [@Klimyk; @GaRa]). In the particular case $q'=q^{n}$, $n=0,1,2,\ldots,$ the formula (\[A1\]) reads $$\begin{aligned} \label{A7} \begin{array}{rcl} \exp_{q}(A)\,B(\exp_{q}(A))^{-1}\!\!&=\!\!&\bigl(q^{-n}(1-q)A;q\bigr)_{n}\,B \\[10pt] \!\!&=\!\!&q^{-n(n+1)/2}\bigl((q-1)A\bigr)^{n}\,\bigl(q/(1-q)A;q\bigr)_{n}\,B~. \end{array}\end{aligned}$$ In the case $q'=q^{-n}$, $n=1,2,\ldots,$ for (\[A1\]) we have $$\begin{aligned} \label{A8} \exp_{q}(A)\,B(\exp_{q}(A))^{-1}\!\!&=\!\!&\bigl((1-q)A;q\bigr)_{n}^{-1}\,B~.\end{aligned}$$ Adopting to the situation in Sect. 5.4 one has to substitute $A\rightarrow\mathbb{A}=\varsigma \E_{+}q_{}^{\H_{}+\bar \H_{}}\otimes q_{}^{\H_{}+\bar \H_{}}\bar \E_{+}$ or $A=\varsigma \E_{+}\bar \E_{+}$ and $n=1,2$ (cf. (\[BD15\]) or (\[BD17\])). [99]{} S. Majid, J. Class. Quant. Grav. **5**, 1587 (1988). S. Doplicher, K. Fredenhagen, J.E. Roberts, Commun. Math. Phys. **172**, 187 (1995); arXiv:hep-th/0303037. J.L. Garay, Int. Jour. Math. Phys. **A10**, 145 (1995); arXiv:gr-qc/9403008. A. Ashtekar, J. Lewandowski, Class.Quant. Grav.**21**, 1253 (2004); arXiv:gr-qc/040418. D. Kaminski, [*Algebras of Quantum Variables for Loop Quantum Gravity, I. Overview*]{}; arXiv:1108.4577. T. Thiemann, [*Modern Canonical Quantum General relativity*]{}, Cambridge Univ. Press, 2007 F. Cianfrani, J. Kowalski-Glikman, D. Pranzetti, G. Rosati, Phys. Rev. **D 94**, 084044 (2016), arXiv:1606.03085. J. Lukierski, H. Ruegg, A. Nowicki, V. N. Tolstoy, Phys. Lett. **B264** 331 (1991). G. Amelino-Camelia, L. Smolin, A. Starodubtsev, Class.Quant.Grav. **21** (2004) 3095; hep-th/0306134. V.G. Drinfeld, [*Quantum Groups*]{}, ed. A. Gleason, Proceedings of the ICM, Berkeley 1985, p. 798, Providence, Rhode Island, 1987. publ. AMS; V. G. Drinfeld, [*Quantum groups*]{} (Leningrad, 1990) 1, Lecture Notes in Math., 1510, Springer, Berlin, 1992. P. I. Etingof, D. A. Kazhdan, Selecta Math. (N.S.) **2** (1996) 1; arXiv:q-alg/9506005. S.L. Woronowicz, Comm. Math. Phys. **111**, (1987), 613 V. Chari, A. Pressley, [*A Guide to Quantum Groups*]{}, Cambridge Univ. Press, 1994. S. Majid, [*Foundations of Quantum Groups*]{}, Cambridge Univ. Press, 1995. A. Klimyk, K. Schmüdgen, [*Quantum Groups and Their Representations*]{}, Springer 1997. P. Etingof, O. Schiffmann, [*Lectures on quantum groups*]{}, Internationa Press 2002. A. Borowiec, J. Lukierski, V.N. Tolstoy, Phys.Lett. **B754** (2016) 176; arXiv:1511.03653\[hep-th\]. J. Lukierski, V.N. Tolstoy, Eur. Phys. J. **C77** (2017) 226; arXiv:1612.03866 \[hep-th\]. A. Borowiec, J. Lukierski, V.N. Tolstoy, Phys.Lett. **B770** (2017) 426; arXiv:1704.06852 \[hep-th\]. A.A. Belavin, V.G. Drinfeld, Funct. Anal. Appl. **16**, 1 (1982). A.A. Belavin, V.G. Drinfeld, Soviet Sci. Rev. Sect. C: Math. Phys. Rev. **4** (1984) 93–165. S.M. Khoroshkin and V.N. Tolstoy, Comm. Math. Phys. **141** 599 (1991). J. Lukierski, A. Nowicki, H. Ruegg, Phys.Lett. **B271** (1991) 321; hep-th/9108018. S.L. Woronowicz, Rep. Math. Phys. **30** (1991) 259. A. Borowiec, J. Lukierski, V.N. Tolstoy, Eur. Phys. J. **C48**, 633 (2006); arXiv:0604144\[hep-th\]. A. Borowiec, J. Lukierski, V.N. Tolstoy, Eur. Phys. J. **C57**, 601 (2008); arXiv:0804.3305\[hep-th\]. E. Celeghini, R. Giachetti, E. Sorace, M. Tarlini, J. Math. Phys., **32** 1155 (1991). C. Klimcik, JHEP **0212**, 051 (2002); arXiv:0210.095\[hep-th\]. C. Klimcik, J. Math. Phys. **50**, 043508 (2009); arXiv:0802.3518\[hep-th\]. B. Vicedo, J. Phys. **A48**, 355203 (2015); arXiv:1504.06303\[hep-th\]. T. Kawaguchi, T. Matsumoto, K. Yoshida, JHEP **1404**, 153 (2014); arXiv:1401.4855\[hep-th\]; see also JHEP **1406**, 146 (2014); arXiv:1402.6147\[hep-th\]. T. Matsumoto, K. Yoshida, J. Phys.: Conf. Ser. **56**, 012020 (2015); arXiv:1410.0575\[hep-th\]. T. Matsumoto, K. Yoshida, Nucl. Phys. **B893**, 287 (2015); arXiv:1501.03665. S.J. van Tongeren, JHEP **1506**, 048 (2015); arXiv:1504.05516\[hep-th\]; see also Nucl.Phys. **B904** (2016) 148; arXiv:1506.01023\[hep-th\]. A. Pachol, S.J. van Tongeren, Phys.Rev. **D93** (2016) 026008; arXiv:1510.02389\[hep-th\]. A. Borowiec, H. Kyono, J. Lukierski, J. Sakamoto, K. Yoshida, JHEP **1604** (2016) 079; arXiv:1510.03083\[hep-th\]. V. G. Drinfeld, Sov. Math. Dokl. **27** (1983)68. M. Blaszak, Physica, **198A** (1993) 637. M. A. Semenov-Tyan-Shanski, [*Integrable Systems and Factorization Problems*]{}, arXiv:nl/0209051. I. Ya. Dorfman, A.S. Fokas, J. Math. Phys. **33** (1997) 2504. A.S. Fokas, I.M. Gelfand, [*Algebraic Aspects of Integrable Systems*]{}, Birkhauser 1997. R. Fioresi, E. Latini, A. Marrani, [*Quantum Klein Space and Superspace*]{}, arXiv:1705.01755. N. Beisert, R. Hecht, B. Hoare, J.Phys. **A50** (2017), 314003; arXiv:1704.05093 \[math-ph\]. M.A. Semenov-Tian-Shansky, Funct. Anal. Appl. **17**, 289 (1983). V.G. Drinfeld, Algebra i Analiz **1** (1989) 114; translation in Leningrad Math. J., **1**, 1419 (1990). P. Kulish, [*Twist deformations of quantum integrable spin chains*]{}, in P. Aschieri, M. Dimitrijevic, P. Kulish, F. Lizzi, J. Wess (Eds.), [*Noncommutative spacetimes. Symmetries in noncommutative geometry and field theory*]{}, Lecture Notes in Physics, 774. Springer-Verlag, Berlin, 2009, p.165. O.V. Ogievetsky, [*Hopf structures on the Borel subalgebra of $sl(2)$*]{}, in Proc. Winter School [*Geometry and Physics*]{}, Zidkov, January 2013, Czech Republic, Rendiconti Circ. Math. Palermo, Serie II **37** (1993) 185; Max Planck Int. prepr. MPI-Ph/92-99. G.W. Delius, A. Hueffmann, J.Phys. **A29**:1703 (1996); arXiv:q-alg/9506017 P. Aschieri, A. Borowiec, A. Pachol, [*Observables and Dispersion Relations in k-Minkowski Spacetime*]{}; arXiv:1703.08726 P.P. Kulish, A.I. Mudrov, Proc. Steklov Inst. Math. **226**, 97 (1999), arXiv:math.QA/9901019 A.P. Isaev, O.V. Ogievetsky, Phys. Atomic Nuclei **64** (2001), 2126; translated from Yadernaya Fiz. **64** (2001) 2216. A. Borowiec, J. Lukierski, V.N. Tolstoy, Modern Physics Letters **A18**, (2003) 1157; hep-th/0301033. M. Samsonov, Lett. Math. Phys. **75** (2006), 63; and Lett. Math. Phys. **72** (2005) 197. V.V. Lyubashenko, [*Real and imaginary forms of Quantum groups*]{}, Proc. of the Euler Institute, St. Petersburg, 1990, Lec. Notes Math., 1510, p. 67, Springer. S. Majid, P.K. Osei, [*Quasitriangular structure and twisting of the 2+1 bicrossproduct model*]{}, arXiv:1708.07999. P. Podles, Adv. Ser. Math. Phys. **16**, 805 (1992). G. Fiore, J. Wess, Phys.Rev. **D75** 105022 (2007). J. Lukierski, M. Woronowicz, J. Phys. **A45** 215402 (2012), arXiv:1105.3612; J. Mod. Phys. **A27** 1250084, arXiv:1206.5656. C. Blohmann, J. Math. Phys. **44** 4736 (2003). P. Kulish, Proc. of Karlstad Conf. AMS, Contemp. Math. **391** 213 (2005), arXiv:0606056\[hep-th\]. S. Zakrzewski, Lett. Math. Phys. **32**, 11 (1994). S. Zakrzewski, Comm. Math. Phys. **185** (1997) 285; q-alg/9602001. V.N. Tolstoy, Bulg. J. Phys. **35**, 441 (2008) (Conference: C07-06-18.13 Proceedings); arXiv:0712.3962. A. Borowiec, A. Pachol, SIGMA **10** (2014) 107; arXiv:1404.2916 A. Borowiec, J. Lukierski, V.N. Tolstoy, 23rd International Conference on Integrable Systems and Quantum Symmetries (ISQS-23), J.Phys.Conf.Ser. **670** (2016), 012013; arXiv:1510.09125\[hep-th\]. E. Witten, Nucl. Phys. **B311** 46 (1988). A. Ballesteros, F.J. Herranz, C. Meusburger, Phys. Lett. **B687** 375 (2010), arXiv:1001.4229; Class. Quantum. Grav. **30** 155012 (2013), arXiv:1303.3080; Phys. Lett. **B732** 201 (2014), arXiv:1402.2884. V.V. Fock, A.A. Rosly, Amer. Math. Soc. Trans. **191** 67 (1999). A.Y. Alekseev, A.Z. Malkin, Comm. Math. Phys. **169** 99 (1995). C. Meusburger, B. J. Schroers, Nucl.Phys. **B806** (2009) 462; arXiv:0805.3318 \[gr-qc\]. G. Papageorgiou, B. J. Schroers, JHEP **1011** (2010) 020; arXiv:1008.0279 \[hep-th\], ibid. JHEP **0911** (2009) 009; arXiv:0907.2880 \[hep-th\]. P. Stachura, J. Phys. **A 31**, no. 19, 4555 (1998). A. Borowiec, J. Lukierski, M. Mozrzymas, V.N. Tolstoy, JHEP **1206**, 154 (2012); arXiv:1112.1936\[hep-th\]. A. Borowiec, J. Lukierski, M. Mozrzymas, V.N. Tolstoy, Proc. XXIX Jnt. Coll. on Group-Theoretical Methods in Physics, Tianjin, August 2012, ed. Cheng-Ming Bai et all., World Scientific, Singapore, p.443 (2013); arXiv:1211.4546\[hep-th\]. G. Gasper and M. Rahman, [*Basic hypergeometric series*]{}, Cambridge University Press, 1990. [^1]: We stress that in LQG spacetime lattice has a dynamical origin, in particular it is not a way to regularize neither the QG action nor the QG functional integral in order to perform effectively the numerical calculations. [^2]: For classification purposes we listed in [@BLTnov15; @LT17; @BLT17] only the antisymmetric $r$-matrices. In the case of standard (or Drinfeld-Jimbo [@Dr2]) r-matrices one quantizes their symmetric Belavin-Drinfeld form [@BD82; @BD84], which satisfies CYBE and describes the leading order in the expansion of quantum $R$-matrix satisfying quantum Yang-Baxter equation [@Dr2; @ChPr94; @Maj95] For general formulae describing universal $R$-matrices see e.g. [@KhTo1]. [^3]: Only five $\mathfrak{o}(4;\mathbb{C})$ $r$-matrices are independent modulo $\mathfrak{o}(4;\mathbb{C})$ automorphism (see [@BLT17]). [^4]: With the cobracket given by the commutator $\delta_r(x)=[x\otimes 1+1\otimes x,r]$, see e.g. [@ChPr94; @Maj95; @EtSch02]. [^5]: For general elements $r_{1}, r_2\in \mathfrak{g}\wedge\mathfrak{g}$ one can extend (\[schouten\]) by bilinearity. [^6]: The symmetric part $r_s$ is $\mathfrak{g}$-invariant and in the case of semi-simple algebra is related to the so-called split Casimir (non-degenerate Cartan-Killing form). [^7]: The importance of quantum $R$-matrices follows from their applications as solutions of qYBE in various branches of theoretical physics e.g. conformal field theory, statistical mechanical models, and in mathematics, e.g. for description of link invariants. [^8]: The parameter $\xi$ in $o(\xi^2)$ should be replaced by $(\xi_1,\ldots,\xi_k)$ in the case of multiparameter deformation, i.e. for multiparameter classical $r$-matrix which is linear in $\xi_i\ (i=1,\dots,k)$ the expansion (\[i1\]) is up to any quadratic term in $\xi_i$. [^9]: They are more tractable, since $dim\,(\mathfrak{g}\wedge\mathfrak{g})<dim\,(\mathfrak{g}\otimes\mathfrak{g})$. [^10]: In a case of Hilbert space realization this condition leads to operators with imaginary spectrum. For this reason some authors do prefer instead Hermitian generators and imaginary structure constants as representing real Lie algebras. [^11]: In other words a real Hopf algebra is identified with a $\star$-Hopf algebra which is a complex Hopf algebra equipped with an additional star operation making the algebraic sector into $\star$-algebra and coalgebraic sector satisfying (\[d-1\]). [^12]: It should be observed that the presence of these two universal $R$-matrices may help to obtain finite contraction limits (see e.g. [@Beisert17]) [^13]: This condition can be rewritten as $\tilde r^{\tau(\divideontimes\otimes\divideontimes)}=\tilde r$, where $(a\otimes b)^{\tau(\divideontimes\otimes\divideontimes)}=b^\divideontimes \otimes a^\divideontimes$ denotes so-called flipped conjugation (cf. [@LNR91; @BLT08] and formula (\[d-1f\]) below). [^14]: From now on all Lie algebras and bialgebras are real if not indicated otherwise. [^15]: There are only two orbit types under the action of $\mathfrak{o}(3;\mathbb{C})$ in $\mathbb{C}^3$: null and non-null, see Appendix A. [^16]: We are working in the Cartan Weyl basis and different reality conditions, cf. [@LT17]. [^17]: Further we shall use specific notation in order to distinguish between real Lie algebras and bialgebras, e.g. $\mathsf{sl}_\gamma(2)$ denotes the triple $(\mathfrak{sl}(2;\mathbb{C}),\#, r_{st}(\gamma))$ [^18]: We remind that a coboundary twist for a given Hopf algebra is constructed out of any invertible element $W$ according to the following prescription $$F_W^{cob}=(W^{-1}\otimes W^{-1})\Delta^{}(W)$$ and leads via twisting (\[blte7\]) to the isomorphic Hopf algebras (see e.g. [@Kulish09] ). [^19]: Here $\chi$ is not an effective deformation parameter. It is a formal variable which enables to write the twist as a formal power series and an invertible element. [^20]: Non-standard, e.g. Jordanian, deformation can be also expressed with the use of nonclassical quantum Lie algebra generators [@Delius95] obtained from twist (\[j1\]) (see also [@ABP17]). [^21]: It is known that complex metric (symmetric, nondegenerate and bilinear form) has no signature [^22]: For the relation with other, physically more meaningful, Cartesian basis see e.g. (\[jl1\]) and [@BLTnov15]. [^23]: The list (\[r1\]) -(\[r5\]) is numbered in different way in comparison with original result [@BLTnov15]; the $r$-matrix $r_6$ in [@BLTnov15] from $r_5=r_V$ by involutive automorphism flipping the chiral sectors. Notation for the parameters is slightly changed as well. [^24]: The same convention will be further used below in the paper. [^25]: The normalization $\frac{\varsigma}{\chi\bar\chi}$ in the deformation parameter is necessary in order to recover correct formula in the limit $\chi, \bar\chi\mapsto 0$. [^26]: One finds $[f(E_+),H]=-E_+f'(E_+)$, $[f(E_+),E_-]=2Hf'(E_+)-E_+f''(E_+)$, where $f$ is an analytic function of one variable. In particular $[\Sigma,H]=-\chi E_+ e^{-\Sigma}=\Lambda e^{\Sigma}$, $[\Sigma,E_-]=2\chi He^{-\Sigma}-\chi\Lambda$. [^27]: Studied first time in [@BLT06]. [^28]: In fact, taking into account (\[ct13bis\]), there are four ways of describing universal $R$- matrix $R_{3'}$. [^29]: In (\[ct19\]) and in other formulas describing classical $r$-matrices, the generators $E_{\pm}$, $\bar E_{\pm}$ are not deformed. [^30]: This method has been e.g. used in [@BLT03] in order to unitarize superextension of the Jordanian deformation. [^31]: For partial results in $D=4$ Euclidean case see e.g. [@BLT15] [^32]: In particular one can easily argue observing that the list of the real $r$-matrices for $\mathfrak{o}(2,2)\ltimes\mathbf{T}(2,2)$ is longer then the Zakrzewski list (see [@Zak97]) for $D=4$ Poincaré algebra.

ArXiv

--- abstract: | The Configuration Interaction (CI) method using orbitals centered on the nucleus has recently been applied to calculate the interactions of positrons interacting with atoms. Computational investigations of the convergence properties of binding energy, phase shift and annihilation rate with respect to the maximum angular momentum of the orbital basis for the $e^+$Cu and PsH bound states, and the $e^+$-H scattering system were completed. The annihilation rates converge very slowly with angular momentum, and moreover the convergence with radial basis dimension appears to be slower for high angular momentum. A number of methods of completing the partial wave sum are compared, an approach based on a $\Delta X_J = a(J + {\scriptstyle \frac{1}{2}})^{-n} + b(J + {\scriptstyle \frac{1}{2}})^{-(n+1)}$ form (with $n = 4$ for phase shift (or energy) and $n = 2$ for the annihilation rate) seems to be preferred on considerations of utility and underlying physical justification. author: - 'J.Mitroy' - 'M.W.J.Bromley' title: 'Convergence of CI single center calculations of positron-atom interactions ' --- Introduction ============ In the last few years there have been a number of calculations of positron-binding to atoms [@mitroy99c; @dzuba99; @dzuba00a; @bromley00a; @bromley02a; @bromley02b; @bromley02d; @bromley02e; @saito03a; @saito03c; @saito05a] and positron-scattering from atoms and ions [@bromley03a; @novikov04a; @gribakin04b] using orthodox CI type methods with large basis sets of single center orbitals. Besides the calculations on atoms, there have some attempts to calculate the properties of positrons bound to molecules with modified versions of standard quantum chemistry methods [@tachikawa03a; @strasburger04a; @buenker05a]. One feature common to all these CI type calculations is the slow convergence of the binding energy, and the even slower convergence of the annihilation rate. The attractive electron-positron interaction leads to the formation of a Ps cluster (i.e. something akin to a positronium atom) in the outer valence region of the atom [@ryzhikh98e; @dzuba99; @mitroy02b; @saito03a]. The accurate representation of a Ps cluster using only single particle orbitals centered on the nucleus requires the inclusion of orbitals with much higher angular momenta than a roughly equivalent electron-only calculation [@strasburger95; @schrader98; @mitroy99c; @dzuba99]. For example, the largest CI calculations on PsH and the group II positronic atoms have typically involved single particle basis sets with 8 radial functions per angular momenta, $\ell$, and inclusion of angular momenta up to $\ell_{\rm max} = 10$ [@bromley02a; @bromley02b; @saito03a]. Even with such large basis sets, between 5-60$\%$ of the binding energy and some 30-80$\%$ of the annihilation rate were obtained by extrapolating from the $\ell_{\rm max} = 10$ to the $\ell_{\rm max} = \infty$ limit. Since our initial CI calculations on group II and IIB atoms [@bromley02a; @bromley02b; @bromley02d], advances in computer hardware mean larger dimension CI calculations are possible. In addition, program improvements have removed the chief memory bottleneck that previously constrained the size of the calculation. As a result, it is now appropriate to revisit these earlier calculation to obtain improved estimates of the positron binding energies and other expectation values. However, as the calculations are increased in size, it has become apparent that the issue of slow convergence of the physical observables with the angular momenta of the basis orbitals is the central technical issue in any calculation. Whilst it is desirable to minimize the amount of mechanical detail in any discussion (so as not to distract from the physics), the ability to draw reliable conclusions from any calculation depends crucially on the treatment of the higher partial waves. For example, the treatment of the higher partial waves in separate calculations by Saito [@saito05a] has already been shown to be flawed [@mitroy05i] while the present work exposes the defects in the methods of Gribakin and Ludlow [@gribakin04b]. The present work, therefore, is solely devoted to an in-depth examination of the convergence properties of mixed electron-positron calculations. In our previous works, [@bromley02a; @bromley02b; @bromley02d; @bromley03a; @novikov04a], a relatively simple solution to this problem was adopted. In effect, it was assumed that the successive increments to any observable scaled as an inverse power series in $\ell_{\rm max}$. This approach does have limitations as do the approaches adopted by other groups [@saito03c; @saito05a; @gribakin04b; @mitroy05i]. In the present article, we examine the convergence properties of positron binding calculations upon PsH and $e^+$Cu, and a positron scattering calculation upon the $e^+$-H system with respect to angular momentum and the dimension of the radial basis sets. The limitations of existing calculations are exhibited, and some improved prescriptions for estimating the variational limit are introduced and tested. Existing methods of performing the angular momentum extrapolation ================================================================= The nature of the problem ------------------------- The positron-atom wave function is written as a linear combination of states created by multiplying atomic states to single particle positron states with the usual Clebsch-Gordan coupling coefficients ; $$\begin{aligned} |\Psi;LS \rangle & = & \sum_{i,j} c_{i,j} \ \langle L_i M_i \ell_j m_j|L M_L \rangle \langle S_i M_{S_i} {\scriptstyle \frac{1}{2}} \mu_j|S M_S \rangle \nonumber \\ & \times & \ \Phi_i(Atom;L_iS_i) \phi_j({\bf r}_0) \ . \label{wvfn} \end{aligned}$$ In the case of a single electron system, e.g. H, $\Phi_i(Atom;L_i S_i)$ is just a single electron wave function, i.e. an orbital. For a di-valent system, $\Phi_i(Atom;L_i S_i)$ is an antisymmetric product of two single electron orbitals coupled to have good $L_i$ and $S_i$ quantum numbers. The function $\phi_j({\bf r}_0)$ is a single positron orbital. The single particle orbitals are written as a product of a radial function and a spherical harmonic: $$\phi({\bf r}) = P(r) Y_{\ell m}({\hat {\bf r}}) \ . \label{orbital}$$ The radial wave functions are a linear combination of Slater Type Orbitals (STO) [@mitroy99f] and Laguerre Type Orbitals (LTOs). Most of the time the radial functions are LTOs, the exceptions occurring for single electron states with angular momenta equal to those of any occupied core orbitals. Since the Hartree-Fock core orbitals are written as a single combination of STOs, some of the active electron basis is written as linear combinations of STOs before the switch to a LTO basis is made. The LTO basis [@bromley02a; @bromley02b] has the property that the basis can be expanded toward completeness without having any linear independence problems. The present discussion is specific to positronic systems with a total orbital angular momentum of zero. It is straight-forward to generalize the discussion to states with $L > 0$, but this just adds additional algebraic complexities without altering any of the general conclusions. For a one electron system, the basis can be characterized by the index $J$, the maximum orbital angular momentum of any single electron or single positron orbital included in the expansion of the wave function. For two electron systems, the $L = 0$ configurations are generated by letting the two electrons and positron populate the single particle orbitals subject to two selection rules, $$\begin{aligned} \max(\ell_0,\ell_1,\ell_2) & \le & J \ , \\ \min(\ell_1,\ell_2) & \le & L_{\rm int} \ . \label{CIselect} \end{aligned}$$ In these rules $\ell_0$ is the positron orbital angular momentum, while $\ell_1$ and $\ell_2$ are the angular momenta of the electrons. The maximum angular momentum of any electron or positron orbital included in the CI expansion is $J$. The other parameter, $L_{\rm int}$ is used to eliminate configurations involving the simultaneous excitation of both electrons into high $\ell$ states. Double excitations of the two electrons into excited orbitals are important for taking electron-electron correlations into account, but the electron-electron correlations converge a lot more quickly with $L_{\rm int}$ than electron-positron correlations do with $J$. Calculations of the positronic bound states of the group II atoms and PsH [@bromley02a; @bromley02b] showed that the annihilation rate changed by less than 1$\%$ when $L_{\rm int}$ was varied from 1 to 3. The present set of calculations upon PsH were performed with $L_{\rm int} = 4$. Further details about the methods used to perform the calculations can be found elsewhere [@bromley02a; @bromley02b]. Various expectation values are computed to provide information about the structure of these systems. All observable quantities can be defined symbolically as $$\langle X \rangle^{J} = \sum_{L=0}^{J} \Delta X^{L} \ , \label{XJ1}$$ where $\Delta X^{J}$ is the increment to the observable that occurs when the maximum orbital angular momentum is increased from $J\!- \! 1$ to $J$, e.g. $$\Delta X^{J} = \langle X \rangle^{J} - \langle X \rangle^{J-1} \ . \label{XJ3}$$ Hence, one can write formally $$\langle X \rangle^{\infty} = \langle X \rangle^{J} + \sum_{L=J+1}^{\infty} \Delta X^{L} \ . \label{XJ2}$$ The first term on the right hand side will be determined by explicit computation while the second term must be estimated. The problem confronting all single center calculations is that most expectation values, $\langle X \rangle^{J}$ converges relatively slowly with $J$ and so the contribution of the second term can be significant. A sensible working strategy is to make $J$ as large as possible while simultaneously trying to use the best possible approximation to mop-up the rest of the partial wave sum. Existing extrapolation techniques and their limitations ------------------------------------------------------- One of the first groups to confront this issue and attempt a solution was the York University group of McEachran and Stauffer. They performed a series of polarized orbital calculations of positron scattering from rare gases [@mceachran78a; @mceachran78b; @mceachran79; @mceachran80]. The decrease in energy when the target atom relaxed in the field of a fixed positron was used to determine the polarization potential as a function of the distance from the nucleus. The polarized orbital method implicitly includes the influence of virtual Ps formation (within an adiabatic approximation), and this means that slow convergence can be expected. McEachran and Stauffer found that the scattering observables, namely the phase shift, and the $Z_{\rm eff}$ annihilation parameter, converged slowly with $J$, the largest angular momentum of the polarized electron orbital set used to represent the adjustment of the atomic charge cloud in the field of the positron. They took this into consideration by assuming their polarization potential scaled as $J^{-p}$ and the polarized orbital scaled as $J^{-q}$ at large $J$. They found $p \approx 3.8$ and $q \approx 1.8$ for the rare gases at $J \approx 12$. The recent CI-type calculations of Mitroy and collaborators also used an inverse power relation of $J^{-p}$, to complete the partial wave sum [@bromley02a; @bromley02b; @bromley03a]. In this case, the observables, $\varepsilon^J$, $\delta^J$, $\Gamma^J$ and $Z^J_{\rm eff}$ were extrapolated. This contrasts with the polarized orbital calculations where the polarization potential and polarized orbital were extrapolated. The value of $p$ was given by $$p = \ln \left( \frac {\Delta X^{J-1}}{\Delta X^J} \right) \biggl/ \ln \left( \frac{J-1}{J} \right) \ , \label{pfraction1}$$ while the constant factor is $$A_X = \Delta X^J J^{p} \ . \label{Avalue}$$ The correction factor was then evaluated by doing the sum $\sum_{L=J+1}^{\infty} A_X/L^p$ explicitly with an upper limit in the thousands. Gribakin and Ludlow [@gribakin02a] applied perturbation theory and the ideas of Schwartz [@schwartz62a; @schwartz62b] to determine the asymptotic behavior of the partial wave increments to the binding energies, phase shifts and annihilation rates of positron-atom systems. This work is largely derived from previous work on the partial wave expansion of two electron atoms [@hill85a; @kutzelnigg92a; @schmidt93a; @ottschofski97a]. They determined that the binding energy $E$, annihilation rate $\Gamma$, phase shift $\delta$, and collisional annihilation parameter $Z_{\rm eff}$ obey $$\begin{aligned} \Delta E^{J} = \langle E \rangle^J - \langle E \rangle^{J-1} &\sim & \frac{B_E}{(J+{\scriptstyle \frac{1}{2}})^4} \label{pE} \\ \Delta \Gamma^J = \langle \Gamma \rangle^J - \langle \Gamma \rangle^{J-1} & \sim & \frac{B_{\Gamma}}{(J+{\scriptstyle \frac{1}{2}})^2} \label{pGamma} \\ \Delta \delta^J = \langle \delta \rangle^J - \langle \delta \rangle^{J-1} & \sim & \frac{B_{\delta}}{(J+{\scriptstyle \frac{1}{2}})^4} \label{pdelta} \\ \Delta Z^J_{\rm eff} = \langle Z _{\rm eff} \rangle^J - \langle Z_{\rm eff} \rangle^{J-1} & \sim & \frac{B_Z}{(J+{\scriptstyle \frac{1}{2}})^2} \label{pZ} \end{aligned}$$ These expressions are merely the leading order terms of a series of the form $$\begin{aligned} \Delta X^J = \frac{B_X}{(J+{\scriptstyle \frac{1}{2}})^n} + \frac{C_X}{(J+{\scriptstyle \frac{1}{2}})^{n+1}} + \ldots \label{Zseries} \end{aligned}$$ To perform the actual extrapolation during a calculation of positron-hydrogen scattering, Gribakin and Ludlow [@gribakin04b] did a fit to calculated values at $J = 9$ and $J = 10$ with the formulae $$\begin{aligned} \delta = \langle \delta \rangle^{\infty} &=& \langle \delta \rangle^J + \sum_{L=J+1}^{\infty} \frac{B_{\delta}}{(L+{\scriptstyle \frac{1}{2}})^4 } \nonumber \\ &\approx& \langle \delta \rangle^J + \frac{B_{\delta}}{3(J+{\scriptstyle \frac{1}{2}})^3 } \ \label{Glebfit1} \ , \\ Z_{\rm eff} = \langle Z_{\rm eff} \rangle^{\infty} &=& \langle Z_{\rm eff} \rangle^J + \sum_{L=J+1}^{\infty} \frac{B_Z}{(L+{\scriptstyle \frac{1}{2}})^2 } \nonumber \\ & \approx & \langle Z_{\rm eff} \rangle^J + \frac{B_Z}{J+{\scriptstyle \frac{1}{2}} } \ , \label{Glebfit2}\end{aligned}$$ and so determined $Z_{\rm eff}$ and $\delta$. They used the approximate identities in eqs. (\[Glebfit1\]) and (\[Glebfit2\]) rather than explicitly evaluating the infinite sum. The identities appear to have been derived as an approximation to the $\int^{\infty}_{J+1} (L+{\scriptstyle \frac{1}{2}})^{-2} \ dL$ integral. However, in equating the sum to the integral they implicitly assume a rectangle rule representation of the integral which is in error of 5-10$\%$ for $J \in [7,10]$ (the net effect of this is that Gribakin and Ludlow state that the increments decrease as $B/(L+{\scriptstyle \frac{1}{2}})^{n}$ but actually assume a $B/L^{n}$ decrease when evaluating the $J \to \infty$ correction). A better approximation to the series is obtained by using a mid-point rule to represent the integral. Doing this leads to $$\sum_{L=J+1}^{\infty} \frac{1}{(L+{\scriptstyle \frac{1}{2}})^n } \approx \frac{1}{(n-1)(J+1)^{n-1} } \ . \label{bettertail}$$ This approximation is accurate to 0.1$\%$ for $n = 2$ and $J = 7$. It will be shown that a more serious problem with the Gribakin and Ludlow methodology is that eqs. (\[Glebfit1\]) and (\[Glebfit2\]) cannot reveal whether the calculated $Z^{J}_{\rm eff}$ are deviating from the expected asymptotic form. For example, successive increments to either the phase shift or $Z_{\rm eff}$ have usually decreased more slowly with $J$ (for $ J$ ranging between 10 and 18) than indicated by eqs. (\[pE\])-(\[pZ\]) [@bromley02a; @bromley02b; @bromley03a; @novikov04a]. Instead of having $p = 2$ or $p = 4$ the successive increments often gave slightly smaller values for $p$. The approach adopted by Gribakin and Ludlow is insensitive to these deviations. Saito has investigated the structure of the PsH and the Ps-halogen systems with the CI method [@saito03a; @saito05a]. A Natural Orbital (NO) truncation algorithm based on the energy was used to reduce the dimensionality of the secular equations, thus making calculations on the heavier halogen atoms viable. Besides using the inverse power series, Saito used the functional form $$\Delta X^{J} = 10^{-\alpha(\log_{10}J)^\beta+\gamma} \ , \label{plog}$$ to complete the partial wave sum for the annihilation rate. This function was not based on any physical principles, and its usage was justified on the grounds that the increments were decreasing faster than $J^{-2}$. However, it has been suggested that the annihilation rate increments were decreasing too quickly because the dimension of the radial basis used in the Ps-halogen calculations was simply too small [@mitroy05i]. So the rationale behind the usage of eq. (\[plog\]) is questionable. Some mention must be made of the difficulties associated with the slower convergence of the annihilation rate. Consider the PsH system, a calculation with $J = 9$ gave 72$\%$ of the total annihilation rate [@bromley02a]. If one doubled the size of $J$, then eq. (\[pGamma\]) suggests that the explicit calculation would only recover 86$\%$ of the total annihilation rate. And it would take a calculation with $J \approx 250$ to recover 99$\%$ of the annihilation rate. The situation is even more sobering when one considers that the annihilation rate converges faster for PsH (since it is the most compact) than for any other positron binding system. Comparison of existing and new approaches to the partial wave extrapolation =========================================================================== The different alternatives -------------------------- In order to expose the strengths and deficiencies of existing approaches, very large calculations have been performed on three mixed electron-positron systems. These are the $e^+$-H scattering system for the $\ell = 0$ partial wave, and the bound PsH and $e^+$Cu systems. It will be seen that the typical calculations on these real-world systems do not agree perfectly with the leading order asymptotic form given by Gribakin and Ludlow, i.e. eqs. (\[pE\]) - (\[pZ\]). Accordingly, six different extrapolation methods for determining the $J \to \infty$ correction were tested. These were: **Method $p$**. The successive increments to all quantities are assumed to obey an $$X^J = \frac{A_X}{(J+{\scriptstyle \frac{1}{2}})^{p} } \ , \label{Xgen}$$ law with the exponent $p$ determined from eq. (\[pdef\]). The value of $p$ derived from three successive calculation of $X^{J-2}$,$X^{J-1}$ and $X^J$ is given by $$p = \ln \left( \frac {\Delta X^{J-1}}{\Delta X^J} \right) \biggl/ \ln \left( \frac{J+{\scriptstyle \frac{1}{2}}}{J-{\scriptstyle \frac{1}{2}}} \right) \ . \label{pdef}$$ (Note, in previous works we have used a $J^{-p}$ series [@mitroy99c; @bromley02a; @bromley02b; @bromley03a]). The notations $p_E$, $p_{\Gamma}$, $p_{\delta}$ and $p_Z$ are used to denote the exponents derived from the partial wave expansions of the energy, annihilation rate, phase shift and $Z_{\rm eff}$ The discrete sum over $L$ in (\[XJ2\]) is done explicitly up to $J = 200$. The remainder of the sum is then estimated using eq. (\[bettertail\]). **Method $p_{\mathrm{av}}$**. This is based on Method $p$. Three successive calculations for $(J\!-\!2)$, $(J\!-\!1)$ and $J$ are once again used with eqs. (\[Xgen\]) and (\[pdef\]) to determine an initial estimate of $p_0$. Then, $p$ is set to the average of $p_0$ and the expected value of either 2 or 4. This method makes an admittedly crude attempt to correct method $p$ in those cases where $p_{\delta}$ and $p_Z$ were significantly different from 4 and 2 [@bromley02b]. Once $p$ has been fixed, eq. (\[Xgen\]) can then be used to determine $A_X$ and the discrete sum over $L$ in (\[XJ2\]) is done explicitly up to $J = 200$. The remainder of the sum is then estimated using eq. (\[bettertail\]). **Method GL**. The relations eq. (\[Glebfit1\]) and (\[Glebfit2\]) are assumed to be exact. The two largest values of $\langle X \rangle ^J$ are used to determine $\langle Z_{\rm eff} \rangle^{\infty}$ and $\langle \delta \rangle^{\infty}$. This method mimics the procedure adopted by Gribakin and Ludlow [@gribakin04b]. **Method I**. The functional form $$\Delta X^J = \frac{B_X}{(J+{\scriptstyle \frac{1}{2}})^n} \label{Lseries1}$$ is assumed to apply and the $\Delta X^{J}$ increment are used to determine $B_X$. The exponent $n$ is set to $2$ for the annihilation rate and $4$ for the energy or phase shift. The discrete sum over $L$ in (\[XJ2\]) is done explicitly up to $J = 200$ and beyond that point eq. (\[bettertail\]) is used. This method has similarities with the GL method. **Method II**. The functional form $$\Delta X^J = \frac{B_X}{(J+{\scriptstyle \frac{1}{2}})^n} + \frac{C_X}{(J+{\scriptstyle \frac{1}{2}})^{n+1}} \ , \label{Lseries2}$$ is assumed to apply and the $\Delta X^{J}$ and $\Delta X^{J-1}$ increments are used to determine $B_X$ and $C_X$. The second term in eq. (\[Lseries2\]) comes from 3rd-order perturbation theory [@hill85a; @kutzelnigg92a; @ottschofski97a]. The exponent $n$ is set to $2$ for the annihilation rate and $4$ for the energy or phase shift. The discrete sum over $L$ in (\[XJ2\]) is done explicitly up to $J = 200$ and beyond that point eq. (\[bettertail\]) is used. **Method III**. The functional form $$\Delta X^J = \frac{B_X}{(J+{\scriptstyle \frac{1}{2}})^n} + \frac{C_X}{(J+{\scriptstyle \frac{1}{2}})^{n+1}} + \frac{D_X}{(J+{\scriptstyle \frac{1}{2}})^{n+2}} \ , \label{Lseries3}$$ is assumed to apply and the $\Delta X^{J}$, $\Delta X^{J-1}$ and $\Delta X^{J-2}$ increments are used to determine $B_X$, $C_X$ and $D_X$. Other particulars are the same as those for Methods I and II. **Method S**. The functional form adopted by Saito, eq. (\[plog\]) is used. The parameters $\alpha$, $\beta$ and $\gamma$ are determined from $\Delta X^{J-2}$, $\Delta X^{J-1}$ and $\Delta X^J$. Then the series is completed by summing to $J=2000$. The $e^+$-H scattering system ----------------------------- The CI-Kohn method has already been used to generate phase shift and annihilation parameter data for the $e^+$-H scattering system [@bromley03a]. New calculations with a radial basis set (i.e. the number of LTOs per $\ell$) of increased dimensionality have been done in order to minimize the influence of the radial basis upon any conclusions that are drawn. The present investigation examines $s$-wave $e^+$H scattering at $k = 0.4 \ \ a_0^{-1}$. [lcc]{} $J$ & $\langle \delta \rangle^J$ (radians) & $\langle Z_{\rm eff} \rangle^J$\ 0 & -0.1992097664 & 0.4529537597\ 1 & -0.01002215705 & 1.018467362\ 2 & 0.05339800596 & 1.466777767\ 3 & 0.08175604707 & 1.798430751\ 4 & 0.09626735469 & 2.043817341\ 5 & 0.1043829191 & 2.228559914\ 6 & 0.1092338304 & 2.370688616\ 7 & 0.1122907225 & 2.482386967\ 8 & 0.1143023638 & 2.571897852\ 9 & 0.1156750002 & 2.644890677\ 10 & 0.1166408472 & 2.705331457\ 11 & 0.1173386269 & 2.756059864\ 12 & 0.1178543969 & 2.799146079\ \ $p$ & 3.6248 & 1.9583\ Method $p$ & 0.1200652 & 3.34020\ Method $p_{av}$ & 0.1199020 & 3.32823\ Method GL & 0.1196691 & 3.29464\ Method I & 0.1197593 & 3.31675\ Method II & 0.1199484 & 3.32750\ Method III & 0.1199456 & 3.30092\ Method S & 0.1198834 & 3.21989\ \ CI-Kohn $J \rightarrow \infty$ [@bromley03a] & 0.1198 & 3.232\ Optical potential [@bhatia71; @bhatia74b]& 0.1201 & 3.327\ Variational [@vanreeth97a; @vanreeth98a; @gribakin04b] & 0.1198 & 3.407\ Close Coupling [@mitroy95a; @ryzhikh00a] & 0.1191 & 3.332\ The largest calculation for $e^+$-H included a minimum of 30 LTOs per $\ell$ with additional LTOs included at small $\ell$. Special attention was given to the $\ell = 1$ positron basis since this channel is responsible for the long-range polarization potential. The dimension of the LTO basis was 80 in this case. All radial integrations were taken to 729 $a_0$ on a composite Gaussian grid. The earlier calculations of Bromley and Mitroy [@bromley03a] with a minimum of 17 LTOs per $\ell$ will be presented for comparison. Table \[tabpH\] gives the $e^+$-H phase shift and $Z_{\rm eff}$ for $s$-wave scattering at $k = 0.4$ $a_0^{-1}$ up to $J = 12$. The larger basis will be referred to as basis 2 while the older basis will be named basis 1. First of all, the values of $p_{\delta}$ and $p_Z$ derived from the earlier [@bromley03a] and present CI-Kohn calculations are shown together in Figure \[pHp\] as a function of $J$. This figure tests whether eqs. (\[pdelta\]) and (\[pZ\]) describe the behavior of a real-world calculation. ![ The exponents $p_{\delta}$ and $p_Z$ as a function of $J$ for 2 different CI-Kohn calculations of $s$-wave $e^+$-H scattering at $k = 0.4 \ a_0^{-1}$. The short curve labeled GL was plotted using $\langle Z_{\rm eff} \rangle^J$ data of GL [@gribakin04b]. []{data-label="pHp"}](pHp.eps){width="8.0cm"} Neither $p_{\delta}$ or $p_Z$ are within 1$\%$ of the expected asymptotic value at $J = 12$ and both are approaching the expected asymptotic limit from below. The two calculations give almost exactly the same $p_{\delta}$ while the larger calculation tends to give smaller values of $p_Z$ with the difference becoming greater as $J$ increases. The increasing difference between $p_Z$ for the two calculations suggests that a converged calculation of $\Delta Z^J_{\rm eff}$ needs an increasingly larger radial basis as $J$ increases. This point is addressed in more detail later. In most of the calculations we have performed, the values of $p$ derived from eq. (\[pdef\]) have been slightly smaller than the expected value at $J \approx 10$ [@bromley02a; @bromley02b; @bromley02d; @novikov04a]. We also note that in all the calculations we have so far performed, the values of $p$ increase steadily (once the broad features of the physical system have been achieved) as $J$ increases. While the asymptotic increments to $\delta$ and $Z_{\rm eff}$ do not agree exactly with eqs. (\[pdelta\]) and (\[pZ\]), the observed trends do appear to be consistent with their derived limits. Figures \[Hphaseinf\] and \[HZinf\] show the behavior of the extrapolated $\delta$ and $Z_{\rm eff}$ as a function of $J$ for some of the different extrapolations. Table \[tabpH\] gives estimates of $\delta$ and $Z_{\rm eff}$ using the calculated values at the largest possible $J$ value (i.e. 12) to determine the $J \to \infty$ corrections. There is one problem in interpreting the results of Figures \[Hphaseinf\] and \[HZinf\] and Table \[tabpH\]. The exact value of $Z_{\rm eff}$ is imprecise at the level of 2-3$\%$. The old calculation of Bhatia *et.al.* using $r_{ij}$ co-ordinates gives 3.327 [@bhatia71; @bhatia74b], the close coupling calculation by Ryzhikh and Mitroy using the $T$-matrix method gives 3.332 [@mitroy95a; @ryzhikh00a], while the variational calculation of van Reeth [*et.al.*]{} gives 3.407 [@vanreeth97a; @vanreeth98a; @gribakin04b]. There is also some scatter in the estimates of the phase shifts, but the degree of difference between the Bhatia [*et.al.*]{} and van Reeth [*et.al.*]{} phase shifts is only 0.3$\%$. (the $T$-matrix phase shift is only expected to have an accuracy of about 1$\%$ [@mitroy95a]). ![ The extrapolated $J \rightarrow \infty$ limit of the phase shift (in radians) $\delta$ for $e^+$-H scattering at $k = 0.4 \ a_0^{-1}$ as a function of $J$. The horizontal solid line shows the phase shift of Bhatia [*et.al.*]{} [@bhatia71]. The phase shift without any $J \rightarrow \infty$ correction is given by the curve labeled $\langle \delta \rangle^J$. []{data-label="Hphaseinf"}](Hphaseinf.eps){width="8.0cm"} ![ The extrapolated $J \rightarrow \infty$ limit using the different methods to complete the partial wave series for the $e^+$-H $Z^J_{\rm eff}$ at $k = 0.4 \ a_0^{-1}$ a function of $J$. The horizontal solid line shows the $Z_{\rm eff}$ of Bhatia [*et.al.*]{} [@bhatia74b] while the horizontal dotted line shows the $Z_{\rm eff}$ of van Reeth [*et.al.*]{} [@vanreeth97a; @vanreeth98a; @gribakin04b]. []{data-label="HZinf"}](HZinf.eps){width="8.0cm"} Figure \[Hphaseinf\] shows that inclusion of the $J \to \infty$ correction leads to greatly improved estimates of $\langle \delta \rangle^{\infty}$. In terms of their impact, the methods belong to 3 operational classes. Firstly, Method $p$ consistently gives the largest values of $\langle \delta \rangle^{\infty}$. Fixing $p_{\delta}$ to a set value at a finite $J$ inevitably results in the $J \to \infty$ correction being overestimated. For example, using $p_{\delta} = 3.343$ at $J = 8$ to fix $p_{\delta}$ for all $J$ results in increments to $\Delta \delta^J$ that do not decrease quickly enough. Methods I and GL, on the other hand, tend to underestimate the size of the $J \to \infty$ correction since $p_{\delta} = 4$ is fixed prior to the sequence of $\Delta \delta^J$ increments achieving the expected $(J+{\scriptstyle \frac{1}{2}})^{-4}$ form. Those methods which attempt to allow for deviations from the leading order behavior, namely Methods II, III and $p_{av}$ approach the expected $J \to \infty$ limit much earlier. Indeed, their estimates of $\langle \delta \rangle^{\infty}$ differ by less than 0.5$\%$ at $J = 6$. Table \[tabpH\] reveals that these 3 methods give $\langle \delta \rangle^{\infty}$ estimates that differ by less than 0.1$\%$ at $J = 12$. Of the 3 alternatives, the 3-term asymptotic series, namely Method III, seems to possess the best convergence properties. Method S also appears to give a reasonable estimate of $\langle \delta \rangle^{\infty}$ when $J \ge 6$. The tabulated estimates of $\langle \delta \rangle^{\infty}$ at $J = 12$ reflect the discussions in the above paragraph. Method $p$ gives the largest phase shift while Methods GL and I give the smallest phase shifts. The maximum difference between any of the phase shifts is only 0.3$\%$ since the net effect of the $J \to \infty$ contribution to the phase shift is only 2$\%$. Figure \[HZinf\] for $Z_{\rm eff}$ shows some features in common with Figure \[Hphaseinf\]. Once again, application of a $J \to \infty$ correction is seen to give much improved estimates of the $\langle Z_{\rm eff} \rangle^{\infty}$ limit. Method $p$ also gives the largest estimate of $Z_{\rm eff}$. Method S gives values of $\langle Z_{\rm eff} \rangle^{\infty}$ that are generally the smallest, and Table \[tabpH\] reveals that it gives a value that is 0.1 smaller than any of the other approaches at $J = 12$. It is not surprising that a method based on a fitting function with no physical justification performs so poorly, and its reasonable estimate of $J \to \infty$ correction to the phase shift can be regarded as a numerical coincidence. No further discussion of Method S will be made although values are reported in the later tables for reasons of completeness. A more detailed analysis and discussion of Figure \[HZinf\] cannot be made until the convergence properties of the underlying radial basis are exposed. Convergence properties of the radial basis ------------------------------------------ In addition to converging very slowly with $J$, the annihilation rate also converges slowly with respect to the number of radial basis functions since the actual wave function has a cusp at the electron-positron coalescence point. A previous CI investigation on helium in an $\ell = 0$ model indicates that the electron-electron $\delta$-function converged as $O(N^{-5/2})$ where $N$ is number of Laguerre orbitals [@mitroy06a]. And it has also been demonstrated that the relative accuracy of the electron-electron $\delta$-function increment for a given size radial basis decreased as $L$ increased [@bromley06a]. Some sample ratios can be used to illustrate these points. The ratio $R^{J}_Z$ compares the two calculations of $\Delta Z^{J}_{\rm eff}$ for $e^+$-H scattering by determining the ratio for basis 1(17 LTOs) and basis 2(30 LTOs). It is defined as $$R^{J}_Z = \frac{(\Delta Z^{J}_{\rm eff})_{30}} {(\Delta Z^{J}_{\rm eff})_{17}} \ . \label{ratio}$$ A similar ratio, $R^{J}_{\delta}$ can be defined for the increment to the phase shifts. A plot of these two ratios is given in Figure \[RdE\], while Table \[radialbasis\] lists values of $R^{J}_{\delta}$ and $R^J_Z$ for some selected $J$ values. Figure \[RdE\] clearly demonstrates the $\Delta Z_{\rm eff}^J$ converges more slowly that the phase shift increments at large $J$. First, the annihilation rate is more sensitive to the size of the radial basis than is the phase shift. Second, the higher partial waves are more sensitive to the size of the radial basis than the lower partial waves. For example there was a 4.0$\%$ increase in $\Delta Z^8_{\rm eff}$ between the basis 1 and basis 2 calculations while there was a larger 7.2$\%$ increase in $\Delta Z^{12}_{\rm eff}$. However, the increase in $\Delta \delta^{12}$ was only 1.7$\%$. ![ The ratio of the increments to $\langle \delta \rangle^J$ and $\langle Z_{\rm eff} \rangle^J$ (refer to eq.(\[ratio\])) for $e^+$-H scattering at $k = 0.4$ $a_0^{-1}$ as a function of $J$ for the 17 and 30 LTO calculations. []{data-label="RdE"}](RdE.eps){width="8.0cm"} Since the lack of completeness in the radial basis has the largest impact at high $J$, it will obviously affect the $J \to \infty$ correction. For example, Method II gives $\Delta Z_{\rm eff}^{J > 12} = 0.4695$ for basis 1. For basis 2, the correction of $\Delta Z_{\rm eff}^{J > 12} = 0.5284$ is significantly larger. [lcc]{} $J$ & $R_\delta$ & $R_Z$\ \ 6 & 1.0043 & 1.0261\ 8 & 1.0075 & 1.0399\ 10 & 1.0117 & 1.0556\ 12 & 1.0169 & 1.0724\ $J$ & $R_E$ & $R_{\Gamma}$\ \ 8 & 1.0100 & 1.0288\ 12 & 1.0258 & 1.0552\ 16 & 1.0477 & 1.0869\ \ 4 & 1.0096 & 1.0478\ 6 & 1.0114 & 1.0825\ 9 & 1.0626 & 1.1357\ The implications of these results can be seen by consideration of the Method II plot of $Z_{\rm eff}^{\infty}$ depicted in Figure \[HZinf\]. This achieves a maximum value of 3.41 at $J = 5$, and then decreases until it is 3.328 at $J = 12$. The question to be addressed is whether the decrease from $J > 6$ is due to convergence properties of the calculation with respect to $J$ or the convergence properties of the basis with respect to $N$, the number of orbitals per $\ell$? Although there are seven estimates of $Z_{\rm eff}$ that lie between 3.30 and 3.34 in Table \[tabpH\], we believe that the true value of $Z_{\rm eff}$ lies closer to 3.407 (the value of van Reeth [*et.al.*]{}) than to 3.327 (the value of Bhatia [*et.al.*]{}). This interpretation is supported by a crude estimate of the variational limit deduced from Figure \[RdE\]. The assumption is made that the $\Delta Z_{\rm eff}^J$ increments converge as $O(N^{-5/2})$ [@mitroy06a]. Consequently, one deduces that the plotted $R^J_{Z}$ ratios comprise some 57$\%$ of the necessary correction to the variational limit. The variational limit for any increment then estimated to be $\Delta Z_{\rm eff}^J \times (1 + 0.43 \times (R^J_{Z}-1)/0.57)$. Applying this correction to the data in Table \[tabpH\] gives $\langle Z_{\rm eff} \rangle^{\infty} = 3.410$ when Method II is used to estimate the $J \to \infty$ limit (the actual correction of $\Delta Z_{\rm eff}^{J>12}$ of 0.578 was about 9$\%$ larger than the basis 2 value). Determination of the variational limit for the phase shift can also be done by assuming that the $\Delta \delta^J$ increments converge as $O(N^{-7/2})$ [@mitroy06a]. In this case, the plotted $R^J_{\delta}$ ratios comprise some 76$\%$ of the correction to the variational limit and thus the final estimates of the phase shift increments would be $\Delta \delta^J \times (1 + 0.24 \times (R^J_{\delta}-1)/0.76)$. The Method II phase shift only increased by $1.3 \times 10^{-5}$ radian giving $\langle \delta \rangle^{\infty} = 0.12008$. One point from Table \[tabpH\] warrants special attention. The 3-term asymptotic series, namely Method 3, gives a smaller $\langle Z_{\rm eff} \rangle^{\infty}$ than Method I! This seems ridiculous given that $p = 1.9583$ at $J = 12$. The tendency for the $\Delta Z_{\rm eff}^J$ increments to be systematically underestimated results in the corruption of the $B_Z$, $C_Z$ and $D_Z$ coefficients extracted from the 3-term fit and renders Method III unreliable for determination of $\langle Z_{\rm eff} \rangle^{\infty}$. Even though data for only one system has been presented, it is possible to make some general comments about the performance of the difference methods since analysis of the $e^+$Cu and PsH data will confirm these conclusions (they are also compatible with the results of large basis CI calculations of He [@bromley06a]). Since the $p_{\delta}$ and $p_Z$ exponents tend to be smaller than 4 and 2 respectively at finite $J$, Method $p$ has an inherent tendency to overestimate the $J \to \infty$ corrections. Application of this approach in the past has not resulted in any gross errors since the problems associated with fixing $p$ at values less than 2 and 4 tend to cancel out the errors associated with a radial basis of finite size $p$ [@bromley02d]. This method should only be applied in situations when the asymptotic form of the expectation value under investigation is unknown. Method I generally underestimates the $J \to \infty$ correction. It gives a useful estimate of the $J \to \infty$ correction and should mainly be applied to give rough estimates for low precision calculations. Method GL can be regarded as a variety of Method I that happens to give inferior $J \to \infty$ corrections. Methods II and $p_{av}$ were seen to give $J \to \infty$ corrections that were close to each other once the calculation reached a certain value of $J$. Method II should be be preferred since it is founded in correct asymptotics. Method III seems to give the earliest reliable estimate of the phase shift. However, it should not be applied to the annihilation rate unless the radial basis is substantially larger than the present basis. Method III should only be applied in situations where the underlying partial wave increments have an accuracy of better than 1$\%$ and in addition the increments should vary smoothly and not exhibit fluctuations. The $e^+$Cu ground state ------------------------ Table \[tabpCu\] gives the $e^+$Cu binding energy and annihilation rate as a function of $J$ up to $J = 18$ for the calculation with 25 LTOs. The table also includes values from a calculation with the fixed core stochastic variation method (FCSVM) [@ryzhikh98f; @bromley02e]. The FCSVM basis includes the electron-positron coordinate explicitly and is very close to convergence. The FCSVM calculation uses a slightly different model potential so it is not expected that the CI energy and $\Gamma$ should be exactly the same. Figure \[pCup\] displays $p_E$ and $p_{\Gamma}$ versus $J$ for two different CI calculations of the $e^+$Cu ground state. One plot is derived from the earlier calculation of Bromley and Mitroy [@bromley03a] which included a minimum of 15 LTOs per $\ell$-value (basis 1). The present calculation (basis 2) is much larger with a minimum of 25 LTOs per $\ell$ value (note, more than 25 LTOs were included for $\ell$ = 0, 1 and 2 since these make the largest contribution to the energy and annihilation rate). [lccc]{} $J$ & $\varepsilon$ & $\Gamma_c$ & $\Gamma_v$\ & Hartree & $10^9$ sec$^{-1}$ & $10^9$ sec$^{-1}$\ 0 & -0.00112467 & 0.000289 & 0.000132\ 1 & -0.00080292 & 0.001443 & 0.001692\ 2 & -0.00037356 & 0.004818 & 0.009728\ 3 & 0.00031179 & 0.011213 & 0.033656\ 4 & 0.00111995 & 0.017605 & 0.070360\ 5 & 0.00187958 & 0.022223 & 0.110121\ 6 & 0.00251736 & 0.025271 & 0.147793\ 7 & 0.00302852 & 0.027272 & 0.181727\ 8 & 0.00343136 & 0.028611 & 0.211657\ 9 & 0.00374774 & 0.029526 & 0.237830\ 10 & 0.00399692 & 0.030168 & 0.260656\ 11 & 0.00419429 & 0.030627 & 0.280571\ 12 & 0.00435174 & 0.030963 & 0.297984\ 13 & 0.00447832 & 0.031212 & 0.313256\ 14 & 0.00458086 & 0.031402 & 0.326694\ 15 & 0.00466456 & 0.031547 & 0.338564\ 16 & 0.00473338 & 0.031660 & 0.349087\ 17 & 0.00479037 & 0.031749 & 0.358454\ 18 & 0.00483788 & 0.031821 & 0.366821\ \ $p$ & 3.2751 & 4.0511 & 2.0295\ Method $p$ & 0.0052012 & 0.032219 & 0.51307\ Method $p_{\mathrm{av}}$ & 0.0051481 & & 0.51526\ Method GL & 0.0050998 & & 0.51325\ Method I & 0.0051080 & & 0.51751\ Method II & 0.0051584 & & 0.51529\ Method III & 0.0051603 & & 0.49364\ Method S & 0.0051394 & & 0.46398\ \ FCSVM [@ryzhikh98f; @bromley02e] & 0.005597 & 0.0339 & 0.544\ CI, $J = 18$ [@bromley02e] & 0.004786 & 0.03173 & 0.35499\ CI, $J \to \infty$ [@bromley02e] & 0.005117 & 0.0321 & 0.4744\ ![ The exponents ($p_E$ and $p_{\Gamma}$) for 2 different CI calculations of $e^+$Cu as a function of $J$. []{data-label="pCup"}](pCup.eps){width="8.0cm"} The plots of $p_E$ and $p_{\Gamma}$ against $J$ for $e^+$Cu are similar to the plots of $p_{\delta}$ and $p_Z$ for $e^-$-H scattering. Both exponents are generally smaller than the expected asymptotic limits but steadily increase as $J$ increases. The actual value of $p_{\Gamma}$ at $J = 18$, namely 2.030, was marginally larger than the expected asymptotic limit of $p_{\Gamma} = 2$. The estimates of $\langle \varepsilon \rangle^{\infty}$ and $\langle \Gamma \rangle^{\infty}$ as a function of $J$ are shown in Figures \[CupEinf\] and \[CupGinf\]. Table \[tabpCu\] reveals an 8$\%$ increase in $\Gamma$ when compared with the earlier CI calculation value of $\Gamma$, namely 0.474 $\times 10^9$ s$^{-1}$ [@bromley02e]. This is a consequence of the bigger radial basis used in the present work. The selected values of $R^J_{\Gamma}$ listed in table \[radialbasis\] reveal an 8.7$\%$ increase in $\Delta \Gamma^{16}$ for basis 2 compared to basis 1. It is expected that further increases in the radial basis would eventually lead to a $p_{\Gamma}$ that was smaller than 2.0 at all $J = 18$. Again it is noticed that $R^J_{\Gamma}$ (and $R^J_E$) increase with increasing $J$. The estimates of the annihilation rate in table \[tabpCu\] are all very close together (with the exception of Method III). This occurs because $p_{\Gamma}= 2.030$ is very close to the expected value of 2.0. The variations between the different approaches are largely concerned with taking care of the deviations from the $p_{\Gamma} = 2$ behavior, and with a minimal deviation at $J =18$, one should expect minimal differences between the final results. Method III is the least accurate (discounting Method S) since it is the most susceptible to the inaccuracies in $\Delta \Gamma^{J}$ Figure \[CupGinf\] shows that Method $p$ systematically overestimates $\langle \Gamma \rangle^{\infty}$ at smaller values of $J$. Method I, on the other hand generally gives the smallest estimates of the $\langle \Gamma \rangle^{\infty}$ and is consistently too small at lower $J$. Method III obtains a reasonable estimate of $\langle \varepsilon \rangle^{\infty}$ the quickest. Beyond $J > 12$ the Method III binding energy does decrease slightly. This may be due to slower convergence of the radial basis at higher $J$. Methods II and $p_{\mathrm{av}}$ give roughly equal estimates of $\langle \varepsilon \rangle^{\infty}$ for $J > 10$, and the spread between the Methods II, III and $p_{\rm av}$ estimates of $\langle \varepsilon \rangle^{\infty}$ is only 0.2$\%$ at $J = 18$. Method $p$ gives the largest estimate of $\langle \varepsilon \rangle^{\infty}$ for all $J$ shown in Figure \[CupEinf\], while Methods I and GL give the smallest values (with Method GL once again being worse that Method I). It is worth mentioning that Methods II, III and $p_{\rm av}$ are all roughly constant after $J = 16$ while Methods I, GL and $p$ are still increasing or decreasing. In summary, the totality of information in Table \[tabpCu\] and Figures \[pCup\], \[CupEinf\] and \[CupGinf\] is very reminiscent of the situation for $e^+$-H scattering and is consistent with the conclusions derived from the analysis of $e^+$-H scattering. ![ The $e^+$Cu binding energy as a function of $J$. The different curves use different algorithms to estimate the $J \rightarrow \infty$ correction as discussed in the text. []{data-label="CupEinf"}](CupEinf.eps){width="8.0cm"} ![ The $e^+$Cu annihilation rate (in units of $10^9$ s$^{-1}$) as a function of $J$. The different curves use different algorithms to estimate the $J \rightarrow \infty$ correction as discussed in the text. []{data-label="CupGinf"}](CupGinf.eps){width="8.0cm"} The PsH ground state -------------------- [lcccccc]{} $J$ & $\lambda$ & $N_{\rm orb}$ & $N_{CI}$ & $E(\text{PsH})$ & $\Gamma$\ 0 & 2.10 & 17 & 2601 & -0.69133618 & 0.374196\ 1 & 2.26 & 33 & 9265 & -0.74705969 & 0.782256\ 2 & 2.36 & 48 & 22810 & -0.76620031 & 1.080456\ 3 & 2.46 & 63 & 44650 & -0.77514128 & 1.292538\ 4 & 2.52 & 78 & 78640 & -0.77995286 & 1.448216\ 5 & 2.72 & 93 & 120265 & -0.78274494 & 1.566206\ 6 & 2.93 & 108 & 165265 & -0.78449597 & 1.658344\ 7 & 3.13 & 123 & 213415 & -0.78565220 & 1.732023\ 8 & 3.34 & 138 & 263365 & -0.78644538 & 1.792061\ 9 & 3.56 & 153 & 314890 & -0.78700639 & 1.841756\ 10 & 3.75 & 168 & 366415 & -0.78741330 & 1.883366\ 11 & 3.95 & 183 & 417940 & -0.78771485 & 1.918595\ 12 & 4.15 & 198 & 469465 & -0.78794247 & 1.948689\ 13 & 4.35 & 213 & 520990 & -0.78811707 & 1.974632\ \ & 3.44547 & 1.92375\ & -0.7889974 & 2.3352\ & -0.7888995 & 2.3232\ & -0.7887894 & 2.2988\ & -0.7888198 & 2.3121\ & -0.7889218 & 2.3225\ & -0.7889231 & 2.2915\ & -0.7888843 & 2.2203\ \ & -0.78919674 & 2.4712\ & 90/91 & 63492 & -0.7866818 & 1.7903\ & 90/91 & 95324 & -0.7867761 & 1.7913\ & $\infty$ & $\infty$ & -0.786949 & 1.8230\ The best two electron system for validation purposes is the positronium-hydride (PsH) system since its properties are very well known as a result of previous investigations [@ryzhikh99a; @yan99a; @usukura98; @mitroy06d]. The stochastic variational method (SVM) expectation values listed in Table \[PsH\] are taken from a new calculation with 1800 ECGs. The energy of this wave function, $E = -0.7891674$ Hartree is the best PsH energy so far reported [@mitroy06d]. The orbital basis used in the present calculation was about twice as large as that used in previous calculations [@bromley02a; @saito03a]. The number of radial functions per $\ell$ was 15 with the exception of $\ell = 0$ and $\ell = 1$ where 17 and 16 functions respectively were used. The largest $J$ was 13 while $L_{\rm int}$ was set to 4. The basis functions for each $\ell$ used a common exponent that had been energy optimized during some preliminary and smaller calculations. It must be emphasized that choosing a common $\lambda$ for both electron and positron states was not arbitrary but was a consequence of energy optimization process. The results of the CI calculation are listed in Table \[PsH\]. ![ The exponents $p_{E}$ and $p_{\Gamma}$ as a function of $J$ for 2 different CI calculations of PsH. Both the basis 1 and basis 2 with $L_{\rm int} = 4$. The third data set was taken from the “full CI” calculations of Saito [@saito03a]. []{data-label="pPsH"}](pPsH.eps){width="8.5cm"} The variation of $p_E$ and $p_{\Gamma}$ with $J$ in Figure \[pPsH\] reflects the behavior seen in Figures \[pHp\] and \[pCup\]. The values of $p$ are smaller than the predicted asymptotic limits and seem to be approaching the correct value. Computational constraints mean that the dimensions of the radial basis, e.g. 15 $e^+$ and $e^-$ LTOs per $\ell$, are smaller than those in the $e^+$-H and $e^+$Cu calculations. The PsH radial basis is further from convergence than the basis sets used for the equivalent calculations upon $e^+$-H and $e^+$Cu. Figure \[pPsH\] gives $p$ values taken from the CI calculations of Saito [@saito03a]. Saito used Natural Orbital techniques to reduce the dimension of the final diagonalization while using an orbital basis with $J = 9$ (this is about the same size as in [@bromley02a]). Saito estimated the variational limit at each $J$ and the Figure \[pPsH\] curves were derived from this “full CI limit” calculation. Although the Saito curves have irregularities, they exhibit a $p$ versus $J$ variation similar to the early Bromley and Mitroy calculation [@bromley02a] ![ The PsH binding energy (in units of Hartree) with respect to the Ps+H threshold of -0.750 Hartree as a function of $J$. The different curves use different algorithms to estimate the $J \rightarrow \infty$ correction as discussed in the text. The close to converged SVM energy [@mitroy06d] is shown as the horizontal line for comparison purposes. []{data-label="PsHEinf"}](PsHEinf.eps){width="8.5cm"} Figure \[PsHEinf\] shows the variation of $\langle \varepsilon \rangle^{\infty}$ vs $J$. Once again, the 3-term series, Method III achieves its asymptotic value at the smallest value of $J$. Methods II and $p_{\rm av}$ achieve their limiting values near $J = 10$. Methods I and GL again tend to underestimate $\langle \varepsilon \rangle^{\infty}$ while Method $p$ overestimates $\langle \varepsilon \rangle^{\infty}$. The best CI estimate of the PsH energy is obtained by adding an $L_{\rm int}$ correction of $9.5 \times 10^{-5}$ Hartree (the difference between the $L_{\rm int} = 4$ and $L_{\rm int} = 9$ energies [@bromley02a]) to the Method III $\varepsilon$ of $0.038923$ Hartree. The resulting binding energy of $\varepsilon = 0.039018$ Hartree is only 0.38 $\%$ smaller than the SVM $\varepsilon$ of 0.0391674 Hartree. ![ The PsH annihilation rate ($\langle \Gamma \rangle^{\infty}$ in units of 10$^{9}$ s$^{-1}$) as a function of $J$. The close to converged SVM annihilation rate is shown as the horizontal line. []{data-label="PsHGinf"}](PsHGinf.eps){width="8.5cm"} Figure \[PsHGinf\] shows the annihilation rate versus $J$. The CI calculation does not converge to the SVM annihilation rate since a radial basis of 15 LTOs per $\ell$ is simply too small. The $R^J_{\Gamma}$ entries in Table \[radialbasis\] reveal a 13.5$\%$ increase in $\Delta {\Gamma}^{9}$ between the basis 1 and basis 2 calculations. The conclusions that can be drawn from Figure \[PsH\] under such circumstances are somewhat limited. But Method III is again susceptible to the accuracy of the $\Delta \Gamma^J $ increments and again gives a final $\langle \Gamma \rangle^{\infty}$ that is smaller than Method I. Methods II and $p_{\rm av}$ again give final estimates of $\langle \Gamma \rangle^{\infty}$ that are close together. Method $p$ gives the largest estimate of $\langle \Gamma \rangle^{\infty}$ while Methods I and GL give the smallest. Another problem with Method III arose from the usage of the Davidson method to perform the matrix diagonalization. This only gives values of $\langle \Gamma \rangle^J$ that are stable to 6-7 significant digits and this leads to the irregularities in the $\langle \Gamma\rangle^{\infty}$ evident in Figure \[PsHGinf\]. The problem of decreased precision when using the Davidson algorithm had been previously noted in CI calculations of helium [@bromley06a] and may be generic to iterative matrix solvers. Comment on the scaling of the annihilation rate =============================================== There is one class of system that has not been studied in the present work, namely the close to threshold scattering of positrons from atoms that can bind a positron. The behavior of the $\Delta Z_{\rm eff}^J$ increments is complicated by a parametric dependence on the scattering length, $A$ [@mitroy02a; @bromley02e; @bromley03a]. Once $J$ is large enough to formally bind a positron, the magnitude of the scattering length decreases as $J$ increases. Since $Z_{\rm eff} \propto A^2$, the decrease in $A$ as $J$ increases impinges on the increase in $Z_{\rm eff}^J$ that would otherwise occur. Indeed, one of the reasons why calculations on $e^+$Cu were originally taken to $J = 18$ was to minimize the disruption that the scattering length had on $Z_{\rm eff}$ [@bromley02e; @bromley03a]. It was not worthwhile to try and analyze the behavior of the partial wave expansion for the $e^+$ + Cu scattering system with its two major complications, the effect of the radial basis set and, secondly, the effect of the $A$ versus $J$ variation. Such an investigation is best delayed until substantially larger basis sets can be deployed. Commentary on the work of Gribakin and Ludlow ============================================= The problems caused by the slow convergence of $Z_{\rm eff}$ can be exposed by a detailed analysis of the recent Gribakin and Ludlow [@gribakin04b] calculation of the annihilation rate for $e^+$-H scattering. This calculation used a variant of many-body perturbation theory (MBPT) to generate the initial set of $\langle Z_{\rm eff}\rangle^J$ data. They then did a fit to eq. (\[Glebfit2\]) over the $J = 7 \to 10$ interval and then plotted the results of that fit under the assumption that this demonstrated that their data obeyed eq. (\[Glebfit2\]). However, this procedure as applied by GL could hardly have been better suited to concealing deviations of the sequence of $\langle Z_{\rm eff}\rangle^J$ values from the leading order $(J+{\scriptstyle \frac{1}{2}})^{-2}$ term (note, in this section the GL acronym refers to the results presented in [@gribakin04b]). In effect, Figures 7 and 8 of GL are actually a demonstration of Taylor’s theorem, namely that any continuous function will approximate a straight line if examined over a sufficiently small domain. To illustrate this point, consider a sequence of synthetic $\langle Z_{\rm eff} \rangle^J$ data generated by the prescription, $\langle Z_{\rm eff}\rangle^7 = 2.5$ and $\Delta Z^J_{\rm eff} = 20 \times (J+{\scriptstyle \frac{1}{2}})^{-2.6}$ (such a data sequence could be generated exactly from a 3-term expansion of eq. (\[Zseries\])). A fit to this sequence was made using eq. (\[Glebfit2\]) and the results of that fit are shown in Figure \[synthetic\]. Even though the data were generated according to $p = 2.6$, a visual inspection (and it should be noted that Figure \[synthetic\] is much higher resolution than Figures 7 and 8 of GL) suggests that the data are in good agreement with a $p = 2$ power law decay whereas this is certainly not the case. ![ Plot of $Z^J_{\rm eff}$ vs $(J+{\scriptstyle \frac{1}{2}})^{-1}$ for the synthetic data as discussed in the text. The line shows the fit to eq. (\[Glebfit2\]) while the $\bullet$ give the synthetic data points with the $\Delta Z^J_{\rm eff} \sim (J+{\scriptstyle \frac{1}{2}})^{-2.6}$ dependence. []{data-label="synthetic"}](synthetic.eps){width="7.6cm"} The $p_Z$ exponent derived from the recent GL calculation [@gribakin04b] is depicted in Figure \[pHp\]. Although GL assume $p_Z = 2$ when making the extrapolation, the actual exponent extracted from their sequence of $Z_{\rm eff}$ values is $p = 2.233$ at $J = 10$. The present basis 2 calculation gives $p_Z = 1.885$ at $J = 10$. It is obvious that the GL calculation overestimates the rate at which $\Delta Z^J_{\rm eff}$ increments are decreasing. This implies that the GL estimates of $\Delta Z^J_{\rm eff}$ will be too small at higher $J$ and this is the case. GL get $\Delta Z^{10}_{\rm eff} = 0.0501$ while the CI-Kohn calculation gives $\Delta Z^{10}_{\rm eff} = 0.0604$. Consequently, the GL calculation gives $\Delta Z^{J>10}_{\rm eff} = 0.48$, while the basis 2 CI-Kohn calculation gives $\Delta Z^{J>10}_{\rm eff} \approx 0.62$. The assertion by GL that their calculation has converged to the region in which the $\Delta Z^J_{\rm eff} = B (J+{\scriptstyle \frac{1}{2}})^{-2}$ formula is valid is incorrect. Besides directly leading to less reliable extrapolation corrections, a related problem with the GL procedure is that it does not have the sensitivity to flag potential problems with the radial basis. A plot of $p_Z$ versus $J$ that crosses the $p_Z = 2$ line is a good indicator of some inadequacy in the basis. There is no indication that GL were aware that their $\Delta Z^{J}_{\rm eff}$ were decreasing much too quickly as $J$ increased, their statement that [*The use of a $B$-spline basis means that fast convergence is achieved with respect to the number of states with a particular angular momentum*]{} is difficult to reconcile with the present analysis. However, it should be noted that do indicate that they could improve the quality of their answers by [*“pushing harder” the numerics*]{}. The tendency for the GL calculation to overestimate the convergence of the annihilation rate increments probably does not arise from MBPT per se, rather it most likely comes from the underlying single electron basis. Besides the inherently slower convergence at higher $J$ mentioned earlier, another possibility is due to the confinement of the basis to a box of radius 15 $a_0$. Confining the basis in this way will result in mean excitation energies (e.g. for predicting the multi-pole polarizabilities) that will eventually increase as $\sim \ell^2$, where $\ell$ is the orbital angular momentum, while for a real H atom the mean excitation energy for any $\ell$ is less than 1.0 Hartree [@dalgarno56a]. Thus the occupancy of the higher $J$ orbitals, which contribute significantly to $Z_{\rm eff}$ will be inhibited, and successive $\Delta Z^{J}_{\rm eff}$ will decrease too rapidly with increasing $J$. The relevance of these issues is best illustrated by a comparison with the exact value of $Z_{\rm eff}$ which will be taken to be 3.407 at $k = 0.4 \ a_0^{-1}$. This is 0.295 larger than the GL value of 3.112. The underestimation of the higher partial wave contribution in the GL calculation, estimated at $0.14=0.62-0.48$ is responsible for about 50$\%$ of the existing discrepancy. The discrepancies of GL with the best calculations are are not that severe for $s$-wave scattering since imposition of the $p = 2$ condition for $J \ge 10$ prevents the inherent deficiencies in their $B$-spline basis from becoming too excessive. Also $s$-wave $e^+$-H scattering system is certainly one of the easier positron annihilation calculations. However, the inadequacy of the GL methodology manifests itself more severely in other positron annihilation situations. The CI expansion converges quicker for electron-positron annihilations that take place at small distances from the nucleus than for annihilations that take place at large distances [@mitroy02b; @bromley02e; @novikov04a]. The presence of the centrifugal barrier for $L > 0$ scattering leads to the electron-positron annihilations occurring further from the nucleus. Consequently the convergence problem is more serious for $p$-wave and $d$-wave scattering since a proportionally larger part of $Z_{\rm eff}$ comes from the $J \to \infty$ correction [@bromley03a; @novikov04a]. We have not repeated the earlier $p$-wave calculations [@bromley03a] with a larger radial basis, but a comparison with the CI-Kohn data at $k = 0.4 \ a_0^{-1}$ indicates that the GL calculation again underestimates the impact of the high $J$ orbitals. The CI-Kohn calculation reported in [@bromley03a] gave $\Delta Z^{10}_{\rm eff} = 0.0480$ while the GL calculation gave 0.0416. In addition, the CI-Kohn calculation gave $p = 1.852$ for $J = 10$ while the GL calculation gave $p = 2.152$ (it is likely that an infinite basis CI-Kohn calculation would have $p < 1.80$ at $J = 10$). The GL calculation (which gave $Z_{rm eff} = 1.607$) underestimates the $L = 1$ $Z_{\rm eff}$ by about 0.18 at $k = 0.4$ $a_0^{-1}$ (the $T$-matrix calculation gives 1.786 [@mitroy95a; @ryzhikh00a] while van Reeth [*et.al.*]{} gave 1.794 [@vanreeth97a; @vanreeth98a; @gribakin04b]). It is likely that at least 0.10 of the discrepancy will arise from orbitals with $J \ge 10$. One of the major results of the GL calculations was their demonstration that the enhancement factor is independent of energy. The enhancement factor can be defined as the factor that the annihilation rate, calculated as a simple product of the electron and positron densities needs to be increased in order to agree with the exact annihilation rate [@puska94; @mitroy02a; @novikov04a]. They (GL) based this conclusion solely on a forensic analysis of the $Z_{\rm eff}$ annihilation rate matrix element. Figure 13 of GL reveals that the variation of the $d$-wave enhancement factor with energy is noticeably larger than either the $s$-wave or $p$-wave enhancement factor [@gribakin04b]. Since a larger fraction of the $d$-wave $Z_{\rm eff}$ comes from $J \ge 10$, the possibility exists that this stronger energy dependence is due to extrapolation issues as opposed to dynamical effects. Although GL seem unaware of the result, the slow variation of the enhancement factor with energy had been demonstrated in a model potential analysis [@mitroy02a]. Comparisons of model potential calculations with ab-initio variational and polarized orbital calculations had shown that a model potential calculation tuned to reproduce the energy dependence of the phase shifts also gave the energy dependence of $Z_{\rm eff}$ [@mitroy02a]. The variation of the $d$-wave enhancement factor was determined by tuning a model potential to the large basis phase shifts of [@mitroy95b] and then normalizing to a similar calculation of $Z_{\rm eff}$ [@ryzhikh00a]. The variation in the $d$-wave enhancement factor over the energy range from $k = 0$ to 0.5 $a_0^{-1}$ was less than 4$\%$. Although the model potential result is not conclusive, it does appear that the variation in the $d$-wave $Z_{\rm eff}$ is less than that indicated by the GL calculation. Another area where application of the GL methodology could lead to larger than anticipated errors is in the determination the angular correlation or the $\gamma$ energy spectrum [@dunlop06a]. These two properties depend on the relative momentum of the annihilating electron-positron pair [@charlton85; @ryzhikh99a]. It is known from investigations of momentum space wave functions that the low momentum part of the wave function largely arise from the large $r$ part of the wave function while the high momentum properties come from the small $r$ part of the wave function [@mccarthy91a]. Under such circumstances, application of the GL method could easily result in errors to the $J \to \infty$ corrections that depend systematically on the $\gamma$-energy or recoil momentum. Summary and Conclusions ======================= Single center methods represent a superficially attractive method to study mixed electron positron systems since existing computer codes can be adapted without too much effort. The penalty associated with this approach is the slow convergence of the binding energy and, more noticeably, the annihilation rate with respect to the partial wave expansion of the single particle basis. The results presented here are generally consistent with the asymptotic limits derived from second-order perturbation theory by Gribakin and Ludlow [@gribakin02a]. The actual calculations at finite $J$ generally give increments to the energy (phase shift) and annihilation rate that decrease slightly slower than the GL limits, but the overall trends are compatible with the GL limits. The tendency for the convergence with respect to the radial basis size to slow down as $J$ increases does have implications for the design of any CI type calculation. Some sort of extrapolation in $J$ is necessary in order to determine the energy and more particularly the annihilation rate. But there is no point in making $J$ bigger if this is done at the expense of the radial basis set. One simply ends up with increments to the energy or annihilation rate which are systematically too small at higher $J$. This problem does not seem to be restricted to the LTO basis used in the present work. Convergence problems at high $J$ are also present for the Gribakin and Ludlow calculations which used a $B$-spline basis [@gribakin04b] and the Saito calculations which used a natural orbital basis [@saito03a; @saito05a; @mitroy05i]. It is amusing to note that the one of the first manifestations of this problem occurred over 40 years ago [@tycko58a; @schwartz62a]. The best methods for estimating the $J \to \infty$ corrections depends on the quality of the underlying calculation. For a low precision calculation, Method I would seem to be appropriate. A low precision calculation can probably be regarded as one with $p_E$ or $p_{\Gamma}$ exceeding 4 or 2 respectively when the $J \to \infty$ correction is evaluated (assuming that $p$ approaches its limiting value from below). Methods II or III would seem to be the preferred options for a high precision calculation. As a general principle, inclusion of the second term in the asymptotic series leads to improved $\langle X \rangle^{\infty}$ predictions when compared with asymptotic series with the single term series. Method III is more susceptible to imperfections in the radial basis and should not be applied to the calculation of the annihilation rate unless a very large radial basis set is employed. Irrespective of how the $J \to \infty$ corrections are evaluated, it is essential that the exponents $p$ relating the changes in the expectation values be examined as a test of the quality of the radial basis. The overall situation regarding the use of single center methods to compute positron-atom phase shifts or energies is that calculation to the sub 1$\%$ accuracy level is achievable for those systems that have a parent atom ionization potential greater than 0.250 Hartree. The $O((J + {\scriptstyle \frac{1}{2}})^{-4})$ convergence means a $J$ of 10 or slightly larger will generally suffice as long as the method used to perform the $J \to \infty$ correction is more sophisticated than those used previously. However, the situation with respect to the annihilation rate is much grimmer and it is not possible to guarantee 1$\%$ accuracy for even the simple $e^+$-H system. Here the $O((J + {\scriptstyle \frac{1}{2}})^{-2})$ convergence means the $J \to \infty$ correction is larger, and moreover the slow convergence with respect to the radial basis is further complicated by the fact that it is slower at high $J$ than low $J$. In this case, it appears that “God is on the side of the big basis set” [@boneparte]. This work was supported by a research grant from the Australian Research Council. The authors would like to thank Shane Caple for providing access to extra computing resources. Dr Gribakin kindly provided tabulations of their calculations of $Z^J_{\rm eff}$. [56]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , , , ****, (). , , , , , , ****, (). , ****, (). , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , ****, (). , , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , p. (). , p. (). , ****, (). , p. (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , p. (), <http://arxiv.org/physics/0512175>. , ****, (). , ****, (). , , , ****, (). .

ArXiv

--- abstract: 'A programmable gate array is a circuit whose action is controlled by input data. In this letter we describe a special–purpose quantum circuit that can be programmed to evaluate the expectation value of any operator $O$ acting on a space of states of $N$ dimensions. The circuit has a program register whose state $|\Psi(O)\rangle_P$ encodes the operator $O$ whose expectation value is to be evaluated. The method requires knowledge of the expansion of $O$ in a basis of the space of operators. We discuss some applications of this circuit and its relation to known instances of quantum state tomography.' author: - 'Juan Pablo Paz$^{1,2}$ and Augusto Roncaglia$^1$' title: A quantum gate array can be programmed to evaluate the expectation value of any operator --- An important feature of classical computers is that they can be programmed. That is to say, a fixed universal device can perform different tasks depending on the state of some input registers. These registers define the program the device is executing. Quantum computers [@NielsenChuang] have a rather different property: Thus, Nielsen and Chuang established in [@NielsenChuangPRL] that a general purpose programmable quantum computer does not exist. Such a device would have to have the following features: It should consist of a fixed gate array with a data register and a program register. The array should work in such a way that the state of the program register encodes the unitary operator ${\cal U}$ that is applied to the state of the data register. As shown in [@NielsenChuangPRL], such devices cannot be universal since different unitary operators require orthogonal states of the program register. However, some interesting examples of programmable devices could still be constructed. For example, non-deterministic programmable gate arrays were first considered in [@NielsenChuangPRL] and analyzed later in a variety of examples [@ProgramableArrays]. More recently, quantum “multimeters” were introduced and discussed in [@DusekBuzek]. Such devices are fixed gate arrays acting on a data register and a program register, together with a final fixed projective measurement on the composite system. They are programmable quantum measurement devices [@DusekBuzek] that act either non–deterministically or in an approximate way (see [@FiurasekDusekFilip]). In this letter we will describe a different kind of programmable quantum gate array that is useful to solve the following problem: Suppose that we are given an operator $O$ acting on an $N$ dimensional Hilbert space and a quantum state $\rho$. By this we mean that someone supplies us with many copies of a quantum system prepared in the same state $\rho$ and defines for us the operator $O$ by specifying its expansion in a basis of the space of operators. Our task is to compute the expectation value of $O$ in the state $\rho$. We will show that it is possible to construct a programmable circuit that evaluates such expectation value by measuring the polarization of a single qubit. The inputs of such circuit are a data register, a program register and an auxiliary qubit. The circuit evaluates the expectation value of an operator $O$ (specified by the program register) in the quantum state $\rho$ of the data register. The expectation value ${\rm Tr}(\rho O)$ is obtained by performing a measurement of the polarization of the auxiliary qubit. We will describe how to construct these circuits and exhibit an interesting example: a programmable array to efficiently solve a class of quantum decision problems concerning properties of quantum phase space distributions. The quantum gate arrays discussed in this letter are designed using the “scattering circuit” shown in Figure 1 as a simple primitive. In such circuit a system, initially in the state $\rho$, is brought in contact with an ancillary qubit prepared in the state $|0\rangle$. This ancilla acts as a probe particle in a scattering experiment. The algorithm consist of the following steps: i) Apply a Hadamard transform $H$ to the ancillary qubit. Since $H|0\rangle=(|0\rangle+|1\rangle)/\sqrt 2$, $H|1\rangle=(|0\rangle-|1\rangle)/\sqrt 2$, the new state of the qubit is $(|0\rangle+|1\rangle)/\sqrt 2$. ii) Apply a “controlled–$A$” operator, which does nothing if the state of the ancilla is $|0\rangle$ and applies the unitary operator $A$ to the system if the ancilla is in state $|1\rangle$, iii) Apply another Hadamard gate to the ancilla and perform measurements of its spin polarizations along the $z$- and $y$-axes. Given sufficiently many independent instances of the experiment, the measurements yield the expectation values $\langle\sigma_z\rangle$ and $\langle\sigma_y\rangle$ of the Pauli spin operators $\sigma_z$ and $\sigma_y$. This algorithm has the following remarkable property: $$\langle\sigma_z\rangle=\mbox{Re}[ \, \mbox{Tr}( A \rho) \, ], \ \langle\sigma_y\rangle=\mbox{Im}[ \, \mbox{Tr}( A \rho) \, ]. \label{polarization}$$ Different versions of this circuit play an important role in many quantum algorithms [@kitaev:qc1997a; @cleve:qc1997b; @Lloyd99; @Knill98]. In particular, the scattering circuit was recently used as a basic tool to interpret tomography and spectroscopy as two dual forms of the same quantum computation [@NatureUS]. For our purpose it is useful to review how to use this scattering circuit as a primitive to design a tomographer (i.e., a device that after a number of experiments determines the quantum state of the system). As a consequence of equations (\[polarization\]) we see that every time we run the algorithm for a known operator $A$, we extract information about the state $\rho$. Doing so for a complete basis of operators $\{A(\alpha)\}$ one gets complete information and determines the full density matrix. Different tomographic schemes are characterized by the basis of operators $A(\alpha)$ they use. Of course, completely determining the quantum state requires an exponential amount of resources. In fact, if the dimensionality of the Hilbert space of the system is $N$ then the complete determination of the quantum state involves running the scattering circuit for a complete basis of $N^2$ operators $A(\alpha)$. However, evaluating any coefficient of the decomposition of $\rho$ in a given basis can be done efficiently provided that the operators $A(\alpha)$ can be implemented by efficient networks. A convenient basis set is defined as (see, for example, [@NatureUS; @MPS02]): $$A(\alpha)= A(q,p)= U^q R V^{-p} \exp(i\pi pq/N).\label{aqp}$$ Here, both $q$ and $p$ are integers between $0$ and $N-1$, $U$ is a cyclic shift operator in the computational basis ($U|n\rangle=|n+1\rangle$), $V$ is the cyclic shift operator in the basis related to the computational one via the discrete Fourier transform, and $R$ is the reflection operator ($R|n\rangle=|N-n\rangle$). It is straightforward to show that the operators $A(q,p)$ are hermitian, unitary and form a complete orthonormal basis of the space of operators satisfying $$\mbox{Tr}[A(\alpha) A(\alpha') ] = N \, \delta_N(q'-q) \delta_N(p'-p),$$ where $\delta_N(x)$ is the periodic Kronecker delta function that is equal to one if $x=0$ (modulo $N$) and vanishes otherwise (the above operators form a ”quorum”, as defined in [@Dariano1]). With this choice for $A(q,p)$ the scattering circuit directly evaluates the discrete Wigner function [@NatureUS; @Wooters; @Leonhardt; @MPS02]. 3.2in We will now show how to design a programmable gate array to evaluate the expectation value of any operator $O$. We will assume that we know how to expand $O$ in a basis such as the one used above: $O=\sum_{q,p} o(q,p) A(q,p)$. As the operators $A(q,p)$ are not only unitary but also hermitian, the real and imaginary parts of the complex coefficients $o(q,p)$ define the expansion of the hermitian and anti-hermitian pieces of $O$ in the basis $A(q,p)$. The expectation value of these two pieces can be evaluated separately using the procedure described below (the results can then be combined to get the expectation of $O$). So, in what follows we will assume that the operator at hand is hermitian and that the coefficients of its expansion in the basis $A(q,p)$ are real numbers. To introduce our method, it is convenient to notice first that the evaluation of the expectation value of the operators $A(q,p)$ can be done using a programmable circuit that is independent of $q$ and $p$. Such circuit is illustrated in Figure 2. This is an application of the scattering circuit shown in Figure 1 with two program registers used to encode the value of $q$ and $p$. When the quantum state of the program is $|\Psi\rangle_P=|q\rangle|p\rangle$ the circuit evaluates the expectation value of $A(q,p)$. This is accomplished by letting the program registers to act as controls of the operators $U$ and $V$ (which, as mentioned, generate cyclic shifts either in the computational or in the conjugated basis). Thus, the action of the circuit is such that when the auxiliary qubit is in state $|\sigma\rangle$ ($\sigma=0,1$) and the state of the program is $|q\rangle|p\rangle$, the operator $A^\sigma(q,p)$ is applied to the system register. The network in Figure 2 is efficient since it can be built using a number of elementary gates which scales polynomially with $\log(N)$ [@MPS02]. The circuit has an obvious property: Different states $|q\rangle|p\rangle$ are used to program the evaluation of the expectation value of orthogonal operators $A(q,p)$. It is clear that by restricting to such program states one has no real advantage with respect to the case in which $q$ and $p$ are stored as classical information. However, we can use more general program states: If the program register is in the state $|\Psi\rangle_{P}=\sum_{q,p}c(q,p) |q\rangle|p\rangle$ then the same circuit evaluates the expectation value of a linear combination of the operators $A(q,p)$ since the final polarization is: $$\langle\sigma_z\rangle={\rm Tr}\left(\rho \sum_{q,p} |c(q,p)|^2 A(q,p) \right) \label{program}$$ Equation (\[program\]) shows that this algorithm can be used to evaluate the expectation value of any operator that can be written as a convex sum of the basis set $A(q,p)$. This is not general enough since the expansion of a hermitian operator can include negative coefficients. For this purpose the method can be extended as follows: The most general hermitian operator can be expressed in the basis $A(q,p)$ as $O=\sum_{q,p} c^2(q,p) \exp(i\pi\phi(q,p))A(q,p)$, where $c(q,p)$ is a real number and $\phi(q,p)$ is either $0$ or $1$ ($\phi(q,p)$ simply stores the information about the sign of each coefficient). We will assume that $\sum_{q,p}c^2(q,p)= 1$ (if this is not the case we can always renormalize the coefficients). For operators with some negative coefficients we can use a third register consisting of a single qubit to store $\phi(q,p)$. The circuit evaluating the expectation value of $O$ is shown in Figure 3. The first two program registers store $q$ and $p$ and are used exactly in the same way as above. The third one, storing $\phi(q,p)$, is acted upon with a $\sigma_z$ operator, introducing the required phase $\exp(i\pi\phi(q,p))$. Then, if the state of the program register is $|\Psi\rangle_P=\sum_{q,p}c(q,p)|q\rangle|p\rangle|\phi(q,p)\rangle$. the final polarization measurement turns out to be $$\langle\sigma_z\rangle={\rm Tr}\left(\rho \sum_{q,p} c^2(q,p) {\rm e}^{i\pi\phi(q,p)}A(q,p)\right)={\rm Tr}\left(\rho O\right). \label{program2}$$ 3.2in 3.2in Summarizing, we showed that the measurement of the expectation value of any operator $O$ can be done using a programmable gate array. The hardware architecture is associated with the particular choice of basis $A(q,p)$, which is just a matter of convenience, and is independent of $O$. The software used to program the array is obviously determined by the choice of hardware. The expectation values of the hermitian and anti–hermitian parts of the operator $O$ are computed separately using a method that requires knowledge of the expansion of these operators in the basis $A(q,p)$. If the coefficients in the expansion are written as $o(q,p)= {\rm Tr}(O A(q,p))/N=c^2(q,p)\exp(i\pi\phi(q,p))$, the program state that needs to be prepared is $|\Psi\rangle_P=\sum_{q,p}c(q,p)\ |q\rangle|p\rangle|\phi(q,p)\rangle$ (where $\phi(q,p)=0$ or $1$). Thus, the coefficients of the expansion of $O$ in the basis $A(q,p)$ define the program state $|\Psi\rangle_P$ required to measure its expectation value. It is clear that in most cases the method will not be efficient. For example, both the task of defining the operator by specifying the coefficients $o(q,p)$ as well as the preparation of the program state $|\Psi\rangle_P$ are likely to be inefficient. The existence of efficient networks to implement “controlled–$A(q,p)$” operations is a less stringent conditions that is fulfilled by the basis defined in (\[aqp\]) [@MPS02]. Having said this, it is worth noticing that there are sets of problems that can be efficiently solved using this method. We will now describe one such example. We will show that the circuit of Figure 2 can be easily adapted to evaluate the sum of values of the Wigner function over various phase space domains. The program register is used to define the domain over which the Wigner function is averaged. If the domain is a line, the algorithm just evaluates the probability for the occurrences of the results of the measurement of a family of observables (see below). However, for more general domains (such as line segments, parallelograms, etc) the circuit evaluates properties that characterize a quantum state that cannot be simply casted in terms of probabilities. In this sense the circuit is as a programmable tomographer measuring various features of phase space distributions. Before going into more details let us briefly review some properties of Wigner functions. Discrete Wigner functions can be used to represent the quantum state of a system in phase space [@NatureUS; @Wooters; @Leonhardt; @MPS02]. For a system with an $N$ dimensional space of states such function is defined on a lattice of $2N\times 2N$ points $(q,p)$ where both $q$ and $p$ take values between $0$ and $2N-1$ (only $N\times N$ of these values are independent). At each phase space point the Wigner function is defined in terms of the operators $A(q,p)$ given in (\[aqp\]) as $$W(q,p)={1\over 2N} {\rm Tr}(A(q,p)\rho). \label{wignerdef}$$ As mentioned above, the measurement of this function can be done by using the scattering circuit [@NatureUS] or the programmable gate array of Figure 2. The program register encodes the value of $q$ and $p$. In general, if the program state is $|\Psi\rangle_P=\sum_{q,p}c(q,p)|q\rangle|p\rangle$, the final polarization measurement is $\langle\sigma_z\rangle=2N \sum_{qp} |c(q,p)|^2 W(q,p)$. Thus, the program defines the region over which we average the value of the Wigner function. In general, preparing the program state associated with a general phase space region can be complicated. However, there are simple procedures to prepare the program states corresponding to general lines, segments and parallelograms. Let us begin with the simplest case: For the program state $|\Psi\rangle_P=\sum_q|q\rangle|p_0\rangle/\sqrt{2N}$, it is clear that the final polarization measurement reveals the sum of values of the Wigner function along the horizontal line defined as $p=p_0$: i.e., $\langle\sigma_z\rangle=\sum_q W(q,p_0)$. To consider more general lines one can notice that the horizontal line $p=c$ can be mapped into one satisfying the equation $p-bq=c$ (mod $2N$, with $b$ and $c$ integers between $0$ and $2N$) by applying a linear area preserving map. Unitary operators corresponding to quantizations of such maps (known as ”quantum cat maps”) act classically in phase space. Thus, Wigner function flows from one point to another according to the classical transformation. Using this feature, one can compute the sum of the Wigner function along any tilted line as follows: One first transforms the state with the appropriate cat map and later evaluates the sum of the Wigner function along a vertical or horizontal line. Moreover, the method can be made fully programmable by adding extra registers to store the integers $b$ and $c$ parametrizing any cat map. Evaluating sums of Wigner functions over phase space lines is particularly interesting because of a crucial property of such functions: Thus, adding $W(q,p)$ along the line $ap-bq=c$ one obtains the probability to detect the eigenstate of the translation operator $T(b,a)=U^aV^b\exp(i\pi ab/N)$ with eigenvalue $\exp(i\pi c/N)$. As a consequence, in this case the programmable gate array evaluates probabilities for the possible results of a set of measurements. Lines are a special case since more general phase space domains cannot be associated with projection operators. However, it is clear that the method described above can be applied to efficiently prepare program states for other phase space domains such as general (tilted) parallelograms: One first trivially prepare the program state for a parallelogram limited by vertical and horizontal segments and later tilt it applying the strategy based on the use of cat maps. Other simple phase space regions can also be programmed using variations of this method. It is also interesting to notice that using variations of the circuit shown in Figure 3 we can also subtract values of the Wigner function in different phase space regions (which could be useful if one is interested in comparing their values). In this letter we established the existence of a gate array that can be programmed to evaluate the expectation value of any operator acting on an $N$ dimensional Hilbert space. The expectation value is obtained by measuring the polarization of a single auxiliary qubit. As an example, we showed how to program the evaluation of sums of values of the discrete Wigner function over various simple phase space domains. It is important to mention that our method is only efficient to determine if the sum of the Wigner function in a phase space domain (with up to $o(N)$ points) is greater than a fixed, $N$–independent, threshold (since this does not require exponential precision). This is a “quantum decision problem” whose input data (encoded in the system’s state $\rho$) is inherently quantum. Due to the nature of the input data, this problem cannot even be formulated on a classical computer. Interest on problems with quantum input data have recently increased, partly due to their significance in connection with the potential detection of entanglement [@Horodecki; @Ekert; @Horodecki2] as well as their relation with tomographic problems like the one described here. The extension of some of the above results to continuous variables is still under investigation [@Horodecki2]. After completing this work we became aware of the related approach to the construction of quantum universal detectors presented in [@Dariano2]. We acknowledge useful discussion with Marcos Saraceno. This work was partially supported with grants from Ubacyt, Anpcyt, Conicet and Fundación Antorchas. , M. Nielsen and I. Chuang (2000), Cambridge University Press. M. Nielsen and I. Chuang, [*Phys. Rev. Lett.*]{} [**79**]{}, 321-324 (1997). G. Vidal L. Masanes and J.I. Cirac, Phys. Rev. Lett. [**88**]{}, 047905 (2002); M. Hillery, V. Buzek and M. Ziman, [*Phys. Rev.*]{} [**A65**]{} 012301 (2002); S.F. Huelga, J.A. Vaccaro, A. Cheflies and M.B. Plenio, [*Phys. Rev.*]{} [**A63**]{}, 042303 (2001). M. Dusek and V. Buzek, Phys. Rev. [**A66**]{}, 022112 (2002). J. Fiurasek, M. Dusek and R. Filip, Phys. Rev. Lett. [**89**]{}, 190401 (2002). A. Yu. Kitaev, [*Quantum measurements and the Abelian Stabilizer Problem*]{}, quant-ph/9511026 R. Cleve, A. Ekert, C. Macchiavello and M. Mosca, Proc. Roy. Soc. Lond. A [**454**]{} (1998), 339 D. Abraham and S. Lloyd, Phys. Rev. Lett. [**83**]{}, 5162-5165 (1999). E. Knill and R. Laflamme, Phys. Rev. Lett. [**81**]{}, 5672-5675 (1999). G.M. D’Ariano and P. Lo Presti, [*Phys. Rev. Lett.*]{} [**86**]{}, 4195 (2001). C. Miquel, J.P. Paz, M. Saraceno, E. Knill, R. Laflamme and C. Negrevergne, [*Nature*]{} [**418**]{}, 59–62 (2002). W. K. Wooters, Ann. Phys. NY [**176**]{}, 1 (1986). U.Leonhardt, [*Pys. Rev.*]{} [**A 53**]{}, 2998–3013 (1996). C. Miquel, J.P. Paz and M. Saraceno, [*Phys. Rev.*]{} [**A 65**]{}, 062309 (2002); J.P. Paz, [*Phys. Rev.*]{} [**A 65**]{}, 062311 (2002). P. Horodecki and A. Ekert, [*Direct detection of quantum entanglement*]{}, e–print available in quant-ph/0111064. A. Ekert, C. Moura–Alves, D. Oi, M. Horodecki, P. Horodecki and L. Kwek, Phys. Rev. Lett. [**88**]{}, 217901 (2002). P. Horodecki, [*Mean of continuous variable observable via measurements on single qubit*]{}, e–print available in quant-ph/0210161. G.M. D’Ariano, P. Perinotti and M. F. Sacchi, arXiv:quant-ph/0306025.

ArXiv

--- abstract: 'We study the phase diagram and the pion modes in the electrically neutral two flavor PNJL model. One of the main result of this paper is that in the model with massive quarks, when electrical neutrality is required, pions do not condense in the ground state of the model: the isospin chemical potential $\mu_I = -\mu_e/2$ is always smaller than the value required for pion condensation to occur. Moreover we investigate on the pions and $\sigma$ mass spectra. We find that the qualitative behavior of the masses resembles that obtained in the NJL model. We close this paper by studying the intriguing possibility that bound states with the quantum numbers of the pions can be formed above the chiral phase transition.' author: - 'H. Abuki' - 'M. Ciminale' - 'R.Gatto' - 'N. D. Ippolito' - 'G. Nardulli' - 'M. Ruggieri' title: Electrical neutrality and pion modes in the two flavor PNJL model --- Introduction ============ The use of effective models to understand the phases of Quantum Chromodynamics (QCD) is nowadays a popular tool. Among them, the Nambu-Jona Lasinio model (NJL) is widely used since it allows for a simple discussion of chiral symmetry breaking [@Nambu:1961tp; @revNJL]. The main defect of the NJL model is that it completely ignores glue dynamics: gluons are not introduced in the lagrangian, and the QCD quark-quark interaction mediated by gluons is replaced by an effective four fermion interaction. An improvement of the NJL model is provided by the introduction of a background temporal static gluon field, coupled to quarks via the QCD covariant derivative. The gluon field is related to the expectation value of Polyakov loop $\Phi$ [@Polyakovetal], and the new model is called Polyakov-NJL (PNJL in the following) [@Meisinger:1995ih; @Fukushima:2003fw; @Ratti:2005jh; @Roessner:2006xn; @Ghosh:2007wy; @Kashiwa:2007hw; @Schaefer:2007pw; @Ratti:2007jf; @Sasaki:2006ww; @Megias:2006bn; @Zhang:2006gu; @Ciminale:2007ei; @Fu:2007xc; @Ciminale:2007sr] It is well known that the Polyakov loop serves as an order parameter for confinement $\rightarrow$ deconfinement transition in the pure glue theory [@Polyakovetal]. In more detail, the confined phase is characterized by $\langle\Phi\rangle=0$ and a $Z_3$ symmetry; on the other hand in the deconfined phase $\langle\Phi\rangle\neq0$ and the symmetry $Z_3$ is broken. In presence of dynamical quarks, $Z_3$ symmetry is explicitly broken and $\Phi$ can not be used as an order parameter. Nevertheless it is commonly used as an indicator of the deconfinement transition. In the PNJL model an effective potential for $\Phi$ is added by hand to the quark lagrangian, and $\Phi$ is coupled to the quarks via the QCD covariant derivative. The value of $\langle\Phi\rangle$ in the ground state, as well as other quantities of interest like the constituent quark masses, are obtained by minimization of the thermodynamic potential, or equivalently by solving coupled gap equations as in the NJL case. Despite the apparent complication due to the increase of the degrees of freedom, the PNJL model offers a better description of QCD then NJL does since it allows to derive, in the framework of field theory, several results obtained in lattice simulations [@Ratti:2005jh; @Roessner:2006xn; @Ghosh:2007wy; @Kashiwa:2007hw; @Ratti:2007jf; @Sasaki:2006ww; @Zhang:2006gu]. The knowledge of the phase diagram of the PNJL model as well as of its bulk properties is of a certain interest. In this paper we study the phase diagram and the pion modes in the electrically neutral two flavor PNJL model. The motivation of our study is straightforward: we are interested to investigate the equilibrium ground state of the model, and in the ground state a net electric charge can not be present, whichever the value of the temperature and of the quark chemical potential is. We consider the possibility of chiral as well as pion condensation in the ground state, at zero and non zero quark chemical potential. A related study without electrical neutrality and at zero chemical potential has been done in Ref. [@Zhang:2006gu]; see also Ref. [@Ebert:2005wr] for a similar study in the framework of the electrically neutral NJL model in the chiral limit. One of the results of our work is that when electrical neutrality is required, pion do not condense in the ground state of the model: the isospin chemical potential $\mu_I = -\mu_e/2$, with $\mu_e$ the electron chemical potential, is always smaller than the value required for pion condensation to occur. Moreover we investigate on the pions and $\sigma$ mass spectra. We find that the qualitative behavior of the masses resembles that obtained in the NJL model. We close this paper by studying the possibility that a bound state with the quantum numbers of the pions can be formed above the chiral phase transition. The plan of the paper is as follows: in Section \[Sec:for\] we briefly describe the PNJL model; in Section \[Sec:pd\] we show the phase diagram of the model, and some results for quantities of interest (in medium quark mass, pion condensate, mean value of the Polyakov loop, electron chemical potential). In Section \[Sec:PGm\] we compute the masses of the pseudo-Goldstone modes of the model and of the $\sigma$ mode. Finally in Section \[Sec:con\] we draw our conclusions. Formalism\[Sec:for\] ==================== The Lagrangian density of the two flavor PNJL model is given by [@Fukushima:2003fw] $${\cal L}^\prime= \bar{e}(i\gamma_\mu\partial^\mu)e + \bar\psi\left(i\gamma_\mu D^\mu + \mu\gamma_0 -m\right)\psi + G\left[\left(\bar\psi \psi\right)^2 + \left(\bar\psi i \gamma_5 \vec\tau \psi\right)^2\right] - {\cal U}[\Phi,\bar\Phi,T]~. \label{eq:LagrP}$$ In the above equation $e$ denotes the electron field; $\psi$ is the quark spinor with Dirac, color and flavor indices (implicitly summed). $m$ corresponds to the bare quark mass matrix; we assume from the very beginning $m_u = m_d$. The covariant derivative is defined as usual as $D_\mu = \partial_\mu -i A_\mu$. The gluon background field $A_\mu=\delta_{0\mu}A_0$ is supposed to be homogeneous and static, with $A_0 = g A_0^a T_a$ and $T_a$, $a=1,\dots,8$ being the $SU(3)$ color generators with the normalization condition $\text{Tr}[T_a T_b]=\delta_{ab}$. $\vec{\tau}$ is a vector of Pauli matrices in flavor space. Finally $\mu$ is the chemical mean quark chemical potential, related to the conserved baryon number. In Eq.  $\Phi$, $\bar\Phi$ correspond to the normalized traced Polyakov loop $L$ and its hermitian conjugate respectively, $\Phi=\text{Tr}L/N_c$, $\bar\Phi=\text{Tr}L^\dagger/N_c$, with $$L={\cal P}\exp\left(i\int_0^\beta A_4 d\tau\right)=\exp\left(i \beta A_4\right)~,~~~~~A_4=iA_0~,$$ and $\beta=1/T$. The term ${\cal U}[\Phi,\bar\Phi,T]$ is the effective potential for the traced Polyakov loop; in absence of dynamical quarks it is built in order to reproduce the pure glue lattice data of QCD, namely thermodynamical quantities (pressure, entropy and energy density) and the deconfinement temperature of heavy (non-dynamical) quarks, $T= 270$ MeV. Several forms of this potential have been suggested in the literature, see for example [@Fukushima:2003fw; @Ratti:2005jh; @Roessner:2006xn; @Ghosh:2007wy]. In this paper we adopt the following logarithmic form [@Roessner:2006xn], $$\frac{{\cal U}[\Phi,\bar\Phi,T]}{T^4} = -\frac{b_2(T)}{2}\bar\Phi\Phi + b(T)\log\left[1-6\bar\Phi\Phi + 4(\bar\Phi^3 + \Phi^3) -3(\bar\Phi\Phi)^2\right]~,\label{eq:Poly}$$ with $$b_2(T) = a_0 + a_1 \left(\frac{\bar T_0}{T}\right) + a_2 \left(\frac{\bar T_0}{T}\right)^2~,~~~~~b(T) = b_3\left(\frac{\bar T_0}{T}\right)^3~.\label{eq:lp}$$ Numerical values of the coefficients are as follows [@Roessner:2006xn]: $$a_0=3.51~,~~~a_1 = -2.47~,~~~a_2 = 15.2~,~~~b_3=-1.75~.$$ If dynamical quarks were not present then one should chose $\bar T_0 = 270$ MeV in order to reproduce the deconfinement transition at $T = 270$ MeV. In presence of quarks $\bar T_0$ gets a dependence on the number of active flavors, as shown in Refs. [@Ratti:2005jh; @Schaefer:2007pw]. Following Ref. [@Schaefer:2007pw] we chose $\bar T_0 = 208$ MeV in Eq. , which is appropriate to deal with two degenerate flavors. As explained in the Introduction we are interested to study the ground state of the model specified by the Lagrangian in Eq. , at each value of the temperature $T$ and the chemical potential $\mu$, corresponding to a vanishing total electric charge. To this end we introduce electrons in Eq.  since a net number of electrons could be needed, beside the pion condensate (if any, see below), to ensure the electrical neutrality of the ground state. In order to build the neutral ground state we work as usual in the gran canonical ensemble formalism, adding to Eq.  the term $\mu_Q N_Q$, $\mu_Q$ being the chemical potential for the total charge $N_Q$, and requiring stationarity of the thermodynamic potential with respect to variations of $\mu_Q$, which is equivalent to the requirement $<N_Q>=0$ in the ground state. To be more specific, since the total charge operator is given in terms of the quark and electron fields by $$N_Q = \frac{2}{3}N_u - \frac{1}{3}N_d - N_e = \frac{2}{3}{u^\dagger}u - \frac{1}{3}{d^\dagger}d -{e^\dagger}e~,$$ it is easy to recognize that the lagrangian ${\cal L}$ in the gran canonical ensemble ${\cal L} = {\cal L}^\prime + \mu_Q N_Q$ can be written as $${\cal L}=\bar{e}(i\gamma_\mu\partial^\mu + \mu_e \gamma_0)e + \bar\psi\left(i\gamma_\mu D^\mu + \hat\mu\gamma_0 -m\right)\psi + G\left[\left(\bar\psi \psi\right)^2 + \left(\bar\psi i \gamma_5 \vec\tau \psi\right)^2\right] - {\cal U}[\Phi,\bar\Phi,T]~, \label{eq:Lagr}$$ where $\mu_e = - \mu_Q$ and the quark chemical potential matrix $\hat\mu$ is defined in flavor-color space as $$\hat\mu=\left(\begin{array}{cc} \mu-\frac{2}{3}\mu_e & 0 \\ 0 & \mu + \frac{1}{3}\mu_e \\ \end{array}\right)\otimes\bm{1}_c~,\label{eq:chemPot}$$ where $\bm{1}_c$ denotes identity matrix in color space. In this paper we work in the mean field approximation. In order to study chiral symmetry breaking and to allow for pion condensation we assume that in the ground state the expectation values for the following operators may develop [@Zhang:2006gu; @Ebert:2005wr], $$\sigma = \left<\bar\psi \psi\right>~,~~~~~\pi = \left<\bar\psi i \gamma_5 \tau_1 \psi\right>~.\label{eq:condensates}$$ In the above equation a summation over flavor and color is understood. We have assumed that the pion condensate aligns along the $\tau_1$ direction in flavor space. This choice is not restrictive. As a matter of fact we should allow for independent condensation both in $\pi^+$ and in $\pi^-$ channels [@Zhang:2006gu]: $$\pi^+\equiv\langle\bar\psi i \gamma_5 \tau_+ \psi\rangle = \frac{\pi}{\sqrt{2}}e^{i\theta}~,~~~~~\pi^-\equiv\langle\bar\psi i \gamma_5 \tau_- \psi\rangle= \frac{\pi}{\sqrt{2}}e^{-i\theta}~,$$ with $\tau_\pm = (\tau_1\pm\tau_2)/\sqrt{2}$; but the thermodynamical potential potential is not dependent on the phase $\theta$, therefore we can assume $\theta=0$ which leaves us with $\pi^+ = \pi^- = \pi/\sqrt{2}$ and introduce only one condensate, specified in Eq. . In the mean field approximation the PNJL lagrangian reads $${\cal L}=\bar{e}(i\gamma_\mu\partial^\mu + \mu_e \gamma_0)e + \bar\psi\left(i\gamma_\mu D^\mu + \hat\mu\gamma_0 -M - N i\gamma_5\tau_1\right)\psi - G\left[\sigma^2 + \pi^2\right] - {\cal U}[\Phi,\bar\Phi,T]~, \label{eq:LagrMF}$$ where the in-medium quark mass $M$ and the pion field $N$ are related to $\sigma$ and $\pi$ by means of the relations $$\begin{aligned} M&=&m-2G\sigma~,\\ N&=&-2G\pi~.\end{aligned}$$ The thermodynamical potential $\Omega$ can be obtained by integration over the fermion fields in the partition function of the model, see for example Ref. [@Ebert:2000pb], $$\Omega = -\left(\frac{\mu_e^4}{12\pi^2} + \frac{\mu_e^2 T^2}{6} + \frac{7\pi^2 T^4}{180}\right) + {\cal U}[\Phi,\bar\Phi,T] + G\left[\sigma^2 + \pi^2\right] - T\sum_n\int_0^\Lambda\frac{d^3{\bm p}}{(2\pi)^3}~\text{Tr}~\text{log}\frac{S^{-1}(i\omega_n,{\bm p})}{T}~,$$ where the sum is over fermion Matsubara frequencies $\omega_n = \pi T(2n+1)$, and the trace is over Dirac, flavor and color indices. $\Lambda$ is an ultraviolet cutoff to ensure convergence of the momentum integral. The inverse quark propagator is defined as $$S^{-1}(i\omega_n,{\bm p})= \left(\begin{array}{cc} (i\omega_n+\mu-\frac{2}{3}\mu_e+iA_4)\gamma_0 -{\bm\gamma}\cdot{\bm p} -M & -i\gamma_5 N \\ -i\gamma_5 N & (i\omega_n+\mu+\frac{1}{3}\mu_e+iA_4)\gamma_0 -{\bm\gamma}\cdot{\bm p} -M\\ \end{array}\right)\otimes{\bm 1}_c~.\label{eq:po}$$ The ground state of the model is defined by the values of $\sigma$, $\pi$, $\Phi$, $\bar\Phi$ that minimize $\Omega$ and that have a vanishing total charge; the latter condition is equivalent to the requirement $$\frac{\partial\Omega}{\partial\mu_e}=0~.$$ The parameters $m$, $G$ and $\Lambda$ are given by [@Roessner:2006xn] $$m=5.5~\text{MeV}~,~~~~~G=5.04~\text{GeV}^{-2}~,~~~~~\Lambda=650.9~\text{MeV}~,$$ which fix, at zero temperature and zero chemical potential, the pion mass $m_\pi=139.3$ MeV, the pion decay constant $f_\pi=92.3$ MeV and the chiral condensate $\langle\bar u u\rangle = -(251~\text{MeV})^3$. Phase diagram of the model\[Sec:pd\] ==================================== ![\[FIG:pd\]Phase diagram of the electrically neutral two flavor PNJL model. Dashed line corresponds to the chiral crossover; solid line describes the first order chiral transition. At $\mu=0$ the critical point is located at $T=206$ MeV; at $T=0$ the chiral transition is found at $\mu=348$ MeV. The black dot denotes the critical end point (CEP), located at $(\mu_E,T_E) = (342,79)$ MeV.](fig1.eps){width="8cm"} In Fig. \[FIG:pd\] we plot the phase diagram of the electrically neutral two flavor PNJL model. Dashed line corresponds to the chiral crossover; solid line describes the first order chiral transition. At each value of $\mu$ the crossover is identified with the inflection point of the chiral condensate. Analogously the first order transition is defined by the discontinuity of $\sigma$. In the region below the lines the chiral symmetry is broken, above the lines it is restored. At $\mu=0$ the chiral symmetry is restored at $T=206$ MeV. For comparison, the inflection point of the Polyakov loop (which is commonly associated to the deconfinement crossover) at $\mu=0$ is located at $T=180$ MeV. Moreover, at $T=0$ the chiral phase transition is of first order and is found at $\mu=348$ MeV. The black dot denotes the critical end point (CEP), located at $(\mu_E,T_E) = (342,79)$ MeV. As explained later, we find a vanishing pion condensate once electrical neutrality is required. ![\[FIG:pd2\]Comparison of the first order transitions of the PNJL model with neutrality (called Model I in the text), of the PNJL model with $\mu_e=0$ (model II) and of the NJL model with electrical neutrality (model III).](phasediagram.eps){width="10cm"} It is instructive to compare the PNJL results discussed above with the electrically neutral NJL model, as well as the PNJL model with $\mu_e=0$, in order to emphasize the role of electrical neutrality and of the Polyakov loop. To this end in Fig. \[FIG:pd2\] we compare the first order transitions of the PNJL model with neutrality (model I), of the PNJL model with $\mu_e=0$ (model II) and of the NJL model with electrical neutrality (model III). Comparison of models I and II shows that adding the electron chemical potential and requiring electrical neutrality enlarges of some MeV the $\mu$ window of the chirally broken phase. Moreover, comparison of models I and III shows that adding the self-consistently evaluated Polyakov loop to the model renders the broken phase more robust, as the critical temperatures are increased. ![\[FIG:nn\]Quark constituent mass $M$ (in units of $M_0$), pion condensate $N$ (in units of $M_0$) and $\Phi$ as a function of temperature (in MeV) for different values of $\mu$. $M_0 = 325$ MeV is the constituent quark mass evaluated at $\mu=0$, $T=0$ and $\mu_e=0$.](nn0.eps "fig:"){width="7cm"}     ![\[FIG:nn\]Quark constituent mass $M$ (in units of $M_0$), pion condensate $N$ (in units of $M_0$) and $\Phi$ as a function of temperature (in MeV) for different values of $\mu$. $M_0 = 325$ MeV is the constituent quark mass evaluated at $\mu=0$, $T=0$ and $\mu_e=0$.](nn342.eps "fig:"){width="7cm"} In Fig. \[FIG:nn\] we plot the quark constituent mass $M$ (in units of $M_0$), pion condensate $N$ (in units of $M_0$) and $\Phi$ as a function of temperature (in MeV) for two values of $\mu$. We have defined $M_0 = 325$ MeV as the constituent quark mass evaluated at $\mu=0$, $T=0$ and $\mu_e=0$. The electron chemical potential evaluated at some values of $\mu$ is shown in Fig. \[FIG:mue\]. We observe that when electrical neutrality is required, pions do not condense in the ground state of the model. As a matter of fact $N=0$ in Fig. \[FIG:nn\]; we have verified this result for several values of $\mu$. Stated in other words the isospin chemical potential $\mu_I = -\mu_e/2$ in the neutral phase is always smaller than the value required for pion condensation to occur. We have verified the robustness of our calculations in several ways. First, we have treated $\mu_e$ as a free parameter, reproducing the results of Ref. [@Zhang:2006gu] finding a pion condensate for high values of $\mu_e$. Second, we have compared our results with those Ref. [@Ebert:2005wr]. The authors of Ref. [@Ebert:2005wr] study the neutral ground state in the NJL model at low temperature. They use a set of parameters in which the values of $G$ and $\Lambda$ are similar to ours, but they consider only the chiral limit $m_u = m_d = 0$. In this limit they find a narrow window in $\mu$ at zero temperature in which the pion condensate can exist. We have explicitly checked that putting by hand $m_u = m_d = 0$ in our calculations we reproduce the aforementioned window. This is a further check since at low temperature the PNJL reduces to the NJL model. Moreover we have verified that the pion condensate is washed out if the current quark mass is increased from zero to the physical value $m_u = m_d = 5.5$ MeV. This is explicitly shown in Fig \[FIG:cmd\] where we plot the pion condensate $N$ and the constituent quark mass $M$ in units of $M_0$, the constituent quark mass at $T=\mu=0$, as a function of the bare quark mass $m_0$ at the representative value of $\mu=330$ MeV (we have checked that the same result hold for other values of $\mu$) and $T=0$. Therefore the current quark mass works as an external field that drives the alignment of the vacuum along the chiral condensate direction. We stress that Fig. 5 has to be taken only as a trick that allows a clear comparison between our results and those of Ref. [@Ebert:2005wr]. A true plot of $M$ and $N$ as a function of the current quark mass requires the self consistent evaluation of $G$ and $\Lambda$, which is beyond the scope of the present paper since we study only the phase diagram and the pion modes at the physical point. Nevertheless the problem of the modification of the ground state by variations of the current quark mass is interesting and deserves further study, therefore we leave this to a future project. ![\[FIG:mue\]Electron chemical potential as a function of the temperature, for different values of the quark chemical potential. ](mue.eps){width="8cm"} ![\[FIG:cmd\]Pion condensate $N$ and the constituent quark mass $M$ in units of $M_0$, the constituent quark mass at $T=\mu=0$, as a function of the bare quark mass $m_0$ at the representative value of $\mu=330$ MeV and $T=0$.](cmd.eps){width="8cm"} Pseudo-Goldstone modes\[Sec:PGm\] ================================= In this Section we compute the masses of the pseudo-Goldstone modes as well as of the $\sigma$ mode in the two flavor and electrically neutral PNJL model. As discussed in the previous Section for each value of $\mu$ and $T$, after requiring electrical neutrality, the difference of chemical potentials between up and down quarks is lower than the critical value required for the realization of the pion condensation. In this Section we therefore put by hand $N=0$ in the quark propagator, Eq. . The pion mass is defined as the solution of the well-known Bethe-Salpeter equation for the bound states. Its derivation in the context of the NJL model is in Ref. [@revNJL]. The same equation is valid in the PNJL model, the difference between PNJL and NJL being only the different quark propagator to be used in the calculation of the pion polarization tensor. The equation reads $$1-2G\Pi_A(m^2_A)=0~,~~~~~A=1,2,3~.\label{eq:bs1}$$ In the above equation $\Pi_A$ is the polarization tensor of the pion $A$, specified later, where $A=1,2,3$ correspond respectively to $\pi^+$, $\pi^-$ and $\pi^0$; $m_A$ denotes the pion mass, and $G$ is the coupling constant introduced in Eq. . The solutions of Eq.  correspond to the poles of the pion propagator, the latter evaluated in the random phase approximation [@revNJL]. We now specify the pion polarization tensor. We introduce the polarization matrix $$\Pi_{ij}(m^2) = -i\int\frac{d^4p}{(2\pi)^4}\text{Tr}\left[i\gamma_5 \tilde\tau_i iS\left(p+\frac{k}{2}\right)i\gamma_5 \tilde\tau_j iS\left(p-\frac{k}{2}\right)\right]~;\label{eq:pp}$$ here $i,j=1,2,3$, $k_\mu = (m,{\bm 0})$, and the trace is over color, flavor and Dirac indices; the matrices $\tilde\tau_i$ are operators in the flavor space, defined by $$\tilde\tau_{1,2}=\frac{1}{\sqrt{2}}\left(\tau_1\pm i \tau_2\right)~,~~~~~ \tilde\tau_3 = \tau_3~,$$ where $\tilde\tau_i$, $i=1,2,3$ are the Pauli matrices; $S$ is the fermion propagator, whose inverse is defined in Eq.  with $N=0$, and masses and chemical potentials evaluated self consistently in the electrically neutral phase. In terms of the matrix $\Pi_{ij}$ the pion polarization tensors appearing in Eq.  are defined as $$\Pi_1 = \Pi_{12}~,~~~~~ \Pi_2= \Pi_{21}~,~~~~~\Pi_3 = \Pi_{33}~.$$ Analogously the equation for the $\sigma$ mass reads $$1-2G\Pi_\sigma(m^2_\sigma)=0~,\label{eq:bss}$$ with the polarization tensor defined as $$\Pi_\sigma(m^2) = -i\int\frac{d^4p}{(2\pi)^4}\text{Tr}\left[ iS\left(p+\frac{k}{2}\right) iS\left(p-\frac{k}{2}\right)\right]~.\label{eq:pps}$$ Once the traces in Eqs.  and  are evaluated, we are left with the loop 4-momentum integral. We replace the integral over energy by a summation over fermion Matsubara frequencies; moreover, the integral over directions of momentum is trivial. Therefore we are left with the integration over the modulus of ${\bm p}$ only, which is evaluated numerically. ![\[FIG:pp\]Masses of the charged pions (left panel) and of the neutral pion (right panel) in the electrically neutral phase, as a function of the temperature, for different values of the quark chemical potential.](PiP.eps "fig:"){width="7cm"}      ![\[FIG:pp\]Masses of the charged pions (left panel) and of the neutral pion (right panel) in the electrically neutral phase, as a function of the temperature, for different values of the quark chemical potential.](Pi0.eps "fig:"){width="7cm"} In Fig. \[FIG:pp\] we plot the masses of the charged pions (left panel) and of the neutral pion (right panel) in the electrically neutral phase, as a function of the temperature, for different values of the quark chemical potential. The behavior of the pseudoscalar modes as the temperature is increased is qualitatively the same observed in the NJL model [@revNJL]. At each value of $\mu$, the charged and neutral pion masses are of the order of the zero temperature value $\approx 140$ MeV below the chiral crossover (or the chiral first order transition); as the chiral transition (either first order or crossover) occurs, the pion masses rapidly increase signaling the disappearing of the pseudoscalar modes from the low energy spectrum of the model. ![\[FIG:sm\]Mass of the $\sigma$ mode in the electrically neutral phase, as a function of the temperature, for different values of the quark chemical potential.](SigmaMass.eps){width="7cm"} In Fig. \[FIG:sm\] we plot the mass of the $\sigma$ mode in the electrically neutral phase, as a function of the temperature, for different values of the quark chemical potential. As in the case of the pions, the behavior of $m_\sigma$ is analogous to that observed in the NJL model [@revNJL]. ![\[FIG:kkk\]The ratio $m_\pi/2M$, with $M$ the constituent quark mass, as a function of temperature, for two representative values of the quark chemical potential. The vertical bold dashed line denotes the chiral transition temperature.](tt1.eps "fig:"){width="7cm"}     ![\[FIG:kkk\]The ratio $m_\pi/2M$, with $M$ the constituent quark mass, as a function of temperature, for two representative values of the quark chemical potential. The vertical bold dashed line denotes the chiral transition temperature.](tt4.eps "fig:"){width="7cm"} It is interesting to evaluate the ratio $m_\pi/2M$, with $M$ the constituent quark mass, as a function of temperature. If it is larger than one, than a bound state with the quantum numbers of the pion is less stable than a state made of a free quark and a free antiquark, and thus the pion melts to its constituent quarks. In Fig. \[FIG:kkk\] we show the results of the computation of $m_\pi/2M$ at different values of $\mu$ in the electrically neutral phase, for values of temperatures close either to the chiral crossover or to the first order chiral transition. We notice that for any value of the quark chemical potential there exists a window in temperature, above the chiral restoration temperature, in which the pions are lighter than a quark and an antiquark. The window shrinks as $\mu$ is increased. These results show that even above the critical temperature a bound state with the pions quantum numbers can be formed. We observe that the window in which the bound state is stable above the chiral phase transition (as well as the chiral crossover) is not a peculiarity of the PNJL model. As a matter of fact the same window was noticed in the NJL model at $\mu=0$ in Ref. [@Hatsuda:1985eb]. We have verified that it exists in the neutral NJL model as well. For example, at $\mu=0$ we find that the bound states exist in the window $(195,212)$ MeV; this result has to be compared with the PNJL one, namely $(205,245)$ MeV, showing that the PNJL window is almost 2.4 times larger than the NJL one. Analogously at $\mu=342$ MeV the NJL window is $(34,44)$ MeV, to be compared with the PNJL result $(73,85)$ MeV. Similar results have been obtained in the framework of the PNJL model in Ref. [@Hansen:2006ee]. The authors of Ref. [@Hansen:2006ee] consider the two flavor PNJL model without the constraint of electrical neutrality. Moreover they use a polynomial form of the Polyakov loop effective potential ${\cal U}$ in Eq.  instead of the logarithmic one, see Eq. . Therefore a quantitative comparison with their results is not feasible. Instead we have computed the ratio $T_M/T_\chi$ in the case $\mu_e=0$ in the PNJL model with the logarithmic form of the effective potential, for several values of $\mu$, in order to quantify the role of the electrical neutrality on the ratio $T_M/T_\chi$. Our results are summarized in Table \[tab:comp\]. $T_\chi~ (\mu_e = 0)$ $T_M~(\mu_e = 0)$ $T_M/T_\chi~(\mu_e=0)$ $T_\chi$ $T_M$ $T_M/T_\chi$ ----------- ----------------------- ------------------- ------------------------ ---------- ------- -------------- $\mu=100$ $218$ $236$ 1.08 $198$ $234$ 1.18 $\mu=200$ $178$ $207$ 1.16 $175$ $211$ 1.20 $\mu=300$ $131$ $145$ 1.10 $131$ $154$ 1.17 $\mu=342$ $44$ $48$ 1.09 $73$ $85$ 1.16 : \[tab:comp\]Comparison of the pion melting temperatures $T_M$ and the chiral transition temperatures $T_\chi$ with electrical neutrality and with $\mu_e = 0$. Quark chemical potentials and temperatures are measured in MeV. It is interesting to superimpose the bound state existence region to the phase diagram in Fig. \[FIG:pd\]. This is done in Fig. \[FIG:ccc\] where we show the region in the $\mu-T$ plane where the bound state can be formed. In the white region below the gray domain the chiral symmetry is broken and pions live as pseudo-Goldstone modes. In the gray region chiral symmetry is restored but $m_\pi/2M_q < 1$, thus bound states can be formed. Finally, in the white region above the gray domain chiral symmetry is restored and free quark states are more stable than bound states (this region is characterized by $m_\pi/2M_q > 1$). The existence of bound states above the chiral transition temperature is interesting because it implies that quark matter above the chiral transition is strongly coupled. It has been suggested [@Shuryak:2003ty; @Brown:2003km] that the formation of bound states just above the chiral transition can explain the experimental results obtained by non-central heavy ion collisions at RHIC facility [@Kolb:2003dz; @Molnar:2001ux; @Teaney:2003kp; @Nakamura:2004sy]; in particular it provides a simple explanation to the low viscosity-to-entropy ratio observed in elliptic flow simulations. The next step would be the computation of the transport coefficients of quark matter just above the chiral transition temperature. We leave this to a future project. ![\[FIG:ccc\]Region of existence of bound states in the $\mu-T$ plane. $\chi SB$ denotes the region where chiral symmetry is broken; $\chi SR$ denotes the region where chiral symmetry is restored. $M$ is the constituent quark mass. The dashed and the solid lines correspond respectively to the chiral crossover and to the chiral first order transition, the dot denotes the CEP. In the gray region chiral symmetry is restored but a bound state with the quantum numbers of the pions can still be formed. See the text for more details.](uu.eps){width="10cm"} Conclusions\[Sec:con\] ====================== In this paper we have studied the two flavor PNJL model in presence of a charge chemical potential, requiring electrical neutrality in the mean field approximation. We have assumed that in the ground state both the chiral and the pion condensates develop, and we have evaluated their values, as well as the value of the Polyakov loop, for each value of the temperature and of the quark chemical potential by minimizing the thermodynamic potential under the condition that the total electric charge of the system vanishes. We have drawn the phase diagram of the model, see Fig. \[FIG:pd\]. One of our main results is that we do not observe pion condensation in the mean field approximation. We relate this to the non vanishing current quark mass, as shown in Fig. 5. This part of our study is completed by the results for the constituent quark masses, the Polyakov loop and the charge chemical potential, see Figs. \[FIG:nn\] and \[FIG:mue\]. We have computed the masses of the pseudo-Goldstone modes and of the $\sigma$-mode. The behavior of the masses as a function of the temperature for different values of the quark chemical potential are shown in Figs. \[FIG:pp\] and \[FIG:sm\]. Furthermore we have investigated on the possibility of existence of bound states with the quantum numbers of the pions above the chiral critical temperature. To this end we have compared the computed pion mass with twice the constituent quark mass. The result is shown in Fig. \[FIG:kkk\]. We have found that for any value of the quark chemical potential there exists a window in temperature, above the chiral restoration temperature, in which the pions are lighter than a state formed by a free quark and a free antiquark. The aforementioned window shrinks as $\mu$ is increased. Our results show that even above the critical temperature a bound state with the pions quantum numbers can be formed. This is summarized in Fig. \[FIG:ccc\]. In this work we have not studied independently the role of the chemical potential on the suppression of the pion condensate. This can be done by leaving the isospin chemical potential as a fixed and free parameter, and varying only the mean quark chemical potential. A further natural extension of our work is the study of the $2+1$ flavor PNJL model with meson condensation, with and without neutrality conditions implemented. Moreover, it would be interesting to include a color superconductive state in the neutral model [@Rapp:1997zu] (studies of the superconductive state without neutrality conditions can be found in Refs. [@Roessner:2006xn; @Ciminale:2007ei]). Even more one should compute the transport coefficients of quark matter just above the chiral transition temperature. We leave these projects to future works. [99]{} Y. Nambu and G. Jona-Lasinio, Phys. Rev.  [**122**]{}, 345 (1961); Phys. Rev.  [**124**]{}, 246 (1961). S. P. Klevansky, Rev. Mod. Phys.  [**64**]{}, 649 (1992); T. Hatsuda and T. Kunihiro, Phys. Rept.  [**247**]{}, 221 (1994) \[arXiv:hep-ph/9401310\]; M. Buballa, Phys. Rept.  [**407**]{}, 205 (2005) \[arXiv:hep-ph/0402234\]. A. M. Polyakov, Phys. Lett.  B [**72**]{}, 477 (1978); L. Susskind, Phys. Rev.  D [**20**]{}, 2610 (1979); B. Svetitsky and L. G. Yaffe, Nucl. Phys.  B [**210**]{}, 423 (1982); B. Svetitsky, Phys. Rept.  [**132**]{}, 1 (1986). P. N. Meisinger and M. C. Ogilvie, Phys. Lett.  B [**379**]{}, 163 (1996) \[arXiv:hep-lat/9512011\]. K. Fukushima, Phys. Lett.  B [**591**]{}, 277 (2004) \[arXiv:hep-ph/0310121\]. C. Ratti, M. A. Thaler and W. Weise, Phys. Rev.  D [**73**]{}, 014019 (2006) \[arXiv:hep-ph/0506234\]. S. Roessner, C. Ratti and W. Weise, Phys. Rev.  D [**75**]{}, 034007 (2007) \[arXiv:hep-ph/0609281\]. S. K. Ghosh, T. K. Mukherjee, M. G. Mustafa and R. Ray, arXiv:0710.2790 \[hep-ph\]; S. K. Ghosh, T. K. Mukherjee, M. G. Mustafa and R. Ray, Phys. Rev.  D [**73**]{}, 114007 (2006) \[arXiv:hep-ph/0603050\]. K. Kashiwa, H. Kouno, M. Matsuzaki and M. Yahiro, arXiv:0710.2180 \[hep-ph\]. B. J. Schaefer, J. M. Pawlowski and J. Wambach, Phys. Rev.  D [**76**]{}, 074023 (2007) \[arXiv:0704.3234 \[hep-ph\]\]. C. Ratti, S. Roessner and W. Weise, Phys. Lett.  B [**649**]{}, 57 (2007) \[arXiv:hep-ph/0701091\]. C. Sasaki, B. Friman and K. Redlich, Phys. Rev.  D [**75**]{}, 074013 (2007) \[arXiv:hep-ph/0611147\]; C. Sasaki, B. Friman and K. Redlich, Phys. Rev.  D [**75**]{}, 054026 (2007) \[arXiv:hep-ph/0611143\]. E. Megias, E. Ruiz Arriola and L. L. Salcedo, Phys. Rev.  D [**74**]{}, 114014 (2006) \[arXiv:hep-ph/0607338\]; E. Megias, E. Ruiz Arriola and L. L. Salcedo, Phys. Rev.  D [**74**]{}, 065005 (2006) \[arXiv:hep-ph/0412308\]. Z. Zhang and Y. X. Liu, Phys. Rev.  C [**75**]{}, 064910 (2007) \[arXiv:hep-ph/0610221\]. M. Ciminale, G. Nardulli, M. Ruggieri and R. Gatto, Phys. Lett.  B [**657**]{}, 64 (2007) \[arXiv:0706.4215 \[hep-ph\]\]. W. j. Fu, Z. Zhang and Y. x. Liu, Phys. Rev.  D [**77**]{}, 014006 (2008) \[arXiv:0711.0154 \[hep-ph\]\]. M. Ciminale, R. Gatto, N. D. Ippolito, G. Nardulli and M. Ruggieri, arXiv:0711.3397 \[hep-ph\]. D. Ebert and K. G. Klimenko, Eur. Phys. J.  C [**46**]{}, 771 (2006) \[arXiv:hep-ph/0510222\]. D. Ebert, K. G. Klimenko and H. Toki, Phys. Rev.  D [**64**]{}, 014038 (2001) \[arXiv:hep-ph/0011273\]; D. Ebert, V. V. Khudyakov, V. C. Zhukovsky and K. G. Klimenko, Phys. Rev.  D [**65**]{}, 054024 (2002) \[arXiv:hep-ph/0106110\]. T. Hatsuda and T. Kunihiro, Phys. Rev. Lett.  [**55**]{} (1985) 158. H. Hansen, W. M. Alberico, A. Beraudo, A. Molinari, M. Nardi and C. Ratti, Phys. Rev.  D [**75**]{}, 065004 (2007) \[arXiv:hep-ph/0609116\]. E. V. Shuryak and I. Zahed, Phys. Rev.  C [**70**]{}, 021901 (2004) \[arXiv:hep-ph/0307267\]. G. E. Brown, C. H. Lee, M. Rho and E. Shuryak, Nucl. Phys.  A [**740**]{}, 171 (2004) \[arXiv:hep-ph/0312175\]. P. F. Kolb and U. W. Heinz, arXiv:nucl-th/0305084. D. Molnar and M. Gyulassy, Nucl. Phys.  A [**697**]{}, 495 (2002) \[Erratum-ibid.  A [**703**]{}, 893 (2002)\] \[arXiv:nucl-th/0104073\]. D. Teaney, Phys. Rev.  C [**68**]{}, 034913 (2003) \[arXiv:nucl-th/0301099\]. A. Nakamura and S. Sakai, Phys. Rev. Lett.  [**94**]{}, 072305 (2005) \[arXiv:hep-lat/0406009\]; A. Peshier and W. Cassing, Phys. Rev. Lett.  [**94**]{}, 172301 (2005) \[arXiv:hep-ph/0502138\]. R. Rapp, T. Schafer, E. V. Shuryak and M. Velkovsky, Phys. Rev. Lett.  [**81**]{}, 53 (1998) \[arXiv:hep-ph/9711396\]; M. G. Alford, K. Rajagopal and F. Wilczek, Phys. Lett.  B [**422**]{}, 247 (1998) \[arXiv:hep-ph/9711395\]; M. G. Alford, K. Rajagopal and F. Wilczek, Nucl. Phys.  B [**537**]{}, 443 (1999) \[arXiv:hep-ph/9804403\].

ArXiv

--- abstract: 'The LULU operators, well known in the nonlinear multiresolution analysis of sequences, are extended to functions defined on continuous domain, namely, a real interval $\Omega\subseteq\mathbb{R}$. Similar to their discrete counterparts, for a given $\delta>0$ the operators $L_\delta$ and $U_\delta$ form a fully ordered semi-group of four elements. It is shown that the compositions $L_\delta\circ U_\delta$ and $U_\delta\circ L_\delta$ provide locally $\delta$-monotone approximations for the bounded real functions defined on $\Omega$. The error of approximation is estimated in terms of the modulus of nonmonotonicity.' author: - | Roumen Anguelov\ Department of Mathematics and Applied Mathematics\ University of Pretoria\ Pretoria 0002, SOUTH AFRICA\ E-mail: [email protected] title: 'LULU operators and locally $\delta$-monotone approximations' --- Introduction ============ The LULU operators remove impulsive noise before a signal extraction from a sequence. They are computationally convenient and conceptually simpler compared to the the median smoothers usually considered to be the “basic” smoothers. The LULU operators have particular properties, e.g. they are fully trend preserving, [@RohwerQM2004], preserve the total variation, [@RohwerQM2002], etc., which make them an essential tool for multiresolution analysis of sequences. Furthermore, it was demonstrated during the last decade or so that these operators, being specific cases of morphological filters, [@Serra], have a critical role in the analysis and comparison of nonlinear smoothers, [@Rohwerbook]. We extend the LULU theory from sequences to functions on a continuous domain, namely, a real interval $\Omega$. The existing LULU theory can be considered as a particular case in this new development, since the discrete LULU operators can be equivalently formulated for splines of order 1 or order 2 on the integer partition of real line, the one-to-one mapping being given by the B-spline basis. Given a sequence $\xi=(\xi_i)_{i\in\mathbb{N}}$ and $n\in\mathbb{N}$ the operators $L_n$ and $U_n$ are defined as follows $$\begin{aligned} (L_n \xi)_i=\max\{\min\{\xi_{i-n},...,\xi_i\},...,\min\{\xi_{i},...,\xi_{i+n}\}\},\ i\in\mathbb{N}\\ (U_n \xi)_i=\min\{\max\{\xi_{i-n},...,\xi_i\},...,\max\{\xi_{i},...,\xi_{i+n}\}\},\ i\in\mathbb{N}\end{aligned}$$ In analogy with the above discrete LULU operators, for a given $\delta>0$ the basic smoothers $L_\delta$ and $U_\delta$ in the LULU theory are defined for functions on $\Omega$ through the concepts of the so called lower and upper $\delta$-envelopes of these functions. These definitions are given in Section 2, where it is also shown that the operators $L_\delta$ and $U_\delta$ preserve essential properties of their discrete counterparts. In particular, the operators $L_\delta$ and $U_\delta$ generate through composition a fully ordered four element semi-group also called a strong LULU structure. This issue is dealt with in Section 3. In section 4 we define the concept of local $\delta$-monotonicity and show that the compositions $L_\delta\circ U_\delta$ and $U_\delta\circ L_\delta$ are smoothers in the sense that the resulting functions are locally $\delta$-monotone. The errors of approximation of real functions $f$ by the these compositions are estimated in terms of the modulus of nonmonotonicity $\mu(f,\delta)$. The basic smoothers $L_\delta$ and $U_\delta$ ============================================= Let $ \mathcal{A}(\Omega )$ denote the set of all bounded real functions defined on the real interval $\Omega\subseteq\mathbb{R}$. Let $B_{\delta}(x)$ denote the closed $\delta$-neighborhood of $x$ in $\Omega $, that is, $B_{\delta }(x)=\{y\in \Omega :|x-y|\leq\delta \}$. The pair of mappings $I$, $S:\mathcal{A}(\Omega )\rightarrow \mathcal{A}(\Omega )$ defined by $$\begin{aligned} I(f)(x) &=&\sup_{\delta >0}\inf \{f(y):y\in B_{\delta }(x)\}, \ x\in\Omega, \label{lbf} \\ S(f)(x) &=&\inf_{\delta >0}\sup \{f(y):y\in B_{\delta }(x)\}, \ x\in\Omega. \label{ubf}\end{aligned}$$are called lower Baire, and upper Baire operators, respectively, [Sendov]{}. We consider on $\mathcal{A}(\Omega)$ the point-wise defined partial order, that is, for any $f,g\in\mathcal{A}(\Omega)$ $$\label{forder} f\leq g\Longleftrightarrow f(x)\leq g(x),\ x\in\Omega.$$ Then the lower and upper Baire operators can be defined in the following equivalent way. For every $f\in \mathcal{A}(\Omega )$ the function $I(f)$ is the maximal lower semi-continuous function which is not greater than $f$. Hence, it is also called lower semi-continuous envelope. In a similar way, $S(f)$ is the smallest upper semi-continuous function which is not less than $f$ and is called the upper semi-continuous envelope of $f$. In analogy with $I(f)$ and $S(f)$ we call the functions $$\begin{aligned} I_{\delta }(f)(x) &=&\inf \{f(y):y\in B_{\delta }(x)\},\ x\in\Omega, \label{ldenvelop} \\ S_{\delta }(f)(x) &=&\sup \{f(y):y\in B_{\delta }(x)\},\ x\in\Omega. \label{udenvelop}\end{aligned}$$a lower $\delta$-envelope of $f$ and an upper $\delta $-envelope of $f$, respectively. It is easy to see from (\[ldenvelop\]) and (\[udenvelop\]) that for every $\delta_1,\delta_2>0$ $$I_{\delta _{1}}\circ I_{\delta _{2}}=I_{\delta _{1}+\delta _{2}}\ \ ,\ \ \ S_{\delta _{1}}\circ S_{\delta _{2}}=S_{\delta _{1}+\delta _{2}}\label{IScomp}$$ Furthermore, the operators $I_\delta$ and $S_\delta$, $\delta>0$, as well as $I$ and $S$ are all monotone increasing with respect to the order (\[forder\]), that is, for every $f,g\in\mathcal{A}(\Omega)$ $$\label{IdSdmon}f\leq g\Longrightarrow I_\delta(f)\leq I_\delta(g),\ S_\delta(f)\leq S_\delta(g),\ I(f)\leq I(g),\ S(f)\leq S(g).$$ The following operators can be considered as continuous analogues of the discrete $LULU$ operators given in the Introduction: $$L_{\delta }=S_{\frac{\delta}{2} }\circ I_{\frac{\delta}{2} }\ ,\ \ U_{\delta }=I_{\frac{\delta}{2} }\circ S_{\frac{\delta}{2}}\ .$$ We will show that these operators have similar properties to their discrete counterparts. Let us note that they inherit monotonicity with respect of the functional argument from the operators $I_\delta$ and $S_\delta$, (\[IdSdmon\]), that is, for $f,g\in\mathcal{A}(\Omega)$ $$\label{LdUdfmon}f\leq g\Longrightarrow L_\delta(f)\leq L_\delta(g),\ U_\delta(f)\leq U_\delta(g).$$ \[tLdfleqf\] For every $f\!\in\!\mathcal{A}(\Omega)$ and $\delta\!>\!0$ we have $L_\delta(f)\leq f$, $U_\delta(f)\geq f$. Let $f\in\mathcal{A}(\Omega)$, $\delta>0$. For any $x\in\Omega$ it follows from the definition of $I_\delta$ that $I_\frac{\delta}{2}(f)(y)\leq f(x),\ y\in B_\frac{\delta}{2}(x)$. Therefore $L_\delta(f)(x)=S_\frac{\delta}{2}(I_\frac{\delta}{2}(f))(x)=\sup\{I_\frac{\delta}{2}(f)(y):y\in B_\frac{\delta}{2}(x)\}\leq f(x)$, $x\in\Omega$. The second inequality in the theorem is proved in a similar way. \[tLdUdmon\] The operator $L_\delta$ is monotone increasing on $\delta$ while the operator $U_\delta$ is monotone decreasing on $\delta$, that is, for any $f\in\mathcal{A}(\Omega)$ and $0<\delta_1\leq\delta_2$ we have $L_{\delta_1}(f)\leq L_{\delta_2}(f)$, $U_{\delta_1}(f)\geq U_{\delta_2}(f)$. Let $\delta_2>\delta_1>0$. Using the properties (\[IScomp\]) the operator $L_{\delta_2}$ can be represented in the form $ L_{\delta_2}=S_\frac{\delta_2}{2}\circ I_\frac{\delta_2}{2}=S_\frac{\delta_1}{2}\circ S_{\frac{\delta_2-\delta_1}{2}}\circ I_{\frac{\delta_2-\delta_1}{2}} \circ I_\frac{\delta_1}{2}=S_\frac{\delta_1}{2}\circ L_{\delta_2-\delta_1} \circ I_\frac{\delta_1}{2}$. It follows from Theorem \[tLdfleqf\] that for every $f\in\mathcal{A}(\Omega)$ we have $L_{\delta_2-\delta_1}(I_\frac{\delta_1}{2}(f))\leq I_\frac{\delta_1}{2}(f)$. Hence using the monotonicity of the operator $S_\delta$ given in (\[IdSdmon\]) we obtain $L_{\delta_2}(f)=S_\frac{\delta_1}{2}(L_{\delta_2-\delta_1}(I_\frac{\delta_1}{2}(f)))\leq S_\frac{\delta_1}{2}(I_\frac{\delta_1}{2}(f))=L_{\delta_1}(f)$, $f\in\mathcal{A}(\Omega)$. The inequality $U_{\delta_1}(f)\geq U_{\delta_2}(f)$ is proved in a similar way. The next lemma is useful in dealing with compositions of $I_\delta$ and $S_\delta$. \[tIdSdId\] We have $I_\delta\circ S_\delta\circ I_\delta=I_\delta$, $S_\delta\circ I_\delta\circ S_\delta=S_\delta$. Using the monotonicity of $I_\delta$, see (\[IdSdmon\]), and Theorem \[tLdfleqf\] for $f\in \mathcal{A}(\Omega)$ we have $(I_\delta\circ S_\delta\circ I_\delta)(f)=I_\delta (L_{2\delta}(f))\leq I_\delta (f)$. On the other side, applying Theorem \[tLdfleqf\] to $U_{2\delta}$ we obtain $(I_\delta\circ S_\delta\circ I_\delta)(f)=U_{2\delta}(I_\delta(f))\geq I_\delta(f)$. Therefore $(I_\delta\circ S_\delta\circ I_\delta)(f)=I_\delta(f)$, $f\in\mathcal{A}(\Omega)$. The second equality is proved similarly. \[tLdUdabsorb\] For every $\delta_1,\delta_2>0$ we have $L_{\delta_1 }\circ L_{\delta_2 } =L_{\max\{\delta_1,\delta_2\}}$ and $U_{\delta_1 }\circ U_{\delta_2 }=U_{\max\{\delta_1,\delta_2\}}$. We will only prove the first equality since the proof of the second one is done in a similar manner. Let first $\delta_2>\delta_1>0$. Using the property (\[IScomp\]) and Lemma \[tIdSdId\] we obtain $$\begin{aligned} L_{\delta_1}\circ L_{\delta_2}&=&(S_\frac{\delta_1}{2}\circ I_\frac{\delta_1}{2})\circ(S_\frac{\delta_2}{2}\circ I_\frac{\delta_2}{2})\ =\ (S_\frac{\delta_1}{2}\circ I_\frac{\delta_1}{2}\circ S_\frac{\delta_1}{2})\circ (S_{\frac{\delta_2-\delta_1}{2}}\circ I_\frac{\delta_2}{2})\\ &=& S_\frac{\delta_1}{2}\circ S_{\frac{\delta_2-\delta_1}{2}}\circ I_\frac{\delta_2}{2}\ =\ S_\frac{\delta_2}{2}\circ I_\frac{\delta_2}{2}\ =\ L_{\delta_2}.\end{aligned}$$ If $\delta_1>\delta_2>0$ in a similar way we have $$\begin{aligned} L_{\delta_1}\circ L_{\delta_2}&=&(S_{\delta_1}\circ I_{\delta_1})\circ(S_{\delta_2}\circ I_{\delta_2})\ =\ (S_{\delta_1}\circ I_{\delta_1-\delta_2})\circ (I_{\delta_2}\circ S_{\delta_2}\circ I_{\delta_2})\\ &=&S_{\delta_1}\circ I_{\delta_1-\delta_2}\circ I_{\delta_2}\ =\ S_{\delta_1}\circ I_{\delta_1}\ =\ L_{\delta_1}.\end{aligned}$$ The proof in the case when $\delta_2\!=\!\delta_1\!>\!0$ follows from either of the above identities where $S_{\delta_2-\delta_1}$ or $I_{\delta_1-\delta_2}$ respectively are replaced by the identity operator. Important properties of smoothing operators are their idempotence and co-idempotence. Hence the significance of the next theorem. \[tLdUdidemp\] The operators $L_\delta$ and $U_\delta$ are both idempotent and co-idempotent, that is, $L_{\delta }\circ L_{\delta } =L_{\delta}$, $U_{\delta }\circ U_{\delta } =U_{\delta}$, $(id-L_{\delta })\circ (id-L_{\delta })=id-L_{\delta }$, $(id-U_{\delta })\circ (id-U_{\delta })=id-U_{\delta }$, where $id$ denotes the identity operator. The idempotence of $L_\delta$ and $U_\delta$ follows directly from Theorem \[tLdUdabsorb\]. The co-idempotence of the operator $L_\delta$ is equivalent to $L_\delta\circ(id-L_\delta)=0$. Using the first inequality in Theorem \[tLdfleqf\] one can easily obtain $L_\delta\circ(id-L_\delta)\geq 0$. Hence, for the co-idempotence of $L_\delta$ it remains to show that $L_\delta\circ(id-L_\delta)\leq 0$. Assume the opposite. Namely, there exists a function $f\in\mathcal{A}(\Omega)$ and $x\in\Omega$ such that $(L_\delta\circ(id-L_\delta))(f)(x)>0$. Let $\varepsilon>0$ be such that $(L_\delta\circ(id-L_\delta))(f)(x)>\varepsilon>0$. Using the definition of $L_\delta$ the above inequality implies that there exists $y\in B_\frac{\delta}{2}(x)$ such that for every $z\in B_\frac{\delta}{2}(y)$ we have $(id-L_\delta)(f)(z)>\varepsilon$, or equivalently $$\label{ineq3tidemp} f(z)>L_\delta(f)(z)+\varepsilon,\ z\in B_\frac{\delta}{2}(y).$$ For every $z\in B_\frac{\delta}{2}(y)$ we also have $L_\delta(f)(z)\geq I_\frac{\delta}{2}(f)(y)=\inf\{f(t):t\in B_\frac{\delta}{2}(y)\}$. Hence there exists $t\in B_\frac{\delta}{2}(y)$ such that $f(t)<I_\frac{\delta}{2}(f)(y)+\varepsilon\leq L_\delta(f)(z)+\varepsilon$, $z\in B_\frac{\delta}{2}(y)$. Taking $z=t$ in the above inequality we obtain $f(t)<L_\delta(f)(t)+\varepsilon$, which contradicts (\[ineq3tidemp\]). The co-idempotence of $U_\delta$ is proved in a similar way. The figures below illustrate graphically the smoothing effect of the operators $L_\delta$, $U_\delta$ and their compositions. The graph of function $f$ is given by dotted lines.  \ The functions $L_\delta(f)$ and $U_\delta(f)$\  \ The functions $(L_\delta\circ U_\delta)(f)$ and $(U_\delta\circ L_\delta)(f)$ The operator $L_\delta$ smoothes the function $f$ from above by removing sharp picks while the operator $U_\delta$ smoothes the function $f$ from below by removing deep depressions. The smoothing effect of the compositions $L_\delta\circ U_\delta$ and $U_\delta\circ L_\delta$ can be described in terms of the local $\delta$-monotonicity discussed in the Section 4. Note that $L_\delta\circ U_\delta$ and $U_\delta\circ L_\delta$ resolve ambiguities in a different way; $L_\delta\circ U_\delta$ treats oscillations of length less then $\delta$ as upward impulses and removes them while $U_\delta\circ L_\delta$ considers such oscillations as downward impulses which are accordingly removed. The inequality $(U_\delta\circ L_\delta)(f)\leq(L_\delta\circ U_\delta)(f)$ which is observed here will be proved in the next section for $f\in\mathcal{A}(\Omega)$. The LULU semi-group =================== In this section we consider the set of the operators $L_\delta$ and $U_\delta$ and their compositions. For operators on $\mathcal{A}(\Omega)$ we consider the point-wise defined partial order. Namely, for operators $P$, $Q$ on $\mathcal{A}(\Omega)$ we have $$P\leq Q\ \Longleftrightarrow \ P(f)\leq Q(f),\ f\in\mathcal{A}(\Omega).$$ Then the inequalities in Theorem \[tLdfleqf\] can be represented in the form $$\label{LleqidleqU} L_\delta\leq id\leq U_\delta,$$ where $id$ denotes the identity operator on $\mathcal{A}(\Omega)$. \[tULleqLU\] For any $\delta>0$ we have $U_\delta\circ L_\delta\leq L_\delta\circ U_\delta$. Let $f\in\mathcal{A}(\Omega)$ and let $x\in\Omega$. Denote $p=(L_\delta\circ U_\delta)(f)(x)=S_\frac{\delta}{2}(I_\delta(S_\frac{\delta}{2}(f)))(x)$. Let $\varepsilon$ be an arbitrary positive. For every $y\in B_\frac{\delta}{2}(x)$ we have $$\label{ineqtULleqLU} I_\delta(S_\frac{\delta}{2}(f))(y)\leq p<p+\varepsilon.$$ There exists $z\in B_\frac{\delta}{2}(x)$ such that $S_\frac{\delta}{2}(f)(z)<p+\varepsilon$. Then $f(t)<p+\varepsilon \ \mbox{ for }\ t\in B_\frac{\delta}{2}(z)$, which implies that $I_\frac{\delta}{2}(f)(t)< p+\varepsilon \ \mbox{ for }\ t\in B_\delta(z)$. Hence $S_\delta(I_\frac{\delta}{2}(f))(z)\leq p+\varepsilon$. Then $(U_\delta\circ L_\delta)(f)(t)=I_\frac{\delta}{2}(S_\delta(I_\frac{\delta}{2}(f)))(t)\leq p+\varepsilon$ for $t\in B_\frac{\delta}{2}(z)$. Since $x\in B_\frac{\delta}{2}(z)$, see the case assumption, from the above inequality we have $(U_\delta\circ L_\delta)(f)(x)\leq p+\varepsilon$.\ For every $z\in B_\frac{\delta}{2}(x)$ we have $S_\frac{\delta}{2}(f)(z)\geq p+\varepsilon$. Denote $D=\left\{z\in\Omega:S_\frac{\delta}{2}(f)(z)< p+\varepsilon \right\}$. We will show that for every $z\in B_\delta(x)$ we have $$\label{intersecttULleqLU} B_\delta(z)\cap D\neq\emptyset$$ Due to the inequality (\[ineqtULleqLU\]) we have that (\[intersecttULleqLU\]) holds for every $z\in B_\frac{\delta}{2}(x)$. Let $z\in B_\delta(x)$ and let $z>x+\frac{\delta}{2}$. This implies that $x+\frac{\delta}{2}\in\Omega$. Using the inequality (\[ineqtULleqLU\]) for $y=x+\frac{\delta}{2}$ as well as the case assumption we obtain that the set $\left(x+\frac{\delta}{2},x+\frac{3\delta}{2}\right]\cap D$ is not empty. Then $B_\delta(z)\cap D\supset \left(x+\frac{\delta}{2},x+\frac{3\delta}{2}\right]\cap D\neq \emptyset$. For $z<x-\frac{\delta}{2}$ condition (\[intersecttULleqLU\]) is proved in a similar way. Hence (\[intersecttULleqLU\]) holds for all $z\in B_\delta(x)$. Let $z\in B_\delta(x)$ and $v\in B_\delta(y)\cap D$. Since $v\in D$ we have $f(t)< p+\varepsilon$, for $t\in B_\frac{\delta}{2}(v)$. Using that $B_\frac{\delta}{2}(z)\cap B_\frac{\delta}{2}(v)\neq \emptyset$ we obtain that $I_\frac{\delta}{2}(f)(z)< p+\varepsilon$, $z\in B_\delta(x)$. Therefore $S_\delta(I_\frac{\delta}{2}(f))(x)\leq p+\varepsilon$. Then $$(U_\delta\circ L_\delta)(f)(x)=I_\frac{\delta}{2}(S_\delta(I_\frac{\delta}{2}(f)))(x)\leq S_\delta(I_\frac{\delta}{2}(f))(x)\leq p+\varepsilon.$$ Combining the results of Case 1 and Case 2 we have $(U_\delta\circ L_\delta)(f)(x)\leq p+\varepsilon$. Since $\varepsilon$ is arbitrary this implies that $(U_\delta\circ L_\delta)(f)(x)\leq p=(L_\delta\circ U_\delta)(f)(x)$. \[tLUULidemp\] For a given $\delta>0$ the operators $L_\delta\circ U_\delta$ and $U_\delta\circ L_\delta$ are both idempotent. The proof is an immediate application of Lemma 1. \[tULU=LU\] We have $U_\delta\circ L_\delta\circ U_\delta=L_\delta\circ U_\delta$, $L_\delta\circ U_\delta\circ L_\delta=U_\delta\circ L_\delta$, $\delta>0$. Using the inequalities (\[LleqidleqU\]) and the monotonicity of the operators $L_\delta$, $U_\delta$, see (\[LdUdfmon\]), we obtain $U_\delta\circ L_\delta\circ U_\delta \geq id\circ L_\delta\circ U_\delta\ =\ L_\delta\circ U_\delta$. For the proof of the inverse inequality we use Theorem \[tULleqLU\] and the idempotence of $U_\delta$ as follows: $$U_\delta\circ L_\delta\circ U_\delta=(U_\delta\circ L_\delta)\circ U_\delta\leq (L_\delta\circ U_\delta)\circ U_\delta=L_\delta\circ (U_\delta\circ U_\delta)=L_\delta\circ U_\delta$$ Therefore $U_\delta\circ L_\delta\circ U_\delta=L_\delta\circ U_\delta$. The second equality is proved in a similar way. It follows from Theorems \[tLUULidemp\] and \[tULU=LU\] that for a fixed $\delta>0$ every composition involving finite number of the operators $L_\delta$ and $U_\delta$ is an element of the set $\{L_\delta,U_\delta,U_\delta\circ L_\delta, L_\delta\circ U_\delta\}$. Hence the operators $L_\delta$ and $U_\delta$ form a semi-group with a composition table as follows: ------------------------------------------------------------------------------------------------------------------------------------------ $ L_\delta$ $ U_\delta$ $U_\delta\circ L_\delta$ $L_\delta\circ U_\delta$ --------------------------- -------------------------- --------------------------- --------------------------- --------------------------- $L_\delta$ $ L_\delta$ $ $U_\delta\circ L_\delta$ $ L_\delta\circ U_\delta$ L_\delta\circ U_\delta$ $ U_\delta$ $ U_\delta\circ $ U_\delta$ $ U_\delta\circ L_\delta$ $L_\delta\circ L_\delta$ U_\delta$ $ U_\delta\circ L_\delta$ $ U_\delta\circ $L_\delta\circ U_\delta$ $U_\delta\circ $L_\delta\circ U_\delta$ L_\delta$ L_\delta$ $ L_\delta\circ $U_\delta\circ L_\delta$ $ L_\delta\circ U_\delta$ $U_\delta\circ L_\delta$ $ L_\delta\circ U_\delta$ U_\delta$ ------------------------------------------------------------------------------------------------------------------------------------------ Furthermore, an easy application of Theorem \[tULleqLU\] shows that this semi-group is completely ordered. Namely, we have $L_\delta\leq U_\delta\circ L_\delta\leq L_\delta\circ U_\delta\leq U_\delta$. Locally $\delta$-monotone approximations ======================================== Let $\delta>0$ and a function $f\in\mathbb{A}(\Omega)$ be given. \(i) The function $f$ is called upwards $\delta$-monotone if for every interval $[x,y]\subset\Omega$ with $y-x\leq\delta$ we have $\displaystyle \sup_{z\in[x,y]} f(z)=\max\{f(x),f(y)\}$. \(ii) The function $f$ is called downwards $\delta$-monotone if for every interval $[x,y]\subset\Omega$ with $y-x\leq\delta$ we have $\displaystyle \inf_{z\in[x,y]} f(z)=\min\{f(x),f(y)\}$. \(iii) The function $f$ is called locally $\delta$-monotone if it is both downwards $\delta$-monotone and upwards $\delta$-monotone. The name locally $\delta$-monotone reflects the following characterization: $$\label{dmonchar} \begin{tabular}{l}$f$ is locally\\ $\delta$-monotone\end{tabular} \!\Longleftrightarrow\! \begin{tabular}{l}On any interval $[x,y]\!\subseteq\! \Omega$, $y\!-\!x\!\leq\!\delta$, $f$ is either\\ monotone increasing or monotone decreasing \end{tabular}$$ For every $\delta>0$ and $f\in\mathcal{A}(\Omega)$ the function $S_\frac{\delta}{2}(f)$ is upwards $\delta$-monotone while the function $I_\frac{\delta}{2}(f)$ is downwards $\delta$-monotone. Let $\delta>0$, $f\in\mathcal{A}(\Omega)$ and $[x,y]\subseteq\Omega$, $y-x\leq\delta$. Denote $g=S_\frac{\delta}{2}(f)$. It is easy to see that for every $z\in[x,y]$ we have $B_\frac{\delta}{2}(z)\subseteq B_\frac{\delta}{2}(x)\cup B_\frac{\delta}{2}(y)$. Therefore $$\begin{aligned} g(z)&=&\sup\{f(t):t\in B_\frac{\delta}{2}(z)\}\ \leq\ \sup\{f(t):t\in B_\frac{\delta}{2}(x)\cup B_\frac{\delta}{2}(y)\}\\ &=&\max\{\sup\{f(t):t\in B_\frac{\delta}{2}(x)\},\sup\{f(t):t\in B_\frac{\delta}{2}(z)\}\}\\ &=&\max\{g(x),g(y)\},\end{aligned}$$ which shows that function $g$ is upwards $\delta$-monotone. The downwards $\delta$-monotonicity of $U_\delta(f)$ is proved in a similar way. \[tLdupmon\] For every $\delta>0$ and $f\in\mathcal{A}(\Omega)$ the function $L_\delta(f)$ is upwards $\delta$-monotone while the function $U_\delta(f)$ is downwards $\delta$-monotone. \[tULlocdmon\] For every $\delta>0$ and $f\in\mathcal{A}(\Omega)$ the functions $(U_\delta\circ L_\delta)(f)$ and $(L_\delta\circ U_\delta)(f)$ are both locally $\delta$-monotone. The proof follows from Corollary \[tLdupmon\] and the composition table in the preceding section. Theorem \[tULlocdmon\] shows that the operators $U_\delta\circ L_\delta$ and $L_\delta\circ U_\delta$ provide locally $\delta$-monotone approximations to the functions in $\mathcal{A}(\Omega)$. The error of the approximation can be estimated in terms of the modulus of nonmonotonicity. Let us recall the definition. \[defmodulus\] Let $f\!\in\!\mathcal{A}(\Omega)$. The mapping $\mu(f,\cdot):\mathbb{R}^+\!\rightarrow\! \mathbb{R}^+\!\cup\!\{0\}$ given by $$\mu(f,\delta)\!=\!\frac{1}{2}\!\!\!\! \sup_{\begin{tabular}{c}$x_{1,2}\in\Omega$\\$0\!<\!x_2\!-\!x_2\!\leq\!\delta$\end{tabular}} \!\!\!\!\!\sup_{\begin{tabular}{c}$x\in[x_1,x_2]$\end{tabular}} \!\!(|f(x_1)\!-\!f(x)|\!+\!|f(x_2)\!-\!f(x)|\!-\!|f(x_1)\!-\!f(x_2)|)$$ is called modulus of nonmonotonicity of $f$. The locally $\delta$-monotone functions can be conveniently characterized through the modulus of nonmonotonicity. For any $f\in\mathcal{A}(\Omega)$ and $\delta>0$ we have $$\label{dmonchar2} f \mbox{ is locally }\delta\mbox{-monotone}\ \Longleftrightarrow \ \mu(f,\delta)=0$$ We will derive error estimates first in the case when $\Omega=\mathbb{R}$. It will prove useful to consider the upper semi-continuous envelope of the modulus of nonmonotonicity. Let $f\in\mathcal{A}(\Omega)$. Using that $\mu(f,\delta)$ is monotone increasing with respect to $\delta$ the upper semi-continuous envelope of $\mu$ can be represented as $\hat{\mu}(f,\delta)=S(\mu(f,\cdot))(\delta)=\lim_{\varepsilon\rightarrow 0^+}\mu(f,\delta+\varepsilon)$. \[testim1\] Let $f\in\mathcal{A}(\mathbb{R})$ and $\delta>0$. Then $f(x)-L_\delta(f)(x)\leq\hat\mu(f,\delta)$, $U_\delta(f)(x)-f(x)\leq\hat\mu(f,\delta)$, $x\in\Omega$. Let $x\in\mathbb{R}$. Denote $p=L_\delta(f)(x)$. If $p=f(x)$ the first inequality of the theorem holds. Assume that $p<f(x)$. Let $\eta>0$ be such that $p+\eta<f(x)$. Then we have $$\label{ineq1testim1} I_\frac{\delta}{2}(f)(y)<p+\eta,\ y\in B_\frac{\delta}{2}(x).$$ Denote $D_1=\{z\leq x:f(z)<p+\eta\}$, $D_2=\{z\geq x:f(z)<p+\eta\}$, $z_1=\sup D_1$, $z_2=\sup D_2$. Using the inequality (\[ineq1testim1\]) with $y=x-\frac{\delta}{2}$ and $y=x+\frac{\delta}{2}$ we obtain that $D_1$ and respectively $D_2$ are not empty and that $x-\delta\leq z_1\leq x\leq z_2\leq x+\delta$. Therefore $z_3=\frac{z_1+z_2}{2}\in\left[\frac{x-\delta+x}{2},\frac{x+x+\delta}{2}\right] =B_\frac{\delta}{2}(x)$. Then the inequality (\[ineq1testim1\]) implies that $B_\frac{\delta}{2}\cap (D_1\cup D_2)\neq\emptyset$. Hence $z_2-z_3=z_3-z_1\leq\frac{\delta}{2}$. Let $\varepsilon>0$ be arbitrary. The neighborhood $B_\frac{\delta+\varepsilon}{2}(z_3)$ has nonempty intersections with both $D_1$ and $D_2$. Let $t_1\in B_\frac{\delta+\varepsilon}{2}\cap D_1$ and $t_2\in B_\frac{\delta+\varepsilon}{2}\cap D_2$. We have $t_2-t_1<\delta+\varepsilon$ and $x\in[t_1,t_2]$. From the definition of the modulus of nonmonotonicity we have $|f(t_1)-f(x)|+|f(t_2)-f(x)|-|f(t_1)-f(t_2)|\leq 2\mu(f,\delta+\varepsilon)$. On the other side $|f(t_1)-f(x)|+|f(t_2)-f(x)|-|f(t_1)-f(t_2)|=2f(x)-f(t_1)-f(t_2)-|f(t_1)-f(t_2)| =2f(x)-\max\{f(t_1),f(t_2)\}>2f(x)-2(p+\eta)$. Therefore $f(x)-p-\eta<\mu(f,\delta+\varepsilon)$. Going with $\varepsilon$ to $0$ we obtain $f(x)-p\leq\hat\mu(f,\delta)+\eta$. Since $\eta$ is arbitrary small this implies the first inequality of the Theorem. The second inequality is proved in a similar manner. Using Theorem \[testim1\] as well as (\[dmonchar\]) and Corollary \[tLdupmon\] we have the following characterization of the fixed points of operators $L_\delta$ and $U_\delta$. For any $f\in\mathcal{A}(\mathbb{R})$ $$\hat\mu(f,\delta)=0\ \Longrightarrow\ (L_\delta(f)=f,\ U_\delta(f)=f)\ \Longrightarrow\ \mu(f,\delta)=0$$ \[testim2\] Let $\delta>0$ and $f\in\mathcal{A}(\mathbb{R})$ . Then $\mu(L_\delta(f),\delta)\leq\mu(f,\delta)$, $\hat\mu(L_\delta(f),\delta)\leq\hat\mu(f,\delta)$, $\mu(U_\delta(f),\delta)\leq\mu(f,\delta)$, $\hat\mu(U_\delta(f),\delta)\leq\hat\mu(f,\delta)$. We will prove the inequalities for $L_\delta(f)$ since the ones for $U_\delta(f)$ are proved in a similar way. Denote $g=L_\delta(f)$. Let $[x_1,x_2]$ be an arbitrary interval of length at most $\delta$ and let $x\in[x_1,x_2]$. We consider the number $q=|f(x_1)-f(x)|+|f(x_2)-f(x)|-|f(x_1)-f(x_2)|$. According to Corollary \[tLdupmon\] the function $g$ is upper $\delta$ monotone, which implies that $g(x)\leq\max\{g(x_1),g(x_2)\}$. If we also have $g(x)\geq\min\{g(x_1),g(x_2)\}$, then the number $q$ is zero and the first inequality of the Theorem is trivially satisfies. Let $g(x)<\min\{g(x_1),g(x_2)\}$ and let $\eta>0$ be such that $g(x)+\eta<\min\{g(x_1),g(x_2)\}$. The number $q$ can then be represented in the form $q=2(\min\{g(x_1),g(x_2)\}-g(x))$. If we assume that $f(y)\geq g(x)+\eta$ for all $y\in B_\frac{\delta}{2}(x)$, then $g(x)\geq I_\frac{\delta}{2}(f)(x)\geq g(x)+\eta$, which is a contradiction. Therefore, there exists $y\in B_\frac{\delta}{2}(x)$ such that $f(y)<g(x)+\eta$. If $y<x_1$ then using that $B_\frac{\delta}{2}(x_1)\subseteq B_\frac{\delta}{2}(y)\cup B_\frac{\delta}{2}(x)$ we obtain $I_\frac{\delta}{2}(z)<g(x)+\eta$ for all $z\in B_\frac{\delta}{2}(x_1)$ which implies $g(x_1)=S_\frac{\delta}{2}(I_\frac{\delta}{2}(f))(x_1)\leq g(x)+\eta<g(x_1)$. This contradiction shows that $y\geq x_1$. In a similar way we show that $y\leq x_2$. Using also the first inequality of Theorem \[tLdfleqf\] we have $q=2(\min\{g(x_1),g(x_2)\}-g(x))\leq 2(\min\{f(x_1),f(x_2)\}-f(y)-2\eta)\leq\mu(f,\delta)-2\eta$. Using that $\eta$ can be arbitrary small we obtain $q\leq \mu(f,\delta)$. Since the interval $[x_1,x_2]$ of length at most $\delta$ and $x\in[x_1,x_2]$ are arbitrary this implies that $\mu(g,\delta)\leq\mu(f,\delta)$. The inequality $\hat\mu(g,\delta)\leq\hat\mu(f,\delta)$ follows from the monotonicity of the operator $S$, see (\[IdSdmon\]). In the next theorem we give estimates for the error of approximation of a function $f\in\mathcal{A}(\mathbb{R})$ in terms of the supremum norm denoted here by $||\cdot||$. \[testim3\] Let $\delta>0$ and $f\in\mathcal{A}(\mathbb{R})$ . Then $||f-(L_\delta\circ U_\delta)(f)||\leq \hat\mu(f,\delta)$, $||f-(U_\delta\circ L_\delta)(f)||\leq \hat\mu(f,\delta)$. Applying Theorems \[testim1\] and \[testim2\] we obtain $(L_\delta\circ U_\delta)(f)(x)\leq U_\delta(f)(x)\ \leq\ f(x)+\hat\mu(f,\delta)$ and $(L_\delta\circ U_\delta)(f)(x)\geq U_\delta(f)(x)-\hat\mu(L_\delta(f),\delta)\ \geq\ f(x)-\hat\mu(f,\delta)$, which implies the first inequality of the Theorem. The second inequality is proved in a similar way. Error estimates similar to Theorem \[testim3\] can be derived in case of $\Omega$ being finite or semi-finite interval using a modification of the modulus of nonmonotonicity. For simplicity we will only consider the case $\Omega=[a,b]$, $a,b\in\mathbb{R}$. Let $f\!\in\!\mathcal{A}(\Omega)$. The mapping $\tilde\mu(f,\cdot):\mathbb{R}^+\!\!\rightarrow\!\! \mathbb{R}^+\!\cup\!\{0\}$ given by $$\tilde\mu(f,\delta)=\sup\left\{\hat\mu(f,\delta), \sup_{x_{1,2}\in[a,a+\frac{\delta}{2}]}(|f(x_1)\!-\!f(x_2)|, \sup_{x_{1,2}\in[b-\frac{\delta}{2},b]}(|f(x_1)\!-\!f(x_2)|\right\}$$ is called modified modulus of nonmonotonicity of $f$. This modulus is similar to the corrected modulus of nonmotonicity in [@Markov]. \[testim4\] Let $f\in\mathcal{A}(\mathbb{R})$ and $\delta>0$. Then $f(x)-L_\delta(f)(x)\leq\tilde\mu(f,\delta)$, $U_\delta(f)(x)-f(x)\leq\tilde\mu(f,\delta)$, $x\in\Omega$. The proof is similar to the proof of Theorem \[testim1\]. It is easy to see that for any $f\in\mathcal{A}[a,b]$ the functions $L_\delta(f)$ and $U_\delta(f)$ are constants on each of the intervals $[a,a+\frac{\delta}{2}]$ and $[b-\frac{\delta}{2}]$. Therefore, using also Theorem \[testim2\] we have $\tilde{\mu}(L_\delta(f),\delta)=\hat\mu(L_\delta(f),\delta)\leq \hat\mu(f,\delta)\leq\tilde\mu(f,\delta)$. In the same way we obtain $\tilde{\mu}(U_\delta(f),\delta)\leq\tilde\mu(f,\delta)$. Hence the modified modulus satisfies the similar inequalities to the ones given in Theorem \[testim2\] for $\mu$ and $\hat\mu$. Using these inequalities and Theorem \[testim4\] we obtain the error estimates in the next theorem, which are similar to the ones in Theorem \[testim3\]. \[testim5\] Let $\delta>0$ and $f\in\mathcal{A}(\mathbb{R})$ . Then $||f-(L_\delta\circ U_\delta)(f)||\leq \tilde\mu(f,\delta)$, $||f-(U_\delta\circ L_\delta)(f)||\leq \tilde\mu(f,\delta)$. Conclusion ========== In this paper we extended the LULU operators from sequences to real functions defined on a real interval using the lower and upper $\delta$-envelopes of functions. The obtained structure, although more general than the well known LULU structure of the discrete operators, retains some of its essential properties. For a fixed $\delta>0$ the compositions $L_\delta\circ U_\delta$ and $U_\delta\circ L_\delta$ provide locally $\delta$-monotone approximations for real functions, the error of approximation being estimated in terms of the modulus of nonmonotonicity of the functions. Further properties of the LULU operators for functions on continuous domains, e.g. trend preservation, will be investigated in the future. Generalizing the theory to functions on multidimensional domains is still an open problem. [9]{} S. Markov, Relations between the integral and Hausdorff distance with applications to differential equations, Pliska [**1**]{} (1977) 112–121. C. H. Rohwer, Variation reduction and $LULU$-smoothing, Quaestiones Mathematicae [**25**]{} (2002) 163–176. C. H. Rohwer, Fully trend preserving operators, Quaestiones Mathematicae [**27**]{} (2004) 217–230. C. H. Rohwer, Nonlinear Multiresolution Analysis, Birkhäuser, 2005. B. Sendov, Hausdorff Approximations, Kluwer, Boston, 1990. J. Serra, Image Analysis and Mathematical Morphology, Academic Press, London, 1982.

ArXiv

--- abstract: 'Results on the study of localized fluctuations in the multiplicity of charged particles and photons produced in 158$\cdot A$ GeV/c Pb+Pb collisions are presented for varying centrality. The charged versus neutral particle multiplicity correlations in common phase space regions of varying azimuthal size are analyzed by two different methods. Various types of mixed events are constructed to probe fluctuations arising from different sources. The measured results are compared to those from simulations and from mixed events. The comparison indicates the presence of non-statistical fluctuations in both the charged particle and photon multiplicities in limited azimuthal regions. However, no correlated charged-neutral fluctuations, a possible signature of formation of disoriented chiral condensates, are observed. An upper limit on the production of disoriented chiral condensates is set.' author: - 'M.M. Aggarwal,$^{1}$ Z. Ahammed,$^{2}$ A.L.S. Angelis,$^{3}$ V. Antonenko,$^{4}$ V. Arefiev,$^{5}$ V. Astakhov,$^{5}$ V. Avdeitchikov,$^{5}$ T.C. Awes,$^{6}$ P.V.K.S. Baba,$^{7}$ S.K. Badyal,$^{7}$ S. Bathe,$^{8}$ B. Batiounia,$^{5}$ T. Bernier,$^{9}$ K.B. Bhalla,$^{10}$ V.S. Bhatia,$^{1}$ C. Blume,$^{8}$ D. Bucher,$^{8}$ H. B[ü]{}sching,$^{8}$ L. Carlén,$^{11}$ S. Chattopadhyay,$^{2}$ M.P. Decowski,$^{12}$ H. Delagrange,$^{9}$ P. Donni,$^{3}$ M.R. Dutta Majumdar,$^{2}$ K. El Chenawi,$^{11}$ A.K. Dubey,$^{13}$ K. Enosawa,$^{14}$ S. Fokin,$^{4}$ V. Frolov,$^{5}$ M.S. Ganti,$^{2}$ S. Garpman,$^{11}$ O. Gavrishchuk,$^{5}$ F.J.M. Geurts,$^{15}$ T.K. Ghosh,$^{16}$ R. Glasow,$^{8}$ B. Guskov,$^{5}$ H. [Å]{}.Gustafsson,$^{11}$ H. H.Gutbrod,$^{17}$ I. Hrivnacova,$^{18}$ M. Ippolitov,$^{4}$ H. Kalechofsky,$^{3}$ K. Karadjev,$^{4}$ K. Karpio,$^{19}$ B. W. Kolb,$^{17}$ I. Kosarev,$^{5}$ I. Koutcheryaev,$^{4}$ A. Kugler,$^{18}$ P. Kulinich,$^{12}$ M. Kurata,$^{14}$ A. Lebedev,$^{4}$ H. L[ö]{}hner,$^{16}$ L. Luquin,$^{9}$ D.P. Mahapatra,$^{13}$ V. Manko,$^{4}$ M. Martin,$^{3}$ G. Martínez,$^{9}$ A. Maximov,$^{5}$ Y. Miake,$^{14}$ G.C. Mishra,$^{13}$ B. Mohanty,$^{2,13}$ M.-J. Mora,$^{9}$ D. Morrison,$^{20}$ T. Mukhanova,$^{4}$ D. S. Mukhopadhyay,$^{2}$ H. Naef,$^{3}$ B. K. Nandi,$^{13}$ S. K. Nayak,$^{7}$ T. K. Nayak,$^{2}$ A. Nianine,$^{4}$ V. Nikitine,$^{5}$ S. Nikolaev,$^{4}$ P. Nilsson,$^{11}$ S. Nishimura,$^{14}$ P. Nomokonov,$^{5}$ J. Nystrand,$^{11}$ A. Oskarsson,$^{11}$ I. Otterlund,$^{11}$ T. Peitzmann,$^{16}$ D. Peressounko,$^{4}$ V. Petracek,$^{18}$ W. Pinganaud,$^{9}$ F. Plasil,$^{6}$ M.L. Purschke,$^{17}$ J. Rak,$^{18}$ R. Raniwala,$^{10}$ S. Raniwala,$^{10}$ N.K. Rao,$^{7}$ F. Retiere,$^{9}$ K. Reygers,$^{16}$ G. Roland,$^{12}$ L. Rosselet,$^{3}$ I. Roufanov,$^{5}$ C. Roy,$^{9}$ J.M. Rubio,$^{3}$ S.S. Sambyal,$^{7}$ R. Santo,$^{8}$ S. Sato,$^{14}$ H. Schlagheck,$^{8}$ H.-R. Schmidt,$^{17}$ Y. Schutz,$^{9}$ G. Shabratova,$^{5}$ T.H. Shah,$^{7}$ I. Sibiriak,$^{4}$ T. Siemiarczuk,$^{19}$ D. Silvermyr,$^{11}$ B.C. Sinha,$^{2}$ N. Slavine,$^{5}$ K. S[ö]{}derstr[ö]{}m,$^{11}$ G. Sood,$^{1}$ S.P. S[ø]{}rensen,$^{20}$ P. Stankus,$^{6}$ G. Stefanek,$^{19}$ P. Steinberg,$^{12}$ E. Stenlund,$^{11}$ M. Sumbera,$^{18}$ T. Svensson,$^{11}$ A. Tsvetkov,$^{4}$ L. Tykarski,$^{19}$ E.C.v.d. Pijll,$^{15}$ N.v. Eijndhoven,$^{15}$ G.J.v. Nieuwenhuizen,$^{12}$ A. Vinogradov,$^{4}$ Y.P. Viyogi,$^{2}$ A. Vodopianov,$^{5}$ S. V[ö]{}r[ö]{}s,$^{3}$ B. Wys[ł]{}ouch,$^{12}$ G.R. Young$^{6}$' title: 'Centrality dependence of charged-neutral particle fluctuations in 158$\cdot$A GeV $^{208}$Pb+$^{208}$Pb collisions' --- INTRODUCTION ============ The large number of particles produced in relativistic heavy-ion collisions at the Super Proton Synchrotron (SPS) and the Relativistic Heavy-Ion Collider (RHIC) provide an opportunity to analyze and study, on an event-by-event basis, fluctuations in physical observables, such as particle multiplicities, transverse momenta, and their correlations. These studies provide information on the dynamics of multi-particle production and may help to reveal the phase transition from hadronic matter to quark-gluon plasma (QGP) [@henning; @stephanov]. The formation of hot and dense matter in high energy heavy-ion collisions also has the possibility to create matter in a chiral symmetry restored phase in the laboratory. After the initial stage of the collision, the system cools and expands returning to the normal QCD vacuum in which chiral symmetry is spontaneously broken. During this process, a metastable state may be formed in which the chiral condensate is disoriented from the true vacuum direction. This transient state would subsequently decay by emitting pions coherently within finite sub volumes or domains of the collision region. The possibility of formation of disoriented chiral condensate (DCC) has been discussed extensively in recent years [@anselm; @bj; @raj; @huang1; @raj2; @gavin1; @asakawa; @koch; @gavin-kapusta]. The detection and study of the DCC state would provide valuable information about the chiral phase transition and the vacuum structure of strong interactions. Theoretical studies [@raj; @huang1; @raj2] suggest that isospin fluctuations caused by formation of a DCC would produce clusters of coherent pions in localized phase space domains. The formation of DCC domains would be associated with large event-by-event fluctuations in the ratio of neutral to charged pions. The probability distribution of the neutral pion fraction, $f$, in such a DCC domain has been shown [@anselm] to follow the relation: $$P(f) = \frac{1}{2\sqrt{f}} {\rm ~~~~~~~~~~where~~~~} f = N_{\pi^0}/N_{\pi} \label{eqn1}$$ which is quite different from that of the normal pion production mechanism. For the normal pion production, where the production of $\pi^{0}$,$\pi^{+}$, $\pi^{-}$ are equally probable, the $f$ distribution is binomial peaking at $1/3$. In the experimental search for localized domains of DCC a practical approach is to search for events with large and localized fluctuations (localized in pseudo-rapidity ($\eta$) and azimuthal angle ($\phi$) ) in the ratio of the number of photons to charged particles which would directly reflect fluctuations in the neutral to charged pion ratio. Typical event structures would be similar to the Centauro and anti-Centauro events reported by the JACEE collaboration [@jacee]. Results from other cosmic ray experiments have not ruled out the possibility of the DCC formation mechanism [@augusto]. The accelerator based studies carried out in $p-\bar{p}$ [@minimax] and heavy ion [@WA98-3; @NA49pt] reactions have investigated particle production over extended regions of phase space. These analysis were not sensitive to the presence of small domains of DCC localized in phase space. A first search for evidence of localized domains of DCC has been carried out at the SPS by the WA98 experiment in a detailed study of central Pb+Pb events  [@WA98-12]. The analysis showed the presence of localized non-statistical fluctuations in the multiplicity of both photons and charged particles. However, the charged-neutral fluctuations were found not to be correlated event-by-event, as would be expected for a DCC production mechanism. An upper limit on the frequency of DCC formation in central Pb+Pb collisions was set. Recently there have been theoretical suggestions to look for DCC formation in events for intermediate centralities [@notsocentral]. In this paper we present first results on the centrality dependence of localized charged-neutral multiplicity fluctuations. It is based on an analysis of event-by-event fluctuation in the relative number of charged particles and photons detected within the common acceptance of the photon and charged particle multiplicity detectors of the WA98 experiment [@wa98]. The paper is organized in the following manner: In the next section we describe the detectors used for the present analysis, the centrality selection criteria, the data reduction, and simulation. Section III deals with the analysis techniques, where two analysis methods are presented, one based on the correlation of photons and charged particles, and the other based on a discrete wavelet transformation analysis. In section IV, we present in detail the construction of mixed events used for this study. Section V discusses the ability of the mixed events to probe specific fluctuations. Final results and discussion are given in section VI. A summary is presented in section VII. EXPERIMENTAL SETUP AND DATA REDUCTION ===================================== In the WA98 experiment at CERN [@wa98], the main emphasis has been on high precision, simultaneous detection of both hadrons and photons. The experimental setup consisted of large acceptance hadron and photon spectrometers, detectors for charged particle and photon multiplicity measurements, and calorimeters for transverse and forward energy measurements. The present study makes use of the data from the photon multiplicity detector (PMD), the silicon pad multiplicity detector (SPMD) and the mid-rapidity calorimeter (MIRAC). Centrality Selection -------------------- The centrality of the interaction was determined from the total transverse energy ($E_{\mathrm T}$) measured by the mid-rapidity calorimeter (MIRAC) [@awes]. The MIRAC measures both the transverse electromagnetic ($E_{\mathrm T}^{em}$) and hadronic ($E_{\mathrm T}^{had}$) energies in the interval $3.5\le\eta\le 5.5$ with a resolution of $17.9\%/\sqrt E$ and $46.1\%/ \sqrt E$, respectively, where $E$ is expressed in GeV. The centrality, or impact parameter of the collision, has a strong correlation with the amount of $E_{\mathrm T}$ produced. Events with large $E_{\mathrm T}$ production correspond to the most central, small impact parameter, collisions [@WA98-10]. The centralities are expressed as fractions of the minimum bias cross section as a function of the measured total $E_{\mathrm T}$. For the present analysis we have used data selections in four centrality bins, the top $5\%$ (henceforth referred to as centrality-1), $5\%$ - $10\%$ (centrality-2), $15\%$ - $30\%$ (centrality-3), and $45\% - 55\%$ (centrality-4) of the minimum bias cross section. The minimum bias distribution of the total $E_{\mathrm T}$ is shown in Fig. \[et\_cen\]. The centrality bins used in this analysis are marked in the figure. Photon Multiplicity Detector ---------------------------- The photon multiplicity is measured using the preshower photon multiplicity detector (PMD) located at a distance of 21.5 meters from the target. The PMD consists of 3 radiation lengths ($X_0$) thick lead converter plates in front of an array of square scintillator pads of four sizes, varying from 15 mm$\times$15 mm to 25 mm$\times$25 mm, placed in 28 box modules. Each box module consists of a matrix of $38\times$50 scintillator pads read out using an image intensifier plus charged coupled device (CCD) camera system. The scintillation light is transmitted to the readout device via a short wavelength shifting fiber spliced to a long extra-mural absorber (EMA) coated clear fiber. The total light amplification of the readout system is $\sim$40000. Digitization of the CCD pixel charge is done by a set of custom built fast-bus modules employing an 8 bit 20 MHz Flash ADC system. Details of the design and characteristics of the PMD may be found in Ref. [@pmd_nim; @WA98-9]. The results presented here make use of the data from the central 22 box modules covering the pseudo-rapidity range of $2.9\le \eta\le 4.2$. The clusters of hit pads, having total ADC content above a hadron rejection threshold are identified as photon-like, the multiplicity being denoted by $N_{\gamma-{\mathrm {like}}}$. If the number of incident photons is denoted by $N_{\gamma}^{inc}$ and the number of photons detected above the hadron rejection threshold as $N_{\gamma}^{th }$, then the photon counting efficiency ($\epsilon_{\gamma}$) and purity of photon sample ($f_p$) are defined as, $\epsilon_{\gamma}~=~N_{\gamma}^{th}/N_{\gamma}^{inc}$ and $f_p~=~N_{\gamma}^{th}/N_{\gamma-{\mathrm {like}}}$ respectively. These are estimated from detector simulations [@pmd_nim; @WA98-9]. The photon counting efficiencies for the central to peripheral cases varies from $68\%$ to $73\%$. The purity of the photon sample in the two cases varies from $65\%$ to $54\%$ [@pmd_nim; @WA98-9]. The acceptance in terms of transverse momentum ($p_{T}$) extends down to about $30~MeV/c$, however the PMD energy resolution is not sufficient for particle-by-particle $p_{T}$ measurement. Silicon Pad Multiplicity Detector --------------------------------- The charged particle multiplicity ($N_{\mathrm {ch}}$) is measured using the circular Silicon Pad Multiplicity Detector (SPMD) located 32.8 cm from the target and having full azimuthal coverage in the region $2.35< \eta < 3.75$, corresponding to the central rapidity region at SPS energies (where $\eta_{CMS} = 2.92$). The detector consists of four overlapping quadrants, each fabricated from a single 300 [$\mu m$]{} thick silicon wafer. The active area of each quadrant is divided into 1012 pads forming 46 azimuthal wedges and 22 radial bins with a pad size increasing with radius to provide equal size pseudo-rapidity bins. The efficiency for detection of a charged particle in the active area has been determined in test beam measurements to be better than $99\%$. Conversely, the detector is transparent to high energy photons, since only about $0.2\%$ are expected to interact in the silicon. During the data recording, $95\%$ of the pads worked properly and are used in this analysis. Details of the characteristics of the SPMD can be found in Ref. [@WA98-3; @spmd_nim]. The SPMD does not provide $p_{T}$ measurement, but provides the multiplicity measurement integrated over transverse momentum ($p_{T}$) with a threshold which extends down to about $20~MeV/c$. Data Reduction -------------- The data presented here were taken during December 1996 at the CERN SPS with the 158$\cdot A$ GeV Pb ion beam on a Pb target of thickness 213 $\mu$m. The WA98 Goliath magnet was switched off during these runs. Events with beam pile-up, downstream interactions, and pile-up in the CCD camera system were rejected in the off-line analysis [@WA98-3; @WA98-9]. The data have been analyzed for the region of common $\eta$ ($2.9< \eta < 3.75$) and $\phi$ coverage of the SPMD charged particle and PMD photon multiplicity detectors. The $N_{\gamma-{\mathrm {like}}}$ and the $N_{\mathrm {ch}}$ distributions for the four centrality bins are shown in Figs. \[ngam\_cen\] and \[nch\_cen\]. The number of events analyzed, the mean number of photons and charged particles along with the root mean square deviations are shown in the figures. The PMD and SPMD detectors provide momentum integrated multiplicity measurements with very low thresholds. Since pions from DCC domains are expected to have small $p_T$ values, below the pion mass, the momentum integration will dilute the DCC signal. On the other hand, the large coverage of the PMD and SPMD are important to overcome the limitations of small number fluctuations. The various sources of systematic errors associated with the $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ distributions have been investigated and described in detail previously [@WA98-9; @WA98-15]. These include: - \(a) The uncertainty in the energy calibration and the associated uncertainty in the energy threshold for hadron rejection in the PMD leads to an error in the efficiency for $N_{\gamma-{\mathrm {like}}}$ clusters. The nominal hadron rejection threshold was set at three times the minimum ionizing particle (MIP) peak. The value of the threshold was reduced by $10\%$ [@WA98-9] in order to estimate the systematic error. The associated error in $N_{\gamma-{\mathrm {like}}}$ is $2.5\%$. - \(b) The error due to the effect of clustering of pad signals in the PMD is a major source of error in $N_{\gamma-{\mathrm {like}}}$. This error is determined from GEANT [@geant] simulation by comparing the number of known photon tracks on the PMD with the total number of reconstructed photon-like clusters. It is found that the number of clusters exceeds the number of tracks by $3\%$ in the case of peripheral events and by $7\%$ for high multiplicity central events. - \(c) The error due to the variation in pad-to-pad gains of the scintillators in PMD was found to be less than $1\%$. - \(d) The uncertainty in the $N_{\mathrm {ch}}$ determination with the SPMD has been estimated to be about $4\%$ [@WA98-3]. - \(e) The error due to the finite resolution in the measurement of the total transverse energy($E_T$) in MIRAC [@WA98-9] translates into an uncertainty in the centrality selection. The effect of this systematic error has been determined by performing the analysis with varying centrality cut within the MIRAC resolution. The contribution of each of these various systematic errors to the final results are discussed in the following sections. Simulated Events ---------------- Simulated events were generated using the VENUS 4.12 event generator [@venus] with the default parameter values. The output was processed through a detector simulation package in the GEANT 3.21 [@geant] framework. This simulation includes the full WA98 experimental setup and includes experimental effects such as photon conversions, downstream interactions, hadron backgrounds in the PMD, etc. which might dilute or enhance the observed fluctuations. The effect of Landau fluctuations in the energy loss of charged particles in silicon was included in the SPMD simulation [@WA98-3]. For the PMD simulation, the GEANT results in terms of energy deposition in pads were converted to the pad ADC values using the MeV-ADC calibration relation. After this the ADC distribution is convoluted with a Gaussian function of proper width taken from the readout resolution curve. If the energy deposition is less than 3 MIP, a Landau distribution is used for convolution. The details of the PMD simulations taking into account the detector and readout effects can be found in Ref. [@pmd_nim]. The centrality selection with the simulated data has been made in an identical manner to the data, determined from the simulated total transverse energy in MIRAC. The minimum bias total $E_{\mathrm T}$ distribution predicted by VENUS is shown by the dashed histogram in Fig. \[et\_cen\]. The agreement with the data is seen to be quite reasonable. A total of 60K VENUS events with simulated detector response were generated for the present study. These simulated events (henceforth referred to simply as VENUS events unless otherwise specified) were then processed with the same analysis codes as used for the analysis of the experimental data. ANALYSIS TECHNIQUES =================== Two different analysis methods have been used in the present study. In the first analysis method, the magnitude of the $N_{\gamma-{\mathrm {like}}}$ versus $N_{\mathrm {ch}}$ multiplicity fluctuations is obtained in decreasing phase space regions. The second method employed the discrete wavelet transformation technique to investigate the relative magnitude of the $N_{\gamma-{\mathrm {like}}}$ versus $N_{\mathrm {ch}}$ fluctuations in adjacent phase space regions. The results from these methods of analysis applied to data, VENUS and various sets of mixed events (to be discussed later) are compared to draw proper conclusions. $N_\gamma$ versus $N_{\mathrm {ch}}$ correlations ------------------------------------------------- In order to search for localized fluctuations in the photon and charged particle multiplicities, which may have non-statistical origin, the correlation between $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ is investigated at various scales in $\phi$. The event-by-event correlation between $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ has been studied in various $\phi$-intervals by dividing the entire $\phi$-space into 2, 4, 8, and 16 bins. The method of analysis is similar to that described in Refs. [@WA98-3; @WA98-12]. Fig. \[ngam\_nch\_cor\] shows scatter plot of the correlation between $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ for the top centrality bin. The correlation plots for each $\phi$ interval size, including the case of the full interval with no segmentation, are shown. The distributions for the other three centrality bins are qualitatively similar. A common correlation axis ($Z$) has been obtained for the full distribution by fitting the $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ correlation with a second order polynomial. The correlation axis with fit parameters is shown in the figure. The distance of separation ($D_{Z}$) between a data point and the correlation axis has been calculated with the convention that $D_{Z}$ is positive for points below the $Z$-axis. The distribution of $D_Z$ represents the relative fluctuations of $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ from the correlation axis for any chosen $\phi$ bin size. In order to compare the fluctuations for different $\phi$ bins on a similar footing, a scaled variable, $S_{Z} = D_Z/s(D_Z)$, is used where $s(D_Z)$ represents the rms deviation of the $D_{Z}$ distribution for VENUS events analyzed in the same manner. The $D_Z$ distributions of data, mixed events and the simulated events for a given centrality and $\phi$ bin size are all scaled by the same $s(D_Z)$ corresponding to the VENUS events for the respective centrality and azimuthal bin size. The presence of events with localized fluctuations in $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$, at a given $\phi$ bin, is expected to result in a broader distribution of $S_Z$ compared to those for normal events. Comparing the rms deviations of the $S_Z$ distributions of data, mixed events (to be discussed later), and VENUS events may allow to infer the presence of non-statistical localized fluctuations. Multi-resolution DWT analysis ----------------------------- A multi-resolution analysis using discrete wavelet transformations (DWT) [@amara] has been shown to be quite powerful in the search for localized domains of DCC [@huang; @dccflow; @dccstr]. The beauty of the DWT technique lies in its power to analyze a spectrum at different resolutions with the ability to identify fluctuations present at any scale. This method has been utilized very successfully in many fields including image processing, data compression, turbulence, human vision, radar, and earthquake prediction [@amara]. It should be noted that the DWT analysis provides different information than the moment analysis of the previous section. It analyzes the event-by-event distribution in phi space to characterize the bin-to-bin fluctuations relative to the average behaviour. For the present DWT analysis the full azimuthal region is divided into smaller bins in $\phi$, the number of bins at a given scale $j$ being $2^j$. The input to the analysis is a spectrum of the sample function at the smallest bin in $\phi$ corresponding to the highest resolution scale, $j_{max}$. In the present case the sample function is chosen to be the photon fraction, given by: $$f^\prime(\phi) = {N_{\gamma-{\mathrm {like}}}(\phi)}/ {(N_{\gamma-{\mathrm {like}}}(\phi)+N_{\mathrm {ch}}(\phi))}$$ A multi-resolution analysis has been carried out using the $D-$4 wavelet basis on the above sample function starting with $j_{max}=5$. It may be mentioned that there are several families of wavelet bases distinguished by the number of coefficients and the level of iteration; we have used the frequently employed $D-$4 wavelet basis [@numeric]. The output of the DWT consists of a set of wavelet or father function coefficients (FFC) at each scale, from $j=1$,...,($j_{max}-1$). The coefficients obtained at a given scale, $j$, are derived from the distribution of the sample function at one higher scale, $j+1$. The FFCs quantify the bin-to-bin fluctuations in the sample function at that higher scale relative to the average behaviour. The presence of localized non-statistical fluctuations will increase the rms deviation of the distribution of FFCs and may result in non-Gaussian tails [@huang; @dccstr]. The DWT technique as used in this analysis has been demonstrated in our earlier publication [@WA98-12]. Once again, comparing the rms deviations of the FFC distributions of data, mixed events, and VENUS events may allow to infer the presence of localized fluctuations. The utility of the mixed events and the response of the analysis technique is demonstrated in Section V. CONSTRUCTION OF MIXED EVENTS ============================ It is possible to search for non-statistical fluctuations in the experimental data in a model independent way by comparison of the data with mixed events generated from the data itself. Furthermore, it is necessary to isolate the various contributions to the fluctuations and to understand all detector related effects in the data. This has been done by generating different types of mixed events which preserve the global multiplicity correlation between $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$. Fluctuations in the ratio of $N_{\gamma}$ to $N_{\mathrm {ch}}$ can arise due to fluctuations in $N_{\gamma}$ only, fluctuations in $N_{\mathrm {ch}}$ only, or fluctuation in both $N_{\gamma}$ and $N_{\mathrm {ch}}$. Furthermore, the fluctuations in $N_{\gamma}$ and $N_{\mathrm {ch}}$ may be correlated event-by-event, as nominally expected in the case of DCC formation. Each of these possibilities is investigated through the construction of four different kinds of mixed events. The method of construction of these mixed events and the type of fluctuations they probe are described next. Maximally Mixed Events ---------------------- The first set of mixed events, referred to as M1 events, are constructed to remove all correlations to the greatest extent possible to provide a baseline for comparison to the real event data. They were generated by mixing hits in both the photon and charged particle detectors separately but still satisfying the global $N_{\gamma-{\mathrm {like}}}$-$N_{\mathrm {ch}}$ correlation of the real event in the full acceptance. This means that on an event-by-event basis the total photon multiplicity and charged particle multiplicity of the mixed event were identical to those of the real event to which it corresponds. Thus, the scatter plot of the mixed events is identical to the real events for the single bin case shown in Fig. \[ngam\_nch\_cor\]. The idea is to constrain the mixed events to be identical to real events globally and then compare them to the real data in localized regions of phase space to search for indications of non-statistical localized fluctuations in the data. The M1 type of mixed events were constructed from the pool of all photon-like and charged particle hits, in which the hit position ($\eta$ and $\phi$) and event information (event number and total multiplicity) was kept for both the photon-like and charged particle hits. For a given real event measured to have multiplicities $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ in the full acceptance region, a mixed event was constructed by randomly selecting $N_{\gamma-{\mathrm {like}}}$ photon-like hits from the pool of photon-like hits and $N_{\mathrm {ch}}$ charged particle hits from the pool of charged particle hits. This procedure was repeated for each real event. Care was taken such that no two hits from the same real event were used in the construction of a mixed event. Also, for the mixed events, hits within either the SPMD or PMD detector were not allowed to lie within the two-track resolution of that detector. In brief, the M1 mixed events randomly distribute the hits in each individual detector but keep the global correlation between the $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ multiplicity. They provide a maximally randomized sample of PMD and SPMD hits. Comparisons of such mixed events to real events will be most sensitive to the presence of localized fluctuations. However, in themselves they would not isolate the source of fluctuations as being due to $N_{\gamma}$ and/or $N_{\mathrm {ch}}$, or correlations between $N_{\gamma}$ and $N_{\mathrm {ch}}$. Minimally Mixed Events ---------------------- A second type of mixed events, referred to as M2 mixed events, were constructed to investigate the presence of correlated event-by-event fluctuations between $N_{\gamma}$ and $N_{\mathrm {ch}}$. These mixed events had a minimal amount of randomization since they were generated by mixing the photon hits taken unaltered from one event with the charged particle hits taken unmodified from another event. As with the M1 mixed events, the global $N_{\gamma-{\mathrm {like}}}$-$N_{\mathrm {ch}}$ multiplicity correlation was maintained to be exactly the same as the data. To construct such mixed events, for a given real event measured to have multiplicities $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ in the full acceptance region, a mixed event was constructed by keeping the PMD $N_{\gamma-{\mathrm {like}}}$ portion of the event intact, but combining it with the unaltered SPMD portion of a different randomly selected event, but constrained to have almost the same charged particle multiplicity $N_{\mathrm {ch}}$. This procedure was repeated for each real event. In brief, this type of mixed event keeps the event-by-event hit structure in each detector identical to that in real events. Thus, such mixed events keep the individual localized fluctuations present in $N_{\gamma}$ or $N_{\mathrm {ch}}$, but remove the event-by-event localized correlated fluctuations between them. Comparison of such mixed events to real events may reveal the presence of correlated localized fluctuations between $N_{\gamma}$ and $N_{\mathrm {ch}}$. Partially Mixed Events ---------------------- Intermediate between the M1 and M2 types of mixed events are a third and fourth type of partially mixed event, referred to as M3-$\gamma$ and M3-${\mathrm {ch}}$ mixed events. These were constructed to provide information regarding the contribution to the localized fluctuations in the $N_{\gamma}$ to $N_{\mathrm {ch}}$ ratio from the individual observables ($N_{\gamma}$ and $N_{\mathrm {ch}}$). They were generated from real events by mixing hits in one of the detectors (following the procedure for construction of M1 mixed events) and keeping the hit structure of the event in the other detector intact. M3-$\gamma$ mixed events correspond to the case where the hits within the photon detector are unaltered while the hits in the charged particle detector are mixed. Similarly in M3-$\mathrm {ch}$ mixed events the hits in the charged particle detector were unaltered and the hits in the photon detector were mixed. In each type of mixed event the global $N_{\gamma-{\mathrm {like}}}$–$N_{\mathrm {ch}}$ correlation is maintained as in the real event. The two track resolution in the detectors where the hits are mixed is kept identical to that in real events. The total number of mixed events is the same as the number of real events. Comparison of such mixed events to real events and the other types of mixed events will reveal the presence of localized fluctuations in $N_{\gamma}$ or $N_{\mathrm {ch}}$ separately. A summary of the different sources of fluctuations in the ratio of $N_{\gamma}$ to $N_{\mathrm {ch}}$ probed by each of the types of mixed events is given in Table 1. --------------------------------------------- -- ---- ------- ------------------- ------------- Fluctuation Mixed Event M1 M2 M3-$\mathrm {ch}$ M3-$\gamma$ $N_{\gamma}$-only No Yes No Yes $N_{\mathrm {ch}}-only$ No Yes Yes No correlated $N_{\gamma}$-$N_{\mathrm {ch}} $ No No No No --------------------------------------------- -- ---- ------- ------------------- ------------- : \[tab:table1\] Type of fluctuations preserved by various mixed events DEMONSTRATION OF ANALYSIS METHOD ================================ In this section we wish to demonstrate the analysis method and illustrate how the relationship of the measured result to that obtained with the various mixed events can be used to provide an essentially model-independent signature of DCC formation. To demonstrate the analysis method and the potential to observe DCC event formation, we have applied the DWT analysis to a simple DCC-like model. The analysis is applied to “real” DCC events from the model as well as the various types of mixed events described in the previous section constructed from the model DCC events. Since event generators which include DCC formation do not exist, we have implemented a simple DCC model in which localized non-statistical $N_{\gamma}$-$N_{ch}$ fluctuations have been introduced by modification of the output of the VENUS event generator. To implement the fluctuations, the final state pions within a localized $\eta$–$\phi$ region from VENUS are interchanged pairwise ($\pi^{+}\pi^{-} \leftrightarrow \pi^{0}\pi^{0}$) according to the DCC probability distribution, $P(f) = 1 / 2 \sqrt{f}$. The fluctuations were generated over a localized region of $\eta=3-4$ and a $\Delta\phi$ interval of $90^{0}$. The $\pi^{0}$’s were then allowed to decay. The resulting events were then passed through the WA98 detector response simulation. The DCC events in the simple model used here give rise to an anti-correlation between $N_{\gamma}$ and $N_{\mathrm ch}$. It also results in non-random fluctuations in both $N_{\gamma}$ and $N_{\mathrm ch}$ individually. Since the probability to produce events with localized charged-neutral fluctuations is unknown, ensembles of events, here referred to as “nDCC events”, were produced as a mixture of normal VENUS events and events with localized fluctuations. The fraction of events with localized DCC-like fluctuations in each nDCC sample was varied as a parameter to be studied. By using VENUS events as the basis to introduce the DCC effect, it is insured that the general features of the event, including the multiplicity, composition, and momenta of the produced particles, as well as correlations in the particle multiplicities due to impact parameter variation, are reasonably well described. Also the GEANT simulation of the detector response to these events insures that other effects such as photon conversions and the response of the PMD to hadrons, which might affect the observed multiplicities and the observed correlations, are also taken into account. While the assumption of a 100% DCC contribution over an interval of fixed size in $\eta-\phi$ is certainly a gross simplification, it provides a well-defined reference to gauge the potential for DCC observation. Other assumptions, such as a varying fraction, which might also be momentum dependent (since DCCs are expected to be a low p$_T$ phenomena), and different or varying sizes might be more realistic. However, without clear theoretical guidance we have chosen this very simple model as a reasonable and well-defined reference. The DWT analysis was carried out on an ensemble of nDCC events and their corresponding mixed event sets created from each set of nDCC events. By ensemble of nDCC events we mean sets of events having different percentages of events with localized fluctuations. The percentage varied from zero, that is, normal VENUS events with no localized fluctuations, to an event set where all events had localized charged-neutral fluctuations. For every nDCC event set the four sets of mixed events (M1,M2, M3-$\gamma$, and M3-ch) were constructed from the nDCC events and analyzed. The rms deviations of the FFC distributions for each set of nDCC events and corresponding mixed event sets were obtained. The results for scale $j=1$ are plotted in Fig. \[rms\] as a function of the percentage of localized DCC-type events in the nDCC event set. It is seen that the rms deviations of the FFC distributions of the nDCC events increase as the percentage of events with localized fluctuations in $N_{\gamma}$-$N_{\mathrm ch}$ increases. This is the expected behaviour and demonstrates the linear response of the DWT to the DCC events when the frequency of events with fluctuations increases. On the other hand, the rms deviations of the FFC distributions of the M1 type mixed events created from the nDCC events are found to be independent of the percentage of events having localized fluctuations. This is also the expected behaviour and demonstrates that the M1 mixed events can be used as a baseline from which to deduce the presence of fluctuations in a model-independent manner. However, the deviation of the real events from the M1 mixed events does not inform about the relative contributions of the individual $N_{\gamma}$ and $N_{\mathrm ch}$ fluctuations. The rms deviations of the FFC distributions of the M3 mixed events are found to be intermediate to those obtained for nDCC events and M1 mixed events. They indicate the separate contributions of the $N_{\gamma}$ or $N_{\mathrm ch}$ fluctuations alone to the ratio. The rms deviations of the M2 mixed events are higher than those of M1 and M3 mixed events. That is because the M2 mixed events keep the separate contributions of both the $N_{\gamma}$ and $N_{\mathrm ch}$ fluctuations. However, the rms deviations of the M2 mixed events are consistently below those for the nDCC events. This is because the M2 events randomize the correlations between $N_{\gamma}$ and $N_{\mathrm ch}$. The difference between the M2 mixed events and the nDCC events indicates the presence of the DCC-like correlated $N_{\gamma}$-$N_{\mathrm ch}$ fluctuations. The relative pattern of rms values for the real events and the various mixed events constructed from those real events seen in Fig. \[rms\] provides a rather unambiguous model-independent signature for DCC-like fluctuations. Similar relative pattern of the rms deviations of the $S_{Z}$ distributions for mixed events and simulated events were also observed for $N_{\gamma}$-$N_{\mathrm ch}$ correlation analysis. In particular, the observation of fluctuations in real events that are greater than the M2 mixed events would constitute what might be called “smoking gun” evidence for DCC formation. Conversely, the lack of a difference between real events and M2 mixed events would indicate the lack of DCC-like correlated charged-neutral fluctuations. Figure \[rms\] also demonstrates an effect which must be taken into account when comparing the measured result to the mixed events. For nDCC events with vanishing fraction of events with fluctuations, which is to say for normal VENUS events, it is seen that the rms deviations of the FFC distributions for all types of mixed events are higher than those for the nDCC events. In the VENUS simulations this is due to the presence of correlations between $N_{\mathrm ch}$ and $N_{\gamma-{\mathrm like}}$. These are primarily due to residual impact parameter correlations (see Fig. \[ngam\_nch\_cor\]) as well as due to the charged particle contamination in the $N_{\gamma-{\mathrm like}}$ data sample whereby the charged particles register in both the PMD and SPMD (see Section II.C) [@bedanga]. These correlations are removed by the event mixing procedure which results in a larger rms deviations for the mixed events. The real data is presumed to have similar residual correlations as observed in the VENUS simulations. In order to correct for the effects of these non-DCC correlations, all mixed event rms values constructed from real events have been rescaled by the percentage difference between the rms deviations of the VENUS distributions and those of the corresponding VENUS mixed events, as also discussed in Ref. [@WA98-12]. For nDCC events with larger percentages of DCC-like events, the anti-correlation between $N_{\mathrm ch}$ and $N_{\gamma-{\mathrm like}}$ overcomes the correlations between $N_{\mathrm ch}$ and $N_{\gamma-{\mathrm like}}$ and hence the rms deviations of the FFC distributions of nDCC events become greater than those of the mixed events, despite the other correlation effects. RESULTS AND DISCUSSION ====================== In the analysis of experimental data, the results from the measured data are compared with simulated and mixed events. Below we discuss the results obtained from such a comparison using two different analysis methods discussed earlier. $N_\gamma$ versus $N_{\mathrm {ch}}$ correlation results -------------------------------------------------------- The $S_{Z}$ distributions calculated for different $\phi$ bin sizes are shown in Fig. \[sz\] for data, M1, and VENUS events, for the four different centrality selections. The distributions for the other types of mixed events are not shown for clarity of presentation. The small differences in the $S_Z$ distributions have been quantified in terms of the corresponding rms deviations, of these distributions shown in Fig. \[sz\_rms\]. The statistical errors on the values are small and are within the size of the symbols. The bars represent statistical and systematic errors added in quadrature. The various sources of systematic error have been discussed in an earlier section. An additional systematic error from the fit errors associated with the determination of the correlation axis ($Z$) is also included. In general, the width of the $S_Z$ distribution increases from the most central to less central event selections and decreases with decreasing bin size. The rms deviations of the $S_Z$ distribution for the VENUS simulated events are $1$ (by definition) for all centrality and all bins in azimuth by definition of $S_{Z}$ and are significantly different from the measured results. This is primarily because the global particle multiplicity as predicted by VENUS and those measured in the experiment within the coverage of the detectors are not same. As seen in Fig. \[sz\_rms\], the widths of the $S_Z$ distributions for mixed events closely follow those of the data. The mixed events have been constructed such that the global $N_{\gamma-{\mathrm {like}}}$ ${\mathrm {vs.}}$ $N_{\mathrm {ch}}$ multiplicity correlations are maintained. Therefore the rms deviations of the data and the mixed event reference are the same in the first $\phi$ bin of Fig. \[sz\_rms\] by construction. Some correlations between $N_{\mathrm {ch}}$ and $N_{\gamma-{\mathrm {like}}}$ are expected, mostly as a result of the charged particle contamination in the $N_{\gamma-{\mathrm {like}}}$ data sample, but are removed by the event mixing procedure and thereby result in a small difference between the real and mixed events, as seen in the analysis of the VENUS events, discussed earlier. All of the mixed event $S_{Z}$ distribution rms values ( Fig. \[sz\_rms\]) have therefore been rescaled by the percentage difference between the rms deviations of the VENUS distributions and those of the corresponding VENUS mixed events for each centrality class (Ref. [@WA98-12]). The rms deviations of $S_{Z}$ distributions of the M2 mixed events are found to agree with those of the experimental data within errors for all four centrality classes and for all azimuthal bin sizes. This indicates the absence of event-by-event localized correlated fluctuations in $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$, such as would be expected for DCC-like fluctuations. On the other hand, the rms deviations of the M1 mixed events are found to be systematically lower than those of the data for 2, 4, and 8 bins in $\phi$ for centrality bins 1, 2, and 3. The results for both types of M3 mixed events are found to be intermediate between those of the data and the M1 mixed events. The results indicate the presence of localized fluctuations in the data in both the photon and charged particle multiplicities. For the case of the most peripheral centrality selection (centrality-4), the rms deviations of the $S_{Z}$ distributions of data and the various mixed events are found to be in close agreement to each other within the quoted errors. Multi-resolution DWT analysis results ------------------------------------- The FFC distributions, at scales $j$=1 to 4, corresponding to 4 to 32 bins in azimuthal angle, are shown in Fig. \[ffc\] for data, M1, and VENUS events for the four centrality classes. The results for other types of mixed events are not shown for clarity of presentation. The widths of the FFC distributions are found to increase in going from the most central to most peripheral centrality class. The rms deviations of these FFC distributions are summarized in Fig. \[ffc\_rms\]. Similar to the case for the $S_{Z}$ distributions discussed above, the rms deviations of the mixed events have been rescaled by the percentage difference between the rms deviations of the VENUS FFC distributions and those of the VENUS mixed events for each centrality class. The statistical errors are small and are within the size of the symbols. The bars represent statistical and systematic errors added in quadrature. The rms deviations of the FFC distributions for the data, VENUS, and mixed events are found to be close to each other (within quoted errors) for the case of $32$ bins in $\phi$ for all of the four centrality classes. The rms deviations for the FFC distribution of M2 mixed events are found to closely follow those of the data for all centrality classes and all bins in $\phi$, while the rms deviations for the M3 mixed events lie between those of the data and M1 mixed events. These results are consistent with those obtained from the analysis of the $S_{Z}$ distributions. These observations indicate the absence of event-by-event localized correlated fluctuations (DCC-like) between $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$. They also suggest the presence of localized fluctuations in both photons and charged particle multiplicities for intermediate bin sizes in azimuth. The rms values of the FFC distributions for VENUS events are close to those of the M1 mixed events for centrality classes 1, 2, and 3. However they are slightly higher for the most peripheral centrality class (centrality-4). Discussion ---------- The results from the two independent methods of analysis are consistent and indicate the absence of event-by-event correlated DCC-like fluctuations in the photon and charged particle multiplicities. However they do suggest the presence of uncorrelated fluctuations in both the photon and charged particle multiplicities for intermediate bin sizes in $\phi$. The data has been compared to various kinds of mixed events and to simulated events which take into account many detector related effects. Still it is worthwhile to explore the extent of other possible experimental effects which might affect the observed rms deviations of the $S_Z$ and FFC distributions. As discussed extensively in Refs. [@WA98-3; @pmd_nim; @WA98-9], care has been taken during the data taking and during the data processing to closely monitor the performance of the PMD and SPMD. The detector uniformity of the PMD was studied in detail by using the minimum ionizing particle (MIP) signal from the data for all pads of the 22 boxes. The fluctuations of the pad-to-pad relative gains were approximately 10$\%$. The gain corrections were made for each pad. Corrections for gain variations during the data taking period were made periodically for both the PMD and SPMD. This reduces the possibility of abrupt gain or threshold changes during the run period. Events with obvious detector readout effects, such as missing or dead regions, were carefully removed from the data sample. It should be recalled that most detector effects are reflected in the mixed events. The effect of local fluctuations in the performance of the PMD has been studied using simulated VENUS events. In one test, the gain of a group of pads corresponding to one or more PMD cameras was randomly varied. A 30$\%$ change of gain in one camera for all events resulted in an increase of 0.7$\%$ in the rms deviations. Changing the gains of three cameras by 30$\%$ for all events resulted in an increase of the rms deviations by 1.7$\%$. Because the camera gains were closely monitored on-line and during the processing of the data this is considered to be a highly unlikely scenario. Still, these changes are within the quoted errors of Figs. \[sz\_rms\] and  \[ffc\_rms\] which indicates that local gain and threshold variations would not account for the observed differences between the data and the M1 mixed events. Several checks were performed to verify the quality of the data obtained with the SPMD. One of the differences between the previous analysis [@WA98-3] and the present one is that, while in the previous case the analysis was performed using the total charged particle multiplicity of the detector deduced from the magnitude of the measured SPMD signals, here we simply used the total number of hit SPMD pads. The reason for using hit pads is that the correction in going from deposited charge to hits for each event and small $\eta$–$\phi$ segments is non-trivial. Also, the effects of two-track resolution and possible shifts of the beam position on the target during the spill were studied in detail. However all of these produced small effects which could not account for the differences observed between the data and M1 mixed events. Strength of localized fluctuations ---------------------------------- In order to quantify the strength of the $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$ fluctuations for various bins in $\phi$ and for different centrality classes, we define a quantity $\chi$ as: $$\chi = \frac{\sqrt{(s^2 - s_1^2)}}{s_1} \label{chi_eqn}$$ where $s_1$ and $s$ correspond to the rms deviations of the FFC distributions of the M1 mixed events and real data, respectively. The results are shown in Fig. \[xi\] as a function of the number of bins in $\phi$ for the four different centrality classes. Qualitatively similar results are obtained when $\chi$ is calculated using the rms deviations of the $S_{Z}$ distributions. The shaded portion indicates the region of $\chi$ where $s$ is one $\sigma$ greater than the rms deviation FFC distributions for M1 events, where $\sigma$ is the total error on the M1 event rms deviation. It represents the limit above which a signal is detectable. Since $\chi$ is calculated from the rms deviations of the FFC distributions for data and M1 mixed events it gives the combined strength of localized fluctuations in both the photon and charged particle multiplicities. We do not present $\chi$ values calculated using M3-type mixed events. However, it is clear from the rms deviation figures (Fig. \[sz\_rms\] and \[ffc\_rms\]) that both photons and charged particles contribute to the observed fluctuations. The result shows that the strength of the fluctuations decreases as the number of bins in $\phi$ increases, with a strength which decreases to below detectable level (within the quoted errors) for 16 and 32 bins. There is an indication that the strength of the signal decreases with decreasing centrality for 4 and 8 bins in azimuthal angle, although the tendency is not very strong. Upper limit on DCC production ----------------------------- It has been shown that the rms deviations of the $S_{Z}$ and FFC distributions for data are very close to those of the M2 mixed events, within the quoted errors. If the DCC-like correlated fluctuations in $N_{\gamma-{\mathrm {like}}}$ ${\mathrm {vs.}}$ $N_{\mathrm {ch}}$ were large, the rms deviations (Fig. \[sz\_rms\] and Fig. \[ffc\_rms\]) of data would have been larger compared to those of the M2 mixed events. Since this is not the case, we may extract an upper limit on the production of DCCs at the $90\%$ confidence level following the standard procedure as discussed in Ref. [@uplim]. The errors are assumed to have Gaussian distribution, although they are asymmetric. The larger of the asymmetric errors is conservatively used for the limit calculation. The $90\%$ C.L upper limit contour has been calculated as $\chi$ + 1.28$e_{\chi}$, where $\chi$ is calculated using Eq.(\[chi\_eqn\]). Here $s_1$ and $s$ correspond to the rms deviations of the FFC (or $S_{Z}$) distributions for M2 mixed events and real data, respectively, and $e_{\chi}$ is the error in $\chi$ from the FFC (or $S_{Z}$) analysis. If the difference between the rms deviation of the FFC distributions for M2 mixed events and real data is negative, we take the value of $\chi$ to be zero. It may be mentioned that for the calculation of the upper limits we have assumed that the total difference in the rms values of data and M2 mixed events is due to DCC-like fluctuations only. To relate the measured upper limit on the size of the fluctuations to a limit on DCC domain size and frequency of occurrence we proceed as follows: Within the context of the simple simulated DCC model described earlier, we obtain the rms deviations of the FFC distributions with various domain sizes in azimuthal angle ($15^{0}$, $30^{0}$, $\cdots$ $180^{0}$) and for each domain size also for different frequency of occurrence of DCC ($0\%$ to $100\%$). The M2 mixed events are then constructed for each of these sets of simulated events. For each set of DCC-type events of a given domain size and frequency of occurrence, the value of $\chi$ is calculated using Eq. \[chi\_eqn\], from the difference in rms deviations of the FFC distribution of the DCC event ($s$) and its corresponding M2 mixed event distribution ($s1$). The upper limit is set at that value of frequency of occurrence for a fixed DCC domain size at which the $\chi$ value from the DCC model matches with that of the $\chi$ + 1.28$e_{\chi}$ upper limit from the experimental data. This is used to set the upper limit contour in terms of domain size and frequency of occurrence of the DCC. The results for centrality classes 1 and 2 are shown in Fig. \[upper\_limit\]. It may be mentioned that the upper limit contour set by a similar analysis of the rms values of the $S_{Z}$ distributions is very similar to that from the FFC analysis. Also, from Fig. \[xi\] it is seen that the total fluctuation is similar or weaker for centrality classes 3 and 4, hence the upper limit for these two classes would be similar or weaker than those shown in Fig. \[upper\_limit\] for centrality classes 1 and 2. Summary ======= A detailed event-by-event analysis of the $\eta-\phi$ phase space distributions of the multiplicity of charged particles and photons in Pb+Pb collisions at 158$\cdot A$ GeV has been carried out using two different analysis methods for four different centrality classes. The results from the two analysis methods were found to be consistent with each other. The first analysis method studied the magnitude of the $N_{\gamma-{\mathrm {like}}}$ versus $N_{\mathrm {ch}}$ multiplicity fluctuations in decreasing phase space regions. The second analysis employed the discrete wavelet transformation technique to investigate the relative magnitude of the $N_{\gamma-{\mathrm {like}}}$ versus $N_{\mathrm {ch}}$ fluctuations in adjacent phase space regions. The results were compared to pure VENUS+GEANT simulation events and to various types of mixed events to search for and identify the source of non-statistical fluctuations. Both analysis methods indicated fluctuations beyond those observed in simulated and fully mixed events for $\phi$ intervals of greater than 45$^\circ$ and which increased weakly in strength with increasing centrality. The additional fluctuations were found to be due to uncorrelated fluctuations in both $N_{\gamma-{\mathrm {like}}}$ and $N_{\mathrm {ch}}$. No significant correlated fluctuations in $N_{\gamma-{\mathrm {like}}}$ versus $N_{\mathrm {ch}}$, a likely signature of formation of disoriented chiral condensates, were observed in all of the four centrality classes studied. Using the results from the data, mixed events, and within the limitations of a simple model of DCC formation, an upper limit on DCC production in 158$\cdot A$ GeV Pb+Pb collisions has been set. [99]{} H. Heiselberg, Phys. Rep. [**351**]{}, 161 (2001). M. Stephanov, K. Rajagopal, E. Shuryak, Phys. Rev. [**D60**]{}, 114028 (1999). A.A. Anselm, M.G. Ryskin, Phys. Lett. [**B266**]{}, 482 (1991). J.D. Bjorken, Int. J. Mod. Phys. [**A7**]{}, 4189 (1992). K. Rajagopal and F. Wilczek, Nucl. Phys. [**B399**]{}, 395 (1993); [**B404**]{}, 577 (1993). Zheng Huang and Xin-Nian Wang, Phys. Rev. [**D49**]{}, 4335 (1994). K. Rajagopal, e-print : hep-ph/9504310, Sean Gavin, Andreas Gocksch and Robert D. Pisarski, Phys. Rev. Lett. [**72**]{}, 2143 (1994). M. Asakawa, Z. Huang and X-N Wang, Phys. Rev. Lett. [**74**]{}, 3126 (1995). James V. Steele and Volker Koch, Phys. Rev. Lett. [**81**]{}, 4096 (1998). Sean Gavin and Joseph I. Kapusta, Phys. Rev. [**C65**]{}, 054910 (2002). Proceedings of VIII International Symposium on Very High Energy Cosmic Ray Interactions (Tokyo, Japan), 24-30 July 1994; C.M.G. Lattes, Y. Fujimoto, and S. Hasegawa, Phys. Rep. [**65**]{}, 151 (1980). C.R.A. Augusto et al., Phys. Rev. [**D59**]{}, 054001 (1999). T.C. Brooks et al., (Minimax Collaboration), Phys. Rev. [**D55**]{}, 5667 (1997). M.M. Aggarwal et al., (WA98 Collaboration), Phys. Lett. [**B420**]{}, 169 (1998). H. Appelshauser et al., (NA49 Collaboration), Phys. Lett. [**B459**]{}, 679 (1999). M.M. Aggarwal et al., (WA98 Collaboration) Phys. Rev. [**C64**]{}, 011901(R) (2001). M. Asakawa, H. Minakata and B. Muller, Nucl. Phys. [**A 638**]{}, 443c (1998). H.H. Gutbrod et al., [*Proposal for a Large Acceptance Hadron and Photon Spectrometer*]{}, Preprint CERN/SPSLC 91-17, SPSLC/P260. T.C. Awes et al., Nucl. Instr. Methods [**A279**]{}, 479 (1989). M.M. Aggarwal et al., (WA98 Collaboration) Eur.Phys.J [**C18**]{}, 651 (2001). M.M. Aggarwal et al., Nucl. Instr. and Methods in Phys. Res. [**A424**]{}, 395 (1999). M.M. Aggarwal et al., (WA98 Collaboration), Phys. Lett. [**B458**]{}, 422 (1999). W.T. Lin et al., Nucl. Instr. Methods [**A389**]{}, 415 (1997). M.M. Aggarwal et al., (WA98 Collaboration) Phys. Rev. [**C65**]{}, 054912 (2002). R. Brun et al., GEANT3 user’s guide, CERN/DD/EE/84-1 (1984). K. Werner, Phys. Rep. [**C232**]{}, 87 (1993). Amara Graps, “An Introduction to Wavelets", IEEE Computational Sciences and Engineering, [**2**]{}, 50 (1995). Zheng Huang, Ina Sarcevic, Robert Thews and Xin-Nian Wang, Phys. Rev. [**D54**]{}, 750 (1996). B.K. Nandi, G.C. Mishra, B. Mohanty, D.P. Mahapatra and T.K. Nayak, Phys. Lett. [**B449**]{}, 109 (1999). B.K. Nandi, T.K. Nayak, B. Mohanty, D.P. Mahapatra and Y.P. Viyogi, Phys. Lett. [**B461**]{}, 142 (1999). Numerical Recipes, Cambridge Univ. Press , 1998. B. Mohanty, Ph. D. thesis, Utkal University (2002). G.J. Feldman and R.D. Cousins, Phys. Rev. [**D57**]{}, 3873 (1998).           ![\[rms\] The rms deviations of the FFC distributions for simulated VENUS events containing a variable fraction of localized DCC-like events with DCC extent $\Delta\phi_{DCC}=90^\circ$, as a function of that fraction. Results are also shown for various mixed events constructed from those events. ](rms_plots_mix.eps)                  

ArXiv

--- abstract: 'We present six new planetary nebulae (PNe) discovered in the Small Magellanic Cloud (SMC) from deep UK Schmidt telescope (UKST) narrow band and broad-band short-red “SR" continuum images and confirmed spectroscopically. These 6 preliminary discoveries provide a 6% increase to the previously known SMC PN population of $\sim$100. Once spectroscopic follow-up of all our newly identified candidates is complete, we expect to increase the total number of known SMC PNe by up to 50%. This will permit a significant improvement to determination of the SMC PN luminosity function (PNLF) and enable further insights into the chemical evolution and kinematics of the SMC PN population.' address: | $^1$ Department of Physics and Astronomy, Macquarie University, NSW, Australia\ $^2$ Department of Physics, The University of Hong Kong, Pokfulam Road, Hong Kong\ $^3$ School of Computing, Engineering and Mathematics, Western Sydney University, Australia\ $^4$ Australian Astronomical Observatory (AAO), North Ryde, NSW, Australia author: - 'D. Drašković$^1$, Q. A. Parker$^{2,4}$, W. A. Reid$^{1,3}$ and M. Stupar$^3$' title: Discovery of new planetary nebulae in the Small Magellanic Cloud --- Introduction ============ Planetary nebulae (PNe) play a crucial role in understanding mass loss for low and intermediate mass stars, they influence chemical evolution of galaxies through interstellar enrichment and are a major Galactic dust factory. Their strong emission lines are laboratories for plasma physics and provide accurate radial velocities for Galactic kinematics while the PNLF in external galaxies is a powerful cosmological distance indicator. The SMC is a gas rich, late type dwarf irregular galaxy with an average gas/dust ratio a factor of $\sim30$ below the Galactic value [@SS2000]. The SMC’s accurately determined distance of 61$\pm1$ kpc [@HHH2005] allows derivation of true PNe physical parameters from optical spectra. With a small angular extent of $5^{\circ}20\arcmin\times3^{\circ}05\arcmin$, it can be easily studied in its entirety. SMC reddening and extinction is relatively low and uniform, enabling absolute nebula luminosity estimates. Discovery technique =================== We have carefully examined UKST  and SR SMC image data for $\sim$120 $\deg^2$. We adopted the same powerful and proven colour merging technique successfully applied to the LMC by [@RP2006a][@RP2006b][@RP2013]. Using the <span style="font-variant:small-caps;">KARMA</span> visualisation software package, we surveyed these fields and created merged false-colour images, combining  (coloured in blue) and SR (coloured in red). All images were systematically visual scanned, looking for a faint, compact or barely resolved objects. By carefully choosing the image combination parameters we could balance the intensity of   and SR red matching images, allowing only specific features of one or other pass-band to be seen. In our scheme emission objects such as PNe and  regions appear with a uniform strong blue colour, whereas normal continuum stars are a uniform pink-purple colour. Emission-line stars usually have a strong continuum component and so are easily detected by the narrow extent of their blue halos around a strong pink core [@RP2006a][@RP2006b]. {width="100mm"} Conclusions =========== Confirmatory spectra of six preliminary PNe candidates are quite homogenous. Each was confirmed as an emission-line source with no obvious stellar continuum. This eliminates confusion with emission-line stars for compact sources. Each candidate also gave high  to  ratios ($\sim$ 2 to 7) [@DPRS2015] not observed in  regions, which are the most likely contaminants [@KBFM2000]. The presence of only narrow diagnostic lines also removes the confusion with supernova remnants. These preliminary results demonstrate the strength and promise of our techniques for providing high-quality PNe candidates. We expect to increase SMC PNe by 20 to 50% after spectroscopic follow-up of our candidates and refinement of identification of all known SMC emission-line sources. This will include multi-wavelength analysis and spectroscopic confirmation (or rejection) of objects previously listed as SMC PNe. Our survey will permit more extreme ends of the PNLF to be explored, especially at the faint end, and enable us to study underrepresented evolutionary stages of SMC PNe. This sample will have a major impact on improving rates of elemental enhancement in lower mass stars within a low-metallicity environment. Other advances will come from the improved SMC kinematical data from over 300 emission-line objects uncovered. References {#references .unnumbered} ========== Stanimirovic S, Staveley-Smith L, van der Hulst J M, Bontekoe T R, Kester D J M and Jones P A 2000 [*MNRAS*]{} [**315**]{} 791 Hilditch R W, Howarth I D and Harries T J 2005 [*MNRAS*]{} [**357**]{} 304 Reid W A and Parker Q A 2006 [*MNRAS*]{} [**365**]{} 401 Reid W A and Parker Q A 2006 [*MNRAS*]{} [**373**]{} 521 Reid W A and Parker Q A 2013 [*MNRAS*]{} [**436**]{} 604 Draskovic D, Parker Q A, Reid W A and Stupar M 2015 [*MNRAS*]{} [**452**]{} 1402 Kennicutt Jr R C, Bresolin F, French H and Martin P 2000 [*ApJ*]{} [**537**]{} 589

ArXiv

--- abstract: 'We obtain the Pairwise Summation Approximation (PSA) of the Casimir energy from first principles in the soft dielectric and soft diamagnetic limit, this analysis let us find that the PSA is an asymptotic approximation of the Casimir energy valid for large distances between the objects. We also obtain the PSA for the electromagnetic (EM) coupling part of the Casimir energy, so we are able to complete the PSA limit for the first time for the complete electromagnetic field.' author: - 'P. Rodriguez-Lopez' - Pablo Rodriguez title: - Pairwise Summation Approximation of Casimir energy from first principles - | [Asymptotic electromagnetic Casimir energy]{}\ --- Introduction ============ Since 1948, when Casimir introduced the energy that got his name [@Casimir; @Placas; @Paralelas], calculation formulas have been looked for. Many analytical and numerical methods have been proposed, such as the zeta function technique, the heat kernel method, semiclassical methods or Green function (local) methods just to mention a few of them [@Review; @Casimir]. However, exact results have been obtained only for some simple geometries. The asymptotic analysis of Casimir energies has also a long history. In fact, Casimir himself and Polder in year 1948 gave the first asymptotic formula for the Casimir energy between two electrically neutral bodies in terms of their electric induced dipoles [@VdW; @int.; @electrica]. Some time later, a generalization of that formula, known as the Pairwise Summation Approximation (PSA) was derived [@Lifshitz] for electric media. The main assumption is a linear superposition of the Casimir - Polder interactions between the induced polarizabilities of each element of volume body. Then, the PSA energy is expressed as an integral over the two object’s volumes and it is proportional to the objects polarizabilities. The formula has been recently reobtained by R. Golestanian in [@Golestanian] and by K. A. Milton et. al. in [@Milton] in the soft dielectric limit. Besides, a new asymptotic method for calculating Casimir energies in term of the induced multipoles of the interacting bodies has been proposed in [@Kardar-Geometrias-Arbitrarias] and [@Functional; @Determinant; @Method; @2]. This formula provides a procedure for the calculation of the Casimir energy between *N* arbitrary shaped compact objects [@Rodriguez-Lopez; @1]. Our goal is the asymptotic calculation of the electromagnetic Casimir energy and a presentation of a systematic asymptotic expansion procedure for the integral dipole formula on higher orders. We will reobtain the classical results of Casimir and Polder [@VdW; @int.; @electrica], and of Feinberg-Sucher in [@VdW; @int.; @magnetica] in the rarified or soft dielectric and soft diamagnetic limit. The result we obtain here is a generalization of Milton et. al. one [@Milton] but assuming that the bodies are soft diamagnetic as well as soft dielectric. For this purpose, we will use the multi-scattering expansion of the Casimir energy formalism given in [@Kardar-Geometrias-Arbitrarias]. To our knownledge, it is the first time this formalism is used to derive the PSA. We obtain that the PSA is the first order of a perturbation expansion in the difference between the electric and magnetic permeability constants of the objects respect the electric and magnetic permeability constants of the medium were the objects are placed. The interest of this result is that we obtain this pairwise summation formula for the complete electromagnetic Casimir energy which is derived as an asymptotic limit of an exact and free of divergences formula. That means that now we can establish the range of validity for that approximation. In fact, we will justify why this formula is valid in the far distance objects limit. This derivation also gives us a perturbative procedure for corrections of this approximation and posterior expansions to more than two objects or finite temperature cases. We will follow this plan for the article: Using the soft dielectric and soft diamagnetic approximation, in Sect. \[sec: 2\] we will obtain the PSA of Casimir energy for the zero temperature case starting from the exact Casimir energy formula given in [@Kardar-Geometrias-Arbitrarias]. We will also study the far distance limit to reobtain the asymptotic Casimir energies given in [@VdW; @int.; @electrica] and [@VdW; @int.; @magnetica]. In Sect. \[sec: 3\] we will obtain PSA Casimir energy formulas for any temperature, studying the high and low temperature limits. In Sect. \[sec: 4\] we will study the PSA Casimir energy for the three bodies system and for the general $N$ bodies system. We will obtain a kind of superposition principle of Casimir energy in the PSA approximation. Finally, in Sect. \[sec: 5\] we will study the first perturbation energy term to the PSA approximation and we will discuss about the nature of the PSA limit of the Casimir energy. Diluted limit at zero temperature {#sec: 2} ================================= Our objective is the calculation of the complete electromagnetic Casimir energy between two bodies in the soft dielectric limit. For this purpose, we will use the Casimir energy formula between two compact bodies at temperature $T = 0$, given in [@Kardar-Geometrias-Arbitrarias] as $$\label{Formula de Emig} E = \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\log\left({\vert \mathbb{I} - \mathbb{N}\vert}\right).$$ Where $k$ is the frequency, $\mathbb{I}$ is the identity matrix, $\mathbb{N}$ is the matrix $\mathbb{N} = \mathbb{T}_{1}\mathbb{U}_{12}\mathbb{T}_{2}\mathbb{U}_{21}$. Here, $\mathbb{T}_{\alpha}$ is the T scattering matrix of the $\alpha$ - body under the electromagnetic field, and $\mathbb{U}_{\alpha\beta}$ is the propagation matrix of the electromagnetic field from $\alpha$ - body to $\beta$ - body. We will use the position representation instead the multipole representation used in [@Kardar-Geometrias-Arbitrarias], so we can identify $\mathbb{U}_{\alpha\beta} = G_{0\alpha\beta}$, where $G_{0\alpha\beta}$ is the free dyadic Green function. Taking into account that $\log({\vert A\vert}) =\textrm{Tr}(\log(A))$ and that $\log(1 - x) = - \sum_{p=1}^{\infty}\frac{x^{p}}{p}$, we transform [Eq. ]{} into $$\label{Formula de Emig en forma de traza} E = - \frac{\hbar c}{2\pi}\sum_{p=1}^{\infty}\frac{1}{p}\int_{0}^{\infty}dk\textrm{Tr}\left(\mathbb{N}^{p}\right).$$ Equation is an asymptotic expansion of [Eq. ]{} in the $p_{max}<\infty$ case, which means that our calculus will be valid in the large bodies distance limit. The T operator is related with the potential $V$ (we will see what $V$ is later) by the Lippmann - Schwinger equation, that can be written as [@Galindo; @y; @Pascual] $$\label{Lippmann - Schwinger equation for T operator} \mathbb{T} = \left(1 - VG_{0}\right)^{-1}V.$$ Applying a Born expansion to [Eq. ]{}, we can obtain an approximation of the $T$ operator in the soft dielectric and soft diamagnetic limit as $$\label{Born expansion of T operator} \mathbb{T} = \sum_{n=0}^{\infty}(VG_{0})^{n}V\simeq V.$$ This approximation is more valid for weaker $V$, so here is where the soft dielectric and soft diamagnetic limit is applied. Now we will study the lowest expansion order. In the lowest expansion order ($p=1$ and $\mathbb{T}=V$) we get the asymptotic approximation of the Casimir energy between two bodies as $$\label{Formula de Emig en soft limit} E = - \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\textrm{Tr}\left(V_{1}G_{012}V_{2}G_{021}\right).$$ We can separate the magnetic and electric part of each body potential (which is a diagonal operator in positions space, because it is local) and each field contribution in the free dyadic Green function. For isotropic dielectrics with constant electric and magnetic permeabilities, it is easy to find that $V_{\alpha}^{E}(\textbf{r}) = (\epsilon_{\alpha} - \epsilon_{0})\chi_{\alpha}(\textbf{r}) = \tilde{\epsilon}_{\alpha}\chi_{\alpha}(\textbf{r})$ and $V_{\alpha}^{H}(\textbf{r}) = (\mu_{\alpha} - \mu_{0})\chi_{\alpha}(\textbf{r}) = \tilde{\mu}_{\alpha}\chi_{\alpha}(\textbf{r})$ in the vacuum, where $\chi_{\alpha}(\textbf{r})$ is the characteristic function of the $\alpha$ - body volume (1 inside the body and 0 in the rest of the space). Note that we represent the electromagnetic properties of the objects by their potential energy instead the complement representation by their boundary conditions [@Balian; @and; @Duplantier; @II]. It will let us to obtain the PSA as an integral over the volume of the bodies as required by the PSA. The contributions of the free dyadic Green function are obtained in the Appendix in [Eq. ]{} - . As seen in [Eq. ]{} - of the Appendix, operators are defined over three different linear spaces: 1) An $EH$ - space, whose components are the electric and the magnetic field; 2) over the space coordinates, because we are working with a vector and a pseudovector; and 3) over positions. We must solve the trace over these three spaces: $EH$ - space, vector coordinate space and position space. First we solve the trace in the $EH$ - space and we obtain: $$\begin{aligned} E & = & - \frac{\hbar c}{2\pi}\int_{0}^{\infty}dkTr(V_{1}^{E}G_{012}^{EE}V_{2}^{E}G_{021}^{EE} \nonumber\\ & & + V_{1}^{E}G_{012}^{EH}V_{2}^{H}G_{021}^{HE} + V_{1}^{H}G_{012}^{HE}V_{2}^{E}G_{021}^{EH} \nonumber\\ & & + V_{1}^{H}G_{012}^{HH}V_{2}^{H}G_{021}^{HH}).\end{aligned}$$ We identify each term with this obvious notation: $$\label{Energia de Milton} E = E_{EE} + E_{EH} + E_{HE} + E_{HH}.$$ Purely electric and purely magnetic Casimir energy -------------------------------------------------- The purely electric case has been solved by K. A. Milton et. al. in [@Derivacion; @caso; @ee; @por; @Milton], where they give the following result: $$E_{EE} = -23\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}\frac{\hbar c}{(4\pi)^{3}}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{{\vert \textbf{r}_{1} - \textbf{r}_{2}\vert}^{7}}.$$ We obtain the same result because the dyadic Green function used in [@Derivacion; @caso; @ee; @por; @Milton] is the pure electric part of the Green function matrix we use here. In fact, the pure magnetic part of the Green function matrix is equal to the pure electric part $G_{0ij}^{HH}(R,k) = G_{0ij}^{EE}(R,k)$, so the calculus of the purely magnetic part of the Casimir energy is similar to the electric one. Consequently we obtain the following result for the purely magnetic part of the Casimir energy: $$E_{HH} = -23\tilde{\mu}_{1}\tilde{\mu}_{2}\frac{\hbar c}{(4\pi)^{3}}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{{\vert \textbf{r}_{1} - \textbf{r}_{2}\vert}^{7}}.$$ To our knowledge, this is the first place where this result is obtained. Coupled electromagnetic Casimir energy -------------------------------------- Here we are going to calculate the contribution of the electromagnetic coupling part of the Casimir energy. Using [Eq. ]{} and [Eq. ]{}, we have to solve: $$E_{EH} = - \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\textrm{Tr}\left(V_{1}^{E}G_{012}^{EH}V_{2}^{H}G_{021}^{HE}\right).$$ Replacing each potential by its value and assuming isotropy, it is followed that they are proportional to the identity matrix. Then we can drop them for the coordinates trace, but not for the spatial positions trace: $$\begin{aligned} E_{EH} & = & - \tilde{\epsilon}_{1}\tilde{\mu}_{2}\frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\int d\textbf{r}_{1}\int d\textbf{r}_{2}\chi_{1}\chi_{2}\nonumber\\ & & \times \textrm{Tr}\left(G_{012}^{EH}G_{021}^{HE}\right).\end{aligned}$$ Using [Eq. ]{} of the Appendix and $R = {\vert \textbf{r} - \textbf{r}'\vert}$ we have, in matrix form: $$\begin{aligned} G_{0ij}^{EH}(R,k) & = & - k\frac{\partial G_{0}}{\partial R}\epsilon_{ijk}\partial_{k}R \nonumber\\ & = & G_{0}(R,k)\frac{k}{R}\left(k + \frac{1}{R}\right)\nonumber\\ & & \times\left( \begin{array}{ccc} 0 & R_{3} & - R_{2}\\ - R_{3} & 0 & R_{1}\\ R_{2} & - R_{1} & 0 \end{array} \right),\end{aligned}$$ so the trace in spatial coordinates is easily solved: $$\label{Traza terminos de acoplo em} \textrm{Tr}\left(G_{012}^{EH}G_{021}^{HE}\right) = - 2k^{2}G_{0}^{2}(R,k)\left(k + \frac{1}{R}\right)^{2}.$$ We can simplify our calculus using the dimensionless variable $u = kR$ to factorize spatial and frequency contributions of the formula $$\label{Energia Casimir acoplo em a T = 0} E_{EH} = \tilde{\epsilon}_{1}\tilde{\mu}_{2}\frac{4\hbar c}{(4\pi)^{3}}\alpha\int d\textbf{r}_{1}\int d\textbf{r}_{2}\frac{\chi_{1}\chi_{2}}{R^{7}},$$ where $$\alpha = \int_{0}^{\infty}due^{-2u}\left(u^{4} + 2u^{3} + u^{2}\right).$$ Finally, using $\int_{0}^{\infty}duu^{n}e^{-au} = \frac{n!}{a^{n+1}}$, the final result is obtained as $$E_{EH} = 7\tilde{\epsilon}_{1}\tilde{\mu}_{2}\frac{\hbar c}{(4\pi)^{3}}\int d\textbf{r}_{1}\int d\textbf{r}_{2}\frac{\chi_{1}\chi_{2}}{R^{7}}.$$ There is an antisymmetry between $G^{EH}$ and $G^{HE}$ shown in [Eq. ]{} and [Eq. ]{} of the Appendix, that is $G_{0ij}^{EH}(R,k) = - G_{0ij}^{HE}(R,k)$. Therefore, we can obtain the coupling between the magnetic part of the first object and the electric part of the second one in a similar way. Then, we can obtain the coupling between the magnetic part of the first object and the electric part of the second one: $$E_{HE} = 7\tilde{\epsilon}_{2}\tilde{\mu}_{1}\frac{\hbar c}{(4\pi)^{3}}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{{\vert \textbf{r}_{1} - \textbf{r}_{2}\vert}^{7}}.$$ These two cross contributions to the Casimir energy in the soft dielectric and diamagnetic limits are the main result of this article. It is interesting to note the sign change in this part of the Casimir energy with respect to the other contributions. Then, these terms can invert the typical attractive nature of the Casimir energy to repulsive in this limit for objects with very different electromagnetic nature. Finally, thanks to these new two terms, we obtain the complete electromagnetic Casimir energy between two objects in the soft dielectric and diamagnetic limit for the first time. Then the global asymptotic electromagnetic Casimir energy is finally obtained: $$\label{Final Result} E = \frac{- \hbar c}{(4\pi)^{3}} \gamma \int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{{\vert \textbf{r}_{1} - \textbf{r}_{2}\vert}^{7}},$$ where $\gamma = 23\tilde{\epsilon}_{1}\tilde{\epsilon}_{2} - 7\tilde{\epsilon}_{1}\tilde{\mu}_{2} - 7\tilde{\epsilon}_{2}\tilde{\mu}_{1} + 23\tilde{\mu}_{1}\tilde{\mu}_{2}$ is the complete multiplicative constant, taking into account all the electromagnetic effects in this limit, nor just electric effects, where $\gamma$ would be $23\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}$. Asymptotic Casimir energy {#sec: 3.C} ------------------------- We can also obtain from [Eq. ]{} (always in the soft electric and magnetic limit) the first asymptotic energy order for two objects. These formulas are the first order of a multipolar expansion of the integrand in [Eq. ]{} in the coordinate system of each object. That means that we have to assume that the distance between the objects is much greater than their characteristic lengths $R_{\alpha}$, that is $R_{\alpha}\ll R$. In this limit we can separate the problem into two scales and we can replace ${\vert \textbf{r}_{1} - \textbf{r}_{2}\vert} = R$, where $R$ is assume to be a constant. Then the integral can be easily solved in this limit to: $$E \simeq \frac{- \hbar c}{(4\pi)^{3}}\gamma\frac{V_{1}V_{2}}{R^{7}},$$ where $V_{\alpha}$ is the volume of the $\alpha$ - object. In the soft limit order we can approximate the electric and magnetic polarizabilities as $\alpha^{E} = \tilde{\epsilon}\frac{V}{4\pi}$ and $\alpha^{H} = \tilde{\mu}\frac{V}{4\pi}$. In [@Optica; @de; @Wolf] it was derived $\alpha^{E}$ for an sphere. By using the same method with the approximation of that the effective field over the dielectric is equal to the induced field in the soft limit, we arrive at $\alpha^{E} = \tilde{\epsilon}\frac{V}{4\pi}$ for any arbitrary shaped object. Equation simplifies into $$E \simeq\frac{- \hbar c}{4\pi R^{7}}\left(23\alpha_{1}^{E}\alpha_{2}^{E} + 23\alpha_{1}^{H}\alpha_{2}^{H} - 7\alpha_{1}^{E}\alpha_{2}^{H} - 7\alpha_{1}^{H}\alpha_{2}^{E}\right).$$ This is the soft response limit of the Feinberg and Sucher potential [@VdW; @int.; @magnetica]. It coincide as well with the soft response limit of the asymptotic Casimir energy obtained in [@Kardar-Geometrias-Arbitrarias] for spheres. The presented scheme looks as we would use a local two points potential (where we substitute the polarizabilities of the bodies by their local susceptibilities) as in the present case with the potential given in [Eq. ]{}, and integrate over the volume of each body to obtain the PSA of the Casimir energy. If we also study the asymptotic distance limit of the PSA, then we reobtain the Feinberg and Sucher potential in the soft response limit, now proportional to the polarizabilities of the objects instead the susceptibilities. Diluted limit at any temperature $T$ {#sec: 3} ==================================== In this section we will calculate the Casimir energy in the diluted limit at any finite temperature. Then we will focus on different approximations to low an high temperature limits, recovering the zero temperature case showed before and giving new formulas of the PSA for any temperature. We begin this study with the Casimir energy formula for any temperature: $$\label{Formula de Emig T} E = k_{B}T{\sum_{n = 0}^{\infty}}'\log\left({\vert \mathbb{I} - \mathbb{N}(k_{n})\vert}\right),$$ with Matsubara frequencies $k_{n} = 2\pi\frac{k_{B}T}{\hbar c}n = \Lambda n$. The tilde means that the $n=0$ case is weighted by a $1/2$ factor. As usual, we apply the Born approximation to the $T$ matrix scattering obtaining at first order $\mathbb{T}_{\alpha} = V_{\alpha}$ (where $\alpha$ labels the body in interaction). We apply again that $\log({\vert A\vert}) = Tr(\log(A))$ and $\log(1 - x) = - \sum_{p=1}^{\infty}\frac{x^{p}}{p}$ , then we transform [Eq. ]{} into: $$E = - k_{B}T\sum_{p=1}^{\infty}\frac{1}{p}{\sum_{n = 0}^{\infty}}'\textrm{Tr}\left(\mathbb{N}^{p}(\Lambda n)\right).$$ At first order in $p$ we obtain: $$\begin{aligned} E & = & - k_{B}T{\sum_{n = 0}^{\infty}}'\textrm{Tr}\left(\mathbb{N}(\Lambda n)\right)\nonumber\\ & = & - k_{B}T{\sum_{n = 0}^{\infty}}'\textrm{Tr}\left(\mathbb{T}_{1}\mathbb{U}_{12}\mathbb{T}_{2}\mathbb{U}_{21}\right).\end{aligned}$$ As we are working in the positions representation space instead in the multipolar representation space, we have to represents the operators $\mathbb{T}_{\alpha}$ and $\mathbb{U}_{\alpha\beta}$ in the positions space. For that issue, we have into account that $\mathbb{U}_{\alpha\beta}$ matrices represent the field propagation between two points. Then they are represented by the free vacuum Green function of the interaction field. Using the Born approximation, we obtain the next formula for the diluted approximation of the Casimir energy: $$E \simeq E_{T} = - k_{B}T{\sum_{n = 0}^{\infty}}'\textrm{Tr}\left(V_{1}G_{012}V_{2}G_{021}\right).$$ As we did with , we can separate the magnetic and electric part of each body potential and each field contribution in the free dyadic Green function. We trace over the $EH$ space obtaining: $$\begin{aligned} E_{T} & = & - k_{B}T{\sum_{n = 0}^{\infty}}'Tr(V_{1}^{E}G_{012}^{EE}V_{2}^{E}G_{021}^{EE} + \nonumber\\ & & + V_{1}^{E}G_{012}^{EH}V_{2}^{H}G_{021}^{HE} + V_{1}^{H}G_{012}^{HE}V_{2}^{E}G_{021}^{EH}\nonumber\\ & & + V_{1}^{H}G_{012}^{HH}V_{2}^{H}G_{021}^{HH}).\end{aligned}$$ As before, we identify each term with this obvious notation: $$E_{T} = E_{EE} + E_{EH} + E_{HE} + E_{HH}.$$ Purely electric and purely magnetic energy ------------------------------------------ Using [Eq. ]{} and [Eq. ]{} of the Appendix, we can solve the trace over the coordinates of $E_{EE}$ and, similarly, of $E_{HH}$. The matricial form of the purely electric part of the dyadic Green function \[Acoplo EE funcion Green\] and of the purely magnetic part of the dyadic Green function\[Acoplo HH funcion Green\] are: $$\begin{aligned} G_{0ij}^{EE}(R,k) & = & - \left( \begin{array}{ccc} R_{x}^{2} & R_{x}R_{y} & R_{x}R_{z}\\ R_{y}R_{x} & R_{y}^{2} & R_{y}R_{z}\\ R_{z}R_{x} & R_{z}R_{y} & R_{z}^{2} \end{array} \right)\nonumber\\ & & \times\left(3 + 3kR + k^{2}R^{2}\right)\frac{e^{-kR}}{4\pi R^{5}}\nonumber\\ & & + \left( \begin{array}{ccc} 1 & 0 & 0\\ 0 & 1 & 0\\ 0 & 0 & 1 \end{array} \right)\nonumber\\ & & \times\left(1 + kR + k^{2}R^{2}\right)\frac{e^{-kR}}{4\pi R^{3}}\nonumber\\ & = & G_{0ij}^{HH}(R,k).\end{aligned}$$ Where $\textbf{R} = \textbf{r}_{\alpha} - \textbf{r}_{\beta}$. The trace in spatial coordinates is easily solved obtaining: $$\begin{aligned} E_{EE} & = & - k_{B}T{\sum_{n = 0}^{\infty}}'\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}\int_{1}d\textbf{r}_{1}\int_{2}d\textbf{r}_{2}\frac{e^{-2kR}}{(4\pi)^{2} R^{6}}\nonumber\\ & & \times[6 + 12kR + 10k^{2}R^{2} + 4k^{3}R^{3} + 2k^{4}R^{4}].\nonumber\\ & & \label{Caso diagonal T neq 0}\end{aligned}$$ Replacing $k$ by $\Lambda n$ we obtain: $$\begin{aligned} E_{EE} & = & - \frac{k_{B}T}{(4\pi)^{2}}\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R^{6}}{\sum_{n = 0}^{\infty}}'e^{-2\Lambda Rn}\nonumber\\ & & \times[6 + 12\Lambda Rn + 10\Lambda^{2}R^{2}n^{2}\nonumber\\ & & + 4\Lambda^{3}R^{3}n^{3} + 2\Lambda^{4}R^{4}n^{4} ]\label{Caso diagonal T neq 0 2}.\end{aligned}$$ Having into account that $\sum_{n = 0}^{\infty}e^{-an} = \frac{1}{1 - e^{-a}}$, it is easily deduced: $$\partial_{a}^{t}\sum_{n = 0}^{\infty}e^{-an} = \sum_{n = 0}^{\infty}(-n)^{t}e^{-an} = \partial_{a}^{t}(1 - e^{-a})^{-1}.$$ Denoting $\lambda = R\Lambda$, we can carry out the sum obtaining: $$\begin{aligned} \label{Caso diagonal T neq 0 resuelto} E_{EE} & = & - \frac{k_{B}T}{(4\pi)^{2}}\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R^{6}}\frac{1}{(e^{2\lambda} - 1)^{5}}\nonumber\\ & & \times[e^{2\lambda}\left( - 9 - 12\lambda + 10\lambda^2 - 4\lambda^3 + 2\lambda^4 \right)\nonumber\\ & & + e^{4\lambda}\left( 6 + 36\lambda - 10\lambda^2 - 12\lambda^3 + 22\lambda^4 \right)\nonumber\\ & & + e^{6\lambda}\left( 6 - 36\lambda - 10\lambda^2 + 12\lambda^3 + 22\lambda^4 \right)\nonumber\\ & & + e^{8\lambda}\left(- 9 + 12\lambda + 10\lambda^2 + 4\lambda^3 + 2\lambda^4 \right)\nonumber\\ & & + 3 e^{10\lambda} + 3].\end{aligned}$$ We obtain a similar result for the purely magnetic Casimir energy replacing $\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}$ by $\tilde{\mu}_{1}\tilde{\mu}_{2}$. Coupled magnetic - electric Casimir energy terms ------------------------------------------------ We perform the same calculations using and of the Appendix as for the purely electric case. The trace over spatial coordinates is already done in , so we obtain the formula for the coupling terms of the Casimir energy just replacing the integral in by a sum: $$\begin{aligned} E_{EH} & = & \frac{2k_{B}T}{(4\pi)^{2}}\tilde{\epsilon}_{1}\tilde{\mu}_{2}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R^{6}}{\sum_{n = 0}^{\infty}}'e^{-2kR}\nonumber\\ & & \times\left(R^{4}k^{4} + 2R^{3}k^{3} + R^{2}k^{2}\right)\label{Energia Casimir acoplo em a T neq 0}.\end{aligned}$$ Here $k_{n} = 2\pi\frac{k_{B}T}{\hbar c}n = \Lambda n$ are the Matsubara frequencies. After solving this sum and denoting $\lambda = R\Lambda$, we obtain: $$\begin{aligned} \label{Energia Casimir acoplo em a T neq 0 resuelto} E_{EH} & = & \frac{2k_{B}T}{(4\pi)^{2}}\tilde{\epsilon}_{1}\tilde{\mu}_{2}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R^{6}}\frac{\lambda^2}{(e^{2\lambda} - 1)^{5}}\nonumber\\ & & \times[e^{2\lambda}\left( 1 - 2\lambda +\lambda^2 \right)\nonumber\\ & & + e^{4\lambda}\left( -1 - 6\lambda + 11\lambda^2\right)\nonumber\\ & & + e^{6\lambda}\left( -1 + 6\lambda + 11\lambda^2\right)\nonumber\\ & & + e^{8\lambda}\left( 1 + 2\lambda +\lambda^2 \right) ].\end{aligned}$$ Replacing $\tilde{\epsilon}_{1}\tilde{\mu}_{2}$ by $\tilde{\mu}_{1}\tilde{\epsilon}_{2}$, we obtain the formula for $E_{HE}$. These formulas for the Casimir energy in the diluted limit are valid for any temperature, but they are too much complicate for analytical analysis at any temperature. Therefore, we study the limits at high and low temperatures. Low and zero temperature limit ------------------------------ Here we will recover the zero temperature limit of the diluted limit. Then we will make a perturbative analysis valid for low temperatures. We can make it easily by expanding the Taylor series in $\lambda = R\Lambda$ of the integrand and studying just the first orders deleting the rest ones. By using and , we reobtain for the zero temperature case. We need to take the fifth order series term of the Taylor expansion of [Eq. ]{} and [Eq. ]{} to get the next non zero perturbation term of the Casimir energy as $$\frac{\Delta_{5}E_{EE}}{\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}} = \frac{\Delta_{5}E_{HH}}{\tilde{\mu}_{1}\tilde{\mu}_{2}} = - \frac{22\pi^{3}}{945}k_{B}T\left(\frac{k_{B}T}{\hbar c}\right)^{5}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R},$$ $$\frac{\Delta_{5}E_{EH}}{\tilde{\epsilon}_{1}\tilde{\mu}_{2}} = \frac{\Delta_{5}E_{HE}}{\tilde{\mu}_{1}\tilde{\epsilon}_{2}} = - \frac{2\pi^{3}}{189}k_{B}T \left(\frac{k_{B}T}{\hbar c}\right)^{5}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R},$$ because $\Delta_{1}E = \Delta_{2}E = \Delta_{3}E = \Delta_{4}E = 0 $, where $\Delta_{n}E$ is the n-th order correction term to the low temperature expansion. High temperature and classical limit ------------------------------------ In this section we will obtain the high temperature limit of the Casimir energy in the diluted limit, whose first term will be the classical limit of the Casimir energy. To obtain this limit, instead of solving the sum in [Eq. ]{} and [Eq. ]{}, we will just keep the first term of the sum. Then the classical limit of the Casimir energy in the diluted limit reads the first sum term: $$\frac{E_{EE}^{cl}}{\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}} = \frac{E_{HH}^{cl}}{\tilde{\mu}_{1}\tilde{\mu}_{2}} = - 3 \frac{k_{B}T}{(4\pi)^{2}}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R^{6}},$$ $$\frac{E_{EH}^{cl}}{\tilde{\epsilon}_{1}\tilde{\mu}_{2}} = \frac{E_{HE}^{cl}}{\tilde{\mu}_{1}\tilde{\epsilon}_{2}} = 0.$$ And the first perturbation to that limit is the next sum term: $$\begin{aligned} \frac{\Delta_{1}E_{EE}^{cl}}{\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}} & = & - \frac{k_{B}T}{(4\pi)^{2}}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R^{6}}e^{-2\Lambda R}\nonumber\\ & & \times\left(6 + 12\Lambda R + 10\Lambda^{2}R^{2} + 4\Lambda^{3}R^{3} + 2\Lambda^{4}R^{4}\right)\nonumber\\ & = &\frac{\Delta_{1}E_{HH}^{cl}}{\tilde{\mu}_{1}\tilde{\mu}_{2}},\end{aligned}$$ $$\begin{aligned} \frac{\Delta_{1}E_{EH}^{cl}}{\tilde{\epsilon}_{1}\tilde{\mu}_{2}} & = & \frac{\Delta_{1}E_{HE}^{cl}}{\tilde{\mu}_{1}\tilde{\epsilon}_{2}} = \frac{2k_{B}T}{(4\pi)^{2}}\int_{1}\int_{2}\frac{d\textbf{r}_{1}d\textbf{r}_{2}}{R^{6}}e^{-2R\Lambda}\nonumber\\ & & \times\left(R^{4}\Lambda^{4} + 2R^{3}\Lambda^{3} + R^{2}\Lambda^{2}\right).\end{aligned}$$ Where $\Delta_{1}E^{cl}$ is the first correction to the classic limit of the Casimir energy $E^{cl}$. With these results we see that we loose the coupling between electric and magnetic energy terms in the classical limit. Finally we must remark that, if we take the asymptotic distance approximation as in Sect. \[sec: 3.C\], we reobtain the results given in [@Barton] in the soft response limit. PSA for three bodies system {#sec: 4} =========================== In this section we are going to calculate the Casimir energy between three bodies in the Pairwise Summation approximation. In this asymptotic limit we will obtain that the energy of the system will be the addition of the PSA energy of each pair of objects. This linear behavior of the Casimir energy was expected in that approximation although we know that it is in general false. We begin this study from the Casimir Energy formula for three objects given in [@Emig; @Casimir; @caso; @escalar]: $$\label{Casimir N objects} E = \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\log\left(\frac{{\vert \mathbb{M}\vert}}{\phantom{_{\infty}}{\vert \mathbb{M}\vert}_{\infty}}\right).$$ The $M$ matrix (whose coefficients are non commutative matrices) is: $$\mathbb{M} = \left(\begin{array}{c c c} \phantom{-}\mathbb{T}_{1}^{-1} & - \mathbb{U}_{12} & - \mathbb{U}_{13}\\ - \mathbb{U}_{21} & \phantom{-}\mathbb{T}_{2}^{-1} & - \mathbb{U}_{23}\\ - \mathbb{U}_{31} & - \mathbb{U}_{32} & \phantom{-}\mathbb{T}_{3}^{-1} \end{array}\right).$$ So, using the logarithm product rule, the Casimir energy between three bodies is $$\begin{aligned} \label{Energia Casimir 3 cuerpos} E_{3} & = & \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\log\left({\vert \mathbb{I} - \mathbb{N}_{12}\vert}\right) \nonumber\\ & & + \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\log\left({\vert \mathbb{I} - \mathbb{N}_{13}\vert}\right) \nonumber\\ & & + \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\log\left({\vert \mathbb{I} - \mathbb{R}\vert}\right).\end{aligned}$$ Here the $N$ matrix is $\mathbb{N}_{\alpha\beta} = \mathbb{T}_{\alpha}\mathbb{U}_{\alpha\beta}\mathbb{T}_{\beta}\mathbb{U}_{\beta\alpha}$, and: $$\begin{aligned} \mathbb{R} & = & \left(\mathbb{I} - \mathbb{N}_{13}\right)^{-1}\left(\mathbb{T}_{3}\mathbb{U}_{32} + \mathbb{N}_{31}\right)\nonumber\\ & & \times\left(\mathbb{I} - \mathbb{N}_{12}\right)^{-1}\left(\mathbb{T}_{2}\mathbb{U}_{23} + \mathbb{N}_{21}\right).\end{aligned}$$ Expanding the inverses we obtain the double series $$\begin{aligned} \mathbb{R} & = & \sum_{n_{1}=0}^{\infty}\left(\mathbb{N}_{13}\right)^{n_{1}}\left(\mathbb{T}_{3}\mathbb{U}_{32} + \mathbb{N}_{31}\right)\nonumber\\ & & \times\sum_{n_{2}=0}^{\infty}\left(\mathbb{N}_{12}\right)^{n_{2}}\left(\mathbb{T}_{2}\mathbb{U}_{23} + \mathbb{N}_{21}\right).\end{aligned}$$ The lowest order of the Born series of the $\mathbb{R}$ matrix will come from the first order expansion series making $n_{1} = n_{2} = 0$, that is $$\mathbb{R} \simeq \left(\mathbb{T}_{3}\mathbb{U}_{32} + \mathbb{N}_{31}\right)\left(\mathbb{T}_{2}\mathbb{U}_{23} + \mathbb{N}_{21}\right).$$ In addition to that, we just consider the sum term with the minimum number of $\mathbb{T}$ matrices products, because that will be the lowest order expansion in susceptibilities of the $\mathbb{R}$ matrix. That means that we reduce the highly non linear $\mathbb{R}$ matrix to $$\label{sol limit R matrix} \mathbb{R} \simeq \left(\mathbb{T}_{3}\mathbb{U}_{32}\mathbb{T}_{2}\mathbb{U}_{23}\right) = \mathbb{N}_{23}.$$ Replacing [Eq. ]{} in [Eq. ]{}, we obtain the next PSA of the Casimir energy between three objects: $$\begin{aligned} \label{3-bodies PSA} E_{3} & = & \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\log\left({\vert \mathbb{I} - \mathbb{N}_{12}\vert}\right) \nonumber\\ & & + \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\log\left({\vert \mathbb{I} - \mathbb{N}_{13}\vert}\right) \nonumber\\ & & + \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\log\left({\vert \mathbb{I} - \mathbb{N}_{23}\vert}\right).\end{aligned}$$ Therefore we obtain that the PSA approximation of the Casimir energy between three objects is the sum of the PSA energy of each pair of objects. In other words, we rederive a kind of superposition law of energies in the diluted PSA limit as expected, because the usual presentation of the PSA approximation is the assumption that we have a superposition behavior in that asymptotic limit. Here we have justified the validity of that approximation. If we take the asymptotic distance limit to [Eq. ]{}, we will obtain a superposition of two bodies PSA energies in the asymptotic limit. The nonlinearity of the Casimir energy must be given by higher orders expansion terms even for systems with three bodies. The same superposition behavior is expected for the general $N$ body case, and it will be proven in the next section. PSA for general N bodies system ------------------------------- In this section we are going to generalize the PSA energy of three bodies given by [Eq. ]{}. For this purpose we are going to use an iterative procedure which will give us the new terms to include to the PSA of the Casimir energy of $n-1$ bodies when we include another new object to our system. Let us represent the $\mathbb{M}$ matrix of [Eq. ]{} as the sum of its diagonal and its non-diagonal parts as [@Emig; @Casimir; @caso; @escalar]: $$\mathbb{M}_{\alpha\beta} = \delta_{\alpha\beta}\mathbb{T}_{\alpha}^{-1} + (\delta_{\alpha\beta} - 1)\mathbb{U}_{\alpha\beta},$$ or symbolically as $$\mathbb{M} = \mathbb{T}^{-1} + \mathbb{U},$$ then it is easy to find that the inverse of $\mathbb{M}$ is the next perturbative series $$\mathbb{M}^{-1} = \mathbb{T}\sum_{n=0}^{\infty}\left(-\mathbb{U}\mathbb{T}\right)^{n}.$$ On the other hand, the $\mathbb{M}$ matrix of the $N$ objects system is related with the $\mathbb{M}$ matrix of the $N-1$ objects system by block matrices in the next way $$\mathbb{M}_{N} = \left(\begin{array}{c c c} \phantom{-}\mathbb{M}_{N-1\phantom{,\gamma}} & - \mathbb{U}_{N-1,\gamma}\\ - \mathbb{U}_{\gamma,N-1} & \phantom{-}\mathbb{T}_{N\phantom{-1,\gamma}}^{-1} \end{array}\right).$$ Where $\gamma$ index goes from 1 to $N-1$. With this result we will calculate the determinant of $\mathbb{M}_{N}$ obtaining $${\vert \mathbb{M}_{N}\vert} = {\vert \mathbb{M}_{N-1}\vert}{\vert \mathbb{T}_{N}\vert}^{-1}{\vert \mathbb{I} - \mathbb{T}_{N}\mathbb{U}_{\gamma ,N-1}\mathbb{M}_{N-1}^{-1}\mathbb{U}_{N-1,\gamma}\vert}.$$ In the PSA approximation is valid the substitution $\mathbb{M}^{-1} \simeq \mathbb{T}$, where $\mathbb{T}$ is a diagonal matrix whose $N-1$ diagonal elements are $\mathbb{T}_{\gamma}$ with $\gamma$ index defined as before. Then, using this approximation we can approximate the $\mathbb{M}_{N}$ determinant as: $${\vert \mathbb{M}_{N}\vert} = {\vert \mathbb{M}_{N-1}\vert}{\vert \mathbb{T}_{N}\vert}^{-1}{\vert \mathbb{I} - \mathbb{T}_{N}\mathbb{U}_{\gamma ,N-1}\mathbb{T}\mathbb{U}_{N-1,\gamma}\vert}.$$ Where we multiply by blocks the matrix of the last determinant obtaining $$\begin{aligned} {\vert \mathbb{M}_{N}\vert} & = & {\vert \mathbb{M}_{N-1}\vert}{\vert \mathbb{T}_{N}\vert}^{-1}{\vert \mathbb{I} - \left(\begin{array}{c c c c} \mathbb{N}_{1N} & 0 & \dots & 0\\ 0 & \mathbb{N}_{2N} & \dots & 0\\ \vdots & \vdots & \ddots & \vdots\\ 0 & 0 & \dots & \mathbb{N}_{N-1,N} \end{array}\right)\vert}.\nonumber\\ & &\end{aligned}$$ Finally, we obtain the desired result: $${\vert \mathbb{M}_{N}\vert} = {\vert \mathbb{M}_{N-1}\vert}{\vert \mathbb{T}_{N}\vert}^{-1}\prod_{\gamma =1}^{N-1}{\vert \mathbb{I} - \mathbb{N}_{\gamma N}\vert}.$$ The other needed matrix is $${\vert \mathbb{M}_{\infty, N}\vert} = {\vert \mathbb{M}_{\infty ,N-1}\vert}{\vert \mathbb{T}_{N}\vert}^{-1}.$$ And with the initial conditions: $${\vert \mathbb{M}_{1}\vert} = {\vert \mathbb{M}_{\infty, 1}\vert} = {\vert \mathbb{T}_{1}\vert}^{-1},$$ we obtain these determinants in a closed form as $${\vert \mathbb{M}_{N}\vert} = \prod_{k = 1}^{N}{\vert \mathbb{T}_{k}\vert}^{-1}\prod_{l = 2}^{N}\prod_{m = 1}^{l-1}{\vert \mathbb{I} - \mathbb{N}_{lm}\vert}$$ and $${\vert \mathbb{M}_{\infty, N}\vert} = \prod_{k = 1}^{N}{\vert \mathbb{T}_{k}\vert}^{-1}.$$ Taking the logarithms of [Eq. ]{}, the Casimir energy for the N bodies system is approximated as $$\label{N-bodies PSA} E_{N} = \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\sum_{l = 2}^{N}\sum_{m = 1}^{l-1}\log\left({\vert \mathbb{I} - \mathbb{N}_{lm}\vert}\right)$$ in the PSA limit. [Eq. ]{} is the generalization of [Eq. ]{} for the $N$ body case, and it show us that in the PSA limit, we always obtain a superposition principle of the two body PSA Casimir energy despite the fact that Casimir energy is not a nonadittive interaction. We will obtain the non-linear effects in the next orders of the expansion. This result is qualitatively different of the usual PSA procedure of integrating each point of a $N$ point potential over the volume of each object. The reason is simple, a $N$ point potential is proportional to $N$ polarizabilities [@Thiru], but in the soft limit approximation, the lower allowed term is proportional to two polarizabilities. Therefore we should study the $(N-1)$ expansion term to obtain the first one proportional to $N$ polarizabilities. So in a consistent calculation of PSA energies for three or more objects, the asymptotic approximation of the PSA energy in the soft material approximation is different of the $N$ point potential function. In the next section we will find where the contribution of the three point potential is relevant in the diluted limit. Second order expansion of diluted limit {#sec: 5} ======================================= In this section we are going to study the second order expansion of the PSA in the diluted limit. This is a complicated long calculus so, instead the complete electromagnetic case, we will restrict ourselves to the purely electric case. That means that we will ignore the magnetic properties of the bodies. It could be interesting to show the complete study of that series term, but it is a long calculation to show here. However, it possesses theoretical utility, as it shows the lack of the superposition principle in the Casimir energy calculations. The next results have been done with the help of [Mathematica]{} [@Mathematica]. Again we start from the Casimir energy between two compact objects as given by [Eq. ]{}. We studied the first order expansion of the energy in Sect. \[sec: 2\]. Now we are interested in the next order expansion term in polarizabilities, so we maintain the study of the fist term of [Eq. ]{} taking just the case $p = 1$. In addition to that, we take the second order approximation of the Born series of the $\mathbb{T}$ matrix: $$\mathbb{T} = \sum_{n=0}^{\infty}(VG_{0})^{n}V\simeq V + VG_{0}V.$$ Applying the linearity of the trace and taking our attention just to the second order expansion in polarizabilities (that is, to the third order term in polarizabilities), we obtain the following result for the second order of the diluted limit result: $$\begin{aligned} E_{2} & = & - \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\textrm{Tr}\left(V_{1'}G_{01'1}V_{1}U_{12}V_{2}U_{21'}\right) \nonumber\\ & & - \frac{\hbar c}{2\pi}\int_{0}^{\infty}dk\textrm{Tr}\left(V_{1}U_{12'}V_{2'}G_{02'2}V_{2}U_{21}\right)\label{segundo orden PSA}.\end{aligned}$$ We will center our study just on the first integral of , because the analysis of these two integrals is the same. First we make the trace over the $EH$ space, which is automatic here because we have cancelled the magnetic properties of the bodies. In other words, we are not studying the coupling between electric and magnetic induced dipoles. After that we trace over the space coordinates, so we need the matricial form of the electric part of the dyadic Green function given in [Eq. ]{} of the Appendix, where $\textbf{R} = \textbf{r} - \textbf{r}'$. We will need to define the function: $$\begin{aligned} \mathcal{R}[\alpha ,\beta]_{i}(k) & = & R_{\alpha\beta}^2 \left(1+k R_{\alpha\beta}+k^2 R_{\alpha\beta}^2\right)\nonumber\\ & & + \left(3+3 k R_{\alpha\beta}+k^2 R_{\alpha\beta}^2\right)R_{i\alpha\beta}^2.\end{aligned}$$ Where $k$ is the frequency, $\alpha$ and $\beta$ labels the bodies coordinates and $i$ shows the vector component of $\textbf{R}_{\alpha\beta}$ used. So we can solve the trace over the space coordinates obtaining the following result for the first term of : $$\begin{aligned} E_{2a} & = & - \frac{\hbar c}{2\pi}\int_{1}d\textbf{r}_{1}\int_{2}d\textbf{r}_{2}\int_{1'}d\textbf{r}_{1'}\int_{0}^{\infty}dk\nonumber\\ & & \times\frac{e^{-k (R_{12} + R_{21'} + R_{1'1})}}{64\pi^{3}R_{12}^{5}R_{21'}^{5}R_{1'1}^{5}}\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}\tilde{\epsilon}_{1'}\nonumber\\ & & \times[\mathcal{R}[1,2]_{x}(k)\mathcal{R}[2,1']_{x}(k)\mathcal{R}[1',1]_{x}(k)\nonumber\\ & & + \mathcal{R}[1,2]_{y}(k)\mathcal{R}[2,1']_{y}(k)\mathcal{R}[1',1]_{y}(k)\nonumber\\ & & + \mathcal{R}[1,2]_{z}(k)\mathcal{R}[2,1']_{z}(k)\mathcal{R}[1',1]_{z}(k)].\end{aligned}$$ Here the space integrations are performed just over the body volumes because it is there where dielectric and diamagnetic potentials are defined. It is possible to perform the $k$ integral, because we have an exponential multiplied by a polynomial integrated between zero and infinity. It is not shown here because it contains 216 sum terms. For any finite temperature we also obtain a long result, but in the classical limit we can at least obtain a tractable formula. This limit consists of keeping only the $k = 0$ mode. In that case we obtain the result $$\begin{aligned} E_{2a}^{cl} & = & - \frac{\hbar c}{256\pi^{4}}\int_{1}d\textbf{r}_{1}\int_{2}d\textbf{r}_{2}\int_{1'}d\textbf{r}_{1'}\frac{\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}\tilde{\epsilon}_{1'}}{R_{12}^{5}R_{21'}^{5}R_{1'1}^{5}}\nonumber\\ & & \times[\mathcal{R}[1,2]_{x}(0)\mathcal{R}[2,1']_{x}(0)\mathcal{R}[1',1]_{x}(0)\nonumber\\ & & + \mathcal{R}[1,2]_{y}(0)\mathcal{R}[2,1']_{y}(0)\mathcal{R}[1',1]_{y}(0)\nonumber\\ & & + \mathcal{R}[1,2]_{z}(0)\mathcal{R}[2,1']_{z}(0)\mathcal{R}[1',1]_{z}(0)].\end{aligned}$$ Where $\mathcal{R}[\alpha ,\beta]_{i}(0) = \left(R_{\alpha\beta}^{2} + 3R_{\alpha\beta i}^{2}\right)$, so we can write: $$\begin{aligned} \label{Divergent integrand example} E_{2a}^{cl} & = & - \frac{\hbar c}{256\pi^{4}}\int_{1}d\textbf{r}_{1}\int_{2}d\textbf{r}_{2}\int_{1'}d\textbf{r}_{1'}\frac{\tilde{\epsilon}_{1}\tilde{\epsilon}_{2}\tilde{\epsilon}_{1'}}{R_{12}^{5}R_{21'}^{5}R_{1'1}^{5}}\nonumber\\ & & \times[\left(R_{12}^{2} + 3R_{12 x}^{2}\right)\left(R_{21'}^{2} + 3R_{21' x}^{2}\right)\left(R_{1'1}^{2} + 3R_{1'1 x}^{2}\right)\nonumber\\ & & + \left(R_{12}^{2} + 3R_{12 y}^{2}\right)\left(R_{21'}^{2} + 3R_{21' y}^{2}\right)\left(R_{1'1}^{2} + 3R_{1'1 y}^{2}\right)\nonumber\\ & & + \left(R_{12}^{2} + 3R_{12 z}^{2}\right)\left(R_{21'}^{2} + 3R_{21' z}^{2}\right)\left(R_{1'1}^{2} + 3R_{1'1 z}^{2}\right)].\nonumber\\ & &\end{aligned}$$ When performing this integral, a term proportional to $R_{1'1}^{-5}$ and another one to $R_{1'1}^{-3}$ appear. These terms can be problematic because $R_{1'1} = 0$ belongs to the integration interval for all the points of the $1$ body. But these singularities can be removed with the appropriate regularization procedure. Note that if we make $1' = 3$ in [Eq. ]{}, this integral would also be one of the contributions of the second term expansion of a three objects system. In that case we have not got any problem in the volumes integration for this term and this expansion term looks quite similar to the local analog of the three points potential given in [@Thiru]. In fact, in the three (or more) bodies system that term is the first term that breaks the superposition behavior founded in \[sec: 4\]. Replacing the $k^{2}$ term by the operator $-\Delta$ in Eqs. and of the Appendix, and replacing the $k$ term by the operator $\sqrt{-\Delta}$ in Eqs. and of the Appendix, which is a change valid on shell because $G_{0}(R,k) = \frac{e^{-kR}}{4\pi R}$, we recover the formalism used in [@Thiru] in the soft dielectric and soft diamagnetic limit. This scheme let us take into account not only electric phenomena as in [@Thiru], but also magnetic and electro-magnetic coupling effects. Taking this argument into account, we can understand the structure of the perturbation terms of the PSA limit of the electromagnetic Casimir energy between $N$ bodies. The Casimir energy in the PSA limit has the structure of a series of infinity terms whose $n$-order term is the sum of integrals over $n+2$ bodies (which can be repeated or not) of the $(n+2)$ points local em potential. Each series term come from the expansions made in [Eq. ]{} and in [Eq. ]{} and is proportional to the product of $n+2$ permittivities. If we take the asymptotic distance limit of these integrals, we will obtain a series of $n$ points em potential now with polarizabilities instead susceptibilities as in [@Thiru], but in the soft response limit. When we use perturbations of the tree term of the Born expansion, it appears divergent terms in the integrand as seen in [Eq. ]{}. These singularities can be removed with the appropriate regularization procedure. Final Remarks ============= We have calculated the full electromagnetic Casimir energy in the diluted limit between two bodies using the formalism given in [@Kardar-Geometrias-Arbitrarias] reobtaining the energy given in [@Milton] in the pure electric case. Although this formalism fails in the perfect metal case, it is valid for soft dielectric and diamagnetic bodies. This formalism can be extended for an arbitrary number of bodies. We have shown that the Casimir energy, which is a nonadittive interaction, has a superposition behavior in the first expansion of the PSA and how the integration of different $N$ point potentials appear in the perturbation series to contribute in the whole energy. Then we have a systematic procedure to calculate Casimir energies valid for soft dielectric and diamagnetic bodies. I acknowledge helpful discussions with R. Brito and T. Emig. This research was supported by projects MOSAICO, UCM/PR34/07-15859 and a FPU MEC grant. Appendix: Obtention of the matricial Green function {#appendix-obtention-of-the-matricial-green-function .unnumbered} =================================================== \[sec:Appendix\] Here we have calculated the Casimir energy using a matricial Green function for the electromagnetic fields instead the dyadic Green function used in [@Kardar-Geometrias-Arbitrarias]. in this appendix we will see that both formalisms are equivalents just using as field sources the induced polarization and magnetization vectors instead the induced currents into the bodies. The induced currents into the bodies are [@Jackson]: $$\textbf{j} = -ik\textbf{P} + \nabla\times\textbf{M}.$$ We introduce this linear change of variable in the partition function, so there is not any new relevant term in the action of the problem. This action will be transformed in the next way: $$S = \int d\textbf{x}\bar{\textbf{j}}\mathcal{G}_{0}\textbf{j},$$ $$\begin{aligned} S & = & \int d\textbf{x}k^{2}\bar{\textbf{P}}\mathcal{G}_{0}\textbf{P} + \int d\textbf{x}\nabla\times\bar{\textbf{M}}\mathcal{G}_{0}\nabla\times\textbf{M}\nonumber\\ & & + \int d\textbf{x}ik\bar{\textbf{P}}\mathcal{G}_{0}\nabla\times\textbf{M} - \int d\textbf{x}\nabla\times\bar{\textbf{M}}\mathcal{G}_{0}ik\textbf{P}.\end{aligned}$$ Where $$\mathcal{G}_{0} = \left[\delta_{ij} - \frac{1}{k^{2}}\nabla_{i}\nabla_{j}\right]G_{0},$$ is the Green dyadic function. We integrate by parts each action term obtaining: $$\begin{aligned} S_{EE} & = & \int d\textbf{x}k^{2}\bar{\textbf{P}}\mathcal{G}_{0}\textbf{P},\nonumber\\ & = & \int d\textbf{x}\bar{\textbf{P}}_{i}\left[k^{2}\delta_{ij} - \nabla_{i}\nabla_{j}\right]G_{0}\textbf{P}_{j},\nonumber\\ & = & \int d\textbf{x}\bar{\textbf{P}}_{i}G_{ij}^{EE}\textbf{P}_{j},\end{aligned}$$ $$\begin{aligned} S_{EH} & = & \int d\textbf{x}ik\bar{\textbf{P}}\mathcal{G}_{0}\vec{\nabla}\times\textbf{M}, \nonumber\\ & = & \int d\textbf{x}ik\bar{\textbf{P}}_{i}\left[\delta_{ij} - \frac{1}{k^{2}}\nabla_{i}\nabla_{j}\right]G_{0}\epsilon_{j\alpha\beta} \nabla^{\alpha}\textbf{M}^{\beta},\nonumber\\ & &\end{aligned}$$ $$\begin{aligned} S_{EH} & = & \int d\textbf{x}\bar{\textbf{P}}_{i}\left[ik\epsilon_{i\alpha\beta}\nabla^{\beta}\right]G_{0}\textbf{M}^{\alpha} \nonumber\\ & & + \int d\textbf{x}\bar{\textbf{P}}_{i}\left[ik\frac{1}{k^{2}}\nabla_{i}\nabla_{j}\epsilon_{j\alpha\beta}\nabla^{\beta}\right]G_{0}\textbf{M}^{\alpha}.\end{aligned}$$ The second term is zero: $$\epsilon_{j\alpha\beta}\nabla_{j}\textbf{M}^{\alpha}\nabla^{\beta}G_{0} = \textbf{M}\cdot\vec{\nabla}\times\vec{\nabla}G_{0} = \textbf{M}\cdot\textbf{0} = 0.$$ So we finally get: $$\begin{aligned} S_{EH} & = & \int d\textbf{x}\bar{\textbf{P}}_{i}\left[-ik\epsilon_{ij\beta}\nabla^{\beta}\right]G_{0}\textbf{M}^{j},\nonumber\\ & = & \int d\textbf{x}\bar{\textbf{P}}_{i}G_{ij}^{EH}\textbf{M}^{j}.\end{aligned}$$ $S_{HE}$ is similar to $S_{EH}$, but with a changed sign because the complex conjugation of the induced current into the action: $$S_{HE} = - \int d\textbf{x}\nabla\times\bar{\textbf{M}}\mathcal{G}_{0}ik\textbf{P},$$ $$S_{HE} = \int d\textbf{x}\bar{\textbf{M}}_{i}G_{ij}^{HE}\textbf{P}^{j} = \int d\textbf{x}\bar{\textbf{M}}_{i}\left[- G_{ij}^{EH}\right]\textbf{P}^{j}.$$ And the last term: $$S_{HH} = \int d\textbf{x}\nabla\times\bar{\textbf{M}}\mathcal{G}_{0}\vec{\nabla}\times\textbf{M},$$ $$S_{HH} = \int d\textbf{x}\epsilon^{i\alpha\beta}\nabla_{\alpha}\bar{\textbf{M}}_{\beta}\left[\delta_{ij} - \frac{1}{k^{2}}\nabla_{i}\nabla_{j}'\right]G_{0}\epsilon_{jab}\nabla'^{a}\textbf{M}^{b},$$ $$S_{HH} = \int d\textbf{x}\epsilon^{i\alpha\beta}\bar{\textbf{M}}_{\alpha}\nabla_{\beta}\left[\delta_{ij} - \frac{1}{k^{2}}\nabla_{i}\nabla_{j}'\right]G_{0}\epsilon_{jab}\textbf{M}^{a}\nabla'^{b},$$ $$\begin{aligned} S_{HH} & = & \int d\textbf{x}\bar{\textbf{M}}_{\alpha}\left(\epsilon^{i\alpha\beta}\nabla_{\beta}\delta_{ij}G_{0}\epsilon_{jab}\nabla'^{b}\right)\textbf{M}^{a} \nonumber\\ & & - \int d\textbf{x}\bar{\textbf{M}}_{\alpha}\left(\epsilon^{i\alpha\beta}\nabla_{\beta}\frac{1}{k^{2}}\nabla_{i}\nabla'^{j}G_{0}\epsilon_{jab}\nabla'^{b}\right)\textbf{M}^{a},\nonumber\\ & &\end{aligned}$$ Using $\epsilon^{i\alpha\beta}\delta_{i}^{j}\epsilon_{jab} = \delta_{a}^{\alpha}\delta_{b}^{\beta} - \delta_{b}^{\alpha}\delta_{a}^{\beta}$ and $\nabla_{b}\nabla'^{b}G_{0} = k^{2}G_{0}$, we get the next result: $$\begin{aligned} \left(\epsilon^{i\alpha\beta}\nabla_{\beta}\delta_{ij}G_{0}\epsilon_{jab}\nabla'^{b}\right) & = & \left[\delta_{a}^{\alpha}\delta_{b}^{\beta} - \delta_{b}^{\alpha}\delta_{a}^{\beta}\right]\nabla_{\beta}\nabla'^{b}G_{0}, \nonumber\\ & = & \left[\delta_{a}^{\alpha}\nabla_{b}\nabla'^{b} - \nabla_{a}\nabla'^{\alpha}\right]G_{0},\nonumber\\ & = & \left[\delta_{a}^{\alpha}k^{2} - \nabla_{a}\nabla'^{\alpha}\right]G_{0}.\end{aligned}$$ The other term requires even a more tedious work, but it is easy to obtain that $$\frac{-1}{k^{2}}\epsilon^{i\alpha\beta}\epsilon_{jab}\nabla_{\beta}\nabla_{i}\nabla'^{j}\nabla'^{b}G_{0} = 0,$$ because this differential operator is zero. So $S_{HH}$ is $$\begin{aligned} S_{HH} & = & \int d\textbf{x}\bar{\textbf{M}}_{\alpha}\left(\epsilon^{i\alpha\beta}\nabla_{\beta}\delta_{ij}G_{0}\epsilon_{jab}\nabla'^{b}\right)\textbf{M}^{a}, \nonumber\\ & = & \int d\textbf{x}\bar{\textbf{M}}_{\alpha}\left[\delta_{a}^{\alpha}k^{2} - \nabla_{a}\nabla'^{\alpha}\right]G_{0}\textbf{M}^{a},\nonumber\\ & = & \int d\textbf{x}\bar{\textbf{M}}_{i}G_{ij}^{HH}\textbf{M}_{j}.\end{aligned}$$ After a Wick rotation, we obtain the used form of the matricial Green function, which components are, using $G_{0}(R,k) = \frac{e^{-kR}}{4\pi R}$ and $R = {\vert \textbf{r} - \textbf{r}'\vert}$: $$\begin{aligned} G_{0ij}^{EE}(R,k) & = & \left[k^{2}\delta_{ij} + \nabla_{i}\nabla_{j}'\right]G_{0}(R,k),\label{Acoplo EE funcion Green}\\ G_{0ij}^{EH}(R,k) & = & - k\epsilon_{ijk}\nabla_{k}G_{0}(R,k),\label{Acoplo EH funcion Green}\\ G_{0ij}^{HE}(R,k) & = & k\epsilon_{ijk}\nabla_{k}G_{0}(R,k),\label{Acoplo HE funcion Green}\\ G_{0ij}^{HH}(R,k) & = & \left[k^{2}\delta_{ij} + \nabla_{i}\nabla_{j}'\right]G_{0}(R,k)\label{Acoplo HH funcion Green}.\end{aligned}$$ Which are the results used in [Eq. ]{}. It is also possible to obtain the same result from fluctuation - dissipation theorem, as made in [@Agarwal]. [20]{} H. B. G. Casimir, Proc. K. Ned. Akad. Wet. 51, 793 (1948). M. Bordag, G.L. Klimchitskaya, U.Mohideen, and V.M. Mostepanenko. *Advances in the Casimir Effect*, Oxford University Press, (2009). ISBN-13: 978-0-19-923874-3. H. B. G. Casimir and D. Polder, Phys. Rev. **73**, 360 (1948). E. M. Lifshitz, Sov. Phys. JETP, **2**, 73 (1956). R. Golestanian, Phys. Rev. Lett. 95, 230601, (2005). K. A. Milton, P. Parashar and J. Wagner, Phys. Rev. Lett. **101**, 160402 (2008). T. Emig, N. Graham, R.L. Jaffe and M. Kardar, Phys. Rev. Lett. 99, 170403 (2007). arXiv:0707.1862 A. Bulgac, P. Magierski and A. Wirzba, Phys. Rev. D, **73**, 025007 (2006). P. Rodriguez-Lopez, S. J. Rahi, and T. Emig, Phys. Rev. A, **80**, 022519 (2009). G. Feinberg and J. Sucher, Phys. Rev. A, **2**, 2395 (1970). A. Galindo and P. Pascual, 1990. *Quantum Mechanics II.* First Edition, Ed. Springer-Verlag. ISBN: 0 387 52309 X. R. Balian and B. Duplantier, Ann. Phys. (N.Y.) **112**, 165 (1978). K. A. Milton, P. Parashar and J. Wagner. arXiv:0811.0128v2 \[math-ph\] M. Born and E. Wolf, 1980. *Principles of Optics. Electromagnetic theory of propagation interference and difraction of light.* Sixth Edition, Ed. Cambridge University Press. ISBN: 0 521 63921 2. G. Barton, Phys. Rev. A, **64**, 032102 (2001). T. Emig, N. Graham, R. L. Jaffe, and M. Kardar, Phys. Rev. D, **77**, 025005 (2008). E. A. Power and T. Thirunamachandran, Proc. R. Soc. Lond. Ser. A, **401**, 267, (1985). Wolfram Research, Inc., Mathematica, Version 6.0, Champaign, IL (2007) J. D. Jackson, 1998. *Classical Electrodynamics* Third Edition, Ed. John Wiley & Sons, Jnc. ISBN: 0 471 30932 X. G. S. Agarwal, Phys. Rev. A, **11**, 230 (1975).

ArXiv

--- abstract: 'We consider random walks (RWs) and self-avoiding walks (SAWs) on disordered lattices directly at the percolation threshold. Applying numerical simulations, we study the scaling behavior of the models on the incipient percolation cluster in space dimensions $d=2, 3, 4$. Our analysis yields estimates of universal exponents, governing the scaling laws for configurational properties of RWs and SAWs.' address: - '$^1$ Institut für Theoretische Physik and Centre for Theoretical Sciences (NTZ),Universität Leipzig, Postfach 100920, D-04009 Leipzig, Germany' - '$^2$ Institute for Condensed Matter Physics, National Academy of Sciences of Ukraine, UA–79011 Lviv, Ukraine' author: - 'V Blavatska$^{1,2}$ and W Janke$^1$' title: 'Walking on fractals: diffusion and self-avoiding walks on percolation clusters' --- Introduction ============ The model of a random walk (RW) provides a good description of diffusion processes, such as for example encountered for electrons in metals or colloidal particles in solution [@RWbook]. The averaged mean square displacement of the diffusive particle at time $t$ (or, equivalently, after $t$ steps on a lattice) scales as $$\langle R^2 \rangle \sim t^{2\nu_{{\rm RW}}}, \label{diff}$$ where in a non-disordered medium $\nu_{{\rm RW}}=1/2$, independently of the space dimension $d$. A RW is a fractal object, with fractal dimension $d_{{\rm RW}}=1/\nu_{RW}$. The number of all possible trajectories $C_t$ for a randomly walking particle of $t$ steps can be found exactly: $C_t=z_0^t$, where $z_0$ is the coordination number of the corresponding lattice. Forbidding the trajectory of a random walk to cross itself, we obtain a self-avoiding walk (SAW), which is one of the most successful in describing the universal configurational properties of a long, flexible single polymer chain in good solvent [@desCloizeaux90]. The average squared end-to-end distance $\langle R^2\rangle$ and the number of configurations $ C_N $ of SAWs with $N$ steps on the underlying lattice obey the scaling laws: $$\label{scaling} \langle R^2 \rangle \sim N^{2\nu_{{\rm SAW}}},\mbox{\hspace{3em}} C_N \sim z^{N} N^{\gamma_{{\rm SAW}}-1},$$ where $\nu_{{\rm SAW}}, \gamma_{{\rm SAW}}$ are universal exponents that only depend on the space dimensionality $d$, and $z$ is a non-universal fugacity, counting the average number of accessible nearest-neighbor sites. The properties of SAWs on a regular lattice have been studied in detail both in analytical approaches [@Guillou80; @Nienhuis82; @Guillou85; @Guida98] and computer simulations [@Rosenbluth55; @Madras88; @MacDonald92; @MacDonald00; @Li95; @Caracciolo98]. For example, in the space dimension $d{=}3$ one finds within the frame of the field-theoretical renormalization group approach $\nu_{\rm SAW}{=}0.5882\pm 0.0011$ [@Guida98] and Monte Carlo simulations give $\nu_{\rm SAW}{=}0.5877\pm0.0006$ [@Li95]. For space dimensions $d$ above the upper critical dimension $d_{\rm up}{=}4$, the effect of self-avoidance becomes irrelevant and SAWs behave effectively as random walks with exponents $\nu_{{\rm RW}}=1/2$, $\gamma_{{\rm RW}}=1$. The problem of random walks in disordered media is of great interest since it is connected with a large amount of physical phenomena: transport properties in fractures and porous rocks, the anomalous density of states in randomly diluted magnetic systems, silica aerogels and in glassy ionic systems, diffusion-controlled fusion of excitations in porous membrane films etc. (see, e.g., Ref. [@Havlin87] for a review). Similarly, SAWs on randomly diluted lattices may serve as a model of linear polymers in a porous medium. Much of our understanding of disordered systems comes from percolation theory [@Stauffer]. A disordered medium can be modelled as randomly diluted lattice, with a given concentration $p$ of lattice sites allowed for walking. Most interesting is the case, when $p$ equals the critical concentration $p_{c}$, the site-percolation threshold (see Table \[dim\]), and an incipient percolation cluster can be found in the system. Studying properties of percolative lattices, one encounters two possible statistical averages. In the first, one considers only percolation clusters with linear size much larger than the typical length of the physical phenomenon under discussion. The other statistical ensemble includes all the clusters which can be found in a percolative lattice. For the latter ensemble of all clusters, the walks can start on any of the clusters, and for an $N$-step walk, performed on the $i$th cluster, we have $\langle R^2 \rangle \sim l_i^2$, where $l_i$ is the averaged size of the $i$th cluster. In what follows, we will be interested in the former case, when trajectories of walks reside only on the percolation cluster. In this regime, the scaling laws (\[diff\]), (\[scaling\]) hold with new exponents $\nu_{{\rm RW}}^{p_c}\neq \nu_{{\rm RW}}$ [@Sahimi83; @Majid84; @Pandey83; @Alexander82; @Avraham82; @Havlin83; @Argyrakis84; @McCarthy88; @Lee00; @Bug86; @Hong84; @Mastorakos93; @Webman81; @Gefen83; @Rammal83; @Mukherjee95], $\nu_{{\rm SAW}}^{p_c}\neq\nu_{{\rm SAW}},\gamma_{{\rm SAW}}^{p_c}\neq\gamma_{{\rm SAW}} $ [@Kremer81; @Lee89; @Kim90; @Woo91; @Grassberger93; @Lee96; @Meir89; @Lam90; @Nakanishi92; @Rintoul94; @Ordemann00; @Nakanishi91; @Barat91; @Sahimi84; @Rammal84; @Kim87; @Roy90; @Roy87; @Aharony89; @Lam84; @Blavatska04; @Janssen07; @Blavatska08]. A hint to the physical understanding of this phenomenon is given by the fact that weak disorder does not change the dimension of a lattice, whereas the percolation cluster itself is a fractal object with fractal dimension $d_{p_c}^F$ dependent on $d$ (see Table \[dim\]). In this way, scaling law exponents of residing walks change with the dimension $d_{p_c}^F$ of the (fractal) lattice on which the walk resides. Since $d_{\rm up}{=}6$ for percolation [@Stauffer], the exponents $\nu_{{\rm SAW}}^{p_c}(d\geq 6){=}1/2$, $\gamma_{{\rm SAW}}^{p_c}(d\geq 6){=}1$. Our present paper aims to supplement the studies of random and self-avoiding walks on percolative lattices by obtaining numerical values for exponents, governing the scaling behavior of the models, up to $d=4$ by computer simulations. The layout of the paper is as follows: in the next section, we will present in detail the procedure of extracting the percolation cluster and its backbone on disordered lattices at the percolation threshold. In section III we describe the pruned-enriched Rosenbluth algorithm, applied to study the scaling of self-avoiding walks, and present the results obtained. In the next section we consider the method for studying random walks on percolation clusters. In Section V, we end up by giving conclusions and an outlook. Construction of percolation cluster =================================== We consider site percolation on a regular lattice of edge length $L=400,200,50$ in dimensions $d=2,3,4$, respectively. Each site of the lattice is assigned to be occupied with probability $p_c$ (values of critical concentration in different dimensions are given in the Table \[dim\]), and empty otherwise. To describe the procedure of extracting the percolation cluster, let us consider schematically the two-dimensional case. We apply an algorithm based on the one proposed by Hoshen and Kopelman [@Hoshen76]. As a first step, a label is prescribed to each of the occupied sites. Such a labeling process is regulated, we start, for example, from the first “column" of the lattice, label the occupied sites upwards, and then turn to the next “column", as shown in Fig. \[gratka\], left. Next, we start the procedure of burning the occupied sites. Again, in the same order, starting from the bottom of the first “column" of the lattice, for each of the labeled sites (say, $i$), we check whether its nearest neighbors are also occupied or not. If yes, two possibilities appear: 1) The label of the neighbor is larger than the label of site $i$. In this case, we change the label of the neighbor to that of site $i$. 2) The label of the neighbor is smaller than that of $i$. Then, we change the label of site $i$ to that of the neighbor. ![\[gratka\]Procedure of site labeling and extracting the percolation cluster.](gratka1.eps "fig:"){width="4cm"} ![\[gratka\]Procedure of site labeling and extracting the percolation cluster.](gratka2.eps "fig:"){width="3.8cm"} Such a burning procedure is applied until no more change of site labels is needed. As a result, we end up with groups of clusters of occupied sites with the same labels (Fig. \[gratka\], right). Finally, we check, whether there exists a cluster, that wraps around the lattice in all $d$ directions. If yes, we have found the percolation cluster (Fig. \[perc\]). If not, this disordered lattice is rejected and a new one is constructed. Note, that on finite lattices the definition of spanning clusters is not unique (e.g., one could consider clusters connecting only two opposite borders), but all definitions are characterized by the same fractal dimension and are thus equally legitimate. The here employed definition has the advantage of yielding the most isotropic clusters. Note also that directly at $p=p_c$ more than one spanning cluster can be found in the system, and the probability $P(k)$ for at least $k$ separated clusters grows with the space dimension as $P(k)\sim \exp(-\alpha k^{ d/(d-1)})$ [@Aizenman97]. In our study, we take into account only one cluster per each disordered lattice constructed, in order to avoid presumable correlations of the data. ![\[perc\]Percolation cluster on a $d=2$-dimensional regular lattice of edge length $L=50$.](perc.eps){width="4cm"} Aiming to investigate the scaling of SAWs on percolative lattices, we are interested in the backbone of percolation clusters, which can be defined as follows. Assume that each bond (or site) of the cluster is a resistor and that an external potential drop is applied at two ends of the cluster. The backbone is the subset of the cluster consisting of all bonds (or sites) through which the current flows; i.e., it is the structure left when all “dangling ends" are eliminated from the cluster. The SAWs can be trapped in “dangling ends", therefore infinitely long chains can only exist on the backbone of the cluster. The algorithm for extracting the backbone of obtained percolation clusters was first introduced in Ref. [@Herrmann84] and improved in Ref. [@Porto97]. We choose the starting point – “seed” – at the center of the cluster, and find the chemical distance $l$ of all the sites belonging to the percolation cluster to this starting point. In Fig. \[himia\], the starting point is numbered with $0$, and the chemical distance of all the other sites are depicted. The burning algorithm is divided into two parts. First, we start from some site of the edge of the lattice belonging to the percolation cluster and consider it as burning site. At the next step, if the nearest neighbor of this site has the chemical distance smaller than the burning site itself, the nearest neighbor site is burnt. Such a procedure ends when the “seed" site is reached. All the thus obtained burnt sites are located along the shortest path between the “seed" and the given site at the edge of the percolation cluster, as is shown in Fig. \[himia\]. The same algorithm is applied successively to all the edge sites. As a result, we obtain the so-called skeleton or elastic backbone [@Havlin84], shown in Fig. \[geom\], left. In the second part of the algorithm, we consider successively each site of the elastic backbone, and check, whether a “loop" starts from this site. “Loop" is a path of sites, belonging to the percolation cluster, which is connected with the elastic backbone by no less than two sites. All sites of the elastic backbone together with the sites of “loops" form finally the geometric backbone of the cluster (see Fig. \[geom\], right). ![\[himia\] For all sites of a percolation cluster the chemical distance from the starting site is calculated. The minimal paths from all the sites on the edge of the percolation cluster to this starting point are found, which form the elastic backbone of the percolation cluster.](himia.eps){width="4cm"} ![\[geom\]Elastic and geometrical backbones of the percolation cluster depicted in Fig. \[perc\].](elas.eps "fig:"){width="4cm"} ![\[geom\]Elastic and geometrical backbones of the percolation cluster depicted in Fig. \[perc\].](geom.eps "fig:"){width="4cm"} The results for fractal dimensions of the percolation cluster and its geometrical backbone in $d{=}2,3,4$ are compiled in Table \[dim\]. Self-avoiding walks on percolation clusters =========================================== The method ---------- For the sampling of SAWs, we use the pruned-enriched Rosenbluth method (PERM), proposed in the work of Grassberger [@Grassberger97]. The algorithm is based on ideas from the very first days of Monte Carlo simulations, the Rosenbluth-Rosenbluth (RR) method [@Rosenbluth55] and enrichment [@Wall59]. Let us consider the growing polymer chain, i.e. the $n$th monomer is placed at a randomly chosen neighbor site of the last placed $(n-1)$th monomer ($n\leq N$, where $N$ is total length of polymer). The growth is stopped, if the total length $N$ of the chain is reached, then the next chain is started to grow from the same starting point. In order to obtain correct statistics, if this new site is occupied, any attempt to place a monomer at it results in discarding the entire chain. This leads to an exponentional “attrition", the number of discarded chains grows exponentially with the chain length, which makes the method useless for long chains. In the RR method, occupied neighbors are avoided without discarding the chain, but the bias is corrected by means of giving a weight $W_n\sim (\prod_{l=1}^n m_l)$ to each sample configuration at the $n$th step , where $m_l$ is the number of free lattice sites to place the $l$th monomer. When the chain of total length $N$ is constructed, the new one starts from the same starting point, until the desired number of chain configurations are obtained. The configurational averaging, e.g., for the end-to-end distance $r(N)\equiv \sqrt{R^2(N)}$, then has the form: $$\begin{aligned} &&\langle r (N) \rangle=\frac{1}{Z_N}{\sum_{{\rm conf}}W_N^{{\rm conf}}(\vec{r}_N-\vec{r}_0)^2}, \,\,\,\,Z_N=\sum_{{\rm conf}} W_N^{{\rm conf}} \label {R},\end{aligned}$$ where $\vec{r}_0$ is the position of the starting point of the growing chains, $\vec{r}_k$ is the position of the $k$th monomer, and $Z_N$ is the partition sum. The Rosenbluth method, however, also suffers from attrition: if all next neighbors at some step ($n<N$) are occupied, i.e., the chain is running into a “dead end", the complete chain has to be discarded and the growth process has to be restarted. Combining the chain growth algorithm with population control, such as PERM (pruned-enriched Rosenbluth method) [@Grassberger97] leads to a considerable improvement of the efficiency by increasing the number of successfully generated chains. The weight fluctuations of the growing chain are suppressed in PERM by pruning configurations with too small weights, and by enriching the sample with copies of high-weight configurations [@Grassberger97]. These copies are made while the chain is growing, and continue to grow independently of each other. Pruning and enrichment are performed by choosing thresholds $W_n^{<}$ and $W_n^{>}$ depending on the estimate of the partition sum for the $n$-monomer chain. These thresholds are continuously updated as the simulation progresses. The zeroes iteration is a pure chain-growth algorithm without reweighting. After the first chain of full length has been obtained, we switch to $W_n^{<}$, $W_n^{>}$. If the current weight $W_n$ of an $n$-monomer chain is less than $W_n^{<}$, a random number $r={0,1}$ is chosen; if $r=0$, the chain is discarded, otherwise it is kept and its weight is doubled. Thus, low-weight chains are pruned with probability $1/2$. If $W_n$ exceeds $W_n^{>}$, the configuration is doubled and the weight of each identical copy is taken as half the original weight. For a value of the weight lying between the thresholds, the chain is simply continued without enriching or pruning the sample. For updating the threshold values we apply similar rules as in [@Hsu03; @Bachmann03]: $W_n^{>}=C(Z_n/Z_1)(c_n/c_1)^2$ and $W_n^{<}=0.2W_n^{>}$, where $c_n$ denotes the number of created chains having length $n$, and the parameter $C$ controls the pruning-enrichment statistics. After a certain number of chains of total length $N$ is produced, the given tour is finished and a new one starts. We adjust the pruning-enrichment control parameter such that on average 10 chains of total length $N$ are generated per each tour [@Bachmann03]. Also, what is even more important for efficiency, in almost all iterations at least one such a chain was created. The number of different trajectories of SAWs with $N$ steps can be then estimated as averaged weight: $$C_N =\frac{1}{t}\sum_{{\rm conf}}W_N^{{\rm conf}}, \label{number}$$ where $t$ is the number of successful tours. PERM has been applied to a wide class of problems, in particular study of $\Theta$-transition in homopolymers [@Grassberger97], trapping of random walkers on absorbing lattices [@Mehra02], study of protein folding [@Frauenkron98; @Bachmann03] etc. Results ------- To study scaling properties of SAWs on percolating lattices, we have to perform two types of averaging: the first average is performed over all SAW configurations on a single percolation cluster according to (\[R\]); the second average is carried out over different realizations of disorder, i.e. over all percolation clusters constructed: $$\begin{aligned} &&\overline{\langle r \rangle}{=}\frac{1}{M}\sum_{i{=}1}^M \langle r \rangle_i,\label{av}\\ &&\overline{ C_N }{=}\frac{1}{M}\sum_{i{=}1}^M C_{N,i},\end{aligned}$$ where $M$ is the number of different clusters and the index $i$ means that a given quantity is calculated on the cluster $i$. ![Disorder averaged end-to-end distance vs number of steps in double logarithmic scale for SAWs on the backbone of percolation clusters in $d{=}2$ (pluses), $d{=}3$ (stars), $d{=}4$ (squares). Lines represent linear fitting, statistical error bars are of the size of symbols.[]{data-label="sawr"}](sawr.eps){width="7cm"} [@r l l l ]{} $\nu_{{\rm SAW}}^{p_c} \setminus d$ & 2 & 3 & 4\ FL Eq. (\[kremer\]) & 7/9 & 0.665& 0.594\ [@Sahimi84] & 0.778 & 0.662 & 0.593\ [@Roy90] & 0.77& 0.66 & 0.62\ [@Roy87] & 0.770 & 0.656 & 0.57\ [@Aharony89] & 0.76 & 0.65 & 0.58\ EE [@Lam90] &0.745(10)& 0.635(10)&\ [@Rintoul94]&0.770(5)& 0.660(5)&\ [@Ordemann00]&0.778(15)& 0.66(1)&\ RS [@Sahimi84] & 0.778&0.724&\ [@Lam84] & 0.767 & &\ RG [@Blavatska04] & 0.785 & 0.678& 0.595\ [@Janssen07] & 0.796 & 0.669& 0.587\ MC [@Woo91] & 0.77(1) & &\ [@Grassberger93] & 0.783(3) & &\ [@Lee96] & & 0.62–0.63 &0.56–0.57\ [@Ordemann00]&0.787(10)& 0.662(6)&\ our results & $ 0.782\pm 0.003$ & $0.667\pm 0.003$ & $0.586\pm 0.003$\ ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/N^{\nu_{{\rm SAW}}^{p_c}}$ in $d{{=}}2$ dimensions. Lattice size $L{{=}}400$, number of SAW steps $N{{=}}140$ (squares), $N{{=}}120$ (triangles), $N{{=}}100$ (crosses).[]{data-label="prsaw2"}](rozpR3d.eps){width="6.8cm"} ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/N^{\nu_{{\rm SAW}}^{p_c}}$ in $d{{=}}3$ dimensions. Lattice size $L{{=}}200$, number of SAW steps $N{{=}}80$ (squares), $N{{=}}60$ (triangles), $N{{=}}50$ (crosses).[]{data-label="prsaw3"}](rozpR2d.eps){width="7cm"} ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/N^{\nu_{{\rm SAW}}^{p_c}}$ in $d{{=}}4$ dimensions. Lattice size $L{{=}}50$, number of SAW steps $N{{=}}30$ (squares), $N{{=}}20$ (triangles), $N{{=}}15$ (crosses).[]{data-label="prsaw4"}](rozpR4d.eps){width="6.6cm"} The SAW statistics crucially depends on the type of disorder averaging, namely, whether the disorder is considered to be “annealed" (positions of defects are in thermodynamical equilibrium with the system) or “quenched" (positions of defects are out of equilibrium). As it was pointed out in Ref. [@Doussal91], the correctness of results, obtained in the picture of “quenched" disorder, depends on whether the location of the starting point of a SAW is fixed while the configurational averaging is performed, or not. In the latter case, one has to average over all locations and effectively this corresponds to the case of annealed disorder. An interesting question arises: what is the difference in statistics between SAWs walking on percolation clusters and the backbone of percolation clusters, after eliminating all the “dead ends"? First Kim [@Kim90] claimed, based on a scaling argument, that the critical behavior on the percolation cluster is the same as that on the backbone, namely $\nu_{{\rm SAW}}^{p_c}=\nu_{{\rm SAW}}^{B}$. This equality was also assumed by Rammal [@Rammal84] in deriving a Flory formula for SAWs on fractal substrates. This can be easily explained: since the walks, which visit the dead ends are eventually terminated after a certain number of steps, the walks that survive in the limit $N\to \infty$ are those confined within the backbone. However, in a numerical study [@Woo91] it was found, that $\nu_{{\rm SAW}}^B>\nu_{{\rm SAW}}^{p_c}$, and, moreover, that $\nu_{{\rm SAW}}^{p_c}$ almost equals the value for SAWs on pure lattices. It was argued, that the averaged end-to-end distance of SAWs on the backbone is significantly enhanced in comparison with averaging on the full percolation cluster. We have checked this, comparing results obtained by us for the averaged end-to-end distance $\overline{\langle r(N) \rangle}$ on percolation clusters and the backbone of percolation clusters. We conclude, that there is practically no difference in scaling for SAWs on both types of clusters, the SAW statistics is determined by the backbone of percolation clusters. To study the scaling properties of SAWs on the backbone of percolation clusters, we choose as the starting point the “seed" of the cluster, and apply the PERM algorithm, taking into account, that a SAW can have its steps only on the sites belonging to the backbone of the percolation cluster. We use lattices of size up to $L_{{\rm max}}{=}400, 200, 50$ in $d{=}2,3,4$, respectively, and performed averages over 1000 percolation clusters in each case. Estimates for the critical exponents $\nu_{{\rm SAW}}^{p_c}$ were obtained by linear least-square fitting (see Tables \[2d\], \[3d\], \[4d\] in the Appendix). Note, that since we can construct lattices only of a finite size $L$, it is not possible to perform very long SAWs on it. For each $L$, the scaling for $\overline{\langle r(N) \rangle}$ holds only up to some “marginal" number of SAWs steps $N_{{\rm marg}}\sim L^{1/\nu_{{\rm SAW}}^{p_c}}$ [@Blavatska08]. We take this into account when analyzing the data obtained; for each lattice size we are interested only in values of $N<N_{{\rm marg}}$, thus avoiding distortions, caused by finite-size effects. Our results for the scaling exponent $\nu_{{\rm SAW}}^{p_c}$ for SAWs on the backbone of percolation clusters [@Blavatska08] are given in Table \[allnu\] and compared with previous estimates obtained by a variety of different techniques. We see that the value of $\nu_{{\rm SAW}}^{p_c}$ is larger than the corresponding exponent on the pure lattice; presence of disorder leads to stretching of the trajectory of self-avoiding walks. A simple modified Flory formula for the exponent of a SAW on a percolation cluster, proposed a long time ago by Kremer [@Kremer81], $$\nu_{\rm SAW}^{p_c}=3/(d_{p_c}^F+2), \label{kremer}$$ gives numbers in an astonishingly good agreement with our numerical data (see Table \[allnu\]). For the estimates we have used the values of the fractal dimension of percolation clusters from Table \[dim\]. Since $d_{\rm up}=6$ for percolation and $d_{p_c}^F(d\geq6)=4$ [@Stauffer], we receive from Eq. (\[kremer\]) that the exponent $\nu_{\rm SAW}^{p_c}(d\geq 6)=1/2$. Note, that there exists a whole family of more sophisticated Flory-like theories [@Kim90; @Sahimi84; @Rammal84; @Kim87; @Roy90; @Roy87; @Aharony89]. The disorder averaged distribution function, defined via $$\overline{\langle r \rangle}=\sum_{r}r {\overline {P(r,N)}} \label{prob}$$ can be written in terms of the scaled variables $r/\overline{\langle r\rangle}$ as $$r\overline{P(r,N)}\sim f(r/\overline{\langle r\rangle})\sim f(r/N^{\nu_{{\rm SAW}}^{p_c}}).$$ The distribution function is normalized according to $\sum_{r}{\overline {P(r,N)}}{{=}}1$. The numerical results for the distribution function in $d=2$,$3$, and $4$ are shown in Figs. \[prsaw2\], \[prsaw3\], and \[prsaw4\] for different $N$. When plotted against the scaling variable $r/N^{\nu_{{\rm SAW}}^{p_c}}$, the data are indeed found to nicely collapse onto a single curve, using our values for the exponent $\nu_{{\rm SAW}}^{p_c}$ reported in Table \[allnu\]. [@r lll ]{} $z^{p_c}\setminus d$ & 2 & 3 & 4\ SS [@Barat91] & $1.31\pm0.03$ &&\ EE [@Lam90] &$1.53\pm 0.05$ & &\ [@Ordemann00] & $ 1.565\pm0.005 $& $1.462\pm0.005$ &\ MC [@Woo91] & $1.459\pm0.003$ & &\ [@Ordemann00] & $1.456\pm0.005$ & $1.317 \pm 0.005$ &\ our results & $1.566\pm0.005$ & $1.459\pm0.005 $ & $1.340\pm0.005$\ $z\cdot p_c$ &1.564&1.460&1.333\ \[allmu\] [@ r l l l ]{} $\gamma_{{\rm SAW}}^{p_c} \setminus d$ & 2 & 3 & 4\ FL [@Roy87] & 1.384 & 1.379& 1.27\ EE [@Lam90] & $1.3\pm0.1 $ & &\ [@Ordemann00] & $1.34\pm0.05$ & $1.29\pm0.05$ &\ MC [@Lee89] & $1.31\pm0.03$ & $1.40 \pm 0.02$ &\ [@Ordemann00] & $1.26\pm0.05$ & $1.19\pm 0.05$ &\ our results & $ 1.350\pm 0.008$ & $1.269 \pm 0.008 $ & $1.250\pm0.008 $\ ![Averaged connectivity constant for SAWs on the backbone of percolation clusters in $d{=}2$ (triangles), $d{=}3$ (squares), $d{=}4$ (stars).[]{data-label="fugpc"}](fugpc.eps){width="8.2cm"} ![Disorder averaged number of SAWs configurations vs number of steps for SAWs on the backbone of percolation clusters in $d{=}2$ (triangles), $d{=}3$ (squares), $d{=}4$ (stars).[]{data-label="gammapc"}](gammapc.eps){width="8cm"} Let us now turn our attention to estimating the number of different possible SAW configurations ${\overline{ C_N }}$, defined by Eq. (\[scaling\]). First, let us note, that the fugacity or connectivity constant $z$ is obviously affected by introducing disorder into the lattice. For the case, when the SAW is not confined only to the percolation cluster, namely when averaging over all the clusters is performed, then $z^{p_c}{=}p_c z$ exactly. In Table \[allmu\] we present results of this estimate, taking values for $p_c$ from Table \[dim\]. However, since each existing bond on the infinite percolation cluster has a non-trivial (correlated) probability of occurrence, a similar argument cannot be applied to the SAWs confined to the infinite percolation cluster only. However, enumeration estimates [@Chakrabarti83] suggested $z^{p_c}{\simeq}p_c z$ up to $p_c$ for SAWs on the percolation cluster. It turns out, that this difference from linear dependence for incipient cluster is subtle and could hardly be detected. We have estimated $z_p$ as the averaged number of possibilities to perform the next step in the PERM procedure for SAWs on the backbone of percolation clusters (see Fig. \[fugpc\]); results are presented in Table \[allmu\]. In the analytical study of Ref. [@Lyklema84], it was argued that the exponent $\gamma$, governing the scaling of the number of SAW configurations, is not changed by the presence of disorder even at $p=p_c$. This was supported by an exact enumeration study [@Lam90]. However, this argument disagrees with results of studies [@Roy87; @Ordemann00; @Lee89], where averaging over single percolation clusters was performed and different values for $\gamma_{{\rm SAW}}^{p_c}$ were found. In Ref. [@Roy87] it was proven, using scaling arguments, that at $p=p_c$ the exponents $\gamma_{{\rm SAW}}$ will crossover to $\gamma_{{\rm SAW}}^{p_c}=\gamma_{{\rm SAW}}+d(\nu_{\rm SAW}^{p_c}-\nu_{{\rm SAW}})$ at $p=p_c$; the estimates based on this equality are given in the first row of Table \[allgamma\]. We obtained numerical estimates for $\gamma_{{\rm SAW}}^p$, studying the behavior of the quantity $$\ln \overline{ C_N }/N=\ln(A)/N+\ln(z^{p_c})+(\gamma_{{\rm SAW}}^{p_c}-1)\frac{\ln N}{N}, \label{ggg}$$ where $A$ is a constant. Figure \[gammapc\] shows these values for the backbone of percolation clusters in $d=2,3,4$. Estimates for $\gamma_{{\rm SAW}}^{p_c}$ are obtained by linear least-square fits (see Tables \[g2d\], \[g3d\], \[g4d\] in the Appendix). Our final results are presented in Table \[allgamma\]. Random walks on percolation clusters ==================================== ![\[3drw\] Averaged end-to-end distance vs number of steps in a double logarithmic scale of RWs on percolation clusters (crosses) and the backbone of percolation clusters (pluses) in $d=2$.](3drw.eps){width="7.8cm"} To simulate the diffusion process in a disordered medium, the picture of the “ant in the labyrinth", proposed by de Gennes [@deGennes76] is traditionally used. Here the walker (an “ant") starts at an arbitrary point on the diluted lattice and tries to move randomly to the nearest site. If the randomly chosen direction leads to an empty site, it moves and the steps increment by 1. If the chosen site is occupied by a defect (in our case, does not belong to the percolation cluster) it stays at the current position for this time step. The growth is stopped, if the total number of steps $N$ is performed, than the next trajectory is started to grow. After averaging the end-to-end distance over RW configurations on a single percolation cluster, the disorder average is carried out as in Eq. (\[av\]) over all constructed percolation clusters. Let us note that, in contrast to the SAW problem, discussed previously, the scaling behavior of RWs on a percolation cluster is different from that on its backbone. Let us remind, that statistics of long SAWs on percolation clusters is nevertheless determined by its backbone, since the walks, which visit the “dead ends" are eventually terminated after a certain number of steps. Simple random walks cannot be trapped in “dead ends", however, visiting these parts of a cluster will lead to some “slowing down" of the diffusion process in comparison with the behavior on the backbone where all the dead ends are removed. This is really confirmed by analyzing our results for the averaged end-to-end distance of random walks on a percolation cluster and its backbone (see Fig. \[3drw\]). ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/\overline{\langle r\rangle}$ in $d=2$ dimensions, left: percolation cluster, right: backbone of percolation cluster. Lattice size $L=400$, number of RW steps $N{{=}}180$ (squares), $N{{=}}160$ (triangles), $N{{=}}140$ (crosses).[]{data-label="prrw2"}](rozrwpc2d.eps "fig:"){width="6.4cm"} ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/\overline{\langle r\rangle}$ in $d=2$ dimensions, left: percolation cluster, right: backbone of percolation cluster. Lattice size $L=400$, number of RW steps $N{{=}}180$ (squares), $N{{=}}160$ (triangles), $N{{=}}140$ (crosses).[]{data-label="prrw2"}](rozrwback2d.eps "fig:"){width="6.8cm"} ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/\overline{\langle r\rangle}$ in $d=3$ dimensions, left: percolation cluster, right: backbone of percolation cluster. Lattice size $L=200$, number of RW steps $N{{=}}100$ (squares), $N{{=}}90$ (triangles), $N{{=}}80$ (crosses).[]{data-label="prrw3"}](rozrwpc3d.eps "fig:"){width="7.1cm"} ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/\overline{\langle r\rangle}$ in $d=3$ dimensions, left: percolation cluster, right: backbone of percolation cluster. Lattice size $L=200$, number of RW steps $N{{=}}100$ (squares), $N{{=}}90$ (triangles), $N{{=}}80$ (crosses).[]{data-label="prrw3"}](rozrwback3d.eps "fig:"){width="6.58cm"} ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/\overline{\langle r\rangle}$ in $d=4$ dimensions, left: percolation cluster, right: backbone of percolation cluster. Lattice size $L=50$, number of RW steps $N{{=}}60$ (squares), $N{{=}}50$ (triangles), $N{{=}}40$ (crosses).[]{data-label="prrw4"}](rozrwpc4d.eps "fig:"){width="6.8cm"} ![Disorder averaged distribution function $r\overline{P(r,N)}$ vs the scaling variable $r/\overline{\langle r\rangle}$ in $d=4$ dimensions, left: percolation cluster, right: backbone of percolation cluster. Lattice size $L=50$, number of RW steps $N{{=}}60$ (squares), $N{{=}}50$ (triangles), $N{{=}}40$ (crosses).[]{data-label="prrw4"}](rozrwback4d.eps "fig:"){width="7.1cm"} We have studied RWs both on the percolation cluster and its backbone, performing $10^7$ realizations on each cluster and average over $1000$ clusters in each space dimensions $d=2,3,4$. Estimates of scaling exponents $\nu_{{\rm RW}}^{p_c}$ and $\nu_{{\rm RW}}^{B}$, describing scaling of walks on percolation cluster and backbone, respectively, are obtained by linear least-square fitting and given in Table \[nurw\]. One can see, that the inequality $\nu_{{\rm RW}}^{p_c} < \nu_{{\rm RW}}^{B}$ holds, and that the quantitative difference between these two values grows with increasing the space dimension $d$. On the other hand, both values are [*smaller*]{} than the corresponding exponent $\nu_{{\rm RW}}=1/2$, governing scaling of random walks on the pure lattice: presence of disorder slows down the diffusion process. The reason for this subdiffusive behavior is intuitively clear: due to the presence of defects, the randomly walking particle returns back to already visited sites more often, thus its walking distance is shorter than on the pure lattice. This has also been observed in recent studies of less disordered deterministic fractals such as two-dimensional Sierpinski carpet composites [@Anh07]. The disorder averaged distribution function, defined in Eq. (\[prob\]), and rewritten in terms of the scaled variables $r/\overline{\langle r\rangle}$ as: $$r\overline{P(r,N)}\sim f(r/\overline{\langle r\rangle})\sim f(r/N^{\nu^{p_c}})$$ is shown in Figs. \[prrw2\], \[prrw3\] and \[prrw4\] for $d=2,3,4$, both for the cases of percolation cluster and backbone. When plotted against the scaling variable $r/N^{\nu^{p_c}}$, the data are indeed found to nicely collapse onto a single curve, using our values for the exponent $\nu_{{\rm RW}}^{p_c}, \nu_{{\rm RW}}^{B}$ reported in Table \[nurw\]. [@ r l l l l ]{} $\nu_{{\rm RW}}^{p_c}\setminus d$ & 2 & 3 & 4\ RS [@Sahimi83]& 0.356 & 0.285 &\ EE [@Majid84] & 0.349$\pm0.002$ &&\ [@Pandey83] & & 0.266$\pm0.01$ &\ analytic [@Alexander82] & 0.352 & 0.268 &\ MC [@Avraham82] & $0.352\pm0.006$ & $0.271\pm0.004$ &\ [@Havlin83] &$0.352\pm0.006$ &$0.271\pm0.004$&\ [@Argyrakis84] & $0.392\pm 0.007$ & $0.282\pm0.003$ &\ [@McCarthy88] & $0.348\pm0.009$ & $0.274\pm0.007$ &\ [@Lee00]& & & $0.222\pm0.007$\ our results &$ 0.353\pm 0.003 $&$ 0.273\pm 0.003$&$ 0.231 \pm 0.003$\ $\nu_{{\rm RW}}^{B}\setminus d$ &2&3&4\ analytic [@Bug86] & $0.371\pm0.001$ &&\ EE [@Hong84] & $0.372\pm0.005$ &&\ MC [@Mastorakos93] &$0.370\pm0.003$ &&\ our results &$0.372\pm 0.002 $&$ 0.306\pm0.002$ & $ 0.262\pm0.002$\ Conclusions =========== We studied the scaling behavior of simple random walks and self-avoiding walks on disordered lattices. Both models are of great interest: RWs provide a good description of diffusion processes, SAWs are successful in describing the universal properties of long flexible polymer macromolecules in a good solvent. We consider the case, when concentration $p$ of lattice sites allowed for walking equals the critical concentration $p_{c}$ and the incipient percolation cluster can be found in the system. Studying properties of percolative lattices, one encounters two possible statistical averages: in the first, one considers only percolation clusters with linear size much larger than the typical length of the physical phenomenon under discussion, the other statistical ensemble includes all the clusters, which can be found on a percolative lattice. In our study, we considered only the first case, being interested in random and self-avoiding walks on a percolation cluster, which has a fractal structure. In this regime, the critical behavior of both models belongs to a new universality class: scaling law exponents change with the dimension of the (fractal) lattice on which the walk resides. We performed numerical simulations of random and self-avoiding walks on percolation clusters and the backbone of percolation clusters on lattice sizes $L=400, 200, 50$ in space dimensions $d=2,3,4$, respectively. Our results bring about the estimates for critical exponents, governing the scaling behavior of the models. We found that the statistics of SAWs is governed by the same scaling exponent both on a percolation cluster and its backbone: since the walks, which visit the dead ends are eventually terminated after a certain number of steps, the walks that survive in the limit $N\to \infty$ on a percolation cluster are those confined within its backbone. For simple random walks, however, the picture is different: they cannot be trapped in “dead ends". However, visiting these parts of a cluster will lead to some “slowing down" of the diffusion process in comparison with the behavior on the backbone where all the dead ends are removed. We found that the inequality $\nu_{{\rm RW}}^{p_c} < \nu_{{\rm RW}}^{B}$ holds, and the quantitative difference between these two values grows with increasing space dimension $d$. The presence of disorder leads to a stretching of the trajectory of self-avoiding walks: the value of $\nu_{{\rm SAW}}^{p_c}$ is larger than the corresponding exponent on the pure lattice. However, the exponent $\nu_{{\rm RW}}^{p_c}$, governing scaling of random walks on percolative lattices is smaller than that on a pure lattice: presence of disorder slows down the diffusion process. This can be explained as follows: due to the presence of defects, the randomly walking particle returns back to already visited sites more often, thus its walking distance is shorter than on the pure lattice. Acknowledgments =============== Work supported in part by the German Science Foundation (DFG) through the Research Group FOR877. V.B. is grateful for support through the “Marie Curie International Incoming Fellowship" EU Programme and to the Institut für Theoretische Physik, Universität Leipzig, for hospitality. Appendix ======== To estimate the critical exponents $\nu_{{\rm SAW}}^{p_c}$, $\gamma_{{\rm SAW}}^{p_c}$, linear least-square fits with varying lower cutoff for the number of steps $N_{{\rm min}}$ are used in order to detect possible corrections to scaling. For estimating $\nu_{{\rm SAW}}^{p_c}$ we use linear fits for the averaged end-to-end distance $\ln({\overline{\langle r(N) \rangle}})=\ln(A)+\nu_{\rm SAW}^{p_c}\ln N$, and for $\gamma_{{\rm SAW}}^{p_c}$ we employ Eq. (\[ggg\]). Since this is an important aspect for assessing the quality of our final exponent estimates discussed in the main text, we have compiled in this Appendix these more detailed results in Tables \[2d\]-\[g4d\]. The $\chi^2$ value (sum of squares of normalized deviation from the regression line) divided by the number of degrees of freedom, DF, given in the last rows, serves as a test of the goodness of fit. [@ rlll]{} $N_{\rm min}$ & $\nu_{{\rm SAW}}^{p_c}$ & $A$ & $\chi^2/DF$\ 11 & 0.790 $\pm$ 0.005 & 0.829 $\pm$ 0.003 & 2.396\ 16 & 0.786 $\pm$ 0.005 & 0.841 $\pm$ 0.005 & 1.910\ 21 & 0.782 $\pm$ 0.004 & 0.847 $\pm$ 0.005 & 1.479\ 26 & 0.783 $\pm$ 0.003 & 0.842 $\pm$ 0.006 & 1.262\ 31 & 0.782 $\pm$ 0.003 & 0.840 $\pm$ 0.007 & 0.839\ \[2d\] [@rlll]{} $N_{{\rm min}}$ & $\nu_{{\rm SAW}}^{p_c}$ & $A$ & $\chi^2/DF$\ 11 & 0.668 $\pm$ 0.003 & 0.935 $\pm$ 0.004 & 2.269\ 16 & 0.669 $\pm$ 0.003 & 0.930 $\pm$ 0.004 & 2.054\ 21 & 0.669 $\pm$ 0.003 & 0.924 $\pm$ 0.004 & 1.345\ 26 & 0.667 $\pm$ 0.002 & 0.930 $\pm$ 0.006 & 0.743\ 31 & 0.668 $\pm$ 0.002 & 0.934 $\pm$ 0.008 & 0.844\ [@rlll]{} $N_{\rm min}$ & $\nu_{{\rm SAW}}^{p_c}$ & $A$ & $\chi^2/DF$\ 8 & 0.588 $\pm$ 0.002 & 1.025 $\pm$ 0.004 & 2.615\ 10 & 0.587 $\pm$ 0.002 & 1.023 $\pm$ 0.006 & 1.777\ 12 & 0.586 $\pm$ 0.003 & 1.021 $\pm$ 0.01 & 0.978\ 14 & 0.586 $\pm$ 0.003 & 1.031 $\pm$ 0.01 & 0.767\ \[4d\] [@rlll]{} $N_{\rm min}$ & $\gamma_{{\rm SAW}}^{p_c}$ & $A$ & $\chi^2/DF$\ 16 & $1.341 \pm 0.005$ & $1.219 \pm 0.003$ & 3.135\ 21 & $ 1.349 \pm 0.005$ & $1.189 \pm 0.003$ & 2.682\ 26 & $1.351 \pm 0.007$ & $1.168 \pm 0.002$ & 1.913\ 31 & $1.352 \pm 0.008$ & $ 1.172 \pm 0.002$ & 1.621\ 36 & $1.350 \pm 0.008 $ & $1.163 \pm 0.001$ & 0.704\ \[g2d\] [@rlll]{} $N_{\rm min}$ & $\gamma_{{\rm SAW}}^{p_c}$ & $A$ & $\chi^2/DF$\ 11 & 1.265 $\pm$ 0.004 & 1.82 $\pm$ 0.003& 2.767\ 16 & 1.268 $\pm$ 0.005 & 1.192 $\pm$ 0.003 & 2.135\ 21 & 1.270 $\pm$ 0.006 & 1.184 $\pm$ 0.003 & 1.968\ 26 & 1.267 $\pm$0.008 & 1.176 $\pm$0.002 & 1.513\ 31 & 1.269 $\pm$0.008 & 1.172 $\pm$ 0.002 $\pm$ & 0.762\ \[g3d\] [@rlll]{} $N_{\rm min}$ & $\gamma_{{\rm SAW}}^{p_c}$ & $A$ & $\chi^2/DF$\ 8 & 1.251 $\pm$ 0.005 & 1.77 $\pm$ 0.003& 1.767\ 10 & 1.252 $\pm$ 0.007 & 1.182 $\pm$ 0.003 & 1.135\ 12 & 1.250 $\pm$ 0.008 & 1.184 $\pm$ 0.003 & 0.968\ \[g4d\] [90]{} See e.g. M F Shlesinger and B West (ed ) [*Random Walks and their Applications in the Physical and Biological Sciences*]{} (AIP Conf Proc vol 109) (AIP New York 1984); F Spitzer [*Principles of Random Walk*]{} (Springer Berlin 1976) J des Cloizeaux and G Jannink [*Polymers in Solution*]{} (Clarendon Press Oxford 1990); P -G  de Gennes [*Scaling Concepts in Polymer Physics*]{} (Cornell University Press Ithaca and London 1979) J C Le Guillou and J Zinn-Justin 1980 Phys. Rev. B [**21**]{} 3976 B  Nienhuis 1982 Phys. Rev. Lett. [**49**]{} 1062 J C Le Guillou and J Zinn-Justin 1985 . Physique. Lett. [**46**]{} L127; 1988 J Physique [**50**]{} 1365 R Guida and J Zinn-Justin 1998 J. Phys. A [**31**]{} 8104 M N Rosenbluth and A W Rosenbluth 1955 J. Chem. Phys. [**23**]{} 356 N Madras and A D Sokal 1988 J. Stat. Phys. [**50**]{} 109 D MacDonald, D L Hunter, K Kelly, and N Jan 1992 J. Phys. A [**25**]{} 1429 D MacDonald, S Joseph, D L Hunter, L L Moseley, N Jan, and A J Guttmann 2000 J. Phys. A [**33**]{} 5973 B Li, N Madras, and A D Sokal 1995 J. Stat. Phys. [**80**]{} 661 S Caracciolo, M S Causo, and A Pelissetto 1998 Phys. Rev. E [**57**]{} R1215 S Havlin and D Ben Abraham 1987 Phys. Adv. [**36**]{} 695 R M Ziff 1994 Phys. Rev. Lett. [**72**]{} 1942 P Grassberger 1992 J. Phys. A [**25**]{} 5867 G Paul, R M Ziff, and H E Stanley 2001 Phys. Rev. E [**64**]{} 026115 P Grassberger 1986 J. Phys. A [**19**]{} [1681]{} C Moukarzel 1998 Int. J. Mod. Phys. C [**8**]{} 887 D Stauffer and A Aharony [*Introduction to Percolation Theory*]{} Taylor and Francis London 1992 M Sahimi and J Jerauld 1983 J. Phys. C: Solid State Phys [**16**]{} L1043 I Majid, D Ben-Avraham, S Havlin, and H E Stanley 1984 Phys. Rev. B [**30**]{} 1626 R B Pandey and D Stauffer 1983 Phys. Rev. Lett. [**51**]{} 527 S Alexander and R Orbach 1982 J. Phys. Paris Lett. [**43**]{} L625 D Ben-Avraham and S Havlin 1982 J. Phys. A [**15**]{} L691; S Havlin, D Ben-Avraham, and D Sompolinsky 1982 Phys. Rev. A [**27**]{} 1730 S Havlin and D Ben-Avraham 1983 J. Phys. A [**16**]{} L483 P Argyrakis and P Kopelman 1984 Phys. Rev. B [**29**]{} 511 J F McCarthy 1988 J. Phys. A [**21**]{} 3379 S B Lee and H Nikanishi 2000 J. Phys. A [**33**]{} 2943 A L Bug, G S Grest, M H Cohen, and I Webman 1986 J. Phys. A [**19**]{} L323 D Hong, S Havlin, and H Herrmann 1984 Phys. Rev. B [**7**]{} 4083 J Mastorakos and P Argyrakis 1993 Phys. Rev. E [**48**]{} 4847 I Webman 1981 Phys. Rev. Lett. [**47**]{} 1497 Y Gefen, A Aharony, and S Alexander 1983 Phys. Rev. Lett. [**50**]{} 77 R Rammal and G Toulouse 1983 J. Phys. Paris Lett. [**44**]{} L13 S Mukherjee, D Jacobs, and H Nakanishi 1995 J. Phys. A [**28**]{} 291 K  Kremer 1981 Z. Phys. B [**45**]{} 149 S  B  Lee and H  Nakanishi 1988 Phys. Rev. Lett. [**61**]{} 2022; S  B  Lee, H  Nakanishi, and Y Kim 1989 Phys. Rev. B [**39**]{} 9561 Y Kim 1990 Phys. Rev. A [**41**]{} 4554 K  Y  Woo and S  B  Lee 1991 Phys. Rev. A [**44**]{} 999 P  Grassberger 1993 J. Phys. A [**26**]{} 1023 S  B  Lee 1996 J. Korean. Phys. Soc. [**29**]{} 1 Y  Meir and A  B  Harris 1989 Phys. Rev. Lett. [**63**]{} 2819 P  M  Lam 1990 J. Phys. A [**23**]{} L831 H  Nakanishi and J  Moon 1992 Physica A [**191**]{} 309 M  D  Rintoul, J  Moon, and H  Nakanishi 1994 Phys. Rev. E [**49**]{} 2790 A  Ordemann, M  Porto, H  E  Roman, S  Havlin, and A  Bunde 2000 Phys. Rev. E [**61**]{} 6858 H  Nakanishi and S  B  Lee 1991 J. Phys. A [**24**]{} 1355 K Barat, S N Karmakar, and B K Chakrabarti 1991 J. Phys. A [**24**]{} 851 M Sahimi 1984 J. Phys. A [**17**]{} L379 R  Rammal, G  Toulouse, and J  Vannimenus 1984 J. Phys. Paris [**45**]{} 389 Y  Kim 1987 J. Phys. A [**20**]{} 1293 A  K  Roy and A  Blumen 1990 J. Stat. Phys. [**59**]{} 1581 A K Roy and B K Chakrabarti 1987 J. Phys. A [**20**]{} 215 A  Aharony and A  B  Harris 1989 J. Stat. Phys. [**54**]{} 1091 P  M  Lam and Z  Q  Zhang 1984 Z. Phys. B [**56**]{} 155 C von Ferber, V Blavatska, R Folk, and Yu Holovatch 2004 Phys. Rev. E [**70**]{} 035104 R H -K Janssen and O Stenull 2007 Phys. Rev. E [**75**]{} 020801 R V Blavatska and W Janke 2008 Europhys. Lett. [**82**]{} 66006 J Hoshen and R Kopelman 1976 Phys. Rev. E [**14**]{} 3438 M Aizenmann 1997 Nucl. Phys. B \[FS\] [**485**]{} 551 ; L N Shchur and T Rostunov 2002 JETP Lett. [**76**]{} 475 H J Herrmann, D C Hong, and H E Stanley 1984 J. Phys. A [**17**]{} L261 M Porto, A Bunde, S Havlin, and H E Roman 1997 Phys. Rev. E [**56**]{} 1667 S Havlin, R Nossal, B Trus, and G H Weiss 1984 J. Phys. A [**17**]{} L957 P Grassberger 1997 Phys. Rev. E [**56**]{} 3682 F T Wall and J J Erpenbeck 1959 J. Chem. Phys. [**30**]{} 634 H P Hsu, V Mehra, W Nadler, and P Grassberger 2007 J. Chem. Phys. [**118**]{} 444 M Bachmann and W Janke 2003 Phys. Rev. Lett. [**91**]{} 208105; J. Chem. Phys. 2004 [**120**]{} 6779 V Mehra and P Grassberger 2002 Physica D [**168**]{} 244 H Frauenkron, U Bastolla, E Gerstner, P Grassberger, and W Nadler 1998 Phys. Rev. Lett. [**80**]{} 3149 P Le Doussal and J Machta 1991 J. Stat. Phys. [**64**]{} 541 B K Chakrabarti, K Bhadra, A K Roy, and S N Karmakar 1983 Phys. Lett. [**93A**]{} 434 A J Guttmann and J Wang 1991 J. Phys. A [**24**]{} 3107 J W Lyklema and K Kremer 1984 Z. Phys. B [**55**]{} 41 B Nienhuis 1984 J. Stat. Phys. [**34**]{} 731 P G de Gennes 1976 La Recherche [**7**]{} 919 D H N Anh, K H Hoffmann, S Seeger, and S Tarafdar 2005 Europhys Lett [**70**]{} 109; D H N Anh, P Blaudeck, K H Hoffmann, J Prehl, and S Tarafdar 2007 J. Phys. A [**40**]{} 11453

ArXiv

--- author: - 'A.M. Hidalgo-Gámez' - 'D. Ramírez-Fuentes' - 'and J. J. Gónzalez' title: 'Spectroscopic observations of a sample of dwarf spiral galaxies. II- Abundance gradients' --- The oxygen gradient of four dS galaxies has been determined using abundances for several H[ii]{} regions determined with four different methods. The gradient slopes of the three non-barred galaxies in the sample are quite steep, larger than $-0.2$ dex/kpc, while the gradient of the barred galaxy is shallower, only $-0.1$ dex/kpc. Although these gradients are quite steep they are real, following all the galaxies the same trend. Moreover, the results obtained here agree with those marked by the late-type, non-dwarf spirals, particularly the relationship between the gradient and the absolute magnitude and the optical size for non-barred galaxies, and the surface density for barred ones. Introduction ============ The existence of differences in the chemical abundances along the galactocentric distance is a well known characteristic of spiral galaxies (see, among others, the works of Searle 1971 as well as Díaz 1989; Zaritsky et al. 1989; Walsh & Roy 1989; Vila-Costas & Edmunds 1992). Normally, the central regions exhibit higher abundances than those located at the outskirts. Such gradients, both nebular and stellar, have been obtained for our Galaxy (e.g. Edvardsson 2001 and references therein), but not without controversy (Fu et al. 2009). The gradients in metallicity are probably the results of billions of years of evolution. As the chemical elements are processed in the interior of the stars, such abundance gradients can be related to the gas mass fraction and the star formation rate in the disc of spiral galaxies (Phillips & Edmunds 1991) or to variations in the Initial Mass Function (Güsten & Mezger 1982). They also could be related with the yield if the closed-box model is considered. A constant yield might give variations in the gas fraction, and vice versa (J. Vílchez, 2010 private communication). Several investigations studied the dependence of the gradient with different properties of the galaxies, such as the morphological type, the absolute magnitude and the rotation velocity (e.g. Zaritsky et al. 1994; Vila-Costas & Edmunds 1992). The main conclusion of these investigations is that gradients in non-barred galaxies seem to be related to the mass surface density and to the morphological type (late-type galaxies have steeper gradients than early-types) while the central abundance are correlated with the galaxy mass. For barred galaxies the gradients seem to be related to the length of the bar compared to the size of the disc and to the ellipticity (Martin & Roy 1994). In addition, barred galaxies seem to show a flatter gradient than the non-barred ones (Pagel et al. 1979; Alloin et al. 1981) for all morphological types. Again, this is not without controversy, because other authors measured large gradients in barred galaxies (Martin & Roy 1994). The relation obtained by Zaritsky et al. (1994) between the gradient and the morphological type is quite interesting in the sense that it might indicate that H[ii]{} regions in galaxies of later morphological types, as Sm and Im, should have important differences in metallicity. A similar conclusion was held also by Kunth & Roy (1996), where they discussed the smoothening of the gradients due to the shear in spiral galaxies, while irregular galaxies, because of their lack of internal movements, might keep their local enrichments. This is contrary to observations, where no significant variations are found in the chemical abundances among the irregular galaxies but only nitrogen local enrichment by Wolf-Rayet stars (see Kobulnicky & Skillman 1997) and only very few galaxies show differences in the abundances larger than $1\sigma$ (Hidalgo-Gámez et al. 2001). There are several reasons for the rejection of the possible existence of metallicity gradients in irregular galaxies. The use of the semi-empirical methods in the abundance determination, the large uncertainties in the metallicity values, the small size of the galaxies, etc, have been given for the non acceptance of differences in the metal content throughout an irregular galaxy (e.g. Pilyugin 2001). A recent example is the Local irregular galaxy IC 10, for which Magrini & Goncalves (2009) obtained differences in the oxygen abundance as large as $0.6$ dex but they still said that ”there is no indications of radial gradients´´. Therefore, there is no irregular galaxy with an accepted gradient or variations in the oxygen abundance. On the contrary, there are several dozens of spiral galaxies where abundance gradients have been determined, some of them as small as $-0.01$ dex/kpc. The main reason argued for the acceptance of such small gradients is the large size of the spiral galaxies, most of the time larger than $10$ kpc. Therefore, although the gradient is small, it gives an important difference between the central part and the outskirts of a certain galaxy. In addition to this situation, there are low-mass, spiral galaxies which does not show any difference in the abundance among the Híi regions. The best studied one is NGC 1313 (Walsh & Roy 1997). Mollá & Roy (1999) concluded that the absence of oxygen gradient in NGC 1313 is due to its low mass, which resembles irregular galaxies. To our concern, NGC 1313 is the only late-type spiral galaxy with no gradient, and therefore unique. If a large population of spiral galaxies with no gradient are found they will be the transition between the normal late-type Sm with larger gradient, as suggested by Zaritsky et al. (1994), and the Im galaxies with no gradients. The main goal of the present investigation is to check if there are any other late-type, low-mass, spiral galaxy without a metallicity gradient. The abundance gradient of four dwarf spiral (dS) galaxies have been determined in order to check if their gradients increase with the morphological type, as suggested by the relationship by Zaristky et al. (1994) or, on the contrary, they got shallower approaching the distribution abundance in irregular galaxies. All the galaxies in the sample are of low mass, both total and gas, low luminosity and small size, very similar to the characteristics of NGC 1313. Moreover, their H[ii]{} regions are extended and their inclinations are not excessive. none of them have previous spectroscopic studies. For more details about the galaxies in the sample, the reader is referred to Hidalgo-Gámez et al. (AJ, submitted; hereafter paper I). In the next section, the determination of the chemical abundance gradients for each of the four galaxies in the sample is presented. In Section 3, a discussion on the accuracy of the values is given. The relationship between these gradients and other global parameters of the galaxies is also discussed. Brief conclusions are presented in Section 4. Gradient abundances in dwarf spiral galaxies ============================================ A total of $29$ H[ii]{} regions in four dS galaxies were studied in a previous paper (Hidalgo-Gámez et al. AJ submitted; hereafter paper I). Two of the galaxies, UGC 5242 and UGC 5296, have a small number of regions while for the other two, UGC 6205 and UGC 6377, the number of regions are larger, of the order of $10$. Although the number of H[ii]{} regions detected in these galaxies is much smaller than in other spiral galaxies, they agree with the observation in other dS galaxies (Hidalgo-Gámez 2005; Reyes-Pérez 2009). The gas mass of all of the galaxies under study is small, about 10$^9$ M$\odot$ (Hidalgo-Gámez, unpublished), and similar to the gas content that the low-mass, late-type spiral NGC 1313 (Walsh & Roy 1996). This galaxy, however is brighter but with a similar chaotic spiral structure than those studied here. Two observables are needed in order to determine the abundance gradient: the galactocentric distance of each H[ii]{} and its chemical abundance. In order to obtain a proper distance from the center of the galaxy to the H[ii]{} regions a zero point should be considered. This could be photometric (the place with the maximum surface brightness) or dynamical (the place with the lowest rotation velocity). In the present investigation, the first one is considered. From $V$ and $R$ images (obtained by one of the authors at the 1.5m telescope of San Pedro Mártir-OAN) a photometric center of the galaxy was determined. To identify the H[ii]{} regions, H$\alpha$ images of each galaxy were compared with the images from the acquisition camera except for UGC 6277 for which no H$\alpha$ image was available. Due to the low number of H[ii]{} regions in this galaxies this was an easy task. Only in UGC 6377, because the lack of the H$\alpha$ image, there might be some problems i n the identification. Once the H[ii]{} regions were identified, the H$\alpha$ images were compared with the $V$ and $R$ ones, in order to determine the photometric center in the H$\alpha$ image. A distance was obtained considering the pixel sizes of every CCD involved and the distance to the galaxy, which might be the greatest source of uncertainties. Inclination corrections for each galaxy were applied, using the values tabulated by NED at the $25$ mag arcsec$^2$. The abundances of the H[ii]{} regions studied here were determined in paper I. Here, a brief summary of the values are given but the reader is referred to paper I for details. In one H[ii]{} region, the forbidden oxygen line at 4363Å  is detected and, therefore the standard method, as described in e.g. Osterbrock 1989, can be used to determine the chemical abundances. In the other H[ii]{} regions, the semi empirical methods are needed. A total of four different semi empirical methods are used. The results are that the abundances are in the range between $8.6$ and $7.7$ dex for most of the H[ii]{} regions. For a study of the feasibility of the semi empirical methods the reader is referred to Hidalgo-Gámez & Ramírez-Fuentes (2009). In order to study as carefully as possible the gradients in these four galaxies, a gradient value will be determined with each set of abundances obtained in paper I: the R$_{23}$, the $P$, the N$2$, the N$3$ and the average abundances. In addition, a comparison among the values determined with each method will be done. This is very interesting because it is believe the bending of the slope observed in some galaxies, as M101 (Zaritsky 1992), is due to an artifact´s in the metallicity determination (Pilyugin 2003). As discussed in paper I, each method has its own advantages and drawbacks: the R$_{23}$ and $P$ used a larger number of H[ii]{} regions whereas only those H[ii]{} regions with the highest S/N are used the $N2$ and $N3$ methods. The gradients will be obtained from a least-square fitting to all the abundance values determined for the H[ii]{} regions for each galaxy. A more robust value of the slope could be obtained from a bi-variate fitting: first considering the galactocentric distance as independent and then treating the abundances as the independent variable (Elmergreen et al. 1996). The fitted slope and constant are a combination of the values for each single fitting. However, in this study a single square-fitting has been used because it provides a simple and quick comparison with other investigations. In this statistical approach the inclusion of one single odd point might lead to an important change in the slope and do not reflect the real tendencies (Vila-Co stas & Edmunds 1992). Therefore, and in addition to this mathematical value of the gradient, we have also considered the differences between the innermost and outermost regions (I/O in Table 1) in each galaxy as well as between the most and less metallic ones (L/M in Table 1). Finally, a gradient can be obtained when averaging all the abundances from all the H[ii]{} regions located in a certain range of distance from the center. Therefore, a single abundance is obtained for a single galactocentric distance. The advantage of this last determination is that the possible odd abundances or uncertainties in the galactocentric distance are smoothed out. We advance that the agreement among all the values of the slopes/gradients is remarkable for most of the galaxies. As said, two galaxies have a very small number of H[ii]{} regions detected and therefore, the gradient values determined might be not very reliable. Indeed, for UGC 5296 and UGC 5242 the total number of H[ii]{} regions is much smaller than the minimum number of data-points needed for a reliable gradient determination according to Dutiful & Roy (1996). There are only five H[ii]{} regions in UGC 5296 (Réyes-Pérez 2009) and about eight in UGC 5242 (Hidalgo-Gámez, in preparation). Therefore, we are aware that the gradient values for these galaxies are less confident than for the rest of the sample. Nevertheless, there are some other galaxies where a gradient is determined from only $4$ regions, as NGC 3351, NGC 5068 and NGC 4395 (Vila-Costas & Edmunds 1992). UGC 6205 -------- UGC 6205 is the galaxy with the largest number of H[ii]{} regions detected in our sample: a total of $11$ regions were studied in this investigation. The gradient determined with the semi empirical methods are shown in the first column of Table 1. Also, the values determined using the most and less metallic regions, as well as the most internal and external ones are presented in rows 6 and 7, respectively. Finally, the slope from the averaged distances is presented in row 8 (See below). If the three latter values of the gradients are similar to those obtained with the least-squared fitting might indicate the robustness of the determination. This is the situation for UGC 6205, with all the values of the gradients ranging from $-0.2$ to $-0.4$ dex/kpc. An average value of $-0.31$ dex/kpc with a dispersion of $0.03$ will be considered. Another remarkable results presented in Table 1 is the steepness of the gradient, being larger than the values obtained for most of the galaxies studied so far (e.g. see Table 4 in Vila-Costas & Edmunds 1992). Why? One might think that this galaxy shows very extreme abundance values, but the most external region, located at 3.1 kpc, has a metallicity of $7.7$ dex while the most internal regions have metallicities of $8.7$ dex. The difference is only of $1$ dex, which is not so large. There are other spiral galaxies with such difference in their abundances, or larger, as M81 (Garnett & Shields 1987). The point here is that M81 is very much larger than UGC 6205. So, UGC 6205 has the same difference in the abundance than other galaxies but the galactocentric distances of the H[ii]{} are smaller. Therefore, so large gradient. The next question to be addressed is which is the region responsible for such a steep gradient. One might think that those regions with low S/N are responsible for the slope because their abundance value might not be so reliable. This is not true for several reasons: firstly, the gradients obtained with the $N2$ and $N3$ methods, which included only those high S/N regions, give only slightly shallower gradients. Moreover, the values determined without the least-squared fitting are of the same order except when the gradient is determined from the most and less metallic regions which is of $-1.1$ dex/kpc, and it will be ignored because of its doubtful meaning. Another way to check the reliability of the steep gradient, is, as previously said, using an average abundance for different ranges of galactocentric distances: $0$-$0.6$ kpc, $0.6$-$1.2$ kpc, $1.2$-$1.8$ kpc, $1.8$-$2.4$ kpc, $2.4$-$3.0$ kpc, and $3.0$-$3.600$ kpc. As said, the variations of the abundance are smoothed and the low S/N regions will not be the responsible of the value of the gradient. These are shown in the eighth row of Table 1 and it is about $-0.3$ dex/kpc for this galaxy. Whatever the method is used for the gradient determination, will end up in a very high value. Then, it can be concluded that such value might be real. Figure  \[fig1\] shows the metallicity vs. the galactocentric radius, in kpc, for each of the five metallicity determinations. The errorbars in the metallicity and in the galactocentric distances are shown but the latter are smaller than the symbol. The solid line is the fitting using all the H[ii]{} regions. The fitting is, in general, very good with all the methods except for regions U05a1, U05c1, and U05b5 which have odd abundance values for their galactocentric distances. If those regions are not considered, the gradients do not change significantly. It can be noticed in figure  \[fig1\] that the gradient seems to change along the radius of the galaxy, with a plateau for radii up to $1.6$ kpc, whatever method is used for the abundance determination. This is motivated mainly because the high metallicity of regions U05c1 and U05a1, which as stood in paper I, they might be something more than simply H[ii]{} regions. In any case, an internal (r $<$ 1.4 kpc) and external gradients were determined without these three anomalous regions, and no differences were found. It can be concluded that the gradient in this small spiral galaxy is very steep, of about $-0.3$ dex/kpc, with no dependence on the metallicity method and no differences along the galactocentric distance. UGC 6377 -------- As mentioned in paper I, two of the nine H[ii]{} regions in this galaxy have abnormal abundances and they might be planetary nebulae instead of normal H[ii]{} regions. Therefore, they are not going to be considered in the determination of the gradients. We will use the other seven regions in order to obtain the gradient. The values from the $N2$ and $N3$ abundances have only two regions and they are located very close each other, thus no reliable values can be obtained. The values of the slopes are shown in the second column of Table 1. It can be seen that the $R_{23}$, the $P$ method and the average abundance gradients are very similar, from $-0.15$ to $-0.20$ dex/kpc. Also, those values are very similar to the slopes determined from the most external/internal regions, with values ranging from $-0.16$ to $-0.13$ dex/kpc. The slopes determined from the most/less metallic regions are larger, of the order of $-0.3$ dex/kpc. The main reason is that the less metallic region, U77c3, is not the most external one. The abundance differences in UGC 6377 are lower than for UGC 6205, only $0.8$ dex, but the distances considered are also smaller, only $2.4$ kpc. Therefore, the gradients are also quite large. A reason for such steep slopes are needed. Similarly as in UGC 6205, one might think that the low S/N regions are changing the slope. Only three of the seven H[ii]{} regions used are of high S/N. When only those three are used (with all the caution due to the low number of data-points) the gradient is $-0.11$ dex/kpc, which is more similar to other late-type spirals galaxies (see next section). Another way to check the influence of the low S/N regions is using an averaged metallicity value for certain distance ranges, as for UGC 6205. In this case, those regions with low S/N are not so critical. The distance ranges were $0$-$0.6$ kpc, $0.6$-$1.2$ kpc, $1.2$-$1.8$ kpc, and $1.8$-$2.4$ kpc. The slope in this case is of the order of $-0.23$ dex/kpc. The (weighted) mean gradient is of $-0.2$ dex/kpc, with a dispersion of $0.02$. Figure  \[fig2\] shows the plot of the metallicity vs. the galactocentric distance along with the fitting for all the metallicity methods for which a gradient has been determined. The fitting is very good for all the cases. From a close inspection of Figure  \[fig2\], one might say that there is a change in the slope of the gradient, being steeper for the internal regions than for the external ones. This situation has been observed in other galaxies such as M101 (Zaritsky 1992) or NGC 2403 (Garnett et al. 1999). Pilyugin (2003) has argued that such bends in the slope of the abundances are artifacts in the abundance determinations. This explanation is not valid here because the flattening of the slope is observed when the abundances are determined with the $P$ method. Considering only those regions most internal than $1.2$ kpc, the gradient is quite steep, of about $-0.35$ dex/kpc, while the value determined with the outer regions is almost flat. This situation is forced by the abundance of U77c1, which is the most external region but its abundance is high. Without this region, the outer gradient is of the order of $-0.15$ dex/kpc and the difference with the internal regions is not that high. Finally, we will discuss about the abundances of U05a1 and U05b1. They have very high abundances for their galactocentric distances. If the low abundance values are considered instead of the high branch ones, they both fit very well the gradient determined by the rest of the data-points. Actually, the value of the gradient does not change at all. So, as concluded in paper I, they are likely PNs embedded in low metallicity H[ii]{} regions. UGC 5296 -------- There are only four H[ii]{} regions for gradient determination in this tiny galaxy. One might think that no gradient can be determined with such a few number of data-points. But there are at least another three galaxies with values determined with only four metallicity determination. Nevertheless, the confidence of the gradient obtained is not as good as for the two galaxies previously studied. The H[ii]{} regions observed in UGC 5296 are located in a line from south to north. For each of these regions, two set of intensities are obtained, as explained in paper I. For those lines obtained from ALICE-MIDAS, the gradient with the $N2$ and $N3$ methods cannot be determined because only two regions are considered. Moreover, the slopes determined with $N2$ and $N3$ VISTA-abundances are positive. This is an indication that the VISTA nitrogen intensities are not very trustworthy. Therefore, no further consideration will be given to those abundances and gradients and only the $P$- and $R_{23}$-abundances determinations will be discussed. In order to obtain a more reliable determination of the gradient, a single value of the abundances is obtained for each H[ii]{} region by averaging the values from the VISTA data-set and the ALICE-MIDAS one. This can be done because the abundances between these two data-set are very similar. The gradients determined in this way are shown in column 3 of Table 1 and in Figure  \[fig3\]. The value from the average abundance is $-0.6$ dex/kpc, which is quite unrealistic as well as those determined from the $R{23}$ and $P$ methods which are even steeper. The reason for such a steep value is the large difference in metallicity among the internal and the external regions in this galaxy, $1.2$ dex, and the small galactocentric distance range, $2$ kpc. Although a gradient at fixed distance has little sense because it will be determined with only two regions, it is useful because it shows that the abundances change dramatically, from $8.5$ dex at the center to $7.7$ dex at the o utskirts. The slope derived using the more/less metallic regions is $-0.6$ dex/kpc while the value using the out most/inmost regions is $-0.3$ dex/kpc. These values are similar to those determined with the mathematical fitting. This is because U96a2 is probably over enriched as discussed in paper I. Such enrichment does not force the steep gradient, because a value of $-0.52$ dex/kpc is obtained without this region. One might think that region U56a1, with its very low metallicity, is forcing such steep gradient. Actually, the $R_{23}$ and $P$ gradients do not change when this region is not considered. On the contrary, the average metallicity gradient does shallow mainly due to the higher abundances for U96a4. Values as large as $-0.6$ dex/kpc are not acceptable because they are quite unrealistic. The value of the gradient adopted for this galaxy is $-0.4$ dex/kpc, as stated in Table 1. The dispersion is about $0.1$. Such gradient is large but similar to the one of UGC 6205. From the results of the averaged distances determination discussed above, it can be concluded there is a real, and large, change in the abundances of this galaxy. Such differences cannot be explained by differences in the SFR (Réyes-Pérez 2009). Finally, it is important to point out that a more accurate determination of the gradient will be very difficult, because there is only one more H[ii]{} region in the galaxy. In any case, more high quality data are needed in order to confirm these results. UGC 5242 -------- The number of H[ii]{} regions is also small in this barred galaxy, but for the same reasons as for UGC 5296, a gradient will be determined. Moreover, as being the only barred galaxy in our sample, it will be very interesting to know if the gradient is similar to the other barred late-type galaxies. The oxygen abundances can be determined for only five regions. Also, as for UGC 5296, there are two set of data for each region. Therefore, a total of eight different estimations of the gradient can be obtained for this galaxy. Those values determined with the ALICE-MIDAS abundances are very similar, but those obtained from the VISTA abundances are not. Two of them are positive. The best way to proceed is, just as in UGC 5296, to determine an average abundance for each H[ii]{} region from those obtained with ALICE-MIDAS and with VISTA data-set of intensities. In this case, the values of the gradient vary from $-0.006$ dex/kpc to $-0.18$ dex/kpc (see Table 1). These values a re slightly larger than those determined for other barred late-type galaxies. Again, the values of the slope obtained with the most/less external regions are very similar to those obtained with the fitting. On the contrary, the slope determined from the most/less metallic regions are very large. This value could not be realistic and the main reason for such a steep gradient is that the less metallic region, U42c2, is not the most external one. The differences in abundances between the regions is of almost $0.5$ dex but with less than $1$ kpc in separation. Moreover, the most metallic region has the lowest S/N. When another region is considered, a gradient of $-0.15$ dex/kpc is obtained, which is more similar to the rest of the values. Although it should be taken with care due to the small number of region involved, a metallicity at averaged distances can be obtained as well as a gradient. Values from $-0.02$ to $-0.14$ dex/kpc are obtained, which are only slightly smaller than those obtained with all the regions. An average gradient of $-0.17$ is obtained from all the values in Table 1, with a dispersion of $0.06$. Discussion ========== How accurate these gradients are? --------------------------------- In a investigation like this there could be a lot of sources of uncertainties. Some of them could be related to uncertainties in the distance determination to the galaxy itself. Normally, distances are very difficult to obtain when no primary candles are considered. This is the case for all the galaxies in this sample. The distance used here were those from Hidalgo-Gámez (2004), normally based on secondary indicators. Distance determinations can change the value of the gradient if the new determination differs greatly from the old one. In general, the distances provided by NED/NASA are very similar to those reported in Hidalgo-Gámez (2004) for the galaxies in the sample. The only important difference is on the distance of UGC 5296, and the gradient changes from $-0.6$ dex/kpc to $-0.45$ dex/kpc, meanwhile for the rest of the galaxies the differences in the gradient determination are smaller than the uncertainties. Therefore, the steep gradients in these galaxies are not due to wrong distance determinations. Also, the galactocentric distances might have influence on the slope of the abundance, but not in this case as indicated by the results of the gradient determined at fixed distances. Probably, the largest source of uncertainty in the slopes obtained in this investigation is the abundance determination. As they cannot be calculated with the standard method (e.g. Osterbrock 1989), the abundance value itself might be very uncertain. As described in paper I, four different semi empirical methods were used in the abundance determination. Moreover, a weighted-average value of the abundance was obtained for each region. Therefore, the gradients determined with such metallicity should be very reliable. Moreover, those regions which are suspected not to be a normal H[ii]{} region are not considered in the gradient determination, as in UGC 6377. There are some regions in UGC 6205, as U05b4 and U05c3, which were considered of high metallicity despite the small value of the log(N/O) ratio (section 3.2 in paper I). Nevertheless, there is not change in the slope because they are located at the intermediate part of the galaxy, and therefore their inclusion is not cru cial. As discussed in paper I, despite the uncertainties in the metallicity determination, it is not quite likely that a high metallicity region was misclassified as of low metallicity one. Another source of uncertainty might be the low number of measurements for the determination of the gradient. Dutil & Roy (2001) said that in order to obtain a robust measurement of the gradient at least $16$ data-points are needed. This is an impossible achievement for the dwarf spiral galaxies, mainly because most of them do not have such a large number of H[ii]{} regions. From the study of Reyes-Pérez (2009) and the H$\alpha$ images obtained for two dozens of dS galaxies by Hidalgo-Gámez, it can be concluded that less than $10$ dwarf spirals out of more than $100$ have more than $15$ H[ii]{} regions. In particular, all the regions were studied for UGC 6205 while only one and two more exist for UGC 5296 and UGC 5242, respectively (Reyes-Pérez & Hidalgo-Gámez, in preparation). Therefore, the results here should be taken with care because of the small number of data-points, but it has to be understood that there are no more regions to be used. Finally, it has to be said that the gradients determined for another 11 galaxies in the literature were determined with only five measurements or less (see table 4 in Vila-Costas & Edmunds 1992). Dutil & Roy (2001) also said that not only the slope is uncertain due to the poor sampling, but also the gradient might change when more data-points are added. This can be studied here comparing the gradients determined with the $R_{23}$ or the $P$ methods and those determined with the $N_2$ or $N_3$ ones, because the number of H[ii]{} regions involved are different, being the latter much lower. The best galaxy is UGC 6205 because the number of abundance determinations are not low even with the $N_2$ and $N_3$ methods. In Table 1 it could be seen that there are differences in the gradient if determined with $11$ measurements or with four, but they are small. One might think that this is the situation for both UGC 5296 and UGC 5242 in the sense that there are important differences in the gradients determined with the different methods. For these galaxies the number of data-points are the same with both methods and it is a problem of false nitrogen detections. Therefore, it can be concluded that none of the possible source of uncertainties have a real influence on the gradient of the abundances. A final criticism that can be made to the data presented here is that as the abundances are determined with few lines, they might be not very reliable. This should be always taken into account when working with these galaxies, as well as with other galaxies for which the abundances have been determined from the \[OIII\]/\[NII\] or the \[NII\]/H$\alpha$ ratios, in the local Universe (e.g. Roy et al. 1996) or for high $z$ galaxies (e.g. Schulte-Ladbeck et al. 2004). As can be seen from Table 1 the value of the gradients are, in general, very similar for each galaxy when the abundances are determined with different methods. Moreover, the values determined with the less/most metallic regions or the internal/external regions are also very similar to those determined from the least-squared fitting. Therefore, we think that, although the abundances for some particular regions can be somehow uncertain, the gradients obtained here are very robust. How real these gradients are? ----------------------------- There are few tests that can be done in order to see if the values obtained here are real or due to some artifacts related to the uncertainties in the abundance determinations. The first one is to determine the gradient, if any, in the \[OIII\]/H$\beta$ ratio. Differences in this ratio were detected in the pioneering work of Searle (1971). Actually, he considered that those were due to differences in the abundances instead of in the excitation. This is shown in Figure  \[fig5\] for those non-barred galaxies in the sample. Along with the data-points and their errorbars, the least-squared fitting is shown (solid line). Only UGC 6205 shows a positive, and steep, gradient (larger ionisation for the most external regions). The dotted lines show the locus where the data-points can be located if the line intensities change by their uncertainties. All but two regions lay inside these lines. On the contrary, UGC 6377 shows a very unexpected behaviour, with a negative and steep gradien t. This is due to the high ionisation of U77b2, and the low ionisation of U77a1, both with low S/N. Considering only those H[ii]{} regions with medium and high S/N, the gradient is positive and very steep, of $-0.53$ dex/kpc, which is shown by the dashed line. Finally, the excitation gradient is almost flat for UGC 5296 (0.11 dex/kpc), with a high value for the most internal region and a low excitation for U96a1. The other gradient that can be determined is the \[NII\]/H$\alpha$ ratio. The only galaxy for which is it meaningful is UGC 6205. The nitrogen gradient is also very steep in UGC 6205, of $-0.44$ dex/kpc. From both line ratios, it can be concluded that the oxygen gradients determined are real, especially for UGC 6205 for which both ratios give very steep gradients. Finally, a total gradient for non-barred dwarf spiral galaxies can be determined. Figure  \[fig6\]a shows the abundances, from the average values, vs. the optical radius, in kiloparsecs, for the three galaxies studied here while the average abundances at a fixed distance for all the H[ii]{} regions of the three non-barred galaxies is shown in figure  \[fig6\]b. Those regions which are not considered in the gradient determination for a particular galaxy are also not considered in the fitting here but they are shown for completeness. Although the dispersion is large, particularly at intermediate distance, the gradient is also very steep, of $-0.3$ dex/kpc. Figure  \[fig6\]c shows the same abundances but to the normalised distance (r/r$_{25}$). In this case, the slope of the fitting is the same, $-0.3$. Two important conclusions can be obtained from this figure. Firstly, the H[ii]{} regions of three dS galaxies are non-distinguishable in the plots, in the sense that all of them have the largest metallicity for the central regions and the lowest for the most external ones. This is more clear in figure  \[fig6\]c, with a very low dispersion. The second conclusion is that these three dS galaxies have an steep abundance gradient, being the gradient obtained from the abundances at a averaged distance, show in figure  \[fig6\]b, the most robust. Therefore, it seems that the oxygen gradient for dwarf late-type spiral galaxies might be larger than for their normal-size counterpart. A larger sample of dwarf spiral galaxies are needed for a more conclusive result. Is there any dependence of the gradient with other parameters? -------------------------------------------------------------- Vila-Costas & Edmunds (1992) found out that the gradient is related with the total mass of the galaxy for non-barred galaxies and any morphological type. They also obtained correlations with the absolute magnitude, when all the morphological types and barred galaxies are included. Also, Zaritsky et al. (1994) found out a good correlation between the gradient and the morphological type, with larger slopes for late spirals. Here, we are going to explore the relationship among the gradient and other characteristics for late-type spirals only. In order to do this, information on the optical radius, surface brightness, absolute magnitude and gas mass and density has been obtained for five late-type (Sd or later), non-barred spiral galaxies and four barred galaxies from different authors, as Webster et al. (1983), McCall et al. (1985), Zaritsky et al. (1994), van Zee et al. (1997) and Hidalgo-Gámez (2004) and references therein. None of these parameters except the absolute magnitude were studied by Vila-Costas & Edmunds (1992) or Zaritsky et al. (1994). The results are shown in Figures from  \[fig7\]a to  \[fig10\]a for non-barred galaxies and in Figures from  \[fig7\]b to  \[fig10\]b for barred ones. In these figures, the stars correspond to the galaxies in the literature and triangles represent the galaxies in this study (dwarf spiral galaxies). Also, the solid line is the fitting to th e non-dwarf galaxies sample while the dotted line is the fitting to all the galaxies, including the dwarf ones. The regression coefficients of the former are shown at the top-right corner of each figure. The largest value of the regression coefficient for non-barred galaxies are those for the M$_b$ and the optical radius, indicating that the gradient of non-dwarf, late-type spirals shows a trend with these two parameters. There is also a weak correlation with the gas mass (total r$_g$ = $-0.61$) but not trend at all for the gas surface density, with a regression coefficient of $-0.3$. When dwarf spirals are included, the regression coefficients do not change much, although the slopes became shallower for all the cases. There might be several reasons for the small regression coefficient in Figure  \[fig10\]a as a much smaller number of data-points. Also, as the values of the dS are determined averaging the gas mass over the area of a disc and not considering an specific disc model, there must be differences between the dS and the Sm gas surface density. In any case, the value obtained for UGC 6205 seems to be similar to the other five Sm galaxies, but not UGC 5296 where the surface density is much smaller. In order to fit the non-dwarf trends, both UGC 6205 and UGC 5296 might have lower gradients, between $-0.15$ and $-0.30$ dex/kcp. They are, in any case, larger than those values for non-dwarf galaxies. They both also seem to have very low content of the gas mass. Figures  \[fig7\]b to  \[fig10\]b show the same as Figures  \[fig7\]a to  \[fig10\]a, but for barred galaxies, with a total of four galaxies from the literature (NGC 4395, NGC 925, NGC 1313, and NGC 5068) and UGC 5242 (triangle). Now, the absolute magnitude (r$_g$ = $-0.93$) and the gas surface density (r$_g$ = $-0.73$) show a correlation with the gradient. The results presented here seem to be in agreement with previous investigations. The correlation of the gradients with both, the radius and the absolute magnitude might be related with the correlation with the morphological type. The late-type galaxies tend to be smaller and with larger magnitudes than earlier galaxies (Hidalgo-Gámez 2004; van der Bergh 2008). Moreover, barred galaxies tend to be larger than non-barred galaxies of the same M$_b$ (Hidalgo-Gámez 2004). The main difference resides in the gas density surface, with a strong correlation for barred galaxies, while none for non-barred. The reason for such difference is not well understood yet. It might be related with some dynamical processes of the bar acting on the gas mass distribution. Or it is just a problem of poor sampling with a small range in values of the gradient. Figures 7-10 can also be used to check the reliability of the gradient of the individual galaxies. Adopting the definition of dwarf spiral galaxy by Hidalgo-Gámez (2004) and Figures 7a and 8a, dwarf spirals might show slopes larger than $-0.15$ dex/kpc, which are in agreement with the results presented here. As said above, the values for UGC 6205 and UGC 5296 seem to be too large. Gradients of about $-0.2$ dex/kpc for UGC 6205 and between $-0.2$ and $-0.3$ dex/kpc for UGC 5296 might fit all the correlations presented here while, considering the uncertainties in the gradient, the value for UGC 6377 seems to be quite good. The fitting of the only barred galaxy in our sample, UGC 5242, is very good for the relationship with the M$_b$ and the M(H[i]{}), while the gas surface density is very large for its gradient. Therefore, it can be concluded that despite all the uncertainties and caveats, the gradient for these four particular dwarf spirals are really quite steep and th ey follow the same relationship as late-type spirals; that is, smaller, and less brighter galaxies have larger gradients, regardless of their gas mass. Conclusions =========== The abundance gradient for four dwarf spiral galaxies have been obtained, none of them previously studied. Contrary to expected, the three non-barred galaxies show very steep gradients, larger than $-0.2$ dex/kpc, while the barred galaxy UGC 4252 shows a shallower one, of only $-0.10$ dex/kpc. Therefore, it seems that barred, dwarf galaxies have smaller slopes than non-barred ones, as previously discussed. Although the gradients look very steep compared with the values for the Milky Way, they follow the trend defined by other late-type, non-dwarf galaxies. The gradients for UGC 6205 and UGC 5296 are, at least, of $-0.2$ and $-0.3$ dex/kpc, respectively. In order to obtain a conclusive answer, more late-type, dwarf and non-dwarf are needed. The increase of the slope with Hubble type has been reported before and it is confirmed with these new gradients of very small spiral galaxies. There are indications than the dwarfer the galaxy is, the steeper the gradient. This might be a very important conclusion. As concluded in Hidalgo-Gámez (2004) it seems that dwarf spiral galaxies do not share the same properties as their normal-size counterparts. Actually, their luminosity functions and star formation processes resemble more those for irregular galaxies (Reyes-Pérez 2009; Reyes-Pérez & Hidalgo-Gámez, in preparation). Therefore, it would be very interesting to ask why normal irregular galaxies, which are similar to dwarf Sm galaxies in many other aspects, do not have strong gradients in oxygen abundance? The authors thank J.M. V[í]{}lchez, E. Terlevich and A. Díaz for interesting discussions on the questions discussed on the paper. This investigation is part of the Master Thesis of Daniel Ramírez Fuentes. This investigation was supported by SIP20100225 and CONACyT CB2006-60526. This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. Aller, L.H. 1984, “Physics of Thermal Gaseous Nebulae”, (Dordrecht: Reidel) Alloin, D., Edmunds, M.G., Lindblad, P.O., & Pagel, B.E.J. 1981, A&A 101, 377 Díaz, A.I. 1989, in “Evolutionary Phenomena in Galaxies”, Ed: Pagel, B.E.J. et al. Cambridge University Press, pp 277 Dutil, Y. & Roy, J.-R. 2001, AJ, 122, 1644 Elmergreen, B.G., Elmergreen D.M., Salzer, J.J., & Mann, H. 1996, ApJ, 467, 579 Fu, J., Hou, J.L., Yin, J., & Chang, R.X. 2009, ApJ, 696, 668 Garnett, D.R. & Shields, G.A. 1987, ApJ, 317, 82 Garnett, D.R., Shields, G.A., Peimbert, M., Torres-Peimbert, S., et al. 1999, ApJ, 513, 168 Güsten, R. & Mezger, P.G. 1982, Vistas in Astronomy, 26, 159 Hidalgo-Gámez, A.M. 2004, Rev.Mex. A&A, 40, 39 Hidalgo-Gámez, A.M. 2005, in “Starburst:from 30 Doradus to Lyman break galaxies” Ed: R. de Grijs & R.M. Gonzalez-Delgado, ApSSL, 329, 23 Hidalgo-Gámez, A.M., Masegosa, J., & Olofsson, K. 2001, A&A, 369, 797 Hidalgo-Gámez, A.M. & Olofsson. K. 2002, A&A, 389, 836 Hidalgo-Gámez, A.M. & Ramírez-Fuentes, D. 2009, AJ, 137, 169 Hidalgo-Gámez, A.M., Ramírez-Fuentes, D., & González, J.J., AJ submitted (paper I) Kobulnicky, H.A., & Skillman, E.D. 1997, ApJ, 489, 636 Kunth, D., & Roy, J.-R. 1996, ApJ, Kennicutt, R.C.Jr., Bresolin, F., & Garnett, D.R. 2003, ApJ, 591, 801 Magrini, L. & Goncalvez, D.R. 2009, MNRAS 398, 280 Martin, P. & Roy, J.-R. 1994, ApJ, 424, 599 Molla, M. & Roy, J.-R. 1999, ApJ 514, 781 Osterbrock, D.E., 1989, “Astrophysics of Gaseous Nebulae and Active Galactic Nuclei”, (University Science Books, Mill Valley, CA) Pagel, B.E.J., Edmunds, M.G., Blackwell, D.E., Chun, M.S., & Smith, G. 1979, MNRAS, 189, 95 Pilyugin, L.S. 2001, A&A 374, 412 Pilyugin, L.S. 2003, ApSS, 284, 775 Phillips, S. & Edmunds, M.G. 1991, MNRAS, 251, 84 Reyes-Pérez, J. 2009, Bachelor Thesis, Instituto Politécnico Nacional Roy, J.-R-. Belley, J., Dutil, Y., & Martin, P. 1996, ApJ, 460, 284 Schulte-Ladbeck, R.E., Rao, S.M., Drozdovsky, I.O., Turnshek, D.A. et al. 2004, ApJ, 600, 613 Searle, L. 1971, ApJ, 168, 327 van den Bergh, S. 2008, A&A, 490, 97 van Zee, L., Haynes, M.P., & Salzer, J.J. 1997, AJ, 114, 2479 Vila-Costas, M.B. & Edmunds, M.G. 1992, MNRAS, 252, 121 Walsh, J.R. & Roy, J.-R. 1997, MNRAS, 288, 726 Webster, B.L., & Smith, M.G. 1983, MNRAS, 204, 743 Zaritsky, D. 1992, ApJL, 390, 73 Zaritsky, D., Elston, R., Hill, J.M. 1989, ApJ, 97, 97 Zaritsky, D., Kennicutt, R.C.Jr., Huchra, J.P. 1994, ApJ, 420, 87 [Method]{} [UGC 6205]{} [UGC 6377]{} [UGC 5296]{} [UGC 5242]{} ----------------------- ------------------- ------------------- ------------------- -------------------- [Z$_{23}$]{} [-0.4]{} [-0.15]{} [-0.8]{} [-0.06]{} [Z$_P$]{} [-0.3]{} [-0.16]{} [-0.7]{} [-0.006]{} [Z$_{N2}$]{} [-0.2]{} [ - ]{} [-]{} [-0.2]{} [Z$_{N3}$]{} [-0.2]{} [ - ]{} [-]{} [-0.14]{} [Z$_{ave}$]{} [-0.3]{} [-0.2]{} [-0.6]{} [-0.1]{} [L/M metallic]{} [-0.4]{} [-0.3]{} [-0.6]{} [-0.6]{} [I/O regions]{} [-0.4]{} [-0.16]{} [-0.33]{} [-0.15]{} [averaged distance]{} [-0.3]{} [-0.25]{} [-0.5]{} [-0.07]{} [gradient]{} [-0.3$\pm$0.03]{} [-0.2$\pm$0.02]{} [-0.43$\pm$0.1]{} [-0.17$\pm$0.06]{} : Gradient abundance of the galaxies here. In columns 2, 3, 4 and 5 the name of the galaxy is given, while column 1 presents the different methods used in the gradient determination. The first row gives the gradients when the abundance is obtained with the $R_{23}$ method. The gradients determined with the $P$ abundances are presented in the second row. Rows 3 and 4 give the gradients obtained with the $N2$ and $N3$ methods, respectively. Row 5 shows the gradient obtained with the average value of the metallicity. Also, the difference in abundance and in distance between the most and the less metallic (L/M) regions and between the most internal and external regions (I/O) in the sample are presented in rows 6 and 7. Finally, the gradient obtained when all the abundance of all regions at a distance bin are averaged is presented in row 8 while the preferred value of the gradient is presented in row 9 with a dispersion value from all of the gradient determinations. All the values are in dex/kpc. See text for more details. ![The abundances along the galactocentric radius for UGC 6205, determined with the four methods described in paper I. From top to bottom are the R$_{23}$, the $P$ method, the $N3$, the $N2$ and the mean metallicity from these four methods (See paper I for details on each method and the abundance determination). The solid line is the least-squared fitting to all the H[ii]{} regions. Errobars in the abundances are also shown.[]{data-label="fig1"}](fig1.ps){width="14cm"} ![The abundances along the galactocentric radius for UGC 6377. Symbols and lines as in figure  \[fig1\]. The abundances determined with the $N2$ and the $N3$ methods are not shown due to a low number of data-points (see text). The high-metalliticy abundances of U77a1 and U77b1 are plotted in order to see if they fits the gradient, although they were not considered in the abundance gradient determination. []{data-label="fig2"}](fig2.ps){width="14cm"} ![The abundances along the galactocentric radius for UGC 5296. Only those abundances determined with the $R_{23}$ (top panel), the $P$ (central panel) methods or the average abundance (bottom panel) are shown, while the abundances determined with the $N2$ and the $N3$ methods are not shown due to a low number of data-points (see text for details). Symbols and lines as in figure  \[fig1\].[]{data-label="fig3"}](fig3.ps){width="14cm"} ![The abundances along the galactocentric radius for UGC 5242. Symbols and lines as in figure  \[fig1\]. Again, only the abundances determined with the $R_{23}$ (top panel), the $P$ (central panel) methods and the average abundance (bottom panel) are shown.[]{data-label="fig4"}](fig4.ps){width="14cm"} ![The excitation gradients for the non-barred galaxies in the sample. The solid line is the least-squared fitting, while the dotted one in the top panel comprises the locus where the data-points can be located. The dashed line in the central panel is the fitting to the high S/N regions in UGC 6377 (see text for details). []{data-label="fig5"}](fig5.ps){width="14cm"} ![The abundance gradient for all H[ii]{} regions for the non-barred galaxies in the sample. In the top panel (a)) the distance is in kpc while the gradient to a normalised distance is shown in the bottom panel. Figure 6b) shows the abundances vs. a fixed distance for all the data-points in the non-barred galaxies.[]{data-label="fig6"}](fig6.ps){width="14cm"} ![The relationship between the absolute magnitude and the gradients for non-barred (a) and barred (b), late-type galaxies. Stars correspond to normal-size Sm spirals while triangles stand for dwarf spirals. The solid line is the fitting to the Sm spirals only, while the dotted line is the fitting to all the data-points in the plot. []{data-label="fig7"}](fig7.ps){width="14cm"} ![The relationship between the optical radius and the gradient for non-barred (panel a) and barred (panel b), late-type galaxies. Symbols and lines as in figure  \[fig7\].[]{data-label="fig8"}](fig8.ps){width="14cm"} ![The relationship between the gas mass and the gradients for non-barred (panel a) and barred (panel b), late-type galaxies. Symbols and lines as in figure  \[fig7\].[]{data-label="fig9"}](fig9.ps){width="14cm"} ![The relationship between the gas surface density and the gradients for non-barred (panel a) and barred (panel b), late-type galaxies. $\Sigma_g$ for the dS galaxies was obtained as the gas mass divided by the area of a disc. Symbols and lines as in figure  \[fig7\].[]{data-label="fig10"}](fig10.ps){width="14cm"}

ArXiv

--- author: - | G. Chanakya, Pradeep Kumar\ Numerical Experiment Laboratory\ (Radiation & Fluid Flow Physics)\ School of Engineering\ Indian Institute of Technology Mandi\ Mandi 175005, India.\ Email: [email protected] bibliography: - 'mybibfile.bib' title: Effects of Diffuse and Collimated Beam Radiation on a Symmetrical Cooling Case of Natural Convection --- Abstract {#abstract .unnumbered} ======== In the present work, the effects of diffuse and collimated radiation on the symmetrical cooling case of natural convection in a two-dimensional cavity heated from the bottom have been investigated, numerically. The collimated beam radiation feature has been developed in OpenFOAM framework and other libraries of fluid flow and heat transfer have been combined together for a coupled simulation of fluid flow and heat transfer. The cavity is convectively heated from the bottom with heat transfer coefficient of 50 $W/m^2$K and free stream temperature 305 K, while both vertical walls of cavity are isothermal at temperature of 296 K. The top wall is adiabatic and all walls are opaque for radiation heat transfer. For collimated case, a small semitransparent window of non-dimensional width of 0.05 at height of 0.7 has been created on the left wall and a collimated irradiation of value 1000 $W/m^2$ at an angle of $45^0$ is applied on this semitransparent window. The study has been performed in two stages, first, the effect of diffuse radiation on the natural convection has been observed and then collimated beam is passed through the semitransparent window. The results reveal that the diffuse radiation has little effect on the dynamics of two rolls inside the cavity, however, collimated beam irradiation changes the dynamics of two rolls significantly and also the heat transfer characteristics. This further changes with the optical thickness of the fluid. The left vortex is bigger in size than the right vortex for collimated beam in transparent fluid, whereas, reverse trend is seen for collimated beam in non-zero optical thickness of the fluid. The size of the left vortex increases with the increase of optical thickness of the fluid. The heat transfer reversal happens at the zone of collimated beam incident on the bottom wall for collimated beam in transparent medium, whereas, this does not happen for participating medium. keywords {#keywords .unnumbered} ======== Semitransparent Window; Natural Convection; Collimated Beam; Symmetrical Cooling; Irradiation; Bottom Heating; Introduction ============ The study of natural convection in a square/rectangular cavities combined with radiation provides the fundamental insight of fluid flow and heat transfer for many engineering applications. This is the prime reason for multi-physics problems in a simple domain that became interest to researchers from last few decades. The natural convection in a square/cylindrical enclosures with heating from the bottom and cooled from the sides, with different inclination angles were analysed by many researchers [@Torrance; @Calcagani; @Ganzorolli; @Aydin] and the formation of two symmetrical vortices were observed. While, the single vortex formation has been observed in the inclined cavity case [@Acharya]. Webb and Viskanta [@Webb] experimentally studied the radiation induced buoyancy flow in a rectangular enclosure which was irradiated from a side. They have observed the formation of thin thermal boundary layers along the vertical walls where the flow structure lost the centrosymmetry characteristics. They also developed a theoretical model for the prediction of radiative heating of the fluid and motion induced due to buoyancy. Mezrhab et al. [@Mezrhab] and Sun et al. [@Hua] had numerically studied the combined natural convection and radiation for a centrally located hot square body in a cavity and observed that the radiation homogenized the temperature inside the cavity, whereas emissivity of the surfaces have affected the fluid flow characteristics inside the cavity. This was due to strengthening of recirculating zones which resultant into the stabilizing of flow fields due to surface radiation exchange. Mukul et al. [@Mukul] have studied the critical assessment of interaction of thermal radiation with natural convection in a square enclosure with different geometric configurations using three different approaches, like, incompressible (with Boussinesq approximation), purely compressible and quasi-incompressible (i.e low Mach-number approximation) approach. It was observed that the presence of corners had affected the heat transfer locally. Kumar and Eswaran [@Kumar] investigated the effect of radiation on natural convection in two slanted cavities of angles $45^0$ and $60^0$ and reported that radiation effect was more pronounced for cavity with slanted angle $60^0$. Saravanan and Sivaraj [@Saravanan] had reported the results of natural convection with thermal radiation for a non-uniformly heated thin plate placed at horizontal and vertical directions in a cavity. They noticed that the overall heat transfer had increased for the horizontal placement of this plate while it has decreased for the vertical placement. The natural convection with the volumetric radiation in a square cavity was studied by Mondal and Mishra [@Bittagopal] where they used lattice Boltzmann method for the fluid flow and finite volume method for radiative transfer equation (RTE). It was observed that scattering coefficient had no significant effect on the stream lines while extinction coefficient had a significant effect on the isothermal lines distribution. The performance of DOM, FVM, P1, SP3 and P3 methods for RTE in a two-dimensional absorbing/emitting medium were studied by Sun et al. [@Yujia]. The Monte Carlo method had been used as a benchmark solution for comparison of performance of above different method of RTE. Xing et al. [@Yuan] have studied the natural convection with heated circular, elliptical, square and triangular geometries inside cylindrical enclosure. The effect of surface radiation was incorporated and concluded that the presence of corners and larger upper space had a major effect on the heat transfer performance. In all the above works diffuse radiation was considered, however, little work is available on collimated beam radiation, like, work by Anand and Mishra [@Anand] and Ben and Dez [@Ben], both the work had tried to capture the collimated beam bending phenomenon in a graded refractive index medium by discrete ordinate method for RTE. The numerical study of a collimated beam in the varying refractive index medium was studied by Ilyushin [@Ya]. Few authors have also studied the short pulse collimated irradiation phenomenon [@Anil; @Rath]. From the above literatures, it has been observed that there is not much work performed for the effects of collimated beam radiation on natural convection. The collimated beam radiation has numerous applications like solar heating of room, solar cavity receiver in solar thermal power plants, optics etc. The collimated beam radiation feature had been developed and integrated with libraries of other fluid flow and heat transfer models in OpenFOAM [@openfoam2017open] framework and an application had also been developed to simulate natural convection with collimated beam irradiation. In the present work, the effect of diffuse and collimated beam radiation has been investigated in symmetrical cooling case of natural convection in a two-dimensional cavity heated at the bottom. The manuscript is organized as follows: The next section deals with the problem statement and mathematical modeling and its numerical procedure has been described in section 3. The results are mostly explained in non-dimensional numbers and hence these non-dimensional numbers are expressed in section 3.1, followed by validation in section 4 and grid independent test study in section 5. The section 6 elaborates results and discussion and finally the present work is concluded in section 7. Problem Statement ================= ![Schematic diagram for the study of collimated irradiation in a cavity[]{data-label="prob_def"}](problem_def.png){width="6cm"} Figure \[prob\_def\] depicts a square cavity (L=H) which has a semitransparent window of a non-dimensional width 0.05 at non-dimensional height of 0.7 on the left wall. Both the vertical walls as well as semitransparent window are isothermal at temperature of 296 K, whereas bottom wall is convectively heated with free stream temperature of 305K and heat transfer coefficient of 50 $W/m^2$K. The top wall is subjected to adiabatic boundary and all four walls of the cavity are opaque to the radiation. The window is treated as opaque for diffuse radiation analysis whereas semitransparent for collimated radiation. Collimated beam has been passed through semitransparent window at an angle of $45^0$. Natural convection is established inside the cavity due to temperature difference. The buoyancy force inside the cavity corresponds to Rayleigh number $10^5$ based on the temperature difference between free stream and vertical wall temperature and the Prandtl number of fluid is 0.71. The Euclidean axis are along the horizontal and the vertical walls of cavity and origin is at lower left corner of the cavity. The gravity force act in the negative Y direction. Mathematical modelling and Numerical procedures =============================================== The following assumptions have been considered for the mathematical modelling of the problem described below: 1. Flow is steady, laminar, incompressible and two dimensional. 2. Flow is driven by buoyancy force that is modeled by Boussinesq approximation. 3. The thermophysical properties of fluid is constant. 4. The fluid may or may not participate in radiative heat transfer. 5. The refractive index of the medium is constant and equal to one. 6. The fluid absorbs and emits but does not scatter the radiation energy. Based on the above assumptions the governing equations in the Cartesian coordinate system are given by $$\label{mass:equN} \frac{\partial u_i}{\partial x_i} = 0$$ $$\label{momentum} \frac{\partial u_i u_j}{\partial x_j}=-\frac{1}{\rho}\frac{\partial p}{\partial x_i} + \nu\frac{\partial^2u_i}{\partial x_j\partial x_j}+g \beta_{T}(T-T_{c})\delta_{i2}$$ $$\label{energy} \frac{\partial u_jT}{\partial x_j} = \frac{k}{\rho C_p}\frac{\partial^2T}{\partial x_j\partial x_j} - \frac{1}{\rho C_p} \frac{\partial q_{R}}{\partial x_i}$$ where $\frac{\partial q_{R_{i}}}{\partial x_{i}}$ is the divergence of the radiative flux, which can be calculated as $$\begin{aligned} \frac{\partial q_{R_{i}}}{\partial x_{i}}=\kappa_{a}(4\pi I_b-G)\label{div_eq}\end{aligned}$$ where $\kappa_{a}$ is the absorption coefficient, $I_b$ is the black body intensity and $G$ is the irradiation which is evaluated by integrating the radiative intensity ($I$) in all directions, i.e., $$\begin{aligned} G=\int_{4\pi} I d\Omega\end{aligned}$$ The intensity field inside the cavity can be obtained by solving the following radiative transfer equation (RTE) $$\label{equN:radiation_1} \frac{\partial I(\bf \hat{r},\bf \hat{s})}{\partial s}=\kappa_{a}I_{b}(\bf\hat{s})-(\kappa_{a}) \text{I} (\bf\hat{r},\bf \hat{s})$$ Where $\bf \hat{r},\bf \hat{s}$ is position and direction vectors. Whereas s is path length. The Navier-Stokes equation and temperature equations are subjected to following boundary conditions 1. *Flow boundary condition* 2. Cavity walls: u=v=0 3. *Thermal boundary conditions* 4. Left wall at x=0; T= 296K 5. Right wall at x=1; T=296K 6. Bottom wall y=0 ; $q_{conv}= h_{free} (T_{free}-T_{w})$ 7. Top wall at y=1; $q_{c}+q_{r}=0$ 8. where $q_c=-k \frac{\partial T}{\partial n}$ and $ q_{r}=\int_{4\pi}I({\bf r_w})({\bf \hat{n}\cdot\hat{s}})\mathrm{d}\Omega $ The radiative transfer equation (\[equN:radiation\_1\]) is subjected to following boundary condition, all cavity wall (assumed black wall) except semitransparent window $$\begin{aligned} \noindent I({\bf r_w,\hat{s}})=\epsilon_w I_b({\bf r_w})+\frac{1-\epsilon_w}{\pi}\int_{\bf \hat{n}\cdot\hat{s}>0}I({\bf r_w,\hat{s}})|{\bf \hat{n}\cdot\hat{s}}|\mathrm{d}\Omega.\nonumber\\ \mbox{for}~~{\bf \hat{n}\cdot\hat{s}}<0 \label{rte:bound2}\end{aligned}$$ where $\hat{n}$ is the surface normal and the emissivity of all walls is considered to be 1. The semitransparent window is subjected to collimated irradiation $(G_{co})$ of value 1000 $W/m^2$. The boundary condition for RTE on semitransparent window is $$\begin{aligned} \noindent I({\bf r_w,\hat{s}})=I_{co}({\bf r_w,\hat{s}}) \delta (\theta-45^{0})+\epsilon_w I_b({\bf r_w})+\frac{1-\epsilon_w}{\pi}\int_{\bf \hat{n}\cdot\hat{s}>0}I({\bf r_w,\hat{s}})|{\bf \hat{n}\cdot\hat{s}}|\mathrm{d}\Omega.\nonumber\\ \mbox{for}~~{\bf \hat{n}\cdot\hat{s}}<0 \label{rte:bound3}\end{aligned}$$ where $\delta (\theta-45^{0})$ is Dirac-delta function, $$\delta(\theta-45^{0})= \begin{cases} 1, & \text{if } \theta=45^{0},\\ 0, & \text{if } \theta \neq 45^{0}. \end{cases}$$ $I_{co}$ is intensity of collimated irradiation and calculated from the irradiation value as below $$I_{co}=\frac{G_{co}}{\mathrm{d}\Omega}$$ Where $d\Omega$ is the collimated beam width. In the current work, the solid angle of discretized angular space (Fig. \[angular\]) in collimated direction is considered as beam width of the collimated beam. The pictorial representation of diffuse emission and reflection and collimated beam radiation from the wall is shown in Fig. \[Diff\_ref\_wall\]. The collimated feature has been developed in OpenFOAM framework, an open source software and coupled with other fluid and heat transfer libraries. The combined application has been used for numerical simulation. The OpenFOAM uses the finite volume method (FVM) to solve the Navier-Stokes and the energy equations. The FVM integrates an equation over a control volume (Fig. \[2d\_fvm\]) to convert the partial differential equation into a set of algebraic equations in the foam $$a_{p}\phi_{p}= \sum_{nb}a_{nb}\phi_{nb}+S$$ where $\phi_{p}$ is any scalar and $a_{p}$ is central coefficient, $a_{nb}$ coefficients of neighbouring cells and S is the source values, whereas, RTE eq (\[equN:radiation\_1\]) is converted into a set of algebraic equations by double integration over a control volume and over a control angle. The set of algebraic equations are solved by Preconditioned bi-conjugate gradient (PBiCG) and the details of the algorithm can be found in the book by Patankar [@patankar] and Moukalled [@Moukalled]. In the present simulation, linear upwind scheme which is second order accurate has been used to interpolate face centred value. The linear upwind scheme is given mathematically as $$\phi_{f}= \begin{cases} \phi_{p}+ \nabla\phi \cdot \nabla r, & \text{if} f_{\phi} > 0,\\ \phi_{nb}+ \nabla\phi \cdot \nabla r, & \text{if} f_{\phi} < 0. \end{cases}$$ and $f_{\phi}$ is the flux of the scalar $\phi$ on a face (Fig \[2d\_fvm\]). [4cm]{} ![Pictorial representation of (a) typical volumetric cell in two dimensional case for finite volume method and (b) Virtual angular discreatization to obtain directions for radiative transfer equation []{data-label="grid_space_ang"}](2D_mesh.png "fig:"){width="5cm"} [5cm]{} ![Pictorial representation of (a) typical volumetric cell in two dimensional case for finite volume method and (b) Virtual angular discreatization to obtain directions for radiative transfer equation []{data-label="grid_space_ang"}](angular.JPG "fig:"){width="6cm"} [4cm]{} ![Pictorial representation of (a) Diffuse reflection of a incident ray and diffuse emission due to wall temperature on opaque wall (b) Diffuse emission and collimated transmission from a semitransparent window[]{data-label="Diff_ref_wall"}](Diff_BC.png "fig:"){width="5.5cm"} [5cm]{} ![Pictorial representation of (a) Diffuse reflection of a incident ray and diffuse emission due to wall temperature on opaque wall (b) Diffuse emission and collimated transmission from a semitransparent window[]{data-label="Diff_ref_wall"}](coli_diff.png "fig:"){width="5cm"} Non-dimensional Parameters -------------------------- The OpenFOAM simulation produces the results in dimensional quantities. To explain the results in more general form, the simulated results are converted into non-dimensional parameters. The scales for length, velocity, temperature, and conductive and radiative fluxes are L, u$_o$, (T$_{free}$-T$_{c}$), $\kappa$(T$_{free}$-T$_{c}$)/L and $\sigma$T$_{free}^{4}$ respectively, where $u_{o}=\sqrt {L g \beta (T_{free}-T_{c})}$ is convective velocity scale. The non-dimensional quantities and parameters involved in the present problem are as follows, $$\begin{aligned} U =\frac{u}{u_{o}} \hspace{0.5cm} V=\frac{v}{u_{o}} \hspace{0.5cm} \hspace{0.5cm} X =\frac{x}{L} \hspace{0.5cm} Y=\frac{y}{L} \hspace{0.5cm} \theta =\frac{T-T_{c}}{T_{free}-T_{c}} \end{aligned}$$ $$\begin{aligned} Ra=\frac{g \beta (T_{free}-T_{c})L^{3}}{\nu \alpha} \hspace{0.5cm} Pr = \frac{\nu}{\alpha} \end{aligned}$$ The optical thickness is defined as $\tau=\kappa_{a} L$ and non-dimensional irradiation is given as $$\begin{aligned} \overline{G}=\frac{G}{\sigma T^{4}_{free}} \hspace{1cm}\end{aligned}$$ The Nu$_{C}$ and Nu$_{R}$ are conductive and radiative Nusselt numbers respectively and defined as $$\begin{aligned} Nu_{C}=\frac{q_{Cw}L}{k(T_{free}-T_{c})} \hspace{1cm} Nu_{R}=\frac{q_{Rw}L}{k(T_{free}-T_{c})} \hspace{1cm} \end{aligned}$$ Thus, the total Nusselt number is defined as below, $$\begin{aligned} Nu=Nu_{C}+Nu_{R} \hspace{0.5cm} \hspace{0.5cm} \end{aligned}$$ Validation ========== In the absence of any standard benchmark test case for the present problem, the validation has been performed with three cases, first, the standalone feature of collimated beam irradiation problem, in second step, pure natural convection problem which is heated from the bottom, in the third and last step combined convection and radiation in differentially heated cavity have been verified with present solver. The collimated irradiation feature [@RAD19] has been tested in a square cavity, as shown in the Fig. (\[collimated\_geo\]). The left side of the wall has a small window of size 0.05 at a height of 0.6. The walls of square cavity are black and cold and also the medium inside the cavity is non-participating. A collimated beam is irradiated on the window in normal direction. It is expected that the beam would travel in normal direction without any attenuation also, the incident should be normal on the opposite wall. Figure (\[collimated\_rayTrav\]) shows the contour of irradiation which clearly shows the travel of collimated in the normal direction without any attenuation. For second step, fluid flow with heat transfer (without radiation) is validated against Aswatha et al. [@Aswatha] and combined diffuse radiation and natural convection in a cavity whose top and bottom walls are adiabatic and vertical walls are isothermal at differential temperatures and radiatively opaque, has been validated against Lari et al. [@Lari]. Figure \[temp\_valid\] shows the combined results of step two and three and they are in good agreement with the published results. [5cm]{} ![Validation of collimated beam feature (a) geometry (b) contour of irradiation shows the travel of the beam[]{data-label="collimated_valid_case"}](collimated_geo.png "fig:"){width="8cm"} [5cm]{} ![Validation of collimated beam feature (a) geometry (b) contour of irradiation shows the travel of the beam[]{data-label="collimated_valid_case"}](2D_coli.png "fig:"){width="6cm"} ![Validation results for pure and combined diffuse radiation with natural convection[]{data-label="temp_valid"}](temp_valid.png){width="8cm"} Grid Independent Test ===================== Numerical solutions of Navier-Stokes equation and radiation transfer equation is sensitive to the spatial discretization. Additionally, radiative transfer equation also requires angular space discretization, which provides directions where radiation transfer equation is solved. Thus, optimum number of grids and directions have been obtained through independent test study in two steps, as below 1. Spatial grids independence test: Three spatial grid sizes, i.e 60$\times$60, 80$\times$80 and 100$\times$100 are chosen to calculate the average Nusselt number on the bottom wall for the present problem of natural convection with diffuse radiation. The calculated Nusselt number values for above three grids arrangement on the bottom wall are 6.89, 7.0 and 7.05, respectively. The percentage error between the first and second is 1.52$\%$, whereas between second and third is 0.7$\%$. Thus, the grid points i.e 80$\times$80 is selected for further study. 2. Angular direction independence test: The polar discretization has no effect on the two-dimensional cases, then OpenFOAM fixed the number of polar directions to 2, in one hemisphere of angular space, while azimuthal considered directions are 3, 5, and 7 for the angular direction independent studied. The total Nusselt numbers on the bottom wall for 80$\times$80 spatial grid and 2$\times$3, 2$\times$5, 2$\times$7 are 6.946, 7.0 and 7.012, respectively. The difference in the second and third angular discreitization is 0.17$\%$. Thus finally $n_\theta \times n_\phi =2 \times 5$ in one hemisphere angular space is selected. Results and Discussions ======================= The present study is performed for natural convection in a cavity that is convective heated from the bottom with heat transfer coefficient 50 $W/m^2$K and free stream temperature 305 K, the side walls are symmetrical cooled at constant temperature 296 K and top wall is adiabatic. All walls are opaque for radiation. The study is performed in two parts, in first part effect of diffuse radiation on natural convection for various optical thicknesses has been explained. In second part, a semitransparent window of width 0.05 is created on the left wall at height of 0.7 and a collimated beam enters into the cavity through this window at an angle of $45^0$. The effect of collimated beam on fluid flow and heat transfer characteristics for different optical thickness of the fluid has been observed. Combined Diffuse Radiation and Natural Convection ------------------------------------------------- The effect of various optical thicknesses on fluid flow and heat transfer characteristics have been presented in the section below ### Non-Dimensional Isothermal and Stream Function Contours 50 --------- ------- ------- ------- ------- ------- ------ Maximum 0.749 0.755 0.753 0.773 0.788 0.80 : Maximum non-dimensional temperature inside the cavity for various optical thicknesses of the fluid[]{data-label="ConDif_iso_table"} Figure (\[CR\_temp\]) represents the non-dimensional isothermal contours inside the cavity for optical thickness $\tau=0$. The isothermal lines are densely placed near to the bottom wall, sparsely near to the isothermal walls, very less isothermal lines at the central part of the cavity near to the top wall. This reveals the temperature gradients are more near to the bottom wall, less near to isothermal wall and temperature remains uniform at the core of the cavity near to the top wall. The isotherms near to the bottom wall are densely populated, and they originate and rise from both the corners and reach maximum height at the middle of the cavity. This creates a non-uniform temperature and temperature gradient on the bottom wall, thus, non-uniform heating from the bottom. The maximum non-dimensional temperature attains at the mid point of the bottom wall. The non-dimensional isothermal contours are symmetrical about the middle vertical line (i.e $ X=1/2$). This isothermal characteristic is found similar in all optical thicknesses cases (contours are not depicted here), but, the different maximum value of non-dimensional temperatures are shown in Table \[ConDif\_iso\_table\]. The maximum non-dimensional temperature increases with the increase of optical thickness of the fluid. [1cm]{} ![Contours of (a)non-dimensional temperature and (b) non-dimensional stream function for optical thickness $\tau=0$ []{data-label="diffuse"}](CR_T_0.png "fig:"){width="6cm"} [6cm]{} ![Contours of (a)non-dimensional temperature and (b) non-dimensional stream function for optical thickness $\tau=0$ []{data-label="diffuse"}](CR_SF_0.png "fig:"){width="6cm"} Being a buoyancy driven flow, the temperature field inside the cavity governs the flow field. The contours of non-dimensional stream function of the flow field are depicted in Fig. (\[CR\_SF\]). The bottom wall is non-uniformly heated and temperature field is symmetric about the mid vertical line, this causes to develop two symmetrical counter rotating vortices are established in the flow field in the steady state condition. These two counter rotating vortices are mirror image about the mid-vertical line, unlike the thermal field. The stream line are closely placed at the junction of the two vortices, whereas these lines are relatively sparsely placed in other locations in the cavity, reveals shear-forces are high on the junction of two vortices. The above flow filed characteristics do not change with the change in optical thickness of the fluid. However, the maximum values of non-dimensional stream function change with optical thickness of fluid. Table \[ConDif\_SF\_table\] represents the maximum value of non-dimensional stream function with optical thicknesses of the fluid inside the cavity. Also, the maximum non-dimensional stream function values is same for both the vortex, but in opposite sign, representing counter rotating vortices exist in the cavity. The maximum values first increases with optical thicknesses upto 10 then start, decreasing. This reveals long distance phenomenon of radiation changes to local phenomenon with increase of optical thickness of fluid. 50 -------------- -------- -------- -------- -------- -------- ------- Left vortex -0.047 -0.048 -0.048 -0.051 -0.051 -0.05 Right vortex 0.047 0.048 0.048 0.051 0.051 0.05 : Maximum non-dimensional stream function for both the vortices inside the cavity for various optical thicknesses of fluid[]{data-label="ConDif_SF_table"} Figure (\[CR\_VV\]) shows the variation of vertical velocity along the horizontal line at mid height of the cavity. The direction of vertical velocity is in downward direction near to the isothermal walls and reaches a maximum downward velocity at non dimensional distance of 0.1 from both the isothermal walls. Afterwards, the downward velocity starts decreasing and reaches to zero at the centre of both the vortices, i.e, at a distance 0.3 from the isothermal walls, further, the direction of vertical velocity is in upward direction and reaches to peak at the mid point, i.e, the point where two counter-rotating vortices meet. The maximum downward and upward non-dimensional vertical velocities are 0.29 and 0.45, respectively. The non-dimensional vertical velocity curves are same for all optical thicknesses case, however, little variation appears at the inflection point of the curves for different optical thicknesses case as shown in the inset of Fig. (\[CR\_VV\]). The maximum velocity in both the directions increase upto optical thickness 10 and then start decreasing for higher optical thicknesses. ![Variation of non-dimensional vertical velocity along the horizontal line at mid height of cavity for various optical thicknesses[]{data-label="CR_VV"}](CR_vertical_vel.png){width="8cm"} The non-dimensional temperature variation on the bottom wall along the horizontal direction, at the mid-height of cavity and on the top wall are shown in Fig. \[CR\_wall\_temp\] (a), (b) and (c), respectively. The non-dimensional temperature starts sudden increasing from the isothermal walls side upto non-dimensional distance of 0.1 after that the slow increase in non-dimensional temperature happens and reaches maximum at middle point only on the bottom wall. Whereas, it reaches platue at the non-dimensional distance 0.2 on mid-height of cavity and remain flat on top wall. These curves are again symmetrical about the mid point. The non-dimensional temperature curves remain the same for all optical thicknesses till the first change in the slope reaches, after that temperature curves start showing the difference for different optical thicknesses. The non-dimensional temperature keeps on increasing with optical thicknesses of the fluid. [6cm]{} ![Non-dimensional temperature variation at (a) bottom wall (b) mid- height of cavity and (c) top wall for various optical thicknesses[]{data-label="CR_wall_temp"}](CR_bot_temp.png "fig:"){width="6cm"} [5cm]{} ![Non-dimensional temperature variation at (a) bottom wall (b) mid- height of cavity and (c) top wall for various optical thicknesses[]{data-label="CR_wall_temp"}](CR_mid_temp.png "fig:"){width="6cm"} [14cm]{} ![Non-dimensional temperature variation at (a) bottom wall (b) mid- height of cavity and (c) top wall for various optical thicknesses[]{data-label="CR_wall_temp"}](CR_top_temp.png "fig:"){width="6cm"} ### Nusselt Number Variation The variation of total Nusselt number on the bottom is shown in Fig. (\[CR\_bot\_Nu\]). A very high Nusselt number is found near to isotherm walls and this value decreases all off sudden within few distance away from the isothermal walls. Afterwords, the gradient of the curve reduces and finally minimum value of Nusselt number is obtained at mid point of the wall. The lowest value of Nusselt number appears due to the fact that stagnation is at the meeting point of counter rotating vortices, thus a lower temperature gradient. The point of change of gradient of the Nusselt number curve also coincides with the point of maximum value of downward velocity (see Fig. \[CR\_VV\]). The Nusselt number curves are similar for all optical thicknesses near to the isothermal walls till the curves change its gradient. Afterwards, the Nusselt number keeps on decreasing with the increase of optical thicknesses. The maximum value of Nusselt number is 18 near to isothermal walls and minimum is 3 for optical thickness 50 at the midpoint on the bottom wall. [1cm]{} ![Variation of total Nusselt number for different optical thicknesses at (a) bottom and (b) left wall[]{data-label="CR_Nu"}](CR_bot_tot_Nu.png "fig:"){width="6cm"} [6cm]{} ![Variation of total Nusselt number for different optical thicknesses at (a) bottom and (b) left wall[]{data-label="CR_Nu"}](CR_left_tot_Nu.png "fig:"){width="6cm"} The Nusselt number variation on the isothermal wall is depicted in Fig. \[CR\_left\_Nu\], a very high Nusselt number 28 is found near to bottom of the cavity on the isothermal wall, this Nusselt number decreases very sudden within small height (Y=0.1) from the bottom and rises a little long the height of the wall. The small difference in Nusselt number curve appears at inflection point of curve otherwise, the optical thickness of the fluid has insignificant effect on the Nusselt number on the isothermal walls. The line average Nusselt number on the bottom and the isothermal walls for the different optical thicknesses are shown in Table \[ConDif\_avgNu\_table\]. The conduction Nusselt number is higher than the radiative Nusselt number on both the walls and the conductive Nusselt number increases little with increase in optical thickness whereas for the radiative Nusselt number decreases drastically on the bottom wall this makes total Nusselt number decreases with increase of optical thickness of fluid. A similar observation also been made for the cold wall while Nusselt number is positive on the bottom wall, but, it is negative for the isothermal wall, reveals that the energy enters inside the cavity through bottom and leaves through the isothermal wall. ----- ------- ------- ------- -------- -------- -------- Total 0 5.530 1.470 7.0 -2.766 -0.734 -3.5 0.5 5.580 1.330 6.91 -2.720 -0.738 -3.458 1 5.613 1.23 6.843 -2.704 -0.713 -3.417 5 5.737 0.827 6.564 -2.817 -0.458 -3.275 10 5.819 0.602 6.421 -2.891 -0.312 -3.203 50 5.929 0.202 6.131 -2.936 -0.095 -3.031 ----- ------- ------- ------- -------- -------- -------- : Average Nusselt number on different walls for convection diffuse ration case[]{data-label="ConDif_avgNu_table"} Combined Collimated/Diffuse Radiation and Natural Convection ------------------------------------------------------------ In the previous section, we have seen that the two vortices inside the cavity remain symmetric for all optical thickness of fluid, which may change with collimated irradiation. To simulate collimated irradiation, a semitransparent window of non-dimensional width 0.05 at a non-dimensional height of 0.7 is created on left wall of the cavity and a collimated beam irradiated at an angle of $45^0$. The beam travels into the fluid in that may have different optical thicknesses and the dynamics of these two vortices and heat transfer characteristics have been studied with collimated irradiation. ### Irradiation Contours Progression of the collimated beam inside the cavity can be best represented by irradiation contours. The non-dimensional irradiation contours inside the cavity for optical thickness ($\tau$) 0, 0.5, 1, 5, 10 and 50 of the fluid are shown in Fig. \[G\_collimated\](a)-(f), respectively. The optical thickness $\tau=0$ corresponds to transparent medium i.e no absorption-emission by fluid, therefore, the irradiation strength remains constant till it reaches to the bottom wall, here at non-dimensional distance of 0.7 from the left corner on bottom wall, whereas, strength of collimated beam reduces for non-zero optical thickness of the fluid. This decrement increases with optical thickness as shown in Fig. \[G\_collimated\](b)-(f). The collimated irradiation does not reach to the bottom wall for optical thickness 5 of fluid onwards. Furthermore the collimated energy gets absorbed near to the window for optical thickness 50 (see Fig. \[G\_collimated\]f). [4cm]{} ![Contours of Non-dimensional irradiation for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="G_collimated"}](C_CR_G_0.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional irradiation for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="G_collimated"}](C_CR_G_pt5.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional irradiation for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="G_collimated"}](C_CR_G_1.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional irradiation for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="G_collimated"}](C_CR_G_5.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional irradiation for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="G_collimated"}](C_CR_G_10.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional irradiation for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="G_collimated"}](C_CR_G_50.png "fig:"){width="6cm"} ### Stream lines and Isothermal line contours The non-dimensional stream function contours inside the cavity for optical thickness 0, 0.5, 1, 5 10 and 50 are shown in Fig. \[SF\_collimated\](a)-(f), respectively. As we have seen in previous section, the cavity contains two symmetrical right and left vortices occupying equally half of the space inside cavity without collimated incidence, the collimated incidence causes asymmetricity in these two vortices which vary with behaviour of the fluid for the radiation heat transfer. As the fluid is transparent for radiation transfer, all the collimated beam energy strikes on the bottom wall at a non-dimensional distance of 0.7 from left corner and creates a hot spot on the wall. There is enhance in buoyancy force in upward direction at that location, this inturn, enhances the upwards direction force in the right vortex, therefore the right vortex becomes thinner and left vortex becomes thicker. The opposite behaviour has been observed with non-zero optical thickness in the fluid. This may be owing to fact that collimated beam is travelling through the left vortex which absorbs the radiation energy and creates local heating of fluid, this enhances upward buoyancy force in the left vortex, whereas some heat is also being transfered to right vortex through absorption in the right vortex and hot spot at the bottom. The energy absorbed by left vortex may be higher due to larger distance travelled by collimated beam in left vortex, this causes decrease in the size of left vortex and increases in size of right vortex. The above scenario for optical thickness continues upto $\tau=10$ of the fluid, however, this changes with further increase of the optical thickness of the fluid. Further,due to the increase in the optical thickness, the size of left and right vortices keeps on increasing and decreasing, respectively. This can be infer from the irradiation contours (Fig. \[G\_collimated\]e and f) that all collimated energy gets absorbed near to the wall only and most of the energy is transferred out of domain due to isothermal condition of the wall, instead of transferring to the fluid. Thus, there is relative decrease in the upward buoyancy force and result in little increase in size of left vortex. There is further increase in the size of left vortex with increase in optical thickness of the fluid. Nevertheless, a situation may arrive with very very high optical thickness case when collimated irradiation case will be similar to without collimated beam case. [4cm]{} ![Contours of Non-dimensional stream function for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="SF_collimated"}](C_CR_SF_0.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional stream function for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="SF_collimated"}](C_CR_SF_pt5.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional stream function for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="SF_collimated"}](C_CR_SF_1.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional stream function for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="SF_collimated"}](C_CR_SF_5.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional stream function for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="SF_collimated"}](C_CR_SF_10.png "fig:"){width="6cm"} [5cm]{} ![Contours of Non-dimensional stream function for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="SF_collimated"}](C_CR_SF_50.png "fig:"){width="6cm"} [|ccccccc|]{} Optical thickness & 0 & 0.5 & 1 & 5 & 10 & 50\ ------------- Left vortex ------------- : Maximum non-dimensional stream function value for various optical thicknesses of fluid[]{data-label="co_SF_table"} & 0.052 & 0.046 & 0.041 & 0.03 & 0.04 & 0.048\ Right vortex & -0.049 & -0.056 & -0.061 & -0.07 & -0.06 & -0.054\ Table \[co\_SF\_table\] shows the value of maximum non-dimensional stream function for left and right vortices with optical thickness of the fluid. The non-dimensional maximum value of stream function of left vortex decreases with the increases of optical thickness of fluid upto 5, then increases. Whereas, this value for right vortex increases upto the optical thickness 5, then decreases. The negative stream function value indicates that direction of rotation of vortex is in clock wise. [4cm]{} ![Contours of non-dimensional isothermal for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="Temp_collimated"}](C_CR_T_0.png "fig:"){width="6cm"} [5cm]{} ![Contours of non-dimensional isothermal for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="Temp_collimated"}](C_CR_T_pt5.png "fig:"){width="6cm"} [5cm]{} ![Contours of non-dimensional isothermal for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="Temp_collimated"}](C_CR_T_1.png "fig:"){width="6cm"} [5cm]{} ![Contours of non-dimensional isothermal for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="Temp_collimated"}](C_CR_T_5.png "fig:"){width="6cm"} [5cm]{} ![Contours of non-dimensional isothermal for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="Temp_collimated"}](C_CR_T_10.png "fig:"){width="6cm"} [5cm]{} ![Contours of non-dimensional isothermal for various optical thicknesses of (a) $\tau=0$ (b) $\tau=0.5$ (c) $\tau=1$ (d) $\tau=5$ (e) $\tau=10$ and (f)$\tau=50$[]{data-label="Temp_collimated"}](C_CR_T_50.png "fig:"){width="6cm"} The effect of collimated beam irradiation on the isothermal contours inside the cavity for optical thickness 0, 0.5, 1, 5, 10 and 50 are shown in Fig. \[Temp\_collimated\](a)-(f), respectively. In absence of collimated irradiance the symmetrical isotherm about the vertical line at middle of the cavity gets tilted either right or left to the vertical line depending upon the behaviour of the fluid for radiation energy. The isothermal lines are tilted towards right for transparent fluid, (see Fig. \[Temp\_collimated\](a)), this is due to the fact that right vortex is smaller than the left vortex and reverse is true for non-zero optical thickness fluid (see Fig. \[Temp\_collimated\](b)-(c)) i.e, the isotherms are tilted towards left. The area near to top adiabatic wall away from isothermal walls is almost at uniform temperature at lower optical thickness, however, the area of uniform temperature reduces with higher optical thickness and also, the isotherm lines do not remain smooth. The isotherm lines are bent in the line of collimated beam at higher optical thickness of fluid. This bend of isothermal lines diminishes with increase of optical thicknesses and remain limited near to the wall for optical thickness 50. Table \[co\_iso\_table\] represents the maximum non-dimensional temperature that exists in side cavity. The maximum non-dimensional temperature decreases with the increase of optical thickness of the fluid upto 1 then starts increasing and this maximum non-dimensional temperatures always exist on the bottom wall (see Fig. \[Temp\_collimated\]). ![Variation of non-dimensional vertical velocity along the horizontal line at the mid height of the cavity[]{data-label="C_CR_VV"}](C_CR_Vy_xdir.png){width="8cm"} Optical thickness 0 0.5 1 5 10 50 ------------------- ------- ------- ------- ------ ------ ------- Maximum 0.894 0.760 0.761 0.78 0.78 0.801 : Maximum non-dimensional temperature inside the cavity for various optical thicknesses of fluid[]{data-label="co_iso_table"} The variation of non-dimensional vertical velocity in the horizontal direction at the mid height of the cavity is shown in Fig. \[C\_CR\_VV\]. The vertical velocity is non-dimensionalised by convective velocity scale as explained in section 3.1. The vertical velocity starts increasing in downward direction and reaches to maximum value at 0.065 distance from wall and 0.09 distance away from the right wall for all optical thickness of the fluid, then its value starts decreasing and reaches to zero at centre point of both the vortices, afterwards, its direction is in upward and again reaches to maximum value. The non-dimensional maximum upward velocity is achieved at different locations for different optical thicknesses of medium, and this location is at junction of the two vortices. It is therefore, the location of maximum upward velocity is little bit towards right to mid point for transparent fluid and left for non-zero optical thickness of fluid, whereas, the location of maximum downward velocity remain almost same for all optical thicknesses. The value of maximum upward velocity also decreases for optical thickness of fluid greater than 5. This could be the owing to fact that less energy is being transferred to the fluid from the bottom wall for case of optical thickness greater than 1. The maximum downward non-dimensional velocity is 0.3 and upwards velocity is 0.4 for optical thickness 0 and minor difference in these values for other optical thicknesses. [6cm]{} ![Variation of non-dimensional temperature in horizontal direction at (a) bottom wall (b) mid-height and (c) top wall for different optical thicknesses[]{data-label="C_CR_BMT"}](C_CR_bot_Temp.png "fig:"){width="6cm"} [5cm]{} ![Variation of non-dimensional temperature in horizontal direction at (a) bottom wall (b) mid-height and (c) top wall for different optical thicknesses[]{data-label="C_CR_BMT"}](C_CR_mid_temp.png "fig:"){width="6cm"} [14cm]{} ![Variation of non-dimensional temperature in horizontal direction at (a) bottom wall (b) mid-height and (c) top wall for different optical thicknesses[]{data-label="C_CR_BMT"}](C_CR_top_temp.png "fig:"){width="6cm"} The non-dimensional temperature variation in the horizontal direction on the bottom wall, at mid height and top wall are depicted in Fig. \[C\_CR\_BMT\] for various optical thicknesses. The non-dimensional temperature curve for the bottom wall has two maxima for optical thickness upto 1 (see Fig. \[C\_CR\_BT\]), that corresponds to strike point of the collimated beam and stagnation point developed at the junction of two vortices. Nevertheless, the location of maxima corresponds to strike point of collimated beam remain fixed, it is the highest for radiatively transparent fluid case and keeps on decreasing with increase of optical thickness of fluid. The hot spot due to beam strike does not occur for the medium of optical thickness higher than 5 as the radiative energy gets absorbed with in the fluid before striking on the bottom wall (Fig. 10). The location of second maxima is decided by the relative size of two vortices. The right vortex is smaller to the left vortex for the transparent case, thus stagnation point shifts to the right of the mid point for non-zero optical thickness. The second maximum point is at non-dimensional distance of 0.55, 0.475, 0.4, 0.425, and 0.43 for optical thicknesses 0, 0.5, 1, 5, 10 and 50, respectively. One point to notice is that the highest value of non-dimensional temperature is greater for the transparent case at the location of beam strikes. The second maxima on the non-dimensional temperature curve at the mid height is only achieved on the mid height appear at the location of collimated beam (Fig. \[C\_CR\_Temp\]) i.e at left of the vertical at the middle of cavity. The non-dimensional temperature increases on the left part of the curve before second maxima arrives for optical thickness upto 5, afterwards it decreases. Then non-dimensional temperature curve remains almost similar for all optical thickness. There is no maxima in the curve for non-dimensional temperature curve for top wall. The non-dimensional temperature increases from both ends upto distance of 0.21 and remain flat afterwards. The maximum non-dimensional temperature increased to 0.37 for optical thickness 5 and little lesser for other optical thicknesses on the top wall. ### Nusselt number [6cm]{} ![Variation of (a) conduction (b) radiation and (c) total Nusselt number on bottom wall []{data-label="Nu_bot_col"}](Co_bot_cond_Nu.png "fig:"){width="6cm"} [5cm]{} ![Variation of (a) conduction (b) radiation and (c) total Nusselt number on bottom wall []{data-label="Nu_bot_col"}](Co_bot_rad_Nu.png "fig:"){width="6cm"} [14cm]{} ![Variation of (a) conduction (b) radiation and (c) total Nusselt number on bottom wall []{data-label="Nu_bot_col"}](Co_bot_tot_Nu.png "fig:"){width="6cm"} The conduction Nusselt number curve on the bottom wall (Fig. \[cond\_nu\]) decreases drastically upto non-dimensional distance 0.1 from the left corner, afterwards the curve changes its slope and keeps on decreasing till lowest point in the curve. This nature of curve remains same for all optical thicknesses of the fluid, afterwords it shows the effect of collimated beam strike. The conduction Nusselt number curve obtains a minimum value of almost 2 at non-dimensional distance of 0.65 from left corner and all of sudden increases to 12 at the strike length of the collimated beam for transparent case. Further, it decreases to value 6 and then starts increasing and reaches to maximum value of 14 on the right side of the isothermal wall for optical thickness zero case. The lowest value for conduction Nusselt number is obtained at distance of 0.45 and remains constant upto the strike point of beam for optical thickness 0.5, then it reaches to value 1, further its behaviour is similar to the transparent case. Furthermore, the minimum value for conduction Nusselt number obtained at same point and then starts increasing for optical thickness 1 and sudden increase to a value of 7 at strike point of beam. Whereas, the Nusselt number curve behaviour remains unaffected by the collimated beam for optical thickness of fluid 5 and above. The radiation Nusselt number curve is almost constant over whole length of the bottom wall expect the length over which collimated beam strikes. The maximum Nusselt number is 12 which is negative indicates that the radiation energy leaves through radiation mode of heat transfer over this length for transparent case. The Nusselt number curve achieves peak at beam strike length and its value decreases with increase of optical thicknesses of the fluid, also no peak appears for optical thicknesses 10 and 50. One interesting thing is to notice that radiative Nusselt number is almost zero for optical thickness 10 and 50. The total Nusselt number which is a linear combination of conduction and radiation Nusselt number is dominated by conduction Nusselt number in the most portion of the length except the length over which collimated beam strikes. The total Nusselt number at the beam strike portion is dominated by the radiation Nusselt number and this portion exhibit the minimum Nusselt number in the curve. The minimum Nusselt number is almost zero which represents the adiabatic condition at the strike point of the collimated beam for the radiation transparent fluid case and this increases to 3 and 5 for optical thickness 0.5 and 1, respectively, whereas no peak appears in the total Nusselt number curve for optical thickness 5, 10 and 50. This is true as collimated energy gets absorbed within the fluid before reaching to the bottom wall. [6cm]{} ![Variation of (a) conduction (b) radiation and (c) total Nusselt number on left wall[]{data-label="Nu_left_col"}](Co_left_cond_Nu.png "fig:"){width="6cm"} [5cm]{} ![Variation of (a) conduction (b) radiation and (c) total Nusselt number on left wall[]{data-label="Nu_left_col"}](Co_left_rad_Nu.png "fig:"){width="6cm"} [14cm]{} ![Variation of (a) conduction (b) radiation and (c) total Nusselt number on left wall[]{data-label="Nu_left_col"}](Co_left_tot_Nu.png "fig:"){width="6cm"} The conduction, radiation and total Nusselt number variations on the left wall which also contains the semitransparent window are depicted in Fig. \[Nu\_left\_col\] (a), (b), and (c), respectively. The conduction Nusselt number sudden decreases to minimum value at height 0.15 and further, small variation happen upto semitransparent window. There is sudden increase in the conduction Nusselt number on the semitransparent window, this is mostly negative, reveals the that energy leaves from this wall through conduction mode of heat transfer. The conduction Nusselt number value is 20 on the semitransparent window on left wall for the optical thickness 50 case, and this reduces to 3 after semitransparent window and remains constant over rest height of the vertical wall. With the increase in the optical thicknesses of the fluid, the conduction Nusselt number decreases on the semitransparent window, however, optical thicknesses does not have much effect on other portion of the left wall. There is no increase in the conduction Nusselt number at the window for optical thickness 0.5 or optically transparent fluid. The radiative Nusselt number is also negative but very small in number all over the length, except on the semitransparent window (Fig. \[rad\_nu\_left\]), where radiative Nusselt number is positive, reveals the radiative flux is coming inside the cavity through the window. There is mimimum variation of radiation Nusselt number with optical thickness over the length except at the window, where Nusselt number is 26 and remains same for all optical thicknesses. The total Nusselt number curve which is linear combination of conduction and radiation Nusselt number is shown in Fig. \[tot\_nu\_left\]. The total Nusselt number is dominated by conduction over all height of the wall except around the window where radiative Nusselt number is in dominance. The total Nusselt number is negative indicates that the energy is going out from the isothermal wall except the window where Nusselt number is positive which indicates that the energy is coming inside from the window. Figure. \[Nu\_right\_col\], depicts the variation of total Nusselt number on the right wall. This wall does not have any semitransparent window, thus the total Nusselt number behaviour is similar to the left wall without any phenomenon that happens on the semitransparent window. The total Nusselt number is negative reveals that the heat is also being transferred outside from the right side wall and most heat is transferred within few small height of the wall afterward, heat transfer rate is small. ![Variation of total Nusselt number on the right wall for various optical thicknesses[]{data-label="Nu_right_col"}](Co_right_tot_Nu.png){width="7cm"} ----- ------------ ----------- ------- ------------ ----------- -------- ------------ ----------- -------- Conduction Radiation Total Conduction Radiation Total Conduction Radiation Total 0 5.84 0.154 5.994 -2.845 0.58 -2.264 -2.988 -0.778 -3.766 0.5 5.662 0.642 6.305 -3.042 0.55 -2.491 -3.021 -0.792 -3.813 1 5.602 0.848 6.45 -3.168 0.546 -2.621 -3.052 -0.777 -3.829 5 5.661 0.8 6.462 -3.602 0.779 -2.823 -3.137 -0.509 -3.646 10 5.724 0.601 6.325 -3.817 0.961 -2.855 -3.127 -0.343 -3.47 50 5.812 0.224 6.036 -4.099 1.208 -2.89 -3.037 -0.109 -3.146 ----- ------------ ----------- ------- ------------ ----------- -------- ------------ ----------- -------- : Average Nusselt number values on the different walls for various value of optical thicknesses of fluid[]{data-label="co_avgNu_table"} Table \[co\_avgNu\_table\] shows the average Nusselt number on the different walls of the cavity. The average conduction Nusselt number on the bottom wall decreases slowly with optical thickness upto 1 then starts increasing whereas treand for average radiation Nusselt number is faster and reverse to the conduction Nusselt number making total Nusselt number increase upto optical thickness 5, then decreases on the bottom wall. The average conduction and radiation Nusselt numbers increase for all optical thicknesses on the left wall but in reverse direction, i.e, conduction Nusselt number is negative and radiative Nusselt number is positive. The rate of increase of both the Nusselt numbers is such that total Nusselt number increases with increase of optical thickness of the fluid on the left wall. Furthermore, the conduction Nusselt number increases upto optical thickness 5 then decreases, whereas radiative Nusselt number only increases upto optical thickness 1 then decreases on the right wall. The linear combination of these two Nusselt numbers make total Nusselt number, that increase upto optical thickness 5 then decreases. All the Nusselt number on the right side wall is negative indicates that energy goes out by both modes of heat transfer through this wall. Conclusions =========== The effect of diffuse and collimated beam radiation on the symmetrical cooling case of natural convection in a cavity heated from bottom has been thoroughly investigated. The collimated beam radiation feature have been developed in OpenFOAM framework and an applications has been made by including libraries of other fluid flow and heat transfer models present in OpenFOAM package. The collimated beam case has been simulated by creating semitransparent window of non-dimensional width 0.05 at non-dimensional height of 0.7 on the left wall and a collimated beam of value 1000 $W/m^2$ is irradiated on semitransparent window at an angle of $45^0$. The following conclusions are drawn from the comprehensive study of fluid flow and heat transfer phenomena in the scenarios of diffuse and collimated beam radiation for various optical thickness of fluid inside the cavity. 1. The isotherm and stream lines are symmetrical about the vertical lines at the middle of the cavity and formation of two counter rotating vortices happen inside the cavity for diffuse radiation case. 2. The maximum values of non-dimensional temperature increases with the increase of optical thickness of the fluid whereas, maximum value of non-dimensional stream function increases upto optical thickness 5 then remains almost constant. 3. These symmetrical vortices change its dynamics by collimated incidence for different values of optical thickness of the fluid. The right side vortex is smaller than the left side for transparent fluid, whereas, left side vortex is smaller for non-zero optical thickness of the fluid. 4. The isotherms are slightly bent in the path of collimated beam for non-zero optical thicknesses of fluid. 5. The variation of Nusselt numbers are mostly dominated by conduction Nusselt number over the length of the walls except at the portion of the collimated irradiation, where radiative Nusselt number is dominating for lower value of optical thickness of fluid. 6. The average total Nusselt number increases with increase of optical thickness upto 5 on the bottom as well asright wall and it increases for all optical thicknesses on the left wall. 7. The energy enters into the cavity by conduction and radiative mode of heat transfer through bottom wall, whereas energy enters through radiation modes of heat transfer and leave by conduction mode of heat transfer from the left wall while energy leaves through both mode of heat transfer from the right wall. Acknowledgements {#acknowledgements .unnumbered} ================ The authors greatly acknowledge the financial support provided by Science and Engineering Research Board (SERB) (Statutory Body of the Government of India) via Grant.No:ECR/2015/000327 to carry out the present work. Declaration of interests ======================== The authors declare that they have no known financial interests or personal relationships that could have appeared to influence the work reported in this paper.

ArXiv

--- abstract: 'In this paper we prove [a precise estimate of the Sobolev embedding constant]{} on general noncompact Lie groups, for sub-Riemannian inhomogeneous Sobolev spaces endowed with relatively invariant measures. Such an estimate appears to be new even in the case of the classical inhomogeneous Sobolev spaces on ${\mathbb{R}}^d$. As an application, we prove local and global Moser–Trudinger inequalities.' address: - 'Department of Mathematics: Analysis, Logic and Discrete Mathematics, Ghent University, Krijgslaan 281, 9000 Ghent, Belgium' - 'Dipartimento di Matematica, Università degli Studi di Milano, Via C. Saldini 50, 20133 Milano, Italy' - 'Dipartimento di Scienze Matematiche “Giuseppe Luigi Lagrange”, Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129 Torino, Italy - Dipartimento di Eccellenza 2018-2022' author: - Tommaso Bruno - 'Marco M. Peloso' - Maria Vallarino title: | Sobolev embedding constants and\ Moser–Trudinger Inequalities on Lie groups --- [^1] [^2] Introduction ============ If $\Delta$ denotes the Laplacian on ${\mathbb{R}}^d$ and $L^p_\alpha=(I+\Delta)^{\alpha/2}L^p $ is the associated inhomogeneous Sobolev space, it is well known that $L^p_{\alpha} \hookrightarrow L^q$ when $1<p<\infty$, $0<\alpha<d/p$ and $1/q=1/p-{\alpha}/{d}$. Interestingly, but also surprisingly to us, the related embedding constant has remained relatively unexplored [@Oz; @CT]. Most of the results, as it sounds natural, deal instead with the constant involved in the embedding $\dot{L}^p_\alpha=\Delta^{\alpha/2}L^p \hookrightarrow L^q$ of the homogeneous spaces. The sharp constant for this embedding has a long history and a multitude of applications, and in some special cases it has been obtained, see e.g. [@Talenti; @Aubin; @Lieb]. A well-established application of the Sobolev embedding theorem, both in the homogeneous and inhomogeneous case, is the classical Moser–Trudinger inequality [@Trudinger; @Moser], which arises as a substitute of boundedness for functions in the Sobolev space $L^{p}_{d/p}$, as this does not embed in $L^\infty$. It has the form $$\label{MT1} \int_{{\mathbb{R}}^d} \left( \exp(\gamma |f|^{p'}(x))-1 \right)\, {\mathrm{d}}x \le C_\Omega$$ for some $\gamma>0$ and all $f\in L^{p}_{d/p}$ supported in a fixed, sufficiently smooth compact set $\Omega$ and with norm not larger than $1$. The Moser–Trudinger inequality on the whole space ${\mathbb{R}}^d$ appeared only at a later time, cf. [@AT; @Og; @Oz], and in this case the exponential needs to be regularized in order to make the integral converge. Nowadays, there exists a vast literature on these inequalities and their generalizations, and a thorough treatment would go out of the scope of the present work. We refer the reader to the recent papers [@FM1; @FM2] for a complete and extensive bibliography. The aim of this paper is to study analogous problems on general noncompact Lie groups. The natural substitutes of the Laplacian in this setting are sub-Laplacians with drift, see [@BPTV], and the measures with respect to which they are symmetric are absolutely continuous with respect to the right Haar measure of the group; their density is any of its continuous positive characters. This setting, and these operators in particular, were first studied in [@HMM], and an associated theory of Sobolev spaces, that we shall denote by $L^p_\alpha$, was developed in [@BPTV]. In such generality, since the Riesz transforms might be unbounded on $L^p$ even when $1<p<\infty$ (see e.g. [@GG]), the natural Sobolev spaces seem to be those endowed with an inhomogeneous norm, which reduces to the Sobolev norm of $L^{p}_\alpha$ in the Euclidean case. Our main attention is focused on the case of Sobolev spaces defined in terms of the [*intrinsic*]{} sub-Laplacian $\cL$ (see [@Agrachev-et-al] and Section \[s: setting\] below for its definition) and the left Haar measure $\lambda$, with respect to which $\cL$ is self-adjoint. In this case, we obtain an estimate for the constant of the embedding $L^p_\alpha(\lambda)\hookrightarrow L^q(\lambda)$, of the form $C \, p\, q^{1-1/p}/(p-1)$, where $C$ depends only on the group and the set of vector fields that define $\cL$, but not on the other parameters. To the best of our knowledge, such a precise estimate is new even in ${\mathbb{R}}^d$. Comparisons with the few known sharp estimates in the homogeneous case (especially [@Talenti]) seem to suggest that the dependence on $p$ and $q$ might be optimal, but we are unable to prove or disprove this at the moment. We leave this [question, which is open even in the case of ${{\mathbb R}}^d$]{}, to future work. As an application of such quantitative Sobolev embeddings, we prove local and global Moser–Trudinger inequalities in our setting, with an explicit description of the threshold $\gamma$ for which the analogue of  and its global version hold. Our approach is close in spirit, and inspired by, [@Oz]. We refer the reader also to the recent work [@RuzYess]. Setting and Preliminaries {#s: setting} ========================= Let $G$ be a noncompact connected Lie group with identity $e$. We denote by $\rho$ a right Haar measure, by $\chi$ a continuous positive character of $G$, and by $\mu_\chi$ the measure with density $\chi$ with respect to $\rho$. As the modular function on $G$, which we denote by $\delta$, is such a character, $\mu_\delta$ is a left Haar measure on $G$. We denote it by $\lambda$. Observe also that $\mu_1=\rho$. Let $\mathbf{X}= \{ X_1,\dots, X_\ell\}$ be a family of left-invariant linearly independent vector fields which satisfy Hörmander’s condition. Let $d_C(\, \cdot\, ,\, \cdot\, )$ be its associated left-invariant Carnot–Carathéodory distance. We let $|x|=d_C(x,e)$, and denote by $B_r$ the ball centred at $e$ of radius $r$. The volume of the ball $B_r$ with respect to the measure $\rho$ will be denoted by $V(r)=\rho(B_r)$; recall that $V(r)= \lambda(B_r)$. We also recall (cf. [@Guiv; @Varopoulos1]) that there exist two constants, $d\in {\mathbb{N}}^*$ depending on $G$ and $\bf{X}$, and $D>0$ depending only on $G$, such that $$\label{pallepiccolegrandi} \begin{split} C^{-1} r^d \leq V(r) &\leq C r^d\qquad \forall r\in (0,1]\,, \\ V(r)&\leq C \e^{Dr}\quad \; \, \forall r\in (1,\infty)\,, \end{split}$$ where $C>0$ is independent of $r$. We also recall that, for any character $\chi$, one has (cf. [@HMM]) $$\sup\nolimits_{|x|\leq r} \chi(x) = \e^{{\mathfrak{c}}(\chi) r}, \quad \mbox{where} \quad {\mathfrak{c}}(\chi) = ( |X_1\chi(e)|^2 + \cdots + |X_\ell\chi(e)|^2)^{1/2}.$$ The metric measure space $(G, d_C, \mu_\chi)$ is then locally doubling, but not doubling in general. Observe moreover that ${\mathfrak{c}}(\chi) = {\mathfrak{c}}(\chi^{-1})$. If $p\in [1,\infty)$, the spaces of (equivalent classes of) measurable functions whose $p$-power is integrable with respect to $\mu_\chi$ will be denoted by $L^p(\mu_\chi)$, and endowed with the usual norm which we shall denote by $\| \cdot \|_{L^p(\mu_\chi)}$. The space $L^\infty$ is defined analogously, but it is independent of $\chi$. The convolution between two functions $f$ and $g$, when it exists, is defined by $$f*g(x) =\int_G f(xy^{-1})g(y)\, {\mathrm{d}}\rho(y),\qquad x\in G\,.$$ We recall that Young’s inequality for the measure $\lambda$ has the following form [@HR]: if $1<p\leq q<\infty$ and $r\geq 1$ is such that $\frac{1}{p}+ \frac{1}{r}=1+\frac{1}{q}$, then $$\label{Young} \begin{split} \|f*g\|_{L^q(\lambda)} &\leq \|f\|_{L^p(\lambda)} \| \widecheck{g}\|_{L^r(\lambda)}^{r/p'} \|g\|_{L^r(\lambda)}^{r/q},\\ \|f*g\|_{L^\infty} & \leq \|f\|_{L^p(\lambda)} \| \widecheck{g}\|_{L^{p'}(\lambda)}, \end{split}$$ where $\widecheck{g}(x)= g(x^{-1})$. We denote by $\Delta_{\chi}$ the sub-Laplacian with drift $$\Delta_\chi = - \sum_{j=1}^\ell (X_j^2 +(X_j\chi)(e) X_j).$$ In particular, we shall denote the operator $\Delta_\delta$ by ${\mathcal{L}}$, and observe that it coincides with the [*intrinsic*]{} sub-Laplacian of [@Agrachev-et-al]. The operator $\Delta_\chi$ generates a diffusion semigroup, i.e.$(\e^{-t\Delta_\chi})_{t>0}$ extends to a contraction semigroup on $L^p(\mu_\chi)$ for every $p \in [1, \infty]$ (see [@HMM]) whose infinitesimal generator, with a slight abuse of notation, we still denote by $\Delta_\chi$. Observe that $\Delta_1$ is the standard left-invariant sum-of-squares sub-Laplacian. We denote by $p_t^\chi$ the convolution kernel of $\e^{-t\Delta_\chi}$, and we recall that by  [@VCS Theorem IX.1.3] and [@BPTV eq. (2.8)] there exist constants $b, c>0$ depending only on $G$ and $\bf{X}$ such that $$\label{heatkernelestimate} p_t^\chi(x) \leq c \, (\delta \chi^{-1})^{1/2}(x)\, (1\wedge t)^{-\frac{d}{2}} \, \e^{-\frac{1}{4} t {\mathfrak{c}}(\chi)^2}\, \e^{-b \frac{|x|^2}{t}}, \qquad x\in G,\,t>0.$$ Let $b_0 = \sqrt{b}/2$, and define $$\label{fixedtranslation} {\mathfrak{\tau}}_\chi = \max \left\{\frac{2}{b} \left[ {\mathfrak{c}}(\delta \chi^{-1}) +2D+b_0 \right]^2 - \frac{1}{4}{\mathfrak{c}}(\chi)^2, 1\right\}.$$ Observe that ${\mathfrak{c}}(\delta\chi^{-1}) =0$ if $\chi = \delta$ or, equivalently, if $\mu_\chi = \lambda$. We refer the reader to [@BPTV; @BPV1; @BPV2] for background and further details on these matters. Following [@BPTV], when $p\in (1,\infty)$ and $\alpha> 0$ we define the Sobolev spaces $L^p_\alpha(\mu_\chi)$ as the set of functions $f\in L^p(\mu_\chi)$ such that $(\tau_\chi I + \Delta_{\chi})^{\alpha/2} f\in L^p(\mu_\chi)$, endowed with the norm $$\label{equivtranslation} \| f\|_{L^p_{\alpha}(\mu_\chi)} = \|(\tau_\chi I + \Delta_\chi)^{\alpha/2}f\|_{L^p(\mu_\chi)}.$$ If $\alpha=0$, we let $L^p_0(\mu_\chi)=L^p(\mu_\chi)$. We recall that  is equivalent to the norm $\|f\|_{L^p(\mu_\chi)} + \|\Delta_\chi^{\alpha/2} f\|_{L^p(\mu_\chi)}$, see [@BPTV]. The reason for choosing the shift $\tau_\chi$ in the definition of $L^p_\alpha(\mu_\chi)$ will be clarified later on; we refer the reader, in particular, to Remark \[rem:Rd\] below. In [@BPTV Theorem 1.1] the Sobolev embeddings $L^{p}_{d/p}(\mu_\chi) \hookrightarrow L^q(\mu_{\chi^{q/p}\delta^{1-q/p}})$ for every $q\geq p$, and $L^p_\alpha(\lambda) \hookrightarrow L^q(\lambda)$ when $0<\alpha<d$ and $q>p$ are such that ${1}/{q}= {1}/{p}-{\alpha}/{d}$, were established. In this paper we estimate the embedding constants in a precise way, as we explain below. Throughout the paper, we shall disregard any dependence of the embedding constants on $G$ and $\bf{X}$, which are assumed to be fixed once and for all from this point on. We shall, instead, obtain explicit results in terms of the dependence on $p$, $q$ and $\alpha$. A generic constant depending only on $G$ and $\bf{X}$ will be denoted by $C$ or $C(G,\bf{X})$, and its value may vary from line to line. Recall in particular that $d=C(G,\bf{X})$ and $D=D(G)$. For $\alpha>0$, let $G_{\chi}^{\alpha}$ be the convolution kernel of $(\tau_\chi I +\Delta_\chi)^{-\alpha/2}$. Let $$\label{Glocglob} G_{\chi}^{\alpha, {\mathrm{loc}}} =G_{\chi}^{\alpha}\mathbf{1}_{B(e,1)}, \qquad G_{\chi}^{\alpha, {\mathrm{glob}}} = G_{\chi}^{\alpha}\mathbf{1}_{B(e,1)^c}.$$ The following is a refined version of [@BPTV Lemma 4.1]. \[Lemma4.1-revised\] There exists $C=C(G, \mathbf{X})>0$ such that, for $\alpha\in (0,d)$ and $x\in G$, $$\begin{aligned} |G_{\chi}^{\alpha, {\mathrm{loc}}} (x)| &\leq C\, \frac{\alpha}{d-\alpha} (\delta \chi^{-1})^{1/2}(x) |x|^{\alpha-d}\mathbf{1}_{B(e,1)}(x) ,\\ |G_{\chi}^{\alpha, {\mathrm{glob}}}(x)| &\leq C \, (\delta \chi^{-1})^{1/2}(x)\, \e^{-(2D+{\mathfrak{c}}(\delta \chi^{-1})+b_0)|x|}\mathbf{1}_{B(e,1)^c}(x).\end{aligned}$$ We recall that the convolution kernel $G_{\chi}^{\alpha}$ can be written as $$G_{\chi}^{\alpha}=\frac{1}{\Gamma(\alpha/2)} \int_0^\infty t^{\alpha/2-1} \e^{-\tau_\chi\, t} p_t^\chi\, {\mathrm{d}}t,$$ so that by  $$G_{\chi}^{\alpha}(x) \le \frac{C}{\Gamma(\alpha/2)} (\delta\chi^{-1})^{1/2}(x) \int_0^\infty t^{\alpha/2-1}(1\wedge t)^{-d/2} \e^{-(\tau_\chi +\frac{1}{4} {\mathfrak{c}}(\chi)^2) t} \e^{-b|x|^2/t}\, {\mathrm{d}}t \,.$$ Set $a=\tau_\chi +\frac{1}{4} {\mathfrak{c}}(\chi)^2$. Since $at +b|x|^2/t \geq \frac{1}{2} ( at + b/t +\sqrt{2ab}|x|)$, we see that when $|x|\ge1$, $$\begin{aligned} G_{\chi}^{\alpha}(x) & \le \frac{C}{\Gamma(\alpha/2)} (\delta\chi^{-1})^{1/2}(x)\, \e^{-\frac{1}{2} \sqrt{2ab} |x|} \int_0^\infty t^{\alpha/2 -1} ( 1 \wedge t )^{-d/2} \e^{-\frac{at}{2}-\frac{b}{2t}}\, {\mathrm{d}}t\\ & \le C (\delta\chi^{-1})^{1/2}(x) \, \e^{-(2D+{\mathfrak{c}}(\delta \chi^{-1})+b_0)|x|} \,.\end{aligned}$$ On the other hand, when $|x|\le 1$, splitting the integral we have $$\begin{aligned} G_{\chi}^{\alpha}(x) &\le C\, \alpha \, (\delta\chi^{-1})^{1/2}(x) \bigg( \int_0^1 t^{(\alpha-d)/2 -1} \e^{-b|x|^2/t} \, {\mathrm{d}}t + \int_1^\infty t^{\alpha/2 -1} \e^{-a t }\e^{-b|x|^2/t} \, {\mathrm{d}}t \bigg)\\ & \eqqcolon{C} \, \alpha\, (\delta\chi^{-1})^{1/2}(x) \left( G_1(x) + G_2(x) \right) .\end{aligned}$$ It is clear, since $\alpha\in (0,d)$ and $a\geq 1$, that $G_2(x)\le C$. Since $\alpha\in (0,d)$, we also have $$\begin{aligned} G_1(x) & = |x|^{\alpha -d}\bigg( \int_{|x|^2}^1+ \int_1^\infty \bigg) u^{(d-\alpha)/2-1} \e^{-bu} \, {\mathrm{d}}u \le C\, |x|^{\alpha -d}\bigg( \frac{1}{d-\alpha} (1-|x|^{d-\alpha})+1 \bigg)\,,\end{aligned}$$ and the conclusion follows. The Sobolev embedding constant ============================== We are now ready to state our main result. We point out that, to the best of our knowledge, such precise dependence of the embedding constant on $p$ and $q$ is new even in the case of the inhomogeneous Sobolev spaces in ${\mathbb{R}}^d$; see Remark \[rem:Rd\] below. Our result should be compared with [@Oz], where the dependence on $p$ is not explicit. \[teo:embed\] Let $p\in (1,\infty)$, $\alpha \in [0,d)$ and $q \in [p,\infty)$ be such that $\frac{1}{q}= \frac{1}{p}-\frac{\alpha}{d}$. There exists $A_1=A_1(G,\mathbf{X})>0$ such that $$\| f\|_{L^q(\lambda)}\leq A_1\, \frac{p}{p-1} q^{1-1/p} \|f\|_{L^p_\alpha(\lambda)}$$ for all $f\in L^{p}_{\alpha}(\lambda)$. Observe first that we may assume $\alpha>0$ and $q>p$, for otherwise the embedding constant is $1$. Then define $$K_\alpha(x) = |x|^{\alpha-d} \mathbf{1}_{B(e,1)}(x) , \qquad \tilde{K}_\alpha(x) = \e^{-(2D+b_0)|x|}\mathbf{1}_{B(e,1)^c}(x).$$ We claim that $$\begin{aligned} \| f\ast K_\alpha \|_{L^q(\lambda)} &\leq C(G,\mathbf{X}) \, \frac{d-\alpha}{\alpha} \, p'\, q^{1-1/p} \|f\|_{L^p(\lambda)},\label{claim}\\ \| f\ast \tilde{K}_\alpha\|_{L^q(\lambda)} &\leq C(G,\mathbf{X}) \|f\|_{L^p(\lambda)}.\label{claim2}\end{aligned}$$ By combining these bounds and Lemma \[Lemma4.1-revised\], we obtain that $$\| (\tau_\delta I +{\mathcal{L}})^{-\alpha/2} f\|_{L^{q}(\lambda)}\leq C \,p'\, q^{1-1/p}\|f\|_{L^{p}(\lambda)},$$ which implies $$\|f\|_{L^{q}(\lambda)}\leq C \, p' \, q^{1-1/p}\|f\|_{L^{p}_\alpha(\lambda)},$$ where $C$ depends only on $G $ and $\bX$. Thus, it remains to prove the claims. The bound  follows by applying Youngs’s inequality   $$\label{tilde K} \|f\ast \tilde K_{\alpha}\|_{L^q(\lambda)}\leq \|f\|_{L^p(\lambda)} \|\tilde K_{\alpha}\|_{L^r(\lambda)}^{r(1/p'+1/q)},$$ where $r\in (1,\infty)$ is such that $\frac1p+\frac1r=1+\frac1q$. We then have $$\begin{aligned} \| \tilde K_{\alpha} \|^r_{L^r(\lambda)} &\leq C \int_{B_1^c} \e^{-r(2D+ b_0)|x|} \,{\mathrm{d}}\lambda(x) \\ & \le C\, \sum_{k=0}^\infty \int_{2^{k}\leq |x|< 2^{k+1}} \e^{-r(2D+ b_0)|x|} \,{\mathrm{d}}\lambda(x) \\ & \le C\, \sum_{k=0}^\infty \e^{-r(2D+b_0) 2^k+D2^{k+1}} \leq C , \end{aligned}$$ which combined with implies . The remainder of the proof will be devoted to show . For $s>0$, define $$K_{\alpha,s}^{(1)} = K_\alpha\, \mathbf{1}_{B(e,s)}, \qquad K_{\alpha,s}^{(2)} = K_\alpha\, \mathbf{1}_{B(e,s)^c}.$$ Let now $\tilde p\in (1,\infty)$ and $\tilde q\in (\tilde p,\infty)$ be such that $\frac{1}{\tilde q}= \frac{1}{\tilde p}-\frac{\alpha}{d}$. By Young’s inequality , there exists $C>0$ depending only on $G$ and $\bf X$ such that $$\begin{aligned} \| f \ast K_{\alpha,s}^{(1)}\|_{L^{\tilde p}(\lambda)} & \leq \|f\|_{L^{\tilde p}(\lambda)} \|\widecheck{K}_{\alpha,s}^{(1)}\|_{L^{1}(\lambda)}^{1/\tilde p} \| {K}_{\alpha,s}^{(1)}\|_{L^{1}(\lambda)}^{1/\tilde p'} \notag \\ & \leq C \|f\|_{L^{\tilde p}(\lambda)} \times \begin{cases} \frac{1}{\alpha} s^\alpha & \mbox{if } s<1 \\ \frac{1}{\alpha} & \mbox{if } s\geq 1, \end{cases} \label{bound1}\end{aligned}$$ and $$\begin{aligned} \| f \ast K_{\alpha,s}^{(2)}\|_{L^\infty} &\leq \|f\|_{L^{\tilde p}(\lambda)} \|\widecheck{K}_{\alpha,s}^{(2)}\|_{L^{\tilde p'}\!(\lambda)} \notag \\ & \leq C \|f\|_{L^{\tilde p}(\lambda)} \times \begin{cases} \left(\frac{\tilde q}{d\tilde p'}\right)^{1/\tilde p'} (s^{(\alpha-d)\tilde p' +d}-1)^{1/\tilde p'}& \mbox{if } s<1 \\ 0 & \mbox{if } s\geq 1. \end{cases} \label{bound02}\end{aligned}$$ Observe that $(\alpha-d)\tilde p' +d<0$ under our assumptions. For $t>0$ we now set $$s(t) = \left[ 1+ \frac{d\tilde p'}{\tilde q} \left( \frac{t}{2} \right)^{\tilde p'}\right]^{\frac{1}{(\alpha-d) \tilde p' + d}},$$ and observe that $s(t)\leq 1$ for every $t>0$. By , $$\label{bound2} \| f \ast K_{\alpha,s(t)}^{(2)}\|_{L^\infty} \leq C \frac{t}{2} \|f\|_{L^{\tilde p}(\lambda)} \qquad \forall t>0\,.$$ Thus, with $C$ the same constant as in  and , $$\begin{aligned} \sup_{t>0} t\, \lambda\! &\left(\left\{ x\colon \: |f* K_\alpha(x)|>t\right\} \right)^{1/\tilde q}\\ & = C \|f\|_{L^{\tilde p}(\lambda)} \: \sup_{t>0} t\, \lambda\!\left(\left\{ x\colon \: |f\ast K_\alpha(x)|>Ct \|f\|_{L^{\tilde p}(\lambda)}\right\} \right)^{1/\tilde q} \\ & \leq C\|f\|_{L^{\tilde p}(\lambda)} \: \sup_{t>0} t\, \lambda\!\left(\left\{ x\colon \: |f\ast K_{\alpha,s(t)}^{(1)}(x)|>C\frac{t}{2} \|f\|_{L^{\tilde p}(\lambda)}\right\} \right)^{1/\tilde q} \\ & \qquad \qquad +C\|f\|_{L^{\tilde p}(\lambda)} \:\sup_{t>0} t\, \lambda\!\left(\left\{ x\colon \: |f\ast K_{\alpha,s(t)}^{(2)}(x)|>C\frac{t}{2} \|f\|_{L^{\tilde p}(\lambda)}\right\} \right)^{1/\tilde q}\\ & = C \|f\|_{L^{\tilde p}(\lambda)}\: \sup_{t>0} t\, \lambda\!\left(\left\{ x\colon \: |f\ast K_{\alpha,s(t)}^{(1)}(x)|>C\frac{t}{2} \|f\|_{L^{\tilde p}(\lambda)}\right\} \right)^{1/\tilde q},\end{aligned}$$ since $s(t)$ was chosen so that the second super-level set was empty. By , we get $$\begin{aligned} \sup_{t>0} t\, \lambda\! & \left(\left\{ x\colon \: |f\ast K_{\alpha,s(t)}^{(1)}(x)|>C\frac{t}{2} \|f\|_{L^{\tilde p}(\lambda)}\right\} \right)^{1/\tilde q}\\ &\leq \sup_{t>0} t\, \left[ \left( \frac{2}{ {C} t \|f\|_{L^{\tilde p}(\lambda)}} \right)^{\tilde p} \| f\ast K_{\alpha,s(t)}^{(1)}\|_{L^{\tilde p}(\lambda)}^{\tilde p} \right]^{1/\tilde q}\\ & \leq \sup_{t>0} t \left( \frac{ {C} t\|f\|_{L^{\tilde p}(\lambda)} }{2} \right)^{-\tilde p/\tilde q} \left( \frac{s(t)^\alpha}{\alpha}\right)^{\tilde p/\tilde q} {C^{\tilde p/\tilde{q}}}\|f\|_{L^{\tilde p}(\lambda)}^{\tilde p/\tilde q}\\ & = \left( \frac{2}{\alpha} \right)^{\tilde p/ \tilde q} \: \sup_{t>0} t^{1-\tilde p/\tilde q} \left[1 + \frac{d \tilde p'}{\tilde q} \left( \frac{t}{2}\right)^{\tilde p'} \right]^{-\frac{1}{\tilde p'}(1- \frac{\tilde p}{\tilde q})} \\ & = \frac{2}{\alpha^{\tilde p/\tilde q}} \left( \frac{\tilde q}{d\tilde p'} \right)^{\frac{1}{\tilde p'}(1- \frac{\tilde p}{\tilde q})} \,\sup_{u>0} u^{1-\tilde p/\tilde q} (1 + u^{\tilde p'})^{-\frac{1}{\tilde p'}(1- \frac{\tilde p}{\tilde q})}. \end{aligned}$$ It is now easy to see that, for every $\tilde p$ and $\tilde q$, $$\sup_{u>0} u^{1-\tilde p/\tilde q} (1 + u^{\tilde p'})^{-\frac{1}{\tilde p'}(1- \frac{\tilde p}{\tilde q})} =\sup_{v>0} \big[ v/(1+v)\big] ^{\frac{1}{\tilde p'}(1- \frac{\tilde p}{\tilde q})} =1.$$ Moreover, by our assumption on $(\tilde p,\tilde q)$, $$\frac{\tilde p}{\tilde q} = 1-\tilde p\frac{\alpha}{d}\quad\text{and}\quad \frac{1}{\tilde p'}\Big(1- \frac{\tilde p}{\tilde q}\Big) = (\tilde p-1)\frac{\alpha}{d} ,$$ so that we end up with the inequality $$\begin{aligned} \|f\ast K_\alpha\|_{L^{\tilde q,\infty}(\lambda)} & = \sup_{t>0} t\, \lambda\! \left( \left\{ x\colon \: |f\ast K_\alpha(x)|>t\right\} \right)^{\frac{1}{\tilde q}} \notag \\ & \leq C \alpha^{\tilde p \alpha/d -1} \left( \frac{\tilde q}{d\tilde p'} \right)^{(\tilde p -1)\alpha/d} \|f\|_{L^{\tilde p}(\lambda)}. \label{debtildeptildeq}\end{aligned}$$ In other words, the operator defined by $\mathcal{K}_\alpha f = f* K_\alpha$ is of weak type $(\tilde{p}, \tilde q)$ for every $\tilde p, \tilde q$ such that $\frac{1}{\tilde q}= \frac{1}{\tilde p}-\frac{\alpha}{d}$, $1< \tilde p < \tilde q <\infty$, $0< \alpha< d$. In a similar way we can also prove that $\mathcal{K}_\alpha$ is of weak type $(1, \tilde q)$ for $\frac{1}{\tilde q}= 1-\frac{\alpha}{d}$ and $0< \alpha< d$. Indeed, the estimate holds also for $\tilde p=1$ and $$\label{K2bis} \| f \ast K_{\alpha,s}^{(2)}\|_{L^\infty} \leq C \|f\|_{L^{1}(\lambda)} \times \begin{cases} s^{\alpha-d}& \mbox{if } s<1 \\ 0 & \mbox{if } s\geq 1. \end{cases}$$ We now set $$s(t) = \begin{cases} \left(1+ \frac{t}{2} \ \right)^{1/(\alpha-d)}&t\geq 2\\ 1&0<t<2\,, \end{cases}$$ which is $\leq 1$. Then holds also in this case and we obtain as above that $$\begin{aligned} \sup_{t>0} t\, \lambda\! &\left(\left\{ x\colon \: |f* K_\alpha(x)|>t\right\} \right)^{1/\tilde q}\\ & \leq C \|f\|_{L^{1}(\lambda)}\: \sup_{t>0} t\, \lambda\!\left(\left\{ x\colon \: |f\ast K_{\alpha,s(t)}^{(1)}(x)|>C\frac{t}{2} \|f\|_{L^{\tilde p}(\lambda)}\right\} \right)^{1/\tilde q}\\ & \leq C \|f\|_{L^{1}(\lambda)}\: \sup_{t>0} t\, \left( \frac{2}{ {C} t \|f\|_{L^{1}(\lambda)}} \| f\ast K_{\alpha,s(t)}^{(1)}\|_{L^{1}(\lambda)} \right)^{1/\tilde q}\,. $$ We now notice that $$\begin{aligned} \sup_{0<t<2} t\, \left( \frac{2}{ {C} t \|f\|_{L^{1}(\lambda)}} \| f\ast K_{\alpha,s(t)}^{(1)}\|_{L^{1}(\lambda)} \right)^{1/\tilde q} & \leq \sup_{0<t<2} t \left( \frac{t\|f\|_{L^{1}(\lambda)} }{2} \right)^{-1/\tilde q} \left( \frac{1}{\alpha}\right)^{1/\tilde q} \|f\|_{L^1(\lambda)}^{1/\tilde q}\\ &= 2\alpha^{-1/{\tilde q}}\,, \end{aligned}$$ while $$\begin{aligned} \sup_{t\geq 2} t\, \left( \frac{2}{ {C} t \|f\|_{L^{1}(\lambda)}} \| f\ast K_{\alpha,s(t)}^{(1)}\|_{L^{1}(\lambda)} \right)^{1/\tilde q}& \leq \sup_{t\geq 2} t \left( \frac{t\|f\|_{L^{1}(\lambda)} }{2} \right)^{-1/{\tilde q}} \left( \frac{s(t)^{\alpha}}{\alpha}\right)^{1/\tilde q} \|f\|_{L^1(\lambda)}^{1/\tilde q}\\ &\leq C\sup_{t\geq 2} t^{1-\frac{1}{\tilde q}}\left( \frac{2}{\alpha} \right)^{1/{\tilde q}}\left(\frac{t}{2}\right)^{-1/d}=C\,\alpha^{-1/{\tilde q}}\,.\end{aligned}$$ This proves that $$\label{deb1tildeq} \|f\ast K_\alpha\|_{L^{\tilde q,\infty}(\lambda)} \leq C \alpha^{-1/\tilde q}\|f\|_{L^{1}(\lambda)}.$$ We shall now use the Marcinkiewicz interpolation theorem for two specific choices of the couple $(\tilde{p}, \tilde q)$. Being $p\in (1,\infty)$, $q \in (p, \infty)$, and $\alpha/d=1/p-1/q$ as in the statement, we define $$\label{p1q1p2q2} \left (\frac{1}{p_{1}},\frac{1}{q_{1}}\right)=\left(1,1-\frac{\alpha}{d}\right), \qquad \left(\frac{1}{p_{2}}, \frac{1}{q_{2}}\right)=\left(\frac{\alpha}{d}+\frac{1}{q+1},\frac{1}{q+1}\right).$$ By the above, $\mathcal{K}_{\alpha}$ is both of weak type $(1,q_{1})$ and $(p_{2},q_{2})$ with norms $M(1,q_1)$ and $M(p_2,q_2)$ respectively, given by $$\begin{aligned} M(1,q_1) & = \alpha ^{-(1-\alpha/d)} , \label{M1} \\ M(p_2,q_2) & =\Big(\frac{d^{\alpha/d}}{\alpha}\Big) \left( \frac{\alpha}{d}\right)^{\frac{\alpha/d}{\alpha/d+1/(q+1)}} \Big[ \Big(1-\frac{\alpha}{d} -\frac{1}{q+1} \Big)(q+1)\Big]^{\frac{1}{1+d/(\alpha(q+1))}- \frac{\alpha}{d}}. \end{aligned}$$ We select $$\label{teta} \theta= \frac{1-\frac{1}{p}}{1-\frac{\alpha}{d} -\frac{1}{q+1} } .$$ Notice that we indeed have $0<\theta<1$, $1/p=(1-\theta)/p_{1}+\theta/p_{2}$ and $1/q=(1-\theta)/q_{1}+\theta/q_{2}$. Thus, $\mathcal{K}_\alpha$ is of strong type $(p,q)$, i.e. bounded from $L^p(\lambda)$ to $L^q(\lambda)$, with norm bounded by $$C M_0(1, q_1, p_2, q_2)^{1/q} M(1,q_1)^{1-\theta} M(p_2,q_2)^{\theta},$$ see e.g. [@Z Ch. XII, (4.18)], where $$\begin{aligned} M_0(1, q_1, p_2, q_2) = \frac{q(p_2/p)^{q_2/p_2}}{q_2-q} + \frac{q/p^{q_1}}{q-q_1} \label{M0} . \end{aligned}$$ If we observe that $$\label{Ipq} M_0(1, q_1, p_2, q_2)^{1/q} M(1,q_1)^{1-\theta} M(p_2,q_2)^{\theta} \leq C\, \frac{d-\alpha}{\alpha}\,p'\, q^{1-1/p},$$ then we get precisely , which concludes the proof of the proposition. We now sketch the proof of . First we consider $M_1=M(1,q_1)$, and simply observe that $$M_1 = \alpha^{-1} d^{\alpha/d}(\alpha/d)^{\alpha/d} \leq d \, \alpha^{-1}$$ as $\alpha/d\leq 1$ and $x^x \leq 1$ for $x\in (0,1]$. Then we consider $M_0=M_0(1, q_1, p_2, q_2)$, and observe that $$M_0 = q \left(y+1+\frac{1}{q}\right)^{1+y} (1+y)^{-(1+y)} + C(p,q)$$ where $$C(p,q) = p^{-p'q/(q+p')} \Big(1+\frac{p'}{q}\Big), \qquad y = \frac{\alpha}{d}(q+1).$$ Moreover $$\begin{aligned} \left(y+1+\frac{1}{q}\right)^{1+y} (1+y)^{-(1+y)}= \left[ \left( 1+ \frac{1}{q(1+y)}\right)^{q(1+y)}\right]^{1/q} \leq \e\end{aligned}$$ since $q(1+y) \geq 1$ and by the estimate $ (1 +\frac{1}{x})^x \leq \e$ for $x\geq 1$. Thus $M_0 \leq \e \, q + C(p,q)$. We then consider $M_2=M(p_2,q_2)$, and estimate $M_2^\theta$. We first observe that $$M_2^\theta \leq d^\theta \alpha^{-\theta} \left( \frac{\alpha}{d}\right)^{\theta \frac{\alpha/d}{\alpha/d+1/(q+1)}} \Big[ \Big(1-\frac{\alpha}{d} -\frac{1}{q+1} \Big)(q+1)\Big]^{\theta\frac{\alpha/d}{\alpha/d+1/(q+1)}- \theta\frac{\alpha}{d}}$$ and that $$\begin{gathered} \left( \frac{\alpha}{d}\right)^{\theta \frac{\alpha/d}{\alpha/d+1/(q+1)}} \Big[ \Big(1-\frac{\alpha}{d} -\frac{1}{q+1} \Big)(q+1)\Big]^{\theta\frac{\alpha/d}{\alpha/d+1/(q+1)}- \theta\frac{\alpha}{d}} \\ = \left( \frac{\alpha}{d}\right)^{\frac{(1-1/p)}{1-z} \frac{\alpha/d}{z}} \Big[ \Big(1-z \Big)(q+1)\Big]^{ \frac{(1-1/p)}{z} \frac{\alpha}{d}}, \label{MM2}\end{gathered}$$ where $z= \frac{\alpha}{d} + \frac{1}{q+1}$. Observe that $0< z < 1/p<1 $. Since $(\alpha/d)/z \leq 1$ and $(1-1/p)/(1-z) \leq 1$, the right hand side of  is smaller than $$\begin{aligned} (q+1)^{1-1/p} = q^{1- 1/p} \left[ \left(1+1/q \right)^q\right]^{\frac{1}{q}(1-\frac{1}{p})}\leq \e\, q^{1- 1/p} .\end{aligned}$$ This proves that $M_2^\theta \leq \e \, d^\theta \, q^{1-1/p} \alpha^{-\theta}$. Putting everything together, we proved that $$M_0^{1/q} M_1^{1-\theta} M_2^{\theta} \leq \e\,d \, \alpha^{-1} (\e \, q+C(p,q))^{1/q} q^{1-1/p}.$$ It remains to estimate the term in the parenthesis in the right hand side. Observe first that $$(\e q+C(p,q))^{1/q} \leq (\e\, q)^{1/q} + C(p,q)^{1/q} \leq 2\e + C(p,q)^{1/q},$$ and then that $$C(p,q)^{1/q} \leq \Big(1+\frac{p'}{q}\Big)^{1/q} = \frac{d-\alpha}{d} \, p' \, \Big(1+\frac{p'}{q}\Big)^{1/q-1} \leq \frac{d-\alpha}{d} \, p' .$$ After observing that $ (d-\alpha) \, p' /d \geq 1$, the proof is complete. \[rem:Rd\] Assume that $G$ has polynomial growth. Then $\delta=1$, and ${\mathcal{L}}= \Delta$ is the sum-of-squares sub-Laplacian associated with ${\bf X}$. Since the exponential dimension $D$ can be taken arbitrarily small, one obtains $\tau_\delta =1$. Thus, in this case the Sobolev norm $\| \cdot\|_{L^p_\alpha(\lambda)}$ is the graph norm of $(I +\Delta)^{\alpha/2}$ in $L^p(\lambda)$. This in particular holds in ${\mathbb{R}}^d$, where ${\bf X} = \{\partial_1, \dots, \partial_d\}$, $\Delta$ is the Laplacian, $\lambda$ is the Lebesgue measure and $L^p_\alpha =L^p_\alpha(\lambda)$ is the classical inhomogeneous Sobolev space. Theorem \[teo:embed\] then reads as $$\| f\|_{L^q}\leq A_1\, \frac{p}{p-1} q^{1-1/p} \|f\|_{L^p_\alpha },$$ where $A_1$ depends only on the dimension $d$. As an application of Theorem \[teo:embed\], we shall prove a Moser–Trudinger inequality for the Sobolev spaces endowed with a left measure. To do this, we will need a precise version of the interpolation inequality [@BPV1 eq. (6.1)] associated to the interpolation space $(L^p(\lambda),L^p_{\alpha}(\lambda))_{[\theta]}=L^p_{\theta\alpha}(\lambda)$, which was originally proved in [@BPTV Lemma 3.1]. To prove this refined estimate, we follow some ideas developed in [@AM]; see also [@PV]. \[teo:interpolation\] Let $p\in(1,\infty)$ and define $$\mathcal C_p= \inf_{\sigma>0} \sup_{t\in{{\mathbb R}}} \e^{\sigma(1- t^2)} \| (\tau_\delta I+\mathcal L)^{it} \|_{L^p(\lambda)\rightarrow L^p(\lambda) } .$$ Then $1\leq \mathcal C_p<\infty$ and for all $f \in L^p_\alpha(\lambda)$, $\alpha\geq 0$, and $\theta\in(0,1)$ we have $$\label{f:interpolationinequality} \|f\|_{L^p_{\theta\alpha}(\lambda)}\leq \mathcal C_p \|f\|_{L^p(\lambda)}^{1-\theta} \, \|f\|_{L^p_{\alpha}(\lambda)}^{\theta}\,.$$ For $\sigma>0$, let $$\mathcal C_{p,\sigma} = \sup_{t\in{{\mathbb R}}} \e^{\sigma(1- t^2)} \| (\tau_\delta I+\mathcal L)^{it} \|_{L^p(\lambda)\rightarrow L^p(\lambda) }.$$ Since $\mathcal C_{p,\sigma}$ is finite for all $\sigma>0$ by [@Cowling Corollary 1], see also [@Meda], it follows that $\mathcal C_{p}$ is finite. Moreover, since $(\tau_\delta I+\mathcal L)^{it} = I $ for $t=0$, one gets $\mathcal C_{p,\sigma} \geq \e^{\sigma} \geq 1$, hence also $\mathcal C_{p}\geq 1$. Suppose that $f=\sum_{j=1}^Na_j\chi_{E_j}$, $h=\sum_{k=1}^{N'}a_k'\chi_{E'_k}$ are two simple functions on $G$. Let $S= \big\{ z\in{{\mathbb C}}:\, 0<\Re z<1\big\}$, and let $\ov S$ denote its closure. For every $z\in \ov S$ we define $$w(z)=\e^{\sigma z^2}\int_G(\tau_\delta I+\mathcal L)^{-\alpha z/2}f(x)h(x)\, {\mathrm{d}}\lambda(x).$$ Then $w$ is holomorphic on $S$, continuous on $\ov S$ and $w$ is bounded on $\ov S$. Indeed, $$\begin{aligned} \sup_{z\in\ov{S}} |w(z)|&\leq \sum_{j=1}^N\sum_{k=1}^{N'}|a_j||a_k'| \sup_{z\in\ov{S}} \Big| \e^{\sigma z^2}\int_{E_k'}(\tau_\delta I+\mathcal L)^{-\alpha z/2}\chi_{E_j}(x)\, {\mathrm{d}}\lambda(x) \Big|\\ &\leq \mathcal C_{p,\sigma} \sum_{j=1}^N\sum_{k=1}^{N'}|a_j||a_{k}'|\lambda (E_{k}')^{1/p'} \sup_{0\leq x\leq 1}\|(\tau_\delta I+\mathcal L)^{-\alpha x/2}\|_{L^p(\lambda)\rightarrow L^p(\lambda)}\lambda(E_j)^{1/p} <\infty\,. \end{aligned}$$ We now observe that for every $t\in\mathbb R$ $$|w(it)|\leq \mathcal C_{p,\sigma} \|f\|_{L^p(\lambda)}\|h\|_{L^{p'}(\lambda)}$$ and $$|w(1+it)|\leq \mathcal C_{p,\sigma} \|(\tau_{\delta}I+\mathcal L)^{-\alpha/2}f\|_{L^p(\lambda)}\|h\|_{L^{p'}(\lambda)}\,.$$ By the classical three lines theorem it follows that $$|w(1-\theta)|\leq \mathcal C_{p,\sigma} \|f\|^{\theta }_{L^p(\lambda)} \|(\tau_{\delta}I+\mathcal L)^{-\alpha/2}f\|^{1-\theta }_{L^p(\lambda)}\|h\|_{L^{p'}(\lambda)}\,.$$ By taking the supremum over all simple functions $h$ such that $\|h\|_{L^{p'}(\lambda)}\leq 1$ we have $$\|(\tau_\delta I+\mathcal L)^{- (1-\theta)\alpha/2}f \|_{L^p(\lambda)}\leq \mathcal C_{p,\sigma} \|f\|^{\theta }_{L^p(\lambda)} \|(\tau_{\delta}I+\mathcal L)^{-\alpha/2}f\|^{1-\theta }_{L^p(\lambda)}.$$ By using the density of simple functions in $L^p(\lambda)$ and choosing $g=(\tau_{\delta}I+\mathcal L)^{-\alpha/2}f$ we get $$\|(\tau_\delta I+\mathcal L)^{\theta \alpha/2}g \|_{L^p(\lambda)}\leq \mathcal C_{p,\sigma} \| (\tau_{\delta}I+\mathcal L)^{\alpha/2} g\|^{\theta }_{L^p(\lambda)} \|g\|^{1-\theta }_{L^p(\lambda)},$$ which is equivalent to $$\|g\|_{L^p_{\theta\alpha}(\lambda)}\leq \mathcal \mathcal C_{p,\sigma} \| g\|^{\theta }_{L^p_{\alpha}(\lambda)} \|g\|^{1-\theta }_{L^p(\lambda)}\,.$$ By taking the infimum over all $\sigma>0$, the inequality follows. As a corollary of the estimate of Theorem \[teo:embed\] and Proposition \[teo:interpolation\], we obtain the following global Moser–Trudinger inequality. Keeping the notation therein, we define $$\gamma_1 =[\e\, (\mathcal C_pA_1\, p')^{p'}p']^{-1}.$$ \[teo:GMT2\] Let $p\in (1,\infty)$. For $\gamma \in [0,\gamma_1)$ and $f\in L^p_{d/p}(\lambda)$ with $\|f\|_{L^p_{d/p}(\lambda)}\leq 1$, $$\label{eq:GMT2} \int_G \Big( \exp(\gamma |f|^{p'})-\sum_{0\leq k<p-1} \frac{\gamma^k}{k!} |f|^{p'k} \Big)\, {\mathrm{d}}\lambda \leq C(G,\mathbf{X},p) \|f\|_{L^p(\lambda)}^p.$$ We point out that, even in the case of the Laplacian in ${\mathbb{R}}^d$, the best constant $\gamma_1$ for which holds is not known, other than in the cases $d/p=1$ [@Li-Ruf] and $d/p=2$ [@Lam-Lu]. By Theorem \[teo:embed\] and the interpolation inequality , when $q>p$ we obtain $$\label{new} \|f\|_{L^q(\lambda)} \leq A_1 \, p'\, q^{1-1/p} \mathcal C_p \|f\|_{L^p_{d/p}(\lambda)}^{1-p/q}\|f\|_{L^p(\lambda)}^{p/q}.$$ Then, if $\|f\|_{L^p_{d/p}(\lambda)}\leq 1$, $$\begin{aligned} \int_G \Big( \exp(\gamma |f|^{p'})-\sum_{0\leq k<p-1} \frac{\gamma^k}{k!} |f|^{p'k} \Big)\, {\mathrm{d}}\lambda &= \sum_{k\geq p-1} \frac{\gamma^k}{k!} \|f\|_{L^{p'k}(\lambda)}^{p'k}\\ & \leq \|f\|_{L^p(\lambda)}^p \sum_{k\geq p-1} \frac{\gamma^k}{k!} (\mathcal C_pA_1\, p')^{p'k} (p'k)^{k} \\ &\leq C(G, {\bf{X}},p) \|f\|_{L^p(\lambda)}^p\end{aligned}$$ if $\gamma<\gamma_1$. The proof of the theorem is complete. The case of general measures ============================ In this final section we consider the case of a general sub-Laplacian $\Delta_\chi$ and relative measure $\mu_\chi$. Recall that an embedding like Theorem \[teo:embed\] fails if $\lambda$ is replaced by any other measure $\mu_\chi$, see [@BPTV], and as we show below in Remark \[rem:muchi\], a global Moser–Trudinger inequality also does not hold if $\mu_\chi\neq \lambda$. Thus we can only prove a local Moser–Trudinger theorem, that is, for compactly supported functions. Define $$\sf(\chi)= \max_{B(e,1)} \chi\delta^{-1} = \e^{{\mathfrak{c}}(\chi\delta^{-1})},$$ and observe that $\sf(\chi) \geq 1$ for all $\chi$’s. \[l: embeddingmuchi\] Let $p\in (1,\infty)$ and $q\in [p,\infty)$. There exists $A_2=A_2(G, \bX)>0$ such that $$\label{embedding} \| f\|_{L^q(\mu_{\chi^{q/p}\delta^{1-q/p}})} \leq \frac{A_2 \, \sf(\chi) }{p-1} \left(1+\frac{q}{p'}\right)^{\frac{1}{q}+\frac{1}{p'}} \| f\|_{ L^{p}_{d/p}(\mu_\chi)}$$ for all $f\in L^{p}_{d/p}(\mu_\chi)$. By Young’s inequality , we obtain that $$\begin{aligned} & \|(\tau_\chi I+\Delta_\chi)^{-d/2p}g\|_{L^q(\mu_{\chi^{q/p}\delta^{1-q/p}})} \notag\\ & \quad= \|(\chi\delta^{-1})^{1/p}g \ast (\chi\delta^{-1})^{1/p} G_\chi^{d/p}\|_{L^q(\lambda)} \nonumber \\ & \quad\leq \|(\chi\delta^{-1})^{1/p} g\|_{L^p(\lambda)} \| (\chi^{-1}\delta)^{1/p} \widecheck{G}_\chi^{d/p}\|_{L^r(\lambda)}^{r/{p'}} \,\| (\chi\delta^{-1})^{1/p} G_\chi^{d/p}\|_{L^r(\lambda)}^{r/{q}} \nonumber \\ & \quad= \| g\|_{L^p(\mu_\chi)} \| (\chi^{-1}\delta)^{1/p} \widecheck{G}_\chi^{d/p}\|_{L^r(\lambda)}^{r/{p'}} \,\| (\chi\delta^{-1})^{1/p} G_\chi^{d/p}\|_{L^r(\lambda)}^{r/{q}}, \label{appl-Young}\end{aligned}$$ where $r\in (1,\infty)$ is such that $\frac1p+\frac1r=1+\frac1q$. We split $G_\chi^{d/p}$ in its local and global part as in , and estimate the integrals of the two terms separately. By Lemma \[Lemma4.1-revised\], we obtain $$\begin{aligned} \| (\chi\delta^{-1})^{1/p} G_\chi^{d/p,{\mathrm{loc}}}\|_{L^r(\lambda)} & \le \frac{C}{p-1}\, \Big(\sum_{k=0}^\infty \int_{2^{-k-1}<|x|\le 2^{-k}} (\delta \chi^{-1})^{r(1/2-1/p)}(x) |x|^{r(d/p-d)} \,{\mathrm{d}}\lambda(x)\Big)^{1/r}\\ &\leq \frac{C}{p-1}\, \sf(\chi) \Big(\sum_{k=0}^\infty 2^{-kr(d/p-d)-kd}\Big)^{1/r} \\ & \leq \frac{C}{p-1}\, \sf(\chi) \Big(\int_0^1 u^{(d/p-d)r} u^{d-1}\, {\mathrm{d}}u\Big)^{1/r} = \frac{C\, \sf(\chi) }{p-1}\,\left(1+ \frac{q}{p'}\right)^{\frac{1}{q}+\frac{1}{p'}},\end{aligned}$$ where we used that $$\begin{aligned} \label{supmodulo} \sup_{y\leq |x| } (\delta \chi^{-1})^{1/2-1/p}(y) = \sup_{y\leq |x|} (\delta \chi^{-1})^{|1/2-1/p|}(x) = \e^{{\mathfrak{c}}(\chi \delta^{-1})|x|},\end{aligned}$$ and that $|1/2-1/p|\leq 1$. As for the global part of the kernel, using again , $$\begin{aligned} \| (\chi\delta^{-1})^{1/p} G_\chi^{d/p,{\mathrm{glob}}}\|_{L^r(\lambda)} &\leq C \Big(\int_0^\infty (\chi\delta^{-1})^{r(1/p-1/2)} \e^{-r(2D+{\mathfrak{c}}(\chi \delta^{-1}) +b_0)|x|} \,{\mathrm{d}}\lambda \Big)^{1/r}\notag \\ & \leq C \Big(\int_0^\infty \e^{-r(2D+b_0)|x|} \,{\mathrm{d}}\lambda \Big)^{1/r}\notag \\ & \le C\, \Big(\sum_{k=0}^\infty \e^{-r(2D+b_0) 2^k+D2^{k+1}} \Big)^{1/r}\leq C . \label{est-G-dp}\end{aligned}$$ The term $\| \widecheck{G}_{d/p}^c\|_{L^r(\lambda)}$ can be estimated in the same way, in view of  and by the radiality of the other terms appearing in the bound of Lemma \[Lemma4.1-revised\]. Keeping the notation of Proposition \[l: embeddingmuchi\], for $1<p<\infty$ we define $$\gamma_2= \left[ \e \, \left(\frac{A_2 \sf(\chi)^{2}}{p-1}\right)^{p'} \right]^{-1}.$$ The following result is inspired by [@Str]. \[teo:LMT\] Let $p\in (1,\infty)$. For $\gamma \in [0,\gamma_2)$, $$\sup_{\| f\|_{L^p_{d/p}(\mu_\chi)}\leq 1,\: \supp f\subseteq B(e,1) } \int_G \left( \exp(\gamma |f|^{p'})-1 \right)\, {\mathrm{d}}\mu_\chi <\infty.$$ We first notice that if $f$ is supported in $B(e,1)$ and $q> p$, then $$\| f\|_{L^q(\mu_\chi)} = \| (\chi \delta^{-1})^{\frac{1}{q}-\frac{1}{p}} f\|_{L^q(\mu_{\chi^{q/p}\delta^{1-q/p}})} \leq \sf(\chi )\| f\|_{L^q(\mu_{\chi^{q/p}\delta^{1-q/p}})} ,$$ so by Proposition \[l: embeddingmuchi\] $$\begin{aligned} \| f\|_{L^q(\mu_\chi)} & \leq \frac{A_2\, \sf(\chi)^2}{p-1} \left(1+\frac{q}{p'}\right)^{\frac{1}{q}+\frac{1}{p'}} \| f\|_{ L^{p}_{d/p}(\mu_\chi)}.\label{Eq1}\end{aligned}$$ If $f$ is supported in $B(e,1)$ and $\|f\|_{L^p_{d/p}(\mu_\chi)} \leq 1$, then $$\|f\|_{L^p(\mu_\chi)} \leq \|(\tau_\chi I+\Delta_\chi)^{-d/{2p}} \|_{L^p(\mu_\chi)\rightarrow L^p(\mu_\chi)}=C(\chi,p),$$ and $$\begin{aligned} \int_G &\left( \exp(\gamma |f|^{p'})-1 \right)\, {\mathrm{d}}\mu_\chi= \sum_{k=1}^\infty \frac{\gamma^k}{k!} \|f\|_{L^{p'k}(\mu_\chi)}^{p'k} \\ & \leq C(\chi,p) \sum_{1\leq k<p/p'} \frac{\gamma^k}{k!} \mu_\chi(B(e,1))^{1- k(p'-1)} + \sum_{k\geq p/p'} \frac{\gamma^k}{k!} \left( \frac{A_2\, \sf(\chi)^2}{p-1}\right)^{p'k} (k+1)^{k+1}\,,\end{aligned}$$ where we applied  when $kp'\geq p$, and Hölder’s inequality and the support condition of $f$ if $kp'< p$. If $\gamma\in [0,\gamma_2)$, then the latter series is convergent and the theorem is proved. \[rem:muchi\] Theorem \[teo:GMT2\] does not hold with any other $\mu_\chi$ in place of $\lambda$. Indeed, if there exist $p\in (1,\infty)$, $C>0$ and $\gamma>0$ such that for all $f\in L^p_{d/p}(\mu_\chi)$, $\|f\|_{L^p_{d/p}(\mu_\chi)}\leq 1$, $$\label{disprove} \int_G \Big( \exp(\gamma |f|^{p'})-\sum_{0\leq k<p-1} \frac{\gamma^k}{k!} |f|^{p'k} \Big)\, {\mathrm{d}}\mu_\chi \leq C\|f\|_{L^p(\mu_\chi)}^p,$$ then necessarily $\mu_\chi = \lambda$. To see this, assume that  holds for all $f\in L^p_{d/p}(\mu_\chi)$, $\|f\|_{L^p_{d/p}(\mu_\chi)}\leq 1$, with $\mu_\chi \neq \lambda$, i.e. $\chi \neq \delta$. We first prove that then  holds for all $f\in L^p_{d/p}(\mu_\chi)$, with no restriction on its norm (other than being finite). Recall, indeed, that for any $y\in G$ and $f\in L^p_{d/p}(\mu_\chi)$, denoting by $L_y$ the left translation by $y\in G$, one has $$\| L_y f\|_{ L^p_{d/p}(\mu_\chi)} = (\chi \delta^{-1})^{1/p}(y) \|f\|_{ L^p_{d/p}(\mu_\chi)}.$$ Since $(\chi \delta^{-1})^{-1/p}$ is a positive nonconstant character, it is unbounded; thus there exists $y\in G$ such that $$(\chi \delta^{-1})^{-1/p}(y) \geq \|f\|_{ L^p_{d/p}(\mu_\chi)}.$$ Equivalently, $(\chi \delta^{-1})^{1/p}(y) \|f\|_{ L^p_{d/p}(\mu_\chi)} \leq 1$, hence $\| L_y f\|_{ L^p_{d/p}(\mu_\chi)}\leq 1$. Thus, we may apply  to $L_yf$; and by a change of variable, one obtains  for $f$ where the constant $C$ does not depend on the norm of $f$. But  cannot hold without restriction on the norm of $f\in L^p_{d/p}(\mu_\chi)$. Indeed, let $\sigma\geq 1$ and consider $\sigma f$, which still belongs to $L^p_{d/p}(\mu_\chi)$ for any $\sigma$. Then, by  applied to $\sigma f$, $$\begin{aligned} \int_ G \sum_{k\geq p-1} \frac{\gamma^k}{k!} \sigma^{p'k} |f|^{p'k} \, {\mathrm{d}}\mu_\chi & \leq C \, \sigma^p \|f\|_{L^p(\mu_\chi)}^p.\end{aligned}$$ Since $$\int_ G \sum_{k\geq p-1} \frac{\gamma^k}{k!} \sigma^{p'k} |f|^{p'k} \, {\mathrm{d}}\mu_\chi \geq \int_ G \sum_{k\geq p} \frac{\gamma^k}{k!} \sigma^{p'k} |f|^{p'k} \, {\mathrm{d}}\mu_\chi \geq \sigma^{pp'} \int_ G \sum_{k\geq p} \frac{\gamma^k}{k!} |f|^{p'k} \, {\mathrm{d}}\mu_\chi,$$ one obtains $$\sigma^{p(p'-1)} \int_ G \sum_{k\geq p} \frac{\gamma^k}{k!} |f|^{p'k} \, {\mathrm{d}}\mu_\chi \leq C \|f\|_{L^p(\mu_\chi)}^p$$ for all $\sigma\geq 1$, which is a contradiction since $p(p'-1)>0$. [9]{} S. Adachi, K. Tanaka, [*Trudinger type inequalities in ${\mathbb{R}}^N$ and their best exponents*]{}, Proc. Amer. Math. Soc. 128 (2000), 2051–2057. D. R. Adams, N. G. Meyers, *Bessel potentials. Inclusion relations among classes of exceptional sets*, Indiana Univ. Math. J. 22 (1972/73), 873–905. A. Agrachev, U. Boscain, J.-P. Gauthier, F. Rossi, [*The intrinsic hypoelliptic Laplacian and its heat kernel on unimodular Lie groups*]{}, J. Funct. Anal. 256(8) (2009) 2621–2655. Th. Aubin, [*Problèmes isoperimétriques et espaces de Sobolev*]{}, J. Diff. Geom. 11 (1976), 573–598. T. Bruno, M. M. Peloso, A. Tabacco, M. Vallarino, *Sobolev spaces on Lie groups: embedding theorems and algebra properties*, J. Funct. Anal. 276 (2019), no. 10, 3014–3050. T. Bruno, M. M. Peloso, M. Vallarino, *Besov and Triebel–Lizorkin spaces on Lie groups*, Math. Ann. 377 (2020), 335–377. , *Potential spaces on Lie groups*, to appear on “Geometric aspects of harmonic analysis: a conference in honour of Fulvio Ricci", arXiv:1903.06415v1. A. Cotsiolis, N. K. Tavoularis, [*Sharp inequalities for Riesz, Bessel and Yukawa potential operators*]{}, Bull. Greek Math. Soc., 52 (2006), 99–118. M. G. Cowling, *Harmonic analysis on semigroups*, Ann. of Math. 117 (1983), 267–283. F.  Fontana, C. Morpurgo, [*Sharp exponential integrability for critical Riesz potentials and fractional Laplacians on ${\mathbb{R}}^n$*]{}, Nonlinear Anal. 167 (2018), 85–122. , [*Adams inequalities for Riesz subcritical potentials*]{}, Nonlinear Anal. 192 (2020), 32 pp. G. Gaudry, P. Sjögren, [*Singular integrals on Iwasawa NA groups of rank $1$*]{}, J. Reine Angew. Math. 479 (1996), 39–66. Y. Guivarc’h, *Croissance polynomiale et périodes des fonctions harmoniques*, Bull. Soc. Math. France 101 (1973), 333–379. W. Hebisch, G. Mauceri, S. Meda, *Spectral multipliers for Sub-Laplacians with drift on Lie groups*, Math. Z. 251 (2005), no. 4, 899–927. E. Hewitt, K. A. Ross, [*[“Abstract harmonic analysis. Vol. I. Structure of topological groups, integration theory, group representations. Second edition."]{}*]{} Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, Berlin-New York, 1979. N.  Lam, G. Lu, [*A new approach to sharp Moser-Trudinger and Adams type inequalities: a rearrangement-free argument*]{}, J. Differential Equations 255 (2013), 298–325. Y. Li, B. Ruf., [*A sharp Trudinger-Moser type inequality for unbounded domains in ${\mathbb{R}}^n$*]{}, Indiana Univ. Math. J. 57 (2008), 451–480. E. Lieb, [*Sharp constants in the Hardy-Littlewood-Sobolev and related inequalities*]{}, Ann. of Math., 118 (1983) 349–374. S. Meda, [*On the Littlewood-Paley-Stein g-function*]{}, Trans. Amer. Math. Soc. 347 (1995), no. 6, 2201–2212. J. Moser, *A sharp form of an inequality by N. Trudinger.* Indiana Univ. Math. J. 20 (1970/71), 1077–1092. T. Ogawa, [*A proof of Trudinger’s inequality and its application to nonlinear Schördinger equations*]{}, Nonlinear Anal. 14 (1990), 765–769. T. Ozawa, [*On critical cases of Sobolev’s inequalities*]{}, J. Funct. Anal. 127 (1995), 259–269. M. M. Peloso, M. Vallarino, *Sobolev algebras on unimodular Lie groups*. Calc. Var. Partial Differential Equations 57 (2018), 34 pp. M. Ruzhansky, N. Yessirkegenov, *Critical Sobolev, Gagliardo–Nirenberg, Trudinger and Brezis–Gallouet–Wainger inequalities, best constants, and ground states on graded groups*, arXiv:1709.08263. R. S. Strichartz, *A note on Trudinger’s extension of Sobolev’s inequalities*. Indiana Univ. Math. J. 21 (1971/72), 841–842. G. Talenti, [*Best constants in Sobolev inequality*]{}, Ann. Mat. Pura Appl. 110 (1976) 353–372. N. S. Trudinger, *On imbeddings into Orlicz spaces and some applications.* J. Math. Mech. 17 (1967) 473–483. N. Th. Varopoulos, *Analysis on Lie groups*, J. Funct. Anal. 76 (1988), no. 2, 346–410. N. Th. Varopoulos, T. Coulhon, L. Saloffe-Coste, *Analysis and geometry on groups*, Cambridge Tracts in Mathematics, 100. Cambridge University Press, Cambridge, 1992. A. Zygmund, “Trigonometric series, vol. II”, Cambridge University Press, Cambridge, (1959). [^1]: [*Math Subject Classification*]{} 26D10, 43A80, 46E35 [^2]: All authors were partially supported by the grant PRIN 2015 [*Real and Complex Manifolds: Geometry, Topology and Harmonic Analysis*]{}, and by the grant [*Fractional Laplacians and subLaplacians on Lie groups and trees*]{} of the Gruppo Nazionale per l’Analisi Matematica, la Probabilità e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM). T. Bruno acknowledges support by the Research Foundation – Flanders (FWO) through the postdoctoral grant 12ZW120N

ArXiv

--- abstract: | A system model is developed where the criterion to partition the world into a system and a rest is based on the functional relation between its states. This approach implies that the gestalt of systems becomes very dynamic. Especially interactions between systems may create temporary super systems ”on the fly”. The reference to the function notion establishes the link to computation. It is shown that being computable is a system property which is compositional in an abstract mathematical sense but is, because of recursion, emergent from a computational point of view. Based on the distinction between the intended determinism versus nondeterminism of the interaction, two classes of system interactions are distinguished: composition and cooperation. Composition means that by interaction super systems are created and the interacting systems become subsystems. Cooperation means that systems sensibly interact loosely without creating any identifiable super system function and therefore without super system creation from the perspective of the interacting systems. Cooperative system interactions are described with the help of the protocol notion based on shared states, interpreted as the stateful exchange of characters via Shannon channels between roles - the projection of systems onto the interaction channels. It is shown that roles can be internally coordinated by a set of rules, leading to a very flexible process notion which unfolds its full potential only in a new type of execution environment capable of executing spontaneous transitions. The system model has immediate implications for componentization which can be viewed as an attempt to either hide functional recursion and provide only functionality whose composition behavior is easy to understand, or to provide loosely coupled interactions via protocols. To be complete, component models should therefore at least support three general classes of components: one for loose coupling, one for hierarchical composition, and one for pipes. author: - 'Johannes Reich, [email protected]' bibliography: - '../../../../bib/soziologie.bib' - '../../../../bib/theoret\_med.bib' - '../../../../bib/philosophy.bib' - '../../../../bib/informatics.bib' - '../../../../bib/develop.bib' title: 'Composition, Cooperation, and Coordination of Computational Systems' --- Introduction ============ Civil, mechanical, electrical, or software engineering - they all deal with the design of systems. The term ”system engineering” was coined in the Bell Laboratories in the 1940s (e.g. [@Buede2011_Engineering]). According to Kenneth J. Schlager [@Schlager1956], a main driver of the field was the complexity of the composition behavior of components. Often seemingly satisfactory working components did not result in satisfactorily working systems. The focus on systems was transferred to other disciplines like biology (e.g. [@Bertalanffy1968_GST]) or sociology (e.g. [@Parson1951_SocialSystems]) Especially with the advent of cyber physical systems (e.g. [@Lee2008_CPS; @Alur2015_Principles]), also the software engineering discipline articulates the importance of a unifying view (e.g. [@Sifakis2011_Vision]) as the different disciplines complement each other more and more. As a base for such a unifying view, I propose a common system notion which is grounded on the notion of a system’s functionality, a notion that is tightly related to the concepts of computability and determinism of computer science. Although systems can be described as isolated entities, their main purpose is to interact. And as will be shown in this article, due to their interactions at least two very different system relations arise: composition and cooperation. System composition means that by their interaction super systems are created and the interacting systems become subsystems. As the system notion is based on the function notion, the concept of system composition can be traced back to the concept of composition of functions which is at the heart of computability. Especially for software components, the borrowed distinction between non-recursive and recursive system composition becomes important. As is illustrated in Fig. \[fig\_system\_composition\_general\], system composition leads to hierarchies of systems which are determined by the ’consists of’-relation between supersystems and their subsystems. ![\[fig\_system\_composition\_general\]An example of a hierarchically composed system, namely system 1. It consists of 3 subsystems which by themselves consist of a couple of further subsubsystems.](fig/system_composition_general){width="7cm"} Quite contrasting, system cooperation means that systems interact sensibly somehow ”loosely” without such super system creation. In essence, this means that no super system function can be identified as the interaction remains nondeterministic. This is typically the case in interaction networks. As an example, Fig. \[fig\_network\_of\_relations\] sketches a cutout of a network of business relations between a buyer, a seller, its stock, a post and a bank. In these networks none of the participants is in total control of all the other participants: the participants’ interactions don’t, in general, determine the participants’ actions. These networks are open in the sense that we will never be able to describe them completely. ![\[fig\_network\_of\_relations\]A cutout of an open business network.](fig/network_of_relations_v01){width="7cm"} The distinction between system composition and cooperation suggests a component model which distinguishes between compositional and cooperative components. The relevance of system composition for component oriented development has already been pointed out for example by Kung-Kiu Lau and Zheng Wang [@LauWang2007_Software]. The desire to use compositional components as building blocks with a strict hierarchical composition behavior leads to the requirement that compositional components represent a complexity border hiding all functionality based on recursion. However, the relevance of system cooperation for component models seems to be less well understood, although the relevant entity to describe these interactions is well known since the 1960s: protocols (e.g. [@Scantlebury1967; @Zimmermann1980; @Holzmann1991]). Protocols describe nondeterministic, asynchronous, and stateful interactions we find in interaction networks showing no formal hierarchy [@Reich2015_kvsi]. In their overview on component models Ivica Crnkovic, Severine Sentilles, Aneta Vulgarakis and Michel R. V. Chaudron [@CrnkovicSentilles2011_Classification] distinguish between ”operation based” and ”port based” interface support and show thereby that many currently important component models indeed do not support protocol declarations. As demonstrated in this article, an important consequence of the recursive nature of computation is that the identification of discrete systems is not computable. The structure of the article is as following. In section \[s\_system\_definition\], a system definition is provided and the concept of computable functionality is recapitulated to give the whole discussion a sound base. In section \[s\_system\_composition\], the issue of system composition is investigated. A distinction is made between the operation of system composition and the property of being compositional for some system properties. Sequential, parallel and recursive Loop- and While-compositions are identified. In section \[s\_system\_cooperation\], the cooperation of systems as the nondeterministic form of their interaction is examined. Cooperative system interactions are described with the help of the protocol notion based on shared states, interpreted as the stateful exchange of characters via Shannon channels between roles - the projection of systems onto the interaction channels. In section \[s\_cooperating\_systems\], the focus is shifted to the inner structure of cooperating systems which are named processes. It is shown that roles can be internally coordinated by a set of rules, leading to a very flexible process notion. Related work is presented in section \[s\_related\_work\]. Here, I dwell on abstract data types, objects, reactive systems, and the system models of Manfred Broy and Rajeev Alur as well as the BIP component framework of Joseph Sifakis et al. In the final section \[s\_discussion\], the concepts and results of this article are discussed with respect to their possible implications for software engineering and componentization as well as for other areas of science. Preliminaries ------------- Throughout this article, elements and functions are denoted by small letters, sets and relations by large letters and mathematical structures by large calligraphic letters. The components of a structure may be denoted by the structure’s symbol or, in case of enumerated structures, index as subscript. The subscript is dropped if it is clear to which structure a component belongs. Character sets and sets of state values are assumed to be enumerable if not stated otherwise. For any character set or alphabet $A$, $A^\epsilon := A\cup\{\epsilon\}$ where $\epsilon$ is the empty character. For state value sets $Q$, $Q^\epsilon := Q\cup\{\epsilon\}$ where $\epsilon$ is the undefined value. If either a character or state value set $A = A_1 \times \dots \times A_n$ is a Cartesian product then $A^\epsilon = A_1^\epsilon \times \dots \times A_n^\epsilon$. Elements of character sets or state value sets can be vectors. There will be no notational distinction between single elements and vectors. However, a state vector $(p_1, \dots, p_n)$ where $p_k$ belongs to structure ${\cal A}_k$ is written as $\vec{p}$ and the change of this vector in a position $k$ from $p$ to $q$ is written as $\vec{p} \left[\frac{q}{p}, k\right]$. An n-dimensional vector of characters with the $k$-th component $v$ and the rest $\epsilon$ is written as $\vec{\epsilon}[v,k]=(\epsilon_1, \dots, \epsilon_{k-1}, v, \epsilon_{k+1}, \dots, \epsilon_n)$. The power set of a set $A$ is written as $\wp(A)$. System definition \[s\_system\_definition\] =========================================== What allows us to draw these nice little boxes to represent systems versus the rest of the world? There seems to be a consensus (e.g. [@HarelPnueli1985_Reactive; @Broy2010_Logical; @Sifakis2011_Vision; @ISO_15288-2014; @IEC_60050]) that a system separates an inside from the rest of the world, the environment. IEC 60050 [@IEC_60050] defines a system (351-42-08) as a ”set of interrelated elements considered in a defined context as a whole and separated from their environment”. So, to gain a well defined system model we have to answer the two questions: What gets separated? And: what separates? What gets separated? It’s time dependent properties each taking a single out of a set of possible values at a given time. These time dependent functions are commonly called ”state variables” and the values are called ”states” [@IEC_60050]. However, in my opinion the term ”variable” should better be restricted to the domain of descriptions, denoting the state functions. To make its meaning unique where necessary, I therefore will qualify the term ”state” either with ”function” or ”value”. If it is not qualified, I usually refer to its meaning as state function. So, a state (function) $s$ in this sense is a function from the time domain $T$ to some set of state values or alphabet $A$ or to a set of state values or alphabet $A^\epsilon$, including the empty character or unknown value $\epsilon$. Some of the states may not be directly accessible from the outside, these are the system’s inner states. We usually assume that the inner states of a system are well defined, i.e. cannot attain the undefined value $\epsilon$ - but that states, representing the input or output of a system may attain the empty character $\epsilon$. What separates? Already in IEC 60050 [@IEC_60050] it is noted that a system is generally defined with the view of achieving a given objective, for example by performing a definite function. The key idea is now to use this function, that is the unique functional relation between the values of states at given time values to separate the state of a system from the rest of the world. Such a relation implies causality and a time scale. Depending on the class of system function or time, different classes of systems can be identified. Important classes of functions are computable functions, finite functions and analytic functions. Important classes of times are discrete and continuous times[^1]. A first example, intentionally not drawn from informatics, is a serial RLC-circuit. It consists of a resistor $R$, an inductor $L$, and a capacitor $C$ and comes with a continuous time domain. Hence, it is described by a differential equation. I chose the current $I$ and the voltage $U_L$ across $L$ as inner state, and the external voltage $U_{ext}$ as input state. With the knowledge that the current is the same at every point in the circuit and that all voltages add up, the following two differential equations result for the time evolution of the two inner states: $$\frac{dI(t)}{dt} = \frac{U_L(t)}{L} \quad \mbox{and} \quad \frac{dU_L(t)}{dt} = \frac{dU_{ext}(t)}{dt} - \frac{R}{L}U_L(t) - \frac{I(t)}{C}$$ In the following I will focus on discrete times. An example for a discrete system is a finite multiplier with its input state $in_1, in_2$ and its output state $out$, all ranging within $\{-2^{31}\dots 2^{31}-1\}$ and the system function $out(t')=f(in_1(t),in_2(t)) = in_1(t)*in_2(t) \,mod\,2^{31}$ . Hence, the formal system definition is as following. \[def\_system\] Be $Q$ a non-empty set of internal state values and $I$ and $O$ the possibly empty sets of input and output state values (alphabets). The state $(q, in, out):T\rightarrow Q \times I \times O$ is said to form a (discrete) [*system*]{} for time step $(t, t')$, if it is aggregated by a function $f:Q \times I \rightarrow Q \times O$ with $f =(f^{int}, f^{ext})$ such that $${q(t') \choose out(t')} = {{{f^{int}}(q(t), i(t))}\choose {f^{ext}}(q(t), i(t))}\,.$$ $t'$ is also called the successor time of $t$ with respect to $f$. The [*system time*]{} is the structure ${\cal T} = (T, succ)$ with $T$ the enumerable set of all time values and $succ:T\rightarrow T$ with $succ(t) = t'$ is the successor function. For an infinite number of time values, $T$ can be identified with the set of natural numbers $\mathbb{N}$. For a finite number of time values, $succ(t_{max})$ is undefined. I also write $t+n = succ^n(t)$. A [*system*]{} thereby is characterized by a tuple ${\cal S}=(T, succ, Q, I, O, q, in, out, f)$. If $Q$, $I$ and $O$ are finite then the system is said to be finite. If $T$ is enumerable, the system is said to be discrete. A system with no or only constant internal state is said to be stateless. ![Diagram of a system as defined in Def. \[def\_system\]. The states $(q, in, out)$ become separated from the rest of the world by the system function $f$. The asymmetric boxes of the input and output state symbolizes their asymmetric use in the system model. The input state $in$ provides input to the system function but cannot be influenced by it. Both, input as well as output state can belong to different systems at a given point in time. The output state can be another system’s input state and vice versa. In case, the inner state $q$ or the system function are of no relevance for illustrative purposes, they can be omitted.[]{data-label="fig_system"}](fig/system_v021.png){width="6cm"} With definition \[def\_system\] we can declare the meaning of these simple system diagrams as shown in Fig \[fig\_system\]. As we will see, this kind of graphical representation is so intuitive, that it is perhaps one of the most stated ”lies” in computer science. Too often, these kind of graphical boxes claim to represent ”systems”, but no system function with respect to the separated state functions can be identified nor probably any other formal system criteria can be fulfilled. With the definition of functional equivalence, we create equivalence classes of isomorphic systems. To do so, we have to extend the system functions to operate also on $O$ and map onto $I$. Be ${\cal S}_1$ and ${\cal S}_2$ two systems and the extended system functions $f_k^*:Q_k\times I_k^\epsilon \times O_k^\epsilon \rightarrow Q_k\times I_k^\epsilon \times O_k^\epsilon$ with $f_k^*(q_k, i_k, o_k) = \left(f^{(int)}_k(q_k,i_k), \mathbb{1}_k(i_k), f^{(ext)}_k(q_k,i_k)\right)$ with $k=1,2$. ${\cal S}_1$ and ${\cal S}_2$ are isomorphic or functionally equivalent (written as ${\cal S}_1 \cong {\cal S}_2$) if two bijective functions $\phi:Q_1\times I_1^\epsilon \times O_1^\epsilon \rightarrow Q_2\times I_2^\epsilon \times O_2^\epsilon$ and $\psi:T_1 \rightarrow T_2$ exists such that $\phi(f_1^*(q_1,i_1,o_1)) = f_2^*(\phi(q_1,i_1,o_1))$ for each $(q_1, i_1, o_1)\in Q_1 \times I_1^\epsilon \times O_1^\epsilon$ and $\psi(succ_1(t)) = succ_2(\psi(t))$ for all $t\in T_1$. Relating to these equivalence classes, we can restrict ourselves to systems operating on representations of natural numbers or a subset thereof. Clocked and unclocked systems ----------------------------- An important differentiator for discrete systems is whether they are clocked or not. A clocked system applies its system function when the clock signal triggers it. An unclocked system applies its system function when an input arrives[^2]. A clocked system thereby can be viewed as an unclocked system with an additional clock ’tick’ as input. It can, in general, react if not to but on an empty input. This is the usual case for systems represented by computer programs running in the same address space. As an unclocked system reacts only to an applied input signal, an empty input cannot trigger any processing. In this case, an empty input of a transition stands for a spontaneous action and an empty output represents no output at all. This is the usual case for systems interacting remotely without access to any common clock signal. Computable functionality ------------------------ Be $F_n$ the set of all functions on natural numbers with arity $n$ and we assume a set of initial computable functions (the successor, the constant and the identity function). Then, based on work of Kurt Gödel, Stephen Kleene [@Kleene1936] showed that there are three rules to create all computable functions: 1. [***Comp:***]{} Be $g_1, \dots, g_n \in F_m$ computable and $h\in F_n$ computable, then $f = h(g_1, \dots, g_n)$ is computable.\[computation\_1st\_rule\] 2. [***PrimRec:***]{} Are $g\in F_n$ and $h\in F_{n+2}$ both computable and $a\in \mathbb{N}^n$, $b\in \mathbb{N}$ then also the function $f\in F_{n+1}$ given by $f(a, 0) = g(a)$ and $f(a, b+1) = h(a, b, f(a,b))$. \[computation\_2nd\_rule\] is computable. 3. [***$\mu$-Rec:***]{} Be $g\in F_{n+1}$ computable and $\forall a\exists b$ such that $g(a,b) = 0$ and the $\mu$-Operation $\mu_b[g(a,b) = 0]$ is defined as the smallest $b$ with $g(a,b) = 0$. Then $f(a) = \mu_b[g(a,b) = 0]$ is computable.\[computation\_3rd\_rule\] System composition \[s\_system\_composition\] ============================================= Systems can be composed from other systems. To compose, systems have to interact. In this theory systems interact by sharing common I/O-states: one system’s output state is another system’s input state. I call such a common state an ”[*ideal Shannon channel*]{}”, following the concept of Claude Shannon [@Shannon1948] that a channel reproduce at one point a state value selected at another point. As he showed, it is possible to reproduce any state value out of a finite set of state values over a channel even in the presence of noise up to any desired level of accuracy if sufficient time is available. Compositionality ---------------- A precise system model allows for a precise definition of the notion of compositionality. Following an idea of Arend Rensink[^3], I distinguish between composition of systems and the property of being compositional for the properties of the systems. \[def\_compositionality\] Be $S$ the set of all systems according to Def. \[def\_system\] and be $comp:S^n \rightarrow S$ a composition operator. A system property $\alpha$ is a partial function $S \rightarrow A$ which attributes values of some attribute set $A$ to systems ${\cal S}\in S$. A property $\alpha$ of a composed system ${\cal S} = comp({\cal S}_1, \dots, {\cal S}_n)$ is called ”compositional” if a function $comp_{\alpha}: A^n \rightarrow A$ exists that fulfills the following relation: $$\alpha\left(comp({\cal S}_1, \dots, {\cal S}_n)\right) = comp_\alpha\left(\alpha({\cal S}_1), \dots, \alpha({\cal S}_n)\right)$$ Otherwise the system property is called ”emergent”. If $comp_{\alpha}$ is computable, then the property is called ”computable compositional”, otherwise it is called ”computational emergent”[^4]. In other words, compositional properties of the composed system result exclusively from the respective properties of the parts. Mathematically, $\alpha$ is a homomorphism. Emergent properties may result also from other properties $\alpha_i$ of the parts [@Reich2001] as $\alpha\left(comp({\cal S}_1, \dots, {\cal S}_n)\right) = comp_\alpha\left(\alpha_1({\cal S}_1), \dots, \alpha_n({\cal S}_n)\right))$. A simple example of a compositional system property is the mass $m({\cal S})$ of a physical system ${\cal S}$. The total mass of the composed system is simply the sum of the masses of the single systems: $m\left(comp({\cal S}_1, \dots, {\cal S}_n)\right) = m({\cal S}_1) + \dots + m({\cal S}_n)$. An example for an emergent property in this sense is the resonance of the RLC-circuit. It’s emergent and not compositional because the building blocks don’t have a resonance frequency - only the RLC-circuit has. Sequential system composition \[ss\_sequential\_system\_composition\] --------------------------------------------------------------------- First, I define what I mean by saying that two systems work sequentially and then I show that under these circumstances, a super system can always be identified. Sequential system composition means that one system’s output is completely fed into another system’s input, see Fig. \[fig\_system\_composition\_sequential\]. ![Diagram of a sequential system composition as defined in Def. \[def\_system\_composition\_sequential\]. The output state of the first system $out_1$ is identical to the input state $in_2$ of the second system.[]{data-label="fig_system_composition_sequential"}](fig/system_composition_sequential_v02.png "fig:"){width="12cm"}\ \[def\_system\_concatenation\] Two discrete systems ${\cal S}_1$ and ${\cal S}_2$ with $O_1 \subseteq I_2$ and $\tau_i, \tau_i' \in T_i$ are said to ”[*work sequentially*]{}” or to be ”[*concatenated*]{}” at $(\tau_1, \tau_2)$ if $out_1(\tau_1') = in_2(\tau_2)$. \[def\_system\_composition\_sequential\] Let ${\cal S}_1$ and ${\cal S}_2$ be two systems that work sequentially at time $(\tau_1, \tau_2)$. Then the states $(q, in, out)$ with $q = (q_1, out_1, q_2)$, $in = in_1$, and $out=out_2$ form a system for time step $((\tau_1, \tau_2), (\tau_1', \tau_2'))$. To prove the proposition we have to identify a system function and a time function for the composed system. Based on the concatenation condition $out_1(\tau_1') = in_2(\tau_2)$ we can replace $in_2(\tau_2)$ by $out_1(\tau_1') = f^{ext}_1(q_1(t_1), i_1(t_1))$. We get $$\begin{gathered} \label{eq_sequential_composition} %\begin{split} \left( \begin{array}{c} q_1(\tau_1'), out_1(\tau_1'), q_2(\tau_2') \\ out_2(\tau_2') \end{array}\right) = \\ \left(\begin{array}{c} f_1 \left(q_1(\tau_1), i_1(\tau_1)\right), f^{int}_2 \left(q_2(\tau_2), f^{ext}_1(q_1(\tau_1), i_1(\tau_1)), out_2(\tau_2)\right)\\ f^{ext}_2 \left(q_2(\tau_2), f^{ext}_1(q_1(\tau_1), i_1(\tau_1)), out_2(\tau_2)\right) \\ \end{array}\right) %\end{split} \end{gathered}$$ where the right hand side depends exclusively on $\tau_1$ and $\tau_2$ and the left hand side, the result, exclusively on $\tau_1'$ and $\tau_2'$. The composed system $C$ exists with the time values $T_C = \{0,1\}$. With the functions $map_i: T_C \rightarrow T_i$, $(i=1,2)$ with $map_i(0) = \tau_i$ and $map_i(1)=\tau_i'$, the states of $C$ are: $out_C(\tau_C) = out_2(map_2(\tau_C))$, $in_C(\tau_C) = in_1(map_1(\tau_C))$, and $q_C(\tau_C) = (q_1(map_1(\tau_C)), out_1(map_1(\tau_C)), q_2(map_2(\tau_C)))$. The concatenation condition $out_1(\tau_1') = in_2(\tau_2)$ implies that there is an important difference between the inner and outer time structure of sequentially composed systems. To compose sequentially, the second system can take its step only after the first one has completed its own. Thus, the concatenation operator $\circ_{(\tau_1, \tau_2, out_1, in_2)}$[^5] composes systems such that ${\cal S} = {\cal S}_2 \circ_{(\tau_1, \tau_2, out_1, in_2)} {\cal S}_1$ is the concatenated or sequentially composed system for one time step. The extension to more then two concatenated systems or time steps is straight forward. In the special case, where $f^{ext}_i: I_i \rightarrow O_i$ is stateless in both systems, the concatenation of these stateless systems results in the simple concatenation of the two system functions: $out_2(\tau_2') = (f^{ext}_2 \circ f^{ext}_1)(in_1(\tau_1))$ Finally, the following proposition holds: For sequential system composition, the system’s functionality composes compositionally. Identifying $\alpha$ of the compositionality definition \[def\_compositionality\] with the mapping from the domain of the systems to the domain of the system functions, we see that $f_{{\cal S}_2 \circ {\cal S}_1} = \alpha({\cal S}_2 \circ {\cal S}_1) = comp_\alpha(\alpha({\cal S}_1), \alpha({\cal S}_2)) = comp_\alpha(f_{{\cal S}_1}, f_{{\cal S}_2})$. The wanted operator $comp_\alpha$ can be directly deduced from the system function as defined in Eq. \[eq\_sequential\_composition\]. Fig. \[fig\_sequential\_vs\_reciprocal\_coupling\] shows an example of an additional interaction, influencing system composition behavior and illustrates the suggestibility of visual descriptions. ![This is an example of visual suggestibility. The left hand side suggest a sequential system composition with an additional feedback channel. The right hand side suggests that we have two reciprocally coupled systems. Actually, both diagrams convey exactly the same information.[]{data-label="fig_sequential_vs_reciprocal_coupling"}](fig/process_composition_I_v01.png "fig:"){width="12cm"}\ Especially important, neither diagram of Fig. \[fig\_sequential\_vs\_reciprocal\_coupling\] conveys any information about the timing of the coupling. If the diagrams are complete and there is no additional input to the systems, then, the resulting ”thing” is no system any longer, as it has neither a dedicated input nor a dedicated output state. Actually, if both systems are unclocked, nothing will happen anyhow. How can it be that linking the output of one system to the input of another system results in sequential composition but creating another link destroys that? Actually, it does not ”destroy” the sequential composition. The additional link just shows the importance of the timing, that is the assignment of time steps to the evaluation of the system function. Assuming clocked systems, we can still see that ${\cal S}_1$ is doing something and hands over its character to ${\cal S}_2$ and say that the interaction resulted in a sequential system composition for these two time steps. But in the next step, its again ${\cal S}_1$ doing something. So for these three consecutive actions, we do not have a sequential composition anymore. What do we have instead? There is no simple answer to this seemingly simple question, as I will discuss in section \[ss\_recursive\_system\_composition\] and \[s\_system\_cooperation\]. Parallel system composition --------------------------- Parallel processing systems can also be viewed as one system if there is a common input to them and the parallel processing operations are independent and therefore well defined as is illustrated in Fig. \[fig\_system\_composition\_parallel\]. ![Diagram of a parallel system composition as defined in Def. \[def\_system\_composition\_parallel\]. The input of both systems is identical.[]{data-label="fig_system_composition_parallel"}](fig/system_composition_parallel_v02.png "fig:"){width="8cm"}\ \[def\_system\_composition\_parallel\] Two discrete systems ${\cal S}_1$ and ${\cal S}_2$ with $I_1=I_2$ are said to ”[*work in parallel*]{}” at $(\tau_1, \tau_2)$ if $q_1 \neq q_2$, $out_1 \neq out_2$ and $\tau_1\in T_1$, $\tau_2\in T_2$ exist such that $in_1(\tau_1) = in_2(\tau_2)$. Let ${\cal S}_1$ and ${\cal S}_2$ be two systems that work in parallel at $(\tau_1, \tau_2)$. Then the state $(q, in, out)$, with $q = (q_1, q_2)$ $in=in_1=in_2$, $out = (out_1, out_2)$ form a system for time step $((\tau_1, \tau_2), (\tau_1', \tau_2'))$. To prove, we have to provide a system function for the composed system $C$ at times $T_C=\{0,1\}$, which is simply $f_C: (Q_1\times Q_2) \times (I_1^\epsilon \times I_2^\epsilon) \rightarrow (Q_1^\epsilon\times Q_2^\epsilon) \times (O_1^\epsilon\times O_2^\epsilon)$ with $$\begin{gathered} f_C(\tau_C') = \\ {f^{int}_1(q_1(map_1(\tau_C)), in(map_1(\tau_C))), f^{ext}_1(q_2(map_2(\tau_C)), in(map_1(\tau_C))) \choose f^{int}_2(q_1(map_1(\tau_C)), in(map_1(\tau_C))), f^{ext}_2(q_2(map_2(\tau_C)), in(map_1(\tau_C)))}\end{gathered}$$ The composed system’s time step does not depend on the exact sequence of the time steps of the subsystems. If both systems don’t perform their step at the same time compared to an external clock then the composed system does not realize its time step in a synchronized manner. If this synchronization takes place, then there is effectively no difference between the inner and outer time structure of parallel composed systems. The parallelization operator $||_{(\tau_1, \tau_2, in_1, in_2)}$ can be defined such that ${\cal S} = {\cal S}_2 ||_{(\tau_1, \tau_2, in_1, in_2)} {\cal S}_1$ is the parallel working system for one time step. The extension to more than two parallel working systems or two time steps is again straight forward. As with sequential system composition, for parallel system composition, the system’s functionality composes compositionally. The proof is similar to that for sequential system composition. Combining sequential and parallel system composition ---------------------------------------------------- As the application of an operation $f$ preceding two parallel operations $g$ and $h$ is equivalent to the parallel execution of $g$ after $f$ and $h$ after $f$, together with some bookkeeping on time steps and i/o-states, the following proposition is easy to prove: For three systems ${\cal P}_1$, ${\cal P}_2$ and ${\cal S}$ where ${\cal P}_1$ and ${\cal P}_2$ are parallel systems and ${\cal P}_1 || {\cal P}_2$ and ${\cal S}$ are sequential systems, then the two systems $({\cal P}_1 || {\cal P}_2) \circ {\cal S}$ and $({\cal P}_1 \circ {\cal S}) || ({\cal P}_2 \circ {\cal S})$ are functional equivalent, that is $({\cal P}_1 || {\cal P}_2) \circ {\cal S}\cong ({\cal P}_1 \circ {\cal S}) || ({\cal P}_2 \circ {\cal S})$. In other words, the right distribution law holds for parallel and sequential composition with respect to functional equivalence. As sequential system composition is non-commutative, it follows that the left distribution law does not hold: ${\cal S} \circ ({\cal P}_1 || {\cal P}_2) \ncong ({\cal S} \circ {\cal P}_1) || ({\cal S} \circ {\cal P}_2)$. Finally I point out that combining parallel and sequential system composition as ${\cal S} \circ ({\cal P}_1 || {\cal P}_2)$ for stateless systems results in a combination of their system functions as $f_{\cal S} (f_{{\cal P}_1}, f_{{\cal P}_2})$, which is exactly the way, the first rule [*Comp*]{} for computable functions was defined. Hence, from a formal point of view, for computational systems, sequential together with parallel system composition represent the first level to compose computable functionality. Recursive system composition \[ss\_recursive\_system\_composition\] ------------------------------------------------------------------- From an interaction perspective we could say that recursive interaction means that at least two systems send each other characters back and forth. But this kind of character ping-pong does not say very much about the created logical system relations as is illustrated in Fig \[fig\_system\_interacting\_reality\]. ![In the middle, two systems interact by sending each other characters. What does that mean from a system perspective? Does a loop- or a while-system result? Or is the interaction loosely coupled? Or is it only pseudo recursive? Or does something else happens?[]{data-label="fig_system_interacting_reality"}](fig/system_interacting_reality_v20.png "fig:"){width="12cm"}\ As sequential together with parallel system composition represent the first level to compose computable functionality, it is interesting to investigate how the system composition operators have to look like for the other two rules [*PrimRec*]{} and [*$\mu$-Rec*]{}. For [*PrimRec*]{} we have to sequentially combine an iterator together with a system which has its output as its own input, see Fig. \[fig\_system\_composition\_loop\]. Additionally, some mechanism has to terminate the computation after the $n$-th step. \[def\_system\_composition\_loop\] Be $g$ and $h$ two functions as defined in [*PrimRec:*]{}. Be ${\cal I}$ a discrete clocked system with with $I=\emptyset$, $Q = \{0, \dots, n+1\}$, $O = \{0, \dots, n\}$, $f:Q \rightarrow Q\times O$ with $(q',out') = f(q) = (q+1, q)$, for $t<n$. And ${\cal S}$ is a discrete (not necessarily clocked) system with $I = \{0, \dots, n\} \times \mathbb{N}$, $Q=\{a\}$, $O=\mathbb{N}$, $f:Q\times I \rightarrow O$ with $f(q, in) = h(q, in)$ are said to work in a for-loop for $n$ times if $(\tau_{\cal I}, \tau_{\cal S})$ exist, such that $out_{\cal I}(\tau_{\cal I}+i+1) = {in_1}_{\cal S}(\tau_{\cal S}+i)$, $out_{\cal S}(\tau_{\cal S}+i) = {in_2}_{\cal S}(\tau_{\cal S}+i)$. \[prop\_composition\_loop\] Let ${\cal I}$ and ${\cal S}$ be two systems as defined in Def. \[def\_system\_composition\_loop\] and $(\tau_{\cal I}, \tau_{\cal S}) \in T_{\cal I}\times T_{\cal S}$ such that $out_{\cal I}(\tau_{\cal I}) = 0$, $q_{\cal S}(\tau_{\cal S}) = a$, $in_{{\cal S},2}(\tau_{\cal S}) = out_{\cal S}(\tau_{\cal S}) = g(a)$, and $in_{{\cal L},2} = n$. Then the state $in_{\cal L} = q_{\cal S}$, $out_{\cal L} = out_{\cal S}$ form a system for time step $((\tau_{\cal I}, \tau_{\cal S}), (\tau_{\cal I}+n, \tau_{\cal S}+n))$. We also write ${\cal L} = Loop_{a,n}(I, S)$ That the loop system calculates $f(a,n)$ of [*PrimRec*]{} in the $n-th$ time step can be deduced from the following table.\ --------------------- --------------------------------- --------------------- -------------------------------------- -------------------------------------- $time_{\cal I}$ $out_{\cal I}(\tau_{\cal I}+i)$ $time_{\cal S}$ $in_{{\cal S},2}(\tau_{\cal S} + i)$ $in_{{\cal S},2}(\tau_{\cal S} + i)$ = $out_{\cal S}(\tau_{\cal S} + i)$ $\tau_{\cal I} + 0$ 0 - - - $\tau_{\cal I} + 1$ 1 $\tau_{\cal S} + 0$ 0 f(a,0) = g(a) $\tau_{\cal I} + 2$ 2 $\tau_{\cal S} + 1$ 1 f(a,1) = h(a, 0, g(a)) $\tau_{\cal I} + 3$ 3 $\tau_{\cal S} + 2$ 2 f(a,2) = h(a, 1, h(a,0,g(a))) $\dots$ $\dots$ $\dots$ $\dots$ $\dots$ --------------------- --------------------------------- --------------------- -------------------------------------- -------------------------------------- ![Diagram of a loop system composition as defined in Def. \[def\_system\_composition\_loop\]. []{data-label="fig_system_composition_loop"}](fig/system_composition_loop_v10.png "fig:"){width="10cm"}\ Please note that there is no explicit end of the loop counter, but the loop has to be terminated by some external mechanism - or to put it the other way around. If some external mechanism guarantees termination of the loop after $n$ steps, then we can always identify a system with a well defined time step. Also, the following proposition holds: \[prop\_loop\_compositionality\] For a loop-system according to Def. \[def\_system\_composition\_loop\], the system function composes compositionally. Be $\alpha({\cal S})$ the system function, then compositionality means that the system function of the loop system of ${\cal I}$ and ${\cal S}$ can be constructed from the system functions of ${\cal I}$ and ${\cal S}$ together with the additional knowledge of the system construction which comprise $n$ and the state identity $out_{\cal I} = {in_1}_{\cal S}$, $out_{\cal S} = {in_2}_{\cal S}$ which can be incorporated into $comp_\alpha$. This is the case as can be seen from the proof of proposition \[prop\_composition\_loop\]. For [*$\mu$-Rec*]{}, we again combine an iterator together with another system, whose only task is to calculate the zeros of $g$ as is illustrated in Fig. \[fig\_system\_composition\_while\]. Again, some mechanism has to explicitly stop the computation after the zero has been computed. ![Diagram of a while system composition as defined in Def. \[def\_system\_composition\_while\]. []{data-label="fig_system_composition_while"}](fig/system_composition_while_v10.png "fig:"){width="10cm"}\ \[def\_system\_composition\_while\] Be $g$ a function as defined in [*$\mu$-Rec*]{}. Two discrete Systems ${\cal I}$ with $I_{\cal I} = \emptyset$, $Q_{\cal I} = O_{\cal I} = \mathbb{N}$, $f:Q_{\cal I} \rightarrow Q_{\cal I} \times O_{\cal I}$ with $(q_{\cal I}', out_{\cal I}') = f_{\cal I}(q_{\cal I}) = (q_{\cal I}+1, q_{\cal I})$ and ${\cal S}$ with $I_{\cal S} = \mathbb{N}$ , $Q_{\cal S} = \{a\}$, $O_{\cal S} = \mathbb{N}$, $f_{\cal S}:I_{\cal S} \times Q_{\cal S} \rightarrow O_{\cal S}$ with $f_{\cal S}(in_{\cal S}, q_{\cal S}) = g(q_{\cal S}, in_{\cal S})$ are said to work in a while-loop if $(\tau_{\cal I}, \tau_{\cal S})$ exist, such that $out_{\cal I}(\tau_{\cal I}+i+1) = {in}_{{\cal S}}(\tau_{\cal S}+i)$. Let ${\cal I}$ and ${\cal S}$ be two systems as defined in Def. \[def\_system\_composition\_while\] and $(\tau_{\cal I}, \tau_{\cal S}) \in T_{\cal I}\times T_{\cal S}$ such that $out_{\cal I}(\tau_{\cal I}) = 0$ and $q(\tau_{\cal S}) = a$. Then the states $in_{\cal L} = q_{{\cal S}}$ and $out_{\cal L} = out_{\cal I}$ form a system if there is a $\delta\in \mathbb{N}$ such that $f_{\cal S}(\delta, q) = 0$. Then the time step is $((\tau_{\cal I}, \tau_{\cal S}), (\tau_{\cal I}+\delta, \tau_{\cal S}+\delta))$. We also write ${\cal L} = While_{a}(I, S)$ The proof is similar to the previous ones. For while-systems the following proposition holds: \[prop\_while\_compositionality\] The system function of a while-composition composes computational emergently. Even if the while functionality composes compositionally, for a system, according to definition \[def\_system\] we have to provide a time step. However, the knowledge, that $g$ somehow has a zero does not suffice. We would need to have a general algorithm to calculate the specific $\delta$, which does not exist, hindering us to provide the time step. As both, the loop- as well as the while system are stateless, another immediate consequence of Kleene’s recursive function theorem is the next proposition: \[prop\_form\_of\_computational\_systems\] Every computable stateless system is functionally equivalent to a system of the form ${\cal S} \circ ({\cal P}_1 || \dots || {\cal P}_n)$, $Loop_n(I, S)$ or $While(I, S)$. This shows that from a functional perspective, ”state” is only needed to represent input and output values and it is also needed internally for loop- and while- computations, but ”state” is not needed for the representation of computational functionality. This fact forms the basis for the influential area of ”functional programming” [@Hudak1989] where manipulating state is termed a ”side-effect”. A point of view which perpetuates itself until nowadays in the form of the so called ”REST-architecture paradigm” [@Fielding2000]. But as we will see below, ”state” does become very important for cooperative interactions and hence, any approach to neglect it will have its weaknesses in describing these kind of system relations. Equipped with a notion of system composition, the concept of super- or subsystem can be defined. \[def\_subsystem\] Let ${\cal S}$ be composed from systems ${\cal U}_k$ ($k = 1,\dots, n$) according to definition \[def\_system\_composition\_sequential\], \[def\_system\_composition\_parallel\], \[def\_system\_composition\_loop\], or \[def\_system\_composition\_while\]. Then ${\cal S}$ is called the [*super system*]{} of the ${\cal U}_k$ and the ${\cal U}_k$ are the [*subsystems*]{} of ${\cal S}$. Pseudo recursion or U-composition --------------------------------- A very common ”pattern” of system composition is shown in Fig \[fig\_system\_pseudo\_recursion\]. Of three consecutive systems, the first and last are attributed to a ”system ${\cal S}$” which seems to interacts recursively with, that is providing output to and receiving input from, the intermediate system ${\cal S}_2$. ![Diagram of a system sequence. Very often there is a box drawn around system ${\cal S}_1$ and ${\cal S}_3$ and this box is treated as a system ”${\cal S}$” as is shown in gray. But this is - although very intuitive - an example of graphical lie. Indeed, all three systems ${\cal S}_1$, ${\cal S}_2$, and ${\cal S}_3$ are subsystems of the resulting system ${\cal S'} = {\cal S}_3\circ{\cal S}_2\circ{\cal S}_1$.[]{data-label="fig_system_pseudo_recursion"}](fig/system_composition_pseudo_recursion_v01.png "fig:"){width="8cm"}\ Actually, ${\cal S}$ could be viewed as ${\cal S}_1 || {\cal S}_3$ - if both systems shared their input and execute in parallel - which obviously is not the case. So, what is termed as ”${\cal S}$” is not a system in the sense of Def. \[def\_system\]. In fact, all three systems ${\cal S}_1$, ${\cal S}_2$, and ${\cal S}_3$ are subsystems of the resulting system ${\cal S'} = {\cal S}_3\circ{\cal S}_2\circ{\cal S}_1$. But we will see that this pattern is of great importance, when it comes to functional components, as both, ”${\cal S}$” and ${\cal S}_2$, can be viewed as components of ${\cal S}'$. I also speak of U-composition. System cooperation \[s\_system\_cooperation\] ============================================= As was illustrated in Fig. \[fig\_system\_interacting\_reality\], from a composition perspective, it is not a priori clear what really happens if systems mutually interact. One important class of interactions that was already mentioned in Fig. \[fig\_system\_interacting\_reality\] but does not lead to system composition in the sense of section \[s\_system\_composition\] are nondeterministic, stateful, asynchronous interactions[^6]. The focus is shifted from hierarchical system composition dominated by an inductive notion of a system function to two complementary issues: First, how can we describe the nondeterministic interactions? How much of the systems do we have to include, how much can we exclude? And second, how can we describe a deterministically working system, where an action is an execution of its system function, but which takes part in multiple nondeterministic interactions? We will see that this leads to the question, how to deal with the problem of interaction coordination. ![Two systems, ${\cal S}_1$ and ${\cal S}_2$ interacting with each other and with multiple other systems. Within a single interaction, only the projection of the system functionality onto the interaction channels becomes relevant.[]{data-label="fig_system_cooperation"}](fig/system_cooperation_v01.png "fig:"){width="8cm"}\ In contrast to a function which represents a unique mapping, a nondeterministic relation between input and output values prompts for an additional criterion to distinguish between success and failure. We also do not expect a successful cooperation to be attributed to a single system, but to all involved systems. Also, we need a tool with which we can describe the relevant parts of the involved systems while ignoring the irrelevant ones. The first issue, how we should describe the nondeterministic interactions, is solved by (partitioned) I/O-automata coupled by ideal Shannon Channels. I/O-automata fulfill both requirements, to provide an additional acceptance component and to be able to represent only the, for the interaction, relevant parts of a system. Their coupling by Shannon channels restricts their transition relation and leads to the protocol concept (e.g. [@Holzmann1991; @Brand1983; @lynch87hierarchical]) as the outer coupling of the interacting systems. Here systems in given roles mutually document their state transitions in a way that is comprehensible to the respective receivers and thereby all systems can achieve a common goal together. As is illustrated in Fig. \[fig\_system\_cooperation\], only projections of the interacting systems onto the interaction channels become visible. The knowledge of the interaction does not suffice anymore to determine the actions of the interaction partners from the perspective of the sender. The second issue of interaction coordination is solved by using the building blocks of the interactions, namely the roles, and by searching for a complementary inner coupling which represents the desired causal linkage between the transitions of the different roles of a given system. That is, we look for logical conditions to restrict the transition relation of the product of all roles until determinism is established and thereby a well defined system function is provided. An extensive description illustrated with the example of a resource administrating process can be found in [@Reich2012_PRI]. Here I present only the most important, slightly modified definitions. The building blocks: input/output automata {#s_building_blocks} ------------------------------------------ Input/output automata will be used to describe both, processes and their interactions. \[def\_NFIOA\] A ”[*nondeterministic I/O automaton (NIOA)*]{}” is a tuple ${\cal A} = (Q, I, O, q_0, Acc, \Delta)$. $Q$ is the non-empty set of state values, $I$ and $O$ are the, possibly empty, input and output alphabets. $q_0$ is the initial state value, $Acc$ is the acceptance component and $\Delta \subseteq Q \times Q \times I^\epsilon\times O^\epsilon$ is the transition relation. Instead of $(p,q,i,o) \in \Delta$ we also write $p \stackrel{i/o}{\rightarrow} q$. In case $Q$, $I$, and $O$ are finite, then the automaton is called ”finite” (NFIOA). The acceptance component $Acc$ represents the information needed to express ”successful” computation, that is, the acceptance condition. It depends on the computational model of ”success”. For a finite computation $Acc_{finite}$ is a finite set of final state values. For an infinite computation of a finite automaton, a representative acceptance condition is to accept all runs in which the finite set of infinitely often occurring state values is an element of the acceptance component (Muller-acceptance), that is for $Acc_{Muller} \subseteq \wp(Q)$ [@DBLP:conf/dagstuhl/Farwer01]. Transitions with an $\epsilon$-character as input are called [*spontaneous*]{}. In case that for each $(p, i) \in Q \times I^\epsilon$ there is at most one transition $(p, q, i, o) \in \Delta$ then $\Delta$ specifies a function $\delta: Q \times I^\epsilon \rightarrow Q \times O^\epsilon$ with $(q,o) = \delta(p, i)$. If no $\epsilon$ is used, that is $\delta: Q \times I \rightarrow Q \times O$, we have a deterministic I/O-automaton or [*DIOA*]{}. A finite DIOA is a Mealy automaton [@Mealy1955]. An automaton is supposed to be executed in a certain way. It is important that some properties of an NIOA depends on the model of execution. It could well be that an automaton permanently provides a character as a possible input, but the execution model never chooses it - which is called ”starvation”. Be ${\cal A}$ an NIOA. A sequence $q_0, q_1, \dots , q_n$ where $q_0 = {q_0}_{\cal A}$ is the initial state and $(q_{j-1}, q_j, i, o)\in\Delta_{\cal A}$ is called an ”[*execution*]{}”. An infinite sequence is called an ”[*$\omega$-execution*]{}”. The pair $(q_{j-1}, q_j)$ is called an [*event*]{}. A computation of an execution is performed according to the following rules: 1. The first state value is the initial state of ${\cal A}$, $q_0$. 2. \[step\_2\] Evaluate the current state value with respect to the acceptance component 3. With $p$ as the current state value, arbitrarily choose one of the possible transitions $(p,q,i,o)\in \Delta_{\cal A}$ and thereby determine the input characters $i\in I_{\cal A}^\epsilon$. 4. $q$ is the new current state value of the execution. Resume with step \[step\_2\] An $\omega$-execution is called [*weakly fair*]{} if a constantly available input character is eventually chosen. An $\omega$-execution is called [*strongly fair*]{} if a repeatedly possible input character is eventually chosen. The reason to introduce NIOAs is their ability to represent projections of systems. An NIOA ${\cal A}$ ”[*specifies a projection*]{}” of a system ${\cal S}$, if $Q_{\cal S} \subseteq Q_{\cal A}$, $I_{\cal S} \subseteq I_{\cal A}$, $O_{\cal S} \subseteq O_{\cal A}$, and a projection function[^7] $\pi = (\pi_Q, \pi_I, \pi_O): Q_{\cal S} \times I_{\cal S}^{\epsilon}\times O_{\cal S}^{\epsilon} \rightarrow Q_{\cal A} \times I_{\cal A}^{\epsilon}\times O_{\cal A}^{\epsilon}$ exists such that $\Delta_{\cal A}$ is the smallest possible set and for each time pair $(t, t') \in T \times T$ in every possible sequence, $(\pi_Q(q(t)), \pi_Q(q(t')), \pi_I(in(t)),\\ \pi_O(out(t'))) \in \Delta_{\cal A}$. It is clear that to any system according to Def. \[def\_system\] there is a corresponding deterministic I/O-automaton (DIOA) and to any DIOA there is a corresponding system. If the system has only one input and output state function, then the projection onto this states reproduces the complete system. The precondition of any coupling of NIOAs is to view them together as a product automaton, stepping synchronously and where all acceptance conditions are to be fulfilled simultaneously. \[def\_weakly\_synchronized\_product\] The [*weakly synchronized product*]{} of a set of $n$ NIOAs ${\cal A}_k$ is defined by NIOA ${\cal B} = (Q, I, O, \vec{q}_0, Acc, \Delta)_{\cal B}$, with $Q_{\cal B} = \mbox{\Large $\times$} Q_k$, $I_{\cal B} = \mbox{\Large $\times$} I_k$, $O_{\cal B} = \mbox{\Large $\times$} O_k$, ${\vec{q}_0}_{\cal B} = ({q_0}_1, \dots, {q_0}_n)$, the common acceptance component represents the logical conjunction of the individual components, symbolized as $Acc_{\cal B} = \bigwedge Acc_k$, $\Delta_{\cal B} := \{(\vec{p}, \vec{q}, \vec{i}, \vec{o})|$ the components of $\vec{p}$ are reachable states of the ${\cal A}_i$ and ${\cal A}_k$ provides a transition $(p_k, q_k, i_k, o_k)$ with $\vec{q} = \vec{p} \left[\frac{q_k}{p_k}, k\right]$ and $\vec{i} =\epsilon[i_k, k]$ and $\vec{o} = \epsilon[o_k, k]$ $\}$. I also write ${\cal B} = \bigotimes_{i=1}^n {\cal A}_i$. We can say that for a state transition $(p,q,i,o) \in \Delta$ the event $(p,q)$ is elicited by $i$ and documented by $o$. This kind of ”event documentation” is an inherent part of the transition semantic. Please note, that the definition \[def\_weakly\_synchronized\_product\] of a weakly synchronized product automaton, which will be the basis for the protocol definition, entails an important restriction with respect to its I/O-behavior compared to a system: it restricts its input or output to a single component of its I/O-vectors (which by themselves could be vectors). The definition of a system does not provide any restrictions on actuating its output components of being driven by multiple input components - the weakly synchronized product automaton does so. ### Partitioned I/O-automata With the help of an equivalence relation $\equiv$ on transitions, the transition relation $\Delta$ can be partitioned into (disjoint) equivalence classes $\Delta_l$ with $l\in L \subseteq \mathbb{N}$. Then there is a function $part:\Delta \rightarrow L$ which maps each tuple of the transition relation to the index of its equivalence class. There are two important equivalence class constructions which are both based on descriptive equivalence, that is two transitions are equivalent according to a given description. #### Extended Automata For nondeterministic I/O-automata, we construct the equivalence relation $\equiv_{ND}$ as following: 1. Be $DocCls$ a set of document classes and $Param$ a set of parameter values. Then be $parse: I\cup O \rightarrow DocCls \times Param$ with $(docCls, param) = parse(a)$ a function assigning a document class and parameter values to each character. We can thereby look at each character as an instance of a document of a given document class representing parameter values. 2. All state values have a mode component and a rest: $Q = Q_{mode} \times Q_{rest}$. 3. Be $Cond$ a set of conditions: $Cond: Q_{rest} \times Param \rightarrow \{true, false\}$. Then, two transitions $t,t'\in\Delta$ are equivalent $t \equiv_{ND} t'$ if both can be attributed the same $docCls_i$, $docCls_o$, $p_{mode}$, $q_{mode}$, and $cond(p_{rest}, param_i)$, that is, if their value of the partition function, set up as $$part(p,q,i,o) = part(docCls_i, docCls_o, p_{mode}, q_{mode}, cond(p_{rest}, param_i))$$ is identical. $p_{rest}$ as well as $param_o$ remain unconsidered. Instead of $\Delta_l$ we also write $$\label{eq_equiv_class_transition} p_{mode}\stackrel{docCls_i, cond(p_{rest}, param_i)/docCls_o}{\xrightarrow{\hspace*{5cm}}} q_{mode}$$ An example is a seller that gets an $Order$ from a customer, transits from $listening$ to $ordered$ if the customer is evaluated as being $trustworthy$ and sends back a $Confirmation$: $$listening\stackrel{Order, isTrustworthy(Customer)/Confirmation}{\xrightarrow{\hspace*{5cm}}} ordered$$ So, ”extended” automata, as they are often called, are actually not so much extended but rather ”abstracted” by equivalence class construction. Please note that in contrast to the notation $p \stackrel{i/o}{\rightarrow} q$ which denotes a single transition, the notation of the expression \[eq\_equiv\_class\_transition\] refers to an equivalence class of transitions. #### Objects/Abstract Data Types In the case of deterministic I/O-automata, also the subautomata are deterministic and each subrelation $\Delta_l$ specifies a function $\delta_l: I\times Q \rightarrow O\times Q$ with $(o,q) = \delta_l(i,p)$ for $(p,q,i,o) \in \Delta_l$. If we treat all I/O-characters as state values then all state transitions relate to externally provided ”data-”inputs. We have $(o, q) = \delta_l(i,p) = \delta_l|_{p}(i)$. Indeed these are two equations: $$\begin{aligned} o & = & \delta_l^{(o)}|_{p}(i) \label{eq_object_I}\\ q & = & \delta_l^{(q)}|_{p}(i) \label{eq_object_II}\end{aligned}$$ The first equation represents the object oriented way of describing state transitions by providing a name for the state function tuple, the object name: outputParameters = objectName.method(inputParameters). The second equation is only implicitly given in the world of objects. If we additional assume all state values to be partitioned into a mode component and a rest: $Q = Q_{mode} \times Q_{rest}$, then we have three equations: $$\begin{aligned} o & = & \delta_l^{(o)}|_{p_{mode}}(i, p_{rest}) \\ q_{mode} & = & \delta_l^{(mode)}|_{p_{mode}}(i, p_{rest}) \\ q_{rest} & = & \delta_l^{(rest)}|_{p_{mode}}(i, p_{rest}) \\ \end{aligned}$$ The definition of a mode-state is known as ”state pattern” in the object oriented world [@Gamma1995] but is usually not directly syntactically supported. It allows the definition of generic events $(p_{mode}, q_{mode})$. But, in contrast to the state transition $p\stackrel{i/o}{\rightarrow}p$ mentioned before, the documentation of these events is not part of the transition semantic! There is a special case of nondeterminism which is treated similar to determinism: exceptions. With exceptions we can syntactically distinguish between a desired ”normal” functioning and undesired ”exceptional” circumstances. An example is the command to write to a file system. Usually it will succeed - except, when the disk is full. So, for most purposes we can assume a deterministic relation and for the rare circumstances, our assumption proves to be wrong, we have to fundamentally change our proceeding anyway. To create exceptions, the transition relation has to be partitioned into two main parts: a deterministic part which gets further partitioned into a set of subrelations as described for the DIOA. And a rest, which gets partitioned into additional - usually only few - equivalence relations, representing the exceptional circumstances. A possible way of syntactically expressing these equivalence classes is the well known exception mechanism, where the execution model is extended by a try-catch-mechanism. Outer coupling: channel based restriction \[ss\_outer\_coupling\] ----------------------------------------------------------------- The outer coupling is concerned with the causal connection between the output and input of different systems, which I call ”outer coupling”, and which was illustrated in Fig. \[fig\_system\_cooperation\]. The essential question is: How can we describe the nondeterministic interaction between deterministic systems. We will aim for a notion, which we call ”protocol”, which is an NIOA that does not receive any additional input - as it already represents all the input and output of the involved systems. With respect to systems, a Shannon channel is a common I/O-state function as defined in section \[s\_system\_composition\]. In the context of I/O-automata representing system projections, a channel does two things: first, it creates a product automaton, and second, it restricts the set of reachable state values of the product automaton by modifying the rules for executing the automaton’s transition rules[^8]. If a character is output on a Shannon channel it has to be processed as an input character in the next step. Be ${\cal A}$ an NIOA where $O_k$ is the $k$-th component of the output alphabet $O = O_1 \times \dots \times O_n$ and $I_l$ is the $l$-th component of the input alphabet $I = I_1 \times \dots \times I_m$ with $O_k \subseteq I_l$. The pair $(k,l)$ is called a [*Shannon channel*]{} if it modifies the execution rules for ${\cal A}$ in the following way: 1. Initially, the automaton is in its initial state $q_0$. 2. \[step\_2\] Evaluate the current state value with respect to the acceptance component 3. Be $p\in Q_{\cal A}$ a reachable state of ${\cal A}$ which was reached by a transition with the $k$-th output component $o \in O_k$. Then only a transition of $\Delta_{\cal A}$ can be chosen next which starts from $p$ and has $o \in I_l$ as input character. I call the transition providing the character [*sending*]{} transition, the ensuing transition [*receiving*]{} transition. 4. Be $p\in Q_{\cal A}$ either the initial state or a reachable state of ${\cal A}$ which was reached by a transition with a character $o\in O_{\cal A}$ with $o_k = \epsilon$ that is no output on the channel. Then only a transition of $\Delta_{\cal A}$ can be chosen next starting from $p$ with $\epsilon$ as its $l$-th component, that is, which is either spontaneous or which has an input not associated with the channel. 5. $q$ is the new current state value of the execution. Resume with step \[step\_2\] Such an execution is also called a ”[*channel based restricted (CBR-)*]{}” execution. The automaton with all reachable states of ${\cal A}$ and only the executable transitions of ${\cal A}$ under CBR is called a ”CBR-automaton”. A CBR-automaton where all input and output states are connected via Shannon channels is called ”[*closed*]{}”, otherwise it is called ”[*open*]{}”. It is not a priori clear that this procedure creates something meaningful. A condition which simplifies the execution of a CBR-automaton is that at most one character has to be considered as input. Hence, the ability of a system to output a vector of characters where multiple positions of this vector are different from $\epsilon$ is to be avoided. This is an important difference to the general definition of a system and its corresponding DIOA. A CBR-automaton is called ”[*linear-executable*]{}” if for each element of $I$ and $O$ at most one component is unequal to $\epsilon$. Otherwise it is called ”[tree-executable]{}” For tree-executable CBR-automata, the result of the execution will in general depend on the execution strategy, for example which branch of the execution tree is calculated first. Another necessary condition obviously is that there has to be a receiving transition for each sending transition for each channel, that each interaction chain terminates and that the acceptance condition can still be met. I therefore define: A CBR automaton is called ”[*well formed*]{}” if for every channel mediated transition which sends a character (different from $\epsilon$) there exists a receiving transition to process it. This is a typical safety property (something bad never happens). To automatically check it, all reachable states have to be explored. Hence, for infinite automata it is generally undecidable. For finite automata it is decidable, but its solution requires exponential effort with the number of state functions. A well formed channel based restricted automaton is called ”[*consistent*]{}” if for each reachable state value either the acceptance condition is met or there is at least one continuation such that the acceptance condition can be met and every interaction chain eventually terminates. This is a typical liveness property (something good eventually happens) and therefore for infinite automata generally undecidable. A ”[*protocol*]{}” ${\cal P}$ is a closed CBR product automaton characterized by the set of factor automata $\{{\cal A}_i\}$ and a set of Shannon Channels $\{c_j\}$. I call the factor automata ”[*roles*]{}”. If not stated otherwise, I assume that a protocol is linear-executable. A protocol is linear-executable if its roles provide at most one character per output vector. As each step of the protocol is achieved by only one component automaton (one role), the input and output components of the product automaton again differ from the empty character epsilon in at most one component. For a product automaton, the properties of being well formed and being consistent are both not compositional as they are only properties of the protocol and not of its parts, the roles. As was already mentioned in [@Reich2010] the definition of ”consistent” directly entails that a consistent protocol neither has deadlocks in the sense that there is a single reachable state without any continuation such that the acceptance condition holds nor livelocks, characterized by a set of periodically reached states without such a continuation. Cooperating systems \[s\_cooperating\_systems\] =============================================== Now we turn our attention to the systems themselves that cooperate, that is, take part in at least two different nondeterministic interactions in two different roles. I call such a system a process. A system is called a ”[*process*]{}” if it takes part in at least two different, well defined, consistent, and nondeterministic (protocol-)interactions in two different roles. A process is called ”[*linear-executable*]{}” if all its roles of all interactions it is involved in are linear-executable. That is, a process is linear-executable, if it produces only one character per Shannon channel at a time. Looking from an interaction perspective onto a process, as a system taking part in multiple nondeterministic interactions, it is an interesting question how to synthesize such a process from its roles. As described in [@Reich2012_PRI], we thereby look for a second mechanism to couple roles, in this case to represent the causal relation between the transitions of the roles of a single system as an inner coupling determined by coordination rules, illustrated in Fig. \[fig\_system\_inner\_coupling\]. Inner coupling: condition based restriction \[ss\_inner\_coupling\] ------------------------------------------------------------------- The goal is to restrict the transition set of a weakly synchronized product automaton such that at least quasi-determinism of the formerly nondeterministic transition set is achieved in a sense that from each reachable state there is at most one transition for each input character and otherwise a single transition with the empty character as input, the collective acceptance condition can still be met, and the factor NIOAs (the roles) still can be regained by projection onto their interaction. ![A single system ${\cal S}$ coordinating multiple roles in different interactions. It shows the necessity to describe the inner coupling between the different roles, indicated by the circle arrow within the opaque area of the system.[]{data-label="fig_system_inner_coupling"}](fig/system_inner_coupling_v01.png "fig:"){width="5cm"}\ An NIOA that has, from any reachable state, at most one transition for each input sign, and otherwise a single transition with the empty character as input is called ”[*quasi-deterministic*]{}” (as is its transition relation). Quasi-determinism means that, although not truly deterministic, there is no real choice anymore, such that an automated execution of all arising state transitions becomes possible. \[def\_projected\_automaton\] Be ${\cal B}$ an NIOA and $\pi = (\pi_Q, \pi_I, \pi_O): Q_{\cal B}^\epsilon \times I_{\cal B}^\epsilon\times O_{\cal B}^\epsilon \rightarrow Q_{\cal B}^\epsilon \times I_{\cal B}^\epsilon \times O_{\cal B}$ a projection function. Then the projected automaton ${\cal A} = \pi({\cal B})$ is given by $Q_{\cal A} = \pi_Q(Q_{\cal B})$, $I_{\cal A} = \pi_I(I_{\cal B})$, $O_{\cal A} = \pi_O(O_{\cal B})$, ${q_0}_{\cal A} = \pi_Q({q_0}_{\cal B})$, $Acc_{\cal A} = \pi_Q(Acc_{\cal B})$, $\Delta_{\cal A} = \{(p',q',i',o')| (p, q, i, o) \in \Delta_{\cal B} \,\mbox{and}\, p'=\pi_Q(p), q'=\pi_Q(q), i'=\pi_I(i), o'=\pi_O(o)\}$. Be ${\cal T}= \bigotimes_{i=1}^n {\cal A}_i$ a weakly synchronized NIOA consisting of $n$ nondeterministic roles ${\cal A}_i$ and be $(\pi_i)$ $n$ projection functions, each projecting $\cal T$ onto the Shannon channels of role ${\cal A}_i$. Then ${\cal P}$ is a ”[*coordinated automaton*]{}” of [T]{} if $Q_{\cal P} = Q_{\cal T}$, $I_{\cal P} = I_{\cal T}$, $O_{\cal P} = O_{\cal T}$, $Acc_{\cal P} = Acc_{\cal T}$, and there exists a transition relation $\Delta_{\cal P} \subseteq \Delta_{\cal T}$ such that it is quasi deterministic, retains the collective acceptance condition, and retains all roles by projection onto their interactions, that is $\pi_i({\cal P}) = {\cal A}_i$. The transitions which are intended not to occur in the coordinated automaton can be described by a set of conditions - why I called this type of coupling ”condition based coupling” in [@Reich2012_PRI]. The fulfillment of all protocols can be seen as an invariant of the transition elimination procedure. Please not that (in contrast to what I have proposed in [@Reich2012_PRI]), it could well be that the single roles still can be regained by projection, but the transition relation of the restricted product does not fulfill the collective acceptance condition any more. An obvious and very interesting research question is, under which conditions such a coordinated automaton can be found. Now that we have achieved a quasi-deterministic automaton by stating the coordinating conditions, we could either advance to a deterministic one and thereby creating a traditional process with an explicit system function. Or we could keep this quasi-deterministic automaton as it is and execute it in a modified execution environment. Achieving full determinism by eliminating the remaining spontaneous transitions generally requires state value elimination, which implies a couple of serious consequences: 1. Eliminating state values possibly deprives all other factor automata from their chance to transit, possibly changing the behavior of the entire product automaton. That is, eliminating the spontaneous transitions changes the ”atomicity of time steps”. 2. As the acceptance components of the roles directly relate to the set of state values, they must be restated. 3. The projection relation between the restricted product automaton and its constituting roles gets lost. 4. If the product automaton was linear-executable before, we might have to aggregate several output characters in one step and thereby loose the property of linear-executability. All these issues can be avoided by modifying the execution environment of a process by not only processing input characters as they become available (the deterministic part), but also possible spontaneous transitions (the nondeterministic part). This leads to an implicit process notion where rules related to the inner state of a system provide the internal glue to internally coupled roles together to become a process. Process composition ------------------- How do processes compose? Can two systems, each taking part in at least two interactions, always be viewed as a single system, taking still part in at least two further interactions? Let us look first at a situation, where, for unclocked systems, no process results: There are three systems interacting as illustrated in Fig. \[fig\_process\_composition\_II\] in a closed chain. ![Three systems, interacting mutually in a closed chain[]{data-label="fig_process_composition_II"}](fig/process_composition_II_v01.png "fig:"){width="6cm"}\ If we look only at ${\cal S}_1$ and ${\cal S}_2$, they seem to compose to a supersystem, having two new interactions. But both interactions refer to the same system ${\cal S}_3$. If ${\cal S}_3$ has no other interfaces than the ones to ${\cal S}_1$ and ${\cal S}_2$, then a closed chain similar to Fig. \[fig\_sequential\_vs\_reciprocal\_coupling\] results. We see that an additional necessary condition beyond the remaining two interactions is, that they are directed towards two different systems. Be ${\cal S}_1$ a linear-executable process with $n \ge 2$ roles $({\cal A}_{1,i})$ and ${\cal S}_2$ a linear-executable process with $m \ge 2$ role $({\cal A}_{2,j})$ which interact only through the consistent protocol ${\cal P} = (\{{\cal A}_{1,1}, {\cal A}_{2,1}\}, \{c_1, c_2\})$. Additionally, ${\cal S}_1$ and ${\cal S}_2$ each interacts with at least one additional linear-executable process ${\cal T}_1$ and ${\cal T}_2$ with ${\cal T}_1 \neq {\cal T}_2$ by consistent protocols. Then the protocol interaction creates a supersystem denoted by ${\cal S} = {\cal P}_{({\cal A}_{1,1}, {\cal A}_{2,1})}({\cal S}_1, {\cal S}_2)$ and this system is again a process. First we have to show that ${\cal S}$ is indeed a system, then we have to show that this system still takes part in at least two interactions. To show that ${\cal S}$ is indeed a system, we have to construct a deterministic automaton from the deterministic automata of ${\cal S}_1$ and ${\cal S}_2$. This is possible, since the well-definedness of ${\cal P}$ guarantees that every interaction chain between ${\cal S}_1$ and ${\cal S}_2$ terminates. The linear-executability of both processes guarantees that any action of both, ${\cal S}_1$ or ${\cal S}_2$, that was initiated by an interaction within ${\cal P}$ terminates either with no output or with an output within another interaction. According to the assumptions, the new system ${\cal S}$ does interact with at least two other systems, So, processes in this sense form unlimited interaction networks as was illustrated in Fig. \[fig\_network\_of\_relations\]. There is no such thing as an ”end-to-end” process. What confines these interaction networks are the roles of the interaction protocols, that is incomplete projections of ”the last interaction partner”. What happens if a process determines the behavior of another system entirely? As the following proposition shows, in the case of a consistent reciprocal system interaction, a system becomes a subsystem if its behavior is entirely determined by its interaction ”partner”. Such an interaction results in a recursive system where some output becomes the input of a next time step, until the chain of recursive interaction terminates Let ${\cal S}$ be a linear executable process with more than one consistent interactions and described by a DIOA ${\cal D}$. It especially interacts with another system ${\cal U}$ described as a DIOA ${\cal B}$ by the consistent linear-executable protocol ${\cal P}({\cal A}, {\cal B})$, where ${\cal A}$ is an NIOA describing only a projection of ${\cal S}$. Then ${\cal S}$ and ${\cal U}$ are subsystems of a larger system ${\cal T}$. The proof follows from the fact that the resulting automaton of two interacting linear executable DIOAs is again a linear executable DIOA and thereby represents a system - if in addition the interaction between the role of the process and the other complete system is a consistent and linear executable protocol. This implies that every interaction chain terminates. , one is determining the other in a consistent protocol interaction, is again deterministic - and thereby represents a system. Related Work \[s\_related\_work\] ================================= There is a vast amount of existing literature on formal system engineering and modeling. Any extensive review would be an ambitious endeavor on its own. I restrict myself to a couple of other system models I came across while I was dealing with the subject. My focus is not so much on truth but on fit: Does a model fit to the phenomena it wants to represent in a way that it leads to adequate questions - and answers. I mainly look at how they represent the system function, how expressive their formal representation is and whether they reproduce the compositionality of the system function. This investigation suggests that these three criteria do provide valid assessment criteria. To some extent, the expressiveness of the investigated system models does not suffice to represent the differences between parallel versus sequential composition or even recursion in a simple enough way or do not make a clear enough distinction between the deterministic versus nondeterministic interaction paradigms. Some lack the compositionality of their system function. And some other do not provide the formal means to distinguish between external and internal coupling. Abstract data types and objects ------------------------------- According to the Dictionary of Algorithms and Data Structures, an abstract data type is ”A set of data values and associated operations that are precisely specified independent of any particular implementation.” Within the object oriented programming paradigm, this construct is in one way or another supplemented by polymorphism and inheritance. According to definition \[def\_system\] such an entity is a system, exposing its system function by its interface. The relation to partitioned I/O-automata was already expressed in Eq. \[eq\_object\_I\] and \[eq\_object\_II\]. As the interaction between objects is expressed by a method call, this kind of interaction usually follows the pseudo-recursive case, illustrated in Fig. \[fig\_system\_pseudo\_recursion\]. However, if the interaction between objects is based on untyped transport functionality then the semantics of the interaction is simply not expressed syntactically. If the interaction is based on mutual typed method calls, things become complex, as I have shown. This explains that especially objects have their stronghold in the area of hierarchical system composition but fail to scale in the area of nondeterministic interactions. Reactive systems ---------------- In [@HarelPnueli1985_Reactive] David Harel and Amir Pnueli distinguish between ”transformational” and ”reactive” systems. They define: "A transformational system accepts input, performs transformation on them and produces outputs” - which is consistent with the definition \[def\_system\] of a system. They further define ”Reactive systems ... are repeatedly prompted by the outside world and their role is to continuously respond to external inputs ... A reactive system, in general, does not compute or perform a function, but is supposed to maintain a certain ongoing relationship, so to speak, with its environment.” In my opinion it is false to conclude from the non-functional relation of the reactive systems to their interaction partners to their somehow non-functional functioning - or ”mode of operation”. Even reactive systems work in a step wise mode and at least for technical systems the simple fact holds that for being constructable, there has to be a function to be implemented - the system function. The main difference between ”transformational” and ”reactive” systems in the sense of David Harel and Amir Pnueli is not their functioning but - as they correctly point out - their relation to their environment. For ”transformational” systems their interfaces towards the interaction provides access to the full system function and thereby provides deterministic interactions. ”transformational” systems are composed hierarchically by composition according to the definition of sub/super systems \[def\_subsystem\]. By contrast, as I proposed already in [@Reich2012_PRI], the interfaces of ”reactive” systems provide access to at least two different projections of the systems, which I call ”roles” and thereby only provide nondeterministic interactions. The system model of Manfred Broy -------------------------------- Manfred Broy describes his component model in several articles (e.g. [@DBLP:books/sp/cstoday95/Broy95; @Broy2009_Relating_Time; @Broy2010_Logical]). The behavior of a system is given by a set of ”input” and ”output” processes, whereby a process in his sense is a finite or infinite set of discrete events. The event concept is not entirely clear, as on the one hand, an event may represent (among others) a state change or some action, but on the other hand he states that an event represents a point in time and thus has no time duration. Also, it remains unclear, whether event sequences like $a,\epsilon, b$ and $a, b$ are semantically equivalent or not. A key concept in his theory is what he calls a [*timed stream*]{} of messages $m\in M$ which is an infinite sequence of finite sequences of messages. Each finite sequence $s\in M^*$ can be attributed to a time interval $t\in \mathbb{N}$ as $s(t)$. The set of timed streams is denoted as $(M^*)^\infty$. If we assume $M=\{0,1,\dots,9\}$, then $M^*$ becomes the set of all natural numbers $\mathbb{N}$ and $(M^*)^\infty$ is the set of all infinite sequences of arbitrary natural numbers. As can be seen from the example, the amount of information to be processes in one time interval is finite, but can be arbitrary large. The messages are transported by typed channels $c\in C$. A [*channel valuation*]{} is a mapping which assigns a typed timed stream to channel. The set of all valuations to a given set of channels $C$ is then defined as $\vec{C} = \{x:C \rightarrow (M^*)^\infty:\forall c\in C: x(c)\in (type(c)^*)^\infty\}$. Each channel valuation defines a communication history $x:C \rightarrow (\mathbb{N} \rightarrow M^*)$ for the channels in $C$. The black box behavior or as Manfred Broy calls it, [*semantic interface*]{}, of a component is then described by a function $f$ mapping the set of the histories of the input channels $\vec{I}$ onto the power set of histories of the output channels $\wp(\vec{O})$ of the component: $f:\vec{I} \rightarrow \wp(\vec{O})$. Manfed Broy then explicitly introduces the properties of causality as well as realizability in the sense that all output histories can indeed be computed. Composition is defined by a single composition operator with feedback. Given two interface behaviors with disjoint set of output channels $O_1 \cap O_2 = \emptyset$: $F_1:\vec{I}_1 \rightarrow \wp(\vec{O}_1)$, $F_2:\vec{I}_2 \rightarrow \wp(\vec{O}_2)$. Then with $I= (I_1 \cup I_2)\setminus (O_1 \cup O_2)$, $O=(O_1 \cup O_2) \setminus (I_1 \cup I_2)$, and $C=I_1 \cup I_2 \cup O_1 \cup O_2$ Manfred Broy defines the interface behavior of the composed interface $F = F_1\otimes F_2: \vec{I} \rightarrow \wp(\vec{O})$ by $F(x) = \{(y\in \vec{C})|O: y|I = x|I \wedge y|O_1 \in F_1(y|I_1) \wedge y|O_2 \in F_2(y|I_2)\}$ where $x|Y$ means the restriction of the valuation $x$ to the channels in $Y$. According to Manfred Broy, the composition operator $\otimes$ preserves realizability. Due to the behavioral approach which focuses on observable behavior, the definition of black box behavior eschew internal states - but not for composed systems. The definition of a composed system allows for hidden channels which become nothing else than internal states in the sense of time dependent attributes of the composed system. Lets look at two simple systems $i=1,2$ each with two input channels $I_i = \{I_{i,1}, I_{i,2}\}$, two output channels $O_i = \{O_{i,1}, O_{i,2}\}$ and a deterministic system function $f_i = (f_{i,1}, f_{i,2}) = (I_{i,1} \wedge I_{i,2}, \neg(I_{i,1} \wedge I_{i,2}))$. The system composition is recursive with $O_{1,2} = I_{2,1}$ and $O_{2,1} = I_{1,2}$. By this equalities, we see that the channels do not just represent the channel names as Manfred Broy states, but the channels themselves. Otherwise the equality of two channel variables relating to two channels with two different names would not make much sense. Following Manfred Broy’s composition rule we then have $I= \{I_{1,1}, I_{2,2}\}$, $O=\{O_{1,1}, O_{2,2}\}$, $C=\{I_{1,1}, I_{1,2}, I_{2,1}, I_{2,2}, O_{1,1}, O_{1,2}, O_{2,1}, O_{2,2}\}$ and $F(x) = \{(y\in \vec{C})|O: y|I = x|I \wedge y|O_1 \in F_1(y|I_1) \wedge y|O_2 \in F_2(y|I_2)\}$. This seems to be pretty simple and straight forward. So one is tempted to ask where the recursion has been hidden. The complexity is hidden in the fact, that without coupling, the streams $I_{1,2}$ and $I_{2,1}$ can take arbitrary values at each point in time, while in the coupled case, only their first value can be arbitrarily chosen and all other values depend in a complex way on these two initial values together with the complete history of input values of the remaining input streams $I_{1,1}$ and $I_{2,2}$. So, the complexity of recursive system relations cannot be avoided. But recurring to the complex structure of timed streams, my impression is that Manfred Broy adds a substantial amount of unnecessary additional complexity whereas it is difficult for me to find the additional benefit. The BIP component framework of Joseph Sifakis et al. ---------------------------------------------------- BIP stands for Behavior, Interaction, and Priority [@BazuBozgaSifakis2006; @BasuEtAl2011_Rigorous; @BensalemBozgaQuilbeufSifakis2012_Knowledge]. It is a general framework to describe reactive systems from certain parts with a set of rules. In my view, BIP is a way to describe computational systems in the sense of this article: a BIP-component provides states and a transition function. By its priority mechanism, together with a given port, there is supposed to be a unique transition to be selected. An [*atomic component*]{} is a labeled finite transition system characterized by the tuple $(State, state_0, P, G, F, Trans)$. $State = S \times V$ gives the possible state values. $S$ is called control state, $V = V_1 \times \dots \times V_n$ is called data state[^9]. $state_0$ is the initial state. $P$ is the set of ports, which are transition labels \[or action names\] used for synchronous state transitions of multiple components. $G$ is a set of ”guard” conditions operating on $V$, that is each $g\in G$ is a function $g:V\rightarrow \{true, false\}$. $F$ is a set of functions operating on V, that is for each $f\in F$, $f:V\rightarrow V$. $Trans$ is the set of transitions with $trans = State \times P \times G \times F \times State$. A transition $t = (state, p, g_p, f_p, state')$ represents a step from state value $(s, v)$ to $(s', v')$ with $v' = f_p(v)$. It can be executed if some interaction including port $p$ is offered by the execution environment and the guard $g_p(v) = true$. Components are glued together by connectors defining interactions and by priorities. For simplicity we assume components with disjoint sets of names of ports, state values and transitions. A [*connector*]{} specifies interactions between its components $\{C_1, \dots, C_n\}$. It is characterized by a tuple $(V, P, I, G, F, Trans)$. $V = V_1 \times \dots \times V_n$ is called data state[^10]. $P$ is the set of ports the connector relates to. $I\subseteq \mathcal{P}(P)$ is the set of feasible interactions, where each $i\in I$, $i \subseteq P$. $G$ is a set of ”guard” conditions operating on $V$, that is each $g\in G$ is a function $g:V\rightarrow \{true, false\}$ having access to the complete data state of all components. $F$ is a set of functions operating on V, that is for each $f\in F$, $f:V\rightarrow V$, also having access to the complete data state of all components and providing ”data exchange” between the components. $Trans$ is the set of transitions with $trans = V \times I \times G \times F \times V$. A transition $t = (v, i, g_i, f_i, v')$ represents a step from state value $v$ to $v' = f_i(v)$. It can be executed if its interaction becomes feasible (that is, the respective components have reached a state such that each provides transitions labeled with the respective port) and the guard $g_p(v) = true$. A [*priority*]{} selects among possible interactions. It is characterized by a tuple $(c, G, I, <)$. $c$ is the component the priority belongs to, which composes $\{C_1, \dots, C_n\}$ with data state $V = V_1 \times \dots \times V_n$. $G$ is a set of ”guard” conditions operating on $V$. $I$ is a set of interactions. $< \subseteq I \times I$ is a strict partial order relation providing the priority. When the condition holds and both interactions are enabled, only the higher one is possible. A [*compound component*]{} is then given by its components, its connectors and its priorities. Neglecting priorities, it is finally characterized by a tuple $(State, P, I, G, F, Trans)$. $State = S\times V = S_1 \times \dots \times S_n \times V_1 \times \dots \times V_n$. $P = \bigcup P_i$, $I = \bigcup I_i$, $G = \bigcup G_i$, and $F = \bigcup F_i$ is the set of functions, operating on $V$. $Trans \subseteq State \times I \times G \times F \times State$ is the set of transitions where a transition is given by $t=(state, \alpha, g, f, state')$. $\alpha$ is a feasible interaction associated with a guard $g_\alpha \in G$ and a function $f_\alpha \in F$ such that there exists a subset $J\subseteq\{1\dots n\}$ of atomic components with transitions $\{(state_j, p_j, g_j, f_j, state_j')\}_{j\in J}$ and $\alpha = \{p_j\}_{j\in J}$. $g = \left(\bigwedge_{j\in J} g_j\right) \wedge g_\alpha$. $f = (f_1, \dots, f_n) \circ f_\alpha$. $v' = f(v)$. $s'(j) = s_j'$ if $j\in J$; otherwise $s'(j) = s_j$. That is, the states from which there are no transitions labeled with ports in $\alpha$, remain unchanged. Without priorities, a move $(s, v) \stackrel{\alpha}{\rightarrow} (s', v')$ is possible if there exists a transition $(state, \alpha, g, f, state')$, such that $g(v) = true$. Most importantly, the functionality of BIP components is not compositional in the sense of definition \[def\_compositionality\]. There is always the global function operating on all states of a composed component at once first. And only afterwards, the components’ genuine functions are applied to the individual states, violating the encapsulation and thereby making this model of reactive systems complex. The purpose of this connector-function is probably to represent data exchange - but obviously it is not limited to reproduce states, but it is allowed to make arbitrary manipulations. So fundamentally, this model does not distinguish clearly between transport and processing of information. The system model of Rajeev Alur ------------------------------- In his excellent book ”Principles of Cyber-Physical Systems” [@Alur2015_Principles] Rajeev Alur distinguishes between functional and reactive components in the tradition of David Harel and Amir Pnueli. Focusing on the timing behavior, he develops several models of computation for reactive components. In the synchronous model all components execute in lock-step, similar to what I have called ”clocked” systems. In the asynchronous model all processes execute at independent speeds, and there is an unspecified delay between the reception of inputs and the production of outputs by a process. In the timed model processes rely on a global physical time to achieve a loose form of synchronization. #### Synchronous Reactive Components (SRC) A [*synchronous reactive component (SRC)*]{} $C$ is defined by a tuple $C = (I, O, S, Init, React)$, mixing semantic elements like sets, states, and values with syntactic elements like variables and descriptions. $I, O, S$ are sets of typed input, output and state variables, each variable taking values only from the sets $Q_I, Q_O, Q_S$. $Init$ is the description of the initialization defining the set $[\![Init]\!] \subseteq Q_S$ of initial states, and [*React*]{} is a description of the reaction defining the transition relation $[\![React]\!] \subseteq Q_S \times Q_I \times Q_O \times Q_S$. An SRC $C$ is said to be [*event triggered*]{}, if a subset $J \subseteq I$ exist, where all the variables in $J$ also can take the value $\bot$ (meaning ”absent”); every output variable is either latched (the value of the output variable is the updated value of a state variable) or can also be absent; and for a reaction with absent input, the non-latched output remains also absent. The reaction $React = (L, {\cal A}, \succ)$ is given by a set $L$ of task local variables and a set ${\cal A}$ of tasks and a binary precedence relation $\succ \subseteq {\cal A} \times {\cal A}$ (The latter two represent a [*task-graph*]{}). Each task $A$ has a read-set $R \subseteq I \cup S \cup O \cup L$, a write-set $W \subseteq O \cup S \cup L$, and an update description $Update$ with $[\![Update]\!] \subseteq Q_R \times Q_W$ such that (1) $\succ$ is acyclic; (2) each output variable belongs to the write-set of exactly one task; (3) if an output or a local variable $y$ belongs to the read-set of a task $A$, then there exists a task $A'$ such that $y$ is in the write-set of $A'$ and $A' \succ^+ A$; and (4) if a state or a local variable $x$ belongs to the write-set of a task $A$ and also to either the read-set or write-set of a different task $A'$, then either $A \succ^+ A'$ or $A' \succ^+ A$. As he says, his approach to ensuring well-behaved composition relies on the syntactic decomposition of the reaction description into tasks given by the designer. He introduce the notion of an interface for $I$, $O$, $\succ \subseteq I \times O$ of a component to ensure consistent composition. Two components $C_1$ and $C_2$ with $I_i$, $O_i$, and $\succ_i$ $(i=1,2)$ are said to be [*compatible*]{} if (1) $O_1$ and $O_2$ are disjoint and (2) the relation $\succ = \succ_1 \cup \succ_2$ is acyclic. Actually, the task graph over the set of tasks of $C_1$ and $C_2$ obtained by retaining the precedence edges in the individual components and adding cross-component edges from a task $A_1$ of one component to a task $A_2$ of another component whenever $A_1$ writes a variable read by $A_2$, is acyclic. For composition, we assume that the input/output variables are named in a way to establishes the intended communication pattern and the internal variables are named to avoid naming conflicts. Let $C_i = (I_i, O_i, S_i, Init_1, React_i)$, $i=1,2$ be two compatible SCR. Suppose that $React_i = (L_i, {\cal A}_i, \succ_i)$. Then the [*parallel composition*]{} $C = C_1|| C_2$ is again an SCR such that $S = S_1 \cup S_2$, $O = O_1\cup O_2$, $I = I_1 \cup I_2 \setminus O$, the initialization for a state variable $x_i$ is given by $Init_i$, $(i=1,2)$; the reaction description of $C$ uses the local variables $L=L_1\cup L_2$ and is given by the task graph such that (1) ${\cal A} = {\cal A}_1 \cup {\cal A}_2$ and (2) $\succ = \succ_1 \cup \succ_2$ together with the task pairs $(A_1, A_2)$, such that $A_1$ and $A_2$ are tasks of different components with some variable occurring in both, the write-set of $A_1$ and the read-set of $A_2$. The commutativity of parallel composition of SCRs indicates that the formalism does not express the different composition schemata described in section \[s\_system\_composition\]. With the compatibility condition recursion is excluded. #### Asynchronous processes In the asynchronous case, Rajev Alur talks about processes and channels instead of components and input/output variables. An [*asynchronous process (AP)*]{} $P$ is defined again by the tuple $P = (I, O, S, $ $Init, React)$. However, the reaction is defined differently with input task $A_x$ for every input channel $x\in I$, an output tasks $A_y$ for every output channel $y$ and a set of internal tasks ${\cal A}$ as $React = (\{{\cal A}_x|x\in I\}, \{{\cal A}_y|y\in O\}, {\cal A_i})$. Each task consist of a guard condition $Guard$ over $S$ and an update rule $Update$. The update rule of each $A_x$ has a read-set $S \cup \{x\}$ and a write-set $S$, each $A_y$ has a read-set $S$ and a write-set $S \cup \{y\}$, and each $A$’s read- and write-set is $S$. The [*parallel composition*]{} $P_1|P_2$ of two AP $P_i = (I_i, O_i, S_i, Init_i, React_i$, $(i=1,2)$ is again an AP if $O_1$ and $O_2$ are disjoint and $S = S_1 \cup S_2$, $O = O_1 \cup O_2$, $I = I_1 \cup I_2 \setminus O$, and $Init = Init_1; Init_2$. For each input channel $x\in I$, (1) if $x\notin I_2$, then the set of input tasks ${\cal A}_x$ is ${\cal A}_x^1$; (2) if $x\notin I_1$, then the set of input tasks ${\cal A}_x$ is ${\cal A}_x^2$; and (3) if $x\in I_1 \cap I_2$, then for each task $A_1 \in {\cal A}_x^1$ and $A_2 \in {\cal A}_x^2$, the set of input tasks ${\cal A}_x$ contains the task described by $Guard_1 \wedge Guard_2 \rightarrow Update_1;Update_2$. For each output channel $y\in O$, (1) if $y\in O_1\setminus I_2$, then the set of output tasks ${\cal A}_y$ is ${\cal A}_y^1$; (2) if $y\in O_2 \setminus I_1$, then the set of output tasks ${\cal A}_y$ is ${\cal A}_y^2$; (3) if $y\in O_1 \cap I_2$, then for each task $A_i\in {\cal A}_y^i$ (i=1,2) the set of output tasks $A_y$ contains the task described by $Guard_1 \wedge Guard_2 \rightarrow Update_1;Update_2$; and (4) if $y\in O_2 \cap I_1$, then for each task $A_i\in A_y^i$ (i=1,2), the set of output tasks ${\cal A}_y$ contains the task described by $Guard_2 \wedge Guard_1 \rightarrow Update_2;Update_1$; The set of internal tasks is ${\cal A} = {\cal A}_1 \cup {\cal A}_2$. #### State machines However, to introduce more advanced matter like safety or liveness requirements, Rajeev Alur turns to state machines, very similar to the I/O-automata in this article. Both, SCRs (p.66) and APs (p.141) have naturally associated transitions systems by declaring the input and output variables to be local variables. With a pair of states $(s, t)$ he identifies $s \stackrel{i/o}{\rightarrow} t$ as the reaction for some input $i$ and some output $o$ for SCRs - which is exactly the notation of a transition of the DIOA. For APs he adds the undefined value $\perp$ to the possible values of input and output variables in case they are not involved in an action - which is one of the essential ideas distinguishing NIOAs from DIOAs. So, from an automata perspective, the difference between SRCs and APs is mapped onto the deterministic versus nondeterministic issue - very similar to the approach of this article. However, what is lacking is recursion in the deterministic case and most importantly the role concept and therewith the distinction between the inner and outer coupling, the essence of the process model of this article. Algebraic approaches based on named actions ------------------------------------------- There are many algebraic approaches to describe processes which are based on named actions. Among others, these are Robin Milner’s calculus of communicating systems (CCS) [@Milner1989; @Milner1992], Charles A. R. Hoare’s calculus of communicating sequential processes (CSP) [@Hoare1985] and Jan A. Bergstra’s and Jan W. Klop’s algebra of communicating processes (ACP) [@BergstraKlopp1987_Universal]. Also, Richard Mayr’s [@Mayr1999] process rewrite systems can be subsumed her. For a recent overview, see the book of Jos C. M. Baeten, Twan Basten, and Michel Reniers [@BaetenBastenReniers2010_ProcessAlgebra]. They all have in common the view of a process as a structure producing events by actions and - this is the main point - are based on providing names for these events as their ’alphabet’ they operate on. So, the semantics of these calculi is provided by transition systems where each transition has a label, the name of the action causing the state change. An interaction becomes the simultaneous execution of actions. Although, it is generally possible to find a mapping from the i/o-pairs of the IO-automata used in this article to some set of unique transition labels, the coupling mechanism thereby gets lost. As described in section \[ss\_outer\_coupling\] the outer coupling was based on identical names of output and input characters: that the output of one transition is the input of another transition (Shannon channel). As described in (section \[ss\_inner\_coupling\]), the basis for the inner coupling are the $\epsilon$ transitions, as these can be eliminated to represent the desired causal relation between the role-transitions. As a result, an important difference between any approach with named actions and the presented approach with anonymous actions but named I/O-characters is their different support for describing composition behavior. As the names of the actions are arbitrary, they do not help to express the fact that some transitions may refer to the same action, but only from a different perspective/projection. Discussion \[s\_discussion\] ============================ This is a rather conceptual contribution where I combined the description of systems with the description of reciprocal system interactions in a network context. Basing the system notion on some identifiable functionality implies that the gestalt of systems becomes very dynamic. Interactions between systems may create temporary super systems. A child that is hit by a car becomes aware of the center-of-mass system including the car only during the very impact. Someone speaking to me at close distance acquires control of my ear tympanum and thereby this part of me temporarily becomes a part of the speaker’s system. Transferring this dynamics to computer science implies that any programming paradigm with a fixed linkage between data and functionality cannot fully represent this kind of system dynamics of the real world. The presented system notion entails that the borders of systems are drawn by logic and sometimes practically restricted by physics. The borders between the internal ”deterministic” parts of an engineered system and the parts that are coupled to it by nondeterministic interactions are drawn somewhat discretionary but are limited - from an engineering perspective - upon the required level of determinism. Effectively, determinism is fiction. In a recent study, Bianca Schroeder, Eduardo Pinheiro, and Wolf-Dietrich Weber [@SchroederPinheiroWeber2009] measured the number of correctable errors of a DRAM-DIMM as nearly 4000 per year or 10 per day. Accordingly, the hardware manufacturer have to use error correction mechanism to bring this rate down below an acceptable limit. On the other hand, unreliable communication can make reaching consensus impossible, as the coordinated attack problem shows [@Gray1978]. So, one effect of the increasingly faster and more reliable communication is to allow engineers to vastly extend the deterministic fiction to control even ”remote” systems. The presented approach shows that there is more to supersystem formation than mere interaction, as for example Robin Milner [@Milner1980; @Milner1989] or Charles Hoare [@Hoare1985] still thought. Supersystem formation in the sense of this article becomes dependent on perspective and knowledge. It requires the presence of a well defined time successor function. Whether such a time successor function exists or not isn’t always obvious and cannot be algorithmically decided in the general case. Recursion makes the question where system borders begin and end, quite complex. As illustrated in Fig. \[fig\_system\_interacting\_reality\] recursively interacting systems could be part of a loop- or a while-system, they could interact loosely coupled or only pseudo recursively - or they could do something thereof we must better be silent. In today’s applications, unclear component relations are daily fare. Either there exist explicitly declared mutual functional dependencies or no declared dependency at all, as in the case of the broadly adopted web-programming paradigm AJAX - a security nightmare. With current practices, where operations represent only send or receive semantics and thereby not determining subsystem relations on the one hand and with protocol based interactions which in the deterministic case may determine subsystem relations, the questions where systems logically end is difficult to decide. From a practical point of view, security mechanisms like encryption and signature could be viewed as tools to make system borders unequivocally visible. Based on asymmetric cryptographic procedures, a signature function uses the ”private” key of a document issuer while a decryption function uses the ”private” key of a document receiver. This notion of ”privacy” relies on valid system borders and means that this kind of functionality is accessible only within a given system. If we look with this perspective on the current web clients, running in unknown environments (e.g. [@LekiesStockJohns2013]), we see that it could be possible for the one who controls the environment to access the private client keys - invalidating the assumed system borders. Well defined system borders are also a prerequisite for the ”end-to-end” argument [@SaltzerReedClark1984]. In their article, Jerome H. Saltzer, David P. Reed and David D. Clark stated that ”choosing the proper boundaries between functions is perhaps the primary activity of the computer system designer” and suggested as a guiding design principle for placement of function among the modules of a distributed system that ”functions placed at low levels of a system may be redundant or of little value compared with the cost of providing them at that low level”. The demonstrated tight connection between our system and our function notion contributes to a better understanding, why approaches that rest mainly on the function notion struggle so much with modeling network-like interacting systems and the necessary ”loose coupling”. One example is functional programming [@Hudak1989], other examples are the many distributed object models like CORBA [@Vinoski1997], Java RMI [@WollrathRiggsWaldo1996], COM/DCOM [@Rogerson1998], SOA [@SchulteNatis1996], etc. Robert B. Glassman [@Glassman1973] defines interacting systems like cells, organisms or even societies as being ”loosely coupled” if they have few states in common, or, to put it more technically, if the state space of the interaction is small against the state space the computed functionality is based on. However, within the object and service oriented communities, it is often stated that ”loose coupling” of systems is created by abstraction (e.g. [@Kaye2003]). Though abstraction decouples from any specific implementation, it does not decouple from the particular implementation which is used during runtime and certainly not from the underlying logical mapping. Actually, this would imply to name a tire ”loosely coupled” to a car, just because it can be changed easily. It seems to me that this kind of ”loose coupling” is more concerned with low versus high effort to replace a subsystem. The presented process model underlines that nondeterminism and therefore nondeterministic automata have an important role in software engineering (e.g. in contrast to Helmut Balzert [@Balzert2001], p.323). Nondeterminism provides the base for the role concept and thereby the differentiation between the roles’ external and internal coupling. It avoids the strong centralization tendencies of the traditional approach to model processes as stepwise executable activities (see [@reich2011; @Reich2012_PRI; @DBLP:journals/tse/DesaiMCS05]). It explains, why execution environments which enforce the construction of deterministic processes also enforce a reduced flexibility on a software architecture level. Hence, the proposed process model signifies a new kind of reuse in software engineering, where processes can easily be configured from predefined roles by simply stating the rules for their internal coupling. I would like to emphasize the break between the representation of systems by general DIOAs and processes, where linear executability becomes important to keep problems analytically tractable. It is the distinction between different interaction partners which makes it impossible to subsume several output states into a single vectorial output state and thereby subsume channels. Another interesting distinction is the one between clocked and unclocked systems. While closed chains of unclocked systems are dead, as is illustrated in Fig. \[fig\_process\_composition\_II\], this is not true for clocked systems. Actually, its enough to have a single clocked system in such a chain to enable activity. As ”clocked” does not necessarily means to be triggered in regular, fixed intervals, or to work in sync with other systems, it could also be interpreted as the difference between systems with internally driven, truly spontaneous activity and systems which are purely passive. It may also be interesting to transfer this insight to other scientific disciplines, like sociology, where both, the role concept (e.g. [@Goffman1959]) as well as the system concept (e.g. [@Parson1951_SocialSystems]) has already been very influential. Consequentially, one could speculate that this kind of rule based, loose inner coupling between roles to processes has its biological analogon within our central nervous system. On a more physiological level, one could speculate that the cerebellum with its important function in learning automated movements [@MantoEtAl2012_Consensus] is somehow a rule engine coordinating primitive movement patterns to more complex composites. Also the economic science of organizing organizations may benefit from the proposed process model, as at least some influential constructs like Lean [@Ono1988_Tyota] view processes as something static, inflexible, which has to be described as meticulous as possible to become amenable to engineered change. This could be extended to business software, whose task is to support the workflows in companies. The main difference between small and large companies is not so much their interface towards the outside, how they sell and buy, but rather the way and the amount of division of labor inside the company. So a business software supporting small as well as large companies alike might benefit enormously from structurally anchoring such a flexible role based process model in its software architecture. The presented point of view on systems and their interactions also implies some important requirements on the completeness of component models. George T. Heineman and Bill Councill ([@HeinemanCouncill2001], p.7) define a software component as ”[*a software element that conforms to a component model and can be independently deployed and composed without modification according to a composition standard*]{}” and a component model as a ”[*specific interaction and composition standard*]{}”. A complete component model must as least support two general classes of components: The first class are components which are intended to behave nondeterministically in interactions. To be described adequately, these interfaces have to be protocol-like. The second class are components which are intended to behave deterministically in interactions. This class can be further partitioned into components designated to work within a pipe with some splitting and merge-mechanism and components designated to work as objects with lumped input and output, or as I called it ”pseudo-recursively” in a component hierarchy. In their overview, Ivica Crnkovic, Severine Sentilles, Aneta Vulgarakis, and Michel R. V. Chaudron [@CrnkovicSentilles2011_Classification] differentiate between ”operation based” and ”port based” interface support and thereby show that there are contemporary widely used, important component models which are not complete in my sense. For the second class of components, an important requirement is to hide functional recursion and thereby to provide compositional behavior that is easy to understand. This requirement is automatically fulfilled if we stick to the proposed classification. Let me conclude this article with the remark, that what computer scientists named ”abstract data types” or ”object” since the 1960s are in fact systems in the sense of this article. So, one could say that finally with the advent of cyber physical systems, it is time to change terminology and switch from ”object” to ”system” orientation, and - while keeping in mind the dynamics of the system notion - use the same terms for the same things as all the other engineering disciplines. [^1]: Referring to quantum physics we could also distinguish between classical and quantum states. [^2]: Some authors like Rajeev Alur [@Alur2015_Principles] name these system ”synchronous” and ”asynchronous”. However, I would like to reserve these terms for the behavior of a ”sending” system: a ”synchronous” working system provides some output and waits for some input - the usual behavior of a component calling another component’s function. An ”asynchronous” working system provides some output and goes on processing without waiting for any further input coming back [@Reich2015_kvsi]. [^3]: Proposed in his talk ”Compositionality huh?” at the Dagstuhl Workshop “Divide and Conquer: the Quest for Compositional Design and Analysis” in December 2012. [^4]: As something is ”computable” means that there could be an algorithm to calculate it, being ”computable emergent” would mean that the property of being emergent is computable - which it is not. Hence, I call it ”computational” emergent, indicating, that from a computational perspective, it is emergent. [^5]: If the time steps and the concatenation states are clear from the context, then they are omitted in the notation. [^6]: An interaction classification, based on these three parameters can be found in [@Reich2015_kvsi] [^7]: A projection function $\pi$ is defined by the property $\pi = \pi \circ \pi$. [^8]: In [@Reich2012_PRI] the same effect war accomplished by introducing ”excited” state values. However, I think that relating to a modified execution is a more adequate way of comprehension. [^9]: the authors use $V$ to denote the variables which represent the data states syntactically. Here the index $i$ relates to the $i$-th data state \[variable\]. [^10]: Now the index $i$ relates to the $i$-th component.

ArXiv

--- abstract: | We use the NIHAO galaxy formation simulations to make predictions for the baryonic budget in present day galaxies ranging from dwarf ($\Mhalo\sim10^{10} \Msun$) to Milky Way ($\Mhalo\sim10^{12} \Msun$) masses. The sample is made of 88 independent high resolution cosmological zoom-in simulations. NIHAO galaxies reproduce key properties of observed galaxies, such as the stellar mass vs halo mass and cold gas vs stellar mass relations. Thus they make plausible predictions for the baryon budget. We present the mass fractions of stars, cold gas ($T<10^4$K), cool gas ($10^4 < T < 10^5$K), warm-hot gas ($10^5 < T < 5\times10^6$K), and hot gas (T$> 5\times10^6$K), inside the virial radius, $R_{200}$. Compared to the predicted baryon mass, using the dark halo mass and the universal baryon fraction, $f_{\rm b}\equiv \Omega_{\rm b}/\Omega_{\rm m}=0.15$, we find that all of our haloes are missing baryons. The missing mass has been relocated past 2 virial radii, and cool gas dominates the corona at low mass (M$_{200} \lta 3 \times 10^{11} \Msun$) while the warm-hot gas dominates at high mass (M$_{200} \gta 3 \times 10^{11} \Msun$). Haloes of mass $\Mhalo\sim 10^{10}\Msun$ are missing $\sim 90\%$ of their baryons. More massive haloes ($\Mhalo\sim 10^{12}\Msun$) retain a higher fraction of their baryons, with $\sim 30\%$ missing, consistent with recent observational estimates. Moreover, these more massive haloes reproduce the observed fraction of cold, warm-hot and hot gas. The fraction of cool gas we predict ($0.11\pm0.06$) is significantly lower than the observation from COS-HALOs (0.3-0.47), but agrees with the alternative analysis of @Stern16. author: - | Liang Wang$^{1,2,4}$[^1], Aaron A. Dutton$^{3,4}$, Gregory S. Stinson$^4$, Andrea V. Macciò$^{3,4}$,\ $^1$International Centre for Radio Astronomy Research (ICRAR) , M468, University of Western Australia, 35 Stirling Hwy,\ Crawley, WA 6009, Australia\ $^2$Purple Mountain Observatory, the Partner Group of MPI für Astronomie, 2 West Beijing Road, Nanjing 210008, China\ $^3$New York University Abu Dhabi, PO Box 129188, Abu Dhabi, UAE\ $^4$Max-Planck-Institut für Astronomie, Königstuhl 17, 69117 Heidelberg, Germany date: submitted to MNRAS title: 'NIHAO VII: Predictions for the galactic baryon budget in dwarf to Milky Way mass haloes' --- \[firstpage\] galaxies: evolution – galaxies: formation – galaxies: dwarf – galaxies: spiral – methods: numerical – cosmology: theory Introduction {#sec:intro} ============ Cosmic structure formation has redistributed the baryons from a nearly uniformly distributed plasma into a variety of states, including stars, stellar remnants, cold (atomic and molecular) gas, and hot (ionized) gas. The theories of galaxy formation can predict the amount of mass in these different states, which can in turn be tested by observational constraints. On cosmological scales, the ratio between the total baryonic and gravitating mass is measured to be $f_{\rm b}\equiv \Omega_{\rm b}/\Omega_{\rm m}\simeq 0.15$ (The Planck Collaboration 2014). However, the cold baryonic mass density implied by several galaxy baryon estimates (mainly stars and cold gas) is only 3-8% of the big bang nucleosynthesis expectation [@Persic92; @Fukugita98; @Bell03; @McGaugh10]. The majority of the cosmic baryons are thought to be in the form of hot gas around or between galaxies [@Cen09]. Until recently only a fraction of these baryons had been detected [@Bregman07; @Shull12]. This discrepancy is referred to as the “missing baryon problem”. Several theoretical studies with cosmological simulations have constrained the phase of the potential reservoirs of the missing baryons in the intergalactic medium (IGM), and find a large fraction of the baryons with low density and high temperature resides between galaxies [@Yoshida05; @He05; @Dave10; @Zhu11; @Haider16]. As a part of the IGM, the circum galactic medium (CGM) is always treated as an major potential reservoir of the missing baryons. and metal absorption lines are expected to signpost such diffuse baryonic content. is mainly from gas with temperature T $\sim 10^4$ K so that it is able to detect cold gas in the CGM. Meanwhile, theoretical work has predicted that a substantial portion of the CGM is in the warm or hot phase with temperature T $> 10^{4.5}$ K. Gas enters this phase through photoionization, accretion shocks or shocks caused by galactic winds [@Voort12]. Such dilute halo gas is at T$\sim 10^{4.5-7}$ K, so the detection is dominated by metal lines, e.g. . Recent advances in the detection of gas in the CGM have come from the COS survey [@Tumlinson11; @Tumlinson13; @Thom12; @Werk12; @Werk13]. On the scale of Milky Way mass haloes $\Mhalo \sim 10^{12}\Msun$ a significant amount of warm ($10^4 < T <5\times 10^6$K) gas has been detected [@Werk14], accounting for 33-88% of the baryon budget. In the future such observations will be extended to a wider range of halo masses. A number of large volume cosmological simulations [@Ford13; @Ford16; @Suresh15; @Oppenheimer16] and zoom-in cosmological simulations [@Stinson12; @Hummels13; @Shull14] have given predictions for the and absorption lines. @Gutcke16 compared the column density profile of and in the CGM of galaxies from the NIHAO [@Wang15] cosmological hydrodynamical simulation suite with observations, studied the covering fraction of dense , looked at the shape of the CGM and its chemical composition. The simulations reproduce the observational covering fraction and column density profile of cool well, and recover the observed trends of column density with luminosity and impact parameter. In common with other simulations, however, the column density of is lower and the extent of optically thick is smaller than observed. The physical properties of the CGM has been shown to be able to test feedback models [@Sharma12; @Marasco13]. @Dave09 predicted galactic halo baryon fractions of galaxies with halo masses ranging from $10^{11} \Msun$ to $10^{13} \Msun$ using cosmological hydrodynamical simulations with a well-constrained model for galactic outflows. They found that, without the outflow model, the baryon fraction inside the virial radius is roughly the cosmic baryonic fraction, but with the outflow model, the baryon fraction is increasingly suppressed in lower mass haloes. By comparing results at $z=3$ and $z=0$, they showed that large haloes remove their baryons at early times while small haloes lose baryons more recently due to the wind material taking longer to return to low-mass galaxies than high-mass galaxies. @Muratov15 showed similar results that the gas and baryon fractions are lower at lower redshift, after powerful outflows at intermediate redshift $z \approx 0.5-2$ remove a large amount of gas from the halo. Several simulations [e.g., @Crain07; @Christensen16; @Voort16] found the baryon fraction and gas fraction are reduced compared to the cosmic baryon fraction, especially in low mass haloes. @Sokolowska16 studied the halo gas of three Milky way-sized galaxies using cosmological zoom-in simulations. They found that most of missing baryons actually resides in warm-hot and hot gas which contribute to 80% of the total gas reservoir. The recovered baryon fraction within 3 virial radii is 90%. The warm-hot medium is sensitive to the feedback model so that a reliable spatial mapping of the warm-hot medium will provide a stringent test for feedback models. In this paper we make predictions for the baryonic budget for stars, cold, warm and hot gas in and around the virial radius of haloes of mass ranging from $\Mhalo\sim 10^{10}\Msun$ to $10^{12}\Msun$. We use a sample of 88 zoom-in galaxy formation simulations from the NIHAO project. Reproducing the stellar mass content in dark matter haloes both today and in the past has been a formidable challenge for cosmological galaxy formation simulations [@Weinmann12; @Hopkins14]. Even the latest state-of-the art simulations have trouble: the ILLUSTRIS simulation [@Vogelsberger14] strongly overpredicts the stellar masses in dwarf galaxy haloes $(M_{200} \lta 10^{11}\Msun)$, while the EAGLE simulations [@Schaye15] underpredict the peak of the star formation efficiency in halos of mass $M_{200}\sim 10^{12}\Msun$. In contrast, the NIHAO galaxies are consistent with the stellar mass vs halo mass relations from halo abundance matching since redshift $z\sim 4$ [@Wang15], the galaxy star formation rate vs stellar mass relation since $z\sim 4$ [@Wang15], and the cold gas mass vs stellar mass relation at $z\sim 0$ [@Stinson15]. Therefore, the simulations make plausible predictions for the mass fractions and physical locations of the warm and hot gas components. We find that all the haloes contain less baryons than expected according to the cosmic baryonic fraction, but the missing fraction is strongly mass dependent. This paper is organized as follows: The cosmological hydrodynamical simulations including star formation and feedback are briefly described in §\[sec:sims\]; In §\[sec:budget\] we present the results including the baryonic budget, baryon distribution, and a comparison with observations; §\[sec:sum\] gives a summary of our results. Simulations {#sec:sims} =========== In this study we use simulations from the NIHAO (Numerical Investigation of a Hundred Astrophysical Objects) project [@Wang15]. The initial conditions are created to keep the same numerical resolution across the whole mass range with typically a million dark matter particles inside the virial radius of the target halo at redshift $z=0$. The halos to be re-simulated at higher resolution with baryons have been extracted from 3 different pure N-body simulations with a box size of 60, 20 and 15 $h^{-1}$ Mpc respectively. We adopted the latest compilation of cosmological parameters from the Planck satellite [@Planck14]. Dark matter particle masses range from $\sim 10^{4} \Msun$ in our lowest mass haloes to $\sim 10^{6} \Msun$ in our most massive haloes, and their force softenings range from $\sim$ 150 pc to $\sim$ 900 pc, respectively Gas particles are less massive by factor of $(\Omega_{\rm dm}/\Omega_{\rm b})\simeq 5.48$, and the corresponding force softenings are 2.34 times smaller. More information on the collisionless parent simulations, the force softenings and particle masses for the highest refinement level for each simulation and sample selection can be found in @Dutton14 and @Wang15. We use the SPH hydrodynamics code [gasoline]{} [@Wadsley04], with a revised treatment of hydrodynamics as described in @Keller14. The code includes a subgrid model for turbulent mixing of metal and energy [@Wadsley08], heating and cooling include photoelectric heating of dust grains, ultraviolet (UV) heating and ionization and cooling due to hydrogen, helium and metals [@Shen10]. The star formation and feedback modeling follows what was used in the MaGICC simulations [@Stinson13]. The gas is converted into stars according to the Kennicutt-Schmidt Law when it satisfies a temperature and density threshold. Stars feed both metals and energy back into the ISM gas surrounding the region where they formed. SN feedback is implemented using the blastwave formalism described in @Stinson06. Pre-SN feedback is an attempt to consider radiation energy from massive stars. Heating is introduced immediately after massive stars form based on how much star light is radiated. Our simulations use thermal feedback to provide pressure support and increase gas temperature above the star formation threshold, and thus to decrease star formation. There are two small changes in NIHAO simulations compared to MaGICC: The change in number of neighbors and the new combination of softening length and particle mass increases the threshold for star formation from 9.3 to 10.3 cm$^{-3}$, the increase of pre-SN feedback efficiency $\epsilon_{\rm ESF}$, from 0.1 to 0.13. More details on the star formation and feedback modeling can be found in @Wang15. Baryon budget {#sec:budget} ============= We define the fiducial baryonic mass as: $$M_{\rm b} \equiv M_{\rm b}(R_{200})= \frac{f_{\rm b}}{1-f_{\rm b}}M_{\rm dm}(R_{200}) \label{equ:mb}$$ where the $M_{\rm dm}$ is the total dark matter mass of the halo, and the $f_{\rm b} = \Omegab/\Omegam \sim 0.15$ is the cosmic baryon fraction (the ratio between baryon density and mass density including baryonic mass plus dark matter), so that $M_{\rm b}$ would be the baryonic mass inside the virial radius if the baryons followed the dark matter closely. Fig. \[fig:budget\] shows the ratio between the mass of each baryon component inside the virial radius to the fiducial baryonic mass for the most massive galaxy in each zoom-in region. We present the fractions of total stellar mass within 20% $R_{200}$ (blue points), and the total baryonic mass including stellar mass plus gas mass (green points). For the stellar mass fraction we also show the relations from halo abundance matching [@Moster13; @Behroozi13; @Kravtsov14]. The grey area is the one sigma scatter around the mean value from @Kravtsov14. Fig. \[fig:budget\] shows that all haloes in our study contain less than the universal fraction of baryons. The upper panel uses a linear y-axis scale, which highlights the large amount of baryons that are missing, especially in low mass haloes. The logarithmic scale in the lower panel highlights the power-law nature of the relations. The trends of each component fraction are similar, in that the fractions are relatively low in low mass haloes, and increase as the halo mass increases. The main difference between the different components is the slope, with the baryonic mass fraction having a shallower slope than the stellar mass fraction. This is because in low mass haloes ($\Mhalo\sim 10^{10}\Msun$) most of the baryons are in the form of gas, while in the highest mass haloes we study ($\Mhalo\sim 10^{12}\Msun$) a substantial amount of gas has been turned into stars. Since most of the haloes we study are above the mass where the cosmic UV background prevents gas from cooling, the missing baryons have most likely been ejected from the central galaxies in supernova/stellar feedback driven winds. Although the lower mass galaxies have converted a smaller fraction of their available baryons into stars, and hence there is proportionally less energy available to drive an outflow, they have expelled a larger fraction of their baryons. This is consistent with expectations from energy driven gas outflows, where the lower star formation efficiency is more than compensated by the shallower potential wells of lower mass halos according to the mass loading factor and circular velocity relation $\eta \propto V^{-2}$ [e.g., @Dutton12; @Christensen16]. The behavior of the baryonic mass fraction, $f_{\rm bar}$, as a function of the halo mass is captured using a double power law formula: $$\frac{f}{f_0} = \left( \frac{M_{200}}{\mathcal{M}_0} \right)^\alpha \left\{ 0.5 \left[ 1+\left( \frac{M_{200}}{\mathcal{M}_0} \right)^\gamma \right] \right\}^{\frac{\beta-\alpha}{\gamma}}.$$ In this formula, the lower and higher mass ends have logarithmic slope $\alpha$ and $\beta$, respectively, while $\gamma$ regulates how sharp the transition is from the lower to the higher ends. Giving all points equal weight, the best fit parameters are as follows: $$\begin{aligned} \mathcal{M}_0 &=& 6.76 \times 10^{10} \nonumber \\ f_0 &=& 0.336 \nonumber \\ \alpha &=& 0.684 \\ \beta &=& 0.205 \nonumber \\ \gamma &=& 3.40\nonumber \end{aligned}$$ The three most massive galaxies in the NIHAO suite have fairly high stellar masses compared to abundance matching results (see blue filled circles in Fig. \[fig:budget\]), and are thus possibly overcooled. However, they don’t significantly bias the baryon fraction fitting formula since the high mass slope ($M_{200} > 10^{11}\Msun$) is constrained by $\sim 40$ other haloes. The green shaded region indicates the scatter about the best fit line, which is 0.151 dex for haloes with mass in the range of $3\times 10^{9}\Msun < M_{200} <2\times 10^{10} \Msun$, 0.236 for halo mass in $2\times 10^{10}\Msun < M_{200} < 7\times 10^{10} \Msun$, 0.125 for halo mass in $7\times 10^{10}\Msun < M_{200} < 3\times 10^{11} \Msun$ and 0.052 for halo mass in $3\times 10^{11}\Msun < M_{200} < 3.5\times 10^{12} \Msun$. As might be expected, haloes with the highest masses we study ($\sim 10^{12}\Msun$) have high baryon fractions ($\sim 0.7$) with relatively small scatter (0.05 dex). As halo mass decreases, the baryon fraction decreases and the scatter increases reaching a maximum of $\sim 0.24$ dex in haloes of mass $\sim 4\times10^{10}\Msun$. However, at the lowest halo masses we study, below $10^{10}\Msun$, the scatter starts to decrease. By contrast the scatter in the stellar mass fraction increases below this scale. Thus it seems unlikely that stellar feedback is primarily responsible for the low baryon fractions. Rather, we suggest an increased importance of the UV background, which heats gas to above the virial temperature, thus preventing it from collapsing into the low mass haloes. The baryon fractions in low mass haloes are thus controlled primarily by the halo masses, and thus are independent of the large scatter in the stellar mass. ### Comparison with other zoom-in simulations In Fig. \[fig:fire\_com\] we compare the NIHAO results for the stellar and baryonic mass fractions with other recent state-of-the-art zoom-in simulations from @Voort16 and @Christensen16. Since these authors both use different halo mass definitions than we adopt here, we re-calculate the NIHAO results using a virial radius defined at $500\times$ critical density (left panels) and $100\times$ critical density (right panels). The simulations analyzed in @Voort16 are from the FIRE project [@Hopkins14], which uses a different hydrodynamical code and different sub-grid model for star formation and feedback. The simulations analyzed in @Christensen16 use a similar code as NIHAO, i.e., [gasoline]{}, but with important differences. NIHAO uses an upgraded version that improves mixing [@Keller14], and includes stronger feedback which was found to be necessary to delay star formation at early times [@Stinson13]. Despite these different codes employed, the baryon fractions in the three sets of simulations are in remarkably good agreement. For the massive haloes with M$_{500} > 10^{12} \Msun$, there are three galaxies from NIHAO and four from FIRE. There is a hint that more massive haloes have lower baryon fractions. Of course the caveat here is that neither NIHAO or FIRE includes AGN feedback, which could have a significant impact on the gaseous content of these haloes. For the stellar fractions, @Voort16 shows almost the same result as our work. Even at the low mass end, their two lowest mass galaxies may suggest the large scatter as found in NIHAO. However, @Christensen16 predicts one magnitude higher stellar fractions than what we find for halo masses below $10^{11.5} \Msun$. This difference is likely due to the additional early stellar feedback included in the NIHAO version of gasoline. In summary, while the stellar mass fractions are dependent on the sub-grid models, the baryon fractions (for haloes in the mass range $10^{10}\lta M_{200} \lta 10^{12}\Msun$) appear insensitive to the details of the simulation code. In particular all three codes predict a greater fraction of missing baryons in lower mass haloes. It suggests that the baryon fraction correlates strongly with halo mass, due to the deeper potential well of the halo and/or stronger ram pressure experienced by the outflowing gas. However, Fig. \[fig:fire\_com\] suggests that the specific implementation of the feedback model does not matter, as long as the feedback is efficient in driving outflows. Mass budget of the corona {#sec:corona} ------------------------- In Fig. \[fig:corona\], we present the radial distribution of gas in different phases at $z=0$, normalized to the total baryonic mass profile, such that, in the region far away from the central galaxy where the stars are rare, for a given halo the four phases add up to unity, each line is colour coded by the virial mass. All simulations share a common attribute. The cold gas (T$<10^4$K) is mostly located near the center ($R<0.2R_{200}$) where most stars in galaxies form. In contrast, the cool ($10^4$K$<$T$<10^5$K) and warm-hot ($10^5$K$<$T$<5\times10^6$K) gas are located at large distances with roughly constant fractions up to 2 times $R_{200}$. The hot gas (T$>5\times10^6$K) is a minority component for all galaxies in the NIHAO sample, with the maximum hot gas fraction at any radius being less than 5%. Despite these similarities, we find a considerably higher proportion of cool gas in lower mass galaxies (M$_{200} < 10^{11} \Msun$) in the whole corona region. For higher mass galaxies, warm-hot gas dominates the corona which signals stronger virial shocks and higher efficiency of feedback. Even beyond the virial radius, the cool and warm-hot gas has similar features as the gas within virial radius which reveals the gas surrounding galaxies within large distances is the major reservoir of baryons. Where are the missing baryons? {#sec:where} ------------------------------ Fig. \[fig:rps\] shows the cumulative fraction of total baryons for each simulation. Here the y-axis is the ratio between the baryonic to dark matter mass, $M_{\rm b}(<R) / M_{\rm dm}(<R)$, enclosed within a sphere of radius, $R$, normalized by the cosmic baryon-to-dark matter ratio, $\Omega_{\rm b}/\Omega_{\rm dm}$. Each solid curve represents a halo, and the curves are coloured by their halo mass (red for high masses to blue for low masses). Broadly speaking, the curves have a similar shape, with a normalization that depends on halo mass. They have a cusp in the central region where the stars and cold gas dominate, then become flat in the outer region. More massive haloes have higher baryon fractions at all radii. At small radii, the baryon to dark matter ratio is higher than the cosmic value due to gas dissipation. Beyond 0.5 virial radii, all haloes are missing baryons. Even beyond the virial radius, there is little change in the baryon fraction up to 2 virial radii. We thus conclude that the missing baryons are well outside of the virial radius. @Ford16 compare results for the cumulative baryon profile from two cosmological hydrodynamic simulations which employed different prescriptions for galactic outflow models. They constructed samples of simulated galaxies with a similar distribution of stellar masses to that of COS-Halos. In the hybrid energy/momentum driven winds model (“ezw”) 65% of all available baryons are inside the halo. This is broadly consistent with estimates of baryonic mass derived from @Werk14 and our finding in Fig. \[fig:budget\]. Comparing the cumulative baryon fraction profile in detail, Fig.10 in @Ford16 shows the baryon fraction within 0.1$\Rvir$ is only $\sim 30\%$ and the fraction gradually increases to 65% at $\Rvir$. In the NIHAO simulations the profile of galaxies with halo masses below $10^{11} \Msun$ in Fig. \[fig:rps\] have similar features, the more massive galaxies all have roughly flat slopes. The simplified constant wind outflow model (“cw”) shows a lower fraction at all radii inside $\Rvir$, even though the “ezw” and “cw” models generally gives similar observational absorption line properties. This suggests that the cumulative baryon fraction profile is complementary to the total amount of CGM gas for distinguishing between competing outflow models. To estimate how far the baryons escape, we measured the radius, $\Rbar$, within which the total baryon mass equals the fiducial baryonic mass defined by Eq. \[equ:mb\]. This is a lower limit to the true extent of the missing baryons since the baryon mass includes gas and stars that belong to nearby lower mass haloes. It is a radius within the high-resolution volume of the simulations since the mass fraction of low resolution dark matter particles in a shell between 0.9 and 1.0 $\Rbar$ shows most NIHAO galaxies have a fraction close to 0, with largest only around 10 - 20%. Fig. \[fig:missvr\] shows the baryon radius of each galaxy as function of the virial mass. In physical units, we find that the baryon radius generally increases with virial mass. When normalized by the virial radius, the distance baryons are ejected gradually decreases as halo mass increases, varying from $\Rbar/R_{200}\sim 5$ at a halo mass of $M_{200}\sim 10^{10}\Msun$ to $\sim 2$ at a halo mass of $M_{200}\sim 10^{12}\Msun$. In Fig. \[fig:inout\], we show the mass fractions of gas (in three phases) inside the virial radius (filled blue points) and between the virial and baryon radius (open red points). All fractions are relative to the fiducial baryonic mass within the virial radius. The fractions of cold gas, $f_{\rm cold}$, are shown in the left panel. Inside the virial radius $f_{\rm cold}$ increases gradually from zero at a halo mass of $10^{10}\Msun$ to $\sim 20\%$ at a halo mass of $10^{12}\Msun$. The gas outside the virial radius has the opposite and much stronger trend: $f_{\rm cold}\sim 50\%$ in haloes of mass $10^{10}\Msun$ and decreases to zero by halo masses of $10^{11}\Msun$. The fractions of cool gas, $f_{\rm cool}$, are shown in the middle panel. Inside the virial radius, $f_{\rm cool}$, has a maximum of 40% at a halo mass of $10^{11} \Msun$, and declines to less than 10% below and above halo masses of $10^{10}\Msun$ and $10^{12}\Msun$, respectively. For most haloes there is more cool gas outside than inside the virial radius. In haloes of mass $10^{10}\Msun$, $f_{\rm cool}\sim 80\%$, and decreases to less than 10% by a halo mass of $10^{12}\Msun$. The fractions for warm-hot gas,$f_{\rm warm}$, are shown in the right panel. The trends of the gas inside and outside the virial radius are quite similar, $f_{\rm warm}$ increases monotonically with halo mass with maximum values of $\sim 30\%$. The hot gas isn’t shown since it is negligible both inside and outside the virial radius across the whole mass range we study. We thus conclude that, for galaxies with halo masses in the range $10^{10} \lta M_{200}\lta 10^{11} \Msun$, the majority of baryons associated with the dark matter halo are in the cold and cool phases, and are located well outside of virial radius. For haloes in the mass range $10^{11} \lta M_{200}\lta 10^{12} \Msun$, the fractions of cold gas, cool gas and warm-hot gas are comparable. Comparison with observations of Milky Way mass haloes ----------------------------------------------------- Observations can gain information of CGM from absorption and emission lines. Although emission lines allow us to directly obtain a 3D picture of the distribution of gas in the CGM, emission line studies preferentially probe the dense gas closer to galaxies, since gas emissivity scales with the square of density. While the situation is improving with new facilities, e.g. @Hayes16, absorption lines are the most common observational constraints on the physical state of the CGM. The COS-HALOs survey is filling in details about the CGM at low redshift [@Peeples14; @Tumlinson11; @Tumlinson13; @Werk12; @Werk13; @Werk14]. For the CGM of low-redshift $\sim L^*$ galaxies ($\Mstar\sim 10^{10.5}\Msun$), @Tumlinson13 and @Peeples14 constrain the mass of the warm-hot CGM ($T \sim 10^{5-6.7}$K), @Werk14 provides a strict lower limit to the mass of cool material ($T \sim 10^{4-5}$K) in the CGM of these galaxies. In a study using X-rays, @Anderson13 place a constraints on the mass of hot gas ($T > 5\times10^6$K) residing in the extended hot halos. These observational constraints are shown in Fig. \[fig:comparison\] with the same colour scheme as in Fig.11 of @Werk14. The stellar mass fractions are based on halo abundance matching as described in @Kravtsov14, while the cold disk gas mass comes from @Dutton11. The upper limits to the missing fraction are calculated using the lower limits to all the fractions. There are two upper limits shown which correspond to the two measurements of the cool-gas fraction [@Werk14; @Stern16]. The lower limit to the missing fraction is consistent with zero. The black points and error bars show the mean values and standard deviation of the mass fraction of stars and different components of gas in our most massive galaxies ($3.49 \times 10^{11} \Msun$ $<$ M$_{200}$ $<$ $3.53 \times 10^{12} \Msun$). The gas is assigned to a range of temperature bins: cold gas (T $< 10^4$ K), cool gas ($10^4$ K $<$ T $< 10^5$ K), warm gas ($10^5$ K $<$ T $<5\times 10^6$ K) and hot gas (T $>5\times 10^6$ K). The observations and the simulations match well in every phase except the cool CGM gas, where the observations from @Werk14 find $3\times$ the mass that NIHAO simulations predict. If the observations are correct, the simulations have either ejected cool gas too far, or they have created a CGM with the wrong mix of gas temperatures. The total gas fractions (0.39 in COS-HALOs, 0.41 in NIHAO) suggest the latter option. However, @Stern16 developed a new method to constrain the physical conditions in the cool CGM from measurements of ionic columns densities. This new method combines the information available from different sight-lines during the photoionization modeling, and was applied to the COS-HALOs data yielding a total cool CGM mass within the virial radius of $1.3\times10^{10}\Msun$. The corresponding cool gas fraction is shown by the green hashed bar in Fig \[fig:budget\] and is in good agreement with our prediction. We should note that @Stern16 does not follow a same assumption on CGM ionization that assume all is produced through photoionization and no collisional ionization, however, @Gutcke16 finds the CGM in $L^*$ galaxies in NIHAO is mostly ionized by collisional ionization. Since @Crain07 and @Voort13 show that soft X-ray emission traces gas with temperatures a few times the virial temperature (T$_{\rm vir} = 10^6$ K for M$_{200} \sim 10^{12} \Msun$), we also estimate the hot gas fraction with lower boundaries ($3\times 10^6$ K and $10^6$ K) between the hot gas and warm-hot gas. The hot gas fractions increase to 0.00511 ($3\times 10^6$K) and 0.0417 ($10^6$ K), while the warm-hot gas fractions decrease by a corresponding amount. Even with these different definitions the NIHAO simulations are in good agreement with the observed warm-hot and hot gas fractions. As the Fig. \[fig:corona\] and Fig. \[fig:inout\] show, the cool gas is the most important component for most haloes in the NIHAO simulations ($10^{10} < M_{200}/\Msun < 10^{12}$). The cold and warm-hot gas only dominate for galaxies at the lower and higher mass ends. As the CGM of lower mass galaxies will soon be observed, Table \[tab:comparison\] lists information about CGM mass fractions of the different components of gas in haloes down to a halo mass of $\sim 10^{10}\Msun$. A recent estimation for the baryonic fractions from a set of eight Milky Way-sized zoom-in cosmological simulations [@Colin16] is generally consistent with our results. The fraction of cool gas from their simulations is even lower ($f_{\rm cool}$ = 0.034) than ours and the result from @Stern16. On the other hand, @Peeples14 gave a conservative observational estimate about warm-hot gas and showed that the fraction of gas in this phase is only 5%, which is the lower limit of the estimates from @Werk14 and less than the average values from @Colin16 ($f_{\rm warm-hot} \approx 24\%$) and our finding ($f_{\rm warm-hot} \approx 17\%$). Therefore, the mix of temperatures of the CGM has large uncertainties. More accurate knowledge of the physical properties of the CGM are necessary to better constrain feedback models and to understand the role of the CGM in galaxy formation. Summary {#sec:sum} ======= We have used the NIHAO galaxy simulation suite [@Wang15] to study the statistical features of the baryonic budget and distribution spanning halo masses of $\sim 10^{10}$ to $\sim 10^{12}\Msun$. NIHAO is a large (currently 88) set of high resolution cosmological zoom-in hydrodynamical galaxy formation simulations. As shown in previous papers the NIHAO galaxies reproduce several key observed scaling relations, and thus they make plausible predictions for the baryon budget in and around galaxies. We summarize our results as follows: - All of the NIHAO haloes have a lower baryon to dark matter ratio, inside the virial radius, than the cosmic baryon fraction (Fig. \[fig:budget\]). We refer to the cosmic baryon fraction associated with each dark matter halo as the fiducial baryons. - Lower mass haloes are missing a larger fraction of their fiducial baryons, even though they convert a much lower fraction of the baryons into stars (Fig. \[fig:budget\]). Similar trends are found by other recent simulations [@Christensen16; @Voort16] using different codes (Fig. \[fig:fire\_com\]). - The missing baryons have been expelled well beyond the virial radius, $R_{200}$, (Fig. \[fig:rps\]). Relative to the virial radius, the baryons are expelled to smaller radii in more massive haloes: $\Rbar \sim 5 R_{200}$ for $M_{200}=10^{10}\Msun$ and $\Rbar \sim 2 R_{200}$ for $M_{200}=10^{12}\Msun$ (Fig. \[fig:missvr\]). - Cold gas ($T<10^4$K) is mostly restricted to be within 0.2 virial radii (Fig. \[fig:corona\]). Cool gas ($10^4 < T < 10^5$K) dominates the baryonic mass outside the virial radius, as well as outside 20% $R_{200}$, at low masses ($M_{200}\lta 3\times 10^{11} \Msun$) while the warm-hot gas ($10^5 < T <5\times 10^6$K) dominates at high masses ($M_{200}\gta 3\times 10^{11} \Msun$) (Figs. \[fig:corona\] & \[fig:inout\]). - For the highest mass haloes in our study $\sim 10^{12}\Msun$ our simulations are consistent with the observed fractions [e.g. @Werk14] of stars, cold gas, warm and hot gas inside the virial radius (Fig. \[fig:comparison\]). - For the cool gas we predict $f_{\rm cool}=0.11\pm0.06$ which is significantly lower than the observations from COS-HALOs ($f_{\rm cool}=0.28-0.48$), but is in excellent agreement with the analysis of @Stern16. We hope our results will motivate observers to obtain more accurate measurements of the mass fractions in different phases of the CGM over a wide range of galaxy masses, and simulators to make the corresponding predictions. [ccccc]{} \[tab:comparison\] Acknowledgments {#acknowledgments .unnumbered} =============== We thank the two anonymous referees whose suggestions greatly improve the paper. We thank Freeke van de Voort for kindly sharing data of the FIRE simulation. [Gasoline]{} was written by Tom Quinn and James Wadsley. Without their contribution, this paper would have been impossible. The simulations were performed on the [theo]{} cluster of the Max-Planck-Institut für Astronomie and the [hydra]{} cluster at the Rechenzentrum in Garching; and the Milky Way supercomputer, funded by the Deutsche Forschungsgemeinschaft (DFG) through Collaborative Research Center (SFB 881) “The Milky Way System” (subproject Z2), hosted and co-funded by the Jülich Supercomputing Center (JSC). We greatly appreciate the contributions of all these computing allocations. AAD, GSS and AVM acknowledge support through the Sonderforschungsbereich SFB 881 “The Milky Way System” (subproject A1) of the German Research Foundation (DFG). The analysis made use of the pynbody package [@Pontzen13]. The authors acknowledge support from the MPG-CAS through the partnership programme between the MPIA group lead by AVM and the PMO group lead by XK. LW acknowledges support of the MPG-CAS student programme. XK acknowledge the support from 973 program (No. 2015CB857003, 2013CB834900), NSFC project No.11333008 and the “Strategic Priority Research Program the Emergence of Cosmological Structures” of the CAS(No.XD09010000). Anderson, M. E., Bregman, J. N., Dai, X. 2013, , 762, 106 Agertz, O., Moore, B., Stadel, J. 2007, , 380, 963 Behroozi, P. S., Wechsler, R. H., & Conroy, C. 2013, , 770, 57 Bell, E. F., McIntosh, D. H., Katz, N., Weinberg, M. D., 2003, , 585, 117 Bregman, J. N. 2007, ARAA, 45, 221 Cen, R. Y., Ostriker, J. P. 1999, , 514, 1 Christensen, C. R., Dav[é]{}, R., Governato, F., et al. 2016, , 824, 57 Colin, P., Avila-Reese, V., Roca-Fabrega, S., et al. 2016, , in press Crain, R. A., Eke, C. R., Frenk, C. S., et al. 2007, , 377, 41 Dav[é]{}, R. 2009, ASPC, 419, 347D Dav[é]{}, R., Oppenheimer, B. D., Katz, N., et al. 2010, , 408, 2051 Dutton, A. A., Conroy, C., van den Bosch, F. C., et al. 2011, , 416, 322 Dutton, A. A. 2012, , 424, 3123 Dutton, A. A., & Macci[ò]{}, A. V. 2014, , 441, 3359 Ford, A. B., Oppenheimer, B. D., Dav[è]{}, R., et al. 2013, , 432, 89 Ford, A. B., Werk, J. W., Dav[è]{}, R., et al. 2016, , 459, 1745 Fukugita, M., Hogan, C. J., Peebles, P. J. F. 1998, , 503, 518 Gutcke, T. A., Stinson, G. S., Macci[ò]{}, A. V., et al. 2016, arXiv:1602.06956, MNRAS in press Haider, M., Steinhauser, D., Vogelsberger, M., et al. 2016, , 457, 3024 Hayes, M., Melinder, J., [Ö]{}stlin, G., et al. 2016, arXiv:1606.04536 He, P., Feng, L. L., Fang, L. Z. 2005, , 623, 601 Hopkins, P. F., Kere[š]{}, D., O[ñ]{}orbe, J., et al. 2014, , 445, 581 Hummels, C. B., Bryan, G. L., Smith, B. D., et al. 2013, , 430, 1548 Keller, B. W., Wadsley, J., Benincasa, S. M., & Couchman, H. M. P. 2014, , 442, 3013 Kravtsov, A., Vikhlinin, A., & Meshscheryakov, A. 2014, arXiv:1401.7329 Marasco, A., Marinacci, F., Fraternali, F. 2013, , 433, 1634 McGaugh, S. S., Schombert, J. M., de Blok, W. J. G., Zagursky, M. J. 2010, , 708, 14 Moster, B. P., Naab, T., & White, S. D. M. 2013, , 428, 3121 Muratov, A. L., Ker[ě]{}s, D., Faucher-Gigu[è]{}re, C., et al. 2015, , 454, 2691 Oppenheimer, B. D., Crain, R. A., Schaye, J., et al. 2016, , 460, 2157 Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2014, , 571, AA16 Peeples, M. S., Werk, J. K., Tumlinson, J., et al. 2014, , 786, 54 Persic, M., Salucci, P. 1992, , 258, 14 Pontzen, A., Ro[š]{}kar, R., Stinson, G., & Woods, R. 2013, Astrophysics Source Code Library, 1305.002 Schaye, J., Crain, R. A., Bower, R. G., et al. 2015, , 446, 521 Shen, S., Wadsley, J., & Stinson, G. 2010, , 407, 1581 Sharma, P., McCourt, M., Parrish, I. J., Quataert, E. 2012, , 427, 1219 Shull, J. M., Smith, B. D., Danforth, C. W. 2012, , 759, 23 Shull, J. M. 2014, , 784, 142 Sokolowska, A., Mayer, L., Babul, A., Madau, P., Shen, S. 2016, , 819, 21 Stern, J., Hennawi, J. F., Prochaska, J. X., & Werk, J. K. 2016, , 830, 87 Stinson, G. S., Seth, A., Katz, N., et al. 2006, , 373, 1074 Stinson, G. S., Brook, C., Prochaska, J. X., et al. 2012, , 425, 129 Stinson, G. S., Brook, C., Macci[ò]{}, A. V., et al. 2013, , 428, 129 Stinson, G. S., Dutton, A. A., Wang, L., et al. 2015, , 454, 1105 Suresh, J., Rubin, K. H. R., Kannan, R., et al. 2015, arXiv:1511.00687 Thom, C., Tumlinson, J., Werk, J. K. 2012, , 758, L41 Tumlinson, J., Thom, C., Werk, J., et al. 2011, Science, 334, 948 Tumlinson, J., Thom, C., Werk, J., et al. 2013, , 777, 59 van de Voort, F., Schaye, J. 2012, , 423, 2991 van de Voort, F., Schaye, J. 2013, , 430, 2688 van de Voort, F., Quataert, E., Hopkins, P. F., et al. 2016, , 463, 4533 Vogelsberger, M., Genel, S., Springel, V., et al. 2014, , 444, 1518 Wadsley, J. W., Stadel, J., & Quinn, T. 2004, , 9, 137 Wadsley, J. W., Veeravalli, G., & Couchman, H. M. P. 2008, , 387, 427 Wang, L., Dutton, A. A., Stinson, G. S., et al. 2015, , 454, 83 Weinmann, S. M., Pasquali, A., Oppenheimer, B. D., et al. 2012, , 426, 2797 Werk, J. k., Prochaska, J. X., Thom, C., et al. 2012, , 198, 3 Werk, J. k., Prochaska, J. X., Thom, C., et al. 2013, , 204, 17 Werk, J. k., Prochaska, J. X., Thom, C., et al. 2014, , 792, 8 Yoshida, N., Furlanetto, S. R., Hernquist, L. 2005, , 618L, 91 Zhu, W., Feng, L. L., Fang, L. Z. 2011, , 415, 1093 Resolution test =============== Fig. \[fig:budget\_res\] shows the budget fraction of stars, cold, cool and warm-hot gas color coded by the number of dark matter particles inside the virial radius. Since one can imagine that with increased resolution, higher densities can be achieved in the CGM, potential leading to more cooling and larger cold/cool gas fractions. In our simulations we find no dependence of stars and gas fractions with particle number, indicating these quantities are not sensitive to numerical resolution using the NIHAO sub-grid model. \[lastpage\] [^1]: [email protected]

ArXiv

--- abstract: 'Recent advances in selected CI, including the adaptive sampling configuration interaction (ASCI) algorithm and its heat bath extension, have made the ASCI approach competitive with the most accurate techniques available, and hence an increasingly powerful tool in solving quantum Hamiltonians. In this work, we show that a useful paradigm for generating efficient selected CI/exact diagonalization algorithms is driven by fast sorting algorithms, much in the same way iterative diagonalization is based on the paradigm of matrix vector multiplication. We present several new algorithms for all parts of performing a selected CI, which includes new ASCI search, dynamic bit masking, fast orbital rotations, fast diagonal matrix elements, and residue arrays. The algorithms presented here are fast and scalable, and we find that because they are built on fast sorting algorithms they are more efficient than all other approaches we considered. After introducing these techniques we present ASCI results applied to a large range of systems and basis sets in order to demonstrate the types of simulations that can be practically treated at the full-CI level with modern methods and hardware, presenting double- and triple-zeta benchmark data for the G1 dataset. The largest of these calculations is Si$_{2}$H$_{6}$ which is a simulation of 34 electrons in 152 orbitals. We also present some preliminary results for fast deterministic perturbation theory simulations that use hash functions to maintain high efficiency for treating large basis sets.' author: - 'Norm M. Tubman, C. Daniel Freeman, Daniel S. Levine, Diptarka Hait, Martin Head-Gordon, K. Birgitta Whaley' title: Modern Approaches to Exact Diagonalization and Selected Configuration Interaction with the Adaptive Sampling CI Method --- Introduction {#sec:intro} ============ Selected configuration interaction techniques (SCI) have seen a recent revival for performing quantum chemistry simulations, especially for treating strongly correlated systems [@tubman2016-1; @holmes2016; @evan2016; @evangelista2014; @wenjian2016]. Much of the recent interest in selected CI was sparked by the demonstration that the adaptive sampling CI method (ASCI) can attain an accuracy comparable to the density matrix renormalization group (DMRG) for Cr$_{2}$ with relatively little computational cost [@tubman2016-1; @holmes2016]. Selected CI techniques can be applied to a wide range of atomic, molecular, and solid state chemical systems. As such, they are a fundamentally different from DMRG, which is often the method of choice for low dimensional solid state systems. Notable selected CI approaches developed in the last few years include ASCI [@tubman2016-1] and the similarly titled adaptive CI approach (ACI) [@evan2016]. An integral driven search extension to ASCI, known as heat bath CI (HBCI), was also recently developed [@holmes2016]. Additionally the classic configuration interaction perturbatively selected iteratively (CIPSI) algorithm has also seen recent advances [@huron1973; @harrison1991; @scemama2013; @giner2013; @evan2016]. The development of the ASCI method [@tubman2016-1] also demonstrated a connection between selected CI and the full configuration interaction quantum Monte Carlo (FCIQMC) technique  [@booth2009; @shepherd2012; @kolodrubetz2013; @hande2014; @booth2013; @daday2012; @thomas2015; @cleland2012; @blunt2015]. In contrast to FCIQMC however, the ASCI method utilizes a significantly more computationally efficient deterministic approach for searching Hilbert space, allowing it to produce nearly identical results at significantly reduced cost. Overall, the ASCI method is complementary to and competitive with density matrix renormalization group (DMRG)[@tubman2016-1] with regards to accuracy and computational cost. For one-dimensional systems DMRG is the method of choice due to the low entanglement of the wave functions [@chan2002; @white1999; @white1993; @amaya2015; @hachmann2007; @mizukami2013; @sharma2014]. However, DMRG is not an optimal choice when the entanglement is large, and virtually all chemically relevant systems have large entanglement [@tubman2012; @tubman2013; @tubman2013-1; @tubman2014; @tubman2015; @amaya2015]. This makes ASCI superior for many systems which include 2D and 3D systems, as well as for systems with large basis sets and for simulations of excited states [@stoudenmire2012; @schollwock2005]. The ASCI algorithm of ref. [@tubman2016-1] improves upon the perturbative approach of the CIPSI algorithm [@huron1973; @buenker1974; @evangelisti1983; @bagus1991; @harrison1991; @giner2013; @caffarel2015; @maynau2011] by introducing approximate search algorithms for finding the important determinants. In the original application of the ASCI algorithm, simulations of up to 48 electrons in 42 orbitals were made [@tubman2016-1; @Lehotola2017]. This work describes advances in the methodology that permit simulations with hundreds of orbitals on single workstations. Few other methods are known to provide this level of speed and accuracy. Auxiliary field quantum Monte Carlo (AFQMC) is a competitive method, but it has several limitations that prevent consistent production of chemically accurate numbers [@zhang2003; @wirawan2015; @chang2016]. However there is always room to improve AFQMC type techniques with better trial wave functions [@borda2018]. The algorithms developed in this work are fundamentally driven by sorting based algorithms that are relevant to all current selected CI approaches. Our approach addresses the critical issue that selected CI methods are largely based on manipulation of large amounts of data. As such, selected CI methods need to be designed to efficiently process and move data to and from the CPU. Many of the improvements presented here make use of optimized tools and libraries that have not yet made their way into the selected CI literature. We demonstrate here all these newly designed algorithms as extensions to the ASCI formalism. This includes algorithms that allow for different levels of parallelization and computing architectures, such as GPUs. With these ideas, we are able optimize the ASCI algorithm beyond what has been done with previous selected CI approaches. System Dets Basis Main ASCI(s) PT2(s) Total time ASCI Energy (Ha) Ref. Energy (Ha) ------------------- ------------- --------- -------------- -------- ------------ ------------------ --------------------------- C$_{2}(12e,28o)$ 10$^{4}$ cc-pVDZ 7 4 11 -75.731895 -75.731958 [@amaya2015] C$_{2}(12e,28o)$ 10$^{5}$ cc-pVDZ 110 44 154 -75.731954 C$_{2}(12e,28o)$ 10$^{6}$ cc-pVDZ 3740 570 4310 -75.731962 C$_{2}(12e,60o)$ 10$^{4}$ cc-pVTZ 45 70 115 -75.808698 -75.809285 [@amaya2015] C$_{2}(12e,60o)$ 10$^{5}$ cc-pVTZ 360 600 960 -75.809190 C$_{2}(12e,110o)$ 10$^{4}$ cc-pVQZ 300 401 701 -75.856781 -75.85728 [@amaya2015] C$_{2}(12e,110o)$ 10$^{5}$ cc-pVQZ 1020 4290 5310 -75.856822 Cr$_{2}$(24e,30o) 10$^{4}$ SVP 18 11 29 -2086.417192 -2086.420948 [@amaya2015] Cr$_{2}$(24e,30o) 10$^{5}$ SVP 100 112 212 -2086.419546 Cr$_{2}$(24e,30o) 10$^{6}$ SVP 1250 1111 2361 -2086.420438 Cr$_{2}$(24e,30o) 2\*10$^{6}$ SVP 2680 2215 4895 -2086.420517 The different techniques presented in this work are outlined in the sections designated below. - Constructing the Hamiltonian and density matrices (Section \[sec:hamcons\]) - Search and pruning based on fast sorting algorithms (Section \[sec:searchmain\]) - Other algorithm improvements (Section \[sec:misc\]) Fast diagonal matrix elements (Section \[sec:fdiag\]) Informed bit string representations (Electron representation and Difference representation) (Section \[sec:ibit\]) Hashed bit string representations (Section \[sec:hbit\]) After we present these techniques, we apply our algorithm to 55 benchmark molecular systems in the G1 set (Section \[sec:results\]). Before presenting the details of the new algorithms however, we first present some timings and accuracy results to demonstrate their capabilities. Comparisons Dets Basis Time PT2 E(Variational) E(HBCI+PT2) E(Exact PT2) ------------------------------------- -------- --------- ---------- ---------------- -------------- -------------- -- C$_{2}$(8e,58o) HBCI [@sharma2017] 142467 cc-pVTZ 80 -75.7738 -75.7846(3) C$_{2}$(8e,58o) ASCI 50000 cc-pVTZ 60 -75.768939 -75.784113 C$_{2}$(8e,58o) ASCI 100000 cc-pVTZ 117 -75.773560 -75.784468 C$_{2}$(8e,58o) ASCI 142467 cc-pVTZ 166 -75.775386 -75.784589 F$_{2}$(14e,58o) HBCI [@sharma2017] 395744 cc-pVTZ 120 -199.2782 -199.2984(9) F$_{2}$(14e,58o) ASCI 20000 cc-pVTZ 60 -199.254301 -199.295491 F$_{2}$(14e,58o) ASCI 100000 cc-pVTZ 300 -199.270670 -199.296289 F$_{2}$(14e,58o) ASCI 300000 cc-pVTZ 891 -199.278140 -199.296686 F$_{2}$(14e,58o) ASCI 395744 cc-pVTZ 1163 -199.279209 -199.296767 New Algorithms and Timings {#sec:algotime} ========================== Figure \[fig:flowchart\] presents an overview of the ASCI algorithm. The main components of the algorithm involve building the Hamiltonian, diagonalization, search, and second order perturbation theory (PT2). We break up the wave function construction into a growth and a refinement process. The growth process involves several steps in which the wave function is increased in size and the orbitals are rotated. The refinement process is a set of final steps in which the size of the wave function is kept fixed, but the quality is improved. In the following sections we introduce new algorithms for ASCI, which include approaches for search and Hamiltonian construction. We present several algorithm choices for these components. Our current recommendation is to use the following algorithms for constructing a selected CI: *New ASCI search* (described in this work), *dynamic bit masking* for Hamiltonian construction (described in this work), *fast orbital rotations* (described in this work) and a *deterministic PT2* based on sorting, which will be described in a future work. We also introduce the residue arrays algorithm for Hamiltonian construction, which may be more efficient than dynamic bit masking for certain Hamiltonians. In particular, the size of a residue array is determined by the number of electrons being simulated, thus a residue array is more efficient in the limit in which a system has a small number of electrons. To illustrate the current set of algorithms, and to compare them to previously published work, we present a series of timings in Tables \[tab:timings\], \[tab:hbcomp\], and \[tab:comparisons\]. In Table \[tab:timings\], we present ASCI timing results for C$_{2}$ and Cr$_{2}$. These timings represent a careful simulation of the variational wave function that is designed to generate highly accurate results. More aggressive approximations can lead to much faster calculation of the variational wave function. In Table \[tab:hbcomp\] we demonstrate the timings for such an aggressive approach in order to make a comparison to HBCI. The timings for the PT2 simulations presented here represent the fastest published timings for deterministic algorithms that we are aware of. Indirect comparisons to stochastic PT2 in Table \[tab:hbcomp\] demonstrate that stochastic algorithms are only competitive if large stochastic errors are acceptable. Recent papers [@sharma2017] have suggested that stochastic PT2 is the best way to efficiently simulate large system with selected CI. However, our comparisons to the stochastic method, indicate that it is unclear when this would be true and for which ranges of accuracy. In fact, the errors on stochastic methods get even larger when energy differences are taken, which is important for chemically relevant results. The details of the sorting based deterministic perturbation theory algorithms will be published in a future work, and we hope in that work to provide a detailed analysis to compare with stochastic approaches to perturbation theory. Comparisons Dets Basis Time(s) E+PT2(Ha) -------------------------------------------- ------------- --------- --------- ------------- -- -- -- old ASCI Cr$_{2}$(24e,30o) [@tubman2016-1] 10$^{4}$ SVP 1000 -2086.4177 old ASCI Cr$_{2}$(24e,30o) [@tubman2016-1] 10$^{6}$ SVP 133000 -2086.4203 CIPSI Cr$_{2}$(24e,30o) [@quantumpackage] 4\*$10^{4}$ SVP 4290 -2086.41818 ASCI Cr$_{2}$(24e,30o) 4\*10$^{4}$ SVP 82 -2086.41808 HBCI [@holmes2016] Cr$_{2}$(24,30) 4\*10$^{4}$ SVP 120 -2086.42130 ASCI (fast) Cr$_{2}$(24e,30o) 4\*10$^{4}$ SVP 55 -2086.42099 ASCI C$_{2}$(12e,60o) 4\*10$^{4}$ cc-pVTZ 280 -75.80898 HBCI [@holmes2016] C$_{2}$(12e,60o) 4\*10$^{4}$ cc-pVTZ 540 -75.80873 : A comparison of the current work with the original ASCI results and other selected CI methods. In the original ASCI paper (labled ’old ASCI’), slow but memory-efficient algorithms were used for the PT2. The Cr$_{2}$ comparisons in the middle columns demonstrate the difference between accurate wave functions (CIPSI and ASCI) and wave functions in which a minimal search algorithm is used such as in HBCI, and ’ASCI (fast)’. For ’ASCI (fast)’ the search parameters are turned down significantly so as to be similar to HBCI. Although the energies for HBCI and ’ASCI (fast)’ are close to the converged DMRG results, this is because, for these inaccurate wave functions, the energy is converging from below. The HBCI energy for Cr$_{2}$ will observe significant non-monotonic behavior before converging. The timings for the different algorithms were performed on different computing architectures. The CIPSI simulation was performed on a single core of an Intel Xeon E5-2680v3 processors of 2.5 GHz. The ASCI simulations were performed on a single core of an Intel Xeon E5-2620 v5 processor of 2.10 GHz. The HBCI timings were taken directly from ref. [@holmes2016]. The timings are meant only to give a semi-quantitative comparison, as the timings for the ASCI includes many extra steps which are not included in the CIPSI and HBCI timings, as discussed in the main text.[]{data-label="tab:comparisons"} The timings for ASCI in Table \[tab:timings\] include the initial SCF calculations, orbital rotations, and extra diagonalization steps that are required after performing an orbital rotation. Additionally after the growth of the wave function we perform several refinement iterations, which includes both search and diagonalization steps, to improve the wave function as much as possible before the PT2. The column labeled ‘Main ASCI’ incorporates all of these different steps. When using the new ASCI algorithms for Hamiltonian building, orbital rotations, and search, the Hamiltonian diagonalization is now the dominant bottleneck in the limit of large determinants. Performing an accurate search step can be done quickly with the new ASCI search algorithms, and thus is no longer the bottleneck of a simulation. In Table \[tab:comparisons\] we present a comparison to selected CI calculations from previous published works, as well as CIPSI simulations as implemented in the Quantum Package suite for the Cr$_2$ dimer [@quantumpackage]. We present a comparison to our original ASCI implementation ’old ASCI’ (the details of which can be found in Ref. [@tubman2016-1]), which uses a memory efficient but computationally slow PT2 correction. The next sets of results are generated with ’HBCI’ and ’ASCI (fast)’, which use a less accurate, but faster, search algorithm than ASCI and CIPSI. These results illustrate problems that arise from using an inaccurate search algorithm and generating medium quality variational wave functions. We observe in some situations that low quality wave functions can result in significant non-variational behavior for the perturbation theory correction. Unconverged wave functions can also lead to slow energy convergence from above, as can be seen in comparing HBCI to ASCI results for the C$_2$ dimer in Table \[tab:comparisons\]. In Table \[tab:hbcomp\], we remove the orbital rotations and extra diagonalizaton steps after the growth algorithm in order to make more direct comparisons (in terms of computational timing) to HBCI. In these comparisons HBCI perturbation theory results are generated using a stochastic algorithm in order to treat large basis sets. This is not necessary for ASCI, and all the ASCI simulations presented in this work are completely deterministic. The results presented here show that ASCI generates lower energy variational wave functions than HBCI for the same number of determinants. For example, the C$_{2}$ ASCI($N_{dets}$=100,000) and F$_{2}$ ASCI($N_{dets}$=300,000) give comparable variational energies to HBCI but with 42,000 and 95,000 fewer determinants, respectively. This means that the HBCI wave function is missing many important determinants, as evidenced by the more compact ASCI wave function. It needs to be emphasized that the PT2 comparisons have to be done with care, since the stochastic HBCI results have both systematic bias and stochastic bias, whereas the ASCI algorithm presented here has no stochastic errors. For timing comparisons of the PT2 results, we will consider the C$_{2}$ ASCI($N_{dets}$=100,000) and F$_{2}$ ASCI($N_{dets}$=300,000) simulations because they have nearly the same variational energy as the HBCI results. The error bar for stochastic perturbation theory decreases like $\frac{1}{\sqrt{N_{samp}}}$, where $N_{samp}$ is the number of samples for computing the stochastic average. For an accuracy of 0.1 mHa, the C$_{2}$ and F$_{2}$ ASCI perturbation theory timings would be 6 times faster and 11 times faster than HBCI, respectively. This is for a 68% likelyhood that the result is within 0.1 mHa of the actual answer. A more definitive result with a 95% likelihood would be would be 24 and 44 times faster with ASCI than HBCI. To calculate energy differences, such as for calculating atomization energies, another factor of two is needed to make sure the energy difference is 0.1 mHa accurate, for which ASCI would be 48 and 88 times faster than HBCI for these two systems respectively. In these scenarios, the ASCI results would still be more accurate than HBCI, as the ASCI results are calculated without stochastic errors. For an accuracy of 0.01 mHa, the ASCI timings presented are several hundred times faster than the HBCI. In the case of C$_{2}$, the fully converged ASCI simulation, which we perform with orbital rotations, has an energy of 75.78508(Ha). The results in Table \[tab:comparisons\] show slow convergence because these results are calculated with Hartree-Fock orbitals. In comparing to ASCI, these results show how both systematic errors and stochastic errors play a role in the HBCI result. The HBCI F$_{2}$ results are significantly non-variational, which is likely due to the large stochastic errors, coupled with a variational wave function that is of only medium quality. We also note that for all ASCI simulations in this work, other than those in Table \[tab:hbcomp\], orbital rotations are always included and significantly improve the convergence. Orbital rotations allow ASCI to generate a more compact wave function with fewer determinants. The calculation of the 1-RDM used in orbital rotations in ASCI are calculated for almost no cost (described in Section \[sec:fdiag\]). We also note here that comparable DMRG simulations use a significant amount of CPU time and memory compared to the results presented here. The largest DMRG calculations in ref. [@amaya2015] (butadiene: 22e,88o) reportedly used 850 GB of memory, 17 TB of disk space, and over a 1000 CPU hours per sweep with a bond dimension of 3000. Although similar sized systems (and even larger) are presented in this work, we note that DMRG is variational whereas ASCI+PT2 is not. Regardless, both methods produce high level results and can be used to provide validation of each other as they use very different approaches to solving quantum Hamiltonians. Theoretical Background: selected CI algorithms {#sec:theory} ============================================== The main idea behind the selected CI approach is to perform diagonalization on a determinant space in which one captures as many important degrees of freedom as possible. This is the principle behind all exact diagonalization and CI techniques, although most methods do not allow for explicit searching for important determinants [@lauchli2011; @vogiatzis2017; @szabo:book; @gan2006; @gan2005; @sherrill1999; @abrams2004; @szalay2012; @bender1969; @buenker1978; @roth2009; @bagus1991]. Thus in contrast to more traditional CI techniques, the idea of using a selected CI approach is to generate determinants that account for 90% or more of the top contributions to the full CI wave function. In the ASCI method, as well as in most selected CI methods, a wave function, $\psi_{k}$, is iteratively improved to reach a desired accuracy. Here we use $k$ to index the current iteration of the algorithm. The search part of the algorithm requires two rules: a selection criterion to determine what part of Hilbert space to search from (pruning) and a ranking criterion to determine the best determinants to include in the improved wave function $\psi_{k+1}$. For the algorithms considered here the ranking criterion is derived from a consistency relationship among the coefficients of eigenstates of the Schrödinger equation. The consistency equation is given as follows if we consider an expansion of an eigenstate in terms of its coefficients $C_i$: $$C_{i} = \frac{\sum_{j \ne i}H_{ij}C_{j}}{(E-H_{ii})}$$ where $H_{ij}$ is the Hamiltonian matrix element between the $i$th and $j$th determinant, and $E$ is the energy of the eigenstate. If we reinterpret this equation, we can use it to predict a new and better set of determinants for expanding $\psi_{k+1}$ by taking the LHS as an estimate of the magnitude of the expansion coefficients, $$A_{i} = \frac{\sum_{j \ne i}H_{ij}C^k_{j}}{(E_k-H_{ii})}, \label{eqn:rankeqn}$$ where $C^k_j$ is the CI expansion coefficient of the $j$th determinant and $E_k$ is the energy of the wave function in the $k$th iteration. It is useful to think of $C^{k}_i$ as the coefficients of an input wave function, and the output coefficients, $A_{i}$, as an estimate of coefficients of an improved wave function (close to the ground state). Since the goal of selected CI is to include the most important weight determinants in the expansion, we use this equation to define a ranking, where $A_i$ is the rank value of the $i$th determinant. These $A_i$ coefficients are related to a first-order perturbation estimate for CI coefficients in many body perturbation theory [@huron1973]. In practice this iterative approach generates all the top contributions to the wave function. Having the top contributions is critical to obtain highly accurate energies, as was recently shown with the ASCI method in combination with second order many body perturbation theory [@tubman2016-1]. The perturbation theory energy is calculated with Epstein-Nesbet PT2 which is given as, $$E_{PT2} = \sum_{i}\frac{|\langle\psi|H|D_{i}\rangle|^{2}}{H_{ii} -E_{}}. \label{eqn:en}$$ To facilitate the following discussion, we introduce the following notation: $\psi_k$ is the wave function at the $k$th iteration of the selected CI algorithm, $C_i$ is the coefficient of the $i$th determinant in the expansion of $\psi_k$, $D_i$ is the $i$th determinant in $\psi_k$, $\{D\}$ is the set of all determinants in $\psi_k$, {$D_{i}^{sd}$} is the set of all determinant which are single or double excitations from determinant $i$. In general, $i\in \{D\}$, and {$D^{sd}$} is the set of all determinants which are single and double excitations from $\{D\}$ (that is {$D^{sd}$}=$\bigcup_i$ {$D_i^{sd}$}). We describe two determinants which differ by a single or double excitation as *connected*. The notation we use here is summarized in Table \[tab:symbols\]. ---------------- -------------------------------------------------------------------------------------------- $\psi_k$ The wave function in the current ASCI step $C_i$ The coefficients of the $i$th deteminant $D_i$ of $\psi_k$ $D_{i}$ The $i$th determinant in $\psi_k$ {$C$} The set of coefficients in $\psi_k$ $\{D\}$ The set of determinants in $\psi_k$ {$D^{sd}$} The set of all single and double excitations that are connected to $\psi_k$ {$D_{i}^{sd}$} The set of all single and double excitations that are connected to the determinant $D_{i}$ $N_{tdets}$ Number of determinants in the current wave function. $N_{cdets}$ Core space size, used for pruning in ASCI ---------------- -------------------------------------------------------------------------------------------- : A reference to a list of the symbols used in this work. []{data-label="tab:symbols"} ### Introduction to the Computational Challenge {#sec:cc} The modern implementations to exact diagonalization and selected CI described in this work are in part focused on calculating Eq. \[eqn:rankeqn\] and Eq. \[eqn:en\] efficiently. The main issue for using these equations is that they involve all the elements in {$D^{sd}$}, which can be extremely large. For general systems of interest a set size of 10$^{12}$ and larger is not uncommon. Furthermore, for each element $i$ in {$D^{sd}$}, all ways in which the Hamiltonian connects from $\{D\}$ to $i$ must be found. This can becomputationally expensive, especially for wave functions that can have between 10$^{5}$ to 10$^{9}$ determinants (although recent full CI simulations have recently gone all the way up to 10$^{12}$ elements [@vogiatzis2017]). The straightforward and slow approach to performing this calculation is as follows: for all $i$ in $\{D\}$ generate the set {$D_{i}^{sd}$}. Then, for each element $a$ in {$D_{i}^{sd}$}, search $\{D\}$ to find all the elements $b$ such that $H_{ab}$ is non-zero (that is $b$ is singly or doubly connected to $a$). This approach scales as $\mathcal{O}((N_{tdets}*N_{occ}*N_{virt})^{2})$, where $N_{occ}$ is the number of occupied orbitals and $N_{virt}$ is the number of virtual orbitals. The asymptotic scaling for $N_{tdets}$ (that is, for a fixed number of orbitals we ignore $N_{occ}$ and $N_{virt}$) can be seen as follows: the number of terms in {$D^{sd}$} is proportional to $N_{tdets}$, and, for each term, a search over the $N_{tdets}$ determinants of the wave function is needed to find all connections. Constructing the Hamiltonian also requires a similar algorithm, where connections among the elements of $\{D\}$ need to be found. Improvements to this problem has been studied previously with the introduction of residue trees [@stampfub2005]. For both the search and Hamiltonian construction, we present new algorithms in the modern implementation sections. Input $\psi_k$ and an initially empty array {$V$} that will hold information in pairs of (determinant bit string, ranking value) Generate $D^{sd}$ from $\psi_{k}$. - Iterate through $i \in \{D_{}\}$, and find {$D_{i}^{sd}$} - If CIPSI search, generate all {$D_{i}^{sd}$} - If original ASCI search, prune terms in {$D_{i}^{sd}$} based on {C$_{}$} - If HB search, prune terms in {$D_{i}^{sd}$} based on {$C_{}$} and $H_{ij}$ Calculate a ranking either approximately or exactly from Eq. \[eqn:rankeqn\]. Gather all the top contributing determinants to determine $\psi_{k+1}$ for use in the next iteration ### CIPSI {#sec:CIPSI} A straightforward idea for searching for important determinants is to generate all terms in $\{D^{sd}\}$, rank them by Eq. \[eqn:rankeqn\], and take the top $N_{tdets}$ of them to form the basis for creating $\psi_{k+1}$. This is the approach used in CIPSI and is defined by a single parameter $N_{tdets}$ that determines the number of determinants that are retained in a simulation. No effort is made to prune the search space in such an approach. As demonstrated explicitly in the ASCI approach, this pruning is important for computational efficiency and results in no significant loss in accuracy. For most CIPSI applications the generation of the search space is the most costly step of the calculation. Pruning the search space and approximating the ranking algorithm become critical for going to larger system sizes. ### ASCI {#sec:ASCI} The ASCI algorithm, introduced in ref. [@tubman2016-1], allows selected CI approaches to be more computationally efficient in the search algorithm. In ASCI, only a subset of {$D^{sd}$} are considered for ranking, resulting in a significant increase in efficiency. This is made possible by having a good criterion (using the structure of the wave function) for quickly determining which sets $\{D_i^{sd}\}$ are likely to have highly ranked connections, allowing one to avoid searching unimportant parts of Hilbert space. The ASCI pruning algorithm uses the magnitude of the {$C_{}$}. Only connections from the top $N_{cdets}$ determinants in $\psi_k$ (where these are ranked by $C_i$ value) are considered. Each ASCI algorithm iteration is parametrized by two determinant subspaces: a *core space* of size $N_{cdets}$ and a *target space* of size $N_{tdets}$. This leads to an iterative algorithm with significantly more efficient performance than CIPSI methodologies. ### Integral driven search extension of ASCI (Heat Bath) {#sec:heatbath} The speed of the search component of ASCI may be accelerated by sacrificing some accuracy. In the HBCI approach, a new pruning criterion was introduced, using a combination of the wave function coefficients and the Hamiltonian matrix elements. This can be understood as an integral driven search approach in the ASCI formalism. The vast majority of connections in the search step are due to double excitations, whose matrix elements are just the antisymmetrized two-electron integrals, which can be sorted once at the beginning of the algorithm run. The HBCI pruning criterion is as follows: all single excitation and only those double excitations such that $c_iH_{ij}>\epsilon$ are selected for ranking, where $\epsilon$ is a parameter that replaces $N_{cdets}$ to define the integral driven search algorithm. In the original heat bath approach [@holmes2016], the ranking criterion was replaced by $A_j=max_{j}(H_{ij}C_{i})$ for the doubles contributions, which essentially allows for the pruning and ranking to occur simulatenously. With this modified ranking criterion, both the denominator and the phase information are ignored. This approximation is less justifiable for more difficult quantum chemistry problems. As shown in Table \[tab:comparisons\], this reduces HBCI’s ability to generate a compact wave function for C$_{2}$ and Cr$_{2}$. Additionally, as the basis set size is increased, the energy denominator can become large and many determinants can become unimportant due to the involvement of high energy orbitals. Thus, the denominator in Eq. \[eqn:rankeqn\] can become increasingly important for pruning Hilbert space. For ASCI, we find that the new search algorithms are sufficiently fast that we do not need to use the approximations in the HBCI formalism. Nevertheless, integral driven search algorithms are compatible with the sorting algorithms presented in this work. Modern Implementation {#sec:algtechs} ===================== In this section, we present the newest algorithms for ASCI. The selected CI algorithms presented here are extremely efficient on modern computing architectures, with added functionality that allows for scaling to large numbers of electrons and basis sets. This section is organized by first presenting several different algorithms for the critical ASCI functions of determinant search and Hamiltonian build, and then describing how the best of these algorithms can be efficiently implemented. The main parts of the algorithm that we describe here are the *Hamiltonian construction step*, presented in Section \[sec:hamcons\], and the *search algorithm*, presented in Section \[sec:searchmain\]. We also briefly discuss the *perturbation theory step*, but the optimal algorithmic implementations of this will be presented in a future work. Sorting as a paradigm for selected CI ------------------------------------- In this section we present evidence for the efficiency of using sorting based algorithms in selected CI on modern computers. We start by first considering both a generic ’search algorithm’ and the ’Epstein-Nesbet perturbation theory’ algorithm in the limit of unlimited memory. The following results were developed through extensive testing of various libraries such as the standard template library (STL) [@plauger2000] and Boost [@boost2011]. We specifically considered different solutions that can be developed through sorting, hash tables, and search trees. The idealized unlimited memory algorithm is as follows: Generate {$D^{sd}$} by going through each $i$ in {$D_{}$} and retain both $i$ and the connected bit string $j\in \{D^{sd}\}$ in an array. Sort the array based on $j$ and now the elements will be grouped together such that that calculation of Eq. \[eqn:rankeqn\] or Eq. \[eqn:en\] can be done with a single pass over the sorted list. We hypothesize that there is no faster way to perform this process than by using a sorting algorithm on modern computing architectures. However, very recent advances with hash tables suggest that such structures might eventually also become competitive, particularly for large basis sets [@hash_benchmark]. For realistic algorithms with limited memory, the sorting algorithm can be modified in different ways. We present the most efficient new ASCI search algorithm using this approach in section \[sec:searchmain\]. ### The world of sorting {#sec:wws} The dominant bottleneck in many selected CI algorithms is cache inefficiency, that is, an inability to get all of the necessary data to the CPU in a way that computation can occur in an efficient manner. Practical sorting algorithms are developed to be as cache efficient as possible and are thus a natural choice for ASCI. Research on sorting algorithms is an active field with new algorithms being developed to work with modern computing architectures [@radulsort; @pdq; @vergesort; @skasort; @musser1997; @timsort]. There have been many significant innovations even within the last few years [@2017arXiv170408579B; @blocksort]. Additionally, parallelization of sorting algorithms is quite different on GPUs versus CPUs, and efficient parallelization approaches are only starting to be developed [@ips4; @2015arXiv151103404B]. Since the selected CI algorithms presented in this work are based on sorting; any developments made in improving sorting algorithms also improves these algorithms. We emphasize that sorting based algorithms allow for easy access to parallelization which include GPU implementations. To demonstrate the efficiency of different sorting algorithms, we present a comparison of different methods in Table \[tab:speedsort\]. The STL implementation is a quicksort and PDQ is a pattern defeating quicksort. The Boost sorting algorithm is spreadsort. Spreadsort is a hybrid algorithm that uses a radix sort in most situations. The IPS$^4$O algorithm is a sorting algorithm designed to be run in parallel. For our small tests, we found the algorithm to be nearly linear scaling up to 8 cores. Our GPU tests were performed with the Thrust library [@thrust] and the timing results include the time is takes to move the data on and off the GPU. Algorithm Timing(s) --------------------------------- ----------- (STL) Quick Sort [@plauger2000] 30 PDQ Sort [@pdq] 29 (Boost) Spreadsort [@boost2011] 31 IPS$^4$O 1-core [@ips4] 37 IPS$^4$O 4-cores [@ips4] 9 IPS$^4$O 8-cores [@ips4] 5 (Thrust) GPU sort [@thrust] 2.3 : Comparison of different sorting techniques over 128 bit determinants from a Cr$_{2}$ SVP ASCI simulation. The test is for sorting 300 million integers. Most of the sorting algorithms presented here have similar performance. STL, Boost, and Thrust are popular libraries that include many application tools. The PDQ is a pattern defeating quick sort. The IPS4O and GPU results show that the sorting parts of the ASCI algorithm can be either parallelized or offloaded to a GPU for enhanced performance. The CPU simulations were performed on an Intel Xeon E5-2620 v5 processor of 2.10 GHz. The GPU calculation was performed on a NVIDIA Kepler K80.[]{data-label="tab:speedsort"} Hamiltonian construction {#sec:hamcons} ------------------------ Constructing the Hamiltonian requires some considerations beyond those required for traditional CI algorithms. At any given step of an iterative selected CI calculation, one has to determine which Hamiltonian matrix elements are non-zero [@szabo:book]. In a typical active space calculation, all determinants within the active space are present and it is trivial to find the non-zero matrix elements. For a selected CI simulation this is not the case. The tests we present in this section take a list of determinants as input, and output a unique set of matrix coordinates that are the non-zero matrix elements. These matrix elements must be ordered by row to be used in a sparse matrix diagonalization routine. There are two straightforward approaches typically used in Hamiltonian construction that are however computationally inefficient in the limit of large numbers of determinants. These algorithms are presented for completeness in algorithms \[alg:doubleloop\] and  \[alg:singledouble\]. The ’double loop’ algorithm is fast for wave functions in which few determinants are being considered, since the number of operations scales as $O(N_{tdets}^{2})$. The ’singles/doubles’ algorithm can be efficient when there are not many orbitals or electrons but a lot of determinants retained. In the limit of a large number of either determinants, electrons, or orbitals, these methods can become inefficient. Generate all determinant pairs in $\psi_k$ Determine if they are singly or double connected If yes, calculate the matrix element Generate $D^{sd}_i$, all singles and doubles from a determinant $D_i$ in current wave function $\psi_k$ For every determinant $j\in D_i^{sd}$, check if $j\in\{D\}$ For each match, calculate the matrix element For each determinant in $\psi_k$, create all bit strings/determinants in which two electrons are removed. Each one of these bit strings is called a residue. Store the following two things together: (residue, list of all determinants that generate the residue) in a tree object {$V$}. Each time a new residue is found, query the tree for its existence. If it is new, add it to the tree. If the residue is already there, add the generating determinant to the residue’s list. Once the tree is finished, go through every residue. All pairs of determinants in a residue list are doubly connected, and have a non-zero Hamiltonian matrix element. Go through the list of Hamiltonian matrix elements and remove any duplicates (Determinant pairs that are singly connected will appear in multiple residues with each other). ### Hamiltonian Construction: Residue Arrays An alternative approach to constructing the Hamiltonian uses a data structure called a residue tree. Residue trees are one of the fastest techniques currently in the literature for constructing Hamiltonians [@stampfub2005]. The residue tree is a simple data structure that makes it straightforward to find connections between determinants. The *residues* that can be generated from a reference determinant is a set of determinants that is generated by removing two electrons in all possible ways from the reference determinant. Each node of a residue tree contains a residue and a list of all the determinants in $\{D_{}\}$ that can generate the residue. The full residue tree consists of all the distinct residues that can be generated from $\{D_{}\}$. The residue tree is created as a tree so that any node in it can queried and found in $\mathcal{O}(log(N))$ time (where $N$ is the number of nodes in the tree). The number of possible residues is approximately $N_{tdets} \binom{N_{elec}}{2}$, where $N_{elec}$ is the number of electrons in the system. In the case of Cr$_{2}$ with 48 electrons and $N_{tdets}$=300,000, the number of residues is on the order of 200 million. For our testing, the residue trees are implemented as red-black trees and related search trees through STL and Boost libraries. After experimenting with residue arrays we realized that the idea of a residue can be made even more efficient with a sorting-based algorithm on such arrays rather than a tree based algorithm. Residue arrays are very similar to residue trees, but the tree structure is completely removed. Instead, each residue, together with the determinant that generated it, is stored in an unsorted array. After all residues have been generated, the array is sorted by residue. It is then in a form in which all non-zero matrix elements of the Hamiltonian can be generated, exactly as in the case of residue trees. Figure \[fig:speedhamil\] and Table \[tab:speed\] provide comparisons of residue arrays to other techniques for finding the non-zero Hamiltonian matrix elements. Figure \[fig:speedhamil\] shows that in our Cr$_{2}$ ASCI test, residue arrays are second only to dynamic bit masking (discussed in Section \[sec:dynbit\]). For each determinant in $\psi_k$, create all bit strings/determinants in which two electrons are removed. Each one of these bit strings is called a residue. Store the following two things together: (residue, determinant that generated residue) in an array {$V$}. Once all the residues have been generated, sort the array by residues. All residues that are equal will now be adjacent in the array. Pairs of determinants that generate a residue are doubly connected, and have a non-zero Hamiltonian matrix element. Store all connections and then remove any duplicates (Determinant pairs that are singly connected will appear in multiple residues with each other). ### Hamiltonian Construction: Dynamic Bit Masking {#sec:dynbit} Determine the number $n$ of bit masks to use, and set the bit masks. This is done by looking at the occupation of the orbitals given by $\psi_k$. The more orbitals that are close to being 50/50 occupied, the more efficient the dynamic bit masking becomes on increasing $n$. Apply each bit mask to each determinant in $\psi_k$, and save the masked value mask(j)& detstring(i) = mv(i,j): i.e., determinant $i$ with mask $j$ applied to it. Pick (n-4) of the bit masks and using the related mv(i,j), create a composite number out of them for each $i$. Find all pairs of determinants with the same composite number. These are non-zero matrix elements of the Hamiltonian (check for false positives). Repeat the last two steps for all $n \choose (n-4)$ possible ways of picking (n-4) bit masks. Go through the list of Hamiltonian matrix elements and remove duplicates. The dynamic bit masking algorithm is based on the property that any two bit strings that are at most doubly connected, differ by at most four orbitals. For a pair of bit strings that are doubly connected, there exists a set of four orbitals such that we can delete those orbitals and the two quadruply-deleted bit strings will become equivalent. Consider the example strings “111000011” and “111001100”. By removing the first four orbitals (right to left ordering), both strings become “11100”. Another way of doing this is simply to mask the bits (to zero), instead of deleting them. Either way, we will call the resulting bit string a ’reduced’ bit string. Thus by checking all possible ways of removing four orbitals, we can determine if two bit strings are singly or doubly connected. To create an algorithm that is efficient and works for large number of bit strings, we have to expand this example in two essential ways. To compare a large list of reduced bit strings, after deleting the four orbitals in all bit strings, the reduced strings are sorted (which requires $\mathcal{O}(n\log(n))$ time), and then a final pass on the reduced strings is performed to look for blocks of adjacent reduced strings that are equal. The final pass can be $\mathcal{O}(n^{2})$ in the limit of dense matrices. For sparse Hamiltonians simulated in this work, this step is essentially linear in the number of determinants. \# dets --------------- --------- ----- ----- ----- ----- ----------- ---- ---- ---- ----- ----------- ---- ----- ----- ------ NNZ DL RT RA DB NNZ DL RT RA DB NNZ DL RT RA DB Cr$_{2}$ ASCI 412336 0.1 4.6 1.7 0.3 7238098 10 59 19 4.2 24095545 75 196 55 14.5 Cr$_{2}$ CISD 5440608 0.1 5.5 3.1 1.4 105791967 10 84 43 29 469933565 75 261 144 147 The second key step is to reduce the number of quadruplets of orbitals to remove. In this current example, we search over all ways to remove four orbitals, of which there are $\binom{N_{orbs}}{4} $. Since this number can be quite large this approach would normally be a very slow way of constructing the Hamiltonian. To reduce the number of combinations of orbital removals, we use a coarse graining of the bit string by considering the bit string as non-overlapping substrings. If we break a pair of bit strings into some number $m$ of non-overlapping substrings (regardless of how this is done), at most four substrings can be different if the bit strings are singly or doubly connected. Hence all pairs of determinants, where the remaining $m-4$ substrings match, are candidates to have a non-zero matrix element between them. The trade off to this coarse graining is that there will be some false positives, which need to be checked for and removed. The way in which the bit strings are subdivided can be represented with bit masks. A simple but inefficient approach, as mentioned at the beginning of this section, is to have a bit mask for each orbital. However, it is easy to understand the approach in this limit. We provide a simple example for how bit masking can allow one to efficiently identify determinants that differ by up to two occupancies in figure \[fig:bit masktable\]. Coarse-graining is most effective when the the orbitals that are closest to being occupied 50% of the time are put on different masks. Thus, one can pre-calculate orbital occupation over all the determinants in $\psi_k$. The “dynamic” part of this bit masking approach is named for this step where we use the wave function information to generate the bit masks. Once the bit masks have been selected, the algorithm proceeds as follows. For all determinants in the determinant list, compute the bitwise AND of the bit string representations of the determinants with the $m$ bit masks. Then, for all possible ways of selecting $m-4$ bit masks, create a unique value that is generated by the $m-4$ masked values. This can be done in any number of ways. A simple way is to create a combined integer by using shifting and adding operations to move all the masked values into the memory of a large integer. Sort this list of $N_\text{dets}$ combined integers to find repeated numbers, which correspond to non-zero elements of the Hamiltonian (checking for any false positives). For the tests we used in this work, we generally used 10 bit masks, which correspond to $\binom{10}{6} = 210$ different bit mask subsets. This corresponds to the number of sorting steps needed in the algorithm, which is one of the main costs. The more bit masks used, the fewer false positives. Timings for a full Hamiltonian build, which includes the calculation of matrix elements, are presented in section \[sec:searchmain\]. ![Spin averaged orbital occupation plot (not weighted by determinant coefficient) for Cr$_2$ in the SVP basis. There are 10 critical spatial orbitals (20 spin orbitals) that are near 50% occupied, and are colored in green. The rest are mostly occupied (blue) or mostly empty (red). For our general Hamiltonian construction algorithms we fixed our calculations to 10 bit masks. Thus, for the Cr$_{2}$ SVP test, 2 critical orbitals are assigned to each bit mask. Conversely, none of the orbitals in the CISD wave function are near 50% occupation, and therefore the orbitals can be assigned without much concern to the different bit masks.[]{data-label="fig:dynbit"}](Cr2WF){width="1.0\columnwidth"} ### Hamiltonian Construction: Algorithm comparisons {#sec:hcresults} We summarize our main algorithmic results for the new diagonalization algorithms in figure \[fig:speedhamil\]. For these tests, we considered Hamiltonian construction for an all-electron Cr$_{2}$ SVP simulation as described in previous work [@tubman2016-1]. As noted above, residue arrays and dynamical bit masking are both based on sorting algorithms. We have also tested many other approaches that do not use sorting algorithms, such as the residue tree algorithm [@stampfub2005]. Search trees and hash tables, which are used in every current published implementation of selected CI that we are aware of, were found for the most part to be not competitive in any of our tests. However, recent advances with hash tables and their related benchmarks need to be investigated further [@hash_benchmark]. ### Orbital rotations and calculating the 2RDM The dynamical bit masking technique is useful not only for calculating the Hamiltonian, but also for generating the 1-RDM and 2-RDM. In the case of a traditional full CI calculation, the single particle orbitals do not affect the final result. In a SCI simulation, which includes ASCI and HBCI, the single particle orbitals can influence the convergence of a simulation with regards to $N_{tdets}$. This suggests that single particle orbitals can be optimized and used to improve the energy convergence. In our initial ASCI implementation, we ran an iterative loop around our ASCI simulations, in which we performed orbital rotations using the 1-RDM at each step. At the end of a given ASCI run, the 1-RDM from the final wave function would be used to rotate to the natural orbital basis and the integrals would be recalculated. We would then use the new integrals as input for the next ASCI simulation. For C$_{2}$ and Cr$_{2}$, we found this procedure converged quite rapidly and needed fewer than four iterations of this in all the simulations presented here. In our current implementation we perform orbital rotations after each iteration during the wave function growth phase (see figure \[fig:flowchart\]). After each diagonalization step in the growth phase, we rotate to the natural orbitals of the current wave function. We then immediately rediagonalize the Hamiltonian. This extra diagonalization step is important, since the expansion coefficients of the wave function are affected by the orbital rotation. After we finish the growth process, we fix the orbitals during the refinement steps. Search Algorithm {#sec:searchmain} ---------------- ### Search: new ASCI search {#sec:searchtech} The new ASCI search combines the pruning and ranking steps together such that the information generated in the pruning step is used to accelerate the ranking step. The new algorithm works as follows: generate all possible contributions of interest and store them in an array along with the generating determinant $i$. For an integral driven search, generate only those $\{D^{sd}\}$ such that $H_{ij} > \epsilon$. For a coefficient driven search, generate all $\{D^{sd}\}$ from the top $N_{cdets}$ determinants in $\{D_{}\}$. Either way, once the pruning is done, sort the array. With the sorted array, it is possible to calculate an approximate ranking of Eq. \[eqn:rankeqn\], $$A_{i} = \frac{\sum^{'}_{j \ne i}H_{ij}C^k_{j}}{(H_{ii}-E_k)}. \label{eqn:rankapprox}$$ The prime on the sum indicates we are only including connections that fit our pruning criteria. For practical cases in which memory is limited, one can sort a partial array (after the maximum number of elements allowed in memory is reached), combine elements with the same bit string representation (that is, carry out a partial sum of Eq. \[eqn:rankapprox\]), and retain some percentage of the largest terms before continuing with the search. The ranking approximation developed for the HBCI were used in Ref. [@holmes2016] to reduce the cost of the search in the original ASCI algorithm. With the improved algorithms described here, this approximation is unnecessary (see Tables \[tab:hbcomp\] and  \[tab:comparisons\]). The accuracy of this new ASCI search is close to that of the original ASCI search algorithm. In the HBCI ranking approximation, Eq. \[eqn:rankeqn\] was replaced with max($|C_{i}H_{ij}|$), which is too approximate and yields less accurate results. To further improve the computational efficiency of the new ASCI search approach, we speed up the calculation of the diagonal matrix elements, which are generally expensive to calculate and slow down the algorithm significantly (see Section \[sec:fdiag\]). The diagonal matrix elements are required for the denominator of Eq. \[eqn:rankeqn\]. The algorithm for calculating fast diagonal matrix elements is described in algorithm \[alg:diag\]. The new ASCI search algorithm (summarized in algorithm \[alg:ASCIsearchbetter\]), is recommended for all compatible selected CI algorithms. In Table  \[tab:hbcomp\] and  \[tab:comparisons\], we present a small number of comparisons of ASCI search to previously published HBCI and CIPSI results. In Table \[tab:alltime\], we demonstrate that the new search algorithm is not a bottleneck in terms of computational time and can generally be performed instead of less accurate searches. See also Table \[tab:speedsort\] for generic timings of sorting integers. ------------------- --------- ---------- --------- -------- --------------- System Basis H(Build) H(Diag) Search NNZ(millions) OH (9e,44o) cc-pVTZ 16 11 3 47 CH (7e,44o) cc-pVTZ 20 17 3 63 Li$_{2}$(6e,60o) cc-pVTZ 19 18 1 65 Li$_{2}$(6e,110o) cc-pVQZ 40 34 4 92 F$_{2}$(18e,28o) cc-pVDZ 6 2 1 14 F$_{2}$(18e,60o) cc-pVTZ 12 7 12 47 N$_{2}$(14e,28o) cc-pVDZ 10 10 1 25 N$_{2}$(18e,60o) cc-pVTZ 19 14 8 60 C$_{2}$(12e,28o) cc-pVDZ 8 6 1 21 C$_{2}$(12e,60o) cc-pVTZ 12 24 7 50 C$_{2}$(12e,110o) cc-pVQZ 26 32 48 68 ------------------- --------- ---------- --------- -------- --------------- : Timing of ASCI steps for various molecules. Timings presented for last step of growth algorithm in which 100,000 determinants are used. All timings are in seconds. NNZ is the number of non-zero matrix elements in the Hamiltonian. The column H(Build) is the timing for finding and calculating all non-zero matrix elements with dynamic bit masking. The column H(Diag) is the timing for matrix diagonalization with the Spectra package [@spectracode; @eigencode]. The search column is the timing for the ASCI search algorithm. These search timings do not represent the fastest set of parameters that make the search accurate and the timings would change slightly between a coefficient driven search versus an integral driven search. However it is evident that the timing for ASCI search is of the same order of magnitude as the other steps in ASCI and is not the bottleneck of the simulation. There is no reason to avoid using an accurate search algorithm such as the ASCI search. []{data-label="tab:alltime"} Input $\psi_k$ and an array {$V$} that will hold information in pairs of (determinant bit string, ranking value ($RV$)) Order the determinants of $\psi_k$ by coefficient magnitude Generate a set $D^{sd}$ from $\psi_k$ as follows: - Generate all single excitations $j$ in {$D^{sd}$} and calculate $C_{i}H_{ij}$ - Generate double connections $j$ from {$D^{sd}$} with either a determinant driven or coefficient driven approach and calculate $c_{i}H_{ij}$ - Add $j$ to an array {$V$} together with $RV$ =$c_{i}H_{ij}$ - Sort the array $\{V\}$ by bit string. For all repeated elements $j$, sum up $RV$ to calculate the numerator of Eq. \[eqn:rankeqn\] - In a situation when memory is limited, perform a partial sort (by $RV$) and retain the top elements in array {$V$}, erase the rest of the array elements and continue with the search Finish calculating Eq. \[eqn:rankeqn\] and gather all the top contributors for use in the next iteration Other algorithmic improvements {#sec:misc} ------------------------------ In this work thus far, we have presented new algorithms that are efficient for performing the search and Hamiltonian construction. Most implementations of selected CI suffer significant performance penalties as a result of the growth of the data structures involved. In particular, the algorithms described in this work generally require the movement and manipulation of bit strings, which grow in size as more orbitals are included in a simulation. By studying and developing the sorting approaches contained in this work, we have been able to understand how most selected CI methods can benefit from cache efficient techniques. Additionally there are many helper algorithms that can be used in combination with the sorting algorithms described above to improve the over all efficiency. ### Fast Diagonal Matrix Elements {#sec:fdiag} During the calculation of the denominator of Eq. \[eqn:rankeqn\], the diagonal matrix element of the connection being considered (H$_{ii}$) is required. Calculating the diagonal matrix elements is a relatively expensive step because these involve sums over the numbers of both electrons and pairs of electrons. However, because the determinant whose diagonal matrix element is being sought (H$_{ii}$) is always a single/double excitation away from a reference determinant, H$_{ref}$, for which the diagonal element is already known, the new matrix element H$_{ii}$ can be calculated quickly. This is to say that H$_{ref}$ - H$_{ii}$ only involves a small subset of terms, and can be calculated much faster than H$_{ii}$ from scratch. Algorithm \[alg:diag\] describes this protocol. Calculating the array of partial contributions described in algorithm \[alg:diag\] only has to be done once per reference determinant. This overhead turns out to be neglibile, as in either the PT2 or search algorithms, many connections from a reference determinant are considered all at once. Thus the initial overhead is small compared to the number of diagonal matrix elements that need to be calculated for a given reference determinant. (Precalculation step) Input Determinant $D_{i}$, the diagonal matrix element $H_{ii}$, and the one-electron integrals $h_{ii}$ (Precalculation step) Calculate the partial contribution: $p(i) = \sum^{occ}_{j} \langle ij||ij\rangle$ Input $D_k$ (connected to $D_i$) with set of orbitals excited into ($A$) and the set of orbitals excited out of ($R$) $E_{rem} = \sum_{i\in R}h_{ii}+p(i)$, $E_{add} = \sum_{i\in A}h_{ii}+p(i)$ $H_{kk}=H_{ii}-E_{rem}+E_{add}-\sum_{i\in R, j\in A}\langle ii||jj\rangle$ $H_{kk} = H_{kk} + \langle R_{1}R_{1}||R_{2}R_{2}\rangle-\langle A_{1}A_{1}||A_{2}A_{2}\rangle-\langle R_{1}R_{1}||A_{2}A_{2}\rangle-\langle R_{2}R_{2}||A_{1}A_{1}\rangle$ $H_{kk} = H_{kk} + \langle R_{\alpha}R_{\alpha}|R_{\beta}R_{\beta}\rangle+\langle A_{\alpha}A_{\alpha}|A_{\beta}A_{\beta}\rangle-\langle A_{\alpha}A_{\alpha}|R_{\beta}R_{\beta}\rangle-\langle R_{\alpha}R_{\alpha}|A_{\beta}A_{\beta}\rangle$ ### Bit string representation Larger basis sets become more costly for ASCI for several reasons. One of the biggest costs comes from the cost associated with manipulating the larger bit strings associated with the larger basis set. Some of this cost can be mitigated by using more compact bit string representations. We first review the standard bit string representation before introducing selected CI adapted representations. Standard bit string representations use one bit for each spin orbital. Often these bits will be divided into a spin up string $\alpha$ and a spin down string $\beta$. When orbital $i$ is occupied by an $\alpha$ or $\beta$ electron, its bit will be set to 1 in either the alpha or beta string, respectively. For simplicity and alignment purposes within modern computing architectures, the number of bits is rounded up to the nearest power of 2. As an example the Cr$_{2}$ SVP basis set has 42 orbitals [@yanai2009]. Thus a 64-bit integer would be used for the alpha and beta bit strings, combining to form a 128-bit string to represent the determinant. \[sec:ibit\] Comparisons and manipulations of larger integers take more time because larger blocks of memory must be compared and copied. Relatedly, parallel algorithms in which communication is the bottleneck also benefit from using smaller datatypes. Moreover, because large integers take up more space, fewer of them may be stored in the small cache available on the CPU. Additionally, most modern CPUs only implement 64-bit integer operations in hardware, so there is additional overhead required to perform bit manipulations in the multi-precision libraries that are implemented for handling integer types with greater than 64 bits. While the effect of the size of the representation of bit strings has not been explored in detail previously in the selected CI literature, this issue will occur in other selected CI approaches. Representation Size Description ---------------- ----------------------------------------------------- ------------------- Standard $2(N_{orbs})$ Regular Electron $N_{elec}*\log(N_{orbs})$ Electron occs Difference $N_{diff}\left(\log(N_{occ})+\log(N_{virt})\right)$ HF diff Hash 64 (32 or 128 possible) Use a 64 bit hash : A list of different bit string representations we considered in this work. See text for details.[]{data-label="tab:bitreps"} In this work, we consider several different representations as well as an approach using hash functions. The length of these alternative representations (see Table \[tab:bitreps\]) can be much shorter than the standard representation. In the standard representation, 2\*($N_{orbs}$) bits are used to represent all possible bit strings with $N_{orbs}$ spatial orbitals, regardless of how many quantum particles there are in the simulation. In the *electron representation*, rather than specifying whether orbitals are occupied or not, the occupied orbital indices are listed. That is, if orbitals 1, 2, and 10 are occupied, this determinant may be represented by concatenating 1, 2, and 10 in binary. The amount of space required for this is $N_{elec}\log{N_{orb}}$ because the space required to store the maximum orbital index in binary is $\log{N_{orb}}$ and this is required for each electron. We use the convention that $\alpha$ electrons are concatenated before $\beta$ electrons, and, moreover, that for each spin the electrons are labeled in order from smallest to largest. This thus provides a unique bit string for a given determinant. The number of $\alpha$ and $\beta$ electrons does not have to be specified in each bit string since this is fixed for a given problem. This representation is best for systems with a small number of electrons but a large number of orbitals. Another alternative is the *difference representation*. In this representation, the determinant is specified by a list of the orbitals excited from the Hartree-Fock determinant and the orbitals into which they are excited. That is, if the electrons in orbitals 1 and 3 have been excited to 8 and 9, then the excited determinant is stored by concatenating 1, 3, 8, 9, and 2 in binary, with the final 2 added to specify that the particular determinant has two excitations. This representation is most useful when a large majority of the determinants are only a few excitations away from the Hartree-Fock determinant. Since the space required to store the maximum orbital excited out of and in to are $\log{N_{elec}}$ and $\log{N_{virt}}$, respectively, the memory requirements for this representation is $N_{diff}(\log(N_{occ})+\log(N_{virt})$. The number of bits needed can be reduced by only working with groups of bit strings that have fixed excitation number at one time. In such a situation, no bits are needed to represent the number of beta excitations, since this would be set by the number of alpha excitations and the total excitation number. ### Generalized compression with hash functions {#sec:hbit} The alternate bit string representations described above will not always reduce to the size of the standard bit string representation. A more generic approach can be considered through hash functions. Hash functions are a ’many to one’ map that reduces large data sets into fixed length integers and are widely known for their use in cryptography. However hash functions also have wide usage for non-cryptographic purposes [@smhasher1; @smhasher2]. The compression aspect of a hash function as well as the general property that hash functions are fast to calculate, are what we aim to exploit for enhancing the efficiency of our sorting algorithms. To understand this, it is important to note that for large basis set calculations, we will often make use of bit strings larger than 64 bits in order to represent a determinant. For example there are over 10$^{36}$ unique numbers that can be represented with a 128 bit integer. Yet we will only ever consider a very sparse number of determinants in this space. Based on the timings in this paper, we might expect to simulate systems on the order of 10$^{13}$ determinants for a large scale PT2 calculation. Thus for purposes for sorting, the question arises whether it is possible to hash (compress) large bit strings, such as 128 and larger, to a smaller size, such as 64 bits, without generating too many collisions. Some collisions are unlikely to effect the overall result, and it is straightforward to quantify how any might be generated. We will present a full benchmarking of hashing for large basis set simulations in future work. ### Perturbation Theory One of the most exciting aspects of the ASCI method is the ability to converge the energy with post processing of the final ASCI wave function, allowing one to produce a wave function that is better than virtually all other approximate CI techniques. One of the main reasons that one might expect perturbation theory to be effective is that selected CI techniques finds all of the most important determinants in the Hilbert space and therefore anything that remains is necessarily small. There are multiple different ways in which the perturbation theory can be applied, including the Epstein-Nesbet perturbation theory that was introduced in Eq. \[eqn:en\]. Recent approaches have tested the use of Monte Carlo sampling of these equations  [@garniron2017; @sharma2017], and have been performed on roughly the same system sizes we presented here. Within the selected CI community, there have been only limited attempts to make fast and scalable deterministic perturbation theory algorithms, since the focus has been on stochastic approaches. However, we find that a deterministic approach will in many cases be faster than a stochastic approach in achieving chemical accuracy. Indeed, if high accuracy is desired, the cost to converge the stochastic error of sampling techniques will likely be higher than that of deterministic approaches. It should also be noted that many of the algorithmic improvements discussed in this work will also improve the stochastic approaches. In future studies we shall present the details of the deterministic PT algorithm we use in this work and a detailed discussion of where it is more efficient than stochastic methods. Results {#sec:results} ======= System E(10$^{5}$) E(3\*10$^{5}$) E(10$^{6}$) E(10$^{5}$)+PT2 E(3\*10$^{5}$)+PT2 E(10$^{6}$)+PT2 E(CCSD(T)) ASCI-CCSD(T)(mHa) ------------------ ------------- ---------------- ------------- ----------------- -------------------- ----------------- ------------- ------------------- BeH -15.189251 -15.189270 -15.189067 -0.202 C$_2$H$_2$ -77.111633 -77.113877 -77.114973 -77.116264 -77.116281 -77.116286 -77.114792 -1.493 C$_2$H$_4$ -78.347897 -78.352665 -78.356526 -78.360534 -78.360736 -78.360870 -78.359679 -1.19 C$_2$H$_6$ -79.562899 -79.569328 -79.575622 -79.586971 -79.587362 -79.587837 -79.587598 -0.238 CH -38.381756 -38.381830 -38.381247 -0.583 CH$_2$\_singlet -39.024713 -39.024890 -39.023981 -0.908 CH$_2$\_triplet -39.043515 -39.043647 -39.043196 -0.451 CH$_3$ -39.718247 -39.718698 -39.718161 -0.536 CH$_3$Cl -499.429423 -499.435363 -499.440527 -499.448171 -499.448369 -499.448604 -499.447848 -0.756 CH$_4$ -40.387284 -40.388797 -40.389683 -40.390483 -40.390461 -40.390449 -40.389881 -0.567 Cl$_2$ -919.261710 -919.267269 -919.271444 -919.275248 -919.275433 -919.275637 -919.274483 -1.154 ClF -559.193146 -559.197394 -559.200247 -559.203185 -559.203196 -559.203258 -559.201909 -1.349 ClO -534.574415 -534.578708 -534.581433 -534.583682 -534.583783 -534.583873 -534.582017 -1.856 CN -92.494095 -92.495493 -92.496013 -92.497001 -92.496990 -92.496993 -92.492776 -4.217 CO -113.055788 -113.058002 -113.05905 -113.060133 -113.060112 -113.060115 -113.058554 -1.561 CO$_2$ -188.124313 -188.133577 -188.142101 -188.155553 -188.155768 -188.155990 -188.154316 -1.673 CS -435.606137 -435.610029 -435.612284 -435.614274 -435.614332 -435.614381 -435.612596 -1.784 F$_2$ -199.096831 -199.099912 -199.101763 -199.103291 -199.103314 -199.103345 -199.101481 -1.863 H$_2$CO -114.213577 -114.217838 -114.221088 -114.224420 -114.224415 -114.224453 -114.222990 -1.462 H$_2$O -76.243432 -76.243908 -76.243266 -0.641 H$_2$O$_2$ -151.180192 -151.185981 -151.191415 -151.199403 -151.199388 -151.199442 -151.197870 -1.571 H$_2$S -398.871385 -398.871798 -398.871891 -398.872360 -398.872357 -398.872357 -398.871682 -0.675 H$_3$COH -115.403262 -115.409733 -115.415589 -115.425482 -115.425525 -115.425695 -115.424943 -0.752 H$_3$CSH -438.038163 -438.045231 -438.051720 -438.063596 -438.063899 -438.064300 -438.063831 -0.469 HCl -460.260248 -460.26072 -460.260217 -0.502 HCN -93.189422 -93.192245 -93.193804 -93.195281 -93.195268 -93.195273 -93.193547 -1.726 HCO -113.571362 -113.575611 -113.578885 -113.582082 -113.58199 -113.581967 -113.580181 -1.785 HF -100.229942 -100.230383 -100.229878 -0.504 HOCl -535.221695 -535.227183 -535.231958 -535.237723 -535.237791 -535.237901 -535.236561 -1.34 Li$_2$ -14.901321 -14.901337 -14.901331 -0.005 LiF -107.157186 -107.157874 -107.157255 -0.618 LiH -8.014688 -8.0147070 -8.014708 0.002 N$_2$ -109.279470 -109.280666 -109.280941 -109.281912 -109.281927 -109.281933 -109.279982 -1.95 N$_2$H$_4$ -111.547435 -111.554763 -111.561876 -111.575558 -111.575552 -111.575733 -111.575055 -0.678 Na$_2$ -323.733949 -323.733997 -323.734047 0.051 NaCl -621.595166 -621.595793 -621.595302 -0.49 NH -55.093412 -55.093530 -55.093131 -0.399 NH$_2$ -55.735042 -55.735304 -55.734737 -0.567 NH$_3$ -56.403846 -56.404593 -56.404853 -56.405301 -56.405298 -56.405299 -56.404668 -0.631 NO -129.597925 -129.600605 -129.601836 -129.603099 -129.60308 -129.603079 -129.60115 -1.926 O$_2$ -149.986002 -149.988413 -149.989404 -149.990669 -149.990697 -149.990712 –149.988242 -2.469 OH -75.561403 -75.561639 -75.561190 -0.449 P$_2$ -681.733926 -681.736715 -681.738451 -681.740076 -681.740139 -681.740138 -681.737699 -2.439 PH$_2$ -342.015274 -342.015912 -342.015208 -0.703 PH$_3$ -342.643041 -342.644264 -342.644881 -342.645539 -342.645537 -342.645538 -342.644777 -0.76 S$_2$ -795.334164 -795.338687 -795.341495 -795.344591 -795.344788 -795.344873 -795.343001 -1.872 Si$_2$ -577.937111 -577.938035 -577.938624 -577.940646 -577.940670 -577.940685 -577.938371 -2.314 Si$_2$H$_6$ -581.596630 -581.602802 -581.609336 -581.622553 -581.623534 -581.623600 -581.625045 1.445 SiH$_2$\_singlet -290.143803 -290.144185 -290.143496 -0.689 SiH$_2$\_triplet -290.101056 -290.101377 -290.100697 -0.68 SiH$_3$ -290.754016 -290.754683 -290.755020 -290.755380 -290.755385 -290.755393 -290.754754 -0.64 SiH$_4$ -291.396355 -291.398117 -291.399246 -291.400364 -291.400415 -291.400435 -291.399825 -0.61 SiO -364.086144 -364.088978 -364.090533 -364.092085 -364.092087 -364.092108 -364.090065 -2.042 SO -472.662216 -472.666953 -472.670542 -472.673837 -472.673857 -472.673893 -472.671759 -2.134 SO$_2$ -547.683822 -547.696538 -547.707992 -547.732183 -547.732764 -547.733404 -547.731453 -1.95 System E(10$^{5}$) E(3\*10$^{5}$) E(10$^{6}$) E(10$^{5}$)+PT2 E(3\*10$^{5}$)+PT2 E(10$^{6}$)+PT2 E(CCSD(T)) ASCI-CCSD(T)(mHa) ------------------ ------------- ---------------- ------------- ----------------- -------------------- ----------------- ------------- ------------------- BeH -15.202968 -15.203059 -15.202841 -0.218 C$_2$H$_2$ -77.203740 -77.209351 -77.213467 -77.219532 -77.219470 -77.219436 -77.218046 -1.39 C$_2$H$_4$ -78.441654 -78.450402 -78.457561 -78.471004 -78.471438 -78.470558 -0.88 C$_2$H$_6$ -79.654252 -79.669309 -79.680275 -79.706885 -79.707123 -79.707843 0.72 CH -38.422000 -38.422323 -38.421572 -0.751 CH$_2$\_singlet -39.074412 -39.074996 -39.075219 -39.075663 -39.075638 -39.075641 -39.074555 -1.086 CH$_2$\_triplet -39.091649 -39.092561 -39.091962 -0.599 CH$_3$ -39.774767 -39.775888 -39.776526 -39.777381 -39.777335 -39.777308 -39.776688 -0.62 CH$_3$Cl -499.558732 -499.574866 -499.586523 -499.613409 -499.613956 -499.614776 0.82 CH$_4$ -40.445943 -40.449959 -40.452376 -40.455999 -40.455844 -40.455727 -40.455065 -0.662 Cl$_2$ -919.443813 -919.456931 -919.467067 -919.489489 -919.490563 -919.491057 0.494 ClF -559.385351 -559.397209 -559.406478 -559.424939 -559.425259 -559.425479 -559.424306 -1.173 ClO -534.743221 -534.75429 -534.763177 -534.779342 -534.779876 -534.780297 -534.778399 -1.898 CN -92.579322 -92.585386 -92.589739 -92.595444 -92.595278 -92.595193 -92.5907969 -4.396 CO -113.164548 -113.170564 -113.174971 -113.181556 -113.18131 -113.181203 -113.179827 -1.376 CO$_2$ -188.30436 -188.320873 -188.334374 -188.366900 -188.367270 -188.367412 0.142 CS -435.711205 -435.720092 -435.727456 -435.739697 -435.740067 -435.740301 -435.738264 -2.037 F$_2$ -199.297893 -199.305614 -199.311568 -199.322174 -199.322044 -199.322013 -199.320490 -1.523 H$_2$CO -114.333293 -114.342815 -114.349969 -114.363196 -114.363265 -114.363298 -114.362161 -1.137 H$_2$O -76.342531 -76.344107 -76.344966 -76.346114 -76.346046 -76.346017 -76.345555 -0.462 H$_2$O$_2$ -151.328267 -151.346149 -151.358343 -151.38593 -151.385453 -151.383997 -1.456 H$_2$S -398.963339 -398.966601 -398.968840 -398.971854 -398.971758 -398.971672 -398.970663 -1.009 H$_3$COH -115.518981 -115.538789 -115.551947 -115.581707 -115.581297 -115.580557 -0.74 H$_3$CSH -438.147677 -438.169313 -438.184286 -438.219185 -438.219443 -438.220721 1.278 HCl -460.365313 -460.367443 -460.368720 -460.370542 -460.370467 -460.370428 -460.369640 -0.788 HCN -93.283038 -93.290436 -93.295931 -93.304741 -93.304506 -93.304428 -93.302782 -1.646 HCO -113.679998 -113.690685 -113.698435 -113.712574 -113.712376 -113.712248 -113.710567 -1.681 HF -100.349364 -100.351308 -100.351011 -0.297 HOCl -535.38483 -535.401621 -535.413545 -535.439971 -535.440168 -535.439573 -0.595 Li$_2$ -14.930785 -14.930786 -14.930734 -0.05 LiF -107.288594 -107.289889 -107.290477 -107.291433 -107.291471 -107.291499 -107.291267 -0.232 LiH -8.036373 -8.036477 -8.036468 -0.009 N$_2$ -109.386980 -109.391765 -109.395149 -109.399681 -109.399578 -109.399513 -109.397769 -1.744 N$_2$H$_4$ -111.653829 -111.677219 -111.693657 -111.730173 -111.728936 -111.727644 -1.292 Na$_2$ -323.768264 -323.769296 -323.769243 -0.053 NaCl -621.715212 -621.718356 -621.720573 -621.723618 -621.723678 -621.723662 -621.723066 -0.596 NH -55.152475 -55.153059 -55.152497 -0.562 NH$_2$ -55.805235 -55.806328 -55.806869 -55.807663 -55.807618 -55.807603 -55.806951 -0.652 NH$_3$ -56.482279 -56.484996 -56.486656 -56.489049 -56.488902 -56.488806 -56.488205 -0.601 NO -129.723293 -129.730435 -129.73569 -129.743992 -129.743756 -129.743632 -129.741868 -1.764 O$_2$ -150.128809 -150.137355 -150.144920 -150.154261 -150.154042 -150.154011 -150.151869 -2.142 OH -75.649012 -75.650479 -75.649945 -0.521 P$_2$ -681.858265 -681.867251 -681.875496 -681.891685 -681.892673 -681.893154 -681.890509 -2.645 PH$_2$ -342.093796 -342.096363 -342.098130 -342.100204 -342.100126 -342.100058 -342.099026 -1.032 PH$_3$ -342.724455 -342.729358 -342.732548 -342.737893 -342.737852 -342.737824 -342.736695 -1.129 S$_2$ -795.479446 -795.492316 -795.501795 -795.523482 -795.524441 -795.52348 -0.961 Si$_2$ -578.064458 -578.073053 -578.078038 -578.091580 -578.092361 -578.0927272 -578.090310 -2.417 Si$_2$H$_6$ -581.742988 -581.757872 -581.808668 -581.810282 -581.816405 6.123 SiH$_2$\_singlet -290.224640 -290.226524 -290.227658 -290.228906 -290.228859 -290.228837 -290.227976 -0.861 SiH$_2$\_triplet -290.179533 -290.181296 -290.182352 -290.183480 -290.183430 -290.183409 -290.182552 -0.857 SiH$_3$ -290.836906 -290.840366 -290.842778 -290.845976 -290.845983 -290.845917 -290.845112 -0.805 SiH$_4$ -291.482012 -291.487852 -291.492150 -291.498424 -291.498589 -291.498717 -291.498002 -0.715 SiO -364.239919 -364.247764 -364.254334 -364.265039 -364.264860 -364.264922 -364.263351 -1.571 SO -472.815828 -472.828155 -472.837710 -472.857033 -472.857111 -472.857246 -472.855188 -2.058 SO$_2$ -547.901866 -547.936060 -547.958730 -548.026183 -548.026880 -548.030305 3.425 Atoms E(cc-pVDZ) E(cc-pVTZ) ------- ------------ ------------ B -24.59062 -24.60580 Be -14.61740 -14.62379 C -37.76190 -37.79003 Cl -459.60432 -459.70385 F -99.52947 -99.63240 H -0.49927 -0.49980 Li -7.43263 -7.44606 N -54.48011 -54.52523 Na -161.85418 -161.86990 O -74.91171 -74.98526 P -340.79727 -340.86128 S -397.60643 -397.68672 Si -288.92066 -288.98829 : Atomic energies for cc-pVDZ and cc-pVTZ atoms in the G1 set. All energies are in units of Ha. All atomic energy calculations are ASCI+PT2 energies. These atomic energies are very easy to simulate with ASCI and all of the presented results are more than 0.1 mHa accurate. []{data-label="tab:atomene"} [|l|l|l|l|l||l|l|l|l|l|]{} System & -------------- ASCI D$_{e}$ cc-pVDZ -------------- & -------------------- CCSDTQ D$_{e}$ cc-pVDZ [@feller1] -------------------- & -------------- ASCI D$_{e}$ cc-pVTZ -------------- & ----------------------------- CCSD(T) D$_{e}$ CBS FC [@feller1; @feller2] ----------------------------- & System & -------------- ASCI D$_{e}$ cc-pVDZ -------------- & -------------------- CCSDTQ D$_{e}$ cc-pVDZ [@feller1] -------------------- & -------------- ASCI D$_{e}$ cc-pVTZ -------------- & ----------------------------- CCSD(T) D$_{e}$ CBS FC [@feller1; @feller2] ----------------------------- \ BeH & 45.546 & & 49.858 & 50.11 & HOCl & 139.668 & & 157.654 & 165.5\ C$_2$H$_2$ & 372.692 & 371.71 & 401.442 & 402.78 & Li$_2$ & 22.635 & & 24.245 & 24.49\ C$_2$H$_4$ & 527.077 & & 560.148 & 561.51 & LiF & 122.846 & & 133.676 & 137.7\ C$_2$H$_6$ & 670.406 & & 707.948 & 710.5 & LiH & 51.955 & & 56.852 & 58.11\ CH & 75.707 & & 83.128 & 83.89 & N$_2$ & 201.877 & 201.44 & 219.031 & 227.14\ CH$_2$\_singlet & 165.932 & 165.83 & 179.457 & 180.68 & N$_2$H$_4$ & 388.051 & & 426.224 & 436.87\ CH$_2$\_triplet & 177.702 & 177.6 & 190.074 & 189.85 & Na$_2$ & 16.081 & & 18.504 & 16.7\ CH$_3$ & 288.002 & 287.62 & 306.124 & 306.71 & NaCl & 86.144 & & 94.063 & 99.3\ CH$_3$Cl & 366.802 & & 389.450 & 394.83 & NH & 71.625 & 71.61 & 80.331 & 82.85\ CH$_4$ & 396.229 & & 418.203 & 419.14 & NH$_2$ & 161.042 & & 177.429 & 182\ Cl$_2$ & 42.029 & 42.13 & 51.987 & 59.87 & NH$_3$ & 268.168 & & 291.255 & 297.2\ ClF & 43.584 & & 55.984 & 62.61 & NO & 132.562 & 132.8 & 146.296 & 151.77\ ClO & 42.561 & & 57.215 & 64.56 & O$_2$ & 104.967 & 105.19 & 115.166 & 119.9\ CN & 160.001 & 160.3 & 175.653 & 179.2 & OH & 94.530 & 94.4 & 101.028 & 107.06\ CO & 242.529 & 242.04 & 254.708 & 258.59 & P$_2$ & 91.361 & 90.43 & 107.038 & 115.85\ CO$_2$ & 358.089 & & 380.715 & 388.12 & PH$_2$ & 138.104 & & 150.068 & 154.0\ CS & 154.395 & 154.1 & 165.372 & 170.93 & PH$_3$ & 219.898 & & 236.637 & 241.65\ F$_2$ & 27.860 & 28.19 & 35.896 & 38.43 & S$_2$ & 82.836 & 82.43 & 94.750 & 103.52\ H$_2$CO & 346.558 & 345.79 & 369.213 & 373.15 & Si$_2$ & 61.811 & 61.94 & 70.958 & 75.95\ H$_2$O & 209.359 & 209.16 & 226.616 & 232.67 & Si$_2$H$_6$ & 493.601 & & 523.189 & 535.0\ H$_2$O$_2$ & 236.853 & & 260.611 & 268.32 & SiH$_2$\_singlet & 141.169 & 140.64 & 151.183 & 153.9\ H$_2$S & 167.776 & & 179.047 & 183.51 & SiH$_2$\_triplet & 15.510 & & 122.676 & 133.5\ H$_3$COH & 473.745 & & 506.248 & 510.9 & SiH$_3$ & 211.405 & & 224.771 & 228.7\ H$_3$CSH & 438.535 & & 466.516 & 472.3 & SiH$_4$ & 302.872 & & 320.774 & 324.3\ HCl & 98.588 & 98.29 & 104.643 & 107.39 & SiO & 162.981 & & 182.836 & 191.77\ HCN & 284.877 & 283.38 & 307.070 & 311.37 & SO & 97.730 & 97.22 & 116.253 & 125.73\ HCO & 256.693 & & 274.308 & 277.1 & SO$_2$ & 190.471 & & 231.949 & 259.14\ HF & 126.525 & 126.52 & 137.468 & 141.59 & & & & &\ System E(10$^{6}$)+PT2 (cc-pVDZ) E(3\*10$^{5}$)+PT2 (cc-pVTZ) ------------ --------------------------- ------------------------------ C$_2$H$_4$ -78.35920 -78.47061 C$_2$H$_6$ -79.58738 -79.70721 CH$_3$Cl -499.44827 -499.61401 CH$_4$ -40.39016 -40.45587 Cl$_2$ -919.27460 -919.48997 ClF -559.20164 -559.42533 ClO -534.58136 -534.77956 CO -113.06008 -113.18187 CO$_2$ -188.15522 -188.36808 CS -435.61384 -435.73994 F$_2$ -199.10299 -199.32210 H$_2$CO -114.22396 -114.36354 H$_2$O$_2$ -151.19915 -151.38581 H$_2$S -398.87226 -398.97178 HCN -93.19409 -93.30518 HCO -113.58083 -113.71129 HOCl -535.23687 -535.44029 N$_2$ -109.28088 -109.40151 NH$_3$ -56.40492 -56.48895 NO -129.60369 -129.74398 P$_2$ -681.73884 -681.89253 PH$_3$ -342.64550 -342.73796 Si$_2$ -577.94012 -578.09017 SiO -364.09010 -364.26490 SO -472.67119 -472.85738 SO$_2$ -547.72710 -548.02811 : Ground state energies calculated with ASCI and PT2 corrections for selected molecules using geometries from Feller *et. al.* [@feller1]. Energies are presented in units of Ha. Most of the energies shown here are within 2 mHa of the energies calculated with the geometries from the original G1 set that are shown in Tables \[tab:dz\] and \[tab:tz\].[]{data-label="tab:feller"} We present several benchmark studies below in order to provide an overview of what can be done with selected CI today. The algorithms described above have been implemented in our own code [@tubman2016-1] and in QChem 5.1 [@shao_advances_2015]. With the exception of a few instances, the results in this section were generated on a single core with an upper limit of 50 CPU (single core) hours. Many of these calculations were made over the course of developing the methods in these papers. G1 dataset ---------- Tables \[tab:dz\] and \[tab:tz\] present ASCI results on the G1 test set of 55 molecules (all electron), a benchmark set of molecules that has been extensively studied with many different methods[@pople1989; @feller1; @feller2; @grossman2002; @ma2005; @mardirossian2017thirty; @hait2018accurate]. Here we present ASCI benchmark data with both cc-pVDZ and cc-pVTZ basis sets. The molecular geometries were taken from the original G1 set [@pople1989; @grossman2002; @ma2005], except for CN and CH$_{2}$ triplet where we use the geometries from Feller *et. al.* [@feller1]. ASCI simulations with different numbers of determinants were performed to demonstrate the convergence with respect to this parameter. Tables \[tab:dz\] and \[tab:tz\] show that nearly all of the molecules are converged to within chemical accuracy for the cc-pVDZ simulations and many are also converged in the cc-pVTZ basis set. These results suggest that cc-pVQZ convergence will be possible for all the G1 molecules in the near future, using only modest computational resources. We also make comparison of the ASCI results with CCSD(T)[@raghavachari1989fifth] and CCSDTQ. Before discussing the full benchmark set, we demonstrate in Figure \[fig:cn-ene\] the convergence of ASCI for the cyanide radical (CN) and compare this with convergence of a set of coupled cluster simulations. The coupled cluster results, for both CCSD and CCSD(T), were performed in QChem 5.1 [@shao_advances_2015]. As seen in this figure, the ASCI + PT2 results are more accurate than the comparable CCSD(T) results, even for calculations with only $10^{3}$ determinants for the variational wave function. Thus while simulations of fully converged ASCI calculations can take a few hours, small simulations that only take a few minutes (or less) already show similar or better accuracy than CCSD(T). In the rest of this section we discuss the energies from our simulations that can serve as benchmarks for future methodological developments. Tables \[tab:dz\] (cc-pVDZ) and Table \[tab:tz\] (cc-pVTZ) present the variational energies and the perturbation results from ASCI, as well as comparisons to CCSD(T). Table \[tab:atom\] presents atomization energies and makes comparisons of these to values obtained from CCSD(T) and CCSDTQ. The results for the cc-pVDZ basis set in Table \[tab:dz\] show apparent convergences of the energy to below 1 mHa accuracy for all molecules. It is possible to extrapolate such results when extra accuracy is needed [@holmes2016]. We note that when the perturbation correction is added to ASCI, neither this nor CCSD(T) is guaranteed to be variational. However, with the exception of Si$_2$H$_6$, all ASCI PT2 energies are below the CCSD(T) energies. We suggest that the difference between ASCI and CCSD(T) is a good estimate for the error in CCSD(T) energies relative to the full configuration interaction (FCI) result. In general these differences are less than 1 mHa. However Table \[tab:tz\] shows that for several molecules it can be higher, with the largest difference being for the CN radical, which has a difference of 4.2 mHa. For many of the molecules, either the Hilbert space is small or the difference between the variational and PT2 estimates is less than 1 mHa for small N$_{tdets}$, and in these cases we do not require testing of convergence with up to 1 million determinants. The results for the cc-pVTZ basis set (Table \[tab:tz\]) indicate that the ground state energies of many of the molecules are also converged in a larger basis set. For molecules not already converged with $N_{dets}$ = 100,000, we were able to run PT2 corrections with 100,000 and 300,000 determinants. The largest calculation made was for Si$_2$H$_6$ with 300,000 determinants, for which extra computing time was needed to converge the simulation. We note that in contrast to results from ASCI in comparing to the cc-pVDZ basis set energies, several CCSD(T) energies lie below the ASCI results. The full significance of this will have to be investigated further with larger determinant calculations in the future. For many systems, the CCSD(T) error is very similar across the basis sets and does not actually get worse when comparing cc-pVDZ to cc-pVTZ. Another way to understand the convergence errors in ASCI is to look at the size of the PT2 energy corrections. These are shown for the G1 dataset in figure \[fig:pt2\]. Within the G1 set, the SO$_{2}$ molecule has the largest PT2 correction. We note that while Si$_2$H$_6$ is one of the most time-consuming simulations here, it does not have the largest PT2 correction. Table \[tab:atomene\] presents atomic ground state energies for the atoms contained in the G1 set that are obtained from ASCI with cc-pVDZ and cc-pVTZ basis sets. These are combined with the molecular data in Tables \[tab:dz\] and \[tab:tz\] to calculate the atomization energies for the molecules in the G1 set, which are shown in Table \[tab:atom\] and compared there to other benchmark simulations. For the cc-pVDZ results (columns 2 and 7), we compare with results from CCSDTQ [@feller1]. These comparisons are necessarily indirect, since different geometries were used between the two sets of results, and the ASCI results are all electron. However the difference in geometries are not so large. Surveying the full set of molecules, we find less than 1 kcal differences between atomization energies. For the cc-pVTZ results (columns 4 and 9), we compare to benchmark complete basis set (CBS) frozen core results [@feller1; @feller2]. Again the comparison is not direct, but it is clear that the atomization energies are trending in the right direction in comparison to the CBS limit. To demonstrate the similarity of geometries between the different calculations, in Table \[tab:feller\] we present results of ASCI with perturbation corrections for selected molecules using the geometries from Feller *et. al.* [@feller1]. These can be compared with the corresponding energies from Table \[tab:dz\] and \[tab:tz\] to estimate the energy difference between the two sets of geometries and confirm the similarity. It is also interesting to compare with recent auxiliary field quantum Monte Carlo results (AFQMC) that are now able to calculate the G1 set [@borda2018]. AFQMC is accurate on many of these systems, but loses accuracy in certain situations such as for ionic systems like LiF and NaCl. This is likely due to deficiencies in the current trial wave functions used in AFQMC. See Ref. [@borda2018] for more details. Further comparisons between these methods will be interesting in the future. We also note that recent work on DMRG has also added perturbative corrections to that methodology [@guo2018-1; @guo2018-2], although it is not clear whether the largest simulations required for the G1 set are feasible with DMRG, even with perturbation theory improvements. Entanglement in chemical systems can be quite large even in systems in which the many body interactions are not necessarily strong, and is thus a problem for efficient simulation with DMRG. Conclusions =========== In this work we have presented a number of new algorithms and techniques for performing selected CI simulations and integrated them into an updated ASCI algorithm. These techniques include new ASCI search, dynamic bit masking, fast orbital rotations, and fast diagonal matrix elements. We also presented other techniques, such as residue arrays and alternative bit string representations, which will be important for simulating certain types of Hamiltonians. Most of the new techniques presented here take advantage of modern sorting techniques that have been designed to be efficient on current computing architectures. Sorting based algorithms provide new avenues for using GPUs and parallel sorting on CPUs. Sorting and hashing techniques can be further combined together, to allow calculations with very large basis sets. Our results indicate that the resulting algorithms are faster than other previously published selected CI approaches. To make comparison with stochastic approaches to CI such as HBCI, we have presented initial benchmarking calculations for C$_{2}$ and F$_{2}$ in Table \[tab:hbcomp\] that show the new ASCI approach is about an order of magnitude faster than HBCI for target accuracies within 0.1 mHa and more than two orders of magnitude faster for target accuracies within 0.01 mHa. With the ideas presented in this work, we have demonstrated the importance of using modern algorithms and understanding computing architectures for design of efficient selected CI algorithms. The speed and accuracy of these new techniques over a wide range of systems were demonstrated with a full set of ground state calculations for the G1 dataset of small molecules. We presented comparisons to coupled cluster (CCSD(T) and CCSDTQ) and benchmarked the accuracy of this compared to ASCI. This demonstration shows the various convergence properties and the accuracy that can be expected from the latest ASCI approach today. Given the results presented here, it is evident that we are now at a new beginning for modern selected CI simulations and that many hitherto intractable new applications are now within reach. Acknowledgements ================ This work was supported through the Scientific Discovery through Advanced Computing (SciDAC) program funded by the U.S. Department of Energy, Office of Science, Advanced Scientific Computing Research and Basic Energy Sciences. We used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by the National Science Foundation Grant No. OCI-1053575. CDF was supported by the NSF Graduate Research Fellowship under Grant DGE-1106400 and by the DOE Office of Science Graduate Student Research (SCGSR) program under contract number DESC0014664. DH was supported by a Berkeley Fellowship. DL and MHG acknowledge support from the Director, Office of Science, Office of Basic Energy Sciences, of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231. Supplemental Information for the Chromium Dimer =============================================== In this section we present an extrapolation for the ground state energy of the Cr$_{2}$ in the SVP basis (24e,30o) and make a comparison to previously published DMRG results. The energies and extrapolations are presented in Figure \[fig:cr2ext\] and show excellent agreement between the methods. Hartree-Fock energies for G1 set molecules ========================================== The energies of the restricted Hartree-Fock determinants used to initialize the ASCI calculations are given in Tables \[tab:hf-all\] and \[tab:hf-feller\]. Stability analysis was performed to ensure that all SCF solutions were local minimas. System cc-pVDZ cc-pVTZ ------------------ -------------- -------------- BeH -15.14945189 -15.15205222 C$_2$H$_2$ -76.82472747 -76.84763525 C$_2$H$_4$ -78.03990264 -78.06354855 C$_2$H$_{6}$ -79.23494277 -79.25973681 CH -38.26877842 -38.27690073 CH$_2$\_singlet -38.88108554 -38.89231068 CH$_2$\_triplet -38.92139886 -38.93217379 CH$_3$ -39.55963482 -39.57296784 CH$_3$Cl -499.1177293 -499.1476064 CH$_4$ -40.19870854 -40.21331465 CN -92.19603079 -92.21741416 CO -112.7461016 -112.7766305 CO$_2$ -187.6463113 -187.7018166 CS -435.3293719 -435.3521612 Cl$_2$ -918.9609562 -918.9983715 ClF -558.8436442 -558.8989343 ClO -534.2516809 -534.2964982 F$_2$ -198.6847963 -198.7508413 H$_2$CO -113.8746242 -113.9098452 H$_2$O -76.02602772 -76.05613647 H$_2$O$_2$ -150.7817569 -150.8331176 H$_2$S -398.6946587 -398.7129979 H$_3$COH -115.0486003 -115.0884306 H$_3$CSH -437.7255248 -437.756402 HCN -92.87969951 -92.90351433 HCO -113.2518818 -113.2848402 HCl -460.0894453 -460.1067487 HF -100.0184682 -100.0569205 HOCl -534.8720321 -534.9177714 Li$_{2}$ -14.87000212 -14.87168961 LiF -106.9451499 -106.9801182 LiH -7.98363507 -7.986532056 N$_2$ -108.9466732 -108.9743976 N$_2$H$_4$ -111.1858923 -111.2234062 NH -54.95953403 -54.97346363 NH$_2$ -55.56273484 -55.58092764 NH$_3$ -56.19548576 -56.21749393 NO -129.254714 -129.2901963 Na$_{2}$ -323.7047434 -323.7151846 NaCl -621.4337887 -621.4537044 O$_2$ -149.6014308 -149.6460477 OH -75.3896954 -75.41403106 P$_2$ -681.4629985 -681.4863607 PH$_2$ -341.8675755 -341.8815522 PH$_3$ -342.4706082 -342.4876397 S$_2$ -795.0490094 -795.0805171 SO -472.330077 -472.3804324 SO$_2$ -547.1725083 -547.2750501 Si$_2$ -577.756499 -577.7739636 Si$_2$H$_{6}$ -581.338866 -581.3706463 SiH$_2$\_singlet -290.0184384 -290.0303487 SiH$_2$\_triplet -289.9952407 -290.0072341 SiH$_3$ -290.6235912 -290.6386838 SiH$_4$ -291.2428929 -291.2605406 SiO -363.7883502 -363.8356554 : Restricted Hartree-Fock energies for molecules in the G1 set (RHF for closed shell species and ROHF for open shell). Energies are presented in units of Ha. The geometries are the same as those employed in Tables \[tab:dz\] and \[tab:tz\] (i.e. are taken from the original G1 set [@pople1989] except for CN and CH$_{2}$ triplet, for which we use the geometries from Ref. [@feller1].)[]{data-label="tab:hf-all"} System cc-pVDZ cc-pVTZ ------------ -------------- -------------- C$_2$H$_4$ -78.03884396 -78.0628233 C$_2$H$_6$ -79.23488258 -79.25993592 CH$_3$Cl -499.1177629 -499.1478118 CH$_4$ -40.19865157 -40.21342364 Cl$_2$ -918.9608245 -918.9987691 ClF -558.8442574 -558.9009543 ClO -534.2495837 -534.295923 CO -112.7480967 -112.7789016 CO$_2$ -187.6511077 -187.7072565 CS -435.3296827 -435.3526846 F$_{2}$ -198.6856445 -198.7520074 H$_2$CO -113.8763719 -113.9119201 H$_2$O$_2$ -150.78415 -150.8360866 H$_2$S -398.694525 -398.7129408 HCN -92.88324631 -92.90805477 HCO -113.2530034 -113.28603 N$_2$ -108.9541417 -108.9834898 NH$_3$ -56.19563482 -56.21791293 NO -129.2536412 -129.2889892 P$_2$ -681.465778 -681.4896859 PH$_3$ -342.4703141 -342.4873996 Si$_2$ -577.7566459 -577.7741915 SiO -363.7899104 -363.8389006 SO -472.3332171 -472.386195 SO$_2$ -547.1791953 -547.2876407 : Restricted Hartree-Fock energies (RHF for closed shell species and ROHF for open shell) for selected molecules using geometries from Feller *et. al.* [@feller1].[]{data-label="tab:hf-feller"} {width="10cm" height="4.5cm"} [87]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, (), http://dx.doi.org/10.1063/1.4955109](\doibase http://dx.doi.org/10.1063/1.4955109) [****,  ()](\doibase 10.1021/acs.jctc.6b00407),  [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/144/16/10.1063/1.4948308) [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/140/12/10.1063/1.4869192) [****,  ()](http://dx.doi.org/10.1021/acs.jctc.5b01099) [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/58/12/10.1063/1.1679199) [****,  ()](\doibase http://dx.doi.org/10.1063/1.460537) @noop [ ()]{},  [****,  ()](http://dx.doi.org/10.1139/cjc-2013-0017) @noop [****,  ()]{} [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/136/24/10.1063/1.4720076) [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/138/2/10.1063/1.4773819) @noop [ ()]{},  [****, ()](http://dx.doi.org/10.1038/nature11770) [****, ()](\doibase 10.1021/ct300486d) [****,  ()](\doibase 10.1103/PhysRevLett.114.033001) [****,  ()](\doibase 10.1021/ct300504f) @noop [****,  ()]{} [****,  ()](\doibase 10.1063/1.1449459) [****,  ()](\doibase 10.1063/1.478295) [****,  ()](\doibase 10.1103/PhysRevB.48.10345) [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/142/3/10.1063/1.4905329) [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/127/13/10.1063/1.2768362) [****, ()](http://dx.doi.org/10.1021/ct3008974) [****,  ()](http://dx.doi.org/10.1038/nchem.2041$\backslash$n10.1038/nchem.2041$\backslash$nhttp://www.nature.com/nchem/journal/v6/n10/abs/nchem.2041.html{#}supplementary-information) @noop [ ()]{},  @noop [****, ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [ ()]{},  @noop [****,  ()]{} [****,  ()](\doibase 10.1103/RevModPhys.77.259) [****,  ()](\doibase 10.1007/BF02394557) [****, ()](\doibase http://dx.doi.org/10.1016/0301-0104(83)85011-3) [****,  ()](\doibase http://dx.doi.org/10.1063/1.461037) [****,  (), http://dx.doi.org/10.1063/1.4905528](\doibase http://dx.doi.org/10.1063/1.4905528) [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/135/1/10.1063/1.3600351) [****,  ()](\doibase 10.1063/1.4996044),  [****, ()](\doibase 10.1103/PhysRevLett.90.136401) [****,  ()](\doibase 10.1063/1.4906829),  [****,  ()](\doibase 10.1103/PhysRevB.94.235144) @noop [ ()]{},  [****, ()](\doibase 10.1021/acs.jctc.6b01028), ,  [“,” ](https://github.com/LCPQ/quantum_package) () [****,  ()](\doibase 10.1103/PhysRevB.83.212401) [****,  ()](\doibase 10.1063/1.4989858), @noop [**]{} (, ) [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/125/12/10.1063/1.2335446) in [**](\doibase 10.1109/SC.2005.17) () pp.  (, ) pp.  [****,  ()](\doibase http://dx.doi.org/10.1063/1.1804498) [****,  ()](\doibase 10.1021/cr200137a) [****,  ()](\doibase 10.1103/PhysRev.183.23) [****,  ()](\doibase 10.1080/00268977800100581) [****,  ()](\doibase 10.1103/PhysRevC.79.064324) [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/122/2/10.1063/1.1829045) @noop [**]{},  ed. (, , ) @noop [**]{} (, ) [“,” ](https://tessil.github.io/2016/08/29/benchmark-hopscotch-map.html) () @noop [ ()]{},  [“,” ](https://github.com/orlp/pdqsort) () [“,” ](https://github.com/Morwenn/vergesort) () [“,” ](https://github.com/skarupke/ska_sort) () @noop [ ()]{} [“,” ](https://en.wikipedia.org/wiki/Timsort) () @noop [ ()]{},  @noop [ ()]{},  @noop [ ()]{},  @noop [ ()]{},  in @noop [**]{} (, ) pp.  [“,” ](https://spectralib.org/) () [“,” ](http://eigen.tuxfamily.org) () [****,  ()](http://scitation.aip.org/content/aip/journal/jcp/130/23/10.1063/1.3152576) [“,” ](https://github.com/aappleby/smhasher) () [“,” ](https://github.com/rurban/smhasher) () [****,  ()](\doibase 10.1063/1.4992127),  [****,  ()](\doibase 10.1063/1.456415), [****,  ()](\doibase 10.1063/1.3008061),  [****,  ()](\doibase 10.1063/1.478747),  [****,  ()](\doibase 10.1080/00268976.2014.952696) [****,  ()](\doibase 10.1063/1.1487829),  [****,  ()](\doibase 10.1063/1.1940588),  @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [ ()]{},  @noop [ ()]{},

ArXiv

--- abstract: 'K. Nikodem and the present author proved in [@NikWas95] a theorem concerning separation by affine functions. Our purpose is to generalize that result for polynomials. As a consequence we obtain two theorems connected with separation of $n$-convex function from $n$-concave function by a polynomial of degree at most $n$ and a stability result of Hyers-Ulam type for polynomials.' author: - bibliography: - 'was\_own.bib' - 'was\_pub.bib' date: January 1996 title: Polynomial selections and separation by polynomials --- Introduction ============ By $\R$, $\N$ we denote the set of all reals and positive integers, respectively. Let $I\subset\R$ be an interval. In this paper we present a necessary and sufficient condition under which two functions $f,g:I\to\R$ can be separated by a polynomial of degree at most $n$, where $n\in\N$ is a fixed number. Our main result is a generalization of the theorem concerning separation by affine functions obtained recently by K. Nikodem and the present author in [@NikWas95]. To get it we use Behrends and Nikodem’s abstract selection theorem (cf. [@BerNik95 Theorem 1]). It is a variation of Helly’s theorem (cf. [@Val64 Theorem 6.1]). By $\ccR$ we denote the family of all non-empty compact real intervals. Recall that if $F:I\to \ccR$ is a set-valued function then a function $f:I\to\R$ is called a *selection* of $F$ iff $f(x)\in F(x)$ for every $x\in I$. Behrends and Nikodem’s theorem states that if $\mathscr{W}$ is an $n$-dimensional space of functions mapping $I$ into $\R$ then a set-valued function $F:I\to \ccR$ has a selection belonging to $\mathscr{W}$ if and only if for every $n+1$ points $x_{1},\dots,x_{n+1}\in I$ there exists a function $f\in\mathscr{W}$ such that $f(x_{i})\in F(x_{i})$ for $i=1,\dots,n+1$. Let us start with the notation used in this paper. Let $n\in\N$. If $x_{1},\dots,x_{n}\in I$ are different points then for $i=1,\dots,n$ we define $$c_{i}(x;x_{1},\dots,x_{n})= \prod_{\substack{j=1\\ j\ne i}}^{n} \frac{x-x_{j}}{x_{i}-x_{j}}.$$ Note that $c_{i}(x_{j};x_{1},\dots,x_{n})$ is equal to 0 if $i\ne j$ and to 1 if $i=j$, $i,j=1,\dots,n$. ${\mathscr{P}}_{n}$ stands for the family of all polynomials of degree at most $n$. If $x_{1},\dots, x_{n+1}\in I$ are different points then the Lagrange interpolating polynomial going through the points $(x_{i},y_{i})$, $i=1,\dots,n+1$, is given by the following formula: $$\label{Lagrange} w(x)=\sum_{i=1}^{n+1} c_{i}(x;x_{1},\dots,x_{n+1})y_{i}.$$ This uniquely determined polynomial is a member of ${\mathscr{P}}_{n}$. Moreover, if $x<x_{1}<\dots<x_{n+1}$ then $c_{i}(x;x_{1},\dots,x_{n+1})$ is positive if $i$ is odd and negative if $i$ is even. Polynomial selections of set-valued functions ============================================= Now we shall prove a selection theorem which will be used to obtain our main result. If $n\in\N$ and $A_{i}\subset\R$, $i=1,\dots,n$, then $$\sum_{i=1}^{n} A_{i}$$ denotes the algebraic sum of the sets $A_{i}$, $i=1,\dots,n$. \[tw1\] Let $n\in\N$. The set-valued function $F:I\to \ccR$ has a selection belonging to ${\mathscr{P}}_{n}$ if and only if for every $x_{0},x_{1},\dots, x_{n+1}\in I$ such that $x_{0}<x_{1}<\dots<x_{n+1}$ the following condition holds: $$\label{selekcja} F(x_{0})\cap \biggl(\sum_{i=1}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1}) F(x_{i})\biggr)\ne\emptyset.$$ If $F$ has a selection belonging to ${\mathscr{W}}_{n}$ then is obvious. We prove that implies an existence of a polynomial selection of $F$. First we note that ${\mathscr{P}}_{n}$ is an $(n+1)$-dimensional space of functions. If we prove that for every $n+2$ points $x_{0},x_{1},\dots,x_{n+1}\in I$ there exists a $w\in {\mathscr{P}}_{n}$ such that $w(x_{i})\in F(x_{i})$, $i=0,1,\dots,n+1$, then by Behrends and Nikodem’s theorem $F$ will have a desired selection. (For another Helly-type theorem which may be used here cf. also [@Val64 Theorem 6.9].) Fix any different points $x_{0},x_{1},\dots,x_{n+1}\in I$ and assume $x_{0}< x_{1}<\dots<x_{n+1}$. Let $L_{i}=c_{i}(x_{0};x_{1},\dots,x_{n+1})$, $i=1,\dots,n+1$. Thus has the form $$\label{sel1} F(x_{0})\cap\biggl(\sum_{i=1}^{n+1} L_{i}F(x_{i})\biggr)\ne\emptyset.$$ It is easy to observe that $L_{i}$ is positive if $i$ is odd and negative if $i$ is even. Put $$y_{0}=\inf F(x_{0}),\;\;\;z_{0}=\sup F(x_{0})$$ and for $i=1,\dots,n+1$, $$y_{i}=\begin{cases} \inf F(x_{i})&\text{if }L_{i}>0,\\ \sup F(x_{i})&\text{if }L_{i}<0, \end{cases}\qquad z_{i}=\begin{cases} \sup F(x_{i})&\text{if }L_{i}>0,\\ \inf F(x_{i})&\text{if }L_{i}<0. \end{cases}$$ Therefore $F(x_{0})=\left[ y_{0},z_{0}\right]$ and for $i=1,\dots,n+1$, $$F(x_{i})=\begin{cases} [y_{i},z_{i}]&\text{if }L_{i}>0,\\ [z_{i},y_{i}]&\text{if }L_{i}<0. \end{cases}$$ Since $-\left[\alpha,\beta\right] =\left[-\beta,-\alpha\right]$ for all $\alpha,\beta\in\R$, we have $L_{i}F(x_{i})=\left[ L_{i}y_{i}, L_{i}z_{i} \right]$, $i=1,\dots,n+1$. If $u=L_{1}y_{1}+\dots+L_{n+1}y_{n+1}$ and $v=L_{1}z_{1}+\dots+L_{n+1}z_{n+1}$ then $u\le v$. Furthermore, $$\begin{gathered} \sum_{i=1}^{n+1} L_{i}F(x_{i}) =[L_{1}y_{1}, L_{1}z_{1}]+\dots+[L_{n+1}y_{n+1},L_{n+1}z_{n+1}]\\ =[ L_{1}y_{1}+\dots+L_{n+1}y_{n+1},L_{1}z_{1}+\dots+L_{n+1}z_{n+1}]=[u,v]\end{gathered}$$ and by we get $$\label{sel2} \left[ y_{0},z_{0}\right]\cap\left[ u,v\right]\ne\emptyset.$$ There are three cases of the location of the above intervals: (a) $u\in[y_{0},z_{0}]$, (b) $v\in[y_{0},z_{0}]$, (c) $[y_{0},z_{0}]\subset[u,v]$. Fix $t\in\left[ 0,1\right]$ and consider the polynomial $\varphi _{t}\in{\mathscr{P}}_{n}$ going through $n+1$ different points: $$\bigl(x_{0},tu+(1-t)v\bigr)\text{\;\;and\;\;} \bigl(x_{i},ty_{i}+(1-t)z_{i}\bigr)\text{\;\;for\;\;}i=1,\dots,n-1,n+1.$$ We shall show later that $$\label{fi_t} \varphi _{t}(x_{n})=ty_{n}+(1-t)z_{n}.$$ Hence, in the case (a) for $w=\varphi _{1}$ we have $$\begin{aligned} w(x_{0})&=u\in\left[ y_{0},z_{0}\right] =F(x_{0}),\\ w(x_{i})&=y_{i}\in F(x_{i}),\;\;i=1,\dots,n-1,n,n+1\end{aligned}$$ and similarly in the case (b) for $w=\varphi _{0}$. In the case (c) $y_{0}=\lambda u+(1-\lambda)v$ for some $\lambda\in\left[ 0,1\right]$. For $w=\varphi _{\lambda}$ we obtain $$\begin{aligned} w(x_{0})&=y_{0}\in F(x_{0}),\\ w(x_{i})&=\lambda y_{i}+(1-\lambda)z_{i}\in F(x_{i}),\;\;i=1,\dots,n-1,n,n+1.\end{aligned}$$ So in all cases there exists a $w\in {\mathscr{P}}_{n}$ such that $w(x_{i})\in F(x_{i})$, $i=0,\dots,n+1$, which will complete the proof if we show that holds true. By we get $$\begin{aligned} \varphi_{t}(x)&=c_{0}(x;x_{0},x_{1},\dots,x_{n-1},x_{n+1})\bigl(tu+(1-t)v\bigr)\\ &+\sum_{i=1}^{n-1}c_{i}(x;x_{0},x_{1},\dots,x_{n-1},x_{n+1})\bigl(ty_{i}+(1-t)z_{i}\bigr)\\ &+c_{n+1}(x;x_{0},x_{1},\dots,x_{n-1},x_{n+1})\bigl(ty_{n+1}+(1-t)z_{n+1}\bigr).\end{aligned}$$ If $M_{i}=c_{i}(x_{n};x_{0},x_{1},\dots,x_{n-1},x_{n+1})$, $i=0,1,\dots, n-1,n+1$, then after a bit of computation $$\varphi _{t}(x_{n})=\sum_{\stackrel{\scriptstyle i=1}{i\ne n}}^{n+1} (M_{0}L_{i}+M_{i})\bigl(ty_{i}+(1-t)z_{i}\bigr)+M_{0}L_{n}\bigl(ty_{n}+(1-t)z_{n}\bigr).$$ One can verify (using the product formula given in Introduction) that $M_{0}L_{n}=1$ and $M_{0}L_{i}+M_{i}=0$ for $i=1,\dots,n-1,n+1$. Hence holds and this finishes the proof. As a consequence of Theorem \[tw1\] we obtain \[wn1\][[@Was95 Theorem 1]]{} A set-valued function $F:I\to \ccR$ has an affine selection iff for every $x,y\in I$, $t\in \left[ 0,1\right]$ $$F\bigl(tx+(1-t)y\bigr)\cap\bigl(tF(x)+(1-t)F(y)\bigr)\ne\emptyset.$$ The above condition is equivalent to for $n=1$, $x<y$, $x_{0}=x$, $x_{2}=y$, $x_{1}=tx_{0}+(1-t)x_{2}$, where $t=\frac{x_{1}-x_{2}}{x_{0}-x_{2}}$. Separation by polynomials ========================= The main result of this paper reads as follows \[tw2\] Let $n\in\N$, $f,g:I\to\R$. The following conditions are equivalent: 1. \[tw2:i\] there exists $w\in {\mathscr{P}}_{n}$ such that $f(x)\le w(x)\le g(x)$, $x\in I$; 2. \[tw2:ii\] $f(b)\le g(b)$, where $b\in I$ is the right side endpoint of $I$ (if exists) and for every $x_{0},x_{1},\dots,x_{n+1}\in I$ such that $x_{0}\le x_{1}<\dots<x_{n+1}$ $$\label{oddziel} \begin{aligned} f(x_{0})&\le \sum_{\substack{i=1\\ i\text{\ \normalfont{odd}}}}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})g(x_{i})+ \sum_{\substack{i=1\\ i\text{\ \normalfont{even}}}}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})f(x_{i}),\\ g(x_{0})&\ge \sum_{\substack{i=1\\ i\text{\ \normalfont{odd}}}}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})f(x_{i})+ \sum_{\substack{i=1\\ i\text{\ \normalfont{even}}}}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})g(x_{i}). \end{aligned}$$ To prove that implies fix any $x_{0},x_{1},\dots,x_{n+1}\in I$ such that $x_{0}\le x_{1}<\dots<x_{n+1}$. Since the polynomial $w$ goes through the points $(x_{i},w(x_{i}))$, $i=1,\dots,n+1$, we have $$w(x_{0})=\sum_{i=1}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})w(x_{i}).$$ Then the ineualities are obvious. To prove the converse implication first note that replacing $x_{0}$ by $x_{1}$ in we have $f(x_{1})\le g(x_{1})$ in both ineualities, i.e. (ii) yields $f\le g$ on $I$. Let $$F(x)=\left[ f(x),g(x)\right],\;\;\;x\in I.$$ We shall show that $F:I\to \ccR$ fulfils . Fix any $x_{0}, x_{1},\dots,x_{n+1}\in I$ such that $x_{0}<x_{1}<\dots<x_{n+1}$. Let $u,v$ be equal to the right hand sides of the upper and lower ineualities , respectively. Therefore $v\le u$ and $$\label{oddz1} \left[ f(x_{0}),g(x_{0})\right]\cap\left[ v,u\right]\ne\emptyset$$ (otherwise $g(x_{0})<v$ or $u<f(x_{0})$ – contradiction with ). Let $L_{i}=c_{i}(x_{0};x_{1},\dots,x_{n+1})$, $i=1,\dots,n+1$. Then $$L_{i}F(x_{i})=\begin{cases} [L_{i}f(x_{i}),L_{i}g(x_{i})]&\text{if $i$ is odd},\\ [L_{i}g(x_{i}),L_{i}f(x_{i})]&\text{if $i$ is even} \end{cases}$$ and $$\left[ v,u\right] =\sum_{i=1}^{n+1} L_{i}F(x_{i}).$$ Thus implies . By Theorem \[tw1\] $F$ has a selection $w\in {\mathscr{P}}_{n}$. This finishes the proof. Ineualities in Theorem \[tw2\] do not guarantee $f(b)\le g(b)$, where $b\in I$ is the right side endpoint of $I$ (if exists). The following two functions $$f(x)=\begin{cases} \frac{1}{2}x&\text{for }0\le x<1,\\ 1&\text{for }x=1 \end{cases}\qquad\text{and}\qquad g(x)=\begin{cases} x&\text{for }0\le x<1,\\ \frac{1}{2}&\text{for }x=1 \end{cases}$$ fulfil for $n=1$ but $f(1)>g(1)$. Of course, $f$ and $g$ can not be separated by a straight line. As a consequence of Theorem \[tw2\] we obtain [@NikWas95 Theorem 1] Let $f,g:I\to\R$. The following conditions are equivalent: 1. there exists an affine function $h:I\to\R$ such that $f(x)\le h(x)\le g(x)$, $x\in I$; 2. for every $x,y\in I$, $t\in\left[ 0,1\right]$ $$\begin{aligned} f\bigl(tx+(1-t)y\bigr)&\le tg(x)+(1-t)g(y)\\ \intertext{and} g\bigl(tx+(1-t)y\bigr)&\ge tf(x)+(1-t)f(y). \end{aligned}$$ The above ineualities are equivalent to in Theorem \[tw2\] (cf. the proof of Corollary \[wn1\]). Applications ============ One can verify that $f:I\to\R$ is convex iff for every $x_{0},x_{1},x_{2}\in I$ such that $x_{0}<x_{1}<x_{2}$ $$f(x_{0})\ge c_{1}(x_{0};x_{1},x_{2})f(x_{1})+ c_{2}(x_{0};x_{1},x_{2})f(x_{2}).$$ We adopt the following definition (cf. [@RobVar73 §83], [@Cie59], [@Pop44], [@Pop34]). Let $n\in\N$. The function $f:I\to\R$ is $n$-*convex* iff for every $x_{0},x_{1},\dots,x_{n+1}\in I$ such that $x_{0}<x_{1}<\dots<x_{n+1}$ $$(-1)^{n}f(x_{0})\le (-1)^{n}\sum_{i=1}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})f(x_{i}).$$ $f$ is $n$-*concave* iff $(-f)$ is $n$-convex. If $f$ is both $n$-convex and $n$-concave then $f$ is a polynomial belonging to ${\mathscr{P}}_{n}$ (going through the points $(x_{i},f(x_{i}))$, $i=0,1,\dots,n+1$). \[wn3\] Let $n\in\N$. If $f:I\to\R$ is $n$-convex, $g:I\to\R$ is $n$-concave and $f(x)\le g(x)$, $x\in I$, then there exists a polynomial $w\in {\mathscr{P}} _{n}$ such that $f(x)\le w(x)\le g(x)$, $x\in I$. Fix any $x_{0},x_{1},\dots,x_{n+1}\in I$ such that $x_{0}\le x_{1}<\dots<x_{n+1}$. If $n$ is even then by $n$-convexity of $f$ $$\begin{aligned} f(x_{0})&\le\sum_{i=1}^{n+1}c_{i}(x_{0};x_{1},\dots,x_{n+1})f(x_{i})\\ &\le\sum_{\substack{i=1\\ i\text{\ \normalfont{odd}}}}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})g(x_{i}) +\sum_{\substack{i=1\\ i\text{\ \normalfont{even}}}}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})f(x_{i}).\end{aligned}$$ If $n$ is odd then by $n$-concavity of $g$ $$\begin{aligned} f(x_{0})&\le g(x_{0})\le\sum_{i=1}^{n+1}c_{i}(x_{0};x_{1},\dots,x_{n+1})g(x_{i})\\ &\le\sum_{\substack{i=1\\ i\text{\ \normalfont{odd}}}}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})g(x_{i}) +\sum_{\substack{i=1\\ i\text{\ \normalfont{even}}}}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1})f(x_{i}).\end{aligned}$$ The proof of the second ineuality in is analogous. Theorem \[tw2\] completes the proof. In the same way we get \[wn4\] Let $n\in\N$. If $f:I\to\R$ is $n$-concave, $g:I\to\R$ is $n$-convex and $f(x)\le g(x)$, $x\in I$, then there exists a polynomial $w\in {\mathscr{P}} _{n}$ such that $f(x)\le w(x)\le g(x)$, $x\in I$. For $n=1$ the above two results are well known and Corollaries \[wn3\] and \[wn4\] generalize them. Finally we prove a stability result for polynomials (cf. a Hyers-Ulam stability theorem for affine functions in [@NikWas95]). First observe that if $n\in\N$ and $w(x)=1$, $x\in I$, then $w\in {\mathscr{P}}_{n}$ and for every different points $x_{1},\dots,x_{n+1}\in I$ has the form $$\sum_{i=1}^{n+1} c_{i}(x;x_{1},\dots,x_{n+1})=1,\;\;x\in I.$$ Let $n\in\N$, $\varepsilon >0$ and $f:I\to\R$. If for every $x_{0},x_{1},\dots,$\ $x_{n+1}\in I$ such that $x_{0}\le x_{1}<\dots<x_{n+1}$ $$\label{modul1} \biggl| f(x_{0})-\sum_{i=1}^{n+1} c_{i}(x_{0};x_{1},\dots,x_{n+1}) f(x_{i})\biggr| \le\varepsilon$$ then there exists a polynomial $w\in {\mathscr{P}}_{n}$ such that $$\label{modul2} \left| f(x)-w(x)\right| \le\frac{\varepsilon}{2},\;\;x\in I.$$ If $f$ satisfies then $(ii)$ in Theorem \[tw2\] holds for $g(x)=f(x)+\varepsilon$, $x\in I$. So there exists a polynomial $\varphi\in {\mathscr{P}}_{n}$ such that $f(x)\le\varphi (x)\le f(x)+\varepsilon$, $x\in I$. For $$w(x)=\varphi (x)-\frac{\varepsilon}{2},\;\;x\in I$$ we obtain .

ArXiv

--- abstract: 'In this paper I introduce a mechanism to derive program transformations from order-preserving transformations of vector types. The purpose of this work is to allow automatic generation of correct-by-construction instances of programs in a streaming data processing paradigm suitable for FPGA processing. We show that for it is possible to automatically derive instances for programs based on combinations of opaque element-processing functions combined using *foldl* and *map*, purely from the type transformations.' author: - Wim Vanderbauwhede bibliography: - 'infering\_programs\_from\_type\_transformations.bib' title: Inferring Program Transformations from Type Transformations for Partitioning of Ordered Sets --- Introduction ============ In this discussion paper I want to introduce a set of *type transformations* on *vector types*. In this work we will use a simple form of dependent types [@bove2009dependent], but the concept can be generalised to transformations on other types, including session types [@honda2008multiparty]. The overall idea is to introduce the transformations and then explore the effect of transforming the types in a program on the program itself, i.e. what are the required corresponding functions that will transform the types of the computations while preserving the results of the computations. The purpose of this work is to allow automatic generation of correct-by-construction instances of programs in a streaming data processing paradigm suitable for data processing using FPGAs (Field Programmable Gate Arrays, [@vanderbauwhede2013high]). Using an optimisation technique such as simulated annealing [@aarts1988simulated] and a cost model for the FPGA implementation, the best instance can be automatically selected. Preliminaries ============= Type Variables -------------- [*a*]{} : is a type variable representing a nullary type constructor. We will call this kind of type variable *atomic*. [*b*,*c*]{} : are general type variables. We call the set of these type variables $\mathcal{B}$ [*k*,*l,m*,*n*]{} : are non-zero natural numbers, i.e. $k,m,n\in\mathbb{N}_{>0}$. We will call these *sizes*. [*p*,*q*]{} : are (unary) type constructor variables, i.e. types that apply to other types and can depend on non-zero natural numbers, e.g. *p n a* or *p k q q b*. Note that we use unary (i.e. one type and one size), right-associative type constructors purely to simplify the discussion, not as a fundamental limitation. The crucial point is however that these are *dependent types*. [*F*,*G*,*H*]{} : are general functions operating on types (we could call them type transformers). We assume the functions take a single type as argument and are right associative. Consequently, we can write *G (F a)* as *G F a* [*S*,*M*,*R* and *I*]{} : are specific functions operating on types, to de defined later. Notations and Definitions ------------------------- [*total size*]{} : The *total size* of a type is the product of all sizes: $\mathcal{N}(p_{1}\, n_{1}\, p_{2}\, n_{2}...p_{i}...p_{k}\, n_{k}\; a)\overset{{\scriptscriptstyle \triangle}}{=}\prod_{i=1}^{k}n_{i}$ [*type transformation*]{} : A *type transformation* is the application of a function from one type to another to a type, i.e. $F\, b\,=\, c$ [*i*,*j*]{} : are used as subscripts to distinguish between type variables of the same class, so that we can write $p_{1}\; p_{2}...p_{i}...p_{k}\; a$ Restrictions on Type Transformations\[Restrictions-on-Type\] ============================================================ Given a type expression of the form $p_{1}\, n_{1}\, p_{2}\, n_{2}...p_{i}...p_{k}\, n_{k}\; a$, the type transformations we want to consider must obey following restrictions: 1. The transformations do not remove any atomic type variables or introduce fresh atomic type variables. It follows that atomic type variables cannot be modified either. 2. The transformations can only remove or add one or more outer type constructors. 3. The type transformations can transform the *sizes*, but the total size of the type is an invariant. Although the types and transformations are more general, our focus is on transformations of types describing ordered sets. The above restrictions intend to reflect that the type transformations should not alter the number or nature of the elements of the set, but only the way the set is partitioned. Vector Types ============ For general types $p_{i}$ it may be hard to prove that the above rules do not alter the number or nature of the elements of the set, but only the way the set is partitioned. However, if we assume a single type representing a vector, then what these restrictions say is that a vector can only be reshaped but not modified in terms of its type or size. This is of course not a sufficient condition to guarantee that the type transformations will not change the computations, but it is a necessary one. We introduce the *vector type* $\upsilon\, k\, b$ where *k* is a non-zero positive integer and *b* is an arbitrary type. This the type representing a vector of length *k* containing values of type *b*. Specifically we define $$b\overset{{\scriptscriptstyle \triangle}}{=}\upsilon\,0\, b$$ Given an atomic type *a*, we can generate the set of all vector types $V(a)$ for $a$ as follows: $$\begin{cases} a\in V(a)\\ \forall\, b\,\in\, V(a),\forall\, k\, in\,\mathbb{N}_{>0}|\upsilon\, k\, b\in V(a) \end{cases}$$ For convenience, we introduce the following notation: *$\upsilon\, k\, b\overset{not.}{=}[b]\langle k\rangle$* And the shorthand: *$[...[[b]\langle k_{1}\rangle]\langle k_{2}\rangle...]\langle k_{n}\rangle\overset{not.}{=}[b]\langle k_{1}\rangle\langle k_{2}\rangle...\langle k_{n}\rangle$* Transformations on Vector Types ================================ For the rest of the paper we consider a specific case of types and type transformations: transformations on *vector types*. I posit three fundamental transformations, each with a corresponding inverse: - converting a type to a singleton vector type - applying a type transformation to the type variable of a vector type (*mapping*) - reshaping a vector type, i.e. modifying the sizes in a vector type such that the total size is remains invariant We can formalise each of these transformations: Singleton vector type ---------------------- The purpose of this operation is to change the dimensionality of a vector. $S\, b\,\overset{{\scriptscriptstyle \triangle}}{=}\,[b]\langle1\rangle$ The inverse operation (reducing dimensionality) is defined trivially as $S^{-1}\,[b]\langle1\rangle\,\overset{{\scriptscriptstyle \triangle}}{=}\, b$ So that $S^{-1}S\, b\,=\, b$ $S\, S^{-1}\,[b]\langle1\rangle\,=\,[b]\langle1\rangle$ Repeated application of S leads to higher-dimensional singleton vectors: $S\, S\, b=[[b]\langle1\rangle]\langle1\rangle$ We introduce a convenient notation $S^{k}\, b=[b]\langle1\rangle^{k}$ Mapping ------- Mapping applies a transformation to the type argument of a vector type. This operation is independent of the size, so I have omitted it: $M\, F\,[b]\,\overset{{\scriptscriptstyle \triangle}}{=}\,[F\, b]$ Note that the inverse operation is the application of the inverse of *F*, not of *M*: $M\, F^{-1}\,[F\, b]\,=\,[b]$ Although of course we can define purely notationally $M^{-1}\, F\overset{{\scriptscriptstyle \triangle}}{=}M\, F^{-1}$ Repeated application of *M* has two cases. The first case is applying different transformations to a 1-D vector $(M\, F)\,(M\, G)\,[b]=[F\, G\, b]$ We can rewrite the lhs as $(M\, F)\,(M\, G)\,[b]=M\,(F\, G)\,[b]$ The second case is applying a single transformation to a multi-dimensional vector $M\,(M\, F)\,[[b]]=[M\, F\,[b]]=[[F\, b]]$ We can rewrite the lhs as $M\,(M\, F)\,[[b]]=M^{2}\, F\,[[b]]$ Reshaping --------- The purpose of this operation is to re-partition a vector. The operation works on a 2-D vector. $R\, m\,[[b]\langle n_{1}\rangle]\langle n_{2}\rangle\overset{{\scriptscriptstyle \triangle}}{=}[[b]\langle n_{1}.m\rangle]\langle n_{2}/m\rangle$ The condition on *m* is of course that *n/m* is a natural number, i.e. $n_{2}$ is a multiple of *m*. The inverse operation can again be defined notationally: $R\, m^{-1}\,[[b]\langle n_{1}\rangle]\langle n_{2}\rangle=[[b]\langle n_{1}/m\rangle]\langle n_{2}.m\rangle$ and $R^{-1}\, m\overset{{\scriptscriptstyle \triangle}}{=}R\, m^{-1}$ Identity Operation ------------------ We define $I\, b\overset{{\scriptscriptstyle \triangle}}{=}b$ for completeness. I contend (but have not formally proven) that the set of operations *S*,*M*,*R*,*I* form a group over V(a). Each of the operations is associative and can be inverted, and any combination of operations on a vector type results in a vector type, i.e. it is closed as well. By adding *I*, the conditions for a group are satisfied. In fact, as we shall show below, *S*,*M*,*R*,*I* form a group over a particular finite subset of *V(a)*: - define *V(a,n)* as the subset of *V(a)* where, for any given vector $[a]\langle k_{1}\rangle\langle k_{2}\rangle...\langle k_{m}\rangle$, $\prod_{i=1}^{m}k_{i}=n$. - then *S,M,R,I* form a group over $V(a,n),\forall a,\, n$ In the next section, we will give a proof of the closure constraint. Operations on Atomic Types -------------------------- We define mapping on or reshaping of an atomic type as identity operations: $$M\, F\, a\overset{{\scriptscriptstyle \triangle}}{=}a$$ $$R\, k\, a\overset{{\scriptscriptstyle \triangle}}{=}a$$ Vector Creation --------------- For what follows, we will need an invertible operation $V$ to create vector types: $V\, k\, b\overset{{\scriptscriptstyle \triangle}}{=}[b]\langle k\rangle$ with its inverse $V^{-1}\, k\,[b]\langle k\rangle\overset{{\scriptscriptstyle \triangle}}{=}b$ In other words, V is equivalent to the vector type constructor but has an inverse. We use this operation to formally extract the argument of a vector type from the constructor. The dependent variable k is not strictly speaking necessary: $V^{-1}\, k\,(\upsilon\, k\, b)=b$ However, for any vector type $c$, $V^{-1}\, k\, c$ will result in a type error unless $c=\upsilon\, k\, b$. Theorem: *V(a,n)* is closed and complete under *S,M,R* ------------------------------------------------------ Any type transformation on any vector type in *V(a,N)* that observes the rules from Section \[Restrictions-on-Type\]: 1. can be expressed as a combination of the operations $S$, $M$ and $R$, and 2. results in a vector type in *V(a,N).*   - The most general expression for a type in our system is and multidimensional vector of type *a*, where the size in every dimension is different. We consider two instances of this type: $$\begin{aligned} T_{1=}[a]\langle n_{1}\rangle\langle n_{2}\rangle...\langle n_{i}\rangle...\langle n_{k}\rangle\\ T_{2}=[a]\langle m_{1}\rangle\langle m_{2}\rangle...\langle m{}_{j}\rangle...\langle m_{l}\rangle\end{aligned}$$ where $$\prod_{i=1}^{k}n_{i}=\prod_{j=1}^{l}m_{j}=N$$ and in general $k\neq l$ and the various $n_{i}$ and $m_{j}$ values can be different. - We aim to show that $T_{1}$ can be transformed into $T_{2}$ through application of a combination of the operations *S*, *M* and *R*. Our approach is to first reduce $T_{1}$ to a one-dimensional vector of size N, and then transform this vector into $T_{2}$. - First we reduce the expression for $T_{1}$ to $[a]\langle N\rangle$as follows 1. First reshape the outer two vectors through application of *R*: $$\begin{aligned} & & R^{-1}\, m_{k-1}\; T_{1}\\ = & & R^{-1}\, m_{k-1}\;[a]\langle n_{1}\rangle\langle n_{2}\rangle...\langle n_{i}\rangle...\langle n_{k-1}\rangle\langle n_{k}\rangle\\ = & & [a]\langle n_{1}\rangle\langle n_{2}\rangle...\langle n_{i}\rangle...\langle1\rangle\langle n_{k-1}.n_{k}\rangle\end{aligned}$$ which can also be written as $$[[[a]\langle n_{1}\rangle\langle n_{2}\rangle...\langle n_{i}\rangle...\langle n_{k-2}\rangle]\langle1\rangle]\langle n_{k-1}.n_{k}\rangle$$ 2. Then apply $S^{-1}$ to the type of the outer vector: $$\begin{aligned} & & M\, S^{-1}\,[[[a]\langle n_{1}\rangle\langle n_{2}\rangle...\langle n_{i}\rangle...\langle n_{k-2}\rangle]\langle1\rangle]\langle n_{k-1}.n_{k}\rangle\\ = & & [S^{-1}\,[[a]\langle n_{1}\rangle\langle n_{2}\rangle...\langle n_{i}\rangle...\langle n_{k-2}\rangle]\langle1\rangle]\langle n_{k-1}.n_{k}\rangle\\ = & & [[a]\langle n_{1}\rangle\langle n_{2}\rangle...\langle n_{i}\rangle...\langle n_{k-2}\rangle]\langle n_{k-1}.n_{k}\rangle\end{aligned}$$ 3. Repeating these steps results in $$[a]\langle n_{1}.n_{2}.\ldots.n_{i}.\,...\,.n_{k}\rangle$$ which can be written as $$[a]\langle N\rangle$$ - Then we perform the reverse process to obtain $T_{2}$: 1. First increase the dimensionality by calling *S* on the type of the outer vector $$M\, S\,[a]\langle N\rangle=[[a]\langle1\rangle]\langle N\rangle$$ 2. Then reshape the outer two vectors using *R* $$R\, m_{1}\,[[a]\langle1\rangle]\langle N\rangle=[[a]\langle m_{1}\rangle]\langle N/m_{1}\rangle$$ 3. Repeating these steps results in $$[a]\langle m_{1}\rangle\langle m_{2}\rangle...\langle m{}_{j}\rangle...\langle m_{l}\rangle=T_{2}$$ An type transformation consisting of a combination of the transformations *S*,*M* and *R* is reversible. Program Transformations ======================= In this section we want to explore how transforming a top-level type impacts on the program. The context is the FPGA architecture developed for the TyTra project[^1], which is similar to the MORA architecture [@chalamalasetti2009mora], and we consider a simple pipeline of computations. Like MORA, and indeed most FPGA architectures, TyTra assumes that data is streamed, and we model this as a map over a vector. I will use the notation $t_{s}$ or $t(s)$ to mean “the type of s”, and denote type-transformed functions and variables using a prime, e.g. *s’*. I will use the notation *{b}* to indicate “*b* may have to be transformed” during the inference process. General Assumption on the Program --------------------------------- The general assumption is that the program is built entirely out of - a set of functions on atomic types $f_{j}\,::\, a_{i}\rightarrow a_{k}$ - the *fold* operation [@hutton1999tutorial] - the *cons* (*:*) operation - tuple construction *(,)* However, we immediately note that *foldl* can be defined in terms of *fold* and the identity function *id*: $$foldl\, f\, v\, xs=foldr\,(\lambda x\, g\rightarrow(\lambda a\rightarrow g\,(f\, a\, x)))\, id\, xs\, v$$ and *map* in terms of *fold* and *cons*: $$map\, f=foldr\,(\lambda v\, x\rightarrow(x:v))$$ and the same goes for all basic list operations, so we take those as given. Essentially, our purpose is to split the program in computational functions and functions which describe the communication. Based on the type transformations, we aim to derive the transformation of those higher-order functions. We start with a few exploratory examples using *map* and *foldl*. Increasing dimensionality – map ------------------------------- We assume a very simple program, we use Haskell syntax [@hudak1992report] augmented with the `\left\langle N\right\rangle ` notation to indicate sizes. $s\,::\,[a]\left\langle N\right\rangle $ $g\,::\,[a]\left\langle N\right\rangle \rightarrow[a]\left\langle N\right\rangle $ $r\,::\,[b]\left\langle N\right\rangle $ $f\,::\, a\rightarrow a$ $r\,=\, g\, s$ $g\,=\, map\, f$ For completeness: $map\,::\,(t_{1}\rightarrow t_{2})\rightarrow[t_{1}]\left\langle n\right\rangle \rightarrow[t_{2}]\left\langle n\right\rangle $ We transform the top-level type: $t_{s}'$*$=R\, k\, M\, S\,$*$t_{s}$ *$=R\, k\, M\, S\,[a]\left\langle N\right\rangle $$=[[a]\left\langle k\right\rangle ]\left\langle N/k\right\rangle $* As *g* is applied to *s*, this leads to a transformation of *g*: $t_{g'}$= $t_{s'}\rightarrow t?$ We assume that we only explicitly transform each of the arguments of *g*. Then we get: $g'\,::\, t_{s}'\rightarrow[b]\left\langle N\right\rangle $ $g'\,=\, map\, f'$ We can substitute $t_{1}$ by the actual type of *s’* in the *map* inside *g’*: $r'\,=\, g'\, s'\,=\, map_{g'}\, f'\, s'$ $map_{g'}\,::\,\left(V^{-1}\,(N/k)\, t_{s'}\right)\rightarrow\left\{ b\right\} \rightarrow t_{s'}\rightarrow\left\{ [b]\left\langle N\right\rangle \right\} $ Clearly, this type can’t work for *map* because the return type *$[b]\left\langle N\right\rangle $* has a different size from $t_{s'}$. So we need to transform that type: $t_{2'}=R\, k\, M\, S\,[b]\left\langle N\right\rangle =[b\left\langle k\right\rangle ]\left\langle N/k\right\rangle $ This means that the signature for the map in *g’* becomes $map_{g'}\,::\,\left(V^{-1}(N/k)\, t_{s'}\rightarrow V^{-1}(N/k)\, t_{2'}\right)\rightarrow t_{s'}\rightarrow t_{2'}$ So that $g'\,::\, t_{s'}\rightarrow t_{2'}$ Consequently $t_{r'}=t(g'\, s')=t_{2'}$ $\rightarrow$ Rewriting the above in a more systematic way: g’ s’ = map f’ s’ map$_{g'}$$\,::\,(t_{1}$ $\rightarrow$ $t_{2}$$)$ $\rightarrow$ $[t_{1}]$ $\rightarrow$ $[t_{2}]$ map$_{g'}$ :: (*{a}* -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}[\[]{}a[\]]{}&lt;k&gt;[\]]{}&lt;N/k&gt; -&gt; *{[\[]{}b&lt;N&gt;}* map$_{g'}$ :: (*{*$V^{-1}\, N/k$ *[\[]{}[\[]{}a[\]]{}&lt;k&gt;[\]]{}&lt;N/k&gt;}* -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}[\[]{}a[\]]{}&lt;k&gt;[\]]{}&lt;N/k&gt; -&gt; *{[\[]{}b&lt;N&gt;}* map$_{g'}$ :: ([\[]{}a[\]]{}&lt;k&gt; -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}[\[]{}a[\]]{}&lt;k&gt;[\]]{}&lt;N/k&gt; -&gt; *{[\[]{}b[\]]{}&lt;N&gt;}* map$_{g'}$ :: ([\[]{}a[\]]{}&lt;k&gt; -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; $R\, k\, M\, S$ *{[\[]{}b[\]]{}&lt;N&gt;}* map$_{g'}$ :: ([\[]{}a[\]]{}&lt;k&gt; -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b&lt;k&gt;[\]]{}&lt;N/k&gt; map$_{g'}$ :: ([\[]{}a[\]]{}&lt;k&gt; -&gt; *{*$V^{-1}\, N/k$ *[\[]{}b&lt;k&gt;[\]]{}&lt;N/k&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b[\]]{}&lt;k[\*]{}m&gt; map$_{g'}$ :: ([\[]{}a[\]]{}&lt;k&gt; -&gt; b&lt;k&gt;) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b[\]]{}&lt;k[\*]{}m&gt; $\Rightarrow$ f’ :: [\[]{}a[\]]{}&lt;k&gt; -&gt; b&lt;k&gt; $\Rightarrow$ r’ :: [\[]{}b&lt;k&gt;[\]]{}&lt;N/k&gt; In other words, we can infer the return type from the single type transformation. What we have so far is $s'\,::\,[a']\left\langle N'\right\rangle $ $g'\,::\,[a']\left\langle N'\right\rangle \rightarrow[b']\left\langle N'\right\rangle $ $r'\,::\,[b']\left\langle N'\right\rangle $ $f'\,::\, a'\rightarrow b'$ $r'\,=\, g'\, s'$ $g'\,=\, map\, f'$ where $\mbox{type}\, a'\,=\,[a]\left\langle k\right\rangle $ $\mbox{type}\, b'\,=\,[b]\left\langle k\right\rangle $ $N'\,=\, N/k$ What we need now is the transformations between *s* and *s’* and *f* and *f’* To transform *s*: $s'\,=\,\mbox{reshapeTo}\, k\, s$ where $\mbox{reshapeTo}\,::\,\mbox{Int}\, k\Rightarrow k\rightarrow[a]\left\langle n\right\rangle \rightarrow[[a]\left\langle k\right\rangle ]\left\langle n/k\right\rangle $ We define the inverse for further use: $\mbox{reshapeFrom}\,::\,\mbox{Int}\, k\Rightarrow k\rightarrow[a\left\langle k\right\rangle ]\left\langle n\right\rangle \rightarrow[a]\left\langle n.k\right\rangle $ So the *R k M S t(s)* type transformation maps directly to *reshapeTo k s* The transformation from *f* to *f’* is even more straightforward, because the transformation of the original type raises the dimensionality $f'\,=\, map\, f$ In general, the original map is replaced by maps over both dimensions. $g'\,=\, map\,map\, f$ and in full, the transformed program becomes $r\,=\, (reshapeFrom \,k) \; .\; (map\,map\, f) \;.\; (reshapeTo\,k)\, s$ Reducing the dimensionality – map --------------------------------- Assume we have $g\,=\, map\, f$ $f\,::\,[a]\left\langle k\right\rangle \rightarrow[b]\left\langle k\right\rangle $ $r\,=\, g\, s$ And we apply the transformation *M* $S^{-1}$ $R^{-1}$*k* to *s*: $R^{-1}\, k\, t(s)=[[a]\left\langle 1\right\rangle ]\left\langle k.m\right\rangle $ $M\, S^{-1}\,[[a]\left\langle 1\right\rangle ]\left\langle k.m\right\rangle =[a]\left\langle k.m\right\rangle $ *$t(s')=[a]\left\langle k.m\right\rangle $* So we obtain $s'\,::\,=[a]\left\langle k.m\right\rangle $ As *g’* is applied to *s’*, we obtain $g'\,::\,[a]\left\langle k.m\right\rangle \rightarrow\left\{ [[b]\left\langle k\right\rangle ]\left\langle m\right\rangle \right\} $ Now we use inference on *map*: g’ s’ = map f’ s’ map$_{g'}$ :: (t\_1 -&gt; t\_2) -&gt; [\[]{}t\_1[\]]{} -&gt; [\[]{}t\_2[\]]{} map$_{g'}$ :: (a -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; *{[\[]{}[\[]{}b[\]]{}&lt;k&gt;[\]]{}&lt;m&gt;}* map$_{g'}$ :: (a -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; *M *$S^{-1}$ $R^{-1}$* k {[\[]{}[\[]{}b[\]]{}&lt;k&gt;[\]]{}&lt;m&gt;}* map$_{g'}$ :: (a -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b[\]]{}&lt;k[\*]{}m&gt; map$_{g'}$ :: (a -&gt; *{*$V^{-1}\, k.m$ *[\[]{}b[\]]{}&lt;k[\*]{}m&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b[\]]{}&lt;k[\*]{}m&gt; map$_{g'}$ :: (a -&gt; b) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b[\]]{}&lt;k[\*]{}m&gt; $\Rightarrow$ f’ :: a -&gt; b $\Rightarrow$ r’ :: [\[]{}b[\]]{}&lt;k[\*]{}m&gt; to express *f’* as a function of *f* , we need a *toVector k x* function $\mbox{toVector}\,::\,\mbox{Int}\, k\Rightarrow k\rightarrow a\rightarrow[a]\left\langle k\right\rangle $ The most intuitive implementation seems to be toVector :: k x = replicate k x Similarly, we need *fromVector k x* (although we don’t really need *k*) $\mbox{fromVector}\,::\,\mbox{Int}\, k\Rightarrow k\rightarrow[a]\left\langle k\right\rangle \rightarrow a$ The most intuitive implementation seems to be fromVector k (x:\_) = x With these, we simply say $f'\, x\,=\,\mbox{fromVector}\, k\,\left(f\,\left(\mbox{toVector}\, k\, x\right)\right)$ ### Correctness condition In general, the above transformation does not necessarily preserve the computation. However, we can see that a sufficient condition to preserves the computation is that *map f’ = f*: \[ConditionOnReducedDim\] Mapping *f’* over *s’* preserves the computation of mapping *f* over *s* iff *f = map h*   1. Observe that *s’ = reshapeFrom k s* and we must show that r’ = *g’ s’ = reshapeFrom k r = reshapeFrom k g s* 2. We show that *map f’ s’ = map h s’* 1. Mapping *f’* to *s’*: *r’ = g’ s’ = map f’ s’* *= [\[]{}f’ x1,f’ x2,...,f’ xk,f’ y1,f’ y2,...,f’ yk,...,f’ z1,f’ z2,...,f’ zk[\]]{}* 2. *f’* is identical to *h*: *f’ x* *= fromVector k (f (toVector k x))* *=head (f [\[]{}x[\]]{})* *= head (map h [\[]{}x[\]]{})* *= head [\[]{}h x[\]]{}* *= h x* $\Rightarrow$*f’ = h* $\Rightarrow$ *r’ = map f’ s’ = [\[]{}h x1,h x2,...,h xk,h y1,h y2,...,h yk,...,h z1,h z2,...,h zk[\]]{}* 3. Mapping *f* to *s*: *r = g s = map f s* *= map f [\[]{}[\[]{}x1,x2,...,xk[\]]{},[\[]{}y1,y2,...,yk[\]]{},...,[\[]{}z1,z2,...,zk[\]]{}[\]]{}* *= [\[]{}f [\[]{}x1,x2,...,xk[\]]{},f [\[]{}y1,y2,...,yk[\]]{},...,f [\[]{}z1,z2,...,zk[\]]{}[\]]{},...[\]]{}* *= [\[]{}map h [\[]{}x1,x2,...,xk[\]]{},map h [\[]{}y1,y2,...,yk[\]]{},...,map h [\[]{}z1,z2,...,zk[\]]{}[\]]{}* *= [\[]{}[\[]{}h x1,h x2,...,h xk[\]]{}, [\[]{}h y1,h y2,...,h yk[\]]{},...,[\[]{}h z1,h z2,...,h zk[\]]{}[\]]{}* 4. Finally, transforming *r* to *r’*: *reshapeFrom k r* *= reshapeFrom k [\[]{}[\[]{}h x1,h x2,...,h xk[\]]{}, [\[]{}h y1,h y2,...,h yk[\]]{},...,[\[]{}h z1,h z2,...,h zk[\]]{}[\]]{}* *= [\[]{}h x1,h x2,...,h xk,h y1,h y2,...,h yk,...,h z1,h z2,...,h zk[\]]{}* *= r’* Preserving the dimensionality – map ----------------------------------- With the same example as above, we apply the transformation ** $R\, n\, R^{-1}\, k$ to *s*: $R^{-1}\, k\, t(s)=[[a]\left\langle 1\right\rangle ]\left\langle k.m\right\rangle $ *$R\, n\,[[a]\left\langle 1\right\rangle ]\left\langle k.m\right\rangle =[a\left\langle n\right\rangle ]\left\langle k.m/n\right\rangle $* $t(s')=[a\left\langle n\right\rangle ]\left\langle k.m/n\right\rangle $ So we obtain $s'\,::\,[a\left\langle n\right\rangle ]\left\langle k.m/n\right\rangle $ As *g’* is applied to *s’*, we obtain $g'\,::\,[a\left\langle n\right\rangle ]\left\langle k.m/n\right\rangle \rightarrow\left\{ [b\left\langle k\right\rangle ]\left\langle m\right\rangle \right\} $ Again we use inference on *map*: g’ s’ = map f’ s’ map$_{g'}$ :: (t\_1 -&gt; t\_2) -&gt; [\[]{}t\_1[\]]{} -&gt; [\[]{}t\_2[\]]{} map$_{g'}$ :: (a&lt;n&gt; -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}a&lt;n&gt;[\]]{}&lt;k[\*]{}m/n&gt; -&gt; *{[\[]{}[\[]{}b[\]]{}&lt;k&gt;[\]]{}&lt;m&gt;}* map$_{g'}$ :: (a&lt;n&gt; -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; $R\, n\, R^{-1}\, k$*  {[\[]{}[\[]{}b[\]]{}&lt;k&gt;[\]]{}&lt;m&gt;}* map$_{g'}$ :: (a&lt;n&gt; -&gt; *{[\[]{}b[\]]{}&lt;k&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b&lt;n&gt;[\]]{}&lt;k[\*]{}m/n&gt; map$_{g'}$ :: (a -&gt; *{*$V^{-1}\, k.m/n$ *[\[]{}b&lt;n&gt;[\]]{}&lt;k[\*]{}m/n&gt;}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b[\]]{}&lt;k[\*]{}m&gt; map$_{g'}$ :: (a&lt;n&gt; -&gt; b&lt;n&gt;) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; [\[]{}b[\]]{}&lt;k[\*]{}m&gt; $\Rightarrow$ f’ :: a&lt;n&gt; -&gt; b&lt;n&gt; $\Rightarrow$ r’ :: [\[]{}b&lt;n&gt;[\]]{}&lt;k[\*]{}m/n&gt; As map is independent of the size of the vector, we have $f'\,=\, f$ Consequently, the computation will always be preserved. Increasing dimensionality – fold -------------------------------- We can easily show that if the operation on f is a fold, then increasing the dimensionality results in applying the fold to every dimension. \[FoldOnNestedList\] Repeated application of *fold* to a nested list is equivalent to applying *fold* to the flattened list *fold (fold f) acc [\[]{},,...[\]]{} = fold f acc [\[]{}x1,x2,...,xk,y1,y2,...,yk,...,z1,z2,...,zk[\]]{}*   *fold (fold f) acc [\[]{},,...,[\[]{}z1,z2,...,zk[\]]{}[\]]{}* *= (fold f ...* ** ** *... [\[]{}z1,z2,...,zk[\]]{})* *= (fold f ...* ** *... [\[]{}z1,z2,...,zk[\]]{})* *= (f ... (f (f ( ...* ** ** *... ) z1) z2) ... zk)* *= fold f acc [\[]{}x1,x2,...,xk,y1,y2,...,yk,...,z1,z2,...,zk[\]]{}* Furthermore, as we consider a streaming operations, we only consider the left fold (*foldl*). We assume the same program as for *map* above: $s\,::\,[a]\left\langle n\right\rangle $ $g\,::\,[a]\left\langle n\right\rangle \rightarrow b$ $r\,::\, b$ $f\,::\, b\rightarrow a\rightarrow b$ $acc\,::\, b$ $r\,=\, g\, s$ $g\,=\,\mbox{fold}\, f\, acc$ For completeness: $\mbox{fold}\,::\,(t_{2}\rightarrow t_{1}\rightarrow t_{2})\rightarrow t_{2}\rightarrow[t_{1}]\left\langle m\right\rangle \rightarrow t_{2}$ We transform the top-level type: $t_{s}'$*= R k M S* $t_{s}$ *= R k M S [\[]{}a[\]]{}&lt;n&gt; = [\[]{}[\[]{}a[\]]{}&lt;k&gt;[\]]{}&lt;n/k&gt;* So we obtain $s'\,::\,[a\left\langle k\right\rangle ]\left\langle n/k\right\rangle $ As *g’* is applied to *s’*, we obtain $g'\,::\,[a\left\langle n\right\rangle ]\left\langle k.m/n\right\rangle \rightarrow\left\{ b\right\} $ Using inference on *fold*: g’ s’ = fold f’ acc s’ fold$_{g'}$ :: (t\_2 -&gt; t\_1 -&gt; t\_2) -&gt; t\_2 -&gt; [\[]{}t\_1[\]]{}&lt;m&gt; -&gt; t\_2 fold$_{g'}$ :: (*{b}* -&gt; *{a}*-&gt;*{b}*) -&gt; *{b}* -&gt; [\[]{}a[\]]{}&lt;k&gt;&lt;n/k&gt; -&gt; *{b}* fold$_{g'}$ :: (*{b} -&gt; {*$V^{-1}\, n/k$ *[\[]{}a[\]]{}&lt;k&gt;&lt;n/k&gt;}* -&gt; *{b}*) -&gt; [\[]{}a[\]]{}&lt;k&gt;&lt;n/k&gt; -&gt; *{b}* fold$_{g'}$ :: (*{b} -&gt;* [\[]{}a[\]]{}&lt;k&gt; -&gt; *{b}*) -&gt; [\[]{}a[\]]{}&lt;k&gt;&lt;n/k&gt; -&gt; *{b}* At this point, the types are valid, so no further transformation is required fold$_{g'}$ :: (b -&gt; [\[]{}a[\]]{}&lt;k&gt; -&gt; b) -&gt; [\[]{}a[\]]{}&lt;k&gt;&lt;n/k&gt; -&gt; b $\Rightarrow$ f’ :: b -&gt; [\[]{}a[\]]{}&lt;k&gt; -&gt; b $\Rightarrow$ r’ :: b $\Rightarrow$ acc’ :: b To transform *f* into *f’*: $f'\,=\, fold\, f$ Decreasing dimensionality – fold -------------------------------- We assume the same program as for map above: $s\,::\,[a\left\langle k\right\rangle ]\left\langle m\right\rangle $ $g\,::\,[a\left\langle k\right\rangle ]\left\langle m\right\rangle \rightarrow b$ $r\,::\, b$ $f\,::\, b\rightarrow a\left\langle k\right\rangle \rightarrow b$ $acc\,::\, b$ $r\,=\, g\, s$ $g\,=\,\mbox{fold}\, f\, acc$ For completeness: $\mbox{fold}\,::\,(t_{2}\rightarrow t_{1}\rightarrow t_{2})\rightarrow t_{2}\rightarrow[t_{1}]\left\langle m\right\rangle \rightarrow t_{2}$ We transform the top-level type: $t_{s}'$*= $M\, S^{-1}\, R^{-1}\, k$* $t_{s}$ *= $M\, S^{-1}\, R^{-1}\, k$ [\[]{}a[\]]{}&lt;k&gt;&lt;m&gt; = [\[]{}a[\]]{}&lt;k.m&gt;* So we obtain $s'\,::\,[a]\left\langle k.m\right\rangle $ $s'\,=\,\mbox{flatten}\, s$ As *g’* is applied to *s’*, we obtain $g'\,::\,[a]\left\langle k.m\right\rangle \rightarrow\left\{ b\right\} $ Using inference on *fold*: g’ s’ = fold f’ acc s’ fold$_{g'}$ :: (t\_2 -&gt; t\_1 -&gt; t\_2) -&gt; t\_2 -&gt; [\[]{}t\_1[\]]{}&lt;m&gt; -&gt; t\_2 fold$_{g'}$ :: (*{b}* -&gt; *{[\[]{}a[\]]{}&lt;k&gt;}*-&gt;*{b}*) -&gt; *{b}* -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; *{b}* fold$_{g'}$ :: (*{b} -&gt; {*$V^{-1}\, k.m$ *[\[]{}a[\]]{}&lt;k.m&gt;}* -&gt; *{b}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; *{b}* fold$_{g'}$ :: (*{b} -&gt;* a -&gt; *{b}*) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; *{b}* At this point, the types are valid, so no further transformation is required fold$_{g'}$ :: (b -&gt; a -&gt; b) -&gt; [\[]{}a[\]]{}&lt;k[\*]{}m&gt; -&gt; b $\Rightarrow$ f’ :: b -&gt; a -&gt; b $\Rightarrow$ r’ :: b $\Rightarrow$ acc’ :: b To transform *f* into *f’*: $f'\, acc\, x\,=\, f\, acc\,\left(\mbox{toVector}\, k\, x\right)$ ### Correctness condition In the case of fold, “preserves the computation” means “produces an identical result”, as from the perspective of the type transformation, the type *b* is opaque. In general, folding *f’* over *s’* is not equal to folding *f* over *s*. However, a sufficient condition for equality is this: Folding *f’* over *s’* is equal to folding *f* over *s* iff *f = fold h*   1. *foldl f’ acc s’ = foldl h acc s’* *foldl f’ acc [\[]{}x1,x2,...,xk,y1,y2,...,yk,...,z1,z2,...,zk[\]]{}* [\[]{}def. of f’[\]]{} *= foldl (\\acc x -&gt; f acc (toVector k x)) acc [\[]{}x1,x2,...,xk,y1,y2,...,yk,...,z1,z2,...,zk[\]]{}* [\[]{}def of toVector[\]]{} *= foldl (\\acc x -&gt; f acc [\[]{}x[\]]{}) acc [\[]{}x1,x2,...,xk,y1,y2,...,yk,...,z1,z2,...,zk[\]]{}* [\[]{}def of foldl[\]]{} foldl h acc [\[]{}x[\]]{} = h acc x *= foldl (\\acc x -&gt; h acc x) acc [\[]{}x1,x2,...,xk,y1,y2,...,yk,...,z1,z2,...,zk[\]]{}* [\[]{}$\eta$ conversion[\]]{} *= foldl h acc [\[]{}x1,x2,...,xk,y1,y2,...,yk,...,z1,z2,...,zk[\]]{}* [\[]{}def. of foldl[\]]{} *= (h (... (h (h (... (h ... (h (h (fh (h ... (h (h acc x1) x2) ... xk) y1) y2) ... yk) ...) z1) z2) ...) zk)* 2. *foldl h acc s’ = foldl f acc s* *foldl f acc s* [\[]{}def. of f[\]]{} *= foldl (foldl h) acc s* [\[]{}Lemma \[FoldOnNestedList\] + def. of s’[\]]{} *= foldl h acc s’* About *zip* and *unzip* ------------------------ We use *zip* and *unzip* to change nested lists of tuples into tuples of nested lists. zip :: [\[]{}a[\]]{}&lt;n&gt; -&gt; [\[]{}b[\]]{}&lt;n&gt; -&gt; [\[]{}(a,b)[\]]{}&lt;n&gt; unzip :: [\[]{}(a,b)[\]]{}&lt;n&gt; -&gt; ([\[]{}a[\]]{}&lt;n&gt;,[\[]{}b[\]]{}&lt;n&gt;) The same type transformation must be applied to both arguments, e.g. for *R k [\[]{}a[\]]{}&lt;n&gt;*. In order to preserve the computation, it is quite clear that zip’ :: [\[]{}[\[]{}a[\]]{}&lt;k&gt;[\]]{}&lt;n/k&gt; -&gt; [\[]{}[\[]{}b[\]]{}&lt;k&gt;[\]]{}&lt;n/k&gt; -&gt; [\[]{}[\[]{}(a,b)[\]]{}&lt;k&gt;[\]]{}&lt;n/k&gt; can be implemented in terms of *zip* as zip’ xs’ ys’ = map (\\(x,y) -&gt; zip x y) (zip xs’ ys’) and similar for *unzip*. To simplify the discussion, we introduce a variant of *zip*, *zipt*, which takes a single tuple as argument, and a corresponding *unzipt*. zipt :: ([\[]{}a[\]]{}&lt;n&gt;,[\[]{}b[\]]{}&lt;n&gt;) -&gt; [\[]{}(a,b)[\]]{}&lt;n&gt; zipt (xs,ys) = zip xs ys and unzipt :: [\[]{}(a,b)[\]]{}&lt;n&gt; -&gt; ([\[]{}a[\]]{}&lt;n&gt;,[\[]{}b[\]]{}&lt;n&gt;) unzipt ltups = (map fst ltups, map snd ltups) then *zipt’* becomes zipt’ :: ([\[]{}a[\]]{}&lt;k&gt;&lt;n/k&gt;,[\[]{}b[\]]{}&lt;k&gt;&lt;n/k&gt;) -&gt; [\[]{}(a,b)[\]]{}&lt;k&gt;&lt;n/k&gt; zipt’ tup = map zipt (zipt tup) and similar for *unzipt*. Conclusion ========== The approach described allows to transform programs consisting of combinations of *map*, *foldl* and *zip* based on transformation of the types of the vectors on which the map or fold acts. We have shown that the the set *V(a,n)* of vectors of type $a$ and size $n$ is closed under the proposed operations for transforming the vector types, *S*,*M* and *R*, with the corollary that every combination of the transformations is reversible. We have shown that, for programs consisting of opaque functions and the operations *map*, *foldl* and *zip*, the program transformations can be automatically derived from the type transformations. This mechanism allows to generate correct-by-construction variants of the programs. The purpose of this works is to allow automatic selection of the variant most suitable for a given platform through optimisation against a platform cost model. This work is supported by the EPSRC through the TyTra project (EP/L00058X/1). [^1]: http://tytra.org.uk/

ArXiv

--- abstract: 'Wheeler’s Superspace is the arena in which Geometrodynamics takes place. I review some aspects of its geometrical and topological structure that Wheeler urged us to take seriously in the context of canonical quantum gravity.' author: - Domenico Giulini bibliography: - 'RELATIVITY.bib' - 'HIST-PHIL-SCI.bib' - 'MATH.bib' - 'QM.bib' date: 'Received: date / Accepted: date' title: The Superspace of Geometrodynamics --- > “The stage on which the space of the Universe moves is certainly not space itself. Nobody can be a stage for himself; he has to have a larger arena in which to move. The arena in which space does its changing is not even the space-time of Einstein, for space-time is the history of space changing with time. The arena must be a larger object: *superspace*$ \dots$ It is not endowed with three or four dimensions—it’s endowed with an *infinite* number of dimensions.” (J.A. Wheeler: *Superspace*, Harper’s Magazine, July 1974, p.9) Introduction {#sec:Intro} ============ From somewhere in the 1950’s on, John Wheeler repeatedly urged people who were interested in the quantum-gravity programme to understand the structure of a mathematical object that he called *Superspace* [@Wheeler:1968][@Wheeler:EinsteinsVision]. The intended meaning of ‘Superspace’ was that of a set, denoted by ${\mathcal{S}(\Sigma)}$, whose points faithfully correspond to all possible Riemannian geometries on a given 3-manifold $\Sigma$. Hence, in fact, there are infinitely many Superspaces, one for each 3-manifold $\Sigma$. The physical significance of this concept is suggested by the dynamical picture of General Relativity (henceforth abbreviated by GR), according to which spacetime is the history (time evolution) of space. Accordingly, in Hamiltonian GR, Superspace plays the rôle of the configuration space the cotangent bundle of which gives the phase space of 3d-diffeomorphism reduced states. Moreover, in Canonical Quantum Gravity (henceforth abbreviated by CQG), Superspace plays the rôle of the domain for the wave function which is still subject to the infamous Wheeler-DeWitt equation. In fact, Bryce DeWitt characterised the motivation for his seminal paper on CQG as follows: > “The present paper is the direct outcome of conversations with Wheeler, during which one fundamental question in particular kept recurring: *What is the structure of the domain manifold for the quantum-mechanical state functional?” ([@DeWittQTGI:1967], p.115)* More than 41 years after DeWitt’s important contribution I simply wish to give a small overview over some of the answers given so far to the question: *What is the structure of Superspace?* Here I interpret ‘structure’ more concretely as ‘metric structure’ and ‘topological structure’. But before answers can be attempted, we need to define the object at hand. This will be done in the next section; and before doing that, we wish to say a few more words on the overall motivation. Minkowski space is the stage for relativistic particle physics. It comes equipped with some structure (topological, affine, causal, metric) that is *not* subject to dynamical changes. Likewise, as was emphasised by Wheeler, the arena for Geometrodynamics is Superspace, which also comes equipped with certain non-dynamical structures. The topological and geometric structures of Superspace are as much a background for GR as the Minkowski space is for relativistic particle physics. Now, Quantum Field Theory has much to do with the automorphism group of Minkowski space and, in particular, its representation theory. For example, all the linear relativistic wave equations (Klein-Gordan, Weyl, Dirac, Maxwell, Proca, Rarita-Schwinger, Dirac-Bargmann, etc.) can be understood in this group-theoretic fashion, namely as projection conditions onto irreducible subspaces in some auxiliary Hilbert space. (In the same spirit a characterisation of ‘classical elementary system’ has been given as one whose phase space supports a transitive symplectic action of the Poincaré group [@Bacry:1967][@Arens:1971a].) This is how we arrive at the classifying meaning of ‘mass’ and ‘spin’. Could it be that Quantum Gravity has likewise much to do with the automorphism group of Superspace? Can we understand this group in any reasonable sense and what has it to do with four dimensional diffeomorphisms? If elementary particles *are* unitary irreducible representations of the Poincaré group, as Wigner once urged, what would the ‘elementary systems’ be that corresponded to irreducible representations of the automorphism group of Superspace? I do not know any reasonably complete answer to any of these questions. But the analogies at least suggests the possibility of some progress *if* these structures and their automorphisms could be be understood in any depth. This is a difficult task, as John Wheeler already foresaw forty years ago: > “Die Struktur des Superraumes enträtseln? Kaum in einem Sprung, und kaum heute!” ([@Wheeler:EinsteinsVision], p.61) Related in spirit is a recent approach in the larger context of 11-dimensional supergravity (see [@Damour.Nicolai:2007; @Damour.etal:2007] and references therein), which is based on the observation that the supergravity dynamics in certain truncations corresponds to geodesic motion of a massless spinning particle on an $E_{10}$ coset space. Here the Wheeler-DeWitt metric (\[eq:WDWmetric-c\]) appears naturally with the right GR-value $\lambda=1$, which in our context is the only value compatible with 4-dimensional diffeomorphism invariance, as we will discuss. This may suggest an interesting relation between $E_{10}$ and spacetime diffeomorphisms. Defining Superspace {#sec:Superspace} =================== As already said, Superspace ${\mathcal{S}(\Sigma)}$ is the set of all Riemannian geometries on the 3-manifold $\Sigma$. Here ‘geometries’ means ‘metrics up to diffeomorphisms’. Hence ${\mathcal{S}(\Sigma)}$ is identified as set of equivalence classes in ${\mathrm{Riem}(\Sigma)}$, the set of all smooth ($C^\infty$) Riemannian metrics in $\Sigma$ under the equivalence relation of being related by a smooth diffeomorphism. In other words, the group of all ($C^\infty$) diffeomorphisms, ${\mathrm{Diff}(\Sigma)}$, has a natural right action on ${\mathrm{Riem}(\Sigma)}$ via pullback and the orbit space is identified with ${\mathcal{S}(\Sigma)}$: $$\label{eq:DefSuperspace} {\mathcal{S}(\Sigma)}:={\mathrm{Riem}(\Sigma)}/{\mathrm{Diff}(\Sigma)}\,.$$ Let us now refine this definition. First, we shall restrict attention to those $\Sigma$ which are connected and closed (compact without boundary). We note that Einstein’s field equations by themselves do not exclude any such $\Sigma$. To see this, recall the form of the constraints for initial data $(h,K)$, where $h\in{\mathrm{Riem}(\Sigma)}$ and $K$ is a symmetric covariant 2nd rank tensor-field (to become the extrinsic curvature of $\Sigma$ in spacetime, once the latter is constructed from the dynamical part of Einstein’s equations) \[eq:Constraints\] $$\begin{aligned} \label{eq:Constraints-a} \Vert K\Vert^2_h-\bigl({\mathrm{Tr}}_h(K)\bigr)^2 -\bigl(R(h)-2\Lambda\bigr)&=-&(2\kappa)\rho_m\,,\\ \label{eq:Constraints-b} \mathrm{div}_h\bigl(K-h\,{\mathrm{Tr}}_h(K)\bigr)&=&(c\kappa)\,j\,,\end{aligned}$$ where $\rho_m$ and $j_m$ are the densities of energy and momentum of matter respectively, $R(h)$ is the Ricci scalar for $h$, and $\kappa=8\pi G/c^4$. Now, it is known that for any smooth function $f:\Sigma\rightarrow{\mathbb{R}}$ which is negative somewhere on $\Sigma$ there exists an $h\in{\mathrm{Riem}(\Sigma)}$ so that $R(h)=f$ [@Kazdan.Warner:1975]. Given that strong result, we may easily solve (\[eq:Constraints\]) for $j=0$ on any compact $\Sigma$ as follows: First we make the Ansatz $K=\alpha h$ for some constant $\alpha$ and some $h\in{\mathrm{Riem}(\Sigma)}$. This solves (\[eq:Constraints-b\]), whatever $\alpha,h$ will be. Geometrically this means that the initial $\Sigma$ will be a totally umbillic hypersurface in spacetime. Next we solve (\[eq:Constraints-a\]) by fixing $\alpha$ so that $\alpha^2>(\Lambda+\kappa\,\sup_\Sigma(\rho_m))/3$ and then choosing $h$ so that $R(h)=2\Lambda+2\kappa\rho_m-6\alpha^2$, which is possible by the result just cited because the right-hand side is negative by construction. This argument can be generalised to non-compact manifolds with a finite number of ends and asymptotically flat data [@Witt:1986a]. Next we refine the definition (\[eq:DefSuperspace\]), in that we restrict the group of diffeomorphisms to the proper subgroup of those diffeomorphisms that fix a preferred point, called $\infty\in\Sigma$, and the tangent space at this point: $$\label{eq:DefDiffF} {\mathrm{Diff_F}(\Sigma)}:=\bigl\{\phi\in{\mathrm{Diff}(\Sigma)}\mid \phi(\infty)=\infty,\, \phi_*(\infty)={\mathrm{id}}\vert_{T_\infty\Sigma}\bigr\}\,.$$ The reason for this is twofold: First, if one is genuinely interested in closed $\Sigma$, the space ${\mathcal{S}(\Sigma)}:={\mathrm{Riem}(\Sigma)}/{\mathrm{Diff}(\Sigma)}$ is not a manifold if $\Sigma$ allows for metrics with non-trivial isometry groups (not all $\Sigma$ do; compare footnote\[foot:DegSym\]). At those metrics ${\mathrm{Diff}(\Sigma)}$ clearly does not act freely, so that the quotient ${\mathrm{Riem}(\Sigma)}/{\mathrm{Diff}(\Sigma)}$ has the structure of a stratified manifold with nested sets of strata ordered according to the dimension of the isometry groups [@Fischer:1970]. In that case there is a natural way to minimally resolve the singularities [@Fischer:1986] which amounts to taking instead the quotient ${\mathrm{Riem}(\Sigma)}\times{\mathrm{F}(\Sigma)}/{\mathrm{Diff}(\Sigma)}$, where ${\mathrm{F}(\Sigma)}$ is the bundle of linear frames over $\Sigma$. The point here is that the action of ${\mathrm{Diff}(\Sigma)}$ is now free since there simply are no non-trivial isometries that fix a frame. Indeed, if $\phi$ is an isometry fixing some frame, we can use the exponential map and $\phi\circ\exp=\exp\circ\phi_*$ (valid for any isometry) to show that the subset of points in $\Sigma$ fixed by $\phi$ is open. Since this set is also closed and $\Sigma$ is connected, $\phi$ must be the identity. Now, the quotient ${\mathrm{Riem}(\Sigma)}\times{\mathrm{F}(\Sigma)}/{\mathrm{Diff}(\Sigma)}$ is isomorphic[^1] to $$\label{eq:DefSupF} {\mathcal{S}_{\mathrm{F}}(\Sigma)}:={\mathrm{Riem}(\Sigma)}/{\mathrm{Diff_F}(\Sigma)}\,,$$ albeit not in a natural way, since one needs to choose a preferred point $\infty\in \Sigma$. This may seem somewhat artificial if really all points in $\Sigma$ are considered to be equally real, but this is irrelevant for us as long as we are only interested in the isomorphicity class of Superspace. On the other hand, if we consider $\Sigma$ as the one-point compactification of a manifold with one end[^2], then (\[eq:DefSupF\]) would be the right space to start with anyway since then diffeomorphisms have to respect the asymptotic geometry in that end, like, e.g., asymptotic flatness. *Therefore, from now on, we shall refer to ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ as defined in (\[eq:DefSupF\]) as Superspace.* In view of the original definition (\[eq:DefSuperspace\]) it is usually called ‘extended Superspace’ [@Fischer:1970]. Clearly, the move from (\[eq:DefSuperspace\]) to (\[eq:DefSupF\]) would have been unnecessary in the closed case if one restricted attention to those manifolds $\Sigma$ which do not allow for metrics with continuous symmetries, i.e. whose degree of symmetry[^3]is zero. Even though these manifolds are not the ‘obvious’ ones one tends to think of first, they are, in a sense, ‘most’ 3-manifolds. On the other hand, in order not to deprive ourselves form the possibility of physical idealisations in terms of prescribed exact symmetries, we prefer to work with ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ defined in (\[eq:DefSupF\]) (called ‘extended superspace’ in [@Fischer:1970], as already mentioned). Let us add a few words on the point-set topology of ${\mathrm{Riem}(\Sigma)}$ and ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$. First, ${\mathrm{Riem}(\Sigma)}$ is an open positive convex cone in the topological vector space of smooth ($C^\infty$) symmetric covariant tensor fields over $\Sigma$. The latter space is a Fréchet space, that is, a locally convex topological vector space that admits a translation-invariant metric, $\overline{d}$, inducing its topology and with respect to which the space is complete. The metric can be chosen such that ${\mathrm{Diff}(\Sigma)}$ preserves distances. ${\mathrm{Riem}(\Sigma)}$ inherits this metric which makes it a metrisable topological space that is also second countable (recall also that metrisability implies paracompactness). ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ is given the quotient topology, i.e. the strongest topology in which the projection ${\mathrm{Riem}(\Sigma)}\rightarrow{\mathcal{S}_{\mathrm{F}}(\Sigma)}$ is continuous. This projection is also open since ${\mathrm{Diff_F}(\Sigma)}$ acts continuously on ${\mathrm{Riem}(\Sigma)}$. A metric $d$ on ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ is defined by $$\label{eq:DefMetricOnSup} d([h_1],[h_2]):=\sup_{\phi_1,\phi_2\in{\mathrm{Diff_F}(\Sigma)}}\, \overline{d}(\phi^*_1h_1,\phi^*_2h_2)\,,$$ which also turns ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ into a connected (being the continuous image of the connected ${\mathrm{Riem}(\Sigma)}$) metrisable and second countable topological space. Hence ${\mathrm{Riem}(\Sigma)}$ and ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ are perfectly decent connected topological spaces which satisfy the strongest separability and countability axioms. For more details we refer to [@Stern:1967][@Fischer:1986][@Fischer:1970]. The basic geometric idea is now to regard ${\mathrm{Riem}(\Sigma)}$ as principal fibre bundle with structure group ${\mathrm{Diff_F}(\Sigma)}$ and quotient ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$: $$\label{eq:PrinFibBundle} {\mathrm{Diff_F}(\Sigma)}\ \mathop{\longrightarrow}^i\ {\mathrm{Riem}(\Sigma)}\ \mathop{\longrightarrow}^p\ {\mathcal{S}_{\mathrm{F}}(\Sigma)}$$ where the maps $i$ are the inclusion and projection maps respectively. This is made possible by the so-called ‘slice theorems’ (see [@Ebin:1968][@Fischer:1970]), and the fact that the group acts freely and properly. This bundle structure has two far-reaching consequences regarding the geometry and topology of ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$. Let us discuss these in turn. Geometry of Superspace {#sec:GeometrySup} ====================== Elements of the Lie algebra ${\mathfrak{diff}_{\mathrm{F}}(\Sigma)}$ of ${\mathrm{Diff_F}(\Sigma)}$ are vector fields on $\Sigma$. For any such vector field $\xi$ on $\Sigma$ there is a vector field $V_\xi$ on ${\mathrm{Riem}(\Sigma)}$, called the vertical (or fundamental) vector field associated to $\xi$, whose value at $h\in{\mathrm{Riem}(\Sigma)}$ is just the infinitesimal change in $h$ generated by $\xi$, that is, $$\label{eq:VertVF} V_\xi(h)=-L_\xi h\,,$$ where $L_\xi$ denoted the Lie derivative with respect to $\xi$. Hence, for each $h\in{\mathrm{Riem}(\Sigma)}$, the map $V(h):\xi\mapsto V_\xi(h)$ is an anti-Lie homomorphism (the ‘anti’ being due to the fact that we have a *right* action of ${\mathrm{Diff}(\Sigma)}$ on ${\mathrm{Riem}(\Sigma)}$), that is $[V_\xi,V_\eta]=-V_{[\xi,\eta]}$, if the Lie structure on ${\mathfrak{diff}_{\mathrm{F}}(\Sigma)}$ is that of ordinary commutators of vector fields. The kernel of the map $V(h):\xi\mapsto V_\xi(h)$ consists of the finite-dimensional subspace of Killing fields on $(\Sigma,h)$. The vertical vectors at $h\in{\mathrm{Riem}(\Sigma)}$ therefore form a linear subspace ${\mathrm{Vert}}_h\subset T_h{\mathrm{Riem}(\Sigma)}$, isomorphic to the vector fields on $\Sigma$ modulo the Killing fields on $(\Sigma,h)$. It is a closed subspace due to the fact that the the operator $\xi\mapsto L_\xi h$ is overdetermined elliptic (cf. [@Besse:EinsteinManifolds], Appendices G-I). The family of ultralocal ‘metrics’ on ${\mathrm{Riem}(\Sigma)}$ is given by $$\label{eq:GeneralMetricRiem} {\mathcal{G}}_{(\alpha,\lambda)}(k,\ell)=\int_\Sigma d^3x\,\alpha\,\sqrt{\det(h)}\, \bigl(h^{ab}h^{cd}\,k_{ac}\ell_{bd} -\lambda\, (h^{ab}k_{ab})( h^{cd}\ell_{cd})\bigr)\,,$$ for each $k,\ell\in T_h{\mathrm{Riem}(\Sigma)}$. Here $\alpha$ is a positive real-valued function on $\Sigma$ and $\lambda$ a real number. An almost trivial but important observation is that ${\mathrm{Diff}(\Sigma)}$ is an isometry group with respect to all ${\mathcal{G}}_{(\alpha,\lambda)}$. The ‘metric’ picked by GR through the bilinear term in the constraint (\[eq:Constraints-a\]) corresponds to $\lambda=1$. The positive real-valued function $\alpha$ is not fixed and corresponds to the free choice of a lapse-function. In what follows we shall focus attention to $\alpha=1$. The pointwise bilinear form $(k,\ell)\mapsto (h\otimes h)(k,\ell)-\lambda{\mathrm{Tr}}_h(k){\mathrm{Tr}}_h(\ell)$ in the integrand of (\[eq:GeneralMetricRiem\]) defines a symmetric bilinear form on the six-dimensional space of symmetric tensors which is positive definite for $\lambda<1/3$, of signature $(1,5)$ for $\lambda>1/3$, and degenerate of signature $(0,5)$ for $\lambda=1/3$. It defines a metric on the homogeneous space $\mathrm{GL}(3)/\mathrm{O}(3)$, where the latter may be identified with the space of euclidean metrics on a 3-dimensional vector space. Parametrising it by $h_{ab}$, we have \[eq:WDWmetric\] $$\label{eq:WDWmetric-a} G_\lambda=G_{\lambda}^{ab\,cd}dh_{ab}\otimes dh_{cd} =-\epsilon d\tau\otimes d\tau+\frac{\tau^2}{c^2} {\mathrm{Tr}}(r^{-1}dr\otimes r^{-1}dr)\,,$$ where $$\label{eq:WDWmetric-b} r_{ab}:=[\det(h)]^{-1/3}h_{ab},\ \ \tau:=c\,[\det(h)]^{1/4},\ \ c^2:=16\vert\lambda-1/3\vert,\ \ \epsilon=\mbox{sign}(\lambda-1/3),$$ and $$\label{eq:WDWmetric-c} G_{\lambda}^{ab\,cd}=\tfrac{1}{2}\sqrt{\det(h)} \bigl(h^{ac}h^{bd}+h^{ad}h^{bc}-2\lambda h^{ab}h^{cd}\bigr)\,.$$ This is a 1+5 – dimensional warped-product geometry in the standard form of ‘cosmological’ models (Lorentzian for $\lambda>1/3$), here corresponding to the 1+5 decomposition $\mathrm{GL}(3)/\mathrm{O}(3)\cong {\mathbb{R}}\times\mathrm{SL}(3)/\mathrm{SO}(3)$ with scale factor $\tau/c$ and homogeneous Riemannian metric on five-dimensional ’space’ $\mathrm{SL}(3)/\mathrm{SO}(3)$, given by ${\mathrm{Tr}}(r^{-1}dr\otimes r^{-1}dr)=r^{ac}r^{bd}\,dr_{ab}\otimes dr_{cd}$. $\tau=0$ is a genuine ‘spacelike’ (‘cosmological’) curvature singularity. An early discussion of this finite-dimensional geometry was given by DeWitt [@DeWittQTGI:1967]. We stress that the Lorentzian nature of the Wheeler-DeWitt metric in GR (i.e. for $\lambda=1$) has nothing to do with the Lorentzian nature of spacetime, as we will see below from the statement of Theorem\[thm:Uniqueness\] and formulae (\[eq:HamConstrFunct\],\[eq:DeWittMetricMomenta\]); rather, it can be related to the attractivity of gravity [@Giulini.Kiefer:1994]. As for the infinite-dimensional geometry of ${\mathrm{Riem}(\Sigma)}$, we remark that an element $h$ of ${\mathrm{Riem}(\Sigma)}$ is a section in $T^*\Sigma\otimes T^*\Sigma$ and so is an element of $T_h{\mathrm{Riem}(\Sigma)}$. The latter has the fibre-metric (\[eq:WDWmetric\]). It is sometimes useful to use $h$ (for index raising) in order to identify $T_h{\mathrm{Riem}(\Sigma)}$ with sections in $T\Sigma\otimes T^*\Sigma\cong\mathrm{End}(T\Sigma)$, also because the latter has a natural structure as associative- (and hence also Lie-) algebra. Then the inner product (\[eq:GeneralMetricRiem\]) for $\alpha=1$ just reads (here and below $d\mu(h)=\sqrt{\det(h)}\,d^3x$) $$\label{eq:GeneralMetricRiem-Alt} {\mathcal{G}}_\lambda(k,\ell)=\int_\Sigma d\mu(h) \bigl({\mathrm{Tr}}(k\cdot\ell)-\lambda\,{\mathrm{Tr}}(k){\mathrm{Tr}}(\ell)\bigr)\,.$$ For $\lambda=0$ the infinite-dimensional geometry of ${\mathrm{Riem}(\Sigma)}$ has been studied in [@Freed.Groisser:1989]. They showed that all curvature components involving one or more pure-trace directions vanish and that the curvature tensor for the trace-free directions is given by (now making use of the natural Lie-algebra structure of $T\Sigma\otimes T^*\Sigma$) $$\label{eq:CurvTensorTracefree} R(k,\ell)m=-\tfrac{1}{4}[[k,\ell],m]\,.$$ In particular, this implies that the sectional curvatures involving pure trace directions vanish and that the sectional curvatures for trace-free directions $k,\ell$ are non-positive: $$\label{eq:SectCurvTracefree} \begin{split} K(k,\ell)&=-\tfrac{1}{4}\int_\Sigma d\mu(h)\, {\mathrm{Tr}}\bigl(k\cdot R(k,\ell)\ell\bigr)\\ &=-\tfrac{1}{4}\int_\Sigma d\mu(h)\, {\mathrm{Tr}}\bigl([k,\ell]\cdot[\ell,k]\bigr)\leq 0\,.\\ \end{split}$$ Similar results hold for other values of $\lambda$, though some positivity statements cease to hold for $\lambda>1/3$. We keep the generality in the value of $\lambda$ for the moment in order to show that the value $\lambda=1$ picked by GR is quite special. Since, as already stated, all elements of ${\mathrm{Diff_F}(\Sigma)}$ are isometries of ${\mathcal{G}}_\lambda$, it is natural to try to define a bundle connection on ${\mathrm{Riem}(\Sigma)}$ by taking the *horizontal subspace* ${\mathrm{Hor}}_h^\lambda$ at each $T_h{\mathrm{Riem}(\Sigma)}$ to be the ${\mathcal{G}}_\lambda$–orthogonal complement to ${\mathrm{Vert}}_h$, as suggested in [@Giulini:1995c]. From (\[eq:GeneralMetricRiem\]) one sees that $k\in T_h{\mathrm{Riem}(\Sigma)}$ is orthogonal to all $L_\xi h$ iff $$\label{eq:HorizontalityCond} (\mathcal{O}_\lambda k)^a:=-\nabla^b(k_{ab}-\lambda h_{ab}h^{cd}k_{cd})=0\,.$$ But note that orthogonality does not imply transversality if the metric is indefinite, as for $\lambda=1$. In that case the intersection ${\mathrm{Vert}}_h\cap{\mathrm{Hor}}_h^\lambda$ may well be non trivial, which implies that there is no well defined projection map $$\label{eq:HorProjMap} {\mathrm{hor}}^\lambda_h: T_h{\mathrm{Riem}(\Sigma)}\rightarrow {\mathrm{Hor}}_h^\lambda\,.$$ The definition of this map would be as follows: Let $k\in T_h{\mathrm{Riem}(\Sigma)}$, find a vector field $\xi$ on $\Sigma$ such that $k-V_\xi$ is horizontal. Then $V_\xi$ is the ($\lambda$ dependent) vertical component of $k$ and the map $k\mapsto k-V_\xi$ is the ($\lambda$ dependent) horizontal projection (\[eq:HorProjMap\]). When does that work? Well, according to (\[eq:HorizontalityCond\]), the condition for $k-V_\xi$ to be horizontal for given $k$ is equivalent to the following differential equation for $\xi$: $$\label{eq:HorizontalProjCond} D_\lambda\xi:= \bigl( \delta d+2(1-\lambda)d\delta-2{\mathrm{Ric}}\bigr)\xi= \mathcal{O}_h k\,.$$ Here we regarded $\xi$ as one-form and $d,\delta$ denote the standard exterior differential and co-differential ($\delta\xi=-\nabla^a\xi_a$) respectively. Moreover, ${\mathrm{Ric}}$ is the endomorphism on one-forms induced by the Ricci tensor ($\xi_a\mapsto R^b_a\xi_b$). Note that the right-hand side of (\[eq:HorizontalProjCond\]) is $L^2$-orthogonal for all $k$ to precisely the Killing fields. The singular nature of the GR value $\lambda=1$ is now seen from writing down the principal symbol of the operator $D_\lambda$ on the left-hand side of (\[eq:HorizontalProjCond\]), $$\label{eq:SymbOpD} \sigma_\lambda(\zeta)^a_b=\Vert\zeta\Vert \left( \delta^a_b+(1-2\lambda)\frac{\zeta^a\zeta_b}{\Vert\zeta\Vert} \right)\,,$$ whose determinant is $\Vert\xi\Vert^62(1-\lambda)$. Hence $\sigma_\lambda$ is positive definite for $\lambda<1$ and indefinite but still an isomorphism for $\lambda >1$. This means that $D_\lambda$ is elliptic for $\lambda\ne 1$ (strongly elliptic[^4] for $\lambda<1$) but fails to be elliptic for precisely the GR value $\lambda=1$. This implies the possibility for the kernel of $D_{\lambda=1}$ to become infinite dimensional. The last remark has a direct implication as regards the intersection of the horizontal and vertical subspaces. Recall that solutions $\xi$ to $D_\lambda\xi=0$ modulo Killing fields (which always solve this equation) correspond faithfully to vertical vectors at $h\in{\mathrm{Riem}(\Sigma)}$ (via $\xi\mapsto V_\xi(h)$) which are also horizontal. Since Killing vectors span at most a finite dimensional space, an infinite dimensional intersection ${\mathrm{Vert}}_h\cap{\mathrm{Hor}}_h^\lambda$ would be implied by an infinite dimensional kernel of $D_1$. That this possibility for $\lambda=1$ is actually realised for any $\Sigma$ is easy to see: Take a metric $h$ on $\Sigma$ that is flat in an open region $U\subset\Sigma$, and consider $k\in T_h{\mathrm{Riem}(\Sigma)}$ of the form $k_{ab}=2\nabla_a\nabla_b\phi$, where $\phi$ is a real-valued function on $\Sigma$ whose support is contained in $U$. Then $k$ is vertical since $k=L_\xi h$ for $\xi={\mathrm{grad}}\phi$ (a non-zero gradient vector-field is never Killing on a compact $\Sigma$), and also horizontal since $k$ satisfies (\[eq:HorizontalityCond\]). Such $\xi$ clearly span an infinite-dimensional subspace in the kernel of $D_1$. On the other hand, for $\lambda=1$ there are also always open sets of $h\in{\mathrm{Riem}(\Sigma)}$ (and of $[h]\in{\mathcal{S}_{\mathrm{F}}(\Sigma)}$) for which the kernel of $D_1$ is trivial (the kernel clearly depends only on the diffeomorphism class $[h]$ of $h$). For example, consider metrics with negative definite Ricci tensor, which exist for any closed $\Sigma$ [@Gao.Yao:1986]. (Note that Ricci-negative geometries never allow for non-trivial Killing fields.) Then it is clear from the definition of $D_\lambda$ that it is a positive-definite operator for $\lambda\leq 1$. Hence the intersection ${\mathrm{Vert}}_h\cap{\mathrm{Hor}}_h^\lambda$ is trivial. In the latter case it is interesting to observe that ${\mathcal{G}}_\lambda$ restricted to ${\mathrm{Vert}}_h$ ($h$ Ricci negative) is positive definite, since $$\label{eq:VerticalMetricRicciFlat} {\mathcal{G}}_\lambda\bigl(V_\xi(h),V_\xi(h)\bigr) =2\int_\Sigma d\mu(h)\,h(\xi,D_\lambda\xi)\,.$$ This means that ${\mathcal{G}}_\lambda$ restricted to the orthogonal complement, ${\mathrm{Hor}}_h^\lambda$, contains infinitely many negative and infinitely may positive directions and the same $(\infty,\infty)$ – signature is then directly inherited by $T_{[h]}{\mathcal{S}_{\mathrm{F}}(\Sigma)}$ for any Ricci-negative geometry $[h]$. Far less generic but still interesting examples for trivial intersections ${\mathrm{Vert}}_h\cap{\mathrm{Hor}}_h^\lambda$ in case $\lambda=1$ are given by Einstein metrics with positive Einstein constants. Since this condition implies constant positive sectional curvature, such metrics only exist on manifolds $\Sigma$ with finite fundamental group, so that $\Sigma$ must be a spherical space form $S^3/G$, where $G$ is a finite subgroup of $SO(4)$ acting freely on $S^3$. Also note that the subspace in ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ of Einstein geometries is finite dimensional. ![\[fig:FigGeomSuperspace\]The rectangle depicts the space ${\mathrm{Riem}(\Sigma)}$ which is fibred by the orbits of ${\mathrm{Diff_F}(\Sigma)}$ (curved vertical lines). The metric ${\mathcal{G}}_1$ on ${\mathrm{Riem}(\Sigma)}$ is such that as we move along ${\mathrm{Riem}(\Sigma)}$ transversal to the fibres the “light-cones” tilt relative to the fibre directions. The process here shows a transition at $[h']$ where some fibre directions are lightlike and no metric can be defined in $T_{[h']}{\mathcal{S}_{\mathrm{F}}(\Sigma)}$, whereas they are timelike at $[h]$ and spacelike at $[h'']$. The parallelogram at $h$ merely indicates the horizontal and vertical components of a vector in $T_h{\mathrm{Riem}(\Sigma)}$.](FigGeomSuperspace "fig:"){width="0.75\linewidth"} (2,0)[${\mathcal{S}_{\mathrm{F}}(\Sigma)}$]{} (2,45)[${\mathrm{Riem}(\Sigma)}$]{} (-224,118)[$h$]{} (-167,62)[$h'$]{} (-136,140)[$h''$]{} (-191,-8)[$[h]$]{} (-164,-8)[$[h']$]{} (-108,-8)[$[h'']$]{} (-62,112) [15]{}[horizontal]{} (-69,122) [93]{}[vertical]{} Now, any solution $\xi$ to $D_1\xi=0$ must be divergenceless (take the co-differential $\delta$ of this equation) and hence Killing. The last statement follows without computation from the fact that $D_1\xi=0$ is nothing but the condition that $L_\xi h$ is ${\mathcal{G}}_1$–orthogonal to all vertical vectors in $T_h{\mathrm{Riem}(\Sigma)}$, which for divergenceless $\xi$ (traceless $L_\xi h$) is equivalent to ${\mathcal{G}}_0$–orthogonality, but then positive definiteness of ${\mathcal{G}}_0$ and ${\mathcal{G}}_0(L_\xi h,L_\xi h)=0$ immediately imply $L_\xi h=0$. This shows the triviality of ${\mathrm{Vert}}_h\cap{\mathrm{Hor}}_h^\lambda$. The foregoing shows that ${\mathcal{G}}_1$ indeed defines a metric at $T_{[h]}{\mathcal{S}_{\mathrm{F}}(\Sigma)}$ for Ricci-positive Einstein geometries $[h]$. How does the signature of this metric compare to the signature $(\infty,\infty)$ at Ricci-negative geometries? The answer is surprising: Take, e.g., for $[h]$ the round geometry on $\Sigma=S^3$. Then it can be shown that ${\mathcal{G}}_1$ defines a *Lorentz geometry* on $T_{[h]}{\mathcal{S}_{\mathrm{F}}(\Sigma)}$, that is with signature $(1,\infty)$, containing exactly one negative direction [@Giulini:1995c]. This means that the signature of the metric defined at various points in Superspace varies strongly, with intermediate transition regions where no metric can be defined at all due to signature change. Figure\[fig:FigGeomSuperspace\] is an attempt to picture this situation. Intermezzo: GR as simplest representation of symmetry {#sec:IntermezzoGRasSymRep} ===================================================== It is well known that the field equations of GR have certain uniqueness properties and can accordingly be ‘deduced’ under suitable hypotheses involving a *symmetry principle* (diffeomorphism invariance), the equivalence principle, and some apparently mild technical hypotheses. More precisely, the equivalence principle suggests to only take the metric as dynamical variable [@ThorneLeeLightman:1973] representing the gravitational field (to which matter then couples universally), whereas diffeomorphism invariance, derivability from an invariant Lagrangian (alternatively: local energy-momentum conservation in the sense of covariant divergencelessness), dependence of the equations on the metric up to at most second derivatives, and, finally, four-dimensionality lead uniquely to the left-hand side of Einstein’s equation, including a possibly non-vanishing cosmological constant [@Lovelock:1972]. Here we will review how this ‘deduction’ works in the Hamiltonian setting on phase space $T^*{\mathrm{Riem}(\Sigma)}$, which goes back to [@Hojman.etal:1973; @Teitelboim:1973; @Kuchar:1973; @Hojman.etal:1976]. 3+1 decomposition ----------------- Since the 3+1 split of Einstein’s equations has already been introduced in Claus Kiefer’s contribution I can be brief on that point. The basic idea is to first imagine a spacetime $(M,g)$ being given, where topologically $M$ is a product ${\mathbb{R}}\times\Sigma$. Spacetime is then considered as the trajectory (history) of space in the following way: Let ${\mathrm{Emb}(\Sigma{,}M)}$ denote the space of smooth spacelike embeddings $\Sigma\rightarrow M$. We consider a curve ${\mathbb{R}}\ni t\rightarrow{\mathcal{E}}_t\in{\mathrm{Emb}(\Sigma{,}M)}$ corresponding to a one-parameter family of smooth embeddings with spacelike images. We assume the images ${\mathcal{E}}_t(\Sigma)=:{\mathcal{E}}_t\subset M$ to be mutually disjoint and moreover that $\hat{\mathcal{E}}:{\mathbb{R}}\times\Sigma\rightarrow M$, $(t,p)\mapsto{\mathcal{E}}_t(p)$, is an embedding (it is sometimes found convenient to relax this condition, but this is of no importance here). The Lorentz manifold $({\mathbb{R}}\times\Sigma,{\mathcal{E}}^*g)$ may now be taken as (${\mathcal{E}}$–dependent) representative of $M$ (or at least some open part of it) on which the leaves of the above foliation simply correspond to the $t=\mathrm{const.}$ hypersurfaces. Let $n$ denote a field of normalised timelike vectors normal to these leaves. $n$ is unique up to orientation, so that the choice of $n$ amounts to picking a ‘future direction’. The tangent vector $d{\mathcal{E}}_t/dt\vert_{t=0}$ at ${\mathcal{E}}_0\in{\mathrm{Emb}(\Sigma{,}M)}$ corresponds to a vector field over ${\mathcal{E}}_0$ (i.e. section in $T(M)\vert_{{\mathcal{E}}_0}$), given by $$\label{eq:LapseShift} \frac{d{\mathcal{E}}_t(p)}{dt}\Big\vert_{t=0} =:\frac{\partial}{\partial t}\Big\vert_{{\mathcal{E}}_0(p)} =\alpha n+\beta$$ with components $(\alpha,\beta)=(\mathrm{lapse},\mathrm{shift})$ normal and tangential to $\Sigma_0\subset M$. Conversely, each vector field $V$ on $M$ defines a vector field $X(V)$ on ${\mathrm{Emb}(\Sigma{,}M)}$, corresponding to the left action of $\mathrm{Diff}(M)$ on ${\mathrm{Emb}(\Sigma{,}M)}$ by composition. In local coordinates $y^\mu$ on $M$ and $x^k$ on $\Sigma$ it can be written as $$\label{eq:XofV} X(V)=\int_\Sigma d^3x\,V^\mu(y(x))\frac{\delta}{\delta y^\mu(x)}\,.$$ One easily verifies that $X:V\mapsto X(V)$ is a Lie homomorphism: $$\label{eq:LieHomo} \bigl[X(V),X(W)\bigr]=X\bigl([V,W]\bigr)\,.$$ In this sense, the Lie algebra of the four-dimensional diffeomorphism group is implemented on phase space of any generally covariant theory whose phase space includes the embedding variables [@Isham.Kuchar:1985a] (so-called ‘parametrised theories’). Alternatively, decomposing (\[eq:XofV\]) into normal and tangential components with respect to the leaves of the embedding at which the tangent-vector field to ${\mathrm{Emb}(\Sigma{,}M)}$ is evaluated, yields an embedding-dependent parametrisation of $X(V)$ in terms of $(\alpha,\beta)$, $$\label{eq:X-alphabeta} X(\alpha,\beta)= \int_\Sigma d^3x \Bigl(\alpha(x)n^\mu[y](x)+\beta^m(x)\partial_m y^\mu(x) \Bigr)\,\frac{\delta}{\delta y^\mu(x)}\,,$$ where $y$ in square brackets indicates the functional dependence of $n$ on the embedding. The functional derivatives of $n$ with respect to $y$ can be computed (see the Appendix of [@Teitelboim:1973]) and the commutator of deformation generators then follows to be, $$\label{eq:X(ab)Comm} \bigl[X(\alpha_1,\beta_1)\,,\,X(\alpha_2,\beta_2)\bigr] =\,-\,X(\alpha',\beta')\,,$$ where \[eq:X(ab)CommValues\] $$\begin{aligned} \label{eq:X(ab)Comm-b} &\alpha' &\,= \beta_1(\alpha_2)-\beta_2(\alpha_1)\,,\\ \label{eq:X(ab)Comm-c} &\beta' &\,= [\beta_1,\beta_2]+\sigma\alpha_1\,\mathrm{grad}_h(\alpha_2) -\sigma\alpha_2\mathrm{grad}_h\,(\alpha_1)\,.\end{aligned}$$ Here we left open whether spacetime $M$ is Lorentzian ($\sigma=1$) or Euclidean ($\sigma=-1$), just in order to keep track how the signature of spacetime, $(-\sigma,+,+,+)$, enters. Note that the $h$-dependent gradient field for the scalar function $\alpha$ is given by $\mathrm{grad}_h(\alpha)=(h^{ab}\partial_b\alpha)\partial_a$. The geometric idea behind (\[eq:X(ab)CommValues\]) is summarised in Figure\[fig:FigCommDia\]. ![\[fig:FigCommDia\]An (infinitesimal) hypersurface deformation with parameters $(\alpha_1,\beta_1)$ that maps $\Sigma\mapsto\Sigma_1$, followed by one with parameters $(\alpha_2,\beta_2)$ that maps $\Sigma_1\mapsto\Sigma_{12}$ differs by one with parameters $(\alpha',\beta')$ given by (\[eq:X(ab)CommValues\]) from that in which the maps with the same parameters are composed in the opposite order.](FigCommDia "fig:"){width="0.32\linewidth"} (-123,51)[$\Sigma$]{} (-68,81)[$\Sigma_2$]{} (-69,21)[$\Sigma_1$]{} (-10,108)[$\Sigma_{21}$]{} (-10,-3)[$\Sigma_{12}$]{} (-113,65) [25]{}$(\alpha_2,\beta_2)$ (-56,94) [25]{}$(\alpha_1,\beta_1)$ (-113,41) [-25]{}$(\alpha_1,\beta_1)$ (-56,13) [-25]{}$(\alpha_2,\beta_2)$ (3,54)[$(\alpha',\beta')$]{} Hamiltonian geometrodynamics ---------------------------- The idea of Hamiltonian Geometrodynamics is to realise these relations in terms of a Hamiltonian system on the phase space of physical fields. The most simple case is that where the latter merely include the spatial metric $h$ on $\Sigma$, so that the phase space is the cotangent bundle $T^*{\mathrm{Riem}(\Sigma)}$ over ${\mathrm{Riem}(\Sigma)}$. One then seeks a correspondence $$\label{eq:PhaseSpaceDist} (\alpha,\beta)\mapsto\bigl( H(\alpha,\beta)\,:\,T^*{\mathrm{Riem}(\Sigma)}\rightarrow{\mathbb{R}}\bigr)\,,$$ where $$\label{eq:NormalDefCorr} H(\alpha,\beta)[h,\pi]:= \int_\Sigma d^3x\bigl( \alpha(x){\mathcal{H}}[h,\pi](x)+h_{ab}(x)\beta^a(x){\mathcal{D}}^b[h,\pi](x)\bigr)\,,$$ with integrands ${\mathcal{H}}[h,\pi](x)$ and ${\mathcal{D}}^b[h,\pi](x)$ yet to be determined. $H$ should be regarded as distribution (here the test functions are $\alpha$ and $\beta^a$) with values in real-valued functions on $T^*{\mathrm{Riem}(\Sigma)}$. Now, the essential requirement is that the Poisson brackets between the $H(\alpha,\beta)$ are, up to a minus sign,[^5] as in (\[eq:X(ab)CommValues\]): $$\label{eq:H(ab)Comm} \bigl\{H(\alpha_1,\beta_1)\,,\,H(\alpha_2,\beta_2)\bigr\} =H(\alpha',\beta')\,.$$ Once the distribution $H$ satisfying (\[eq:H(ab)Comm\]) has been found, we can turn around the arguments given above and recover the action of the Lie algebra of four-dimensional diffeomorphism on the extended phase space including embedding variables [@Isham.Kuchar:1985b]. That such an extension is indeed necessary has been shown in [@Pons:2003], where obstructions against the implementation of the action of the Lie algebra of four-dimensional diffeomorphisms have been identified in case the dynamical fields include non-scalar ones. Why constraints --------------- From this follows a remarkable uniqueness result. Before stating it with all its hypotheses, we show why the constraints ${\mathcal{H}}[h,\pi]=0$ and ${\mathcal{D}}^b[h,\pi]=0$ must be imposed. Consider the set of smooth real-valued functions on phase space, $F:T^*{\mathrm{Riem}(\Sigma)}\rightarrow{\mathbb{R}}$. They are acted upon by all $H(\alpha,\beta)$ via Poisson bracketing: $F\mapsto\bigl\{F,H(\alpha,\beta)\bigr\}$. This defines a map from $(\alpha,\beta)$ into the derivations of phase-space functions. We require this map to also respect the commutation relation (\[eq:H(ab)Comm\]), that is, we require $$\label{eq:Der(ab)Comm-1} \bigl\{\bigl\{F,H(\alpha_1,\beta_1)\bigr\},H(\alpha_2,\beta_2)\bigr\}- \bigl\{\bigl\{F,H(\alpha_2,\beta_2)\bigr\},H(\alpha_1,\beta_1)\bigr\} =\bigl\{F,H\bigr\}(\alpha',\beta')\,.$$ The subtle point to be observed here is the following: Up to now the parameters $(\alpha_1,\beta_1)$ and $(\alpha_2,\beta_2)$ were considered as given functions of $x\in\Sigma$, independent of the fields $h(x)$ and $\pi(x)$, i.e. independent of the point of phase space. However, from (\[eq:X(ab)Comm-c\]) we see that $\beta'(x)$ does depend on $h(x)$. This dependence may not give rise to extra terms $\propto\{F,\alpha'\}$ in the Poisson bracket, for, otherwise, the extra terms would prevent the map $(\alpha,\beta)\mapsto\bigl\{-,H(\alpha,\beta)\bigr\}$ from being a homomorphism from the algebraic structure of hypersurface deformations into the derivations of phase-space functions. This is necessary in order to interpret $\bigl\{-,H(\alpha,\beta)\bigr\}$ as a generator (on phase-space functions) of a *spacetime* evolution corresponding to a normal lapse $\alpha$ and tangential shift $\beta$. In other words, the evolution of observables from an initial hypersurface $\Sigma_i$ to a final hypersurface $\Sigma_f$ must be independent of the intermediate foliation (‘integrability’ or ‘path independence’  [@Teitelboim:1973; @Hojman.etal:1973; @Hojman.etal:1976]). Therefore we placed the parameters $(\alpha',\beta')$ outside the Poisson bracket on the right-hand side of (\[eq:Der(ab)Comm-1\]), to indicate that no differentiation with respect to $h,\pi$ should act on them. To see that this requirement implies the constraints, rewrite the left-hand side of (\[eq:Der(ab)Comm-1\]) in the form $$\label{eq:Der(ab)Comm-2} \begin{split} & \bigl\{\bigl\{F,H(\alpha_1,\beta_1)\bigr\},H(\alpha_2,\beta_2)\bigr\}- \bigl\{\bigl\{F,H(\alpha_2,\beta_2)\bigr\},H(\alpha_1,\beta_1)\bigr\}\\ &\quad=\,\bigl\{F,\bigl\{H(\alpha_1,\beta_1),H(\alpha_2,\beta_2)\bigr\}\bigr\}\\ &\quad=\,\bigl\{F,H(\alpha',\beta')\bigr\}\\ &\quad=\,\bigl\{F,H\bigr\}(\alpha',\beta') +H\bigl(\{F,\alpha'\}\,,\,\{F,\beta'\}\bigr)\,, \end{split}$$ where the first equality follows from the Jacobi identity, the second from (\[eq:H(ab)Comm\]), and the third from the Leibniz rule. Hence the requirement (\[eq:Der(ab)Comm-1\]) is equivalent to $$\label{eq:Der(ab)Comm-3} H\bigl(\{F,\alpha'\}\,,\,\{F,\beta'\}\bigr)=0$$ for all phase-space functions $F$ to be considered and all $\alpha',\beta'$ of the form (\[eq:X(ab)CommValues\]). Since only $\beta'$ depends on phase space, more precisely on $h$, this implies the vanishing of the phase-space functions $H\bigl(0,\{F,\beta'\}\bigr)$ for all $F$ and all $\beta'$ of the form (\[eq:X(ab)Comm-c\]). This can be shown to imply $H(0,\beta)=0$, i.e. ${\mathcal{D}}[h,\pi]=0$. Now, in turn, for this to be preserved under all evolutions we need $\bigl\{H(\alpha,\tilde \beta),H(0,\beta)\bigr\}=0$, and hence in particular $\bigl\{H(\alpha,0),H(0,\beta)\bigr\}=0$ for all $\alpha,\beta$, which implies $H(\alpha,0)=0$, i.e. ${\mathcal{H}}[h,\pi]=0$. So we see that the constraints indeed follow. Sometimes the constraints $H(\alpha,\beta)=0$ are split into the *Hamiltonian (or scalar) constraints*, $H(\alpha,0)=0$, and the *diffeomorphisms (or vector) constraints*, $H(0,\beta)=0$. The relations (\[eq:H(ab)Comm\]) with (\[eq:X(ab)CommValues\]) then show that the vector constraints form a Lie-subalgebra which, because of $\{H(0,\beta),H(\alpha,0)\} =H\bigl(\beta(\alpha),0\bigr)\ne H(0,\beta')$, is not an ideal. This means that the Hamiltonian vector fields for the scalar constraints are not tangent to the surface of vanishing vector constraints, except where it intersects the surface of vanishing scalar constraints. This implies that the scalar constraints do not act on the solution space for the vector constraints, so that one simply cannot first reduce the vector constraints and then, on the solutions of that, search for solutions to the scalar constraints. Also, it is sometimes argued that the scalar constraints should *not* be regarded as generators of gauge transformations but rather as generators of physically meaningful motions whose effect is to change the physical state in a fashion that is, in principle, observable. See [@Kuchar:1993] and also [@Barbour.Foster:2008] and Sect.2.3 of Claus Kiefer’s contribution for a recent revival of that discussion. However, it seems inconsistent to me to simultaneously assume 1) physical states to always satisfy the scalar constraints and 2) physical observables to exist which do not Poisson commute with the scalar constraints: The Hamiltonian vector field corresponding to such an ‘observable’ will not be tangent to the surface of vanishing scalar constraints and hence will transform physical to unphysical states upon being actually measured. Uniqueness of Einstein’s geometrodynamics ----------------------------------------- It is sometimes stated that the relations (\[eq:H(ab)Comm\]) together with (\[eq:X(ab)CommValues\]) determine the function $H(\alpha,\beta):T^*{\mathrm{Riem}(\Sigma)}\rightarrow{\mathbb{R}}$, i.e. the integrands ${\mathcal{H}}[h,\pi]$ and ${\mathcal{D}}[h,\pi]$, uniquely up to two free parameters, which may be identified with the gravitational and the cosmological constants. This is a mathematical overstatement if read literally, since the result can only be shown if certain additional assumptions are made concerning the action of $H(\alpha,\beta)$ on the basic variables $h$ and $\pi$. The first such assumption concerns the intended (‘semantic’ or ‘physical’) meaning of $H(0,\beta)$, namely that the action of $H(0,\beta)\}$ on $h$ or $\pi$ is that of an infinitesimal spatial diffeomorphism of $\Sigma$. Hence it should be the spatial Lie derivative, $L_\beta$, applied to $h$ or $\pi$. It then follows from the general Hamiltonian theory that $H(0,\beta)$ is given by the *momentum map* that maps the vector field $\beta$ (viewed as element of the Lie algebra of the group of spatial diffeomorphisms) into the function on phase space given by the contraction of the momentum with the $\beta$-induced vector field $h\rightarrow L_\beta h$ on ${\mathrm{Riem}(\Sigma)}$: $$\label{eq:MomentiumMap} H(0,\beta)=\int_\Sigma d^3x\, \pi^{ab}(L_\beta h)_{ab} =-2\int_\Sigma d^3x (\nabla_a\pi^{ab})h_{bc}\beta^c\,.$$ Comparison with (\[eq:NormalDefCorr\]) yields $$\label{eq:DiffConstrFunct} {\mathcal{D}}^b[h,\pi]=-2\nabla_a\pi^{ab}\,.$$ The second assumption concerns the intended (‘semantic’ or ‘physical’) meaning of $H(\alpha,0)$, namely that $\{-,H(\alpha,0)\}$ acting on $h$ or $\pi$ is that of an infinitesimal ‘timelike’ diffeomorphism of $M$ normal to the leaves ${\mathcal{E}}_t(\Sigma)$. If $M$ were given, it is easy to prove that we would have $L_{\alpha n}h=2\alpha\,K$, where $n$ is the timelike field of normals to the leaves ${\mathcal{E}}_t(\Sigma)$ and $K$ is their extrinsic curvature. Hence one requires $$\label{eq:Anticip} \{h,H(\alpha,0)\}=2\alpha\,K\,.$$ Note that both sides are symmetric covariant tensor fields over $\Sigma$. The important fact to be observed here is that $\alpha$ appears without differentiation. This means that $H(\alpha,0)$ is an ultralocal functional of $\pi$, which is further assumed to be a polynomial. (Note that we do not assume any relation between $\pi$ and $K$ at this point). Quite generally, we wish to stress the importance of such ’semantic’ assumptions concerning the intended meanings of symmetry operations when it comes to ‘derivations’ of physical laws from ‘symmetry principles’. Such derivations often suffer from the same sort of overstatement that tends to give the impression that the mere requirement that some group $G$ acts as symmetries alone distinguishes some dynamical laws from others. Often, however, additional assumptions are made that severely restrict the form in which $G$ is allowed to act. For example, in field theory, the requirement of locality often enters decisively, like in the statement that Maxwell’s vacuum equations are Poincaré- but not Galilei invariant. In fact, without locality the Galilei group, too, is a symmetry group of vacuum electrodynamics [@Fushchich.Shtelen:1991]. Coming back to the case at hand, I do not know of a uniqueness result that does not make the assumptions concerning the spacetime interpretation of the generators $H(\alpha,\beta)$. Compare also the related discussion in [@Samuel:2000b; @Samuel:2000a]. The uniqueness result for Einstein’s equation, which in its space-time form is spelled out in Lovelock’s theorem [@Lovelock:1972] already mentioned above, now takes the following form in Geometrodynamics [@Kuchar:1973]: \[thm:Uniqueness\] In four spacetime dimensions (Lorentzian for $\sigma=1$, Euclidean for $\sigma=-1$), the most general functional (\[eq:NormalDefCorr\]) satisfying (\[eq:H(ab)Comm\]) with (\[eq:X(ab)CommValues\]), subject to the conditions discussed above, is given by (\[eq:DiffConstrFunct\]) and the two-parameter $(\kappa,\Lambda)$ family $$\label{eq:HamConstrFunct} {\mathcal{H}}[h,\pi]=\sigma\,(2\kappa)\, G_{ab\,cd}\pi^{ab}\pi^{cd} -(2\kappa)^{-1}\sqrt{\det(h)}\bigl(R(h)-\Lambda\big)\,,$$ where $$\label{eq:DeWittMetricMomenta} G_{ab\,cd}=\tfrac{1}{2\sqrt{\det(h)}}\bigl( h_{ac}h_{bd}+h_{ad}h_{bc}-\tfrac{1}{2}h_{ab}h_{cd}\bigr)\,,$$ and $R(h)$ is the Ricci scalar of $(h,\Sigma)$. Note that (\[eq:DeWittMetricMomenta\]) is just the “contravariant version” of the metric (\[eq:WDWmetric-c\]) for $\lambda=1$, i.e., $G_{ab\,nm}G^{nm\,cd}=\frac{1}{2}(\delta^c_a\delta^d_b+\delta^d_a\delta^c_b)$. The Hamiltonian evolution so obtained is precisely that of General Relativity (without matter) with gravitational constant $\kappa=8\pi G/c^4$ and cosmological constant $\Lambda$. The proof of the theorem is given in [@Kuchar:1973], which improves on earlier versions [@Teitelboim:1973; @Hojman.etal:1976] in that the latter assumes in addition that ${\mathcal{H}}[h,\pi]$ be an even function of $\pi$, corresponding to the requirement of time reversibility of the generated evolution. This was overcome in [@Kuchar:1973] by the clever move to write the condition set by $\{H(\alpha_1,0),H(\alpha_2,0)\}=H(0,\beta')$ (the right-hand side being already known) on $H(\alpha,0)$ in terms of the corresponding Lagrangian functional $L$, which is then immediately seen to turn into a condition which is *linear* in $L$, so that terms with even powers in velocity decouple form those with odd powers. However, a small topological subtlety remains that is neglected in all these references and which potentially introduces a little more ambiguity that that encoded in the two parameters $\kappa$ and $\Lambda$, though its significance is more in the quantum theory. To see this recall that we can always perform a canonical transformation of the form $$\label{eq:CanTrans} \pi\mapsto \pi':=\pi+\Theta$$ where $\Theta$ is a closed one-form on ${\mathrm{Riem}(\Sigma)}$. The latter condition ensures that all Poisson brackets remain the same if $\pi$ is replaced with $\pi'$. Since ${\mathrm{Riem}(\Sigma)}$ is an open positive convex cone in a vector space and hence contractible, it is immediate that $\Theta=d\theta$ for some function $\theta:{\mathrm{Riem}(\Sigma)}\rightarrow{\mathbb{R}}$. However, $\pi$ and $\pi'$ must satisfy the diffeomorphism constraint, which is equivalent to saying that the kernel of $\pi$ (considered as one-form on ${\mathrm{Riem}(\Sigma)}$) contains the vertical vector fields, which implies that $\Theta$, too, must annihilate all $V_\xi$ so that $\theta$ is constant on each connected component of the ${\mathrm{Diff_F}(\Sigma)}$ orbit in ${\mathrm{Riem}(\Sigma)}$. But unless the ${\mathrm{Diff_F}(\Sigma)}$ orbits in ${\mathrm{Riem}(\Sigma)}$ are connected, this does not mean that $\theta$ is the pull back of a function on Superspace, as assumed in [@Kuchar:1973]. We can only conclude that $\Theta$ is the pull back of a closed but not necessarily exact one-form on Superspace. Hence there is an analogue of the Bohm-Aharonov-like ambiguity that one always encounters if the configuration space is not simply connected. Whether this is the case depends in a determinate fashion on the topology of $\Sigma$: One has, due to the contractibility of ${\mathrm{Riem}(\Sigma)}$, $$\label{eq:SuperspaceHomotopyGroups} \pi_n\bigl({\mathrm{Riem}(\Sigma)}/{\mathrm{Diff_F}(\Sigma)}\bigr)\cong\pi_{n-1}\bigl({\mathrm{Diff_F}(\Sigma)}\bigr) \qquad (n\geq 1)\,.$$ For $n=1$ the right hand side is $$\label{eq:SuperspaceFundGroup} \pi_{0}\bigl({\mathrm{Diff_F}(\Sigma)}\bigr):={\mathrm{Diff_F}(\Sigma)}/{\mathrm{Diff}_{\mathrm{F}}^0(\Sigma)}=:{\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$$ where ${\mathrm{Diff}_{\mathrm{F}}^0(\Sigma)}$ is the identity component of ${\mathrm{Diff_F}(\Sigma)}$ and where we introduced the name ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$ (Mapping-Class Group for Frame fixing diffeomorphisms) for the quotient group of components. In view of the uniqueness result above, one might wonder what goes wrong when using (the contravariant version of) the metric $G_\lambda$ for $\lambda\ne 1$ in (\[eq:DeWittMetricMomenta\]). The answer is that it would spoil (\[eq:H(ab)Comm\]). More precisely, it would contradict $\{H(\alpha_1,0)\,,\,H(\alpha_2,0)\} =H(0,\alpha_1\nabla\alpha_2-\alpha_2\nabla\alpha_1)$ due to an extra term $\propto (h_{ab}\pi^{ab})^2$ in ${\mathcal{H}}[h,\pi]$, unless the additional constraint $h_{ab}\pi^{ab}=0$ were imposed, which is equivalent to ${\mathrm{Tr}}_h(K)=0$ and hence to the condition that only maximal slices are allowed [@Giulini:1995c]. But this is clearly unacceptable (cf. Sect.6 of [@Barbour.etal:2002]). As a final comment about uniqueness of representations of (\[eq:H(ab)Comm\]) we mention the apparently larger ambiguity—labelled by an additional ${\mathbb{C}}$-valued parameter, the *Barbero-Immirzi parameter*—that one gets if one uses connection variables rather than metric variables (cf. [@Barbero:1995; @Immirzi:1997], Sect.4.2.2 of [@Thiemann:MCQGR], and Sect.4.3.1 of [@Kiefer:QuantumGravity]). However, in this case one does not represent (\[eq:H(ab)Comm\]) but a semi-direct product of it with the Lie algebra of $SU(2)$ gauge transformations, so that after taking the quotient with respect to the latter (which form an ideal) our original (\[eq:H(ab)Comm\]) is represented non locally. Also, unless the Barbero-Immirzi parameter takes the very special values $\pm i$ (for Lorentzian signature; $\pm 1$ for Euclidean signature) the connection variable does not admit an interpretation as a space-time gauge field restricted to spacelike hypersurfaces (cf. [@Immirzi:1997; @Samuel:2000a]). For example, the holonomy of a spacelike curve $\gamma$ varies with the choice of the spacelike hypersurface containing $\gamma$, which would be impossible if the spatial connection were the restriction of a space-time connection [@Samuel:2000a]. Accordingly, the dynamics generated by the constraints does then not admit the interpretation of being induced by appropriately moving a hypersurface through a spacetime with fixed geometric structures on it. Consequently, the argument provided here for why one should require (\[eq:H(ab)Comm\]) in the first place does, strictly speaking, not seem to apply in case of connection variables. It is therefore presently unclear to me on what set of assumptions a uniqueness result could be based in this case. Topology of configuration space {#sec:TopologySup} =============================== Much of the global topology of ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ is encoded in its homotopy groups, which, in turn are given by those of ${\mathrm{Diff_F}(\Sigma)}$ according to (\[eq:SuperspaceHomotopyGroups\]). Their structures were investigated in [@Witt:1986b; @Giulini:1995a; @Giulini:1997a]. Early references as to their possible relevance in quantum gravity are [@Friedman.Sorkin:1980; @Isham:1982; @Sorkin:1986; @Sorkin:1989]. We start by remarking that topological invariants of ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ are also topological invariants of $\Sigma$, which need not be homotopy invariant of $\Sigma$ even if they are homotopy invariants of ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$. This is, e.g., the case for the mapping-class group of homeomorphisms [@McCarty:1963] and hence (in 3 dimensions) also for the mapping-class group ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$. Remarkably, this means that we may distinguish homotopy equivalent but non homeomorphic 3-manifolds by looking at homotopy invariants of their associated Superspaces. Examples for this are given by certain types of lens spaces. First recall the definition of lens spaces $L(p,q)$ in 3 dimensions: $L(p,q)=S^3/\!\!\sim$, where $(p,q)$ is a pair of positive coprime integers with $p>1$, $S^3=\{(z_1,z_2)\in{\mathbb{C}}^2\mid\vert z_1\vert^2+\vert z_2\vert^2=1\}$, and $(z_1,z_2)\sim (z'_1,z'_2)\Leftrightarrow z'_1=\exp(2\pi i/p)z_1$, and $z'_2=\exp(2\pi i\,q/p)z_2$. One way to picture them is to take a solid ball in ${\mathbb{R}}^3$ and identify each point on the upper hemisphere with a points on the lower hemisphere after a rotation by $2\pi q/p$ about the vertical symmetry axis. (Usually one depicts the ball in a way in which it is slightly squashed along the vertical axis so that the equator develops a sharp edge and the whole body looks like a lens; see e.g. Fig.in [@Seifert.Threlfall:Topology].) In this way each set of $p$ equidistant points on the equator is identified to a single point. The fundamental group of $L(p,q)$ is ${\mathbb{Z}}_p$, independent of $q$, and the higher homotopy groups are those of its universal cover, $S^3$. Moreover, for connected closed orientable 3-manifolds the homology and cohomology groups are also determined by the fundamental group in an easy fashion: If $A$ denotes the operation of abelianisation of a group, $F$ the operation of taking the free part of a finitely generated abelian group, then the first four (zeroth to third, the only non-trivial ones) homology and cohomology groups are respectively given by $H_*=({\mathbb{Z}},A\pi_1,FA\pi_1,{\mathbb{Z}})$ and $H^*=({\mathbb{Z}},FA\pi_1,A\pi_1,{\mathbb{Z}})$ respectively. Hence, if taken of $L(p,q)$, all these standard invariants are sensitive only to $p$. However, it is known that $L(p,q)$ and $L(p,q')$ are 1. homotopy equivalent iff $qq'=\pm n^2\,(\mathrm{mod}\,p)$ for some integer $n$, 2. homeomorphic iff (all four possibilities) $q'=\pm q^{\pm 1}\,(\mathrm{mod}\,p)$, and 3. orientation-preserving homeomorphic iff $q'=q^{\pm 1}\,(\mathrm{mod}\,p)$. The first statement is Theorem10 in [@Whitehead:1941] and the second and third statement follow, e.g., from the like combinatorial classification of lens spaces [@Reidemeister:1935] together with the validity of the ‘Hauptvermutung’ (the equivalence of the combinatorial and topological classifications) in 3 dimensions [@Moise:1952]. So, for example, $L(15,1)$ is homotopy equivalent but not homeomorphic to $L(15,4)$. On the other hand, it is known that the mapping-class group ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$ for $L(p,q)$ is ${\mathbb{Z}}\times{\mathbb{Z}}$ if $q^2=1\,(\mathrm{mod}\,p)$ with $q\not =\pm 1\,(\mathrm{mod}\,p)$, which applies to $p=15$ and $q=4$, and that in the remaining cases for $p>2$ it is just ${\mathbb{Z}}$ (see TableIV on p.591 of [@Witt:1986b]). Hence ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}\cong{\mathbb{Z}}\times{\mathbb{Z}}$ for $\Sigma=L(15,4)$ and ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}\cong{\mathbb{Z}}$ for $\Sigma=L(15,1)$, even though $L(15,1)$ and $L(15,4)$ are homotopy equivalent! Quite generally it turns out that Superspace stores much information about the topology of the underlying 3-manifold $\Sigma$. This can be seen from the table in Figure\[Fig:3ManifoldsTable\], which we reproduced form [@Giulini:1994a], and where properties of certain prime manifolds (see below for an explanation of ‘prime’) are listed. There is one interesting observation from that list which we shall mention right away: From gauge theories it is known that there is a relation between topological invariants of the classical configuration space and certain features of the corresponding quantum-field theory [@Jackiw:LesHouches1983], in particular the emergence of certain anomalies which represent non-trivial topological invariants [@Alvarez-GaumeGinsparg:1985]. By analogy one could conjecture similar relations to hold quantum gravity. An interesting question is then whether there are preferred manifolds $\Sigma$ for which all these invariants are trivial. From those represented on the table there is indeed a unique pair of manifolds for which this is the case, namely the 3-sphere and the 3-dimensional real projective space. To understand more of the information collected in the table we have to say more about general 3-manifolds. Of particular interest is the fundamental group of Superspace. Experience with ordinary quantum mechanics (cf. [@Giulini:1995b] and references therein) already suggests that its classes of inequivalent irreducible unitary representations correspond to a superselection structure which here might serve as fingerprint of the topology of $\Sigma$ in the quantum theory. The sectors might, e.g., correspond to various statistics (in the presence of diffeomorphic primes) that preserve or violate a naively expected spin-statistics correlation [@Aneziris.etal:1989a; @Aneziris.etal:1989b; @Dowker.Sorkin:1998; @Dowker.Sorkin:2000] (see also below). General three-manifolds and specific examples {#sec:GeneralThreeManifolds} --------------------------------------------- ![\[Fig:3ManifoldsTable\] This table, taken from [@Giulini:1994a], lists various properties of certain prime 3-manifolds. The manifolds are grouped into those of finite fundamental group, which are of the form $S^3/G$, the exceptional one, $S^1\times S^2$, which is prime but not irreducible ($\pi_2(S^1\times S^2)=\mathbb{Z}$), those six which can carry a flat metric and which are of the form $\mathbb{R}^3/G$, and so-called Haken manifolds (sufficiently large $K(\pi,1)$ primes). For the lens spaces $q_1$ stands for $q=\pm 1$, $q_2$ for $q\ne\pm 1$ and $q^2=1$, $q_3$ for $q^2=-1$, and $q_4$ for the remaining cases, where all equalities are taken $\mathrm{mod}\,p$. The 3rd and 4th column list spinoriality and chirality, the last three columns the homotopy groups of their corresponding superspace. We refer to [@Giulini:1994a] for the meanings of the other columns.](Table3Manifolds){width="\linewidth"} The way to understand general 3-manifolds is by cutting them along certain embedded two manifolds so that the remaining pieces are simpler in an appropriate sense. Here we shall only consider those simplifications that are achieved by cutting along embedded 2-spheres. (Further decompositions by cutting along 2-tori provide further simplifications, but these are not directly relevant here.) The 2-spheres should be ‘essential’ and ‘splitting’. An essential 2-sphere is one which does not bound a 3-ball and a splitting 2-sphere is one whose complement has two (rather than just one) connected components. Figure\[fig:FigRiemannSurface\] is intended to visualise the analogues of these notions in two dimensions. ![\[fig:FigRiemannSurface\] A Riemann surface of genus3 with three pairs of embedded 1-spheres (circles) of type A, B, and C. Type A is essential and splitting, type B is essential but not splitting, and type C is splitting but not essential. Any third essential and splitting 1-sphere can be continuously deformed via embeddings into one of the two drawn here.](FigRiemannSurface){width="0.8\linewidth"} Given a closed 3-manifold $\Sigma$, consider the following process: Cut it along an essential splitting 2-sphere and cap off the 2-sphere boundary of each remaining component by a 3-disk. Now repeat the process for each of the remaining closed 3-manifolds. This process stops after a finite number of steps [@Kneser:1929] where the resulting components are uniquely determined up to diffeomorphisms (orientation preserving if oriented manifolds are considered) and permutation [@Milnor:1962]; see [@Hatcher:3-manifolds] for a lucid discussion. The process stops at that stage at which none of the remaining components, $\Pi_1,\cdots,\Pi_n$, allows for essential splitting 2-spheres, i.e. at which each $\Pi_i$ is a prime manifold. A 3-manifold is called *prime* if each embedded 2-sphere either bounds a 3-disc or does not split; it is called *irreducible* if each embedded 2-sphere bounds a 3-disc. In the latter case the second homotopy group, $\pi_2$, must be trivial, since, if it were not, the so-called sphere theorem (see, e.g., [@Hatcher:3-manifolds]) ensured the existence of a non-trivial element of $\pi_2$ which could be represented by an *embedded* 2-sphere. Conversely, it follows from the validity of the Poincaré conjecture that a trivial $\pi_2$ implies irreducibility. Hence irreducibility is equivalent to a trivial $\pi_2$. There is precisely one non-irreducible prime 3-manifold, and that is the handle $S^1\times S^2$. Hence a 3-manifold is prime iff it is either a handle or if its $\pi_2$ is trivial. Given a general 3-manifold $\Sigma$ as connected sum of primes $\Pi_1,\cdots,\Pi_n$, there is a general method to establish ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$ in terms of the individual mapping-class groups of the primes. The strategy is to look at the effect of elements in ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$ on the fundamental group of $\Sigma$. As $\Sigma$ is the connected sum of primes, and as connected sums in $d$ dimensions are taken along $d-1$ spheres which are simply-connected for $d\geq 3$, the fundamental group of a connected sum is the free product of the fundamental groups of the primes for $d\geq 3$. The group ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$ now naturally acts as automorphisms of $\pi(\Sigma)$ by simply taking the image of a based loop that represents an element in $\pi(\Sigma)$ by a based (same basepoint) diffeomorphism that represents the class in ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$. Hence there is a natural map $$\label{eq:MCGmapstoAut} d_F:{\mathrm{MCG}_{\mathrm{F}}(\Sigma)}\rightarrow\mathrm{Aut}\bigl(\pi_1(\Sigma)\bigr)\,.$$ The known presentations[^6] of automorphism groups of free products in terms of presentations of the automorphisms of the individual factors and additional generators (basically exchanging isomorphic factors and conjugating whole factors by individual elements of others) can now be used to establish (finite) presentations of ${\mathrm{MCG}_{\mathrm{F}}(\Sigma)}$, provided (finite) presentations for all prime factors are known.[^7] Here I wish to stress that this situation would be more complicated if ${\mathrm{Diff}(\Sigma)}$ rather than ${\mathrm{Diff_F}(\Sigma)}$ (or at least the diffeomorphisms fixing a preferred point) had been considered; that is, had we not made the transition from (\[eq:DefSuperspace\]) to (\[eq:DefSupF\]). Only for ${\mathrm{Diff_F}(\Sigma)}$ (or the slightly larger group of diffeomorphisms fixing the point) is it generally true that the mapping-class group of a prime factor injects into the mapping-class group of the connected sum in which it appears. For more on this, compare the discussion on p.182-3 in [@Giulini:2007a]. Clearly, one also needs to know which elements are in the kernel of the map (\[eq:MCGmapstoAut\]). This will be commented on below in connection with Fig \[fig:FigDehnTwists\]. ![\[fig:TwoRP3\]The connected sum of two real projective spaces may be visualised by the shaded region obtained from the spherical shell that is obtained by rotating the shaded annulus about the vertical symmetry axis as indicated. Antipodal points on the outer 2-sphere boundary $S_1$, as well as on the inner 2-sphere boundary $S_2$, are pairwise identified. This results in the connected sum of two ${\mathbb{R}}P^3$ along the connecting sphere $S$. Due to the antipodal identifications, the two thick horizontal segments in the shaded region become a single loop, showing that the entire space is fibred by circles over ${\mathbb{R}}P^2$. Remarkably, the handle $S^1\times S^2$, which is prime, is a double cover of this reducible manifold.](FigTwoRP3 "fig:"){width="0.64\linewidth"} (-57,134)[$S_1$]{} (-116,71)[$S_2$]{} (-78,111)[$S$]{} ![\[fig:GenFundGrTwoRP3\]The shown loops represent the generators $a$ and $b$ of the fundamental group ${\mathbb{Z}}_2*{\mathbb{Z}}_2=\langle a,b\mid a^2=b^2=1\rangle$. The product loop $c=ab$ is then seen to be one of the fibres mentioned in the caption of the previous Figure. In terms of $a$ and $c$ we have the presentation $\langle a,c\mid a^2=1,\, aca^{-1}=c\rangle$, showing the isomorphicity ${\mathbb{Z}}_2*{\mathbb{Z}}_2\cong{\mathbb{Z}}_2\ltimes{\mathbb{Z}}$.](FigRP3x2-1 "fig:"){width="0.70\linewidth"} (-58,75)[$a$]{} ![\[fig:GenFundGrTwoRP3\]The shown loops represent the generators $a$ and $b$ of the fundamental group ${\mathbb{Z}}_2*{\mathbb{Z}}_2=\langle a,b\mid a^2=b^2=1\rangle$. The product loop $c=ab$ is then seen to be one of the fibres mentioned in the caption of the previous Figure. In terms of $a$ and $c$ we have the presentation $\langle a,c\mid a^2=1,\, aca^{-1}=c\rangle$, showing the isomorphicity ${\mathbb{Z}}_2*{\mathbb{Z}}_2\cong{\mathbb{Z}}_2\ltimes{\mathbb{Z}}$.](FigRP3x2-2 "fig:"){width="0.70\linewidth"} (-58,75)[$b$]{} The connected sum of two real-projective spaces ----------------------------------------------- In some (in fact many) cases the map $d_F$ is an isomorphism. For example, this is the case if $\Sigma$ is the connected sum of two ${\mathbb{R}}P^3$, so that $\pi_1(\Sigma)$ is the free product ${\mathbb{Z}}_2*{\mathbb{Z}}_2$, a presentation of which is $\langle a,b\mid a^2=b^2=1\rangle$. For the automorphisms we have $\mathrm{Aut}({\mathbb{Z}}_2*{\mathbb{Z}}_2)\cong {\mathbb{Z}}_2*{\mathbb{Z}}_2=\langle E,S\mid E^2=S^2=1\rangle$, where $E:(a,b)\rightarrow (b,a)$ and $S:(a,b)\rightarrow (a,aba^{-1})$. In this sense the infinite discrete group ${\mathbb{Z}}_2*{\mathbb{Z}}_2$ is a quotient of the automorphism group of Superspace ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$ for $\Sigma$ being the connected sum of two real projective spaces. It is therefore of interest to study its unitary irreducible representations. This can be done directly in a rather elementary fashion, or more systematically by a simple application of the method of induced representations (Mackey theory) using the isomorphicity ${\mathbb{Z}}_2*{\mathbb{Z}}_2\cong{\mathbb{Z}}_2\ltimes{\mathbb{Z}}$ (cf. caption to Fig.\[fig:GenFundGrTwoRP3\]). The result is that, apart form the obvious four one-dimensional ones, given by $(E,S)\mapsto (\pm{\mathrm{id}},\pm{\mathrm{id}})$, there is a continuous set of mutually inequivalent two-dimensional ones, given by $$\label{eq:MCGRepTwoRP3s} E\mapsto \left( \begin{array}{lr} 1&0\\ 0&-1 \end{array}\right)\,,\qquad S\mapsto \left( \begin{array}{lr} \cos\tau & \sin\tau\\ \sin\tau & -\cos\tau\\ \end{array}\right)\,,\quad \tau\in(0,\pi)\,.$$ Already in this most simple example of a non-trivial connected sum we have an interesting structure in which the two ‘statistics sectors’ corresponding to the irreducible representations of the permutation subgroup (here just given by the ${\mathbb{Z}}_2$ subgroup generated by $E$) get mixed by $S$, where the ‘mixing angle’ $\tau$ uniquely characterises the representation. This behaviour can also be studied in more complicated examples [@Giulini:1994a][@SorkinSurya:1998]. For a more geometric understanding of the maps representing $E$ and $S$, see [@Giulini:2007a]. ![\[fig:FigDehnTwists\]Both pictures show rotations parallel to spheres $S_1$ and $S_2$: On the left, a rotation of a prime manifold in a connected sum parallel to the connecting sphere, on the right a rotation parallel to two meridian spheres in a ‘handle’ $S^1\times S^2$. The support of the diffeomorphism is on the cylinder bound by $S_1$ and $S_2$. In either case its effect is depicted by the two curves connecting the two spheres. The two-dimensional representation given here is deceptive insofar, as in two dimensions the original and the mapped curves are not homotopic (keeping their endpoints fixed), due to the one-sphere not being simply connected, whereas they are in three (and higher) higher dimensions they are due to the higher-dimensional spheres being simply connected. ](FigDehnTwist1){width="0.85\linewidth"} ![\[fig:FigDehnTwists\]Both pictures show rotations parallel to spheres $S_1$ and $S_2$: On the left, a rotation of a prime manifold in a connected sum parallel to the connecting sphere, on the right a rotation parallel to two meridian spheres in a ‘handle’ $S^1\times S^2$. The support of the diffeomorphism is on the cylinder bound by $S_1$ and $S_2$. In either case its effect is depicted by the two curves connecting the two spheres. The two-dimensional representation given here is deceptive insofar, as in two dimensions the original and the mapped curves are not homotopic (keeping their endpoints fixed), due to the one-sphere not being simply connected, whereas they are in three (and higher) higher dimensions they are due to the higher-dimensional spheres being simply connected. ](FigDehnTwist2){width="0.85\linewidth"} (-50,12)[$S_1$]{} (-130,12)[$S_2$]{} (-290,50)[$\Pi$]{} (-290,-21)[$S_1$]{} (-290,-64)[$S_2$]{} Spinoriality ------------ I also wish to mention one very surprising observation that was made by Rafael Sorkin and John Friedman in 1980 [@Friedman.Sorkin:1980] and which has to do with the physical interpretation of the elements in the kernel of the map (\[eq:MCGmapstoAut\]), leading to the conclusion that pure (i.e. without matter) quantum gravity should already contain states with half-integer angular momenta. The reason being a purely topological one, depending entirely on the topology of $\Sigma$. In fact, given the right topology of $\Sigma$, its one-point decompactification used in the context of asymptotically flat initial data will describe an isolated system whose asymptotic (at spacelike infinity) symmetry group is not the ordinary Poincaré group [@Beig.Murchadha:1987] but rather its double (= universal) cover. This gives an intriguing answer to Wheeler’s quest to find a natural place for spin1/2 in Einstein’s standard geometrodynamics (cf. [@Misner.Thorne.Wheeler:Gravitation] Box44.3). I briefly recall that after introducing the concept of a ‘Geon’ (‘gravitational-electromagnetic entity’) in 1955 [@Wheeler:1955], and inspired by the observation that electric charge (in the sense of non-vanishing flux integrals of $\star F$ over closed 2-dimensional surfaces) could be realised in Einstein-Maxwell theory without sources (‘charge without charge’), Wheeler and collaborators turned to the Einstein-Weyl theory [@Brill.Wheeler:1957] and tried to find a ‘neutrino analog of electric charge’ [@Klauder.Wheeler:1957]. Though this last attempt failed, the programme of ‘matter as geometry’ in the context of geometrodynamics, as outlined in the contributions to the anthology [@Wheeler:Geometrodynamics], survived in Wheeler’s thinking well into the 1980s [@Wheeler:1982]. Back to the ‘spin without spin’ topologies, the elements of the kernel of (\[eq:MCGmapstoAut\]) can be pictured as rotation parallel to certain spheres, as depicted in Figure\[fig:FigDehnTwists\]. (In many—and possibly all—cases the group generated by such maps actually exhaust the kernel; compare Theorem.1.5 in [@McCullough:1990] and footnote21 in [@Giulini:2007a]). The point we wish to focus on here is that for *some* prime manifolds the diffeomorphism depicted on the left in Figure\[fig:FigDehnTwists\] is indeed not in the identity component of all diffeomorphisms that fix a frame exterior to the outer ($S_2$) 2-sphere. Such manifolds are called *spinorial*. For each prime it is known whether it is spinorial or not, and the easy-to-state but hard-to-prove result is, that the only non-spinorial manifolds[^8] are the lens spaces $L(p,q)$, the handle $S^1\times S^2$, and connected sums amongst them. That these manifolds are not spinorial is, in fact, very easy to visualise. Hence, given the proof of the ‘only’ part and of the fact that a connected sum is spinorial iff it contains at least one spinorial prime, one may summarise the situation by saying that that the only non-spinorial manifolds are the ‘obvious’ ones. Even though being a generic property in the sense just stated, spinoriality is generally hard to prove in dimensions three or greater. This is in marked contrast to two dimensions, where the corresponding transformation shown in the left picture of Figure\[fig:FigDehnTwists\] acts non trivially on the fundamental group. Indeed, consider a base point outside (below) $S_2$ in the left picture in Figure\[fig:FigDehnTwists\], then the rotation acts by conjugating each of the $2g$ generators $(a_1\cdots a_g,b_1\cdots b_2)$ that $\Pi$ adds to $\pi_1(\Sigma)$ by the element $\prod_{i=1}^g a_ib_ia^{-1}_ib^{-1}_i$, which is non-trivial in $\pi_1(\Sigma)$ if other primes exist or otherwise if the point with the fixed frame is removed (as one may do, due to the restriction to diffeomorphisms fixing that point). An example of a spinorial manifold is the spherical space form $S^3/D^*_8$, where $S^3$ is thought of as the sphere of unit quaternions and $D_8^*$ is the subgroup in the group of unit quaternions given by the eight elements $\{\pm 1,\pm i, \pm j, \pm k\}$. The coset space $S^3/D^*_8$ may be visualised as solid cube whose opposite faces are identified after a $90$-degree rotation by either a right- or a left-handed screw motion; see Figure\[fig:Q-Space\]. Drawings of fundamental domains in form of (partially truncated) solid polyhedra with suitable boundary identifications for spaces $S^3/G$ are given in [@DuVal:Homographies]. ![\[fig:Q-Space\]Fundamental domain for the space $S^3/D^*_8$. Opposite faces are identified after a right-handed screw motion with $90$-degree rotation, as indicated by the coinciding labels for the edges and vertices. The corresponding space with a left-handed identification is not orientation-preserving diffeomorphic to this one.](FigQ-Space){width="0.4\linewidth"} Let us take the basepoint $\infty\in S^3/D^*_8$ to be the centre of the cube in Figure\[fig:Q-Space\]. The two generators, $a$ and $b$, of the fundamental group $$\label{eq:FundGrPresQSpace} \pi_1(S^3/D^*_8)\cong D^*_8=\langle a,b\mid a^2=b^2=(ab)^2\rangle$$ are then represented by two of the three oriented straight segments connecting the midpoints of opposite faces. The third corresponds to the product $ab$. A rotation of the cube about its centre by any element of its crystallographic symmetry group defines a diffeomorphism of $S^3/D^*_8$ fixing $\infty$ since it is compatible with the boundary identification. It is not in the identity component of $\infty$-fixing diffeomorphisms since it obviously acts non-trivially on the generators of the fundamental group. Clearly, each such rigid rotation may be modified in an arbitrarily small neighbourhood of $\infty$ so as to also fix the tangent space at this point. That $S^3/D^*_8$ is spinorial means that in going from the point-fixing to the frame-fixing diffeomorphisms one acquires more diffeomorphisms not connected to the identity. More precisely, the mapping-class group of frame-fixing diffeomorphisms is a ${\mathbb{Z}}_2$-extension of the mapping-class group of merely point-fixing diffeomorphisms. The generator of this extending ${\mathbb{Z}}_2$ is a full $360$-degree rotation parallel to two small concentric spheres centred at $\infty$. In this way, the spinoriality of $S^3/D^*_8$ extends the crystallographic symmetry group $O\subset SO(3)$ of the cube to its double cover $O^*\subset SU(2)$ and one finally gets $$\label{eq:MCGofQSpace} \mathrm{MCG}_{\mathrm{F}}(\Pi)\cong O^*\quad\mathrm{for}\quad \Pi=S^3/D^*_8\,.$$ This is precisely what one finds at the intersection of the 2nd row and 7th column of the table in Figure\[Fig:3ManifoldsTable\], with corresponding results for the other spherical space forms. Coming back to the previous example of the connected sum of two (or more [@Giulini:1994a]) real projective spaces, we can make the following observation: First of all, real projective 3-space is a non-spinorial prime. This is obvious once one visualises it as a solid 3-ball whose 2-sphere boundary points are pairwise identified in an antipodal fashion, since this identification is compatible with a rigid rotation. A full rotation about the, say, centre-point of the ball may therefore be continuously undone by a rigid rotation outside a small ball about the centre, suitably ‘bumped off’ towards the centre. Second, as we have seen above, the irreducible representations of the mapping-class group of the connected sum contains both statistics sectors independently for one-dimensional representations and in a mixed form for the continuum of two-dimensional irreducible representations. This already shows [@Aneziris.etal:1989b] that there is no general kinematical spin-statistics relation as in other non-linear theories [@Finkelstein.Rubinstein:1968; @Sorkin:1988]. Such a relation may at best be re-introduced for some manifolds by restricting the way in which states are constructed, e.g., via the sum-over-histories approach [@Dowker.Sorkin:1998]. Chirality --------- There is one last aspect about diffeomorphisms that can be explained in terms of Figure\[fig:Q-Space\]. As stated in the caption of this figure, there are two versions of this space: one where the identification of opposite faces is done via a $90$-degree right-handed screw motion and one where one uses a left-handed screw motion. These spaces are not related by an orientation preserving diffeomorphism. This is equivalent to saying that, say, the first of these spaces has no orientation-reversing self-diffeomorphism. Manifolds for which this is the case are called *chiral*. There are no examples in two dimensions. To see this, just consider the usual picture of a Riemannian genus $g$ surface embedded into ${\mathbb{R}}^3$ and map it onto itself by a reflection at any of its planes of symmetry. So chiral manifolds start to exist in 3 dimensions and continue to do so in all higher dimensions, as was just recently shown [@Muellner:2008]. If one tries to reflect the cube in Figure\[fig:Q-Space\] at one of its symmetry planes one finds that this is incompatible with the boundary identifications, that is, pairs of identified points are not mapped to pairs of identified points. Hence this reflection simply does not define a map of the quotient space. This clearly does not prove the nonexistence of orientation reversing maps, since there could be others than these obvious candidates. In fact, following an idea of proof in [@Witt:1986b], the chirality of $S^G/D^*_8$ (and others of the form $S^3/G$) can be reduced to that of $L(4,1)$. That the latter is chiral follows from the following argument: Above we have already stated that $L(p,q)$ and $L(p,q')$ are orientation preserving diffeomorphic iff $q'=q^{\pm 1}\,(\mathrm{mod}\,p)$. Since taking the mirror image in ${\mathbb{R}}^3$ of the lens representing $L(p,q)$ gives the lens representing $L(p,-q)$, $L(p,q)$ admits an orientation reversing diffeomorphism iff $L(p,q)$ and $L(p,-q)$ are orientation preserving diffeomorphic. But, as just stated, this is the case iff $-q=q^{\pm 1}\,(\mathrm{mod}\,p)$, i.e. if either $p=2,\,q=1$ (recall that $p$ and $q$ must be coprime) or $q^2=-1\,(\mathrm{mod}\,p)$.[^9] Hence, in particular, all $L(p,1)$ are chiral. Now, $G=D_8^*$ has three subgroups isomorphic to ${\mathbb{Z}}_4$, the fundamental group of $L(4,1)$, namely the ones generated by $i$, $j$, and $k$. They are normal so that we have a regular covering $L(4,1)\displaystyle{\mathop{\rightarrow}^p} S^3/G$. Now suppose $f: S^3/G\rightarrow S^3/G$ were orientation reversing, i.e. a diffeomorphism with $\deg(f)=-1$. Consider the diagram $$\label{eq:DiagChirality} \bfig \qtriangle|amr|/.>`>`>/<970,600>[L(4{,}1)`L(4{,}1)`S^3/G;\tilde f`f\circ p`p] \btriangle|lmb|/>`>`>/<970,600>[L(4{,}1)`S^3/G`S^3/G;p`f\circ p`f] \efig$$ If the lift $\tilde f$ existed we would immediately get a contradiction since from commutativity of (\[eq:DiagChirality\]) we would get $\deg(\tilde f)\cdot\deg(p)=\deg(p)\cdot\deg(f)$ and hence $\deg(\tilde f)=-1$, which contradicts chirality of $L(4,1)$. Now, according to the theory of covering spaces the lift $\tilde f$ of $f\circ p$ exists iff the image of $\pi_1\bigl(L(4,1)\bigr)$ under $(f\circ p)_*$, which is a subgroup ${\mathbb{Z}}_4\subset D_8^*$, is conjugate to the image of $\pi_1\bigl(L(4,1)\bigr)$ under $p_*$. This need not be the case, however, as different subgroups ${\mathbb{Z}}_4$ in $D_8^*$ are normal and hence never conjugate (here we deviate from the argument in [@Witt:1986b] which seems incorrect). However, by composing a given orientation reversing $f$ with an orientation preserving diffeomorphism that undoes the ${\mathbb{Z}}_4$ subgroup permutation introduced by $f$, we can always create a new orientation reversing diffeomorphism that does not permute the ${\mathbb{Z}}_4$ subgroups. That new orientation reversing diffeomorphism—call it again $f$—now indeed has a lift $\tilde f$, so that finally we arrive at the contradiction envisaged above. As prime manifolds, the two versions of a chiral prime corresponding to the two different orientations count as different. This means the following: Two connected sums which differ only insofar as a particular chiral prime enters with different orientations are not orientation-preserving diffeomorphic; they are not diffeomorphic at all if the complement of the selected chiral prime also chiral (i.e. iff another chiral prime exists). For example, the connected sum of two oriented $S^3/D_8^*$ is not diffeomorphic to the connected sum of $S^3/D_8^*$ with $\overline{S^3/D_8^*}$, where the overbar indicates the opposite orientation. Note that the latter case also leads to two non-homeomorphic 3-manifolds whose classic invariants (homotopy, homology, cohomology) coincide. This provides an example of a topological feature of $\Sigma$ that is not encoded into the structure of ${\mathcal{S}_{\mathrm{F}}(\Sigma)}$. Summary and outlook {#sec:SummaryOutlook} =================== Superspace is for geometrodynamics what gauge-orbit space is for non-abelian gauge theories, though Superspace has generally a much richer topological and metric structure. Its topological structure encodes much of the topology of the underlying 3-manifold and one may conjecture that some of its topological invariants bear the same relation to anomalies and sectorial structure as in the case of non-abelian gauge theories. Recent progress in 3-manifold theory now allows to make more complete statements, in particular concerning the fundamental groups of Superspaces associated to more complicated 3-manifolds. Its metric structure is piecewise nice but also suffers from singularities, corresponding to signature changes, whose physical significance is unclear. Even for simple 3-manifolds, like the 3-sphere, there are regions in superspace where the metric is strictly Lorentzian (just one negative signature and infinitely many pluses), like at the round 3-sphere used in the FLRW cosmological models, so that the Wheeler-DeWitt equation becomes strictly hyperbolic, but there are also regions with infinitely many negative signs in the signature. Note that the cotangent bundle over Superspace is not the fully reduced phase space for matter-free General Relativity. It only takes account of the vector constraints and leaves the scalar constraint unreduced. However, under certain conditions, the scalar constraints can be solved by the ‘conformal method’ which leaves only the conformal equivalence class of 3-dimensional geometries as physical configurations. In those cases the fully reduced phase space is the cotangent bundle over conformal superspace, whose analog to (\[eq:DefSuperspace\]) is given by replacing ${\mathrm{Diff}(\Sigma)}$ by the semi-direct product ${\mathrm{C}(\Sigma)}\rtimes{\mathrm{Diff}(\Sigma)}$, where ${\mathrm{C}(\Sigma)}$ is the abelian group of conformal rescalings that acts on ${\mathrm{Riem}(\Sigma)}$ via $(f,h)\mapsto f\,h$ (pointwise multiplication), where $f:\Sigma\rightarrow{\mathbb{R}}_+$. The right action of $(f,\phi)\in{\mathrm{C}(\Sigma)}\rtimes{\mathrm{Diff}(\Sigma)}$ on $h\in{\mathrm{Riem}(\Sigma)}$ is then given by by $R_{(f,\phi)}(h)=f\phi^*h$, so that, using $R_{(f_2,\phi_2)}R_{(f_1,\phi_1)}=R_{(f_1,\phi_1)(f_2,\phi_2)}$, the semi-direct product structure is seen to be $(f_1,\phi_1)(f_2,\phi_2)=\bigl(f_2(f_1\circ\phi_2),\phi_1\circ\phi_2\bigr)$. Note that because of $(f_1f_2)\circ\phi=(f_1\circ\phi)(f_2\circ\phi)$ ${\mathrm{Diff}(\Sigma)}$ indeed acts as automorphisms of ${\mathrm{C}(\Sigma)}$. Conformal superspace and extended conformal superspace would then, in analogy to (\[eq:DefSuperspace\]) and (\[eq:DefSupF\]), be defined as ${\mathcal{CS}(\Sigma)}:={\mathrm{Riem}(\Sigma)}/{\mathrm{C}(\Sigma)}\rtimes{\mathrm{Diff}(\Sigma)}$ and ${\mathcal{CS}_{\mathrm{F}}(\Sigma)}:={\mathrm{Riem}(\Sigma)}/{\mathrm{C}(\Sigma)}\rtimes{\mathrm{Diff_F}(\Sigma)}$ respectively. The first definition was used in [@Fischer.Moncrief:1996] as applied to manifolds with zero degree of symmetry (cf. footnote\[foot:DegSym\]). In any case, since ${\mathrm{C}(\Sigma)}$ is contractible, the topologies of ${\mathrm{C}(\Sigma)}\rtimes{\mathrm{Diff}(\Sigma)}$ and ${\mathrm{C}(\Sigma)}\rtimes{\mathrm{Diff_F}(\Sigma)}$ are those of ${\mathrm{Diff}(\Sigma)}$ and ${\mathrm{Diff_F}(\Sigma)}$ which also transcend to the quotient spaces analogously to (\[eq:SuperspaceHomotopyGroups\]) whenever the groups act freely. In the first case this is essentially achieved by restricting to manifolds of vanishing degree of symmetry, whereas in the second case this follows almost as before, with the sole exception being $(S^3,h)$ with $h$ conformal to the round metric.[^10] Hence the topological results obtained before also apply to this case. In contrast, the geometry for conformal superspace differs insofar from that discussed above as the conformal modes that formed the negative directions of the Wheeler-DeWitt metric (cf. (\[eq:WDWmetric-a\]) are now absent. The horizontal subspaces (orthogonal to the orbits of ${\mathrm{C}(\Sigma)}\rtimes{\mathrm{Diff_F}(\Sigma)}$) are now given by the transverse and traceless (rather than just obeying (\[eq:HorizontalityCond\])) symmetric two-tensors. In that sense the geometry of conformal superspace, if defined as before by some ultralocal bilinear form on ${\mathrm{Riem}(\Sigma)}$, is manifestly positive (due to the absence of trace terms) and hence less pathological than the superspace metric discussed above. It might seem that its physical significance is less clear, as there is now no constraint left that may be said to induce this particular geometry; see however [@Barbour.Murchadha:1999]. Whether it is a realistic hope to understand superspace and conformal superspace (its cotangent bundle being the space of solutions to Einstein’s equations) well enough to actually gain a sufficiently complete understanding of its automorphism group is hard to say. An interesting strategy lies in the attempt to understand the solution space directly in a group- (or Lie algebra-) theoretic fashion in terms of a quotient $G_\infty/H_\infty$, where $G_\infty$ is an infinite dimensional group (Lie algebra) that (locally) acts transitively on the space of solutions and $H_\infty$ is a suitable subgroup (algebra), usually the fixed-point set of an involutive automorphism of $G$. The basis for the hope that this might work in general is the fact that it works for the subset of stationary and axially symmetric solutions, where $G^\infty$ is the Geroch Group; cf. [@Breitenlohner.Maison:1987]. The idea for generalisation, even to $d=11$ supergravity, is expressed in [@Nicolai:1999] and further developed in [@Damour.etal:2007]. **Acknowledgements** I thank Hermann Nicolai and Stefan Theisen for inviting me to this stimulating 405th WE-Heraeus-Seminar on Quantum Gravity and for giving me the opportunity to contribute this paper. I also thank Hermann Nicolai for pointing out [@Nicolai:1999] and Ulrich Pinkall for pointing out [@Lelong-Ferrand:1971]. [^1]: “Isomorphic as what?” one may ask. The answer is: as ILH (inverse-limit Hilbert) manifolds. In the ILH sense the action of ${\mathrm{Diff}(\Sigma)}$ on ${\mathrm{Riem}(\Sigma)}\times{\mathrm{F}(\Sigma)}$ is smooth, free, and proper; see [@Fischer:1986] for more details and references. [^2]: The condition of ‘asymptotic flatness’ of an end includes the topological condition that the one-point compactification is again a manifold. This is the case iff there exists a compact subset in the manifold the complement of which is homeomorphic to the complement of a closed solid ball in ${\mathbb{R}}^3$. [^3]: \[foot:DegSym\]Let $\mathcal{I}(\Sigma,h):=\{\phi\in{\mathrm{Diff}(\Sigma)}\mid\phi^*h=h\}$ be the isometry group of $(\Sigma,h)$, then it is well known that $\dim\mathcal{I}(\Sigma,h)\leq\tfrac{1}{2}n(n+1)$, where $n=\dim \Sigma$. $\mathcal{I}(\Sigma,h)$ is compact if $\Sigma$ is compact (see, e.g., Sect.5 of [@Myers.Steenrod:1939]). Conversely, if $\Sigma$ allows for an effective action of a compact group $G$ then it clearly allows for a metric $h$ on which $G$ acts as isometries (just average any Riemannian metric over $G$.) The degree of symmetry of $\Sigma$, denoted by $\mathrm{deg}(\Sigma)$, is defined by $\mathrm{deg}(\Sigma):=\sup_{h\in{\mathrm{Riem}(\Sigma)}}\{\dim\mathcal{I}(\Sigma,h)\}$. For compact $\Sigma$ the degree of symmetry is zero iff $\Sigma$ cannot support an action of the circle group $SO(2)$. A list of 3-manifolds with $\mathrm{deg}>0$ can be found in [@Fischer:1970] whereas [@Fischer.Moncrief:1996] contains a characterisation of $\mathrm{deg}=0$ manifolds. [^4]: We follow the terminology of Appendix I in [@Besse:EinsteinManifolds]. [^5]: Due to the standard convention that the Hamiltonian action being defined as a *left* action, whereas the Lie bracket on a group is defined by the commutator of left-invariant vector fields which generate *right* translations. [^6]: A (finite) presentation of a group is its characterisation in terms of (finitely many) generators and (finitely many) relations. [^7]: This presentation of the automorphism group of free products is originally due to Fouxe-Rabinovitch [@Fouxe-Rabinovitch:1940; @Fouxe-Rabinovitch:1941]. Modern forms with corrections are given in [@McCullough.Miller:1986] and [@Gilbert:1987] [^8]: We remind the reader that ‘manifold’ here stands for ‘3-dimensional closed orientable manifold’. [^9]: Remarkably, this result was already stated in footnote1 of p.256 of [@Kneser:1929]. An early published proof is that in Sect.77 of [@Seifert.Threlfall:Topology] [^10]: \[foot:ConfIsom\] Let $\mathcal{CI}(\Sigma,h):=\{\phi\in{\mathrm{Diff}(\Sigma)}\mid\phi^*h=fh,\,f:\Sigma\rightarrow{\mathbb{R}}_+\}$ be the group of conformal isometries. For compact $\Sigma$ it is known to be compact except iff $\Sigma=S^3$ and $h$ conformal to the round metric [@Lelong-Ferrand:1971]. Hence, for $\Sigma\ne S^3$, we can average $h$ over the compact group $\mathcal{CI}(\Sigma,h)$ and obtain a new Riemannian metric $h'$ in the conformal equivalence class of $h$ for which $\mathcal{CI}(\Sigma,h)$ acts as proper isometries. Therefore, by the argument presented in Sect.\[sec:Superspace\], it cannot contain non-trivial elements fixing a frame.

ArXiv

--- abstract: 'We apply some tools developed in categorical logic to give an abstract description of constructions used to formalize constructive mathematics in foundations based on intensional type theory. The key concept we employ is that of a Lawvere hyperdoctrine for which we describe a notion of quotient completion. That notion includes the exact completion on a category with weak finite limits as an instance as well as examples from type theory that fall apart from this.' author: - 'Maria Emilia Maietti[^1]' - 'Giuseppe Rosolini[^2]' title: Quotient completion for the foundation of constructive mathematics --- [**MSC 2000**]{}: 03G30 03B15 18C50 03B20 03F55 [**Keywords:**]{} quotient completion, split fibration, type theory, setoid. Introduction {#intro} ============ Category theory provides a language to investigate the syntax and the semantics of formal systems on the same ground, as it provides an appropriate abstraction useful to bring to the foreground an algebraic structure that usually remains hidden behind both. In fact, the present paper is a plain example of how category theory offers a language which is suitable to describe a key property that foundations of constructive mathematics should have according to [@mtt]. In the following part of the Introduction, we address the relevance of such abstract properties. After that we describe the category-theoretic concepts that are dealt with in the paper. Finally, we examine two quotient models based on intensional type theory. #### The need of quotient completion to found constructive mathematics {#the-need-of-quotient-completion-to-found-constructive-mathematics .unnumbered} There are various foundations for constructive mathematics available in the literature: some are formulated in axiomatic set theory, others in category theory, yet others in type theory. In fact there is no foundation for constructive mathematics as standard as the theory ZFC is for classical mathematics. The authors of [@mtt] propose to look for a minimalist foundation which could form a common core for the most relevant constructive theories. Its finalized construction is in [@m09]. In [*loc. cit.*]{} the authors also state that a foundation for constructive mathematics should make it evident which key aspects differentiate it from classical mathematics. For instance, contrary to classical proofs, constructive proofs enjoy the [*existence property*]{}, [*i.e.*]{} one can extract programs that compute witnesses of existential statements occurring in them. Even more, any proof of a constructive system should be seen as a program. Hence, ideally, a foundation for constructive mathematics should be at the same time a set theory, in which to formalize mathematical theorems, and a programming language in which to extract the computational contents of mathematical proofs. Type theory provides examples of such formal systems, such as Martin-L[ö]{}f’s Intensional Type Theory [@PMTT] or Coquand’s Calculus of Constructions [@tc90]. But there is a problem in adopting such type theories as a foundation for constructive mathematics. First of all they do not validate extensional features used in the everyday practice of mathematics such as extensional equality of sets or of functions, nor do they offer quotient constructions. Indeed, if one wants that these systems act as useful functional programming languages, they must meet decidability issues on typing of proofs which are incompatible with extensional features. This is argued in more formal terms (with the notion of proofs-as-programs theory) in [@mtt]. The solution adopted in practice in the formalization of mathematics in type theory is to represent extensional concepts in a model built on top of the type theory, for example by using setoids, see [@disttheshof; @ven]. A main drawback of this approach for a constructive mathematician is that working with setoids—and especially with dependent setoids—is extremely complicated compared to adopting a foundation in the same vein as ZFC for classical mathematics, like for instance Aczel’s CZF or Friedman’s IZF. A natural solution is to work with an [*axiomatization of the quotient model of setoids supported by the intensional type theory*]{} instead of working [*directly in the model*]{}. In a sense, someone who would like to have a foundation for constructive mathematics based on a type theory is naturally led to abandon the traditional view of having a unique system to formalize mathematics in favour of a two-level foundation where one level is used for program extraction and the other to formalize mathematics. This is the central idea of the [*notion of constructive foundation*]{} put forward in [@mtt]. There it is required that a foundation for constructive mathematics should be [*a two-level theory*]{}: one level, named [*intensional*]{}, should be used as a programming language; the other, which is called [*extensional*]{}, should be closed under standard extensional constructs in order to be used as the actual formal system in which to perform mathematical proofs, and it should be seen as an abstraction of the intensional level according to Sambin’s forget-restore principle, see [@toolbox]. In [@m09] it was stated that to satisfy the link between the two levels in [@mtt] it is enough to interpret the extensional level in the intensional one [*by means of a quotient completion of the latter*]{}, [*i.e.*]{} to see the extensional level as a fragment of the internal language of a quotient completion built on the intensional one. The two-level minimalist foundation in [@m09] provides an example of such a constructive foundation. What remains to specify in the notion of constructive foundation in [@m09] is what one means “abstractly” by quotient completion. In particular one wants to see whether the construction performed in [@m09]—the quotient model built over the intensional level to interpret the extensional level—is an instance of a free construction on categories with structure. In the literature on category theory various constructions of quotient completion have been studied, for example in [@CarboniA:regec; @RosoliniG:typtec; @RosoliniG:loccce]. These constructions all rely on defining quotients as stable effective coequalizers of monic equivalence relations. Hence, they all produce exact categories—and indeed they usually go under the name of [*exact completions*]{}. But, as we observe in the present paper, the construction of quotients adopted in [@m09] does not necessarily lead to an exact category and therefore it cannot be an exact completion. This motivates the quest for a more general notion of quotient completion than the exact completion. In this paper we accomplish this task by relativizing the notion of quotient to that of a suitable hyperdoctrine: the fibers act as the “logic” in which to consider equivalence relations. With respect to those, we introduce a notion of quotient in the base category of the hyperdoctrine, and we prove that that notion is algebraic. In other words, there is a universal construction that “freely adds” quotients for the equivalence relations without adding any further power to the logic. We use a weakened notion of Lawvere hyperdoctrine [@LawvereF:adjif; @LawvereF:equhcs; @LawvereF:diaacc; @LawvereF:setfm], here simply called “elementary doctrine”, with respect to which we present a universal construction of quotient completion, which we call “elementary quotient completion”. Instances of this construction include both the quotient model in [@m09] and the exact completion of a category with finite limits in [@CarboniA:freecl]. Indeed, the study of the elementary quotient completion helps to isolate, in the doctrine setting, the properties of a model “with quotients” and to handle those properties independently from one another. Thanks to the more general setting than that, say, of categories with (weak) finite limits, it is also possible to analyse the properties of the elementary quotient completion that are similar to the exact completion, such as closure under exponentials in [@RosoliniG:loccce]. As a biproduct of all this, we also obtain a clear explanation of the well-known result that the category of total setoids à la Bishop built over Martin-L[ö]{}f’s type theory is the exact completion of an appropriate category with weak finite limits, see [@notepal; @CarboniA:somfcr; @RosoliniG:typtec]. Doctrines ========= We introduce the notion of doctrine that will be used to specify that of [*quotient*]{}. This notion is an obvious generalization of that of a hyperdoctrine. Hyperdoctrines were introduced, in a series of seminal papers, by F.W. Lawvere to synthesize the structural properties of logical systems, see [@LawvereF:adjif; @LawvereF:equhcs; @LawvereF:diaacc; @LawvereF:setfm]. His crucial intuition was to consider logical languages and theories as indexed categories and to study their 2-categorical properties. For instance, connectives and quantifiers are determined by adjunctions. That approach proved to be extremely fruitful, see [@MakkaiM:firocl; @LambekJ:inthoc; @jacobbook; @TaylorP:prafom; @OostenJ:reaait] and the references therein. Recall from [@LawvereF:adjif] that a hyperdoctrine is a contravariant functor $F:\cat{C}\op\longrightarrow\Cat{Heyt}$ from a cartesian closed category $\cat{C}$ to the category of Heyting algebras satisfying some further conditions: for every arrow $f:A\to B$ in $\cat{C}$, the homomorphism $F_f:F(B)\to F(A)$ of Heyting algebras—$F_f$ denotes the action of the functor $F$ on the arrow $f$—has a left adjoint $\D_{f}$ and a right adjoint $\B_{f}$ satisfying the Beck-Chevalley condition. The intuition is that a hyperdoctrine determines an appropriate categorical structure to abstract both notions of first order theory and of interpretation. A many-sorted first order theory gives rises directly to a hyperdoctrine $F:\cat{C}\op\longrightarrow\Cat{Heyt}$—a detailed presentation is given as example \[lta\]: - the objects of $\cat{C}$ are declarations of sort variables - a morphism $f:A\to B$ of $\cat{C}$ is a list of terms of the sorts in $B$, in the variables in the declaration $A$ - an object $P$ in $F(A)$ is a property written with the variables declared in $A$ - a morphism $P\leq Q$ in $F(A)$ shows that property $Q$ follows from the property $P$ - a functor $F_f:F(B)\to F(A)$ represents the substitution, in properties of the sort $B$, of the terms $f$ for the variables in $B$ - the adjoints $\D_f$ and $\B_f$ represent the existential quantifier and the universal quantifier, respectively. Another instance of hyperdoctrine $F:\cat{C}\op\longrightarrow\Cat{Heyt}$ is the following: - the objects of $\cat{C}$ are sets - a morphism $f:A\to B$ of $\cat{C}$ is a function into the set $B$ from the set $A$ - an object $P$ in $F(A)$ is a subset of the set $A$ - a morphism $P\leq Q$ in $F(A)$ indicates that $Q$ contains $P$ - a functor $F_f:F(B)\to F(A)$ acts by inverse image along $f$ on subsets of the set $B$ - the adjoints $\D_f$ and $\B_f$ must be evaluated, on a subset $P$ of $A$, respectively as $\begin{array}{l}\D_f(P)=\left\{b\in B\mid \exists a\in A [b=f(a)\Land a\in P]\right\}\\ \B_f(P)=\left\{b\in B\mid \forall a\in A[b=f(a)\Implies a\in P]\right\}\end{array}$ Thus a model of the many-sorted first order theory determines precisely a functor from the former hyperdoctrine to the latter, as already pointed out in [@LawvereF:adjif; @LawvereF:equhcs]. Our aim is to take advantage of the algebraic presentation of logic offered by hyperdoctrines and we shall consider a more general notion with the structure needed to define a quotient of an equivalence relation from that perspective. We shall follow [@LawvereF:equhcs] and use the word “doctrine” with some attribute to christen the more general notion and others derived from it. We shall be able to separate the logical components producing a universal construction of completion by quotients of doctrines. Somehow reflecting the essential logical structure that is needed in order to present the theory of an equivalence relation, the basic concept is that of a contravariant functor from a category with finite products into the category of inf-semilattices and inf-preserving maps. \[ph\] A is a functor $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ from (the opposite of) a category with finite products to the category of inf-semilattices, [*i.e.*]{} a contravariant functor $P:\cat{C}\op\longrightarrow\Cat{Pos}$ in the category of partial orders $\Cat{Pos}$ such that - for every object $A$ in , the partial order $P(A)$ has finite infs - for every arrow $f:A\to B$ in , the monotone map $P_f:P(B)\to P(A)$ preserves them.[^3] The structure of a primary doctrine is just what is needed to handle a many-sorted logic with binary conjunctions and a true constant, as seen in the following example. \[lta\] The leading logical example is the indexed order $LT:\cat{V}\op\longrightarrow\Cat{InfSL}$ given by the Lindenbaum-Tarski algebras of well-formed formulae of a first order theory (with only one sort). Given a theory $\mathscr{T}$ in a first order language $\mathscr{L}$, the domain category of the functor is the category of lists of variables and term substitutions: object of : are lists[^4] of distinct variables $\vec x=(x_1,\ldots,x_n)$ arrows : are lists of substitutions[^5] for variables $[\vec t/\vec y]:\vec x\to \vec y$ where each term $t_j$ in $\vec t$ is built in $\mathscr{L}$ on the variables $x_1,\ldots,x_n$ composition : $\xymatrix@1@=4em{\vec x\ar[r]^{[\vec t/\vec y]}& \vec y\ar[r]^{[\vec s/\vec z]}&\vec z}$ is given by simultaneous substitutions $$\xymatrix@=10em{\vec x \ar[r]^{\left[s_1[\vec t/\vec y]/z_1,\ldots,s_k[\vec t/\vec y]/z_k\right]} &\vec z}$$ The product of two objects $\vec x$ and $\vec y$ is given by a(ny) list $\vec w$ of as many distinct variables as the sum of the number of variables in $\vec x$ and of that in $\vec y$. Projections are given by substitution of the variables in $\vec x$ with the first in $\vec w$ and of the variables in $\vec y$ with the last in $\vec w$. The functor $LT:\cat{V}\op\longrightarrow\Cat{InfSL}$ is given as follows: for a list of distinct variables $\vec x$, the category $LT(\vec x)$ has objects : equivalence classes $\ec{W\kern.2ex}$ of well-formed formulae $W$ of $\mathscr{L}$ with no more free variables than $x_1$,…,$x_n$ with respect to provable reciprocal consequence $W\dashv\vdash_{\mathscr T}W'$ in $\mathscr T$.[^6] arrows : $\ec{W\kern.2ex}\to\ec{V\kern.2ex}$ are the provable consequences $W\vdash_{\mathscr T}V$ in $\mathscr T$ for some pair of representatives (hence for any pair) composition : is given by the cut rule in the logical calculus identities : $\ec{W\kern.2ex}\to\ec{W\kern.2ex}$ are given by the logical rules $W\vdash_{\mathscr T}W$ Observe that, in particular, for a list of distinct variables $\vec x$, the category $LT(\vec x)$ has finite limits: products are given by conjunctions of formulae and a terminal object is any provable formula, such as $\vec x=\vec x$, that is any formula equivalent to the true constant. \[model\] The following example of primary doctrine $S:\cat{S}\longrightarrow\Cat{InfSL}$ is the set-theoretic hyperdoctrine described in the introduction and it can be considered in an(y) axiomatic set theory such as ZF. We briefly recall its definition: $\cat{S}$ is the category of sets and functions, $S(A)$ is the poset category of subsets of the set $A$ whose morphisms are inclusions, a functor $S_f:S(B)\to S(A)$ acts as the inverse image $f^{-1}U $ on a subset $U$ of the set $B$. The example \[lta\] suggests that, by considering only doctrines, from a logical point of view one restricts attention to the mere existence of a proof of a consequence, [*i.e.*]{} one only deals with proof irrelevance. As already pointed out in [@LawvereF:adjif; @LawvereF:equhcs], a set-theoretic model of a first order theory determines precisely a functor from the doctrine $LT$ to the doctrine $S$ that preserves all the structure of a primary doctrine. As the example \[model\], also the example \[lta\] gives rise to a Lawvere hyperdoctrine when performed on a many-sorted first order theory giving rise to a cartesian closed base category. And the characterization of set-theoretic model extends directly, see [*loc.cit.*]{} \[wow2\] In many senses it is more general—and more elegant—to treat the abstract theory of the relevant structures for the present paper in terms of fibrations. For instance, a different, but equivalent presentation of the structure above is as a faithful fibration $p:\cat{A}\to\cat{C}$ between categories with binary products such that $p$ preserves them and has a right inverse right adjoint. In fact, a primary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ determines a faithful fibration $p_P:\Gr(P)\to\cat{C}$ by a well-known very general construction due to Grothendieck, see [@GrothendieckA:catfd; @jacobbook], which applies to indexed categories. We recall very briefly that construction in the present situation. The data for the category $\Gr(P)$ of $P$ are as follows: objects of $\Gr(P)$ : are pairs $(A,\alpha)$ such that $A$ is an object in and $\alpha$ is an object in $P(A)$. an arrow $(f,\phi):(A,\alpha)\to(B,\beta)$ : is a pair of an arrow $f:A\to B$ in and an arrow $\phi:\alpha\to P_f(\beta)$ composition : of $(f,\phi):(A,\alpha)\to(B,\beta)$ and $(g,\psi):(B,\beta)\to(C,\gamma)$ is $(g\circ f,P_f(\psi)\circ\phi)$. One checks that a product of objects $(A,\alpha)$ and $(B,\beta)$ is given by $$\xymatrix@C=3em{(A,\alpha)&(C,\chi)\ar[l]_{(\pr_1,\pi_1)} \ar[r]^{(\pr_2,\pi_2)}&(B,\beta)}$$ where $$\xymatrix@C=3em{A&C\ar[l]_{\pr_1}\ar[r]^{\pr_2}&B}$$ is a product in and $$\xymatrix@C=3em{P_{\pr_1}(\alpha)& \chi\ar[l]_(.35){\pi_1}\ar[r]^(.35){\pi_2}& P_{\pr_2}(\beta)}$$ is a product in $P(C)$. The first projection extends to a functor $p_F:\Gr(P)\to\cat{C}$ which is easily seen to be faithful with a right inverse right adjoint. On the other hand, given a faithful fibration $p:\cat{A}\to\cat{C}$, one considers the functor $\ch{p}:\cat{C}\op\longrightarrow\Cat{InfSL}$ which maps an object $A$ in to the partial order which is the poset reflection of the preorder of the vertical arrows on $A$, see [@jacobbook], [*i.e.*]{} one first considers the subcategory $p^A$ of consisting of those objects $\alpha$ such that $p(\alpha)=A$ and a map $g:\alpha\to\alpha'$ of is in $p^A$ if $p(g)=\id{A}$; faithfulness of $p$ ensures that the category $p^A$ is a preorder. Product preservation ensures that $p^A$ has binary products, the right inverse right adjoint ensures that $p^A$ has a terminal object. So the poset reflection of $p^A$ produces the partial order $\ch{p}(A)$ on the equivalence classes of objects of $p^A$ with respect to the equivalence given by isomorphism, where $\ec{\alpha}\leq\ec{\alpha'}$ if there is an arrow $g:\alpha\to\alpha'$ in $p^A$ for some pair of representatives (hence for any pair), and the partial order has finite infs. For an arrow $f:B\to A$ in , the functor $\ch{p}_f\, :\ch{p}(A)\to\ch{p}(B)$ sends an equivalence class $\ec{\alpha}$ to the equivalence class $\ec{\beta}$ such that there is a cartesian lifting $g:\beta\to\alpha$ of $f$. Setting up an appropriate 2-category for each structure (one for primary doctrines, one for faithful fibrations as above), it is easy to see that the two constructions extend to an equivalence between those 2-categories. The total category $\Gr(LT)$ is the syntactic presentation of the $\mathscr{L}$-definable subsets of (the finite powers of) a set underlying a model of the theory $\mathscr{T}$ with functions defined by terms in $\mathscr{L}$. The total category $\Gr(S)$ is the full subfibration on subset inclusions of the codomain fibration $\textrm{cod}:\cat{S}^{\to}\longrightarrow\cat{S}$ on the category of sets and functions. \[elh\] A primary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is if, for every $A$ and $C$ in , the functor $P_{id_C\times \Delta_A}: P(C\times (A\times A))\to P(C\times A)$[^7] has a left adjoint $\D_{id_C\times \Delta_A}$, and these satisfy : : for every $A$ and $C$ in , for $\alpha$ in $P(C\times (A\times A))$, $\beta$ in $P(C\times A)$, the canonical arrow $\D_{id_C\times \Delta_A}(P_{id_C\times \Delta_A}(\alpha) \Land_{C\times A}\beta)\leq \alpha\Land_{C\times (A\times A)}\D_{id_C\times \Delta_A}(\beta)$ in $P(C\times (A\times A))$ is iso (hence an identity). We refer the reader to [@LawvereF:equhcs; @jacobbook] for a thorough analysis of the concepts in the definition just given. \[equ\] For an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$, for any object $A$ in , taking $C$ a terminal object, the conditions in \[elh\] ensure the existence of a left adjoint $\D_{\Delta_A}$ to $P_{\Delta_A}:P(A\times A)\to P(A)$. On an object $\alpha$ in $P(A)$ it can be written as $$\label{delta} \D_{\Delta_A}(\alpha)=P_{\pr_1}(\alpha)\Land_{A\times A}\D_{\Delta_A}(\tt_A)= P_{\pr_2}(\alpha)\Land_{A\times A}\D_{\Delta_A}(\tt_A).$$ where $\tt_A$ is the terminal object of $P(A)$. Because of (\[delta\]), an abbreviation like $\delta_A$ for the object $\D_{\Delta_A}(\tt_A)$ is useful. \[ltae\] For $\mathscr{T}$ a first order theory, the primary fibration $LT:\cat{V}\op\longrightarrow\Cat{InfSL}$, as defined in \[lta\], is elementary exactly when $\mathscr{T}$ has an equality predicate. \[monoe\] The standard example of an elementary doctrine is the fibration of subobjects. Consider a category with finite products and pullbacks, [*i.e.*]{} with right adjoints to the diagonal functors $\cat{X}\to\cat{X}^2$ and $\cat{X}\to \cat{X}^{\strut\smash{\begin{array} {@{}r@{}}\scriptstyle\downarrow\\[-1.8ex]\scriptstyle\to\cdot\kern.2ex \end{array}}}$. The functor $S:\cat{X}\to\Cat{InfSL}$ assigns to any object $A$ in the poset $S(A)$ whose objects : are subobjects $\ec{\smash{\xymatrix@1{\alpha:X\ \ar@{>->}[r]&A}}}$ in with codomain $A$ $\ec{\alpha}\leq\ec{\alpha'}$ : if there is a commutative diagram $$\xymatrix@C=1em{ X\strut\ar@{>->}[rd]_{\alpha}\ar[rr]^{x}&&X'\strut\ar@{>->}[ld]^{\alpha'}\\ &A}$$ for a (necessarily unique) arrow $x:X\to X'$. For an arrow $f:B\to A$, the assignment mapping an equivalence $\ec{\alpha}$ in $S(A)$ to that represented by the left-hand arrow in the (chosen) pullback $$\xymatrix{Y\ar[d]_{\beta}\ar[r]&X\ar[d]^{\alpha}\\B\ar[r]_f&A}$$ produces a functor $S_f:S(A)\to S(B)$ which preserves products. Post-composition with an equalizer provides the elementary structure since equalizers are monic. In example \[monoe\], we used the same notation for the functor $S$ as in example \[model\] because that is a particular instance of \[monoe\] when $\cat{X}$ is the category $\cat{S}$ of sets and functions since each subobject $\ec{\smash{\xymatrix@1{\alpha:X\ \ar@{>->}[r]&A}}}$ has a unique inclusion $\xymatrix@1{U\ar@<-.3ex>@{^(->}[r]&A}$ among its representatives. \[weaklhy\] Consider a cartesian category with weak pullbacks, [*i.e.*]{} for every pair $f:B\to A$, $g:C\to A$ of arrows in , there is a commutative diagram $$\xymatrix@=4em{V_{f,g}\ar[d]_{\pr_{2,f,g}}\ar[r]^{\pr_{1,f,g}}& C\ar[d]^{g}\\B\ar[r]_{f}&A}$$ such that, for any $t:T\to B$, $s:T\to C$ satisfying $f\circ s=g\circ t$, there is $u:T\to V_{f,g}$ such that $t=\pr_{1,f,g}\circ u$ and $s=\pr_{2,f,g}\circ u$. One can consider the functor $\Psi:\cat{C}\to\Cat{InfSL}$ given by the poset reflection of each comma category $\cat{C}/A$ as $A$ varies over the objects of . So, for a given object $A$ in , one first considers the preorder whose objects : are arrows $\alpha:X\to A$ in with codomain $A$ $\alpha\leq\alpha'$ : if there is an arrow $x:X\to X'$ in providing a commutative diagram $$\xymatrix@C=1em{X\ar[rd]_{\alpha}\ar[rr]^{x}&&X'\ar[ld]^{\alpha'}\\&A}$$ and then quotients that preorder with respect to the reciprocal relation $\alpha\lge\alpha'$ to obtain the poset $\Psi(A)$. For an arrow $f:B\to A$, the assignment that maps an equivalence class $\ec{\alpha}$ in $\Psi(A)$ to that represented by the left-hand arrow $\pr_{2,f,\alpha}:V_{f,\alpha}\to B$ in the weak (chosen) pullback $$\xymatrix@=4em{V_{f,\alpha}\ar[d]_{\pr_{2,f,\alpha}}\ar[r]^{\pr_{1,f,\alpha}}& C\ar[d]^{\alpha}\\B\ar[r]_{f}&A}$$ produces a functor $\Psi_f:\Psi(A)\to\Psi(B)$. It is easy to check that it preserves products—which are given by weak pullbacks over $A$. The elementary structure is given by post-composition with the equalizer. This example is a slight generalization of a similar one given in [@LawvereF:equhcs]. Note that the apparently minor difference between the example in \[monoe\] and that in \[weaklhy\] depends crucially on the possibility of factoring an arbitrary arrow as a retraction followed by a monomorphism: for instance, in the category of sets and functions, the fact that the two doctrines are equivalent can be achieved thanks to the Axiom of Choice. Consider the 2-category of elementary doctrines: the 1-arrows : are pairs $(F,b)$ $$\xymatrix@C=4em@R=1em{ {\cat{C}\op}\ar[rd]^(.4){P}_(.4){}="P"\ar[dd]_F&\\ & {\Cat{InfSL}}\\ {\cat{D}\op}\ar[ru]_(.4){R}^(.4){}="R"&\ar"P";"R"_b^{\kern-.4ex\cdot}}$$ where the functor $F$ preserves finite products and, for every object $A$ in , the functor $b_A:P(A)\to R(F(A))$ preserves all the structure. More explicitly, $b_A$ preserves finite meets and, for every object $A$ in , $$\label{two} b_{A\times A}(\delta_A)\gel R_{<F(\pr_1),F(\pr_2)>}(\delta_{F(A)}).$$ the 2-arrows : are natural transformations $d$ such that $$\xymatrix@C=6em@R=1em{ {\cat{C}\op}\ar[rd]^(.4){P}_(.4){}="P" \ar@<-1ex>@/_/[dd]_F^{}="F"\ar@<1ex>@/^/[dd]^G_{}="G"&\\ & {\Cat{InfSL}}\\ {\cat{D}\op}\ar[ru]_(.4){R}^(.4){}="R"& \ar@/_/"P";"R"_{b\kern.5ex\cdot\kern-.5ex}="b" \ar@<1ex>@/^/"P";"R"^{\kern-.5ex\cdot\kern.5ex c}="c" \ar"F";"G"^{.}_d\ar@{}"b";"c"|{\leq}}$$ so that, for every object $A$ in and every $\alpha$ in $P(A)$, one has $R_{d_A}(b_A(\alpha))\leq c_A(\alpha)$ [^8]. The following definition is similar to \[elh\] and contribute the final part of the essential “logical” structure of an indexed poset. \[exh\] A primary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is if, for $A_1$ and $A_2$ in , for a(ny) projection $\pr:A_1\times A_2\to A_i$, $i=1,2$, the functor $P_{\pr_i}:P(A_i)\to P(A_1\times A_2)$ has a left adjoint $\D_{\pr_i}$, to which we shall unimaginatively refer as , and these satisfy : : for any pullback diagram $$\xymatrix{X'\ar[r]^{\pr'}\ar[d]_{f'}&A'\ar[d]^f\\X\ar[r]^{\pr}&A}$$ with $\pr$ a projection (hence also $\pr'$ a projection), for any $\beta$ in $P(X)$, the canonical arrow $\D_{\pr'}P_{f'}(\beta)\leq P_f\D_\pr(\beta)$ in $P(A')$ is iso; : : for $\pr:X\to A$ a projection, $\alpha$ in $P(A)$, $\beta$ in $P(X)$, the canonical arrow $\D_\pr(P_\pr(\alpha)\Land_A\beta)\leq\alpha\Land_X\D_\pr(\beta)$ in $P(A)$ is iso. About this notion we refer the reader to [@LawvereF:adjif; @jacobbook]. \(a) For $\mathscr{T}$ a first order theory, the primary fibration $LT:\cat{V}\op\longrightarrow\Cat{InfSL}$, as defined in \[lta\], is existential. An existential left adjoint to $P_\pr$ is computed by quantifying existentially the variables that are not involved in the substitution given by the projection, [*e.g.*]{} for the projection $\pr=[x/z]:(x,y)\to(z)$ and a formula $W$ with free variables at most $x$ and $y$, $\D_\pr(W)$ is $\Exists y.(W[z/x])$. We stop to note that the example reveals the meaning of the Beck-Chevalley condition: suppose $S$ and $T$ are sorts and consider the morphism $[x/z]:(x,y)\to(z)$. On a formula $W$ with free variables at most $x$ and $y$, for any morphism $[t/z]:(\ldots,w_i,\ldots)\to(z)$, the diagram $$\xymatrix@C=5em{ (\ldots,w_i,\ldots,y) \ar[d]_{[t/x,y]}\ar[r]^(.55){[\ldots,w_i/w_i,\ldots]}& (\ldots,w_i,\ldots)\ar[d]^{[t/z]}\\ (x,y)\ar[r]^(.55){[x/z]}&(z)}$$ is a pullback and the Beck-Chevalley condition rewrites the fact that substitution commutes with quantification as $$\Exists y.(W[t/x])\equiv(\Exists y.W[z/x])[t/z]$$ since the declaration $(\ldots,w_i,\ldots)$ ensures that $y$ does not appear in $t$. \(b) For a cartesian category with weak pullbacks, the elementary doctrine $\Psi:\cat{C}\to\Cat{InfSL}$ given in \[weaklhy\] is existential. Existential left adjoints are given by post-composition. \(c) The primary doctrine in example \[model\] is existential: on a subset $P$ of $A$, the adjoint $\D_\pr$, for a projection $\pr$, must be evaluated as $\D_\pr(P)=\left\{b\in B\mid \exists a\in A [a\in \pr^{-1}\left\{b\right\}\cap P]\right\}$, usually called the $\pr$. \[here\] In an existential elementary doctrine, for every map $f:A\to B$ in the functor $P_f$ has a left adjoint $\D_f$ that can be computed as $$\D_{\pr_2}(P_{f\times \id B}(\delta_B)\wedge P_{\pr_1}(\alpha))$$ for $\alpha$ in $P(A)$, where $\pr_1$ and $\pr_2$ are the projections from $A\times B$. For a category with products and pullbacks, the elementary doctrine $S:\cat{X}\to\Cat{InfSL}$ in \[monoe\] is existential if and only if has a stable proper factorization system $(\mathcal{E},\mathcal{M})$, see [@HughesJ:facsft; @PavlovicD:mapsii]. So, in particular, for regular, the subobject doctrine $S:\cat{X}\to\Cat{InfSL}$ is elementary existential. Consider the 2-full 2-subcategory of whose objects are elementary existential doctrines. The 1-arrows : are those pairs $(F,b)$ in $$\xymatrix@C=4em@R=1em{ {\cat{C}\op}\ar[rd]^(.4){P}_(.4){}="P"\ar[dd]_F&\\ & {\Cat{InfSL}}\\ {\cat{D}\op}\ar[ru]_(.4){R}^(.4){}="R"&\ar"P";"R"_b^{\kern-.4ex\cdot}}$$ such that $b$ preserves the left adjoints along projections. Hence, by \[here\], the second functor of a 1-arrow in preserves left adjoints along all arrows in $\cat{C}$. Quotients in an elementary doctrine =================================== The structure of elementary doctrine is suitable to describe the notion of an equivalence relation and that of a quotient for such a relation. \[per\] Given an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$, an object $A$ in and an object $\rho$ in $P(A\times A)$, we say that $\rho$ is a if it satisfies : : $\delta_A\leq\rho$ : : $\rho\leq P_{<\pr_2,\pr_1>}(\rho)$, for $\pr_1,\pr_2:A\times A\to A$ the first and second projection, respectively : : $P_{<\pr_1,\pr_2>}(\rho)\Land P_{<\pr_2,\pr_3>}(\rho)\leq P_{<\pr_1,\pr_3>}(\rho)$, for $\pr_1,\pr_2,\pr_3:A\times A\times A\to A$ the projections to the first, second and third factor, respectively. \[exer\] \(a) Given an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ and an object $A$ in , the object $\delta_A$ is a $P$-equivalence relation on $A$. \(b) Given a first order theory $\mathscr{T}$ with equality predicate, consider the elementary doctrine $LT:\cat{V}\op\longrightarrow\Cat{InfSL}$ as in \[ltae\]. An $LT$-equivalence relation is a $\mathscr{T}$-provable equivalence relation. \(c) For a category with products and pullbacks, consider the elementary doctrine of subobjects $S:\cat{X}\to\Cat{InfSL}$ as in \[monoe\]. An $S$-equivalence relation is an equivalence relation in the category . \(d) For a cartesian category with weak pullbacks, consider the elementary doctrine $\Psi:\cat{C}\to\Cat{InfSL}$. A $\Psi$-equivalence relation is a pseudo-equivalence relation in , see [@CarboniA:freecl]. \[reme\] Let $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ be an elementary doctrine. For an arrow $f:A\to B$ in , the functor $P_{f\times f}:P(B\times B)\to P(A\times A)$ takes a $P$-equivalence relation $\sigma$ on $B$ to a $P$-equivalence relation on $A$. Let $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ be an elementary doctrine. Let $\rho$ be a $P$-equivalence relation on $A$. A is an arrow $q:A\to C$ in such that $\rho\leq P_{q\times q}(\delta_C)$ and, for every arrow $g:A\to Z$ such that $\rho\leq P_{g\times g}(\delta_Z)$, there is a unique arrow $h:C\to Z$ such that $g=h\circ q$. Such a quotient is when, for every arrow $f:C'\to C$ in , there is a pullback $$\xymatrix{A'\ar[d]_{f'}\ar[r]^{q'}&C'\ar[d]^{f}\\A\ar[r]_q&C}$$ in and the arrow $q':A'\to C'$ is a quotient of the $P$-equivalence relation $P_{f'\times f'}(\rho)$. Let $f:A\to B$ be an arrow in . The is the $P$-equivalence relation $P_{f\times f}(\delta_B)$. A quotient $q:A\to B$ of the $P$-equivalence relation $\rho$ is if its $P$-kernel is $\rho$. \(a) Given an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ and an object $A$ in , a quotient of the $P$-equivalence relation $\delta_A$ on $A$ is the identity arrow $\id{A}:A\to A$. It is trivially stable and effective by definition. \(b) In the elementary doctrine $S:\cat{X}\to\Cat{InfSL}$ obtained from a category with products and pullbacks, a quotient of the $S$-equivalence relation $\ec{\smash{\xymatrix@=2.3ex@1{r:R\ \ar@{>->}[r]&A\times A}}}$ is precisely a coequalizer of the pair of $$\xymatrix@C=5em{R\ar@<.5ex>[r]^(.45){\pr_1\circ r} \ar@<-.5ex>[r]_(.45){\pr_2\circ r}&A}$$ —hence of any such pair obtained from the class $\ec{\smash{\xymatrix@=2.3ex@1{r:R\ \ar@{>->}[r]&A\times A}}}$. In particular, all $S$-equivalence relations have quotients which are stable and effective if and only if the category is exact. Set-like doctrines ================== We intend to develop doctrines that may interpret constructive theories for mathematics. We shall address two crucial properties that an elementary doctrine should verify in order to sustain such interpretations. One relates to the axiom of comprehension and to equality, the other to quotients. \[compd\] Let $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ be a primary doctrine. Let $A$ be an object in and $\alpha$ an object in $P(A)$. A is an arrow $\cmp\alpha:X\to A$ in such that $\tt_{X}\leq P_{\cmp\alpha}(\alpha)$ and, for every arrow $g:Y\to A$ such that $\tt_{Y}\leq P_g(\alpha)$ there is a unique $h:Y\to X$ such that $g=\cmp\alpha\circ h$. Such a comprehension is when, for every arrow $f:A'\to A$ in , $P_{f}(\alpha)$ has a comprehension. We say that $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ if, for every object $A$ in , every $\alpha$ in $P(A)$ has a comprehension. Again we refer the reader to [@LawvereF:equhcs]. The primary doctrine $S:\cat{S}\longrightarrow\Cat{InfSL}$ of \[model\] has comprehensions given by the trivial remark that a subset determines an actual function by inclusion. Among the examples listed in \[exer\], only example (c), the doctrine of subobjects $S:\cat{X}\to\Cat{InfSL}$ for a category with products and pullbacks, has comprehensions. Example (d), the elementary doctrine $\Psi:\cat{C}\to\Cat{InfSL}$ constructed as in \[weaklhy\] for a cartesian category with weak pullbacks, suggests to modify the requirements in \[compd\] by dropping uniqueness of the mediating arrows. Before doing that, we note the following. \[str\] For $f:A'\to A$ in , the mediating arrow $f'$ between the comprehensions $\cmp\alpha:X\to A$ and $\cmp{P_{f}(\alpha)}:X'\to A'$ produces a pullback $$\xymatrix@=3em{X'\ar[d]_{f'}\ar[r]^{\cmp{P_{f}(\alpha)}}&A'\ar[d]^{f}\\ X\ar[r]_{\cmp\alpha}&A.}$$ Hence a primary doctrine with comprehensions has comprehensions stable under pullbacks. \[wcd\] Let $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ be a primary doctrine. Let $A$ be an object in and $\alpha$ an object in $P(A)$. A is an arrow $\cmp\alpha:X\to A$ in such that $\tt_{X}\leq P_{\cmp\alpha}(\alpha)$ and, for every arrow $g:Y\to A$ such that $\tt_{Y}\leq P_g(\alpha)$ there is a (not necessarily unique) $h:Y\to X$ such that $g=\cmp\alpha\circ h$.[^9] Such a comprehension is when, for every arrow $f:A'\to A$ in , $P_{f}(\alpha)$ has a weak comprehension and there is a weak pullback $$\xymatrix@=3em{X'\ar[d]_{f'}\ar[r]^{\cmp{P_{f}(\alpha)}}&A'\ar[d]^{f}\\ X\ar[r]_{\cmp\alpha}&A.}$$ We say that $P:\cat{C}\op\longrightarrow\Cat{InfSL}$, if, for every object $A$ in , every $\alpha$ in $P(A)$ has a weak comprehension. \[cmpm\] Suppose $\cmp\alpha:X\to A$ is a weak comprehension of $\alpha$. The arrow $\cmp\alpha$ is monic if and only if it is a strict comprehension. Thus, when a diagonal is a weak comprehension, the arrow itself satisfies the condition strictly (as in \[compd\]). But some of its reindexings may satisfy the weaker condition without uniqueness. For a cartesian category with weak pullbacks, the elementary doctrine $\Psi:\cat{C}\to\Cat{InfSL}$ as in \[weaklhy\] has weak comprehensions. \[dswc\] Suppose $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is an elementary doctrine. If the diagonal arrow $\Delta_A:A\to A\times A$ is a stable $($weak$)$ comprehension of $\delta_A$, then every pair of parallel arrows $\xymatrix@1{X\ar@<.5ex>[r]^f\ar@<-.5ex>[r]_g&A}$ in has a $($weak$)$ equalizer. It follows immediately from the construction of a weak equalizer of $\xymatrix@1{X\ar@<.5ex>[r]^f\ar@<-.5ex>[r]_g&A}$ as a weak pullback of $\Delta_A$: in the diagram $$\xymatrix@=3em{ E\ar[d]_{f\circ\cmp{P_{<f,g>}(\delta_A)}}\ar[r]^{\cmp{P_{<f,g>}(\delta_A)}} &X\ar[d]_{<f,g>}\ar@<1ex>[rd]^f\ar@<.1ex>[rd]_g\\ A\ar[r]_(.45){\Delta_A}& A\times A\ar@<.5ex>[r]^{\pr_1}\ar@<-.5ex>[r]_{\pr_2}&\strut A&}$$ the square is a weak pullback. Since the bottom horizontal arrow is the equalizer of the parallel pair that follows it, the top horizontal arrow is a weak equalizer. We say that an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$, has if, for every object $A$ in , the diagonal $\Delta_A:A\to A\times A$ is a stable (weak) comprehension of $\delta_A$. Let $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ be a primary doctrine. Let $A$ be an object in and $\alpha$ an object in $P(A)$. A (weak) comprehension $\cmp\alpha:X\to A$ of $\alpha$ is if $\alpha\leq_A\beta$ whenever $\tt_X\leq_X P_{\cmp\alpha}(\beta)$ for $\beta$ in $P(A)$. Note that the notion of full (weak) comprehension ensures that $\alpha\leq_A\beta$ is equivalent to $\tt_X\leq_X P_{\cmp\alpha}(\beta)$ for $\beta$ in $P(A)$. In an elementary doctrine it follows directly from the definition of $\delta_A$ that the diagonal arrow $\Delta_A:A\to A\times A$ is a full comprehension if and only if it is the comprehension of $\delta_A$. \[dm\] Suppose that $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is an elementary doctrine with full comprehensions and comprehensive equalizers. If $f:A\to B$ is monic, then $P_{f\times f}(\delta_B)=\delta_A$. It follows by fullness of comprehensions after noting that the comprehension of each side is the kernel of $f$, which in turn follows from \[dswc\] and the fact that the kernel of a monic is the diagonal. The next lemma will be needed in section \[appl\]. \[wfs\] Let $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ be a primary doctrine with full weak comprehensions. Suppose also that, for a given $\alpha$ in $P(A)$, a weak comprehension $\cmp\alpha:X\to A$ is such that the functor $P_{\cmp\alpha}:P(A)\to P(X)$ has a right adjoint $\B_{\cmp\alpha}:P(X)\to P(A)$. Then $\alpha\Land\blank:P(A)\to P(A)$ has a right adjoint $\alpha\Implies\blank:P(A)\to P(A)$. Consider $\alpha\Implies\beta\colon=\B_{\cmp\alpha}(P_{\cmp\alpha}(\beta))$. To see that $\gamma \leq \B_{\cmp\alpha}(P_{\cmp\alpha}(\beta))$ if and only if $\alpha\Land\gamma\leq\beta$ proceed as follows. If $\gamma \leq \B_{\cmp\alpha}(P_{\cmp\alpha}(\beta))$ then $P_{\cmp\alpha}(\gamma)\leq P_{\cmp\alpha}(\beta)$ and, since a weak comprehension $\cmp{\alpha\Land\gamma}:Z\to A$ of $\alpha\Land\gamma$ factors through $\cmp\alpha:X\to A$, then $$\tt_Z\leq P_{\cmp{\alpha\Land\gamma}}(\alpha\Land\gamma)\leq P_{\cmp{\alpha\Land\gamma}}(\gamma)\leq P_{\cmp{\alpha\Land\gamma}}(\beta)$$ and, finally full comprehension yields that $\alpha\Land\gamma\leq\beta$. Next, if $\alpha\Land\gamma\leq\beta$, then $$P_{\cmp\alpha}(\gamma)\leq\tt_X\Land P_{\cmp\alpha}(\gamma)\leq P_{\cmp\alpha}(\alpha)\Land P_{\cmp\alpha}(\gamma)\leq P_{\cmp\alpha}(\alpha\Land\gamma)\leq P_{\cmp\alpha}(\beta).\eqno{\qedhere}$$ Consider the 2-full 2-subcategory of whose objects are elementary doctrines with full comprehensions and comprehensive equalizers. The 1-arrows : are those pairs $(F,b)$ in $$\xymatrix@C=4em@R=1em{ {\cat{C}\op}\ar[rd]^(.4){P}_(.4){}="P"\ar[dd]_F&\\ & {\Cat{InfSL}}\\ {\cat{D}\op}\ar[ru]_(.4){R}^(.4){}="R"&\ar"P";"R"_b^{\kern-.4ex\cdot}}$$ such that $F$ preserves comprehensions. The functor $F$ in a pair $(F,b)$ in preserves all finite limits. The other aspect that we shall consider about set-like doctrines is that every quotient should be of effective descent. We recall the notion of descent data for a $P$-equivalence relation: Given an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ and a $P$-equivalence relation $\rho$ on an object $A$ in , the partial order of descent data is the sub-order of $P(A)$ on those $\alpha$ such that $$P_{\pr_1}(\alpha)\Land_{A\times A}\rho\leq P_{\pr_2}(\alpha),$$ where $\pr_1,\pr_2:A\times A\to A$ are the projections. Given an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$, for $f:A\to B$ in , let $\rho$ be the $P$-kernel ${P_{f\times f}(\delta_B)}$. The functor $P_f:P(B)\to P(A)$ takes values in $\des\rho\subseteq P(A)$. \[efdes\] Given an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ and an arrow $f:A\to B$ in , let $\rho$ be the $P$-kernel ${P_{f\times f}(\delta_B)}$. The arrow $f$ is if the functor $P_f:P(B)\to\des\rho$ is an isomorphism. In the example of the doctrine $S:\cat{S}\to\Cat{InfSL}$ on the category of sets and functions, as in \[monoe\], every canonical surjection $f:A\to A/\sim$, in the quotient of an equivalence relation $\sim$ on $A$, is of effective descent. The condition in \[efdes\] recognizes the fact that the subsets of the $A/\sim$ are in bijection with those subsets $U$ of $A$ that are the equivalence relation, in the sense that, for $a_1,a_2\in A$ such that $a_1\sim a_2$ and $a_2\in U$, one has also that $a_1\in U$. Consider the 2-full 2-subcategory of whose objects are elementary doctrines $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ in with stable effective quotients of $P$-equivalence relations and of effective descent. The 1-arrows : are those pairs $(F,b)$ in $$\xymatrix@C=4em@R=1em{ {\cat{C}\op}\ar[rd]^(.4){P}_(.4){}="P"\ar[dd]_F&\\ & {\Cat{InfSL}}\\ {\cat{D}\op}\ar[ru]_(.4){R}^(.4){}="R"&\ar"P";"R"_b^{\kern-.4ex\cdot}}$$ such that $F$ preserves quotients and comprehensions. \[regq\] If $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is an elementary doctrine in , then the category is regular. Moreover, if $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is also existential, every equivalence relation in has a stable coequalizer. The category has pullbacks and these can be computed by means of comprehensions: in the case of interest, given an arrow $f:A\to B$, the comprehension $\cmp{\rho}\equiv\cmp{P_{f\times f}(\delta_B)}:K\to A\times A$ of the $P$-kernel of $f$ gives a pullback $$\xymatrix@=4em{ K\ar[r]^{\pr_1\circ\cmp\rho}\ar[d]_{\pr_2\circ\cmp\rho}&A\ar[d]^f\\ A\ar[r]_f&B}$$ in by \[dswc\]. Moreover, thanks to fullness of the comprehension of $\rho$, the quotient $q:A\to C$ of $\rho$ provides a coequalizer $$\xymatrix@=4em{K\ar@<.5ex>[r]^{\pr_1\circ\cmp\rho} \ar@<-.5ex>[r]_{\pr_2\circ\cmp\rho}&A\ar[r]^q&C}$$ which is stable thanks to the stability of quotients and comprehensions in $P$. For the second part, suppose that the doctrine $P$ is existential. An equivalence relation $\xymatrix@1{r:R\ \ar@{>->}[r]&A\times A}$ in determines the $P$-equivalence relation $\D_{r}(\tt_R)$ on $A$. To conclude note that the mediating arrow $R\to\cmp{\D_{r}(\tt_R)}$ is monic. The following result is a direct consequence of those accomplished in [@HughesJ:facsft], see also [@PavlovicD:mapsii]. \[exq\] Let $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ be a primary doctrine. $P$ is an existential elementary doctrine in where every monomorphism in is a comprehension if and only if is exact and $P$ is equivalent to the doctrine $S:\cat{C}\op\longrightarrow\Cat{InfSL}$ of subobjects. Completing with quotients as a free construction ================================================ There is a fairly obvious construction that produces an elementary doctrine with quotients. We shall present it in the following and prove that it satisfies a universal property. for the rest of the section. Consider the category $\cat{Q}_P$ of “quotients in $P$”, the , defined as follows: an object of $\cat{Q}_P$ : is a pair $(A,\rho)$ such that $\rho$ is a $P$-equivalence relation on $A$ an arrow [$\ec{f}:(A,\rho)\to(B,\sigma)$]{} : is an equivalence class of arrows $f:A\to B$ in (with a chosen representative) such that $\rho\leq_{A\times A}P_{f\times f}(\sigma)$ in $P(A\times A)$ with respect to the relation determined by the condition that $\rho\leq_{A\times A}P_{f\times g}(\sigma)$ Composition is given by that of on representatives, and identities are represented by identities of . The indexed partial inf-semilattice $\Q{P}:\cat{Q}_P\op\longrightarrow\Cat{InfSL}$ on $\cat{Q}_P$ will be given by categories of descent data: on an object $(A,\rho)$ it is defined as $$\Q{P}(A,\rho)\colon=\des{\rho}$$ and the following lemma is instrumental to give the assignment on arrows using the action of $P$ on arrows. With the notation used above, let $(A,\rho)$ and $(B,\sigma)$ be objects in $\cat{Q}_P$, and let $\beta$ be an object in . If $f,g:A\to B$ are arrows in such that $\rho\leq_{A\times A}P_{f\times g}(\sigma)$, then $$P_f(\beta)\gel P_g(\beta).$$ Since $\beta$ is in , it is $$P_{\pr_1'}(\beta)\Land\sigma\leq_{B\times B}P_{\pr_2'}(\beta)$$ where $\pr_1',\pr_2':B\times B\to B$ are the two projections. Hence $$P_{f\times g}(P_{\pr_1'}(\beta))\Land P_{f\times g}(\sigma)\leq_{A\times A}P_{f\times g}(P_{\pr_2'}(\beta))$$ since $P_{f\times g}$ preserves the structure. By the hypothesis that $\rho\leq_{A\times A}P_{f\times g}(\sigma)$, $$P_{f\circ\pr_1}(\beta)\Land\rho\leq_{A\times A}P_{g\circ\pr_2}(\beta)$$ where $\pr_1,\pr_2:A\times A\to A$ are the two projections. Taking $P_{\Delta_A}$ of both sides and recalling reflexivity of $\rho$ $$P_f(\beta)\gel P_f(\beta)\Land\tt_A\gel P_{\Delta_A}(P_{f\circ\pr_1}(\beta))\Land P_{\Delta_A}(\rho)\leq P_{\Delta_A}(P_{g\circ\pr_2}(\beta))\gel P_g(\beta).$$ The other direction follows by symmetry. With the notation used above, $\Q{P}:\cat{Q}_P\op\longrightarrow\Cat{InfSL}$ is a primary doctrine. For $(A,\rho)$ and $(B,\sigma)$ in $\cat{Q}_P$ let $\pr_1,\pr_3:A\times B\times A\times B\to A$ and $\pr_2,\pr_4:A\times B\times A\times B\to B$ be the four projections. The meet of two $P$-equivalence relations on $A\times B$ $$\rho\boxtimes\sigma\colon=P_{<\pr_1,\pr_3>}(\rho) \Land_{A\times B\times A\times B} P_{<\pr_2,\pr_4>}(\sigma)$$ is a $P$-equivalence relation on $A\times B$ and it provides an object $(A\times B,\rho\boxtimes\sigma)$ which, together with the arrows determined by the two projections from $A\times B$, gives a product of $(A,\rho)$ and $(B,\sigma)$ in $\cat{Q}_P$.For each $(A,\rho)$, the sub-partial order $\des\rho\subseteq P(A)$ is closed under finite meets. for the rest of the section. \[wscl\] With the notation used above, is an elementary doctrine with comprehensions and comprehensive equalizers. If $P$ has full weak comprehensions, then has full comprehensions. First we show that $\cat{Q}_P$ has equalizers, hence all finite limits. Consider a parallel pair $\ec{f},\ec{g}:(A,\rho)\to(B,\sigma)$, and let $e\colon=\cmp{P_{<f,g>}(\sigma)}:E\to A$ be a weak comprehension. It is easy to see that $\ec{e}:(E,P_{e\times e}(\rho))\to(A,\rho)$ is an equalizer as required.A similar argument shows that $\Q{P}$ has comprehensions. More precisely, given $(A,\rho)$ and $\beta$ in $\des\rho$, let $\cmp\beta:X\to A$ be a weak comprehension for $\beta$ over $A$ in the doctrine $P$. A comprehension for $\beta$ over $(A,\rho)$ in is $\ec{\cmp\beta}:(X,P_{\cmp\beta\times\cmp\beta}(\rho))\to(A,\rho)$. Fullness of weak comprehensions in $P$ implies fullness of comprehensions in because objects of on $(A,\rho)$ are descent data related to $P$. A left adjoint $\QD_{\ec{\Delta_A}}$ for $\Q{P}_{\ec{\Delta_A}}$ is computed by $$\QD_{\ec{\Delta_A}}(\alpha)\colon= P_{\pr_1}(\alpha)\Land_{A\times A}\rho$$ for $\alpha$ in . Indeed, let $\theta$ be in such that $\alpha\leq_{(A,\rho)}\Q{P}_{\ec{\Delta_A}}(\theta)$, [*i.e.*]{} $\alpha\leq_AP_{\Delta_A}(\theta)$. Thus $\D_{\Delta_A}(\alpha)\leq_{A\times A}\theta$ and one has by remark \[equ\] $$\begin{array}{r@{}l} P_{\pr_1'}(\alpha)\Land P_{<\pr_1',\pr_2'>}(\delta_A)\Land P_{<\pr_2',\pr_3'>}(\rho)&{}\leq_{A\times A\times A} P_{<\pr_1',\pr_2'>}(\theta)\Land P_{<\pr_2',\pr_3'>}(\rho)\\[1.5ex] &{}\leq_{A\times A\times A}P_{<\pr_1',\pr_3'>}(\theta) \end{array}$$ for $\pr_i':A\times A\times A\to A,\quad i=1,2,3$, the three projections. Hence $P_{\pr_1}(\alpha)\Land\rho\leq_{A\times A}\theta$, which gives $\QD_{\ec{\Delta_A}}(\alpha) \leq_{(A\times A,\rho\boxtimes\rho)}\theta$. It is easy to prove the converse that, if $\QD_{\ec{\Delta_A}}(\alpha)\leq\theta$, then $\alpha\leq\Q{P}_{\ec{\Delta_A}}(\theta)$. More generally, a left adjoint $\QD_{\ec{id_C\times \Delta_A}}$ for $\Q{P}_{\ec{id_C\times\Delta_A}}$ is computed analogously by $$\QD_{\ec{id_C\times\Delta_A}}(\alpha)\colon= P_{<\pr_1,\pr_2>}(\alpha)\Land_{C\times A\times A}P_{<\pr_2,\pr_3>}(\rho)$$ for $\alpha$ in $\Q{P}((C,\sigma)\times (A,\rho))$ with $\pr_1: C \times A\times A\rightarrow C$ and $\pr_i: C \times A\times A\rightarrow A$ for $i=2,3$ the corresponding projections. Finally, $\Q{P}$ has comprehensive equalizers by construction. A careful reader may have noticed that the hypothesis on weak comprehensions in $P$ in \[wscl\] was needed in order to construct all pullbacks and characterize regular monos in $\cat{Q}_P$. With the notation used above, has effective quotients of -equivalence relations and those are of effective descent. Since the sub-partial order $\des\rho\subseteq P(A)$ is closed under finite meets, a -equivalence relation $\tau$ on $(A,\rho)$ is also a $P$-equivalence relation on $A$. It is easy to see that $\ec{\id{A}}:(A,\rho)\to(A,\tau)$ is an effective quotient and of effective descent since $\rho\leq_{A\times A}\tau$. With the notation used above, a quotient of -equivalence is stable. Let $\tau$ be a -equivalence relation on $(A,\rho)$, let $\ec{\id{A}}:(A,\rho)\to(A,\tau)$ be its quotient, and let $\ec{f}:(B,\sigma)\to(A,\tau)$ be an arrow in $\cat{Q}_P$. By the previous lemma \[wscl\], in $\cat{Q}_P$ there is a pullback diagram $$\xymatrix@C=1.5ex@R=2.5em{(E,\upsilon)\ar[rr]\ar[dd] \ar[rd]|{\EC{\cmp{P_{f\times\id{A}}(\tau)}}} &&(B,\sigma)\ar[dd]^{\EC{f}}\\ &(B\times A,\sigma\boxtimes\rho) \ar[ru]_(.6){\EC{\pr_1}}\ar[ld]^(.4){\EC{\pr_2}}\\ (A,\rho)\ar[rr]_{\EC{\id{A}}}&&(A,\tau)}$$ where $\upsilon\colon= P_{\cmp{P_{f\times\id{A}}(\tau)}\times\cmp{P_{f\times\id{A}}(\tau)}} (\sigma\boxtimes\rho)$. Let $e\colon=\cmp{P_{f\times\id{A}}(\tau)}$ and $\omega\colon=P_{e\times e}(\sigma\boxtimes\tau)$. The arrow $<\id{B},f>:B\to B\times A$ of $\cat{C}$ factors through the comprehension $\cmp{P_{f\times\id{A}}(\tau)}:E\to B\times A$ because $$\tt_B\leq P_{f}(P_{\Delta_A}(\tau))= P_{<f,f>}(\tau)=P_{<\id{B},f>}(P_{f\times\id{A}}(\tau)),$$ say $<\id{B},f>=\cmp{P_{f\times\id{A}}(\tau)}\circ g= e \circ g$. Moreover $$\sigma\leq P_{<\id{B},f>\times<\id{B},f>}(\sigma\boxtimes\tau)= P_{g\times g}(P_{e\times e}(\sigma\boxtimes\tau))= P_{g\times g}(\omega)$$ hence producing a commutative diagram $$\xymatrix@C=.5ex@R=3em{ (E,\omega)\ar[rd]_(.4){\EC{\cmp{P_{f\times\id{A}}(\tau)}}}&& (B,\sigma)\ar[ll]^{\EC{g}}\ar[dl]^(.4){\EC{<\id{B},f>}}\\ &(B\times A,\sigma\boxtimes\tau)}$$ in $\cat{Q}_P$. It is easy to see that $\ec{g}:(B,\sigma)\to(E,\omega)$ has a retraction $$\ec{\pr_1\circ\cmp{P_{f\times\id{A}}(\tau)}}:(E,\omega)\to(B,\sigma)$$ which is monic and hence an inverse of $\ec{g}$. Therefore $(B,\sigma)$ becomes a quotient of $\omega$ on $(E,\upsilon)$. There is a 1-arrow $(J,j):P\to\Q{P}$ in . The functor $J:\cat{C}\to\cat{Q}_P$ sends an object $A$ in to $(A,\delta_A)$ and an arrow $f:A\to B$ to $\ec{f}:(A,\delta_A)\to(B,\delta_B)$ since $\delta_A\leq_{A\times A}P_{f\times f}(\delta_B)$ by \[exer\](a) and \[reme\]. The functor $J$ is full. For $A$ in , the partial order $\Q{P}(A,\delta_A)=\des{\delta_A}$ is $P(A)$ since $P_{\pr_1}\alpha\Land\delta_A\leq_{A\times A}P_{\pr_2}\alpha$ for any $\alpha$ in $P(A)$. Take the function $j_A:P(A)\to\Q{P}(A,\delta_A)$ to be the identity. \[cov\]With the notation used above, 1. every object in $\cat{Q}_P$ is a quotient of a -equivalence relation on an object in the image of $J$ 2. every object in the image of $J$ is projective with respect to quotients of -equivalence relation. \(i) The object $(A,\rho)$ is a quotient of $(A,\delta_A)$.(ii) A quotient of a -equivalence relation is isomorphic to one that has an identity as a representative. for the rest of the section. \[wec\] With the notation used above, the functor $J:\cat{C}\to\cat{Q}_P$ is faithful. For $f,g:A\to B$ suppose that $J(f)=J(g)$. In other words, $\D_{\Delta_A}(\tt_A)=\delta_A\leq_{A\times A}P_{f\times g}(\delta_B)$, and equivalently $$\tt_A\leq_AP_{\Delta_A}(P_{f\times g}(\delta_B))\gel_AP_{<f, g>}(\delta_B)$$ since $\D_{\Delta_A}\dashv P_{\Delta_A}$. It follows that the identity on $A$ equalizes $f$ and $g$, thanks to lemma \[dswc\] and the hypothesis that $P$ has comprehensive weak equalizers. The doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ in \[wec\] is obtained from $\Q{P}:\cat{Q}_P\op\longrightarrow\Cat{InfSL}$ by change of base along $J$ since $j$ is the identity natural transformation. In logical terms, this states that the theory expressing $P$ extended with a quotient constructor is conservative over the original theory. We can now prove that the assignment $P\mapsto\Q{P}$ gives a left bi-adjoint to the forgetful 2-functor $U:\QH\to\CH$. \[mthm\] For every elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ in pre-composition with the 1-arrow $$\xymatrix@C=4em@R=1em{ {\cat{C}\op}\ar[rd]^(.4){P}_(.4){}="P"\ar[dd]_J&\\ & {\Cat{InfSL}}\\ {\cat{Q}_P\op}\ar[ru]_(.4){\Q{P}}^(.4){}="R"&\ar"P";"R"_j^{\kern-.4ex\cdot}}$$ in induces an essential equivalence of categories $$\label{eqv} -\circ(J,j):\QH(\Q{P},X)\equiv\CH(P,X)$$ for every $X$ in . Suppose $X$ is a doctrine in . As to full faithfulness of the functor in (\[eqv\]), consider two pairs $(F,b)$ and $(G,c)$ of 1-arrows from to $X$. By \[cov\], the natural transformation $\theta:F\stackrel.\to G$ in a 2-arrow from $(F,b)$ to $(G,c)$ in is completely determined by its action on objects in the image of $J$. And, since a quotient $q:U\to V$ of an $X$-equivalence relation on $U$ is of effective descent, $X(V)$ is a full sub-partial order of $X(U)$. Essential surjectivity of the functor in (\[eqv\]) follows from \[cov\]. Preservation of comprehensions for $J$ follows from \[dm\] because a comprehension is monic. The reason why $-\circ(J,j)$ need not be a strong equivalence in \[mthm\] is that quotients in the doctrine $X$ are determined only up to iso. Since in the doctrine quotients are determined by a specified construction, it is possible to strengthen the result by restricting the 2-categories and to doctrines with a choice of the categorical structure. \[eqc\] \(a) The structure of the elementary quotient completion for the doctrine $LT:\cat{V}\op\longrightarrow\Cat{InfSL}$ from a first order theory $\mathscr{T}$ with an equality predicate depends heavily on the choice of basic operation symbols used to present the theory $\mathscr{T}$, see \[wfs\] and \[ccqc\]. \(b) For a cartesian category with weak pullbacks, the elementary quotient completion of the elementary doctrine $\Psi:\cat{C}\op\to\Cat{InfSL}$ given in \[weaklhy\] is essentially equivalent to the exact completion of , see [@CarboniA:regec]. \[exlexcom\] By \[regq\], the elementary quotient completion $\cat{Q}_S$ of an elementary doctrine $S:\cat{X}\op\to\Cat{InfSL}$ in , obtained from a category with products and pullbacks as in \[monoe\], is regular. Then, it may seem natural to compare $\cat{Q}_S$ with the regular completion of , see [@CarboniA:somfcr], or with the exact/regular completion of in case is regular, see [@FreydP:cata; @CarboniA:somfcr]. The three constructions are in general not equivalent; that will be treated in future work. Also, an example of an elementary quotient completion that is not exact will be presented in section \[twl\]. In case an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ with comprehensive equalizers fails to have *all* comprehensions, it is possible to add these freely (hence fully) preserving the rest of the structure by taking the fibration of vertical maps on the total category of $P$, see [@jacobbook; @TaylorP:prafom]. Applications to richer doctrines {#appl} ================================ We shall now analyse the extent to which the elementary quotient completion construction produced in the previous section behaves with respect to further logical structure. We start with existential doctrines. \[exi\] Suppose $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is an existential elementary doctrine. Then the doctrine $\Q{P}:\cat{Q}_P\op\longrightarrow\Cat{InfSL}$ is existential and the pair $(J,j):P\to\Q{P}$ is a 1-arrow in . The left adjoint along a projection with respect to $\Q{P}$ is given by the left adjoint along the underlying projection in $\cat{C}$ with respect to $P$. We now consider other logical structures such as implications and universal quantification, always remaining within Lawvere’s unifying schema of “logical constructs as adjoints”, see [@LawvereF:adjif; @jacobbook]. \[imh\] A primary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is if, for every object $A$ in , every $\alpha$ in $P(A)$, the functor $\alpha\Land\blank:P(A)\to P(A)$ has a right adjoint $\alpha\Implies\blank:P(A)\to P(A)$. \(a) For $\mathscr{T}$ a first order theory, the primary fibration $LT:\cat{V}\op\longrightarrow\Cat{InfSL}$, as defined in \[lta\], is implicational as, for $V$ and $W$ in $LT(\vec x)$, the formula $V\Implies W$ gives the value of the right adjoint of $V\Land\blank$ on $W$. \(b) The primary doctrine in example \[model\] is implicational: on a subset $P$ of $A$, the right adjoint $P\Implies\blank$ on a subset $Q$ of $A$ is computed as $$P\Implies Q\colon=\left\{a\in A\mid a\in P\Implies a\in Q\right\}$$ \[unh\] A primary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is if, for $A_1$ and $A_2$ in , for a(ny) projection $\pr_i:A_1\times A_2\to A_i$, $i=1,2$, the functor $P_{\pr_i}:P(A_i)\to P(A_1\times A_2)$ has a right adjoint $\B_{\pr_i}$, and these satisfy the Beck-Chevalley condition:for any pullback diagram $$\xymatrix{X'\ar[r]^{\pr'}\ar[d]_{f'}&A'\ar[d]^f\\X\ar[r]^{\pr}&A}$$ with $\pr$ a projection (hence also $\pr'$ a projection), for any $\beta$ in $P(X)$, the canonical arrow $ P_f\B_\pr(\beta)\leq \B_{\pr'}P_{f'}(\beta)$ in $P(A')$ is iso; \(a) For $\mathscr{T}$ a first order theory, the primary fibration $LT:\cat{V}\op\longrightarrow\Cat{InfSL}$, as defined in \[lta\], is universal. A right adjoint to $P_\pr$ is computed by quantifying universally the variables that are not involved in the substitution given by the projection, [*e.g.*]{} for the projection $\pr=[z/x]:(x,y)\to(z)$ and a formula $W$ with free variables at most $x$ and $y$, $\B_\pr(W)$ is $\Forall y.(W[z/x])$. As in the case of the left adjoints in \[exh\], the Beck-Chevalley condition expresses the correct interplay between term substitution and universal quantification. \(b) The primary doctrine in example \[model\] is universal: on a subset $U$ of $A$, the adjoint $\B_\pr$, for a projection $\pr$, is given by $\B_\pr(U)\colon=\left\{b\in B\mid \pr^{-1}\left\{b\right\}\subseteq U\right\}.$ \[hure\] In an elementary doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ which is implicational and universal for every map $f:A\to B$ in the functor $P_f$ has a right adjoint $\B_f$ that can be computed as $$\B_{\pr_2}(P_{f\times \id B}(\delta_B)\Implies P_{\pr_1}(\alpha))$$ for $\alpha$ in $P(A)$, where $\pr_1$ and $\pr_2$ are the projections from $A\times B$—similarly to \[here\]. Moreover, if $P$ has full weak comprehensions, then the converse also holds by \[wfs\]. \#1\#2[$\mathop{\mathrm{exp}}({#1},{#2})$]{} \#1\#2[$\mathop{\mathrm{wexp}}({#1},{#2})$]{} \#1\#2[$[{#1}\to_w{#2}]$]{} \#1\#2[$\llparenthesis{#2}^{#1}\rrparenthesis$]{} \#1\#2[$\left[{#2}^{#1}\right]$]{} The following is an instance of a general situation for the quotient completion: a weak representation in the doctrine $P$ gives rise to a strict representation in the completion . Recall from [@RosoliniG:loccce] that a is an arrow $w:\W{A}{B}\times A\to B$ such that for every arrow $f:X\times A\to B$ there is an arrow $f':X\to\W{A}{B}$ with $f=w\circ(f'\times\id{A})$. In [*loc.cit.*]{}, the notion above was instrumental to give a characterization of those exact completions which are locally cartesian closed; it helps to show how the more general structure in the present paper extends that of exact completion, see also [@OostenJ:reaait]. We follow [@RosoliniG:loccce] and say that is if it has a weak evaluation for each pair of objects. \[ccqc\] Let $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ be an elementary, implicational, universal doctrine with weak comprehensions. If is weakly cartesian closed, then $\cat{Q}_P$ is cartesian closed and $\Q{P}$ is implicational and universal. Moreover, the 1-arrow $(J,j):P\to\Q{P}$ preserves the implicational and universal structure. Suppose that $(A,\rho)$ and $(B,\sigma)$ are objects in $\cat{Q}_P$ and $w:\W{A}{B}\times A\to B$ is a weak evaluation from $A$ to $B$ in . Recalling \[wfs\], consider the object $\varphi$ of $P(\W{A}{B}\times\W{A}{B})$: $$\varphi\, =\, \B_{\pr_3}\B_{\pr_4}(P_{<\pr_3,\pr_4>}(\rho)\Implies P_{<w<\pr_1,\pr_3>,w<\pr_2,\pr_4>>}(\sigma))$$ where $\pr_1,\pr_2:\W{A}{B}\times\W{A}{B}\times A\times A\to\W{A}{B}$ and $\pr_3,\pr_4:\W{A}{B}\times\W{A}{B}\times A\times A\to A$ are the projections. Within the logic provided by the doctrine, it can be described as the intension of those pairs $(e_1,e_2)$ of “elements of $\W{A}{B}$” such that $$\Forall a_3\in A.\Forall a_4\in A.(a_3\mathrel{\rho}a_4\Implies w(e_1,a_3)\mathrel{\sigma}w(e_2,a_4))$$ where we have used the same indices for the variables as those for the projections. It is easy to see that $\varphi$ satisfies the conditions of symmetry and transitivity in \[per\] so that, considered a weak comprehension $\cmp{P_{\Delta_{\W{A}{B}}}(\varphi)}:F\to\W{A}{B}$ of $P_{\Delta_{\W{A}{B}}}(\varphi)$ in $P(\W{A}{B})$, the object $\psi\colon= P_{\cmp{P_{\Delta_{\W{A}{B}}}(\varphi)}\times\cmp{P_{\Delta_{\W{A}{B}}}(\varphi)}} (\varphi)$ is a $P$-equivalence relation. It is straightforward to prove that $w\circ(\cmp{P_{\Delta_{\W{A}{B}}}(\varphi)}\times\id{A}): (F,\psi)\times(A,\rho)\to(B,\sigma)$ is a (strict) evalutation from $(A,\rho)$ to $(B,\sigma)$. The right adjoints required for the implicational and universal structures on $\Q{P}$ are computed by the corresponding right adjoints with respect to $P$. As an application, we shall consider the Axiom of Unique Choice (AUC) within an existential doctrine $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ with the same structure as in \[ccqc\]. Recall that (AUC) states that, in a cartesian closed regular category, a total relation, which is [*also*]{} single-valued, contains the graph of a function, see [*e.g.*]{} [@streicher]. We shall adapt it to the more general frame of existential doctrines as in \[ccqc\]. Let $w:\W{A}{B}\times A\to B$ be a weak evaluation from $A$ to $B$ in . We say that (AUC) if, for every $\rho$ in $P(A\times B)$, $$\def\mp#1_#2{{\displaystyle\mathop{#1}_{\scriptstyle\kern1ex #2}}} \begin{array}{@{}l@{}l@{}} \mp\B_{A\to1}\mp\D_{A\times B\to A}(\rho)&\Land \mp\B_{B\times B\to1}\mp\B_{A\times B\times B\to B\times B} ((P_{<\pr_1,\pr_2>}(\rho)\Land P_{<\pr_1,\pr_3>}(\rho))\Implies P_{<\pr_2,\pr_3>}(\delta_B))\leq\\[1ex] &\leq\mp\D_{\W{A}{B}\to1}\mp\B_{\W{A}{B}\times A\to\W{A}{B}} P_{<\pr_2',w >}(\rho) \end{array}$$ where we dispensed with the labels on the various appropriate projection arrows as these are easily reconstructed. Using plain logical notation (as in the proof of \[ccqc\]), $$\begin{array}{@{}l@{}l@{}} \Forall x_1\in A. \, \Exists y_2\in B. \ \ x_1\mathrel{\rho}y_2&{}\ \Land\ \Forall y_2\in B.\, \Forall y_3\in B. \, \Forall x_1\in A. \, (\, (x_1\mathrel{\rho}y_2\Land x_1\mathrel{\rho}y_3)\Implies y_2\mathrel{\delta_B} y_3\, ) \leq\\[1ex] &\leq\Exists h_1\in\W{A}{B}. \, \Forall x_2\in A. \ \ x_2\mathrel{\rho}w(h_1,x_2). \end{array}$$ We skip the easy proof that, if (AUC) holds for a weak evaluation from $A$ to $B$, then it holds for every weak evaluation. For that reason, from now on we say that (AUC) holds from $A$ to $B$. \[uac\]Suppose that $P:\cat{C}\op\longrightarrow\Cat{InfSL}$ is an elementary existential, implicational and universal doctrine with weak comprehensions. Suppose that is weakly cartesian closed and that $A$ and $B$ are objects in . The following are equivalent: 1. (AUC) holds from $A$ to $B$ in $P$ 2. (AUC) holds from $(A,\delta_A)$ to $(B,\delta_B)$ in $\Q{P}$. It follows directly from the preservation properties of the 1-arrow $(J,j):P\to\Q{P}$ by proposition \[exi\] and \[ccqc\]. Examples from constructive foundations ====================================== The main applications of the elementary quotient completion appear in the formalization of constructive mathematics in type theory. As we already explained, the ability of constructing quotients of equivalence relations is an essential tool of standard mathematics; any formalization of constructive mathematics that intends to achieve foundational relevance must allow treatment of such quotients in some way, in particular if it is inherently intensional. We shall sketch two examples of elementary quotient completion in type theory: one is given over intensional Martin-L[ö]{}f’s type theory [@PMTT] and the other over the intensional level of the minimalist foundation [@m09]. Both models are based on total setoids à la Bishop, with proof-terms as morphisms. The setoid model over Martin-L[ö]{}f’s Type Theory -------------------------------------------------- The model of “setoids” over intensional Martin-L[ö]{}f’s Type Theory, see [@notepal; @dyb], is an instance of the exact completion of a category with finite products and weak equalizers. We already know that it is an elementary quotient completion, so we simply review the elementary doctrine one can construct in order to obtain it as such a completion. The base category is a syntactic construction $\cat{ML}$ is defined as follows. The objects of $\cat{ML}$ : are closed sets of MLTT. An arrow $\ec{t\in B\ [x\in A]}:A\to B$ : is an equivalence class of terms $t\in B\ [x\in A]$[^10] derivable in MLTT where two terms $t\in B\ [x\in A]$ and $t'\in B\ [x\in A]$ are equivalent if there is a proof-term in MLTT $$p\in \mathsf{Id}(B, t, t')\ [x\in A]$$ The composition of $\ec{t\in B\ [x\in A]}:A\to B$ and $\ec{s(x)\in C\ [x\in B]}$ is defined as $$\ec{s(t)\in C\ [x\in A]}:A\to C.$$ The identity morphism on $A$ is $\ec{x\in A\ [x\in A]}:A\to A$. \[synmltt\] The category $\cat{ML}$ has finite products and weak equalizers. A terminal object is the singleton $N_1$. A product of closed sets $A,B$ is $\ds_{x\in A}B$ with its first and second projections since $$\mathsf{idpeel}(z,(x,y).\mathsf{id}(<x,y>))\in \mathsf{Id}(\ds_{x\in A}B,<\pi_1(z),\pi_2(z)>,z)$$ thanks to the $\beta$-rules.A weak equalizer of $\ec{t\in B\ [x\in A]}, \ec{t'\in B\ [x\in A]}:A\to B$ is $$\ec{\pi_1(z)\in A\ [z\in \ds_{x\in A} \mathsf{Id}(B, t,t')]}: \ds_{x\in A} \mathsf{Id}(B, t,t')\to A\qquad\qedhere$$ The functor $\F:\cat{ML}\op\longrightarrow\Cat{InfSL}$ is defined on a closed set $A$ as the partial order consisting of equivalence classes $\ec{\phi\prp [x\in A]}$ of predicates : in MLTT depending on $A$ with respect to equiprovability, [*i.e.*]{} $\phi\prp [x\in A]\sim\phi'\prp [x\in A]$ if there is a proof of $\phi\leftrightarrow\phi'\prp [x\in A]$ in MLTT $\ec{\phi\prp [x\in A]}\leq\ec{\psi\prp [x\in A]}$ : if there is a proof-term $r\in\psi\ [x\in A, w\in \phi]$ in MLTT. The action of the functor on arrows in is given by substitution. \[mlttd\] The doctrine $\F:\cat{ML}\op\longrightarrow\Cat{InfSL}$ is existential elementary, with full weak comprehensions and comprehensive weak equalizers. Products of propositions and the singleton set provide finite infs thanks to the identification of propositions with sets typical of MLTT. The left adjoint to substitution with $\ec{t(x)\in B\ [x\in A]}:A\to B$ is computed, for $\phi(x)\prp [x\in A]$, as the equivalence class represented by the proposition $$%% \ds_{x\in A}\ds_{z\in\mathsf{Id}(A,t,y)}\phi\prp [y\in B]. \ds_{x\in A}(\mathsf{Id}(B,t(x),y)\wedge\phi(x))\prp [y\in B].$$ A weak comprehension $\cmp\phi$ of $\phi\prp [x\in A]$ is given by $$\ec{\pi_1(w)\in A\ [w\in \ds_{x\in A}\phi]}: \ds_{x\in A}\phi\to A.$$ Note that weak comprehensions are full: it is enough to show a proof of $\phi\leq\exists_{\cmp\phi}\top_W$ (with $W$ codomain of $\cmp\phi$) given by $$<<y,p>, <\mathsf{id}(y), \ast>>\in \ds_{w\in W}\left(\mathsf{Id}(A,\pi_1(w),y) \wedge N_1 \right) [y\in A,p\in \phi].$$ with $\ast$ the canonical element of the singleton set $N_1$. It follows by an immediate, direct calculation that weak equalizers are comprehensions. The doctrine $\F:\cat{ML}\op\longrightarrow\Cat{InfSL}$ enjoys more properties. \[mlbc\] The category is weakly cartesian closed. For every arrow $f$ in , the functor $\F_f$ has also a right adjoint $($which necessarily satisfies Beck-Chevalley condition$)$. The category is weakly cartesian closed because given sets $A$ and $B$, a weak evaluation is defined via the dependent product as $$\ec{\pi_1(z)(\pi_2(z))\in B\ \left[z\in\left(\dpd_{x\in A} B\right)\times A\right]}: \left(\dpd_{x\in A} B\right)\times A\to B.$$ It is only a weak evalutation because inhabitation of $\mathsf{Id}(B(x),f(x),g(x))\ [x\in A]$ cannot ensure inhabitation of $\mathsf{Id}(\dpd_{x\in A}B(x),\lambda x.f(x),\lambda x. g(x))$, see [@disttheshof]. A right adjoint to substitution is given by the dependent product set. In fact, the doctrine $\F$ is isomorphic to the doctrine $\Psi^{\cat{ML}}:\cat{ML}\op\longrightarrow\Cat{InfSL}$, constructed as in \[weaklhy\] from the category . The isomorphism $(\Id{\cat{ML}},h):\F\to\Psi^{\cat{ML}}$ has the identity functor in the first component and, for $A$ a closed set, the functor $h(A):\F(A)\to\Psi^{\cat{ML}}(A)$ maps the equivalence class $\ec{\phi(x)\textit{ prop\/}\ [x\in A]}$ to the equivalence class represented by the arrow $$\label{ir} \ec{\pi_1(w)\in A\ [w\in \ds_{x\in A}\phi(x)]}: \ds_{x\in A}\phi(x)\to A$$ which has codomain $A$. Note that a proof-term $t(y)\in \psi\ [x\in A, y\in \phi]$ produces a map and a commutative triangle $$\xymatrix@C=7em{{\ds_{x\in A}\phi} \ar[rr]^{\ec{<\pi_1(w),t(\pi_2(w))>\in\ds_{x\in A}\psi\ [w\in\ds_{x\in A}\in \phi]}} \ar[rd]_{\ec{\pi_1(w)\in A\ [w\in \ds_{x\in A}\phi]}\kern5em}&& {\ds_{x\in A}\psi} \ar[ld]^{\kern5em\ec{\pi_1(w)\in A\ [w\in \ds_{x\in A}\psi]}}\\ & A}$$ which shows that the assignment for $h(A)$ given in (\[ir\]) does not depend on the choice of representatives and it extends to a functor. Moreover, one can easily check that the functor $k(A):\Psi^{\cat{ML}}(A)\to\F(A)$ mapping $f(y)\in A\ [y\in B]$ to $\ds_{y\in B}\mathsf{Id}(A,f(y),x)$ for $x\in A$ provides an inverse to $h(A)$. \[mlttequiv\] The base category $\cat{Q}_{\F}$ of the elementary quotient completion $\Q{\F}$ of $\F:\cat{ML}\op\longrightarrow\Cat{InfSL}$ is essentially equivalent to the model of total setoids over Martin-Löf type theory as given in [@notepal]. The quotient completion in the two-level minimalist foundation {#twl} -------------------------------------------------------------- One of the motivations for developing the notion of elementary quotient completion is to describe abstractly the “quotient model” given in [@m09]. That model is built over a dependent type theory acting as the intensional level of the two-level minimalist foundation for constructive mathematics in [@m09]. It is used to give an interpretation of the extensional level of the foundation—denoted for [*extensional minimal type theory*]{}—in the intensional level—denoted for [*minimal type theory*]{}. Quite similar to that built over Martin-L[ö]{}f’s type theory, also this model is based on the idea of total setoids à la Bishop, but it appears to have different properties from that. For instance, the setoid model built over a set-theoretic fragment of is not exact as we note in the following—hence it cannot be an exact completion. Recall from [@m09] that is a predicative version of Coquand’s Calculus of Constructions , see [@tc90]. Indeed, as , it resembles a [*many sorted logic*]{} with propositions given primitively and depending on sorts. Propositions are equipped with proof-terms and sorts include dependent sets as in Martin-L[ö]{}f’s type theory. But, contrary to , in there are two types of sorts on which propositions depend: these are sets and collections, where the first are included in the latter; this distinction resembles the usual distinction between sets and classes in NBG and is instrumental to represent the power collection of a set which, from a predicative point of view, need not be a set. However the whole can be interpreted in if one interprets both sets and collections in as sets in . Indeed, if in collections were identified with sets then one would just get . Therefore, many properties of the quotient model over can be extended to the corresponding quotient model over . Corresponding to sets and collections in we consider two doctrines: one is a sub-doctrine of the other. The larger has the base category formed by collections with their typed terms, the other has the base category restricted to the full subcategory on sets. The “syntactic category of collections in ” $\cat{CM}$ is constructed like the previous example. It has closed collections of as objects. An arrow $\ec{t\in B\ [x\in A]}:A\to B$ from the closed collection $A$ to the closed collection $B$ is an equivalence class of proof-terms in where $t\in A\ [x\in B]$ is equivalent to $t'\in A\ [x\in B]$ if there is a proof-term $$p\in \mathsf{Id}(A,t,t')\ [x\in B]$$ in . The composition of $\ec{t\in B\ [x\in A]}:A\to B$ and $\ec{s(y)\in C\ [y\in B]}:B\to C$ is given as the equivalence class $$\ec{s(t)\in C\ [x\in A]}:A\to C$$ with identity arrows given by $\ec{x\in A\ [x\in A]}:A\to A$. The “syntactic category of sets in ” is the full subcategory of on the closed sets. Like in the previous example, the two categories just introduced have weak limits. \[synmtt\] The categories $\cat{CM}$ and $\cat{SM}$ are cartesian and with weak equalizers. The proof is analogous to that of \[synmltt\]. We now define a doctrine on each syntactic category defined on in a similar way to that for Martin-L[ö]{}f’s type theory. The functor $$\Fm:\cat{CM}\op\longrightarrow\mathsf{InfSl}$$ is defined on a closed collection $A$ as the partial order consisting of equivalence classes $\ec{\phi\prp [x\in A]}$ of predicates : in depending on $A$ with respect to equiprovability, [*i.e.*]{} $\phi\prp [x\in A]\sim\phi'\prp [x\in A]$ if there is a proof of $\phi\leftrightarrow\phi'\prp [x\in A]$ in $\ec{\phi\prp [x\in A]}\leq\ec{\psi\prp [x\in A]}$ : if there is a proof-term $r\in\psi\ [x\in A, w\in \phi]$ in . The action of the functor on arrows in is given by substitution. The doctrine $$\Fs:\cat{SM}\op\longrightarrow \mathsf{InfSl}$$ is defined on a closed set $A$ as the partial order consisting of equivalence classes $\ec{\phi\prps [x\in A]}$ : of propositions $\phi(x)\prps [x\in A]$ depending on $A$, [*i.e.*]{} propositions closed only under quantification on sets, with respect to equiprovability in $\ec{\phi\prps [x\in A]}\leq\ec{\phi'\prps [x\in A]}$ : if there is a proof-term of $r\in \phi'\ [x\in A, w\in \phi]$ in . The action of the functor on arrows in is given by substitution. \[mttd\] The two doctrines $\Fm:\cat{CM}\op\longrightarrow\mathsf{InfSl}$ and $\Fs:\cat{SM}\op\longrightarrow\mathsf{InfSl}$ are existential elementary, implicational and universal, with full weak comprehensions and comprehensive weak equalizers. Moreover, the doctrine $\Fs$ is weakly cartesian closed. The proof for each is analogous to that for Martin-L[ö]{}f’s type theory in propositions \[mlttd\] and \[mlbc\]. Similarly to \[mlttequiv\], there is an essential equivalence between the completion $\Q{\Fm}$ of the doctrine $\Fm$ and the model of extensional collections in [@m09], and there is another between the completion $\Q{\Fs}$ of the doctrine $\Fs$ and the model of extensional sets in [@m09]. Then, if one recalls that is the extensional version of with the addition of effective quotients, proof-irrelevance of propositions and extensional equality of functions, it is easy to see that there is an obvious interpretation of in the doctrine $\Q{\Fm}$ provided that a choice of the structure is given. Actually, by means of logical constructors of , following the paradigm proposed in [@LawvereF:adjif; @jacobbook], we can describe the logical theories modelled by most of the doctrines considered in this paper. For example, the notion of elementary existential doctrine with full comprehensions and comprehensive equalizers is the necessary structure to interpret the set-theoretic fragment of where sets just include the singleton set, binary products and indexed sums of a set with a proposition, and whose propositions are just closed under conjunctions, extensional equality, existential quantifiers and the truth constant. Then the extension with -quotients captures the logical theory behind existential doctrines in . \[notex\] The quotient model $\cat{Q}_{\Fs}$ of $\Fs$ for the fragment of in [@iac] provides a genuine example that the elementary quotient completion is not an exact completion. Indeed, consider that in an exact, locally cartesian closed category, (AUC) holds for any pair of objects, as shown in [@tumscs]. Now, if $\cat{Q}_{\Fs}$ were exact, then (AUC) would hold for any pair of objects in the original doctrine $\Fs$ by \[uac\]. But [@iac] shows that (AUC) does not hold for the natural numbers in $\Fs$. Knowing that (AUC) does not hold in  [@streicher], with the same argument as above we get another example of elementary quotient completion that is not exact: take $\cat{Q}_{G^{\CTT}}$ with $G^{\CTT}$ defined as $\Fs$ but within . Conclusions and future work =========================== The notion of elementary quotient completion developed in the previous sections was inspired by the need to give an abstract presentation of the quotient model used in [@m09]. This notion is more general than that of exact completion. As remarked in \[exlexcom\], an exact completion of a cartesian category with weak pullbacks is an instance of the elementary quotient completion construction. On the other hand, as remarked in \[notex\], there are elementary quotient completions which are not exact. Relevant instances of elementary quotient completion are used to formalize mathematics within an intensional type theory. They are applied to turn a theory with [*intensional*]{} and [*weak*]{} structures into one with [*extensional*]{} and [*strong*]{} ones. For example, the category $\cat{Q}_{\Fs}$ of extensional sets over the intensional theory , as well as the category $\cat{Q}_{\F}$ of setoids over Martin-L[ö]{}f’s type theory, are cartesian closed, while the corresponding categories of sets of and of Martin-L[ö]{}f’s type theory are not. Moreover the doctrines $\F$, $\Fs$ and $\Fm$ associated to and to Martin-L[ö]{}f’s type theory satisfy only weak comprehension, because their propositions may be equipped with more than one proof. But it is worth mentioning that to view an elementary quotient completion, as that over in [@m09], as a model of an extensional theory as in [@m09], one needs to solve a coherence problem. Indeed, the base category of an elementary quotient completion, as well as of an exact completion in [@CarboniA:freecl], does not provide explicit construction for the categorical structure used to interpret logical sorts. In [@m09] we solved the problem by using an [*ad hoc*]{} coherence theorem to select the model structure necessary to interpret the syntax of in the elementary quotient completion over . Since that theorem heavily relies on the fact that the syntax of is defined inductively, a possible direction of further investigation is to explore how to interpret a logical theory corresponding to doctrines in in (the doctrine $\Q{P}$ of) an elementary quotient completion of a doctrine $P$ in . In future work we shall investigate the case of completing an existential elementary doctrine with respect to quotients with a more relaxed notion of arrow and compare it with the elementary quotient completion introduced here. #### Acknowldgements {#acknowldgements .unnumbered} The authors would like to acknowledge Giovanni Sambin and Ruggero Pagnan for very many, useful and intensive discussions on the subject. Discussions with Fabio Pasquali and Lorenzo Malatesta were also very useful for the second author. , G., V. [Capretta]{}, and O. [Pons]{} (2003). Setoids in type theory.  [*13*]{}(2), 261–293. Special issue on [“Logical frameworks and metalanguages”]{}. Birkedal, L., A. Carboni, G. Rosolini, and D. Scott (1998). Type theory via exact categories. In V. Pratt (Ed.), [*Proc. 13th [S]{}ymposium in [L]{}ogic in [C]{}omputer [S]{}cience*]{}, Indianapolis, pp.  188–198. I.E.E.E. Computer Society. Carboni, A. (1995). Some free constructions in realizability and proof theory.  [*103*]{}, 117–148. Carboni, A. and R. C. Magno (1982). The free exact category on a left exact one.  [*33*]{}(A), 295–301. Carboni, A. and G. Rosolini (2000). Locally cartesian closed exact completions.  [*154*]{}, 103–116. Carboni, A. and E. Vitale (1998). Regular and exact completions.  [*125*]{}, 79–117. , T. (1990). Metamathematical investigation of a calculus of constructions. In P. [Odifreddi]{} (Ed.), [*Logic in Computer Science*]{}, pp. 91–122. Academic Press. , P. (1996). Internal type theory. In [*TYPES ’95*]{}, Volume 1158 of [*LNCS*]{}, pp.  120–134. Freyd, P. and A. Scedrov (1991). . North Holland Publishing Company. Grothendieck, A. (1971). Catégories fibrées et descent ([E]{}xposé [VI]{}). In A. Grothendieck (Ed.), [*Revêtements etales et groupe fondamental - [SGA]{} 1*]{}, Number 224 in Lecture Notes in Mathematics, pp. 145–194. Springer-Verlag. , M. (1997). Distinguished Dissertations. Springer-Verlag. Hughes, J. and B. Jacobs (2002). Factorization systems and fibrations: [T]{}oward a fibred birkhoff variety theorem.  [ *11*]{}, [\#]{}. , B. (1999). , Volume 141 of [*Studies in Logic*]{}. Elsevier. Lambek, J. and P. Scott (1986). . Cambridge University Press. Lawvere, F. W. (1969a). Adjointness in foundations.  [*23*]{}, 281–296. Lawvere, F. W. (1969b). Diagonal arguments and cartesian closed categories. In [*Category Theory, Homology Theory and their Applications, II (Battelle Institute Conference, Seattle, Wash., 1968, Vol. Two)*]{}, pp. 134–145. Springer-Verlag. Lawvere, F. W. (1970). Equality in hyperdoctrines and comprehension schema as an adjoint functor. In A. Heller (Ed.), [*Proc. [N]{}ew [Y]{}ork [S]{}ymposium on [A]{}pplication of [C]{}ategorical [A]{}lgebra*]{}, pp.  1–14. Amer.[M]{}ath.[S]{}oc. Lawvere, F. W. and R. Rosebrugh (2003). . Cambridge University Press. , M. (2005). Modular correspondence between dependent type theories and categories including pretopoi and topoi.  [*15*]{}(6), 1089–1149. , M. (2009). A minimalist two-level foundation for constructive mathematics.  [*160*]{}(3), 319–354. , M. (2010). Consistency of the minimalist foundation with [C]{}hurch thesis and [B]{}ar [I]{}nduction. submitted. , M. and G. [Sambin]{} (2005). . In [L. Crosilla and P. Schuster]{} (Ed.), [*From Sets and Types to Topology and Analysis: Practicable Foundations for Constructive Mathematics*]{}, Number 48 in [Oxford Logic Guides]{}, pp.  91–114. [Oxford University Press]{}. Makkai, M. and G. Reyes (1977). , Volume 611 of [*Lecture Notes in Mathematics*]{}. Springer-Verlag. ordstr[ö]{}m, B., K. [P]{}etersson, and J. [S]{}mith (1990). Clarendon Press. , E. (2005). Bishop’s set theory. Slides for lecture at the TYPES summer school. Pavlović, D. (1996). Maps [II]{}: chasing proofs in the [L]{}ambek-[L]{}awvere logic.  [*4*]{}(2), 159–194. , G. and S. [Valentini]{} (1998). Building up a toolbox for [M]{}artin-[L]{}[ö]{}f’s type theory: subset theory. In G. [Sambin]{} and J. [Smith]{} (Eds.), [*Twenty-five years of constructive type theory*]{}, pp.  221–244. Oxford University Press. , T. (1992). Independence of the induction principle ad the axiom of choice in the pure calculus of constructions.  [*103*]{}(2), 395–408. Taylor, P. (1999). . Cambridge University Press. van Oosten, J. (2008). , Volume 152. North-Holland Publishing Co. [^1]: Dipartimento di Matematica Pura ed Applicata, Università di Padova, via Trieste 63, 35121 Padova, Italy, email: `[email protected]` [^2]: DIMA, Università di Genova, via Dodecaneso 35, 16146 Genova, Italy, email: `[email protected]`. Project MIUR-PRIN McTAFI provided support for the research presented in the paper. [^3]: Here and in the sequel we write the action of a doctrine $P$ on an arrow as $P_f$. [^4]: The empty list is included. [^5]: We shall employ a vector notation for lists of terms in the language as well as for simultaneous substitutions such as $[\vec t/\vec y]$ in place of $[t_1/y_1,\ldots,t_m/y_m]$. We shall also extend vectorial notation to conjunctions and quantifiers writing $\vec t=\vec s$ for the conjunction $t_1=s_1\Land\ldots\Land t_\ell=s_\ell$, provided the lists $\vec t$ and $\vec s$ are the same length, and writing $\Exists\vec x.$ instead of $\Exists x_1.\ldots\Exists x_n.$. [^6]: We shall denote an equivalence class with representative $x$ as $\ec{x\kern.2ex}$ in order to leave plain square brackets available for other situations. [^7]: We write $\Delta_A$ and $f\times f'$ respectively for the map $<\id{A},\id{A}>$ and for the map $<f\circ\pr_1,f'\circ\pr_2>:A\times A'\to B\times B'$, provided $f:A\to B$ and $f':A'\to B'$. [^8]: Note that the functors $F,G$ are thought from ${\cat{C}\op}$ to ${\cat{D}\op}$. [^9]: When necessary to distinguish between the notions in definitions \[compd\] and \[wcd\], we shall refer to one as in \[compd\] with the further attribute . [^10]: To represent a term depending on a variable $x$, we shall indifferently use the metavariable $t$ or $t(x)$. We use the latter when we need to refer to a substitution of a variable $x$ that may appear in $t$.

ArXiv

--- abstract: 'We show that a certain weighted mean of the Liouville function $\lambda(n)$ is negative. In this sense, we can say that the Liouville function is negative “on average”.' address: - 'Mathematical Sciences Institute, Australian National University, Canberra, ACT 0200, Australia' - 'Langebuorren 49, 9074 CH Hallum, The Netherlands (Formerly at CWI, Amsterdam ) ' author: - 'Richard P. Brent$\,\,$' - '$\,\,$Jan van de Lune' title: 'A note on Pólya’s observation concerning Liouville’s function' --- Introduction ============ For $n\in {{\mathbb N}}$ let $n=\prod_{p|n} p^{e_p(n)}$ be the canonical prime factorization of $n$ and let $\Omega(n):=\sum_{p\mid n} e_p(n)$. Here (as always in this paper) $p$ is prime. Thus, $\Omega(n)$ is the total number of prime factors of $n$, counting multiplicities. For example: $\Omega(1)=0$, $\Omega(2)=1$, $\Omega(4)=2$, $\Omega(6)=2$, $\Omega(8)=3$, $\Omega(16)=4$, $\Omega(60)=4$, etc. Define Liouville’s multiplicative function $\lambda(n)=(-1)^{\Omega(n)}$. For example $\lambda(1)=1$, $\lambda(2)=-1$, $\lambda(4)=1$, etc. The Möbius function $\mu(n)$ may be defined to be $\lambda(n)$ if $n$ is square-free, and $0$ otherwise. It is well-known, and follows easily from the Euler product for the Riemann zeta-function $\zeta(s)$, that $\lambda(n)$ has the Dirichlet generating function $$\sum_{n=1}^\infty \frac{\lambda(n)}{n^s} = \frac{\zeta(2s)}{\zeta{(s)}}$$ for $\Re(s) > 1$. This provides an alternative definition of $\lambda(n)$. Let $L(n) := \sum_{k \le n}\lambda(k)$ be the summatory function of the Liouville function; similarly $M(n) := \sum_{k \le n}\mu(k)$ for the Möbius function. The topic of this note is closely related to Pólya’s conjecture [@Polya 1919] that $L(n) \le 0$ for $n\ge2$. Pólya verified this for $n \le 1500$ and Lehmer [@Lehmer 1956] checked it for $n \le 600\,000$. However, Ingham [@Ingham 1942] cast doubt on the plausibility of Pólya’s conjecture by showing that it would imply not only the Riemann Hypothesis and simplicity of the zeros of $\zeta(s)$, but also the linear dependence over the rationals of the imaginary parts of the zeros $\rho$ of $\zeta(s)$ in the upper half-plane. Ingham cast similar doubt on the Mertens conjecture $|M(n)| \le \sqrt{n}$, which was subsequently disproved in a remarkable *tour de force* by Odlyzko and te Riele [@OR 1985]. More recent results and improved bounds were given by Kotnik and te Riele [@KR 2006]; see also Kotnik and van de Lune [@KL 2004]. In view of Ingham’s results, it was no surprise when Haselgrove showed [@Haselgrove 1958] that Pólya’s conjecture is false. He did not give an explicit counter-example, but his proof suggested that $L(u)$ might be positive in the vicinity of $u \approx 1.8474 \times 10^{361}$. Sherman Lehman [@Lehman 1960] gave an algorithm for calculating $L(n)$ similar to Meissel’s [@Meissel 1885] formula for the prime-counting function $\pi(x)$, and found the counter-example $L(906\,180\,359) = +1$. Tanaka [@Tanaka 1980] found the smallest counter-example $L(n) = +1$ for $n = 906\,150\,257$. Walter M. Lioen and Jan van de Lune \[*circa* 1994\] scanned the range $n \le 2.5 \times 10^{11}$ using a fast sieve, but found no counter-examples beyond those of Tanaka. More recently, Borwein, Ferguson and Mossinghoff [@BFM 2008] showed that $L(n) = +1\,160\,327$ for $n = 351\,753\,358\,289\,465$. Humphries [@Humphries1; @Humphries2] showed that, under certain plausible but unproved hypotheses (including the Riemann Hypothesis), there is a limiting logarithmic distribution of $L(n)/\sqrt{n}$, and numerical computations show that the logarithmic density of the set $\{n\in{{\mathbb N}}| L(n) < 0\}$ is approximately $0.99988$. Humphries’ approach followed that of Rubinstein and Sarnak [@RS], who investigated “Chebyshev’s bias” in prime “races”. Here we show in an elementary manner, and without any unproved hypotheses, that $\lambda(n)$ is (in a certain sense) “negative on average”. To prove this, all that we need are some well-known facts about Mellin transforms, and the functional equation for the Jacobi theta function (which may be proved using Poisson summation). Our main result is: \[thm:1\] There exists a positive constant $c$ such that for every (fixed) $N\in{{\mathbb N}}$ $$\sum_{n=1}^\infty \frac{\lambda(n)}{e^{n\pi x}+1}=-\frac{c}{\sqrt{x}} +\frac12+O(x^N) \qquad \text{as}\quad x\downarrow 0.$$ Thus, a weighted mean of $\{\lambda(n)\}$, with positive weights initially close to a constant $(1/2)$ and becoming small for $n \gg 1/x$, is negative for $x < x_0$ and tends to $-\infty$ as $x \downarrow 0$. In the final section we mention some easy results on the Möbius function $\mu(n)$ to contrast its behaviour with that of $\lambda(n)$. Proof of Theorem \[thm:1\] ========================== We prove Theorem \[thm:1\] in three steps. *Step 1.* For $x > 0$, $$\sum_{n=1}^\infty \frac{\lambda(n)}{e^{n\pi x}-1}=\phi(x) = \frac{\theta(x) -1}{2}\,,$$ where $$\phi(x):=\sum_{k=1}^\infty e^{-k^2\pi x}, \qquad \theta(x):= \sum_{k \in {{\mathbb Z}}} e^{-k^2\pi x}.$$ *Step 2.* For $x > 0$, $$\sum_{n=1}^\infty \frac{\lambda(n)}{e^{n\pi x}+1}=\phi(x)-2\phi(2x).$$ *Step 3.* Theorem \[thm:1\] now follows from the functional equation $$\theta(x) = \frac{1}{\sqrt{x}}\,\theta\left(\frac{1}{x}\right)$$ for the Jacobi theta function $\theta(x)$.  \ (1) In the following, we assume that $\Re(s) > 1$, so the Dirichlet series and integrals are absolutely convergent, and interchanging the orders of summation and integration is easy to justify. As mentioned above, it is well-known that $$\sum_{n=1}^\infty\frac{\lambda(n)}{n^s} =\prod_p\left({1+p^{-s}}\right)^{-1} =\prod_p\frac{1-p^{-s}} {1-p^{-2s}} = \frac{\zeta(2s)}{\zeta(s)}.$$ Define $$f(x):=\sum_{n=1}^\infty \frac{\lambda(n)}{e^{n x}-1},\qquad (x>0).$$ We will use the well known fact that if two sufficiently well-behaved functions (such as ours below) have the same Mellin transform then the functions are equal. The Mellin transform of $f(x)$ is $$\begin{aligned} {F}(s) &:=& \int_0^\infty f(x)x^{s-1}\,{\rm d}x=\int_0^\infty \sum_{n=1}^\infty \frac{\lambda(n)}{e^{nx}-1}x^{s-1}\,{\rm d}x\\ &=&\sum_{n=1}^\infty \lambda(n)\int_0^\infty\frac{x^{s-1}}{e^{nx}-1} \,{\rm d}x =\Bigl(\sum_{n=1}^\infty\frac{\lambda(n)}{n^s}\Bigr)\times \int_0^\infty\frac{x^{s-1}}{e^{x}-1}\,{\rm d}x\\ &=&\frac{\zeta(2s)}{\zeta(s)}\times \zeta(s)\Gamma(s) =\zeta(2s)\Gamma(s). \end{aligned}$$ We also have $$\begin{aligned} \int_0^\infty \phi\Bigl(\frac{x}{\pi}\Bigr)x^{s-1}\,{\rm d}x &=&\int_0^\infty \Bigl(\sum_{n=1}^\infty e^{-n^2x}\Bigr)x^{s-1}\,{\rm d}x\\ &=&\Bigl(\sum_{n=1}^\infty \frac{1}{n^{2s}}\Bigr)\times\int_0^\infty e^{-x}x^{s-1}\,{\rm d}x= \zeta(2s)\Gamma(s), \end{aligned}$$ so the Mellin transforms of $f(x)$ and of $\phi(x/\pi)$ are identical. Thus $f(x)=\phi({x}/{\pi})$. Replacing $x$ by $\pi x$, we see that $$\sum_{n=1}^\infty \frac{\lambda(n)}{e^{n\pi x}-1}=\sum_{k=1}^\infty e^{-k^2\pi x},$$ completing the proof of step (1). \(2) Observe that $$\frac{1}{e^{n\pi x}+1}=\frac{1}{e^{n\pi x}-1}- \frac{2}{e^{2n\pi x}-1},$$ so, from step (1), $$\sum_{n=1}^\infty \frac{\lambda(n)} {e^{n\pi x}+1}=\phi(x)-2\phi(2x).$$ \(3) Using the functional equation for $\theta(x)$, we easily find that $$\phi(x)-2\phi(2x)=-\frac{c}{\sqrt{x}}+\frac12+ \frac{1}{\sqrt{x}}\Bigl(\phi\Bigl(\frac{1}{x}\Bigr)-\sqrt{2}\, \phi\Bigl(\frac{1}{2x}\Bigr)\Bigr)$$ with $c={(\sqrt{2}-1)}/{2} > 0$, proving our claim, since the “error” term is bounded by $\phi(1/x)/\sqrt{x} \sim \exp(-\pi/x)/\sqrt{x} = O(x^N)$ as $x \downarrow 0$ (for any fixed exponent $N$). Remarks on the Möbius function ============================== We give some further applications of the identity $$\tag{$*$}\label{eq:starred} \frac{1}{z+1}=\frac{1}{z-1}-\frac{2}{z^2-1}$$ that we used (with $z = e^{n\pi x}$) in proving step (2) above. \[lemma:2\] For $|x| < 1$, we have $${\sum_{n=1}^\infty \mu(n)\frac{x^n}{x^n+1}} = x - 2x^2.$$ Assume that $|x| < 1$. It is well known that $${\sum_{n=1}^\infty \mu(n)\frac{x^n}{1-x^n} = x},$$ in fact this “Lambert series” identity is equivalent to the Dirichlet series identity $\sum \mu(n)/n^s = 1/\zeta(s)$. Writing $y = 1/x$, we have $$\sum_{n=1}^\infty \frac{\mu(n)}{y^n-1} = 1/y.$$ If follows on taking $z = y^n$ in our identity (\[eq:starred\]) that $$\sum_{n=1}^\infty \frac{\mu(n)}{y^n+1}= \sum_{n=1}^\infty \frac{\mu(n)}{y^n-1}- 2\sum_{n=1}^\infty \frac{\mu(n)}{y^{2n}-1} = y^{-1} - 2y^{-2}.$$ Replacing $y$ by $1/x$ gives the result. $${\sum_{n=1}^\infty \frac{\mu(n)}{2^n+1}} = 0.$$ Take $x = 1/2$ in Lemma \[lemma:2\]. If follows from Lemma \[lemma:2\] that $$\lim_{x\uparrow1} \sum_{n=1}^\infty \mu(n)\frac{x^n}{x^n+1} =-1,$$ so that one might say that in this sense $\mu(n)$ is negative on average. However, this is much weaker than what we showed in Theorem \[thm:1\] for $L(n)$, where the corresponding sum tends to $-\infty$. The “complex-analytic” reason for this difference is that $\zeta(2s)/\zeta(s)$ has a pole (with negative residue) at $s = 1/2$, but $1/\zeta(s)$ is regular at $s=1$. [99]{} <span style="font-variant:small-caps;">P. Borwein, R. Ferguson and M. J. Mossinghoff</span>, *Sign changes in sums of the Liouville function*, Math. Comp. **77** (2008), 1681–1694. <span style="font-variant:small-caps;">C. B. Haselgrove</span>, *A disproof of a conjecture of Pólya*, Mathematika **5** (1958), 141–145. <span style="font-variant:small-caps;">P. B. Humphries</span>, *The Summatory Function of Liouville’s Function and Pólya’s Conjecture*, Honours Thesis, Department of Mathematics, The Australian National University, Canberra, October 2010. <span style="font-variant:small-caps;">P. B. Humphries</span>, *The Distribution of Weighted Sums of the Liouville’s Function and Pólya’s Conjecture*, arXiv:1108.1524, August 2011. <span style="font-variant:small-caps;">A. E. Ingham</span>, *On two conjectures in the theory of numbers*, Amer. J. of Mathematics **64** (1942), 313–319. <span style="font-variant:small-caps;">T. Kotnik and J. van de Lune</span> *On the order of the Mertens function*, Experimental Mathematics **13**:4 (2004), 473–481. <span style="font-variant:small-caps;">T. Kotnik and H. J. J. te Riele</span> *The Mertens conjecture revisited*, Proc. ANTS 2006, Lecture Notes in Computer Science **4076** (2006), 156–167. <span style="font-variant:small-caps;">R. S. Lehman</span>, *On Liouville’s function*, Math. Comp. **14** (1960), 311–320. <span style="font-variant:small-caps;">D. H. Lehmer</span>, *Extended computation of the Riemann zeta function*, Mathematika **3** (1956), 102–108. <span style="font-variant:small-caps;">E. D. F. Meissel</span>, *Berechnung der Menge von Primzahlen, welche innerhalb der ersten Milliarde naturlicher Zahlen vorkommen*, Math. Ann. **25** (1885), 251-257. <span style="font-variant:small-caps;">A. M. Odlyzko and H. J. J. te Riele</span>, *Disproof of the Mertens conjecture*, J. für die reine und angewandte Mathematik **357** (1985), 138–160. <span style="font-variant:small-caps;">G. Pólya</span>, *Verschiedene Bemerkungen zur Zahlentheorie*, Jahresbericht der deutschen Math.-Vereinigung **28** (1919), 31–40. <span style="font-variant:small-caps;">M. Rubinstein and P. Sarnak</span>, *Chebyshev’s bias*, Experimental Mathematics **3** (1994), 173–197. <span style="font-variant:small-caps;">M. Tanaka</span>, *A numerical investigation on cumulative sum of the Liouville function*, Tokyo J. Math. **3** (1980), 187–189.

ArXiv

--- abstract: 'I describe and present the results of a newly developed fitting methodology optimized for very long time series. The development of this new methodology was motivated by the fact that we now have more than half a decade of nearly uninterrupted observations by GONG and MDI, with fill factors as high as 89.8% and 82.2% respectively. It was recently prompted by the availability of a 2088-day-long time series of spherical harmonic coefficients produced by the MDI team. The fitting procedure uses an optimal sine-multi-taper spectral estimator – whith the number of tapers based on the mode linewidth, the complete leakage matrix ([[*i.e.*]{}]{}, horizontal as well as vertical components), and an asymmetric mode profile to fit simultaneously all the azimuthal orders with individually parametrized profiles. This method was applied to 2088-day-long time series of MDI and GONG observations, as well as 728-day-long subsets, and for spherical harmonic degrees between 1 and 25. The values resulting from these fits are inter-compared (MDI versus GONG) and compared to equivalent estimates from the MDI team and the GONG project. I also compare the results from fitting the 728-day-long subsets to the result of the 2088-day-long time series. This comparison shows the well known change of frequencies with solar activity – and how it scales with a nearly constant pattern in frequency and $m/\ell$. This comparison also shows some changes in the mode linewidth and the constancy of the mode asymmetry.' author: - 'S. G. Korzennik' title: | A Mode Fitting Methodology\ Optimized for Very Long Time Series --- Motivation ========== Several methodologies for the precise measurements of p-mode frequencies have been developed over the years . These techniques have evolved as the quality of the observations have improved over the years. More recently, such methods have produced frequency and rotational frequency splitting tables – typically based on 36-day-long and 72-day-long time series for GONG and MDI data respectively – that have been the basis for countless inferences of the structure and dynamics of the solar interior. Some eight years after the deployment of the GONG network and the launch of the SOHO spacecraft, we now have access to more than half a decade of nearly uninterrupted times series of solar observations. The MDI team has recently produced a 2088-day-long time series of spherical harmonic coefficients at a one minute sampling interval, namely a time series in excess of 3 millions points. For low-degree and low-frequency modes the availability of such a very long time series provides a unique opportunity to develop a mode fitting methodology that exploits the properties of such an exquisite data set. I present here a new methodology that I have developed to fit modes using very long time series. This methodology includes an optimal multi-taper spectral estimator, the complete leakage matrix ([[*i.e.*]{}]{}, horizontal as well as vertical components), an asymmetric profile and the simultaneous fitting of individual profiles at all the azimuthal orders ($m$) for a given ($n$, $\ell$) mode. The contamination by nearby modes ($n'$, $\ell'$) within the fitting range is also included. Since simultaneous fitting on these contaminants is impractical, the fitting procedure is iterated – [[*i.e.*]{}]{}, the characteristics of the contaminants used in the fitting correspond to the values fitted at the previous iteration. The primary goal for this work was to extend the mode fitting to low-order and low-degree modes. The low degree modes carry information on the structure and dynamics of the deep interior while the low-order low-degree modes are long lived ([[*i.e.*]{}]{}, show very narrow peaks) and can thus be measured with very high precision. Since their amplitude is small, they will emerge above the background noise only for very long time series. The secondary goals were to fit simultaneously all the modes instead of using the traditional polynomial expansion in ${m}/{\ell}$, while using the complete leakage matrix, as well as to include the well known mode profile asymmetry. I describe the data sets I used in Section 2, and present the details of the methodology in Section 3. In Section 4 I present the results of fitting both MDI and GONG 2088-day-long co-eval time series. These results are compared to equivalent estimates from the MDI team and the GONG project. I also present the results of fitting two-year-long subsets of that time series for both data sets, and how the fitted parameters change with epoch and henceforth with solar activity. Data Sets Used ============== The work presented here is based on times series of spherical harmonic coefficients computed from full-disk observations by the MDI and GONG instruments and limited to $\ell \le 25$. The MDI 2088-day-long time series was used to develop the methodology. It was then applied to the co-eval GONG data set. Both data sets were also subdivided in five 728-day-long overlapping segments (actually 1,048,576 minute long, [[*i.e.*]{}]{}: $1024^2$), each offset by some 364 days ([[*i.e.*]{}]{}, 524,288 minutes) from the previous one. The last 728-day-long segment was completed by augmenting the 2088-day-long time series by an extra 97 days ([[*i.e.*]{}]{}, 139,008 minutes) to fill it completely. The respective ranges of all the time series analyzed are shows in Table \[table:TimeRanges\]. Length Segment No From To ---------- ------------ --------------------------- --------------------------- 2088-day n/a 04/30/1996 at 23:59:30 UT 01/17/2002 at 23:59:26 UT 728-day 1 04/30/1996 at 23:59:30 UT 04/29/1998 at 04:14:30 UT 728-day 2 04/30/1997 at 02:07:30 UT 04/28/1999 at 06:22:30 UT 728-day 3 04/29/1998 at 04:15:30 UT 04/26/2000 at 08:30:30 UT 728-day 4 04/28/1999 at 06:23:30 UT 04/25/2001 at 10:38:30 UT 728-day 5 04/26/2000 at 08:31:30 UT 04/24/2002 at 12:46:30 UT : Respective time ranges for the time-series analyzed \[table:TimeRanges\] The fill factors of the 2088-day-long time series, before detrending, are 89.8% and 82.2% for MDI and GONG observations respectively. The time series of spherical harmonic coefficients were detrended using a 21-minute-long running mean and clipped with a 3$\sigma$ rejection threshold. The rejection reduced the fill factor by less than one percent for the MDI time series and by less than two percents for the GONG time series. For the MDI observations, discontinuities resulting from instrumental reconfigurations were removed by detrending over sub-intervals delineated by these reconfigurations. Methodology =========== The key original aspects of this new methodology are the use of an [*optimal*]{} multi-tapered spectral estimator, the simultaneous fitting of all $m$ spectra and the use of an asymmetric profile. The fitted model includes both radial and horizontal leakage components and the leakage of nearby modes inside the fitting range. Spectral Estimator ------------------ The $N^{\rm th}$ order sine multi-taper, defined as $$P_{\ell, m}^{(N)}(\nu) = \sum_{k=1}^{N} \left|{\rm FFT}\left[ \sin(\frac{\pi\,k\,i}{M+1}) c_{\ell, m}(t_i) \right]\right|^2$$ was used as power spectrum estimator, where $c_{\ell, m}(t_i)$ represents the spherical harmonic coefficient for $\ell$ and $m$ at the time $t_i$, and $M$ is the length of the time series. Sine multi-tapered power spectra were computed with an oversampling factor of 2, and for a pre-selected list of number of tapers. For the 2088-day-long time series, I used 5, 9, 21, 45 and 91 tapers, while for the 728-day-long time series I used 3, 7, 15, 31 and 63 tapers. The choice of the optimal number of tapers amongst that list is explained below. Mode Fitting ------------ The fitting procedure uses a downhill simplex minimization [@Nelder+Mead:65; @PressEtAl:92] to fit simultaneously, and in the least-squares senses, all the multiplets for a given mode, [[*i.e.*]{}]{} all $m$ for a given $n, \ell$. The fitting is done iteratively, over a frequency range limited to only encompass the closest spatial leaks ($\delta m = \pm 2$, $\delta \ell = 0$), using the optimal multi-taper power spectrum and fitting only for the modes whose amplitudes are above some prescribed threshold. The fitted profile is an asymmetric Lorentzian: $$P_{n, \ell,m}(\nu) = \frac{1+\alpha_{n,\ell} (x_{n,\ell,m}-\alpha_{n,\ell}/2)}{x_{n,\ell,m}^2+1}$$ where $$x_{n,\ell,m} = \frac{\nu - \nu_{n,\ell,m}}{\Gamma_{n,\ell}/2}$$ The mode linewidth, $\Gamma_{n,\ell}$, and its asymmetry, $\alpha_{n,\ell}$, are assumed to be independent of the azimuthal order, $m$. The power spectrum is thus modeled as the superposition of the mode profile and the spatial leaks present in the fitting range: $$\begin{aligned} P_{\ell, m}(\nu) = & &A_{n,\ell,m} P_{n, \ell, m}(\nu) + B_{n, \ell, m} \\\nonumber &+& \sum_{m'} A_{n,\ell,m'}\, C(n, \ell, m'; n, \ell, m)\, P_{n, \ell, m'}(\nu) \\\nonumber &+& \sum_{n', \ell', m'} A_{n',\ell',m'}\, C(n', \ell', m'; n, \ell, m)\, P_{n', \ell', m'}(\nu)\end{aligned}$$ where $A$ represents the respective mode power amplitudes, $B$ the noise background levels and $C$ the leakage coefficients. For obvious procedural reasons, I will refer to the terms in the first summation as spatial leaks (same $n$ and $\ell$) and the ones in the second summation as mode contamination (leaks from a different $n$ and $\ell$). The sum on $m'$ is actually limited to $m'=m-2$ and $m'=m+2$ by the choice of the fitting range (see below). The sum on $n',\ell',m'$ is included only if a nearby mode ($|\ell-\ell'| \le 3$ & $|m-m'| \le 3$) fall within the widened fitting window ([[*i.e.*]{}]{}, the fitting range expanded by 40% to include the tail of nearby leaks). In the absence of mode contamination in the fitting range, there are 2$\ell$+1 amplitude, frequency, and background parameters plus one linewidth and asymmetry parameters, or $3\times(2\ell+1)+2$ coefficients to fit using $2\ell+1$ sections of spectra. By performing a simultaneous fit, the main peak as well as the spatial leaks are used to constrain the mode parameters, under the implicit assumption that the leakage coefficients are perfectly well known. For $\ell \le 25$, the mode contamination is dominated by $n-n' = \pm 1$ and $\ell-\ell' =\mp 3$, and is almost confined to the high frequency modes. When present, this mode contamination was added in the fitted model by using the mode parameters resulting from the previous iteration, and leaving them fixed. The fitting procedure is therefore repeated — [[*i.e.*]{}]{} iterated — until the mean change in the mode frequencies between iterations drops below some prescribed threshold. ### Fitting Range The fitting frequency range is set to be $\tilde\nu_{n,\ell,m} \pm \delta\nu$, where $\tilde\nu_{n,\ell,m}$ represents some estimate of the multiplet frequency at the previous fitting step (see below) and $\delta\nu$ is given by $$\delta\nu = 4\, \Gamma^{(\rm eff)}_{n, \ell} + \Delta\nu$$ where $\Gamma^{(\rm eff)}$ is the effective mode linewidth and $\Delta\nu=800$ nHz. The factor 4 ensures a good sampling of the mode profile while the additional 800 nHz is added to always include the $\delta m = \pm 2$, $\delta \ell = 0$ spatial leaks. This range is set to never be smaller than 1.2 at low frequencies and reduced, at high frequencies, to not exceed the mid-point from the frequency of the previous order and to the next order. The effective mode linewidth, $\Gamma^{(\rm eff)}_{n, \ell}$, is estimated by $$(\Gamma^{(\rm eff)}_{n, \ell})^2 = \tilde\Gamma^2_{n, \ell} + \Gamma_{r, N}^2$$ where $ \tilde\Gamma_{n, \ell}$ is some estimate of the mode linewidth (see below) and $\Gamma_{r, N}$ is the resolution of the $N^{\rm th}$ order multi-taper power spectrum, given by $$\Gamma_{r,N} = N\ \Gamma_r = \frac{N}{T}$$ and where $T$ is the time interval span by the time series. ### Optimal Multi-Taper The optimal $N^{\rm th}$ order multi-tapered power spectrum is defined as the highest order multi-taper spectrum from a pre-selected list having a resolution at least five times better than the effective mode linewidth, whenever possible. The value of $N$ is thus selected to satisfy: $$\frac{\tilde\Gamma_{n,\ell}}{5} \ge \Gamma_{r,N} = N\,\Gamma_r$$ Mainly for convenience, an estimate of the mode linewidth was used for this selection and for the determination of the mode fitting range. This estimate was computed from a polynomial parameterization as a function of frequency, based on previous published estimates [@Schou:99]. The logarithm of the mode linewidth was estimated by a 6th order polynomial in $\nu$. Figure \[figure:whichmt\] compares the estimate of the mode linewidth to the $N^{\rm th}$ order multi-tapered spectrum resolution, for the pre-selected list of values of $N$, and indicates the number of tapers selected for the fitting. ### Detectability Threshold At low frequencies, the mode amplitude becomes comparable if not smaller than the background noise all the while the mode linewidth becomes smaller than the spectral resolution, forcing the fitting procedure to hunt for small and narrow peaks. This can easily lead to confusing a noise spike with a mode peak and can lead to what is sometimes referred as [*fitting the grass*]{}. To prevent such “grass fitting”, only modes with a power amplitude 3 times greater than the root-mean-squares (RMS) of the residuals to the fit were kept. As the fitting proceeds a sanity check rejects any mode whose amplitude has dropped below that threshold. If this happens, the amplitude of the mode is set to zero and its amplitude and frequency are no longer fitted, leaving the fitting procedure to only adjust the background term to the noise. As a result of the line of sight apodization of any radial velocity observation, the near sectoral modes ($m \approx \pm\ell$) emerge above the noise background before the near zonal ones ($m \approx 0$). By not using a polynomial expansion in $m$, I end up fitting only the azimuthal degrees that are large enough, and I avoid any potential bias in the estimate of the frequency splittings based on an expansion that would be unevenly constrained. ### Fitting Procedure & Initial values At each iteration the fitting is done in steps, [[*i.e.*]{}]{} not all the parameters are adjusted simultaneously right away. The whole procedure is iterated to include progressively better and better estimates of the mode contamination. An initial value of all the mode parameters is first needed. The initial guess for the mode frequency was computed by using previous estimates of the singlet frequencies ($\nu_{\ell,m}$) and of the polynomial expansion coefficients for the frequency splittings. The frequency table was extended at low frequencies by looking at “derotated” $m$-averaged spectra while the frequency splittings table was extended with values that produced nearly “straight” derotated $m$-averaged spectra. The initial values for the linewidth and the asymmetry were computed from a polynomial parameterization in term of frequency, also based on previous estimates, and extrapolated when needed. The initial amplitude and background parameters were computed from averaging appropriate portions of the section of the spectrum being fitted and then performing a polynomial fit of these values as a function of $m$ with a 3$\sigma$ rejection of outliers. To make sure that the result is not biased by the initial guess of the parameters, the fitting steps at the very first iteration were different from the ones taken during the consecutive iterations. The first iteration included additional steps where the fitting was done on individual spectra ([[*i.e.*]{}]{}, one $m$ at a time) instead of the simultaneous fitting of all the $m$. A sanity check is performed after each step to remove from the fitting any peak with too low of an amplitude. The amplitude of such peaks are set to zero and their amplitude and frequency is no longer fitted. This check is followed by a readjustment of the fitting interval by using for the center of the interval smoothed values resulting from a polynomial fit of the frequencies as a function of $m$, with a 3$\sigma$ rejection of outliers. ### Leakage Coefficients The first leakage coefficients I used for the MDI observations were based on the direct computation of the spatial leakage from the leakage equations (see Section 3.1 of [@KorzennikEtAl:2004]). Such calculation leads to an approximation of the actual leakage coefficients as it remains oversimplified by ignoring effects like the detector finite pixel size and the on-board Gaussian weighting performed on the MDI [*Structure Program*]{} images. A more sophisticated leakage matrices computation was carried out by @Schou:99, by constructing simulated images corresponding to the line-of-sight contribution of each component of a single spherical harmonic mode. These images were then decomposed into spherical harmonic coefficients using the same numerical decomposition used to process the observations. This approach allows to include the effect of the finite pixel size of the detector and incorporates the above mentioned Gaussian weighting. It also takes into account [*de facto*]{} the effects of both the foreshortening at high lattitudes and the image apodization performed in the spatial decomposition. In both cases, the horizontal and vertical components were computed and the horizontal to vertical displacement ratio, $\beta$, taken to be the theoretical prediction. Using a simple outer boundary condition, [[*i.e.*]{}]{} that the Lagrangian pressure perturbation vanishes ($\delta p = 0$), the small amplitude oscillation equations for the adiabatic and non-magnetic case lead to an estimate of the ratio $\beta$, given by: $$\beta_{n, \ell} = \frac{G\;M_{\odot}\;L}{R_{\odot}^3\; \omega^2_{n, \ell}} = \frac{\nu_{0, \ell}^2}{\nu_{n, \ell}^2} \label{eq:beta}$$ where $G$ is the gravitational constant, $M_{\odot}$ is the solar mass, $R_{\odot}$ is the solar radius, $\omega$ is the cyclic frequency ($\omega=2\pi\nu$) and $L^2 = {\ell(\ell+1)}$. Leakage matrices were also computed by Schou (private communication, 2004) specifically for the GONG observations. The GONG project has also recently computed leakage matrices (Howe & Hill, private communication, 2004) for the GONG observations, but have only included the vertical component. I have therefore used for the fitting of the GONG data Schou’s complete leakage matrix. In all cases, since the time series are long compared to a year I used leakage matrices computed for $B_{\rm o}=0$. Figure \[fig:compare\_leaks\] compares the leakage matrix, for selected values of $\ell$ and $\beta$, resulting from my direct and approximate computation to the values computed explicitly by Schou for the MDI instrument. Figure \[fig:compare\_leaks\_gong\] compares the radial component of the leakage matrix as computed by the GONG project and by Schou, for the GONG observations. The differences seen in Figure \[fig:compare\_leaks\] should not come as a surprise: the wider leakage seen in the explicit computations is primarly the result of the Gaussian weighting. The RMS of the differences for the radial component of the GONG leakage matrices, as shown in Figure \[fig:compare\_leaks\_gong\], is less than 0.2% above $\ell=4$. The discrepancy at $\ell=1$ is due to the fact that the GONG values includes already the effect of subtracting the image average. ### Error Bars Computation Uncertainties on all the fitted parameters were estimated from the covariance matrix of the problem. This covariance matrix was estimated from the Hessian matrix [@PressEtAl:92] using numerical estimates of the second derivative of the merit function. The increments appropriate for the estimate of these derivatives were based on the size of the shrunk simplex resulting from the fitting. Results ======= Figures \[fig:example1\] to \[fig:example5\] illustrate the fitting of the MDI 2088-day-long time series for $\ell=9$ and a selection of values of $n$. It shows that for $n=4$ not all azimuthal orders emerge above the background; it illustrates with $n=7$ the case where modes and leaks are well resolved; it shows for $n=10$ a case where the leaks nearly merge with the main peak; for $n=18$ it illustrates the $n'=n+1$, $\ell'= \ell-3$ contamination and finally for $n=25$ it presents an example of fitting a high order mode where the optimal spectral estimator uses a large number of multi-tapers. I have produced frequency tables of individual multiplets based on MDI and GONG observations for 2088-day-long time series as well as five 728-day-long time series, for $1 \le \ell \le 25$. The coverage in the $\ell - \nu$ diagram is illustrated in Figure \[fig:lnu\] and corresponds to some 15,491 & 14,883 $(n,\ell,m)$ multiplets or some 595 & 554 $(n,\ell)$ singlets for MDI and GONG 2088-day-long time series respectively. The higher fill factor and lower signal-to-noise ratio (SNR) of the MDI time series has allowed me to push the fitting of that time series towards lower frequencies. Figure \[fig:compare1\] compares the results from fitting the two 2088-day-long co-eval time series, [[*i.e.*]{}]{} MDI and GONG. The frequency differences are small ($2.8 \pm 59.7$ nHz) and, as expected, increase with frequency – since the accuracy of the fitting decreases as the mode linewidth increases. The reduced frequency differences (${\delta\nu}/{\sigma_{\nu}}$) are uniform and correspond to a nearly Gaussian distribution with a mean of 0.017 and a standard deviation of 0.394. The smaller than expected value of the standard deviation suggests that my estimate of the uncertainties might be too conservative. At low frequency, the error bars on the frequency for the GONG observations are larger than the corresponding one for the MDI observations again as the results of the difference in fill factor and SNR. Figure \[fig:compare2\] compares the mode linewidth ($\Gamma_{n,\ell}$) and the asymmetry coefficients ($\alpha_{n,\ell}$), as well as the singlet mode frequencies. The mode singlet was computed by fitting a Clebsch-Gordan polynomial expansion [@Ritzwoller+Lavely:91] to the multiplets, with a 3$\sigma$ rejection. Mode linewidths and asymmetries between both data sets are nearly identical. Comparisons for the five 728-day-long segments show similar results, and are quantitatively summarized in Table \[tab:comparison\]. The larger average frequency difference seen for segment no. 2 results most likely from the data gap at the end of that time series present in the MDI observations[^1]. Since the gap is at the end of that time series, the results for MDI and GONG observations do not correspond to the same mean solar activity level. ------------------------- -------- --------- ---------- --------- -------- ------- ---------- --------- 2088-day-long $ 2.8$ $59.7 $ $ 0.017$ $0.394$ $-0.6$ $2.7$ $-0.001$ $0.003$ 728-day-long seg. no. 1 $-2.9$ $93.5 $ $-0.009$ $0.389$ $ 1.5$ $3.1$ $ 0.002$ $0.005$ [no. 2]{} $25.9$ $118.1$ $ 0.124$ $0.569$ $-1.5$ $3.6$ $-0.002$ $0.003$ [no. 3]{} $-5.7$ $113.4$ $-0.035$ $0.660$ $-1.8$ $1.3$ $-0.001$ $0.003$ [no. 4]{} $ 2.8$ $85.8 $ $ 0.013$ $0.791$ $ 0.2$ $1.1$ $-0.002$ $0.004$ [no. 5]{} $ 5.4$ $100.0$ $ 0.023$ $0.392$ $-0.1$ $2.0$ $-0.003$ $0.006$ ------------------------- -------- --------- ---------- --------- -------- ------- ---------- --------- : Quantitative Comparison, MDI versus GONG. \[tab:comparison\] Comparisons with Previous Estimates ----------------------------------- ### MDI Observations Figure \[fig:cmp\_w\_mdi1\] compares results from my fit to the MDI 2088-day-long time series to average values computed from 27 tables resulting from fitting 72-day-long times series that covers the same time span[^2] and are routinely computed by the MDI team [@Schou:99]. The averaging was weighted by the uncertainty – to reflect the relative fill factor of each period. The comparison is for singlets, since the MDI fitting procedure uses a polynomial expansion in $m$. The top panel illustrates the region of the $\ell$ – $\nu$ diagram where fitting a very long time series allows for detecting additional low order modes. The comparison of the resulting fitting uncertainty (lower right) shows – as expected – that at high frequency the uncertainty is dominated by the mode linewidth while at low frequency by the length of the time series. This figure also shows a systematic difference between frequencies, with a specific frequency dependence. One obvious reason for this difference is the fact that Schou’s fitting uses a symmetric profile while I am fitting an asymmetric one. To estimate the effect of fitting a symmetric profile to an asymmetric peak I have computed a grid of isolated asymmetric profiles and fitted them with symmetric ones. The resulting offset in frequency varies nearly linearly with the asymmetry coefficient as defined in my parameterization[^3], for a given FWHM, as shown in Figure \[fig:sym\_vs\_asym\]. The systematic error introduced by fitting an asymmetric peak with a symetric profile is thus given by $$\nu_{\rm asymmetric} - \nu_{\rm symmetric} = - \alpha\,\Gamma/2 \label{eq:dnuasym}$$ Roughly half of the frequency differences seen in Figure \[fig:cmp\_w\_mdi1\] can be explained by this model, as shown Figure \[fig:cmp\_w\_mdi2\]. The residual differences after correcting for the asymmetry using Eq. \[eq:dnuasym\], are marginaly significant, [[*i.e.*]{}]{} at the 4$\sigma$ level, and still show a systematic trend with frequency. Back in 1991, Schou (private communication) fitted one 72-day-long MDI time series[^4] using an asymmetric profile as well as a symmetric one. The comparison between his symmetric and asymmetric fitting for that time series and for modes up to $\ell = 25$ is shown in Figure \[fig:cmp\_sym\_vs\_asym\]. The frequency differences compare very well with the systematic differences seen in Figure \[fig:cmp\_w\_mdi1\] and are nearly twice as large as the values predicted by Eq. \[eq:dnuasym\]. While one could be tempted to rescale the prediction of Eq. \[eq:dnuasym\] by some [*ad hoc*]{} factor, I must point out that not only there is no rationale for this – unless either the leakage matrix or the horizontal to vertical ratio or both are substantially wrong – but also that a single factor would not match the observed differences (such factor would have to be quite different for frequencies above and below 3 mHz). Comparison of mode linewidths, amplitudes and background levels are shown in Figure \[fig:cmp\_w\_mdi3\]. The mode linewidth comparison shows a small systematic discrepancy ([[*i.e.*]{}]{}, a factor of 1.2) above 2.5 mHz, while below 2.5 mHz it indicates that the MDI estimates are too large – most likely as a result of the resolution limit of the 72-day-long time series – despite being corrected for the intrinsic frequency resolution of the time series. The mode power and background level estimates compare rather well. ### GONG Observations Figure \[fig:cmp\_w\_gong1\] compares results from my fit to the GONG 2088-day-long time series to average values based on 58 tables resulting from fitting 36-day-long times series that covers the same time span and are routinely computed by the GONG project [@HillEtAl:96]. The averaging was weighted by the uncertainty – to reflect the relative fill factor of each period. Singlets were then computed – by fitting a Clebsch-Gordan polynomial expansion to the multiplets, with a 3$\sigma$ rejection of outliers – to produce comparisons similar to the case of the MDI observations. Unfortunately the GONG fitting methodology does not produce consistent frequency tables: for a substantial number of modes the same multiplets are not fitted every time, as illustrated in Figure \[fig:gong\_incons\]. The averages used in the comparison in Figure \[fig:cmp\_w\_gong1\] correspond to the case where the multiplet is measured at least 10 times out of 58. Using a higher threshold – to obtain better temporal averages – would reduce substantially the resulting frequency set. The lower SNR of the GONG observations at low frequency does not allow to push mode fitting down to orders as low as for the MDI observations. But using a long time series still allowed me to fit low order modes that are rarely (less than 10 out of 58 times) fitted by the GONG project. The frequency comparison, presented in Figure \[fig:cmp\_w\_gong2\], shows systematic differences – whether using singlets or multiplets[^5]. The GONG project also fits a symmetric profile, hence some of the discrepency can be attributed to fitting a symmetric profile to an asymmetric peak. The residual differences, after correcting for the asymmetry using Eq. \[eq:dnuasym\], are also shown in Fig. \[fig:cmp\_w\_gong2\] for both singlets and multiplets. The corrected singlets show marginally significant differences (at the 4$\sigma$ level) with a systematic trend with frequency as for the MDI comparison (see Fig. \[fig:cmp\_w\_mdi2\]). The corrected multiplets show the same trend with frequency, but with less significance (0.8$\sigma$) simply because the uncertainties on the multiplets are larger. Comparison of mode linewidth, presented in Figure \[fig:cmp\_w\_gong3\], indicates that below 2.5 mHz the GONG values are dominated by the frequency resolution of the time series – not surprisingly, since by contrast to MDI’s estimates they are not corrected for that effect – while agreeing rather well above 2.5 mHz – except for the dip between 4 and 4.5 mHz. Mode power levels compare rather well – except at low and high frequencies – while estimates of the background level do not compare as well. ### MDI Observations - II Schou (private communication, 2004) has carried out a fitting of the 2088-day-long time series as well – using the same methodology he uses for fitting the 72-day-long time series. He has graciously provided me with the results of that fit for a direct comparison that is shown in Figures \[fig:cmp\_w\_mdi2088d\_1\] and \[fig:cmp\_w\_mdi2088d\_2\]. Coverage over the $\ell - \nu$ diagram is comparable, while the comparison of the uncertainties on the frequency suggests – as indicated earlier – that my estimates might be too conservative, by as much as a factor 3. The comparison of the mode frequencies (singlets) shows again systematic differences, that are in part explained by the effect of not including the mode asymmetry. The residual frequency differences after correcting for the asymmetry using Eq. \[eq:dnuasym\] are at the 3$\sigma$ level and present a residual systematic variation with frequency. The mode linewidths agree remarkably well (the discrepancy factor of $1.2$ is no longer present) – except at very low frequency where my uncorrected estimates are biased by the frequency resolution. The mode power and background level estimates compare rather well. Changes with Epoch ------------------ ### MDI & GONG Observations Changes in frequency, linewidth and asymmetry, with respect to the values estimated from the 2088-day-long time series, and using MDI observations over the five 728-day-long segments are illustrated in Figure \[fig:tc\_mdi\]. This figure shows that the frequency changes with epoch scale with frequency and range from 5 to 800 nHz, a now well established property that can be explained by changes with activity levels concentrated near the solar surface. The frequency changes at low degrees and very low frequency are very small and barely significant. These changes are thus small enough to justify using very long time series to fit these modes. Changes in linewidth with epoch also scale with frequency, with relative changes as large as 20% observed around 3 mHz. Changes at low frequency are hard to measure since the measured width is dominated by the spectral resolution. The mode asymmetry shows no significant sign of change with time. Changes in frequency, linewidth and asymmetry, with respect to the values estimated from the 2088-day-long time series, and using GONG observations over the five 728-day-long segments are shown in Figure \[fig:tc\_gong\]. These are very similar to the ones for the MDI data. ### Changes as a Function of Frequency and Azimuthal Order Changes in frequency, computed from multiplets rather than singlets, with respect to epoch are shown for both data set in Figure \[fig:tc\_multiplets\], as a function of frequency and azimuthal order. The individual frequency differences were binned over an equispaced grid in $\nu$ and $m/\ell$ to generate this figure. This figure shows that the change in frequencies is concentrated near the sectoral modes and that it is dominated by a pattern mostly symmetric in $m$ and nearly constant in time – given a time dependant scaling factor. This is clearly illustrated in Figure \[fig:tc\_multiplets\_2\] where the top panels show the average of the absolute value of the binned frequency differences, while the midle panels show how well the changes for each segment scale with that average. The bottom panels show the variation of the mean scaled changes and the nearly constancy of the RMS of these scaled changes. Conclusions =========== Fitting very-long time series has allowed me to push the precise characterization of low degree modes to lower frequencies. The use of such a very long time series is well justified for fitting of these modes since their variation with activity remains small and comparable to the fitting uncertainty itself. The resulting table of frequencies[^6] will allow us to further improve inferences on the structure and dynamics of the solar deep interior. The methodology I have developed includes the most up-to-date procedural elements: the use of an optimized sine multi-tapered power spectrum estimator; the simultaneous fitting of all azimuthal orders – each individually parameterized; the use of the complete leakage matrix and the inclusion of an asymmetric mode profile. The inter-comparison of the values resulting from my fit to the co-eval MDI and GONG time series is very good. The comparison with equivalent values produced by the MDI team and the GONG project is not as good. It is complicated by the inclusion in my fit of the mode profile asymmetry. The observed differences in mode frequencies are most likely dominated by the effect of including or not the mode profile asymmetry. Unfortunately a simple model for this effect does not fully account for the discrepancies. Acknowledgments {#acknowledgments .unnumbered} =============== I am very grateful to J. Schou for providing his leakage matrix coefficients and results from his mode fitting, to R. Howe and F. Hill from providing the GONG leakage matrix coefficients. The Solar Oscillations Investigation - Michelson Doppler Imager project on SOHO is supported by NASA grant NAG5–8878 and NAG5–10483 at Stanford University. SOHO is a project of international cooperation between ESA and NASA. This work utilizes data obtained by the Global Oscillation Network Group (GONG) program, managed by the National Solar Observatory, which is operated by AURA, Inc. under a cooperative agreement with the National Science Foundation. The data were acquired by instruments operated by the Big Bear Solar Observatory, High Altitude Observatory, Learmonth Solar Observatory, Udaipur Solar Observatory, Instituto de Astrofísico de Canarias, and Cerro Tololo Interamerican Observatory. SGK was supported by NASA grant NAG5–9819 & NAG5–13501 and by NSF grant ATM–0318390. Anderson, E. R., Duvall, T. L., & Jefferies, S. M. 1990, , 364, 699 Appourchaux, T., Rabello-Soares, M.-C., & Gizon, L. 1998, , 132, 121 Duvall, T. L., Jefferies, S. M., Harvey, J. W., Osaki, Y., & Pomerantz, M. A. 1993, , 410, 829 Hill, F., et al. 1996, Science, 272, 1292 Jimenez-Reyes, S. 2001, Ph.D. Thesis, La Laguna Univ. Korzennik, S. G. 1990, Ph.D. Thesis, UCLA. Korzennik, S. G., Rabello-Soares, M. C., & Schou, J. 2004, , 602, 481 Libbrecht, K. G. 1988, , 334, 510 Nelder, J.A., Mead, R. 1965, Computer Journal, 7, 308. Nigam, R. & Kosovichev, A. G. 1998, , 505, L51 Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. 1992, Cambridge: University Press, |c1992, 2nd ed., Rabello-Soares, M. C. & Appourchaux, T. 1999, , 345, 1027 Ritzwoller, M. H. & Lavely, E. M. 1991, , 369, 557 Schou, J. 1992,Ph.D. Thesis, Aarhus Univ. Schou, J. 1999, , 523, L181 Thiery, S. 2000, Ph.D. Thesis, Paris XI Univ. Toutain, T., Appourchaux, T., Fr[" o]{}hlich, C., Kosovichev, A. G., Nigam, R., & Scherrer, P. H. 1998, , 506, L147 ![Mode linewidth (long dash) and effective linewidth (solid bold curve), compared to multi-tapers spectral resolution (dotted lines). The optimal number of multi-tapers (see description in text) is indicated by the stepwise solid line, and is such that the spectral resolution remains 5 times smaller than the effective linewidth, whenever possible. Top and bottom panels correspond to 2088-day-long and 768-day-long time series, respectively. \[figure:whichmt\]](f1) ![Comparison of leakage matrix, for selected values of $\ell$ and $\beta$, resulting from a direct computation (analytical approximation) to the values computed explicitly by Schou for the MDI instrument. The RMS of the differences is around 4 to 7% below $\ell=10$ and less than 2% above. The “wider” leakage seen in the explicit computation is primarly the result of including the Gaussian weighting of the MDI images for the [*Structure Program*]{} data. \[fig:compare\_leaks\]](f2) ![Comparison of the radial component of the leakage matrix computed for the GONG observations by the GONG project (black) and by Schou (red). The RMS of the differences is less than 0.2% above $\ell=4$. \[fig:compare\_leaks\_gong\]](f3) ![Example of fitting, for MDI 2088-day-long time series, for $\ell=9$, $n=4$ and $N=5$. The top panel shows a fraction of the derotated power spectra (5th sine multi-taper). The second panel from the top shows the model of the power spectra, diamonds indicate the location of the mode frequencies. The third panel from the top shows the residuals and the bottom panel show $m$-averaged profiles (data and model). The vertical dash lines delineates the frequency range used for the fitting. Notice how for this low order case some of the modes amplitudes were not large enough to be fitted. \[fig:example1\]](f4) ![Example of fitting, as in Figure \[fig:example1\], but for $n=7$. In this case the mode and the leaks are still well resolved and all the modes are well above the background. \[fig:example2\]](f5) ![Example of fitting, as in Figure \[fig:example1\], but for $n=10$. In this case the closest spatial leaks ($\delta m = \pm 2$, $\delta\ell=0$) are barely resolved, and blends with the main peak in the $m$-averaged spectrum. \[fig:example3\]](f6) ![Example of fitting, as in Figure \[fig:example1\], but for $n=18$, to illustrate the $n'=n\pm1$, $\ell'=\ell\mp3$ contamination. The contamination – illustrated by itself in an additional panel – is in this case relatively well separated from the main peak. The $m$-averaged model without including this contamination is illustrated in the bottom panel by the dashed line. \[fig:example4\]](f7) ![Example of fitting, as in Figure \[fig:example1\], but for $n=25$, where the mode linewidth is very large and a large number of tapers has been used to estimate the power spectrum ($N=91$). \[fig:example5\]](f8) ![The coverage, in a $\ell - \nu$ diagram, of the fitting range for the 2088-day-long time series, with diamonds for the MDI observations and crosses for the GONG observations. The higher fill factor and lower signal-to-noise ratio of the MDI time series allowed to extend the fitting towards lower frequencies. \[fig:lnu\]](f9) ![Comparison of fitting results (multiplets) from the two 2088-day-long co-eval MDI and GONG time series. Top panel shows frequency differences while the bottom panel shows the frequency differences scaled to their uncertainties (blue points); in both plots the solid line results from bining the points over 20 equispaced frequency intervals, with the error bars representing the standard deviation within the bin. A histogram of the scaled difference is drawn in the bottom panel. The middle panel shows the frequency uncertainties. \[fig:compare1\]](f10) ![Comparison of fitting results (singlets) from the two 2088-day-long co-eval MDI and GONG time series (top panels). Middle panels compare the mode linewidth ($\Gamma$) while the bottom panels compare the mode asymmetry ($\alpha$). The solid lines correspond to bining individual points (shown as dots) over 10 equispaced frequency intervals, with the error bars representing the standard deviation within the bin. \[fig:compare2\]](f11) ![Comparison between singlets resulting from fitting the 2088-day-long MDI time series (diamonds) and the corresponding MDI average values computed from 27 tables resulting from fitting 72-day-long times series. Top panel shows the respective coverage in an $\ell$ – $\nu$ diagram. The lower left panel compares the frequency uncertainties, while the lower right shows the frequency differences (dots) and these differences binned over 10 equispaced frequency bins – the error bars represent the standard deviation inside each bin. \[fig:cmp\_w\_mdi1\]](f12) ![Offset resulting from fitting a symmetric profile to an asymmetric peak, as a function of the asymmetry coefficient, $\alpha$, for a fixed FWHM. The predicted offset (crosses) scales nearly lineary with the asymmetry coefficient (dash line). \[fig:sym\_vs\_asym\]](f13) ![Observed asymmetry (top left) and estimate of systematic frequency offsets due to fitting a symmetric profile to an asymmetric peak (green points in top right panel) compared to observed differences (blue points and solid line in top right panel), for MDI values. The bottom panels show residual differences after correcting for the effect of not including the mode asymmetry. The solid lines represent the values (blue dots) binned over 10 equispaced frequency bins. The standard deviation inside each bin is represented by the error bars. \[fig:cmp\_w\_mdi2\]](f14) ![(Left) Comparison of observed differences between fitting the 2088-day-long MDI time series and the corresponding MDI average values computed from 27 tables resulting from fitting 72-day-long times series (raw values as blue dots and binned values shown as the red line) with the predicted differences resulting from a simple model (green points and line) of fitting an isolated symmetric profile to an asymmetric one. (Right) Frequency differences resulting from symmetric and asymmteric fits carried out by Schou using a 72-day-long MDI time series (individual values shown as dots, solid line represents binned values) compared to the frequency differences shown in the left panel (red line, binned values) and the simple model prediction (green line, binned values).\[fig:cmp\_sym\_vs\_asym\]](f15) ![Comparison between results from fitting the 2088-day-long MDI time series (red dots & diamonds) and the corresponding MDI average values computed from 27 tables resulting from fitting 72-day-long times series (black dots & crosses). The top panel compares linewidths, the horizontal lines correspond to the respective time series frequency resolution, while the blue curve illustrates the factor 1.2 between the two sets. The middle panel compares the mode power ($A\times\Gamma$). The bottom panel compares the background power level. \[fig:cmp\_w\_mdi3\]](f16) ![Comparison between results from fitting the 2088-day-long GONG time series (diamonds or black dots) and the corresponding GONG average values computed from 58 tables resulting from fitting 36-day-long times series (crosses or red dots). Top panel shows the respective coverage in an $\ell$ – $\nu$ diagram, when computing averages only if the multiplet is present in at least 10 of the 58 tables. The lower left panel compares the frequency uncertainties for the multiplets, while the lower right panel shows the frequency differences (dots, multiplets) and the differences (singlets and multiplets) binned over equispaced frequency bins – the error bars represent the standard deviation inside each bin. The binned singlets differences is shown as the green curve. \[fig:cmp\_w\_gong1\]](f17) ![This figure illustrates a consistency problem present in the GONG frequency tables. The top panel shows in an $\ell$–$\nu$ diagram averaged multiplets (dots) and the resulting average singlets when only considering multiplets that are present in at least 10 of the 58 tables. The lower left panel shows the number of times each multiplet is present – as a function of its frequency. The lower right panel shows the cumulative histogram of the number of times a multiplet is present. Dashed lines are draw at 10 (the value I ended up using) and at 2/3 and 3/4 of 58. When using 10 as a threshold only some 15% of the mode set has to be dropped, while if I would use a more conservative threshold nearly half of the mode set would be ignored. Note also how using only such a threshold can skew singlet estimates at the edge of the covered $\ell$–$\nu$ diagram. \[fig:gong\_incons\]](f18) ![Observed asymmetry (top left) and estimate of systematic frequency offsets due to fitting a symmetric profile to an asymmetric peak (in green, top right) compared to observed differences (blue dots and solid line) for GONG singlets. The bottom left panel compares directly the multiplets (blue dots and solid line) to the estimate of systematic offsets (green curve). The bottom right panels show the reduced frequency differences, after correcting for the effect of not including the mode asymmetry, for singlets and multiplets. \[fig:cmp\_w\_gong2\]](f19) ![Comparison between results from fitting the 2088-day-long GONG time series (red dots & diamonds) and the corresponding GONG average values computed from 58 tables resulting from fitting 36-day-long times series (black dots & crosses). The top panel compares linewidths, the horizontal lines correspond to the respective time series frequency resolution. The middle panel compares the mode power ($A\times\Gamma$). The bottom panel compares the background power level. \[fig:cmp\_w\_gong3\]](f20) ![Comparison of MDI singlets resulting from this work (crosses) and from Schou’s fitting to the same 2088-day-long time series (diamonds). Coverage in the $\ell$ – $\nu$ diagram is very similar. My estimate of frequency uncertainties appears too conservative while the frequency differences show a systematic pattern with frequency. \[fig:cmp\_w\_mdi2088d\_1\]](f21) ![Comparison of MDI singlets resulting from this work (crosses) and from Schou’s fitting to the same 2088-day-long time series (diamonds). The frequency differences (blue dots and black curves for binned values) are not fully explained by the effect of not including an asymmetry in the mode fitting (green points, top left panel). Residual differences, after correcting for the effect of not including the mode asymmetry, remain at the 3$\sigma$ level. Mode linewidth and power as well as estimates of background level compares remarkably well. \[fig:cmp\_w\_mdi2088d\_2\]](f22) ![Changes in singlet parameters, with respect to the values estimated from the 2088-day-long time series, using MDI observations over the five 728-day-long segments. Panels in the top row show change in frequency, panels in the middle row show changes in linewidth, while the panels in the bottom row compare estimates of asymmetry. Binned values, over equispaced interval in frequency are shown, with the standard deviation within the bin indicated by the error bars. \[fig:tc\_mdi\]](f23) ![Changes in singlet parameters, with respect to the values estimated from the 2088-day-long time series, using GONG observations over the five 728-day-long segments. Panels in the top row show change in frequency, panels in the middle row show changes in linewidth, while the panels in the bottom row compare estimates of asymmetry. Binned values, over equispaced interval in frequency are shown, with the standard deviation within the bin indicated by the error bars. \[fig:tc\_gong\]](f24) ![Changes in frequency for each 728-day-long segment with respect to the values estimated from the 2088-day-long time series based on multiplets and plotted as a function of frequency and the ratio $m/\ell$, for MDI (left column) and GONG (right column). The frequency differences were binned over a grid equispaced in $\nu$ and $m/\ell$. \[fig:tc\_multiplets\]](f25) ![Top panels: average of the absolute value of the changes in frequency for each 728-day-long segment with respect to the values estimated from the 2088-day-long time series, as shown in Fig. \[fig:tc\_multiplets\]. Middle 5 panels: frequency changes divided by the average of the absolute values, for each 728-day-long segment. Bottom panels: average and RMS of the changes scaled by the average of the absolute values as a function of segment number. The RMS remains nearly constant while the average varies with time, indicating that the changes in frequencies can be described by a fixed pattern nearly constant in time scaled by a time dependant factor. \[fig:tc\_multiplets\_2\]](f26) [^1]: This gap results from the loss of contact with the SOHO spacecraft. [^2]: The missing 2 tables (since $2088=29\times72$) correspond to the time interval when contact with the SOHO spacecraft was lost. [^3]: This parameterization is equivalent to the one defined by Equation 4 of @Nigam+Kosovichev:1998. After some rudimentary algebra, one can identify their asymmetry coefficient, $B$, to the one I use, $\alpha$, namely $\frac{\alpha}{2} = B\,(1-\frac{\alpha^2}{2})$. [^4]: It is the 72-day-long time series starting at mission day no.2368, namely on June 27th, 1999 at 0 UT. [^5]: The GONG project produces tables of multiplets. [^6]: Available at [ftp://cfa-ftp.harvard.edu/pub/sylvain/tables/]{}.

ArXiv

--- abstract: 'We investigate granular particles in a shaken quasi two dimensional box in molecular dynamics computer simulations. After a sudden change (quench) of the shaking amplitude, transient density correlations are observed orders of magnitude beyond the steady state correlation length scale. Propagation of the correlations is ballistic, in contrast to recently investigated quenches of Brownian particles that show diffusive propagation [@rohwer2017transient; @rohwer2018nonequilibrium]. At sufficiently strong cooling of the fluid the effect is overlaid by clustering instability of the homogeneous cooling state with different scaling behavior. We are able to identify different quench regimes. In each regime correlations exhibit remarkably universal position dependence. In simulations performed with side walls we find confinement effects for temperature and pressure in steady state simulations, and an additional transient wall pressure contribution upon changing the shaking amplitude. The transient contribution is ascribed to enhanced relaxation of the fluid in the presence of walls. From incompatible scaling behavior we conclude that the observed effects with and without side walls constitute distinct phenomena.' author: - Thomas Schindler - 'Christian M. Rohwer' bibliography: - 'literature-paper.bib' title: Ballistic propagation of density correlations and excess wall forces in quenched granular media --- Introduction {#sec:intro} ============ A dynamic system of macroscopic particles tends to dissipate kinetic energy due to inelastic collisions. In order to maintain particle motion, energy input by an external source is needed. One of the setups commonly employed to this end is the quasi granular shaker, which consists of a flat box filled with typically millimeter sized beads (usually made of metal or glass), that is vibrated vertically. The directed energy input is randomized in particle-particle collisions yielding dynamical steady states reminiscent of thermal equilibrium. In particular, the parameter space for formation of regular lattices, fluids and coexistence thereof [@prevost2004nonequilibrium; @melby2005dynamics; @reis2006crystallization; @clerc2008liquid; @vega2008effect; @rivas2011segregation; @guzman2018critical; @schindler2019nonequilibrium] bears analogy to the corresponding equilibrium system [@schmidt1997phase]. However several properties reveal the nonequilibrium nature of the steady states, such as inelastic collapse at the bottom of the container [@olafsen1998clustering; @nie2000dynamics; @olafsen2005two; @khain2011hydrodynamics], inhomogeneous granular temperatures [@prevost2004nonequilibrium; @lobkovsky2009effects], non-Gaussian velocity distributions [@losert1999velocity; @olafsen1999velocity; @kawarada2004non], segregation of mixtures [@rivas2011sudden; @rivas2011segregation; @rivas2012characterization], and inelastic hydrodynamic modes [@brito2013hydrodynamic]. In the present study we disturb the steady state by changing the driving strength, in order to search for further evidence of its nonequilibrium origin. This technique has proven fruitful as several anomalies in response functions have been reported, [e.g., ]{}in the Kovacs memory effect [@prados2014kovacs; @trizac2014memory; @brey2014memory] or in the compaction behavior [@nicodemi1999dynamical; @caglioti1997tetris; @brey2001linear; @brey2002memory]. Part of the article is devoted to forces between distant walls, mediated by the granular medium. When the properties of the medium are altered by the confinement, this can result in nontrivial macroscopic forces [@rodriguez2016clustering; @aumaitre2001segregation; @villanueva2010casimir; @denisov2011simulation; @zuriguel2005role]. In analogy to the corresponding quantum effect [@bordag2009advances], these are commonly termed as classical Casimir forces [@kardar1999friction]. The paper is organized as follows. In Sec. \[sec:model\] we introduce the examined setup and give technical details about simulation parameters. Our results for density correlations in bulk simulations are presented in Sec. \[sec:quench-bulk\]. We observe large scale correlations after a change in driving amplitude, for which we carve out similarities and differences to a recent theory. Section \[sec:casimir\] treats temperature and pressure in a geometry with additional side walls, and discusses their finite size scaling in steady state and after quenches. In Sec. \[sec:con\] we conclude by classifying the observed phenomena. Setup and model {#sec:model} =============== ![\[fig:set\] Sketches of the setup [(a)]{} in bulk geometry and [(b)]{} with side walls (red) confining the simulation box in the $x$-direction. The bold arrows indicate the directions in which periodic boundary conditions are applied. Dashed arrows mark the shaking in $z$-direction. Symbols are declared in the main text. ](figure1.png){width="\halfnl"} We consider a system of $N$ hard spheres of diameter $\sigma$ and mass $m_\mathrm{p}$ in a shallow cuboidal box with dimensions $L_x\times L_y\times L_z$ ($L_z=2\sigma$) with hard bottom and top plates (see [Fig. \[fig:set\]]{}). Gravitational acceleration $g$ acts in the negative $z$-direction, which induces a time scale $\tau_0\equiv\sqrt{\sigma/g}$ and an energy scale $\epsilon\equiv m_\mathrm{p}g\sigma$. The plates are oscillating with a time dependent displacement $A\sin(\omega t + \varphi)$ in the $z$-direction with an amplitude $A$, an angular frequency $\omega=50\tau_0^{-1}$, time $t$, and a phase shift $\varphi$. The sole purpose of introducing $\varphi$ here is to clarify that quenches (explained below) are not in sync with the plate oscillation; this is achieved by averaging over $\varphi$. The area density of particles is fixed to $\rho\equiv N/L_x L_y=0.5\sigma^{-2}$ throughout the paper. The main control parameter in this article is $A$, and three different amplitudes, $A_1\equiv 0.002\sigma$, $A_2\equiv 0.005\sigma$, and $A_3\equiv 0.05\sigma$ are considered. We conduct steady state simulations at constant $A$, as well as quenched simulations. Quenches are performed by suddenly increasing or decreasing the shaking amplitude at $t=0$ from an initial value $A{_\mathrm{I}}$ to a final value $A{_\mathrm{F}}$, and observing the granular fluid after the quench. Henceforth we denote quench protocols between $A_1$ and $A_2$ as “moderate quenches” and protocols starting from or ending at $A_3$ (with $A_3$ being an order of magnitude larger than $A_1$ and $A_2$) as “strong quenches”. As the setup is very shallow, we treat it as an effective -system and calculate observables only from the $x$- and $y$-components of particle positions and velocities. Two different geometries are investigated. On the one hand we employ the described setup with periodic boundary conditions in the $x$- and $y$-directions [\[see [Fig. \[fig:set\][(a)]{}]{}\]]{}, henceforth referred to as (-)bulk. The lateral box dimensions here are square-shaped, with $L\equiv L_x=L_y$ ranging from $100\sigma$ to $400\sigma$. The other setup considered has periodic boundary conditions in $y$-direction (with $L_y=400\sigma$) but is confined between two vertical immovable side walls in the $x$-direction, separated by a distance $L_x$ ranging from $5\sigma$ to $200\sigma$ [\[[Fig. \[fig:set\][(b)]{}]{}\]]{}. The well known Casimir setup would also include an infinitely extended exterior domain beyond the side walls. The main goal here is the measurement of differences in temperature and pressure between the interior and exterior. However, we can not simulate the latter explicitly. Instead we extrapolate our finite simulation results to $L_x\rightarrow\infty$. All temperature differences and net pressures on the side walls are calculated as differences between the actually simulated interior and the extrapolated exterior. The setup is studied in event driven molecular dynamics computer simulations with the DynamO [@bannerman2011dynamo] package. The central idea of the algorithm is to predict collisions of particles from their current positions and velocities. These are entered into a schedule. The system is then evolved by forwarding to the next collision in the schedule and calculating the new velocities and the next collisions of the collision partners. The algorithm is suitable for a system with short interaction times and parabolic trajectories in between, such as the hard macroscopic marbles studied here. The interactions are modeled as instantaneous billiard-like collisions with momentum conservation [@rapaport2004art]. Energy loss is accounted for by rescaling the relative particle velocities after collisions (either with other particles or with walls) by a coefficient of restitution of 0.95. Coulomb friction ([i.e., ]{}friction due to relative tangential motion at contact) is neglected, and hence there is no transfer of angular momentum. Therefore, we do not need to simulate rotations of the spheres. While this model is simplistic, it captures the essential mechanisms of energy input and dissipation, and thus creates the nonequilibrium steady states that are also found in experiments or more sophisticated simulations. In our previous studies [@schindler2019nonequilibrium] we found the phase behavior to be consistent with simulations employing rotating spheres [@prevost2004nonequilibrium; @melby2005dynamics; @reis2006crystallization; @clerc2008liquid; @vega2008effect; @rivas2011segregation; @guzman2018critical]. Quenched bulk {#sec:quench-bulk} ============= ![\[fig:PT-bqA\] [(a)]{} Relaxation of the temperature $T$ towards the final steady state value $T{_\mathrm{F}}$ in the quenched fluid, as a function of time $t$ after the quench on logarithmic scale for several initial and final amplitudes $A_i$ ($i=1,2,3$) as indicated. The insets show the values of $T$ itself. The $T{_\mathrm{F}}$ at each $A_i$ are obtained in separate steady state simulations. [(b)]{} shows the same analysis as [(a)]{} but for the internal pressure ${\ensuremath{P_\mathrm{int}}\xspace}$. ](figure2a.pdf "fig:"){width="\halfnl"} ![\[fig:PT-bqA\] [(a)]{} Relaxation of the temperature $T$ towards the final steady state value $T{_\mathrm{F}}$ in the quenched fluid, as a function of time $t$ after the quench on logarithmic scale for several initial and final amplitudes $A_i$ ($i=1,2,3$) as indicated. The insets show the values of $T$ itself. The $T{_\mathrm{F}}$ at each $A_i$ are obtained in separate steady state simulations. [(b)]{} shows the same analysis as [(a)]{} but for the internal pressure ${\ensuremath{P_\mathrm{int}}\xspace}$. ](figure2b.pdf "fig:"){width="\halfnl"} In this section we characterize the bulk system [\[see [Fig. \[fig:set\]]{}[a]{}\]]{} after a quench and show how large scale transient correlations emerge. At first, however, we look at the granular temperature and internal mechanical pressure, defined as $$\label{eq:Temp} T\equiv\left\langle \frac{1}{Nd}\sum_{i=1}^{N}\frac{m_\mathrm{p}}{2}{\mathbf{v}}_i^2\right\rangle$$ and $$\label{eq:Pint} {\ensuremath{P_\mathrm{int}}\xspace}\equiv\left\langle\frac{1}{d L_x L_y\Delta t}\sum_{\mathrm{PP}}\Delta{\mathbf{p}}_i\cdot({\mathbf{r}}_i-{\mathbf{r}}_j)\right\rangle\,,$$ respectively. Here ${\mathbf{v}}_i$ is the velocity of particle $i$, $d=2$ is the spatial dimension, $\mathrm{PP}$ indicates summation over all particle-particle collisions between particles $i$ and $j$ during a time interval $\Delta t$, $\Delta{\mathbf{p}}_i$ is the change of momentum of particle $i$ during the collision, and ${\mathbf{r}}_i$ is its position. Angular brackets denote averages over a large time interval in steady state simulations, or over a small time interval $[t-\Delta t/2,t+\Delta t/2]$ and multiple quench realizations in quenched simulations. We stress once more that all vectors in the above equations are projections onto the $xy$-plane. Figure \[fig:PT-bqA\] shows the two quantities defined in [Eqs. ]{} and as functions of time $t$ after the quench, for several $A{_\mathrm{I}}$ and $A{_\mathrm{F}}$. We observe two qualitatively different types of behaviors. In the case of heating or moderate cooling (21), $T$ and ${\ensuremath{P_\mathrm{int}}\xspace}$ relax exponentially in time towards the final steady state values $T{_\mathrm{F}}$ and $P_\mathrm{int,F}$, respectively. The relaxation time is inversely proportional to $A{_\mathrm{F}}$, since $A{_\mathrm{F}}$ is proportional to the root mean square particle velocity. This proportionality is not necessarily true in general but it does apply in the strong shaking regime ($A\omega^2\gg g$) employed in this work. Here $\omega$ provides the predominant time scale (see, [e.g., ]{}Refs. [@rivas2011segregation; @melby2005dynamics]) and therefore the particle velocities scale with the peak velocity of the plates, $A\omega$. Only in the case of strong cooling, ([i.e., ]{}with initial amplitude $A_3$,) do $T$ and ${\ensuremath{P_\mathrm{int}}\xspace}$ relax algebraically in time. This behavior is known for a freely cooling granular gas [@haff1983grain] which is comparable to the present situation as long as $T(t)\gg T{_\mathrm{F}}$. The asymmetry between strong heating and strong cooling originates from the asymmetry of energy gain and dissipation of the horizontal degrees of freedom via particle-particle collisions. Cooling through inelastic dissipation takes place in every collision. Heating of the horizontal directions, on the other hand, only occurs, if a particle has been accelerated by the oscillating plates and then transfers its energy to horizontal motion in a particle-particle collision as has been demonstrated for steady state fluctuations in a setup with particles of differing masses [@rivas2011sudden; @rivas2012characterization]. Moderate quenches constitute only weak disturbances of the steady state where the described asymmetry plays only a minor role. As we shall see, the two described cases are distinct by other observables as well. Hence, we will refer to heating or moderate cooling as type I and to strong cooling as type II behavior in the remainder of the paper. {width="\textwidth"} The different types can even be distinguished when comparing simulation snapshots by eye. [Figure \[fig:snap-bq\]]{} shows two different time series, where the particle color encodes the local density. In the type I simulation (top row) the system remains homogeneous and only the distribution of local densities of the particles becomes more heterogeneous with increasing $T$, which can be seen by the number of small dense patches increasing. This is a consequence of faster particles exploring the upper half of the box and hence displaying larger overlaps in projection ([cf. ]{}Appendix \[sec:steady-bulk\]). The bottom row shows the reverse process of strong cooling and has switched initial and final states. At intermediate times, however, we observe the formation of dense and dilute domains on the scale of the box size. (See [e.g., ]{}the diluted region in the upper right part in the snapshots of $t=30\tau_0$ and $100\tau_0$.) This constitutes a clustering instability (see, [e.g., ]{}Refs. [@goldhirsch1993clustering; @poschel2005transient]), which is ultimately dissolved by the weak shaking at $A{_\mathrm{F}}$. The clustering instability is not necessarily but commonly observed in free cooling states. It therefore serves as an indicator of free cooling, and nicely illustrates the asymmetry between cooling and heating discussed before. {width="\thirdnl"} {width="\thirdnl"} {width="\thirdnl"}\ {width="\thirdnl"} {width="\thirdnl"} {width="\thirdnl"} Now, we turn to two point correlations measured by the transient total correlation function [@HMcD] $$h(r,t) \equiv\left\langle \frac{1}{\rho^2}\sum_{i=1}^{N}\sum_{j\neq i}^{N}\delta({\mathbf{r}}_i - {\mathbf{r}})\delta({\mathbf{r}}_j - {\mathbf{r}}')\right\rangle-1\,,$$ where $\delta$ is the Dirac-delta-distribution, ${\mathbf{r}}$ and ${\mathbf{r}}'$ are two position vectors, and $r\equiv |{\mathbf{r}}-{\mathbf{r}}'|$. [Figure \[fig:corr-bqA\]]{} shows time series of $h$ as a function of $r$ for all considered quench protocols. In the range $r<5\sigma$ there are exponentially decaying oscillations (off scale), which constitute the fluid structure also present in the steady state ([cf. ]{}Fig. \[fig:bs\] in Appendix \[sec:steady-bulk\]). The focus of this work is not this well known feature of any dense liquid [@HMcD], but rather the transient contributions that are observed at larger distances $r>5\sigma$. In the initial state $t=0$ (black curves) there are no large scale correlations. After the quench, however, these build up in time and decay to zero again as $t\rightarrow\infty$. There are two rather distinct types of behaviors for type I and II protocols as classified previously in this section. In type I simulations [\[[Fig. \[fig:corr-bqA\][(a)]{}]{}–[(d)]{}\]]{} we find oscillating correlations with local maxima and/or minima (the monotony is discussed later in this section). The extrema are propagating in time towards $r\rightarrow\infty$ and are of the order of at most $\sim 10^{-3}$ in magnitude. The shape of each function does not exhibit finite size scaling, [i.e., ]{}it is stable against utilizing different box sizes (aside from boundary effects at $r \approx L/2$). Type II protocols [\[[Fig. \[fig:corr-bqA\][(e)]{}]{} and [(f)]{}\]]{} exhibit larger correlations up to $10^{-2}$ in magnitude that are not propagating and are positive in sign. (The fact that the correlations do not vanish for $r\rightarrow L/2$, but approach a small negative value, indicates density inhomogeneities at the scale of the box size, and clearly constitutes a finite size effect.) This confirms the conclusions from the inspection of the snapshots and characterization of type II behavior as inelastic collapse. ![\[fig:XS-CS-t\] [(a)]{} Distance $r{_\mathrm{E}}$ of the rightmost local extremum of the total correlation function $h$ as a function of time $t$ (multiplied by the post-quench peak velocity of the oscillating plates $A{_\mathrm{F}}\omega$ for comparability), extracted from type I protocols [\[[Figs. \[fig:corr-bqA\][(a)]{}]{}-[(d)]{}\]]{}. Straight lines are linear fits obtained from the data ranges, where the lines are continuous. [(b)]{} Absolute values $|h{_\mathrm{E}}|$ of the total correlation function at the extrema as a function of $r{_\mathrm{E}}$ on double logarithmic scales. Straight lines are the theoretical predictions for diffusive systems with instantaneous temperature quenches according to . The black dashed lines corresponding to power laws with exponents $-0.7$ and $-2$ are guides to the eye. Data points of both panels at $r<50\sigma$ have been obtained in simulation boxes with $L=100\sigma$, data points at $r>50\sigma$ have been obtained in simulation boxes with $L=400\sigma$. ](figure5a.pdf "fig:"){width="\halfnl"}\ ![\[fig:XS-CS-t\] [(a)]{} Distance $r{_\mathrm{E}}$ of the rightmost local extremum of the total correlation function $h$ as a function of time $t$ (multiplied by the post-quench peak velocity of the oscillating plates $A{_\mathrm{F}}\omega$ for comparability), extracted from type I protocols [\[[Figs. \[fig:corr-bqA\][(a)]{}]{}-[(d)]{}\]]{}. Straight lines are linear fits obtained from the data ranges, where the lines are continuous. [(b)]{} Absolute values $|h{_\mathrm{E}}|$ of the total correlation function at the extrema as a function of $r{_\mathrm{E}}$ on double logarithmic scales. Straight lines are the theoretical predictions for diffusive systems with instantaneous temperature quenches according to . The black dashed lines corresponding to power laws with exponents $-0.7$ and $-2$ are guides to the eye. Data points of both panels at $r<50\sigma$ have been obtained in simulation boxes with $L=100\sigma$, data points at $r>50\sigma$ have been obtained in simulation boxes with $L=400\sigma$. ](figure5b.pdf "fig:"){width="\halfnl"} The remainder of the section is devoted to quantitative evaluation of the type I behavior via the positions of the rightmost local extrema [\[marked with black crosses in [Figs. \[fig:corr-bqA\][(a)]{}]{}-[(d)]{}\]]{}. [Figure \[fig:XS-CS-t\][(a)]{}]{} shows the distances $r{_\mathrm{E}}$ of the extrema as functions of $t$. We identify two distinct regimes, namely a short time regime during which $T$ still adjusts to $A{_\mathrm{F}}$ ([cf. ]{}[Fig. \[fig:PT-bqA\]]{}), and a long time regime at constant $T$. In the short time regime, the propagation velocity $v{_\mathrm{E}}$ of the extremum increases or decreases as $T$ increases or decreases in heating or cooling protocols ([cf. ]{}[Fig. \[fig:PT-bqA\]]{}). As $T$ relaxes to the final steady state value, $v{_\mathrm{E}}$ takes a constant value approximately proportional to $A{_\mathrm{F}}$, which lies in the range $$v{_\mathrm{E}}=(11.4\pm 0.5)A{_\mathrm{F}}\omega$$ for the different protocols. Hence we obtain the dynamic scaling exponent $\alpha=1$ (defined via $r{_\mathrm{E}}\propto t^\alpha$) of ballistic motion. We stress that $v{_\mathrm{E}}$ is not the speed of sound $c_\mathrm{s}$ of the final steady state fluid, which we calculated in supplemental steady state simulations via the dynamic structure factor according to the method described in Ref. [@HMcD] as $c_\mathrm{s}(A_1)=(7.8\pm0.4)A\omega$ and $c_\mathrm{s}(A_3)=(4.2\pm0.2)A\omega$. The fact that $c_\mathrm{s}\not\propto A\omega$ clearly disqualifies the speed of sound as possible interpretation for $v{_\mathrm{E}}$. Instead, a physical interpretation of $v{_\mathrm{E}}$ could be provided by the following possible origin of the correlations, which–in the case of heating–is similar to bursts caused by collisions of heavy particles reported in Ref. [@rivas2012characterization]. Upon heating (weak cooling), particles are accelerated (decelerated) in $z$-direction by the oscillating plates to a velocity $\propto A{_\mathrm{F}}\omega$. Accelerated (decelerated) particles may transfer their kinetic energy to (recover vertical kinetic energy from) horizontal directions in particle-particle collisions, which creates pairs of excess (depleted) momenta in opposite directions of the involved particles. These pairs of momenta induce particle currents, which create correlations that are transported at constant speed $\propto A{_\mathrm{F}}\omega$. In this manner, momentum is transferred without loss to other particles, as momentum is conserved in collisions, which explains the constant speed in the long time regime. In the intermediate regime the velocity may be reduced (enhanced), as particles are not accelerated (decelerated) completely by the plates before they undergo particle-particle collisions. This mechanism requires that particles are thermalized at a variety of rates, which is not given in the case of strong cooling, where $T{_\mathrm{F}}\omega\ll T{_\mathrm{I}}$, which means that plates are practically immovable and energy loss is dominated by particle inelasticity. The value $h{_\mathrm{E}}$ of the correlation function at the extremum [\[see [Fig. \[fig:XS-CS-t\][(b)]{}]{}\]]{} also exhibits a crossover like $r{_\mathrm{E}}$ with the same crossover times. In each of the regimes the dynamics is describable by an algebraic scaling. In the short time regime $h{_\mathrm{E}}\propto r{_\mathrm{E}}^{-0.7}$ and in the long time regime $h{_\mathrm{E}}\propto r{_\mathrm{E}}^{-2}$ for all protocols, which yields the scaling exponent of the correlation strength $\beta=-2$. We compare our results to a recently developed theory [@rohwer2017transient; @rohwer2018nonequilibrium] for fluctuation induced correlations after instantaneous temperature changes in diffusive systems. This theory predicts $$\label{eq:C-X-t-anal} h(r,t)=\frac{S{_\mathrm{I}}-S{_\mathrm{F}}}{\rho}\frac{e^{-\frac{r^2}{2r{_\mathrm{D}}(t)^2}}}{\sqrt{2\pi r{_\mathrm{D}}(t)^2}^{\,d}}\,,$$ where $S{_\mathrm{I}}$ and $S{_\mathrm{F}}$ are the zero wavelength limits of the static structure factors of the initial and final steady state fluid, respectively, and $r{_\mathrm{D}}$ is the correlation length. For quantitative comparison, we extract the inflection point of this function, giving us the typical strength and length scale of correlations. The typical strength is obtained by setting $r=r{_\mathrm{D}}$, $$\label{eq:CS-XS}h{_\mathrm{D}}(r{_\mathrm{D}})=\frac{S{_\mathrm{I}}-S{_\mathrm{F}}}{\rho\sqrt{e}2\pi r{_\mathrm{D}}^2}\propto r{_\mathrm{D}}^{-2}\,.$$ yielding a scaling exponent $\beta=-2$. Note that this is independent of the dynamical scaling of $r{_\mathrm{D}}$ and only depends on the spatial dimension. Therefore, $\beta$ can be seen as a geometric property that ensures constant normalization of the Gaussian distribution. The correlation length as a function of time reads $$\label{eq:XS-t}r{_\mathrm{D}}(t)=\sqrt{4D{_\mathrm{F}}t}\propto t^{1/2}\,,$$ with the long-time single-particle diffusion coefficient $D{_\mathrm{F}}$ of the final steady state. This implies $\alpha=1/2$ – the dynamic scaling exponent of diffusive motion. The prefactor in [Eq. ]{} is proportional to the difference of the initial and final steady state static structure factors, which are connected to the respective compressibilities $\chi$ via $S=\rho T \chi\equiv T \partial \rho/\partial P$. Hence, we only expect a nonzero effect for thermal particles with variable softness. In the present setup we employ hard spheres, which are athermal by themselves. As scrutinized in Appendix \[sec:steady-bulk\], however, we create an effective softness via variation of the stratification of the particles at different $A$, which influences their overlaps in the projection. This makes it possible to observe the predicted effect in our system despite the hard core model. Note that we use the term “softness” for structural properties of the fluid rather than material properties of the individual particles. Overlaps due to actual compression of glass or metal spheres in experiments or different simulation models would be orders of magnitude smaller than the overlaps due to stratification. A first notable observation upon comparing our results to the diffusive case, is the rather different functional form of $h$ in simulation and theory. While the theory predicts a universal Gaussian shape of the correlations, we observed a more complicated function shape with oscillating behavior that depends on the applied protocol (see also below). Therefore, we only compare the scaling exponents of the extracted extrema shown in [Fig. \[fig:XS-CS-t\]]{} to [Eqs. ]{} and , and not the prefactors. Even though the theory is strictly speaking not applicable here, the value $\alpha=1$ obtained in simulation is remarkable. $T$ is fully relaxed in the long time regime and thus one could expect the post-quench fluid to behave like a steady state fluid. Indeed, we performed preparatory steady state simulations where we find diffusive motion of the individual particles for $r>50\sigma$. However, the correlation function exhibits the ballistic scaling of propagating waves. The picture is quite different when considering $\beta$. Here the long time limit $\beta=-2$ fits the diffusive theory well. This universality supports the notion that $\beta$ is a purely geometric quantity that is independent of the dynamic details and designates the correlations as a conserved quantity in steady state. The scaling $|h{_\mathrm{E}}|\propto r{_\mathrm{E}}^{-0.7}$ during thermalization, on the other hand, indicates a normalization which increases in time or in other words a source of correlations. This observation supports the picture of a build-up of correlations during temperature equilibration as proposed before. ![\[fig:C-X-rescaled\] The same data as [Figs. \[fig:corr-bqA\][(a)]{}]{}-[(d)]{}, but with abscissa and ordinate rescaled by $r{_\mathrm{E}}(t)$ and $|h{_\mathrm{E}}(t)|$, respectively [\[such that the extrema collapse at (1,1) or (1,-1)\]]{}. Protocols are indicated in the labels. Only data exhibiting an extremum in $h$ is shown. Panels [(a)]{} and [(b)]{} show moderate heating and cooling, respectively. Panels [(c)]{} and [(d)]{} show strong heating. Time is encoded in the color of each curve as depicted in the colored bars of the panels. ](figure6a.pdf "fig:"){width="\quarternl"} ![\[fig:C-X-rescaled\] The same data as [Figs. \[fig:corr-bqA\][(a)]{}]{}-[(d)]{}, but with abscissa and ordinate rescaled by $r{_\mathrm{E}}(t)$ and $|h{_\mathrm{E}}(t)|$, respectively [\[such that the extrema collapse at (1,1) or (1,-1)\]]{}. Protocols are indicated in the labels. Only data exhibiting an extremum in $h$ is shown. Panels [(a)]{} and [(b)]{} show moderate heating and cooling, respectively. Panels [(c)]{} and [(d)]{} show strong heating. Time is encoded in the color of each curve as depicted in the colored bars of the panels. ](figure6b.pdf "fig:"){width="\quarternl"}\ ![\[fig:C-X-rescaled\] The same data as [Figs. \[fig:corr-bqA\][(a)]{}]{}-[(d)]{}, but with abscissa and ordinate rescaled by $r{_\mathrm{E}}(t)$ and $|h{_\mathrm{E}}(t)|$, respectively [\[such that the extrema collapse at (1,1) or (1,-1)\]]{}. Protocols are indicated in the labels. Only data exhibiting an extremum in $h$ is shown. Panels [(a)]{} and [(b)]{} show moderate heating and cooling, respectively. Panels [(c)]{} and [(d)]{} show strong heating. Time is encoded in the color of each curve as depicted in the colored bars of the panels. ](figure6c.pdf "fig:"){width="\quarternl"} ![\[fig:C-X-rescaled\] The same data as [Figs. \[fig:corr-bqA\][(a)]{}]{}-[(d)]{}, but with abscissa and ordinate rescaled by $r{_\mathrm{E}}(t)$ and $|h{_\mathrm{E}}(t)|$, respectively [\[such that the extrema collapse at (1,1) or (1,-1)\]]{}. Protocols are indicated in the labels. Only data exhibiting an extremum in $h$ is shown. Panels [(a)]{} and [(b)]{} show moderate heating and cooling, respectively. Panels [(c)]{} and [(d)]{} show strong heating. Time is encoded in the color of each curve as depicted in the colored bars of the panels. ](figure6d.pdf "fig:"){width="\quarternl"} In order to characterize its features, we collapse $h(r)$ for different $t$ by rescaling abscissa and ordinate with the values of the extrema, [i.e., ]{}we plot $h / |h{_\mathrm{E}}|$ versus $r / r{_\mathrm{E}}$ as shown in [Figs. \[fig:C-X-rescaled\][(a)]{}]{} - [(d)]{} for the different quench protocols. By definition, this scaling function has a local maximum at $(1,1)$ for heating or minimum at $(1,-1)$ for cooling protocols. What is striking here is the stability of the function shapes, which is in sharp contrast to the strong crossover of the scaling exponents. There are only slight shifts of the functions left and right of the main extrema at the times of the crossover. At moderate heating [\[panel [(a)]{}\]]{} only a single maximum is present. For $r\rightarrow 0$, $h$ attains negative values and for $r\rightarrow\infty$, $h$ decays towards zero. The strong heating protocols [(b)]{} and [(c)]{} both show the same qualitative behavior, which differs from moderate heating by an additional local minimum at $\approx(0.4,-4)$ before or $\approx(0.5,-3)$ after the crossover. Consequently there is an additional zero at $r/r{_\mathrm{E}}\approx0.2$ and $h>0$ for $r\rightarrow0$. This new feature could be attributed to an additional process that only takes place when the fluid temperature is changed strongly. The feature is located at smaller $r$ than the original extremum, indicating that the additional process takes place at a later time than the process that creates the maximum at $(1,1)$. The data, however, does not reveal the nature of this process. Moderate cooling [\[panel [(d)]{}\]]{} exhibits exactly the same behavior as moderate heating but with the opposite sign. This is a clear signature for a linear response regime. Casimir geometry {#sec:casimir} ================ {width="\thirdnl"} {width="\thirdnl"} {width="\thirdnl"}\ This section treats the setup with side walls as depicted in [Fig. \[fig:set\][(b)]{}]{}. The aim here is to report boundary effects on global observables and to determine whether these are caused by the bulk post-quench correlations described in the previous section as is the case in diffusive systems [@rohwer2017transient]. We start by describing confinement effects in the steady state by means of $T$, [$P_\mathrm{int}$]{}and the pressure on the side walls, $${\ensuremath{P_\mathrm{wall}}\xspace}\equiv\left\langle\frac{1}{2L_y\Delta t}\sum_{\mathrm{PW}}\Delta{\mathbf{p}}_i\cdot{\mathbf{n}}_{\mathrm{w}}\right\rangle\,,$$ where the sum is performed over all particle-wall collisions of any particle $i$ with either of the side walls (with normals ${\mathbf{n}}_{\mathrm{w}}=\pm{\mathbf{e}}_x$) during $\Delta t$. [Figure \[fig:P-cs\]]{} shows $T$, [$P_\mathrm{int}$]{}, and [$P_\mathrm{wall}$]{}as functions of $L_x^{-1}$. We observe a linear dependence on $L_x^{-1}$ in all three functions. Extrapolations to $L_x\rightarrow\infty$ agree with the values $T_\mathrm{bulk}$ and $P_\mathrm{bulk}$ of temperature and internal pressure, respectively, of separate bulk simulations performed beforehand (dashed lines). Linear fitting and averaging over amplitudes yields $$\label{eq:sfy-cs} \begin{split} T(L_x)=&T_\mathrm{bulk}[1-(0.8\pm 0.1)\sigma L_x^{-1}]\,,\\ {\ensuremath{P_\mathrm{int}}\xspace}(L_x)=&P_\mathrm{bulk}[1-(1.4\pm 0.1)\sigma L_x^{-1}]\,,\\ {\ensuremath{P_\mathrm{wall}}\xspace}(L_x)=&P_\mathrm{bulk}[1+(0.3\pm 0.1)\sigma L_x^{-1}]\,. \end{split}$$ (Note that ${\ensuremath{P_\mathrm{int}}\xspace}(L_x)$ and ${\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ converge to the same value in the limit $L_x\rightarrow\infty$.) $T$ decreases with decreasing $L_x$ and [$P_\mathrm{int}$]{}follows the temperature behavior. [$P_\mathrm{wall}$]{}, however, increases with decreasing $L_x$, which constitutes a nontrivial effect: In a setup, where the side walls were placed in an infinite system with fluid both inside and outside the side walls, the exterior domain would exert a side wall pressure of the extrapolated value. Hence, there would be a net *expanding* force on the side walls even though the granular temperature is smaller at the interior. Qualitatively, the finite size scaling can be described in terms of an excess particle accumulation at the side walls [@rohwer2018nonequilibrium] that changes the density between the walls by a contribution $\propto L_x^{-1}$. ![\[fig:P-cq\] Data for temperature $T$ and side wall pressure ${\ensuremath{P_\mathrm{wall}}\xspace}$ as functions of time $t$ after the quench (logarithmic scale) for several side wall distances $L_x$ as indicated in each panel. [(a)]{} and [(b)]{} $T$ on logarithmic scale. [(c)]{} and [(d)]{} ${\ensuremath{P_\mathrm{wall}}\xspace}$ on logarithmic scale. [(e)]{} and [(f)]{} Temperature difference $\Delta T\equiv T(\infty)-T(L_x)$ between the exterior and interior of the slit on logarithmic scale. [(g)]{} and [(h)]{} Net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}\equiv {\ensuremath{P_\mathrm{wall}}\xspace}(\infty)-{\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ ([i.e., ]{}difference between pressures on outer and inner surfaces of the side walls). Left column [(a)]{}, [(c)]{}, [(e)]{}, and [(g)]{}: strong heating 13 (type I). Right column [(b)]{}, [(d)]{}, [(f)]{}, and [(h)]{}: strong cooling 31 (type II). Data for $L_x=\infty$ is extrapolated from finite $L_x$ at each time, as was done for the steady state values ([cf. ]{}[Fig. \[fig:P-cs\]]{}). ](figure8a.pdf "fig:"){width="\quarternl"} ![\[fig:P-cq\] Data for temperature $T$ and side wall pressure ${\ensuremath{P_\mathrm{wall}}\xspace}$ as functions of time $t$ after the quench (logarithmic scale) for several side wall distances $L_x$ as indicated in each panel. [(a)]{} and [(b)]{} $T$ on logarithmic scale. [(c)]{} and [(d)]{} ${\ensuremath{P_\mathrm{wall}}\xspace}$ on logarithmic scale. [(e)]{} and [(f)]{} Temperature difference $\Delta T\equiv T(\infty)-T(L_x)$ between the exterior and interior of the slit on logarithmic scale. [(g)]{} and [(h)]{} Net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}\equiv {\ensuremath{P_\mathrm{wall}}\xspace}(\infty)-{\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ ([i.e., ]{}difference between pressures on outer and inner surfaces of the side walls). Left column [(a)]{}, [(c)]{}, [(e)]{}, and [(g)]{}: strong heating 13 (type I). Right column [(b)]{}, [(d)]{}, [(f)]{}, and [(h)]{}: strong cooling 31 (type II). Data for $L_x=\infty$ is extrapolated from finite $L_x$ at each time, as was done for the steady state values ([cf. ]{}[Fig. \[fig:P-cs\]]{}). ](figure8b.pdf "fig:"){width="\quarternl"}\ ![\[fig:P-cq\] Data for temperature $T$ and side wall pressure ${\ensuremath{P_\mathrm{wall}}\xspace}$ as functions of time $t$ after the quench (logarithmic scale) for several side wall distances $L_x$ as indicated in each panel. [(a)]{} and [(b)]{} $T$ on logarithmic scale. [(c)]{} and [(d)]{} ${\ensuremath{P_\mathrm{wall}}\xspace}$ on logarithmic scale. [(e)]{} and [(f)]{} Temperature difference $\Delta T\equiv T(\infty)-T(L_x)$ between the exterior and interior of the slit on logarithmic scale. [(g)]{} and [(h)]{} Net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}\equiv {\ensuremath{P_\mathrm{wall}}\xspace}(\infty)-{\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ ([i.e., ]{}difference between pressures on outer and inner surfaces of the side walls). Left column [(a)]{}, [(c)]{}, [(e)]{}, and [(g)]{}: strong heating 13 (type I). Right column [(b)]{}, [(d)]{}, [(f)]{}, and [(h)]{}: strong cooling 31 (type II). Data for $L_x=\infty$ is extrapolated from finite $L_x$ at each time, as was done for the steady state values ([cf. ]{}[Fig. \[fig:P-cs\]]{}). ](figure8c.pdf "fig:"){width="\quarternl"} ![\[fig:P-cq\] Data for temperature $T$ and side wall pressure ${\ensuremath{P_\mathrm{wall}}\xspace}$ as functions of time $t$ after the quench (logarithmic scale) for several side wall distances $L_x$ as indicated in each panel. [(a)]{} and [(b)]{} $T$ on logarithmic scale. [(c)]{} and [(d)]{} ${\ensuremath{P_\mathrm{wall}}\xspace}$ on logarithmic scale. [(e)]{} and [(f)]{} Temperature difference $\Delta T\equiv T(\infty)-T(L_x)$ between the exterior and interior of the slit on logarithmic scale. [(g)]{} and [(h)]{} Net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}\equiv {\ensuremath{P_\mathrm{wall}}\xspace}(\infty)-{\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ ([i.e., ]{}difference between pressures on outer and inner surfaces of the side walls). Left column [(a)]{}, [(c)]{}, [(e)]{}, and [(g)]{}: strong heating 13 (type I). Right column [(b)]{}, [(d)]{}, [(f)]{}, and [(h)]{}: strong cooling 31 (type II). Data for $L_x=\infty$ is extrapolated from finite $L_x$ at each time, as was done for the steady state values ([cf. ]{}[Fig. \[fig:P-cs\]]{}). ](figure8d.pdf "fig:"){width="\quarternl"}\ ![\[fig:P-cq\] Data for temperature $T$ and side wall pressure ${\ensuremath{P_\mathrm{wall}}\xspace}$ as functions of time $t$ after the quench (logarithmic scale) for several side wall distances $L_x$ as indicated in each panel. [(a)]{} and [(b)]{} $T$ on logarithmic scale. [(c)]{} and [(d)]{} ${\ensuremath{P_\mathrm{wall}}\xspace}$ on logarithmic scale. [(e)]{} and [(f)]{} Temperature difference $\Delta T\equiv T(\infty)-T(L_x)$ between the exterior and interior of the slit on logarithmic scale. [(g)]{} and [(h)]{} Net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}\equiv {\ensuremath{P_\mathrm{wall}}\xspace}(\infty)-{\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ ([i.e., ]{}difference between pressures on outer and inner surfaces of the side walls). Left column [(a)]{}, [(c)]{}, [(e)]{}, and [(g)]{}: strong heating 13 (type I). Right column [(b)]{}, [(d)]{}, [(f)]{}, and [(h)]{}: strong cooling 31 (type II). Data for $L_x=\infty$ is extrapolated from finite $L_x$ at each time, as was done for the steady state values ([cf. ]{}[Fig. \[fig:P-cs\]]{}). ](figure8e.pdf "fig:"){width="\quarternl"} ![\[fig:P-cq\] Data for temperature $T$ and side wall pressure ${\ensuremath{P_\mathrm{wall}}\xspace}$ as functions of time $t$ after the quench (logarithmic scale) for several side wall distances $L_x$ as indicated in each panel. [(a)]{} and [(b)]{} $T$ on logarithmic scale. [(c)]{} and [(d)]{} ${\ensuremath{P_\mathrm{wall}}\xspace}$ on logarithmic scale. [(e)]{} and [(f)]{} Temperature difference $\Delta T\equiv T(\infty)-T(L_x)$ between the exterior and interior of the slit on logarithmic scale. [(g)]{} and [(h)]{} Net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}\equiv {\ensuremath{P_\mathrm{wall}}\xspace}(\infty)-{\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ ([i.e., ]{}difference between pressures on outer and inner surfaces of the side walls). Left column [(a)]{}, [(c)]{}, [(e)]{}, and [(g)]{}: strong heating 13 (type I). Right column [(b)]{}, [(d)]{}, [(f)]{}, and [(h)]{}: strong cooling 31 (type II). Data for $L_x=\infty$ is extrapolated from finite $L_x$ at each time, as was done for the steady state values ([cf. ]{}[Fig. \[fig:P-cs\]]{}). ](figure8f.pdf "fig:"){width="\quarternl"}\ ![\[fig:P-cq\] Data for temperature $T$ and side wall pressure ${\ensuremath{P_\mathrm{wall}}\xspace}$ as functions of time $t$ after the quench (logarithmic scale) for several side wall distances $L_x$ as indicated in each panel. [(a)]{} and [(b)]{} $T$ on logarithmic scale. [(c)]{} and [(d)]{} ${\ensuremath{P_\mathrm{wall}}\xspace}$ on logarithmic scale. [(e)]{} and [(f)]{} Temperature difference $\Delta T\equiv T(\infty)-T(L_x)$ between the exterior and interior of the slit on logarithmic scale. [(g)]{} and [(h)]{} Net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}\equiv {\ensuremath{P_\mathrm{wall}}\xspace}(\infty)-{\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ ([i.e., ]{}difference between pressures on outer and inner surfaces of the side walls). Left column [(a)]{}, [(c)]{}, [(e)]{}, and [(g)]{}: strong heating 13 (type I). Right column [(b)]{}, [(d)]{}, [(f)]{}, and [(h)]{}: strong cooling 31 (type II). Data for $L_x=\infty$ is extrapolated from finite $L_x$ at each time, as was done for the steady state values ([cf. ]{}[Fig. \[fig:P-cs\]]{}). ](figure8g.pdf "fig:"){width="\quarternl"} ![\[fig:P-cq\] Data for temperature $T$ and side wall pressure ${\ensuremath{P_\mathrm{wall}}\xspace}$ as functions of time $t$ after the quench (logarithmic scale) for several side wall distances $L_x$ as indicated in each panel. [(a)]{} and [(b)]{} $T$ on logarithmic scale. [(c)]{} and [(d)]{} ${\ensuremath{P_\mathrm{wall}}\xspace}$ on logarithmic scale. [(e)]{} and [(f)]{} Temperature difference $\Delta T\equiv T(\infty)-T(L_x)$ between the exterior and interior of the slit on logarithmic scale. [(g)]{} and [(h)]{} Net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}\equiv {\ensuremath{P_\mathrm{wall}}\xspace}(\infty)-{\ensuremath{P_\mathrm{wall}}\xspace}(L_x)$ ([i.e., ]{}difference between pressures on outer and inner surfaces of the side walls). Left column [(a)]{}, [(c)]{}, [(e)]{}, and [(g)]{}: strong heating 13 (type I). Right column [(b)]{}, [(d)]{}, [(f)]{}, and [(h)]{}: strong cooling 31 (type II). Data for $L_x=\infty$ is extrapolated from finite $L_x$ at each time, as was done for the steady state values ([cf. ]{}[Fig. \[fig:P-cs\]]{}). ](figure8h.pdf "fig:"){width="\quarternl"}\ Next, we turn to quenches of the Casimir geometry, and investigate whether there is an effect beyond these steady state confinement effects. The dynamics of $T$ is shown for a type I quench in [Fig. \[fig:P-cq\][(a)]{}]{} and a type II quench in [Fig. \[fig:P-cq\][(b)]{}]{}. As in the bulk setup (cf. [Fig. \[fig:PT-bqA\]]{}), $T$ relaxes to the steady state value $T{_\mathrm{F}}$ exponentially in the type I quench but algebraically in the type II quench. Other type I quenches between $A_1\leftrightarrow A_2$ (not shown) also show the same behavior as [Fig. \[fig:P-cq\][(a)]{}]{}. [$P_\mathrm{wall}$]{}shown for heating in [Fig. \[fig:P-cq\][(c)]{}]{} and cooling in [Fig. \[fig:P-cq\][(d)]{}]{} exhibits similar behavior as $T$. The main difference is the inverted finite size scaling ([i.e., ]{}deviation from the limit $L_x\rightarrow\infty$) of the initial and final states, which is consistent with the steady state results ([cf. ]{}[Fig. \[fig:P-cs\]]{}). The temperature differences $\Delta T$ between the exterior and interior [\[[Fig. \[fig:P-cq\][(e)]{}]{} and [(f)]{}\]]{} transition monotonically from the initial to the final steady state values plotted in [Fig. \[fig:P-cs\][(a)]{}]{} and [(c)]{}. The net side wall pressure $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}$ [\[[Fig. \[fig:P-cq\][(g)]{}]{} and [(h)]{}\]]{}, however, shows an undershoot or overshoot in the case of heating or cooling, respectively. The undershoots in the heating simulations [\[panel [(g)]{}\]]{} for all $L_x$ take place at $t\approx2-3\tau_0$, which is roughly the time at which the fluid is fully heated [\[[cf. ]{}panel [(a)]{}\]]{}. The times $t\approx1-2\tau_0$ of the overshoots at cooling [\[panel [(h)]{}\]]{} coincide with the start of the cooling of the fluid [\[[cf. ]{}panel [(b)]{}\]]{}. Moderate heating 12 shows the same behavior as strong heating and moderate cooling 21 shows the same behavior as strong cooling (not shown). The extremum is, however, not as pronounced upon moderate amplitude changes. We would like to stress that the qualitative behavior here does not correspond to the type I and II classification of the previous section but depends on whether the fluid is cooled or heated. {width="\thirdnl"} {width="\thirdnl"} {width="\thirdnl"}\ In general, non-monotonic behavior of the pressure and temperature may indicate a separation of time scales as regards dynamics in the vicinity of the surfaces and dynamics of the bulk. Overshoots of this type have, [e.g., ]{}been observed in drift-diffusion systems [@khalilian2020interplay]. To explain the net pressure overshoot in our system, we collapse the data and obtain the finite size scaling. We assume that the data follows the scaling relation $$\label{eq:scaling-walls} y(t/{\ensuremath{s_t}\xspace}(L),L)={\ensuremath{s_y}\xspace}(L)y(t,\infty)$$ for all considered observables $y=T, {\ensuremath{P_\mathrm{int}}\xspace}, {\ensuremath{P_\mathrm{wall}}\xspace}$, with scaling factors [$s_t$]{}and [$s_y$]{}describing the scaling of time and of the final steady state value, respectively. The scaling factors are determined by minimizing an error function defined via $$E({\ensuremath{s_t}\xspace},{\ensuremath{s_y}\xspace}) \equiv \frac{1}{t{_\mathrm{F}}}\int_{0}^{t{_\mathrm{F}}} \mathrm{d}t\left[\frac{y(t/{\ensuremath{s_t}\xspace},L)}{{\ensuremath{s_y}\xspace}y(t,\infty)} -1\right]^2,$$ with an upper bound of integration $t{_\mathrm{F}}$, which is in practice given by the range of simulation data. This error function is defined such that $E=0$ if [Eq. ]{} is satisfied and $E>0$ otherwise. With the input simulation data for $y(t,L)$ and $y(t,\infty)$ being noisy, the true values of [$s_t$]{}and [$s_y$]{}are the ones that minimize $E$. These are determined simply by sampling a fine grid in the ${\ensuremath{s_t}\xspace}{\ensuremath{s_y}\xspace}$-plane and taking the minimum value. The position of the minimum $({\ensuremath{s_t}\xspace},{\ensuremath{s_y}\xspace})$ does not depend on the choice of $t{_\mathrm{F}}$ as long as $t{_\mathrm{F}}$ is greater than the time the fluid needs to relax. We judge the quality of this numerical data collapse by the minimal value of $E$ being small. This criterion is fulfilled for type I protocols with $E < 0.001 (L_x/\sigma)^{-1}$. For type  II protocols however we obtain values of up to $E \approx 0.1 (L_x/\sigma)^{-1}$. Therefore, we conclude that the finite size scaling of cluster instability is not captured adequately by the simple two-parameter scaling of [Eq. ]{}. [Figures \[fig:scaling-cq\][(a)]{}]{} and [(b)]{} show [$s_t$]{}and [$s_y$]{}as functions of $L_x^{-1}$. To a good approximation all scaling factors are linear functions of $L_x^{-1}$ with slopes $a_t$ and $a_y$ plotted against each other in [Fig. \[fig:scaling-cq\][(c)]{}]{}. The $a_y$ are equal to our steady state results in [Eq. ]{} within the uncertainties (which serves as an additional consistency check for the minimization procedure). Values for $a_t$ are in the range $0.2\leq a_t \leq 0.5$ for the different quench protocols, which implies an enhanced relaxation in the presence of side walls. With this, the non-monotonic ${\ensuremath{P_\mathrm{wall}}\xspace}$ can be explained as follows. The undershoot in the side wall pressure observed upon increasing $A$ [\[[Fig. \[fig:P-cq\][(g)]{}]{}\]]{} takes place when the fluid between the side walls is already completely heated, while the pressure of the fluid outside is smaller as it is still adjusting to the post-quench amplitude. The overshoot in the reverse process [\[[Fig. \[fig:P-cq\][(h)]{}]{}\]]{} takes place when the inner fluid has started to cool faster than the fluid outside. The obtained values for the slopes of the scaling factors also explain why the overshoot is not observed in $T$ or [$P_\mathrm{int}$]{}: Here the overshoot through the enhanced relaxation is overlaid by decreasing of the (initial and final) steady state values (as $a_t<0$ and $a_y<0$ and $|a_y|>|a_t|$ for $y=T,{\ensuremath{P_\mathrm{int}}\xspace}$). The slightly positive value of $a_{{\ensuremath{P_\mathrm{wall}}\xspace}}$ on the other hand slightly enhances the overshoot. ![\[fig:t-XL\] Extremal times $t{_\mathrm{E}}$ of the side wall pressure under-/overshoots (filled symbols) versus side wall distance $L_x$ and time $t$ [\[open symbols; same data as [Fig. \[fig:XS-CS-t\][(a)]{}]{}\]]{} versus distance $r{_\mathrm{E}}$ of the local extrema of the total correlation function $h$. Same quench protocols are depicted as same symbol shapes and colors. Lines are guides to the eye. ](figure10.pdf){width="\halfnl"} In order to affirm that the bulk correlations described in Sec. \[sec:quench-bulk\] are not the origin of the transient net pressure overshoot, we compare the dynamic scaling of both effects. To this end, we extract the times $t{_\mathrm{E}}$ at which the extrema of $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}$ are attained by fitting quadratic functions in the vicinity of the extremum (displayed in [Fig. \[fig:t-XL\]]{}). For most protocols $t{_\mathrm{E}}$ increases only slightly with $L_x$, which does not match the scaling obtained for the bulk correlations (also plotted in [Fig. \[fig:t-XL\]]{}). This observation disqualifies the density correlations as origin of the non-monotonic $\Delta {\ensuremath{P_\mathrm{wall}}\xspace}$. Discussion and Conclusion {#sec:con} ========================= The quenched granular quasi system considered here exhibits many nontrivial phenomena, several of which were observed upon disturbing the steady state by changing the driving amplitude. The first is the emergence of propagating density correlations in bulk on length scales beyond the commonly known fluid structure upon heating or moderate cooling [\[type I; [Figs. \[fig:corr-bqA\][(a)]{}]{}–[(d)]{}\]]{}. Qualitatively such an effect has been predicted recently for soft Brownian particles. However, we identified several crucial differences regarding the dynamic scaling and the functional form of the correlations. An assumption in the theory that is not met by our setup is that of instantaneous temperature change. Here we change the driving amplitude instantaneously and the temperature of the inertial particles slowly adapts to the new amplitude. This was shown to make an important difference as saturation of the temperature causes a crossover in the scaling behavior ([Fig. \[fig:XS-CS-t\]]{}). Remarkably, correlation functions were shown to collapse onto universal position-dependent curves ([i.e., ]{}scaling functions) when rescaled appropriately ([Fig. \[fig:C-X-rescaled\]]{}). This universality is robust across different simulation sizes. Although we did not decipher the various features of these curves, we also ascribe their emergence to the gradual temperature change. In the saturated ([i.e., ]{}long time) regime we could identify the scaling of the correlation strength $\beta=-2$ as a universal geometric property. On the other hand, the ballistic scaling of the dynamic exponent $\alpha=1$ is remarkable because the individual particle motion is diffusive at the considered length scales. This observation, as well as the origin of the salient functional forms of the scaling functions remain open puzzles for future studies. A fruitful path to this end may emerge via the testing of a local conservation law with a possible source term via explicit sampling of particle currents [@schindler2016dynamic]. The asymmetry between energy gain and dissipation induces an asymmetry between heating and cooling. Hence, when cooling the fluid starting from a large $A$ we observe an additional effect named clustering instability[\[type II; [Figs. \[fig:corr-bqA\][(e)]{}]{} and [(f)]{}\]]{}. This effect, which is well known for inelastic systems, is greater and hence overlays the former one. In particular, the true length scale of the clustering instability could not be determined in our simulations, as it exceeded even our largest simulation boxes. We further investigated the finite size scaling of temperature and pressure by adding side walls to the setup. While the internal pressure follows the scaling of the temperature, the pressure exerted on the side walls behaves differently. In the steady state, the side wall pressure deviation from the infinite size limit has the opposite sign than the temperature difference ([Fig. \[fig:P-cs\]]{}). Moreover, after a quench we observe nonmonotonic behavior in the side wall pressure ([Fig. \[fig:P-cq\]]{}). Numerical data collapse revealed that this is a consequence of an enhanced relaxation speed of the fluid between the side walls ([Fig. \[fig:scaling-cq\]]{}). In combination with the anomalous steady state behavior, this results in an over-/undershoot in the side wall pressure that is not observed in the temperature or the internal pressure. Future plans include investigations of a setup with tuned densities inside and outside the slit such that net side wall pressure in steady state is zero, in order to isolate the transient contribution. A further open question and a possible next objective involves studying forces between compact inclusions immersed in the fluid. TS was supported by the Deutsche Forschungsgemeinschaft as part of the Forschergruppe GPSRS under Grant No. ME1361/13-2. We thank Sebastian Kapfer for fruitful discussions throughout the genesis of this article. Steady state bulk {#sec:steady-bulk} ================= ![\[fig:bs\] [(a)]{} Vertical density distribution $\rho_z$ as a function of the $z$ coordinate, and [(b)]{} projected total correlation function $h$ as a function of distance $r$, for steady state fluids at amplitudes $A_1$, $A_2$, and $A_3$. ](figure11a.pdf "fig:"){width="\halfnl"} ![\[fig:bs\] [(a)]{} Vertical density distribution $\rho_z$ as a function of the $z$ coordinate, and [(b)]{} projected total correlation function $h$ as a function of distance $r$, for steady state fluids at amplitudes $A_1$, $A_2$, and $A_3$. ](figure11b.pdf "fig:"){width="\halfnl"} Here, we discuss the steady state fluid as present in the initial state prior to the quench and in the final state infinitely long time after the quench. The aim is to demonstrate how stratification creates effectively soft particles. [Figure \[fig:bs\][(a)]{}]{} shows the vertical density distribution $$\rho_z(z)\equiv \frac{L_z}{L^2}\left\langle\sum_{i=1}^{N}\delta(z_i(t) - z)\right\rangle$$ at the three investigated $A$ (normalized such that its mean equals $\rho$). At high amplitude $A_3$, $\rho_z$ is almost symmetric and the particles fill the whole space between the plates. The peaks at the top and bottom plates originate from the mutual repulsion of the particles. At low amplitude $A_1$, however, we observe strongly barometric ([i.e., ]{}exponentially decaying) distribution of particles, where most particles are located near the bottom plate and are only slightly hopping. At intermediate amplitude $A_2$ the particles are partly stratified. The differences in stratification also make an impact on $h$ [\[shown in [Fig. \[fig:bs\][(b)]{}]{}\]]{}. As mentioned before, we obtain $h$ from the projected $xy$-coordinates of the particles. Therefore, even though the particles are hard and cannot penetrate each other, we can observe a nonzero contribution of $h$ at $r<\sigma$ originating from particles that are (partly) on top of each other. This contribution is larger at high amplitudes $A_2$ and $A_3$, where particles fill the whole space between the plates, and smaller at low amplitude $A_1$, where most of the particles populate a single layer near the bottom plate. In a description of the system one can therefore consider the fluid as effectively soft.

ArXiv

--- abstract: 'Many of the systems that are traditionally analyzed as complex networks have natural interpretations as multiplex structures. While these formulations retain more information than standard network models, there is not yet a fully developed theory for computing network metrics and statistics on these objects. As many of the structural representations associated to these models can distort the underlying characteristics of dynamical process, we introduce an algebraic method for modeling arbitrary dynamics on multiplex networks. Since several network metrics are based on generalized notions of dynamical transfer, we can use this framework to extend many of the standard network metrics to multiplex structures in a consistent fashion.' author: - - bibliography: - 'Multiplex.bib' title: Network models that reflect multiplex dynamics --- Multiplex Networks, Network Dynamics Introduction ============ Mathematical analyses of simple network models of complex systems provide a surprising amount of information - determining important components of the system [@borgatti_centrality_2005], finding hidden communities [@newman_detecting_2004], demonstrating the robustness of the system to failures [@barabasi_scale-free_2003; @foti_stability_2013], among many others. But the simplest network models, where nodes represent components of the system and edges represent interaction or association between two nodes, are limited as they conflate different types of interaction between nodes. A nascent theory of multiplex networks (see [@kivela_multilayer_2014] for a thorough review of the current research) deals with this limitation by allowing for an all-encompassing structure with multiple layers, one for each type of interaction, where each layer is a network on the same set of nodes. These structures arise naturally in many settings: trade networks in economics [@barigozzi_multinetwork_2010; @barigozzi_community_2011], social networks [@kapferer_strategy_1972; @krackhardt_cognitive_1987], and transportation networks [@domenico_navigability_2014; @gallotti_multilayer_2015], among others. While researchers have recently focused on describing structural representations of multiplex networks and analyzing their properties [@de_domenico_mathematical_2013; @cozzo_structure_2013], only a relatively small portion of the literature engages questions about dynamics on multiplex networks. As structural representations of simple networks that reflect dynamics are essential to some aspects of network analysis - including diffusion processes, random walk probabilities, clustering, and percolation - transporting these ideas to the multiplex setting requires structural representations consistent with multiplex dynamics. Studying multiplex structures using tools from complex networks exposes a tension between structural representations of the entire multiplex and the dynamical interpretations of the individual layers. Since the edges in traditional networks only express a single type of relationship, representing a multiplex as a single network can lead to distortion of the very topological properties we are trying to understand. A meaningful global dynamic model should respect the individual topological behaviors of the layers without introducing confounding structural effects. In this paper, we revisit some recent multiplex structural representations, examining them through the lens of dynamical consistency to show that in many cases the structural definitions introduce factors that confound the analysis of the multiplex dynamics. To remedy this, we develop a new method for modeling multiplex systems which respects dynamics and derive spectral results similar to those known for monoplex networks. Related Work {#related-work .unnumbered} ------------ Kivelä et al. [@kivela_multilayer_2014] provide a descriptive overview of many of the methods and techniques that are being developed for studying multiplex structures. Currently, these methods favor tensorial representations [@de_domenico_mathematical_2013], as they contain all available data, in contrast to earlier attempts which aggregated multiplex data into a single network [@zachary_information_1977]. Many standard monoplex metrics and statistics have been extended to multiplex structures. For example Cozzo et al. describe a generalized notion of clustering coefficients for multiplex structures in [@cozzo_structure_2013], while De Domenico et al. generalized a wider variety of single network techniques to the multilayer setting [@de_domenico_mathematical_2013]. A study of the redundant information captured by these representations is presented in [@de_domenico_structural_2015]. A first approach, [@leicht_percolation_2009], studied dynamics on disaggregated multiplex structures via percolation theory. Buldyrev et al. [@buldyrev_catastrophic_2010] considered cascading failures on related networks. Building on this work, Gao et al. characterized robustness of interconnected networks [@gao_robustness_2011] and developed models for studying the connectivity of interconnected networks [@gao_networks_2012]. Recently, a $k-$core approach to these problems in the multiplex setting was discussed in . Several studies have extended monoplex centralities, such as eigenvector centrality, Katz centrality, and various geodesic centralities, to multiplex networks [@de_domenico_mathematical_2013; @de_domenico_centrality_2015; @de_domenico_ranking_2015]. Our dynamical method presented in this paper is a generalization of the Khatri-Rao product approach to centrality problems introduced by Sol et al. in [@sola_eigenvector_2013]. To study distributed consensus problems on multiplex structures Trpevski et al. [@trpevski_discrete-time_2014] proposed a supratransition matrix as an analogue of the standard stochastic random walk matrix for networks. Their stochastic model relates to De Domenico et al.’s model of random walks on multiplex structures [@domenico_navigability_2014]. Both of these constructions are special cases of the method we develop, see Section IV. Additionally, our work provides a general framework for interpreting similar results on arbitrary multiplex structures not arising from these specific applications. In [@gomez_diffusion_2013], the authors describe a process for modeling diffusion in multiplex networks by constructing a generalized Laplacian. They used pertubation theory to characterize the eigenvalues of the supra–Laplacian in order to discuss rates of diffusion across multiplex networks [@sole-ribalta_spectral_2013]. Further spectral considerations of multiplex structures were studied in [@sanchez-garcia_dimensionality_2014]. Additionally, the structural behavior of multiplex networks has been connected to the eigenvalues of this generalized Laplacian [@radicchi_driving_2014; @radicchi_abrupt_2013], although these results may be artifacts of the composition of the operator under consideration [@garrahan_comment_2014]. Topological Consequences of Structural Representations ====================================================== There are two main constructions or structural models in the literature for generating such a network, aggregation and matched summation. Aggregation combines all edges between pairs of nodes into a single edge, while matched summation (called diagonal, categorical multilayer networks in [@kivela_multilayer_2014]) creates a single network from the layers by introducing new edges connecting the nodes to copies of themselves in other layers. Figure 1 shows aggregate and matched sum networks for an example two-layer multiplex. Both of these structural models distort network statistics relative to those of the original layers. The main drawback to aggregation is that it loses information by conflating different types of interactions. The World Trade Web network provides an important example where aggregation obscures much of the topological heterogeneity of the individual layers, particularly with respect to clustering and reciprocity [@barigozzi_multinetwork_2010]. Unlike aggregate networks, no information is lost when forming the matched sum network of a multiplex. Matched summation corresponds to unfolding/flattening the tensorial representation of a multiplex structure [@de_domenico_mathematical_2013], which underlies the centrality measures presented in [@de_domenico_centrality_2015], as well as many of the network metrics discussed in [@de_domenico_mathematical_2013; @kivela_multilayer_2014]. Although the tensorial construction preserves all of the structural information about the multiplex, unfolding distorts the original data. There are two main problems with matched summation. First, the added interlayer edges have a different role than the intralayer edges, but the two are conflated. Second, connecting all the copies of a given node to one another creates dense cliques throughout the matched sum. Many of the dynamics based on flattening/unfolding tensorial representations of multiplex structures implicitly incorporate these distortions in a fundamental way. We analyze the global network statistics of the Krackhardt High–Tech Managers data [@krackhardt_cognitive_1987] and a collection of random multiplex networks, comparing the values derived from the original layers with the values associated to the single network representations. For the random networks, we follow [@trpevski_discrete-time_2014] and use a mixture of Erdös–Renyi, Barabasi–Albert, and Watts–Strogatz networks on 100 nodes per layer. The results, summarized in Table 1 and Figure 2 respectively, show that these multiplex models obscure the heterogeneity of the layer statistics. The Krackhardt data gives rise to a multiplex with directed layers, which poses a particular problem for the matched sum model as the structural conflation can be particularly misleading. Each of the three layers in the Krackhardt data represents a particular type of association: advice relations, friendship relations, and direct report relations in the company. However, the matched sum does not distinguish between these types of connections. Additionally, the edges between copies of the same individual do not have a natural interpretation. Statistic Advice Friend Report Aggregate Matched Sum -------------- -------- -------- -------- ----------- ------------- Density 0.452 0.243 0.0476 0.552 0.112 Transitivity 0.465 0.276 0.000 0.561 0.311 Reciprocity 0.236 0.225 0.000 0.314 0.299 Mean Degree 18.1 9.71 1.90 22.1 13.9 : Comparison of Multiplex Global Statistics The distortions we witness in these examples are not isolated: given an arbitrary multiplex with $n$ nodes and $k$ layers, we compute bounds on many of the global statistics of the aggregate and matched sum networks in terms of the layer statistics. **Density:** *The density of the aggregate monoplex is greater than or equal to the density of the densest layer. Matched sum constructions tend to be less dense than the original layers since the only inter–layer connections occur between copies of the same node. Thus, the maximum number of neighbors of a node is $(n-1)+(k-1)$ out of the $nk-1$ total nodes in the matched sum.* **Transitivity:** *The transitivity of the aggregate monoplex is greater than or equal to the transitivity of the most transitive layer. Matched sum networks tend to be less transitive than the original layers since the only inter layer triangles occur between the copies of each node. None of the connected triples formed by one inter–layer edge and one intra-layer edge can be closed, diluting the proportion of transitive triples.* **Reciprocity:** *The reciprocity of the aggregate monoplex is greater than or equal to the reciprocity of the most reciprocal layer. The reciprocity of the matched sum is greater than or equal to the reciprocity of the least reciprocal layer since all of the connecting arcs are bidirectional.* **Mean Degree:** *The mean degree of the aggregate monoplex is greater than or equal to the maximum mean degree of the layers. The mean degree of the aggregate monoplex is greater than or equal to the minimum mean degree of the layers because each node receives $k-1$ new incident edges. Although the mean degree and density capture the same information about the original layers, the matched sum perturbs each statistic differently.* **Average Path Length:** *The average path length of the aggregate monoplex is less than or equal to the shortest average path length of the layers because the shortest paths from each layer are still viable in the aggregate network. In matched sum networks, the average path length also tends to decrease since it requires at most two interlayer steps to make use of the shortest intra–layer path between two nodes.* **Clique Numbers:** *The clique numbers of the aggregate and matched sum networks are both be greater than or equal to the maximum clique number among the layers.* Except in the case of degenerate examples, such as duplicate or empty layers, the bounds above become strict inequalities. We observe that increasing the number of layers magnifies the affects on these statistics. For example, as the number of layers increases with a fixed number of nodes and average density, the aggregate density increases while the matched sum density decreases. Similar results hold for the other statistics. Taken together, these results emphasize the toplogical damage inherent in using aggregate or matched sum representations of multiplex networks. Although for some applications, such as the transportation networks studied in [@domenico_navigability_2014], these structural additions accurately model the system, researchers should consider possible confounding effects before applying these techniques. Dynamical Consequences of Structural Representations ==================================================== In addition to distorting the basic topological properties, forming aggregate or matched sum networks from multiplex data also distorts multiplex dynamics. We identify two issues – instances where interpretations of the dynamics are difficult, and where these models provide incorrect generalizations of standard dynamics. The structural distortions we observed in the previous section interfere with such optimizations. As the problems with aggregate models have been previously noted by Barigozzi et al. in the context of the World Trade Web network [@barigozzi_multinetwork_2010; @barigozzi_community_2011], and more generally by Kivelä et al. in their survey [@kivela_multilayer_2014], we focus here on the effects of matched summation models and, consequently, of tensorial representations. To illustrate the problem of interpretability of tensorial representations, we demonstrate the consequences of conflating edge types by considering the simplest associated dynamical model. Here, the tensor acts on supra–vector of quantities, representing flow along the “edges” of the structure. The resulting quantity at node $j$ layer $i$ is the sum of the neighbors of $j$ along layer $i$ combined with the sum of the values at $\beta$ for all other layers. Although the intra–layer component corresponds directly to familiar monoplex dynamics, the inter–layer component reflects the effect of adding edges between the node copies in the matched sum. Additionally, all of the copies of node $j$ represent the same object: it is unclear what meaning should be attached to the edges connecting the layers. Together these problems of interpretation make it difficult to describe meaningful objective functions by directly viewing the tensorial representation as an operator. This problem also arises when considering heterogeneous multiplex data, for example when some of the layers of the multiplex are directed while others are undirected. While it is true that an undirected edge may be considered as a pair of opposing directed edges with no loss of structural information, the dynamical interpretation of such relations can be quite different and is entirely obscured by a naïve model. Indeed, the more heterogeneous the layers, the more important it is not to add confounding structural factors that obscure the effects of the topology on dynamical models. Questions of interpretation are not the only issue inherent in these tensorial dynamics. The structural distortions described in the previous section have clear dynamical consequences. In a matched sum, each node belongs to a clique representing all its copies across the layers, vastly distoring local topologies in networks with sparse layers. This in turn distorts metrics and dynamics computed with walks and geodesics, since the number of possible paths between two nodes is greatly magnified. For example, measures like Katz centrality and Pagerank penalize longer paths by an exponential factor, so most of the centrality value accrues directly from the other copies of the same node, overwhelming the actual connections of interest. The combinatorial supra–Laplacian, as defined in [@de_domenico_mathematical_2013] and [@kivela_multilayer_2014] suffers from these deficiencies. Interpreted as a linear differential operator on the tensor space, it represents the Laplacian of the matched sum of the multiplex and hence incorporates these structural distortions at a fundamental level. Although possibly a reasonable model in some applications, it is not a clear dynamical generalization of the standard graph Laplacian. Consequently, we cannot derive the relations between the eigenvalues of this structural representation and the multiplex properties analogously to the monoplex derivation. Finally, as discussed above as the number of layers increases, the effects of these distortions increase. This explains the behaviors noted in several papers based on dynamical models associated to tensorial or supra– constructions that for large portions of the parameter space, the dynamical process is controlled by the layer mixing [@de_domenico_centrality_2015; @de_domenico_ranking_2015; @domenico_navigability_2014; @gomez_diffusion_2013; @radicchi_driving_2014; @radicchi_abrupt_2013; @sole-ribalta_spectral_2013; @sanchez-garcia_dimensionality_2014; @trpevski_discrete-time_2014]. It is unclear at this point what information about multiplex models is actually related to the spectral structures of the associated tensor representations since the corresponding dynamic operators do not always have a natural interpretation. Multiplex Dynamics ================== We present a model for multiplex dynamics that avoids the structural conflation of single network models by allowing effects to pass between layers without artificially adding edges to the network. For a multiplex structure on $n$ nodes and $k$ layers with a matrix dynamical operator $D_i$ associated to each layer $1\leq i\leq k$ we model the intra–layer dynamics by defining the block diagonal matrix $D=\operatorname{diag}(D_1,D_2,\ldots,D_k)$. The matrix $D$ acts on an $nk\times 1$ vector $v$ representing the “quantities” contained at $k$ copies of each if the $n$ nodes. Notationally, $v_{(i-1)n+j}$ is $ \textrm{the quantity at node}\ j$ $\textrm{in layer}\ i$. To extract layer quantities, we write $v= \begin{bmatrix}v^1,v^2,\cdots v^k\end{bmatrix}^T$ where each $v^i$ is a $1\times n$ vector associated to layer $i$. That is, $v^i_j \equiv v^i(j)=v_{(i-1)n+j}$ is $\textrm{the amount at node}\ j\ \textrm{in layer}\ i$. Motivation {#motivation .unnumbered} ---------- We want to interleave the intra–layer dynamics with an operation that gathers the quantity stored at each copy of a given node and redistributes among the other copies of that node. We model this as an iterative two step process: we first apply $D$ to $v$ and second enforce the interlayer dynamics by setting $(v')_j^i$ to a proportion of a convex combination of the $\{(Dv)_j^\ell\}_{\ell=1}^k$, $$(v')_j^i=\alpha^i_j\sum_{\ell=1}^k c_j^{i,\ell}(Dv)_j^\ell.$$ The constants $c_j^{i,\ell}$ provide the rate at which quantities pass from layer $\ell$ to layer $i$ through node $j$, while the $\alpha^i_j$ allow node $j$ in layer $i$ to send or receive quantities external to the network. The restriction to a convex combination of the $(Dv)^\ell_j$ enforces that $c^{i,\ell}_j\geq0$ and $\sum_{\ell=1}^kc_j^{i,\ell}=1$ for all $1\leq j\leq n$ and $1\leq i,\ell\leq k$. Scaling by $\alpha^i_j$ allows any linear combination of the $(Dv)^i_j$ with non–negative coefficients. As we will see, we can perform this second operation algebraically using a collection of scaled orthogonal projection operators, one for each node in the multiplex. This dynamic process preserves the intra–layer dynamics, while also allowing effects to pass between layers through the copies of each node. Unless the intra–layer dynamics have a component that incorporates the value at the node, the copies of a node do not interact directly. Instead the dynamical effects acting on any particular copy are passed through to the other copies, with proportions dictated by the coupling constraints, the $\alpha^i_j$ and $c^{i,\ell}_j$. Consequently, we avoid the pitfalls of aggregate and matched sum networks described above. Derivation of the Operator {#derivation-of-the-operator .unnumbered} -------------------------- We accomplish step one of our method with the block diagonal matrix $D$, so it suffices to construct a matrix that redistributes the values between the copies of the node as described in equation (1). Let $P^i_j=P_{(i-1)n+j}$ be the projection onto the $(i-1)n+j$ coordinate, $P^i_j$ is a $1\times nk$ vector with a one in the $(i-1)n+j$ entry and zeros elsewhere. Then, $P^i_j Dv=(Dv)^i_j$. Using the $\alpha^i_j$ and $c^{i,\ell}_j$, we proportionally distribute this value among all copies of node $j$ using the $nk\times 1$ column vector $$\alpha_j^i\begin{bmatrix}c^{1,i}_j\delta_1(j),\ldots, c^{1,i}_j\delta_{n}(j),\ldots, c_j^{k,i}\delta_1(j), \ldots, c_j^{k,i}\delta_{n}(j)\end{bmatrix}^T.$$ The projections are pairwise orthogonal, so we perform the projection and redistribution steps concurrently using the matrix $$M=\sum_{i=1}^k \sum_{j=1}^n \alpha_j^i\begin{bmatrix}c^{1,i}_j\delta_1(j)\\ \vdots\\ c^{1,i}_j\delta_{n}(j)\\ \vdots \\ c_j^{k,i}\delta_1(j) \\ \vdots \\ c_j^{k,i}\delta_{n}(j)\end{bmatrix}P^i_j.$$ We realize $M$ as block matrix, $$M=\begin{bmatrix} C^{1,1}&C^{1,2}&\cdots&C^{1,k}\\ C^{2,1}&C^{2,2}&\cdots&C^{2,k}\\ \vdots&\ddots&\ddots&\vdots\\ C^{k,1}&C^{k,2}&\cdots&C^{k,k} \end{bmatrix},$$ where $C^{i,\ell}=\operatorname{diag}(\alpha^{\ell}_1c^{i,\ell}_1,\alpha^{\ell}_2c^{i,\ell}_2,\ldots,\alpha^{\ell}_nc^{i,\ell}_n)$. The final multiplex dynamic operator is a product of the layer dynamics matrix $D$ and the redistribution matrix $M$ $$\mathfrak{D}=MD=\begin{bmatrix} C^{1,1}D_1&C^{1,2}D_2&\cdots&C^{1,k}D_k\\ C^{2,1}D_1&C^{2,2}D_2&\cdots&C^{2,k}D_k\\ \vdots&\ddots&\ddots&\vdots\\ C^{k,1}D_1&C^{k,2}D_2&\cdots&C^{k,k}D_k \end{bmatrix}.$$ Before examining the spectral properties of $\mathfrak{D}$ we discuss three special cases of $\mathfrak{D}$ reflecting natural modeling constraints. We begin with a useful definition, describing a hypothesis that enforces a conservation law for the quantities under consideration. The operator $\mathfrak{D}$ is called closed if $\alpha^i_j=1$ for all $1\leq i\leq k$ and $1\leq j\leq n$ When $\mathfrak{D}$ is closed, the matrix $M$ is stochastic, which we will use in the next section when analyzing multiplex random walk dynamics. When $\mathfrak{D}$ is closed and the $D_i$ are random walk matrices we recover the models used in [@domenico_navigability_2014] and [@trpevski_discrete-time_2014]. We motivate our first simplification, which we call the [*unified node*]{} model, with an example using economic exchange networks. Consider trade relationships between countries, where bilateral trade occurs within commodity based layers, but parties inside the countries redistribute money internally redistributed before international trade begins anew. This allows countries to use surpluses in one area to fund deficits in another. Further, the $\alpha$ coefficients represent unmodeled sources or sinks of income external to the trade network. To adjust our model, we enforce a homogeneity condition at each node: instead of computing $M$ from $nk$ projections we use the $n$ projections $Q_j=\sum_{\ell=1}^k\alpha_j^\ell P_j^\ell, j \in \{1,\dots,n\}$. The value of the product $Q_jDv$ is the weighted sum of the quantities at node $j$ on each layer. When $\mathfrak{D}$ is closed, $Q_j$ simply becomes the uniform projection onto the subspace spanned by all of the copies of node $j$. For each node $j$, we redistribute $Q_jDv$ proportionally among its copies according to their relative importance as defined by the network structure or application. This is equivalent to the condition $c_j^{i,\ell}=c_j^{i,m}$ for all $1\leq i,\ell,m\leq k$ and $1\leq j\leq n$. For our second simplification, the [*hierarchical layer*]{} model, we assume there is a hierarchy of layers where the effect of layer $i$ on layer $\ell$ is fixed for all the nodes. In this case, each of the matrices $C^{i,\ell}$ is just a linear homothety or scalar multiple of the identity, i.e. $c^{i,\ell}_a=c^{i,\ell}_b$ and $\alpha^\ell_a=\alpha^\ell_b$ for all layers $a$ and $b$. We see two natural approaches to defining these coefficients in the absence of an application specific determination. The first is to use the density of the layers as the $C^{i,\ell}$. Second, we set $C^{i,\ell}$ to be the ratio of the number of edges in layer $i$ to the number of edges in layer $\ell$ or more globally define $C^{i,\ell}$ to be equal to the proportion of all of the edges in the multiplex that occur in layer $\ell$. This model is the assymmetric influence matrix $W$ introduced in [@sola_eigenvector_2013] for measuring eigenvector centrality in multiplex networks (we recover the asymmetric version by assuming $c^{i,\ell}_a=c^{i,\ell}_b=c^{\ell,i}_a=c^{\ell,i}_b$ for all $1\leq a,b\leq n$). When the $D_i$ are the adjacency matrices for the layers, we recover their computation of the global heterogeneous eigenvector centrality. Thus, our model is a direct generalization which allows for arbitrary dynamics and more complex layer ranking behavior individualized to each node. In the next section, we use a still simpler version of this model where we further assume that $c^{i,\ell}_j \equiv c^i$ for $1 \le \ell \le k, 1 \le j \le n$. Our third and simplest version, the [*equidistribution*]{} model, occurs when $\mathfrak{D}$ is closed and the quantities at each copy of a node are distributed equally among other copies. We set $c^{i,\ell}_j=\frac1k$ for all $1\leq j\leq n$ and $1\leq i,\ell\leq k$ and name the resulting operator as $\mathfrak{E}$. This formulation is a natural model for applications where the dynamic flow is equally likely to move between layers, or when the quantities at each node represent the outcomes of a binary process. In the next section we will show that $\mathfrak{E}$ has a particularly simple spectral structure related to that of the aggregate network. Spectral Results ================ We present basic spectral results for our multiplex formulation and show that our operators arise as natural multiplex interpretations of standard network dynamic models, such as random walks and heat diffusion, derived from first principles. One particularly valuable aspect of $\mathfrak{D}$ is that under reasonable assumptions, it preserves matrix properties of the original layer dynamics, such as primitivity, stochasticness, and postive semi–definiteness. These properties provide the link between the spectrum of the operator and the multiplex structure. We begin with a simple result, relating the eigenvalues of the equidistribution operator $\mathfrak{E}$ to the spectrum of the aggregate network. Except for zero, the eigenvalues of $\mathfrak{E}$ are exactly the eigenvalues of $\frac1k\sum_{\ell=1}^k D_\ell$. Similar results exist for the other formulations, but they are more complex. In the simplest version of the hierarchical layer model, where $c^{i,\ell}_j\equiv c^i$, we have that the non–zero eigenvalues of $\mathfrak{D}$ are exactly the eigenvalues of $\sum_{i=1}^k c^i D_i$. Although determining the spectral structure of sums of matrices in terms of the spectra of the summands is difficult, when the matrices are symmetric, we can use standard results due to Weyl described in [@fulton_eigenvalues_2000; @knutson_honeycomb_1999] to obtain upper and lower bounds on the individual eigenvalues of the derived operators (as in Proposition 4). Stochastic Dynamics {#stochastic-dynamics .unnumbered} ------------------- A left stochastic matrix is a non–negative matrix where the entries in each column sum to one. Such a matrix defines a Markov process on a network, with the entries $D_{i,j}$ corresponding to the probability of moving from from node $j$ to node $i$ at each time step. Stochastic processes on networks are commonly used for determining centrality and community detection. If a stochastic operator is primitive or irreducible, the Perron–Fröbenius Theorem guarantees that the largest eigenvalue of that matrix is $1$ and that the entries in one of the corresponding eigenvectors are non–negative. This vector represents the limiting state of the dynamical system for arbitrary input. If each $D_i$ is stochastic and $\mathfrak{D}$ is closed then $\mathfrak{D}$ is stochastic. Additionally, if each $D_i$ is irreducible then $\mathfrak{D}$ is irreducible and if each $D_i$ is primitive then $\mathfrak{D}$ is primitive. The first statement follows from the fact that the hypotheses guarantee that $\mathfrak{D}$ is the product of two stochastic operators, and the second two conclusions can be easily shown from the graph theoretic interpretation of the conditions. We view a stochasitic $\mathfrak{D}$ as describing a random walk on the multiplex with the probability of transitioning from the copy of node $j$ on layer $\ell$ to a neighbor of node $j$ on any arbitrary layer $i$ given by $c^{i,\ell}_j\frac{1}{\operatorname{deg}(n^{i}_j)}$, where $n^i_j$ is the copy of node $i$ in layer $j$. This Markov formulation is equivalent to the problem considered in [@trpevski_discrete-time_2014] for modeling distributed concensus which provides additional confirmation of our methods. Under similar hypotheses to those in [@domenico_navigability_2014; @trpevski_discrete-time_2014], our operator reduces to a version of their supratransition matrix. Diffusion Dynamics {#diffusion-dynamics .unnumbered} ------------------ We often model diffusion on networks as a discretization of the continuous heat flow problem, which yields the graph Laplacian, $L=D-A$, whose spectral structure has strong connections to important graph properties such as connectivity, communities, and random walks [@brualdi_mutually_2011; @chung_spectral_1996; @newman_networks:_2010]. To extend this model to the multiplex setting, we consider an initial vector $v$ representing the temperature at each node copy and define the change in the value of $v^i_j$ with respect to time to be proportional to the sum of the differences in temperature between each copy of node $j$ (including $n^i_j$) and its neighbors: $$\frac{dv^i_j}{dt}=K\sum_{\ell=1}^kc_j^{i,\ell}\sum_{n^\ell_j\sim n^\ell_m}(v^\ell_j-v^\ell_m).$$ Here $K$ is the diffusion constant and the $c^{i,\ell}_j$ represent the proportion of the effect on layer $\ell$ that passes through to $n^{i}_j$. Linear algebraically, this is: $$\frac{dv^i_j}{dt}=K\left[\sum_{\ell=1}^kc_j^{i,\ell}L^\ell v^\ell \right]_j$$ which agrees with (1) when the layer dynamics are given by the respective layer Laplacians, $L^{i}$. This model implicitly assumes that each copy of each node can be assigned a separate temperature. This is a reasonable assumption for applications such as some economic exchange networks or the transportation models considered in [@domenico_navigability_2014]. However, in examples such as our social multiplex where the node copies are only representing different interaction types associated to the same individual, we enforce the additional condition that $v^i_j=v^\ell_j$ for all $1\leq j\leq n$ and $1\leq i,\ell\leq k$. Intuitively, at each time step, each individual has some fixed amount of information regardless of what types of interactions they are performing. The equidistribution operator $\mathfrak{E}$, encodes this additional condition. As we’ve reduced to $n$ independent variables, we restate Equation (2) in terms of a $n\times q$ vector $w$ recording the temperature at each node as $ \frac{dw_j}{dt}=\frac{K}{k}\left[\sum_{\ell=1}^kL^\ell w \right]_j. $ The ability to models these types of dynamics distinguishes our analysis from previous studies of diffusion on multiplex networks [@gomez_diffusion_2013]. Using tools from the theory of Hermitian matrices we can prove the following bounds for the operator $\mathfrak{E}$ when the intra–layer dynamics are Laplacians. If each $D_i$ is the graph Laplacian then $\mathfrak{E}$ is positive semi–definite and the eigenvectors of $\mathfrak{E}$ corresponding to distinct non–zero eigenvalues are orthogonal. Consequently, the solutions to the differential equation $\frac{d\hat{\varphi}}{dt}+\mathfrak{E}\hat{\varphi}=0$ have the same algebraic and analytic structure as the standard Laplacian: the solution consists of constant elements and terms that decay exponentially. We now find eigenvalue bounds for $\mathfrak{E}$ in this case. Our first result shows that diffusion across the multiplex network should happen at least as fast as diffusion across the best connected layer. We assume that the layer networks are connected - although similar bounds exist in the case where the layer networks are disconnected, the formulas become more complex. We also introduce some additional notation. Let $\lambda^i_j$ be the $j^{th}$ eigenvalue of $D_i$ written in descending order with $\lambda_f^i$ representing the Fiedler value corresponding to $D_i$. Let the eigenvalues of $\mathfrak{E}$ be $\lambda_1\geq\lambda_2\cdots\geq \lambda_{kn}$. As $\operatorname{rank}(\mathfrak{E})=n-1$, for $\ell>n-1$ we have $\lambda_\ell=0$. Finally, let $m$ be the index such that $\lambda^m_1=\operatorname{max}_i(\lambda_1^i)$. If each $D_i$ is the graph Laplacian then we have the following eigenvalue bounds for the operator $\mathfrak{E}$ - Fiedler Value: $ \operatorname{max}_i(\lambda_f^i)\leq k\lambda_f \leq \lambda_f^m+\sum_{j\neq \ell} \lambda_1^j$, - Leading Value: $\operatorname{max}_i(\lambda_1^i)\leq k\lambda_1\leq \sum_{i}\lambda_1^i$, These bounds are special cases of the following more general but less computationally feasible bounds: $$\operatorname{max}_i(\lambda_{n-\ell}^i)\leq k\lambda_{n-\ell}\leq \min_{J\vdash n+k-(\ell+1)}\left(\min_{\sigma\in S_n}\left(\sum_{i=1}^k \lambda_{j_i}^{\sigma(i)}\right)\right),$$ where $J=(j_1,j_2,\ldots,j_k)$ such that $\sum_{i=1}^kj_i=n+k-(\ell+1)$. As remarked after Proposition 1, we can give a similar characterization of the eigenvalues of the simplest case of the hierarchical layer model, where we obtain results equivalent to Propositions 3 and 4, replacing $\lambda^i_j$ with $(c^i\lambda^i)_j$ (after possible reordering) in each occurrence. Discussion and Conclusions ========================== The field of multiplex networks is rapidly becoming an important segment of the complex systems community. Although many natural structural representations for multiplex structures have been presented, there is still a great deal left to learn about multiplex structures and their associated dynamics. To faithfully represent dynamical processes on multiplex networks, we have shown it is sometimes necessary to avoid the confounding structural features introduced by tensorial and supra– constructions. To obtain meaningful results in the multiplex setting we must verify that the correct analogy with the monoplex network case holds. While there may be particular applications for which these structurally motivated dynamics are well suited, in general we recommend that researchers choose dynamical models that do not introduce structural features in order to more accurately capture the dynamics of interest. It is particularly important when relying on spectral computations and invariants that the associated operator have a natural interpretation that allows the appropriate Rayleigh quotient to be derived as a solution to the given objective function. Our model avoids the structural defects inherent in single network multiplex representations and allows the inherent layer dynamics to interact without explicitly confounding structural effects, generalizing several successful models to a wide class of network dynamics.

ArXiv

--- abstract: 'In the standard theory of delay equations, the fundamental solution does not ‘live’ in the state space. To eliminate this age-old anomaly, we enlarge the state space. As a consequence, we lose the strong continuity of the solution operators and this, in turn, has as a consequence that the Riemann integral no longer suffices for giving meaning to the variation-of-constants formula. To compensate, we develop the Stieltjes-Pettis integral in the setting of a norming dual pair of spaces. Part I provides general theory, Part II deals with “retarded” equations, and in Part III we show how the Stieltjes integral enables incorporation of unbounded perturbations corresponding to neutral delay equations.' author: - 'O. Diekmann and S.M. Verduyn Lunel' title: Twin semigroups and delay equations --- [*Dedicated, with considerable but finite delay, to John Mallet-Paret\ on the occasion of his sixtieth birthday*]{} Introduction ============ A delay equation is a rule for extending a function of time towards the future on the basis of the (assumed to be) known past. The shift along the extended function (i.e., the introduction of current-time-specific past) defines a dynamical system. Delay equations come in two kinds: delay differential equations (DDE) [@Diek95; @HVL93] and renewal equations (RE) [@DGMT98; @DGHKMT01; @DGG07]. From a PDE oriented semigroup perspective, delay equations are eccentric: one first constructively defines the semigroup and only then determines the generator, in order to relate to an abstract ODE. Subsequently the abstract ODE guides the development, along a well-established path, of the qualitative theory, comprising in particular local stability and bifurcation results. The variation-of-constants formula is a key instrument for developing the qualitative theory. It relates semigroups that correspond to (slightly) different rules for extension to each other. The formula involves an integral. The great advantage of strongly continuous semigroups is that one can work with the Riemann integral. But on many function spaces the shift operator is NOT strongly continuous. Also note that, ideally, the function space should allow us to choose the value in one point independently of the values in nearby points, in order to represent explicitly the rule for extension (which acts in one point). But in many function spaces the value in one point is NOT independent of the value in nearby points. Is it possible to avoid both Scylla and Charybdis? In [@Diek95] and [@DGG07] an affirmative answer is presented. One starts with the simplest rule for extension and a Banach space $X$ on which the semigroup is strongly continuous. The representation of the rule for extension is facilitated by embedding the ’small’ space $X$ into a ’big’ space $\BX$, obtained as the dual of the subspace $X^\odot$ of $X^\ast$ on which the adjoint operators are strongly continuous. Perturbations are bounded maps from $X$ into $\BX$ and the integral is now a weak-star Riemann integral taking values in $\BX$. Since one can show that the values belong to the image of $X$ under the embedding, they can be re-interpreted as elements of $X$. The framework of the four spaces $X$, $X^\ast$, $X^\odot$, $\BX$ is stable under perturbations at the generator level that are described by [**bounded**]{} maps from $X$ to $\BX$. Thus sun-star calculus yields a satisfactory theory for semilinear problems (see [@MR09; @MR18] for an alternative approach using integrated semigroups). [*Note on terminology:*]{} In the context of delay equations we call a space of functions of one real variable (time) “small” if translation along an (extended) element is continuous and “big” if it is not. So the spaces of continuous functions $C\gb{[-1,0],\BR^n}$ and integrable functions $L^1\gb{[-1,0],\BR^n}$ are small, while the spaces of bounded Borel measurable functions $B\gb{[-1,0],\BR^n}$ and bounded variation functions $NBV\gb{[-1,0],\BR^n}$ are big. The aim of the present paper is to establish the variation-of-constants formula for a semigroup of linear operators $\{S(t)\}$ on a big state space $Y$ that accommodates the fundamental solution. The motivation has four components: - We anticipate that such a formula should hold; indeed, an integrated version was verified in [@Diek95 Theorem III.2.16], so it seems merely a matter of making sense of the integral. - Strong continuity is a blessing, but the need to have it can be a curse; already in 1953 Feller emphasized that measurability and integrability of (in matrix inspired notation) $t \mapsto y^\ast S(t)y$, for $y$ belonging to $Y$ and for a sufficiently rich collection of $y^\ast$ in the dual space $Y^\ast$, might be a natural starting point for defining integrals [@Fel53]; more recently Kunze [@Kun11], building on Feller’s ideas, emphasized that it is natural to work with a norming dual pair of spaces, see the beginning of Section 2 below, such as $B(E)$, the space of all bounded measurable function on a measurable space $E$ and $M(E)$, the space of all bounded measures on $E$, in the theory of Markov processes; in delay equations the Markov process is trivial (just aging), but numbers change; can one incorporate the change of numbers via the variation-of-constants formula? - For Renewal Equations corresponding to population models, the space NBV of normalized functions of bounded variation is a very natural state space (see [@Fel76 Chapter XI]) with jumps capturing cohorts, cf. [@Huy97]. - This is a first step towards covering neutral delay equations in Part III. Neutral delay equations correspond to [**unbounded**]{} (yet relatively bounded) maps from $X$ to $\BX$ and as a consequence the spaces $X^\odot$ and $\BX$ depend on the particular perturbation; this undermines the strength (and beauty) of sun-star calculus. We shall heavily exploit that the extension can be defined in terms of the solution of a finite dimensional renewal equation, for which the powerful (Lebesgue) integration theory of real valued functions provides a wealth of results. In other words, we exploit that the rule for extension is represented by an operator with finite dimensional range (so abstract delay equations are not (yet) included). But the variation-of-constants formula itself involves an abstract integral. To define it, we fine-tune the Pettis integral developed by Kunze [@Kun11] in the context of a norming dual pair of spaces. In Sections 2–4 we introduce twin semigroups defined on a norming dual pair of spaces and we show how Retarded Functional Differential Equations (RFDE), with the space of bounded measurable functions as the state space, fit into this framework. In the second part, Sections 5–7, we deal with bounded finite rank perturbations of twin semigroups and show that the theory covers both RFDE and Renewal Equations (RE) with “smooth" kernels. In the third and final part we turn to relatively bounded (but still finite rank) perturbations. We use “cumulative output” [@DGT93] and the Stieltjes integral to extend our approach to cover Neutral Functional Differential Equations (NFDE) and RE with bounded variation kernels. **Part I: Twin semigroups** Twin semigroups on a norming dual pair ====================================== Conceptually, the linear space $Y$ is the state space for the dynamical systems that we want to study and the linear space $\DY$ is an auxiliary space that helps us to perform such studies. But this difference in role is more or less hidden in the linear situation considered in this paper (it will clearly manifest itself in follow-up work on nonlinear problems that we plan to do). A related remark is that our formulation employs the field $\BR$ of real numbers, even though conceptually there is no difference with vector spaces over the field $\BC$ of complex numbers (also see the beginning of Section 5). Two Banach spaces $Y$ and $\DY$ are called a [*norming dual pair*]{} (cf. [@Kun11]) if a bilinear map $$\pa{\novar}{\novar} : \DY \times Y \to \BR$$ exists such that, for some $M \in [1,\infty)$, $${{\vert \pa{\dy}{y} \vert}} \le M \|\dy\|\|y\|$$ and, moreover, $$\begin{aligned} \|y\| &= \sup \Bigl\{|\pa{\dy}{y}| \mid \dy \in \DY,\ \|\dy\| \le 1\,\Bigr\}\\ \|\dy\| &= \sup \Bigl\{|\pa{\dy}{y}| \mid y \in Y,\ \|y\| \le 1\,\Bigr\}.\end{aligned}$$ So we can consider $Y$ as a closed subspace of $Y^{\diamond\ast}$ and $\DY$ as a closed subspace of $Y^{\ast}$ and both subspaces are necessarily weak$^\ast$ dense. A [*twin operator*]{} $L$ on a norming dual pair $(Y,\DY)$ is a bounded linear map $$L : \DY \times Y \to \BR\qquad (\dy,y) \mapsto \dy Ly$$ such that - for given $y \in Y$ the map $\dy \mapsto \dy Ly$ is continuous with respect to the $\si(\DY,Y)$-topology and - for given $\dy \in \DY$ the map $y \mapsto \dy Ly$ is continuous with respect to the $\si(Y,\DY)$-topology Since $(\DY,\si(\DY,Y))' = Y$ the map in $(1)$ defines an element of $Y$ that we denote by $Ly$, and since $(Y,\si(Y,\DY))' = \DY$ the map in $(2)$ defines an element of $\DY$ that we denote by $\dy L$. Associated with a twin operator $L$ are two maps, one mapping $Y$ to $Y$ and one mapping $\DY$ to $\DY$, both also denoted by the symbol $L$, but to indicate on which space $L$ acts we write, inspired by [@Fel53], the point on which $L$ acts to the left of $L$ when this is a point of $\DY$. In other words, $Ly \in Y$ and $\dy L \in \DY$ are such that for all $y \in Y$ and $\dy \in \DY$ we have $$\pa{\dy}{Ly} = \dy Ly = \pa{\dy L}{y}.$$ Note that if our starting point is an operator $L : Y \to Y$ then there is an associated twin operator $L$ if and only if the adjoint of $L$ leaves the embedding of $\DY$ into $Y^\ast$ invariant. Similarly, if our starting point is an operator $L : \DY \to \DY$ then there is an associated twin operator $L$ if and only if the adjoint of $L$ leaves the embedding of $Y$ into $Y^{\diamond\ast}$ invariant. \[def:2.1\] A family $\{S(t)\}_{t \ge 0}$ of twin operators on a norming dual pair $(Y,\DY)$ is called a [*twin semigroup*]{} if - $S(0) = I$, and  $S(t+s) = S(t)S(s)\quad$ for $t,s \ge 0$; - there exist constants $M \ge 1$ and $\om \in \BR$ such that $${{\vert \dy S(t)y \vert}} \le Me^{\om t} \|y\|\|\,\dy\|;$$ - for all $y \in Y$, $\dy \in \DY$ the function $$t \mapsto \dy S(t)y$$ is measurable; - for $\Re \la > \om$ $with $\om$ as introduced in ii)$ there exists a twin operator $\Rbar(\la)$ such that $$\label{eq:2.2} \dy \Rbar(\la)y = \int_0^\infty e^{-\la t} \dy S(t)y\,dt.$$ Note that the combination of $ii)$ and $iii)$ allows us to conclude that the right hand side of ${\eqref{eq:2.2}}$ is well-defined, but not that it defines a bounded linear operator on $Y$ and $\DY$. We call $\Rbar(\la)$ defined on $\{\la \mid \Re \la > \om\}$ the [*Laplace transform*]{} of $\{S(t)\}$. It actually suffices to assume that the assertion of iv) holds for $\la = \la_0$ with $\Re \la_0 > \om$. This assumption allows us to introduce the multi-valued operator $$\label{eq:2.3a} C = \la_0 I - \Rbar(\la_0)^{-1}$$ on $Y$ and next define the function $\la \mapsto \Rbar(\la)$ by $$\label{eq:2.3} \Rbar(\la) = (\la I - C)^{-1}$$ on an open neighbourhood of $\la_0$. As Proposition A.2.3 of [@Haa06] shows, the function $R$ is holomorphic with Taylor series given by $$\Rbar(\la) = \sum_{k=0}^\infty (\mu - \la)^k \Rbar(\mu)^{k+1}$$ and the resolvent identity $$\Rbar(\la) - \Rbar(\mu) = (\mu - \la)\Rbar(\la)\Rbar(\mu)$$ holds. In Proposition 5.2 of [@Kun11] these facts are used to prove that $$\Om_0 := \bigl\{ \la \mid \Rbar(\la) \hbox{ is a twin operator and } {\eqref{eq:2.2}} \hbox{ holds } \bigr\}$$ contains the half plane $\{ \la \mid \Re \la > \om \}$. In Definition 2.6 of [@Kun09] Kunze calls $C$ the [*generator*]{} of the semigroup provided the Laplace transform is injective and hence $C$ is single-valued. Here we adopt a more pliant position and call $C$ the generator even when it is multi-valued. Note that we might equally well call the operator $\DC$, defined on $\DY$ as the inverse of the Laplace transform, but now considered as an operator mapping $\DY$ into $\DY$, the generator. As long as one realises that the two have the same twin operator as their resolvent, this cannot lead to confusion. By combining [@Kun09 Prop. 2.7] and [@Kun11 Thm. 5.4] one obtains that the twin semigroup is uniquely determined by the generator if both $C$ and $\DC$ are single-valued. We now list some basic results from [@Kun11]. For completeness we provide proofs, even though these are, in essence, copied from [@Kun11]. \[lem:2.2\] The following statements are equivalent - $y \in \DOM{C}$ and $z \in Cy$; - there exist $\la \in \BC$ with $\Re \la > \om$ and $\om$ as introduced in ii) and $y,z \in Y$ such that $$\label{eq:2.4} y = \Rbar(\la)(\la y - z)$$ - $y,z \in Y$ and for all $t > 0$ $$\label{eq:2.5} \int_0^t S(\tau)z\,d\tau = S(t)y - y.$$ Here it should be noted that 3. includes the assertions - the integral $\int_0^t S(\tau)z\,d\tau$ defines an element of $Y$ (even though at first it only defines an element of $Y^{\diamond\ast}$); - the integral $\int_0^t S(\tau)z\,d\tau$ does not depend on the choice of $z \in Cy$ in case $C$ is multi-valued. The observation $y \in \DOM{C}$ if and only if $y = \Rbar(\la)\tilde y$ and in that case $(\la I - C)y = \tilde y$, establishes the equivalence of the items 1. and 2. The integrals below derive their meaning by pairing the integrand with arbitrary $\dy \in \DY$. But in order to enhance readability, we do not actually write these pairings. Let $\Re \la > \om$. The identity $$\label{eq:2.6} \int_0^t e^{-\la \tau}S(\tau)y\,d\tau = \Rbar(\la)\gb{y - e^{-\la t}S(t)y}$$ follows straightforwardly by considering $\int_0^t = \int_0^\infty - \int_t^\infty$ and next shifting the integration variable in the second integral over $t$. If we multiply by $\la$, assume that 2. holds, and use to rewrite $\la R(\la)y$, we obtain $$\la \int_0^t e^{-\la \tau}S(\tau)y\,d\tau = y + \Rbar(\la)\gb{z - \la e^{-\la t}S(t)y}.$$ Next use with $y$ replaced by $z$, as well as the fact that $S(t)$ and $\Rbar(\la)$ commute, to arrive at $$\la \int_0^t e^{-\la \tau}S(\tau)y\,d\tau = y + \int_0^t e^{-\la \tau} S(\tau)z\,d\tau - e^{-\la t}S(t)\Rbar(\la)(\la y - z)$$ or, on account of $$\label{eq:2.7} \la \int_0^t e^{-\la \tau}S(\tau)y\,d\tau = y + \int_0^t e^{-\la \tau} S(\tau)z\,d\tau - e^{-\la t}S(t)y.$$ The identity does not involve any improper integral, so we can extend by analytic continuation and, in particular, take $\la = 0$. This yields . Thus we have proved that 2. implies 3. Finally, assume that 3. holds. Then $$\begin{aligned} \la \Rbar(\la)y - y &= \int_0^\infty \la e^{-\la \tau}\gb{S(\tau)y - y}\,d\tau\\ &= \int_0^\infty \la e^{-\la \tau}\int_0^\tau S(\si)z\,d\si\,d\tau\\ &= \int_0^\infty \int_\si^\infty \la e^{-\la \tau}\,d\tau\,S(\si)z\,d\si\\ &= \int_0^\infty e^{-\la \si}S(\si)z\,d\si = \Rbar(\la)z\end{aligned}$$ which amounts to \[lem:2.3\] For all $t > 0$ and $y \in Y$, we have $\int_0^t S(\tau)y\,d\tau \in \DOM{C}$ and $$\label{eq:2.8} S(t)y - y \in C\int_0^t S(\tau)y\,d\tau.$$ Again we omit the pairing with $\dy$. Yet, we keep in mind that the integrals define elements in $Y^{\diamond\ast}$ for which we subsequently check that they are represented by elements in $Y$. Since $y \in (\la I - C)\Rbar(\la)y$ we have $$\int_0^t S(\tau)y\,d\tau \in \int_0^t S(\tau)(\la I - C)\Rbar(\la)y\,d\tau$$ and $$\begin{aligned} \int_0^t S(\tau)(\la I - C)\Rbar(\la)y\,d\tau &= \la \int_0^t S(\tau)\Rbar(\la)y\,d\tau - \int_0^t S(\tau)C\Rbar(\la)y\,d\tau\\ &= \la \int_0^t S(\tau)\Rbar(\la)y\,d\tau - S(t)\Rbar(\la)y + \Rbar(\la)y,\end{aligned}$$ where we have used . Note that the right hand side is single valued. We claim that the right hand side belongs to $Y$. This is clear for the last two terms. Concerning the first, observe that implies that $t \mapsto S(t)y$ is continuous if $y \in \DOM{C}$. Hence we can interpret the integral $\int_0^t S(\tau)\Rbar(\la)y\,d\tau$ as a Bochner integral of a continuous $Y$-valued function. Since $S(\tau)\Rbar(\la) = \Rbar(\la)S(\tau)$ and $\Rbar(\la)$ is a twin operator we have $$\int_0^t S(\tau)\Rbar(\la)y\,d\tau = \Rbar(\la)\int_0^t S(\tau)y\,d\tau.$$ So the identity above can be written in the form $$\int_0^t S(\tau)y\,d\tau = \Rbar(\la)\gb{\la\int_0^t S(\tau)y\,d\tau + y - S(t)y}.$$ Comparing this to we conclude that $\int_0^t S(\tau)y\,d\tau \in \DOM{C}$ and that holds. In the proof of Lemma \[lem:2.3\] we used the assumption that $\Rbar(\la)$ is a twin operator (cf. Definition \[def:2.1\], iv) to prove that the same is true for local integrals of the orbit $t \mapsto S(t)y$ for arbitrary $y \in Y$. In Theorem 5.8 of [@Kun11] Kunze proves that these two properties are equivalent. In order to obtain information about the asymptotic behaviour of the twin semigroup $S(t)$, we adapt a result for strongly continuous semigroups from [@AreBat88]. It was observed by Batty in [@Batty94] that in case $\si(C) \mcap i\BR = \emptyset$, the asymptotic behaviour actually follows from Korevaar’s proof of the Ingham theorem [@Korevaar]. Here we adapt this argument from [@Batty94] to the case of twin semigroups. \[thm:Tauber\] Let $S(t)$ be a twin semigroup on a norming dual pair $(Y,\DY)$ and assume that $S(t)$ is bounded. If $\si(C) \mcap i\BR = \emptyset$, then $$\label{eq:Tauber} \|S(t)C^{-1}\| \to 0 \qquad \hbox{as}\quad t \to \infty.$$ As a consequence we have that $S(t)y \to 0$ as $t \to \infty$ for every $y$ in the norm-closure of $\DOM{C}$. Let $\Ga_R := \{ z \in \BC \mid |z| = R \}$ and $\Ga_R^{-}$ and $\Ga_R^+$ denote the part of $\Ga_R$ in the, respectively, left and right closed half plane of $\BC$. Define $\Ga_0$ to be a path in the intersection of $\rho(C)$ and the open left half plane connecting $iR$ and $-iR$ such that the closed contour $\Ga$ given by the union of $\Ga_R^+$ and $\Ga_0$ does not encircle any pole of $(zI - C)^{-1}$. From Cauchy’s Residue Theorem it follows that we can write $$\label{eq:cauchy-id} \dy S(t)C^{-1}y = -\frac{1}{2\pi i} \int_{\Ga} \gb{1 + \frac{z^2}{R^2}}\dy (zI - C)^{-1}S(t)y\,\frac{dz}{z},$$ where the factor $\gb{1 + \frac{z^2}{R^2}}$ is chosen because for $z \in \Ga_R$ the identity $$\label{eq:fudge} \bigl| 1 + \frac{z^2}{R^2} \bigr| = \frac{2\bigl| \Re z\bigr|}{R}$$ holds. Fix $t \ge 0$ and observe that from the identity we have for $\Re z \ge 0$ $$\label{eq:2.6new} e^{zt}\int_0^t e^{-z\tau} \dy S(\tau)y\,d\tau = \dy (zI - C)^{-1}\gb{e^{zt}y - S(t)y}.$$ Define the entire function $g_t : \BC \to Y$ by $$g_t(z) := \int_0^t e^{-z\tau} S(\tau)y\,d\tau$$ and use to deduce the identity $$\begin{aligned} \label{eq:cauchy-id2} &\frac{1}{2\pi i} \int_{\Ga_0}\gb{1 + \frac{z^2}{R^2}}\dy (zI - C)^{-1}S(t)y\,\frac{dz}{z}\nonumber\\ &\qquad\qquad\qquad= \frac{1}{2\pi i} \int_{\Ga_0}\gb{1 + \frac{z^2}{R^2}}e^{zt}\dy (zI - C)^{-1}y\,\frac{dz}{z}\nonumber\\ &\qquad\qquad\qquad\qquad\qquad\qquad -\frac{1}{2\pi i} \int_{\Ga_R^{-}}\gb{1 + \frac{z^2}{R^2}} e^{zt} \dy g_t(z)\,\frac{dz}{z}.\end{aligned}$$ Since along $\Ga_0$ we have $\Re z < 0$, it follows from the dominated convergence theorem that the first integral on the right hand side of tends to zero as $t \to \infty$. Using the fact that $|\dy S(t)y| \le M\|\dy\|\,\|y\|$, we have for $z \in \Ga_R^-$, $$\bigl|e^{zt}\dy g_t(z)\bigr| = \bigl| \int_0^t e^{z(t-\tau)} \dy S(\tau)y\,d\tau \bigr| \le \frac{M}{\bigl|\Re z\bigr|}\|\dy\|\,\|y\|.$$ Similarly, for $z \in \Ga_R^+$ $$\bigl| \dy (z I - C)^{-1}S(t)y \bigr| = \bigl| \int_0^\infty e^{-z\tau} \dy S(t+\tau)y\,d\tau \bigr| \le \frac{M}{\Re z}\|\dy\|\,\|y\|,$$ From the property it follows that both the integral over $\Ga_R^-$ in and the integral over $\Ga_R^+$ in are bounded. Using these estimates in combination with the identities and yields $$\label{eq:cauchy-id3} \limsup_{t \to \infty}\ \bigl|\dy S(t)C^{-1}y \bigr| \le \frac{2M}{R}\|\dy\|\,\|y\|.$$ By letting $R \to \infty$, we conclude . Since $C^{-1}$ has dense range in the norm-closure of $\DOM{C}$, the final observation follows from and the fact that $S(t)$ is bounded. In this paper we will see that our perturbation results are well suited to verify the conditions of Theorem \[thm:Tauber\] in terms of the given data. The subspace of strong continuity ================================= We define the subspace $X$ of $Y$ by $$\label{eq:3.1} X := \bigl\{y \in Y \mid t \mapsto S(t)y\hbox{ is continuous}\, \bigr\}$$ and note, first of all, that the semigroup property of $\{S(t)\}_{t \ge 0}$ yields as an equivalent characterization $$\label{eq:3.2} X := \bigl\{y \in Y \mid \lim_{t \da 0} \| S(t)y - y \| = 0 \bigr\}.$$ As $S(t)$ maps $X$ into $X$, the restriction $$\label{eq:3.3} T(t) = S(t)\big\vert_{X}$$ defines a strongly continuous semigroup $\{T(t)\}_{t \ge 0}$ on the Banach space $X$ ($X$ is norm-closed in $Y$, see Theorem \[thm:3.1\]). The main results of this section are the following theorems. \[thm:3.1\] The subspace $X$ of strong continuity equals the norm closure of $\DOM{C}$ $$X = \clo{\DOM{C}}.$$ \[thm:3.2\] The generator $A$ of the strongly continuous semigroup $\{T(t)\}_{t \ge 0}$ on $X$ is the part of $C$ in $X$. It should be noted here that, as we shall prove below, the generator $A$ is single-valued even if $C$ is a multi-valued map. In order to prove Theorem \[thm:3.1\] and \[thm:3.2\] we first provide an auxiliary result that is of independent interest, cf. [@Cran73]. \[lem:3.3\] If $y \in \DOM{C}$ then $$\limsup_{h \da 0} \frac{1}{h}\|S(h)y - y\| < \infty.$$ By Lemma \[lem:2.2\] we have for $z \in Cy$ the identity $$\frac{1}{h}\gb{\dy S(h)y - \pa{\dy}{y}} = \frac{1}{h}\int_0^h \dy S(\tau)z\,d\tau$$ and consequently $$\begin{aligned} \frac{1}{h}\bigl|\dy S(h)y - \pa{\dy}{y}\bigr| &\le \frac{1}{h}\int_0^h Me^{\om\tau}\|\dy\|\|z\|\,d\tau\\ &= M\frac{e^{\om h}-1}{\om h}\|\dy\|\|z\|.\end{aligned}$$ It follows that $$\frac{1}{h}\|S(h)y - y\| \le M\frac{e^{\om h}-1}{\om h}\|z\|$$ and so $$\limsup_{h \da 0} \frac{1}{h}\|S(h)y - y\| \le M\|z\|.$$ \[col:3.4\] The domain of the generator $C$ of the semigroup $\{S(t)\}_{t \ge 0}$ satisfies $$\DOM{C} \subset X.$$ \[lem:3.5\] For $X$ defined by ${\eqref{eq:3.1}}$ we have $$X \subset \clo{\DOM{C}}.$$ For arbitrary $y \in Y$ $$\begin{aligned} \bigl\|\frac{1}{t}\int_0^t S(\tau)y\,d\tau - y \bigr\| &= \sup_{\|\dy\|\le 1}\bigl|\frac{1}{t}\int_0^t \gb{\dy S(\tau)y - \pa{\dy}{y}}\,d\tau\bigr|\\ &\le \frac{1}{t}\int_0^t \| S(\tau)y - y\|\,d\tau.\end{aligned}$$ If $y \in X$, then the integrand at the right hand side is a continuous function of $\tau$ vanishing at $\tau = 0$. It follows that in that case the right hand side converges to zero for $t \da 0$. Since $$\int_0^t S(\tau)y\,d\tau \in \DOM{C},$$ cf. Lemma \[lem:2.3\], we conclude that in any $\ep$-neighbourhood of $y \in X$, there is an element of $\DOM{C}$. By combining Corollary \[col:3.4\] and Lemma \[lem:3.5\] we obtain a proof of Theorem \[thm:3.1\]. Note that the semigroups $\{S(t)\}_{t \ge 0}$ and $\{T(t)\}_{t \ge 0}$ are [*intertwined*]{} in the sense that $$\label{eq:3.4} S(t)y \in (\la I - C)T(t)\Rbar(\la)y.$$ [*Proof of Theorem $3.2$.*]{} If $y \in \DOM{C}$ and $z \in Cy \cap X$ then, by Lemma \[lem:2.2\], $$T(t)y - y = \int_0^t T(\tau)z\,d\tau$$ and it follows that $t^{-1}(T(t)y - y) \to z$ for $t \da 0$. In particular this shows that $Cy \cap X$ is, when non-empty, a singleton. Moreover, $Ay \in Cy$. Now assume that $y \in \DOM{A}$ and $Ay = z \in X$. Then $$T(t)y - y = \int_0^t T(\tau)z\,d\tau$$ and we conclude from Lemma \[lem:2.2\] that $y \in \DOM{C}$ and $z \in Cy$. Note on notation: the analogue of $X$ at the $\diamond$ side we shall denote by $X^{\odot}$. So in this paper $$\label{eq:3.5} X^{\odot} := \bigl\{ \dy \in \DY \mid \lim_{t \da 0} \|\DS(t)\dy - \dy\| = 0\, \bigr\}.$$ RFDE - Retarded Functional Differential\ Equations ======================================== We adopt the standard notation $x_t(\th) = x(t+\th)$ and the only slightly less standard notation $$\pa{\ze}{\ph} := \int_{[0,1]} d\ze(\si)\ph(-\si)$$ for $\ze \in NBV\gb{[0,1],\BR^{n \times n}}$ and $\ph \in B\gb{[-1,0],\BR^n}$. An equation of the form $$\label{eq:6.1} \dot x(t) = \pa{\ze}{x_t} = \int_{[0,1]} d\ze(\si)\,x(t-\si)$$ is called a RFDE. If we pose an initial value problem, we require to hold for $t \ge 0$ and supplement the equation by the initial condition $$\label{eq:6.2} x(\th) = \ph(\th),\qquad -1 \le \th \le 0,$$ for a known/given function $\ph$. The standard theory assumes that $\ph \in X$ with $X = C\gb{[-1,0],\BR^n}$, but here we allow $$\label{eq:6.3} \ph \in Y = B\gb{[-1,0],\BR^n}.$$ Concerning the given kernel $\ze$ we assume that for $i = 1,\ldots,n$ $$\label{eq:6.4} \ze_i \in \DY = NBV\gb{[0,1],\BR^n},$$ where $\ze_i$ is the $i$-th row of the matrix $\ze$. In Appendix B, it is shown that $Y$ and $\DY$ given by and form a norming dual pair. Once we solve –, we can define a $Y$-valued function $u : [0,\infty) \to Y$ by $$\label{eq:6.5} u(t)(\th) = x(t+\th; \ph),\qquad -1 \le \th \le 0,\ t \ge 0$$ and bounded linear operators $S(t) : Y \to Y$ by $$\label{eq:6.5a} S(t)\ph = u(t;\ph) = x(t + \novar;\ph).$$ The initial condition translates into $$\label{eq:6.6} S(0)\ph = u(0;\ph) = \ph$$ and reflects that we define a dynamical system on $Y$ by translating along the function $\ph$ extended according to . Below we show that $\{S(t)\}$ is a twin semigroup and we characterize its generator $C$. But first we present some heuristics. In order to motivate an abstract ODE for the $Y$-valued function $u$, we first observe that the infinitesimal formulation of the translation rule amounts to the PDE $$\label{eq:6.7} \frac{\partial u}{\partial t} - \frac{\partial u}{\partial \th} = 0$$ We need to combine this with , in terms of $u(t)(0) = x(t)$, and we need to specify the domain of definition of the derivative with respect to $\th$. The latter is actually rather subtle. An absolutely continuous function has almost everywhere a derivative and when the function is Lipschitz continuous this derivative is bounded. Thus a Lipschitz function specifies a unique $L^\infty$-equivalence class by the process of differentiation. But [*not*]{} a unique element of $Y$. In fact the set $$\label{eq:6.8} C\psi = \bigl\{ \psi' \in Y \mid \psi(\th) = \psi(-1) + \int_{-1}^\th \psi'(\si)\,d\si,\ \psi'(0) = \pa{\ze}{\psi} \bigr\}$$ is, for a given Lipschitz continuous function $\psi$, very large indeed. Nota bene that the condition $\psi'(0) = \pa{\ze}{\psi}$ takes care of and that, in the context of the space $Y$, we can simply take this as the definition of $\psi'(0)$ without having to worry about an influence of this choice on $\psi'(\si)$ for $\si$ near zero (such in sharp contrast to the space $X$ of continuous functions). Anyhow, we define $C$ as a multi-valued, unbounded, operator on $Y$ by $$\label{eq:6.9} \DOM{C} = Lip\gb{[-1,0],\BR^n},\qquad C\psi \hbox{ given by } {\eqref{eq:6.8}}.$$ We claim that - and correspond to $$\label{eq:6.10} \frac{du}{dt} \in Cu.$$ To substantiate this claim, we shall first derive (following essentially Section I.2 of [@Diek95]) a representation of the solution of – in terms of $\ph$, $\ze$ and the resolvent $\rho$ of $\ze$, next verify that $\{S(t)\}_{t \ge 0}$ defined by is a twin semigroup and, finally, that $C$ is the corresponding generator in the sense of –. \[lem:6.1\] The solution of – is given explicitly by $$\begin{aligned} \label{eq:6.12} x(t;\ph) &= \bigl(1 + \int_0^t \rho(\si)\,d\si\bigr)\ph(0) + \int_0^1 \Bigl\{ \ze(t+\si) - \ze(\si) +\nonumber\\ &\qquad\qquad\qquad\qquad\int_0^t \rho(\tau)\gb{\ze(t-\tau+\si) - \ze(\si)}\,d\tau\Bigr\}\ph(-\si)\,d\si,\end{aligned}$$ where the [*resolvent*]{} $\rho$ of the kernel $\ze$ is the unique solution of $$\label{eq:6.13} \rho \ast \ze + \ze = \rho = \ze \ast \rho + \ze$$ and hence given by $$\label{eq:6.14} \rho = \sum_{l=1}^\infty \ze^{l\ast}.$$ (See Section I.2 of [@Diek95] for more detail). We integrate from $0$ to $t$ and interchange the order of the two integrals at the right hand side. This yields $$\label{eq:6.15} x = \ze \ast x + f$$ with $$\begin{aligned} \label{eq:6.16} f(t) &= \ph(0) + \int_0^t\gb{\int_s^1 d\ze(\th)\,\ph(s-\th)}\,ds\nonumber\\ &= \ph(0) + \int_0^t \gb{\ze(t+\si)-\ze(\si)}\ph(-\si)\,d\si.\end{aligned}$$ The solution of is given by $$\label{eq:6.17} x = f + \rho * f$$ which leads, after another change of integration order, to . Please observe that $x$ depends on the value of $\ph$ in $\th = 0$ and the $L^\infty$-equivalence class to which $\ph$ belongs, but not on the precise point values of $\ph$ in points $\th<0$. \[col:6.2\] The definition amounts to $$\label{eq:6.18} \gb{S(t)\ph}(\th) = \int_0^1 K_t(\th,d\si)\,\ph(-\si)$$ with for $\si > 0$ $$\begin{aligned} \label{eq:6.19} K_t(\th,\si) &= H(\si+t+\th) + H(t+\th)\Bigl\{\int_0^{t+\th} \rho(\tau)\,d\tau + \int_0^\si \Bigl[\ze(t+\th+\tau)\nonumber\\ &\qquad\qquad - \ze(\tau) + \int_0^{t+\th} \rho(\xi)\gb{\ze(t+\th+\tau-\xi) - \ze(\tau)}\,d\xi\Bigr]\,d\tau\Bigr\}\end{aligned}$$ and $K_t(\th,0) = 0$. $Here $H$ is the standard Heaviside function.$ For $t+\th < 0$ the second term in the expression for $K$ does not contribute and the first term yields $$\gb{S(t)\ph}(\th) = \ph(t+\th)$$ which is in accordance with because of . Now assume that $t+\th \ge 0$. Clearly the first term contributes a unit jump at $\si = 0$ and $H(t+\th) = 1$. The second factor has, as a function of $\si$, a jump of magnitude $\int_0^{t+\th} \rho(\tau)\,d\tau$ at $\si = 0$, but is otherwise absolutely continuous with derivative $$\ze(t+\th+\si) - \ze(\si) + \int_0^{t+\th} \rho(\xi)\gb{\ze(t+\th+\si-\xi) - \ze(\si)}\,d\xi.$$ The jumps yield the first term at the right hand side of evaluated at $t+\th$ and the absolutely continuous part yields the second term. \[col:6.3\] The operator $S(t)$ belongs to a twin operator pair. This is a general property of kernel operators. Explicitly we have $$\label{eq:6.20} \gb{\DS(t)\dy}(\si) = \int_0^1 \dy(d\tau)\,K_t(-\tau,\si).$$ \[thm:6.4\] The semigroup $\{S(t)\}_{t \ge 0}$ defined by is a twin semigroup. With reference to Definition \[def:2.1\] we note that $S(0)=I$ follows directly from , while the semigroup property follows from the uniqueness of solutions to – and the fact that $S(t)$ corresponds to translation along the solution (so essentially it follows from the corresponding property for translation, and uniqueness of extension). The exponential estimates ii) are well-established in the theory of RFDE, for instance Sections I.5, IV.2 and IV.3 of [@Diek95] or the proof of Theorem \[thm:5.1\]. Property iii), the measurability of $t \mapsto \dy S(t)y$, is a direct consequence of the way $K_t(\th,\si)$ defined in depends on $t$. It remains to verify that the Laplace transform defines a twin operator pair. By Fubini’s Theorem, the Laplace transform is a kernel operator with kernel $$\int_0^\infty e^{-\la t} K_t(\th,\si)\,dt.$$ \[thm:6.5\] The operator $C$ defined by – is the generator $in the sense of {\eqref{eq:2.3}}$ of $\{S(t)\}_{t \ge 0}$ defined by . Assume $\ph \in (\la I - C)\psi$. Then there exists $\psi' \in Y$ which is a.e. derivative of $\psi$ such that $$\begin{aligned} \la\psi - \psi' &= \ph,\qquad -1 \le \th < 0\\ \la\psi(0) - \pa{\ze}{\psi} &= \ph(0).\end{aligned}$$ Solving the differential equation yields that $$\label{eq:6.21} \psi(\th) = e^{\la\th}\bigl\{\int_\th^0 e^{-\la\si}\ph(\si)\,d\si + \psi(0)\bigr\}$$ and accordingly the boundary condition for $\th = 0$ boils down to $$\label{eq:6.22} \psi(0) = \De(\la)^{-1}\bigl[\ph(0) + \int_0^1 d\ze(\si)e^{-\la\si}\int_{-\si}^0 e^{-\la\tau}\ph(\tau)\,d\tau\bigr]$$ which requires that $\det \De(\la) \not= 0$ with $$\De(\la) = \la I - \int_0^1 d\ze(\si)e^{-\la\si}.$$ Our claim is that the identity $$\psi(\th) = \int_0^\infty e^{-\la t}\gb{S(t)\ph}(\th)\,dt$$ holds. To verify this, we first note that $$\int_0^\infty e^{-\la t}\gb{S(t)\ph}(\th)\,dt = e^{\la\th}\bigl\{\int_\th^0 e^{-\la\si}\ph(\si)\,d\si + \bar x(\la;\ph)\bigr\}$$ (where $\bar x(\la;\ph) := \int_0^\infty e^{-\la t}x(t;\ph)\,dt$, with $x(t;\ph)$ the solution of – given by ) since $$\begin{aligned} \int_0^\infty e^{-\la t}x(t+\th;\ph)\,dt &= \int_0^{-\th} e^{-\la t}\ph(t+\th)\,dt + \int_{-\th}^\infty e^{-\la t}x(t+\th)\,dt\\ &= e^{\la\th}\bigl\{\int_\th^0 e^{-\la\si}\ph(\si)\,d\si + \bar x(\la;\ph)\bigr\}.\end{aligned}$$ So, since holds, we need to check that $\psi(0) = \bar x(\la;\ph)$. From we deduce that $$\bar x = (1 - \bar \ze)^{-1}\bar f.$$ Therefore, using the first representation of $f$ in , it follows that $$\begin{aligned} \la \bar f(\la) &= \ph(0) + \int_0^\infty \la e^{-\la t}\int_0^t\gb{\int_s^1 d\ze(\th)\ph(s-\th)}\,ds dt\\ &= \ph(0) + \int_0^\infty e^{-\la t}\int_t^1 d\ze(\th)\ph(t-\th)\, dt\\ &= \ph(0) + \int_0^1 d\ze(\th)\int_0^\th e^{-\la t}\ph(t-\th)\,dt\\ &= \ph(0) + \int_0^1 d\ze(\th)e^{-\la\th}\int_{-\th}^0 e^{-\la\si}\ph(\si)\,d\si\end{aligned}$$ which equals the vector at the right hand side of on which the matrix $\De(\la)^{-1}$ acts. Since $$\la \bar\ze(\la) = \int_0^1 d\ze(\th)e^{-\la\th},$$ we arrive at the conclusion that indeed $\psi(0) = \bar x(\la;\ph)$. It is a direct consequence of that $$\label{eq:6.23} X = \clo{\DOM{C}} = C\gb{[-1,0],\BR^n}.$$ Clearly $C\psi \cap X$ is either empty or a singleton, cf. , and for the set to be nonempty we need that $\psi \in C^1$ and $\psi'(0) = \pa{\ze}{\psi}$. So the generator $A$ of the restriction $\{T(t)\}_{t \ge 0}$ of $\{S(t)\}_{t \ge 0}$ to $X$ is given by $$\begin{aligned} \label{eq:6.24} \begin{split} \DOM{A} &= \bigl\{ \psi \in C^1 \mid \psi'(0) = \pa{\ze}{\psi} \bigr\}\\ A\psi &= \psi' \end{split}\end{aligned}$$ in complete agreement with the standard theory. As $S(t)$ maps $Y$ into $X$ for $t \ge 1$, one might wonder whether we gained anything at all by the extension from $X$ to $Y$? Already in the pioneering first version of his book [@Hale71], Jack Hale emphasized that if one adds a forcing term to , one needs $$\label{eq:6.25} q(\th) = \begin{cases} 1 &\th = 0\\ 0 &-1 \le \th < 0 \end{cases}$$ to describe the solution by way of the variation-of-constants formula. Indeed, the solution of $$\begin{aligned} \label{eq:6.26} \begin{split} \dot x(t) &= \pa{\ze}{x_t} + f(t),\qquad t \ge 0\\ x(\th) &= \ph(\th),\qquad -1 \le \th \le 0 \end{split}\end{aligned}$$ is explicitly given by $$\label{eq:6.27} x_t = S(t)\ph + \int_0^t S(t-\tau)q f(\tau)\,d\tau$$ since corresponds to the initial value problem $$\label{eq:6.28} \frac{du}{dt} \in Cu + qf,\qquad u(0) = \ph,$$ where $u(t) = x_t$. The integration theory of Section 5 provides a precise underpinning of the integral in . In the original approach of Hale, the hidden argument $\th$ in is inserted and thus the integral reduces to the integration of an $\BR^n$-valued function. Note that evaluation in a point corresponds to the application of a Dirac functional, so our approach yields, in a sense, a rather late theoretical underpinning of Hale’s approach. The $\odot\ast$-calculus approach of [@Diek95] amounts, for RFDE, to the observation just before Corollary \[eq:6.2\] and its consequences. More precisely, one embeds $X$ into $\BR^n \times L^\infty\gb{[-1,0],\BR^n}$, interprets $q$ as $(1,0)$, considers $\BR^n \times L^\infty\gb{[-1,0],\BR^n}$ as the dual space of $\BR^n \times L^1\gb{[0,1],\BR^n}$, interprets the integral as a weak$^\ast$-integral and checks that the integral belongs to the range of the embedding, so defines an element of $X$. As long as one restricts attention to RFDE, the current approach has its more sophisticated integration theory as a drawback and no clear advantage to compensate. However, this changes when one extends the theory, as we shall do in Section 11, to neutral equations. Neutral equations correspond to an unbounded change in the rule for extension, even within a functional analytic framework where $q$ is well-defined. In the $\odot\ast$-setting this manifests itself in dependence of $X^\odot$, and hence $X^{\odot\ast}$, on the particular perturbation obstructing a satisfactory sun-star perturbation theory for neutral equations. In contrast, the present approach allows us to keep working with the norming dual pair $Y$ and $\DY$ and to develop a variation-of-constants formula. At the end of Section 6 we shall briefly indicate how, alternatively, one can use a perturbation approach to derive the results presented above. As a final remark, we emphasize that the variation-of-constants formula is the key first step towards a local stability and bifurcation theory for nonlinear problems, as shown in detail in [@Diek95]. **Part II: Bounded perturbations** **describing retarded equations** The variation-of-constants formula for forcing functions with finite dimensional range ====================================================================================== When the ultimate aim is to study nonlinear problems, one usually focuses on real-valued functions and functionals. Spectral theory, on the other hand, benefits from complexification. The formulation below considers real functionals acting on a real vector space, but when $Y$, $\DY$ is a norming dual pair, the same holds for their complexifications.[^1] Motivated by RFDE, in particular , we want to define an element $u(t)$ of $Y$ by way of the action on $\DY$ expressed in the formula $$\label{eq:4.2} \pa{\dy}{u(t)} = \dy S(t)u_0 + \int_0^t \dy S(t-\tau)q\,f(\tau)d\tau,$$ where - $(Y,\DY)$ is a norming dual pair; - $q \in Y$; - $f : [0,T] \to \BR$ is bounded and measurable; - $\bigl\{S(t)\bigr\}$ is a twin semigroup, and where $u_0$ (corresponding to $\ph$ in ) is an arbitrary element of $Y$. The first term at the right hand side of is no problem at all, it contributes $S(t)u_0$ to $u(t)$. The second term defines an element of $Y^{\diamond\ast}$, but it is not clear that this element is, without additional assumptions, represented by an element of $Y$. \[lem:4.1\] In addition to (i)-(iv) assume that $$\label{eq:4.1} \gb{Y,\si(Y,\DY)}\quad\hbox{is sequentially complete.}$$ Then $$\label{eq:4.3} \dy \mapsto \int_0^t \dy S(t-\tau)q\,f(\tau)d\tau$$ is represented by an element of $Y$, to be denoted as $$\label{eq:4.4} \int_0^t S(t-\tau)q\,f(\tau)\,d\tau$$ There exists a sequence of step functions $f_m$ such that $|f_m| \le |f|$ and $f_m \to f$ pointwise. Lemma \[lem:2.3\] shows that $$\int_0^t S(t-\tau)q\,f_m(\tau)d\tau$$ belongs to $Y$ (in fact even to $\DOM{C}$). Since (see Definition \[def:2.1\](ii)) $$\babs{\dy S(t-\tau)q f_m(\tau)} \le M e^{\om(t-\tau)}\|q\|\,\|\dy\|\,\sup_{\si} |f(\si)|,$$ the dominated convergence theorem implies that for every $\dy \in \DY$ $$\lim_{m \to \infty} \int_0^t \dy S(t-\tau)q\,f_m(\tau)d\tau = \int_0^t \dy S(t-\tau)q\,f(\tau)d\tau.$$ The sequential completeness next guarantees that the limit too is represented by an element of $Y$. In Section 8 we shall, as a step towards treating neutral equations, replace $f(\tau)\,d\tau$ by $F(d\tau)$ with $F$ of bounded variation. Then approximation by step functions no longer works. This observation motivates to look for an alternative sufficient condition. \[lem:4.2\] In addition to (i)-(iv) assume that $$\begin{aligned} \label{eq:4.5} &\hbox{a linear map } \gb{\DY, \si(\DY,Y)} \to \BR\hbox{ is continuous}\nonumber\\ &\hbox{if it is sequentially continuous}.\end{aligned}$$ Then the assertion of Lemma \[lem:4.1\] holds. Again we are going to make use of the dominated convergence theorem. Consider a sequence $\{\dy_m\}$ in $\DY$ such that for every $y \in Y$ the sequence $\pa{\dy_m}{y}$ converges to zero in $\BR$. Then for all relevant $t$ and $\tau$ we have $$\lim_{m \to \infty} \dy_m S(t-\tau)q = 0$$ and consequently $$\lim_{m \to \infty} \int_0^t \dy_m S(t-\tau)q \, f(\tau)\,d\tau = 0.$$ So the linear map is, in the sense described in , sequentially continuous and therefore, by the assumption, continuous. Since $$\gb{\DY,\si(\DY,Y)}' = Y,$$ we conclude that is represented by an element of $Y$. In the next section we are going to use these results to show that a certain type of perturbation of a twin semigroup yields again a [*twin*]{} semigroup and then we will also need that with (ii) replaced by - $\dq \in \DY$, we have that $$\label{eq:4.6} y \mapsto \int_0^t \dq S(t-\tau)y\, f(\tau)\,d\tau$$ is represented by an element of $\DY$, to be denoted as $$\label{eq:4.7} \int_0^t \dq S(t-\tau)\, f(\tau)\,d\tau.$$ Applying the two lemmas above, with the role of $Y$ and $\DY$ interchanged, we find that this is indeed the case if either $$\label{eq:4.8} \gb{\DY,\si(\DY,Y)}\quad\hbox{is sequentially complete.}$$ or $$\begin{aligned} \label{eq:4.9} &\hbox{a linear map } \gb{Y, \si(Y,\DY)} \to \BR\hbox{ is continuous}\nonumber\\ &\hbox{if it is sequentially continuous}.\end{aligned}$$ In our treatment of delay differential equations we shall assume and , but in our treatment of renewal equations we shall assume and . This difference is a consequence of what we stated at the start of Section 2: we want that $Y$ is the state space and $\DY$ is an auxiliary space. For delay differential equations we take $Y = B([-1,0])$ and $\DY = NBV([0,1])$, while for renewal equations we take $Y = NBV([0,1])$ and $\DY = B([-1,0])$. So in terms of the two function spaces involved, the assumptions are identical (and these assumptions are substantiated in Appendix B), but because their roles are interchanged the formulations are a mirror image of each other. As in the next section we shall use both properties, we state \[def:4.3\] We say that a norming dual pair $(Y,\DY)$ is suitable for twin perturbation if - at least one of and holds; and - at least one of and holds Finite dimensional range perturbation of twin semigroups ======================================================== In this section we consider the following situation: - $(Y,\DY)$ is a norming dual pair that is suitable for twin perturbation, cf. Definition \[def:4.3\]; - $\{S_0(t)\}$ is a twin semigroup on $(Y,\DY)$ with generator $C_0$; - For $j=1,\ldots,n$ the elements $q_j \in Y$ and $\dq_j \in \DY$ are given. Our aim is to define constructively a twin semigroup $\{S(t)\}$ with generator $C$ defined by $$\label{eq:5.1} \DOM{C} = \DOM{C_0}.\qquad Cy = C_0y + \sum_{j=1}^n \pa{\dq_j}{y}q_j.$$ The first step is to introduce a $n \times n$-matrix valued function $k$ on $[0,\infty)$ via $$\label{eq:5.2} k_{ij}(t) = \dq_i S_0(t) q_j.$$ Note that, by assumption, $t \mapsto k(t)$ is locally bounded and measurable. With the kernel $k$ we associate its resolvent $r$. This is by definition the unique solution of the matrix renewal equation $$\label{eq:5.3} k + k \ast r = r = k + r \ast k$$ or, equivalently, $$\label{eq:5.4} r = \sum_{j=1}^\infty k^{j\ast},$$ where $k^{1\ast} := k$ and $k^{m\ast} := k \ast k^{(m-1)\ast}$ for $m \ge 2$. Here $\ast$ denotes the usual convolution product of functions. Motivated by we consider the equation $$\label{eq:5.5} S(t) = S_0(t) + \int_0^t S_0(t-\tau)BS(\tau)\,d\tau,$$ where $$\label{eq:5.6} By := \sum_{j=1}^n \pa{\dq_j}{y}q_j.$$ By letting $B$ act on we obtain, for given initial point $y \in Y$, a finite dimensional renewal equation. To formulate this equation, we first write as $$\label{eq:5.7} By = \pa{\dq}{y}\cdot q$$ where $\dq$ is the $n$-vector with $\DY$-valued components $\dq_j$ and similarly $q$ is the $n$-vector with $Y$-valued components $q_j$ and where $\cdot$ denotes the inner product in $\BR^n$. We can factor (a rank factorization) $B$ as $B= B_1B_2$ with $B_1 : Y \to \BR^n$ and $B_2 : \BR^n \to Y$ defined by $$\label{eq:5.7a} B_1y = \pa{\dq}{y},\qquad B_2x = \sum_{j=1}^n x_jq_j$$ Now let act on $y \in Y$ and next act on the resulting identity with the vector $\dq$. This yields the equation $$\label{eq:5.8} v(t)y = \dq S_0(t)y + \int_0^t k(t-\tau)v(\tau)y\,d\tau,$$ where $v(t)y$ corresponds to $\dq S(t)y = B_1S(t)y$. The solution of can be expressed in terms of the resolvent $r$ of the kernel $k$ and the forcing function $t \mapsto \dq S_0(t)y$ by the formula $$\label{eq:5.9} v(t)y = \dq S_0(t)y + \int_0^t r(t-\tau)\dq S_0(\tau)y\,d\tau.$$ And now that $v(\novar)y$, representing $\dq S(\novar)y$, can be considered as known, becomes an explicit formula $$\label{eq:5.10} S(t) = S_0(t) + \int _0^t S_0(t-\tau)q\cdot v(\tau)\,d\tau.$$ Please note that, with this definition of $S(t)$, we do indeed have that $$v(t)y = \dq S(t)y$$ (compare to ). Formula is well suited for proving, on the basis of Lemma \[lem:4.1\] or Lemma \[lem:4.2\], that $S(t)$ maps $Y$ into $Y$. But as long as we do not know yet that the perturbed family $\{S(t)\}$ has the semigroup property, we cannot use the “dual” variant of these lemmas to prove that $S(t)$ maps $\DY$ into $\DY$. Therefore we now first consider $$\label{eq:5.12} S(t) = S_0(t) + \int_0^t S(t-\tau) BS_0(\tau)\,d\tau$$ which is the variant of in which the roles of $S$ and $S_0$ are interchanged. Let act (from the right) on $\dy \in \DY$ and next let the resulting identity act on the vector $q$. This yields the equation $$\label{eq:5.13} \dy w(t) = \dy S_0(t)q + \int_0^t \dy w(t-\tau) k(\tau)\,d\tau,$$ where $\dy w(t)$ corresponds to $\dy S(t)q$. The formula $$\label{eq:5.14} \dy w(t) = \dy S_0(t)q + \int_0^t \dy S_0(t-\tau)q\,r(\tau)\,d\tau$$ expresses the solution of in terms of the forcing function $\dy S_0(t)q$ and the resolvent $r$ of the kernel $k$. Next we rewrite in the form $$\label{eq:5.15} S(t) = S_0(t) + \int_0^t w(t-\tau) \cdot \dq S_0(\tau)\,d\tau.$$ Please note that indeed $\dy w(t) = \dy S(t)q$ (compare to ). Of course we should now verify that the integrals in and do indeed define the same object. Writing the integral in as $w_0 \ast v$ and the integral in as $w \ast v_0$, equality follows from written in the form $$v = v_0 + r \ast v_0$$ and written in the form $$w = w_0 + w_0 \ast r$$ since $$\begin{aligned} w_0 \ast v &= w_0 \ast (v_0 + r \ast v_0) = w_0 \ast v_0 + w_0 \ast r \ast v_0\\ &= (w_0 + w_0 \ast r) \ast v_0 = w \ast v_0.\end{aligned}$$ \[thm:5.1\] The combination – or, equivalently, the combination of –, defines a twin semigroup $\bigl\{S(t)\bigr\}$ with generator $C$ defined in . Since $(Y,\DY)$ is suitable for twin perturbation, we can use and either Lemma \[lem:4.1\] or Lemma \[lem:4.2\] to deduce that $S(t)$ maps $Y$ into $Y$. Similarly we can use and the observation concerning to deduce that $S(t)$ maps $\DY$ into $\DY$. So $\bigl\{S(t)\bigr\}$ is a twin operator. With a view to deriving the semigroup property $$\label{eq:5.16} S(t+s) = S(t)S(s),\qquad t,s \ge 0,$$ we first formulate the auxiliary result \[lem:5.2\] The solution $v(\novar)y$ of has the property $$\label{eq:5.17} v(t+s)y = v(t)S(s)y$$ From it follows that $$\begin{aligned} v(t+s)y &= \dq S_0(t)S_0(s)y + \int_0^s k(t+s-\tau)v(\tau)y\,d\tau\\ &\qquad + \int_0^t k(t-\si)v(s+\si)y\,d\si\end{aligned}$$ and by uniqueness follows provided $$\dq S_0(t)S_0(s)y + \int_0^s k(t+s-\tau)v(\tau)y\,d\tau = \dq S_0(t)S(s)y.$$ Noting that $$k(t+s-\tau) = \dq S_0(t+s-\tau)q = \dq S_0(t)S_0(s-\tau)q,$$ we conclude from that this identity does indeed hold. To verify , we start from and write $$\begin{aligned} S(t+s)y &= S_0(t)S_0(s)y + \int_0^s S_0(t+s-\tau)q \cdot v(\tau)y\,d\tau\\ &\qquad + \int_0^t S_0(t-\si)q \cdot v(\si+s)y\,d\si\\ &= S_0(t)S(s)y + \int_0^t S_0(t-\si)q \cdot v(\si)S(s)y\,d\si\\ &= S(t)S(s)y.\end{aligned}$$ Both the property $S(0)=I$ and the measurability, for all $y \in Y$, $\dy \in \DY$, of $t \mapsto \dy S(t)y$ follow from and the corresponding properties of $\{S_0(t)\}$. The exponential estimate for $\dy S_0(t) y$ yields exponential estimates for both the kernel $k$ and the forcing function $\dq S_0(\novar)y$ in the renewal equation . Therefore, see Theorem \[thm:resolvent\] with $\mu(dt)=k(t)dt$, we obtain an exponential estimate for the resolvent $\rho(dt)=r(t)dt$, and hence via an exponential bound for $v(t)y$. Finally, using we obtain an exponential bound for $\dy S(t)y$. It remains to compute the Laplace transform, cf. . Since $$\label{eq:5.18} \int_0^\infty e^{-\la t} \dy S_0(t)y\,dt = \dy (\la I - C_0)^{-1}y,$$ we obtain by Laplace transformation of the identity $$\int_0^\infty e^{-\la t} \dy S(t)y\,dt = \dy (\la I - C_0)^{-1}y + \dy (\la I - C_0)^{-1}q \novar \bar{v}(\la)y.$$ Laplace transformation of either or and yields $$\begin{aligned} \label{eq:5.18a} \bar{v}(\la)y &= \bigl[ I - \dq(\la I - C_0)^{-1}q\bigr]^{-1}\,\dq(\la I - C_0)^{-1}y\nonumber\\ &= \bigl[ I - \bar k(\la) \bigr]^{-1}\,\dq(\la I - C_0)^{-1}y\end{aligned}$$ By combining the last two identities we arrive at $$\begin{aligned} \label{eq:5.19} \int_0^\infty e^{-\la t}\dy S(t)y\,dt &= \dy (\la I - C_0)^{-1}y + \dy (\la I - C_0)^{-1}q\, \cdot\nonumber \\ &\qquad \cdot\, \bigl[ I - \dq(\la I - C_0)^{-1}q\bigr]^{-1}\,\dq(\la I - C_0)^{-1}y.\end{aligned}$$ It remains to check that the right hand side of is exactly $\dy (\la I - C)^{-1}y$ when $C$ is defined by . So consider the equation $$(\la I - C)\eta = y.$$ By this is equivalent to $$(\la I - C_0)\eta = y + \pa{\dq}{\eta}\cdot q$$ and hence to $$\eta = (\la I - C_0)^{-1}y + \sum_{j=1}^n \pa{\dq_j}{\eta}(\la I - C_0)^{-1}q_j.$$ In particular, $$\pa{\dq_k}{\eta} = \dq_k(\la I - C_0)^{-1}y + \sum_{j=1}^n \dq_k (\la I - C_0)^{-1}q_j\,\pa{\dq_j}{\eta}$$ or, in vector form, $$\pa{\dq}{\eta} = \dq (\la I - C_0)^{-1}y + \dq (\la I - C_0)^{-1}q \pa{\dq}{\eta}.$$ Hence $$\begin{aligned} \label{eq:5.19a} \eta = (\la I - C)^{-1}y &= (\la I - C_0)^{-1}y + (\la I - C_0)^{-1}q \nonumber\\ &\hspace{1.75cm} \times\gb{ I - \dq (\la I - C_0)^{-1}q}^{-1}\,\dq (\la I - C_0)^{-1}y\end{aligned}$$ and comparison with shows that indeed $$\label{eq:5.20} \int_0^\infty e^{-\la t} \dy S(t)y\,dt = \dy (\la I - C)^{-1}y.$$ This completes the proof of Theorem \[thm:5.1\]. \[col:5.3\] The subspace $X$ of strong continuity is the same for $\{S_0(t)\}$ and $\{S(t)\}$. Likewise $\{\DS_0(t)\}$ and $\{\DS(t)\}$ do have the same subspace of strong continuity $X^{\odot}$. The special representation of the perturbed semigroup $S(t)$ given in respectively – and – allows us to use Theorem \[thm:Tauber\] to derive a result about the asymptotic behaviour of $S(t)$ without using a spectral mapping theorem, eventual compactness or eventual norm continuity of the semigroup $S(t)$. \[thm:5.4\] Under the assumptions of this section let $k$, given by , be integrable. Suppose that $S_0(t)$ is bounded and that $(\la I - C_0)^{-1}$ is bounded for $\Re \la \ge 0$. If $$\label{eq:5.22} \det\gb{I - \bar k(\la)}\quad\hbox{has no zeros for } \Re \la \ge 0,$$ then $$\label{eq:5.23} \|S(t)C^{-1}\| \to 0\qquad\hbox{as}\quad t \to \infty.$$ As a consequence we have that $S(t)y \to 0$ as $t \to \infty$ for every $y$ in the norm-closure of $\DOM{C}$. We first show that $S(t)$ is bounded. From the half-line Gel’fand theorem, see Theorem \[thm:Gelfand\], applied to the absolutely continuous measure $\mu(dt) = k\,dt$, it follows that the resolvent $\rho(dt) = r\,dt$ is an absolutely continuous bounded measure. Fix $y$ in $Y$ and $\dy \in \DY$. From Theorem \[thm:convo-meas-Borel-fun\] and it follows that $v\,dt$ is a bounded measure and another application of Theorem \[thm:convo-meas-Borel-fun\] shows that $\dy S(t) y$ as defined via , is a bounded Borel function and hence $S(t)$ is a bounded twin semigroup. If $(\la I - C_0)^{-1}$ is bounded for $\Re \la \ge 0$ and holds, then it follows from and that $(\la I - C)^{-1}$ is bounded for $\Re \la \ge 0$. This completes the proof that $S(t)$ is bounded and that $\si(C) \mcap i\BR = \emptyset$. So an application of Theorem \[thm:Tauber\] yields the proof. The following variant of Theorem \[thm:5.4\] is motivated by RFDE and various boundary value problems \[thm:5.5\] Under the assumptions of this section let $k$, given by , be of bounded variation with $k(0) = 0$. Suppose that $S_0(t)$ is bounded and that $(\la I - C_0)^{-1}$ has a simple pole at $\la = 0$ but is otherwise bounded for $\Re \la \ge 0$. Assume that $$\label{eq:5.24} \pa{\dq}{P_0q}P_0y = \pa{\dq}{P_0y}P_0q,$$ where $P_0 : Y \to Y$ denotes the spectral projection onto the eigenspace of $C_0$ at $\la = 0$. If $$\label{eq:5.25} \det\gb{\la I - \hat k(\la)}\quad\hbox{has no zeros for } \Re \la \ge 0,$$ then $$\label{eq:5.26} \|S(t)C^{-1}\| \to 0\qquad\hbox{as}\quad t \to \infty.$$ As a consequence we have that $S(t)y \to 0$ as $t \to \infty$ for every $y$ in the norm-closure of $\DOM{C}$. We first show that $S(t)$ is bounded. Let $\mu(dt) = k\,dt$ and observe that $$\what\mu(\la) = \bar k(\la) = \frac{1}{\la}\hat k(\la).$$ So it follows from and the half-line Gel’fand theorem, see Theorem \[thm:Gelfand\], applied to $\mu$, that the resolvent $\rho$ is an absolutely continuous bounded measure $\rho = r\,dt$. Fix $y$ in $Y$ and $\dy \in \DY$. From Theorem \[thm:convo-meas-Borel-fun\] and it follows that $v\,dt$ is a bounded measure and another application of Theorem \[thm:convo-meas-Borel-fun\] shows that $\dy S(t) y$ as defined via , is a bounded Borel function and hence $S(t)$ is a bounded twin semigroup. To show that $(\la I - C)^{-1}$ is bounded for $\Re \la \ge 0$ first observe that $$\begin{aligned} \label{eq:5.27} (\la I - C)^{-1}y &= \la \bigl[\la I - \hat k(\la)\bigr]^{-1}\gb{(\la I - C_0)^{-1}y - \dq (\la I - C_0)^{-1}q\,(\la I - C_0)^{-1}y\nonumber\\ &\qquad\qquad\qquad + \dq (\la I - C_0)^{-1}y\, (\la I - C_0)^{-1}q}.\end{aligned}$$ Using the assumption on $(\la I - C_0)^{-1}$ we can write $$(\la I - C_0)^{-1}y = \frac{1}{\la}P_0y + H(\la)y,$$ where $P_0 : Y \to Y$ denotes the spectral projection onto the eigenspace of $C_0$ at $\la = 0$ and $H(\la) : Y \to Y$ is a bounded linear operator for $\Re \la \ge 0$. Using this we can expand $$\begin{aligned} &-\dq (\la I - C_0)^{-1}q\,(\la I - C_0)^{-1}y + \dq (\la I - C_0)^{-1}y\, (\la I - C_0)^{-1}q\\ &\qquad=\\ &\ \frac{1}{\la^2}\gb{\dq P_0y\, P_0q-\dq P_0q\,P_0y }\\ &\quad +\frac{1}{\la}\gb{\dq\, P_0y H(\la)q + \dq H(\la)y P_0q -\dq P_0q\,H(y)y - \dq H(\la)q\, P_0y}\\ &\quad + \dq H(\la) y \,H(\la)q - \dq H(\la)q\,H(\la)y \end{aligned}$$ Condition shows that the term $\la^{-2}$ vanishes and since $(\la I - C_0)^{-1}y$ has a simple pole at $\la = 0$ and $H(\la)y$ is bounded for $\Re \la \ge 0$, it follows from that $(\la I - C)^{-1}$ is bounded for $\Re \la \ge 0$. This completes the proof that $S(t)$ is bounded and $\si(C) \mcap i\BR = \emptyset$. So an application of Theorem \[thm:Tauber\] yields the proof. Note that condition is automatically satisfied if the null space of $C$ is one-dimensional. In general, the condition that $\la = 0$ is a simple pole of $(\la I - C)^{-1}$ implies that the generalized null space of $C$ equals the null space of $C$, but does not give any information about the dimension of the null space of $C$. In the example of RFDE, the null space of $C$ is $n$-dimensional. In Section 4 there was no need to use the perturbation approach developed in this section. Yet, alternatively, we can first concentrate on the special case $\ze = 0$, calling the corresponding twin semigroup $\{S_0(t)\}$ and its generator $C_0$. Next we define for $i = 1,\ldots,n$ elements $q_i \in Y$ and $\dq_i \in \DY$ by $$\label{eq:6.29} q_i(\th) = \begin{cases} 0 &\th < 0\\ e_i &\th = 0, \end{cases}$$ where $e_i$ is the $i$-th unit vector in $\BR^n$ and $$\label{eq:6.30} \dq_i(\th) = \ze_i(\th),$$ where $\ze_i$ is the $i$-th row of the matrix valued function $\ze$. For the matrix $k$ introduced in we find $$\label{eq:6.31} k_{ij}(t) = \dq_i S_0(t) q_j = \int_0^1 \ze_i(d\tau) \chi_{t-\tau \ge 0} e_j = \ze_{ij}(t).$$ With the convention that $\ze(\tau) = \ze(1)$ for $\tau \ge 1$, we can also write (with $y$ corresponding to $\ph$) $$\begin{aligned} \label{eq:6.32} \dq S_0(t)y &= \int_0^t \ze(d\tau)y(0) + \int_t^1 \ze(d\tau) y(t-\tau)\nonumber\\ &= \ze(t)y(0) + \int_t^1 \ze(d\tau) y(t-\tau).\end{aligned}$$ Comparing the right hand side to the right hand side of (2.5) on page 16 of [@Diek95], we see that the RE of the present paper is identical to equation (2.4a) on page 16 of [@Diek95]: $$\dot x(t) = \int_0^t \ze(\th) \dot x(t-\th)\,d\th + \ze(t)y(0) + \int_t^1 \ze(d\tau) y(t-\tau).$$ We conclude that $v(\cdot)y$ in corresponds to $\dot x$ in this equation. Next apply to the element of $\DY$ that corresponds to the Dirac measure in $-\th \in [0,1]$. This yields $$\label{eq:6.33} \gb{S(t)y}(\th) = y(t+\th) + \int_0^t r(t-\tau+\th)\cdot v(\tau)y\,d\tau,$$ where we adopted the convention that both $y$ and $q$ are extended by their value in zero. It follows that $$\label{eq:6.34} \gb{S(t)y}(\th) = \begin{cases} y(t+\th), &t+\th \le 0\\ y(0) + \int_0^{t+\th} v(\tau)y\,d\tau, &t+\th \ge 0. \end{cases}$$ Since $$y(0) + \int_0^{t+\th} v(\tau)y\,d\tau = y(0) + \int_0^{t+\th} \dot x(\tau;y)\,d\tau = x(t+\th;y),$$ this corresponds exactly to . We conclude that the direct approach and the perturbation approach are fully consistent. We conclude by showing that the assumptions of Theorem \[thm:5.5\] are satisfied for RFDE. From it follows that $k$ is of bounded variation. Furthermore, $$\gb{(\la I - C_0)^{-1}y}(\th) = \frac{e^{\la\th}}{\la}y(0) + \int_{\th}^0 e^{\la(\th-\si)} y(\si)\,d\si$$ and $P_0 : Y \to Y$ is given by $$P_0 y = y(0) \mathbbm{1},$$ where $\mathbbm{1} \in Y$ denotes the function that is identically one. Using and , observe that $$\pa{\dq}{P_0q}P_0y = \pa{\ze}{I\,\mathbbm{1}}\,y(0)\,\mathbbm{1} = \ze(1)y(0)\,\mathbbm{1}$$ and $$\pa{\dq}{P_0y}P_0q = \pa{\ze}{y(0)\,\mathbbm{1}}\,\mathbbm{1} = \ze(1)y(0)\,\mathbbm{1}.$$ This shows that condition is satisfied for RFDE. Therefore an application of Theorem \[thm:5.5\] yields that if $$\det \gb{\la I - \int_0^1 e^{-z\si}\,d\ze(\si)}\qquad\hbox{ has no zeros for } \Re \la \ge 0,$$ then for $y \in C\gb{[-1,0];\BR^n}$ we have $S(t)y \to 0$ as $t \to \infty$. Since for RFDE $S(1)y \in \DOM{C}$ for every $y \in B\gb{[-1,0];\BR^n}$, we conclude that $S(t)y \to 0$ as $t \to \infty$ for every $y \in B\gb{[-1,0];\BR^n}$. RE - Renewal equations with “smooth” kernels ============================================ The RE $$\label{eq:7.1} b(t) = \int_0^1 k(a)b(t-a)\,da$$ arises in the context of age-structured population dynamics. In that context, $b(t)$ is the rate at which newborn individuals are added to the population at time $t$ and $$k(a) = \F(a)\be(a),$$ with $\F(a)$ the probability to survive to at least age $a$ and $\be(a)$ the age-specific fecundity (it is helpful to think in terms of mothers and daughters, with the male subpopulation implicitly included via a fixed sex ratio). Note that we have scaled the time variable such that the maximum age at which reproduction is possible equals one. It is convenient to define $k(a) = 0$ for $a > 1$. To facilitate statements and arguments based on the interpretation, we focus in this section our attention on a scalar equation. Generalization to $n$-vector valued functions $b$ and $n \times n$-matrix valued kernels $k$ is straightforward. We consider as a rule for extending the function $b$ and, to get started, supplement it by prescribing the history of $b$ at a particular time, say $t=0$: $$\label{eq:7.2} b(\th) = \ph(\th),\qquad -1 \le \th \le 0.$$ By translation along the extended function, i.e., by putting $$\label{eq:7.3} T(t)\ph = b_t$$ which is a shorthand for $$\label{eq:7.4} \gb{T(t)\ph}(\th) = b(t+\th;\ph)$$ with $b(\novar;\ph)$ the unique solution of –, we define a dynamical system. But what do we choose for the state space $X$ on which the dynamical system acts? Since $b$ is a rate, we get numbers by integrating with respect to time. So the interpretation suggests to take $$\label{eq:7.5} X = L^1\gb{[-1,0];\BR}$$ as is indeed done in [@DGG07]. The bonus is that the semigroup $\{T(t)\}$ defined by is strongly continuous. But when we compute the infinitesimal generator $A$, we find (with $AC$ standing for “absolutely continuous”) $$\label{eq:7.6} \DOM{A} = \bigl\{\ph \in AC \mid \ph(0) = \int_0^1 k(a)\ph(-a)\,da\,\bigr\},\qquad A\ph = \ph'$$ showing that all information about the rule for extension is in the domain of $A$ and that the action of $A$ only reflects the translation. The trouble with this is that even small changes in the rule for extension correspond, at the generator level, to unbounded perturbations. In [@DGG07] it is shown how perturbation theory of dual semigroups, aka sun-star calculus, can be used to overcome this difficulty. Here we show that the formalism of twin semigroups on a norming dual pair of spaces provides an alternative approach. In Section 12 we shall show that this new approach allows us to cover “neutral” RE as well, where the adjective neutral expresses that we replace $k(a)da$ by a measure. Soon we will assume that $k$ is a given bounded measurable function (defined on $[0,\infty)$ but with support in $[0,1]$), but for the time being, while discussing the representation of the solution of –, it suffices that $k$ is in $L^1$. Combining and we obtain $$\label{eq:7.7} b = k \ast b + f$$ with $$\label{eq:7.8} f(t) = \int_t^1 k(a)\ph(t-a)\,da = \int_{t-1}^0 k(t-\th)\ph(\th)\,d\th$$ In the theory of RE, cf. Section 4, [@GLS90] and Appendix A, the solution $r$ of $$\label{eq:7.9} k * r + k = r = r * k + k$$ is called the [*resolvent*]{} of the kernel $k$, in particular since the solution of is given by $$\label{eq:7.10} b = f + r * f.$$ Note that $$\label{eq:7.11} r = \sum_{j=1}^\infty k^{j*}.$$ More importantly, note that would lead to $f(t) = k(t)$ if we replace $\ph(\th)\,d\th$ by the Dirac mass in $\th = 0$. So the situation is reminiscent of the situation for delay differential equations, where, when working with $C$, one finds that the fundamental solution corresponds to a discontinuous initial condition. Here, while working with $L^1$, we find that the resolvent corresponds to a measure as initial condition (corresponding to a cohort of newborn individuals in the population dynamical context). And our strategy will be the same: enlarge the state space, even though this entails the loss of strong continuity. In the tradition of delay equations we will represent measures by $NBV$ functions. So let now $$\label{eq:7.12} Y = NBV\gb{[-1,0];\BR}$$ but with the normalization convention that the elements are zero in the right end point $\th = 0$. Let $$\label{eq:7.13} \DY = B\gb{[0,1];\BR}$$ with pairing defined by $$\label{eq:7.13a} \pa{\dy}{y} = \int_{-1}^0 \dy(-\th)\,y(d\th)$$ and let $$\label{eq:7.14} k \in \DY$$ be given. We still consider but replace by $$\label{eq:7.15} f(t) = \int_{t-1}^0 k(t-\th)\psi(d\th)$$ with $\psi \in Y$ considered as the initial condition. (So in the population dynamical context one should interpret $\psi$ as the cumulative number of newborns, but otherwise nothing changes. In particular there is still a population level birth [*rate*]{} for $t > 0$. This will change in Section 12, where we work with cumulative quantities throughout.) For given $\psi \in Y$ equation with $f$ given by has a unique solution given explicitly by . We define $$\label{eq:7.16} B(t) = \int_0^t b(\tau)\,d\tau,\qquad t > 0,$$ $$\label{eq:7.17} B(\th) = \psi(\th),\qquad \th \le 0,$$ (where we have suppressed the dependence of $b$ on $\psi$ in the notation) and next $S(t) : Y \to Y$ by $$\label{eq:7.18} \gb{S(t)\psi}(\th) = B(t+\th) - B(t).$$ Note that we subtract $B(t)$ in order to comply with the normalization that the value in $\th = 0$ should be zero. It is very well possible to check that $\{S(t)\}$ is a twin semigroup on the norming dual pair $(Y,\DY)$ specified by and and to determine the generator via the Laplace transform. Here, however, we establish the relevant facts via the perturbation theory of Section 6. This allows us to show that and are identical (to call the kernel in $k$, introduces the risk of ambiguity when invoking Section 6, but in fact there is no need to worry). The twin semigroup $\{S_0(t)\}$ defined by $$\label{eq:7.19} \gb{S_0(t)\psi}(\th) = \begin{cases} \psi(t+\th) &t+\th \le 0\\ 0 &t+\th > 0 \end{cases}$$ corresponds to a kernel $k$ that is identically equal to zero, and so to trivial extension of the initial function. A straightforward calculation reveals that $$\int_0^\infty e^{-\la t} \gb{S_0(t)\psi}(\th)\,dt = e^{\la\th}\int_\th^0 e^{-\la\si} \psi(\si)\,d\si$$ and next that $$\int_0^\infty e^{-\la t}\dy S_0(t)\psi\,dt = \dy (\la I - C_0)^{-1}\psi,$$ where $$\begin{aligned} \label{eq:7.20} \begin{split} \DOM{C_0} &= \bigl\{ \psi \mid \exists \psi' \in Y \mid \psi(\th) = \int_0^\th \psi'(\si)\,d\si\,\bigr\}\\ C_0\psi &= \psi' \end{split}\end{aligned}$$ Note that the subspace of strong continuity is given by $$\begin{aligned} \label{eq:7.21} X &= AC\gb{[-1,0];\BR}\nonumber\\ &= \bigl\{ \psi \mid \exists \psi' \in L^1\gb{[-1,0];\BR} \mid \psi(\th) = \int_0^\th \psi'(\si)\,d\si\,\bigr\}\end{aligned}$$ according to Theorem \[thm:3.1\] and the fact that NBV functions are dense in $L^1$ (admittedly we ignore an isometric isomorphism when using the same symbol $X$ in and ). To capture the true rule for extension, we introduce $q \in Y$ and $\dq \in \DY$ according to $$\label{eq:7.22} q(\th) = \begin{cases} 0 & \th = 0\\ -1 & -1 \le \th < 0 \end{cases}$$ (i.e., $q$ is the Heaviside function that represents the Dirac measure in $\th = 0$) and the requirement, inspired by , that $$\label{eq:7.23} \dq S_0(t)q = k(t).$$ It follows from and that $$\dq S_0(t)q = \begin{cases} \dq(t) &\mbox{for } 0 \le t \le 1\\ 0 &\mbox{otherwise} \end{cases}$$ so we can in fact identify $\dq$ and $k$. The perturbed semigroup is defined by and this involves the solution of , which is a RE with kernel $k$ and forcing function $$\begin{aligned} \dq S_0(t)\psi &= \int_{-1}^0 k(-\th)\gb{S_0(t)\psi}(d\th) = \int_{-1}^{-t} k(-\th)\psi(t+d\th)\\ &=\int_{t-1}^0 k(t-\si)\psi(d\si)\end{aligned}$$ which is exactly equal to $f(t)$ defined in . We conclude that in the present setting is simply another way of writing and that, accordingly, we may replace $v(\tau)y$ in by $b(\tau)$. It only remains to verify that amounts to . With $Y$ and $\DY$ given by, respectively, and , one can turn into a pointwise equality (just use step functions from $\DY$ in the pairing that provides the precise meaning of the integral). It reads $$\gb{S(t)\psi}(\th) = \gb{S_0(t)\psi}(\th) + \int_0^t \gb{S_0(t-\tau)q}(\th) b(\tau)\,d\tau$$ with $$\gb{S_0(t)\psi}(\th) = \begin{cases} \psi(t+\th) & t+\th \le 0\\ 0 & t+\th > 0 \end{cases}$$ and $$\begin{aligned} \int_0^t \gb{S_0(t-\tau)q}(\th) b(\tau)\,d\tau &= \int_0^t - \chi_{t-\tau+\th < 0}\, b(\tau)\,d\tau\\ &= -\int_{\max\{t+\th,0\}}^t b(\tau)\,d\tau\\ &= \int_0^{\max\{t+\th,0\}} b(\tau)\,d\tau - \int_0^t b(\tau)\,d\tau.\end{aligned}$$ On account of – we conclude that the twin semigroup defined by is equivalently described by . In terms of $B$ we can rewrite as the delay differential equation $$\label{eq:7.24} B'(t) = \int_{-1}^0 k(-\si)B_t(d\si).$$ If we formally differentiate with respect to $t$ and next evaluate at $t=0$, we obtain for $\th < 0$ $$\frac{d}{dt}\gb{S(t)\psi}(\th)\big|_{t=0} = \psi'(\th) - B'(0) = \psi'(\th) - \int_{-1}^0 k(-\si)\psi(d\si)$$ which is completely in line with the characterization of the generator $C$ in when , and $\dq = k$ are taken into account. **Part III: Unbounded perturbations** **describing neutral equations** The variation-of-constants formula for forcing functions with finite dimensional range revisited ================================================================================================ As will become clear in Sections 9 and 10 below, the analysis of relatively bounded perturbations requires to consider linear functionals $$\label{eq:8.2} \dy \mapsto \int_0^t\dy S(t-\tau)q \, F(d\tau)$$ for a given $\BR$-valued $(N)BV$ function $F$. When is such a functional represented by an element of $Y$? The proof of Lemma \[lem:4.2\] carries over verbatim if we replace $f(\tau)\,d\tau$ by $F(d\tau)$. The proof of Lemma \[lem:4.1\], on the other hand, breaks down. To save the underlying idea, we perform integration by parts and first rewrite as $$\label{eq:8.3} \dy \mapsto \int_0^t d_\si\bigl[\dy S(\si)q \bigr] F(t-\si) + F(t)\pa{\dy}{q}$$ and next incorporate the last term into the first term by redefining $\dy S(\si)q$ as zero for $\si = 0$. In we can allow $F$ to be a bounded measurable function, but we have to require that for every $\dy \in \DY$ the function $$t \mapsto \dy S(t)q\qquad\hbox{for } t > 0$$ with value zero for $t=0$, is of bounded variation. Once this is assumed, the proof of Lemma \[lem:4.1\] can be copied in order to show \[lem:8.1\] Let $\bigl\{S(t)\bigr\}$ be a twin semigroup on a norming dual pair $(Y,\DY)$. Assume holds, i.e., assume that $(Y,\si(Y,\DY))$ is sequentially complete. Let $q \in Y$ be given. For $\dy \in \DY$ define $$\label{eq:8.4} \dy W(\si) = \begin{cases} 0 &\hbox{for } \si = 0\\ \dy S(\si)q &\hbox{for } \si > 0. \end{cases}$$ Assume that for all $\dy \in \DY$ the function $$\si \mapsto \dy W(\si)$$ belongs to $\NBV_{loc}\gb{[0,\infty),\BR}$. Let $F : [0,\infty) \to \BR$ be locally bounded and measurable. Then there exists $u(t) \in Y$ such that for all $\dy \in \DY$ $$\label{eq:8.5} \int_0^t d_\si\bigl[\dy W(\si)\bigr]\, F(t-\si) = \pa{\dy}{u(t)}.$$ For completeness we also state \[lem:8.2\] Let $\bigl\{S(t)\bigr\}$ be a twin semigroup on a norming dual pair $(Y,\DY)$. Assume holds, i.e., assume that a linear map $\gb{\DY,\si(\DY,Y)} \to \BR$ is continuous if it is sequentially continuous. Let $q \in Y$ be given. Let $F : [0,\infty) \to \BR$ be of locally bounded variation. Then there exists $u(t) \in Y$ such that for all $\dy \in \DY$ $$\label{eq:8.6} \int_0^t \dy S(t-\tau)q \, F(d\tau) = \pa{\dy}{u(t)}.$$ \[lem:8.3\] Let $\bigl\{S(t)\bigr\}$ be a twin semigroup on a norming dual pair $(Y,\DY)$. Assume holds, i.e., assume that $(\DY,\si(\DY,Y))$ is sequentially complete. Let $\dq \in \DY$ be given. For $y \in Y$ define $$\label{eq:8.7} V(\si)y = \begin{cases} 0 &\hbox{for } \si = 0\\ \dq S(\si)y &\hbox{for } \si > 0. \end{cases}$$ Assume that for all $y \in Y$ the function $$\si \mapsto V(\si)y$$ belongs to $\NBV_{loc}\gb{[0,\infty),\BR}$. Let $F : [0,\infty) \to \BR$ be locally bounded and measurable. Then there exists $\du(t) \in \DY$ such that for all $y \in Y$ $$\label{eq:8.8} \int_0^t F(t-\si)\,d_\si\bigl[V(\si)y\bigr]\, = \pa{\du(t)}{y}.$$ \[lem:8.4\] Let $\bigl\{S(t)\bigr\}$ be a twin semigroup on a norming dual pair $(Y,\DY)$. Assume holds, i.e., assume that a linear map $\gb{Y,\si(Y,\DY)} \to \BR$ is continuous if it is sequentially continuous. Let $\dq \in \DY$ be given. Let $F : [0,\infty) \to \BR$ be of locally bounded variation. Then there exists $\du(t) \in \DY$ such that for all $y \in Y$ $$\label{eq:8.9} \int_0^t F(d\tau) \,\dq S(t-\tau)y \, = \pa{\du(t)}{y}.$$ The main ideas explained by formula manipulation ================================================ In Section 6 we perturbed the abstract ODE $$\label{eq:9.1} \frac{du}{dt} \in C_0u$$ by adding at the right hand side a [*bounded*]{} finite rank perturbation. Here, instead, we shall add a [*relatively bounded*]{} finite rank perturbation. Again we introduce $q_j \in Y$, $j = 1,\ldots, n$, to span the range of the perturbation. But the coefficients are now of the form $\pa{\DQ_j}{C_0u}$ for given $\DQ_j \in \DY$ (so we use a capital letter to alert the reader that the element of $\DY$ does now act on $C_0u$, rather than on $u$ itself). Thus the aim is to study $$\label{eq:9.2} \frac{du}{dt} \in C_0u + q \cdot \DQ C_0u$$ with $$\label{eq:9.3} q \cdot \DQ C_0 u = \sum_{j=1}^n \pa{\DQ_j}{C_0u} q_j.$$ We assume that $C_0$ is the generator of a twin semigroup $\{S_0(t)\}$ and our aim is to construct a twin semigroup $\{S(t)\}$ with a generator that has $\DOM{C_0}$ as its domain of definition and action given by the right hand side of . The construction starts from the variation-of-constants formula (cf. ) $$\label{eq:9.4} S(t) = S_0(t) + \int_0^t S_0(t-\tau)q \cdot \DQ C_0 S(\tau)\,d\tau,$$ or from the variant (cf. ) $$\label{eq:9.5} S(t) = S_0(t) + \int_0^t S(t-\tau)q \cdot \DQ C_0 S_0(\tau)\,d\tau$$ in which the roles of $\{S_0(t)\}$ and $\{S(t)\}$ are interchanged. Again the construction is based on solving a finite dimensional RE, but now this RE involves the Stieltjes integral and a bounded variation kernel (or, equivalently, a measure, cf. Appendix A). To see how the Stieltjes integral might come in, please recall Lemma \[lem:2.3\] and note that this suggests to replace the second term at the right hand side of by $$\int_0^t S(t-\tau)q \cdot d_\tau\DQ\gb{S_0(\tau) - I}$$ when indeed $\tau \mapsto \DQ \gb{S_0(\tau) - I}y$ is $(N)BV$ for all $y \in Y$. Another option, again motivated by Lemma \[lem:2.3\], is to integrate the second term at the right hand side of by parts while assuming that $\tau \mapsto \dy S_0(\tau)q$ is $BV$ for all $\dy \in \DY$. The point of both options is to “neutralize" the unbounded operator $C_0$ and the price we pay is that we have to work with Stieltjes integrals. Define $V_0(t) : Y \to \BR^n$ by $$\begin{aligned} \label{eq:9.6} V_0(t)y &= \DQ C_0 \int_0^t S_0(\si)\,d\si y\nonumber\\ &= \DQ\gb{S_0(t) - I}y\end{aligned}$$ and define a $\BR^{n \times n}$-valued kernel $K$ by $$\label{eq:9.7} K(t) = \DQ\gb{S_0(t) - I}q$$ or, in more detail, $$\label{eq:9.8} K_{ij}(t) = \DQ_i\gb{S_0(t) - I}q_j.$$ If we first change the integration variable in to $\si = t - \tau$, next integrate both sides of the equation with respect to time, and finally apply $\DQ C_0$ to both sides, we obtain the equation $$\label{eq:9.9} V(t) = V_0(t) + \int_0^t K(d\si) V(t-\si)$$ with $$\label{eq:9.10} V(t) = \DQ C_0 \int_0^t S(\tau)\,d\tau.$$ Here is short hand for $$\label{eq:9.11} V(t)y = V_0(t)y + \int_0^t K(d\si) V(t-\si)y$$ which is, for given $y \in Y$, an equation for the $\BR^n$-valued function $t \mapsto V(t)y$. Introducing the notation (cf. Theorem \[thm:convo-meas-Borel-fun\]) $$\label{eq:9.12} \gb{K \bast V}(t) := \int_0^t K(d\si)V(t-\si),$$ we can write , and hence , in the even more compact form $$\label{eq:9.13} V = V_0 + K \bast V.$$ Let $R$ be the resolvent of $K$, i.e., the unique solution of (cf. Theorem \[thm:resolvent\]) $$\label{eq:9.14} K + R \bast K = R = K + K \bast R$$ then the solution of is given by $$\label{eq:9.15} V = V_0 + R \bast V_0.$$ Define $W_0(t) : \DY \to \BR^n$ by $$\label{eq:9.16} \dy W_0(t) = \dy S_0(t)q,\qquad\hbox{for } t > 0,$$ with value zero for $t=0$, where we allow ourselves once more the freedom of writing the element of $\DY$, on which the operator acts, to the left of the operator itself. By applying to $q$ we obtain the equation $$\label{eq:9.17} W(t) = W_0(t) + \int_0^t W(t-\tau)K(d\tau)$$ and accordingly we find for $$\label{eq:9.18} W(t) = S(t)q,\qquad\hbox{for } t > 0,$$ with value zero for $t=0$, the formula $$\label{eq:9.19} W(t) = W_0(t) + \int_0^t W_0(t-\tau)R(d\tau).$$ Again we abbreviate and write as $$\label{eq:9.20} W = W_0 + W \bast K$$ and as $$\label{eq:9.21} W = W_0 + W_0 \bast R.$$ (Please note a notational difficulty: in principle we would like to indicate by the order of the factors in the product which of the two factors is considered as a measure, but, on the other hand, we also want to indicate by the order how the matrix acts on the vector. In and we sacrified the first in order to realize the second.) Motivated by we now define $$\label{eq:9.22} S(t) = S_0(t) + \int_0^t W_0(t-\tau) \cdot V(d\tau)$$ by which we mean that $$\dy S(t)y = \dy S_0(t)y + \int_0^t \dy W_0(t-\tau) \cdot V(d\tau)y.$$ This is compatible with the formula $$\label{eq:9.23} S(t) = S_0(t) + \int_0^t W(t-\tau) \cdot V_0(d\tau)$$ that is motivated by . (Because of and the checking amounts to verifying $$W_0 \bast \gb{V_0 + R \bast V_0} = \gb{W_0 + W_0 \bast R} \bast V_0$$ which is a direct consequence of the associativity and distributivity of the $\bast$-convolution product.) To make all this work, we need in any case that the kernel $K$ defined by ${\eqref{eq:9.7}}$ is of bounded variation. In the next section we shall indeed assume that $t \mapsto K(t)$ belongs to $NBV_{loc}$ (note that is compatible with $K(0) = 0$). The formulas and are based on the additional assumption that for all $y \in Y$ the function $$\label{eq:9.23a} t \mapsto V_0(t)y = \DQ\gb{S_0(t) - I}y\hbox{ belongs to } NBV_{loc}.$$ Note that guarantees that this property of $V_0$ is inherited by $V$, making also well-defined. In Section 12, when applying the theory to renewal equations involving a measure as kernel, we shall find that this assumption indeed holds. But in Section 11, when dealing with NFDE (neutral functional differential equations), we shall need to replace and by their counterparts $$\label{eq:9.24} S(t) = S_0(t) + \int_0^t W_0(d\si) \cdot V(t-\si)$$ and $$\label{eq:9.25} S(t) = S_0(t) + \int_0^t W(d\si) \cdot V_0(t-\si)$$ that are obtained by partial integration. Note carefully that and are based on the definition $$\label{eq:9.26} W_0(0) = 0\quad\hbox{and}\quad W(0) = 0$$ and therefore both of these have a jump of size $q$ in zero, cf. and . By this we mean that both $\dy W_0(t)$ and $\dy W(t)$ have $\pa{\dy}{q}$ as limit for $t \da 0$. When working with or , we replace the earlier additional assumption by the new additional assumption that for all $\dy \in \DY$ $$\label{eq:9.26a} \hbox{the function } t \mapsto \dy W_0(t) = \dy S_0(t)q\hbox{ belongs to } NBV_{loc},$$ where we define the function to be zero at $t=0$, cf. . Relatively bounded finite dimensional range perturbation of twin semigroups =========================================================================== Throughout this section we assume - $\{S_0(t)\}$ is a twin semigroup with generator $C_0$; - the elements $q_j \in Y$ and $\DQ_j \in \DY$, $j=1,2,\ldots,n$, are such that for $i,j = 1,\ldots,n$ the function $$\label{eq:10.1} t \mapsto K_{ij}(t) := \DQ_i \gb{S_0(t) - I}q_j$$ belongs to $NBV_{loc}\gb{[0,\infty);\BR}$ and is continuous in $t = 0$. Moreover, we use the following notation and definitions - $R$ denotes the resolvent of $K$, i.e., the solution of . Note that $R$ too is continuous in $t=0$. - $V_0(t) : Y \to \BR^n$ is for $t \ge 0$ defined by $$\label{eq:10.2} V_0(t) = \DQ\gb{S_0(t) - I}.$$ - $W_0(t) : \DY \to \BR^n$ is for $t > 0$ defined by $$\label{eq:10.3} W_0(t) = S_0(t)q$$ and $W_0(0) = 0$. - $V(t) : Y \to \BR^n$ is for $t \ge 0$ defined by $$\label{eq:10.4} V(t) = V_0(t) + \int_0^t R(d\tau)\,V_0(t-\tau).$$ - $W(t) : \DY \to \BR^n$ is for $t > 0$ defined by $$\label{eq:10.5} W(t) = W_0(t) + \int_0^t W_0(t-\tau)\, R(d\tau)$$ and $W(0) = 0$. \[thm:10.1\] Let $(Y,\DY)$ be a norming dual pair such that and hold. Assume that $t \mapsto V_0(t)y$ belongs to $NBV_{loc}\gb{[0,\infty); \BR^n}$ for all $y \in Y$. Then the same holds for the function $t \mapsto V(t)y$ and $$\label{eq:10.6} S(t) = S_0(t) + \int_0^t W_0(t-\tau) \cdot V(d\tau)$$ defines a twin semigroup with generator $C$ given by $$\label{eq:10.7} \DOM{C}= \DOM{C_0},\qquad Cy = C_0y + \pa{\DQ}{C_0y}\cdot q.$$ \[thm:10.2\] Let $(Y,\DY)$ be a norming dual pair such that and hold. Assume that $t \mapsto \dy W_0(t)$ belongs to $NBV_{loc}\gb{[0,\infty); \BR^n}$ for all $\dy \in \DY$. Then the same holds for the function $t \mapsto \dy W(t)$ and $$\label{eq:10.8} S(t) = S_0(t) + \int_0^t W(d\tau) \cdot V_0(t-\tau)$$ defines a twin semigroup with generator $C$ given by . [*Proof of Theorem $\ref{thm:10.1}$.*]{} We follow the lines of the proof of Theorem \[thm:5.1\], but adapt the details to the somewhat different situation. In order to show that $$(\dy,y) \mapsto \int_0^t \dy W_0(t-\tau) \cdot V(d\tau)y$$ defines a twin operator, we have to verify - for given $y \in Y$, the linear functional on $\DY$ defined by $$\dy \mapsto \int_0^t \dy W_0(t-\tau) \cdot V(d\tau)y$$ is represented by an element of $Y$ and - for given $\dy \in \DY$, the linear functional on $Y$ defined by $$y \mapsto \int_0^t \dy W_0(t-\tau) \cdot V(d\tau)y$$ is represented by an element of $\DY$. To verify (i) we invoke Lemma \[lem:8.2\] and to verify (ii) we invoke Lemma \[lem:8.3\]. So $S(t)$ defined by is a twin operator and we proceed by verifying properties (i)–(iv) of Definition \[def:2.1\]. We start with the semigroup property (i). Clearly $S(0) = S_0(0) = I$. To verify the semigroup property, we first derive , i.e., we show that $\bigl\{V(t)\bigr\}$ is a cumulative output family [@DGT93] for the semigroup $\bigl\{S(t)\bigr\}$. \[lem:10.3\] We have $$\label{eq:10.9} V(t+s) - V(t) = V(s)S(t).$$ From we deduce that $$\label{eq:10.10} V(t+s)y - V(t)y = f(t,y) + \int_0^s K(d\si)\,\bigl[V(t+s-\si)y - V(t)y\bigr]$$ with $$\begin{aligned} \label{eq:10.11} f(t,y) &= V_0(t+s)y - V_0(t)y + \int_s^{t+s} K(d\si)\,V(t+s-\si)y\nonumber\\ &\qquad\qquad+ K(s)V(t)y - \int_0^t K(d\si)\, V(t-\si)y.\end{aligned}$$ We claim that $$\label{eq:10.12} f(t,y) = V_0(s)S(t)y.$$ If the claim is justified, we can write as $$\label{eq:10.13} U(s)y = V_0(s)S(t)y + \int_0^s K(d\si)\, U(s-\si)y$$ with $$\label{eq:4.19bis} U(s)y := V(t+s)y - V(t)y.$$ Comparing to we conclude that $$\label{eq:4.19} U(s)y = V(s)S(t)y$$ which, on account of , amounts to . To verify the claim, we first rewrite as $$f(t,y) = V_0(s)S_0(t)y + \int_0^t \bigl[K(s+d\tau)-K(d\tau)\bigr]\,V(t-\tau)y + K(s)V(t)y$$ and observe that we need to show that $$V_0(s)\bigl[S(t)y - S_0(t)y\bigr] = \int_0^t \bigl[K(s+d\tau)-K(d\tau)\bigr]V(t-\tau)y + K(s)V(t)y.$$ The left hand side equals $$\begin{aligned} &\DQ\gb{S_0(s)-I}\int_0^t S_0(t-\tau)q V(d\tau)y\\ &\qquad\qquad\qquad= \int_0^t \DQ\gb{S_0(t+s-\tau) - I - S_0(t-\tau) + I}q\,V(d\tau)y\\ &\qquad\qquad\qquad= \int_0^t \bigl[K(t+s-\tau) - K(t-\tau)\bigr]\,V(d\tau)y.\end{aligned}$$ Partial integration shows that this is equal to the right hand side. To prove the exponential estimate (ii) for $\dy S(t)y$, first note that the exponential estimates for $\dy S_0(t)y$ directly yield exponential estimates for $V_0(t)y$, $\dy W_0(t)$ and $K(t)$. The exponential estimate for $R(t)$ follows from Theorem \[thm:resolvent\] and the exponential estimate for $K(t)$. Using the exponential estimate for $V(t)$ follows from the exponential estimates for $V_0(t)y$ and $R(t)$ together with Theorem \[thm:convo-meas-prop\]. The exponential estimate for $\dy S(t)y$ now follows from the exponential estimates for $\dy W_0(t)$ and $V(t)$ and again Theorem \[thm:convo-meas-prop\]. The proof of (iii) that $t \mapsto \dy S(t)y$ is measurable follows from the measurability of both $t \mapsto \dy S_0(t)y$ and, using Theorem \[thm:convo-meas-Borel-fun\], $$t \mapsto \int_0^t \dy W_0(t-\tau) \cdot V(d\tau)y.$$ It remains to prove (iv). In order to compute the Laplace transform, we use the notation $$\label{eq:10.15} \what{K}(\la) := \int_0^\infty e^{-\la \tau}\,K(d\tau)$$ and note that, since $K(0) = 0$, we have $$\label{eq:10.16} \what{K}(\la) = \la \wbar{K}(\la).$$ The relation $$\label{eq:10.17} \wbar{V}(\la) = \gb{I - \what{K}(\la)}^{-1}\wbar{V}_0(\la)$$ can either be derived from or from in combination with . From we deduce that $$\label{eq:10.18} \wbar{V}_0(\la) = \DQ\gb{\wbar{S}_0(\la) - \la^{-1} I}$$ and from that $$\label{eq:10.19} \what{K}(\la) = \la \wbar{K}(\la) = \la \DQ \wbar{S}_0(\la)q - \pa{\DQ}{q}.$$ Laplace transformation of yields, when using , $$\label{eq:10.20} \wbar{S}(\la) = \wbar{S}_0(\la) + \la \wbar{S}_0(\la) q \wbar{V}(\la).$$ Combining , , and we arrive at $$\begin{aligned} \label{eq:10.21} \wbar{S}(\la) &= \wbar{S}_0(\la) + \la \wbar{S}_0(\la)q\gb{I + \pa{\DQ}{q} - \la \DQ \wbar{S}_0(\la)q}^{-1}\nonumber\\ &\qquad\qquad\times \DQ\gb{\wbar{S}_0(\la) - \la^{-1} I}.\end{aligned}$$ We claim that the right hand side of is equal to $(\la I - C)^{-1}$ when $C$ is as defined in . To substantiate the claim, we rewrite the equation $$(\la I - C)z = y$$ in the form $$(\la I - C_0)z = y + \pa{\DQ}{C_0z}q.$$ It follows that $$z = (\la I - C_0)^{-1}y + \pa{\DQ}{C_0z}(\la I - C_0)^{-1}q$$ and hence that $\pa{\DQ}{C_0z}$ should satisfy the equation $$\begin{aligned} \pa{\DQ}{C_0z} &= \pa{\DQ}{C_0(\la I - C_0)^{-1}y}\\ &\qquad\qquad + \pa{\DQ}{C_0z}\pa{\DQ}{C_0(\la I - C_0)^{-1}q}.\end{aligned}$$ So necessarily $$\begin{aligned} \pa{\DQ}{C_0z} &= \gb{ I - \pa{\DQ}{C_0(\la I - C_0)^{-1}q}}^{-1}\\ &\qquad\qquad \times \pa{\DQ}{C_0(\la I - C_0)^{-1}y}.\end{aligned}$$ Using $$C_0(\la I - C_0)^{-1} = \la(\la I - C_0)^{-1} - I$$ we find that $$\begin{aligned} \label{eq:10.21a} z &= (\la I - C)^{-1}y\nonumber\\ &= (\la I - C_0)^{-1}y + \bigl[ I + \pa{\DQ}{q} - \la \pa{\DQ}{(\la I - C_0)^{-1}q}\bigr]^{-1}\nonumber\\ &\qquad\qquad\quad \times \la (\la I - C_0)^{-1}q\pa{\DQ}{(\la I - C_0)^{-1}y - \la^{-1}y}\nonumber\\ &= (\la I - C_0)^{-1}y + \bigl[ I - \what K(\la)\bigr]^{-1}\nonumber\\ &\qquad\qquad\quad \times \la (\la I - C_0)^{-1}q\pa{\DQ}{(\la I - C_0)^{-1}y - \la^{-1}y}\end{aligned}$$ Since $(\la I - C_0)^{-1} = \wbar{S}_0(\la)$ this is identical to the right hand side of . This completes the proof of the verification of (i)-(iv) of Definition \[def:2.1\]. [*Proof of Theorem $\ref{thm:10.2}$.*]{} We follow the lines of the proof of Theorem \[thm:10.1\], but adapt the details. In order to show that $$(\dy,y) \mapsto \int_0^t \dy W(d\tau) \cdot V_0(t-\tau)y$$ defines a twin operator, we have to verify - for given $y \in Y$, the linear functional on $\DY$ defined by $$\dy \mapsto \int_0^t \dy W(d\tau) \cdot V_0(t-\tau)y$$ is represented by an element of $Y$ and - for given $\dy \in \DY$, the linear functional on $Y$ defined by $$y \mapsto \int_0^t \dy W(d\tau) \cdot V_0(t-\tau)y$$ is represented by an element of $\DY$. To verify (i) we invoke Lemma \[lem:8.1\] and to verify (ii) we invoke Lemma \[lem:8.4\]. So $S(t)$ defined by is a twin operator and the verification of properties (ii)–(iv) of Definition \[def:2.1\] proceeds as in the proof of Theorem \[thm:10.1\]. Exactly as in Section 6, the special representation of the perturbed semigroup $S(t)$ given in respectively and allows us to derive a strong result about the asymptotic behaviour of $S(t)$. \[thm:10.4\] Let $K$ be given by and let $S(t)$ be given by or with generator $C$ given by . Suppose that $S_0(t)$ is bounded and that $(\la I - C_0)^{-1}$ is bounded for $\Re \la \ge 0$. If $$\label{eq:cond-asymp-behav} \inf_{\Re z \ge 0} \bigl | \det \gb{ I - \what K(z)} \bigr| > 0,$$ then $$\label{eq:10.25} \|S(t)C^{-1}\| \to 0\qquad\hbox{as}\quad t \to \infty.$$ As a consequence we have that $S(t)y \to 0$ as $t \to \infty$ for every $y$ in the norm-closure of $\DOM{C}$. We first show that the semigroup $S(t)$ is bounded. From the half-line Gel’fand theorem, see Theorem , it follows that the resolvent $R$ of $K$ belongs to $NBV\gb{[0,\infty); \BR^n}$. Fix $y \in Y$ and $\dy \in \DY$. Since $S_0(t)$ is a bounded semigroup, $t \mapsto V_0(t)y$ and $t \mapsto \dy W_0(t)$ are bounded Borel functions on $[0,\infty)$. Suppose $S(t)$ is given by . From the assumption that $t \mapsto V_0(t)y$ belongs to $NBV_{loc}\gb{[0,\infty); \BR^n}$ and the fact that $t \mapsto V_0(t)y$ is bounded, it follows from Theorem \[thm:convo-meas-fun2\] that $t \mapsto V(t)y$ defined by belongs to $NBV\gb{[0,\infty); \BR^n}$ as well. Therefore, it follows from Theorem \[thm:convo-meas-Borel-fun\] that there exists $M \ge 0$ such that $|\dy S(t)y| \le M\|\dy\|\,\|y\|$. Suppose $S(t)$ is given by . From the assumption that $t \mapsto \dy W_0(t)$ belongs to $NBV_{loc}\gb{[0,\infty); \BR^n}$ and the fact that $t \mapsto \dy W_0(t)$ is bounded, it follows from Theorem \[thm:convo-meas-fun2\] that $t \mapsto \dy W(t)$ defined by belongs to $NBV\gb{[0,\infty); \BR^n}$ as well. Therefore it follows from Theorem \[thm:convo-meas-Borel-fun\] that there exists $M \ge 0$ such that $|\dy S(t)y| \le M\|\dy\|\,\|y\|$ and this proves that $S(t)$ is bounded. Finally, the representation implies that, under the assumptions of the theorem, $(\la I - C)^{-1}$ is bounded, for $\Re \la \ge 0$. This completes the proof that $S(t)$ is bounded and that $\si(C) \mcap i\BR = \emptyset$. So an application of Theorem \[thm:Tauber\] yields the proof. NFDE - Neutral Functional Differential Equations ================================================ Much of our motivation for developing the abstract perturbation theory of Section 10 came from our interest in the NFDE $$\label{eq:11.1} \frac{d}{dt}\bigl[x(t) - \int_{[0,1]} d\eta(\si)x(t-\si)\bigr] = \int_{[0,1]} d\ze(\si)x(t-\si),\qquad t > 0,$$ with initial condition $$\label{eq:11.2} x(\th) = \ph(\th),\qquad -1 \le \th \le 0.$$ Here both $\eta$ and $\ze$ belong to $NBV\gb{[0,1],\BR^{n \times n}}$ and $\ph \in B\gb{[-1,0],\BR^n}$. So we work with the norming dual pair $$\label{eq:11.3} Y = B\gb{[-1,0],\BR^n},\qquad \DY = NBV\gb{[0,1],\BR^{n}}$$ with pairing $$\label{eq:11.4} \pa{\dy}{y} = \int_{[0,1]} d\dy(\si) \cdot y(-\si).$$ The rows of both $\ze$ and $\eta$ are considered as elements of $\DY$. Concerning $\eta$ we additionally assume that $$\label{eq:11.5} \eta\hbox{ is continuous at zero},$$ the idea being that we normalize the jump at zero and write its contribution separately as the term $x(t)$ at the left hand side of . The special case that $\eta$ in is identically zero was considered in Section 4. Here we take the twin semigroup constructed in that section as our starting point. In order to stay in line with the framework of Section 10, we add an index zero when referring to this “unperturbed" semigroup $S_0(t)$: $$\label{eq:11.6} S_0(t)\hbox{ is defined by the right hand side of {\eqref{eq:6.18}}}$$ with generator $$\begin{aligned} \label{eq:11.7} \DOM{C_0} &= \hbox{Lip}\gb{[-1,0],\BR^n}\nonumber\\ C_0\ph &= \bigl\{ \ph' \in Y \mid \ph(\th) = \ph(-1) + \int_{-1}^\th \ph'(\si)\,d\si,\quad \ph'(0) = \pa{\ze}{\ph}\,\bigr\}.\end{aligned}$$ Equivalently $$\label{eq:11.8} S_0(t)\ph = z(t+\novar;\ph),$$ where $z$ is the unique solution of – with $\eta = 0$. As in we define, for $i=1,\ldots,n$, the elements $q_i \in Y$ by $$\label{eq:11.9} q_i(\th) = \begin{cases} 0 &\th < 0\\ e_i &\th = 0, \end{cases}$$ where $e_i$ is the $i$-th unit vector in $\BR^n$. The elements $Q_i^\diamond$ of $\DY$ are defined by $$\label{eq:11.10} Q_i^\diamond(\th) = \eta_i(\th),$$ where $\eta_i$ is the $i$-th row of the matrix valued function $\eta$. The aim of the present section is to show that, with these definitions, the twin semigroup $\bigl\{S(t)\bigr\}$ defined in Theorem \[thm:10.2\] is exactly the semigroup of solution operators of –. At a formal level this is immediate: if we rewrite as $$\dot x(t) = \int_{[0,1]} d\eta(\si)\dot x(t-\si) + \int_{[0,1]} d\ze(\si)x(t-\si)$$ and proceed as in the formal derivation of from , we obtain $$\frac{du}{dt} \in C_0u + q \cdot \pa{\DQ}{C_0u}$$ by making the crucial observation that, on account of and , the value of $\gb{C_0u}(0)$ is irrelevant when evaluating the second term at the right hand side. The rigorous proof of the general case, presented below, involves an unpleasant amount of formula manipulation. We therefore first present the proof for the relatively simple situation that the kernel $\ze$ in is identically zero. In that case we have $$\label{eq:11.11} \gb{S_0(t)y}(\th) = y(t+\th),$$ where by definition $y(t) = y(0)$ for $t \ge 0$. Hence $$\begin{aligned} \label{eq:11.12} V_0(t)y &= \DQ\gb{S_0(t) - I}y\nonumber\\ &= \int_{[0,1]} d\eta(\si)\bigl[y(t-\si) - y(-\si)\bigr]\nonumber\\ &= \int_{(t,1]} d\eta(\si)y(t-\si) + \eta(t)y(0) - \int_{[0,1]} d\eta(\si)y(-\si).\end{aligned}$$ It follows that $$\label{eq:11.13} K(t) = V_0(t)q = \eta(t).$$ Moreover $$\begin{aligned} \label{eq:11.14} \dy W_0(t) = \dy S_0(t)q &= \int_{[0,1]} d\dy(\si)q(t-\si)\nonumber\\ &= \int_{[0,t]} d\dy(\si) = \dy(t)\end{aligned}$$ (where now $\dy(t) = \dy(1)$ for $t \ge 1$, by definition) and accordingly $$\label{eq:11.15} \int_0^t \dy W_0(d\tau) \cdot V(t-\tau)y = \int_0^t d\dy(d\tau) \cdot V(t-\tau)y.$$ Let, for $\th \in [0,1]$, $$\label{eq:11.16} y^\diamond_\th(\si) = \begin{cases} 0, &0 \le \si < \th;\\ (1,1,\ldots,1)^T, &\th \le \si \le 1. \end{cases}$$ The identity $$y_\th^\diamond S(t)y = y_\th^\diamond S_0(t)y + \int_0^t y_\th^\diamond W_0(d\tau) \cdot V(t-\tau)y$$ reads $$\begin{aligned} \label{eq:11.17} \gb{S(t)y}(-\th) &= \gb{S_0(t)y}(-\th) + \begin{cases} 0 &\hbox{if } \th \not\in [0,t]\\ V(t-\th)y &\hbox{if } \th \in [0,t] \end{cases}\nonumber\\ &= y(t-\th) + V(t-\th)y\, \chi_{[0,t]}(\th).\end{aligned}$$ So in order to establish that the twin semigroup of Theorem \[thm:10.2\] is indeed the semigroup of solution operators of , for the special case $\ze = 0$, we need to verify that $$\label{eq:11.18} V(t)y = x(t) - y(0).$$ By elementary operations one derives from the equation $$\label{eq:11.19} x(t) - y(0) = \int_{[0,t)} d\eta(\si)\bigl[x(t-\si) - y(0)\bigr] + V_0(t)y$$ with $V_0(t)y$ as specified in . From , and it follows, by uniqueness, that holds. This completes the proof in the special case when $\zeta = 0$. For the general case we have to replace by $$\label{eq:11.20} \gb{S_0(t)y}(\th) = z(t+\th),$$ with $z$ the solution of , as given in in terms of the resolvent $\rho$ of $\ze$ and $f$ defined by . So is replaced by $$\label{eq:11.21} V_0(t)y = \gb{\eta \star z}(t) + g(t)$$ with $$\label{eq:11.21a} g(t) = \int_{(t,1]} d\eta(\si)y(t-\si) - \int_{[0,1]} d\eta(\si)y(-\si).$$ For $y=q$ we have $f = I$ and hence for $t \ge 0$ $$z(t) = I + \int_0^t \rho(\tau)\,d\tau.$$ Since $g(0) = 0$ for $y=q$ we find $$\begin{aligned} \label{eq:11.22} K(t) = V_0(t)q &= \int_{[0,t]} d\eta(\si)\bigl[I + \int_0^{t-\si} \rho(\tau)\,d\tau\bigr]\nonumber\\ &= \eta(t) + \int_0^t \eta(\si) \rho(t-\si)\,d\si.\end{aligned}$$ For $t > 0$ $$\begin{aligned} \label{eq:11.23} \dy W_0(t) = \dy S_0(t)q &= \int_{[0,1]} d\dy(\si)\bigl[I + \int_0^{t-\si} \rho(\tau)\,d\tau\bigr]\,\chi_{\si \le t}(\si)\nonumber\\ &= \dy(t) + \int_0^t \dy(\si) \rho(t-\si)\,d\si\end{aligned}$$ (so note that $t \mapsto \dy W_0(t)$ is actually continuous in $t=0$!) and accordingly $$\begin{aligned} \label{eq:11.24} &\int_0^t \dy W_0(d\tau) \cdot V(t-\tau)y\nonumber\\ &\qquad= \int_0^t \dy(d\tau) \cdot V(t-\tau)y + \int_0^t d_\tau\bigl[\int_0^\tau \dy(\si)\rho(\tau-\si)\,d\si\bigr] \cdot V(t-\tau)y\nonumber\\ &\qquad= \int_0^t \dy(d\tau) \cdot \bigl[ V(t-\tau)y + \int_0^{t-\tau} \rho(\th)V(t-\tau-\th)y\,d\th\bigr]\end{aligned}$$ (where in the last step we have used integration by parts). So, repeating the argument embodied in and , we find that the general version of reads $$\label{eq:11.25} V(t)y + \int_0^t \rho(\th)V(t-\th)\,d\th = x(t) - z(t).$$ We rewrite – as $$x = \eta \star x + \ze \ast x + f +g$$ with $f$ given by and $g$ by . Subtracting $$z = \ze \ast z + f$$ we obtain (using in the second step) $$\begin{aligned} x - z &= \eta \star x + \ze \ast (x-z) + g\\ &= \eta \star (x-z) + \ze \ast (x-z) + V_0.\end{aligned}$$ Applying $\rho \ast$ to both sides and using we find $$\ze \ast (x-z) = \rho \ast \eta \star (x-z) + \rho \ast V_0$$ and accordingly we can rewrite the equation for $x-z$ in the form $$\label{eq:11.27} x -z = \eta \star (x-z) + \rho \ast \eta \star (x-z) + V_0 + \rho \ast V_0.$$ The general equation amounts, when $K$ is given by , to $$\begin{aligned} V &= \eta \star V + \eta \star \rho \ast V + V_0\\ &= \eta \star (V + \rho \ast V) + V_0.\end{aligned}$$ So $$\rho \ast V = \rho \ast \eta \star V + \rho \ast \eta \star \rho \ast V + \rho \ast V_0$$ and $$\label{eq:11.28} V + \rho \ast V = \eta \star (V + \rho \ast V) + \rho \ast \eta \star (V + \rho \ast V) + V_0 +\rho \ast V_0.$$ Comparing and we deduce from the uniqueness of a solution that holds. We summarize our conclusions as \[thm:11.1\] The semigroup of solution operators of –, with the assumption , is identical to the twin semigroup of Theorem $\ref{thm:10.2}$ when the specifications , , ${\eqref{eq:11.6}}\slash{\eqref{eq:11.8}}$,,, and are made. Motivated by Theorem \[thm:10.4\] we add a result about the asymptotic behaviour for $t \to \infty$. \[thm:11.2\] Suppose that $\eta$ has no singular part $see {\eqref{eq:decomp-meas}}$. The semigroup of solution operators of – restricted to $C\gb{[-1,0];\BR^n}$ is asymptotically stable if the following two conditions are satisfied 1. $\det \bigl[ zI - \int_0^1 e^{-z\si} d\ze(\si) \bigr] \not= 0$ for $\Re z \ge 0$; 2. $\inf_{\Re z \ge 0}\ \bigl|\det \bigl[ I - \int_0^1 e^{-z\si} d\eta(\si) \bigr] \bigr| > 0$. The first condition and Theorem \[thm:5.5\] imply that the unperturbed semigroup $\{S_0(t)\}$ is bounded. Furthermore, in the present setting $K$ = $\eta$ and the second condition is equivalent to . Since in the present setting the norm closure of $\DOM{C}$ equals $C\gb{[-1,0];\BR^n}$, the result follows from an application of Theorem \[thm:10.4\]. In contrast to RFDE, general NFDE do not have smoothing properties and it is a delicate question whether the semigroup of solution operators of – is asymptotically stable under the assumptions of Theorem \[thm:11.2\] on $B\gb{[-1,0];\BR^n}$. The solutions of NFDE with an absolutely continuous measure $\ze$ do become continuous and Theorem \[thm:11.2\] can be used to study the asymptotic stability of the semigroup on $B\gb{[-1,0];\BR^n}$. As an illustration, consider the following NFDE $$\label{eq:nfde-example} \frac{d}{dt}\bigl[x(t) - \int_0^1 a(s)x(t-s)\,ds\bigr] = -c x(t)$$ with $c > 0$ and $\int_0^1 |a(s)|\,ds < 1$. Note that $c > 0$ implies that the first condition in Theorem \[thm:11.2\] is satisfied and that $\int_0^1 |a(s)|\,ds < 1$ implies that the second condition in Theorem \[thm:11.2\] is satisfied as well. Therefore, the zero solution of is asymptotically stable. RE - Renewal equations with BV kernels ====================================== If a cell divides into two daugther cells after a cell cycle of fixed length, and we take this length as the unit of time, we may replace by $$\label{eq:12.1} b(t) = 2b(t-1)$$ when mortality is negligible. More generally we may consider $$\label{eq:12.2} b(t) = \int_0^1 L(da) b(t-a),$$ where the model ingredient $L$ specifies the age-specific expected [*cumulative*]{} number of offspring. The assumption $$\label{eq:12.3} L \hbox{ is continuous in } a = 0$$ reflects that instantaneous reproduction by a newborn individual is impossible. It turns out to be useful to extend the domain of definition of $L$ via $$\label{eq:12.15} L(a) = L(1)\qquad \hbox{for } a \ge 1.$$ In terms of $B$ defined by (cf. ) $$\label{eq:12.4} B(t) = \int_0^t b(\tau)\,d\tau,\qquad t > 0,$$ we can write as the neutral delay differential equation $$\label{eq:12.5} B'(t) = \int_0^1 L(da) B'(t-a)$$ Provided $B'$ is of bounded variation and $B_t'$ is normalized to be zero in zero, we can rewrite as $$\label{eq:12.6} B'(t) = \int_{-1}^0 \gb{L(-\si) - L(1)} B_t'(d\si).$$ To verify this transformation, we fix $t$ and show that we can interpret and as convolution of measures. Note that $L(-\cdot) - L(1) \in NBV\gb{\BR}$ and so by Theorem \[thm:bv2\] there exists a measure $\mu$ such that $$\label{eq:12.6a} L(a) - L(1) = \mu\gb{(-\infty,a]}.$$ Also note that we can normalize $B'$ such that $B'(a) = 0$ for $a \ge t$ and again by Theorem \[thm:bv2\] there exists a measure $\nu$ such that $$\label{eq:12.6b} B'(a) = \nu\gb{(-\infty,a]}.$$ An application of now yields $$\begin{aligned} B'(t) = \int_0^1 L(da)\,B'(t-a) &= \int_\BR L(da)\, B'(t-a)\\ &= \int_\BR \mu(da)\, \nu\gb{(-\infty,t-a]}\\ &= \int_\BR \mu\gb{(-\infty,t-a]}\,\nu(da)\quad(\mbox{by {\eqref{eq:convo-meas-ident}}})\\ &= \int_\BR \gb{L(t-a) - L(1)}\, B'(da)\\ &= \int_\BR \gb{L(-\si) - L(1)}\, B_t'(d\si)\quad(\mbox{with } t-a = -\si)\\ &= \int_{-1}^0 \gb{L(-\si) - L(1)}\, B_t'(d\si)\end{aligned}$$ and this shows that the equations and are equivalent. Now recall the definition of $Y$, $\DY$ and $C_0$ in , and , respectively. It appears that the right hand side of can be written as $$\pa{\DQ}{C_0B_t}$$ when we define $$\label{eq:12.7} \DQ(\th) = L(\th) - L(1),\qquad 0 \le \th \le 1.$$ Thus we are led to believe that the theory of Section 10 applies to equation . The aim of this section is to show that this is indeed the case by elaborating the details. We supplement by the initial condition $$\label{eq:12.8} B(\th) = \psi(\th),\qquad -1 \le \th \le 0$$ with $\psi \in Y$, so in particular $B(0) = \psi(0) = 0$. Integrating both sides of with respect to time from $0$ to $t$, we obtain first $$\label{eq:12.9} B(t) = \int_0^1 L(da)\bigl[B(t-a) - B(-a)\bigr]$$ and next, using , $$\label{eq:12.10} B(t) = \int_{[0,t]} L(da)B(t-a) + f(t)$$ with $$\label{eq:12.11} f(t) := \int_{(t,1]} L(da)\psi(t-a) - \int_{[0,1]} L(da)\psi(-a).$$ The [*resolvent*]{} $R$ of $L$ is the solution of (cf. Theorem \[thm:resolvent\]) $$\label{eq:12.12} R(a) = \int_{[0,a]} L(a-\si) R(d\si) + L(a)$$ which is consistent with $R(0)=0$ and shows that $R$, just like $L$ (recall ), is continuous from the right in $a=0$. Because of this property of both $R$ and $L$ we have $$\int_{[0,a]} L(a-\si)R(d\si) = \int_{[0,a]} L(d\si)R(a-\si).$$ We also note that in general, i.e., even for systems, so for functions taking values in $\BR^n$, $$\int_{[0,a]} L(a-\si)R(d\si) = \int_{[0,a]} R(d\si)L(a-\si)$$ whenever $R$ is the resolvent of $L$, cf. Theorem \[thm:resolvent\]. According to Theorem \[thm:renewal\] the solution of is given by $$\label{eq:12.13} B(t) = f(t) + \int_{[0,t]} R(da)f(t-a).$$ Starting from the initial condition we thus provided a constructive definition of $B(t)$ for $t > 0$. Clearly the definition of the operators $S(t)$ in extends to the current situation. We want to identify these operators with the semigroup of Theorem \[thm:10.1\] when $Y$, $\DY$, $\{S_0(t)\}$, $C_0$, $q$ and $\DQ$ are given by, respectively, , , , , and . In a family of maps $V_0(t) : Y \to \BR$ was defined by $$\label{eq:12.14} V_0(t)\psi = \DQ\gb{S_0(t)-I}\psi.$$ Our first step will be to spell out the right hand side for the current situation. \[lem:12.1\] Let $V_0$ be defined by , then $$\label{eq:12.16} V_0(t)\psi = \int_{-1}^0 \psi(d\th)\,\bigl[L(t-\th) - L(-\th)\bigr].$$ Since $$\begin{aligned} \DQ\gb{S_0(t)-I}\psi &= \int_{-1}^{-t} d_\th \psi(t+\th)\bigl[ L(-\th) - L(1)\bigr]\\ &\qquad\qquad -\int_{-1}^0 \psi(d\th)\bigl[L(-\th) - L(1)\bigr],\end{aligned}$$ the claim follows from $$\begin{aligned} \int_{-1}^{-t} d_\th \psi(t+\th)\bigl[ L(-\th) - L(1)\bigr] &= \int_{t-1}^0 \psi(d\si)\bigl[ L(t-\th) - L(1)\bigr]\\ &= \int_{-1}^0 \psi(d\th)\,\bigl[L(t-\th) - L(1)\bigr]\end{aligned}$$ (where in the last step we used ). \[col:12.2\] For the kernel $K = V_0(\novar)q$, cf. , we find $$\label{eq:12.17} K(t) = L(t)$$ \[lem:12.3\] Let $f$ be defined by , then $$\label{eq:12.18} f(t) = V_0(t)\psi.$$ Extending $\psi$ by zero for positive arguments, we can write as $$f(t) = \int_0^1 L(da)\,\bigl[\psi(t-a) - \psi(-a)\bigr]$$ and next use partial integration to obtain $$\begin{aligned} f(t) &= L(1)\bigl[\psi(t-1) - \psi(-1)\bigr] + \int_{t-1}^0 L(t-\th)\,\psi(d\th) - \int_{-1}^0 L(-\th)\,\psi(d\th)\\ &= \int_{-1}^0 \bigl[L(t-\th) - L(-\th)\bigr]\,\psi(d\th) = V_0(t)\psi.\end{aligned}$$ \[col:12.4\] Let $B$ be defined by then, by comparing to , we find $$\label{eq:12.19} B(t) = V(t)\psi.$$ \[thm:12.5\] Let $\{S(t)\}$ be the twin semigroup defined by , then $$\label{eq:12.20} \gb{S(t)\psi}(\th) = B(t+\th) - B(t)$$ holds for the special case of $Y$, $\DY$, $\{S_0(t)\}$, $C_0$, $q$ and $\DQ$ considered in this section. By pairing with step functions from $\DY$ we deduce from the pointwise definition $$\label{eq:12.21} \gb{S(t)\psi}(\th) = \gb{S_0(t)\psi}(\th) + \int_0^t \gb{S_0(t-\tau)q}(\th) V(d\tau)\psi.$$ For $t+\th \le 0$ we have $$\gb{S_0(t)\psi}(\th) = \psi(t+\th)$$ and $\gb{S_0(t-\tau)q}(\th) = -1$ for $0 \le \tau \le t$. Hence $$\gb{S(t)\psi}(\th) = \psi(t+\th) - V(t)\psi = \psi(t+\th) - B(t).$$ For $t+\th > 0$ we have $\gb{S_0(t)\psi}(\th) = 0$ and $$\gb{S_0(t-\tau)q}(\th) = -1\quad \hbox{for } t+\th < \tau \le t$$ (and zero otherwise), showing that holds. An application of Theorem \[thm:10.4\], note that $S_0(t)$ defined by is bounded and identically zero for $t \ge 1$, yields the following asymptotic stability result. \[thm:12.6\] Suppose that $L$ has no singular part $see {\eqref{eq:decomp-meas}}$. The twin semigroup $\{S(t)\}$ defined by is asymptotically stable if $$\label{eq:12.22} \inf_{\Re z \ge 0} \bigl| \det\gb{I - \int_0^1 e^{-z\tau}L(d\tau)} \bigr| > 0.$$ Discussion ========== When supplemented by an appropriate initial condition, a delay equation has, as a rule, a unique solution. The proof consists of formulating a fixed point problem and verifying the conditions of the contraction mapping theorem. Next a semigroup of solution operators is defined by translation along the constructed solution. In pioneering fundamental work [@Hale71], J.K. Hale developed the qualitative theory of delay equations along the lines of the corresponding theory for ODE, but with due attention for the infinite dimensional character of the state space. The variation-of-constants formula is an essential instrument for building such a theory. This formula involves both the right hand side of the equation (corresponding to the derivative of the point value in zero of the function that describes the current state, taking values in $\BR^n$) and integration. If one wants to work with the Riemann-integral, the state space needs to be such that the semigroup is strongly continuous. If one wants that the right hand side of the equation corresponds to a well-defined operator on the state space, this space needs to be such that point evaluation is well defined and that point values are not constrained by values in nearby points. As explained in the introduction, these requirements are incompatible. So a fundamental difficulty arises. (In our opinion, the challenge arising from this difficulty actually gives the theory of delay equations its charm.) As far as we know, until now state spaces have been chosen such that one can work with the Riemann integral. In [@Hale71] the semigroup is strongly continuous and the difficulty is addressed by introducing the fundamental solution (corresponding to an initial condition that does NOT belong to the state space) and letting the formula define the point values of the function that represents the state. In [@Diek95], first an auxiliary space is introduced. This is in fact a dual space ‘containing’ the fundamental solution. Next one checks that the weak\* Riemann integral defines an element of the original state space. In [@MR09; @MR18] integrated semigroups are used to avoid the need of considering elements that do not belong to the state space. Here we have chosen to work with a state space $Y$ that is ‘big’ enough to contain the fundamental solution. This has two consequences - we lose strong continuity of the semigroups on $Y$ - the dual space $Y^\ast$ does not allow a useful characterization. To overcome these difficulties, we have in a first step singled out an explicitly characterized subspace $\DY$ of the dual space $Y^\ast$ that is both rich enough and not too rich. By this we mean that the combination of $Y$ and $\DY$ forms a norming dual pair, i.e., an element of $Y$ is completely determined by the action of the elements of $\DY$ on it and, vice versa, an element of $\DY$ is completely determined by the action of the elements of $Y$ on it. Integrals of functions in either one of these spaces are next defined by integrating (after requiring measurability) the scalar functions obtained by pairing with elements of the other space. This yields elements of, respectively, $Y^\ast$ and $Y^{\diamond\ast}$ and a priori it is not guaranteed that these are represented by elements of, respectively, $\DY$ and $Y$. To verify that actually they are, we equip both spaces with a second topology, the weak topology generated by the other space. Viewed thus as locally convex spaces, one space is the dual of the other and all we need to check is the continuity of linear functionals with respect to the right topology. This is where the dominated convergence theorem and additional assumptions enter the story. In this paper we developed the relevant linear theory and showed that, with appropriate choice of $Y$ and $\DY$, it covers perturbation theory for both delay differential equations and renewal equations, not only in the retarded, but also in the neutral case. We plan to extend our work in several directions. We are confident that equations with infinite delay can be dealt with in the spirit of [@DiekGyll12] and that the proofs in [@Diek95] of the Principle of Linearized Stability, the Centre Manifold Theorem et cetera, generalize, mutatis mutandae, to the nonlinear version of the present setting. But this has to be checked, with special attention for the neutral case. For Renewal Equations it is not yet entirely clear what exactly qualifies as ‘the nonlinear version of the present setting’. And, on top of that, in population dynamical models with individuals characterized by a multi-dimensional variable (e.g., age and size) that can assume a continuum of birth values, we have to deal with an infinite dimensional Renewal Equation. Ideally, we connect the modelling and bookkeeping approach of [@DGMT98; @DGHKMT01] to our nonlinear extension. Renewal equations and their resolvents ====================================== Let $\SB$ denote the Borel $\si$-algebra on $\BR$. A complex Borel measure is a map $\mu : \SB \to \BC$ such that $\mu(\emptyset) = 0$ and if $\{E_j\}$ is a sequence of disjoint sets in $\SB$, then $$\mu\gb{\Mcup_{j=1}^\infty E_j} = \sum_{j=1}^\infty \mu(E_j),$$ where the series converges absolutely. The total variation measure ${{\vert \mu \vert}}$ of a complex Borel measure $\mu$ is given by $$\label{eq:bv3} {{\vert \mu \vert}}(E) = \sup\big\{\sum_{j=0}^n {{\vert \mu(E_j) \vert}} \mid n \in \BN,\ \{E_j\} \hbox{ a partition of } E \hbox{ in } \SB\,\big\}.$$ The vector space of complex Borel measures $\mu$ on $\BR$ of bounded total variation is denoted by $M\gb{\BR}$. Provided with the total variation norm given by $$\label{eq:norm-TV2} \|\mu\|_{TV} = {{\vert \mu \vert}}\gb{\BR},$$ the vector space $M\gb{\BR}$ becomes a Banach space. We embed $L^1\gb{\BR}$ into $M\gb{\BR}$ by identifying $f \in L^1\gb{\BR}$ with the measure $\mu$ defined by $\mu(E) = \int_E f(x)\,dx$ or, in short, $\mu(dx) = f(x)dx$. We define , for $-\infty < a < b < \infty$, the Banach space $M\gb{[a,b]}$ by restriction. Let $f : \BR \to \BC$. For $t \in \BR$, we define $T_f : \BR \to [0,\infty]$ by $$\label{eq:bv1} T_f(t) := \sup \sum_{j=1}^n {{\vert f(t_j)-f(t_{j-1}) \vert}},$$ where the supremum is taken over $n \in \BN$ and all partitions $-\infty < t_0 < t_1 < \cdots < t_n = t$ of $(-\infty,t]$. The extended real function $T_f$ is called the *total variation function* of $f$. Note that if $a < b$, then $T_f(b) - T_f(a) \ge 0$ and hence $T_f$ is an increasing function. If $\lim_{t \to \infty} T_f(t)$ is finite, then we call $f$ a function of bounded variation. We denote the space of all such functions by $BV$. The space $NBV(\BR)$ of normalized functions of bounded variation is defined by $$NBV(\BR) = \{f \in BV \mid f \hbox{ is continuous from the right and } f(-\infty) = 0\,\}.$$ Provided with the norm $$\label{eq:bv2} \|f\|_{TV} := \lim_{t \to \infty} T_f(t)$$ the space $NBV(\BR)$ becomes a Banach space. More generally, we define for $-\infty < a < b < \infty$, the vector space $NBV\gb{[a,b]}$ to be the space of functions $f : \BR \to \BC$ such that $f(t) = f(a) = 0$ for $t \le a$, $f$ is continuous from the right on the open interval $(a,b)$, $f(t) = f(b)$ for $t \ge b$, and whose total variation on $[a,b]$, given by $T_f(b)-T_f(a) = T_f(b)$, is finite. Provided with the norm given by , the space $NBV\gb{[a,b]}$ becomes a Banach space. The following fundamental result, see [@Fol90 Theorem 3.29] provides the correspondence between functions of bounded variation and complex Borel measures. \[thm:bv2\] Let $\mu$ be a complex Borel measure on $\BR$. If $f : \BR \to \BC$ is defined by $f(t) = \mu((-\infty,t])$, then $f \in NBV(\BR)$. Conversely, if $f \in NBV(\BR)$ is given, then there is a unique complex Borel measure $\mu_f$ such that $\mu_f((-\infty,t]) = f(t)$. Moreover ${{\vert \mu_f \vert}} = \mu_{T_f}$. Given a function $f \in NBV\gb{[a,b]}$ with corresponding measure $\mu_f$, we define the Lebesgue-Stieltjes integral $\int g\,df$ or $\int g(x)\,f(dx)$ to be $\int g\,d\mu_f$. Thus, a Lebesgue-Stieltjes integral is a special Lebesgue integral and the theory for the Lebesgue integral applies to the Lebesgue-Stieltjes integral. In this appendix we collect some results about the convolution of a measure and a function and the convolution of two measures needed to study renewal equations. For details and further results we refer to [@Fol90; @GLS90]. Let $B\gb{\BR}$ denote the vector space of all bounded, Borel measurable functions. Provided with the supremum norm (denoted by $\|\cdot\|$), the space $B\gb{\BR}$ becomes a Banach space. We define $B\gb{[a,b]}$ by restriction, to functions that vanish outside $[a,b]$. The convolution $\mu \bast f$ of a measure $\mu \in M(\BR)$ and a Borel measurable function $f$ is the function $$\label{eq:convo-meas-fun} (\mu \bast f) (t) = \int_{\BR} \mu(ds)f(t-s)$$ defined for those values of $t$ for which the function $s \mapsto f(t-s)$ is $|\mu|$-integrable. The following result can be found in [@GLS90 Theorem 3.6.1(ii)]. \[thm:convo-meas-Borel-fun\] If $f \in B\gb{\BR}$ and $\mu \in M\gb{\BR}$, then the convolution $\mu \bast f \in B\gb{\BR}$ and $$\|\mu \bast f \| \le \|\mu\|_{TV}\|f\|.$$ The convolution $\mu \ast \nu$ of two measures $\mu,\nu \in M\gb{\BR}$ is the completion of the measure that to each Borel set $E \subset \BR$ assigns the value $$\label{eq:convo-meas} (\mu \ast \nu) (E) = \int_{\BR} \mu(ds)\nu(E-s),$$ where $E - s = \{e-s \mid e \in E\}$ (cf. [@GLS90 Definition 4.1.1]). If $\chi_E$ is the characteristic function of the set $E$, then $\si \mapsto \chi_E(\si + s)$ is the characteristic function of $E-s$, and $$\nu(E-s) = \int_\BR \chi_E(\si + s)\nu(d\si).$$ It follows from Theorem \[thm:convo-meas-Borel-fun\] that the function $s \mapsto \nu(E-s)$ belongs to $B\gb{\BR}$ and hence the definition $\mu \ast \nu$ in makes sense. Furthermore, using Fubini’s Theorem, we have the following useful identity $$\begin{aligned} \label{eq:convo-meas-ident} \mu \ast \nu (E) &= \int_{\BR} \mu(ds)\nu(E-s)\nonumber\\ &= \int_{\BR}\int_{\BR} \chi_E(\si+s) \mu(ds)\nu(d\si)\nonumber\\ &= \int_{\BR} \mu(E-s)\nu(ds)\end{aligned}$$ The following result can be found in [@GLS90 Theorem 4.1.2(ii)]. \[thm:convo-meas-prop\] Let $\mu,\nu \in M\gb{\BR}$ and let the convolution $\mu \ast \nu$ be defined by . - The convolution $\mu \ast \nu$ belongs to $M\gb{\BR}$ and $$\|\mu \ast \nu\|_{TV} \le \|\mu\|_{TV}\|\nu\|_{TV}.$$ - For any bounded Borel function $h \in B\gb{\BR}$, we have $$\int_\BR h(t) \gb{\mu \ast \nu}(dt) = \int_\BR\int_\BR h(t+s)\,\mu(dt)\nu(ds).$$ Using the one-to-one correspondence between complex Borel measures and functions of bounded variation, see Theorem \[thm:bv2\], we can combine the above results to obtain the following theorem. \[thm:convo-meas-fun2\] If $f \in NBV(\BR)$ and $\mu \in M(\BR)$, then $\mu \bast f \in NBV(\BR)$ and $$\|\mu \bast f\|_{TV} \le \|\mu\|_{TV} \|f\|_{TV}.$$ If $\nu$ is the unique complex Borel measure such that $f(t) = \nu\gb{(-\infty,t]}$ for every $t \in \BR$, then with $E = (-\infty,t]$ $$\begin{aligned} \label{eq:convo-meas-fun2} \mu \bast f(t) &= \int_{\BR} \mu(ds) f(t-s)\nonumber\\ &= \int_{\BR} \mu(ds)\nu\gb{(-\infty,t-s]}\nonumber\\ &= \int_{\BR} \mu(ds)\nu\gb{E-s}\nonumber\\ &= \gb{\mu \ast \nu}(E),\end{aligned}$$ where we have used . Since $\mu \ast \nu \in M\gb{\BR}$, we can use to write $$\mu \bast f(t) = g(t),$$ where $g(t) = \mu \ast \nu\gb{(-\infty,t]}$. According to Theorem \[thm:bv2\], $g$ belongs to $NBV(\BR)$. Finally, the norm estimate follows from Theorem \[thm:convo-meas-prop\](i). We also need the following result concerning convolution on half-lines. \[thm:basic-est\] Let $\mu \in M\gb{[0,\infty)}$ and let $f : [0,\infty) \to \BC$ be a bounded continuous function. - If $f(0) = 0$, then $\mu \bast f$ is a bounded continuous function and $$\|\mu \bast f \| \le \|\mu\|_{TV}\|f\|.$$ - If $\mu$ has no discrete part, then $\mu \bast f$ is a bounded continuous function and $$\|\mu \bast f \| \le \|\mu\|_{TV}\|f\|.$$ Since $\mu$ and $f$ are supported on $[0,\infty)$, it follows from that $$\label{eq:convo-halfline} \gb{\mu \bast f}(t) = \int_{[0,t]} \mu(ds)\, f(t-s).$$ To prove (i), observe first that if $f(0) = 0$, then we can extend $f$ to a continuous function on $\BR$ by defining $f(t) = 0$ for $t < 0$. Therefore, we have $$\gb{\mu \bast f}(t) = \int_{\BR} \mu(ds)f(t-s).$$ To prove that $\mu \bast f$ is continuous, we fix $t$ and $T$ such that $0 \le t < T < \infty$ and use Theorem \[thm:convo-meas-Borel-fun\]. This yields $$\begin{aligned} \sup_{t \in \BR} \big| \gb{\mu \bast f}(t+h) - \gb{\mu \bast f}(t) \big| &= \sup_{t \in \BR} \big|\mu \bast \gb{\tau_h f - f}(t) \big|\\ &\le \|\mu\|_{TV} \|\tau_h f - f\|,\end{aligned}$$ where $\tau_h f(t) = f(t+h)$ and $h \in \BR$ such that $0 \le t+h \le T$. Since $f$ is a bounded continuous function, we have that $\tau_h f \to f$ uniformly on compact sets as $h$ tends to zero. Thus, in particular $\tau_h f \to f$ on $[0,T]$ as $h$ tends to zero. This shows that $\gb{\mu \bast f}(t)$ is continuous at $t$. Together with the estimate this shows that $\mu \bast f$ is a bounded continuous function and the norm estimate in the supremum norm follows from the corresponding estimate given in Theorem \[thm:convo-meas-Borel-fun\]. To prove (ii), we first write $$\begin{aligned} \label{eq:basic-est} \gb{\mu \bast f}(t) &= \int_{[0,t]} \mu(ds)f(t-s)\nonumber\\ &= \int_{[0,t]} \mu(ds)\gb{f(t-s) - f(0)} + \mu\gb{[0,t]}f(0).\end{aligned}$$ If $g(s) = f(s) - f(0)$, then $g(0) = 0$ and the first term on the right of becomes $$\int_{[0,t]} \mu(ds)\gb{f(t-s) - f(0)} = \int_{[0,t]} \mu(ds)g(t-s)$$ and by the first part it follows that this term is continuous. Since $\mu$ has no discrete part, the function $t \mapsto \mu\gb{[0,t]}f(0)$ is also continuous. This shows that $\mu \bast f$ is a bounded continuous function and the norm estimate follows from the corresponding estimate given in Theorem \[thm:convo-meas-Borel-fun\]. Let $M_{loc}\gb{[0,\infty)}$ be the vector space of measures on $[0,\infty)$ that belong to $M\gb{[a,b]}$ for every $0 \le a < b < \infty$ and let $NBV_{loc}\gb{[0,\infty)}$ be the vector space of functions of bounded variation on $[0,\infty)$ that belong to $NBV\gb{[0,b]}$ for every $0 < b < \infty$. We continue with the existence of the resolvent $\rho$ of a complex Borel measure $\mu$ supported on $[0,\infty)$. See [@GLS90 Theorem 4.1.5]. \[thm:resolvent\] Suppose that $\mu \in M_{loc}\gb{[0,\infty), \BC^{n \times n}}$. There exists a unique measure $\rho \in M_{loc}\gb{[0,\infty), \BC^{n \times n}}$ satisfying either one of the following identities $$\label{eq:resol-eqn-rho} \rho - \mu \ast \rho = \mu = \rho - \rho \ast \mu$$ if and only if $\det\bigl[I - \mu(\{0\})\bigr] \not= 0$. Furthermore, if there exists a positive real $\ga$ such that the measure $e^{-\ga t}\mu(dt)$ is a bounded Borel measure, then there exists $\al$ with $\al \ge \ga$ such that $e^{-\al t}\rho(dt)$ is a bounded Borel measure. Suppose that there exists a measure $\rho$ such that $\rho - \mu \ast \rho = \mu$, then $\gb{\de - \mu} \ast \gb{\de + \rho} = \de$, where $\de$ denotes the point mass at zero. Therefore, $$\bigl[1 - \mu(\{0\})\bigr] \bigl[1+\rho(\{0\})\bigr] = I$$ and hence $\det\bigl[I - \mu(\{0\})\bigr] \not= 0$. Next assume that $\det\bigl[I - \mu(\{0\})\bigr] \not= 0$. We first show that if $\rho \in M_{loc}\gb{[0,\infty)}$ exists such that holds, then it is unique. Indeed, if there exist $\hat\rho$ such that $\hat\rho - \mu \ast \hat\rho = \mu$ , then $$\begin{aligned} \rho = \mu + \rho \ast \mu &= \mu + \rho \ast \gb{ \hat\rho - \mu \ast \hat\rho }\\ &= \mu + \rho \ast \hat\rho - \gb{\rho \ast \mu}*\hat\rho\\ &= \mu + \gb{\rho - \rho \ast \mu}\ast\hat\rho\\ &= \mu + \mu \ast \hat\rho = \hat\rho.\end{aligned}$$ Furthermore, if $\ga \in \BR$ and $\rho \in M_{loc}\gb{[0,\infty), \BC^{n \times n}}$ exists such that holds, then $\hat\rho(dt)= e^{-\ga t}\rho(dt)$ is a solution of $$\label{eq:resol-eqn-rho-scaled} \hat\rho - \hat\mu \ast \hat\rho = \hat\mu = \hat\rho - \hat\rho \ast \hat\mu,$$ where $\hat\mu(dt) = e^{-\ga t}\mu(dt)$. Indeed, this follows from multiplication of by $e^{-\ga t}$ on both sides and using $$\begin{aligned} \mu \ast \rho\gb{(-\infty,t]} &= \int_{\BR}\mu(ds)\,\rho\gb{(-\infty,t-s]}\\ &= \int_{\BR}\mu(ds)\,e^{\ga(t-s)}\hat\rho\gb{(-\infty,t-s]}\\ &= e^{\ga t}\int_{\BR}\hat\mu(ds)\,\hat\rho\gb{(-\infty,t-s]}\\ &= e^{\ga t} \mu \ast \rho\gb{(-\infty,t]}\end{aligned}$$ Because of the uniqueness of the solution $\rho \in M_{loc}\gb{[0,\infty), \BC^{n \times n}}$, it suffices to show that for each $T \in (0,\infty)$ there is a measure $\rho_T \in M\gb{[0,T]}$ satisfying the resolvent equation on $[0,T]$. Fix $T > 0$ and assume at first that $\mu$ has no mass at zero. By replacing $\mu$ by $e^{-\ga t}\mu$, we can assume without loss of generality that $$\label{eq:contraction-estimate} \babs{\mu}\gb{[0,T]} < 1.$$ Using this fact, we have that the map $$\rho \mapsto \mu + \mu \ast \rho$$ defines a contraction on the Banach space $M\gb{[0,T]}$ for every $T > 0$. The Banach contraction principle implies that the solution $\rho$ of is the unique fixed point of this map. Furthermore using the iteration method to approximate the fixed point, we have the following representation for $\rho$ $$\label{eq:rep-resol-rho} \rho = \sum_{j=1}^\infty \mu^{\ast j},$$ where $\mu^{\ast j}$ denotes the $j$-times convolution of $\mu$ with itself. Next assume that $\mu$ has mass at zero and put $A = \mu\gb{\{0\}} \not= 0$. It follows from $ \det \bigl[I -A\bigr] \not = 0$ that we can rewrite the resolvent equation $$\rho - \mu \ast \rho = \mu$$ as $$\label{eq:resol-eqn-rho2} \rho = A(I-A)^{-1}\de + \nu + \nu \ast \rho,$$ where $$\label{eq:loc-nu} \nu = (I-A)^{-1}\gb{\mu - A\de}$$ has no mass at zero. Note from that $\rho\gb{\{0\}} = A(I-A)^{-1}$ and since $\nu$ has no mass at zero, it follows from representation with $\mu = \nu$ that in case $A = \mu\gb{\{0\}} \not= 0$, we have the following representation for $\rho$ $$\label{eq:rep-resol-rho3} \rho = A(I-A)^{-1}\de + \sum_{j=1}^\infty \nu^{\ast j},$$ where $\nu$ is given by . This completes the proof of the theorem. Finally, we prove the exponential estimate for the resolvent by modifying the above contraction argument. If there exists a positive real $\ga$ such that the measure $\hat\mu = e^{-\ga t}\mu(dt)$ is a bounded Borel measure, then we can modify and replace $\mu$ by $\hat\mu(dt) = e^{-\al t}\mu(dt)$ with $\al \ge \ga$ such that $$\label{eq:contraction-estimate-large} \babs{\hat\mu}\gb{[0,\infty)} < 1.$$ so that the map $\hat\rho \mapsto \hat\mu + \hat\mu \ast \hat\rho$ is a contraction in $M\gb{[0,\infty)}$. This shows that there exists a $\hat\rho \in M\gb{[0,\infty)}$ and hence $e^{-\al t}\rho(dt)$ is bounded Borel measure. In order to give the precise asymptotic behaviour of the resolvent $\rho$, i.e., the case that $\al = \ga$ in Theorem \[thm:resolvent\], we have to impose additional conditions on $\mu$, see Theorem \[thm:Gelfand\]. We first need some preparations. From the Radon-Nikodym theorem it follows that we can split a matrix-valued Borel measure $\mu$ on $\BR$ into three parts, the absolutely continuous part, the discrete part, and the singular part: $$\label{eq:decomp-meas} \mu(ds) = b(s)\,ds + \sum_{k=1}^\infty a_k \de_{t_k}(ds) + \mu_s(ds),$$ where $b$ represents the absolutely continuous part of $\mu$, $a_k$ are constant matrices and $\de_{t_k}$ is the unit point mass at $t_k$, and $\mu_s$ denotes the singular part of $\mu$. The Laplace transform $\what\mu : \BC \to \BC^{n \times n}$ of a matrix-valued Borel measure $\mu$ on $\BR$ is the function $$\label{eq:Laplace} \what\mu(\la) = \int_{[0,\infty)} e^{-\la t}\,\mu(dt)$$ defined for those values of $\la \in \BC$ for which the integral converges absolutely. If $\mu \in M_{loc}\gb{[0,\infty),\BC^{n \times n}}$ and $\what\mu(\la_0)$ exists for some $\la_0 \in \BC$, then $\what\mu(\la)$ is defined in the closed half plane $\Re \la \ge \Re \la_0$. Furthermore, if $f \in B\gb{[0,\infty),\BC^n}$, then $$\gb{\mu \bast f}\what\, (\la) = \what\mu(\la)\bar f(\la)$$ for all $\la \in \BC$ for which both $\what\mu(\la)$ and $$\label{eq:Laplace-f} \bar f(\la) = \int_{[0,\infty)} e^{-\la t}f(t)\,dt$$ are defined. The following result, the so-called half-line Gel’fand theorem (see [@GLS90 Theorem 4.4.3 and Corollary 4.4.7]), gives a precise estimate for the growth of the resolvent of $\mu$. \[thm:Gelfand\] Suppose $\mu \in M_{loc}\gb{[0,\infty), \BC^{n \times n}}$ has no singular part and is such that $e^{-\ga t}\mu(dt)$ is a bounded Borel measure. Let $\rho \in M_{loc}\gb{[0,\infty), \BC^{n \times n}}$ denote the unique solution of . If $$\label{eq:Gelfand1} \det \gb{I - \what\mu(z)} \not= 0\qquad\hbox{for } \Re z \ge \ga$$ and $$\label{eq:Gelfand2} \inf_{\Re z \ge \ga} \Big| \det\gb{I - \what\mu_d(z)} \Big| > 0,$$ or combined in one condition $$\label{eq:Gelfand3} \inf_{\Re z \ge \ga} \Big| \det\gb{I - \what\mu(z)} \Big| > 0,$$ then $e^{-\ga t}\rho(dt)$ is a bounded Borel measure. We conclude this appendix summarizing the results developed in this section when applied to the renewal equation $$\label{eq:renewal1} x(t) = \int_{[0,t]} \mu(ds)x(t-s) + f(t),\quad\hbox{for } t \ge 0,$$ for various classes of forcing functions $f$. The following theorem summarizes some relevant results [@GLS90 Theorem 4.1.7]. \[thm:renewal\] Let $\mu \in M_{loc}\gb{[0,\infty),\BC^{n \times n}}$ with $\det\bigl[I - \mu(\{0\}\bigr] \not= 0$. - For every $f \in B_{loc}\gb{[0,\infty),\BC^n}$, the renewal equation has a unique solution $x \in B_{loc}\gb{[0,\infty),\BC^n}$ given by $$x = f + \rho \bast f,$$ where $\rho$ satisfies and is given by . - For every $f \in NBV_{loc}\gb{[0,\infty),\BC^n}$, the renewal equation has a unique solution $x \in NBV_{loc}\gb{[0,\infty),\BC^n}$ given by $$x = f + \rho \bast f,$$ where $\rho$ satisfies and is given by . Furthermore, if $f$ is locally absolutely continuous, then the solution $x$ is locally absolutely continuous as well. - If $f \in C\gb{[0,\infty),\BC^n}$ and $f(0) = 0$, then $x \in C\gb{[0,\infty), \BC^n}$. - Suppose that the kernel $\mu$ has no discrete part. If $f \in C\gb{[0,\infty),\BC^n}$, then $x \in C\gb{[0,\infty),\BC^n}$. Standard arguments show that the solution of the renewal equation is given by $x = f + \rho \bast f$, where $\rho$ denotes the resolvent of $\mu$ given by Theorem \[thm:resolvent\]. So (i) follows from Theorem \[thm:convo-meas-Borel-fun\]. To prove (ii), first note that it follows from Theorem \[thm:convo-meas-fun2\] that $x$ is locally of bounded variation. If $f$ is locally absolutely continuous, then $f$ is the integral of a locally $L^1$-function. Using the representation $x = f + \rho \bast f$ and Fubini’s Theorem, we derive that $x$ is the integral of a locally $L^1$-function as well. Therefore it follows that $x$ is locally absolutely continuous. Furthermore, (iii) follows from Theorem \[thm:basic-est\] (i). Finally, if $\mu, \nu \in M_{loc}\gb{[0,\infty)}$, then the discrete part of $\mu \ast \nu$ is given by the sum $$\label{eq:discrete-part} \gb{\mu \ast \nu}_d = \sum_{k=1}^\infty \sum_{l=1}^\infty p_kq_l\de_{t_k+t_l}.$$ In particular, we conclude that if either $\mu$ or $\nu$ has no discrete part, then the convolution $\mu \ast \nu$ also has no discrete part. In particular, if $\mu$ has no discrete part, then it follows from that the resolvent $\rho$ has no discrete part. Thus (iv) follows from Theorem \[thm:basic-est\] (ii) If the measure $\mu$ has no singular part, then an application of Theorem \[thm:Gelfand\] implies the following corollary. \[col:Gelfand\] Suppose that $\mu \in M\gb{\BR;\BC^{n \times n}}$ has no singular part and satisfies $$\label{eq:Gelfand2} \inf_{\Re z \ge 0} \Big| \det\gb{I - \hat\mu(z)} \Big| > 0.$$ - For every $f \in B\gb{[0,\infty),\BC^n}$, the renewal equation has a unique solution $x \in B\gb{[0,\infty),\BC^n}$ given by $$x = f + \rho \bast f,$$ where $\rho$ satisfies . - For every $f \in NBV\gb{[0,\infty),\BC^n}$, the renewal equation has a unique solution $x \in NBV\gb{[0,\infty),\BC^n}$ given by $$x = f + \rho \bast f,$$ where $\rho$ satisfies . Furthermore, if $f$ is absolutely continuous, then the solution $x$ is absolutely continuous as well. The norming dual pair $\bf{(B,NBV)}$ ==================================== In the study of delay differential equations, the natural dual pair is given by $$\label{eq:mot-ex0} Y = B\gb{[-1,0],\BR^n}\quad\mbox{and}\quad\DY = NBV\gb{[0,1],\BR^n}$$ with the pairing $$\label{eq:mot-ex1} \pa{\dy}{y} = \int_{[0,1]} \dy(d\si) \cdot y(-\si),$$ and in study of renewal equations, the natural dual pair is given by $$\label{eq:mot-ex0r} Y = NBV\gb{[-1,0],\BR^n}\quad\mbox{and}\quad\DY = B\gb{[0,1],\BR^n}$$ with the pairing $$\label{eq:mot-ex1r} \pa{\dy}{y} = \int_{[-1,0]} y(d\si) \cdot \dy(-\si).$$ We first make two trivial, yet useful, observations $$\int_{[0,1]} \dy(d\si) \cdot y(-\si) = y(-\th)$$ if $\dy(\si) = 0$ for $\si < \th$ and $\dy(\si) = {\bf 1}$ for $\si \ge \th$ (where ${\bf 1}$ equals the $n$-vector with all components equal to one), and similarly $$\int_{[-1,0]} y(d\si) \cdot \dy(-\si) = y(-\th)$$ if $\dy(\si) = 0$ for $\si < \th$ and $\dy(\si) = {\bf 1}$ for $\si \ge \th$ The point is that, consequently, convergence in $\gb{Y,\si(Y,\DY)}$ with, respectively, $Y = B\gb{[-1,0],\BR^n}$ or $Y = NBV\gb{[-1,0],\BR^n}$ entails pointwise convergence. In the first case, the dominated convergence theorem implies that this is in fact an *if and only if* characterization. For $Y = NBV\gb{[-1,0],\BR^n}$, this is not so clear. It is true that the pointwise limit of a sequence of functions of bounded variation is again of bounded variation (Helly’s theorem), but there is no dominated convergence theorem for measures. The purpose of this appendix is to show that the dual pairs given by, respectively, and and by and are norming dual pairs suitable for twin perturbation, cf. Definition \[def:4.3\]. \[thm:suit-twin-perI\] The dual pair given by and is a norming dual pair such that and hold, i.e., - $(Y,\si(Y,\DY))$ is sequentially complete; - a linear map $(Y,\si(Y,\DY)) \to \BR$ is continuous if it is sequentially continuous. Before we can prove the theorem we need to present some notions from the theory of Riesz spaces. A Riesz space $Y$ is a real vector space equipped with a lattice structure, i.e., a partial ordering compatible with the vector space structure such that each pair of vectors in $Y$ has a supremum in $Y$. For a given vector $y$ in a Riesz space, the absolute value $|y| \in Y$ is defined by $|y| = \sup\{y, -y\}$. The Banach spaces $Y = B\gb{[-1,0],\BR^n}$ and $Y = NBV\gb{[-1,0],\BR^n}$ are Riesz spaces when the ordering is defined pointwise and componentwise, i.e., $f \ge g$ whenever $P_jf(x) \ge P_jg(x)$ for each $x$ and $1 \le j \le n$ where $P_j : \BR^n \to \BR$ denotes the projection onto the $j^{th}$-coordinate of a $n$-vector. The corresponding absolute value function $|f| : [-1,0] \to \BR^n$ is defined componentwises by $$P_j |f|(x) := \sup \{f_j(x), -f_j(x)\}\quad\hbox{for } x \in [-1,0]\hbox{ and } 1 \le j \le n.$$ A sequence $\{f_n\}$ in a Riesz space $Y$ is order bounded from above if there is a $g \in Y$ such that $f_n \le g$. A sequence $\{f_n\}$ is called decreasing to zero if $n \ge m$ implies $f_m \ge f_n \ge 0$. Furthermore, a sequence $\{f_n\}$ in a Riesz space $Y$ converges in order to $f \in Y$ if there is a sequence $\{g_n\}$ in $Y$ decreasing to zero such that $$|f - f_n| \le g_n$$ For the Riesz spaces $Y = B\gb{[-1,0],\BR^n}$ and $Y = NBV\gb{[-1,0],\BR^n}$ an order bounded sequence $f_n$ converges in order to $f$ if and only if $f_n(x) \to f(x)$ for all $x \in [-1,0]$. A linear functional $\La : Y \to \BR$ on a Riesz space $Y$ is $\si$-order continuous if $\La(f_n) \to 0$ in $\BR$ for every sequence $\{f_n\}$ in $Y$ that converges to zero in order. The vector space of all $\si$-order continuous linear functionals is called the $\si$-order continuous dual of $Y$, cf. [@AliBor06 Definition 8.26]. The following result [@AliBor06 Theorem 14.5] is an essential ingredient of the proof of Theorem \[thm:suit-twin-perI\]. \[thm:sigma-order-dual\] The $\si$–order continuous dual of $B\gb{[-1,0],\BR^n}$ is represented by $NBV\gb{[0,1],\BR^n}$. In the proof we use the fact that the norm dual of a Riesz Banach space is again a Riesz Banach space (cf, [@AliBor06 Theorem 9.27 and Theorem 14.2]). So, in particular, if $\La$ is a bounded linear functional on $B\gb{[-1,0],\BR^n}$, then is has an absolute value ${{\vert \La \vert}}$ in the norm dual of $B\gb{[-1,0],\BR^n}$. Step 1. We first show that if $\La$ is a bounded linear functional on $B\gb{[-1,0],\BR^n}$, then the set function $\mu_\La$ defined by $$\mu_\La(A) = \La(\chi_A)\quad\hbox{ for any Borel set } A$$ is a finitely additive signed measure of bounded variation. Indeed from the linearity of $\La$ it is clear that $\mu_\La$ is a finitely additive real-valued set function. To see that $\mu_\La$ is of bounded variation, let $\{E_1,\ldots,E_n\}$ be a partition of $[-1,0]$, then $$\begin{aligned} \sum_{i=1}^n |\mu_\La(E_i)| &= \sum_{i=1}^n |\La| (\chi_{E_i})\\ &= \babs{\La}\gb{\sum_{i=1}^n \chi_{E_i}} = \babs{\La}({\bf 1}) = \|\La\|\end{aligned}$$ which implies that $\mu_\La$ is of bounded variation. (Here ${\bf 1} \in B\gb{[-1,0],\BR^n}$ denotes the function which is constant one in all components.) As a side remark we mention that the norm dual of $B\gb{[-1,0],\BR^n}$ is actually represented by the Riesz Banach space of all finitely additive signed measures of bounded variation (cf. [@AliBor06 Theorem 14.4]). Step 2. We next show that $\mu_\La$ is a Borel measure if and only if $\La$ is a $\si$-order continuous linear functional. Assume first that $\La$ is $\si$-order continuous and let $\{E_i\}$ be a pairwise disjoint sequence of Borel measurable sets. Put $$E= \mcup_{i=1}^\infty E_i\quad\hbox{and}\quad F_n = \mcup_{i=1}^n E_i$$ and note that $\chi_{F_n}$ converges in order to $\chi_E$. Since $\La$ is $\si$-order continuous, it follows that $$\sum_{i=1}^n \mu_{\La}(E_i) = \La(\chi_{F_n}) \to \La(\chi_E) = \mu_{\La}(E),$$ which shows that $\mu_{\La}$ is $\si$-additive. Conversely, assume that $\mu_\La$ is a complex Borel measure. Let $f_n$ be a sequence decreasing to zero in $B\gb{[-1,0],\BR^n}$ . So $f_n(x)$ decreases to zero pointwise and the Lebesgue dominated convergence theorem implies that $$\La(f_n) = \int_{[-1,0]} f_n\,d\mu_\La \da 0,$$ proving that $\La$ is $\si$-order continuous. [*Proof of Theorem $\ref{thm:suit-twin-perI}$.*]{} We first prove that $(Y,\DY)$ is a norming dual pair. From Theorem \[thm:convo-meas-Borel-fun\] it follows that for every $\dy \in \DY$ and $y \in Y$ $$\label{eq:mot-ex2} \babs{ \int_{[0,1]} \dy(d\si) \cdot y(-\si) } \le \|\dy\|\,\|y\|.$$ By considering step functions for $\dy$, i.e., Dirac point measures by Theorem \[thm:bv2\], we obtain $$\label{eq:mot-ex5} \|y\| = \sup_{x \in [-1,0]} |y(x)| = \sup\bigl\{{{\vert \pa{\dy}{y} \vert}} \mid \dy \in \DY,\ \|\dy\| \le 1 \bigr\}.$$ On the other hand, fix $\dy \in \DY$ and let $\mu = \mu_{\dy}$ be the corresponding Borel measure according to Theorem \[thm:bv2\]. If $\SE$ is a partition of $[-1,0]$ into finitely many, pairwise disjoint, measurable sets, then $$\label{eq:mot-ex6} y_\SE = \sum_{E \in \SE} {\rm sgn}\, \mu(-E)\chi_E$$ is a bounded Borel function on $[-1,0]$ of norm at most one. Furthermore, $$\label{eq:mot-ex7} \pa{\dy}{y_\SE} = \sum_{E \in \SE} {{\vert \mu(-E) \vert}},$$ and taking the supremum over all such partitions $\SE$ of $[-1,0]$ we arrive at $$\label{eq:mot-ex8} \|\dy\| = \sup\bigl\{{{\vert \pa{\dy}{y_\SE} \vert}} \mid \SE \hbox{ a partition of } [-1,0] \bigr\}.$$ This shows that that the pair $(Y,\DY)$ is a norming dual pair. We next prove that $(Y,\si(Y,\DY))$ is sequentially complete. Let $\{y_n\}$ be a Cauchy sequence in $(Y,\si(Y,\DY))$. Since step functions belong to $\DY$ it follows that $\{y_n(x)\}$ is, for every $x \in [-1,0]$, a Cauchy sequence in $\BR$. Since $\BR$ is complete, we have that $$\lim_{n \to \infty} y_n(x) \ \hbox{ exists pointwise for } x \in [-1,0].$$ The pointwise limit of measurable functions is measurable, so it only remains to check the uniform boundedness of the sequence. From the Cauchy property, it follows that the sequence $\{y_n\}$ is bounded in $(Y,\si(Y,\DY))$, i.e., $$\sup_n {{\vert \pa{y}{y_n} \vert}} < \infty\quad\hbox{for any } y \in Y$$ and by considering $\{y_n\}$ as elements of $Y^{\diamond\ast}$, the uniform boundedness principle implies that $$\sup_{n \ge 1} \|y_n\| \ \hbox{ is bounded}.$$ Therefore the sequence $\{y_n\}$ is bounded in the supremum norm and hence the pointwise limit defines a bounded Borel function. This shows that $(Y,\si(Y,\DY))$ is sequentially complete. Finally consider a sequentially continuous linear map $$\La : (Y,\si(Y,\DY)) \to \BR.$$ An application of Theorem \[thm:sigma-order-dual\] shows that in order to prove that $\La$ belongs to $\DY$ it suffices to prove that $\La$ is $\si$-order continuous. But this follows from - $y_n \to 0$ in order implies that $y_n \to 0$ in $\gb{Y,\si(Y,\DY)}$; - the sequential continuity of $\La$ next implies the $\si$-order continuity.   Since reflection is harmless and induces an isometric isomorphism, it follows that $B\gb{[0,1],\BR^n}$ and $NBV\gb{[-1,0],\BR^n}$ form a norming dual pair as well. Next note that $(Y,\DY)$ is a norming dual pair if and only if $(\DY,Y)$ is a norming dual pair. Therefore, we also have the following corollary to Theorem \[thm:suit-twin-perI\]. \[thm:suit-twin-perII\] The dual pair given by and is a norming dual pair such that and hold, i.e., - a linear map $(\DY,\si(\DY,Y)) \to \BR$ is continuous if it is sequentially continuous. - $(\DY,\si(\DY,Y))$ is sequentially complete; Note that if the dual pair is given by and , then the weak topology $\si(Y,\DY)$ on $Y$ is strictly stronger than the weak$^\ast$ topology on $Y$ as can be seen from the fact that $\de_{x_n} \to \de$ in the weak$^\ast$ topology on $Y$ if $x_n \to x$ in $[-1,0]$, but $\de_{x_n} \not\to \de$ in $\si(Y,\DY)$. We end this appendix with some more detailed information about norming dual pairs and their topologies. Given a norming dual pair $(Y,\DY)$, we call a topology $\tau$ on $Y$ [*consistent*]{} (with the duality) if $\DY$ is the dual space of $(Y,\tau)$. By the Mackey-Arens theorem [@AliBor06 Theorem 5.112], a consistent topology $\tau$ is finer than the weak topology $\si(Y,\DY)$ and coarser than the Mackey topology $\tau(Y,\DY)$, the finest topology on $Y$ that preserves the continuous dual. Note that the Mackey topology $\tau(Y,\DY)$ allows the greatest number of continuous functions on $Y$ and all consistent topologies have the same bounded sets [@AliBor06 Theorem 6.30]. Furthermore, if $\DY = Y^\ast$, then the Mackey topology $\tau(Y,Y^\ast)$ on $Y$ corresponds to the norm topology on $Y$, cf. [@AliBor06 Corollary 6.23]. For the dual pair given by and , the topological space $(Y,\tau(Y,\DY))$ has been studied in [@Schaefer89] and in [@BS96], and plays an important role in the theory of Markov processes, cf. [@AliBor06 Chapter 19] and [@Kraaij16]. A topological space $(Y,\tau(Y,\DY))$ is called [*semi-bornological*]{} whenever full and sequential continuity of its linear forms is equivalent, cf. [@Schaefer86 IV.3, p. 131]. The following result [@BS96 Proposition E.2.4] shows that and hold as well with respect to the Mackey topology. \[thm:P12\] Let $$Y = B\gb{[-1,0],\BR^n}\quad\mbox{and}\quad\DY = NBV\gb{[0,1],\BR^n}.$$ The topological space $(Y,\tau(Y,\DY))$ is semi-bornological and $\tau(Y,\DY)$-sequentially complete. In this paper we have formulated our assumptions with respect to the weak topology $\si(Y,\DY)$, but we could have formulated and or and with respect to any consistent topology and hence, in particular, with respect to the Mackey topology $\tau(Y,\DY)$. This yields, strictly speaking, stronger results. But we feel that the formulation in terms of the weak topology is easier to digest by people working with delay equations. [99]{} Aliprantis, C. D. and K.C. Border, *Infinite dimensional analysis. A hitchhiker’s guide. Third edition.* Springer, Berlin, 2006. Arendt, W. and C. J. K. Batty, Tauberian theorems and stability of one-parameter semigroups, *Trans. Amer. Math. Soc.* **306** (1988), 837–852. Arendt, W. and N. Nikolski, Vector-valued holomorphic functions revisited, *Math. Z.* **252**, (2006), 687–689. Batty, C.J.K., *Asymptotic behaviour of semigroups*, Banach Center Publications **30**, Warsaw 1994. Beatty, T.A. and H.H. Schaefer, Semi-bornological spaces, *Math. Z.* **221** (1996), 337–351. Clément, Ph., Diekmann, O., Gyllenberg, M., Heijmans, H. J. A. M., H.R. Thieme, Perturbation theory for dual semigroups. I. The sun-reflexive case, *Math. Ann.* **277** (1987), 709–725. Clément, Ph., Diekmann, O., Gyllenberg, M., Heijmans, H. J. A. M., H.R. Thieme, A Hille-Yosida theorem for a class of weakly\* continuous semigroups, *Semigroup Forum* **38** (1989), 157–178. Crandall, M.G., A gereralized domain for semigroups generators, *Proc. Amer. Math. Soc.* **37** (1973) 434–440. Diekmann, O., Modeling and analysing physiologically structured populations, in: Mathematics Inspired by Biology , Springer LNiM 1714 , V. Capasso & O. Diekmann (eds.) 1999, 1–37. Diekmann, O., and M. Gyllenberg, Abstract delay equations inspired by population dynamics, in: Functional Analysis and Evolution Equations. The Günter Lumer Volume (H. Amann, W. Arendt, M. Hieber, F. Neubrander, S. Nicaise, J. von Below, eds.), Birkhauser 2007, 187–200. Diekmann, O., Gyllenberg, M., and H.R. Thieme, Perturbing semigroups by solving Stieltjes renewal equations, *Differential Integral Equations* **6** (1993), 155–181. Diekmann, O., S.A. van Gils, S.M. Verduyn Lunel and H.O. Walther, *Delay Equations: Functional-, Complex-, and Nonlinear Analysis*, Springer-Verlag, New York, 1995. Diekmann, O., Gyllenberg, M., Metz, J.A.J. , and H.R. Thieme, On the formulation and analysis of general deterministic structured population models. I. Linear theory, *J. Math. Biol.* **36** (1998), 349–388. Diekmann, O., Gyllenberg, M., Huang, H., Kirkilionis, M., Metz, J.A.J. , and H.R. Thieme, On the formulation and analysis of general deterministic structured population models. II. Nonlinear theory, *J. Math. Biol.* **43** (2001), 157–189. Diekmann, O., Getto, P. and M. Gyllenberg, Stability and bifurcation analysis of Volterra functional equations in the light of suns and stars, *SIAM J. Math. Anal.* **39** (2007), 1023–1069. Diekmann, O., and M. Gyllenberg, Equations with infinite delay: blending the abstract and the concrete *J. Differential Equations* **252** (2012), 819–851. Engel, K.-J. and R. Nagel, *One-Parameter Semigroups for Linear Evolution Equations*, Springer-Verlag, New York, 2000. Feller, W., Semigroups of transformations in general weak topologies, *Annals of Math.* **57** (1953), 287–308. W. Feller, *An Introduction to Probability Theory and its Applications*, Vol. II, second edition, 1976. Folland, G.B., *Real Analysis*, Second Edition, Wiley-Interscience, 1999. Gripenberg, G., Londen, S-O. and O. Staffans, *Volterra Integral and Functional Equations*, Cambridge University Press, Cambridge, 1990. Haase, M., *The Functional Calculus for Sectorial Operators*, vol. 169 of Operator Theory: Advances and Applications, Birkäuser, 2006. Hale, J.K., *Functional Differential Equations*, Springer-Verlag, New York, 1971. Hale, J.K. and S.M. Verduyn Lunel, *Introduction to Functional Differential Equations*, Springer-Verlag, New York, 1993. Hille, E. and R. Phillips, *Functional Analysis and Semigroups*, American Mathematical Society, Providence, RI, 1957. Hönig, C. S. , *Volterra Stieltjes-integral equations*, Mathematics Studies 16, North-Holland Publ. Co., Amsterdam, 1975. W. Huyer, On periodic cohort solutions of a size-structured population model, *J. Math. Biol.* **35** (1997), 908-934. Jefferies, B., The generation of weakly integrable semigroups, *J. Funct. Anal.* **73** (1987), 195–215. Kato, T., *Perturbation Theory for Linear Operators $2nd edn.$*, Springer-Verlag, Berlin, 1976. Kunze, M., Continuity and equicontinuity of semigroups on norming dual pairs, *Semigroup Forum* **79** (2009), 540–560. Kunze, M., A Pettis-type integral and applications to transition semigroups, *Czech. Math. J.* **61** (2011), 437–459. Korevaar, J., On Newman’s quick way to the prime number theorem, *Math. Intelligencer* **4** (1982), 108–115. Kraaij, R., Strongly continuous and locally equi-continuous semigroups on locally convex spaces, *Semigroup Forum* **92** (2016), 158–185. Lant, T. and H.R. Thieme, Markov Transition Functions and Semigroups of Measures, *Semigroup Forum* **74** (2007), 337–369. Magal P. and S. Ruan, *Center manifolds for semilinear equations with non-dense domain and applications to Hopf bifurcation in age structured models*, Memoirs of the American Mathematical Society 2009; Volume 202, Number 951. Magal P. and S. Ruan, *Theory and Applications of Abstract Semilinear Cauchy Problems*, Springer-Verlag, Berlin, 2018. de Pagter, B., A characterization of sun-reflexivity, *Math. Ann.* **283** (1989), 511–518. Pazy, A. *Semigroups of Linear Operators and Applications to Partial Differential Equations*, Springer-Verlag, Berlin, 1983. Rudin, W., *Real and Complex Analysis $2nd edition$*, McGraw-Hill, New York, 1974. Rudin, W., *Functional Analysis $2nd edition$*, McGraw-Hill, New York, 1991. Schaefer, H.H., *Topological Vector Spaces*, Springer-Verlag, Berlin, 1986. Schaefer, H.H., Radon, Baire, and Borel measures on compact space, I, *Hokkaido Math. Journal*, **18** (1989), 231-243. Walther, H.O. , Autonomous linear neutral equations with bounded Borel functions as initial data, preprint, 2010. [^1]: Complexification of a norming dual pair entails some subtle difficulties regarding the choice of norms. These subtleties are explained in [@Diek95 Section III.7]. But when we deal with function spaces, complexification can be represented by allowing the functions to take values in $\BC$ or $\BC^n$ and subsequently the norm can be defined by copying the definition for the real functions, while replacing the real absolute value by the complex modulus. In the present paper the two relevant norms are the supremum norm and the total variation norm, see Appendix B.

ArXiv

--- abstract: 'We explicitly describe infinitesimal deformations of cyclic quotient singularities that satisfy one of the deformation conditions introduced by Wahl, Kollár–Shepherd-Barron and Viehweg. The conclusion is that in many cases these three notions are different from each other. In particular, we see that while the KSB and the Viehweg versions of the moduli space of surfaces of general type have the same underlying reduced subscheme, their infinitesimal structures are different.' address: - 'Institut für Mathematik, FU Berlin, Arnimalle 3, 14195 Berlin, Germany' - 'Department of Mathematics, Princeton University, Fine Hall, Washington Road Princeton, NJ 08544-1000, USA' author: - Klaus Altmann - János Kollár bibliography: - 'qG.bib' title: 'The dualizing sheaf on first-order deformations of toric surface singularities' --- [^1] Introduction {#intro} ============ In order to compactify the moduli space of surfaces of general type, one has to consider singular surfaces but for a long time it was not clear which class of singularities should be allowed. Building on Mori’s program, [@KSB] described such a class, named semi-log-canonical singularities. These include quotient singularities, cusps and a few others; see [@kk-singbook Sec.2.2] for a complete list. A new feature of the theory is that not every flat deformation of a surface with such singularities should be allowed in moduli theory. In essence this observation can be traced back to Bertini who observed that the cone over the degree 4 rational normal curve admits two distinct smoothings. One is the Veronese surface the other is a ruled surface; see [@pin-def]. For the Veronese the self-intersection of the canonical class is 9 for the ruled surface it is 8. Since we would like the basic numerical invariants to be locally constant in families, one of these deformations should not be allowed. It is not obvious how to obtain the right class of deformations. Three variants have been investigated in the past. Their common feature is that they all study the compatibility of deformations with powers of the dualizing sheaf $\omega$. In order to define these 3 versions, we need some definitions. General setup {#QG.V.defn} ------------- We are ultimately interested in schemes with semi-log-canonical singularities ${S}$, but for the basic definitions we need to assume only that ${S}$ is a pure dimensional $S_2$ scheme over a field $k$ such that 1. there is a closed subset $Z\subset{S}$ of codimension $\geq 2$ such that $\omega_{{S}\setminus Z}$ is locally free and 2. there is an ${m}>0$ such that $\omega_{S}^{[{m}]}$ is locally free, where $\omega_{S}^{[{m}]}$ denotes the reflexive hull of $\omega_{S}^{\otimes {m}}$. The smallest such ${m}>0$ is called the [*index*]{} of $\omega_{S}$. (Both of these conditions are satisfied by schemes with semi-log-canonical singularities.)\ Let $(0,T)$ be a local scheme such that $k(0)\cong k$ and $p:X_T\to T$ a flat deformation of ${S}\cong X_0$. For every ${g}\in {{\mathbb Z}}$ we have natural restriction maps $${\operatorname{\mathcal R}}^{[{g}]}:\omega_{X_T/T}^{[{g}]}|_{X_0}\to \omega_{X_0}^{[{g}]}.$$ These maps are isomorphisms over ${S}\setminus Z$ and we are interested in understanding those cases when they are isomorphisms over ${S}$. The local criterion of flatness shows (see [@k-modbook] for details) that if $T$ is Artinian then $${\operatorname{\mathcal R}}^{[{g}]} \mbox{ is an isomorphism } {\Leftrightarrow}{\operatorname{\mathcal R}}^{[{g}]} \mbox{ is surjective } {\Leftrightarrow}\omega_{X_T/T}^{[{g}]} \mbox{ is flat over } T.$$ We will denote this condition by ${\mbox{$(*)_{g}$}}$ (with ${g}\in{\mathbb Z}$). Definitions of - and V- and VW-deformations {#QG.V.defn.2} ------------------------------------------- Let $p:X_T\to T$ be a flat deformation as in (\[QG.V.defn\]). ### ${\operatorname{qG}}$-deformations {#def-qG} We call $p:X_T\to T$ a [*${\operatorname{qG}}$-deformation*]{} if the conditions defined in (\[QG.V.defn\]) hold for every ${g}\in{\mathbb Z}$. It is enough to check these for ${g}=1,\dots, {\operatorname{index}}(\omega_{S})$. (${\operatorname{qG}}$ is short for “Quotient of Gorenstein,” but this is misleading if $\dim {S}\geq 3$.)\ These deformations were introduced and studied by Kollár and Shepherd-Barron [@KSB] as the class most suitable for compactifying the moduli of varieties of general type. A list of log canonical surface singularities with ${\operatorname{qG}}$-smoothings is given in [@KSB]. In the key case of cyclic quotient singularities the list (of the so-called T-singularities) was earlier established by Wahl [@MR82d:14004 2.7], though he viewed them as examples of W-deformations (see below). ### V-deformations {#def-V} We call $p:X_T\to T$ a [*Viehweg-type deformation*]{} (or V-deformation) if the conditions from (\[QG.V.defn\]) hold for every ${g}$ divisible by ${\operatorname{index}}(\omega_{S})$. It is enough to check this for ${g}={\operatorname{index}}(\omega_{S})$.\ These deformations form the natural class suggested by the geometric invariant theory methods used in the monograph [@vieh-book]. Actually, [@vieh-book] considers the—a priori weaker—condition: ${\operatorname{\mathcal R}}^{[{g}]}$ is an isomorphism for some ${g}>0$ divisible by ${\operatorname{index}}(\omega_{S})$. One can see that in this case holds for every ${g}$ divisible by ${\operatorname{index}}(\omega_{S})$, at least in characteristic 0; see [@k-modbook]. V-deformations are problematic in positive characteristic, see [@hac-kov 14.7]. ### W-deformations {#def-W} We call $p:X_T\to T$ a [*Wahl-type deformation*]{} (or W-deformation) if the condition holds for ${g}=-1$. These deformations were considered in [@MR82d:14004; @MR83h:14029] and called $\omega^*$-constant deformations there. ### VW-deformations {#def-VW} We call $p:X_T\to T$ a VW-deformation if it is both a V-deformation and a W-deformation. Relations between , V and VW {#rel-qG-VW} ---------------------------- It is clear that every ${\operatorname{qG}}$-deformation is also a VW-deformation. Understanding the precise relationship between the four classes (\[def-qG\]) – (\[def-VW\]) has been a long standing open problem. For reduced base spaces we have the following very strong result. \[th-VqG\] A flat deformation of a log canonical scheme over a reduced, local scheme of characteristic 0 is a V-deformation if and only if it is a ${\operatorname{qG}}$-deformation. When $T$ is the spectrum of a DVR, $\dim {S}=2$ and ${S}$ has quotient singularities, this was proved in [@MR541025] and [@k-flat 14.2]. If $\dim {S}>2$ and ${S}$ has log terminal singularities, this is a special case of inversion of adjunction as proved in [@k-etal Sec.17] and the log canonical case similarly follows from [@kawakita] and the normality of log canonical centers [@ambro; @fujinobook]; see also [@kk-singbook Sec.4.3]. These imply the claim for arbitrary reduced base schemes using [@k-hh]; see [@k-modbook] for more details.\ This raised the possibility that every V-deformation of a log-canonical singularity is also a ${\operatorname{qG}}$-deformation over arbitrary base schemes. It would be enough to check this for Artinian bases. In this note we focus on first order deformations and prove that these two classes are quite different from each other. \[def-TqGVW\] Let ${S}$ be a scheme satisfying the conditions (\[QG.V.defn\])(i)-(ii). Let ${{T^1}}({S})$ be the (possibly infinite dimensional) $k$-vector space of deformations of ${S}$ over $\operatorname{Spec}k[\epsilon]$. We denote by ${{T^1_{{\operatorname{qG}}}}}({S})\subset {{T^1}}({S})$ the space of first order ${\operatorname{qG}}$-deformations, ${{T^1_V}}({S})$ the space of first order V-deformations, ${{T^1_W}}({S})$ the space of first order W-deformations, and ${{T^1_{{\operatorname{V\!W}}}}}({S})$ the space of first order VW-deformations. We have obvious inclusions $${{T^1_{{\operatorname{qG}}}}}({S})\;\subset\; {{T^1_{{\operatorname{V\!W}}}}}({S})\;\subset\; {{T^1_V}}({S}), {{T^1_W}}({S})\;\subset\; T^1({S}),$$ but the relationship between ${{T^1_V}}({S})$ and $ {{T^1_W}}({S}) $ is not clear. The case of cyclic quotient singularities {#resultCQS} ----------------------------------------- We completely describe first order V-, - and ${\operatorname{qG}}$-deformations of two-dimensional cyclic quotient singularities. The precise answers are stated in Sections \[V-deformations\] and \[compJanos\]. The main conclusion is that V-deformations and -deformations, and even more V-deformations and -deformations are quite different over Artinian bases. \[VW-GQ.thm.1\] Let $S_{n,q}:={{\mathbb A}}^2/\tfrac1{n}(1,q) $ denote the quotient of ${{\mathbb A}}^2$ by the cyclic group action generated by $(x,y)\mapsto (\eta x, \eta^q y)$, where $\eta$ is a primitive $n$th root of unity. Then, if $\,q\neq -1$ in $({\mathbb Z}/n{\mathbb Z})^*$, i.e.if $\,{\operatorname{embdim}}\bigl(S_{n,q}\bigr)\geq 4$, $$\dim {{T^1_V}}\bigl(S_{n,q}\bigr)- \dim {{T^1_{{\operatorname{V\!W}}}}}\bigl(S_{n,q}\bigr)= {\operatorname{embdim}}\bigl(S_{n,q}\bigr) - 4 \hspace{0.5em}\mbox{or}\hspace{0.5em} {\operatorname{embdim}}\bigl(S_{n,q}\bigr) - 5.$$ In particular, if ${\operatorname{embdim}}\bigl(S_{n,q}\bigr)\geq 6$ then $S_{n,q}$ has V-deformations that are not ${\operatorname{V\!W}}$-deformations, hence also not ${\operatorname{qG}}$-deformations. This is a direct consequence of the more detailed Theorem \[th-IAB\]. By contrast, ${\operatorname{qG}}$-deformations and VW-deformations are quite close to each other, as shown by the next result. This will be proved in (\[proofIntro\]). \[V-VW.thm.1\] Let $S_{n,q}$ be as in the previous theorem. Then 1. If $\gcd(n, q+1)=1$ then ${{T^1_{{\operatorname{qG}}}}}\bigl(S_{n,q}\bigr)={{T^1_{{\operatorname{V\!W}}}}}\bigl(S_{n,q}\bigr)=\{0\}$. 2. If $S_{n,q}$ admits a ${\operatorname{qG}}$-smoothing then ${{T^1_{{\operatorname{qG}}}}}\bigl(S_{n,q}\bigr)={{T^1_{{\operatorname{V\!W}}}}}\bigl(S_{n,q}\bigr)$. 3. In general $\,\dim {{T^1_{{\operatorname{qG}}}}}\bigl(S_{n,q}\bigr)\leq \dim {{T^1_{{\operatorname{V\!W}}}}}\bigl(S_{n,q}\bigr)\leq \dim {{T^1_{{\operatorname{qG}}}}}\bigl(S_{n,q}\bigr)+1$. Using the interval language {#usingInt} --------------------------- Besides the description of cyclic quotient singularities in terms of the invariants $n$ and $q$, there is an alternative possibility by using rational intervals $I=[-{A},{B}]\subseteq {\mathbb Q}$ with uniform denominators at the end points, i.e. ${A}$ and ${B}$ have the same denominator in reduced form. We call $I$ or the resulting singularity $S_I$ [*grounded*]{} if $I$ contains an integer in its interior. Since integral shifts of $I$ will be neglected, this leads to ${A},{B}>0$. See (\[coneInterval\]) and (\[groundedCones\]) for details. This language allows a much more detailed description of the situation: \[th-IAB\] Assume that $\,{\operatorname{embdim}}\bigl(S_I\bigr)\geq 4$. Then\ [**1)**]{} If the interval $I$ is not grounded, then the associated surface singularity $S_I$ has neither - nor -deformations. The dimension of ${{T^1_V}}(S_I)$ is ${\operatorname{embdim}}(S_I)-4$.\ [**2)**]{} If ${A},{B}>0$, then $\dim {{T^1_V}}(S_I)={\operatorname{embdim}}(S_I)-4 + {{\lfloor {A}\rfloor}}+{{\lfloor {B}\rfloor}}$.\ [**3)**]{} If ${A},{B}>0$ with fractional parts $\{{A}\}=\frac{1}{{m}}$ or $\{{B}\}=\frac{1}{{m}}$, then $$\dim {{T^1_{{\operatorname{V\!W}}}}}(S_I)=\dim {{T^1_{{\operatorname{qG}}}}}(S_I)={{\lfloor {A}+{B}\rfloor}}.$$ [**4)**]{} If ${A},{B}>0$ with both fractional parts $\{{A}\}$ and $\{{B}\}$ different from $\frac{1}{{m}}$, then $$\dim {{T^1_{{\operatorname{V\!W}}}}}(S_I)={{\lfloor {A}\rfloor}}+{{\lfloor {B}\rfloor}}+1 \hspace{0.8em}\mbox{and}\hspace{0.8em} \dim {{T^1_{{\operatorname{qG}}}}}(S_I)={{\lfloor {A}+{B}\rfloor}}.$$ The first two parts, i.e. the description of the V-deformations follows from (\[restrictPhi\]). The remaining two parts are just another formulation of Theorem \[th-lastDef\]. Implications for moduli spaces ------------------------------ One can construct compactified moduli spaces for surfaces of general type using either KSB-deformations or V-deformations. Let us denote these by ${\mathbf M}(\mbox{KSB})$ and ${\mathbf M}(\mbox{V})$. By Theorem \[th-VqG\], the underlying reduced structures of these moduli spaces are isomorphic. As a consequence of our computations we can say that the scheme structures are not isomorphic. More generally, let $X$ be a projective variety with isolated singularities $x_1,\dots, x_m$. Any flat deformation of $X$ restricts to a deformation of the singularities $(x_i, X)$. This induces a map of the local deformation spaces $$\Re: \operatorname{Def}(X)\to \operatorname{Def}(x_1, X)\times \cdots \times \operatorname{Def}(x_m, X).$$ A direct consequence of the definition of -deformations given in [@KSB] is that $\operatorname{Def}_{{\operatorname{qG}}}(x, X)$ is smooth for 2-dimensional quotient singularities. Our computations show that, by contrast, $\operatorname{Def}_{V}(x, X)$ is usually non-reduced but $$\operatorname{red}\bigl(\operatorname{Def}_{V}(x, X)\bigr)=\operatorname{Def}_{{\operatorname{qG}}}(x, X).$$ We thus expect that if $X$ is a surface with quotient singularities then $\operatorname{Def}_V(X)$ can be non-reduced but $\operatorname{Def}_{{\operatorname{qG}}}(X)$ should be smooth. This is not true in general, but there are many examples when local-to-global obstructions vanish and the map $\Re$ is smooth. The situation is not well understood for surfaces of general type, but [@MR2581246 Prop 3.1] shows that local-to-global obstructions vanish for Del Pezzo surfaces. Thus we obtain that if $S$ is a Del Pezzo surface with quotient singularities then $\operatorname{Def}_{{\operatorname{qG}}}(S)$ is smooth but $\operatorname{Def}_{V}(S)$ is nonreduced as soon as $S$ has at least 1 singular point of multiplicity $\geq 5$. Five descriptions of cyclic quotient singularities {#fiveCQS} ================================================== In (\[normAct\]) – (\[coneInterval\]) we present several ways of representing two-dimensional cyclic quotient singularities $S={\mathbb A}^2_k/G$, i.e. those coming from a cyclic group $G$ acting on ${\mathbb A}^2_k$. While most of them are quite classic, the description (\[coneInterval\]) seems to be not common so far. At the end, in (\[groundedCones\]), we introduce the notion of grounded singularities. In the language of (\[coneInterval\]) this becomes especially simple. Normalizing the action {#normAct} ---------------------- Let $G$ denote a cyclic group of order $n$ with ${\operatorname{char}}k\nmid n$. Then, by [@Brieskorn §2], every linear action of $G$ on ${\mathbb A}^2_k$ is isomorphic to some action $\tfrac1{n}(1,q)$ generated by $$(x,y)\mapsto (\eta x, \eta^q y),$$ where $q\in({\mathbb Z}/n{\mathbb Z})^*$ and $\eta$ is a primitive $n$-th root of unity. The corresponding ring of invariants is $R_{n,q}:=k[x,y]^G$ and the corresponding quotient singularity is $$S_{n,q}:={{\mathbb A}}^2/\tfrac1{n}(1,q)=\operatorname{Spec}R_{n,q}.$$ While we work with this affine model, all the results apply to its localization, Henselisation or completion at the origin. We can also choose $\eta'=\eta^q$ as our primitive $n$-th root of unity. This shows the isomorphism $$S_{n,q}\cong S_{n,q'} \mbox{ where } qq'=1 \mbox{ in } ({\mathbb Z}/n{\mathbb Z})^*.$$ Note that we can and will choose a representative for $q$ such that $1\leq q\leq n-1$. The case $q=n-1$ encodes the $A_{n-1}$-singularities. These are exceptional for many of the subsequent formulas, so we assume from now on that $q\neq -1$ in $({\mathbb Z}/n{\mathbb Z})^*$. The abc notation {#abcNotation} ---------------- Here we just rename the invariants $n$ and $q$. Denote $b:=\gcd(n,q+1)$, $a:=n/b$, and $c:=(q+1)/b$. Hence we know that $\gcd(a,c)=1$, and $n$ and $q$ can be recovered as $n=ab$ and $q=bc-1$. When using these invariants, we might write $S_{abc}=\frac{1}{ab}(1,bc-1)$ instead of $S_{n,q}=\frac{1}{n}(1,q)$.\ Note that the case $q=n-1$ which was just excluded at the end of (\[normAct\]) can be recovered in the abc language as the case $a=1$.\ The isomorphic singularities $S_{n,q}$ and $S_{n,q'}$ from (\[normAct\]) share the same $a$ and $b$, i.e. $a'=a$ and $b'=b$. This follows from the fact that $qq'\equiv 1\hspace{-0.3em}\mod n$ implies $qq'\equiv 1\hspace{-0.3em}\mod b$ and that $q\equiv -1\hspace{-0.3em}\mod b$ becomes then equivalent to $1\equiv -q'\hspace{-0.3em}\mod b$. The third invariants $c$ and $c'$ differ. However, it is in general not true that they are mutually inverse within $({\mathbb Z}/a{\mathbb Z})^*$. See the discussion at the end of (\[coneInterval\]). The toric nature of $S_{n,q}$ {#toricSnq} ----------------------------- Dealing with toric varieties involves a standardized language, cf. [@CoxLittleSchenck] for details: Assume that $N$ and $M$ are mutually dual free abelian groups of finite rank; with $N_{\mathbb Q}$ and $M_{\mathbb Q}$ we denote the associated ${\mathbb Q}$-vector spaces; similarly we often write $N_{k}$ and $M_{k}$ for $N\otimes_{\mathbb Z}{k}$ and $M\otimes_{\mathbb Z}{k}$, respectively. Let $\sigma\subseteq N_{\mathbb Q}$ be a polyhedral cone and denote by $\sigma{^{\scriptscriptstyle\vee}}:=\{r\in M_{\mathbb Q}{\,|\;}\langle \sigma,r\rangle\geq 0\}$ the dual one. Then, $\sigma{^{\scriptscriptstyle\vee}}\cap M$ is a finitely generated semigroup, and its $M$-graded semigroup ring (with ${k}$-basis $\{x^r{\,|\;}r\in \sigma{^{\scriptscriptstyle\vee}}\cap M\}$) provides the affine toric variety $${{\mathbb T}{\mathbb V}}(\sigma,N):=\operatorname{Spec}{k}[\sigma{^{\scriptscriptstyle\vee}}\cap M].$$ Since we are going to deal with surface singularities, $N$ and $M$ will be of rank two. Hence, the primitive generators of $\sigma$ and $\sigma{^{\scriptscriptstyle\vee}}$ are just pairs ${\alpha},{\beta}\in N$ and ${r^1},{r^e}\in M$ (with $\langle{\alpha},{r^1}\rangle=\langle{\beta},{r^e}\rangle=0$), respectively. The relation to (\[normAct\]) is the well-known \[prop-toricCQS\] $\,S_{n,q}={{\mathbb T}{\mathbb V}}(\sigma,{\mathbb Z}^2)$ where $ \,\sigma=\langle (1,0),\,(-q,n)\rangle\subseteq{\mathbb Q}^2=N_{\mathbb Q}$. Note that we use $\langle{{{\scriptscriptstyle}\bullet}},{{{\scriptscriptstyle}\bullet}}\rangle$ to denote both the pairing $N\times M\to{\mathbb Z}$ and the generation of a polyhedral cone. Moreover, when using coordinates, we try to distinguish between $M$ and $N$ by using the different brackets $[{{{\scriptscriptstyle}\bullet}},{{{\scriptscriptstyle}\bullet}}]$ and $({{{\scriptscriptstyle}\bullet}},{{{\scriptscriptstyle}\bullet}})$, respectively. So we will write $\sigma{^{\scriptscriptstyle\vee}}=\langle [0,1],\,[n,q]\rangle\subseteq{\mathbb Q}^2=M_{\mathbb Q}$. That is, ${\alpha}=(1,0)$, ${\beta}=(-q,n)$, ${r^1}=[0,1]$, and ${r^e}=[n,q]$. The group order $n$ may be recovered as $\det\sigma:=\det({\alpha},{\beta})$. Writing $k[x,y]$ as the semigroup ring $k[{\mathbb N}^2]$, we know that $R_{n,q}=k[x,y]^G=k[{\mathbb N}^2\cap M]$ where $M\subseteq{\mathbb Z}^2$ is the sublattice freely generated by, e.g. $[-q,1], [n,0]\in{\mathbb Z}^2$. Now, a linear combination $ \lambda\cdot[-q,1]+\mu\cdot [n,0]=[-\lambda q + \mu n,\, \lambda] $ has non-negative entries if and only if $ \langle (-q,n), [\lambda,\mu]\rangle\geq 0 $ and $ \langle (1,0), [\lambda,\mu]\rangle\geq 0 $. Equations of $S_{n,q}$ via continued fractions {#equSnq} ---------------------------------------------- Let $S:={{\mathbb T}{\mathbb V}}(\sigma)$ for some two-dimensional cone $\sigma=\langle{\alpha},{\beta}\rangle$ as in (\[toricSnq\]). Denote by $E\subset\sigma{^{\scriptscriptstyle\vee}}\cap M$ the set of indecomposable elements within this semigroup (“Hilbert basis”). This finite set coincides with the lattice points on the compact edges of $\operatorname{conv}(\sigma{^{\scriptscriptstyle\vee}}\cap M\setminus 0)$. In particular, we can naturally list its elements as $E=\{{r^1},r^2,\ldots,r^{e-1},{r^e}\}$ with $e\geq 4$ (the cases $e=2$ and $e=3$ refer to $S$ being smooth or an $A_{n-1}$-singularity). Any two adjacent elements of this set do always form a ${\mathbb Z}$-basis of $M\cong{\mathbb Z}^2$. Hence, for $i=2,\ldots,e-1$, we can write $$r^{i-1}+r^{i+1}=a_i\cdot r^i$$ with natural numbers $a_i\geq 2$. The continued fraction $[a_2,\ldots,a_{e-1}]:= a_2-\frac{1}{a_3-\ldots}$ recovers $\frac{n}{n-q}$. Moreover, $E$ provides an embedding $S\hookrightarrow{\mathbb A}_{k}^e$. Among the equations one finds $x_{i-1}x_{i+1}-x_i^{a_i}$, see [@riems209] for more details. \[rem-Ji\] With growing $i$, the values $\langle{\alpha},r^i\rangle$ and $\langle{\beta},r^i\rangle$ increase and decrease, respectively. Hence, defining $\eta_i:=\min\{\frac{\langle{\alpha},r^{i+1}\rangle}{\langle{\alpha},r^i\rangle}, \frac{\langle{\beta},r^{i-1}\rangle}{\langle{\beta},r^i\rangle}\}\in{\mathbb Q}_{\geq 1}$, we obtain that ${{\lfloor \eta_i \rfloor}}=a_i-1\in{\mathbb Z}_{\geq 1}$. Replacing cones by intervals {#coneInterval} ---------------------------- Let $\sigma=\langle{\alpha},{\beta}\rangle$ and $\sigma{^{\scriptscriptstyle\vee}}=\langle{r^1},{r^e}\rangle$ be mutually dual (two-dimensional, rational) cones as before. The primitive elements $R\in\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}\cap M$ (we will call them [*primitive degrees*]{} of $\sigma$) give rise to affine crosscuts $Q(\sigma,R):=\sigma\cap[R=1]$. Since the affine line $[R=1]$ can be identified with the rational line ${\mathbb Q}^1$ (canonically, up to integral shifts), we can and will understand $Q(\sigma,R)$ as an interval in ${\mathbb Q}$.\ Reciprocally, every closed interval $I\subseteq {\mathbb Q}$ provides a cone via $C(I):={\mathbb Q}_{\geq 0}\cdot (I,1)\subseteq{\mathbb Q}^2$ and a primitive degree $R:=[0,1]$. These two constructions provide a natural one-one correspondence $$\big\{\mbox{pairs $(\sigma,R)$}\big\}\big/{\operatorname{SL}}(2,{\mathbb Z}) \longleftrightarrow \big\{\mbox{bounded intervals $I\subseteq{\mathbb Q}$}\big\}\big/ \{\mbox{${\mathbb Z}$-shifts}\}.$$ On the other hand, every cone $\sigma$ provides a canonical primitive degree ${\overline}{R}$, called the central degree. It is defined as the primitive generator of the ray ${\mathbb Q}_{\geq 0}\cdot ({r^1}+{r^e})$. It is the only primitive degree such that $\langle{\alpha},{\overline}{R}\rangle = \langle {\beta},{\overline}{R}\rangle$. Using coordinates via the $(n,q)/(a,b,c)$ language discussed in (\[abcNotation\]), one obtains that ${r^1}+{r^e}=[n,q+1]=b\cdot[a,c]$, hence ${\overline}{R}=[a,c]=\frac{{r^1}+{r^e}}{b}$. \[rem-index\] Actually, $S={{\mathbb T}{\mathbb V}}(\sigma)$ is ${\mathbb Q}$-Gorenstein with index $a$ (and we suppose that $a>1$). The corresponding power $\omega_S^{[a]}$ equals the ideal $(x^{{\overline}{R}})\subseteq{{\mathcal O}}_S$ represented by the shifted semigroup ${\overline}{R}+(\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}\cap M)$. Thus, properly speaking, not ${\overline}{R}$ but the non-integral $\frac{1}{a}{\overline}{R}$ is the truly canonical degree. Using this special central degree ${\overline}{R}$, the previous correspondence yields \[prop-CentrDeg\] There is a one-one correspondence $$\big\{\mbox{\rm cones $\sigma$}\big\}\big/{\operatorname{SL}}(2,{\mathbb Z}) \longleftrightarrow \big\{\mbox{\rm intervals $I\subseteq{\mathbb Q}$ with uniform denominators}\big\}\big/ \{\mbox{\rm ${\mathbb Z}$-shifts}\}.$$ We call $I$ to have “[*uniform denominators*]{}” (at the end points) if both become equal in the reduced forms, i.e. if $I=[\frac{g}{{m}},\frac{h}{{m}}]$ with $g,h,{m}\in{\mathbb Z}$ and $\gcd(g,{m})=\gcd(h,{m})=1$. $({\Rightarrow}$) After a possible coordinate change, we may assume that ${\overline}{R}=[0,1]$. Setting ${m}:=\langle {\alpha},[0,1]\rangle = \langle {\beta},[0,1]\rangle$ we obtain that ${\alpha}=(g,{m})$ and ${\beta}=(h,{m})$, hence $Q(\sigma,\,[0,1])=[\frac{g}{{m}},\frac{h}{{m}}]$ for some $g,h$ as asked for in the claim.\ (${\Leftarrow}$) If $I=[\frac{g}{{m}},\frac{h}{{m}}]$, then $C(I)=\langle (\frac{g}{{m}},1),\,(\frac{h}{{m}},1)\rangle = \langle (g,{m}),\,(h,{m})\rangle$, i.e. its primitive generators are ${\alpha}=(g,{m})$ and ${\beta}=(h,{m})$. Thus, $R=[0,1]$ coincides with ${\overline}{R}$. In (\[toricSnq\]) we had considered cones $\sigma=\langle {\alpha},{\beta}\rangle=\langle (1,0),\,(-q,n)\rangle$, i.e.$n=|\det({\alpha},{\beta})|$, and $q$ was characterized by $n|(q{\alpha}+{\beta})$. Alternatively we had used $b:=\gcd(n,q+1)$ to write $n=ab$ and $q+1=bc$ in (\[abcNotation\]). Now, given an interval $I=[\frac{g}{{m}},\frac{h}{{m}}]$ as in Proposition \[prop-CentrDeg\], it has length $|I|=\frac{h-g}{{m}}$, and we may obtain the invariants $(a,b,c)$ for $\sigma:=C(I)$ via \[prop-abcInv\] For $I=[\frac{g}{{m}},\frac{h}{{m}}]$ one has $a={m}$, $b=h-g$, and $c=-1/g\in ({\mathbb Z}/{m}{\mathbb Z})^*$. In particular, $b/a=n/{m}^2=|I|$ and $\,{\operatorname{index}}(\omega_S)={m}$ with $S={{\mathbb T}{\mathbb V}}(C(I))$. Let $a,b,c$ be as in the claim. By definition, we have $\gcd(a,c)=1$. We have to show that the generators ${\alpha}=(g,{m})$ and ${\beta}=(h,{m})$ of $C(I)$ and the invariants $n:=ab$, $q:=bc-1$ yield isomorphic cones: First, we clearly obtain that $|\det({\alpha},{\beta})|=(h-g){m}=n$. It remains to check the characterizing relation $n|(q{\alpha}+{\beta})$. But this follows from $$q{\alpha}+{\beta}=\big( (h-g)\,c-1\big)\cdot (g,{m}) + (h,{m}) \;=\; \big((h-g)(cg+1),\; (h-g)\,c{m}\big)$$ which is indeed divisible by $n=(h-g){m}$. Finally, $\,{\operatorname{index}}(\omega_S)=a$ by Remark \[rem-index\]. In (\[abcNotation\]) we mentioned the invariant $c'$ associated to $(n,q')$ as it was $c$ to $(n,q)$. In the “interval language”, to switch $q$ and $q'$ means to replace $I$ by $-I$, i.e. to keep ${m}$ and to replace $g$ and $h$ by $-h$ and $-g$, respectively. In particular, this implies that $c'=1/h\in({\mathbb Z}/{m}{\mathbb Z})^*$.\ Moreover, there is a way to visualize both $c$ and $c'$: The points $[c,\,-\frac{gc+1}{{m}}]$ and $[-c',\,\frac{hc'-1}{{m}}]$ appear as the first lattice points on the two rays of the shifted cone $C(I){^{\scriptscriptstyle\vee}}-[0,\frac{1}{{m}}]$, cf. the proof of Proposition \[prop-finalVW\]. Grounded cones and intervals {#groundedCones} ---------------------------- To represent two-dimensional cones $\sigma$ by intervals $I$ via Proposition \[prop-CentrDeg\], the central degree ${\overline}{R}$ played an important role. This leads to the following notion: \[def-grounded\] A two-dimensional, polyhedral cone $\sigma$ (or the associated interval $I$, or the associated singularity $S_{n,q}=S_{abc}={{\mathbb T}{\mathbb V}}(\sigma)$) is called [*grounded*]{} $:{\Leftrightarrow}$ the central degree ${\overline}{R}$ belongs to the Hilbert basis $E=\{{r^1},r^2,\ldots,{r^e}\}$ of $\sigma{^{\scriptscriptstyle\vee}}\cap M$, i.e. ${\overline}{R}$ is irreducible within this semigroup. If ${\overline}{R}=r^{\nu}$, then ${\nu}$ is called the central index. \[prop-piercInt\] An interval $I\subseteq{\mathbb Q}$ with uniform denominators is grounded if and only if it contains an interior integer. $({\Leftarrow})$ We may assume that $0\in\operatorname{int}I$, i.e.$I=[\frac{g}{{m}},\frac{h}{{m}}]$ with ${m}>0$, $g<0$, and $h>0$. Then, the dual cone of $\sigma=C(I)$ equals $\sigma{^{\scriptscriptstyle\vee}}=\langle [-{m},h],\,[{m},-g]\rangle\subseteq ({\mathbb Q}\times {\mathbb Q}_{>0})\cup\{[0,0]\}$. On the other hand, the central degree ${\overline}{R}$ coincides with $[0,1]$, and it is obvious that this is irreducible even within the semigroup $({\mathbb Z}\times {\mathbb Z}_{>0})\cup\{[0,0]\}$.\ (${\Rightarrow}$) Let $I=[\frac{g}{{m}},\frac{h}{{m}}]$ with ${m}>0$ and $g<h$, i.e. $\sigma{^{\scriptscriptstyle\vee}}=\langle [-{m},h],\,[{m},-g]\rangle$. We are going to show that we can obtain $g<0$ and $h>0$ by an integral shift of $I$. Obviously, we can assume that $0<h<{m}$ implying that $$[-1,1]\in\operatorname{int}\langle [-{m},h],\,[0,1]\rangle \subseteq\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}.$$ On the other hand, if we had $g>0$, then this would similarly imply that $$[1,0]\in\operatorname{int}\langle [0,1],\,[{m},-g]\rangle \subseteq\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}.$$ Then ${\overline}{R}=[0,1]=[-1,1] + [1,0]$ would be a decomposition within $\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}$. Grounded intervals can always be shifted by integers to look like $I=[-{A},{B}]$ with ${A},{B}\in{\mathbb Q}_{>0}$ (sharing the same denominator). Then, if ${\nu}$ denotes the central index, we can directly express the invariants $\eta_{\nu}$ and $a_{\nu}$ from (\[equSnq\]) in terms of $I$: \[prop-etaci\] Let $I=[-{A},{B}]$ with ${A},{B}\in{\mathbb Q}_{>0}$ be a [(]{}grounded[)]{} interval with uniform denominators. Then $$\eta_{\nu}=1+\min\{{{\lfloor {A}\rfloor}}+{B},\, {A}+{{\lfloor {B}\rfloor}}\}, \hspace{0.8em} a_{\nu}=2+{{\lfloor {A}\rfloor}}+{{\lfloor {B}\rfloor}}, \hspace{0.6em} \mbox{and} \hspace{0.5em} |I|={A}+{B}. \vspace{0.3ex}$$ With ${A}=\frac{-g}{{m}}$ and ${B}=\frac{h}{{m}}$, we have $\sigma{^{\scriptscriptstyle\vee}}=\langle [-{m},h],\,[{m},-g]\rangle$ as usual. Now, since $\,r^{\nu}={\overline}{R}=[0,1]$ and $\{r^{{\nu}-1},r^{\nu}\}$ forms a basis of ${\mathbb Z}^2$, we know that $r^{{\nu}-1}=[1,{{{\scriptscriptstyle}\bullet}}]$, and it has to be the lowest lattice point above the ray ${\mathbb Q}_{\geq 0}\cdot [{m},-g]={\mathbb Q}_{\geq 0}\cdot [1,{A}]$. Thus, $r^{{\nu}-1}=[1,{{\lfloor {A}\rfloor}}+1]$ and, similarly, $r^{{\nu}+1}=[-1,{{\lfloor {B}\rfloor}}+1]$. Now, the claim for $\eta_{\nu}=\min\{\frac{\langle{\alpha},r^{{\nu}+1}\rangle}{\langle{\alpha},r^{\nu}\rangle}, \frac{\langle{\beta},r^{{\nu}-1}\rangle}{\langle{\beta},r^{\nu}\rangle}\}$ follows from ${\alpha}={m}\cdot(-{A},1)$ and ${\beta}= {m}\cdot ({B},1)$. The dualizing sheaf on infinitesimal deformations of $S$ {#canX} ======================================================== The (isomorphism classes of) infinitesimal ${k}[{\varepsilon}]$-deformations (with ${\varepsilon}^2=0$) of a ${k}$-algebra are gathered in a vector space called ${{T^1}}$, see [@janS] for a detailed introduction to deformation theory. In case of toric varieties such as $S={{\mathbb T}{\mathbb V}}(\sigma)$ from (\[toricSnq\]), the torus ${{{\mathbb T}}}:=\operatorname{Spec}{k}[M]$ acts on the variety, on the functions, and on all naturally defined modules. In particular, the vector space ${{T^1}}$ becomes $M$-graded. This can be made explicit by comparing the $M$-degrees of the defining equations $f$ with those of the perturbation $g$ arising in $f+{\varepsilon}g$, cf. (\[constX\]). Thus, the distribution along the degrees of $M$ becomes the essential information. We will study the dualizing sheaf $\omega_X$ on the total spaces $X=X_\xi$ for homogeneous elements $\xi\in {{T^1}}(S)$. Degrees carrying ${{T^1}}$ {#degT1} -------------------------- \[SchlessT1\] Let $\sigma$ be a two-dimensional cone – we will adopt the notation of (\[toricSnq\]) and (\[equSnq\]). The dimensions of the homogeneous components ${{T^1}}(S,-R)$ ($R\in M$) of the finite-dimensional vector space ${{T^1}}(S)$ (abbreviated as ${{T^1}}(-R)\subseteq{{T^1}}$) are, [@pinkham_CQS]: 1. \[eckdegs\] $R=r^2$ or $R=r^{e-1}$: $\;\dim_{k}{{T^1}}(-R)=1$, 2. \[maindegs\] $R=r^i$ for $i=3,\ldots,e-2$: $\;\dim_{k}{{T^1}}(-R)=2$, and 3. \[multipledegs\] $R=k\cdot r^i$ for $i=2,\ldots,e-1$ with $2\leq k\leq a_i-1$: $\;\dim_{k}{{T^1}}(-R)=1$. We would like to recall Pinkham’s method to obtain this – this approach will also provide the major tool for our own calculations of $\omega_X$. However, unlike the original reference, we will consequently use the toric language. It leads to a slightly more structured description than just naming the dimensions. ### Puncturing The main point is to consider deformations of the smooth, but non-affine $S\setminus 0$ first. They are always locally trivial, and some of them lift to deformations of $S$. The exact statement for the ${k}[{\varepsilon}]$- level is encoded in the exact sequence $$0 \to {{T^1}}\to {\operatorname{H}}^1(S\setminus 0,\theta_S) \to {\operatorname{H}}^1(S\setminus 0, {{\mathcal O}}_S^e)$$ where the latter map is given by $\sum_{i=1}^e dx^{r^i}$. For the upcoming calculations it is helpful to use this sequence for redoing the calculation of ${{T^1}}(-R)$ for $R\in M$. Moreover, since we have a very nice open affine covering $S\setminus 0={{\mathbb T}{\mathbb V}}({\alpha})\cup{{\mathbb T}{\mathbb V}}({\beta})$ where we identify ${\alpha}$ and ${\beta}$ with the rays they are generating, hence ${{\mathbb T}{\mathbb V}}({\alpha})$ and ${{\mathbb T}{\mathbb V}}({\beta})$ are defined similarly to ${{\mathbb T}{\mathbb V}}(\sigma)$ in (\[toricSnq\]). Since $${{\mathbb T}{\mathbb V}}({\alpha})\cap{{\mathbb T}{\mathbb V}}({\beta})={{\mathbb T}{\mathbb V}}({\alpha}\cap{\beta})= {{\mathbb T}{\mathbb V}}(0)=\operatorname{Spec}{k}[M]={{{\mathbb T}}},$$ this is easily done by using Čech cohomology: ### ${\operatorname{H}}^1(S\setminus 0, {{\mathcal O}}_S)$ {#cechO} The $1$-Čech cocycles are ${\Gamma}({{{\mathbb T}}},{{\mathcal O}}_S|_{{{\mathbb T}}})={k}[M]$, and the $1$-Čech coboundaries are generated by the monomials $x^{-R}\in{k}[M]$ with $\langle {\alpha},R\rangle\leq 0$ or $\langle {\beta},R\rangle\leq 0$. That is, $${\operatorname{H}}^1(S\setminus 0, {{\mathcal O}}_S)(-R)={k}{\Leftrightarrow}R\in\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}\mbox{\ (and $=0$ otherwise).}$$ ### ${\operatorname{H}}^1(S\setminus 0, \theta_S)$ {#cechTheta} Here we use the derivations $\partial_a\in\theta(-\log\partial S)\subseteq\theta_S$ (with $\partial S:=S\setminus {{{\mathbb T}}}$) defined by $x^r\mapsto \langle a,r\rangle\cdot x^r$. Via $x^s\otimes a\mapsto x^s\partial_a$ they provide an isomorphism ${{\mathcal O}}_{S}\otimes_{\mathbb Z}N\stackrel{\sim}{\to} \theta(-\log\partial S)$ and, restricted to $S\setminus 0$, the latter sheaf equals $\theta_S$.\ In particular, the $1$-Čech cocycles of $\theta_S$ on $S\setminus 0$ are ${k}[M]\otimes N$, and the $1$-Čech coboundaries (in degree $-R$) are generated by $x^{-R}\partial_a$ with $\langle {\alpha},R\rangle\leq 1$ or $\langle {\beta},R\rangle\leq 1$. Thus, $${\operatorname{H}}^1(S\setminus 0, \theta_S)(-R)=N_{k}\mbox{ for } \langle {\alpha},R\rangle, \langle {\beta},R\rangle\geq 2,$$ and the remaining cases where ${\operatorname{H}}^1(S\setminus 0, \theta_S)(-R)$ is non-vanishing are $\langle {\alpha},R\rangle=1$, $\langle {\beta},R\rangle\geq 2$ (then it becomes $N_{k}/{k}\cdot{\alpha}$) and $\langle {\alpha},R\rangle\geq 2$, $\langle {\beta},R\rangle=1$ (yielding $N_{k}/{k}\cdot{\beta}$). Note that these cases include $R=r^2$ and $R=r^{e-1}$, respectively. ### The kernel {#cechKer} The $i$-th summand $dx^{r^i}$ maps $x^{-R}\partial_a$ to $\langle a,r^i\rangle\cdot x^{-R+r^i}$. In particular, whenever $R-r^i\in\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}$, then $dx^{r^i}$ imposes the codimension one condition $\langle a,r^i\rangle=0$ on the derivation $x^{-R}\partial_a$.\ [*Case 1*]{}. Assume that $\langle {\alpha},R\rangle, \langle {\beta},R\rangle\geq 2$. Each occurrence of at least two conditions $R-r^i,R-r^j\in\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}$ enforces $a=0$. Using the numbering of the beginning of (\[degT1\]), the remaining cases are\ (ii) where this does not happen at all yielding ${{T^1}}(-R)=N_{k}$, and\ (iii) where $(k\cdot r^i)-r^i=(k-1)\cdot r^i\in\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}$ leads to the single condition $\langle a, r^i\rangle =0$. There we are left with a one-dimensional ${{T^1}}(-R)=(r^i)^\bot\subset N_{k}$.\ [*Case 2*]{}. Assume that $\langle {\alpha},R\rangle=1$ and $\langle {\beta},R\rangle\geq 2$. Then, either $\langle {\beta},R\rangle > \langle{\beta},r^1\rangle=n$, i.e.$R-r^1\in\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}$ implying the condition $\langle a, r^1\rangle =0$ forcing $a\in N/{\alpha}{\mathbb Z}$ to become $0$, or, using the numbering of (\[degT1\]) again,\ (i) $R=r^2$ with ${{T^1}}(-R)=N_{k}/{k}\cdot{\alpha}$.\ The case $\langle {\alpha},R\rangle\geq 2$, $\langle {\beta},R\rangle=1$ (yielding $R=r^{e-1}$) works similar. The construction of $X_\xi\setminus 0$ {#constX} -------------------------------------- Let $\xi\in{\operatorname{H}}^1(S\setminus 0,\theta_S)$ be given by the $1$-Čech cocycle $\xi=x^{-R}\partial_a\in {\Gamma}({{{\mathbb T}}},\theta_S|_{{{\mathbb T}}})={k}[M]\otimes N$, cf. (\[cechTheta\]). The associated infinitesimal deformation $X_\xi\setminus 0$ of $S\setminus 0$ arises from glueing the trivial pieces ${{\mathbb T}{\mathbb V}}({\alpha})\otimes{k}[{\varepsilon}]$ and ${{\mathbb T}{\mathbb V}}({\beta})\otimes{k}[{\varepsilon}]$ along the ${k}[{\varepsilon}]$-algebra map $$\xymatrix@R=2.6ex@C=3.5em{ {k}[{{\alpha}}{^{\scriptscriptstyle\vee}}\cap M,\;{\varepsilon}] \ar@{_(->}[d] & {k}[{{\beta}}{^{\scriptscriptstyle\vee}}\cap M,\;{\varepsilon}] \ar@{^(->}[d] \\ {k}[M,{\varepsilon}] \ar[r]^-{\varphi_\xi} & {k}[M,{\varepsilon}] } \vspace{0.5ex}$$ with $\,\varphi_\xi(x^r):=x^r+{\varepsilon}\cdot\xi(x^r)= x^r + {\varepsilon}\cdot\langle a,r\rangle\cdot x^{r-R}$. Note that we have decided to use the notation $X_\xi\setminus 0$ even in the case when there is no extension of this to some deformation $X_\xi$ of the non-punctured $S$. The dualizing sheaf on $X_\xi\setminus 0$ {#omegaXPunct} ----------------------------------------- Let $\xi=x^{-R}\partial_a$ as before. Since $X_\xi\setminus 0$ is smooth over $\operatorname{Spec}{k}[{\varepsilon}]$ and since $\omega_{{k}[{\varepsilon}]}={k}[{\varepsilon}]$, it follows from [@HartRD p.140] that $$\omega_{X\setminus 0}=\omega_{(X\setminus 0)|{k}[{\varepsilon}]} = \Lambda^2\Omega^1_{(X\setminus 0)|{k}[{\varepsilon}]}.$$ Choosing a ${\mathbb Z}$-basis $\{{A},{B}\}$ of $M$, the local pieces of the latter equal $$\textstyle \omega_{{\alpha}}= \oplus_{\langle {\alpha},r\rangle\geq 1}\,{k}[{\varepsilon}]\cdot x^r \cdot\frac{dx^{A}}{x^{A}}\wedge\frac{dx^{B}}{x^{B}} \cong \oplus_{\langle {\alpha},r\rangle\geq 1}\,{k}[{\varepsilon}]\cdot x^r\subseteq {k}[M,{\varepsilon}]$$ and similarly for $\omega_{\beta}$, cf. [@CoxLittleSchenck Prop. 8.2.9]. Note that the isomorphism does, up to sign, not depend on the choice of $\{{A},{B}\}$. Now, we determine the impact of the ${k}[{\varepsilon}]$-algebra isomorphism $\varphi_\xi^{[0]}:=\varphi_\xi$ on the glueing $\varphi_\xi^{[1]}$ of the modules $\omega_{{\alpha}}|_{{{\mathbb T}}}$ and $\omega_{{\beta}}|_{{{\mathbb T}}}$. Since $$\textstyle \frac{dx^{A}}{x^{A}}\mapsto \frac{d(x^{A}+{\varepsilon}\langle a,{A}\rangle x^{{A}-R})} {(x^{A}+{\varepsilon}\langle a,{A}\rangle x^{{A}-R}} = \frac{dx^{A}}{x^{A}} + {\varepsilon}\langle a,{A}\rangle\, dx^{-R},$$ we obtain that $x^r\cdot \frac{dx^{A}}{x^{A}}\wedge\frac{dx^{B}}{x^{B}}$ maps to $$\textstyle \big(x^r\cdot \frac{dx^{A}}{x^{A}}\wedge\frac{dx^{B}}{x^{B}}\big) + {\varepsilon}\,x^{r-R}\big( \langle a,r\rangle \frac{dx^{A}}{x^{A}}\wedge\frac{dx^{B}}{x^{B}} + \langle a,{A}\rangle \frac{dx^{-R}}{x^{-R}}\wedge\frac{dx^{B}}{x^{B}} + \langle a,{B}\rangle \frac{dx^{{A}}}{x^{{A}}}\wedge\frac{dx^{-R}}{x^{-R}} \big).$$ Expressing $R$ within the basis $\{{A},{B}\}$ (and suppressing $\frac{dx^{A}}{x^{A}}\wedge\frac{dx^{B}}{x^{B}}$) finally yields $$\textstyle \varphi_\xi^{[1]}:\hspace{1em} \omega_{{\alpha}}|_{{{\mathbb T}}}\ni \; x^r \;\longmapsto\; x^r+{\varepsilon}\cdot\langle a,r-R\rangle\cdot x^{r-R} \; \in \omega_{{\beta}}|_{{{\mathbb T}}}.$$ This description enables us to determine the class $[\omega_{X\setminus 0}]\in {\operatorname{H}}^1(S\setminus 0, {{\mathcal O}}_X^*)$. If $x^{s}$ and $x^{t}$ are generators of $\omega_{{\alpha}}$ and $\omega_{{\beta}}$, respectively, i.e. if $\langle {\alpha},{s}\rangle = \langle {\beta},{t}\rangle=1$, then $[\omega_{X\setminus 0}]$ is represented by the $1$-Čech cocycle $\varphi(x^{s})/x^{t}\in{\Gamma}({{{\mathbb T}}},{{\mathcal O}}_{{{{\mathbb T}}}\otimes{k}[{\varepsilon}]}^*)$. It is equal to $$\psi_1:=x^{{s}-{t}} + {\varepsilon}\cdot\langle a,{s}-R\rangle\cdot x^{{s}-{t}-R}.$$ Similarly, we might consider the glueing map $\varphi_\xi^{[{g}]}$ for a reflexive power $\omega_{X\setminus 0}^{[{g}]}$ instead of just for $\omega_{X\setminus 0}$. Then, the previous calculations yield \[lem-cocycle\] $ \textstyle \,\varphi_\xi^{[{g}]}: x^r \mapsto x^r+{\varepsilon}\cdot\langle a,r-{g}R\rangle\cdot x^{r- R}, $ and the $1$-Čech cocycle becomes $$\psi_{g}:= x^{{s}({g})-{t}({g})} + {\varepsilon}\cdot\langle a,{s}({g})-{g}R\rangle\cdot x^{{s}({g})-{t}({g})-R} \in {\Gamma}({{{\mathbb T}}},{{\mathcal O}}_{{{{\mathbb T}}}\otimes{k}[{\varepsilon}]}^*) \vspace{-0.5ex}$$ with ${s}({g}),{t}({g})\in M$ satisfying $\langle {\alpha},{s}({g})\rangle = \langle {\beta},{t}({g})\rangle={g}$. Note that one might take, if some ${s}={s}(1)$ and ${t}={t}(1)$ are available, the multiples ${s}({g})={g}\cdot{s}$ and ${t}({g})={g}\cdot{t}$ for a general ${g}\in{\mathbb Z}$. However, in (\[emailVersion\]) we will prefer a different choice for ${g}={m}$. Extending functions along codimension two {#extCodim2} ----------------------------------------- Let $S={{\mathbb T}{\mathbb V}}(\sigma)$ be as before. Since it is normal, it carries the Hartogs property $S_2$ as it was asked for in (\[QG.V.defn\]). Now, if $A$ is an Artinian ${k}$-algebra and $X$ is a deformation of $S$ over $A$ (we just need the case $A={k}[{\varepsilon}]$ here), we would like to keep this property. \[lem-reflexive\] If ${{\mathcal F}}$ is a reflexive ${{\mathcal O}}_X$-module, i.e.if ${{\mathcal F}}={Hom}_{{{\mathcal O}}_X}({{\mathcal G}},{{\mathcal O}}_X)$ for some ${{\mathcal O}}_X$-module ${{\mathcal G}}$, then $\,{\Gamma}(S,{{\mathcal F}})\to{\Gamma}(S\setminus 0,\,{{\mathcal F}})$ is an isomorphism. It suffices to check this for ${{\mathcal F}}={{\mathcal O}}_X$. We proceed by induction. Choosing a non-trivial element ${\varepsilon}\in A$ with ${\varepsilon}\cdot{\mathfrak{m}}_A=0$, we obtain an exact sequence of $A$-modules $$0 \to {k}\stackrel{\cdot{\varepsilon}}{\to} A \to {\overline}{A}\to 0.$$ Denoting ${\overline}{X}:=X\otimes_A{\overline}{A}$, flatness, restriction to $S\setminus 0$, and taking global sections provides the the following commutative diagram with exact rows: $$\xymatrix@R=2.6ex@C=1.5em{ 0 \ar[r] & {\Gamma}(S,{{\mathcal O}}_S) \ar[r]^-{\cdot{\varepsilon}} \ar[d]^-{\sim}& {\Gamma}(S,{{\mathcal O}}_X) \ar[r] \ar[d] & {\Gamma}(S,{{\mathcal O}}_{{\overline}{X}}) \ar[r] \ar[d]^-{\sim}& 0 \ar[d] \\ 0 \ar[r] & {\Gamma}(S\setminus 0,\,{{\mathcal O}}_S) \ar[r]^-{\cdot{\varepsilon}} & {\Gamma}(S\setminus 0,\,{{\mathcal O}}_X) \ar[r] & {\Gamma}(S\setminus 0,\,{{\mathcal O}}_{{\overline}{X}}) \ar[r]^-{0} & {\operatorname{H}}^1(S\setminus 0,\,{{\mathcal O}}_S). }$$ Now, the claim follows from the 5-lemma. The dualizing sheaf on $X_\xi$ {#omegaX} ------------------------------ In contrast to (\[constX\]) we now start with a $$\xi=x^{-R}\,\partial_a \in {{T^1}}(-R)\subseteq {\operatorname{H}}^1(S\setminus 0,\,\theta_S)(-R) {\leftarrow\hspace{-0.8em}\leftarrow}{k}[M] \otimes_{\mathbb Z}N$$ yielding a true $X=X_\xi$ and not just a punctured $X_\xi\setminus 0$. The extension theorem along two-codimensional subsets gives us the right tool to understand $\omega_X^{[{g}]}$ out of $\omega_{X\setminus 0}^{[{g}]}$ we have obtained in (\[omegaXPunct\]), namely $$\textstyle \omega^{[{g}]}_{X} = {\Gamma}(S\setminus 0,\,\omega^{[{g}]}_{X}) = \{f= \sum_{{\alpha}\geq {g}} (c_r+{\varepsilon}d_r) \,x^r{\,|\;}\varphi_\xi^{[{g}]}(f)\in \oplus_{{\beta}\geq {g}} \,{k}[{\varepsilon}]\cdot x^r\}.$$ From Lemma \[lem-cocycle\] we know that $$\textstyle \varphi_\xi^{[{g}]}\big(\sum_{r} (c_r+{\varepsilon}d_r) x^r\big) = \sum_{r} c_r\,x^r + {\varepsilon}\cdot\big( \sum_r d_r\,x^r + c_r \langle a,r-{g}R\rangle\cdot x^{r-R}\big),$$ and the combination of these two statements yields \[lem-omegaXg\] The sum $f= \sum_{r\in M} (c_r+{\varepsilon}d_r)\, x^r$ belongs to $\omega^{[{g}]}_{X}$ if and only if the following two assertions hold 1. $\langle{\alpha},r\rangle<{g}$ implies $c_r=d_r=0$ 2. $\langle{\beta},r\rangle<{g}$ implies $c_r=0$ and $d_r=c_{r+R}\cdot\big\langle a,\,({g}-1)R-r\big\rangle$. Surjectivity of the restriction map ${{\mathcal R}}_\xi^{[{g}]}$ {#surjOmega} ---------------------------------------------------------------- Recall from (\[QG.V.defn\]) that one of the characterizations of the property ${\mbox{$(*)_{g}$}}$ for $X=X_\xi$ was the surjectivity of the restriction map ${{\mathcal R}}_\xi^{[{g}]}:\omega_X^{[{g}]}\to\omega_S^{[{g}]}$. Since this map just sends ${\varepsilon}\mapsto 0$, i.e.$$\textstyle {{\mathcal R}}_\xi^{[{g}]}:\; \sum_{r} (c_r+{\varepsilon}d_r) \,x^r \mapsto \sum_{r} c_r \,x^r$$ and $\omega_S^{[{g}]}=\{ \sum_{\langle{\alpha},r\rangle, \langle{\beta},r\rangle \geq {g}} c_r\,x^r\}$ by [@CoxLittleSchenck Prop. 8.2.9], we can characterize this property by \[lem-surjOmega\] $\xi=x^{-R}\,\partial_a$ satisfies ${\mbox{$(*)_{g}$}}$ ${\Leftrightarrow}$ each $r\in M$ with both $${g}\leq\langle {\alpha},r\rangle<{g}+\langle {\alpha},R\rangle \hspace{0.8em}\mbox{and}\hspace{0.8em} {g}\leq\langle {\beta},r\rangle<{g}+\langle {\beta},R\rangle$$ leads to $\langle a, {g}R-r\rangle =0$. The second part of Condition (ii) for $f\in \omega^{[{g}]}_{X}$ in Lemma \[lem-omegaXg\] can be read as that $\langle{\beta},r-R\rangle<{g}$ implies $d_{r-R}=c_{r}\cdot\big\langle a,\,{g}R-r\big\rangle$. Hence, together with $\langle{\alpha},r-R\rangle<{g}$, this would enforce that $c_{r}\cdot\big\langle a,\,{g}R-r\big\rangle=0$. On the other hand, if $\langle{\alpha},r\rangle, \langle{\beta},r\rangle \geq {g}$, then $c_r\neq 0$ is allowed in $\omega_S^{[{g}]}$. For given $R\in M$ and ${g}\in{\mathbb Z}$ we define the following zones within $M_{\mathbb Q}$: $$Z_{R,\,{g}}:= \{r\in M_{\mathbb Q}{\,|\;}{g}\leq\langle {\alpha},r\rangle<{g}+\langle {\alpha},R\rangle \hspace{0.5em}\mbox{and}\hspace{0.5em} {g}\leq\langle {\beta},r\rangle<{g}+\langle {\beta},R\rangle\}.$$ Then, for $\xi=x^{-R}\,\partial_a$, the previous lemma says that $${\mbox{$(*)_{g}$}} \;{\Leftrightarrow}\; {{\mathcal R}}_\xi^{[{g}]} \mbox{ is surjective } \;{\Leftrightarrow}\; Z_{R,\,{g}}\cap M\subseteq a^\bot+{g}R.$$ Actually, up to the point that $X_\xi$ does not make sense otherwise, we did not use $\xi\in {{T^1}}$ so far. This property is equivalent to ${\mbox{$(*)_0$}}$, and it will be discussed in (\[zonesT1\]). V-deformations {#V-deformations} ============== Let $\sigma=\langle{\alpha},{\beta}\rangle$ be as before, e.g.it can be obtained as the cone $C(I)$ over an interval with uniform denominators $I=[\frac{g}{{m}},\frac{h}{{m}}]$ as in (\[coneInterval\]).\ In the present section, we will approach the V-deformations of $S={{\mathbb T}{\mathbb V}}(\sigma)$ defined in (\[def-V\]). Since ${\operatorname{index}}(\omega_S)={m}$, we will mostly study the property ${\mbox{$(*)_{m}$}}$ for a given infinitesimal deformation $\xi=x^{-R}\partial_a$. Shifting the zones {#shiftZones} ------------------ Recall from Remark \[rem-index\] that $\frac{{r^1}+{r^e}}{n}=\frac{1}{{m}}\cdot{\overline}{R}\in M_{\mathbb Q}$ is the truly canonical (but rational) degree. Moreover, depending on $R\in M$ we denote $$\textstyle Z_R := Z_{R,\,0}= \sigma{^{\scriptscriptstyle\vee}}\cap(R-\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}) \subseteq M_{\mathbb Q}.$$ The degrees $R$ we are interested in are always elements of $\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}$. In particular, $Z_R$ is then a bounded region – it is a half-open parallelogram having $0$ and $R$ as opposite vertices. While these two vertices belong to the lattice $M$, the remaining ones usually do not. The relation to the zones $Z_{R,\,{g}}$ from (\[surjOmega\]) is $$\textstyle Z_{R,\,{g}}= \frac{{g}}{{m}}\,{\overline}{R} +Z_R \hspace{1.5em}\mbox{for every}\hspace{0.5em} {g}\in{\mathbb Z}.$$ In particular, $Z_{R,\,{g}+ {m}}= {\overline}{R} + Z_{R,\,{g}}$, i.e. the zones $Z_{R,\,{g}+{\mathbb Z}{m}}$ just differ by integral translation. This gives rise to define the “stable” condition $$\textstyle {\mbox{${\overline}{(*)}_{g}$}} := \bigcap_{\ell\in{\mathbb Z}} \hspace{0.2em} {\mbox{$(*)_{{g}+\ell {m}}$}} \hspace{1.5em}\mbox{(still being a condition for $\xi=x^{-R}\partial_a$).}$$ That is, the condition that ${\mbox{$(*)_{g}$}}$ is true for all ${g}\in{\mathbb Z}$ can be replaced by the finite one asking for ${\mbox{${\overline}{(*)}_{g}$}}$ for all ${g}\in{\mathbb Z}/{m}{\mathbb Z}$. Moreover, to be a V-deformation in the sense of (\[def-V\]) means to fulfill the condition ${\mbox{${\overline}{(*)}_0$}}$. \[prop-shift\] [1)]{} If $Z_{R,\,{g}}\cap M =\emptyset$, then ${\mbox{${\overline}{(*)}_{g}$}}$ is fulfilled for each $a\in N$.\ [2)]{} Assume that $Z_{R,\,{g}}\cap M\neq\emptyset$. Then the following conditions are equivalent: $${\mbox{${\overline}{(*)}_{g}$}} \;{\Leftrightarrow}\; {\mbox{$(*)_{{g}+\ell {m}}$}} \mbox{\rm\ for two different } \ell\in{\mathbb Z}\;{\Leftrightarrow}\; {\mbox{$(*)_{g}$}} \mbox{\rm\ and } a\in({\overline}{R}-{m}R)^\bot.$$ Let $\ell\in{\mathbb Z}$. If $r\in Z_{R,\,{g}}\cap M$, then $r+\ell\, {\overline}{R}\in Z_{R,\,{g}+\ell {m}}$. Hence, the conditions ${\mbox{$(*)_{g}$}}$ and ${\mbox{$(*)_{{g}+\ell {m}}$}}$ mean that $$\langle a,\,{g}R-r\rangle =0 \hspace{1em}\mbox{and}\hspace{1em} \langle a,\,({g}+\ell {m})R-(r+\ell {\overline}{R})\rangle =0,$$ respectively. However, the difference of the two left hand sides equals $\langle a, \,-\ell {m}R +\ell {\overline}{R}\rangle = \ell\cdot\langle a,\,{\overline}{R}-{m}R\rangle$. \[cor-shift\] [1)]{} If $\,\xi=x^{-R}\partial_a$ is a V-deformation, then $a\in({\overline}{R}-{m}R)^\bot$. [(See the upcoming Corollary \[cor-nqG\] for a stronger statement.)]{}\ [2)]{} Assume that $a\in({\overline}{R}-{m}R)^\bot\setminus\{0\}$. Then, for any ${g}\in{\mathbb Z}$, ${\mbox{$(*)_{g}$}}$ [(]{}or even ${\mbox{${\overline}{(*)}_{g}$}} $[)]{} is equivalent to $(Z_{R,\,{g}}\cap M)-{g}R\subseteq{\mathbb Q}\cdot({\overline}{R}-{m}R)$. Likewise, this condition is equivalent to $(Z_{R,\,{g}}\cap M)-\frac{{g}}{{m}}\,{\overline}{R}\subseteq{\mathbb Q}\cdot({\overline}{R}-{m}R)$. \(1) follows from the fact that there are non-empty $(Z_{R,\,\ell{m}}\cap M)$ whenever $R\in\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}$ (and only those $R$ matter for ${{T^1}}(-R)\neq 0$): Just take $\ell=0$.\ (2) ${\mbox{$(*)_{g}$}}$ means that for each $r\in Z_{R,\,{g}}\cap M$ we have $a\in (r-{g}R)^\bot$. Together with $a\in({\overline}{R}-{m}R)^\bot$ this means that $a$ can be non-trivial if and only if both $r-{g}R$ and ${\overline}{R}-{m}R$ are collinear. Moreover, ${\overline}{R}-{m}R$ does never vanish (since ${{\mathbb T}{\mathbb V}}(\sigma)\neq A_k$). Focusing on ${{T^1}}$-degrees {#zonesT1} ----------------------------- For investigating the ${\mbox{$(*)_{g}$}}$ property we did not use yet that the set of degrees $R\in M$ with ${{T^1}}(-R)\neq 0$ is very restricted. Taking this into account implies \[lem-deg0\] Every deformation $x^{-R}\partial_a\in {{T^1}}(-R)$ satisfies ${\mbox{$(*)_0$}}$. Actually, this statement is trivial – the condition ${\mbox{$(*)_0$}}$ means that $\omega_X^{[0]}={{\mathcal O}}_X$ is flat over $k[{\varepsilon}]$, i.e. it even characterizes the elements of ${{T^1}}(-R)$. Nevertheless, e.g. to practice our new language involving the zones $Z_R$, we would like to present a direct argument, too:\ Condition ${\mbox{$(*)_0$}}$ means $Z_R\cap M\subseteq a^\bot$ with $Z_R=\sigma{^{\scriptscriptstyle\vee}}\cap(R-\sigma{^{\scriptscriptstyle\vee}})$. According to (\[degT1\]), we distinguish between two cases:\ (i)+(ii) $\,R=r^i$ with $i=2,\ldots,e-1$: Since these elements are irreducible in the semigroup $\sigma{^{\scriptscriptstyle\vee}}\cap M$, we obtain $Z_R\cap M=\{0\}$, and this belongs to every $a^\bot$.\ (iii) $\,R=k\cdot r^i$ for $i=2,\ldots,e-1$ with $2\leq k\leq a_i-1$: Here we have $$Z_R\cap M=\{0,r^i,\ldots,(k-1)r^i\},$$ i.e. Condition ${\mbox{$(*)_0$}}$ means $\langle a,r^i\rangle =0$. However, by (\[cechKer\]), Case 1, exactly this is ensured to hold true within ${{T^1}}(-k\cdot r^i)$. [*Remark.*]{} Actually, the condition $${\mbox{$(*)_0$}} \;{\Leftrightarrow}\; Z_R\cap M\subseteq a^\bot \;{\Leftrightarrow}\; a\in (Z_R\cap M)^\bot$$ should be understood as an alternative description of ${{T^1}}(-R)$. However, this is not new – it coincides with the description in [@flip (2.2)]. There, one has defined the finite subsets $$E_{\alpha}^R:=\{r\in E{\,|\;}\langle{\alpha}, r\rangle < \langle{\alpha}, R\rangle\} \hspace{1em}\mbox{and}\hspace{1em} E_{\beta}^R:=\{r\in E{\,|\;}\langle{\beta}, r\rangle < \langle{\beta}, R\rangle\}$$ of $M$, and this lead to an exact sequence $$0 \to {{T^1}}(-R) \to\big({\operatorname{span}}_k E_{\alpha}^R\cap {\operatorname{span}}_k E_{\beta}^R\big)^* \to {\operatorname{span}}_k(E_{\alpha}^R\cap E_{\beta}^R)^* \to 0.$$ In particular, ${{T^1}}(-R)$ is a subquotient of $N_k$, and $a\in {{T^1}}(-R)$ if and only if $a\in(E_{\alpha}^R\cap E_{\beta}^R)^\bot$. Now, the relation to our condition ${\mbox{$(*)_0$}}$ is that $E_{\alpha}^R\cap E_{\beta}^R=Z_R\cap E$. \[cor-nqG\] $\xi=x^{-R}\partial_a\in {{T^1}}(-R)$ is a V-deformation, i.e. it fulfills the stable condition ${\mbox{${\overline}{(*)}_0$}}$, if and only if $a\in ({\overline}{R}-{m}R)^\bot$. This follows from Proposition \[prop-shift\](2). The implication (${\Rightarrow}$) was already stated in Corollary \[cor-shift\]. The reversed implication (${\Leftarrow}$) makes use of Lemma \[lem-deg0\]. Counting V-deformations {#restrictPhi} ----------------------- We run through the list (i)-(iii) of (\[degT1\]) and especially (\[cechKer\]) to determine $({\overline}{R}-{m}R)^\bot=({r^1}+{r^e}-nR)^\bot$, i.e. the V-deformations within each homogeneous summand ${{T^1}}(-R)$.\ (i) $R=r^2$ (and similarly $R=r^{e-1}$): $\,{{T^1}}(-r^2)=N_{k}/{k}\cdot{\alpha}$. The element ${r^1}+{r^e}-n r^2$ is contained in ${{\alpha}}^\bot\subset M$, hence it provides a linear map $N/{\mathbb Z}\cdot{\alpha}\to{\mathbb Z}$ where ${{T^1_V}}(-r^2)$ is generated by the kernel. Since we had excluded the $A_{n-1}$-singularity, this linear map is also non-trivial, i.e. there is no V-deformations in degree $-r^2$ (and $-r^{e-1}$).\ (ii) $R=r^i$ for $i=3,\ldots,e-2$: $\; {{T^1}}(-r^i)=N_{k}$. We know that ${r^1}+{r^e}-n r^i$ is again non-trivial, hence it provides a one-dimensional kernel within the two-dimensional ${{T^1}}(-r^i)$. Altogether, this yields an $(e-4)$-dimensional space of V-deformations.\ (iii) $R=k\cdot r^i$ for $i=2,\ldots,e-1$ with $2\leq k\leq a_i-1$: $\; {{T^1}}(-kr^i)=(r^i)^\bot\subset N_{k}$. Here we obtain ${{T^1_V}}(-kr^i)=({r^1}+{r^e},\, r^i)^\bot$, i.e.this is non-trivial if and only if ${r^1}+{r^e}\in{\mathbb N}\cdot r^i$, i.e. if $\sigma$ is grounded (see Definition \[def-grounded\]) and $r^i={\overline}{R}$ is the central degree (i.e. $i={\nu}$ is the central index). If this is the case, and if $\sigma$ stems from an interval $I=[-{A},{B}]$ as in Proposition \[prop-etaci\], then we gather another $a_{\nu}-2={{\lfloor {A}\rfloor}}+{{\lfloor {B}\rfloor}}\,$ . Representing ${{T^1_V}}$ as a kernel {#emailVersion} ------------------------------------ An alternative approach to visualize the V-deformations of $S={{\mathbb T}{\mathbb V}}(\sigma)$ is to consider the following map $\Phi:{{T^1}}\to {\operatorname{H}}^2_0(S,{{\mathcal O}}_S)$: If $\xi\in {{T^1}}$ is represented by an infinitesimal deformation $X=X_\xi\to\operatorname{Spec}{k}[{\varepsilon}]$, then there is an exact sequence $$\xymatrix@R=0.4ex@C=1.5em{ 0 \ar[r] & {{\mathcal O}}_S \ar[r]^{} & {{\mathcal O}}_X^* \ar[r] & {{\mathcal O}}_S^* \ar[r] & 1 \\ & f \ar@{|->}[r] & 1+{\varepsilon}f. }$$ Since the map ${\operatorname{H}}^0(S\setminus 0,{{\mathcal O}}_X^*)\to {\operatorname{H}}^0(S\setminus 0,{{\mathcal O}}_S^*)$ equals ${{\mathcal O}}_X^*{\rightarrow\hspace{-0.8em}\rightarrow}{{\mathcal O}}_S^*$ on the affine $S$, i.e. it is notably surjective, this implies the exactness of $$0 \to {\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S) \to {\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_X^*) \to {\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S^*).$$ Thus, using that $\omega_S^{[{m}]}$ is trivial, we may define $\Phi(\xi)$ as the class of $\omega_X^{[{m}]}$ in $$\xymatrix@R=0.6ex@C=1.5em{ {\operatorname{H}}^2_0(S,{{\mathcal O}}_S)={\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S) = \ker\Big( \hspace{-2em} & {\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_X^*) \ar[r] & {\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S^*)& \hspace{-3.5em}\Big) \\ & {\operatorname{Pic}}(X_\xi\setminus 0)\ar@{=}[u] \ar[r] & {\operatorname{Pic}}(S\setminus 0). \ar@{=}[u] }$$ Comparing with the flatness part of the definition at the end of (\[QG.V.defn\]), it follows that the kernel $\ker\Phi\subseteq {{T^1}}$ consists exactly of the deformations satisfying ${\mbox{$(*)_{m}$}}$, i.e., of the V-deformations, cf. (\[def-V\]).\ It turns out that $\Phi$ can be extended to ${\operatorname{H}}^1(S\setminus 0,\theta_S)$, i.e.we consider (locally trivial) deformations $X_\xi\setminus 0$ of (the smooth) $S\setminus 0$ again. Using the descriptions of ${\operatorname{H}}^1(S\setminus 0,\theta_S)$ and ${\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S)$ given in \[cechTheta\] and \[cechO\], respectively, the final result fits perfectly with Corollary \[cor-nqG\]: \[prop-calcPhi\] Let $R\in\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}\cap M$. Then, the degree $-R$ part of $\Phi$ $$\Phi(-R):{\operatorname{H}}^1(S\setminus 0,\theta_S)(-R)\to {\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S)(-R)$$ is given by $\,x^{-R}\partial_a\mapsto \langle a,{\overline}{R}-{m}R\rangle\cdot x^{-R}$. In other words, using the natural maps $N_{k}{\rightarrow\hspace{-0.8em}\rightarrow}{\operatorname{H}}^1(S\setminus 0,\theta_S)(-R)$ and ${\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S)(-R)={k}$, the map $\Phi(-R):N_k\to k$ equals ${\overline}{R}-{m}R\in M$. In (\[omegaXPunct\]) we have dealt with 1-cocycles of ${{\mathcal O}}_X^*$, and in Lemma \[lem-cocycle\] we have obtained an element $\psi_{m}$ describing the class of $\omega_X^{[{m}]}$ after using the surjection $\Gamma({{{\mathbb T}}},{{\mathcal O}}_{{{{\mathbb T}}}\otimes k[{\varepsilon}]}^*){\rightarrow\hspace{-0.8em}\rightarrow}{\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_X^*)$. Restricting $\psi_{m}$ via ${\varepsilon}\mapsto 0$ to $\Gamma({{{\mathbb T}}},{{\mathcal O}}_{{{\mathbb T}}}^*){\rightarrow\hspace{-0.8em}\rightarrow}{\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S^*)$ yields $x^{{s}({m})-{t}({m})}$.\ Since $\omega_S^{[{m}]}={{\mathcal O}}_S$, this is a $1$-Čech coboundary. One can see this directly by the possibility of choosing ${s}({m})={t}({m})={\overline}{R}$ – then $x^{{s}({m})-{t}({m})}$ becomes $1$ right away. Applying this recipe to the original cocycle $\psi_{m}$ of Lemma \[lem-cocycle\] as well, we obtain that $$\psi_{m}= 1 + {\varepsilon}\cdot\langle a,\,{\overline}{R}-{m}R\rangle\cdot x^{-R}.$$ Recall that the second map within the exact sequence $$0 \to {{\mathcal O}}_{S\setminus 0} \to {{\mathcal O}}_{X\setminus 0}^* \to {{\mathcal O}}_{S\setminus 0}^* \to 1$$ sends $f\mapsto 1+{\varepsilon}\cdot f$. Thus, $\Phi(\xi)=[\omega_{X\setminus 0}^{[{m}]}]\in {\operatorname{H}}^1(S\setminus 0,{{\mathcal O}}_S)(-R)$ is given by the $1$-Čech cocycle $\langle a,\,{\overline}{R}-{m}R\rangle\cdot x^{-R}\in {k}[M]={\Gamma}({{{\mathbb T}}},{{\mathcal O}}_S)$. - and -deformations {#compJanos} =================== V-deformations are understood, by Corollary \[cor-nqG\] and (\[restrictPhi\]). Next we will turn to the stronger - and -deformations, cf. (\[def-qG\]) and (\[def-VW\]) for the definition of these notions. Extending the lattice $M$ {#extendM} ------------------------- We adopt the notation from (\[coneInterval\]). For $I=[\frac{g}{{m}},\frac{h}{{m}}]$ we know that $\frac{1}{{m}}{\overline}{R}\in M_{\mathbb Q}$ is the truly canonical, but rational degree. This gives rise to an enlargement of our lattice $M$, namely $$\textstyle {\widetilde{M}}:=M+{\mathbb Z}\cdot \frac{1}{{m}}{\overline}{R}.$$ Let us assume that $\xi=x^{-R}\partial_a\in {{T^1}}(-R)$ is a V-deformation, i.e. $\langle a,\,{\overline}{R}-{m}R\rangle=0$. Using the zone $Z_R=\sigma{^{\scriptscriptstyle\vee}}\cap(R-\operatorname{int}\sigma{^{\scriptscriptstyle\vee}})$ defined in (\[shiftZones\]), we obtain \[prop-WqG\] [1)]{} $\xi$ is a -deformation ${\Leftrightarrow}$ ${\widetilde{M}}\cap Z_R\subseteq {\mathbb Q}\cdot ({\overline}{R}-{m}R)$, and\ [2)]{} $\xi$ is a -deformation ${\Leftrightarrow}$ $(M+\frac{1}{{m}}{\overline}{R})\cap Z_R\subseteq {\mathbb Q}\cdot ({\overline}{R}-{m}R)$. Note that the difference between both cases just arises from the tiny difference between ${\widetilde{M}}=M+{\mathbb Z}\cdot \frac{1}{{m}}{\overline}{R}$ and $M+\frac{1}{{m}}{\overline}{R}$. Recall from (\[shiftZones\]) that $Z_{R,\,{g}}= \frac{{g}}{{m}}\,{\overline}{R} +Z_R$ for ${g}\in{\mathbb Z}$. Thus, Corollary \[cor-shift\](2) says that the conditions ${\mbox{$(*)_{g}$}}$ and ${\mbox{${\overline}{(*)}_{g}$}}$ are equivalent to $$\textstyle Z_R\cap (M-\frac{{g}}{{m}}{\overline}{R})= ((Z_{R} + \frac{{g}}{{m}}{\overline}{R})\cap M)- \frac{{g}}{{m}}{\overline}{R} \subseteq{\mathbb Q}\cdot({\overline}{R}-{m}R).$$ While ${g}=-1$ directly leads to (2), one uses $\bigcup_{{g}\in{\mathbb Z}}(M-\frac{{g}}{{m}}{\overline}{R})={\widetilde{M}}$ for (1). Now, we are going to scan the degrees of ${{T^1_V}}$ listed in (\[restrictPhi\])(ii) and (iii) for - and -deformations. (Note that the deformations in (\[restrictPhi\])(i) are not even V-deformations.)\ Actually, it is convenient to proceed with a minor change to the division into the two cases: We will shift (and this applies only to the grounded case) the central degree ${\overline}{R}=r^{\nu}$ from Class (ii) to (iii). Thus, in (ii) we now collect exactly the non-central $R=r^i$ ($i=3,\ldots,e-2$), and Class (iii) will gather all $R=k\cdot r^{\nu}$ with $1\leq k\leq a_{\nu}-1$. Note that this set is empty unless $\sigma$ is grounded, i.e. $r^{\nu}={\overline}{R}$. The degrees of (\[restrictPhi\])(ii) {#scanII} ------------------------------------ Let $R=r^i$ with $i=3,\ldots,e-2$ be a non-central degree. The latter property can be expressed by $$\textstyle \frac{1}{{m}}{\overline}{R}\;\notin\;{\mathbb Q}\cdot ({\overline}{R}-{m}r^i).$$ On the other hand, we know that $$\textstyle \langle{\alpha},\frac{1}{{m}}{\overline}{R}\rangle= \langle{\beta},\frac{1}{{m}}{\overline}{R}\rangle=1 \hspace{0.7em}\mbox{and}\hspace{0.7em} \langle{\alpha},r^i\rangle, \langle{\beta},r^i\rangle >1$$ which implies that $\frac{1}{{m}}{\overline}{R}\,\in\, \sigma{^{\scriptscriptstyle\vee}}\cap(r^i-\operatorname{int}\sigma{^{\scriptscriptstyle\vee}}) = Z_{r^i}$. Hence, $$\textstyle \frac{1}{{m}}{\overline}{R}\;\in\; (M+\frac{1}{{m}}{\overline}{R})\cap Z_{r^i}.$$ Applying Proposition \[prop-WqG\](2), this shows that the deformations of degree $r^i$ cannot be -deformations, let alone -deformations. In other words, the property of being a grounded singularity is a necessary condition for the existence of - or -deformations. The degrees of (\[restrictPhi\])(iii) {#scanIII} ------------------------------------- Let $\sigma$ be a grounded cone with central degree ${\overline}{R}=r^{\nu}$. From (\[coneInterval\]) and (\[groundedCones\]) we know that $\sigma=\langle{\alpha},{\beta}\rangle$ can be obtained as $C(I)=\langle (g,{m}),\,(h,{m})\rangle$ from a grounded interval $I=[\frac{g}{{m}},\frac{h}{{m}}]=[-{A},{B}]$ with ${m}>0$, $\,g<0<h$, and $\gcd(g,{m})=\gcd(h,{m})=1$. In particular, ${A},{B}\in{\mathbb Q}_{>0}$, the central degree ${\overline}{R}\in M$ becomes $[0,1]\in{\mathbb Z}^2$, and $\sigma{^{\scriptscriptstyle\vee}}=\langle [-{m},h],\,[{m},-g]\rangle$.\ Let $R=k\cdot {\overline}{R}$ with $k=1,\ldots,(a_{\nu}-1)=1+{{\lfloor {A}\rfloor}}+{{\lfloor {B}\rfloor}}$ (cf. Proposition \[prop-etaci\]). Note that we have included the case $k=1$ originally belonging to (\[restrictPhi\])(ii). The zone $Z_{k{\overline}{R}}$ is the half open parallelogram in $M_{\mathbb Q}={\mathbb Q}^2$ with the vertices $$\textstyle [0,0],\hspace{1.2em} \frac{k}{h-g}\cdot [{m},-g], \hspace{1.2em} [0,k], \hspace{1.2em} \frac{k}{h-g}\cdot [-{m},h].$$ Moreover, the line ${\mathbb Q}\cdot({\overline}{R}-{m}R)={\mathbb Q}\cdot{\overline}{R}$ we are interested in by Proposition \[prop-WqG\] is given by the diagonal ${\overline}{[0,0]\,[0,k]}$. ### qG-deformations {#IIIqG} From (\[restrictPhi\])(iii) we know that each $k=1,\ldots,(a_{\nu}-1)$ gives rise to a one-dimensional ${{T^1_V}}(-k\cdot{\overline}{R})={\overline}{R}^\bot\subseteq N_k$. For each of these $k$ we have to decide whether ${{T^1_{{\operatorname{qG}}}}}(-k\cdot{\overline}{R})=0$ or ${\overline}{R}^\bot$. \[prop-finalQG\] The -deformations of $S$ consist exactly of the one-dimensional subspaces ${\overline}{R}^\bot\subseteq {{T^1}}(-k\cdot{\overline}{R})$ with $1\leq k\leq \min\{a_{\nu}-1,\,|I|\}$. We consider the embedding $\iota:{\widetilde{M}}\hookrightarrow{\mathbb Z}^2$ obtained by evaluating $({\alpha},{\beta})$. Actually, restricting to $M={\mathbb Z}^2$, this reflects the original situation of $M=({\mathbb Z}^2)^G$, and $\iota|_M$ is given by the matrix ${\left(\begin{array}{@{}*{2}{c}@{}} g & {m}\\ h & {m}\end{array}\right)}$. The rational $\iota_{\mathbb Q}$ is an isomorphism, we can detect $\iota_{\mathbb Q}(\sigma{^{\scriptscriptstyle\vee}})={\mathbb Q}^2_{\geq 0}$, and the new, truly canonical degree $\frac{1}{{m}}{\overline}{R}=[0,\frac{1}{{m}}]\in{\widetilde{M}}$ maps to $[1,1]$.\ We are going to apply Proposition \[prop-WqG\]. The description by $\iota$ implies that $({\widetilde{M}}\cap Z_{k{\overline}{R}})\setminus{\mathbb Q}{\overline}{R}$ is non-empty if and only if $Z_{k{\overline}{R}}$ contains an ${\widetilde{M}}$-lattice point on the boundary $\partial\sigma{^{\scriptscriptstyle\vee}}\setminus\{0\}$ – just subtract $\iota(\frac{1}{{m}}{\overline}{R})=[1,1]$ whenever the boundary is not reached yet.\ While $r^1=[{m},-g]$ used to be a primitive generator of one ray of $\sigma{^{\scriptscriptstyle\vee}}$ within the lattice $M={\mathbb Z}^2$, this is no longer true in ${\widetilde{M}}={\mathbb Z}\times{\mathbb Z}\frac{1}{{m}}$. Here, the element $[1,\frac{-g}{{m}}]$ does the job instead. Thus, it remains to check whether this generator belongs to $Z_{k{\overline}{R}}$. Since $\langle{\alpha},[1,\frac{-g}{{m}}]\rangle=0$ and $\langle{\beta},[1,\frac{-g}{{m}}]\rangle=h-g$, this leads to the condition $h-g<km$. The situation for the other ray generated by $r^e=[-{m},h]$ is the same. ### -deformations {#IIIVW} Recall from (\[normAct\]) that $q'\in({\mathbb Z}/n{\mathbb Z})^*$ denotes the multiplicative inverse of $q$. It is, like $q$ itself, assumed to be normalized as $1\leq q'<n-1$. The singularities $S_{n,q}$ and $S_{n,q'}$ are isomorphic, by (\[abcNotation\]) they share $a'=a$ (that is ${m}'={m}$) and $b'=b$, and at the end of (\[coneInterval\]) we have seen that $c=-1/g$ and $c'=1/h$ in $({\mathbb Z}/{m}{\mathbb Z})^*$.\ As before, we are in the grounded case, and we consider a $k\in\{1,\ldots,a_{\nu}-1\}$ with $a_{\nu}-1=1+{{\lfloor {A}\rfloor}}+{{\lfloor {B}\rfloor}}$. \[prop-finalVW\] $\,{{T^1_{{\operatorname{V\!W}}}}}(-k\cdot{\overline}{R})\neq 0$ $\;{\Leftrightarrow}\;$ $k\leq \min\{\frac{q+1}{a}, \frac{q'+1}{a}\}= \min\{c\cdot |I|, \,c'\cdot |I|\}$. By Corollary \[cor-shift\](2) or Proposition \[prop-WqG\], a degree $k{\overline}{R}$ fails to meet the -property if and only if $(M+\frac{1}{{m}}{\overline}{R})\cap Z_{k{\overline}{R}}$ or, equivalently, $M\cap (Z_{k{\overline}{R}}-\frac{1}{{m}}{\overline}{R})$ has points outside the diagonal ${\mathbb Q}\cdot{\overline}{R}$.\ First, we check that $M\cap Z_{k{\overline}{R}}\subseteq M\cap Z_{(a_{\nu}-1){\overline}{R}}$ (i.e. without the translation) is always contained in the diagonal. If not, then we could find $r^i,r^j\in E$ with, w.l.o.g., $i<{\nu}$ such that $(a_{\nu}-1)r^{\nu}-(r^i+r^j)\in\sigma{^{\scriptscriptstyle\vee}}$. This implies ${\nu}<j$, and we choose an element ${\gamma}\in\operatorname{int}\sigma$ such that $\langle{\gamma}, r^i\rangle = \langle{\gamma}, r^j\rangle \,(>0)$. Since $r^1,\ldots,r^e$ run along the boundary of the convex polygon $\operatorname{conv}(\sigma{^{\scriptscriptstyle\vee}}\cap M\setminus 0)$, it follows that $r^{{\nu}-1}, r^{\nu}, r^{{\nu}+1}\in\operatorname{conv}\{r^i,r^j\}$. Thus $$\textstyle \langle {\gamma},\, \frac{r^{{\nu}-1}+r^{{\nu}+1}}{2}\rangle \leq \langle {\gamma},\, r^i\rangle = \langle {\gamma},\, r^j\rangle = \langle {\gamma},\, \frac{r^{i}+r^{j}}{2}\rangle,$$ and we obtain a contradiction via $$\textstyle \langle {\gamma},\, r^{{\nu}-1}+r^{{\nu}+1}\rangle \leq \langle {\gamma},\, r^{i}+r^{j}\rangle \leq \langle {\gamma},\, (a_{\nu}-1)r^{\nu}\rangle < \langle {\gamma},\, a_{\nu}r^{\nu}\rangle = \langle {\gamma},\, r^{{\nu}-1}+r^{{\nu}+1}\rangle.$$ Hence, $({M}\cap Z_{k{\overline}{R},\,-1})\setminus{\mathbb Q}{\overline}{R}\,$ is non-empty if and only if $Z_{k{\overline}{R},\,-1}=Z_{k{\overline}{R}}-\frac{1}{{m}}{\overline}{R}\,$ contains an ${M}$-lattice point on the boundary $\partial\sigma{^{\scriptscriptstyle\vee}}-\frac{1}{{m}}{\overline}{R}\,$. So we have to determine the smallest $\lambda\in{\mathbb Q}_{>0}$ such that $$\textstyle \lambda\cdot[{m},-g]-[0,\frac{1}{{m}}]\in M \hspace{1.0em} \mbox{(and similarly with $[-{m},h]$).}$$ This condition is equivalent to $\lambda{m}\in{\mathbb Z}$ and $\lambda\cdot g + \frac{1}{{m}}\in{\mathbb Z}$, i.e.${m}|(\lambda m\cdot g+1)$. By Proposition \[prop-abcInv\], this means $\lambda{m}=c$. Hence, the first lattice point on the shifted ray ${\mathbb Q}_{>0}\cdot r^1-\frac{1}{{m}}{\overline}{R}\,$ is $[c,\,-\frac{gc+1}{{m}}]$. Its value under ${\beta}$ is $$\textstyle \langle{\beta},\,[c,\,-\frac{gc+1}{{m}}]\rangle = \langle (h,{m}),\,[c,\,-\frac{gc+1}{{m}}]\rangle = c(h-g)-1.$$ This leads to the condition $c(h-g)-1<k{m}-1$ for $Z_{k{\overline}{R},\,-1}$-membership. Thus, the -condition coming from the ray $r^1$ is exactly the opposite, namely $\,k\cdot{m}\leq c\cdot(h-g)$. Similarly, the first lattice point on the shifted ray ${\mathbb Q}_{>0}\cdot r^e-\frac{1}{{m}}{\overline}{R}\,$ is $[-c',\,\frac{hc'-1}{{m}}]$. It leads to the inequality $\,k\cdot{m}\leq c'\cdot(h-g)$. Comparison of - and -deformations {#compVWqG} --------------------------------- In [@KSB Definition 3.7] the so-called T-singularities are defined as those cyclic quotient singularities that admit a ${\mathbb Q}$-Gorenstein one-parameter smoothing. Their toric characterization can be found in [@minkDef (7.3)] and [@PResol (1.1)]: The toric variety ${{\mathbb T}{\mathbb V}}(\sigma)$ is a T-singularity with Milnor number $\mu$ if and only if $\sigma$ is the cone over a rational interval of integral length $\mu+1$ placed in height one.\ Since an integral length does automatically imply the uniform denominator property of (\[coneInterval\]), this description of T-singularities can directly be compared to our Proposition \[prop-finalQG\]. Looking at $k=1$, it implies that $S={{\mathbb T}{\mathbb V}}(C(I))$ allows a -deformation at all if and only if $|I|\geq 1$. Altogether, we obtain the following chain of properties of an interval $I\not\cong[0,1]$ with uniform denominators: $$(|I|=1) \hspace{{0.5em}}{\Longrightarrow}\hspace{{0.5em}} (|I|\in{\mathbb Z}_{\geq 1}) \hspace{{0.5em}}{\Longrightarrow}\hspace{{0.5em}} (|I|\geq 1)\hspace{{0.5em}}{\Longrightarrow}\hspace{{0.5em}} (\operatorname{int}(I)\cap{\mathbb Z}\neq\emptyset)$$ translating into $$(\mbox{T$_0$-singularity}){\Rightarrow}(\mbox{T-singularity}){\Rightarrow}(\exists \mbox{ {\mbox{qG}}-deformation}){\Rightarrow}(\mbox{grounded CQS}).$$ The last deformation {#lastDef} -------------------- Let $I=[\frac{g}{{m}},\frac{h}{{m}}]=[-{A},{B}]$ be a grounded interval as in (\[scanIII\]). By Proposition \[prop-etaci\] we know that $|I|\geq a_{\nu}-2$ (with equality exactly for the T-singularities). Hence, Proposition \[prop-finalQG\] implies that all subspaces ${\overline}{R}^\bot\subseteq {{T^1}}(-k{\overline}{R})$ with $k=1,\ldots,a_{\nu}-2$ are -deformations (hence -deformations, too).\ We will call the remaining deformation in degree $-(a_{\nu}-1)\cdot {\overline}{R}$ the “[*last deformation*]{}”. This is the only degree where - and -deformations might differ at all. Note that the last deformation might also be the first one, i.e.$k=1$. This happens if and only if $a_{\nu}=2$, i.e. if and only if $\,0<{A},{B}<1$.\ For the following theorem, we will denote by $\{C\}:=C-{{\lfloor C \rfloor}}$ the fractional part of a (positive, rational) number $C$. Recall that ${A},{B}\in{\mathbb Q}_{>0}$. \[th-lastDef\] The one-dimensional subspaces ${\overline}{R}^\bot\subseteq {{T^1}}(-k\cdot{\overline}{R})$ for a grounded $S={{\mathbb T}{\mathbb V}}(C(I))$ with $k=1,\ldots,a_{\nu}-2={{\lfloor {A}\rfloor}}+{{\lfloor {B}\rfloor}}$ are - and -deformations. Moreover, the “last” deformation from ${\overline}{R}^\bot$ in degree $-k\cdot{\overline}{R}$ with $k=a_{\nu}-1$ is a - or -deformation in the following cases:\ [1)]{} The last deformation of $S={{\mathbb T}{\mathbb V}}(C(I))$ is  if and only if $\,\{{A}\}+\{{B}\}\geq 1$.\ [2)]{} If $\{{A}\}, \{{B}\}\neq \frac{1}{{m}}$, then the last deformation is .\ [3)]{} Otherwise, i.e. if $\{{A}\}=\frac{1}{{m}}$ or $\{{B}\}=\frac{1}{{m}}$, then the last deformation is  if and only if it is . Hence, every -deformation is  in this case. \(1) By Proposition \[prop-etaci\] and \[prop-finalQG\], both sides are equivalent to $|I|\geq a_{\nu}-1$.\ (3) The condition $\{{A}\}=\frac{1}{{m}}$ means $g\equiv -1$ (mod ${m}$), and since $c\cdot (-g)=1$ in $({\mathbb Z}/{m}{\mathbb Z})^*$, this translates into $c=1$. Similarly, $\{{B}\}=\frac{1}{{m}}$ is equivalent to $c'=1$. Thus, the bounds in Proposition \[prop-finalQG\] and \[prop-finalVW\] coincide.\ (2) We distinguish two cases. First, if $a_{\nu}\geq 3$, then $|I|\geq 1$. Hence $$a_{\nu}-1 \leq |I|+1 \leq \min\{c\cdot |I|, \;c'\cdot |I|\} \;\mbox{ since } c,c'\geq 2.$$ Otherwise, if $a_{\nu}=2$, then $c\geq 2$ together with $c\cdot (-g)\equiv 1\, ({m})$ implies that $$c\cdot (h-g)\geq c\cdot (-g)\geq {m}+1 > {m},$$ hence $\,c\cdot |I|>1=a_{\nu}-1$. Similarly we use $c'\cdot h \equiv 1\, ({m})$ to obtain $\,c'\cdot |I|>a_{\nu}-1$. Proof of Theorem \[V-VW.thm.1\] {#proofIntro} ------------------------------- We are going to proof Theorem \[V-VW.thm.1\] of the introduction.\ (1) Since $b=\gcd(n,\, q+1)$, the assumption implies $b=1$, hence $a={m}=n$. Thus, $|I|=\frac{1}{{m}}$, and this does not leave space for $I=[\frac{g}{{m}},\frac{h}{{m}}]$ to become grounded, i.e. to allow an integer as an interior point of $I$.\ (2) Singularities admitting a -smoothing are called T-singularities. In (\[compVWqG\]) we have seen that they correspond exactly to the intervals of integral length, i.e. $\{{A}\}+\{{B}\}=1$. Now, the claim follows directly from Theorem \[th-lastDef\](1).\ (3) This follows because the - and -deformations can at most differ by the “last deformation”. This was just addressed in (\[lastDef\]). Alternatively, it follows directly from Theorem \[th-IAB\]. An example of a -deformation which is not {#exVWnotQG} ------------------------------------------ Let $\,I=[-\frac{2}{5},\frac{2}{5}]$, i.e. ${A}={B}=\frac{2}{5}$. This implies that $\{{A}\}=\{{B}\}=\frac{2}{5}$, i.e. by Theorem \[th-lastDef\](1), the last deformation is not . Another way to see this is the criterion from (\[compVWqG\]): Since $|I|<1$, there is no -deformation at all.\ On the other hand, both $\{{A}\}$ and $\{{B}\}$ are different from $\frac{1}{5}$. Thus, Theorem \[th-lastDef\](2) implies that the last deformation is . Moreover, since there is no -deformation at all, this has to be the “first” deformation ${\overline}{R}^\bot\subsetneq {{T^1}}(-{\overline}{R})$ (i.e. with $k=1$) as well. The other invariants are $$\renewcommand{{0.5em}}{0.7em} n=20,\hspace{{0.5em}} q=11,\hspace{{0.5em}} {m}=a=5,\hspace{{0.5em}} b=4,\hspace{{0.5em}} \mbox{and }\; c=c'=3.$$ The continued fraction $\frac{n}{n-q}=\frac{20}{9}$ yields $[a_2,\ldots,a_6]=[3,2,2,2,3]$, i.e. $e=7$ and ${\overline}{R}=r^4$. The associated $a_4=2$ was already known from our observation that the “first” equals the “last” deformation. Finally, we obtain the following dimensions: $$\dim {{T^1}}=10 \mbox{ with } \dim {{T^1}}(-k\cdot r^i)=\left\{\begin{array}{ll} 1 & \mbox{ if $k=1,2$ and $i=2,6$}\\ 2 & \mbox{ if $k=1$ and $i=3,4,5$}, \end{array}\right.$$ $\dim {{T^1_V}}=3$ (degrees $-r^3,-r^4,-r^5$), and $\dim {{T^1_{{\operatorname{V\!W}}}}}=1$ (in degree $-r^4$). Unobstructed -families {#unobQG} ---------------------- While the focus of the paper is on the infinitesimal level, we would just like to add how the first order -deformations of $S={{\mathbb T}{\mathbb V}}(\sigma)$ extend to an unobstructed global family. Assume that $I=[-{A},{B}]$ is an interval with uniform denominators giving rise to a cyclic quotient singularity $S_I$. From Theorem \[th-IAB\] we know that $d:=\dim {{T^1_{{\operatorname{qG}}}}}(S_I)$ equals ${{\lfloor {A}+{B}\rfloor}}$. This number vanishes unless $I$ is grounded. In particular, we may write $$I = I' + d\cdot [0,1]$$ for some interval $I'$ (with uniform denominators) of length $|I'|<1$. In [@flip (3.2)], such decompositions gave rise to so-called homogeneous toric deformations of $S_I$ over the parameter space ${\mathbb A}^d_k$. Its total space arises from the cone ${\widetilde{\sigma}}$ taken over the [Cayley]{}-construction, i.e. from $$\textstyle {\widetilde{\sigma}}:={\mathbb Q}_{\geq 0} \cdot \big(I',\,e^0\big) + \sum_{j=1}^d {\mathbb Q}_{\geq 0} \cdot \big([0,1],\,e^j\big)\subseteq {\mathbb Q}\times{\mathbb Q}^{d+1}$$ where $\{e^j{\,|\;}j=0,\ldots,d\}$ denotes the canonical basis of ${\mathbb Q}^{d+1}$. As it is $S_I={{\mathbb T}{\mathbb V}}(\sigma)$, also ${{\mathbb T}{\mathbb V}}({\widetilde{\sigma}})$ is ${\mathbb Q}$-Gorenstein. The (non-toric) flat map ${{\mathbb T}{\mathbb V}}({\widetilde{\sigma}})\to{\mathbb A}^d_k$ arises from the toric map ${{\mathbb T}{\mathbb V}}({\widetilde{\sigma}})\to{\mathbb A}_k^{d+1}$ assigned to the projection ${\mathbb Z}\times{\mathbb Z}^{d+1}{\rightarrow\hspace{-0.8em}\rightarrow}{\mathbb Z}^{d+1}$ composed with the linear projection ${\mathbb A}_k^{d+1}{\rightarrow\hspace{-0.8em}\rightarrow}{\mathbb A}_k^{d+1}/k\cdot (1,1,\ldots,1)\cong{\mathbb A}_k^d$.\ [**Acknowledgment:**]{} We would like to thank F.-O. Schreyer for many fruitful discussions and initiating the contact on this topic. Thanks to Jan Stevens for finding mistakes in the originally submitted arXiv version and to the anonymous referee for valuable suggestions. [^1]: Partial financial support to KA was provided by the DFG via the CRC 647 and to JK by the NSF under grant number DMS-1362960.

ArXiv

--- abstract: | A CCD $V$,$V-I$ colour-magnitude diagram (CMD) of the Galactic globular cluster NGC 2808 has been obtained with the ESO-NTT, reaching down to $V \sim 24$. The highly populated Main Sequence (MS) presents a significant broadening redward to the MS ridge line, larger than expected from photometric errors alone, which could be interpreted as due to binary candidates. author: - 'F.R. Ferraro , E. Carretta , F. Fusi Pecci , A. Zamboni' date: '$^{\dag}$Based on observations obtained at the European Southern Observatory, La Silla, Chile' title: 'Binary stars in globular clusters: detection of a binary sequence in NGC 2808?$^{\dag}$' --- Introduction ============ There are various (direct and indirect, but always difficult) approaches followed so far to detect binaries in Galactic Globular Custers (GGCs) (see Hut et al. 1992 and Bailyn 1995 for a discussion). Most of them have been unsuccesful, but this may well be due to the fact that the surviving binaries are likely segregated in the inner cluster regions, hardly observable from the ground because of crowding. One possible path to explore consists in the study of the intrinsic width of the Main Sequence (MS), looking for some sort of “parallel" sequence due to the photometric combination of the binary components (see Romani and Weimberg 1991). This approach requires the achievement of a very accurate photometry (to $\sim 0.01 - 0.02$ mag) for a very wide sample of faint MS stars. On the other hand, other methods imply time-consuming and distance-constrained spectroscopic observations, usually sensitive only to the brighter objects with present day instrumentation. The true fraction and the physical nature of binary stars in a GGC is a clue element that brings information on a variety of topics like the environment in which the cluster was born, the details of the star formation, and the subsequent internal dynamical evolution. While theoretical models are presently only poorly constrained by any observational (and quantitative) data-set, it is now well established that a strict connection does exist between stellar and dynamical evolution of a GC star, as revealed for instance by the morphology of the evolved branches, in particular the Horizonthal Branch (HB; see e.g. Fusi Pecci et al. 1993 for discussion and references). NGC 2808 ($\alpha_{1950} = 9^h 10.9^m, \delta_{1950} = -64^o 39'$) is one of the most interesting objects to study, in this respect, since its very high concentration is likely to enhance any effect due to the dynamical evolution of binary or multiple star systems. This cluster is also the typical template of a growing class of GGCs with a net, bimodal HB morphology (see Harris 1975, Ferraro et al. 1990-F90, and Rood et al. 1993). This, in turn, could well be heavily affected by the evolution of binary systems that at later stages can populate both the red and the blue HB extremes (Fusi Pecci et al. 1993). Further, to explain its bimodal HB, van den Bergh (1997) has recently suggested that this cluster might have formed by mergers (but see Catelan 1997). Any new (though preliminar) information can thus be useful to have a deeper insight on its properties. In this research note we present a very preliminary evidence of a secondary sequence running parallel to the MS of this cluster that, if confirmed, could be interpreted as composed of unresolved, candidate binary systems. Observations and data reduction =============================== A set of deep $V$ and $I$ frames (3 exposures of 600 secs in each filter) were acquired under good seeing conditions (0.7-0.9" FWHM) on January 1995 at the NTT telescope (ESO, La Silla, Chile) with the EMMI camera equipped with a 2048 $\times$ 2048 CCD detector. The image scale is 0.25 arcsec/pixel, yielding a total field of view of 8.5’ $\times$ 8.5’. In order to avoid severe crowding conditions, the results reported here are based only on a sub-field of about 4’ $\times$ 4’, whose center lies about 6’ southward and 2’ eastward away from the cluster center (see Figure 1). All reductions were performed using the package ROMAFOT (Buonanno et al. 1983). Details of the image analysis are very similar to those described elsewhere (see F90, Ferraro et al. 1992) and will not repeated here. The data-files are available upon request. The instrumental magnitudes (from profile-fitting) were first tied to fixed aperture photometry and then referred to the photometric Johnson system using 15 stars in the Landolt (1992) standard fields Rubin 149 and SA98. The resulting calibration equations are: $$V = v + 0.024 (v-i) + 23.793$$ $$I = i -0.077 (v-i) + 23.260$$ where $V$, $I$ are standard magnitudes and $v$, $i$ indicate the instrumental ones. Atmospheric extinction was taken into account by means of average extinction curves appropriate for La Silla. Based on comparison of 70 stars in common, we note here that our new calibration is fainter (0.12 mag) in $V$ than previously obtained by F90; no comparison was feasible in the $I$ band. Results ======= Figure 2 shows the ($V$, $V-I$) CMD for 3174 stars detected in the selected sub-frame. Both the turn-off (TO) region and MS are well defined. However, [*the most intriguing evidence is the presence of a clear broadening above and to the red of the MS, which can be interpreted as a secondary sequence, running parallel to the MS, but at brighter magnitudes and redder colour.*]{} In Figure 2, to make more evident this feature, we marked the claimed parallel sequence with a dashed line. The distance in magnitude of this “secondary" sequence from the MS is somewhat less than 0.75 mag. Note that equal component binaries would lie at the same colour and 0.75 mag brighter than the two individual members (Romani and Weinberg 1991), and this magnitude difference should thus be considered as an upper limit. Consequently, taking also into account the photometric errors, the average difference here measured for the two parallel sequences is substantially compatible with the above limits. To test the existence of the candidate binary sequence we have considered various luminosity bins along the MS (between 21 and 22.5, see Table 1). Then for all the stars in each bin we derived the distribution of the distances ($\delta x$) from the MS ridge line (MSRL). The observable ($\delta x$) is defined as the geometrical distance in the ($V$, $V-I$)-plane of each star from the adopted MSRL, with $\delta x > 0$ or $\delta x < 0$ if the star is redder or bluer than the MSRL, respectively. In Figure 3, the distributions of the ($\delta x$)-values in the ranges $V = 21.5 \div 22.0$ and $V = 21.0 \div 22.0$ are shown for illustration and clarity. [lcc]{}\ Section & Mag. range & K-S\ \ \ A & 21.0$-$22.0& 0.7%\ B & 21.5$-$22.0& 0.5%\ C & 21.5$-$22.5& 3.0%\ \ As can easily be seen, both panels in Figure 3 show that the distributions of the ($\delta x$)-values are clearly asymmetric. In fact, besides the main peak centered on the MS, there is a quite evident secondary peak, displaced by about $\delta$x = 0.12 from the main one. In our view, this is a direct evidence for the presence of a secondary sequence along the MS as the asymmetric shape itself of these distributions is a proof that the broadening of the MS is not just due to photometric errors. Before discussing at any level the possible origin of the stars located in the parallel sequence, we have tested the statistical significance of this claimed feature. To this aim we have compared the number of stars distributed over equal intervals in distance ($|\delta x|$) from the MSRL lying in the blue and red side of the distributions. Over the magnitude interval $21.5 < V < 22.0$ (where the effect is best evident), the number of stars within $|\delta x| < 0.1$ from the MSRL is almost the same (within 1 $\sigma$) on both side of the distribution (138 stars on the blue side and 159 on the red one, respectively). On the contrary, if one consider the range $0.1 < |\delta x| < 0.2$, one finds only 15 stars bluer than the MSRL to be compared with 48 objects located on the red side. Hence, the number of red stars is more than 3 times that found in the blue side of the distribution. Moreover, a Kolmogorov-Smirnov (KS) test applied to evaluate the significance of the detected difference in the red and blue distributions yields a very low probability that the two distributions have been extracted from the same $parent$ distribution. In fact, as can be seen in the last column of Table 1 for the 3 intervals of magnitudes between $V = 21.0$ and $V = 22.5$, the probability that the presence of the secondary peak is not due to a statistical fluctuation ranges from $\sim 97\%$ (2.2 $\sigma$) to $\sim 99.5\%$ (2.8 $\sigma$). The ($V$,$V-I$)-CMD shown in Figure 2 and the statistical tests carried out above indicate thus that several stars in the MS of NGC 2808 lie on a secondary sequence running parallel and redward to the MS. From the areas of the interpolating Gaussian fits (see Figure 3) we can also obtain a (rough) estimate of the percentage of stars populating the two sequences over the total number of MS stars. In the different magnitude bins considered above we obtain for the parallel sequences figures ranging from 20.0 to 28.0 $\%$, with a mean values of $\sim (24 \pm 4) \%$. Optical and/or physical binaries? ================================= Concerning the possible nature of the stars populating the parallel sequence, the label [*unresolved binary*]{} has to be taken with particular caution as it may mean either unresolved $optical$ binaries or unresolved $physical$ binaries. In fact, since we are making photometric measurements in a very crowded field, it has to be seen as quite normal the detection of some blended images. On the other hand, the fraction of candidate binaries we have found ($\sim 24\%$) is somewhat too high compared to the number of unresolved blends one can predict to find (just because of crowding) “making artificial star” experiments (which would yield here values close to $\sim 10\%$). As a consequence, one could also conclude as a working hypothesis that a non neglegible fraction of the stars located on the parallel MS could represent reliable intrinsic binary candidates. Obviously, since stellar crowding mimics the effect of the physical dynamical relaxation acting in the cluster and concentrates the optical blends toward the cluster center, as does mass segregation with true (and more massive) binary systems, it is too early to extend deeply the discussion of the possible implications of this observational result, if confirmed. However, it may be interesting to recall that NGC 2808 presents one of the most intriguing HB-morphologies found in a Galactic globular cluster so far. As reported in the Introduction, since the early observations carried out by Harris (1974, 1975), it was evident that the HB of NGC 2808 is highly bimodal. Moreover, the latest HST data presented by Djorgovski et al. (1996) have shown that, besides the already detected bimodality, the blue HB of NGC 2808 is actually made by 3 sub-portions, separated by quite evident gaps. Within this framework still so uncertain and worth of further exploration, the detection of a fraction of $physical$ binaries could open new perspectives to the possible interpretation of the detected peculiarities in the HB morphology. For instance, one could imagine that at least a group of the HB stars could represent the descendants of the binary systems. The discussion of how this could occur is however beyond the purposes of the present work, and we refer to the review papers by Hut et al. (1992) and Bailyn (1995) for further analysis of this aspect. This work has been supported by the [*Agenzia Spaziale Italiana*]{} (ASI) and by the [*Ministero della Università e della Ricerca Scientifica e Tecnologica*]{} (MURST). Bailyn, C.D. 1995, ARA&A, 33, 133 Buonanno, R., Corsi, C.E., Fusi Pecci, F. 1985, A&A, 145, 97 Buonanno, R., Buscema, G., Corsi, C.E., Ferraro, I., Iannicola, G. 1983, A&A, 126, 278 Crocker, D.A., Rood, R.T., O’Connell, R.W. 1988, ApJ, 332, 236 Djorgovski, S.G. et al. 1996, in [*Stellar Ecology Workshop*]{}, ed. R.T. Rood (Cambridge: CUP) Ferraro, F.R., Clementini, G., Fusi Pecci, F., Buonanno, R. 1992, MNRAS, 252, 357 Ferraro, F.R., Clementini, G., Fusi Pecci, F., Buonanno, R., Alcaino, G. 1990, A&ASS, 84, 59 (F90) Fusi Pecci, F., Ferraro, F.R., Bellazzini, M., Djorgovski, S., Piotto, G., Buonanno, R. 1993, AJ, 105, 1145 Harris, W.E. 1975, ApJS, 29, 397 Harris, W.E. 1974, ApJ, 192, L14 Hut, P., McMillan, S., Goodman, J., Mateo, M., Phinney, E.S., Pryor, C., Richer, H.B., Verbunt, F., Weinberg, M. 1992, PASP, 104, 981 Landolt, A.U. 1992, AJ, 104, 340 Leonard, P.J.T., Fahlman, G.G. 1991, AJ, 102, 994 Romani, R.W, Weinberg, M.D. 1991, ApJ, 372, 487 Rood, R.T., Crocker, D.A., Fusi Pecci, F., Ferraro, F.R., Clementini, G., Buonanno, R. 1993, in [*The Globular Cluster-Galaxy Connection*]{}, A.S.P. Conf. Ser., 48, 218, eds. G.H. Smith and J.P. Brodie

ArXiv

--- author: - 'Frédéric Campana, Jungkai A. Chen and Thomas Peternell' date: 'october 25, 2005' title: Strictly nef divisors --- .2cm Introduction {#introduction .unnumbered} ============ Given a line bundle $L$ on a projective manifold $X$, the Nakai-Moishezon criterion says that $L$ is ample if and only if $$L^s \cdot Y > 0$$ for all $s$ and all irreducible subvarieties $Y \subset X$ of dimension $s.$ Examples show that it is not sufficient to assume that $L \cdot C > 0$ for all curves; line bundles with this property are called [*strictly nef* ]{}. If however $L = K_X$ is strictly nef, then standard conjectures predict that $K_X$ is already ample; this is proved by “Abundance” in dimension up to 3 (Kawamata, Miyaoka). If $L = -K_X$ is strictly nef in dimension 3, then Serrano \[Se95\] showed that $-K_X$ is ample, i.e. $X$ is a Fano threefold. This lead him to set up the following Let $X_n$ be a projective manifold and $L$ a strictly nef line bundle on $X$. Then $K_X + tL $ is ample for any real $t > n+1.$ Serrano established the conjecture in dimension 2, and also in dimension 3, with the following possible exceptions: - $X$ is Calabi-Yau and $L \cdot c_2(X) = 0;$ - $X$ is uniruled with irregularity $q(X) \leq 1$, in particular $X$ is rationally connected; - $X$ is uniruled with irregularity $q(X) = 2$ and $\chi({{\mathcal O}}_X) = 0$. As said, he also settled the case $L = -K_X$ in dimension 3. In this paper we rule out the two last cases and establish also results in higher dimensions: Let $X_n$ be a projective manifold and $L$ a strictly nef line bundle on $X$. Then $K_X + tL $ is ample if $t > n+1$ in the following cases. 1. $\dim X = 3$ unless (possibly) $X$ is Calabi-Yau with $L \cdot c_2 = 0;$ 2. $\kappa (X) \geq n-2;$ 3. $\dim \alpha(X) \geq n-2$, with $\alpha: X \to A$ the Albanese map. Statement 2) (resp. 3), resp. 1)) will be proved in §2 (resp. §3, resp. §4-5). The remaining three-dimensional case that $X$ is Calabi-Yau with $L \cdot c_2 = 0$ is a very hard problem in Calabi-Yau theory and definitely requires very different methods. Basic definitions, known results and main problems ================================================== For technical reasons we have to consider not only strictly nef line bundles, but also a slight generalization of this notion. Let $X$ be a normal projective variety. 1. A line bundle $L$ over $X$ is strictly nef, if $L \cdot C > 0$ for all irreducible curves $C \subset X.$ 2. $L$ is almost strictly nef, if there is a normal projective variety $X'$, a surjective birational holomorphic map $f: X \to X'$ and a strictly nef line bundle $L'$ on $X'$ such that $L = f^*(L').$ The main problem on strictly nef line bundles is Serrano’s: Let $X_n$ be a projective manifold and let $L$ be a strictly nef line bundle on $X$. Then $K_X + tL$ is ample for $t > n+1$. [**Remark:**]{} *More generally, one might conjecture that if $X_n$ is a normal projective variety with canonical singularities and index $i(X)$, and if $L$ is a strictly nef line bundle on $X$, then $K_X+ tL$ is ample for all $t > i(X)(n+1).$* By definition, the index $i(X)$ is the smallest number $i$ such that $iK_X$ is Cartier. One could add (in the smooth case) that $K_X + nL$ is always nef, and not ample if and only if $X = {{\mathbb P}}_n, L = {{\mathcal O}}_X(1).$ It is known since a long time that strictly nef divisors need not be ample; even if moreover big. See Ramanujam’s example in \[Ha70\]. There are however three important special cases of the conjecture, namely when $L = K_X$ (resp. $L = -K_X)$, resp. $K_X \equiv 0$. In the first case the abundance conjecture predicts that $mK_X$ is spanned for a suitable large $m$ so that $K_X$ will be ample as soon as $K_X$ is strictly nef. This is known in dimension up to $3.$ In the second case $X$ should be Fano if $-K_X$ is strictly nef. In the last case, $L$ should be ample. Perhaps the best justification for the above conjecture (1.2) is that it holds for $L$ if and only if $$L^{\perp} \cap K_X^{\perp} \cap {\overline {NE}}(X) = \{0\},$$ in $N_1(X)$, see Proposition 1.4 below. So the conjecture should be viewed as a statement on the cone ${\overline {NE}}(X),$ at the points where the intersection number with $K$ and $L$ simultaneously vanish. Observe thus that the crucial cases are precisely the three “special" cases above, where $L= K_X$, $L=-K_X$, and $ K_X \equiv 0$. Notice also that, if ${\overline {NE}}(X)$ is generated by the classes of irreducible curves (i.e. without taking limits), then the conjecture is true since $K_X+t L$ is again strictly nef, for $t > (n+1)$ (1.6). This holds in particular if $X$ is Fano. By $ME(X)$ we will always denote the cone of movable curves. Its closure is the cone dual to the cone of effective divisors; see \[BDPP04\] for details. Let $L$ be strictly nef and $\alpha \in {\overline {NE}}(X) $ such that $(K_X + tL) \cdot \alpha = 0$ ($ t > n+1)$. Then 1. $K_X \cdot \alpha = L \cdot \alpha = 0.$ 2. $\alpha \in \partial ME(X)$ for a suitable choice of $\alpha.$ \(1) Suppose $L \cdot \alpha \ne 0.$ Then $L \cdot \alpha > 0$ and $K_X \cdot \alpha < 0.$ By the cone theorem we can write $$\alpha = \sum_{i=1}^N a_i C_i + R$$ with $C_i$ extremal and $K_X \cdot R \geq 0.$ Since $-K_X \cdot C_i \leq n+1$, and $tL.C_i\geq t>n+1$, for all $i$, we have $(K_X + tL) \cdot C_i > 0,$ which gives a contradiction.\ (2) If there is no nonzero $ \alpha \in \partial ME(X)$ with $(K_X + tL) \cdot \alpha = 0,$ then by \[BDPP04\] $K_X + tL$ is big. But then $K_X + tL$ is ample, by (1.6(2)) below. The following cases have been settled by Serrano \[Se95\] ([*Serrano*]{}) 1. Let $X$ be a irreducible reduced projective Gorenstein surface and $L$ strictly nef on $X$. Then $K_X+tL$ is ample for any real $t > 3.$ 2. Let $X$ be a smooth projective threefold and $L$ strictly nef. Then $K_X+tL$ is ample for $t > 4$ with the following possible exceptions only: $X$ is Calabi-Yau and $L \cdot c_2 = 0;$ or $X$ is uniruled with $q \leq 1$; or $X$ is uniruled, $q = 2$ and $\chi({{\mathcal O}}_X) = 0.$ Moreover $X$ is Fano as soon as $-K_X$ is strictly nef. The following more technical results are also due to Serrano. Let $X$ be an n-dimensional connected projective manifold and $L$ a strictly nef line bundle on $X$. 1. For every real number $t > n+1,$ $K_X+tL$ is a strictly nef ${{\mathbb R}}-$divisor. This also holds for $t >> 0$ and $X$ a normal projective variety with only canonical singularities. 2. If $K_X+tL$ is not ample for some real number $t > n+1,$ then $K_X^j \cdot L^{n-j} = 0$ for all $j \geq 0.$ So if $(K_X+tL)^n \ne 0$ for some real number $t > n+1$ (i.e. if $ K_X+tL$ is big and strictly nef), then $K_X+tL$ is ample. 3. If $\dim X = 3$ and $\vert pK_X+qL\vert$ contains an effective non-zero divisor for some integers $p,q$, then $K_X+t L$ is ample for $t > 4.$ The last proposition says in particular that to prove Conjecture 1.2 for each $t > n+1,$ it is sufficient to prove it for some positive integer $t> n+1.$ Results in case of positive Kodaira dimension ============================================= If $X$ is of general type, then Conjecture (1.2) easily holds: Let $X$ be a projective n-dimensional manifold with $\kappa (X) = n.$ Let $L$ be strictly nef on $X$. Then $K_X+(n+1)L$ is ample. Let $t > n+1$ be a rational number. By (1.6), $K_X+tL$ is strictly nef. Then $2(K_X+tL) - K_X$ is big and nef, hence by the base point free theorem, $K_X+tL$ is semi-ample and strictly nef, hence ample. If $X$ is not of general type, things are more complicated; here we want to use the Iitaka fibration. For technical reasons we introduce ${\bf (C_d):}$ Let $F_d$ be a projective manifold with $\kappa (F) = 0. $ Let $L$ be almost strictly nef. Then, $K_F+tL$ is big for $t > d+1.$ Let $X$ be an n-dimensional connected projective manifold with $\kappa (X) = k \geq 0.$ Let $L$ be a strictly nef divisor on $X$. Suppose that $C_d$ holds for $d = n-k.$ Then, $K_X+tL$ is ample for $t > n+1.$ Let $f: X \rightharpoonup Y$ be the Iitaka fibration; we may assume $\dim Y = k \geq 1$, because otherwise there is nothing to prove. Let $\pi: \hat X \to X$ be a sequence of blow-ups such that the induced map $\hat f: \hat X \to Y$ is holomorphic, and moreover we can write: $$\pi^*(mK_X) = \hat f^*(A) + E \eqno (*)$$ with an ample divisor $A$ on $Y$ and an effective divisor $E.$ We also have an equality: $K_{\hat X} = \pi^*(K_X) + E'$, for some effective $E'$. Let us set: $\hat L = \pi^*(L).$\ By $(C_d)$ applied to the general fiber $F$ of $\hat f$, the divisor $K_F + t \hat L$ is big, for $t > d+1,$ even.\ Thus $\pi^*(K_X) + E' + t \hat L$ is $\hat f-$big. The following Lemma (2.4) therefore applies, with $N = \pi^*(K_X) + t \hat L$ and with $D = E'$ and shows the bigness of $$B:=\pi^*(K_X) + E' + t \hat L + \hat f^*(A)=(f^*(A)+E)+(\pi^*(K_X)+E'+t\hat{L})$$ Thus by (\*), $\pi^*((m+1)(K_X)) + E' + t \hat L = B+E $ is big, and so does $$\pi^*((m+1)K_X) + E' + (m+1)t \hat L=(B+E)+mt \hat L,$$ being the sum of two divisors, $B+E$, which is big, and $mt \hat L$, which is nef.\ Therefore $K_{ \hat X} + t \hat L$ is also big and thus ample, by (1.6). Let $g: X \to Y$ be a holomorphic map of projective varieties. Let $A,N,D$ be ${{\mathbb Q}}$-divisors, with $A$ ample on $Y$, $N$ nef on $X$ and $D$ effective on $X$. Suppose that $D+N$ is $g-$big, i.e. big on the general fiber. Then $D+N+g^*(A)$ is big. Choose and fix $k$ large such that $D+N+g^*(kA)$ is big.\ (This is a standard fact, seen as follows, by a relative version of Kodaira’s Lemma: let $H$ be $g-$ample on $X$. Then choose $m$ such that $g_*(m(D+N)-H)$ has positive rank. This is obviously possible, by the coherence of direct image sheaves, since $(D+N)$ is $g-$big. See \[KMM87,0-3-4\], for example. Now choose $k$ large enough, such that $g_*(m(D+N)-H) + kA$ has a section. Thus $E:=m(D+N)-H + g^*(kA)$ is effective, and $m(D+N) + g^*(kA) =H+E$ is of the form: ample plus effective, and thus big, as claimed). .2cm Thus $D+N+g^*(kA) = aH+E$ with $H$ ample, $a$ a positive rational number and $E$ an effective ${{\mathbb Q}}-$divisor. Since $N$ is nef, $N+\epsilon H$ is ample for all positive numbers $\epsilon;$ choose $\epsilon$ such that $(k-1)\epsilon < a.$ Next observe, introducing the effective divisor $E' = (k-1)(D+N+\epsilon H)$ that $$(k-1)(D+N+\epsilon H) + D + N + g^*(kA) = E' + aH + E =: aH + E''$$ with $E''$ effective. On the other hand, $$(k-1) (D+N+\epsilon H) + D + N + g^*(kA) = k(D+N+g^*(A)) + (k-1)\epsilon H,$$ hence, substracting $(k-1).\epsilon.H$ from both sides, we get the equality: $$k(D+N+g^*(A))= aH + E'' - (k-1)\epsilon H = (a-(k-1)\epsilon)H + E''.$$ Since $(a-(k-1)\epsilon) > 0$, by the choice of $\epsilon$, the right hand hand side divisor is big, hence $D+N+g^*(A)$ is also big. Conjecture $C_1$ being obvious, we are now going to prove $C_2.$ Let $X$ be a smooth projective surface with $\kappa (F) = 0$, $L$ almost strictly nef. Then $K_X+tL$ is big for $t > 3.$ Fix a rational number $t > 3$ and suppose that $K_X + tL$ is not big. By blowing down the $(-1)-$curves $E_i$ with $L \cdot E_i = 0,$ we may assume that $K_X + tL$ is nef. So if $(K_X + tL)$ is not big, we must have $(K_X + tL)^2 = 0.$ If $L^2 > 0,$ then of course our claim is clear, so suppose $L^2 = 0.$ Hence $$( K_X^2 + 2t K_X) \cdot L = 0.$$ This holds also for all rational numbers $3 < t_0 < t,$ because otherwise $K_X + t_0L$ would be big and then also $K_X + tL$ is big. Thus: $K_X^2=0=K_X\cdot L$. The surface $X$ is thus minimal. By taking a finite étale cover, we can assume $X$ to be either an abelian, or a $K3$-surface. But the argument used in \[Se95\] for abelian varieties shows that an almost strictly nef divisor on an abelian variety is ample. On the other hand, Riemann-Roch shows that a nef line bundle on a $K3$-surface is either effective or trivial. An effective almost strictly nef line bundle on a surface is immediately seen to be big, and thus ample. Claim $(C_2)$ trivially holds also on surfaces of general type and is very easily checked in case $\kappa = 1.$ It should also hold in case $\kappa = - \infty$ but we don’t need it. Combining (2.3) and (2.5) we obtain: Let $X$ be an n-dimensional connected projective manifold with $\kappa (X) \geq n-2.$ Let $L$ be a strictly nef line bundle on $X$. Then $K_X+t L$ is ample for $t > n+1.$ The Albanese map ================ We now study Conjecture 1.2 on projective manifolds $X$ with $q(X) > 0.$ Since our most complete result is in dimension 3, we will do this case first and then examine what can be done in higher dimensions. Let $X$ be a smooth projective threefold, $L$ strictly nef. Suppose there exists a non-constant map $g: X \to A$ to a abelian variety. Then $K_X + tL$ is ample for $t > 4.$ Let $D_t:= 2K_X+2tL$. We claim that $\mathcal{F} = \mathcal{F}_t:=g_*(2K_X+2tL)$ satisfies a [*generic vanishing theorem*]{} (cf. [@Ha04] Theorem 1.2) for $t$ a sufficiently large integer. That is, we have a chain of inclusions $$V^0( \mathcal{F}) \supset V^1(\mathcal{F})\ldots \supset V^n(\mathcal{F}),$$ where $$V^i(\mathcal{F}):=\{ P \in {\rm Pic}^0(X)| h^i(A, \mathcal{F} \otimes P) \ne 0\}.$$ Grant the claim for the time being. Since $\mathcal{F}$ is a non-zero sheaf for $t \gg 0$, one concludes that $V^0(\mathcal{F}) \ne \emptyset$. For otherwise, $V^i(\mathcal{F}) =\emptyset$ for all $i$, which implies that the Fourier-Mukai transform of $\mathcal{F}$ is zero. This is absurd e.g. by \[Mu81,2.2\]. Therefore $h^0(X, 2K_X+2tL+P) \ne 0$ for some $P \in {{\rm Pic}}^0(X)$. Now $L'$ be a divisor such that $2tL'=2tL+P$, then by Proposition 1.6.(3), $K_X+tL'$ is ample for $t > 4$ and hence so is $K_X+tL$ (notice that if $2K_X + 2tL + P$ has a section without zeroes, then $-K_X $ is strictly nef, hence $X$ is Fano and $q(X) = 0,$ so that 1.6(3) really applies). To see the claim, first note that $K_X+t_0L$ is $g$-big for some $t_0 >0$ (1.5(1)). Fix any ample line bundle $H$ on $A$. By Lemma 2.4, it is easy to see that $a(K_X+2tL)+g^*H$ is nef and big for $a>0$ and $t >t_0$. Set $D_0:= 2(K_X+2tL)+g^*H$. then $D_0-K_X$ is again nef and big. By the Base Point Free Theorem, $mD_0$ is spanned for some $m \gg 0$. Take $D$ a general smooth member in $|mD_0|$. Then we have $$2K_X+2tL+g^*H \equiv K_X+\frac{2m-1}{2m}g^*H+\frac{1}{2m}D,$$ where $(X,\frac{1}{2m}D)$ is klt. By the vanishing theorem of Kollár, we have $$H^j(A, g_*(2K_X+2tL) \otimes H))=0, \text{ for all } j >0$$ and moreover $$H^j(A,\mathcal {F} \otimes H \otimes P) = H^j(A, g_*(2K_X+2tL) \otimes H \otimes P))=0 \text{ for all } j >0, \ P \in \rm{Pic}^0(X).$$ In other words, [*per definitionem*]{} the sheaf $\mathcal{F} \otimes H$ is $IT^0$ for all ample line bundles $H$. Next, let $M$ be any ample line bundle on the dual abelian variety $\hat{A}$ and $\phi: \hat{A} \to A$ is the isogeny defined by $M$. Let $\hat{M}$ be the Fourier-Mukai transform of $M$ on $A$ and let $\hat{M}^\vee$ be its dual. By [@Mu81] Proposition 3.11, $$\phi^*( \hat{M}^\vee) \cong \oplus^{h^0(M)} M.$$ Let $\hat{g}: \hat{X}:=X \times_A \hat{A} \to \hat{A}$ be the base change with $\varphi: \hat{X} \to X$ being étale. Clearly, $K_{\hat{X}}=\varphi^*K_X$ and $\varphi^*L$ is strictly nef on $\hat{X}$. Let $\mathcal{G}:=\hat{g}_*(\varphi^*(2K_X+2tL))$. By applying the above argument to $\varphi^*D_t$, we see that $\mathcal{G} \otimes M$ is $IT^0$ for all $M$. Thus $$\begin{array}{rl} \phi^*(\mathcal{F} \otimes \hat{M}^{\vee}) & = \phi^*( g_*( D_t \otimes \hat{M}^{\vee}))\\ &= \phi^*g_*( ( D_t \otimes g^* \hat{M}^{\vee})) \\ &= \hat{g}_* \varphi^*( ( D_t \otimes g^* \hat{M}^{\vee})) \\ &= \hat{g}_* (\varphi^* D_t \otimes \varphi^*g^* \hat{M}^{\vee})\\ &= \hat{g}_* (\varphi^* D_t \otimes \hat{g}^* \phi^* \hat{M}^{\vee})\\ &=\hat{g}_* (\varphi^* D_t \otimes \hat{g}^* ( \oplus M))\\ &=\oplus (\hat{g}_* \varphi^* D_t \otimes M) \end{array}$$ which is $IT^0$. Since $\mathcal{O}_A$ is a direct summand of $\phi_* \phi^* \mathcal{O}_A$, it follows that $\mathcal{F} \otimes \hat{M}^{\vee}$ is $IT^0$. By [@Ha04] Theorem 1.2, our claim follows. Let $X$ be a smooth projective threefold, $L$ strictly nef. Suppose that $\tilde q(X) > 0.$ Then $K_X + tL$ is ample for $t > 4.$ By the previous theorem we only have to treat the case that $q(X) = 0$. Then we choose a finite étale cover $h: \tilde X \to X$ such that $q(\tilde X) > 0.$ Hence $K_{\tilde X} + h^*(L) $ is ample for $t > n+1$ and so does $K_X + tL.$ There are two obstacles for extending Theorem 3.1 to all dimensions. The first is the use of 1.6(3) which has to be extended to higher dimensions. We will do this below. The second is the $g-$bigness of $K_X + tL.$ This means that $K_F + tL_F$ is big for the general fiber $F$ of $g.$ Thus we need to argue by induction on the dimension, but of course we are far from proving the conjecture for arbitrary manifolds (with vanishing irregularity). Let $X$ be an irreducible reduced projective Gorenstein variety with desingularization $\pi: \hat X \to X.$ Let $g: X \to A$ non-constant and $L$ be a strictly nef line bundle on $X$. Suppose that $K_X + t_0 L$ is $g-$big for some $t_0$ and set $\hat L = \pi^*(L).$ 1. The sheaf $\hat {{\mathcal F}}= g_*\pi_*(2K_{\hat X} + 2t \hat L) $ satisfies the generic vanishing theorem $$V^0(\hat {{\mathcal F}}) \supset V^1( \hat {{\mathcal F}}) \ldots \supset V^n(\hat {{\mathcal F}}).$$ 2. If $t \gg 0,$ then there exists $P \in {\rm Pic}^0(X)$ such that $$H^0(2K_X + 2tL + P) \ne 0.$$ \(1) This is just what the second part of the proof of Theorem 3.1 gives. (2) By (1) and the first arguments in the proof of Theorem 3.1 we obtain that $$H^0(2K_{\hat X} + 2 \hat L + \hat P) \ne 0$$ for some $\hat P \in {\rm Pic}^0(\hat X).$ Since $\hat P$ comes from $A$, it is of the form $\pi^*(P).$ Moreover we have $\pi_*(2K_{\hat X}) \subset 2K_X$ since $X$ is Gorenstein, hence claim (2) follows. Let $X_n$ be an irreducible reduced projective variety with a non-constant map $g: X \to A.$ Let $L$ be a strictly nef line bundle on $X$ and assume that $K_X + t_0L$ is $g-$big for some $t_0$ (e.g., Conjecture 1.2 holds in dimension $< n).$ Then $K_X + tL $ is ample for $t > n+1.$ We prove the claim by induction on $n.$ Since we argue numerically, we may ignore $P$ and choose by Lemma 3.4 $$D = \sum m_i D_i \in \vert 2K_X + 2tL \vert$$ for large $t.$ We may select a component, say $D_1,$ such that $\dim g(D_1) \ne 0$ and consider the non-constant map $g_1: D_1 \to A.$ By induction $K_{D_1} + t L_{D_1}$ is ample, if $t > n.$ We now adopt the methods of \[Se95,3.1\]. The equation $$0 = D \cdot (K_X + tL)^{n-1}$$ leads, via the nefness of $K_X + tL,$ to $$0 = D_1 \cdot K_X^j \cdot L^{n-1-j}, \ 0 \leq j \leq n-1. \eqno (*)$$ Choose $k$ such that $L_{D_1}^{n-k} \not \equiv 0,$ e.g., $k = n-1.$ Then $$0 < (K_{D_1} + tL_{D_1})^k \cdot L_{D_1}^{n-k} = D_1 \cdot (K_{D_1} + tL_{D_1})^k \cdot L_{D_1}^{n-k}.$$ By (\*) we obtain $$0 < D_1^2 \cdot L^{n-2}.$$ On the other hand, (\*) yields $$0 = D_1 \cdot (2K_X + tL) \cdot L^{n-2} = D_1 \cdot (\sum m_i D_i) \cdot L^{n-2} =$$ $$= m_1 D_1^2 \cdot L^{n-2} + \sum_{i \geq 2} m_i D_1 \cdot D_i \cdot L^{n-2} \geq m_1 D_1^2 L^{n-2},$$ a contradiction. Let $X_n$ be a projective manifold with Albanese map $\alpha: X \to A.$ Let $L$ be strictly nef on $X.$ Assume that $\dim \alpha (X) \geq n-2,$ or $\dim \alpha (X) = n-3 $ but the general fiber $F$ is not Calabi-Yau with $L_F \cdot c_2(F) = 0.$ Then $K_X + tL$ is ample for $t > n+1.$ Fano fibrations =============== We shall now (in particular) complete the proof of Theorem 0.2 (1). Observe that due to 1.5(2), 2.3 and 3.1, the only cases left are uniruled threefolds with $q=0$. These cases are thus settled by 4.1-2 and 5.1-2 below. In this section we settle the cases of del Pezzo fibrations over curves and elementary conic bundles over surfaces. Let $X$ be a smooth projective threefold, $L$ strictly nef on $X.$ Suppose that $X$ carries an extremal contraction $f: X \to B$ to a curve $B.$ Then $K_X + tL $ is ample for large $t.$ Since $K_{X} + tL$ and $L$ are strictly nef and since $\rho(X) = 2,$ $K_{X} + tL$ is clearly ample for large $t$ unless $-K_{X}$ and $L$ are proportional. Hence $X$ is Fano by Serrano’s theorem (1.5) which ends the proof. .2cm Let $f: X_3 \to S$ be a conic bundle with $\rho(X/S) = 1.$ If $L$ is strictly nef on $X$, then $K_X + tL$ is ample, for $t > 4.$ By Corollary 3.2 we may assume that $q(S) = 0, $ even after a finite étale cover of the smooth surface $S.$\ Since $\rho(X/S) = 1,$ we find a positive number $t_0$ such that $$K_X + t_0L = \phi^*(K_S+M)$$ with a ${{\mathbb Q}}-$divisor $M$ on $S.$ We cube the equation $t_0L = -K_X + f^*(K_S+M) $ to obtain $$0 = 3K_X^2 \cdot f^*(K_S+M) - 3 K_X \cdot f^*(K_S+M)^2 = 3 K_X^2 \cdot f^*(K_S+M) + 6 (K_S+M)^2.$$ From $K_X \cdot t_0L^2 = 0$ we get $K_X^2 \cdot f^*(K_S+M) = 0,$ hence in total $$(K_S+M)^2 = 0.$$ By applying (1.4), we find $\alpha \in {\overline {ME}}(X)$ such that $$K_X \cdot \alpha = L \cdot \alpha = 0,$$ in particular $ D \cdot \alpha = 0.$ Introducing $\gamma = \phi_*(\alpha) \in {\overline {ME}}(S),$ we obtain $$(K_S + M) \cdot \gamma = 0.$$ Notice that ${\overline {ME}}(S) $ is nothing than the nef cone, so $\gamma$ is a nef class. Next notice that we may choose $\gamma$ rational. In fact, since the rational points are dense in the nef cone on $S$ and since neither $K_S+M$ nor $-(K_S+M)$ are strictly positive functionals on the nef cone, we find rational points $x$ and $y$ in the nef cone such that $$(K_S+M) \cdot x \geq 0; \ (K_S+M) \cdot y \leq 0.$$ We may assume strict inequality in both cases, otherwise we are already done. Then choose $\lambda > 0$ such that $$(K_S+M) \cdot (x+\lambda y) = 0.$$ Noticing that $\lambda \in {{\mathbb Q}},$ we may substitute $\gamma$ by $x + \lambda y.$ Now multiply $\gamma$ suitably to obtain a nef line bundle $G$ such that $$(K_S+M) \cdot G = 0.$$ If now $G^2 > 0,$ then Hodge Index gives $K_S+M = 0$, so that $H^0(m(K_X+t_0L)) \ne 0$ for positive integers $m$ such that $mt_0 \in {{\mathbb N}}.$ Thus we may assume that $G^2=0$. Together with $(K_S+M)^2=(K_S+M) \cdot G=0$, one has $(K_S+M+ \tau G)^2=0$ for all $\tau$. Let $C \subset S$ be an irreducible curve. Then $$t_0^2 L^2 \cdot f^*(C) = (f^*(K_S+M) - K_{X})^2 \cdot f^*(C) = -2f^*(K_S+M)\cdot K_{X} \cdot C + K_{X}^2 \cdot f^*(C) =$$ $$= 4 (K_S+M) \cdot C - (4K_S + \Delta) \cdot C = (M - \Delta) \cdot C. \eqno (1)$$ The last equation is explained as follows. Outside the singular locus of $S,$ the map $f$ is a conic bundle; let $\Delta$ denote the closure of the discriminant locus. Then it is well-known that $$f_*(K_{X}^2) = -(4K_S + \Delta).$$ Now we restrict ourselves to curves $C$ with $C^2 \geq 0.$ Then clearly $L^2 \cdot f^*(C) \geq 0,$ hence $$(M-\Delta) \cdot C \geq 0, \eqno (2)$$ in particular $$(M-\Delta) \cdot G \geq 0. \eqno (3)$$ Moreover we have a strict inequality in (2) unless $C_0 = \emptyset $ and $L^2 \cdot f^*(C) = 0.$ The inequality (3) says im particular that $M$ is pseudo-effective. Thus the equation $(K_S+M) \cdot G = 0$ forces $\kappa (S) \leq 1.$ .2cm [**(I)**]{} We first assume $\kappa (S) = - \infty.$ Then $S$ is a rational surface. The case that $S=\mathbb{P}^2$ is easy. So we may assume that $\pi: S=S_n \to S_{n-1} \to ... \to S_0$ is a succession of blow-ups, where $S_0$ is a ruled surface with minimal section $C_0$ that $C_0^2=-e$. Now we write $$K_S+M = \pi^*( \alpha_1 C_0 + \beta_1 F ) + E_1$$ $$G=\pi^*( \alpha_2 C_0 + \beta_2 F ) + E_2,$$ where $E_1,E_2$ are divisors supported on exceptional curves. If $\alpha_2=0$, then it is clear that $E_2=0$ and $G=\beta_2 \pi^*F$. Then $(K_S+M) \cdot G =0$ gives $\alpha_1=0$ and $(K_S+M)^2=0$ gives $E_1=0$. So $K_S+M= \beta_1 \pi^* F$, and we are done. If $\alpha_2 \ne 0$, take $\tau =\frac{ -\alpha_1}{\alpha_2}$, then $$K_S+M+\tau G = (\beta_1 +\tau \beta_2) \pi^*F + E_1 + \tau E_2.$$ $(K_S+M+\tau G)^2=0$ gives $(E_1 + \tau E_2)^2=0$. It implies $E_1 + \tau E_2 =0$ by the negativity of intersection form of exceptional divisors. Let $\delta:=\beta_1 +\tau \beta_2$. If $\delta \ne 0$, then $K_S+M= -\tau G + \delta \pi^* F$. Again, one has $\alpha_2=G \cdot \pi^*F=0$ which is absurd. Therefore, $K_S+M = -\tau G$. Now since, by $(3)$, $M \cdot G \ge 0$, we have $K_S \cdot G \le 0$. By Riemann-Roch and the obvious vanishing $H^0(K_S-G) = 0,$ we have $$h^0(S,G) \ge \chi(\mathcal{O}_S)=1.$$ Hence $G$ is effective. $G$ is non-zero for otherwise $K_X + t_0L \equiv 0,$ hence $-K_X$ is strictly nef and thus $X$ is Fano. Therefore $m(K_X+t_0L)$ is effective for some $m \in \mathbb{Z}$ and we are done in Case (I). .2cm [**(II)**]{} Now suppose that $\kappa (S) \geq 0.$ Let $$\sigma: S \to S_0$$ be the minimal model. Since $\kappa (S) \geq 0,$ we conclude by (2) that $$K_S \cdot G = M \cdot G = 0.$$ Hence $$0 = \sigma^*(K_{S_0}) \cdot G + \sum a_iA_i \cdot G$$ with $A_i$ the $\sigma-$exceptional curves and $a_i$ suitable positive rational numbers. Thus $G = \sigma^*(G_0)$ with a nef line bundle $G_0$ on $S_0$; observe that $K_{S_0} \cdot G_0 = 0$ and that $G_0^2 = 0.$ .2cm Suppose that $\kappa (S) = 1.$ Then we consider the Iitaka fibration $g: S_0 \to B$ to the curve $B$ (necessarily $B = {{\mathbb P}}_1).$ We conclude that $G_0$ is a sum of fibers of $g$. Thus $G$ is a sum of fibers of $g \circ \sigma.$ Now consider the composed map $h: X \to B.$ Then it follows that $h_*(\alpha)$ consists of finitely many points. This means that we can find a fiber of $h$ such that $K_X+tL \vert F$ is not ample for large $t$. Thus $K_F+tL_F$ is not ample. If (the reduction of) $F$ is irreducible, this contradicts (1.6). If $F_i$ is a component of $F$ with multiplicity $a_i,$ then $a_iK_{F_i} + tL_{F_i}$ is a subsheaf of $K_F + tL_F \vert L_{F_i}$, and the contradiction is the same. .2cm Finally we have to treat the case $\kappa (S) = 0.$ Here we may assume that $S_0$ is K3. If $G_0^2 = 0$, then by Riemann-Roch $\kappa (G_0) = 1.$ Hence some multiple of $G_0$ is spanned, defining a morphism $g: S \to B.$ Since the divisor $M_0$ must be supported on fibers of $g,$ so does $\Delta.$ Thus we conclude by (3) for $b \in B$ that $$L^2 \cdot X_b = 0.$$ But for general $b$, the fiber $S_{0,b}$ is an elliptic curve and $X_b$ is a ${{\mathbb P}}_1-$bundle over $S_{0,b}$ since $\Delta$ does not meet $S_{0,b}.$ Moreover $L \vert X_b$ is strictly nef, hence ample, contradiction. [Suppose in (4.2) that $\phi: X \to S$ is a conic bundle, but not necessarily with $\rho(X/S) = 1.$ Then all arguments still remain valid if $ K_X + t_0 L$ is the $\phi-$pull-back of a ${{\mathbb Q}}-$bundle on $S$, for some rational $t_0$. ]{} Birational maps =============== In order to prove Conjecture 1.2 in the remaining uniruled cases, it is natural to consider the Mori program. If $X$ admits a contraction contracting a divisor to a point, the situation is easily understood. Let $X$ be a smooth projective threefold, $L$ strictly nef on $X$. Suppose that $X$ admits a birational Mori contraction $\phi: X \to Y$ contracting the exceptional divisor $E$ to a point. Then $K_X + tL$ is ample for $t > n+1.$ Suppose that $K_X + tL$ is not ample. Write $$K_X = \phi^*(K_Y) + aE;$$ then $a \in \{2,1,{{1}\over {2}} \}.$ Possibly after replacing $L$ by $2L$ in case $a = {{1}\over {2}},$ we can moreover write $$L = \phi^*(L') - bE$$ with a line bundle $L'$ on $Y.$ Notice that $b > 0$ since $L$ is strictly nef. Introduce $$D = b K_X + aL; \ D' = bK_Y + aL'.$$ Since $L'$ is again strictly nef, $K_Y + tL'$ is strictly nef for $t >> 0.$ Using (1.6)(1) on $X$ it is a simple matter to verify $$(K_Y + tL')^3 > 0$$ for large $t$, so that $K_Y + tL'$ is ample. Hence we find positive integers $p,q$ such that $pK_Y + qL'$ is spanned. Choose $S \in \vert pK_Y + qL' \vert $ smooth. Now a simple calcluation shows that $$D'^2 \cdot (pK_Y + qL')= D' \cdot (pK_Y + qL')^2 = 0.$$ Thus $D'_S \cdot (pK_Y + tL')_S = 0.$ Moreover $(D'_S)^2 = 0.$ Hence $D'_S \equiv 0$ by the Hodge index theorem. Thus $D' \equiv 0$ Hence $D \equiv 0$ so that $aL \equiv -bK_X$. Therefore $X$ is Fano by by Serrano (1.5) and $K_X + tL$ is ample for $t > 4,$ contradiction. In case that the contraction $\phi: X \to Y$ contracts a divisor to a curve $C$, the situation is more involved. The reason is that the induced line bundle $L'$ on $Y$ is not necessarily strictly nef, in fact we can have $L' \cdot C \leq 0.$ We have already shown that if $X$ admits a Mori fibration or a divisorial contraction to a point, then the conjecture holds. Since $X$ is smooth, it remains to consider the case that [*all*]{} the extremal rays produce a divisorial contraction to a nonsingular curve. Let $X$ be a smooth uniruled threefold, $L$ strictly nef on $X$. Suppose that all extremal contractions on $X$ contract a divisor to a curve. Then $K_X+tL$ is ample for large $t$. \(a) Let us fix some notations first. Let ${\phi_i},\ i \in I \subset {{\mathbb N}}$ be the extremal contractions on $X,$ with exceptional divisor $E_i$. Let $C_i:=\phi_i(E_i)$ so that $E_i$ is a $\mathbb{P}_1$ bundle over $C_i$. Let $[l_i] \in K_X^{<0}$ denotes the class of the contracted ruling lines in $E_i$. Let $$\mu:= \min\{ \frac{L \cdot l_i}{-K_X \cdot l_i}\}=\min\{ L \cdot l_i\} \in \mathbb{N}.$$ Reorder $I$ so that $\phi_1, \ldots, \phi_n$ are exactly those contractions with $$L \cdot l_i = \mu.$$ Then the divisor $$D:= L+ \mu K_X$$ is nef, as a consequence of the cone theorem and the definition of $\mu.$ Moreover, if $D \cdot B = 0$ for some $B \in \overline{NE}(X),$ then $K_X \cdot B \leq 0.$ In other words, $$D^{\perp} \cap \overline{NE}(X) \subset K_X^{\leq 0}.$$ In particular, if $B$ is an effective curve, then $D \cdot B = 0$ forces $K_X \cdot B < 0,$ because otherwise $K_X \cdot B = 0,$ hence $L \cdot B = 0,$ contradicting the strict nefness of $L.$ .2cm Our goal is to show that some multiple $mD=mL+m \mu K_X$ is effective, so that we are done by (1.6.3).\ Let $\phi = \phi_1 : X \to X_1 = X'$ be the contraction of $E = E_1.$ Let $[l]=[l_1]$ and set $L':= (\phi_*L)^{**}$, $$D':=L'+\mu K_{X'}, D:=\phi^*(D')=L+\mu K_{X}$$ and let $C = \phi(E)$. .2cm (b) We introduce the following numbers $$\tau := L' \cdot C, \sigma := K_{X'} \cdot C, \gamma = c_1(N^*_{C/X}).$$ Furthermore, let $g$ be the genus of $C$ and $\chi = 2-2g. $ .2cm First we treat the case $L' \cdot C > 0$ so that $L'$ is strictly nef. Then by induction on $\rho,$ the bundle $K_{X'} + tL'$ is ample, for $t > 4.$ Let $t_0 = {{1} \over {\mu}}.$ Then $K_{X'} + t_0 L'$ is nef, since $D'$ is nef. Let $\epsilon > 0$ be a small positive number. Then $$K_{X'} + {{t_0} \over {1 - \epsilon}} L'$$ is big (otherwise we would have $(K_{X'} + tL')^3 = 0$ for all $t$ which is absurd). Now the base point free theorem implies that some multiple $m(K_{X'} + t_0L')$ is spanned, hence $m'D'$ is spanned, and we are done. .2cm Thus we are reduced to $$L' \cdot C \leq 0.$$ Hence $K_{X'} \cdot C \geq 0$, and $C$ is rigid, since $L'.C'>0$ for every irreducible effective curve $C'\neq C$ on $X'$. We claim that: $$D' \cdot C \geq 1. \eqno (*)$$ In fact, we need only to exclude the case: $D' \cdot C = 0.$ Assuming that, we obtain $$L' \cdot C + \mu K_{X'} \cdot C = 0$$ and $$L_E \equiv - \mu K_X \vert E.$$ Since $L \cdot C_0 > 0,$ we have $K_{X'} \cdot C_0 < 0,$ hence $C_0$ moves. Since $C$ is rigid, $C_0$ can move only inside $E,$ hence $e \leq 0.$ Write $N^*_E \equiv C_0 + \lambda l.$ Then it is easily checked that $\lambda = {{1} \over {2}} \gamma + {{1} \over {2}} e $, in the notations of \[Ha 77\]; so that $$N^*_E = C_0 + ({{1} \over {2}}\gamma + {{1} \over {2}} e)l.$$ Since $L_E$ is strictly nef, so is $-K_X \vert E - N^*_E = C_0 + (e+2-2g-\lambda)l,$ so that we conclude: $$e + 2-2g - {{1} \over {2}}\gamma - {{1} \over {2}} e \geq {{e} \over 2}, \eqno (**)$$ hence $$2-2g \geq {{1} \over {2}} \gamma,$$ with strict inequality for $e = 0,$ since on those ruled surfaces all strictly nef line bundles are ample. .2cm By the adjunction formula we have $ \gamma = \sigma + (2-2g),$ hence $\sigma \leq 2-2g.$ Since $\sigma \geq 0,$ we obtain $g \leq 1.$ But a strictly nef divisor on a ruled surface over a rational or an elliptic curve is ample, hence the inequality (\*\*) is strict. Thus $g = e = 0$ and $\sigma < 2, \gamma \leq 3.$ So $$N^*_C = {{\mathcal O}}(k) \oplus {{\mathcal O}}(k)$$ with $0 < \gamma = 2k \leq 3,$ hence $k = 1$ and $\sigma = 0.$ So $K_{X'} \cdot C = 0 = L' \cdot C,$ and $L'$ is nef. If for large $t$, the nef bundle $K_{X'} + tL'$ is big, then we conclude as in the case $L' \cdot C > 0.$ So we may assume that $K_{X'} + tL')^3 = 0$ for all $t$. Then $K_{X'}^3 = 0.$ However $K_X^3 = 0$ forces $K_{X'}^3 = -2,$ contradiction. Thus we must have $$D' \cdot C > 0.$$ .2cm (c) Case: $D'^{\perp} \cap {\overline {NE}}(X) \subset K_{X'}^{\perp}.$\ We are going to rule out this case. Assume there is an irreducible curve $B' \in {\overline {NE}}(X')$ such that $D' \cdot B' = 0.$ Necessarily $B' \ne C.$ By assumption, $K_{X'} \cdot B' = 0.$ Let $B$ be the strict transform of $B'$ in $X.$ Then $D \cdot B = 0.$ Since $E \cdot B \geq 0,$ we also get $K_X \cdot B \geq 0.$ Since $L \cdot B > 0$ and $D \cdot B = 0,$ this is impossible. Hence $D'$ is strictly nef and by induction, $K_{X'} + tL'$ is ample for large $t$. On the other hand, $D'$ is not ample, hence there exists a nonzero class $B^*\in {\overline {NE}}(X')$ with $D' \cdot B^* = 0,$ hence $K_{X'} \cdot B^* = 0,$ by assumption.This is absurd. .2cm (d) Case: $D'^{\perp} \cap {\overline {NE}}(X) \not\subset K_{X'}^{\perp}.$\ Then we find $B' \in {\overline {NE}}(X')$ such that $D' \cdot B' = 0 $ and $K_{X'} \cdot B' < 0.$ Since $D'$ is nef, we also find an extremal curve $l'$ with $D' \cdot l' = 0$. Let $\phi': X' \to X''$ be the associated contraction. .2cm (d.1) Suppose that $\dim X'' \leq 2.$ Observe that $D' = \phi'^*(D'')$ with a nef bundle $D'$ on $X''.$ So if $\dim X'' \leq 1,$ the bundle $D'$ has a section and we are done. The same argument works if $\dim X'' = 2$ and $D''^2 \ne 0.$ In the remaining case we need more arguments. Let $l'$ be a smooth conic and assume that $l'$ meets $C.$ Let $l$ be its strict transform in $X.$ Then $K_X \cdot l \geq -1.$ Since $D \cdot l = 0$ and $L \cdot l > 0,$ necessarily $K_X \cdot l = -1$ and $E \cdot l = 1.$ Thus $l$ meets $C$ transversely in one point. The same computations show that $C$ cannot meet a singular conic. Thus $C$ is a section of $X' \to X''$ and $X \to X''$ is still a conic bundle. Then we conclude by Lemma 4.2 and Remark 4.3. .2cm (d.2) Suppose $\phi'$ is birational with exceptional divisor $E'.$\ If $C \subset E',$ then, $C$ being rigid, $E'$ must be ruled and $C$ is the exceptional section in $E'.$ Let $l'$ be a ruling line and $l$ its strict transform in $X.$ Then $K_X \cdot l = 0.$ Since $D \cdot l = 0,$ we have $L \cdot l = 0$, which is absurd.\ Things are more complicated when $E' \cap C$ is a finite non-empty set. Suppose first that $E'$ is not ${{\mathbb P}}_2$ with normal bundle ${{\mathcal O}}(-1).$ In this situation we find a rational curve $l' \subset E'$ meeting $C$ with $K_{X'} \cdot l' = -1.$ Let $\hat l$ be the strict transform in $X.$ Then $$\phi^*(l') = \hat l + al$$ with some positive integer $a.$ Since $D' \cdot l' = D \cdot l = 0,$ it follows $D \cdot \hat l = 0.$ Now $$K_X \cdot \hat l = -1 + a \geq 0.$$ Hence $D \cdot \hat l = (L + \mu K_X) \cdot l > 0,$ contradiction.\ It remains to do the case $E' = {{\mathbb P}}_2$ with normal bundle ${{\mathcal O}}(-1). $ Fixing a line $l' \subset E'$ which meets $C,$ the same computations as above show that $L \cdot \hat l = 1, \mu = 1, K_X \cdot \hat l = -1$ and $a = 1.$ Notice that $E'$ can meet $C$ only in one point (transversely). In fact, otherwise we choose two points in $E' \cap C$ and a line $l^*$ through these two points. Then the strict transform $\hat l^*$ satisfies $K_X \cdot \hat l^* \geq 0$, which is impossible, as already observed. Hence $\hat E'$ is ruled over ${{\mathbb P}}_1$ with fibers $\hat l.$ Since $\hat E' \cdot \hat l = -1, $ we can blow down $X$ along the projection $\hat E' \to {{\mathbb P}}_1$ to obtain $\psi: X \to Y,$ the blow-up of $Y$ along a smooth curve $C' \simeq {{\mathbb P}}_1.$ A priori it is not clear that $Y$ is projective. Let $L_Y = (\psi_*(L))^{**}.$ Then $$L = \psi^*(L_Y) - \hat E'.$$ Denoting by $C_0$ the exceptional section of $\hat E'$ and noticing that $N^*_{\hat E'} = C_0 + \hat l,$ we obtain $$L \vert \hat E' = C_0 + (L_Y \cdot C')+1) \hat l.$$ Since $L \vert \hat E'$ is ample, it follows that $L_Y \cdot C' > 0$ so that $L_Y$ is strictly nef on the Moishezon manifold $Y$. Then $Y$ has to be projective: otherwise by \[Pe86\] we find an irreducible curve $D$ and a positive closed current $T$ on $Y$ such that $[D+T] = 0.$ But $L_Y \cdot D > 0$ and $L_Y \cdot T \geq 0.$ Now, $Y$ being projective, we conclude by the first part of (b).\ If finally $E' \cap C = \emptyset, $ then the strict transform of $E'$ in $X$ is some $E_j, 2 \leq j \leq n,$ hence defines an extremal contraction on $X$ with the same properties as $\phi$ and we can continue by induction. Since we assume $X$ uniruled, after finitely many steps we arrive at $\dim X^{[m]} \leq 2$ and argue as above, Higher dimensions ================= .2cm In higher dimensions it is certainly very difficult to deal with Fano fibrations; however it is instructive to look at ${{\mathbb P}}_k-$bundles to get an idea on the higher dimensional case. Here we can calculate explicitly. Let $X$ be a ${{\mathbb P}}_k-$bundle over a smooth surface $S.$ Suppose that $L$ is strictly nef on $X$. Then $K_X+tL$ is ample for $t > k+3.$ After possibly performing a finite étale cover, we may assume that $X$ is the projectivisation of a rank $(r+1)$-bundle ${{\mathcal E}}$ on $S.$ If we allow ${{\mathcal E}}$ to be a ${{\mathbb Q}}-$bundle, we may assume that $$L = {{\mathcal O}}_{{{\mathbb P}}({{\mathcal E}})}(k)$$ with some positive number $k.$ We also introduce $\zeta = {{\mathcal O}}_{{{\mathbb P}}({{\mathcal E}})}(1).$ Notice that $\det {{\mathcal E}}$ is strictly nef and suppose that $K_X+tL$ is not ample. Then $$K_X^j \cdot L^{r+2-j} = 0$$ for all $j$ by (1.5). First recall the following $$\zeta^{r+1} - \pi^*c_1({{\mathcal E}}) \zeta^{r} +\pi^*c_2({{\mathcal E}}) \zeta^{r-1}=0,$$ and $$K_X= -(r+1) \zeta + \pi^*(\det {{\mathcal E}}+ K_S).$$ The equation $L^{r+2}=0$ immediately leads to $$\zeta^{r+2}=c_1({{\mathcal E}})^2-c_2({{\mathcal E}})=0. \eqno(5)$$ Secondly, combining with $\zeta^{r+2}=0$, the equation $L^{r+1} \cdot K_X=0$ leads to $$\zeta^{r+1} \cdot \pi^*(c_1({{\mathcal E}}) + K_S) = c_1({{\mathcal E}}) \cdot (c_1({{\mathcal E}}) + K_S)=0.\eqno(6)$$ Moreover, the equation $L^{r} \cdot K_X^2=0$ leads to $$\zeta^{r} \cdot \pi^*(c_1({{\mathcal E}}) + K_S)^2 = (c_1({{\mathcal E}}) + K_S)^2=0.\eqno(7)$$ By $(6),(7)$, we have $K_S \cdot (c_1({{\mathcal E}}) + K_S)=0$ and hence $K_S^2 =c_1({{\mathcal E}})^2$. Since $\det {{\mathcal E}}$ is strictly nef, equation $(6)$ yields that $K_S^2=c_1({{\mathcal E}})^2 \ge 0$ and $c_1({{\mathcal E}}) \cdot K_S \le 0$. First suppose that $\kappa (S) \geq 0.$ Then $K_S \cdot \det {{\mathcal E}}= 0$ and $K_S^2 = 0$ for $\det {{\mathcal E}}$ being strictly nef. Hence $K_S \equiv 0.$ Then by (1.5) $\det {{\mathcal E}}$ is ample, contradicting $c_1({{\mathcal E}})^2 = K_S^2.$\ It remains to consider $\kappa (S) = - \infty.$ Since $K_S^2 \geq 0,$ $S$ is either rational or a minimal ruled surface over an elliptic curve. In the latter case, $K_S^2 = 0,$ hence $c_1({{\mathcal E}})^2 = 0.$ On the other hand, any strictly nef divisor on a ruled surface over an elliptic curve is ample (use \[Ha77,V.2\]), a contradiction.\ In case of a rational surface $S$, choose a positive integer $m$ such that $m \det {{\mathcal E}}$ is Cartier. Then Riemann-Roch and $(K_S+\det {{\mathcal E}})^2 = 0$ show that $h^0(m(K_S+\det {{\mathcal E}})) > 0.$ This contradicts via $(K_S+\det {{\mathcal E}}) \cdot \det {{\mathcal E}}= 0$ the strict nefness of $\det {{\mathcal E}}.$ [BDPP04]{} Boucksom,S.;Demailly,J.P.;Paun,M.;Peternell,Th.: The pseudo-effective cone of a compact Kähler manifold and varieties of negative Kodaira dimension. math.AG/0405285 Hacon, C.: A derived category approach to generic vanishing. J. Reine Angew. Math. [**575**]{}, 173-187 (2004). Hartshorne, R.:Ample subvarieties of algebraic varieties. Lecture Notes in Mathematics, [**156**]{}, Springer Verlag, (1970). Hartshorne, R.:Algebraic geometry. Graduate Texts in Mathematics, [**52**]{}, Springer-Verlag, (1977). Kawamata, Y.; Matsuki, K.; Matsuda, K.: Introduction to the minimal model program. Adv. Stud. Pure Math. [**10**]{} (1987), 283-360 Mukai,S.: Duality between $D(X)$ and $D(\hat X)$ with application to Picard sheaves. Nagoya Math. J. [**81**]{}, 153-175 (1981) Serrano,F.: Strictly nef divisors and Fano threefolds. J. Reine Angew. Math. [**464**]{}, 187-206 (1995). -- -- -- -- -- -- .2cm .2cm Frédéric Campana, Département de Mathématiques, Université de Nancy, F-54506 Vandoeuvre-les-Nancy, France,\ [email protected] .2cm Jungkai Alfred Chen, Department of Mathematics, National Taiwan University, Taipei 106, Taiwan,\ [email protected] .2cm Thomas Peternell, Mathematisches Institut, Universit" at Bayreuth, D-95440 Bayreuth,\ [email protected]

ArXiv

--- abstract: | Building upon ideas of the second and third authors, we prove that at least $2^{(1-\varepsilon)\frac{\log s}{\log \log s}}$ values of the Riemann zeta function at odd integers between 3 and $s$ are irrational, where $\varepsilon$ is any positive real number and $s$ is large enough in terms of $\varepsilon$. This lower bound is asymptotically larger than any power of $\log s$; it improves on the bound $ \frac{1-\varepsilon}{1+\log 2} \log s $ that follows from the Ball–Rivoal theorem. The proof is based on construction of several linear forms in odd zeta values with related coefficients. author: - 'Stéphane Fischler, Johannes Sprang and Wadim Zudilin' title: Many odd zeta values are irrational --- Introduction {#introduction .unnumbered} ============ When $s\geq 2$ is an even integer, the value $\zeta(s)$ of the Riemann zeta function is a non-zero rational multiple of $\pi^s$ and, therefore, a transcendental number. On the other hand, no such relation is expected to hold for $\zeta(s)$ when $s\geq 3$ is odd; a folklore conjecture states that the numbers $\pi$, $\zeta(3)$, $\zeta(5)$, $\zeta(7),\ldots$ are algebraically independent over the rationals. This conjecture is predicted by Grothendieck’s period conjecture for mixed Tate motives. But both conjectures are far out of reach and we do not even know the transcendence of a single odd zeta value. It was only in 1978 when Apéry astonished the mathematics community by his proof [@Apery] of the irrationality of $\zeta(3)$ (see [@SFBou] for a survey). The next breakthrough was taken in 2000 by Ball and Rivoal [@BR; @RivoalCRAS] who proved the following: \[thBR\] Let ${\varepsilon}> 0$. Then for any $s\geq 3$ odd and sufficiently large with respect to ${\varepsilon}$, we have $$\dim_{\mathbb{Q}}{\operatorname{Span}}_{\mathbb{Q}}( 1, \, \zeta(3), \, \zeta(5), \, \zeta(7), \, \ldots , \zeta(s)) \geq \frac{1-{\varepsilon}}{1+\log 2}\,\log s.$$ Their corresponding result for small $s$ has been refined several times [@Zudilincentqc; @SFZu], but the question whether $\zeta(5)$ is irrational remains open. The proof of Theorem \[thBR\] involves the well-poised hypergeometric series $$\label{eqBR} n!^{s-2r}\, \sum_{t=1}^\infty \frac{ \prod_{j=0}^{ (2r+1)n} (t-rn+j) }{ \prod_{j=0}^{ n} (t+j)^{s+1}},$$ which happens to be a ${\mathbb{Q}}$-linear combination of 1 and odd zeta values when $s$ is odd and $n$ is even, and Nesterenko’s linear independence criterion [@Nesterenkocritere]. The bound $\frac{1-{\varepsilon}}{1+\log 2}\log s$ follows from comparison of how small the linear combination is with respect to the size of its coefficients, after multiplying by a common denominator to make them integers. To improve on this bound using the same strategy, one has to find linear combinations that are considerably smaller, with not too large coefficients,—it comes out to be a rather difficult task. This may be viewed as an informal explanation of why the lower bound in Theorem \[thBR\] has never been improved for large values of $s$, whereas the theorem itself has been generalized to several other families of numbers. Using (with $s=20$) the series $$n!^{s-6}\, \sum_{k=1}^\infty \left.\Big( \frac{{\mathrm{d}}}{{\mathrm{d}}t}\Big)^2 \bigg( \Big(t+\frac{n}{2}\Big) \frac{\prod_{j=0}^{ 3n} (t-n+j)^3 }{ \prod_{j=0}^{ n} (t+j)^{s+3}}\bigg)\right|_{t=k} ,$$ which is a ${\mathbb{Q}}$-linear combination of 1 and odd zeta values starting from $\zeta(5)$, Rivoal has proved [@vingtetun] that among the numbers $\zeta(5)$, $\zeta(7)$, …, $\zeta(21)$, at least one is irrational. This result has been improved by the third author [@Zudilinonze]: among the four numbers $\zeta(5)$, $\zeta(7)$, $\zeta(9)$, $\zeta(11)$, at least one is irrational; and he also showed [@Zudilincentqc] that, for any odd $\ell \geq 1$, there is an irrational number among $\zeta(\ell+2)$, $\zeta(\ell+4)$, …, $\zeta(8\ell-1)$. Proofs of these results do not require use of linear independence criteria: if a sequence of ${\mathbb{Z}}$-linear combinations of real numbers from a given (fixed) collection tends to 0, and is non-zero infinitely often, then at least one of these numbers is irrational. A drawback of this approach is that it only allows one to prove that [*one*]{} number in a family is irrational. The situation has drastically changed when the third author introduced [@Zudilintrick] a new method (see also [@KrattZ]). He casts (with $s=25$) the rational function in the form $$R(t) = 2^{6n}n!^{s-5}\, \frac{\prod_{j=0}^{6n} (t-n+\frac{j}{2})}{ \prod_{j=0}^{ n} (t+j)^{s+1}}$$ and proves that both series $$\sum_{t=1}^\infty R(t) \quad\mbox{and}\quad \sum_{t=1}^\infty R\Big(t+\frac12\Big)$$ are ${\mathbb{Q}}$-linear combinations of 1, $\zeta(3)$, $\zeta(5)$, …, $\zeta(s)$ with [*related*]{} coefficients. This allows him to eliminate one odd zeta value, and to prove that [*at least two*]{} zeta values among $\zeta(3)$, $\zeta(5)$, …, $\zeta(25)$ are irrational. In view of Apéry’s Theorem, the result means that one number among $\zeta(5)$, …, $\zeta(25)$ is irrational—nothing really novel, but the method of proof is new and more elementary than the ones in [@vingtetun] and [@Zudilinonze] as it avoids use of the saddle point method. More importantly, the method allows to prove the irrationality of at least two zeta values in a family without having to produce very small linear forms. The same strategy has been adopted by Rivoal and the third author [@RZnote] to prove that among $\zeta(5)$, $\zeta(7)$, …, $\zeta(69)$, at least two numbers are irrational. The method in [@Zudilintrick] has been generalized by the second author [@Sprang], who introduces another integer parameter $D>1$ and considers the rational function $$\label{eqSprang} R(t) = D^{6(D-1)n}n!^{s-3D-1}\,\frac{ \prod_{j=0}^{3Dn} (t-n+\frac{j}{D})}{ \prod_{j=0}^{ n} (t+j)^{s+1}}.$$ He proves that for any divisor $d$ of $D$ the series $$\sum_{j=1}^d \sum_{t=1}^\infty R\Big(t+\frac{j}{d}\Big)$$ is a ${\mathbb{Q}}$-linear combination of 1, $\zeta(3)$, $\zeta(5)$, …, $\zeta(s)$. The crucial point of this construction is that each $\zeta(i)$ appears in this ${\mathbb{Q}}$-linear combination with a coefficient that depends on $d$ in a very simple way. This makes it possible to eliminate from the entire collection of these linear combinations as many odd zeta values as the number of divisors of $D$. Finally, taking $D$ equal to a power of 2 and $s$ sufficiently large with respect to $D$, the second author proves that at least $\frac{\log D}{\log 2}$ numbers are irrational among $\zeta(3)$, $\zeta(5)$, …, $\zeta(s)$. This strategy represents a new proof that $\zeta(i)$ is irrational for infinitely many odd integers $i$. Building upon the approach in [@Zudilintrick] and [@Sprang] we prove the following result. \[th1\] Let ${\varepsilon}> 0$, and $s\geq 3$ be an odd integer sufficiently large with respect to ${\varepsilon}$. Then among the numbers $$\zeta(3), \, \zeta(5), \, \zeta(7), \, \ldots , \zeta(s),$$ at least $$2^{(1-{\varepsilon})\frac{\log s}{\log \log s}}$$ are irrational. In this result, the lower bound is asymptotically greater than $\exp(\sqrt{\log s})$, and than any power of $\log s$; “to put it roughly, \[it is\] much more like a power of $s$ than a power of $\log s$” [@HW Chapter XVIII, §1]. In comparison, Theorem \[thBR\] gives only $\frac{1-{\varepsilon}}{1+ \log 2} \log s$ irrational odd zeta values, but they are linearly independent over the rationals, whereas Theorem \[th1\] ends up only with their irrationality. Our proof of Theorem \[th1\] follows the above-mentioned strategy of the second and third authors. The main new ingredient, compared to the proof in [@Sprang], is taking $D$ large (about $s^{1-2{\varepsilon}}$) and equal to the product of the first prime numbers (the so-called primorial)—such a number has asymptotically the largest possible number of divisors with respect to its size (see [@HW Chapter XVIII, §1]). To perform the required elimination of a prescribed set of odd zeta values, we need to establish that a certain auxiliary matrix is invertible. Whereas the second author’s choice of $D$ in [@Sprang] allows him to deal with elementary properties of a Vandermonde matrix, we use at this step a generalization of the corresponding result. We give three different proofs of the latter, based on arguments from combinatorics of partitions, from linear algebra accompanied with a lemma of Fekete, and from analysis using Rolle’s theorem. The structure of this paper is as follows. In §\[sec2\] we construct linear forms in values of the Hurwitz zeta function. Denominators of the coefficients are studied in §\[sec3\]; and the asymptotics of the linear forms are dealt with in §\[sec4\]. Section \[sec5\] is devoted to the proof that an auxiliary matrix is invertible. Finally, we establish Theorem \[th1\] in §\[sec6\]. Construction of linear forms {#sec2} ============================ From now on we let $s$, $D$ be positive integers such that $s \geq 3D$; we assume that $s$ is odd. Let $n$ be a positive integer, such that $Dn $ is even. Consider the following rational function: $$R_n(t) = D^{3Dn} \, \, n!^{s+1-3D} \, \, \frac{ \prod_{j=0}^{3Dn} (t-n+\frac{j}{D})}{ \prod_{j=0}^{ n} (t+j)^{s+1}}$$ which, of course, depends also on $s$ and $D$. Notice that the difference of the function from the corresponding one in [@Sprang] is in the factor $D^{3Dn}$ instead of $D^{6(D-1)n}$ (see Eq. ). Similar rational functions have already been considered, see [@Catalan] for the case $D=2$ and [@Nash; @Nishimoto; @SFcaract] for general $D$. However the “central” factors $t-n+\frac{j}{D}$ with $Dn < j < 2Dn$ are missing, and (as the second author noticed [@Sprang]) they play a central role in the arithmetic estimates (see Lemma \[lemarith\] below). \[Rem-r\] Though one can implement an additional parameter $r$ in the definition of the rational function $R_n(t)$, in a way similar to the one for the Ball–Rivoal series , we have verified that this does not bring any improvement to the result of Theorem \[th1\]. The rational function $R_n(t)$ has a partial fraction expansion $$\label{eqelsples} R_n(t) = \sum_{i=1}^s \sum_{k=0}^n \frac{a_{i,k}}{(t+k)^i}.$$ For any $j\in \{1,\ldots,D\}$, take $$r_{n,j} = \sum_{m=1}^\infty R_n\Big(m+\frac{j}{D}\Big).$$ We recall that the Lerch and Hurwitz zeta functions are defined by $$\Phi(z,i,\alpha) = \sum_{n=0}^{\infty}\frac{z^n}{(n+\alpha)^i} \quad\mbox{and}\quad \zeta(i,\alpha) = \Phi(1,i,\alpha) = \sum_{n=0}^\infty \frac1{(n+\alpha)^i},$$ where $\alpha>0$ and also $i\geq 2$ for the latter. The following is precisely [@Sprang Lemma 1.5]; the change of the normalizing factor $D^{3Dn}$ does not affect the statement. \[lemconstru\] For each $j\in \{1,\ldots,D\}$, we have $$r_{n,j} = \rho_{0,j}+\sum_{\substack{3\leq i \leq s\\i \;\mbox{\scriptsize odd}}} \rho_i \, \zeta\Big(i, \frac{j}{D}\Big),$$ where $$\rho_i = \sum_{k=0}^n a_{i,k} \quad\mbox{for $3\leq i \leq s$, \ $i$ odd},$$ does not depend on $j$, and $$\label{eqdefrhoz} \rho_{0,j} = - \sum_{k=0}^n \sum_{\ell=0}^k \sum_{i=1}^s \frac{ a_{i,k}}{(\ell+\frac{j}{D})^i} .$$ We follow the strategy of proofs in [@Zudilintrick Lemma 3] and [@Sprang Lemma 1.5]. Let $z$ be a real number such that $0<z<1$. We have $$\begin{aligned} \sum_{m=1}^\infty R_n\Big(m+\frac{j}{D}\Big) z^m &= \sum_{m=1}^\infty \sum_{i=1}^s \sum_{k=0}^n \frac{a_{i,k} z^m }{(m+k+\frac{j}{D})^i} \\ &= \sum_{i=1}^s \sum_{k=0}^n a_{i,k} z^{-k} \sum_{m=1}^\infty\frac{ z^{m+k} }{(m+k+\frac{j}{D})^i} \\ &= \sum_{i=1}^s \sum_{k=0}^n a_{i,k} z^{-k} \bigg( \Phi\Big(z, i, \frac{j}{D}\Big) - \sum_{\ell = 0}^k \frac{z^\ell}{(\ell+\frac{j}{D})^i} \bigg). \end{aligned}$$ Now we let $z$ tend to 1 in the equality we have obtained; the left-hand side tends to $r_{n,j} $. On the right-hand side, the term involving the Lerch function with $i=1$ has coefficient $\sum_{k=0}^n a_{1,k} z^{-k} $. Since $ \Phi(z,1,\frac{j}{D}) $ has only a logarithmic divergence as $z\to 1$ and $$\sum_{k=0}^n a_{1,k} =\lim_{t\to \infty } t R_n(t) =0,$$ this term tends to 0 as $z\to1$. All other terms have finite limits as $z\to 1$, so that $$r_{n,j} = \rho_{0,j}+\sum_{i=2}^s \rho_i \, \zeta\Big(i, \frac{j}{D}\Big),$$ where $ \rho_{0,j} $ is given by Eq. , and $ \rho_i = \sum_{k=0}^n a_{i,k} $ for any $i\in \{2,\ldots,s\}$. To complete the proof, we apply the symmetry phenomenon of [@BR; @RivoalCRAS]. Since $s$ is odd and $Dn $ is even we have $R_n(-n-t) = - R_n(t)$. Now the partial fraction expansion is unique, so that $a_{i,n-k} = (-1)^{i+1} a_{i,k}$ for any $i$ and $k$. This implies that $\rho_i = 0$ when $i$ is even, and Lemma \[lemconstru\] follows. Arithmetic estimates {#sec3} ==================== As usual we let $d_n = \operatorname{lcm}(1,2,\ldots,n)$. \[lemarith\] We have $$\label{arith-I} d_n^{s+1-i} \rho_i \in{\mathbb{Z}}\quad\mbox{for}\; i = 3, 5, \ldots, s,$$ and $$\label{arith-II} d_{n+1}^{s+1} \rho_{0,j} \in{\mathbb{Z}}\quad\mbox{for any}\; j\in \{1,\ldots,D\}.$$ For part we use the strategy of the proof of [@SFcaract Lemma 4.5]; note that [@Sprang Lemma 1.3] does not apply in our present situation because of the different normalization of the rational function $R_n(t)$ compared to the one in . To establish we follow the proof of [@Sprang Lemma 1.4]; we use $d_{n+1}$ here instead of $d_n$ to include the case corresponding to $j=D$. For any $\alpha\in\frac1{D}{\mathbb{Z}}$ we introduce $$F_\alpha(t) = D^n \, \, \frac{\prod_{j=1}^n (t+\alpha+\frac{j}{D})}{\prod_{j=0}^n( t+j)} = \sum_{k=0}^n \frac{A_{\alpha,k}}{t+k},$$ where $A_{\alpha,k}$ is an integer in view of the explicit formulas $$(-1)^k A_{\alpha,k} = \binom{n}{k} \frac{\prod_{j=1}^n (D( \alpha-k)+j )}{n!} = \begin{cases} \binom{n}{k} \binom{D( \alpha-k) +n}{n} &\mbox{if}\; \alpha-k\geq 0, \\ 0 &\mbox{if}\; \frac{-n}{D} \leq \alpha-k < 0, \\ (-1)^n \binom{n}{k} \binom{D(k- \alpha) -1}{n} &\mbox{if}\; \alpha-k < \frac{-n}{D} . \end{cases}$$ We also consider $$G(t) = \frac{n!}{\prod_{j=0}^n (t+j)} = \sum_{k=0}^n \frac{ (-1)^k \binom{n}{k} }{t+k},$$ so that $$\label{eqpro} R_n(t) = (t-n) \, G(t)^{s+1-3D} \, \prod_{\ell = 0 }^{3D-1} F_{-n+\frac{\ell n}{D}}(t).$$ From this expression we compute the partial fraction expansion of $R_n(t)$ using the rules $$\frac{t-n}{t+k} = 1 - \frac{k+n}{t+k} \quad\mbox{and}\quad \frac{1}{(t+k)(t+k')} = \frac{1}{(k'-k)(t+k)}+\frac{1}{(k- k')(t+k')} \quad\mbox{for}\; k\neq k'.$$ A denominator appears each time the second rule is applied, and the denominator is always a divisor of $d_n$ (see [@Colmez] or [@Zudilintrick Lemma 1]). This happens $s+1-i$ times in each term that contributes to $a_{i,k}$ because there are $s+1$ factors in the product (apart from $t-n$). Therefore, $$d_{n}^{s+1-i} a_{i,k} \in {\mathbb{Z}}\quad\mbox{for any $i$ and $k$},$$ implying . We now proceed with the second part of Lemma \[lemarith\], that is, with demonstrating the inclusions . Recall from Lemma \[lemconstru\] that $$\label{eqrhoz} d_{n+1}^{s+1} \rho_{0,j} = - \sum_{k=0}^n \sum_{\ell=0}^k \bigg( \sum_{i=1}^s \frac{d_{n+1}^{s+1} a_{i,k}}{(\ell+\frac{j}{D})^i}\bigg).$$ If $j=D$ then $$d_{n+1}^{s+1-i} a_{i,k} \quad\text{and}\quad \frac{d_{n+1}^i}{(\ell+\frac{j}{D})^i}$$ are integers for any $k$, $\ell$ and $i$, so that $d_{n+1}^{s+1} \rho_{0,j} \in{\mathbb{Z}}$. From now on, we assume that $1\leq j \leq D-1$ and we prove that for any $k$ and any $\ell$ the internal sum over $i$ in Eq.  is an integer. With this aim in mind, fix integers $k_0$ and $\ell_0$, with $0\leq \ell_0 \leq k_0 \leq n$, and assume that the corresponding sum is not an integer. Since $1\leq j \leq D-1$ we have $R_n(\ell_0-k_0+\frac{j}{D}) = 0$, so that $$\label{eqentier} \sum_{i=1}^s \frac{d_{n+1}^{s+1} a_{i,k_0}}{(\ell_0+\frac{j}{D})^i} = - \sum_{\substack{k=0\\k\neq k_0}}^n \sum_{i=1}^s \frac{d_{n+1}^{s+1} a_{i,k}}{(\ell_0-k_0+k +\frac{j}{D})^i} .$$ This rational number is not an integer: it has negative $p$-adic valuation for at least one prime number $p$. Therefore, on either side of there is at least one term with negative $p$-adic valuation: there exist $i_0,i_1\in\{1,\ldots,s\}$ and $k_1\in \{0,\ldots,n\}$, $k_1\neq k_0$, such that $$v_p\bigg( \frac{d_{n+1}^{s+1} a_{i_0,k_0}}{( \ell_0+\frac{j}{D})^{i_0}} \bigg) < 0 \quad \mbox{and} \quad v_p\bigg( \frac{d_{n+1}^{s+1} a_{i_1,k_1}}{( \ell_0-k_0+k_1+\frac{j}{D})^{i_1}} \bigg) < 0.$$ Since $d_{n+1}^{s+1-i} a_{i,k}\in{\mathbb{Z}}$ for any $i$ and $k$, this leads to $$v_p\bigg( \frac{d_{n+1}^{i_0} }{( \ell_0+\frac{j}{D})^{i_0}} \bigg) < 0 \quad \mbox{and} \quad v_p\bigg( \frac{d_{n+1}^{i_1} }{( \ell_0-k_0+k_1+\frac{j}{D})^{i_1}} \bigg) < 0,$$ implying $$\min\bigg( v_p \Big( \ell_0+\frac{j}{D} \Big) , \, v_p \Big( \ell_0-k_0+k_1+\frac{j}{D} \Big)\bigg)> v_p(d_{n+1}).$$ As $k_0 - k_1 = ( \ell_0+\frac{j}{D} ) - ( \ell_0-k_0+k_1+\frac{j}{D} )$, we deduce that $v_p(k_0 - k_1 ) > v_p(d_{n+1})$, which is impossible in view of the inequality $0 < |k_0 - k_1 | \leq n$. The contradiction completes the proof of Lemma \[lemarith\]. \[remarith\] It is made explicit in [@RZnote], for a particular situation considered there, that the inclusions in Lemma \[lemarith\] can be sharpened as follows: $$\Phi_n^{-1}d_n^{s+1-i} \rho_i \in{\mathbb{Z}}\quad\mbox{for}\; i = 3, 5, \ldots, s,$$ and $$\Phi_n^{-1}d_{n+1}^{s+1} \rho_{0,j} \in{\mathbb{Z}}\quad\mbox{for any}\; j\in \{1,\ldots,D\},$$ where $\Phi_n=\Phi_n(D)$ is a certain product over primes in the range $2\le p\le n$, whose asymptotic behavior $$\phi=\phi(D)=\lim_{n\to\infty}\frac{\log\Phi_n}n$$ can be controlled by means of the prime number theorem. It is possible to show that the quantity $\phi(D)/D$ increases to $\infty$ and at the same time $\phi(D)/(D\log^\varepsilon D)\to0$ as $D\to\infty$, for any choice of $\varepsilon>0$. Later, we choose $D$ such that $ D\log D <s $, implying that the arithmetic gain coming from the factors $\Phi_n^{-1}$ is asymptotically negligible as $s\to\infty$. Asymptotic estimates of the linear forms {#sec4} ======================================== The following lemma is proved along the same lines as [@Sprang Lemma 2.1] (see also [@Zudilintrick Lemma 4] and the second proof of [@BR Lemme 3]). The difference is that here we only assume $\frac{s}{ D\log D} $ to be sufficiently large, whereas in [@Sprang] parameter $D$ is fixed and $ s\to\infty$. \[lemasy\] Assume that $$\label{eqhyp} \frac{s}{ D\log D} \quad\mbox{is sufficiently large.}$$ Then we have $$\label{eqasyun} \lim_{n\to\infty} r_{n,j}^{1/n} = g(x_0) <3^{-(s+1)} \quad\mbox{and}\quad \lim_{n\to\infty} \frac{r_{n,j'}}{r_{n,j}} = 1 \quad\mbox{for any}\; j,j'\in\{1,\ldots,D\},$$ where $$g(x) = D^{ 3D } \, \frac{ (x+3)^{3D} (x+1)^{ s+1 }}{ ( x+2)^{2(s+1)}}$$ and $x_0$ is the unique positive root of the polynomial $$(X+3)^D(X+1)^{s+1} - X^D (X+2)^{s+1}.$$ For $j\in \{1,\ldots,D\}$ and $k\geq 0$, let $$c_{k,j} = R_n\Big( n+k+\frac{j}{D}\Big) = D^{ 3D n} \, n!^{s+1-3D} \, \frac{ \prod_{\ell=0}^{3Dn} ( k +\frac{j+\ell }{D})}{ \prod_{\ell=0}^{ n} (n+k+\ell+\frac{j}{D})^{s+1}},$$ so that $$r_{n,j} = \sum_{m=1}^\infty R_n\Big(m+\frac{j}{D}\Big) = \sum_{k=0}^\infty c_{k,j}$$ is a sum of positive terms. We have $$\label{eqquok} \frac{c_{k+1,j}}{c_{k,j}} = \bigg( \prod_{\ell = 1}^D \frac{k+3n+\frac{j+\ell}{D}}{k+\frac{j+\ell-1}{D}}\bigg) \, \bigg( \frac{k+n+\frac{j}{D}}{k+2n+1+ \frac{j}{D}}\bigg)^{s+1}$$ implying that, for any $j$, the quotient $\frac{c_{k+1,j}}{c_{k,j}}$ tends to $ f(\kappa)$ as $n\to\infty$ assuming $k\sim \kappa n$ for $\kappa>0$ fixed, where $$f(x) = \Big( \frac{ x+ 3 }{x}\Big) ^D \Big( \frac{ x+ 1 }{x+2}\Big) ^{s+1}.$$ For the logarithmic derivative of this function we have $$\frac{f'(x)}{f(x)} = \frac{D}{x+3} - \frac{D}{x }+\frac{s+1}{x+ 1 } - \frac{s+1}{x+ 2} = \frac{ ax^2+bx+c}{x(x+1)(x+2)(x+3)}$$ with $a = s+1-3D > 0$ and $c = -6 D < 0$, hence the derivative $f'(x)$ vanishes exactly at one positive real number $x_1$. This means that the function $f(x)$ decreases on $(0, x_1]$ and increases on $[x_1,+\infty)$. Since $\lim_{x\to 0^+} f(x) = +\infty$ and $\lim_{x\to+\infty} f(x) = 1$, we deduce that there exists a unique positive real number $x_0$ such that $f(x_0 ) = 1$. Let us now prove . As in [@Bruijn §3.4] we wish to demonstrate that the asymptotic behaviour of $r_{n,j}$ is governed by the terms $c_{k,j}$ with $k$ close to $x_0n $ (see Eq.  below). To begin with, notice that $$\begin{aligned} c_{k,j} &= D^{-1} \, n!^{s+1-3D} \, \frac{ \prod_{\ell=0}^{3Dn} ( Dk+j+\ell )}{ \prod_{\ell=0}^{ n} (n+k+\ell+\frac{j}{D})^{s+1}}\\ &= D^{-1} \, n!^{s+1-3D} \, \frac{(3Dn+Dk+j)!}{(Dk+j-1)!} \, \frac{\Gamma(n+k+\frac{j}{D})^{s+1}}{\Gamma (2n+k +1+\frac{j}{D})^{s+1}}.\end{aligned}$$ Denoting by $k_0 (n) $ the integer part of $x_0 n$ and applying the Stirling formula to the factorial and gamma factors we obtain, as $n\to\infty$, $$\begin{aligned} c_{k_0(n),j}^{1/n} &\sim \Big(\frac{n}{e}\Big) ^{s+1-3D} \, \bigg(\frac{3Dn+Dk_0(n)+j }{e}\bigg) ^{ 3D+Dx_0} \, \bigg(\frac{e}{Dk_0(n)+j-1 }\bigg) ^{ Dx_0} \nonumber \\ & \qquad \times \bigg(\frac{n+k_0(n)+\frac{j}{D} -1 }{e}\bigg) ^{ (s+1)( x_0+1)} \, \bigg(\frac{e}{2n+k_0(n) +\frac{j}{D}}\bigg) ^{ (s+1)( x_0+2)} \nonumber \\ &\sim \frac{((x_0+3)D)^{ (x_0+3)D}}{(x_0D)^{x_0D}}\, \frac{(x_0+1)^{(s+1)(x_0+1)} }{(x_0+2)^{(s+1)( x_0+2)} } \nonumber \\ & = g(x_0) f(x_0)^{x_0} = g(x_0). \label{eqasykz}\end{aligned}$$ We shall now give details that the asymptotic behavior of $r_{n,j}$ as $n\to\infty$ is determined by the terms $c_{k,j}$ with $k$ close to $x_0n $. Given $D$ and $s$, we take ${\varepsilon}>0$ sufficiently small to accommodate the condition $$b({\varepsilon}) = \max\Big( f(x_0+{\varepsilon}), \frac1{f(x_0-{\varepsilon})}\Big)<1.$$ Then there exists $A({\varepsilon}) > x_1$, where $x_1$ is the unique positive root of $f'(x)=0$, such that $f(A({\varepsilon}))=b({\varepsilon})$. We have $f(x) \geq \frac1{b({\varepsilon})}$ for any $x\in (0, x_0-{\varepsilon}]$ and $f(x) \leq b({\varepsilon}) $ for any $x\in [ x_0+{\varepsilon}, A({\varepsilon})]$. For any $k$ such that $(x_0+2{\varepsilon})n \leq k\leq (A({\varepsilon})-{\varepsilon})n$, Eq.  implies that $c_{k,j} \leq b({\varepsilon}) c_{k-1,j}$ provided $n$ is large (in terms of $D$, $s$ and ${\varepsilon}$), so that taking $k_1 = \lfloor (x_0+2{\varepsilon})n \rfloor$ and $k_2 = \lfloor (x_0+3{\varepsilon})n \rfloor$ we obtain $$\label{eqsommeun} \sum_{k_2\leq k\leq (A({\varepsilon})-{\varepsilon})n} c_{k,j} \leq c_{k_1,j}\sum_{k=k_2}^{+\infty} b({\varepsilon})^{k-k_1}\leq c_{k_1,j}\frac{b({\varepsilon})^{k_2-k_1}}{1-b({\varepsilon})} \leq {\varepsilon}\, c_{k_1,j}$$ for all $n$ sufficiently large. In the same way, we get the estimate $$\label{eqsommede} \sum_{1\leq k\leq \lfloor (x_0-3{\varepsilon})n \rfloor } c_{k,j} \leq {\varepsilon}c_{ \lfloor (x_0-2{\varepsilon})n \rfloor , j}$$ for all $n$ large (in terms of $D$, $s$ and ${\varepsilon}$). At last, choosing ${\varepsilon}$ small we can assume that $A({\varepsilon})$ is sufficiently large (in terms of $D$ and $s$), so that for $k \geq (A({\varepsilon})-{\varepsilon})n$ we have $$c_{k,j} \leq (2D)^{3Dn} \bigg( \frac{n!}{k^{n+1}}\bigg)^{ s+1 - 3D}$$ for $n$ large. Using hypothesis and the Stirling formula, the latter estimate implies $$\begin{aligned} \sum_{k = \lceil (A({\varepsilon})-{\varepsilon})n\rceil} ^{+\infty} c_{k,j} &\leq (3D)^{3Dn} \frac{n!^{ s+1 - 3D}}{((A({\varepsilon})-{\varepsilon})n)^{(s+1-3D)(n+1)-1}} \nonumber\\ &\leq \bigg( \frac{3D}{e (A({\varepsilon})-{\varepsilon})}\bigg)^{sn/2} \leq \Big( \frac12 g(x_0)\Big)^n \label{eqsommetr}\end{aligned}$$ provided $n$ is sufficiently large. Combining Eqs. , , and we obtain $$\label{eqenca} (1-3{\varepsilon}) r_{n,j} \leq \sum_{ (x_0-3{\varepsilon})n \leq k \leq (x_0+3{\varepsilon})n } c_{k,j} \leq r_{n,j}.$$ Now for any $k$ in the range $ (x_0-3{\varepsilon})n \leq k \leq (x_0+3{\varepsilon})n $ it follows from the proof of Eq.  that $$g(x_0)-h({\varepsilon}) \leq c_{k,j}^{1/n} \leq g(x_0)+h({\varepsilon})$$ for $n$ large (in terms of $D$, $s$ and ${\varepsilon}$), where $h$ is a positive function of ${\varepsilon}$ such that $\lim_{{\varepsilon}\to0^+} h({\varepsilon}) = 0$. This implies $$(g(x_0)-2h({\varepsilon}))^n \leq 5{\varepsilon}n (g(x_0)-h({\varepsilon}))^n \leq r_{n,j} \leq \frac{7{\varepsilon}n}{1-3{\varepsilon}} (g(x_0)+h({\varepsilon}))^n \leq (g(x_0)+2h({\varepsilon}))^n$$ for $n$ sufficiently large, and finishes the proof of $\lim_{n\to\infty} r_{n,j}^{1/n} = g(x_0)$ for any $j$. To establish $$\lim_{n\to\infty} \frac{r_{n,j'}}{r_{n,j}} = 1$$ for any $j,j'\in\{1,\ldots,D\}$, we can assume that $1\leq j \leq D-1$ and $j'=j+1$. For any $k $ we have $$\frac{c_{k,j+1}}{c_{k,j}} = \frac{k+3D+\frac{j+1}{D}}{k+ \frac{j}{D}} \bigg( \frac{ \Gamma( n+k+\frac{j+1}{D})}{\Gamma( n+k +\frac{j }{D})} \, \frac{ \Gamma( 2n+k+1+\frac{j }{D})}{\Gamma( 2n+k+1+\frac{j+1}{D})}\bigg)^{s+1}.$$ It follows from the Stirling formula that $\Gamma(x+\frac1{D}) \sim x^{1/D}\Gamma(x)$ as $x\to\infty$, so that for $k = \lfloor x_0 n\rfloor $ we have, as $n\to\infty$, $$\frac{c_{k,j+1}}{c_{k,j}} \sim \frac{x_0+3}{x_0}\bigg( \frac{(x_0+1)^{1/D}}{(x_0+2)^{1/D}}\bigg)^{s+1} = f(x_0)^{1/D} = 1.$$ More generally, for $k$ in the range $ (x_0-3{\varepsilon})n \leq k \leq (x_0+3{\varepsilon})n $ and $n$ sufficiently large we have $$1-\tilde h({\varepsilon}) \leq \frac{c_{k,j+1}}{c_{k,j}} \leq 1-\tilde h({\varepsilon})$$ with $\lim_{{\varepsilon}\to0^+} \tilde h({\varepsilon}) = 0$. Using Eq.  this concludes the proof of , except for the upper bound on $g(x_0)$ that we shall verify now. To estimate $g(x_0)$ from above, we first show that $x_0<a$, where $a = 4\cdot 2^{-\frac{s+1}{D}}$. Observe that $a < \frac12$, since $\frac{s}{D}\geq 3$. For any $x>0$ we have $$\frac{x+1}{x+2} \leq \frac{1}{2} \Big( 1 + \frac{x}{2}\Big)$$ implying $$f(a) \leq \Big( \frac78\Big) ^D \Big( 1 + \frac{a}{2}\Big)^{s+1}.$$ As $\frac sD$ is large and $\log(\frac78) < -\frac{1}{10}$, we deduce that $$f(a)^{1/(s+1)} \leq \Big( \frac78\Big) ^{\frac{D}{s+1}} \Big( 1 + \frac{a}{2}\Big) < \bigg( 1 - \frac{1}{10} \frac{D}{s+1}\bigg) \big( 1 + 2^{1-\frac{s+1}{D}}\big) < 1,$$ so that we indeed have $x_0<a<\frac12$. Now this upper bound for $x_0$ implies $$\log g(x_0) \leq 3D\log D + 3D \log \Big(\frac72\Big) + (s+1) \big( \log(a+1) - 2 \log (a+2)\big).$$ By taking $\frac{s}{D\log D}$ sufficiently large, we may ensure that the first two terms are sufficiently small in comparison with $s$ and that $a$ is sufficiently close to 0, so that $\log g(x_0) < - (s+1) \log 3$. This completes our proof of Lemma \[lemasy\]. \[remexample\] For $s=77$ and $D=4$ one computes $g(x_0) < \exp(-78)$. Thus, the suitable linear combinations $$\hat{r}_{n,1}=r_{n,4}, \quad \hat{r}_{n,2}=r_{n,2}+r_{n,4} \quad\mbox{and}\quad \hat{r}_{n,4}=r_{n,1}+r_{n,2}+r_{n,3}+r_{n,4}$$ of the corresponding linear forms allow us to eliminate three of the odd zeta values on the list $$\{\zeta(3),\zeta(5),\ldots,\zeta(77)\}.$$ In particular, we obtain that two out of $\{\zeta(5),\zeta(7),\ldots,\zeta(77)\}$ are irrational. This result is slightly weaker than the result of Rivoal and the third author [@RZnote], but it drops out as a byproduct of the construction above. The arithmetic gain given by $\Phi_n(4)$ for $\Phi_n(D)$ defined in Remark \[remarith\] can be used to slightly reduce the bound of $77$ to $73$, still weaker than the one in [@RZnote]. A non-vanishing determinant {#sec5} =========================== The following lemma is used to eliminate irrational zeta values in §\[sec6\] below. \[lemdet\] For $t\geq 1$, let $x_1<\ldots<x_t$ be positive real numbers and $\alpha_1<\ldots<\alpha_t$ non-negative integers. Then the generalized Vandermonde matrix $[x_j^{\alpha_i}]_{1\leq i,j \leq t}$ has positive determinant. We remark that, subject to the hypothesis that $x_1,\ldots,x_t$ are real and positive, Lemma \[lemdet\] is a stronger version of [@LMN Lemme 1] and, therefore, has potential applications to the zero estimates for linear forms in two logarithms. The above result is quite classical and known to many people. While writing this paper we have found various proofs of rather different nature, three given below. We leave it to the readers to choose their favorite proof. As pointed out in [@Krattdet §2.1], the generalized Vandermonde determinant in question is closely related to Schur polynomials. Let $\Delta := \det [x_j^{\alpha_{i}}]_{1\leq i,j \leq t}$, and $$V = \det [x_j^{i-1}]_{1\leq i,j \leq t} = \prod_{1\leq i < j \leq t} (x_j-x_i) >0$$ be the Vandermonde determinant of $x_1,\ldots, x_t$. For any $i\in\{1,\ldots,t\}$, we take $\lambda_i = \alpha_{t+1-i}+i - t $, so that $\lambda_1 \geq \ldots \geq \lambda_t \geq 0$; then $\lambda = (\lambda_1,\ldots,\lambda_t)$ is a partition of the integer $\lambda_1+\ldots+\lambda_t$. The associated Schur polynomial $$s_\lambda = s_\lambda (x_1,\ldots, x_t)= \frac{\Delta}{V}$$ possesses the following expression: $$s_\lambda = \sum_\mu K_{\lambda,\mu} m_\mu$$ with the sum over partitions $\mu$ (see, for instance, [@FH Appendix A] or [@Macdonald], and [@Proctor] for a direct proof). Here $m_\mu$ denotes the monomial symmetric polynomial $\sum_\sigma x_1^{\mu_{\sigma(1)}}\ldots x_t^{\mu_{\sigma(t)}}$, where the sum is over the distinct permutations of $\mu$, while the coefficients $K_{\lambda,\mu} $ are non-negative integers and $K_{\lambda,\lambda}=1$. From this we deduce that $ s_\lambda$ is a positive real number, thus $\Delta=s_\lambda V>0$. Write $A_{J,K}$ for the minor of an $n\times m$-matrix $A$, where $n\le m$, determined by ordered index sets $J$ and $K$. A classical result due to Fekete [@fekete] asserts that if all $(n-1)$-minors $$A_{(1,2,\ldots,n-1),K},\quad K=(k_1,\ldots,k_{n-1}) \quad\text{with}\; 1\leq k_1<\ldots<k_{n-1}\leq m$$ are positive, and all minors of size $n$ with consecutive columns are positive, then all $n$-minors of $A$ are positive. Thus, Lemma \[lemdet\] follows by induction on $t$ from Fekete’s result applied to the matrix $[x_j^k]_{1\leq j \leq t,\, 0\leq k < m}$, using the positivity of the Vandermonde determinant. By induction on $t$ one proves the following claim: [*A non-zero function $$f(x)=\sum_{i=1}^t c_i x^{\alpha_i},$$ with $c_i,\alpha_i\in{\mathbb{R}}$, has at most $t-1$ positive zeros.*]{} Indeed, if $f$ has $t$ positive zeros then Rolle’s theorem provides $t-1$ positive zeros of the derivative $\frac{{\mathrm{d}}}{{\mathrm{d}}x} (x^{-\alpha_1} f(x))$. The non-vanishing of the determinant in Lemma \[lemdet\] is an immediate consequence of this claim. Since the determinant depends continuously on the parameters $\alpha_i$, we deduce the required positivity from the positivity of the Vandermonde determinant. Elimination of odd zeta values {#sec6} ============================== Let $0 < {\varepsilon}< \frac13$, and let $s$ be odd and sufficiently large with respect to ${\varepsilon}$. We take $D$ to be the product of all primes less than or equal to $(1-2{\varepsilon}) \log s$ (such a product has asymptotically the largest possible number of divisors with respect to its size, see [@HW Chapter XVIII, §1]). We have $$\log D \, = \sum_{\substack{p \; \mbox{\scriptsize prime}\\ p\leq (1-2{\varepsilon}) \log s}} \log p \, \leq ( 1-{\varepsilon}) \log s$$ by the prime number theorem, that is, $D\leq s^{1-{\varepsilon}}$. Then $D \log D \leq s^{1-{\varepsilon}}\log s $: the assumption of Lemma \[lemasy\] holds. Notice that $D$ has precisely $\delta = 2^{\pi( (1-2{\varepsilon}) \log s)}$ divisors, with $$\log \delta = \pi( (1-2{\varepsilon}) \log s)\, \log 2 \geq (1-3{\varepsilon}) (\log 2) \,\frac{\log s}{\log \log s}.$$ Assume that the number of irrational odd zeta values between $\zeta(3)$ and $\zeta(s)$ is less than $\delta$. Let $3 = i_1 < i_2 < \ldots < i_{\delta-1}\leq s$ be odd integers such that if $\zeta(i)\not\in{\mathbb{Q}}$ and $i$ is odd, $3\leq i \leq s$, then $i=i_j$ for some $j$. We set $i_0=1$, and consider the set ${{\mathcal D}}$ of all divisors of $D$, so that ${\operatorname{Card}}{{\mathcal D}}= \delta$. Lemma \[lemdet\] implies that the matrix $[d^{i_j}]_{d\in{{\mathcal D}}, 0\leq j \leq \delta-1}$ is invertible. Therefore, there exist integers $w_d \in {\mathbb{Z}}$, where $d\in{{\mathcal D}}$, such that $$\label{eqwun} \sum_{d\in{{\mathcal D}}}w_d \, d^{i_j} = 0 \quad\mbox{for any}\; j\in\{1,\ldots,\delta-1\}$$ and $$\label{eqwde} \sum_{d\in{{\mathcal D}}}w_d \, d^{i_0} = \sum_{d\in{{\mathcal D}}}w_d \, d \neq 0.$$ With the help of Lemma \[lemconstru\] we construct the linear forms $$r_{n,j} = \rho_{0,j}+\sum_{\substack{3\leq i \leq s\\i \;\mbox{\scriptsize odd}}} \rho_i \, \zeta\Big(i, \frac{j}{D}\Big)$$ for $n \geq 1$ and $1\leq j \leq D$. The crucial point (as in [@Sprang §3]) is that for any $d\in{{\mathcal D}}$ and any $i\geq 2$, $$\sum_{j=1}^d \zeta\bigg(i, \frac{j \frac{D}{d}}{D}\bigg) = \sum_{j=1}^d \zeta\Big(i, \frac{j}{d}\Big) = \sum_{n=0}^\infty \sum_{j=1}^d \frac{d^i}{(dn+j)^i} = d^i \zeta(i)$$ implying that $$\widehat r_{n,d} = \sum_{j=1}^d r_{n, j\frac{D}{d}} = \sum_{j=1}^d \rho_{0,j\frac{D}{d}} +\sum_{\substack{3\leq i \leq s\\i \;\mbox{\scriptsize odd}}} \rho_i \, d^i \, \zeta(i),$$ are linear forms in the odd zeta values with asymptotic behavior $$\widehat r_{n,d} = (d+o(1))r_{n,1} \quad\text{as}\; n\to\infty, \quad\mbox{where}\; \lim_{n\to\infty} r_{n,1}^{1/n} = g(x_0) <3^{-(s+1)},$$ by Lemma \[lemasy\]. We shall use now the integers $w_d$ to eliminate the odd zeta values $\zeta(i_j)$ for $j=1,\dots,\delta-1$, including all irrational ones, as in [@Zudilintrick] and [@Sprang]. For that, consider $$\widetilde r_n = \sum_{d\in{{\mathcal D}}} w_d \, \widehat r_{n,d}.$$ Eqs.  imply that $$\widetilde r_n = \sum_{d\in{{\mathcal D}}} w_d \sum_{j=1}^d \rho_{0,j\frac{D}{d}} +\sum_{i\in I } \rho_i \bigg( \sum_{d\in{{\mathcal D}}} w_d \, d^i \bigg) \zeta(i),$$ where $I = \{3,5,7,\ldots,s\}\setminus \{i_1,\ldots,i_{\delta-1}\}$; in particular, no irrational zeta value $\zeta(i)$, where $3\leq i\leq s$, appears in this linear combination. Using Eq.  we obtain $$\widetilde r_n = \bigg( \sum_{d\in{{\mathcal D}}}w_d \, d +o(1) \bigg) r_{n,1} \quad \mbox{ with } \quad \sum_{d\in{{\mathcal D}}}w_d \, d \neq 0,$$ so that $$\lim_{n\to\infty} | \widetilde r_n | ^{1/n} = g(x_0) <3^{-(s+1)}.$$ Now all $\zeta(i)$, $i\in I$, are assumed to be rational. Denoting by $A$ their common denominator, we deduce from Lemma \[lemarith\] that $A d_{n+1}^{s+1} \widetilde r_n $ is an integer. From the prime number theorem we have $\lim_{n\to\infty} d_{n+1}^{1/n} = e $, hence the sequence of integers satisfies $$0 < \lim_{n\to\infty} | A d_{n+1}^{s+1} \widetilde r_n | ^{1/n} = e^{s+1} g(x_0) < \Big(\frac{e}{3}\Big)^{s+1} < 1.$$ This contradiction concludes the proof of Theorem \[th1\]. Acknowledgements {#acknowledgements .unnumbered} ---------------- We thank Michel Waldschmidt for his advice, Ole Warnaar for his comments on an earlier draft of the paper, and Javier Fresán for educating us about the state of the art in Grothendieck’s period conjecture and its consequences. [10]{} – [[“]{}Irrationalité de $\zeta(2)$ et $\zeta(3)$[”]{}]{}, in *Journ[é]{}es Arithm[é]{}tiques (Luminy, 1978)*, Ast[é]{}risque, no. 61, 1979, p. 11–13. – [[“]{} Irrationalit[é]{} d’une infinit[é]{} de valeurs de la fonction zêta aux entiers impairs[”]{}]{}, *Invent. Math.* **146** (2001), no. 1, p. 193–207. – *Asymptotic methods in analysis*, Dover Publications, 1981. – [[“]{}Arithm[é]{}tique de la fonction zêta[”]{}]{}, in *Journ[é]{}es math[é]{}matiques X-UPS 2002*, [é]{}ditions de l’[é]{}cole [P]{}olytechnique, 2003, p. 37–164. – [[“]{}Über ein [P]{}roblem von [L]{}aguerre[”]{}]{}, *Rendiconti del Circolo Matematico di Palermo* **34** (1912), p. 89–120. – [[“]{}Shidlovsky’s multiplicity estimate and irrationality of zeta values[”]{}]{}, preprint arXiv 1609.09770 \[math.NT\], J. Austral. Math. Soc., to appear. [— ]{}, [[“]{}Irrationalit[é]{} de valeurs de zêta (d’après [A]{}p[é]{}ry, [R]{}ivoal, ...)[”]{}]{}, in *S[é]{}m. Bourbaki 2002/03*, Ast[é]{}risque, no. 294, 2004, exp. no. 910, p. 27–62. – [[“]{}A refinement of [N]{}esterenko’s linear independence criterion with applications to zeta values[”]{}]{}, *Math. Ann.* **347** (2010), p. 739–763. – *Representation theory: a first course*, Graduate Texts in Math., no. 129, Springer-Verlag, 1991. – *Oscillation matrices and kernels and small vibrations of mechanical systems*, Graduate Texts in Math., AMS Chelsea Publishing, Providence, RI, 2002. – *An introduction to the theory of numbers*, fifth ed., Oxford Science Publications, 1979. – [[“]{}Advanced determinant calculus[”]{}]{}, *Sém. Lotharingien Combin.* **42** (1999), Article B42q, 67 pp. – [[“]{} Hypergeometry inspired by irrationality questions[”]{}]{}, preprint arXiv:1802.08856 \[math.NT\], 2018. – [[“]{}Formes linéaires en deux logarithmes et déterminants d’interpolation[”]{}]{}, *J. Number Th.* **55** (1995), p. 285–321. – *Symmetric functions and [H]{}all polynomials*, Oxford Univ. Press, 1979. – [[“]{}Special values of [H]{}urwitz zeta functions and [D]{}irichlet ${L}$-functions[”]{}]{}, Ph.[D]{}. thesis, Univ. of Georgia, Athens, U.S.A., 2004. – [[“]{}On the linear independence of numbers[”]{}]{}, *Vestnik Moskov. Univ. Ser. I Mat. Mekh. \[Moscow Univ. Math. Bull.\]* **40** (1985), no. 1, p. 46–49 \[69–74\]. – [[“]{}On the linear independence of the special values of a [D]{}irichlet series with periodic coefficients[”]{}]{}, preprint arXiv:1102.3247 \[math.NT\], 2011. – [[“]{}Equivalence of the combinatorial and the classical definitions of [S]{}chur functions[”]{}]{}, *J. Combinatorial Th., Series A* **51** (1989), p. 135–137. – [[“]{}La fonction zêta de [R]{}iemann prend une infinit[é]{} de valeurs irrationnelles aux entiers impairs[”]{}]{}, *C. R. Acad. Sci. Paris, Ser. I* **331** (2000), no. 4, p. 267–270. [— ]{}, [[“]{}Irrationalit[é]{} d’au moins un des neuf nombres $\zeta(5)$, $\zeta(7)$, …, $\zeta(21)$[”]{}]{}, *Acta Arith.* **103** (2002), no. 2, p. 157–167. – [[“]{}Diophantine properties of numbers related to [C]{}atalan’s constant[”]{}]{}, *Math. Annalen* **326** (2003), no. 4, p. 705–721. [— ]{}, [[“]{}A note on odd zeta values[”]{}]{}, preprint arXiv:1803.03160 \[math.NT\], 2018. – [[“]{}Infinitely many odd zeta values are irrational. [B]{}y elementary means[”]{}]{}, preprint arXiv:1802.09410 \[math.NT\], 2018. – [[“]{}One of the numbers $\zeta(5)$, $\zeta(7)$, $\zeta(9)$, $\zeta(11)$ is irrational[”]{}]{}, *Uspekhi Mat. Nauk \[Russian Math. Surveys\]* **56** (2001), no. 4, p. 149–150 \[774–776\]. [— ]{}, [[“]{}Irrationality of values of the [R]{}iemann zeta function[”]{}]{}, *Izvestiya Ross. Akad. Nauk Ser. Mat. \[Izv. Math.\]* **66** (2002), no. 3, p. 49–102 \[489–542\]. [— ]{}, [[“]{}One of the odd zeta values from $\zeta(5)$ to $\zeta(25)$ is irrational. [B]{}y elementary means[”]{}]{}, *SIGMA* **14** (2018), no. 028, 8 pages. Stéphane Fischler, Laboratoire de Mathématiques d’Orsay, Univ. Paris-Sud, CNRS, Université Paris-Saclay, 91405 Orsay, France Johannes Sprang, Fakultät für Mathematik, Universität Regensburg, 93053 Regensburg, Germany Wadim Zudilin, Department of Mathematics, IMAPP, Radboud University, PO Box 9010, 6500 GL Nijmegen, Netherlands;\ School of Mathematical and Physical Sciences, The University of Newcastle, Callaghan, NSW 2308, Australia

ArXiv

--- abstract: 'While considering a class of generalized negative binomial states, we verify that the basic minimum properties for these states to be considered as coherent states are satisfied. We particularize them for the case of the Hamiltonian of the isotonic oscillator and we determine the corresponding Husimi’s Q-function. This function may be used to determine a lower bound for the thermodynamical potential of the Hamiltonian by applying a Berezin-Lieb inequality.' address: ' Faculty of Sciences & Technics (M’Ghila) BP.523, Béni Mellal, Morocco' author: - '[**Zouhaïr MOUAYN** ]{}' --- Introduction ============ The negative binomial states (NBS) are the field states that are superposition of the number states with appropriately chosen coefficients [@1]. Precisely, these labeling coefficients are such that the associated photon-counting distribution is a negative binomial probability distribution [@2]. As matter of fact, these coefficients turn out to be an orthonormal basis of a weighted Bergman space of analytic functions on the complex unit disk satisfying a certain growth condition. Furthermore, the NBS are considered as intermediate states between pure coherent states and pure thermal states [@3] and reduce to Susskind-Glogower phases states for a particular limitof the parameter [@4]. Now, as in [@5; @6], we replace the labeling coefficients by an orthonormal basis of a Hilbert space that generalize the weighted Bergman space of analytic functions on the unit disk we have mentioned to consider a class of generalized negative binomial states (GNBSs) in this sense. Here, we precisely verify that the basic minimum properties for the constructed states to be considered as coherent states are satisfied. Namely, the conditions which have been formulated by Klauder [@7]: $\left( a\right) $ the continuity of labeling, $\left( b\right) $ the fact that these states are normalizable but not orthogonal and $\left( c\right) $ these states fulfilled the resolution of the identity with a positive weight function. Next, we particularize the GNBSs formalism for the case of the isotonic oscillator (IO) [@8] whose importance consists in the fact that it admits exact analytic solutions an being in a certain sense an intermediate potential between the three dimensional harmonic oscillator potential and other anharmonic potentials such as Pöschl-Teller or Morse potentials [@9]. Next, we determine the Husimi’s $Q$-function [@10] which turns out to be the expectation value in the GNBSs representation of the heat semigroup operator associated with the IO. The obtained $Q$-function can be considered as a lower symbol for this heat semigroup operator. Finally, a lower bound for the thermodynamical potential of the IO may be obtained by applying a Berezin-Lieb inequality [@11; @12]. The paper is organized as follows. In Section 2, we review briefly the coherent states formalism we will be using. Section 3 deals with some needed facts on the generalized weighted Bergman spaces on the disk. In Section 4, we attach to each of these spaces a set of coherent states generalizing the negative binomial states and verify that they satisfy the basic minimum properties of coherent states. In Section 5, we recall briefly some needed spectral properties of the isotonic oscillator Hamiltonian . In Section 6, we obtain the Husimi’s Q-function associated with the IO in the coherent state representation and we deduce a lower bound for the thermodynamical potential of the IO. Coherent states and Berezin-Lieb inequalities ============================================= Here, we review a coherent states formalism starting from a measure space ”[it as a set of data]{}” as presented in [@13]. Let $X=\left\{ x\mid x\in X\right\} $ be a set equipped with a measure $d\mu $ and $L^{2}(X,d\mu )$ the space of $d\mu$-square integrable functions on $X$. Let $\mathcal{A}^{2}\subset L^{2}(X,d\mu )$ be a subspace of infinite dimension with an orthonormal basis $\left\{ \Phi _{j}\right\} _{j=0}^{\infty }$. Let $\mathcal{H}$ be another (functional) space with $\dim \mathcal{H}=\infty $ and $\left\{ \phi _{j}\right\} _{j=1}^{\infty }$ is a given orthonormal basis of $\mathcal{H}$. Then consider the family of states $\left\{ \mid x>\right\} _{x\in X}$ in $\mathcal{H}$, through the following linear superpositions: $$\mid x>:=\left( \mathcal{N}\left( x\right) \right) ^{-\frac{1}{2}}\sum_{j=0}^{+\infty }\Phi _{j}\left( x\right) \mid \phi _{j}>\quad \label{2.1}$$ where $$\mathcal{N}\left( x\right) =\sum_{j=0}^{+\infty }\Phi _{j}\left( x\right) \overline{\Phi _{j}\left( x\right) }. \label{2.2}$$ These coherent states obey the normalization condition $$\left\langle x\mid x\right\rangle _{\mathcal{H}}=1 \label{2.3}$$ and the following resolution of the identity of $\mathcal{H}$$$\mathbf{1}_{\mathcal{H}}=\int _{X}\mid x><x\mid \mathcal{N}\left( x\right) d\mu \left( x\right) \label{2.4}$$ which is expressed in terms of Dirac’s bra-ket notation $\mid x><x\mid $ meaning the rank-one -operator $\varphi \mapsto \left\langle \varphi \mid x\right\rangle _{\mathcal{H}}.\mid x>$.  The choice of the Hilbert space $\mathcal{H}$ define in fact a quantization of the space $X$  by the coherent states in , via the inclusion map $X\ni x\mapsto \mid x>\in \mathcal{H}$ and the property is crucial in setting the bridge between the classical and the quantum mechanics. Now, given a set of coherent states $\left\{ \mid x>\right\} $, the concept of upper and lower symbols of an operator $A$ was separately introduced by Berezin [@11] and Lieb [@12] by $$A=\int _{X}d\mu \left( x\right) \widehat{A}\mid x><x\mid \label{2.5}$$ to define the upper symbol $\widehat{A}$ of $A,$ and the expectation value $$\widetilde{A}\left( x\right) :=\left\langle x\mid A\mid x\right\rangle \label{2.6}$$ for the definition of the lower symbol $\widetilde{A}$ of $A.$ Note that given an operator $A$ its upper symbol is not unique in general. It can be proved [@12] that given any convex function $\phi ,$ the following inequalities $$\int _{X}\phi \left( \widetilde{A}\right) d\mu \left( x\right) \leq Tr\left( \phi \left( A\right) \right) \leq \int _{X}\phi \left( \widehat{A}\right) d\mu \left( x\right) \label{2.7}$$ hold and are called Berezin-Lieb inequalities. Generalized Bergman spaces on $\mathbb{D}$ ========================================== Let $\mathbb{D}=\left\{ z\in \mathbb{C},\left| z\right| <1\right\} $ be unit disk endowed with its usual Khäler metric $ds^{2}=-\partial \overline{\partial }Log\left( 1-z\overline{z}\right) dz\otimes d\overline{z}.$ The Bergman distance on $\mathbb{D}$ is given by $$\cosh ^{2}d\left( z,w\right) =\frac{(1-z\overline{w})(1-\overline{z}w)}{\left( 1-z\overline{z}\right) \left( 1-w\overline{w}\right) } \label{3.1}$$ and the volume element reads $$d\mu \left( z\right) =\frac{1}{\left( 1-z\overline{z}\right) ^{2}}d\nu \left( z\right) \label{3.2}$$ with the Lebesgue measure $d\nu \left( z\right) .$ Let us consider the $1-$form on $\mathbb{D}$ defined by $\theta =-i\left( \partial -\overline{\partial }\right) Log\left( 1-z\overline{z}\right) $ to which the Schrödinger operator $$H_{\sigma }:=\left( d+i\frac{\sigma }{2}ext\left( \theta \right) \right) ^{\ast }\left( d+i\frac{\sigma }{2}ext\left( \theta \right) \right) \label{3.3}$$ can be associated. Here $\sigma \geq 0$ is a fixed number, $d$ denotes the usual exterior derivative on differential forms on $\mathbb{D}$ and $ext\left( \theta \right) $ is the exterior multiplication by $\theta $ while the symbol $\ast $ stands for the adjoint operator with respect to the Hermitian scalar product induced by the Bergman metric $ds^{2}$ on differential forms. Actually, the operator $H_{\sigma }$ is acting on the Hilbert space $L^{2}\left( \mathbb{D},d\mu \left( z\right) \right) $ and can be unitarly intertwined as $$\left( 1-z\overline{z}\right) ^{\frac{1}{2}\sigma }\Delta _{\sigma }\left( 1-z\overline{z}\right) ^{-\frac{1}{2}\sigma }=H_{\sigma } \label{3.4}$$ in terms of the second order differential operator $$\Delta _{\sigma }:=-4\left( 1-z\overline{z}\right) \left( \left( 1-z\overline{z}\right) \frac{\partial ^{2}}{\partial z\partial \overline{z}}-\sigma \overline{z}\frac{\partial }{\partial \overline{z}}\right) . \label{3.5}$$ The latter one is acting on the Hilbert space $$L^{2,\sigma }\left( \mathbb{D}\right) =L^{2}\left( \mathbb{D},\left( 1-z\overline{z}\right) ^{\sigma -2}d\nu \left( z\right) \right) . \label{3.6}$$ The spectral analysis of $\Delta _{\sigma }$ have been studied by many authors, see [@5] and references therein. Note that this operator is an elliptic densely defined operator on $L^{2,\sigma }\left( \mathbb{D}\right) $ and admits a unique self-adjoint realization that we denote also by $\Delta _{\sigma }.$ The part of its spectrum is not empty if and only if $\sigma >1.$ This discrete part consists of eigenvalues occurring with infinite multiplicities and having the expression $$\epsilon _{m}^{\sigma }:=4m\left( \sigma -m-1\right) \label{3.7}$$ for varying $m=0,1,...,\left[ (\sigma -1)/2\right] .$ Here, $[x]$ denotes the greatest integer not exceeding $x.$ Moreover, it is well known that the functions given in terms of Jacobi polynomials [@15; @17] by $$\begin{aligned} \Phi _{k}^{\sigma ,m}\left( z\right) = &\sqrt{\frac{\left( \sigma -2m-1\right) k!\Gamma \left( \sigma -m\right) }{\pi m!\Gamma \left( \sigma -2m+k\right) }} \label{3.8} \\ & \times \left( -1\right) ^{k}\left( 1-z\overline{z}\right) ^{-m}\overline{z}^{m-k}P_{k}^{\left( m-k,\sigma -2m-1\right) }\left( 1-2z\overline{z}\right) \nonumber\end{aligned}$$ constitute an orthonormal basis of the eigenspace $$\mathcal{A}_{m}^{2,\sigma }\left( \mathbb{D}\right) :=\left\{ \phi \in L^{2,\sigma }\left( \mathbb{D}\right) ,\Delta _{\sigma }\phi =\epsilon _{m}^{\sigma }\phi \right\} . \label{3.9}$$ of $\Delta _{\sigma }$ associated with the eigenvalue $\epsilon _{m}^{\sigma }$ in . Finally, the $L^{2}-$eigenspace $\mathcal{A}_{0}^{2,\sigma }\left( \mathbb{D}\right) =\left\{ \phi \in L^{2,\sigma }\left( \mathbb{D}\right) ,\Delta _{\sigma }\phi =0\right\} $ corresponding to $m=0$ and associated to $\epsilon _{0}^{\sigma }=0$ in reduces further to the weighted Bergman space consisting of holomorphic functions $\phi $ : $\mathbb{D\rightarrow C}$ with the growth condition $$\int _{\mathbb{D}}\left| \phi \left( z\right) \right| ^{2}\left( 1-z\overline{z}\right) ^{\sigma -2}d\nu \left( z\right) <+\infty . \label{3.10}$$ This is why the eigenspaces in have been called generalized Bergman spaces on the complex unit disk. \[Rem3.1\] In [@5], we have used the basis in to perform a class of coherent states belonging to the Hilbert space $L^{2}\left( \mathbb{R}_{+}^{\ast },x^{-1}dx\right) $. The associated coherent state transform have been considered as a generalization of the second Bargmann transform ([@18 p.203]). Generalized negative binomial states ==================================== The negative binomial states are labeled by points $z \in \mathbb{D}$ and are of the form $$\mid z\text{ },\sigma ,0>:=\left( 1-z\overline{z}\right) ^{\frac{1}{2}\sigma }\sum\limits_{k=0}^{+\infty }\sqrt{\frac{\Gamma \left( \sigma +k\right) }{ \Gamma \left( \sigma \right) k!}}z^{k}\mid \psi _{k}>, \label{4.1}$$ where $\sigma >1$ is a fixed parameter and $\mid \psi _{k}>$ are Fock states. Their photon probability distribution $$\left| <\psi _{k}\mid z\text{ },\sigma >\right| ^{2}=\left( 1-z\overline{z}\right) ^{\sigma }\left( z\overline{z}\right) ^{k}\frac{\Gamma \left( \sigma +k\right) }{\Gamma \left( \sigma \right) k!} \label{4.2}$$ obeys the negative binomial probability distribution with parameters $\lambda =z\overline{z}$ and $\sigma .$ Also, observe that the coefficients in the superposition : $$\Phi _{k}^{\sigma ,0}\left( z\right) :=\sqrt{\frac{\Gamma \left( \sigma +k\right) }{\pi \Gamma \left( \sigma \right) k!}}z^{k},k=0,1,2,... \label{4.3}$$ constitute an orthonormal basis of the eigenspace $\mathcal{A}_{0}^{2,\sigma }\left( \mathbb{D}\right) $ associated with the first eigenvalue $\epsilon _{0}^{\sigma }=0$ and consisting of analytic functions on $\mathbb{D}$ with the growth condition . For instance, let $\sigma >1$ and $m=0,1,...,\left[ \left( \sigma -1\right) /2\right]$ be fixed parameters and let $\left\{ \mid \psi _{k}>\right\} _{k=0}^{\infty }$ be a set of Fock states in a Hilbert space $\mathcal{H}$. Then a class of generalized negative binomial states (GNBS) can be defined as in [@5; @6] by $$\mid z\text{ },\sigma ,m>=\left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-\frac{1}{2}}\sum\limits_{k=0}^{+\infty }\Phi _{k}^{\sigma ,m}\left( z\right) \mid \psi _{k}> , \label{4.4}$$ where $\mathcal{N}_{\sigma ,m}\left( z\right) $ is a normalization factor and $\left\{ \Phi _{k}^{\sigma ,m}\left( z\right) \right\} _{k=0}^{\infty }$  is the orthonormal basis of the generalized Bergman space in . Now, one of the important task to do is to determine is the overlap relation between two GNBSs. \[Prop4.1\] Let $\sigma >1$ and $m=0,1,..., \left[ \left( \sigma -1\right) /2\right] .$ Then, for every $z,w\in \mathbb{D}$, the overlap relation between two GNBSs is given through the scalar product $$\begin{aligned} &<w,\sigma ,m\mid z\text{ },\sigma ,m>_{\mathcal{H}} =\frac{(\sigma -2m-1)\Gamma \left( \sigma -m\right) \left( \mathcal{N}\left( z\right) \mathcal{N}\left( w\right) \right) ^{-\frac{1}{2}}}{\pi m!\left( -1\right) ^{m}\Gamma \left( \sigma -2m\right) \left( 1-z\overline{w}\right) ^{\sigma }} \label{4.5} \\ &\times \left( \frac{(1-z\overline{w})\left( 1-\overline{w}z\right) }{\left( 1-z\overline{z}\right) \left( 1-w\overline{w}\right) }\right) ^{m}._{2}\digamma _{1}\left( -m,\sigma -m,\sigma -2m;\frac{\left( 1-z\overline{z}\right) \left( 1-w\overline{w}\right) }{(1-z\overline{w})\left( 1-\overline{w}z\right) }\right), \nonumber\end{aligned}$$ where $_{2}\digamma _{1}$ is a terminating Gauss hypergeometric sum. **Proof.** In view of Eq. , the scalar product of two GNBS $\mid z$ $,\sigma ,m>$ and $\mid w$ $,\sigma ,m>$ in $\mathcal{H}$ reads $$\begin{aligned} <w,\sigma ,m\mid z\text{ },\sigma ,m>_{\mathcal{H}}&=\left( \mathcal{N}_{\sigma ,m}\left( z\right) \mathcal{N}_{\sigma ,m}\left( w\right) \right) ^{-\frac{1}{2}}\sum\limits_{k=0}^{+\infty }\Phi _{k}^{\sigma ,m}\left( z\right) \overline{\Phi _{k}^{\sigma ,m}\left( w\right) } \nonumber \\ &=\frac{(\sigma -2m-1)\left( \mathcal{N}\left( z\right) \mathcal{N}\left( w\right) \right) ^{-\frac{1}{2}}}{\pi \left( \left( 1-z\overline{z}\right) \left( 1-w\overline{w}\right) \right) ^{m}}\mathcal{S}_{z,w}^{\sigma ,m}, \label{4.6}\end{aligned}$$ where $$\begin{aligned} \mathcal{S}_{z,w}^{\sigma ,m}&=\frac{\Gamma \left( \sigma -m\right) \left( \overline{z}w\right) ^{m}}{m!\Gamma \left( \sigma -2m\right) }\sum_{k=0}^{+\infty }\frac{k!}{(\sigma -2m)_{k}}\left( \frac{1}{\overline{z}w}\right) ^{k} \label{4.7} \\ &\times P_{k}^{\left( m-k,\sigma -2m-1\right) }\left( 1-2z\overline{z}\right) P_{k}^{\left( m-k,\sigma -2m-1\right) }\left( 1-2w\overline{w}\right) . \nonumber\end{aligned}$$ Making use of the following identity ([@19 p.1329]): $$\begin{aligned} &\sum_{n=0}^{+\infty }\frac{n!t^{n}}{\left( 1+\alpha \right) _{n}} P_{n}^{(\gamma -n,\alpha )}\left( x\right) P_{n}^{(\gamma -n,\alpha )}\left( y\right) =\left( 1-\frac{1}{4}\left( x-1\right) \left( y-1\right) t\right) ^{1+\gamma +\alpha } \label{4.8} \\ & \times \left( 1-t\right) ^{\gamma }._{2}\digamma _{1}\left( 1+\gamma +\alpha ,-\gamma ,1+\alpha ;\frac{-\left( x+1\right) \left( y+1\right) t}{\left( 1-t\right) \left( 4-\left( x-1\right) \left( y-1\right) t\right) }\right) \nonumber\end{aligned}$$ for $n=k,t=1/\overline{z}w,\gamma =m,\alpha =\sigma -2m-1,x=1-2z\overline{z}$ and $y=1-2w\overline{w},$ we obtain, after calculations, the expression $$\begin{aligned} \mathcal{S}_{z,w}^{\sigma ,m}&=\frac{\Gamma \left( \sigma -m\right) \left( -1\right) ^{m}}{m!\Gamma \left( \sigma -2m\right) }\frac{\left( (1-z\overline{w})\left( 1-\overline{w}z\right) \right) ^{m}}{\left( 1-z\overline{w}\right) ^{\sigma }} \label{4.9} \\ &\times _{2}\digamma _{1}\left( -m,\sigma -m,\sigma -2m;\frac{\left( 1-z\overline{z}\right) \left( 1-w\overline{w}\right) }{(1-z\overline{w})\left( 1-\overline{w}z\right) }\right) .\nonumber\end{aligned}$$ Returning back to Eq. and inserting the expression we arrive at the announced formula. [$\blacksquare$]{} \[Cor4.1\] The normalization factor in is given by $$\mathcal{N}_{\sigma ,m}\left( z\right) =\frac{(\sigma -2m-1)}{\pi \left( 1-z\overline{z}\right) ^{\sigma }}, \label{4.10}$$ for every $z\in \mathbb{D}.$ **Proof.** We first make appeal to the relation ([@17 p.212]): $$_{2}\digamma _{1}\left( -n,n+\kappa +\varrho +1,1+\kappa ;\frac{1-\tau }{2}\right) =\frac{n!\Gamma \left( 1+\kappa \right) }{\Gamma \left( 1+\kappa +n\right) }P_{n}^{\left( \kappa ,\varrho \right) }\left( \tau \right) \label{4.11}$$ connecting the $_{2}\digamma _{1}-$sum with the Jacobi polynomial for the parameters $n=m,\kappa =\sigma -2m-1,\varrho =0$ and the variable $$\tau =1-2\frac{\left( 1-z\overline{z}\right) \left( 1-w\overline{w}\right) }{(1-z\overline{w})\left( 1-\overline{w}z\right) } \label{4.12}$$ to rewrite Eq. as $$\begin{aligned} \sqrt{\mathcal{N}_{\sigma ,m}\left( z\right) \mathcal{N}_{\sigma ,m}\left( w\right) } & =\frac{\left( -1\right) ^{m}(\sigma -2m-1)\left( 1-z\overline{w}\right) ^{-\sigma }}{\pi <w;\sigma ,m\mid z\text{ },\sigma ,m>_{\mathcal{H}}} \label{4.13} \\ & \times P_{m}^{\left( \sigma -2m-1,0\right) }\left( 1-2\frac{\left( 1-z\overline{z}\right) \left( 1-w\overline{w}\right) }{(1-z\overline{w})\left( 1-\overline{w}z\right) }\right) . \nonumber\end{aligned}$$ The factor $\mathcal{N}_{\sigma ,m}\left( z\right) $ should be such that $$<z,\sigma ,m\mid z,\sigma ,m>_{\mathcal{H}}=1. \label{4.14}$$ So that we put $z=w$ in and we use the symmetry identity ([@17 p.210]): $$P_{m}^{\left( \gamma ,\varrho \right) }\left( \xi \right) =\left( -1\right) ^{m}P_{m}^{\left( \varrho ,\gamma \right) }\left( -\xi \right) \label{4.15}$$ to obtain the expression $$\mathcal{N}_{\sigma ,m}\left( z\right) =\frac{(\sigma -2m-1)}{\pi \left( 1-z\overline{z}\right) ^{\sigma }}P_{m}^{\left( 0,\sigma -2m-1\right) }\left( 1\right) \label{4.16}$$ Finally, we apply the fact that ([@17 p.1329]): $$P_{n}^{\left( \alpha ,\varrho \right) }\left( 1\right) =\frac{\Gamma \left( n+\alpha +1\right) }{n!\Gamma \left( \alpha +1\right) } \label{4.17}$$ in the case of $\alpha =0,n=m$ and $\varrho =\sigma -2m-1$. This ends the proof.[$\blacksquare$]{} \[Prop4.2\] Let $\sigma >1$ and $m=0,1,...,\left[ \left( \sigma -1\right) /2\right] .$ Then, the GNBS in satisfy the following resolution of the identity $$\mathbf{1}_{\mathcal{H}}=\int _{\mathbb{D}}\mid z,\sigma ,m><z,\sigma ,m\mid d\mu _{\sigma ,m}\left( z\right), \label{4.18}$$ where $1_{\mathcal{H}}$ is the identity operator and $ d\mu _{\sigma ,m}\left( z\right) $ is a measure which can be expressed through a Meijer’s $G$-function as $$d\mu _{\sigma ,m}\left( z\right) :=\pi ^{-1}(\sigma -2m-1)G_{11}^{11}\left( -z\overline{z}\mid \begin{array}{c} -1 \\ 0 \end{array} \right) d\nu \left( z\right) , \label{4.18}$$ and $d\nu \left( z\right) $ being the Lebesgue measure on $\mathbb{D}$. **Proof**. Let us assume that the measure takes the form $$d\mu _{\sigma ,m}\left( z\right) =\mathcal{N}_{\sigma ,m}\left( z\right) \Omega \left( z\right) d\nu \left( z\right) , \label{4.19}$$ where $\Omega \left( z\right) $ is an auxiliary density to be determined. Let $\varphi \in \mathcal{H}$ and let us start by writing the following action $$\begin{aligned} \mathcal{O}\left[ \varphi \right] &:=\left( \int _{\mathbb{D}}\mid z,\sigma ,m><z,\sigma ,m\mid d\mu _{\sigma ,m}\left( z\right) \right) \left[ \varphi \right] \label{4.20} \\ &=\int _{\mathbb{D}}<\varphi \mid z,\sigma ,m><z,\sigma ,m\mid d\mu _{\sigma ,m}\left( z\right) . \label{4.21}\end{aligned}$$ Making use Eq. , we obtain successively $$\begin{aligned} \mathcal{O}\left[ \varphi \right] &=\int _{\mathbb{D}}<\varphi \mid \left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-\frac{1}{2}}\sum_{k=0}^{+\infty }\Phi _{k}^{\sigma ,m}\left( z\right) \mid \psi _{k}>><z,\sigma ,m\mid d\mu _{\sigma ,m}\left( z\right) \label{4.22} \\ &=\int _{\mathbb{D}}\sum_{k=0}^{+\infty }\Phi _{k}^{\sigma ,m}\left( z\right) <\varphi \mid \psi _{k}><z,\sigma ,m\mid \left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-\frac{1}{2}}d\mu _{\sigma ,m}\left( z\right) \label{4.23} \\ &=\left( \sum_{j,k=0}^{+\infty }\int _{\mathbb{D}}\overline{\Phi _{j}^{\sigma ,m}\left( z\right) }\Phi _{k}^{\sigma ,m}\left( z\right) \mid \psi _{k}><\psi _{j}\mid \left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-1}d\mu _{\sigma ,m}\left( z\right) \right) \left[ \varphi \right] . \label{4.24}\end{aligned}$$ We replace the measure $d\mu _{\sigma ,m}\left( z\right) $ by the expression in the right hand side of Eq. , then Eq. can be written without $\varphi $ as follows $$\mathcal{O}=\sum_{j,k=0}^{+\infty }\left[ \int _{\mathbb{D}}\overline{\Phi _{j}^{\sigma ,m}\left( z\right) }\Phi _{k}^{\sigma ,m}\left( z\right) \Omega \left( z\right) d\nu \left( z\right) \right] \mid \psi _{j}><\psi _{k}\mid . \label{4.25}$$ Therefore, we need to have $$\int _{\mathbb{D}}\overline{\Phi _{j}^{\sigma ,m}\left( z\right) }\Phi _{k}^{\sigma ,m}\left( z\right) \Omega \left( z\right) d\nu \left( z\right) =\delta _{jk}. \label{4.26}$$ For this we recall the orthogonality relation of the $\Phi _{k}^{\sigma ,m}\left( z\right) $ in the Hilbert space $L^{2,\sigma }\left( \mathbb{D}\right) $, which reads $$\int _{\mathbb{D}}\overline{\Phi _{j}^{\sigma ,m}\left( z\right) }\Phi _{k}^{\sigma ,m}\left( z\right) \left( 1-z\overline{z}\right) ^{\sigma -2}d\nu \left( z\right) =\delta _{jk}. \label{4.27}$$ This suggests us to set $\Omega \left( z\right) :=\left( 1-z\overline{z}\right) ^{\sigma -2}.$ Therefore, we get that $$d\mu _{\sigma ,m}\left( z\right) =\frac{(\sigma -2m-1)}{\pi \left( 1-z\overline{z}\right) ^{2}}d\nu \left( z\right) . \label{4.28}$$ By making us of the identity [@20]: $$G_{11}^{11}\left( \zeta \mid \begin{array}{c} a \\ b \end{array} \right) =\Gamma \left( 1-a+b\right) \zeta ^{b}\left( 1+\zeta \right) ^{a-b-1} \label{4.29}$$ for $\zeta =-z\overline{z}$, $a=-1$ and $b=0$, we arrive at the expression of the measure in . Therefore, Eq. reduces to $$\mathcal{O}=\sum\limits_{j,k=0}^{+\infty }\delta _{jk}\mid \psi _{j}><\psi _{k}\mid =\mathbf{1}_{\mathcal{H}}. \label{4.30}$$ The proof is finished. [$\blacksquare$]{} \[Prop4.3\] Let $\sigma >1$ *and* $m=0,1,...,\left[ \left( \sigma -1\right) /2\right] .$ Then, the states $\mid z,\sigma ,m>$ satisfy the continuity property with respect to the label $z\in \mathbb{D}.$ That is, the norm of the difference of two states $$d_{\sigma ,m}\left( z,w\right) \mathit{\ :}=\left\| \left( \mid z,\sigma ,m>-\mid w,\sigma ,m>\right) \right\| _{\mathcal{H}} \label{4.31}$$ goes to zero whenever $z\rightarrow w.$ **Proof.** By using the fact that any GNBS is normalized by the factor given in , direct calculations enable us to write the square of the quantity in as $$\mathit{\ }d_{\sigma ,m}^{2}\left( z,w\right) \mathit{\ }=2\left( 1-\Re e <z,\sigma ,m\mid w,\sigma ,m>\right) . \label{4.32}$$ Next, we use of the expression of the scalar product in form which it is clear that the overlap takes the value $1$ as $z\rightarrow w$ and consequently $d_{\sigma ,m}\left( z,w\right) \rightarrow 0$. [$\blacksquare$]{}\ We end this section by the following remarks. \[Rem3.2\] By a general fact on reproducing kernels [@21], the proof of proposition also says that the knowledge of the explicit orthonormal basis in leads directly to expression of the reproducing kernel of the generalized Bergman space in via calculations using the formula due to A. Srivastava and A. B. Rao [@19]. \[Rem3.2\] In [@6], we have used the same basis under another form as labeling coefficients in order to consider the photon-counting probability distribution with the mass function $$p_{k}^{(\sigma ,m,\lambda )}:=\gamma _{\sigma ,m,k}\left( 1-\lambda \right) ^{\sigma -2m}\lambda ^{\left| m-k\right| }\left( P_{\frac{1}{2}\left( m+k-\left| m-k\right| \right) }^{\left( \left| m-k\right| ,\sigma -2m-1\right) }\left( 1-2\lambda \right) \right) ^{2} , \label{4.33}$$ where $k=0,1,2,...,$ $\gamma _{\sigma ,m,k}>0$ is a constant, $m=0,1,...,\left[ \sigma -1)/2\right]$ and $\lambda =z\overline{z}$. We have calculated the associated Mandel parameter [@22] and we have discussed the classicality/nonclassicality of the GNBS with respect to the location of their labeling points $z$ inside the hyperbolic disk $\mathbb{D} $. Similar results, in the Euclidean plane and the Riemann sphere settings have been obtained respectively in [@23] and [@24]. \[Rem3.3\] The fact that we have written the measure $d\mu _{\sigma ,m}\left( z\right) $ in in terms of the Meijer’s G-function could be of help when tackling the ”*photon-added coherent states* (*PACS*)” problem for the GNBS under consideration. The isotonic oscillator $\mathbf{L}_{\protect\alpha }$ ====================================================== Not all quantum Hamiltonians are known to have exact solutions. An important model of a solvable class is the isotonic oscillator [@25] $$\mathbf{L}_{\alpha }:=\frac{1}{2}\left( -\frac{d^{2}}{dx^{2}}+x^{2}+\frac{\left( \alpha ^{2}-1/4\right) }{x^{2}}\right) ,\alpha \geq 1/2 \label{5.1}$$ acting in the Hilbert space $\mathcal{H}:=L^{2}\left( \mathbb{R}_{+},dx\right) $ and the eigenfunctions $\psi \in \mathcal{H}$ satisfy the Dirichlet boundary condition $\psi \left( 0\right) =0.$ This operator appears in the literature under many names such as Gol’dman-Krivchenkov Hamiltonian [@26] or pseudoharmonic oscillator [@8] or Laguerre operator [@27]. It is the generalization the harmonic oscillator in three dimensions where the generalization lies in the parameter $\nu =\alpha ^{2}-1/4$ ranging over $\left(0,+\infty \right) $ instead of the angular momentum quantum numbers $l=0,1,2,...$ . This operator may be factorized as follows $$\mathbf{L}_{\alpha }=D_{\alpha }^{+}D_{\alpha }^{-}+\alpha +1 \label{5.2}$$ in terms of the operators, having the form $$D_{\alpha }^{\pm }=\frac{1}{\sqrt{2}}\left( -\frac{\alpha +1/2}{x}+x\pm \frac{d}{dx}\right) . \label{5.3}$$ It is well known that the Hamiltonian $\mathbf{L}_{\alpha }$ admits exact solutions of the form $$\psi _{k}^{\alpha }\left( x\right) =\left( \frac{2\Gamma \left( k+1\right) }{\Gamma \left( k+\alpha +1\right) }\right) ^{\frac{1}{2}}x^{\alpha +\frac{1}{2}}e^{-\frac{1}{2}x^{2}}L_{k}^{\left( \alpha \right) }\left( x^{2}\right), \label{5.4}$$ where $L_{k}^{\left( \alpha \right) }\left( .\right) $ denotes the Laguerre polynomial [@16] and the corresponding eigenvalues are given by ([@28]): $$\lambda _{k}^{\alpha }=2\alpha +k+1,k=0,1,2,...\text{ \ }. \label{5.5}$$ Note also that the functions $\psi _{k}^{\alpha }$ can be obtained by $k-$fold application of a creation operator to the ground state wavefunction $\psi _{0}^{\alpha }$. The vectors $\left\{ \mid \psi _{k}^{\alpha }>\right\} _{k=0}^{\infty }$ satisfy the orthogonality relation $$\left\langle \psi _{k}^{\alpha }\mid \psi _{k}^{\alpha }\right\rangle _{\mathcal{H}}=\int _{0}^{+\infty }\psi _{k}^{\alpha }\left( x\right) \psi _{j}^{\alpha }\left( x\right) dx=\delta _{kj} \label{5.6}$$ and constitute a complete orthonormal basis for the Hilbert space $\mathcal{H}$. Furthermore, they can be used together with the eigenvalues in to define the heat semigroup associated with $\mathbf{L}_{\alpha }$ as * * $$e^{-t\mathbf{L}_{\alpha }}\left[ f\right] :=\sum\limits_{k=0}^{+\infty }e^{-t\lambda _{k}^{\alpha }}<f\mid \psi _{k}^{\alpha }>_{\mathcal{H}}.\psi _{k}^{\alpha } \label{5.7}$$ for any function $f\in L^{2}\left( \mathbb{R}_{+},dx\right) $ .It is also well known that by using the Hille-Hardy formula ([@17 p.242]), this semigroup has an integral representation, i.e., $$e^{-t\mathbf{L}_{\alpha }}\left[ f\right] \left( x\right) =\int _{0}^{+\infty }W_{t}\left( x,y\right) f\left( y\right) dy, \label{5.9}$$ where $$W_{t}\left( x,y\right) =\frac{2\sqrt{xy}e^{-t}}{(1-e^{-2t})}I_{\alpha }\left( \frac{2xye^{-t}}{1-e^{-2t}}\right) \exp \left( -\frac{1}{2}\left( x^{2}+y^{2}\right) \frac{1+e^{-2t}}{1-e^{-2t}}\right) . \label{5.10}$$ Here $I_{\alpha }\left( .\right) $ denotes the modified Bessel function of the first kind and order $\alpha $ ([@17 p.66]). Husimi’s Q-function attached to $L_{\protect\alpha }$ ===================================================== For $\sigma >1$ and $m=0,1,...,\left[ \left( \sigma -1\right) /2\right] .$ A class of generalized negative binomial states (GNBS) attached to the isotonic oscillator $L_{\alpha }$ can be defined by setting $$\mid z\text{ },\sigma ,m,\alpha >=\left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-\frac{1}{2}}\sum\limits_{k=0}^{+\infty }\Phi _{k}^{\sigma ,m}\left( z\right) \mid \psi _{k}^{\alpha }> \label{6.1}$$ where $\mathcal{N}_{\sigma ,m}\left( z\right) $ is the factor in , $\left\{ \Phi _{j}^{\sigma ,m}\left( z\right) \right\} $ are defined in and $\left\{ \mid \psi _{k}^{\alpha }>\right\} $ are the Fock vectors given in . The diagonal representation of $e^{-t\mathbf{L}_{\alpha }}$ in the GNBSs representation is now precised as follows. \[Def6.1\] The Husimi’s Q-function attached to the operator $\mathbf{L}_{\alpha }$ is given through the mean value $$Q_{m}^{t}\left( \mathbf{L}_{\alpha }\right) \left( z\right) =\mathbb{E}_{\left\{ \mid z\text{ },\sigma ,m,\alpha >\right\} }\left( e^{-t\mathbf{L}_{\alpha }}\right) =<z ,\sigma ,m,\alpha \mid e^{-t\mathbf{L}_{\alpha }}\mid z\text{ },\sigma ,m,\alpha > \label{6.2}$$ with respect to the set of GNBSs defined in . \[Prop6.1\] The mean value defined in has the  following expression $$\begin{aligned} \mathbb{E}_{\left\{ \mid z ,\sigma ,m,\alpha >\right\} } & \left( e^{-t\mathbf{L}_{\alpha }}\right) =\frac{\pi \left( 1-z\overline{z}\right) ^{\sigma }e^{-\left( 2\alpha +1\right) t}}{(\sigma -2m-1)}\left( \frac{\left( z\overline{z}-e^{-t}\right) \left( 1-z\overline{z}e^{-t}\right) }{\left( 1-z\overline{z}\right) ^{2}}\right) ^{m} \nonumber \\ & \times \left( \frac{1-z\overline{z}}{1-z\overline{z}e^{-t}}\right) ^{\sigma }P_{m}^{\left( \sigma -2m,0\right) }\left( 1+\frac{2e^{-t}\left( 1-z\overline{z}\right) ^{2}}{\left( z\overline{z}-e^{-t}\right) \left( 1-z\overline{z}e^{-t}\right) }\right) . \label{6.3}\end{aligned}$$ **Proof**. We start by inserting the expression of the operator $e^{-t\mathbf{L}_{\alpha }}$ into the equation in which we also replace the GNBS by their definition in . We obtain successively $$\begin{aligned} Q_{m}^{t}\left( \mathbf{L}_{\alpha }\right) \left( z\right) &=\sum\limits_{k=0}^{+\infty }\exp \left( -t\lambda _{k}^{\alpha }\right) <z\text{ },\sigma ,m,\alpha \mid \psi _{k}^{\alpha }><\psi _{k}^{\alpha }\mid z\text{ },\sigma ,m,\alpha > \label{6.4} \\ &=\sum\limits_{k=0}^{+\infty }\exp \left( -t\lambda _{k}^{\alpha }\right) \left| <z\text{ },\sigma ,m,\alpha \mid \psi _{k}^{\alpha }>\right| ^{2} \label{6.5} \\ &=\sum\limits_{k=0}^{+\infty }\exp \left( -t\lambda _{k}^{\alpha }\right) \left| \left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-\frac{1}{2}}\Phi _{k}^{\sigma ,m}\left( z\right) \right| ^{2} \label{6.6} \\ &=\left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-1}\sum\limits_{k=0}^{+\infty }\exp \left( -t\left( 2\alpha +k+1\right) \right) \left| \Phi _{k}^{\sigma ,m}\left( z\right) \right| ^{2} \label{6.7} \\ &=\left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-1}e^{-\left( 2\alpha +1\right) t}\sum\limits_{k=0}^{+\infty }\left( e^{-t}\right) ^{k}\Phi _{k}^{\sigma ,m}\left( z\right) \overline{\Phi _{k}^{\sigma ,m}\left( z\right) }. \label{6.8}\end{aligned}$$ Now, to calculate the sum in we make use of the expression of the functions $\Phi _{k}^{\sigma ,m}\left( z\right) $ in involving Jacobi polynomials and we apply the identity $\left( 4.8\right) $. This allows us to obtain the expression $$\begin{aligned} Q_{m}^{t}\left( \mathbf{L}_{\alpha }\right) \left( z\right) &=\left( \mathcal{N}_{\sigma ,m}\left( z\right) \right) ^{-1}e^{-\left( 2\alpha +1\right) t}\left( \frac{\left( z\overline{z}-e^{-t}\right) \left( 1-z\overline{z}e^{-t}\right) }{\left( 1-z\overline{z}\right) ^{2}}\right) ^{m} \label{6.9} \\ &\times \left( \frac{1-z\overline{z}}{1-z\overline{z}e^{-t}}\right) ^{\sigma }P_{m}^{\left( \sigma -2m,0\right) }\left( 1+\frac{2e^{-t}\left( 1-z\overline{z}\right) ^{2}}{\left( z\overline{z}-e^{-t}\right) \left( 1-z\overline{z}e^{-t}\right) }\right) . \nonumber\end{aligned}$$ Finally, we replace the factor $\mathcal{N}_{\sigma ,m}\left( z\right) $ by its expression in . [$\blacksquare$]{}\ In the following we will use the $Q_{m}-$function presented above in order to write an inequality involving the thermodynamical potential associated with the operator $\mathbf{L}_{\alpha }.$ This potential reads $$\Omega _{\alpha }:=\frac{-1}{\beta }Tr\left( Log\left( 1+e^{-\beta \left( \mathbf{L}_{\alpha }-\eta \right) }\right) \right) \label{6.10}$$ where $\eta $ is the chemical potential and $\beta =1/k_{B}T,$ $k_{B}$ is the Boltzman constant and $T$ denotes the temperature. Let us put $\epsilon =e^{\beta \eta }>0$ and state the following inequality. \[Prop6.2\] Let $\sigma >1.$ Then, the thermodynamical potential in satisfy the inequality $$\max_{m\in \mathbb{Z}_{+}\cap \left[ 0,\left( \sigma -1\right) /2\right] }\left[ \frac{1}{\beta }\int _{\mathbb{D}}Log\left( \frac{1}{1+\epsilon Q_{m}^{t}\left( \mathbf{L}_{\alpha }\right) \left( z\right) }\right) d\mu _{\sigma ,m}\left( z\right) \right] \leq \Omega _{\alpha } \label{6.11}$$ for every $\beta >0.$ **Proof**. The form of  the potential $\Omega _{\alpha }$ in suggests us to consider the function $$\phi _{\epsilon }\left( u\right) =-Log\left( 1+\epsilon u\right) . \label{6.12}$$ So that we can rewrite as $$t\Omega _{\alpha }=Tr\left( \phi _{\epsilon }\left( e^{-t\mathbf{L}_{\alpha }}\right) \right) , \label{6.13}$$ where $\beta =t\in \mathbb{R}_{+}$. We now apply the Berezin-Lieb inequality for the lower symbol $Q_{m}$ of the operator $e^{-t\mathbf{L}_{\alpha }}$ in the GNBSs representation to obtain the following inequality $$\int _{\mathbb{D}}\left[ \phi _{\epsilon }\circ Q_{m}^{t}\left( \mathbf{L}_{\alpha }\right) \right] \left( z\right) d\mu _{\sigma ,m}\left( z\right) \leq Tr\left( \phi _{\epsilon }\left( e^{-t\mathbf{L}_{\alpha }}\right) \right) . \label{6.14}$$ Making use of and replacing the right hand side of by $t\Omega _{\alpha }$ as in , we get an inequality that holds for every $m=0,1,...,\left[ \left( \sigma -1\right) /2\right] .$ Therefore, we consider the maximum with respect to the integer $m$ of the quantity in the left hand side of in order to be close as possible to the value of $\Omega _{\alpha }$. [$\blacksquare$]{} [99]{} S. M. Barnett,* J. Mod. Opt. A*, p.2201 (1998) W. Feller, An introduction to probability: theory and its applications, Vol.1 2nd ed., John Wiley, 1957 Ts Gantsog, Amitabh Joshi and R. Tanas, Quantum opt. **6** (1994) pp.517-526 H-C Fu and R. Sasaki, Negative binomial states of quantized Radiation fields, Preprint YIPT-96-54, arXiv: quant-ph/ 9610024v1. F. ELWassouli, A. Ghanmi, A. Intissar and Z. Mouayn, Generalized second Bargmann transforms associated with the hyperbolic Landau levels on the Poincaré disk, *Ann. Henri Poincaré*, **13** pp.513-524 (2012) N. Askour & Z. Mouayn, Probability distributions attached to generalized Bergman spaces on the Poincaré disk, arXiv:*1003.4323v1*math-ph 23 Mar 2010 J. R. Klauder, Continuous Representation theory I. Postulates of continuous representation theory, *J. Math. Phys.* 4 1055-1058 D. Popov, Barut-Girardello coherent states of the pseudoharmonic oscillator. *J. Phys. A: Math. Gen*** 34** 5283-5296 (2001) D. Popov, Gazeau-Klauder quasi-coherent states for the Morse oscillator, Phys. Lett. A., 316 (6) pp.369-381 (2003) K. Husimi, *Proc. Phys. Soc. Japan* 22 264 (1940) F. A. Berezin, Covariant and contravariant symbols of operators, *Izv. Akad. SSSR Ser. Mat.* **6** 1134 (1972) E. H. Lieb, The classical limit of quantum spin systems Comm. Math. Phys.31 615 (1973) J. P. Gazeau, Coherent states in quantum physics, WILEY-VCH Verlag GmbH & Co. KGaA Weinheim 2009 Dodonov V V, ’Noncalssical’ states in quantum optics: a ’squeezed review of the first 75 years, *J.Opt.B: Quantum Semiclass.opt.* **4**, R1-R33 (2002) Mourad E.H.Ismail, Classical and Quantum Orthogonal Polynomials in one variable, Encyclopedia of Mathematics and its applications, Cambridge university press (2005)  Gradshteyn I S and Ryzhik I M, ”Table of Integrals, Series and Products”, Academic Press, INC, Seven Edition 2007 W.Magnus, F.Oberhettinger & R.P.Soni, Formulas and Theorems for the Special Functions of Mathematical Physics, Springer-Verlag Berlin Heidelberg New York, 1966. V. Bargman, On a Hilbert space of analytic functions and an associated integral transform, Part I. *Comm. Pure. Appl. Math.*, **14** 187-214 (1961) A. Sirvastava and A. B. Rao, A polynomial of the form $\digamma _{4},$ *Indian Jour. Pure and App. Math*, **6** (1), pp. 1326-1339 (1975) A. M. Mathai and R. K. Saxena, Generalized hypergeometric functions with applications in statistics and physical sciences, Lect. Notes. Math. Vol 348, Springer-Verlag, Berlin, 1973 Aronszajn N, Theory of reproducing kernels, *Trans. Am. Math. Soc*.** 68,** pp.337-404 (1950) L. Mandel, Sub-Poissonian photon stattistics in resonance fluorescence, *Opt. Lett.*, **4**, 205-207 (1979) Z. Mouayn and A. Touhami, Probability distributions attached to generalized Bargmann spaces in the complex plane, *Infinite dimensional Analysis, Quantum Probability & Related fields*, **13,** No.2, 2010 A. Ghanmi, A. Hafoud and Z. Mouayn, Generalized binomial distributions attached to Landau levels on the Riemann sphere, *Adv. Math. Phys*., vol 2011, article ID 39417. K. Thirulogasantar and N. Saad, *J. Phys. A: Math. Gen.* **37** (2004), 4567-4577 I. I. Gol’dman and D. V.Krivchenkov, Problems in Quantum Mechanics, Pergamon, London, 1961 J. Bentacor, Tranference of $L^{p}-$boundedness between harmonic analysis operators for Laguerre and Hermite settings,*Revista de la Union Mathematica, Argentina,* Vol. **50**, No.2, pp.39-46 (2009) R. L. Hall, N. Saad and A.B. Von Keviczky, *J. Phys. A: Math. Gen.* **34** (2001), 11287-11300

ArXiv

--- abstract: 'We quantify the total, quantum, and classical correlations with entropic measures, and quantitatively compare these correlations in a quantum system, as exemplified by a Heisenberg dimer which is subjected to the change of environmental parameters: temperature and nonuniform external field. Our results show that the quantum correlation may exceed the classical correlation at some nonzero temperatures, though the former is rather fragile than the later under thermal fluctuation. The effect of the external field to the classical correlation is quite different from the quantum correlation.' author: - 'Wen-Ling Chan$^1$' - 'Jun-Peng Cao$^{1,2}$' - 'Dong Yang$^{1,3}$' - 'Shi-Jian Gu$^1$' title: 'Effects of environmental parameters to total, quantum and classical correlations' --- Introduction ============ Correlation effect plays an important role in physical phenomena. Many interesting properties of the quantum systems are attributed to the existence of the entanglement [@EPR], which is intrinsically related to the superposition principle of quantum mechanics and the direct product structure of the Hilbert space [@Nielsen1; @AGalindo02]. Entanglement is a kind of pure quantum correlation which does not exist in any classical systems, and regarded as a significant resource in quantum information processing, such as quantum teleportation, dense coding, and quantum cryptography [@Bennett]. Due to the central role of the entanglement in quantum information, various issues on the entanglement have been studied intensively in recent fifteen years. Among these issue, the effects of environmental parameters (such as thermal fluctuation and external field) to the entanglement in quantum spin systems have been attracted much attention [@XWang01; @PZanardi; @XQXi02; @CAnteneodo03; @YSun03]. Moreover, some interesting properties of the entanglement which is beyond the traditional physical intuition were found. For examples, the thermal fluctuation can enhance the entanglement in some special cases, the external field does not always suppress the entanglement [@YSun03]. Therefore, these studies shed new light on our understanding of the entanglement. However, besides the quantum entanglement, a quantum system possesses the classical correlation [@HV; @BGroisman05]. A simple example is the spin singlet state $({|\uparrow\downarrow\rangle}-{|\downarrow\uparrow\rangle})/\sqrt{2}$, besides the entanglement with value 1, the state also has the classical correlation of the value 1 (see section II). Another example is the mixed state with the density matrix as $\rho=(|\uparrow\uparrow\rangle\langle\uparrow\uparrow| +|\downarrow\downarrow\rangle\langle\downarrow\downarrow|)/2$. In this state, the quantum correlation between two spins is zero, while the classical correlation is 1. Therefore, some interesting questions arise. For examples, What is the difference between the classical correlation and quantum correlation in a realistic system? Is the classical correlation always larger than the quantum one? Why we live in a classical world rather than a quantum world? etc.. Answering these questions from the point of view of different correlations are our main motivations in this work. Our paper is organized as follows. In section \[sec:definition\], we define the measurements of total, quantum and classical correlations. In section \[sec:enviroment\], we use a Heisenberg dimer which interacts with the thermal environment as an example to study the effects of the temperature, external fields, and anisotropic interaction on the the correlations. Finally, a summary is given in section \[sec:sum\]. Definitions and measures of bipartite correlations {#sec:definition} ================================================== Total correlation ----------------- In quantum information theory [@Nielsen1], for two subsystems 1 and 2, the mutual information is defined as $$S(1:2)=S(1)+S(2)-S(1\cup 2), \label{eq:S(1:2)}$$ where $S(i)=-{\rm tr}(\rho_i\,{\rm log}_2\,\rho_i),i=1,2,1\cup 2$ is the entropy of the corresponding reduced density matrix. Since the entropy is used to quantify the physical resource (in unit of classical bit due to $\log_2$ in its expression) needed to store information of a system, the mutual entropy then measures additional physical resource required if we store two subsystems respectively rather than store them together. Let us look at a very simple example: a two-qubit system in a singlet state $({|\uparrow\downarrow\rangle}-{|\downarrow\uparrow\rangle})/\sqrt{2}$. We have $S(1)=S(2)=1$ and $S(1\cup 2)=0$, hence $S(1:2)=2$. Obviously, there is no information in a given singlet state. However, each spin in this state is completely uncertain. So we need two bits to store them respectively. Here the mutual information is twice the entanglement, as measured by the von Neumann entropy of either subsystem. This is due to the reason that besides quantum correlation, the state has also classical correlation between the two subsystems. Therefore, the mutual information can be used to measure the total correlation between two subsystems. We will call the quanity $S(1:2)$ as the “total correlation entropy” or simply “total correlation” hereafter. Quantum correlation ------------------- The quantum correlation only exists in the quantum world, and usually is called entanglement. For bipartite state, there are a few measures to quantify the entanglement of a general mixed state [@PV]. Among the measures, the entanglement of formation [@CHBennett96] is well known and a analytic formula for two-qubit system is found [@SHill97]. Consider a density matrix $\rho$ of two subsystems 1 and 2. There are infinite pure-state ensembles $\{\psi_i,p_i\}$ of $\rho$, where $p_i$ is the probability of $\psi_i$, such that $$\rho=\sum_i p_i {|\psi_i\rangle}{\langle \psi_i|}. \label{decomposition}$$ For each pure state ${|\psi_i\rangle}$, the entanglement $E$ is measured by the von Neumann entropy [@AWehrl94]. Then the entanglement of formation $E_f$ of the density matrix $\rho$ is the average entanglement of the pure states of the decomposition, minimized over all the possible ensembles: $$E_f(\rho)={\rm min}\sum_i p_i E(\psi_i). \label{eq:Ef}$$ (Note that if the system is in a pure state, $E_f$ is just $E$.) For a mixed state, it is usually difficult to evaluate $E_f$. However, for a two-qubit system, it can be readily obtained from the concurrence of the system. Given the density matrix $\rho$ of the pair qubits, the concurrence is given by [@SHill97] $$C={\rm max}\{\lambda_1-\lambda_2-\lambda_3-\lambda_4,0\}, \label{eq:con}$$ where $\lambda_i$ are the square roots of the eigenvalues of the operator $$\varrho=\rho(\sigma_1^y\otimes\sigma_2^y)\rho^*(\sigma_1^y\otimes\sigma_2^y), \label{eq:varrho}$$ with $\lambda_1\geq\lambda_2\geq\lambda_3\geq\lambda_4$, $\sigma_i^y$ are the normal Pauli operators, and $\rho^*$ is the complex conjugate operator of $\rho$. The entanglement of formation can then be evaluated as [@SHill97] $$\begin{aligned} && E_f=h\left({1+\sqrt{1-C^2}}\over2\right); \nonumber \\ && h(x)=-x\,{\rm log}_2 \, x-(1-x)\,{\rm log}_2 \, (1-x). \label{eq:Eof}\end{aligned}$$ $E_f$ is monotonically increasing and ranges from 0 to 1 as $C$ goes from 0 to 1. $E_f=C=0$ if the system is unentangled and $E_f=C=1$ if it is maximally entangled. In fact, one can take the concurrence itself as a measurement of entanglement. Since the mutual information has the unit of bit, for comparison purpose, we will take the entanglement of formation instead of concurrence to be our measurement standard in this paper. We will call the quantity $E_f$ as the “quantum correlation entropy” or simply “quantum correlation” hereafter. Classical correlation --------------------- The classical correlation of a bipartite system is defined in different scenarios [@HV; @BGroisman05]. The measure defined in [@HV] reflects the effect of one party’s measurement on the other party’s state. The measure defined in [@BGroisman05] attempts to explain the total correlation coming from quantum part and classical part based on the distance concept of relative entropy. Both these two measure coincides in the case of pure states. Consider a pure state of bipartite system, ${|\psi\rangle}=\sum_i\alpha_i{|u_i\rangle}\otimes{|v_i\rangle}$ unpon Schmidt decomposition. The quantum correlation actually defines the amount of immediate effect on one subsystem during the performing measurement on another subsystem. For pure state, it is just the entropy of one subsystem. After the measurement, the density matrix becomes diagonal in the basis of the Schmidt decomposition. Then the classical correlation between these two subsystems corresponds to the maximum amount of change of uncertainty in one subsystem after knowing some new information of another subsystem through a classical channel. Such a correlation equals to the entropy of one subsystems too. For the mixed state, the total correlation cannot neatly divided into the quantum part and the classical part. These two parts are much more “entangled" with each other. In this paper, we follow the lines of [@BGroisman05]. That is, roughly speaking, the total correlation comes from the quantum part and the classical one. Intuitively, the quantum correlation is more flimsy than the classical one, and the classical correlation should be larger than the quantum correlation in the mixed state. An obvious instance is that for the separable state, there is no entanglement while classical correlation exists, in which case all the total correlation comes from the classical part. Is it possible that the quantum correlation is larger than the classical part? For this purpose, we would like to adopt the quantum entanglement measure as large as possible. It is proved that all the reasonable entanglement measures is not larger than the entanglement of formation [@3H]. So, we take the entanglement of formation as the quantum correlation and the classical correlation is defined as the total correlation minus the quantum part. Before we discuss the main result, we argue that the total correlation minus the entanglement of formation is the non-erasable correlation under the constraint that entanglement is preserved. Recall that the entanglement of formation is originally proposed to describe the process of preparation of an entangled state under local operation and classical communication (LOCC). From Eq. (\[eq:Ef\]), we can see that the entanglement of formation corresponds to a specified decomposition of the density matrix. In experiment, to have a given decomposition, one of experimentalists, say Alice, prepares the states $\{ |\psi_i\rangle\}$ according the probability distribution $\{p_i\}$. Therefore, the actual state that describe the initial state of the preparation process is $$\bar\rho=\sum_i p_i |i\rangle\langle i |\otimes {|\psi_i\rangle}{\langle \psi_i|}, \label{eq:varrho}$$ where $\{|i\rangle\}$ are the flags, a set of orthogonal basis for Alice to distinguish $|\psi_i\rangle$. After compression, Alice send the subsystem through an ideal quantum channel to the another experimentalist Bob who need to know which one he receives because in general he cannot decompress the state without destroying the entanglement of the state. He requires the information to distinguish $|\psi_i\rangle$. Therefore, the flags state are also needed to be sent though this task does not require an ideal quantum channel. A classical channel is enough. In order to obtain the goal state $\rho$ from the prepared state $\bar\rho$, both Alice and Bob are required to erase the flag memory which is used to store the information of the set of orthogonal basis. This procedure decreases the classical correlation but preserves the entanglement of formation. However any more information cannot be erased further or entanglement will be destroyed. Therefore, the remaining part of the correlation represents the non-erasable classical correlation between two subsystems $\rho$ under the preservation of quantum correlation, and is calculated by $$\begin{aligned} S_C=S(1:2)-S_f. \label{eq:Sc}\end{aligned}$$ In short, we argue that the mutual information is taken as the total correlation, the entanglement of formation is taken as the quantum correlation, and difference of them is the classical correlation in the meaning of non-erasable classical correlation. Especially for an arbitrary two-qubit system, the total, quantum and classical correlations can be easily calculated and we can compare the quantum correlation and the classical one quantitatively. Environment’s effects on correlations {#sec:enviroment} ===================================== In this section, we use the Heisenberg dimer as a prototype model to show the interesting behavior of the total, quantum and classical correlations under different environment. The model Hamiltonian reads $$\begin{aligned} H= && J\left[{1-\gamma \over 2}(\sigma_1^x\sigma_2^x+\sigma_1^y\sigma_2^y) +{1+\gamma \over 2}\sigma_1^z\sigma_2^z\right] \nonumber \\ && +B_1\sigma_1^z+B_2\sigma_2^z, \label{eq:H}\end{aligned}$$ where $\sigma_i^\alpha$ ($\alpha=x,y,z$) are the Pauli matrices, $J$ is the strength of Heisenberg interaction, and $B_1,B_2$ are the external magnetic fields. For simplicity, we choose $J$ as the energy unit. The parameter $\gamma$, which ranges from $-1$ to 1, adjusts the anisotropic interactions. Anisotropic Heisenberg model ---------------------------- We first consider the case of $B_1=B_2=0$. The eigenstates and eigenvalues are $$\begin{aligned} {|\psi_0\rangle} &=& {1 \over \sqrt{2}}({|\uparrow \downarrow \rangle}- {|\downarrow \uparrow \rangle}), \; E_0={-3+\gamma \over 2}; \nonumber \\ {|\psi_1\rangle} &=& {1 \over \sqrt{2}}({|\uparrow \downarrow \rangle} +{|\downarrow \uparrow \rangle}), \; E_1={1-3\gamma \over 2}; \nonumber \\ {|\psi_2\rangle} &=& {|\uparrow \uparrow \rangle}, \; E_2={1+\gamma \over 2}; \nonumber \\ {|\psi_3\rangle} &=& {|\downarrow \downarrow \rangle}, \; E_3={1+\gamma \over 2}.\end{aligned}$$ The ground state is ${|\psi_0\rangle}$ for $\gamma \neq 1$. At the thermal equilibrium, the density matrix of the system is $$\rho(T)=\eta \left( \begin{array}{cccc} e^{-(1+\gamma) \over T} & 0 & 0 & 0 \\ 0 & {\rm cosh}{1-\gamma \over T} & -{\rm sinh}{1-\gamma \over T} & 0 \\ 0 & -{\rm sinh}{1-\gamma \over T} & {\rm cosh}{1-\gamma \over T} & 0 \\ 0 & 0 & 0 & e^{-(1+\gamma) \over T} \end{array} \right), \label{eq:rho1}$$ where the Boltzmann’s constant $k_B$ is set as one and $$\eta = {1 \over 2\left[{\rm cosh}{1-\gamma \over T} + e^{-(1+\gamma)/T}\right]}. \nonumber$$ ![\[fig:anisotropic\]The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the anisotrpic Heisenberg dimer with differnt values of parameter $\gamma$. In the Fig. (d), the $E_f=0$ and the curves of $S(1:2)$ and $S_C$ are overlap. (a) $\gamma=-1$, (b) $\gamma=0$, (c) $\gamma=0.9$ and (d) $\gamma=1$. ](an_a "fig:"){width="4cm"} ![\[fig:anisotropic\]The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the anisotrpic Heisenberg dimer with differnt values of parameter $\gamma$. In the Fig. (d), the $E_f=0$ and the curves of $S(1:2)$ and $S_C$ are overlap. (a) $\gamma=-1$, (b) $\gamma=0$, (c) $\gamma=0.9$ and (d) $\gamma=1$. ](an_b "fig:"){width="4cm"} ![\[fig:anisotropic\]The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the anisotrpic Heisenberg dimer with differnt values of parameter $\gamma$. In the Fig. (d), the $E_f=0$ and the curves of $S(1:2)$ and $S_C$ are overlap. (a) $\gamma=-1$, (b) $\gamma=0$, (c) $\gamma=0.9$ and (d) $\gamma=1$. ](an_c "fig:"){width="4cm"} ![\[fig:anisotropic\]The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the anisotrpic Heisenberg dimer with differnt values of parameter $\gamma$. In the Fig. (d), the $E_f=0$ and the curves of $S(1:2)$ and $S_C$ are overlap. (a) $\gamma=-1$, (b) $\gamma=0$, (c) $\gamma=0.9$ and (d) $\gamma=1$. ](an_d "fig:"){width="4cm"} From the density matrix (\[eq:rho1\]), the total, quantum and classical correlations can be calculated directly. The results is shown in Fig. \[fig:anisotropic\]. We see several interesting features from this figure. First, at high temperature, all correlations approach zero. This is because that the occupation probabilities of the unentangled states will be enhanced and the correlations will be diluted. The thermal fluctuation is the leading effects. Second, at certain temperature range, the classical correlation exceeds the quantum correlation. This is obvious in cases of small $\gamma$ (Fig. \[fig:anisotropic\](a) and (b)). Third, when $\gamma$ is close to 1 (Fig. \[fig:anisotropic\](c)), the classical correlation may exhibit a local minimum at low temperature. It is worth noting that the quantum correlation is smaller for a larger $\gamma$. The physical interpretation is that a larger $\gamma$ corresponds to a more classical model, hence less amount of entanglement. When $\gamma$ reaches 1, the quantum correlation vanishes for all temperature (Fig. \[fig:anisotropic\](d)). This is expected because when $\gamma=1$ all eigenstates of the Hamiltonian are unentangled states, hence the resulting density matrix $\rho(T)$ is separable for all temperature. The above features can be illustrated more clearly by defining a threshold temperature $T_{\rm th}$, which is a function of $\gamma$. Above the $T_{\rm th}$, the quantum correlation completely vanishes. The quantum correlation reaches zero when the concurrence reaches zero. From Eqs. (\[eq:con\]),(\[eq:varrho\]) and (\[eq:rho1\]), the concurrence of the system is $$C={\rm max} \left\{ {{\rm sinh}{1-\gamma \over T}-e^{-(1+\gamma)/T} \over {\rm cosh}{1-\gamma \over T}+e^{-(1+\gamma)/T}},0 \right\}. \nonumber$$ The $C=0$ requires ${\rm sinh}{1-\gamma \over T} \leq e^{-(1+\gamma)/T}$. Then the threshold temperature $T_{\rm th}$ should satisfy $$\gamma={T_{\rm th} \over 2}{\rm ln}\left( e^{2/T_{\rm th}}-2 \right).$$ The plot of $T_{\rm th}$ versus $\gamma$ is shown in Fig. \[fig:thre1\]. It is obvious that $T_{\rm th}$ drops when $\gamma$ increases. From Fig.\[fig:thre1\], we can divide the whole plane into two regions. Below the line of $T_{\rm th}$, the system has both quantum and classical correlations, while above the line, the system has only classical correlation. ![\[fig:thre1\]The threshold temperature $T_{\rm th}$ versus anisotropic parameter $\gamma$ in the Heisenberg dimer. The quantum correlation of the system vanishes if $T>T_{\rm th}$.](thre1){width="7cm"} ![\[fig:XY\_1\] The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the $XY$ model under a uniform magnetic fields $B_1=B_2$. Varying the strength of the fields, four typical cases are shown: (a) $B_1=B_2=0.5$, (b) $B_1=B_2=0.95$, (c) $B_1=B_2=1.05$ and (d) $B_1=B_2=1.5$.](XY1a "fig:"){width="4cm"} ![\[fig:XY\_1\] The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the $XY$ model under a uniform magnetic fields $B_1=B_2$. Varying the strength of the fields, four typical cases are shown: (a) $B_1=B_2=0.5$, (b) $B_1=B_2=0.95$, (c) $B_1=B_2=1.05$ and (d) $B_1=B_2=1.5$.](XY1b "fig:"){width="4cm"} ![\[fig:XY\_1\] The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the $XY$ model under a uniform magnetic fields $B_1=B_2$. Varying the strength of the fields, four typical cases are shown: (a) $B_1=B_2=0.5$, (b) $B_1=B_2=0.95$, (c) $B_1=B_2=1.05$ and (d) $B_1=B_2=1.5$.](XY1c "fig:"){width="4cm"} ![\[fig:XY\_1\] The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the $XY$ model under a uniform magnetic fields $B_1=B_2$. Varying the strength of the fields, four typical cases are shown: (a) $B_1=B_2=0.5$, (b) $B_1=B_2=0.95$, (c) $B_1=B_2=1.05$ and (d) $B_1=B_2=1.5$.](XY1d "fig:"){width="4cm"} ![\[fig:XY\_2\]The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the $XY$ model under nonuniform field: (a) $B_1=0.5, B_2=-0.5$ and (b) $B_1=2, B_2=-2$. ](XY2a "fig:"){width="4cm"} ![\[fig:XY\_2\]The total correlation $S(1:2)$, quantum correlation $E_f$ and classical correlation $S_C$ versus temperature $T$, in the $XY$ model under nonuniform field: (a) $B_1=0.5, B_2=-0.5$ and (b) $B_1=2, B_2=-2$. ](XY2b "fig:"){width="4cm"} XY model with nonuniform magnetic field --------------------------------------- ----------------------------- ----------------------------- ----------------------------- {width="4cm"} {width="4cm"} {width="4cm"} {width="4cm"} {width="4cm"} {width="4cm"} {width="4cm"} {width="4cm"} {width="4cm"} \[2mm\] (a) (b) (c) ----------------------------- ----------------------------- ----------------------------- Now we investigate the correlation effects of the external magnetic fields. We only consider the case $\gamma=-1$ and other $\gamma$ can be obtained similarly. Then the eigenstates and eigenvalues of the Hamiltonian are $$\begin{aligned} && {|\Psi_1\rangle} = {|\uparrow \uparrow \rangle}, \; E_1=B_1+B_2; \nonumber \\ && {|\Psi_2\rangle} = {|\downarrow \downarrow \rangle}, \; E_2=-(B_1+B_2); \nonumber \\ && {|\Psi_\pm\rangle} = {1 \over N_\pm}\left[ {(B_1-B_2)\pm \sqrt{D} \over 2} {|\uparrow \downarrow\rangle}+{|\downarrow \uparrow\rangle}\right], \nonumber \\ && E_\pm = \pm \sqrt{D},\end{aligned}$$ where $D=(B_1-B_2)^2+4$ and $N_\pm$ are the normalization factors. The thermal equilibrium state can be described by the density matrix $$\rho(T)={1 \over Z}\left( \begin{array}{cccc} d & 0 & 0 & 0 \\ 0 & b-c & -s & 0 \\ 0 & -s & b+c & 0 \\ 0 & 0 & 0 & d^{-1} \end{array} \right), \label{eq:rho2}$$ where $Z=2\{{\rm cosh}{[(B_1+B_2)/T]} + {\rm cosh}{(\sqrt{D}/ T)}\}$, $b={\rm cosh}(\sqrt{D}/T)$, $c={\rm sinh}{(\sqrt{D} /T)}(B_1-B_2)/\sqrt{D}$, $s={2 }{\rm sinh}{(\sqrt{D} / T)}/\sqrt{D}$ and $d=\exp{[-(B_1+B_2)/T]}$. We first study the correlations under uniform magnetic fields at finite temperatures. The results are shown in Fig. \[fig:XY\_1\] (a-d). Clearly, if $B_1$ is small, the ground state ${|\Psi_-\rangle}$ is a superposition of two antiferromagnetic basis, and is entangled. If the system is subject to a thermal environment, the contribution from the other eigenstates (two of them are separable) will suppress both the quantum and classical correlation. However, if $B_1$ is large enough, the ground state becomes ${|\Psi_2\rangle}$, which is fully polarized and not entangled. The classical correlation, whose value is equal to the quantum one for a pure state, is also zero. The thermal fluctuation, as can see from Fig. \[fig:XY\_1\] (c and d), increases both the quantum and classical correlations at low temperatures. Moreover, an interesting observation is that there exists a range where the quantum correlation exceeds the classical one. In addition, the larger the external field, the smaller the quantum correlation. It is because a large-field setting corresponds to a more classical model. It may be interesting to note that the threshold temperature of the quantum correlation is independent of the field [@YSun03]. All three correlations approach or equal zero at high temperatures. If the directions of two external fields are opposite to each other and the strength are the same, we find that all the total, classical and quantum correlations show comparatively gentle changes against the temperature and fields (Fig. \[fig:XY\_2\] (a) and (b)). The figure is not difficult to interpret, as we argued for the case of $B_1=B_2$. The main difference is that the a larger $B_1$ here may leads to a higher threshold temperature $T_{\rm th}$. In order to see the role of the nonuniform field, we show three correlations against fields at some fixed temperatures in Fig. \[fig:XY\_3\] (The results of quantum correlation, which have already been obtained by Sun [*etal*]{} [@YSun03], are also presented for comparison). At low temperature, we notice that the three correlations are sharply peaked at zero fields. They decay rapidly with the increasing fields if the fields have the same direction, while decay comparatively slowly if the fields have opposite directions. This means that the correlation effects can be adjusted by the uniform fields. At some higher temperature, the peak of the quantum correlation splits into two in the region $B_1B_2<0$ (Fig. \[fig:XY\_3\](b)). Therefore, the nonuniform fields may enhance the quantum correlation, while the uniform fields always destructs it. At very high temperature, the peaks are completely separated as shown in Fig. \[fig:XY\_3\](c). A region with zero quantum correlation appears between the peaks. This implies that the nonuniform field can be used as a switch to turn on and off the quantum correlation [@YSun03]. Meanwhile, unlike the quantum correlation, the classical correlation always decreases with the increasing external magnetic fields. Which means that the external magnetic fields have different effects to the quantum correlation and to the classical correlation. Discussions and summary {#sec:sum} ======================= In this paper, we provide quantification of the total, quantum and classical correlations in a general bipartite system. In order to see their properties in a realistic system, we study them in an anisotropic Heisenberg model at finite temperatures. We find that the quantum correlation always decreases with the increasing temperature, while the classical one may increase in some temperature range. More interestingly, the classical correlation is not always larger than the quantum one, which actually is beyond the general physical intuition [@BGroisman05]. We also investigate the three correlations in the XY model under a nonuniform magnetic field. We find that the fields may enhance the quantum correlation, which is very different from the effect of fields to the classical correlation. In short, our results imply that the environmental parameters (temperature, magnetic fields) demonstrate obviously different effects to the quantum correlation and the classical correlation. This work is partially supported by Direct Grant of CUHK (A/C 2060286), the Earmarked Grant for Research from the Research Grants Council of HKSAR, China (Project CUHK N\_CUHK204/05), and NSFC under Grant No. 10574150. D. Yang acknowledges the financial support from the C. N. Yang Foundation. J. P. Cao and D. Yang are grateful for the hospitality of the Department of Physics at CUHK. A. Einstein, B. Podolsky and N. Rosen, Phys. Rev. [**47**]{}, 777 (1935). M. A. Nilesen and I. L. Chuang, [*Quantum Computation and Quantum Information*]{} (Cambridge University Press, Cambridge, England, 2000) A. Galindo and M. A. Martin-Delgado, Rev. Mod. Phys. [**74**]{}, 347 (2002). C. H. Bennett and S. J. Wiesner, Phys. Rev. Lett., [**68**]{}, 557 (1992); C. H. Bennett, G. Brassard, C. Crepeau, R. Jozsa, A. Peres, and W. Wootters, Phys. Rev. Lett., [**70**]{}, 1895 (1993); A. K. Ekert, J. G. Rarity, P. R. Tapster, and G. M. Palma, Phys. Rev. Lett. [**69**]{}, 1293 (1992). X. Wang, Phys. Rev. A [**64**]{}, 012313 (2001); X. Wang, Phys. Rev. A [**66**]{}, 044305 (2002). P. Zanardi, X. Wang, J. Phys. A: Math. and Gen. [**35**]{} 7947 (2002). X. Q. Xi, S. R. Hao, W. X. Chen, and R. H. Yue, Chin. Phys. Lett. [**19**]{}, 1044 (2002). C. Anteneodo and A. M. C. Souza, J. Opt. B: Quantum Semiclass. Opt. [**5**]{} 73 (2003). Y. Sun, Y. G. Chen and H. Chen, Phys. Rev. A [**68**]{}, 044301 (2003). L. Henderson, V. Vedral, quant-ph/0105028 B. Groisman, S. Popescu and A. Winter, Phys. Rev. A [**72**]{}, 032317 (2005). M. B. Plenio, S. Virmani, Quant. Inf. Comp. [**7**]{}, 1 (2007), quant-ph/0504163 C. H. Bennett, D. P. DiVincenzo, J. Smolin, and W. K. Wooters, Phys. Rev. A [**54**]{}, 3824 (1996). S. Hill and W. K. Wootters, Phys. Rev. Lett. [**78**]{}, 5022 (1997); W. K. Wootters, Phys. Rev. Lett. [**80**]{}, 2245 (1998). A. Wehrl, Rev. Mod. Phys. [**66**]{}, 129 (1994). M. Horodecki, P. Horodecki, R. Horodecki, Phys. Rev. Lett. [**84**]{}, 2014 (2000)

ArXiv

--- abstract: 'We consider an optimal control problem for an abstract nonlinear dissipative evolution equation. The differential constraint is penalized by augmenting the target functional by a nonnegative global-in-time functional which is null-minimized iff the evolution equation is satisfied. Different variational settings are presented, leading to the convergence of the penalization method for gradient flows, noncyclic and semimonotone flows, doubly nonlinear evolutions, and GENERIC systems.' address: - 'IST Austria, Am Campus 1, 3400 Klosterneuburg, Austria.' - 'Faculty of Mathematics, University of Vienna, Oskar-Morgenstern-Platz 1, 1090 Wien, Austria and Istituto di Matematica Applicata e Tecnologie Informatiche *[E. Magenes]{}*, v. Ferrata 1, 27100 Pavia, Italy.' author: - Lorenzo Portinale - Ulisse Stefanelli title: | Penalization via global functionals of\ optimal-control problems for dissipative evolution --- =1 Introduction ============ We are concerned with the abstract optimal control problem $$\label{eq:1} \min\{F(u,y) \ : \ y \in S(u)\}.$$ Here, $u:[0,T]\to H $ stands for a time-dependent admissible control, $H$ is a Hilbert space, and $y:[0,T]\to H$ belongs to the set $S(u)$ of a nonlinear evolution equation with datum $u$ to be specified below. The nonnegative [*target*]{} functional $F$ is defined on the trajectories $u$ and $y$. Relation $y \in S(u)$ corresponds to different models of dissipative evolution. In particular, we will consider the case of $u$-forced $$\begin{aligned} &\text{Gradient flows:}&&y'+\partial \phi(y) = u,\\ &\text{Monotone and pseudomonotone flows:}&&y'+A(y) = u,\\ &\text{Generalized gradient flows:}&&\partial_{y'} \psi(y,y')+\partial \phi(y) = u,\\ &\text{GENERIC flows:}&&y' =L(y) \,DE(y) - K(y) (\partial \phi(y) -u).\end{aligned}$$ The reader is referred to the following sections for all necessary details. In all of these cases, the abstract relation $y \in S(u)$ stands for the variational formulation of a nonlinear partial differential problem of parabolic type, possibly being singular or degenerate. The differential constraint $y\in S(u)$ will be equivalently reformulated as $$y \in S(u) \ \ \Leftrightarrow \ \ G(u,y)=0,$$ where the [*constraining functional*]{} $G$ is a nonnegative functional on entire trajectories. This characterization is not new. In the specific case of a gradient flow $y'+\partial \phi(y) = u$, where $\partial \phi$ stands for the subdifferential of the convex energy $\phi:H \to (-\infty,\infty]$, two possible choices of the constraint functional $G$ are given by the [*Brezis-Ekeland-Nayroles*]{} functional $$G_{\rm BEN}(u,y) = \int_0^T \Big(\phi(y)+\phi^*(u-y') -(u,y) \Big) {{\mathrm{d}}t}+ \frac12\| y(T)\|^2 - \frac12\| y_0\|^2$$ and the [*De Giorgi*]{} functional $$G_{\rm DG}(u,y) = \int_0^T \left(\frac12 \| y'\|^2 +\frac12 \|\partial \phi(y)-u\|^2 - (u,y') \right){{\mathrm{d}}t}+ \phi(y(T)) - \phi(y_0).$$ Here, $(\cdot,\cdot)$ and $\| \cdot \|$ denote the scalar product and the norm in $H$, respectively. The trajectory $y$ is forced to assume the initial value $y(0)=y_0$ by defining $G(u,y)=\infty$ otherwise. The focus of this note is on the penalization of problem by $$\label{eq:2} \min E_{\varepsilon}(u,y) \quad \text{for} \ \ E_{\varepsilon}(u,y):= F(u,y) + \frac{1}{{\varepsilon}}G(u,y).$$ This corresponds to approximate the constrained minimization of problem by means of a family of unconstrained minimizations. This approach is indeed classical and has to be traced back to [Lions]{} [@Lions68], who proposed to penalize the constraint by the residual of the equation. This has already been investigated, both in the stationary and the evolutive case, see [@Bergounioux92; @Bergounioux94; @Bergounioux98; @Gariboldi09; @Gunzburger00; @Mophou11] among many others. We follow this line by penalizing the minimization by the De Giorgi functional $G_{\rm DG}$, which corresponds to the residual by nonetheless exploiting the variational structure of the equation in order to simplify the energy. On the other hand, penalization in coordination with the Brezis-Ekeland-Nayroles functional $G_{\rm BEN}$ is not directly related with residual minimization and, to our knowledge, has not been studied yet. Note that the actual choice of the constraining functional $G$ strongly influences the properties of the problem, so that the considering different options for $G$ is a sensible issue. In the case of the Brezis-Ekeland-Nayroles functional $G_{\rm BEN}$, problem turns out to be a [*separately convex*]{} minimization problem. This allows for the implementation of an alternate minimization procedure, where $E_{\varepsilon}$ is alternatively minimized in the state and the control until convergence. The case of the De Giorgi functional $G_{\rm DG} $ bears its interest in the fact that it is not restricted to convex functionals $\phi$. In fact, $G_{\rm DG} $ is suited for nonconvex potentials as well and it can be easily modified to accommodate additional nonlinear features, such as nonlinear dissipative or conservative terms (see Section \[sec:DG\] below). Our aim is that of checking the solvability of the penalized minimization problem and the convergence of its minimizers to minimizers of the constrained problem as ${\varepsilon}\to 0$. This will be achieved by proving the $\Gamma$-convergence of the penalized functional $E_{\varepsilon}$ to the limit $E_0$ defined by $$E_0(u,y)= F(u,y) \ \ \text{if} \ \ G(u,y)=0 \ \ \text{and} \ \ E_0(u,y)=\infty \ \ \text{otherwise}$$ under different variational settings, corresponding to the above-mentioned different evolution models. The paper is organized as follows. The abstract functional setup is detailed in Section \[sec:abstract\]. Then, the application of the abstract theory to the case of the Brezis-Ekeland-Nayroles variational principle for gradient, noncyclic and semimonotone flows, and doubly nonlinear flows is addressed in Section \[sec:BEN\]. Eventually, Section \[sec:DG\] deals with the applications of De Giorgi principle in the context of gradient, doubly nonlinear, and GENERIC flows. Abstract setup {#sec:abstract} ============== Let us start by specifying some notation. In the following, $H$ stands for a real separable Hilbert space with scalar product $(\cdot,\cdot)$ and norm $\| \cdot \|$. The norm in the general Banach space $E$ will be denoted by $\| \cdot \|_E$. Given the reference time $T>0$, we make use of the standard Bochner spaces $L^p(0,T;E)$, $W^{1,p}(0,T;E)$, $C([0,T];E)$ and so on. A caveat on notation: we will use the same symbol $c$ to indicate positive universal constants, possibly depending on data, and changing from line to line. Given a topological space $(X,\tau)$, we recall that a sequence of functionals ${\mathcal E}_{\varepsilon}: (X,\tau)\to [0,\infty]$ is said to $\Gamma$-converge [@DeGiorgi79] to the limit ${\mathcal E}_0 : (X,\tau)\to [0,\infty]$ if ${\mathcal E}_0(x) \leq \liminf_{{\varepsilon}\to 0} {\mathcal E}_{\varepsilon}(x_{\varepsilon})$ for any $x_{\varepsilon}\to x$ and for all $\hat x\in X$ there exists a sequence $\hat x_{\varepsilon}\to \hat x$ such that ${\mathcal E}_{\varepsilon}(x_{\varepsilon})\to {\mathcal E}_0(\hat x)$. The reader is referred to [Dal Maso]{} [@DalMaso93] for a thorough presentation. We record here the following elementary lemma, which serves as basis for proving convergence of the minimizers of problem throughout. \[lem\] Let $(X,\tau)$ be a sequential topological space and the functionals ${\mathcal F},\, {\mathcal G}: (X,\tau) \to [0,\infty]$ be lower semicontinuous. Assume ${\mathcal E}_{\varepsilon}: = {\mathcal F}+{\varepsilon}^{-1} {\mathcal G}$ to be proper (${\mathcal E}_{\varepsilon}\not \equiv \infty$) and equicoercive for ${\varepsilon}>0$ small enough, namely that there exists ${\varepsilon}_0>0$, $\lambda>0$, and a compact $K \subset X $ such that $ \{x\in X \ : \ {\mathcal E}_{\varepsilon}(x) <\lambda \} \subset K$ for all ${\varepsilon}<{\varepsilon}_0$. Then, - ${\mathcal E}_{\varepsilon}\stackrel{\Gamma}{\to} {\mathcal E}_0 $ where ${\mathcal E}_0(x) := {\mathcal F}(x) $ if $ {\mathcal G}(x)=0$ and ${\mathcal E}_0=\infty$ otherwise; - $\min {\mathcal E}_{\varepsilon}$ can be solved for all ${\varepsilon}<{\varepsilon}_0$. Any sequence $x_{\varepsilon}$ of quasiminimizers, namely $\liminf_{{\varepsilon}\to 0} ({\mathcal E}_{\varepsilon}(x_{\varepsilon}) {-} \inf{\mathcal E}_{\varepsilon})=0$, admits a subsequence converging to a minimizer of ${\mathcal E}_0$; - If $ {\mathcal E}_0$ admits a unique minimizer $x_0$, any sequence of quasiminimizers of ${\mathcal E}_{\varepsilon}$ converges to $x_0$. Ad 1. Let $x_{\varepsilon}\to x$ and assume with no loss of generality that $\sup_{{\varepsilon}} {\mathcal E}_{\varepsilon}(x_{\varepsilon})\leq c <\infty$. In particular, $0\leq {\mathcal G}(x) \leq \liminf_{{\varepsilon}\to 0}{\varepsilon}{\mathcal E}_{\varepsilon}(x_{\varepsilon}) \leq \liminf_{{\varepsilon}\to 0}{\varepsilon}c=0$. Then ${\mathcal E}_0(x)={\mathcal F}(x) \leq \liminf_{{\varepsilon}\to 0}{\mathcal F}(x_{\varepsilon}) \leq \liminf_{{\varepsilon}\to 0}{\mathcal E}_{\varepsilon}(x_{\varepsilon}) $. Fix now any $ \hat x\in X$. As ${\varepsilon}^{-1}{\mathcal G}( \hat x) \to\infty$ if ${\mathcal G}( \hat x)>0$, one has that ${\mathcal E}_{\varepsilon}( \hat x) \to {\mathcal E}_0(\hat x)$. This proves the $\Gamma$-convergence ${\mathcal E}_{\varepsilon}\stackrel{\Gamma}{\to} {\mathcal E}_0$. Ad 2. The existence of a minimizer $x_{\varepsilon}$ of ${\mathcal E}_{\varepsilon}$ for ${\varepsilon}<{\varepsilon}_0$ follows from the equicoercivity and the lower semicontinuity of the sum ${\mathcal F}+{\varepsilon}^{-1} {\mathcal G}$. Any sequence $x_{\varepsilon}$ of quasiminimizers belongs to $K$ for ${\varepsilon}$ small enough. As such, it admits a subsequence (not relabeled) converging to $x_0$ and, for any $x\in X$, we have that ${\mathcal E}_0(x_0)\leq \liminf_{{\varepsilon}\to 0} {\mathcal E}_{\varepsilon}(x_{\varepsilon}) = \liminf_{{\varepsilon}\to 0} \min {\mathcal E}_{\varepsilon}\leq \liminf_{{\varepsilon}\to 0}{\mathcal E}_{\varepsilon}(x)={\mathcal E}_0(x)$. In particular, $x_0$ minimizes ${\mathcal E}_0$. Ad 3. This follows from the uniqueness of the minimizer of $\mathcal E_0$ and from the fact that the topology is assumed to be sequential. Brezis-Ekeland-Nayroles principle {#sec:BEN} ================================= In this section, we investigate penalization by letting the constraining functional to be of Brezis-Ekeland-Nayroles type. Let us start by presenting a result in the case of the classical gradient flow with forcing $u$ $$\label{eq:gf0} y'+\partial \phi(y)\ni u \ \ \text{in $H$, a.e. in} \ (0,T), \ \ y(0)=y_0.$$ As usual, the prime denotes here derivation with respect to time. The potential $\phi:H \to (-\infty,\infty]$ is assumed to be convex, proper, and lower semicontinuous, and we denote by $D(\phi)=\{y\in H \ : \ \phi(y)<\infty\}$ its essential domain. The symbol $\partial \phi$ denotes the corresponding subdifferential in the sense of convex analysis. This is defined as $$\xi \in \partial \phi(y) \ \ \Leftrightarrow \ \ y \in D(\phi) \ \ \text{and} \ \ ( \xi,x-y ) \leq \phi(x) - \phi(y) \ \ \forall x \in H.$$ The initial datum $y_0$ is assumed to belong to $D(\phi)$. Given $u\in L^2(0,T;H)$, the solution $y\in H^1(0,T;H)$ of exists uniquely [@Brezis73]. The celebrated result by Brezis & Ekeland [@Brezis-Ekeland76; @Brezis-Ekeland76b] and Nayroles [@Nayroles76; @Nayroles76b] implies that $y$ solves iff $G_{\rm BEN} (u,y) =0$, where the constraining functional $G_{\rm BEN} (u,y) : L^2(0,T;H)\times H^1(0,T;H)$ is given by $$G_{\rm BEN} (u,y) = \left\{ \begin{array}{ll} {\displaystyle}\int_0^T \Big( \phi(y) {+} \phi^*(u{-}y'){-}( u,y)\Big) \, {{\mathrm{d}}t}+ \frac12 \| y(T)\|^2 - \frac12\|y_0 \|^2 & \ \text{if} \ \ y(0)=y_0\\ \infty&\quad \text{otherwise}. \end{array} \right.\label{eq:BEN}$$ Here, $\phi^*$ denotes the conjugate to $\phi$, namely, $\phi^*(y^*) = \sup_y( (y^*,y) - \phi(y)$. Note that, for all $(u,y) \in L^2(0,T;H)\times H^1(0,T;H)$ the functions $t \mapsto \phi(y'(t))$ and $y \mapsto \phi^*(u(t){-}y'(t))$ are measurable, so that $G_{\rm BEN} (u,y)$ is well defined. Still, $G_{\rm BEN} (u,y)$ takes the value $\infty$ if $t \mapsto \phi(y'(t))$ or $y \mapsto \phi^*(u(t){-}y'(t))$ do not belong to $L^1(0,T)$. Existence results based in the Brezis-Ekeland-Nayroles principle have been obtained by [Rios]{} [@Rios76], [Auchmuty]{} [@Auchmuty93], [Roubíček]{} [@Roubicek00], and [Ghoussoub & Tzou]{} [@Ghoussoub-Tzou04] among others. In [@Ghoussoub-Tzou04], the authors recast the problem within the far-reaching theory of (anti-)selfdual Lagrangians [@Ghoussoub08]. A variety of extensions have been proposed, including perturbations [@Ghoussoub-McCann04], long-time dynamics [@Lemaire96], measure data [@Mabrouk00], time discretizations [@be2], second-order [@Mabrouk03], doubly-nonlinear [@be], monotone [@visintin08], pseudomonotone equations and their structural compactness [@visintin18], and rate-independent flows [@plas]. Note however that deriving existence via these extensions may call for more stringent assumptions on the data of the problem. In the following, we will assume that the set of admissible controls $U$ is a compact subset of $L^2(0,T;H)$. Moreover, we ask the target functional $F : L^2(0,T;H)\times H^1(0,T;H)\to [0,\infty)$ to be lower semicontinuous with respect to the strong $\times$ weak topology of $L^2(0,T;H)\times H^1(0,T;H)$. An example in this class is $$F(u,y) =\frac12 \int_0^T \| y - y_{\rm target}\|^2{{\mathrm{d}}t}+\frac12 \int_0^T \| y' - y_{\rm target}'\|^2{{\mathrm{d}}t}+ \frac12\int_0^T\| u \|^2{{\mathrm{d}}t}$$ for some given $y_{\rm target}\in H^1(0,T;H)$. The main result of this section is the following. \[thm:BEN\] Let $\phi:H\to (-\infty,\infty]$ be convex, proper, and lower semicontinuous, $y_0\in D(\phi)$, $\emptyset \not = U\subset\subset L^2(0,T;H)$, $F: L^2(0,T;H) \times H^1(0,T;H)\to [0,\infty]$ lower semicontinuous and coercive w.r.t. the strong $\times$ weak topology $\tau$ of $L^2(0,T;H)\times H^1(0,T;H)$, $F(u,y) <\infty$ only if $u\in U $, $G_{\rm BEN}$ defined as in , and $E_{\varepsilon}:= F+{\varepsilon}^{-1}G_{\rm BEN}$ for ${\varepsilon}>0$. Then, $\min E_{\varepsilon}$ admits a solution for all ${\varepsilon}>0$. Moreover, $E_{\varepsilon}\stackrel{\Gamma}{\to} E_0$ with respect to topology $\tau$ where $E_0 =F$ on $\{G_{\rm BEN}=0\}$ and $E_0=\infty$ otherwise, and any sequence of quasiminimizers converges, up to a subsequence, to a solution of $\min E_0$. In case $\min E_0$ admits a unique minimizer, any sequence of quasiminimizers $\tau$-converges to it. In order to prove the statement we apply Lemma \[lem\] with the choices $X= L^2(0,T;H)\times H^1(0,T;H)$ and $\tau =$ strong $\times$ weak topology in $X$. We start by checking that $E_{\varepsilon}$ is proper. In fact, by letting $u \in U$ and $y\in H^1(0,T;H)$ be the unique solution of $y'+\partial \phi(y) \ni u$ with $y(0)=y_0$ we have that $E_{\varepsilon}(u,y) = F(u,y)<\infty$. In order to prove the lower semicontinuity of $G_{\rm BEN}$, assume that $(u_n,y_n)\stackrel{\tau}{\to} (u,y)$. As $H^1(0,T;H)\subset C([0,T];H)$ and $U$ is compact in $L^2(0,T;H)$ we have that $$\begin{aligned} &u_n - y_n' \to u- y' \ \ \text{weakly in} \ L^2(0,T;H), \\ &( u_n ,y_n ) \to ( u,y) \ \ \text{in} \ \ L^1(0,T), \\ &y_n(T)\to y(T) \ \ \text{weakly in} \ H.\end{aligned}$$ This implies that $G_{\rm BEN}(u,y) \leq \liminf_{n \to \infty}G_{\rm BEN}(u_n,y_n)$. The equicoercivity of $E_{\varepsilon}$ follows from that of $F$. A remarkable feature of the penalization of problem via the Brezis-Ekeland-Nayroles functional relies in the possibility of exploiting convexity. Indeed, in case $F$ is convex, the penalized $F + {\varepsilon}^{-1}G_{\rm BEN}$ turns out to be separately convex, the only nonconvexity coming from the bilinear term $(u,y)$. This in turn suggests the possibility of implementing some alternate minimization procedure. Note that, in relation with applications to PDEs, the bilinear term $(u,y)$ is usually of lower order. In the statement of Theorem \[thm:BEN\] we have assumed $F$ to be coercive. In fact, the functional $G_{\rm BEN}$ itself cannot be expected to be coercive with respect to topology $\tau$. In particular, this would follow by asking $\phi^*$ to be superquadratic. This would however induce a quadratic bound to $\phi$, a quite restrictive assumption, especially in relation to PDEs. An alternative possibility is that of augmenting $G_{\rm BEN}$ by a coercive term, which would still vanish on solutions of . A proposal in this direction is in [@be], where the following variant of the Brezis-Ekeland-Nayroles functional is presented $$\tilde G_{\rm BEN} (u,y) = G_{\rm BEN} (u,y) + \left( \int_0^T\big( \|y'\|^2{-}(u,y')\big) \, {{\mathrm{d}}t}+ \phi(y(t)) - \phi(y_0) \right)^+\label{eq:BEN2}$$ with $r^+:=\max \{r,0\}$. By letting now $E_{\varepsilon}= F + {\varepsilon}^{-1}\tilde G_{\rm BEN}$ one can prove the statement of Theorem \[thm:BEN\] also for a noncoercive functional $F$, for coercivity for $y$ with respect to the weak topology of $H^1(0, T ;H)$ is provided by $\tilde G_{\rm BEN}$. Before closing this subsection, let us remark that a time-dependent potential $\phi$ can be considered as well, namely $$\label{eq:gf0t} y'(t)+\partial \phi(t,y(t))\ni u(t) \ \ \text{in $H$, for a.e.} \ t \in (0,T), \ \ y(0)=y_0.$$ Here, $\phi:(0,T)\times H \to (-\infty,\infty]$ is asked to be measurable with respect to ${\mathcal L}\otimes {\mathcal B}(H)$, where ${\mathcal L} $ is the Lebesgue $\sigma$-algebra in $(0,T)$ and ${\mathcal B}(H)$ is the Borel $\sigma$-algebra in $H$, and such that $y \mapsto \phi(t,y)$ is proper, convex, and lower semicontinuous for a.e. $t \in (0,T)$. Problem can be equivalently reformulated as $G_{\rm BEN}(u,y)=0$ where $$G_{\rm BEN} (u,y) = \left\{ \begin{array}{ll} {\displaystyle}\int_0^T \Big( \phi(t,y(t)) {+} \phi^*(t,u(t){-}y(t)'){-}( u(t),y(t))\Big) \, {{\mathrm{d}}t}\\ \qquad{} +{\displaystyle}\frac12 \| y(T)\|^2 - \frac12\|y_0 \|^2 & \quad \text{if} \ \ y(0)=y_0\\ \infty&\quad \text{otherwise}. \end{array} \right.$$ where of course conjugation in $\phi^*$ is taken with respect to the second variable only. In order to be sure, however, that pairs $(u,y)$ exist with that $G_{\rm BEN}(u,y) =0$, some additional assumptions on the time dependence $t \mapsto \phi(t,y)$ is required. The reader is referred to [@Kenmochi77; @Kenmochi91; @Moreau77; @Yamada76] for a collection of classical results in this direction. An example ---------- With the aim of illustrating the statement of Theorem \[thm:BEN\], we investigate the ODE optimal control problem $$\begin{aligned} &\min\Bigg\{\frac12 \int_0^1 (y(t) - {{\rm e}}^{-t})^2{{\mathrm{d}}t}+\frac12 \int_0^1t^2 (u(t) - {{\rm e}}^{-t})^2{{\mathrm{d}}t}\ :\\ &\quad \quad \quad y'(t)+y(t)=u(t)\equiv u_0 {{\rm e}}^{-t}, \ u_0\in [0,1], \ y(0)=1\Bigg\}. \label{eq:case}\end{aligned}$$ Here, by taking advantage of the linearity of the constraint one can directly compute $y(t)=S(u_0 {{\rm e}}^{-t})(t) = {{\rm e}}^{-t}(1+tu_0)$ and $$\begin{aligned} &u_0\mapsto F(S(u_0 {{\rm e}}^{-t}), u_0 {{\rm e}}^{-t}) :=\frac12 \int_0^1 (S(u_0 {{\rm e}}^{-t})(t)- {{\rm e}}^{-t})^2{{\mathrm{d}}t}+\frac12 \int_0^1t^2 (u_0 {{\rm e}}^{-t} - {{\rm e}}^{-t})^2{{\mathrm{d}}t}\\ &\quad= \left(\frac12 \int_0^1t^2 {{\rm e}}^{-2t}{{\mathrm{d}}t}\right)\left( u_0^2 + (u_0-1)^2 \right)=:\gamma \left( u_0^2 + (u_0-1)^2 \right)\end{aligned}$$ In particular, the optimal control corresponds to $u_0=1/2$, the optimal solution is $y(t) = {{\rm e}}^{-t}(1+t/2)$, and the minimum of $E_0$ is $$F({{\rm e}}^{-t}(1+t/2),{{\rm e}}^{-t}/2)=\gamma/2 = 1/16 -5/(16{{\rm e}}^2)\sim 0.0202.$$ The ODE is the gradient flow of the potential $\phi(y)=y^2/2$ under the additional forcing $u$. Correspondingly, the Brezis-Ekeland-Nayroles functional $G_{\rm BEN}$ is given by $$G_{\rm BEN}(u,y) = \left\{ \begin{array}{ll} {\displaystyle}\int_0^1\left(\frac12y^2 + \frac12 (u-y')^2 -uy \right){{\mathrm{d}}t}+\frac12 y^2(1) - \frac12& \ \ \text{if} \ y(0)=1,\\ \infty&\ \ \text{otherwise}. \end{array} \right.$$ The penalized optimal control problem reads then $$\begin{aligned} &\min\Bigg\{\int_0^1 \left( \frac12 (y(t) - {{\rm e}}^{-t})^2 +\frac{t^2}{2} (u(t) - {{\rm e}}^{-t})^2 + \frac{1}{2{\varepsilon}}y^2(t) + \frac{1}{2{\varepsilon}} (u(t)-y'(t))^2 -\frac{1}{{\varepsilon}}u(t)y(t)\right){{\mathrm{d}}t}\\ &\quad \quad \quad \quad +\frac{1}{2{\varepsilon}}y^2(1) - \frac{1}{2{\varepsilon}}\ \ : \ \ u(t)\equiv u_0 {{\rm e}}^{-t}, \ u_0 \in [0,1], \ y(0)=1\Bigg\}.\end{aligned}$$ For all given $u$, the Euler-Lagrange equation for $E_{\varepsilon}= F + {\varepsilon}^{-1}G_{\rm BEN}$ in terms of $y_{\varepsilon}$ is $$\begin{aligned} &y''(t)-y(t)-{\varepsilon}y(t) = -(2u_0+{\varepsilon}){{\rm e}}^{-t}, \quad y'(1) + y(1) = u_0/{{\rm e}}. \end{aligned}$$ Complemented with the initial condition $ y(0)=1$, these linear relations uniquely identify a critical point $y_{\varepsilon}$ of $E_{\varepsilon}$. In fact, this is necessarily the unique minimizer of the convex functional $y \mapsto E_{\varepsilon}(u,y)$ and can be explicitly determined in terms of $u_0$ as $$y_{{\varepsilon},u_0}(t) = c_{1{\varepsilon}}{{\rm e}}^{-\alpha_{\varepsilon}t} + c_{2{\varepsilon}}{{\rm e}}^{\alpha_{\varepsilon}t} + \left(\frac{2u_0}{{\varepsilon}} + 1 \right) {{\rm e}}^{-t}$$ where we have used the shorthand notation $$\begin{aligned} \alpha_{\varepsilon}&:=(1+{\varepsilon})^{1/2},\\ c_{1{\varepsilon}}&:=\left(\frac{u_0}{{{\rm e}}} - (1+\alpha_{\varepsilon})\left( \frac{2u_0}{{\varepsilon}}+1\right) \right)\left((1-\alpha_{\varepsilon}){{\rm e}}^{-\alpha_{\varepsilon}} - (1+\alpha_{\varepsilon}){{\rm e}}^{\alpha_{\varepsilon}} \right)^{-1},\\ c_{2{\varepsilon}} &:= - \frac{2u_0}{{\varepsilon}} - c_{1{\varepsilon}}.\end{aligned}$$ The value of $E_{\varepsilon}(u_0{{\rm e}}^{-t} ,y_{{\varepsilon},u_0})$ can be explicitly evaluated. An elementary but tedious computation gives $$\begin{aligned} &E_{\varepsilon}(u_0{{\rm e}}^{-t} ,y_{{\varepsilon},u_0}) = \left(\frac{c_{1{\varepsilon}}^2}{2} +\frac{c_{1{\varepsilon}}^2}{2{\varepsilon}} +\frac{\alpha_{\varepsilon}^2c_{1{\varepsilon}}^2}{2{\varepsilon}}\right) \frac{{{\rm e}}^{-2\alpha_{\varepsilon}}-1}{-2\alpha_{\varepsilon}} + \left(\frac{c_{2{\varepsilon}}^2}{2} +\frac{c_{2{\varepsilon}}^2}{2{\varepsilon}}+\frac{\alpha_{\varepsilon}^2c_{2{\varepsilon}}^2}{2{\varepsilon}}\right) \frac{{{\rm e}}^{2\alpha_{\varepsilon}}-1}{2\alpha_{\varepsilon}} \nonumber\\ & \quad + \left( \frac{2u_0^2}{{\varepsilon}^2}+\frac{1}{2{\varepsilon}} \left(\frac{2u_0}{{\varepsilon}}+1 \right)^2 + \frac{1}{2{\varepsilon}}\left(\frac{2u_0}{{\varepsilon}}+1+u_0 \right)^2-\frac{u_0}{{\varepsilon}}\left(\frac{2u_0}{{\varepsilon}}+1 \right)\right) \frac{{{\rm e}}^{-2}-1}{-2} \nonumber\\ & \quad + \left(\frac{2c_{1{\varepsilon}}u_0}{{\varepsilon}} +\frac{c_{1{\varepsilon}}}{{\varepsilon}}\left(\frac{2u_0}{{\varepsilon}}+1 \right) +\frac{\alpha_{\varepsilon}c_{1{\varepsilon}}}{{\varepsilon}}\left(\frac{2u_0}{{\varepsilon}}+1+u_0 \right) - \frac{c_{1{\varepsilon}}u_0}{{\varepsilon}}\right) \frac{{{\rm e}}^{-\alpha_{\varepsilon}-1}-1}{-\alpha_{\varepsilon}-1} \nonumber\\ & \quad + \left(\frac{2c_{2{\varepsilon}}u_0}{{\varepsilon}} +\frac{c_{2{\varepsilon}}}{{\varepsilon}}\left(\frac{2u_0}{{\varepsilon}}+1 \right) -\frac{\alpha_{\varepsilon}c_{2{\varepsilon}}}{{\varepsilon}}\left(\frac{2u_0}{{\varepsilon}}+1+u_0 \right) - \frac{c_{2{\varepsilon}}u_0}{{\varepsilon}}\right) \frac{{{\rm e}}^{\alpha_{\varepsilon}-1}-1}{\alpha_{\varepsilon}-1} \nonumber\\ & \quad + \left(1+\frac1{\varepsilon}-\frac{\alpha_{\varepsilon}^2}{{\varepsilon}}\right) c_{1{\varepsilon}}c_{2{\varepsilon}} + \frac{1}{2{\varepsilon}}\left(c_{1{\varepsilon}} {{\rm e}}^{-\alpha_{\varepsilon}} + c_{2{\varepsilon}} {{\rm e}}^{\alpha_{\varepsilon}} + \left(\frac{2u_0}{{\varepsilon}}+1\right){{\rm e}}^{-1} \right)^2 - \frac{1}{2{\varepsilon}} \nonumber\\ & \quad + \gamma (u_0-1)^2. \label{ill}\end{aligned}$$ Different curves $u_0 \mapsto E_{\varepsilon}(u_0{{\rm e}}^{-t} ,y_{{\varepsilon},u_0})$ for different choices of ${\varepsilon}$ are depicted in Figure \[illustration\]. We observe that the minimizer and the minimum approach $1/2$ and $0.0202$, respectively, as ${\varepsilon}\to 0$, as expected. Gradient flows in dual space ---------------------------- The statement of Theorem \[thm:BEN\] can be extended to the case of gradient-flow dynamics in dual spaces. Let us introduce a real reflexive Banach space $W$, densely and continuously embedded into $H$, so that $W\subset H \subset W^*$ is a classical Gelfand triplet. We consider the problem $$\label{eq:gfff} y'+\partial \phi(y)\ni u \ \ \text{in $W^*$, a.e. in} \ (0,T), \ \ y(0)=y_0.$$ The potential $\phi:W \to {{\mathbb R}}$ is assumed to be everywhere defined, convex, proper, and lower semicontinuous. The symbol $\partial \phi$ in denotes now the subdifferential between $W$ and $W^*$. This is defined as $$\xi \in \partial \phi(y) \ \ \Leftrightarrow \ \ {\langle}\xi,x-y {\rangle}\leq \phi(x) - \phi(y) \ \ \forall x \in W$$ where ${\langle}\cdot, \cdot {\rangle}$ is the duality pairing between $W^*$ and $W$. We assume $\phi$ to be bounded as follows $$\begin{aligned} &\phi(y)\geq c\|y\|_W^{ m} - \frac1c \quad \forall y \in W, \quad \phi^*(y^*)\geq c\|y^*\|_{W^*}^{ m'} - \frac{1}{c} \quad \forall y^* \in W^*\\ & \| \xi \|_{W^*}^{ m'} \leq c (1+ \| y \|_W^{ m}) \quad \forall y \in W, \, \xi \in \partial \phi(y)\end{aligned}$$ where $m>1$ and $m'=m/(m-1)$. In particular, the above bounds entail a polynomial control on $\phi$ of the form $$\phi(y)\leq c \| y\|_W^{ m} + c \quad \forall y \in W, \quad \phi^* (y^* )\leq c \| y^* \|_{W^* }^{ m'} + c \quad \forall y^* \in W^*$$ which is now compatible with PDE applications. Given the initial datum $y_0\in W$ (recall that $D(\phi) =W$), for all $u\in L^{ m'} (0,T;W^*)$, the solution $y\in W^{1,m'} (0,T;W^*)\cap L^{ m}(0,T,W)$ of exists uniquely. In particular, $y$ solves iff $G_{\rm BEN} (u,y) =0$ where $$G_{\rm BEN} (u,y) = \left\{ \begin{array}{ll} {\displaystyle}\int_0^T \big( \phi(y) + \phi^*(u-y')-{\langle}u,y{\rangle}\Big) \, {{\mathrm{d}}t}+ \frac12 \| y(T)\|^2 - \frac12\|y_0 \|^2 & \quad \text{if} \ \ y(0)=y_0\\ \infty&\quad \text{otherwise}. \end{array} \right.$$ The result of Theorem \[thm:BEN\] can be reformulated in this setting by assuming the set of admissible controls $U$ to be a compact subset of $L^{ m'}(0,T;W^*)$ and $F : L^{ m'}(0,T;W^*)\times W^{1,m'} (0,T;W^*)\cap L^{ m}(0,T,W)\to [0,\infty]$ to be lower semicontinuous with respect to the strong $\times$ weak topology of $L^{ m'}(0,T;W^*)\times W^{1,m'} (0,T;W^*)\cap L^{ m'}(0,T,W)$, with $F(u,y)<\infty$ only if $u \in U$. Note that here no coercivity of $F$ is actually needed, for in this case $G_{\rm BEN} $ itself turns out to be coercive, due to the lower bounds on $\phi$ and $\phi^*$. Once again, $G_{\rm BEN}$ is proper, since it vanishes on solutions to , which are known to exist. In order to check for the lower semicontinuity of $G_{\rm BEN}$ one would need to recall the embedding $ W^{1,m'} (0,T;W^*)\cap L^{ m}(0,T,W)\subset C([0,T];H)$. In particular, the term $\| y(T)\|^2$ turns out to be lower semicontinuous. Nonpotential and nonmonotone flows ---------------------------------- Originally limited to gradient flows of convex functionals, the Brezis-Ekeland-Nayroles variational approach has been extended to classes of nonpotential monotone flows by [ Visintin]{} [@visintin08]. By replacing Fenchel duality by the representation theory by [ Fitzpatrick]{} [@Fitzpatrick], he noticed that solutions of the nonpotential flow $$\label{eq:gf1} y'+Ay\ni u \quad \text{a.e. in} \ W^*, \quad y(0)=y_0,$$ where $A: W \to 2^{W^*}$ is a maximal monotone, coercive, and [*representable*]{} operator and $y_0\in D(A)$, can be characterized by $G_{\rm BEN}(u,y)=0$, where $G_{\rm BEN} : L^2(0,T;W^*)\times H^1(0,T;W^*)\cap L^{ 2}(0,T,W)\to [0,\infty]$ is now given as $$G_{\rm BEN}(u,y)= \left\{ \begin{array}{ll} {\displaystyle}\int_0^T\Big(f_A(y, u{-}y') {-} {\langle}u,y{\rangle}\Big)\, {{\mathrm{d}}t}+\frac12\|y(T)\|^2 - \frac12 \| y_0\|^2& \ \ \text{if} \ y(0)=y_0,\\ \infty &\ \ \text{otherwise.} \end{array} \right.\label{eq:semi}$$ The function $f_A:W\times W^*\to (-\infty,\infty]$ is convex, lower semicontinuous, with $f_A(y,y^*)\geq {\langle}y^*,y{\rangle}$ for all $(y,y^*) \in W\times W^*$, and [ *represents*]{} the operator $A$ in the following sense $$y^* \in Ay \ \ \Leftrightarrow \ \ f_A (y,y^*)={\langle}y^*,y{\rangle}.\label{eq:rep}$$ An operator is said to be [*representable*]{} when it admits a representing function. All maximal monotone operators are representable, for instance via their [*Fitzpatrick function*]{} $$f_A(y,y^*) := {\langle}y^*,y{\rangle}+ \sup\{{\langle}y^* - \tilde y^*, \tilde y- y{\rangle}\ : \ \tilde y \in W, \ \tilde y^* \in A\tilde y\}.$$ A monotone operator need however not be cyclic nor maximal to be representable. The reader is referred to [@visintin17; @visintin18] for a full account on this theory. By taking advantage of position , the assertion of Theorem \[thm:BEN\] can hence be modified to include the case of the differential constraint as well. More generally, the reach of the penalization via the Brezis-Ekeland-Nayroles functional extends even beyond monotone situations. Assume to be given $B:H \times W \to 2^{W^*}$ such that $$\begin{aligned} &B(h,\cdot): W \to 2^{W^*} \ \ \text{is maximal monotone}, \ \forall h \in H,\\[2mm] & \forall (h,y)\in H\times W, \ \forall y^* \in B(h,y), \ \forall h_n \to h \ \text{in} \ H \\ &\text{there exists} \ y^*_n \ \text{such that} \ y^*_n \in B(h_n,y_n) \ \text{and} \ y^*_n \to y^* \ \text{in} \ W^*.\end{aligned}$$ This class of nonmonotone operators $A(y) : = B(y,y)$, called [ *semimonotone*]{} [@visintin18], includes the class of [*pseudomonotone*]{} operators [@browder], and it is representable [@visintin18 Thm. 4.4] in the sense of by means of a weakly lower semicontinuous albeit nonconvex function $f_A$ $$f_A(y,y^*) := {\langle}y^*,y{\rangle}+ \sup\{{\langle}y^* - \tilde y^*, \tilde y- y{\rangle}\ : \ \tilde y \in W, \ \tilde y^* \in B(y,\tilde y)\}.\label{eq:fi}$$ On this basis, the nonmonotone flow $$\label{eq:gf11} y'+A(y)\ni u \quad \text{a.e. in} \ W^*, \quad y(0)=y_0,$$ driven by the semimonotone operator $A(y)$ can be variationally reformulated as $G_{\rm BEN}=0$, where $G_{\rm BEN}$ is defined in from , where however $f_A$ is now defined by . Note that $G_{\rm BEN}$ is proper and lower semicontinuous with respect to the strong $\times$ weak topology of $L^2(0,T;W^*)\times H^1(0,T;W^*)\cap L^{ 2}(0,T,W)$. By letting $E_{\varepsilon}= F+{\varepsilon}^{-1}G_{\rm BEN}$ and assuming again that $F$ is coercive and $F(u,y)<\infty$ only if $u\in U$, the results of Theorem \[thm:BEN\] can be extended to the case of optimal control problems driven by as well. Doubly nonlinear flows ---------------------- A gradient flow can be seen as a particular case of the doubly nonlinear evolution $$\label{eq:gf3} \partial\psi(y')+\partial \phi(y)\ni u \ \ \text{in $V^*$, a.e. in} \ (0,T), \ \ y(0)=y_0.$$ Here, $V$ is a real reflexive Banach space with $W \subset \subset V$, the symbol $\partial$ refers to the subdifferential between $V$ and $V^*$, and $\psi:V \to [0,\infty)$ is a second convex, proper, lower semicontinuous functional defined on the whole $V$. More precisely, we assume $\psi$ to fulfill $0\in \partial \psi(0)$ and to be of polynomial growth, namely $$\begin{aligned} & c\|y'\|^p_V -\frac1c \leq {\langle}w , y' {\rangle}, \quad \| w\|_{V^*}^{p'} \leq c(1 + \| y'\|^p_V) \quad \forall y' \in V, \, w \in \partial \psi(y')\\ & \psi^*(w)\geq c \| w\|^{p'}_{V^*} -\frac1c \quad \forall w \in V^*\end{aligned}$$ for $p >1$ and $p'=p/(p-1)$. Additionally, we assume $D(\phi)=W$ and the coercivity $$\phi(y) \geq c \| y \|^m_W - \frac1c \quad \forall y \in W$$ for some $m>1$. In [@be] a doubly nonlinear version of the Brezis-Ekeland-Nayroles functional is addressed. In particular, one has that $(u,y,w) \in L^{ p'}(0,T;V^*) \times W^{1,p'}(0,T;V^*)\cap L^m(0,T;W)\times L^{ p'}(0,T;V^*)$ solve $$w\in \partial \psi(y'), \quad \partial \phi(y)\ni u -w\ \ \text{a.e. in} \ (0,T), \ \ y(0)=y_0$$ iff $G_{\rm BEN}(u,y,w)=0$, where $G_{\rm BEN}: L^{ p'}(0,T;V^*) \times W^{1,p'}(0,T;V^*)\cap L^m(0,T;W)\times L^{ p'}(0,T;V^*)\to [0,\infty]$ is now defined as $$G_{\rm BEN}(u,y,w)= \left\{ \begin{array}{ll} \left({\displaystyle}\int_0^T\Big (\psi(y') + \psi^*(w) - {\langle}u,y'{\rangle}\Big)\, {{\mathrm{d}}t}+ \phi(y(T)) - \phi(y_0) \right)^+& \\ \quad+ {\displaystyle}\int_0^T \Big (\phi(y) + \phi^*(u-w) - {\langle}u-w, y {\rangle}\Big)\, {{\mathrm{d}}t}& \ \text{if} \ y(0)=y_0\\ \infty& \ \text{otherwise}. \end{array} \right.$$ Indeed, the two nonnegative integrals in the definition of $G_{\rm BEN}$ correspond to the two relations $ w\in \partial \psi(y')$ and $ \partial \phi(y)\ni u -w$, respectively. At the price of introducing the new variable $w$, one can penalize the differential constraint by minimizing $(u,y,w)\mapsto E_{\varepsilon}(u,y,w) = F(u,y,w)+{\varepsilon}^{-1}G_{\rm BEN}(u,y,w)$. Again, the results of Theorem \[thm:BEN\] can be extended to this situation. In particular, it can be proved that $G_{\rm BEN}$ is proper and lower semicontinuous with respect to the strong $\times$ weak $\times$ weak topology of $L^{ p'} (0,T;V^*) \times W^{1,p'}(0,T;V^*)\cap L^m(0,T;W)\times L^{ p'} (0,T;V^*)$. Moreover, it turns out to be coercive as well, as soon as it is restricted to $u\in U$. In particular, no coercivity has to be assumed on $F$ in this case. Indeed, $G_{\rm BEN}$ is here the doubly nonlinear version of the former , which was in fact introduced to ensure coercivity. De Giorgi principle {#sec:DG} =================== Let us now turn out attention to the penalization by means of a variational reformulation of dissipative evolution, following the general approach to gradient flows from [@DeGiorgi80]. Consider again the classical gradient flow in a Hilbert space where now the potential $\phi:H \to (-\infty,\infty]$ is asked to be lower semicontinuous and proper, possibly being nonconvex. To keep notation to a minimum, let us assume $\phi = \phi_1 + \phi_2$ with $\phi_1$ convex, proper, and lower semicontinuous, and $\phi_2\in C^{1,1}$. Then, by letting $\partial \phi$ denote the classical [*Fréchet subdifferential*]{}, namely $$\xi \in \partial \phi(y) \ \ \Leftrightarrow \ \ y \in D(\phi) \ \ \text{and} \ \ \liminf_{w \to y} \frac{\phi(w) - \phi(y) - (\xi,w-y)}{\| y-w\|}\geq 0,$$ (note that the Fréchet subdifferential coincides with the subdifferential of convex analysis on convex functions) we have that $\partial \phi= \partial \phi_1 + D\phi_2$. We will additionally assume $\partial \phi_1$ to be single-valued, whenever nonempty. More general settings are discussed in Subsection \[sec:more\] below. Solutions to correspond to $G_{\rm DG}(u,y)=0$, where the functional $G_{\rm DG}: L^2(0,T;H)\times H^1(0,T;H) \to [0,\infty]$ is defined as $$G_{\rm DG}(u,y)= \left\{ \begin{array}{ll} &{\displaystyle}\int_0^T \left({\displaystyle}\frac12 \| y'\|^2+ \frac12 \|\partial \phi(y){-}u\|^2 - (u,y') \right){{\mathrm{d}}t}+ \phi(y(T))- \phi(y_0) \\[2mm] {}& \qquad \text{if} \ y \in D(\partial \phi) \ \text{a.e.} \ \text{and} \ y(0)=y_0\\[2mm] &\infty \quad \text{otherwise}. \end{array} \right.\label{eq:DG0}$$ Due to its ties with the variational theory of steepest decent in metric spaces from [@DeGiorgi80] we call $G_{\rm DG}$ [*De Giorgi*]{} functional. In we used the notation $D(\partial \phi)$ to indicate the essential domain of $\partial \phi$, namely $D(\partial \phi)=\{y \in H \ : \ \partial \phi(y)\not =\emptyset\}$. Note that, by [@Rossi-Savare06 Lemma 3.4], the map $t\mapsto \partial\phi(y(t))$ is measurable whenever $y\in H^1(0,T;H)$ with $y \in D(\partial \phi)$ a.e. The reformulation of the gradient flow via $G_{\rm DG}$ is based on the computation of the squared residual of , namely, $$\begin{aligned} & \int_0^T \frac12\| y' +\partial \phi(y) -u\|^2{{\mathrm{d}}t}= \int_0^T\left(\frac12\|y'\|^2+\frac12\| \partial \phi(y) -u\|^2+ (y', \partial \phi(y) -u)\right){{\mathrm{d}}t}\\ &\quad= G_{\rm DG}(u,y) \ \ \ \text{if} \ y(0)=y_0.\end{aligned}$$ The latter computation hinges on the chain rule $(\partial \phi(y),y') = (\phi \circ y)'$, which holds in the case of $\phi=\phi_1+ \phi_2$ in the following precise form [@Brezis73 Lemme 3.3] $$\begin{aligned} &y \in H^1(0,T;H), \ \ \xi \in L^2(0,T;H), \ \ \xi \in \partial \phi(y) \ \ \text{a.e. in} \ (0,T)\nonumber\\ & \Rightarrow \ \ \phi \circ y \in AC(0,T) \ \ \text{and} \ \ (\phi \circ y)' = (\xi,y') \ \ \text{a.e. in} \ (0,T). \label{eq:chain}\end{aligned}$$ Indeed, note that $\partial \phi(y) \in L^2(0,T;H)$ if $G_{\rm BEN}(u,y)<\infty$. The main result of this section is the following. \[thm:DG\] Let $\phi=\phi_1+\phi_2:H\to (-\infty,\infty]$ have compact sublevels and fulfill the chain rule , with $\phi_1 $ proper, convex, and lower semicontinuous, $\partial \phi_1$ single-valued, and $\phi_2\in C^{1,1}$. Moreover, let $y_0\in D(\phi)$, $\emptyset\not =U \subset\subset L^2(0,T;H)$, $F: L^2(0,T;H) \times H^1(0,T;H)\to [0,\infty]$ be lower semicontinuous w.r.t. the strong $\times$ weak topology $\tau$ of $L^2(0,T;H)\times H^1(0,T;H)$, $F(u,y)<\infty$ only if $u\in U$, $G_{\rm DG}$ be defined as in , and $E_{\varepsilon}:= F+{\varepsilon}^{-1}G_{\rm DG}$ for ${\varepsilon}>0$. Then, $\min E_{\varepsilon}$ admits a solution for all ${\varepsilon}>0$. Moreover $E_{\varepsilon}\stackrel{\Gamma}{\to} E_0$ with respect to topology $\tau$ where $E_0 =F$ on $\{G_{\rm DG}=0\}$ and $E_0=\infty$ otherwise, and any sequence of quasiminimizers converges, up to a subsequence, to a solution of $\min E_0$. In case $\min E_0$ admits a unique minimizer, any sequence of quasiminimizers converge to it with respect to $\tau$. The statement follows by applying Lemma \[lem\] in the space $X=L^2(0,T;H)\times H^1(0,T;H)$ endowed with its strong $\times$ weak topology $\tau$. Let $u \in U$ and let $y\in H^1(0,T;H)$ be the unique solution of $y'+\partial \phi(y) \ni u$ with $y(0)=y_0$. As we have that $E_{\varepsilon}(u,y) = F(u,y)<\infty$, the functional $E_{\varepsilon}$ is clearly proper. Functional $F$ is $\tau$-lower semicontinuous by assumption. In order to check the $\tau$-lower semicontinuity of $G_{\rm DG}$ let $(u_{\varepsilon},y_{\varepsilon})\stackrel{\tau}{\to} (u,y)$ be given. With no loss of generality, one can assume $\sup_{\varepsilon}G_{\rm DG}(u_{\varepsilon},y_{\varepsilon})\leq c<\infty$. In particular, we can assume that $y_{\varepsilon}'$ and $\partial \phi(u_{\varepsilon})$ are uniformly bounded in $L^2(0,T;H)$. By means of the chain rule we obtain that for all $t \in [0,T]$ $$\begin{aligned} &\phi(y_{\varepsilon}(t)) -\phi(y_0) = \int_0^t (\phi\circ y)'\, {{\mathrm{d}}t}= \int_0^t (\partial \phi(y),y') {{\mathrm{d}}t}\nonumber\\ & \quad\leq \|\partial \phi(y) \|_{L^2(0,T,H)}\|y' \|_{L^2(0,T,H)}<\infty\label{eq:bound}\end{aligned}$$ independently of $t\in [0,T]$ and ${\varepsilon}>0$. This implies that $t\mapsto \phi(y_{\varepsilon}(t))$ is uniformly bounded. As the sublevels of $\phi$ are compact, this yields that there exists $K \subset \subset H$ such that $y_{\varepsilon}(t)\in K$ for all $t\in [0,T]$ and ${\varepsilon}>0$. The uniform bound on $y_{\varepsilon}'$ gives that $y_{\varepsilon}$ are equicontinuous and the Ascoli-Arzelà Theorem implies that, up to not relabeled subsequences, $y_{\varepsilon}\to y$ strongly in $C([0,T];H)$. This entails that $\partial \phi(y_{\varepsilon}) \to \partial \phi(u)$ in $L^2(0,T,H)$ since $\partial \phi$ is strongly $\times$ weakly closed as subset of $L^2(0,T;H)\times L^2(0,T;H)$. Moreover, the strong convergence of $y_{\varepsilon}$ in $C([0,T];H)$ implies that $y_{\varepsilon}(T)\to y(T)$ strongly in $H$, so that $\phi(y(T)) \leq \liminf_{{\varepsilon}\to 0}\phi(y_{\varepsilon}(T))$ as $\phi$ is lower semicontinuous. Since $(u_{\varepsilon},y_{\varepsilon}')\to (u,y')$ strongly in $L^1(0,T)$, we can pass to lower limits in all terms in $G_{\rm DG}(u_{\varepsilon},y_{\varepsilon})$ and thus check that $G_{\rm DG}(u,y)\leq \liminf_{{\varepsilon}\to 0}G_{\rm DG}(u_{\varepsilon},y_{\varepsilon})$. The $\tau$-equicoercivity of $E_{\varepsilon}$ follows as $U$ is compact in $L^2(0,T;H)$ and $G_{\rm DG}(u,y)$ controls the $L^2(0,T;H)$ norm of $y'$. Before closing this subsection, let us record that in the former case of the two functionals $G_{\rm BEN}$ and $G_{\rm DG}$ coincide. In particular, Figure \[illustration\] illustrates the convergence of the penalization via $G_{\rm DG}$ as well. By considering in that same linear ODE example $\phi(y)=\lambda y^2/2$ with $\lambda>0$ instead of $\phi(y)=y^2/2$ one finds the relation $G_{\rm BEN}(u,y)=\lambda G_{\rm DG}(u,y)$, which implies that the minimizers of $F+{\varepsilon}^{-1}G_{\rm BEN}$ and $F+({\varepsilon}/\lambda)^{-1}G_{\rm DG}$ coincide. Hence, for fixed ${\varepsilon}>0$ one has that $G_{\rm BEN}$, respectively $G_{\rm DG}$, delivers the best approximation in terms of minimum and minimizer if $\lambda<1$, respectively $\lambda>1$. This in particular proves that, in general, no functional a priori dominates the other in terms of accuracy of the approximation for fixed ${\varepsilon}$. A numerical simulation ---------------------- In order to present a second illustration of the penalization procedure, let us resort to a nonlinear ODE. We consider the optimal control problem $$\min \left\{ \frac12\int_0^1(y(t)-1)^2{{\mathrm{d}}t}+ \frac12 (u-2)^2 \ : \ y'(t)+y^3(t)=u \ \text{for} \ t \in [0,1], \ y(0)=1 \right\}.\label{ill2}$$ with $u \in {{\mathbb R}}$. By evaluating $u \mapsto F(u,S(u))$ with Matlab, where $y=S(u)$ is the unique solution to $y'+ y^3=u$ with $y(0)=1$, one finds a unique optimal $u \sim 1.016$ and, correspondingly, $F(u,S(u))\sim 0.4917$. The De Giorgi penalized problem for ${\varepsilon}>0$ reads $$\begin{aligned} &\min\left(F+ {\varepsilon}^{-1}G_{\rm DG} \right) = \min \Bigg\{ \frac12\int_0^1(y(t)-1)^2{{\mathrm{d}}t}+ \frac12 (u-2)^2\\ &\quad +\frac{1}{{\varepsilon}}\left(\int_0^1\left(\frac12(y'(t))^2+\frac12(y^3(t){-}u)^2 -u y'(t) \right) \, {{\mathrm{d}}t}+ \frac14 y^4(1)-\frac14\right) \ : \ y(0)=1 \Bigg\}.\end{aligned}$$ The corresponding Euler-Lagrange equations, complemented by the initial condition, reads $$\begin{aligned} - y''(t) + 3(y^3(t)-u)y^2(t) = 0 \ \ \text{for} \ t \in (0,1), \quad y'(1) +y^3(1) = u, \quad y(0)=1.\label{eq:el}\end{aligned}$$ Given $u$, by numerically solving the latter boundary-value problem with Matlab, one finds a critical point $y_{{\varepsilon},u}$ of $E_{\varepsilon}$ and evaluates $u \mapsto E_{\varepsilon}(u,y_{{\varepsilon},u})$. The results of this simulation are illustrated in Figure \[illustration2\], showing convergence of minima and minimizers as ${\varepsilon}\to 0$. More general potentials {#sec:more} ----------------------- The proof of Theorem \[thm:DG\] can be extended to include some more general classes of potentials. A first generalization of the theory allows to treat the case of $\phi=\phi_1 + \phi_2$ with $\partial \phi_1$ not single-valued. In this case, one starts by equivalently rewriting problem as $$\label{eq:gf0circ} y' =\big( u- \partial \phi(y)\big)^\circ \quad \text{in $H$, a.e. in} \ (0,T), \ \ y(0)=y_0.$$ Here, $\big( u- \partial \phi(y)\big)^\circ$ denotes the unique element of minimal norm in the convex and closed set $ u- \partial \phi(y) = u- \partial \phi_1(y) - D\phi_2(y)$. Let us briefly comment on the equivalence of problems and . On the one hand, a solution to clearly solves as well. On the other hand, solutions to are unique: Let $y_1$ and $y_2$ be two solutions, and write $$y_1'-y_2' +\xi_1 - \xi_2 = D\phi_2(y_1) - D\phi_2(y_2) \quad \text{in $H$, a.e. in} \ \ (0,T),$$ where $\xi_i \in \partial \phi_1(y_i)$ a.e. in $(0,T)$, for $i=1,2$. Test the latter equality by $y_1 - y_2$ and integrate on $(0,t)$. By the monotonicity of $\partial \phi_1$ and the Lipschitz continuity of $D\phi_2$ we obtain $$\frac12\| y_1(t) - y_2(t)\|^2 \leq \| D^2 \phi\|_{L^\infty} \int_0^t \| y_1(s) - y_2(s)\|^2{\mathrm{d}}s$$ and $y_1=y_2$ follows by the Gronwall Lemma. Equality can then be equivalently recast as $G_{\rm DG}(u,y)=0$ along with the choice $$G_{\rm DG}(u,y)= \left\{ \begin{array}{ll} &{\displaystyle}\int_0^T \left({\displaystyle}\frac12 \| y'\|^2+ \frac12 \| (\partial\phi(y)-u)^\circ\|^2 - (u,y') \right){{\mathrm{d}}t}+ \phi(y(T))- \phi(y_0) \\ &{}\qquad \text{if} \ y \in D(\partial\phi_1) \ \text{a.e. and} \ y(0)=y_0\\ &\infty \quad \text{otherwise}. \end{array} \right.$$ Note that $G_{\rm DG}$ is proper, as it vanishes on solutions of the gradient flow. In particular, if $G_{\rm DG}(u,y)<\infty$ we have $y\in H^1(0,T;H)$ and we can find $\xi \in L^2(0,T;H)$ such that $\xi -u = (\partial\phi(y)-u)^\circ$ and $\xi \in \partial \phi(y)$ a.e. Then, by means of the chain rule one computes $$(\phi\circ y)' = (\xi,y') = (\xi -u,y') + (u,y') \quad \text{a.e. in} \ \ (0,T).$$ as well as the chain of equivalences $$\begin{aligned} & y' = (u - \partial \phi(y))^\circ \ \ \text{a.e.} \\ &\quad \Leftrightarrow 0 = \frac12 \|y' + \xi-u\|^2 = \frac12\|y'\|^2+\frac12 \| \xi-u\|^2 + (\xi-u,y') \ \ \text{a.e.} \\ &\quad \Leftrightarrow 0 = \frac12\|y'\|^2+\frac12 \| \xi-u\|^2-(u,y') + (\phi\circ y)' \ \ \text{a.e.} \\ & \quad \Leftrightarrow G_{\rm DG}(u,y)=0.\end{aligned}$$ In order to extend the results of Theorem \[thm:DG\] to this case, one just needs to check that, by replacing the term $ \partial \phi(y)-u$ with $( \partial \phi(y)-u)^\circ$ in the functional, coercivity and lower semicontinuity still hold. As for the first, one still has that $\phi$ is controlled along trajectories as in , since $( \partial \phi(y)-u)^\circ= \xi -u $ a.e., for some $\xi \in \partial \phi(u)$ a.e. As for lower semicontinuity, one just needs to be able to pass to the $\liminf$ in the term containing $(\partial \phi(y)-u)^\circ$. By letting $y_{\varepsilon}\to y$ strongly in $C([0,T];H)$ and $\eta_{\varepsilon}=(u - \partial \phi(y_{\varepsilon}))^\circ \to \eta$ weakly in $L^2(0,T;H)$ one finds that $ \xi_{\varepsilon}:=u - \eta_{\varepsilon}\in \partial \phi(y_n)$ a.e. are such that $\xi_{\varepsilon}\to u - \eta =: \xi$ weakly in $L^2(0,T;H)$. Moreover, by the strong $\times$ weak closure of $\partial \phi$ we have that $\xi \in \partial \phi(y)$ a.e. We conclude that $$\begin{aligned} &\frac12\int_0^T \| (\partial \phi(y)-u)^\circ\|^2 {{\mathrm{d}}t}\leq \frac12\int_0^T \| \xi -u\|^2 {{\mathrm{d}}t}= \frac12 \int_0^T \| \eta\|^2 {{\mathrm{d}}t}\\ & \quad\leq \liminf_{{\varepsilon}\to 0}\frac12\int_0^T \| \eta_{\varepsilon}\|^2 {{\mathrm{d}}t}=\liminf_{{\varepsilon}\to 0}\frac12\int_0^T \| ( \partial \phi(y_{\varepsilon})-u)^\circ\|^2 {{\mathrm{d}}t}\end{aligned}$$ and lower semicontinuity of $G_{\rm DG}$ follows. Even more generally, the theory could be adapted to potential which are not $C^{1,1}$ perturbations of convex functions. The reader is referred to [Rossi & Savaré]{} [@Rossi-Savare06] where a general frame for existence of solutions to gradient flows on nonconvex functionals is addressed. In this context, weaker notions of (sub)differential are introduced and the validity of a corresponding chain rule as in is discussed. In particular, examples of operators fulfilling a suitable chain rule are presented, including classes of dominated concave perturbations of convex functions. Let us mention that the validity of a chain rule [*equality*]{}, albeit of a paramount importance in order to relate the minimization of $G_{\rm DG}$ to the solution of , is actually not needed to prove Theorem \[thm:DG\]. In fact, the chain rule has been used there just to check that the potential $\phi$ remains uniformly bounded along trajectories. In particular, a suitable chain-rule [*inequality*]{} would serve for this purpose as well. Generalized gradient flows -------------------------- The De Giorgi functional approach can be adapted to encompass [ *generalized gradient flows*]{}, namely relations of the form $$\label{eq:gf33} \partial \psi (y,y') + \partial \phi(y)\ni u \quad \text{for a.e.} \ t\in (0,T), \quad y(0)=y_0.$$ Here, $\psi: H \times H \to [0,\infty)$ and $\partial \psi (y, y') $ denotes partial subdifferentiation with respect to the second variable only. More precisely, we assume that the map $v\in H\mapsto \psi(y,v)$ is convex and lower semicontinuous for all $y\in H$, the map $ (y,v,w)\in H\times H\times H \mapsto \psi(y,v) + \psi^*(y,w)$ is weakly lower semicontinuous and $$\begin{aligned} \psi (y,v)+\psi^*(y,w)\geq c \|v\|^p+ c\|w\|^{p'} \quad \forall y,\, v,\, w\in H \label{eq:psi5}\end{aligned}$$ and some $p>1$ where $p'=p/(p-1)$ and the Legendre-Fenchel conjugation is taken with respect to the second variable only. An example for $\psi$ satisfying is $\psi(y,y) = \beta(y)|y|^p$, where $p>1$ and $\beta$ is sufficiently smooth, uniformly positive, and bounded. Note that this includes the case of [*doubly nonlinear*]{} flows. As in Theorem \[thm:DG\], we assume for simplicity that $\partial\phi = \partial \phi_1 + D\phi_2$ and is single-valued. Solutions to can be characterized via $G_{\rm DG}(u,y)=0$ where $G_{\rm DG}: L^p(0,T;H)\times W^{1,q}(0,T;H) \to [0,\infty]$ is defined as $$G_{\rm DG}(u,y)= \left\{ \begin{array}{ll} & {\displaystyle}\int_0^T \left({\displaystyle}\psi(y,y') +\psi^*(y, u{-}\partial \phi(y))- (u,y') \right){{\mathrm{d}}t}+ \phi(y(T)) - \phi(y_0) \\ &\qquad \text{if} \ y \in D(\partial \phi) \ \text{a.e and} \ y(0)=y_0\\[2mm] &\infty \quad \text{otherwise}. \end{array} \right.\label{eq:DG}$$ This can be checked by equivalently rewriting $$\begin{aligned} &\partial \psi (y,y') + \partial \phi(y)\ni u \ \ \text{a.e.} \\ & \Leftrightarrow \psi(y,y')+ \psi^*(y,u-\partial \phi(y)) - (u-\partial \phi(u),y')=0 \ \ \text{a.e.} \\ & \Leftrightarrow \psi(y,y')+ \psi^*(y,u-\partial \phi(y)) - (u,y') + (\phi\circ y)'=0 \ \ \text{a.e.} \\ &\Leftrightarrow G_{\rm DG}(u,y)=0.\end{aligned}$$ Indeed, the last equivalence follows by integrating the second-last relation in time, in one direction, and by realizing that the integrand is always nonnegative, in the other direction. By replacing $\|\cdot \|^2/2$ by $\psi(y,\cdot)$ under assumption , an analogous statement to Theorem \[thm:DG\] holds. More precisely, by assuming $F:L^p(0,T;H) \times W^{1,p'}(0,T;H) \to [0,\infty)$ to be lower semicontinuous in $X=L^p(0,T;H) \times W^{1,p'}(0,T;H)$ with respect to the strong$\times$ weak topology and $U$ to be compact in $ L^p(0,T;H)$, one can reproduce the former argument. Note however that extra conditions have to be imposed in such a way that pairs with $G_{\rm BEN}(u,y)=0$ exist. GENERIC flows ------------- The applicability of the penalization technique via the De Giorgi functional can be extended to classes of so-called GENERIC flows (General Equations for Non-Equilibrium Reversible-Irreversible Coupling). These are systems of the form $$\label{eq:gf4} y' = L(y) \,D E(y) -K(y) (\partial \phi(y)-u) \quad \text{for a.e.} \ t\in (0,T), \quad y(0)=y_0.$$ Here, $-\phi$ is to be interpreted as the [*entropy*]{} and will have the property of being nondecreasing in time. The functional $E: H \to {{\mathbb R}}$ represents an [*energy*]{}, to be conserved along trajectories instead. For the sake of simplicity, we assume $E$ to be Fréchet differentiable, with a linearly bounded, strongly $\times$ weakly closed differential $DE$. The mapping $K:H \to {\mathcal L}(H)$ (linear and continuous operators) is the so called [ *Onsager*]{} operator and is asked to be continuous with symmetric and positive semidefinite values. On the other hand, the operator $L: H \to {\mathcal L}(H)$ is required to be continuous with antiselfadjoint values, namely $L^*(y) = -L(y)$. The GENERIC formalism [@Grmela] is a general approach to the variational formulation of physical models and is particularly tailored to the unified treatment of coupled conservative and dissipative dynamics. Potentials and operators are related by the following structural assumptions $$\label{eq:gen} L^*(y)\partial \phi (y) = K^*(y) D E(y)=0.$$ These guarantee that solutions of are such that $(E \circ y)'=0$ and $(-\phi\circ y)'\geq 0$, namely energy is conserved and entropy increases along trajectories. To date, GENERIC has been successfully applied to a variety of situations ranging from complex fluids [@Grmela], to dissipative quantum mechanics [@Mielke13], to thermomechanics [@generic_souza; @Mielke11], and to the Vlasov-Fokker-Planck equation [@Peletier]. By defining the convex potential $\xi \mapsto \psi^*(y,\xi)=(K(y)\xi,\xi)/2$, so that $K(y)=\partial \psi^*(y,\cdot)$ (subdifferential with respect to the second variable only), problem can be reformulated as $G_{\rm DG}(u,y)=0$ where now $G_{\rm DG}: L^2(0,T;H) \times H^1(0,T;H) \to [0,\infty]$ is defined as $$G_{\rm DG}(u,y)= \left\{ \begin{array}{ll} &{\displaystyle}\int_0^T \Big({\displaystyle}\psi(y,y'{-}L(y)\,DE(y)) +\psi^*(y, u{-}\partial \phi(y))\Big){{\mathrm{d}}t}\\ &\quad {-} {\displaystyle}\int_0^T (u,y'{-}L(y)\, DE(y)) \, {\mathrm{d}}t + \phi(y(T)) - \phi(y_0) \\[4mm] &\qquad \text{if} \ y \in D(\partial \phi) \ \text{a.e. and} \ y(0)=y_0\\[3mm] &\infty \quad \text{otherwise}. \end{array} \right.\label{eq:DG}$$ In fact, we have the following chain of equivalencies $$\begin{aligned} &y' = L(y)\,D E(y) -K(y) (\partial \phi(y)-u) \ \ \text{a.e.}\\ &\Leftrightarrow\ \psi(y,y'{-}L(y)\,DE(y))+\psi^*(y,u{-}\partial\phi(y)) - \big( y'{-}L(y)\,DE(y), u{-} \partial \phi(y)\big) =0 \ \ \text{a.e.}\\ & \Leftrightarrow \ \psi(y,y'{-}L(y)\,D E(y))+\psi^*(y,u{-}\partial\phi(y)) \\ &\qquad - (u,y'-L(y)\, DE(y)) -(DE(y),L^*(y)\partial \phi(y)) + ( \phi\circ y)' =0 \ \ \text{a.e.} \\ &\Leftrightarrow\ G_{\rm DG}(u,y)=0.\end{aligned}$$ Again, the last equivalence follows by integration in time. The statement of Theorem \[thm:DG\] can be extended to cover the case of GENERIC flows as well. Let us assume from the very beginning that for all $u\in H$ there exists $y$ such that $G_{\rm DG}(u,y)=0$. In applications $K$ and $\phi$ are often degenerate (see below). Coercivity for the sole $G_{\rm DG}$ is hence not to be expected. In order to state a general result, let us hence assume $F$ itself to be lower semicontinuous and coercive with respect to the strong $\times$ weak topology of $L^2(0,T;H) \times H^1(0,T;H)$. Moreover, let $F$ be coercive with respect to the strong $\times$ strong topology of $L^2(0,T;H)\times C([0,T];H)$ on sublevels of $\phi$ and to control the $L^2(0,T;H) $ norm of $\partial \phi(y)$ (alternatively, let $\partial \phi(y)$ be linearly bounded). Eventually, we ask $\psi^*$ and $\psi$ to be lower semicontinuous in the following sense $$\begin{aligned} &\psi(y,\eta)+\psi^*(y,\xi) \leq \liminf_{{\varepsilon}\to 0}\big(\psi(y_{\varepsilon},\eta_{\varepsilon})+\psi^*(y_{\varepsilon},\xi_{\varepsilon}) \big)\nonumber\\ &\quad\forall y_{\varepsilon}\to y \ \ \text{strongly in} \ \ C([0,T];H) \ \ \text{with} \ \ \sup \phi(y_{\varepsilon}(t)) <\infty\nonumber\\ &\quad \text{and} \ \ (\eta_{\varepsilon},\xi_{\varepsilon}) \to (\eta,\xi) \ \ \text{weakly in} \ \ L^2(0,T;H)^2.\label{eq:mosco}\end{aligned}$$ Owing to the assumptions on $F$, in order to reproduce the argument of Theorem \[thm:DG\] in this setting, one is left to check the lower semicontinuity of $G_{\rm DG}$. Let $(u_{\varepsilon},y_{\varepsilon})\to (u,y)$ strongly $\times$ weakly in $L^2(0,T;H) \times H^1(0,T;H)$ and assume with no loss of generality that $\partial \phi(y_{\varepsilon})$ is bounded in $L^2(0,T;H)$. By arguing as in one can bound $t\mapsto \phi(y_{\varepsilon}(t))$ so that all trajectories belong to a sublevel of $\phi$. From the strong coercivity of $F$ on sublevels of $\phi$ we deduce strong compactness in $C([0,T];H)$ for $y_{\varepsilon}$, so that $y_{\varepsilon}\to y$ uniformly, up to not relabeled subsequences. As $DE$ is assumed to be strongly $\times$ weakly closed and $L$ is continuous, we have that $y_{\varepsilon}'-L(y_{\varepsilon})\, DE(y_{\varepsilon}) \to y'-L(y)\, DE(y)$ weakly in $L^2(0,T;H)$. On the other hand, the strong $\times$ weak closure of $\partial \phi$ ensures that, again without relabeling, $\partial \phi(y_{\varepsilon}) \to \partial \phi(y)$ weakly in $L^2(0,T;H)$. We can hence make use of and deduce the lower semicontinuity of $G_{\rm DG}$. Before closing this discussion, let us give an example of an elementary GENERIC system fitting into this abstract setting. Consider the thermalized oscillator problem $$\begin{aligned} & q'' + \nu q' + \lambda q + \theta =0,\label{eq:o1}\\ &\kappa \theta' = \nu ( q')^2 + \theta q'. \label{eq:o2}\end{aligned}$$ Here, $y=(q,p,\theta)\in {{\mathbb R}}^3=:H$ where $q$ represents the state of the oscillator, $p$ is its momentum, and $\theta>0$ is the absolute temperature. The nonnegative constants $\nu$, $\lambda$, and $\kappa$ are the viscosity parameter, the elastic modulus, and the heat capacity, respectively. Relations and express the conservation of momentum and energy, respectively. In order to reformulate - as a GENERIC system, we specify the free energy of the system as $$\Psi(y)= \frac{\lambda}{2}q^2 + q\theta - \kappa\theta \ln \theta.$$ Moving from this, the entropy $-\phi$ and the total energy $E$ are derived by the classical Helmholtz relations as $$\begin{aligned} -\phi(y)= -\partial_\theta \Psi= -q+\kappa\ln \theta +\kappa, \quad E(y)= \frac{1}{2}p^2+ \Psi+\theta \phi = \frac{1}{2}p^2 +\frac{\lambda}{2} q^2+ \kappa\theta.\end{aligned}$$ In particular, we have that $$DE(y) = (\lambda q, p, \kappa), \quad \partial \phi(y) = (-1,0,\kappa/\theta).$$ By defining the mappings $K$ and $L$ as $$K(y)=\nu \theta \left( {\displaystyle}\begin{matrix} 0 & 0 & 0 \\ 0 & 1 & -p/\kappa\\ 0 & -p/\kappa & p^2/\kappa^2 \end{matrix} \right), \quad L(y)= \left( {\displaystyle}\begin{matrix} 0 & 1& 0 \\ -1 &0 &{\displaystyle}-\theta/\kappa\\ 0 &{\displaystyle}\theta/\kappa& 0 \end{matrix} \right),$$ we readily check that the compatibility conditions hold and that system - takes the form in . By computing the conjugate we find $$\psi^*(y,\xi) = \frac{\nu \theta}{2}(\xi_2 - p \xi_3/\kappa)^2, \quad \psi(y,\eta) = \left\{ \begin{array}{ll} {\displaystyle}\frac{1}{2\nu \theta}\eta_2^2& \quad \text{if} \ \eta_1=\eta_3+py_2/\kappa=0,\\ \infty&\quad \text{otherwise} \end{array} \right.$$ for all $y=(q,p,\theta)\in {{\mathbb R}}^3$ with $\theta>0$ and for all $(\xi,\eta)\in {{\mathbb R}}^2$. In particular, the lower semicontinuity follows as $\sup \phi(y_{\varepsilon}(t))<\infty$ implies that $\theta_{\varepsilon}\geq c>0$ for some $c$, hence $1/\theta_{\varepsilon}\to 1/\theta$ in $C([0,T])$. In order to give a concrete example of target functional $F$ choose $$\begin{aligned} F(u,y) &= \frac12\int_0^T |y-y_{\rm target}|^2{\mathrm{d}}t + \frac12\int_0^T |y'-y_{\rm target}'|^2{\mathrm{d}}t +\int_0^T |1/\theta - 1/\theta_{\rm target}|^2{\mathrm{d}}t\\ &+ \int_0^T |u|^2 \, {{\mathrm{d}}t}+ \int_0^T |u'|^2 \, {{\mathrm{d}}t}\end{aligned}$$ for some given $y_{\rm target} = (q_{\rm target},p_{\rm target},\theta_{\rm target}) \in H^1(0,T;H)$ with $1/\theta_{\rm target} \in L^2(0,T)$. The functional $F$ is coercive with respect to the strong $\times$ weak topology of $L^2(0,T;H) \times H^1(0,T;H)$, as well as to the strong$\times$strong topology of $L^2(0,T;H)\times C([0,T];H)$ on sublevels of $\phi$. Moreover, it controls the $L^2(0,T;H) $ norm of $\partial \phi(y)$. Hence, the abstract setting described above applies. Acknowledgement {#acknowledgement .unnumbered} =============== This work has been funded by the Vienna Science and Technology Fund (WWTF) through Project MA14-009 and by the Austrian Science Fund (FWF) projects F65 and I2375. [99]{} . [Saddle-points and existence-uniqueness for evolution equations]{}. [*Differential Integral Equations*]{}, 6 (1993), 1161–1171. F. Auricchio, E. Boatti, A. Reali, U. Stefanelli. [Gradient structures for the thermomechanics of shape-memory materials]{}. [*Comput. Methods Appl. Mech. Engrg.*]{} 299 (2016), 440–469. M. Bergounioux. A penalization method for optimal control of elliptic problems with state constraints. [*SIAM J. Control Optim.*]{} 30 (1992), 305–323. M. Bergounioux. Optimal control of parabolic problems with state constraints: a penalization method for optimality conditions. [ *Appl. Math. Optim.*]{} 29 (1994), 285–307. M. Bergounioux. Optimal control of problems governed by abstract elliptic variational inequalities with state constraints. [*SIAM J. Control Optim.*]{} 36 (1998), 273–289. . . (1973). . [Un principe variationnel associé à certaines équations paraboliques. [L]{}e cas indépendant du temps]{}. [*C. R. Acad. Sci. Paris Sér. A-B*]{}, 282 (1976), A971–A974. . [Un principe variationnel associé à certaines équations paraboliques. [L]{}e cas dépendant du temps]{}. [*C. R. Acad. Sci. Paris Sér. A-B*]{}, 282 (1976), A1197–A1198. F. Browder and P. Hess. Nonlinear mappings of monotone type in Banach spaces. [*J. Funct. Anal.*]{} 11 (1972), 251–294. G. [Dal Maso]{}. . Progress in Nonlinear Differential Equations and their Applications, 8. Birkhäuser Boston Inc., Boston, MA, 1993. . On a type of variational convergence. In [*Proceedings of the Brescia Mathematical Seminar*]{}, 9, 63–101, Milan, 1979. E. De Giorgi, A. Marino, and M. Tosques. Problems of evolution in metric spaces and maximal decreasing curve. . 68 (1980), 180–187. M. H. Duong, M. A. Peletier, J. Zimmer. GENERIC formalism of a Vlasov–Fokker–Planck equation and connection to large-deviation principles. [*Nonlinearity*]{}, 26 (2013), 2951–2971. S. Fitzpatrick. Representing monotone operators by convex functions. Workshop/Miniconference on Functional Analysis and Optimization (Canberra, 1988), 59–65, [*Proc. Centre Math. Anal. Austral. Nat. Univ.*]{} 20, Austral. Nat. Univ., Canberra, 1988. C. M. Gariboldi and D. A. Tarzia. Convergence of distributed optimal controls in mixed elliptic problems by the penalization method. [*Math. Notae*]{}, 45 (2007/08), 1–19. . [A least action principle for steepest descent in a non-convex landscape]{}. In Partial differential equations and inverse problems, vol. 362 of [*Contemp. Math.*]{}, Amer. Math. Soc., Providence, RI, 2004, pp. 177–187. N. Ghoussoub, L. Tzou. A variational principle for gradient flows. , 330 (2004), 519–549. N. Ghoussoub. . Universitext. Springer, New-York, 2009. M. Grmela and H. C. Öttinger. [Dynamics and thermodynamics of complex fluids. I. Development of a general formalism]{}. [ *Phys. Rev. E*]{}, 56 (1997), 6620–6632. M. D. Gunzburger and H.-C. Lee. A penalty/least-squares method for optimal control problems for first-order elliptic systems. [ *Appl. Math. Comput.*]{} 107 (2000), 57–75. N. Kenmochi. Nonlinear evolution equations with variable domains in Hilbert spaces. [*Proc. Japan Acad. Ser. A Math. Sci.*]{} 53 (1977), 163–166. N. Kenmochi, T. Koyama. Nonlinear functional variational inequalities governed by time-dependent subdifferentials. [*Nonlinear Anal.*]{} 17 (1991), 863–883. . [An asymptotical variational principle associated with the steepest descent method for a convex function]{}. [*J. Convex Anal.*]{} 3 (1996), 63–70. J.-L. Lions. [*Contrôle optimal de systèmes gouvernés par des équations aux dérivées partielles*]{}. Gauthier-Villars, Paris 1968. M. Mabrouk. A variational approach for a semi-linear parabolic equation with measure data. [*Ann. Fac. Sci. Toulouse Math.*]{} 9 (2000), 91–112. M. Mabrouk. A variational principle for a nonlinear differential equation of second order. [*Adv. in Appl. Math.*]{} 31 (2003), 388–419. A. Mielke. [Formulation of thermoelastic dissipative material behavior using GENERIC]{}. [*Contin. Mech. Thermodyn.*]{} 23 (2011), 233–256. A. Mielke. Dissipative quantum mechanics using GENERIC. Proc. of the conference on [*Recent Trends in Dynamical Systems*]{}, vol. 35 of Proceedings in Mathematics & Statistics, Springer, 2013, pp. 555–585. G. Mophou and G. M. N’Guérékata. Optimal control of a fractional diffusion equation with state constraints. [*Comput. Math. Appl.*]{} 62 (2011), 1413–1426. J.-J. Moreau. Evolution problem associated with a moving convex set in a Hilbert space. [*J. Differential Equations*]{}, 26 (1977), 347-374. . [ Deux théorèmes de minimum pour certains systèmes dissipatifs]{}. [*C. R. Acad. Sci. Paris Sér. A-B*]{}, 282 (1976), A1035–A1038. . [ Un théorème de minimum pour certains systèmes dissipatifs. [V]{}ariante hilbertienne]{}. [*Travaux Sém. Anal. Convexe*]{}, 6 (1976), 22. . [Étude de la question d’existence pour certains problèmes d’évolution par minimisation d’une fonctionnelle convexe]{}. [*C. R. Acad. Sci. Paris Sér. A-B*]{}, 283 (1976), A83–A86. R. Rossi and G. Savaré. . 12 (2006), 564–614. . [Direct method for parabolic problems]{}. [ *Adv. Math. Sci. Appl.*]{} 10 (2000), 57–65. U. Stefanelli. The Brezis-Ekeland principle for doubly nonlinear equations. [*SIAM J. Control Optim.*]{} 47 (2008), 1615–1642 . . 16 (2009), 71–87. . . 40 (2008), 623–652. A. Visintin. Extension of the Brezis-Ekeland-Nayroles principle to monotone operators. [*Adv. Math. Sci. Appl.*]{} 18 (2008), 633–650. A. Visintin. On the variational representation of monotone operators. [*Discrete Contin. Dyn. Syst. Ser. S*]{}, 10 (2017), 909-–918. A. Visintin. Structural compactness and stability of semi-monotone flows. [*SIAM J. Math. Anal.*]{} 50 (2018), 2628–2663. Y. Yamada. On evolution equations generated by subdifferential operators. [*J. Fac. Sci. Univ. Tokyo Sect. IA Math.*]{} 23 (1976), 491–515.

ArXiv

--- author: - 'Bart[ł]{}omiej Czech,' - 'Lampros Lamprou,' - 'Samuel McCandlish,' - James Sully title: Tensor Networks from Kinematic Space --- Introduction ============ Gravitational physics presents us with a paradox. On the one hand, its most successful formulation to date – the general theory of relativity – relies on differential geometry, which emphasizes local dynamics. On the other hand, all gauge-invariant observables in gravity live on the asymptotic boundary and are therefore global in character. While the local approach has been pursued with undiminished success for one century [@einstein1915], a more global strategy has not yet congealed into a unified formalism. The best developed attempt to fill this gap is the AdS/CFT correspondence [@adscft], which organizes the gauge-invariant quantities in a gravitational spacetime into a field theory living on its asymptotic boundary. A key challenge facing the holographic program – one that has come to the spotlight in recent years [@samirstheorem; @amps] – is this: how do we reconcile the CFT-based, global formulation of gravity with the local language of general relativity? In Ref. [@lastpaper], we outlined a semantically evident answer to this question: to complement Einstein’s apparatus of differential geometry, we need an approach based on integral geometry [@santalo]. This beautiful field of mathematics is concerned with translating between local and global properties of geometric spaces. A well-known application is to recover a function from its integrals along straight lines [@helgason], a problem that occurs in seismology [@seismology], medical imaging [@ctscans] and the reconstruction of bulk operators in holographic duality [@ooguri]. In [@lastpaper], we focused on a prequel to this problem: determining the geometry of an asymptotically AdS$_3$ spacetime from data in the dual conformal field theory. Generally, the input to the reconstruction problem consists of all correlation functions in the CFT. However, recent years have taught us that information-theoretic CFT data are particularly robust probes of the bulk geometry. The foremost among them are entanglement entropies of boundary regions, which compute areas of bulk minimal surfaces [@rt1; @rt2; @hrt].[^1] In Ref. [@lastpaper], we used entanglement entropies to define an auxiliary, Lorentzian geometry, whose points are in one-to-one correspondence with boundary intervals and, by the Ryu-Takayanagi proposal [@rt1], with spacelike geodesics in the dual geometry. The resulting object, called kinematic space, is an intermediary in the AdS/CFT translation, providing a natural volume form on the space of bulk geodesics. Integrals of that form compute lengths of all bulk curves in a generalization of the famous Crofton formula, which tells us how likely a dropped needle is to land on a single bathroom tile [@bouffon]. But the problems solved by kinematic space are not confined to holographic duality. Even in the absence of a gravitational dual, $d$-dimensional conformal field theory intertwines space and scale (RG direction), as is evident from its global symmetry group SO$(d,2)$. One may imagine that kinematic space, which organizes the entanglement structure of a state by location and scale, may have already found use in the study of conformal field theories, independently of holographic considerations. If so, in what form has kinematic space previously appeared? The answer is the Multi-Scale Entanglement Renormalization Ansatz (MERA) [@mera; @mera2]. The present paper explains the merits of viewing the MERA network as a discrete version of the vacuum kinematic space. The argument makes crucial use of the auxiliary causal structure of MERA, which originates from working with unitary and isometric tensors as part of the ansatz. This causal structure was independently exploited to argue that MERA most naturally lives on de Sitter space [@beny]. Our key insight is to recognize that this de Sitter space is the vacuum kinematic space, which carries an information metric determined by entanglement. This allows us to propose a generalization of MERA to excited states [@compression]. ![MERA naturally lives on half of two-dimensional de Sitter space, the kinematic space for an equal-time slice of AdS$_3$.[]{data-label="fig:MERAondS"}](Images/dsTexture){width="60.00000%"} Our results do not preclude other types of networks, for example ones studied in [@errorcorrectingnetwork; @patrickxiaoliang], from discretizing a time slice of the dual geometry directly. #### Reading guide In an effort to keep the paper self-contained, we begin with a review of integral geometry and MERA (Sec. \[sec:review\]). Secs. \[sec:therelation\] and \[sec:BHmera\] contain the main arguments for identifying MERA with the kinematic geometry. Sec. \[sec:therelation\] is set in the broad context of the ground state MERA while Sec. \[sec:BHmera\] discusses the recently reported [@quotientmera] MERA construction of the thermofield double state. Sec. \[compressionsummary\] briefly motivates the results of the second part of this work, which will be presented in [@compression]. Sec. \[summary\] summarizes our core arguments and discusses the main results of this work. Review {#sec:review} ====== We begin by reviewing the properties of kinematic space and the MERA tensor network. The reader is encouraged to look for commonalities. Kinematic space --------------- A more complete discussion of the ensuing material was given in [@lastpaper]. ### Crofton’s formula in flat space Crofton’s formula states that the length of a curve is measured by the number of straight lines that intersect it. To state this result formally, we need to clarify how to count straight lines. Straight lines on the plane form a two-dimensional manifold $K$ known as ‘kinematic space.’ To quantify ‘how many’ straight lines $g$ satisfy some condition, we need a homogeneous measure $\mathcal{D}g$ on kinematic space. Using translations and rotations fixes the measure, up to a multiplicative constant, to be $$\mathcal{D}g = dp \wedge d\theta. \label{flatmeasure}$$ Here $p$ is the distance of the straight line from the origin and $\theta$ is the angle it makes with some fixed axis. Allowing $p$ to be negative extends the measure to the set of oriented straight lines. Crofton’s formula states that, for every curve $\gamma$ of finite length, $$\textrm{length of }\gamma = \frac{1}{4} \int_K n(g,\gamma)\,\,\mathcal{D}g. \label{croftonformula}$$ Here $n(g, \gamma)$ is the number of intersections between the straight line $g$ and $\gamma$. This result can be used to solve Buffon’s needle problem [@bouffon], which we referenced in the Introduction. Our primary interest is in an extension of this formula to holographic spacetimes. ### Crofton’s formula in holographic geometries ![The kinematic coordinates $\alpha$ and $\theta$ correspond to the half-opening angle of the geodesic and the angular location of its center-point repectively. Geodesics in the hyperbolic plane are mapped to points on kinematic space.[]{data-label="fig:KScoords"}](Images/KScoords "fig:"){width="30.00000%"} ![The kinematic coordinates $\alpha$ and $\theta$ correspond to the half-opening angle of the geodesic and the angular location of its center-point repectively. Geodesics in the hyperbolic plane are mapped to points on kinematic space.[]{data-label="fig:KScoords"}](Images/kinematic "fig:"){width="30.00000%"} To extend eq. (\[croftonformula\]) to a static slice of a holographic geometry, one need only supply the correct measure on the generalized kinematic space. In pure AdS$_{3}$ of curvature radius $L$, the measure is again uniquely selected by invariance under the isometries of $\mathbb{H}_2$ [@solanes], and takes the form $$\mathcal{D}g=\frac{L\,d\theta\wedge d\alpha}{\sin^{2}\alpha}.\label{hypmeasure}$$ This expression gives a ‘density of geodesics’ near a geodesic centered at a boundary point $\theta$, with opening angle $\alpha$ (see Fig. \[fig:KScoords\]). Looking ahead to a connection with the MERA network, we note that $\theta$ and $\alpha$ play the role of position and scale in the CFT, respectively. In a general, static, holographic geometry, the measure cannot be found from symmetry alone. In Ref. [@lastpaper], we showed that when the tangent bundle of a bulk time-slice is covered by boundary-anchored geodesics, the measure takes a remarkably simple form, $$\mathcal{D}g=\frac{\partial^{2}S(u,v)}{\partial u\,\partial v}\,du\wedge dv,\label{kinvolume}$$ where $S(u,v)$ is the length of a geodesic connecting the boundary points $u$ and $v$. Here, we have defined ‘light-cone’ coordinates on kinematic space, $$u=\theta-\alpha\qquad{\rm and}\qquad v=\theta+\alpha,\label{nullcoords}$$ which label a geodesic by its asymptotic endpoints. Then eq. (\[kinvolume\]) agrees with eq. (\[hypmeasure\]) after substituting $S(u,v)=2L\log\sin\frac{v-u}{2}+{\rm const}$. When $S(u,v)$ refers to the length of the shortest geodesic connecting points $u$ and $v$ satisfying the homology condition, its length in units of $4G$ is the entanglement entropy of the boundary interval $(u,v)$ [@rt1]. Thus, it is convenient to divide both sides of eq. (\[croftonformula\]) by $4G$ and obtain: $$\frac{\textrm{length of }\gamma}{4G} = \frac{1}{4} \int_K n(g,\gamma)\, \frac{\partial^2 S_{\rm ent}(u,v)}{\partial u \, \partial v}\, du\wedge dv \label{croftonformula4G}$$ In what follows, we set $4G \equiv 1$ and do not distinguish between $S(u,v)$ as a length and $S_{\rm ent}(u,v)$ as an entanglement entropy. #### Differential entropy For a closed curve, it is instructive to carry out the integral in (\[croftonformula4G\]) explicitly in one direction: $$\textrm{length of }\gamma = - \int_0^{2\pi} du\, \frac{\partial S(u,v)}{\partial u}\,\Big|_{v = v(u)} = S_{\rm diff} \label{sdiff}$$ This expression is the differential entropy, first reported in [@holeography] (see also [@robproof]). It localizes on the set of geodesics tangent to $\gamma$, which is the boundary of the set of its intersecting geodesics. The tangency condition appears through $v(u)$, which is defined by demanding that the geodesic connecting $u$ and $v(u)$ be tangent to $\gamma$. In Sec. \[cutcount\] we will find an analogue of this expression in the cut-counting prescription for estimating entropies in MERA. ### Causal structure and the kinematic metric {#kincausality} The kinematic space for a static slice of an asymptotically AdS$_{3}$ geometry has a richer structure than just a density form: It can also be equipped with a metric with mixed signature. To see this, note that the space of geodesics maps naturally to the space of boundary intervals via the R-T prescription. The causal structure of kinematic space descends not from the causal structure of AdS$_{3}$, but from the partial ordering of boundary intervals by containment. In particular, given two boundary intervals $A,B$ corresponding to two points $a,b$ in kinematic space, we say that $a$ causally precedes $b$ if $A\subset B$. Any pair of geodesics may then be classified as timelike, lightlike, or spacelike-separated: - Timelike: geodesic $(u_{1},v_{1})$ is said to live in the past of geodesic $(u_{2},v_{2})$ if $$[u_{1},v_{1}] \subset [u_{2},v_{2}]\label{containment}$$ as intervals on the asymptotic boundary. Note that the direction of kinematic ‘time’ reverses under changes of orientation. For the same geodesics with opposite orientation, we have $[v_{2},u_{2}]\subset[v_{1},u_{1}]$. - Spacelike: geodesics $(u_{1},v_{1})$ and $(u_{2},v_{2})$ are spacelike separated when neither interval contains the other. - Lightlike: This is the borderline case between spacelike and timelike separation. It occurs when one of the intervals subtended by the geodesics contains the other, but only marginally. This means that the intervals share an endpoint – on the left or on the right: $$u_1=u_2 \,\,\,\text{ or }\,\,\, v_1=v_2$$ This is the reason why above eq. (\[nullcoords\]) we referred to the endpoint coordinates of kinematic space as ‘light-cone’ coordinates. #### Kinematic Metric For a time-slice of pure AdS$_{3}$, we can now see that symmetry fixes the metric on kinematic space to be the two-dimensional de Sitter metric [@solanes; @lampros]: $$ds_{{\rm kin}}^{2}=\frac{L}{\sin^{2}\alpha}\left(-d\alpha^{2}+d\theta^{2}\right).\label{adskinmetric}$$ To see this, note that $\text{dS}_{2}$ is the only metric space with $\text{SO}\left(2,1\right)$ isometry group that realizes the requisite causal structure. With the coefficient above, the volume form $d^{2}V_{\text{kin}}$ in kinematic space is equal to the geodesic density $\mathcal{D}g$ of eq. (\[hypmeasure\]). Moving to a general holographic geometry, specifying the causal structure and the volume form $d^{2}V_{\text{kin}}=\mathcal{D}g$ yields a unique Lorentzian metric: $$ds_{{\rm kin}}^{2}=\frac{\partial^{2}S(u,v)}{\partial u\,\partial v}\,du\,dv\label{kinmetric}$$ The relevance of this metric for reconstructing local features of the bulk geometry was reported in [@lampros; @lastpaper]. ### Conditional mutual information in kinematic space {#kincmi} ![Volumes of causal diamonds in kinematic space compute conditional mutual informations of triples of contiguous intervals. As a special case, causal diamonds with one vertex on the boundary compute mutual informations of adjacent intervals.[]{data-label="fig:KSmutual"}](Images/KSmutualinfo.pdf){width="60.00000%"} The volume form (\[kinvolume\]) has a meaning in information theory. Conditional mutual information is defined as the following combination of entanglement entropies: $$I(A,C|B) = S(AB) + S(BC) - S(ABC) - S(B) \label{defcmi}$$ Mutual information $I(A, C)$ is a special case of this quantity, conditioned on . Conditional mutual information is also familiar from the strong subadditivity of entanglement entropy, which guarantees that it is non-negative [@ssa]. For the special choice $$A = (u-du, u) \qquad {\rm and} \qquad B = (u, v) \qquad {\rm and} \qquad C = (v, v+dv)\,, \label{3intervals}$$ we have: $$S(u-du,v) + S(u,v+dv) - S(u-du,v+dv) - S(u,v) = \frac{\partial^2 S(u,v)}{\partial u \,\partial v}\,du\,dv = d^2V_{\rm kin}\,. \label{infinitesimalcmi}$$ Eq. (\[infinitesimalcmi\]) states that the Lorentzian area of an infinitesimal causal diamond in kinematic space computes the conditional mutual information of a triple of neighboring intervals (\[3intervals\]). Owing to the chain rule for conditional mutual information $$I(A, CD|B) = I(A,C|B) + I(A, D|BC)\,,$$ this conclusion automatically extends to all causal diamonds in kinematic space, regardless of size (see Fig. \[fig:KSmutual\]). A special case is a causal diamond with one of its vertices on the boundary of kinematic space, whose volume is equal to the mutual information of the two adjacent intervals. Thus, eq. (\[croftonformula4G\]) states that the length of any curve on a static slice of an asymptotically AdS$_3$ geometry computes a combination of conditional mutual informations. The MERA network ---------------- Our presentation will be brief, because good reviews exist elsewhere [@mera-review; @mera-revresults]. We highlight those aspects of MERA, which are key for appreciating the connection with kinematic space. A reader familiar with MERA may skip over to Sec. \[sec:therelation\]. ### Tensor network generalities The wavefunction of a general $N$-body system defines a tensor with $N$ indices: $$| \Psi \rangle = \sum_{i_1i_2\ldots i_N} \Psi_{i_1i_2\ldots i_N}|i_1 i_2 \ldots i_N\rangle \label{genwavefn}$$ The walloping number of components of this tensor –exponential in $N$– reflects the complexity of an arbitrary many-body wavefunction. However, imposing physical constraints such as locality and symmetry ought to simplify the description of the wavefunction drastically. This simplification is the objective of tensor network techniques. Tensor networks are graphs, which consist of vertices and edges. Every vertex stands for a tensor with as many indices, as there are edges incident on it. The indices range from 1 to $\chi$, the ‘bond dimension’ of a given edge. An edge connecting two vertices denotes a common index of two tensors, which is contracted (traced out.) Some examples of tensor networks, including the featureless wavefunction from eq. (\[genwavefn\]), are shown in Fig. \[fig:tensors\]. ![Examples of tensor networks. (a) A featureless tensor network composed of a single tensor. This can prepare a generic state, as in eq. (\[genwavefn\]). (b) A tensor network composed of a chain of tensors contracted together (a matrix product state). (c) The unitary (resp. isometric) character of the disentanglers and isometries in MERA means that these tensors cancel out when contracted with their hermitian conjugates.[]{data-label="fig:tensors"}](Images/tensors.pdf){width="95.00000%"} ### Structure of the MERA network {#merastructure} The MERA network is a successful ansatz for the ground state wavefunction of a conformal field theory [@mera-revresults]. For a CFT$_2$—the case of interest in the present paper—it is a two-dimensional array of tensors shown in Fig. \[fig:mera\]. While the horizontal direction corresponds to the spatial axis of the CFT, the vertical direction is meant to encode scale (RG direction). In a true CFT, which has no characteristic scale, the vertical direction ought to be infinite. In practice, however, MERA networks are presented with a finite number of layers, which is tantamount to fixing a UV cutoff. Because the axes of MERA correspond to space and scale, the network provides a graphical representation of renormalization in real space. For example, cutting the network one layer higher takes the wavefunction at scale $\mu$ to the wavefuction at the coarser scale $2\mu$. More generally, cutting the network in an inhomogeneous way can be understood as enacting a local scale transformation [@quotientmera]. In this way, the hierarchical structure of MERA encodes an iterative application of local coarse-graining transformations. To understand the rationale underlying the MERA ansatz, it is useful to examine a single layer of the network and ask how it is intended to coarse-grain the wavefunction. #### Disentanglers and isometries A layer of MERA consists of two types of tensors laid out in two rows. The tensors with four legs are called disentanglers. They are $\chi^2 \times \chi^2$ unitary transformations, which select bases wherein incoming UV degrees of freedom will appear locally unentangled. This change of basis is performed in order to prevent UV entanglement from accumulating in the IR wavefunctions defined on higher cuts. In this way, through the action of disentanglers, the MERA network partitions entanglement entropies of intervals into scale-specific contributions. ![The MERA lattices for states on a line and a circle.[]{data-label="fig:mera"}](Images/mera_flat.pdf "fig:"){width="60.00000%"} ![The MERA lattices for states on a line and a circle.[]{data-label="fig:mera"}](Images/mera.pdf "fig:"){width="34.00000%"} The second component in every layer of MERA is a row of isometries. These project the locally disentangled UV degrees of freedom into the effective IR Hilbert space. This transformation is isometric, which means that it can be extended to a unitary map $$\mathcal{H}_{\rm UV} \to \mathcal{H}_{\rm IR} \otimes \mathcal{H}_{\rm frozen}\,, \label{frozen}$$ with $\mathcal{H}_{\rm frozen}$ not participating in further steps of the renormalization [@mera2; @xiaoliang]. Our diagrams ignore $\mathcal{H}_{\rm frozen}$, showing isometries as maps from two UV lines (a $\chi^2$-dimensional vector space) to a single IR line (a $\chi$-dimensional vector space.) Of course, not every wavefunction can be prepared with this ansatz. This is the price we pay for efficiency—by varying the tensors in this fixed network, we scan an $\mathcal{O}(\chi^4 N)$-dimensional corner of the full Hilbert space, which we hope includes the ground state wavefunction. This hope has been validated in numerous computations, with the optimized MERA (the state of lowest energy in the variational class [@mera-opt]) correctly reproducing the spectrum and OPE coefficients of CFT$_2$s such as the critical Ising model [@mera-revresults; @mera-expl; @mera-moredata]. ### Causal structure {#MERAcausality} This fundamental feature of the MERA network, noticed and exploited already in the initial papers on the subject [@mera2; @mera-expl], offers the first hint of a relation to kinematic space. A prescient proposal relating MERA to de Sitter space appeared in [@beny]. Consider the reduced density matrix of an interval $\mathcal{I}$ in a pure state $|\Psi\rangle$: $$\rho_\mathcal{I}={\rm Tr}_{\mathcal{I}^c}|\Psi\rangle\langle\Psi | \, . \label{rhoi}$$ In the language of MERA, we compute it by putting together the tensor network representations of the bra and ket states and joining (tracing out) indices not contained in $\mathcal{I}$. Tracing out these indices means that disentanglers from the $|\Psi\rangle$ network get contracted with their hermitian conjugates from the $\langle\Psi |$ network and cancel out (compare with Fig. \[fig:tensors\](c)). A similar cancellation occurs in the isometries above them, then in the next row of disentanglers, and so on. The ensuing cascade of cancellations divides the network into two parts: the region that determines $\rho_\mathcal{I}$, and the region that drops out from it. In analogy with the propagation of signals in a Lorentzian spacetime, we call the former region the ‘inclusive causal cone’ of interval $\mathcal{I}$. ![The MERA causal structure. The inclusive causal cone of eight terminal sites is indicated by the expanding yellow region of the network. The exclusive causal cone of the same eight UV sites is indicated by the contracting orange region of the network. The tensors in this region perform a change of basis, which takes the state living on the five sites on the lightlike cut to the eight sites in the UV.[]{data-label="fig:cone"}](Images/mera_cone.pdf){width="75.00000%"} We shall see in a moment that this notion of ‘causality’ is the same as in Sec. \[kincausality\]. Before explaining this, let us consider $\rho_{\mathcal{I}^c}$, the reduced density matrix of the complement of $\mathcal{I}$. It too splits up the MERA network into two regions – the inclusive causal cone of $\mathcal{I}^c$ and the rest. Altogether, the division of the Hilbert space into localized tensor factors $$\mathcal{H} = \mathcal{H}_\mathcal{I} \otimes \mathcal{H}_{\mathcal{I}^c} \label{factors}$$ partitions the MERA network into three components: a region that only affects $\mathcal{I}$, an analogous region for $\mathcal{I}^c$, and a region that affects the reduced states of both. This division is shown in Fig. \[fig:cone\]. For obvious reasons, the first two regions are often called the exclusive causal cones of their respective intervals. ![The reduced density matrix $\rho_A$ for an interval $A$ in the vacuum can be represented as a tensor network. This network is obtained by taking two copies of MERA, then tracing over $A^c$. This causes a cascade of cancellations of disentanglers and isometries. The remaining tensors are only those in the inclusive causal cone of $A$.[]{data-label="fig:density-matrix"}](Images/density-matrix.pdf){width="75.00000%"} #### Lightlike coordinates To keep the nomenclature consistent, we ought to call the boundaries of the aforementioned regions ‘lightlike.’ These lightlike directions – one left-going and one right-going – are linear combinations of the two axes of MERA: $${\rm lightlike} = {\rm location}\pm {\rm scale} \quad \leftrightarrow \quad u, v~\textrm{from eq. }(\ref{nullcoords}) \, . \label{MERAcoords}$$ As in kinematic space, it is convenient to use them as coordinates on MERA. Doing so canonically assigns an interval to every tensor in the network. Specifically, a tensor at lightlike coordinates $u$ and $v$ is the topmost component of the exclusive causal cone of interval $(u,v)$. Notice that the notion of causality defined by eq. (\[MERAcoords\]) in MERA is exactly the same as that in Sec. \[kincausality\]: if a tensor at $(u_2, v_2)$ is in the MERA-future (past) of the tensor at $(u_1, v_1)$, the corresponding interval contains (is contained in) its counterpart. The privileged role of the lightlike directions in MERA is a consequence of working with unitary tensors; without unitarity, cancellations discussed below eq. (\[rhoi\]) would not occur and all parts of the network would affect $\rho_{\mathcal{I}}$ and $\rho_{\mathcal{I}^c}$. This marriage of unitarity and causality is displayed by the exclusive causal cones of intervals, whose role amounts to a change of basis. Observe that the action of tensors in the exclusive causal cone of $\mathcal{I}$ is undetectable by observables in $\mathcal{I}^c$, so it is a transformation within $\mathcal{H}_\mathcal{I}$. After concatenating with the exclusive causal cone, the rather abstract state defined on its lightlike edges is mapped into a local basis of $\mathcal{H}_\mathcal{I}$. Although the linear map effected by the exclusive causal cone is an isometric embedding of a smaller Hilbert space in a larger one, when the frozen degrees of freedom from eq. (\[frozen\]) are taken into account, it is manifestly unitary. ### Entanglement entropies from cut-counting {#cutcount} A central motif of the present work—and one that motivated holographers’ initial interest in MERA [@briansessay]—is the simple way the network encodes entanglement entropies. For intervals of less than half system size, a good estimate is obtained by counting the number of lines emanating from the exclusive causal cone of the interval. If each line is counted with weight $\log \chi$, this amounts to computing the logarithm of the dimension of the Hilbert space living on the edge of the exclusive causal cone; see Fig. \[fig:cutcounting\]. In what follows, we will refer to this edge as the ‘causal cut,’ though the term ‘minimal curve’ has been used in prior literature [@briansessay; @brianspaper]. On the one hand, the cut-counting prescription gives a manifest upper bound on the entanglement entropy. We saw in Sec. \[MERAcausality\] that the spectrum of the reduced density matrix of the interval is prepared above the causal cut. The tensors below the cut merely choose a basis in which the state is expressed and therefore have no effect on the entanglement entropy. The maximal value of the entanglement entropy is the logarithm of the dimension of the Hilbert space, in which the state prepared by the network lives. This is precisely what the cut-counting prescription computes. ![Up to an overall coefficient, the entanglement entropy of a CFT interval $A$ can be estimated by the number of network lines crossing the exclusive causal cone of $A$ times the logarithm of the bond dimension $\log \chi$.[]{data-label="fig:cutcounting"}](Images/cut-counting.pdf){width="70.00000%"} Though there is no similar argument bounding the entanglement entropy from below, in practice counting cuts gives a good estimate of the entanglement entropy up to a multiplicative constant. In the vacuum, we can surmise this from the logarithmic scaling of entanglement entropy with interval size [@cardycalabrese], which agrees with the number of cuts. This heuristic reasoning was verified numerically in [@mera; @mera2; @mera-expl]. More formal evidence was given in [@quotientmera], which confirmed that the state on the exclusive causal cones of complementary semi-infinite lines has a thermal entaglement spectrum (see Sec. \[sec:BHmera\] for a more detailed summary). In light of this fact, the proportionality of entanglement entropy and the number of cuts is equivalent to the extensivity of thermal entropy. For any state built from networks with suitably generic tensors, the proportionality of entanglement entropy and the number of cuts will be established in upcoming work [@patrickmichael], which draws on Page’s theorem [@donpage]. In the present discussion of the MERA network, we treat the cut-counting prescription as an empirical fact. The reader should remember, however, that no fundamental principle protects this relation and it cannot be expected to hold in full generality. A case in point [@glenprivate] is the minimally updated MERA network [@minupdates], which models a CFT in the presence of an impurity. In such circumstances, any connection between MERA and holography will involve the incremental entanglement entropy per bond instead of a na[ï]{}ve count of bonds. MERA and Kinematic Geometry {#sec:therelation} =========================== Kinematic space encodes the data about CFT subsystems in an elegant geometric way. CFT intervals are organized by location and scale in a Lorentzian space whose metric structure is supplied by conditional mutual information. An analogous representation of CFT subsystem data is given by the MERA network whose tensors are canonically associated with contiguous collections of UV sites. In this section, we outline a series of commonalities that motivate the identification of the two structures. In particular, we propose to view MERA as a discrete counterpart of kinematic space. Our proposal to associate the MERA network with kinematic space runs contrary to a long-held belief that MERA ought to discretize a spatial slice of the bulk geometry. This idea, first put forward by Swingle [@briansessay; @brianspaper], gave the impetus to the prolific program of investigating tensor networks vis-[à]{}-vis holographic duality [@tadashisnetwork; @hartmanmaldacena; @xiaoliang; @shocks; @errorcorrectingnetwork; @donspaper; @patrickxiaoliang], of which the present paper is a part. It is, thus, worthwhile to contrast our novel kinematic proposal with ‘the traditional view’ of MERA as a discretized spatial geometry. In the discussion to follow, we comment on the conceptual drawbacks of a direct connection to the bulk, which are manifestly absent from the kinematic space perspective.[^2] Partial order of MERA and kinematic causality --------------------------------------------- The space of geodesics is a partially ordered set. This is an intrinsic property of kinematic space that follows from the containment relation of their boundary support—a property that is invariant under symmetry transformations. The signature of the kinematic metric is the geometric reflection of this structure. The same applies to MERA: the tensors in the network are partially ordered with respect to their domains of influence. The locality of the tensor contractions, which is built into the skeleton of the network, makes each tensor capable of affecting only a subset of the spatial degrees of freedom. This immediately induces a hierarchy among them in that the regions affected by certain tensors are strictly enclosed within the domain of other tensors’ influence. This property of MERA makes no reference to a UV cutoff. Moreover, the unitarity of the tensors promotes this ordering to a true notion of causality: Not only do individual tensors affect the state of well-defined spatial intervals, but also the state on given intervals is influenced only by specific network subregions. We can, therefore, draw light-like directions which restrict the propagation of information in the network. We observe that the two notions of causality—network and kinematic—coincide. This structure is absent from the hyperbolic plane, all points in which are treated on equal footing. Only upon introducing a cutoff can points on $\mathbb{H}_2$ be partially ordered with reference to their distance from the boundary. We shall see that this structural difference has interesting consequences. ### Spacelike versus timelike paths An immediate consequence of the Lorentzian signature of kinematic space is a qualitative distinction between kinematic paths that are *spacelike*, *null* or *timelike*. This classification is robust under the action of symmetries and suggests that only certain types of curves, i.e. spacelike, can be used as good kinematic cutoff surfaces. The stipulation that cutoffs must not be timelike is evident in the holographic view of kinematic space. The reason is illustrated in Fig. \[fig:KScutoff\]. A spacelike kinematic trajectory selects a family of geodesics, which has a well-defined outer envelope in the bulk. This envelope acts as a (diffeomorphism invariant) cutoff surface in the spatial geometry. But whenever a trajectory in kinematic space becomes timelike, the bulk cutoff surface is no longer defined [@robfirst]. ![Spacelike (left) and timelike (right) curves in kinematic space as seen from the bulk point of view.[]{data-label="fig:KScutoff"}](Images/KSspacelike.pdf "fig:"){width="47.00000%"} ![Spacelike (left) and timelike (right) curves in kinematic space as seen from the bulk point of view.[]{data-label="fig:KScutoff"}](Images/KStimelike.pdf "fig:"){width="47.00000%"} Tensor networks, on the other hand, prepare wavefunctions on selected cutoff surfaces in the network. The causal structure of MERA, however, makes some cutoffs inadmissible: The cut on which the state is defined must be, like in kinematic space, piecewise spacelike or lightlike and never timelike [@quotientmera]. A MERA network ending on a locally timelike cut cannot be associated to a coarse-grained version of the vacuum wavefunction. The failure of MERA to prepare the wavefunction on a timelike cutoff surface is analogous to the failure of timelike-separated (in kinematic space) geodesics to form a curve in AdS. The Euclidean signature of a spatial AdS slice is compatible with any convex bulk cutoff surface [@AdSMERAtakayanagi]. The mixed signature of MERA is therefore in obvious tension with a direct AdS/MERA connection. This is best seen from the simple example of a cutoff that follows a radial geodesic in $\mathbb{H}_2$. Symmetry demands that such a cutoff surface, although geometrically admissible in the bulk, map to a vertical (and therefore timelike) cut in MERA. This conclusion can only be averted by introducing extra assumptions into a putative MERA/AdS correspondence. ### Representation of symmetries In studying MERA representations of CFT states, an important question concerns the action of conformal symmetry on the tensor network. Because the choice of a cutoff surface on which the state is defined breaks conformal invariance, the cutoff transforms under the conformal group. In two dimensions, conformal symmetry acts locally and can reset the cutoff to an arbitrary function of position.[^3] #### Local conformal transformations The primary focus of the discussion in [@quotientmera] were local conformal transformations in MERA. A conformally transformed wavefunction was recognized as the state living on an inhomogeneous cut in the network. In this way, conformal maps in MERA are implemented by locally changing the cut on which the wavefunction is defined. Importantly, this operation does not affect the rest of the network away from the UV cut. In particular, the causal structure of the MERA network is fixed and independent of conformal transformations. To summarize, the action of the conformal group in MERA can take a uniform UV cutoff to some other, inhomogeneous cutoff, but without affecting the null directions. Consistency then requires that conformal maps take *spacelike* cutoffs to other *spacelike* cutoffs. Such a constraint is guaranteed when MERA is associated with the kinematic geometry. Interpreted in the bulk, however, this seems to impose an artificial restriction on the set of allowed (or MERA-representable) cutoff surfaces: they can never become approximately radial. Since conformal symmetry transforms radial and other bulk surfaces into one another, such a limitation would be a radical breaking of conformal symmetry. #### Inhomogeneity of a causal cut Causal cuts in MERA are not homogeneous. Their lightlike segments are uniform, but the top of a causal cut where left-going and right-going cuts meet is distinct from the rest. The non-uniform shape of a causal cut in MERA is readily understood in the kinematic interpretation. The top corresponds to the geodesic $g$ supported on the base of the chosen lightcone while other points on a causal cut correspond to narrower geodesics that share one endpoint with $g$ and are otherwise contained within it (compare e.g. Fig. 1 in [@protocol]). The insensitivity of both the MERA and kinematic partial order to the UV-cutoff ensures that this point will remain special under local conformal maps. The AdS isometries, on the other hand, map different points on the same geodesic to one another. In other words, geodesics are homogeneous, a fact that forbids special points. When identifying the causal cut with an AdS geodesic -as the direct AdS/MERA connection suggests- one might try to assuage this discrepancy by declaring that the special point on a cut in MERA corresponds to some select point on a bulk geodesic, chosen according to some prescription. Any such prescription, however, must refer to a UV cutoff; in the absence of a UV cutoff there is no reference with respect to which a special point may be chosen. Because conformal symmetry acts on the cutoff, it must also affect the choice of a preferred point on a geodesic. Yet in MERA, the top of a causal cut is fixed, its location blind to any changes in the cutoff. This reveals that the conformally invariant notion of causality in MERA disfavors a na[ï]{}ve partial ordering of the hyperbolic plane induced by a UV cutoff. But it is in full agreement with the causal structure of kinematic space, which is likewise conformally invariant. Localization of information --------------------------- ### Crofton form and volumes in MERA {#meravolume} In Sec. \[kincmi\], we observed that the notion of volume of kinematic space (eq. \[kinvolume\]) hails from information theory: it is the conditional mutual information (\[defcmi\]) of three contiguous intervals. Let us inspect the same quantity in MERA. #### Conditional mutual information localizes in MERA When we apply the cut-counting prescription reviewed in Sec. \[cutcount\] to $$I(A, C|B) = S(AB) + S(BC) - S(ABC) - S(B)\,, \tag{\ref{defcmi}}$$ we obtain Fig. \[fig:meraCMI\]. The cuts associated with the positive terms in (\[defcmi\]) are in large part the same as the cuts for the negative terms, leading to cancellations. The net result comes from a localized part of the network, whose boundaries are lightlike. In other words, the conditional mutual information of neighboring intervals localizes in a causal diamond. For intervals with endpoints at $$A = (u - \Delta u, u) \qquad {\rm and} \qquad B = (u, v) \qquad {\rm and} \qquad C = (v, v + \Delta v)\,, \label{3INT}$$ the relevant causal diamond resides between $u$ and $u-\Delta u$ in the left-moving coordinate and between $v$ and $v + \Delta v$ for the right-moving one. ![Conditional Mutual Information in MERA. Most cuts that contribute to the computation of $I(A,C|B)$ cancel in the alternating sum. The net contribution to the conditional mutual information arises from a localized region of the network.[]{data-label="fig:meraCMI"}](Images/localization.pdf){width="1.\textwidth"} In the end, the entire MERA network is a tilted chessboard of causal diamonds, each of which computes some conditional mutual information. The grid of lightlike coordinates demarcates conditional mutual informations of different triples of intervals. Fig. \[fig:meratiles\], which displays these facts, is a faithful copy of Fig. \[fig:KSmutual\], which highlights the analogous characteristics of kinematic space. #### What does conditional mutual information count? Figs. \[fig:meraCMI\] and \[fig:meratiles\] give a crisp answer: conditional mutual information counts how many isometries live in the appropriate causal diamond. Eq. (\[defcmi\]) asks for the net reduction in the number of lines passing through the causal diamond as we go from the bottom up. The only way we can register a net loss of lines is if a line is soaked up by an isometry. Indeed, every isometry accounts for precisely one line, which enters the diamond from the bottom but does not emerge at the top. Counting the decrease in the number of lines is equivalent to counting isometries. #### Conditional mutual information as volume We propose to adopt conditional mutual information as a definition of volume in MERA. The two facts highlighted above guarantee that this is a reasonable proposal: conditional mutual information localizes in MERA and counts a crisply defined object—the isometries contained in a causal diamond. In other words, we observe that for $A, B, C$ defined in eq. (\[3INT\]): $$\mathcal{D}\textrm{(isometries)} = I(A, C | B) \,. \label{disometries}$$ We declare this quantity a discrete volume form, in analogy to eq. (\[kinvolume\]) in kinematic space. In the upcoming second part of this work [@compression], where we discuss our more general compression networks, we will appreciate better the rationale for working with eq. (\[disometries\]). The volume of a causal diamond computed by (\[disometries\]) evaluates the amount by which the tensors in the diamond compress the state living on its past edges. This is how eq. (\[disometries\]) should be viewed in applications beyond the standard MERA. In the special case of the vacuum MERA, this ‘density of compression’ is directly proportional to a na[ï]{}ve count of isometries. We give a short summary of the compression networks in Sec. \[compressionsummary\], referring to [@compression] for details. It is worth noting that in the traditional holographic view of MERA the connection between conditional mutual information and localized volumes of the network is puzzling. If we represent the terms in eq. (\[defcmi\]) by geodesics in the bulk, no such localization occurs. Instead, the calculation involves an extended region in the spatial geometry, which reaches all the way to the asymptotic boundary. When $A$ and $C$ are taken to be small as in eq. (\[3INT\]), the bulk region associated with $I(A, C | B)$ becomes a fattened geodesic subtending $B$. In MERA this limit shrinks the relevant causal diamond to a small number of tensors. This again motivates relating small regions in MERA to bulk geodesics. ![Localization of mutual information in MERA. We indicate the local regions of the network that control the computation of the mutual information of two neighboring intervals, and the conditional mutual information of three neighboring intervals.[]{data-label="fig:meratiles"}](Images/meratiles.pdf){width=".7\textwidth"} #### A metric for MERA In Sec. \[kincausality\], we assembled the kinematic metric (\[kinmetric\]) from two ingredients: the causal structure (eq. \[nullcoords\]) and the volume form (\[kinvolume\]). In Sec. \[MERAcausality\] we recognized that MERA has an identical causal structure. Now eq. (\[disometries\]) gives us a notion of volume, which is a direct analogue of eq. (\[kinvolume\]). These reasons justify conceptualizing MERA as a discrete version of kinematic space. More explicitly, we may write down a discrete tensor network metric $$ds^2_{\rm T.N.} = I(\Delta u, \Delta v | B) \,\,\stackrel{\rm MERA}{\xrightarrow{\hspace*{12mm}}}\,\, (\# {\rm isometries}) \, \Delta u \, \Delta v \label{ds2tn}$$ which in the case of the familiar MERA simply counts isometries in causal diamonds. This metric is the obvious counterpart to eq. (\[kinmetric\]) in kinematic space. #### Differential entropy and cut-counting in MERA One attractive feature of kinematic space is that volumes in it reproduce the differential entropy formula [@holeography]; see eq. (\[sdiff\]). Metric (\[ds2tn\]) ought to give rise to a similar relation in MERA. Indeed, any spacelike cut across MERA defines a (possibly non-uniform) UV cutoff and a coarse-grained Hilbert space; see Sec. \[merastructure\]. The logarithm of the dimension of that Hilbert space is proportional to the number of indices living on the cut. Because every line ends on some isometry in the UV part of the network, the logarithm of the dimension of the Hilbert space defined by a cut is equal to the volume of MERA living above that cut, counted according to eq. (\[disometries\]). The equality between the ‘volume’ of a subregion of MERA and the number of lines on its boundary follows from a discrete version of Stokes’ theorem. This argument is an exact analogue of the reasoning articulated below eq. (\[sdiff\]). Thus, computing the size of a coarse-grained Hilbert space by counting indices on its defining cut is the MERA version of the differential entropy formula. As a special case, this recovers the cut-counting prescription for entanglement entropy, which we revisit in Sec. \[entanglementMERA\]. More generally, counting cuts assigns an entropic quantity to any (possibly non-uniform) spacelike UV cutoff, which in the bulk is represented by a collection of tangent geodesics (see Fig. \[fig:KScutoff\]). ### Entanglement entropy {#entanglementMERA} The feature of MERA that makes it especially relevant for holography is the way it geometrizes entanglement entropies. In Sec. \[cutcount\] we reviewed the cut-counting prescription in MERA: estimating the entanglement entropy of an interval by counting the lines which cross the causal cut. This special property of the optimized network was used in Sec. \[meravolume\] to place a physical metric on MERA (eq. \[ds2tn\]) and recognize it as a faithful representation of the kinematic geometry. ![The entanglement entropy of $A$ is given by half of the total kinematic volume of the ‘causal wings’ depicted in the figure.[]{data-label="fig:KSentropy"}](Images/KSentropy.pdf){width="57.00000%"} If MERA discretizes the kinematic space, however, the causal cut in MERA becomes a discrete version of the kinematic causal cone. The integral geometric computation of entanglement entropies then ought to be consistent with the cut-counting prescription along this causal cut. Recall that in a pure state, the entanglement entropy of an interval $\mathcal{I}$ is half the mutual information of $\mathcal{I}$ and its complement, $\mathcal{I}^c$: $$S(\mathcal{I}) = \frac{1}{2}\, I(\mathcal{I}, \mathcal{I}^c) \label{eecmi}$$ As we saw in Sec. \[kincmi\], the mutual information of two adjacent intervals can be read off from kinematic space as the volume of a causal diamond, which includes the common endpoint of both intervals. In the case at hand, we actually have two causal diamonds, because $\mathcal{I}$ and $\mathcal{I}^c$ have two endpoints in common (see Fig. \[fig:KSentropy\].) We may use eq. (\[sdiff\]) and convert the volume of the two causal diamonds to a differential entropy, that is a one-dimensional integral over the causal cut: $$S(u,v) = \frac{1}{2} \int_u^v d\tilde{v}\,\,\frac{\partial S(u,\tilde{v})}{\partial \tilde{v}} + \frac{1}{2} \int_v^u d\tilde{u}\,\,\frac{\partial S(\tilde{u},v)}{\partial \tilde{u}} \label{eece}$$ The two terms in this formula come from the two causal diamonds in Fig. \[fig:KSentropy\]. In the context of MERA, their integrands become densities of lines that cross the causal cut (Sec. \[meravolume\]). This is precisely what the cut-counting prescription mandates. ![The entanglement entropy of a CFT interval $A$ is computed by the number of network isometries responsible for correlating $A$ with its complement. This is consistent with the integral geometric computation of the corresponding geodesic length (Fig. \[fig:KSentropy\]). An application of the discrete Stokes’ theorem recovers the cut-counting prescription of Sec. \[cutcount\]. The figure shows the network volume relevant for the EE on the line (LEFT) and the circle (RIGHT). It contains all tensors in the inclusive causal cones of the two endpoints of $A$, excluding their intersection. This network region prepares the entanglement spectrum between $A$ and $A^c$.[]{data-label="fig:meraEE"}](Images/meraEE.pdf "fig:"){width="57.00000%"} ![The entanglement entropy of a CFT interval $A$ is computed by the number of network isometries responsible for correlating $A$ with its complement. This is consistent with the integral geometric computation of the corresponding geodesic length (Fig. \[fig:KSentropy\]). An application of the discrete Stokes’ theorem recovers the cut-counting prescription of Sec. \[cutcount\]. The figure shows the network volume relevant for the EE on the line (LEFT) and the circle (RIGHT). It contains all tensors in the inclusive causal cones of the two endpoints of $A$, excluding their intersection. This network region prepares the entanglement spectrum between $A$ and $A^c$.[]{data-label="fig:meraEE"}](Images/EEcircle.pdf "fig:"){width="40.00000%"} To see this more clearly, consult Fig. \[fig:meraEE\], which is the MERA analogue of Fig. \[fig:KSentropy\]. Taking advantage of the localization of mutual information in MERA illustrated in Fig. \[fig:meratiles\], we know that the entanglement entropy of $A$ is given by the number of isometries living in the highlighted part of the network. But the same count yields the number of lines crossing the causal cut of $A$. In effect, we are learning that the cut-counting prescription of Sec. \[cutcount\] secretly enumerates the isometries responsible for correlating the interval with its complement. This is in direct analogy with the way kinematic space encodes the length of a Ryu-Takayanagi geodesic as the ‘number of geodesics’ connecting the boundary interval with its complement. The counting of geodesics is done with the Crofton measure, which is a geometric counterpart of the density of isometries in MERA. In contrast, if we place the network directly on the time-slice of AdS, the relation between volumes of ‘causal wings’ and entanglement entropies appears mysterious. It seems to imply that a special region in the bulk—which lacks an independent motivation in the AdS/CFT correspondence—quantifies the correlation between a given interval and its complement in terms of its volume; see Fig. \[fig:meraEE\]. Insisting on an AdS/MERA correspondence appears to add another peculiar property to its putative dictionary, a peculiarity that is readily resolved by the kinematic proposal. MERA as renormalization {#merarg} ----------------------- ### Coarse-graining with MERA {#coarsemera} As we reviewed in Sec. \[merastructure\], MERA provides a graphical representation of renormalization in real space. The vertical direction corresponds to scale in the field theory. Cutting MERA on different levels defines states, which are related to one another by coarse-graining or fine-graining. As we go higher up in MERA, the successively coarse-grained states live in Hilbert spaces of exponentially decreasing sizes (entropies). The same features are observed in kinematic space; its identification with the space of CFT intervals makes it a natural domain for real space cutoffs. The two coordinates of kinematic space, $\theta$ and $\alpha$ (see eq. \[nullcoords\]) also correspond to location and scale. The role of $\alpha$ as setting a scale is evident from its definition as the half-width of a field theory interval. Cutting off kinematic space at $\alpha = \alpha_*$ imposes a real space cutoff—it amounts to declaring $2\alpha_*$ to be the smallest resolution in the field theory. The spatial size of a cutoff surface in kinematic space also varies exponentially with the cutoff; in the vacuum on a circle, metric (\[kinmetric\]) expressed in terms of $\theta$ and $\tilde\rho = -\log (\csc\alpha + \cot\alpha)$ is: $$ds_{\rm kin}^2 = \frac{c}{3}\, (-d\tilde\rho^2 + \cosh^2\tilde\rho \, d\theta^2)\,. \label{metrickin}$$ Holographically, every real space cutoff defined by a curve in kinematic space selects a set of bulk geodesics. These in turn identify a bulk cutoff surface by their outer envelope as we illustrated in Fig. \[fig:KScutoff\]. This proposal for the holographic cutoff has the appealing feature that it is manifestly diffeomorphism invariant, because it is implemented on bulk geodesics that make no reference to AdS coordinate systems. Interestingly, the kinematic cutoff can be further promoted to a precise coarse-graining prescription for CFT operators, which exploits the structure of the operator product expansion (OPE). A detailed discussion of this point will be presented in [@kinematicoperators], where we formulate 1+1-dimensional CFTs in the language of kinematic space and derive the connection to the effective field theory in the bulk from first AdS/CFT principles. On a spatial slice of the bulk geometry, the radial direction $\rho$ is also dual to changes of scale in the CFT [@uvir1; @uvir2; @uvir3]. Regulating large scale divergences on the gravity side with a radial cutoff $\rho = \rho_*$ is dual to selecting an ultraviolet cutoff in the CFT. When we push the radial cutoff $\rho_*$ to infinity, the area of the cutoff surface grows exponentially. This is captured by the spatial metric: $$ds^2 = L^2 (d\rho^2 + \sinh^2\rho\, d\theta^2) \label{metricspatial}$$ The interpretation of MERA as a real space RG transformation, however, can be leveraged to distinguish between the two types of geometric coarse-graining suggested above. As we explain in the next section, the constraints that causality imposes on the RG operation of MERA act in favor of the kinematic proposal. ### Real space RG and causal cuts {#rgcausalcuts} Consider two points on the 1-D boundary slice where the CFT state lives. Any such choice splits the CFT into two regions: an interval $A$ and its complement $A^c$. In a pure state such as the vacuum the entanglement entropies $S(A)$ and $S(A^c)$ are equal. This fact is nicely captured by the Ryu-Takayangi proposal: a unique minimal geodesic homologous to both $A$ and $A^c$ joins the two boundary points. In MERA, for any selection of two spatial points there are two distinct causal cuts in the network, which bound the exclusive causal cones of $A$ and $A^c$, respectively (see Fig. \[fig:radial-cut\]). The two cuts typically do not cross the same number of links; only the minimal one has the correct count of links to match the entanglement entropy. Nevertheless, both lightcones are physically meaningful. In the view of MERA as a real space RG transformation, every local application of disentanglers and isometries performs a local coarse-graining of the wavefunction. Such coarse-grainings can be understood as a change of basis, but only if the cutoff surface is piecewise spacelike or null at every RG step. In this way, we obtain an upper bound for the allowed local coarse-graining of an interval. The two causal cuts encode the maximally coarse-grained state of $A$ and $A^c$, respectively. ![Causal cuts for a region $A$ and its complement $A^c$. For both MERA on the line and circle, the causal cuts for $A$ and its complement $A^c$ are distinct.[]{data-label="fig:radial-cut"}](Images/lineradial.pdf "fig:"){width="53.00000%"} ![Causal cuts for a region $A$ and its complement $A^c$. For both MERA on the line and circle, the causal cuts for $A$ and its complement $A^c$ are distinct.[]{data-label="fig:radial-cut"}](Images/radialcut.pdf "fig:"){width="40.00000%"} To appreciate the importance of the two distinct MERA lightcones in a different way, observe that the tensors contained in the exclusive causal cone of an interval build local correlations among the interval’s lattice sites. This amounts to selecting a local basis for representing the state and leaves the entanglement spectrum unaffected. In other words, the entanglement spectrum of $A$ is solely prepared in the region between the two causal cones. This feature of MERA is naturally included in the kinematic proposal: the two complementary intervals possess distinct causal cones, each of which bounds the set of bulk geodesics anchored on the respective boundary region. There exists, moreover, a finite ‘volume’ of geodesics that connect the two intervals, a fact reflected by the separation of the two lightcones in kinematic space. When approached from the traditional AdS/MERA perspective, however, no meaningful geometric counterpart exists for the non-minimal causal cut. This contradicts the equal treatment of the two cuts in the network and seems to select a peculiar, IR-probing curve associated to the coarse-grained state of the larger interval. By the AdS/MERA interpretation, that coarse-grained state should have been instead associated to the minimal geodesic. MERA for boundary gravitons and two-sided black holes {#sec:BHmera} ===================================================== Boundary gravitons ------------------ Thus far, we have argued for identifying the vacuum MERA with the kinematic space of a time slice of pure AdS$_3$. This conclusion automatically extends to conformal descendants of the vacuum—states related to the vacuum by a local conformal transformation. In MERA, wavefunctions of such states can be read off from inhomogeneous UV cuts [@quotientmera]. In particular, going from the vacuum to a descendant does not change local properties of the network. On the bulk side, descendant states are represented by so-called boundary gravitons [@banados]. They are locally AdS$_3$ geometries, which differ from global AdS$_3$ by large diffeomorphisms. Importantly, a large diffeomorphism changes lengths of geodesics by two additive pieces, which carry no joint dependence on the two endpoints [@mandal; @joan]: $$S(u,v) \to S(u,v) + \Delta \mu(u) + \Delta \mu(v)$$ This change leaves kinematic volumes (\[kinvolume\]) invariant. We reach the same conclusion by noting that a boundary conformal transformation that preserves a time slice of the CFT maps $x \to \tilde{x} = f(x)$. Applying this transformation to $u$ and $v$ in the kinematic metric (\[kinmetric\]) gives: $$ds_{\rm kin}^2= \frac{\partial^2 S(u,v)}{\partial u \,\partial v}\, du\,dv = \frac{\partial^2 S(\tilde{u},\tilde{v})}{\partial \tilde{u} \,\partial \tilde{v}}\, d\tilde{u}\,d\tilde{v}$$ This illustrates that the kinematic space defined in eq. (\[kinmetric\]) is invariant under all conformal transformations which preserve a time slice of the CFT. The only dependence on the conformal frame is introduced by the UV cutoff. The thermofield double state and the two-sided BTZ black hole ------------------------------------------------------------- ![The quotient of the optimized vacuum MERA, which prepares the thermofield double state. In different parts of the network the identifications act in a timelike, lightlike and spacelike manner, respectively. The network displayed here identifies lines that are $k=2$ layers apart, so $s = 2^k = 2^2$ in eq. (\[ltobeta\]). The figure is reproduced with permission from [@quotientmera].[]{data-label="quotientTN"}](Images/quotient-KS1.pdf){width=".7\textwidth"} #### The thermofield double state in the CFT and in MERA For a non-trivial application of local conformal transformations, consider a map that acts not on the full line $\mathbb{R}$, but the line minus a point, $\mathbb{R}\setminus\{0\}$. The two semi-infinite lines on either side of the excluded point can each be mapped to an infinite line by the logarithmic map: $$x \to (\beta / 2 \pi) \log |x| \label{logmap}$$ In this way, we view the vacuum on $\mathbb{R}$ as an entangled state on $\mathbb{R} \times \mathbb{R}$. Famously, in a conformal field theory this entangled state is the thermofield double state. One may further quotient the two $\mathbb{R}$s by a discrete translation $\log x \sim \log x + \log s$ to obtain the thermofield double state on $S^1 \times S^1$. The circumference of the $S^1$ sets a natural scale, in which to express the otherwise dimensionless inverse temperature: $$(\beta / 2\pi) \log s \equiv 2 \pi L \label{defs}$$ In Ref. [@quotientmera] our collaborators and we performed these operations in the optimized MERA network. The conformal transformation (\[logmap\]) was enacted by cutting the network along the two null rays emanating from $x=0$. After the cut, the quotient identifies identical pieces of the causal cone of the origin. Alternatively, we can apply the quotient prior to the conformal map (\[logmap\]). This produces the entire network shown in Fig. \[quotientTN\], including the regions living below the null rays. In this view, the two semi-infinite lines are modded out by a discrete scaling transformation $x \sim s\, x$. #### The BTZ black hole as a quotient of AdS$_3$ The field theory operations outlined above reflect a famous fact in 3-d gravity: that the two-sided BTZ black hole is a quotient of pure anti-de Sitter space [@btz]. Consider the Poincar[é]{}-AdS$_3$ metric restricted to $t=0$: $$ds^2 = \frac{dx^2 + dz^2}{z^2}$$ In order to quotient $\mathbb{R}_-$ and $\mathbb{R}_+$ by a discrete scale transformation, select a family of geodesics centered at $x = 0$ whose radii are related by powers of $s$: $$x^2 + z^2 = s^{2n} r^2 \qquad {\rm where}~n \in \mathbb{Z} \label{identgeos}$$ Identifying these geodesics with one another produces a topological cylinder, which is the static slice of the two-sided BTZ geometry; see Fig. \[BTZidents\]. This identification can be canonically extended away from the time slice to produce the full, 2+1-dimensional BTZ space-time with two asymptotic regions. The inverse temperature of the black hole in units of the AdS$_3$ curvature scale is given by eq. (\[defs\]): $$\beta / L = 4\pi^2 / \log s \label{ltobeta}$$ In the limit $n \to -\infty$, the geodesics in Fig. \[BTZidents\] zoom on a point on the boundary at $x = 0$. This location, which is a fixed point of the quotiented discrete scale transformation, separates the two semi-infinite lines into which the $x$-axis decomposes under map (\[logmap\]). After the quotient, every fundamental domain in $x > 0$ represents a copy of one asymptotic boundary while fundamental domains in $x < 0$ are copies of the other asymptotic boundary. In the bulk, the line $x = 0$ is also meaningful. It connects points of closest approach of identified geodesics and, therefore, comprises images of the bifurcation horizon. Many good reviews of these facts exist, including [@multibd]. The quotient MERA is the kinematic space of the two-sided black hole -------------------------------------------------------------------- Let us compare the tensor network shown in Fig. \[quotientTN\] with the space of geodesics on a static slice of the two-sided BTZ black hole. Due to the discrete nature of the tensor network we may only quotient MERA by discrete scalings with $s = 2^k$ for $k \in \mathbb{Z}_+$. #### Structure of the identifications Observe that in the tensor network quotient in Fig. \[quotientTN\] not all identifications of indices are on the same footing. In the middle of the network we connect lines which are timelike-separated in the MERA sense. In the UV the identification joins indices that are spacelike-separated. The network contains two such regions, one on each side of the thermofield double. Separating the spacelike-identified regions from the timelike-identified one are single lines of tensors, which after the quotient form closed lightlike curves. These distinct components of the thermofield double MERA correspond to analogously distinguished classes of geodesics in the two-sided BTZ geometry.[^4] We marked the three classes in Fig. \[BTZidents\]. ![Identifying concentric geodesics on a static slice of AdS$_3$ produces a static slice of the two-sided BTZ black hole. Three BTZ geodesics are shown, along with the AdS$_3$ geodesics from which they descend. The green spacelike-identified geodesic is anchored on a single boundary. The red lightlike-identified geodesic circles the horizon indefinitely. The blue timelike-identified geodesic connects the two sides.[]{data-label="BTZidents"}](Images/btzQuotient.pdf "fig:"){height="35.00000%"} ![Identifying concentric geodesics on a static slice of AdS$_3$ produces a static slice of the two-sided BTZ black hole. Three BTZ geodesics are shown, along with the AdS$_3$ geodesics from which they descend. The green spacelike-identified geodesic is anchored on a single boundary. The red lightlike-identified geodesic circles the horizon indefinitely. The blue timelike-identified geodesic connects the two sides.[]{data-label="BTZidents"}](Images/btzGeodesics.pdf "fig:"){height="35.00000%"} The first of these are the *timelike-identified geodesics*. A canonical example of these are the geodesics in eq. (\[identgeos\]), which define the geometric quotient in Fig. \[BTZidents\]. More generally, a geodesic becomes identified with a timelike-separated image of itself if one of its endpoints is negative ($u < 0$) while the other one is positive ($v > 0$). This means that timelike-identified geodesics connect opposite sides of the two-sided black hole. They are horizon-crossing geodesics. The *spacelike-identified geodesics* remain on one side of the horizon. Their endpoints are either both positive ($0 < u < v$) or both negative ($u < v < 0$). In the bulk, such geodesics do not reach the horizon. The marginal case separating the previous two are *lightlike-identified geodesics*. Recall that $u$ and $v$, the left and right endpoint of a geodesic, are lightlike coordinates in kinematic space. Thus, the geodesic $(u,v)$ is lightlike-separated from its scaled image $(su, sv)$ if and only if $u = su = 0$ or $v = sv = 0$. This is consistent with the scope of the timelike-identified ($u < 0 < v$) and spacelike-identified regions ($u < v < 0$ and $0 < u < v$). The lightlike-identified geodesics are the borderline case, which separates horizon-crossing geodesics from those which remain a finite distance apart from the horizon. They are tangent to the horizon. A boundary-anchored geodesic can only become tangent to the horizon after spiraling around it infinitely many times. The infinite winding of the lightlike-identified geodesics can be seen in Fig. \[BTZidents\]. In the covering space, such geodesics cross infinitely many copies of one type of asymptotic boundary. For a more extensive discussion of infinitely winding geodesics in the BTZ geometry, consult [@lampros]. ![The kinematic space for the two-sided BTZ black hole, to be compared with Figs. \[quotientTN\],\[BTZidents\]. The BTZ kinematic space is obtained as a quotient of the vacuum kinematic space, where two causal cuts are identified. A fundamental domain is labeled above, which separates into horizon-crossing, winding (entwinement), and minimal (entropy) geodesics. The points on the lightlike lines indicated correspond to infinitely winding geodesics.[]{data-label="btzKS"}](Images/btzKS.pdf){width=".7\textwidth"} #### Timelike-identified regions In Ref. [@quotientmera], our co-authors and we explained that the timelike-identified region of the quotient network prepares the spectrum of the thermofield double state. In other words, this region is solely responsible for fixing the correlations between the two sides. This is exactly what we expect from the kinematic interpretation of MERA, which relates this region of the network to geodesics that cross the horizon and connect the two asymptotic boundaries. As an example, such geodesics were used to compute two-sided correlators in the thermofield double state in [@excursions]. #### Black hole entropy Following Fig. \[fig:KSmutual\], the volume of the timelike-identified region in kinematic space computes the mutual information between the two sides: $$I(L, R) = 2 S_{\rm BH} \label{mithermal}$$ Referring to the Crofton formula, this equation states that the area of the black hole horizon counts the geodesics that cross the horizon and connect the two sides of the wormhole geometry. It is instructive to recover this result explicitly in kinematic space by the use of the differential entropy formula. The latter asks for a complete set of geodesics tangent to the horizon, i.e. the lightlike-identified geodesics. Thus, the contour of integration is one full closed lightlike curve in kinematic space, for example $s u_0 < u \leq u_0$ and $v(u) = 0$. Substituting this into eq. (\[sdiff\]), we obtain: $$S_{\rm BH} = - \int_{s\, u_0}^{u_0} du\,\,\frac{\partial S(u,0)}{\partial u} = S(s\, u_0, 0) - S(u_0, 0) \label{sbhsdiff}$$ We recognize this as the difference of the lengths of an $\infty$-wound geodesic and an ‘$(\infty-1)$-wound’ geodesic. Indeed, taking the boundary-anchored endpoint of the geodesic from $u_0$ to $s u_0$ winds the already infinitely wound geodesic one additional time. The extra winding happens on the horizon of the black hole, which justifies eq. (\[sbhsdiff\]). Of course, the other closed lightlike-curve in the kinematic space of the BTZ black hole gives a similar result. There, we substitute geodesics $v_0 < v < s v_0$ and $u(v) = 0$ into the analogue of (\[sdiff\]) appropriate for integrating over the $v$ coordinate: $$S_{\rm diff} = \int_{v_0}^{s\, v_0} dv\,\,\frac{\partial S(u,v)}{\partial v} \Big|_{u = u(v)} = S(0, s\, v_0) - S(0, v_0) = S_{\rm BH} \label{sbhsdiff2}$$ Interpreted in MERA, this computation recovers the minimal cut prescription in a novel setting: when the region whose entanglement we compute does not have endpoints. But as a bonus, we have explained why the thermal entropy can be read off from two different minimal cuts, on either side of the timelike-identified region. This is because we have two distinct families of lightlike-identified geodesics, which asymptote to the black hole horizon from either side of the wormhole. Their contributions add up to account for the factor of 2 in eq. (\[mithermal\]). Furthermore, our network construction obtains the exact spectrum of the thermofield double state and not just the scaling of the entropy with the number of cuts. Indeed, in Ref. [@quotientmera] our co-authors and we confirmed that the spectrum of the quotient network agrees quantitatively with the entanglement spectrum of the thermofield double state, including the numerical factors in eq. (\[ltobeta\]). This comparison was conducted in the critical Ising model, a decidedly non-holographic theory, in which case the kinematic space ought to be understood as the space of CFT intervals rather than the space of bulk geodesics. #### Lightlike-identified geodesics and entwinement Eqs. (\[sbhsdiff\]) and (\[sbhsdiff2\]) involve geodesics, which wrap around the black hole horizon. These geodesics do not compute the entanglement entropy of any spatial interval in the thermofield double state. If we allow the use of their lengths in the differential entropy formula, however, we obtain correct geometric quantities, including some information-theoretically meaningful ones such as the entropy of the BTZ black hole [@lampros]. Emboldened by this, Ref. [@entwinement] named a conjectured CFT avatar of the length of a non-minimal geodesic ‘entwinement.’ Working in the conical defect geometry, the authors of [@entwinement] studied entwinement and concluded that it is related to entanglement among internal degrees of freedom. But an intrinsic definition of entwinement has remained an open question since then. The quotient MERA manifests the relevance of entwinement in the CFT in the form of a lightlike-identified ray of tensors. Changing any one of these tensors affects the state on the entire CFT circle uniformly. Therefore, we may think of them as acting in the s-wave sector of the CFT, where no further spatial coarse-graining can be performed. Isometries in the lightlike-identified region separate the degrees of freedom which are internally entangled within the s-wave sector on one side from those which carry entanglement with the thermofield image. In the continuous geometry, entwinement is manifested by geodesics that wrap once or more around the black hole. Such ‘long geodesics’ are sensitive to the internal organization of the CFT degrees of freedom, but also exhibit some degree of localization on the CFT circle. The discrete nature of MERA collapses the entire family of long geodesics into one line of tensors, which live on a lightlike-identified ray. Entwinement is related to the structure of the CFT thermal state in the far infrared. At fixed temperature, it should therefore be more important for smaller circle sizes. By eq. (\[ltobeta\]) the effect of entwinement should wash out when $s \to \infty$ and gain in importance as $k = \log_2 s$ becomes of order 1. In Ref. [@quotientmera] our co-authors and we confirmed these expectations. In particular, the lightlike-identified region in the quotient network is approximately isometric, with the approximation improving exponentially in $k$. For these reasons, we view the lightlike-identified regions in MERA as a tangible CFT realization of entwinement: because of their nearly isometric character and because their effect is completely delocalized in the CFT. #### Spacelike-identified regions These prepare correlations between spatial regions on one side, just as they would in the vacuum. As unitary transformations between $\mathcal{H}_{\rm IR} \otimes \mathcal{H}_{\rm frozen}$ and $\mathcal{H}_{\rm UV}$, they select a local basis on each side in which the thermofield double state is presented. Different choices of local bases correspond to different conformal frames, distinguished by different UV cutoffs. Toward excited states: Geometry as compression {#compressionsummary} ============================================== The intimate relation between integral geometry and information theory [@lastpaper] prompted us to look for an analogous structure on the CFT side of the holographic duality. We found it in the MERA tensor network. The two key properties, which MERA shares with kinematic space are the causal structure and the representation of entanglement entropy as a ‘flux’ through a causal cut (counting lines in MERA and eq. (\[eece\]) in kinematic space). All our arguments originate from these two starting points. We would like to extend our conclusions beyond the ground state MERA and its sub-networks—the cases discussed in Secs. \[sec:therelation\] and \[sec:BHmera\]. We begin with the following observations: 1. In the ground state MERA, counting lines that cross a causal cut computes an entanglement entropy only in the optimized network. This feature is not a built-in property of the network; it is an emergent feature that arises after optimization and should not be expected to hold in excited states. 2. The kinematic metric is not rigid. As seen in eq. (\[kinmetric\]), it depends on the state under consideration. When we consider excited states, the kinematic metric can only be relevant to the optimized network whose structure is adjusted for an efficient description of the state. In consequence, we must look for a flexible network that can incorporate the entanglement pattern in its structure. We propose a holographically motivated generalization of MERA which (a) shares the causal structure of MERA and kinematic space and (b) maintains the approximate relation between entanglement entropies and counting lines on causal cuts. Property (b) will require the state on a causal cut to be an approximate product state, a condition that is generally achievable only in holographic theories. In information theoretic terms, such a tensor network is an *iterative compression* algorithm: it maps the density matrix of every interval to a compressed state on its exclusive causal cone. Conditional mutual information—which for the ground state MERA was the number of isometries in a causal diamond—now provides a local *density of compression*, namely the net reduction of the local Hilbert space dimension upon an isometric coarse-graining. Details of the construction, properties and limitations of the compression algorithm, as well as new insights about holographic geometries that follow from it, will be presented in an upcoming paper [@compression]. Summary ======= Holographic duality posits that gravity in an asymptotically anti-de Sitter space-time can be studied in a conformal field theory living on its asymptotic boundary. Of the many CFT quantities which encode the physics in AdS, entanglement entropies play a privileged role: they directly characterize the background space-time on top of which dynamics unfold [@rt1]. In the special case of the AdS$_3$/CFT$_2$ duality, where this characterization comes in the form of lengths of spacelike geodesics, it is particularly easy to convert it into a conventional picture of the bulk geometry in terms of points and distances. In [@lampros; @lastpaper] we studied this conversion in detail and, in the process, discovered kinematic space. From the CFT point of view, kinematic space is the geometry of intervals; in the bulk, it becomes the space of geodesics which is of interest in integral geometry. Kinematic space is a metric space of mixed signature: its causal structure reflects the containment relation among intervals while its volume form relates to variations of lengths of geodesics (eq. \[kinvolume\]). This last point is particularly significant in light of the Ryu-Takayanagi proposal: kinematic volumes compute conditional mutual informations of contiguous triples of intervals. The strikingly simple interpretation of kinematic space in information theory suggests that this concept will be useful for more general purposes. We will discuss one such use in an upcoming paper [@kinematicoperators], which is concerned with defining a convenient ‘kinematic’ basis of CFT operators that exploits the operator product expansion (OPE). #### Kinematic space and the MERA network In the present paper, we report another application of kinematic space, which was independently discovered a decade ago in [@mera]. In order to write down the ground state wavefunction of a critical system in 1+1 dimensions, Vidal proposed MERA, a tensor network with a structure tailored to the scale dependence of entanglement entropies. The same dependence fixes the metric of kinematic space. This is more than a superficial similarity; we found that the properties of MERA are identical to the analogous properties of kinematic space, modulo the obvious limitations induced by discretization. The key common features of MERA and kinematic space are the following: 1. *Causal propagation of information.* The locality and unitarity of the tensors comprising MERA implies that information propagates only within subregions dubbed ’causal cones.’ This auxiliary notion of causality in MERA is identical to the causal structure of kinematic space, which encapsulates the containment relation among CFT intervals. 2. *Localization of conditional mutual information.* In MERA, the conditional mutual information of three contiguous intervals localizes in a causal diamond. The same quantity defines the volume of a causal diamond in kinematic space. It is sensible to interpret conditional mutual information as a discrete volume form in MERA, because it simply counts isometries in a causal diamond. Using Stokes’ theorem we can relate the number of isometries in a MERA region to the number of lines entering it. This is a MERA version of the differential entropy formula (eq. \[sdiff\]), a special case of which reproduces the commonly used cut counting prescription for estimating entanglement entropies. 3. *Real space renormalization.* The MERA network performs RG transformations in real space and every spacelike cut defines a coarse-grained lattice. Kinematic space, understood as the space of CFT intervals, is a natural domain of real space cutoffs in the continuum. As in MERA, cutoff surfaces in kinematic space are not allowed to be timelike. An interesting type of cutoff is one that maximally coarse-grains two complementary intervals $A$ and $A^c$ (Sec. \[rgcausalcuts\]). Both in MERA and in kinematic space such a cutoff retains a finite portion of the network (resp. space of geodesics), which prepares (resp. represents) the correlations between $A$ and $A^c$. This is in stark contrast to what a direct bulk interpretation of MERA would suggest. An important ingredient that underlies this detailed agreement between the kinematic geometry and MERA is the realization of entanglement entropies in the network as the size of causal cuts. This is a property that emerges in the optimized network for the vacuum but does not hold in general. Since it is a crucial prerequisite for the compatibility of the kinematic metric with the network, the simple MERA construction needs to be refined for general states. This can done by promoting features 1 and 2 above to principles for building the network. Adopting these principles leads to a representation of kinematic space as a compression algorithm, the details of which are reserved for a separate upcoming publication [@compression]. #### Black Holes Cutting the MERA network inhomogeneously effects local conformal transformations [@quotientmera]. In the kinematic proposal, this type of truncated MERA network discretizes the space of geodesics of a locally AdS$_3$ geometry. We discussed these kinematic spaces in Sec. \[sec:BHmera\], with a particular emphasis on the BTZ black hole which involves an additional, non-trivial quotient. Both the thermofield double MERA and the BTZ kinematic space divide into three regions, distinguished by the timelike versus spacelike action of the quotient. The agreement of MERA and kinematic space goes beyond this structural similarity. The timelike-identified region of MERA, which relates to geodesics that connect the two asymptotic boundaries, is responsible for preparing the correlations (entanglement spectrum) between the two sides. The thermal entropy of each asymptotic boundary equals half the volume of this timelike-identified region, which equates the black hole entropy with the total number of geodesics connecting the two sides. On the other hand, the spacelike-identified sectors of the network, which correspond to geodesics that stay on one side of the black hole, prepare spatial correlations within each asymptotic region. An intriguing part of the thermofield double MERA are two single lines of tensors on which the quotient acts in a lightlike manner. Geometrically, they relate to geodesics which asymptote to the black hole horizon after winding around it infinitely many times. These tensors do not perform any type of coarse-graining of local degrees of freedom, just as winding geodesics do not compute entanglement entropies of localized intervals. We view these tensors as a concrete realization of the concept of entwinement—entanglement of internal or gauged degrees of freedom, which is conjecturally related to the length of non-minimal geodesics [@entwinement]. #### Comparison to other tensor network proposals Recent work on applications of tensor networks to holography includes novel networks distinct from MERA, which discretize the spatial geometry of anti-de Sitter space directly [@errorcorrectingnetwork; @patrickmichael]. At present, it is unclear whether these networks can prepare the wavefunction of the CFT ground state. Likewise, the question of how such networks can be extended to encode wavefunctions of excited states and their dual geometries remains open. On the other hand, the AdS/CFT correspondence essentially guarantees that a network of this type—one whose structure mimics the fabric of the bulk geometry—must exist. After all, the change of basis that takes CFT degrees of freedom into low energy effective fields in the bulk can always be presented in the form of a tensor network. While the quest for such a network continues, we believe that it is also important to work with those tensor networks, which are known to prepare the correct CFT states. No ansatz is successful by accident; if MERA prepares the CFT ground state, it does so because its design was guided by the correct set of principles. As we explained above, the principles underlying MERA are also the defining features of kinematic space, which in holographic theories becomes the space of bulk geodesics. In this way, MERA makes contact with geometric concepts without any input from the bulk side; it is a showcase example of an *emergent* geometry. At the same time, because MERA requires no input from the bulk and because kinematic space can be merely a space of CFT intervals (rather than a space of geodesics), our analysis applies to non-holographic theories. Ref. [@quotientmera] illustrates that this approach can benefit the more traditional, CFT-centered applications of tensor networks. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Glen Evenbly, Patrick Hayden, Esperanza Lopez, Don Marolf, Rob Myers, John Preskill, Xiao-Liang Qi, Joan Sim[ó]{}n, Leonard Susskind, Brian Swingle and Guifr[é]{} Vidal for useful discussions. BC, SM and JS thank Caltech, BC thanks the University of Amsterdam and the University of Edinburgh, and JS thanks Princeton University for hospitality. BC thanks the organizers of “Holographic duality for condensed matter physics” and KITPC-CAS in Beijing. We all thank the organizers of “Quantum Gravity Foundations: UV to IR,” “Closing the Entanglement Gap: Quantum Information, Quantum Matter and Quantum Fields,” and the Follow-On Program held at KITP (supported in part by the National Science Foundation under Grant No. NSF PHY11-25915), and of “Quantum Information Theory in Quantum Gravity II” meeting held at the Perimeter Institute for Theoretical Physics (supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research and Innovation). BC and JS thank the organizers of the workshop “AdS/CFT and Quantum Gravity” at Centre de Recherches Math[é]{}matiques at the University of Montreal. SM was supported in part by an award from the Department of Energy (DOE) Office of Science Graduate Fellowship Program. [99]{} A. Einstein, “Erkl[ä]{}rung der Perihelbewegung des Merkur aus der allgemeinen Relativit[ä]{}tstheorie,” K[ö]{}niglich Preussische Akademie der Wissenschaften (Berlin), Sitzungsberichte 831-839 (25 November 1915). J. Maldacena, “The large $N$ limit of superconformal field theories and supergravity,” Int. J. Theor. Phys.  [**38**]{}, 1113 (1999) \[Adv. Theor. Math. Phys.  [**2**]{}, 231 (1998)\] \[hep-th/9711200\]. S. D. Mathur, “The information paradox: A pedagogical introduction,” Class. Quant. Grav.  [**26**]{}, 224001 (2009) \[arXiv:0909.1038 \[hep-th\]\]. A. Almheiri, D. Marolf, J. Polchinski and J. Sully, “Black holes: Complementarity or firewalls?,” JHEP [**1302**]{}, 062 (2013) \[arXiv:1207.3123 \[hep-th\]\]. B. Czech, L. Lamprou, S. McCandlish and J. Sully, “Integral geometry and holography,” arXiv:1505.05515 \[hep-th\]. L. Santal[ó]{}, “Integral geometry and geometric probability,” Cambridge University Press, 1976. S. Helgason, “Radon transform,” Birkh[ä]{}user Boston, 1983. G. Beylkin, “Discrete Radon transform,” Acoustics, Speech and Signal Processing, IEEE Transactions on [**35**]{}, no 2, 162 (1987). “The Nobel Prize in Physiology or Medicine 1979,” Nobelprize.org. Nobel Media AB 2014, retriveved on 26 Sep 2015. J. Lin, M. Marcolli, H. Ooguri and B. Stoica, “Locality of gravitational systems from entanglement of conformal field theories,” Phys. Rev. Lett.  [**114**]{}, no. 22, 221601 (2015) \[arXiv:1412.1879 \[hep-th\]\]. S. Ryu and T. Takayanagi, “Holographic derivation of entanglement entropy from AdS/CFT,” Phys. Rev. Lett.  [**96**]{}, 181602 (2006) \[hep-th/0603001\]. S. Ryu and T. Takayanagi “Aspects of holographic entanglement entropy,” JHEP [**0608**]{}, 045 (2006) \[hep-th/0605073\]. V. E. Hubeny, M. Rangamani and T. Takayanagi, “A covariant holographic entanglement entropy proposal,” JHEP [**0707**]{}, 062 (2007) \[arXiv:0705.0016 \[hep-th\]\]. N. Lashkari, C. Rabideau, P. Sabella-Garnier and M. Van Raamsdonk, “Inviolable energy conditions from entanglement inequalities,” JHEP [**1506**]{}, 067 (2015) \[arXiv:1412.3514 \[hep-th\]\]. N. Lashkari, M. B. McDermott and M. Van Raamsdonk, “Gravitational dynamics from entanglement ’thermodynamics’,” JHEP [**1404**]{}, 195 (2014) \[arXiv:1308.3716 \[hep-th\]\]. T. Faulkner, M. Guica, T. Hartman, R. C. Myers and M. Van Raamsdonk, “Gravitation from entanglement in holographic CFTs,” JHEP [**1403**]{}, 051 (2014) \[arXiv:1312.7856 \[hep-th\]\]. B. Swingle and M. Van Raamsdonk, “Universality of gravity from entanglement,” arXiv:1405.2933 \[hep-th\]. N. Lashkari and M. Van Raamsdonk, “Canonical Energy is Quantum Fisher Information,” arXiv:1508.00897 \[hep-th\]. V. Balasubramanian, B. D. Chowdhury, B. Czech, J. de Boer and M. P. Heller, “A hole-ographic spacetime,” Phys. Rev. D [**89**]{}, 086004 (2014) \[arXiv:1310.4204 \[hep-th\]\]. M. Headrick, R. C. Myers and J. Wien, “Holographic holes and differential entropy,” JHEP [**1410**]{}, 149 (2014) \[arXiv:1408.4770 \[hep-th\]\]. B. Czech, P. Hayden, N. Lashkari and B. Swingle, “The information theoretic interpretation of the length of a curve,” JHEP [**1506**]{}, 157 (2015) \[arXiv:1410.1540 \[hep-th\]\]. B. Czech and L. Lamprou, “Nuts and bolts for creating space,” Phys. Rev. D [**90**]{}, 106005 (2014) \[arXiv:1409.4473 \[hep-th\]\]. M. Van Raamsdonk, “Comments on quantum gravity and entanglement,” arXiv:0907.2939 \[hep-th\]. M. Van Raamsdonk, “Building up spacetime with quantum entanglement,” Gen. Rel. Grav.  [**42**]{}, 2323 (2010) \[Int. J. Mod. Phys. D [**19**]{}, 2429 (2010)\] \[arXiv:1005.3035 \[hep-th\]\]. G. L. Leclerc, Histoire de l’Acad. Roy. des. Sciences, 43 (1733). G. Vidal, “Entanglement renormalization,” Phys. Rev. Lett.  [**99**]{}, 220405 (2007) \[cond-mat/0512165\]. G. Vidal, “A class of quantum many-body states that can be efficiently simulated,” Phys. Rev. Lett.  [**101**]{}, 110501 (2008) \[cond-mat/0610099\]. F. Pastawski, B. Yoshida, D. Harlow and J. Preskill, “Holographic quantum error-correcting codes: Toy models for the bulk/boundary correspondence,” JHEP [**1506**]{}, 149 (2015) \[arXiv:1503.06237 \[hep-th\]\]. Z. Yang, P. Hayden and X. L. Qi, “Bidirectional holographic codes and sub-AdS locality,” arXiv:1510.03784 \[hep-th\]. B. Czech, G. Evenbly, L. Lamprou, S. McCandlish, X.-L. Qi, J. Sully and G. Vidal, “A tensor network quotient takes the vacuum to the thermal state,” arXiv:1510.07637 \[cond-mat.str-el\]. B. Czech, P. Hayden, L. Lamprou, S. McCandlish and J. Sully, “Geometry as a compression algorithm,” [*to appear*]{}. G. Solanes, “Integral geometry and curvature integrals in hyperbolic space,” Ph.D. thesis, Universitat Autonoma de Barcelona, 2003. E. H. Lieb and M. B. Ruskai, “Proof of the strong subadditivity of quantum mechanical entropy,” J. Math. Phys. [**14**]{}, 1938 (1973). G. Vidal, “Entanglement renormalization: An introduction,” arXiv:0912.1651 \[cond-mat.str-el\]. G. Evenbly and G. Vidal, “Quantum criticality with the Multi-scale Entanglement Renormalization Ansatz,” arXiv:1109.5334 \[quant-ph\]. X. L. Qi, “Exact holographic mapping and emergent space-time geometry,” arXiv:1309.6282 \[hep-th\]. G. Evenbly and G. Vidal, “Algorithms for entanglement renormalization,” Phys. Rev. B [**79**]{}, 144108 (2009) \[arXiv:0707.1454 \[cond-mat.str-el\]\]. R. Pfeifer, G. Evenbly and G. Vidal, “Entanglement renormalization, scale invariance, and quantum criticality,” Phys. Rev. A [**79(4)**]{}, 040301(R) (2009) \[arXiv:0810.0580 \[cond-mat.str-el\]\]. G. Evenbly, P. Corboz and G. Vidal, “Non-local scaling operators with entanglement renormalization,” Phys. Rev. B [**82**]{}, 132411 (2010) \[arXiv:0912.2166 \[cond-mat.str-el\]\]. C. Beny, “Causal structure of the entanglement renormalization ansatz,” New J. Phys.  [**15**]{}, 023020 (2013) \[arXiv:1110.4872 \[quant-ph\]\]. B. Swingle, “Entanglement renormalization and holography,” Phys. Rev. D [**86**]{}, 065007 (2012) \[arXiv:0905.1317 \[cond-mat.str-el\]\]. B. Swingle, “Constructing holographic spacetimes using entanglement renormalization,” arXiv:1209.3304 \[hep-th\]. P. Calabrese and J. L. Cardy, “Entanglement entropy and quantum field theory,” J. Stat. Mech.  [**0406**]{}, P06002 (2004) \[hep-th/0405152\]. P. Hayden, S. Nezami, X.-L. Qi, N. Thomas, M. Walter, and Z. Yang, “Holographic duality defined by random tensor networks,” [*to appear*]{}. D. N. Page, “Average entropy of a subsystem,” Phys. Rev. Lett.  [**71**]{}, 1291 (1993) \[gr-qc/9305007\]. G. Evenbly, [*private communication*]{}. G. Evenbly and G. Vidal, “A theory of minimal updates in holography,” Phys. Rev. B [**91**]{}, 205119 (2015) \[arXiv:1307.0831 \[quant-ph\]\]. M. Nozaki, S. Ryu and T. Takayanagi, “Holographic geometry of entanglement renormalization in quantum field theories,” JHEP [**1210**]{}, 193 (2012) \[arXiv:1208.3469 \[hep-th\]\]. T. Hartman and J. Maldacena, “Time evolution of entanglement entropy from black hole interiors,” JHEP [**1305**]{}, 014 (2013) \[arXiv:1303.1080 \[hep-th\]\]. D. A. Roberts, D. Stanford and L. Susskind, “Localized shocks,” JHEP [**1503**]{}, 051 (2015) \[arXiv:1409.8180 \[hep-th\]\]. D. Marolf, H. Maxfield, A. Peach and S. F. Ross, “Hot multiboundary wormholes from bipartite entanglement,” arXiv:1506.04128 \[hep-th\]. N. Bao, C. Cao, S. M. Carroll, A. Chatwin-Davies, N. Hunter-Jones, J. Pollack and G. N. Remmen, “Consistency conditions for an AdS multiscale entanglement renormalization ansatz correspondence,” Phys. Rev. D [**91**]{}, no. 12, 125036 (2015) \[arXiv:1504.06632 \[hep-th\]\]. R. C. Myers, J. Rao and S. Sugishita, “Holographic Holes in Higher Dimensions,” JHEP [**1406**]{}, 044 (2014) \[arXiv:1403.3416 \[hep-th\]\]. M. Miyaji and T. Takayanagi, “Surface/State Correspondence as a Generalized Holography,” PTEP [**2015**]{}, no. 7, 073B03 (2015) \[arXiv:1503.03542 \[hep-th\]\]. B. Czech, L. Lamprou, S. McCandlish and J. Sully, [*to appear*]{}. E. T. Akhmedov, “A Remark on the AdS / CFT correspondence and the renormalization group flow,” Phys. Lett. B [**442**]{}, 152 (1998) \[hep-th/9806217\]. V. Balasubramanian and P. Kraus, “Space-time and the holographic renormalization group,” Phys. Rev. Lett.  [**83**]{}, 3605 (1999) \[hep-th/9903190\]. J. de Boer, E. P. Verlinde and H. L. Verlinde, “On the holographic renormalization group,” JHEP [**0008**]{}, 003 (2000) \[hep-th/9912012\]. M. Ba[ñ]{}ados, “Three-dimensional quantum geometry and black holes,” AIP Conf. Proc.  [**484**]{}, 147 (1999) \[hep-th/9901148\]. G. Mandal, R. Sinha and N. Sorokhaibam, “The inside outs of AdS$_{3}$/CFT$_{2}$: exact AdS wormholes with entangled CFT duals,” JHEP [**1501**]{}, 036 (2015) \[arXiv:1405.6695 \[hep-th\]\]. M. Sheikh-Jabbari, J. Sim[ó]{}n and H. Yavartanoo, [*to appear*]{}. M. Ba[ñ]{}ados, M. Henneaux, C. Teitelboim and J. Zanelli, “Geometry of the (2+1) black hole,” Phys. Rev. D [**48**]{}, 1506 (1993) \[Phys. Rev. D [**88**]{}, no. 6, 069902 (2013)\] \[gr-qc/9302012\]. V. Balasubramanian, P. Hayden, A. Maloney, D. Marolf and S. F. Ross, “Multiboundary wormholes and holographic entanglement,” Class. Quant. Grav.  [**31**]{}, 185015 (2014) \[arXiv:1406.2663 \[hep-th\]\]. A. de la Fuente and R. Sundrum, “Holography of the BTZ black hole, inside and out,” JHEP [**1409**]{}, 073 (2014) \[arXiv:1307.7738 \[hep-th\]\]. P. Kraus, H. Ooguri and S. Shenker, “Inside the horizon with AdS / CFT,” Phys. Rev. D [**67**]{}, 124022 (2003) \[hep-th/0212277\]. V. Balasubramanian, B. D. Chowdhury, B. Czech and J. de Boer, “Entwinement and the emergence of spacetime,” JHEP [**1501**]{}, 048 (2015) \[arXiv:1406.5859 \[hep-th\]\]. [^1]: Numerous other gravitational quantities also have information-theoretic dual descriptions, including a version of the null energy condition [@nec], Einstein’s equations [@einsteineq1; @einsteineq2; @einsteineq3], canonical energy [@canenergy], lengths of curves [@holeography; @robproof; @protocol], the triangle inequality [@lampros] and even connectedness of spacetime [@marksessay; @mark2]. [^2]: Our arguments are structural in character and differ fundamentally from the reasoning followed in [@critique], which was based on counting degrees of freedom. [^3]: Of course these statements hold up to the usual artifacts of discretization. [^4]: They are also reminiscent of the ‘torus with whiskers’ analyzed in [@funnytorus].

ArXiv

--- abstract: 'The alt-right is a far-right movement that has uniquely developed on social media, before becoming prominent in the 2016 United States presidential elections. However, very little research exists about their discourse and organization online. This study aimed to analyze how a sample of alt-right supporters organized themselves in the week before and after the 2018 midterm elections in the US, along with which topics they most frequently discussed. Using community finding and topic extraction algorithms, results indicated that the sample commonly used racist language and anti-immigration themes, criticised mainstream media and advocated for alternative media sources, whilst also engaging in discussion of current news stories. A subsection of alt-right supporters were found to focus heavily on white supremacist themes. Furthermore, small groups of alt-right supporters discussed anime, technology and religion. These results supported previous results from studies investigating the discourse of alt-right supporters.' author: - | Ańgel Panizo-LLedot\ Computer Science Department\ Universidad Autónoma de Madrid, Spain\ `[email protected]` Javier Torregrosa\ Biological and Health Psychology Department\ Universidad Autónoma de Madrid, Spain\ `[email protected]` Gema Bello-Orgaz\ Departmento de Sistemas Informáticos\ Universidad Politécnica de Madrid, Spain\ `[email protected]` Joshua Thorburn\ RMIT University\ Melbourne, Victoria, Australia\ `[email protected]` David Camacho\ Departmento de Sistemas Informáticos\ Universidad Politécnica de Madrid, Spain\ `[email protected]` bibliography: - 'template.bib' title: 'Describing Alt-Right communities and their discourse on Twitter during the 2018 US mid-term elections' --- Acknowledgements {#acknowledgements .unnumbered} ================ This work has been supported by several research grants: Spanish Ministry of Science and Education under TIN2014-56494-C4-4-P grant (DeepBio) and Comunidad Autónoma de Madrid under S2013/ICE-3095 grant (CYNAMON).

ArXiv

--- author: - 'W. Luis Mochán^,^, Raksha Singla^^, Lucila Juárez^^, and Guillermo P. Ortiz^^' bibliography: - 'biblio.bib' title: Recursive Calculation of the Optical Response of Multicomponent Metamaterials --- Introduction {#intro} ============ Metamaterials made up of a repeated pattern of one or more ordinary materials within a host have optical properties that might differ substantially from those of its components [@shalaev_optical_2007]. According to their geometry and composition they might display both electric and magnetic resonances of dipolar and quadrupolar nature [@cho_contribution_2008] around which their macroscopic permittivity and permeability may become negative, yielding an exotic negative refraction [@shelby2001experimental; @smith2004metamaterials; @aydin2005investigation; @hoffman2007negative; @peng2007experimental]. A usual typical geometry of these left-handed and other exotic metamaterials is that of pairs of wires and split conducting rings within a dielectric. Nevertheless, to avoid the dissipation inherent within the conducting phases, all-dielectric structures employing high index of refraction have also been investigated [@jahani_all-dielectric_2016; @kivshar_all-dielectric_2018], and it has been shown that their Mie like resonances may also be employed for guiding light and for enhancing non-linear optical effects. All dielectric structures of appropriate shapes may also exhibit negative dispersion [@vynck_all-dielectric_2009]. The microscopic field within metamaterials may have small regions where the field is very high. Small modifications of the composition at these regions may produce notable macroscopic effects. Thus, metamaterials have been used to develop different kinds of sensors for different spectral regions [@ming_huang_microwave_2011; @la_spada_metamaterial-based_2011; @chen_metamaterials_2012; @wongkasem_chiral_2017]. Many other known and emerging applications of metamaterials have been reviewed recently [@pendry_photonics:_2006; @baev_metaphotonics:_2015]. The permeability and permittivity, as well as chiral properties of a metamaterial may be obtained from its reflection and transmission properties [@smith_determination_2002] and from the dispersion of guided modes within metamaterial waveguides [@chen_experimental_2006], but more fundamentally, from the frequency and wavevector dependence of an appropriately defined spatially-dispersive macroscopic dielectric function [@agranovich_linear_2004; @silveirinha_metamaterial_2007; @agranovich_electrodynamics_2009; @alu_restoring_2011; @alu_first-principles_2011; @konovalenko_nonlocal_2019]. A very efficient scheme for the calculation of the optical properties of metamaterials has been developed for binary metamaterials by exploiting an analogy between the macroscopic dielectric tensor and the projected Green’s function corresponding to a Hermitian Hamiltonian [@mochan_efficient_2010] which may be obtained through Haydock’s recursive procedure [@haydock_electronic_1972; @haydock_recursive_1980; @pettifor_recursion_1984]. The method has been used to study extraordinary transmission through perforated metallic slabs [@mochan_efficient_2010], to calculate plasmonic properties of odd shaped metallic inclusions [@cortes_optical_2010], to study enhanced birrefrigency and dichroism in anisotropic metamaterials [@mendoza_birefringent_2012], for the design and optimization of optical devices to control the absorption [@ortiz_effective_2014] and polarization of light [@mendoza_tailored_2016], and to optimize electrical and optical properties of semitransparent contacts [@toranzos_optical_2017]. The method has also been generalized to account for retardation, yielding a non-local macroscopic response from which the complete band structure of photonic crystals may be obtained [@perez-huerta_macroscopic_2013] and from which magnetic properties may be extracted [@juarez-reyes_magnetic_2018]. There has also surged interest in the nonlinear properties of metamaterials [@lapine_colloquium:_2014; @czaplicki] and metasurfaces [@czaplicki]. The Haydock’s recursive approach has been extended to calculate the microscopic field and from it the macroscopic non-linear optical response of metamaterials. In particular, to obtain the second harmonic generation spectra of metamaterials with centrosymmetric components but noncentrosymmetric shapes [@meza_second-harmonic_2019]. Unfortunately, the efficient computational approach developed in [@mochan_efficient_2010] is directly applicable only to binary metamaterials, that is, to systems composed of exactly two different materials $A$ and $B$. The reason for this limitation is that the geometry of such systems may be decoupled from their composition and described by a characteristic function $B(\bm r)$ whose value is 1 for those points $\bm r$ that belong to region $B$, and 0 when $\bm r$ does not, i.e., within region $A$. It is from this characteristic function that a Hermitian operator is built, regardless of the actual composition of $A$ and $B$, and of their dielectric or conducting nature, their dispersion and dissipation. The Haydock coefficients for this operator are readily obtained and from them a closed expression for its macroscopic response may be built. However, for multicomponent systems, one cannot find such a Hermitian operator to describe the geometry, and the dielectric response itself is not Hermitian in the presence of dissipation. Given this limitation, many interesting systems seem to lie beyond the possibilities of the recursive approach. For example, metasurfaces are arrangements of patterned particles on a substrate [@yu_flat_2014] that have been used to manipulate the refraction of light [@yu_light_2011] and produce flat lenses [@khorasaninejad_metalenses_2016], compound lenses [@khorasaninejad_metalenses_2016] and even fabricate spin switchable holograms [@chen_spin-controlled_2018]. A numerical study of metasurfaces would require at least three materials corresponding to the particles, the substrate and the ambient. Similarly, it has been shown that the field enhancement due to resonant excitation of plasmonic particles may not decrease when protected by a dielectric, if the dielectric presents a coexisting Mie resonance [@liren_deng_enhancing_2019]. Arrays of metallic cores coated by semiconductors may also display negative index of refraction as an electric dipole plasmonic resonance might coexist with a magnetic dipole Mie resonance [@paniagua-dominguez_ultra_2013; @paniagua-dominguez_metallo-dielectric_2011; @abujetas_photonic_2015]. The study of these coupled plasmonic-Mie resonances requires accounting at least for a core, a coating and the ambient. The purpose of the present paper is to generalize the efficient homogenization procedure using Haydock’s recursion, as presented in Ref. [@mochan_efficient_2010], in order to deal with periodic metamaterials of arbitrary geometry and composition and with an arbitrary number of components, or even with a dielectric response that varies continuously in space. To that end, we realize that though the dielectric response is not in general an Hermitian operator, it corresponds to a symmetrical complex operator. Thus, we can cautiously employ well know theorems of linear algebra provided we define an appropriate Euclidean-like metric instead of the usual Hermitian metric. However, this metric couples Bloch waves moving in opposite directions, requiring us to introduce a spinor-like two-component representation of the Bloch states, with one component for each of the opposing propagation directions. Besides obtaining the macroscopic dielectric response of the system, we can also calculate the microscopic electric field, so our procedure can further be employed in non-linear calculations. We restrict ourselves to the non-retarded, long-wavelength approximation, though the same ideas can be applied to fully retarded calculations. In order to verify the suitability of our computational procedure, we calculate the macroscopic response of various 2D multicomponent systems and verify that our results are consistent with a generalized [@ortiz_kellers_2018] Keller’s theorem [@keller_theorem_1964], and with Mortola and Steffé’s exact expression [@s._mortola_two-dimensional_1985; @milton_proof_2001] for four-component chess-board systems. The structure of the paper is the following. In Sec. \[theory\] we present our theory: In Subsec. \[ssmulti\] we develop our recursive approach to the calculation of the dielectric function. In order to test our results, in Subsec. \[sscoated\] we obtain analytical approximate formulae for the response of a simple multicomponent system, in Subsec. \[ssKeller\] we present a generalized Keller’s theorem for multicomponent 2D systems and in Subsec. \[ssmortola\] we discuss a 2D four component system for which exact analytical expressions are available. In Sec. \[results\] we present numerical results for a variety of 2D systems and verify that they agree with analytical results in the appropriate limits, that in general they obey Keller’s theorem and that they are consistent with Mortola and Steffe’s expression. Finally, Sec. \[conclusions\] is devoted to conclusions. Theory ====== Multicomponent metamaterials {#ssmulti} ---------------------------- In the non-retarded, long-wavelength limit, the longitudinal projection of the macroscopic dielectric function of a periodic system may be obtained from [@mochan_electromagnetic_1985; @mochan_electromagnetic_II_1985] $$\label{epsM} (\hat\epsilon_M^{LL})^{-1}=(\hat\epsilon^{LL})^{-1}_{aa},$$ where the superscript $LL$ and the subscript $aa$ on an operator $\hat{\mathcal O}$ denote the application of longitudinal projectors $\hat{\mathcal P}^L$ and the application of spatial average projectors $\hat{\mathcal P}_a$ on both sides of $\hat{\mathcal O}$. For a periodic system with its fields represented in reciprocal space we may express the longitudinal and average projectors by the matrices $$\label{PL} \mathcal P^L_{\bm G\bm G'}=\hat{\bm G}_{\bm k}\hat{\bm G}_{\bm k}\delta_{\bm G\bm G'},$$ and $$\label{Pa} \mathcal P^a_{\bm G\bm G'}=\delta_{\bm G\bm 0}\delta_{\bm G'\bm 0},$$ where $\{\bm G\}$ is the reciprocal lattice and we abbreviate the unit vectors, $$\label{hatG} \hat{\bm G}_{\bm k}\equiv \frac{\bm k+\bm G}{|\bm k+\bm G|}$$ where $\bm k$ is a small Bloch’s wavevector which in the long-wavelength approximation is assumed to be much smaller that $G$, except for the case $\bm G=0$, for which we define the direction $\hat{\bm 0}_{\bm k}\equiv\hat{\bm k}/k$. We remark that we may simplify our calculations by reinterpreting the $LL$ [*projection*]{} in Eq. (\[epsM\]) and similar equations below by the $LL$ [*component*]{}, representing any operator $\hat{\mathcal O}^{LL}$ by the matrix $\hat {\bm G}_{\bm k}\cdot\mathcal O_{\bm G\bm G'} \hat {\bm G}'_{\bm k}$. For a system with only two components $A$ and $B$ we define a characteristic function $B(\bm r)$ which takes the values 0 when $\bm r\in A$ and 1 when $\bm r\in B$. In this case, we may write the [ *microscopic*]{} dielectric function $$\label{epsvsu} \epsilon(\bm r)=\frac{\epsilon_A}{u}(u-B(\bm r)),$$ where $\epsilon_\alpha$ is the dielectric function of component $\alpha=A,B$ and $$\label{u} u=\frac{1}{1-\epsilon_B/\epsilon_A},$$ is the [*spectral variable*]{}. From Eqs. (\[epsM\]) and (\[epsvsu\]) it is clear that we only need the average projection of the operator $$\label{Gu} \hat{\mathcal G}(u)=(u-\hat B^{LL})^{-1},$$ which plays the role of a Green’s function for the operator $\hat B^{LL}$, the longitudinal projection of the charateristic function. The spectral variable $u$ would then play the role of a complex [ *energy*]{} which depends on the dielectric functions of both media, which in turn are generally complex valued functions of the frequency. As $\hat B^{LL}$ is a Hermitian operator, it can be represented as a tridiagonal real matrix with diagonal elements $a_n$, and subdiagonal and supradiagonal elements $b_n$, its Haydock coefficients, in a basis of Haydock states $\ket{n}$ obtained from an initial [*macroscopic*]{} state $\ket{0}$ by repeatedly applying $\hat B^{LL}$ and orthonormalizing the resulting state, i.e., defining $$\label{haydockstep} \hat B^{LL}\ket{n}\equiv b_{n+1}\ket{n+1}+a_n\ket{n}+b_{n}\ket{n-1},$$ with the condition $$\label{orthonorm} \braket{n|m}=\delta_{nm}.$$ The resulting response is given by the continued fraction $$\label{epsMH} \epsilon_M^{LL}= \frac{\epsilon_A}{u}\left(u-a_0 - \frac{b_1^2}{u-a_1- \frac{b_2^2}{u-a_2-\frac{b_3^2}{\ddots}}}\right).$$ Details of this procedure may be seen in Ref. [@mochan_efficient_2010]. For multicomponent metamaterials the procedure above does not work, as the geometry of the system would no longer be described by a single characteristic function, and if we introduce several characteristic functions, one for each component, then it wouldn’t be possible to represent all of them by tridiagonal matrices in the same basis. One way out of this difficulty is to use the longitudinal part of the microscopic dielectric function $\hat\epsilon^{LL}$ as the operator to use in Haydock’s recursion. If we replace the recursion (\[haydockstep\]) by $$\label{haydockstep1} \hat \epsilon^{LL}\ket{n}\equiv b_{n+1}\ket{n+1}+a_n\ket{n}+b_{n}\ket{n-1},\quad \text{(ND)}$$ then the macroscopic response would be given by $$\label{epsM1} \epsilon_M^{LL}= \left(a_0 - \frac{b_1^2}{a_1- \frac{b_2^2}{a_2-\frac{b_3^2}{\ddots}}}\right).\quad \text{(ND)}$$ Nevertheless, this procedure would only work in the absence of dissipation, when $\epsilon(\bm r)$ is real and $\hat\epsilon^{LL}$ is a Hermitian operator. Otherwise, there would be no reason for Eq. (\[haydockstep1\]) to contain only three terms on its RHS with real coefficients nor for its first and third terms to contain coefficients from the same set $\{b_n\}$. We would have instead $$\label{haydockstep2} \begin{split} \hat \epsilon^{LL}\ket{n}\equiv& b_{n+1}\ket{n+1}+a_n\ket{n}+c_{n}\ket{n-1}\\ &+d_{n}\ket{n-2}+\ldots \end{split}$$ with complex coefficients $a_n$, $b_n$, $c_n$, $d_n$…, and Eq. (\[epsM1\]) would no longer hold. For this reason we flagged Eq. (\[haydockstep1\]) and (\[epsM1\]) with ND (no dissipation). We notice that even when there is dissipation, the longitudinal dielectric function is a symmetrical operator. To show this, we chose an [*Euclidean*]{} scalar product between states $$\label{scalarr} \braket{\phi|\psi}\equiv \int d^3\bm r\, \phi(\bm r)\psi(\bm r)$$ where $\phi(\bm r)$ and $\psi(\bm r)$ are the [*wavefunctions*]{} that represent the [*states*]{} $\ket{\phi}$ and $\ket{\psi}$ in real space. Notice that in Eq. (\[scalarr\]) we didn’t conjugate $\phi(\bm r)$ as we would have done had we chosen a [*Hermitian*]{} product. We can express this scalar product in reciprocal space as $$\label{scalarq} \braket{\phi|\psi}\equiv \int \frac{d^3\bm q}{(2\pi)^3} \phi(-\bm q)\psi(\bm q)=\int \frac{d^3\bm q}{(2\pi)^3} \phi(\bm q)\psi(-\bm q),$$ where we define the Fourier transform $\zeta(\bm q)$ of any function $\zeta(\bm r)$ through $$\label{Fourier} \zeta(\bm r)\equiv \int \frac{d^3\bm q}{(2\pi)^3} \zeta(\bm q)e^{i\bm q\cdot\bm r}.$$ Notice in Eq. (\[scalarq\]) the minus sign in the argument of $\phi$ instead of the its conjugate as in Parseval’s theorem. Then, we may compute a matrix element of $\hat\epsilon^{LL}$ as $$\label{phiepspsi} \braket{\phi|\hat\epsilon^{LL}|\psi}=-\int\frac{d^3\bm q}{(2\pi)^3}\int\frac{d^3\bm q'}{(2\pi)^3}\phi(\bm q)\hat{\bm q}\cdot \epsilon(-\bm q-\bm q')\hat{\bm q}\psi(\bm q),$$ where $\epsilon(\bm q)$ is the Fourier transform of $\epsilon(\bm r)$. Clearly, $\braket{\phi|\hat\epsilon^{LL}|\psi}=\braket{\psi|\hat\epsilon^{LL}|\phi}$ showing that the operator $\hat\epsilon^{LL}$ is symmetric under the appropriate scalar product. Notice that for a periodic system, $\epsilon(\bm r)=\epsilon(\bm r+\bm R)$ may be written as a Fourier [*series*]{} with coefficients $$\label{epsG} \epsilon_{\bm G}=\int_{\text{UC}} \frac{d^3r}{\Omega} \epsilon(\bm r) e^{-i\bm G\cdot\bm r},$$ related to the Fourier transform $\epsilon(\bm q)=(2\pi)^3 \sum_{\bm G} \epsilon_{\bm G} \delta(\bm q-\bm G)$, where $\{\bm R\}$ is the Bravais lattice, $\{\bm G\}$ its reciprocal lattice, and UC indicates that the integral is over a unit cell, whose volume is $\Omega$. Thus, we can write Eq. (\[phiepspsi\]) as $$\label{phiepspsi1} \begin{split} \braket{\phi|\hat\epsilon^{LL}|\psi}=&\int_{\text{BZ}}\frac{d^3\bm k}{(2\pi)^3}\sum_{\bm G}\sum_{\bm G'} \phi(-\bm k-\bm G)\\ &\hat{\bm G}\cdot \epsilon_{\bm G-\bm G'}\hat{\bm G'}\psi(\bm k+\bm G'), \end{split}$$ where we replaced the wavevector $\bm q$ by the sum of a Bloch’s vector $\bm k$ and some reciprocal vector $\bm G$, and BZ indicates that the integral is over the first Brillouin zone. Ordinarily, in a periodic system, the normal modes of the system may be chosen as Bloch waves with a single Bloch’s vector $\bm k$. However, our chosen metric couples $\bm k$ to $-\bm k$. Thus we will consider simultaneously states with given Bloch’s vectors $\pm\bm k$ and denote them using a spinor-like notation as, $$\label{spinor} \ket{\zeta}\to\left( \begin{array}{c} \zeta(\bm k+\bm G)\\ \zeta(-\bm k+\bm G) \end{array} \right).$$ Consequently, we represent the dielectric response as a $2\times2$ matrix, $$\label{epsspin} \hat\epsilon^{LL}\to\left( \begin{array}{cc} \hat{\bm G}_{\bm k}\cdot\epsilon_{\bm G-\bm G'}\hat{\bm G'}_{\bm k}&0\\ 0&\hat{\bm G}_{-\bm k}\cdot\epsilon_{\bm G-\bm G'}\hat{\bm G'}_{-\bm k} \end{array} \right).$$ The scalar product (\[scalarq\]) becomes $$\label{scalarspin} \begin{split} \braket{\phi|\psi}=\sum_G&\left(\phi(-\bm k-\bm G)\psi(\bm k +\bm G) \right. \\ & +\left.\phi(\bm k-\bm G)\psi(-\bm k+\bm G)\right). \end{split}$$ From Eqs. (\[spinor\])–(\[scalarspin\]) we get $$\label{phiepspsis} \begin{split} \braket{\phi|\hat\epsilon^{LL}|\psi}&= \sum_{\bm G\bm G'}\left(\phi(-\bm k-\bm G) \hat{\bm G}_{\bm k}\cdot\epsilon_{\bm G-\bm G'}\hat{\bm G'}_{\bm k}\psi(\bm k+\bm G')\right. \\ &+ \left. \phi(\bm k-\bm G) \hat{\bm G}_{-\bm k}\cdot\epsilon_{\bm G-\bm G'}\hat{\bm G'}_{-\bm k} \psi(-\bm k+\bm G')\right). \end{split}$$ Using Eqs. (\[spinor\])–(\[phiepspsis\]) we can proceed to build a Haydock’s representation of the operator $\hat\epsilon^{LL}$. We start from a couple of macroscopic states representing longitudinal waves propagating in the directions $\pm\hat{\bm k}$, corresponding to the starting spinor $$\label{zero} \ket{0}\to\frac{1}{\sqrt2}\left( \begin{array}{c} 1\\1 \end{array} \right)\delta_{\bm G\bm 0},$$ normalized according to Eq. (\[scalarspin\]). We also define a state $\ket{-1}\to 0$. Then we repeatedly apply $\hat\epsilon^{LL}$ using the matrix representation (\[epsspin\]) and we orthonormalize the resulting states to the previously obtained states, through Haydock’s recursion $$\label{haydockstep3} b_{n+1}\ket{n+1}=\hat \epsilon^{LL}\ket{n}-a_n\ket{n}-b_{n}\ket{n-1},$$ where we demand $$\label{orthonorm1} \braket{n|m}=\delta_{nm}$$ using the product (\[scalarspin\]). Thus, $$\label{an} a_n=\braket{n|\hat\epsilon^{LL}|n}$$ and $$\label{bnp1} \begin{split} b_{n+1}^2=&(\bra{n}\hat\epsilon^{LL}-a_n\bra{n}-b_{n}\bra{n-1}) \\ &(\hat\epsilon^{LL}\ket{n}-a_n\ket{n}-b_{n}\ket{n-1}). \end{split}$$ We remark that the symmetry of $\hat\epsilon^{LL}$ guarantees that the coefficient of $\ket{n-1}$ is $b_n$, that there are no more terms in Eq. (\[haydockstep3\]) and that the resulting state $\ket{n+1}$ is implicitly orthogonal to all previous states $\ket{0}\ldots\ket{n-1}$ even though we only orthogonalize it explicitly to $\ket{n}$, except for the accumulation of numerical errors, which would have to be handled in the implementation [@simon_analysis_1984; @horst_d._simon_lanczos_1984; @z._bai_lanczos_2000]. In analogy to Ref. [@mochan_efficient_2010], the products by $\hat{\bm G}_{\pm\bm k}$ and $\hat{\bm G}'_{\pm\bm k}$ in Eq. (\[phiepspsis\]) may be performed in reciprocal space, while the convolution with $\epsilon_{\bm G-\bm G'}$ may be replaced by a simple multiplication with $\epsilon(\bm r)$ in real space, so that we may apply the operator $\hat\epsilon^{LL}$ without involving any large matrix product. The Haydock coefficients in Eqs. (\[an\]) and (\[bnp1\]) are not guaranteed to be real and positive as those in Eq. (\[haydockstep\]) and may be complex valued. As in Ref. [@mochan_efficient_2010], in the orthonormal basis $\{\ket{n}\}$ the microscopic longitudinal dielectric function is represented by a tridiagonal symmetric matrix $$\label{tridiag} \hat\epsilon^{LL}\to T_{nn'}=\left( \begin{array}{ccccc} a_0&b_1& 0& 0& \ldots\\ b_1&a_1&b_2& 0& \ldots\\ 0 &b_2&a_2&b_3& \ldots\\ \vdots&\vdots&\vdots&\vdots&\ddots \end{array} \right),$$ from which Eq. (\[epsM\]) allows to extract the macroscopic response $$\label{epsM2} \epsilon_M^{LL}= \left(a_0 - \frac{b_1^2}{a_1- \frac{b_2^2}{a_2-\frac{b_3^2}{\ddots}}}\right).$$ Notice that Eq. (\[epsM2\]) seems identical to (\[epsM1\]), but its Haydock coefficients are different, as they are obtained by using spinor-like states and an Euclidean metric. Thus, Eq. (\[epsM2\]) may be used for arbitrary compositions, including multiple disperssive and dissipative media or even a continuosly varying complex response $\epsilon(\bm r)$. By identifying the longitudinal displacement field $\bm D^L$ with an external macroscopic field and thus with no spatial fluctuations, we may represent it in Haydock’s basis as a column vector with components $\bm D^L\to d_n=D^L \delta_{n0}$. We may expand the longitudinal electric field in the same basis as $\bm E^L\to e_n$, and solve the tridiagonal system $$\label{TEvsD} \sum_{n'} T_{nn'} e_{n'}=d_{n'}$$ for the unknowns $e_n$ to obtain a representation of the microscopic electric field $\bm E^L\to\sum e_n\ket{n}$ which may be translated into reciprocal or real space to obtain $\bm E^L(\bm k+\bm G)$ or $\bm E^L(\bm r)$. We have implemented the formalism above as a set of modules written in the Perl programming language, using its Perl Data Language (PDL) [@glazebrook97pdl] extension for efficient numerical calculations, and the Moose [@moose] object system, and we have incorporated them into the publicly available package [ *Photonic*]{} [@photonic]. Coated clylinders {#sscoated} ----------------- A simple system to test our approach above is that of a lattice of coated cylindrical particles. Consider a single multilayered cylindrical particle with a core ($p=1$) covered by $N-1$ coaxial shells ($p=2\ldots N$) within vacuum ($p=N+1$). Each layer is characterized by an outer radius $a_p$ and a dielectric function $\epsilon_p$. The system is subject to an external field $\bm E_{\text{ex}}=E_{\text{ex}} \hat x$. The potential within each layer may be written as $$\label{layer} \phi_p(\bm r)=(A_p r + B_p/r)\cos\theta$$ in polar coordinates, where, using the symmetry of the system, we restricted ourselves to the angular momentum $l=1$ of the external potential. The boundary conditions at the $p$-th boundary may be written as $$\label{transfer} \left( \begin{array}{c} A_{p+1}\\B_{p+1} \end{array} \right) = \mathcal M_p \left( \begin{array}{c} A_{p}\\B_{p} \end{array} \right),$$ where we introduced the [*transfer matrix*]{} $$\label{transferM} M_p=\frac{1}{2} \left( \begin{array}{cc} \frac{\epsilon_{p+1} + \epsilon_{p}} {\epsilon_{p+1}} & \frac{\epsilon_{p+1} - \epsilon_{p}} {a_p^2\epsilon_{p+1}}\\ \frac{\epsilon_{p+1} - \epsilon_{p}} {\epsilon_{p+1}}a_p^2 & \frac{\epsilon_{p+1} + \epsilon_{p}} {\epsilon_{p+1}} \end{array} \right).$$ Using Eq. (\[transfer\]) repeatedly, we may relate $$\label{Np1} \left( \begin{array}{c} A_{N+1}\\B_{N+1} \end{array} \right) = \mathcal M \left( \begin{array}{c} A_{1}\\B_{1} \end{array} \right),$$ with $\mathcal M=\mathcal M_N\mathcal M_{N-1}\ldots\mathcal M_2\mathcal M_1$. As we may identify $A_{N+1}=-E_{\text{ex}}$ and $B_{N+1}=2p$, with $p$ the total dipole moment per unit length, and as $B_1=0$ to avoid a singularity at $r=0$, from Eq. (\[Np1\]) we may obtain the polarizability per unit length of the particle $$\label{alpha} \alpha=\frac{p}{E_{\text{ex}}}=-\frac{\mathcal M_{21}}{2\mathcal M_{11}}.$$ For a square array of such coated cylinders we may approximate the macroscopic dielectric response through the Claussius-Mossotti 2D relation $$\label{epsM3} \epsilon_M=\frac{1+2\pi n\alpha}{1-2\pi n\alpha}$$ where $n$ is the number density. We expect this expression to hold as long as the distance between cylinders is not so short as to allow exciting multipoles higher than the dipole. Keller’s theorem {#ssKeller} ---------------- In order to further test our result (\[epsM2\]) we will show below that they satisfy a generalization of Keller’s theorem [@ortiz_kellers_2018] for multicomponent metamaterials, which we prove in a simple (limited) form below. Consider a 2D metamaterial with three or more components $A$, $B$, $C$…, each characterized by a dielectric function $\epsilon_A$, $\epsilon_B$, $\epsilon_C$…Then, we write its dielectric function as $$\label{epsABC} \epsilon\equiv\epsilon(\epsilon_A,\epsilon_B,\epsilon_C,\ldots) = \epsilon_A A + \epsilon_B B + \epsilon_C C + \ldots,$$ where we introduced characteristic functions $A(\bm r)$, $B(\bm r)$, $C(\bm r)$…, that take the value 1 when $\bm r$ lies within the corresponding region $A$, $B$, $C$…and 0 otherwise. We expect the use of the same letters to denote materials, regions and characteristic functions will not be confusing, as their use may be distinguished by context. In the absence of external charge and neglecting retardation, the microscopic displacement and electric fields are solenoidal and irrotational respectively, $$\label{divD} \begin{array}{cc} \nabla\cdot \bm D=0,&\nabla\times\bm E=0, \end{array}$$ and they are related through $$\label{DvsE} \bm D=\epsilon(\epsilon_A,\epsilon_B,\epsilon_C\ldots)\bm E=\epsilon\bm E.$$ Now consider the rotated fields $\bm D^R(\bm r)\equiv \mathcal R \bm D(\bm r)$ and $\bm E^R(\bm r)\equiv \mathcal R \bm E(\bm r)$, where $$\label{rotation} \mathcal R\equiv\left( \begin{array}{cc} 0&1\\-1&0 \end{array}\right)$$ is a rotation matrix by $\pi/2$. Notice that we rotate the fields but not their application point $\bm r$. As $\mathcal R$ coincides with the Levi-Civita symbol in 2D, then $$\label{divER} \begin{array}{cc} \nabla\cdot \bm E^R=0,&\nabla\times\bm D^R=0, \end{array}$$ which are similar to Eqs. (\[divD\]) but with the substitutions $$\label{subst} \begin{array}{cc} \bm D\to\tilde {\bm D}\equiv\bm E^R,&\bm E\to\tilde{\bm E}\equiv \bm D^R. \end{array}$$ Notice that $$\label{ERvsDR} \tilde{\bm D} = \epsilon(1/\epsilon_A,1/\epsilon_B,1/\epsilon_C\ldots)\tilde{\bm E}\equiv \tilde\epsilon \tilde{\bm E}.$$ Thus, $\tilde{\bm D}$ and $\tilde{\bm E}$ obey the same equations as $\bm D$ and $\bm E$ but they are related through a microscopic dielectric response $\tilde\epsilon$ identical to that in Eq. (\[epsABC\]) but for the replacements $\epsilon_A\to1/\epsilon_A$, $\epsilon_B\to1/\epsilon_B$, $\epsilon_C\to1/\epsilon_C$… Through a homogenization procedure, such as using Eq. (\[epsM\]), from Eqs. (\[DvsE\]) and (\[ERvsDR\]) we obtain $\bm D_M=\bm\epsilon_M\bm E_M$ and $\tilde{\bm D}_M = \tilde{\bm \epsilon}_M \tilde{\bm E}_M$, where $\bm\epsilon_M$ is the macroscopic dielectric tensor of the original system and $\tilde{\bm \epsilon}_M$ is the corresponding response of the system obtained from the original one by replacing the response of each component by its inverse. Then we may write $$\label{preK} \begin{split} \bm E_M=&\mathcal R^{-1}\tilde{\bm D}_M= \mathcal R^{-1}\tilde{\bm \epsilon}_M\tilde{\bm E}_M \\ &= \mathcal R^{-1}\tilde{\bm \epsilon}_M\mathcal R{\bm D_M}= \mathcal R^{-1}\tilde{\bm \epsilon}_M\mathcal R\bm \epsilon_M{\bm E_M}, \end{split}$$ and cancelling $\bm E_M$ we finally obtain $$\label{Keller} \tilde{\bm\epsilon}_M^R\bm \epsilon_M=\bm 1,$$ where $\tilde{\bm \epsilon}_M^R=\mathcal R^{-1}\tilde{\bm\epsilon}_M\mathcal R$. Thus the original macroscopic response times the rotated macroscopic response of the system with the reciprocal dielectric functions yields the unit tensor. This is Keller’s theorem for multicomponent metamaterials in 2D. Mortola and Steffe’s theorem {#ssmortola} ---------------------------- Consider now a 2D system made of a square lattice whose unit cell is divided into four identical squares that are occupied from left to right, top to bottom, by four materials $A$, $B$, $C$, $D$. Mortola and Steffe proposed an expression [@s._mortola_two-dimensional_1985] for the macroscopic conductivity of this system in terms of the conductivities of its components. This expression was later proved by Milton [@milton_proof_2001]. However, as argued in [@ortiz_kellers_2018], we expect that the correct expression for finite frequencies is that written in terms of the dielectric response, i.e., $$\label{Mortola} \begin{split} \epsilon_M^{xx}=&\{[(\epsilon_A+\epsilon_C) (\epsilon_B+\epsilon_D) (\epsilon_A\epsilon_B\epsilon_C + \epsilon_B\epsilon_C\epsilon_D \\ &+ \epsilon_C\epsilon_D\epsilon_A + \epsilon_D\epsilon_A\epsilon_B)]/[(\epsilon_A+\epsilon_B) (\epsilon_C+\epsilon_D) \\ &\times (\epsilon_A+\epsilon_B +\epsilon_C+\epsilon_D)]\}^{1/2}. \end{split}$$ A similar expression holds for $\epsilon_M^{yy}$ obtained from Eq. (\[Mortola\]) by exchanging $B\leftrightarrow C$. Results ======= In Figure \[fcoated\] we show the imaginary part of the dielectric function of a square lattice of thin coated and uncoated Ag cylinders of radius $a_{\text{Ag}}=0.1L$ within vacuum, where $L$ is the lattice parameter. As the cylinders are very thin, their mutual interaction is negligible. Thus, in the case of the uncoated cylinders, there is a peak around $\hbar\omega\approx 3.7$eV which corresponds to the surface plasmon of an isolated Ag cylinder, given by $\epsilon_{\text {Ag}}=-1$. If the cylinder is coated by a SiO$_2$ layer of outer radius $a_{\text{SiO}_2}=0.25L$ the peak is redshifted. The analytical result based on the Claussius-Mossotti relation using the polarizability given by Eq. (\[alpha\]) based on a transfer matrix formalism agrees quite closely with the numerical calculation based on Haydock’s recursion for the case of coated cylinders and is indistinguishable for the case of uncoated cylinders. The numerical calculation was done using a $401\times401$ grid and with up to 200 pairs of Haydock coefficients. {width=".7\textwidth"} {width=".7\textwidth"} Fig. \[fglassAg\] shows the real and imaginary parts of the macroscopic dielectric function for a system similar to that in Fig. \[fcoated\] but with an SiO$_2$ core of radius $a_{\text{SiO}_2}=0.3L$ covered by an Ag shell of outer radius $a_{\text{Ag}}=0.45L$. As neighboring cylinders are closer together than in Fig. \[fcoated\] dipolar and higher multipoles may couple together. Thus, the extension (\[epsM3\]) of the Claussius-Mossotti formalism may not be accurate. The response obtained from the numerical calculation has a peak around 1.92eV further red-shifted from that of the isolated cylinder than the peak of the analytical calculation around 2.04eV. Nevertheless, a numerical calculation based on Keller’s theorem, Eq. (\[Keller\]), obtained by inverting the dielectric functions of the components, calculating the corresponding dielectric function using our recursive formalism and inverting the result, seems to agree perfectly with the straightforward numerical calculation. Thus, our recursive procedure agrees with Keller’s theorem even for large inclusions and strong interactions. In Fig. \[fmortola\] we show the real and imaginary parts of a component $\epsilon_M^{xx}$ of the macroscopic dielectric tensor of a metamaterial made up of four materials, Au, Ag, TiO$_2$ and SiO$_2$ filling square prisms occupying a 2x2 block and repeated periódically in a checkerboard geometry, as illustrated in Fig. \[fcheckerboard\]. The calculation was performed using the procedure described in Subsec. \[ssmulti\] using a $201 \times 201$ grid and with up to 300 Haydock coefficient pairs. In the figure we also show the results of an analytical calculation using the formula presented in Subsec. \[ssmortola\]. {width=".7\textwidth"} ![ Geometry of the metamaterial corresponding to Fig. \[fmortola\], consisting of the periodic repetition of a unit cell made up of four square prisms of Au, Ag, TiO$_2$ and SiO$_2$.[]{data-label="fcheckerboard"}](paper-0){width=".4\textwidth"} Notice the good agreement for both the real and imaginary parts for a wide energy range. Conclusions =========== We have developed a recursive procedure based on a Haydock’s representation that allows the efficient calculation of the macroscopic dielectric function and the microscopic fields of multicomponent metamaterials of arbitrary composition and geometry. Our formalism admits materials that can be insulating, conducting, transparent, opaque, dissipative, and/or dispersive. Although the response of the system may be non-Hermitian, we could take advantage of its symmetric nature by introducing an appropriate scalar product and using a spinor-like representation of the fields. Though efficient, the procedure developed here is not as fast as that for only two materias, as in the current case the Haydock coefficients depend on the composition and not only on the geometry. The results presented here correspond to the non-retarded limit, though we have verified that the same ideas may be extended to the retarded region where they may even be applied to chiral systems. We have prepared computational modules implementing our procedures and added them to a publicly available software package. We tested our formalism by calculating the response of simple systems for which approximate analytical formulae are available, and by demonstrating that our results are consistent with some exact conditions, namely, Keller’s and Mortola and Steffe’s theorems. This work was supported by DGAPA-UNAM under grant IN111119. RS acknowledges a scholarship from CONACyT. Graphical Table of Contents\ {#graphical-table-of-contents .unnumbered} ============================ GTOC image: ![ We develop an efficient procedure to compute recursively the fields and the response of metamaterials made of any number of components with an arbitrary geometry and composition, metallic or dielectric, transparent or dissipative, using a spinor-like representation of the field states. We test the procedure against systems with analytically determined properties. The figure shows the field of a square array of Ag covered SiO$_2$ cylinders close to resonance. []{data-label="GTOC"}](field){width="4cm" height="4cm"}

ArXiv

--- abstract: 'We report on a search for SM Higgs Boson in the Higgs to two photons decay channel conducted by the CMS experiment with the data accumulated during the 2010 and 2011 running of the LHC at sqrt(s) = 7 TeV.' author: - Christopher Palmer title: 'A Search for the Higgs Boson in the $H\rightarrow\gamma\gamma$ Channel with CMS DPF-2011 Proceedings' --- Introduction - $H\rightarrow\gamma\gamma$ in a Nutshell ======================================================= We are motivated to explore this signal because of its narrow resonance over a smoothly falling background. Despite the enormous background and the somewhat low number of expected events, this decay channel is extremely relevant to the favored, low-mass Higgs search as long as the resolution of the peak can be measured with reasonable resolution. Our background can be categorized in two components: irreducible and reducible. Irreducible background is more difficult to eliminate because these are events which have two real photons in them. Reducible background events are typically poorly reconstructions electrons or jets which are mistakenly taken to be photons. In the final selection invariant mass plot ( Figure \[mgg\]) one can see the various components from each. ![Final invariant mass selection with simulation.[]{data-label="mgg"}](mgg_label.png){width="11cm"} Necessary CMS Design/ECAL ========================= Compact Muon Solenoid (CMS) --------------------------- CMS is a general purpose detector. It is designed to detect and reconstruct numerous physics objects (e.g. photons, electrons, muons, hadronic jets, etc). From this general design we are compatible of seeking out new physics (e.g. Higgs and SUSY). Below ( Figure \[cms\_xsec\]) is a cross sectional schematic of the detectored with the relevant components highlighted (i.e. the trackers and the electromagnetic calorimeter (ECAL)). ![Cross sectional schematic of CMS.[]{data-label="cms_xsec"}](cms_slice_spotlight.png){width="11cm"} CMS - Electromagnetic Calorimeter (ECAL) ---------------------------------------- The ECAL is made of $\sim$76K $PbWO_{4}$ crystals in the barrel ($|\eta|<1.48$) and the endcap ($1.48<|\eta|<3.$) of CMS. It was reconstructed in order to determine the energy of photons and electrons to high precision. The design resolution of unconverted photons in the barrel with energy greater than 100 GeV was $\sim$0.5% There are two main critical issues that impact the resolution that CMS is able to determine. One is the calibration of the crystals. Two techniques, which are exploited, are $\pi_{0}\rightarrow\gamma\gamma$ events for inter-crystal calibration and $Z\rightarrow e^{+}e^{-}$ for global energy scale calibration. The other critical issue is transparency corrections for radiation damage. An integrated laser system measures the transparency of crystals and that information is used to correct the transparency loss. Below is a schematic of the ECAL ( Figure \[ecal\_sch\]) and a photo of one of the crystals ( Figure \[ecal\_cry\]). ![A photograph of one of the crystals from the ECAL.[]{data-label="ecal_cry"}](crystal_0803027_02-A4-at-144-dpi.jpg){width="5.5cm"} ![A schmatic of CMS’s ECAL.[]{data-label="ecal_sch"}](ecal_colorless.png){width="6cm"} Analysis Strategy ================= DiPhoton Selection ------------------ Photon identification is important for removal of reducible background. The primary tools used to distinguish between signal and background are isolation, cluster shape and electron veto. The cuts set in these photon identification variables are set separately for in categories defined by the photon being in the barrel or one of the endcaps and by whether or not the photon converted to an electron-positron before before reaching the ECAL. The efficiency versus $\eta$ is plotted below for these four categories ( Figure \[mc\_eff\]). Also, note that our selection contains minimum thresholds on the transverse momentum of the two photons (40 and 30 GeV). ![Efficiency of selection in categories as a function of $\eta$ with simulated Higgs’ photons.[]{data-label="mc_eff"}](eff_eta.pdf){width="8cm"} Vertex Selection ---------------- In our data, numerous proton-proton interactions occur each time bunches of protons pass through each other. If the wrong vertex is chosen to reconstruct the photons then the energy of the photons and the mass of the di-photon will be incorrect. Effectively smearing the mass of the Higgs signal. We use the following quantities to increase our probability of determining the correct vertex. For unconverted photons we use $\displaystyle\sum\limits_{Tracks} P_{T}^{2}$, projection of tracks onto $\gamma\gamma$ and the balance between $\gamma\gamma$ and vertex’s tracks. For converted $\gamma$’s the conversion-tracks are used to point back to vertex. The efficiency of this method is plotted below as a function of di-photon transverse momentum ( Figure \[vertex\_eff\]). ![Efficiency of vertex selection within 1 cm as function of Higgs’ transverse momentum.[]{data-label="vertex_eff"}](Vertex_eff.pdf){width="4.3cm"} Efficiencies from Data ====================== Making the basic assumption that electrons and photons have similar shower properties we use the tag and probe method with $Z\rightarrow e^{+}e^{-}$ electrons to determine the efficiency of our photon selection. First events with robustly identified electrons are selected and we seek an additional electron passing a minimal transverse energy requirement, the probe. The photon reconstructed with the same energy deposit has the photon cuts applied to it and efficiency of passing events to selected events is measured in each category. Below is a table of these efficiencies. Please note the majority of our photons are in the unconverted, barrel category, which is more than 90% efficient. To measure the efficiency of the electron veto cuts, photons from $Z\rightarrow \mu^{+}\mu^{-}\gamma$ events have the selection applied. Efficiencies are listed in the table below. [| c | c | c | c|]{} Category & $\epsilon_{data}$ (%) & $\epsilon_{MC}$ (%) & $\epsilon_{data}$/$\epsilon_{MC}$\ \ 1 & 91.77$\pm$0.14 & 92.43$\pm$0.07 & 0.993$\pm$0.002\ 2 & 72.67$\pm$0.43 & 71.89$\pm$0.08 & 1.011$\pm$0.007\ 3 & 80.33$\pm$0.47 & 80.04$\pm$0.18 & 1.004$\pm$0.008\ 4 & 57.80$\pm$1.26 & 55.09$\pm$0.15 & 1.049$\pm$0.025\ \ 1 & 99.78$^{+0.13}_{-0.16}$ & 99.59$^{+0.13}_{-0.17}$ & 1.002$^{+0.002}_{-0.002}$\ 2 & 98.77$^{+0.59}_{-0.73}$ & 97.70$^{+0.32}_{-0.37}$ & 1.011$^{+0.007}_{-0.008}$\ 3 & 99.32$^{+0.51}_{-1.02}$ & 99.29$^{+0.30}_{-0.42}$ & 1.000$^{+0.006}_{-0.011}$\ 4 & 93.0$^{+2.1}_{-2.3}$ & 93.34$^{+0.79}_{-0.86}$ & 0.996$^{+0.024}_{-0.027}$\ \[example\_table\] [| c | c |]{}\ 2 Unconverted & 1,2 Converted\ 100.00$^{+0.00}_{-0.01}$% & 99.53$\pm$0.04%\ \ 2 Unconverted & 1,2 Converted\ 100.00$^{+0.00}_{-0.02}$% & 98.86$\pm$0.07%\ \[example\_table\] Resolution Results ================== ECAL resolution measured from $Z\rightarrow e^{+}e^{-}$ ( Figure \[Zee\_res\])is applied to simulated Higgs’ $\gamma$’s ( Figure \[Hgg\_res\]). The simulated Higgs’ $\gamma$’s with data resolution are used in signal modeling for CL limits. Suboptimal transparency loss corrections may be responsible for degraded resolution. An example of the measured resolution and its application to the best Higgs’ photons category are below. ![Resolution from $Z\rightarrow e^{+}e^{-}$ events in data.[]{data-label="Zee_res"}](out-all_R9_golden-Histos-sum_region_0_0_boldZee.png){width="5.5cm"} ![Signal resolution in $H\rightarrow\gamma\gamma$ estimated from data.[]{data-label="Hgg_res"}](resolution_hipt_hir9_EB_boldHgg.png){width="5.5cm"} Cateogories and Limit Setting ============================= Event Classes Used for CL Evaluation ------------------------------------ There are eight event classes in which confidence levels are computed. There are both photons in the barrel or either in an endcap times both converted or either unconverted times high/low diphoton transverse momentum classifications. Below is a table which contains the fraction of signal and background in each event class. ---------------------------------------------------- --------------- --------------- --------------- --------------- 2 Unconverted 1,2 Converted 2 Unconverted 1,2 Converted [**$\mathbf{P_{T}^{\gamma\gamma} <}$ 40 GeV/c**]{} Signal 0.209 0.271 0.094 0.116 Background 0.167 0.263 0.129 0.203 Signal $\sigma_{effective}$ (GeV/$c^{2}$) 1.58 2.33 3.14 3.60 [**$\mathbf{P_{T}^{\gamma\gamma} >}$ 40 GeV/c**]{} Signal 0.102 0.122 0.035 0.051 Background 0.043 0.079 0.043 0.074 Signal $\sigma_{effective}$ (GeV/$c^{2}$) 1.37 2.12 2.95 3.26 ---------------------------------------------------- --------------- --------------- --------------- --------------- \[example\_table\] Second order polynomial fits are performed in each of the eight categories on the data. The resulting fit is the background distribution used for toy experiments. The two both barrel, both unconverted categories are below ( Figure \[datafit1\] and  Figure \[datafit2\]). ![High $P_{T}^{\gamma\gamma}$, both high $R_{9}$, barrel-barrel .[]{data-label="datafit1"}](datafit_wSig_hipt_hir9_EB.pdf){width="5.5cm"} ![Low $P_{T}^{\gamma\gamma}$, both high $R_{9}$, barrel-barrel .[]{data-label="datafit2"}](datafit_wSig_lopt_hir9_EB.pdf){width="5.5cm"} Systematics ----------- Below ( Figure \[syst\]) is a table that summarizes the systematics that are applied to the signal models. Since the background model is from data, no systematics are applied to it. {width="11.cm"} \[syst\] Evaluated Limits ---------------- Confidence Level (CL) limits are determined in two ways with extremely consistent results. Our official limits are set with the modified frequentist approach (CLs) using profile likelihood. As a cross check with all use the Bayesian method with flat prior in cross section. In the following two plots show SM exclusions between cross sections of between 0.06 and 0.26 pb ( Figure \[sm\_ex\]) and between 1.9 and 7.0 $*\sigma_{SM}$ ( Figure \[sm\_ex\_rel\]). While the following plot shows Fermiophobic Higgs cross section excluded between 0.04 and 0.18 pb ( Figure \[fp\_ex\]) and its mass constrained to be greater than 111 GeV. {width="9cm"} \[sm\_ex\] {width="9cm"} \[sm\_ex\_rel\] {width="9cm"} \[fp\_ex\] Conclusions/Outlook =================== Our analysis is defined by our use of photon selection in categories and our use of various methods to select the best vertex possible. Data is used to determine the resolution of our signal (via $Z\rightarrow e^{+}e^{-}$ events) model and our background model is determined directly from the fit of the data in event classes. The CL evaluation in event classes improves our sensitivity to both the SM and Fermiophobic Higgs. We are keen to improve this analysis by taking as much data as possible and by improving the resolution of CMS’s ECAL. Improving resolution could be the signal greatest improvement to the SM analysis and work is ongoing in this field. Vivek Sharma and Christoph Paus have put forth great effort in coordinating the Higgs effort within CMS over the past year. The Higgs result from CMS would not have been achieved without them. Christopher Seez and Paolo Meridiani have been directing the research specifically outlined in this proceeding and should be commended for their efforts and the group’s achievements. Finally, I would like to personally thank my advisor Jim Branson and Marco Pieri for their direction of my research. [9]{} CMS PAS HIG-11-010 ($H\rightarrow\gamma\gamma$ PAS) - dsweb.cern.ch/record/1369553/files/HIG-11-010-pas.pdf CMS PAS HIG-11-011 (Higgs combination PAS - cdsweb.cern.ch/record/1370076/files/HIG-11-011-pas.pdf Other public plots - https://twiki.cern.ch/twiki/bin/view/CMSPublic/Hig11010TWiki

ArXiv

--- abstract: 'Abstract. We define bilateral series related to Ramanujan-like series for $1/\pi^2$. Then, we conjecture a property of them and give some applications.' address: 'Univ. de Zaragoza (Spain)' author: - Jesús Guillera title: 'Bilateral sums related to Ramanujan-like series' --- Introduction ============ In [@GuiRo] we constructed bilateral summations related to Ramanujan-type series for $1/\pi$ and applied them to prove a new kind of identities, which we called “the upside-down" counterpart. In this paper we consider bilateral sums related to Ramanujan-like series for $1/\pi^2$, conjecture a property of them and give some applications. We will need the following theorems \[per\] Suppose that $f(x)$ is the sum over all integers $n$ of $g(n+x)$. Then, clearly $f(x)$ is a periodic function of period $x=1$. \[fou\] Let $f: \mathbb{C} \longrightarrow \mathbb{C}$ be a holomorphic function such that $f(x)=\mathcal{O}(e^{c \pi |{\rm Im}(x)|})$, where $c \geq 0$ is a constant, and suppose that $f(x)$ admits a Fourier series: $$f(x)=\sum_{n=-\infty}^{+\infty} a_n e^{2 \pi i n x}.$$ Then $|2n| > c \, \Rightarrow a_n=0$. It is known that the coefficients of a Fourier series are given by $$a_n=\int_{a+it}^{a+1+it} f(x) e^{-2 \pi i n x} dx.$$ Then for $n<0$ let $t \to+\infty$ and for $n \geq 0$ let $t \to -\infty$. Ramanujan-like series for $1/\pi^2$ =================================== Let $s_0=1/2 \,$, $s_3=1-s_1 \,$, $s_4=1-s_2 \,$. We recall that a Ramanujan-like series for $1/\pi^2$ is a series of the form $$\label{gui-general} \sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_i)_n}{(1)_n}\right] (a+bn+cn^2) z^n =\frac{1}{\pi^2},$$ where $z$, $a$, $b$ and $c$ are algebraic numbers and the possible couples $(s_1, s_2)$ are $(1/2,1/2)$, $(1/2,1/3)$, $(1/2,1/4)$, $(1/2,1/6)$, $(1/3,1/3)$, $(1/3,1/4)$, $(1/3,1/6)$, $(1/4,1/4)$, $(1/4,1/6)$, $(1/6,1/6)$, $(1/5,2/5)$, $(1/8,3/8)$, $(1/10,3/10)$, $(1/12,5/12)$. Up to date $11$ convergent and $6$ “divergent" formulas are known in this new family (see all in the Appendix). The value $$\label{tau} \tau=\frac{c}{\sqrt{1-z}},$$ plays an important roll in the theory [@Gui-matrix eq. 3.47]. Bilateral series related to Ramanujan-like series for $1/\pi^2$ =============================================================== We introduce the function $$\label{rama-like-gui} f(x)=\prod_{i=0}^{4} \frac{\cos \pi x - \cos \pi s_i}{1-\cos \pi s_i} \sum_{n \in \mathbb{Z}} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}}\right] [a+b(n+x)+c(n+x)^2] (-z)^{n+x}.$$ Applying Theorem \[per\], we see that it is a periodic function of period $x=1$ because $\cos \pi x - \cos \pi s_0 =\cos \pi x$ and $(-1)^n \cos \pi x = \cos \pi (n+x)$. Besides it is analytic due to the factor with the cosenus, which cancels the poles at $x=-s_i$ and therefore all the other poles due to the periodicity of $f(x)$. On the other hand we conjecture that $f(x)=\mathcal{O}(e^{5 \pi |{\rm Im}(x)|})$. Hence by Theorem \[fou\] we deduce that $$\begin{gathered} \label{bilateral-general} \prod_{i=0}^{4} \frac{\cos \pi x - \cos \pi s_i}{1-\cos \pi s_i} \sum_{n \in \mathbb{Z}} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}}\right] [a+b(n+x)+c(n+x)^2] (-z)^{n+x} \\ = \frac{1}{\pi^2} \left( \frac{}{} u_1 \cos 2\pi x + u_2 \cos 4\pi x + (1-u_1-u_2) + v_1 \sin 2 \pi x + v_2 \sin 4 \pi x \right),\end{gathered}$$ where $u_1$, $u_2$, $v_1$, $v_2$ do not depend on $x$, and we can determine their values by giving values to $x$. We conjecture that for the bilateral series corresponding to a Ramanujan-like series for $1/\pi^2$ (that is, when $z$, $a$, $b$, $c$ are algebraic numbers), the values of $u_1$, $u_2$, $v_1$ and $v_2$ are rational numbers. In addition, for the case of alternating series we conjecture that $v_1=v_2=0$. These bilateral series are “divergent", because the sum to the left or the sum to the right “diverges". However we turn them convergent by analytic continuation. We can write (\[bilateral-general\]) in the following way: $$\begin{gathered} \label{bilateral-general-split} \sum_{n=0}^{\infty} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i+x)_{n}}{(1+x)_{n}} \right] [a+b(n+x)+c(n+x)^2] (-z)^{n+x} \\ + x^5 \sum_{n=1}^{\infty} (-1)^n \left[\prod_{i=0}^{4} \frac{(1-x)_{n}}{(s_i-x)_{n}} \right] \frac{a+b(-n+x)+c(-n+x)^2}{(-n+x)^5} (-z)^{-n+x} = \\ \frac{1}{\pi^2} \left[\prod_{i=0}^{4} \frac{(1)_x}{(s_i)_{x}} \right] \frac{u_1 \cos 2\pi x + u_2 \cos 4\pi x + (1-u_1-u_2) + v_1 \sin 2 \pi x + v_2 \sin 4 \pi x}{\csc^2 \pi s_1 \csc^2 \pi s_2 \cos \pi x (\cos^2 \pi x - \cos^2 \pi s_1)(\cos^2 \pi x - \cos^2 \pi s_2)},\end{gathered}$$ where we have splitted the summation in two sums and used the properties $$\frac{1}{(1)_{-n+x}}=(-1)^{n+1} \frac{(1-x)_n}{(1)_x} \frac{x}{n-x}, \quad (s_i)_{-n+x}=(-1)^n \frac{(s_i)_x}{(1-s_i-x)_n}.$$ As the second sum is $\mathcal{O}(x^5)$, formula (\[bilateral-general-split\]) implies an expansion of the form $$\sum_{n=0}^{\infty} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i+x)_{n}}{(1+x)_{n}} \right] [a+b(n+x)+c(n+x)^2] (-z)^{n+x} = \frac{1}{\pi^2} - k \frac{x^2}{2!} + j \pi^2 \frac{x^4}{4!}+ \mathcal{O}(x^5),$$ if $z<0$ (alternating series), and an expansion of the form $$\sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_i+x)_{n}}{(1+x)_{n}} \right] [a+b(n+x)+c(n+x)^2] z^{n+x} = \frac{1}{\pi^2} - k \frac{x^2}{2!} + j \pi^2 \frac{x^4}{4!}+ \mathcal{O}(x^5),$$ if $0<z<1$ (convergent series of positive terms). Hence, the conjecture that the coefficients of the Fourier expansion are rational is equivalent to the conjecture that $k$ and $j$ are rational stated in [@Gui-matrix; @AlGu]. The relation of $\tau$ with $k$ and $j$ is [@Gui-matrix eq. 3.48]: $$\tau^2=\frac{j}{12}+\frac{k^2}{4}+\frac{5k}{3}+1+(\cot^2 \pi s_1)(\cot^2 \pi s_2)+(1+k)(\cot^2 \pi s_1+\cot^2 \pi s_2).$$ From (\[bilateral-general-split\]) we deduce that $$\sum_{n \in \mathbb{Z}} \left[\prod_{i=0}^{4} \frac{(s_i)_n}{(1)_n}\right] (a+b n+c n^2) z^n= \sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_i)_n}{(1)_n}\right] (a+b n+c n^2) z^n,$$ and as the function (\[rama-like-gui\]) has period $x=1$, we see that for all integer $x$, we have $$\begin{gathered} \label{x-integer} \sum_{n \in \mathbb{Z}} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}} \right] [a+b(n+x)+c(n+x)^2] (-z)^{n+x} \\ = (-1)^x \sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_i)_n}{(1)_n}\right] (a+b n+c n^2) z^n.\end{gathered}$$ If $z>0$, we can write $$\begin{gathered} \label{z-mayor0} \sum_{n \in \mathbb{Z}} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}} \right] [a+b(n+x)+c(n+x)^2] (-z)^{n+x} \\ =e^{-i \pi x} \sum_{n \in \mathbb{Z}} \left[\prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}} \right] [a+b(n+x)+c(n+x)^2] z^{n+x}.\end{gathered}$$ Taking into account (\[x-integer\]) and (\[z-mayor0\]), we have written the following Maple procedure: bilater:=proc(s1,s2,z0,x,j) local p,s0,s3,s4,z1,h,hh,r,u,v; p:=(a,b)->pochhammer(a,b): s0:=1/2; s3:=1-s1; s4:=1-s2: h:=n->p(s0,n)*p(s1,n)*p(s2,n)*p(s3,n)*p(s4,n)/p(1,n)^5: if type(x,integer) then; hh:=n->h(n)*n^j: r:=evalf((-1)^x*subs(z=z0,sum(hh(n)*z^n,n=0..infinity))): return r; else; if z0<0 then; z1:=z0; hh:=(j,n)->(-1)^n*h(n+x)*(n+x)^j; else z1:=-z0; hh:=(j,n)->exp(-I*Pi*x)*h(n+x)*(n+x)^j; fi; u:=j->subs(z=z1,sum(hh(j,n)*(-z)^(n+x),n=0..infinity)): v:=j->subs(z=z1,sum(hh(j,-n)*(-z)^(-n+x),n=1..infinity)): r:=evalf(u(j)+v(j)): return r: fi; end: This procedure calculates the bilateral sum $$\sum_{n \in \mathbb{Z}} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}} \right] (-z)^{n+x} \, (n+x)^j,$$ and helps the reader to check the examples. Examples of bilateral series ============================ We give some examples of bilateral series related to Ramanujan-like series for $1/\pi^2$. Formula (\[f3\]) in the Appendix is $$\label{gui-1} \frac18 \sum_{n=0}^{\infty} (-1)^n \frac{(\frac12)_n^5}{(1)_n^5}\frac{20n^2+8n+1}{4^n} = \frac{1}{\pi^2}.$$ Taking $v_1=v_2=0$ and giving two values to $x$ we get numerical approximations of $u_1$, $u_2$. We observe that these values look rational and do not change using other values of $x$. Hence, we conjecture that $$\label{Z-gui-1} \frac18 \sum_{n \in \mathbb{Z}} (-1)^n \frac{(\frac12)_{n+x}^5}{(1)_{n+x}^5}\frac{20(n+x)^2+8(n+x)+1}{4^{n+x}}= \frac{1-\frac12 \cos 2\pi x + \frac12 \cos 4\pi x}{\pi^2 \cos^5 \pi x},$$ after replacing $u_1$, $u_2$ with their guessed rational values. \[other-example\] Formula (\[f10\]) in the Appendix is $$\sum_{n=0}^{\infty} (-1)^n \frac{\left(\frac12\right)_n\left(\frac13\right)_n \left(\frac23\right)_n\left(\frac16\right)_n\left(\frac56\right)_n}{(1)_n^5} \left( \frac34 \right)^{6n} (1930n^2+549n+45) = \frac{384}{\pi}.$$ From it we conjecture the bilateral form $$\begin{gathered} \frac{1}{384} \sum_{n \in \mathbb{Z}} (-1)^n \frac{\left(\frac12\right)_{n+x}\left(\frac13\right)_{n+x}\left(\frac23\right)_{n+x}\left(\frac16\right)_{n+x}\left(\frac56\right)_{n+x}}{(1)_{n+x}^5} \left( \frac34 \right)^{6(n+x)} \\ \times \left( 1930(n+x)^2+549(n+x)+45 \right) = \frac{11-14 \cos 2\pi x + 6 \cos 4\pi x}{\pi^2 \cos \pi x \, (4\cos^2 \pi x-1)(4\cos^2 \pi x -3)},\end{gathered}$$ after identifying the coefficients. From the following formula: $$\label{gui-16} \sum_{n=0}^{\infty} \frac{\left(\frac12\right)_n^3\left(\frac14\right)_n\left(\frac34\right)_n}{(1)_n^5}\frac{120n^2+34n+3}{16^n} = \frac{32}{\pi^2},$$ which is (\[f2\]) in the Appendix, we get the bilateral form $$\begin{gathered} \frac{1}{32} \sum_{n \in \mathbb{Z}} \frac{\left(\frac12\right)_{n+x}^3\left(\frac14\right)_{n+x}\left(\frac34\right)_{n+x}}{(1)_{n+x}^5} \left( \frac{1}{16} \right)^{n+x} (120(n+x)^2+34(n+x)+3) = \\ e^{i \pi x} \, \frac{3-\frac72 \cos 2 \pi x + \frac32 \cos 4 \pi x + \left( \frac12 \sin 2\pi x - \frac12 \sin 4 \pi x \right) i}{\pi^2 \cos^3 \pi x \, (2\cos^2 \pi x -1)},\end{gathered}$$ after identifying the coefficients of the Fourier expansion from their numerical approximations. Observe that $v_1$ and $v_2$ are not null because the series in (\[gui-16\]) is of positive terms. Observe also the factor $e^{i \pi x}$, which comes from $(-1)^n / (-16)^{n+x}$ For the formula (\[f7\]) in the Appendix we have the bilateral identity $$\begin{gathered} \frac{\sqrt 7}{392} \sum_{n \in \mathbb{Z}} \frac{\left(\frac12 \right)_{n+x} \left( \frac18 \right)_{n+x} \left( \frac38 \right)_{n+x} \left( \frac58 \right)_{n+x} \left( \frac78 \right)_{n+x}}{(1)_{n+x}^5} \frac{1}{7^{4(n+x)}} (1920(n+x)^2+304(n+x)+15) \\ = e^{i \pi x} \, \frac{ 29-\frac{79}{2} \cos 2 \pi x + \frac{23}{2} \cos 4 \pi x + \left( \frac52 \sin 2\pi x - \frac32 \sin 4 \pi x \right) i }{\pi^2 \csc^2 \frac{\pi}{8} \csc^2 \frac{3\pi}{8} \cos \pi x \left( \cos^2 \pi x - \cos^ 2 \frac{\pi}{8} \right)\left( \cos^2 \pi x - \cos^ 2 \frac{3\pi}{8} \right)},\end{gathered}$$ after identifying the coefficients. Looking at the formula (\[f9\]) in the Appendix which has $z=(3/\phi)^3$, we had the intuition that another series with the other sign of the square root, namely $z=-(3\phi)^3$, could exist and we were right because we discovered it by using the PSLQ algorithm. Here we recover it in a different way by writing the bilateral identity $$\begin{gathered} \label{diverg-phi} \sum_{n \in \mathbb{Z}} \frac{\left(\frac12\right)_{n+x}^3\left(\frac13\right)_{n+x}\left(\frac23 \right)_{n+x}}{(1)_{n+x}^5} (-1)^n \left(3 \phi\right)^{3(n+x)} (c(n+x)^2+b(n+x)+a) = \\ \frac{3}{4\pi^2} \frac{u_1 \cos 2 \pi x + u_2 \cos 4 \pi x + (1-u_1-u_2)}{\cos^3 \pi x \left(\cos^2 \pi x - \frac14 \right)}.\end{gathered}$$ Giving five values to $x$ we get numerical approximations of $a$, $b$, $c$, $u_1$ and $u_2$, which we could identify: $$\label{sol-diverg-phi} u_1=\frac{17}{36}, \quad u_2=\frac{3}{16}, \quad c=2408 + \frac{216}{\phi}, \quad b=1800 + \frac{162}{\phi}, \quad a= 333 + \frac{30}{\phi}.$$ Replacing these values and taking $x=0$, we arrive at the “divergent" formula (\[d5\]), in the second list of the Appendix. From the “divergent" series (\[d3\]) in the Appendix, namely $$\sum_{n=0}^{\infty}\frac{\left(\frac12 \right)_n^3 \left( \frac13 \right)_n \left( \frac23 \right)_n}{(1)_n^5} (28n^2+18n+3) (-1)^n 3^{3n} = \frac{6}{\pi^2},$$ we have the following bilateral sum evaluation: $$\begin{gathered} \label{Z-gui-27} \frac16 \sum_{n \in \mathbb{Z}} \frac{(\frac12)_{n+x}^3(\frac13)_{n+x}(\frac23)_{n+x}}{(1)_{n+x}^5} \left[ 28(n+x)^2+18(n+x)+3 \right] \, (-1)^n \, 27^{n+x} \\ = \frac{5 + 4 \cos 2\pi x + 3 \cos 4\pi x}{4 \pi^2 \cos^3 \pi x (4\cos^2 \pi x - 1)},\end{gathered}$$ after identifying the coefficients of the Fourier terminating expansion from their numerical approximations obtained giving values to $x$. Applications of the bilateral series ==================================== If $|z|>1$, then letting $x \to -1/2$ or $x \to -s_1$, etc in (\[bilateral-general-split\]), we obtain the evaluation of some convergent series. In general, taking the limit of (\[bilateral-general-split\]) as $x \to -s_j$, where $j \in \{0,1,2,3,4\}$, we deduce the following identity: $$\begin{gathered} \label{diverg-to converg} \sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_j)_{n}}{(s_i+s_j)_{n}} \right] \left[ a+b(-n-s_j)+c(-n-s_j)^2 \right] z^{-n} = \\ \frac{(-z)^{s_j}}{\pi^2}\lim_{x \to -s_j}\left[\prod_{i=0}^{4} \frac{(1)_x}{(s_i)_{x}} \right] \frac{u_1 \cos 2\pi x + u_2 \cos 4\pi x + (1-u_1-u_2) + v_1 \sin 2 \pi x + v_2 \sin 4 \pi x}{\csc^2 \pi s_1 \, \csc^2 \pi s_2 \, \cos \pi x (\cos^2 \pi x - \cos^2 \pi s_1)(\cos^2 \pi x - \cos^2 \pi s_2)}.\end{gathered}$$ Another application is explained in [@GuiRo], but only proved for bilateral sums related to Ramanujan-type series for $1/\pi$. From the formula (\[Z-gui-27\]), and taking $s_j=1/2$ in (\[diverg-to converg\]), we get $$\sum_{n=0}^{\infty} \frac{\left(\frac12\right)_n^5}{\left(1\right)_n^3 \left(\frac16\right)_n \left(\frac56 \right)_n} \frac{28n^2+10n+1}{6n+1} \left( \frac{-1}{27} \right)^n = \frac{3}{\pi},$$ which is a new convergent series for $1/\pi$. In the same way, from the ‘divergent" series (\[diverg-phi\]-\[sol-diverg-phi\]), we get the convergent evaluation $$\sum_{n=0}^{\infty} \frac{\left(\frac12\right)_n^5}{\left(1\right)_n^3 \left(\frac16\right)_n \left(\frac56 \right)_n} \frac{ (2408 \! + \! \frac{216}{\phi}) n^2 \! + \! (608 + \frac{54}{\phi})n + (35 + \frac{3}{\phi})}{6n+1} \left( \frac{-1}{3\phi} \right)^{3n} = \frac{3\sqrt{\phi^3}}{\pi},$$ where $\phi$ is the fifth power of the golden ratio. From (\[Z-gui-1\]), and letting $s_j \to -1/2$ in (\[diverg-to converg\]), we recover formula (\[d1\]) in the Appendix. We observe that a formula with $z$ implies another one with $z^{-1}$ in the family $s_1=s_2=1/2$, and we will refer to this property as “duality". The mirror map ============== All the parameters of the bilateral sums related to Ramanujan-like series for $1/\pi^2$ are algebraic, but the value of $q$ (related to $z$ by the mirror map) is not. Now we will define a function $w(x)$ in which $q$ is easily related to the coefficients of the Fourier expansion. First, we let $u(x)$ and $v(x)$ be the analytic functions $$u(x) =\prod_{i=0}^{4} \frac{\cos \pi x - \cos \pi s_i}{1-\cos \pi s_i} \sum_{n \in \mathbb{Z}} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}}\right] (-z)^{n+x},$$ and $$v(x) =\prod_{i=0}^{4} \frac{\cos \pi x - \cos \pi s_i}{1-\cos \pi s_i} \sum_{n \in \mathbb{Z}} (-1)^n \left[\prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}}\right] (n+x) (-z)^{n+x}.$$ Then, we define the analytic function $$w(x) = v(0)u(x)-u(0)v(x).$$ We have the following Fourier expansion: $$w(x)=\frac{}{} a_1 \cos 2\pi x + a_2 \cos 4\pi x + (1-a_1-a_2) + b_1 \sin 2 \pi x + b_2 \sin 4 \pi x,$$ The above identity for the function $w(x)$ implies an expansion of the form $$\begin{gathered} \label{q-alternating} \sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_i)_{n}}{(1)_{n}} \right] n z^n \sum_{n=0}^{\infty} (-1)^n \left[ \prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}}\right] (-z)^{n+x} \\ - \sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_i)_{n}}{(1)_{n}} \right] z^n \sum_{n=0}^{\infty} (-1)^n \left[ \prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}}\right] (n+x) (-z)^{n+x} \\ = p_1 x + p_2 x^2 + p_3 x^3 + p_4 x^4 + \mathcal{O}(x^5)\end{gathered}$$ if $z<0$ (alternating series), and an expansion of the form $$\begin{gathered} \label{q-positive} \sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_i)_{n}}{(1)_{n}} \right] n z^n \sum_{n=0}^{\infty} \left[ \prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}}\right] z^{n+x} \\ - \sum_{n=0}^{\infty} \left[\prod_{i=0}^{4} \frac{(s_i)_{n}}{(1)_{n}} \right] z^n \sum_{n=0}^{\infty} \left[ \prod_{i=0}^{4} \frac{(s_i)_{n+x}}{(1)_{n+x}}\right] (n+x) z^{n+x} \\ = p_1 x + p_2 x^2 + p_3 x^3 + p_4 x^4 + \mathcal{O}(x^5)\end{gathered}$$ if $z>0$ and $z \neq 1$ (series of positive terms). From the differential equations for Calabi-Yau threefolds [@YaZu; @Gui-matrix; @AlGu; @Zu4], we deduce that $$-\frac{1}{\pi} \frac{p_2}{p_1}=t, \quad q=-e^{-\pi t} \quad \text{or} \quad \, q=e^{-\pi t}$$ for (\[q-alternating\]) and (\[q-positive\]) respectively, where $z=z(q)$ is the mirror map, and $$k = 2 \left( \frac{1}{\pi^2} \frac{p_4}{p_2} - \frac53 - \cot^2 \pi s_1 - \cot^2 \pi s_2 \right).$$ The method used in [@AlGu] is good for convergent Ramanujan series, but not when the series is “too divergent". However the method based on bilateral series permits to calculate $q$ with many digits in all the cases, when we know the value of $z$. Taking $s_1=1/2$, $s_2=1/3$ and $z=27\phi^{-3}$ (formula (\[f9\]) in the Appendix), and using the formulas in [@AlGu], and $q=e^{-\pi t}$, we get $$t_1=t(27 \phi^{-3})=3.619403396730928522140860042453285904901.$$ However for $z=(-3 \phi)^3$ (formula (\[d5\]) in the Appendix), we need to use the method based on bilateral sums. Then, we get that the value of $t$ corresponding to (\[d5\]), is with $40$ correct digits (we can calculate many more) equal to $$t_2=t(-27\phi^3)=1.233412165189594043723756118628903242841.$$ We get the values of $\tau$ from (\[tau\]), and they are $\tau_1=4/3 \cdot \sqrt{10}$ and $\tau_2=1/9 \cdot \sqrt{10}$. We thought that the values of $t_1$ and $t_2$ would be easily related but we have not succeeded in finding any relation. For $s_1=s_2=1/2$, the series with $z$ and with $z^{-1}$ satisfy duality. For the formulas (\[f1\]) and (\[d2\]) in the Appendix, we have $$\begin{aligned} t_1=t(-2^{-10}) &=4.412809109031200738238212268698423552548, \\ t_2=t(-2^{10}) &=1.252302434231184754606061614044396018505.\end{aligned}$$ We get $\tau_1=\tau(-2^{-10})=\sqrt{41}$ and $\tau_2=1/16 \cdot \sqrt{41}$ from (\[tau\]), and we observe that $$t_1 t_2 = 2 \frac{\sqrt{41}+3}{\sqrt{41}-3}.$$ Hence the values of $t_1$ and $t_2$ are nicely related. Conclusion ========== We have conjectured that the terminating Fourier expansion of the function (\[rama-like-gui\]) corresponding to a Ramanujan-like series for $1/\pi^2$ has rational coefficients. A challenging problem is to prove it. That is, to obtain rigourosly the bilateral form corresponding to any Ramanujan-like series for $1/\pi^2$. This would suppose to give another step towards understanding the family of formulas for $1/\pi^2$. Below, we show how to solve the problem for the Ramanujan-type series for $1/\pi$. Consider, for example the Ramanujan series $$\label{rama-882} \sum_{n=0}^{\infty} (-1)^n \frac{\left(\frac12\right)_n\left(\frac14\right)_n \left(\frac34\right)_n}{(1)_n^3} \frac{21460n+1123}{882^{2n}} = \frac{3528}{\pi}.$$ To solve the problem of finding the bilateral form we need to determine the value of $u$ in $$\label{bilateral-rama-882} \sum_{n \in \mathbb{Z}} (-1)^n \frac{\left(\frac12\right)_{n+x} \left(\frac14\right)_{n+x} \left(\frac34\right)_{n+x}}{(1)_{n+x}^3} \frac{21360(n+x)+1123}{3528 \cdot 882^{2(n+x)}} = \frac{1-u+u\cos 2 \pi x}{\pi \cos \pi x (2\cos^2 \pi x-1)}.$$ Expanding the right side, comparing with [@Gui-matrix Expansion 1.1], and using [@Gui-matrix eq. 2.30-2.32], we rigorously obtain that $u=10$, which is a rational number. Appendix ======== In the following two lists $\phi$ means the fifth power of the golden ratio. That is $$\phi = \left (\frac{1+\sqrt 5}{2} \right)^5.$$ The notation ${\overset{?}=}$ means that the formula remains unproved, and the notation $``="$ means that we get the equality by analytic continuation. List of convergent formulas --------------------------- $$\label{f1} \sum_{n=0}^{\infty}\frac{\left(\frac12 \right)_n^5}{(1)_n^5}\frac{(-1)^n}{2^{10n}} (820n^2+180n+13)=\frac{128}{\pi^2},$$ $$\label{f2} \sum_{n=0}^{\infty} \frac{\left( \frac12 \right)_n^3 \left( \frac14 \right)_n \left( \frac34 \right)_n}{(1)_n^5} \frac{1}{2^{4n}} (120n^2+34n+3)=\frac{32}{\pi^2},$$ $$\label{f3} \sum_{n=0}^{\infty} \frac{\left( \frac12 \right)_n^5}{(1)_n^5} \frac{(-1)^n}{2^{2n}} (20n^2+8n+1)={8 \over \pi^2},$$ $$\label{f4} \sum_{n=0}^{\infty} \frac{\left( \frac12 \right)_n \left( \frac14 \right)_n \left( \frac34 \right)_n \left( \frac16 \right)_n \left( \frac56 \right)_n}{(1)_n^5} \frac{(-1)^n}{2^{10n}} (1640n^2+278n+15)\, {\overset{?}=} \, \frac{256 \sqrt 3}{\pi^2},$$ $$\label{f5} \sum_{n=0}^{\infty} \frac{\left( \frac12 \right)_n \left( \frac13 \right)_n \left( \frac23 \right)_n \left( \frac14 \right)_n \left( \frac34 \right)_n}{(1)_n^5} \frac{(-1)^n}{48^n} (252n^2+63n+5) \, {\overset{?} =} \, \frac{48}{\pi^2},$$ $$\label{f6} \sum_{n=0}^{\infty} \frac{\left( \frac12 \right)_n \left( \frac13 \right)_n \left( \frac23 \right)_n \left( \frac16 \right)_n \left(\frac56 \right)_n}{(1)_n^5} \frac{(-1)^n}{80^{3n}} (5418n^2+693n+29)\, {\overset{?} =} \, \frac{128 \sqrt 5}{\pi^2},$$ $$\label{f7} \sum_{n=0}^{\infty} \frac{\left(\frac12 \right)_n \left( \frac18 \right)_n \left( \frac38 \right)_n \left( \frac58 \right)_n \left( \frac78 \right)_n}{(1)_n^5} \frac{1}{7^{4n}} (1920n^2+304n+15) \, {\overset{?} =} \, \frac{56 \sqrt 7}{\pi^2},$$ $$\label{f8} \sum_{n=0}^{\infty} \frac{\left(\frac12 \right)_n^3 \left( \frac13 \right)_n \left( \frac23 \right)_n}{(1)_n^5} \left( \frac34 \right)^{3n} (74n^2+27n+3)\, = \, \frac{48}{\pi^2},$$ $$\label{f9} \sum_{n=0}^{\infty} \frac{\left(\frac12 \right)_n^3 \! \left( \frac13 \right)_n \! \left( \frac23 \right)_n}{(1)_n^5} \! \left( \frac{3}{\phi} \right)^{3n} \! \left[ (32 \! - \! \frac{216}{\phi}) n^2 \! + \! (18 \! - \! \frac{162}{\phi})n \! + \! (3 \! -\! \frac{30}{\phi}) \right] \, {\overset{?} =} \, \frac{3}{\pi^2},$$ $$\label{f10} \sum_{n=0}^{\infty} \frac{\left(\frac12 \right)_n \left( \frac13 \right)_n \left( \frac23 \right)_n \left( \frac16 \right)_n \left( \frac56 \right)_n}{(1)_n^5} (-1)^n \left( \frac{3}{4} \right)^{6n} (1930n^2+549n+45) \, {\overset{?}=} \, \frac{384}{\pi^2},$$ $$\label{f11} \sum_{n=0}^{\infty} \frac{\left(\frac12 \right)_n \left( \frac13 \right)_n \left( \frac23 \right)_n \left( \frac16 \right)_n \left( \frac56 \right)_n}{(1)_n^5} \left( \frac{3}{5} \right)^{6n} (532n^2+126n+9) \, {\overset{?}=} \, \frac{375}{4\pi^2}.$$ List of “divergent" formulas ---------------------------- $$\label{d1} \sum_{n=0}^{\infty}\frac{\left(\frac12 \right)_n^5}{(1)_n^5} (10n^2+6n+1) (-1)^n 4^n \, \, ``\!=\!" \, \, \frac{4}{\pi^2},$$ $$\label{d2} \sum_{n=0}^{\infty}\frac{\left(\frac12 \right)_n^5}{(1)_n^5} (205n^2+160n+32) (-1)^n 2^{10n} \, \, ``\!=\!" \, \, \frac{16}{\pi^2},$$ $$\label{d3} \sum_{n=0}^{\infty}\frac{\left(\frac12 \right)_n^3 \left( \frac13 \right)_n \left( \frac23 \right)_n}{(1)_n^5} (28n^2+18n+3) (-1)^n 3^{3n} \, `` \, {\overset{?}=} \, " \, \frac{6}{\pi^2},$$ $$\label{d4} \sum_{n=0}^{\infty}\frac{\left(\frac12 \right)_n \left( \frac13 \right)_n \left( \frac23 \right)_n\left( \frac14 \right)_n \left( \frac34 \right)_n}{(1)_n^5} (172n^2+75n+9) (-1)^n \left( \frac{27}{16} \right)^n \, \, ``=" \, \, \frac{48}{\pi^2},$$ $$\label{d5} \sum_{n=0}^{\infty} \frac{\left(\frac12 \right)_n^3 \! \left( \frac13 \right)_n \! \left( \frac23 \right)_n}{(1)_n^5} \! (-3 \phi)^{3n} \! \left[ (2408 \! + \! \frac{216}{\phi}) n^2 \! + \! (1800 \! + \! \frac{162}{\phi})n \! + \! (333 \! + \! \frac{30}{\phi}) \right] \, `` \, {\overset{?}=} \, " \, \frac{36}{\pi^2},$$ $$\label{d6} \sum_{n=0}^{\infty}\frac{\left(\frac12 \right)_n \left( \frac15 \right)_n \left( \frac25 \right)_n\left( \frac35 \right)_n \left( \frac45 \right)_n}{(1)_n^5} (483n^2+245n+30) (-1)^n \left( \frac{5^5}{2^8} \right)^n \, ``\, {\overset{?}=}\," \, \frac{80}{\pi^2}.$$ Formulas (\[f1\], \[f2\], \[f3\], \[f8\], \[d1\]), were proved by the author using the WZ-method. The upside-down of (\[d4\]) was proved in , and we can extract from it the WZ-pair that we need to prove (\[d4\]). All the other formulas are conjectured. The conjectured formula (\[f11\]) is joint with Gert Almkvist. In [@AlGu] we recovered all the known convergent series for $1/\pi^2$ and also two “divergent" ones: (\[d1\]) and (\[d4\]). However, the method used in [@AlGu] does not allow to discover “too divergent" series such as (\[d2\]) or (\[d5\]). We can check that the mosaic supercongruences pattern [@Gu6] hold for the formula (\[d5\]). Inspired by these congruences, by the conjectured formulas [@Sun Conj. 1.1–1.6] and by [@GuiRo], we observe that $$\begin{aligned} \nonumber \sum_{n=0}^{\infty} \frac{(1)_n^5}{\left(\frac12 \right)_n^3 \left( \frac13 \right)_n \left( \frac23 \right)_n} \left( \frac{-1}{3\phi} \right)^{3n} & \frac{(2408+216 \phi^{-1})n^2-(1800+162 \phi^{-1})n+(333+30\phi^{-1})}{n^5} \\ {\overset{?} =} & \frac{1125}{4}\sqrt{5}L_5(3)-448\zeta(3),\end{aligned}$$ seems true. As the convergence is fast, we can use it to get many digits of $L_5(3)$. We discovered the formula (\[d6\]) very recently. Related to it are the Zudilin-type supercongruences [@Zu0; @GuiZu]: $$\label{sc1} \sum_{n=0}^{p-1}\frac{\left(\frac12 \right)_n \left( \frac15 \right)_n \left( \frac25 \right)_n\left( \frac35 \right)_n \left( \frac45 \right)_n}{(1)_n^5} (483n^2+245n+30) (-1)^n \left( \frac{5^5}{2^8} \right)^n \, \, {\overset{?} \equiv} \, \, 30 p^3 \pmod{p^5},$$ for primes $p>2$, and the “upside-down" counterpart [@GuiRo], namely $$\label{ud1} \sum_{n=1}^{\infty}\frac{(1)_n^5}{\left(\frac12 \right)_n \left( \frac15 \right)_n \left( \frac25 \right)_n\left( \frac35 \right)_n \left( \frac45 \right)_n} \frac{-483n^2+245n-30}{n^5} (-1)^n \left( \frac{2^8}{5^5} \right)^n \, \, {\overset{?}=} \, \, 896 \, \zeta(3),$$ which is a convergent series for $\zeta(3)$. [99]{} <span style="font-variant:small-caps;">G. Almkvist and J. Guillera</span>, Ramanujan-like series for $1/\pi^2$ and String theory, *Exp. Math.*, **21**, (2012), 223-234. (eprint arXiv:1009.5202). <span style="font-variant:small-caps;">J. Guillera</span>, A matrix form of Ramanujan-type series for $1/\pi$; in Gems in Experimental Mathematics T. Amdeberhan, L.A. Medina, and V.H. Moll (eds.), *Contemp. Math.* **517** (2010), Amer. Math. Soc., 189–206; (eprint arXiv:0907.1547). <span style="font-variant:small-caps;">J. Guillera</span>, Mosaic supercongruences of Ramanujan-type, *Exp. Math.* **21**, (2012), 65-68. (e-print arXiv:1007.2290). <span style="font-variant:small-caps;">J. Guillera</span> and <span style="font-variant:small-caps;">W. Zudilin</span>, “Divergent” Ramanujan-type supercongruences. *Proc. Amer. Math. Soc.* **140** (2012), 765–777. (e-print arXiv:1004.4337). <span style="font-variant:small-caps;">J. Guillera</span> and <span style="font-variant:small-caps;">M. Rogers</span>, Ramanujan series upside-down. *Journal of the Australian Math. Soc.* **97**, 78–106; (e-print arXiv:1206.3981). <span style="font-variant:small-caps;">Zhi-Wei Sun</span>, List of conjectural formulas for powers of $\pi$ and other constants. (arXiv:1102.5649). <span style="font-variant:small-caps;">W. Chu and W. Zhang</span> Accelerating Dougall’s $_5F_4$-sum and infinite series involving $\pi$, *Journal: Math. Comp.* **83** (2014), 475–512. <span style="font-variant:small-caps;">Y. Yang and W. Zudilin</span>, On $\operatorname{Sp}_4$ modularity of Picard–Fuchs differential equations for Calabi–Yau threefolds, (with an appendix by V. Pasol); in Gems in Experimental Mathematics T. Amdeberhan, L.A. Medina, and V.H. Moll (eds.), *Contemp. Math.* **517** (2010), Amer. Math. Soc., 381–413; (e-print arXiv:0803.3322). <span style="font-variant:small-caps;">W. Zudilin</span>, Ramanujan-type supercongruences. *J. Number Theory* 129:8 (2009), 1848–1857; (e-print arXiv:0805.2788). <span style="font-variant:small-caps;">W. Zudilin</span>, Arithmetic hypergeometric series. *Russian Math. Surveys* 66:2 (2011), 369–420. *Russian version in Uspekhi Mat. Nauk* 66:2 (2011), 163–216; available at the author’s web site.

ArXiv