Datasets:

NousResearch

/

dolma-v1_7-arxiv

like 1

\section{Introduction} Deep networks, when trained on large-scale labeled datasets, can learn transferable representations which are generically useful across diverse tasks and application domains \cite{cite:ICML14DeCAF,cite:NIPS14CNN}. However, due to a phenomenon known as dataset bias or domain shift \cite{cite:CVPR11DB}, predictive models trained with these deep representations on one large dataset do not generalize well to novel datasets and tasks. The typical solution is to further fine-tune these networks on task-specific datasets, however, it is often prohibitively expensive to collect enough labeled data to properly fine-tune the high-capacity deep networks. Hence, there is strong motivation to establishing effective algorithms to reduce the labeling consumption by leveraging readily-available labeled data from a different but related source domain. This promising transfer learning paradigm, however, suffers from the shift in data distributions across different domains, which poses a major obstacle in adapting classification models to target tasks \cite{cite:TKDE10TLSurvey}. Existing transfer learning methods assume shared label space and different feature distributions across the source and target domains. These methods bridge different domains by learning domain-invariant feature representations without using target labels, and the classifier learned from source domain can be directly applied to target domain. Recent studies have revealed that deep neural networks can learn more transferable features for domain adaptation \cite{cite:ICML14DeCAF,cite:NIPS14CNN}, by disentangling explanatory factors of variations behind domains. The latest advances have been achieved by embedding domain adaptation modules in the pipeline of deep feature learning to extract domain-invariant representations \cite{cite:Arxiv14DDC,cite:ICML15DAN,cite:ICML15RevGrad,cite:ICCV15SDT,cite:NIPS16RTN,cite:NIPS16SDSN,cite:ICML17JAN}. \begin{figure*}[tbp] \centering \includegraphics[width=0.8\textwidth]{MADAprob.pdf} \caption{The difficulty of domain adaptation: discriminative structures may be mixed up or falsely aligned across domains. As an intuitive example, in this figure, the source class \textit{cat} is falsely aligned with target class \textit{dog}, making final classification wrong.} \label{fig:MADAproblem} \end{figure*} Recently, adversarial learning has been successfully embedded into deep networks to learn transferable features to reduce distribution discrepancy between the source and target domains. Domain adversarial adaptation methods \cite{cite:ICML15RevGrad,cite:ICCV15SDT} are among the top-performing deep architectures. These methods mainly align the whole source and target distributions, without considering the complex multimode structures underlying the data distributions. As a result, not only all data from the source and target domains will be confused, but also the discriminative structures could be mixed up, leading to false alignment of the corresponding discriminative structures of different distributions, with intuitive example shown in Figure~\ref{fig:MADAproblem}. Hence, matching the whole source and target domains as previous methods without exploiting the discriminative structures may not work well for diverse domain adaptation scenarios. There are two technical challenges to enabling domain adaptation: {(1)} enhancing \emph{positive} transfer by maximally matching the multimode structures underlying data distributions across domains, and {(2)} alleviating \emph{negative} transfer by preventing false alignment of modes in different distributions across domains. Motivated by these challenges, we present a multi-adversarial domain adaptation (MADA) approach, which captures multimode structures to enable fine-grained alignment of different data distributions based on multiple domain discriminators. A key improvement over previous methods is the capability to simultaneously promote positive transfer of relevant data and alleviate negative transfer of irrelevant data. The adaptation can be achieved by stochastic gradient descent with the gradients computed by back-propagation in linear-time. Empirical evidence demonstrates that the proposed MADA approach outperforms state of the art methods on standard domain adaptation benchmarks. \section{Related Work} Transfer learning \cite{cite:TKDE10TLSurvey} bridges different domains or tasks to mitigate the burden of manual labeling for machine learning~\cite{cite:TNN11TCA,cite:TPAMI12DTMKL,cite:ICML13TCS,cite:NIPS14FTL}, computer vision \cite{cite:ECCV10Office,cite:CVPR12GFK,cite:NIPS14LSDA} and natural language processing \cite{cite:JMLR11MTLNLP}. The main technical difficulty of transfer learning is to formally reduce the distribution discrepancy across different domains. Deep networks can learn abstract representations that disentangle different explanatory factors of variations behind data \cite{cite:TPAMI13DLSurvey} and manifest invariant factors underlying different populations that transfer well from original tasks to similar novel tasks \cite{cite:NIPS14CNN}. Thus deep networks have been explored for transfer learning \cite{cite:ICML11DADL,cite:CVPR13MidLevel,cite:NIPS14LSDA}, multimodal and multi-task learning \cite{cite:JMLR11MTLNLP,cite:ICML11MDL}, where significant performance gains have been witnessed against prior shallow transfer learning methods. However, recent advances show that deep networks can learn abstract feature representations that can only reduce, but not remove, the cross-domain discrepancy \cite{cite:ICML11DADL,cite:Arxiv14DDC}, resulting in unbounded risk for target tasks \cite{cite:COLT09DAT,cite:ML10DAT}. Some recent work bridges deep learning and domain adaptation \cite{cite:Arxiv14DDC,cite:ICML15DAN,cite:ICML15RevGrad,cite:ICCV15SDT,cite:NIPS16RTN,cite:NIPS16SDSN,cite:ICML17JAN}, which extends deep convolutional networks (CNNs) to domain adaptation by adding adaptation layers through which the mean embeddings of distributions are matched \cite{cite:Arxiv14DDC,cite:ICML15DAN,cite:NIPS16RTN}, or by adding a subnetwork as domain discriminator while the deep features are learned to confuse the discriminator in a domain-adversarial training paradigm \cite{cite:ICML15RevGrad,cite:ICCV15SDT}. While performance was significantly improved, these state of the art methods may be restricted by the fact that the discriminative structures as well as complex multimode structures are not exploited for fine-grained alignment of different distributions. Adversarial learning has been explored for generative modeling in Generative Adversarial Networks (GANs) \cite{cite:NIPS14AdversarialNet}. Recently, several difficulties of GANs have been addressed, e.g. ease training \cite{cite:Arxiv17WGAN,cite:ICLR17GAN}, avoid mode collapse \cite{cite:arXiv14CGAN,cite:ICLR17MGAN,cite:ICLR17UGAN}. In particular, Generative Multi-Adversarial Network (GMAN) \cite{cite:ICLR17GMAN} extends GANs to multiple discriminators including formidable adversary and forgiving teacher, which significantly eases model training and enhances distribution matching. \begin{figure*}[tbp] \centering \includegraphics[width=0.8\textwidth]{MADAfig.pdf} \caption{The architecture of the proposed Multi-Adversarial Domain Adaptation (MADA) approach, where $\mathbf{f}$ is the extracted deep features, ${\hat{\mathbf{y}}}$ is the predicted data label, and ${\hat{\mathbf{d}}}$ is the predicted domain label; $G_f$ is the feature extractor, $G_y$ and $L_y$ are the label predictor and its loss, $G_d^k$ and $L_d^k$ are the domain discriminator and its loss; GRL stands for Gradient Reversal Layer. The blue part shows the multiple adversarial networks (each for a class, $K$ in total) crafted in this paper. \emph{Best viewed in color.}} \label{fig:MADAfig} \end{figure*} \section{Multi-Adversarial Domain Adaptation} In unsupervised domain adaptation, we are given a source domain $\mathcal{D}_s = \{(\mathbf{x}_i^s,{\bf y}^s_i)\}_{i=1}^{n_s}$ of $n_s$ labeled examples and a target domain ${{\cal D}_t} = \{ {\bf{x}}_j^t\} _{j = 1}^{{n_t}}$ of $n_t$ unlabeled examples. The source domain and target domain are sampled from joint distributions $P(\mathbf{X}^s, \mathbf{Y}^s)$ and $Q(\mathbf{X}^t, \mathbf{Y}^t)$ respectively, and note that $P \ne Q$. The goal of this paper is to design a deep neural network that enables learning of transfer features $\mathbf{f} = G_f\left( {\bf{x}} \right)$ and adaptive classifier $y = G_y\left( {\bf{f}} \right)$ to reduce the shifts in the joint distributions across domains, such that the target risk ${\Pr _{\left( {{\mathbf{x}},y} \right) \sim q}}\left[ {G_y \left( G_f({\mathbf{x}}) \right) \ne {\bf y}} \right]$ minimized by jointly minimizing source risk and distribution discrepancy by multi-adversarial domain adaptation. There are two technical challenges to enabling domain adaptation: \textbf{(1)} enhancing \emph{positive} transfer by maximally matching the multimode structures underlying data distributions $P$ and $Q$ across domains, and \textbf{(2)} alleviating \emph{negative} transfer by preventing false alignment of different distribution modes across domains. These two challenges motivate the multi-adversarial domain adaptation approach. \subsection{Domain Adversarial Network} Domain adversarial networks have been successfully applied to transfer learning~\cite{cite:ICML15RevGrad,cite:ICCV15SDT} by extracting transferable features that can reduce the distribution shift between the source domain and the target domain. The adversarial learning procedure is a two-player game, where the first player is the domain discriminator $G_d$ trained to distinguish the source domain from the target domain, and the second player is the feature extractor $G_f$ fine-tuned simultaneously to confuse the domain discriminator. To extract domain-invariant features $\mathbf{f}$, the parameters $\theta_f$ of feature extractor $G_f$ are learned by maximizing the loss of domain discriminator $G_d$, while the parameters $\theta_d$ of domain discriminator $G_d$ are learned by minimizing the loss of the domain discriminator. In addition, the loss of label predictor $G_y$ is also minimized. The objective of domain adversarial network \cite{cite:ICML15RevGrad} is the functional: \begin{equation}\label{eqn:GRL} \begin{aligned} C_{0} \left( {{\theta _f},{\theta _y},{\theta _d}} \right) &= \frac{1}{{{n_s}}}\sum\limits_{{{\mathbf{x}}_i} \in {\mathcal{D}_s}} {{L_y}\left( {{G_y}\left( {{G_f}\left( {{{\mathbf{x}}_i}} \right)} \right),{y_i}} \right)} \\ &- \frac{\lambda }{n}\sum\limits_{{{\mathbf{x}}_i} \in \left( {{\mathcal{D}_s} \cup {\mathcal{D}_t}} \right)} {{L_d}\left( {{G_d}\left( {{G_f}\left( {{{\mathbf{x}}_i}} \right)} \right),{d_i}} \right)} , \end{aligned} \end{equation} where $n = n_s + n_t$ and $\lambda$ is a trade-off parameter between the two objectives that shape the features during learning. After training convergence, the parameters $\hat\theta_f$, $\hat\theta_y$, $\hat\theta_d$ will deliver a saddle point of the functional~\eqref{eqn:GRL}: \begin{equation}\label{eqn:param1} \begin{gathered} (\hat\theta_f, \hat\theta_y) = \arg \mathop {\min }\limits_{{\theta _f},{\theta _y}} C_0 \left( {{\theta _f},{\theta _y},{\theta _d}} \right), \\ (\hat\theta_d) = \arg \mathop {\max }\limits_{{\theta_d}} C_0 \left( {{\theta _f},{\theta _y},{\theta _d}} \right). \end{gathered} \end{equation} Domain adversarial networks \cite{cite:ICML15RevGrad,cite:ICCV15SDT} are the top-performing architectures for standard domain adaptation when the distributions of the source domain and target domain can be aligned successfully. \subsection{Multi-Adversarial Domain Adaptation} In practical domain adaptation problems, however, the data distributions of the source domain and target domain usually embody complex multimode structures, reflecting either the class boundaries in supervised learning or the cluster boundaries in unsupervised learning. Thus, previous domain adversarial adaptation methods that only match the data distributions without exploiting the multimode structures may be prone to either under transfer or negative transfer. Under transfer may happen when different modes of the distributions cannot be maximally matched. Negative transfer may happen when the corresponding modes of the distributions across domains are falsely aligned. To promote positive transfer and combat negative transfer, we should find a technology to reveal the multimode structures underlying distributions on which multi-adversarial domain adaptation can be performed. To match the source and target domains upon the multimode structures underlying data distributions, we notice that the source domain labeled information provides strong signals to reveal the multimode structures. Therefore, we split the domain discriminator $G_d$ in Equation~\eqref{eqn:GRL} into $K$ class-wise domain discriminators $G_d^k, {k=1, \ldots, K}$, each is responsible for matching the source and target domain data associated with class $k$, as shown in Figure~\ref{fig:MADAfig}. Since target domain data are fully unlabeled, it is not easy to decide which domain discriminator $G_d^k$ is responsible for each target data point. Fortunately, we observe that the output of the label predictor ${\hat{\bf{y}}}_i = G_y(\mathbf{x}_i)$ to each data point $\mathbf{x}_i$ is a probability distribution over the label space of $K$ classes. This distribution well characterizes the probability of assigning $\mathbf{x}_i$ to each of the $K$ classes. Thus, it is a natural idea to use ${\hat{\bf{y}}}_i$ as the probability to indicate how much each data point $\mathbf{x}_i$ should be attended to the $K$ domain discriminators $G_d^k, {k=1, \ldots, K}$. The attention of each point $\mathbf{x}_i$ to a domain discriminator $G_d^k$ can be modeled by weighting its features $G_f(\mathbf{x}_i)$ with probability $\hat y_i^k$. Applying this to all $K$ domain discriminators $G_d^k, {k=1, \ldots, K}$ yields \begin{equation}\label{eqn:Ld} {L_d} = \frac{1}{n}\sum\limits_{k = 1}^K {\sum\limits_{{{\mathbf{x}}_i} \in {\mathcal{D}_s} \cup {\mathcal{D}_t}} {L_d^k\left( {G_d^k\left( {\hat y_i^k{G_f}\left( {{{\mathbf{x}}_i}} \right)} \right),{d_i}} \right)} } , \end{equation} where $G_d^k$ is the $k$-th domain discriminator while $L_d^k$ is its cross-entropy loss, and $d_i$ is the domain label of point $\mathbf{x}_i$. We note that the above strategy shares similar ideas with the attention mechanism. Compared with the previous single-discriminator domain adversarial network in Equation~\eqref{eqn:GRL}, the proposed multi-adversarial domain adaptation network enables fine-grained adaptation where each data point $\mathbf{x}_i$ is matched only by those relevant domain discriminators according to its probability ${\hat{\bf{y}}}_i$. This fine-grained adaptation may introduce three benefits. \textbf{(1)} It avoids the hard assignment of each point to only one domain discriminator, which tends to be inaccurate for target domain data. \textbf{(2)} It circumvents negative transfer since each point is only aligned to the most relevant classes, while the irrelevant classes are filtered out by the probability and will not be included in the corresponding domain discriminators, hence avoiding false alignment of the discriminative structures in different distributions. \textbf{(3)} The multiple domain discriminators are trained with probability-weighted data points $\hat y_i^k G_f(\mathbf{x}_i)$, which naturally learn multiple domain discriminators with different parameters $\theta_d^k$; these domain discriminators with different parameters promote \emph{positive transfer} for each instance. Integrating all things together, the objective of the Multi-Adversarial Domain Adaptation (MADA) is \begin{equation}\label{eqn:MultiA} \begin{aligned} C\left( {{\theta _f},{\theta _y},\theta _d^k|_{k = 1}^{K}} \right) & = \frac{1}{{{n_s}}}\sum\limits_{{{\mathbf{x}}_i} \in {\mathcal{D}_s}} {{L_y}\left( {{G_y}\left( {{G_f}\left( {{{\mathbf{x}}_i}} \right)} \right),{y_i}} \right)} \\ & - \frac{\lambda }{n}\sum\limits_{k = 1}^K {\sum\limits_{{{\mathbf{x}}_i} \in {\mathcal{D}}} {L_d^k\left( {G_d^k\left( {\hat y_i^k{G_f}\left( {{{\mathbf{x}}_i}} \right)} \right),{d_i}} \right)} }, \end{aligned} \end{equation} where $n = n_s + n_t$, $\mathcal{D} = {\mathcal{D}_s} \cup {\mathcal{D}_t}$ and $\lambda$ is a hyper-parameter that trade-offs the two objectives in the unified optimization problem. The optimization problem is to find the parameters ${\hat\theta_f}$, ${\hat\theta_y}$ and ${\hat\theta_d^k}(k=1,2,...,K)$ that jointly satisfy \begin{equation}\label{eqn:parameter1} \begin{gathered} ({\hat\theta_f}, {\hat\theta_y}) = \arg \mathop {\min }\limits_{{\theta _f},{\theta _y}} C\left( {{\theta _f},{\theta _y},\theta _d^k|_{k = 1}^{K}} \right), \\ ({\hat\theta_d^1},...,{\hat\theta_d^{K}}) = \arg \mathop {\max }\limits_{{\theta_d^1},...,{\theta_d^K}} C\left( {{\theta _f},{\theta _y},\theta _d^k|_{k = 1}^{K}} \right). \end{gathered} \end{equation} The multi-adversarial domain adaptation (MADA) model simultaneously enhances \emph{positive} transfer by maximally matching the multimode structures underlying data distributions across domains, and circumvents \emph{negative} transfer by avoiding false alignment of the distribution modes across domains. \section{Experiments} We evaluate the proposed multi-adversarial domain adaptation (MADA) model with state of the art transfer learning and deep learning methods. The codes, datasets and configurations will be available online at \url{github.com/thuml}. \begin{table*}[!htbp] \centering \caption{Accuracy (\%) on \emph{Office-31} for unsupervised domain adaptation (AlexNet and ResNet)} \label{table:office31} \begin{small} \begin{tabular}{|c|c|c|c|c|c|c|c|} \hline Method & A $\rightarrow$ W & D $\rightarrow$ W & W $\rightarrow$ D & A $\rightarrow$ D & D $\rightarrow$ A & W $\rightarrow$ A & Avg \\ \hline \hline AlexNet \cite{cite:NIPS12CNN} & 60.6$\pm$0.4 & 95.4$\pm$0.2 & 99.0$\pm$0.1 & 64.2$\pm$0.3 & 45.5$\pm$0.5 & 48.3$\pm$0.5 & 68.8\\ TCA \cite{cite:TNN11TCA} & 59.0$\pm$0.0 & 90.2$\pm$0.0 & 88.2$\pm$0.0 & 57.8$\pm$0.0 & 51.6$\pm$0.0 & 47.9$\pm$0.0 & 65.8\\ GFK \cite{cite:CVPR12GFK} & 58.4$\pm$0.0 & 93.6$\pm$0.0 & 91.0$\pm$0.0 & 58.6$\pm$0.0 & 52.4$\pm$0.0 & 46.1$\pm$0.0 & 66.7\\ DDC \cite{cite:Arxiv14DDC} & 61.0$\pm$0.5 & 95.0$\pm$0.3 & 98.5$\pm$0.3 & 64.9$\pm$0.4 & 47.2$\pm$0.5 & 49.4$\pm$0.4 & 69.3\\ DAN \cite{cite:ICML15DAN} & 68.5$\pm$0.3 & 96.0$\pm$0.1 & 99.0$\pm$0.1 & 66.8$\pm$0.2 & 50.0$\pm$0.4 & 49.8$\pm$0.3 & 71.7\\ RTN \cite{cite:NIPS16RTN} & 73.3$\pm$0.2 & 96.8$\pm$0.2 & 99.6$\pm$0.1 & 71.0$\pm$0.2 & 50.5$\pm$0.3 & 51.0$\pm$0.1 & 73.7 \\ RevGrad \cite{cite:ICML15RevGrad} & 73.0$\pm$0.5 & 96.4$\pm$0.3 & 99.2$\pm$0.3 & 72.3$\pm$0.3 & 52.4$\pm$0.4 & 50.4$\pm$0.5 & 74.1 \\ \textbf{MADA} & \textbf{78.5}$\pm$0.2 & \textbf{99.8}$\pm$0.1 & \textbf{100.0}$\pm$.0 & \textbf{74.1}$\pm$0.1 & \textbf{56.0}$\pm$0.2 & \textbf{54.5}$\pm$0.3 & \textbf{77.1} \\ \hline \hline ResNet \cite{cite:CVPR16DRL} & 68.4$\pm$0.2 & 96.7$\pm$0.1 & 99.3$\pm$0.1 & 68.9$\pm$0.2 & 62.5$\pm$0.3 & 60.7$\pm$0.3 & 76.1 \\ TCA \cite{cite:TNN11TCA} & 74.7$\pm$0.0 & 96.7$\pm$0.0 & 99.6$\pm$0.0 & 76.1$\pm$0.0 & 63.7$\pm$0.0 & 62.9$\pm$0.0 & 79.3 \\ GFK \cite{cite:CVPR12GFK} & 74.8$\pm$0.0 & 95.0$\pm$0.0 & 98.2$\pm$0.0 & 76.5$\pm$0.0 & 65.4$\pm$0.0 & 63.0$\pm$0.0 & 78.8 \\ DDC \cite{cite:Arxiv14DDC} & 75.8$\pm$0.2 & 95.0$\pm$0.2 & 98.2$\pm$0.1 & 77.5$\pm$0.3 & 67.4$\pm$0.4 & 64.0$\pm$0.5 & 79.7 \\ DAN \cite{cite:ICML15DAN} & 83.8$\pm$0.4 & 96.8$\pm$0.2 & 99.5$\pm$0.1 & 78.4$\pm$0.2 & 66.7$\pm$0.3 & 62.7$\pm$0.2 & 81.3 \\ RTN \cite{cite:NIPS16RTN} & 84.5$\pm$0.2 & 96.8$\pm$0.1 & 99.4$\pm$0.1 & 77.5$\pm$0.3 & 66.2$\pm$0.2 & 64.8$\pm$0.3 & 81.6 \\ RevGrad \cite{cite:ICML15RevGrad} & 82.0$\pm$0.4 & 96.9$\pm$0.2 & 99.1$\pm$0.1 & 79.7$\pm$0.4 & 68.2$\pm$0.4 & \textbf{67.4}$\pm$0.5 & 82.2 \\ \textbf{MADA} & \textbf{90.0}$\pm$0.1 & \textbf{97.4}$\pm$0.1 & \textbf{99.6}$\pm$0.1 & \textbf{87.8}$\pm$0.2 & \textbf{70.3}$\pm$0.3 & {66.4}$\pm$0.3 & \textbf{85.2} \\ \hline \end{tabular} \end{small} \end{table*} \subsection{Setup} \textbf{Office-31} \cite{cite:ECCV10Office} is a standard benchmark for visual domain adaptation, comprising 4,652 images and 31 categories collected from three distinct domains: \textit{Amazon} (\textbf{A}), which contains images downloaded from \url{amazon.com}, \textit{Webcam} (\textbf{W}) and \textit{DSLR} (\textbf{D}), which contain images respectively taken by web camera and digital SLR camera with different environments. We evaluate all methods across three transfer tasks \textbf{A} $\rightarrow$ \textbf{W}, \textbf{D} $\rightarrow$ \textbf{W} and \textbf{W} $\rightarrow$ \textbf{D}, which are widely used by previous deep transfer learning methods \cite{cite:Arxiv14DDC,cite:ICML15RevGrad}, and another three transfer tasks \textbf{A} $\rightarrow$ \textbf{D}, \textbf{D} $\rightarrow$ \textbf{A} and \textbf{W} $\rightarrow$ \textbf{A} as used in \cite{cite:ICML15DAN,cite:ICCV15SDT,cite:NIPS16RTN}. \textbf{ImageCLEF-DA}\footnote{\url{http://imageclef.org/2014/adaptation}} is a benchmark dataset for ImageCLEF 2014 domain adaptation challenge, which is organized by selecting the 12 common categories shared by the following three public datasets, each is considered as a domain: \textit{Caltech-256} (\textbf{C}), \textit{ImageNet ILSVRC 2012} (\textbf{I}), and \textit{Pascal VOC 2012} (\textbf{P}). The 12 common categories are {aeroplane}, {bike}, {bird}, {boat}, {bottle}, {bus}, {car}, {dog}, {horse}, {monitor}, {motorbike}, and {people}. There are 50 images in each category and 600 images in each domain. We use all domain combinations and build 6 transfer tasks: \textbf{I} $\rightarrow$ \textbf{P}, \textbf{P} $\rightarrow$ \textbf{I}, \textbf{I} $\rightarrow$ \textbf{C}, \textbf{C} $\rightarrow$ \textbf{I}, \textbf{C} $\rightarrow$ \textbf{P}, and \textbf{P} $\rightarrow$ \textbf{C}. Different from the \emph{Office-31} dataset where different domains are of different sizes, the three domains in this dataset are of equal size, making it a good alternative dataset. We compare the proposed multi-adversarial domain adaptation (\textbf{MADA)} with both shallow and deep transfer learning methods: Transfer Component Analysis (\textbf{TCA}) \cite{cite:TNN11TCA}, Geodesic Flow Kernel (\textbf{GFK}) \cite{cite:CVPR12GFK}, Deep Domain Confusion (\textbf{DDC}) \cite{cite:Arxiv14DDC}, Deep Adaptation Network (\textbf{DAN}) \cite{cite:ICML15DAN}, Residual Transfer Network (\textbf{RTN}) \cite{cite:NIPS16RTN}, and Reverse Gradient (\textbf{RevGrad}) \cite{cite:ICML15RevGrad}. TCA learns a shared feature space by Kernel PCA with linear-MMD penalty. GFK interpolates across an infinite number of intermediate subspaces to bridge the source and target subspaces. For these shallow transfer methods, we adopt SVM as the base classifier. DDC maximizes domain confusion by adding to deep networks a single adaptation layer that is regularized by linear-kernel MMD. DAN learns transferable features by embedding deep features of multiple domain-specific layers to reproducing kernel Hilbert spaces (RKHSs) and matching different distributions optimally using multi-kernel MMD. RTN jointly learns transferable features and adapts different source and target classifiers via deep residual learning \cite{cite:CVPR16DRL}. RevGrad enables domain adversarial learning \cite{cite:NIPS14AdversarialNet} by adapting a single layer of deep networks, which matches the source and target domains by making them indistinguishable for a domain discriminator. We follow standard evaluation protocols for unsupervised domain adaptation \cite{cite:ICML15DAN,cite:ICML15RevGrad}. For both \emph{Office-31} and \emph{ImageCLEF-DA} datasets, we use all labeled source examples and all unlabeled target examples. We compare the average classification accuracy of each method on three random experiments, and report the standard error of the classification accuracies by different experiments of the same transfer task. For all baseline methods, we either follow their original model selection procedures, or conduct \emph{transfer cross-validation} \cite{cite:ECML10TCV} if their model selection strategies are not specified. We also adopt transfer cross-validation \cite{cite:ECML10TCV} to select parameter $\lambda$ for the MADA models. Fortunately, our models perform very stably under different parameter values, thus we fix $\lambda = 1$ throughout all experiments. For MMD-based methods (TCA, DDC, DAN, and RTN), we use Gaussian kernel with bandwidth set to the median pairwise squared distances on the training data, i.e. median trick \cite{cite:JMLR12MMD,cite:ICML15DAN}. We examine the influence of deep representations for domain adaptation by exploring \textbf{AlexNet} \cite{cite:NIPS12CNN} and \textbf{ResNet} \cite{cite:CVPR16DRL} as base architectures for learning deep representations. For shallow methods, we follow DeCAF \cite{cite:ICML14DeCAF} and use as deep representations the activations of the $fc7$ (AlexNet) and $pool5$ (ResNet) layers. We implement all deep methods based on the \textbf{Caffe} \cite{cite:MM14Caffe} framework, and fine-tune from AlexNet \cite{cite:NIPS12CNN} and ResNet \cite{cite:CVPR16DRL} models pre-trained on the ImageNet dataset \cite{cite:Arxiv14ImageNet}. We fine-tune all convolutional and pooling layers and train the classifier layer via back propagation. Since the classifier is trained from scratch, we set its learning rate to be 10 times that of the lower layers. We employ the mini-batch stochastic gradient descent (SGD) with momentum of 0.9 and the learning rate strategy implemented in RevGrad \cite{cite:ICML15RevGrad}: the learning rate is not selected by a grid search due to high computational cost---it is adjusted during SGD using these formulas: ${\eta _p} = \frac{{{\eta _0}}}{{{{\left( {1 + \alpha p} \right)}^\beta }}}$, where $p$ is the training progress linearly changing from $0$ to $1$, $\eta_0 = 0.01, \alpha=10$ and $\beta=0.75$, which is optimized to promote convergence and low error on source domain. To suppress noisy activations at the early stages of training, instead of fixing parameter $\lambda$, we gradually change it by multiplying $\frac{2}{{1 + \exp \left( { - \delta p} \right)}} - 1$, where $\delta = 10$ \cite{cite:ICML15RevGrad}. This progressive training strategy significantly stabilizes parameter sensitivity of the proposed approach. \begin{table*}[!htbp] \centering \caption{Accuracy (\%) on \emph{ImageCLEF-DA} for unsupervised domain adaptation (AlexNet and ResNet)} \label{table:imageclefda} \begin{small} \begin{tabular}{|c|c|c|c|c|c|c|c|} \hline Method & I $\rightarrow$ P & P $\rightarrow$ I & I $\rightarrow$ C & C $\rightarrow$ I & C $\rightarrow$ P & P $\rightarrow$ C & Avg \\ \hline \hline AlexNet \cite{cite:NIPS12CNN} & 66.2$\pm$0.2 & 70.0$\pm$0.2 & 84.3$\pm$0.2 & 71.3$\pm$0.4 & 59.3$\pm$0.5 & 84.5$\pm$0.3 & 73.9 \\ DAN \cite{cite:ICML15DAN} & 67.3$\pm$0.2 & 80.5$\pm$0.3 & 87.7$\pm$0.3 & 76.0$\pm$0.3 & 61.6$\pm$0.3 & 88.4$\pm$0.2 & 76.9 \\ RTN \cite{cite:NIPS16RTN} & 67.4$\pm$0.3 & 82.3$\pm$0.3 & 89.5$\pm$0.4 & 78.0$\pm$0.2 & 63.0$\pm$0.2 & 90.1$\pm$0.1 & 78.4 \\ RevGrad \cite{cite:ICML15RevGrad} & 66.5$\pm$0.5 & 81.8$\pm$0.4 & 89.0$\pm$0.5 & 79.8$\pm$0.5 & 63.5$\pm$0.4 & 88.7$\pm$0.4 & 78.2 \\ \textbf{MADA} & \textbf{68.3}$\pm$0.3 & \textbf{83.0}$\pm$0.1 & \textbf{91.0}$\pm$0.2 & \textbf{80.7}$\pm$0.2 & \textbf{63.8}$\pm$0.2 & \textbf{92.2}$\pm$0.3 & \textbf{79.8} \\ \hline \hline ResNet \cite{cite:CVPR16DRL} & 74.8$\pm$0.3 & 83.9$\pm$0.1 & 91.5$\pm$0.3 & 78.0$\pm$0.2 & 65.5$\pm$0.3 & 91.2$\pm$0.3 & 80.7 \\ DAN \cite{cite:ICML15DAN} & 75.0$\pm$0.4 & 86.2$\pm$0.2 & 93.3$\pm$0.2 & 84.1$\pm$0.4 & 69.8$\pm$0.4 & 91.3$\pm$0.4 & 83.3 \\ RTN \cite{cite:NIPS16RTN} & \textbf{75.6}$\pm$0.3 & 86.8$\pm$0.1 & 95.3$\pm$0.1 & 86.9$\pm$0.3 & 72.7$\pm$0.3 & 92.2$\pm$0.4 & 84.9 \\ RevGrad \cite{cite:ICML15RevGrad} & 75.0$\pm$0.6 & 86.0$\pm$0.3 & \textbf{96.2}$\pm$0.4 & 87.0$\pm$0.5 & 74.3$\pm$0.5 & 91.5$\pm$0.6 & 85.0 \\ \textbf{MADA} & {75.0}$\pm$0.3 & \textbf{87.9}$\pm$0.2 & 96.0$\pm$0.3 & \textbf{88.8}$\pm$0.3 & \textbf{75.2}$\pm$0.2 & \textbf{92.2}$\pm$0.3 & \textbf{85.8} \\ \hline \end{tabular} \end{small} \end{table*} \begin{table*}[!htbp] \centering \caption{Accuracy (\%) on \emph{Office-31} for domain adaptation from 31 classes to 25 classes (AlexNet)} \label{table:office25} \begin{small} \begin{tabular}{|c|c|c|c|c|c|c|c|} \hline Method & A $\rightarrow$ W & D $\rightarrow$ W & W $\rightarrow$ D & A $\rightarrow$ D & D $\rightarrow$ A & W $\rightarrow$ A & Avg \\ \hline \hline AlexNet \cite{cite:NIPS12CNN} & 58.2$\pm$0.4 & 95.9$\pm$0.2 & 99.0$\pm$0.1 & 60.4$\pm$0.3 & 49.8$\pm$0.5 & 47.3$\pm$0.5 & 68.4\\ RevGrad \cite{cite:ICML15RevGrad} & 65.1$\pm$0.5 & 91.7$\pm$0.3 & 97.1$\pm$0.3 & 60.6$\pm$0.3 & 42.1$\pm$0.4 & 42.9$\pm$0.5 & 66.6 \\ \textbf{MADA} & \textbf{70.8}$\pm$0.2 & \textbf{96.6}$\pm$0.1 & \textbf{99.5}$\pm$.0 & \textbf{69.6}$\pm$0.1 & \textbf{51.4}$\pm$0.2 & \textbf{54.2}$\pm$0.3 & \textbf{73.7} \\ \hline \end{tabular} \end{small} \end{table*} \subsection{Results} The classification accuracy results on the \textit{Office-31} dataset for unsupervised domain adaptation based on AlexNet and ResNet are shown in Table \ref{table:office31}. For fair comparison, the results of DAN \cite{cite:ICML15DAN}, RTN \cite{cite:NIPS16RTN}, and RevGrad \cite{cite:ICML15RevGrad} are directly reported from their original papers. MADA outperforms all comparison methods on most transfer tasks. It is noteworthy that MADA promotes the classification accuracies substantially on hard transfer tasks, e.g. \textbf{A $\rightarrow$ W}, \textbf{A $\rightarrow$ D}, \textbf{D $\rightarrow$ A}, and \textbf{W $\rightarrow$ A}, where the source and target domains are substantially different, and produce comparable classification accuracies on easy transfer tasks, \textbf{D $\rightarrow$ W} and \textbf{W $\rightarrow$ D}, where the source and target domains are similar \cite{cite:ECCV10Office}. The three domains in the \textit{ImageCLEF-DA} dataset are balanced in each category. As reported in Table \ref{table:imageclefda}, the MADA approach outperforms the comparison methods on most transfer tasks. The encouraging results highlight the importance of multi-adversarial domain adaptation in deep neural networks, and suggest that MADA is able to learn more transferable representations for effective domain adaptation. The experimental results reveal several insightful observations. \textbf{(1)} Standard deep learning methods (AlexNet and ResNet) either outperform or underperform traditional shallow transfer learning methods (TCA and GFK) using deep features as input. This confirms the current practice that deep networks, even the extremely deep ones (ResNet), can learn abstract feature representations that only reduce but not remove the cross-domain discrepancy \cite{cite:NIPS14CNN}. \textbf{(2)} Deep transfer learning methods substantially outperform both standard deep learning methods and traditional shallow transfer learning methods with deep features as input. This validates that explicitly reducing the cross-domain discrepancy by embedding domain-adaptation modules into deep networks (DDC, DAN, RTN, and RevGrad) can learn more transferable features. \textbf{(3)} MADA substantially outperforms previous methods based on either multilayer adaptation (DAN), semi-supervised adaptation (RTN), and domain adversarial training (RevGrad). Although both MADA and RevGrad \cite{cite:ICML15RevGrad} perform domain adversarial adaptation, the improvement from RevGrad to MADA is crucial for domain adaptation: RevGrad matches data distributions across domains without exploiting the complex multimode structures; MADA enables domain adaptation by making the source and target domains indistinguishable multiple domain discriminators, each responsible for matching the source and target data associated with the same class, which can essentially reduce the shift in the data distributions of complex multimode structures. \begin{figure*}[h] \centering \subfigure[RevGrad: source=\textbf{A}]{ \includegraphics[width=0.22\textwidth]{RevGrad1.pdf} \label{fig:RevGrad1} }\hfil \subfigure[RevGrad: target=\textbf{W}]{ \includegraphics[width=0.22\textwidth]{RevGrad2.pdf} \label{fig:RevGrad2} }\hfil \subfigure[MADA: source=\textbf{A}]{ \includegraphics[width=0.22\textwidth]{MADA1.pdf} \label{fig:MADA1} }\hfil \subfigure[MADA: target=\textbf{W}]{ \includegraphics[width=0.22\textwidth]{MADA2.pdf} \label{fig:MADA2} } \caption{The t-SNE visualization of deep features extracted by RevGrad (a)(b) and MADA (c)(d).} \end{figure*} \begin{figure*}[h] \centering \subfigure[Sharing Strategies]{ \includegraphics[width=0.3\textwidth]{sharing.pdf} \label{fig:sharing} }\hfil \subfigure[${\cal A}$-distance]{ \includegraphics[width=0.3\textwidth]{Adist.pdf} \label{fig:Adist} }\hfil \subfigure[Convergence]{ \includegraphics[width=0.3\textwidth]{error.pdf} \label{fig:error} }\hfil \caption{Empirical analysis: (a) Sharing strategies, (b) ${\cal A}$-distance, and (c) Convergence performance.} \end{figure*} Negative transfer is an important technical bottleneck for successful domain adaptation. Negative transfer is more likely to happen when the source domain is substantially larger than the target domain, in which there exist many source data points that are irrelevant to the target domain. To evaluate the robustness against negative transfer, we randomly remove 6 classes from all transfer tasks constructed from the \emph{Office-31} dataset. For example, we perform domain adaptation on transfer task \textbf{A 31 $\rightarrow$ W 25}, where the source domain \textbf{A} has 31 classes but the target domain \textbf{W} has only 25 classes. In this more general and challenging scenario, we observe from Table~\ref{table:office25} that the top-performing adversarial adaptation method, RevGrad, significantly underperforms standard AlexNet on most transfer tasks. This is an evidence of the negative transfer difficulty. The proposed MADA approach significantly exceeds the performance of both AlexNet and RevGrad, and successfully avoids the negative transfer trap. These positive results imply that the multi-adversarial adaptation can alleviate negative transfer. \subsection{Analysis} \textbf{Feature Visualization:} We go deeper into the feature transferability by visualizing in Figures~\ref{fig:RevGrad1}--\ref{fig:MADA2} the network activations of task \textbf{A} $\rightarrow$ \textbf{W} (10 classes) learned by RevGrad (the bottleneck layer $fcb$) and MADA (the bottleneck layer $fcb$) respectively using t-SNE embeddings \cite{cite:ICML14DeCAF}. The visualization results reveal several interesting observations. \textbf{(1)} Under RevGrad features, the source and target domains are made indistinguishable; however, different categories are not well discriminated clearly. The reason is that domain adversarial learning is performed only at the feature layer $fcb$, while the discriminative information is not taken into account by the domain adversary. \textbf{(2)} Under MADA features, not only the source and target domains are made more indistinguishable but also different categories are made more discriminated, which leads to the best adaptation accuracy. This superior results benefit from the integration of discriminative information into multiple domain discriminators, which enables matching of complex multimode structures of the source and target data distributions. \textbf{Sharing Strategies:} Besides the proposed multi-adversarial strategy, one may consider other sharing strategies for multiple domain discriminators. For example, one can consider sharing all network parameters in the multiple domain discriminators, which is similar to previous domain adversarial adaptation methods with single domain discriminator; or consider sharing only a fraction of the network parameters for more flexibility. To examine different sharing strategies, we compare different variants of MADA: \textbf{MADA-full}, which shares all parameters of the multiple domain discriminator networks; \textbf{MADA-partial}, which shares only the lowest layers of the multiple discriminator networks. The accuracy results of tasks \textbf{A $\rightarrow$ W} and \textbf{A $\rightarrow$ D} in Figure~\ref{fig:sharing} reveal that the transfer performance decreases when we share more parameters of multiple discriminators. This confirms our motivation that multiple domain discriminators are necessary to establish fine-grained distribution alignment. \textbf{Distribution Discrepancy:} The domain adaptation theory~\cite{cite:ML10DAT,cite:COLT09DAT} suggests $\mathcal{A}$-distance as a measure of cross-domain discrepancy, which, together with the source risk, will bound the target risk. The proxy ${\cal{A}}$-distance is defined as ${d_{\cal A}} = 2\left( {1 - 2\epsilon } \right)$, where $\epsilon$ is the generalization error of a classifier (e.g. kernel SVM) trained on the binary task of discriminating source and target. Figure~\ref{fig:Adist} shows ${d_{\cal A}}$ on tasks \textbf{A} $\rightarrow$ \textbf{W}, \textbf{W} $\rightarrow$ \textbf{D} with features of ResNet, RevGrad, and MADA. We observe that ${d_{\cal A}}$ using MADA features is much smaller than ${d_{\cal A}}$ using ResNet and RevGrad features, which suggests that MADA features can reduce the cross-domain gap more effectively. As domains \textbf{W} and \textbf{D} are similar, ${d_{\cal A}}$ of task \textbf{W} $\rightarrow$ \textbf{D} is smaller than that of \textbf{A} $\rightarrow$ \textbf{W}, which well explains better accuracy of \textbf{W} $\rightarrow$ \textbf{D}. \textbf{Convergence Performance:} Since MADA involves alternating optimization procedures, we testify the convergence performance with ResNet and RevGrad. Figure~\ref{fig:error} demonstrates the test errors of different methods on task \textbf{A $\rightarrow$ W}, which suggests that MADA has similarly stable convergence performance as RevGrad while significantly outperforming RevGrad in the whole process of convergence. Also, the computational complexity of MADA is similar to RevGrad since the multiple domain discriminators only occupy a small fraction of the overall computational complexity. \section{Conclusion} This paper presented a novel multi-adversarial domain adaptation approach to enable effective deep transfer learning. Unlike previous domain adversarial adaptation methods that only match the feature distributions across domains without exploiting the complex multimode structures, the proposed approach further exploits the discriminative structures to enable fine-grained distribution alignment in a multi-adversarial adaptation framework, which can simultaneously promote positive transfer and circumvent negative transfer. Experiments show state of the art results of the proposed approach. \section*{Acknowledgments} This work was supported by the National Key Research and Development Program of China (2016YFB1000701), National Natural Science Foundation of China (61772299, 61325008, 61502265, 61672313) and Tsinghua National Laboratory (TNList) Key Project. \begin{small}

83da63734139449c75d55942d6f42cc9173a3a16

\section{\label{sec:level1}Introduction} Experimentally measurable high energy scattering processes have played important roles in revealing the internal structures of hadrons and the quark-gluon dynamics inside them over several decades. Recently, the data taken at LHC provide us great opportunities to deepen our knowledge about the Standard Model, especially about the strong interaction. Focusing on QCD, cross sections of high energy processes can be factorized into the soft and hard parts, according to the factorization theorem. The hard part can be calculated in principle by using the perturbation technique in QCD. However, on the other hand, the soft part cannot be rigorously calculated because of its nonperturbative nature, and the only way to treat it is the phenomenological parameterization of the parton distribution functions (PDFs). PDFs are expressed with two kinematic variables, the Bjorken scaling variable $x$ and the energy scale $Q^2$. Since the recent improvements in experimental techniques enable us to access the higher energy regime, a solid understanding of the small Bjorken $x$ physics is certainly required. Due to this background, improving nonperturbative methods to describe QCD in such a kinematic region has been one of the most important subjects in hadron physics. The holographic QCD, which is constructed based on the AdS/CFT correspondence~\cite{Maldacena:1997re,Gubser:1998bc,Witten:1998qj}, is one of the effective approaches to QCD, and has gathered a lot of theoretical interests so far~\cite{Kruczenski:2003be,Son:2003et,Kruczenski:2003uq,Sakai:2004cn,Erlich:2005qh,Sakai:2005yt,DaRold:2005zs}. Since Polchinski and Strassler first applied the AdS/CFT correspondence to the analysis on high energy scattering~\cite{Polchinski:2001tt}, many related studies have been done. In particular, the deep inelastic scattering (DIS) has been intensively studied~\cite{Polchinski:2002jw,BoschiFilho:2005yh,Brower:2006ea,Hatta:2007he,Brower:2007qh,BallonBayona:2007rs,Brower:2007xg,Cornalba:2008sp,Pire:2008zf,Cornalba:2010vk,Brower:2010wf,Watanabe:2012uc,Watanabe:2013spa,Watanabe:2015mia}, and understanding its small $x$ dynamics is especially important as mentioned above. To describe the complicated partonic dynamics in the small $x$ region, assuming the Pomeron exchange, which can be interpreted as a multi-gluon exchange, is useful~\cite{Donnachie:1992ny,ForshawRoss,PomeronPhysicsandQCD}. Brower, Polchinski, Strassler, and Tan (BPST) performed the gauge/string duality based analysis in the higher dimensional AdS space, and proposed a kernel which gives a Pomeron exchange contribution to the cross sections in high energy scattering phenomena~\cite{Brower:2006ea}. The BPST kernel was further improved subsequently~\cite{Brower:2007qh,Brower:2007xg}, and applied to the analysis on structure functions in DIS at small $x$~\cite{Brower:2010wf,Watanabe:2012uc,Watanabe:2013spa,Watanabe:2015mia}. The most striking result may be that the nontrivial scale dependence of the Pomeron intercept, which were experimentally measured at HERA~\cite{Breitweg:1998dz}, was well reproduced with this kernel. In this work, we apply the BPST kernel to the analysis on the total hadronic cross sections of hadron-hadron scattering at high energies. Compared to DIS, hadron-hadron scattering seems to be a simpler process, but this is not necessarily correct in the theoretical study in particular. In DIS, correctly producing the $Q^2$ dependence, where $Q^2$ is the four-momentum squared of the probe photon, may be most important, but the description of the photon is established in both the QCD and string theory, which relies on the fact that the photon is an elementary particle. On the other hand, in hadron-hadron scattering, since the involved scales are the masses of the participants only, the descriptions of the involved hadrons, which are nonperturbative composites, become more important. Hence, describing the hadron-hadron scattering is also challenging. Recently, new data for the proton-proton total cross sections at the TeV scale have been provided by the TOTEM collaboration at LHC~\cite{Antchev:2017dia}. The energy 13~TeV, at which the latest data were taken, corresponds to $x \sim 10^{-8}$, considering the proton mass $m_p \sim 1$~GeV. This $x$ value is much smaller compared to the $x$ range which was investigated at HERA. Therefore, it is expected that testing a phenomenological model in this kinematic region may provide us a new insight to deepen our understanding of the Pomeron nature. In our model setup, we combine the BPST Pomeron exchange kernel and two density distributions of the involved hadrons in the five-dimensional AdS space, and calculate the total cross sections. Since the Pomeron exchange can be realized as the Reggeized graviton exchange in the AdS space, the density distributions are described by the gravitational form factors, which are obtained by using the bottom-up AdS/QCD models~\cite{Abidin:2008hn,Abidin:2008ku,Abidin:2009hr}. It is known that the hard-wall model, in which the AdS geometry is sharply cut off in the infrared (IR) region to introduce the QCD scale, can give a reasonable description unless we consider an excited state of the hadron, so we employ the hard-wall models to describe the hadrons in this study. It is shown that both the recent TOTEM data and other pp and $\bar{\rm{p}}$p data can be well reproduced within our model. In the framework, once the adjustable model parameters are fixed with the data, one can predict the hadron-hadron cross sections involving other hadrons without any additional parameter, because a hadron is a normalizable mode and its density distribution is normalized. As examples, besides the nucleon-nucleon case, we also present the pion-nucleon and pion-pion results, which shows the versatility of our model and may support further applications to other high energy scattering processes in which the Pomeron exchange gives a reasonable approximation. In the next section, we explain the model setup, and then present our numerical results in Sec.~\ref{sec:level3}. The summary and discussion are given in Sec.~\ref{sec:level4}. \section{\label{sec:level2}Holographic description of hadron-hadron total cross sections} Here we present our model setup to obtain the total cross sections of hadron-hadron scattering, employing the BPST Pomeron exchange kernel denoted by $\chi$ in the equations hereafter~\cite{Brower:2006ea}. Following the preceding study~\cite{Brower:2010wf}, the scattering amplitude of the two-body process, $1 + 2 \to 3 + 4$, in the five-dimensional AdS space can be expressed in the eikonal representation as \be {\cal A} (s, t) = 2 i s \int d^2 b e^{i \bm{k_\perp } \cdot \bm{b}} \int dzdz' P_{13}(z) P_{24}(z') \left[ 1-e^{i \chi (s, \bm{b}, z, z')} \right], \label{eq:amplitude} \ee where $s$ and $t$ are the Mandelstam variables, $\bm{b}$ is the two-dimensional impact parameter, and $z$ and $z'$ are fifth (or bulk) coordinates for the incident and target particles, respectively. $P_{13}(z)$ and $P_{24}(z')$ in Eq.~\eqref{eq:amplitude} represent the density distributions of the two hadrons in the AdS space, which are to be specified for the numerical evaluations, and satisfy the normalizable conditions, \be \int dz P_{13}(z) = \int dz' P_{24}(z') = 1, \label{eq:normalization_condition} \ee because the both particles are normalizable modes in this study. Utilizing the optical theorem and picking up the leading contribution to the eikonal approximation in Eq.~\eqref{eq:amplitude}, which means that we consider the single-Pomeron exchange, the total cross section is expressed as \be \sigma_{tot} (s) = 2 \int d^2b \int dzdz' P_{13} (z) P_{24} (z') \mbox{Im} \chi (s,\bm{b},z,z'). \label{eq:tcs_original} \ee In the conformal limit, the analytical expression of the imaginary part of $\chi$ can be obtained, and the impact parameter integration in Eq.~\eqref{eq:tcs_original} can be performed analytically. Hence, Eq.~\eqref{eq:tcs_original} is rewritten as \begin{align} &\sigma_{tot} (s) = \frac{g_0^2 \rho^{3/2}}{8 \sqrt{\pi}} \int dzdz' P_{13} (z) P_{24}(z') (zz') \mbox{Im} [\chi_{c}(s,z,z')], \label{eq:tcs_with_CK} \\ &\mbox{Im} [\chi_c(s,z,z') ] \equiv e^{(1-\rho)\tau} e^ {-[({\log ^2 z/z'})/{\rho \tau}]} / {\tau^{1/2}}, \label{eq:CK} \end{align} where \be \tau = \log (\rho z z' s / 2). \ee Here $g_0^2$ and $\rho$ are adjustable parameters which control the magnitude and the energy dependence of the cross sections, respectively. It is known that the inclusion of the confinement effect is necessary to reproduce the experimental data of the structure functions measured at HERA with the BPST Pomeron exchange kernel, unless we focus on the hard scattering, i.e., the high $Q^2$ region in the DIS case~\cite{Brower:2010wf,Watanabe:2012uc,Watanabe:2013spa}. Considering the nucleon-nucleon scattering, the scale of the process can be characterized by the nucleon mass, $m_N \sim 1$~GeV, which implies that the low energy dynamics with a strong coupling in QCD becomes dominant. Hence, instead of the conformal kernel, in this study we utilize the modified BPST kernel, in which the confinement effect is mimicked, given with the same functional form as the conformal kernel by \begin{align} &\mbox{Im} [\chi_{mod} (s, z, z')] \equiv \mbox{Im} [\chi_c (s, z, z') ] \nonumber \\ &\hspace{27mm} + \mathcal{F} (s, z, z') \mbox{Im} [\chi_c (s, z, z_0 z_0' / z') ],\label{eq:MK} \\ &\mathcal{F} (s, z, z') = 1 - 2 \sqrt{\rho \pi \tau} e^{\eta^2} \mbox{erfc}( \eta ), \\ &\eta = \left( -\log \frac{z z'}{z_0 z_0'} + \rho \tau \right) / {\sqrt{\rho \tau}}, \end{align} where $z_0$ and $z'_0$ are the cutoffs of the fifth coordinates which characterize the QCD scale. Note that these two parameters are uniquely fixed with hadron masses, so not adjustable. To perform the numerical evaluations of the total cross sections, one needs to specify the density distributions, $P_{13}(z)$ and $P_{24}(z')$, which characterize the involved hadrons. These distributions are expressed by the gravitational form factors which can be extracted from the hadron-Pomeron-hadron three point functions, using the bottom-up AdS/QCD models of hadrons. Since we consider the nucleon-nucleon, pion-nucleon, and pion-pion scattering in this study, we need to specify the density distributions of the nucleon and the pion. The gravitational form factors of the both hadrons were obtained by the authors of Refs.~\cite{Abidin:2008hn,Abidin:2009hr}, so we can utilize those results. In the model discussed in Refs.~\cite{Henningson:1998cd,Muck:1998iz,Contino:2004vy,Hong:2006ta}, the nucleon is described as a solution to the five-dimensional Dirac equation. Following the previous study~\cite{Watanabe:2012uc}, the density distribution of the nucleon is expressed with the left-handed and right-handed components of the Dirac field, $\psi_L$ and $\psi_R$, respectively in terms of the Bessel function as \begin{align} &P_N (z) = \frac{1}{2z^{3}} \left[ \psi_L^2 (z) + \psi_R^2 (z) \right], \\ &\psi_L (z) = \frac{\sqrt{2} z^2 J_2 (m_N z)}{z_0^N J_2 (m_N z_0^N)}, \ \psi_R (z) = \frac{\sqrt{2} z^2 J_1 (m_N z)}{z_0^N J_2 (m_N z_0^N)}, \end{align} where the cutoff parameter $z_0^N$ is fixed by the condition, $J_1 (m_N z_0^N) = 0$. The value used in this study is $z_0^N = 1 / (245$~MeV$)$ with $m_N = (m_p + m_n)/2$, where $m_p$ and $m_n$ are the proton and neutron physical masses, respectively. On the other hand, focusing on the chiral limit case, the pion wave function $\Psi$ can also be analytically obtained as a solution to the equation of motion derived from the bottom-up AdS/QCD model of mesons~\cite{Erlich:2005qh}. Following the procedure in Ref.~\cite{Watanabe:2012uc} again, the density distribution of the pion is given with the Bessel function by \begin{align} &P_\pi (z) = \frac{ \left[ \partial _{z} \Psi (z) \right] ^2 }{4 \pi^2 f_\pi ^2 z} + \frac{\sigma^2 z^6 \Psi (z)^2 }{ f_\pi ^2 z^3 }, \\ &\Psi \left( z \right) = z\Gamma \left[ {\frac{2}{3}} \right] \left( {\frac{\alpha }{2}} \right)^{1/3} \Biggl[ I_{ - 1/3} \left( {\alpha z^3 } \right) - I_{1/3} \left( {\alpha z^3 } \right)\frac{{I_{2/3} \left( {\alpha (z_0^\pi)^3 } \right)}}{{I_{ - 2/3} \left( {\alpha (z_0^\pi)^3 } \right)}} \Biggr] , \end{align} where $f_\pi$ is the pion decay constant, $\alpha = 2 \pi \sigma / 3$, $\sigma = (332$~MeV$)^3$, and the cutoff parameter $z_0^\pi$ is fixed with the $\rho$ meson mass $m_\rho$ by the condition, $J_0 (m_\rho z_0^\pi) = 0$. The actual value used in this work is $z_0^\pi = 1/(322$~MeV$)$. \section{\label{sec:level3}Numerical results} As seen in the previous section, since the expressions for the density distributions do not include any adjustable parameter, there are only two parameters in total, $g_0^2$ and $\rho$, in the BPST Pomeron exchange kernel, to be determined with experimental data. Focusing on the energy range, $10^2 < \sqrt{s} < 10^5$~GeV, and considering the recent proton-proton collision data measured by the TOTEM collaboration at LHC~\cite{Antchev:2017dia,Antchev:2013gaa,Antchev:2013iaa,Antchev:2013paa,Antchev:2015zza,Antchev:2016vpy,Nemes:2017gut} and other pp~\cite{Baltrusaitis:1984ka,Honda:1992kv,Collaboration:2012wt} and $\bar{\rm{p}}$p~\cite{Battiston:1982su,Hodges:1983oba,Bozzo:1984rk,Alner:1986iy,Amos:1991bp,Abe:1993xy,Augier:1994jn,Avila:2002bp} data which were summarized by the Particle Data Group in 2010~\cite{Nakamura:2010zzi}, we perform the numerical fitting. The best fit values we obtained are $g_0^2 = 6.27 \times 10^2$ and $\rho = 0.824$. It should be mentioned here that $g_0^2$ is just an overall factor and the slope of the total cross section is totally governed by the other parameter $\rho$. Also, the values we obtained in this work are different from those in Ref.~\cite{Watanabe:2012uc}, which is because the kinematics of hadron-hadron scattering is different from that of DIS. We display in Fig.~\ref{fig:TCS_NN} \begin{figure*}[tb!] \begin{center} \includegraphics[width=0.75\textwidth]{fig/TCS_NN.pdf} \caption{ The nucleon-nucleon total cross section as a function of $\sqrt{s}$. The solid and dashed curves represent our calculation and the empirical fit by COMPETE collaboration~\cite{Cudell:2002xe}, respectively. The experimental data measured by the TOTEM collaboration at LHC~\cite{Antchev:2017dia,Antchev:2013gaa,Antchev:2013iaa,Antchev:2013paa,Antchev:2015zza,Antchev:2016vpy,Nemes:2017gut} and other data for pp~\cite{Baltrusaitis:1984ka,Honda:1992kv,Collaboration:2012wt} and $\bar{\rm{p}}$p~\cite{Battiston:1982su,Hodges:1983oba,Bozzo:1984rk,Alner:1986iy,Amos:1991bp,Abe:1993xy,Augier:1994jn,Avila:2002bp} collisions are depicted with error bars. } \label{fig:TCS_NN} \end{center} \end{figure*} the resulting nucleon-nucleon total cross section, compared with the experimental data and the empirical fit by COMPETE collaboration~\cite{Cudell:2002xe}. One can see from the figure that our calculation is in agreement with the TOTEM data and other pp and $\bar{\rm{p}}$p data, although the data at $\sqrt{s} > 10$~TeV except for the TOTEM's have huge uncertainties because they were extracted from the cosmic-ray experiments. Also, it can be seen that our result is consistent with the empirical fit at $\sqrt{s} < 10$~TeV, but a substantial deviation is observed in the higher $s$ region. Next, we show our results for the total cross sections of the pion-nucleon and pion-pion scattering in Fig.~\ref{fig:TCS_summary}. \begin{figure}[tb!] \includegraphics[width=0.47\textwidth]{fig/TCS_summary.pdf} \caption{ The resulting total cross sections as a function of $\sqrt{s}$. The solid, dashed, and dotted curves represent our calculations for the nucleon-nucleon, pion-nucleon, and pion-pion scattering, respectively. } \label{fig:TCS_summary} \end{figure} The Pomeron exchange is a universal picture, and the adjustable parameters included in the BPST kernel do not depend on the involved hadron properties. Also, in our present model setup both the nucleon and the pion are normalizable modes and follow the normalization conditions, Eq.~\eqref{eq:normalization_condition}. Hence, once we fix the model parameters with the nucleon-nucleon data, we can predict the pion involved processes without any additional parameter. Since the scales appeared in the considered scattering processes are characterized by the involved hadron masses, which are constants, the $s$ dependence is common to all the three curves in Fig.~\ref{fig:TCS_summary}. Therefore, the total cross section ratios which can be obtained from the results are also constant. Our predictions are as follows: \be \frac{\sigma_{tot}^{\pi N}}{\sigma_{tot}^{N N}} = 0.63, \ \frac{\sigma_{tot}^{\pi \pi}}{\sigma_{tot}^{N N}} = 0.45. \ee Donnachie and Landshoff studied various hadron-hadron total cross sections, considering the so-called soft Pomeron with the intercept 1.0808~\cite{Donnachie:1992ny}, and the ratio extracted from their results is $\sigma_{tot}^{\pi N} / \sigma_{tot}^{N N} = 0.63$ which agrees with our prediction. \section{\label{sec:level4}Summary and discussion} In this work, we have investigated the total cross sections of hadron-hadron scattering at high energies in the framework of holographic QCD, assuming the Pomeron exchange to describe the involved strong interaction. The cross sections are expressed by combining the BPST Pomeron exchange kernel and the two density distributions of the involved hadrons. The density distributions are described by the gravitational form factors which can be obtained from the bottom-up AdS/QCD models. The comparison between our calculations and experimental data, including the recent ones measured by the TOTEM collaboration at LHC, has been explicitly demonstrated. The resulting nucleon-nucleon total cross section is in agreement with the experimental data in the considered kinematic range, $10^2 < \sqrt{s} < 10^5$~GeV, which shows that our model works well in the high energy hadron-hadron scattering in addition to the previously studied DIS processes. In DIS, the density distribution of the probe photon is concentrated around the ultraviolet (UV) boundary (small $z$), and the peak position of the target hadron density distribution is located in the IR region (large $z$). This UV-IR scattering realizes the larger Pomeron intercept depending on the probe photon virtuality $Q^2$, compared to the soft Pomeron with a smaller constant intercept. On the other hand, since the hadron-hadron scattering considered in this work is the IR-IR scattering, its kinematics is different from that of DIS. Therefore, the observation that the total cross section data can be well reproduced within the model implies its wide applicability. Our nucleon-nucleon result is also consistent with the empirical fit proposed by the COMPETE collaboration at $\sqrt{s} < 10$~TeV. However, a substantial deviation is observed in the higher $s$ regime. One possible reason for this may be a lack of the saturation effect in our model. Although the interaction in the hadron-hadron scattering is soft, the gluonic saturation may occur in the TeV scale. We have not taken into account this effect in the model so far, but the present observation implies the importance of considering this. Our calculation of the nucleon-nucleon scattering is larger than that the empirical fit shows, which seems reasonable, because the saturation effect may suppress the cross section. Besides the nucleon-nucleon scattering, we have studied the pion-nucleon and pion-pion scattering also. Since the energy dependence of the cross section in our model setup is totally governed by the BPST kernel, the observed $s$ dependencies are common to all the results. There is no data for such pion involved processes in the high $s$ region at this moment, but the predicted total cross section ratio $\sigma_{tot}^{\pi N} / \sigma_{tot}^{N N}$ agrees with the value extracted from the soft Pomeron based study done by Donnachie and Landshoff. To pin down this, experimental data are expected to be taken with a high intensity pion beam in the future. The results we obtained in this work suggest that the BPST Pomeron exchange kernel can be a useful and powerful theoretical tool in the analysis on various high energy scattering processes in which the nonperturbative gluonic dynamics can be approximated by the Pomeron exchange. One of the important extensions may be to consider differential cross sections within the present framework. The TOTEM collaboration has measured the proton-proton differential cross sections, and those data would be useful to further constrain the model parameters. Moreover, applications to other processes, such as the deeply virtual Compton scattering or the vector meson production, may also be interesting. Further studies are certainly needed. \section*{Acknowledgements} A.W. acknowledges Marco Ruggieri for fruitful discussions. This work was supported in part by the NSFC under Grant Nos. 11725523, 11735007, and 11261130311 (CRC 110 by DFG and NSFC). \bibliographystyle{elsarticle-num} \section*{References}

14592fe92d46878b7c9229974526adf4bcf225be

\section{Introduction}\label{sec:intro} Solving recursive domain equations $$X\cong TX$$ is an important method to define mathematical models of computation in which recursive equations have unique solutions \cite{smyth-plotkin}. The technique was introduced by Dana Scott \cite{scott:continuouslattices} in order to give a model of the untyped lambda calculus. Following a suggestion of Lawvere, Scott also emphasises that this solution arises as a limit of a chain constructed in a systematic fashion from a certain functor $T$ on a category of continuous lattices. In fact, Scott's solution is then the final $T$-coalgebra $$X\to TX$$ constructed in the by now standard way called the final coalgebra sequence. In domain theory \cite{abra-jung:dt}, this construction has been employed to give solutions to domain equations $X\cong TX$ for functors $T$ on sophisticated base categories such as metric spaces, measurable spaces and many more. Indeed, there is a considerable variety of interesting categories of domains, each supporting different properties and type constructors. It may not always be obvious which category of domains is the most appropriate for a given modelling task, and the research into finding new good classes of domains still continues. \medskip\noindent While an unusual property in most areas of mathematics, in domain theory final coalgebras often coincide with initial algebras. So final coalgebras as solutions of domain equations became more prominent only after Aczel in his work on non-well founded sets~\cite{aczel:book} showed that many important domain equations can be solved in the category of sets, a category in which initial algebras and final coalgebras do not coincide and in which the final coalgebras are the ones that provide the desired solutions to recursive equations. Subsequent work by Rutten \cite{rutten:uc} and many others showed that streams, automata, probabilistic and other systems can be successfully described by functors on ${\mathsf{Set}}$, the category of sets and functions. Moreover, working with coalgebras over ${\mathsf{Set}}$ instead of over more complicated domains, allowed many powerful results to emerge, in particular on modal logics for coalgebraic systems and on formats of systems of equations that have unique solutions in final coalgebras. \medskip\noindent On the other hand, studying these techniques for richer base categories than ${\mathsf{Set}}$ has also been an important topic. Coalgebras over categories of algebras often correspond to more powerful automata in which the states are equipped with extra memory. And coalgebras over partial orders and metric spaces allow us to capture simulation instead of bisimulation \cite{hugh-jaco:simulation} or metric bisimulations for quantitative systems \cite{rutten:cmcs98,worrell:cmcs00}. To summarise, while coalgebra successfully promoted simpler set-based models of computation, it also extended the class of interesting domains in which to solve domain equations. \medskip\noindent This increasing variety of categories of domains makes it necessary to find systematic ways to relate them. For example, in this paper, one of the questions we ask is the following. What does it mean to solve the same domain equation $$X\cong TX$$ in different categories? On the face of it, this question does not make sense as the definition of a functor $T:{\mathbb{X}}\to {\mathbb{X}}$ depends on the base category ${\mathbb{X}}$. So what does it mean to be the ``same'' functor on two different base categories ${\mathbb{X}}$ and ${\mathbb{Y}}$? We take our answer from Kelly's work on enriched category theory and consider two functors $T:{\mathbb{X}}\to{\mathbb{X}}$ and $\overline T:{\mathbb{Y}}\to{\mathbb{Y}}$ the same if there is a dense ``discrete'' functor $D:{\mathbb{X}}\to{\mathbb{Y}}$ and $\overline T$ is the left-Kan extension of $DT$ along $D$ \begin{equation*} \xymatrix{ {\mathbb{Y}} \ar[r]^{\overline T} & {\mathbb{Y}}\\ {\mathbb{X}} \ar[r]_T \ar[u]^D& {\mathbb{X}}\ar[u]_D } \end{equation*} For example, we know already from previous work \cite{lmcs:bkv} that sets are dense in posets and that the convex powerset functor $\overline{\mathcal P}$ on ${\mathsf{Pos}}$, the category of partial orders and monotone maps, is the extension (posetification) of the powerset functor ${\mathcal P}$ on ${\mathsf{Set}}$. \begin{equation*} \xymatrix{ {\mathsf{Pos}} \ar[r]^{\overline{\mathcal P}} & {\mathsf{Pos}}\\ {\mathsf{Set}} \ar[r]_{\mathcal P} \ar[u]^D& {\mathsf{Set}}\ar[u]_D } \end{equation*} As we will show in Section~\ref{sec:beh}, this implies that the final coalgebra solution of the domain equation $Y\cong \overline{\mathcal P} Y$ is the same, in a suitable sense, as the the one of $X\cong \mathcal PX$. \medskip\noindent Another question that we study is how to define type constructors on ``richer'' categories from type constructors on ``easier'' categories. For example, it is well-known that a power domain for simulation is given by the ``up-set functor'' $\mathcal U:{\mathsf{Pos}}\to{\mathsf{Pos}}$, which is the left Kan extension of $\mathcal P_u$ along $D$ \begin{equation*} \xymatrix{ {\mathsf{Pos}} \ar[r]^{{\mathcal U}} & {\mathsf{Pos}}\\ {\mathsf{Set}} \ar[ru]_{\mathcal P_u} \ar[u]^D& } \end{equation*} where $\mathcal P_u(X)$ is again the powerset of $X$, but now \emph{ordered by inclusion}. Section~\ref{sec:beh} shows that the solution of a ``richer'' equation such as $Y\cong \mathcal U Y$ classifies the same notion of bisimulation as the solution of the corresponding ``easier'' equation, here $X\cong \mathcal P X$, but also carries an ordered structure that classifies simulation. \medskip\noindent The purpose then of this paper is to extend the observations above from posets to metric spaces and, more generally, to the category $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ of categories enriched over a commutative quantale $\kat{V}$. The essence is to show that the discrete functor $$D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$$ is dense. Importantly, the proof will exhibit a formula that allows us to compute the left Kan extensions along $D$ in many concrete examples. \medskip\noindent What is modelled by $\kat{V}$ and what role does $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ play? \medskip\noindent To say that $\kat{V}$ is a commutative quantale is to say that $\kat{V}$ is a lattice of ``truth values'' or ``distances''. The leading example is the lattice of real numbers $[0,\infty]$ which is promoted from a lattice to a quantale when equipped with the extra structure of addition of distances, important to capture the triangle inequality. \medskip\noindent Objects of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ are then ``categories enriched over $\kat{V}$'', that is, sets $\kat{X}$ equipped with a distance $\kat{X}(x,y)\in\kat{V}$. To say that $\kat{X}$ is a ``category enriched over $[0,\infty]$ with $+$'' is to say that $\kat{X}$ satisfies the triangle inequality $$ \kat{X}(x,y)+\kat{X}(y,z) \ge_\mathbb R \kat{X}(x,z).$$ Thus, one reason to work with enriched categories is that many structures of interest such as posets and metric spaces appear themselves as categories. Another one, emphasised in \cite{smyth-plotkin}, is the importance of homsets carrying ordered/metric structure. A third reason, important for this paper, will be discussed now in more detail, namely that enriched categories make available a richer variety of colimits. \medskip\noindent That $D:{\mathsf{Set}}\to{\mathsf{Pos}}$ is dense implies that every poset is a colimit of sets. But how can we construct ordered sets by taking quotients of discrete sets? The crucial point must be to work with a more general notion of quotient, namely with quotients that do not merely add equations between elements (coequalisers), but with quotients that add inequations. \medskip\noindent Such a richer notion of quotient is automatically provided to us in ${\mathsf{Pos}}$ or $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ as a special case of a so-called weighted (or indexed) colimit, a notion native to the theory of enriched categories. In this paper, we only need some special cases of weighted colimits that are reviewed in detail in Section~\ref{sec:preliminaries}. \medskip\noindent In the case of posets, the basic idea is easily explained. Every poset is a weighted colimit of discrete posets. How? Let $X$ be a poset and write $X_0$ for its set of elements and $X_1\subseteq X_0\times X_0$ for its order. Then $X$ is the ``ordered coequalizer'', or rather ``coinserter'', as it is known in the literature \begin{equation} \label{eq:poset_coins1} \xymatrix@C=35pt{ X_1 \ar@<0.5ex>[r]^{} \ar@<-0.5ex>[r]_{} & X_0 \ar[r]^-{} & X}. \end{equation} It is important here that the coinserter turns the discrete posets $X_0$, $X_1$ into a genuine poset $X$. This also makes it clear how to compute the left Kan extension $\overline T$ of $T$ by applying $T$ to the parallel pair of \eqref{eq:poset_coins1} and then taking the coinserter in ${\mathsf{Pos}}$ \begin{equation} \label{eq:poset_coins2} \xymatrix@C=35pt{ T X_1 \ar@<0.5ex>[r]^{} \ar@<-0.5ex>[r]_{} & T X_0 \ar[r]^-{} &\overline T X} \end{equation} This formula, describing the left-Kan extension as a certain weighted colimit of discrete $\kat{V}$-categories, allows us to compute such Kan extensions in many concrete situations. For example, in \cite{lmcs:bkv} we showed that it follows easily from \eqref{eq:poset_coins2} that the extension of $\mathcal P$ from ${\mathsf{Set}}$ to ${\mathsf{Pos}}$ is the convex powerset functor. \medskip\noindent A conceptual way of summarising this construction is to say that the coinserters of \eqref{eq:poset_coins1} form a ``density presentation'' of $D:{\mathsf{Set}}\to{\mathsf{Pos}}$. But density presentations work best when $D$ is fully faithful, which coincides with the quantal being integral, whereas the generalisation of the formula \eqref{eq:poset_coins2} also applies in case $D$ is not fully faithful. \paragraph{\bf Summary of main results.} \begin{enumerate} \item We prove in Theorem~\ref{thm:Lan} that \eqref{eq:poset_coins2} can be generalised to compute left Kan extensions along $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ for any commutative quantale $\kat{V}$. As a corollary, we obtain that $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is dense. If $D$ is fully faithful, we obtain a density presentation in Theorem~\ref{thm:dense-pres}. Those functors that arise as left Kan extensions along $D$ are characterised by preserving certain colimits in Theorems \ref{thm:char-discrete-arities} and \ref{thm:char-Vcatification}. \item It is well known that the fully-faithfulness of $D$ is equivalent to the unit $H\to (\Lan{D}H)D$ of the left Kan extension of $H$ along $D$ being an isomorphism for {\em all} $H$. We show in Proposition~\ref{prop:D-ff} that $D$ is fully faithful if and only if the quantale $\kat{V}$ is integral, which is the case in all the examples we pursue. We also characterise, in the general case, those functors $H$ for which the unit is an isomorphism in Theorem~\ref{thm:unitiso}. \item We use that $\kat{V}$-categories are relational presheaves to show in Theorem~\ref{thm:wpbcontrelpresh} how the extension of weak pullback preserving ${\mathsf{Set}}$-functors can be computed from their relation lifting. This is often the easiest route to concrete computations, the Pompeiu-Hausdorff metric being Example~\ref{ex:Vcatif-powerset}. \item Theorems \ref{thm:beh-vcatification} and \ref{thm:beh-discrete-arities} formalise the intuition that the final coalgebra over $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ of a ``functor $T:{\mathsf{Set}}\to{\mathsf{Set}}$ equipped with a $\kat{V}$-metric'' is the final $T$-coalgebra over ${\mathsf{Set}}$ equipped with an appropriate metric. In particular, both final coalgebras determine the same equivalence relation (bisimulation), but the final coalgebra over $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ gives refined metric information about states that are not bisimilar in the sense of measuring how far such states are from being equal. \end{enumerate} \noindent The formulas for computing left Kan extensions are \eqref{eq:shortest-path} in the general case and \eqref{eq:wpb} for weak pullback preserving functors. Examples of left Kan extensions are in Sections~\ref{sec:examples}--\ref{sec:Vmetric-functors}. Closure properties of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications are studied in Section~\ref{sec:closure} and 2-categorical properties of the category of coalgebras over $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ in Section~\ref{sec:2-cat}. \medskip\noindent The present paper is a revised and extended version of \cite{bkv:V-cat}. \section{Preliminaries} \label{sec:preliminaries} In this section we gather all the necessary technicalities and notation from category theory enriched in a complete and cocomplete symmetric monoidal closed category that we will use later. For the standard notions of enriched categories, enriched functors and enriched natural transformations we refer to Kelly's book~\cite{kelly:book} and for their importance to logical methods in computer science see for example~\cite{lawvere:gen_metric_spaces,turi-rutten,wagner,worrell:cmcs00,worrellPhD}. Readers familiar with enriched category theory are invited to skip these preliminaries and pass directly to Section~\ref{sec:extension} after taking note of Section~\ref{sec:discrete functor} on the fully-faithfulness of the discrete functor $D:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \medskip\noindent We will mainly use two prominent enrichments: that in a commutative quantale $\kat{V}$ and that in the category $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ of {\em small\/} $\kat{V}$-categories and $\kat{V}$-functors for a commutative quantale $\kat{V}$. We spell out in more detail how the relevant notions look like and carefully write all the enrichment-prefixes. In particular, the underlying category of an enriched category will be denoted by the same symbol, followed by the subscript ``$o$'' as usual. \subsection{$\kat{V}$-categories as generalised metric spaces} \ \bigskip\noindent Suppose $\kat{V}=(\kat{V}_o,\otimes,e,[{-},{-}])$ is a {\em commutative quantale\/}. More in detail: $\kat{V}_o$ is a complete lattice, equipped with a commutative and associative monotone binary operation $\otimes$, called the {\em tensor\/}. We require the element $e$ to be a {\em unit\/} of tensor. Furthermore, we require every monotone map ${-}\otimes r:\kat{V}_o\to\kat{V}_o$ to have a right adjoint $[r,{-}]:\kat{V}_o\to\kat{V}_o$. We call $[{-},{-}]$ the {\em internal hom\/} of $\kat{V}_o$. We will suppose that $\kat{V}_o$ is non-trivial, that is, $\kat{V}_o$ is not a singleton (or equivalently, $e \neq \bot$). Recall that the quantale is said to be {\em integral} if $e = \top$.\footnote{% A commutative integral quantale is sometimes called a {\em complete residuated lattice.} } \medskip\noindent Commutative quantales are complete and cocomplete symmetric monoidal closed categories. Therefore, one can define $\kat{V}$-categories, $\kat{V}$-functors, and $\kat{V}$-natural transformations. Before we say what these are, let us mention several examples of commutative quantales. \begin{exas} \label{ex:quantale} \hfill \begin{enumerate} \item \label{ex:quantale_two} The two-element chain ${\mathbb{2}}=\{0,1\}$ with the usual order and tensor $r\otimes s=r\wedge s$. The internal hom is implication. \item \label{ex:quantale_infinity} The real half line $([0,\infty], \geq_\mathbb R)$, with (extended) addition as tensor product. The internal hom $[r,s]$ is truncated minus, that is, $[r,s]=\textsf{ if } r\ge_\mathbb R s\textsf{ then } 0 \textsf{ else } s-r$. \item \label{ex:quantale_unit} The unit interval $([0,1], \geq_\mathbb R)$ with tensor product $r\otimes s=\max(r,s)$. The internal hom is given by $[r,s]=\textsf{ if }r\ge_\mathbb R s\textsf{ then } 0 \textsf{ else } s$. \item \label{ex:quantale_prob} The poset of all monotone functions $f:[0,\infty]\to [0,1]$ such that the equality $f(r)=\bigvee_{s < r} f(s)$ holds, with the pointwise order. It becomes a commutative quantale with the tensor product $$ (f \otimes g) (r) = \bigvee_{s + s' \leq r} f(s)\cdot g(s') $$ having as unit the function mapping all nonzero elements to $1$, and $0$ to itself~\cite{hofmann+reis}. \item \label{ex:quantale_three} Let $\underline n=\{0,\ldots n-1\}$ be an ordinal equipped with a monotone, commutative, idempotent operation satisfying $0\otimes i=0$ for all $i\in\underline n$. Then there are unique $e$ and $[-,-]$ such that $(\underline n, \otimes, e, [-,-])$ is a quantale \cite[Prop 2.5]{Casley-phd}. In the case $n=2$ we just obtain Item~\eqref{ex:quantale_two} above. In case $n=3$ there are exactly two ways of turning $\underline 3$ into a commutative quantale with an idempotent tensor, determined by choosing either $1\otimes 2=2$ or $1\otimes 2=1$~\cite[Cor 2.6]{Casley-phd}. % We briefly explain the associated $\kat{V}$-categories in Example~\ref{exle:casley}. The three-element quantale $\underline 3$ which is determined by $e=1$ is the smallest non-integral quantale, in the sense that it embeds into any other non-integral $\kat{V}$.% \item \label{ex:free-quant} Let $(\mathsf M, \cdot, e)$ be a monoid. Then $\mathcal P(\mathsf M)$, the powerset of $\mathsf M$, becomes a quantale ordered by inclusion with tensor $S \otimes S' = \{x \cdot x'\mid x\in S, x'\in S'\}$ and unit $\{e\}$. In fact, $\mathcal P(\mathsf M)$ is the free quantale over the monoid $\mathsf M$~\cite{rosenthal} and $\mathcal P(\mathsf M)$ is commutative if $\mathsf M$ is commutative. If $\mathsf M$ is the free monoid over an alphabet of size at least two, then $\mathcal P(\mathsf M)$ is the non-commutative non-integral quantale of languages over the alphabet. \end{enumerate} \end{exas} \begin{rem} If we think of the elements of the quantale as truth values, then $\le$ is external implication, $e$ is true, $\bot$ is false, $\otimes$ is a conjunction and $[-,-]$ is internal implication. Some standard logical laws such as $[e,r]=r$ and $r \otimes [r,s]\le s$ and $r\le[s,s\otimes r]$ and $s\le r \ \Leftrightarrow \ e\le [s,r]$ hold in all quantales. If we consider the elements of the quantale as representing distances, the order of the quantale is opposite to the natural order of distances, with $0$ as top and $\infty$ as bottom. To reconcile the two points of view, we can think of distance as ``distance from truth'', so that a distance of 0 corresponds to true and a distance of $\infty$ to false. \qed \end{rem} \medskip\noindent A ({\em small\/}) {\em $\kat{V}$-category\/} $\kat{X}$ consists of a (small) set of objects, together with an object $\kat{X}(x',x)$ in $\kat{V}_o$ for each pair $x'$, $x$ of objects, subject to the following axioms $$ e\leq\kat{X}(x,x), \quad \kat{X}(x',x)\otimes\kat{X}(x'',x')\leq\kat{X}(x'',x) $$ for all objects $x''$, $x'$ and $x$ in $\kat{X}$. \noindent A $\kat{V}$-category $\kat{X}$ is called \emph{discrete} if \begin{equation}\label{eq:discrete V-category} \kat{X}(x',x) = \begin{cases} \, e \ \ , \ \ x'=x \\ \bot \ \ , \ \ \mbox{otherwise} \end{cases} \end{equation} \medskip\noindent A {\em $\kat{V}$-functor\/} $f:\kat{X}\to\kat{Y}$ is given by the object-assignment $x\mapsto fx$, such that $$ \kat{X}(x',x)\leq\kat{Y}(fx',fx) $$ holds for all $x'$, $x$. \medskip\noindent A {\em $\kat{V}$-natural transformation\/} $f\to g$ is given whenever $$ e\leq\kat{Y}(fx,gx) $$ holds for all $x$. Thus, there is at most one $\kat{V}$-natural transformation between $f$ and $g$. \medskip\noindent $\kat{V}$-categories and $\kat{V}$-functors form a category which we denote by $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o$ (actually, a 2-category having $\kat{V}$-natural transformations as 2-cells). \begin{exa}\label{exle:V-categories} \hfill \begin{enumerate} \item \label{exle:Twocat} The two-element chain ${\mathbb{2}}$ is a commutative quantale. A small ${\mathbb{2}}$-category\footnote{ Not to be confounded with the notion of a 2-category, that is, a $\mathsf{Cat}$-enriched category.} $\kat{X}$ is precisely a {\em preorder\/}, where $x'\leq x$ iff $\kat{X}(x',x)=1$, while a ${\mathbb{2}}$-functor $f:\kat{X} \to \kat{Y}$ is a monotone map. A ${\mathbb{2}}$-natural transformation $f\rightarrow g$ expresses that $fx\leq gx$ holds for every $x$. Thus ${\mathbb{2}}\mbox{-}\mathsf{cat}$ is the category ${\mathsf{Preord}}$ of preorders and monotone maps. It plays an important role not only because we study $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ by generalising from ${\mathbb{2}}$-cat, but also because the definition \begin{equation} \label{eq:V-metric=>order} x\le_\kat{X} y \ \Longleftrightarrow \ e\le \kat{X}(x,y). \end{equation} equips every $\kat{V}$-category $\kat{X}$ with the structure of a preorder $\le_\kat{X}$. In case $\kat{X}$ is $\kat{V}$, we have ${\le_\kat{X}}={\le}$. In the following, we will drop the subscript in $\le_\kat{X}$. \medskip \item \label{ex:V as V-category itself} $\kat{V}$ is itself a $\kat{V}$-category with $\kat{V}(r,s)=[r,s]$. This is of fundamental importance, at least because the internal hom $\kat{V}(r,s)$ usually has a richer structure than the external hom $\kat{V}_o(r,s)$. For example in case $\kat{V}=[0,\infty]$, the external hom is only two-valued whereas the internal hom is $[0,\infty]$-valued. \end{enumerate} \end{exa} \noindent A good intuition is that $\kat{V}$-categories are possibly non-symmetric metric spaces and $\kat{V}$-functors are non-expanding maps. This intuition goes back to Lawvere~\cite{lawvere:gen_metric_spaces}. We show next some examples that explain this intuition. A good resource is also~\cite{rutten:ultrametric_spaces}. \begin{exas}\label{exs:Vcats=metric_spaces} \begin{enumerate} \item \label{ex:V=[0,infty]} Let $\kat{V}$ be the real half line $([0,\infty],\geq_\mathbb R, +, 0)$ as in Example~\ref{ex:quantale}\eqref{ex:quantale_infinity}. A small $\kat{V}$-category can be identified with a set $X$ equipped with a mapping $d_X:X\times X\to [0,\infty]$ such that $(X,d_X)$ is a {\em generalised metric space\/}. The generalisation of the usual notion is three-fold. First, $d_X$ is a pseudo-metric in the sense that two distinct points may have distance $0$. Second, $d_X$ is a quasi-metric in the sense that distance is not necessarily symmetric. Third, distances are allowed to be infinite, which has the important consequence that the category of generalised metric spaces has colimits (whereas metric spaces do not even have coproducts). \noindent A $\kat{V}$-functor $f:(X,d_X)\to (Y,d_Y)$ is then a exactly a {\em non-expanding mapping\/}, that is, one satisfying the inequality $d_Y(fx',fx)\leq_\mathbb R d_X(x',x)$ for every $x,x'\in X$. \noindent The existence of a $\kat{V}$-natural transformation $f\to g$ means that $\bigvee_x d_Y(fx,gx)=0$, that is, the distance $d_Y(fx,gx)$ is $0$, for every $x\in X$, or also that $fx\le gx$ with respect to the order discussed in Example~\ref{exle:V-categories}\eqref{exle:Twocat}. \item For the unit interval $\kat{V}=([0,1], \geq_\mathbb R, \max, 0)$ from Example~\ref{ex:quantale}\eqref{ex:quantale_unit}, a $\kat{V}$-category is a {\em generalised ultrametric space\/} $( X,d_X:X \times X \to [0,1])$~\cite{rutten:ultrametric_spaces,worrell:cmcs00}. Again, the slight generalisation of the usual notion lies in the fact that the distance function $d_X$ is not necessarily symmetric and $d_X(x',x)=0$ does not necessarily entail $x=x'$. Similarly, $\kat{V}$-functors are precisely the non-expanding maps, and the existence of a $\kat{V}$-natural transformation $f\to g:( X,d_X) \to ( Y, d_Y)$ means, again, that $\bigvee _x d_Y (fx,gx)=0$, that is, the distance $d_Y(fx,gx)$ is $0$, for every $x\in X$. \item Using the quantale $\kat{V}$ from Example~\ref{ex:quantale}\eqref{ex:quantale_prob} leads to {\em probabilistic metric spaces\/}: for a $\kat{V}$-category $\kat{X}$, and for every pair $x$, $x'$ of objects of $\kat{X}$, the hom-object is a function % $\kat{X}(x',x):[0,\infty]\to [0,1]$ % with the intuitive meaning $\kat{X}(x',x)(r)=s$ holds iff $s$ is the probability that the distance from $x'$ to $x$ is smaller than $r$. See~\cite{flag-kopp:continuity-spaces,hofmann+reis}. \item The categories enriched in the free quantale $\mathcal P(\mathsf M)$ of Example~\ref{ex:quantale}\eqref{ex:free-quant} can be perceived as $\mathsf M$-labeled automata~\cite{betti,rosenthal:paper}. The ``distance'' between two states is the language connecting them. In particular, if $\mathsf M$ is the free monoid over an alphabet, the enriched hom $\kat{X}(x,y)$ between two states $x,y$ of an automaton $\kat{X}$ is the language accepted by the automaton $\kat{X}$, considered with initial state $x$ and final state $y$. However, this is an example outside the scope of the paper, since in this case $\kat{V}=\mathcal P(\mathsf M)$ is not commutative. \end{enumerate} \end{exas} \subsection{$\kat{V}$-categories as relational presheaves} \label{sec:relpresh} \ \bigskip\noindent A $\kat{V}$-category $\kat{X}$ cannot only be considered as generalised metric space, but also as a set $X$ equipped with a collection $(X_r)_{r\in\kat{V}_o}$ of binary relations given by \begin{equation}\label{eq:Xr} X_r = \{(x',x)\in X \times X \mid r\le \kat{X}(x',x)\}. \end{equation} In the case of the generalised metric spaces discussed above, the collection% \footnote{ $X_\infty, X_0$ are redundant since $\infty$ is bottom and $X_0=\bigcap\{X_r\mid r>_{\mathbb R} 0\}$.} $(X_r)_{0<r<\infty}$ can be considered as a basis for the quasi-uniformity \cite{fletcher} associated with $\kat{X}$. But interest in the collection $(X_r)_{r\in\kat{V}}$ can also arise from considerations independent of the view of $\kat{V}$-categories as metric spaces as the following example from concurrency theory demonstrates. \begin{exa}\label{exle:casley} A category $\kat{X}$ enriched over the quantale $\underline n$ of Example~\ref{ex:quantale}\eqref{ex:quantale_three} can be understood as a set equipped with $n-1$ transitive relations $$X_{n-1}\subseteq \ldots \subseteq X_i\subseteq \ldots \subseteq X_1.$$ given by $X_i=\{(x',x)\mid i\le \kat{X}(x',x) \}$, $0<i<n$. Following \cite{Casley-phd,Casley-etal}, the $\kat{V}$-enriched categories for the three-element quantales $\kat{V}=(\underline 3, \otimes, e, [-,-])$ of Example~\ref{ex:quantale}\eqref{ex:quantale_three} can be interpreted as well-known models of concurrency given by sets equipped with two relations (in the terminology of the $(X_r)_{r\in\kat{V}_o}$ above, the relation $X_0$ is redundant since $0$ is bottom). \medskip\noindent The first case, which is determined by $e=1$, accounts for the {\em prossets} of Gaifman and Pratt~\cite{gaifman-pratt}. Explicitly, the objects of a $\kat{V}$-category can be seen as events subject to a schedule, endowed with a preorder ${x\leq y}$ given by $X_1$ (with the interpretation that ``$x$ happens no later than $y$'') and a binary relation $x\prec y$ given by $X_2$ (which is intended to mean ``$x$ happens strictly earlier than $y$''). Then $X_2\subseteq X_1$ says that strict precedence implies weak precedence, while the multiplication law $1\otimes 2=2$ reflects the prosset-law that $x\le y \prec z \le w$ implies $x\prec w$.% \footnote{See also~\cite[Definition I-1.11]{ghklms:ContLatDom-book}, where $\prec$ is called an auxiliary relation.} % \medskip\noindent The second case, which is determined by $e=2$, is due to Gaifman~\cite{gaifman}. The relation $x<^ty$ given by $X_1$ is interpreted as ``$x$ precedes $y$ in time'' and the relation $x<^c y$ given by $X_2$ is interpreted as ``$x$ causally precedes $y$''. $X_2\subseteq X_1$ captures that causal precedence implies temporal precedence and the multiplication law $1\otimes 2=1$ reflects that $x<^t y <^c z <^t w$ implies $x<^t w$. \qed \end{exa} \medskip\noindent The idea illustrated in the two examples above can be formalised as a relational presheaf \cite[Chapter~3.4]{rosenthal}, that is, as a (lax) monoidal functor \begin{align*} X_\_ : \kat{V} & \to\mathsf{Rel}\\ r \, & \ \mapsto \ X_r \end{align*} which satisfies \begin{gather} \label{eq:monoidal1} {\mathsf{Id}} \subseteq X_e\\ \label{eq:monoidal2} X_r \cdot X_s\subseteq X_{r\otimes s} \end{gather} Moreover such a functor comes from a $\kat{V}$-category iff $$ X_{\bigvee_{i\in I} r_i} = \bigcap_{i\in I} X_{r_i} $$ that is, the presheaf is continuous. We present this result in some detail, because it is related to the $\kat{V}$-nerves of Definition~\ref{def:Vcoinserter} and because it prepares the grounds for Theorem~\ref{thm:wpbcontrelpresh}. \medskip\noindent We will follow Chapter~3.4 of~\cite{rosenthal}, specialised to an one-object quantaloid, that is, to a commutative quantale $\kat{V}=(\kat{V}_o,\otimes,[-,-],e)$. \medskip\noindent Let $\Sigma\kat{V}$ denote the {\em suspension\/} of $\kat{V}$. That is, $\Sigma\kat{V}$ is an {\em ordered\/} category on one object $*$, with $\Sigma\kat{V}(*,*)=\kat{V}_o$. The composition in $\Sigma\kat{V}$ is given by $\otimes$ and the unit $e$ serves as an identity. Further, let $\mathsf{Rel}$ be the 2-category of {\em ordinary\/} relations. \medskip\noindent The {\em ordinary\/} category $\mathsf{RelPresh}$ of {\em relational presheaves\/} and their morphisms is defined as follows: \begin{enumerate} \item A relational presheaf $\frak{X}:(\Sigma\kat{V})^\mathsf{coop}\to\mathsf{Rel}$ is a {\em lax\/} 2-functor. That is, if we denote $X=\frak{X}(*)$, the inclusions $$ {\mathsf{Id}}_X\subseteq \frak{X}(e), \quad \quad \frak{X}(r)\cdot \frak{X}(s)\subseteq \frak{X}(r\otimes s) $$ hold in $\mathsf{Rel}(X,X)$, with ``$\cdot$'' denoting relational composition. Moreover, the inclusion \begin{equation}\label{eq:monot-rel-presheaf} \frak{X}(r)\subseteq \frak{X}(s) \end{equation} holds whenever $r\geq s$ holds in $\kat{V}$. \item A relational morphism from $\frak{X}$ to $\frak{Y}$ is a lax natural transformation $\frak f$ from $\frak{X}$ to $\frak{Y}$ that has maps as components. That is, if we denote $X=\frak{X}(*)$, $Y=\frak{Y}(*)$ and write $f$ for $\frak{f}_{*}$, then for every $r$ the inclusion \begin{equation}\label{eq:rel-morph} \vcenter{ \xymatrix{X \ar[d]|-{\object @{/}}_{\frak{X}(r)} \ar[r]|-{\object @{/}}^{f_\diamond} \ar@{}[dr]|{\subseteq} & Y \ar[d]|-{\object @{/}}^{\frak{Y}(r)} \\ X \ar[r]|-{\object @{/}}_{f_\diamond} & Y }} \end{equation} holds. Above, $f_\diamond$ is the graph relation of $f:X\to Y$. \end{enumerate} \begin{exa} \label{ex:cat-as-presheaf} Every small $\kat{V}$-category $\kat{X}$ can be turned into a relational presheaf $$ \Phi(\kat{X}):(\Sigma\kat{V})^\mathsf{coop}\to\mathsf{Rel} $$ as follows: \begin{enumerate} \item Let $\Phi(\kat{X})(*)=X_{{\sf o}}$, where we write $X_{{\sf o}}$ for the set of all objects of $\kat{X}$.\footnote{ In the introduction, in the special case of preorders, we used the more familiar notation $X_0$ instead of $X_{{\sf o}}$, but in the general case of $\kat{V}$-categories we now have examples where 0 is an element of $\kat{V}$ and the notation $X_0$ has a meaning given by \eqref{eq:Xr}.} \item Put $\Phi(\kat{X})(r)=\{ (x',x)\mid r\leq \kat{X}(x',x)\}$ for every $r$. Then the relations $$ {\mathsf{Id}}_{X_{{\sf o}}}\subseteq \Phi(\kat{X})(e), \quad \quad \Phi(\kat{X})(r)\cdot \Phi(\kat{X})(s) \subseteq \Phi(\kat{X})(r\otimes s) $$ hold in $\mathsf{Rel}(X_{{\sf o}},X_{{\sf o}})$ precisely because $\kat{X}$ is a $\kat{V}$-category. \end{enumerate} Observe that the presheaf $\Phi(\kat{X})$ satisfies an additional condition: the equality $$ \Phi(\kat{X})(\bigvee_{i\in I} r_i) = \bigcap_{i\in I} \Phi(\kat{X})(r_i) $$ holds for any family $\{r_i\mid i\in I\}$ of elements of $\kat{V}$, since $$ \{ (x',x)\mid \bigvee_{i\in I} r_i\leq \kat{X}(x',x)\} = \bigcap_{i\in I} \{ (x',x)\mid r_i\leq \kat{X}(x',x)\} $$ It is easy to see that any $\kat{V}$-functor $f:\kat{X}\to\kat{Y}$ yields a morphism $\Phi(f):\Phi(\kat{X}) \to \Phi(\kat{Y})$ of relational presheaves. Indeed, denote by $f_{{\sf o}}:X_{{\sf o}}\to Y_{{\sf o}}$ the object-assignment of $f$. Then the inclusion $$ (f_{{\sf o}})_\diamond\cdot \Phi(\kat{X})(r) \subseteq \Phi(\kat{Y})(r)\cdot (f_{{\sf o}})_\diamond $$ means the following \[ \mbox{if $r\leq\kat{X}(x',x)$, then $r\leq\kat{Y}(fx',fx)$, for all $x'$, $x$ in $X_{{\sf o}}$} \] But this holds precisely since $f$ is a $\kat{V}$-functor. \end{exa} \begin{defiC}[{\cite[Definition~3.4.1]{rosenthal}}] A relational presheaf $\frak{X}$ is called {\em continuous\/}, if $$ \frak{X}(\bigvee_{i\in I} r_i) \supseteq \bigcap_{i\in I} \frak{X}(r_i) $$ holds for any family $\{r_i\mid i\in I\}$ of elements of $\kat{V}$. The full subcategory of $\mathsf{RelPresh}$ spanned by continuous relational presheaves is denoted by $\mathsf{RelPresh}_c$. \end{defiC} \begin{rem} The inclusion $$ \frak{X}(\bigvee_{i\in I} r_i) \subseteq \bigcap_{i\in I} \frak{X}(r_i) $$ holds for any family $\{r_i\mid i\in I\}$ of elements of $\kat{V}$ and {\em any\/} relational presheaf $\frak{X}$, because \eqref{eq:monot-rel-presheaf} implies that $ \frak{X}(\bigvee_{i\in I} r_i)\subseteq \frak{X}(r_i) $ holds for any $i$. \end{rem} \begin{propC}[{\cite[Proposition~3.4.1]{rosenthal}}]\label{prop:relpresh} The assignment $\kat{X}\mapsto\Phi(\kat{X})$ extends to an {\em ordinary\/} functor $\Phi:(\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}})_o\to\mathsf{RelPresh}_c$ which is an equivalence of categories. \end{propC} \begin{proof} It is easy to see that the processes $\kat{X}\mapsto\Phi(\kat{X})$ and $H\mapsto\Phi(H)$ of Example~\ref{ex:cat-as-presheaf} extend to a functor $\Phi:(\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}})_o\to\mathsf{RelPresh}_c$. Its pseudoinverse $\Psi:\mathsf{RelPresh}_c\to (\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}})_o$ sends a continuous relational presheaf $\frak{X}:(\Sigma\kat{V})^\mathsf{coop}\to\mathsf{Rel}$ into the following $\kat{V}$-category $\Psi(\frak{X})$: \begin{enumerate} \item The set of objects of $\Psi(\frak{X})$ is the set $\frak{X}(*)$. \item For every pair $x'$, $x$ in $\frak{X}(*)$ we put $$ \Psi(\frak{X})(x',x) = \bigvee\{r\mid (x',x)\in \frak{X}(r) \}. $$ Then $\Psi(\frak{X})$ is a $\kat{V}$-category for the following reasons: \begin{enumerate} \item The inequality $$ e \leq \bigvee\{r\mid (x,x)\in \frak{X}(r) \} = \Psi(\frak{X})(x,x) $$ holds since ${\mathsf{Id}}\subseteq \frak{X}(e)$ holds. \item The inequality $$ \Psi(\frak{X})(x'',x')\otimes\Psi(\frak{X})(x',x) \leq \Psi(\frak{X})(x'',x) $$ holds, since \begin{eqnarray*} \Psi(\frak{X})(x'',x')\otimes\Psi(\frak{X})(x',x) &=& \bigvee\{r\mid (x'',x')\in \frak{X}(r) \} \otimes \bigvee\{s\mid (x',x)\in \frak{X}(s) \} \\ &=& \bigvee_{(x'',x')\in \frak{X}(r) } \ \bigvee_{(x',x)\in \frak{X}(s) } r\otimes s \\ &\leq& \bigvee_{(x'',x)\in \frak{X}(r\otimes s) } r\otimes s \\ &=& \Psi(\frak{X})(x'',x) \end{eqnarray*} Above, we have used that $\frak{X}(r)\cdot \frak{X}(s)\subseteq \frak{X}(r\otimes s)$. \end{enumerate} \end{enumerate} That $\Phi$ and $\Psi$ are essentially inverse to each other is verified as follows: \begin{enumerate} \item Suppose a $\kat{V}$-category $\kat{X}$ is given. Then the $\kat{V}$-category $\Psi\Phi(\kat{X})$ has the same set of objects as $\kat{X}$. Moreover, $$ \Psi\Phi(\kat{X})(x',x) = \bigvee\{r\mid r\leq\kat{X}(x',x)\} = \kat{X}(x',x) $$ \item Suppose $\frak{X}:(\Sigma\kat{V})^\mathsf{coop}\to\mathsf{Rel}$ is a relational presheaf. Then the presheaf $\Phi\Psi(\frak{X})$ has the value $\frak{X}(*)$ at $*$ by definition of $\Psi$ and $\Phi$. \medskip\noindent We need to prove that $(x',x) \in \frak{X}(r)$ holds iff $r\leq \bigvee\{ s\mid (x',x)\in \frak{X}(s) \}$ holds. The implication from left to right is trivial. For the converse implication, apply $\frak{X}$ to the inequality $r\leq \bigvee\{ s\mid (x',x) \in \frak{X}(s) \}$ and use continuity of $\frak{X}$: $$ \frak{X}(r) \supseteq \frak{X}\left(\bigvee\{ s\mid (x',x) \in \frak{X}(s) \}\right) \supseteq \bigcap\{ \frak{X}(s)\mid (x',x) \in \frak{X}(s) \} \ni (x',x) $$ \end{enumerate} This finishes the proof. \end{proof} \begin{rem}\label{rem:enrichedRelPresh} The functor $\Psi$ can be extended to an equivalence of {\em $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories\/} (see the next section for an introductory exposition on $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched categories). This means that every $$ \mathsf{RelPresh}_c(\frak{X},\frak{Y}) $$ should have the structure of a $\kat{V}$-category. We do it by transferring this structure from $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, using the equivalence of $(\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}})_o$ and $\mathsf{RelPresh}_c$. More precisely, write $\frak{X}=\Phi(\kat{X})$ and $\frak{Y}=\Phi(\kat{Y})$ for some unique $\kat{V}$-categories $\kat{X}$ and $\kat{Y}$ and put $$ \mathsf{RelPresh}_c(\frak{X},\frak{Y}) = [\kat{X},\kat{Y}]. $$ \end{rem} \subsection{Categories enriched in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$}\label{sec:V-cat-cat} \ \bigskip\noindent Suppose that $\kat{V}=(\kat{V}_o,\otimes,e,[{-},{-}])$ is a commutative quantale and recall that we denoted by $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o$ the (ordinary) category of all small $\kat{V}$-categories and all $\kat{V}$-functors between them. The category $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o$ has a monoidal closed structure: The {\em tensor product\/} is inherited from $\kat{V}_o$. Namely, for $\kat{V}$-categories $\kat{X}, \kat{Y}$, put $\kat{X}\otimes \kat{Y}$ to be the $\kat{V}$-category having as objects the corresponding pairs of objects and $\kat{V}$-homs given by $$ (\kat{X}\otimes \kat{Y})((x',y'),(x,y))= \kat{X}(x',x)\otimes \kat{Y}(y',y) $$ The {\em unit\/} for the tensor product is the $\kat{V}$-category ${\mathbb{1}}$, with one object $0$ and corresponding $\kat{V}$-hom given by ${\mathbb{1}}(0,0)=e$. \medskip\noindent The $\kat{V}$-functor ${-}\otimes \kat{Y}:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o$ has a right adjoint $[\kat{Y},{-}]$. Explicitly, $[\kat{Y},\kat{Z}]$ is the following $\kat{V}$-category: \begin{enumerate} \item Objects of $[\kat{Y},\kat{Z}]$ are $\kat{V}$-functors from $\kat{Y}$ to $\kat{Z}$. \item The $\kat{V}$-``distance'' between two $\kat{V}$-functors $f,g:\kat{Y} \to \kat{Z}$ is \begin{equation}\label{eq:Vcat internal hom} [\kat{Y},\kat{Z}](f,g) = \bigwedge_y \kat{Z}(fy,gy) \end{equation} \end{enumerate} \noindent We will sometimes write $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{Y},\kat{Z})$ or even $\kat{Z}^\kat{Y}$ instead of $[\kat{Y},\kat{Z}]$. \medskip\noindent By~\cite{kelly+lack:locally-bounded}, the symmetric monoidal closed category $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}=(\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o,\otimes,{\mathbb{1}},[{-},{-}])$ is complete and cocomplete, with generators consisting of $\kat{V}$-categories of the form ${\mathbb{2}}_r$, $r\in \kat{V}_o$. Here, every ${\mathbb{2}}_r$ has two objects $0$ and $1$, with $\kat{V}$-homs \begin{equation}\label{eq:Two_r} {\mathbb{2}}_r(0,0)={\mathbb{2}}_r(1,1)=e \, , \, {\mathbb{2}}_r(0,1)=r \, , \, {\mathbb{2}}_r(1,0)=\bot \end{equation} \medskip\noindent Since $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is a symmetric monoidal closed category, we can define $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched categories, $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors and $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformations. \medskip\noindent A ({\em small\/}) {\em $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category\/} ${\mathbb{X}}$ consists of a (small) set of objects $X$, $Y$, $Z$, \dots, a small $\kat{V}$-category ${\mathbb{X}}(X,Y)$ for every pair $X$, $Y$ of objects, and $\kat{V}$-functors $$ u_X:{\mathbb{1}}\to {\mathbb{X}}(X,X), \quad c_{X,Y,Z}:{\mathbb{X}}(Y,Z)\otimes{\mathbb{X}}(X,Y)\to{\mathbb{X}}(X,Z) $$ that represent the identity and composition and satisfy the usual axioms~\cite{kelly:book}: $$ \xymatrix@R=30pt@C=40pt{ {\mathbb{X}}(Z,U)\otimes {\mathbb{X}}(Y,Z) \otimes {\mathbb{X}}(X,Y) \ar[r]^-{{\mathsf{Id}} \otimes c_{X,Y,Z}} \ar[d]_{c_{Y,Z,U} \otimes {\mathsf{Id}}} & {\mathbb{X}}(Z,U)\otimes {\mathbb{X}}(X,Z) \ar[d]^{c_{X,Z,U}} \\ {\mathbb{X}}(Y,U)\otimes {\mathbb{X}}(X,Y) \ar[r]^{c_{X,Y,U}} & {\mathbb{X}}(X,U) } $$ $$ \xymatrix@R=30pt@C=27pt{ {\mathbb{1}} \otimes {\mathbb{X}}(X,Y) \ar[r]^-{u_Y \otimes {\mathsf{Id}}} \ar[dr]_{\cong} & {\mathbb{X}}(Y,Y)\otimes {\mathbb{X}}(X,Y) \ar[d]^{c_{X,Y,Y}} & {\mathbb{X}}(X,Y)\otimes {\mathbb{X}}(X,X) \ar[d]_{c_{X,X,Y}} & {\mathbb{X}}(X,Y) \otimes {\mathbb{1}} \ar[l]_-{{\mathsf{Id}} \otimes u_X} \ar[dl]^\cong \\ & {\mathbb{X}}(X,Y) & {\mathbb{X}}(X,Y) } $$ Objects of ${\mathbb{X}}(X,Y)$ will be denoted by $f:X\to Y$ and their $\kat{V}$-distance by ${\mathbb{X}}(X,Y)(f,g)$ in $\kat{V}$. The action of $c_{X,Y,Z}$ at objects $(f',f)$ in ${\mathbb{X}}(Y,Z)\otimes{\mathbb{X}}(X,Y)$ is denoted by $f'\cdot f$, and for their distances the inequality below (expressing that $c_{X,Y,Z}$ is a $\kat{V}$-functor) holds: $$ \left({\mathbb{X}}(Y,Z)\otimes{\mathbb{X}}(X,Y)\right)( (f',f),(g',g) ) \leq {\mathbb{X}}(X,Z)(f'\cdot f,g'\cdot g) $$ \begin{exa} \label{ex:Vcat-categories} \hfill \begin{enumerate} \item Start with the simplest quantale, namely $\kat{V}={\mathbb{2}}$. We have seen in Example~\ref{exle:V-categories}\eqref{exle:Twocat} that preorders are ${\mathbb{2}}$-categories, and that monotone maps are ${\mathbb{2}}$-functors, hence that $\Two\mbox{-}\hspace{.9pt}{\mathsf{cat}}={\mathsf{Preord}}$. Then $\Two\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories are categories with ordered homsets, such that composition is monotone in both arguments. Examples of $\Two\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories are the category $\Two\mbox{-}\hspace{.9pt}{\mathsf{cat}}={\mathsf{Preord}}$ of preorders itself, the category ${\mathsf{Set}}$ of sets with the discrete enrichment (see Section~\ref{sec:discrete functor} on the discrete enrichment of ${\mathsf{Set}}$, discussed for general $\kat{V}$), as well as any $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category, by~\eqref{eq:V-metric=>order}. In particular, $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ as such can be seen $\Two\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched (for more details, we refer to Section~\ref{sec:beh-cat}, where this situation, known as {\em change-of-base}, is discussed). \item The $\boldsymbol{O}$-categories of Smyth and Plotkin~\cite{smyth-plotkin} are special cases of ${\mathsf{Preord}}$-enriched categories, in the sense that the hom-sets are not only preorders, but actual partial orders such that every ascending $\omega$-sequence has a 1.u.b. and composition of morphisms is $\omega$-continuous. \item \label{ex:Vcat as enriched over itself} $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ itself is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category (see~\cite[Section~1.6]{kelly:book} and also Example~\ref{exle:V-categories}\eqref{ex:V as V-category itself}). \item\label{ex:metric-enriched} Let $\kat{V}=([0,\infty],\geq_{\mathbb R},+,0)$ be the quantale of Example~\ref{exs:Vcats=metric_spaces}\eqref{ex:V=[0,infty]}. As explained there, $\kat{V}$-categories are generalised metric spaces, and $\kat{V}$-functors are non-expanding maps, while $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories are known as locally metric categories, or metric-enriched categories -- an example of such being the (sub)category of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ consisting of complete and bounded-by-1 metric spaces $\mathsf{CMS}$, with embedding-projection pairs as arrows~\cite{america-rutten,wagner}. $\mathsf{CMS}$ provides an appropriate context for studying reflexive quantitative domain equations~\cite{america-rutten,flag-kopp:continuity-spaces,turi-rutten,wagner}. The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category structure of $\mathsf{CMS}$ is inherited from $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ itself. \end{enumerate} \end{exa} \medskip\noindent A {\em $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor\/} $F:{\mathbb{X}}\to{\mathbb{Y}}$ is given by: \begin{enumerate} \item The assignment $X\mapsto FX$ on objects. \item For each pair of objects $X,Y$ in ${\mathbb{X}}$, a $\kat{V}$-functor $F_{X,Y}:{\mathbb{X}}(X,Y)\to{\mathbb{Y}}(FX,FY)$, whose action on objects $f:X\to Y$ is denoted by $Ff:FX\to FY$. For the distances we have the inequality $$ {\mathbb{X}}(X,Y)(f,g) \leq {\mathbb{Y}}(FX,FY)(Ff,Fg) $$ \end{enumerate} \noindent Of course, the diagrams of $\kat{V}$-functors below, expressing the preservation of unit and composition, should commute: $$ \xymatrix@R=17pt@C=10pt{ {\mathbb{X}}(X,X) \ar[0,2]^-{F_{X,X}} & & {\mathbb{Y}}(FX,FX) \\ & {\mathbb{1}} \ar[-1,-1]^{u_X} \ar[-1,1]_{u_{FX}} & } \quad \xymatrix@R=17pt@C=17pt{ {\mathbb{X}}(Y,Z)\otimes{\mathbb{X}}(X,Y) \ar[0,2]^-{F_{Y,Z}\otimes F_{X,Y}} \ar[1,0]_{c_{X,Y,Z}} & & {\mathbb{Y}}(FY,FZ)\otimes{\mathbb{Y}}(FX,FY) \ar[1,0]^{c_{ FX,FY,FZ }} \\ {\mathbb{X}}(X,Z) \ar[0,2]_-{F_{X,Z}} & & {\mathbb{Y}}(FX,FZ) } $$ \medskip\noindent Given $F,G:{\mathbb{X}}\to{\mathbb{Y}}$, a {\em $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation} $\tau:F\to G$ is given by a collection of $\kat{V}$-functors $\tau_X:{\mathbb{1}}\to{\mathbb{Y}}(FX,GX)$, such that the diagram $$ \xymatrix@C=37pt@R=20pt{ & {\mathbb{1}}\otimes{\mathbb{X}}(X,Y) \ar[0,1]^-{\tau_{Y}\otimes F_{X,Y}} & {\mathbb{Y}}(FY,GY)\otimes {\mathbb{Y}}(FX,FY) \ar[1,1]^{\phantom{MM}c_{FX,FY,GY}} & \\ {\mathbb{X}}(X,Y) \ar[-1,1]^{\cong} \ar[1,1]_{\cong} & & & {\mathbb{Y}}(FX,GY) \\ & {\mathbb{X}}(X,Y)\otimes {\mathbb{1}} \ar[0,1]_-{G_{X,Y}\otimes\tau_{X}} & {\mathbb{Y}}(GX,GY)\otimes{\mathbb{Y}}(FX,GX) \ar[-1,1]_{\phantom{MM}c_{FX,GX,GY}} & } $$ of $\kat{V}$-functors commutes. We will abuse the notation and denote by $\tau_X:FX\to GX$ the image in ${\mathbb{Y}}(FX,GX)$ of the object $0$ from ${\mathbb{1}}$ under the $\kat{V}$-functor $\tau_X:{\mathbb{1}}\to{\mathbb{Y}}(FX,GX)$. The above diagram (when read at the object-assignments of the ambient $\kat{V}$-functors) then translates as the equality \begin{equation}\label{eq:Vcat-nat} Gf \cdot \tau_{X}=\tau_{Y} \cdot Ff \end{equation} of objects of the $\kat{V}$-category ${\mathbb{Y}}(FX,GY)$, for every object $f:X \to Y$. On hom-objects, the above diagram says nothing\footnote{ This is well-known for ${\mathsf{Preord}}$-natural transformations: one only needs to verify ordinary naturality.} (recall that $\kat{V}_o$ is a poset, hence there are no parallel pairs of morphisms in $\kat{V}_o$). \begin{exa}\label{ex:V-cat-functor} \hfill \begin{enumerate} \item \label{ex:Two-cat-functor} We consider again first the case $\kat{V}={\mathbb{2}}$. A $\Two\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor, that is, a ${\mathsf{Preord}}$-functor, is also known as a locally monotone functor. For example, the usual discrete functor $D:{\mathsf{Set}}\to{\mathsf{Preord}}$ is a ${\mathsf{Preord}}$-functor, but the forgetful functor $V:{\mathsf{Preord}}_o\to{\mathsf{Set}}_o$ is not, because $V$ is not locally monotone (the mapping $V_{X,Y}:{\mathsf{Preord}}(X,Y)\to{\mathsf{Set}}(VX,VY)$ is not order-preserving). \item The hom-contracting functors on $\mathsf{CMS}$, employed in the metric domain equations of~\cite{america-rutten}, are in fact $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched functors, for $\kat{V}=[0,\infty]$ (see Example~\ref{ex:Vcat-categories}\eqref{ex:metric-enriched}). \end{enumerate} \end{exa} \subsection{On the fully-faithfulness of the discrete functor $D:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} $}\label{sec:discrete functor} \ \bigskip\noindent This section studies in detail the observation that the discrete functor $$D:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$$ is fully faithful if and only if $\kat{V}$ is integral, that is, $e=\top$.% \footnote{All our examples are integral with the exception of the last two items of Example~\ref{ex:quantale}.} Indeed, if $e<\top$ then $${\mathsf{Set}}(\emptyset,\emptyset)({\mathsf{Id}},{\mathsf{Id}}) = e \quad\quad\textrm{and}\quad\quad \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\mathbb 0,\mathbb 0)({\mathsf{Id}},{\mathsf{Id}}) = \top$$ where $\mathbb 0 = D\emptyset$ denotes the empty $\kat{V}$-category. \medskip\noindent This contradicts the fully faithfulness of $D$, which is defined in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ via the existence of an isomorphism ${\mathsf{Set}}(A,B)\cong \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DA,DB)$ and implies $${\mathsf{Set}}(A,B)(f,g) = \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DA,DB)(Df,Dg).$$ \noindent To make the argument above precise, we need to explain in what sense ${\mathsf{Set}}$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category and in what sense $D$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor. The remainder of this section formalises the concept of discreteness within $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ following~\cite{eilenberg-kelly, kelly:book} and may be skipped by a reader who accepts the argument as sketched above. \medskip\noindent Let ${\mathsf{Set}}$, for now, denote the ordinary category of sets and functions. The ordinary forgetful functor $V:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o\to{\mathsf{Set}}$, mapping each $\kat{V}$-category to its set of objects, has a left adjoint~\cite[Section~2.5]{kelly:book}, known as the free $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category functor (or, as we will call it, {\em the discrete functor}) $$D:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o$$ \medskip\noindent The functor $D$ maps each set $X$ to the $\kat{V}$-category $X\cdot \mathbb 1$, the coproduct in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ of $X$-copies of the unit $\kat{V}$-category $\mathbb 1$ (recall that $\mathbb 1$ has one object $0$, with $\kat{V}$-self-distance $\mathbb 1(0,0)=e$). That means that each $\kat{V}$-category $DX$ is discrete, as in~\eqref{eq:discrete V-category}. In particular, $D\emptyset$ is the empty $\kat{V}$-category $\mathbb 0$. \medskip\noindent By~\cite[Section~2.5]{kelly:book}, $D$ is strong monoidal: \[ D(X \times Y) \cong DX \otimes DY, \quad D1 \cong {\mathbb{1}} \] hence it induces a (2-)functor $$D_\ast:{\mathsf{Set}}\hspace{.9pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\hspace{.9pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$$ between ordinary categories (i.e., ${\mathsf{Set}}$-enriched) and $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched categories~\cite[Section~6]{eilenberg-kelly}. The functor $D_\ast$ maps an ordinary category $\mathcal C$ to the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category with the same set of objects as $\mathcal C$, and $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-homs $(D_\ast\mathcal C)(f,g)=D(\mathcal C(f,g))$. In particular, ${\mathsf{Set}}$ itself gets enriched to a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category $D_\ast{\mathsf{Set}}$ with sets as objects, and for any sets $X,Y$, $$(D_\ast{\mathsf{Set}})(X,Y)=D({\mathsf{Set}}(X,Y))={\mathsf{Set}}(X,Y)\cdot \mathbb 1$$ Notice that $(D_\ast{\mathsf{Set}})(\emptyset,\emptyset) =\mathbb 1$ (as there is only one map $\emptyset\to\emptyset$, namely the identity) \medskip\noindent Then $D$ can be perceived as a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $$\widehat{D}:D_\ast {\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$$ with same action as $D$ on objects (i.e. it maps a set $X$ to the $\kat{V}$-category $DX=X\cdot \mathbb 1$). Saying that $\widehat D$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor means that for each sets $X,Y$, there is a $\kat{V}$-functor $$\widehat{D}_{X,Y}:(D_\ast{\mathsf{Set}})(X,Y)\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DX,DY)$$ induced by the monoidal (closed) structure of $D$, and this collection of $\kat{V}$-functors is compatible with identity and composition. Explicitly, $\widehat{D}_{X,Y}$ maps a function $f:X\to Y$ to the $\kat{V}$-functor $Df:DX \to DY$, the $\kat{V}$-functor structure of $\widehat{D}_{X,Y}$ being witnessed by the inequality $$(D_\ast{\mathsf{Set}})(X,Y)(f,g)\leq \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DX,DY)(Df,Dg)$$ In particular, there is a $\kat{V}$-functor $$\widehat{D}_{\emptyset,\emptyset}:(D_\ast{\mathsf{Set}})(\emptyset,\emptyset)=\mathbb 1 \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(D\emptyset,D\emptyset)$$ Observe now that the only $\kat{V}$-functor $D\emptyset \to D\emptyset$ is the identity ${\mathsf{Id}}$, and that $$\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(D\emptyset,D\emptyset)({\mathsf{Id}},{\mathsf{Id}})=\top$$ hence $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(D\emptyset, D\emptyset)$ is (isomorphic to) the $\kat{V}$-category $\mathbb 1_\top$ with one object $0$ and $\mathbb 1_\top(0,0)=\top$. Therefore we have $$\widehat{D}_{\emptyset,\emptyset}:\mathbb 1 \to \mathbb 1_\top$$ This situation extends to \begin{equation}\label{eq:D not ff} \widehat{D}_{\emptyset, Y}: (D_\ast{\mathsf{Set}})(\emptyset, Y)= {\mathsf{Set}}(\emptyset, Y)\cdot \mathbb 1 = \mathbb 1 \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(D\emptyset,DY)=\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\mathbb 0,DY)=\mathbb 1_\top \end{equation} for arbitrary set $Y$. \medskip\noindent In view of the observation above, we may formulate the following result, vital for the development of the paper: \begin{prop}\label{prop:D-ff} The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $\widehat{D}:D_\ast{\mathsf{Set}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is fully-faithful if and only if the quantale $\kat{V}$ is integral. \end{prop} \begin{proof} It is easy to see that for non-empty $X$, $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DX,DY)$ is a discrete $\kat{V}$-category, hence \[ \widehat{D}_{X, Y}: (D_\ast{\mathsf{Set}})(X, Y)= {\mathsf{Set}}(X, Y)\cdot \mathbb 1 \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DX,DY) \] is an isomorphism of $\kat{V}$-categories. The result now follows from~\eqref{eq:D not ff}, using that $e=\top$ iff $\mathbb 1 \cong \mathbb 1_\top$. \end{proof} \begin{rem}\label{ex:discrete V-cat-functor} \hfill \begin{enumerate} \item Notice that not only $D$, but actually the whole adjunction $D\dashv V$ lifts to an adjunction \[ D_\ast \dashv V_\ast : \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\hspace{.9pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to {\mathsf{Set}}\hspace{.9pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \] where $V_\ast$ maps a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category to its underlying ordinary category. That means that ordinary functors, like ${\mathsf{Set}}\to {\mathsf{Set}} = V_\ast D_\ast {\mathsf{Set}}$ or ${\mathsf{Set}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o = V_\ast \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, and ordinary natural transformations between such, automatically get $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ enriched to $D_\ast {\mathsf{Set}} \to D_\ast {\mathsf{Set}}$ or $D_\ast {\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, and so on. \item It will be important in the sequel that although $D$ acquires an enrichment to the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $\widehat D$, its ordinary right adjoint $V$ does not. Hence $\widehat D$ lacks a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched right adjoint. \end{enumerate} \end{rem} \paragraph{\bf Notation.} For simplicity, we write in the sequel ${\mathsf{Set}}$ for the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched category $D_\ast{\mathsf{Set}}$ and $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ for the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $\widehat{D}:D_\ast {\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. When referring to ${\mathsf{Set}}$ or $D$ as an ordinary category/functor, we will denote them by ${\mathsf{Set}}_o$ and $D_o:{\mathsf{Set}}_o\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o$, respectively. This agrees with the convention set at the beginning of the paper, because $V_\ast D_\ast = {\mathsf{Id}}$, as the reader can easily check. \subsection{On $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched colimits}\label{sec:colimit} \ \bigskip\noindent The last bit of standard notation from enriched category theory concerns colimits. We recall it for % $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories. \begin{defi}\label{def:colimit} A {\em colimit\/} of a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $F:{\mathbb{A}}\to{\mathbb{X}}$, where ${\mathbb{A}}$ is small, weighted by a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $W:{\mathbb{A}}^{\mathit{op}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, consists of an object $\Colim{W}{F}$ of ${\mathbb{X}}$, together with an isomorphism in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ \[ {\mathbb{X}}(\Colim{W}{F},X)\cong [{\mathbb{A}}^{\mathit{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W,{\mathbb{X}}(F-,X)) \] which is $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural in $X$.% \footnote{ Above, $[{\mathbb{A}}^{\mathit{op}}, \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]$ is the usual $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category of presheaves over ${\mathbb{A}}$.}% \qed \end{defi} \medskip\noindent In case ${\mathbb{A}}$ is the unit $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category,% \footnote{ Having only one object, with corresponding $\kat{V}$-category-hom ${\mathbb{1}}$.} % we may identify $F$ with an object $X$ of ${\mathbb{X}}$ and $W$ with a $\kat{V}$-category $\kat{A}$. % The resulting weighted colimit, usually called the {\em copower} (or the {\em tensor}) of $\kat{A}$ with $X$, shall be denoted then $\kat{A}\bullet X$ instead of $\Colim{W}{F}$. \medskip\noindent To emphasise the importance of \emph{weighted} colimits, recall that in ordinary category theory, in particular in algebraic theories, quotienting by a set of equality constraints produces a coequalizer (an ordinary colimit). Contrary, in the simplest $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched setting -- that is, ${\mathsf{Preord}}$-enriched, quotienting by a set of inequality constraints is an example of a weighted colimit called {\em coinserter}, which cannot be obtained using only ordinary colimits. % \begin{exa}\label{exle:coinserter} In the case $\kat{V}={\mathbb{2}}$, for which $\Two\mbox{-}\hspace{.9pt}{\mathsf{cat}}={\mathsf{Preord}}$, a coinserter is a colimit that has as weight $\varphi:{\mathbb{A}}^{\mathit{op}} \to {\mathsf{Preord}}$, where $\mathbb A$ is \[ \xymatrix{\cdot \ar@<0.5ex>[r] \ar@<-0.5ex>[r] & \cdot} \] and $W$ maps ${\mathbb{A}}$ to the parallel pair \[ \xymatrix{\mathbb 2 &\ar@<-0.5ex>[l]_1 \ar@<0.5ex>[l]^0 \mathbb 1} \] in ${\mathsf{Preord}}$, with arrow $0$ mapping to $0\in\mathbb 2$ and arrow $1$ mapping to $1\in\mathbb 2$ (recall that $\mathbb 2$ is the preorder $\{0 < 1\}$). \end{exa} \noindent This example is of importance to us because the next section is based on the $\kat{V}$-generalisation of the observation that every preorder is a coinserter of discrete preorders, which was used in \cite{lmcs:bkv} to show that ${\mathsf{Set}}$ is dense in ${\mathsf{Pos}}$ (and also in ${\mathsf{Preord}}$). \begin{exa}\label{exle:preord-coins-discrete} Let $X$ be a poset (or a preordered set). Denote by $X_0=DVX$ the discrete preorder of the elements of $X$, and by $X_1$ the discrete preorder of all comparable pairs, $X_1=\{ (x',x)\in X \mid x'\leq x\}=DV(X^{\mathbb{2}})$.\footnote{ $X^{\mathbb{2}}={\mathsf{Preord}}({\mathbb{2}}, X)$ is the ordered set of monotone maps ${\mathbb{2}} \to X$.} Let $d_0, d_1:X_1\to X_0$ be the two projections. Then the obvious map $q:X_0\to X$, $q(x)=x$, exhibits $X$ as the coinserter \begin{equation} \label{eq:poset_coins} \xymatrix@C=35pt{ X_1 \ar@<0.5ex>[r]^{d_0} \ar@<-0.5ex>[r]_{d_1} & X_0 \ar[r]^-q & X} \end{equation} in ${\mathsf{Preord}}$ of the diagram \begin{equation} \xymatrix@C=35pt{ X_1 \ar@<0.5ex>[r]^{d_0} \ar@<-0.5ex>[r]_{d_1} & X_0. } \end{equation} of discrete posets $X_1,X_0$. \qed \end{exa} \medskip\noindent To follow the technical developments of the next section, it is worth understanding which part of Definition~\ref{def:colimit} forces the colimit $X$ in \eqref{eq:poset_coins} to be ordered. If we instantiate Definition~\ref{def:colimit} with ${\mathbb{X}}={\mathsf{Preord}}$ and $W$, $F$ as in Example~\ref{exle:coinserter}, then it is the monotonicity of the component ${\mathbb{2}}\to{\mathsf{Preord}}(X_1,X)$ of the natural transformation in $$[{\mathbb{A}}^{\mathit{op}},{\mathsf{Preord}}](W,{\mathsf{Preord}}(F-,X))$$ that gives $c \circ d_0\le c \circ d_1$. \bigskip\noindent Having introduced weighted colimits, we are now in the position to talk about Kan extensions: \begin{defi}[Kan extension] \label{def:Lan} Let $J:\mathbb A \to \mathbb B$, $H:\mathbb A \to \mathbb X$ be $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors. A $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched \emph{left Kan extension of $H$ along $J$}, is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $\Lan{J}{H}:\mathbb B\to \mathbb X$, such that there is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural isomorphism \begin{equation}\label{eq:lan} (\Lan{J}{H}) B \cong \mathbb B(J-, B) * H \end{equation} for each $B$ in $\mathbb B$. \end{defi} \begin{rem} \label{rem:lan} \hfill \begin{enumerate} \item According to Definition~\ref{def:colimit}, Equation~\eqref{eq:lan} above simply says that there is an isomorphism \[ \mathbb X ((\Lan{J}{H})B,X) \cong [\mathbb A^{\mathit{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](\mathbb B(J-, B), \mathbb X(H-, X)) \] in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, natural in $B$ and $X$. \item \label{rem:weak_def_lan} For any $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H':\mathbb B\to \mathbb X$, there is a bijection between $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformations $\Lan{J}{H}\to H'$ and $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformations $H \to H' J$, in analogy to the case of ordinary left Kan extensions. In particular, there is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $\alpha:H\to (\Lan{J}{H} ) J$, called the {\em unit\/} of the left Kan extension, which is {\em universal\/} in the sense that for a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H':\mathbb B \to \mathbb X$, any $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $H\to H' J$ factorises through $\alpha$. If the codomain ${\mathbb{X}}$ is $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ itself, the above bijection can be taken as an alternative definition of the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched left Kan extension (see the discussion after Theorem~4.43 in~\cite{kelly:book}). \item \label{unit of Lan iso} If $J:\mathbb A \to \mathbb B$ is fully faithful, then the unit $\alpha:H \to ( \Lan{J}{H} )J$ of the left Kan extension is an isomorphism~\cite[Proposition~4.23]{kelly:book}. \item The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched left Kan extension $\Lan{J}{H}$ exists whenever $\mathbb A$ is small and $\mathbb X$ is cocomplete~\cite[Section~4.1]{kelly:book}. But it might exist even when $\mathbb A$ is not small, as we will see later in a special case (Theorem~\ref{thm:Lan}). \qed \end{enumerate} \end{rem} \section{Extending functors from ${\mathsf{Set}}$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$} \label{sec:extension} As explained in the introduction, we are interested in left Kan extensions $$ \xymatrix@R=20pt{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \ar[0,2]^-{\Lan{D}{(DT)}} & & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \\ {\mathsf{Set}} \ar[rr]_-{T} \ar[-1,0]^-{D} & & {\mathsf{Set}}\ar[u]_D } \quad\quad \quad\quad \xymatrix@R=20pt{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \ar[0,2]^-{\Lan{D}{H}} & & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \\ {\mathsf{Set}} \ar[-1,2]_-{H} \ar[-1,0]^-{D} & & } $$ where $\kat{V}=(\kat{V}_o,\otimes,e,[{-},{-}])$ is again a commutative quantale and $D:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is the discrete functor mapping a set $X$ to the discrete $\kat{V}$-category $DX$, that is, the $\kat{V}$-category that has $X$ as set of objects, all self-distances $e$, and all other distances bottom, as in Section~\ref{sec:discrete functor}. \medskip\noindent We call $\Lan{D}{(DT)}$ in the left-hand diagram the \emph{$\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification} of $T$. \medskip\noindent It is now important to note that we are interested in $\Lan{D}{(DT)}$ and, more generally, in $\Lan{D} H$, in the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched sense and not in the ordinary sense. Given that the ordinary functor $D_o$ has the forgetful functor $V:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o\to{\mathsf{Set}}_o$ as a right adjoint, the {\em ordinary} left Kan extension of $H_o$ along $D_o$, is $H_oV$. We observe that $\Lan{D_o} H_o =H_oV$ is not interesting from a metric point of view as $H_oV(\kat{X})$ does not depend on the metric of $\kat{X}$. But, crucially, $D$ does not have an \emph{enriched} right adjoint and, as we will see, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched left Kan extension $\Lan{D}H$ is rather more interesting than $\Lan{D_o}H_o=H_oV$. \subsection{Preliminaries on extensions and liftings} \ \bigskip\noindent Here we will have a closer look at extensions along $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ and liftings along $V:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o\to{\mathsf{Set}}_o$. Notice that the ordinary adjunction $D_o\dashv V:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o\to{\mathsf{Set}}_o$ exhibits ${\mathsf{Set}}_o$ as a full coreflective subcategory of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o$. \begin{defi}\label{def:ext-lift} Let $T:{\mathsf{Set}}\to {\mathsf{Set}}$, $\ol T:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ be $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors. \begin{enumerate} \item \label{def:ext} We say that a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural isomorphism $$ \xymatrix@R=20pt{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \ar[0,2]^-{\ol{T}} & & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \\ {\mathsf{Set}} \ar[0,2]_-{T} \ar[-1,0]^{D} \ar@{}[-1,2]|{\nwarrow\alpha} & & {\mathsf{Set}} \ar[-1,0]_{D} } $$ of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors exhibits $\ol{T}$ as an {\em extension\/} of $T$. \noindent If a natural (iso)morphism $\alpha:DT\to \ol T D$ happens to be the unit of a left Kan extension, that is, if $\ol{T}\cong \Lan{D}{(D T)}$ holds, then we say that $\alpha$ exhibits $\ol{T}$ as the {\em $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification\/} of $T$, and we will denote it by $T_\kat{V}$. \item \label{def:lift} We say that a natural isomorphism $$ \xymatrix@R=20pt{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o \ar[0,2]^-{\ol{T}_o} & & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o \\ {\mathsf{Set}}_o \ar[0,2]_-{T_o} \ar@{<-}[-1,0]^{V} \ar@{}[-1,2]|{\nearrow\beta} & & {\mathsf{Set}}_o \ar@{<-}[-1,0]_{V} } $$ of ordinary functors exhibits $\ol{T}$ as a {\em lifting\/} of $T$. \end{enumerate} \end{defi} \begin{exas}\label{ex:extensions} \hfill \begin{enumerate} \item The identity $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor ${\mathsf{Id}}:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is an extension and a lifting of the identity functor on ${\mathsf{Set}}$. We will see later that ${\mathsf{Id}}:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is universal among all extensions of ${\mathsf{Id}}:{\mathsf{Set}}\to{\mathsf{Set}}$, in that it is the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification (for an arbitrary commutative quantale $\kat{V}$) of the identity functor on ${\mathsf{Set}}$. \item \label{ex:unicity-of-extens} If $D:{\mathsf{Set}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ has a left adjoint $C\dashv D$,% \footnote{ In case $\kat{V}={\mathbb{2}}$, the left adjoint $C:{\mathsf{Preord}}\to{\mathsf{Set}}$ takes connected components. For the general situation see Remark~\ref{rem:change-of-base-cd}. } % then not only the identity functor ${\mathsf{Id}}$ on $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ but also the composite $DC$ is an extension of the identity functor (but not a lifting!). \item The convex powerset functor $\overline{\mathcal P}:{\mathsf{Pos}}\to{\mathsf{Pos}}$ is an extension, in fact the posetification \cite{lmcs:bkv}, of $\mathcal P:{\mathsf{Set}}\to{\mathsf{Set}}$, but is not a lifting. On the other hand, the extension of the powerset functor to ${\mathsf{Preord}}\to{\mathsf{Preord}}$ is also a lifting, as are all $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications (see ). \item Every functor $T:{\mathsf{Set}} \to {\mathsf{Set}}$ admits the lifting $\ol T_\top:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ mapping a $\kat{V}$-category $\kat{X}$ to the $\kat{V}$-category with set of objects $TX_{{\sf o}}$, where $X_{{\sf o}}$ is the set of objects of $\kat{X}$,% \footnote{% See Proposition~\ref{prop:Vcoinserter}.} % and $\ol T_\top\kat{X}(A',A)=\top$ for all objects $A', A $. In fact, $\ol T_\top$ is final among all liftings of $T$, in in the sense that for any other lifting $\ol T$ with $\beta:T_o V \to V\ol T_o$, there is a unique $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $\gamma:\ol T \to \ol T_\top$ such that $\beta_\top = V\gamma_o \cdot \beta$, where $\gamma _o$ is the ordinary natural transformation associated to $\gamma$, and $\beta_\top:T_o V \to V{\ol T_\top}_o$ is the isomorphism corresponding to $\ol T_\top$. \item The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $T_\kat{V}$ exists for every {\em accessible\/} functor $T:{\mathsf{Set}}\to{\mathsf{Set}}$ for general reasons. More in detail, if $T$ is $\lambda$-accessible for a regular cardinal, then $T=\Lan{J_\lambda}{(T J_\lambda)}$, where $J_\lambda:{\mathsf{Set}}_\lambda\to{\mathsf{Set}}$ is the inclusion of the full subcategory ${\mathsf{Set}}_\lambda$ spanned by $\lambda$-small sets. Consequently, $$ T_\kat{V}=\Lan{D J_\lambda}{(D T J_\lambda)} $$ exhibits $T_\kat{V}$ as $\Lan{D}{(D T)}$ by~\cite[Theorem~4.47]{kelly:book}. In particular, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $(T_\Sigma)_\kat{V}$ exists for every polynomial functor $$ T_\Sigma X=\coprod_n {\mathsf{Set}}(n,X)\bullet\Sigma n $$ where $\Sigma:|{\mathsf{Set}}_\lambda|\to{\mathsf{Set}}$ is a $\lambda$-ary signature. We will give an explicit formula for the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $(T_\Sigma)_\kat{V}$ later. \end{enumerate} \end{exas} \subsection{The extension theorem} \ \bigskip\noindent Referring to the last item of the examples above, one of the aims of this paper is to show that $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications exist even if $T$ is not accessible. This is a consequence of a more general result, which we are going to prove as Theorem~\ref{thm:Lan}, namely that the left Kan-extension $$ \xymatrix@R=20pt{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \ar[0,2]^-{\Lan{D}{H}} & & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \\ {\mathsf{Set}} \ar[-1,2]_-{H} \ar[-1,0]^-{D} & & } $$ exists for any $H:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \medskip\noindent We will give an explicit construction of these left Kan-extensions, the idea of which is as follows. \begin{itemize} \item Every $\kat{V}$-category $\kat{X}$ can be represented as a certain small colimit of discrete $\kat{V}$-categories. \item Since $H$ is defined on discrete $\kat{V}$-categories, $H$ can be extended to $\kat{X}$ by applying $H$ to the corresponding discrete categories and then computing the colimit. \item Since the colimit is small, the extension of $H$ always exists, even though ${\mathsf{Set}}$ is not a small subcategory of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \end{itemize} \medskip\noindent Recall from Example~\ref{exle:coinserter} the definition of a coinserter, and from Example~\ref{exle:preord-coins-discrete} how any ${\mathbb{2}}$-category (preorder) appears as the coinserter \eqref{eq:poset_coins} of discrete ${\mathbb{2}}$-categories. Before adapting the notion of coinserter to enrichment over $\kat{V}$, we first explain why the obvious modification of the $\Two\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-situation leads to a notion that is too strong for our purposes. \begin{rem} Given that we want to generalise the construction \eqref{eq:poset_coins} of a preorder as a coinserter \begin{equation}\label{eq:canonical-coinserter} \xymatrix@C=35pt{ DV(X^{\mathbb{2}}) \ar@<0.5ex>[r]^{d_0} \ar@<-0.5ex>[r]_{d_1} & DVX \ar[r]^-q & X} \end{equation} it is tempting to ask what it could mean for \begin{equation}\label{eq:XtotheV} \xymatrix@C=45pt{ DV(\kat{X}^\kat{V}) \ar[r]^{(d^r)_{r\in\kat{V}_o} } & DV\kat{X} \ar[r]^-q & \kat{X}} \end{equation} to be a colimit in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, where we denoted $\kat{X}^\kat{V} = \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{V},\kat{X})$. In detail, let ${\mathbb{A}}$ be the free $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category built on the ordinary category with two objects $\{1,0\}$ and with arrows \[ \xymatrix{1 \ar[r]^{\delta^r} & 0} \] for all $r\in\kat{V}_o$. In analogy with Example~\ref{exle:coinserter}, we consider the weight $W:{\mathbb{A}}^{\mathit{op}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ mapping $0$ to ${\mathbb{1}}$ and $1$ to $\kat{V}$.% \footnote{ Recall from Section~\ref{sec:V-cat-cat} that ${\mathbb{1}}$ is the unit $\kat{V}$-category, with only one object $0$ and with ${\mathbb{1}}(0,0)=e$.} On arrows, $W\delta^r$ sends the unique object of ${\mathbb{1}}$ to $r\in \kat{V}_o$ \[ \xymatrix{{\mathbb{1}} \ar[r]^{0\mapsto r} & \kat{V}} \] Furthermore, let $F:{\mathbb{A}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ be the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor which maps the arrow $\delta^r:1\to 0$ to the ``evaluation at $r$'' $\kat{V}$-functor \[ \xymatrix{DV(\kat{X}^\kat{V}) \ar[r] & DV\kat{X}} \] that is, $F\delta^r(f)=f(r)$. Now spelling out Definition~\ref{def:colimit} of what it would mean for $\kat{X}$ to be the colimit $W * F$, we see that for each $\kat{V}$-category $\kat{Y}$, % $$\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X},\kat{Y})\cong [{\mathbb{A}}^{\mathit{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W,\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(F-,\kat{Y}))$$ an object of the $\kat{V}$-category on the right hand side amounts to a map on objects $g:\kat{X}\to \kat{Y}$ satisfying \begin{equation}\label{eq:rsle} [s,r]\le \kat{Y}(g(f(s)),g(f(r))) \end{equation} for all $\kat{V}$-functors $f:\kat{V} \to \kat{X}$. For $\kat{X}$ to be a colimit, one needs the implication \[ r\le \kat{X}(x',x) \ \Rightarrow r\le \kat{Y}(g(x'),g(x)) \] to hold for all $r\in\kat{V}$, which follows from \eqref{eq:rsle}, if one can find $f:\kat{V}\to\kat{X}$ such that $f(e)=x'$ and $f(r)=x$. But note that, in the case of $\kat{V}=[0,\infty]$ and $\kat{X}=\{x',x\}$ with $0<_\mathbb R\kat{X}(x',x)<_\mathbb R\infty$ no such $f$ exists. Indeed, it would be interesting to define $\kat{X}$ to be ``path-connected'', or $\kat{V}$-connected, if for all $r\le \kat{X}(x',x)$ there is a $\kat{V}$-functor $f:\kat{V}\to\kat{X}$ such that $f(e)=x', f(r)=x$. \qed \end{rem} \noindent Intuitively, what goes wrong with trying to set up \eqref{eq:XtotheV} as a colimit is that $\kat{V}$-functors $f:\kat{V}\to \kat{X}$ can be perceived as ``paths'' and that a general $\kat{X}$ need not be ``path-connected''. While this may be interesting to pursue in the future, the solution of our problem is to replace \eqref{eq:XtotheV} by \begin{equation}\label{eq:Vnerve} \xymatrix@C=45pt{ DX_r \ar@<0.5ex>[r]^{\partial^r_0} \ar@<-0.5ex>[r]_{\partial^r_1} & DV\kat{X} \ar[r]^-q & \kat{X}} \end{equation} where the sets $X_r$ were introduced in \eqref{eq:Xr} as $X_r = \{(x',x) \mid r\le \kat{X}(x',x)\}$ and can equally be described as $V\kat{X}^{{\mathbb{2}}_r}$ (see~\eqref{eq:Two_r} and Example~\ref{ex:non-Vcatif}) in total accordance now with~\eqref{eq:canonical-coinserter}. \medskip \noindent Below we give the full details and prove that every $\kat{X}$ is a colimit as in \eqref{eq:Vnerve}, from which the existence of left-Kan extensions along $D$ then follows. \medskip \noindent We use the letter ``N'' in various fonts and notations to indicate the analogy with the nerve of an ordinary category. \begin{defi}\label{def:Vcoinserter} A $\kat{V}$-coinserter is a colimit that has the weight $$N:\mathbb{N}^{\mathit{op}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}},$$ where ${\mathbb{N}}$ is the {\em free} $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category built upon the following ordinary category $\mathsf N$: the objects are all $r$ in $\kat{V}_o$, together with an extra symbol ${{\sf o}}$, with arrows \[ \xymatrix{r \ar@<0.45ex>[r]^{\delta_1^r} \ar@<-0.45ex>[r]_{\delta_0^r} & {{\sf o}}} \] for $r$ ranging over the elements of $\kat{V}_o$. The weight $N$ is the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor sending ${{\sf o}}$ to ${\mathbb{1}}$, and $r$ to ${\mathbb{2}}_r$. The action of $N$ on arrows is defined as follows: $N\delta^r_0:{\mathbb{1}}\to{\mathbb{2}}_r$ sends $0$ to $0$, while $N\delta^r_1:{\mathbb{1}}\to{\mathbb{2}}_r$ sends $0$ to $1$: \[ \xymatrix{\mathbb 2_r &\ar@<-0.45ex>[l]_{N\delta^r_1} \ar@<0.45ex>[l]^{N\delta^r_0} \mathbb 1} \] \end{defi} \begin{rem} \label{rem:V-coins-exist} As $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is complete and cocomplete as a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category itself~\cite[Sections 3.2,~3.10]{kelly:book}, $\kat{V}$-coinserters do exist in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ and can be computed using copowers and conical colimits, the latter of which, on the level of objects, are computed as in ${\mathsf{Set}}_o$ because of the topologicity of the forgetful functor $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o\to {\mathsf{Set}}_o$~\cite{hst:book}. But this general recipe is not so easy to use in practice, and we will provide in Corollary~\ref{cor:Lan by zigzag paths} and Lemma~\ref{lem:wpb} alternatives for computing the $N$-weighted colimits of interest to us. \qed \end{rem} \noindent We are now going to show that every $\kat{V}$-category $\kat{X}$ is a $\kat{V}$-coinserter of discrete $\kat{V}$-categories which we would like to think of as components of ``its $\kat{V}$-nerve''. \begin{prop} \label{prop:Vcoinserter} Let $\kat{X}$ be a $\kat{V}$-category and let $F\!_\kat{X}:{\mathbb{N}}\to{\mathsf{Set}}$ be given by \begin{align*} \mathsf N & \to{\mathsf{Set}}_o \\ {{\sf o}} & \ \mapsto \ X_{{\sf o}} \mbox{ , the set of objects of $\kat{X}$} \\ r & \ \mapsto \ X_r =\{(x',x)\in X_{{\sf o}} \times X_{{\sf o}} \mid r\leq \kat{X}(x',x)\} \end{align*} with $F\!_\kat{X}\delta^r_0$ and $F\!_\kat{X}\delta^r_1$ the evident projections. Then the colimit $$\Colim{N}{(D F_\kat{X})}$$ of the diagram \begin{equation*}\label{eq:DFX} (N:\mathbb{N}^{\mathit{op}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}},DF_\kat{X}:{\mathbb{N}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}) \end{equation*} in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is isomorphic to $\kat{X}$. \end{prop} \begin{proof} The colimit $\Colim{N}{(DF\!_\kat{X})}$ exists in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, since the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category ${\mathbb{N}}$ is small. \medskip\noindent To ease the notation, we put \begin{gather*} \kat{X}_{{\sf o}}=D F_\kat{X}({{\sf o}} ) \quad\quad \quad\quad \kat{X}_r=DF_\kat{X} (r)\\ \ \ \partial^r_0=DF_\kat{X}\delta^r_0 \quad\quad \quad\quad \ \partial^r_1=D F_\kat{X}\delta^r_1 \end{gather*} \noindent Let us analyse the defining isomorphism $$ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\Colim{N}{(DF_\kat{X})},\kat{Y}) \cong [{\mathbb{N}}^{\mathit{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](N,\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DF_\kat{X}{-},\kat{Y})) $$ of $\kat{V}$-categories, natural in $\kat{Y}$. The $\kat{V}$-category $$[{\mathbb{N}}^{\mathit{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](N,\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DF_\kat{X}{-},\kat{Y}))$$ of $N$-weighted ``cocones'' for $DF_\kat{X}$ is described explicitly as follows. \begin{enumerate} \item The objects are $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformations $$\tau:N\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DF_\kat{X}{-},\kat{Y}).$$ Each such $\tau$ consists of the following $\kat{V}$-functors: \begin{itemize} \item $\tau_{{\sf o}}: N{{\sf o}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X}_{{\sf o}},\kat{Y})$. Since $N{{\sf o}}={\mathbb{1}}$, $\tau_{{\sf o}}$ picks up a $\kat{V}$-functor $$f_{{\sf o}}:\kat{X}_{{\sf o}}\to \kat{Y}.$$ No other restrictions are imposed since ${\mathbb{1}}(0,0)=e$. \item $\tau_r:Nr\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X}_r,\kat{Y})$. This $\kat{V}$-functor picks up two $\kat{V}$-functors $$f^r_0:\kat{X}_r\to \kat{Y}$$ and $$f^r_1:\kat{X}_r\to \kat{Y}.$$ Since $\kat{X}_r$ is discrete, both $f^r_0$ and $f^r_1$ are defined by their object-assignments only. There is, however, the constraint below, because $Nr={\mathbb{2}}_r$: $$ r\leq\bigwedge_{r \leq \kat{X}(x',x)} \kat{Y}(f^r_0 (x',x),f^r_1 (x',x)) $$ \end{itemize} In addition to the above, there are various commutativity conditions since $\tau$ is natural. Explicitly, for $\delta^r_0:r\to {{\sf o}}$, we have the commutative square $$ \xymatrix@R=17pt{ N{{\sf o}} ={\mathbb{1}} \ar[0,2]^-{\tau_{{\sf o}}} \ar[1,0]_{N\delta^r_0} & & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X}_{{\sf o}},\kat{Y}) \ar[1,0]^{\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\partial^r_0,\kat{Y})} \\ Nr={\mathbb{2}}_r \ar[0,2]_-{\tau_r} & & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X}_r,\kat{Y}) } $$ that, on the level of objects, is the requirement $$f_{{\sf o}} \cdot \partial^r_0 = f^r_0 $$ Similarly, for $\delta^r_1:r \to {{\sf o}}$, we obtain $$f_{{\sf o}} \cdot \partial^r_1 = f^r_1$$ We conclude that to give $\tau:N\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DF_\kat{X}{-},\kat{Y})$ reduces to a $\kat{V}$-functor $$f_{{\sf o}}:\kat{X}_{{\sf o}}\to\kat{Y}$$ (and, recall, this $\kat{V}$-functor is given just by the object-assignment $x\mapsto f_{{\sf o}} x$, since $\kat{X}_{{\sf o}}$ is discrete) such that $r\leq \kat{Y}(f_{{\sf o}} x',f_{{\sf o}} x)$ holds for every object $(x',x)$ in $\kat{X}_r$ and every $r$, which means precisely that $$\kat{X}(x',x)\leq \kat{Y}(f_{{\sf o}} x',f_{{\sf o}} x)$$ holds. \medskip \item Given $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformations $\tau$ and $\tau'$, then the $\kat{V}$-distance between them is given by $$ [{\mathbb{N}}^{\mathit{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](N,\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DF_\kat{X}{-},\kat{Y}))(\tau,\tau') = \bigwedge_{x} \kat{Y}(f_{{\sf o}} x,f'_{{\sf o}} x) $$ where $f_{{\sf o}}$ corresponds to $\tau$ and $f'_{{\sf o}}$ corresponds to $\tau'$. \end{enumerate} \medskip\noindent From the above description of the $\kat{V}$-category of $N$-weighted ``cocones'' for $DF_\kat{X}$, it follows that the $\kat{V}$-functor \begin{equation}\label{eq:couniversal cocone} q_\kat{X}:\kat{X}_{{\sf o}}\to \kat{X} \end{equation} that sends each object $x$ to itself is the couniversal such ``cocone''. More precisely, the inequality $r\leq \kat{X}(q_\kat{X} x',q_\kat{X} x)$ holds for every $(x',x)$ in $\kat{X}_r$ and every $r$. \medskip\noindent Furthermore, given any $\kat{V}$-functor $f_{{\sf o}}:\kat{X}_{{\sf o}}\to \kat{Y}$ with the above properties, there is a unique $\kat{V}$-functor $f_{{\sf o}}^\sharp:\kat{X}\to \kat{Y}$ such that $f_{{\sf o}}^\sharp q_\kat{X}=f_{{\sf o}}$ holds. \medskip\noindent The ``2-dimensional aspect'' of the colimit, that is, $$ \bigwedge_x \kat{Y}(f^\sharp_{{\sf o}} x,f'^\sharp_{{\sf o}} x) = \bigwedge_x \kat{Y}(f_{{\sf o}} x,f'_{{\sf o}} x). $$ is satisfied because $f^\sharp_{{\sf o}}$ and $f_{{\sf o}}$ coincide on objects. Hence we have proved that $\kat{X}$ is isomorphic to $\Colim{N}{(DF_\kat{X})}$. \end{proof} \begin{rem} \hfill \begin{enumerate} \item Assume that the quantale $\kat{V}$ is integral. Then following the same steps as in the proof above, one can see that the colimit $N\ast F_{DX}$ exists in ${\mathsf{Set}}$ and equals $X$. Intuitively, what is happening is that if $e=\top$, the diagram is non-empty everywhere and $D$ is fully faithful on non-empty domains, so it reflects colimits in general and the colimit $DX=N\ast DF_{DX}$ in particular. \item The ordinary category $\mathsf N$ carries information about the elements of $\kat{V}_o$, but not about its order. This may seem peculiar at first sight, but the order is actually captured within the observation that the identity-on-objects ${\mathsf{Id}}_{r,s}:{\mathbb{2}}_r =N(r) \to {\mathbb{2}}_s = N(s)$ is a $\kat{V}$-functor iff $r\leq s$, and it is the only $\kat{V}$-functor such that ${\mathsf{Id}}_{r,s} \circ N\delta^r_i = N\delta^s_i$, $i=0,1$. Hence nothing is gained if additionally $\mathbb N$ is enhanced with arrows witnessing the order relation in $\kat{V}_o$, compatible with the existing arrows $r\rightrightarrows {{\sf o}}$ and with (the transitivity of) the order relation in $\kat{V}_o$. \end{enumerate} \end{rem} \medskip\noindent The lemma above allows us to compute left Kan extensions along $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$: \begin{thm} \label{thm:Lan} Every functor $H:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ has a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched left Kan extension $H^\sharp:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ along $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ given by $H^\sharp\kat{X} = \Colim{N}{(HF\!_\kat{X})}$. \end{thm} \begin{proof} Based on Proposition~\ref{prop:Vcoinserter}, we will prove the following. If we define $H^\sharp \kat{X}$ as the colimit $\Colim{N}{(H F\!_\kat{X})}$, then the assignment $\kat{X}\mapsto H^\sharp\kat{X}$ can be extended to a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor that is a left Kan extension of $H$ along $D$. \medskip\noindent Suppose $H:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is given. We will use the notation employed in the proof of Proposition~\ref{prop:Vcoinserter}. \medskip \begin{enumerate} \item \label{item 2.a} We first define a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H^\sharp:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \medskip\noindent For every small $\kat{V}$-category $\kat{X}$, let $F_\kat{X}$ be as in Proposition~\ref{prop:Vcoinserter}. Recall that for every $r\in \kat{V}_o$, we put $F_\kat{X}(r)=X_r$, the {\em set} of pairs $(x',x)$ such that $r\leq\kat{X}(x',x)$, and $F_\kat{X}({{\sf o}})=X_{{\sf o}}$, the {\em set} of objects of $\kat{X}$. Analogously, for a $\kat{V}$-functor $f:\kat{X}\to\kat{Y}$, we now denote by $f_r:X_r\to Y_r$ and $f_{{\sf o}}:X_{{\sf o}}\to Y_{{\sf o}}$ the {\em maps} corresponding to $(x',x)\mapsto (fx',fx)$ and the object assignment of $f$, respectively. Let also put $d^r_0=F_\kat{X}\delta^r_0$ and $d^r_1=F_\kat{X} \delta^r_1$. \medskip\noindent We define \begin{equation}\label{eq:H(Vnerve)} H^\sharp\kat{X}=\Colim{N}{(H F_\kat{X})}, \end{equation} where $N$ is as in Definition~\ref{def:Vcoinserter}. \medskip\noindent Unravelling the definition of the weighted colimit, the $1$-dimensional aspect says that to give a $\kat{V}$-functor $f^\sharp : H^\sharp \kat{X} \to \kat{Y}$ is the same as to give a $\kat{V}$-functor $f:HX_{{\sf o}} \to \kat{Y}$ such that \begin{equation} \label{eq:couniversal-property} r \leq \bigwedge_{C\in HX_r} \kat{Y} (f Hd^r_0 (C), f Hd^r_1(C)) \end{equation} holds for all $r$.% \footnote{ By slight abuse of language, we will use here and subsequently notation like $C\in HX_r$ to mean that $C$ runs through all objects in \emph{the $\kat{V}$-category} $HX_r$. } In particular, there is a ``quotient'' $\kat{V}$-functor $c_\kat{X} :HX_{{\sf o}} \to H^\sharp \kat{X}$ such that \begin{equation} \label{eq:Hsharp-quotient} r \leq \bigwedge_{C \in HX_r} H^\sharp \kat{X} (c_\kat{X} Hd^r_0 (C), c_\kat{X} Hd^r_1(C))\end{equation} holds for all $r$, with the property that any $\kat{V}$-functor $HX_{{\sf o}} \to \kat{Y}$ satisfying~\eqref{eq:couniversal-property} uniquely factorizes through $c_\kat{X}$. \medskip\noindent The 2-dimensional aspect of the colimit says that given any $f,g:HX_{{\sf o}} \to \kat{Y}$ satisfying~\eqref{eq:couniversal-property}, the relation \begin{equation} \label{eq:2-dim-aspect} \bigwedge_{B \in HX_{{\sf o}}} \kat{Y}( f (B),g (B)) = \bigwedge_{A \in H^\sharp \kat{X}} \kat{Y}( f^\sharp (A),g^\sharp (A)) \end{equation} holds. \medskip\noindent For a $\kat{V}$-functor $f:\kat{X}\to\kat{Y}$ we recall that the diagram $$ \xymatrix{ X_r \ar@<.5ex> [0,1]^-{d_1^r} \ar@<-.5ex> [0,1]_-{d_0^r} \ar[1,0]_{f_r} & X_{{\sf o}} \ar[1,0]^-{f_{{\sf o}}} \\ Y_r \ar@<.5ex> [0,1]^-{d_1^r} \ar@<-.5ex> [0,1]_-{d_0^r} & Y_{{\sf o}} } $$ commutes serially. Hence $f$ induces a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $F_f:F_\kat{X}\to F_\kat{Y}$. Therefore we can define $H^\sharp f:H^\sharp\kat{X}\to H^\sharp\kat{Y}$ as the unique mediating $\kat{V}$-functor $$ \Colim{N}{(H F_f)} : \Colim{N}{(H F_\kat{X})} \to \Colim{N}{(H F_\kat{Y})} $$ In particular, we have the commutative diagram below: \begin{equation*} \vcenter{\xymatrix{ HX_{{\sf o}} \ar[r]^{c_\kat{X}} \ar[d]_{Hf_{{\sf o}}} & H^\sharp \kat{X} \ar[d]^{H^\sharp f} \\ HY_{{\sf o}} \ar[r]^{c_\kat{Y}} & H^\sharp \kat{Y} }} \end{equation*} Also, from the 2-dimensional aspect of the colimit (see Eq.~\eqref{eq:2-dim-aspect}), we have that for any $f,g:\kat{X} \to \kat{Y}$, the equality below holds: \begin{equation} \label{eq:Hsharp-2-dim-aspect} \bigwedge_{B \in HX_{{\sf o}}} H^\sharp \kat{Y}( c_\kat{Y} Hf_{{\sf o}}(B), c_\kat{Y} Hg_{{\sf o}} (B)) = \bigwedge_{A \in H^\sharp \kat{X}} H^\sharp \kat{Y}( H^\sharp f (A), H^\sharp g (A)) \end{equation} It remains to prove that the inequality $$ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X},\kat{Y})(f,g) \leq \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(H^\sharp \kat{X},H^\sharp \kat{Y})(H^\sharp f,H^\sharp g) $$ is satisfied. To that end, suppose that $r\leq \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X},\kat{Y})(f,g)$ holds. This is equivalent to the fact that there is a mapping $t:X_{{\sf o}}\to Y_r$ such that the triangles \begin{equation} \label{eq:internal_nat1} \vcenter{ \xymatrix{ X_{{\sf o}} \ar[0,1]^-{t} \ar[1,1]_{f_{{\sf o}}} & Y_r \ar[1,0]^{d_0^r} & & X_{{\sf o}} \ar[0,1]^-{t} \ar[1,1]_{g_{{\sf o}}} & Y_r \ar[1,0]^{d_1^r} \\ & Y_{{\sf o}} & & & Y_{{\sf o}} } } \end{equation} commute. In fact, $t(x)=(f(x),g(x))$. To prove that $ r \leq \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(H^\sharp \kat{X},H^\sharp \kat{Y})(H^\sharp f,H^\sharp g) $ holds, we need to prove the inequality $$ r \leq \bigwedge_{A\in H^\sharp \kat{X}} H^\sharp \kat{Y} ( H^\sharp f(A), H^\sharp g(A)) $$ This follows from: \allowdisplaybreaks \begin{align*} r \leq & \bigwedge_{C \in HY_r} H^\sharp \kat{Y} (c_\kat{Y} Hd^r_0(C), c_\kat{Y} Hd^r_1 (C)) & \mbox{by~\eqref{eq:Hsharp-quotient}} \\ \leq & \bigwedge_{B \in HX_{{\sf o}}} H^\sharp \kat{Y} (c_\kat{Y} Hd^r_0 Ht(B), c_\kat{Y} Hd^r_1 Ht (B)) \\ = & \bigwedge_{B \in HX_{{\sf o}}} H^\sharp \kat{Y} (c_\kat{Y} Hf_{{\sf o}} (B), c_\kat{Y} Hg_{{\sf o}} (B)) & \mbox{by~\eqref{eq:internal_nat1}} \\ = & \bigwedge_{A \in H^\sharp \kat{X}} H^\sharp \kat{Y} (H^\sharp f (A), H^\sharp g (A)) & \mbox{by~\eqref{eq:Hsharp-2-dim-aspect}} \end{align*} Preservation of composition and identity follows easily from the colimit property, hence we obtain that the correspondence $\kat{X}\mapsto H^\sharp\kat{X}$ can be extended to a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H^\sharp:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \medskip \item We show now that $H^\sharp\cong\Lan{D}{H}$ holds. As explained in~\cite[Section~4.3]{kelly:book} (see also Remark~\ref{rem:lan}\eqref{rem:weak_def_lan}), it is enough to check that for any $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $K:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, there is a one-to-one correspondence between $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformations $H^\sharp\to K$ and $H\to KD$. \medskip\noindent Due to the definition of $H^\sharp$, there is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $\alpha:H\to H^\sharp D$ having as components the $\kat{V}$-functors $\alpha_X = c_{DX}$. We will prove that $\alpha$ is the unit of a left Kan extension. \medskip\noindent Let $\tau:H^\sharp\to K$ be a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation. Then the composite $$ \xymatrix{ H \ar[0,1]^-{\alpha} & H^\sharp D \ar[0,1]^-{\tau D} & K D } $$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $\tau^\flat:H\to K D$. \medskip\noindent Conversely, let $\sigma: H\to K D$ be a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation. To give a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $\sigma^\sharp:H^\sharp \to K$ means to give a family of $\kat{V}$-functors $$\sigma^\sharp:H^\sharp \kat{X} \to \kat{K}\kat{X}$$ satisfying the naturality condition~\eqref{eq:Vcat-nat}. In turn, each $\kat{V}$-functor $\sigma^\sharp:H^\sharp \kat{X} \to \kat{K}\kat{X}$ is uniquely determined by a $\kat{V}$-functor $HX_o\to K\kat{X}$ such that~\eqref{eq:couniversal-property} holds. \noindent For this, we choose the composite $\kat{V}$-functor $$HX_o\overset{\sigma_{X_o}}{\to} KDX_o\overset{Kq_{\kat{X}}}{\to} K\kat{X}$$ where $q_\kat{X}:DX_o \to \kat{X}$ is the canonical cocone~\eqref{eq:couniversal cocone} associated to each $\kat{V}$-category $\kat{X}$. \noindent We verify~\eqref{eq:couniversal-property} for each $r\in \kat{V}_o$: \begin{align*} && r && \leq &&& \ \ \bigwedge_{(x',x)\in X_r} \kat{X}(x',x) \\ && && = &&& \bigwedge_{(x',x)\in \mathsf{ob} DX_r} \kat{X}(q_{\kat{X}} d^r_0(x',x), q_{\kat{X}} d^r_1(x',x)) \\ && && = &&& [DX_r, \kat{X}](q_{\kat{X}} d^r_0, q_{\kat{X}} d^r_1) \\ && && \leq &&& [KDX_r, K\kat{X}](K(q_\kat{X} d^r_0), K(q_\kat{X} d^r_1)) \\ && && = &&& \bigwedge_{B\in \mathsf{ob} KDX_r} K \kat{X} (Kq_{\kat{X}} Kd^r_0(B), Kq_{\kat{X}} Kd^r_1(B)) \\ && && \leq &&& \bigwedge_{C\in \mathsf{ob}HX_r} K\kat{X}(Kq_{\kat{X}} Kd^r_0 \sigma_{X_r}(C),Kq_{\kat{X}} Kd^r_1 \sigma_{X_r}(C)) \\ && && = &&& \bigwedge_{C\in \mathsf{ob}HX_r} K\kat{X}(Kq_{\kat{X}} \sigma_{X_o} Hd^r_0(C),Kq_{\kat{X}} \sigma_{X_o} Hd^r_1(C)) \end{align*} In the above, we have used that $K$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor and the naturality of $\sigma$. \medskip\noindent Hence there is a unique $\kat{V}$-functor $\sigma^\sharp_\kat{X} :H^\sharp \kat{X} \to K\kat{X}$ such that $\sigma^\sharp_\kat{X} c_\kat{X} = Kq_\kat{X} \sigma_{X_o}$ holds. Recall that $c_\kat{X}:HX_{{\sf o}} \to H^\sharp \kat{X}$ is the colimiting cocone. \medskip\noindent We leave to the reader the verifications that the $\kat{V}$-functors $\sigma_\kat{X}$ are the components of a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $\sigma^\sharp :H^\sharp \to K$, and that the correspondences $\tau\mapsto\tau^\flat$ and $\sigma\mapsto\sigma^\sharp$ are inverses to each other. \qedhere \end{enumerate} \end{proof} \begin{cor}\label{cor:dense} $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is dense. \end{cor} \begin{proof} It follows from Proposition~\ref{prop:Vcoinserter} and Theorem~\ref{thm:Lan} that the identity $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is the left Kan extension of $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ along $D$. But this is one of the equivalent definitions of density in~\cite[Thm.5.1]{kelly:book} \end{proof} \medskip\noindent To be able to compute these left Kan extensions in concrete examples, we show that from the proof above one can extract a more explicit construction in terms of ``shortest paths''. \footnote{ Left Kan extensions can be computed pointwise as colimits, hence using copowers and conical colimits. Due to the topologicity of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_0$ over ${\mathsf{Set}}_0$, the latter are obtained (on object-level) as in ${\mathsf{Set}}_0$ and endowed with the corresponding $\kat{V}$-metric structure. This is why the subsequent formula~\eqref{eq:shortest-path} is reminiscent of the usual construction of quotient metric spaces~\cite{smyth}. } \begin{cor} \label{cor:Lan by zigzag paths} The left Kan extension $H^\sharp = \Lan{D}{H}$ of a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is obtained as follows: \begin{enumerate} \item For a $\kat{V}$-category $\kat{X}$, $H^\sharp \kat{X}$ is the $\kat{V}$-category having the same objects as $HX_{{\sf o}}$ (that is, the underlying set of objects of the $\kat{V}$-category obtained by applying $H$ to $X_{{\sf o}}$, the set of objects of $\kat{X}$). For any objects $A'$ and $A$ of $H^\sharp\kat{X}$, their corresponding $\kat{V}$-hom $H^\sharp \kat{X} (A', A)$ is given by \begin{equation}\label{eq:shortest-path} \begin{gathered} H^\sharp \kat{X} (A', A) = \bigvee \{ {HX_{{\sf o}}(A'_0,A_0)} {\otimes} r_1 {\otimes} HX_{{\sf o}} (A'_1,A_1) {\otimes} {\ldots} {\otimes} r_n {\otimes} HX_{{\sf o}}(A_n', A_n) \} \end{gathered} \end{equation} where the suprema is computed over all paths $$ (A'_0,A_0), (A'_1, A_1), \ldots , (A'_n, A_n) $$ where $A'=A'_0$, $A=A_n$, and all (possibly empty) tuples of elements $$ (r_1, \ldots, r_n) $$ such that there are objects $C_i$ in $HX_{r_i}$ with $Hd^{r_i}_0(C_i) = A_{i-1}$, $Hd^{r_i}_1(C_i) = A'_i$, for all $i=1,n$:% $$ \xymatrix@C=-12pt@R=30pt{ & C_1 \ar[dl]|{Hd^{r_1}_0} \ar[dr]|{Hd^{r_1}_1} & & C_2 \ar[dl]|{Hd^{r_2}_0} \ar[dr]|{Hd^{r_2}_1} & & \quad \quad \ar@{}[d]|{.\ . \ .} & & \quad C_n \, \, \ar[dr]|{Hd^{r_n}_1} & & \\ *+[l]{A' =A_0' , A_0} & & {\quad A'_1 , A_1 \quad} & & {\quad A'_2 , A_2 \quad} & & {\quad A'_{n-1} , A_{n-1}\quad} \ar@{<-}[]!<5ex,1ex>;[ur]!<0ex,-1ex>|{Hd^{r_n}_0} & & *+[r]{A_n' , A_n=A} } $$ \item The couniversal cocone $c_\kat{X} : HX_{{\sf o}} \to H^\sharp \kat{X}$ is the identity on objects. \item For a $\kat{V}$-functor $f:\kat{X}\to\kat{Y}$, $H^\sharp f$ acts as $Hf_{{\sf o}}$ on objects. \end{enumerate} \ \end{cor} \begin{proof} First, notice that the construct $H^\sharp\kat{X}$ described above satisfies \begin{equation}\label{eq:quotient V-functor} HX_{{\sf o}} (A',A) \leq H^\sharp \kat{X} (A',A) \end{equation} for all objects $A', A$. In particular, $e\leq H^\sharp \kat{X} (A,A)$ holds. Next, the inequality $H^\sharp \kat{X}(A', A) \otimes H^\sharp \kat{X} (A'', A') \leq H^\sharp \kat{X}(A'', A)$ can be established by path concatenation, using that $HX_{{\sf o}}$ is a $\kat{V}$-category. Hence $H^\sharp \kat{X}$ is a $\kat{V}$-category. \medskip\noindent We will verify that $H^\sharp \kat{X}$ as above is indeed $\Colim{N}{(H F\!_\kat{X})}$. Let $c_\kat{X}:HX_{{\sf o}} \to H^\sharp \kat{X}$ be the identity on objects. It is a $\kat{V}$-functor by~\eqref{eq:quotient V-functor}. \medskip\noindent We show now that~\eqref{eq:Hsharp-quotient} holds. Let $r$ be an arbitrary element of the quantale and fix an object $C$ in $HX_r$. Then by employing the path $$ \xymatrix@C=-2pt@R=30pt{ & & C \ar[dl]|{Hd^{r}_0} \ar[dr]|{Hd^{r}_1} & \\ Hd^r_0(C) \ar@{}[r]_{,} & Hd^r_0(C) & & Hd^r_1(C) \ar@{}[r]_{,} & Hd^r_1(C) } $$ one can see that \begin{align*} r = & \ e \otimes r\otimes e \\ \leq & \ HX_{{\sf o}} (Hd^r_0(C), Hd^r_0(C)) \otimes r\otimes HX_{{\sf o}} (Hd^r_1(C), Hd^r_1(C)) \\ \leq & \ H^\sharp \kat{X} (Hd^r_0(C), Hd^r_1(C)) \end{align*} \medskip\noindent We check the 1-dimensional aspect of the colimit: Let $f:HX_{{\sf o}} \to \kat{Y}$ be a $\kat{V}$-functor such that~\eqref{eq:couniversal-property} holds. Then the object-assignment $A \mapsto f(A)$ (uniquely!) extends to a $\kat{V}$-functor $f^\sharp : H^\sharp \kat{X} \to \kat{Y}$ such that $f^\sharp c_\kat{X} = f$. This is because for each path $(A'=A'_0,A_0), (A'_1, A_1), \ldots , (A'_n, A_n=A)$ and each tuple of elements $(r_1, \ldots, r_n)$ we have that \begin{align*} &HX_{{\sf o}}(A_0',A_0) \otimes r_1 \otimes HX_{{\sf o}} (A'_1,A_1) \otimes \ldots \otimes r_n \otimes HX_{{\sf o}}(A'_n, A_n) \\ & = \\ & HX_{{\sf o}}(A_0',Hd^{r_1}_0(C_1)) \otimes r_1 \otimes HX_{{\sf o}} (Hd^{r_1}_1(C_1),Hd^{r_2}_0(C_2)) \otimes \ldots \otimes r_n \otimes HX_{{\sf o}}(Hd^{r_n}_0(C_n), A_n) \\ & \leq \\ & \kat{Y}(f(A_0'),fHd^{r_1}_0(C_1)) \otimes \kat{Y}(fHd^{r_1}_0(C_1), fHd^{r_1}_1(C_1)) \otimes \kat{Y}( f Hd^{r_1}_1(C_1) , f Hd^{r_2}_0(C_2)) \otimes \ldots \\ & \otimes \kat{Y}(fHd^{r_n}_0(C_n), fHd^{r_n}_1(C_n)) \otimes \kat{Y} ( f Hd^{r_n}_0(C_n), f(A_n) ) \\ & \leq \\ & \kat{Y}(f(A'), f(A)) \end{align*} where we used that $f$ is a $\kat{V}$-functor and~\eqref{eq:couniversal-property}. This proves that $$H^\sharp \kat{X}(A',A) \leq \kat{Y}(f^\sharp (A'), f^\sharp (A))$$ holds for all objects $A',A$. \medskip\noindent Finally, the 2-dimensional aspect of the colimit~\eqref{eq:2-dim-aspect} is trivial, because $c_\kat{X}$ is the identity on objects. \medskip\noindent Hence we have proved that $H^\sharp \kat{X} = \Colim{N}{(H F\!_\kat{X})}$. The action of $H^\sharp$ on $\kat{V}$-functors is as follows: For $f:\kat{X} \to \kat{Y}$, $H^\sharp f:H^\sharp \kat{X} \to H^\sharp \kat{Y}$ is the unique $\kat{V}$-functor satisfying \[ (H^\sharp f) c_\kat{X} = c_\kat{Y} (H f_{{\sf o}}) \] Given that both $c_\kat{X}$ and $c_\kat{Y}$ are identity on objects, we see that $H^\sharp f$ and $H f_{{\sf o}}$ coincide on objects. \end{proof} \begin{cor} \label{cor:Vcat-ifications_exist} Every $T:{\mathsf{Set}}\to{\mathsf{Set}}$ has a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $T_\kat{V}:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \end{cor} \begin{proof} Apply Theorem~\ref{thm:Lan} to the composite $H=D T:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \end{proof} \subsection{When the unit of the left-Kan extension is an isomorphism}\label{sec:unit-kan-iso} \ \bigskip\noindent Recall from Proposition~\ref{prop:D-ff} that the discrete functor $D:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is fully faithful (hence left Kan extensions along $D$ are genuine extensions) if and only if the quantale is integral. In particular, for $e<\top$, there is no guarantee that the unit $H\to (\mathsf{Lan}_D H)D$ of the left-Kan extension of $H$ along $D$ is an isomorphism. \medskip\noindent The reason why $D$ is not fully-faithful for non-integral $\kat{V}$ is that, for a discrete $\kat{V}$-category $\kat{X}=DX$, the relation $X_\top$ as defined in Proposition~\ref{prop:Vcoinserter} is empty in case $e<\top$, which in turn forces the self-distances in $(\mathsf{Lan}_{D}H)D\emptyset$ to be $\top$ whereas self-distances in $H\emptyset$ may be $<\top$. \medskip\noindent In this section we shall see necessary and sufficient conditions for $\Lan{D}H$ to coincide with $H$ on discrete $\kat{V}$-categories. \medskip\noindent Recall that ${\mathbb{1}}$ is the one-element $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ with self-distance $e$ and ${\mathbb{1}}_\top$ is the one-element $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ with self-distance $\top$. We use the same notation to denote the respectice constant functors ${\mathbb{1}}:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ and ${\mathbb{1}}_\top:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. The main technical observation is contained in the following \begin{exa}\label{exle:One} The left Kan extension of the functor ${\mathbb{1}}:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is the functor ${\mathbb{1}}_\top:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \end{exa} \begin{defi} A {\em constant} of a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $K:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation ${\mathbb{1}}_\top\to K$. A {\em constant} of a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H:{\mathsf{Set}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation ${\mathbb{1}}\to H$. \end{defi} \begin{rem} The constants of a functor $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ form a $\kat{V}$-category in which all self-distances are $\top$. --- Let $\tau:{\mathbb{1}}_\top\to K$ be a constant. Then each $\tau_\kat{X}: {\mathbb{1}}_\top \to K \kat{X}$ is a $\kat{V}$-functor, picking an object $\tau_\kat{X}(0)$ in $K \kat{X}$ with \[ {\mathbb{1}}_\top(0,0)=\top \leq K\kat{X}(\tau_\kat{X}(0),\tau_\kat{X}(0)) \] hence we obtain the self-distance \[ [{\mathbb{1}}_\top, K \kat{X}](\tau_\kat{X}, \tau_\kat{X}) = \top \] and consequently \[ [\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]({\mathbb{1}}_\top,K)(\tau,\tau)=\bigwedge_{\kat{X}}[{\mathbb{1}}_\top, K \kat{X}](\tau_\kat{X}, \tau_\kat{X}) = \top \] \end{rem} \begin{prop}\label{prop:constantsvcat} The $\kat{V}$-category of constants of a functor $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is isomorphic to its value on the empty $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \end{prop} \begin{proof} There is an isomorphism of $\kat{V}$-categories $$ [\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]({\mathbb{1}}_\top, K) \cong [{\mathsf{Set}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]({\mathbb{1}}, KD)\cong [{\mathsf{Set}}, \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]({\mathsf{Set}}(\emptyset,-),KD)\cong KD\emptyset $$ where the the first isomorphism is due to Example~\ref{exle:One} and the last isomorphism is the Yoneda lemma. \end{proof} \noindent The remark and the proposition combine to the result that, for any $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $K:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, the self-distances in $KD\emptyset$ must be $\top$. In particular, we have \begin{cor} If $e<\top$, then there is no constant functor $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ with value ${\mathbb{1}}$, or with value $DX$ for any $X\in{\mathsf{Set}}$. \end{cor} \noindent More generally, what is at stake here is that $e$ is not a retract of $\top$, see the discussion towards the end of~\cite[Ch.3.9]{kelly:book}. \medskip\noindent Essentially the same argument as for Proposition~\ref{prop:constantsvcat} also proves \begin{prop}\label{prop:constantsset} The $\kat{V}$-category of constants of a functor ${\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is isomorphic to its value on the empty set. \end{prop} \noindent To summarise, for any functor $K:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, the constants of $K$ are isomorphic to the constants of $KD$, and isomorphic to $KD\emptyset$. In particular, self-distances on elements of $KD\emptyset$ are $\top$. \medskip\noindent Therefore, in order to have $H \cong (\Lan{D}H)D$, self-distances in $H\emptyset$ must be $\top$. We are now going to show that this necessary condition is also sufficient. \medskip\noindent An analysis of why, in Example~\ref{exle:One}, the left Kan extension of ${\mathbb{1}}$ along $D$ is ${\mathbb{1}}_\top$ but not ${\mathbb{1}}$, suggests that the reason for the failure of the unit being an isomorphism resides in the fact that ${\mathbb{1}}$ is not a cocone over the $\kat{V}$-nerve $(N, {\mathbb{1}})$ of ${\mathbb{1}}$. This leads to \begin{defi} The functor $H:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ {\em preserves $\kat{V}$-nerves of sets} if the identity $\kat{V}$-functor on $HX$ induces a cocone of $HX$ over the diagram $(N, HF_{DX})$ for each $X$. \end{defi} \begin{rem}\label{rem:pres-discr-nerves} \hfill \begin{enumerate} \item $H$ preserves $\kat{V}$-nerves of sets iff ${\mathsf{Id}}_{HX}$ satisfies~\eqref{eq:couniversal-property}, that is, iff \[ r\leq \bigwedge_{C \in H(DX)_r} HX(Hd^r_0(C), Hd^r_1(C)) \] for all $r\in\kat{V}$. \item \label{rem:pres-discr-nerv-impl-unit-iso} If $H$ preserves $\kat{V}$-nerves of sets, then by the universal property of the colimit there is a $\kat{V}$-functor $\beta_X:(\Lan{D}H) DX \to HX$ such that the identity of $HX$ factorises as \[ \xymatrix@C=35pt{HX \ar[r]^{\alpha_{X}} \ar[dr]_{{\mathsf{Id}}} & H^\sharp DX \ar[d]^-{\beta_X}\\ & HX} \] Since $\alpha_X$ is the identity on objects, $\beta_X$ will again be so. Moreover, because both $\alpha_X$ and $\beta_X$ are $\kat{V}$-functors, the inequalities \[ HX(A',A) \leq H^\sharp DX(A',A) \leq HX(A',A) \] hold for each $A',A$. Hence $(\Lan{D}H) DX(A',A) = HX(A',A)$, that is, the unit of the left Kan extension $\alpha_X$ is an isomorphism (actually the identity). \end{enumerate} \end{rem} \begin{thm}\label{thm:unitiso} Let $H:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ a functor. The following are equivalent. \begin{enumerate} \item The unit of the left-Kan extension $H\to(\Lan{D}H)D$ is an isomorphism. \item Self-distances on elements of $H\emptyset$ are $\top$. \item All self-distances of all constants of $H$ are $\top$. \item $H$ preserves $\kat{V}$-nerves of sets. \end{enumerate} These conditions always hold if $\kat{V}$ is integral, or, for general $\kat{V}$, if $H\emptyset=\emptyset$. \end{thm} \begin{proof} For $(1)\Rightarrow (3)$ and $(2)\Leftrightarrow (3)$ see the comment after Proposition~\ref{prop:constantsset}. $(4)\Rightarrow (1)$ has been established in Remark~\ref{rem:pres-discr-nerves}. For $(2)\Rightarrow (4)$, let $X$ be an arbitrary set and assume that $(2)$ holds. We will show that the identity of $HX$ satisfies~\eqref{eq:couniversal-property} \[ r\leq \bigwedge_{C \in H(DX)_r} HX(Hd^r_0(C), Hd^r_1(C)) \] for each $r$, hence $H$ preserves $\kat{V}$-nerves of sets by Remark~\ref{rem:pres-discr-nerves}. We start by examining the the associated binary relations ($r$-level sets) of the (discrete) $\kat{V}$-category $DX$: \[ (DX)_r = \begin{cases} X \times X & r=\bot \\ \Delta_X=\{(x,x)\mid x\in X\} & \bot <r\leq e \\ \emptyset & \mbox{otherwise}\end{cases} \] Condition~\eqref{eq:couniversal-property} is automatically satisfied for $r=\bot$. \noindent If $\bot<r\leq e$, then $(DX)_r$ is the diagonal of $X$ and $d_0^r=d^r_1$ is a bijection. Hence $Hd_0^r=Hd_1^r$, being an isomorphism of $\kat{V}$-categories, implies that \[ r \leq e\leq \bigwedge_{C\in H\Delta_X} H\Delta_X(C,C)=\bigwedge_{C\in H(DX)_r} HX(Hd^r_0(C),Hd^r_1(C)) \] proving that~\eqref{eq:couniversal-property} holds for $\bot<r\leq e$. \noindent We will now consider the case $r\nleq e$. Then $(DX)_r=\emptyset$ and $d^r_0=d^r_1$ is the unique map $!:\emptyset \to X$. By hypothesis, $H\emptyset$ has all self-$\kat{V}$-distances equal to $\top$. Then \[ r \leq \top = \bigwedge_{C\in H\emptyset}H\emptyset(C,C)\leq \bigwedge_{C \in H\emptyset} HX(H! (C), H! (C)) = \bigwedge_{C\in H(DX)_r} HX(Hd^r_0(C),Hd^r_1(C)) \] which establishes \eqref{eq:couniversal-property} and, therefore, that $H$ preserves $\kat{V}$-nerves of sets. \end{proof} \subsection{Characterisation theorems and density} \ \bigskip\noindent There is a well-known connection between left Kan extensions and colimits. For example, the finitary (that is, filtered colimit preserving) functors ${\mathsf{Set}}\to{\mathsf{Set}}$ are precisely those which are left Kan extensions of their restrictions along the inclusion of finite sets \begin{equation}\label{eq:finitary} \vcenter{ \xymatrix@R=20pt@C=32pt{ {\mathsf{Set}} \ar[r]^{} & {\mathsf{Set}}\\ {\mathsf{Fin}} \ar[ru]_{} \ar[u]& }} \end{equation} In this section the functors that arise as $\Lan{D}{H}$ \begin{equation}\label{eq:discretearities} \vcenter{ \xymatrix{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \ar[r]^{} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\\ {\mathsf{Set}} \ar[ru]_(.58){H} \ar[u]^(.45)D& }} \end{equation} are characterised as those functors that preserve $\kat{V}$-nerves~\eqref{eq:Vnerve}. The proof proceeds by showing that the colimits exhibited in~\eqref{eq:Vnerve} and subsequently in Proposition~\ref{prop:Vcoinserter} constitute a density presentation of $D$~\cite[Section~5.4]{kelly:book}. \medskip\noindent The functors ${\mathsf{Set}}\to{\mathsf{Set}}$ that arise as left Kan extensions in diagram \eqref{eq:finitary} are precisely those that have a presentation by finitary operations and equations. Note that the elements $n\in{\mathsf{Fin}}$ appear as the arities $X^n\to X$ of the operations of the presentation. Similarly, we may speak of the functors that arise as left Kan extensions in diagram \eqref{eq:discretearities} as functors that have a \emph{presentation in discrete arities}, with the presentation given by the usual formula $(\Lan{D} H)\kat{X}=\int^S \kat{X}^{DS} \otimes HS$,% \footnote{Note that although ${\mathsf{Set}}$ is not small, the coend does exist, by~\cite[Section~4.2]{kelly:book} } % but which is more easily computed by applying $H$ to the $\kat{V}$-nerve as in Equation~\eqref{eq:H(Vnerve)}. \medskip\noindent A complication, studied in Proposition~\ref{prop:D-ff}, is that $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is not fully faithful if $\kat{V}$ is not integral, that is, if $e<\top$. \medskip\noindent Nevertheless we know from Corollary~\ref{cor:dense}, even without restricting to integral $\kat{V}$, that the discrete functor ${\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is dense. Here we show that this result also follows for general reasons that do not depend on the particular presentation of $\kat{V}$-categories as colimits of $\kat{V}$-nerves. \begin{prop}\label{prop:D-dense} The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is dense. \end{prop} \begin{proof} Let $\mathbb I$ denote the unit $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category (which has only one object $*$ and $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-hom $\mathbb I (*,*) = \mathbb 1$). Consider the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors $F:\mathbb I \to {\mathsf{Set}}, \quad F(*)=\underline{1}$, where $\underline{1}$ denotes a one-element set, and $G:\mathbb I \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}, \quad G(*)=\mathbb 1$. Then the usual coend formula gives $\mathsf{Lan}_F G = D$, using that $\mathbb I$ is small and $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is cocomplete. Now by~\cite[(5.17)]{kelly:book}, $G$ is dense, and by~\cite[Proposition~5.10]{kelly:book}, $D$ is also dense, given that $F$ is fully faithful. \end{proof} \medskip\noindent The result on density presentations in \cite[Chapter 5]{kelly:book} requires Kan extensions along fully faithful functors. We will henceforth demand in the remaining of this Section that the quantale $\kat{V}$ is integral, that is, $e=\top$, see Proposition~\ref{prop:D-ff}. \medskip\noindent Recall from Proposition~\ref{prop:Vcoinserter} that every $\kat{V}$-category $\kat{X}$ is the colimit $N\ast DF_\kat{X}$ of the diagram $(N,DF_\kat{X})$ of discrete $\kat{V}$-categories. \begin{thm}\label{thm:dense-pres} The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $D :{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is dense and the diagrams $(N,DF_\kat{X})$ form a density presentation of $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. % \end{thm} \begin{proof} We know that $D$ is dense from Proposition~\ref{prop:D-dense}. Using Proposition~\ref{prop:Vcoinserter}, we see that all $\kat{V}$-coinserters $N\ast DF_\kat{X}$ exist in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ and the category $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is the closure of ${\mathsf{Set}}$ under these colimits. % By~\cite[Theorem~5.19]{kelly:book}, in order to establish the density presentation of $D$, it remains to show that the colimits $N\ast DF_\kat{X}$ are preserved by the functors $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DS,-):\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ for all sets $S$. Abbreviate as earlier $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DS,\kat{X})=\kat{X}^{DS}$. Going back to the notation of the proof of Proposition~\ref{prop:Vcoinserter}, we notice that $(DX_r)^{DS}=(\kat{X}^{DS})_r$ and $(DX_{{\sf o}})^{DS}=(\kat{X}^{DS})_{{\sf o}}$, so that the diagram $(DF_\kat{X})^{DS}$ \[ \xymatrix{ (DX_r)^{DS} \ar@<0.5ex>[rr]^{(Dd_1^r)^{DS}} \ar@<-0.5ex>[rr]_{(Dd_0^r)^{DS}} && (DX_{{\sf o}})^{DS}} \] is equally the diagram $DF_{\kat{X}^{DS}}$ \[ \xymatrix{ (\kat{X}^{DS})r \ar@<0.5ex>[rr]^{} \ar@<-0.5ex>[rr]_{} && (\kat{X}^{DS})_{{\sf o}}}. \] This gives us, using Proposition~\ref{prop:Vcoinserter}, that \begin{align*} \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DS,N\ast DF_\kat{X}) & \cong \kat{X}^{DS} \\ & \cong N\ast DF_{\kat{X}^{DS}}\\ & \cong N\ast (DF_\kat{X})^{DS}\\ & \cong N\ast \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DS,DF_\kat{X}) \end{align*} showing that $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DS,-)$ preserves $N\ast DF_\kat{X}$. \end{proof} \noindent We next characterise the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors with presentations in discrete arities, that is, the functors that are left Kan extensions of their restrictions along $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, as those functors that preserve colimits of $\kat{V}$-nerves. \begin{thm}[Characterisation of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors with presentations in discrete arities]\label{thm:char-discrete-arities} For $G:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ the following are equivalent: \begin{enumerate} \item There exists a functor $H:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ such that $G\cong \Lan{D}{H}$. \item $G$ preserves the colimits of all diagrams $(N,DF_\kat{X})$. \end{enumerate} \end{thm} \begin{proof} Since the $(N,DF_\kat{X})$ form a density presentation (Theorem~\ref{thm:dense-pres}), this is the equivalence of items~(i) and~(iii) of \cite[Theorem~5.29]{kelly:book}. \end{proof} \noindent $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications are similarly characterised but additionally preserve discreteness. \begin{thm}[Characterisation of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications]\label{thm:char-Vcatification} For $G:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ the following are equivalent. \begin{enumerate} \item There exists a functor $T:{\mathsf{Set}}\to{\mathsf{Set}}$ such that $G\cong \Lan{D}{(DT)}$, that is, $G$ is \hspace{.2em}$T_\kat{V}$, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of $T$. \item $G$ preserves the colimits of all diagrams $(N,DF_\kat{X})$, and $G$ preserves discrete $\kat{V}$-categories. \end{enumerate} \end{thm} \begin{proof} We use Theorem~\ref{thm:char-discrete-arities} and observe that $G$ mapping discrete $\kat{V}$-categories to discrete $\kat{V}$-categories is equivalent to $GD$ factoring as $DT$ for some $T$. \end{proof} \begin{exa} \label{ex:non-Vcatif} We will now give an example of a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $G:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ which is not a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification, even though it preserves discreteness of $\kat{V}$-categories. $G$ is presented in non-discrete arities as follows: For a $\kat{V}$-category $\kat{X}$, put $G\kat{X} = \kat{X}^{{\mathbb{2}}_r}$, where the $\kat{V}$-category ${\mathbb{2}}_r$ was introduced in~\eqref{eq:Two_r}, for $r$ an element of the quantale. Notice that $G$ is the identity on discrete $\kat{V}$-categories, for every $\bot<r\leq e$. Hence, if $G$ was $\Lan{D}(DT)$ for some $T:{\mathsf{Set}} \to {\mathsf{Set}}$, we could choose $T$ to be the identity and then $G$ would also have to be the identity, by Theorem~\ref{thm:dense-pres} and~\cite[Theorem~5.1]{kelly:book}. \end{exa} \subsection{Extending polynomial and finitary functors}\label{sec:examples} \ \bigskip\noindent Let now $\kat{V}$ be an arbitrary commutative quantale. We show that functors defined by finite powers and colimits extend from ${\mathsf{Set}}$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ via their universal properties. In particular, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of a finitary functor is presented by the same operations and equations as its underlying ${\mathsf{Set}}$-functor. \begin{exas}[\bf The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of polynomial functors] \label{exs:polynomial_functors} \hfill \begin{enumerate} \item \label{ex:Vcatif-ct} Let $T:{\mathsf{Set}}\to{\mathsf{Set}}$, be a constant functor at a set $S$. Then $T_\kat{V}$ acts as follows: For any $\kat{V}$-category $\kat{X}$, $T_\kat{V}\kat{X}$ is the constant $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor to the $\kat{V}$-category with $S$ as set of objects, with $\kat{V}$-distances \[ T_\kat{V}\kat{X}(x',x) = \begin{cases} \, \top \ \ , \ \ x'=x \\ \bot \ \ , \ \ \mbox{otherwise} \end{cases} \] That is, $T_\kat{V} \kat{X} = S \cdot {\mathbb{1}}_\top$ is the coproduct in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ of $S$ copies of the terminal $\kat{V}$-category ${\mathbb{1}}_\top$. Observe that in case $\kat{V}$ is integral, we obtain $T_\kat{V} \kat{X}=D S$ for any $\kat{V}$-category $\kat{X}$. \item Let $T:{\mathsf{Set}}\to{\mathsf{Set}}$ be the functor $TX= X^n$, for $n$ a natural number. Then $T_\kat{V}$ maps a $\kat{V}$-category $\kat{X}$ to its $n$-th power $\kat{X}^n$, where an easy computation shows that \begin{equation}\label{eq:V-dist-tuples} \kat{X}^n((x'_0,\dots,x'_{n-1}),(x_0,\dots,x_{n-1})) = \kat{X}(x'_0,x_0)\wedge\dots\wedge\kat{X}(x'_{n-1},x_{n-1}). \end{equation} \item \label{ex:Vcatif-power} If $n$ is an {\em arbitrary\/} cardinal number, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $T_\kat{V}$ of $T:{\mathsf{Set}}\to {\mathsf{Set}}$, $TX=X^n$ also exists and $T_\kat{V}\kat{X}((x'_i),(x_i))=\bigwedge_i \kat{X} (x'_i,x_i)$. That is, $T_\kat{V}\kat{X}=\kat{X}^n$. \item The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of a finitary polynomial functor $ X\mapsto \coprod_n X^n\bullet \Sigma n $ is the ``strongly polynomial'' $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $ \kat{X}\mapsto \coprod_n \kat{X}^n \otimes D\Sigma n $, where $n$ ranges through finite sets. In particular, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of the list functor $LX = X^* = \coprod_n X^n$ maps a $\kat{V}$-category $\kat{X}$ to the $\kat{V}$-category having as objects tuples of objects of $\kat{X}$, with non-trivial $\kat{V}$-distances only between tuples of same order, given by~\eqref{eq:V-dist-tuples}. \end{enumerate} \end{exas} \smallskip\noindent % Our next aim is to show that quotients of polynomial functors can be $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ified by taking the ``same'' quotients in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \begin{rem} \label{rem:VV} The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ preserves conical colimits. This follows from the $D _o$ being an \emph{ordinary} left adjoint. However, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is not a left \emph{$\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-adjoint}, as its ordinary right adjoint functor $V$ cannot be extended to a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor. \end{rem} \begin{prop} \label{prop:Vcatification=functor} The assignment $({-})_\kat{V}:[{\mathsf{Set}},{\mathsf{Set}}]\to [\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]$, $T\mapsto T_\kat{V}$ of the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification preserves all colimits preserved by $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. % In particular, $T\mapsto T_\kat{V}$ preserves conical colimits. \end{prop} \begin{proof} Any natural transformation $\tau:T\to S$ induces a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $$ (\tau_\kat{V})_\kat{X} = \Colim{N}{(D\tau F\!_\kat{X})} : \Colim{N}{(D T F\!_\kat{X})} \to \Colim{N}{(D S F\!_\kat{X})} $$ Since any colimit is cocontinuous in its weight and since $$ \Colim{N}{(D T F\!_\kat{X})} \cong \Colim{(D T F\!_\kat{X})}{N} $$ holds~\cite[Section 3.1]{kelly:book}, the assignment $T\mapsto T_\kat{V}$ preserves all colimits that are preserved by $D:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. The last statement follows from Remark~\ref{rem:VV}. \end{proof} \begin{cor} \label{cor:presentation} Suppose that the coequalizer $$ \xymatrix{ T_\Gamma \ar@<.5ex>[0,1]^-{\lambda} \ar@<-.5ex>[0,1]_-{\rho} & T_\Sigma \ar[0,1]^-{\gamma} & T } $$ is the equational presentation of an accessible functor $T:{\mathsf{Set}}\to{\mathsf{Set}}$. Then the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $T_\kat{V}$ can be obtained as the coequalizer $$ \xymatrix{ (T_\Gamma)_\kat{V} \ar@<.5ex>[0,1]^-{\lambda_\kat{V}} \ar@<-.5ex>[0,1]_-{\rho_\kat{V}} & (T_\Sigma)_\kat{V} \ar[0,1]^-{\gamma_\kat{V}} & T_\kat{V} } $$ in $[\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]$. \end{cor} \begin{rem}[\bf The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of finitary functors] Corollary~\ref{cor:presentation} allows us to say that the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $T_\kat{V}$ of a finitary functor $T$ presented by operations and equations is given by imposing the ``same'' operations and equations in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$.% \footnote{ This explains~\cite[Remark~6.5(1)]{ammu:finitary-functors}. } This gives us an alternative way of computing the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of a finitary (or accessible) functor $T$: First, one presents $T$ as the quotient of a polynomial functor. Then one extends the polynomial functor to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ as in Example~\ref{exs:polynomial_functors}. Finally, one computes the relevant coequalizers to take a quotient in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ (see~\cite[Proposition~2.11]{wolff:Vcat&Vgraph} on coequalizers in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$). \end{rem} \begin{exa} The finite powerset functor is a quotient of the list functor by the familiar equations stating that sets are invariant under the order and multiplicity of elements. It is shown in \cite{VelebilK11} that adding to these equations the requirement that lists are monotone, one obtains the finite convex powerset functor, which is indeed the posetification of the finite powerset functor \cite{lmcs:bkv}. \end{exa} \subsection{Extending functors via relation lifting}\label{sec:wp} \ \bigskip\noindent Extending set-functors to various (ordered) categories using the so-called relation lifting is an established topic of research going back to~\cite{barr} and \cite{trnkova}. Using the representation of quantale-enriched categories as relational presheaves of Section~\ref{sec:relpresh}, we will see here that this idea can also be applied to the extension of ${\mathsf{Set}}$-functors to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ and, moreover, that $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications of weak-pullback preserving functors are more easily computed in this way. In particular, \eqref{eq:wpb} simplifies \eqref{eq:shortest-path} significantly. As an application, we will compute the extension of the multiset functor and show that the well-known Pompeiu-Hausdorff metric arises as a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of the powerset functor. \medskip\noindent As in ~\cite{wagner}, ``what we do in effect is to reason about metric spaces (or other structures) in a universe where they {\em are} pre-orders, viz.\ in sheaves over $[0, \infty]$ (or other $\kat{V}$).'' This point of view is summarised by Theorem~\ref{thm:wpbcontrelpresh} stating mild conditions under which the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification is obtained by applying the relation lifting to these preorders. \begin{rem}[Relation lifting, see for example~\cite{barr,trnkova,freyd-scedrov:allegories,HermidaJ98,kv:rellift}] Let $\mathsf{Rel}$ be the 2-category of sets with relations ordered by inclusion as arrows and $(-)_\diamond:{\mathsf{Set}}\to\mathsf{Rel}$ the functor mapping a function to its graph. Given a functor $T:{\mathsf{Set}}\to{\mathsf{Set}}$ the relation lifting $\mathsf{Rel}_T:\mathsf{Rel}\to\mathsf{Rel}$ extends $T$ to a colax functor $$ \xymatrix{ \mathsf{Rel} \ar[0,2]^-{\mathsf{Rel}_T} & & \mathsf{Rel} \\ {\mathsf{Set}} \ar[0,2]_-{T} \ar[-1,0]^{(-)_\diamond} & & {\mathsf{Set}} \ar[-1,0]_{(-)_\diamond} } $$ and is defined as follows. Let $p:Z\to X$ and $q:Z\to Y$ two arrows (called a span from $X$ to $Y$). A \emph{span $(p,q)$ represents the relation} \begin{equation}\label{eq:relofspan} q_\diamond \cdot (p_\diamond)^{-1} = \{(x,y) \mid \mbox{ there is $z\in Z$ s.t. $x=p(z)$ and $q(z)=y$}\} \end{equation} where $\cdot$ is composition and $^{-1}$ is relational inverse. Composition of two relations is represented by the pullback of their spans. Let $R$ be a relation represented by the span $(p,q)$. Then \emph{$\mathsf{Rel}_T(R)$ is the relation represented by $(Tp,Tq)$}. This definition is independent of the choice of span. In the following we only need to know that $\mathsf{Rel}_T$ is monotone, preserves identities, and that it preserves function composition iff $T$ preserves weak pullbacks. \qed \end{rem} \medskip\noindent The following lemma considerably simplifies the formula in Corollary~\ref{cor:Lan by zigzag paths} for computing the left-Kan extension $\Lan{D}{H}$ in case $H=DT$ with $T$ preserving weak pullbacks. \begin{lem}\label{lem:wpb} Let $T:{\mathsf{Set}}\to{\mathsf{Set}}$ be a weak pullback preserving functor. Then the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of $T$ is given by \begin{equation}\label{eq:wpb} T_\kat{V}\kat{X}(A',A) = \bigvee_s \{ s\mid (A',A)\in \mathsf{Rel}_T(X_s) \} \end{equation} \end{lem} \begin{proof} To show ``$\ge$'' assume $(A',A)\in \mathsf{Rel}_T(X_s)$. Then there is some $C\in T(X_s)$ such that $Td_0^s(C)=A'$ and $Td_1^s(C)=A$. Since this is a (very special) zig-zag of the form of Corollary~\ref{cor:Lan by zigzag paths} it follows that $T_\kat{V}\kat{X}(A',A) \ge s$. \medskip\noindent Conversely, for ``$\le$'', consider a zig-zag \begin{equation*} (A',A_0, C_1\in DTX_{r_1}, A'_1, A_2, C_1\in DTX_{r_2},\ldots , A_n') \end{equation*} as in Corollary~\ref{cor:Lan by zigzag paths}, contributing with $$r=r_1\otimes\ldots r_n$$ to the distance $T_\kat{V}\kat{X}(A',A)$. We want to see that $$(A',A)\in \mathsf{Rel}_T(X_r).$$ Note that $A'=A_0$ and $A'_i=A_{i+1}$, $0<i<n$, and $A'_n=A$, so that we are looking at the zig-zag \begin{equation}\label{eq:zigzag} \vcenter{ \xymatrix@C=0pt@R=45pt{ & DT X_{r_1} \ar[dl]_{DTd^{r_1}_0} \ar[dr]^{} & & DTX_{r_2} \ar[dl]_{} \ar[dr]^{} & & \quad \cdots \quad \ar[dl]_{\cdots} \ar[dr]^{\cdots} & & DTX_{r_n} \ar[dl]_{} \ar[dr]^{DTd^{r_n}_1} & & \\ DTX & & DTX & & DTX & & DTX & & DTX }} \end{equation} Consequently, $(A',A)$ is in the relation represented by the zig-zag~\eqref{eq:zigzag}, that is, in $$\mathsf{Rel}_T (X_{r_1})\cdot \mathsf{Rel}_T (X_{r_2})\cdot\ldots \cdot \mathsf{Rel}_T (X_{r_n})$$ We calculate \begin{align*} \mathsf{Rel}_T (X_{r_1})\cdot \mathsf{Rel}_T (X_{r_2})\cdot\ldots \cdot \mathsf{Rel}_T (X_{r_n}) & = \mathsf{Rel}_T(X_{r_1}\cdot X_{r_2}\cdot\ldots \cdot X_{r_n}) \\ & \subseteq \mathsf{Rel}_T(X_{r_1\otimes r_2\otimes\ldots \otimes r_n})\\ & =\mathsf{Rel}_T(X_r) \end{align*} where the first step is due to $T$ preserving weak pullbacks and the second step is due to $X_\_$ being lax monoidal, see \eqref{eq:monoidal2}. \end{proof} \begin{exa}[\bf The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of the multiset functor] The (finitary) multiset functor $M:{\mathsf{Set}}\to {\mathsf{Set}}$ can be presented as $MX = \coprod_{n}X^n/S_n$, where $S_n$ denotes the full $n$-permutation group~\cite{hasegawa}. Denote by $\widehat{[x_1, \ldots, x_n]}$ the equivalence class (multiset) of the tuple $(x_1, . . . , x_n)$ under the action of $S_n$. \noindent By Lemma~\ref{lem:wpb} and Equation~\eqref{eq:relofspan}, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $M_\kat{V}$ admits the following description: it maps a $\kat{V}$-category $\kat{X}$ to the $\kat{V}$-category having as objects multisets of objects of $\kat{X}$ -- formally, $(M_\kat{V} \kat{X})_{{\sf o}} = M X_{{\sf o}}$, and the only possible non-trivial $\kat{V}$-distances are between multisets of same cardinal, namely \begin{alignat*}{4} M_\kat{V} \kat{X}(\widehat{[x'_1,\ldots,x'_{n}]},\widehat{[x_1, \ldots, x_{n}]}) & = && \bigvee_s \{ s \mid \mbox{ there is } C \in MX_s \mbox{ s.t. } Md_0^s(C) = \widehat{[x'_1,\ldots,x'_{n}]} \\ &&& \mbox{ and } Md_1^s(C) = \widehat{[x_1,\ldots,x_{n}]}\, \} \\ & = && \bigvee_s \{ s \mid \mbox{ there is }\sigma\in S_n \mbox{ s.t. } s\leq \kat{X}(x'_i,x_{\sigma(i)}) \\ &&& \mbox{ for all } i=1,n \, \} \\ & = && \bigvee_{\sigma\in S_n} \bigvee_s \{s \mid s\leq \kat{X}(x'_i,x_{\sigma(i)}) \mbox{ for all } i=1,n \, \} \\ & = && \bigvee_{\sigma\in S_n} \bigwedge_i \kat{X}(x'_i, x_{\sigma(i)}) \end{alignat*} \noindent The reader might notice that for $\kat{V}=([0,\infty],\geq_{\mathbb R} +,0)$, we regain by the above formula the matching metric on multisets~\cite{dd:encycl-dist} employed in image recognition techniques. Also, it generalises~\cite{dahlqvist-kurz:calco17}, where the case $\kat{V}={\mathbb{2}}$ (that is, $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}={\mathsf{Preord}}$) was considered, in the sense that given a (pre)ordered set $(X,\leq)$, multisets on $X$ are ordered by \[ \widehat{[x'_1,\ldots,x'_{n}]} \leq \widehat{[x_1, \ldots, x_{n}]} \Longleftrightarrow \mbox{ there is } \sigma\in S_n \mbox{ s.t. } x'_i \leq x_{\sigma(i)} \mbox{ for all } i=1,n \] In fact, \cite{dahlqvist-kurz:calco17} corroborated with Lemma~\ref{lem:wpb} and \eqref{eq:relofspan} gives us the recipe to compute the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of any analytic ${\mathsf{Set}}$-functor,% \footnote{ Analytic functors ${\mathsf{Set}} \to {\mathsf{Set}}$ preserve weak pullbacks~\cite[Lemma~1.8]{hasegawa}. } % not just of the multiset functor. \end{exa} \begin{exa}[The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of the powerset]\label{ex:Vcatif-powerset} Let $\mathcal P:{\mathsf{Set}}\to{\mathsf{Set}}$ be the powerset functor. Using Lemma~\ref{lem:wpb} and Equation~\eqref{eq:relofspan}, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $\mathcal P_\kat{V}$ is described as follows. Let $\kat{X}$ be any small $\kat{V}$-category. Then the objects of $\mathcal P_\kat{V}\kat{X}$ are subsets of the set of objects of $\kat{X}$, while the $\kat{V}$-``distances'' in $\mathcal P_\kat{V} \kat{X}$ are computed as follows: \begin{alignat*}{4} \mathcal P_\kat{V}\kat{X}(A',A) & = && \bigvee_s % \{ % s\mid \mbox{ there is $C$ in $\mathcal PX_s$ s.t. $\mathcal Pd_0^s(C)=A'$ and $\mathcal Pd_1^s(C)=A$} % \} \\ & = && \bigvee_s \{ s \mid \forall \, x'\in A' \, \exists \, x\in A . \, s \leq \kat{X}(x',x) \mbox{ and } \forall \, x\in A \, \exists\, x'\in A' . \, s \leq \kat{X}(x',x) \} \end{alignat*} Remembering that the join in $\kat{V}$ denotes the infimum over all distances, we recognise the familiar Pompeiu-Hausdorff metric as seen in~\cite[Definition~2.2]{debakker-devink}, now generalised to categories enriched over an arbitrary commutative quantales $\kat{V}$. We will discuss the relationship to the more customary formulation in the following. \qed \end{exa} \medskip\noindent Recall from~\cite{fawcett-wood:ccd,raney} that a quantale (more generally, a poset) $\kat{V}$ is called {\em completely distributive} if the function from the poset of downsets of $\kat{V}$ to $\kat{V}$, $\frak D \mapsto \bigvee \frak{D}$, has a left adjoint. This left adjoint maps each $r\in \kat{V}$ to the downset $\{s \mid s\ll r\}$ of all elements totally-below it, where the ``totally-below" relation is $$ s\ll r \quad \Longleftrightarrow \quad (\, \forall \, \frak{D}\subseteq \kat{V} \mbox{ downset such that} \, \, r \leq \bigvee \frak{D} \Rightarrow \, s \in \frak{D}\, ) $$ Examples of completely distributive quantales are $\kat{V}=[0,1]$ and $\kat{V}=[0,\infty]$. For such a quantale, we provide below another characterisation of the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of the powerset functor: \begin{prop}[The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of the powerset, continued]\label{prop} If the quantale $\kat{V}$ is completely distributive, then the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of the powerset functor can be described as follows: for a $\kat{V}$-category $\kat{X}$, $\mathcal P_\kat{V}\kat{X}$ is the $\kat{V}$-category with objects $\mathcal PX_{{\sf o}}$, and $\kat{V}$-homs \begin{equation}\label{Vcat-powerset} \mathcal P_\kat{V} \kat{X}(A', A) = \big(\bigwedge_{x'\in A'} \bigvee_{x\in A} \kat{X}(x',x) \big) \bigwedge \big(\bigwedge_{x\in A} \bigvee_{x'\in A'} \kat{X}(x',x)\big) \end{equation} \end{prop} \noindent We will first need a lemma: \begin{lem}\label{lemma} Let $r\in \kat{V}$ and $S\subseteq \kat{V}$. Then $r\ll \bigvee S$ if and only if there is some $s\in S$ with $r\ll s$. \end{lem} \begin{proof}[Proof of Lemma~\ref{lemma}] The ``if'' implication is immediate. To see the other implication, assume that for all $s\in S$, $r\not\ll s$. That is, for each $s\in S$ there is a downset $\frak{D}_s$ with $s\leq \bigvee \frak{D}_s$ and $r \not\in \frak{D}_s$. But then $\bigcup_{s\in S} \frak{D}_s$ is again a downset and $$ \bigvee S \leq \bigvee \bigcup_{s\in S} \frak{D}_s \quad \mbox{ but } r \not \in \bigcup_{s\in S} \frak{D}_s $$ which contradicts $r \ll \bigvee S$. \end{proof} \medskip\noindent We are now ready to prove Proposition~\ref{prop}. \begin{proof}[Proof of Proposition~\ref{prop}] Remember that in~Example~\ref{ex:Vcatif-powerset}, we have obtained that \begin{alignat*}{3} \mathcal P_\kat{V} \kat{X}(A',A) &=& \bigvee_s \{ s\mid \mbox{% $\forall\, x'\in A' \ \exists\, x\in A.\, s\leq\kat{X}(x',x)$ and $\forall\, x\in A\ \exists\, x'\in A'.\, s\leq\kat{X}(x',x)$% } \} \end{alignat*} Denote for simplicity $$ S=\{ s\mid \mbox{ $\forall\, x'\in A' \ \exists\, x\in A.\, s\leq\kat{X}(x',x)$ and $\forall\, x\in A\ \exists\, x'\in A'.\, s\leq\kat{X}(x',x)$ } \} $$ and notice that $S$ is a downset; % also put $$ r=\big(\bigwedge_{x'\in A'} \bigvee_{x\in A} \kat{X}(x',x) \big) \bigwedge \big(\bigwedge_{x\in A} \bigvee_{x'\in A'} \kat{X}(x',x)\big) $$ Thus we want to see that $r=\displaystyle{\bigvee} S$. The inequality ``$\geq$'' follows easily with no assumption on the quantale $\kat{V}$. For the other inequality, use complete distributivity to see that $$ r \leq \bigvee S $$ is equivalent with $$ \{s \mid s\ll r\} \subseteq S $$ Let now $s\ll r$. Then for each $x'\in A'$, it follows that $s \ll \bigvee_{x\in A} \kat{X}(x',x)$, thus by the above lemma we find some $x\in A$ with $s \ll \kat{X}(x',x)$, which implies $s \leq \kat{X}(x',x)$. Similarly, for all $x\in A$ there is $x'\in A'$ with $s \leq \kat{X}(x',x)$. Consequently, $s\in S$ and the proof is finished. \end{proof} \begin{rem} \begin{enumerate} \item Let us switch notation to the dual order (that is, the natural ``less-or-equal'' order in case of the reals). So we write $\inf$ for $\bigvee$ and $\sup$ for $\bigwedge$, in order to emphasise the interpretation of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$'s as metric spaces. Then formula~\eqref{Vcat-powerset} becomes \begin{equation}\label{eq:Pompeiu-Hausdorff} \sup \{ \ \sup_{x'\in A'}\ \underset{x\in A}{\vphantom{p}\inf\vphantom{p}} \ \kat{X}(x',x) \ , \ \sup_{x\in A}\ \underset{x'\in A'}{\vphantom{p}\inf\vphantom{p}} \ \kat{X}(x',x) \} \end{equation} As mentioned earlier, this metric given by Equation~\eqref{eq:Pompeiu-Hausdorff} above, is known as the Pompeiu-Hausdorff metric (\cite[\textsection~28]{Hausdorff},~\cite[\textsection~21]{Pompeiu},~\cite[\textsection~21.VII]{kuratowski}). We should mention also the connection with the work of~\cite{akvhlediani+clementino+tholen:hausdorff}. Finally, observe that in case $\kat{V}={\mathbb{2}}$ (so $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}={\mathsf{Preord}}$), the above specialises to the locally monotone functor $\mathcal P_{\mathbb{2}}:{\mathsf{Preord}}\to{\mathsf{Preord}}$ which sends a preorder $(X,\leq)$ to the Egli-Milner preorder $$ A'\sqsubseteq A \quad \mbox{iff} \quad \forall\, x'\in A' \ \exists\, x\in A.\, x'\leq x \mbox{ and } \forall\, x'\in A \ \exists\, x\in A'.\, x'\leq x $$ on the powerset $\mathcal PX$. \item Notice that the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of the powerset functor obtained in Proposition~\ref{prop} for completely distributive quantale is the (symmetrised version of the) free cocompletion monad with respect to (the saturated class of) conical colimits in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$~\cite{stubbe:hausdorff}. \end{enumerate} \end{rem} \medskip\noindent Example~\ref{ex:Vcatif-powerset} raises the question whether the relations $(T_\kat{V}\kat{X})_r$ can be computed simply by applying the relation lifting $\mathsf{Rel}_T$ to the $X_r$. The next example shows that we need to be careful with this. \begin{exa} Let $\kat{X}$ be the metric space of real numbers considered as a generalised metric space over $\kat{V}=[0, \infty]$. Let $A'$ be the subset of irrational reals and $A$ the subset of rational reals. Then $\mathcal P_\kat{V}\kat{X}(A',A)=0$ but $(A',A)\notin \mathsf{Rel}_\mathcal P(X_0)$. Indeed, $\mathsf{Rel}_\mathcal P(X_0)$ is the diagonal, since $X_0$ is the diagonal. \end{exa} \noindent In terms of Section~\ref{sec:relpresh}, the example above shows that the relational presheaf \begin{align*} (\Sigma\kat{V})^\mathsf{coop} & \to \ \mathsf{Rel}\\ r \ \ \ & \ \mapsto \ \mathsf{Rel}_\mathcal P(X_r) \end{align*} is not continuous as we have $(A',A)\in \mathsf{Rel}_\mathcal P(X_\varepsilon), \varepsilon >_{\mathbb R} 0$, but $(A',A)\notin \mathsf{Rel}_\mathcal P(X_0)$. We can repair this defect by closing the presheaf as follows. Given a relational presheaf $F:(\Sigma\kat{V})^\mathsf{coop}\to\mathsf{Rel}$, and assuming that $\kat{V}$ is completely distributive, define its closure as \begin{align*} \mathcal CF: (\Sigma\kat{V})^\mathsf{coop} & \to \ \mathsf{Rel} \\ * \ \ \ & \ \mapsto F(*) \\ r \ \ \ & \ \mapsto \ \bigcap\{F(s) \mid s \ll r\} \end{align*} It is immediate from the respective definitions that $\mathcal CF$ is a continuous relational presheaf, assuming the tensor laxly preserves the totally-below relation, in the sense that $r\ll s$ and $r'\ll s'$ imply $s \otimes s'\ll r \otimes r'$. Also, that $\mathcal C(\mathcal C F) = \mathcal CF$. That is, $\mathcal C$ acts like a closure operator on relational presheaves. This construct extends to a functor $\mathcal C:\mathsf{RelPresh} \to \mathsf{RelPresh}_c$. \medskip\noindent The idea that the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of a weak pullback preserving ${\mathsf{Set}}$-functor $T$ can be computed on a $\kat{V}$-category $\kat{X}$ by applying the relation lifting of $T$ to the ``uniformity'' $(X_r)_{r\in\kat{V}}$ of $\kat{X}$ can now be formalised as in next result. It is a consequence of Proposition~\ref{prop:relpresh} and of Lemma~\ref{lem:wpb}. \begin{thm}\label{thm:wpbcontrelpresh} Suppose $\kat{V}$ is a completely distributive quantale, such that the tensor laxly preserves the totally-below relation. Let $T:{\mathsf{Set}}\to{\mathsf{Set}}$ be a functor which preserves weak pullbacks. Then post-composition with $\mathsf{Rel}_T:\mathsf{Rel}\to\mathsf{Rel}$ yields a functor $$ \mathsf{Rel}_T \circ (-): \mathsf{RelPresh}_c\to\mathsf{RelPresh} $$ such that the diagram $$ \xymatrix{ \mathsf{RelPresh}_c \ar[0,2]^-{\mathsf{Rel}_T \circ (-)} && \mathsf{RelPresh} \ar[0,2]^-{\mathcal C} && \mathsf{RelPresh}_c \\ {\mathsf{Set}} \ar[0,2]^-{T} \ar[-1,0]^{D} && {\mathsf{Set}} \ar[-1,0]^{D} \ar[0,2]^-{{\mathsf{Id}}} && {\mathsf{Set}} \ar[-1,0]_{D} } $$ commutes. Moreover, the composite in the upper row is the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of $T$. \end{thm} \begin{proof} Let $\kat{X}$ be a $\kat{V}$-category and let $\Phi(\kat{X})$ be its continuous relational presheaf. We compute \begin{align*} (A',A)\in (T_\kat{V}\kat{X})_r & \ \Longleftrightarrow \ r \leq \bigvee_s \{ s\mid (A',A)\in \mathsf{Rel}_T(X_s) \} \\ & \ \Longleftrightarrow \ \forall s \, .\, s \ll r \, \Longrightarrow \, (A',A)\in \mathsf{Rel}_T(X_s)\\ & \ \Longleftrightarrow \ (A',A)\in\bigcap \{ \mathsf{Rel}_T(X_s) \mid s \ll r \} \\ & \ \Longleftrightarrow \ (A',A)\in \mathcal C (\mathsf{Rel}_T\circ \Phi(\kat{X}))(r) \end{align*} where the respective steps are due to the following. (i) Lemma~\ref{lem:wpb}. (ii) For all downsets $\frak{D}$ in a completely distributive $\kat{V}$ one has $r \leq \bigvee \frak{D} \Longleftrightarrow \forall s(s\ll r \Longrightarrow s\in \frak{D})$. (iii) Definition of $\bigcap$. (iv) Definition of $\mathcal C$. \end{proof} \subsection{Extending ${\mathsf{Set}}$-functors equipped with a $\kat{V}$-metric} \label{sec:Vmetric-functors} \ \bigskip\noindent In Section~\ref{sec:beh-cat}, we will see that the behaviour of coalgebras for a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification is metrically trivial. But often ${\mathsf{Set}}$-functors carry a natural metric and thus can be considered to be functors ${\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, which in turn produce $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-endofunctors with non-trivial metric coalgebraic behaviour. \begin{exa}\label{ex:functors-with-V-metric} \begin{enumerate} \item Consider $\kat{V}={\mathbb{2}}$ and let $H:{\mathsf{Set}}\to{\mathsf{Preord}}$ be the powerset functor with the usual inclusion relation on subsets $HX=(\mathcal P(X),\subseteq)$. Then by~\cite{calco2011:balan-kurz}, the extension $H^\sharp$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}={\mathsf{Preord}}$ maps a (pre)ordered set $(X,\leq)$ to $\mathcal P(X)$, ordered with the lower half of the Egli-Milner order on subsets, namely \[ A' \leq A \iff \forall \, x'\in A' \,.\, \exists \, x\in A \,.\, x'\le x \] for all $A',A\in \mathcal P(X)$. \item \label{ex:machine} Consider the functor $T:{\mathsf{Set}}\to {\mathsf{Set}}$, $TX=X^{A}\times B$. Recall that $T$-coalgebras are known as deterministic automata with input set $A$ and outputs in $B$: the transition map of a coalgebra $X\longrightarrow X^A\times B$ sends a state $x\in X$ to a pair: the associated next-state function $next(x):A\longrightarrow X$ that works when receiving an input from $A$, and an output $out(x)$ in $B$. \medskip\noindent Assume that the output set $B$ carries the additional structure of a $\kat{V}$-category -- that, is, there is a $\kat{V}$-category $\kat{B}$ with underlying set of objects $B$. This is the case for example whenever there is a natural order on $B$, or a metric, which measures how far apart two outputs might be. \medskip\noindent Then the ordinary functor $T_o$ can be written as the composite $V H_o$, where $H:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $HX= D X^A \otimes \kat{B}$. Now it is immediate to see that the latter extends to the functor $H^\sharp$ on $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ over the ``generalised metric space'' $\kat{B}$, mapping a $\kat{V}$-category $\kat{X}$ to the tensor product of $\kat{V}$-categories $H^\sharp \kat{X} = \kat{X}^A \otimes \kat{B}$, where $\kat{X}^A= \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DA,\kat{X})$. \medskip\noindent There are two particular cases of this functor that are worth mentioning from a coalgebraic point of view: \begin{enumerate} \item \label{ex:stream} We start with the easiest case, when $A$ is a singleton. The coalgebras for the functor $TX=X\times B$ are usually called stream automata; under the assumption that $B$ carries the structure of a $\kat{V}$-category $\kat{B}$, we obtain as above $H^\sharp \kat{X} = \kat{X} \otimes \kat{B}$ the stream functor on $\kat{V}$-metric spaces. \item \label{ex:det-autom} If we take an arbitrary set of inputs $A$, but particularise $B$ to $\{0,1\}$, then the transition function of a coalgebra provides binary outputs, deciding if a state is accepting (response $1$) or not -- that is, $T$-coalgebras are deterministic automata. Enhance now the set of outputs with a $\kat{V}$-category structure ${\mathbb{2}}_{r,s}$ generalising~\eqref{eq:Two_r} \[ {\mathbb{2}}_{r,s}(0,0)={\mathbb{2}}_{r,s}(1,1)=e\, , \, {\mathbb{2}}_{r,s}(0,1)=r\, , \, {\mathbb{2}}_{r,s}(1,0)=s \] where the elements $r,s$ of the quantale satisfy $r\otimes s \leq e$, in order to produce the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-endofunctor $H^\sharp \kat{X} = \kat{X}^A \otimes {\mathbb{2}}_{r,s}$. \end{enumerate} \end{enumerate} \end{exa} \medskip\noindent The above examples are typical. It happens quite often for endofunctors $T$ on ${\mathsf{Set}}$ to carry an interesting $\kat{V}$-metric. Then $TX$ is a $\kat{V}$-category rather than a mere set. The following generalises the notion of an order on a functor \cite{hugh-jaco:simulation} from ${\mathbb{2}}$ to an arbitrary quantale $\kat{V}$. \begin{defi} Let $T:{\mathsf{Set}} \to {\mathsf{Set}}$ be a functor. We say that $T$ {\em carries a $\kat{V}\hspace{-1.5pt}$-metric\/} if there is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ such that $T_o$ coincides with the composite $$ \xymatrix{{\mathsf{Set}}_o \ar[r]^-{H_o}& \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o \ar[r]^-{V} & {\mathsf{Set}}_o}. $$ \end{defi} \medskip \noindent Let $T$ and $H$ be as in the above definition. How are $T$ and $H^\sharp$, the left Kan extension of $H$ along $D$ as provided by Theorem~\ref{thm:Lan}, related? \medskip\noindent Recall from Theorem~\ref{thm:Lan} and Corollary~\ref{cor:Lan by zigzag paths} that the unit $\alpha:H \to H^\sharp D$ of the left Kan extension is the identity on objects, in particular that $\alpha$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural isomorphism in case $D$ is fully faithful, that is, if the quantale is integral. Hence $T_o = V H_o \cong V H_o^\sharp D_o$ always holds; using now that the counit of the ordinary adjunction $D_o \dashv V$ is again the identity on objects, we obtain an ordinary natural transformation $$ \xymatrix@R=20pt{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o \ar[0,2]^-{H^\sharp_o} & & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o \\ {\mathsf{Set}}_o \ar[0,2]_-{T_o} \ar@{<-}[-1,0]^{V} \ar@{}[-1,2]|{\nearrow\beta} & & {\mathsf{Set}}_o \ar@{<-}[-1,0]_{V} } $$ \begin{prop}\label{prop:hsharp-lifting} The natural transformation $\beta$ is component-wise bijective. \end{prop} \noindent Consequently, $H^\sharp$ is a lifting of $T$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ as in Definition~\ref{def:ext-lift}\eqref{def:lift}. \medskip\noindent We will exhibit below another possible way of lifting $T$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \begin{exa}[\bf The Kantorovich lifting]\label{exle:kantorovich} Let $T:{\mathsf{Set}} \to {\mathsf{Set}}$ be a functor and let $\heartsuit: T\kat{V}_o \to \kat{V}_o$ be an arbitrary map (a $\kat{V}$-valued predicate lifting), where by slight abuse we identify the quantale with its underlying set of elements. \medskip\noindent Using the $\kat{V}$-valued predicate lifting $\heartsuit$, we can endow $T$ with a $\kat{V}$-metric $H:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ as follows: for each set $X$, put $HX$ to be the $\kat{V}$-category with set of objects $TX$, and $\kat{V}$-distances \begin{equation}\label{eq:discrete-kant} (HX) (A',A) = \bigwedge_{h: X \to \kat{V}_o} [(\heartsuit \circ Th)(A'), (\heartsuit \circ Th)(A)] \end{equation} Remark that the $\kat{V}$-category $HX$ is precisely the initial lift of $TX$ with respect to the family of maps $\heartsuit \circ Th:TX \to \kat{V}_o$ indexed by all $h:X \to \kat{V}_o$. In particular, the $\kat{V}$-distances $(HX) (A',A)$ are the greatest such that all maps $\heartsuit \circ Th$ are actually $\kat{V}$-functors $HX \to \kat{V}$. \noindent For a map $f:X \to Y$, we let $Hf$ act as $Tf$ on objects. It is easy to see that $Hf:HX \to HY$ is a $\kat{V}$-functor. \medskip\noindent The above defines a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ (the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enrichment being a consequence of ${\mathsf{Set}}$ being free as a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category) with $V H_o=T_o$, for every map \linebreak $\heartsuit:T\kat{V}_o\to \kat{V}_o$. \medskip\noindent That is, $H$ is a $\kat{V}$-metric for $T$ and we may consider the lifting of $T$ given by the left Kan extension $H^\sharp$, as discussed above. \noindent In order to obtain (possibly) another lifting of $T$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, we make an additional assumption: $\heartsuit$ is $\kat{V}$-monotone,% \footnote{ This generalises the notion of monotone predicate lifting from the two-element quantale to arbitrary $\kat{V}$, see~\cite[Section~7]{lmcs:bkv}.} % in the following sense: for every set $X$ and maps $h,k: X \to \kat{V}_o$, the inequality \begin{equation}\label{eq:pred-V-monot} \bigwedge_{x\in X} [h(x),k(x)] \leq \bigwedge_{A \in TX} [(\heartsuit \circ Th)(A), (\heartsuit \circ Tk)(A)] \end{equation} holds.% \footnote{ In categorical terms, it says that the correspondence $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DX,\kat{V}){\rightarrow}\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(DTX,\kat{V})$, $h \mapsto \heartsuit \circ Th$ is a $\kat{V}$-functor.} % In particular, taking $X=\emptyset$ and $h=k$ the unique map from the empty set, we see that $\kat{V}$-monotonicity of $\heartsuit$ entails that self-distances on objects of $H\emptyset$ are $\top$, hence $H^\sharp D=H$ by Theorem~\ref{thm:unitiso}. \medskip\noindent Consider the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $\bar H:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ which maps a $\kat{V}$-category $\kat{X}$ to the $\kat{V}$-category $\bar H \kat{X}$ with set of objects $T X_{{\sf o}}$, and $\kat{V}$-homs \begin{equation}\label{eq:kant} \bar H \kat{X}(A',A) = \bigwedge_{h: \kat{X} \to \kat{V}} [(\heartsuit \circ Th_{{\sf o}})(A'), (\heartsuit \circ Th_{{\sf o}})(A)] \end{equation} for every $A', A$ in $TX_{{\sf o}}$, where this time $h$ ranges over $\kat{V}$-functors and the notation $h_{{\sf o}}$ refers to the object assignment of a $\kat{V}$-functor $h$, as in item~\eqref{item 2.a} of the proof of Theorem~\ref{thm:Lan}. \noindent As above for $H$, the $\kat{V}$-category $\bar H \kat{X}$ is precisely the initial lift of $TX_{{\sf o}}$ with respect to the family of maps $\heartsuit \circ Th_{{\sf o}}:TX_{{\sf o}} \to \kat{V}_o$ (indexed by all $\kat{V}$-functors $h:\kat{X} \to \kat{V}$), in particular~\eqref{eq:kant} provides the greatest $\kat{V}$-distances such that for all $\kat{V}$-functors $h:\kat{X} \to \kat{V}$, the composite maps $\heartsuit \circ Th_{{\sf o}}$ are also $\kat{V}$-functors. \noindent For $f:\kat{X} \to \kat{Y}$ a $\kat{V}$-functor, let $\bar Hf$ act as $Tf_{{\sf o}}$ on objects. Observe that it is the $\kat{V}$-monotonicity of $\heartsuit$ which ensures that $\bar H$ is indeed a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor. It is clear that this $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $\bar H$ is a lifting for $T$. The particular case $\kat{V}=[0,\infty]$ provides a generalised Kantorovich lifting as in~\cite{bbkk:calco2015-ext}. \medskip\noindent Relations~\eqref{eq:discrete-kant} and~\eqref{eq:kant} ensure that for each $\kat{V}$-category $\kat{X}$, the identity on objects is a $\kat{V}$-functor $\bar c_\kat{X}:HX_{{\sf o}}\to \bar H\kat{X}$. We will now prove that $\bar c_\kat{X}$ satisfies~\eqref{eq:couniversal-property}, thus it induces a $\kat{V}$-functor $\gamma_\kat{X}: H^\sharp \kat{X} \to \bar H\kat{X}$ such that $\gamma_\kat{X} \circ c_\kat{X} = \bar c_\kat{X}$. \noindent Let thus $r$ an arbitrary element of the quantale. Then we have \begin{equation}\label{eq:proof1} \begin{aligned} & \bigwedge_{C \in HX_r} \bar H \kat{X} (\bar c_\kat{X} Hd^r_0(C), \bar c_\kat{X} Hd^r_1(C)) & = & \bigwedge_{(A',A)\in \mathsf{Rel}_T(X_r)} \bar H \kat{X} (\bar c_\kat{X} (A'), \bar c_\kat{X} (A)) \\ & & = & \bigwedge_{(A',A)\in \mathsf{Rel}_T(X_r)} \bigwedge_{h:\kat{X} \to \kat{V}} [\heartsuit\circ Th_{{\sf o}}(A'),\heartsuit\circ Th_{{\sf o}}(A)] \\ & & \geq & \quad \bigwedge_{w\in T\kat{V}_r} [\heartsuit\circ Td^r_0(w),\heartsuit\circ Td^r_1(w)] \end{aligned} \end{equation} because \[ (Th_{{\sf o}} \times Th_{{\sf o}})(\mathsf{Rel}_T(X_r)) \subseteq \mathsf{Rel}_T((h_{{\sf o}} \times h_{{\sf o}})(X_r)) \subseteq \mathsf{Rel}_T(\kat{V}_r) \] In the above, we have used that $h:\kat{X} \to \kat{V}$ is a $\kat{V}$-functor and that relation lifting preserves inclusions, and denoted $\kat{V}_r = \{(s',s) \mid r\leq [s',s]\}$ and $d^r_0, d^r_1:\kat{V}_r \to \kat{V}_{{\sf o}}$ as for any $\kat{V}$-category. \noindent Using now~\eqref{eq:pred-V-monot} for the set $\kat{V}_r$ and for the maps $d^r_0, d^r_1$, we see that \begin{equation}\label{eq:proof2} r \leq \bigwedge_{(s',s)\in \kat{V}_r} [d^r_0(s',s),d^r_1(s',s)] \leq \bigwedge_{w \in T\kat{V}_r} [\heartsuit\circ Td^r_0(w),\heartsuit\circ Td^r_1(w)] \end{equation} Putting together~\eqref{eq:proof1} and~\eqref{eq:proof2} shows that $\bar c_\kat{X}$ satisfies~\eqref{eq:couniversal-property}. \medskip\noindent In conclusion, there is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $H^\sharp \to \bar H$ whose components act as identity on objects, that is, the $\kat{V}$-distances in $H^\sharp \kat{X}$ are always smaller than in $\bar H\kat{X}$. \medskip\noindent Corroborating the above with the particular example of the Hausdorff metric $\bar H$ on the (finite) powerset functor $T=\mathcal P$ exhibited in~\cite{bbkk:calco2015-ext}, it raises the question under which conditions $\bar H \cong H^\sharp$ holds, very much in the spirit of the Kantorovich duality~\cite{villani}. Currently this aspect is still under research and will not be treated in this paper. \qed \end{exa} \subsection{Closure properties of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications} \label{sec:closure} \ \bigskip\noindent We have seen in Proposition~\ref{prop:Vcatification=functor} that $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications are closed under conical colimits. In this section, we give sufficient conditions for the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications to be closed under composition and finite products. \medskip\noindent We will first discuss when $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications are closed under composition: Generalising a result from \cite[Remark 4.8(4)]{lmcs:bkv} for the case $\kat{V}={\mathbb{2}}$, given weak pullback preserving functors $T,T':{\mathsf{Set}}\to {\mathsf{Set}}$, we will see that the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of $TT'$ is the composite of the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ifications $T_\kat{V}$ and $T'_\kat{V}$ if the quantale $\kat{V}$ satisfies the additional assumption of being completely distributive. \medskip \noindent Observe that by the universal property of left Kan extensions, there is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $(TT')_\kat{V} \to T_\kat{V} T'_\kat{V}$, for any functors $T,T':{\mathsf{Set}} \to {\mathsf{Set}}$. The component of this natural transformation at a $\kat{V}$-category $\kat{X}$, namely the $\kat{V}$-functor $(TT')_\kat{V}\kat{X} \to T_\kat{V} T'_\kat{V}\kat{X}$, is easily seen to act as identity on objects,\footnote{ Notice that both $\kat{V}$-categories $(TT')_\kat{V} \kat{X}$ and $T_\kat{V} T'_\kat{V} \kat{X}$ share the same set of objects, namely $TT' X_{{\sf o}}$.} and on $\kat{V}$-homs it provides the inequality \[ (TT')_\kat{V} \kat{X} (A',A) \leq T_\kat{V} T'_\kat{V} \kat{X}(A',A) \] for all objects $A',A \in TT'X_{{\sf o}}$. \medskip \begin{prop} Assume that the quantale $\kat{V}$ is completely distributive. Let $T,T':{\mathsf{Set}}\to {\mathsf{Set}}$ be functors which preserve weak pullbacks. Then the inequality \[ T_\kat{V} T'_\kat{V} \kat{X}(A',A) \leq (TT')_\kat{V} \kat{X}(A',A) \] holds for all $A',A \in TT'X_{{\sf o}}$. \end{prop} \begin{proof} Because $T$ preserves weak pullbacks, $T_\kat{V} T'_\kat{V} \kat{X} (A',A) = \bigvee \{r \mid (A',A) \in \mathsf{Rel}_T((T'_\kat{V} \kat{X})_r)\}$ by Lemma~\ref{lem:wpb}. Let $r\in \kat{V}_o$ such that $(A',A)\in \mathsf{Rel}_T((T'_\kat{V} \kat{X})_r)$. Given that $\kat{V}$ is completely distributive, to show that \[ r\leq (TT')_\kat{V} \kat{X}(A',A) = \bigvee\{s \mid (A',A)\in \mathsf{Rel}_{TT'}(X_s) \} \] is the same as \[ \forall \, s\in \kat{V}_o\,. \, s \ll r \, \Longrightarrow (A',A) \in \mathsf{Rel}_{TT'}(X_s) \] because $\{s \mid (A',A)\in \mathsf{Rel}_{TT'}(X_s) \}$ is a downset and $\kat{V}$ is completely distributive. Also notice that \begin{equation} \label{eq:r-cut} (T'_\kat{V} \kat{X})_r \subseteq \mathsf{Rel}_T'(X_s) \ \ \ \end{equation} holds for all $s \ll r$. Indeed, for $(B',B) \in (T'_\kat{V} \kat{X})_r$, we have that \[ r \leq T'_\kat{V} \kat{X} (B',B) = \bigvee \{ s \mid (B',B)\in \mathsf{Rel}_T'(X_s) \} \] using that $T'$ preserves weak pullbacks and Lemma~\ref{lem:wpb}. But the relation above is equivalent to \[ \forall \, s\in \kat{V}_o\,. \, s \ll r \, \Longrightarrow (B',B) \in \mathsf{Rel}_T'(X_s) \] Hence $(T'_\kat{V} \kat{X})_r \subseteq \mathsf{Rel}_T'(X_s)$ holds for all $s \ll r$. Therefore we have \[ (A',A)\in \mathsf{Rel}_T((T'_\kat{V} \kat{X})_r) \subseteq \mathsf{Rel}_T(\mathsf{Rel}_{T'}(X_s)) = \mathsf{Rel}_{TT'}(X_s) \] for all $s\ll r$, where the last equality follows from~\cite[Section~4.4]{ckw}. \end{proof} \begin{cor} Let $\kat{V}$ be a completely distributive commutative quantale and $T,T':{\mathsf{Set}} \to {\mathsf{Set}}$ functors which preserve weak pullbacks. Then \[ (TT')_\kat{V} = T_\kat{V} T'_\kat{V} \] holds. \end{cor} \noindent Next we show that the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification process is also closed under finite products, if some additional conditions are imposed to $\kat{V}$. \begin{prop} \label{prop:presfinprod} Let $\kat{V}$ be an integral commutative quantale such that finite meets distribute over arbitrary joins in $\kat{V}$ (the underlying lattice of $\kat{V}$ is a frame). Then the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification assignment $({-})_\kat{V}:[{\mathsf{Set}},{\mathsf{Set}}]\to [\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]$, $T\mapsto T_\kat{V}$, preserves finite products of weak pullback preserving functors. \end{prop} \begin{proof} The empty product, that is, the constant functor ${\mathsf{Set}} \to {\mathsf{Set}}$ at the one-element set has as $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification the constant functor at ${\mathbb{1}}$. Because $e=\top$ holds in $\kat{V}$ by hypothesis, the latter is the terminal object in $[\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}, \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]$~\cite{bird:PhD}. \medskip\noindent Let us look now at binary products of weak-pullback preserving functors $T,T':{\mathsf{Set}} \to {\mathsf{Set}}$. We will denote by $T\times {T'}$ the product in $[{\mathsf{Set}},{\mathsf{Set}}]$, sending $X$ to $TX \times {T'}X$ and shall extend this notation to $[\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}]$. \medskip\noindent Now observe that for a $\kat{V}$-category $\kat{X}$, the $\kat{V}$-categories $T_\kat{V}\kat{X} \times T'_\kat{V}\kat{X}$ and $(T\times {T'})_\kat{V} \kat{X}$ share the same objects. The universal property of products induces a $\kat{V}$-functor $(T \times T')_\kat{V} \kat{X} \to T_\kat{V} \kat{X} \times T'_\kat{V} \kat{X}$, which on the corresponding $\kat{V}$-homs says that the inequality \[ (T\times {T'})_\kat{V}(\kat{X})((A',B'),(A,B)) \le (T_\kat{V} \times T'_\kat{V})(\kat{X})((A',B'),(A,B)) \] holds for each objects $A',A$ of $T_\kat{V} \kat{X}$ and $B', B$ of $T'_\kat{V} \kat{X}$. The converse inequality is shown below, using Lemma~\ref{lem:wpb} and the frame law: \begin{align*} & (T_\kat{V} \times T'_\kat{V})(\kat{X})((A',B'),(A,B)) = \\ &T_\kat{V}(\kat{X})(A',A) \wedge T'_\kat{V}(\kat{X})(B',B) = \\ &\bigvee\{ t \mid (A',A) \in \mathsf{Rel}_T(X_t))\} \wedge \bigvee\{ s \mid (B',B) \in \mathsf{Rel}_{T'}(X_s))\} = \\ &\bigvee\{ t \wedge s \mid (A',A) \in \mathsf{Rel}_T(X_t), (B',B) \in \mathsf{Rel}_{T'}(X_s)\} \leq \\ &\bigvee\{ t \wedge s \mid (A',A) \in \mathsf{Rel}_T(X_{t\wedge s}), (B',B) \in \mathsf{Rel}_{T'}(X_{t\wedge s})\} \leq \\ &\bigvee\{ r \mid (A',A) \in \mathsf{Rel}_T(X_r), (B',B) \in \mathsf{Rel}_{T'}(X_r)\} = \\ &\bigvee\{ r \mid ((A',B'),(A,B)) \in \mathsf{Rel}_{T\times T'}(X_r)\} = \\ &(T\times T')_\kat{V}(\kat{X})((A',B'),(A,B)) \end{align*} That is, $(T \times T')_\kat{V} = T_\kat{V} \times T'_\kat{V}$ holds for all functors $T,T'$ which preserve weak pullbacks. \end{proof} \section{Solving domain equations across different base categories} \label{sec:beh} \ \medskip\noindent In the previous section, we studied those functors on $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ that arise as left Kan extensions of functors on ${\mathsf{Set}}$, with Theorems~\ref{thm:char-Vcatification} and \ref{thm:char-discrete-arities} accounting for the situations depicted on the left and right below. \begin{equation*} \xymatrix{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\ar[r]^{\overline T} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \\ {\mathsf{Set}}\ar[r]^T\ar[u]^D & {\mathsf{Set}} \ar[u]_D } \quad\quad \quad\quad \quad\quad \xymatrix{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\ar[r]^{\overline T} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \\ {\mathsf{Set}}\ar[u]^D \ar[ur]_H & } \end{equation*} In this section, we study the relationship between solving recursive equations in ${\mathsf{Set}}$ and in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. In both cases, accounted for by Theorems~\ref{thm:beh-vcatification} and \ref{thm:beh-discrete-arities}, if we have a coalgebra $X\to TX$ or $X\to VHX$, then two states in $X$ are behaviourally equivalent (bisimilar) iff they are behaviourally equivalent in $DX\to \overline TDX$. Moreover, while in the first case the behavioural distance of two states is either $0$ or $\infty$ (the states are behaviourally equivalent or not), in the second case the enrichment typically gives rise to non-trivial behavioural distances. \medskip\noindent We first review in Section~\ref{sec:2-cat} what in relation to~\cite{street:monad} could be called the ``formal theory of endofunctors''. Relevant references for this section include \cite{dubuc,street:monad} in case of algebras for $\kat{V}$-enriched monads and \cite{HermidaJ98,lenisa-power-watanabe} in case of (co)algebras for ($\kat{V}$-enriched) endofunctors. \medskip\noindent Section~\ref{sec:beh-cat} carries out an analysis of categories of coalgebras for both liftings and extensions, in case of ``change-of-base'' for the quantale $\kat{V}$, extending results of \cite{calco2011:balan-kurz,lmcs:bkv}. As corollaries we get information about final coalgebras of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-endofunctors $\overline T$ as in the diagrams above. \subsection{A 2-categorical approach to coalgebras} \label{sec:2-cat} \ \bigskip\noindent In order to relate coalgebras for different type functors on different categories, it will be convenient to work in a (2-)category where the objects are types (endofunctors) and the arrows are certain natural transformations that allow to connect types on different base categories. For the basic notions of $2$-categories that we need in this section we refer to~\cite{kelly-street:review-2-cats}. \begin{defi}\label{def:endo} Let $\kat{C}$ be a 2-category. The category $\mathsf{Endo}(\kat{C})$ has objects $(X,T)$ given by 1-cells $T:X\to X$ in $\kat{C}$. An arrow $(F,\phi):(X,T)\to (X',T')$ is given by a 1-cell $F:X\to X'$ and a 2-cell $\phi:FT\to T'F$. A 2-cell $\sigma:(F,\phi)\to (F',\phi')$ is a 2-cell $\sigma:F\to F'$ such that $T'\sigma\circ\phi=\phi'\circ\sigma T$. $$ \xymatrix{ X \ar[r]^T \ar[d]_F & X \ar[d]^F \ar@{}[dl]|{\swarrow \phi} && X \ar`l[d]`[d]_{F'}[d] \ar[r]^T \ar[d]^F_{\sigma \swarrow} & X \ar[d]^F \ar@{}[dl]|{\swarrow \phi} \ar@{}[drr]|{=} && X \ar[r]^T \ar[d]_{F'} & X \ar`r[d]`[d]^{F}[d] \ar[d]_{F'}^{\swarrow \sigma} \ar@{}[dl]|{\swarrow \phi} \\ X' \ar[r]^{T'} & X' && X' \ar[r]^{T'} & X' && X' \ar[r]^{T'} & X' } $$ Notice that there is a canonical inclusion 2-functor $\mathsf{Incl}: \kat{C}\to\mathsf{Endo}(\kat{C})$ mapping each object $X$ to the pair $(X,{\mathsf{Id}})$. \end{defi} \medskip\noindent The above definition is set-up precisely for the next proposition, which allows us to say, in analogy to the formal theory of monads~\cite{street:monad}, that a general 2-category $\kat{C}$ admits coalgebras (for endo-1-cells) if the inclusion $\mathsf{Incl}: \kat{C}\to\mathsf{Endo}(\kat{C})$ has a right 2-adjoint.% \footnote{ Notice that changing the orientation of the 2-cells in Definition~\ref{def:endo} would lead to a different 2-category of endo-1-cells, such that the right 2-adjoint to the inclusion (if exists) produces instead the ``category of algebras'' for the endofunctor $T$.} % This result has been independently mentioned by several authors at different moments of time (see for example~\cite{lenisa-power-watanabe} for the 2-categorical story, but also the early~\cite{cockett-spencer} which deals with algebras instead of coalgebras): \begin{prop} Let $\kat{C}$ be a 2-category with inserters.% \footnote{ Inserters are dual to coinserters, see Example~\ref{exle:coinserter}.} Then the inclusion 2-functor $\mathsf{Incl}$ has a right 2-adjoint $\mathsf{Coalg}:\mathsf{Endo}(\kat{C})\to\kat{C}$, mapping an object $(X,T)$ to the inserter of ${\mathsf{Id}}:X \to X$ and $T$. \end{prop} \begin{proof} This follows from the correspondences \renewcommand{\arraystretch}{1.2} \begin{tabular}{l} (1-cell in $\kat{C}$) $\ol F : Z \to \mathsf{Coalg}(X,T)$ \\ \hline (2-cell in $\kat{C}$) $\phi: F \to TF$ (where $F : Z \to X$ is a 1-cell in $\kat{C}$) \\ \hline (1-cell in $\mathsf{Endo}(\kat{C})$) $(F,\phi) : (Z,{\mathsf{Id}})\to (X,T)$ \end{tabular} \smallskip\noindent In the above, the second bijection is the definition of 1-cells in $\mathsf{Endo}(\kat{C})$, while the first bijection is due to the fact that $\mathsf{Coalg}(X,T)$ is an inserter of ${\mathsf{Id}}$ and $T$ (in that order). More precisely, $$ \phi:F\to TF $$ is an ``inserter cone'', and therefore $\phi$ gives rise to $\ol F$ (that is, to the factorisation through the ``inserter cone''). \end{proof} \begin{rem} In case $\kat{C} $ is the 2-category of $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories, $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors, and $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformations, recall from~\cite[Sect.~6.1]{bird:PhD} how to compute the inserter between two $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors $F,G:{\mathbb{X}} \to {\mathbb{Y}}$: \begin{itemize} \item $\mathsf{Ins}(F,G)$ is the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category having as objects pairs $(X,\tau)$, where $X$ is an object of ${\mathbb{X}}$ and $\tau:{\mathbb{1}} \to {\mathbb{Y}}(FX,GX)$ is a $\kat{V}$-functor (which only picks an object, still denoted $\tau:FX\to GX$). The $\kat{V}$-category-hom $\mathsf{Ins}(F,G)((X, \tau),(X', \sigma))$ is the equaliser in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ of the parallel pair \[ \xymatrix@C=30pt@R=15pt{ & {\mathbb{1}}\otimes{\mathbb{X}}(X,X') \ar[0,1]^-{\sigma\otimes F_{X,X'}} & {\mathbb{Y}}(FX',GX')\otimes {\mathbb{Y}}(FX,FX') \ar[1,1]^{\phantom{MM}c_{FX,FX',GX'}} & \\ {\mathbb{X}}(X,X') \ar[-1,1]^{\cong} \ar[1,1]_{\cong} & & & {\mathbb{Y}}(FX,GX') \\ & {\mathbb{X}}(X,X')\otimes {\mathbb{1}} \ar[0,1]_-{G_{X,X'}\otimes\tau} & {\mathbb{Y}}(FX,GX')\otimes{\mathbb{Y}}(FX,GX) \ar[-1,1]_{\phantom{MM}c_{FX,GX,GX'}} & } \] Explicitly, the objects of $\mathsf{Ins}(F,G)((X, \tau),(X', \sigma))$ are objects $h:X \to X'$ of the $\kat{V}$-category ${\mathbb{X}}(X,X')$ such that $\sigma \, \cdot \, Fh = Gh \, \cdot \, \tau$, with $\kat{V}$-distances \[ \mathsf{Ins}(F,G)((X, \tau),(X', \sigma))(h,k) = {\mathbb{X}}(X,X')(h,k) \] for $h,k:X\to X'$ as above. Composition and identity in $\mathsf{Ins}(F,G)((X, \tau),(X', \sigma))$ are induced from the composition in ${\mathbb{X}}(X,X')$. \item There is an obvious $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $J: \mathsf{Ins}(F,G) \to {\mathbb{X}}$ mapping $(X,\tau)$ to $X$, and a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $\iota:FJ \Rightarrow GJ$ with components $\iota_{(X,\tau)}:{\mathbb{1}} \to {\mathbb{Y}}(FJ(X,\tau),GJ(X,\tau))$ mapping the unique object $0$ to $\tau$. \end{itemize} \end{rem} \medskip\noindent In particular, the above yields the description of the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category $\mathsf{Coalg}(T)$ as $\mathsf{Ins}({\mathsf{Id}}, T)$, for any $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $T:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. More in detail: \begin{itemize} \item Objects are $T$-coalgebras, that is, pairs $(\kat{X},c)$ where $\kat{X}$ is a $\kat{V}$-category and $c:\kat{X} \to T\kat{X}$ is a $\kat{V}$-functor. \item The $\kat{V}$-category $\mathsf{Coalg}(T)((\kat{X},c),(\kat{Y},d))$ is the equaliser of \[ \xymatrix@C=65pt{% \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X},\kat{Y}) \ar@<+.5ex>[r]^-{\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X},d)} \ar@<-.5ex>[r]_{\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(c, T\kat{Y}) \circ T_{\kat{X},\kat{Y}}} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(\kat{X},T\kat{Y}) } \] that is, it has as objects $\kat{V}$-functors $f:\kat{X} \to \kat{Y}$ such that $d \circ f = T(f) \circ c$, with $\kat{V}$-distances between two such $\kat{V}$-functors $f,g: \kat{X} \to \kat{Y}$ being given by $[\kat{X},\kat{Y}](f,g) = \bigwedge_x \kat{Y}(f(x),g(x))$. \item There is a forgetful $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $\mathsf{Coalg}(T) \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ sending a $T$-coalgebra $(\kat{X},c)$ to the underlying $\kat{V}$-category $\kat{X}$. \end{itemize} \medskip\noindent One of the reasons to introduce the formal category $\mathsf{Endo}(\kat{C})$ is that it allows the following characterisation of when adjunctions in $\kat{C}$ can be lifted. \begin{prop}\label{prop:doctrinaladjunction} $(L,\phi)\dashv (R,\psi)$ in the 2-category $\mathsf{Endo}(\kat{C})$ if and only if $L\dashv R$ holds in $\kat{C}$, $\psi$ is iso, and $\phi$ is the mate of $\psi^{-1}$ under the adjunction $L\dashv R$. \end{prop} \begin{proof} By standard doctrinal adjunction~\cite{kelly:doctrinal}. \end{proof} \noindent Since $\mathsf{Coalg}$ is a 2-functor, it preserves adjunction of 1-cells, thus we obtain a well-known corollary allowing us to lift an adjunction $L\dashv R$ between base objects to an adjunction between ``categories of coalgebras'': \begin{cor}\label{cor:lift-adj} Let $\kat{C}$ be a 2-category with inserters, and let $L\dashv R:X' \to X$ an adjunction in $\kat{C}$. For any $T:X\to X$, $T':X'\to X'$, such that there is an iso-2-cell $\psi:RT' \to TR$, it follows that: \begin{enumerate} \item $R$ lifts to $\tilde R=\mathsf{Coalg}(R,\psi):\mathsf{Coalg}(X',T')\to\mathsf{Coalg}(X,T)$ \item There is an adjunction $\tilde L \dashv\tilde R:\mathsf{Coalg}(X',T')\to\mathsf{Coalg}(X,T)$, where $\tilde L$ is the image under $\mathsf{Coalg}$ of the 1-cell in $\mathsf{Endo}(\kat{C})$ given by $L$ and the mate of the inverse of $\psi: RT'\to TR$. \end{enumerate} \end{cor} \noindent We end this section with observations on the existence and the computation of limits and colimits in $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories of coalgebras. Again, these results seem to be folklore, but being unable to find them explicitly in the literature, we provide them below. \begin{prop}\label{prop:vcatcocomplete} The $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category $\mathsf{Coalg}(T)$, for a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $T:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, is cocomplete, and has all (weighted) limits that $T$ preserves. \end{prop} \begin{proof} Let $W:\mathcal K^{\mathsf{op}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ and $\ol{F}:\mathcal K \to \mathsf{Coalg}(T)$ be $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functors, denoting the weight and the diagram of shape $\mathcal K$, a small $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-category. By the inserter's universal property, to give $\ol{F}:\mathcal K \to \mathsf{Coalg}(T)$ is the same as to give a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $F:\mathcal K \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, together with a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-natural transformation $\xi:F\to TF$. Then the colimit $W*F$ becomes a $T$-coalgebra with structure $\alpha:W*F \to T(W*F)$ being the $\kat{V}$-functor which corresponds to the identity on $W*F$ under the composite \begin{align*} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(W*F,W*F) & \cong && [\mathcal K^{\mathsf{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W,\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(F -, W*F)) \\ && \overset{T}{\to} && [\mathcal K^{\mathsf{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W,\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(TF -, T(W*F))) \\ && \overset{\xi}{\to} && [\mathcal K^{\mathsf{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W,\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(F -, T(W*F))) \\ && \cong && \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(W*F,T(W*F)) \end{align*} \noindent Now, let $(\kat{X},c:\kat{X} \to T\kat{X})$ be an arbitrary $T$-coalgebra and consider the diagram below: \[ \xymatrix@C=30pt@R=35pt{ \mathsf{Coalg}(T)((W*F,\alpha), (\kat{X},c)) \ar[d] \ar@{.>}[r] & [\mathcal K^{\mathsf{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W, \mathsf{Coalg}(T)(\ol{F}-, (\kat{X}, c))) \ar[d] \\ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(W*F, \kat{X}) \ar[r]^-{\cong} \ar@<-.65ex>[d] \ar@<+.65ex>[d] & [\mathcal K^{\mathsf{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W, \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(F-,\kat{X})) \ar@<-.65ex>[d] \ar@<+.65ex>[d] \\ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(W*F, T\kat{X}) \ar[r]^-{\cong} & [\mathcal K^{\mathsf{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W, \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}(F-,T\kat{X})) } \] where the bottom square commutes serially and both columns are equalizers -- the left one by construction of $\kat{V}$-category-homs in the inserter $\mathsf{Coalg}(T)$, and the right one because the representable $[\mathcal K^{\mathsf{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W,-)$ preserves limits, in particular equalizers. Consequently, there is a unique arrow (isomorphism) \[ \xymatrix{ \mathsf{Coalg}(T)((W*F,\alpha), (\kat{X},c))\ar[r]^-{\cong} & [\mathcal K^{\mathsf{op}},\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}](W, \mathsf{Coalg}(T)(\ol{F}-, (\kat{X}, c))) } \] exhibiting $(W*F, \alpha)$ as the $W$-weighted colimit of $\ol{F}$. \medskip\noindent A similar computation that we leave to the reader shows that $\mathsf{Coalg}(T)$ has all weighted limits that $T$ preserves. \end{proof} \subsection{Relating behaviours across different base categories} \label{sec:beh-cat} \ \bigskip\noindent We consider ``change of base'' across different $\kat{V}$. We therefore now write $D^\kat{V}:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ and $V^\kat{V}:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o\to{\mathsf{Set}}_o$ to distinguish the discrete and forgetful functors for different $\kat{V}$. We drop the superscript $(-)^\kat{V}$ in case $\kat{V}={\mathbb{2}}$. \medskip\noindent In the previous section, we have shown that every $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor $H:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ has a left Kan extension along $D^\kat{V}$, denoted $H^\sharp$. Now, each such functor induces an ordinary set-endofunctor simply by forgetting the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-structure $$ \xymatrix{{\mathsf{Set}}_o \ar[r]^-{H_o} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}_o \ar[r]^-{V^\kat{V}} & {\mathsf{Set}}_o } $$ Moreover, $H^\sharp$ is a lifting of $V^\kat{V} H_o$ according to Proposition~\ref{prop:hsharp-lifting}. \medskip\noindent In the special case when $H$ is $D^\kat{V} T$, the above composite gives back $T$ (formally, it is $T_o$), and $H^\sharp$ is $T_\kat{V}$, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of $T$. \medskip\noindent We will see how the corresponding behaviours are related. In particular, we will show at the end of the section that $H^\sharp$ and $V^\kat{V} H_o$, hence $T_\kat{V}$ and $T$, induce the same behavioural equivalence (bisimilarity), while $H^\sharp$ may additionally induce a behavioural pseudometric. \begin{rem}\label{rem:change-of-base} For each commutative quantale $\kat{V}$, the inclusion (quantale morphism) \linebreak $\mathsf{d}:{\mathbb{2}} \to \kat{V}$ given by $0\mapsto \bot, \ 1 \mapsto e$, has a right adjoint (as it preserves suprema), denoted $\mathsf{v}:\kat{V} \to {\mathbb{2}}$ , which maps an element $r$ of $\kat{V}$ to $1$ if $e\le r$, and to $0$ otherwise.\footnote{ Notice that $\mathsf{v}$ is only a \emph{lax} morphism of quantales, being right adjoint.} This induces as usual the \emph{change-of-base} adjunction (even a 2-adjunction, see~\cite{eilenberg-kelly}) $$ \xymatrix@C=45pt{ {\mathbb{2}} \ar@/^1.5ex/[r]^{\mathsf{d}} \ar@{}[r]|{\perp} & \kat{V} \ar@/^1.5ex/[l]^{\mathsf{v}} & \mapsto & {\mathsf{Preord}} \ar@/^1.5ex/[r]^{\mathsf{d}_*} \ar@{}[r]|{\perp} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \ar@/^1.5ex/[l]^{\mathsf{v}_*} } $$ Explicitly, the functor $\mathsf{d}_*$ maps a preordered set $X$ to the $\kat{V}$-category $\mathsf{d}_* X$ with same set of objects, and $\kat{V}$-homs given by $\mathsf{d}_* X(x',x)=e$ if $x'\leq x$, and $\bot$ otherwise. Its right adjoint transforms a $\kat{V}$-category $\kat{X}$ into the preorder $\mathsf{v}_*\kat{X}$ with same objects again, and order $x'\leq x$ iff $e \leq \kat{X}(x',x)$ holds. Hence $\mathsf{d}_* X$ is the free $\kat{V}$-category on the preorder $X$, while $\mathsf{v}_* \kat{X}$ is the underlying ordinary category (which happens to be a preorder, due to the simple nature of quantales) of the $\kat{V}$-category $\kat{X}$. \qed \medskip\noindent Note that $\mathsf{d}_*$ is both a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor and a ${\mathsf{Preord}}$-functor, while its right adjoint $\mathsf{v}_*$ (in fact, the whole adjunction $\mathsf{d}_*\dashv \mathsf{v}_*$) is only ${\mathsf{Preord}}$-enriched. \end{rem} \begin{rem}\label{rem:change-of-base-cd} If $\kat{V}$ is nontrivial and integral, the embedding $\mathsf{d}:{\mathbb{2}} \to \kat{V}$ has also a left adjoint $\mathsf{c}:\kat{V} \to {\mathbb{2}}$, given by $\mathsf{c}(r)=0$ iff $r=\bot$, otherwise $\mathsf{c}(r)=1$. Notice that $\mathsf{c}$ is only a colax morphism of quantales, in the sense that $\mathsf{c}(e)\leq 1$ (in fact, here we have equality) and $\mathsf{c}(r\otimes s) \leq \mathsf{c}(r) \land \mathsf{c}(s)$, for all $r,s$ in $\kat{V}$. \medskip\noindent We will in the sequel assume that $\mathsf{c}$ is actually a morphism of quantales. The reader can check that this boils down to the requirement that $r\otimes s =\bot$ in $\kat{V}$ implies $r=\bot$ or $s=\bot$. That is, the quantale has no {\em zero-divisors}. All our examples satisfy this assumption. \medskip\noindent If the quantale has no zero-divisors, $\mathsf{d}_*$ also has a left adjoint $\mathsf{c}_*$ mapping a $\kat{V}$-category $\kat{X}$ to the preorder $\mathsf{c}_* \kat{X}$ with same objects and $x'\leq x$ in ${\mathsf{c}_*\kat{X}}$ iff $\kat{X}(x',x)\neq \bot$. Moreover, the adjunction $\mathsf{c}_* \dashv \mathsf{d}_*$ is $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched: $$ \xymatrix@C=45pt{ {\mathbb{2}} \ar@/_1.5ex/[r]_{\mathsf{d}} \ar@{}[r]|{\perp} & \kat{V} \ar@/_1.5ex/[l]_{\mathsf{c}} & \mapsto & {\mathsf{Preord}} \ar@/_1.5ex/[r]_{\mathsf{d}_*} \ar@{}[r]|{\perp} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \ar@/_1.5ex/[l]_{\mathsf{c}_*} } $$ \end{rem} \medskip\noindent From the above remark we obtain the following propositions. \begin{prop} Let $\kat{V}$ be an arbitrary commutative quantale and let $\widehat T:{\mathsf{Preord}} \to {\mathsf{Preord}}$ be a locally monotone functor (that is, ${\mathsf{Preord}}$-enriched) and $\ol{T}:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ be a lifting of $\widehat T$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ (meaning that $\ol T$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor such that $\mathsf{v}_* \ol{T} \cong \widehat T \mathsf{v}_*$ holds). Then the locally monotone adjunction $\mathsf{d}_*\dashv \mathsf{v}_*$ lifts to a locally monotone adjunction $\widetilde \mathsf{d}_*\dashv \widetilde \mathsf{v}_*$ between the associated ${\mathsf{Preord}}$-categories of coalgebras. $$ \xymatrix@R=33pt{ & \mathsf{Coalg}(\widehat T) \ar[d] \ar@/^8pt/[rr]^{\widetilde \mathsf{d}_*} \ar@{}[rr]|\bot \ar@{<-}@/_8pt/[rr]_{\widetilde \mathsf{v}_*} && \mathsf{Coalg}(\ol T ) \ar[d] & \\ & {\mathsf{Preord}} \ar@/^8pt/[rr]^{\mathsf{d}_*} \ar@{}[rr]|\bot \ar@{<-}@/_8pt/[rr]_{\mathsf{v}_*} \POS!L(.7),\ar@(ul, dl)_{\widehat T} && \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \POS!R(.7),\ar@(ur, dr)^{\ol T} & } $$ \end{prop} \begin{prop} Assume that $\kat{V}$ is a non-trivial integral commutative quantale without zero divisors. Let again $\widehat T:{\mathsf{Preord}} \to {\mathsf{Preord}}$ be a locally monotone functor, but this time consider $\ol{T}:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ be an extension of $\widehat T$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ (meaning that $\ol T$ is a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-functor, such that $\ol{T} \mathsf{d}_* \cong \widehat T \mathsf{d}_*$ holds). Then the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-adjunction $\mathsf{c}_*\dashv \mathsf{d}_*$ lifts to a $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-adjunction $\widetilde \mathsf{c}_*\dashv \widetilde \mathsf{d}_*$ between the associated $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories of coalgebras. $$ \xymatrix@R=33pt{ & \mathsf{Coalg}(\widehat T) \ar[d] \ar@/_8pt/[rr]_{\widetilde \mathsf{d}_*} \ar@{}[rr]|\bot \ar@{<-}@/^8pt/[rr]^{\widetilde \mathsf{c}_*} && \mathsf{Coalg}(\ol{T}) \ar[d] & \\ & {\mathsf{Preord}} \ar@/_8pt/[rr]_{\mathsf{d}_*} \ar@{}[rr]|\bot \ar@{<-}@/^8pt/[rr]^{\mathsf{c}_*} \POS!L(.7),\ar@(ul, dl)_{\widehat T} && \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \POS!R(.7),\ar@(ur, dr)^{\ol T} & } $$ \end{prop} \noindent We return now to the discrete functor $D^\kat{V}:{\mathsf{Set}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. It is easy to see that it decomposes as $\xymatrix@C=17pt@1{{\mathsf{Set}} \ar[r]^-D & {\mathsf{Preord}} \ar[r]^-{\mathsf{d}_*} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}}$. Additionally, recall the following (see also Example~\ref{ex:extensions}\eqref{ex:unicity-of-extens}): \begin{enumerate} \item There are locally monotone functors $D:{\mathsf{Set}}\to{\mathsf{Preord}}$, $C:{\mathsf{Preord}}\to{\mathsf{Set}}$, where $D$ maps a set to its discrete preorder and $C$ maps a preorder to its set of connected components. \item There is a chain $C_o \dashv D_o \dashv V:{\mathsf{Preord}} \to {\mathsf{Set}}$ of ordinary adjunctions where $V$ is the underlying-set forgetful functor. \item The locally monotone adjunction $C \dashv D$ is $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched. \end{enumerate} \medskip\noindent The next two propositions from \cite{calco2011:balan-kurz} are similar to the two above, but connect ${\mathsf{Set}}$ with ${\mathsf{Preord}}$ instead of ${\mathsf{Preord}}$ with $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. \begin{prop} Let $T:{\mathsf{Set}}\to {\mathsf{Set}}$ and $\widehat T:{\mathsf{Preord}} \to {\mathsf{Preord}}$ an extension of $T$ (a locally monotone functor such that $DT \cong \widehat T D$). Then the locally monotone adjunction $C\dashv D$ lifts to a locally monotone adjunction $\widetilde C \dashv \widetilde D$ between the associated categories of coalgebras: $$ \xymatrix@R=33pt{ & \mathsf{Coalg}(T) \ar[d] \ar@/_8pt/[rr]_{\widetilde D} \ar@{}[rr]|\bot \ar@{<-}@/^8pt/[rr]^{\widetilde C} && \mathsf{Coalg}(\widehat T) \ar[d] & \\ & {\mathsf{Set}} \ar@/_8pt/[rr]_{D} \ar@{}[rr]|\bot \ar@{<-}@/^8pt/[rr]^{C} \POS!L(.7),\ar@(ul, dl)_{T} && {\mathsf{Preord}} \POS!R(.7),\ar@(ur, dr)^{\widehat T} & } $$ \end{prop} \noindent Consequently, $\widetilde D$ will preserve limits, in particular, the final coalgebra (if it exists). \begin{prop} Let $T:{\mathsf{Set}}\to {\mathsf{Set}}$ and $\widehat T:{\mathsf{Preord}} \to {\mathsf{Preord}}$ a lifting of $T$ (an ordinary functor such that $TV \cong V\widehat T$). Then the ordinary adjunction $D_o\dashv V$ lifts to an ordinary adjunction $\widetilde D_o \dashv \widetilde V$ between the associated categories of coalgebras. $$ \xymatrix@R=33pt{ \mathsf{Coalg}(T) \ar[d] \ar@/^8pt/[rr]^{\widetilde D_o} \ar@{}[rr]|\bot \ar@{<-}@/_8pt/[rr]_{\widetilde V} && \mathsf{Coalg}(\widehat T) \ar[d] \\ {\mathsf{Set}} \ar@/^8pt/[rr]^{D_o} \ar@{}[rr]|\bot \ar@{<-}@/_8pt/[rr]_{V} \POS!L(.7),\ar@(ul, dl)_{T} && {\mathsf{Preord}} \POS!R(.7),\ar@(ur, dr)^{\widehat T} \\ } $$ \end{prop} \noindent Consequently, $\widetilde V$ will preserve ordinary limits; in particular, the underlying set of a final $\widehat T$-coalgebra will be a final $T$-coalgebra. \medskip\noindent Let now $T:{\mathsf{Set}}\to {\mathsf{Set}}$ and denote by $T_{\mathbb{2}}$ is ${\mathbb{2}}\mbox{-}\mathsf{cat}$-ification, that is, its ${\mathsf{Preord}}$-ification~\cite{lmcs:bkv}. We plan to see how $T_{\mathbb{2}}$ and $T_\kat{V}$, the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of $T$, are related. We start by the following observation: \begin{prop}\label{prop:preordvcatdense} The embedding $\mathsf{d}_\ast:{\mathsf{Preord}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ is dense. \end{prop} \begin{proof} We have shown in Theorem~\ref{thm:dense-pres} that $D^\kat{V} = \mathsf{d}_* D$ is $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-dense. Using that $\mathsf{d}_\ast$ is fully faithful, it follows from~\cite[Theorem~5.13]{kelly:book} that both $D$ and $\mathsf{d}_*$ are $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-dense and that $\mathsf{d}_*=\mathsf{Lan}_D (D^\kat{V})$ holds. \end{proof} \begin{rem}\label{rem:two-step-vcatification} Let $T:{\mathsf{Set}}\to {\mathsf{Set}}$ be a set-functor and $T_{\mathbb{2}}$ its ${\mathsf{Preord}}$-ification as above. Then the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $T_\kat{V}$ of $T$ can be computed in two stages, as follows: \begin{align*} T_\kat{V} \ & = && \mathsf{Lan}_{D^\kat{V}} (D^\kat{V} T) \\ & = && \mathsf{Lan}_{(\mathsf{d}_* D)} (\mathsf{d}_* D T) = \mathsf{Lan}_{\mathsf{d}_*} (\mathsf{Lan}_D (\mathsf{d}_* DT)) & \mbox{\quad (by \cite[Theorem~4.47]{kelly:book},} \\ & && & \mbox{because $\mathsf{d}_\ast$ is fully faithful)} \\ & \cong && \mathsf{Lan}_{\mathsf{d}_*} (\mathsf{Lan}_D (\mathsf{d}_* T_{\mathbb{2}} D)) & \mbox{ (because $DT \cong T_{\mathbb{2}} D$)} \\ & \cong && \mathsf{Lan}_{\mathsf{d}_*} (\mathsf{d}_* T_{\mathbb{2}}) \end{align*} where the last isomorphism holds because $\mathsf{Lan}_D (\mathsf{d}_* T_{\mathbb{2}} D) \cong \mathsf{d}_* T_{\mathbb{2}}$. To verify the latter isomorphism, notice first that the ${\mathsf{Preord}}$-enriched left Kan extension $\mathsf{Lan}_D (\mathsf{d}_* T_{\mathbb{2}} D)$ is also the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-enriched left Kan extension of $\mathsf{d}_* T_{\mathbb{2}} D$ along $D$, by a change-of-base argument as in~\cite[Theorem~1.7.1]{verity}. Next, apply~\cite[Theorem~5.29]{kelly:book} to the composite ${\mathsf{Preord}}$-functor $\mathsf{d}_* T_{\mathbb{2}}$, using the density presentation of $D$ exhibited in~\cite{lmcs:bkv} and the fact that $\mathsf{d}_*$ is ${\mathsf{Preord}}$-left adjoint. \end{rem} \noindent The remark above says that the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification of an endofunctor $T$ of ${\mathsf{Set}}$ can be obtained by taking first the ${\mathsf{Preord}}$-ification $T_{\mathbb{2}}:{\mathsf{Preord}} \to {\mathsf{Preord}}$ and then the left Kan extension of $\mathsf{d}_\ast T_{\mathbb{2}}$ along $\mathsf{d}_*$, as in \begin{equation*} \xymatrix{ \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \ar[r]^{T_\kat{V}} & \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \\ {\mathsf{Preord}} \ar[r]^{T_{\mathbb{2}}}\ar[u]^{\mathsf{d}_\ast} & {\mathsf{Preord}}\ar[u]_{\mathsf{d}_\ast} \\ {\mathsf{Set}} \ar[r]^{T}\ar[u]^{D} & {\mathsf{Set}}\ar[u]_{D} } \end{equation*} \medskip\noindent Putting things together we now obtain \begin{thm}\label{thm:beh-vcatification} Let $\kat{V}$ be a non-trivial integral commutative quantale without zero divisors, and $T:{\mathsf{Set}} \to {\mathsf{Set}}$ an arbitrary endofunctor, with $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $T_\kat{V}:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. Then the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-adjunctions $C\dashv D:{\mathsf{Set}} \to {\mathsf{Preord}}$, $\mathsf{c}_* \dashv \mathsf{d}_*:{\mathsf{Preord}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ lift to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-adjunctions between the associated $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories of coalgebras: $$ \xymatrix@R=33pt{ & \mathsf{Coalg}(T) \ar[d] \ar@/_8pt/[rr]_{\widetilde D} \ar@{}[rr]|\bot \ar@{<-}@/^8pt/[rr]^{\widetilde C} && \mathsf{Coalg}(T_{\mathbb{2}}) \ar[d] \ar@/_8pt/[rr]_{\widetilde \mathsf{d}_*} \ar@{}[rr]|\bot \ar@{<-}@/^8pt/[rr]^{\widetilde \mathsf{c}_*} && \mathsf{Coalg}(T_\kat{V})\ar[d] & \\ & {\mathsf{Set}} \ar@/_8pt/[rr]_{D} \ar@{}[rr]|\bot \ar@{<-}@/^8pt/[rr]^{C} \POS!L(.7),\ar@(ul, dl)_{T} && {\mathsf{Preord}} \ar@/_8pt/[rr]_{\mathsf{d}_*} \ar@{}[rr]|\bot \ar@{<-}@/^8pt/[rr]^{\mathsf{c}_*} && \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \POS!R(.7),\ar@(ur, dr)^{T_\kat{V}} & } $$ \end{thm} \medskip\noindent Since the $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-ification $T_\kat{V}$ of an endofunctor $T$ on ${\mathsf{Set}}$ is supposed to be ``$T$ in the world of $\kat{V}$-categories'', the theorem above confirms the expectation that final $T_\kat{V}$-coalgebras have a discrete metric. In fact, we can say that the final $T$-coalgebra is the final $T_\kat{V}$-coalgebra, if we consider $\mathsf{Coalg}(T)$ as a full (enriched-reflective) subcategory of $\mathsf{Coalg}(T_\kat{V})$. \medskip\noindent The next theorem deals with a more general situation where the final metric-coalgebra is the final set-coalgebra with an additional metric. This includes in particular the case where $\ol T$ is $H^\sharp$ for some $H:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ with $V^\kat{V} H_o=T_o$. \begin{thm}\label{thm:beh-discrete-arities} Let $\kat{V}$ be a commutative quantale, $T:{\mathsf{Set}} \to {\mathsf{Set}}$ be an arbitrary endofunctor, $\widehat T:{\mathsf{Preord}} \to {\mathsf{Preord}}$ a lifting of $T$ to ${\mathsf{Preord}}$, and $\ol T:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ be a lifting of $\widehat T$ to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$. Then the ordinary adjunction $D_o\dashv V: {\mathsf{Preord}}\to{\mathsf{Set}}$ and the ${\mathsf{Preord}}$-adjunction $\mathsf{d}_* \dashv \mathsf{v}_*:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}\to{\mathsf{Preord}}$ lift to adjunctions between the associated $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-categories of coalgebras $$ \xymatrix@R=33pt{ & \mathsf{Coalg}(T) \ar[d] \ar@{<-}@/_8pt/[rr]_{\widetilde V} \ar@{}[rr]|\bot \ar@/^8pt/[rr]^{\widetilde D_o} && \mathsf{Coalg}(\widehat T) \ar[d] \ar@{<-}@/_8pt/[rr]_{\widetilde \mathsf{v}_*} \ar@{}[rr]|\bot \ar@/^8pt/[rr]^{\widetilde \mathsf{d}_*} && \mathsf{Coalg}(\overline T)\ar[d] & \\ & {\mathsf{Set}} \ar@{<-}@/_8pt/[rr]_{V} \ar@{}[rr]|\bot \ar@/^8pt/[rr]^{D_o} \POS!L(.7),\ar@(ul, dl)_{T} && {\mathsf{Preord}} \ar@{<-}@/_8pt/[rr]_{\mathsf{v}_*} \ar@{}[rr]|\bot \ar@/^8pt/[rr]^{\mathsf{d}_*} && \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \POS!R(.7),\ar@(ur, dr)^{\ol T} & } $$ \end{thm} \noindent It follows that $\overline T$ and $T$ induce the same notion of behavioural equivalence (bisimilarity). Nevertheless, the final $\overline T$-coalgebra can provide additional information about order and metric of non-bisimilar elements. \begin{exa}\label{ex:beh-metric-stream} Recall from Example~\ref{ex:functors-with-V-metric}\eqref{ex:machine} the functor $TX=X^A \times B$ and its lifting $H^\sharp \kat{X} = \kat{X}^A \otimes \kat{B}$. Assume that the quantale is integral. Then the final $H^\sharp$-coalgebra is the power $\kat{V}$-category $\kat{B}^{A^*}$ having as objects all functions mapping each finite sequence of inputs to the last observable output in $\kat{B}$, with $\kat{V}$-distances \[ \kat{B}^{A^*}(h,k) = \bigwedge_{\mathsf l\in A^*} \kat{B}(h(\mathsf l),k(\mathsf l)) \] for each pair of behaviour functions $h,k:A^* \to B$. \medskip\noindent In particular, for the lifting $H^\sharp$ of the stream functor from Example~\ref{ex:functors-with-V-metric}\eqref{ex:stream} we obtain the final $H^\sharp$-coalgebra as the $\kat{V}$-category $\kat{B}^{\mathbb N} $ of streams over $B$. The lifting $H^\sharp$ of the deterministic automata functor from Example~\ref{ex:functors-with-V-metric}\eqref{ex:det-autom} given by $H^\sharp \kat{X} = \kat{X}^A \otimes {\mathbb{2}}_{r,s}$ has now as final coalgebra the ``generalised metric space'' ${\mathbb{2}}_{r,s}^{A^*}$ of languages over the alphabet $A$. \end{exa} \section{Conclusion} \medskip\noindent This paper is part of a larger endeavour extending set-based coalgebra to $\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$-based coalgebra, see for example \cite{VelebilK11,KurzV13,bkpv:rel-lift,lmcs:bkv,KurzV17,BabusK16,dahlqvist-kurz:calco17}. Here, we showed that every functor $H:{\mathsf{Set}}\to\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$ has a left-Kan extension $H^\sharp:\V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}} \to \V\hspace{-1.5pt}\mbox{-}\hspace{.9pt}{\mathsf{cat}}$, and that the final $H^\sharp$-coalgebra is the corresponding final coalgebra over ${\mathsf{Set}}$ equipped with a $\kat{V}$-metric. \medskip\noindent There are several directions in which to expand our results. For example, it would be interesting to move from commutative to general quantales and to quantaloids~\cite{rosenthal, stubbe:quant}. There is also the question whether one can extend not only functors but also monads in a uniform way, which could be related to (metric) trace (bi)simulation as in~\cite{bbkk:calco2015-ext}. \medskip\noindent Coalgebraically, it would be interesting to further develop the topic, barely touched upon in the last section, of behavioural pseudo-metrics~\cite{approx-bisim-prob,dgjp:metric-markov-proc,rutten:cmcs98,worrell:cmcs00}, while on the logical side, we aim to combine this paper with \cite{KurzV17,BabusK16} in the pursuit of an equational approach to quantitative reasoning about coalgebras/transition systems, an objective related to recent work on quantitative algebraic reasoning~\cite{mpp:quant-alg-reason}. \section*{Acknowledgements} \medskip\noindent We thank the anonymous referees of CALCO and LMCS for their valuable comments and patience that allowed us to improve the presentation of our results.

c78b8ad3f3eef7cf97157158313d7b6c12bf4e88

\section{Introduction} Let $\A$ be a locally finite arrangement of affine hyperplanes in $\R^n$. The complement $M(\A) \subseteq \mathbb{C}^n$ of the complexified arrangement $\A_\mathbb{C}$ is a well studied topological space. As proved by Salvetti \cite{salvetti1987topology,salvetti1994homotopy}, $M(\A)$ has the homotopy type of an $n$-dimensional CW complex. This complex is usually called the \emph{Salvetti complex} of $\A$, and we denote it by $\sal(\A)$. For a finite arrangement $\A$, in \cite{randell2002morse,dimca2003hypersurface,yoshinaga2007hyperplane} it was proved that the complement $M(\A)$ has the homotopy type of a minimal CW complex, i.e.\ with a number of $k$-cells equal to the $k$-th Betti number. This minimality result was later made more explicit with discrete Morse theory, in \cite{salvetti2007combinatorial} (for finite affine arrangements), \cite{delucchi2008shelling} (for finite central arrangements and oriented matroids in general), \cite{gaiffi2009morse} (for finite line arrangements), \cite{dantonio2015minimality} (for affine arrangements with a finite number of directions). In this work we consider a (possibly infinite) affine arrangement $\A$, and construct a minimal CW model for the complement $M(\A)$. This is obtained applying discrete Morse theory to the Salvetti complex of $\A$. For a (possibly infinite) CW complex, by ``minimal'' we mean that all the incidence numbers vanish. As in the well known case of finite arrangements, we obtain a geometrically meaningful bijection between cells in the minimal CW model and chambers of $\A$. Our starting point is the work of Delucchi on the minimality of oriented matroids \cite{delucchi2008shelling}. Specifically, we build on the idea of decomposing the Salvetti complex according to some ``good'' total order of the chambers. For a general affine arrangement, however, the combinatorial order used in \cite{delucchi2008shelling} does not yield a decomposition with the desired properties. In Section \ref{sec:decomposition} we introduce a class of total orders of the chambers for which we are able to extend the construction of Delucchi, and we call them \emph{valid orders}. We remark that in \cite[Question 4.18]{delucchi2008shelling} it was explicitly asked for one such extension to affine arrangements. For a finite affine arrangement, the polar order of Salvetti and Settepanella \cite{salvetti2007combinatorial} is valid (Remark \ref{rmk:polar-order}). Therefore our work contributes to linking the constructions of \cite{salvetti2007combinatorial} and \cite{delucchi2008shelling} (see also \cite[Remark 3.8]{delucchi2008shelling}). In Section \ref{sec:matching} we show how to construct an acyclic matching on $\sal(\A)$ for any given valid order. \begin{customthm}{\ref{thm:matching}} Let $\A$ be a locally finite hyperplane arrangement, with a given valid order of the set of chambers. Then there exists a proper acyclic matching on $\sal(\A)$ with critical cells in bijection with the chambers. \end{customthm} In the same section we also prove the following two results. The first one can be regarded as a generalization of \cite[Theorem 3.6]{delucchi2008shelling}, and the second one is the analogue of \cite[Theorem 3.6]{delucchi2008shelling} where (finite) oriented matroids are replaced by locally finite hyperplane arrangements. \begin{customthm}{\ref{teo:fiber}} Let $X$ be a $k$-dimensional polytope in $\R^k$, and let $y \in \R^k$ be a point outside $X$ that does not lie in the affine hull of any facet of $X$. Then there exists an acyclic matching on the poset of faces of $X$ visible from $y$, such that no face is critical. \end{customthm} \begin{customthm}{\ref{thm:matching-faces}} Let $\A$ be a locally finite hyperplane arrangement. For every chamber $C \in \C(\A)$, there is a proper acyclic matching on the poset of faces $\F(\A)$ such that $C$ is the only critical face. \end{customthm} In Section \ref{sec:euclidean} we construct valid orders for any locally finite arrangement $\A$, considering the Euclidean distance of the chambers from a fixed generic point $x_0 \in \R^n$. In this way, we obtain a family of matchings on $\sal(\A)$ that we call \emph{Euclidean matchings}. The idea of constructing a minimal complex that depends on a ``generic point'' appears to be new, as opposed to the more classical approach of using a ``generic flag'' \cite{yoshinaga2007hyperplane, salvetti2007combinatorial, gaiffi2009morse}. The critical cells are in bijection with the chambers, and can be described explicitly. \begin{customthm}{\ref{thm:euclidean-matching}} Let $\A$ be a locally finite arrangement in $\R^n$. For every generic point $x_0 \in \R^n$, there exists a Euclidean matching on $\sal(\A)$ with base point $x_0$. Such a matching has exactly one critical cell $\cell{C, F_C}$ for every chamber $C \in \C(\A)$, where $F_C$ is the smallest face of $C$ that contains the projection of $x_0$ onto $C$. \end{customthm} We prove that, also for infinite arrangements, the Morse complex of a Euclidean matching is minimal. \begin{customthm}{\ref{thm:minimality}} Let $\A$ be a locally finite hyperplane arrangement in $\R^n$, and let $\mathcal{M}$ be a Euclidean matching on $\sal(\A)$ with base point $x_0$. Then the associated Morse complex $\sal(\A)_{\mathcal{M}}$ is minimal (i.e.\ all the incidence numbers vanish). \end{customthm} In particular, we obtain a new geometric way to read the Betti numbers and the Poincaré polynomial of $M(\A)$ from the arrangement $\A$. \begin{customcor}{\ref{cor:betti-numbers}} Let $\A$ be a (locally) finite hyperplane arrangement in $\R^n$, and let $x_0 \in \R^n$ be a generic point. The $k$-th Betti number of the complement $M(\A)$ is equal to the number of chambers $C$ such that the projection of $x_0$ onto $C$ lies in the relative interior of a face $F_C$ of codimension $k$. Equivalently, the Poincaré polynomial of $\A$ is given by \[ \pi(\A,t) = \sum_{C \in \,\C(\A)} t^{\,\codim F_C}. \] \end{customcor} In Section \ref{sec:brieskorn} we use Euclidean matchings to obtain a proof of Brieskorn's Lemma (for locally finite complexified arrangements) which makes no use of algebraic geometry. In addition, we show that for every flat $X$ there exist Euclidean matchings on $\sal(\A)$ for which the Morse complex of the subarrangement $\A_X$ is naturally included into the Morse complex of $\A$. Finally, in Section \ref{sec:homology} we give an explicit description of the algebraic Morse complex that computes the homology of $M(\A)$ with coefficients in an abelian local system, for any locally finite line arrangement $\A$ in $\R^2$. We compare our result with the one of Gaiffi and Salvetti \cite{gaiffi2009morse}, where similar formulas are obtained in the case of finite line arrangements (using the polar matchings of Salvetti and Settepanella \cite{salvetti2007combinatorial}). \section{Background and notations} In this section we briefly recall some basic definitions and results about hyperplane arrangements, discrete Morse theory, polytopes, and shellability. \subsection{Hyperplane arrangements} \label{sec:hyperplane-arrangements} See \cite{orlik2013arrangements} for a general reference about hyperplane arrangements. Our notations mostly agree with those of \cite{salvetti2007combinatorial} and \cite{delucchi2008shelling}. Let $\A$ be a locally finite arrangement of affine hyperplanes in $\R^n$. Denote by $M(\A) \subseteq \mathbb{C}^n$ the complement of the complexified arrangement $\A_\mathbb{C}$. The arrangement $\A$ gives rise to a stratification of $\R^n$ into subspaces called \emph{faces} (see \cite[Chapter 5]{bourbaki1968elements}). It is more convenient for us to work with the closure of these subspaces, so we assume from now on that the faces are closed. By \emph{relative interior} of a face $F$ we mean the topological interior of $F$ inside the affine span of $F$. The faces of codimension $0$ are called \emph{chambers}. Denote the set of faces by $\F = \F(\A)$, and the set of the chambers by $\C = \C(\A)$. The set $\F$ has a natural partial order: $F \preceq G$ if and only if $F \supseteq G$. The poset $\F$ is called \emph{face poset} of $\A$, and it is ranked by codimension. Given two chambers $C,C' \in \C$, let $s(C,C') \subseteq \A$ be the set of hyperplanes which separate $C$ and $C'$. Also, denote by $\mathcal{W}_C \subseteq \A$ the set of hyperplanes that intersect $C$ in a face of codimension $1$. These hyperplanes are called \emph{walls} of $C$. For every chamber $C$, the set $\C$ can be endowed with a partial order $\leq_C$ defined as follows: $D' \leq_C D$ if and only if $s(C,D') \subseteq s(C,D)$. In the language of oriented matroids, $(\C, \leq_C)$ is called the \emph{tope poset based at $C$} \cite[Definition 4.2.9]{bjorner1999oriented}. Let $\L = \L(\A)$ be the poset of intersections of the hyperplanes in $\A$, ordered by reverse inclusion. An element $X \in \L$ is called a \emph{flat}. Notice that the entire space $\R^n$ is an element of $\L$ (being the intersection of zero hyperplanes), and it is in fact the unique minimal element of $\L$. The poset $\L$ is a geometric semilattice called \emph{poset of flats}, and it is also ranked by codimension. Denote by $\L_k(\A) \subseteq \L(\A)$ the set of flats of codimension $k$. For a subset $U \subseteq \R^n$ (usually a face or a flat), let $\supp(U) \subseteq \A$ be the subarrangement of $\A$ consisting of the hyperplanes that contain $U$. This is called the \emph{support} of $U$. Also, denote by $|U| \subseteq \R^n$ the affine span of $U$. Notice that, for a face $F \in \F$, we have $|F| \in \L$. Given a flat $X \in \L$, we also use the notation $\A_X$ to indicate the support of $X$ (this operation is called \emph{restriction}). Denote by $\A^X$ the arrangement in $X$ given by $\{ H \cap X \mid H \not\in \A_X \}$ (this operation is called \emph{contraction}). Let $\pi_X\colon \C(\A) \to \C(\A_X)$ be the natural projection, which maps a chamber $C \in \C(\A)$ to the unique chamber of $\A_X$ that contains $C$. For a chamber $C \in \C$ and a face $F \in \F$, denote by $C.F$ the unique chamber $C' \preceq F$ such that $\pi_{|F|}(C) = \pi_{|F|}(C')$. In other words, this is the unique chamber containing $F$ and lying in the same chamber as $C$ in $\A_{|F|}$. In addition, if $C \preceq F$, denote by $C^F$ the chamber opposite to $C$ with respect to $F$. The Salvetti complex of $\A$, first introduced in \cite{salvetti1987topology}, is a regular CW complex homotopy equivalent to the complement $M(\A)$ in $\mathbb{C}^n$ (see also \cite{gelfand1989algebraic, bjorner1992combinatorial, salvetti1994homotopy, orlik2013arrangements}). Its poset of cells $\sal(\A)$ is defined as follows. There is a $k$-cell $\cell{C,F}$ for each pair $(C,F)$ where $C \in \C$ is a chamber and $F \in \F$ is a face of $C$ of codimension $k$. A cell $\cell{C,F}$ is in the boundary of $\cell{D,G}$ if and only if $F \prec G$ and $D.F = C$. \begin{theorem}[\cite{salvetti1987topology}] The poset $\sal(\A)$ is the poset of cells of a regular CW complex homotopy equivalent to $M(\A)$. \end{theorem} \subsection{Discrete Morse theory} We recall here the main concepts of Forman's discrete Morse theory \cite{forman1998morse,forman2002user}. We follow the point of view of Chari \cite{chari2000discrete}, using acyclic matchings instead of discrete Morse functions, and we make use of the generality of \cite[Section 3]{batzies2002discrete} for the case of infinite CW complexes. Let $(P,<)$ be a ranked poset. If $q < p$ in $P$ and there is no element $r\in P$ with $q<r<p$, then we write $q \lessdot p$. Given $p \in P$ we define $P_{\leq p}= \{q \in P \mid q \leq p\}$. Let $G$ be the Hasse diagram of $P$, i.e.\ the graph with vertex set $P$ and having an edge $(p, q)$ whenever $q \lessdot p$. Denote by $\mathcal{E} = \{ (p,q) \in P\times P \mid q \lessdot p \}$ the set of edges of $G$. Given a subset $\mathcal{M}$ of $\mathcal{E}$, we can orient all edges of $G$ in the following way: an edge $(p,q) \in \mathcal{E}$ is oriented from $p$ to $q$ if the pair does not belong to $\mathcal{M}$, otherwise in the opposite direction. Denote this oriented graph by $G_{\mathcal{M}}$. \begin{definition}[Acyclic matching \cite{chari2000discrete}] A \emph{matching} on $P$ is a subset $\mathcal{M} \subseteq \mathcal{E}$ such that every element of $P$ appears in at most one edge of $\mathcal{M}$. A matching $\mathcal{M}$ is \emph{acyclic} if the graph $G_{\mathcal{M}}$ has no directed cycle. \end{definition} Given a matching $\mathcal{M}$ on $P$, an \emph{alternating path} is a directed path in $G_{\mathcal{M}}$ such that two consecutive edges of the path do not belong both to $\mathcal{M}$ or both to $\mathcal{E} \setminus \mathcal{M}$. The elements of $P$ that do not appear in any edge of $\mathcal{M}$ are called \emph{critical} (with respect to the matching $\mathcal{M}$). \begin{definition}[Grading \cite{batzies2002discrete}] Let $Q$ be a poset. A poset map $\varphi \colon P \to Q$ is called a \emph{$Q$-grading} of $P$. The $Q$-grading $\varphi$ is \emph{compact} if $\varphi^{-1}(Q_{\leq q}) \subseteq P$ is finite for all $q \in Q$. A matching $\mathcal{M}$ on $P$ is \emph{homogeneous} with respect to the $Q$-grading $\varphi$ if $\varphi(p) = \varphi(p')$ for all $(p,p') \in \mathcal{M}$. An acyclic matching $\mathcal{M}$ is \emph{proper} if it is homogeneous with respect to some compact grading. \end{definition} The following is a direct consequence of the definition of a proper matching (cf.\ \cite[Definition 3.2.5 and Remark 3.2.17]{batzies2002discrete}). \begin{lemma}[\cite{batzies2002discrete}]\label{lemma:alternating} Let $\mathcal{M}$ be a proper acyclic matching on a poset $P$, and let $p \in P$. Then there is a finite number of alternating paths starting from $p$, and each of them has a finite length. \end{lemma} We are ready to state the main theorem of discrete Morse theory. This particular formulation follows from \cite[Theorem 3.2.14 and Remark 3.2.17]{batzies2002discrete} \begin{theorem}[\cite{forman1998morse,chari2000discrete,batzies2002discrete}] Let $X$ be a regular CW complex, and let $P$ be its poset of cells. If $\mathcal{M}$ is a proper acyclic matching on $P$, then $X$ is homotopy equivalent to a CW complex $X_\mathcal{M}$ (called the \emph{Morse complex} of $\mathcal{M}$) with cells in dimension-preserving bijection with the critical cells of $X$. \end{theorem} The construction of the Morse complex is explicit in terms of the CW complex $X$ and the matching $\mathcal{M}$ (see for example \cite{batzies2002discrete}). This allows to obtain relations between the incidence numbers with $\Z$ coefficients in the Morse complex and incidence numbers in the starting complex. \begin{theorem}[{\cite[Theorem 3.4.2]{batzies2002discrete}}] \label{thm:incidence} Let $X$ be a regular CW complex, $P$ its poset of cells and $\mathcal{M}$ a proper acyclic matching on $P$. Let $X_\mathcal{M}$ be the Morse complex of $\mathcal{M}$. Given two critical cells $\sigma, \tau \in X$ with $\dim \sigma = \dim \tau +1$, denote by $\sigma_\mathcal{M}$ and $\tau_\mathcal{M}$ the corresponding cells in $X_\mathcal{M}$. Then the incidence number between $\sigma_\mathcal{M}$ and $\tau_\mathcal{M}$ in $X_\mathcal{M}$ is given by \begin{equation*} [\sigma_\mathcal{M} : \tau_\mathcal{M}]_{X_\mathcal{M}} \,= \sum_{\gamma \in \Gamma(\sigma,\tau)}{m(\gamma)}, \end{equation*} where $\Gamma(\sigma,\tau)$ is the set of all alternating paths between $\sigma$ and $\tau$. If $\gamma \in \Gamma(\sigma,\tau)$ is of the form \begin{equation*} \sigma=\sigma_0 \searrow \tau_1 \nearrow \sigma_1 \searrow \ldots \tau_k \nearrow \sigma_k \searrow \tau, \end{equation*} then $m(\gamma)$ is given by \begin{equation*} m(\gamma) \,=\, (-1)^{k}[\sigma_k : \tau] \, \prod_{i=1}^k{[\sigma_{i-1}:\tau_i][\sigma_i:\tau_i]}. \end{equation*} \end{theorem} Finally, recall the following standard tool to construct acyclic matchings. \begin{theorem}[Patchwork theorem {\cite[Theorem 11.10]{kozlov2007combinatorial}}]\label{teo:patchwork} Let $\varphi \colon P \to Q$ be a $Q$-grading of $P$. For all $q \in Q$, assume to have an acyclic matching $\mathcal{M}_q \subseteq \mathcal{E}$ that involves only elements of the subposet $\varphi^{-1}(q) \subseteq P$. Then the union of these matchings is itself an acyclic matching on $P$. \end{theorem} \subsection{Polyhedra, polytopes and shellability} \label{sec:shellability} In this section we briefly recall some notions and results from \cite{ziegler2012lectures}. \begin{definition} A \emph{polyhedron} is an intersection of finitely many closed halfspaces in some $\R^d$. A \emph{polytope} is a bounded polyhedron. \end{definition} Given a polyhedron $P$, denote by $\F(P)$ the set of its faces (considering the polyhedron $P$ itself as a trivial face). The faces of codimension $1$ are called \emph{facets}. In addition, denote by $\F(\partial P)$ the \emph{boundary complex} of $P$, i.e.\ the complex that contains only the proper faces of $P$. \begin{definition} We say that a facet $G \in \F(P)$ is \emph{visible} from a point $p \in \R^d$ if every line segment from $p$ to a point of $G$ does not intersect the interior of $P$ (cf.\ \cite[Theorem 8.12]{ziegler2012lectures}). We say that a face $F \in \F(P)$ is \emph{visible} from $p$ if all the facets $G \supseteq F$ of $P$ are visible from $p$. In particular, notice that the entire polyhedron $P$ is always visible. \end{definition} We are now able to recall the notion of shellability of the boundary complex of a polytope. \begin{definition}[{\cite[Definition 8.1]{ziegler2012lectures}}] A \emph{shelling} of the boundary complex of a polytope $P$ is a linear ordering $F_1,F_2,\ldots,F_s$ of the facets of $P$ such that either the facets are points, or the following conditions are satisfied. \begin{enumerate} \item The boundary complex $\F(\partial F_1)$ of the first facet has a shelling. \item For $1<j \leq s$, the intersection of the facet $F_j$ with the previous facets is nonempty and is a beginning segment of a shelling of $\F(\partial F_j)$, that is \begin{equation*} F_j \cap \left(\bigcup_{i=1}^{j-1}{F_i} \right)=G_1 \cup G_2 \cup\dots \cup G_r \end{equation*} for some shelling $G_1,G_2,\ldots,G_r,\ldots,G_t$ of $F_j$, with $1 \leq r \leq t$. A facet $F_j$ is called a \emph{spanning facet} if $r=t$. \end{enumerate} A polytope is \emph{shellable} if its boundary complex has a shelling. \end{definition} To conclude, recall the following two results about shellability of the boundary complex of a polytope. \begin{lemma}[{\cite[Lemma 8.10]{ziegler2012lectures}}] \label{lemma:reverse-shelling} If $F_1,F_2,\ldots,F_s$ is a shelling order for the boundary of a polytope $P$, then so is the reverse order $F_s,F_{s-1},\ldots,F_1$. \end{lemma} \begin{theorem}[\cite{bruggesser1972shellable}, {\cite[Theorem 8.12]{ziegler2012lectures}}] \label{thm:line-shelling} Let $P \subseteq \R^d$ be a $d$-polytope, and let $x \in \R^d$ be a point outside $P$. If $x$ lies in general position (that is, not in the affine hull of a facet of $P$), then the boundary complex of the polytope has a shelling in which the facets of $P$ that are visible from $x$ come first. \end{theorem} \section{Decomposition of the Salvetti complex} \label{sec:decomposition} Our aim is to construct an acyclic matching on the Salvetti complex of a locally finite affine arrangement $\A$, with critical cells in explicit bijection with the chambers of $\A$. Following the ideas of Delucchi \cite{delucchi2008shelling}, we want to decompose the Salvetti complex into ``pieces'' (one piece for every chamber) and construct an acyclic matching on each of these pieces with exactly one critical cell. More formally, we are going to decompose the poset of cells $\sal(\A)$ as a disjoint union \[ \sal(\A) = \bigsqcup_{C \in \C} N(C), \] so that every subposet $N(C) \subseteq \sal(\A)$ admits an acyclic matching with one critical cell. \begin{definition} Given a chamber $C \in \C$, let $S(C) \subseteq \sal(\A)$ be the set of all the cells $\cell{C',F} \in \sal(\A)$ such that $C' = C.F$. In other words, a cell is in $S(C)$ if all the hyperplanes in $\supp(F)$ do not separate $C$ and $C'$. \end{definition} Notice that the cells in $S(C)$ form a subcomplex of the Salvetti complex (using poset terminology, $S(C)$ is a lower ideal in $\sal(\A)$). This subcomplex is dual to the stratification of $\R^n$ induced by $\A$. Also, the natural map $S(C) \to \F$ which sends $\cell{C',F}$ to $F$ is a poset isomorphism. Now fix a total order $\dashv$ of the chambers: \[ \C = \{ C_0 \dashv C_1 \dashv C_2 \dashv \dots \} \] (when $\C$ is infinite, the order type is that of natural numbers). \begin{definition} For every chamber $C \in \C$, let $N(C) \subseteq S(C)$ be the subset consisting of all the cells not included in any $S(C')$ with $C' \dashv C$. \end{definition} The union of the subcomplexes $S(C)$, for $C \in \C$, is the entire complex $\sal(\A)$. Then the subsets $N(C)$, for $C \in \C$, form a partition of $\sal(\A)$. All the $0$-cells are contained in $N(C_0) = S(C_0)$. Therefore, for $C \neq C_0$, the cells of $N(C)$ do not form a subcomplex of the Salvetti complex. If $\A$ is a (finite) central arrangement, this definition of $N(C)$ coincides with the one given in \cite[Section 4]{delucchi2008shelling}. We want now to choose the total order $\dashv$ of the chambers so that each $N(C)$ admits an acyclic matching with exactly one critical cell. In \cite{delucchi2008shelling}, this is done taking any linear extension of the partial order $\leq_{C_0}$, for any base chamber $C_0$. Such a total order works well for central arrangements but not for general affine arrangements, as we see in the following two examples. \begin{example} Consider a non-central arrangement of three lines in the plane, as in Figure \ref{fig:two-arrangements} on the left. Choose $C_0$ to be one of the three simplicial unbounded chambers. In any linear extension of $\leq_{C_0}$, the last chamber $C_6$ must be the non-simplicial unbounded chamber opposite to $C_0$. However, $S(C_6) \subseteq \bigcup_{C \neq C_6} S(C)$, so $N(C_6)$ is empty, and therefore it does not admit an acyclic matching with one critical cell. Figure \ref{fig:non-valid} shows the decomposition of the Salvetti complex for one of the possible linear extensions of $\leq_{C_0}$. \label{example:non-valid} \end{example} \begin{example} Consider the arrangement of five lines depicted on the right of Figure \ref{fig:two-arrangements}. For every choice of a base chamber $C_0$ and for every linear extension of $\leq_{C_0}$, there is some chamber $C$ such that $N(C)$ is empty. \end{example} \newcommand{\arrangement}{ \draw (3.5,0) -- (3.5,4); \draw (0,1.25) -- (5,3.5); \draw (0,2.75) -- (5,0.5); } \newcommand{\chambers}{ \coordinate (C0) at (0.5,2) {}; \coordinate (C1) at (2,1) {}; \coordinate (C2) at (2,3) {}; \coordinate (C3) at (2.9,2) {}; \coordinate (C4) at (4.1,0.4) {}; \coordinate (C5) at (4.1,3.6) {}; \coordinate (C6) at (4.3,2) {}; } \begin{figure} \begin{tikzpicture} \arrangement \node (C0) at (0.5,2) {$C_0$}; \node (C1) at (2,1) {}; \node (C2) at (2,3) {}; \node (C3) at (2.9,2) {}; \node (C4) at (4.1,0.4) {}; \node (C5) at (4.1,3.6) {}; \node (C6) at (4.3,2) {$C_6$}; \end{tikzpicture} \qquad\qquad \begin{tikzpicture} \draw (4.1,0) -- (4.1,4); \draw (0,0.2) -- (5,2.7); \draw (0,3.8) -- (5,1.3); \draw (0.5,4) -- (4.5,0); \draw (0.5,0) -- (4.5,4); \end{tikzpicture} \caption{Two line arrangements.} \label{fig:two-arrangements} \end{figure} We are now going to state a condition on the total order $\dashv$ on $\C$ that produces a decomposition of the Salvetti complex with the desired properties. First recall the following definition from \cite{delucchi2008shelling}. \begin{definition} Given a chamber $C$ and a total order $\dashv$ on $\C$, let \[ \J(C) = \{ X \in \L \mid \supp(X) \cap s(C,C')\neq \emptyset \;\; \forall\, C' \dashv C \}. \] \end{definition} Notice that $\J(C)$ is an upper ideal of $\L$, and it coincides with $\L$ for $C=C_0$. In \cite[Theorem 4.15]{delucchi2008shelling} it is proved that, if $\A$ is a (finite) central arrangement and $\dashv$ is a linear extension of $\leq_{C_0}$ (for any choice of $C_0 \in \C$), then $\J(C)$ is a principal upper ideal for every chamber $C \in \C$. This is the condition we need. \begin{definition}[Valid order] A total order $\dashv$ on $\C$ is \emph{valid} if, for every chamber $C \in \C$, $\J(C)$ is a principal upper ideal generated by some flat $X_C = |F_C| \in \L$ where $F_C$ is a face of $C$. \label{def:valid-order} \end{definition} The total orders of Example \ref{example:non-valid} are not valid, because $\J(C_6)$ is empty. A valid order that begins with the chamber $C_0$ of Example \ref{example:non-valid} is showed in Figure \ref{fig:valid}. The previous definition is the starting point of our answer to \cite[Question 4.18]{delucchi2008shelling}, where it was asked for an extension of the arguments of \cite{delucchi2008shelling} to affine arrangements. Sections \ref{sec:matching} and \ref{sec:euclidean} will motivate this definition. \begin{remark} \label{rmk:polar-order} If $\A$ is a finite affine arrangement, the polar order of the chambers defined by Salvetti and Settepanella \cite[Definition 4.5]{salvetti2007combinatorial} is valid. Indeed, $\J(C)$ is a principal upper ideal generated by $X_C = |F_C|$, where $F_C$ is the smallest face of $C$ with respect to the polar order of the faces. Therefore Definition \ref{def:valid-order} highlights the link between the constructions of \cite{salvetti2007combinatorial} and \cite{delucchi2008shelling} (see also \cite[Remark 3.8]{delucchi2008shelling}). The results of Section \ref{sec:matching}, if applied to polar orders, give rise to acyclic matchings that are related to the polar matchings of \cite{salvetti2007combinatorial}. \end{remark} \tikzstyle{salvettivertices}=[every node/.style={circle,inner sep=2pt,fill=black}] \tikzstyle{salvettiedges}=[every path/.style={->,very thick}] \tikzstyle{XC}=[every path/.style={red, dashed, very thick}, every node/.style={red}] \begin{figure} \subfigure[Total order of the chambers]{ \begin{tikzpicture} \arrangement \node (C0) at (0.5,2) {$C_0$}; \node (C1) at (2,1) {$C_1$}; \node (C2) at (2,3) {$C_2$}; \node (C3) at (2.9,2) {$C_3$}; \node (C4) at (4.1,0.4) {$C_4$}; \node (C5) at (4.1,3.6) {$C_5$}; \node (C6) at (4.3,2) {$C_6$}; \end{tikzpicture} } \qquad\qquad \subfigure[$N(C_0)$]{ \begin{tikzpicture}[salvettivertices] \arrangement \chambers \node[circle,inner sep=2pt,fill=white] (c0) at (0.15,2) {$C_0$}; \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C0) -- (C1) -- (C3) -- (C2); \fill (C2) -- (C3) -- (C6) -- (C5); \fill (C1) -- (C3) -- (C6) -- (C4); \end{scope} \foreach \i in {0,...,6} \node (P\i) at (C\i) {}; \begin{scope}[salvettiedges] \draw (P0) -> (P1); \draw (P0) -> (P2); \draw (P1) -> (P3); \draw (P2) -> (P3); \draw (P1) -> (P4); \draw (P2) -> (P5); \draw (P3) -> (P6); \draw (P4) -> (P6); \draw (P5) -> (P6); \end{scope} \end{tikzpicture} } \bigskip \subfigure[$N(C_1)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C0) -- (C1) -- (C3) -- (C2); \fill (C2) -- (C3) -- (C6) -- (C5); \end{scope} \begin{scope}[salvettiedges] \draw (C1) -> (C0); \draw (C3) -> (C2); \draw (C6) -> (C5); \end{scope} \node (C1) at (2,0.7) {$C_1$}; \end{tikzpicture} } \qquad\qquad \subfigure[$N(C_2)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C0) -- (C1) -- (C3) -- (C2); \fill (C3) -- (C1) -- (C4) -- (C6); \end{scope} \begin{scope}[salvettiedges] \draw (C2) -> (C0); \draw (C3) -> (C1); \draw (C6) -> (C4); \end{scope} \node (C2) at (2,3.3) {$C_2$}; \end{tikzpicture} } \bigskip \subfigure[$N(C_3)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3,color=blue}] \fill (C0) -- (C1) -- (C3) -- (C2); \end{scope} \node (C3) at (3.15,2) {$C_3$}; \end{tikzpicture} } \qquad\qquad \subfigure[$N(C_4)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C5) -- (C2) -- (C3) -- (C6); \fill (C3) -- (C1) -- (C4) -- (C6); \end{scope} \begin{scope}[salvettiedges] \draw (C4) -> (C1); \draw (C6) -> (C3); \draw (C5) -> (C2); \end{scope} \node (C4) at (4.35,0.25) {$C_4$}; \end{tikzpicture} } \bigskip \subfigure[$N(C_5)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C5) -- (C2) -- (C3) -- (C6); \fill (C3) -- (C1) -- (C4) -- (C6); \end{scope} \node (C5) at (4.25,3.75) {$C_5$}; \draw[opacity=0.2] (C6) -- (C3); \end{tikzpicture} } \qquad\qquad \subfigure[$N(C_6)$ is empty]{ \begin{tikzpicture} \arrangement \chambers \node (C6) at (4.3,2) {$C_6$}; \end{tikzpicture} } \caption{A non-central arrangement of three lines in the plane, with a linear extension of $\leq_{C_0}$. Here $N(C_5)$ and $N(C_6)$ do not admit acyclic matchings with one critical cell.} \label{fig:non-valid} \end{figure} \begin{figure} \subfigure[Total order of the chambers]{ \begin{tikzpicture} \arrangement \node (C0) at (0.5,2) {$C_0$}; \node (C1) at (2,1) {$C_1$}; \node (C2) at (2,3) {$C_2$}; \node (C3) at (2.9,2) {$C_3$}; \node (C4) at (4.1,0.4) {$C_5$}; \node (C5) at (4.1,3.6) {$C_6$}; \node (C6) at (4.3,2) {$C_4$}; \end{tikzpicture} } \qquad\qquad \subfigure[$N(C_0)$]{ \begin{tikzpicture}[salvettivertices] \arrangement \chambers \node[circle,inner sep=2pt,fill=white] (c0) at (0.15,2) {$C_0$}; \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C0) -- (C1) -- (C3) -- (C2); \fill (C2) -- (C3) -- (C6) -- (C5); \fill (C1) -- (C3) -- (C6) -- (C4); \end{scope} \foreach \i in {0,...,6} \node (P\i) at (C\i) {}; \begin{scope}[salvettiedges] \draw (P0) -> (P1); \draw (P0) -> (P2); \draw (P1) -> (P3); \draw (P2) -> (P3); \draw (P1) -> (P4); \draw (P2) -> (P5); \draw (P3) -> (P6); \draw (P4) -> (P6); \draw (P5) -> (P6); \end{scope} \end{tikzpicture} } \bigskip \subfigure[$N(C_1)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C0) -- (C1) -- (C3) -- (C2); \fill (C2) -- (C3) -- (C6) -- (C5); \end{scope} \begin{scope}[salvettiedges] \draw (C1) -> (C0); \draw (C3) -> (C2); \draw (C6) -> (C5); \end{scope} \begin{scope}[XC] \draw (0,1.25) -- (5,3.5); \node at (0.5,1) {$X_{C_1}$}; \end{scope} \node (C1) at (2,0.7) {$C_1$}; \end{tikzpicture} } \qquad\qquad \subfigure[$N(C_2)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C0) -- (C1) -- (C3) -- (C2); \fill (C3) -- (C1) -- (C4) -- (C6); \end{scope} \begin{scope}[salvettiedges] \draw (C2) -> (C0); \draw (C3) -> (C1); \draw (C6) -> (C4); \end{scope} \begin{scope}[XC] \draw (0,2.75) -- (5,0.5); \node at (0.5,2.9) {$X_{C_2}$}; \end{scope} \node (C2) at (2,3.3) {$C_2$}; \end{tikzpicture} } \bigskip \subfigure[$N(C_3)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C0) -- (C1) -- (C3) -- (C2); \end{scope} \filldraw[red] (1.667,2) circle (2pt); \begin{scope}[XC] \node at (1.6,2.4) {$X_{C_3}$}; \end{scope} \node (C3) at (3.15,2) {$C_3$}; \end{tikzpicture} } \qquad\qquad \subfigure[$N(C_4)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C5) -- (C2) -- (C3) -- (C6); \fill (C3) -- (C1) -- (C4) -- (C6); \end{scope} \begin{scope}[salvettiedges] \draw (C4) -> (C1); \draw (C6) -> (C3); \draw (C5) -> (C2); \end{scope} \begin{scope}[XC] \draw (3.5,0) -- (3.5,4); \node at (3.1,3.7) {$X_{C_4}$}; \end{scope} \node (C4) at (4.6,2) {$C_4$}; \end{tikzpicture} } \bigskip \subfigure[$N(C_5)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C3) -- (C1) -- (C4) -- (C6); \end{scope} \filldraw[red] (3.5,1.175) circle (2pt); \begin{scope}[XC] \node at (3,1.1) {$X_{C_5}$}; \end{scope} \node (C5) at (4.35,0.25) {$C_5$}; \end{tikzpicture} } \qquad\qquad \subfigure[$N(C_6)$]{ \begin{tikzpicture} \arrangement \chambers \begin{scope}[every path/.style={opacity=0.3, color=blue}] \fill (C5) -- (C2) -- (C3) -- (C6); \end{scope} \filldraw[red] (3.5,2.825) circle (2pt); \begin{scope}[XC] \node at (3,2.95) {$X_{C_6}$}; \end{scope} \node (C6) at (4.3,3.75) {$C_6$}; \end{tikzpicture} } \caption{A non-central arrangement of three lines in the plane, with a valid order of the chambers. For every chamber $C$ except $C_0$, the generator $X_C$ of $\J(C)$ is highlighted.} \label{fig:valid} \end{figure} \section{Construction of the acyclic matching} \label{sec:matching} Throughout this section we assume to have an arrangement $\A$ together with a valid order $\dashv$ of $\C$ (as in Definition \ref{def:valid-order}). Using the decomposition \[ \sal(\A) = \bigsqcup_{C \in \C} N(C) \] of Section \ref{sec:decomposition} (induced by the valid order $\dashv$), we are going to construct a proper acyclic matching on $\sal(A)$ with critical cells in bijection with the chambers. More precisely, we are going to construct an acyclic matching on every $N(C)$ with exactly one critical cell, and then attach these matchings together using the Patchwork Theorem (Theorem \ref{teo:patchwork}). This strategy is the same as the one employed in \cite{delucchi2008shelling}, but our proofs are different since we deal with affine and possibly infinite arrangements. \begin{lemma} \label{lemma:NC} Suppose that $\dashv$ is a valid order of $\C$, in the sense of Definition \ref{def:valid-order}. Then \begin{equation*} N(C) = \{ \cell{D,F} \in S(C) \mid F \subseteq X_C\}. \end{equation*} \end{lemma} \begin{proof} To prove the inclusion $\subseteq$, assume by contradiction that there exists some cell $\cell{D,F} \in N(C)$ with $F \nsubseteq X_C$. By minimality of $X_C$ in $\J(C)$, we have that $|F| \notin \J(C)$. This means that there exists a chamber $C' \dashv C$ such that $\supp(F) \cap s(C,C') = \emptyset$. Then $C$ and $C'$ are contained in the same chamber of $\A_{|F|}$, which implies $C'.F=C.F$. By definition of $S(C)$, we have that $C.F=D$. Then $C'.F=D$, so $\cell{D,F} \in S(C')$. This is a contradiction, since $\cell{D,F} \in N(C)$ and $C' \dashv C$. For the opposite inclusion, consider a cell $\cell{D,F} \in S(C)$ with $F \subseteq X_C$. Then $|F| \in \J(C)$, so for every chamber $C' \dashv C$ there exists an hyperplane in $\supp(F) \cap s(C,C')$. By the same argument as before we can deduce that $D=C.F \neq C'.F$ for all $C' \dashv C$, which means that $\cell{D,F} \notin S(C')$ for all $C' \dashv C$. Therefore $\cell{D,F} \in N(C)$. \end{proof} Recall that, for a chamber $D \in \C$ and a face $F \succeq D$, we denote by $D^F$ the chamber opposite to $D$ with respect to $F$. For every chamber $C \in \C$, consider the map \begin{equation*} \tilde \eta_C \colon S(C) \to \mathcal{C} \end{equation*} that sends a cell $\cell{D,F}$ to $D^F$. We now show that $\eta_C$ is a poset map, if we endow $\C$ with the partial order $\leq_C$. \begin{lemma} The map $\tilde \eta_C\colon S(C) \to (\C, \leq_C)$ is order-preserving. \label{lemma:eta-order-preserving} \end{lemma} \begin{proof} Let $\cell{D,F}, \cell{D',F'} \in S(C)$, and suppose that $\cell{D',F'} \leq \cell{D,F}$ (see Figure \ref{Fig:lemma4.2}). Then $F' \preceq F$ and therefore $\supp(F') \subseteq \supp(F)$. Call $E=D^F$ and $E'=D'^{F'}$. By definition of $S(C)$, we have that $s(C,E)=s(C,D) \cup \supp(F)$ and $s(C,E')=s(C,D') \cup \supp(F')$. In addition, $F' \preceq F$ implies that $s(D,D') \subseteq \supp(F) \setminus \supp(F')$. Since $s(C,D') \subseteq s(C,D) \cup s(D,D')$, we conclude that \begin{align*} s(C,E') &= s(C,D') \cup \supp(F') \subseteq s(C,D) \cup s(D,D') \cup \supp(F') \\ &\subseteq s(C,D) \cup \supp(F) = s(C,E). \end{align*} Therefore $E' \leq_C E$. \end{proof} \begin{figure} \begin{tikzpicture} \draw (0,0) -- (10,5); \draw (0,4.5) -- (10,2); \draw (5,5) -- (7.5,0); \draw (2,0) -- (3,5); \node[circle,inner sep=2pt,fill=black] (f) at (6,3) {}; \node (F) at (5.9,2.65) {$F$}; \node (C) at (1,2.5) {$C$}; \node (D) at (3.7,2.7) {$D$}; \node (E) at (8.5,3.2) {$E$}; \node[text=blue] (D') at (4,4.5) {$D'$}; \node[text=blue] (F') at (5.8,4.1) {$F'$}; \node[text=blue] (E') at (7,4.5) {$E'$}; \draw[draw=blue, thick] (5,5) -- (6,3); \end{tikzpicture} \caption{Proof of Lemma \ref{lemma:eta-order-preserving}.} \label{Fig:lemma4.2} \end{figure} Consider the restriction $\eta_C = \tilde \eta_C |_{N(C)} \colon N(C) \to \C$. The matching on $N(C)$ will be obtained as a union of acyclic matchings on each fiber $\eta_C^{-1}(E)$ of $\eta_C$. Lemma \ref{lemma:eta-order-preserving}, together with the Patchwork Theorem, will ensure that the matching on $N(C)$ is acyclic. We now fix two chambers $C$ and $E$, and study the fiber $\eta_{C}^{-1}(E)$. \begin{lemma} Let $\dashv$ be a valid order of $\C$, and let $C,E$ be two chambers. A cell $\cell{D, F} \in \sal(\A)$ is in the fiber $\eta_C^{-1}(E)$ if and only if $D=E^F$, $F \subseteq X_C$ and $\supp(F) \subseteq s(C,E)$. \label{lemma:fiber} \end{lemma} \begin{proof} Suppose that $\cell{D,F} \in \eta_C^{-1}(E)$. In particular, $\cell{D,F} \in N(C)$, thus by Lemma \ref{lemma:NC} we have that $F \subseteq X_C$. By definition of $\eta_C$, $D^F=E$ and so $E^F=D$. Finally, we have $\supp(F) \subseteq s(D,E)$ by definition of $\eta_C$, and $\supp(F) \cap s(C,D) = \emptyset$ by definition of $S(C)$, so $\supp(F) \subseteq s(D,E) \setminus s(C,D) \subseteq s(C,E)$. We want now to prove that a cell $\cell{D,F}$ that satisfies the given conditions is in the fiber $\eta_C^{-1}(E)$. Since $D$ is opposite to $E$ with respect to $F$, we deduce that $\supp(F) \subseteq s(D,E)$. Then, using the hypothesis $\supp(F) \subseteq s(C,E)$, we obtain $\supp(F) \cap s(C,D) = \emptyset$. This means that $C.F = D$, i.e.\ $\cell{D,F} \in S(C)$. By Lemma \ref{lemma:NC}, we conclude that $\cell{D,F} \in N(C)$. The fact that $\eta_C(\cell{D,F}) = E$ follows directly from the definition of $\eta_C$. \end{proof} A cell $\cell{D,F}$ in the fiber $\eta_C^{-1}(E)$ is determined by $F$, because $D=E^F$. Then we immediately have the following corollary. \begin{provedcorollary}\label{corollary:visible-unrestricted} The fiber $\eta_{C}^{-1}(E)$ is in order-preserving (and rank-preserving) bijection with the set of faces $F \succeq E$ such that $F \subseteq X_C$ and $\supp(F) \subseteq s(C,E)$. In particular, if $\eta_C^{-1}(E)$ is non-empty, then $\supp(X_C) \subseteq s(C,E)$. \end{provedcorollary} Assume from now on that the fiber $\eta_C^{-1}(E)$ is non-empty. The above corollary can be restated as follows, restricting to the flat $X_C$. \begin{corollary}\label{corollary:visible} Suppose that the fiber $\eta_C^{-1}(E)$ is non-empty. Then $C' = C \cap X_C$ and $E' = E \cap X_C$ are chambers of the arrangement $\A^{X_C}$, and $\eta_{C}^{-1}(E)$ is in order-preserving bijection with the set of faces $F \succeq E'$ such that $\supp(F) \subseteq s(C',E')$ in $\A^{X_C}$. \end{corollary} \begin{proof} By Definition \ref{def:valid-order}, $X_C = |F_C|$ for some face $F_C$ of $C$. Then $C' = C \cap X_C = F_C$ is a chamber of $\A^{X_C}$. Consider now any cell $\cell{D,F} \in \eta_{C}^{-1}(E)$, and let $D'=D \cap X_C$. If we prove that $D'$ is a chamber of $\A^{X_C}$, then the same is true for $E'$, since they are opposite with respect to $F$ and $F \subseteq X_C$ (by Lemma \ref{lemma:NC}). Let $F'_C=F_C.F$ in the arrangement $\A^{X_C}$ (so $F'_C$ is a chamber of $\A^{X_C}$), and consider the chamber $\tilde D = C.F'_C$ in $\A$. Then $\tilde D = C.F = D$ (the first equality holds because $F'_C \preceq F$, and the second equality because $D \in S(C)$). Therefore $D' = D \cap X_C = \tilde D \cap X_C = F'_C$ is a chamber of $\A^{X_C}$. The second part is mostly a rewriting of Corollary \ref{corollary:visible-unrestricted}, but some care should be taken since we are passing from the arrangement $\A$ to the arrangement $\A^{X_C}$. To avoid confusion, in $\A^{X_C}$ write $\supp'$ and $s'$ in place of $\supp$ and $s$. Given a face $F \subseteq X_C$, we need to prove that $\supp(F) \subseteq s(C,E)$ in $\A$ if and only if $\supp'(F) \subseteq s'(C',E')$ in $\A^{X_C}$. This is true because \begin{align*} \supp'(F) &= \{ H \cap X_C \mid H \in \supp(F) \text{ and } H \nsupseteq X_C \}; \\ s'(C',E') &= \{ H \cap X_C \mid H \in s(C,E) \text{ and } H \nsupseteq X_C \}. \qedhere \end{align*} \end{proof} Constructing an acyclic matching on $\eta_C^{-1}(E)$ is then the same as constructing an acyclic matching on the set of faces of $E'$ given by Corollary \ref{corollary:visible}. We start by considering the special case $E'=C'$. \begin{lemma} \label{lemma:special-fiber} Suppose that the fiber $\eta_C^{-1}(E)$ is non-empty. Then $E' = C'$ if and only if $E$ is the chamber opposite to $C$ with respect to $X_C$. In this case, $\eta_C^{-1}(E)$ contains the single cell $\cell{C,F_C}$. \end{lemma} \begin{proof} If $E$ is opposite to $C$ with respect to $X_C$, then clearly $E' = C'$. Conversely, suppose that $E' = C' = F_C$. Let $\cell{D,F}$ be any cell in $\eta_C^{-1}(E)$. As in the proof of Corollary \ref{corollary:visible}, we have that $D \cap X_C = F_C'$ where $F_C' = F_C.F$ in $\A^{X_C}$. Notice that $F \subseteq E \cap X_C = E' = F_C$, so $F_C' = F_C.F = F_C$. In other words, the chambers $C$, $D$ and $E$ all contain the face $F_C$. Since $F \subseteq F_C \subseteq C \cap D$, we have that $s(C,D) \subseteq \supp(F)$. But $D \in S(C)$ implies that $D = C.F$, i.e.\ $s(C,D) \cap \supp(F) = \emptyset$. Therefore $s(C,D) = \emptyset$, so $C=D$. Now, $E$ is the opposite of $D$ with respect to $F$, and $E \cap X_C = D \cap X_C = F_C$, so $F = F_C$. This means that $E$ is the opposite of $C$ with respect to $X_C$. The previous argument also shows that $\eta_C^{-1}(E)$ contains the single cell $\cell{C,F_C}$. \end{proof} In particular, for every chamber $C$ there is exactly one fiber $\eta_C^{-1}(E)$ for which $E'=C'$. This fiber contains exactly one cell, which is going to be critical with respect to our matching. Consider now the case $E' \neq C'$. In view of Corollary \ref{corollary:visible}, we work with the restricted arrangement $\A^{X_C}$ in $X_C$. Until Lemma \ref{lemma:unb-bounded}, all our notations (for example, $\supp(F)$ and $s(C',E')$) are intended with respect to the arrangement $\A^{X_C}$. In what follows, we make use of the definitions and facts of Section \ref{sec:shellability}. \begin{lemma} \label{lemma:visible-faces} Let $y_{C'}$ be a point in the interior of $C'$. The faces $F \succeq E'$ such that $\supp(F) \subseteq s(C',E')$ are exactly the faces of $E'$ that are visible from $y_{C'}$. \end{lemma} \begin{proof} Suppose that $\supp(F) \subseteq s(C',E')$. In particular, for every facet $G \supseteq F$ of $E'$, the hyperplane $|G| \in \A^{X_C}$ separates $C'$ and $E'$ and so $G$ is visible from $y_{C'}$. Then $F$ is visible from $y_{C'}$. Conversely, suppose that $F$ is visible from $y_{C'}$. Denote by $\mathcal{B} \subseteq \supp(F)$ the set of hyperplanes $|G|$ where $G\supseteq F$ is a facet of $E'$. All the facets $G \supseteq F$ of $E'$ are visible from $y_{C'}$, so the hyperplanes $|G|$ separate $C'$ and $E'$. In other words, $\mathcal{B} \subseteq s(C',E')$. In the central arrangement $\A^{X_C}_{|F|} = \supp(F)$, the chambers $\pi_{|F|}(C')$ and $\pi_{|F|}(E')$ are therefore opposite to each other, and $\mathcal{B}$ is the set of their walls. Then every hyperplane in $\supp(F)$ separates $C'$ and $E'$. \end{proof} Fix an arbitrary point $y_{C'}$ in the interior of $C'$. By the previous lemma, the faces $F$ given by Corollary \ref{corollary:visible} are exactly the faces of $E'$ that are visible from $y_{C'}$. See Figure \ref{fig:visible} for an example. \begin{figure} \begin{tikzpicture} \draw[name path=L1] (0,0.9) -- (12,4.5); \draw[name path=L3] (0,5.2) -- (12,1.5); \draw[name path=L2] (4.55,0) -- (4.55,6); \draw[name path=L4] (0,2.4) -- (6,6); \draw[name path=L5] (0,3.6) -- (6,0); \path[name intersections={of=L1 and L2, by=E0}]; \path[name intersections={of=L3 and L2, by=E1}]; \path[name intersections={of=L1 and L5, by=E2}]; \path[name intersections={of=L3 and L4, by=E3}]; \path[name intersections={of=L4 and L5, by=E4}]; \node (E) at (3,3) {$E'$}; \node (C) at (10.8,3) {$C'$}; \node at (E2) {}; \node at (E3) {}; \node at (E4) {}; \begin{scope}[every node/.style={circle,inner sep=2.2pt,fill=darkgreen}] \node at (E0) {}; \node at (E1) {}; \end{scope} \node at (9.7,2.7) {$y_{C'}$}; \begin{scope}[every node/.style={circle,inner sep=1.8pt,fill=blue}] \node at (10,2.6) {}; \end{scope} \fill[opacity=0.15, color=darkgreen] (E0) -- (E1) -- (E3) -- (E4) -- (E2); \begin{scope}[every path/.style={draw=darkgreen,ultra thick, fill=darkgreen}] \draw (E0) -- (E1); \draw (E0) -- (E2); \draw (E1) -- (E3); \end{scope} \end{tikzpicture} \caption{The faces of $E'$ that are visible from a point $y_{C'}$ in the interior of $C'$.} \label{fig:visible} \end{figure} The idea now is that, if $E'$ is bounded, the boundary of $E'$ is shellable and we can use a shelling to construct an acyclic matching on the set of visible faces. We first need to reduce to the case of a bounded chamber (i.e.\ a polytope). \begin{lemma}\label{lemma:unb-bounded} There exists a finite set $\A'$ of hyperplanes in $X_C$, and a bounded chamber $\tilde E \subseteq E'$ of the hyperplane arrangement $\A' \cup \A^{X_C}$, such that the poset of faces of $\tilde{E}$ that are visible from $y_{C'}$ is isomorphic to the poset of faces of $E'$ that are visible from $y_{C'}$. \end{lemma} \begin{proof} Let $X_C \cong \R^k$. Let $Q$ be a finite set of points which contains $y_{C'}$ and a point in the relative interior of each visible face of $E'$. For $i=1,\dots, k$, define $q_i \in \R$ as the minimum of all the $i$-th coordinates of the points in $Q$, and $q^i$ as the maximum. Choose $\A'$ as the set of the $2k$ hyperplanes of the form $\{ x_i=q_i-1 \}$ and $\{ x_i=q^i+1 \}$, for $i=1,\dots, k$. Let $\tilde{E}$ be the chamber of $\A^{X_C} \cup \A'$ that contains $Q \setminus \{y_{C'}\}$. By construction, $\tilde{E}$ is bounded and is contained in $E'$. See Figure \ref{fig:Lemma} for an example. The walls of $E'$ and of $\tilde E$ are related as follows: $\mathcal{W}_{\tilde E} = \mathcal{W}_{E'} \cup \A''$ for some $\A'' \subseteq \A'$. The hyperplanes in $\mathcal{W}_{E'}$ separate $y_{C'}$ and $\tilde E$, whereas the hyperplanes in $\A''$ do not. This means that a facet $\tilde G$ of $\tilde E$ is visible if and only if $|\tilde G| \in \mathcal{W}_{E'}$. There is a natural order-preserving (and rank-preserving) injection $\varphi$ from the set $\mathcal{V}$ of the visible faces $F$ of $E'$ to the set of faces of $\tilde{E}$, which maps a face $F$ to the unique face $\tilde F$ of $\tilde E$ such that $F \cap Q \subseteq \tilde F \subseteq F$. We want to show that the image of $\varphi$ coincides with the set of visible faces of $\tilde E$. Consider a facet $\tilde G$ of $\tilde E$. Then $\tilde G$ is in the image of $\varphi$ if and only if $|\tilde G| \not\in \A''$, which happens if and only if $\tilde G$ is visible. Consider now a generic face $\tilde F$ of $\tilde E$. If $\tilde F = \varphi(F)$ for some $F \in \mathcal{V}$, then $Q \cap F \subseteq \tilde F$ and so $\tilde F$ is not contained in any hyperplane of $\A''$. Then all the facets $\tilde G \supseteq \tilde F$ of $\tilde E$ are visible, and so $\tilde F$ is visible. Conversely, if $\tilde F$ is not in the image of $\varphi$, then $\tilde F$ is contained in some hyperplane of $\A''$ and therefore also in some non-visible facet $\tilde G$. Then $\tilde F$ is not visible. \end{proof} \begin{figure} \begin{tikzpicture} \draw[name path=L1] (0,0.9) -- (12,4.5); \draw[name path=L3] (0,5.2) -- (12,1.5); \draw[name path=L2] (4.55,0) -- (5.55,6); \path[name intersections={of=L1 and L2, by=E0}]; \path[name intersections={of=L3 and L2, by=E1}]; \node (E') at (3,3) {$E'$}; \node (C') at (11.4,3) {$C'$}; \node at (9.7,2.7) {$y_{C'}$}; \begin{scope}[every path/.style={dashed, thick}] \draw (0.7,0) -- (0.7,6); \draw (10.7,0) -- (10.7,6); \draw (0,0.64) -- (12,0.64); \draw (0,5) -- (12,5); \end{scope} \begin{scope}[every path/.style={draw=darkgreen, ultra thick, fill=darkgreen}] \draw (E0) -- (E1); \draw (E0) -- (0,0.9); \draw (E1) -- (0,5.2); \end{scope} \fill[opacity=0.15, color=darkgreen] (E0) -- (E1) -- (0,5.2)-- (0,0.9); \begin{scope}[every node/.style={circle,inner sep=1.8pt,fill=blue}] \node at (E0) {}; \node at (E1) {}; \node at (5.05,3) {}; \node at (1.5,1.34) {}; \node at (3,4.27) {}; \node at (2,3.5) {}; \node at (10,2.6) {}; \end{scope} \end{tikzpicture} \caption{Construction of the bounded chamber $\tilde E \subseteq E'$ in the proof of Lemma \ref{lemma:unb-bounded}. The points of $Q$ are colored in blue, and the hyperplanes of $\A'$ are the dashed lines.} \label{fig:Lemma} \end{figure} We now show that the poset of visible faces of a polytope admits an acyclic matching such that no face is critical. We will use this result on the polytope $\tilde E$, in order to obtain a matching on the fiber $\eta_C^{-1}(E)$. \begin{theorem}\label{teo:fiber} Let $X$ be a $k$-dimensional polytope in $\R^k$, and let $y \in \R^k$ be a point outside $X$ that does not lie in the affine hull of any facet of $X$. Then there exists an acyclic matching on the poset of faces of $X$ visible from $y$ such that no face is critical. \end{theorem} \begin{proof} By \cite[Theorem 8.12]{ziegler2012lectures} and \cite[Lemma 8.10]{ziegler2012lectures}, there is a shelling $G_1,\dots,G_s$ of $\partial X$ such that the facets visible from $y$ are the last ones. Suppose that $G_t,G_{t+1},\dots, G_s$ are the visible facets. Notice that there is at least one visible facet and at least one non-visible facet. In particular, the first facet $G_1$ is not visible and the last facet $G_s$ is visible. In other words, we have $2\leq t \leq s$. In \cite[Proposition 1]{delucchi2008shelling} it is proved that a shelling of a regular CW complex $Y$ induces an acyclic matching on the poset of cells $(P, <)$ of $Y$ (augmented with the empty face $\emptyset$), with critical cells corresponding to the spanning facets of the shelling. In our case, $Y = \partial X$ is a regular CW decomposition of a sphere, so the only spanning facet of a shelling is the last one (see for example \cite[Lemma 2.13]{delucchi2008shelling}). Let $\mathcal{M}$ be an acyclic matching on $\partial X$ induced by the shelling $G_1,\dots,G_s$, as in \cite[Proposition 1]{delucchi2008shelling}. We claim that the construction of \cite{delucchi2008shelling} produces a matching which is homogeneous with respect to the grading $\varphi\colon (P, <) \to \{1,\dots,s\}$ given by \[ \varphi(F) = \min \{ i \in \{1,\dots,s\} \mid F \leq G_i \}. \] To prove this, we need to briefly go through the construction of $\mathcal{M}$. The first step \cite[Lemma 2.10]{delucchi2008shelling} is to construct a total order $\sqsubset_i$ on each $P_i$ (the set of faces of codimension $i$). The order $\sqsubset_0$ is simply the shelling order of the facets. It follows from the recursive construction of $\sqsubset_i$ that each $\varphi|_{P_i} \colon (P_i, \sqsubset_i) \to \{1,\dots,s\}$ is order-preserving. Then the linear extension $\vartriangleleft$ of $P$ constructed in \cite[Definition 2.11]{delucchi2008shelling} is such that $\varphi\colon (P,\vartriangleleft) \to \{1,\dots,s\}$ is also order-preserving. By construction of the matching \cite[Lemma 2.12]{delucchi2008shelling}, if $(p,q) \in \mathcal{M}$ (with $p \geq q$) then $p \vartriangleleft q$. From this we obtain $\varphi(p) \geq \varphi(q)$ and $\varphi(p) \leq \varphi(q)$, so $\varphi(p) = \varphi(q)$. Therefore the matching is homogeneous with respect to $\varphi$. The set of visible faces of $X$ is $\varphi^{-1}(\{t,\dots,s\}) \cup \{ X \}$. Notice that the empty face $\emptyset$ belongs to $\varphi^{-1}(1)$, so it does not appear in $\varphi^{-1}(\{t,\dots,s\})$ because $t \geq 2$. Let $\mathcal{M}'$ be the restriction of $\mathcal{M}$ to $\varphi^{-1}(\{t,\dots,s\})$. This is an acyclic matching on $\varphi^{-1}(\{t,\dots,s\})$ with exactly one critical face, the facet $G_s$. Then $\mathcal{M}' \cup \{(X, G_s) \}$ is an acyclic matching on the poset of visible faces of $X$ such that no face is critical. \end{proof} We are finally able to attach the matchings on the fibers $\eta_C^{-1}(E)$, using the previous results of this section. \begin{theorem} \label{thm:matching} Let $\A$ be a locally finite hyperplane arrangement, and let $\dashv$ be a valid order of the set of chambers $\C(\A)$. For every chamber $C \in \C(\A)$, there exists a proper acyclic matching on $N(C)$ such that the only critical cell is $\cell{C, F_C}$. The union of these matchings forms a proper acyclic matching on $\sal(\A)$ with critical cells in bijection with the chambers. \end{theorem} \begin{proof} Consider the map $\eta\colon \sal(\A) \to \C \times \C$ defined as \[ \eta(\cell{D,F}) = (C, D^F), \] where $C \in \C$ is the chamber such that $\cell{D,F} \in N(C)$. Corollary \ref{corollary:visible} provides a description of the non-empty fibers $\eta^{-1}(C,E)$, since by definition we have $\eta^{-1}(C,E)=\eta_C^{-1}(E)$. By Lemma \ref{lemma:special-fiber}, we know that for every $C \in \C$ there is exactly one non-empty fiber such that $E \cap X_C = C \cap X_C$, and this fiber contains the single cell $\cell{C,F_C}$. By Lemma \ref{lemma:visible-faces} and Lemma \ref{lemma:unb-bounded}, every other non-empty fiber $\eta^{-1}(C,E)$ is isomorphic to the poset of visible faces of some polytope in $X_C$ (with respect to some external point not lying on the affine hull of the facets). Finally, by Theorem \ref{teo:fiber}, this poset admits an acyclic matching with no critical faces. We want to use the Patchwork Theorem (Theorem \ref{teo:patchwork}) to attach these matchings together. To do so, we first need to define a partial order on $\C \times \C$ that makes $\eta$ a poset map. The order $\leq$ on $\C \times \C$ is the transitive closure of: \[ (C',E') \leq (C,E) \quad \text{if and only if} \quad C' \dashveq C \text{ and } E' \leq_C E \] (here we denote by $\dashveq$ the ``less than or equal to'' with respect to the total order $\dashv$). To prove that $\eta$ is a poset map, suppose to have $\cell{D',F'} \leq \cell{D,F}$ in $\sal(\A)$. Let $\eta(\cell{D',F'})=(C',E')$ and $\eta(\cell{D,F})=(C,E)$. Since $S(C)$ is a lower ideal of $\sal(\A)$, we immediately obtain that $\cell{D',F'} \in S(C)$ and thus $C' \dashveq C$. Then, Lemma \ref{lemma:eta-order-preserving} implies that $E' \leq_C E$. Therefore $(C',E') \leq (C,E)$. By the Patchwork Theorem, the union of the matchings on the fibers of $\eta$ forms an acyclic matching on $\sal(\A)$, with critical cells in bijection with the chambers. We now need to prove that this matching is proper. To do so, we prove that the $(\C \times \C)$-grading $\eta$ is compact. Since every fiber $\eta^{-1}(C,E)$ is finite by Lemma \ref{lemma:fiber}, we only need to show that the poset $(\C \times \C)_{\leq (C,E)}$ is finite for every pair of chambers $(C,E)$. We prove this by double induction, first on the chamber $C$ (with respect to the order $\dashv$) and then on $m=|s(C,E)|$. The base case $C=C_0$ and $m=0$ is trivial, since $E=C_0$. We want now to prove the induction step. Given a pair $(C,m) \in \C \times \N$, suppose that the claim is true for every pair $(C',m')$ such that either $C' \dashv C$, or $C'=C$ and $m'<m$. For every chamber $E$ with $|s(C,E)| = m$ we have that \begin{equation*} (\C \times \C)_{\leq (C,E)} \; = \!\!\! \bigcup_{ \substack{C' \dashveq C \\ E' \leq_C E \\ (C',E')\neq (C,E)} }{\!\!\!\!\!\!\! (\C \times \C)_{\leq (C',E')}} \;\;\cup\; \{(C,E)\} \end{equation*} This is a union of a finite number of sets, and by induction hypothesis every set $(\C \times \C)_{\leq (C',E')}$ is finite. Therefore the set $(\C \times \C)_{\leq (C,E)}$ is finite. By the Patchwork Theorem, the matchings on the fibers $\eta^{-1}(C,E)$ can be attached together to form a proper acyclic matching on $\sal(\A)$. By construction, this matching is a union of proper acyclic matchings on the subposets $N(C)$ for $C \in \C$, each of them having $\cell{C,F_C}$ as the only critical cell. \end{proof} We end this section with a few remarks. We are not going to use them in the rest of this paper, but they are interesting by themselves (especially in relation with \cite{delucchi2008shelling}). The first remark is that, without the need of a valid order, the results of this section allow to obtain a proper acyclic matching on $S(C_0)$ (for any chamber $C_0 \in \C$) with the single critical cell $\cell{C_0,C_0}$. This is because $N(C_0) = S(C_0)$, and in the construction of the matching on $N(C_0)$ we do not use the existence of a valid order that begins with $C_0$. As noted in Section \ref{sec:decomposition}, there is a natural poset isomorphism $S(C_0) \cong \F$ for every chamber $C_0 \in \C$. Then the existence of an acyclic matching on $S(C_0)$ can be stated purely in terms of $\F$, without speaking of the Salvetti complex. This result appears in \cite[Theorem 3.6]{delucchi2008shelling} in the case of the face poset of an oriented matroid. \begin{provedtheorem} \label{thm:matching-faces} Let $\A$ be a locally finite hyperplane arrangement. For every chamber $C \in \C(\A)$, there is a proper acyclic matching on the poset of faces $\F(\A)$ such that $C$ is the only critical face. \end{provedtheorem} The second remark is that, given a valid order $\dashv$ of $\C$ and a chamber $C \in \C$, the poset $N(C)$ is isomorphic to $\F(\A^{X_C})$. This is the analogue of \cite[Lemma 4.20]{delucchi2008shelling}. \begin{lemma} \label{lemma:NC-F} Suppose that $\dashv$ is a valid order of $\C$. For every chamber $C \in \C$ there is a poset isomorphism \[ N(C) \cong \F\big(\A^{X_C}\big). \] \end{lemma} \begin{proof} The isomorphism in the left-to-right direction sends a cell $\cell{D,F} \in N(C)$ to the face $F$, which is in $\F(\A^{X_C})$ by Lemma \ref{lemma:NC}. The inverse map sends a face $F \in \F(\A^{X_C})$ to the cell $\cell{C.F,F}$, which is in $N(C)$ by definition of $S(C)$ and by Lemma \ref{lemma:NC}. These maps are order-preserving. \end{proof} Together, Lemma \ref{lemma:NC-F} and Theorem \ref{thm:matching-faces} give an alternative (but equivalent) construction of our matching on $\sal(\A)$, closer to the approach of \cite{delucchi2008shelling}. \section{Euclidean matchings} \label{sec:euclidean} In this section we are going to construct a valid order $\eu$ of the set of chambers $\C$, for any locally finite arrangement $\A$, using the Euclidean distance in $\R^n$. Then we are going to prove that the matching induced by this order (given by Theorem \ref{thm:matching}) yields a minimal Morse complex. Denote by $d$ the Euclidean distance in $\R^n$. Also, if $K$ is a closed convex subset of $\R^n$, denote by $\rho_K(x)$ the projection of a point $x\in \R^n$ onto $K$. The point $\rho_K(x)$ is the unique point $y \in K$ such that $d(x,y) = d(x, K)$. The first step is to prove that there exist a lot of \emph{generic points} with respect to the arrangement $\A$. For this, we need the following technical lemma. By \emph{measure} we always mean the Lebesgue measure in $\R^n$. \renewcommand{\S}{\mathcal{S}} \begin{lemma}\label{lemma:measure} Let $K_1$ and $K_2$ be two closed convex subsets of $\R^n$. Let \[ \S = \{ x \in \R^n \mid d(x,K_1) = d(x,K_2) \text{ and } \rho_{K_1}(x) \neq \rho_{K_2}(x) \}. \] Then $\S$ has measure zero. \end{lemma} \begin{proof} This proof was suggested by Federico Glaudo. Let $d_i(x) = d(x,K_i)$ for $i=1,2$. Each function $d_i \colon \R^n \to \R$ is differentiable on $\R^n \setminus K_i$ by \cite[Lemma 2.19]{giaquinta2012convex}, and its gradient in a point $x \not\in K_i$ is the versor with direction $x-\rho_{K_i}(x)$. Let $f(x) = d_1(x) - d_2(x)$. Denote by $A$ the open set of points $x \in \R^n \setminus (K_1 \cup K_2)$ such that $\rho_{K_1}(x) \neq \rho_{K_2}(x)$. On this set, the function $f$ is differentiable and its gradient does not vanish. It is known that the gradient of $f$ must vanish almost everywhere on $A\cap f^{-1}(0)$ \cite[Corollary 1 of Section 3.1]{lawrence1992measure}, hence $A\cap f^{-1}(0)$ has measure zero. It is easy to check that the points in $K_1 \cup K_2$ cannot belong to $\S$. Then $\S = A \cap f^{-1}(0)$ has measure zero. \end{proof} \newcommand{\mathcal{G}}{\mathcal{G}} \newcommand{\mathcal{T}}{\mathcal{T}} \begin{lemma}[Generic points]\label{lemma:eucpt} Given a locally finite hyperplane arrangement $\A$ in $\R^n$, let $\mathcal{G} \subseteq \R^n$ be the set of points $x \in \R^n$ such that: \begin{enumerate}[(i)] \item for every $C,C' \in \mathcal{C}$ with $d(x,C)=d(x,C')$, we have $\rho_C(x) = \rho_{C'}(x) \in C \cap C'$; \item for every $L,L' \in \L$ with $L' \subsetneq L$, we have $d(x,L') > d(x,L)$. \end{enumerate} Then the complement of $\mathcal{G}$ has measure zero. In particular, $\mathcal{G}$ is dense in $\R^n$. \end{lemma} \begin{proof} Given $C,C' \in \C$, let $\S_{C,C'}$ be the set of points $x \in \R^n$ such that $d(x,C_1) = d(x,C_2)$ and $\rho_{C_1}(x) \neq \rho_{C_2}(x)$. By Lemma \ref{lemma:measure}, every $\S_{C,C'}$ has measure zero. Similarly, for every $L,L' \in \L$ with $L' \subsetneq L$, denote by $\mathcal{T}_{L,L'}$ the set of points $x \in \R^n$ such that $d(x, L') = d(x,L)$. We have that $\mathcal{T}_{L,L'}$ is an affine subspace of $\R^n$ of codimension at least $1$, and in particular it has measure zero. The complement of $\mathcal{G}$ is the union of all the sets $\S_{C,C'}$ for $C,C' \in \C$ and $\mathcal{T}_{L,L'}$ for $L,L' \in \L$ with $L' \subsetneq L$. This is a finite or countable union of sets of measure zero, hence it has measure zero. \end{proof} We call \emph{generic points} the elements of $\mathcal{G}$, as defined in Lemma \ref{lemma:eucpt}. Notice that, by condition (ii) with $L=\R^n$, a generic point must lie in the complement of $\A$. \begin{remark} \label{rmk:generic-point-equivalent} An equivalent definition of a generic point is the following: $x_0 \in \R^n$ is generic with respect to $\A$ if and only if every flat of $\A$ has a different distance from $x_0$. Indeed, this definition immediately implies condition (ii) of Lemma \ref{lemma:eucpt}. It also implies condition (i), because for any chamber $C$ we have $d(x_0, C) = d(x_0, L)$ where $L$ is the smallest flat that contains $\rho_C(x_0)$. Conversely, suppose that $x_0$ satisfies both conditions (i) and (ii). Given two flats $L, L' \in L$ with $d(x_0, L) = d(x_0, L')$, by condition (ii) the projections $\rho_L(x_0)$ and $\rho_{L'}(x_0)$ must lie in the relative interior of faces $F, F' \in \F$ with $|F| = L$ and $|F'| = L'$. Defining $C$ as the chamber containing $F$ and with the greatest distance from $x_0$, we immediately obtain that $\rho_L(x_0)=\rho_C(x_0)$. If $C'$ in defined in the same way (using $F'$ and $L'$), the chambers $C$ and $C'$ violate condition (ii) unless $L=L'$. With this equivalent definition, it is possible to prove Lemma \ref{lemma:eucpt} in an alternative way without using Lemma \ref{lemma:measure} (cf.\ Lemma \ref{lemma:add-generic-hyperplane}). \end{remark} We are now able to define Euclidean orders. \begin{definition}[Euclidean orders] A total order $\eu$ of the set of chambers $\C$ is \emph{Euclidean} if there exists a generic point $x_0$ such that $C \eu C'$ implies that $d(x_0,C)\leq d(x_0,C')$. The point $x_0$ is called a \emph{base point} of the Euclidean order $\eu$. \end{definition} In other words, a Euclidean order is a linear extension of the partial order on $\C$ given by $C < C'$ if $d(x_0, C) < d(x_0, C')$, for some fixed generic point $x_0 \in \R^n$. In particular, for every generic point $x_0$ there exists at least one Euclidean order with $x_0$ as a base point. Since the set of generic points is dense, we immediately get the following corollary. \begin{corollary} For every chamber $C_0 \in \C$, there exists a Euclidean order $\eu$ that starts with $C_0$. \end{corollary} \begin{proof} It is enough to take the base point $x_0$ in the interior of the chamber $C_0$. \end{proof} See Figure \ref{fig:eucl-order} for an example of a Euclidean order. We now prove that every Euclidean order is valid, in the sense of Definition \ref{def:valid-order}. \begin{theorem} \label{thm:euclidean-valid} Let $\eu$ be a Euclidean order with base point $x_0$. For every chamber $C$, let $x_C = \rho_C(x_0)$ and let $F_C$ be the smallest face of $C$ that contains $x_C$. Then $\J(C)$ is the principal upper ideal generated by $X_C = |F_C|$. Therefore $\eu$ is a valid order. \end{theorem} \begin{proof} First we want to prove that $X_C \in \J(C)$. This is equivalent to proving that for every chamber $C' \eu C$ there exists a hyperplane $H \in \supp(X_C)\cap s(C,C')$. We have that $\rho_{X_C}(x_0) = x_C$ because $F_C$ is the smallest face that contains $x_C$. Then it is also true that $\rho_{\pi_{X_C}(C)}(x_0)=x_C$. Given a chamber $C'\eu C$, we have two possibilities. \begin{itemize} \item $d(x_0,C')<d(x_0,C)$. Then $C' \nsubseteq \pi_{X_C}(C)$, because all the points of $\pi_{X_C}(C)$ have distance at least $d(x_0,C)$ from $x_0$. This means that there exists a hyperplane $H \in \supp(X_C)=\A_{X_C}$ which separates $C$ and $C'$. \item $d(x_0,C')=d(x_0,C)$. Since $x_0$ is a generic point, we have that $x_C = x_{C'} \in C \cap C'$. Then $F_C$ is a common face of $C$ and $C'$, and every hyperplane in $s(C,C')$ contains $F_C$. \end{itemize} Now we want to prove that $X \subseteq X_C$ for every $X \in \J(C)$. Suppose by contradiction that $X \nsubseteq X_C$ for some $X \in \J(C)$. In particular, $X_C \neq \R^n$ and thus $x_0 \neq x_C$. We first prove that $\supp(X_C \cup X)$ is non-empty. Let $C'$ be the chamber of $\A$ such that $x_0 \in \pi_{X_C}(C')$ and $C' \prec F_C$. Since $x_C \in X_C \subseteq \pi_{X_C}(C')$, the entire line segment $\ell$ from $x_0$ to $x_C$ is contained in $\pi_{X_C}(C')$. Then there is a neighbourhood of $x_C$ in $\ell$ which is contained in $C'$, hence $d(x_0, C') < d(x_0, x_C)$ and therefore $C' \eu C$. Since $X \in \J(C)$, there exists a hyperplane $H \in \supp(X) \cap s(C,C')$. We also have that $F_C \subseteq C \cap C'$, and thus $X_C \subseteq H$. Consider now the flat $X' = \cap \,\{Z \in \L \mid X_C \cup X \subseteq Z\}$, i.e.\ the meet of $X_C$ and $X$ in $\L$. The flat $X'$ is contained in the hyperplane $H$ constructed above, so in particular $X' \neq \R^n$. In addition, since $X \nsubseteq X_C$, $X'$ is different from $X_C$. Then the point $y_0=\rho_{X'}(x_0)$ is different from $x_C$, and we have $d(x_0, y_0) < d(x_0, x_C)$, because $x_0$ is generic (see condition (ii) of Lemma \ref{lemma:eucpt}). Let $F$ be the smallest face that contains the line segment $[x_C, \, x_C+ \epsilon (y_0-x_C)]$ for some $\epsilon>0$. By construction, for every chamber $C''$ such that $C'' \preceq F$ we have that $C'' \eu C$. This holds in particular for $C'' = C.F$. Then we have $\supp(F) \cap s(C,C'') = \emptyset$. Since $X \in \J(C)$ and $C'' \eu C$, there exists a hyperplane $H \in \supp(X) \cap s(C,C'')$. By construction, $x_C \in C \cap C''$ and then $X_C$ is contained in every hyperplane of $s(C,C'')$. In particular, $X_C \subseteq H$. Therefore $X_C \cup X \subseteq H$, which means that $H \in \supp(X_C \cup X) \subseteq \supp(X')$. Both $x_C$ and $y_0$ belong to $X'$, hence $F \subseteq X'$. Putting everything together, we get $H \in \supp(X') \cap s(C,C'') \subseteq \supp(F) \cap s(C,C'') = \emptyset$. This is a contradiction. \end{proof} Since Euclidean orders are valid, we are able to construct acyclic matchings on the Salvetti complex of any arrangement. \begin{definition}[Euclidean matchings] \label{def:euclidean-matching} Let $\A$ be a locally finite hyperplane arrangement in $\R^n$. We say that an acyclic matching $\mathcal{M}$ on $\sal(\A)$ is a \emph{Euclidean matching} with base point $x_0 \in \R^n$ if: \begin{enumerate}[(i)] \item the point $x_0$ is generic with respect to $\A$; \item $\mathcal{M}$ is homogeneous with respect to the poset map $\eta \colon \sal(\A) \to \C \times \C$ induced by a Euclidean order $\eu$ with base point $x_0$ (defined as in the proof of Theorem \ref{thm:matching}); \item there is exactly one critical cell $\cell{C, F_C}$ for every chamber $C \in \C$, where $F_C$ is the smallest face of $C$ that contains $\rho_C(x_0)$. \end{enumerate} Notice that, by condition (ii), a Euclidean matching is also proper. \end{definition} \begin{theorem} \label{thm:euclidean-matching} Let $\A$ be a locally finite arrangement in $\R^n$. For every generic point $x_0 \in \R^n$, there exists a Euclidean matching on $\sal(\A)$ with base point $x_0$. \end{theorem} \begin{proof} It follows from Theorems \ref{thm:matching} and \ref{thm:euclidean-valid}. \end{proof} \begin{remark} \label{rmk:matching-independent} For a given generic point $x_0$, there might be more than one Euclidean order $\eu$ with base point $x_0$. Nonetheless, all Euclidean orders with a given base point produce the same faces $F_C$ (by Theorem \ref{thm:euclidean-valid}) and the same critical cells (by Theorem \ref{thm:matching}). The decomposition \[ \sal(\A) = \bigsqcup_{C \in \C} N(C) \] also depends only on $x_0$ (by Lemma \ref{lemma:NC}), and therefore the definition of a Euclidean matching is not influenced by the choice of $\eu$ (once the base point $x_0$ is fixed). \end{remark} We are going to prove that a Euclidean matching yields a minimal Morse complex. In order to do so, we first prove two lemmas about generic points. \begin{lemma} \label{lemma:generic-point-subarrangement} Let $x_0 \in \R^n$. If $x_0$ is generic with respect to an arrangement $\A$, then it is also generic with respect to any subarrangement $\A' \subseteq \A$. \end{lemma} \begin{proof} Condition (i) for $\A'$ holds because a chamber of $\A'$ is a union of chambers of $\A$. Condition (ii) follows from the fact that $\L(\A') \subseteq \L(\A)$. \end{proof} \begin{lemma} \label{lemma:add-generic-hyperplane} Let $\A$ be a locally finite arrangement in $\R^n$, and let $x_0 \in \R^n$ be a generic point with respect to $\A$. Let $\H \subseteq (\R^n\setminus \{0\}) \times \R \subseteq \R^{n+1}$ be the set of elements $(a_1,\dots,a_n,b)$ such that $x_0$ is generic also with respect to the arrangement $\A \cup \{H\}$, where $H$ is the hyperplane defined by the equation $a_1x_1 + \dots + a_nx_n = b$. Then the complement of $\H$ in $\R^{n+1}$ has measure zero. In particular, $\H$ is dense in $\R^{n+1}$. \end{lemma} \begin{proof} In this proof we use the equivalent definition of a generic point given in Remark \ref{rmk:generic-point-equivalent}. Assume that $H$ intersects generically every flat $X \in \L(\A)$, i.e.\ $\codim(X \cap H) = \codim(X) + 1$. These conditions exclude a subset of measure zero in $\R^{n+1}$. Since $x_0$ is generic with respect to $\A$, the distances between $x_0$ and the flats of $\A$ are all distinct. Consider now a flat of $\A \cup \{H\}$ of the form $X \cap H$, for some flat $X \in \L(\A)$ of dimension $\geq 1$. The squared distance $d^2(x_0, X \cap H)$ is a rational function of the coefficients $(a_1,\dots,a_n,b)$ that define $H$. Given two flats $X,Y \in \L(\A)$ with $\dim(X) \geq 1$, the condition $d^2(x_0, X \cap H) = d^2(x_0, Y)$ can be written as a polynomial equation $p(a_1,\dots,a_n,b)=0$. This equation is not satisfied if $d(x_0, H) > d(x_0, Y)$, therefore the polynomial $p$ is not identically zero. Then the zero locus of $p$ has measure zero. Similarly, given two flats $X,Y \in \L(\A)$ with $\dim(X) \geq 1$ and $\dim(Y) \geq 1$, the condition $d^2(x_0, X \cap H) = d^2(x_0, Y \cap H)$ can be written as a polynomial equation $q(a_1,\dots,a_n,b) = 0$. Up to exchanging $X$ and $Y$, we can assume that $\rho_X(x_0) \not \in Y$, because $d(x_0,X) \neq d(x_0, Y)$. If $H$ is the hyperplane orthogonal to the vector $\rho_X(x_0) - x_0$ that passes through $\rho_X(x_0)$, then we have $d(x_0, X\cap H) = d(x_0, X)$ and $d(x_0, Y \cap H) > d(x_0, H) = d(x_0, X)$ (the inequality is strict because $Y$ does not contain $\rho_H(x_0) = \rho_X(x_0)$). Therefore the polynomial $q$ is not identically zero, and the zero locus of $q$ has measure zero. Then the complement of $\H$ is contained in a finite or countable union of sets of measure zero, thus it has measure zero. \end{proof} \begin{theorem}[Minimality] \label{thm:minimality} Let $\A$ be a locally finite hyperplane arrangement in $\R^n$, and let $\mathcal{M}$ be a Euclidean matching on $\sal(\A)$ with base point $x_0$. Then the associated Morse complex $\sal(\A)_{\mathcal{M}}$ is minimal (i.e., all the incidence numbers va\-nish). \end{theorem} \begin{proof} If the arrangement $\A$ is finite, it is well know that the sum of the Betti numbers of $\sal(\A)$ is equal to the number of chambers \cite{orlik1980combinatorics, zaslavsky1997facing}. By Theorem \ref{thm:matching}, the critical cells of $\mathcal{M}$ are in bijection with the chambers. Then the Morse complex is minimal. Suppose from now on that $\A$ is infinite. Fix a chamber $C \in \C$, and consider the associated critical cell $\cell{C,F_C} \in N(C)$. Recall from the proof of Theorem \ref{thm:matching} the definition of the poset map $\eta \colon \sal(\A) \to \C \times \C$, and let $(C,E)=\eta(\cell{C,F_C})$. Since the matching is proper, the set $\eta^{-1}((\C \times \C)_{\leq (C,E)})$ is finite. Consider now the finite set of faces \[ \mathcal{U} = \; \{ F \in \F \mid \cell{D,F} \in \eta^{-1}((\C \times \C)_{\leq (C,E)}) \text{ for some chamber } D \in \C \}. \] Let $B \subseteq \R^n$ be an open Euclidean ball centered in $x_0$ that contains the projection $\rho_F(x_0)$ for every face $F \in \mathcal{U}$. Let $\bar \A$ be a set of $n+1$ hyperplanes that do not intersect $B$, such that: $x_0$ is still generic with respect to $\A \cup \bar \A$; the chamber $K$ of the arrangement $\bar \A$ containing $B$ is bounded. Such an arrangement $\bar \A$ exists thanks to Lemma \ref{lemma:add-generic-hyperplane}. Consider the finite arrangement \[ \A' = \{ H \in \A \mid H \cap K \neq \emptyset \} \cup \bar \A, \] and let $\F_K \subseteq \F(\A)$ be the set of faces of $\A$ that intersect the interior of $K$. Notice that, by construction, we have $\mathcal{U} \subseteq \F_K$. In addition, there is a natural order-preserving and rank-preserving injection $\varphi\colon \F_K \to \F(\A')$ given by $\varphi(F) = F \cap K$. The image of $\varphi$ consists of the faces of $\A'$ that intersect the interior of $K$. By construction and Lemma \ref{lemma:generic-point-subarrangement} $x_0$ is still generic with respect to $\A'$ and all the chambers $D \in \C(\A)$ with $d(x_0, D) \leq d(x_0, C)$ intersect the interior of $K$. Then, given a Euclidean order $\eu$ of $\C(\A)$ with base point $x_0$, there exists a Euclidean order $\dashv'_{\text{eu}}$ of $\C(\A')$ with base point $x_0$ such that $\varphi$ is an order-preserving bijection between the initial segment of $(\C(\A), \eu)$ up to $C$ and the initial segment of $(\C(\A'), \dashv'_{\text{eu}})$ up to $\varphi(C)$. Consider the subcomplex $S = \eta^{-1}((\C\times\C)_{\leq (C,E)})$ of $\sal(\A)$. Since $\mathcal{U} \subseteq \F_K$, the map $\varphi$ induces an order-preserving and orientation-preserving injection $\psi \colon S \to \sal(\A')$ that maps a cell $\cell{D,G} \in S$ to the cell $\cell{\varphi(D),\varphi(G)} \in \sal(\A')$. Let $S' = \psi(S)$ be the copy of $S$ inside $\sal(\A')$. By definition of $S$, a fiber of $\eta$ is either disjoint from $S$ or entirely contained in $S$. Therefore, a non-critical cell of $S$ is matched with another cell of $S$. We now use the order $\dashv'_{\text{eu}}$ to construct a Euclidean matching $\mathcal{M}'$ on $\sal(\A')$. Denote by $\eta' \colon \sal(\A') \to \C(\A') \times \C(\A')$ the analogue of $\eta$ for the arrangement $\A'$ (see the proof of Theorem \ref{thm:matching}). Consider a fiber $\eta'^{-1}(C',E')$ that intersects $S'$. Then there is some cell $\cell{D',G'} \in \eta'^{-1}(C', E') \cap S'$, with $\cell{D',G'} = \psi(\cell{D,G})$ for some $\cell{D,G} \in S$. If we define $(\bar{C},\bar{E})=\eta(\cell{D,G})$, we have that $\varphi(\bar{C})=C'$ and $\varphi(\bar{E})=E'$, because by construction the cell $\cell{\varphi(D),\varphi(G)}$ is in the fiber $\eta^{-1}(\varphi(\bar{C}),\varphi(\bar{E}))$. By Corollary \ref{corollary:visible} and Lemma \ref{lemma:visible-faces}, the fiber $\eta'^{-1}(C',E')$ is isomorphic to the poset of faces of $E' \cap X_{C'}$ visible from some point $y_{C'}$ in the relative interior of $\C' \cap X_{C'}$. By construction of $\A'$, the map $\varphi$ induces a bijection between the faces of $E' \cap X_{C'}$ visible from $y_{C'}$ and the faces of $\bar E \cap X_{\bar C} = \bar E \cap X_{C'}$ visible from $y_{C'}$: if $F$ is a visible face of $\bar E \cap X_{C'}$, then $F \in \mathcal{U}$ and so $\varphi(F)$ is still visible; conversely, a visible face $F'$ of $E' \cap X_{\bar C}$ cannot be contained in any hyperplane of $\bar{\A}$, and by construction of $\A'$ it must also be a face of $\bar E$. Therefore the fiber $\eta'^{-1}(C',E')$ is the isomorphic image of the fiber $\eta^{-1}(\bar C, \bar E)$ under the map $\psi$. We have proved that a fiber of $\eta'$ is either disjoint from $S'$ or entirely contained in $S'$. Then we can choose the Euclidean matching $\mathcal{M}'$ so that its restriction to $S'$ coincides with the image of the restriction of $\mathcal{M}$ to $S$ under the isomorphism $\psi \colon S \to S'$. In particular, a cell $\cell{D,G} \in S$ is $\mathcal{M}$-critical if and only if $\psi(\cell{D,G}) \in S'$ is $\mathcal{M}'$-critical. Consider now a $\mathcal{M}$-critical cell $\cell{D,G} \in \sal(\A)$ such that there is at least one alternating path from $\cell{C,F_C}$ to $\cell{D,G}$. Since $\mathcal{M}$ is homogeneous with respect to $\eta$, every alternating path starting from $\cell{C,F}$ is entirely contained in $S$. In particular, $\cell{D,G} \in S$. Then the map $\psi\colon S \to S'$ induces a bijection between the alternating paths from $\cell{C,F}$ to $\cell{D,G}$ in $\sal(\A)$ (with respect to the matching $\mathcal{M}$) and the alternating paths from $\psi(\cell{C,F})$ to $\psi(\cell{D,G})$ in $\sal(\A')$ (with respect to the matching $\mathcal{M}'$). In particular, the incidence number between $\cell{C,F}$ and $\cell{D,G}$ in the Morse complex $\sal(\A)_\mathcal{M}$ is the same as the incidence number between $\psi(\cell{C,F})$ and $\psi(\cell{D,G})$ in the Morse complex $\sal(\A')_{\mathcal{M}'}$. Since $\A'$ is finite, the Morse complex $\sal(\A')_{\mathcal{M}'}$ is minimal and all its incidence numbers vanish. Therefore the incidence number between $\cell{C,F}$ and $\cell{D,G}$ in $\sal(\A)_\mathcal{M}$ also vanishes. \end{proof} \tikzfading[name=fade outside, inner color=transparent!0, outer color=transparent!100] \begin{figure} \begin{tikzpicture}[label distance=-3] \draw[name path=L3] (0,5) -- (12,0); \fill[white] (4.1,3.25) circle(0.55); \draw[name path=L1] (0,1) -- (12,6); \draw[name path=L2] (9,0) -- (10,6); \draw[name path=L4] (0,3) -- (12,3); \path[name intersections={of=L1 and L2, by=x6}]; \path[name intersections={of=L3 and L2, by=x7}]; \path[name intersections={of=L1 and L3, by=x8}]; \path[name intersections={of=L2 and L4, by=x4}]; \node[circle,inner sep=1.5pt, fill=black, label=left:{$x_0$}] (x0) at (8,3.5) {}; \node (C0) at (8.7,4) {$C_0=$\color{red}$\,F_0$}; \node (C1) at (8.5,2.2) {$C_1$}; \node (C2) at (6,5) {$C_2$}; \node (C3) at (11,4) {$C_3$}; \node (C4) at (11,1.8) {$C_4$}; \node (C5) at (6,1) {$C_5$}; \node (C6) at (10.8,5.9) {$C_6$}; \node (C7) at (10.6,0.2) {$C_7$}; \node (C8) at (1,3.8) {$C_8$}; \node (C9) at (1,2.2) {$C_9$}; \begin{scope}[every node/.style={circle,inner sep=1.5pt,fill=red}] \node at (x4) {}; \node at (x6) {}; \node at (x7) {}; \node at (x8) {}; \end{scope} \begin{scope}[every path/.style={draw=red, very thick, fill=red}] \draw (x6) -- (x8); \draw (x6)-- (x4); \draw (x7) -- (x8); \draw (x4) -- (x8); \end{scope} \begin{scope}[every node/.style={text=red}] \node (F1) at (7,2.75) {$F_1$}; \node (F2) at (7.1,4.23) {$F_2$}; \node (F3) at (9.9,4) {$F_3$}; \node (F4) at (9.7,2.7) {$F_4$}; \node (F5) at (7,1.8) {$F_5$}; \node (F6) at (10.13,5.48) {$F_6$}; \node (F7) at (9.35,0.85) {$F_7$}; \node (F8) at (4.17,3.25) {$F_8=F_9$}; \end{scope} \fill[opacity=0.13, color=red] (x6) -- (x4) -- (x8); \end{tikzpicture} \caption{Euclidean order with respect to $x_0$. The faces $F_i = F_{C_i}$ defined in Theorem \ref{thm:euclidean-valid} are highlighted in red.} \label{fig:eucl-order} \end{figure} The following result is a direct consequence of Theorems \ref{thm:euclidean-matching} and \ref{thm:minimality}. It gives a simple geometric way to compute the Betti numbers of the complement of an arrangement. \begin{provedcorollary}[Betti numbers] \label{cor:betti-numbers} Let $\A$ be a (locally) finite hyperplane arrangement in $\R^n$, and let $x_0 \in \R^n$ be a generic point. The $k$-th Betti number of the complement $M(\A)$ is equal to the number of chambers $C$ such that the projection $\rho_C(x_0)$ lies in the relative interior of a face $F_C$ of codimension $k$. Equivalently, the Poincaré polynomial of $\A$ is given by \[ \pi(\A,t) = \sum_{C \in \,\C(\A)} t^{\,\codim F_C}. \qedhere \] \end{provedcorollary} \begin{example} Consider the line arrangement $\A$ of Figure \ref{fig:eucl-order}. For the given generic point $x_0$ in the interior of $C_0$, the computation of the Betti numbers $b_i$ according to Corollary \ref{cor:betti-numbers} goes as follows: there is one chamber (namely $C_0$) such that the projection of $x_0$ lies in its interior, so $b_0 = 1$; there are four chambers (namely $C_1$, $C_2$, $C_3$ and $C_5$) such that the projection of $x_0$ lies in the interior of a $1$-dimensional face, so $b_1 = 4$; finally, for the remaining chambers ($C_4$, $C_6$, $C_7$, $C_8$ and $C_9$) the projection of $x_0$ is a $0$-dimensional face, so $b_2 = 5$. \end{example} \begin{remark} For any choice of the generic point $x_0$, the only chamber that contributes to the $0$-th Betti number is the one containing $x_0$. In addition, for every hyperplane $H \in \A$ there is exactly one chamber $C$ such that $\rho_C(x_0) \in H$ and $\rho_C(x_0) \not\in H'$ for every $H' \in \A \setminus \{H\}$. Therefore Corollary \ref{cor:betti-numbers} immediately implies the well-known facts that $b_0(\A) = 1$ and $b_1(\A) = |\A|$. \end{remark} \section{Brieskorn's Lemma and naturality} \label{sec:brieskorn} In this section we are going to relate the Morse complex of $\A$, constructed using a Euclidean matching, to the Morse complexes of subarrangements $\A_X$. Given a flat $X \in \L(\A)$, for every face $\bar F \in \F(\A)$ such that $|\bar F|=X$ there is a natural inclusion of $\sal(\A_X)$ into $\sal(\A)$. It maps a cell $\cell{D, G} \in \sal(\A_X)$ to the unique cell $\cell{C, F} \in \sal(\A)$ such that $\bar F \subseteq F \subseteq G$, $\dim F = \dim G$, and $C \subseteq D$. We call this the inclusion of $\sal(\A_X)$ into $\sal(\A)$ \emph{around} $\bar F$. Geometrically, this corresponds to including the complement of $\A_X^\mathbb{C}$, intersected with a neighbourhood of some point in the interior of $\bar F$, into $M(\A)$. The inclusions $\sal(\A_X) \hookrightarrow \sal(\A)$ that we are going to consider in this section are always of this type, for some face $\bar F \in \F(\A)$ with $|\bar F| = X$. Recall from Definition \ref{def:euclidean-matching} that a Euclidean matching has a critical cell $\cell{C, F_C} \in \sal(\A)$ for every chamber $C$, where $F_C$ is the smallest face of $C$ containing $\rho_C(x_0)$. Every critical cell $\cell{C, F_C}$ is thus associated to a flat $X_C = |F_C|$. Conversely, given a flat $X \in \L(\A)$, the critical cells $\cell{C, F_C}$ associated to $X$ are exactly those for which $\rho_C(x_0) = \rho_X(x_0)$. This simple observation yields a proof of Brieskorn's Lemma, a classical result in the theory of hyperplane arrangements due to Brieskorn \cite{brieskorn1973groupes}. See also \cite[Lemma 5.91]{orlik2013arrangements} and \cite[Proposition 3.3.3]{callegaro2017integer}. \begin{lemma}[Brieskorn's Lemma \cite{brieskorn1973groupes}] \label{lemma:brieskorn} Let $\A$ be a locally finite arrangement in $\R^n$. For every $k\geq 0$, there is an isomorphism \[ \theta_k \colon \bigoplus_{X \in \L_k} H_k(M(\A_X); \Z) \to H_k(M(\A); \Z) \] induced by suitable inclusions $j_X\colon \sal(\A_X) \hookrightarrow \sal(\A)$ of CW complexes. The inverse isomorphism $\theta_k^{-1}$ is induced by the natural inclusion maps $i_X \colon M(\A) \hookrightarrow M(\A_X)$. \end{lemma} \begin{proof} Let $x_0 \in \R^n$ be a generic point with respect to $\A$, and let $\mathcal{M}$ be a Euclidean matching on $\sal(\A)$ with base point $x_0$. Let $X \in \L_k$ be a flat of codimension $k$. By Lemma \ref{lemma:generic-point-subarrangement}, the point $x_0$ is generic also with respect to the subarrangement $\A_X$. Consider the inclusion $j_X \colon \sal(\A_X) \hookrightarrow \sal(\A)$ around the unique face of $\A$ containing the projection $\rho_X(x_0)$. Let $\mathcal{M}_X$ be a Euclidean matching on $\sal(\A_X)$ with base point $x_0$. All homology groups in this proof are with integer coefficients. By Theorem \ref{thm:minimality}, we have that $H_k(\sal(\A))$ is a free abelian group generated by elements of the form \[ [ \cell{C,F_C} + \text{a finite sum of non-critical $k$-cells}] \] for each critical $k$-cell $\cell{C,F_C}$ of $\sal(\A)$. Similarly, for every flat $X \in \L_k$, we have that $H_k(\sal(\A_X))$ is a free abelian group generated by elements of the same form as above, one for every critical $k$-cell of $\sal(\A_X)$. The critical $k$-cells of $\sal(\A_X)$ are in bijection (through the map $j_X$) with the critical $k$-cells $\cell{C,F}$ of $\sal(\A)$ such that $|F| = X$. Then the inclusions $j_X$ induce an isomorphism \[ \bar\theta_k \colon \bigoplus_{X \in \L_k} H_k(\sal(\A_X)) \to H_k(\sal(\A)). \] Let $\varphi\colon \sal(\A) \xhookrightarrow{\simeq} M(\A)$ and $\varphi_X\colon \sal(\A_X) \xhookrightarrow{\simeq} M(\A_X)$ be the homotopy equivalences constructed in \cite{salvetti1987topology}. Then the composition \[ \bigoplus_{X \in \L_k} H_k(M(\A_X)) \xrightarrow{\bigoplus (\varphi_X)_*^{-1}} \bigoplus_{X \in \L_k} H_k(\sal(\A_X)) \xrightarrow{\bar \theta_k} H_k(\sal(\A)) \xrightarrow{\varphi_*} H_k(M(\A)) \] is the isomorphism $\theta_k$ as in the statement. By naturality of Salvetti's construction, the following diagram is commutative up to homotopy. \begin{center} \begin{tikzcd} \sal(\A_X) \arrow[r, hook, "j_X"] \arrow[d, hook', "\varphi_X"] & \sal(\A) \arrow[d, hook', "\varphi"] \\ M(\A_X) & M(\A) \arrow[l, hook', "i_X"] \end{tikzcd} \end{center} Looking at the induced commutative diagram in homology, we obtain that the inverse isomorphism $\theta_k^{-1}$ is induced by the inclusion maps $i_X$. \end{proof} \begin{remark} With an idea similar to the one employed in the previous proof, it might be possible to derive a version of Brieskorn's Lemma for abstract oriented matroids. We leave this as a potential direction for future works. \end{remark} If we fix a flat $X \in \L(\A)$, it is possible to choose the base point $x_0$ so that the Morse complex of $\sal(\A_X)$ injects into the Morse complex of $\sal(\A)$. We prove this naturality property in the following lemma. \begin{lemma} Let $X \in \L(\A)$ be a flat, and fix an inclusion $j \colon \sal(\A_X) \hookrightarrow \sal(\A)$ around some face $\bar F$ with $|\bar F| = X$. There exist Euclidean matchings $\mathcal{M}_X$ and $\mathcal{M}$, on $\sal(\A_X)$ and $\sal(\A)$ respectively, such that: \begin{enumerate}[(i)] \item they share the same base point $x_0$; \item $j(\mathcal{M}_X) \subseteq \mathcal{M}$; \item the inclusion $j$ induces an inclusion of the Morse complex of $\sal(\A_X)$ into the Morse complex of $\sal(\A)$. \end{enumerate} \end{lemma} \begin{proof} Let $x_0 \in \R^n$ be a generic point such that $d(x_0, \bar F) < d(x_0, H)$ for every hyperplane $H \in \A \setminus \A_X$ (the existence of $x_0$ follows from Lemma \ref{lemma:eucpt}). For example, we can choose a point $y$ in the relative interior of $\bar F$, and then take $x_0$ in a small neighbourhood of $y$. Let $\eu$ and $\dashv'_{\text{eu}}$ be Euclidean orders with base point $x_0$ on $\C(\A)$ and $\C(\A_X)$, respectively. Notice that, by construction of $x_0$, the total order $\eu$ starts with the chambers containing $\bar F$. Let $\eta\colon \sal(\A) \to \C(\A) \times \C(\A)$ be the poset map defined in the proof of Theorem \ref{thm:matching}, induced by the total order $\eu$. Let $\eta'\colon \sal(\A_X) \to \C(\A_X) \times \C(\A_X)$ be the analogous poset map for the arrangement $\A_X$, induced by the total order $\dashv'_{\text{eu}}$. Then, for every pair of chambers $C,E \in \C(\A)$ containing $\bar F$, we have \[ \eta^{-1}(C,E) = j( {\eta'}^{-1} (\pi_X(C), \pi_X(E))). \] In other words, the inclusion $j$ maps fibers of $\eta'$ to fibers of $\eta$. Notice that, by Remark \ref{rmk:matching-independent}, these fibers only depend on $x_0$ and not on the particular choices of the Euclidean orders $\eu$ and $\dashv'_{\text{eu}}$. Let $\mathcal{M}_X$ be a Euclidean matching on $\sal(\A_X)$ with base point $x_0$. Recall that a Euclidean matching on $\sal(\A)$ with base point $x_0$ is constructed on the fibers of $\eta$ (see Definition \ref{def:euclidean-matching}). Then there exists a Euclidean matching $\mathcal{M}$ on $\sal(\A)$ with base point $x_0$ that contains $j(\mathcal{M}_X)$. The alternating paths in $\sal(\A)$ starting from cells in the subcomplex $j(\sal(\A_X))$ remain in this subcomplex. Therefore $j$ induces an inclusion of the Morse complex of $\sal(\A_X)$ (with respect to the matching $\mathcal{M}_X$) into the Morse complex of $\sal(\A)$ (with respect to the matching $\mathcal{M}$). \end{proof} \section{Local homology of line arrangements} \label{sec:homology} Let $\A$ be a locally finite line arrangement in $\R^2$. In this section we are going to describe the algebraic Morse complex (associated to a Euclidean matching) that computes the homology of the complement $M(\A)$ with coefficients in an abelian local system. Then we are going to compare the obtained complex with the algebraic complex of Gaiffi and Salvetti \cite{gaiffi2009morse}, which is based on the polar matching of Salvetti and Settepanella \cite{salvetti2007combinatorial}. An abelian local system $L$ on $M(\A)$ is determined by the elements $t_\ell \in \Aut(L)$ associated to elementary positive loops around every line $\ell \in \A$ (cf.\ \cite[Section 2.4]{gaiffi2009morse}). The boundaries $\partial_i$ of the algebraic Morse complex are determined by the incidence numbers $[\cell{D,G}, \cell{C,F}]^\mathcal{M} \in \Z[t_\ell^{\pm 1}]_{\ell \in \A}$, between critical $i$-cells $\cell{D,G}$ and critical $(i-1)$-cells $\cell{C,F}$, in the Morse complex. We refer to \cite[Section 5]{salvetti2007combinatorial} for a detailed explanation of how to compute these incidence numbers, given an acyclic matching on the Salvetti complex $\sal(\A)$. We only make the following substantial change of convention with respect to \cite{salvetti2007combinatorial, gaiffi2009morse}: given a cell $\cell{C,F}$, we choose as its representative point the $0$-cell $\cell{C^F,C^F}$, where $C^F$ is the chamber opposite to $C$ with respect to $F$ (the role of the representative point is thoroughly described in \cite[Section 9]{steenrod1943homology}). It is more convenient to choose $\cell{C^F,C^F}$ instead of $\cell{C,C}$, because in this way two matched cells have the same representative point. We recall some useful definitions and facts from \cite[Chapter 5]{salvetti2007combinatorial}, adapting them to our different convention on the representative point. Given two chambers $D$ and $C$, denote by $u(D,C)$ a combinatorial positive path of minimal length from $\cell{D,D}$ to $\cell{C,C}$, in the $1$-skeleton of $\sal(\A)$. In particular, let $\Gamma(C) = u(C, C_0)$ be a minimal positive path from the chamber $C$ to a base chamber $C_0$. Every path $u(D,C)$ crosses each hyperplane at most once by \cite[Lemma 5.1]{salvetti2007combinatorial}. Consider the closed path $\Gamma(D)^{-1} \, u(D,C) \, \Gamma(C)$, which starts from $C_0$, passes through $D$ and $C$, and then goes back to $C_0$. This path determines an element $\bar u(D,C) \in H_1(M(\A))$ which is equal to the product of the positive loops around the hyperplanes in $s(C_0,C) \cap s(D,C)$. Then the incidence number $[\cell{D,G}, \cell{C,F}] \in \Z[t_\ell^{\pm 1}]_{\ell \in \A}$ between an $i$-cell $\cell{D,G}$ and an $(i-1)$-cell $\cell{C,F}$ in $\sal(\A)$ is given by \begin{equation*} [\cell{D,G}: \cell{C,F}]= [\cell{D,G}: \cell{C,F}]_{\Z} \; \bar{u}(D^G,C^F), \end{equation*} where $[\cell{D,G} : \cell{C,F}]_{\Z} = \pm 1$ denotes the integral incidence number in $\sal(\A)$. Let $x_0 \in \R^2$ be a generic point with respect to the line arrangement $\A$, and fix a Euclidean matching $\mathcal{M}$ on the Salvetti complex $\sal(\A)$ with base point $x_0$. Let $C_0$ be the chamber containing $x_0$ (this is the first chamber in any Euclidean order with base point $x_0$). Recall that the matching $\mathcal{M}$ is constructed on the fibers of the map $\eta \colon \sal(\A) \to \C \times \C$. To compute the algebraic Morse complex (see \cite[Definition 11.23]{kozlov2007combinatorial}), we first need to describe the alternating paths between critical cells. The alternating paths between a critical $1$-cell $\cell{C,F}$ and the only critical $0$-cell $\cell{C_0,C_0}$ are particularly simple, since all the $0$-cells are in $N(C_0)$. \begin{lemma}\label{Lemma:1path} Let $\cell{C,F}$ be a critical $1$-cell. Denote by $C'$ the unique chamber containing $F$ other than $C$. There are exactly two alternating paths from $\cell{C,F}$ to the only critical $0$-cell $\cell{C_0,C_0}$: \begin{itemize} \item $\cell{C,F} \searrow \cell{C,C} \nearrow \cell{C',F} \searrow \cell{C',C'} \nearrow \dots \searrow \cell{C_0,C_0}$, \item $\cell{C,F} \searrow \cell{C',C'} \nearrow \dots \searrow \cell{C_0,C_0}$ \end{itemize} (after $\cell{C',C'}$, they continue in the same way). \end{lemma} \begin{proof} Since $\cell{C,F}$ is critical, the line $|F|$ separates $C$ and $C_0$. In the boundary of the $1$-cell $\cell{C,F}$ there are the two $0$-cells $\cell{C,C}$ and $\cell{C',C'}$. The $0$-cell $\cell{C,C}$ is matched with the $1$-cell $\cell{C',F}$, because these are the unique cells in the fiber $\eta^{-1}(C_0, C)$. Then an alternating path starting with $\cell{C,F} \searrow \cell{C,C}$ is forced to continue with $\nearrow \cell{C',F} \searrow \cell{C',C'}$. After $\cell{C',C'}$ there is exactly one way to continue the path, because every non-critical $0$-cell is matched with some $1$-cell, and this $1$-cell has exactly one other $0$-cell in the boundary. Since the matching is proper, one such path must eventually reach the critical $0$-cell $\cell{C_0,C_0}$. \end{proof} We can use the previous lemma to compute the boundary $\partial_1$. The resulting formula coincides with the one of \cite[Proposition 4.1]{gaiffi2009morse}. \begin{proposition} The incidence number between a critical $1$-cell $\cell{C,F}$ and the only critical $0$-cell $\cell{C_0,C_0}$ in the Morse complex is given by \[ [\cell{C,F} : \cell{C_0, C_0}]^\mathcal{M} = (1-t_{\,|F|}). \] \end{proposition} \begin{proof} The orientation of a $1$-cell $\cell{\tilde{C},\tilde{F}}$ is defined so that $[\cell{\tilde{C}, \tilde{F}}, \cell{\tilde C^{\tilde F}, \tilde C^{\tilde F}}]_\Z = 1$. Now, if $\cell{\tilde{C},\tilde{F}} \in N(C_0)$, then $\tilde{C}$ is closer to $C_0$ with respect to $\tilde{C}^{\tilde{F}}$ and so we have that: \begin{equation*} [\cell{\tilde{C},\tilde{F}},\cell{\tilde{C},\tilde{C}}] = -1; \qquad [\cell{\tilde{C},\tilde{F}}:\cell{\tilde{C}^{\tilde{F}},\tilde{C}^{\tilde{F}}}] = 1. \end{equation*} By Lemma \ref{Lemma:1path} we see that there are exactly two alternating paths between $\cell{C,F}$ and $\cell{C_0,C_0}$, and by \cite[Definition 11.23]{kozlov2007combinatorial} the incidence number in the Morse complex is given by \begin{equation*} [\cell{C,F} : \cell{C_0, C_0}]^\mathcal{M} = [\cell{C,F}:\cell{C,C}]+[\cell{C,F}:\cell{C',C'}]. \end{equation*} Since $|F| \in s(C_0,C) \cap s(C^F,C)$, the first term is \begin{equation*} [\cell{C,F}:\cell{C,C}]=[\cell{C,F}:\cell{C,C}]_{\Z} \bar{u}(C^F,C)=-t_{|F|}, \end{equation*} The second term is given by \begin{equation*} [\cell{C,F}:\cell{C',C'}]=[\cell{C,F}:\cell{C',C'}]_{\Z} \bar{u}(C^F,C')=\bar{u}(C',C')=1. \qedhere \end{equation*} \end{proof} Now we want to compute the boundary $\partial_2$. To simplify the notation, denote a $2$-cell $\cell{D,\{p\}}$ also by $\cell{D,p}$, where $p \in \R^2$ is the intersection point of two or more lines of $\A$. It is convenient to assign the orientation of the $2$-cells so that they behave well with respect to the matching. Given a $2$-cell $\cell{D,p} \not\in N(C_0)$, we choose the orientation in the following way. Let $\ell, \ell'$ be the two walls of $D$ that pass through $p$. Let $\ell$ be the one that does not separate $D$ from $C_0$ if it exists, or otherwise the closest one to $x_0$. Then the orientation of $\cell{D,p}$ is the one for which $[\cell{D,p}, \cell{D,\ell}]_\Z = 1$. The orientation of the $2$-cells in $N(C_0)$ is assigned arbitrarily. The reason of this choice is that the incidence number between two matched cells is always $+1$. Indeed, if $C'$ is the chamber such that $X_{C'}=\ell'$, then $\cell{D,p} \in N(C')$ by construction. We are going to show that there is a correspondence between alternating paths from critical $2$-cells to critical $1$-cells and certain sequences of elements of $\L_1(\A)$. Consider an alternating path of the form \begin{equation}\label{eq:AltPath} \cell{D,p} \searrow \cell{C_1,F_1} \nearrow \cell{D_1,p_1} \searrow \cell{C_2,F_2} \dots \searrow \cell{C_n,F_n}, \end{equation} where $\cell{D,p}$ is a critical $2$-cell and $\cell{C_n,F_n}$ is a critical $1$-cell. By construction of the matching, none of the cells in \eqref{eq:AltPath} belongs to $N(C_0)$. We have that the starting cell $\cell{D,p}$ and the sequence $(F_1,\ldots,F_n)$ uniquely determine the alternating path. This is because for each $i$ there are only two cells of the form $\cell{C',F_i}$ for some $C' \in \C$, and one of these cells is in $N(C_0)$. Then $C_i$ is uniquely determined by $F_i$ for every $i$, and $\cell{D_i, p_i}$ is the cell matched with $\cell{C_i, F_i}$. We are now going to describe which sequences in $\L_1(\A)$ give rise to an alternating path. Given a face $F \in \L_1(\A)$, let $\ell = |F|$. If $\rho_\ell(x_0)\notin F$, we denote by $p(F)$ the endpoint of $F$ which is closer to $\rho_\ell(x_0)$. In addition, let $C(F)$ be the unique chamber containing $F$ such that $\cell{C(F),F} \notin N(C_0)$. \begin{definition} \label{Def:valid-sequence} Given two different faces $F,\, G \in \L_1(\A)$, we write $F \to G$ if \begin{itemize} \item $F \cap G = \{p(F)\}$; \item $|F|=|G|$, or $F$ and $C_0$ lie in the same half-plane with respect to $|G|$. \end{itemize} \end{definition} \begin{lemma}\label{Lemma:sequence} Let $\cell{D,p}$ be a critical $2$-cell and $\cell{C,F}$ a critical $1$-cell. The alternating paths between $\cell{D,p}$ and $\cell{C,F}$ are in one to one correspondence with the sequences in $\L_1(\A)$ of the form $(F_1 \to F_2 \to \ldots \to F_n=F)$ such that $\cell{C(F_1),F_1} < \cell{D,p}$. \end{lemma} \begin{proof} Consider an alternating path as in \eqref{eq:AltPath}. We have already seen that such a path is completely determined by the starting cell $\cell{D,p}$ and by the sequence $(F_1,\ldots,F_n)$. Clearly the condition $\cell{C(F_1),F_1} < \cell{D,p}$ must be satisfied. We want to show that $F_i \to F_{i+1}$ for each $i=1,\dots,n-1$. Let $E_i$ be the chamber opposite to $C(F_i)$ with respect to $F_i$. By construction of the matching, it is immediate to see that the cell $\cell{C(F_i),F_i}$ is matched with $\cell{D(F_i),p(F_i)}$, where $D(F_i)$ is the chamber opposite to $E_i$ with respect to $p(F_i)$. By hypothesis we have that $\cell{C(F_{i+1}),F_{i+1}} < \cell{D(F_i),p(F_i)}$ which implies that $F_i \cap F_{i+1} = \{p(F_i)\}$ and that $D(F_i).F_{i+1}=C(F_{i+1})$. Since $\cell{C(F_{i+1}),F_{i+1}} \notin N(C_0)$, we have that $C_0$ and $C(F_{i+1})$ are in opposite half-planes with respect to $|F_{i+1}|$. The same is true for $F_i$ and $C(F_{i+1})$, because $D(F_i)$ and $F_i$ are in opposite half-planes with respect to $|F_{i+1}|$, unless $F_i \subset |F_{i+1}|$. Then we have that $F_i \to F_{i+1}$. Conversely, we now prove that every sequence $(F_1 \to F_2 \to \dots \to F_n=F)$ satisfying $\cell{C(F_1), F_1} < \cell{D,p}$ has an associated alternating path. We do this by induction on the length $n$ of the sequence. The case $n=1$ is trivial, since we already know that $\cell{C(F_1),F_1}<\cell{D,p}$. In the induction step, we need only to prove that $F \to G$ implies $\cell{C(G),G} < \cell{D(F),p(F)}$. From the first condition of Definition \ref{Def:valid-sequence}, we have that $G \prec \{p(F)\}$. We need to check that $D(F).G=C(G)$. By definition of $C(G)$, this is equivalent to proving that $D(F)$ and $C_0$ lie in opposite half-planes with respect to $|G|$. This is true because $F$ and $C_0$ lie in the same half-plane with respect to $|G|$. \end{proof} Now that we have a description of the alternating paths, we can use it to compute the boundary of the Morse complex. \begin{definition} Given two different faces $F, \,G \in \L_1(\A)$, let \begin{equation*} [F \to G]= \frac{[\cell{D(F),p(F)}:\cell{C(G),G}]}{[\cell{D(F),p(F)}:\cell{C(F),F}]}, \end{equation*} where the incidence numbers on the right are taken in the Salvetti complex $\sal(\A)$, and $D(F)$ is defined as in the proof of Lemma \ref{Lemma:sequence}. \end{definition} \begin{lemma} \label{Lemma:boundary-line} Given two different faces $F,G \in \F_1(\A)$ such that $F \to G$, we have \begin{equation*} [F \to G]=\pm \prod t_\ell, \end{equation*} where the product is on the set of lines $\ell \neq |G|$ passing through $p(F)$, such that $G$ and $C_0$ lie in opposite half-planes, whereas $F$ and $C_0$ lie in the same closed half-plane (with respect to $\ell$). The sign is $+1$ if $p(F)=p(G)$, and $-1$ otherwise. \end{lemma} \begin{proof} Denote by $E(F)$ and $E(G)$ the chambers $C(F)^F$ and $C(G)^G$, respectively. Notice that $E(F)=D(F)^{p(F)}$, and therefore $[\cell{D(F),p(F)}:\cell{C(F),F}]= 1$. See Figure \ref{fig:F-to-G} for an example. Now we need to determine $\bar{u}(E(F),E(G))$, which is the product of the positive loops around the hyperplanes in $s(C_0,E(G)) \cap s(E(F),E(G))$. By definition of $E(G)$, we have that $s(C_0,E(G))$ is the set of lines different from $|G|$ for which $G$ and $C_0$ in opposite half-planes. Since every line in $s(E(F),E(G))$ goes through $p(F)$, it is now easy to see that $s(C_0,E(G)) \cap s(E(F),E(G))$ is the set described in the statement. We now need to determine the sign. If $p(G)=p(F)$, then we immediately see that $G$ is in the half-plane delimited by $|F|$ that contains $D(F)$. The opposite is true if $p(G) \neq p(F)$. By our choice of the orientation, we obtain the stated result. \end{proof} \begin{figure} \begin{center} \begin{tikzpicture}[label distance=-3] \draw[name path=L1] (-5,3) -- (7,3); \draw[name path=L2] (-5,0) -- (7,0); \draw[name path=L3] (-4,-2) -- (7,3.5); \draw[name path=L4] (-2,-3) -- (3,4.5); \draw[name path=L5] (3,-3) -- (-4.2,4.2); \node[circle,inner sep=1.5pt, fill=black, label=left:{$x_0$}] (x0) at (-3,0.5) {}; \node (C0) at (-4,1.5) {$C_0$}; \node (CF) at (2.3,1.9) {$C(F)$}; \node (EF) at (-0.4,1.9) {$E(F)$}; \node (CG) at (-2,-1.8) {$C(G)$}; \node (EG) at (-4,-0.8) {$E(G)$}; \path[name intersections={of=L1 and L4, by=a}]; \begin{scope}[every path/.style={draw=blue, very thick, fill=red}] \draw (a) -- (0,0); \draw (0,0) -- (-4,-2); \end{scope} \begin{scope}[every node/.style={text=blue}] \node (F) at (0.8,1.6) {$F$}; \node (G) at (-2.7,-1.05) {$G$}; \end{scope} \fill[white] (0.3,-0.31) circle(0.25); \node[circle,inner sep=1.5pt, fill=black, label=below right:{$\!p(F)$}] (p) at (0,0) {}; \end{tikzpicture} \end{center} \caption{Faces $F,G \in \F_1(\A)$ such that $F \to G$, as in Lemma \ref{Lemma:boundary-line}.} \label{fig:F-to-G} \end{figure} \begin{theorem}\label{Thm:line-complex} Let $\A$ be a locally finite line arrangement in $\R^2$. Let $\cell{D,p}$ be a critical $2$-cell and $\cell{C,F}$ a critical $1$-cell. Then their incidence number in the Morse complex is given by \begin{equation*} [\cell{D,p}:\cell{C,F}]^\mathcal{M} =\sum_{s \in \Seq}{\omega(s)}, \end{equation*} where $\Seq$ is the set of sequences of Lemma \ref{Lemma:sequence}, and for each sequence $s=(F_1 \to F_2 \to \dots \to F_n=F) \in \Seq$ we define \begin{equation*} \omega(s)=(-1)^n \, [\cell{D,p}:\cell{C(F_1),F_1}] \, \prod_{i=1}^{n-1}{[F_i \to F_{i+1}]}. \end{equation*} \end{theorem} \begin{proof} It follows directly from \cite[Definition 11.23]{kozlov2007combinatorial} and Lemma \ref{Lemma:sequence}. \end{proof} \begin{example}[Deconing $A_3$] Consider the line arrangement $\A$ of Figure \ref{fig:dec-A3}, obtained by deconing the reflection arrangement of type $A_3$. Given a chamber $C_i$, denote by $\cell{C_i,F_i}$ the associated critical cell if it is of dimension $1$, or by $\cell{C_i,p_i}$ if it is of dimension $2$. Applying Theorem \ref{Thm:line-complex} and Lemma \ref{Lemma:boundary-line}, we obtain the boundary matrix $\partial_2$ of Table \ref{tab:decA3-2}. This matrix is slightly simpler than the one computed in \cite[Section 7]{gaiffi2009morse}, but there are many similarities. Specializing to the case $t_1=\ldots=t_5=t$, we obtain that \begin{equation*} H_1(M(\A); \, \mathbb{Q}[t^{\pm 1}]) \cong \left( \frac{\mathbb{Q}[t^{\pm 1}]}{t-1} \right)^3 \oplus \frac{\mathbb{Q}[t^{\pm 1}]}{t^3-1}, \end{equation*} as already computed for example in \cite{gaiffi2009morse}. \end{example} \begin{figure} \begin{tikzpicture}[label distance=-3] \draw[name path=L1] (0,6.5) -- (6.5,0); \draw[name path=L2] (10,6.5) -- (3.5,0); \draw[name path=L3] (0,4) -- (10,4); \draw[name path=L4] (10,1.5) -- (3.5,8); \draw[name path=L5] (0,1.5) -- (6.5,8); \path[name intersections={of=L1 and L2, by=x6}]; \path[name intersections={of=L3 and L2, by=x7}]; \path[name intersections={of=L1 and L3, by=x8}]; \path[name intersections={of=L2 and L4, by=x4}]; \node[circle,inner sep=1.5pt, fill=black, label=left:{$x_0$}] (x0) at (6,3.6) {}; \node (C0) at (5,2.9) {$C_0$}; \node (C1) at (5,5.1) {$C_1$}; \node (C2) at (7.5,2) {$C_2$}; \node (C3) at (7.5,6) {$C_3$}; \node (C4) at (9,4.6) {$C_4$}; \node (C5) at (9,3.4) {$C_5$}; \node (C6) at (2.5,2) {$C_6$}; \node (C7) at (5,0.5) {$C_7$}; \node (C8) at (2.5,6) {$C_8$}; \node (C9) at (5,7.5) {$C_9$}; \node (C10) at (1,4.6) {$C_{10}$}; \node (C11) at (1,3.4) {$C_{11}$}; \node (l1) at (0.5,6.5) {$l_1$}; \node (l2) at (9.5,1.5) {$l_2$}; \node (l3) at (9.7,3.7) {$l_3$}; \node (l4) at (9.5,6.5) {$l_4$}; \node (l5) at (0.5,1.5) {$l_5$}; \end{tikzpicture} \caption{Deconing $A_3$.} \label{fig:dec-A3} \end{figure} \begin{table}[htbp] \begin{tabular}{l|cccccc|} & $\cell{C_4,p_4}$ & $\cell{C_5,p_5}$ & $\cell{C_7,p_7}$ & $\cell{C_9,p_9}$ & $\cell{C_{10},p_{10}}$ & $\cell{C_{11},p_{11}}$ \\ &&&&&&\\[-1em] \hline &&&&&&\\[-1em] $\cell{C_1,F_1}$ & $1-t_4$ & $t_4(t_2-1)$ & $0$ & $0$ & $t_1-1$ & $t_1(1-t_5)$ \\ $\cell{C_2,F_2}$ & $t_2t_3-1$ & $t_2-1$ & $1-t_1$ & $0$ & $0$ & $0$ \\ $\cell{C_3,F_3}$ & $t_3(1-t_4)$ & $1-t_3t_4$ & $0$ & $t_5-1$ & $0$ & $0$ \\ $\cell{C_6,F_6}$ & $0$ & $0$ & $t_4-1$ & $0$ & $1-t_3t_5$ & $1-t_5$\\ $\cell{C_8,F_8}$ & $0$ & $0$ & $0$ & $1-t_2$ & $t_3(t_1-1)$ & $t_1t_3-1$ \\ \hline \end{tabular} \bigskip \caption{The boundary $\partial_2$ of the deconing of $A_3$} \label{tab:decA3-2} \end{table} \section*{Acknowledgements} We would like to thank our supervisor, Mario Salvetti, for always giving valuable advice. We also thank Emanuele Delucchi, Simona Settepanella, Federico Glaudo, and Daniele Semola, for the useful discussions. \bibliographystyle{amsalpha-abbr}

50435d80865c829c04aa3f0ad3838198d3624dd8

\section{Introduction} Observations of the Milky Way (MW) and the haloes of nearby galaxies have revealed spatial sub-structures \citep{majewski2003,belokurov2006, morrison2009,bonaca2012, sesar2012, gilbert2013,ibata2014}, as well as differences in the kinematical and chemical properties of the stars in the inner region of the Galactic halo with respect to the outer region \citep{carollo2007,carollo2010,beers2012,fernandezalvar2015}. These results suggest that the stellar populations dominating the inner- and outer-halo regions formed in systems with different star-formation histories. One explanation proposed is that the outer halo was built mainly from mergers of satellite galaxies, whereas the inner halo might have received contributions of stars formed in-situ in the Galaxy \citep[for example stars from a heated disc due to merger events, e.g. ][]{sheffield2012, nissen2010, schuster2012} The formation of the stellar halo has been tackled using dynamical and hydrodynamical simulations \citep[e.g.,][]{bullock2010,cooper2010, tissera2013,pillepich2015}. All of these works predict the formation of the stellar halo mainly from the contribution of satellites with a variety of masses. However, hydrodynamical simulations also reported the possible existence of an in-situ component \citep{zolotov2009, tissera2012, cooper2015}. Disc-heated stars are a possible candidate to explain the origin of the in-situ components \citep[as proposed from both numerical simulations as well as observational works, e.g.][]{tissera2014, bonaca2017}, but they might have also formed from gas transported in by gas-rich satellites \citep{zolotov2009, tissera2013}. \citet{harmsen2017} found that these simulations were able to reproduce global trends of the stellar haloes of nearby galaxies, although most of them predicted more massive stellar haloes than inferred from observations, and larger contributions of in-situ stars compared to the observed estimates. However, recently observations have shown evidence for the accretion of a relatively massive satellite contributing stars to the inner halo of the Milky Way. \citet{belokurov2018} observed that the orbits of stars with [Fe/H] $> -1.7$ are strongly radial, in a sausage-like distribution, and, by comparing with N-body simulations, deduced that they come from the merger of a satellite with virial mass M $> 10^{10}$ M$_{\odot}$, which was called in following works the ``Gaia sausage". Using data from the Gaia second data release (DR2), \citet{deason2018} obtained that stars with very radial orbits (like in the ``Gaia sausage") have apocentres which pile up at $r \sim$ 20 kpc, where the spatial break in the halo was measured in numerous previous papers. They conclude that these stars likely come from a satellite with stellar mass at least M $> 10^{8}$ M$_{\odot}$, or from a group of dwarf satellites accreted at similar times. \citet{helmi2018} identified stars in retrograde orbits, and they also constrained \citep[from the star-formation rates obtained from similar stars in][] {fernandezalvar2018a} the stellar mass of their progenitor satellite to be $\sim 6\times 10^{8}$ M$_{\odot}$ (which they called ``Gaia-Enceladus"). From the analysis of globular clusters in the Galactic halo, other evidence was found for relatively massive mergers. \citet{myeong2018} confirmed that the configuration of a group of globular clusters in action space is compatible with their coming from the ``Gaia sausage" merger event. \citet{kruijssen2018} used the age-metallicity relation in halo clusters, compared with the E-MOSAICS hydrodynamical simulations, to infer that the MW may have undergone mergers with three satellites with stellar masses between $10^{8}$ and $10^{9}$ (the most massive of which they refer to as ``Kraken"). Other evidences have been also provided by \citet{lancaster2018,haywood2018,fernandezalvar2018b,simion2018,iorio2018,wegg2018,mackereth2018}. An important piece of information is provided by $\alpha$-element abundances \citep[e.g.,][]{zolotov2009}. Using the Aquarius haloes, \citet{tissera2013} showed that stars in the stellar haloes have different $\alpha$-enrichments that could provide hints on the mass and star-formation history (SFH) of the accreted satellites, as well as the relative contributions of stars formed through different processes. In their simulations, the authors identifed accreted stars (or debris) formed in systems outside the virial radius of the main galaxy progenitor, and later on, in-situ stars, formed within the virial radius. Within the latter population, they also identified the so-called endo-debris stellar populations, formed from gas carried in by accreted satellites, and disc-heated stars, born in-situ in the disc and later, dynamically heated into the halo. Their different origin predicts different chemical and kinematical patterns in the inner regions of the stellar haloes, in particular different $\alpha$-enhancement patterns. Indeed, a promising path for investigation is consideration of the [$\alpha$/Fe] abundance ratios that have been derived from observations of large numbers of stars in the MW stellar halo \citep{fernandezalvar2015, fernandezalvar2017}. The $\alpha$-elements are ejected into the interstellar medium (ISM) on short timescales (Myr), because these elements are produced mainly by massive stars that explode as Type II supernovae (SNeII). In contrast, release of the bulk of Fe occurs on longer timescales (Gyr), because iron is synthesised by binary stars of low and intermediate mass that explode as Type Ia supernovae (SNeIa) . The patterns in the relative abundances [$\alpha$/Fe]\footnote{[$\alpha$/Fe] = $\log$($\alpha$/Fe) $- \log$($\alpha$/Fe)$_{\odot}$} vs. [Fe/H]) are well-known to be affected by the star-formation rate and the initial mass function (e.g., see Fig. 1 of \citet{fernandezalvar2018a}). \citet{fernandezalvar2017} identified abundance gradients for $\alpha$-elements and other chemical species relative to the iron abundance with distance from the Galactic center, based on an analysis of over 400 stars from the APOGEE twelfth data release (DR12). They found that the nature of these gradients depended on the [Fe/H] range, flattening as [Fe/H] decreases, suggesting a variation of the chemical-enrichment history for stars as a function of distance. The identification of Galactic halo simulations able to reproduce the observed trends can help to unravel the assembly history of our Galaxy. In this work, we examine the [$\alpha$/Fe] trends with Galactocentric distance observed in the halo of the MW by the APOGEE fourteenth data release (DR14), which has tripled the number of stars with this information available. We compare them with the trends inferred from a set of simulated haloes of MW mass-size galaxies in order to constrain the assembly history that led to the stellar populations in the inner-halo region with the observed chemical compositions. We employ the set of Aquarius haloes analysed by \citet{tissera2012, tissera2013,tissera2014} and \citet{carollo2018} that correspond to MW mass-size galaxies to study the chemical enrichment of baryons during Galactic assembly. These authors found that the stellar halo of the MW is characterised by combination of stellar populations with different chemistry, kinematics, ages, and binding energies depending on their origin. We acknowledge the fact that the Aquarius haloes \citep{scan09} are known to over-produce stars at high redshift. However, they comprise a diversity of assembly histories for galaxies of similar virial masses that allow us to explore which assembly history provides the best representation of the observations. Our approach is to use them as case studies to identify the main processes that could explain the origin of the $\alpha$-abundance patterns detected in observations of the MW's halo. This paper is organised as follows. In section 2 we describe the main characteristics of the observational and simulated data used in this work. In section 3 we infer the chemical trends from both sets, and analyse their similarities and differences. Finally, in section 4 we summarise our main results and conclusions. \section{Observational and simulated data} \subsection{Observations} We make use of the APOGEE fourteenth data release, DR14 \citep{blanton2017}. This stellar spectra database includes all the APOGEE observations gathered from May 2011 to July 2014 \citep{eisenstein2011}, and the observations from the first two years of its extension, APOGEE-2. Using a multi-object infrared spectrograph \citep{wilson2010} coupled to the 2.5-meter telescope at the Apache Point Observatory \citep{Gunn2006}, high-resolution ($R\sim22,500$) spectra were obtained with a typical signal-to-noise S/N$\sim 100$. Targets were spread over the main Galactic components, including the halo \citep{zasowski2013,zasowski2017}. DR14 offers elemental-abundance determinations for $\sim$19 chemical especies determined by the APOGEE Stellar Parameters and Abundances pipeline \citep[ASPCAP; ][]{garcia2016}. We reproduce the analysis performed by \citet{fernandezalvar2017}, now using the more extended sample of APOGEE-2 stars included in DR14. As in their work, we select stars at distances from the Galactic plane, $|h|$, larger then 5 kpc to avoid contamination from thick-disc stars. We consider distances derived by the Brazilian Participation Group \citep[BPG][]{santiago2016, queiroz2018} from the spectrophotometric stellar atmospheric parameters (effective temperature, $T_{\rm eff}$, and surface gravity, $\log g$) determined in DR14. We reject objects with less reliable measured parameters, warned by the flags STAR\_BAD, GRIDEDGE\_BAD, PERSIST\_LOW, PERSIST\_MED, PERSIST\_HIGH, TEFF\_BAD, LOGG\_BAD,METALS\_BAD, ALPHAFE\_BAD, CHI2\_BAD, SN\_BAD, and NO\_ASPCAP\_RESULT in the catalogue. In addition, we only consider stars with spectra having S/N $> 80$, $T_{\rm eff} > 4000$ K, and $1.0 < \log g < 3.5$. These were the selection criteria chosen by \citet{fernandezalvar2017} for analysing the APOGEE DR12, to which we aim to compare our inferred chemical trends. Thus, we would like to maintain the same $T_{\rm eff}$ and $\log g$ selection cuts for consistence. Besides, although improvements in the ASPCAP were performed in DR14 (see \citet{holtzman18}), chemical issues have been still detected related to stars at lower $T_{\rm eff}$ and large $\log g$. For instance, \citet{zasowski19} revealed a strange high [O/Fe] trend at Solar metallicities in stars with $T_{\rm eff} < 3800$ K. And due to the lack of surface gravity estimations from asteroseismology at the time DR14 was released, calibrated $\log g$ values for stars at $\log g > 4$ are still not provided in DR14. We also exclude stars likely belonging to clusters. We do not exclude targets in streams, because substructures will also be considered in the simulation analysis. Our final sample comprises 1185 stars. The mean uncertainties estimated and their standard deviations, for the stellar parameters chemical abundances and distances in this stellar sample are the following: $\sigma(T_{\rm eff}) = 83 \pm 17$ K, $\sigma(\log g) = 0.09 \pm 0.02$, $\sigma($[O/Fe]$) = 0.07 \pm 0.04$, $\sigma($[Mg/Fe]$) = 0.05 \pm 0.02$, and $\sigma(d_{\rm sun})/d_{\rm sun} = 12 \pm 3 \%$. \subsection{Simulations} We analysed six Milky-Way mass-sized haloes from the Aquarius project \citep{scan09}. The initial conditions of the Aquarius haloes are consistent with a $\Lambda$CDM model with $\Omega_{m} = 0.25$, $\Omega_{\Lambda} = 0.75$, $\sigma_{8}=0.9$, $n_{S}=1$, and $H_{0}=100 h$ km $\rm s^{-1}$ Mp$\rm c^{-1}$ with $h = 0.73$. Dark matter particles have masses on the order of $\sim 10^{6} M_{\odot}h^{-1}$; initially the particles have masses of $\sim2 \times 10^{5} M_{\odot} h^{-1}$. The analysed haloes correspond to level-5 resolution. The analysed haloes are Aq-C-5 and Aq-D-5. Hereafter, we will use them Aq-C and Aq-D for simplicity. The simulations were performed using a modified version of GADGET-3 \citep{scan05,scan06} that includes the physics of the baryons, chemical evolution, and galactic mass-loaded outflows triggered by SNe explosions. The chemical model follows the enrichment from both SNeII and SNeIa. The galactic systems are identified using the Friends-of-Friends algorithm, and have virial masses in the range $\sim 5-11 \times 10^{11} M_{\odot}h^{-1}$. Within the virial radius the simulations are numerically resolved using $\sim 1$ million total particles at z=0. Satellite galaxies that could be identified by the SUBFIND \citep{springel2005} were taken out. Streams and substructures were not removed. To define the stellar haloes, we applied the same criteria described in \citet{tissera2012}, with an extra condition on the minimum height ($|h|$) above the disc plane. We define the disc components by considering particles dominated by rotation (i.e., using the ratio ($\epsilon$) between the angular momentum along the z-axis and the total angular momentum at that binding energy, so that star particles with $\epsilon>0.65$ are taken as part of the disc components). Star particles that are dominated by dispersion and have a height $|h| < 5$ kpc are excluded from the analysis. This condition allows us to better mimic the observations that, as described above, apply a similar condition \citep[see also][]{carollo2018, monachesi2018}. For this analysis we consider the entire stellar haloes (i.e., no separation between inner and outer haloes is applied). In order to explore in more detail the origin of the $\alpha$-element patterns in the inner region (within 30 kpc) of the stellar haloes, we adopt the same classification as \citet{tissera2013}: (1) Stars born from gas accreted in the first stages of assembly, (2) Disc-heated stars formed in the disc structure of the main progenitor galaxy, then heated kinematically, and (3) Stars formed from gas carried in by gas-rich satellite galaxies ('endo-debris'). \begin{figure*} \includegraphics[scale=0.8]{whole_sample.pdf} \includegraphics[scale=0.8]{reduced_sample.pdf} \caption{Top panels: Median [O/Fe] and [Mg/Fe] ratios determined for the selected whole sample in bins $r$ (from 5 kpc to 30 kpc) for stars in three different ranges in [Fe/H] ($-2.5 <$ [Fe/H] $< -1.8$, $-1.8 <$ [Fe/H] $< -1.1$, and -$1.1 <$ [Fe/H] $< 0.0$). The error bars are the statistical error for each median determined, the median absolute deviation (MAD). Bottom panels: Same as top panels but considering only stars with 4600 < $T_{\rm eff}$ < 4800 (reduced sample).} \label{figure1} \end{figure*} \begin{figure*} \includegraphics[scale=0.8]{histo_only_norm_sn80.pdf} \caption{Distribution of [X/Fe] of the whole sample for each $r$ range (each histogram: 5 $< r < 10$ kpc in black, 10 $< r <$ 15 kpc in grey, and 15 $< r < 30$ kpc in red) at the three different [Fe/H] bins considered, from the first to the third rows of panels, respectively. The y-axis corresponds to the number of stars ($N$) in each [X/Fe] bin, normalized to the total number of stars ($N_{tot}$) in each bin of $r$ and [Fe/H].} \label{histo_norm} \end{figure*} We acknowledge the fact that these simulated haloes appear to be more massive than suggested by observations \citep{harmsen2017}. In part, this could be due to an excess of in-situ stars or a misclassification of thick-disc stars as disc-heated stars. Depending on the halo, the in-situ fraction contributes about 30 per cent of the inner stellar halo \citep{tissera2012}. A better agreement between observations and simulations is found if the in-situ components are not considered \citep{carollo2018}. This is also the case for other simulations that have been performed with different numerical codes and baryonic physics \citep[e.g.][]{monachesi2018}. Nevertheless, the comparative analysis of the assembly histories and the $\alpha$-element patterns of these two haloes provides useful clues for the interpretation of the current observations, as they have been performed with the same subgrid physics and numerical resolution, being the initial conditions (i.e. assembly histories) the only difference between them. In this paper, we focus on the simulated [O/Fe] distributions as a function of radius, which we take to be representative of [Mg/Fe] as well. The yields adopted for SNII are those of \citet{woosleyweaver1995} and they have some issues with Mg so that they produce less elements than expected. However the O abundances are well-reproduced. As a consequence we decided to use O abundances for the simulations. \section{The analysis} \subsection{Observed chemical patterns} We aim to properly study the [X/Fe] abundance variations across the halo for $\alpha$-elements, as a function of metallicity, which we identify with the [Fe/H] abundance ratio. We perform this analysis for the [O/Fe] and [Mg/Fe] abundances released in DR14. These two elements have a similar origin, and are mainly released into the interstellar medium through the explosion of Type II supernovae. As in \citet{fernandezalvar2017}, we calculate the median [X/Fe] from the abundances available in the DR14 catalogue in bins of distance from the Galactic center, $r$, of 2.5 kpc, assuring a minimum of 5 stars per bin. We calculate the medians for different metallicity ranges: $-2.5 <$ [Fe/H] $< -1.8$ (low-metallicity population), $-1.8 <$ [Fe/H] $< -1.1$ (intermediate-metallicity population), and $-1.1 <$ [Fe/H] $< 0.0$ (high-metallicity population). The derived chemical trends for our entire halo sample are shown in Figure \ref{figure1}. The panels exhibit the resulting [X/Fe] medians, as well as their corresponding median absolute deviation (MAD), as a function of $r$. The chemical trends for both elemental-abundance ratios exhibit a significant decrease with $r$ for stars at $-1.1 <$ [Fe/H] $< 0.0$. We first evaluate the possibility that this gradient is the result of our selection criteria. We obtain the $\log g$ distribution at each [Fe/H] and radial bin. We note that for $r > 15$ kpc the $\log g$ distribution is skewed toward lower values (peak at $\log g \sim 1.0$), whereas at lower distances the distribution peaks at higher $\log g$ values. The [O/Fe] abundances do not show a trend with $\log g$, except a slightly increasing trend with $\log g$ in the $-1.1 <$ [Fe/H] $<0.0$ bin at $r > 15$ kpc. Stars in this bin exhibit lower [O/Fe] values at $\log g \sim 1.0$ than stars at $r < 15$ kpc. However, this trend with $\log g$ is not observed in the case of [Mg/Fe], for which the decreasing trend of this ratio with distance is the same as observed for [O/Fe]. For this reason, we consider that the decrease detected for the [$\alpha$-elements/Fe] ratio with distance from the Galactic center is not due to the shift in the $\log g$ distribution toward lower values at the largest distances. The oxygen abundances are determined from 50 regions of the spectra covering mostly OH molecular bands (Holtzman et al. 2018). The OH bands are very sensitive to the adopted $T_{\rm eff}$. We detect a bias in [Fe/H] towards high values at $T_{\rm eff} >$ 4600\,K. We also detect an increasing trend of [O/Fe] with $T_{\rm eff}$, and an important increase in the abundance dispersion for stars with $T_{\rm eff} > 5000$\,K. [Mg/Fe] abundances also show an increasing trend with $T_{\rm eff}$, although much lower than for [O/Fe]. In order to avoid artificial distortions in the derived trends with distance due to these systematic effects, we reduce our sample to a narrower range of $T_{\rm eff}$, [4600 , 4800]\,K, where there is no bias toward higher [Fe/H] and no dependence of the chemical abundances with $T_{\rm eff}$ is observed. This reduced sample comprised 183 stars. The bottom panels in Fig. \ref{figure1} show the resulting trends inferred from this smaller group. Due to the lower number of stars, the dispersion in the resulting median trends is larger and the statistical results are less robust. However, the trends inferred are qualitatively the same as those derived from the whole sample, in particular the steeper decreasing trend observed in the most metal-rich bin. The fact that we detected the same patterns with $r$ using a sample where the chemical abundances do not show systematic correlations with stellar parameters suggests that these trends are real, and not an artifact. Figure \ref{histo_norm} shows the [X/Fe] distribution in each of the [Fe/H] ranges considered, and three radial bins, $5 < r < 10$ kpc, $10 < r < 15$ kpc, and $15 < r < 30$ kpc, normalised to the total number of stars in each radial bin. These trends, inferred from a larger number of halo stars and improved abundance determinations and distances, confirm the decrease in [$\alpha$/Fe] with $r$ for stars at [Fe/H] $> -1.1$ reported by \citet{fernandezalvar2017}. \subsection{Simulated chemical trends} \begin{figure} \includegraphics[scale=0.4]{OFE-binFEHAq-C-5.pdf} \includegraphics[scale=0.4]{OFE-binFEHAq-D-5.pdf} \caption{Median [O/Fe] values derived from the Aq-C (top panel) and the Aq-D (bottom panel) halos obtained from the Aquarius cosmological simulations, in the same [Fe/H] and $r$ bins, and using the same color coding, as the observations.} \label{figure2} \end{figure} We estimated the [O/Fe] profiles for the seven Aquarius haloes (A, B, C, D, F, G, H) analysed by \citet{tissera2012}, calculating medians in the same [Fe/H] and $r$ intervals used for the observations. We only considered stars located at $|h| > 5$ kpc, the same requirement imposed for the APOGEE observations. In the case of the simulations, $^{16}$O is our $\alpha$-element reference. After exploring the trends in all the haloes, we choose two of them to make a detailed analysis in relation to their assembly histories, Aq-C and Aq-D. The simulated halo that best represents the observational [O/Fe] and [Mg/Fe] trends shown in Fig. ~\ref{figure1} is Aq-C, which exhibits a decrease in these trends with $r$ for the most metal-rich bin, stars with [Fe/H]$ > -1.1$. On the contrary, the halo that differs the most from the observations is Aq-D, exhibiting a slightly increasing $\alpha$-enrichment with increasing radius. We show the trends for both haloes in Fig.\ref{figure2}. From these figures, we also note that, for the intermediate-[Fe/H] interval, Aq-D also exhibits an increase of [O/Fe], while Aq-C has a flat trend with radius. In the low-[Fe/H] interval, the trends with radius are flat for both simulated haloes. In order to further compare the observations and simulations, Fig.~\ref{histosimus} shows the [O/Fe] distributions for the defined radial intervals. At first sight, the distributions appear different from the observed ones shown in Fig.~\ref{histo_norm}\footnote{The sharp increase in [O/Fe] close to Solar values is produced by the enrichment from SNIa relative to SNII. This occurrs in the low-metallicity sub-samples were the star formation is starting, and the mixing of chemical elements in the ISM is more inhomogeneous. More efficient metal mixing will solve this numerical artefact.}. However, when examined in detail, some similarities emerge. In the case of Aq-C, the trends from the the inner to outer radii for the three metallicity sub-samples are similar: the inner radii are more dominated by $\alpha$-rich stars with respect to the $\alpha$-poor stars, principally in the outer radial interval. Althougth at all radii and metallicities, there is a larger contribution of $\alpha$-rich stars in Aq-C, the relative contribution of $\alpha$-poor stars increases remarkably for high metallicities in outer radial interval. It is this different relative contribution of $\alpha$-poor stas that results in a slope more comparable to the MW abundance patterns. A similar change in the relative contribution of $\alpha$-rich to $\alpha$-poor stars can be seen in Fig. \ref{histo_norm}. However, we acknowledge the fact that the observations show a sharper decrease of $\alpha$-rich and high-metallicity stars in the outer region than that predicted by Aq-C. Conversely, Aq-D does not exhibit such a significant contribution from $\alpha$-rich stars in the high-metallicity interval at any radius, because there is a larger contribution of $\alpha$-poor stars at all radii. In addition, the contribution of $\alpha$-rich stars increases for $r > 15$ kpc. This is why the median [O/Fe] in the central regions of Aq-D are lower than in the outer radial interval, producing the opposite trend to that found in Aq-C. Below we explore why this happens in Aq-D but not in Aq-C, the simulated halo that is most like the observed MW [O/Fe] pattern. \begin{figure*} \includegraphics[scale=0.8]{distrib_massfract_sim.pdf} \caption{Mass fractions of stars within the three metallicity intervals: $-2.5 <$ [Fe/H] $< -1.8$ (upper), $-1.8 <$ [Fe/H] $< -1.1$ (middle), and $-1.1 <$ [Fe/H] $< 0.0$ (lower panels) and in three radial intervals: $5 < r < 10$ kpc (blue lines), $10 < r < 15$ kpc (green lines), and $15 < r < 30$ kpc (red lines), for haloes Aq-C and Aq-D (left and right columns, respectively), normalized to the total mass in each bin of $r$ and [Fe/H]. } \label{histosimus} \end{figure*} \begin{figure*} \includegraphics[scale=0.8]{distrib_ofe_sim.pdf} \caption{Mass fractions of stars, F, at $-1.1 < $ [Fe/H]$ < 0.0$ in the three radial intervals: $5 < r < 10$ kpc (green lines), $10 < r < 15$ kpc (blue lines), and $15 < r < 30$ kpc (red lines), normalized to the total mass in this [Fe/H] bin, for haloes Aq-C and Aq-D (left and right columns, respectively), considering separately each sub-population with a different origin. } \label{figure3} \end{figure*} \subsection{The high-metallicity stellar population} To understand which stellar populations (according to their origin) are responsible of the [O/Fe] gradient in stars with $-1.1 < $ [Fe/H] $ < 0.0$, we examine the contribution of each sub-population according to their origin. We focus on this metallicity interval because it is the one which shows the strongest signature in the observations (see Fig.~\ref{figure1}). Figure \ref{figure3} shows the contribution of the stellar populations to the [O/Fe] distributions for those stars with $-1.1 < $ [Fe/H]$ < 0.0$, divided into the three defined radial intervals for the Aq-C and Aq-D simulated haloes. Each sub-sample has been normalised to the total stellar mass within a radial interval. This figure reveals that the Aq-C halo is characterised by a conspicuous fraction of accreted stars with high-[O/Fe] ($\sim +0.3$) for each $r$ interval. The contribution of in-situ stars is almost negligible compared with the contribution of accreted stars. The fraction of accreted stars with [O/Fe] $< +0.2$ increases with $r$, whereas that corresponding to high [O/Fe] decreases, causing the negative [O/Fe] gradient with $r$. Therefore, the stellar population responsible for the negative [O/Fe] gradient with distance is the large fraction of accreted stars with low $\alpha$-values. In Aq-D, there is smaller relative contribution of accreted high-[O/Fe] stars. There is also a larger contribution of disc-heated stars with low [O/Fe]. This fraction of in-situ stars decreases as $r$ increases, so the contribution of stars with high [O/Fe] increases with $r$. In addition, the distribution of [O/Fe] is broader, without a strong peak at high [O/Fe], as in the case of Aq-C. The relative fraction of high-[O/Fe] stars tend to increase with $r$. Consequently, the median [O/Fe] value for stars within this metallicity range is higher, as they are located farther away. In this simulated halo, the $\alpha$-poor stars are more concentrated in the inner regions. \subsection{The assembly history behind the $\alpha$-element trends} In the case of Aq-C, the distribution of [O/Fe], with a peak at high [O/Fe] $\sim +0.35$ at high metallicities ([Fe/H] $> -1.1$), implies that these stars should have formed from a short and intense burst of star formation. The ISM would have reached large [Fe/H] and [O/H], both mainly due to the contribution of SNeII. The stars contributing to the nearer regions ($r < 15$ kpc) should be old and contemporaneous, otherwise SNeIa would have had time to explode, enriching the ISM so that the subsequently born stars would have lower [O/Fe]. On the other hand, the increase of low [O/Fe] at more distant radii implies that these stars would be comparatively younger. The broader distribution of [O/Fe] displayed by the accreted sample in the halo Aq-D, with a more similar fraction of low-[O/Fe] stars with respect to high-[O/Fe] stars, indicates that they would have formed during an extended star-formation era, allowing for a significant fraction of stars to be formed from gas enriched by SNeIa. To further probe this hypothesis, we examine the age distribution of each stellar population. As can be seen from Fig. \ref{figure4}, the accreted stars in the halo Aq-C populating the nearer region are concentrated in an age range [$\sim11.5, 13.0$] Gyr. A secondary peak at $\sim 11$ Gyr, which would correspond to stars with lower [O/Fe] values, is considerably lower with respect to the rest of the distribution, but increases with increasing radii and extends towards lower ages. This age distribution explains the increment of low-[O/Fe] stars, which are indeed slightly younger stars, formed at an epoch when the ISM would have been already enriched by SNeIa. There is also a fraction of in-situ endo-debris stars peaked at $\sim 12$ Gyr for stars at $r < 10$ kpc, which diminishes at higher radii. Conversely, the distribution of stellar ages for the halo Aq-D peaks at slightly lower values, $\sim 11.5$ Gyr, and extends over the range [$\sim 9.5$, $\sim 13$] Gyr. These stars formed during a more extending period compared with Aq-C. From Fig.~\ref{figure5} we can also see that there are disc-heated stars with ages in the range [$\sim 7$, $\sim 9.5$] Gyr. In the intermediate radial range, there is also an significant contribution from disc-heated stars in this halo. The main difference between the age distributions of the high-metallicity stars in Aq-C and Aq-D is that the former has older stellar populations in the inner regions, and in the outer radial interval the age distributions are consistent with two star-formation episodes. This is not the case for Aq-D, which exhibits a more extended period of star formation in all three radial intervals. \begin{figure*} \includegraphics[scale=0.8]{distrib_age_sim.pdf} \caption{Distribution of simulated ages from the Aq-C and Aq-D haloes for each stellar population with different origin and [Fe/H] > -1.1. We show the distributions for each adopted radial interval as in Fig.~\ref{figure3}.} \label{figure4} \end{figure*} \subsection{Accreted satellites} To further understand the contributions of the accreted satellite galaxies to the formation of the inner regions of these haloes and, therefore, their role in setting the [O/Fe] profiles, Fig.~\ref{figure5} and Fig.~\ref{figure6} show the contribution of stars with different metallicities to the different radial intervals of the inner stellar haloes, by small (M $< 10^{9}$M$_{\odot}$) , intermediate ($10^{9}$M$_{\odot} <$ M $<10^{10}$M$_{\odot}$), and massive (M $>10^{10}$M$_{\odot}$) satellites (where M denotes the dynamical mass of the accreted satellites at the time it enters the virial radius of the main progenitor galaxy). As shown in Fig.~\ref{figure5}, Aq-C has a larger contribution from intermediate-mass satellites for the lower metallicity interval. The higher-metallicity accreted stars are formed in massive and intermediate-mass satellites. For increasing radius, the contributions from massive satellites increases. This is consistent with the previous trends showing these stars have lower [O/Fe] abundances. From figure 4 in \citet{tissera2014}, we know that the Aq-C inner region did not receive satellites larger than about $\sim 2\times10^{10}$M$_{\odot}$. Our analysis shows that three satellites, with dynamical masses not larger than $2\times 10^{10}$ M$_{\odot}$, contribute stars particularly in the outer radial interval. Regarding Aq-D, the trends in Fig.~\ref{figure6} show a different assembly history, with a more significant contribution coming from massive satellites, also in agreement with figure 4 in \citet{tissera2014}. These satellites are more massive, and contribute to all radial and metallicity intervals. These more massive satellite accretion affects the inner region and contributes less $\alpha$-enhanced stars to the inner radial bins. We identify contributions from up to four satellites with mean dynamical masses $\sim 2.5\times10^{10} $M$_{\odot}$ (i.e., in the range [$\sim 2.5$ , $\sim 4.3\times10^{10}$] M$_{\odot}$). For both Aq-C and Aq-D, these satellites did not deposit all of their stars within these regions. There are partial contributions as the satellites are disrupted. In the case of Aq-D, the more-massive satellites contribute $\sim 75\%$, $\sim 66\%$, and $\sim 56\%$ of the stars in the high-, intermediate-, and low-metallicity populations. Conversely, for Aq-C, the higher-mass satellites contribute $\sim 39\%$, $\sim 28\%$, and $\sim 21 \%$ for the same metallicity populations. However, this halo has the largest contributions coming from intermediate-mass satellites:$\sim 52\%$, $\sim 53\%$, and $\sim 47\%$ of the stars in the high-, intermediate-, and low-metallicity populations, respectively. \begin{figure*} \includegraphics[scale=0.8]{cum_msat_sim_C.pdf} \caption{Mass-weighted distributions of the accreted stellar populations of Aq-C, grouped according to the mass of their host satellites, normalized to each $r$ and [Fe/H] bin. Accreted stars are shown according to different metallicity intervals (left:high-metallicity, middle: intermediate-metallicity, right: low-metallicity), and different radial intervals (upper: inner regions, middle: intermediate regions, lower panels: outer regions). } \label{figure5} \end{figure*} \begin{figure*} \includegraphics[scale=0.8]{cum_msat_sim_D.pdf} \caption{Mass-weighted distributions of the accreted stellar populations of Aq-D, grouped according to the mass of their host satellites, normalized to each $r$ and [Fe/H] bin. Accreted stars are shown according to different metallicity intervals (left:high-metallicity, middle: intermediate-metallicity, right: low-metallicity), and different radial intervals (upper: inner regions, middle: intermediate regions, lower panels: outer regions).} \label{figure6} \end{figure*} In summary, Aq-C shows a similar [$\alpha$/Fe] trend as a function of radius to the observations; it has been already reported that a late-infalling massive satellite with an extended SFH is contributing significant mass at intermediate distances -- for a very detailed description of that satellite orbit see \citet{parry2012} and Appendix A1 of \citet{cooper2017}. Aq-D shows the opposite trend, supporting the idea that incomplete mixing of a massive progenitor is responsible for what is observed in the MW. In fact, considering recent observational results in the MW regarding the existence of remnants of accreted satellites in the stellar halo \citep{belokurov2018,helmi2018}, we carried out the exercise of selecting high-metallicity star particles with large radial velocities ($|v_{r}| > 150$ km/s) and small tangential ones $|v_{t}| < 30$ km/s, contributed by satellites of different masses in order to find out clues on their origins. We find that, in Aq-C, 50, 32 and 18 per cent of high-metallicity stars were contributed by massive, intermediate and small mass galaxies. While for Aq-D these percentages are 73, 19, and 7 per cent, respectively, in agreement with the trends shown in Fig.6 and Fig.7. It is clear that in both cases the majority of the high-metallicity stars are associated to massive satellites. However, in the case of Aq-D, the massive satellite clearly has the principal role. Nevertheless, this last halo do not have a distribution of $\alpha$-elements consistent with observations in the MW because most of the stars are $\alpha$-poor and they have been able to reach the center. So that the $\alpha$-profiles is inverted with respect to that of Aq-C. It is very interesting to take into account the results of \citet{cooper2017} that tracked the satellite in time and found it has gone several pass-bye in the last 8 Gyrs. We cannot prevent the obvious question: in the case of the MW, could be the satellite responsible of imprinting this $\alpha$-pattern still around \citep[e.g.][]{belokurov2018,koposov2018}? \section{Conclusions} The elemental-abundance ratios for the two $\alpha$-elements O and Mg reported in the DR14 APOGEE/APOGEE-2 database for relatively metal-rich halo stars ([Fe/H] $> -1.1$) at Galactocentric radii between 5 and $\sim$30 kpc exhibit a gradient with distance, which becomes flatter at lower metallicity. The simulated halo that best reproduces this observational feature is Aq-C. This halo exhibits a decreasing trend in [O/Fe] with $r$ for stars at [Fe/H] $> -1.1$ and flat trends at lower [Fe/H], for stars at $r > 5$ kpc, qualitatively similar to the [O/Fe] trends inferred from the APOGEE/APOGEE-2 DR14 observations. The results from the comparative analysis of Aq-C and Aq-D indicates that the decreasing [O/Fe] with radius for more metal-rich stars is due to the contribution of accreted stars from satellite galaxies with about $10^{10}$ M$\sun$ dynamical mass that did not get all way to the central region. The assembly history for Aq-C is characterised by the larger contribution from intermediate-mass satellites in the inner regions ($r < 15$ kpc) with old high-$\alpha$ stars, and the increase of the contribution from more-massive satellites, populating it with younger low-$\alpha$ stars, at $r > 15$ kpc. The large fraction of old high-[O/Fe] stars at $r < 15$ kpc is due to short and intense bursts of star formation, while the low-[O/Fe] stellar populations are associated with a later second burst. There is a clear gap between the two starbursts, which provides enough time for SNIa enrichment. Note that these two $\alpha$-enriched populations might have formed in different accreted satellites. An intermediate/massive satellite would be able to retain gas to have a second starburst, after the first one quenched star formation by heating up the ISM (Aq-C). However, if it is sufficiently massive, then the period of star formation would be more extended, and the satellite would also be able to reach the inner regions \citep{amorisco2017a,fattahi2018} as is the case in Aq-D. The resulting [O/Fe] patterns will be different: more enriched material would be able to reach the central region in the latter case, thus producing a positive increases of [O/Fe] with radius. Note that these satellites could also heat up the old disc, contributing with more disc-heated stars in the halo. Our results are in agreement with those recently reported in the literature pointing to the accretion of a relatively massive satellite. In particular, our results agree with \citet{mackereth2018} who analysed simulated galaxies from the EAGLE project and inferred a merger history for the Milky Way of satellites with stellar masses in the range $10^{8}$-$10^{9}$M$_{\odot}$. The comparative analysis between Aq-C and Aq-D we have carried out suggests that, in order to reproduce the observed $\alpha$-element patterns in the inner region of the MW, in particular the decrease of [O/Fe] for increasing radius, intermediate- and low-mass satellites would be expected to contribute to the very inner regions but a more-massive satellite (i.e., dynamical mass about $\sim 10^{10}$ M$_{\odot}$ at the time the satellite enters the virial radius of the MW progenitor) is expected to contribute low-$\alpha$ stars farther away. Considering our results and those of Cooper et al . (2017) we speculate that the surviving remnant of this massive satellite could be still orbiting the MW. \section*{Acknowledgements} We thank Andrew Cooper for his detailed comments. E.F.A. acknowleges financial support from the ANR 14-CE33-014-01, and partial support provided by CONACyT of M\'exico (grant 247132). P.B.T. acknowledges financial support from UNAB project D667/2015 and Fondecyt Regular 1150334. Part of the analysis was done in RAGNAR cluster of the Numerical Astrophysics group at UNAB. L.C. thanks for the financial supports provided by CONACyT of M \'exico (grant 241732), by PAPIIT of M \'exico (IG100115, IA101215, IA101517) and by MINECO of Spain (AYA2015-65205-P). This project has received funding from the European Union Horizon 2020 Research and Innovation Programme under the Marie Sklodowska-Curie grant agreement No 734374. T.C.B. acknowledges partial support from grant PHY 14-30152 (Physics Frontier Center/JINA-CEE) awarded by the U.S. National Science Foundation (NSF). Funding for the Sloan Digital Sky Survey IV has been provided by the Alfred P. Sloan Foundation, the U.S. Department of Energy Office of Science, and the Participating Institutions. SDSS-IV acknowledges support and resources from the Center for High-Performance Computing at the University of Utah. The SDSS web site is www.sdss.org. SDSS-IV is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS Collaboration including the Brazilian Participation Group, the Carnegie Institution for Science, Carnegie Mellon University, the Chilean Participation Group, the French Participation Group, Harvard-Smithsonian Center for Astrophysics, Instituto de Astrof\'isica de Canarias, The Johns Hopkins University, Kavli Institute for the Physics and Mathematics of the Universe (IPMU) / University of Tokyo, Lawrence Berkeley National Laboratory, Leibniz Institut f\"ur Astrophysik Potsdam (AIP), Max-Planck-Institut f\"ur Astronomie (MPIA Heidelberg), Max-Planck-Institut f\"ur Astrophysik (MPA Garching), Max-Planck-Institut f\"ur Extraterrestrische Physik (MPE), National Astronomical Observatories of China, New Mexico State University, New York University, University of Notre Dame, Observat\'ario Nacional / MCTI, The Ohio State University, Pennsylvania State University, Shanghai Astronomical Observatory, United Kingdom Participation Group, Universidad Nacional Aut\'onoma de M\'exico, University of Arizona, University of Colorado Boulder, University of Oxford, University of Portsmouth, University of Utah, University of Virginia, University of Washington, University of Wisconsin, Vanderbilt University, and Yale University. \bibliographystyle{mnras}

33b72ded31f83b126cb69fd3e4f402a70986a562

\section{Introduction} \label{S1} The quantification of risk is a central topic of study in finance. The need to guard against unforeseeable events with adverse and often catastrophic consequences has led to an extensive and burgeoning literature on risk measures. The two most popular financial risk measures, the value-at-risk (VaR) and the expected shortfall (ES), are extensively used for risk and portfolio management as well as regulation in the finance industry. Parallel to the developments in risk estimation, there has been longstanding interest in detecting changes in financial time series, especially changes in the tail structure, which is essential for effective risk and portfolio management. Indeed, empirical findings strongly suggest that financial time series frequently exhibit changes in their underlying statistical structure due to changes in economic conditions, e.g. monetary policies, or critical social events. Although there is an established literature on structural change detection for parametric time series models, including monitoring of proxies for risk such as tail index, there are no studies or tools concerning monitoring of general tail structure, or of risk measures in particular. To underscore this point, existing literature on risk measure estimation assume stationarity of time series observations over a time period of interest, with the stationarity of the risk measure being key for the estimation to make sense. However, to the best of our knowledge, no tools have been provided to verify this assumption. We provide tools to detect general and potentially multiple changes in the tail structure of time series, and in particular, tools for monitoring for changes in risk measures such as ES and related measures such as conditional tail moments (CTM) \citep{methni_etal2014} over time periods of interest. Specifically, we develop retrospective change-point tests to detect changes in ES and related risk measures. Additionally, we offer new ways of constructing confidence intervals for these risk measures. Our methods are applicable to a wide variety of time series models as they depend on functional central limit theorems for the risk measures, which we develop under weak assumptions. As will be described, our methods complement and extend the existing literature in several ways. As mentioned previously, the literature lacks tools to monitor general tail structure or risk measures such as ES or CTM over time. This deficiency appears to be two-fold. (1) Although there are studies on VaR change-point testing, for example, \cite{qu2008}, often one is interested in characterizations of tail structure that are more informative than simple location measures. Indeed, the introduction of ES as an alternative risk measure to VaR was, to a great extent, driven by the need to quantify tail structure, particularly, the expected magnitude of losses conditional on losses being in the tail. Aligned with this goal, a popular measure of tail structure is the tail index, which describes tail thickness and governs distributional moments. In tail index estimation, an extreme value theory approach is typically taken with the assumption of so-called regularly-varying Pareto-type tails. However, tail index estimation is very sensitive to the choice of which fraction of sample observations is classified to be ``in the tail''. \footnote{The Hill estimator \citep{hill1975} is widely used and requires the user to choose the fraction of sample observations deemed to be ``in the tail'' to use for estimation. However, generally, there appears to be no ``best'' way to select this fraction. In the specific setting of change-point testing for tail index, recommendations for this fraction in the literature range from the top 20th percentile to the top 5th percentile of observations \citep{kim_etal2009,kim_etal2011,hoga2017}. Such choice heavily influences the quality of tail index estimation and change detection and must be made on a case-by-case basis \citep{rocco2014}. It is a delicate matter as choosing too small of a fraction results in high estimation variance and choosing too large of a fraction often results in high bias due to misspecification of where the tail begins.} Moreover, the typical regularly-varying Pareto-type tail assumption may not even be valid in some situations. And even if they are valid, the tail index is invariant to changes of the location and scale types, and thus these types of structural changes in the tail would remain undetected using tail index-based change-point tests. (2) As mentioned before, all previous studies on risk measure estimation implicitly assume the risk measure is constant over some time period of interest---otherwise, risk measure estimation and confidence interval construction could behave erratically. For instance, if there is a sudden change in VaR (at some level) in the middle of a time series, naively estimating VaR using the entire time series could result in a wrong estimate of VaR or ES. Hence, given the importance of ES and related risk measures, a simple test for ES change over a time period of interest is a useful first step prior to follow-up statistical analysis. To simultaneously address both deficiencies, we propose retrospective change-point testing for ES and related risk measures such as CTM. We introduce, in particular, a consistent test for a potential ES change at an unknown time based on a variant of the widely used cumulative sum (CUSUM) process. \footnote{The cumulative sum (CUSUM) process was first introduced by \cite{page1954} and is discussed in detail in \cite{csorgo_etal2011}.} We subsequently generalize this test to the case of multiple potential ES change points, leveraging recent work by \cite{zhang_etal2018}, and unlike existing change-point testing methodologies in the literature, our test does not require the number of potential change points to be specified in advance. Our use of a risk measure such as ES is attractive in many ways. First, the fraction of observations used in ES estimation, for example, the upper 5th percentile of observations, directly has meaning, and is often comparable to the fraction of observations used in tail index estimation, as discussed previously. Second, use of ES does not require parametric-type tail assumptions, which is not the case with use of the tail index. Third, ES can detect much more general structural changes in the tail such as location or scale changes, while such changes go undetected when using tail index. Moreover, our change-point tests can be used to check in a statistically principled way whether or not ES and related risk measures are constant over a time period of interest, and provide additional validity when applying existing estimation methods for these risk measures. Additionally, a key detail that has been largely ignored in previous studies is standard error estimation. Here, the issue is two-fold. (1) In the construction of confidence intervals for risk measures, consistent estimation of standard errors is nontrivial due to standard errors involving the entire time series correlation structure at all integer lags, and thus being infinite-dimensional in nature. A few studies in which standard error estimation has been addressed (or bypassed) are \cite{chen_etal2005,chen2008,wang_etal2008,xu2016}. However, confidence interval construction in these studies all require user-specified tuning parameters such as the bandwidth in periodogram kernel smoothing or window width in the blockwise bootstrap and empirical likelihood methods. Although the choice of these tuning parameters significantly influences the quality of the resulting confidence intervals, it is not always clear how best to select them. (2) Choosing the tuning parameters is not only difficult, but data-driven approaches may result in non-monotonic test power, as pointed out in numerous studies \citep{vogelsang1999,crainiceanu_etal2007,deng_etal2008,shao_etal2010}.\footnote{For general change-point tests with time series observations, consistent estimation of standard errors is typically done using periodogram kernel smoothing, and the performance of such tests is heavily influenced by the choice of kernel bandwidth.} We offer the following solutions to the above issues. (1) To address the issue of often problematic selection of tuning parameters for confidence interval construction with time series data, we make use of ratio statistics to cancel out unknown standard errors and form pivotal quantities, thereby avoiding the often difficult estimation of such nuisance parameters. We examine confidence interval construction using a technique originally referred to in the simulation literature as sectioning \citep{asmussen_etal2007}, which involves splitting the data into equal-size non-overlapping sections, separately evaluating the estimator of interest using the data in each section, and relying on a normal approximation to form an asymptotically pivotal t-statistic. We also examine its generalization, referred to in the simulation literature as standardized time series \citep{schruben1983,glynn_etal1990} and in the time series literature as self-normalization \citep{lobato2001,shao2010}, which uses functionals different from the t-statistic to create asymptotically pivotal quantities. (2) In the context of change-point testing using ES and related risk measures, to avoid potentially troublesome standard error estimation, we follow \cite{shao_etal2010} and \cite{zhang_etal2018} and apply the method of self-normalization for change-point testing by dividing CUSUM-type processes by corresponding processes designed to cancel out the unknown standard error. The processes we divide by take into account potential change point(s) and thus avoids the problem of non-monotonic power which often plagues change-point tests that rely on consistent standard error estimation, as discussed previously. The outline of our paper is as follows. In Section \ref{S2p2}, we develop asymptotic theory for VaR and ES, specifically, functional central limit theorems, which provide the theoretical basis for the proposed confidence interval construction and change-point testing methodologies. In introducing our statistical methods, we first discuss the simpler task of confidence interval construction for risk measures in Section \ref{S3p1}. Then, with several fundamental ideas in place, we take up testing for a single potential change point for ES in Section \ref{S3p2}. We extend the change-point testing methodology to an unknown, possibly multiple, number of change points in Section \ref{S3p3}. In Section \ref{S4}, we examine the finite-sample performance of our proposed methods through simulations. We conclude with an empirical study in Section \ref{S5} of returns data for the S\&P 500 exchange-traded fund (ETF) SPY and also US 30-Year Treasury bonds. Proofs of our theoretical results are delegated to the Appendix. \section{Asymptotic Theory} \label{S2} \subsection{Model Setup} \label{S2p1} Suppose the random variable $X$ and the stationary sequence of random variables $X_1, \dots, X_n$ have marginal distribution function $F$. For some level $p \in (0,1)$, we wish to estimate the risk measures VaR and ES defined by \begin{eqnarray*} & VaR(p) = \inf \{ x \in \mathbb{R} : F(x) \ge p\} \\ & ES(p) = \Earg{X \mid X \ge VaR(p)}. \end{eqnarray*} Let $\hat{F}_n(\cdot) = n^{-1} \sum_{i=1}^n \indic{X_i \le \cdot}$ be the sample distribution function. We consider the following ``natural'' sample-based nonparametric estimators. \begin{eqnarray} \label{E1} & \widehat{VaR}_n(p) = \inf \{x \in \mathbb{R} : \hat{F}_n(x) \ge p\} \\ \label{E2} & \widehat{ES}_n(p) = \frac{1}{1-p} \frac{1}{n} \sum_{i=1}^n X_i \indic{X_i \ge \widehat{VaR}_n(p)}. \end{eqnarray} For $m > l$, we will also consider the following nonparametric estimators based on samples $X_l,\dots, X_m$, with $\hat{F}_{l:m}(\cdot) = (m-l+1)^{-1} \sum_{i=l}^m \indic{X_i \le \cdot}$. \begin{eqnarray} \label{E8} & \widehat{VaR}_{l:m}(p) = \inf \{x \in \mathbb{R} : \hat{F}_{l:m}(x) \ge p\} \\ \label{E9} & \widehat{ES}_{l:m}(p) = \frac{1}{1-p} \frac{1}{m-l+1} \sum_{i=l}^m X_i \indic{X_i \ge \widehat{VaR}_{l:m}(p)} \end{eqnarray} Also, let $\mathcal{F}_l^m$ denote the $\sigma$-algebra generated by $X_l, \dots, X_m$, and let $\mathcal{F}_l^\infty$ denote the $\sigma$-algebra generated by $X_l, X_{l+1}, \dots$. The $\alpha$-mixing coefficient introduced by \cite{rosenblatt1956} is $$\alpha(k) = \sup_{A \in \mathcal{F}_1^j, B \in \mathcal{F}_{j+k}^\infty, j \ge 1} \abs{\Parg{A}\Parg{B} - \Parg{AB}},$$ and a sequence is said to be $\alpha$-mixing if $\lim_{k \to \infty} \alpha(k) = 0$. The dependence described by $\alpha$-mixing is the least restrictive as it is implied by the other types of mixing; see \cite{doukhan1994} for a comprehensive discussion. In what follows, $D[0,1]$ denotes the space of real-valued functions on $[0,1]$ that are right-continuous and have left limits, and convergence in distribution on this space is defined with respect to the Skorohod topology \citep{billingsley1999}. Also, for some index set $\Delta$, $\ell^\infty(\Delta)$ denotes the space of real-valued bounded functions on $\Delta$, and convergence in distribution on this space is defined with respect to the uniform topology \citep{pollard1984}. We denote the integer part of a real number $x$ by $[x]$ and the positive part by $[x]_+$. Standard Brownian motion on the real line is denoted by $W$. In each of our theoretical results, we use one of the following two types of assumptions. \begin{assumption} \label{A1} There exists $a > 1$ such that the following hold. ~\begin{enumerate} \item[(i)] $X_1, X_2, \dots$ is $\alpha$-mixing with $\alpha(k) = O(k^{-a})$ \item[(ii)] $X$ has positive and continuous density $f$ in a neighborhood of $VaR(p)$, and for each $k \ge 2$, $(X_1,X_k)$ has joint density in a neighborhood of $(VaR(p),VaR(p))$. \end{enumerate} \end{assumption} \begin{assumption} \label{A2} There exists $\delta > 0$ such that Assumption \ref{A1} holds with $a > (2+\delta)/\delta$ along with $\Earg{\abs{X}^{2+\delta}} < \infty$. \end{assumption} Assumption \ref{A1} condition (i) is a form of ``asymptotic independence'', which ensures that the time series is not too serially dependent so that non-degenerate limit distributions are possible. A very wide range of commonly used financial time series models such as ARCH models and diffusion models satisfy this condition. Condition (ii) is a standard condition in VaR and ES estimation, and the joint density condition ensures that there are not too many ties among the time series observations. Assumption \ref{A2} indicates that the strength of the moment condition of the underlying marginal distribution and the rates of $\alpha$-mixing trade off, in that weaker $\alpha$-mixing conditions require stronger moment conditions and vice versa. We point out, in particular, that while our results are illustrated for the most widely used risk measures, VaR and ES, they can be adapted to many other important functionals of the underlying marginal distribution. One straightforward adaptation (by assuming stronger $\alpha$-mixing and moment conditions) is to CTM: $\Earg{X^\beta \mid X > VaR(p)}$ for some level $p \in (0,1)$ and some $\beta > 0$ \citep{methni_etal2014}. Our results also easily extend to multivariate time series, but for simplicity of illustration, we focus on univariate time series. \subsection{Functional Central Limit Theorems} \label{S2p2} We develop functional central limit theorems for Var and ES under weak assumptions. These functional central limit theorems allow the construction of general change point statistics. Our results essentially model the ``progress'' of the estimation (for fixed $n$, allowing $t$ to increase from zero to one) as a random walk, which for large $n$ and after rescaling the random walk steps by $n^{-1/2}$, can be approximated by a Brownian motion. \begin{thm} \label{T1} Under Assumption \ref{A1} with the modifications: $a \ge 3$ and $X$ has positive and differentiable density at $VaR(p)$, the process \begin{align} \label{E3} \{n^{1/2}t(\widehat{VaR}_{[nt]}(p) - VaR(p)), \: t \in [0,1]\} \end{align} converges in distribution in $D[0,1]$ to $\sigma_{VaR} W$, where $$\sigma_{VaR}^2 = \frac{1}{f^2(VaR(p))}\left(\Earg{g_p^2(X_1)} + 2 \sum_{i=2}^\infty \Earg{g_p(X_1)g_p(X_i)}\right)$$ with $g_p(X) = \indic{X \le VaR(p)} - p$. Under Assumption \ref{A2}, the process \begin{align} \label{E4} \left\{n^{1/2}t(\widehat{ES}_{[nt]}(p) - ES(p)), \: t \in [0,1]\right\} \end{align} converges in distribution to $\sigma_{ES} W$ in $D[0,1]$, where $$\sigma_{ES}^2 = \frac{1}{(1-p)^2}\left(\Earg{h_p^2(X_1)} + 2 \sum_{i=2}^\infty \Earg{h_p(X_1)h_p(X_i)}\right)$$ with $h_p(X) = \max(X, VaR(p)) - \Earg{\max(X, VaR(p))}$. \end{thm} From this, we immediately have the following central limit theorems for VaR and ES. \begin{cor} \label{C2} Under Assumption \ref{A1}, $$n^{1/2}(\widehat{VaR}_n(p) - VaR(p)) \overset{d}{\to} \mathcal{N}(0,\sigma_{VaR}^2),$$ where $\sigma_{VaR}^2$ is the same as in Theorem \ref{T1} for VaR. Under Assumption \ref{A2}, $$n^{1/2} (\widehat{ES}_n(p) - ES(p)) \overset{d}{\to} \mathcal{N}(0,\sigma_{ES}^2),$$ where $\sigma_{ES}^2$ is the same as in Theorem \ref{T1} for ES. \end{cor} We point out that our Assumptions $\ref{A1}$ and $\ref{A2}$ are weaker than those of existing central limit theorems in the literature (c.f. \cite{chen_etal2005,chen2008}, who require $\alpha$-mixing coefficients to decay exponentially fast as well as additional regularity of the marginal and pairwise joint densities of the observations). The above functional central limit theorem for VaR is a modification of a result due to \cite{sen1972}. The condition $a \ge 3$ for the VaR functional central limit theorem is due to our use of a so-called Bahadur representation for $\widehat{VaR}_n(p)$ by \cite{wendler2011}. Although such a condition can likely be weakened, we do not pursue that here. Moreover, the above functional central limit theorem for ES does not require such a condition; Assumption \ref{A2} is all that is needed. In deriving the ES functional central limit theorem, we used the following Bahadur representation for $\widehat{ES}_n(p)$. \begin{prop} \label{P1} Under Assumption \ref{A2}, for any $\delta' > 0$ satisfying $-1/2 + 1/(2a) + \delta' < 0$, $$\widehat{ES}_n(p) - \biggl(VaR(p) + \frac{1}{1-p} \frac{1}{n} \sum_{i=1}^n [X_i - VaR(p)]_+ \biggr) = o_{a.s.}(n^{-1 + 1/(2a) + \delta'} \log n).$$ \end{prop} \cite{sun_etal2010} developed such a Bahadur representation in the setting of independent, identically-distributed data, but to the best of our knowledge, no such representation exists in the stationary, $\alpha$-mixing setting. Such a Bahadur representation is generally useful for developing limit theorems in many different contexts. We also have the following extension of the functional central limit theorems for VaR and ES (Theorem \ref{T1}), where we have two ``time'' indices instead of a single ``time'' index as in the standard functional central limit theorems. The standard functional central limit theorems are useful for detecting a single change point in a time series, but the following extension will allow us to detect an unknown, possibly multiple, number of change points, as we will discuss later. We point out that this result does not follow automatically from Theorem \ref{T1} and an application of the continuous mapping theorem because the estimators in Equations \ref{E8} and \ref{E9} are not additive, for instance, for $m > l$, $\widehat{ES}_{l:m}(p) \ne \widehat{ES}_{1:m}(p) - \widehat{ES}_{1:l-1}(p)$. \begin{thm} \label{T3} Fix any $\delta > 0$ and consider the index set $\Delta = \{ (s,t) \in [0,1]^2 : t-s \ge \delta \}$. Under Assumption \ref{A1}, the process \begin{align} \label{E14} \{n^{1/2}(t-s)(\widehat{VaR}_{[ns]:[nt]}(p) - VaR(p)), \: (s,t) \in \Delta \} \end{align} converges in distribution in $\ell^\infty(\Delta)$ to $\{\sigma_{VaR} (W(t) - W(s)), \: (s,t) \in \Delta \}$, where $\sigma_{VaR}^2$ is the same as in Theorem \ref{T1} for VaR. Under Assumption \ref{A2} with the modification that $(X_1, X_k)$ has a joint density for all $k \ge 2$, the process \begin{align} \label{E15} \{n^{1/2}(t-s)(\widehat{ES}_{[ns]:[nt]}(p) - ES(p)), \: (s,t) \in \Delta \} \end{align} converges in distribution in $\ell^\infty(\Delta)$ to $\{\sigma_{ES} (W(t) - W(s)), \: (s,t) \in \Delta \}$, where $\sigma_{ES}^2$ is the same as in Theorem \ref{T1} for ES. \end{thm} \section{Statistical Inference} \subsection{Confidence Intervals} \label{S3p1} In time series analysis, confidence interval construction for an unknown quantity is often difficult, due to dependence. Indeed, from Theorem \ref{T1} and Corollary \ref{C2}, we see that the standard errors appearing in the normal limiting distributions depend on the time series autocovariance at all integer lags. To construct confidence intervals using the central limit theorem in Corollary \ref{C2} directly, these standard errors must be estimated. One approach, taken in \cite{chen_etal2005,chen2008}, is to estimate using kernel smoothing the spectral density at zero frequency of the transformed time series $g_p(X_1),g_p(X_2),\dots$ and $h_p(X_1),h_p(X_2),\dots$, where $g_p$ and $h_p$ are from Theorem \ref{T1} and Corollary \ref{C2}. Although it is known that under certain moment and correlation assumptions, spectral density estimators are consistent for stationary processes \citep{brockwell_etal1991,anderson1971}, in practice it is nontrivial to obtain quality estimates due to the need to select tuning parameters for the kernel smoothing-based approach. As for other approaches, under certain conditions, resampling methods such as the moving block bootstrap \citep{kunsch1989,liu_etal1992} and the subsampling method for time series \citep{politis_etal1999} bypass direct standard error estimation and yield confidence intervals that asymptotically have the correct coverage probability. However, these approaches also require user-chosen tuning parameters such as block length in the moving block bootstrap and window width in subsampling. We therefore investigate other ways of obtaining pivotal limiting distributions to construct asymptotic confidence intervals, where the quality of the obtained confidence intervals is less sensitive to the choice of tuning parameters and are thus, in some sense, more robust. We first examine a technique called sectioning from the simulation literature \citep{asmussen_etal2007}, which can be used to construct confidence intervals for risk measures in time series. Although the method can be used to construct confidence intervals for general risk measures, in light of our results from Section \ref{S2}, we apply the method to VaR and ES in particular. As a general overview of the method, let $Y_1(\cdot), Y_2(\cdot), \dots$ be a sequence of random bounded real-valued functions on $[0,1]$. For some user-specified integer $m \ge 2$, suppose we have the joint convergence in distribution \begin{align} \label{E5} \biggl(Y_n(1/m) - Y_n(0), Y_n(2/m) - Y_n(1/m), \dots, Y_n(1) - Y_n((m-1)/m) \biggr) \overset{d}{\to} \frac{\sigma}{m^{1/2}} \biggl( \mathcal{N}_1, \mathcal{N}_2, \dots, \mathcal{N}_m \biggr), \end{align} where $\sigma > 0$ and $\mathcal{N}_1, \dots, \mathcal{N}_m$ are independent standard normal random variables. Then, with \begin{align*} \bar{Y}_n &= m^{-1} \sum_{i=1}^m (Y_n(i/m) - Y_n((i-1)/m)) \\ S_n &= \left((m-1)^{-1} \sum_{i=1}^m (Y_n(i/m) - Y_n((i-1)/m) - \bar{Y}_n)^2 \right)^{1/2}, \end{align*} by the continuous mapping theorem, as $n \to \infty$ with $m$ fixed, $m^{1/2}\bar{Y}_n/S_n$ converges in distribution to the Student's $t$-distribution with $m-1$ degrees of freedom. With the limiting Student's $t$-distribution, we may construct confidence intervals for VaR or ES by taking the random functions $Y_n(\cdot)$ to be the processes in Equations \ref{E3} or \ref{E4}, respectively. The distributional convergence result in Equation \ref{E5} is easily obtained by modifying the proof of Corollary \ref{C2} or by directly applying Theorem \ref{T1}. Next, we examine a generalization of sectioning, called self-normalization, which has been studied recently in the time series literature \citep{lobato2001,shao2010} (and earlier in the simulation literature, where it is known as standardization \citep{schruben1983,glynn_etal1990}). As with sectioning, the method applies to confidence interval construction for general risk measures, but we specialize to the case of VaR and ES. In self-normalization, the idea is to use a ratio-type statistic where the unknown standard error appears in both the numerator and the denominator and thus cancels, resulting in a pivotal limiting distribution. Once again, let $Y_1(\cdot), Y_2(\cdot), \dots$ be a sequence of random bounded real-valued functions on $[0,1]$. Suppose we have the distributional convergence in $D[0,1]$ (for example, via a functional central limit theorem): $Y_n(\cdot) \overset{d}{\to} \sigma W(\cdot)$, where $\sigma > 0$. For some positive homogeneous functional $T : D[0,1] \mapsto \mathbb{R}$ (i.e., $T(\sigma Y) = \sigma T(Y)$ for $\sigma > 0$ and $Y \in D[0,1]$) to which the continuous mapping theorem applies, we get \begin{eqnarray} \label{E6} \frac{Y_n(1)}{T(Y_n)} \overset{d}{\to} \frac{W(1)}{T(W)}, \end{eqnarray} which is a pivotal limiting distribution whose critical values may be computed via simulation and are tabulated in \cite{lobato2001}. As before, we may construct confidence intervals for VaR or ES by taking the random functions $Y_n(\cdot)$ to be the processes in Equations \ref{E3} or \ref{E4}, respectively. The distributional convergence result in Equation \ref{E6} follows directly from Theorem \ref{T1}. As an example, considering the functional $T(Y) = \left(\int_0^1(Y(t) - tY(1))^2 dt\right)^{1/2}$, we have the following result for ES. \begin{eqnarray} \label{E7} \frac{\widehat{ES}_n(p) - ES(p)}{\left(\int_0^1 t^2 \left(\widehat{ES}_{[nt]}(p) - \widehat{ES}_n(p)\right)^2 dt\right)^{1/2}} \overset{d}{\to} \frac{W(1)}{(\int_0^1 (W(t) - t W(1))^2 dt)^{1/2}}. \end{eqnarray} \subsection{Change-Point Testing} \label{S3p2} As is the case with confidence interval construction with time series data, change-point testing in time series based on statistics constructed from functional central limit theorems and the continuous mapping theorem is often nontrivial due to the need to estimate standard errors such as $\sigma$ in Theorem \ref{T1}. Motivated by the maximum likelihood method in the parametric setting, variants of the \cite{page1954} CUSUM statistic are commonly used for nonparametric change-point tests \citep{csorgo_etal2011}, and generally rely on asymptotic approximations (via functional central limit theorems and the continuous mapping theorem) to supply critical values of pivotal limiting distributions under the null hypothesis of no change. As discussed in \cite{vogelsang1999,shao_etal2010,zhang_etal2018}, testing procedures where standard errors are estimated directly, for example, by estimating the spectral density of transformed time series via a kernel-smoothing approach, can be biased under the change-point alternative. Such bias can result in nonmonotonic power, i.e., power can decrease in some ranges as the alternative deviates from the null. To avoid this issue, \cite{shao_etal2010} and \cite{zhang_etal2018} propose using self-normalization techniques to general change-point testing. We adopt this idea as discussed in Section \ref{S3p1} to our specific problem of detecting changes in tail risk measures. As motivated in the Introduction (Section \ref{S1}), it is important to perform hypothesis tests for abrupt changes of risk measures in the time series setting. We introduce the methodology for ES, but note that it can be applied to risk measures in general. For simplicity, we consider the case of at most one change point; the case of a unknown, potentially multiple, number of change points can be treated using the approach in \cite{zhang_etal2018}. For time series sample $X_1, \dots, X_n$, let $ES_{X_i}(p)$ be the ES at level $p$ for the marginal distribution of $X_i$. We test the following null and alternative hypotheses. \begin{align*} & \mathcal{H}_0 : X_1, \dots, X_n \text{ is stationary, and in particular, } ES_{X_1}(p) = \dots = ES_{X_n}(p) \\ & \mathcal{H}_1 : \text{There is } t^* \in (0,1) \text{ such that } ES_{X_1}(p) = \dots = ES_{X_{[nt^*]}}(p) \ne ES_{X_{[nt^*]+1}}(p) = \dots = ES_{X_n}(p) \\ & \qquad \text{ and } X_1, \dots, X_{[nt^*]} \text{ and } X_{[nt^*]+1}, \dots, X_n \text{ are separately stationary} \end{align*} We base our change-point test on the following variant of the CUSUM process. \begin{align} \label{E10} \left\{n^{1/2}t(1-t)\left( \widehat{ES}_{1:[nt]}(p) - \widehat{ES}_{[nt]+1:n}(p) \right), \: t \in [0,1] \right\} \end{align} Note that we split the above process for all possible break points $t \in (0,1)$ into a difference of two ES estimators, an estimator using samples $X_1, \dots, X_{[nt]}$ and an estimator using samples $X_{[nt]+1}, \dots, X_n$. As the ES estimator in Equation \ref{E9} involves the VaR estimator in Equation \ref{E8}, splitting the process as in Equation \ref{E10} avoids potential VaR estimation using a sample sequence containing the change point, which could have undesirable behavior. In Proposition \ref{P3} below, we self-normalize the process in Equation \ref{E9} using the approach of \cite{shao_etal2010}. Note the denominator of Equation \ref{E11} below takes into account the potential change point and is split into two separate integrals involving samples $X_1, \dots, X_{[nt]}$ and $X_{[nt]+1}, \dots, X_n$. \begin{prop} \label{P3} Assume Assumption \ref{A2} holds. Under the null hypothesis $\mathcal{H}_0$ \small \begin{align} \label{E11} G_n := \sup_{t \in [0,1]} \frac{t^2(1-t)^2\left(\widehat{ES}_{1:[nt]}(p) - \widehat{ES}_{[nt]+1:n}(p) \right)^2}{n^{-1}\sum_{i=1}^{[nt]} (\frac{i}{n})^2 \left(\widehat{ES}_{1:i}(p) - \widehat{ES}_{1:[nt]}(p) \right)^2 + n^{-1}\sum_{i=[nt]+1}^n (\frac{n-i+1}{n})^2 \left(\widehat{ES}_{i:n}(p) - \widehat{ES}_{[nt]+1:n}(p) \right)^2} \end{align} \normalsize converges in distribution to \small \begin{eqnarray} \label{E12} G := \sup_{t \in [0,1]} \frac{(W(t) - tW(1))^2}{\int_0^t \left(W(s) - \frac{s}{t}W(t) \right)^2 ds + \int_t^1 \left(W(1) - W(s) - \frac{1-s}{1-t}(W(1) - W(t))\right)^2 ds}. \end{eqnarray} Assume the alternative hypothesis $\mathcal{H}_1$ is true with the change point occurring at some fixed (but unknown) $t^* \in (0,1)$. For any fixed difference $ES_{X_{[nt^*]}}(p) = c_1 \ne c_2 = ES_{X_{[nt^*]+1}}(p)$, we have $G_n \overset{P}{\to} \infty$ as $n \to \infty$. Furthermore, if the difference varies with $n$ according to $c_1 - c_2 = n^{-1/2 + \epsilon} L$ for some $L \ne 0$ and $\epsilon \in (0,1/2)$, then $G_n \overset{P}{\to} \infty$ as $n \to \infty$. \normalsize \end{prop} The distribution of $G$ is pivotal, and its critical values may be obtained via simulation and are tabulated in \cite{lobato2001}. For testing $\mathcal{H}_0$ versus $\mathcal{H}_1$ at some level, we reject $\mathcal{H}_0$ if the test statistic $G_n$ exceeds some corresponding critical value of $G$. In subsequent discussions concerning Equation \ref{E11}, we will refer to the process appearing in the numerator as the ``CUSUM process'', the process appearing in the denominator as the ``self-normalizer process'', and the entire ratio process as the ``self-normalized CUSUM process''. \footnote{To theoretically evaluate the efficiency of statistical tests, an analysis based on sequences of so-called local limiting alternatives (for example, sequences $c_1 - c_2 = O(n^{-1/2})$ in Proposition \ref{P3} above) can be considered (see, for example, \cite{vandervaart1998}). However, such an analysis would be considerably involved, and we save it for future study.} \subsection{Extension to Multiple Change Points} \label{S3p3} We extend our single change-point testing methodology to the case of multiple change points. Typically, the number of potential change points in the alternative hypothesis must be prespecified. However, we leverage the recent work of \cite{zhang_etal2018} and introduce change-point tests of ES and related risk measures that can accommodate an unknown, possibly multiple, number of change points in the alternative hypothesis. For illustration, we introduce the method using ES for univariate time series, but the method extends easily to related risk measures and also to multivariate time series. We fix some small $\delta > 0$ and consider the following null and alternative hypotheses (following the notation from Section \ref{S3p2}). \begin{align*} & \mathcal{H}_0 : X_1, \dots, X_n \text{ is stationary, and in particular, } ES_{X_1}(p) = \dots = ES_{X_n}(p) \\ & \mathcal{H}_1 : \text{There are } 0 = t_0^* < t_1^* < \dots < t_k^* < t_{k+1}^* = 1 \text{ with } t_{j}^* - t_{j-1}^* > \delta \text{ for } j=1,\dots,k+1 \\ & \qquad \text{ such that } ES_{X_{[nt_j^*]}}(p) \ne ES_{X_{[nt_j^*]+1}}(p) \text{ and } X_{[nt_j^*]+1}, \dots, X_{[nt_{j+1}^*]} \text{ are separately stationary} \\ & \qquad \text{ for } j=0,\dots,k \end{align*} Consider the index set $\Delta = \{ (s,t) \in [\delta,1-\delta]^2 : t-s \ge \delta \}$ and the test statistic \begin{align*} H_n & = \sup_{(s_1,s_2) \in \Delta} \frac{C_{n,f}(s_1,s_2)}{D_{n,f}(s_1,s_2)} + \sup_{(t_1,t_2) \in \Delta} \frac{C_{n,b}(t_1,t_2)}{D_{n,b}(t_1,t_2)}, \end{align*} where \begin{align*} C_{n,f}(s_1,s_2) =& \frac{[ns_1]^2([ns_2] - [ns_1])^2}{[ns_2]^3} \left(\widehat{ES}_{1:[ns_1]}(p) - \widehat{ES}_{[ns_1]+1:[ns_2]}(p)\right)^2 \\ D_{n,f}(s_1,s_2) =& \sum_{i=1}^{[ns_1]} \frac{i^2 ([ns_1] - i)^2}{[ns_2]^2 [ns_1]^2} \left(\widehat{ES}_{1:i}(p) - \widehat{ES}_{i+1:[ns_1]}(p)\right)^2 \\ &+ \sum_{i=[ns_1]+1}^{[ns_2]} \frac{(i - 1 - [ns_1])^2 ([ns_2] - i + 1)^2}{[ns_2]^2 ([ns_2] - [ns_1])^2} \left(\widehat{ES}_{[ns_1]+1:i-1}(p) - \widehat{ES}_{i:[ns_2]}(p)\right)^2 \\ C_{n,b}(t_1,t_2) =& \frac{([nt_2] - [nt_1])^2(n-[nt_2]+1)^2}{(n-[nt_1]+1)^3} \left(\widehat{ES}_{[nt_2]:n}(p) - \widehat{ES}_{[nt_1]:[nt_2]-1}(p)\right)^2 \\ D_{n,b}(t_1,t_2) =& \sum_{i=[nt_1]}^{[nt_2]-1} \frac{(i-[nt_1]+1)^2 ([nt_2]-1-i)^2}{(n-[nt_1]+1)^2([nt_2]-[nt_1])^2} \left(\widehat{ES}_{[nt_1]:i}(p) - \widehat{ES}_{i+1:[nt_2]-1}(p)\right)^2 \\ &+ \sum_{i=[nt_2]}^{n} \frac{(i-[nt_2])^2 (n-i+1)^2}{(n-[nt_1]+1)^2 (n-[nt_2]-1)^2} \left(\widehat{ES}_{i:n}(p) - \widehat{ES}_{[nt_2]:i-1}(p)\right)^2. \end{align*} Then, under $\mathcal{H}_0$, applying Theorem 3 of \cite{zhang_etal2018}, our Theorem \ref{T3} above yields \begin{cor} Assume Assumption \ref{A2} holds. Under the null hypothesis $\mathcal{H}_0$, it holds \begin{align*} H_n \overset{d}{\to} \sup_{(s_1,s_2) \in \Delta} \frac{C(0,s_1,s_2)}{D(0,s_1,s_2)} + \sup_{(t_1,t_2) \in \Delta} \frac{C(t_1,t_2,1)}{D(t_1,t_2,1)} := H, \end{align*} where \begin{align*} C(r_1,r_2,r_3) =& \frac{1}{(r_3 - r_1)^2} \left(W(r_2) - W(r_1) - \frac{r_2 - r_1}{r_3 - r_1}(W(r_3) - W(r_1)) \right)^2 \\ D(r_1,r_2,r_3) =& \frac{1}{(r_3 - r_1)^2} \Biggl( \int_{r_1}^{r_2} \left[ W(s) - W(r_1) - \frac{s - r_1}{r_2 - r_1}(W(r_2) - W(_1)) \right]^2 ds \qquad \qquad \qquad \\ &+ \int_{r_2}^{r_3} \left[ W(r_3) - W(s) - \frac{r_3 - s}{r_3 - r_2}(W(r_3) - W(r_2))\right]^2 ds \Biggr). \end{align*} Under the alternative $\mathcal{H}_1$, $H_n \overset{P}{\to} \infty$ as $n \to \infty$. \end{cor} We reject $\mathcal{H}_0$ if $H_n$ exceeds the critical value corresponding to a desired test level of the pivotal quantity $H$, which may be obtained via simulation. Moreover, under $\mathcal{H}_1$, our test is asymptotically consistent, and the analogous version of Proposition \ref{P3} holds. To reduce the computational burden of the method, we use a grid approximation suggested by \cite{zhang_etal2018}, where in the doubly-indexed set $\Delta$, one index is reduced to a coarser grid. Specifically, let $\mathcal{G}_\delta = \{(1+k\delta)/2 : k \in \mathbb{Z}\} \cap [0,1]$ and consider the modified statistic \begin{align} \label{E15} \tilde{H}_n & = \sup_{(s_1,s_2) \in \Delta \cap ([0,1] \times \mathcal{G}_\delta)} \frac{C_{n,f}(s_1,s_2)}{D_{n,f}(s_1,s_2)} + \sup_{(t_1,t_2) \in \Delta \cap (\mathcal{G}_\delta \times [0,1])} \frac{C_{n,b}(t_1,t_2)}{D_{n,b}(t_1,t_2)}. \end{align} As before, under $\mathcal{H}_0$, we have \begin{align} \label{E16} \tilde{H}_n \overset{d}{\to} \sup_{(s_1,s_2) \in \Delta \cap ([0,1] \times \mathcal{G}_\delta)} \frac{C(0,s_1,s_2)}{D(0,s_1,s_2)} + \sup_{(t_1,t_2) \in \Delta \cap (\mathcal{G}_\delta \times [0,1])} \frac{C(t_1,t_2,1)}{D(t_1,t_2,1)} := \tilde{H}. \end{align} Note that simply using the original doubly-indexed set $\Delta$, for a sample of size $n$ and an arbitrary number of change points, we would need to search for maxima over $O(n^2)$ points. However, using the grid approximation, we need only search for maxima over $O(n)$ points. In contrast, if we were to use a direct extension of the single-change point detection methodology of Section \ref{S3p2}, with $m$ change points (which needs to be specified in advance), we would need to search for maxima over $O(n^m)$ points. Hence, the methodology introduced in this section offers significant computational savings. To obtain critical values for hypothesis testing, we simulate standard Brownian motion paths on $[0,1]$, with each path consisting of appropriately scaled 5,000 independent standard normal random variables. Figure \ref{F11} below shows the approximate distribution based on 10,000 samples of the test statistic from Equation \ref{E16}, and the estimated 0.95 quantile is 138.19. \begin{figure}[H] \centering \minipage{0.45\textwidth} \includegraphics[width=\linewidth]{figure11.jpg} \endminipage \caption{Estimated density of the test statistic from Equation \ref{E16} based on 10,000 samples, with each sample utilizing 5,000 independent standard normal random variables to approximate standard Brownian motion on $[0,1]$. The 0.95 quantile is 138.19 and is indicated by the vertical red dashed line.} \label{F11} \end{figure} \section{Simulations} \label{S4} We perform a simulation study to investigate the finite sample performance of ES confidence interval construction using the sectioning and self-normalization methods (Section \ref{S3p1}) as well as upper tail change detection (Section \ref{S3p2}) using ES. We consider two data generating processes, AR(1): $X_{i+1} = \phi X_i + \epsilon_i$ and ARCH(1): $X_{i+1} = \sqrt{\beta + \lambda X_i^2}\epsilon_i$. We take the innovations $\epsilon_i$ to be i.i.d. standard normal, and we use parameters $\phi = 0.5$, $\beta = 1$ and $\lambda = 0.3$. The stationary distribution of the AR(1) process is mean-zero normal with variance $1/(1-\phi^2)$. According to \cite{embrechts_etal1997}, the above choice of parameters for the ARCH(1) process yields a stationary distribution $F$ with right tail $1-F(x) \sim x^{-8.36}$ as $x \to \infty$. \subsection{Confidence Intervals} \label{S4p1} In Figure \ref{F1} below, we vary the time series sample size from 200 to 2000 and examine the widths and empirical coverage probability of the 95\% confidence intervals for ES at the 0.95 level (in the upper tail past the 95th percentile) produced by the sectioning and self-normalization methods (Section \ref{S3p1}) for the AR(1) and ARCH(1) processes introduced above. Each data point is the averaged result over 10,000 replications. For each replication, we initialize the AR(1) process with its stationary distribution, and for the ARCH(1) process we use a burn-in period of 5,000 to approximately reach stationarity. Generally, the sectioning method produces smaller confidence intervals, but yields lower empirical coverage probability compared to the self-normalization method. This is especially pronounced for small sample sizes such as 200 or 400. As sample size increases, the performance of the two methods becomes more similar, and the desired coverage probability of 0.95 is approximately reached. \begin{figure}[H] \centering \minipage{0.45\textwidth} \includegraphics[width=\linewidth]{figure1a.jpg} \endminipage \minipage{0.45\textwidth} \includegraphics[width=\linewidth]{figure1b.jpg} \endminipage \caption{Left: relationship between sample size and empirical coverage probability of 95\% confidence intervals for ES at the 0.95 level computed for stationary AR(1) and ARCH(1) processes. Right: relationship between time series sample size and width of 95\% confidence intervals for ES at the 0.95 level computed for stationary AR(1) and ARCH(1) processes. Each plotted point is the averaged result over 10,000 replications.} \label{F1} \end{figure} To further examine the finite-sample performance of the sectioning method, in Figure \ref{F2} below, we show normalized histograms (with total area one) of the pivotal t-statistics formed when using the method with $m=10$ sections, as discussed following Equation \ref{E5}. We show histograms for different time series sample sizes (200-1,200). Each histogram uses 10,000 t-statistics. We compare the histograms with the density of a t-distribution with 9 degrees of freedom, which according to our theory, is the asymptotic (in the sense of time series sequence length tending to infinity) sampling distribution of the t-statistics formed when using the sectioning method with 10 sections. In accordance with our theory, for both the AR(1) and ARCH(1) processes, the sampling distribution of the t-statistics are well approximated by the asymptotic sampling distribution, even at small time series sample sizes. However, there is a small, but noticeable bias at small sample sizes where the sampling distribution of the t-statistics appears to be shifted to the left of the asymptotic sampling distribution. This bias becomes less noticeable as sample size increases. \begin{figure} \centering \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{arch1.jpg} \endminipage \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{arch2.jpg} \endminipage \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{arch3.jpg} \endminipage \\ \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{arch4.jpg} \endminipage \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{arch5.jpg} \endminipage \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{arch6.jpg} \endminipage \\ \centering \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{ar1.jpg} \endminipage \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{ar2.jpg} \endminipage \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{ar3.jpg} \endminipage \\ \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{ar4.jpg} \endminipage \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{ar5.jpg} \endminipage \minipage{0.25\textwidth} \includegraphics[width=\linewidth]{ar6.jpg} \endminipage \caption{``Goodness-of-fit'' in constructing confidence intervals for ES at the 0.95 level for (1) stationary AR(1) process (top 6 histograms) and (2) approximately stationary ARCH(1) process (bottom 6 histograms) using the sectioning method with 10 sections. Each normalized histogram: 10,000 pivotal t-statistics resulting from the sectioning method with 10 sections. ``Samples'' refers to the sample sizes of the time series used. Red curve: density of t-distribution with 9 degrees of freedom.} \label{F2} \end{figure} \subsection{Detection of Location Change in Tail} \label{S4p2} We also investigate through simulations the detection of abrupt location changes using ES at the 0.9 level (in the upper tail past the 90th percentile), as discussed in Section \ref{S3p2}. We consider the same AR(1) and ARCH(1) processes introduced previously. Figure \ref{F3} below shows the approximate power of change-point tests using the self-normalized CUSUM statistic (Equations \ref{E11} and \ref{E12}) at the 0.05 significance level as the magnitude of the abrupt location change varies between 0 (the null hypothesis of no change) and 3. For each data point, we perform 1,000 replications of change-point testing using times series sequences of length 400 with the potential changes occurring in the middle of the sequences. For each replication, we initialize the AR(1) process with its stationary distribution, and for the ARCH(1) process we use a burn-in period of 5,000 to approximately reach stationarity. As expected, for both processes, the power is a monotonic function of the magnitude of location change, and moreover the power curves are very similar and almost coincide. In accordance with the desired 0.05 significance level of our procedure, for both processes, the probability of false positive detection is approximately 0.05 (0.044 and 0.042 for the AR(1) and ARCH(1) processes, respectively, as indicated by the points with zero magnitude of location change in Figure \ref{F3}). \begin{figure}[H] \centering \includegraphics[width = 0.7\textwidth]{figure3.jpg} \caption{Relationship between empirical detection probability and magnitude of location change for change-point detection with 0.05 significance level using ES at the 0.9 level. For both the AR(1) and ARCH(1) processes, the abrupt location change occurs in the middle of the time series sequence. Each plotted point is the average over 1,000 replications with time series sequences of length 400 in each replication.} \label{F3} \end{figure} In Figure \ref{F4} below, we show the sample path behavior of the self-normalized CUSUM process from Equation \ref{E11}. Here, we consider a single realization of the ARCH(1) process used previously. The red lines correspond to the process sample paths under the alternative hypothesis with a unit magnitude location change in the middle of the ARCH(1) sequence of length 400. The blue dotted lines correspond to the process sample paths for the same realized ARCH(1) path, but under the null hypothesis of no location change. The horizontal black dashed line is the threshold for a 0.05 significance level, which if exceeded by the maximum of the self-normalized CUSUM process, results in rejection of the null hypothesis of no location change. As discussed in Section \ref{S3p2}, the ratio form of the self-normalized CUSUM process allows unknown nuisance scale parameters to be canceled out, thereby allowing us to avoid the often problematic task of estimating standard errors in the setting of dependent data. In the case there is a location change, both the CUSUM process in the numerator and the self-normalizer process in the denominator contribute to the threshold exceedance of the self-normalized CUSUM process. \begin{figure}[H] \centering \minipage{0.95\textwidth} \includegraphics[width=\linewidth]{figure4a.jpg} \endminipage \\ \quad \minipage{0.4\textwidth} \includegraphics[width=\linewidth]{figure4b.jpg} \endminipage \: \; \minipage{0.4\textwidth} \includegraphics[width=\linewidth]{figure4c.jpg} \endminipage \: \caption{Change-point testing at the 0.05 significance level using ES at the 0.9 level for ARCH(1) process with sequence length 400. Location change is of unit magnitude and occurs in middle of time series sequence. Top: sample path of self-normalized CUSUM process. Horizontal black dashed line is rejection threshold corresponding to the 0.05 significance level. Bottom left: sample path of CUSUM process. Bottom right: sample path of self-normalizer process.} \label{F4} \end{figure} \subsection{Detection of General Change in Tail} \label{S4p3} We also investigate detection of general structural changes in the upper tail of the underlying marginal distribution. Although the relationship between power and the ``magnitude" of the change in the upper tail is not as simple as in the case of pure location changes, nevertheless, with Proposition \ref{P3} we will detect the change with high probability as our sample size increases. In our simulations, we study the detection of general structural changes in the tail using ES at the 0.95 level (in the upper tail past the 95th percentile), as discussed in Section \ref{S3p2}. We consider variants of the AR(1) and ARCH(1) processes introduced previously. $$\text{AR(1) process:} \qquad X_{i+1} = \begin{cases} 0.5 X_i + t_i(16.5) \qquad \text{for } i \le [n/2] \\ 0.5 X_i + t_i(v) \qquad \text{for } i > [n/2] \end{cases}$$ $$\text{ARCH(1) process:} \qquad X_{i+1} = \begin{cases} \sqrt{1 + 0.2 X_i^2}\epsilon_i \qquad \text{for } i \le [n/2] \\ \sqrt{1 + \lambda X_i^2}\epsilon_i \qquad \text{for } i > [n/2] \end{cases}$$ Here, each $t_i(v)$ is a sample from the t-distribution with $v$ degrees of freedom, and each $\epsilon_i$ is a sample from the standard normal distribution. The parameter values $v$ and $\lambda$ after the change point in the two processes are adjusted, and we examine the effect on power. Figure \ref{F5} below shows the approximate power of change-point tests using the self-normalized CUSUM statistic (Equation \ref{E11}) at the 0.05 significance level. For each data point, we perform 100 replications using times series sequences of length 1,000 and 2,000. For each replication of the AR(1) and ARCH(1) processes, we use an initial burn-in period of 5,000 to approximately reach stationarity. As expected, for both processes, the power is a monotonic function of the magnitude of change. In accordance with the desired 0.05 significance level of our procedure, for both processes, the probability of false positive detection is below 0.05 (0.03 and 0.026 for the AR(1) and ARCH(1) processes, respectively, as indicated by the leftmost point in each of the two plots of Figure \ref{F5}). \begin{figure}[H] \centering \minipage{0.38\textwidth} \includegraphics[width=\linewidth]{figure5a.jpg} \endminipage \minipage{0.38\textwidth} \includegraphics[width=\linewidth]{figure5b.jpg} \endminipage \caption{Left: Change-point tests for AR(1) process. Relationship between empirical detection probability and degrees of freedom $v$ of t-distributed innovations after the change point. Before the change point, $v = 16.5$. Right: Change-point tests for ARCH(1) process. Relationship between empirical detection probability and ARCH(1) parameter $\lambda$ after the change point. Before the change point, $\lambda = 0.2$. In both cases, change-point testing is conducted with 0.05 significance level using ES at the 0.95 level. For both the AR(1) and ARCH(1) processes, the abrupt change occurs in the middle of the time series sequence. Each plotted point is an average over 100 replications, and $n$ refers to the time series sequence length.} \label{F5} \end{figure} \subsection{Detection of Multiple Changes in Tail} \label{S4p4} We additionally investigate detection of multiple structural changes in the upper tail of the underlying marginal distribution. Here, we again use ES at the 0.95 level (in the upper tail past the 95th percentile) and compare the practical performance of the single change-point methodology discussed in Section \ref{S3p2} versus the unsupervised multiple change-point methodology discussed in Section \ref{S3p3}. We consider the following variant of the AR(1) process introduced previously. \begin{align} \label{E17} \text{AR(1) process:} \qquad X_{i+1} = \begin{cases} 0.5 X_i + t_i(16.5) \qquad \text{for } i \le [n/3] \\ 0.5 X_i + t_i(v) \qquad \text{for } [n/3] < i \le [2n/3] \\ 0.5 X_i + t_i(16.5) \qquad \text{for } i > [2n/3], \end{cases} \end{align} where, as before, each $t_i(v)$ is a sample from the t-distribution with $v$ degrees of freedom. In this AR(1) process, the innovations initially have relatively light tails, then change to heavier tails ($v < 16.5$), and finally revert back to the original lighter tails. Figure \ref{F10} below shows the approximate detection power as $v$ is varied using the single change-point methodology (Equations \ref{E11} and \ref{E12}) versus the unsupervised multiple change-point methodology (Equations \ref{E15} and \ref{E16}) at the 0.05 significance level. For each data point, we perform 100 replications using times series sequences of length 1,500. For each replication, we use an initial burn-in period of 5,000 to approximately reach stationarity in the AR(1) process. We see that the single change-point testing method is unable to detect a change in the process in Equation \ref{E17}, even for extremely strong deviations from the null such as the case $v = 2.1$. In fact, its power decays to zero as the magnitude of the change increases. On the other hand, the unsupervised multiple change-point testing method exhibits the desired performance with increasing power as the magnitude of the change increases. Hence, it is a promising candidate for detecting more complex patterns of changes in the tails of time series. In accordance with the desired 0.05 significance level of our procedure, the probability of false positive detection is below 0.05 (0.03 and 0.01 for the single change-point and unsupervised multiple change-point methodologies, respectively, as indicated by the leftmost points in Figure \ref{F10}). \begin{figure}[H] \centering \minipage{0.4\textwidth} \includegraphics[width=\linewidth]{figure10.jpg} \endminipage \caption{Change-point tests for AR(1) process (Equation \ref{E17}) conducted at the 0.05 significance level using ES at the 0.95 level. The relationship between empirical detection probability and degrees of freedom $v$ of t-distributed innovations (as in Equation \ref{E17}) is plotted. The single change-point and unsupervised multiple change-point methodologies are compared. Each plotted point is an average over 100 replications using time series sequences of length 1,500.} \label{F10} \end{figure} \section{Empirical Applications} \label{S5} We first apply the two methods of confidence interval construction from Section \ref{S3p1} to daily log returns of the SPY ETF, which tracks the S\&P 500 Index, for the years 2004-2016. In Figure \ref{F6} below, we show ES estimates of the lower 10th percentile of log returns throughout this time period along with 95\% confidence bands computed using the sectioning and self-normalization methods. We use a rolling window of 100 days with 80 days of overlap between successive windows. The self-normalization method appears to be more conservative and yields a wider confidence band compared to the sectioning method, which agrees with the results presented in Figure \ref{F1}. Overall, the ES estimates and confidence bands appear to capture well the increased volatility of returns during periods of financial instability such as during the 2008 Financial Crisis. \begin{figure}[H] \centering \minipage{0.95\textwidth} \includegraphics[width=\linewidth]{figure6a.jpg} \endminipage \\ \minipage{0.95\textwidth} \includegraphics[width=\linewidth]{figure6b.jpg} \endminipage \caption{Top: Log returns for SPY ETF between January 7, 2004 and December 30, 2016, along with ES estimate and 95\% confidence bands for lower 10th percentile. ES is computed using a rolling window of 100 days with 10 day shifts. Bottom: Cumulative log returns between January 7, 2004 and December 30, 2016.} \label{F6} \end{figure} We next apply the two methods of confidence interval construction from Section \ref{S3p1} to monthly log returns of US 30-Year Treasury bonds for the years 1942-2017. In Figure \ref{F7} below, we show ES estimates of the lower 20th percentile of log returns throughout this time period along with 95\% confidence bands computed using the sectioning and self-normalization methods. We use a rolling window of 40 months with 20 months of overlap between successive windows. Again, the self-normalization method appears to be more conservative and yields a wider confidence band compared to the sectioning method, which agrees with the results presented in Figure \ref{F1}. Although time series for government bond returns have a greater degree of autocorrelation compared to S\&P 500 returns, our methods are still applicable. \begin{figure}[H] \centering \centering \minipage{0.95\textwidth} \includegraphics[width=\linewidth]{figure7.jpg} \endminipage \caption{Log returns for US 30-Year Treasury between March 31, 1942 and December 29, 2017, along with ES estimate and 95\% confidence bands for lower 20th percentile of log returns. ES is computed using a rolling window of 40 months with 10 month shifts.} \label{F7} \end{figure} Next, we apply our single change-point testing methodology at the 0.05 significance level to detect ES changes in the lower 5th percentile of SPY ETF log returns and also the lower 5th percentile of US 30-Year Treasury log returns. Five time series are shown in Figure \ref{F8} below, and for each, a change was detected using our method from Equations \ref{E11} and \ref{E12}. First, we examine change-point tests for SPY ETF time series. Figure \ref{F8} (a) and (b) contain the 2008 Financial Crisis and 2011 August Stock Markets Fall, respectively. Figure \ref{F8} (c) also contains the 2011 August Stock Markets Fall, but with the change point located towards the end of the time series. Our methods are generally robust in situations where change points are located near the extremes of the time series. We also examine change-point tests for US 30-Year Treasury time series. Figure \ref{F8} (d) contains the 1980-1982 US recession (due in part to government restrictive monetary policy, and to a less degree the Iranian Revolution of 1979, which resulted in significant oil price increases). Figure \ref{F8} (e) contains the 2008 Financial Crisis. \begin{figure}[H] \textbf{\qquad SPY ETF} \begin{center} (a) \minipage{0.85\textwidth} \includegraphics[width=\linewidth]{figure8a.jpg} \endminipage \\ (b) \minipage{0.85\textwidth} \includegraphics[width=\linewidth]{figure8b.jpg} \endminipage \\ (c) \minipage{0.85\textwidth} \includegraphics[width=\linewidth]{figure8c.jpg} \endminipage \end{center} \textbf{\qquad US 30-Year Treasury} \begin{center} (d) \minipage{0.85\textwidth} \includegraphics[width=\linewidth]{figure8d.jpg} \endminipage \\ (e) \minipage{0.85\textwidth} \includegraphics[width=\linewidth]{figure8e.jpg} \endminipage \end{center} \caption{Log returns for (a) SPY ETF between May 15, 2008 and December 17, 2008 (b) SPY ETF between May 6, 2011 and September 28, 2011 (c) SPY ETF between February 24, 2011 and August 16, 2011 (d) US 30-Year Treasury between June 29, 1973 and June 30, 1983 (e) US 30-Year Treasury between April 30, 2004 and August 31, 2012.} \label{F8} \end{figure} In Figure \ref{F9} below, we show the plots associated with change-point testing for the time series in Figure \ref{F8} (d). Similar to our findings for Figure \ref{F4}, we see that the CUSUM process in the numerator and the self-normalizer process in the denominator (recall Equation \ref{E11}) contribute to the threshold exceedance of the self-normalized CUSUM process. \begin{figure}[H] \centering \minipage{0.95\textwidth} \includegraphics[width=\linewidth]{figure9a.jpg} \endminipage \\ \quad \: \minipage{0.4\textwidth} \includegraphics[width=\linewidth]{figure9b.jpg} \endminipage \: \: \minipage{0.4\textwidth} \includegraphics[width=\linewidth]{figure9c.jpg} \endminipage \: \caption{Change-point testing at 0.05 significance level using ES for lower 5th percentile of US 30-Year Treasury log returns between June 29, 1973 and June 30, 1983. Top: sample path of self-normalized CUSUM process. Horizontal black dashed line is rejection threshold corresponding to the 0.05 significance level. Bottom left: sample path of CUSUM process. Bottom right: sample path of self-normalizer process.} \label{F9} \end{figure} Finally, we compare the effectiveness of the unsupervised multiple change-point testing and single change-point testing methodologies for detecting ES changes in the lower 5th percentile of SPY ETF log returns for longer time horizons. In agreement with our findings from the simulation study in Section \ref{S4p4}, at the 0.05 significance level, we confirm that the unsupervised multiple change-point test is able to detect the presence of one or more change points in the SPY ETF time series between the start of 2007 and the end of 2010 (thus including the 2008 Financial Crisis) as shown in Figure \ref{F12}, while the single change-point test is unable to detect any change.\footnote{This could be due to the effect of multiple change points canceling out over the this longer time period.} In particular, the value of the unsupervised multiple change-point test statistic in Equation (\ref{E15}) is 299.4, which well exceeds the test's rejection threshold of 138.2 (corresponding to the 0.05 significance level) and gives a strong signal for rejection of the null hypothesis that there is no change point. However, the single change-point test performs poorly on this longer time series, and the value of the test statistic in Equation (\ref{E11}) is 1.915, which is well below the test's rejection threshold of 40.1 (corresponding to the 0.05 significance level). Thus, the single change-point test is reliable on shorter time horizons where only one change point occurs, but can fail on longer time horizons with multiple change points. Our unsupervised multiple change-point test provides a solution to this problem. \begin{figure}[H] \centering \minipage{0.95\textwidth} \includegraphics[width=\linewidth]{figure12.jpg} \endminipage \caption{Log returns for SPY ETF between January 3, 2007 and December 20, 2010. A change in the tail distribution of this time series is clearly present between the end of 2008 through early 2009 (during the 2008 Financial Crisis). At the 0.05 significance level, the single change-point test is unable to detect a change in ES in the lower 5th percentile of the log returns, while the unsupervised multiple change-point test finds strong evidence of one or more changes.} \label{F12} \end{figure} \section{Conclusion} \label{C} We propose methodology to perform confidence interval construction and change-point testing for fundamental nonparametric estimators of risk such as ES. This allows for evaluation of the homogeneity of ES and related measures such as the conditional tail moments, and in particular allows the investigator to detect general tail structural changes in time series observations. While current approaches to tail structural change testing typically involve quantities such as the tail index and thus require parametric modeling of the tail, our approach does not require such assumptions. Moreover, we are able to detect more general structural changes in the tail using ES, for example, location and scale changes, which are undetectable using tail index. Hence, we advocate the use of ES for general purpose monitoring for tail structural change. We note that our proposed sectioning and self-normalization methods for confidence interval construction and change-point testing still require some user choice, for example, the number of sections to use in sectioning or which particular functional to use in the self-normalization. Simulations suggest that our method is robust to these user choices. Therefore, we view our method as more robust compared to extant approaches which involve consistent estimation of standard errors or blockwise versions of the bootstrap or empirical likelihood, which can be more sensitive to tuning parameters. Our simulations illustrate the promising finite-sample performance of our procedures. Furthermore, we are able to construct pointwise ES confidence bands for SPY ETF returns in the period 2004-2016 and for US 30-Year Treasury returns in the period 1942-2017 that well capture periods of market distress such as the 2008 Financial Crisis. In similar spirit, our change-point tests are able to detect tail structural changes through the ES for both the SPY ETF returns and the US 30-Year Treasury returns during key times of financial instability in the recent past. \bibliographystyle{plainnat}

72f866fda08a52181b9c480c16c256f45c40db43

\section{Introduction} \label{sec:Introduction} The use of models to support informed decision making is ubiquitous. However, the size and nature of the decision variable solution space, and the model runtime, may make a comprehensive -- exhaustive or simply extensive -- evaluation of the problem space computationally infeasible. In such cases an efficient approach is required to search for global optima. Mathematical programs are one form of model that are explicitly formulated as optimisation problems, where the model representation imposes assumptions on the structure of the decision variable space and objective function. Such models are well suited to efficient solution, and identification of global optima may be theoretically guaranteed when feasible solutions exist. However many real-world problems are not suited to expression as a mathematical program (e.g., a solution is evaluated by using a simulation tool). From an optimisation perspective models where no assumptions are made about the model structure can be thought of as a black-box, where decision variables values are input and outputs generated for interpretation as an objective. In this case optimisation search techniques such as metaheuristics are required, i.e., general rule-based search techniques that can be applied to any model. An additional widespread feature of many real-world problems is the consideration of uncertainty which may impact on model outputs, and so on corresponding objective function values. One strategy is to simply ignore any uncertainty and perform a standard search, possibly assessing and reporting on the sensitivity of the optimum after it has been identified. However it has been established that optimal solutions which are sensitive to parameter variations within known bounds of uncertainty may substantially degrade the optimum objective function value, meaning that solutions sought without explicitly taking account of uncertainty are susceptible to significant sub-optimality, see \cite{BenTalElGhaouiNemirovski2009, GoerigkSchobel2016}. In the face of uncertainty the focus of attention for an optimisation analysis shifts from the identification of a solution that just performs well in the expected case, to a solution that performs well over a range of scenarios. In this paper we develop a new algorithm for box-constrained robust black-box global optimisation problems taking account of implementation uncertainty, i.e., the solution that a decision maker wants to implement may be slightly perturbed in practice, and the aim is to find a solution that performs best under the worst-case perturbation. Our method is based on an exploration technique that uses largest empty hyperspheres (LEHs) to identify regions that can still contain improving robust solutions. In a computational study we compare our method with a local search approach from the literature (see \cite{BertsimasNohadaniTeo2010}) and a standard particle swarm approach. We find that our approach considerably outperforms these methods, especially for higher-dimensional problems. \paragraph{Structure of this paper.} We begin with a review of the literature on metaheuristics for robust optimisation in Section~\ref{sec:literature} before outlining the formal description of robust min max problems in Section~\ref{sec:RobustOptimisation}. We also consider some of the details of the established local robust search technique due to \cite{BertsimasNohadaniTeo2010}. In Section~\ref{sec:LargestEmptyHypersphere} we introduce a novel approach, an exploration-focused movement through the search space identifying areas that are free of previously identified poor points. We include a discussion and descriptions of the algorithms used to identify empty regions of the decision variable search space. The approach is then tested against alternative heuristics in Section~\ref{sec:ExperimentsResults}, on test problems of varying dimension. The experimental set up is described and the results of this analysis presented. Finally we summarise and consider further extensions of this work in Section~\ref{sec:SummaryConcusionsFurtherWork}. \section{Literature review} \label{sec:literature} \subsection{Robust optimisation} Different approaches to model uncertainty in decision making problems have been explored in the literature. Within robust optimisation, a frequent distinction is made between parameter uncertainty (where the problem data is not known exactly) and implementation uncertainty (where a decision cannot be put into practice with full accuracy). Implementation uncertainty is also known as decision uncertainty \cite{BenTalElGhaouiNemirovski2009, Ghazali2009, BertsimasNohadaniTeo2010}. A common approach to the incorporation of uncertainty for black-box problems is stochastic optimisation. Here knowledge of the probability distributions of the uncertain parameters is assumed and some statistical measure of the fitness of a solution assessed, e.g. using Monte Carlo simulation to estimate the expected fitness. This may be the expected value, or a more elaborate model such as the variance in the fitness of a solution, or even a multi-objective optimisation setting, see \cite{PaenkeBrankeJin2006, HomemdeMelloBayraksan2014}. An alternative to a stochastic approach is robust optimisation, whose modern form was first developed in \cite{KouvelisYu1997} and \cite{BenTalNemirovski1998}. Whereas with stochastic optimisation a knowledge of probability distributions over all possible scenarios is typically assumed, in robust optimisation it is only assumed that some set is identified containing all possible uncertainty scenarios (potentially infinite in number). A classic robust approach is then to find a solution across all scenarios that is always feasible (strictly robust) and optimises its performance in the worst case. This is known as min max. For a given point in the decision variable space there is an `inner' objective to identify the maximal (worst case) function value in the local uncertainty neighbourhood, and an overall `outer' objective to identify the minimum such maximal value. The field of robust optimisation has been primarily aligned with mathematical programming approaches. There the methodology is based around the definition of reasonable uncertainty sets and the reformulation of computationally tractable mathematical programming problems. For specific forms of convex optimisation problems, the problem incorporating uncertainty can be re-formulated to another tractable, convex problem, see \cite{BertsimasNohadaniTeo2007, GohSim2010}. To overcome concerns that the strictly robust worst case approach may be overly conservative, the concept of robustness can be expanded in terms of both the uncertainty set considered and the robustness measure \cite{GoerigkSchobel2016}. On the assumption that it is overly pessimistic to assume that all implementation errors take their worst value simultaneously \cite{BertsimasSim2004} consider an approach where the uncertainty set is reduced, and a robust model defined where the optimal solution is required to remain feasible for uncertainty applied to only a subset of the decision variables at any given time. Min max regret, see \cite{AissiBazganVanderpooten2009}, is an alternative to min max, seeking to minimise the maximum deviation between the value of the solution and the optimal value of a scenario, over all scenarios. \cite{BenTalBertsimasBrown2010} considers soft robustness, which utilises a nested family of uncertainty sets. The distributionally robust optimisation approach, see \cite{GohSim2010}, attempts to bridge robust and stochastic techniques by utilizing uncertainty defined as a family of probability distributions, seeking optimal solutions for the worst case probability distribution. \cite{ChasseinGoerigk2016} use a bi-objective approach to balance average and worst case performance by simultaneously optimising both. Robust optimisation in a mathematical programming context has been application-driven, so considerable work has been undertaken in applying robustness techniques to specific problems or formulations, see \cite{BeyerSendhoff2007, GoerigkSchobel2016}. There has also been some cross-over into the application of specific heuristics, for example see \cite{GoldenLaporteTaillard1997, ValleMartinezdaCunhaMateus2011}. However application to general problems has been less well addressed \cite{GoerigkSchobel2016}. Furthermore robust approaches applied to black-box models are much less widely considered than approaches for mathematical programming problems, see \cite{MarzatWalterPietLahanier2013, GoerigkSchobel2016, MarzatWalterPietLahanier2016}. Recently, robust optimisation with implementation uncertainty has also been extended to multi-objective optimisation, see \cite{eichfelder2017decision}. \subsection{Metaheuristic for robust optimisation} The min max approach has been tackled with standard metaheuristic techniques applied to both the inner maximisation and outer minimisation problems. In co-evolutionary approaches two populations (or swarms) evolve separately but are linked. The fitness of individuals in one group is informed by the performance of individuals in the other, see \cite{CramerSudhoffZivi2009}. \cite{Herrmann1999, Jensen2004} use such a two-population genetic algorithm (GA) approach, whilst \cite{ShiKrohling2002, MasudaKuriharaAiyoshi2011} consider two-swarm co-evolutionary particle swarm optimisation (PSO) techniques for min max problems. A brute force co-evolutionary approach is to employ complete inner maximisation searches to generate robust values for each individual in each generation of the outer minimisation, however this is expensive in terms of model runs (i.e., function evaluations), see \cite{MarzatWalterPietLahanier2016}. More practical co-evolutionary approaches, for example using only small numbers of populations for the outer search and the inner (uncertainty) search which share information between populations from generation to generation, or following several generations, require the application of additional simplifications and assumptions, see \cite{CramerSudhoffZivi2009, MasudaKuriharaAiyoshi2011}. One general area of research is the use of emulation to reduce the potential burden of computational run times and the number of model-function evaluations, see \cite{VuDAmbrosioHamadiLiberti2016}. \cite{ZhouZhang2010} use a surrogate-assisted evolutionary algorithm to tackle the inner search for black-box min max problems. \cite{MarzatWalterPietLahanier2013, MarzatWalterPietLahanier2016} employs Kriging meta-modelling coupled with an expected improvement (EI) metric, as well as a relaxation of the inner maximisation search. The EI metric is used to efficiently choose points in the decision variable space where nominal (expensive) function evaluation should be undertaken, see \cite{JonesSchonlauWelch1998}, here with a view to most efficiently improving the estimate of the robust global minimum. The relaxation involves iteratively performing the outer minimisation on a limited inner uncertainty neighbourhood followed by an extensive inner maximisation search in the region of the identified outer minimum. This continues whilst the inner search sufficiently deteriorates the outer solution, with the inner maximum point being added to the limited inner uncertainty set with each iteration. A second approach due to \cite{urRehmanLangelaarvanKeulen2014, urRehmanLangelaar2017} also uses Kriging and an EI metric, building on a meta-model of the expensive nominal problem by applying a robust analysis directly to the Kriging model and exploiting the fact that many more inexpensive function evaluations can be performed on this meta-model. A modified EI metric is calculated for the worst case cost function of the meta-model, to efficiently guide the search in the nominal expensive function space. In \cite{urRehmanLangelaar2017} the approach is applied to a constrained non-convex 2 dimensional problem due \cite{BertsimasNohadaniTeo2010, BertsimasNohadaniTeo2010nonconvex}, the unconstrained version of which is also considered here. The Kriging-based approach is shown to significantly outperform the approaches outlined here, in terms of the number of expensive function evaluations required to converge towards the robust optimum. In general we would expect the approach from \cite{urRehmanLangelaarvanKeulen2014, urRehmanLangelaar2017} to outperform the approaches considered here, in terms of efficiency when applied to low dimensional non-convex problems. However the primary challenge with meta-model based approaches is their application to higher dimensional problems. The test cases considered in \cite{MarzatWalterPietLahanier2013, MarzatWalterPietLahanier2016, urRehmanLangelaarvanKeulen2014, urRehmanLangelaar2017} have either been restricted to low dimensional non-convex problems, or simpler convex and convex-concave problems of up to 10 dimensions. One local black-box min max approach is due to \cite{BertsimasNohadaniTeo2007, BertsimasNohadaniTeo2010, BertsimasNohadaniTeo2010nonconvex}. Here a search is undertaken by iteratively moving along 'descent directions'. Uncertainty around individual points is assessed using local gradient ascents, based on which undesirable 'high cost points' (hcps) are identified. Steps are taken in directions which point away from these hcps, until no direction can be found. Our approach is inspired by both elements of the descent directions technique and the concept of relaxation of the inner maximisation search. We extend the idea of locally moving away from identified hcps to a global perspective, seeking regions of the solution space currently empty of such undesirable points. Furthermore the nature of our outer approach enables the curtailing of an inner maximisation search if it is determined that the current point under consideration cannot improve on the current best robust global solution. \section{Notation and previous results} \label{sec:RobustOptimisation} \subsection{Problem description} \label{sec:ProblemDescription} We consider a general optimisation problem of the form \begin{align*} \quad \min\ & f(\pmb{x}) \\ \text{s.t. } & \pmb{x} \in \mathcal{X} \end{align*} where $\pmb{x}=(x_{1}, x_{2}, \ldots, x_{n})^T$ denotes the $n$-dimensional vector of decision variables, $f: \mathbb{R}^n \to \mathbb{R}$ is the objective function, and $\mathcal{X} \subseteq \mathbb{R}^n$ is the set of feasible solutions. We write $[n]:=\{1,\ldots,n\}$. In this paper, we assume box constraints $\mathcal{X} = \prod_{i\in[n]} [l_i,u_i]$. Any other potential feasibility constraints are assumed to be ensured through a penalty in the objective. In implementation uncertainty, we assume that a desired solution $\pmb{x}$ might not be possible to put into practice with full accuracy. Instead, a ''close'' solution $\tilde{\pmb{x}}=\pmb{x}+\Delta\pmb{x}$ may be realised. The aim is to find a robust $\pmb{x}$ such that for any such solution $\tilde{\pmb{x}}$ from the neighbourhood of $\pmb{x}$, the worst case performance is optimised. More formally, we follow the setting of \cite{BertsimasNohadaniTeo2010} and consider the so-called uncertainty set \[ \mathcal{U}:=\{ \Delta \pmb{x}\in\mathbb{R}^n \mid \| \Delta \pmb{x} \| \leq \Gamma \} \] where $\Gamma > 0$ defines the magnitude of the uncertainty, and $\|\cdot\|$ refers to the Euclidean norm. The worst case costs of a solution $\pmb{x}\in\mathcal{X}$ are then given as \[ g(\pmb{x}):=\max_{\Delta \pmb{x} \in \mathcal{U}} f(\pmb{x} + \Delta \pmb{x}) \] and so the robust optimisation problem is given by: \[ \min_{\pmb{x}\in\mathcal{X}} g(\pmb{x}) = \min_{\pmb{x}\in\mathcal{X}} \max_{\Delta\pmb{x} \in \mathcal{U}} f(\pmb{x} + \Delta \pmb{x}) \tag{ROP}\] We therefore have an inner maximisation and outer minimisation problem, such that the identification of the robust global optimum is based on finding the (outer) minimum worst case cost objective function value in the decision variable space, and that objective is determined by the (inner) maximisation of the nominal objective function in the uncertainty neighbourhood around each point in the decision variable space. This type of problem is also known as min max. Note that $\pmb{x}+\Delta\pmb{x}$ may not be in $\mathcal{X}$, for which reason we assume that the definition of $f$ is not limited to $\mathcal{X}$. However, if it is desired that $\pmb{x}+\Delta\pmb{x}\in\mathcal{X}$ for all $\Delta\pmb{x}\in\mathcal{U}$, then this can be ensured by reducing the size of the feasible search space by $\Gamma$. As an example for our problem setting, consider the 2-dimensional polynomial function due to \cite{BertsimasNohadaniTeo2010}: \begin{align*} f(x, y) = &2x^6 - 12.2x^5 + 21.2x^4 + 6.2x - 6.4x^3 - 4.7x^2 - y^6 \\ - &11y^5 + 43.3y^4 - 10y - 74.8y^3 + 56.9y^2 - 4.1xy \\ - &0.1y^2x^2 + 0.4y^2x + 0.4x^2y \tag{poly2D} \end{align*} For a feasible solution space within bounds $[-1, 4]$ in each dimension, and uncertainty defined by a $\Gamma$-radius value of $0.5$, the nominal and worst case plots for (poly2D) are shown in Figure~\ref{fig:NominalWorstCaseBertsimasPoly}. In min max the problem is one of finding the global minimum for the worst case cost function. If uncertainty is ignored the problem is just one of finding the global minimum of the (nominal) objective as shown in Figure~\ref{BertsimasNominal}, whereas including uncertainty the problem becomes one of finding the (worst case cost) objective as shown in Figure~\ref{BertsimasWorstCase}. In both cases the search proceeds based on generating nominal objective values but for the worst case cost we must further undertake some assessment of the impact of uncertainty on those objective outputs. \vspace*{-6mm} \begin{figure}[htbp] \centering \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=2.8in, height=3.0in]{BertsimasNom.pdf} \vspace*{-7mm} \caption{Nominal problem} \label{BertsimasNominal} \end{subfigure}% \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=2.8in, height=3.0in]{BertsimasWorst.pdf} \vspace*{-7mm} \caption{Worst case problem with $\Gamma$=0.5} \label{BertsimasWorstCase} \end{subfigure} \caption{Nominal and worst case cost for (poly2D) from \cite{BertsimasNohadaniTeo2010}. Marked in purple are the respective optima.} \label{fig:NominalWorstCaseBertsimasPoly} \end{figure} Here the global optimum value for the nominal problem is -20.8 at (2.8, 4.0). The worst case plot is estimated by randomly sampling large numbers of points in the $\Gamma$-uncertainty neighbourhood around each plotted point. The worst case cost at each point is then approximated as the highest value of $f(x)$ identified within each $\Gamma$-uncertainty neighbourhood. The global optimum for the worst case problem is approximately 4.3 at (-0.18, 0.29). The significant shift in the nominal versus robust optima, both in terms of its location and the optimum objective, emphasises the potential impact of considering uncertainty in decision variable values. The difference between the nominal and robust optimal objective function values is the `price of robustness', see \cite{BertsimasSim2004}. \subsection{Local robust search using descent directions} \label{sec:LocalRobustSearchDescentDirections} We briefly summarise the local search approach for (ROP) that was developed in \cite{BertsimasNohadaniTeo2010}. Here, (ROP) is solved using a local robust optimisation heuristic illustrated by Figure~\ref{fig:BertsimasDescription}. An initial decision variable vector $\hat{\pmb{x}}$ is randomly sampled. Then a series of gradient ascent searches are undertaken within the $\Gamma$-uncertainty neighbourhood of this candidate solution to identify hcps, see Figure~\ref{fig:algo2}. This approximates the inner maximisation problem $\max_{\Delta\pmb{x}} f(\hat{\pmb{x}} + \Delta\pmb{x})$. Using a threshold value that is dynamically adjusted during the algorithm, a subset $H(\hat{\pmb{x}})$ of all evaluated points is identified, see Figure~\ref{fig:algo3}. \begin{figure}[htbp] \centering \begin{subfigure}{.44\textwidth} \centering \includegraphics[width=.7\textwidth]{algo2.pdf} \caption{Candidate point $\pmb{x}$ (centre), and points evaluated for the inner maximisation problem (blue).} \label{fig:algo2} \vspace*{9mm} \end{subfigure}% \hspace{.1\textwidth} \begin{subfigure}{.44\textwidth} \centering \includegraphics[width=.7\textwidth]{algo3.pdf} \caption{Subset $H(\pmb{x})$ of critical high cost points.} \label{fig:algo3} \vspace*{9mm} \end{subfigure} \begin{subfigure}{.44\textwidth} \centering \includegraphics[width=.7\textwidth]{algo5.pdf} \caption{A descent direction is identified by solving a second order cone problem.} \label{fig:algo5} \end{subfigure}% \hspace{.1\textwidth} \begin{subfigure}{.44\textwidth} \centering \includegraphics[width=.7\textwidth]{algo6.pdf} \caption{The step size is determined.} \label{fig:algo6} \end{subfigure} \caption{Description of the descent direction robust local search approach \cite{BertsimasNohadaniTeo2010}.} \label{fig:BertsimasDescription} \end{figure} In the next step, a descent direction is identified that points away from the set $H(\hat{\pmb{x}})$, see Figure~\ref{fig:algo5}. To this end, a mathematical programming approach is used, minimising the angle between the hcps and the candidate solution. This leads to the following second order cone problem. \begin{align} \min_{\pmb{d},\beta} \ & \beta \label{soc1} \\ \text{s.t. } &\| \pmb{d}\| \le 1 \label{soc2} \\ & \pmb{d}^T \pmb{h} \le \beta & \forall \pmb{h}\in H(\hat{\pmb{x}}) \label{soc3}\\ & \beta \le -\varepsilon \label{soc4} \end{align} Here, $\pmb{d}$ is the descent direction, which is normalised by Constraint~\eqref{soc2}. Constraints~\eqref{soc3} ensure that $\beta$ is the maximum angle between $\pmb{d}$ and all high cost points $\pmb{h}$. Through Constraint~\eqref{soc4}, we require a feasible descent direction to point away from all points in $H(\hat{\pmb{x}})$. When an optimal direction cannot be found, the algorithm stops -- a robust minimum has been reached. Next the size of the step to be taken is calculated, see Figure~\ref{fig:algo6}. A step size just large enough to ensure that all of the hcps are outside of the $\Gamma$-uncertainty neighbourhood of the next candidate solution is used. Using the identified descent direction and step size the algorithm moves to a new candidate point, and so the heuristic repeats iteratively until a robust minimum has been identified. \section{A new largest empty hypersphere approach} \label{sec:LargestEmptyHypersphere} \subsection{Algorithm overview} \label{sec:GlobalRobustSearchLEH} \noindent Building on the notion of a search that progresses locally by moving away from already identified poor (high cost) points, we develop a global approach that iteratively moves to the region of the decision variable solution space furthest away from recognised hcps. This is an exploration-focused approach, although rather than concentrating on examining unvisited regions the intention here is to identify and visit regions devoid of hcps. Assuming uncertainty as considered previously in terms of a single value $\Gamma$ that defines a radius of uncertainty in all decision variables, we associate the idea of the largest empty region (empty of hcps) with the idea of the largest empty hypersphere (LEH), or largest empty circle in 2D. The approach is then to locate the next point in the search at the centre of the identified LEH, and to iteratively repeat this as more regions are visited and hcps identified. The approach is described in Figure~\ref{fig:LEHDescription}. We start by randomly sampling one or more points and evaluating the objective function $f$ at each. From these start points a candidate point is selected and an inner analysis undertaken in the candidate's $\Gamma$-uncertainty neighbourhood with a view to identifying the local maximum, Figure~\ref{fig:newalgo3}. This local worst case cost for the candidate is the first estimate of a robust global minimum, that is a global min max, and is located at the candidate point. The aim is now to move to a point whose uncertainty neighbourhood has a lower worst case cost than the current global value. We seek to achieve this by identifying the largest hypersphere of radius at least $\Gamma$ within the defined feasibility bounds which is completely empty of hcps, and moving to the centre of that LEH, see Figures~\ref{fig:newalgo4} -~\ref{fig:newalgo5}. All points evaluated are recorded in a history set, a subset of which forms the high cost set. The high cost set contains a record of all points evaluated so far with an objective function value greater or equal to a high cost threshold, and here the high cost threshold is set as the current estimate of the robust global minimum. Both the history set and the high cost set are updated as more points are visited and the high cost threshold reduces, see Figures~\ref{fig:newalgo6} -~\ref{fig:newalgo7}. On performing all inner searches after the first candidate, a candidate's robust value may be no better than the current estimate of the robust global minimum (and therefore the current high cost threshold), in which case at least one point will be added to the high cost set. Alternatively if a candidate's robust value is better than the current estimate of the robust global minimum, this current recorded optimum is overwritten and the high cost threshold reduced accordingly. Again this introduces at least one high cost point to the high cost set, but the reducing threshold may also introduce additional points from the history set; this is suggested in Figure~\ref{fig:newalgo7}. The search stops when no LEH of radius greater than $\Gamma$ exists or some pre-defined resource limit has been reached. Then the candidate point around which the current estimate of the robust global minimum has been determined is deemed the robust global minimum. Otherwise the search repeats, performing analysis in the $\Gamma$-uncertainty neighbourhood around candidates to estimate the local (inner) max, updating the global minimum worst case cost if appropriate, and moving to the next identified LEH, Figure~\ref{fig:newalgo8}. \begin{figure}[H] \centering \begin{subfigure}{.48\textwidth} \centering \includegraphics[width=2.2in]{newalgo3xC.pdf} \caption{The decision variable space is seeded randomly. Perform an inner search around one candidate point.} \label{fig:newalgo3} \vspace*{5mm} \end{subfigure}% \hfill \begin{subfigure}{.48\textwidth} \centering \includegraphics[width=2.2in]{newalgo4xC.pdf} \caption{The current high cost set, including one point from the previous inner search and some of the seed points.} \label{fig:newalgo4} \vspace*{5mm} \end{subfigure} \begin{subfigure}{.48\textwidth} \centering \includegraphics[width=2.2in]{newalgo5xC.pdf} \caption{Identify the largest empty hypersphere, the centre of which is the next candidate point.} \label{fig:newalgo5} \vspace*{5mm} \end{subfigure}% \hfill \begin{subfigure}{.48\textwidth} \centering \includegraphics[width=2.2in]{newalgo6yC.pdf} \caption{Inner search around the new candidate. The robust value here is less than the current global minimum.} \label{fig:newalgo6} \vspace*{5mm} \end{subfigure} \begin{subfigure}{.48\textwidth} \centering \includegraphics[width=2.2in]{newalgo7yC.pdf} \caption{The current high cost set, including more previously evaluated points due to the reduced high cost threshold.} \label{fig:newalgo7} \vspace*{5mm} \end{subfigure}% \hfill \begin{subfigure}{.48\textwidth} \centering \includegraphics[width=2.2in]{newalgo8yC.pdf} \caption{Identify the largest empty hypersphere, the centre of which is the next candidate point.} \label{fig:newalgo8} \vspace*{5mm} \end{subfigure} \caption{Description of largest empty hypersphere (LEH) approach.} \label{fig:LEHDescription} \end{figure} The critical feature of such an approach is the identification of regions of the solution space that are currently empty of, and furthest from, the undesirable hcps. As defined here this corresponds to identifying the largest hypersphere devoid of all hcps. Given a discrete history set $H$ of all points evaluated so far, high cost points are those members of $H$ with objective value which is at least the current high cost threshold $\tau$, i.e., \[ H_\tau := \{ \pmb{h}\in H \mid f(\pmb{h}) \geq \tau \} \] We denote $N_{\tau} = |H_\tau|$ as the cardinality of $H_\tau$, and write $H_\tau = \{\pmb{h}^1,\ldots,\pmb{h}^{N_\tau}\}$. The identification of a point $\pmb{p}\in\mathcal{X}$ which is furthest from all $N_\tau$ high cost points in $H_\tau$ is a max min problem: \[ \max_{\pmb{p}\in\mathcal{X}} \min_{i\in[N_\tau]} d(\pmb{p}, \pmb{h}^i), \tag{LEHP}\] where $d(\pmb{p}, \pmb{q})$ is the Euclidean distance between two points $\pmb{p}$ and $\pmb{q}$, see \cite{OkabeSuzuki1997}. In the following, we specify this general LEH approach by considering two aspects in more detail: The \textit{outer search} is concerned with placing the next candidate point $\pmb{x}$ by solving (LEHP). The \text{inner search} then evaluates this candidate by calculating $g(\pmb{x})$ approximately. \subsection{Outer search methods} Here we will introduce different approaches to identifying the largest empty hypersphere, given a set of high cost points $H_\tau$. It should be noted that none of these approaches requires additional function evaluations, which is usually considered the limiting resource in black-box settings. \subsubsection{Randomly sampled LEH algorithm} \label{secLEH1} A very simple approach is to generate potential candidates randomly within the feasible region, then determine whether they are more than $\Gamma$ away from all hcps. If so they are a valid candidate, if not re-sample up to some defined maximum number of times beyond which it is assumed that no such candidate point can be found and the solution has converged on a robust global minimum. Rather than being a largest empty hypersphere approach this is just a valid empty hypersphere approach, and the size of the identified empty hypersphere might vary considerably from one candidate to the next. \subsubsection{Genetic Algorithm for LEH} \label{secLEH2} The solution of (LEHP) is an optimisation problem. Furthermore, given a point $\pmb{p}$ which is a potential candidate for the centre of the largest empty hypersphere, the inner minimisation calculation in (LEHP) involves just an enumeration over the $N_\tau$ Euclidean distance calculations between each hcp and $\pmb{p}$ to identify the minimum distance $d(\pmb{p}, \pmb{h}^k)$, where $\pmb{h}^k$ is the closest hcp. Therefore the focus for the solution of (LEHP) is the outer maximisation, for which we may consider an approximate heuristic approach. We employ a genetic algorithm (GA), a commonly cited evolutionary algorithm (EA) \cite{Ghazali2009}. Here each individual represents a point $\pmb{p}$ in the decision variable space, and the objective function $f_{LEH}(\pmb{p}):=\min_{\pmb{h}\in H_\tau} d(\pmb{p},\pmb{h})$ is the minimum distance between a given point $\pmb{p}$ and all hcps in $H_\tau$. We seek to maximise this minimal distance by evolving a population of points starting from randomly selected feasible points in the decision variable space $\mathcal{X}$. The best point generated by the GA is the next candidate point -- that is estimated centre of the LEH, for the current $H$, $\tau$ and $H_\tau$. \subsubsection{Voronoi based LEH} \label{secLEH3} Within the literature a widely referenced approach for tackling low dimensional LEH problems is due to \cite{Toussaint1983}, and is based on the geometric Voronoi diagram approach, see \cite{Chazelle1993, OkabeSuzuki1997}. The Voronoi approach partitions a space into regions (cells). For a given set of points each cell corresponds to a single point such that no point in the cell is closer to any other point in the set. Points on the edges between cells are equidistant between the set points which lie on either side of that edge. For our LEH problem the set of points is $H_\tau$, and the Voronoi diagram approach corresponds to segmenting the feasible space $\mathcal{X}$ into $N_\tau$ separate cells, one for each hcp. The (Voronoi) vertices that lie at the intersection of these cell (Voronoi) edges maximise the minimum distance to the nearby set points, see \cite{Chazelle1993, OkabeSuzuki1997}. So for a given $H_\tau$ if we can determine the Voronoi diagram we can use the identified Voronoi vertices as potential candidate points $\pmb{p}$. The solution of (LEHP) is then simply a matter of enumeration, for each $\pmb{p}$ calculating the (inner) minimum Euclidean distance to all hcps, and then selecting the (outer) maximum such minimal distance. The original approach due to \cite{Toussaint1983} includes the identification of vertices (candidate centres of LEHs) that can be sited outside of defined boundaries, in infeasible regions. This is not exactly as required here. To deal with this edges that cross feasibility boundaries are identified and the associated vertices which are outside of $\mathcal{X}$ are relocated to an appropriate point on the boundary of $\mathcal{X}$. Here any coordinate $i\in[n]$ of such an external vertex that is either less than $l_i$ or greater than $u_i$ is re-set to $l_i$ or $u_i$ as appropriate. However the Voronoi approach has exponential dependence on $n$, as constructing the Voronoi diagram of $N_\tau$ points requires $O(N_\tau log N_\tau + N_\tau^{\ceil*{n/2}})$ time \cite{Chazelle1993}. This suggests that such an approach in not computationally viable for anything other than low dimensional problems. On the basis that a Voronoi diagram based approach is the primary recognised heuristic for identifying the largest empty \textit{circle} we will consider a Voronoi based robust LEH heuristic here only in the context that for 2D problems in our experimental analysis this approach will serve as a good direct comparator for our other robust LEH heuristics. \subsection{Inner search methods} Discussions of the LEH approach have so far focussed on the outer minimisation search, assuming some form of inner search that provides the inner robust maximum for each candidate point in the minimisation search. In \cite{BertsimasNohadaniTeo2010} a two-stage gradient ascent search is recommended for each inner search around a candidate point. This assumes gradient information is available and proposes $(n+1)$ individual two-stage gradient ascents for each candidate. For a 100-dimensional problem this would require several thousand function evaluations around each candidate point. In practical terms both the number of function evaluations required to undertake a global search and the requirement for gradient information may make such extensive inner searches prohibitive. Given, for example, budgetary restrictions on the number of function evaluations, some trade-off must be achieved between the extent of each inner $\Gamma$-radius uncertainty neighbourhood search and globally exploring the search space. But this trade-off between robustness in terms of the extent of the inner searches, and performance in terms of the outer global search, is complex, see \cite{MirjaliliLewisMostaghim2015, DiazHandlXu2017}. For example the determination of an appropriate inner approach -- type of search, extent of search and parameter settings -- may be both instance (problem and dimension) dependent and dependent on the outer approach. Here we do not propose to recommend a definitive inner search approach. From a theoretical point of view we assume the information is provided by some oracle. From an experimental point of view in the algorithm testing and comparisons below we assume the same basic inner $\Gamma$-radius uncertainty neighbourhood analysis for all heuristics, to ensure a consistency when comparing results for alternative search approaches. There is, however, an aspect of our LEH approach that enables an additional feature, the forcing of an early end to an inner search. The LEH approach is exploration-led, the objective being to locate and move to the candidate point in the decision variable space furthest from all hcps. Hcps are designated based on the determination of a high cost threshold $\tau$, set here as the current estimate of the robust global minimum (min max) value. The nature of this approach enables (inner) uncertainty neighbourhood searches around each candidate point to be restricted when appropriate. If an inner search identifies a local point with objective function value above $\tau$ the inner search can be immediately curtailed on the basis that the candidate is not distant from hcps. This equates to the recognition that the candidate point is not an improvement on the current estimated robust optima. Such regulating of inner searches has the potential to significantly reduce the number of function evaluations expended on local neighbourhood analysis. In the case of budgetary limitations on numbers of function evaluations this further enables more exploration of the decision variable space. \subsection{Algorithm summary} Given one of our three approaches to identifying the LEH devoid of hcps, random, GA or Voronoi, the overarching algorithm for the robust exploratory LEH heuristic is given in Algorithm~\ref{LEHAlgorithm}. Here one of these three approaches to the outer search is applied in line~\ref{LEHSubRoutine} as $LEH\textunderscore Calculator(H_\tau)$, for a defined high cost set $H_\tau$. It is assumed that this routine will return a candidate point $\pmb{x}_{LEH}$ and an associated radius $r_{LEH}$, that is the minimal distance between $\pmb{x}_{LEH}$ and all points in $H_\tau$. The heuristic will halt if $r_{LEH}$ is not greater than $\Gamma$. For a defined number of initialisation points, random points in $\mathcal{X}$ are selected and the function $f$ evaluated at these points. The points and their function evaluations are recorded in history sets $H$ and $F_H$, lines~\ref{InitialisationStart} -~\ref{InitialisationEnd}. Having randomly selected a candidate point $\pmb{x}_c$ from $H$ we perform an inner maximisation in the $\Gamma$-uncertainty neighbourhood around $\pmb{x}_c$, see line~\ref{InnerSubRoutine}. The description of the inner maximisation is given below as Algorithm~\ref{InnerAlgorithm}. If this is the first candidate point, or the local robust value for this candidate $\tilde{g}(\pmb{x}_c)$ is less than the current best solution $\tau$, this minimum is updated and the associated global minimum point $\pmb{x}_{Op}$ replaced by $\pmb{x}_c$, see lines~\ref{RobustGlobalStart} -~\ref{RobustGlobalEnd}. Next the high cost set $H_\tau$ is established as all members of $H$ with corresponding function values in $F_H$ that are greater than or equal to the current high cost threshold $\tau$, see line~\ref{HighCost}. Based on $H_\tau$, the next candidate point is identified via one of the outer search approaches, see line~\ref{LEHSubRoutine}. If the heuristic is halted at this stage due to an inability to identify a valid LEH or at any stage due to the budget being exceeded, the extant estimate for the robust global minimum $\pmb{x}_{Op}$ is returned. \begin{algorithm}[htbp] \caption{Robust global exploration using Largest Empty Hyperspheres} \label{LEHAlgorithm} \vspace{2mm} \hspace*{\algorithmicindent} \textbf{Input:} $f$, $\mathcal{X}$, $\Gamma$ \\ \hspace*{\algorithmicindent} \textbf{Parameters:} $Num\textunderscore Initial$, $Budget$, $Max\textunderscore Search$ \vspace{2mm} \begin{algorithmic}[1] \ForAll{$i$ in $[Num\textunderscore Initial]$} \label{InitialisationStart} \State Choose random point $\pmb{x}^i\in\mathcal{X}$ \State Calculate $f(\pmb{x}^i)$ and store in $F_H$ \State $Budget \gets Budget-1$ \State $H \gets H \cup \{\pmb{x}^i\}$ \EndFor \label{InitialisationEnd} \State Select random point $\pmb{x}_c\in H$ \State $r_{LEH} \gets \infty$; $\tau \gets \infty$ \While{$r_{LEH} > \Gamma$} \State $\tilde{g}(\pmb{x}_c) \gets$ \textbf{CALL} Algorithm~\ref{InnerAlgorithm} \label{InnerSubRoutine} \If{$\tilde{g}(\pmb{x}_c) < \tau$} \label{RobustGlobalStart} \State $\pmb{x}_{Op} \gets \pmb{x}_c$ \State $\tau \gets \tilde{g}(\pmb{x}_c)$ \EndIf \label{RobustGlobalEnd} \State $H_{\tau} \gets \{ \pmb{x}\in H : F_H(\pmb{x}) \ge \tau \}$ \label{HighCost} \State Find ($\pmb{x}_{LEH},r_{LEH})$ by calling LEH\textunderscore Calculator($H_\tau$) \label{LEHSubRoutine} \State $\pmb{x}_c \gets \pmb{x}_{LEH}$ \EndWhile \State \Return A robust solution $\pmb{x}_{Op}$ and robust objective estimate $\tau$ \label{EndAlgorithm} \end{algorithmic} \end{algorithm} \begin{algorithm}[htbp] \caption{$\Gamma$-uncertainty neighbourhood inner maximisation} \label{InnerAlgorithm} \vspace{2mm} \hspace*{\algorithmicindent} \textbf{Input:} $Budget$, $Max\textunderscore Search$, $\pmb{x}_c$, $\Gamma$, $\tau$ \vspace{2mm} \begin{algorithmic}[1] \If{$\tau<\infty$} \State Calculate $f(\pmb{x}_c)$ and store in $F_H$ \State $H \gets H \cup \{\pmb{x}_c\}$ \State $Budget \gets Budget-1$ \If{$Budget==0$} \State GOTO line~\ref{EndAlgorithm} of Algorithm 1 \EndIf \EndIf \State Set $Local\textunderscore Robust \gets f(\pmb{x_c})$ \ForAll{$i$ in $[Max\textunderscore Search]$} \label{InnerStart} \State Choose $\Delta\pmb{x}^i_c \in \mathcal{U}$, set $\pmb{x}^i \gets \pmb{x}_c + \Delta\pmb{x}^i_c$ \label{RandomInHyper} \State Calculate $f(\pmb{x}^i_c)$ and store in $F_H$ \State $H \gets H \cup \{\pmb{x}^i_c\}$ \State $Budget \gets Budget-1$ \If{$Budget==0$} \State GOTO line line~\ref{EndAlgorithm} of Algorithm 1 \EndIf \State $Local\textunderscore Robust \gets \max\{Local\textunderscore Robust,f(\pmb{x}^i_c)\}$ \label{UpdateRobust} \If{$Local\textunderscore Robust>\tau$} \label{ThresholdStart} \State GOTO line~\ref{DefineOutput} \EndIf \label{ThresholdEnd} \EndFor \label{InnerEnd} \State $\tilde{g}(\pmb{x}_c) \gets Local\textunderscore Robust$ \label{DefineOutput} \State \Return $\tilde{g}(\pmb{x}_c)$: estimated worst case cost at $\pmb{x}_c$ \end{algorithmic} \end{algorithm} \medskip Algorithm~\ref{InnerAlgorithm}, the $\Gamma$-uncertainty neighbourhood inner maximisation called in line~\ref{InnerSubRoutine} of Algorithm~\ref{LEHAlgorithm}, requires several inputs: $Budget$ the current count of function evaluations completed, $Max\textunderscore Search$ the maximum number of function evaluations permitted in an inner search, $\pmb{x}_c$ the current candidate point (centre of an LEH) around which the inner search is to be performed, $\Gamma$ to define the uncertainty neighbourhood of $\pmb{x}_c$, and $\tau$ the high cost threshold for stopping the inner search if appropriate. Algorithm~\ref{InnerAlgorithm} proceeds by looping through up to $Max\textunderscore Search$ inner search points, identifying a point in the $\Gamma$-uncertainty neighbourhood of $\pmb{x}_c$ and evaluating the function at each point visited, lines~\ref{InnerStart} -~\ref{InnerEnd}. Here the point to be evaluated is determined by random sampling in the $\Gamma$-radius hypersphere centred on $\pmb{x}_c$, line~\ref{RandomInHyper}. Under other inner maximisation rules this would be determined by some explicit maximisation search heuristic. As the function is evaluated at the inner search points the local robust value (inner maximum) $Local\textunderscore Robust$ is updated as appropriate, line~\ref{UpdateRobust}. If $Local\textunderscore Robust$ exceeds the high cost threshold $\tau$ the inner maximisation is immediately terminated, lines~\ref{ThresholdStart} -~\ref{ThresholdEnd}. Algorithm~\ref{InnerAlgorithm} ends by returning an estimate for the worst case cost value at $\pmb{x}_c$, $\tilde{g}(\pmb{x}_c)$ into Algorithm~\ref{LEHAlgorithm}. \subsection{Example LEH application} \label{sec:NatureLEHSearch} In order to give some indication of the nature of our LEH search we have applied it to the 2-dimensional problem (poly2D) and plotted the points evaluated and associated search path of the current estimate of the robust global minimum in Figures~\ref{fig:LEHVoronoiPoints} and~\ref{fig:LEHVoronoiSearch}. Here the LEH Voronoi algorithm is used. For comparison we have also plotted corresponding results for two alternative heuristics, a robust Particle Swarm Optimisation (PSO) approach shown in Figures~\ref{fig:PSOPoints} and~\ref{fig:PSOSearch}, and the local descent directions approach from Section~\ref{sec:LocalRobustSearchDescentDirections} shown in Figures~\ref{fig:DesDirPoints} and~\ref{fig:DesDirSearch}. Here the robust PSO is used as a proxy to a brute force or co-evolutionary approach. The basic global PSO formulations have been used, as described in \cite{ShiEberhart1998}. The descent directions approach has been extended by using random re-starts, as a proxy to extending it to a global approach. In all cases inner random sampling in a hypersphere of 100 $\Gamma$-uncertainty neighbourhood points is used, and a maximum budget of 10,000 function evaluations employed. The plots shown in Figure~\ref{fig:AlternativeRobust2DSearches} are for only a single run of each heuristic, and as such should only be seen as exemplars intended to give some indication of the different natures of these outer search approaches. It can be seen that whilst the robust PSO explores the decision variable space somewhat, and the re-starting descent directions follows (exploits) a series of local paths, the LEH approach features both considerable exploration globally and more intense analysis of promising points. It is clear that the curtailing of the inner searches in the LEH approach enables much wider exploration for fewer function evaluations. In this example less than 1,000 function evaluations have been required before the LEH heuristic has stopped because an LEH of radius greater than $\Gamma$ cannot be found, but for larger (dimensional) problems such stopping prior to reaching the budgetary limit will not apply. One striking feature of Figure~\ref{fig:LEHVoronoiPoints} is how many of the inner searches stop immediately on the evaluation of a candidate point. This is because the objective value at these candidate points exceeds the current threshold $\tau$. \begin{figure}[htbp] \centering \begin{subfigure}{.38\textwidth} \centering \includegraphics[width=2.5in, height=2.6in]{PSOBertP2.pdf} \vspace{-11mm} \caption{PSO points} \label{fig:PSOPoints} \end{subfigure}% \hspace{7mm} \begin{subfigure}{.38\textwidth} \centering \includegraphics[width=2.5in, height=2.6in]{PSOBertC.pdf} \vspace{-11mm} \caption{PSO search} \label{fig:PSOSearch} \end{subfigure} \vspace{-5mm} \begin{subfigure}{.38\textwidth} \centering \includegraphics[width=2.5in, height=2.6in]{DDBertP2.pdf} \vspace{-11mm} \caption{DD points} \label{fig:DesDirPoints} \end{subfigure}% \hspace{7mm} \begin{subfigure}{.38\textwidth} \centering \includegraphics[width=2.5in, height=2.6in]{DDBertC.pdf} \vspace{-11mm} \caption{DD search} \label{fig:DesDirSearch} \end{subfigure} \vspace{-5mm} \begin{subfigure}{.38\textwidth} \centering \includegraphics[width=2.5in, height=2.6in]{LEHBertP2.pdf} \vspace{-11mm} \caption{LEH Vor points} \label{fig:LEHVoronoiPoints} \end{subfigure}% \hspace{7mm} \begin{subfigure}{.38\textwidth} \centering \includegraphics[width=2.5in, height=2.6in]{LEHBertC.pdf} \vspace{-11mm} \caption{LEH Vor search} \label{fig:LEHVoronoiSearch} \end{subfigure} \caption{Contour plots of example searches of the 2-dimensional problem (poly2D), for $\Gamma$=0.5. Plots on the left show all points evaluated. Plots on the right show the progress of the current best robust solution. The heuristics used are: (top) outer PSO, (middle) outer descent directions with re-start, and (bottom) outer LEH using the Voronoi based approach.} \label{fig:AlternativeRobust2DSearches} \end{figure} The Voronoi based search exemplified by Figures~\ref{fig:LEHVoronoiPoints} and~\ref{fig:LEHVoronoiSearch} is a good indicator of the nature of the searches due to all three LEH approaches, random, GA and Voronoi. However the radii of the LEH identified for each candidate will vary with the use of each of these algorithms. Figure~\ref{fig:AlternativeLEHSearchesRadii} in Appendix~\ref{sec:RadiiLEH} gives some indication of how the radii of the hyperspheres generated by each of these LEH heuristics progress as the exploration proceeds. \section{Computational experiments} \label{sec:ExperimentsResults} \subsection{Set up} \label{sec:SetUp} In order to assess the effectiveness of the LEH approach the heuristic has been applied to eight test problems, and results compared against the two alternative search heuristics described in Section~\ref{sec:ComparatorHeuristics}. Experiments have been performed on 2D, 4D, 7D, 10D and 100D instances of these test problems; results have also been generated for (poly2D). Both the genetic algorithm and random forms of the LEH heuristic have been assessed for all instances. The LEH Voronoi has additionally been applied to the 2D instances, with the intention of giving some indication of the differences due to a `best' LEH identifier algorithm (Voronoi) versus the alternatives. All LEH approaches are initialised by randomly sampling a single point in $\mathcal{X}$. Assuming that for most real-world problems the optimisation analysis will be limited by resources, a fixed budget of 10,000 function evaluations (model runs) is assumed. The same inner approach is employed for all heuristics. A simple random sampling in a hypersphere of 100 points in a point's local $\Gamma$-uncertainty neighbourhood is used for all instances, and the local robust maximum is estimated as the maximum due to this sampling. For the LEH approaches this inner sampling is curtailed if a point is identified in the uncertainty neighbourhood that has objective value exceeding the current high cost threshold $\tau$. All experiments have have been performed using Java, on an HP Pavilion 15 Notebook laptop computer, with 64 bit operating system, an Intel Core i3-5010U, 2.10GHz processor, and 8GB RAM. Each heuristic search has been applied to each test problem-dimension instance 50 times to reduce variability. For the solution of the Second Order Cone Problem as part of the descent directions algorithm \cite{BertsimasNohadaniTeo2010}, the IBM ILOG CPLEX Optimization Studio V12.6.3 package is called from Java. \subsection{Comparator heuristics} \label{sec:ComparatorHeuristics} Our experiments have been conducted on LEH, a re-starting descent directions, and robust PSO metaheuristics. We have applied parameter tuning to 3 of the 5 comparator heuristics -- LEH Voronoi and LEH Random do no have tunable parameters -- employing an evolutionary tuning approach using a genetic algorithm to generate a single set of parameters for each heuristic, for all test problems. For each of the 3 tuned heuristics the same subset of the test instances was used, running each member of an evolving population on each of these instances multiple times to generate mean result for each member of a population on each test instance. The performance of each individual in a population was ranked separately for each test instance, across the members of the population, leading to mean overall ranks which were used as the utility measure in tournament selection; see e.g. \cite{Eiben2012}. The effectiveness of the local descent directions approach \cite{BertsimasNohadaniTeo2010} suggests that extending this to a global search by using random re-starts will provide a reasonable comparator. A local descent directions search is undertaken from a random start point, and when this is complete it is repeated from another random start point. This is repeated until the function evaluations budget is reached. In descent directions a set of high cost points leads to the identification of an optimal stepping direction and step size, if a valid direction exists. However the algorithm includes a number of dynamically changing parameters which adapt the high cost set and enforce a minimum step size. Here we have tuned 5 parameters relating to these stages of the heuristic; see \cite{BertsimasNohadaniTeo2010} for further information. Labelled `d.d.\,Re' in the results section. As a proxy to a brute force or co-evolutionary approach an outer particle swarm search is considered. The basic formulations for the global PSO approach have been used as described in \cite{ShiEberhart1998} and 5 parameters have been tuned: swarm size, number of iterations, and for the velocity equation the $C_1$ and $C_2$ acceleration parameters and inertia weight parameter $\omega$. The combined swarm size times number of iterations was limited to 100 in order to align with the budget of 10,000 function evaluations and the level of inner sampling. Labelled `PSO' in the results section. Our robust LEH metaheuristic is considered for the three alternative ways of identifying the largest hypersphere that is empty of hcps: \begin{itemize}[leftmargin=*] \item Randomly sampled valid empty hypersphere, see Section~\ref{secLEH1}. This includes re-sampling up to 1,000 potential candidates in an attempt to identify a valid empty hypersphere, otherwise it is assumed that a valid point cannot be found and a robust global minimum has been reached. Labelled `LEH Rnd' in the results section. \item Genetic algorithm LEH, see Section~\ref{secLEH2}. Here we have tuned 6 parameters: the size of the population, number of generations, number of elites, tournament size, and mutation probability and size; we have fixed the use of tournament selection and the choice of mid-point crossover. The combined population size times number of generations was limited to 100, which is somewhat based on runtime considerations associated with the large value of $N_\tau$, the number of candidate points visited with a budget of 10,000 function evaluations. Labelled `LEH GA' in the results section. \item Voronoi based \cite{Toussaint1983} LEH, see Section~\ref{secLEH3}. Here the construction of the Voronoi diagram for the input points $H_\tau$ is performed using the Java library due to \cite{Nahr2017}. This generates geometric data, Voronoi vertices and edges, which are used to determine a set of potential candidate points -- Voronoi vertices, including those originally outside of $\mathcal{X}$ relocated to the boundary of $\mathcal{X}$ -- for the centre of the LEH. Labelled `LEH Vor' in the results section. \end{itemize} \subsection{Test functions} \label{sec:TestFunctions} \noindent A large number of test functions are available for benchmarking optimisation algorithms, and posing a variety of difficulties, see \cite{Kruisselbrink2012, JamilYang2013}. Here eight are considered, plus (poly2D) as outlined in Section~\ref{sec:ProblemDescription}. In each case a single $\Gamma$-uncertainty value is used: \begin{itemize}[leftmargin=*] \item Ackleys: feasible region [-32.768, 32.768]; $\Gamma$=3.0. \item Multipeak F1: feasible region [0, 1]; $\Gamma$=0.0625. \item Multipeak F2: feasible region [0, 10]; $\Gamma$=0.5. \item Rastrigin: feasible region [-5.12, 5.12]; $\Gamma$=0.5. \item Rosenbrock: feasible region [-2.048, 2.048]; $\Gamma$=0.25. \item Sawtooth: feasible region [-1, 1]; $\Gamma$=0.2. \item Sphere: feasible region [-5, 5]; $\Gamma$=1.0. \item Volcano: feasible region [-10, 10]; $\Gamma$=1.5. \end{itemize} The full description of these eight test functions is given in Appendix~\ref{sec:TestFunctionFormulae}. To give some indication of the nature of these functions contour plots of the 2D instances are shown in Figure~\ref{fig:NominalAndWorstCase2DTestFunctions}, for both the nominal and worst cases. \begin{figure}[htbp] \captionsetup[subfigure]{font=scriptsize,labelfont=scriptsize} \centering \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{AckleyNom.pdf} \label{fig:AckleyContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{MultipeakF1Nom.pdf} \label{fig:MultipeakF1Contour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{MultipeakF2Nom.pdf} \label{fig:MultipeakF2Contour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{RastriginNom.pdf} \label{fig:RastriginContour} \end{subfigure}% \vspace{-10mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{RosenbrockNom.pdf} \label{fig:RosenbrockContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{SawtoothNom.pdf} \label{fig:SawtoothContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{SphereNom.pdf} \label{fig:SphereContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{VolcanoNom.pdf} \label{fig:VolcanoContour} \end{subfigure}% \vspace{-10mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{AckleyWorst.pdf} \label{fig:AckleyWorstContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{MultipeakF1Worst.pdf} \label{fig:MultipeakF1WorstContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{MultipeakF2Worst.pdf} \label{fig:MultipeakF2WorstContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{RastriginWorst.pdf} \label{fig:RastriginWorstContour} \end{subfigure}% \vspace{-10mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{RosenbrockWorst.pdf} \label{fig:RosenbrockWorstContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{SawtoothWorst.pdf} \label{fig:SawtoothWorstContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{SphereWorst.pdf} \label{fig:SphereWorstContour} \end{subfigure} \hspace{4mm} \begin{subfigure}{.20\textwidth} \centering \includegraphics[width=1.5in, height=1.7in]{VolcanoWorst.pdf} \label{fig:VolcanoWorstContour} \end{subfigure}% \caption{Contour plots of nominal (top 8) and worst case (bottom 8) 2D test functions. Left to right, top to bottom: Ackley, Multipeak F1, Multipeak F2, Rastrigin, Sawtooth, Sphere and Volcano.} \label{fig:NominalAndWorstCase2DTestFunctions} \end{figure} \subsection{Results} \label{sec:Results} Results of the 50 samples runs for each heuristic applied to each test problem-dimension instance are presented here. In each run the best solution as identified by the heuristic is used. However the points in the decision variable space that have been identified as best have robust values generated using the simple inner random sampling approach, with a budget of up to 100 sample points. To better approximate the true robust values at these points their robust values have been re-estimated based on randomly sampling a large number of points (nominally 1,000,000) in the $\Gamma$-uncertainty neighbourhood of the identified robust point. This is a post processing exercise and does not affect the min max search. Mean results due to each set of 50 sample runs are shown in Tables~\ref{fig:BertMeanResults} and~\ref{fig:MeanResults}. We have applied the Wilcoxon rank-sum test with 95\% confidence to identify the statistically best approaches. Results highlighted in bold indicate the approaches that are statistically equivalent to the best one observed, for a given problem-dimension instance. Corresponding box plots, giving some indication of how the results are distributed across the 50 samples, are shown in Figures~\ref{fig:BertBoxPlotResults},~\ref{fig:BoxPlotResultsa} and~\ref{fig:BoxPlotResultsb}. Additional results, the standard deviations due to each set of 50 sample runs, the average number of candidate points visited and average number of function evaluations undertaken, are given in Appendix~\ref{sec:AdditionalResults}. \begin{table}[htbp] \begin{center} \begin{tabular}{rr|r} & & (poly2D) \\ \hline \multirow{5}{*}{2D} & PSO & 5.57 \\ & d.d.\,Re & \textbf{5.11} \\ & LEH Vor & \textbf{5.52} \\ & LEH GA & 5.50 \\ & LEH Rnd & \textbf{5.26} \end{tabular} \caption{Mean results due to 50 sample runs for the 2-dimensional polynomial function (poly2D) due to \cite{BertsimasNohadaniTeo2010}.}\label{fig:BertMeanResults} \end{center} \end{table} \begin{sidewaystable}[htbp] \begin{center} \begin{tabular}{rr|r|r|r|r|r|r|r|r} & & Ackley's & MultipeakF1 & MultipeakF2 & Rastrigin & Rosenbrock & Sawtooth & Sphere & Volcano \\ \hline \multirow{5}{*}{2D} & PSO & 11.44 & -0.36 & -0.49 & 38.04 & 10.00 & \textbf{0.49} & 1.47 & 0.39 \\ & d.d.\,Re & 12.78 & -0.40 & -0.44 & 36.42 & \textbf{7.71} & 0.54 & \textbf{1.01} & \textbf{0.24} \\ & LEH Vor & \textbf{9.36} & \textbf{-0.61} & \textbf{-0.68} & \textbf{34.67} & \textbf{7.71} & 0.59 & 1.05 & \textbf{0.24} \\ & LEH GA & 9.62 & -0.60 & -0.65 & \textbf{35.17} & \textbf{7.68} & \textbf{0.48} & 1.14 & 0.27 \\ & LEH Rnd & 9.77 & -0.59 & -0.65 & 35.52 & 7.92 & \textbf{0.47} & 1.21 & 0.29 \\ \hline \multirow{4}{*}{4D} & PSO & 13.50 & -0.30 & -0.36 & 65.91 & 34.20 & 0.50 & 3.35 & 0.75 \\ & d.d.\,Re & 17.32 & -0.33 & -0.32 & 60.43 & \textbf{11.94} & 0.60 & \textbf{1.02} & \textbf{0.46} \\ & LEH GA & \textbf{8.73} & \textbf{-0.64} & \textbf{-0.68} & \textbf{54.34} & 12.17 & \textbf{0.45} & 1.39 & \textbf{0.34} \\ & LEH Rnd & 12.21 & -0.50 & -0.57 & 61.39 & 23.18 & 0.46 & 1.70 & 0.57 \\ \hline \multirow{4}{*}{7D} & PSO & 15.36 & -0.29 & -0.23 & 102.35 & 123.42 & 0.51 & 8.21 & 1.27 \\ & d.d.\,Re & 19.72 & -0.30 & -0.24 & \textbf{88.44} & \textbf{17.47} & 0.63 & \textbf{1.03} & 1.21 \\ & LEH GA & \textbf{12.35} & \textbf{-0.51} & \textbf{-0.57} & \textbf{88.07} & 48.75 & \textbf{0.42} & 2.94 & \textbf{0.77} \\ & LEH Rnd & 16.19 & -0.42 & -0.48 & 104.31 & 126.28 & 0.52 & 9.49 & 1.37 \\ \hline \multirow{4}{*}{10D} & PSO & 16.17 & -0.31 & -0.15 & 142.99 & 238.36 & 0.51 & 14.66 & 1.63 \\ & d.d.\,Re & 20.69 & -0.30 & -0.19 & \textbf{112.61} & \textbf{41.12} & 0.63 & \textbf{1.40} & 1.93 \\ & LEH GA & \textbf{14.08} & \textbf{-0.48} & \textbf{-0.56} & \textbf{115.06} & 103.31 & \textbf{0.43} & 7.34 & \textbf{1.19} \\ & LEH Rnd & 18.11 & -0.39 & -0.43 & 145.52 & 322.27 & 0.55 & 20.62 & 1.92 \\ \hline \multirow{4}{*}{100D} & PSO & 19.02 & -0.35 & -0.17 & 1,215.34 & 7,989.77 & 0.49 & 226.66 & 4.45 \\ & d.d.\,Re & 21.38 & -0.32 & -0.32 & 1,386.77 & 36,141.80 & 0.70 & 656.86 & 6.18 \\ & LEH GA & \textbf{17.30} & \textbf{-0.44} & \textbf{-0.42} & \textbf{1,065.44} & \textbf{3,264.49} & \textbf{0.43} & \textbf{136.18} & \textbf{3.79} \\ & LEH Rnd & 21.12 & -0.36 & -0.28 & 1,577.84 & 26,526.42 & 0.66 & 588.03 & 5.93 \end{tabular} \caption{Mean results due to 50 sample runs.}\label{fig:MeanResults} \end{center} \end{sidewaystable} \begin{figure}[htbp] \centering \begin{subfigure}{\textwidth} \centering \includegraphics[width=.5\textwidth]{Bertsimas2D.pdf} \label{fig:Bertsimas2Dboxplot} \vspace*{-1.0mm} \end{subfigure} \caption{Box plots of robust objective values due to multiple sample runs for the 2-dimensional polynomial function (poly2D) due to \cite{BertsimasNohadaniTeo2010}.} \label{fig:BertBoxPlotResults} \end{figure} \begin{figure}[htbp] \centering \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim2a.pdf} \label{fig:Aboxplots2D} \vspace*{-1.0mm} \end{subfigure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim4a.pdf} \label{fig:Aboxplots4D} \vspace*{-1.0mm} \end{subfigure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim7a.pdf} \label{fig:Aboxplots7D} \vspace*{-1.0mm} \end{subfigure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim10a.pdf} \label{fig:Aboxplots10D} \vspace*{-1.0mm} \end{subfigure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim100a.pdf} \label{fig:Aboxplots100D} \vspace*{-1.0mm} \end{subfigure} \caption{Box plots of robust objective values due to multiple sample runs. Left to right: Ackleys, Multipeak F1, Multipeak F2, Rastrigin; Top to bottom: 2D, 4D, 7D, 10D, 100D.} \label{fig:BoxPlotResultsa} \end{figure} \begin{figure}[htbp] \centering \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim2b.pdf} \label{fig:Bboxplots2D} \vspace*{-1.0mm} \end{subfigure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim4b.pdf} \label{fig:Bboxplots4D} \vspace*{-1.0mm} \end{subfigure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim7b.pdf} \label{fig:Bboxplots7D} \vspace*{-1.0mm} \end{subfigure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim10b.pdf} \label{fig:Bboxplots10D} \vspace*{-1.0mm} \end{subfigure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=1.0\textwidth]{ExpDim100b.pdf} \label{fig:Bboxplots100D} \vspace*{-1.0mm} \end{subfigure} \caption{Box plots of robust objective values due to multiple sample runs. Left to right: Rosenbrock, Sawtooth, Sphere, Volcano; Top to bottom: 2D, 4D, 7D, 10D, 100D.} \label{fig:BoxPlotResultsb} \end{figure} From Table~\ref{fig:MeanResults} we see that for 100D instances the LEH GA approach is best for all test problems, and in several cases the mean LEH GA result is substantially better than all of the alternative heuristics. From Tables~\ref{fig:BertMeanResults} and~\ref{fig:MeanResults} the LEH approach is among the best in at least 6 of the instances for all other dimensions. For 2D instances the LEH Voronoi approach is among the best results for 7 of the 9 problems, whilst LEH GA and LEH Rnd are each amongst the best results for 3 and 2 problems respectively. It should also be noted that in 5 of the 7 instances where LEH Voronoi is among the best, LEH GA is either statistically equivalent or the mean value is second best. For the 2D Sphere instance {d.d.\,Re} is marginally better than LEH Voronoi, whilst {d.d.\,Re} and LEH heuristics are statistically equivalent for the (poly2D) and 2D Volcano and Rosenbrock instances. The robust PSO approach is statistically equivalent to LEH heuristics for the 2D Sawtooth instance. For the 4D -- 10D instances {d.d.\,Re} is statistically equivalent to LEH GA in the 4D Volcano problem and the 7D and 10D instances of the Rastrigin problem, and better than LEH GA for the Rosenbrock and Sphere problems. For the 4D Rosenbrock and Sphere problems the differences between {d.d.\,Re} and LEH GA are reasonably small, however in the 7D and 10D instances {d.d.\,Re} is substantially better. Considering the shape of the Rosenbrock and Sphere functions it can be expected that a local search will perform particularly well for these problems. LEH GA is better than LEH Rnd for all instances excluding (poly2D). In a number of instances LEH GA is substantially better than LEH Rnd. The number of candidate points that LEH can visit is substantially increased by the early stopping of inner searches as soon as the high cost threshold is exceeded, see Tables~\ref{fig:BertOtherResults} and~\ref{fig:CandidateResults} in Appendix~\ref{sec:AdditionalResults}. Although this feature must unquestionably play a role in the success of the LEH GA approach, the fact that LEH Rnd visits a comparable number of candidate points indicates that the additional pro active seeking of the largest hypersphere devoid of high cost points is also a significant factor in the success of LEH GA. \section{Conclusions and further work} \label{sec:SummaryConcusionsFurtherWork} We have introduced a new metaheuristic for box-constrained robust optimisation problems with implementation uncertainty. We do not assume any knowledge on the structure of the original objective function, making the approach applicable to black-box and simulation-optimisation problems. We do assume that the solution is affected by uncertainty, and the aim is to find a solution that optimises the worst possible performance in this setting. This is the min max problem. Previously, few generic search methods have been developed for this setting. We introduce a new approach for a global search based on distinguishing undesirable high cost -- high objective value -- points (hcps), identifying the largest hypersphere in the decision variable space that is completely devoid of hcps, and exploring the decision variable space by stepping between the centres of these largest empty hyperspheres. We demonstrated the effectiveness of the approach using a series of test problems, considering instances of varying dimension, and comparing our LEH approach against one metaheuristic that employs an outer particle swarm optimisation and one from the literature that uses multiple re-starts of the local descent directions approach. For low and moderate dimensional instances the approach shows competitive performance; for high-dimensional problems the LEH approach significantly outperforms the comparator heuristics for all problems. There are several ways in which this work can be developed. Further consideration can be given to the inner maximisation search approach in order to better understand the trade-off between expending function evaluations on the local $\Gamma$-radius uncertainty neighbourhood search versus globally exploring the search space, in the context of our LEH approach. The repeated calculation of large numbers of Euclidean distances each time a new LEH needs to be identified within the LEH GA heuristic is computationally expensive. Rather than only calculating a single next candidate point each time the GA is performed, identifying multiple points could speed up computation or alternatively enable the use of larger population-generation sizes to improve the estimation of the largest empty hypersphere. Results of the mid-dimension experiments on the Rosenbrock and Sphere test problems suggest that an exploitation based approach works well in these instances, indicating a direction for extending our exploration focussed LEH approach. It is clear that within the LEH algorithm the early stopping of the inner searches when it is established that the current robust global value cannot be improved upon has significant advantages. It is worth considering whether alternative search approaches could take advantage of this feature.

7475b247a0ca02c9d3024e3f15a14de3db3ed087

\section{Introduction} \label{s:intro} Community detection is a fundamental question in network analysis \citep{Goldenberg2010, NewmanPNAS, Fortunato2010}. Traditional approaches consider the adjacency matrix, whose elements equal one or zero indicating whether there is a connection between two nodes, as the input. Then the nodes are partitioned into cohesive groups, that is, communities, with homogeneous linkage probabilities within and heterogeneous probabilities between the groups. Current community detection methods assume all nodes belong to certain communities of interests. However, this assumption is not always true in real applications. For example, when we are looking for pathways involving genes related to the risk of target disease, connections between candidate genes regardless their involvement in the disease process are collected. Furthermore, whether a gene has impacts on the disease origination and development is usually unknown. Often, information on the characteristics of the nodes/genes can help to differentiate the nodes related to the outcome of interests and the unrelated ones. Novel two-stage models with one joint likelihood are proposed to incorporate the node-specific information which isolate irrelevant nodes from relevant ones and in return improve detection accuracy of communities related to a specific outcome. Our study is motivated by the problem to discover gene pathways leading to complex diseases in genomic studies. Highly correlated gene expression levels and experimentally verified protein-protein interactions provide useful information on connections between genes. However, not all genes are related to the disease under study. In fact, most genes are ``household" genes with functions to maintain normal metabolic processes within healthy human bodies. Mixing genes and pathways for normal life processes with those leading to the target disease in community detection models will introduce noise to disease-generating pathways which are the true interests of clinicians and biologists. De novo mutations refer to gene mutations that occur for the first time in a family compared to mutations inherited from parents. We believe that discrepancy in the numbers of de novo mutations on the same gene in patients and the number in healthy controls would help differentiate genes related to the disease from those unrelated to the disease, which we call the ``background''. The three kinds of data, gene expression, protein-protein interaction and number of de novo mutations, can be downloaded from different online data consortiums and combined using unique gene names. The stochastic blockmodel is the most used statistical tool for modeling and detecting communities \citep{Holland83, Snijders&Nowicki1997, Nowicki2001}. We model the relationship between the unobserved indicator whether a gene is related to the target disease or not and gene-specific covariates by logistic regression in the first stage, then cluster disease-related genes into several pathways in the second stage. Both indicators for disease relevance in the first stage and community labels in the second stage are latent variables, and the expectation-maximization algorithm is employed. However, this approach is intractable due to the numerous possible label assignments in the E-step. \citet{AAA} proposed a fast pseudo-likelihood algorithm for fitting blockmodels and we adapt this algorithm in Section \ref{sec:alg} to the joint pseudo-likelihoods incorporating both the logistic regression and the block models. The pseudo-likelihood may also be optimized by other alternative approaches such as the EMM algorithm by \citet{gormley2008mixture}. Another distinct feature of the proposed method is the extension to the robust community detection allowing heterogeneous linkage probabilities in the background, which relaxes the assumption of homogeneous linkage probability within each group in the stochastic blockmodel. For instance, the background can be a mixture of multiple strongly or weakly connected groups. These groups all belong to the background because they are not related to the target disease, but their linkage rates are not necessarily homogeneous. In Section \ref{sec:robust}, we further develop the model in section 3 to allow for arbitrary linkage patterns within the background. Interestingly, when the linkage probabilities within the background are unspecified, the pseudo-likelihood algorithm can be modified to leave the likelihood of the links in the background out while the classical likelihood approach cannot. Recently there have been works on community detection which utilize covariates information. These papers use the additional covariates information to improve the accuracy of community detection. Some papers combine a similarity or kernel matrix based on covariates with the adjacency matrix \citep{Rohe2014,Zhang2015,Sarkar2016,Xu2012}. Other papers build likelihoods of linkage probabilities incorporating auxiliary nodal information \citep{Tallberg2004,Yang13,Newman2016,Handcock2007,Hoff2009,gormley2010mixture}. However, none of these works follow the same framework as our method. In short, in our method, the sole reason of using auxiliary information on nodal characteristics is to distinguish the disease related nodes from unrelated ones, then we carry out community detection within the disease-related nodes. On the contrary, in the literature, auxiliary information is used to facilitate partition of all nodes into communities. For example, \citet{Tallberg2004} used covariates to predict the probabilities into each homogeneous community in a Bayesian framework, while we use covariates to predict the probability into the heterogeneous background in a pseudo-likelihood framework. \section{Methods} \label{sec:method} We begin by introducing the data structure and notation. A network with $n$ nodes can be represented by an $n\times n$ adjacency matrix $A=[A_{ij}]$, where \[ A_{ij}=\begin{cases} 1 & $if there is an edge between $ i $ and $ j, \\ 0 & $otherwise$ \\ \end{cases} \] In addition to the adjacency matrix $A$, some covariate information on nodes is also available. These covariates are represented by an $n \times P$ matrix $X=[x_{ip}]$, where $x_{ip}$ denotes the value of the $p$th covariate on node $i$. We model networks with a particular community structure where the network is composed of multiple cohesive communities, together with some \textit{background} nodes. Unlike the usual definition of background set which is diffuse within itself or weakly connected to other parts of the network \citep{Zhao.et.al.2011}, we assume that the probability of a node belonging to the background set depends on its covariates. Suppose there are $K$ communities besides the background set. Let $\V{c}=(c_1,c_2,...,c_n)$ denote the community that each of the $n$ nodes/genes belongs to, thus $c_i=k$ if nodes $i$ belongs to community $k$, for $k \in \{1,2,...,K\}$, and $c_i=K+1$ if node $i$ is a background gene. Moreover, let $\V{y}=[y_i]$ be a vector indicating whether the node belongs to one of the $K$ communities or the background, i.e. $y_i=1$ if $c_i\leq K$, $y_i=0$ otherwise. The network is generated in three steps. \begin{enumerate} \item The random variable $y_i$ is independent for $i=1,\cdots,n$ and follows a logistic regression \begin{align} \textnormal{pr}(y_i=1 \mid X)=\frac{e^{\V{x}_i \V{\beta}}}{1+e^{\V{x}_i\V{\beta}}}, \nonumber \end{align} where $\V{\beta}=(\beta_1,...,\beta_P)^T$ is the coefficients vector, and $\V{x}_i$ is the $i$th row of $X$. Here the logistic model has an intercept, that is, the first column of $X$ is $(1,1,...,1)^T$. \item The probability that a node with $y_i=1$ belongs each of the $K$ communities is given by the independent multinomial distribution with parameter $\V{\pi}=(\pi_1,...,\pi_K)$, \begin{align} & \textnormal{pr}(c_i=k \mid y_i=1)= \pi_k, \quad (i=1,...,n; k=1,...,K). \nonumber \end{align} In addition, $c_i=K+1$ if $y_i=0$. \item Conditional on the labels, $A_{ij}$ for $i<j$ are independent Bernoulli variables with \begin{align} \textnormal{pr}(A_{ij}=1 \mid \V{c})=P_{c_ic_j}, \nonumber \end{align} where $P$ is a $(K+1)\times (K+1)$ symmetric matrix. \end{enumerate} The total number of genes in the $k$th community is $n_k=\sum_{i=1}^n 1(c_i=k)$ and the number of links between the $k$th and $l$th commuity is given by $O_{kl}=\sum_{1\leq i, j\leq n} A_{ij}1(c_i=k,c_j=l)$, where $1(\cdot)$ is the indicator function. Moreover, let $n_{kl}=n_k n_l$ if $k \neq l$, and $n_{kk}=n_k(n_k-1)$. Then the joint log-likelihood of $\V{c}$ and $A$ is \begin{align}\label{block} \mathcal{L} (\V{\beta},\V{\pi},P ; \V{c},A)= & \sum_{i=1}^n \{ y_i \V{x}_i\V{\beta}- \log(1+e^{\V{x}_i\V{\beta}} )\}+ \sum_{k=1}^K n_k \log \pi_k \nonumber \\ & + \frac{1}{2}\sum_{1\leq k,l \leq K+1} \left \{ O_{kl} \log P_{kl}+(n_{kl}-O_{kl})\log (1-P_{kl}) \right \}. \end{align} \section{Estimating Procedures}\label{sec:alg} The community labels $\V{c}$ are unobserved in a community detection problem. Furthermore, the E-step of such algorithm requires evaluating all possible label assignments, which makes the algorithm intractable \citep{AAA,Zhaoetal2012}. We adopt the idea of pseudo-likelihood in \cite{AAA} which partitions each row of $A$ into blocks and assumes the independence between rows. We use the same notation as those in \cite{AAA}. The vector $\V{e}=(e_1,...,e_n)$ denotes an initial blocking vector, where $e_i\in \{1,...,K+1\}$. And $b_{ik}$ denotes the number of edges associated with node $i$ in the $k$th block, that is, $b_{ik}=\sum_{j=1}^n A_{ij} 1(e_j=k) \,\, (i=1,..,n; k=1,...,K+1)$. Let $B=[b_{ik}]_{1\leq i \leq n, l \leq k \leq K+1}$ and $\Lambda=[\lambda_{lk}]_{1\leq l,k \leq K+1}$, where $\lambda_{lk}$ is the expected total number of edges in the $k$-th block for a node $i$ in community $l$, i.e., $c_i=l$. When $n$ is large, $b_{ik}$ can be approximated by a Poisson distribution given $c_i$, and the dependence of $B$ between different rows is weak. Assuming $b_{ik}$ are independence for $i=1,\cdots,n$ and $k=1,\cdots,K+1$ and using the Poisson approximation, the log-pseudolikelihood of $\V{c}$ and $B$ (up to a constant) is \begin{align} \sum_{i=1}^n \{ y_i \V{x}_i\V{\beta}- \log(1+e^{\V{x}_i\V{\beta}} )\}+ \sum_{k=1}^K n_k \log \pi_k +\sum_{i=1}^n \sum_{l=1}^{K+1} 1(c_i=l) \left ( -\mu_l +\sum_{k=1}^{K+1} b_{ik} \log \lambda_{lk} \right ), \nonumber \end{align} where $\mu_{l}=\sum_{k} \lambda_{lk} \,\, (l=1,...,K+1)$. And the log-likelihood for the marginal distribution of $B$ (up to a constant) is \begin{align}\label{Poisson} \mathcal{L}_{\mbox{Poisson}} (\V{\beta},\V{\pi},\Lambda ; B)= & \sum_{i=1}^n \log \left \{ \sum_{l=1}^K \frac{e^{\V{x}_i \V{\beta} }}{1+e^{\V{x}_i \V{\beta} }}\pi_l e^{-\mu_l} \left ( \prod_{k=1}^{K+1} \lambda_{lk}^{b_{ik}} \right ) \right . \nonumber \\ & + \left . \frac{1}{1+e^{\V{x}_i \V{\beta} }} e^{-\mu_{K+1}} \left ( \prod_{k=1}^{K+1} \lambda_{K+1,k}^{b_{ik}} \right ) \right \}. \end{align} Given initial labels $\V{e}$, equation \eqref{Poisson} can be maximized by a standard expectation-maximization algorithm. The details of the E-step and M-step are given in Algorithm 1. \textbf{Algorithm 1:} (The expectation-maximization algorithm under Poisson distribution) \begin{itemize} \item E-step: Let $\hat{\V{\beta}}, \hat{\V{\pi}}$ and $\hat{\Lambda}$ be the estimates at the current iteration, and $\hat{\mu}_l=\sum_{k} \hat{\lambda}_{lk}\,\, (l=1,...,K+1)$. The posterior probability of label assignment is \begin{align} z_{il} & =\textnormal{pr}(c_i=l \mid B ) \nonumber \\ & =\frac{\frac{e^{\V{x}_i\hat{\V{\beta} }}}{1+e^{\V{x}_i \hat{\V{\beta} }}}\hat{\pi}_l e^{-\hat{\mu}_l} \left ( \prod_{k=1}^{K+1} \hat{\lambda}_{lk}^{b_{ik}} \right )}{\sum_{l=1}^K \frac{e^{\V{x}_i\hat{\V{\beta}} }}{1+e^{\V{x}_i\hat{\V{\beta} }}}\hat{\pi}_l e^{-\hat{\mu}_l} \left ( \prod_{k=1}^{K+1} \hat{\lambda}_{lk}^{b_{ik}} \right ) + \frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} e^{-\hat{\mu}_{K+1}} \left ( \prod_{k=1}^{K+1} \hat{\lambda}_{K+1,k}^{b_{ik}} \right ) } \nonumber \\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad (i=1,...,n; l=1,...,K), \nonumber \\ z_{i,K+1}& =\textnormal{pr}(c_i=K+1 \mid B ) \nonumber \\ & =\frac{\frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} e^{-\hat{\mu}_{K+1}} \left ( \prod_{k=1}^{K+1} \hat{\lambda}_{K+1,k}^{b_{ik}} \right )}{\sum_{l=1}^K \frac{e^{\V{x}_i\hat{\V{\beta}} }}{1+e^{\V{x}_i\hat{\V{\beta} }}}\hat{\pi}_l e^{-\hat{\mu}_l} \left ( \prod_{k=1}^{K+1} \hat{\lambda}_{lk}^{b_{ik}} \right ) + \frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} e^{-\hat{\mu}_{K+1}} \left ( \prod_{k=1}^{K+1} \hat{\lambda}_{K+1,k}^{b_{ik}} \right ) } \nonumber \\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad (i=1,...,n). \nonumber \end{align} \item M-step: Given $z_{il} \,\, (i=1,...n; l=1,...,K+1)$, $\hat{\pi}$ and $\hat{\Lambda}$ can be updated by the following formulae, \begin{align} \hat{\pi}_{l} & = \frac{\sum_{i} z_{il}}{\sum_{i} \sum_{l=1}^K z_{il} } \quad (l=1,...,K), \nonumber \\ \hat{\lambda}_{lk} & =\frac{\sum_i z_{il} b_{ik}}{\sum_i z_{il}} \quad(l=1,...,K+1; k=1,...,K+1) . \nonumber \end{align} $\hat{\V{\beta}}$ can be updated by solving the logistic regression, \begin{align} \hat{\V{\beta}} & = \argmax_{\V{\beta}} \sum_{i=1}^n \left \{ \left (\sum_{l=1}^K z_{il} \right ) \V{x}_i\V{\beta}- \log(1+e^{\V{x}_i\V{\beta}} ) \right \}. \nonumber \end{align} Note $\sum_{l=1}^Kz_{il}$ is the sum of the estimated conditional probabilities of gene $i$ belonging to one of the $K$ communities. \end{itemize} Once the expectation-maximization algorithm converges, we can update the labels $\V{e}$ by $e_i=\argmax_{1\leq l\leq K+1} z_{il}$. We repeat this procedure several times until $\V{e}$ becomes stable. \cite{AAA} also introduced a pseudo-likelihood conditional on the node degrees. We generalize this conditional pseudo-likelihood to our scenario. Denote the node degree by $d_i=\sum_{k} b_{ik} \,\, (i=1,...,n)$. Then $(b_{i1},...,b_{i,K+1})$ follows multinomial distribution conditional on label $\V{c}$ and $d_i$. The multinomial log pseudo-likelihood (up to a constant) is \begin{align}\label{Multinomial} \mathcal{L}_{\mbox{Multinomial}} (\V{\beta},\V{\pi},\Theta ; B)= & \sum_{i=1}^n \log \left \{ \sum_{l=1}^K \frac{e^{\V{x}_i \V{\beta} }}{1+e^{\V{x}_i \V{\beta} }}\pi_l\left ( \prod_{k=1}^{K+1} \theta_{lk}^{b_{ik}} \right ) \right . \\ & + \left . \frac{1}{1+e^{\V{x}_i \V{\beta} }} \left ( \prod_{k=1}^{K+1} \theta_{K+1,k}^{b_{ik}} \right ) \right \}, \nonumber \end{align} where $\Theta=[\theta_{lk}]$ $(l=1,...,K+1;k=1,...,K+1)$ is the parameter in the multimomial distribution satisfying $\sum_{k=1}^{K+1}\theta_{lk}=1 (l=1,...,K+1)$. The algorithm is similar to that for the Poisson pseudo-likelihood. We give the details of the expectation-maximization algorithm under the multinomial distribution in Algorithm 2. \textbf{Algorithm 2:} (The expectation-maximization algorithm under multinomial distribution) \begin{itemize} \item E-step: Based on current estimates $\hat{\V{\beta}}, \hat{\V{\pi}}$ and $\hat{\Theta}$, the posterior probability of label assignment is \begin{align} & z_{il} =\frac{\frac{e^{\V{x}_i\hat{\V{\beta} }}}{1+e^{\V{x}_i \hat{\V{\beta} }}}\hat{\pi}_l \left ( \prod_{k=1}^{K+1} \hat{\theta}_{lk}^{b_{ik}} \right )}{\sum_{l=1}^K \frac{e^{\V{x}_i\hat{\V{\beta}} }}{1+e^{\V{x}_i\hat{\V{\beta} }}}\hat{\pi}_l \left ( \prod_{k=1}^{K+1} \hat{\theta}_{lk}^{b_{ik}} \right ) + \frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} \left ( \prod_{k=1}^{K+1} \hat{\theta}_{K+1,k}^{b_{ik}} \right ) } \quad (i=1,...,n; l=1,...,K), \nonumber \\ & z_{i,K+1} =\frac{\frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} \left ( \prod_{k=1}^{K+1} \hat{\theta}_{K+1,k}^{b_{ik}} \right )}{\sum_{l=1}^K \frac{e^{\V{x}_i\hat{\V{\beta}} }}{1+e^{\V{x}_i\hat{\V{\beta} }}}\hat{\pi}_l \left ( \prod_{k=1}^{K+1} \hat{\theta}_{lk}^{b_{ik}} \right ) + \frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} \left ( \prod_{k=1}^{K+1} \hat{\theta}_{K+1,k}^{b_{ik}} \right ) } \quad (i=1,...,n). \nonumber \end{align} \item M-step: Given $z_{il} \,\, (i=1,...n; l=1,...,K+1)$, $\hat{\pi}$, $\hat{\Theta}$ and $\hat{\V{\beta}}$ can be updated by \begin{align} \hat{\pi}_{l} & = \frac{\sum_{i} z_{il}}{\sum_{i} \sum_{l=1}^K z_{il} } \quad (l=1,...,K), \nonumber \\ \hat{\theta}_{lk} & =\frac{\sum_i z_{il} b_{ik}}{\sum_i z_{il} d_i} \quad(l=1,...,K+1; k=1,...,K+1), \nonumber \\ \hat{\V{\beta}} & = \argmax_{\V{\beta}} \sum_{i=1}^n \left \{ \left (\sum_{l=1}^K z_{il} \right ) \V{x}_i\V{\beta}- \log(1+e^{\V{x}_i\V{\beta}} ) \right \}. \nonumber \end{align} \end{itemize} \section{Robust Community Detection }\label{sec:robust} So far we assume that all the disease-related communities and the background satisfy the stochastic blockmodel assumption. In this section, we propose a new pseudo-likelihood method that allows for arbitrary structure in the background, for example, a mixture of strongly and weakly connected groups, or nodes with high degree variations. In other words, we keep the stochastic blockmodel assumption in the disease-related communities, but make no assumption on the structure within the background. A network with the heterogeneous background is generated in three steps, of which the first two steps are identical to the first two steps in Section 2. The last step has been modified as follows. \textsc{Step} 3$^*$: Conditional on the labels, when $k \leq K$ or $l \leq K$, $A_{ij}$ for $i<j$ are independent Bernoulli variables with \begin{align} \textnormal{pr}(A_{ij}=1 \mid c_i=k,c_j=l)=P_{k l}.\nonumber \end{align} The link probabilities within the background set, i.e., when $k=K+1$ and $l=K+1$, are not specified. The joint likelihood (\ref{block}) cannot be used as the criteria to estimate c in this situation because it is maximized when all nodes belong to group $K+1$. By contrast, the pseudo-likelihood method introduced in Section \ref{sec:alg} can be extended to this new scenario. Recall the setup in Section \ref{sec:alg}. Let $\V{e}=(e_1,...,e_n)$ be an initial blocking vector. And $b_{ik}$ denotes the number of edges associated with node $i$ in the $k$th block $(i=1,..,n; k=1,...,K+1)$. When $\V{e}$ is a reasonable initial vector, $b_{ik}$ can be approximated by a mixture of Poisson distributions as before when $k=1,...,K$. However, when $k=K+1$, the distribution of $b_{ik}$ is unknown since the link probabilities within the background are unspecified. Therefore, we exclude this part of unreliable information, and propose the following pseudo-likelihood for robust community detection, \begin{align}\label{Robust} \mathcal{L}_{\mbox{Robust}} (\V{\beta},\V{\pi},\Lambda ; B, \V{c})= & \sum_{i=1}^n \{ y_i \V{x}_i\V{\beta}- \log(1+e^{\V{x}_i\V{\beta}} )\}+ \sum_{k=1}^K n_k \log \pi_k \nonumber \\ & +\sum_{i=1}^n \sum_{l=1}^{K+1} 1(c_i=l) \left ( -\mu_l +\sum_{k=1}^{K} b_{ik} \log \lambda_{lk} \right ), \end{align} where $\mu_l= \sum_{k=1}^K \lambda_{lk}\,\, (k=1,..,K)$. Notice equation \eqref{Robust} is indeed a valid likelihood function conditional on $\V{e}$ because the blocking vector $\V{e}$ and the community labeling vector $\V{c}$ are treated differently in our algorithm. The proof of its identifiability is given in the supplementary material. The blocking vector $\V{e}$ partitions the columns of $A$ into $K+1$ blocks and $b_{ik}$ is the sum in the $i$th row $k$th block. Likelihood \eqref{Robust} does not include $B_{\cdot,K+1}$ -- the last column of $B$ since the Poisson approximation are inappropriate. But this does not affect the range of $c_i$, which is still $\{1,...,K+1\}$. Community detection based on \eqref{Robust} can be viewed as a classic clustering problem on $B$. We need to assign a label from 1 to $K+1$ to each row data point, i.e., each $B_{i\cdot}$, which contains $K+1$ features. But we only use the first $K$ features since the last one is not reliable. Then we update the labelling of each gene in the columns and iterate several times. In each iteration, because the groups in the columns and the grouping results in the rows are considered separately, dropping one column of noise will not result in all genes (rows) falling into the $(K+1)$th group. After each iteration of the outer loop updating $e_i$, genes with similar linkage probabilities with the first $K$ groups are classified into the same group with higher and higher accuracy. The algorithm is therefore similar to Algorithm 1 and given in the following. \textbf{Algorithm 3:} (The expectation-maximization algorithm for robust community detection) \begin{itemize} \item E-step: Let $\hat{\V{\beta}}, \hat{\V{\pi}}$ and $\hat{\Lambda}$ be the estimates at the current iteration, and $\hat{\mu}_l=\sum_{k=1}^K \hat{\lambda}_{lk}\,\, (l=1,...,K+1)$. The posterior probability of label assignment is \begin{align} z_{il} & =\textnormal{pr}(c_i=l \mid B ) \nonumber \\ & =\frac{\frac{e^{\V{x}_i\hat{\V{\beta} }}}{1+e^{\V{x}_i \hat{\V{\beta} }}}\hat{\pi}_l e^{-\hat{\mu}_l} \left ( \prod_{k=1}^{K} \hat{\lambda}_{lk}^{b_{ik}} \right )}{\sum_{l=1}^K \frac{e^{\V{x}_i\hat{\V{\beta}} }}{1+e^{\V{x}_i\hat{\V{\beta} }}}\hat{\pi}_l e^{-\hat{\mu}_l} \left ( \prod_{k=1}^{K} \hat{\lambda}_{lk}^{b_{ik}} \right ) + \frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} e^{-\hat{\mu}_{K+1}} \left ( \prod_{k=1}^{K} \hat{\lambda}_{K+1,k}^{b_{ik}} \right ) } \nonumber \\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad (i=1,...,n; l=1,...,K), \nonumber \\ z_{i,K+1}& =\textnormal{pr}(c_i=K+1 \mid B ) \nonumber \\ & =\frac{\frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} e^{-\hat{\mu}_{K+1}} \left ( \prod_{k=1}^{K} \hat{\lambda}_{K+1,k}^{b_{ik}} \right )}{\sum_{l=1}^K \frac{e^{\V{x}_i\hat{\V{\beta}} }}{1+e^{\V{x}_i\hat{\V{\beta} }}}\hat{\pi}_l e^{-\hat{\mu}_l} \left ( \prod_{k=1}^{K} \hat{\lambda}_{lk}^{b_{ik}} \right ) + \frac{1}{1+e^{\V{x}_i\hat{\V{\beta}} }} e^{-\hat{\mu}_{K+1}} \left ( \prod_{k=1}^{K} \hat{\lambda}_{K+1,k}^{b_{ik}} \right ) } \nonumber \\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad (i=1,...,n). \nonumber \end{align} \item M-step: Given $z_{il} \,\, (i=1,...n; l=1,...,K+1)$, $\hat{\pi}$, $\hat{\Lambda}$ and $\hat{\V{\beta}}$ can be updated by, \begin{align} \hat{\pi}_{l} & = \frac{\sum_{i} z_{il}}{\sum_{i} \sum_{l=1}^K z_{il} } \quad (l=1,...,K), \nonumber \\ \hat{\lambda}_{lk} & =\frac{\sum_i z_{il} b_{ik}}{\sum_i z_{il}} \quad(l=1,...,K+1; k=1,...,K) , \nonumber \\ \hat{\V{\beta}} & = \argmax_{\V{\beta}} \sum_{i=1}^n \left \{ \left (\sum_{l=1}^K z_{il} \right ) \V{x}_i\V{\beta}- \log(1+e^{\V{x}_i\V{\beta}} ) \right \}. \nonumber \end{align} \end{itemize} As before, once the expectation-maximization algorithm converges, $\V{e}$ is updated by $e_i= \\ \argmax_{1\leq l\leq K+1} z_{il}$. We repeat this procedure until $\V{e}$ becomes stable. We do not consider robust community detection using multinomial approximation because the condition $\sum_{k=1}^{K+1}\theta_{lk}=1 (l=1,...,K+1)$ is invalid if the last column is removed. \section{Asymptotic Properties}\label{sec:proof} In this section we study the consistency under stochastic blockmodels. Equation \eqref{Multinomial} has slightly simpler form and theoretical derivations than \eqref{Poisson}. The theoretical analysis in this section will focus on the multinomial pseudo-likelihood. We begin with the setup, which closely follow \cite{AAA}. The true community labels $\V{c}$ are the parameters of interests, where $\pi_k=1/n\sum\limits_{i}1(c_i=k) \,\, (k=1,2)$. We focus on the case of directed blockmodel. A coupling technique can be used to extend the result to the undirected case analogous to that in \cite{AAA}. Consider the edge matrix \begin{align*} P=\frac{1}{n}\left( \begin{array}{cc} a_1 & b \\ b & a_2 \\ \end{array} \right)=\frac{b}{n}\left( \begin{array}{cc} \rho_1 & 1 \\ 1 & \rho_2 \\ \end{array} \right), \end{align*} where $\rho_k={a_k}/{b}$. Here $\rho_1$ and $\rho_2$ remain constant, while $b$ can scale with $n$. The directed blockmodel assumes that all the entries in the adjacency matrix are independent Bernoulli variables without forcing $P$ to be symmetric, that is, $A_{ij} \sim \mbox{Bernoulli}(P_{c_ic_j}) \,\, (i=1,...,n;j=1,...,n)$. For simplicity, a univariate covariate $\V{x}$ taking values in $(1/n,2/n,...,1)$ is assumed. We illustrate the consistency of one-step expectation-maximization of the multinomial pseudo-likelihood. Starting from some initial labels $\V{e}$ and initial estimates $\hat{b}$, $\hat{\rho}_1, \hat{\rho}_2$ of the parameters $b$, $\rho_1$ and $\rho_2$, the initial estimates of $\hat{\beta}_0$ and $\hat{\beta}_1$ are obtained from the logistic regression, that is, \begin{align*} (\hat{\beta}_0, \hat{\beta}_1) & = \argmax_{\beta_0,\beta_1} \sum_{i=1}^n \left \{ y_i(\beta_0+x_i\beta_1)- \log(1+e^{\beta_0+x_i\beta_1} ) \right \}. \end{align*} Define \begin{align*} \hat{\pi}_{i1} & = \frac{e^{\hat{\beta}_0+x_i\hat{\beta}_1} }{1+e^{ \hat{\beta}_0+x_i\hat{\beta}_1 }} \quad (i=1,...,n), \\ \hat{\pi}_{i2} & = \frac{1}{1+e^{ \hat{\beta}_0+x_i\hat{\beta}_1 }} \quad (i=1,...,n). \end{align*} Let \begin{align*} \hat{P}=\frac{\hat{b}}{n}\left( \begin{array}{cc} \hat{\rho}_1 & 1 \\ 1 & \hat{\rho}_2 \\ \end{array} \right), \end{align*} and $R$ be the 2 by 2 matrix with entries $\{R_{ka}\}$ given by $R_{ka}=(1/n)\sum\limits_{i=1}^n 1 (e_i=k,c_i=a)$. The initial estimates $\hat{\Theta}$ is obtained by row normalization of $\hat{\Lambda}=[n R \hat{P}]^T $, that is, \begin{align*} \hat{\Theta}=\left( \begin{array}{cc} \frac{\hat{\lambda}_{11}}{\hat{\lambda}_{11}+\hat{\lambda}_{12}} & \frac{\hat{\lambda}_{12}}{\hat{\lambda}_{11}+\hat{\lambda}_{12}} \\ \frac{\hat{\lambda}_{21}}{\hat{\lambda}_{21}+\hat{\lambda}_{22}} & \frac{\hat{\lambda}_{22}}{\hat{\lambda}_{21}+\hat{\lambda}_{22}} \\ \end{array} \right). \end{align*} With the notation defined above, the output of one-step expectation-maximization is \begin{align*} \hat{c}_i(\V{e})=\argmax_{k \in \{1,2\}} \left (\log \hat{\pi}_{ik}+\sum_{l=1}^2 b_{il}\log \hat{\theta}_{kl} \right ) \quad (i=1,...,n). \end{align*} We use the mis-classification error rate \citep{Choietal2011,Zhaoetal2012,AAA} to measure the performance of $\hat{c}_i$. That is, define \begin{align*} M_n(\V{e})=\min_{\phi \in \{ (1 2), (2 1) \} }\frac{1}{n} \sum_{i=1}^n 1 \{ \hat{c}_i(\V{e}) \neq \phi(c_i) \}, \end{align*} where $\{(1 2),(2 1)\}$ is the set of permutations of $\{1,2\}$. In this definition we consider all $\phi$ values that are permutations of each other because they result in the same community structure. Consider the class of initial labels that correctly classify the node $i$ as a member of community $k$. The fraction of such nodes among all nodes belonging to community $k$, $\gamma_k$, is formally given by \begin{align*} \mathcal{E}=\{\V{e}: \sum\limits_{i}1(e_i=k,c_i=k)=\gamma_k n_k, k=1,2\}, \end{align*} where $n_k=\sum\limits_{i}1(c_i=k)$ is the size of community $k$. An extra condition is introduced to avoid perfect separation of $\V{e}$ in the logistic fit. We define the following class \begin{align*} \mathcal{F}=\{\V{e}: \sum_{i=\hat{n}_2+1}^n 1(e_i=1 )\leq \hat{n}_1 \tilde{\gamma}_1, \sum_{i=1}^{\hat{n}_1} 1(e_i=1 )\leq \hat{n}_1 \tilde{\gamma}_2 \}, \end{align*} where $\hat{n}_k=\sum\limits_{i}1(e_i=k)$ is the size of initial estimate of community $k$. The uniform consistency of $\hat{c}_i$ within the class $\mathcal{E} \cap \mathcal{F}$ is established by the following theorem. \begin{thm}[Main result] \label{mainThm} Assume $\gamma_1, \gamma_2 \neq 1/2$ and $0<\tilde{\gamma}_1, \tilde{\gamma}_2<1$. Then under some regularity condition, with sufficiently large $\hat{\rho}_1$, $\hat{\rho}_2$ and $b \rightarrow \infty$, for any $\epsilon$, \begin{align*} \textnormal{pr} \left [\sup_{\V{e}\in \mathcal{E} \cap \mathcal{F} } M_n(\V{e})>\epsilon \right ] \rightarrow 0, \quad \mbox{as} \,\, n \rightarrow \infty. \end{align*} \end{thm} The details of the regularity condition and the proof is given in the supplementary material. The proof of the main theorem depends on a key fact that the log ratio of the estimated probabilities $\hat{\pi}_{i1}$ and $\hat{\pi}_{i2}$ has a uniform bound independent with $n$, for $\V{e} \in \mathcal{F} \cap \mathcal{E}$. This is summarized in the following lemma. \begin{lemma} \label{lemma1} Assume $0<\tilde{\gamma}_1, \tilde{\gamma}_2<1$. Then if $\V{e} \in \mathcal{F} \cap \mathcal{E}$, there exist $M$ such that for sufficiently large $n$, \begin{align*} \left | \log \frac{\hat{\pi}_{i1}}{\hat{\pi}_{i2}} \right |<M, \end{align*} where $M$ is independent with $n$. \end{lemma} The proof is given in the supplementary material. \section{Simulations} \label{sec:sim} We first examine the performance of the proposed methods under standard stochastic blockmodel. Each network contains $n=500$ nodes and each setup is repeated 500 times. There are three groups including two disease-related communities and one disease-irrelevant background set. The probability a gene is related to the disease follows a logistic regression with $\mbox{logit} \mbox{ pr}(y_i=1 \mid x_i) = 4 x_i+\beta_0$. Here $y_i$ is the indicator for the $i$th node belonging to a disease-related community and covariate $x_i \sim U(-1,1)$. And $\beta_0=-1,0,1$ correspond to the percentages background $62\%$, $50\%$ and $38\%$, respectively. Nodes with $y_i=1$ are assigned to two non-overlapping communities with equal probabilities $\pi_1=\pi_2=1/2$. Pairs within the background, as well as pairs composed of one node in the background and the other node in a disease-related community are linked with probability $0$$\cdot$$1$. The linkage probability between the two non-background communities is 0$\cdot$05, while the linkage probability for pairs within the same community ranges from 0$\cdot$15 to 0$\cdot$25. \begin{table} \def~{\hphantom{0}} \caption{Comparison of average adjusted rand index (ARI) $\times 100$ under stochastic blockmodels. Numbers within parentheses are empirical standard deviations of ARI $\times 100$. } \begin{center} \scalebox{0.7}{ \begin{tabular}{lccccccc} \\ \hline \hline &\multicolumn{3}{c}{With Logistic Models}&&\multicolumn{3}{c}{Without Logistic Models} \\ \cline{2-4} \cline{6-8} $p_{11}$ &Poisson & Multinomial & Robust && Poisson & Multinomial & Robust \\ \hline \multicolumn{8}{l}{$62\%$ Background Nodels} \\ 15 & 58 (12) & 57 (13) & 59 (12) && 15 (7) & 15 (8) & 15 (8) \\ 16 & 66 (8) & 66 (9) & 67 (8) && 23 (11) & 24 (11) & 23 (11) \\ 17 & 72 (7) & 72 (6) & 73 (7) && 34 (13) & 33 (13) & 33 (13) \\ 18 & 77 (5) & 76 (5) & 77 (5) && 48 (14) & 45 (13) & 46 (15) \\ 19 & 81 (5) & 80 (5) & 81 (4) && 61 (11) & 55 (13) & 60 (11) \\ 20 & 85 (4) & 83 (4) & 85 (4) && 70 (8) & 66 (9) & 70 (9) \\ 21 & 88 (3) & 86 (4) & 88 (3) && 78 (6) & 73 (7) & 78 (7) \\ 22 & 91 (3) & 88 (3) & 91 (3) && 83 (4) & 79 (6) & 83 (5) \\ 23 & 93 (3) & 90 (3) & 93 (3) && 87 (4) & 83 (5) & 87 (4) \\ 24 & 94 (2) & 92 (3) & 94 (2) && 90 (3) & 86 (4) & 90 (3) \\ 25 & 96 (2) & 93 (2) & 96 (2) && 93 (3) & 89 (3) & 93 (3) \\ \hline \multicolumn{8}{l}{$50\%$ Background Nodels} \\ 15 & 74 (5) & 74 (5) & 74 (5) && 44 (10) & 44 (10) & 43 (11) \\ 16 & 78 (4) & 78 (4) & 79 (4) && 56 (8) & 56 (8) & 55 (10) \\ 17 & 82 (4) & 82 (4) & 82 (4) && 66 (6) & 64 (7) & 66 (7) \\ 18 & 86 (3) & 85 (4) & 86 (3) && 74 (6) & 72 (6) & 74 (6) \\ 19 & 89 (3) & 88 (3) & 89 (3) && 80 (5) & 78 (5) & 80 (5) \\ 20 & 91 (3) & 90 (3) & 92 (3) && 86 (4) & 82 (4) & 86 (4) \\ 21 & 94 (2) & 92 (3) & 94 (2) && 89 (3) & 86 (4) & 89 (3) \\ 22 & 95 (2) & 93 (2) & 95 (2) && 92 (3) & 89 (3) & 92 (3) \\ 23 & 96 (2) & 95 (2) & 97 (2) && 94 (2) & 91 (3) & 94 (2) \\ 24 & 98 (1) & 96 (2) & 97 (1) && 96 (2) & 93 (3) & 96 (2) \\ 25 & 98 (1) & 96 (2) & 98 (1) && 97 (1) & 94 (2) & 97 (1) \\ \hline \multicolumn{8}{l}{$38\%$ Background Nodels} \\ 15 & 82 (4) & 82 (4) & 82 (3) && 67 (6) & 67 (6) & 67 (6) \\ 16 & 86 (3) & 86 (3) & 86 (3) && 74 (5) & 74 (5) & 74 (5) \\ 17 & 89 (3) & 88 (3) & 89 (3) && 81 (4) & 80 (4) & 80 (4) \\ 18 & 91 (3) & 91 (3) & 91 (3) && 85 (3) & 84 (4) & 85 (3) \\ 19 & 94 (2) & 92 (2) & 94 (2) && 89 (3) & 87 (3) & 89 (3) \\ 20 & 96 (2) & 94 (2) & 96 (2) && 92 (2) & 90 (3) & 92 (2) \\ 21 & 97 (1) & 95 (2) & 97 (1) && 95 (2) & 92 (2) & 95 (2) \\ 22 & 98 (1) & 96 (2) & 98 (1) && 96 (2) & 94 (2) & 96 (2) \\ 23 & 98 (1) & 97 (1) & 98 (1) && 97 (1) & 95 (2) & 97 (1) \\ 24 & 99 (1) & 97 (1) & 99 (1) && 98 (1) & 96 (2) & 98 (1) \\ 25 & 99 (1) & 98 (1) & 99 (1) && 99 (1) & 97 (2) & 99 (1) \\ \hline \hline \end{tabular} } \end{center} \label{table1} \end{table} Table \ref{table1} compares the performance of three models - the pseudo-likelihood methods with Poisson and multinomial approximation introduced in Section \ref{sec:alg} as well as the robust community detection method introduced in Section \ref{sec:robust}. For each model, we further compare the two versions where auxiliary nodal information, i.e, logistic regression, is either used or unused. The community detection accuracy is measured by the adjusted rand index (\textsc{ari}) \citep{Vinh10}, a widely-used measure for comparing two partitions. The value of the index is 0 for two independent partitions, and higher values indicate better agreement. The performance of all methods improves as the linkage probability within disease-related community increases, or as the percentage of background nodes decreases. More importantly, the proposed method incorporating auxiliary information through logistic regression always outperforms its counterpart without logistic regression. Moreover, the robust method gives the same performance as the Poisson pseudo-likelihood which suggests the robust method does not lose discriminatory accuracy when data follow standard stochastic block models. On the other hand, the algorithm fitting multinomial distributions performs slightly worse than the other two methods. Rigorously speaking, the multinomial pseudo-likelihood is an approximation to the degree-corrected blockmodel, which is a generalization of standard blockmodel by allowing more variation on degrees \citep{Zhaoetal2012,Karrer10,AAA}. Therefore, the finite sample performance of multinomial pseudo-likelihood has slightly lower ARI on average since it fits a more complicated model. Next we consider the setup with heterogeneous background nodes. For any node $i$ in background, we generate $u_i$ from $U(0, 0$$\cdot$$2)$. The linkage probability between a background node $i$ and a disease-related node is $u_i$. For two background nodes $i$ and $j$, the linkage probability is $\sqrt{u_iu_j}$. The rest of the model setups such as the generation mechanism of communities labels, the linkage probabilities within/between communities and linkage probabilities between a community and the background remain the same. \begin{table} \def~{\hphantom{0}} \caption{Comparison of average adjusted rand index (ARI) $\times 100$ under heterogeneous backgrounds. Numbers within parentheses are empirical standard deviations of ARI $\times 100$. } \begin{center} \scalebox{0.7}{ \begin{tabular}{lccccccc} \\ \hline \hline &\multicolumn{3}{c}{With Logistic Models}&&\multicolumn{3}{c}{Without Logistic Models} \\ \cline{2-4} \cline{6-8} $p_{11}$ &Poisson & Multinomial & Robust && Poisson & Multinomial & Robust \\ \hline \multicolumn{8}{l}{$62\%$ Background Nodes} \\ 15 & 20 (11) & 58 (14) & 54 (21) && 15 (6) & 17 (8) & 12 (10) \\ 16 & 23 (13) & 65 (9) & 63 (18) && 18 (5) & 23 (10) & 17 (12) \\ 17 & 25 (13) & 70 (8) & 69 (16) && 20 (5) & 30 (12) & 24 (17) \\ 18 & 29 (14) & 74 (6) & 76 (11) && 23 (5) & 39 (13) & 31 (21) \\ 19 & 35 (18) & 78 (5) & 81 (8) && 24 (5) & 50 (12) & 43 (26) \\ 20 & 39 (20) & 80 (5) & 85 (6) && 27 (6) & 57 (12) & 50 (27) \\ 21 & 43 (23) & 83 (5) & 88 (5) && 29 (5) & 63 (10) & 61 (27) \\ 22 & 48 (25) & 85 (4) & 91 (3) && 30 (6) & 66 (10) & 71 (25) \\ 23 & 53 (27) & 86 (4) & 93 (3) && 32 (7) & 69 (10) & 78 (22) \\ 24 & 60 (29) & 87 (4) & 94 (2) && 34 (9) & 72 (10) & 84 (19) \\ 25 & 67 (30) & 89 (4) & 95 (2) && 37 (13) & 74 (10) & 89 (15) \\ \hline \multicolumn{8}{l}{$50\%$ Background Nodes} \\ 15 & 62 (19) & 73 (5) & 74 (5) && 34 (12) & 44 (9) & 42 (12) \\ 16 & 70 (15) & 77 (4) & 79 (4) && 41 (15) & 53 (9) & 55 (11) \\ 17 & 75 (14) & 81 (4) & 83 (4) && 46 (15) & 62 (8) & 66 (8) \\ 18 & 81 (10) & 84 (4) & 86 (3) && 52 (15) & 69 (6) & 74 (7) \\ 19 & 85 (9) & 86 (4) & 89 (3) && 58 (15) & 74 (6) & 80 (6) \\ 20 & 89 (7) & 89 (3) & 92 (3) && 63 (17) & 78 (5) & 85 (5) \\ 21 & 92 (4) & 90 (3) & 93 (2) && 72 (18) & 82 (5) & 89 (3) \\ 22 & 95 (2) & 92 (3) & 95 (2) && 82 (16) & 84 (5) & 92 (2) \\ 23 & 96 (2) & 93 (2) & 96 (2) && 88 (15) & 87 (4) & 94 (2) \\ 24 & 97 (2) & 94 (2) & 97 (1) && 93 (10) & 88 (4) & 96 (2) \\ 25 & 98 (1) & 94 (2) & 98 (1) && 96 (7) & 90 (4) & 97 (1) \\ \hline \multicolumn{8}{l}{$38\%$ Background Nodes} \\ 15 & 81 (5) & 82 (4) & 82 (4) && 65 (8) & 66 (6) & 67 (7) \\ 16 & 85 (4) & 85 (3) & 86 (3) && 71 (7) & 72 (5) & 74 (5) \\ 17 & 89 (3) & 88 (3) & 89 (3) && 77 (7) & 78 (5) & 81 (4) \\ 18 & 91 (3) & 90 (3) & 91 (2) && 83 (6) & 82 (4) & 85 (4) \\ 19 & 93 (2) & 92 (3) & 94 (2) && 88 (4) & 86 (4) & 90 (3) \\ 20 & 95 (2) & 93 (2) & 95 (2) && 91 (3) & 88 (3) & 92 (3) \\ 21 & 96 (2) & 94 (2) & 97 (2) && 94 (2) & 90 (3) & 94 (2) \\ 22 & 98 (1) & 95 (2) & 98 (1) && 96 (2) & 92 (3) & 96 (2) \\ 23 & 98 (1) & 96 (2) & 98 (1) && 97 (2) & 93 (2) & 97 (1) \\ 24 & 99 (1) & 97 (2) & 99 (1) && 98 (1) & 94 (2) & 98 (1) \\ 25 & 99 (1) & 97 (1) & 99 (1) && 99 (1) & 95 (2) & 99 (1) \\ \hline \hline \end{tabular} } \end{center} \label{table2} \end{table} \begin{figure} \begin{center} \includegraphics[width=400pt]{back62.eps} \end{center} \caption{Comparison of the average ARI for Poisson pseudo-likelihood, multinomial pseudo-likelihood and robust community detection with and without logistic regressions under 62$\%$ of background nodes. This figure appears in color in the electronic version of this article.} \label{fig1} \end{figure} \begin{figure} \begin{center} \includegraphics[width=400pt]{back50.eps} \end{center} \caption{Comparison of the average ARI for Poisson pseudo-likelihood, multinomial pseudo-likelihood and robust community detection with and without logistic regressions under 50$\%$ of background nodes. This figure appears in color in the electronic version of this article.} \label{fig2} \end{figure} \begin{figure} \begin{center} \includegraphics[width=400pt]{back38.eps} \end{center} \caption{Comparison of the average ARI for Poisson pseudo-likelihood, multinomial pseudo-likelihood and robust community detection with and without logistic regressions under 38$\%$ of background nodes. This figure appears in color in the electronic version of this article.} \label{fig3} \end{figure} The \textsc{ari} of the six methods are shown in Table \ref{table2} and Figures \ref{fig1} - \ref{fig3}. Similar to what we observed in Table \ref{table1}, the average ARIs of all methods increases as the linkage probability within community increases, or as the percentage of background nodes decreases. And the method with logistic regression outperforms their counterparts without logistic regression. The robust method with logistic regression gives the best performance in most scenarios. The Poisson pseudo-likelihood has the worst performance when the stochastic blockmodel assumption is violated in the heterogeneous background. Especially, under the case of high percentage of background nodes, the Poisson pseudo-likelihood performs poorly even when the linkage probability within community is much higher than the linkage probability between communities. The multinomial pseudo-likelihood slightly outperforms the robust method when the percentage of background nodes is high, in which case the robust method discards lots of information, while the multinomial pseudo-likelihood (or correspondingly degree corrected stochastic blockmodel) accounts for high variations on degrees. On the other hand, the robust method outperforms the multinomial pseudo-likelihood in all the other cases. In summary, the robust method has the best performance in terms of both accuracy and efficacy in almost all the setups we examined regardless the data follows stochastic blockmodels or not. In the only exception where the multinomial pseudo-likelihood method with logistic regression performs slightly better, the discrepancies between the two methods are small. Therefore, the robust community detection method is our recommended method. In real applications, the number of communities is often unknown a priori. \citet{saldana2017many} proposed a modified Bayesian information criterion (BIC) for community detection: \begin{align} -2\mathcal{L} (\V{\hat{\beta}},\V{\hat{\pi}},\hat{P} ; \V{\hat{c}},A)+\frac{(K+1)(K+2)}{2} \log \left ( \frac{n(n-1)}{2} \right ). \nonumber \end{align} \citet{daudin2008mixture} proposed another model selection criterion -- integrated classification likelihood (ICL) with a heavier penalty: \begin{align} -2\mathcal{L} (\V{\hat{\beta}},\V{\hat{\pi}},\hat{P} ; \V{\hat{c}},A)+\frac{(K+1)(K+2)}{2} \log \left ( \frac{n(n-1)}{2} \right ) + K \log(n) . \nonumber \end{align} We use simulation studies to verify the performance of BIC and ICL for our case. Since BIC and ICL are designed for the stochastic blockmodel, we compute ${L} (\V{\hat{\beta}},\V{\hat{\pi}},\hat{P} ; \V{\hat{c}},A)$ by \eqref{block} in the present studies although $\V{\hat{c}}$ is estimated by the robust method. We follow the aforementioned setup for heterogeneous background nodes and only consider the case with 50\% background nodes. For each network, we vary the assumed number of communities from $1$ to $8$, and report in the first three columns of Table \ref{model_selection} the percentages of selected numbers of communities in 50 replicates by BIC and ICL, respectively. Both BIC and ICL perfectly identifies the true community number ($K=2$). In the last set of simulation studies, we consider the model selection for networks with 5 clusters plus the background. The setup is the same as the previous study except that $n=1000$ and $\beta_0=1$. With this setup, the average size of clusters is approximately 125 as in the previous study. The results are shown in the last three column of Table \ref{model_selection}: BIC and ICL almost perfectly identifies the community number except for one replicate. \begin{table} \def~{\hphantom{0}} \caption{Proportions of the Numbers of Communities Selected by BIC and ICL} \begin{center} \begin{tabular}{ccccccc} \\ \hline \multicolumn{3}{c}{$K=2$} && \multicolumn{3}{c}{$K=5$} \\ $\hat{K}$ & BIC & ICL && $\hat{K}$ & BIC & ICL \\ \hline 1 & 0 & 0 && 1 & 0 & 0 \\ \textbf{2} & \tb{1} & \tb{1} && 2 & 0 & 0 \\ 3 & 0 & 0 && 3 & 0 & 0 \\ 4 & 0 & 0 && 4 & 0 & 0 \\ 5 & 0 & 0 && \tb{5} & \tb{0.98} & \tb{0.98} \\ 6 & 0 & 0 && 6 & 0.02 & 0.02 \\ 7 & 0 & 0 && 7 & 0 & 0 \\ 8 & 0 & 0 && 8 & 0 & 0 \\ \hline \end{tabular} \end{center} \label{model_selection} \end{table} \section{Application} With the development of improved sequencing techniques, more and more de novo mutations in candidate genes associated with neurodevelopmental or neuropshychiatric diseases are being reported. Here we focus on autism spectrum disorder and related neurological disorders. Most identified de novo mutations are rare and patients with the same clinical symptoms often carry heterogeneous mutation loci on different genes. Most probably, the pathophysiology mechanism underpinning autism involves perturbed molecular pathways. There is evidence of enrichment of de novo mutations in gene groups connected by protein-protein interactions, co-expression patterns, or pathways defined by common functions, annotations or evolutional patterns \citep{Allen13}. Our study targets at interactive groups of biomarkers that form biological pathways related to autism origination and development. Autism and related disorder data from \citet{Hormozdiari15} are employed, which reports four types of information (clinically diagnosed disease status, RNA expression levels, de novo mutations, protein-protein interactions) from three major data consortiums including BrainSpan Atlas, published autism studies, protein-protein interaction databases. There are 52,801 verified protein-protein interaction links and 192,499 mRNA pairs with Pearson's correlation coefficient between their expression levels higher than 0$\cdot$5, with an overlap of 1060 links. Together, there are 244,240 unique links from both data sources. These links involve 13,243 genes. \citet{Hormozdiari15} further gathered the de novo mutation and length information on 796 out of the 13,243 genes. In total, 796 genes with de novo mutations are employed in our analysis with 1334 mutual links between them, among which 602 genes have at least one link and 194 have none. Synonymous mutations that differ at the DNA level but produce the same protein products are excluded. The frequencies of each type of mutation in a gene in all cases are summed up as well as the total number in the controls. Two covariates are employed in estimating the probability that a gene is involved in the occurrence or progression of autism and related neurological disorders -- frequency of missense or loss of function mutations in cases, and number of mutations in controls. The choice of the covariates is based on biological beliefs on their involvement on autism development, hence decided a priori. As in the simulation study, the BIC \citep{saldana2017many} and ICL \citep{daudin2008mixture} select the same number of communities for this data -- five autism related modules plus one irrelevant background group produces. We then use three community detection methods to cluster genes into seven clusters: the robust community detection (Section 4), the pseudo-likelihood method with nodal covariates (Section 3) and the standard stochastic blockmodel fitted by the profile-likelihood \citep{bickel2009nonparametric}. We run the algorithm with a number of random initial values for the pseudo-likelihood method with nodal covariates and pick the solution with the largest value of the likelihood \eqref{block}. For a fair comparison, we use this solution as the initial value for the other two methods. Table 4 shows the estimated link probabilities within gene groups and between gene group pairs for the five autism-related gene groups (group 1-5) and the background gene group (group 6). We compare the estimates from the three methods side by side. According to the table, the standard stochastic blockmodel classifies the nodes with zero connection as a cluster. The pseudo-likelihood method with nodal covariates gives a very similar partition. The adjusted Rand index for the partitions of these two methods are 0.963. Higher values of this index indicate better agreement and 1 means perfect agreement \citep{hubert1985comparing}. The robust method gives a more different partition. The adjusted Rand index for the robust method and the standard stochastic blockmodel is 0.635, which may result from the fact that the robust method allows heterogeneous linkage rates in background genes while the two stochastic blockmodels do not. Table \ref{table_density} shows the estimated link probabilities for the three methods. Furthermore, the estimated odds ratio from the robust method for mutation numbers in cases and mutation numbers in controls are 1.4874 (P-value=0.2808) and -1.0335 (P-value=0.0173), respectively. On the contrary, the two odds ratio estimates are -0.5332 (P-value=0.0419) and -0.2464 (P-value=0.5770) from the stochastic blockmodels, which disagrees with the prior that genes with higher number of mutations in cases are more likely to be related to neurological disorders. Therefore, we employ the results from the robust method. \begin{table} \def~{\hphantom{0}} \caption{Estimated Link Probabilities between Groups} \begin{center} \begin{tabular}{cccccc}\\ \hline \multicolumn{5}{l}{Stochastic Blockmodel:} \\ \multicolumn{5}{c}{Group 1--5} &\textit{Group 6}\\ \hline 0.377 & 0.014 & 0 & 0.002 & 0.003 & 0.002 \\ 0.014 & 0.410 & 0 & 0.001 & 0.034 & 0.069 \\ 0.000 & 0.000 & 0 & 0.000 & 0.000 & 0.000 \\ 0.002 & 0.001 & 0 & 0.009 & 0.004 & 0.000 \\ 0.003 & 0.034 & 0 & 0.004 & 0.027 & 0.081 \\ 0.002 & 0.069 & 0 & 0.000 & 0.081 & 0.529 \\ \hline \multicolumn{5}{l}{Pseudo Likelihood with Covariates:} \\ \multicolumn{5}{c}{Group 1--5} &\textit{Group 6}\\ \hline 0.414 & 0.014 & 0 & 0.002 & 0.004 & 0.002 \\ 0.014 & 0.433 & 0 & 0.003 & 0.046 & 0.065 \\ 0.000 & 0.000 & 0 & 0.000 & 0.000 & 0.000 \\ 0.002 & 0.003 & 0 & 0.007 & 0.006 & 0.003 \\ 0.004 & 0.046 & 0 & 0.006 & 0.035 & 0.095 \\ 0.002 & 0.065 & 0 & 0.003 & 0.095 & 0.514 \\ \hline \multicolumn{5}{l}{Robust Community Detection:} \\ \multicolumn{5}{c}{Group 1--5} &\textit{Group 6}\\ \hline 0.414 & 0.002 & 0.000 & 0.004 & 0.012 & 0.000 \\ 0.002 & 0.618 & 0.000 & 0.006 & 0.190 & 0.028 \\ 0.000 & 0.000 & 0.000 & 0.001 & 0.000 & 0.003 \\ 0.004 & 0.006 & 0.001 & 0.009 & 0.009 & 0.027 \\ 0.012 & 0.190 & 0.000 & 0.009 & 0.087 & 0.055 \\ 0.000 & 0.028 & 0.003 & 0.027 & 0.055 & 0.104 \\ \hline \end{tabular} \end{center} \label{table_density} \end{table} The gene set enrichment analysis (GSEA) of the selected gene modules compared with the curated gene sets in the Molecular Signatures Database are listed in Table 5. P-values are calculated assuming a hypergeometric distribution for the number of overlapping genes between the selected group and the curated gene set. Given the large number of multiple comparisons, stringent P-value threshold $10^{-10}$ is employed. The five autism-related groups overlaps significantly with gene sets in essential cellular functions or abnormal conditions such as cancer, apoptosis, cell structure, circulatory system, nervous system, multicellular organismal development. Group four overlaps with gene sets related to neurological functions or disorders. Gene set ``GO NEUROGENESIS" are genes involved in generation of cells within the nervous system. Gene set ``GO REGULATION OF NERVOUS SYSTEM DEVELOPMENT" concerns processes that modulate the frequency, rate or extent of nervous system development, the origin and formation of nervous tissue. Gene sets ``GO NEURON PROJECTION" and ``GO SYNAPSE" are composed of genes involved in nerve cell prolongation and nerve fiber junction, respectively. Furthermore, our results are compared to those from the Merging Aaffected Genes into Integrated-Nnetworks method in \citet{Hormozdiari15}. The Merging Affected Genes into Integrated-Networks method was not able to detect group four. P-values from gene set enrichment analysis for the two best sets identified by their method against known neurodevelopmental diseases sets are 4$\cdot$2$\times 10^{-5}$ and 1$\cdot$0$\times 10^{-4}$, failing to reach the $10^{-10}$ threshold. \begin{table} \def~{\hphantom{0}} \caption{Gene Set Enrichment Analysis of Selected Groups. The first column is the group number identified by the proposed method; Size refers to the number of genes in the identified group, or gene set in the GSEA or their overlap. \scalebox{0.6}{ \begin{tabular}{llccccc} \\ \hline Group &Gene Set &Group &GeneSet & Overlap &Nominal & FDR \\ Number &Name &Size & Size & Size & P-value & q-value \\ \hline 1& FISCHER DREAM TARGETS & 18 &929 &16 &1$\cdot$01 $\times 10^{-25}$ &1$\cdot$07 $\times 10^{-21}$\\ 1& GOBERT OLIGODENDROCYTE DIFFERENTIATION & 18 &570 &11 &2$\cdot$86 $\times 10^{-17}$ &1$\cdot$52 $\times 10^{-13}$\\ 1& DUTERTRE ESTRADIOL RESPONSE 24HR UP & 18 &324 &9 &1$\cdot$77 $\times 10^{-15}$ &6$\cdot$29 $\times 10^{-12}$\\ 1& FISCHER G2 M CELL CYCLE & 18 &225 &8 &1$\cdot$22 $\times 10^{-14}$ &3$\cdot$26 $\times 10^{-11}$\\ 1& PUJANA BRCA2 PCC NETWORK & 18 &423 &9 &1$\cdot$97 $\times 10^{-14}$ &4$\cdot$19 $\times 10^{-11}$\\ 1& PUJANA XPRSS INT NETWORK & 18 &168 &7 &2$\cdot$37 $\times 10^{-13}$ &4$\cdot$20 $\times 10^{-10}$\\ 1& GEORGES TARGETS OF MIR192 AND MIR215 & 18 &893 &10 &2$\cdot$78 $\times 10^{-13}$ &4$\cdot$23 $\times 10^{-10}$\\ 1& NUYTTEN EZH2 TARGETS DN & 18 &1024 &10 &1$\cdot$08 $\times 10^{-12}$ &1$\cdot$43 $\times 10^{-9}$\\ 1& PUJANA CHEK2 PCC NETWORK & 18 &779 &9 &4$\cdot$68 $\times 10^{-12}$ &5$\cdot$54 $\times 10^{-9}$\\ 1& GO CHROMOSOME & 18 &880 &9 &1$\cdot$38 $\times 10^{-11}$ &1$\cdot$48 $\times 10^{-8}$\\ 2& GO CHROMOSOME ORGANIZATION & 29 &1009 &11 &1$\cdot$31 $\times 10^{-11}$ &8$\cdot$62 $\times 10^{-8}$\\ 2& GRAESSMANN APOPTOSIS BY DOXORUBICIN DN & 29 &1781 &13 &1$\cdot$62 $\times 10^{-11}$ &8$\cdot$62 $\times 10^{-8}$\\ 2& DACOSTA UV RESPONSE VIA ERCC3 DN & 29 &855 &10 &6$\cdot$86 $\times 10^{-11}$ &2$\cdot$44 $\times 10^{-7}$\\ 3& GO INTRINSIC COMPONENT OF PLASMA MEMBRANE & 518 &1649 &71 &4$\cdot$89 $\times 10^{-22}$ &5$\cdot$21 $\times 10^{-18}$\\ 3& GO RIBONUCLEOTIDE BINDING & 518 &1860 &75 &1$\cdot$37 $\times 10^{-21}$ &7$\cdot$29 $\times 10^{-18}$\\ 3& GO TRANSPORTER ACTIVITY & 518 &1276 &60 &1$\cdot$53 $\times 10^{-20}$ &5$\cdot$43 $\times 10^{-17}$\\ 3& DODD NASOPHARYNGEAL CARCINOMA UP & 518 &1821 &71 &1$\cdot$16 $\times 10^{-19}$ &3$\cdot$10 $\times 10^{-16}$\\ 3& GO ION TRANSPORT & 518 &1262 &58 &2$\cdot$04 $\times 10^{-19}$ &4$\cdot$34 $\times 10^{-16}$\\ 3& GO CELL PROJECTION & 518 &1786 &69 &6$\cdot$53 $\times 10^{-19}$ &1$\cdot$07 $\times 10^{-15}$\\ 3& GO ADENYL NUCLEOTIDE BINDING & 518 &1514 &63 &7$\cdot$01 $\times 10^{-19}$ &1$\cdot$07 $\times 10^{-15}$\\ 3& GO TRANSMEMBRANE TRANSPORTER ACTIVITY & 518 &997 &49 &1$\cdot$07 $\times 10^{-17}$ &1$\cdot$43 $\times 10^{-14}$\\ 3& GO PLASMA MEMBRANE REGION & 518 &929 &47 &1$\cdot$68 $\times 10^{-17}$ &1$\cdot$99 $\times 10^{-14}$\\ 3& GO TRANSMEMBRANE TRANSPORT & 518 &1098 &51 &2$\cdot$29 $\times 10^{-17}$ &2$\cdot$44 $\times 10^{-14}$\\ 4& GO CELL PROJECTION & 138 &1786 &39 &5$\cdot$66 $\times 10^{-23}$ &6$\cdot$03 $\times 10^{-19}$\\ 4& GO REGULATION OF CELL DEVELOPMENT & 138 &836 &28 &2$\cdot$62 $\times 10^{-21}$ &1$\cdot$39 $\times 10^{-17}$\\ 4& GO CIRCULATORY SYSTEM DEVELOPMENT & 138 &788 &27 &8$\cdot$15 $\times 10^{-21}$ &2$\cdot$89 $\times 10^{-17}$\\ 4& GO NEUROGENESIS & 138 &1402 &32 &2$\cdot$29 $\times 10^{-19}$ &6$\cdot$11 $\times 10^{-16}$\\ 4& GO REGULATION OF MULTICELLULAR ORGANISMAL DEVELOPMENT& 138 &1672 &34 &4$\cdot$94 $\times 10^{-19}$ &1$\cdot$05 $\times 10^{-15}$\\ 4& GO VASCULATURE DEVELOPMENT & 138 &469 &21 &1$\cdot$12 $\times 10^{-18}$ &1$\cdot$98 $\times 10^{-15}$\\ 4& GO REGULATION OF CELL DIFFERENTIATION & 138 &1492 &32 &1$\cdot$40 $\times 10^{-18}$ &2$\cdot$12 $\times 10^{-15}$\\ 4& GO REGULATION OF NERVOUS SYSTEM DEVELOPMENT & 138 &750 &24 &6$\cdot$78 $\times 10^{-18}$ &9$\cdot$03 $\times 10^{-15}$\\ 4& GO NEURON PROJECTION & 138 &942 &26 &8$\cdot$96 $\times 10^{-18}$ &1$\cdot$06 $\times 10^{-14}$\\ 4& GO SYNAPSE & 138 &754 &23 &9$\cdot$88 $\times 10^{-17}$ &1$\cdot$04 $\times 10^{-13}$\\ 5& GO CHROMOSOME ORGANIZATION & 62 &1009 &20 &2$\cdot$17 $\times 10^{-18}$ &2$\cdot$31 $\times 10^{-14}$\\ 5& GO CHROMATIN MODIFICATION & 62 &539 &15 &5$\cdot$06 $\times 10^{-16}$ &2$\cdot$69 $\times 10^{-12}$\\ 5& GO CHROMATIN ORGANIZATION & 62 &663 &15 &1$\cdot$04 $\times 10^{-14}$ &3$\cdot$69 $\times 10^{-11}$\\ 5& PILON KLF1 TARGETS DN & 62 &1972 &19 &7$\cdot$01 $\times 10^{-12}$ &1$\cdot$87 $\times 10^{-8}$\\ 5& GO CELL CYCLE & 62 &1316 &16 &1$\cdot$48 $\times 10^{-11}$ &3$\cdot$15 $\times 10^{-8}$\\ 5& GO COVALENT CHROMATIN MODIFICATION & 62 &345 &10 &3$\cdot$80 $\times 10^{-11}$ &6$\cdot$74 $\times 10^{-8}$\\ 5& DACOSTA UV RESPONSE VIA ERCC3 DN & 62 &855 &13 &1$\cdot$05 $\times 10^{-10}$ &1$\cdot$50 $\times 10^{-7}$\\ 5& GO NUCLEAR CHROMOSOME & 62 &523 &11 &1$\cdot$13 $\times 10^{-10}$ &1$\cdot$50 $\times 10^{-7}$\\ 5& GO CELL CYCLE PROCESS & 62 &1081 &14 &1$\cdot$49 $\times 10^{-10}$ &1$\cdot$59 $\times 10^{-7}$\\ 5& GO CHROMOSOME & 62 &880 &13 &1$\cdot$49 $\times 10^{-10}$ &1$\cdot$59 $\times 10^{-7}$\\ \hline \vspace{10mm} \footnote{} \end{tabular} } \label{table5} \vspace{1ex} \end{table} \section{Discussion} A major improvement of the proposed method over previous ones is the integration of network topology and auxiliary node information. The proposed analysis pools rich epigenomic information from heterogeneous online resources, such as expression/co-expression profiles from BrainSpan Atlas, de novo mutations in cases and controls from autism or related neurological disorder studies, protein-protein interactions in protein databases. Although these three types of information are measured on different cohorts, they describe distinct aspects of the candidate genes. They can be linked by unique genes, which are the unit of our analysis. In the era of big data, statistical methods need not be restricted to one data source or single clinical trial. Instead, methods should incorporate information from multiple related resources. The estimation method is non-standard. For a fixed initial label assignment, we use the expectation-maximization algorithm to fit a pseudo-likelihood. Then the label assignment is updated according to the expectation-maximization results, and used as initial label assignment in the next iteration. Taking advantage of the pseudo-likelihood, we are able to allow heterogeneous linkage probabilities in the background. The consistency of the label assignments is proved for a simple version of this complicated procedure -- one-step expectation-maximization. Further research is needed to understand the statistical properties of the algorithm in more complex settings. Furthermore, the pseudo-likelihood approach for the robust setup is very unique. Usually, a regularization term is used to penalize the log-likelihood when the MLE is degenerate, for instance, the roughness penalty in smoothing splines and $L_1$ penalty in lasso for ``large-$p$-small-$n$'' problems. Our approach for the robust setup does not follow exactly the ``loss+penalty'' framework. Due to the special two-dimensional structure (adjacency matrix) of network data, the pseudo-likelihood with the noisy data being removed can give accurate community labels while the MLE of the true likelihood is degenerate. In network analysis the pseudo-likelihood approach not only provides computational efficiency but can also serve as a convenient likelihood formulation that can discard some columns of variables (noise/background) while still remain a valid likelihood because all rows of observations (nodes/genes) are kept. The full likelihood will fail in this case. Researchers have suggested that a node may belong to multiple communities in a biological networks. For example, \cite{airoldi2008mixed} proposed a mixed membership stochastic blockmodels and applied this model into a network of protein-protein interactions. We will explore the extension of the logistic regression augmented model to overlapping community detection in our future work.

a9a5ec94e0b7bfbd44c1ea48cbb05f0788e2c9f3

\section{Introduction} Classification is the process of finding and analyzing underlying group structure in heterogenous data. This problem can be framed as the search for class labels of unlabelled observations. In general, some (non-trivial) proportion of observations have known labels. A special case of classification, known as clustering, occurs when none of the observations have known labels. One common approach for clustering is mixture model-based clustering, which makes use of a finite mixture model. In general, a $G$-component finite mixture model assumes that a multivariate random variable $\mathbf{X}$ has density \begin{equation}\label{eqn:lll} f(\mathbf{x}~|~{\boldsymbol{\vartheta}})=\sum_{g=1}^G\pi_g f_g(\mathbf{x}~|~{\boldsymbol{\theta}}_g), \end{equation} where ${\boldsymbol{\vartheta}}=\left(\pi_1,\pi_2,\ldots,\pi_G,{\boldsymbol{\theta}}_1,{\boldsymbol{\theta}}_2,\ldots,{\boldsymbol{\theta}}_G\right)$, $f_g(\cdot)$ is the $g$th component density, and $\pi_g>0$ is the $g$th mixing proportion such that $\sum_{i=1}^G\pi_g=1$. Note that the notation used in \eqref{eqn:lll} corresponds to the multivariate case and, save for Appendix~A, $\mathscr{X}$ will hereafter represent a matrix variate random variable with $\mathbf{X}$ denoting its realization. \cite{mcnicholas16a} traces the relationship between mixture models and clustering to \cite{tiedeman55}, who uses a component of a mixture model to define a cluster. The mixture model was first used for clustering by \cite{wolfe65} who considered a Gaussian mixture model. Other early uses of Gaussian mixture models for clustering can be found in \cite{baum70} and \cite{scott71}. Although the Gaussian mixture model is attractive due to its mathematical properties, it can be problematic when dealing with outliers and asymmetry in clusters and thus there has been an interest in non-Gaussian mixtures in the multivariate setting. Some examples of mixtures of symmetric distributions that parameterize tail weight include the $t$ distribution \citep{peel00,andrews11a,andrews12,lin14} and the power exponential distribution \citep{dang15}. There has also been work in the area of skewed distributions such as the skew-$t$ distribution, \citep{lin10,vrbik12,vrbik14,lee14,murray14b,murray14a}, the normal-inverse Gaussian (NIG) distribution \citep{karlis09}, the shifted asymmetric Laplace (SAL) distribution \citep{morris13b,franczak14}, the variance-gamma distribution \citep{smcnicholas17}, the generalized hyperbolic distribution \citep{browne15}, the hidden truncation hyperbolic distribution \citep{murray17b}, the joint generalized hyperbolic distribution \citep{tang18}, and the coalesced generalized hyperbolic distribution \citep{tortora19}. There has also been an increased interest in model-based clustering of matrix variate, or three-way, data such as multivariate longitudinal data and images. Examples include the work of \cite{viroli11a} and \cite{Anderlucci15}, who consider mixtures of matrix variate normal distributions for clustering. \cite{viroli11b} further considers model-based clustering with matrix variate normal distributions in the Bayesian paradigm. More recently, \cite{gallaugher18a} investigate mixtures of four skewed matrix distributions--- namely, the matrix variate skew-$t$, generalized hyperbolic, variance-gamma, and normal inverse Gaussian (NIG) distributions \citep[see][for details about these matrix variate distributions]{gallaugher17b}---and consider classification of greyscale Arabic numerals. \cite{melnykov18,melnykov19} consider modelling skewness by means of transformations. The main problem with all of the aforementioned methods, for both the multivariate and matrix variate cases, arises when the dimensionality of the data increases. Although the problem of dealing with high-dimensional data has been thoroughly addressed in the case of multivariate data, there is relative paucity of work for matrix variate data. In the matrix variate case, matrix variate bilinear probabilistic principal component analysis was developed by \cite{zhao12}. More recently, \cite{gallaugher18b} considered the closely-related mixture of matrix variate bilinear factor analyzers (MMVBFA) model for clustering. The MMVBFA model can be viewed as a matrix variate generalization of the mixture of factor analyzers model \citep{ghahramani97} in the multivariate case. Although these methods allow for simultaneous dimension reduction and clustering, they both assume matrix variate normality which is not sensible if cluster skewness or heavy tails are present. Herein we present an extension of the MMVBFA model to skewed distributions; specifically, the matrix variate skew-$t$, generalized hyperbolic, variance-gamma, and NIG distributions. \section{Background} \subsection{Generalized Inverse Gaussian Distribution} The generalized inverse Gaussian distribution has two different parameterizations, both of which will be utilized herein. A random variable $Y$ has a generalized inverse Gaussian (GIG) distribution parameterized by $a, b$ and $\lambda$, and denoted $\text{GIG}(a,b,\lambda)$, if its probability density function can be written as $$ f(y|a, b, \lambda)=\frac{\left({a}/{b}\right)^{\frac{\lambda}{2}}y^{\lambda-1}}{2K_{\lambda}(\sqrt{ab})}\exp\left\{-\frac{ay+{b}/{y}}{2}\right\}, $$ for $y>0$, $a,b\in\mathbb{R}^+$ and $\lambda\in\mathbb{R}$, where $$ K_{\lambda}(u)=\frac{1}{2}\int_{0}^{\infty}y^{\lambda-1}\exp\left\{-\frac{u}{2}\left(y+\frac{1}{y}\right)\right\}dy $$ is the modified Bessel function of the third kind with index $\lambda$. Expectations of some functions of a GIG random variable have a mathematically tractable form, e.g.: \begin{equation} \mathbb{E}(Y)=\sqrt{\frac{b}{a}}\frac{K_{\lambda+1}(\sqrt{ab})}{K_{\lambda}(\sqrt{ab})}, \label{eq:ai} \end{equation} \begin{equation} \mathbb{E}\left({1}/{Y}\right)=\sqrt{\frac{a}{b}}\frac{K_{\lambda+1}(\sqrt{ab})}{K_{\lambda}(\sqrt{ab})}-\frac{2\lambda}{b}, \label{eq:bi} \end{equation} \begin{equation} \mathbb{E}(\log Y)=\log\left(\sqrt{\frac{b}{a}}\right)+\frac{1}{K_{\lambda}(\sqrt{ab})}\frac{\partial}{\partial \lambda}K_{\lambda}(\sqrt{ab}). \label{eq:ci} \end{equation} Although this parameterization of the GIG distribution will be useful for parameter estimation, the alternative parameterization given by \begin{equation} g(y|\omega,\eta,\lambda)= \frac{\left({w}/{\eta}\right)^{\lambda-1}}{2\eta K_{\lambda}(\omega)}\exp\left\{-\frac{\omega}{2}\left(\frac{w}{\eta}+\frac{\eta}{w}\right)\right\}, \label{eq:I} \end{equation} where $\omega=\sqrt{ab}$ and $\eta=\sqrt{a/b}$, is used when deriving the generalized hyperbolic distribution \citep[see][]{browne15}. For notational clarity, we will denote the parameterization given in \eqref{eq:I} by $\text{I}(\omega,\eta,\lambda)$. \subsection{Matrix Variate Distributions} As in the multivariate case, the most mathematically tractable matrix variate distribution is the matrix variate normal. An $n\times p$ random matrix $\mathscr{X}$ follows an $n\times p$ matrix variate normal distribution with $n\times p$ location matrix $\mathbf{M}$ and scale matrices $\mathbf\Sigma$ and $\mathbf\Psi$, of dimensions $n\times n$ and $p\times p$, respectively, denoted $\mathcal{N}_{n\times p}(\mathbf{M}, \mathbf\Sigma, \mathbf\Psi)$, if the density of $\mathscr{X}$ is \begin{equation} f(\mathbf{X}~|~\mathbf{M}, \mathbf\Sigma, \mathbf\Psi)=\frac{1}{(2\pi)^{\frac{np}{2}}|\mathbf\Sigma|^{\frac{p}{2}}|\mathbf\Psi|^{\frac{n}{2}}}\exp\left\{-\frac{1}{2}\,\mbox{tr}\left(\mathbf\Sigma^{-1}(\mathbf{X}-\mathbf{M})\mathbf\Psi^{-1}(\mathbf{X}-\mathbf{M})'\right)\right\}. \end{equation} The matrix variate normal distribution is related to the multivariate normal distribution, as discussed in \cite{harrar08}, via $\mathscr{X}\sim \mathcal{N}_{n\times p}(\mathbf{M},\mathbf\Sigma,\mathbf\Psi) \iff \text{vec}(\mathscr{X})\sim \mathcal{N}_{np}(\text{vec}(\mathbf{M}),\mathbf\Psi\otimes \mathbf\Sigma),$ where $\mathcal{N}_{np}(\cdot)$ is the multivariate normal density with dimension $np$, $\text{vec}(\mathbf{M})$ is the vectorization of $\mathbf{M}$, and $\otimes$ denotes the Kronecker product. Although the matrix variate normal distribution is popular, there are other well-known examples of matrix variate distributions. For example, the Wishart distribution \citep{Wishart} is the distribution of the sample covariance matrix in the multivariate normal case. There are also a few formulations of a matrix variate skew normal distribution \citep{chen2005,dominguez2007,harrar08}. More recently, \cite{gallaugher17a,gallaugher17b} derived a total of four skewed matrix variate distributions using a variance-mean matrix variate mixture approach. This assumes that a random matrix $\mathscr{X}$ can be written as \begin{equation} \mathscr{X}=\mathbf{M}+W\mathbf{A}+\sqrt{W}\mathscr{V}, \label{eq:mvmix} \end{equation} where $\mathbf{M}$ and $\mathbf{A}$ are $n\times p$ matrices representing the location and skewness, respectively, $\mathscr{V} \sim \mathcal{N}_{n \times p}\left(\bf{0}, \mathbf\Sigma , \mathbf\Psi \right)$, and $W>0$ is a random variable with density $h(w|{\boldsymbol{\theta}})$. \cite{gallaugher17a}, show that the matrix variate skew-$t$ distribution, with $\nu$ degrees of freedom, arises from \eqref{eq:mvmix} with $W^{\text{ST}}\sim \text{IGamma}(\nu/2,\nu/2)$, where $\text{IGamma}(\cdot)$ denotes the inverse-gamma distribution with density $$ f(y~|~a,b)=\frac{b^a}{\Gamma(a)}y^{-a-1}\exp\left\{-\frac{b}{y}\right\}, $$ for $y>0$ and $a,b \in \mathbb{R}^+$. The resulting density of $\mathscr{X}$ is \begin{align*} f_{\text{\tiny MVST}}(\mathbf{X}~|~{\boldsymbol{\vartheta}})=&\frac{2\left(\frac{\nu}{2}\right)^{\frac{\nu}{2}}\exp\left\{\,\mbox{tr}(\mathbf\Sigma^{-1}(\mathbf{X}-\mathbf{M})\mathbf\Psi^{-1}\mathbf{A}') \right\} }{(2\pi)^{\frac{np}{2}}| \mathbf\Sigma |^{\frac{p}{2}} |\mathbf\Psi |^{\frac{n}{2}}\Gamma(\frac{\nu}{2})} \left(\frac{\delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)+\nu}{\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)}\right)^{-\frac{\nu+np}{4}} \\ & \qquad\qquad\qquad\qquad\times K_{-\frac{\nu+np}{2}}\left(\sqrt{\left[\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)\right]\left[\delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)+\nu\right]}\right), \end{align*} where $ \delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)=\,\mbox{tr}(\mathbf\Sigma^{-1}(\mathbf{X}-\mathbf{M})\mathbf\Psi^{-1}(\mathbf{X}-\mathbf{M})')$, $\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)=\,\mbox{tr}(\mathbf\Sigma^{-1}\mathbf{A}\mathbf\Psi^{-1}\mathbf{A}') $ and $\nu>0$. For notational clarity, this distribution will be denoted by $\text{MVST}(\mathbf{M},\mathbf{A},\mathbf\Sigma,\mathbf\Psi,\nu)$. In \cite{gallaugher17b}, one of the distributions considered is a matrix variate generalized hyperbolic distribution. This again arises from \eqref{eq:mvmix} with $W^{\text{GH}}\sim\text{I}(\omega,1,\lambda)$. This distribution will be denoted by $\text{MVGH}(\mathbf{M},\mathbf{A},\mathbf\Sigma,\mathbf\Psi,\lambda,\omega)$, and the density is \begin{align*} f_{\text{\tiny MVGH}}(\mathbf{X}|{\boldsymbol{\vartheta}})=&\frac{\exp\left\{\,\mbox{tr}(\mathbf\Sigma^{-1}(\mathbf{X}-\mathbf{M})\mathbf\Psi^{-1}\mathbf{A}') \right\}}{(2\pi)^{\frac{np}{2}}| \mathbf\Sigma |^{\frac{p}{2}} |\mathbf\Psi |^{\frac{n}{2}}K_{\lambda}(\omega)} \left(\frac{\delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)+\omega}{\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)+\omega}\right)^{\frac{\left(\lambda-\frac{np}{2}\right)}{2}} \\ & \times K_{\left(\lambda-{np}/{2}\right)}\left(\sqrt{\left[\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)+\omega\right]\left[\delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)+\omega\right]}\right), \end{align*} where $\lambda\in \mathbb{R}$ and $\omega\in\mathbb{R}^+$. The matrix variate variance-gamma distribution, also derived in \cite{gallaugher17b} and denoted $\text{MVVG}(\mathbf{M},\mathbf{A},\mathbf\Sigma,\mathbf\Psi,\gamma)$, arises from \eqref{eq:mvmix} with $W^{\text{VG}}\sim\text{gamma}(\gamma,\gamma)$, where $\text{gamma}(\cdot)$ denotes the gamma distribution with density $$ f(y~|~a,b)=\frac{b^a}{\Gamma(a)}y^{a-1}\exp\left\{-by\right\}, $$ for $y>0$ and $a,b\in\mathbb{R}^+$ The density of the random matrix $\mathscr{X}$ with this distribution is \begin{align*} f_{\text{\tiny MVVG}}(\mathbf{X}|{\boldsymbol{\vartheta}})=&\frac{2\gamma^{\gamma}\exp\left\{\,\mbox{tr}(\mathbf\Sigma^{-1}(\mathbf{X}-\mathbf{M})\mathbf\Psi^{-1}\mathbf{A}') \right\}}{(2\pi)^{\frac{np}{2}}| \mathbf\Sigma |^{\frac{p}{2}} |\mathbf\Psi |^{\frac{n}{2}}\Gamma(\gamma)} \left(\frac{\delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)}{\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)+2\gamma}\right)^{\frac{\left(\gamma-{np}/{2}\right)}{2}} \\ &\times K_{\left(\gamma-\frac{np}{2}\right)}\left(\sqrt{\left[\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)+2\gamma\right]\left[\delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)\right]}\right), \end{align*} where $\gamma>0$. Finally, the matrix variate NIG distribution arises when $W^{\text{NIG}}\sim\text{IG}(1,\kappa)$, where $\text{IG}(\cdot)$ denotes the inverse-Gaussian distribution with density $$ f(y~|~\delta,\kappa)=\frac{\delta}{\sqrt{2\pi}}\exp\{\delta\kappa\}y^{-\frac{3}{2}}\exp\left\{-\frac{1}{2}\left(\frac{\delta^2}{y}+\kappa^2y\right)\right\}, $$ for $y>0$, $\delta,\kappa\in\mathbb{R}^+$. The density of $\mathscr{X}$ is \begin{align*} f_{\text{\tiny MVNIG}}(\mathbf{X}|{\boldsymbol{\vartheta}})&=\frac{2\exp\left\{\,\mbox{tr}(\mathbf\Sigma^{-1}(\mathbf{X}-\mathbf{M})\mathbf\Psi^{-1}\mathbf{A}')+\kappa\right\} }{(2\pi)^{\frac{np+1}{2}}| \mathbf\Sigma |^{\frac{p}{2}} |\mathbf\Psi |^{\frac{n}{2}}}\left(\frac{\delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)+1}{\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)+\kappa^2}\right)^{-{\left(1+np\right)}/{4}}\\ &\times K_{-{(1+np)}/{2}}\left(\sqrt{\left[\rho(\mathbf{A},\mathbf\Sigma,\mathbf\Psi)+\kappa^2\right]\left[\delta(\mathbf{X};\mathbf{M},\mathbf\Sigma,\mathbf\Psi)+1\right]}\right), \end{align*} where $\kappa>0$. This distribution is denoted by $\text{MVNIG}(\mathbf{M},\mathbf{A},\mathbf\Sigma,\mathbf\Psi,\kappa)$. \subsection{Matrix Variate Factor Analysis} Readers who may benefit from the context provided by the mixture of factor analyzers model should consult Appendix A. \cite{xie08} and \cite{yu08} consider a matrix variate extension of probabilistic principal components analysis (PPCA), which assumes an $n\times p$ random matrix $\mathscr{X}$ can be written \begin{equation} \mathscr{X}=\mathbf{M}+\mathbf\Lambda\mathscr{U}\mathbf\Delta'+\mathscr{E}, \label{eq:MVPPCA} \end{equation} where $\mathbf{M}$ is an $n\times p$ location matrix, $\mathbf\Lambda$ is an $n\times q$ matrix of column factor loadings, $\mathbf\Delta$ is a $p\times r$ matrix of row factor loadings, $\mathscr{U}\sim \mathcal{N}_{q\times r}({\bf 0},\mathbf{I}_q,\mathbf{I}_r)$, and $\mathscr{E}\sim\mathcal{N}_{n\times p}({\bf 0},\sigma^2\mathbf{I}_n,\sigma^2\mathbf{I}_p)$. It is assumed that $\mathscr{U}$ and $\mathscr{E}$ are independent of each other. The main disadvantage of this model is that, in general, $\mathscr{X}$ does not follow a matrix variate normal distribution. \cite{zhao12} present bilinear probabilistic principal component analysis (BPPCA), which extends \eqref{eq:MVPPCA} by adding two projected error terms. The resulting model assumes $\mathscr{X}$ can be written $\mathscr{X}=\mathbf{M}+\mathbf\Lambda\mathscr{U}\mathbf\Delta'+\mathbf\Lambda\mathscr{E}^B+\mathscr{E}^A\mathbf\Delta'+\mathscr{E},$ where $\mathscr{U}$ and $\mathscr{E}$ are defined as in \eqref{eq:MVPPCA}, $\mathscr{E}^B\sim \mathcal{N}_{q\times p}({\bf 0},\mathbf{I}_q,\sigma_B\mathbf{I}_p)$, and $\mathscr{E}^A\sim \mathcal{N}_{n\times r}({\bf 0}, \sigma_A\mathbf{I}_n,\mathbf{I}_r)$. In this model, it is assumed that $\mathscr{U}, \mathscr{E}^B, \mathscr{E}^A$, and $\mathscr{E}$ are all independent of each other. \cite{gallaugher18b} further extend this to matrix variate factor analysis and consider a mixture of matrix variate bilinear factor analyzers (MMVBFA) model. For MMVBFA, \cite{gallaugher18b} generalize BPPCA by removing the isotropic constraints so that $\mathscr{E}^B\sim \mathcal{N}_{q\times p}({\bf 0},\mathbf{I}_q,\mathbf\Psi)$, $\mathscr{E}^A\sim \mathcal{N}_{n\times r}({\bf 0},\mathbf\Sigma,\mathbf{I}_r)$, and $\mathscr{E}\sim\mathcal{N}_{n\times p}({\bf 0},\mathbf\Sigma,\mathbf\Psi)$, where $\mathbf\Sigma=\,\mbox{diag}\{\sigma_1,\sigma_2,\ldots,\sigma_n\}$ with $\sigma_i>0$, and $\mathbf\Psi=\,\mbox{diag}\{\psi_1,\psi_2, \ldots, \psi_p\}$ with $\psi_i>0$. With these slight modifications, it can be shown that $\mathscr{X}\sim\mathcal{N}_{n\times p}(\mathbf{M},\mathbf\Lambda\loada'+\mathbf\Sigma,\mathbf\Delta\loadb'+\mathbf\Psi)$, similarly to its multivariate counterpart (Appendix~A). It is important to note that the term ``column factors" refers to reduction in the dimension of the columns, which is equivalent to the number of rows, and not a reduction in the number of columns. Likewise, the term ``row factors" refers to the reduction in the dimension of the rows (number of columns). As discussed by \cite{zhao12}, the interpretation of the terms $\mathscr{E}^B$ and $\mathscr{E}^A$ are the row and column noise, respectively, whereas the term $\mathscr{E}$ is the common noise. \section{Mixture of Skewed Matrix Variate Bilinear Factor Analyzers} \subsection{Model Specification} We now consider a mixture of skewed bilinear factor analyzers according to one of the four skewed distributions discussed previously. Each random matrix $\mathscr{X}_i$ from a random sample distributed according to one of the four distributions can be written as $$ \mathscr{X}_i=\mathbf{M}_g+W_{ig}\mathbf{A}_g+\mathscr{V}_{ig} $$ with probability $\pi_g$ for $g\in\{1,2,\ldots,G\}$, $\pi_g>0$, $\sum_{i=1}^G\pi_g=1$, where $\mathbf{M}_g$ is the location of the $g$th component, $\mathbf{A}_g$ is the skewness, and $W_{ig}$ is a random variable with density $h(w_{ig}|{\boldsymbol{\theta}}_g)$. The distribution of the random variable $W_{ig}$---and so the density $h(w_{ig}|{\boldsymbol{\theta}}_g)$---will change depending on the distribution of $\mathscr{X}_i$, i.e., skew-$t$, generalized hyperbolic, variance-gamma, or NIG. Assume also that $\mathscr{V}_{ig}$ can be written as $$ \mathscr{V}_{ig}=\mathbf\Lambda_g\mathscr{U}_{ig}\mathbf\Delta_g'+\mathbf\Lambda_g\mathscr{E}_{ig}^B+\mathscr{E}_{ig}^A\mathbf\Delta_g'+\mathscr{E}_{ig}, $$ where $\mathbf\Lambda_g$ is a $n\times q$ matrix of column factor loadings, $\mathbf\Delta_g$ is a $p\times r$ matrix of row factor loadings, and \begin{align*} \mathscr{U}_{ig}|w_{ig}&\sim\mathcal{N}_{q\times r}({\bf 0},w_{ig}\mathbf{I}_q,\mathbf{I}_p), &\mathscr{E}_{ig}^B|w_{ig}\sim\mathcal{N}_{q\times p}({\bf 0},w_{ig}\mathbf{I}_q,\mathbf\Psi_g),\\ \mathscr{E}_{ig}^A|w_{ig}&\sim\mathcal{N}_{n\times r}({\bf 0},w_{ig}\mathbf\Sigma_g,\mathbf{I}_r), &\mathscr{E}_{ig}|w_{ig}\sim\mathcal{N}_{n\times p}({\bf 0},w_{ig}\mathbf\Sigma_g,\mathbf\Psi_g). \end{align*} Note that $\mathscr{U}_{ig},\mathscr{E}_{ig}^B,\mathscr{E}_{ig}^A$, and $\mathscr{E}_{ig}$ are all independently distributed and independent of each other. To facilitate clustering, introduce the indicator $z_{ig}$, where $z_{ig}=1$ if observation $i$ belongs to group $g$, and $z_{ig}=0$ otherwise. Then, it can be shown that $$ \mathscr{X}_i~|~z_{ig}=1\sim\text{D}_{n\times p}(\mathbf{M}_g,\mathbf{A}_g,\mathbf\Sigma_g+\mathbf\Lambda_g\mathbf\Lambda_g',\mathbf\Psi_g+\mathbf\Delta_g\mathbf\Delta_g',{\boldsymbol{\theta}}_g), $$ where $\text{D}$ is the distribution in question, and ${\boldsymbol{\theta}}_g$ is the set of parameters related to the distribution of $W_{ig}$. As in the matrix variate normal case, this model has a two stage interpretation given by \begin{equation*}\begin{split} \mathscr{X}_i&=\mathbf{M}_g+W_{ig}\mathbf{A}+\mathbf\Lambda_g\mathscr{Y}_{ig}^B+\mathscr{R}_{ig}^B,\\ \mathscr{Y}_{ig}^B&=\mathscr{U}_{ig}\mathbf\Delta_g'+\mathscr{E}_{ig}^B,\\ \mathscr{R}_{ig}^B&=\mathscr{E}_{ig}^A\mathbf\Delta_{g}'+\mathscr{E}_{ig}, \end{split}\end{equation*} and \begin{equation*}\begin{split} \mathscr{X}_i&=\mathbf{M}_g+W_{ig}\mathbf{A}+\mathscr{Y}_{ig}^A\mathbf\Delta_g'+\mathscr{R}_{ig}^A,\\ \mathscr{Y}_{ig}^A&=\mathbf\Lambda_g\mathscr{U}_{ig}+\mathscr{E}_{ig}^A,\\ \mathscr{R}_{ig}^A&=\mathbf\Lambda_g\mathscr{E}_{ig}^B+\mathscr{E}_{ig}, \end{split}\end{equation*} which will be useful for parameter estimation. \subsection{Parameter Estimation} Suppose we observe the $N$ $n\times p$ matrices $\mathbf{X}_1,\ldots,\mathbf{X}_N$ distributed according to one of the four distributions. We assume that these data are incomplete and employ an alternating expectation conditional maximization (AECM) algorithm \citep{meng97}. This algorithm is now described after initialization. \paragraph{AECM 1st Stage} The complete-data in the first stage consists of the observed data $\mathbf{X}_i$, the latent variables ${\bf W}_i=(W_{i1},\ldots,W_{iG})'$, and the unknown group labels ${\bf z}_{i}=(z_{i1},\ldots,z_{iG})'$ for $i=1,2,\ldots,N$. In this case, the complete-data log-likelihood is \begin{equation*}\begin{split} \ell_{\text{C1}}=C+&\sum_{i=1}^N\sum_{g=1}^Gz_{ig}\bigg[\log \pi_g+\log h(w_{ig}|{\boldsymbol{\theta}}_g)-\frac{1}{2}\,\mbox{tr}\bigg\{\frac{1}{W_{ig}}(\mathbf\Sigma_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)(\mathbf\Psi_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)'\\ &-(\mathbf\Sigma_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)(\mathbf\Psi_g^*)^{-1}\mathbf{A}_g'-(\mathbf\Sigma_g^*)^{-1}\mathbf{A}_g(\mathbf\Psi_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)'+W_{ig}(\mathbf\Sigma_g^*)^{-1}\mathbf{A}_g(\mathbf\Psi_g^*)^{-1}\mathbf{A}_g'\bigg\}\bigg], \end{split}\end{equation*} where $\mathbf\Sigma_g^{*}=\mathbf\Sigma_g+\mathbf\Lambda_g\mathbf\Lambda_g'$, $\mathbf\Psi_g^*=\mathbf\Psi_g+\mathbf\Delta_g\mathbf\Delta_g'$ and $C$ is constant with respect to the parameters. In the E-step, we calculate the following conditional expectations: \begin{equation*}\begin{split} \hat{z}_{ig}&=\frac{\pi_g f(\mathbf{X}_i~|~\hat{{\boldsymbol{\vartheta}}}_g)} {\sum_{h=1}^G\pi_h f(\mathbf{X}_i~|~\hat{{\boldsymbol{\vartheta}}}_h)},\qquad\qquad a_{ig}=\mathbb{E}(W_{ig}~|~\mathbf{X}_i,z_{ig}=1,\hat{{\boldsymbol{\vartheta}}}_g),\\ b_{ig}&=\mathbb{E}\left(\frac{1}{W_{ig}}~\bigg|~\mathbf{X}_i,z_{ig}=1,\hat{{\boldsymbol{\vartheta}}}_g\right),\qquad c_{ig}=\mathbb{E}(\log W_{ig}~|~\mathbf{X}_i,z_{ig}=1,\hat{{\boldsymbol{\vartheta}}}_g).\\ \end{split}\end{equation*} As usual, all expectations are conditional on current parameter estimates; however, to avoid cluttered notation, we do not use iteration-specific notation. Although these expectations are dependent on the distribution in question, it can be shown that \begin{align*} W_{ig}^{\text{ST}}~|~\mathbf{X}_i, z_{ig}=1&\sim \text{GIG}\left(\rho(\mathbf{A}_g,\mathbf\Sigma^*_g,\mathbf\Psi^*_g),\delta(\mathbf{X};\mathbf{M}_g,\mathbf\Sigma^*_g,\mathbf\Psi^*_g)+\nu_g,-(\nu_g+np)/2\right),\\ W_{ig}^{\text{GH}}~|~\mathbf{X}_i, z_{ig}=1&\sim \text{GIG}\left(\rho(\mathbf{A}_g,\mathbf\Sigma^*_g,\mathbf\Psi^*_g)+\omega_g,\delta(\mathbf{X};\mathbf{M}_g,\mathbf\Sigma^*_g,\mathbf\Psi^*_g)+\omega_g,\lambda_g-{np}/{2}\right),\\ W_{ig}^{\text{VG}}~|~\mathbf{X}_i, z_{ig}=1&\sim \text{GIG}\left(\rho(\mathbf{A}_g,\mathbf\Sigma^*_g,\mathbf\Psi^*_g)+2\gamma_g,\delta(\mathbf{X};\mathbf{M}_g,\mathbf\Sigma^*_g,\mathbf\Psi^*_g),\gamma_g-{np}/{2}\right),\\ W_{ig}^{\text{NIG}}~|~\mathbf{X}_i, z_{ig}=1&\sim \text{GIG}\left(\rho(\mathbf{A}_g,\mathbf\Sigma^*_g,\mathbf\Psi^*_g)+\kappa_g^2,\delta(\mathbf{X};\mathbf{M}_g,\mathbf\Sigma^*_g,\mathbf\Psi^*_g)+1,-{(1+np)}/{2}\right). \end{align*} Therefore, the exact updates are obtained by using the expectations given in \eqref{eq:ai}--\eqref{eq:ci} for appropriate values of $\lambda, a$, and $b$. In the M-step, we update $\hat{\pi}_g,\hat\mathbf{M}_g,\hat\mathbf{A}_g$, and $\hat{\boldsymbol{\theta}}_g$ for $g=1,\ldots,G$. We have: $$ \hat{\pi}_g=\frac{N_g}{N}, \qquad \hat{\mathbf{M}}_g=\frac{\sum_{i=1}^N\hat{z}_{ig}\left(\overline{a}_gb_{ig}-1\right)\mathbf{X}_i}{\sum_{i=1}^N\hat{z}_{ig}\overline{a}_gb_{ig}-N_g}, \qquad \hat{\mathbf{A}}=\frac{\sum_{i=1}^N\hat{z}_{ig}\left(\overline{b}_g-b_{ig}\right)\mathbf{X}_i}{\sum_{i=1}^N\hat{z}_{ig}\overline{a}_gb_{ig}-N_g}, $$ where $$ N_g=\sum_{i=1}^N\hat{z}_{ig},\qquad\overline{a}_g=\frac{1}{N_g}{\sum_{i=1}^N\hat{z}_{ig}a_{ig}},\qquad \overline{b}_g=\frac{1}{N_g}{\sum_{i=1}^N\hat{z}_{ig}b_{ig}}. $$ The update for ${\boldsymbol{\theta}}_g$ is dependent on the distribution and will be identical to one of those given in \cite{gallaugher18a}. \paragraph{AECM Stage 2} In the second stage, the complete-data consists of the observed data $\mathbf{X}_i$, the latent variables ${\bf W}_i$, the unknown group labels ${\bf z}_{i}$, and the latent matrices ${\boldsymbol \mathscr{Y}}^B_i=(\mathscr{Y}^B_{i1},\ldots,\mathscr{Y}^B_{iG})$ for $i=1,\ldots,N$. The complete-data log-likelihood at this stage is \begin{align*} \ell_{\text{C}2}&=C+\sum_{i=1}^N\sum_{g=1}^Gz_{ig}\big[\log \pi_g+\log h(W_{ig}|\nu_g)+\log \phi_{q\times p}(\mathscr{Y}^B_{ig}|{\bf 0},W_{ig}\mathbf{I}_q,\mathbf\Psi^{*}_g)\\& \quad+\log \phi_{n\times p}(\mathbf{X}_i|\mathbf{M}_g+W_{ig}\mathbf{A}_g+\mathbf\Lambda_g\mathscr{Y}_{ig}^B,W_{ig}\mathbf\Sigma_g,\mathbf\Psi_g^*)\big]\\ &=C+\sum_{i=1}^N\sum_{g=1}^G-\frac{1}{2}z_{ig}\bigg[-p\log|\mathbf\Sigma_g|+\,\mbox{tr}\bigg\{\frac{1}{W_{ig}}\mathbf\Sigma_g^{-1}(\mathbf{X}_i-\mathbf{M}_g)(\mathbf\Psi_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)'\\ &\quad-\mathbf\Sigma_g^{-1}(\mathbf{X}_i-\mathbf{M}_g)(\mathbf\Psi_g^*)^{-1}\mathbf{A}_g'\bigg\} -\frac{1}{W_{ig}}\mathbf\Sigma_g^{-1}(\mathbf{X}_i-\mathbf{M}_g)(\mathbf\Psi_g^*)^{-1}{\mathscr{Y}_{ig}^B}'\mathbf\Lambda_g'-\mathbf\Sigma_g^{-1}\mathbf{A}_g(\mathbf\Psi_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)'\\ &\quad+W_{ig}\mathbf\Sigma_g^{-1}\mathbf{A}_g(\mathbf\Psi_g^*)^{-1}\mathbf{A}_g'+\mathbf\Sigma_g^{-1}\mathbf{A}_g(\mathbf\Psi_g^*)^{-1}{\mathscr{Y}_{ig}^B}'\mathbf\Lambda_g' -\frac{1}{W_{ig}}\mathbf\Sigma_g^{-1}\mathbf\Lambda_g\mathscr{Y}_{ig}^B(\mathbf\Psi_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)'\\ &\quad+\mathbf\Sigma_g^{-1}\mathbf\Lambda_g\mathscr{Y}_{ig}^B(\mathbf\Psi_g^*)^{-1}\mathbf{A}_g'+\frac{1}{W_{ig}}\mathbf\Sigma_g^{-1}\mathbf\Lambda_g{\mathscr{Y}_{ig}^B}(\mathbf\Psi_g^*)^{-1}{\mathscr{Y}^B_{ig}}'\mathbf\Lambda_g'\bigg]. \end{align*} In the E-step, it can be shown that \begin{equation*}\begin{split} \mathscr{Y}_{ig}^B~|~\mathbf{X}_i,&W_{ig},z_{ig}=1\sim\\& \mathcal{N}_{q\times p}((\mathbf{I}_q+\mathbf\Lambda_g'\mathbf\Sigma_g^{-1}\mathbf\Lambda_g)^{-1}\mathbf\Lambda_g'\mathbf\Sigma_g^{-1}(\mathbf{X}_i-\mathbf{M}_g-W_{ig}\mathbf{A}_g),W_{ig}(\mathbf{I}_q+\mathbf\Lambda_g'\mathbf\Sigma_g^{-1}\mathbf\Lambda_g)^{-1},\mathbf\Psi^*_g) \end{split}\end{equation*} and so we can calculate the expectations \begin{equation*}\begin{split} {\bf E}_{1ig}^{(2)}&\coloneqq\mathbb{E}[\mathscr{Y}_{ig}^B|\hat{{\boldsymbol{\vartheta}}},\mathbf{X}_i,z_{ig}=1 =\mathbf{L}_g(\mathbf{X}_i-\hat{\mathbf{M}}_g-a_{ig}\hat{\mathbf{A}}_g),\\ {\bf E}_{2ig}^{(2)}&\coloneqq\mathbb{E}\left[\frac{1}{W_{ig}}\mathscr{Y}_{ig}^B\bigg|\hat{{\boldsymbol{\vartheta}}},\mathbf{X}_i,z_{ig}=1\right =\mathbf{L}_g(b_{ig}(\mathbf{X}_i-\hat{\mathbf{M}}_g)-\hat{\mathbf{A}}_g),\\ {\bf E}_{3ig}^{(2)}&\coloneqq\mathbb{E}\left[\frac{1}{W_{ig}}\mathscr{Y}_{ig}^B(\mathbf\Psi_g^*)^{-1}{\mathscr{Y}_{ig}^B}'\bigg|\hat{{\boldsymbol{\vartheta}}},\mathbf{X}_i,z_{ig}=1\right]\\%&=\mathbb{E}\left[\mathbb{E}\left[\frac{1}{W_{ig}}\mathscr{Y}_{ig}^B(\mathbf\Psi_g^*)^{-1}{\mathscr{Y}_{ig}^B}'|\mathbf{X}_i,W_{ig},z_{ig}=1\right]|\mathbf{X}_{i},z_{ig}=1\right]\\ &=p(\mathbf{I}_q+\hat{\mathbf\Lambda}_g'\hat{\mathbf\Sigma}_g^{-1}\hat{\mathbf\Lambda}_g)^{-1}+b_{ig}\mathbf{L}_g(\mathbf{X}_i-\hat{\mathbf{M}}_g)(\mathbf\Psi_g^*)^{-1}(\mathbf{X}_i-\hat{\mathbf{M}}_g)'\mathbf{L}_g'\\ &\quad-\mathbf{L}_g((\mathbf{X}_i-\hat{\mathbf{M}}_g)(\hat{\mathbf\Psi}_g^*)^{-1}\hat{\mathbf{A}}_g'+\hat{\mathbf{A}}_g(\hat{\mathbf\Psi}_g^*)^{-1}(\mathbf{X}_i-\hat{\mathbf{M}}_g)')\mathbf{L}_g'+a_{ig}\mathbf{L}_g\hat{\mathbf{A}}_g(\hat{\mathbf\Psi}_g^*)^{-1}\hat{\mathbf{A}}_g'\mathbf{L}_g', \end{split}\end{equation*} where $\mathbf{L}_g=(\mathbf{I}_q+\hat{\mathbf\Lambda}_g'\hat{\mathbf\Sigma}_g^{-1}\hat{\mathbf\Lambda}_g)^{-1}\hat{\mathbf\Lambda}_g'\hat{\mathbf\Sigma}_g^{-1}$. In the M-step, the updates for $\mathbf\Lambda_g$ and $\mathbf\Sigma_g$ are calculated. These updates are given by $$ \hat{\mathbf\Lambda}_g=\sum_{i=1}^N\hat{z}_{ig}\left[(\mathbf{X}_i-\hat{\mathbf{M}}_g)(\hat{\mathbf\Psi}_g^*)^{-1}{{\bf E}_{2ig}^{(2)}}'-\hat{\mathbf{A}}_g(\hat{\mathbf\Psi}_g^*)^{-1}{{\bf E}_{1ig}^{(2)}}'\right]\left(\sum_{i=1}^Nz_{ig}{\bf E}_{3ig}^{(2)}\right)^{-1} $$ and $ \hat{\mathbf\Sigma}_g=\,\mbox{diag}(\mathbf{S}^L_g), $ respectively, where \begin{equation*}\begin{split} \mathbf{S}^L_g=\frac{1}{N_gp}&\sum_{i=1}^N\hat{z}_{ig}\big[b_{ig}(\mathbf{X}_i-\hat{\mathbf{M}}_g)(\hat{\mathbf\Psi}_g^*)^{-1}(\mathbf{X}_i-\hat{\mathbf{M}}_g)'-(\hat{\mathbf{A}}_g+\hat{\mathbf\Lambda}_g{\bf E}_{2ig}^{(2)})(\hat{\mathbf\Psi}_g^*)^{-1}(\mathbf{X}_i-\hat{\mathbf{M}}_g)'\\ &-(\mathbf{X}_i-\hat{\mathbf{M}}_g)(\hat{\mathbf\Psi}_g^*)^{-1}\hat{\mathbf{A}}_g'+a_{ig}\hat{\mathbf{A}}_g(\hat{\mathbf\Psi}_g^*)^{-1}\hat{\mathbf{A}}_g+\hat{\mathbf\Lambda}_g{\bf E}_{1ig}^{(1)}(\hat{\mathbf\Psi}_g^*)^{-1}\hat{\mathbf{A}}_g'\\ &-(\mathbf{X}_i-\hat{\mathbf{M}}_g)(\hat{\mathbf\Psi}_g^*)^{-1}{{\bf E}_{2ig}^{(2)}}'\hat{\mathbf\Lambda}_g'+\hat{\mathbf{A}}_g(\hat{\mathbf\Psi}_g^*)^{-1}{{\bf E}_{1ig}^{(2)}}'\hat{\mathbf\Lambda}_g'+\hat{\mathbf\Lambda}_g{\bf E}_{3ig}^{(2)}\hat{\mathbf\Lambda}_g'\big]. \end{split}\end{equation*} \paragraph{AECM Stage 3} In the third stage, the complete-data consists of the observed data $\mathbf{X}_i$, the latent variables ${\bf W}_i$, the labels ${\bf z}_{i}$ and the latent matrices ${\boldsymbol \mathscr{Y}}^A_i=(\mathscr{Y}^A_{i1},\ldots,\mathscr{Y}^A_{iG})$ for $i=1,\ldots,N$. The complete-data log-likelihood at this stage is \begin{equation*}\begin{split} \ell_{\text{C}3}&=C+\sum_{i=1}^N\sum_{g=1}^Gz_{ig}\big[\log \pi_g+\log h(W_{ig}|\nu_g)+\log \phi_{q\times p}(\mathscr{Y}^A_{ig}|{\bf 0},W_{ig}\matsig_g^*,\mathbf{I}_p)\\& \quad+\log \phi_{n\times p}(\mathbf{X}_i|\mathbf{M}_g+W_{ig}\mathbf{A}_g+\mathscr{Y}_{ig}^A\mathbf\Delta_g',W_{ig}\matsig_g^*,\mathbf\Psi_g)\big]\\ &=C+\sum_{i=1}^N\sum_{g=1}^G-\frac{1}{2}z_{ig}\bigg[-n\log|\mathbf\Psi_g|+\,\mbox{tr}\bigg\{\frac{1}{W_{ig}}\mathbf\Psi_g^{-1}(\mathbf{X}_i-\mathbf{M}_g)'(\matsig_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)\\& \quad-\mathbf\Psi_g^{-1}(\mathbf{X}_i-\mathbf{M}_g)'(\matsig_g^*)^{-1}\mathbf{A}_g\bigg\} -\frac{1}{W_{ig}}\mathbf\Psi_g^{-1}(\mathbf{X}_i-\mathbf{M}_g)'(\matsig_g^*)^{-1}{\mathscr{Y}_{ig}^A}\mathbf\Delta_g'-\mathbf\Psi_g^{-1}\mathbf{A}_g'(\matsig_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)\\ &\quad+W_{ig}\mathbf\Psi_g^{-1}\mathbf{A}_g'(\matsig_g^*)^{-1}\mathbf{A}_g+\mathbf\Psi_g^{-1}\mathbf{A}_g'(\matsig_g^*)^{-1}{\mathscr{Y}_{ig}^A}\mathbf\Delta_g' -\frac{1}{W_{ig}}\mathbf\Psi_g^{-1}\mathbf\Delta_g{\mathscr{Y}_{ig}^A}'(\matsig_g^*)^{-1}(\mathbf{X}_i-\mathbf{M}_g)\\& \quad+\mathbf\Psi_g^{-1}\mathbf\Delta_g{\mathscr{Y}_{ig}^A}'(\matsig_g^*)^{-1}\mathbf{A}_g+\frac{1}{W_{ig}}\mathbf\Psi_g^{-1}\mathbf\Delta_g{\mathscr{Y}_{ig}^A}'(\matsig_g^*)^{-1}{\mathscr{Y}_{ig}^A}\mathbf\Delta_g'\bigg]. \end{split}\end{equation*} In the E-step, it can be shown that \begin{equation*}\begin{split} \mathscr{Y}_{ig}^A|\mathbf{X}_i,&W_{ig},z_{ig}=1\sim\\& \mathcal{N}_{n\times r}((\mathbf{X}_i-\mathbf{M}_g-W_{ig}\mathbf{A}_g)\mathbf\Psi_g^{-1}\mathbf\Delta_g(\mathbf{I}_r+\mathbf\Delta_g'\mathbf\Psi_g^{-1}\mathbf\Delta_g)^{-1},W_{ig}\matsig_g^*,(\mathbf{I}_r+\mathbf\Delta_g'\mathbf\Psi_g^{-1}\mathbf\Delta_g)^{-1}) \end{split}\end{equation*} and so we can calculate the expectations \begin{equation*}\begin{split} {\bf E}_{1ig}^{(3)}&\coloneqq\mathbb{E}[\mathscr{Y}_{ig}^A|\hat{{\boldsymbol{\vartheta}}},\mathbf{X}_i,z_{ig}=1 =(\mathbf{X}_i-\hat{\mathbf{M}}_g-a_{ig}\hat{\mathbf{A}}_g)\mathbf{D}_g,\\ {\bf E}_{2ig}^{(3)}&\coloneqq\mathbb{E}\left[\frac{1}{W_{ig}}\mathscr{Y}_{ig}^A|\hat{{\boldsymbol{\vartheta}}},\mathbf{X}_i,z_{ig}=1\right =(b_{ig}(\mathbf{X}_i-\hat{\mathbf{M}}_g)-\hat{\mathbf{A}}_g)\mathbf{D}_g,\\ {\bf E}_{3ig}^{(3)}&\coloneqq\mathbb{E}\left[\frac{1}{W_{ig}}{\mathscr{Y}_{ig}^A}'(\mathbf\Sigma_g^*)^{-1}{\mathscr{Y}_{ig}^A}|\hat{{\boldsymbol{\vartheta}}},\mathbf{X}_i,z_{ig}=1\right]\\%&=\mathbb{E}\left[\mathbb{E}\left[\frac{1}{W_{ig}}\mathscr{Y}_{ig}^B(\mathbf\Psi_g^*)^{-1}{\mathscr{Y}_{ig}^B}'|\mathbf{X}_i,W_{ig},z_{ig}=1\right]|\mathbf{X}_{i},z_{ig}=1\right]\\ &=n(\mathbf{I}_r+\hat{\mathbf\Delta}_g'\hat{\mathbf\Psi}_g^{-1}\hat{\mathbf\Delta}_g)^{-1}+b_{ig}\mathbf{D}_g'(\mathbf{X}_i-\hat{\mathbf{M}}_g)'(\hat{\matsig}_g^*)^{-1}(\mathbf{X}_i-\hat{\mathbf{M}}_g)\mathbf{D}_g\\ &\quad-\mathbf{D}_g'((\mathbf{X}_i-\hat{\mathbf{M}}_g)'(\hat{\matsig}_g^*)^{-1}\hat{\mathbf{A}}_g+\hat{\mathbf{A}}_g'(\hat{\matsig}_g^*)^{-1}(\mathbf{X}_i-\hat{\mathbf{M}}_g))\mathbf{D}_g+a_{ig}\mathbf{D}_g'\hat{\mathbf{A}}_g'(\hat{\matsig}_g^*)^{-1}\hat{\mathbf{A}}_g\mathbf{D}_g, \end{split}\end{equation*} where $\mathbf{D}_g=\hat{\mathbf\Psi}_g^{-1}\hat{\mathbf\Delta}_g(\mathbf{I}_r+\hat{\mathbf\Delta}_g'\hat{\mathbf\Psi}_g^{-1}\hat{\mathbf\Delta}_g)^{-1}$. In the M-step, the updates for $\mathbf\Delta_g$ and $\mathbf\Psi_g$ are calculated. These updates are given by $$ \hat{\mathbf\Delta}_g=\sum_{i=1}^N\hat{z}_{ig}[(\mathbf{X}_i-\hat{\mathbf{M}}_g)'(\hat{\matsig}_g^*)^{-1}{\bf E}_{2ig}^{(3)}-\hat{\mathbf{A}}_g'(\hat{\matsig}_g^*)^{-1}{\bf E}_{1ig}^{(3)}](\sum_{i=1}^Nz_{ig}{\bf E}_{3ig}^{(3)})^{-1} $$ and $ \hat{\mathbf\Psi}_g=\,\mbox{diag}(\mathbf{S}^D_g), $ respectively, where \begin{equation*}\begin{split} \mathbf{S}^D_g=\frac{1}{N_gp}&\sum_{i=1}^N\hat{z}_{ig}[b_{ig}(\mathbf{X}_i-\hat{\mathbf{M}}_g)'(\hat{\matsig}_g^*)^{-1}(\mathbf{X}_i-\hat{\mathbf{M}}_g)-(\hat{\mathbf{A}}_g'+\hat{\mathbf\Delta}_g{{\bf E}_{2ig}^{(3)}}')(\hat{\matsig}_g^*)^{-1}(\mathbf{X}_i-\hat{\mathbf{M}}_g)\\ &-(\mathbf{X}_i-\hat{\mathbf{M}}_g)'(\hat{\matsig}_g^*)^{-1}\hat{\mathbf{A}}_g+a_{ig}\hat{\mathbf{A}}_g'(\hat{\matsig}_g^*)^{-1}\hat{\mathbf{A}}_g+\hat{\mathbf\Delta}_g{{\bf E}_{1ig}^{(3)}}'(\hat{\matsig}_g^*)^{-1}\hat{\mathbf{A}}_g\\ &-(\mathbf{X}_i-\hat{\mathbf{M}}_g)'(\hat{\matsig}_g^*)^{-1}{\bf E}_{2ig}^{(3)}\hat{\mathbf\Delta}_g'+\hat{\mathbf{A}}_g'(\hat{\matsig}_g^*)^{-1}{\bf E}_{1ig}^{(3)}\hat{\mathbf\Delta}_g'+\hat{\mathbf\Delta}_g{\bf E}_{3ig}^{(3)}\hat{\mathbf\Delta}_g']. \end{split}\end{equation*} Details on initialization, convergence, model selection, and performance criteria are given in Appendix~B. \subsection{Reduction in the Number of Free Parameters in the Scale Matrices} The reduction in the number of free parameters in the scale matrices for each of these models is equivalent to the Gaussian case discussed in \cite{gallaugher18b}. The reduction in the number of free covariance parameters for the row scale matrix is \begin{equation}\label{eqn:cond1} \frac{1}{2}n(n+1)-nq-n+\frac{1}{2}q(q-1)=\frac{1}{2}\big[(n-q)^2-(n+q)\big], \end{equation} which is positive for $(n-q)^2>n+q$. Likewise, for the column scale matrix the reduction in the number of parameters is \begin{equation}\label{eqn:cond2} \frac{1}{2}p(p+1)-pr-p+\frac{1}{2}r(r-1)=\frac{1}{2}\big[(p-r)^2-(p+r)\big], \end{equation} which is positive for $(p-r)^2>p+r$. In applications herein, each model is fit for a range of row factors and column factors. If the number of factors chosen by the BIC is the maximum within the range, the number of factors is increased so long as the conditions \eqref{eqn:cond1} and \eqref{eqn:cond2} are met. \subsection{Semi-Supervised Classification} Each of the four models presented herein may also be used in the context of semi-supervised classification. Suppose $N$ matrices are observed and $K$ of these observations have known labels from one of $G$ classes. Following \cite{mcnicholas10c}, and without loss of generality, the matrices are ordered so that the first $K$ have known labels and the remaining observations have unknown labels. Using the MMVBFA model for illustration, the observed likelihood is then \begin{equation*}\begin{split} L({\boldsymbol{\vartheta}})=\prod_{i=1}^K\prod_{g=1}^G&\big[\pi_g\varphi_{n\times p}(\mathbf{X}_i~|~\mathbf{M}_g,\mathbf\Sigma_g+\mathbf{A}_g\mathbf{A}_g',\mathbf\Psi_g+\mathbf{B}_g\mathbf{B}_g')\big]^{z_{ig}}\\ &\times \prod_{j=K+1}^N\sum_{h=1}^H\pi_h\varphi_{n\times p}(\mathbf{X}_i~|~\mathbf{M}_h,\mathbf\Sigma_h+\mathbf{A}_h\mathbf{A}_h',\mathbf\Psi_h+\mathbf{B}_h\mathbf{B}_h'). \end{split}\end{equation*} Whilst it is possible that $H\ne G$, we assume that $H=G$ for the analyses herein. Parameter estimation then proceeds in a similar manner as for the clustering scenario. For more information on semi-supervised classification refer to \cite{mcnicholas10c,mcnicholas16a}. \subsection{Computational Issues} One situation that needs to be addressed for all four of these distributions, but particularly the variance-gamma distribution, is the infinite likelihood problem. This occurs as a result of the update for $\hat{\mathbf{M}}_g$ becoming very close, and in some cases equal to, an observation $\mathbf{X}_i$ when the algorithm gets close to convergence. A similar situation occurs in the multivariate case for the mixture of SAL distributions described in \cite{franczak14} and we follow a similar procedure when faced with this issue. While iterating the algorithm, when the likelihood becomes numerically infinite, we set the estimate of $\hat{\mathbf{M}}_g$ to the previous estimate which we will call $\hat{\mathbf{M}}^*_g$. We then update $\hat{\mathbf{A}}_g$ according to $$ \hat{\mathbf{A}}^*_g=\frac{\sum_{i=1}^N\hat{z}_{ig}(\mathbf{X}_i-{\hat{\mathbf{M}}}^*_g)}{\sum_{i=1}^N\hat{z}_{ig}a_{ig}}. $$ The updates for all other parameters remain the same. As mentioned in \cite{franczak14}, this solution is a little naive; however, it does generally work quite well. It is not surprising that this problem is particularly prevalent in the case of the variance-gamma distribution because the SAL distribution arises as a special case of the variance-gamma distribution. Another computational concern is in the evaluation of the Bessel functions. In the computation of the GIG expected values and the component densities, it may be the case that the argument is far larger than the magnitude of the index---especially in higher dimensional cases. Therefore, in these situations, the result is computationally equivalent to zero which causes issues with other computations. In such a situation, we calculate the exponentiated version of the Bessel function, i.e., we calculate $\exp(u)K_{\lambda}(u)$ and subsequent calculations can be easily adjusted. \section{Simulation Study} A simulation study was performed for each of the four models presented herein. For each of the four models, we consider $d\times d$ matrices with $d\in\{10,30\}$ and, for each value of $d$, we consider datasets coming from a mixture with two components and $\pi_1=\pi_2=0.5$. The datasets have sample sizes $N\in\{100,200,400\}$ and the following parameters are used for all four models for each combination of $d$ and $N$. We take $\mathbf{M}_1={\bf 0}$ and $\mathbf{M}_2=\mathbf{M}_1+\mathbf{C}$, where $\mathbf{C}$ is a matrix with all entries equal to $c$ for $c\in\{1,2,4\}$. All other parameters are held constant. We take $\mathbf\Sigma_1=2\mathbf{I}_d$, $\mathbf\Sigma_2=\mathbf{I}_d$, $\mathbf\Psi_1=\mathbf{I}_d$, $\mathbf\Psi_2=2\mathbf{I}_d$, and $\mathbf{A}_1=\mathbf{A}_2={\bf 1}$, where ${\bf 1}$ is a matrix of 1's. Three column factors and two row factors are used with their values being randomly drawn from a uniform distribution on $[-1,1]$. See Table~\ref{tab:params} for distribution-specific parameters. \begin{table}[!htb] \centering \caption{Distribution-specific parameters used for the simulations, where the acronyms all take the form MMVDFA and denote ``mixture of matrix variate D factor analyzers'' with $\text{D}\in\{\text{skew-$t$ (ST)},\text{generalized hyperbolic (GH)},\text{variance-gamma (VG)},\text{NIG}\}$.} {\scriptsize\begin{tabular}{lrr} \hline & Component 1 & Component 2\\ \hline MMVSTFA & $\nu_1=4$ & $\nu_2=20$\\ MMVGHFA& $\omega_1=4,\lambda_1=-4$ & $\omega_2=10,\lambda_2=4$\\ MMVVGFA& $\gamma_1=4$ & $\gamma_2=10$\\ MMVNIGFA& $\kappa_1=2$ & $\kappa_2=4$\\ \hline \end{tabular}} \label{tab:params} \end{table} We fit the MMVSTFA model to data that is simulated from the MMVSTFA model using the parameters above together with the distribution-specific parameters in Table~\ref{tab:params}. We take an analogous approach with the MMVGHFA, MMVVGFA, and MMVNIGFA models. However, we fit the MMVBFA model to data that is simulated from the MMVVGFA model---this is done to facilitate an illustration that uses data simulated from a mixture of skewed matrix variate distributions. We fit all models for $G\in\{1,2,3,4\}$ and $q,r\in\{1,2,3,4,5\}$. In Tables~\ref{tab:GQR1} and~\ref{tab:GQR2}, we show the number of times that the BIC correctly chooses the number of groups, row factors, and column factors. In Table~\ref{tab:ARIsim}, the average ARI and corresponding standard deviation for each setting is shown. As one would expect, for each model introduced herein, the classification performance generally improves as $N$ increases. However, this is not the case for the MMVBFA model. In the case $d=10$, it is interesting to note that the number of correct choices made by the BIC for the row and column factors generally decreases as we increase the separation (Table~\ref{tab:GQR1}). However, when $d$ is increased to 30, there is no clear trend in this regard (Table~\ref{tab:GQR2}). The classification performance for the four models introduced here in is excellent overall (Table~\ref{tab:ARIsim}). However, when fitting the MMVBFA model to data simulated from the MMVVGFA model, the BIC never chooses the correct number of groups for $N\in\{200,400\}$. Furthermore, although not apparent from the tables, the model generally overfits the number of groups which, as in the multivariate case, is to be expected when using a Gaussian mixture model in the presence of skewness or outliers. \begin{table}[!htb] \centering \caption{Number of datasets for which the BIC correctly chose the number of groups, row factors, and column factors ($d=10$).} {\scriptsize\begin{tabular}{ll|ccc|ccc|ccc|ccc|ccc} \hline &&\multicolumn{3}{c|}{MMVSTFA}&\multicolumn{3}{|c|}{MMVGHFA}&\multicolumn{3}{|c|}{MMVVGFA}&\multicolumn{3}{|c|}{MMVNIGFA}&\multicolumn{3}{|c}{MMVBFA}\\ $c$&$N$&$G$&$q$&$r$&$G$&$q$&$r$&$G$&$q$&$r$&$G$&$q$&$r$&$G$&$q$&$r$\\ \hline \multirow{3}{*}{$1$} &$100$&18&15&19&25&16&24&18&10&12&21&21&20&17&15&19\\ &$200$&23&18&21&25&25&25&21&11&8&24&24&24&0&19&23\\ &$400$&25&21&21&25&25&25&22&17&15&24&24&25&0&19&14\\ \hline \multirow{3}{*}{$2$} &$100$&18&14&17&25&9&22&16&7&4&17&18&19&16&17&15\\ &$200$&24&18&19&25&22&22&23&10&2&23&23&24&0&20&20\\ &$400$&25&23&23&25&25&25&19&20&19&25&24&25&0&21&14\\ \hline \multirow{3}{*}{$4$} &$100$&8&13&14&23&5&10&17&11&0&24&2&7&19&23&9\\ &$200$&22&9&16&25&4&16&21&8&8&24&7&24&0&18&18\\ &$400$&25&12&21&25&22&12&17&10&19&21&0&14&0&17&19\\ \hline \end{tabular}} \label{tab:GQR1} \end{table} \begin{table}[!htb] \centering \caption{Number of datasets for which the BIC correctly chose the number of groups, row factors, and column factors ($d=30$).} {\scriptsize\begin{tabular}{ll|ccc|ccc|ccc|ccc|ccc} \hline &&\multicolumn{3}{c|}{MMVSTFA}&\multicolumn{3}{|c|}{MMVGHFA}&\multicolumn{3}{|c|}{MMVVGFA}&\multicolumn{3}{|c|}{MMVNIGFA}&\multicolumn{3}{|c}{MMVBFA}\\ $c$&$N$&$G$&$q$&$r$&$G$&$q$&$r$&$G$&$q$&$r$&$G$&$q$&$r$&$G$&$q$&$r$\\ \hline \multirow{3}{*}{$1$} &$100$&24&11&12&25&15&18&25&12&12&25&20&21&15&6&2\\ &$200$&25&17&18&25&22&23&25&21&20&25&23&25&0&4&3\\ &$400$&25&22&23&25&25&24&25&25&20&25&25&25&0&10&1\\ \hline \multirow{3}{*}{$2$} &$100$&24&15&17&25&17&18&25&13&11&25&22&23&14&5&0\\ &$200$&25&22&19&25&19&22&25&20&22&25&23&25&0&5&3\\ &$400$&25&19&20&25&22&24&25&23&24&25&24&25&0&9&8\\ \hline \multirow{3}{*}{$4$} &$100$&24&17&17&25&12&14&25&17&14&25&23&16&18&2&2\\ &$200$&25&18&20&25&18&23&25&21&22&25&21&21&0&3&8\\ &$400$&25&15&15&25&20&24&25&19&20&25&21&22&0&5&7\\ \hline \end{tabular}} \label{tab:GQR2} \end{table} \begin{table}[!htb] \centering \caption{Average ARI values over 25 runs for each setting with standard deviations in parentheses.} {\scriptsize \scalebox{0.8}{\begin{tabular}{ll|cc|cc|cc|cc|cc} \hline && \multicolumn{2}{c|}{MMVSTFA}& \multicolumn{2}{c|}{MMVGHFA}& \multicolumn{2}{c|}{MMVVGFA}& \multicolumn{2}{c|}{MMVNIG}& \multicolumn{2}{c}{MMVBFA}\\ $c$&$N$ &$d=10$&$d=30$&$d=10$&$d=30$&$d=10$&$d=30$&$d=10$&$d=30$&$d=10$&$d=30$\\ \hline \multirow{3}{*}{$1$}&$100$&0.91(0.08)&0.96(0.01)&0.97(0.03)&0.97(0.02)&0.90(0.1)&0.97(0.02)&0.98(0.05)&1.00(0.0)&0.90(0.1)&0.91(0.1)\\ &$200$&0.98 (0.03)&0.99(0.009)&1.00(0.006)&1.00(0.007)&0.97(0.03)&0.99(0.01)&1.00(0.007)&1.00(0.0)&0.75(0.05)&0.76(0.01)\\ &$400$&1.00 (0.005)&1.00(0.004)&1.00(0.0)&1.00(0.0)&0.99(0.03)&1.00(0.0)&1.00(0.006)&1.00(0.0)&0.69(0.1)&0.54(0.07)\\ \hline \multirow{3}{*}{$2$} &$100$&0.94(0.03)&0.96(0.02)&0.96(0.03)&0.97(0.03)&0.88(0.1)&0.98(0.02)&0.96(0.07)&1.00(0.0)&0.91(0.1)&0.90(0.1)\\ &$200$&0.98 (0.02)&0.99(0.009)&1.00(0.0)&1.00(0.0)&0.97(0.05)&1.00(0.007)&0.99(0.02)&1.00(0.0)&0.76(0.01)&0.76(0.02)\\ &$400$&1.00 (0.005)&1.00(0.003)&1.00(0.0)&1.00(0.0)&0.98(0.05)&1.00(0.0)&1.00(0.0)&1.00(0.0)&0.66(0.1)&0.53(0.07)\\ \hline \multirow{3}{*}{$4$} &$100$&0.84(0.08)&0.97(0.03)&0.94(0.04)&0.97(0.02)&0.92(0.1)&1.00(0.01)&0.98(0.08)&1.00(0.0)&0.94(0.1)&0.93(0.1)\\ &$200$&0.98 (0.02)&1.00(0.004)&1.00(0.0)&1.00(0.0)&0.97(0.05)&1.00(0.0)&0.99(0.02)&1.00(0.0)&0.76(0.02)&0.76(0.01)\\ &$400$&1.00 (0.004)&1.00(0.002)&1.00(0.0)&1.00(0.0)&0.98(0.03)&1.00(0.0)&1.00(0.03)&1.00(0.0)&0.73(0.08)&0.54(0.07)\\ \hline \end{tabular}}} \label{tab:ARIsim} \end{table} \section{MNIST Digits} \cite{gallaugher18a,gallaugher18b} consider the MNIST digits dataset; specifically, looking at digits 1 and 7 because they are similar in appearance. Herein, we consider the digits 1, 6, and 7. This dataset consists of 60,000 (training) images of Arabic numerals 0 to 9. We consider different levels of supervision and perform either clustering or semi-supervised classification. Specifically we look at 0\% (clustering), 25\%, and 50\% supervision. For each level of supervision, 25 datasets consisting of 200 images each of digits 1, 6, and 7 are taken. As discussed in \cite{gallaugher18a}, because of the lack of variability in the outlying rows and columns of the data matrices, random noise is added to ensure non-singularity of the scale matrices. Each of the four models developed herein, as well as the MMVBFA model, are fitted for 1 to 17 row and column factors. In Table~\ref{tab:ARIMNIST}, the average ARI and misclassification rate (MCR) values are presented for each model and each level of supervision. \begin{table}[!htb] \centering \caption{Average ARI and MCR values for the MNIST dataset for each level of supervision, with respective standard deviations in parentheses for digits 1,6, and 7.} {\scriptsize\begin{tabular}{ll|rrrrr} \hline Supervision&&MMVSTFA&MMVGHFA&MMVVGFA&MMVNIGFA&MMVBFA\\ \hline \multirow{2}{*}{0\% (clustering)}&${\text{ARI}}$&0.58(0.09)&0.58(0.09)&0.62(0.1)&0.47(0.1)&0.36(0.09)\\ &${\text{MCR}}$&0.17(0.04)&0.17(0.08)&0.15(0.04)&0.22(0.05)&0.28(0.09)\\ \hline \multirow{2}{*}{25\%}&${\text{ARI}}$&0.72(0.1)&0.72(0.1)&0.75(0.1)&0.64(0.2)&0.51(0.16)\\ &${\text{MCR}}$&0.10(0.04)&0.10(0.04)&0.094(0.04)&0.14(0.07)&0.20(0.07)\\ \hline \multirow{2}{*}{50\%}&${\text{ARI}}$&0.83(0.07)&0.85(0.03)&0.83(0.07)&0.81(0.1)&0.72(0.06)\\ &${\text{MCR}}$&0.059(0.03)&0.052(0.02)&0.061(0.03)&0.067(0.05)&0.10(0.06)\\ \hline \end{tabular}} \label{tab:ARIMNIST} \end{table} In the completely unsupervised case, three of the skewed models have a MCR of around 16\%. However, at 25\% supervision, this decreases to around 10\% and, at 50\% supervision, this falls again to around 5\%. At all three levels of supervision, it is clear that all four skewed mixture models introduced herein outperform the MMVBFA model. In fact, the performance of the MMVBFA model at 25\% supervision is not as good as that of the MMVVGFA, MMVGHFA or MMVSTFA models in the completely unsupervised case (i.e., 0\% supervision). \begin{figure}[!htb] \centering \includegraphics[width=0.5\textwidth,height=0.5\textwidth]{Image.pdf} \caption{Heat maps of estimated location matrices for the MMVBFA and MMVVGFA models for each class in the unsupervised case.} \label{fig:MNIST} \end{figure} It is of interest to compare heat maps of the estimated location matrices for the MMVBFA and MMVVGFA models for one of the datasets in the unsupervised case (Figure~\ref{fig:MNIST}). It can be seen that the images are a lot clearer for the MMVVGFA model compared to the MMVBFA model. This is particularly prominent when considering the the results for the digit 6, for which one can see a possible 1 or 7 in the background for the MMVBFA heat map. Moreover, for digit 1, one can see a faint 6 in the background when looking at the MMVBFA heat map. \section{Discussion} The MMVBFA model has been extended to four skewed distributions; specifically, the matrix variate skew-$t$, generalized hyperbolic, variance-gamma, and NIG distributions. AECM algorithms were developed for parameter estimation, and the novel approaches were illustrated on real and simulated data. In the simulations, the models introduced herein generally exhibited very good performance under various scenarios. As expected, the MMVBFA model did not perform well when applied to data from the MMVVGFA model. In the real data example, all four of the skewed matrix variate models introduced herein performed better than the MMVBFA model. As one would expect, the difference in performance was most stark in the clustering case. Software to implement the approaches introduced herein, written in the Julia language \citep{bezanson17,mcnicholas19}, is available in the {\tt MatrixVariate.jl} repository \citep{pocuca19}. Future work will include considering a family of models similar to the parsimonious Gaussian mixture models of \cite{mcnicholas08,mcnicholas10d}. Another area of future work would be to compare this method of directly modelling skewness to using transformations such as those found in \cite{melnykov18, melnykov19}. It might also be of interest to consider matrix variate data of mixed type, which would allow, for example, analysis of multivariate longitudinal data where the variables are of mixed type. Finally, this methodology could be extended to mixtures of multidimensional arrays \citep[see][]{tait19}, which would be useful for studying several data types including coloured images and black and white movie clips. \section*{Acknowledgements} This work was supported by a Vanier Canada Graduate Scholarship (Gallaugher), the Canada Research Chairs program (McNicholas), and an E.W.R.\ Steacie Memorial Fellowship (McNicholas). {\small

a40b4c95bdac736690f33632162f5ff6756b8995

\section{Introduction} Consider the feasibility problem: \begin{eqnarray} \mbox{Find}~~ &X \in \mathcal{R}^{N\times N}\nonumber \\ \mbox{such~that}~~ &\mathcal{A}(X) = b, ~ X \succeq 0 \mbox{~~and~~} \mbox{Rank} (X) = k, \label{main_prob} \end{eqnarray} where $\mathcal{A}$ is a linear transformation from $\mathcal{R}^{N\times N}$ to $\mathcal{R}^{m}$ and $b\in \mathcal{R}^{m\times 1}$. It is also assumed that $m < N^2$. The case where the rank of the solution is constrained to be 1 is of particular interest to many applications such as combinatorics and signal processing. In the following discussions, (i) $[e;~f]$ would mean the concatenated vector by appending the vector $f$ below $e$, (ii) $\mbox{diag}(X)$ would mean the column vector of the diagonal entries of the matrix $X$, (iii) $X(m:n,p:q)$ would the sub-matrix of $X$ comprised of the elements with row index between $m$ and $n$ and column index between $p$ and $q$, (iv) $X(:,i)$ would be the $i^{\rm th}$ column of $X$, $X(i,:)$ would be the $i^{\rm th}$ row of $X$, (v) $X(m:n,i)$ would be the $i^{\rm th}$ column with elements from row $m$ to row $n$ (similar interpretation would hold for $X(i,m:n)$), (vi) $X(:)$ or $\mbox{\textbf{vec}}(X)$ would mean the vector obtained by vectorizing $X$, and (vii) $\mbox{Sym}^N$ and $S_+^N$ are the spaces of $N\times N$ symmetric matrices and positive semidefinite matrices, respectively. Continuing, a typical decision problem in combinatorics would have the following form: \begin{eqnarray} \mbox{Find}~~ &x \in \{0,1\}^{N\times 1} \nonumber\\ \mbox{such~that}~~ &\mathcal{C}x = b, \label{combi_prob} \end{eqnarray} where $C\in \mathcal{R}^{m\times N}$ and $b\in \mathcal{R}^{m\times 1}$. A standard method to convert this into a form given in \eqref{main_prob} is to firstly recognize the simple fact that $z \in \{0,1\}$ if and only if $z(z-1) = 0$. Now suppose $X = [1; ~x][1; ~x]^{\top}$. By construction, the matrix $X$ is positive semidefinite and has unity rank. Moreover, the constraint $x_i(x_i-1) = 0$, where $x_i$ is the $i^{\mbox th}$ element of the vector $x$, can be equivalently written as $\mbox{diag}(X) = X(:,1)$. With all these facts in place, the feasibility problem \eqref{combi_prob} can be written as \begin{eqnarray} \mbox{Find}~~ &X \in \mathcal{R}^{N\times N}\nonumber \\ \mbox{such~that}~~ &\mathcal{C}X_{(\small{2:N,1})} = b, ~\mbox{diag}(X) = X_{(:,1)}, ~X \succeq 0 \mbox{~~and~~} \mbox{Rank} (X) = 1. \label{combi_prob_conversion} \end{eqnarray} A problem from the field of non-linear equations which can be converted to rank constrained feasibility problem is the following: \begin{eqnarray} \mbox{Find}~~ &[\theta_1,~\theta_2,\cdots,~\theta_N] \in [-\pi,\pi]^{N\times 1}\nonumber \\ \mbox{such~that}~~ & Ax = b, \mbox{~~where~} x = \left[e^{\mbox{j}\theta_1},~e^{\mbox{j}\theta_2},\cdots,~e^{\mbox{j}\theta_N}\right]^\top, \label{nl_prob1} \end{eqnarray} and $A\in \mathcal{R}^{m\times N}$ and $b\in \mathcal{R}^{m\times 1}$. Here it is assumed that the given system of equations has a solution. As in the example in combinatorics, suppose that $X = [1; ~x][1; ~x]^{\mbox H}$. Again by construction, $X \succeq 0$ and $X$ has unity rank. A minor difference, as compared to the previous example, is the fact that the $X$ here is defined on the complex field. Note that $\mbox{diag}(X)=\textbf{1}$, a column of ones. With this, the problem in \eqref{nl_prob1} can be cast as: \begin{eqnarray} \mbox{Find}~~ &X \in \mathcal{R}^{N\times N}\nonumber \\ \mbox{such~that}~~ &\mathcal{A}X(\small{2:N,1}) = b, ~\mbox{diag}(X) = \textbf{1}, ~X \succeq 0 \mbox{~~and~~} \mbox{Rank} (X) = 1. \label{combi_prob_conversion} \end{eqnarray} It is also worth mentioning an instance where a non-unity rank comes of use is in a very elegant formulation of the Optimal Power Flow problem proposed in \cite{lavaei2012zero}. It is shown in \cite{lavaei2012zero} that a sufficient condition for a solution to the dual problem to be optimal to the primal is that a specific positive semidefinite matrix (affine in the optimization variables) has a null space of dimension 2. The aforementioned examples highlight the advantage of using same rank-constrained feasibility formulation to represent a 0/1 programming problem on one hand, and a continuous domain feasibility problem on the other. Many more applications can be found in the paper \cite{fazel2004rank,recht2010guaranteed}, which serve as the motivation for studying this problem in detail. In this paper, the general rank constrained problem is also considered, which is given by: \begin{eqnarray} \mbox{Find}~~ &X \in \mathcal{R}^{N\times M}\nonumber \\ \mbox{such~that}~~ &\mathcal{A}(X) = b \mbox{~~and~~} \mbox{Rank} (X) = k. \label{main_prob_gen} \end{eqnarray} A related problem is the \textit{Minimum-Rank-Problem} (MRP) given by: \begin{eqnarray} \mbox{Find}~~ &X \in \mathcal{R}^{M\times N} \mbox{~~with ~the ~minimum~ rank} \nonumber \\ \mbox{such~that}~~ &\mathcal{A}(X) = b \mbox{~and~} X \succeq 0. \label{min_rank_prob} \end{eqnarray} A popular heuristic employed to solve \eqref{min_rank_prob} is called the \textbf{log-det} heuristic, first proposed in the seminal paper \cite{fazel2003log}. This method relies on the intuition that minimizing $\log(\det(X))$ would naturally reduce the singular values and hence lead to rank minimization. In particular, this is achieved by the following iterative scheme, which derived using gradient descent: \begin{eqnarray} X^{k+1} = \displaystyle \mbox{argmin}_{X \mbox{~s.t.~} \mathcal{A}(X)=b}~~\mbox{Trace}((X^k + \delta I_N)^{-1}X), \end{eqnarray} where $\delta>0$ is a regularization parameter to ensure invertibility of $(X^k + \delta I_N)$, and $X^0$ is typically set to the identity matrix. However, note that this matrix might become il-conditioned with the iterations thereby making the inverse error prone. Moreover, inversion is also computationally intensive as the dimension of the problem grows. \\\\ A variant of \eqref{min_rank_prob} problem is the \textit{General Minimum-Rank-Problem} (MRP) given by: \begin{eqnarray} \mbox{Find}~~ &X \in \mathcal{R}^{M\times N} \mbox{~~with ~the ~minimum~ rank} \nonumber \\ \mbox{such~that}~~ &\mathcal{A}(X) = b. \label{min_rank_prob_gen} \end{eqnarray} In the seminal paper \cite{recht2010guaranteed}, it has been proved that minimizing nuclear norm of $X$ results in the minimum rank solution provided the linear operator $\mathcal{A}$ satisfies the Restricted Isometry Property (RIP). The nuclear norm minimization is the following convex program: \begin{equation} \begin{array}{l} \displaystyle \min_{X\in \mathcal{R}^{N\times M}} ~~~\sigma_*(X)\\ \mbox{subject to}\\ \hspace*{1.5cm} \mathcal{A}(X) = b, \end{array} \end{equation} where $\sigma_*(X)$ is the sum of the singular values of $X$. However, checking RIP for a given linear operator is itself NP-Hard in general. In certain special cases where $\mathcal{A}$ is sampled from Gaussian ensembles, it has been shown that it satisfies RIP with a high probability \cite{recht2010guaranteed}. Loosely speaking, this would mean that given an optimization problem where it is known that the linear constraint operator is sampled from special ensembles, one can find the minimum rank solution with a high probability. In addition, another common method employed to solve \eqref{min_rank_prob_gen} is the coordinate descent method, in which firstly, $X$ is parameterized as $L\times R$, where $L\in \mathcal{R}^{M\times r}$ and $R\in \mathcal{R}^{M\times r}$. Then $||\mathcal{A}(X) - b||_2$ is minimized iteratively; in each iteration either $L$ or $R$ is alternatively held to the value of the previous iteration. Note that the optimization problem \eqref{min_rank_prob} is devoid of the positive semidefinite constraint. But then one can modify it to the following equivalent problem (as shown in \cite{fazel2003log}): \begin{eqnarray} \mbox{Find}~~ &X \in \mathcal{R}^{M\times N}, Y \in \mathcal{R}^{M\times M} \mbox{~and~} Z \in \mathcal{R}^{N\times N} \mbox{~~with ~the ~minimum~ \mbox{Rank}(Y)+\mbox{Rank}(Z)} \nonumber \\ \mbox{such~that}~~ &\mathcal{A}(X) = b \mbox{~~and}~~ \begin{bmatrix} Y&X\\X^\top&Z \end{bmatrix} \succeq 0. \label{min_rank_prob_equi} \end{eqnarray} The modified problem falls into the class of problems defined by \eqref{main_prob}. Therefore, \eqref{main_prob} is more general than \eqref{min_rank_prob}, in the sense that if there is an algorithm to solve \eqref{main_prob}, the one can also solve \eqref{min_rank_prob} by iterating over all possible ranks and choosing the one with the lowest rank. Moreover, with this formulation one can also used the \textbf{log-det} heuristic. Also note that minimizing nuclear norm with positive definite constraint boils down to minimizing the trace of the matrix. \section{The proposed method} In this section, the proposed heuristic will be outlined. In addition, the intuition leading to the development of the heuristic will also be discussed, a few interesting observations will be made and a few relevant questions will be posed. The author would like to begin by stating the following two standard lemmas.\\\\ \textbf{\textit{Lemma 1}}: For a positive semidefinite matrix $X$, $X_{i,i}X_{j,j}\geq X^2_{i,j}$ for any $1\leq i,j \leq N$.\\ \textbf{\textit{Proof}}: For a positive semidefinite matrix, any principal $2\times 2$ minor is positive definite. Consider the $2\times 2$ principal minor given by $\displaystyle \begin{bmatrix}X_{i,i}& X_{i,j}\\X_{j,i} & X_{j,j}\end{bmatrix}$. Since this is positive semidefinite, its determinant is also positive, that is, $X_{i,i}X_{j,j}-X^2_{i,j}\geq 0$. \hfill \textbf{Q.E.D.}\\\\ \textbf{\textit{Definition 2}}: Note the following definition motivated from \textit{Lemma 1}, termed as the \textit{Low-Rank-Functional} (LRF) $\mathcal{L}: S^N_+\rightarrow \mathcal{R}$: \begin{equation} \displaystyle \mathcal{L}(X) = \sum_{i,j}\left(X_{i,i}X_{j,j}-X^2_{i,j}\right) = X^\top(:)QX(:). \label{lrf_def} \end{equation} A theorem follows immediately from this definition.\\\\ \textbf{\textit{Theorem 3}}: A non-zero positive semidefinite matrix $X$ has unity rank if and only if $\mathcal{L}(X)=0$.\\ \textbf{\textit{Proof}}: Suppose $X$ has unity rank. Then X = $xx^\top$ and therefore $X_{i,j} = x_ix_j$. It is immediately clear that $\mathcal{L}(X) = 0$. The proof of the converse is by induction on the dimension $N$ of the matrix $X$. For $N=2$, it is immediately clear that if for some non-zero matrix $X\succeq 0$, $\mathcal{L}(X)=det(X)=0$, rank of $X$ has to be unity. Suppose this is true for the case $N\leq M$, for some positive integer $M$ greater than 2. Then one proceeds to prove the statement for dimension $(M+1)$. Suppose $X\in \mbox{Sym}^{(M+1)}$, $X\succeq 0$ and that $\mathcal{L}(X)=0$. Write $\displaystyle X = \begin{bmatrix} X_M & v\\ v^\top & z \end{bmatrix}$. Note that $\displaystyle \mathcal{L}(X) = \mathcal{L}(X_M) + \sum_{i}\left(zX_{i,i}-v_i^2\right)$. Since $\mathcal{L}(X)$ is 0 and $\displaystyle \sum_{i}\left(zX_{i,i}-v_i^2\right)$ is non-negative (due to Lemma 1), $\mathcal{L}(X_M)$ is also equal to zero. Suppose $X_M$ is a zero matrix. Then, since $\left(zX_{i,i}-v_i^2\right)=0$ for each $i$, $v=0$. Moreover, $z>0$ as it is assumed $X$ is a non-zero matrix. In this case, the rank of $X$ is certainly unity. Now suppose $X_M$ is not a zero matrix. Again, since $\mathcal{L}(X_M)=0$, by the induction hypothesis, $X_M$ has unity rank. Say $X_M = xx^\top$. Now, there are two cases possible: (i) $z = 0$ and (ii) $z>0$. For case (i), $v=0$ and therefore the rank of $X$ still remains unity. For case (ii), by Schur's complement, $X\succeq 0$ if and only if $\displaystyle \left(xx^\top - \frac{1}{z}vv^\top\right) \succeq 0$. This condition implies that $v = kx$, such that $k^2\leq z$. Now suppose $x_r$ is a non-zero element of $x$. Then, $\left(zx_r^2 - k^2x_r^2\right) = 0$ would imply $z = k^2$. This in turn means that $X = \begin{bmatrix} xx^\top & kx\\kx^\top & k^2 \end{bmatrix}$. Therefore, $X$ has unity rank, and the proof concludes. \hfill \textbf{Q.E.D.}\\\\ \textbf{\textit{Lemma 4}}: For any two positive semidefinite matrices $X$ and $Y$, $\mbox{Tr}(XY) \geq 0$.\\ \textbf{\textit{Proof}}: Consider two unity rank positive semidefinite matrices $A$ and $B$. Since each has unity rank, $A = aa^\top$ and $B = bb^\top$. Now, it is easy to see that $\mbox{Trace}(AB) = (a^\top b)^2 \geq 0$. By Singular Value Decomposition, $\displaystyle X = \sum_{i}^{r_X} X_i$ and $\displaystyle Y = \sum_{i}^{r_Y} Y_i$. Therefore, $\displaystyle \mbox{Trace}(XY) = \sum_{i,j}\mbox{Trace}(X_iY_j)$. Since each of the terms in the summation is non-negative by the last statement on trace of product of the two unity rank PSD matrices, $\mbox{Trace}(XY)\geq 0$. \hfill \textbf{Q.E.D.}\\\\ \textbf{\textit{Definition 5}}: Now for each integer $1\leq r\leq N$, one can make also the following definition, termed here as the \textit{Particular-Rank-Functional} (PRF), $\mathcal{P}_r:\underbrace{S^n_+\times\cdots\times S^n_+}_{\mbox{r-times}} \rightarrow \mathcal{R}$ as: \begin{equation} \mathcal{P}_r(X^{(1)},\cdots,~X^{(r)}) = \sum_{k,i,j}\left(X^{(k)}_{i,i}X^{(k)}_{j,j}-\left(X^{(k)}_{i,j}\right)^2\right) + \sum_{p\ne q}\mbox{Tr}\left(X^{(p)}X^{(q)}\right) = \tilde{X}^\top Q_r\tilde{X}, \label{eqn:part_rank_def} \end{equation} where $\tilde{X}=[X^{(1)}(:);\cdots;X^{(r)}(:)]$. Note that $\mathcal{P}_1(X) = \mathcal{L}(X)$ for a PSD matrix $X$. Now one can derive the following theorem:\\\\ \textbf{\textit{Theorem 6}}: A positive semidefinite matrix $X$ defined as: \begin{equation} X = \sum_{k=1}^{r}X^{(k)}, \end{equation} where each matrix in the summation is also PSD, has rank $r$ if $\mathcal{P}_r(X^{(1)},\cdots,~X^{(r)})=0$. Conversely, if a matrix $X$ has rank $r$, then it can be written as a sum of $r$ positive semidefinite matrices $\{X^{(1)},\cdots,~X^{(r)}\}$, such that $\mathcal{P}(X^{(1)},\cdots,~X^{(r)})=0$.\\ \textbf{\textit{Proof}}: Suppose that for a set of PSD matrices $\{X^{(1)},\cdots,~X^{(r)}\}$, $\mathcal{P}_r(X^{(1)},\cdots,~X^{(r)})$ is zero. Since each of the summands in \eqref{eqn:part_rank_def} is nonnegative (by \textit{Lemma 1} and \textit{Lemma 4}), each term has to be equal to 0. Note that the first summation term in \eqref{eqn:part_rank_def} is equal to $\displaystyle \sum_{k=1}^{r}\mathcal{L}(X^{(k)})$. Since each of these terms is equal to 0, it implies that each $X^{(k)}$ has unity rank. Similarly, the summands in the second summation term in \eqref{eqn:part_rank_def} equal to 0 imply that the $X^{(i)}$ and $X^{(j)}$ are orthogonal for all $i\neq j$. Hence, the rank of $X = \sum_{k=1}^{r}X^{(k)}$ is equal to $r$. The second statement of the theorem is a corollary of Singular Value Decomposition. \hfill Q.E.D.\\\\ \textbf{\underline{Construction of the heuristic}}: With the aforementioned results, it is quite intuitive that given an instance of the feasibility problem \eqref{main_prob}, one can replace it by: \begin{eqnarray} \nonumber \displaystyle \min_{X^{(1)},\cdots,X^{(k)}}~~& \mathcal{P}_k(X^{(1)},\cdots,X^{(k)})\\ \displaystyle \mbox{such~that}~~ &\mathcal{A}\left(\sum_{j=1}^{k}X^{(j)}\right) = b, ~ X^{(j)} \succeq 0, ~\forall j. \label{main_prob_heuristic} \end{eqnarray} Note that \eqref{main_prob_heuristic} is a non-convex quadratic optimization problem. One standard method to deal with non-convex problems is the gradient descent. This would constitute the first heuristic. Note that the gradient here would be a linear functional which is computationally efficient. The other method is to convert the quadratic problem into a bilinear form and solve a sequence of convex problems fixing one of the variables in the bilinear form to the previous iterate. This would constitute the basic idea for the second heuristic. Note that convex problem in each iteration would be linear semidefinite program which can be solved efficiently using standard optimizers. \\\\ The gradient based heuristic for finding a rank $r$ constrained solution to \eqref{main_prob} is shown in Figure. \ref{fig:rank_gradient}. For \eqref{main_prob_gen}, a similar gradient based heuristic is shown in Figure. \ref{fig:rank_gradient_gen}. It has been noted that constraining both $F$ and $G$ matrices in \ref{fig:rank_gradient_gen} leads to a rank $r$ solution, although by the semidefinite-embedding theorem (see \cite{fazel2003log}), the rank can be lower than $r$. In both these heuristics, a solution is said to be reached if the smallest $(N-r)$ singular values are lesser than a tolerance, say $10^{-8}$. Similarly, the heuristic based on the bilinear formulation for the problem in \eqref{main_prob} is presented in Figure. \ref{fig:rank_bilinear}. The calculation of SVD before the update step has been observed to improve the convergence rates. Moreover, it has also been observed that if one wants to find a rank $r$ solution and if one starts with an initial guess which has a rank $r+1$, the convergence is quicker. Similarly, the bilinear formulation based heuristic for solving \eqref{main_prob_gen} is presented in Figure. \ref{fig:rank_bilinear_gen}.\\\\ Note that the aforementioned heuristics can also be used with linear matrix inequalities (in this case, $A\mbox{\textbf{vec}}(\Delta J)=0$ is replaced by $A\mbox{\textbf{vec}}(\Delta J) \geq A\mbox{\textbf{vec}}(X J)-b$). Once these heuristics converge onto a matrix $X$ which has $N-r$ eigenvalues lesser than a tolerance, say $10^{-8}$, then one can perform the following polishing step: compute the SVD of $X = USU^\top$ and then use $U(:,1:r)Z^*$, where $Z^*$ solves \begin{equation} \begin{array}{l} \displaystyle \min_{V \in \mathcal{R}^{r\times N}} ||A\mbox{\textbf{vec}}(U(:,1:r)Z) - b||_2\\ \mbox{subject to}\\ \hspace*{2cm} U(:,1:r)Z \succeq 0. \end{array} \label{polish_step} \end{equation} It has also been observed that the eigenvectors of the matrix $Q$ in \eqref{lrf_def}, when reshaped into a square matrix are either symmetric or skew-symmetric. For $Q_r$, suppose it that the eigenvectors are divided into $r$ consecutive vectors. Then, it has been observed that each of these parts reshapes either into a symmetric matrix or skew-symmetric matrix. These observations does not have a theoretical basis yet and further properties need to be explored. However, for the matrix $Q$ as defined in \eqref{lrf_def}, the following properties can also be shown easily: \begin{itemize} \item For any two PSD matrices $X$ and $Y$ of dimension $N\times N$, $X(:)^\top QY(:)\geq 0$. \item For any two PSD matrices $X$ and $Y$ of dimension $N\times N$, \newline $(X(:)+Y(:))^\top Q(X(:)+Y(:))\geq X(:)^\top QX(:) + Y(:)^\top QY(:)$. \end{itemize} \begin{figure}[t] \centering \fbox{\begin{minipage}{40em} Given $A \in \mathcal{R}^{m\times N^2}$ and $b\in \mathcal{R}^{m\times 1}$, define $Q_r\in \mbox{Sym}^{N^2\times r}$ as shown in \eqref{eqn:part_rank_def}. \\\\ Find a feasible solution $X^0$ such that $X^0\succeq 0$ and $AX_0(:)=b$. Let $X^{(k)}=\frac{1}{r}X^0$, for all integers $k$ such that $1\leq k \leq r$. Let $J = \textbf{1}^{r\times 1}\bigotimes I_N$. Let $X = [X^{(1)},\cdots,X^{(r)}]$ and $\Delta = [\Delta^{(1)},\cdots,\Delta^{(r)}]$. Set Max-Iteration-Count to a moderately large positive integer, say 100.\\\\ \begin{equation} \begin{array}{l} \hspace*{1cm}\mbox{count} = 1\\ \hspace*{1cm}\textbf{while} (\mbox{count} \leq \mbox{Max-Iteration-Count}) \\ \hspace*{2cm}\displaystyle \max_{\tiny \Delta^{(k)} \in \mbox{Sym}^N, ~\forall 1\leq k\leq r}~~~~ \left(\Delta(:)\right)^\top Q_rX(:) \\ \hspace*{2cm}\displaystyle \mbox{subject~to}\\ \hspace*{4cm}\displaystyle X^{(k)}-\Delta^{(k)} \succeq 0, ~\forall~ 1\leq k\leq r \\ \hspace*{4cm}\displaystyle A\mbox{\textbf{vec}}\left(\Delta J\right) = 0\\\\ \hspace*{2cm} t = \min\left(1, \frac{\left(\Delta(:)\right)^\top Q_rX(:)}{\left(X(:)\right)^\top Q_rX(:)}\right)\\ \hspace*{2cm} X = X - t\Delta\\ \hspace*{2cm}\mbox{count} = \mbox{count} + 1 \\ \hspace*{1cm}\textbf{end} \end{array} \label{rank_one_gradient} \end{equation} If the smallest $(N-r)$ eigenvalues of $X$ are all smaller than, say, $10^{-8}$, then consider this a rank $r$ solution satisfying the linear matrix equations. \end{minipage}} \caption{Heuristic based on gradient descent for finding a rank $r$ solution to \eqref{main_prob}.} \label{fig:rank_gradient} \end{figure} \begin{figure}[t] \centering \fbox{\begin{minipage}{40em} Given $A \in \mathcal{R}^{m\times NM}$ and $b\in \mathcal{R}^{m\times 1}$, define $Q^F_r\in \mbox{Sym}^{N^2\times r}$ and $Q^G_r\in \mbox{Sym}^{M^2\times r}$ as in \eqref{eqn:part_rank_def}. \\\\ Find $X^0$, $F^0$ and $G^0$ such that $\begin{bmatrix}F^0 & X^0\\\left(X^0\right)^\top & G^0\end{bmatrix}\succeq 0$ and $AX^0(:)=b$. Let $X=X^0$, $F^{(k)}=F^0/r$ and $G^{(k)}=G^0/r$ $\forall~ 1\leq k\leq r$. Let $F = [F^{(1)},\cdots,F^{(r)}]$ and $G = [G^{(1)},\cdots,G^{(r)}]$. Let $J_F = \textbf{1}^{r\times 1}\bigotimes I_N$ and $J_G = \textbf{1}^{r\times 1}\bigotimes I_M$. Let $Z = \begin{bmatrix}FJ_F & X\\X^\top & GJ_G\end{bmatrix}$. Let $\Delta_X\in \mathcal{R}^{N\times M}$. Let $\Delta_F^{(k)}\in \mathcal{R}^{N\times N}$ and $\Delta_G^{(k)}\in \mathcal{R}^{M\times M}$ $\forall~ 1\leq k\leq r$. Let $\Delta_F = [\Delta_F^{(1)},\cdots,\Delta_F^{(r)}]$ and $\Delta_G = [\Delta_G^{(1)},\cdots,\Delta_G^{(r)}]$. Suppose $\Delta = \begin{bmatrix}\Delta_FJ_F & \Delta_X\\\Delta_X^\top & \Delta_GJ_G\end{bmatrix}$. Set Max-Iteration-Count to a moderately large positive integer, say 100.\\\\ \begin{equation} \begin{array}{l} \hspace*{1cm}\mbox{count} = 1\\ \hspace*{1cm}\textbf{while} (\mbox{count} \leq \mbox{Max-Iteration-Count}) \\ \hspace*{2cm}\displaystyle \max_{\tiny\substack{\Delta_F^{(k)} \in \mbox{Sym}^N, ~\forall 1\leq k\leq r\\\Delta_G^{(k)} \in \mbox{Sym}^M, ~\forall 1\leq k\leq r}}~~ \left(\Delta_F(:)\right)^\top Q^F_rF(:) + \left(\Delta_G(:)\right)^\top Q^G_rG(:) \\ \hspace*{2cm}\displaystyle \mbox{subject~to}\\ \hspace*{4cm}\displaystyle Z-\Delta \succeq 0\\ \hspace*{4cm}\displaystyle F^{(k)}-\Delta^{(k)}_F \succeq 0, ~\forall~ 1\leq k\leq r\\ \hspace*{4cm}\displaystyle G^{(k)}-\Delta^{(k)}_G \succeq 0, ~\forall~ 1\leq k\leq r\\ \hspace*{4cm}\displaystyle A\Delta_X(:) = 0\\\\ \hspace*{2cm} t = \min\left(1, \frac{\left(\Delta_F(:)\right)^\top Q^F_rF(:)}{\left(F(:)\right)^\top Q^F_rF(:)},\frac{\left(\Delta_G(:)\right)^\top Q^G_rG(:)}{\left(G(:)\right)^\top Q^G_rG(:)}\right)\\ \hspace*{2cm} F = F - t\Delta_F\\ \hspace*{2cm} G = G - t\Delta_G\\ \hspace*{2cm} X = X - t\Delta_X\\ \hspace*{2cm}\mbox{count} = \mbox{count} + 1 \\ \hspace*{1cm}\textbf{end} \end{array} \label{rank_one_gradient_general} \end{equation} If the smallest $(\min(N,M)-r)$ singular values of $X$ are all smaller than, say, $10^{-8}$, then consider this a rank $r$ solution satisfying the linear matrix equations. \end{minipage}} \caption{Heuristic based on gradient descent for finding a rank $r$ constrained solution to \eqref{main_prob_gen}.} \label{fig:rank_gradient_gen} \end{figure} \begin{figure}[t] \centering \fbox{\begin{minipage}{40em} Given $A \in \mathcal{R}^{m\times N^2}$ and $b\in \mathcal{R}^{m\times 1}$, define $Q_r\in \mbox{Sym}^{N^2}$ as shown in \eqref{eqn:part_rank_def}. \\\\ Find a feasible solution $X^0$ such that $X^0\succeq 0$ and $AX^0(:)=b$. Set Max-Iteration-Count to moderately large positive integer, say 100.\\\\ Compute the SVD of $X^0$, \textit{i.e.}, $X^0 = USU^\top$. Define $X^{(k)} = U(:,k)S_{k,k}U^\top(:,k)$ and let $X = [X^{(1)},\cdots,X^{(r)}]$. Let $J = \textbf{1}^{r\times 1}\bigotimes I_N$.\\\\ \begin{equation} \begin{array}{l} \hspace*{1cm}\mbox{count} = 1\\ \hspace*{1cm}\textbf{while} (\mbox{count} \leq \mbox{Max-Iteration-Count}) \\ \hspace*{2cm}\displaystyle \min_{Y^{(1)}, \cdots,~Y^{(r)} \in \mathcal{R}^{N\times N}}~~~~ Y(:)^\top Q_r X(:)\\ \hspace*{2cm}\displaystyle \mbox{subject~to}\\ \hspace*{3cm}\displaystyle Y = [Y^{(1)},\cdots,Y^{(r)}]\\ \hspace*{3cm}Y^{(k)}\succeq 0;~~1\leq k\leq r\\ \hspace*{3cm}\displaystyle A\mbox{\textbf{vec}}(YJ) = b\\\\ \hspace*{2cm}\mbox{Compute the SVD of}~YJ\mbox{~, that is,~} YJ = USU^\top\\ \hspace*{2cm}X^{(k)} = U(:,k)S_{k,k}U^\top(:,k)\\ \hspace*{2cm}\mbox{count} = \mbox{count} + 1 \\ \hspace*{1cm}\textbf{end} \end{array} \label{all_rank_hill_climbing} \end{equation} If the smallest $(N-r)$ eigenvalues of $XJ$ are all smaller than, say, $10^{-8}$, then consider this a rank $r$ solution satisfying the linear matrix equations. \end{minipage}} \caption{Heuristic based on a bilinear formulation for finding a rank $r$ constrained solution to \eqref{main_prob}.} \label{fig:rank_bilinear} \end{figure} \begin{figure}[t] \centering \fbox{\begin{minipage}{40em} Given $A \in \mathcal{R}^{m\times NM}$ and $b\in \mathcal{R}^{m\times 1}$, define $Q^F_r\in \mbox{Sym}^{N^2}$ and $Q^G_r\in \mbox{Sym}^{M^2}$ as in \eqref{eqn:part_rank_def}. \\\\ Find $X^0$, $F^0$ and $G^0$ such that $\begin{bmatrix}F^0 & X^0\\\left(X^0\right)^\top & G^0\end{bmatrix}\succeq 0$ and $AX^0(:)=b$. Set Max-Iteration-Count to moderately large positive integer, say 100.\\\\ Compute the SVD of $\begin{bmatrix}F^0 & X^0\\\left(X^0\right)^\top & G^0\end{bmatrix}$, that is, $\begin{bmatrix}F^0 & X^0\\\left(X^0\right)^\top & G^0\end{bmatrix} = USU^\top$. Define $F^{(k)} = U(1:N,k)S_{k,k}U^\top(1:N,k)$ and $G^{(k)} = U((N+1):(N+M),k)S_{k,k}U^\top((N+1):(N+M),k)$. Let $F = [F^{(1)},\cdots,F^{(r)}]$ and $G = [G^{(1)},\cdots,G^{(r)}]$. Let $J_Y = \textbf{1}^{r\times 1}\bigotimes I_N$ and $J_Z = \textbf{1}^{r\times 1}\bigotimes I_M$.\\\\ \begin{equation} \begin{array}{l} \hspace*{1cm}\mbox{count} = 1\\ \hspace*{1cm}\textbf{while} (\mbox{count} \leq \mbox{Max-Iteration-Count}) \\ \hspace*{2cm}\displaystyle \min_{\substack{X\in \mathcal{R}^{N\times M},\\ Y^{(1)}, \cdots,~Y^{(r)} \in \mathcal{R}^{N\times N},\\Z^{(1)}, \cdots,~Z^{(r)} \in \mathcal{R}^{M\times M}}}~~ F(:)^\top Q^F_rY(:) + G(:)^\top Q^F_rZ(:)\\ \hspace*{2cm}\displaystyle \mbox{subject~to}\\ \hspace*{4cm}\displaystyle Y = [Y^{(1)},\cdots,Y^{(r)}],~~Z = [Z^{(1)},\cdots,Z^{(r)}]\\ \hspace*{4cm}Y^{(k)}\succeq 0,~~ Z^{(k)}\succeq 0;~\forall~1\leq k\leq r\\ \hspace*{4cm}\displaystyle AX(:) = b\\ \hspace*{4cm}\begin{bmatrix}YJ_Y & X\\ X^\top & ZJ_Z\end{bmatrix} \succeq 0\\ \hspace*{2cm}\mbox{Compute the SVD of}~\begin{bmatrix}F & X\\ X^\top & G\end{bmatrix}\mbox{~, that is,~} \begin{bmatrix}F & X\\X^\top & G\end{bmatrix} = USU^\top\\ \hspace*{2cm}F^{(k)} = U(1:N,k)S_{k,k}U^\top(1:N,k),~\forall~1\leq k\leq r\\ \hspace*{2cm} G^{(k)} = U((N+1):(N+M),k)S_{k,k}U^\top((N+1):(N+M),k), ~\forall~1\leq k\leq r\\ \hspace*{2cm}\mbox{count} = \mbox{count} + 1 \\ \hspace*{1cm}\textbf{end}\\\\ \end{array} \label{all_rank_general_hill_climbing} \end{equation} If the smallest $(\min(N,M)-r)$ singular values of $X$ are all smaller than, say, $10^{-8}$, then consider this a rank $r$ solution satisfying the linear matrix equations. \end{minipage}} \caption{Heuristic based on a bilinear formulation for finding a rank $r$ constrained solution to \eqref{main_prob_gen}.} \label{fig:rank_bilinear_gen} \end{figure} \section{Simulations} In this section, the following specific cases are considered and simulation examples are presented explicitly, that is, (i) finding rank constrained solution to problems of the form given in \eqref{main_prob} and \eqref{main_prob_gen}, and a comparison with the log-det heurisitic and nuclear norm heuristic, (ii) application to Knapsack decision problem, (iii) solving a class of non-linear equations and (iv) application to Fourier Phase Retrieval with amplitude constraints. In additions, two cases where the heuristic fails have also been discussed. Having said this, these cases certainly do not span the entire range of problems which can be tackled with this approach, neither do the cases presented here promise the effectiveness of the proposed method in all scenarios. The author would urge the interested reader to explore the strengths and limitations of the proposed methods. \subsection{Rank $r$ constrained solutions} Here, the problem of finding rank $r$ solutions to LME is considered, with and without the positive semidefinite constraint. For these cases, the heuristics based on the bilinear formulations are used. Note that the heuristic is followed by the polishing step given in \eqref{polish_step}. Two simulation results have been given in the files "PSD\_sim.txt" and "GEN\_sim.txt" uploaded with this document. In both simulations, the matrix $A$, the matrix which was used to generate $b$ (just to ensure that $AX(:)=b$ has at least one solution), the vector $b$ are given at the beginning. In "PSD\_sim.txt" this is followed by the initial guess to start the bi-linear heuristic (here $=0$ implies a zero matrix), results for every rank (and the solution obtained after polishing step), the singular values of the solutions and the norm of $AX(:)=b$. Finally, the results obtained using the trace minimzation and log-det heuristic are also shown. The content of the "GEN\_sim.txt" are similar, except for the fact that the diagonal elements are constrained to 100 here, and that the initial guess for finding rank $r$ solution is the rank $r+1$ solution, if it is found using the proposed heuristic. \\\\ One can note that in "PSD\_sim.txt", solution corresponding to all ranks are obtained using the proposed heuristic, where as the solution obtained using the trace minimization method and log-det heuristic have a rank of 3. On the other hand, one can see that "GEN\_sim.txt" shows solutions corresponding to every rank except unity rank are obtained using the proposed heuristic, where as the nuclear norm minimization leads to a full rank solution. \subsection{Knapsack Problem} The classical problem of Knapsack asks the following question: given a set of objects, each with a weight and a value, what is the subset which yields the maximum sum total value subject to an upper bound on the total weight. The Knapsack decision problem is a minor modification: given a lower bound on the sum total value and an upper bound on the weight, does there exist a subset which satisfies both the constraints. As mentioned earlier, in terms of a rank constrained problem, it can be posed as: \begin{equation} \begin{array}{l} \mbox{Find~} X \in \mathcal{R}^{(N+1)\times (N+1)} \mbox{~with unity rank~}\\ \mbox{subject to~}\\ \hspace*{1cm} X_{1,1} = 1, ~~X \succeq 0,\\ \hspace*{1cm} \mbox{diag}(X) = X_{(:,1)},\\ \hspace*{1cm} \textbf{v}^\top X_{(2:(N+1),1)} \geq V,\\ \hspace*{1cm} \textbf{w}^\top X_{(2:(N+1),1)} \leq W, \end{array} \end{equation} where $\textbf{v}$ and $\textbf{w}$ are the vectors of values and weights respectively, whereas $V$ and $W$ denote the lower bound on total value and upper bound on total weight, respectively. A simulation instance for this is provided.\\\\ A simulation instance has been provided in the file "Knapsack.txt" provided with this document. In the beginning, it shows that weigths and the values vectors, the minimum total value desired and the maximum allowed total weight and the initial guess for the bilinear heuristic to be applied to this problem. This is followed by the matrix obtained using the proposed heuristic, its singular values, the result for the knapsack problem and the total weight and value of the objects chosen. In addition, it can be seen that the matrix obtained using the trace heuristic has a rank of 15. However, the solution obtained using the log-det heuristic has unity rank, and thus this qualifies as a valid solution. \subsection{Non-linear Equations} In this subsection, a solution to a class of non-linear equations (mentioned earlier) is obtained using the proposed method. The problem considered here is: \begin{eqnarray} \mbox{Find}~~ &[\theta_1,~\theta_2,\cdots,~\theta_N] \in [-\pi,\pi]^{N\times 1}\nonumber \\ \mbox{such~that}~~ & Ax = b, \mbox{~~where~} x = \left[e^{\mbox{j}\theta_1},~e^{\mbox{j}\theta_2},\cdots,~e^{\mbox{j}\theta_N}\right]^\top. \label{nl_prob} \end{eqnarray} This can be converted to a rank constrained feasibility problem given by: \begin{eqnarray} \mbox{Find}~~ &X \in \mathcal{R}^{N\times N}\nonumber \\ \mbox{such~that}~~ &\mathcal{A}X(\small{2:N,1}) = b, ~\mbox{diag}(X) = \textbf{1}, ~X \succeq 0 \mbox{~~and~~} \mbox{Rank} (X) = 1. \label{combi_prob_conversion} \end{eqnarray} A simulation instance for this case is provided in the file "Nonlinear\_Equations.txt" available with this document. In the beginning, it comprises of the matrix $A$, a vector of unit magnitude complex numbers used to generate $b$ (just to ensure that $AX(2:N,1)=b$ has at least one solution) and the initial guess for the proposed bilinear heuristic. This is followed by the solution obtained using the proposed heuristic, its singular values and the solution. Finally, it can be seen that the trace minimization leads to a a full rank solution and log-det heuristic leads to a rank 2 solution. \subsection{Fourier Phase Retrieval with Amplitude Constraints} Fourier Phase Retrieval is a classical problem in signal processing, which asks the following: given the magnitude spectrum of a signal (discrete time/frequency), can one find the time domain signal? As the problem as such is ill-posed (has infinitely many solutions, each corresponding to a phase spectrum one chooses), one looks for conditions which renders the solution unique. Here, amplitude constraint are chosen. These constraints need not necessarily lead to a unique solution, but this example serves as a good application for the proposed method. Mathematically, the Fourier Phase Retrieval problem with amplitude constraints can be cast as the following: \begin{equation} \begin{array}{l} \mbox{Find~} X \in \mathcal{R}^{(N+1)\times (N+1)} \mbox{~with unity rank~}\\ \mbox{subject to~}\\ \hspace*{1cm} X_{1,1} = 1, ~~X \succeq 0,\\ \hspace*{1cm} \mbox{diag}(DX(2:(N+1),2:(N+1))D^{\mbox{H}}) = z,\\ \hspace*{1cm} \mbox{diag}(X(2:(N+1),2:(N+1))) \leq A, \end{array} \end{equation} where $D$ is the DFT matrix of dimension $N\times N$, $z$ is the vector of the magnitudes at different frequencies and $A$ is the amplitude constraint on the time domain signal.\\\\ The simulation instance for this problem has been provided in the file "Phase\_Retrieval.txt" available with this document. It shows the given magnitude squared spectrum (a vector) in the beginning, followed by the amplitude constraint desired. The initial guess used for the bilinear heuristic is shown next. Then the matrix obtained using the proposed heuristic is given, its singular values are shown and the final signal adhering to the magnitude and amplitude constraints is provided. The matrix obtained using trace minimization has full rank and that obtained using log-det has a rank of 4. Hence, both of these do not qualify as a valid solution the Fourier Phase Retrieval problem instance. \subsection{Shortcomings of the proposed method} Having mentioned the applications where the proposed method seems to work satisfactorily, it has also been observed that there are applications where the method seems to fail to find a unity rank solution. Two such cases are: (i) the subset sum problem and (ii) the linear complementarity problem in the general symmetric case. However, in both these cases, it has also been observed that the heuristic converges to a rank 2 solution. \begin{itemize} \item {The subset sum problem asks the following question: given a finite set of integers, does there exist a subset whose elements add up to a desired value? Mathematically, the problem can be posed as: \begin{equation} \begin{array}{l} \mbox{Find~} X \in \mathcal{R}^{(N+1)\times (N+1)} \mbox{~with unity rank~}\\ \mbox{subject to~}\\ \hspace*{1cm} X_{1,1} = 1, ~~X \succeq 0,\\ \hspace*{1cm} \mbox{diag}(X) = X(:,1),\\ \hspace*{1cm} \textbf{s}^\top X(2:(N+1),1) = D, \end{array} \end{equation} where $\textbf{s}$ is the vector of values of the elements in the finite set and $D$ is the desired value. } \item{The linear complementarity problem in the general symmetric case is the following problem: given a matrix $M\in \mbox{Sym}^N$ and $q\in \mathcal{R}^{N\times 1}$, find non-negative vectors $w\in \mathcal{R}^{N\times 1}$ and $z\in \mathcal{R}^{N\times 1}$ such that $w^\top z = 0$ and $w = Mz + q$. Mathematically, it can be posed as the following feasibility problem: \begin{equation} \begin{array}{l} \mbox{Find~} X \in \mathcal{R}^{(2N+1)\times (2N+1)} \mbox{~with unity rank~}\\ \mbox{subject to~}\\ \hspace*{1cm} X_{1,1} = 1, ~~X \succeq 0,\\ \hspace*{1cm} \textbf{1}^\top \mbox{diag}(X_{(2:(N+1),(N+2):(2N+1))}) = 0,\\ \hspace*{1cm} X_{(2:(N+1),1)} = MX_{((N+2):(2N+1),1)} + q. \end{array} \end{equation} } \end{itemize} \section{Conclusions and Acknowledgements} The author believes that the method outlined in this paper will pave the way for further research in the domain of rank constrained optimization. The author would also like to extend an invitation to all the interested readers to enrich the proposed method with theoretical guarantees. \\\\ The author wishes to thank Dr. Bharath Bhikkaji of \textit{IIT Madras}, Dr. Arun Ayyar of \textit{Santa Fe Research}, Dr. Girish Ganesan and Dr. Kumar Vijay Mishra for all their support. \bibliographystyle{alpha}

05b3fcc26dbad297ba6a8bd5bf88c1419f76ba8a

\section{Introduction} \emph{Clustering} is a fundamental problem in machine learning and data science. A general clustering task is to partition the given datapoints such that points inside the same cluster are ``similar'' to each other. More formally, consider a set of datapoints $\mathcal{C}$ and a set of ``potential cluster centers'' $\mathcal{F}$, with a metric $d$ on $\mathcal{C} \cup \mathcal{F}$. We define $n := |\mathcal{C} \cup \mathcal{F}|$. Given any set $\S \subseteq \mathcal{F}$, each $j \in \mathcal{C}$ is associated with the key statistic $d(j, \S) = \min_{i \in \S} d(i, j)$. The typical clustering task is to select a set $\S \subseteq \mathcal{F}$ which has a small size and which minimizes the values of $d(j, \S)$. The size of the set $\S$ is often fixed to a value $k$, and we typically ``boil down'' the large collection of values $d(j, \S)$ into a single overall objective function. A variety of objective functions and assumptions on sets $\mathcal{C}$ and $\mathcal{F}$ are used. The most popular problems include\footnote{In the original version of the $k$-means problem, $\mathcal{C}$ is a subset of $\mathbb{R}^\ell$ and $\mathcal{F} = \mathbb{R}^\ell$ and $d$ is the Euclidean metric. By standard discretization techniques (see, e.g., \cite{matousek, feldman}), the set $\mathcal{F}$ can be reduced to a polynomially-bounded set with only a small loss in the value of the objective function.} \begin{itemize} \item the $k$-center problem: minimize the value $\max_{j \in \mathcal{C}} d(j, \S)$ given that $\mathcal{F} = \mathcal{C}$. \item the $k$-supplier problem: minimize the value $\max_{j \in \mathcal{C}} d(j, \S)$ (where $\mathcal{F}$ and $\mathcal{C}$ may be unrelated); \item the $k$-median problem: minimize the summed value $\sum_{j \in \mathcal{C}} d(j, \S)$; and \item the $k$-means problem: minimize the summed square value $\sum_{j \in \mathcal{C}} d(j, \S)^2$. \end{itemize} An important special case is when $\mathcal{C} = \mathcal{F}$ (e.g. the $k$-center problem); since this often occurs in the context of data clustering, we refer to this as the \textbf{self-contained clustering (SCC)} setting. These classic NP-hard problems have been studied intensively for the past few decades. There is an alternative interpretation from the viewpoint of operations research: the sets $\mathcal{F}$ and $\mathcal{C}$ can be thought of as ``facilities'' and ``clients'', respectively. We say that $i \in \mathcal{F}$ is \emph{open} if $i$ is placed into the solution set $\S$. For a set $\S \subseteq \mathcal{F}$ of open facilities, $d(j, \S)$ can then be interpreted as the connection cost of client $j$. This terminology has historically been used for clustering problems, and we adopt it throughout for consistency. However, our focus is on the case in which $\mathcal{C}$ and $\mathcal{F}$ are arbitrary abstract sets in the data-clustering setting. Since these problems are NP-hard, much effort has been paid on algorithms with ``small'' provable \emph{approximation ratios/guarantees}: i.e., polynomial-time algorithms that produce solutions of cost at most $\alpha$ times the optimal. The current best-known approximation ratio for $k$-median is $2.675$ by Byrka et.\ al.\ \cite{byrka} and it is NP-hard to approximate this problem to within a factor of $1+2/e \approx 1.735$ \cite{jms}. The recent breakthrough by Ahmadian et.\ al.\ \cite{svensson} gives a $6.357$-approximation algorithm for $k$-means, improving on the previous approximation guarantee of $9+\epsilon$ based on local search \cite{kanungo}. Finally, the $k$-supplier problem is ``easier'' than both $k$-median and $k$-means in the sense that a simple $3$-approximation algorithm \cite{hochbaum} is known, as is a $2$-approximation for $k$-center problem: we cannot do better than these approximation ratios unless $\text{P} = \text{NP}$ \cite{hochbaum}. While optimal approximation algorithms for the center-type problems are well-known, one can easily demonstrate instances where such algorithms return a worst-possible solution: (i) all clusters have the same worst-possible radius ($2T$ for $k$-center and $3T$ for $k$-supplier where $T$ is the optimal radius) and (ii) almost all data points are on the circumference of the resulting clusters. Although it is NP-hard to improve these approximation ratios, our new randomized algorithms provide significantly better ``per-point'' guarantees. For example, we achieve a new ``per-point'' guarantee $\ensuremath{\mathbf{E}}[d(j,\S)] \leq (1+2/e)T \approx 1.736T$, while respecting the usual guarantee $d(j,\S) \leq 3T$ with probability one. \emph{Thus, while maintaining good global quality with probability one, we also provide superior stochastic guarantees for each user.} The general problem we study in this paper is to develop approximation algorithms for center-type problems where $\S$ is drawn from a probability distribution over $k$-element subsets of $\mathcal{F}$; we refer to these as \emph{$k$-lotteries.} We aim to construct a $k$-lottery $\Omega$ achieving certain guarantees on the distributional properties of $d(j,\S)$. The classical $k$-center problem can be viewed as the special case where the distribution $\Omega$ is deterministic, that is, it is supported on a single point. Our goal is to find an \emph{approximating distribution} $\tilde \Omega$ which matches the \emph{target distribution} $\Omega$ as closely as possible for each client $j$. Stochastic solutions can circumvent the approximation hardness of a number of classical center-type problems. There are a number of additional applications where stochasticity can be beneficial. We summarize three here: smoothing the integrality constraints of clustering, solving repeated problem instances, and achieving fair solutions. \smallskip \noindent \textbf{Stochasticity as interpolation.} In practice, \emph{robustness} of the solution is often more important than achieving the absolute optimal value for the objective function. One potential problem with the (deterministic) center measure is that it can be highly non-robust. As an extreme example, consider $k$-center with $k$ points, each at distance $1$ from each other. This clearly has value $0$ (choosing $\S = \mathcal{C}$). However, if a single new point at distance $1$ to all other points is added, then the solution jumps to $1$. Stochasticity alleviates this discontinuity: by choosing $k$ facilities uniformly at random among the full set of $k+1$, we can ensure that $\ensuremath{\mathbf{E}}[d(j,\S)] = \frac{1}{k+1}$ for every point $j$, a much smoother transition. \smallskip \noindent \textbf{Repeated clustering problems.} Consider clustering problems where the choice of $\mathcal{S}$ can be changed periodically: e.g., $\mathcal{S}$ could be the set of $k$ locations in the cloud chosen by a service-provider. This set $\mathcal{S}$ can be shuffled periodically in a manner transparent to end-users. For any user $j \in \mathcal{C}$, the statistic $d(j,\S)$ represents the latency of the service $j$ receives (from its closest service-point in $\mathcal{S}$). If we aim for a fair or minmax service allocation, then our $k$-center stochastic approximation results ensure that, with high probability, \emph{every client} $j$ gets long-term average service of at most around $1.736 T$. The average here is taken over the periodic re-provisioning of $\mathcal{S}$. (Furthermore, we have the risk-avoidance guarantee that in no individual provisioning of $\mathcal{S}$ will any client have service greater than $3T$.) \smallskip \noindent \textbf{Fairness in clustering.} The classical clustering problems combine the needs of many different points (elements of $\mathcal{C}$) into one metric. However, clustering (and indeed many other ML problems) are increasingly driven by inputs from parties with diverse interests. Fairness in these contexts has taken on greater importance in the current environment where decisions are increasingly made by algorithms and machine learning. Some examples of recent concerns include the accusations of, and fixes for, possible racial bias in Airbnb rentals \cite{wpost:Airbnb} and the finding that setting the gender to ``female" in Google's \emph{Ad Settings} resulted in getting fewer ads for high-paying jobs \cite{ad-privacy}. Starting with older work such as \cite{doi:10.1056/NEJM199902253400806}, there have been highly-publicized works on bias in allocating scarce resources -- e.g., racial discrimination in hiring applicants who have very similar resum\'{e}s \cite{bertrand-mullainathan}. Additional work discusses the possibility of bias in electronic marketplaces, whether human-mediated or not \cite{RAND:RAND12115,wpost:Airbnb}. A fair allocation should provide good service guarantees \emph{to each user individually}. In data clustering settings where a user corresponds to a datapoint, this means that every point $j \in \mathcal{C}$ should be guaranteed a good value of $d(j, \S)$. This is essentially the goal of $k$-center type problems, but the stochastic setting broadens the meaning of good per-user service. Consider the following scenarios. Each user, either explicitly or implicitly, submits their data (corresponding to a point in $\mathcal{C}$) to an aggregator such as an e-commerce site. A small number $k$ of users are then chosen as ``influencer'' nodes; for instance, the aggregator may give them a free product sample to influence the whole population in aggregate, as in \cite{DBLP:journals/toc/KempeKT15}, or the aggregator may use them as a sparse ``sketch'', so that each user gets relevant recommendations from a influencer which is similar to them. Each point $j$ would like to be in a cluster that is ``high quality'' \emph{from its perspective}, with $d(j, \S)$ being a good proxy for such quality. Indeed, there is increasing emphasis on the fact that organizations monetize their user data, and that users need to be compensated for this (see, e.g., \cite{lanier-book,ibarra-etal:data}). This is a transition from viewing data as capital to viewing \emph{data as labor}. A concrete way for users (i.e., the data points $j \in \mathcal{C}$) to be compensated in our context is for each user to get a guarantee on their solution quality: i.e., bounds on $d(j, \S)$. \subsection{Our contributions and overview} In Section~\ref{chance-sec}, we encounter the first clustering problem which we refer to as \emph{chance $k$-coverage}: namely, every client $j$ has a distance demand $r_j$ and probability demand $p_j$, and we wish to find a distribution satisfying $\Pr[d(j,\S) \leq r_j] \geq p_j$. We show how to obtain an approximation algorithm to find an approximating distribution $\tilde \Omega$ with\footnote{Notation such as ``$S \sim \tilde \Omega$" indicates that the random set $S$ is drawn from the distribution $\tilde \Omega$.} $$ \Pr_{\S \sim \tilde \Omega}[d(j, \S) \leq 9 r_j] \geq p_j. $$ In a number of special cases, such as when all the values of $p_j$ or $r_j$ are the same, the distance factor $9$ can be improved to $3$, \emph{which is optimal}; it is an interesting question to determine whether this factor can also be improved in the general case. In Section~\ref{sec2}, we consider a special case of chance $k$-coverage, in which $p_j = 1$ for all clients $j$. This is equivalent to the classical (deterministic) $k$-supplier problem. Allowing the approximating distribution $\tilde \Omega$ to be stochastic yields significantly better distance guarantees than are possible for $k$-supplier or $k$-center. For instance, we find an approximating distribution $\tilde \Omega$ with $$ \forall j \in \mathcal C \qquad \ensuremath{\mathbf{E}}_{\S \sim \tilde \Omega}[d(j, \S)] \leq 1.592 T ~\mbox{and}~ \Pr[d(j, \S) \leq 3T] = 1 $$ where $T$ is the optimal solution to the (deterministic) $k$-center problem. By contrast, deterministic polynomial-time algorithms cannot guarantee $d(j, \S) < 2 T$ for all $j$, unless $\text{P} = \text{NP}$ \cite{hochbaum}. In Section~\ref{lb-sec}, we show a variety of lower bounds on the approximation factors achievable by efficient algorithms (assuming $\text{P} \neq \text{NP}$). For instance, we show that our approximation algorithm for chance $k$-coverage with equal $p_j$ or $r_j$ has the optimal distance approximation factor 3, that our approximation algorithm for $k$-supplier has optimal approximation factor $1 + 2/e$, and that the approximation factor $1.592$ for $k$-center cannot be improved below $1 + 1/e$. In Section~\ref{approx-alg}, we consider a different type of stochastic approximation problem based on expected distances: namely, every client has a demand $t_j$, and we seek a $k$-lottery $\Omega$ with $\ensuremath{\mathbf{E}}[d(j,\S)] \leq t_j$. We show that we can leverage \emph{any given} $\alpha$-approximation algorithm for $k$-median to produce a $k$-lottery $\tilde \Omega$ with $\ensuremath{\mathbf{E}}[d(j,\S)] \leq \alpha t_j$. (Recall that the current-best $\alpha$ here is $2.675$ as shown in \cite{byrka}.) In Section~\ref{determinism-sec}, we consider the converse problem to Section~\ref{sec2}: if we are given a $k$-lottery $\Omega$ with $\ensuremath{\mathbf{E}}[d(j,\S)] \leq t_j$, can we produce a single deterministic set $\S$ so that $d(j, \S) \approx t_j$ and $|\S| \approx k$? We refer to this as a \emph{determinization} of $\Omega$. We show a variety of determinization algorithms. For instance, we are able to find a set $\S$ with $|\S| \leq 3 k$ and $d(j, \S) \leq 3 t_j$. We also show a number of nearly-matching lower bounds. \subsection{Related work} With algorithms increasingly running our world, there has been substantial recent interest on incorporating fairness systematically into algorithms and machine learning. One important notion is \emph{disparate impact}: in addition to requiring that \emph{protected attributes} such as gender or race not be used (explicitly) in decisions, this asks that decisions not be disproportionately different for diverse protected classes \cite{DBLP:conf/kdd/FeldmanFMSV15}. This is developed further in the context of clustering in the work of \cite{DBLP:conf/nips/Chierichetti0LV17}. Such notions of \emph{group fairness} are considered along with \emph{individual fairness} -- treating similar individuals similarly -- in \cite{DBLP:conf/icml/ZemelWSPD13}. See \cite{DBLP:conf/innovations/DworkHPRZ12} for earlier work that developed foundations and connections for several such notions of fairness. In the context of the location and sizing of services, there have been several studies indicating that proactive on-site provision of healthcare improves health outcomes significantly: e.g., mobile mammography for older women \cite{reuben2002randomized}, mobile healthcare for reproductive health in immigrant women \cite{guruge2010immigrant}, and the use of a community mobile-health van for increased access to prenatal care \cite{edgerley2007use}. Studies also indicate the impact of distance to the closest facility on health outcomes: see, e.g., \cite{mccarthy2007veterans,mooney2000travel,schmitt2003influence}. Such works naturally suggest tradeoffs between resource allocation (provision of such services, including sizing -- e.g., the number $k$ of centers) and expected health outcomes. While much analysis for facility-location problems has focused on the static case, other works have examined a similar lottery model for center-type problems. In~\cite{outliers, srdr}, Harris et.\ al.\ analyzed models similar to chance $k$-coverage and minimization of $\ensuremath{\mathbf{E}}[d(j,\S)]$, but applied to knapsack center and matroid center problems; they also considered robust versions (in which a small subset of clients may be denied service). While the overall model was similar to the ones we explore here, the techniques are somewhat different. Furthermore, these works focus only on the case where the target distribution is itself deterministic. Similar stochastic approximation guarantees have appeared in the context of approximation algorithms for static problems, particularly $k$-median problems. In \cite{charikar-li}, Charikar \& Li discussed a randomized procedure for converting a linear-programming relaxation in which a client has \emph{fractional} distance $t_j$, into a distribution $\Omega$ satisfying $\ensuremath{\mathbf{E}}_{\S \sim \Omega}[d(j, \S)] \leq 3.25 t_j$. This property can be used, among other things, to achieve a $3.25$-approximation for $k$-median. However, many other randomized rounding algorithms for $k$-median only seek to preserve the \emph{aggregate} value $\sum_j \ensuremath{\mathbf{E}}[d(j,\S)]$, without our type of per-point guarantee. We also contrast our approach with a different stochastic $k$-center problem considered in works such as \cite{huang, alipour}. These consider a model with a fixed, deterministic set $\S$ of open facilities, while the client set is determined stochastically; this model is almost precisely opposite to ours. \subsection{Publicly verifying the distributions} Our approximation algorithms will have the following structure: given some target distribution $\Omega$, we construct a randomized procedure $\mathcal A$ which returns some random set $\S$ with good probabilistic guarantees matching $\Omega$. Thus the algorithm $\mathcal A$ is \emph{itself} the approximating distribution $\tilde \Omega$. In a number of cases, we can convert the randomized algorithm $\mathcal A$ into a distribution $\tilde \Omega$ which has a sparse support (set of points to which it assigns nonzero probability), and which can be enumerated directly. This may cause a small loss in approximation ratio. The distribution $\tilde \Omega$ can be publicly verified, and the users can then draw from $\tilde \Omega$ as desired. Recall that one of our main motivations is fairness in clustering; the ability for the users to verify that they are being treated fairly in a stochastic sense (although not necessarily in any one particular run of the algorithm) is particularly important. \subsection{Notation} We define $\binom{\mathcal{F}}{k}$ to be the collection of $k$-element subsets of $\mathcal{F}$. We assume throughout that $\mathcal{F}$ can be made arbitrarily large by duplicating its elements; thus, whenever we have an expression like $\binom{\mathcal{F}}{k}$, we assume without loss of generality that $|\mathcal{F}| \geq k$. We will let $[t]$ denote the set $\{1, 2, \ldots, t\}$. For any vector $a = (a_1, \dots, a_t)$ and a subset $X \subseteq [t]$, we write $a(X)$ as shorthand for $\sum_{i \in X} a_i$. We use the Iverson notation throughout, so that for any Boolean predicate $\mathcal P$ we let $[[\mathcal P]]$ be equal to one if $\mathcal P$ is true and zero otherwise. For a real number $q \in [0,1]$, we define $\overline q = 1 - q$. Given any $j \in \mathcal{C}$ and any real number $r \geq 0$, we define the ball $B(j, r) = \{ i \in \mathcal{F} \mid d(i,j) \leq r \}$. We let $\theta(j)$ be the distance from $j$ to the nearest facility, and $V_j$ be the facility closest to $j$, i.e. $d(j, V_j) = d(j, \mathcal{F}) = \theta(j)$. Note that in the SCC setting we have $V_j = j$ and $\theta(j) = 0$. For a solution set $\S$, we say that $j \in \mathcal{C}$ is \emph{matched} to $i \in \S$ if $i$ is the closest facility of $\S$ to $j$; if there are multiple closest facilities, we take $i$ to be one with least index. \subsection{Some useful subroutines} We will use two basic subroutines repeatedly: \emph{dependent rounding} and \emph{greedy clustering}. In dependent rounding, we aim to preserve certain marginal distributions and negative correlation properties while satisfying some constraints with probability one. Our algorithms use a dependent-rounding algorithm from \cite{srin:level-sets}, which we summarize as follows: \begin{proposition} \label{prop:dep-round} There exists a randomized polynomial-time algorithm $\textsc{DepRound}(y)$ which takes as input a vector $y\in [0,1]^n$, and outputs a random set $Y \subseteq [n]$ with the following properties: \begin{enumerate} \item[(P1)] $\Pr[i \in Y]=y_i$, for all $i\in [n]$, \item[(P2)] $\lfloor \sum_{i=1}^n y_i \rfloor \leq |Y| \leq \lceil \sum_{i=1}^n y_i \rceil$ with probability one, \label{depround:cardinality} \item[(P3)] $\Pr[Y \cap S = \emptyset] \leq \prod_{i \in S} (1 - y_i)$ for all $S \subseteq [n]$. \label{depround:negcorr} \end{enumerate} \end{proposition} We adopt the following additional convention: suppose $(y_1, \dots, y_n) \in [0,1]^n$ and $S \subseteq [n]$; we then define $\textsc{DepRound}(y,S) \subseteq S$ to be $\textsc{DepRound}(x)$, for the vector $x$ defined by $x_i = y_i [[ i \in S]]$. The greedy clustering procedure takes an input a set of weights $w_j$ and sets $F_j \subseteq \mathcal{F}$ for every client $j \in \mathcal{C}$, and executes the following procedure: \begin{algorithm}[H] \caption{\sc GreedyCluster$(F, w)$} \begin{algorithmic}[1] \STATE Sort $\mathcal{C}$ as $\mathcal{C} = \{j_1, j_2, \dots, j_{\ell} \}$ where $w_{j_1} \leq w_{j_2} \leq \dots \leq w_{j_{\ell}}$. \STATE Initialize $C' = \emptyset$ \FOR{$t = 1, \dots, \ell$} \STATE\textbf{if} $F_{j_t} \cap F_{j'} = \emptyset$ for all $j' \in C'$ \textbf{then} update $C' \leftarrow C' \cup \{j_t \}$ \ENDFOR \STATE Return $C'$ \end{algorithmic} \end{algorithm} \begin{observation} \label{greedy-prop} If $C' = \textsc{GreedyCluster}(F, w)$ then for any $j \in \mathcal{C}$ there is $z \in C'$ with $w_z \leq w_j$ and $F_z \cap F_j \neq \emptyset$. \end{observation} \section{The chance $k$-coverage problem} \label{chance-sec} In this section, we consider a scenario we refer to as the \emph{chance $k$-coverage problem}: every point $j \in \mathcal{C}$ has demand parameters $p_j, r_j$, and we wish to find a $k$-lottery $\Omega$ such that \begin{equation} \label{stt1} \Pr_{\S \sim \Omega}[ d(j, \S) \leq r_j] \geq p_j. \end{equation} If a $k$-lottery satisfying (\ref{stt1}) exists, we say that parameters $p_j, r_j$ are \emph{feasible.} We refer to the special case wherein every client $j$ has a common value $p_j = p$ and a common value $r_j = r$, as \emph{homogeneous}. Homogeneous instances naturally correspond to fair allocations, for example, $k$-supplier is a special case of the homogeneous chance $k$-coverage problem, in which $p_j= 1$ and $r_j$ is equal to the optimal $k$-supplier radius. Our approximation algorithms for this problem will be based on a linear programming (LP) relaxation which we denote $\mathcal P_{\text{chance}}$. It has fractional variables $b_i$, where $i$ ranges over $\mathcal{F}$ ($b_i$ represents the probability of opening facility $i$), and is defined by the following constraints: \begin{enumerate}[label=(B\arabic*)] \item $\sum_{i \in B(j, r_j)} b_i \ge p_j$ for all $j \in \mathcal{C}$, \label{cmkcP:masspj} \item $b(\mathcal{F}) = k$, \label{cmkcP:cardinality} \item $b_i \in [0,1]$ for all $i \in \mathcal{F}$. \label{cmkcP:unit} \end{enumerate} \begin{proposition} If parameters $p, r$ are feasible, then $\P_{\text{chance}}$ is nonempty. \end{proposition} \begin{proof} Consider a distribution $\Omega$ satisfying (\ref{stt1}). For each $i \in \mathcal{F}$, set $b_i = \Pr_{\S \sim \Omega}[i \in \S]$. For $j \in \mathcal{C}$ we have $p_j = \Pr \bigl[ \bigvee_{i \in B(j, r_j)} i \in \S \bigr] \leq \sum_{i \in B(j,r_j)} \Pr[i \in S] = \sum_{i \in B(j,r_j)} b_i$ and thus (B1) is satisfied. We have $k = \ensuremath{\mathbf{E}}[|\S|] = \sum_{i \in \mathcal{F}} \Pr[i \in S] = b(\mathcal{F})$ and so (B2) is satisfied. (B3) is clear, so we have demonstrated a point in $\P_{\text{chance}}$. \end{proof} For the remainder of this section, we assume we have a vector $b \in \P_{\text{chance}}$ and focus on how to round it to an integral solution. By a standard facility-splitting step, we also generate, for every $j \in \mathcal{C}$, a set $F_j \subseteq B(j,r_j)$ with $b(F_j) = p_j$. We refer to this set $F_j$ as a \emph{cluster}. In the SCC setting, it will also be convenient to ensure that $j \in F_j$ as long as $b_j \neq 0$. As we show in Section~\ref{lb-sec}, any approximation algorithm must either significantly give up a guarantee on the distance, or probability (or both). Our first result is an approximation algorithm which respects the distance guarantee exactly, with constant-factor loss to the probability guarantee: \begin{theorem} \label{ud1} If $p, r$ is feasible then one may efficiently construct a $k$-lottery $\Omega$ satisfying $$ \Pr_{\S \sim \Omega} [d(j, \S) \leq r_j] \geq (1 - 1/e) p_j. $$ \end{theorem} \begin{proof} Let $b \in \P_{\text{chance}}$ and set $\S = \textsc{DepRound}(b)$. This satisfies $|\S| \leq \lceil \sum_{i=1}^n b_i \rceil \leq \lceil k \rceil = k$ as desired. Each $j \in\mathcal{C}$ has $$ \Pr[\S \cap F_j = \emptyset] \leq \prod_{i \in F_j}(1-b_i) \le \prod_{i\in F_j}e^{-b_i}= e^{-b(F_i)}= e^{-p_j}. $$ and then simple analysis shows that \[ \Pr[ d(j, \S) \leq r_j] \geq \Pr[ \S \cap F_j \neq \emptyset] \geq 1 - e^{-p_j} \geq (1-1/e) p_j \qedhere \] \end{proof} As we will later show in Theorem~\ref{ath0}, this approximation constant $1 - 1/e$ is optimal. We next turn to preserving the probability guarantee exactly with some loss to distance guarantee. As a warm-up exercise, let us consider the special case of ``half-homogeneous'' problem instances: all the values of $p_j$ are the same, or all the values of $r_j$ are the same. A similar algorithm works both these cases: we first select a set of clusters according to some greedy order, and then open a single item from each cluster. We summarize this as follows: \begin{algorithm}[H] \begin{algorithmic}[1] \STATE Set $C' = \textsc{GreedyCluster}(F_j, w_j)$ \STATE Set $Y = \textsc{DepRound}(p, C')$ \STATE Form solution set $\S = \{ V_j \mid j \in Y \}$. \end{algorithmic} \caption{Rounding algorithm for half-homogeneous chance $k$-coverage} \label{kcov1} \end{algorithm} Algorithm~\ref{kcov1} opens at most $k$ facilities, as the dependent rounding step ensures that $|Y| \leq \lceil \sum_{j \in C'} p_j \rceil = \lceil \sum_{j \in C'} b(F_j) \rceil \leq \lceil \sum_{i \in \mathcal{F}} b_i \rceil \leq k$. The only difference between the two cases is the choice of weighting function $w_j$ for the greedy cluster selection. \begin{proposition} \label{ud2} \begin{enumerate} \item Suppose that $p_j$ is the same for every $j \in \mathcal{C}$. Then using the weighting function $w_j = r_j$ ensures that every $j \in \mathcal{C}$ satisfies $\Pr[d(j,\S) \leq 3 r_j] \geq p_j.$ Furthermore, in the SCC setting, it satisfies $\Pr[d(j,\S) \leq 2 r_j] \geq p_j.$ \item Suppose $r_j$ is the same for every $j \in \mathcal{C}$. Then using the weighting function $w_j = 1 - p_j$ ensures that every $j \in \mathcal{C}$ satisfies $\Pr[d(j,\S) \leq 3 r_j] \geq p_j.$ Furthermore, in the SCC setting, it satisfies $\Pr[d(j,\S) \leq 2 r_j] \geq p_j.$ \end{enumerate} \end{proposition} \begin{proof} Let $j \in \mathcal{C}$. By Observation~\ref{greedy-prop} there is $z \in C'$ with $w_z \leq w_j$ and $F_j \cap F_z \neq \emptyset$. In either of the two cases, this implies that $p_z \geq p_j$ and $r_z \leq r_j$. Letting $i \in F_j \cap F_z$ gives $d(j,z) \leq d(j,i) + d(z,i) \leq r_j + r_z \leq 2 r_j$. This $z \in C'$ survives to $Y$ with probability $p_z \geq p_j$, and in that case we have $d(z,\S) = \theta(z)$. In the SCC setting, this means that $d(z,\S) = 0$; in the general setting, we have $\theta(z) \leq r_z \leq r_j$. \end{proof} \subsection{Approximating the general case} We next show how to satisfy the probability constraint exactly for the general case of chance $k$-coverage, with a constant-factor loss in the distance guarantee. Namely, we will find a probability distribution with $$ \Pr_{\S \sim \Omega} [d(j, \S) \leq 9 r_j] \geq p_j. $$ The algorithm is based on iterated rounding, in which the entries of $b$ go through an unbiased random walk until $b$ becomes integral (and, thus corresponds to a solution set $\S$). Because the walk is unbiased, the probability of serving a client at the end is equal to the fractional probability of serving a client, which will be at least $p_j$. In order for this process to make progress, the number of active variables must be greater than the number of active constraints. We ensure this by periodically identifying and discarding clients which will be automatically served by serving other clients. This is similar to a method of \cite{krishnaswamy}, which also uses iterative rounding for (deterministic) approximations to $k$-median with outliers and $k$-means with outliers. The sets $F_j$ will remain fixed during this procedure. We will maintain a vector $b \in [0,1]^{\mathcal{F}}$ and maintain two sets $C_{\text{tight}}$ and $C_{\text{slack}}$ with the following properties: \begin{enumerate} \item[(C1)] $C_{\text{tight}} \cap C_{\text{slack}} = \emptyset$. \item[(C2)] For all $j, j' \in C_{\text{tight}}$, we have $F_j \cap F_{j'} = \emptyset$ \item[(C3)] Every $j \in C_{\text{tight}}$ has $b(F_j) = 1$, \item[(C4)] Every $j \in C_{\text{slack}}$ has $b(F_j) \leq 1$. \item[(C5)] We have $b(\bigcup_{j \in C_{\text{tight}} \cup C_{\text{slack}}} F_j) \leq k$ \end{enumerate} Given our initial solution $b$ for $\P_{\text{chance}}$, setting $C_{\text{tight}} = \emptyset, C_{\text{slack}} = \mathcal{C}$ will satisfy criteria (C1)--(C5); note that (C4) holds as $b(F_j) = p_j \leq 1$ for all $j \in \mathcal{C}$. \begin{proposition} \label{ggcor} Given any vector $b \in [0, 1 ]^{\mathcal{F}}$ satisfying constraints (C1)---(C5) with $C_{\text{slack}} \neq \emptyset$, it is possible to generate a random vector $b' \in [0,1]^{\mathcal{F}}$ such that $\ensuremath{\mathbf{E}}[b'] = b$, and with probability one $b'$ satisfies constraints (C1) --- (C5) as well as having some $j \in C_{\text{slack}}$ with $b'(F_j) \in \{0, 1 \}.$ \end{proposition} \begin{proof} We will show that any basic solution $b \in [0, 1 ]^{\mathcal{F}}$ to the constraints (C1)---(C5) with $C_{\text{slack}} \neq \emptyset$ must satisfy the condition that $b(F_j) \in \{0, 1 \}$ for some $j \in C_{\text{slack}}$. To obtain the stated result, we simply modify $b$ until it becomes basic by performing an unbiased walk in the nullspace of the tight constraints. So consider a basic solution $b$. Define $A = \bigcup_{j \in C_{\text{tight}}} F_j$ and $B = \bigcup_{j \in C_{\text{slack}}} F_j$. We assume that $b(F_j) \in (0,1)$ for all $j \in C_{\text{slack}}$, as otherwise we are done. First, suppose that $b(A \cap B) > 0$. So there must be some pair $j \in C_{\text{slack}}, j' \in C_{\text{tight}}$ with $i \in F_j \cap F_{j'}$ such that $b_i > 0$. Since $b(F_{j'}) = 1$, there must be some other $i' \in F_{j'}$ with $b_{i'} > 0$. Consider modifying $b$ by incrementing $b_i$ by $\pm \epsilon$ and decrementing $b_{i'}$ by $\pm \epsilon$, for some sufficiently small $\epsilon$. Constraint (C5) is preserved. Since $F_{j'} \cap F_{j''} = \emptyset$ for all $j'' \in C_{\text{tight}}$, constraint (C3) is preserved. Since the (C4) constraints are slack, then for $\epsilon$ sufficiently small they are preserved as well. This contradicts that $b$ is basic. Next, suppose that $b(A \cap B) = 0$ and $b(A \cup B) < k$ strictly. Let $j \in C_{\text{slack}}$ and $i \in F_j$ with $b_i > 0$; if we change $b_i$ by $\pm \epsilon$ for sufficiently small $\epsilon$, this preserves (C4) and (C5); furthermore, since $i \notin A$, it preserves (C3) as well. So again $b$ cannot be basic. Finally, suppose that $b(A \cap B) = 0$ and $b(A \cup B) = k$. So $b(B) = k - |A|$ and $b(B) > 0$ as $C_{\text{slack}} \neq \emptyset$. Therefore, there must be at least two elements $i, i' \in B$ such that $b_i > 0, b_{i'} > 0$. If we increment $b_i$ by $\pm\epsilon $ while decrementing $b_{i'}$ by $\pm \epsilon$, this again preserves all the constraints for $\epsilon$ sufficiently small, contradicting that $b$ is basic. \end{proof} We can now describe our iterative rounding algorithm, Algorithm~\ref{algo:it3}. \begin{algorithm}[H] \caption{Iterative rounding algorithm for chance $k$-coverage} \begin{algorithmic}[1] \STATE Find a fractional solution $b$ to $\P_{\text{chance}}$ and form the corresponding sets $F_j$. \STATE Initialize $C_{\text{tight}} = \emptyset, C_{\text{slack}} = \mathcal{C}$ \WHILE{ $C_{\text{slack}} \neq \emptyset$ } \STATE Draw a fractional solution $b'$ with $\ensuremath{\mathbf{E}}[b'] = b$ according to Proposition~\ref{ggcor}. \STATE Select some $v \in C_{\text{slack}}$ with $b'(F_v) \in \{0, 1 \}$. \STATE Update $C_{\text{slack}} \leftarrow C_{\text{slack}} - \{ v \}$ \IF{$b'(F_v) = 1$} \STATE Update $C_{\text{tight}} \leftarrow C_{\text{tight}} \cup \{ v \}$ \IF{there is any $z \in C_{\text{tight}} \cup C_{\text{slack}}$ such that $r_{z} \geq r_v/2$ and $F_{z} \cap F_v \neq \emptyset$} \STATE{Update $C_{\text{tight}} \leftarrow C_{\text{tight}} - \{ z \}, C_{\text{slack}} \leftarrow C_{\text{slack}} - \{ z \}$} \ENDIF \ENDIF \STATE Update $b \leftarrow b'$. \ENDWHILE \STATE For each $j \in C_{\text{tight}}$, open an arbitrary center in $F_j$. \end{algorithmic} \label{algo:it3} \end{algorithm} To analyze this algorithm, define $C_{\text{tight}}^t, C_{\text{slack}}^t, b^t$ to be the values of the relevant variables at iteration $t$. Since every step removes at least one point from $C_{\text{slack}}$, this process must terminate in $T \leq n$ iterations. We will write $b^{t+1}$ to refer to the random value $b'$ chosen at step (4) of iteration $t$, and $v^t$ denote the choice of $v \in C_{\text{slack}}$ used step in step (5) of iteration $t$. \begin{proposition} \label{bprop} The vector $b^t$ satisfies constraints (C1) --- (C5) for all times $t = 1, \dots, T$. \end{proposition} \begin{proof} The vector $b^0$ does so since $b$ satisfies $\P_{\text{chance}}$. Proposition~\ref{ggcor} ensures that step (4) does not affect this. Also, removing points from $C_{\text{tight}}$ or $C_{\text{slack}}$ at step (6) or (1) will not violate these constraints. Let us check that adding $v^t$ to $C_{\text{tight}}$ will not violate the constraints. This step only occurs if $b^{t+1}(v^t) = 1$, and so (C3) is preserved. Since we only move $v^t$ from $C_{\text{slack}}$ to $C_{\text{tight}}$, constraints (C1) and (C5) are preserved. Finally, to show that (C2) is preserved, suppose that $F_{v^t} \cap F_{v^s} \neq \emptyset$ for some other $v^s$ which was added to $C_{\text{tight}}$ at time $s < t$. If $r_{v^t} \geq r_{v^s}$, then step (10) would have removed $v^t$ from $C_{\text{slack}}$, making it impossible to enter $C_{\text{tight}}^t$. Thus, $r_{v^t} \leq r_{v^s}$; this means that when we add $v^t$ to $C_{\text{tight}}^t$, we also remove $v^s$ from $C_{\text{tight}}^t$. \end{proof} \begin{corollary} Algorithm~\ref{algo:it3} opens at most $k$ facilities. \end{corollary} \begin{proof} At the final step (12), the number of open facilities is equal to $|C_{\text{tight}}|$. By Proposition~\ref{bprop}, the vector $b^T$ satisfies constraints (C1) --- (C5). So $b(F_j) = 1$ for $j \in C_{\text{tight}}$ and $F_j \cap F_{j'} = \emptyset$ for $j, j \in C_{\text{tight}}$; thus $|C_{\text{tight}}| = \sum_{j \in C_{\text{tight}}} b(F_j) \leq k$. \end{proof} \begin{proposition} \label{hh1} If $j \in C_{\text{tight}}^t$ for any time $t$, then $d(j,\S) \leq 3 r_j$. \end{proposition} \begin{proof} Let $t$ be maximal such that $j \in C_{\text{tight}}^t$. We show the desired claim by induction on $t$. When $t = T$, then this certainly holds as step (12) will open some facility in $F_j$ and thus $d(j,\S) \leq r_j$. Suppose that $j$ was added into $C_{\text{tight}}^s$, but was later removed from $C_{\text{tight}}^{t+1}$ due to adding $z = v^t$. Thus there is some $i \in F_z \cap F_j$. When we added $j$ in time $s$, we would have removed $z$ from $C_{\text{tight}}^s$ if $r_z \geq r_j/2$. Since this did not occur, it must hold that $r_z < r_j/2$. Since $z$ is present in $C_{\text{tight}}^{t+1}$, the induction hypothesis implies that $d(z,\S) \leq 3 r_z$ and so \[ d(j,\S) \leq d(j, i) + d(i, z) + d(z, \S) \leq r_j + r_z + 3 r_z \leq r_j + (r_j/2) (1 + 3) = 3 r_j. \qedhere \] \end{proof} \begin{theorem} Every $j \in \mathcal{C}$ has $\Pr[ d(j,\S) \leq 9 r_j] \geq p_j$. \end{theorem} \begin{proof} We will prove by induction on $t$ the following claim: suppose we condition on the full state of Algorithm~\ref{algo:it3} up to time $t$, and that $j \in C_{\text{tight}}^t \cup C_{\text{slack}}^t$. Then \begin{equation} \label{y1} \Pr[ d(j,\S) \leq 9 r_j] \geq b^t(F_j). \end{equation} At $t = T$, this is clear; since $C_{\text{slack}}^T = \emptyset$, we must have $j \in C_{\text{tight}}^T$, and so $d(j,\S) \leq r_j$ with probability one. For the induction step at time $t$, note that as $\ensuremath{\mathbf{E}}[b^{t+1}(F_j)] = b(F_j)$, in order to prove (\ref{y1}) it suffices to show that if we also condition on the value of $b^{t+1}$, it holds that \begin{equation} \label{y2} \Pr[ d(j,\S) \leq 9 r_j \mid b^{t+1}] \geq b^{t+1}(F_j). \end{equation} If $j$ remains in $C_{\text{tight}}^{t+1} \cup C_{\text{slack}}^{t+1}$, then we immediately apply the induction hypothesis at time $t+1$. So the only non-trivial thing to check is that (\ref{y2}) will hold even if $j$ is removed from $C_{\text{tight}}^{t+1} \cup C_{\text{slack}}^{t+1}$. If $j = v^t$ and $b^{t+1}(F_j) = 0$, then (\ref{y2}) holds vacuously. Otherwise, suppose that $j$ is removed from $C_{\text{tight}}^t$ at stage (10) due to adding $z = v^t$. Thus $r_j\geq r_z/2$ and there is some $i \in F_j \cap F_z$. By Proposition~\ref{hh1}, this ensures that $d(z,\S) \leq 3 r_z$. Thus with probability one we have $$ d(j,\S) \leq d(j,i) + d(i,z) + d(z, \S)\leq r_j + r_z + 3 r_z \leq r_j + (2 r_j) (1 + 3) = 9 r_j. $$ The induction is now proved. The claimed result follows since $b^0(F_j) = p_j$ and $C_{\text{slack}}^0 = \mathcal{C}$. \end{proof} \section{Chance $k$-coverage: approximating the deterministic case} \label{sec2} An important special case of $k$-coverage is where $p_j = 1$ for all $j \in \mathcal{C}$. Here, the target distribution $\Omega$ is just a single set $\S$ satisfying $\forall j d(j, \S) \leq r_j$. In the homogeneous case, when all the $r_j$ are equal to the same value, this is specifically the $k$-supplier problem. The usual approximation algorithm for this problem chooses a single approximating set $\S$, in which case the best guarantee available is $d(j, \S) \leq 3 r_j$. We improve the distance guarantee by constructing a $k$-lottery $\tilde \Omega$ such that $d(j, \S) \leq 3 r_j$ with probability one, and $\ensuremath{\mathbf{E}}_{\S \sim \tilde \Omega}[d(j, \S)] \leq c r_j$, where the constant $c$ satisfies the following bounds: \begin{enumerate} \item In the general case, $c = 1 + 2/e \approx 1.73576$; \item In the SCC setting, $c = 1.60793$; \item In the homogeneous SCC setting, $c = 1.592$.\footnote{This value was calculated using some non-rigorous numerical analysis; we describe this further in what we call ``Pseudo-Theorem''~\ref{thm:k-center-1.592}} \end{enumerate} We show matching lower bounds in Section~\ref{lb-sec}; the constant value $1 + 2/e$ is optimal for the general case (even for homogeneous instances), and for the third case the constant $c$ cannot be made lower than $1 + 1/e \approx 1.367$. We remark that this type of stochastic guarantee allows us to efficiently construct publicly-verifiable lotteries. \begin{proposition} Let $\epsilon > 0$. In any of the above three cases, there is an expected polynomial time procedure to convert the given distribution $\Omega$ into an explicitly-enumerated $k$-lottery $\Omega'$, with support size $O(\frac{\log n}{\epsilon^2})$, such that $\Pr_{\S \sim \Omega'} [d(j, \S) \leq 3 r_j] = 1$ and $\ensuremath{\mathbf{E}}_{\S \sim \Omega'} [d(j, \S)] \leq c (1 + \epsilon) r_j$. \end{proposition} \begin{proof} Take $X_1, \dots, X_t$ as independent draws from $\Omega$ for $t = \frac{6 \log n}{c \epsilon^2}$ and set $\Omega'$ to be the uniform distribution on $\{ X_1, \dots, X_t \}$. To see that $\ensuremath{\mathbf{E}}_{\S \sim \Omega'} [d(j, \S)] \leq c (1+\epsilon) r_j$ holds with high probability, apply a Chernoff bound, noting that $d(j,X_1), \dots, d(j,X_t)$ are independent random variables in the range $[0, 3 r_j]$. \end{proof} We use a similar algorithm to Algorithm~\ref{kcov1} for this problem: we choose a covering set of clusters $C'$, and open exactly one item from each cluster. The main difference is that instead of opening the nearest item $V_j$ for each $j \in C'$, we instead open a cluster according to the solution $b$ of $\P_{\text{chance}}$. \begin{algorithm}[H] \begin{algorithmic}[1] \STATE Set $C' = \textsc{GreedyCluster}(F_j, r_j)$. \STATE Set $F_0 = \mathcal{F} - \bigcup_{j \in C'} F_j$; this is the set of ``unclustered'' facilities \FOR{$j \in C'$} \STATE Randomly select a point $W_j \in F_j$ according to the distribution $\Pr[W_j = i] = b_i$ \\ \textit{// This is a valid probability distribution, as $b(F_j) = 1$} \ENDFOR \STATE Let $\S_0 \gets \textsc{DepRound}(b, F_0)$ \STATE Return $\S = \S_0 \cup \{ W_j \mid j \in C' \}$ \end{algorithmic} \caption{Rounding algorithm with clusters} \label{algo:round0} \end{algorithm} We will need the following technical result in order to analyze Algorithm~\ref{algo:round0}. \begin{proposition} \label{sec2-bound} For any set $U \subseteq \mathcal{F}$, we have $$ \Pr[\S \cap U = \emptyset] \leq \prod_{i \in U \cap F_0}(1-b_i) \prod_{j \in C'}(1-b(U \cap F_j)) \leq e^{-b(U)}. $$ \end{proposition} \begin{proof} The set $U$ contains each $W_j$ independently with probability $b(U \cap F_j)$. The set $\S_0$ is independent of them and by (P3) we have $\Pr[U \cap \S_0 = \emptyset] \leq \prod_{i \in U \cap F_0} (1 - b_i)$. So \begin{align*} \Pr[ \S \cap U = \emptyset ] &\leq \negthinspace \prod_{i \in U \cap F_0} \negthinspace (1 - b_i) \prod_{j \in C'} (1 - b(U \cap F_j)) \leq \negthinspace \prod_{i \in U \cap F_0} e^{-b_i} \negthinspace \prod_{j \in C'}e^{-b(U \cap F_j)} = e^{-b(U)} \qedhere \end{align*} \end{proof} At this point, we can show our claimed approximation ratio for the general (non-SCC) setting: \begin{theorem} \label{simple-bnd-thm0} For any $j \in \mathcal{C}$, the solution set $\S$ of Algorithm~\ref{algo:round0} satisfies $d(j,\S) \leq 3 r_j$ with probability one and $\ensuremath{\mathbf{E}}[d(j,\S)] \leq (1 + 2/e) r_j$. \end{theorem} \begin{proof} By Observation~\ref{greedy-prop}, there is some $v \in C'$ with $F_j \cap F_v \neq \emptyset$ and $r_v \leq r_j$. Letting $i \in F_j \cap F_v$, we have $$ d(j,\S) \leq d(j,i) + d(i,v) + d(v,\S) \leq r_j + r_v + r_v \leq 3 r_j. $$ with probability one. If $\S \cap F_j \neq \emptyset$, then $d(j, \S) \leq r_j$. Thus, a necessary condition for $d(j,\S) > r_j$ is that $\S \cap F_j = \emptyset$. Applying Proposition~\ref{sec2-bound} with $U = F_j$ gives $$ \Pr[ d(j,\S) > r_j ] \leq \Pr[\S \cap F_j = \emptyset] \leq e^{-b(F_j)}= e^{-1} $$ and so $\ensuremath{\mathbf{E}}[d(j,\S)] \leq r_j + 2 r_j \Pr[d(j,\S) > r_j] \leq (1 + 2/e) r_j$. \end{proof} \subsection{The SCC setting} To motivate the algorithm for the SCC setting (where $\mathcal{C} = \mathcal{F}$), note that if some client $j \in \mathcal{C}$ has some facility $i$ opened in a nearby cluster $F_{v}$, then this guarantees that $d(j, \S) \leq d(j, v) + d(v, i) \leq 3 r_j$. This is what we used to analyze the non-SCC setting. But, if instead of opening facility $i$, we opened $v$ itself, then this would ensure that $d(j, \S) \leq 2 r_j$. Thus, opening the centers of a cluster can lead to better distance guarantees compared to opening any other facility. We emphasize that this is only possible in the SCC setting, as in general we do not know that $v \in \mathcal{F}$. We use the following Algorithm~\ref{algo:round1}, which takes a parameter $q \in [0,1]$ which we will discuss how to set shortly. We recall that we have assumed in this case that $j \in F_j$ for every $j \in \mathcal{C}$. (Also, note our notational convention that $\overline q = 1 - q$; this will be used extensively in this section to simplify the formulas.) \begin{algorithm}[H] \begin{algorithmic}[1] \STATE Set $C' = \textsc{GreedyCluster}(F_j, r_j)$. \STATE Set $F_0 = \mathcal{F} - \bigcup_{j \in C'} F_j$; this is the set of ``unclustered'' facilities \FOR{$j \in C'$} \STATE Randomly select $W_j \in F_j$ according to the distribution $\Pr[ W_j = i ] = \overline q b_i + q [[i =j ]]$ \ENDFOR \STATE Let $\S_0 = \textsc{DepRound}(b, F_0)$ \STATE Return $\S = \S_0 \cup \{ W_j \mid j \in C' \}$ \end{algorithmic} \caption{Rounding algorithm for $k$-center} \label{algo:round1} \end{algorithm} This is the same as Algorithm~\ref{algo:round0}, except that some of the values of $b_i$ for $i \in F_j$ have been shifted to the cluster center $j$. In fact, we can think of Algorithm~\ref{algo:round1} as a two-part process: we first modify the fractional vector $b$ to obtain a new fractional vector $b'$ defined by $$ b'_i = \begin{cases} \overline q b_i + q & \text{if $i \in C'$} \\ \overline q b_i & \text{if $i \in F_{j} - \{j \}$ for $j \in C'$} \\ b_i & \text{if $i \in F_0$} \end{cases}. $$ and we then execute Algorithm~\ref{algo:round0} on the resulting vector $b'$. In particular, Proposition~\ref{sec2-bound} remains valid with respect to the modified vector $b'$. \begin{theorem} \label{simple-bnd-thm} Let $j \in \mathcal{C}$. After running Algorithm~\ref{algo:round1} with $q = 0.464587$ we have $d(j,\S) \leq 3 r_j$ with probability one and $\ensuremath{\mathbf{E}}[d(j,\S)] \leq 1.60793 r_j$. \end{theorem} \begin{proof} Let $D = \{v \in C' \mid F_j \cap F_{v} \neq \emptyset, r_v \leq r_j \}$; note that $D \neq \emptyset$ by Observation~\ref{greedy-prop}. For each $v \in D \cup \{0 \}$, set $a_v = b(F_j \cap F_v)$ and observe that $a_0 + \sum_{v \in D} a_v = b(F_j) = 1$. As before, a necessary condition for $d(j,\S) > r_j$ is that $F_j \cap \S = \emptyset$. So by Proposition~\ref{sec2-bound}, \begin{align*} \Pr[ d(j,\S) > r_j ] &\leq \Pr[ F_j \cap \S = \emptyset ] \leq \prod_{i \in F_j \cap F_0} (1-b'_i) \prod_{v \in C'} (1 - b'(F_v \cap F_j)) \\ &\leq \prod_{i \in F_j \cap F_0} (1-b_i) \prod_{v \in D} (1 - \overline q b(F_v \cap F_j)) \leq e^{-b(F_j \cap F_0)} \prod_{v \in D} (1 - \overline q b(F_v \cap F_j)) \\ &= e^{-a_0} \prod_{v \in D} e^{a_v} (1 - \overline q a_v) = (1/e) \prod_{v \in D} e^{a_v} (1 - \overline q a_v). \end{align*} where the last equality comes from the fact that $a_0 + a(D) = 1$. Similarly, if there is some $i \in \S \cap D$, then $d(j, \S) \leq d(v,i) \leq 2 r_j$. Thus, a necessary condition for $d(j,\S) > 2 r_j$ is that $\S \cap (D \cup F_j) = \emptyset$. Applying Proposition~\ref{sec2-bound} with $U = D \cup F_j$ gives: \begin{align*} \Pr[ d(j,\S) > 2 r_j] &\leq \prod_{v \in (D \cup F_j) \cap F_0} (1 - b_v) \prod_{v \in C'} ( 1 - b'( (D \cup F_j) \cap F_v) ) \\ &\leq e^{-b(F_j \cap F_0)} \prod_{v \in D} \overline q (1 - a_v) = (1/e) \prod_{v \in D} e^{a_v} \overline q (1 - a_v) \end{align*} Putting these together gives: \begin{equation} \label{yy1} \ensuremath{\mathbf{E}}[d(j,\S)] \leq r_j \Bigl( 1 + 1/e \prod_{v \in D} e^{a_v} (1 - \overline q a_v) + 1/e \prod_{v \in D} e^{a_v} \overline q (1 - a_v) \Bigr) \end{equation} Let us define $s = a(D)$ and $t = |D|$. By the AM-GM inequality we have: \begin{align*} \ensuremath{\mathbf{E}}[d(j,\S)] \leq r_j \Bigl( 1 + e^{s-1} \bigl( 1 - \overline q s/t \bigr)^t + e^{s -1} \overline q^t (1 - s/t)^t \Bigr) \end{align*} This is a function of a single real parameter $s \in [0,1]$ and a single integer parameter $t \geq 1$. Some simple analysis, which we omit here, shows that $\ensuremath{\mathbf{E}}[d(j,\S)] \leq 1.60793 r_j$. \end{proof} \subsection{The homogeneous SCC setting} \label{sec:k-center-partial-clusters} From the point of view of the target distribution $\Omega$, this setting is equivalent to the classical $k$-center problem. We may guess the optimal radius, and so we do not need to assume that the common value of $r_j$ is ``given'' to us by some external process. By rescaling, we assume without loss of generality here that $r_j = 1$ for all $j$. We will improve on Theorem~\ref{simple-bnd-thm} through a more complicated rounding process based on greedily-chosen partial clusters. Specifically, we select cluster centers $\pi(1), \dots, \pi(n)$, wherein $\pi(i)$ is chosen to maximize $b(F_{\pi(i)} - F_{\pi(1)} - \dots - F_{\pi(i-1)})$. By renumbering $\mathcal{C}$, we may assume without loss of generality that the resulting permutation $\pi$ is the identity; therefore, we assume throughout this section that $\mathcal{C} = \mathcal{F} = [n]$ and for all pairs $i < j$ we have \begin{equation} \label{ttr1} b(F_i - F_1 - \dots - F_{i-1}) \geq b(F_j - F_1 - \dots - F_{i-1}) \end{equation} For each $j \in [n]$, we let $G_j = F_{j} - F_{1} - \dots - F_{j-1}$ and we define $z_j = b(G_j)$. We say that $G_j$ is a \emph{full cluster} if $z_j = 1$ and a \emph{partial cluster} otherwise. We note that the values of $z$ appear in sorted order $1 = z_1 \geq z_2 \geq z_3 \geq \dots \geq z_n \geq 0$. We use the following randomized Algorithm~\ref{algo:round2} to select the centers. Here, the quantities $Q_{\text{f}}, Q_{\text{p}}$ (short for \emph{full} and \emph{partial}) are drawn from a joint probability distribution which we discuss later. \begin{algorithm}[H] \caption{$\textsc{Round2}\left(y, z,\bigcup_{j=1}^n G_j, Q_{\text{f}}, Q_{\text{p}} \right)$} \begin{algorithmic}[1] \STATE Draw random variables $Q_{\text{f}}, Q_{\text{p}}$. \STATE $Z \gets \textsc{DepRound}(z)$ \FOR{$j \in Z$} \STATE Randomly select $W_j \in G_j$ according to the distribution $\Pr[W_j=i] = ( \overline{q_j} y_i + q_j [[ i =j ]] )/z_j$ where $q_j$ is defined as $$ q_j = \begin{cases} Q_{\text{f}} & \text{if $z_j = 1$} \\ Q_{\text{p}} & \text{if $z_j < 1$} \end{cases} $$ \ENDFOR \STATE Return $\S = \{W _j \mid j \in Z \}$ \end{algorithmic} \caption{Partial-cluster based algorithm} \label{algo:round2} \end{algorithm} Before the technical analysis of Algorithm~\ref{algo:round2}, let us provide some intuition. It may be beneficial to open the center of some cluster near a given client $j \in \mathcal{C}$ as this will ensure $d(j,\S) \leq 2$. However, there is no benefit to opening more than one such cluster center. So, we would like a significant negative correlation between opening the centers of distinct clusters near $j$. Unfortunately, the full clusters all ``look alike,'' and so it seems impossible to enforce any significant negative correlation among them. Partial clusters break the symmetry. There is at least one full cluster near $j$, and possibly some partial clusters as well. We will create a probability distribution with significant negative correlation between the event that partial clusters open their centers and the event that full clusters open their centers. This decreases the probability that $j$ will see multiple neighboring clusters open their centers, which in turn gives an improved value of $\ensuremath{\mathbf{E}}[d(j,\S)]$. The dependent rounding in Algorithm~\ref{algo:round2} ensures that $|Z| \leq \lceil \sum_{j=1}^n z_j \rceil = \sum_{j=1}^n b(F_{j} - F_{1} - \dots - F_{j-1}) = b(\mathcal{F}) \leq k$, and so $|\S| \leq k$ as required. We also need the following technical result; the proof is essentially the same as Proposition~\ref{sec2-bound} and is omitted. \begin{proposition} \label{sec3-bound} For any $U \subseteq \mathcal{C}$, we have $$ \Pr[\S \cap U = \emptyset \mid Q_{\text{f}}, Q_{\text{p}}] \leq \prod_{j = 1}^n \bigl( 1 - \overline q_j b(U \cap G_j) - q_j z_j [[j \in U]] \bigr). $$ \end{proposition} Our main lemma to analyze Algorithm~\ref{algo:round2} is the following: \begin{lemma} \label{prop31} Let $i \in [n]$. Define $J_{\text{f}}, J_{\text{p}} \subseteq [n]$ as \begin{align*} J_{\text{f}} &= \{ j \in [n] \mid F_i \cap G_j \neq \emptyset, z_j = 1\} \\ J_{\text{p}} &= \{ j \in [n] \mid F_i \cap G_j \neq \emptyset, z_j < 1\} \end{align*} Let $m = |J_{\text{f}}|$ and $J_{\text{p}} = \{j_1, \dots, j_t \}$ where $j_1 \leq j_2 \leq \dots \leq j_t$. For each $\ell = 1, \dots, t+1$ define $$ u_{\ell} = b(F_i \cap G_{j_{\ell}}) + b(F_i \cap G_{j_{\ell+1}}) + \dots + b(F_i \cap G_{j_t}) $$ Then $1 \geq u_1 \geq u_2 \geq \dots \geq u_t \geq u_{t+1} = 0$ and $m \geq 1$. Furthermore, if we condition on a fixed value of $(Q_{\text{f}}, Q_{\text{p}})$ then we have \begin{align*} \ensuremath{\mathbf{E}}[d(i,\S)] & \leq 1 + \left( 1 - \overline Q_{\text{f}} \frac{\overline{ u_1}}{m} \right)^m \prod_{\ell=1}^t (1 - \overline Q_{\text{p}} (u_{\ell} - u_{\ell+1})) + \left( \overline Q_{\text{f}} \bigl( 1 - \frac{\overline{u_1}}{m} \bigr) \right)^m \prod_{\ell=1}^t (\overline{u_{\ell}} + \overline Q_{\text{p}} u_{\ell + 1}). \end{align*} \end{lemma} \begin{proof} For $\ell = 1, \dots, t$ we let $a_{\ell} = b(F_i \cap G_{j_{\ell}}) = u_{\ell} - u_{\ell+1}$. For $j \in J_{\text{f}}$, we let $s_j = b(F_i \cap G_j)$. First, we claim that $z_{j_\ell} \geq u_{\ell}$ for $\ell = 1, \dots, t$. For, by (\ref{ttr1}), we have \begin{align*} z_{j_\ell} &\geq b(F_i - F_{1} - \dots - F_{j_{\ell}-1}) \geq b(F_i) - \sum_{j \in J_{\text{f}}} b(F_i \cap G_j) - \sum_{v<j_{\ell}} b(F_i \cap G_v) \\ &= b(F_i) - \sum_{j \in J_{\text{f}}} b(F_i \cap G_j) - \sum_{v<{\ell}} b(F_i \cap G_{j_v}) \qquad \text{as $F_i \cap G_v = \emptyset$ for $v \notin J_{\text{p}} \cup J_{\text{f}}$} \\ &= 1 - \sum_{j \in J_{\text{f}}} s_j - \sum_{v = 1}^{\ell-1} a_v = u_{\ell}. \end{align*} Next, observe that if $m = 0$, we have $u_1 = \sum_{\ell = 1}^t b(F_i \cap G_{j_{\ell}}) = \sum_{j=1}^n b(F_i \cap G_j) = b(F_i) = 1$. Applying the above bound at $\ell = 1$ gives $z_{j_1} \geq u_1 = 1$. But, by definition of $J_{\text{p}}$ we have $z_{j_1} < 1$, which is a contradiction. This shows that $m \geq 1$. We finish by showing the bound on $\ensuremath{\mathbf{E}}[d(i, \S)]$. All the probabilities here should be interpreted as conditioned on fixed values for parameters $Q_{\text{f}}, Q_{\text{p}}$. A necessary condition for $d(i,\S) > 1$ is that no point in $F_i$ is open. Applying Proposition~\ref{sec3-bound} with $U = F_i$ yields \begin{align*} \Pr[\S \cap F_i = \emptyset] &\leq \prod_{j \in J_{\text{f}}} (1 - \overline Q_{\text{f}} b(F_i \cap G_j) - Q_{\text{f}} [[j \in F_i]]) \prod_{j \in J_{\text{p}}} (1 - \overline Q_{\text{p}} b(F_i \cap G_j)) - Q_{\text{p}} z_j [[j \in F_i]] ) \\ &\leq\prod_{j \in J_{\text{f}}} (1 - \overline Q_{\text{f}} b(F_i \cap G_j)) \prod_{j \in J_{\text{p}}} (1 - \overline Q_{\text{p}} b(F_i \cap G_j)) =\prod_{j \in J_{\text{f}}} (1 - \overline Q_{\text{f}} s_j) \prod_{\ell = 1}^t (1 - \overline Q_{\text{p}} a_{\ell}) . \end{align*} A necessary condition for $d(i,\S) > 2$ is that we do not open any point in $F_i$, nor do we open center of any cluster intersecting with $F_i$. Applying Proposition~\ref{sec3-bound} with $U = F_i \cup J_{\text{f}} \cup J_{\text{p}}$, and noting that $z_{j_{\ell}} \leq u_{\ell}$, we get: \begin{align*} \Pr[\S \cap U = \emptyset] &\leq \prod_{j \in J_{\text{f}}} (1 - {\overline Q_{\text{f}}} b(U \cap G_j) - Q_{\text{f}}) \prod_{j \in J_{\text{p}}} (1 - {\overline Q_{\text{p}}} b(U \cap G_j)) - Q_{\text{p}} z_j ) \\ &\leq\prod_{j \in J_{\text{f}}} (1 - \overline Q_{\text{f}} b(F_i \cap G_j) - Q_{\text{f}}) \prod_{\ell = 1}^t (1 - \overline Q_{\text{p}} b(F_i \cap G_{j_\ell}) - Q_{\text{p}} z_{j_\ell}) \\ &=\prod_{j \in J_{\text{f}}} \overline Q_{\text{f}} (1 - s_j) \prod_{\ell = 1}^t (1 - \overline Q_{\text{p}} a_{\ell} - Q_{\text{p}} z_{j_\ell}) \leq \prod_{j \in J_{\text{f}}} \overline Q_{\text{f}} (1 - s_j) \prod_{\ell = 1}^t (1 - \overline Q_{\text{p}} a_{\ell} - Q_{\text{p}} u_{\ell}) \end{align*} Thus, \begin{equation} \label{yy3} \ensuremath{\mathbf{E}}[d(i,\S)] \leq 1 +\prod_{j \in J_{\text{f}}} (1 - \overline Q_{\text{f}} s_j) \prod_{\ell = 1}^t (1 - \overline Q_{\text{p}} a_{\ell}) + \prod_{j \in J_{\text{f}}} \overline Q_{\text{f}} (1 - s_j) \prod_{\ell = 1}^t (1 - \overline Q_{\text{p}} a_{\ell} - Q_{\text{p}} u_{\ell}) \end{equation} The sets $G_j$ partition $[n]$, so $\sum_{j \in \text{f}} s_j = 1 - \sum_{\ell=1}^t a_{\ell} = 1 - u_1$. So by the AM-GM inequality, we have \begin{equation} \label{yy4} \ensuremath{\mathbf{E}}[d(i,\S)] \leq 1 + \left( 1 - \overline Q_{\text{f}} \frac{\overline{ u_1}}{m} \right)^m \prod_{\ell = 1}^t (1 - \overline Q_{\text{p}} a_{\ell}) + \left( \overline Q_{\text{f}} \bigl( 1 - \frac{\overline{u_1}}{m} \bigr) \right)^m \prod_{\ell = 1}^t (1 - \overline Q_{\text{p}} a_{\ell} - Q_{\text{p}} u_{\ell}). \end{equation} The claim follows as $a_{\ell} = u_{\ell} - u_{\ell+1}$. \end{proof} Lemma~\ref{prop31} gives an upper bound on $\ensuremath{\mathbf{E}}[d(i,\S)]$ for a fixed distribution on $Q_{\text{f}}, Q_{\text{p}}$ and for fixed values of the parameters $m, u_1, \dots, u_t$. By a computer search, along with a number of numerical tricks, we can obtain an upper bound over all values for $m$ and all possible sequences $u_1, \dots, u_t$ satisfying $1 = u_1 \geq u_2 \geq \dots \geq u_t$. This gives the following result: \begin{pseudo-theorem} \label{thm:k-center-1.592} Suppose that $Q_{\text{f}}, Q_{\text{p}}$ has the following distribution: $$ ( Q_{\text{f}}, Q_{\text{p}}) = \begin{cases} (0.4525, 0) & \text{with probability $p = 0.773436$} \\ (0.0480, 0.3950) & \text{with probability $1-p$} \end{cases}. $$ Then for all $j \in \mathcal{C}$ we have $d(j,\S) \leq 3 r_j$ with probability one, and $\ensuremath{\mathbf{E}}[d(j,\S)] \leq 1.592 r_j$. \end{pseudo-theorem} We call this a ``pseudo-theorem'' because some of the computer calculations used double-precision floating point for convenience, without carefully tracking rounding errors. In principle, this could have been computed using rigorous numerical analysis, giving a true theorem. Since there are a number of technical complications in this calculation, we defer the details to Appendix~\ref{proof1592}. \section{Lower bounds on approximating chance $k$-coverage} \label{lb-sec} We next show lower bounds for the chance $k$-coverage problems of Sections~\ref{chance-sec} and \ref{sec2}. These constructions are adapted from lower bounds for approximability of $k$-median \cite{guha1999greedy}, which in turn are based on the hardness of set cover. Formally, a set cover instance consists of a collection of sets $\mathcal B = \{S_1, \dots, S_m \}$ over a ground set $[n]$. For any set $X \subseteq [m]$ we define $S_X = \bigcup_{i \in X} S_i$. The goal is to select a collection $X \subseteq [m]$ of minimum size such that $S_X = [n]$. The minimum value of $|X|$ thus obtained is denoted by $\text{OPT}$. We quote a result of Moshovitz \cite{moshkovitz} on the inapproximability of set cover. \begin{theorem}[\cite{moshkovitz}] \label{hardness-th1} Assuming $\text{P} \neq \text{NP}$, there is no polynomial-time algorithm which guarantees a set-cover solution $X$ with $S_X = [n]$ and $|X| \leq (1 - \epsilon) \ln n \times \text{OPT}$, where $\epsilon > 0$ is any constant. \end{theorem} We will need a simple corollary of Theorem~\ref{hardness-th1}, which is a (slight reformulation) of the hardness of approximating max-coverage. \begin{corollary} \label{hardprop1} Assuming $\text{P} \neq \text{NP}$, there is no polynomial-time algorithm which guarantees a set-cover solution $X$ with $|X| \leq \text{OPT}$ and $\bigl | S_X \bigr | \geq c n$ for any constant $c > 1 - 1/e$. \end{corollary} \begin{proof} Suppose for contradiction that $\mathcal A$ is such an algorithm. We will repeatedly apply $\mathcal A$ to solve residual instances, obtaining a sequence of solutions $X_1, X_2, \dots, $. Specifically, for iterations $i \geq 1$, we define $U_i = [n] - \bigcup_{j < i} S_{X_j}$ and $\mathcal B_i = \{S \cap U_i \mid S \in \mathcal B \}$, and we let $X_i$ denote the output of $\mathcal A$ on instance $\mathcal B_i$. Each $\mathcal B_i$ has a solution of cost at most $\text{OPT}$. Thus, $|X_i| \leq \text{OPT}$ and $|U_i \cap S_{X_i} | \geq c |U_i|$. Thus $|U_{i+1}| = |U_i - S_{X_i}| \leq (1-c) |U_i|$. So, for $s = \lceil 1 + \frac{\ln n}{\ln(\frac{1}{1-c})} \rceil$ we have $U_s = \emptyset$, and so the set $X = X_1 \cup \dots \cup X_s$ solves the original set-cover instance $\mathcal B$, and $|X| \leq \sum_{i=1}^s |X_i| \leq (1 + \frac{\ln n}{\ln (\frac{1}{1 - c})}) \text{OPT}$. Since $c > 1 - 1/e$, we have $\frac{1}{\ln (\frac{1}{1 - c})} \leq (1 - \Omega(1))$, which contradicts Theorem~\ref{hardness-th1}. \end{proof} \begin{theorem} \label{ath0} Assuming $\text{P} \neq \text{NP}$, there is a family of homogeneous chance $k$-coverage instances, with a feasible demand of $p_j = r_j = 1$ for all clients $j$, such that no polynomial-time algorithm can guarantee a distribution $\Omega$ with either of the following: \begin{enumerate} \item $\forall j \ensuremath{\mathbf{E}}_{\S \sim \Omega} [d(j,\S)] \leq c r_j$ for constant $c < 1 + 2/e$ \item $\forall j \Pr_{\S \sim \Omega} [ d(j, \S) < 3 r_j ] \geq c p_j$ for constant $c > 1 - 1/e$ \end{enumerate} In particular, approximation constants in Theorem~\ref{ud1}, Proposition~\ref{ud2}, and Theorem~\ref{simple-bnd-thm0} cannot be improved. \end{theorem} \begin{proof} Consider a set cover instance $\mathcal B = \{S_1, \dots, S_m \}$. We begin by guessing the value $\text{OPT}$ (there are at most $m$ possibilities, so this can be done in polynomial time). We define a $k$-center instance with $k = \text{OPT}$ and with disjoint client and facility sets, where $\mathcal{F}$ is identified with $[m]$ and $\mathcal{C}$ is identified with $[n]$. We define $d$ by $d(i,j) = 1$ if $j \in S_i$ and $d(i,j) = 3$ otherwise. If $X$ is an optimal solution to $\mathcal B$ then $d(j,X) \leq 1$ for all points $j \in \mathcal{C}$. So there exists a (deterministic) distribution with feasible demand parameters $p_j = 1, r_j = 1$. Also, note that $d(j, \S) \in \{1, 3 \}$ with probability one for any client $j$. Thus, if $j$ satisfies the second property $\Pr_{\S \sim \Omega} [ d(j, \S) < 3 r_j ] \geq c p_j$ for constant $c > 1 - 1/e$, then it satisfies $\ensuremath{\mathbf{E}}[d(j, \S)] \leq 1 + 2 (1 - c p_j) = 3 - 2 c = (1 + 2/e - \Omega(1)) r_j$. So it will satisfy the first property as well. So it suffices to show that no algorithm $\mathcal A$ can satisfy the first property. Suppose that $\mathcal A$ does satisfy the first property. The resulting solution set $\S \subseteq \mathcal F$ can be regarded as a solution $X$ to the set cover instance, where $d(j, \S) = 1 + 2 [[ j \notin S_X ]]$. Thus $$ \sum_{j \in [n]} d(j, \S) = |S_X| + 3 (n - |S_X|), $$ and so $|S_X| = \frac{3 n - \sum_{j \in [n]} d(j, \S)}{2}$. As $\ensuremath{\mathbf{E}}[d(j, \S)] \leq c r_j = c$ for all $j$, this implies that $\ensuremath{\mathbf{E}}[ |S_X| ] \geq \frac{(3 - c) n}{2}$. After an expected constant number of repetitions of this process we can ensure that $|S_X| \geq c' n$ for some constant $c' > \frac{3 - (1 + 2/e)}{2} = 1 - 1/e$. This contradicts Corollary~\ref{hardprop1}. \end{proof} A slightly more involved construction applies to the homogeneous SCC setting. \begin{theorem} \label{ath1} Assuming $\text{P} \neq \text{NP}$, there is a family of homogeneous SCC chance $k$-coverage instances, with a feasible demand of $p_j = r_j = 1$ for all $j$, such that no polynomial-time algorithm can guarantee a distribution $\Omega$ with either of the following: \begin{enumerate} \item $\forall j \ensuremath{\mathbf{E}}_{\S \sim \Omega} [d(j,\S)] \leq c r_j$ for constant $c < 1 + 1/e$ \item $\forall j \Pr_{\S \sim \Omega} [ d(j, \S) < 2 r_j ] \geq c p_j$ for constant $c > 1 - 1/e$ \end{enumerate} In particular, the approximation constants in Theorem~\ref{ud1} and Proposition~\ref{ud2} cannot be improved for SCC instances, and the approximation factor $1.592$ in Pseudo-Theorem~\ref{thm:k-center-1.592} cannot be improved below $1 + 1/e$. \end{theorem} \begin{proof} Consider a set cover instance $\mathcal B = \{S_1, \dots, S_m\}$, where we have guessed the value $\text{OPT} = k$. We define a $k$-center instance as follows. For each $i \in [m]$, we create an item $v_i$ and for each $j \in [n]$ we create $t = n^2$ distinct items $w_{j,1}, \dots, w_{j,t}$. We define the distance by setting $d(v_i, w_{j,t}) = 1$ if $j \in S_i$ and $d(v_i, v_{i'}) = 1$ for all $i, i' \in [m]$, and $d(x,y) = 2$ for all other distances. This problem size is polynomial (in $m,n$), and so $\mathcal A$ runs in time $\text{poly}(m,n)$. If $X$ is an optimal solution to the set cover instance, the corresponding set $\S = \{ v_i \mid i \in X \}$ satisfies $d(j, \S) \leq 1$ for all $j \in \mathcal{C}$. So the demand vector $p_j = r_j = 1$ is feasible. Also, note that that $d(j, \S) \in \{0, 1, 2 \}$ with probability one for any $j$. Thus, if $j$ satisfies the second property $\Pr_{\S \sim \Omega} [ d(j, \S) < 2 r_j ] \geq c p_j$ for constant $c > 1 - 1/e$, then it satisfies $\ensuremath{\mathbf{E}}[d(j, \S)] \leq 1 + 1 (1 - c p_j) = (1 + 1/e - \Omega(1)) r_j$. So it will satisfy the first property as well. So it suffices to show that no algorithm $\mathcal A$ can satisfy the first property. Suppose that algorithm $\mathcal A$ satisfies the first property. From the solution set $\S$, we construct a corresponding set-cover solution by $X = \{ i \mid v_i \in \S \}$. For $w_{j,\ell} \notin \S$, we can observe that $d( w_{j, \ell}, \S) = 1 + [[ j \notin S_X ]]$. Therefore, we have \begin{align*} \sum_{j \in [n]} \sum_{\ell=1}^t d(w_{j,\ell}, \S) &\geq \sum_{j, \ell: w_{j,\ell} \notin \S} (1 + [[j \notin S_X]]) \geq \sum_{j, \ell} (1 + [[j \notin S_X]]) - 2 |\S| \geq n^2 (2 n - |S_X|) - 2 k, \end{align*} and so $|S_X| \geq 2 n - \frac{\sum_{j, \ell} d(w_{j,\ell}, \S)}{n^2} - 2 k/n^2.$ Taking expectations and using our upper bound on $\ensuremath{\mathbf{E}}[d(j,\S)]$, we have $\ensuremath{\mathbf{E}}[ |S_X| ] \geq 2 n - c n - 2 k/n^2 \geq (2 - c) n - 2/n$. Thus, for $n$ sufficiently large, after an expected constant number of repetitions of this process we get $|S_X| \geq (2 - c - o(1))n \geq (1 - 1/e + \Omega(1)) n$. This contradicts Corollary~\ref{hardprop1}. \end{proof} \section{Approximation algorithm for $\ensuremath{\mathbf{E}}[d(j,\S)]$} \label{approx-alg} In the chance $k$-coverage problem, our goal is to achieve certain fixed values of $d(j,\S)$ with a certain probability. In this section, we consider another criterion for $\Omega$; we wish to achieve certain values for the expectation $\ensuremath{\mathbf{E}}_{\S \sim \Omega}[d(j, \S)]$. We suppose we are given values $t_j$ for every $j \in C$, such that the target distribution $\Omega$ satisfies \begin{equation} \label{stt5} \ensuremath{\mathbf{E}}_{\S \sim \Omega} [d(j, \S)] \leq t_j. \end{equation} In this case, we say that the vector $t_j$ is \emph{feasible.} As before, if all the values of $t_j$ are equal to each other, we say that the instance is \emph{homogeneous}. We show how to leverage any approximation algorithm for $k$-median with approximation ratio $\alpha$, to ensure our target distribution $\tilde \Omega$ will satisfy $$ \ensuremath{\mathbf{E}}_{\S \sim \tilde \Omega} [d(j, \S)] \leq (\alpha + \epsilon) t_j. $$ More specifically, we need an approximation algorithm for weighted $k$-median. In this setting, we have a problem instance $\mathcal{I} = \mathcal{F}, \mathcal{C}, d$ along with non-negative weights $w_j$ for $j \in \mathcal{C}$, and we wish to find $\S \in \binom{\mathcal{F}}{k}$ minimizing $\sum_{j \in \mathcal{C}} w_j d(j, \S).$ (Nearly all approximation algorithms for ordinary $k$-median can be easily adapted to the weighted setting, for example, by replicating clients.) If we fix an approximation algorithm $\mathcal A$ for (various classes of) weighted $k$-median, then for any problem instance $\mathcal{I}$ we define $$ \alpha_{\mathcal{I}} = \sup_{\text{weights $w$}} \frac{ \sum_{j \in \mathcal{C}} w_j d(j, \mathcal A(\mathcal{I}, w)) }{ \min_{\S \in \binom{\mathcal{F}}{k}} \sum_{j \in \mathcal{C}} w_j d(j, \S) }. $$ We first show how to use the $k$-median approximation algorithm to achieve a set $\S$ which ``matches'' the desired distances $t_j$: \begin{proposition} \label{kmedianprop1} Given a weighted instance $\mathcal{I}$ and a parameter $\epsilon > 0$, there is a polynomial-time algorithm to produce a set $\S \in \binom{\mathcal{F}}{k}$ satisfying: \begin{enumerate} \item $\sum_{j \in \mathcal{C}} w_j \frac{d(j, \S)}{t_j} \leq (\alpha_{\mathcal{I}} + O(\epsilon)) \sum_{j \in \mathcal{C}} w_j$, \item Every $j \in \mathcal{C}$ has $d(j, \S) \leq n t_j/\epsilon$. \end{enumerate} \end{proposition} \begin{proof} We assume $\alpha_{\mathcal{I}} = O(1)$, as constant-factor approximation algorithms for $k$-median exist. By rescaling $w$, we assume without loss of generality that $\sum_{j \in \mathcal{C}} w_j = 1$. By rescaling $\epsilon$, it suffices to show that $d(j, \S) \leq O(n t_j/\epsilon)$. Let us define the weight vector $z_j = \frac{\epsilon/n + w_j}{t_j}$. Letting $\Omega$ be a distribution satisfying (\ref{stt5}), we have \begin{align*} \ensuremath{\mathbf{E}}_{\S \sim \Omega} \bigl[ \sum_{j \in \mathcal{C}} z_j d(j, \S) \bigr] &= \sum_{j \in \mathcal{C}} z_j t_j \leq \sum_{j \in \mathcal{C}} (\frac{\epsilon}{n t_j} + \frac{w_j}{t_j}) t_j = \epsilon | \mathcal{C} |/n + \sum_{j \in \mathcal{C}} w_j = 1 + \epsilon. \end{align*} In particular, there exists some $\S \in \binom{\mathcal{F}}{k}$ with $\sum_{j \in \mathcal{C}} z_j d(j, \S) \leq 1 + \epsilon$. When we apply algorithm $\mathcal A$ with weight vector $z$, we thus get a set $\S \in \binom{\mathcal{F}}{k}$ with $\sum_{j \in \mathcal{C}} z_j d(j, \S) \leq \alpha_{\mathcal{I}} (1 + \epsilon)$. We claim that this set $\S$ satisfies the two conditions of the theorem. First, we have \begin{align*} \sum_{j \in \mathcal{C}} \frac{w_j d(j, \S)}{t_j} &\leq \sum_{j \in \mathcal{C}} z_j d(j,\S) \leq \alpha_{\mathcal{I}} (1 + \epsilon) \leq (\alpha_{\mathcal{I}} + O(\epsilon)) \sum_j w_j. \end{align*} Next, for any given $j \in \mathcal{C}$, we have \[ \frac{d(j, \S)}{t_j} \leq d(j, \S) z_j (n/\epsilon) \leq (n/\epsilon) \sum_{w \in \mathcal{C}} z_w d(w,\S) \leq (n/\epsilon) \alpha_{\mathcal{I}} (1 + \epsilon) \leq O(n/\epsilon). \qedhere \] \end{proof} \begin{theorem} \label{main-approx-thm} There is an algorithm which takes as input an instance $\mathcal{I}$, a parameter $\epsilon > 0$ and a feasible vector $t_j$, runs in time $\text{poly}(n,1/\epsilon)$, and returns an explicitly enumerated distribution $\tilde \Omega$ with support size $n$ and $\ensuremath{\mathbf{E}}_{S \sim \tilde \Omega}[d(j, \S)] \leq (\alpha_{\mathcal{I}} + \epsilon) t_j$ for all $j \in \mathcal{C}$. \end{theorem} \begin{proof} We assume without loss of generality that $\epsilon \leq 1$; by rescaling $\epsilon$ it suffices to show that $\ensuremath{\mathbf{E}}[d(j,\S)] \leq (\alpha_\mathcal{I} + O(\epsilon)) t_j$. We begin with the following Algorithm~\ref{algo:app0}, which uses a form of multiplicative weights update with repeated applications of Proposition~\ref{kmedianprop1}. \begin{algorithm}[H] \begin{algorithmic}[1] \FOR{$\ell = 1, \dots, r = \frac{n \ln n}{\epsilon^3}$} \STATE Let $X_{\ell} \in \binom{\mathcal{F}}{k}$ be the resulting of applying the algorithm of Proposition~\ref{kmedianprop1} with parameters $\epsilon, t_j$ and weight vector $w$ given by \vspace{-0.16in} $$ w_j = \exp \Bigl( \epsilon^2 \sum_{s = 1}^{\ell-1} \frac{d(j,X_s)}{n t_j} \Bigr), $$ \vspace{-0.12in} \ENDFOR \STATE Set $\tilde \Omega'$ to be the uniform distribution on $X_1, \dots, X_r$ \end{algorithmic} \caption{Approximation algorithm for $\ensuremath{\mathbf{E}}[d(j,\S)]$: first phase} \label{algo:app0} \end{algorithm} Let us define $\phi = \epsilon^2/n$. For each iteration $\ell = 1, \dots, r+1$ let $u_j^{(\ell)} = \phi d(j,X_{\ell}) / t_j$, and let $w^{(\ell)}_j = e^{\sum_{s=1}^{\ell-1} u_j^{(s)}}$ denote the weight vector. Proposition~\ref{kmedianprop1} ensures that $u^{(\ell)}_j \leq \epsilon$, and thus $e^{u_j^{(\ell)}} \leq 1 + \frac{e^\epsilon - 1}{\epsilon} u_j^{(\ell)} \leq 1 + (1+\epsilon) u_j^{(\ell)}$, as well as ensuring that $\sum_j w_j^{(\ell)} u_j^{(\ell)} \leq \phi (\alpha_I + O(\epsilon)) \sum_{j} w_j^{(\ell)}$. Now let $\Phi_{\ell} = \sum_{j \in \mathcal{C}} w^{(\ell)}_j$. Note that $\Phi_1 =n$, and for each $\ell \geq 1$, we have \begin{align*} \Phi_{\ell+1} &= \sum_{j \in \mathcal{C}} w^{(\ell)}_j e^{u_j^{(\ell)}} \leq \sum_{j \in \mathcal{C}} w^{(\ell)}_j \bigl( 1 + (1 + \epsilon) u_j^{(\ell)} \bigr) \leq \sum_{j \in \mathcal{C}} w^{(\ell)}_j + (1 + \epsilon) \sum_{j \in \mathcal{C}} w^{(\ell)}_j u_j^{(\ell)} \\ &\leq \Phi_{\ell} \bigl( 1 + (1 + \epsilon) \phi (\alpha_{\mathcal{I}} + O(\epsilon)) \bigr) \leq \Phi_{\ell} e^{ \phi (\alpha_{\mathcal{I}} + O(\epsilon))} \end{align*} This recurrence relation implies that $\Phi_{\ell} \leq n e^{ (\ell-1) \phi (\alpha_{\mathcal{I}} + O(\epsilon)) }$. Since $w^{(r+1)}_j \leq \Phi^{r+1}$, this implies $$ \sum_{\ell=1}^r \phi d(j,X_{\ell}) / t_j = \ln w_j^{(r+1)} \leq \ln \Phi_{r+1} \leq \ln n + r \phi ( \alpha_{\mathcal{I}} + O(\epsilon)). $$ or equivalently, $$ \sum_{\ell=1}^r \frac{d(j,X_{\ell})}{r} = t_j \Bigl( \frac{\ln n}{r \phi} + (\alpha_{\mathcal{I}} + O(\epsilon)) \Bigr) $$ As $r = \frac{n \ln n}{\epsilon^3} = \frac{\ln n}{\epsilon \phi}$, we thus have $\frac{\sum_{\ell=1}^r d(j,X_{\ell})}{r} \leq (\alpha_{\mathcal{I}} + O(\epsilon)) t_j$. Thus, the distribution $\tilde \Omega'$ satisfies \begin{equation} \label{bv13} \forall j \in \mathcal C \qquad \ensuremath{\mathbf{E}}_{\S \sim \tilde \Omega'}[d(j, \S)] \leq (\alpha_I + O(\epsilon)) t_j. \end{equation} Now the distribution $\tilde \Omega'$ satisfies the condition on $\ensuremath{\mathbf{E}}[d(j, \S)]$, but its support is too large. We can reduce the support size to $|\mathcal{C}|$ by moving in the nullspace of the $| \mathcal{C} |$ linear constraints (\ref{bv13}). \end{proof} Byrka et al. \cite{byrka} have shown a $2.675 + \epsilon$-approximation algorithm for $k$-median, which automatically gives a $2.675 + \epsilon$-approximation algorithm for $k$-lottery as well. Some special cases of $k$-median have more efficient approximation algorithms. For instance, Cohen-Addad, Klein \& Mathieu \cite{cohen2016local} gives a PTAS for $k$-median problems derived from a planar graph, and Ahmadian et al. \cite{svensson} gives a $2.633 + \epsilon$-approximation for Euclidan distances. These immediately give approximation algorithms for the corresponding $k$-lotteries. We also note that, by Theorem~\ref{ath0}, one cannot obtain a general approximation ratio better than $1 + 2/e$ (or $1 + 1/e$ in the SCC setting). \section{Determinizing a $k$-lottery} \label{determinism-sec} Suppose that we have a set of feasible weights $t_j$ such some $k$-lottery distribution $\Omega$ satisfies $\ensuremath{\mathbf{E}}_{\S \sim \Omega }[d(j, \S)] \leq t_j$; let us examine how to find a \emph{single, deterministic} set $\S$ with $d(j, \S) \approx t_j$. We refer to this as the problem of \emph{determinizing} the lottery $\Omega$. Note that this can be viewed as a converse to the problem considered in Section~\ref{sec2}. We will see that, in order to obtain reasonable approximation ratios, we may need $|\S|$ to be significantly larger than $k$. We thus define an \emph{$(\alpha, \beta)$-determinization} to be a set $\S \in \binom{\mathcal{F}}{k'}$ with $k' \leq \alpha k$ and $d(j, \S) \leq \beta t_j$ for all $j \in \mathcal{C}$. We emphasize that we cannot necessarily obtain $(1,1)$-determinizations, even with unbounded computational resources. The following simple example illustrates the tradeoff between parameters $\alpha$ and $\beta$: \begin{observation} \label{lb-obs} Let $\alpha, \beta, k \geq 1$. If $\beta < \frac{\alpha k + 1}{(\alpha - 1) k + 1}$, there is a homogeneous SCC instance for which no $(\alpha, \beta)$-determinization exists. \end{observation} \begin{proof} Let $k' = \alpha k$ and consider a problem instance with $\mathcal{F} = \mathcal{C} = \{1, \dots, k'+1 \}$, and $d(i,j) = 1$ for every distinct $i,j$. Clearly, every $\S \in \binom{\mathcal{F}}{k'}$ satisfies $\min_j d(j, \S) = 1$. When $\Omega$ is the uniform distribution on $\binom{\mathcal{F}}{k}$, we have $\ensuremath{\mathbf{E}}[d(j, \S)] = 1 - \frac{k}{k'+1}$. Thus $t_j = \frac{k}{k'+1}$ is feasible and therefore $\beta \geq \frac{1}{1 - \frac{k}{k'+1}} = \frac{ \alpha k + 1}{(\alpha - 1) k + 1}$. \end{proof} In particular, when $\alpha = 1$ we must have $\beta \geq k+1$ and when $k \rightarrow \infty$, we must have $\beta \gtrsim \frac{\alpha}{\alpha - 1}$. We examine three main regimes for the parameters $(\alpha, \beta)$: (1) the case where $\alpha, \beta$ are scale-free constants; (2) the case where $\beta$ is close to one, in which case $\alpha$ must be of order $\log n$; (3) the case where $\alpha = 1$, in which case $\beta$ must be order $k$. Our determinization algorithms for the first two cases will based on the following LP denoted $\mathcal P_{\text{expectation}}$, defined in terms of fractional vectors $b_i, a_{i,j}$ where $i$ ranges over $\mathcal{F}$ and $j$ ranges over $\mathcal{C}$: \begin{enumerate} \item[(A1)] $\forall j \in \mathcal{C}, \qquad \sum_{i \in \mathcal{F}} a_{i,j} d(i,j) \leq t_j$, \item[(A2)] $\forall j \in \mathcal{C}, \qquad \sum_{i \in \mathcal{F}} a_{i,j} = 1$, \item[(A3)] $\forall i \in \mathcal{F}, y \in \mathcal{C}, \qquad 0 \leq a_{i,j} \leq b_i$, \item[(A4)] $\forall i \in \mathcal{F}, \qquad 0 \leq b_i \leq 1$, \item[(A5)] $\sum_{i \in \mathcal{F}} b_i \leq k$. \end{enumerate} \begin{theorem} If $t_j$ is feasible, then $\mathcal P_{\text{expectation}}$ has a fractional solution. \end{theorem} \begin{proof} Let $\Omega$ be a probability distribution with $\ensuremath{\mathbf{E}}[d(j,\S)] \leq t_j$. For any draw $\S \sim \Omega$, define random variable $Z_{j}$ to be the facility of $\S$ matched by $j$. Now consider the fractional vector defined by $$ b_i = \Pr_{\S \sim \Omega} [ i \in \S ], \qquad \qquad a_{i,j} = \Pr_{\S \sim \Omega} [ Z_{j} = i ] $$ We claim that this satisfies (A1) --- (A5). For (A1), we have $$ \ensuremath{\mathbf{E}}[d(j, \S)] = \ensuremath{\mathbf{E}}[d(j, Z_{j})] = \sum_{i \in \mathcal{F}} d(i,j) \Pr[Z_{j} = i] = \sum_{i \in F} d(i,j) a_{i,j} \leq t_j. $$ For (A2), note that $\sum_i \Pr[Z_j = i] = 1$. For (A3), note that $Z_j = i$ can only occur if $i \in \S$. (A4) is clear, and (A5) holds as $|\S| = k$ with probability one. \end{proof} We next describe upper and lower bounds for these three regimes. \subsection{The case where $\alpha, \beta$ are scale-free constants.} For this regime (with all parameters independent of problem size $n$ and $k$), we may use the following Algorithm~\ref{algo:det1}, which is based on greedy clustering using a solution to $\mathcal P_{\text{expectation}}$. \begin{algorithm}[H] \begin{algorithmic}[1] \STATE Let $a,b$ be a solution to $\mathcal P_{\text{expectation}}$. \STATE For every $j \in \mathcal{C}$, select $r_j \geq 0$ to be minimal such that $\sum_{i \in B(j, r_j)} a_{i,j} \geq 1/\alpha$ \STATE By splitting facilities, form a set $F_j \subseteq B(j, r_j)$ with $b(F_j) = 1/\alpha$. \STATE Set $C' = \textsc{GreedyCluster}(F_j, \theta(j) + r_j)$ \STATE Output solution set $\S = \{ V_j \mid j \in C' \}$. \end{algorithmic} \caption{$(\alpha,\beta)$-determinization algorithm} \label{algo:det1} \end{algorithm} Step (3) is well-defined, as (A3) ensures that $b(B(j, r_j)) \geq \sum_{i \in B(j,r_j)} a_{i,j} \geq 1/\alpha$. Let us analyze the resulting approximation factor $\beta$. \begin{proposition} \label{rybound} Every client $j \in \mathcal{C}$ has $r_j \leq \frac{\alpha t_j - \theta(j)}{\alpha - 1}$. \end{proposition} \begin{proof} Let $s = \frac{\alpha t_j - \theta(j)}{\alpha - 1}$. It suffices to show that $$ \sum_{i \in \mathcal{F}, d(i,j) > s} a_{i,j} \leq 1 - 1/\alpha. $$ As $d(i,j) \geq \theta(j)$ for all $i \in \mathcal{F}$, we have {\allowdisplaybreaks \begin{align*} \sum_{\substack{i \in \mathcal{F} \\ d(i,j) > s}} a_{i,j} &\leq \sum_{\substack{i \in \mathcal{F} \\ d(i,j) > s}} a_{i,j} \frac{d(i,j) - \theta(j)}{s - \theta(j)} \leq \sum_{i \in \mathcal{F}} a_{i,j} \frac{d(i,j) - \theta(j)}{s - \theta(j)} = \frac{\sum_{i \in \mathcal{F}} a_{i,j} d(i,j) - \theta(j) \sum_{i \in \mathcal{F}} a_{i,j}}{s-\theta(j)} \\ &\leq \frac{t_j - \theta(j)}{s-\theta(j)} = 1 - 1/\alpha, \qquad \text{by (A1), (A2).} \qedhere \end{align*} } \end{proof} \begin{theorem} \label{algo:det1-result} Algorithm~\ref{algo:det1} gives an $(\alpha, \beta)$-determinization with the following parameter $\beta$: \begin{enumerate} \item In the general setting, $\beta = \max(3, \frac{2 \alpha}{\alpha - 1})$. \item In the SCC setting, $\beta = \frac{2 \alpha}{\alpha - 1}$. \end{enumerate} \end{theorem} \begin{proof} We first claim that the resulting set $\S$ has $|\S| \leq \alpha k$. The algorithm opens at most $|C'|$ facilities. The sets $F_j$ are pairwise disjoint for $j \in C'$ and $b(F_j) = 1/\alpha$ for $j \in C'$. Thus $\sum_{j \in C'} b(F_j) = |C'|/\alpha$. On the other hand, $b(\mathcal{F}) = k$, and so $k \geq |C'|/\alpha$. Next, consider some $j \in \mathcal{C}$; we want to show that $d(j, \S) \leq \beta t_j$. By Observation~\ref{greedy-prop}, there is $z \in C'$ with $F_j \cap F_z \neq \emptyset$ and $\theta(z) + r_z \leq \theta(j) + r_j$. Thus $d(j, \S) \leq d(z,\S) + d(z,i) + d(j,i)$ where $i \in F_j \cap F_z$. Step (5) ensures $d(z,\S) = \theta(z)$. We have $d(z,i) \leq r_z$ and $d(i,j) \leq r_j$ since $i \in F_j \subseteq B(j, r_j)$ and $i \in F_z \subseteq B(z, r_z)$. So $$ d(j, \S) \leq \theta(z) + r_z + r_j \leq 2 r_j + \theta(j). $$ By Proposition~\ref{rybound}, we therefore have \begin{equation} \label{hhg1} d(j, \S) \leq \frac{2 \alpha t_j - 2 \theta(j)}{\alpha - 1} + \theta(j) = \frac{2 \alpha t_j}{\alpha-1} + \frac{\alpha - 3}{\alpha - 1} \theta(j) \end{equation} This immediately shows the claim for the SCC setting where $\theta(j)= 0$. In the general setting, for $\alpha \leq 3$, the second coefficient in the RHS of (\ref{hhg1}) is non-positive and hence the RHS is at most $\frac{2 \alpha t_j}{\alpha-1}$ as desired. When $\alpha \geq 3$, then in order for $t$ to be feasible we must have $t_j \geq \theta(j)$; substituting this upper bound on $\theta(j)$ into (\ref{hhg1}) gives \[ d(j, \S) \leq \frac{2 \alpha t_j}{\alpha-1} + \frac{\alpha - 3}{\alpha - 1} t_j = 3 t_j \qedhere \] \end{proof} We note that these approximation ratios are, for $\alpha$ close to $1$, within a factor of $2$ compared to the lower bound of Observation~\ref{lb-obs}. As $\alpha \rightarrow \infty$, the approximation ratio approaches to limiting values $3$ (or $2$ in the SCC setting). \subsection{The case of small $\beta$} We now consider what occurs when $\beta$ becomes smaller than the critical threshold values $3$ (or $2$ in the SCC setting). We show that in this regime we must take $\alpha = \Omega(\log n)$. Of particular interest is the case when $\beta$ approaches $1$; here, in order to get $\beta = 1 +\epsilon$ for small $\epsilon$ we show it is necessary and sufficient to take $\alpha = \Theta( \frac{\log n}{\epsilon})$. \begin{proposition} \label{det1prop} For any $\epsilon < 1/2$, there is a randomized polynomial-time algorithm to obtain a $(\frac{3 \log n}{\epsilon}, 1 + \epsilon)$ determinization. \end{proposition} \begin{proof} First, let $a,b$ be a solution to $\mathcal P_{\text{expectation}}$. Define $p_i = \min(1, \frac{2 \log n}{\epsilon} b_i)$ for each $i \in \mathcal{F}$ and form $\S = \textsc {DepRound}(p)$. Observe then that $|\S| \leq \lceil \sum_i p_i \rceil \leq \lceil \frac{2 \log n}{\epsilon} \sum b_i \rceil \leq 1 + \frac{2 k \log n}{\epsilon} \leq \frac{3 k \log n}{\epsilon}$. For $j \in \mathcal{C}$, define $A = B_{j, (1+\epsilon) t_j}$. Let us note that, by properties (A3), (A1) and (A2), we have $$ \sum_{i \in A} b_i \geq \sum_{i \in A} a_{i,j} = 1 - \sum_{i: d(i,j) > (1+\epsilon) t_j} a_{i,j} \geq 1 - \sum_{i: d(i,j) > (1+\epsilon) t_j} a_{i,j} \frac{d(i,j)}{(1+\epsilon) t_j} \geq 1 - \frac{1}{1 + \epsilon} $$ So by property (P3) of \textsc{DepRound}, and using the bound $\epsilon < 1/2$, have {\allowdisplaybreaks \begin{align*} \Pr[d(j, \S) > (1 + \epsilon) t_j] &= \Pr[ A \cap \S = \emptyset] \leq \prod_{i \in A} (1 - p_i) \leq \prod_{i \in A} e^{-\frac{2 \log n}{\epsilon} b_i} \leq e^{ \frac{-2 \log n}{\epsilon} (1 - \frac{1}{1+\epsilon})} \leq n^{-4/3} \end{align*} } A union bound over $j \in \mathcal{C}$ shows that solution set $\S$ satisfies $d(j, \S) \leq (1+\epsilon) t_j$ for all $j$ with high probability. \end{proof} The following shows matching lower bounds: \begin{proposition} \label{rr4a} \begin{enumerate} \item There is a universal constant $K$ with the following properties. For any $k \geq 1, \epsilon \in (0,1/3)$ there is some integer $N_{k,\epsilon}$ such that for $n > N_{k,\epsilon}$, there is a homogeneous SCC instance of size $n$ in which \emph{every} $(\alpha, 1+\epsilon)$-determinization satisfies $\alpha \geq \frac{K \log n}{\epsilon}.$ \item For each $\beta \in (1,2)$ and each $k \geq 1$, there is a constant $K'_{\beta, k}$ such that, for all $n \geq 1$, there is a homogeneous SCC instance of size $n$ in which \emph{every} $(\alpha, \beta)$-determinization satisfies $\alpha \geq K'_{\beta, k} \log n.$ \item For each $\beta \in (1,3)$ and each $k \geq 1$, there is a constant $K''_{\beta, k}$ such that, for all $n \geq 1$, there is a homogeneous instance of size $n$ in which \emph{every} $(\alpha, \beta)$-determinization satisfies $\alpha \geq K''_{\beta, k} \log n$ \end{enumerate} \end{proposition} \begin{proof} These three results are very similar, so we show the first one in detail and sketch the difference between the other two. Consider an Erd\H{o}s-R\'{e}nyi random graph $G \sim \mathcal G(n,p)$, where $p = 3 \epsilon / k$; note that $p \in (0,1)$. As shown by \cite{glebov} asymptotically almost surely the domination number $J$ of $G$ satisfies $J = \Omega( \frac{k \log n}{\epsilon})$. We construct a related instance with $\mathcal{F} = \mathcal{C} = [n]$, and where $d(i,j) = 1$ if $(i,j)$ is an edge, and $d(i,j) = 2$ otherwise. Note that if $X$ is not a dominating set of $G$, then some vertex of $G$ has distance at least $2$ from it; equivalently, $\max_j d(j,X) \geq 2$ for every set $X$ with $|X| < J$. Chernoff's bound shows that every vertex of $G$ has degree at least $u = 0.9 n p$ with high probability. Assuming this event has occured, we calculate $\ensuremath{\mathbf{E}}[d(j, \S)]$ where $\S$ is drawn from the uniform distribution on $\binom{\mathcal{F}}{k}$. Note that $d(j, \S) \leq 1$ if $j$ is a neighbor of $X$ and $d(j, \S) = 2$ otherwise, so $$ \ensuremath{\mathbf{E}}[d(j, \S)] \leq 1 + \frac{ \binom{n - u}{k}}{\binom{n}{k}} \leq 1+ e^{-0.9 p k} = 1 + e^{-2.7 \epsilon}. $$ Both the bound on the domination number and the minimum degree of $G$ hold with positive probability for $n$ sufficiently large (as a function of $k, \epsilon$). In this case, $t_j = 1 + e^{-2.7 \epsilon}$ is a feasible homogeneous demand vector. At the same time, every set $\S \in \binom{F}{J-1}$ satisfies $\min_{j \in \mathcal{C}} d(j, \S) \geq 2$. Thus, an $(\alpha, \beta)$-determinization cannot have $\alpha < \frac{J}{k} = \Theta(\frac{\log n}{\epsilon})$ and $\beta \leq \frac{2}{1 + e^{-2.7 \epsilon}}$. Note that $\frac{2}{1 + e^{-2.7 \epsilon}} \geq 1 + \epsilon$ for $\epsilon < 1/3$. Thus, whenever $\beta \leq 1 + \epsilon$, we have $\alpha \geq \Theta( \frac{\log n}{\epsilon} )$. For the second result, we use the same construction as above with $p = 1 - \tfrac{1}{2} (\lambda/2)^{1/k}$ where $\lambda = 2 - \beta$. A similar analysis shows that the vector $t_j = 1 + \lambda/2$ is feasible with high probability and $|J| \geq \Omega(k \log n)$ (where the hidden constant may depend upon $\beta, k$). Thus, unless $\alpha \geq \Omega(\log n)$, the approximation ratio achieved is $\frac{2}{1 + \lambda/2} \geq \beta$. The third result is similar to the second one, except that we use a random bipartite graph. The left-nodes are associated with $\mathcal{F}$ and the right-nodes with $\mathcal{C}$. For $i \in \mathcal{F}$ and $j \in \mathcal{C}$, we define $d(i,j) = 1$ if $(i,j)$ is an edge and $d(i,j) = 3$ otherwise. \end{proof} \subsection{The case of $\alpha = 1$} We finally consider the case $\alpha = 1$, that is, where the constraint on the number of open facilities is respected \emph{exactly}. By Observation~\ref{lb-obs}, we must have $\beta \geq k+1$ here. The following greedy algorithm gives a $(1, k+2)$-determinization, nearly matching this lower bound. \begin{algorithm}[H] \begin{algorithmic}[1] \STATE Initialize $\S = \emptyset$ \FOR{$\ell = 1, \dots, | \mathcal{F} |$} \STATE Let $\mathcal C_{\ell}$ denote the set of points $j \in \mathcal{C}$ with $d(j, \S) > (k+2) t_j$ \STATE If $\mathcal C_{\ell} = \emptyset$, then return $\S$. \STATE Select the point $j_{\ell} \in \mathcal C_{\ell}$ with the smallest value of $t_{j_{\ell}}$. \STATE Update $\S \leftarrow \S \cup \{ V_{j_{\ell}} \}$ \ENDFOR \end{algorithmic} \caption{$(1, k+2)$-determinization algorithm} \label{algo:det2} \end{algorithm} \begin{theorem} If the values $t_j$ are feasible, then Algorithm~\ref{algo:det2} outputs a $(1, k+2)$-determinization in $O(|\mathcal \mathcal{F}| | \mathcal \mathcal{C} |)$ time. \end{theorem} \begin{proof} For the runtime bound, we first compute $V_j$ for each $j \in \mathcal{C}$; this requires $O( |\mathcal F| |\mathcal C|)$ time upfront. When we update $\S$ at each iteration $\ell$, we update and maintain the quantities $d(j,\S)$ quantities by computing $d(j, V_{j_{\ell}})$ for each $j \in \mathcal{C}$. This takes $O(|\mathcal{C}|)$ time per iteration. To show correctness, note that if this procedure terminates at iteration $\ell$, we have $\mathcal C_{\ell} = \emptyset$ and so every point $j \in \mathcal{C}$ has $d(j, \S) \leq (k+2) t_j$. The resulting set $\S$ at this point has cardinality $\ell - 1$. So we need to show that the algorithm terminates before reaching iteration $\ell = k+2$. Suppose not; let the resulting points be $j_1, \dots, j_{k+1}$ and for each $\ell = 1, \dots, k+1$ let $w_{\ell} = t_{j_\ell}$. Because $j_{\ell}$ is selected to minimze $t_{j_{\ell}}$ we have $w_1 \leq w_2 \leq \dots \leq w_{k+1}$. Now, let $\Omega$ be a $k$-lottery satisfying $\ensuremath{\mathbf{E}}_{\S \sim \Omega}[ d(j, \S)] \leq t_j$ for every $j \in \mathcal C$, and consider the random process of drawing $\S$ from $\Omega$. Define the random variable $D_{\ell} = d(j_{\ell}, \S)$ for $\ell =1, \dots, k+1$. For any such $\S$, by the pigeonhole principle there must exist some pair $j_{\ell}, j_{\ell'}$ with $1 \leq \ell < \ell' \leq k+1$ which are both matched to a common facility $i \in \S$, that is $$ D_{\ell} = d(j_{\ell}, \S) = d(j_{\ell}, i), D_{\ell'} = d(j_{\ell'}, \S) = d(j_{\ell'}, i). $$ By the triangle inequality, $$ d(j_{\ell'}, V_{j_{\ell}}) \leq d(j_{\ell'}, i) + d(i, j_{\ell}) + d(j_{\ell}, V_{j_{\ell}}) = D_{\ell'} + D_{\ell} + \theta(j_{\ell}) $$ On the other hand, $j_{\ell'} \in C_{\ell'}$ and yet $V_{j_{\ell}}$ was in the partial solution set $\S$ seen at iteration $\ell'$. Therefore, it must be that $$ d(j_{\ell'}, V_{j_{\ell}}) > (k+2) t_{j_{\ell'}} = (k+2) w_{\ell'} $$ Putting these two inequalities together, we have shown that $$ D_{\ell} + D_{\ell'} + \theta(j_{\ell}) > (k+2) w_{\ell'}. $$ As $\theta(j_{\ell}) \leq w_{\ell} \leq w_{\ell'} $, this implies that $$ \frac{D_{\ell}}{w_{\ell}} + \frac{D_{\ell'}}{w_{\ell'}} \geq \frac{D_{\ell} + D_{\ell'}}{w_{\ell'}} > \frac{ (k+2) w_{\ell'} - \theta(j_{\ell})}{w_{\ell'}} \geq \frac{ (k+2) w_{\ell'} - w_{\ell}}{w_{\ell'}} \geq \frac{ (k+2) w_{\ell'} - w_{\ell'}}{w_{\ell'}} = k+1. $$ We have shown that, with probability one, there is some pair $\ell < \ell'$ satisfying this inequality $D_{\ell}/w_{\ell} + D_{\ell'}/w_{\ell'} > k+1$. Therefore, with probability one it holds that \begin{equation} \label{tr77} \sum_{\ell=1}^{k+1} D_{\ell}/w_{\ell} > k+1. \end{equation} But now take expectations, observing that $\ensuremath{\mathbf{E}}[D_{\ell}] = \ensuremath{\mathbf{E}}[ d(j_{\ell}, \S) ] \leq t_{j_{\ell}} = w_{\ell}$. So the LHS of (\ref{tr77}) has expectation at most $k+1$. This is a contradiction. \end{proof} We remark that it is possible to obtain an optimal $(1, k+1)$-determinization algorithm for the SCC or homogeneous settings, but we omit this since it is very similar to Algorithm~\ref{algo:det2}. \section{Acknowledgments} Our sincere thanks to Brian Brubach and to the anonymous referees, for many useful suggestions and for helping to tighten the focus of the paper. \bibliographystyle{abbrv}

52a76888ed3f7379083097b87c9506f55c00b098

\section{Introduction} The \emph{tree-width} of a graph is a parameter quantifying how ``close'' the graph is to being a tree. While tree-width has its roots in structural graph theory, and in particular in the Robertson-Seymour theory of graph minors, it has become a key tool in algorithm design in the past decades. This is because many algorithmic problems, on small tree-width graphs, can be solved efficiently using dynamic programming techniques. The algorithmic success of tree-width has also been demonstrated in the realm of topology. Makowsky and Mari\~no showed that, despite being $\#P$-hard to compute in general, the Jones and Kauffman polynomials can be computed efficiently on knot diagrams where the underlying graphs have small tree-width~\cite{MakowskyMarino03}. Recently Burton obtained a similar result for the HOMFLY-PT polynomial~\cite{Burton18}. In a different vein, Bar-Natan used divide-and-conquer techniques to design an algorithm to compute the Khovanov homology of a knot; he conjectures that his algorithm is very efficient on graphs of low cut-width~\cite{BarNatan07}, a graph parameter closely related to tree-width. Similarly, many topological invariants can be computed efficiently on manifold triangulations for which the face-pairing graphs have small tree-width; this is nicely captured by a generalization of Courcelle's theorem due to Burton and Rodney~\cite{BurtonDowney17}. A \emph{diagram} $D$ of a knot $K$ is a four-valent graph embedded in $S^2$, the two-sphere, with some ``crossing'' information at the vertices. Thus, we can apply the definition of tree-width directly to $D$. The \emph{diagrammatic tree-width} of $K$ is then the minimum of the tree-widths of its diagrams. Burton~\cite[page~2694]{BurtonEtAl15} and Makowsky and Mari\~no~\cite[page~755]{MakowskyMarino03} pose a natural question: is there a family of knots where the tree-width increases without bound? For example, while the usual diagrams of torus knots $T(p,q)$ (see Figure~\ref{F:torus}) can be easily seen to have tree-width going to infinity with $p$ and $q$, it could be that other diagrams of these torus knots have a uniformly bounded tree-width. \begin{figure}[h] \centering \includegraphics[width=7cm]{Figures/TorusKnot97.pdf} \caption{The usual diagram of a torus knot $T(9,7)$.} \label{F:torus} \end{figure} In this article, we give several positive answers. Our main tool comes from the analysis of surfaces in three-manifolds. We defer the precise definitions to \refsec{Background}. \begin{restate}{Theorem}{Thm:Main} Suppose that $k$ is an integer and $K$ is a knot having a diagram with tree-width at most $k$. Then either \begin{enumerate} \item there exists an essential planar meridional surface for $K$ with at most $8k + 8$ boundary components or \item $K$ has bridge number at most $4k + 4$. \end{enumerate} \end{restate} The main idea in the proof of \refthm{Main} is as follows. Suppose that $D$ is the given small tree-width diagram of a knot $K$. In \reflem{Sphere} we transform the given decomposition of $D$ into a family of disjoint spheres, each meeting $K$ in a small number of points. This partitions the ambient space into simple pieces. We call this a \emph{sphere-decomposition} of $K$. In \reflem{Tubing} we turn this into a \emph{multiple Heegaard splitting}, as introduced by Hayashi and Shimokawa~\cite[page~303]{HayashiShimokawa01}. Their thin position arguments, roughly following \cite{Gabai87iii}, establish the existence of the surfaces claimed by \refthm{Main}. To find families of knots with diagrammatic tree-width going to infinity, we use the contrapositive of \refthm{Main}. The following families of knots have neither small essential planar meridional surfaces nor small bridge spheres. \begin{itemize} \item Torus knots $T(p,q)$ where $p$ and $q$ are large, coprime, and roughly equal. These have no essential planar meridional surface by work of Tsau~\cite[page~199]{Tsau94}. Also, the bridge number is $\min (p, q)$, by work of Schubert~\cite[Satz~10]{Schubert54}. \item Knots with a high distance Heegaard splitting, for example following Minsky, Moriah, and Schleimer~\cite[Theorem~3.1]{MinskyMoriahSchleimer07}. A theorem of Scharlemann~\cite[Theorem~3.1]{Scharlemann06} shows that such knots cannot have essential planar meridional surfaces in their complements with a small number of boundary components. A theorem of Tomova~\cite[Theorem~1.3]{Tomova08} shows that such knots cannot have small bridge spheres. \item Highly twisted plats, for example following Johnson and Moriah~\cite[Theorem~1.2]{JohnsonMoriah16}. They show that the given bridge sphere has high distance. A theorem of Bachman and Schleimer~\cite[Theorem~5.1]{BachmanSchleimer05a} excludes small essential planar meridional surfaces. A theorem of Tomova~\cite[Theorem~10.3]{Tomova07} excludes bridge spheres with smaller bridge number than the given one. \end{itemize} \begin{corollary} For every integer $k > 0$, there is a knot $K$ with diagrammatic tree-width at least $k$. \end{corollary} \begin{proof} By \refthm{Main}, the three families of knots above have diagrammatic tree-width going to infinity. \end{proof} \subsection{Relations with the tree-width of the complement} Two recent works~\cite{HuszarSpreerWagner17, MariaPurcell18} have investigated the interplay between the tree-width of a three-manifold and its geometric or topological properties. Here the tree-width of a three-manifold $M$ is the minimum of the tree-widths of the face-pairing graphs of its triangulations. Interestingly, both articles find it more convenient to work with \emph{carving-width} (also called \emph{congestion}) than with tree-width directly. This is also the case in our paper; see \refsec{Background} for definitions. Husz\'{a}r, Spreer and Wagner~\cite[Theorem~2]{HuszarSpreerWagner17} give a result similar to ours, but for three-manifolds. They prove that a certain family of manifolds $\{M_n\}$, constructed by Agol~\cite{Agol03}, has the property that any triangulation of $M_n$ has tree-width at least $n$. Their result differs from ours in two significant ways. For one, Agol's examples are non-Haken, while we work with knots and their complements, which are necessarily Haken. Furthermore, there is an inequality between the diagrammatic tree-width of a knot and the tree-width of its complement which cannot be reversed. That is, given a knot diagram of tree-width $k$, there are standard techniques to build a triangulation of its complement of tree-width $O(k)$. For example, one may use the method embedded in SnapPy~\cite[\texttt{link\_complement.c}]{SnapPy}; see the discussion of~\cite[Remark~6.2]{MariaPurcell18}. To see that this inequality cannot be reversed recall that \refthm{Main} gives a sequence of torus knots with tree-width going to infinity. On the other hand we have the following. \begin{restate}{Lemma}{Lem:Torus} The complement of the torus knot $T(p,q)$ admits a triangulation of constant tree-width. \end{restate} \noindent This is well-known to the experts; we include a proof in \refsec{Background}. There is much more to say about the tree-widths of triangulations of knot complements; however in this article we will restrict ourselves to knot diagrams. \subsection*{Acknowledgements} We thank Ben Burton for his very provocative questions, and Kristof Husz{\'a}r and Jonathan Spreer for enlightening conversations. We thank the Mathematisches Forschungsinstitut Oberwolfach and the organizers of Workshop 1542, as well as the Schlo\ss{} Dagstuhl Leibniz-Zentrum f\"ur Informatik and the organizers of Seminar 17072. Parts of this work originated from those meetings. The work of A. de Mesmay is partially supported by the French ANR project ANR-16-CE40-0009-01 (GATO) and the CNRS PEPS project COMP3D. Purcell was partially supported by the Australian Research Council. \section{Background} \label{Sec:Background} \subsection{Graph Theory} Unless otherwise indicated, all graphs will be simple and connected. The tree-width of a graph measures quantitatively how close the graph is to a tree. Although we provide a definition below for completeness, we will not use it in the proof of \refthm{Main}, as we rely instead on a roughly equivalent variant. \begin{definition} Suppose $G = (V, E)$ is a graph. A \emph{tree-decomposition} of $G$ is a pair $(\mathcal{T}, \mathcal{B})$ where $\mathcal{T}$ is a tree, $\mathcal{B}$ is a collection of subsets of $V$ called \emph{bags}, and the vertices of $\mathcal{T}$, called \emph{nodes}, are the members of $\mathcal{B}$. Additionally, the following properties hold: \begin{itemize} \item $\bigcup_{B \in \mathcal{B}} B = V$ \item For each $uv \in E$, there is a bag $B \in \mathcal{B}$ with $u, v \in B$. \item For each $u \in V$, the nodes containing $u$ induce a connected subtree of $\mathcal{T}$. \end{itemize} \noindent The \emph{width} of a tree-decomposition is the size of its largest bag, minus one. The \emph{tree-width} of $G$, denoted by $\operatorname{tw}(G)$ is the minimum width taken over all possible tree-decompositions of $G$. \end{definition} Tree-width has many variants, which all turn out to be equivalent up to constant factors. For our purposes, we will rely on the concept of carving-width; it has geometric properties that are well-suited to our setting. \begin{definition} Suppose that $G$ is a graph. A \emph{carving-decomposition} of $G$ is a pair $(\mathcal{T}, \phi)$ where $\mathcal{T}$ is a binary tree and $\phi$ is a bijection from the vertices of $G$ to the leaves of $\mathcal{T}$. For an edge $e$ of $\mathcal{T}$, the \emph{middle set} of $e$, denoted $\operatorname{mid}(e)$ is defined as follows. Removing $e$ from $\mathcal{T}$ breaks $\mathcal{T}$ into two subtrees $S$ and $T$. The leaves of $S$ and $T$ are mapped via $\phi$ to vertex sets $U$ and $V$. Then $\operatorname{mid}(e)$ is the set of edges connecting a vertex of $U$ to a vertex of $V$. The \emph{width} of an edge $e$ is the size of $\operatorname{mid}(e)$. The width of the decomposition $(\mathcal{T}, \phi)$ is the maximum of the widths of the edges of $\mathcal{T}$. Finally the \emph{carving-width} $\operatorname{cw}(G)$ is the minimum possible width of a carving-decomposition of $G$. \end{definition} The carving-width of a graph of constant degree is always within a constant factor of its tree-width, as follows. \begin{theoremQED}\cite[page~111 and Theorem~1]{Bienstock90} \label{Thm:CW} Suppose that $G$ is a graph with degree at most $d$. Then we have: \[ \frac{2}{3} \cdot (\operatorname{tw}(G) + 1) \leq \operatorname{cw}(G) \leq d \cdot (\operatorname{tw}(G) + 1). \qedhere \] \end{theoremQED} Recall that a \emph{bridge} in a connected graph $G$ is an edge separating $G$ into more than one connected component. A \emph{bond carving-decomposition} is one where, for any edge $e$ in the tree $\mathcal{T}$, the associated vertex sets $U$ and $V$ induce connected subgraphs. One of the strengths of the notion of a carving-decomposition is the following theorem of Seymour and Thomas (see also the discussion in Marx and Piliczuk~\cite[Section~4.6]{MarxPilipczuk15}). \begin{theorem}\cite[Theorem~5.1]{SeymourThomas94} \label{Thm:SeymourThomas} Suppose that $G$ is a simple connected bridgeless graph with at least two vertices and with carving-width at most $k$. Then there exists a bond carving-decomposition of $G$ having width at most $k$. \qed \end{theorem} Suppose that $G$ has a bond carving-decomposition. Suppose also that $G$ is \emph{planar}: it comes with an embedding into $S^2$. Then each middle set $\operatorname{mid}(e)$ for $G$ gives a Jordan curve $\gamma_e \subset S^2$ separating the two vertex sets $U$ and $V$. One can take, for example, the simple cycle in the dual graph $G^*$ corresponding to the cut. After a small homotopy of each, we can assume that these Jordan curves are pairwise disjoint. We say that a family of pairwise disjoint Jordan curves, crossing $G$ transversely, \emph{realizes} a carving-decomposition $(\mathcal{T}, \phi)$ if the partitions of the vertex sets that it induces correspond to those of $(\mathcal{T}, \phi)$. \refthm{SeymourThomas} and the above discussion yield the following. \begin{corollary} \label{Cor:Realize} Suppose that $G$ is a bridgeless planar graph with at least two vertices and with carving-width at most $k$. Then there exists a family of pairwise disjoint Jordan curves realizing a bond carving-decomposition of $G$ of width at most $k$. \qed \end{corollary} \subsection{Knots} We refer to Rolfsen's book~\cite{Rolfsen90} for background on knot theory. Here we will use the equatorial two-sphere $S^2$, and its containing three-sphere $S^3$, as the canvas for our knot diagrams. In this way we avoid choosing a particular point at infinity. If one so desires, they may place the point at infinity on the equatorial two-sphere and so obtain the $xy$--plane in $\mathbb{R}^3$. A \emph{knot} $K$ is a regularly embedded circle $S^1$ inside of $S^3$. One very standard way to represent $K$ is by a \emph{knot diagram}: we move $K$ to be generic with respect to geodesics from the north to south poles and we then project $K$ onto the equatorial two-sphere. We add a label at each double point of the image recording which arc was closer to the north pole. Any knot has infinitely many knot diagrams, even when considering diagrams up to isotopy. The four-valent graph underlying a knot diagram need not be simple, but it is always connected. To avoid technical issues we may subdivide edges, introducing vertices of valence two, to ensure our graphs are simple. Since knot diagrams are four-valent, this makes at most a constant difference in any relevant quantity. \begin{definition} Suppose that $K$ is a knot with diagram $D$. The \emph{tree-width} of $D$ is defined to be the tree-width of the underlying graph (after any subdivision needed to ensure simplicity). We define the \emph{carving-width} of $D$ similarly. The \emph{diagrammatic tree-width} (respectively \emph{diagrammatic carving-width}) of $K$ is defined to be the minimum tree-width (respectively carving-width) taken over all diagrams of $K$. \end{definition} \begin{remark} It is relatively easy to find knots with small diagrammatic tree-width. For example, any connect sum of trefoil knots has uniformly bounded diagrammatic tree-width~\cite[Section~2.G]{Rolfsen90}. For example, any two-bridge knot has uniformly bounded diagrammatic tree-width~\cite[Section~4.D and Exercise~10.C.6]{Rolfsen90}. On the other hand, it is an exercise to show that any knot (even the unknot) has \emph{some} diagram of high tree-width. This then motivates the question of Burton~\cite[page~2694]{BurtonEtAl15} and of Makowsky and Mari\~no~\cite[page~755]{MakowskyMarino03}: is there a family of knots $\{K_n\}$ so that \emph{every} diagram of $K_n$ has tree-width at least $n$? \end{remark} \subsection{Tangles} \begin{figure}[t] \centering \includegraphics[width=12cm]{Figures/Tangles} \caption{Three tangles represented by two-dimensional projections. The first one is trivial in a three-ball, the second one is non-trivial in a three-ball and the third one is flat in the solid pants.} \label{Fig:Tangles} \end{figure} Suppose that $M$ is a compact, connected three-manifold. A \emph{tangle} $T \subset M$ is a properly embedded finite collection of pairwise disjoint arcs and loops. We call $(M, T)$ a \emph{three-manifold/tangle pair}, see \reffig{Tangles} for examples. We call the components of $T$ the \emph{strands} of the tangle. We take $N(T) \subset M$ to be a small, closed, tubular neighborhood of $T$. Let $n(T)$ be the relative interior of $N(T)$. So if $\alpha \subset T$ is a strand then $n(\alpha)$ is homeomorphic to an open disk crossed with a closed interval (or with a circle). We define the \emph{tangle complement} to be $M_T = M - n(T)$. Now suppose that $F \subset M$ is a properly embedded surface (two-manifold) which is transverse to $T$. In particular, $T \cap \bdy F = \emptyset$. Shrinking $n(T)$ if needed, we define $F_T = F - n(T)$. Note that $F_T$ is properly embedded in $M_T$. For any surface $F$, a simple closed properly embedded curve $\alpha \subset F$ is \emph{essential} on $F$ if it does not cut a disk off of $F$. A properly embedded arc $\alpha \subset F$ is \emph{essential} on $F$ if it does not cut a \emph{bigon} off of $F$: a disk $D \subset F$ with $\bdy D = \alpha \cup \beta$ and with $\beta \subset \bdy F$. Suppose that $(M, T)$ is a three-manifold/tangle pair. A \emph{meridian} for a tangle $T$ is an essential simple closed curve in $\bdy N(T) - \bdy M$ that bounds a properly embedded disk in $N(T)$. The given disk is called a \emph{meridional disk}. A surface $F_T$ is \emph{meridional} if $\bdy F$ is empty and thus every component of $\partial F_T$ is a meridian of $T$. A surface is \emph{planar} if it is a subsurface of the two-sphere. For example, if $(B^3,T)$ is a three-ball/tangle pair, the boundary of the three-ball is a planar meridional surface. Similarly, a meridian for a solid torus $D \times S^1$ is an essential simple closed curve on the boundary $S^1 \times S^1$ bounding a properly embedded disk. For example, any disk of the form $D \times \{\pt\}$, with product structure as above, is a meridional disk, see~\reffig{compDisks}, right. \begin{figure}[t] \centering \includegraphics[width=12cm]{Figures/compDisks} \caption{Left: a compressing disk for a planar meridional surface. Right: a meridional disk for the boundary of a solid torus.} \label{Fig:compDisks} \end{figure} With the notion of meridians in hand, we can give the promised proof of \reflem{Torus}. \begin{lemma} \label{Lem:Torus} The complement of the torus knot $T(p,q) \subset S^3$ admits a triangulation of constant tree-width. \end{lemma} We prove this by building an explicit triangulation of the complement of torus knots which is very path-like. We use Jaco and Rubinstein's layered triangulations~\cite{JacoRubinstein06}. \begin{proof}[Proof of \reflem{Torus}] Recall that the two-torus $T \homeo S^1 \times S^1$ has a triangulation with exactly one vertex, three edges, and two triangles. We call this the \emph{standard triangulation} of $T$; we label the edges by $a, b, c$. We form $P = T \times [-1, 1]$ and triangulate it with two triangular prisms, each made out of three tetrahedra. Without changing the upper or lower boundaries ($T \times \{\pm 1\}$) we subdivide the triangulation of $P$ to make $n(a \times \{0\})$ an open subcomplex. Finally, we form $Q = P - n(a \times \{0\})$, together with its triangulation, by removing those open cells. Since the torus knot $T(p, q)$ is connected, the integers $p$ and $q$ are coprime. Using B\'ezout's identity, pick $u$ and $v$ such that $pv - qu = 1$. Following Jaco and Rubinstein~\cite[Section~4]{JacoRubinstein06}, build two layered solid tori $U$ and $V$ of slopes $p/u$ and $q/v$: these are one-vertex triangulations of $D \times S^1$ so that \begin{itemize} \item the boundary of each is a standard triangulation of the two-torus, \item the face-pairing graph for the triangulation of $U$ and $V$ is a daisy chain graph (see Figure~\ref{F:daisy}), and \item the meridian $\mu \subset \bdy U$ crosses the edges $(a, b, c)$ in $\bdy U$ exactly $(p, u, p + u)$ times while \item the meridian $\nu \subset \bdy V$ crosses the edges $(a, b, c)$ in $\bdy V$ exactly $(q, v, q + v)$ times. \end{itemize} \begin{figure}[t] \centering \includegraphics[width=8cm]{Figures/daisy} \caption{The daisy chain graph.} \label{F:daisy} \end{figure} We now glue $U$ to $P$ by identifying $\bdy U$ with the upper boundary of $P$, respecting the $a$, $b$, $c$ labels. Similarly we glue $V$ to $P$ along the lower boundary, again respecting labels. Let $M = U \cup P \cup V$ be the result. Since the upper and lower halves of $P$ are products, we may isotope $\mu$ down, and $\nu$ up, to lie in the middle torus $T \times \{0\}$. These curves now cross $|pv - qu| = 1$ times: that is, once. We deduce that $M$ is homeomorphic to $S^3$; see for example Rolfsen~\cite[Section~9.B]{Rolfsen90}. Now, the edge $a \times \{0\} \subset P \subset M$ crosses these copies of $\mu$ and $\nu$ exactly $p$ and $q$ times, respectively. So this edge gives a copy of $T(p,q)$ in $S^3$. We now form $X(p, q) = S^3 - n(T(p, q))$ by drilling $a \times \{0\}$ out of $M$. We do this by replacing the triangulation of $P$ by that of $Q$. The resulting triangulation of $X(p, q)$ has a standard triangulation in $Q$ and a daisy chain triangulation in each of $U$ and $V$. So the overall triangulation has constant tree-width, with the model tree being a path. \end{proof} \subsection{Compression bodies} Before defining sphere-decompositions of knots, we recall a few more notions from three-manifold topology. \begin{definition} A \emph{one-handle} is a copy of the disk cross an interval, $D \times [-1, 1]$, with \emph{attaching regions} $D \times \{-1\}$ and $D \times \{1\}$. Suppose that $F$ is a disjoint (and possibly empty) union of closed oriented surfaces. We form a \emph{compression body} $C$ by starting with $F \times [0, 1]$, taking the disjoint union with a three-ball $B$, and attaching one-handles to $(F \times \{1\}) \cup \bdy B$. We attach enough one-handles to ensure that $C$ is connected. Then $\bdy_- C = F \times \{0\}$ is the \emph{lower boundary} of $C$. Also, $\bdy_+ C = \bdy C - \bdy_- C$ is the \emph{upper boundary} of $C$. If $\bdy_- C = \emptyset$ then $C$ is a \emph{handlebody}. \end{definition} Here is a concrete example which we will use repeatedly. \begin{definition} Fix $n > 0$. Suppose that $F$ is a disjoint union of $n - 1$ copies of $S^2$. We thicken, take the disjoint union with a three-ball $B$, and attach $n - 1$ one-handles in such a way to obtain a compression body $C(n)$. Note that $\bdy_+ C(n) \homeo S^2$. We call $C(n)$ an \emph{$n$--holed three-sphere}. Equally well, $C(n)$ has exactly $n$ boundary components, all spheres, and $C(n)$ embeds in $S^3$. Equally well, $C(n)$ is obtained from $S^3$ by deleting $n$ small open balls with disjoint closures. \end{definition} The four simplest $n$-holed three-spheres are the three-sphere (by convention), the three-ball, the shell $S^2 \times [0, 1]$, and the \emph{solid pants}: the thrice-holed three-sphere $C(3)$. Suppose that $B$ is the closed unit three-ball, centered at the origin of $\mathbb{R}^3$. Remove the open balls of radius $1/4$ centered at $(\pm 1/2, 0, 0)$, respectively to obtain the \emph{standard model} for $C(3)$. The \emph{equatorial} pair of pants $P \subset C(3)$ is the intersection of $C(3)$ and the $xy$--plane. \begin{definition} Suppose that $C(3)$ is the standard model of the solid pants. Suppose that $T \subset C(3)$ is a tangle. We say $T$ is \emph{flat} if \begin{itemize} \item $T$ is properly embedded in $P \subset C(3)$, the equatorial pair of pants, \item $T$ has no loop component, and \item all strands of $T$ are essential in $P$. \end{itemize} \end{definition} \begin{definition} Suppose that $C$ is a compression body. The construction of $C$ gives us a homeomorphism $C \homeo N \cup H$ where $N$ is either a three-ball or a copy of $\bdy_- C \times [0, 1]$ and where $H$ is a disjoint union of one-handles. Suppose that $\alpha \subset C$ is a properly embedded arc. \begin{itemize} \item We call $\alpha$ a \emph{vertical} arc if $\alpha$ lies in $N$ and there has the form $\{\pt\} \times [0, 1]$. \item We call $\alpha$ a \emph{bridge} if there is an embedded disk $D \subset C$ so that $\bdy D = \alpha \cup \beta$ with $\beta \subset \bdy_+ C$. We call $D$ a \emph{bigon} for $\alpha$. \end{itemize} A tangle $T \subset C$ is \emph{trivial} if \begin{itemize} \item every strand of $T$ is either vertical or is a bridge and \item every bridge $\alpha \subset T$ has a bigon $D$ so that $D \cap T = \alpha$. \end{itemize} \end{definition} Note that when $C$ is a handlebody, a trivial tangle consists solely of bridges. We refer to~\reffig{Tangles} for examples. Now we can begin to define the two types of surfaces that appear in \refthm{Main}. \begin{definition} Suppose that $K$ is a knot in $S^3$. A two-sphere $S \subset S^3$ is a \emph{bridge sphere} for $K$ if \begin{itemize} \item $S$ is transverse to $K$ and \item the induced tangles in the two components of $S^3 - n(S)$ are trivial. \end{itemize} The \emph{bridge number} of $K$ with respect to $S$ is the number of bridges in either trivial tangle. \end{definition} \begin{definition} Suppose that $M$ is a compact connected oriented three-manifold. Suppose that $F \subset M$ is a properly embedded two-sided surface. An embedded disk $(D, \bdy D) \subset (M, F)$ is a \emph{compressing disk} for $F$ if $D \cap \bdy M = \emptyset$, if $D \cap F = \bdy D$, and if $\bdy D$ is an essential loop in $F$, see~\reffig{compDisks} left for an illustration. If $F$ does not admit any compressing disk, then $F$ is \emph{incompressible}. A \emph{boundary compressing bigon} is an embedded disk $D \subset M$ with boundary $\bdy D = \alpha \cup \beta$ being the union of two arcs where $\alpha = D \cap F$ is an essential arc in $F$ and where $\beta = D \cap \bdy M$. If $F$ does not admit any boundary compressing disk, then $F$ is \emph{boundary-incompressible}. A closed surface $F$, embedded in the interior of $M$ is \emph{boundary parallel} if $F$ cuts a copy of $F \times [0, 1]$ off of $M$. A surface $F \subset M$ that is incompressible, is boundary-incompressible, and is not boundary-parallel is called \emph{essential}. \end{definition} \subsection{Multiple Heegaard splittings} Here we recall the concept of a multiple Heegaard splitting, which is central to the work of Hayashi and Shimokawa~\cite{HayashiShimokawa01}. We also state one of their theorems that we will rely on. Suppose that $M$ is a compact connected oriented three-manifold. A \emph{Heegaard splitting} $\mathcal{C}$ of $M$ is a decomposition of $M$ as a union of two compression bodies $C$ and $C'$, disjoint on their interiors, with $\bdy_+ C = \bdy_+ C'$ and with $\bdy M = \bdy_- C \cup \bdy_- C'$. Following~\cite{ScharlemannSchultensSaito16, ScharlemannThompson94b}, a \emph{generalized Heegaard splitting} $\mathcal{C}$ of $M$ is a path-like version of a Heegaard splitting; it is a decomposition of $M$ into a sequence of compression bodies $C_i$ and $C_i'$, disjoint on their interiors, where $\bdy_+ C_i = \bdy_+ C_i'$, where $\bdy_- C_i' = \bdy_- C_{i+1}$, and where $\bdy M = \bdy_- C_1 \cup \bdy_- C_n'$. Following Hayashi and Shimokawa~\cite[page~303]{HayashiShimokawa01} we now generalize generalized Heegaard splittings. We model the decomposition of $M$ on a graph instead of just a path. We also must respect a given tangle $T$ in $M$. Here is the definition. \begin{definition} Suppose that $(M, T)$ is a three-manifold/tangle pair. A \emph{multiple Heegaard splitting} $\mathcal{C}$ of $(M, T)$ is a decomposition of $M$ into a union of compression bodies $\{C_i\}$, with the following properties. \begin{enumerate} \item For each index $i$ there is some $j \neq i$ so that $\partial_+ C_i$ is identified with $\partial_+ C_j$. \item For each index $i$ and for each component $F \subset \partial_- C_i$ either $F$ is a component of $\bdy M$ or there is an index $j$ so that $F$ is attached to some component $G \subset \partial_- C_j$. Here we allow $j = i$ but require $G \neq F$. \item The surfaces $\cup \bdy_\pm C_i$ are transverse to $T$. The tangle $T \cap C_i$ is trivial in $C_i$. \end{enumerate} \end{definition} When discussing the union of surfaces we will use the notation $\bdy_\pm \mathcal{C} = \cup \bdy_\pm C_i$. Again, following~\cite[page~303]{HayashiShimokawa01} we give a complexity of multiple Heegaard splittings. \begin{definition} Suppose that $(M, T)$ is a three-manifold/tangle pair. Suppose that $F \subset M$ is a connected closed surface embedded in the interior of $M$. Suppose also that $F$ is transverse to $T$. The \emph{complexity} of $F$ is the ordered pair $c(F) = (\text{genus(F)}, |F \cap T|)$. Now, suppose that $\mathcal{C}$ is a multiple Heegaard splitting of $(M, T)$. \begin{itemize} \item The \emph{cost} of $\mathcal{C}$ is the maximal number of intersections between a component of $\bdy_+ \mathcal{C}$ and $T$. That is, the cost is \[ \max\{ |F \cap T| : \mbox{$F$ is a component of $\bdy_+ \mathcal{C}$} \}. \] \item The \emph{width} of $\mathcal{C}$ is the ordered multiset of complexities \[ w(\mathcal{C}) = \{ c(F) : \mbox{$F$ is a component of $\bdy_+ \mathcal{C}$} \}. \] Here the complexities are listed in lexicographically non-increasing order. \end{itemize} The \emph{width} of a pair $(M,T)$ is the minimum possible width of a multiple Heegaard splitting of $(M,T)$, ordered lexicographically. A multiple Heegaard splitting of $(M,T)$ is \emph{thin} if it achieves the minimal width over all possible multiple Heegaard splittings of $(M,T)$. \end{definition} The following definitions, still from Hayashi and Shimokawa~\cite[page~304]{HayashiShimokawa01}, provide a notion of essential surfaces in the setting of a three-manifold/tangle pair. Recall that if $X \subset M$ is transverse to $T$ then $X_T$ denotes $X - n(T)$. \begin{definition} Let $X$ be a compact orientable $3$-manifold, $T$ a $1$-manifold properly embedded in $X$, and $F$ a closed orientable $2$-manifold embedded transversely to $T$ in $X$. Let $\tilde{X}$ be the $3$-manifold obtained from $X$ by capping off all the spherical boundary components disjoint from $T$ with balls. An embedded disk $Q$ is said to be a \emph{thinning disk} of $F$ if $T \cap Q = T \cap \partial Q= \alpha$ is an arc and $Q \cap F$ contains the arc $cl(\partial Q \setminus \alpha)=\beta$ as a connected component. The surface $F$ is \emph{$T$--essential} if \begin{enumerate} \item $F_T$ is incompressible in $X_T$, \item $F$ has no thinning disk, \item no component $F_0$ of $F$ cobounds with a component $F_1$ of $\partial X$ in $\tilde{X}$ a $3$-manifold homeomorphic to $F_0 \times I$, possibly intersecting $T$ in vertical arcs, where $F_0=F_0 \times \{0\}$ and $F_0=F_1 \times \{1\}$, and \item no sphere component of $F$ bounds a ball disjoint from $T$ in $\tilde{X}$. \end{enumerate} \end{definition} Note that when $\partial X =\emptyset$, which will be the case throughout this paper, a $T$--essential surface $F$ is incompressible and boundary-incompressible in $X_T$. We can now state the necessary result from~\cite{HayashiShimokawa01}. \begin{theorem} \label{Thm:HS} Suppose that $\mathcal{C}$ is a thin multiple Heegaard splitting of a three-manifold/tangle pair $(M,T)$. Then every component of $\bdy_- \mathcal{C}$ is $T$--essential in $(M,T)$. \end{theorem} \begin{proof} This follows from Theorem~1.1 and Lemma~2.3 of~\cite{HayashiShimokawa01}. \end{proof} \section{Sphere-decompositions} We connect multiple Heegaard splittings to tree-width by turning the Jordan curves of \refcor{Realize} into spheres, and broadening these into (almost) a multiple Heegaard splitting. \begin{definition} Let $K$ be a knot in $S^3$. A \emph{sphere-decomposition} $\mathcal{S}$ of $K$ is a finite collection of pairwise disjoint, embedded spheres in $S^3$ meeting $K$ transversely, and so that every component $X$ of $S^3 - n(\mathcal{S})$ is either \begin{itemize} \item a three-ball, and $(X, X \cap K)$ is a trivial tangle, or \item a solid pants, and $(X, X \cap K)$ is a flat tangle. \end{itemize} \end{definition} We now define various notions of complexity of a sphere-decomposition. \begin{definition} The \emph{weight} of a sphere $S$ in a sphere-decomposition $\mathcal{S}$ is the number of intersections $S \cap K$. The \emph{cost} of $\mathcal{S}$ is the weight of its heaviest sphere. The \emph{width} of $\mathcal{S}$ is the list of the weights of its spheres, with multiplicity, given in non-increasing order. \end{definition} As an example, a bridge sphere $S$ for a knot in $S^3$ gives a sphere-decomposition with one sphere; the width is $\{2b\}$, where $b$ is number of bridges of $K$ on either side of $S$. As another example consider a pretzel knot, as in \reffig{Pretzel}, or even more generally a Montesinos knot consisting of three rational tangles. Then the three two-spheres about the three rational tangles give a sphere-decomposition of width $\{4, 4, 4\}$. \begin{figure} \centering \def5cm{5cm} \import{Figures/}{Pretzel_knot.pdf_tex} \caption{The pretzel knot $P(-2,3,7)$ with a sphere-decomposition made of three spheres $S_1$, $S_2$ and $S_3$\protect\footnotemark.} \label{Fig:Pretzel} \end{figure} A carving-decomposition of a knot diagram yields a sphere-decomposition of the knot, as follows. \begin{lemma} \label{Lem:Sphere} Suppose that $K$ is a knot. Suppose that $D$ is a diagram of $K$ with carving-width at most $k$. Then there is a sphere-decomposition of $K$ with cost at most $k$. \end{lemma} \begin{proof} Let $S$ be the equatorial sphere containing the graph $D$. Note that $D$ is bridgeless. Thus, by \refcor{Realize}, there is a family $\Gamma$ of pairwise disjoint Jordan curves in $S$ realizing a minimal width carving-decomposition of $D$. By assumption this has width at most $k$. We obtain the knot $K$ by moving the overstrands of $D$ slightly towards the north pole of $S^3$ while moving the understrands slightly towards the south pole. Every Jordan curve $\gamma \in \Gamma$ can be realized as the intersection of a two-sphere $S(\gamma)$ with the equatorial sphere $S$: we simply cap $\gamma$ off with disks above and below $S$. We do this in such a way that resulting spheres $S(\gamma)$ are pairwise disjoint and only meet $K$ near $S$. We claim that the resulting family of spheres $\mathcal{S} = \{S(\gamma)\}$ is a sphere-decomposition of $K$ of cost at most $k$. First, the bound on the cost follows from the bound on the width of the initial carving-decomposition realized by $\Gamma$. The arborescent and tri-valent structure of the carving-decomposition translates into the fact that every sphere in $\mathcal{S}$ is adjacent, on each side, to either a pair of spheres (respectively to no other sphere) as the corresponding half-edge is not (respectively is) a leaf of the carving-tree. Since we are in $S^3$, this implies that the connected components of $S^3 - n(\mathcal{S})$ are either three-balls (at the leaves) or solid pants. At the leaves, the tangle in the three-ball is a small neighborhood of a single crossing in the original diagram $D$ (or of a valence two vertex). Thus the tangles at the leaves are trivial. Inside each solid pants $C$, the tangle lies in the equatorial pair of pants $P$. This is because there are no crossings of $D$ in $P$. Finally, the tangle is flat in $P$ as otherwise we could find a carving of lower width. \end{proof} \footnotetext{This picture and the next one are adapted from a public domain \href{https://commons.wikimedia.org/wiki/File:Pretzel_knot.svg}{figure} from the Wikimedia commons by Sakurambo.} A sphere-decomposition is not in general a multiple Heegaard splitting, as the example of \reffig{Pretzel} shows. However, the following lemma shows that a sphere-decomposition of small cost can be upgraded to a multiple Heegaard splitting of small cost. \begin{lemma} \label{Lem:Tubing} Suppose that $K \subset S^3$ is a knot, and suppose that $K$ admits a sphere-decomposition $\mathcal{S}$ of cost at most $k$. Then $(S^3, K)$ admits a multiple Heegaard splitting $\mathcal{C}$ of cost at most $2k$, where all components of $\bdy_\pm \mathcal{C}$ are spheres. \end{lemma} \begin{proof} The construction simply adds a ``thick'' sphere into every component of $S^3 - n(\mathcal{S})$. Let $C$ be a solid pants of $S^3 - n(\mathcal{S})$; label its boundaries $U$, $V$, and $W$ and set $u = |K \cap U|$ and so on. Suppose that $\alpha$ is a strand of $K \cap C$ so that, relabelling as needed, $\alpha$ meets $U$ and $V$. We form a two-sphere $S(C)$ by tubing $U$ to $V$ using an unknotted tube parallel to the strand $\alpha$ but such that $\alpha$ is outside of the tube. See \reffig{Pretzel2} for an example. \begin{figure} \centering \def5cm{5cm} \import{Figures/}{Pretzel_knot2.pdf_tex} \caption{The pretzel knot $P(-2,3,7)$ with its natural sphere-decomposition, and with the addition of a ``thick'' sphere $S'$ in the solid-pants component, obtained by tubing $S_1$ to $S_2$. The spheres of $\mathcal{S}$ are dashed and the thick sphere is dotted (compare with \reffig{Pretzel}).} \label{Fig:Pretzel2} \end{figure} The sphere $S(C)$ meets $K$ in $u + v$ points. By the definition of the cost, we have $\max \{ u, v, w \} \leq k$, thus $u + v \leq 2k$. We perform this tubing construction in every solid pants $C$ of $S^3 - n(\mathcal{S})$; we place the sphere $S(C)$ into the set $\mathcal{R}$. For every three-ball component $B$ of $S^3 - n(\mathcal{S})$ we take $S(B)$ to be a push-off of $\bdy B$ into $B$. We place the sphere $S(B)$ into the set $\mathcal{R}$. We claim that $\mathcal{C} = S^3 - n(\mathcal{R} \cup \mathcal{S})$ is a multiple Heegaard splitting, where $\bdy_+ \mathcal{C} = \mathcal{R}$ and where $\bdy_- \mathcal{C} = \mathcal{S}$. Note that every component $C' \subset \mathcal{C}$ is either a three-ball, a shell, or a solid pants. So there is a unique (up to isotopy) compression body structure on $C'$ that has $\bdy_+ C' \subset \mathcal{R}$ and $\bdy_- C' \subset \mathcal{S}$. Let $C$ be the component of $S^3 - n(\mathcal{S})$ containing $C'$. \begin{itemize} \item Suppose that $C'$ is a three-ball. Then $K \cap C'$ is either one or two boundary parallel bridges, by the construction of $\mathcal{S}$. \item Suppose that $C'$ is a shell and $C$ is a three-ball. Then $K \cap C'$ is four vertical arcs, because $\bdy_+ C'$ is parallel to $\bdy C$. \item Suppose that $C'$ is a shell and $C$ is a solid pants. We use the notation of the construction: $\bdy C = U \cup V \cup W$ and $\bdy_+ C'$ is obtained by tubing $U$ to $V$. The tangle $K \cap C'$ consists of vertical arcs (coming from strands of $K \cap C$ connecting $U$ or $V$ to $W$) and bridges parallel into the upper boundary (coming from strands of $K \cap C$ connecting $U$ to $V$, or $U$ to itself, or $V$ to itself). \item Suppose that $C'$ is a solid pants and thus $C$ is a solid pants. Then $K \cap C'$, by construction, consists of vertical arcs only. \end{itemize} This concludes the proof. \end{proof} We are now ready to prove our main result. \begin{theorem} \label{Thm:Main} Suppose that $k$ is an integer and $K$ is a knot having a diagram with tree-width at most $k$. Then either \begin{enumerate} \item there exists an essential planar meridional surface for $K$ with at most $8k + 8$ boundary components or \item $K$ has bridge number at most $4k + 4$. \end{enumerate} \end{theorem} \begin{proof} Let $D$ be a diagram of $K$ that has tree-width at most $k$. Then by \refthm{CW}, it has carving-width at most $4k + 4$. Applying successively Lemmas~\ref{Lem:Sphere} and~\ref{Lem:Tubing} to this carving-decomposition, we obtain a multiple Heegaard splitting $\mathcal{C}$ of $(S^3, K)$ of cost at most $8k + 8$ and in which all the surfaces are spheres. Now let us consider a thin multiple Heegaard splitting $\mathcal{C}'$ of $(S^3, K)$. Since $\mathcal{C}'$ is thin, its width is less than or equal to that of $\mathcal{C}$. Thus all of the surfaces of $\bdy_+ C'$ must be spheres, as genus is the first item in our measure of complexity. We next deduce that the cost of $\mathcal{C}'$ is at most that of $\mathcal{C}$, so is at most $8k + 8$. We now have two cases, depending on whether or not $\bdy_- \mathcal{C}'$ is empty. If it is non-empty, then, taking $T=K$ in \refthm{HS} tells us that every component of $\bdy_- \mathcal{C}'$ is $K$--essential in $S^3$. Thus every component of $\bdy_- \mathcal{C}'_K$ is incompressible and boundary-incompressible in $S^3_K$. Furthermore, since all the surfaces of $\bdy_+ C'$ are spheres, the components of $\bdy_- \mathcal{C}'_K$ are not boundary-parallel, therefore they are essential. Since the cost of $\mathcal{C}'$ is at most $8k + 8$, we deduce that $K$ admits an essential planar meridional surface with at most $8k + 8$ boundary components. Suppose instead that $\bdy_- \mathcal{C}'$ is empty. Thus $\bdy_+ \mathcal{C}'$ consists of a single sphere $S$, which is necessarily a bridge sphere. Again, the cost of $\mathcal{C}'$ is at most $8k + 8$, so $S$ has at most $4k + 4$ bridges. This concludes the proof. \end{proof} \bibliographystyle{hyperplain}

97a451ca8e0b0bdb9e47e80255fc891802787313

\section{Algortihms}\label{sec:alg} In this section, we present \textsf{Replacement Greedy}~\citep{Stan2017} and \textsf{Replacement OMP}{} for dictionary selection with generalized sparsity constraints. \subsection{Replacement Greedy} \textsf{Replacement Greedy}{} was first proposed as an algorithm for a different problem, \emph{two-stage submodular maximization}~\citep{Balkanski2016}. In two-stage submodular maximization, the goal is to maximize \begin{align} h(X) = \sum_{t = 1}^T \max_{Z_t \subseteq X \colon Z_t \in \mathcal{I}_t} f_t(Z_t), \end{align} where $f_t$ is a nonnegative monotone submodular function ($t \in [T]$) and $\mathcal{I}_t$ is a matroid. Despite the similarity of the formulation, in dictionary selection, the functions $f_t$ are not necessarily submodular, but come from the continuous function $u_t$. Furthermore, in two-stage submodular maximization, the constraints on $Z_t$ are individual for each $t \in [T]$, while we pose a global constraint $\mathcal{I}$. In the following, we present an adaptation of \textsf{Replacement Greedy}{} to dictionary selection with generalized sparsity constraints. \textsf{Replacement Greedy}{} stores the current dictionary $X$ and supports $Z_t \subseteq X$ such that $(Z_1, \dots, Z_T) \in \mathcal{I}$, which are initialized as $X = \emptyset$ and $Z_t = \emptyset$ ($t\in [T]$). At each step, the algorithm considers the gain of adding an element $a \in V$ to $X$ with respect to each function $f_t$, i.e., the algorithm selects $a$ that maximizes $\max_{(Z_1', \dots, Z_T') \in \mathcal{F}_a} \sum_{t=1}^T \{f_t(Z_t') - f(Z_t) \}$. See \Cref{alg:replacement-greedy} for a pseudocode description. Note that for the individual matroid constraint $\mathcal{I}$, the algorithm coincides with the original \textsf{Replacement Greedy}~\citep{Stan2017}. \begin{algorithm} \caption{\textsf{Replacement Greedy}{} \& \textsf{Replacement OMP}}\label{alg:replacement-greedy} \begin{algorithmic}[1] \STATE Initialize $X \gets \emptyset$ and $Z_t \gets \emptyset$ for $t = 1, \dots, T$. \FOR{$i = 1, \dots, k$} \STATE Pick $a^* \in V$ that maximizes\\ $ \begin{cases} \max_{(Z'_1,\cdots,Z'_T) \in \mathcal{F}_{a^*}} \sum_{t = 1}^T \left\{ f_t(Z'_t) - f_t(Z_t) \right\} \quad \text{(\textsf{Replacement Greedy}{})}\\ \max_{(Z'_1,\cdots,Z'_T) \in \mathcal{F}_{a^*}} \left\{ \frac{1}{M_{s,2}} \sum_{t = 1}^T \norm{\nabla u_t(\mathbf{w}^{(Z_t)}_t)_{Z'_t \setminus Z_t}}^2 - M_{s,2} \sum_{t = 1}^T \norm{(\mathbf{w}^{(Z_t)}_t)_{Z_t \setminus Z'_t}}^2 \right\}\\ \end{cases} $ \\\hfill(\textsf{Replacement OMP}{})\\ and let $(Z'_1, \cdots, Z'_T)$ be a replacement achieving a maximum. \STATE Set $X \gets X + a^*$ and $Z_t \gets Z'_t$ for each $t \in [T]$. \ENDFOR \STATE \textbf{return} $X$. \end{algorithmic} \end{algorithm} \citet{Stan2017} showed that \textsf{Replacement Greedy}{} achieves an $((1-1/\sqrt{e})/2)$-approximation when $f_t$ are monotone submodular. We extend their analysis to our non-submodular setting. The proof can be found in \Cref{sec:alg-appendix}. \begin{theorem}\label{thm:replacement-greedy-1} Assume that $u_t$ is $m_{2s}$-strongly concave on $\Omega_{2s}$ and $M_{s,2}$-smooth on $\Omega_{s,2}$ for $t \in [T]$ and that the sparsity constraint $\mathcal{I}$ is $p$-replacement sparse. Let $(Z_1^*,\cdots,Z_T^*) \in \mathcal{I}$ be optimal supports of an optimal dictionary $X^*$. Then the solution $(Z_1, \cdots, Z_T) \in \mathcal{I}$ of \textsf{Replacement Greedy}{} after $k'$ steps satisfies \begin{equation*} \sum_{t = 1}^T f_t(Z_t) \ge \frac{m^2_{2s}}{M_{s,2}^2} \left( 1 - \exp \left( - \frac{k'}{p} \frac{M_{s,2}}{m_{2s}} \right) \right) \sum_{t = 1}^T f_t(Z^*_t) \end{equation*} \end{theorem} \subsection{Replacement OMP}\label{sec:replacement-OMP} Now we propose our algorithm, \textsf{Replacement OMP}{}. A down-side of \textsf{Replacement Greedy}{} is its heavy computation: in each greedy step, we need to evaluate $\sum_{t=1}^Tf_t(Z_t')$ for each $(Z_1',\dots,Z_t') \in\mathcal{F}_a(Z_1,\dots,Z_t)$, which amounts to solving linear regression problems $snT$ times if $u$ is the $\ell^2$-utility function. To avoid heavy computation, we propose a proxy of this quantity by borrowing an idea from orthogonal matching pursuit. \textsf{Replacement OMP}{} selects an atom $a \in V$ that maximizes \begin{equation}\label{eq:replacement-omp-marginal-gain-1} \max_{(Z'_1,\cdots,Z'_T) \in \mathcal{F}_{a}(Z_1,\cdots,Z_T)} \left\{ \frac{1}{M_{s,2}} \sum_{t = 1}^T \norm{\nabla u_t(\mathbf{w}^{(Z_t)}_t)_{Z'_t \setminus Z_t}}^2 - M_{s,2} \sum_{t = 1}^T \norm{(\mathbf{w}^{(Z_t)}_t)_{Z_t \setminus Z'_t}}^2 \right\}. \end{equation} This algorithm requires the smoothness parameter $M_{s,2}$ before the execution. Computing $M_{s,2}$ is generally difficult, but this parameter for the squared $\ell^2$-utility function can be bounded by $\sigma_{\mathrm{max}}^2(\mathbf{A}, 2)$. This value can be computed in $O(n^2d)$ time. \begin{theorem}\label{thm:replacement-omp-general} Assume that $u_t$ is $m_{2s}$-strongly concave on $\Omega_{2s}$ and $M_{s,2}$-smooth on $\Omega_{s,2}$ for $t \in [T]$ and that the sparsity constraint $\mathcal{I}$ is $p$-replacement sparse. Let $(Z_1^*,\cdots,Z_T^*) \in \mathcal{I}$ be optimal supports of an optimal dictionary $X^*$. Then the solution $(Z_1, \cdots, Z_T) \in \mathcal{I}$ of \textsf{Replacement OMP}{} after $k'$ steps satisfies \begin{equation*} \sum_{t = 1}^T f_t(Z_t) \ge \frac{m^2_{2s}}{M_{s,2}^2} \left( 1 - \exp \left( - \frac{k'}{p} \frac{M_{s,2}}{m_{2s}} \right) \right) \sum_{t = 1}^T f_t(Z^*_t). \end{equation*} \end{theorem} \subsection{Complexity} Now we analyze the time complexity of \textsf{Replacement Greedy}{} and \textsf{Replacement OMP}. In general, $\mathcal{F}_a$ has $O(n^T)$ members, and therefore it is difficult to compute $\mathcal{F}_a$. Nevertheless, we show that \textsf{Replacement OMP}{} can run much faster for the examples presented in \Cref{sec:general}. In \textsf{Replacement Greedy}{}, it is difficult to find an atom with the largest gain at each step. This is because we need to maximize a nonlinear function $\sum_{t=1}^T f_t(Z_t')$. Conversely, in \textsf{Replacement OMP}{}, if we can calculate $\mathbf{w}^{(Z_t)}_t$ and $\nabla u_t(\mathbf{w}^{(Z_t)}_t)$ for all $t \in [T]$, the problem of calculating gain of each atom is reduced to maximizing a linear function. In the following, we consider the $\ell_2$-utility function and average sparsity constraint because it is the most complex constraint. A similar result holds for the other examples. In fact, we show that this task reduces to maximum weighted bipartite matching. The Hungarian method returns the maximum weight bipartite matching in $\mathrm{O}(T^3)$ time. We can further improve the running time to $O(T\log T)$ time by utilizing the structure of this problem. Due to the limitation of space, we defer the details to \Cref{sec:alg-appendix}. In summary, we obtain the following: \begin{theorem}\label{thm:replacement-omp-runtime} Assume that the assumption of \Cref{thm:replacement-omp-general} holds. Further assume that $u$ is the $\ell^2$-utility function and $\mathcal{I}$ is the average sparsity constraint. Then \textsf{Replacement OMP}{} finds the solution $(Z_1, \cdots, Z_T) \in \mathcal{I}$ \begin{equation*} \sum_{t = 1}^T f_t(Z_t) \ge \left( \frac{\sigma_{\mathrm{max}}^2(\mathbf{A}, 2s)}{\sigma_{\mathrm{min}}^2(\mathbf{A},2)} \right)^2 \left( 1 - \exp \left( - \frac{1}{3} \frac{\sigma_{\mathrm{min}}^2(\mathbf{A},2)}{\sigma_{\mathrm{max}}^2(\mathbf{A}, 2s)} \right) \right) \sum_{t = 1}^T f_t(Z^*_t) \end{equation*} in $O(Tk(n \log T + ds))$ time. \end{theorem} \begin{remark} If finding an atom with the largest gain is computationally intractable, we can add an atom whose gain is no less than $\tau$ times the largest gain. In this case, we can bound the approximation ratio with replacing $k'$ with $\tau k'$ in \Cref{thm:replacement-greedy-1} and \ref{thm:replacement-omp-general}. \end{remark} \section{Proofs for \textsf{Replacement Greedy}{} and \textsf{Replacement OMP}{}}\label{sec:alg-appendix} \subsection{Proof for \textsf{Replacement Greedy}{}} \begin{lemma}\label{lem:replacement-greedy-marginal} Assume $\mathcal{I}$ is $p$-replacement sparse. Suppose that at some step, the solution is updated from $(Z_1,\cdots,Z_T)$ to $(Z'_1,\cdots,Z'_T)$ by \textsf{Replacement Greedy}{}. Let $(Z^*_1,\cdots,Z^*_T) \in \argmax_{(Z_1,\cdots,Z_T) \in \mathcal{I} \colon Z_t \subseteq X^*} f_t(Z)$ where $X^*$ is an optimal solution for dictionary selection. Then, the marginal gain of \textsf{Replacement Greedy}{} is bounded from below as follows: \begin{equation*} \sum_{t = 1}^T f_t(Z'_t) - \sum_{t = 1}^T f_t(Z_t) \ge \frac{1}{p} \left\{ \frac{m_{2s}}{M_{s,2}} \sum_{t = 1}^T f_t(Z^*_t) - \frac{M_{s,2}}{m_{2s}} \sum_{t = 1}^T f_t(Z_t) \right\} \end{equation*} where $s = \max_{(Z_t)_{t=1}^T \in \mathcal{I}} \max_{t \in [T]} |Z_t|$. \end{lemma} \begin{proof} Note that from the condition on feasible replacements, we have $|Z_t \triangle Z'_t| \le 2$. Since $u_t$ is $M_{s,2}$-smooth on $\Omega_{s,2}$, it holds that for any $\mathbf{z} \in \mathbb{R}^n$ with $\mathrm{supp}(\mathbf{z}) \subseteq Z'_t \setminus Z_t$, \begin{align*} f_t(Z'_t) - f_t(Z_t) &= u_t (\mathbf{w}^{(Z'_t)}) - u_t (\mathbf{w}^{(Z_t)})\\ &\ge u_t ((\mathbf{w}^{(Z_t)})_{Z_t \cap Z'_t} + \mathbf{z}) - u_t (\mathbf{w}^{(Z_t)})\\ &\ge \left\langle \nabla u_t (\mathbf{w}^{(Z_t)}), \mathbf{z} - (\mathbf{w}^{(Z_t)})_{Z_t \setminus Z'_t} \right\rangle - \frac{M_{s,2}}{2} \norm{ \mathbf{z} - (\mathbf{w}^{(Z_t)})_{Z_t \setminus Z'_t}}^2 \end{align*} Since this inequality holds for every $\mathbf{z}$ with $\mathrm{supp}(\mathbf{z}) \subseteq Z'_t \setminus Z_t$, by optimizing it for $\mathbf{z}$, we obtain \begin{equation}\label{eq:replacement-greedy-general-ineq1} f_t(Z'_t) - f_t(Z_t) \ge \frac{1}{2M_{s,2}} \norm{\nabla u_t(\mathbf{w}_t^{(Z_t)})_{Z'_t \setminus Z_t}}^2 - \frac{M_{s,2}}{2} \norm{(\mathbf{w}_t^{(Z_t)})_{Z_t \setminus Z'_t} }^2. \end{equation} In addition, due to the strong concavity of $u_t$, we have \begin{align}\label{eq:replacement-greedy-general-ineq2} f_t(Z_t^*) - f_t(Z_t) &= u_t(\mathbf{w}^{(Z_t^*)}) - u_t(\mathbf{w}^{(Z_t)}) \nonumber\\ &\le \left\langle \nabla u_t (\mathbf{w}_t^{(Z_t)}), \mathbf{w}_t^{(Z^*_t)} - \mathbf{w}_t^{(Z_t)} \right\rangle - \frac{m_{2s}}{2} \left\| \mathbf{w}_t^{(Z_t^*)} - \mathbf{w}_t^{(Z_t)} \right\|^2 \nonumber \\ &\le \max_{\mathbf{z} \colon \mathrm{supp}(\mathbf{z}) \subseteq Z_t^*} \left\{ \left\langle \nabla u_t (\mathbf{w}_t^{(Z_t)}), \mathbf{z} - \mathbf{w}_t^{(Z_t)} \right\rangle - \frac{m_{2s}}{2} \left\| \mathbf{z} - \mathbf{w}_t^{(Z_t)} \right\|^2 \right\} \nonumber \\ &= \frac{1}{2m_{2s}} \left\| (\nabla u_t (\mathbf{w}^{(Z_t)}))_{Z_t^* \setminus Z_t} \right\|^2 - \frac{m_{2s}}{2} \left\| (\mathbf{w}^{(Z_t)})_{Z_t \setminus Z_t^*} \right\|^2. \end{align} Similarly, due to the strong concavity of $u_t$, we have \begin{align}\label{eq:replacement-greedy-general-ineq3} - f_t(Z_t) &= u_t(\mathbf{0}) - u_t(\mathbf{w}_t^{(Z_t)}) \nonumber \\ &\le \left\langle \nabla u_t (\mathbf{w}_t^{(Z_t)}), - \mathbf{w}_t^{(Z_t)} \right\rangle - \frac{m_{2s}}{2} \left\| \mathbf{w}_t^{(Z_t)} \right\|^2 \nonumber \\ &= - \frac{m_{2s}}{2} \left\| \mathbf{w}_t^{(Z_t)} \right\|^2 \nonumber \\ &\le - \frac{m_{2s}}{2} \left\| (\mathbf{w}_t^{(Z_t)})_{Z_t \setminus Z^*_t} \right\|^2 \end{align} Since $\mathcal{I}$ is $p$-replacement sparse, we can take a sequence of $p$ replacements $(Z^{p'}_1, \cdots, Z^{p'}_T)_{p'=1}^p$ such that \begin{itemize} \item $(Z^{p'}_1, \cdots, Z^{p'}_T) \in \mathcal{F}(Z_1, \cdots, Z_T)$, \item each element in $Z^*_t \setminus Z_t$ appears at least once in sequence $(Z^{p'}_t \setminus Z_t)_{p'=1}^p$ for each $t \in [T]$, \item each element in $Z_t \setminus Z_t^*$ appears at most once in sequence $(Z_t \setminus Z^{p'}_t)_{p'=1}^p$ for each $t \in [T]$. \end{itemize} Now we prove the lemma by utilizing these properties. \begin{align*} &\sum_{t = 1}^T f_t(Z'_t) - \sum_{t = 1}^T f_t(Z_t) \\ &\ge \frac{1}{p} \sum_{p' = 1}^{p} \left\{ \sum_{t = 1}^T f_t(Z^{p'}_t) - \sum_{t = 1}^T f_t(Z_t) \right\} \tag{by the choice of $(Z_1', \dots, Z_T')$ and the feasibility of $(Z^{p'}_1,\cdots,Z^{p'}_T)$} \\ &\ge \frac{1}{p} \sum_{p'=1}^p \sum_{t = 1}^T \left\{ \frac{1}{2M_{s,2}} \norm*{\nabla u_t(\mathbf{w}_t^{(Z_t)})_{Z^{p'}_t \setminus Z_t}}^2 - \frac{M_{s,2}}{2} \norm*{\left( \mathbf{w}_t^{(Z_t)} \right)_{Z_t \setminus Z^{p'}_t}}^2 \right\} \tag{by \eqref{eq:replacement-greedy-general-ineq1}} \\ &\ge \frac{1}{p} \sum_{t = 1}^T \left\{ \frac{1}{2M_{s,2}} \norm*{\nabla u_t(\mathbf{w}_t^{(Z_t)})_{Z^*_t \setminus Z_t}}^2 - \frac{M_{s,2}}{2} \norm*{\left( \mathbf{w}_t^{(Z_t)} \right)_{Z_t \setminus Z_t^*}}^2 \right\} \tag{by the property of $(Z_t^{p'})_{p'=1}^p$} \\ &\ge \frac{1}{p} \sum_{t = 1}^T \left\{ \frac{m_{2s}}{M_{s,2}} \left( f_t(Z_t^*) - f_t(Z_t) \right) - \left( \frac{M_{s,2}}{m_{2s}} - \frac{m_{2s}}{M_{s,2}} \right) f_t(Z_t) \right\} \tag{by \eqref{eq:replacement-greedy-general-ineq2} and \eqref{eq:replacement-greedy-general-ineq3}} \\ &= \frac{1}{p} \sum_{t = 1}^T \left\{ \frac{m_{2s}}{M_{s,2}} f_t(Z_t^{*}) - \frac{M_{s,2}}{m_{2s}} f_t(Z_t) \right\}. \end{align*} \end{proof} Combined with \Cref{lem:misc}, \Cref{thm:replacement-greedy-1} is obtained. \subsection{Proof for \textsf{Replacement OMP}{}} \begin{lemma}\label{lem:replacement-omp-marginal} Assume $\mathcal{I}$ is $p$-replacement sparse. Suppose at some step, the solution is updated from $(Z_1,\cdots,Z_T)$ to $(Z'_1,\cdots,Z'_T)$ by \textsf{Replacement OMP}{}. Let $(Z^*_1,\cdots,Z^*_T) \in \argmax_{(Z_1,\cdots,Z_T) \in \mathcal{I} \colon Z_t \subseteq X^*} f_t(Z)$ where $X^*$ is an optimal solution for dictionary selection. Then, the marginal gain of \textsf{Replacement OMP}{} is bounded from below as follows: \begin{equation*} \sum_{t = 1}^T f_t(Z'_t) - \sum_{t = 1}^T f_t(Z_t) \ge \frac{1}{p} \left\{ \frac{m_{2s}}{M_{s,2}} \sum_{t = 1}^T f_t(Z^*_t) - \frac{M_{s,2}}{m_{2s}} \sum_{t = 1}^T f_t(Z_t) \right\} \end{equation*} where $s = \max_{(Z_t)_{t=1}^T \in \mathcal{I}} \max_{t \in [T]} |Z_t|$. \end{lemma} \begin{proof} We can obtain the following inequalities from the strong concavity and smoothness of $u_t$ in the same way as the above proof of \Cref{lem:replacement-greedy-marginal}. \begin{align} f_t(Z'_t) - f_t(Z_t) &\ge \frac{1}{2M_{s,2}} \norm*{\nabla u_t(\mathbf{w}_t^{(Z_t)})_{Z'_t \setminus Z_t}}^2 - \frac{M_{s,2}}{2} \norm*{(\mathbf{w}_t^{(Z_t)})_{Z_t \setminus Z'_t} }^2.\label{eq:replacement-omp-general-ineq1}\\ f_t(Z_t^*) - f_t(Z_t) &\le \frac{1}{2m_{2s}} \left\| (\nabla u_t (\mathbf{w}^{(Z_t)}))_{Z_t^* \setminus Z_t} \right\|^2 - \frac{m_{2s}}{2} \left\| (\mathbf{w}^{(Z_t)})_{Z_t \setminus Z_t^*} \right\|^2.\label{eq:replacement-omp-general-ineq2}\\ - f_t(Z_t) &\le - \frac{m_{2s}}{2} \left\| (\mathbf{w}_t^{(Z_t)})_{Z_t \setminus Z^*_t} \right\|^2.\label{eq:replacement-omp-general-ineq3} \end{align} Since $\mathcal{I}$ is $p$-replacement sparse, we can take a sequence of $p$ replacements $(Z^{p'}_1, \cdots, Z^{p'}_T)_{p'=1}^p$ that satisfies the properties mentioned in the proof of \Cref{lem:replacement-greedy-marginal}. With these properties, we obtain \begin{align*} &\sum_{t = 1}^T f_t(Z'_t) - \sum_{t = 1}^T f_t(Z_t) \\ &\ge \sum_{t = 1}^T \left\{ \frac{1}{2M_{s,2}} \norm*{\nabla u_t(\mathbf{w}_t^{(Z_t)})_{Z'_t \setminus Z_t}}^2 - \frac{M_{s,2}}{2} \norm*{(\mathbf{w}_t^{(Z_t)})_{Z_t \setminus Z'_t}}^2 \right\} \tag{by \eqref{eq:replacement-omp-general-ineq1}} \\ &\ge \frac{1}{p} \sum_{p'=1}^p \sum_{t = 1}^T \left\{ \frac{1}{2M_{s,2}} \norm*{\nabla u_t(\mathbf{w}_t^{(Z_t)})_{Z^{p'}_t \setminus Z_t}}^2 - \frac{M_{s,2}}{2} \norm*{\left( \mathbf{w}_t^{(Z_t)} \right)_{Z_t \setminus Z^{p'}_t}}^2 \right\} \tag{by the choice of $(Z_1', \dots, Z_T')$ and the feasibility of $(Z^{p'}_1,\cdots,Z^{p'}_T)$} \\ &\ge \frac{1}{p} \sum_{t = 1}^T \left\{ \frac{1}{2M_{s,2}} \norm*{\nabla u_t(\mathbf{w}_t^{(Z_t)})_{Z^*_t \setminus Z_t}}^2 - \frac{M_{s,2}}{2} \norm*{\left( \mathbf{w}_t^{(Z_t)} \right)_{Z_t \setminus Z_t^*}}^2 \right\} \tag{by the property of $(Z_t^{p'})_{p'=1}^p$} \\ &\ge \frac{1}{p} \sum_{t = 1}^T \left\{ \frac{m_{2s}}{M_{s,2}} \left( f_t(Z_t^*) - f_t(Z_t) \right) - \left( \frac{M_{s,2}}{m_{2s}} - \frac{m_{2s}}{M_{s,2}} \right) f_t(Z_t) \right\} \tag{by \eqref{eq:replacement-omp-general-ineq2} and \eqref{eq:replacement-omp-general-ineq3}} \\ &= \frac{1}{p} \sum_{t = 1}^T \left\{ \frac{m_{2s}}{M_{s,2}} f_t(Z_t^{*}) - \frac{M_{s,2}}{m_{2s}} f_t(Z_t) \right\}. \end{align*} \end{proof} Combined with \Cref{lem:misc}, we obtain \Cref{thm:replacement-omp-general}. \subsubsection{About greedy selection at each step} Next we consider how to find the atom with the largest gain at each step of \textsf{Replacement OMP}{} for the average sparsity constraints. First we show that this task reduces to weighted bipartite matching. Let us fix an atom $a^*$ because we can simply check all the atoms in $V$. Let $g_t = (\nabla u_t(\mathbf{w}^{(Z_t)}))^2_{a^*}$ and $c_t = \min_{a \in Z_t} (\mathbf{w}^{(Z_t)})^2_{a}$ for each $t \in [T]$. Let $S = \{t \in [T] \mid |Z_t| = s_t\}$ be the set of $t \in [T]$ such that the constraint on $|Z_t|$ is tight. For each $a^* \in V$, the problem of finding the best replacement can be formulated as follows: The goal is to maximize $\sum_{t \in A} g_t - \sum_{t \in B} c_t$ by selecting $A \subseteq [T]$ (the set of indices $t$ such that $a^*$ is added to $Z_t$) and $B \subseteq [T]$ (the set of indices $t$ such that an atom is removed from $Z_t$). We have two constraints on $A$ and $B$. The first constraint is $|A| - |B| \le \theta$ where $\theta = s' - \sum_{t = 1}^T |Z_t|$, derived from the total number constraint $\sum_{t=1}^T |Z_t|$. The second constraint is $A \cap S \subseteq B$, derived from the individual constraint $|Z_t| \le s_t$. In summary, the formulation as an optimization problem is: \begin{align*} \max_{A,B \subseteq [T]} \quad & \sum_{t \in A} g_t - \sum_{t \in B} c_t\\ \text{subject to} \quad & |A| - |B| \le \theta\\ & A \cap S \subseteq B. \end{align*} This problem can be regarded as a special case of maximum weight bipartite matching problem. Let $U = [T]$ and $V = [T] \cup \{d_1,\cdots,d_\theta\}$ be the set of vertices where $d_1,\cdots,d_\theta$ are dummy elements with zero cost, i.e., $c_{d_i} = 0$ for all $i \in [\theta]$. Let $E = \{(t, t) \mid t \in S\} \cup (U \setminus S) \times V$ be the set of edges. The weight of each edge $(\alpha, \beta) \in E$ is defined as $w((\alpha, \beta)) = g_\alpha - c_\beta$. Then any matching $M \subseteq E$ in this graph corresponds to a solution $A = \partial M \cap U$ and $B = \partial M \cap V \setminus \{d_1,\cdots,d_\theta\}$ in the above optimization problem. \begin{algorithm} \caption{Calculation of the gain for average sparsity constraints}\label{alg:greedy-selection-average-sparsity} \begin{algorithmic}[1] \REQUIRE $S = \{t \in [T] \mid |Z_t| = s_t\}$ the set of indices $t$ such that $Z_t$ is tight, $g_t = (\nabla u_t(\mathbf{w}^{(Z_t)}))^2_{a^*}$, $c_t = \min_{a \in Z_t} (\mathbf{w}^{(Z_t)})^2_{a}$ for each $t \in [T]$, and $\theta = s' - \sum_{t = 1}^T |Z_t|$. \ENSURE $A, B \subseteq [T]$ maximizing $\sum_{t \in A} g_t - \sum_{t \in B} c_t$ subject to $A \cap S \subseteq B$ and $|A| \le |B| + \theta$. \STATE Initialize $A_0 \gets \emptyset$ and $B_0 \gets \emptyset$. \STATE Let $S = \{t \in [T] \mid |Z_t| = s_t\}$. \STATE Sort $t \in [T] \setminus S$ according to $g_t$ into the priority queue $Q_1$ in descending order. \STATE Sort $t \in [T]$ according to $c_t$ into the priority queue $Q_2$ in ascending order. \STATE Sort $t \in S$ according to $g_t - c_t$ into the priority queue $Q_3$ in descending order. \FOR{$i = 1,\cdots,T$} \STATE Let $\alpha$, $\beta$ and $\gamma$ be the top elements in $Q_1$, $Q_2$, and $Q_3$, respectively. \IF{$g_\alpha - c_\beta \mathbf{1} \{|A_{i-1}| = |B_{i-1}| + \theta\} \le 0$ and $g_\gamma - c_\gamma \le 0$} \STATE \textbf{return} $A_{i-1}$ and $B_{i-1}$ \ELSE \IF{$g_\alpha - c_\beta \mathbf{1} \{|A_{i-1}| = |B_{i-1}| + \theta\} \ge g_\gamma - c_\gamma$} \STATE $A_i \gets A_{i-1} + \alpha$ and remove $\alpha$ from $Q_1$. \IF{$|A_{i-1}| = |B_{i-1}| + \theta$} \STATE $B_i \gets B_{i-1} + \beta$ and remove $\beta$ from $Q_2$. \IF{$\beta \in S$} \STATE Remove $\beta$ from $Q_3$ and add $\beta$ to $Q_1$. \ENDIF \ENDIF \ELSE \STATE $A_i \gets A_{i-1} + \gamma$ and $B_i \gets B_{i-1} + \gamma$. \STATE Remove $\gamma$ from $Q_3$. \ENDIF \ENDIF \ENDFOR \STATE \textbf{return} $A_T$ and $B_T$ \end{algorithmic} \end{algorithm} Here we give a fast greedy method for calculating the gain of each atom. This algorithm can be executed in $\mathrm{O}(T \log T)$ time. The detailed description of this algorithm is given in \Cref{alg:greedy-selection-average-sparsity}. \begin{proposition} \Cref{alg:greedy-selection-average-sparsity} returns an optimal solution in $\mathrm{O}(T \log T)$ time. \end{proposition} \begin{proof} First we show the validity of the algorithm. Before proving the optimality of the output, we note that the marginal gain of each step of the algorithm is largest among all the feasible updates. Let us consider the addition of $\alpha$ to $A_{i-1}$. There are three cases of updates. If $\alpha \in S \setminus B_{i-1}$ is added to $A_{i-1}$, we must also add $\alpha$ to $B_{i-1}$. If $\alpha \not\in S \setminus B_{i-1}$ and $|A_{i-1}| = |B_{i-1}| + \theta$, adding $\beta \not\in B_{i-1}$ with smallest cost $c_t$ is the best choice. If $\alpha \not\in S \setminus B_{i-1}$ and $|A_{i-1}| < |B_{i-1}| + \theta$, not changing $B_{i-1}$ is the best choice. \Cref{alg:greedy-selection-average-sparsity} selects the best one from these cases. We show $(A_i, B_i)$ be optimal among feasible solutions such that $|A| = i$ by induction on $i$. It is clear that $(A_0, B_0)$ is optimal among feasible solutions such that $|A| = 0$. Now we assume $(A_{i-1}, B_{i-1})$ is optimal among feasible solutions such that $|A| = i-1$. Let $(A'_i, B'_i)$ be an optimal solution among feasible solutions such that $|A| = i$. If there exist $\alpha \in A'_i \setminus A_{i-1}$ and $\beta \in B'_i \setminus B_{i-1}$ such that $(A_{i-1} + \alpha, B_{i-1} + \beta)$ and $(A'_i - \alpha, B'_i - \beta)$ are both feasible, we obtain \begin{align*} \sum_{t \in A_i} g_t - \sum_{t \in B_i} c_t &\ge \left(\sum_{t \in A_{i-1}} g_t - \sum_{t \in B_{i-1}} c_t \right) + \left( g_{\alpha_i} - c_{\beta_i} \right)\\ &\ge \left(\sum_{t \in A'_{i-1}} g_t - \sum_{t \in B'_{i-1}} c_t \right) + \left(g_{\alpha'} - c_{\beta'} \right)\\ &\ge \left(\sum_{t \in A'_{i}} g_t - \sum_{t \in B'_{i}} c_t \right), \end{align*} which proves the optimality of $(A_i,B_i)$. The second inequality is because the marginal gain of $\alpha_i$ (or possibly $\alpha_i$ and $\beta_i$) is largest among feasible additions. In the same way, if there exist $\alpha \in A'_i \setminus A_{i-1}$ such that $(A_{i-1} + \alpha, B_{i-1})$ and $(A'_i - \alpha, B'_i)$ are both feasible, then $(A_i,B_i)$ is optimal. We show the existence of such an $\alpha$ or pair $(\alpha, \beta)$. Since $|A'_i| > |A_{i-1}|$, we have $A'_i \setminus A_{i-1} \neq \emptyset$. Let $\alpha \in A'_i \setminus A_{i-1}$ be an arbitrary element. If $\alpha \in B'_i \setminus B_{i-1}$, the pair $(\alpha, \alpha)$ satisfies the condition. If $\alpha \not\in B'_i \setminus B_{i-1}$ and $|A_{i-1}| < |B_{i-1}| + \theta$, then $\alpha$ satisfies the condition. If $\alpha \not\in B'_i \setminus B_{i-1}$ and $|A_{i-1}| = |B_{i-1}| + \theta$, we have $|B'_i| \ge |A'_i| - \theta > |B_{i-1}|$, then $B'_i \setminus B_{i-1} \neq \emptyset$. Therefore a pair of $\alpha$ and an arbitrary $\beta \in B'_i \setminus B_{i-1}$ satisfies the condition. Finally we consider the running time of this algorithm. Sorting requires $\mathrm{O}(T \log T)$ time. Each iteration requires $\mathrm{O}(\log T)$ time. Thus, the total running time is $\mathrm{O}(T \log T)$. \end{proof} \subsection{Proof of \Cref{thm:replacement-omp-runtime}} \begin{proof} In each iteration, we need to find an atom with the largest gain and the corresponding new supports $(Z_1', \dots, Z_t')$. This can be done in $\mathrm{O}(nT\log T)$ time. Furthermore, we need to compute a new coefficient $\mathbf{w}^{(Z'_t)}_t = \mathbf{A}_{{Z'_t}}^+ \mathbf{y}_t$ for the new support ${Z'_t}$ ($t \in [T]$), where $\mathbf{A}^+$ is the pseudo inverse. This can be done efficiently via maintaining the QR-decomposition of $\mathbf{A}_{Z_t}$ under rank-two update~\citep{Golub2012matrix} with a cost of $\mathrm{O}(s^2 + ds) = \mathrm{O}(ds)$ time for each matrix. Thus each iteration requires $\mathrm{O}(T(n\log T + ds))$ time, which proves the theorem. \end{proof} \section{Missing proofs for generalized sparsity constraints} \label{sec:general-appendix} \subsection{Individual matroids} \begin{proposition}\label{prop:matroid-sparsity} An individual matroid constraint is $k$-replacement sparse. \end{proposition} \begin{proof} Let $(Z_1,\cdots,Z_T), (Z^*_1,\cdots,Z^*_T) \in \mathcal{I}$ be arbitrary sparse subsets. First we consider the case where $Z_t$ and $Z^*_t$ are both bases\footnote{For any matroid $(V, \mathcal{I})$, a set $X \in \mathcal{I}$ is called \textit{a base} if it is maximal in $\mathcal{I}$.} of the matroid for all $t \in [T]$. For such $Z_t$ and $Z^*_t$, we can make $k$ replacements as follows: For each $t \in [T]$, there exists a bijection $\pi_t \colon Z^*_t \to Z_t$ by the exchange property of matroids. For each atom $a^* \in \bigcup_{t=1}^T Z^*_t$, we make a replacement that adds $a^*$ to and removes $\pi_t(a^*)$ from $Z_t$ for all $t \in [T]$ such that $a^* \in Z^*_t$. If $Z_t$ or $Z^*_t$ is not a base of the matroid, we can add arbitrary atoms to $Z_t$ and $Z^*_t$ until they are both bases, and make $k$ replacements for them in the same way as described above. Removing the atoms that do not exist in $Z_t$ and $Z^*_t$ from these $k$ replacements, we obtain replacements for original $Z_t$ and $Z^*_t$. \end{proof} \subsection{Block sparsity} \begin{proposition} A block sparsity constraint is $k$-replacement sparse. \end{proposition} \begin{proof} Let $(Z_1,\cdots,Z_T), (Z^*_1,\cdots,Z^*_T) \in \mathcal{I}$ be arbitrary sparse subsets. We can make $k$ replacements as follows: Let $Z_{b'} = \bigcup_{t \in B_{b'}} Z_t$ and $Z^*_{b'} = \bigcup_{t \in B_{b'}} Z^*_t$. If $|Z_{b'}| < s_{b'}$ or $|Z_{b'}| < s_{b'}$, we can add arbitrary atoms until these inequalities are tight. For each block $b' \in [b]$, we can make a bijection $\pi_t \colon Z^*_b \to Z_b$. For each atom $a^* \in \bigcup_{t=1}^T Z^*_t$, we make one replacement that adds $a^*$ for all $t \in [T]$ such that $a^* \in Z^*_t$ and removes $\pi_t(a^*)$ from all blocks such that $a^* \in \bigcup_{t \in B_{b'}} Z^*_t$. \end{proof} We can show the common generalization of an individual matroid sparsity and block sparsity is also $k$-replacement sparse by combining the proofs. \subsection{Average sparsity without individual sparsity} First we consider an easier case with only a total number constraint, that is, $\mathcal{I} = \{(Z_1,\cdots,Z_T) \mid \sum_{t = 1}^T |Z_t| \le s' \}$. We call it an average sparsity constraint without individual sparsity. \begin{proposition}\label{prop:average-sparsity-replacement} An average sparsity constraint without individual sparsity is $(2k-1)$-replacement sparse. \end{proposition} \begin{proof} Let $(Z_1,\cdots,Z_T), (Z^*_1,\cdots,Z^*_T) \in \mathcal{I}$ be arbitrary feasible sparse subsets. We assume $(Z_1,\cdots,Z_T)$ and $(Z^*_1,\cdots,Z^*_T)$ are maximal in $\mathcal{I}$, but we can deal with non-maximal ones by filling them with dummy elements in the same way as the proof of \Cref{prop:matroid-sparsity}. Here we show it is possible to greedily make a sequence of $2k - 1$ feasible replacements $(Z^{r'}_1,\cdots,Z^{r'}_T)_{r'=1}^{2k-1}$ such that each atom in $Z^*_t \setminus Z_t$ appears at least once in the sequence $(Z^{r'}_t \setminus Z_t)_{r'=1}^{2k-1}$ and each atom in $Z_t \setminus Z_t^*$ appears at most once in the sequence $(Z_t \setminus Z^{r'}_t)_{r'=1}^{2k-1}$. Let $X$ and $X^*$ be the sets of atoms appearing in $(Z_1, \cdots, Z_T)$ and $(Z^*_1, \cdots, Z^*_T)$, respectively. We arrange the atoms in each of $X$ and $X^*$ in an arbitrary order and consider them one by one in parallel. Let us suppose we currently consider $a \in X$ and $a^* \in X^*$. We make a replacement that adds $a^*$ for several $t \in [T]$ and removes $a$ for the other several $t \in [T]$ in the following way. Let $\tau$ be the number of $Z_t \setminus Z^*_t$ that contains $a$, i.e., $\tau = |\{t \in [T] \mid a \in Z_t \setminus Z^*_t\}|$ and $\tau^*$ the number of $Z^*_t \setminus Z_t$ that contains $a^*$, i.e., $\tau = |\{t \in [T] \mid a^* \in Z^*_t \setminus Z_t\}|$. If $\tau > \tau^*$, we can let this replacement add $a^*$ for all $t \in [T]$ such that $a^* \in Z^*_t \setminus Z_t$ and remove $a$ for any subset of $\{t \in [T] \mid a \in Z_t \setminus Z^*_t\}$ with size $\tau^*$. Conversely, if $\tau \le \tau^*$, we can let this replacement add $a^*$ for an arbitrary subset of $\{t \in [T] \mid a^* \in Z^*_t \setminus Z_t\}$ of size $\tau$ and remove $a$ for all $t \in [T]$ such that $a \in Z_t \setminus Z^*_t$. We proceed to a next replacement after removing $a^*$ from $Z^*_t$ for all $t \in [T]$ such that $a^*$ is added in this replacement, and $a$ from $Z_t$ for all $t \in [T]$ such that $a$ is removed in this replacement. If $a \not\in Z_t \setminus Z^*_t$ for all $t \in [T]$, we move the focus from $a$ to the next atom. Similarly, if $a^* \not\in Z^*_t \setminus Z_t$ for all $t \in [T]$, we move the focus from $a$ to the next atom. This procedure ends after at most $2k - 1$ iterations. This is because at each iteration we move the focus from $a$ to the next atom in $X$ or from $a^*$ to the next atom in $X^*$, and we obtain $|X| \le k$ and $|X^*| \le k$. \end{proof} Here we show this bound is tight for an average sparsity constraint without individual sparsity by giving an example. \begin{example} Assume $T \ge k^2$. For simplicity, we further assume $T$ is a multiple of $k$. Let us consider the case of $s' = T$, i.e., $\mathcal{I} = \{(Z_1,\cdots,Z_T) \mid \sum_{t = 1}^T |Z_t| \le T\}$. Let $V = \{v_1,\cdots,v_{2k}\}$ be the ground set. Here we show the replacement sparsity parameter of this sparsity constraint is at least $2k - 1$ by giving $(Z_1,\cdots,Z_T)$ and $(Z^*_1,\cdots,Z^*_T)$ that require $2k - 1$ replacements. Suppose $Z_t = \{v_1,\cdots,v_k\}$ for $1 \le t \le T / k$ and $Z_t = \emptyset$ for other $t$. Let $Z^*_t = \{v_{k+1}\}$ for $1 \le t \le T - k + 1$ and $Z^*_{T - k + i} = \{v_{k+i}\}$ for each $i = 2, \cdots, k$. . It can be seen that we must use $k - 1$ different replacements for $Z^*_{T-k+2}, \cdots, Z^*_{T}$. In each replacement, an added element is restricted to a single atom, but $Z^*_{T-k+2}, \cdots, Z^*_{T}$ are all singleton sets of different atoms. Then elements in $Z^*_{T-k+2}, \cdots, Z^*_{T}$ must be dealt with by different replacements, and $k-1$ replacements are needed. In addition, we must use $k$ other replacements for $Z^*_1,\cdots,Z^*_{T-k+1}$. Since $(Z_1,\cdots,Z_T)$ is maximal in $\mathcal{I}$, the total number of added atoms of each replacement must be at most the total number of removed atoms of this replacement. However, in each replacement, the number of atoms removed from each $Z_t$ is at most one, and only $Z_1,\cdots,Z_{T/k}$ are non-empty, hence at most $T / k$ elements can be removed in each replacement. Therefore, we must use $k$ different replacements for $Z^*_1,\cdots,Z^*_{T-k+1}$ because there are $T - k + 1$ singleton sets $Z^*_1,\cdots,Z^*_{T-k+1}$ and $T \ge k^2$. In conclusion, the replacement sparsity parameter of this sparsity constraint is at least $2k - 1$. \end{example} \subsection{Average sparsity} We bound the replacement sparsity parameter of an average sparsity constraint based on the analysis on average sparsity without individual sparsity. \begin{proposition} An average sparsity constraint is $(3k-1)$-replacement sparse. \end{proposition} \begin{proof} Here we give a sequence of $3k - 1$ replacements that satisfies the conditions for replacement sparsity. First we use $k$ replacements for dealing with the individual sparsity constraints. Let $S \subseteq [T]$ be the set of indices such that $|Z_t| = s_t$. For each $a^* \in X^*$, we make a replacement that adds $a^*$ for all $t \in S$ such that $a^* \in Z^*_t \setminus Z_t$ and possibly removes an atom in $Z_t \setminus Z^*_t$ for all $t \in S$. By selecting the removed atoms so that they do not overlap, we can define these $k$ replacements such that, for all $t \in S$, each atom in $Z^*_t \setminus Z_t$ is added once and each atom in $Z_t \setminus Z^*_t$ is removed once. For the rest of the elements, we need not consider the individual sparsity constraints, therefore the rest elements can be dealt with $2k-1$ replacements in the same way as the proof of \Cref{prop:average-sparsity-replacement}. \end{proof} \section{Miscellaneous fact} The following folklore result is often useful for proving an approximate ratio. \begin{lemma}\label{lem:misc} Suppose that $\Delta_i, r_i \geq 0$ ($i = 1, 2, \dots$) satisfies \begin{align} \Delta_i \geq C \left(v^* - \sum_{j = 1}^{i-1} \Delta_j \right) - r_i, \end{align} for $i = 1, 2, \dots$, for some constants $C \in [0,1]$ and $v^* \geq 0$. Then \begin{align} \sum_{i=1}^l \Delta_i \geq \left[1 - (1 - C)^l \right] v^* - \sum_{i=1}^l r_i \geq (1- \exp(-Cl)) v^* - \sum_{i=1}^l r_i. \end{align} for any nonnegative integer $l$. \end{lemma} \begin{proof} We show \begin{align}\label{eq:misc} v^* - \sum_{i=1}^l \Delta_i \leq \left(1 - C \right)^l v^* + \sum_{i=1}^l r_i \end{align} for $l = 0, 1, 2, \dots$ by the induction on $l$. For $l = 0$, \eqref{eq:misc} is trivial. For $l\geq 1$, we have \begin{align*} v^* - \sum_{i=1}^l \Delta_i &= v^* - \sum_{i=1}^{l-1} \Delta_i - \Delta_l \\ &\leq v^* - \sum_{i=1}^{l-1} \Delta_i - C\left(v^* - \sum_{j = 1}^{l-1} \Delta_j \right) + r_l \\ &= \left(1 - C \right)\left(v^* - \sum_{i=1}^{l-1} \Delta_i \right) + r_l. \end{align*} Now \eqref{eq:misc} follows from the induction and $1 - C \in [0, 1]$. \end{proof} \section{Experiments}\label{sec:exp} \begin{figure*} \centering \subfigure[synthetic, $T = 100$, time]{ \includegraphics[width=0.31\textwidth]{./figures/offline_synthetic_T100_time.pdf}\label{fig:offline_synthetic_T100_time} } \subfigure[synthetic, $T = 100$, residual]{ \includegraphics[width=0.31\textwidth]{./figures/offline_synthetic_T100_error.pdf}\label{fig:offline_synthetic_T100_error} } \subfigure[voc, $T = 100$, residual]{ \includegraphics[width=0.31\textwidth]{./figures/offline_voc_T100_error.pdf}\label{fig:offline_voc_T100_error} } \subfigure[synthetic, $T = 1000$, time]{ \includegraphics[width=0.31\textwidth]{./figures/offline_synthetic_T1000_time.pdf}\label{fig:offline_synthetic_T1000_time} } \subfigure[synthetic, $T = 1000$, residual]{ \includegraphics[width=0.31\textwidth]{./figures/offline_synthetic_T1000_error.pdf}\label{fig:offline_synthetic_T1000_error} } \subfigure[voc, $T = 1000$, residual]{ \includegraphics[width=0.31\textwidth]{./figures/offline_voc_T1000_error.pdf}\label{fig:offline_voc_T1000_error} } \subfigure[synthetic, $T = 1000$, time]{ \includegraphics[width=0.31\textwidth]{./figures/offline_synthetic_average_T1000_time.pdf}\label{fig:offline_synthetic_average_T1000_time} } \subfigure[synthetic, $T = 1000$, residual]{ \includegraphics[width=0.31\textwidth]{./figures/offline_synthetic_average_T1000_error.pdf}\label{fig:offline_synthetic_average_T1000_error} } \subfigure[voc, $T = 1000$, residual]{ \includegraphics[width=0.31\textwidth]{./figures/offline_voc_average_T1000_error.pdf}\label{fig:offline_voc_average_T1000_error} } \caption{The experimental results for the offline setting. In all figures, the horizontal axis indicates the size of the output dictionary. \subref{fig:offline_synthetic_T100_time}, \subref{fig:offline_synthetic_T100_error}, and \subref{fig:offline_voc_T100_error} are the results for $T = 100$. \subref{fig:offline_synthetic_T1000_time}, \subref{fig:offline_synthetic_T1000_error}, and \subref{fig:offline_voc_T1000_error} are the results for $T = 1000$. \subref{fig:offline_synthetic_average_T1000_time}, \subref{fig:offline_synthetic_average_T1000_error}, and \subref{fig:offline_voc_average_T1000_error} are the results for $T = 1000$ with an average sparsity constraint. For each setting, we provide the plot of the running time for the synthetic dataset, test residual variance for the synthetic dataset, and test residual variance for VOC2006 image dataset.} \end{figure*} In this section, we empirically evaluate our proposed algorithms on several dictionary selection problems with synthetic and real-world datasets. We use the squared $\ell^2$-utility function for all of the experiments. Since evaluating the value of the objective function is NP-hard, we plot the approximated residual variance obtained by orthogonal matching pursuit. \paragraph{Ground set} We use the ground set consisting of several orthonormal bases that are standard choices in signal and image processing, such as 2D discrete cosine transform and several 2D discrete wavelet transforms (Haar, Daubechies 4, and coiflet). In all of the experiments, the dimension is set to $d = 64$, which corresponds to images of size $8 \times 8$ pixels. The size of the ground set is $n = 256$. \paragraph{Machine} All the algorithms are implemented in Python 3.6. We conduct the experiments in a machine with Intel Xeon E3-1225 V2 (3.20 GHz and 4 cores) and 16 GB RAM. \paragraph{Datasets} We conduct experiments on two types of datasets. The first one is a synthetic dataset. In each trial, we randomly pick a dictionary with size $k$ out of the ground set, and generate sparse linear combinations of the columns of this dictionary. The weights of the linear combinations are generated from the standard normal distribution. The second one is a dataset of real-world images extracted from PASCAL VOC2006 image datasets \citep{pascal-voc-2006}. In each trial, we randomly select an image out of 2618 images and divide it into patches of $8 \times 8$ pixels, then select $T$ patches uniformly at random. All the patches are normalized to zero mean and unit variance. We make datasets for training and test in the same way, and use the training dataset for obtaining a dictionary and the test dataset for measuring the quality of the output dictionary. \subsection{Experiments on the offline setting} We implement our proposed methods, \textsf{Replacement Greedy}{} (\textsf{RG}) and \textsf{Replacement OMP}{} (\textsf{ROMP}), as well as the existing methods for dictionary selection, $\textsf{SDS}_\textsf{MA}${} and $\textsf{SDS}_\textsf{OMP}${}. We also implement a heuristically modified version of \textsf{ROMP}, which we call \textsf{ROMPd}. In \textsf{ROMPd}, we replace $M_{s,2}$ with some parameter that decreases as the size of the current dictionary grows, which prevents the gains of all the atoms from being zero. Here we use $M_{s,2} / \sqrt{i}$ as the decreasing parameter where $i$ is the number of iterations so far. In addition, we compare these methods with standard methods for dictionary learning, \textsf{MOD} \citep{Engan1999} and \textsf{KSVD} \citep{Aharon2006}, which is set to stop when the change of the objective value becomes no more than $10^{-6}$ or 200 iterations are finished. Orthogonal matching pursuit is used as a subroutine in both methods. First, we compare the methods for dictionary selection with small datasets of $T = 100$. The parameter of sparsity constraints is set to $s = 5$. The results averaged over 20 trials are shown in \Cref{fig:offline_synthetic_T100_time}, \subref{fig:offline_synthetic_T100_error}, and \subref{fig:offline_voc_T100_error}. The plot of the running time for VOC2006 datasets is omitted as it is much similar to that for synthetic datasets. In terms of running time, $\textsf{SDS}_\textsf{MA}${} is the fastest, but the quality of the output dictionary is poor. \textsf{ROMP} is several magnitudes faster than $\textsf{SDS}_\textsf{OMP}${} and \textsf{RG}, but its quality is almost the same with $\textsf{SDS}_\textsf{OMP}${} and \textsf{RG}. In \Cref{fig:offline_synthetic_T100_error}, test residual variance of $\textsf{SDS}_\textsf{OMP}${}, \textsf{RG}, and \textsf{ROMP} are overlapped, and in \Cref{fig:offline_voc_T100_error}, test residual variance of \textsf{ROMP} is slightly worse than that of $\textsf{SDS}_\textsf{OMP}${} and \textsf{RG}. From these results, we can conclude that \textsf{ROMP} is by far the most practical method for dictionary selection. Next we compare the dictionary selection methods with the dictionary learning methods with larger datasets of $T = 1000$. $\textsf{SDS}_\textsf{OMP}${} and \textsf{RG} are omitted because they are too slow to be applied to datasets of this size. The results averaged over 20 trials are shown in \Cref{fig:offline_synthetic_T1000_time}, \subref{fig:offline_synthetic_T1000_error}, and \subref{fig:offline_voc_T1000_error}. In terms of running time, \textsf{ROMP} and \textsf{ROMPd} are much faster than \textsf{MOD} and \textsf{KSVD}, but their performances are competitive with \textsf{MOD} and \textsf{KSVD}. Finally, we conduct experiments with the average sparsity constraints. We compare \textsf{ROMP} and \textsf{ROMPd} with \Cref{alg:greedy-selection-average-sparsity} in \Cref{sec:alg-appendix} with a variant of $\textsf{SDS}_\textsf{MA}${} proposed for average sparsity in \citet{Cevher2011}. The parameters of constraints are set to $s_t = 8$ for all $t \in [T]$ and $s' = 5T$. The results averaged over 20 trials are shown in \Cref{fig:offline_synthetic_average_T1000_time}, \subref{fig:offline_synthetic_average_T1000_error}, and \subref{fig:offline_voc_average_T1000_error}. \textsf{ROMP} and \textsf{ROMPd} outperform $\textsf{SDS}_\textsf{MA}${} both in running time and quality of the output. In \Cref{sec:further}, We provide further experimental results. There we provide examples of image restoration, in which the average sparsity works better than the standard dictionary selection. \subsection{Experiments on the online setting} Here we give the experimental results on the online setting. We implement the online version of $\textsf{SDS}_\textsf{MA}${}, \textsf{RG} and \textsf{ROMP}, as well as an online dictionary learning algorithm proposed by \citet{Mairal2010}. For all the online dictionary selection methods, the hedge algorithm is used as the subroutines. The parameters are set to $k = 20$ and $s = 5$. The results averaged over 50 trials are shown in \Cref{fig:online_synthetic_k20}, \subref{fig:online_voc_k20}. For both datasets, Online \textsf{ROMP} shows a better performance than Online $\textsf{SDS}_\textsf{MA}${}, Online \textsf{RG}, and the online dictionary learning algorithm. \begin{figure*} \centering \subfigure[synthetic]{ \includegraphics[width=0.31\textwidth]{./figures/online_synthetic_k20_error.pdf}\label{fig:online_synthetic_k20} } \subfigure[voc]{ \includegraphics[width=0.31\textwidth]{./figures/online_voc_k20_error.pdf}\label{fig:online_voc_k20} } \caption{The experimental results for the online setting. In both figures, the horizontal axis indicates the number of rounds. \subref{fig:online_synthetic_k20} is the result with synthetic datasets, and \subref{fig:online_voc_k20} is the result with VOC2006 image datasets.} \end{figure*} \section{Extensions to the online setting}\label{sec:extension} Our algorithms can be extended to the following online setting. The problem is formalized as a two-player game between a player and an adversary. At each round $t = 1, \dots, T$, the player must select (possibly in a randomized manner) a dictionary $X_t \subseteq V$ with $\abs{X_t} \leq k$. Then, the adversary reveals a data point $\mathbf{y}_t \in \mathbb{R}^d$ and the player gains $f_t(X_t) = \max_{\mathbf{w} \in \mathbb{R}^k: \norm{\mathbf{w}}_0 \leq s}u(\mathbf{y}_t, \mathbf{A}_X\mathbf{w})$. The performance measure of a player's strategy is the \emph{expected $\alpha$-regret}: \begin{align*} \regret_\alpha(T) = \alpha\max_{X^*: \abs{X^*} \leq k}\sum_{t=1}^T f_t(X^*) - \E\left[ \sum_{t=1}^T f_t(X_t)\right], \end{align*} where $\alpha > 0$ is a constant independent from $T$ corresponding to the offline approximation ratio, and the expectation is taken over the randomness in the player. For this online setting, we present an extension of \textsf{Replacement Greedy}{} and \textsf{Replacement OMP}{} with sublinear $\alpha$-regret, where $\alpha$ is the corresponding offline approximation ratio. The details are provided in \Cref{sec:online}. \section{Experiments on dimensionality reduced data}\label{sec:further} In this section, we conduct experiments on the task called \emph{image restoration}. In this task, we are given an incomplete image, that is, a portion of its pixels are missing. First, we divide this incomplete image into small patches of $8 \times 8$ pixels. Then we regard each of these patches as a data point $\mathbf{y}_t$, and aim to select a dictionary that yields a sparse representation of these patches. In the procedure of the algorithms, the loss is evaluated only on the given pixels. Finally, we restore the original image by replacing each patch with a sparse approximation using the selected dictionaries, and the loss is evaluated on the whole pixels. First we conduct experiments with synthetic datasets to investigate the behavior of the algorithms. For each of the training and test datasets, we generate a bit mask such that each value takes $0$ or $1$ with equal probability. We give the masked training dataset to the algorithms and let them learn a dictionary. With this dictionary, we create the sparse representation of each data point in the test dataset with only unmasked elements and evaluate its residual variance with the whole elements. \Cref{fig:offline_synthetic_inpainting_T100_time} and \ref{fig:offline_synthetic_inpainting_T100_error} are the results for smaller datasets of $T = 100$, and \Cref{fig:offline_synthetic_inpainting_T1000_time} and \ref{fig:offline_synthetic_inpainting_T1000_error} are the results for larger datasets of $T = 1000$. In both experiments, we can see the relationship of the algorithms' performance is similar to the one in the non-masked settings, \Cref{fig:offline_synthetic_T100_time}, \ref{fig:offline_synthetic_T100_error}, \ref{fig:offline_synthetic_T1000_time}, and \ref{fig:offline_synthetic_T1000_error}. \begin{figure*} \centering \subfigure[$T = 100$, time]{ \includegraphics[width=0.4\textwidth]{./figures/offline_synthetic_inpainting_T100_time.pdf}\label{fig:offline_synthetic_inpainting_T100_time} } \subfigure[$T = 100$, residual]{ \includegraphics[width=0.4\textwidth]{./figures/offline_synthetic_inpainting_T100_error.pdf}\label{fig:offline_synthetic_inpainting_T100_error} } \subfigure[$T = 1000$, time]{ \includegraphics[width=0.4\textwidth]{./figures/offline_synthetic_inpainting_T1000_time.pdf}\label{fig:offline_synthetic_inpainting_T1000_time} } \subfigure[$T = 1000$, residual]{ \includegraphics[width=0.4\textwidth]{./figures/offline_synthetic_inpainting_T1000_error.pdf}\label{fig:offline_synthetic_inpainting_T1000_error} } \caption{The experimental results for dimensionality reduced synthetic datasets. In all figures, the horizontal axis indicates the size of the output dictionary. \subref{fig:offline_synthetic_inpainting_T100_time} and \subref{fig:offline_synthetic_inpainting_T100_error} are the results for $T = 100$. \subref{fig:offline_synthetic_inpainting_T1000_time} and \subref{fig:offline_synthetic_inpainting_T1000_error} are the results for $T = 1000$. For each setting, we give the plot of the running time and the test residual variance.} \end{figure*} In order to illustrate the advantage of the average sparsity to ordinary dictionary learning (the individual sparsity), we give image restoration examples with real-world images. We use \textsf{Replacement OMP}{} for both the individual sparsity and the average sparsity. With setting $s_t = s$ for all $t \in [T]$, the parameters $k$, $s$, and $s'$ are determined with the grid search. We apply \textsf{Replacement OMP}{} to incomplete images and obtain a dictionary. Then with this dictionary, we repeatedly compute the sparse representation of patches in the input image while shifting a single pixel. OMP is used for obtaining the sparse representation. When calculating the coefficients of the sparse representation of each patch, we use only the observed pixels and restore the whole pixels with these coefficients. We take the median value of all the restored patches for each pixel. In \Cref{fig:image}, the input image, the image restored with the individual sparsity, and the image restored with the individual sparsity are shown with PSNR ratios. The method with the average sparsity obtains higher PSNR ratios than one with the individual sparsity for all the images. \begin{figure*}[h] \centering \subfigure[Input]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/lena80_noisy.png} } \subfigure[individual, 34.42dB]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/lena80_individual.png} } \subfigure[average, 34.62dB]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/lena80_average.png} } \subfigure[Input]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/barbara80_noisy.png} } \subfigure[individual, 32.18dB]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/barbara80_individual.png} } \subfigure[average, 32.25dB]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/barbara80_average.png} } \subfigure[Input]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/house80_noisy.png} } \subfigure[individual, 33.94dB]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/house80_individual.png} } \subfigure[average, 34.14dB]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/house80_average.png} } \subfigure[Input]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/pepper80_noisy.png} } \subfigure[individual, 33.16dB]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/pepper80_individual.png} } \subfigure[average, 33.40dB]{ \includegraphics[width=0.25\textwidth]{./figures/restoration80/pepper80_average.png} } \caption{The results of the image restoration experiment from images with 80\% of pixels missing.}\label{fig:image} \end{figure*} \section{Gap between Dictionary Selection and Dictionary Learning}\label{sec:dictlearn} In this section, we analyze a theoretical gap between dictionary learning and dictionary selection. Recall that in dictionary learning, all the unit vectors are possible atoms, and therefore we cannot apply dictionary selection algorithms directly. In the following, we fix a dictionary with normalized columns $\mathbf{D}^* \in \mathbb{R}^{d\times k}$. Let $f_t \colon \mathbb{R}^{d \times k} \to \mathbb{R}$ be $f_t(\mathbf{D}) = \max_{\mathbf{w}\colon \norm{\mathbf{w}}_0 \leq s} u(\mathbf{y}_t, \mathbf{D}\mathbf{w})$. We define the distance between two dictionaries $\mathbf{D}$ and $\mathbf{D}'$ by $d(\mathbf{D}, \mathbf{D}') = \max_i \min_j \norm{\mathbf{d}_i \pm \mathbf{d}'_j}_2^2$, where $\mathbf{d}_i$ and $\mathbf{d}'_i$ are the $i$th column of $\mathbf{D}$ and $\mathbf{D}'$, respectively. The following lemma relates the regret in online dictionary selection to that in online dictionary learning. We need some notations to state the lemma. Let $\mathbf{A} \in \mathbb{R}^{d \times n}$ be a fixed matrix with normalized columns, and denote by $\bar{\mathbf{D}} \in \mathbb{R}^{d \times k}$ a column submatrix of $\mathbf{A}$ minimizing $d(\bar{\mathbf{D}}, \mathbf{D}^*)$. For $t \in [T]$, let $\mathbf{w}^*_t \in \argmax_{\mathbf{w}: \norm{\mathbf{w}}_0 \leq s} u(\mathbf{y}_t, \mathbf{D}^*\mathbf{w})$ and define $\bar{\mathbf{w}}_t$ similarly for $\bar{\mathbf{D}}$. Define $\Theta_t \subseteq \mathbb{R}^d$ as the convex hull of four points $\mathbf{D}^*\mathbf{w}^*_t$, $\mathbf{D}^*\bar\mathbf{w}_t$, $\bar\mathbf{D}\mathbf{w}^*_t$, and $\bar\mathbf{D}\bar\mathbf{w}_t$ for $t \in [T]$. \begin{theorem}\label{thm:discretization} Assume that $u(\mathbf{y}_t, \mathbf{x})$ is Lipschitz continuous in $\mathbf{x}$: there exists a constant $L > 0$ such that $\abs{u(\mathbf{y}_t, \mathbf{x}) - u(\mathbf{y}_t, \mathbf{x}')} \leq L \norm{\mathbf{x} - \mathbf{x}'}_2$ for arbitrary $\mathbf{x}, \mathbf{x}' \in \Theta_t$ and $t \in [T]$. Let $\mathbf{D}_t$ be the outputs of an online dictionary selection algorithm with the columns in $\mathbf{A}$ being the ground set. Then \begin{align*} \alpha &\sum_{t=1}^T f_t(\mathbf{D}^*) - \sum_{t=1}^T f_t(\mathbf{D}_t) \\ &\leq r_\alpha(T) + \alpha LTM\left[ \sqrt{s\epsilon} + 2\sigma_{\mathrm{max}}(\bar\mathbf{D}, 2s) \right], \end{align*} where $d(\mathbf{D}^*, \bar\mathbf{D}) \leq \epsilon$, $r_\alpha(T)$ is the $\alpha$-regret of the dictionary selection algorithm, and $M \geq \max_t \max \{ \norm{\mathbf{w}^*_t}_2, \norm{\bar\mathbf{w}_t}_2 \}$. \end{theorem} \begin{proof} Let us fix $t \in [T]$ arbitrary. For $\mathbf{y}_t$, let $\mathbf{w}_t^*, \bar{\mathbf{w}}_t \in \argmin_{\mathbf{w}\colon \norm{\mathbf{w}}\leq s} u(\mathbf{y}_t, \mathbf{D}\mathbf{w})$ for $\mathbf{D} = \mathbf{D}^*$ and $\bar\mathbf{D}$, respectively. Then, \begin{align*} &u(\mathbf{y}_t, \mathbf{D}^*\mathbf{w}_t^*) \\ &= u(\mathbf{y}_t, \bar{\mathbf{D}}\bar{\mathbf{w}}_t - (\bar{\mathbf{D}} - \mathbf{D}^*)\mathbf{w}^*_t + \bar{\mathbf{D}}(\mathbf{w}^*_t - \bar{\mathbf{w}}_t)) \\ &\leq u(\mathbf{y}_t, \bar{\mathbf{D}}\bar{\mathbf{w}}_t) + L \norm{(\bar{\mathbf{D}} - \mathbf{D}^*)\mathbf{w}^*_t + \bar{\mathbf{D}}(\mathbf{w}^*_t - \bar{\mathbf{w}}_t)}_2 \tag{by the Lipschitz continuity of $u$}\\ &\leq u(\mathbf{y}_t, \bar{\mathbf{D}}\bar{\mathbf{w}}_t) + L \norm{(\bar{\mathbf{D}} - \mathbf{D}^*)\mathbf{w}^*_t}_2 + L \norm{\bar{\mathbf{D}}(\mathbf{w}^*_t - \bar{\mathbf{w}}_t)}_2. \tag{by the triangle inequality} \end{align*} Since $\mathbf{w}^*_t$ is $s$-sparse, $\norm{(\bar{\mathbf{D}} - \mathbf{D}^*)\mathbf{w}^*_t}_2 \leq \sqrt{s\epsilon}M$. On the other hand, $\norm{\bar{\mathbf{D}}(\mathbf{w}^*_t - \bar{\mathbf{w}}_t)}_2 \leq \sigma_{\mathrm{max}}(\bar{\mathbf{D}}, 2s)\norm{\mathbf{w}^*_t - \bar\mathbf{w}_t}_2 \leq \sigma_{\mathrm{max}}(\bar{\mathbf{D}}, 2s) \cdot 2M$. Thus, \begin{align*} u(\mathbf{y}_t, \mathbf{D}^*\mathbf{w}^*) \leq u(\mathbf{y}_t, \bar{\mathbf{D}}\bar{\mathbf{w}}_t) + LM(\sqrt{s\epsilon} + 2\sigma_{\mathrm{max}}(\bar{\mathbf{D}},2s)) \end{align*} for $t \in [T]$. Therefore, \begin{align*} &\alpha \sum_{t=1}^T u(\mathbf{y}_t, \mathbf{D}^*\mathbf{w}^*_t) \\ &\leq \alpha \sum_{t=1}^T u(\mathbf{y}_t, \bar{\mathbf{D}}\bar{\mathbf{w}}_t) + \alpha LMT(\sqrt{s\epsilon} + 2\sigma_{\mathrm{max}}(\bar{\mathbf{D}},2s)) \\ &\leq \alpha \sum_{t=1}^T u(\mathbf{y}_t, \mathbf{D}_t\mathbf{w}_t) + \alpha LMT(\sqrt{s\epsilon} + 2\sigma_{\mathrm{max}}(\bar{\mathbf{D}},2s)) + r_{\alpha}(T), \end{align*} where $\mathbf{w}_t \in \argmin_{\mathbf{w}\colon \norm{\mathbf{w}}_0 \leq s} u(\mathbf{y}_t, \mathbf{D}_t \mathbf{w})$ for $t \in [T]$. By the definition of $f_t$, we obtain the desired bound. \end{proof} \begin{corollary} Let $u$ be the squared $\ell^2$-utility function. Assume that there exists $D > 0$ such that $\norm{\mathbf{y}_t}_2 \leq D$ for $t \in [T]$. Then $L \leq 2\max\{\sigma_{\mathrm{max}}(\mathbf{D}^*, s), \sigma_{\mathrm{max}}(\bar\mathbf{D}, s)\}M$ and $M \leq D/\min\{\sigma_{\mathrm{min}}(\mathbf{D}^*,s), \sigma_{\mathrm{min}}(\bar\mathbf{D}, s)\}$. \end{corollary} In the rest of this section, we focus on the squared $\ell^2$-utility function. A natural approach for reducing dictionary learning to dictionary selection is to take a discretization of the unit sphere. A finite subset $U$ of the unit sphere is called an \emph{$\epsilon$-net} for $\epsilon > 0$, if for any unit vector $\mathbf{x}$, there exists $\mathbf{y} \in U$ such that $\norm{\mathbf{x} - \mathbf{y}}^2_2 \leq \epsilon$. \begin{definition} Let $\mathbf{A} \in \mathbb{R}^{d \times n}$ be a matrix with normalized columns. The \emph{coherence} of $\mathbf{A}$ is defined as \begin{align} \mu = \max_{i \neq j} \abs{\inprod{\mathbf{a}_i, \mathbf{a}_j} }, \end{align} where $\mathbf{a}_i$ is the $i$th column of $A$. For a positive integer $s$, the \emph{restricted isometry constant (RIP constant)} $\delta_s$ with respect to $s$ is the smallest number such that \begin{align} (1-\delta_s)\norm{\mathbf{x}}_2^2 \leq \norm{\mathbf{A}\mathbf{x}}_2^2 \leq (1+\delta_s)\norm{\mathbf{x}}_2^2, \end{align} for all $\mathbf{x} \in \mathbb{R}^m$ with $\norm{\mathbf{x}}_0 \leq s$. \end{definition} It is well-known that $\mu \leq \delta_s \leq s \mu$ and $\frac{\sigma_{\mathrm{max}}(\mathbf{A}, s)}{\sigma_{\mathrm{min}}(\mathbf{A},s)} \leq \sqrt{\frac{1+\delta_s}{1-\delta_s}}$ (see \citet{Foucart2013} for example). From this perspective, an $\epsilon$-net has the coherence $1-\epsilon/2$, and therefore $\frac{\sigma_{\mathrm{max}}(\mathbf{A}, s)}{\sigma_{\mathrm{min}}(\mathbf{A},s)}$ can be arbitrary large unless $s$ is a constant. This makes the approximate-regret guarantees in \Cref{thm:modular-approx,thm:online-replacement-greedy} trivial. The following shows that for a certain range of $d$, $k$, $n$ and $s$, we can take a discretization with high probability that preserves the utility approximately with controlled dependence on the restricted condition number. \begin{proposition}\label{thm:gap} Let $\epsilon, \delta > 0$ and $p \in (0,1)$ be fixed constants and $\mathbf{D} \in \mathbb{R}^{d \times k}$ be a fixed dictionary. Let $\tilde\mathbf{A} \in \mathbb{R}^{d \times n}$ be a random matrix whose entries are the standard Gaussians and $\mathbf{A}$ be the matrix obtained by normalizing columns of $\tilde\mathbf{A}$. If $2n(\epsilon - \epsilon^2/4)^{d/2} - \log{k} \geq \log(2/p)$ and $d \geq \frac{C}{\delta^2}(s\log(n/s) + \log(4/p))$ for some constant $C \approx 80.98$, then $\mathbf{A}$ satisfies the following two conditions, with probability at least $1-p$: (i) $\delta_s \leq \delta$, and (ii) there exists a submatrix $\mathbf{D}' \in\mathbb{R}^{d\times k}$ of $\mathbf{A}$ such that $d(\mathbf{D}, \mathbf{D}') \leq \epsilon$. \end{proposition} \begin{proof} Let $\mathbf{a}_1, \dots, \mathbf{a}_n \in \mathbb{R}^d$ be the columns of $\mathbf{A}$. First, observe that for any fixed $\mathbf{d}\in S^{d-1}$, the probability of $\abs{\inprod{\mathbf{a}_i, \mathbf{d}}} \geq 1 - \epsilon/2$ is the twice of the fraction of spherical area of the set $\{ \mathbf{x} \in S^{d-1} \colon \inprod{\mathbf{a}_i, \mathbf{d}} \geq 1-\epsilon/2 \}$, which is at least $2 (\epsilon-\epsilon^2/4)^{d/2}$. Let $\mathbf{d}_1, \dots, \mathbf{d}_k \in \mathbb{R}^d$ be the columns of $\mathbf{D}^*$. Then \begin{align*} \Pr\left(\max_{i=1}^n \abs{\inprod{\mathbf{a}_i, \mathbf{d}_j}} < 1- \epsilon/2 \right) \leq (1 - 2 (\epsilon-\epsilon^2/4)^{d/2})^n \leq \exp(-2n(\epsilon-\epsilon^2/4)^{d/2}). \end{align*} By the union bound, \begin{align*} \Pr\left(\bigvee_{j=1}^k [\max_{i=1}^n\abs{\inprod{\mathbf{a}_i, \mathbf{d}_j}} < 1- \epsilon/2] \right) \leq k \exp(-2n(\epsilon-\epsilon^2/4)^{d/2}) = \exp(\log k -2n(\epsilon-\epsilon^2/42)^{d/2}). \end{align*} Therefore, by the assumption of the theorem, the probability of the condition (i) is violated is at most $p/2$. On the other hand, it is known (see~\citet{Foucart2013}, for example) that if $d \leq \frac{C}{\delta^2}\left( s\log(en/s) + \log(4/p) \right)$, then $\delta_s(\mathbf{A}) \leq \delta$ with probability at least $1-p/2$. Therefore, the conditions (i) and (ii) are both satisfied with probability at least $1-p$. \end{proof} \section{Dictionary selection with generalized sparsity constraints}\label{sec:general} In this section, we formalize our problem, \emph{dictionary selection with generalized sparsity constraints}. In this setting, the supports $Z_t$ for each $t \in [T]$ cannot be independently selected, but we impose a global constraint on them. We formally write such constraints as a down-closed~\footnote{A set family $\mathcal{I}$ is said to be down-closed if $X \in \mathcal{I}$ and $Y \subseteq X$ then $Y \in \mathcal{I}$.} family $\mathcal{I} \subseteq \prod_{t = 1}^T 2^V$. Therefore, we aim to find $X \subseteq V$ with $\abs{X} \leq k$ maximizing \begin{align} h(X) = \max_{Z_1, \dots, Z_t \subseteq X \mid (Z_1, \dots, Z_t) \in \mathcal{I}} \sum_{t=1}^T f_t(Z_t) \end{align} Since a general down-closed family is too abstract, we focus on the following class. First, we define the set of \emph{feasible replacements} for the current support $Z_1,\cdots,Z_T$ and an atom $a$ as \begin{equation}\label{eq:feasible-replacement} \mathcal{F}_{a}(Z_1,\cdots,Z_T) = \left\{(Z'_1, \cdots, Z'_T) \in \mathcal{I} \mid Z'_t \subseteq Z_T + a, \, |Z_t \setminus Z'_t| \le 1 ~ (\forall t \in [T]) \right\}. \end{equation} That is, the set of members in $\mathcal{I}$ obtained by adding $a$ and removing at most one element from each $Z_t$. Let $\mathcal{F}(Z_1,\cdots,Z_T) = \bigcup_{a \in V} \mathcal{F}_{a}(Z_1,\cdots,Z_T)$. If $Z_1, \dots, Z_T$ are clear from the context, we simply write it as $\mathcal{F}_{a}$. \begin{definition}[{$p$-replacement sparsity}] A sparsity constraint $\mathcal{I} \subseteq \prod_{t = 1}^T 2^V$ is \emph{$p$-replacement sparse} if for any $(Z_1,\dots,Z_T), (Z_1^*, \dots, Z_T^*) \in \mathcal{I}$, there is a sequence of $p$ feasible replacements $(Z^{p'}_1,\dots,Z^{p'}_T) \in \mathcal{F}(Z_1,\dots,Z_T)$ ($p' \in [p]$) such that each element in $Z^*_t \setminus Z_t$ appears at least once in the sequence $(Z^{p'}_t \setminus Z_t)_{p'=1}^p$ and each element in $Z_t \setminus Z_t^*$ appears at most once in the sequence $(Z_t \setminus Z^{p'}_t)_{p'=1}^{p}$. \end{definition} The following sparsity constraints are all $p$-replacement sparsity families. See \Cref{sec:general-appendix} for proof. \begin{example}[individual sparsity] The sparsity constraint for the standard dictionary selection can be written as $\mathcal{I} = \{(Z_1,\cdots,Z_T) \mid |Z_t| \le s ~ (\forall t \in [T])\}$. We call it \textit{the individual sparsity constraint}. This constraint is a special case of an individual matroid constraint, described below. \end{example} \begin{example}[individual matroids] This was proposed by \citep{Stan2017} as a sparsity constraint for two-stage submodular maximization. An \emph{individual matroid constraint} can be written as $\mathcal{I} = \{(Z_1,\cdots,Z_T) \mid Z_t \in \mathcal{I}_t ~ (\forall t \in [T])\}$ where $(V, \mathcal{I}_t)$ is a matroid\footnote{A \textit{matroid} is a pair of a finite ground set $V$ and a non-empty down-closed family $\mathcal{I} \subseteq 2^V$ that satisfy that for all $Z, Z' \in \mathcal{I}$ with $|Z| < |Z'|$, there is an element $a \in Z' \setminus Z$ such that $Z \cup \{a\} \in \mathcal{I}$} for each $t \in [T]$. An individual sparsity constraint is a special case of an individual matroid constraint where $(V, \mathcal{I}_t)$ is the uniform matroid for all $t$. \end{example} \begin{example}[block sparsity] Block sparsity was proposed by \citet{Krause2010}. This sparsity requires that the support must be sparse within each prespecified block. That is, disjoint blocks $B_1,\cdots,B_b \subseteq [T]$ of data points are given in advance, and an only small subset of atoms can be used in each block. Formally, $\mathcal{I} = \{(Z_1,\cdots,Z_T) \mid | \bigcup_{t \in B_{b'}} Z_t | \le s_{b'}~(\forall b' \in [b]) \}$ where $s_{b'} \in \mathbb{Z}_+$ for each $b' \in [b]$ are sparsity parameters. \end{example} \begin{example}[{average sparsity~\citep{Cevher2011}}] This sparsity imposes a constraint on the average number of used atoms among all data points. The number of atoms used for each data point is also restricted. Formally, $\mathcal{I} = \{(Z_1,\cdots,Z_T) \mid |Z_t| \le s_t, \sum_{t = 1}^T |Z_t| \le s' \}$ where $s_t \in \mathbb{Z}_{+}$ for each $t \in [T]$ and $s' \in \mathbb{Z}_{+}$ are sparsity parameters. \end{example} \begin{proposition} The replacement sparsity parameters of individual matroids, block sparsity, and average sparsity are upper-bounded by $k$, $k$, and $3k - 1$, respectively. \end{proposition} \section{Introduction} Learning sparse representations of data and signals has been extensively studied for the past decades in machine learning and signal processing~\citep{Foucart2013}. In these methods, a specific set of basis signals (atoms), called a \emph{dictionary}, is required and used to approximate a given signal in a sparse representation. The design of a dictionary is highly nontrivial, and many studies have been devoted to the construction of a good dictionary for each signal domain, such as natural images and sounds. Recently, approaches to construct a dictionary from data have shown the state-of-the-art results in various domains. The standard approach is called \emph{dictionary learning}~\citep{Arora2014,Zhou2009,Agarwal2016}. Although many studies have been devoted to dictionary learning, it is usually difficult to solve, requiring a non-convex optimization problem that often suffers from local minima. Also, standard dictionary learning methods (e.g., MOD \citep{Engan1999} or $k$-SVD \citep{Aharon2006}) require a heavy time complexity. \citet{Krause2010} proposed a combinatorial analogue of dictionary learning, called \emph{dictionary selection}. In dictionary selection, given a finite set of candidate atoms, a dictionary is constructed by selecting a few atoms from the set. Dictionary selection could be faster than dictionary learning due to its discrete nature. Another advantage of dictionary selection is that the approximation guarantees hold even in agnostic settings, i.e., we do not need stochastic generating models of the data. Furthermore, dictionary selection algorithms can be used for \emph{media summarization}, in which the atoms must be selected from given data points~\cite{Cong2012,Cong2017}. The basic dictionary selection is formalized as follows. Let $V$ be a finite set of candidate atoms and $n = \abs{V}$. Throughout the paper, we assume that the atoms are unit vectors in $\mathbb{R}^d$ without loss of generality. We represent the candidate atoms as a matrix $\mathbf{A} \in \mathbb{R}^{d\times n}$ whose columns are the atoms in $V$. Let $\mathbf{y}_t \in \mathbb{R}^d$ ($t \in [T]$) be data points, where $[T] = \{1, \dots, T\}$, and $k$ and $s$ be positive integers with $k \geq s$. We assume that a utility function $u : \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}_+$ exists, which measures the similarity of the input vectors. For example, one can use the $\ell^2$-utility function $u(\mathbf{y}, \mathbf{x}) = \norm{\mathbf{y}}_2^2 - \norm{\mathbf{y} - \mathbf{x}}_2^2$ as in \citet{Krause2010}. Then, the dictionary selection finds a set $X \subseteq V$ of size $k$ that maximizes \begin{align} h(X) = \sum_{t = 1}^T \max_{\mathbf{w} \in \mathbb{R}^k \colon \norm{\mathbf{w}}_0 \leq s} u(\mathbf{y}_t, \mathbf{A}_X \mathbf{w}), \end{align} where $\norm{\mathbf{w}}_0$ is the number of nonzero entries in $\mathbf{w}$ and $\mathbf{A}_X$ is the column submatrix of $\mathbf{A}$ with respect to $X$. That is, we approximate a data point $\mathbf{y}_t$ with a sparse representation in atoms in $X$, where the approximation quality is measured by $u$. Letting $f_t(Z_t) := \max_{\mathbf{w}} u(\mathbf{y}_t, \mathbf{A}_{Z_t}\mathbf{w})$ ($t \in [T]$), we can rewrite this as the following \emph{two-stage optimization}: $h(X) = \sum_{t = 1}^T \max_{Z_t \subseteq X \colon \abs{Z_t} \leq s} f_t(Z_t)$. Here $Z_t$ is the set of atoms used in a sparse representation of data point $\mathbf{y}_t$. The main challenges in dictionary selection are that the evaluation of $h$ is NP-hard in general~\citep{Natarajan1995}, and the objective function $h$ is not submodular~\citep{Fujishige2005} and therefore the well-known greedy algorithm~\citep{Nemhauser1978a} cannot be applied. The previous approaches construct a good proxy of dictionary selection that can be easily solved, and analyze the approximation ratio. \subsection{Our contribution} Our main contribution is a novel and efficient algorithm called the \emph{replacement orthogonal matching pursuit (\textsf{Replacement OMP}{})} for dictionary selection. This algorithm is based on a previous approach called \textsf{Replacement Greedy}{}~\citep{Stan2017} for \emph{two-stage submodular maximization}, a similar problem to dictionary selection. However, the algorithm was not analyzed for dictionary selection. We extend their approach to dictionary selection in the present work, with an additional improvement that exploits techniques in orthogonal matching pursuit. We compare our method with the previous methods in \Cref{table:comparison}. \textsf{Replacement OMP}{} has a smaller running time than $\textsf{SDS}_\textsf{OMP}$~\citep{Das2011} and \textsf{Replacement Greedy}. The only exception is $\textsf{SDS}_\textsf{MA}${}~\citep{Das2011}, which intuitively ignores any correlation of the atoms. In our experiment, we demonstrate that \textsf{Replacement OMP}{} outperforms $\textsf{SDS}_\textsf{MA}${} in terms of test residual variance. We note that the constant approximation ratios of $\textsf{SDS}_\textsf{MA}${}, \textsf{Replacement Greedy}, and \textsf{Replacement OMP}{} are incomparable in general. In addition, we demonstrate that \textsf{Replacement OMP}{} achieves a competitive performance with dictionary learning algorithms in a smaller running time, in numerical experiments. \paragraph{Generalized sparsity constraint} Incorporating further prior knowledge on the data domain often improves the quality of dictionaries~\citep{Rubinstein2010,Rusu2014,Dumitrescu2018dictionary}. A typical example is a combinatorial constraint independently imposed on each support $Z_t$. This can be regarded as a natural extension of the \emph{structured sparsity}~\cite{Huang2009} in sparse regression, which requires the support to satisfy some combinatorial constraint, rather than a cardinality constraint. A \emph{global structure} of supports is also useful prior information. \citet{Cevher2011} proposed a global sparsity constraint called the \emph{average sparsity}, in which they add a global constraint $\sum_{t=1}^T \abs{Z_t} \leq s'$. Intuitively, the average sparsity constraint requires that the most data points can be represented by a small number of atoms. If the data points are patches of a natural image, most patches are a simple background, and therefore the number of the total size of the supports must be small. The average sparsity has been also intensively studied in dictionary learning~\cite{Dumitrescu2018dictionary}. To deal with these generalized sparsities in a unified manner, we propose a novel class of sparsity constraints, namely \textit{$p$-replacement sparsity families}. We prove that \textsf{Replacement OMP}{} can be applied for the generalized sparsity constraint with a slightly worse approximation ratio. We emphasize that the OMP approach is essential for \emph{efficiency}; in contrast, \textsf{Replacement Greedy}{} cannot be extended to the average sparsity setting because it can only handle local constraints on $Z_t$, and yields an exponential running time. { \begin{table}[t] \rowcolors{1}{gray!15}{white} \centering \caption{Comparison of known methods with \textsf{Replacement OMP}. The constants $m_s$, $M_s$, and $M_{s,2}$ are the restricted concavity and smoothness constants of $u(\mathbf{y}_t, \cdot)$ ($t \in [T]$); see \Cref{sec:pre}. The running time is from the $\ell^2$-utility function $u$ and the individual sparsity constraint.}\label{table:comparison} \begin{tabular}{cccp{1.5cm}} \hline Method & Approximation ratio & Running time & Generalized sparsity \\\hline $\textsf{SDS}_\textsf{MA}$~\cite{Krause2010} & $\frac{m_1 m_s}{M_1 M_s} (1-1/\mathrm{e})$~\cite{Das2011} & $\mathrm{O}((k+d)nT)$ & No \\ $\textsf{SDS}_\textsf{OMP}$~\cite{Krause2010} & $\mathrm{O}(1/k)$~\cite{Das2011} & $\mathrm{O}((s+k)sdknT)$ & No \\ \textsf{Replacement Greedy}~\cite{Stan2017} & $\left(\frac{m_{2s}}{M_{s,2}}\right)^2\left(1-\exp\left(-\frac{M_{s,2}}{m_{2s}}\right)\right)$ & $\mathrm{O}(s^2dknT)$ & No \\ \bf this paper & $\left(\frac{m_{2s}}{M_{s,2}}\right)^2\left(1-\exp\left(-\frac{M_{s,2}}{m_{2s}}\right)\right)$ & \bf $\mathrm{O}((n+ds)kT)$ & Yes \end{tabular} \end{table} } \paragraph{Online extension} In some practical situations, it is not always feasible to store all data points $\mathbf{y}_t$, but these data points arrive in an online fashion. We show that \textsf{Replacement OMP}{} can be extended to the online setting, with a sublinear approximate regret. The details are given in \Cref{sec:extension}. \subsection{Related work} \citet{Krause2010} first introduced dictionary selection as a combinatorial analogue of dictionary learning. They proposed $\textsf{SDS}_\textsf{MA}${} and $\textsf{SDS}_\textsf{OMP}${}, and analyzed the approximation ratio using the \emph{coherence} of the matrix $\mathbf{A}$. \citet{Das2011} introduced the concept of the \emph{submodularity ratio} and refined the analysis via the \emph{restricted isometry property}~\citep{Candes2005}. A connection to the restricted concavity and submodularity ratio has been investigated by \citet{Elenberg2016,Khanna2017} for sparse regression and matrix completion. \citet{Balkanski2016} studied two-stage submodular maximization as a submodular proxy of dictionary selection, devising various algorithms. \citet{Stan2017} proposed \textsf{Replacement Greedy}{} for two-stage submodular maximization. It is unclear that these methods provide an approximation guarantee for the original dictionary selection. To the best of our knowledge, there is no existing research in the literature that addresses online dictionary selection. For a related problem in sparse optimization, namely \emph{online linear regression}, \citet{Kale2017} proposed an algorithm based on \emph{supermodular minimization}~\citep{Liberty2017} with a sublinear approximate regret guarantee. \citet{Elenberg2017} devised a streaming algorithm for weak submodular function maximization. \citet{Chen2018} dealt with online maximization of weakly DR-submodular functions. \paragraph{Organization} The rest of this paper is organized as follows. \Cref{sec:pre} provides the basic concepts and definitions. \Cref{sec:general} formally defines dictionary selection with generalized sparsity constraints. \Cref{sec:alg} presents our algorithm, \textsf{Replacement OMP}{}. \Cref{sec:extension} sketches the extension to the online setting. The experimental results are presented in \Cref{sec:exp}. \section*{Acknowledgement} The authors would thank Taihei Oki and Nobutaka Shimizu for their stimulating discussions. K.F. was supported by JSPS KAKENHI Grant Number JP 18J12405. T.S. was supported by ACT-I, JST. This work was supported by JST CREST, Grant Number JPMJCR14D2, Japan. \bibliographystyle{abbrvnat} \section{Online dictionary selection}\label{sec:online} Online dictionary selection is the problem of selecting a dictionary at each round. At each round $t$, the player selects a dictionary $X_t \subseteq V$ with $|X_t| \le k$, then the adversary reveals a data point $\mathbf{y}_t \in \mathbb{R}^d$. Then the player gains with respect to the best $s$-sparse approximation to $\mathbf{y}_t$ with the selected dictionary $X_t$: \begin{align} \max_{\mathbf{w} \colon \norm{\mathbf{w}}_0 \le s} u(\mathbf{y}_t, \mathbf{A}_{X_t} \mathbf{w}), \end{align} where $\mathbf{A}_{X_t}$ is the matrix obtained by arranging all vectors contained in $X_t$. Let $g_t(X) = \max_{Z \subseteq X \colon |Z| \le s} f_t(Z)$ be the objective function at the $t$th round, where $f_t(Z) = \max_{\mathbf{w} \colon \norm{\mathbf{w}}_0 \leq s} u(\mathbf{y}_t, A_{Z} \mathbf{w})$. In the following, we provide the online versions of algorithms for offline dictionary selection: Online $\textsf{SDS}_\textsf{MA}${}, Online \textsf{Replacement Greedy}{}, and Online \textsf{Replacement OMP}{}. \subsection{Online $\textsf{SDS}_\textsf{MA}${}}\label{subsec:modular-approx} The first algorithm is based on $\textsf{SDS}_\textsf{MA}${} for offline dictionary selection, which was proposed by \citet{Krause2010} and given an improved analysis by \citet{Das2011}. At each round $t$, we consider a function $\tilde{f}_t(Z) = \sum_{a \in Z} f_t(a|\emptyset)$, which is a modular approximation of $f_t$. Intuitively, the modular approximation $\tilde{f}_t$ ignores the interactions among the atoms. We define the surrogate objective $\tilde{g}_t$ as \begin{align} \tilde{g}_t(X) = \max_{Z \subseteq X \colon |Z| \le s} \tilde{f}_t(Z). \end{align} It is easy to show that $\tilde{g}_t$ is monotone submodular. Hence, we can apply the online greedy algorithm~\citep{Streeter2009} to these surrogate functions. Let $u_t(\mathbf{w}) \coloneqq u(\mathbf{y}_t, \mathbf{A}\mathbf{w})$ for $t \in [T]$. Assuming the strong concavity and smoothness of $u_t$, the original objective function $g_t$ can be bounded from lower and upper with the surrogate function $\tilde{g}_t$. A similar result is given in \citet{Elenberg2016} for offline sparse regression. \begin{lemma}\label{lem:surrogate} Suppose $u_t$ is $m_1$-strongly concave and $M_1$-smooth on $\Omega_1$, and $m_s$-strongly concave and $M_s$-smooth on $\Omega_s$. Then, \[ \frac{m_1}{M_s} \tilde{g}_t(X) \le g_t(X) \le \frac{M_1}{m_s} \tilde{g}_t(X). \] \end{lemma} \begin{proof} Let $Z \subseteq V$ be an arbitrary subset such that $|Z| \le s$. Since the submodularity ratio $\gamma_{\emptyset, s}$ of $f$ is no less than $m_s / M_1$~\citep{Elenberg2016}, \begin{equation*} \frac{m_s}{M_1} f_t(Z) \le \sum_{a \in Z} \tilde{f}_t(a) = \tilde{f}_t(Z). \end{equation*} As this bound holds for any $Z \subseteq V$ of size no more than $s$, we have \begin{equation*} g_t(X) = \max_{Z \subseteq X \colon |Z| \le s} f_t(Z) \le \frac{M_1}{m_s} \max_{Z \subseteq X \colon |Z| \le s} \tilde{f}_t(Z) = \frac{M_1}{m_s} \tilde{g}_t(X). \end{equation*} Next we prove the lower bound of $g_t(X)$. From the optimality of $\mathbf{w}^{(Z)}$, for any $\mathbf{z}$ such that $\mathrm{supp}(\mathbf{z}) \subseteq Z$, \begin{align*} f_t(Z) &= u_t(\mathbf{w}^{(Z)}) - u_t(\mathbf{0})\\ &\ge u_t(\mathbf{z}) - u_t(\mathbf{0})\\ &\ge \langle \nabla u_t (\mathbf{0}), \mathbf{z} \rangle - \frac{M_s}{2} \| \mathbf{z} \|^2 \end{align*} where the last inequality is due to the strong concavity of $u_t$. Using $\mathbf{z} = \frac{1}{M_s} (\nabla u_t(\mathbf{0}) )_Z$, we obtain \begin{equation}\label{eq:surrogate-bound-concave} f_t(Z) \ge \frac{1}{2M_s} \| (\nabla u_t (\mathbf{0}) )_Z \|^2. \end{equation} On the other hand, from the smoothness of $u_t$, we have for all $a \in Z$, \begin{align*} f_t(a) &= u_t(\mathbf{w}^{(a)}) - u_t(\mathbf{0})\\ &\le \langle \nabla u_t (\mathbf{0}), \mathbf{w}^{(a)} \rangle - \frac{m_1}{2} \| \mathbf{w}^{(a)} \|^2\\ &\le \max_{c \in \mathbb{R}} \langle \nabla u_t (\mathbf{0}), c \mathbf{e}_a \rangle - \frac{m_1}{2} \| c \mathbf{e}_a \|^2\\ &= \frac{1}{2m_1} (\nabla u_t(\mathbf{0}))_a^2. \end{align*} Summing up for all $a \in Z$, we obtain \begin{equation}\label{eq:surrogate-bound-smooth} \tilde{f}_t(Z) = \sum_{a \in Z} f_t(a) \le \frac{1}{2m_1} \| (\nabla u_t(\mathbf{0}))_Z \|^2. \end{equation} Combining \eqref{eq:surrogate-bound-concave} and \eqref{eq:surrogate-bound-smooth}, we obtain the lower bound \begin{equation*} f_t(Z) \ge \frac{m_1}{M_s} \tilde{f}_t(Z), \end{equation*} which proves the lower bound of $g_t(X)$ in the same way as the upper bound. \end{proof} The expected regret of this algorithm can be bounded as follows. \begin{theorem}\label{thm:modular-approx} Let $\alpha = (1 - \frac{1}{\mathrm{e}}) \frac{m_1 m_s}{M_1 M_s}$. The expected $\alpha$-regret of the modular approximation algorithm after $T$ rounds is bounded as follows: \[ \regret_\alpha(T) \le \frac{k \Delta_{\max} m_1}{M_s} \sqrt{2T \ln n} \] where $n = |V|$ and $\Delta_{\max} = \max_{a \in V} \max_{t \in [T]} f_t(a|\emptyset)$. \end{theorem} \begin{proof} Applying the regret bound for online submodular maximization \cite{Streeter2009}, we obtain \begin{align} \left( 1 - \frac{1}{\mathrm{e}} \right) \sum_{t=1}^T \tilde{g}_t(X^*) - \sum_{t=1}^T \tilde{g}_t(X_t) \le k \Delta_{\max} \sqrt{2 T \ln n}. \end{align} since the gains for each subroutine are bounded by $\Delta_{\max}$. From Lemma~\ref{lem:surrogate}, we obtain the bound in the statement. \end{proof} For the squared $\ell^2$-utility function $u(\mathbf{y},\mathbf{x}) = \frac{1}{2}\norm{\mathbf{y}}_2^2 - \frac{1}{2}\norm{\mathbf{y}-\mathbf{x}}_2^2$, $\alpha$ is equal to an approximation ratio shown in \citet{Das2011}. \begin{corollary} For the squared $\ell^2$-utility function, the expected regret of the modular approximation algorithm is \[ \regret_\alpha(T) \le \frac{k \Delta_{\max} }{\sigma_{\mathrm{max}}^2(\mathbf{A}, s)} \sqrt{2T \ln n}, \] where $\alpha = (1 - \frac{1}{\mathrm{e}})\frac{\sigma_{\mathrm{min}}^2(\mathbf{A}, s)}{\sigma_{\mathrm{max}}^2(\mathbf{A}, s)}$. \end{corollary} \subsection{Online \textsf{Replacement Greedy}{}}\label{app:online-replacement-greedy} In the following, we provide online adoptation of \textsf{Replacement Greedy}{}. Similarly to \citet{Streeter2009}, we use $k$ expert algorithms $\mathcal{A}^1, \cdots, \mathcal{A}^k$ as subroutines. At each round, online \textsf{Replacement Greedy}{} selects a set of $k$ elements $a_t^1, \cdots, a_t^k$ according to the expert algorithms $\mathcal{A}^1,\cdots,\mathcal{A}^k$, respectively. After the target point $\mathbf{y}_t$ is revealed, the algorithm decides the feedback to the subroutines by considering how $Z_t$ changes if $a_t^1,\cdots,a_t^k$ are added to $X$ sequentially. As in the offline version of \textsf{Replacement Greedy}{}, we start with $Z_t^0 = \emptyset$ and consider adding $a_t^i$ to $Z_t$ or not with keeping $|Z_t| \le s$ for each $i = 1, \cdots, k$. Denoting $Z_t$ at the $i$th step by $Z_t^i$, we can write the feedback given to the subroutine $\mathcal{A}^i$ as $\Delta_t(\cdot, Z_t^{i-1})$ where \begin{equation*} \Delta_t(a, Z_t^{i}) = \begin{cases} \displaystyle f_t(Z_t^{i} + a) - f_t(Z_t^{i}) & (i < s)\\ \displaystyle \max \left\{ 0, \max_{a' \in Z_t^{i}} \left\{ f_t(Z_t^{i} - a' + a) - f_t(Z_t^{i}) \right\} \right\} & (i \ge s) \end{cases} \end{equation*} is the gain obtained by adding $a$ to $Z_t^i$. If $\Delta_t(a^i_t, Z_t^{i-1}) > 0$, the algorithm updates $Z_t$ by adding $a_t^i$ and, if $i > s$, removing $a'$ that maximizing $f_t(Z_t^{i-1} - a' + a_t^i)$. For each $a \in V$, the value of gain $\Delta_t(a, Z_t^{i-1})$ is given to $\mathcal{A}^i$ as the feedback about $a$. A pseudocode description of our algorithm is shown in \Cref{alg:online-replacement}. \begin{theorem}\label{thm:online-replacement-greedy} Assume that $u_t$ is $m_{2s}$-strongly concave on $\Omega_{2s}$ and $M_{s,2}$-smooth on $\Omega_{s,2}$ for $t \in [T]$. Then the online replacement greedy algorithm achieves the regret bound $\regret_{\alpha}(T) \le \sum_{i=1}^k r_{i}$, where $r_i$ is the regret of the online greedy selection subroutine $\mathcal{A}^i$ for $i \in [k]$ and \[ \alpha = \left( \frac{m_{2s}}{M_{s,2}} \right)^2 \left(1 - \exp\left(- \frac{M_{s,2}}{m_{2s}} \right)\right). \] In particular, if we use the hedge algorithm as the online greedy selection subroutines, we obtain $\regret_\alpha(T) \leq k\sqrt{2T\ln{n}}$. \end{theorem} \begin{corollary} For the squared $\ell^2$-utility function, \[ \alpha \ge \left( \frac{\sigma_{\mathrm{min}}^2(\mathbf{A}, 2s)}{\sigma_{\mathrm{max}}^2(\mathbf{A}, 2)} \right)^2 \left(1 - \exp\left(- \frac{\sigma_{\mathrm{max}}^2(\mathbf{A}, 2)}{\sigma_{\mathrm{min}}^2(\mathbf{A}, 2s)} \right)\right). \] \end{corollary} \begin{proof}[Proof of \Cref{thm:online-replacement-greedy}] We provide a lower bound on the sum of the $i$th step marginal gains of the algorithm. Let $Z_t^*$ be an optimal sparse subset of $X^*$ for $f_t$, i.e., $Z_t^* \in \argmax_{Z \subseteq X^* \colon |Z| \le s} f_t(Z)$. Then we have \begin{align} \sum_{t = 1}^T \Delta_t (a_t^i, Z_{t}^{i-1}) &\ge \max_{x \in V} \sum_{t = 1}^T \Delta_t (x, Z_{t}^{i-1}) - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{a \in X^*} \sum_{t = 1}^T \Delta_t (a, Z_{t}^{i-1}) - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{t = 1}^T \sum_{a \in Z_t^*} \Delta_t (a, Z_{t}^{i-1}) - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{t = 1}^T \left(C_1 f_t(Z_i^*) - C_2 f_t(Z_{t}^{i-1}) \right) - r_i \label{eq:online-replacement-mwu-bound} \end{align} where $C_1 = \frac{m_{2s}}{M_{s,2}}$ and $C_2 = \frac{M_{s,2}}{m_{2s}}$. The first inequality is due to the regret bound for the subroutine $\mathcal{A}^i$. The last inequality is due to Lemma \ref{lem:replacement-greedy-marginal}. Now the theorem directly follows from \Cref{lem:misc}. \end{proof} \subsection{Online \textsf{Replacement OMP}{}} In this section, we consider an online version of \textsf{Replacement OMP}{}. This algorithm is the same as Online \textsf{Replacement Greedy}{} except the gain at each step. The gain obtained when $a$ is added to $Z_t^{i}$ is \begin{equation*} \displaystyle \frac{1}{2M_{s,2}} \left( \nabla u_t(\mathbf{w}_t^{(Z^i_t)})\right)^2_a \end{equation*} when $i < s$, and \begin{equation*} \displaystyle \max \left\{0, \frac{1}{2M_{s,2}} \left( \nabla u_t(\mathbf{w}_t^{(Z^i_t)})\right)^2_a - \min_{a' \in Z^i_t} \frac{M_{s,2}}{2} \left( \mathbf{w}_t^{(Z^i_t)} \right)^2_{a'} \right\} \end{equation*} when $i \ge s$, where $\mathbf{w}_t^{(Z^i_t)} \in \argmax_{\mathbf{w}: \mathrm{supp}(\mathbf{w}) \subseteq Z^i_t} u_t(\mathbf{w})$. \begin{theorem}\label{thm:online-replacement-omp} Assume that $u_t$ is $m_{2s}$-strongly concave on $\Omega_{2s}$ and $M_{s,2}$-smooth on $\Omega_{s,2}$ for $t \in [T]$. Then Online \textsf{Replacement OMP}{} algorithm achieves the regret bound $\regret_{\alpha}(T) \le \sum_{i=1}^k r_{i}$, where $r_i$ is the regret of the online greedy selection subroutine $\mathcal{A}^i$ for $i \in [k]$ and \[ \alpha = \left( \frac{m_{2s}}{M_{s,2}} \right)^2 \left(1 - \exp\left(- \frac{M_{s,2}}{m_{2s}} \right)\right). \] In particular, if we use the hedge algorithm as the online greedy selection subroutines, we obtain $\regret_\alpha(T) \leq k\sqrt{2T\ln{n}}$. \end{theorem} \begin{proof} Since $f_t$ is $M_{s,2}$-smooth on $\Omega_{s,2}$, it holds that for any $a, a' \in V$ and $Z_t \subseteq V$ of size at most $s$, \begin{equation*} f_t(Z_t - a' + a) - f_t(Z_t) \ge \frac{1}{2M_{s,2}} \left( \nabla u_t(\mathbf{w}_t^{(Z_t)}) \right)_a^2 - \frac{M_{s,2}}{2} \left( \mathbf{w}_t^{(Z_t)} \right)_{a'}^2. \end{equation*} In addition, we have \begin{equation*} \frac{1}{2M_{s,2}} \| ( \nabla u_t (\mathbf{w}^{(Z)}) )_{X \setminus Z} \|^2 - \frac{M_{s,2}}{2} \| (\mathbf{w}^{(Z)})_{Z \setminus X} \|^2 \le \frac{m_{2s}}{M_{s,2}} f(X) - \frac{M_{s,2}}{m_{s2}} f(Z) \end{equation*} from the proof of \Cref{lem:replacement-greedy-marginal}. We provide a lower bound on the $i$th step marginal gain of the algorithm. Let $Z_t^*$ be an optimal sparse subset of $X^*$ for $f_t$, i.e., $Z_t^* \in \argmax_{Z \subseteq X^* \colon |Z| \le s} f_t(Z)$. If $i \le s$, then $|Z_t^{i-1}| < s$ holds for all $t$. Then we have \begin{align} \sum_{t = 1}^T \Delta_t (a_t^i, Z_{t}^{i-1}) &= \sum_{t = 1}^T \left\{ f_t(Z_t^{i-1} + a_t^i) - f_t( Z_{t}^{i-1}) \right\}\nonumber \\ &\ge \sum_{t = 1}^T \frac{1}{2M_{s,2}} \left( \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right)^2_{a_t^i} \nonumber \\ &\ge \max_{a^i \in V} \sum_{t = 1}^T \frac{1}{2M_{s,2}} \left( \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right)^2_{a^i} - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{a \in X^*} \sum_{t = 1}^T \frac{1}{2M_{s,2}} \left( \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right)^2_{a} - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{t = 1}^T \sum_{a \in Z_t^* \setminus Z_t} \frac{1}{2M_{s,2}} \left( \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right)^2_{a} - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{t = 1}^T \frac{1}{2M_{s,2}} \left\| \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right\|^2_{Z^*_t \setminus Z_t} - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{t = 1}^T \left(\frac{m_{s,2}}{M_{s,2}} f_t(Z_i^*) - \frac{M_{s,2}}{m_{2s}} f_t(Z_{t}^{i-1}) \right) - r_i. \nonumber \end{align} Otherwise, $|Z_t^{i-1}| = s$ holds for all $t$, therefore \begin{align} \sum_{t = 1}^T \Delta_t (a_t^i, Z_{t}^{i-1}) &\ge \sum_{t = 1}^T \max \left\{ 0, \max_{a'_t \in Z_t} \left\{ f_t(Z_t - a'_t + a_t^i) - f_t(Z_t) \right\} \right\} \nonumber \\ &\ge \sum_{t = 1}^T \max \left\{ 0, \frac{1}{2M_{s,2}} \left( \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right)^2_{a_t^i} - \min_{a'_t \in Z_t} \frac{M_{s,2}}{2} \left( \mathbf{w}_t^{(Z_t)} \right)^2_{a'_t} \right\} \nonumber \\ &\ge \max_{a^i \in V} \sum_{t = 1}^T \max \left\{ 0, \frac{1}{2M_{s,2}} \left( \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right)^2_{a^i} - \min_{a'_t \in Z_t} \frac{M_{s,2}}{2} \left( \mathbf{w}_t^{(Z_t)} \right)^2_{a'_t} \right\} - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{a \in X^*} \sum_{t = 1}^T \max \left\{ 0, \frac{1}{2M_{s,2}} \left( \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right)^2_a- \min_{a'_t \in Z_t} \frac{M_{s,2}}{2} \left( \mathbf{w}_t^{(Z_t)} \right)^2_{a'_t} \right\} - r_i \nonumber\\ &\ge \frac{1}{k} \sum_{t = 1}^T \sum_{a \in Z_t^* \setminus Z_t} \left\{ \frac{1}{2M_{s,2}} \left( \nabla u_t (\mathbf{w}_t^{(Z_t)}) \right)^2_a- \frac{M_{s,2}}{2} \left( \mathbf{w}_t^{(Z_t)} \right)^2_{\pi_t(a)} \right\} - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{t = 1}^T \left\{ \frac{1}{2M_{s,2}} \left\| ( \nabla u_t (\mathbf{w}_t^{(Z_t)}) )_{Z^*_t \setminus Z_t} \right\|^2 - \frac{M_{s,2}}{2} \left\| (\mathbf{w}_t^{(Z_t)})_{Z_t \setminus Z_t^*} \right\|^2 \right\} - r_i \nonumber \\ &\ge \frac{1}{k} \sum_{t = 1}^T \left(\frac{m_{2s}}{M_{s,2}} f_t(Z_i^*) - \frac{M_{s,2}}{m_{2s}} f_t(Z_{t}^{i-1}) \right) - r_i. \label{eq:replacement-omp-mwu-bound} \end{align} where a map $\pi_t \colon Z^*_t \setminus Z_t \to Z_t \setminus Z_t^*$ is an arbitrary bijection for each $t$. Combining with \Cref{lem:misc}, we obtain the theorem. \end{proof} \begin{algorithm} \caption{Online \textsf{Replacement Greedy}{} \& Online \textsf{Replacement OMP}}\label{alg:online-replacement} \begin{algorithmic}[1] \STATE Initialize online greedy selection subroutines $\mathcal{A}^i$ for $i = 1, \dots, k$. \FOR{$t= 1, \dots, T$} \STATE Initialize $X_t^0 \gets \emptyset$ and $Z_t^0 \gets \emptyset$ for all $t \in [T]$. \FOR{$i = 1, \dots, k$} \STATE Pick $a_t^i \in V$ according to $\mathcal{A}^i$. \STATE Set $X_t^i \gets X_t^{i-1} + a_t^i$. \ENDFOR \STATE Play $X^k_t$ and observe $\mathbf{y}_t$. \FOR{$i = 1, \dots, k$} \STATE To the subroutine $\mathcal{A}_i$, feed the gain of $a$ defined as \begin{itemize} \item $\Delta_t(a, Z_{t}^{i-1})$ \hfill (Online \textsf{Replacement Greedy}) \\ \item $ \begin{cases} \displaystyle \frac{1}{M_{s,2}} \left(\nabla u_t\left(\mathbf{w}_t^{(Z_t)}\right)\right)^2_a & \text{ if $i \leq s$, } \\ \displaystyle \max \left\{0, \frac{1}{M_{s,2}} \left( \nabla u_t(\mathbf{w}_t^{(Z^{i-1}_t)})\right)^2_a - M_{s,2} \min_{a'_t \in Z^{i-1}_t} \left( \mathbf{w}_t^{(Z^{i-1}_t)} \right)^2_{a_t'} \right\} & \text{otherwise} \end{cases} $ \\\hfill(Online \textsf{Replacement OMP}) \end{itemize} \STATE Do the optimal replacement of $Z_t^{i-1}$ with respect to $a_t^i$ that achieves the above gain for \textsf{Replacement Greedy}{} or \textsf{Replacement OMP}{}, and obtain $Z_t^i$. \ENDFOR \ENDFOR \end{algorithmic} \end{algorithm} \section{Preliminaries}\label{sec:pre} \paragraph{Notation} For a positive integer $n$, $[n]$ denotes the set $\{1,2,\dots, n\}$. The sets of reals and nonnegative reals are denoted by $\mathbb{R}$ and $\mathbb{R}_{\ge 0}$, respectively. We similarly define $\mathbb{Z}$ and $\mathbb{Z}_{\ge 0}$. Vectors and matrices are denoted by lower and upper case letters in boldface, respectively: $\mathbf{a}, \mathbf{x}, \mathbf{y}$ for vectors and $\mathbf{A}, \mathbf{X}, \mathbf{Y}$ for matrices. The $i$th standard unit vector is denoted by $\mathbf{e}_i$; that is, $\mathbf{e}_i$ is the vector such that its $i$th entry is equal to one and all other entries are zero. For a matrix $\mathbf{A} \in \mathbb{R}^{d \times n}$ and $X \subseteq [n]$, $\mathbf{A}_X$ denotes the column submatrix of $\mathbf{A}$ with respect to $X$. The maximum and minimum singular values of a matrix $\mathbf{A}$ are denoted by $\sigma_{\mathrm{max}}(\mathbf{A})$ and $\sigma_{\mathrm{min}}(\mathbf{A})$, respectively. For a positive integer $k$, we define $\sigma_{\mathrm{max}}(\mathbf{A}, k) \coloneqq \max_{X \subseteq [n]\colon \abs{X} \leq k} \sigma_{\mathrm{max}}(\mathbf{A}_X)$. We define $\sigma_{\mathrm{min}}(\mathbf{A}, k)$ in a similar way. For $t \in [T]$, let $u_t(\mathbf{w}):= u(\mathbf{y}_t, \mathbf{A} \mathbf{w})$. Let $\mathbf{w}^{(Z_t)}_t$ denote the maximizer of $u_t(\mathbf{w})$ subject to $\mathrm{supp}(\mathbf{w}) \subseteq Z_t$. Throughout the paper, $V$ denotes the fixed finite ground set. For $X \subseteq V$ and $a \in V \setminus X$, we define $X + a \coloneqq X \cup \{a\}$. Similarly, for $a \in V \setminus X$ and $b \in X$, we define $X - b + a \coloneqq (X \setminus \{b\})\cup\{a\}$. \subsection{Restricted concavity and smoothness} The following concept of restricted strong concavity and smoothness is crucial in our analysis. \begin{definition}[{Restricted strong concavity and restricted smoothness~\citep{negahban2012}}] Let $\Omega$ be a subset of $\mathbb{R}^d \times \mathbb{R}^d$ and $u \colon \mathbb{R}^d \to \mathbb{R}$ be a continuously differentiable function. We say that $u$ is \emph{restricted strongly concave} with parameter $m_\Omega$ and \emph{restricted smooth} with parameter $M_\Omega$ if, \begin{align*} - \frac{m_\Omega}{2} \norm{\mathbf{y} - \mathbf{x}}_2^2 \ge u(\mathbf{y}) - u(\mathbf{x}) - \inprod{\nabla u (\mathbf{x}), \mathbf{y} - \mathbf{x}} \ge - \frac{M_\Omega}{2} \norm{\mathbf{y} - \mathbf{x}}^2_2 \end{align*} for all $(\mathbf{x}, \mathbf{y}) \in \Omega$. \end{definition} We define $\Omega_{s, p} \coloneqq \{ (\mathbf{x}, \mathbf{y}) \in \mathbb{R}^d \times \mathbb{R}^d \colon \norm{\mathbf{x}}_0, \norm{\mathbf{y}}_0 \leq s, \norm{\mathbf{x} - \mathbf{y}}_0 \leq p \}$ and $\Omega_s \coloneqq \Omega_{s,s}$ for positive integers $s$ and $p$. We often abbreviate $M_{\Omega_s}$, $M_{\Omega_{s,p}}$, and $m_{\Omega_s}$ as $M_{s}$, $M_{s,p}$, and $m_{s}$, respectively.

226fd5e90a4f958d8871c4c256eb46477c75d2f4

\section{Introduction} \label{sec:intro} The Galactic supernova remnant (SNR) \mbox {RX\,J0852.0--4622}\ was first discovered in 1998 \citep{Aschenbach:1998}, and studies of a tenuous detection of the radioactive decay line of $^{44}$Ti suggested \mbox {RX\,J0852.0--4622}\ to be young ($\sim$680\,yr) and nearby ($\sim$200\,pc) \citep{Iyudin:1998}. The $^{44}$Ti detection was and remains controversial \citep{Renaud:2006}. More recent studies estimate ages of $\sim$1 to 3\,kyr and 2.4 to 5.1\,kyr \citep[][respectively]{Katsuda:2009,Allen:2015}, with distances of $\sim$700\,pc ($\pm$200\,pc). A new relevance for this object came to light when HESS observations at TeV energies \citep[HESS\,J10852$-$463,][]{Aharonian:2005,Aharonian:2007,Abdalla:2016} revealed that \mbox {RX\,J0852.0--4622}\ is in fact a member of a class of gamma-ray SNRs that have a shell-type morphology resolved at gamma-ray energies. This reinforced the SNR's status as a key object for the study of $>$TeV cosmic-ray (CR) acceleration, which was first suggested by the presence of TeV electron acceleration \citep[][]{Slane:2001}. A key candidate production mechanism for the gamma-ray shell of \mbox {RX\,J0852.0--4622}\ is neutral pion production via CR interactions with gas and subsequent pion-decay \citep{Aharonian:2005,Aharonian:2007,Abdalla:2016}. Indeed, \textit{Fermi-LAT} GeV gamma-ray observations \citep{Tanaka:2011} reveal a spectrum that is compatible with cosmic-ray acceleration, but current spectral studies cannot distinguish this scenario from one where the gamma-rays are generated by a high energy electron population \citep{Abdalla:2016}, even if the spectral model is fitted to a broad wavelength range \citep[e.g. radio continuum data from][]{Stupar:2005}. Since neutral pion production requires the interaction of CRs with gas, attempts to identify associated gas clouds may hold the key to understanding the nature of the \mbox {RX\,J0852.0--4622} . Recent work by \citet{Fukui:2017} has successfully identified a void in atomic gas that has a near-perfect spatial match with \mbox {RX\,J0852.0--4622} , likely implying a core-collapse progenitor wind-blown bubble (see Section\,\ref{sec:gas}). \citet{Slane:2001} had previously noted that \mbox {RX\,J0852.0--4622}\ was most likely a core collapse event, although no compact object has been conclusively associated for \mbox {RX\,J0852.0--4622} . A coincident gamma-ray emitting Pulsar Wind Nebula (PWN), PSR\,J0855$-$4644, was identified in the south-east \citep{Acero:2013}, but it is believed to be unrelated despite having a compatible distance. A central X-ray source CXOU\,J085201.4$-$461753 at $\sim$1\,kpc was investigated as an association \citep[e.g.][]{Kargaltsev:2002}. \citet{Reynoso:2006} concluded that the object was likely unassociated with \mbox {RX\,J0852.0--4622}, instead favouring a planetary nebula counterpart for the compact object. Molecular gas clumps pervade the SNR boundary \citep{Fukui:2017} in a scenario where the progenitor star is argued to be associated with evaporating gaseous globules, and mirroring the molecular clumps of sister SNR, RX\,J1713.7$-$3946, which are well-studied in literature \citep[e.g.][]{Fukui:2003,Sano:2010,Maxted:2012,Maxted:2013,Fukui:2012}. \citet{Fukui:2017} find that the gamma-ray distribution traces the gas distribution, which is strong evidence for a significant hadronic gamma-ray component in the \mbox {RX\,J0852.0--4622}\ gamma-ray emission.\footnote{An alternative view is suggested by \citet{Sushch:2018} - leptonic gamma-ray emission would also exhibit a gas-gamma-ray correlation if electron-injection becomes more efficient in parts of the SNR shock moving through high gas densities. Regardless, the kinematic distance solution would likely remain valid independent of gamma-ray mechanism.} With the newly-proposed \mbox {RX\,J0852.0--4622}\ gas association, the time is right to test this scenario using new and archival multi-wavelength data-sets. \subsection{The Vela SNR}\label{sec:introVela} The Vela\,SNR is larger and overlaps \mbox {RX\,J0852.0--4622} , leading to the alternate names for \mbox {RX\,J0852.0--4622}\ - Vela\,Jr. and Vela\,Z (in addition to G266.2$-$1.2). The PWN PSR\,J0855$-$4644 emits at TeV energies, but the Vela SNR itself has not been detected in HESS gamma-ray images. It follows that the Vela SNR is not an object considered for the study of $>$TeV CR acceleration. Nevertheless, the SNR may create some radio continuum structures within the \mbox {RX\,J0852.0--4622}\ shell that may cause confusion. The Vela SNR is foreground to Vela\,Jr at a distance of $\sim$250-350\,pc \citep[e.g.][]{Cha:1999,Dubner:1998,Caraveo:2001}. The best distance estimate probably comes from parallax measurements of the associated Central Compact Object (CCO), the Vela\,pulsar, at 287$^{+19}_{-17}$\,pc \citep{Dodson:2003pulsar}. The corresponding spin-down age \citep{Reichley:1970} of 11.4\,kyr is in agreement with the age of 18$\pm$9\,kyr, derived from the angular separation between the Vela\,pulsar and the Vela SNR centroid (as indicated by `explosion fragments') given its measured proper motion \citep{Aschenbach:1995}. \citet{Sushch:2011} proposed that the evolution of Vela SNR took place inside the wind-blown bubble of the Wolf-Rayet star, $\gamma ^2$\,Velorum. The model put the Vela SNR at the east-north-eastern side of the bubble, with interactions taking place on the foreground side of the bubble. The resulting density difference between the shock component expanding into the bubble and the shock component expanding into the stellar bubble boundary might account for the asymmetry of the Vela SNR shell \citep{Sushch:2011}. \citet{Kim:2012} found this model to be consistent with the discovery of FUV filaments along regions proposed to have a higher density. The authors also noted the possible first detection of Vela\,Jr at FUV wavelengths through the examination of atomic line ratios (primarily OIII]/OIV] and OIII]/CIV]) that are more indicative of non-radiative shocks than other parts of the Vela SNR shell. The implication is that Vela\,Jr, which has non-radiative shocks, may be responsible for a component of oxygen ion emission seen by \citet[][]{Miller:1973} or \citet[][]{Nichols:2004}. This is despite low-ionisation oxygen emission generally being associated with cooling in post-shocked gas associated with older radiative shocks. While Vela\,Jr generally exhibits shock speeds of $\sim$3000\,kms$^{-1}$ \citep[][]{Katsuda:2009}, the measured velocity of the foreground Vela SNR shock front ranges between $\sim$100 and 280\,kms$^{-1}$ \citep{Cox:1972,Raymond:1991,Jenkins:1995,Bocchino:1999,Cha:2000,Bocchino:2000,Pakhomov:2012}, while so-called `explosion fragments' exhibit velocities of 660-1020\,kms$^{-1}$ \citep{Aschenbach:1995,Sushch:2011}. \citet{Redman:2000} find optical [SII] emission from one such fragment, RCW\,37. The large associated speed and temperature lead the authors to suggest that Vela\,Jr may be responsible for the fragment feature. This is in contrast to the majority of filamentary structures seen in optical atomic and ionic emission lines in the region. Emission lines such as [OIII] have been attributed to the cooling of low-density shock-heated regions associated with shock speeds of $\sim$100\,kms$^{-1}$ in the Vela SNR \citep[e.g.][]{Raymond:1997,Sankrit:2004}.\footnote{We note that such emission is also characteristic of the similar-speed shocks of Herbig-Haro objects \citep[e.g.][]{Schwartz:1978,Hartigan:1987}.} UV emission of higher transition emission lines are attributed to faster components (150-170\,kms$^{-1}$) within the Vela SNR shock \citep[e.g.][]{Raymond:1981,Slavin:2004,Nichols:2004,Sankrit:2004,Kim:2012}. We conduct a morphological investigation using radio-continuum data (including radio spectral index), HI, CO, H$\alpha$, [SII], [OIII], UV, X-ray and gamma-ray emission. We attempt to attribute components within the \mbox {RX\,J0852.0--4622}\ field of view to either \mbox {RX\,J0852.0--4622}\ (i.e. Vela\,Jr) or the overlapping Vela SNR. \subsection{An Expanded History of \mbox {RX\,J0852.0--4622}} First imaged in X-rays by the \textit{ROSAT} all-sky survey, \mbox {RX\,J0852.0--4622}\ (i.e. Vela\,Jr) was initially apparent at E$>$1.3\,keV \citep{Aschenbach:1998}. Before this X-ray discovery there were no reports of a radio SNR coincident with \mbox {RX\,J0852.0--4622}\, although many Vela and Galactic Plane surveys covered the area. \citet{Duncan:1996} and \citet{Bock:1998} surveyed the SNR region with both the Parkes radio telescope and the Molonglo Observatory Synthesis Telescope (MOST) at 2420\,MHz and 843\,MHz, respectively. Before this time, the Parkes-MIT-NRAO (PMN) survey \citep{Griffith:1993} covered the area containing \mbox {RX\,J0852.0--4622}\ at 4850\,MHz, and a candidate non-thermal source coinciding with the north-eastern limb of \mbox {RX\,J0852.0--4622}\ was identified at 408\,MHz as long ago as 1968 \citep{Milne:1968}. \citet{Combi:1999} presented a reanalysis of 2420\,MHz data and a low angular resolution (30$'$) detection of Vela\,Jr at 1420\,MHz in light of its discovery overlapping the Vela SNR at keV wavelengths. The authors concluded that the overall spectral index is $\sim -0.3$ which is flatter than expected for young SNRs (typically $\alpha \sim -0.7$)\footnote{where $S_\nu \propto \nu^\alpha$}. Follow-up observations \citep{Duncan:2000} measured spectral index ($\alpha =-0.4\pm0.15$) for \mbox {RX\,J0852.0--4622}\ that are consistently flatter than similar age SNRs in the Milky Way and Magellanic Clouds \citep[MCs,][]{Bozzetto:2017}. \citet{Filipovic:2001} showed that the structure of \mbox {RX\,J0852.0--4622}\ is shell-like (barrel-shaped/bilateral) with good correlation between radio-continuum, EUV and \textit{ROSAT} PSPC X-ray emission. The radio-continuum emission coinciding with X-rays confirmed that synchrotron radiation is responsible for the north brightened X-ray limb \citep{Bamba:2005,Pannuti:2010}. \citet{Stupar:2005} presented a multi-frequency radio-continuum study of \mbox {RX\,J0852.0--4622}\ based on low-resolution mosaic observations with the Australia Telescope Compact Array (ATCA) radio interferometer at 1384 and 2496\,MHz, Parkes 4850\,MHz and MOST 843\,MHz survey data. They determined the radio spectral index for several prominent features of this SNR and found a sudden spectral turn over at 1384\,MHz, but based on a closer inspection of the data leading to this turnover feature, we argue that observational shortcomings put the existence of this feature in doubt. This is due to the lack of short spacings within the array configuration used. In response to this, as part of this investigation we present new spectral index data towards \mbox {RX\,J0852.0--4622} . \section{Observational data} \label{sec:obs} Various observations have been carried out on \mbox {RX\,J0852.0--4622}\ over a period of 15 years. According to literature, \mbox {RX\,J0852.0--4622}\ is centred at RA(J2000) = $8^{h}52^{m}3^{s}$ and DEC(J2000) = --46\D22$'$ \citep{Aschenbach:1998} which is within the bounds of the larger, Vela SNR. In this study, we highlight GHz radio continuum, optical emission line and ultraviolet (UV) data. \subsection{ATCA Radio Continuum}\label{sec:cont} The Australia Telescope Compact Array (ATCA) is an array of six 23\,m dishes in Narrabri, New South Wales in Australia. The observations that were analysed in this study were obtained from the ATNF online archive -- ATOA\footnote {http://atoa.atnf.csiro.au/}. A list containing the observations that were analysed, which span the frequency range between 1384 and 2868\,MHz, are displayed in Table \ref{tab:observations_table}. \begin{table*} \caption{Summary of the ATCA project file radio-continuum observations of \mbox {RX\,J0852.0--4622}\ used in this study.} \center \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} l l c c c c } \hline Project & Dates & Array & $\nu$ & Bandwidth & Channels \\ Code & & & (MHz) & (MHz) & \\ \hline C789 & 14-15 Nov 1999 & 210 & 2496 & 128 & 33 \\ C2449 & 26-27 Feb 2011 & EW352 & 2100 & 2048 & 2049 \\ C2449 & 29-30 Mar 2011 & EW367 & 2100 & 2048 & 2049 \\ \hline \end{tabular*} \label{tab:observations_table} \end{table*} ATCA data from 1999 were taken as part of the Southern Galactic Plane Survey \cite[SGPS,][]{McClure:2005} and had full coverage of the the \mbox {RX\,J0852.0--4622}\ shell, while more recent measurements from 2011 were from a campaign targeting HI emission in the SNRs south-west \citep{Fukui:2017}. The latter observations were comprised of 43 pointings taken in mosaic mode and arranged in a hexagonal grid that covers approximately half of the SNR. The introduction of the Compact Array Broadband Backend system \citep[CABB,][]{Wilson:2011} to ATCA provided a factor 16 increase in the observing bandwidth compared to earlier observations of \mbox {RX\,J0852.0--4622}\ \citep{Stupar:2005,Pannuti:2010}. From 2$\times$128\,MHz to 2$\times$2048\,MHz IF bands, the addition of 16 zoom windows significantly improved the RMS noise and therefore detections of features of this SNR. With the added bandwidth and functionality from the inclusion of CABB, separate spectral line observations can be made using CABBs zoom band mode. This allowed for observations of both continuum and spectral lines simultaneously from 2011 onwards. These data were taken with an increased image-size for each pointing, facilitating effective image cleaning techniques \citep[e.g. `Peeling', as described in][]{Hughes:2006,Crawford:2011}. The properties of each image examined in this analysis can be found in Table \ref{tab:image_table}. Figure\,\ref{fig:ATCAcoverage} indicates the coverage of the 1999 and 2011 ATCA observation campaigns. Images were processed using the MIRIAD\footnote{http://www.atnf.csiro.au/computing/software/miriad/} software package \citep{Sault:1995}. Frequencies range between 1332\,MHz and 2868\,MHz and typical RMS noise levels are $\sim$1\,mJy/beam. \begin{table*} \caption{ Summary of various image properties of \mbox {RX\,J0852.0--4622}\ used in this study.} \center \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} h l c c c c c c } \hline Fig. & Project & Frequency & Bandwidth & Pixel Size & Beam Size & Position Angle & RMS \\ No. & Code & (MHz) & (MHz) & (arcsec) & (arcsec) & (degrees) & (mJy/beam) \\ \hline \ref{fig:Rad_C789} & C789 & 1384 & 128 & 51.4 & 247.6 $\times$ 179.4 & 60.4 & 1.5\phantom{0} \\ \ref{fig:Rad_C789} & C789 & 2496 & 128 & 28.9 & 118.8 $\times$ 89.0 & 55.4 & 1.0\phantom{0} \\ \ref{fig:Rad_C2449} & C2449 & 1332 & 512 & 27.2 & 120.4 $\times$ 99.0 & \p07.3 & 1.0\phantom{0} \\ \ref{fig:Rad_C2449} & C2449 & 2100 &2048 & 14.1 & 80.8 $\times$ 69.0 & \p07.7 & 0.4\phantom{0} \\ \ref{fig:Rad_C2449} & C2449 & 1844 & 512 & 20.3 & 91.6 $\times$ 77.4 & \p01.9 & 1.0\phantom{0} \\ \ref{fig:Rad_C2449} & C2449 & 2356 & 512 & 16.3 & 71.9 $\times$ 61.5 & 10.9 & 1.0\phantom{0} \\ \ref{fig:Rad_C2449} & C2449 & 2868 & 512 & 14.1 & 59.5 $\times$ 50.0 & \p08.7 & 1.0\phantom{0} \\ \hline \end{tabular*} \label{tab:image_table} \end{table*} \begin{figure}[h] \centering \includegraphics[angle=0, trim=0 0 0 0, width=0.75\columnwidth]{Vela_RadioCombo_Coverage.eps} \caption{A noise-weighted combination of ATCA radio continuum data spanning frequency bands of 1332 to 2868\,MHz. Regions of ATCA coverage from 1999 and 2011 observation campaigns are indicated by blue dashed and solid green regions, respectively. \textit{ROSAT} broad-band (0.1-2.4\,keV) X-ray count contours (5, 15, 25, 35, 45\,arcmin$^{-2}$) are overlaid, dominated by the soft thermal component of Vela SNR X-ray emission \citep{Aschenbach:1998}. \label{fig:ATCAcoverage}} \end{figure} Figure\,\ref{fig:Rad_combo} shows a final a noise-weighted combination of the 1999 and 2011 ATCA radio continuum data-sets (project codes C789 and C2449) spanning frequency bands between 1332 to 2868\,MHz. The image largely replicates features observed in images by \citet{Stupar:2005} and was produced to encapsulate the morphological features seen in individual images. The very bright, radio-loud \hbox{H\,{\sc ii}}\ region outside of the Vela\,Jr field, RCW\,38, produced large side lobes, so image boundaries were cropped prior to merging and data-sets were normalised to ensure that morphological features were continuous. The final image, which encompasses $\sim$320$^{\circ}$ of the azimuthal angle of the SNR radio shell, is displayed in Figure\,\ref{fig:Rad_combo}. This image has limitations in coverage in the South-East and in the interpretation of flux, but can clearly illustrate filaments discussed in Section\,\ref{sec:filaments}, so is ideal for morphological studies. \begin{figure}[h] \centering \includegraphics[angle=0, trim=0 0 0 0, width=0.45\columnwidth]{Vela_RadioCombo_wPMM.eps} \includegraphics[angle=0, trim=0 0 0 0, width=0.45\columnwidth]{Vela_RadioCombo_SoftX.eps}\\ \includegraphics[angle=0, trim=0 0 0 0, width=0.45\columnwidth]{Vela_RadioCombo_HardX.eps} \includegraphics[angle=0, trim=0 0 0 0, width=0.45\columnwidth]{Vela_RadioCombo.eps} \caption{A noise-weighted combination of ATCA radio continuum data spanning frequency bands of 1332 to 2868\,MHz, and years 1999 to 2011 (project codes C789 and C2449). A circle indicates the position of the strong radio source, RCW\,38, and diamond indicates the position of PSR\,J0855$-$4644. Dashed lines, labelled A-D, indicate filamentary structures discussed in Section\,\ref{sec:filaments}. In the top left image, 13\,cm Parkes radio continuum contours (0.4, 0.6, 0.8 1.2, 1.4\,Jy/beam) are overlaid \citep{Duncan:2000}. In the top right image, \textit{ROSAT} broad-band (0.1-2.4\,keV) X-ray count contours (5, 15, 25, 35, 45\,arcmin$^{-2}$) are overlaid \citep{Aschenbach:1998}. In the bottom left image, \textit{ROSAT} hard ($>$1.3\,keV) X-ray count contours (0.8, 1.2, 1.6, 2.0, 2.4, 2.8) are overlaid \citep{Aschenbach:1998}. In the bottom right image, HESS TeV gamma-ray excess count (65, 80, 95) contours are overlaid. \label{fig:Rad_combo}} \end{figure} In addition to intensity images, we also present a radio spectral index map of \mbox {RX\,J0852.0--4622}\ in Figure\,\ref{fig:SpecIndex}. The 1999 and 2011 data were examined independently, but only the 2011 results are displayed. The array configuration used to take the 1999 data lacked short-spacings and systematic effects were introduced into the corresponding spectral index map. \begin{figure}[h] \centering \includegraphics[width=0.8\columnwidth]{Vela_GasSpecIndex.eps} \caption{Spectral index derived from project C2449 (year 2011) towards the North-Western rim of \mbox {RX\,J0852.0--4622}\ (Vela\,Jr). Black, dashed radio continuum contours from a single frequency band are overlaid. Blue 2, 4, 6, 8, 10 and 12\,K\,km\,s$^{-1}$ Nanten $^{12}$CO(1-0) integrated emission (28-33\,km\,s$^{-1}$) contours are also overlaid \citep[as seen in Figure\,1 of][]{Fukui:2017}. The estimated uncertainty in spectral index is $\sim$15-20\%. \label{fig:SpecIndex}} \end{figure} \subsection{Optical Data} \label{sec:opt} This study utilises measurements from the Australia National University 16\,inch Boller \& Chivens Telescope\footnote{http://rsaa.anu.edu.au/observatories/telescopes/anu-16-inch-boller-chivens-telescope} in February 1999 \citep{Filipovic:2001}. Filters targeted the doublet [OIII], H$\alpha$ and doublet [SII] transitions at 4861.3/5006.9, 6562.8 and 6718.3/6732.7\,\r{A}, respectively, over an exposure time of 600\,seconds. We display [OIII] emission in Figure\,\ref{fig:OIII}. \begin{figure*}[h] \centering \includegraphics[width=1.0\columnwidth]{Vela_OIIIpics.eps} \caption{A 2-color image of 500\,nm [OIII] emission image \citep{Filipovic:2001} in blue and ATCA 1332-2868\,MHz radio continuum emission in red. The approximate location of the Vela SNR is indicated by a dashed-green circle. In the right-hand image, yellow dashed lines, labelled A-D, indicate filamentary structures.\label{fig:OIII}} \end{figure*} \subsection{UV Data} 58-174\,\r{A} UV images from the Extreme Ultraviolet Explorer \citep{Welsh:1990} are displayed in Figure\,\ref{fig:GasUV}. UV data was collected using 2 Wolter-Schwarzschild Type I grazing incidence mirrors with a microchannel plate detector\footnote{https://heasarc.gsfc.nasa.gov/docs/euve/euve.html}. The average exposure across the Galaxy is generally greater than 500\,s. The data was taken between 1991 and 1993 \citep{Edelstein:1993} and was utilised in previous studies by \citet{Filipovic:2001}. The data have a natural angular resolution of $\sim$6$\times$6$^{\prime}$, and was smoothed to $\sim$10$^{\prime}$ resolution. \begin{figure}[h] \centering \includegraphics[width=0.75\columnwidth]{Vela_UV.eps} \caption{A 3-colour image featuring 58-174\,\r{A} Ultra Violet emission from the EUVE satellite (red), and gas tracers at the velocity of a void. ATCA and Parkes HI emission integrated between 22 and 33\,km\,s$^{-1}$ is green, with dashed green contour levels of 150, 200, 250 and 300\,K\,km\,s$^{-1}$ is displayed \citep[as seen in Figure\,1 of][]{Fukui:2017}. Nanten $^{12}$CO(1-0) emission integrated between 28 and 33\,km\,s$^{-1}$ is blue in the image, with blue 2, 4, 6, 8, 10 and 12\,K\,km\,s$^{-1}$ contours overlaid \citep[as seen in Figure\,1 of][]{Fukui:2017}.\label{fig:GasUV}} \end{figure} \subsection{Spectral CO and HI data} \citet{Fukui:2017} presented an analysis of the interstellar medium towards \mbox {RX\,J0852.0--4622}\ with a focus on HI emission, and molecular clumps traced by CO(1-0). The authors identified a candidate interstellar medium association for \mbox {RX\,J0852.0--4622}\ and their data are used in our multi-wavelength investigation. As detailed in \citet{Fukui:2017}, 21\,cm HI data were taken with ATCA (see Section\,\ref{sec:cont}) and the 64\,m Parkes telescope. The resultant combined data-set has a beam FWHM of 245$^{\prime\prime}\times$130$^{\prime\prime}$. CO(1-0) data at 115.290\,GHz were taken with the Nanten telescope with beam FWHM of 160$^{\prime\prime}$ \section{Results and Discussion} Like in radio continuum images by \citet{Stupar:2005}, in Figure\,\ref{fig:Rad_combo} the circular structure can be clearly discerned in the north and north-west of the \mbox {RX\,J0852.0--4622}\ shell, in addition to the southern section. This \mbox {RX\,J0852.0--4622}\ (Vela\,Jr) circular structure appears consistent with Parkes 13\,cm data \citep{Duncan:2000}, hard $>$1.3\,keV X-ray emission \citep{Aschenbach:1998} and TeV gamma-ray emission \citep{Abdalla:2016}, whereas, as noted in the \citeauthor{Aschenbach:1998} Vela\,Jr discovery paper, the broadband X-ray (0.1-2.4\,keV) structure does not reflect the circular structure of Vela\,Jr. We characterise the radio emission towards \mbox {RX\,J0852.0--4622}\ as a circular shell in Section\,\ref{sec:shell} and highlight several filamentary radio continuum structures in Figure\,\ref{fig:Rad_combo} - Filament A, B, C and D. These are discussed in Section\,\ref{sec:filaments}. \subsection{Parametrising the SNR shell}\label{sec:shell} We perform $\chi^2$-minimisation fits of a 2-dimensional shell model\footnote{http://cxc.harvard.edu/sherpa/ahelp/shell2d.html} to each radio continuum data-set in Table\,\ref{tab:image_table} using SHERPA\footnote{http://cxc.cfa.harvard.edu/sherpa/} software. We masked rectangular regions encompassing three filamentary structures (A, B and C, see Section\,\ref{sec:filaments}) then employed a Neldermead fitting method \citep{Lagarias:1998}, assuming a constant noise level across each image. The centre position, SNR radius and shell width were free parameters in the fit. The best-fit parameters are displayed in Table\,\ref{tab:sherpa}. \begin{table*} \centering \caption{Results of a $\chi^2$-minimisation fit of a 2-dimensional shell function to radio continuum data. Central position (J2000 R.A.,Dec.), Radius and Width were solved for each radio continuum image. The displayed width has been deconvolved assuming a beam FWHM that is the average of the minor and major axis of the beam FWHM (see Table\,\ref{tab:image_table}).\label{tab:sherpa}} \begin{tabular}{|l|l|l|l|l|l|l|} \hline year & Central & Right & Declination & Radius$^{\dagger}$ & Width \\ & Frequency & Ascension & & & \\ & (MHz) & (deg) & (deg) & (arcsec) & (arcsec) \\ \hline 1999 & 1384 & 132.98 $_{ -0.002 } ^{+ 0.001 }$ & -46.26 $_{ -0.001 } ^{+ 0.001 }$ & 2943 $_{ -4 } ^{+ 12 }$ & 50 $_{ -2 } ^{+ 13 }$ \\ & 2496 & 132.99 $_{ -0.004 } ^{+ 0.003 }$ & -46.27 $_{ -0.002 } ^{+ 0.003 }$ & 2965 $_{ -12 } ^{+ 17 }$ & 125 $_{ -14 } ^{+ 7 }$ \\ \noalign{\smallskip} 2011 & 1332 & 133.19 $_{ -0.013 } ^{+ 0.008 }$ & -46.29 $_{ -0.004 } ^{+ 0.010 }$ & 3380 $_{ -35 } ^{+ 49 }$ & 185 $_{ -46 } ^{+ 12 }$ \\ & 1844 & 133.22 $_{ -0.010 } ^{+ 0.013 }$ & -46.27 $_{ -0.009 } ^{+ 0.007 }$ & 3433 $_{ -46 } ^{+ 44 }$ & 160 $_{ -22 } ^{+ 41 }$ \\ & 2100 & 133.19 $_{ -0.009 } ^{+ 0.010 }$ & -46.28 $_{ -0.008 } ^{+ 0.007 }$ & 3377 $_{ -48 } ^{+ 38 }$ & 157 $_{ -36 } ^{+ 26 }$ \\ & 2356 & 133.23 $_{ -0.014 } ^{+ 0.018 }$ & -46.26 $_{ -0.010 } ^{+ 0.015 }$ & 3471 $_{ -78 } ^{+ 85 }$ & 152 $_{ -46 } ^{+ 60 }$ \\ & 2868 & 133.23 $_{ -0.013 } ^{+ 0.017 }$ & -46.26 $_{ -0.013 } ^{+ 0.009 }$ & 3456 $_{ -73 } ^{+ 71 }$ & 127 $_{ -59 } ^{+ 27 }$ \\ \noalign{\smallskip} 1999+2011 & 1332-2868 & 133.08 $_{ -0.007 } ^{+ 0.010 }$ & -46.34 $_{ -0.004 } ^{+ 0.005 }$ & 3242 $_{ -34 } ^{+ 35 }$ & -$^{\ddagger}$ \\ \hline \end{tabular}\\ \footnotesize{\textit{$^{\dagger}$The radius displayed is the inner shell radius derived from the SHERPA fit plus half of the non-deconvolved shell width.}\\ \textit{$^{\ddagger}$The deconvolved shell width is not calculated, because the beam FWHM is unclear for this merged data, but an non-deconvolved width of 310$_{-34 }^{+22}$\,arcsec is found.}} \end{table*} The optimum 2-dimensional shell fit to the radio continuum suggests a centre of ($\alpha$,$\delta$ J2000)=(133.08$^{\circ}$ $_{-0.007}^{+0.010}$, $-$46.34$^{\circ}$ $_{-0.004}^{+0.005}$) (8h\,52m\,19.2\,s, $-$46$^{\circ}$20$^{\prime}$24.0$^{\prime\prime}$), and a radius of 3242$_{-34}^{+35}$\,arcsec ($\sim$0.9$^{\circ}$). We do not have sufficient coverage in the south-east of the SNR to warrant attempting an egg-shaped functional fit as suggested by X-ray emission studies \citep[e.g. ][]{Aschenbach:1998,Fukui:2017}. The $\sim$10\,arcsecond-scale errors associated with the SNR radius and the differing coverage of the 1999 and 2011 observations (see Figure\,\ref{fig:ATCAcoverage}) make the data unsuitable for expansion rate studies, particularly because previous X-ray expansion rate measurements \citep{Katsuda:2009,Allen:2015} suggest that sub-arcsecond\,yr$^{-1}$ precision is required for this purpose. \subsection{Filamentary Features}\label{sec:filaments} We have identified filamentary structures within the 1999 and 2011 ATCA data-sets that are not aligned with the circular \mbox {RX\,J0852.0--4622}\ shell. We highlight these with dashed lines in Figure\,\ref{fig:Rad_combo}. These features are also indicated on Figure\,\ref{fig:OIII}, which displays [OIII] emission, and Figure\,\ref{fig:GasUV}, which displays short-wavelength UV emission. We note that none of the identified filaments correspond to the TeV gamma-ray emission or hard X-ray emission from the \mbox {RX\,J0852.0--4622}\ shell. The unassociated PWN PSR\,J0855$-$4644 is also indicated in Figure\,\ref{fig:Rad_combo}. Coincident gamma-ray emission at this location is associated \citep{Acero:2013} with PWN PSR\,J0855$-$4644, not \mbox {RX\,J0852.0--4622} . Filament\,A is a vertical filament of emission at right-ascension $\sim$134$^{\circ}$ that partially overlaps with the north-eastern edge of \mbox {RX\,J0852.0--4622} . This feature has previously been identified as `feature\,C' by \citet{Combi:1999}. Filament\,B is a $\sim$17$^{\prime}$-length radio continuum structure centred at approximately [133.73,-46.25]. Filament\,C is a prominent feature of the 1999 and 2011 radio continuum maps which extends almost vertically along right ascension $\sim$132.5$^{\circ}$ for a length of $\sim$1.5$^{\circ}$ inside the perimeter of the \mbox {RX\,J0852.0--4622}\ shell. Filament D is a half-arcminute long filament feature, centred on approximately [133.45,-46.62], at an angle of $\sim$24$^{\circ}$ to the ecliptic plane. \subsubsection{High Energy Correspondence}\label{sec:highE} The broad-band X-ray structure in the top right image of Figure\,\ref{fig:Rad_combo} is dominated by thermal emission from the Vela SNR \citep{Aschenbach:1995,Aschenbach:1998}. This means that the broadband X-ray structure can help to distinguish the radio continuum structure of the Vela\,SNR versus Vela\,Jr, allowing us to investigate the origin of Filaments A, B, C and D. Similarly, UV emission in Figure\,\ref{fig:GasUV} highlights the older, foreground Vela SNR with perhaps a smaller contamination from the Vela\,Jr shocks due to the larger distance of Vela\,Jr. Referring to Figures \ref{fig:Rad_combo} and \ref{fig:GasUV}, filaments A, C and D have corresponding soft X-ray and UV emission. Since this emission is believed to be dominated by thermal emission from the Vela SNR, Filaments A, C and D have clear thermal counterparts, strongly suggesting an association with the Vela SNR. Furthermore, UV and soft X-ray emission towards Filament A appear to extend northwards beyond the \mbox {RX\,J0852.0--4622}\ perimeter, towards more vertical features that are likely associated with the Vela SNR (see Figures \ref{fig:Rad_combo}, \ref{fig:OIII} and \ref{fig:GasUV}). Conversely, all filaments, A, B, C and D, do not have clear correspondences with hard X-rays or gamma-ray emission, suggesting no association with \mbox {RX\,J0852.0--4622} . Filament\,B does not have any UV, X-ray or gamma-ray counterpart, so we are unable to favour either a Vela\,SNR or Vela\,Jr origin for this feature. \subsubsection{Optical Correspondence} [OIII] emission is considered a good tracer of cooling post-shock gas associated with radiative-phase SNRs, consistent with emission from the Vela\,SNR. In Figure\,\ref{fig:OIII}, the structure of the Vela\,SNR can be seen to overlap and align well with filaments A and D. These have corresponding filamentary [OIII] emission structure, as seen in Figure\,\ref{fig:GasOIII}, which shows [OIII] with [SII] and H$\alpha$ emission. Both filaments B and C have no optical counterpart. The eastern side of the Filament\,D radio continuum emission is coincident with an arc in [OIII] emission (Figure\,\ref{fig:GasOIII}). This feature, which we have coined the `Vela\,Claw' may be related to a shock interaction with gas, and is discussed in Section\,\ref{sec:gas}. Generally, a lack of [OIII] emission towards the outer circular shell of \mbox {RX\,J0852.0--4622}\ suggests that [OIII] does not generally trace the fast $\sim$10$^3$\,kms$^{-1}$ \mbox {RX\,J0852.0--4622}\ shocks, consistent with expectations. \subsection{Correspondence with the Interstellar Medium Associated with \mbox {RX\,J0852.0--4622}}\label{sec:gas} Figure\,\ref{fig:GasRadio} shows the ATCA radio continuum morphology against a backdrop of a HI-dip identified by \citet{Fukui:2017} to be likely associated with \mbox {RX\,J0852.0--4622} . The SNR shell corresponds well to this dip in HI emission (velocity $\sim$30\,kms$^{-1}$), with the eastern side partially-overlapping the contours of atomic gas, while regions of the north-east and south sit outside the lowest HI contour level within the HI-dip. It follows that ATCA radio continuum emission supports the gas-association found by \citet{Fukui:2017}. \begin{figure}[h] \centering \includegraphics[width=0.75\columnwidth]{Vela_GasAndRadio.eps} \caption{A 3-colour image featuring radio continuum (red, also in Figure\,\ref{fig:Rad_combo}), and gas tracers at the velocity of a void. ATCA and Parkes HI emission integrated between 22 and 33\,km\,s$^{-1}$ is green, with dashed green contour levels of 150, 200 and 250\,K\,km\,s$^{-1}$ is displayed \citep[as seen as Figure\,1 of][]{Fukui:2017}. Nanten $^{12}$CO(1-0) emission integrated between 28 and 33\,km\,s$^{-1}$ is blue, with blue 2, 4, 6, 8, 10 and 12\,K\,km\,s$^{-1}$ contours overlaid \citep[as seen in Figure\,1 of][]{Fukui:2017}. \label{fig:GasRadio}} \end{figure} On examination of the [OIII]-traced, Filament\,D, a forking structure was identified: we denote this feature as the Vela Claw and indicate its position and morphology in Figure\,\ref{fig:GasOIII}. The Claw appears to be part of a larger structure that is connected to the [OIII] emission coincident with Filament\, D. The [OIII] filament diverges into two filamentary structures to form a claw-like structure. This appears to occur at a location near a stripped CO clump referred to as `CO30E' by \citet{Fukui:2017}, which the authors suggested to be associated with Vela Jr. The diverging filament forms a crescent around clump CO30E from the north-west to the east, suggesting that the [OIII] emission may be associated with the Vela Jr shock. Two scenarios are considered to explain the origin of the Vela Claw: (i) Vela\,Jr generating the Vela\,Claw, and (ii) the Vela\,SNR generating the Vela\,Claw. \begin{figure}[h] \centering \includegraphics[width=0.65\columnwidth]{Vela_GasOIIIzoom.eps} \caption{\textbf{Top:} A 3-colour image featuring H$\alpha$ (red), [SII] (green) and [OIII] (blue) from the Boller \& Chivens Telescope. Red 150 and 200\,K\,km\,s$^{-1}$ contours of ATCA and Parkes HI emission integrated between 22 and 33\,km\,s$^{-1}$ are overlaid \citep[as seen as Figure\,1 of]{Fukui:2017}. Yellow 2, 4, 6, 8, 10 and 12\,K\,km\,s$^{-1}$ contours of Nanten $^{12}$CO(1-0) emission integrated between 28 and 33\,km\,s$^{-1}$ are also overlaid \citep[as seen in Figure\,1 of][]{Fukui:2017}. \textbf{Bottom:} Same as top, only red HI contours have been replaced by magenta contours that indicate \textit{ROSAT} broadband (0.2-2.4\,keV) X-ray emission \citep{Aschenbach:1998}. \label{fig:GasOIII}} \end{figure} Scenario (ii) is consistent with the accepted picture of how optical [OIII] emission is produced from SNRs, i.e. the radiative cooling of diffuse gas in the wake of a $\sim$100\,km$^{-1}$ shockwave (see Section\,\ref{sec:introVela}). Furthermore the Vela\,SNR is a known emitter of this transition, as is clearly seen extensively throughout the Vela\,SNR shell (see Figure\,\ref{fig:OIII}). It follows that if scenario (ii) is correct, the Vela Claw correspondence with clump CO30E might simply be coincidental. Indeed, this scenario is consistent with the Vela\,Claw being associated with the coincident radio continuum feature, Filament\,D, for which a Vela\,SNR origin is perhaps favoured by the coincident UV/soft X-ray structure attributable to the Vela\,SNR (see Section\,\ref{sec:highE}). Alternatively, the correspondence might be evidence for an association of clump CO30E with the Vela SNR - a scenario which is disfavoured by the remarkable correlation between gas near the CO30E Galactic velocity and the Vela\,Jr gamma-ray emission \citep{Fukui:2017}, assuming that Vela\,Jr and the Vela\,SNR are indeed at different distances ($\sim$750 and $\sim$300\,pc, respectively), as currently believed. Scenario (i), an association of the Vela Claw with Vela\,Jr, would be a surprising result because the conditions within the $\sim$3000\,kms$^{-1}$ Vela\,Jr shock are not considered conducive to stimulate optical [OIII] emission, or indeed any significant detectable thermal cooling lines at other wavelengths. Naively, an association of Vela\,Jr with clump CO30E and the Vela\,Claw might require that the SNR shock (forward or reverse) is slowed significantly by the clump CO30E density gradient. If we assume a constant ram pressure model, $P=\rho v^2$, where $\rho$ is the gas density and $v$ is the shock speed, the Vela\,Jr shock might be slowed to $\sim$10\% of the initial speed in a localised region with a sharp density gradient of $\sim$100$\times$. This is plausibly occurring for Vela\,Jr. Clump CO30E has a mass of 180\,M$_{\odot}$ and an approximate radius of 3\,pc \citep[Table 1][]{Fukui:2017}. Assuming a spherical geometry, the average H$_{2}$ density of clump CO30E is approximately n$\sim$30\,cm$^{-3}$, which would represent a $>$100$\times$ increase in density with respect to the density expected for a wind-blown cavity region (e.g. $\sim$0.1\,cm$^{-3}$), like that proposed for the evolution of Vela\,Jr \citep{Fukui:2017}. \citet{Sutherland:1995} proposed another mechanism to explain thermal emission from seven young SNRs (including Cassiopeia\,A, SNR G292.0$+$1.8 and Puppis\,A) that may also be able to describe scenario (i) in Vela\,Jr. In the \citeauthor{Sutherland:1995} model, knots of oxygen-rich ejecta material move through a low-density medium until encountering a density discontinuity at a relative velocity of several$\times$1000\,km$^{-1}$. A density increase of 100$\times$ is said to translate to internal cloud shock speeds of $\sim$100\,kms$^{-1}$, which could lead to the observed [OIII] emission. The validity of scenario (i) is complicated by the expectation that a thermal UV/X-ray emission counterpart to the [OIII] emission might be expected, as is the case for shell segments of the SNR RCW\,86, which exhibits shock velocities that vary by an order of magnitude due to sharp localised density gradients \citep[e.g.][]{Vink:2006_RCW86,Broersen:2014_RCW86}. As discussed in Section\,\ref{sec:highE}, the Vela\,SNR likely dominates the thermal UV/X-ray emission in this region, therefore scenario (i) is difficult to test with UV or X-ray data. Figure\,\ref{fig:GasOIII} (bottom) shows the significant overlap between the Vela\,Claw and \textit{ROSAT} soft thermal X-ray emission. Small-scale spectral studies of the X-ray emission might help to distinguish a thermal emission component from Vela\,Jr towards the Vela\,Claw. Such a detection would not only strengthen the association between the SNR and clump CO30E, but would be further model-dependent evidence of the core-collapse nature of the progenitor event, since the \citeauthor{Sutherland:1995} model requires oxygen-rich ejecta material inherent to core-collapse events. Also displayed in Figure\,\ref{fig:GasRadio} are CO-traced molecular gas clumps at the outskirts of the HI dip. Towards the north-west region of \mbox {RX\,J0852.0--4622}\ the signal/noise level and resolution of the 2011 ATCA data-set was sufficient across the bands to derive a reliable spectral index map. In Figure\,\ref{fig:SpecIndex}, radio continuum emission towards the rim of \mbox {RX\,J0852.0--4622}\ may flatten ($\gtrsim -$0.5) towards a dense molecular region traced by CO(1-0) emission - a scenario indicative of a SNR-cloud shock interaction \citep[e.g.][]{Keohane:1997,Ingallinera:2014}. No other suggestion of a shock-cloud interaction in this region is seen at other wavelengths, including thermal UV and X-ray emission. \section{CONCLUSION} We suggest that the 1332-2868\,MHz radio continuum emission of \mbox {RX\,J0852.0--4622}\ is well-characterised by a 2-dimensional shell of 3242$\pm{35}$\,arcsec centred at (l,b)=(133.08$^{\circ}$ $\pm{0.01}^{\circ}$,-46.34$^{\circ}\pm{0.005}^{\circ}$). Several filamentary structures are identified and multi-wavelength data is examined to investigate their relation to \mbox {RX\,J0852.0--4622} . Based \textbf{only} on morphological studies of radio continuum, UV, X-ray and gamma-ray emission, three radio filaments towards \mbox {RX\,J0852.0--4622}\ show no indication of an \mbox {RX\,J0852.0--4622}\ origin, while one filament has no clear multi-wavelength counterpart. An investigation of [OIII] emission, however, lead to the identification of a feature we coined the Vela\,Claw, which is possibly associated with one of the radio filaments. The feature corresponds to a south-east molecular clump previously suggested to be stripped by the progenitor of \mbox {RX\,J0852.0--4622}\ \citep{Fukui:2017}. Although the [OIII] feature is consistent with an origin in the shocks of the coincident Vela\,SNR, motivated by morphological correspondence with \mbox {RX\,J0852.0--4622} , we propose the possibility of \mbox {RX\,J0852.0--4622}\ triggering the [OIII] emission of the Vela\,Claw. Proof of an association would reinforce the gas association found by \citet{Fukui:2017}. \acknowledgments \section{acknowledgments} We thank the anonymous referee for their considered and constructive feedback which improved the quality of our manuscript. The Australia Telescope Compact Array is part of the Australia Telescope National Facility which is funded by the Commonwealth of Australia for operation as a National Facility managed by CSIRO. This paper includes archived data obtained through the Australia Telescope Online Archive (http://atoa.atnf.csiro.au). This research has made use of software provided by the Chandra X-ray Center (CXC) in the application packages CIAO, ChIPS, and Sherpa. \bibliographystyle{aasjournal}

fb3678975c43484365fccf8d15d217f84844eecf

\section*{Introduction} Toric degenerations as well as other flat degenerations of flag varieties have been studied extensively over the past thirty years. A fascinating aspect of this field is the large assortment of constructions from diverse research areas that produce these degenerations as well as the relationships between these constructions. We recommend~\cite{knutson} for a short overview and~\cite{fafltoric} for a more detailed survey. One of the earliest and most popular results in this regard is the Gelfand--Tsetlin (GT) toric degeneration which was first obtained implicitly (without establishing the connection with Gelfand--Tsetlin theory) in~\cite{GL} and later explicitly in~\cite{KM}. The reader is also referred to~\cite[Section 14]{MS} for an exposition and to~\cite{BCKS,Ch,Ca,NNU,L} for various generalizations and interpretations. A different source of degenerations of flag varieties is the recently popular field of PBW degenerations (overviewed below). The main goal of this paper is to link the two constructions by answering the following question: \textbf{can the Gelfand--Tsetlin toric degeneration be realized within a context similar to the theory of PBW degenerations?} Let us outline the general setting of the latter theory (we will refer to such and similar settings as ``degenerate representation theories''). For a complex semisimple Lie algebra $\mathfrak{g}$ choose a triangular decomposition and let $\mathfrak{n}_-$ be the nilpotent subalgebra spanned by negative root vectors. Consider a filtration of the universal enveloping algebra $\mathcal U(\mathfrak{n}_-)$ which respects the multiplication and provides an associated graded algebra $\gr\mathcal U(\mathfrak{n}_-)$. The chosen filtration induces a filtration on a finite-dimensional irreducible representation $L_\lambda$, the associated graded space $\gr L_\lambda$ is then naturally a $\gr\mathcal U(\mathfrak{n}_-)$-module, the degeneration of $L_\lambda$. Furthermore, for certain well-behaved filtrations it turns out that $\gr\mathcal U(\mathfrak{n}_-)$ is itself a universal enveloping algebra, hence $\gr L_\lambda$ is acted upon by the corresponding Lie group $N^{\gr}$. In $\mathbb{P}(\gr L_\lambda)$ consider the point corresponding to the line of highest weight vectors. The closure $F^{\gr}$ of the orbit of this point under the $N^{\gr}$-action is the corresponding degeneration of the flag variety. Now it should be said that the bulk of the research carried out on this subject is concerned specifically with the standard filtration by PBW degree. In this case $\gr\mathcal U(\mathfrak{n}_-)$ is the symmetric algebra $S(\mathfrak{n}_-)$ and $N^{\gr}$ is simply $\mathbb{C}^{\dim\mathfrak{n}_-}$ under addition, while $F^{\gr}$ is, in general, some non-toric singular variety known as the abelian PBW degeneration. The (ongoing) studies of this particular setting have resulted in numerous works by a wide range of authors, with some of the key results being obtained in the papers~\cite{Fe2,FFL1,ABS,FFL2,CFR,CL}. However, to us it is important that this scenario can be generalized beyond the standard PBW filtration to obtain a wider class of flat degenerations. In particular, the papers~\cite{favourable} and~\cite{fafefom} show that certain toric varieties can be obtained as the degeneration $F^{\gr}$. Both papers obtain the same concrete example: the toric variety associated with the FFLV polytope (named after the authors of~\cite{FFL1} and Vinberg who was the first to consider it conjecturally). This raises the problem of obtaining other toric degenerations with similar methods, in particular, the GT toric degeneration. Below we summarize our proposed solution(s). Since our results bear structural similarities to those in~\cite{fafefom}, we first recall the key points in that paper. In~\cite{fafefom} a $\mathbb{Z}$-degree is assigned to every negative root vector and the algebra $\mathcal U(\mathfrak{n}_-)$ is filtered by this modified PBW degree. It is shown there that when the degrees satisfy certain linear inequalities, $\gr\mathcal U(\mathfrak{n}_-)$ is indeed a universal enveloping algebra and $F^{\gr}$ is a flat Gr\"obner degeneration of the flag variety (i.e. the variety cut out by an initial ideal of the ideal of Pl\"ucker relations). Furthermore, the mentioned inequalities cut out a polyhedral cone, for points in the interior of which $F^{\gr}$ is the FFLV toric variety. This cone parametrizes the Gr\"obner degenerations in consideration and turns out to be a maximal cone in the tropical flag variety. An important role in~\cite{fafefom} is played by monomial bases in $L_\lambda^{\gr}$ that are parametrized by the integer points in FFLV poytopes (FFLV bases). Now, in this paper we also consider a family of Gr\"obner degenerations $F^S$ (generalizing the GT toric degeneration), that are given by $\mathbb{Z}$-gradings $S$ on the Pl\"ucker variables satisfying certain linear inequalities. We show that each such Gr\"obner degeneration $F^S$ defines a filtration on every $L_\lambda$ which produces the degenerate representations $L_\lambda^S$. These filtrations may also be obtained representation-theoretically, however, the key difference from~\cite{fafefom} is that we first define a certain deformation $\Phi_n$ of $\mathcal U(\mathfrak{n}_-)$ that acts on every $L_\lambda$ and degenerate the action of this algebra rather than that of $\mathcal U(\mathfrak{n}_-)$. The algebra $\Phi_n$ is generated by $\varphi_{i,j}$ with $1\le i<j\le n$ (i.e. corresponding to the root vectors in $\mathfrak{n}_-$) and is filtered by assigning $\mathbb{Z}$-degrees to these $\varphi_{i,j}$. Thus we obtain a $\Phi_n$ action on every $L_\lambda^S$, since the associated graded algebras are also isomorphic to $\Phi_n$. However, $\Phi_n$ is not a universal enveloping algebra and we need to provide a strategy of recovering the degenerate flag variety from of the degenerate representation theory that doesn't rely on a group action. We propose two such strategies, the first one is to imitate the group action by taking exponentials. \begin{customthm}{A}[cf. Theorem~\ref{main}]\label{thma} $F^S$ is isomorphic to the closure of the set of points of the form $\prod_{i,j}\exp(c_{i,j}\varphi_{i,j})\bm v_\lambda^S\in\mathbb{P}(L_\lambda^S)$ where $c_{i,j}\in\mathbb{C}$ and $v_\lambda^S$ is the highest weight line. \end{customthm} The second strategy relies on the existence of tensor products and Cartan components for our degenerate representations, which is also one of the key properties of the degenerate representation theories in~\cite{FFL1,favourable,fafefom} et al. In our case we have a $\Phi_n$-module structure on $L_\lambda^S\otimes L_\mu^S$ such that the submodule $\Phi_n(\bm v_\lambda^S\otimes \bm v_\mu^S)$ is isomorphic to $L_{\lambda+\mu}^S$. This provides a map $(L_\lambda^S)^*\otimes(L_\mu^S)^*\to (L_{\lambda+\mu}^S)^*$ and a commutative algebra structure on $\mathcal P^S=\bigoplus_\lambda (L_\lambda^S)^*$. The degenerate flag variety can be defined as the ``multigraded $\mathrm{Proj}$'' of $\mathcal P^S$, in other words, we prove the following \begin{customthm}{B}[cf. Theorem~\ref{mainproj}]\label{thmb} $\mathcal P^S$ is generated by the components $(L_{\omega_i}^S)^*$ (for fundamental weights $\omega_i$) and the kernel of the surjection $\mathbb{C}[\bigoplus_i (L_{\omega_i}^S)^*]\to \mathcal P^S$ cuts out a subvariety in $\bigtimes_i\mathbb{P}(L_{\omega_i}^S)$ that is isomorphic to $F^S$. \end{customthm} This characterization of the variety is implicit in the earlier work on various PBW degenerations. It is certainly not as simple as the orbit closure definition but has the advantage of being independent of the highest weight. In our setting, however, it is arguably simpler and more natural than the above trick with exponentials. Let us provide some details concerning the structure of this paper. In Section~\ref{generalities} we consider a general Gr\"obner degeneration $F^S$ and show that in this generality we already obtain a certain $\mathbb{Z}$-filtration on each irreducible representation $L_\lambda$ which provides the corresponding degeneration $L_\lambda^S=\gr L_\lambda$. Moreover, $F^S$ embeds naturally into $\mathbb{P}(L_\lambda^S)$. This shows that once we have realized our toric variety as a Gr\"obner degeneration (which was done already by~\cite{GL}) the associated graded spaces as well as the embeddings into their projectivizations are already fixed and we are to establish a fitting representation theory. In Section~\ref{bases} we define Gelfand--Tsetlin polytopes together with certain unimodular transforms $P_\lambda$ of theirs. In a sense, the toric varieties we consider are defined by the latter rather than the former. Moreover, in analogy with FFLV bases, the integer points in $P_\lambda$ also enumerate a monomial basis in the $L_\lambda$ (and its degenerations). These bases are instrumental to the subsequent arguments just like FFLV bases in~\cite{fafefom}. It should be pointed out that these bases in $L_\lambda$ itself are not new and were, apparently, first obtained in~\cite{R} but without the connection with Gelfand--Tsetlin polytopes (also cf.~\cite{MY}). While Sections~\ref{generalities} through~\ref{gtdegens} can be viewed as largely preparatory, Section~\ref{gtdegens} does establish some key ideas crucial to the main results. Here we define the set of Gr\"obner degenerations $F^S$ we will be considering (the ``Gelfand--Tsetlin degenerations'') and then prove the existence of the monomial bases defined in the previous section with the use of this family of Gr\"obner degenerations. We also show that in the interior of the parametrizing cone these Gelfand--Tsetlin degenerations are isomorphic to the Gelfand--Tsetlin toric variety. It should be acknowledged that this latter fact could have already been obtained with the methods of~\cite{KM}. One could say that~\cite{KM} considers one specific point in the cone's interior but their argument would be valid for any other point. (This is in contrast with the new converse observation from Section~\ref{tropical} that no point {\it outside} of the cone would work.) In Section~\ref{gtdegens} we also show that $\mathbb{Z}$-filtrations on the $L_\lambda$ provided by a Gelfand--Tsetlin degeneration are all induced by a certain $\mathbb{Z}$-filtration on $\mathcal U(\mathfrak{n}_-)$. The latter filtration has a simple description in terms of a particular PBW basis. However, we are still lacking a degenerate representation theory, since the filtration on $\mathcal U(\mathfrak{n}_-)$ does not respect multiplication and we do not have an associated graded algebra acting on the $L_\lambda^S$. This issue is addressed in Section~\ref{degenact} which contains our main results. There we define the algebra $\Phi_n$, its action and the degenerations of this action as well as tensor products of degenerate representations. Then we formulate and prove the full versions of Theorems~\ref{thma} and~\ref{thmb}. The next three sections contain several extensions and consequences of our results. In Section~\ref{singular} we generalize results from Sections~\ref{gtdegens} and~\ref{degenact} to partial flag varieties. In Section~\ref{tropical} we show that the cone parametrizing our degenerations can be viewed as a cone in the Gr\"obner fan that is also a maximal cone in the tropical flag variety. We thus obtain the minimal H-description of the cone corresponding to the Gelfand--Tsetlin toric degeneration (this cone is discussed in~\cite[Example 7.3]{KaM} where a redundant H-description is given). This also characterizes the other degenerations in our family (corresponding to proper faces of the cone) as those interpolating between the flag variety and the toric variety. In Section~\ref{dual} we explain how the whole construction can be dualized via the Dynkin diagram automorphism. Finally, Section~\ref{addendum} contains later findings that were inspired by the attempts to iron out some kinks in the above theory. They constitute an alternative approach to the original problem of providing a degenerate representation-theoretic context for the Gelfand--Tsetlin toric variety. While the approach developed in the above sections is comparable to the theory in~\cite{fafefom}, the approach in Section~\ref{addendum} bears certain parallels to~\cite{favourable}. In the latter paper a degenerate representation theory is constructed by choosing a PBW basis in $\mathcal U(\mathfrak{n}_-)$ (in natural bijection with $\mathbb{Z}_{\ge0}^{\{1\le i<j\le n\}}$) together with a total order on the monoid $\mathbb{Z}_{\ge0}^{\{1\le i<j\le n\}}$. This provides filtrations by the totally ordered monoid on $\mathcal U(\mathfrak{n}_-)$ and all $L_\lambda$. The associated graded spaces are $\mathbb{Z}_{\ge0}^{\{1\le i<j\le n\}}$-graded, i.e. acted upon by a $n\choose 2$-dimensional torus. Certain additional restrictions are imposed on the total order so that the filtrations respect multiplication and the associated graded algebra is a universal enveloping algebra (which is also required to be commutative in the paper). The orbit closure then turns out to be a toric variety and a specific order is given for which the FFLV toric variety is obtained. Our idea is, in a sense, to do this without fixing a PBW basis, i.e. to choose a total order on the non-abelian free monoid with $n\choose 2$ generators which is in natural bijection with the set of all PBW monomials in $\mathcal U(\mathfrak{n}_-)$. We obtain a degenerate representation theory with the associated graded algebra and the degenerate representations graded by the free monoid. We then define a specific total order which produces the Gelfand--Tsetlin toric variety via analogs of Theorems~\ref{thma} and~\ref{thmb}. The main advantage of this approach is that the degenerate algebra action appears directly as a degeneration of the $\mathcal U(\mathfrak{n}_-)$-action without a need to first define an action ad hoc (as was the case with $\Phi_n$). A disadvantage, however, is that the intermediate Gr\"obner degenerations are not recovered, nor is the connection with tropical geometry. \section{Generalities on Gr\"obner degenerations of flag varieties}\label{generalities} For a fixed $n\ge 2$ consider the Lie group $G=SL_n(\mathbb{C})$ with Borel subgroup $B$ and tangent algebra $\mathfrak{g}=\mathfrak{sl}_n(\mathbb{C})$. Let $\mathfrak b\subset\mathfrak{g}$ be the Borel subalgebra tangent to $B$, let $\mathfrak{h}\subset\mathfrak b$ be the Cartan subalgebra and let $\mathfrak{g}=\mathfrak b\oplus\mathfrak{n}_-$ for nilpotent subalgebra $\mathfrak{n}_-$. For $1\le k\le n-1$ denote the simple roots $\alpha_k\in\mathfrak{h}^*$ and let $\omega_k\in\mathfrak{h}^*$ be the corresponding fundamental weights. Denote the positive roots \[\alpha_{i,j}=\alpha_i+\ldots+\alpha_{j-1}\] for $1\le i<j\le n-1$. Let $\mathfrak{n}_-$ be spanned by negative root vectors $f_{i,j}$ with weight $-\alpha_{i,j}$. Our choice of basis in $\mathfrak{h}^*$ will be the set of fundamental weights, i.e.\ $(a_1,\ldots,a_{n-1})$ will denote the weight $a_1\omega_1+\ldots+a_{n-1}\omega_{n-1}$. For a dominant integral $\mathfrak{g}$-weight $\lambda$ let $L_\lambda$ be the irreducible representation of $\mathfrak{g}$ with highest weight $\lambda$ and highest weight vector $v_\lambda$. Let the $n$-dimensional complex space $V$ be the tautological representation of $\mathfrak{g}=\mathfrak{sl}_n$ with basis $e_1,\ldots,e_n$. The irreducible representations with fundamental highest weights can be explicitly described as $L_{\omega_k}=\wedge^k V$ with a basis consisting of the vectors \[e_{i_1,\ldots,i_k}=e_{i_1}\wedge\ldots\wedge e_{i_k}.\] We may assume that $v_{\omega_k}=e_{1,\ldots,k}$. Consider the variety of complete flags $F=G/B$ and the Pl\"ucker embedding \[F\subset\mathbb{P}=\mathbb{P}(L_{\omega_1})\times\ldots\times\mathbb{P}(L_{\omega_{n-1}}).\] The product $\mathbb{P}$ is equipped with the Pl\"ucker coordinates $X_{i_1,\ldots,i_k}$ with $1\le k\le n-1, 1\le i_1<\ldots<i_k\le n$, coordinate $X_{i_1,\ldots,i_k}$ corresponding to $e_{i_1,\ldots,i_k}\in L_{\omega_k}$. The homogeneous coordinate ring of $\mathbb{P}$ is $R=\mathbb{C}[\{X_{i_1,\ldots,i_k}\}]$. The homogeneous coordinate ring of $F$ is then $\mathcal P=R/I$ (known as the Pl\"ucker algebra), where $I$ is the ideal of Pl\"ucker relations. Note that $R$ is naturally graded by the semigroup of dominant integral weights with the homogeneous component $R_\lambda$ corresponding to weight $\lambda=(a_1,\ldots,a_{n-1})$ spanned by monomials with total degree in variables of the form $X_{i_1,\ldots,i_k}$ equal to $a_k$. We will denote this grading $\wt$. Since the ideal $I$ is $\wt$-homogeneous, so is $\mathcal P$. In the latter, the homogeneous component $\mathcal P_\lambda$ of degree $\lambda$ is identified with the dual representation $L_\lambda^*$. These classical definitions and results concerning $\mathfrak{sl}_n$-representations and flag varieties can be found in~\cite{carter} and~\cite[Chapter 9]{fulton}. Now consider a collection of integers $S=(s_{i_1,\ldots,i_k})$, one for each Pl\"ucker variable. This provides a $\mathbb{Z}$-grading on $R$ by setting $\grad^S X_{i_1,\ldots,i_k}=s_{i_1,\ldots,i_k}$. Consider the initial ideal $\initial_{\grad^S}I$ (spanned by nonzero components of minimal grading of elements of $I$). We will be considering the subvariety $F^S$ in $\mathbb{P}$ defined by this ideal. Varieties of the form $F^S$ are known as {\it Gr\"obner degenerations} of $F$. We have a decreasing $\mathbb{Z}$-filtration on $R$ with the $m$th filtration component $R_{\ge m}$ being spanned by monomials in $R$ of $\grad^S$ no less than $m$. This induces a decreasing $\mathbb{Z}$-filtration on $\mathcal P$ with components $\mathcal P_m$, note that this is a filtered algebra structure. We denote the associated $\mathbb{Z}$-graded algebra $\mathcal P^S=\bigoplus_m \mathcal P_{m-1}/\mathcal P_m$. \begin{proposition}\label{initdegen} $\mathcal P^S$ and $R/\initial_{\grad^S}I$ are isomorphic as $\mathbb{Z}$-graded algebras. \end{proposition} \begin{proof} The associated graded algebra of $R$ with respect to the filtration $(\cdot)_{\ge m}$ is again $R$ with the same grading $\grad^S$. This associated graded algebra, however, projects naturally onto $\mathcal P^S$. We obtain a surjection of $\mathbb{Z}$-graded algebras from $R$ onto $\mathcal P^S$ and are left to show that the kernel of this surjection is $\initial_{\grad^S}I$. A $\grad^S$-homogeneous element $p\in R$ lies in this kernel if and only if there exists some $q\in I$ such that $p+q\in R_{\ge \grad^S(p)+1}$ which simply means that $p$ is the initial part of $-q$. \end{proof} In particular, we obtain isomorphisms between the $\wt$-homogeneous components, i.e.\ we have identified every $\wt$-homogeneous component $\mathcal P^S_\lambda$ of a Gr\"obner degeneration of the Pl\"ucker algebra with a certain associated graded space of the dual irreducible representation. \begin{remark} Isomorphisms between quotients by initial ideals and associated graded rings are a very general phenomenon with Proposition~\ref{initdegen} being a special case. This phenomenon will reappear in this paper as Proposition~\ref{adinitdegen}. However, somewhat surprisingly, the author was not able to find this proved in the literature as a general fact that would imply either of the two propositions, see also MathOverflow question~\cite{MO}. \end{remark} Now, for an integral dominant weight $\lambda=(a_1,\ldots,a_{n-1})$ consider the tensor product \[U_\lambda=L_{\omega_1}^{\otimes a_1}\otimes\ldots\otimes L_{\omega_{n-1}}^{\otimes a_{n-1}}.\] The subrepresentation of $U_\lambda$ generated by the highest weight vector \[u_\lambda=v_{\omega_1}^{\otimes a_1}\otimes\ldots\otimes v_{\omega_{n-1}}^{\otimes a_{n-1}}\] is the irreducible representation $L_\lambda$ (naturally dual to $\mathcal P_\lambda$). If we grade $L_{\omega_k}$ by setting $\grad^S e_{i_1,\ldots,i_k}=s_{i_1,\ldots,i_k}$, a grading (which we also denote $\grad^S$) on $U_\lambda$ is induced. We may consider an increasing $\mathbb{Z}$-filtration on $U_\lambda$ with the $m$th component $(U_\lambda)_{\le m}$ being spanned by $\grad^S$-homogeneous elements of $\grad^S$ no greater than $m$. This induces a $\mathbb{Z}$-filtration on $L_\lambda\subset U_\lambda$ with components $(L_\lambda)_m$, denote the associated graded space $L_\lambda^S$ with homogeneous components $(L_\lambda^S)_m$. \begin{proposition}\label{dualspaces} $L_\lambda^S$ and $\mathcal P_\lambda^S$ are dual as $\mathbb{Z}$-graded vector spaces. \end{proposition} \begin{proof} Let us consider the subrepresentation \[W_\lambda=\Sym^{a_1}(L_{\omega_1})\otimes\ldots\otimes\Sym^{a_{n-1}}(L_{\omega_{n-1}})\subset U_\lambda,\] note that $W_\lambda$ is a graded subspace and $L_\lambda\subset W_\lambda$. The space $W_\lambda$ is dual to $R_\lambda$ where the symmetrization of a tensor product of some vectors $e_{i_1,\ldots,i_k}$ is the basis element dual to the product of the corresponding variables $X_{i_1,\ldots,i_k}$. The increasing $\mathbb{Z}$-filtration on $W_\lambda$ (given by $(W_\lambda)_{\le m}=(U_\lambda)_{\le m}\cap W_\lambda$) is dual to the decreasing $\mathbb{Z}$-filtration on $R_\lambda$ in the sense that the subspace $(W_\lambda)_{\le m}$ is dual to the subspace $(R_\lambda)_{\ge m+1}$. This provides a duality between their associated graded spaces which are again $W_\lambda$ and $R_\lambda$ with the same gradings. The space $L_\lambda^S$ is, by definition, embedded into the associated graded space $W_\lambda$ and the space $\mathcal P_\lambda^S$ is (as noted in the proof of Proposition~\ref{initdegen}) a projection of the associated graded space $R_\lambda$. We are to show that the kernel $\initial_{\grad^S} I_\lambda$ of the latter projection is the orthogonal of the subspace $L_\lambda^S\subset W_\lambda$. Now, it is known (see, for instance, Theorem 8.2.2 and Proposition 9.1.1 in~\cite{fulton}) that $L_\lambda\subset W_\lambda$ is the orthogonal of the kernel $I_\lambda$ of the projection of $R_\lambda$ onto $\mathcal P_\lambda$. However, if we consider an element of $v\in L_\lambda\subset W_\lambda$ and take its projection onto the $\grad^S$-homogeneous component of maximal grading for which the projection is nonzero we will obtain an element of $L_\lambda^S\subset W_\lambda$ and $L_\lambda^S$ is spanned by elements of this form. By definition $\initial_{\grad^S} I_\lambda$ is spanned by the $\grad^S$-initial parts of elements of $I_\lambda$. One sees that such an initial part annihilates the mentioned projection in $L_\lambda^S$ and the orthogonality follows. \end{proof} As discussed above, we have an embedding $L_\lambda^S\subset W_\lambda\subset U_\lambda$. We also have the Segre embedding $$\mathbb P(L_{\omega_1})^{a_1}\times\ldots\times\mathbb P(L_{\omega_{n-1}})^{a_{n-1}}\subset\mathbb P(U_\lambda)$$ and, for regular $\lambda$ (i.e.\ all $a_k>0$), the embedding $$\mathbb P\subset \mathbb P(L_{\omega_1})^{a_1}\times\ldots\times\mathbb P(L_{\omega_{n-1}})^{a_{n-1}}$$ where $\mathbb P(L_{\omega_k})$ is embedded diagonally into $\mathbb P(L_{\omega_k})^{a_k}$. We obtain an embedding \[F^S\subset\mathbb{P}\subset\mathbb{P}(U_\lambda).\] \begin{proposition}\label{FSinPLS} For a regular $\lambda$ the image of $F^S$ under this embedding is contained in $\mathbb{P}(L_\lambda^S)\subset\mathbb{P}(U_\lambda)$. \end{proposition} \begin{proof} The image of the Segre embedding and, therefore, of $F^S$ lies in $\mathbb{P}(W_\lambda)$. In view of Proposition~\ref{dualspaces} and its proof, to show that a point in $F^S$ is contained in $\mathbb{P}(L_\lambda^S)$ we are to show that the corresponding line in $W_\lambda$ is annihilated by every element of $\initial_{\grad^S} I_\lambda\subset R_\lambda$ (where these elements are viewed as functionals on $W_\lambda$). This, however, is straightforward from the definitions. \end{proof} The above proposition provides an embedding of $F^S$ into $\mathbb{P}(L_\lambda^S)$. When the degeneration is trivial, i.e.\ all $s_{i_1,\ldots,i_k}=0$ and $F^S=F$, we obtain the usual embedding of $F$ into $\mathbb{P}(L_\lambda)$ as the closure of the orbit of the point corresponding to $\mathbb{C} v_\lambda$ under the action of the group $\exp(\mathfrak{n}_-)$. \begin{remark}\label{projlim} For integral dominant weights $\lambda$ and $\lambda'$ such that $\lambda-\lambda'$ is also dominant we have a surjection $L_\lambda\mapsto L_{\lambda'}$ of $\mathcal U(\mathfrak{n}_-)$-modules. This is a projective system with projective limit $\mathcal U(\mathfrak{n}_-)$. Now, setting $s_{i_1,\ldots,i_{k}}:=s_{i_1,\ldots,i_{k}}-s_{1,\ldots,k}$ for all tuples $i_1,\ldots,i_{k}$ does not change the degeneration $F^S$, therefore we may assume that all $s_{1,\ldots,k}=0$. With this assumption in place one can see that the said projective system respects the $\mathbb{Z}$-filtrations on the $L_\lambda$ in the sense that it induces an increasing $\mathbb{Z}$-filtration on the limit $\mathcal U(\mathfrak{n}_-)$. We will recover an explicit description of this filtration on $\mathcal U(\mathfrak{n}_-)$ for the case of Gelfand--Tsetlin degenerations in Theorem~\ref{Lfiltration}. However, it would be nice to have a more direct definition of this filtration in the general case. \end{remark} \begin{remark} All the results found here as well as in the sections below can be formulated for real (rather than integer) gradings. The reason for us to assume that $s_{i_1,\ldots,i_k}\in\mathbb{Z}$ is that the real case would require us to consider spaces graded by the semigroup generated by the $s_{i_1,\ldots,i_k}$ rather than by $\mathbb{Z}$. This would complicate the notations with virtually no gain in mathematical merit. (The only disadvantage integer gradings give us is not being able to work directly with cones in the Gr\"obner fan in the proof of Theorem~\ref{maxcone}, only with their sets of integer points. This, however, is easily circumvented.) \end{remark} \begin{remark} Proposition~\ref{FSinPLS} is stated only for regular $\lambda$ since a version of this theorem for singular $\lambda$ would concern degenerations of partial flag varieties rather than $F$. Sections~\ref{gtdegens} and~\ref{degenact} will contain more results concerned only with complete flag varieties and/or regular highest weights. This is, again, done to avoid overcomplicating the notations in these key sections. However, partial flag varieties and singular highest weights will be discussed in Section~\ref{singular} and the corresponding generalizations of the results will be given there. \end{remark} \section{Polytopes and monomial bases}\label{bases} Consider $\Theta=\mathbb{R}^{\{1\le i<j\le n\}}$, for a point $T\in\Theta$ denote its coordinates $T_{i,j}$. For each tuple $1\le i_1<\ldots<i_k\le n$ with $1\le k\le n-1$ we define a vector $T(i_1,\ldots,i_k)\in\Theta$ coordinatewise by setting \begin{equation} T(i_1,\ldots,i_k)_{l,m}= \begin{cases} 1\text{ if }l\le k\text{ and }m=i_l,\\ 0\text{ otherwise}. \end{cases} \end{equation} In other words, we have the coordinate corresponding to pair $(l,i_l)$ equal to $1$ for every $1\le l\le k$ with $i_l>l$ and all other coordinates zero. For a $1\le k\le n-1$ let $\Pi_{\omega_k}\subset\Theta$ be the set of all $T(i_1,\ldots,i_k)$. Next, for an integral dominant weight $\lambda=(a_1,\ldots,a_{n-1})$ consider the Minkowski sum \begin{equation}\label{minksum} \Pi_\lambda=\underbrace{\Pi_{\omega_1}+\ldots+\Pi_{\omega_1}}_{a_1}+\ldots+\underbrace{\Pi_{\omega_{n-1}}+\ldots+\Pi_{\omega_{n-1}}}_{a_{n-1}}. \end{equation} We also introduce a convex lattice polytope $P_\lambda\subset\Theta$ consisting of points $T$ such that \begin{enumerate}[label=(\roman*)] \item $T_{i,j}\ge0$ for all $1\le i<j\le n$; \item $\sum_{l=j}^n T_{i,l}-\sum_{l=j+1}^n T_{i+1,l}\le a_i$ for all $1\le i<j\le n$. \end{enumerate} The second sum in (ii) is empty if $j=n$. Before we proceed, let us recall the definition of Gelfand--Tsetlin polytopes introduced in~\cite{GT}. For each integral dominant weight $\lambda=(a_1,\ldots,a_{n-1})$ the corresponding GT polytope $GT_\lambda$ is a convex lattice polytope in $\Theta$ composed of points $T$ such that \begin{enumerate}[label=(\roman*)] \setcounter{enumi}{2} \item $\lambda_i\ge T_{i,i+1}\ge\lambda_{i+1}$ for all $1\le i\le n-1$ where $\lambda_i=a_i+\ldots+a_{n-1}$ and $\lambda_n=0$; \item $T_{i,j-1}\ge T_{i,j}\ge T_{i+1,j}$ for all pairs $1\le i<j\le n$ with $j>i+1$. \end{enumerate} Let $\Gamma_\lambda$ denote the set of integer points in $GT_\lambda$. A key property of GT polytopes established in~\cite{GT} is that $\Gamma_\lambda$ enumerates a basis in $L_\lambda$, hence $|\Gamma_\lambda|=\dim L_\lambda$. We will also make use of the following {\it Minkowski sum property}. \begin{proposition}\label{gtminkowski} For integral dominant weights $\lambda$ and $\mu$ the Minkowski sum $\Gamma_\lambda+\Gamma_\mu$ coincides with $\Gamma_{\lambda+\mu}$. \end{proposition} \begin{proof} A proof can, for instance, be found in~\cite[Theorem 2.5]{chainorder} in the much more general context of marked chain-order polytopes of which the GT polytopes are a special case (as discussed in the Introduction of~\cite{chainorder}). \end{proof} We return to the polytopes $P_\lambda$ to prove the following. \begin{lemma}\label{gtequiv} $\Pi_\lambda$ is the set of integer points in $P_\lambda$. Furthermore, $P_\lambda$ is unimodularly equivalent to the Gelfand--Tsetlin polytope $GT_\lambda$. \end{lemma} \begin{example}\label{sl3polytopes} Before we proceed with the proof let us illustrate the definitions and the Lemma with an example. Let $n=3$, we visualize $T\in\Theta$ as ${T_{1,2}\,T_{2,3}}\atop{T_{1,3}}$. We have \[\Pi_{\omega_1}=\left\{T(1)={{0\,0}\atop0},T(2)={{1\,0}\atop0},T(3)={{0\,0}\atop1}\right\}\] and \[\Pi_{\omega_2}=\left\{T(1,2)={{0\,0}\atop0},T(1,3)={{0\,1}\atop0},T(2,3)={{1\,1}\atop0}\right\}.\] For $\lambda=\omega_1+\omega_2$ we obtain \[\Pi_\lambda=\left\{{{0\,0}\atop0},{{0\,1}\atop0},{{1\,1}\atop0},{{1\,0}\atop0},{{2\,1}\atop0},{{0\,0}\atop1},{{0\,1}\atop1},{{1\,1}\atop1}\right\}.\] We see that $\Pi_{\omega_1}$ is the set of integer points in the polytope $P_{\omega_1}$ defined by the inequalities $T_{i,j}\ge 0$, $T_{1,2}+T_{1,3}-T_{2,3}\le 1$, $T_{1,3}\le 1$ and $T_{2,3}\le 0$. Polytope $P_{\omega_2}$ is given by $T_{i,j}\ge 0$, $T_{1,2}+T_{1,3}-T_{2,3}\le 0$, $T_{1,3}\le 0$ and $T_{2,3}\le 1$ and $P_{\lambda}$ is given by $T_{i,j}\ge 0$, $T_{1,2}+T_{1,3}-T_{2,3}\le 1$, $T_{1,3}\le 1$ and $T_{2,3}\le 1$. Now, we have $(\lambda_1,\lambda_2,\lambda_3)=(2,1,0)$ and the set $\Gamma_\lambda$ is seen to be \[\left\{{{2\,1}\atop2},{{2\,0}\atop2},{{2\,0}\atop1},{{2\,1}\atop1},{{2\,0}\atop0},{{1\,1}\atop1},{{1\,0}\atop1},{{1\,0}\atop0}\right\}.\] Consider the affine transformation $\psi$ of $\Theta$ given by $\psi:T\mapsto{{2-T_{1,3}\:1-T_{2,3}}\atop{2-T_{1,2}-T_{1,3}}}$. Observe that $\psi(\Pi_\lambda)=\Gamma_\lambda$. Moreover, if one takes the inequalities defining $GT_\lambda$ (i.e.\ $2\ge T_{1,2}\ge 1\ge T_{2,3}\ge 0$ and $T_{1,2}\ge T_{1,3}\ge T_{2,3}$) and substitutes every occurrence of $T_{i,j}$ with the expression $\psi(T)_{i,j}$ defined above, one will end up with the 6 inequalities defining $P_\lambda$. For instance, $2\ge T_{1,2}$ turns into $2\ge 2-T_{1,3}\Leftrightarrow T_{1,3}\ge 0$ or $T_{1,3}\ge T_{2,3}$ turns into $2-T_{1,2}-T_{1,3}\ge 1-T_{2,3}\Leftrightarrow T_{1,2}+T_{1,3}-T_{2,3}\le 1$. To prove the Lemma we generalize this map $\psi$. \end{example} \begin{proof}[Proof of Lemma~\ref{gtequiv}] Consider the affine transformation $\psi$ of $\Theta$ given by \begin{equation}\label{psidef} \psi(T)_{i,j}=\lambda_i-\sum_{l=i+n+1-j}^n T_{i,l}. \end{equation} It is evident that $\psi$ is unimodular and preserves the lattice of integer points, let us show that $\psi(P_\lambda)=GT_\lambda$. Indeed, if $j<n$, then the inequality in (i) is equivalent to \begin{equation}\label{psi1} \psi(T)_{i,i+n-j}\ge \psi(T)_{i,i+n+1-j} \end{equation} and the inequality in (ii) is equivalent \begin{equation}\label{psi2} \psi(T)_{i,i+n+1-j}\ge \psi(T)_{i+1,i+n+1-j}. \end{equation} (\ref{psi1}) and (\ref{psi2}) combined over all $1\le i<j<n$ give (iv). If $j=n$, then the inequality in (i) is equivalent to \begin{equation}\label{psi3} \psi(T)_{i,i+1}\le\lambda_i \end{equation} and the inequality in (ii) is equivalent to \begin{equation}\label{psi4} \psi(T)_{i,i+1}\ge\lambda_{i+1}. \end{equation} Combining~(\ref{psi3}) and~(\ref{psi4}) over all $1\le i\le n-1$ gives (iii). This proves the second part of the lemma. Now, with the second part established, the first part follows from the definition of $\Pi_\lambda$, Proposition~\ref{gtminkowski} and the claim that $\Pi_{\omega_k}$ is the set of integer points in $P_{\omega_k}$ for all $1\le k\le n-1$. To verify this last claim note that $\Pi_{\omega_k}\subset P_{\omega_k}$ is immediate from the definitions and that, in view of the second part, $P_{\omega_k}$ has exactly $n\choose k$ integer points and therefore has no integer points outside of $\Pi_{\omega_k}$. \end{proof} \begin{cor}\label{dimension} $|\Pi_\lambda|=\dim L_\lambda$. \end{cor} \begin{remark} Lemma~\ref{gtequiv} is immediate from the results found in~\cite[Section 14.4]{MS}, in particular, a variation of our map $\psi$ is also constructed there. We give a self-contained proof for the sake of completeness. The formula~(\ref{psidef}) for $\psi$ is easily derived via additivity with respect to $\lambda$, a much more involved question is \emph{why} a polytope given by the fairly simple expression~(\ref{minksum}) would turn out to be equivalent to the GT polytope. One possible explanation can be found in~\cite[Section 5]{KM}. \end{remark} Let us now define the aforementioned monomial bases. We make use of the following terminology. Call a monomial $M\in\mathcal U(\mathfrak{n}_{-})$ {\it ordered} if the factors $f_{i,j}$ appearing in M are ordered by $i$ increasing from left to right. For every $T\in\Theta$ with nonnegative integer coordinates we consider the ordered monomial $M_T\in \mathcal U(\mathfrak{n}_-)$ that contains $f_{i,j}$ in degree $T_{i,j}$. Note that $M_T$ is defined uniquely since any two elements of the form $f_{i,j_1}$ and $f_{i,j_2}$ commute. \begin{theorem}[{\cite[Introduction]{R}, also proved in the next section}]\label{basis} The set $\{M_T v_\lambda, T\in\Pi_\lambda\}$ is a basis in $L_\lambda$. \end{theorem} This theorem will be proved in Section~\ref{gtdegens} after we introduce the relevant degree functions on monomials in $\mathcal U(\mathfrak{n}_-)$. \section{Gelfand--Tsetlin degenerations}\label{gtdegens} In this section we will define a specific family of Gr\"obner degenerations of $F$ and list several properties of the corresponding objects introduced in Section~\ref{generalities} (as well as proving Theorem~\ref{basis}). First, consider a collection of integers $A=(a_{i,j}|1\le i<j\le n)$ such that \begin{enumerate}[label=(\alph*)] \item $a_{i,i+1}+a_{i+1,i+2}\le a_{i,i+2}$ for any $1\le i\le n-2$ and \item $a_{i,j}+a_{i+1,j+1}\le a_{i,j+1}+a_{i+1,j}$ for any $1\le i<j-1\le n-2$ \end{enumerate} or, equivalently, \begin{enumerate}[label=(\Alph*)] \item $a_{i,j}+a_{j,k}\le a_{i,k}$ for any $1\le i<j<k\le n$ and \item $a_{i,j}+a_{k,l}\le a_{i,l}+a_{k,j}$ for any $1\le i<k<j<l\le n$. \end{enumerate} The proof that the inequalities in (A) and (B) can be deduced from those in (a) and (b) is straightforward and almost identical (up to reversing all inequalities) to the proof of Proposition~2.1 in~\cite{fafefom}. We will view $A$ as an element of $\Theta^*$ equipped with the basis dual to the one chosen in $\Theta$. We define \[\sigma(A)=(\sigma(A)_{i_1,\ldots,i_k})\in \mathbb{R}^{\{1\le i_1<\ldots<i_k\le n|1\le k\le n-1\}}\] with $\sigma(A)_{i_1,\ldots,i_k}=A(T(i_1,\ldots,i_k))$ (as a functional on $\Theta$). We will refer to Gr\"obner degenerations given by $\sigma(A)$ with $A$ satisfying (a) and (b) as Gelfand--Tsetlin (or GT) degenerations of $F$ and to the corresponding associated graded spaces $L_\lambda^{\sigma(A)}$ as GT degenerations of $L_\lambda$. From now on and through Section~\ref{singular} we fix $A$ and $S=(s_{i_1,\ldots,i_k})=\sigma(A)$. \begin{example}\label{sl3ideal} For $n=3$ one may set $A={{a_{1,2}\,a_{2,3}}\atop{a_{1,3}}}={{-1\,-1}\atop{-1}}$. The only inequality here is $a_{1,2}+a_{2,3}\le a_{1,3}$ and it, evidently, holds. All points $T(i_1,\ldots,i_k)$ are listed in Example~\ref{sl3polytopes} and one can compute $\grad^S X_1=\grad^S X_{1,2}=0$, $\grad^S X_2=\grad^S X_3=\grad^S X_{1,3}=-1$ and $\grad^S X_{2,3}=-2$. The ideal $I$ is generated by the element $X_1X_{2,3}-X_2X_{1,3}+X_3X_{1,2}$ and the initial part of this element with respect to $\grad^S$ is $X_1X_{2,3}-X_2X_{1,3}$ which is the sole generator of $\initial_{\grad^S}I$. \end{example} For a monomial $M=f_{i_1,j_1}\ldots f_{i_N,j_N}\in\mathcal U(\mathfrak{n}_{-})$ denote \[\deg^A M=a_{i_1,j_1}+\ldots+a_{i_N,j_N}.\] Note that when $n=3$ there are two ordered monomials in $\mathcal U(\mathfrak{n}_-)$ which map $v_{\omega_2}$ to a nonzero multiple of $e_{2,3}$, these are $M_{T(2,3)}$ and (in the notations of Example~\ref{sl3polytopes}) $M_{{0\,0}\atop{1}}$. We have \[\deg^A M_{T(2,3)}=a_{1,2}+a_{2,3}\le a_{1,3}=\deg^A M_{{0\,0}\atop{1}}.\] This inequality between the degrees of the monomials is a special case of the following key lemma which shows why we want $A$ to satisfy (A) and (B). \begin{lemma}\label{minmon} For a tuple $1\le i_1<\ldots<i_k\le n$ and an ordered monomial $M$ such that $M v_{\omega_k}\in\mathbb{C}^*e_{i_1,\ldots,i_k}$ we have $\deg^A M\ge s_{i_1,\ldots,i_k}$. \end{lemma} \begin{proof} First of all, note that $f_{i,j}$ maps $e_{i_1,\ldots,i_k}$ to $\pm e_{j_1,\ldots,j_k}$ where \[\{j_1,\ldots,j_k\}=\{i_1,\ldots,i_k\}\cup\{j\}\backslash\{i\}\] if $i\in\{i_1,\ldots,i_k\}$ and $j\notin\{i_1,\ldots,i_k\}$, otherwise $f_{i,j}$ maps $e_{i_1,\ldots,i_k}$ to 0. In particular, that means that for any monomial $M$ in the $f_{i,j}$ the vector $Mv_{\omega_k}$ is either zero or of the form $\pm e_{j_1,\ldots,j_k}$. Recall the ordered monomials $M_T$ defined in Section~\ref{bases}. Note that $M_{T(i_1,\ldots,i_k)} v_{\omega_k}=\pm e_{i_1,\ldots,i_k}$ and that $\deg^A M_{T(i_1,\ldots,i_k)}=s_{i_1,\ldots,i_k}$. Now, consider an ordered monomial \[M=f_{l_1,m_1}\ldots f_{l_N,m_N}\] such that $Mv_{\omega_k}=\pm e_{i_1,\ldots,i_k}$ with minimal $\deg^A M$ and out of these with the minimal possible sum $\sum_i (m_i-l_i)^2$. We prove the lemma by showing that $M=M_{T_{i_1,\ldots,i_k}}$. Any product of the form $f_{i,j_1}f_{i,j_2}$ annihilates $L_{\omega_k}$, therefore all $l_i$ are pairwise distinct. Suppose that for some $i$ we have $m_i\ge m_{i+1}$. The product $f_{l_i,m_i}f_{l_{i+1},m_i}$ annihilates $L_{\omega_k}$, hence $m_i>m_{i+1}$. For any $e_{j_1,\ldots,j_k}$ we have \[f_{l_i,m_i}f_{l_{i+1},m_{i+1}}(e_{j_1,\ldots,j_k})=\pm f_{l_i,m_{i+1}}f_{l_{i+1},m_i}(e_{j_1,\ldots,j_k}).\] Therefore, by replacing $f_{l_i,m_i}f_{l_{i+1},m_{i+1}}$ in $M$ with $f_{l_i,m_{i+1}}f_{l_{i+1},m_i}$ we would obtain a monomial also mapping $v_{\omega_k}$ to $\pm e_{i_1,\ldots,i_k}$ and of no greater $\deg^A$-degree due to (B). However, $$(m_i-l_{i+1})^2+(m_{i+1}-l_i)^2<(m_i-l_i)^2+(m_{i+1}-l_{i+1})^2$$ which would contradict our choice of $M$. We see that both sequences $(l_1,\ldots,l_N)$ and $(m_1,\ldots,m_N)$ are strictly increasing. Now define a tuple $(j_1,\ldots,j_k)$ by setting $j_l=m_i$ if $l=l_i$ for some $i$ and $j_l=l$ otherwise. $Mv_{\omega_k}=\pm e_{i_1,\ldots,i_k}$ implies that $(j_1,\ldots,j_k)$ is a permutation of $(i_1,\ldots,i_k)$ (note that all $l_i\le k$ since $M v_{\omega_k}\neq 0$), we are to prove that $(j_1,\ldots,j_k)=(i_1,\ldots,i_k)$. Suppose the contrary, i.e.\ that $j_l>j_{l+1}$ for some $l$. The sequences $(l_1,\ldots,l_N)$ and $(m_1,\ldots,m_N)$ increasing implies that $j_{l+1}=l+1$ while $j_l=m_i$ for some $i$. In particular, $l+1\notin\{l_1,\ldots,l_N\}$ and replacing $f_{l,m_i}$ in $M$ with the product $f_{l,l+1}f_{l+1,m_i}$ we would obtain a monomial $M'$ with $M'v_{\omega_k}=\pm Mv_{\omega_k}$ and $\deg^AM'\le\deg^A M$ due to (A). However, $1+(m_i-l-1)^2<(m_i-l)^2$ which again achieves a contradiction. \end{proof} The above proof has the following two implications. \begin{proposition}\label{minmonstrict} Suppose that all the inequalities in (A) and (B) are strict. If $T\in\mathbb{Z}_{\ge0}^{\{1\le i<j\le n\}}\subset\Theta$ is such that $M_Tv_{\omega_k}=\pm e_{i_1,\ldots,i_k}$, then either $T=T(i_1,\ldots,i_k)$ or $\deg^A M_T>s_{i_1,\ldots,i_k}$. \end{proposition} For $T\in\Theta$ denote $\sq(T)=\sum_{i,j} T_{i,j}(j-i)^2$. \begin{proposition}\label{sumsquares} If $T\in\mathbb{Z}_{\ge0}^{\{1\le i<j\le n\}}\subset\Theta$ is such that $M_Tv_{\omega_k}=\pm e_{i_1,\ldots,i_k}$ and $\deg^A M_T=s_{i_1,\ldots,i_k}$, then either $T=T(i_1,\ldots,i_k)$ or $\sq(T)>\sq(T(i_1,\ldots,i_k))$. \end{proposition} We are now ready to prove Theorem~\ref{basis}. \begin{proof}[Proof of Theorem~\ref{basis}] In view of Corollary~\ref{dimension} it suffices to show that the set $\{M_T v_\lambda,T\in\Pi_\lambda\}$ is linearly independent. We make use of $L_\lambda$ being embedded into $U_\lambda$ as the subrepresentation generated by $u_\lambda$. For $T\in\Pi_\lambda$ let $U_T\subset U_\lambda$ be the subspace spanned by products of the form \[v^1_1\otimes\ldots\otimes v^1_{a_1}\otimes\ldots\otimes v^{n-1}_1\otimes\ldots\otimes v^{n-1}_{a_{n-1}}\] with $v^i_j=e_{l^{i,j}_1,\ldots,l^{i,j}_i}$ for which the total of all $T(l^{i,j}_1,\ldots,l^{i,j}_i)$ is equal to $T$. Then $U_\lambda$ is the direct sum of $U_T$ with $T$ ranging over $\Pi_\lambda$ and we see that every $U_T$ is $\grad^S$-homogeneous with $U_T\subset (U_\lambda)_{A(T)}$. Now, choose $T\in\Pi_\lambda$ and decompose $M_T u_\lambda$ into a sum of tensor products. Every summand is obtained by partitioning the set of factors in $M_T$ into $a_1+\ldots+a_{n-1}$ subsets, one for every tensor factor, applying the ordered product of each subset to the corresponding $v_{\omega_k}$ and then taking the tensor product of the results. We now see, by applying Lemma~\ref{minmon} and Proposition~\ref{sumsquares}, that one of the following holds. Every summand in this decomposition lies in $U_T$, lies in some $U_{T'}$ with $A(T')=A(T)=\deg^A M_T$ and $\sq(T')<\sq(T)$ or lies in some $U(T')$ with $A(T')<A(T)$. In view of the Minkowski sum property, at least one summand lies in $U_T$ and the linear independence follows via a triangularity argument. \end{proof} For $T\in\mathbb{Z}_{\ge0}^{\{1\le i<j\le n\}}$ say that the monomial $M_T$ is {\it $L_\lambda$-optimal} if $M_T v_\lambda$ lies in $(L_\lambda)_{A(T)}$ but not in $(L_\lambda)_{A(T)-1}$. The above proof implies the following fact which we will make use of later. \begin{proposition}\label{basiscor} For every $T\in\Pi_\lambda$ the monomial $M_T$ is $L_\lambda$-optimal. \end{proposition} We now explicitly describe a filtration on $\mathcal U(\mathfrak{n}_-)$ which induces the filtrations on $L_\lambda$ given by a GT degeneration. The increasing $\mathbb{Z}$-filtration (but not a filtered algebra structure!) on $\mathcal U(\mathfrak{n}_{-})$ is defined by component $\mathcal U(\mathfrak{n}_{-})_m$ being spanned by ordered monomials $M$ with $\deg^AM\le m$. Recall the filtration $((L_\lambda)_m, m\in\mathbb{Z})$ defined in Section~\ref{generalities} (with respect to the chosen $S=\sigma(A)$). \begin{theorem}\label{Lfiltration} $(\mathcal U(\mathfrak{n}_-))_mv_\lambda=(L_\lambda)_m$ for every $m\in\mathbb{Z}$. \end{theorem} \begin{proof} Similarly to the proof of Theorem~\ref{basis}, for a $T\in\mathbb{Z}_{\ge0}^{\{1\le i<j\le n\}}$ the vector $M_T u_\lambda$ is a sum of tensor products each lying in some $U_m$ with $m\le \deg^A M_T$. This gives the inclusion $(\mathcal U_\lambda)_mv_\lambda\subset(L_\lambda)_m$. For the reverse inclusion consider a vector $v\in(L_\lambda)_m$ and express \[v=\sum_{T\in\Pi_\lambda} c_TM_Tv_\lambda.\] Among the $T$ with $c_T\neq 0$ choose a $T_0$ which has the maximal $A(T)=\deg^A M_T$ and among those with maximal $A(T)$ has the minimal $\sq(T)$. From the proof of Theorem~\ref{basis} we see that the projection of $v$ onto the direct summand $U_{T_0}$ is nonzero. This implies that $A(T_0)=\deg^A M_{T_0}\le m$ and, consequently, $v\in(\mathcal U(\mathfrak{n}_-))_m$. \end{proof} We proceed to give two characterizations of the initial ideal $\initial_{\grad^S} I$ which mimic (and follow from) well known characterizations of the ideal of Pl\"ucker relations $I$. Let us consider the polynomial ring $Q=\mathbb{C}[\{z_{i,j},1\le i\le j\le n\}]$. On this ring we have a grading $\grad^A$ given by $\grad^A z_{i,j}=a_{i,j}$ if $i<j$ and $\grad^A z_{i,i}=0$. Let $\zeta$ be the $n\times n$ matrix with $\zeta_{i,j}=z_{i,j}$ if $i\le j$ and $\zeta_{i,j}=0$ otherwise. Let $D_{i_1,\ldots,i_k}\in Q$ be the determinant of the submatrix of $\zeta$ spanned by the first $k$ rows and columns $i_1,\ldots,i_k$. First, a fact concerning non-degenerate flag varieties. \begin{proposition}\label{kerneldetclassic} $I$ is the kernel of the map $\delta$ from $R$ to $Q$ taking $X_{i_1,\ldots,i_k}$ to $D_{i_1,\ldots,i_k}$. \end{proposition} \begin{proof} This is a variation of the following classical fact which can be interpreted as a definition of the Pl\"ucker embedding. If we introduce $n\choose2$ more variables $z_{i,j}$ for $1\le j<i\le n$, consider the matrix $\zeta'$ with $\zeta_{i,j}=z_{i,j}$ and let $D'_{i_1,\ldots,i_k}$ be the same minor but in $\zeta'$, then $I$ is the kernel of the map $\delta^0$ from $R$ to $\mathbb{C}[z_{i,j},1\le i,j\le n]$ taking $X_{i_1,\ldots,i_k}$ to $D'_{i_1,\ldots,i_k}$. The map $\delta$ is the composition of $\delta^0$ and the map from $\mathbb{C}[z_{i,j},1\le i,j\le n]$ to Q taking $z_{i,j}$ to $0$ if $j<i$ and to $z_{i,j}$ if $i\le j$. Therefore, the kernel of $\delta$ contains $I$. Now, $F$ can be viewed as $GL_n/B'$, where $B'$ is the set of lower triangular matrices. If we consider a matrix $z\in GL_n$ and specialize the variables $z_{i,j}$ to the elements of this matrix, then the image of $z$ under the projection $GL_n\to F\subset \mathbb{P}$ will have homogeneous coordinates $(D'_{i_1,\ldots,i_k},1\le i_1<\ldots<i_k\le n)$ which coincides with $(D_{i_1,\ldots,i_k},1\le i_1<\ldots<i_k\le n)$ if $z$ is upper triangular. Therefore, any polynomial $p\in R$ with $\delta(p)=0$ vanishes on the subset of $F$ that is the image of the set of upper triangular matrices in $GL_n$. However, the latter image is Zariski dense in F and we obtain $p\in I$. \end{proof} Now we give an analogous fact for GT degenerations. \begin{theorem}\label{kerneldet} $\initial_{\grad^S} I$ is the kernel of the map $\delta^S$ from $R$ to $Q$ sending $X_{i_1,\ldots,i_k}$ to $\initial_{\grad^A} D_{i_1,\ldots,i_k}$. \end{theorem} \begin{proof} For a monomial $p\in Q$ let $T(p)\in\Theta$ be the point with coordinate $T(p)_{i,j}$ equal to the degree of $z_{i,j}$ in $p$. Observe that for every monomial $p$ appearing in the polynomial $D_{i_1,\ldots,i_k}$ we have $M_{T(p)} v_{\omega_k}=\pm e_{i_1,\ldots,i_k}$. Exactly one of those monomials $q$ has $T(q)=T(i_1,\ldots,i_k)$, therefrom we see that $\grad^A(\initial_{\grad^A} D_{i_1,\ldots,i_k})=s_{i_1,\ldots,i_k}$ and that $\delta^S(X_{i_1,\ldots,i_k})=\initial_{\grad^A} D_{i_1,\ldots,i_k}$ is a sum of $q$ and other monomials $p$ with $\sq(T(p))>\sq(T(i_1,\ldots,i_k))$. The fact that $\grad^A(\initial_{\grad^A} D_{i_1,\ldots,i_k})=\grad^S(X_{i_1,\ldots,i_k})$ implies (via Proposition~\ref{kerneldetclassic}) that the kernel of $\delta^S$ contains $\initial_{\grad^S} I$. To prove the reverse inclusion we show that the graded components of $\delta^S(R)$ have dimensions no less than those of $\mathcal P$. Namely, for a integral dominant weight $\lambda$ let $Q(\lambda)$ be spanned by those monomials that for every $1\le i\le n$ contain all variables of the form $z_{i,j}$ in total degree $\lambda_i$ (the $\lambda_i$ were defined in (iii) in Section~\ref{bases}). One sees that $\delta^S$ maps $R_\lambda$ into $Q(\lambda)$. (The somewhat inconsistent notation is caused by the fact that a slightly different grading on $Q$ by weights will be considered below.) Choose a $\lambda=(a_1,\ldots,a_{n-1})$ and and some $T\in\Pi_\lambda$. Combining Proposition~\ref{gtminkowski} and Lemma~\ref{gtequiv} we can decompose \[T=T^1_1+\ldots+T^1_{a_1}+\ldots+T^{n-1}_1+\ldots+T^{n-1}_{a_{n-1}}\] where $T^i_j\in\Pi_{\omega_i}$. For $T^i_j=T(i_1,\ldots,i_k)$ denote $X^i_j=X_{i_1,\ldots,i_k}$ and consider the monomial $Y_T=\prod_{i,j} X^i_j\in R_\lambda$. From the first paragraph of the proof we see that $\delta^S(Y_T)$ is the sum of a monomial $q$ with $T(q)=T$ and other monomials $p$ with $\sq(T(p))<\sq(T)$. Consequently, the expressions $\delta^S(Y_T)$ with $T$ ranging over $\Pi_\lambda$ are linearly independent and the proposition follows. \end{proof} In Example~\ref{sl3polytopes} we see that the only point in $\Pi_\lambda$ that can be decomposed into a sum of points in $\Pi_{\omega_1}$ and $\Pi_{\omega_2}$ in two different ways is ${{1\,1}\atop0}=T(1)+T(2,3)=T(2)+T(1,3)$. Herefrom once can deduce that the toric variety associated with the polytope $P_\lambda$ can be embedded into $\mathbb{P}$ as the set of zeros of the ideal $\langle X_1X_{2,3}-X_2X_{1,3}\rangle$. However, this ideal coincides with $\initial_{\grad^S}I$ obtained in Example~\ref{sl3ideal}. We generalize this to a fact that is one of our main reasons for considering these degenerations and terming them ``Gelfand--Tsetlin degenerations''. \begin{theorem}\label{toric} If all inequalities in (a) and (b) (equivalently, all inequalities in (A) and (B)) are strict and $\lambda$ is regular, then the GT degeneration $F^S$ is the toric variety associated with the polytope $P_\lambda$. This is isomorphic to the toric variety associated with the Gelfand--Tsetlin polytope $GT_\lambda$. \end{theorem} \begin{proof} As pointed out in the proof of Theorem~\ref{kerneldet}, for every monomial $p$ appearing in the polynomial $D_{i_1,\ldots,i_k}$ we have $M_{T(p)} v_{\omega_k}=\pm e_{i_1,\ldots,i_k}$. However, in view of Proposition~\ref{minmonstrict}, if all inequalities in (a) and (b) are strict, then $M_{T(i_1,\ldots,i_k)}$ is the only ordered $L_{\omega_k}$-optimal monomial mapping $v_{\omega_k}$ to $\pm e_{i_1,\ldots,i_k}$. We deduce that $\initial_{\grad^A} D_{i_1,\ldots,i_k}=\prod_{l=1}^k z_{l,i_l}$. The fact that the subring in $Q$ generated by the monomials $\prod_{l=1}^k z_{l,i_l}$ is the coordinate ring of the toric variety in question is essentially proved in~\cite[Chapter 14]{MS}. However, we can observe that this subring is the semigroup ring of the semigroup in $\mathfrak{h}^*\oplus\Theta$ generated by points of the form $(\omega_k,T(i_1,\ldots,i_k))$. This semigroup ring is the homogeneous coordinate ring of the toric variety associated with $P_\lambda$. The second claim in the proposition follows from the unimodular equivalence proved in Lemma~\ref{gtequiv}. \end{proof} The toric ideal $J$ obtained as the kernel of the map taking $X_{i_1,\ldots,i_k}$ to $\prod_{l=1}^k z_{l,i_l}$ is precisely the ideal considered in the works~\cite{GL} and~\cite{KM}. We will come back to this ideal in Sections~\ref{tropical} and~\ref{addendum}. We move on to the second characterization. Choose a complex vector $c=(c_{i,j})\in\mathbb{C}^{\{1\le i<j\le n\}}$ and consider the $G$ action $v_k(c)=\prod_{i,j} \exp(c_{i,j}f_{i,j})v_{\omega_k}\in L_{\omega_k}$, where factors in the product are ordered by $i$ increasing from left to right (which defines $v_k(c)$ uniquely in view of the commutation relations). The coordinate of $v_k(c)$ corresponding to basis vector $e_{i_1,\ldots,i_k}$ is equal to $C_{i_1,\ldots,i_k}(c)$ for some polynomial $C_{i_1,\ldots,i_k}\in\mathbb{C}[z_{i,j},1\le i<j\le n]$. In the non-degenerate case the following holds. \begin{proposition}\label{expprodclassic} $I$ is the kernel of the map $\varepsilon$ from $R$ to $Q$ sending $X_{i_1,\ldots,i_k}$ to $z_{k,k}C_{i_1,\ldots,i_k}$. \end{proposition} \begin{proof} For an integral dominant weight $\lambda$ let $\bm v_\lambda\in\mathbb{P}(L_\lambda)$ be the point corresponding to $\mathbb{C} v_\lambda$. Let $N\subset G$ be the unipotent subgroup with tangent algebra $\mathfrak{n}_-$. $N$ acts on $\mathbb{P}$ and the closure of the orbit $N\bm v$ is $F$ where $\bm v=\bm v_{\omega_1}\times\ldots\times\bm v_{\omega_{n-1}}$. Now, the Pl\"ucker coordinates of the point $\prod_{i,j}\exp(c_{i,j}f_{i,j})(\bm v)$ are precisely $C_{i_1,\ldots,i_k}$(c). In view of the additional factor $z_{k,k}$, the kernel of $\varepsilon$ is a $\wt$-homogeneous ideal that contains $I$. We are left to show that the set of points of the form $\prod_{i,j}\exp(c_{i,j}f_{i,j})(\bm v)$ is open in $F$ or, sufficiently, that the set of products of the form $\prod_{i,j}\exp(c_{i,j}f_{i,j})$ is open in $N$. In fact, it easily seen by induction on $n$ that the set of such products is all of $N$. For the induction step one writes $N=N_{n-1}\exp(\mathfrak{n}_1)$ where $N_{n-1}$ is the exponential of the subalgebra spanned by $f_{i,j}$ with $i>1$. \end{proof} Now, our analog for GT degenerations. \begin{theorem}\label{kernelexp} $\initial_{\grad^S} I$ is the kernel of the map $\varepsilon^S$ from $R$ to $Q$ sending $X_{i_1,\ldots,i_k}$ to $\initial_{\grad^A}(z_{k,k}C_{i_1,\ldots,i_k})$. \end{theorem} \begin{proof} Consider a grading on $Q$ with $Q_\lambda$ being spanned by those monomials that for every $1\le i\le n-1$ contain the variable $z_{i,i}$ in degree $a_i$. Once again, for every monomial $p$ appearing in the polynomial $z_{k,k}C_{i_1,\ldots,i_k}$ we have $M_{T(p)} v_{\omega_k}=\pm e_{i_1,\ldots,i_k}$ and exactly one of these monomials $q$ has $T(q)=T(i_1,\ldots,i_k)$. The rest of the proof repeats that of Theorem~\ref{kerneldet} verbatim modulo the appropriate substitutions. \end{proof} \section{The degenerate action}\label{degenact} In this section we define an associative algebra that acts on the GT degenerate representation spaces $L_\lambda^S$ and give an explicit description of the embedding of $F^S$ into $L_\lambda^S$ in terms of this action. Let us consider the associative algebra $\Phi_n$ generated by elements $\{\varphi_{i,j}|1\le i<j\le n\}$ with relations $\varphi_{i_1,j_1}\varphi_{i_2,j_2}=0$ whenever $i_1>i_2$ and $\varphi_{i_,j_1}\varphi_{i_,j_2}=\varphi_{i,j_2}\varphi_{i,j_1}$ for all $1\le i<j_1<j_2\le n$. For $T\in\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$ let $\varphi^T\in\Phi_n$ be the product $\prod_{i,j}\varphi_{i,j}^{T_{i,j}}$ with the factors ordered by $i$ increasing from left to right (which defines $\varphi^T$ uniquely). The elements $\varphi^T$ form a basis in $\Phi_n$. We define an action of $\Phi_n$ on the vector space $L_\lambda$. To do so for $1\le k\le n-1$ consider the Lie algebra $\mathfrak{n}_-(k)\subset\mathfrak{n}_-$ spanned by $f_{i,j}$ with $i\ge k$, we see that $\mathfrak{n}_-(1)=\mathfrak{n}_-$ and that $\mathfrak{n}_-(k)$ is a nilpotent subalgebra in $\mathfrak{sl}_{n-k+1}$. Denote $L_\lambda(k)=\mathcal U(\mathfrak{n}_-(k))v_\lambda\subset L_\lambda$. Note that the root vectors $-\alpha_{i,j}$ with $i\ge k$ generate a simple cone $\mathfrak c(k)\subset\mathfrak{h}^*$ of dimension $n-k$ with edges generated by $\alpha_i$ with $i\ge k$. One sees that $L_\lambda(k)$ is precisely the sum of all weight subspaces in $L_\lambda$ of weights $\mu$ for which $\mu-\lambda\in\mathfrak c(k)$. Our action is defined as follows. For each $\varphi_{i,j}$ and a weight vector $v\in L_\lambda$ we have $\varphi_{i,j}v=f_{i,j}v$ if $v\in L_\lambda(i)$ and $\varphi_{i,j}v=0$ otherwise. \begin{proposition}\label{phionL} This is a well-defined $\Phi_n$-module structure on $L_\lambda$. For every $T\in\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$ we have $\varphi^T v_\lambda=M_T v_\lambda$. \end{proposition} \begin{proof} We are to verify that the considered endomorphisms of $L_\lambda$ satisfy the defining relations for $\Phi_n$. The image of $f_{i,j}$ intersects $L_\lambda(i-1)$ trivially, consequently, so does the image of $\varphi_{i,j}$. This implies that the first set of relations is satisfied. The actions of $\varphi_{i,j_1}$ and $\varphi_{i,j_2}$ commute since they both annihilate anything outside of $L_\lambda(i)$ and therefore may be viewed as commuting endomorphisms of $L_\lambda(i)$. The second claim is easily obtained by induction on $\sum_{i,j} T_{i,j}$ via the fact that $\varphi_{i,j}$ preserves $L_\lambda(i)$. \end{proof} Now we introduce a $\mathbb{Z}$-grading on the algebra $\Phi_n$ by setting $\deg^A\varphi_{i,j}=a_{i,j}$, denote $\Phi_{n,m}$ the homogeneous components of this grading. Subsequently we obtain an increasing $\mathbb{Z}$-filtration on $\Phi_n$ with components $\Phi_{n,\le m}=\bigoplus_{l\le m}\Phi_{n,l}$. On one hand, this is a filtered algebra and the associated graded algebra is again $\Phi_n$ with the same grading $\deg^A$. On the other, this filtration induces a filtration on $L_\lambda$ via $(L_\lambda)_{\le m}=\Phi_{n,\le m}v_\lambda$. \begin{proposition}\label{filtrationsame} $(L_\lambda)_{\le m}=(L_\lambda)_m$, i.e.\ the newly introduced filtration coincides with the one considered previously. \end{proposition} \begin{proof} This is immediate from the second part of Proposition~\ref{phionL} and Theorem~\ref{Lfiltration}. \end{proof} Let us turn the associated graded space $L_\lambda^S$ into a $\Phi_n$-module by degenerating the action on $L_\lambda$. We have the surjections $(L_\lambda)_m\to (L_\lambda^S)_m$ with kernels $(L_\lambda)_{m-1}$ and the maps $\varphi_{i,j}:(L_\lambda)_m\to(L_\lambda)_{m+a_{i,j}}$. This induces maps $\varphi_{i,j}:(L_\lambda^S)_m\to(L_\lambda^S)_{m+a_{i,j}}$ which are summed over $m$ to provide maps $\varphi_{i,j}:L_\lambda^S\to L_\lambda^S$. Let $v_\lambda^S$ be the image of $v_\lambda\in (L_\lambda)_0$ in $(L_\lambda^S)_0$. \begin{proposition}\label{phionLS} This is a well-defined $\Phi_n$-module structure on $L_\lambda^S$. The action on $v_\lambda^S$ is described as follows. For $T\in\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$ the vector $\varphi^T v_\lambda^S$ is the projection of $M_T v_\lambda\in(L_\lambda)_{A(T)}$ to $(L_\lambda^S)_{A(T)}$ (thus $\varphi^T v_\lambda^S=0$ if $M_T$ is not $L_\lambda$-optimal). \end{proposition} \begin{proof} To compute the image of a vector $v\in(L_\lambda^S)_m$ under the action of $\varphi_{i,j}$ one may choose a preimage $v'\in(L_\lambda)_m$ of $v$ and take the image of $\varphi_{i,j}v'\in(L_\lambda)_{m+a_{i,j}}$ in $(L_\lambda^S)_{m+a_{i,j}}$. The first claim now follows from the first part of Proposition~\ref{phionL}. $\varphi^T v_\lambda^S$ is the projection of $\varphi^T v_\lambda\in (L_\lambda)_{A(T)}$ to $(L_\lambda^S)_{A(T)}$ and $\varphi^T v_\lambda=M_T v_\lambda$, therefrom we obtain the second claim. \end{proof} In particular, the above proposition combined with Proposition~\ref{basiscor} have the following consequence. \begin{cor}\label{degenbasis} For an integral dominant weight $\lambda$ the set of vectors $\{\varphi^Tv_\lambda^S,T\in\Pi_\lambda\}$ is a basis in $L_\lambda^S$. \end{cor} Our next goal is, given a $\lambda=(a_1,\ldots,a_{n-1})$, to define a $\Phi_n$-module structure on \[U_\lambda^S=(L_{\omega_1}^S)^{\otimes a_1}\otimes\ldots\otimes (L_{\omega_{n-1}}^S)^{\otimes a_{n-1}}.\] In fact, we define a monoidal structure on the category $\mathcal C$ of finite-dimensional $\Phi_n$-modules $L$ that are also equipped with an $\mathfrak{h}$-action with the following properties. First, $L$ is a direct sum of its $\mathfrak{h}$-weight spaces. Second, for a $\mathfrak{h}$-weight vector $v\in L$ and $h\in\mathfrak{h}$ we one has $h(\varphi_{i,j}(v))=\varphi_{i,j}(h(v))-h(\alpha_{i,j})\varphi_{i,j}(v)$, i.e.\ $\varphi_{i,j}$ decreases the weight by $\alpha_{i,j}$. Finally, among all weights with nonzero multiplicities in $L$ there exists a single highest weight $\lambda$ (i.e. $\lambda$ is obtained from any other by adding a sum of positive roots) such that $\varphi_{i,j}(v)=0$ for any $\varphi_{i,j}$ and any weight vector $v\in L$ of weight $\mu$ such that $\mu-\lambda\notin\mathfrak c(i)$. We refer to this $\lambda$ as the highest weight of $L$. It is easily seen that all the $L_\lambda^S$ inherit a weight decomposition from $L_\lambda$ and lie in $\mathcal C$ (with highest weight $\lambda$). For modules $L_1$, $L_2$ in $\mathcal C$ with highest weights $\lambda_1$, $\lambda_2$ denote $U=L_1\otimes L_2$. Note that a weight decomposition is induced on $U$ and for $1\le k\le n-1$ let $U(k)$ be the sum of weight subspaces in $U$ of weights $\mu$ such that $\mu-(\lambda_1+\lambda_2)\in\mathfrak c(k)$. For weight vectors $v_1\in L_1$, $v_2\in L_2$ and each $\varphi_{i,j}$ we set $\varphi_{i,j}(v_1\otimes v_2)=0$ if $v_1\otimes v_2\notin U(i)$, otherwise we set \begin{equation}\label{phionUdef} \varphi_{i,j}(v_1\otimes v_2)=\varphi_{i,j}(v_1)\otimes v_2+v_1\otimes\varphi_{i,j}(v_2) \end{equation} \begin{proposition} This is a well-defined $\Phi_n$-module structure on $U$. \end{proposition} \begin{proof} We see that the action of $\varphi_{i,j}$ subtracts $\alpha_{i,j}$ from the weight of a weight vector in $U$ and that the image of this action lies in $U(i)$ but intersects $U(i+1)$ trivially. We deduce $\varphi_{i_1,j_1}\varphi_{i_2,j_2}U=0$ whenever $i_1>i_2$. The commutation of the actions of $\varphi_{i,j_1}$ and $\varphi_{i,j_2}$ on $U$ follows from the definition~(\ref{phionUdef}) and the fact that they commute on $L_1$ and $L_2$. \end{proof} It is now obvious that $U$ lies in $\mathcal C$ (with highest weight $\lambda_1+\lambda_2$) and that the defined tensor product in $\mathcal C$ is associative and symmetric. In particular, this lets us view $U_\lambda^S$ as a $\Phi_n$-module. Next, note that a vector $e_{i_1,\ldots,i_k}$ lies in $(L_{\omega_k})_{s_{i_1,\ldots,i_k}}$ but not in $(L_{\omega_k})_{s_{i_1,\ldots,i_k}-1}$ (due to Lemma~\ref{minmon}), let $e_{i_1,\ldots,i_k}^S$ be the image of this vector in $(L_{w_k}^S)_{s_{i_1,\ldots,i_k}}$. We have linear isomorphisms between $L_{\omega_k}$ and $L_{\omega_k}^S$ sending $e_{i_1,\ldots,i_k}$ to $e_{i_1,\ldots,i_k}^S$ which induce a linear isomorphism between $U_\lambda$ and $U_\lambda^S$. Now recall the embedding $L_\lambda^S\subset U_\lambda$ from Section~\ref{generalities}. Denote the composition of the latter embedding and former isomorphism $\iota:L_\lambda^S\hookrightarrow U_\lambda^S$. Note that \[\iota(v_\lambda^S)=u_\lambda^S=(v_{\omega_1}^S)^{\otimes a_1}\otimes\ldots\otimes (v_{\omega_{n-1}}^S)^{\otimes a_{n-1}}.\] \begin{lemma} The embedding $\iota$ is a homomorphism of $\Phi_n$-modules. \end{lemma} \begin{proof} Since $L_\lambda^S$ is generated by $v_\lambda^S$ as a $\Phi_n$ module (due to Corollary~\ref{degenbasis}), it suffices to show that for any $\varphi^T$ we have $\iota(\varphi^T v_\lambda^S)=\varphi^T u_\lambda^S$. The vector $\varphi^T u_\lambda^S$ can be written explicitly as \begin{equation}\label{phiTulaS} \sum_{\sum T^k_l=T} (\varphi^{T_1^1} v_{\omega_1}^S)\otimes \ldots\otimes(\varphi^{T_l^k} v_{\omega_k}^S)\otimes\ldots\otimes (\varphi^{T^{n-1}_{a_{n-1}}} v_{\omega_{n-1}}^S). \end{equation} Here we sum over all decompositions of $T$ into a sum of $T^k_l\in\mathbb{Z}^{\{1\le i<j\le n\}}$ with $1\le k\le n-1$ and $1\le l\le a_k$. Note that, in view Proposition~\ref{phionLS}, only those summands are nonzero in which each of the monomials $M_{T^k_l}$ is $L_{\omega_k}$-optimal. Now consider $\iota(\varphi^T v_\lambda^S)$. The image of $\varphi^T v_\lambda^S\in L_\lambda^S$ under the embedding into $U_\lambda$ is seen to coincide with the projection of $M_T u_\lambda\in(L_\lambda)_{A(T)}\subset(U_\lambda)_{\le A(T)}$ to $(L_\lambda^S)_{A(T)}\subset(U_\lambda)_{A(T)}$ due to Proposition~\ref{phionLS}. Now, \[M_T u_\lambda=\sum_{\sum T^k_l=T} (M_{T_1^1} v_{\omega_1})\otimes \ldots\otimes(M_{T_l^k} v_{\omega_k})\otimes\ldots\otimes (M_{T^{n-1}_{a_{n-1}}} v_{\omega_{n-1}})\] with $\{T^k_l\}$ ranging over the same set of partitions as in~(\ref{phiTulaS}). Observe that unless each $M_{T^k_l}$ is $L_{\omega_k}$-optimal in a summand, this summand lies in $(U_\lambda)_{\le A(T)-1}$. Therefore, when taking the projection onto $(L_\lambda)_{A(T)}\subset(U_\lambda)_{\le A(T)}$ only those summands in which each $M_{T^k_l}$ is $L_{\omega_k}$-optimal remain. Finally observe that if for a $L_{\omega_k}$-optimal $M_{T^k_l}$ we have $M_{T_l^k} v_{\omega_k}=\pm e_{i_1,\ldots,i_k}$, then, due to Proposition~\ref{phionLS}, $\varphi^{T_l^k} v_{\omega_k}^S=\pm e_{i_1,\ldots,i_k}^S$. Thus our bijection from $L_{\omega_k}$ to $L_{\omega_k}^S$ maps $M_{T_l^k} v_{\omega_k}$ to $\varphi^{T_l^k} v_{\omega_k}^S$ and the assertion follows. \end{proof} \begin{cor}\label{degcartcomp} The $\Phi_n$-submodule in $U_\lambda^S$ generated by $u_\lambda^S$ is isomorphic to $L_\lambda^S$ (as a $\Phi_n$-module). \end{cor} Next, consider a complex vector $c=(c_{i,j},1\le i<j\le n)$. It is evident that each element $c_{i,j}\varphi_{i,j}$ acts nilpotently in the $\Phi_n$-modules $L_\lambda^S$ and $U_\lambda^S$ which allows us to consider the exponential of its action. We denote this exponential simply $\exp(c_{i,j}\varphi_{i,j})$. Furthermore, in each of these $\Phi_n$-modules we introduce the operator \[\exp(c)=\prod_{i,j}\exp(c_{i,j}\varphi_{i,j})\] where the factors are are ordered by $i$ increasing from left to right (which defines $\exp(c)$ uniquely). We may now straightforwardly transfer the actions of $\exp(c_{i,j}\varphi_{i,j})$ and $\exp(c)$ to the projectivizations of said $\Phi_n$-modules. Let and $\bm v_\lambda^S$ be the point in $\mathbb{P}(L_\lambda^S)$ corresponding to $\mathbb{C} v_\lambda^S$. The following theorem is what we view as the main result of this paper. \begin{theorem}\label{main} For an integral dominant regular weight $\lambda$ let $E_\lambda$ be the image of $\mathbb{C}^{\{1\le i<j\le n\}}$ in $\mathbb{P}(L_\lambda^S)$ under the map taking $c$ to $\exp(c)\bm v_\lambda^S$. The Zariski closure of $E_\lambda$ is the degenerate flag variety $F^S$. This embedding of $F^S$ into $\mathbb{P}(L_\lambda^S)$ coincides with the one given by Proposition~\ref{FSinPLS} \end{theorem} The following fact is crucial to our proof of the theorem. \begin{lemma}\label{exptensor} For any $c\in\mathbb{C}^{\{1\le i<j\le n\}}$ and $\lambda=(a_1,\ldots,a_{n-1})$ we have \[\exp(c)u_\lambda^S=(\exp(c)v_{\omega_1}^S)^{\otimes a_1}\otimes\ldots\otimes (\exp(c)v_{\omega_{n-1}}^S)^{\otimes a_{n-1}}\in U_\lambda^S.\] \end{lemma} \begin{proof} Since $\varphi_{i,j}$ acts on all $L_{\omega_k}^S(i)$ and on $U_\lambda^S(i)$, so does the one-dimensional Lie algebra $\mathbb{C}\varphi_{i,j}$. By definition, the action of this Lie algebra on $U_\lambda^S(i)$ is the tensor product of its actions on the $L_{\omega_k}^S(i)$. Therefore the Lie group $\exp(\mathbb{C}\varphi_{i,j})=\mathbb G_a$ (i.e.\ $\mathbb{C}$ under addition) also acts on these spaces and its action on $U_\lambda^S(i)$ is the tensor product of its actions on the $L_{\omega_k}^S(i)$. This means that for any \[v=v^1_1\otimes\ldots\otimes v^{n-1}_{a_{n-1}}\in U_\lambda^S(i)\] we have \[\exp(c_{i,j}\varphi_{i,j})v=\exp(c_{i,j}\varphi_{i,j})v^1_1\otimes\ldots\otimes \exp(c_{i,j}\varphi_{i,j})v^{n-1}_{a_{n-1}}.\] Now for $1\le k\le n$ denote $c(k)$ the vector with $c(k)_{i,j}=c_{i,j}$ whenever $i\ge k$ and $c(k)_{i,j}=0$ otherwise. In particular, $c(n)=0$ and $c(1)=c$. The vector $\exp(c(k))u_\lambda^S$ lies in $U_\lambda^S(k)$ and we obtain by (decreasing) induction on $k$ that \[\exp(c(k))u_\lambda^S=(\exp(c(k))v_{\omega_1}^S)^{\otimes a_1}\otimes\ldots\otimes (\exp(c(k))v_{\omega_{n-1}}^S)^{\otimes a_{n-1}}.\qedhere\] \end{proof} \begin{proof}[Proof of Theorem~\ref{main}] In view of Corollary~\ref{degcartcomp} it suffices to prove that $F^S$ coincides with the closure of the set $\{\exp(c)\bm u_\lambda^S\}\subset \mathbb{P}(U_\lambda^S)$ where $\bm u_\lambda^S$ corresponds to $\mathbb{C} u_\lambda^S$. Let us write out $\exp(c)u_\lambda^S$ as in Lemma~\ref{exptensor} and consider the tensor factor $\exp(c)v_{\omega_k}^S$. We may rewrite every $\exp(c_{i,j}\varphi_{i,j})$ as the series $1+\varphi_{i,j}+\frac{\varphi_{i,j}^2}2+\ldots$, expand the product $\exp(c)$ and then retain only those monomials $\varphi^T$ in the result for which $M_T$ is $L_{\omega_k}^S$-optimal, since all others act trivially. For a $L_{\omega_k}^S$-optimal monomial $M_T$, if $M_Tv_{\omega_k}=\pm e_{i_1,\ldots,i_k}$, then $\varphi^T v_{\omega_k}^S=\pm e_{i_1,\ldots,i_k}^S$. Now let us consider $\prod_{i,j} \exp(c_{i,j}f_{i,j})v_{\omega_k}\in L_{\omega_k}$. Let us expand every $\exp(c_{i,j}f_{i,j})$ as $1+f_{i,j}+\frac{f_{i,j}^2}2+\ldots$, then expand the product and retain only the actions of $L_{\omega_k}$-optimal monomials. Then the coordinate of the result corresponding to $e_{i_1,\ldots,i_k}$ will be equal to $\initial_{\grad^A}(C_{i_1,\ldots,i_k})(c)$ where $C_{i_1,\ldots,i_k}$ are the polynomials considered in Theorem~\ref{kernelexp}. This shows that the coordinate of $\exp(c)v_{\omega_k}^S$ corresponding to $e_{i_1,\ldots,i_k}^S$ will be equal to $\initial_{\grad^A}(C_{i_1,\ldots,i_k})(c)$. If we now compose the embedding $\mathbb{P}\subset \mathbb{P}(U_\lambda)$ from Section~\ref{generalities} with the isomorphism between $\mathbb{P}(U_\lambda)$ and $\mathbb{P}(U_\lambda^S)$, then we see that $\exp(c)\bm u_\lambda^S$ lies in $\mathbb{P}\subset\mathbb{P}(U_\lambda^S)$ and its Pl\"ucker coordinates are precisely the values $\initial_{\grad^S}(C_{i_1,\ldots,i_k})(c)$. Finally, we know from Theorem~\ref{kernelexp} that $\initial_{\grad^S}I$ is precisely the ideal of polynomials vanishing in all points with Pl\"ucker coordinates of this form with $c$ ranging over $\mathbb{C}^{\{1\le i<j\le n\}}$. This concludes the proof. The obtained embedding of $F^S$ into $\mathbb{P}(L_\lambda^S)$ coincides with the one obtained in Proposition~\ref{FSinPLS}, since we have considered the same embedding of $\mathbb{P}(L_\lambda^S)$ into $\mathbb{P}(U_\lambda)$, the same embedding of $\mathbb{P}$ into $\mathbb{P}(U_\lambda)$ and the same embedding of $F^S$ into $\mathbb{P}$. \end{proof} \begin{remark} In the case of abelian PBW degenerations as well as in~\cite{fafefom} and~\cite{favourable} the degenerate flag variety was defined as an orbit closure for a degenerate Lie group. This was then shown to coincide with a certain Gr\"obner degeneration. There is no degenerate group to be seen here, however, it turns out that considering the exponentials of generators of the degenerate algebra is sufficient. (And the fact that this embedding provided by representation theory coincides with the geometric one given by Proposition~\ref{FSinPLS} shows that we have constructed the ``correct'' embedding.) Now, in the mentioned earlier works one could define the degenerate flag variety in complete analogy with Theorem~\ref{main} without mentioning the Lie group formed by the exponentials. Since our degenerate representation theory shares many nice properties with the earlier theories of PBW degenerations (the existence of tensor products, Corollary~\ref{degcartcomp}, Lemma~\ref{exptensor} and, of course, Theorem~\ref{main}), one could argue that a degenerate Lie group is not inherent to a degenerate representation theory but is an additional nice feature of the earlier theories. This idea is strengthened by the below approach which gets rid of the exponentials altogether, requiring only tensor products with a ``Cartan component property'' analogous to Corollary~\ref{degcartcomp}. \end{remark} We now present an alternative way of characterizing $F^S$ in terms of the representation theory of $\Phi^n$. First of all, for integral dominant weights $\lambda$ and $\mu$ consider the tensor product $U_\lambda^S\otimes U_\mu^S=U_{\lambda+\mu}^S$. On one hand, by Corollary~\ref{degcartcomp} applied to $\lambda$ and $\mu$, this product contains $L_\lambda^S\otimes L_\mu^S$. On the other, it contains $L_{\lambda+\mu}^S$ as the submodule generated by $u_\lambda^S\otimes u_\mu^S=u_{\lambda+\mu}^S$. We obtain an embedding $L_{\lambda+\mu}^S\subset L_\lambda^S\otimes L_\mu^S$ and the dual surjection $(L_\lambda^S)^*\otimes (L_\mu^S)^*\twoheadrightarrow(L_{\lambda+\mu}^S)^*$. This gives a commutative algebra structure on $\mathcal Q^S=\bigoplus_\lambda (L_\lambda^S)^*$. \begin{theorem} $\mathcal Q^S$ is isomorphic to $\mathcal P^S$. \end{theorem} \begin{proof} Both of the algebras decompose as $\bigoplus_\lambda (L_\lambda^S)^*$, we are to identify the multiplicative structures. In view of Corollary~\ref{degcartcomp} the algebra $\mathcal Q^S$ is generated by the components $(L_{\omega_i}^S)^*$ and we have a surjection $R\twoheadrightarrow\mathcal Q^S$. We are to show that the kernel of this surjection is $\initial_{\grad^S} I$. For $\lambda=(a_1,\ldots,a_{n-1})$ the subspace \[W_\lambda^S=\Sym^{a_1}(L_{\omega_1}^S)\otimes\ldots\otimes\Sym^{a_{n-1}}(L_{\omega_{n-1}}^S)\subset U_\lambda^S\] is seen to be a $\Phi_n$-submodule. When restricted to homogeneity degree $\lambda$, the kernel of the above surjection is the orthogonal of the submodule $L_\lambda^S\subset W_\lambda^S$ generated by $u_\lambda^S\in W_\lambda^S$ (the bases in $L_{\omega_k}^S$ composed of the $e_{i_1,\ldots,i_k}^S$ provide a duality between $W_\lambda^S$ and $R_\lambda$). However, in the proof of Proposition~\ref{dualspaces} it was shown that $L_\lambda^S\subset W_\lambda$ is orthogonal to $\initial_{\grad^S} I_\lambda$ and the theorem follows via the linear isomorphism between $U_\lambda$ and $U_\lambda^S$. \end{proof} The above theorem can be rephrased as the following characterization of $F^S$ (where $\{(e_{i_1,\ldots,i_k}^S)^*\}\subset (L_\lambda^S)^*$ is the basis dual to $\{e_{i_1,\ldots,i_k}^S\}$). \begin{theorem}\label{mainproj} There exists a surjection $R\twoheadrightarrow\mathcal Q^S$ mapping $X_{i_1,\ldots,i_k}$ to $(e_{i_1,\ldots,i_k}^S)^*$ the kernel of which cuts out $F^S\subset\mathbb{P}$. \end{theorem} We have thus established two ways of characterizing a Gelfand--Tsetlin degeneration in terms of the representation theory of $\Phi_n$. While Theorem~\ref{main} mimics the traditional way of defining PBW degenerations as orbit closures, Theorem~\ref{mainproj} seems more natural and also has the advantage of being independent of the highest weight. \section{Singular highest weights and partial flag varieties}\label{singular} The purely representation-theoretic results in the above sections such as Theorem~\ref{basis}, Theorem~\ref{Lfiltration}, Proposition~\ref{filtrationsame} or Corollary~\ref{degenbasis} hold equally well for regular and singular highest weight $\lambda$. However, results concerned with the geometry of $F^S$ and its defining ideal $I^S$ (Proposition~\ref{FSinPLS}, Theorems~\ref{kerneldet} and~\ref{kernelexp}, Theorem~\ref{toric}, Theorems~\ref{main} and~\ref{mainproj}) are limited to the consideration of the complete flag variety $F$ and its degenerations and, therefore, only deal with regular highest weights. In this section we will recall the necessary facts concerning partial flag varieties and then generalize said results to this setting. Fix (within this section) a set ${\bf d}=\{d_1,\ldots,d_l\}\subset\{1,\ldots,n-1\}$ and an integral dominant weight $\lambda=\sum_j a_{d_j}\omega_{d_j}$ with all $a_{d_j}>0$ (i.e.\ having nonzero coordinates precisely at positions $d_1,\ldots,d_l$). The subgroup in $G$ stabilizing $v_\lambda\in L_\lambda$ is the standard parabolic subgroup $P_{\bf d}$ (depending only on $\bf d$ and not on the chosen $\lambda$) and $F_{\bf d}=G/P_{\bf d}$ is the corresponding partial flag variety. Here we obtain $P_{\bf d}=B$ and $F_{\bf d}=F$ when ${\bf d}=\{1,\ldots,n-1\}$, i.e.\ $\lambda$ is regular. Consider the subring $R_{\bf d}\subset R$ generated by all Pl\"ucker variables of the from $X_{i_1,\ldots,i_{d_j}}$, this subring is the homogeneous coordinate ring of \[\mathbb{P}_{\bf d}=\mathbb{P}(L_{\omega_{d_1}})\times\ldots\times\mathbb{P}(L_{\omega_{d_l}}).\] The Pl\"ucker embedding $F_{\bf d}\subset\mathbb{P}_{\bf d}$ of the partial flag variety is given the ideal $I_{\bf d}=I\cap R_{\bf d}$, denote the homogeneous coordinate ring $\mathcal P_{\bf d}=R_{\bf d}/I_{\bf d}$. The grading $\wt$ restricts to $R_{\bf d}$ as a grading by the semigroup generated by all $\omega_{d_j}$. If $\mu$ is a weight in this semigroup, then the corresponding homogeneous components can be identified: $R_{{\bf d},\mu}=R_\mu$, $I_{{\bf d},\mu}=I_\mu$ and $\mathcal P_{{\bf d},\mu}=\mathcal P_\mu$. The grading $\grad^S$ can also be restricted to $R_{\bf d}$ and we set $\mathcal P_{\bf d}^S=R_{\bf d}/\initial_{\grad^S}I_{\bf d}$. This is the homogeneous coordinate ring of the Gr\"obner degeneration $F_{\bf d}^S\subset\mathbb{P}_{\bf d}$ of $F_{\bf d}$ given by the ideal $\initial_{\grad^S}I_{\bf d}$. The decreasing filtration on $R$ given by $\grad^S$ induces a filtration on $R_{{\bf d}}$ and, subsequently, on $\mathcal P_{{\bf d}}$. For a weight $\mu$ in the semigroup generated by all $\omega_{d_j}$ the obtained filtration on $\mathcal P_{{\bf d},\mu}$ coincides with that on $\mathcal P_\mu$ and we see (via Proposition~\ref{initdegen}) that the homogeneous components of $\mathcal P_{\bf d}^S$ are associated graded spaces of duals of irreducible representations. Now, the Segre embedding provides an embedding $\mathbb{P}_{\bf d}\subset\mathbb{P}(U_\lambda)$ and, subsequently, $F_{\bf d}^S\subset\mathbb{P}(U_\lambda)$. We have the following generalization of Proposition~\ref{FSinPLS} which is proved in the same manner. \begin{proposition} The image of the embedding $F_{\bf d}^S\subset\mathbb{P}(U_\lambda)$ is contained in the image of $\mathbb{P}(L_\lambda^S)\subset\mathbb{P}(U_\lambda)$ (defined in Section~\ref{generalities}). \end{proposition} From the fact that $\initial_{\grad^S}I_{\bf d}=\initial_{\grad^S}I\cap R_{\bf d}$ we immediately obtain generalizations of Theorems~\ref{kerneldet} and~\ref{kernelexp}. \begin{theorem}\label{kerneldetsing} $\initial_{\grad^S} I_{\bf d}$ is the kernel of the map from $R_{\bf d}$ to $Q$ sending $X_{i_1,\ldots,i_{d_j}}$ to $\initial_{\grad^A} D_{i_1,\ldots,i_{d_j}}$. \end{theorem} \begin{theorem}\label{kernelexpsing} $\initial_{\grad^S} I_{\bf d}$ is the kernel of the map from $R_{\bf d}$ to $Q$ sending $X_{i_1,\ldots,i_{d_j}}$ to $\initial_{\grad^A}(z_{d_j,d_j}C_{i_1,\ldots,i_{d_j}})$. \end{theorem} Like in the proof of Theorem~\ref{toric}, we see (via Theorem~\ref{kerneldetsing}) that if all inequalities in (a) and (b) are strict, then $\mathcal P_{\bf d}^S$ is the semigroup ring of the semigroup in $\mathfrak{h}^*\oplus\Theta$ generated by the union of all $(\omega_{d_j},\Pi_{\omega_{d_j}})$. This is precisely the homogeneous coordinate ring of the toric variety associated with the polytope $P_\lambda$ and we obtain the following generalization. \begin{theorem} If all inequalities in (a) and (b) (equivalently, all inequalities in (A) and (B)) are strict, then the Gelfand--Tsetlin degeneration $F_{\bf d}^S$ is the toric variety associated with the polytope $P_\lambda$. This is isomorphic to the toric variety associated with the Gelfand--Tsetlin polytope $GT_\lambda$. \end{theorem} Finally, we have the generalizations of Theorems~\ref{main} and~\ref{mainproj}. The former is deduced from Theorem~\ref{kernelexpsing} just like Theorem~\ref{main} is deduced from Theorem~\ref{kernelexp} while the latter is a direct consequence of Theorem~\ref{mainproj}. \begin{theorem} Let $E_\lambda$ be the image of $\mathbb{C}^{\{1\le i<j\le n\}}$ in $\mathbb{P}(L_\lambda^S)$ under the map taking $c$ to $\exp(c)\bm v_\lambda^S$. The Zariski closure of $E_\lambda$ is the degenerate flag variety $F_{\bf d}^S$. \end{theorem} Let $\mathcal P_{\bf d}^S$ be the subalgebra in $\mathcal P^S$ spanned by $(L_\mu^S)^*$ such that $\mu$ being a sum of the $\omega_{d_j}$. \begin{theorem} There exists a surjection $R_{\bf d}\twoheadrightarrow\mathcal P_{\bf d}^S$ mapping $X_{i_1,\ldots,i_{d_j}}$ to $(e_{i_1,\ldots,i_{d_j}}^S)^*$ the kernel of which cuts out $F_{\bf d}^S\subset\mathbb{P}_{\bf d}$. \end{theorem} \section{Gr\"obner fans and tropical flag varieties}\label{tropical} It is evident that all $A$ with properties (a) and (b) (i.e.\ (A) and (B)) form the set $K$ of integer points inside a convex rational polyhedral cone $\mathcal K\subset\Theta^*$. Now, $\sigma$ can be viewed as a linear map from $\Theta^*$ to the space of Gr\"obner degenerations $\Xi=\mathbb{R}^{\{1\le i_1<\ldots<i_k\le n|1\le k\le n-1\}}$. \begin{proposition}\label{sigmaK} The linear map $\sigma$ is injective and each of $\mathcal K$ and $\sigma(\mathcal K)$ is a product of $\mathbb{R}^{n-1}$ and a simpicial cone of dimension ${n-1}\choose2$. Furthermore, the map $\sigma$ is unimodular in the sense that is establishes a bijection between the set integer points in $\Theta^*$ and the set of integer points in its image. \end{proposition} \begin{proof} The map $\sigma$ can be represented by a $(2^n-2)\times{n\choose2}$-matrix. Choose a pair $1\le i<j\le n$ and consider the row of this matrix corresponding to the tuple $(1,\ldots,i-1,j)$ (i.e.\ the integers between 1 and $i$ with $i$ replaced by $j$). One sees that $T(1,\ldots,i-1,j)_{i,j}=1$ while all other coordinates of $T(1,\ldots,i-1,j)$ are zero. This means that this row in our matrix has exactly one nonzero entry in the column corresponding to the pair $(i,j)$. This shows that the matrix has maximal rank and, therefore, $S$ is injective. One easily sees that altogether there are ${n-1}\choose2$ inequalities in (a) and (b) and that they (the functionals on $A$ these inequalities bound) are linearly independent. This immediately implies that $\mathcal K$ has the claimed form. The claim concerning $\sigma(\mathcal K)$ follows from the injectivity of $S$. The final claim can be obtained as follows. Obviously, the image of any integer point under $\sigma$ is an integer point. Conversely, consider an integer point $S=(s_{i_1,\ldots,i_k})\in \sigma(\Theta^*)$. We claim that $S=\sigma(A)$ where the coordinate $a_{i,j}$ of $A$ is equal to $s_{1,\ldots,i-1,j}$. Indeed, we know that the coordinate $\sigma(A)_{1,\ldots,i-1,j}=a_{i,j}$ and that $S$ is the unique point in the image $\sigma$ with the given coordinates $s_{1,\ldots,i-1,j}$, since all other coordinates are expressed as linear combinations of these. \end{proof} Let us briefly introduce the Gr\"obner fan and the tropicalization of the flag variety $F$, the details can be found in~\cite[Chapter 3]{MLS}. Every point in $S=(s_{i_1,\ldots,i_k})\in\Xi$ defines a Gr\"obner degeneration of $F$ but, as mentioned in Remark~\ref{projlim}, increasing all $s_{i_1,\ldots,i_k}$ for a chosen $k$ by the same constant does not change the degeneration. Therefore we can restrict our attention to the subspace $\Xi_0$ in which $s_{1,\ldots,k}=0$ for all $k$. Note that $\sigma(\mathcal K)\subset\Xi_0$. Let us define an equivalence relation on $\Xi_0$ by setting $S\sim S'$ if and only if $\initial_{\grad^{S'}}I=\initial_{\grad^S}I$. Each equivalence class is the relative interior of a closed convex rational polyhedral cone in $\Xi_0$. Together all these cones form a complete fan in $\Xi_0$ known as the Gr\"obner fan of the variety $F$. Let us consider all the cones in this fan such that for a point $S$ in the relative interior of the cone the initial ideal $\initial_{\grad^{S}}I$ does not contain any monomials in the Pl\"ucker variables. These cones form a subfan in the Gr\"obner fan known as the tropicalization of $F$ (with respect to a trivial valuation) or the {\it tropical flag variety}. \begin{theorem}\label{maxcone} $\sigma(\mathcal K)$ is a cone in the Gr\"obner fan of $F$. Moreover, $\sigma(\mathcal K)$ is a maximal cone in the tropicalization of $F$. \end{theorem} \begin{proof} For every $A\in K$ we know from Theorem~\ref{kerneldet} that $\mathcal P^{\sigma(A)}$ can be embedded into a polynomial ring, therefore $\initial_{\grad^{\sigma(A)}}I$ is prime and hence monomial free. From Proposition~\ref{sigmaK} we now deduce that all the integer points in $\sigma(\mathcal K)$ are contained in (the support of) the tropical flag variety and, consequently, so is all of $\sigma(\mathcal K)$ since it is a rational cone. Furthermore, as shown in~\cite[Lemma 3.2.10]{MLS}, a maximal cone in the tropicalization of $F$ can have dimension at most $\dim F={n\choose 2}$ which is precisely the dimension of $\sigma(\mathcal K)$. We see that it suffices to prove the first claim and the second will follow. We know that for every integer point $S$ in the relative interior of $\sigma(\mathcal K)$ the initial ideal $\initial_{\grad^S}I$ is the toric ideal $J$ discussed in Theorem~\ref{toric}, hence the same holds for every (not necessarily integer) point in the relative interior. To prove that $\sigma(\mathcal K)$ is a cone in the Gr\"obner fan we are to show that for every point $S$ in its relative boundary the ideal $\initial_{\grad^S}I$ differs from $J$. Again, since $\sigma(\mathcal K)$ and all of its proper faces are rational cones, it suffices to prove the last assertion for integer points $S$, i.e.\ points of the form $\sigma(A)$ where $A=(a_{i,j})\in K$ is such that at least one of the inequalities in (a) and (b) is an equality. Now,~\cite[Theorem 14.16]{MS} provides an explicit description of $J$. It is generated by binomials of the form \begin{equation}\label{Jrelation} X_{i_1,\ldots,i_k}X_{j_1,\ldots,j_l}-X_{\max(i_1,j_1),\ldots,\max(i_k,j_k)}X_{\min(i_1,j_1),\ldots,\min(i_k,j_k),j_{k+1},\ldots,j_l} \end{equation} where $k\le l$. Choose an $A=(a_{i,j})\in K$. Suppose that we have $a_{i,i+1}+a_{i+1,i+2}=a_{i,i+2}$ for some $1\le i\le n-2$. Let us show that $\initial_{\grad^{\sigma(A)}}I$ differs from $J$ by presenting a binomial of the form~(\ref{Jrelation}) which is not contained in $\initial_{\grad^{\sigma(A)}}I$. Indeed, we have \begin{align*} &\initial_{\grad^A}(D_{1,\ldots,i})=z_{1,1}\ldots z_{i,i},\\ &\initial_{\grad^A}(D_{1,\ldots,i-1,i+1,i+2})=z_{1,1}\ldots z_{i-1,i-1}z_{i,i+1}z_{i+1,i+2}+z_{1,1}\ldots z_{i-1,i-1}z_{i,i+2}z_{i+1,i+1},\\ &\initial_{\grad^A}(D_{1,\ldots,i-1,i+1})=z_{1,1}\ldots z_{i-1,i-1}z_{i,i+1}\text{ and}\\ &\initial_{\grad^A}(D_{1,\ldots,i,i+2})=z_{1,1}\ldots z_{i,i}z_{i+1,i+2}. \end{align*} Theorem~\ref{kerneldet} is now seen to imply \[X_{1,\ldots,i}X_{1,\ldots,i-1,i+1,i+2}-X_{1,\ldots,i-1,i+1}X_{1,\ldots,i,i+2}\notin \initial_{\grad^{\sigma(A)}}I.\] Now, suppose that $a_{i,j}+a_{i+1,j+1}=a_{i,j+1}+a_{i+1,j}$ for some $1\le i<j-1\le n-2$. Similarly to the above we observe that in this case \[X_{1,\ldots,i}X_{1,\ldots,i-1,j,j+1}-X_{1,\ldots,i-1,j}X_{1,\ldots,i,j+1}\notin \initial_{\grad^{\sigma(A)}}I.\qedhere\] \end{proof} Let us stress that, in view of the above theorem, the relative interior of $\sigma(\mathcal K)$ is the set of {\it all} Gr\"obner degenerations such that $\initial_{\grad^S}I=J$. General properties of Gr\"obner fans found in~\cite[Section 3.3]{MLS} can now be used to obtain the following. \begin{cor} The degeneration $F^{\sigma(A)}$ depends only on the minimal face of $\mathcal K$ containing $A$. Furthermore, if $A,B\in K$ are such that the minimal face of $\mathcal K$ containing $A$ contains the minimal face of $\mathcal K$ containing $B$, then $F^{\sigma(A)}$ is a Gr\"obner degeneration of $F^{\sigma(B)}$. \end{cor} \begin{remark} The toric degeneration is seen to be a Gr\"obner degeneration of any other Gelfand--Tsetlin degeneration. This allows us to use general properties of Gr\"obner degenerations and initial ideals to generalize various facts known about the toric degeneration to all GT degenerations. For instance, one can now easily deduce that any of the ideals $\initial_{\grad^{\sigma(A)}}I$ with $A\in K$ is generated by its quadratic part. Or that the set of all monomials $X_{i_1^1,\ldots,i_{k_1}^1}\ldots X_{i_1^N,\ldots,i_{k_N}^N}\in R$ such that the tuples $(i_1^1,\ldots,i_{k_1}^1),\ldots,(i_1^N,\ldots,i_{k_N}^N)$ are the columns of a semistandard Young tableau projects to a basis in $\mathcal P^{\sigma(A)}$ (see~\cite[Corollary 14.9]{MS}). \end{remark} Without going into specifics let us explain how Theorem~\ref{maxcone} can be generalized to partial flag varieties. For ${\bf d}\subset\{1,\dots,n-1\}$ one can consider the subspace $\Xi_{\bf d}\subset\Xi$ consisting of $S=(s_{i_1,\ldots,i_k})$ with $s_{i_1,\ldots,i_k}=0$ whenever $k\notin{\bf d}$. We then define the map $\sigma_{\bf d}:\Theta^*\to\Xi_{\bf d}$ as the composition of $\sigma$ and coordinatewise projection onto $\Xi_{\bf d}$. The assertion is that $\sigma_{\bf d}(\mathcal K)$ is a maximal cone in the tropicalization of $F_{\bf d}$, the latter being contained in $\Xi_{\bf d}\cap\Xi_0$. This can first be proved for Grassmannians by induction on $n$ and then generalized to arbitrary ${\bf d}$ by considering a point in the relative boundary of $\sigma_{\bf d}(\mathcal K)$ and showing that it projects into the relative boundary of some $\sigma_{\{i\}}(\mathcal K)$ with $i\in{\bf d}$. To complete the section let us give a fully explicit description of the maximal cone $\sigma(K)$ in the tropicalization, i.e. its minimal H-description. \begin{proposition} The cone $\sigma(\mathcal K)$ consists of such $S=(s_{i_1,\ldots,i_k})\in\Xi$ that all \[s_{i_1,\ldots,i_k}=\sum_{i,j} T(i_1,\ldots,i_k)_{i,j}s_{1,\ldots,i-1,j}\] and \begin{enumerate}[label=({\alph*}$'$)] \item $s_{1,\ldots,i-1,i+1}+s_{1,\ldots,i,i+2}\le s_{1,\ldots,i-1,i+2}$ for any $1\le i\le n-2$, \item $s_{1,\ldots,i-1,j}+s_{1,\ldots,i,j+1}\le s_{1,\ldots,i-1,j+1}+s_{1,\ldots,i,j}$ for any $1\le i<j-1\le n-2$. \end{enumerate} \end{proposition} \begin{proof} The inequalities in (a) and (b) provide a minimal H-description of the cone $\mathcal K$. One then applies the map $\sigma$ to these inequalities as described in the proof of Proposition~\ref{sigmaK} to obtain the proposition. \end{proof} \section{The dual construction}\label{dual} The results in Sections \ref{bases}--\ref{tropical} can be dualized via the Dynkin diagram automorphism for type $\mathrm{A}_{n-1}$. Let us show how this dualization works and why most of it, in a sense, reduces to the results already obtained. For a tuple $1\le i_1<\ldots<i_k\le n$ let $j_{k+1}<\ldots<j_n$ be the elements of $\{1,\ldots,n-1\}\backslash\{i_1,\ldots,i_k\}$. We define the points $\widetilde T(i_1,\ldots,i_k)\in\Theta$ by \begin{equation} \widetilde T(i_1,\ldots,i_k)_{l,m}= \begin{cases} 1\text{ if }m\ge k+1\text{ and }l=j_m,\\ 0\text{ otherwise}. \end{cases} \end{equation} In other words, we have the coordinate corresponding to pair $(j_m,m)$ equal to $1$ for every $k+1\le m\le n$ with $j_m<m$ and all other coordinates zero. We then define $\widetilde\Pi_{\omega_k}$ as the set of all $\widetilde T(i_1,\ldots,i_k)$ and $\widetilde\Pi_\lambda$ as the corresponding Minkowski sum. Let $\widetilde P_\lambda$ be the convex hull of $\widetilde\Pi_\lambda$. Consider the involution $\eta$ of $\Theta$ with $\eta(T)_{i,j}=T_{n+1-j,n+1-i}$. In terms of Example~\ref{sl3polytopes} this is simply reflection across a vertical line. \begin{proposition} $\widetilde P_\lambda$ is unimodularly equivalent to $GT_{\lambda}$. \end{proposition} \begin{proof} One sees that in the above notation we have \[\widetilde T(i_1,\ldots,i_k)=\eta(T(n+1-j_n,\ldots,n+1-j_{k+1})).\] Hence, $\widetilde\Pi_{\omega_k}=\eta(\Pi_{\omega_{n-k}})$ and $\widetilde\Pi_\lambda=\eta(\Pi_{\widetilde\lambda})$ where $\widetilde\lambda$ is the image of $\lambda$ under the linear involution of $\mathfrak{h}^*$ that transposes $\omega_k$ and $\omega_{n-k}$. We see that $\eta(\widetilde P_\lambda)=P_{\widetilde \lambda}$, i.e.\ $\widetilde P_\lambda$ is unimodularly equivalent to $P_{\widetilde \lambda}$ and hence $GT_{\widetilde \lambda}$. However, $GT_{\widetilde \lambda}$ is easily seen to be unimodularly equivalent to $GT_\lambda$. \end{proof} Now consider the linear invoultion $\zeta$ of $\mathfrak{n}_-$ that maps $f_{i,j}$ to $-f_{n+1-j,n+1-i}$, this is a Lie algebra automorphism. The representations $L_\lambda$ and $L_{\widetilde \lambda}$ are conjugate under this automorphism, meaning that there exists a linear isomorphism $\zeta_\lambda:L_\lambda\to L_{\widetilde \lambda}$ such that $f\zeta_\lambda(v)=\zeta_\lambda(\zeta(f)v)$ for any $f\in\mathfrak{n}_-$ and $v\in L_\lambda$. (These representations are also contragradient duals of each other but we will not be making direct use of this here.) \begin{proposition}\label{dualbasis} The set $\{\prod f_{i,j}^{T_{i,j}}v_{\widetilde\lambda},T\in\Pi_{\widetilde\lambda}\}\subset L_{\widetilde\lambda}$ with the products ordered by $j$ decreasing from left to right is (up to signs) the image of the set $\{M_Tv_\lambda,T\in\Pi_\lambda\}$ under $\zeta_\lambda$. In particular the former set constitutes a basis in $L_{\widetilde\lambda}$. \end{proposition} \begin{proof} Since $v_\lambda$ and $v_{\widetilde\lambda}$ are the only highest weight vectors in the respective representations up to a scalar factor, we can assume that $\zeta_\lambda(v_\lambda)=v_{\widetilde \lambda}$. We see that, in view of the definitions of $\eta$ and $\zeta_\lambda$, for $T\in\Pi_\lambda$ the image $\zeta_\lambda(M_Tv_\lambda)$ is $\pm\prod f_{i,j}^{\eta(T)_{i,j}}v_{\widetilde\lambda}$. It remains to recall that $\eta(\Pi_\lambda)=\Pi_{\widetilde\lambda}$. \end{proof} To dualize the results in Section~\ref{gtdegens} one considers $A\in\Theta^*$ satisfying the same inequalities (a) and (b) as before and sets \[\widetilde\sigma(A)_{i_1,\ldots,i_k}=A(\widetilde T(i_1,\ldots,i_k)).\] Let $\eta^*$ be the involution of $\Theta^*$ dual to $\eta$, i.e.\ given by $\eta^*(A)_{i,j}=A_{n+1-j,n+1-i}$. We see that for a monomial $M\in\mathcal U(\mathfrak{n}_-)$ we have $\deg^A M=\deg^{\eta^*(A)}\zeta(M)$ where we extend $\zeta$ to the universal enveloping algebra. This provides dual versions of Lemma~\ref{minmon} and Theorem~\ref{Lfiltration} via the conjugation between $L_\lambda$ and $L_{\widetilde\lambda}$, we omit the details. Furthermore, let $\Upsilon$ be the involution of $R$ mapping $X_{i_1,\ldots,i_k}$ to $X_{n+1-j_n,\ldots,n+1-j_{k+1}}$ where again \[\{j_{k+1},\ldots,j_n\}=\{1,\ldots,n-1\}\backslash\{i_1,\ldots,i_k\}.\] \begin{proposition} The ideals $\initial_{\grad^{\widetilde\sigma(A)}}I$ and $\Upsilon(\initial_{\grad^{\sigma(\eta^*(A))}}I)$ coincide. \end{proposition} \begin{proof} This follows from $\Upsilon(I)=I$ and $\grad^{\widetilde\sigma(A)}X_{i_1,\ldots,i_k}=\grad^{\sigma(\eta^*(A))}\Upsilon(X_{i_1,\ldots,i_k})$. \end{proof} Herefrom the dual versions of the results concerned with the Gr\"obner degeneration $F^{\widetilde\sigma(A)}$ in Sections~\ref{gtdegens} and~\ref{singular} are obtained straightforwardly and we again do not go into details. We point out, however, that in the dual version of Theorem~\ref{toric} the initial ideal obtained when all inequalities in (a) and (b) are strict will not be the toric ideal $J$, instead we will have $\initial_{\grad^{\widetilde\sigma(A)}I}=\Upsilon(J)$. The variety $F^{\widetilde\sigma(A)}$ will again be the toric variety of the GT polytope but $\Upsilon(J)$ provides a different projective embedding thereof. This means that when dualizing the results in Section~\ref{tropical} we obtain a different maximal cone in the tropical flag variety: \begin{theorem} The cone $\widetilde\sigma(K)$ is a maximal cone in the tropicalization of $F$ which is different from $\sigma(K)$ when $n\ge 4$. \end{theorem} Thus we have explicit descriptions of two different series of maximal cones in tropical flag varieties. This pair of cones is transposed by the action of $\mathbb{Z}/2\mathbb{Z}$ on the tropical flag variety, see~\cite{BLMM}. (This is in contrast with the maximal cone in in the tropicalization that was obtained in~\cite{fafefom}, since the latter is invariant under the $\mathbb{Z}/2\mathbb{Z}$-action.) To dualize the results in Section~\ref{degenact} one considers the associative algebra $\widetilde\Phi_n$ with generators $\widetilde\varphi_{i,j},1\le i<j\le n$ and relations $\widetilde\varphi_{i_1,j_1}\widetilde\varphi_{i_2,j_2}=0$ whenever $j_1<j_2$ and $\widetilde\varphi_{i_1,j}\widetilde\varphi_{i_2,j}=\widetilde\varphi_{i_2,j}\widetilde\varphi_{i_1,j}$ for all $1\le i_1<i_2<j\le n$. This algebra acts on $L_\lambda$ by $\widetilde\varphi_{i,j}$ acting like $f_{i,j}$ on $\mathcal U(\bigoplus_{m\le j}\mathbb{C} f_{l,m})v_\lambda$ and annihilating all weight vectors outside of this space. There is an isomorphism $\Psi$ between $\Phi_n$ and $\widetilde\Phi_n$ mapping $\varphi_{i,j}$ to $-\widetilde\varphi_{n+1-j,n+1-i}$. \begin{proposition} Define another action of $\Phi_n$ on the space $L_{\lambda}$ by letting $\varphi_{i,j}$ act as $\Psi(\varphi_{i,j})$ in the above $\widetilde\Phi_n$-action. The obtained $\Phi_n$-module is isomorphic to $L_{\widetilde\lambda}$ with the action considered in Section~\ref{degenact}. \end{proposition} \begin{proof} The isomorphism is given by the map $\zeta_\lambda$. Indeed, for $2\le j\le n$ the involution $\zeta$ maps $\bigoplus_{m\le j}\mathbb{C} f_{l,m}$ bijectively onto $\mathfrak{n}_-(n+1-j)$ and, therefore, $\zeta_\lambda$ maps $\mathcal U(\bigoplus_{m\le j}\mathbb{C} f_{l,m})v_{\lambda}$ bijectively onto $L_{\widetilde\lambda}(n+1-j)$. We now see that for a weight vector $v\in L_\lambda$ if $v\in\mathcal U(\bigoplus_{m\le j}\mathbb{C} f_{l,m})v_{\lambda}$, then \[\zeta_\lambda(\Psi(\varphi_{n+1-j,n+1-i})v)=-\zeta_\lambda(f_{i,j}v)=f_{n+1-j,n+1-i}\zeta_\lambda(v)=\varphi_{n+1-j,n+1-i}\zeta_\lambda(v),\] and if $v\notin\mathcal U(\bigoplus_{m\le j}\mathbb{C} f_{l,m})v_{\lambda}$, then \[\zeta_\lambda(\Psi(\varphi_{n+1-j,n+1-i})v)=0=\varphi_{n+1-j,n+1-i}\zeta_\lambda(v).\qedhere\] \end{proof} In other words, $\zeta_\lambda$ establishes an isomorphism between the $\Phi_n$-module $L_\lambda$ and the $\widetilde\Phi_n$-module $L_{\widetilde\lambda}$ modulo the isomorphism $\Psi$. Further details regarding the dualization of results concerned with the action of $\Phi_n$ are now recovered straightforwardly. \section{Addendum: approach via non-abelian gradings}\label{addendum} In this section we propose an alternative solution to the original problem of realizing the Gelfand--Tsetlin toric variety in a context of degenerate representation theory. Let us first recall the following standard definitions concerning initial ideals in free associative algebras. These notions originate from~\cite{Be} et al. Let $\mathcal F_n$ be the free associative $\mathbb{C}$-algebra generated by the symbols $\hat f_{i,j}$ with $1\le i<j\le n$. We have the surjection $\mathcal F_n\twoheadrightarrow\mathcal U(\mathfrak{n}_-)$ mapping $\hat f_{i,j}$ to $f_{i,j}$, denote $\mathcal I$ the kernel of this surjection. Now consider the set of monomials (i.e. products of $\hat f_{i,j}$) in $\mathcal F_n$. Under multiplication these monomials form the free monoid $\{f_{i,j}\}^*$, let $\prec$ be a total ordering of this monoid, i.e. a total order on the monomials such that for any monomials $x$, $y$ and $z$ the condition $x\prec y$ implies $xz\prec yz$ and $zx\prec zy$. Then for any $f\in\mathcal F_n$ we may define its initial part $\initial_\prec f$ as the $\prec$-\emph{maximal} monomial occurring in $f$. The initial (two-sided) ideal $\initial_{\prec}\mathcal I$ is then spanned by the monomials $\initial_\prec f$ with $f$ ranging over $\mathcal I$. Furthermore, for an integral dominant weight $\lambda$ let $\mathcal I_\lambda\subset\mathcal F_n$ be the left ideal annihilating $L_\lambda$. Then $\initial_\prec\mathcal I_\lambda$ defined as the span of the initial parts of elements of $\mathcal I_\lambda$ will be a left ideal containing $\initial_{\prec}\mathcal I$. In other words, we obtain a left module $L_\lambda^\prec=\mathcal F_n/\initial_\prec\mathcal I_\lambda$ over $\mathcal U^\prec=\mathcal F_n/\initial_{\prec}\mathcal I$. This module is generated by the vector $v_\lambda^\prec$, the image of $1$. A key property of the algebra $\mathcal U^\prec$ is that it is graded by the non-abelian monoid $\{f_{i,j}\}^*$ in a way that respects the non-commutative multiplication. The modules $L_\lambda^\prec$ are also $\{f_{i,j}\}^*$-graded in a way that respects the left $\mathcal U^\prec$-action. There is another way of defining these graded objects: as associated graded spaces. The order $\prec$ on $\{f_{i,j}\}^*$ defines a filtration of $\mathcal U(\mathfrak{n}_-)$ by this totally ordered set. For $x\in \{f_{i,j}\}^*$ the filtration component $\mathcal U(\mathfrak{n}_-)_{\preceq x}$ is defined as the span of all PBW monomials $f_{i_1,j_1}\ldots f_{i_m,j_m}$ such that $\hat f_{i_1,j_1}\ldots \hat f_{i_m,j_m}\preceq x$. We also define the space $\mathcal U(\mathfrak{n}_-)_{\prec x}$ as the span of all PBW monomials $f_{i_1,j_1}\ldots f_{i_m,j_m}$ such that $\hat f_{i_1,j_1}\ldots \hat f_{i_m,j_m}\prec x$. We then have the associated $\{f_{i,j}\}^*$-graded space \[\gr\mathcal U(\mathfrak{n}_-)=\bigoplus_{x\in\{f_{i,j}\}^*} \mathcal U(\mathfrak{n}_-)_{\preceq x}/\mathcal U(\mathfrak{n}_-)_{\prec x}.\] In view of the condition on $\prec$ the space $\gr\mathcal U(\mathfrak{n}_-)$ inherits a multiplicative structure from $\mathcal U(\mathfrak{n}_-)$ and is, in fact, an associated graded algebra. Further, the $\{f_{i,j}\}^*$-filtration on $\mathcal U(\mathfrak{n}_-)$ induces a $\{f_{i,j}\}^*$-filtration on every $L_\lambda$ by acting on $v_\lambda$ and we may again consider the associated $\{f_{i,j}\}^*$-graded spaces $\gr L_\lambda$. The following fact is proved in complete analogy with Proposition~\ref{initdegen}. \begin{proposition}\label{adinitdegen} $\gr\mathcal U(\mathfrak{n}_-)$ and $\mathcal U^\prec$ are isomorphic as $\{f_{i,j}\}^*$-graded algebras while $\gr L_\lambda$ and $L_\lambda^\prec$ are isomorphic as graded modules over these algebras. \end{proposition} \begin{remark}\label{notmonomial} To obtain abelian PBW degenerations as well as the degenerations from~\cite{favourable} and~\cite{fafefom} one needs to generalize this setting. Namely, one needs to consider total orderings of an arbitrary semigroup $\Delta$ equipped with a homomorphism $\{f_{i,j}\}^*\to\Delta$ (in the mentioned cases $\Delta$ is either $\mathbb{Z}$ or $\mathbb{Z}^{n(n-1)/2}$). One then obtains a $\Delta$-grading on $\mathcal F_n$ and the total order on $\Delta$ induces a partial monomial order on $\mathcal F_n$. The key novelty of the present construction is that we consider a non-abelian $\Delta$ (namely, all of $\{f_{i,j}\}^*$). \end{remark} We now specialize to a particular order $\prec$ which is defined as follows. For a monomial $x=\hat f_{i_1,j_1}\ldots\hat f_{i_m,j_m}$ define $|x|=\sum_k (j_k-i_k)$. Now consider two monomials $x=\hat f_{i_1,j_1}\ldots\hat f_{i_m,j_m}$ and $y=\hat f_{p_1,q_1}\ldots\hat f_{p_r,q_r}$. First, we set $x\prec y$ whenever $|x|<|y|$. Now, if $|x|=|y|$ we compare the monomials lexicographically. Namely we consider the least such $k$ that $(i_k,j_k)\neq(p_k,q_k)$ and set $x\prec y$ whenever $i_k<p_k$ or $i_k=p_k$ and $j_k<q_k$. Note that $|x|=|y|$ ensures that neither of $x$ and $y$ is a prefix of the other. This is seen to be a monomial order with the above properties. \begin{remark}\label{wthomogeneous} We invoke the function $|x|$ rather than simply comparing monomials lexicographically in order to avoid indefinite situations in which one monomial is a prefix of the other. However, we could consider any other increasing function $g$ on $[1,n]$ and set $|x|=\sum_k (g(j_k)-g(i_k))$ instead. It is easily seen that this would not alter the initial ideals $\initial_\prec\mathcal I$ and $\initial_\prec\mathcal I_\lambda$. This is a special case of the following general principle: if an ideal is homogeneous with respect to some grading, then its initial ideal is determined by the order relations between pairs of monomials of the same grading. In our case the ideals are homogeneous with respect to weight, i.e. the grading $\wt(\hat f_{i,j})=\alpha_{i,j}$. \end{remark} The algebra $\mathcal U^\prec$ is easy to describe. Denote $\chi_{i,j}\in\mathcal U^\prec$ the image of $\hat f_{i,j}$, the $\chi_{i,j}$ generate $\mathcal U^\prec$. \begin{proposition}\label{nonzeromons} A product $\chi_{i_1,j_1}\ldots\chi_{i_m,j_m}$ is nonzero if and only if $i_1\le\ldots\le i_m$ and $j_{k}\le j_{k+1}$ whenever $i_k=i_{k+1}$. These nonzero products form a basis in $\mathcal U^\prec$. \end{proposition} \begin{proof} Consider a monomial $x=\hat f_{i_1,j_1}\ldots\hat f_{i_m,j_m}$ such that for some $k$ either $i_k>i_{k+1}$ or $i_k=i_{k+1}$ and $j_k>j_{k+1}$. Let $y$ be obtained from $x$ by exchanging $f_{i_k,j_k}$ and $f_{i_{k+1},j_{k+1}}$. The ideal $\mathcal I$ contains an element of the form $x-y-z$ such that either $z=0$ or $z\prec x$ (due to the commutation relations in $\mathfrak{n}_-$). Since we also have $y\prec x$, we see that $x\in\initial_\prec\mathcal I$. The fact that such $x$ span $\initial_\prec\mathcal I$ follows from $\initial_\prec\mathcal I$ having the same character as $\mathcal I$ (i.e. the same dimensions of $\wt$-graded components in terms of Remark~\ref{wthomogeneous}). \end{proof} For $T\in\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$ denote $\chi^T=\prod_{i,j} \chi_{i,j}^{T_{i,j}}$ with the factors ordered so that the product is nonzero. The following fact explains the usefulness of the order $\prec$. \begin{lemma}\label{annideal} For any integral dominant weight $\lambda$ the $\mathcal U^\prec$-module $L_\lambda^\prec$ is isomorphic to $\mathcal U^\prec/\mathcal J_\lambda$ where $\mathcal J_\lambda$ is the left ideal spanned by $\chi^T$ with $T\notin\Pi_\lambda$. \end{lemma} \begin{proof} This can be proved using the notion of essential signatures (see~\cite{favourable,Go,essential,MY}). Namely, choose a PBW basis $\mathcal B$ in $\mathcal U(\mathfrak{n}_-)$ and a bijection $\theta:\mathcal B\to \mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$ taking each basis element to its exponent vector. Consider a total order $\prec'$ of the semigroup $\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$. Then for a dominant integral weight $\lambda$ the set of essential signatures $\essential(\lambda)\subset \mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$ (with respect to the choice of $\mathcal B$ and $\prec'$) consists of such $T$ that $\theta^{-1}(T)v_\lambda$ is not in the linear span of all $\theta^{-1}(T')v_\lambda$ with $T'\prec' T$. Obviously, the vectors $\theta^{-1}(T)v_\lambda$ with $T\in\essential(\lambda)$ form a basis in $L_\lambda$. For $T\in\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$ denote $\hat f^T=\prod_{i,j} \hat f_{i,j}^{T_{i,j}}$ with the factors ordered so that the image of this product in $\mathcal U^\prec$ is nonzero. We may now set $T_1\prec T_2$ whenever $\hat f^{T_1}\prec\hat f^{T_2}$ to define a total order on the semigroup $\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$. Also note that the image of $\hat f^T$ in $\mathcal U(\mathfrak{n}_-)$ is the monomial $M_T$ considered in the previous sections. Now, since the ideal $\initial_\prec\mathcal I_\lambda$ is monomial, we see that $L_\lambda^\prec=\mathcal U^\prec/\tilde{\mathcal J}_\lambda$ where $\tilde{\mathcal J}_\lambda$ (the annihilator of $v_\lambda^\prec$) is spanned by some set of the $\chi^T$. The characterization of $L_\lambda^\prec$ as of an associated graded space given by Proposition~\ref{adinitdegen} implies that the $\chi^T$ that are not contained in $\tilde{\mathcal J_\lambda}$ compose the set $\essential(\lambda)$ of essential signatures with respect to the chosen PBW basis and $\prec$. Since we are to prove that $\tilde{\mathcal J_\lambda}=\mathcal J_\lambda$, we are to show that $\essential(\lambda)=\Pi_\lambda$. It is known that $\essential(\lambda+\mu)$ contains the Minkowski sum $\essential(\lambda)+\essential(\mu)$ (see~\cite[Proposition 2]{Go}). Therefore, it suffices to prove that $\essential(\omega_k)=\Pi_{\omega_k}$ for all $1\le k\le n-1$. This amounts to showing that whenever $M_T v_{\omega_k}=\pm e_{i_1,\ldots,i_k}$, we have $T\succeq T(i_1,\ldots,i_k)$. In the proof of Lemma~\ref{minmon} it was established that if $M_T v_{\omega_k}=\pm e_{i_1,\ldots,i_k}$, then $M_{T(i_1,\ldots,i_k)}$ can be obtained from $M_T$ by a series of operations each of which either replaces $f_{l,m}f_{i,j}$ with $f_{l,j}f_{i,m}$ for some $l<i<j<m$ or replaces $f_{i,l}$ with $f_{i,j}f_{j,l}$ for some $i<j<l$. One sees that both of these operations decrease the exponent vector of the monomial with respect to $\prec$ and the theorem ensues. \end{proof} \begin{remark} There other ways of interpreting $\Pi_\lambda$ as a set of essential signatures. Proposition~\ref{minmonstrict} implies that when all inequalities in (A) and (B) are strict, the $M_T$ with $T\in\Pi_\lambda$ are the only ordered monomials that are $L_\lambda$-optimal. This means that in this case $\Pi_\lambda$ is the set of essential signatures with respect to, again, the PBW basis consisting of the monomials $M_T$ and the degree lexicographic order given by degree $A(T)$ and any lexicographic order. Furthermore, \cite{MY} describes a whole family of orders with respect to which (and yet again the same PBW basis) $\Pi_\lambda$ is the set of essential signatures. This family does not seem to contain the orders considered here, however. \end{remark} Lemma~\ref{annideal} shows that the structure of $L_\lambda^\prec$ is very simple. The only $\chi^T$ acting nontrivially on $v_\lambda^\prec$ are those with $T\in\Pi_\lambda$ and $\{\chi^Tv_\lambda^\prec,T\in\Pi_\lambda\}$ is a basis in $L_\lambda^\prec$. This allows us to immediately obtain an analog of Theorem~\ref{main} without defining tensor products (which will then be introduced to give an analog of Theorem~\ref{mainproj}). For a complex vector $c=(c_{i,j})\in\mathbb{C}^{\{1\le i<j\le n\}}$ and any $\lambda$ define the operator $\exp(c)$ on $L_\lambda^\prec$ as the product \[\exp(c_{1,2}\chi_{1,2})\dots\exp(c_{1,n}\chi_{1,n})\exp(c_{2,3}\chi_{2,3})\ldots\exp(c_{n-1,n}\chi_{n-1,n})\] (with the factors again ordered first by $i$ and then by $j$). These operators are invertible and induce automorphisms of $\mathbb{P}(L_\lambda^\prec)$ which we also denote $\exp(c)$. We denote $\bm v_\lambda^\prec\in\mathbb{P}(L_\lambda^\prec)$ the point corresponding to $v_\lambda^\prec$. \begin{theorem} For an integral dominant $\lambda$ the closure of the set of points $\exp(c)\bm v_\lambda^\prec\in\mathbb{P}(L_\lambda^\prec)$ with $c$ ranging over $\mathbb{C}^{\{1\le i<j\le n\}}$ is isomorphic to the toric variety associated with $P_\lambda$. \end{theorem} \begin{proof} Consider the homogeneous coordinates in $\mathbb{P}(L_\lambda^\prec)$ given by the basis $\{\chi^Tv_\lambda^\prec,T\in\Pi_\lambda\}$. We see that the homogeneous coordinate of $\exp(c)\bm v_\lambda^\prec$ corresponding to $\chi^Tv_\lambda^\prec$ is equal to $b_T\prod_{i,j} c_{i,j}^{T_{i,j}}$ for a certain constant $b_T$ independent of $c$. This means that by scaling the chosen basis we can make the homogeneous coordinate of $\exp(c)\bm v_\lambda^\prec$ corresponding to $\chi^Tv_\lambda^\prec$ simply equal to $\prod_{i,j} c_{i,j}^{T_{i,j}}$. The closure of the set of points with such homogeneous coordinates is precisely the desired toric variety. \end{proof} We move on to defining tensor products. Note that $\mathcal U^\prec$ and all $L_\lambda^\prec$ are graded by $\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$ via $\grad(\chi^T)=T$. We consider the category $\mathcal D$ of finite-dimensional $\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$-graded $\mathcal U^\prec$-modules $L$ with the following property. For any $\chi_{i,j}$ and any $\grad$-homogeneous $v\in L$ we have $\chi_{i,j}v=0$ whenever $\grad(v)_{l,m}\neq 0$ for some $(l,m)$ with $l<i$ or $l=i$ and $m<j$ (in other words, $\hat f^{\grad(v)}\prec\hat f_{i,j}$). Evidently, the $L_\lambda^\prec$ lie in this category. For $L_1$ and $L_2$ in $\mathcal D$ we see that a $\mathbb{Z}_{\ge 0}^{\{1\le i<j\le n\}}$-grading $\grad$ is induced on $L_1\otimes L_2$. For a $\grad$-homogeneous vector $v_1\otimes v_2\in L_1\otimes L_2$ we set $\chi_{i,j}(v_1\otimes v_2)=0$ if $\hat f^{\grad(v_1\otimes v_2)}\prec\hat f_{i,j}$, otherwise we set $\chi_{i,j}(v_1\otimes v_2)=\chi_{i,j}(v_1)\otimes v_2+v_1\otimes \chi_{i,j}(v_2)$. \begin{proposition} This makes $L_1\otimes L_2$ a $\mathcal U^\prec$-module lying in $\mathcal D$. The defined tensor product in $\mathcal D$ is associative and symmetric. \end{proposition} \begin{proof} We see that whenever $l<i$ or $l=i$ and $m<j$ the image $\chi_{l,m} L$ is contained in the span of vectors $v$ with $\hat f^{\grad(v)}\prec\hat f_{i,j}$ and therefore $\chi_{i,j}\chi_{l,m} L=0$. The remaining assertions are straightforward. \end{proof} The Cartan components are easily found. \begin{proposition} For integral dominant $\lambda$ and $\mu$ the $\mathcal U^\prec$-submodule in $L_\lambda^\prec\otimes L_\mu^\prec$ generated by $v_\lambda^\prec\otimes v_\mu^\prec$ is isomorphic to $L_{\lambda+\mu}^\prec$. \end{proposition} \begin{proof} $\chi^T(v_\lambda^\prec\otimes v_\mu^\prec)$ is a linear combination with positive coefficients of all the vectors of the form $\chi^{T_1}(v_\lambda^\prec)\otimes\chi^{T_2}(v_\mu^\prec)$ with $T_1\in\Pi_\lambda$, $T_2\in\Pi_\mu$ and $T_1+T_2=T$. We see that the annihilator of $v_\lambda^\prec\otimes v_\mu^\prec$ is precisely $\mathcal J_{\lambda+\mu}$ and the proposition follows. \end{proof} This proposition provides a commutative algebra structure on $\mathcal P^\prec=\bigoplus_\lambda (L_\lambda^\prec)^*$ and we have an alternative way of extracting the GT toric degeneration from the representation theory of $\mathcal U^\prec$. Denote \[e_{i_1,\ldots,i_k}^\prec=\chi^{T(i_1,\ldots,i_k)}(v_{\omega_k}^\prec)\in L_{\omega_k}^\prec\] (i.e. the image of $e_{i_1,\ldots,i_k}$ in the corresponding graded component) and let $(e_{i_1,\ldots,i_k}^\prec)^*$ compose the dual basis in $(L_{\omega_k}^\prec)^*$. Recall the toric ideal $J\subset R$ that cuts out the GT toric variety in $\mathbb{P}$. \begin{theorem} There exists a surjection $R\twoheadrightarrow\mathcal P^\prec$ with kernel $J$ mapping $X_{i_1,\ldots,i_k}$ to $(e_{i_1,\ldots,i_k}^\prec)^*$. \end{theorem} \begin{proof} It is evident that the components $(L_{\omega_k}^\prec)^*$ generate $\mathcal P^\prec$. For $\lambda=(a_1,\ldots,a_{n-1})$ consider \[W_\lambda^\prec=\Sym^{a_1}(L_{\omega_1}^\prec)\otimes\ldots\otimes\Sym^{a_{n-1}}(L_{\omega_{n-1}}^\prec)\subset (L_{\omega_1}^\prec)^{\otimes a_1}\otimes\ldots\otimes (L_{\omega_{n-1}}^\prec)^{\otimes a_{n-1}}.\] The space $W_\lambda^\prec$ is seen to be a $\mathcal U^\prec$-submodule. $W_\lambda^\prec$ is also dual to $R_\lambda$ via the chosen bases (similarly to the proof of Proposition~\ref{dualspaces}). We are to show that $L_\lambda^\prec\subset W_\lambda^\prec$ is the orthogonal of $J_\lambda\subset R_\lambda$. However, $J_\lambda$ is the span of binomials \[\prod_{k=1}^{n-1}\prod_{l=1}^{a_k}X_{i^{k,l}_1,\dots,i^{k,l}_k}-\prod_{k=1}^{n-1}\prod_{l=1}^{a_k}X_{j^{k,l}_1,\dots,j^{k,l}_k}\] with \[\sum_{k=1}^{n-1}\sum_{l=1}^{a_k}T(i^{k,l}_1,\dots,i^{k,l}_k)=\sum_{k=1}^{n-1}\sum_{l=1}^{a_k}T(j^{k,l}_1,\dots,j^{k,l}_k)\in\Pi_\lambda.\] Meanwhile, $\chi^T$ maps the highest weight vector $\bigotimes_k(v_{\omega_k}^\prec)^{a_k}$ in $W_\lambda^\prec$ to (a scalar multiple of) the sum of all \[\bigotimes_{k=1}^{n-1} \prod_{l=1}^{a_k}e_{i^{k,l}_1,\dots,i^{k,l}_k}\] with $\sum_{k,l}T(i^{k,l}_1,\dots,i^{k,l}_k)=T$ (where we refer to the symmetric multiplication in $\Sym(L_{\omega_k}^\prec)$). The theorem follows. \end{proof} \begin{remark}\label{final} To conclude let us point out that the constructions in this section appear to have a certain potential for generalization. One could define a total order $\prec'$ on $\{\hat f_{i,j}\}^*$ analogous to $\prec$ but for a different ordering of the generators $\hat f_{i,j}$. If exactly one of $\hat f_{i,l}\prec'\hat f_{i,j}$ and $\hat f_{i,l}\prec'\hat f_{j,l}$ holds for any $i<j<l$, we have a description of $\mathcal U^{\prec'}$ similar to Proposition~\ref{nonzeromons}. This lets us define the sets $\Pi'_\lambda$ consisting of exponent vectors of monomials acting nontrivially on $L_\lambda^{\prec'}$. If, in addition, we have $\Pi'_\lambda+\Pi'_\mu=\Pi'_{\lambda+\mu}$, then we also obtain analogs of other results in this section. It would be interesting to construct other such examples, especially such where non-abelian gradings are necessary (as they are here). Another question is whether some generalization of the above setting would let one obtain the degenerate representation theories from Section~\ref{degenact} directly. One would certainly need to invoke gradings by semigroups different from $\{\hat f_{i,j}\}^*$ as discussed in Remark~\ref{notmonomial} (since the annihilating ideals are not monomial). More generally, it would be interesting to adjust the construction in any way that produces the intermediate Gelfand-Tseltin degenerations (i.e. given by points lying on proper faces of $\mathcal K$). \end{remark} \section*{Acknowledgements} The author would like to thank Evgeny Feigin, Xin Fang, Ievgen Makedonskyi and Oksana Yakimova for helpful discussions of these subjects. The work was partially supported by the grant RSF 19-11-00056. This research was supported in part by the Young Russian Mathematics award.

d9dbbc9a2ad3bc788dfd46cd9ff84a936ff57113

\section{Introduction} We study inference on scaling parameters of a conditionally Gaussian process under discrete noisy observations over a fixed time interval. There are still many open questions in the field of covariance estimation of Gaussian processes under high-frequency asymptotics. Existing results reveal surprising phenomena, such as unusual convergence rates and unexpected emergences of parameters in the asymptotic covariance of estimators, which calls for a better understanding of how the underlying signal process drives asymptotic quantities of interest. Particularly, the multidimensional interplay of estimation targets encumbers the understanding of central object, such as asymptotic information. Moreover, for covariance operators that depend on high-dimensional or possibly even infinite-dimensional parameters, the mathematical analysis is not trivial. Gaussian processes constitute a versatile class with a wide range of applications. Finance marks a major field of interest in practice, where usually models driven by Brownian motions are regarded. Fractional processes yield a more controversial approach, cf. \citet{Rogers[1997]}, but are also highly relevant in, for example, geophysics and biomechanics, cf. \citet{Mandelbrot[1970]} and \citet{Bardet[2007]}. Integrated Gaussian processes are used in Physics and Biology, e.g. for modelling particles, cf. \citet{Tory[2000]}, or in the meteorological literature, cf. \citet{Boughton[1987]}. The increasing usage of sophisticated Gaussian processes, such as multifractional Brownian motions, cf. \citet{Bianchi[2013]}, calls for a general understanding of lower and upper bounds, at least for benchmark cases. As mentioned conditionally Gaussian models play a major role in finance, where inference is commonly performed conditionally on the underlying volatility process, cf. \citet{Mykland[2012]} for a general framework. A fundamental estimation problem is the extraction of the quadratic covariation (or integrated covolatility) of a continuous martingale in terms of a Brownian motion under microstructure noise. Moreover, some even consider application-driven generalisations, such as asynchronous and irregular (non-equidistant) observation schemes with varying sample sizes. Several famous approaches exist, e.g. \citet{Zhang[2005]}, \citet{Jacod[2009]}, \citet{Barndorff[2011]}, \citet{Bibinger[2014]}, \citet{Hayashi[2005]} and \citet{Christensen[2013]}, with varying limiting behaviours depending on the employed estimation techniques. These variations make a comparison of the existing approaches difficult. Additionally and importantly, the asymptotic lower bounds are not yet completely understood, even under regular observation schemes. The reason for this lies in the fact that the underlying statistical properties in these models are mathematically highly involved, which can be seen by regarding the results on efficiency in the literature. Notable works in the one-dimensional field exist, for a parametric set-up by \citet{Gloter[2001]}, and in a semi-parametric case by \citet{Reiss[2011]}, whose results are based on the verification of local asymptotic normality (LAN) and use sophisticated arguments such as asymptotic equivalences of experiments. An interesting finding in both cases, parametric and semi-parametric, is that due to the noise the optimal rate is of the unusual order $n^{-1/4}$. A multidimensional extension of these results marks the semi-parametric Cram\'er-Rao lower bound derived by \citet{Bibinger[2014]}. As the latter is provided under rather strong assumptions for synchronous and regular finite samples, in which non-parametric estimators are biased, an asymptotic characterisation of efficiency under asynchronicity is required. Moreover, \cite{Ogihara[2018]} derives asymptotic lower bounds for $d=2$. Little is known about efficient estimation if the assumption that the signal is driven by a Brownian motion is dropped. The one-dimensional Cram\'er-Rao bound derived by \citet{Sabel[2014]} is noteworthy, where the signal is given by a fractional Brownian motion. However, an asymptotic and particularly multidimensional lower bound and its dependence on the Hurst parameter remain an open question. \newpage Estimation of scaling parameters of Gaussian processes under noise also attracts interest in other fields. Related models appear in nonparametric Bayesian problems, where Gaussian process priors subject to an unknown parameter (hyperparameter) are used, cf. \citet{Szabo[2013]}. The difference in their setting lies in the asymptotic behaviour of the scaling parameter itself, whose estimation is carried-out pathwise. Latent variance estimation can also be found in genetic fields, e.g. \citet{Verzelen[2018]}. Here, the task of estimating the heritability bears structural similarities to the problems in this work. The aim of this paper is to provide a general asymptotic theory for Gaussian covariance estimation models. In the following Section~\ref{SecMainResults} the fundamental parametric model is introduced, in which the superposition of a scaled multivariate Gaussian process with additive errors is observed in equation \eqref{DiscPar}. A main contribution of this paper is the universal Convolution Theorem~\ref{ThmConvPara}, which gives a precise asymptotic characterisation of efficient estimation and includes the set-ups of \citet{Gloter[2001]} and \citet{Sabel[2014]} as special cases but also applies to more models of practical relevance given as examples below. Even though an idealised parametric model might not be as such utilisable for practical purposes, its asymptotic lower bounds provide a basic case benchmark for comparing estimation procedures of more general models. Moreover, the insight gained in the fundamental model might be used in far more complex models. This phenomenon resembles the approach with which the second main result, Theorem~\ref{ThmConvNonpara}, is derived, which marks a semi-parametric convolution theorem for estimating the integrated covolatility matrix. This result not only extends the set-up in \citet{Reiss[2011]} by multidimensionality and asynchronicity, but also weakens smoothness assumptions to Sobolev regularity $\beta>1/2$. The following section gives an overview of the main results along with their proof techniques, imposed assumptions and examples. Section~\ref{SecParametric} contains the parametric analysis, particularly the verification of Theorem~\ref{ThmConvPara}. The construction of efficient estimators is followed by further asymptotic equivalences that provide further insight on the estimation problem. Section~\ref{SecSemiPara} concludes this work by the stepwise deduction of Theorem~\ref{ThmConvNonpara}. Most of the proofs and reviews of several mathematical concepts can be found in the \hyperref[Appendix]{Appendix}. \newpage \section{Methodology and main results}\label{SecMainResults} \subsection{\textbf{Notation}} We introduce spaces of matrix-valued functions as they appear as canonical parameter sets. For $A,B\in\mathbb R^{v\times w}$ and $C\in\mathbb R^{vw\times vw}$, let \[\langle A,B\rangle_C:=\text{vec}(A)^{\top}C\text{vec}(B),\] and set $\langle\cdot,\cdot\rangle:=\langle\cdot,\cdot\rangle_{I_{vw}}$, where $\text{vec}(A)\in\mathbb R^{vw}$ is the vectorisation of $A$ and $I_k$ denotes the identity matrix in $\mathbb R^{k\times k}$. Denote the corresponding induced norms by $\|\cdot\|_C$ and $\|\cdot\|$, given that $C>0$, i.e., if $C$ is positive-definite. Note that $\|\cdot\|$ is just the Hilbert-Schmidt norm. Further let for $u\in\mathbb N,\ \Omega:=[0,1]$ and $f,g:\Omega^u\to\mathbb R^{v\times w}$ the inner product \[\langle f,g\rangle_{L^2}:=\int_{\Omega^u}\langle f(t),g(t)\rangle dt\] induce the norm $\|\cdot\|_{L^2}$ and the space $L^2=L^2(\Omega^u,\mathbb R^{v\times w})$. For $\beta\in(0,2)$ the $L^2$-subspace $H^{\beta}=H^{\beta}(\Omega^u,\mathbb R^{v\times w})$ consists of all $f:\Omega^u\to\mathbb R^{v\times w}$ such that \[\|f\|_{H^{\beta}}:=\sum_{k:|k|<\beta}\|f^{(k)}\|_{L^2}+|f|_{H^{\beta}}<\infty.\] Here $|\cdot|_{H^{\beta}}$ denotes the Sobolev-Slobodeckij semi-norm given for $\beta\neq1$ by \[|f|^2_{H^{\beta}}:=\sup_{k:|k|=\lfloor\beta\rfloor}\int_{\Omega^u}\int_{\Omega^u}\frac{\|f^{(k)}(x)-f^{(k)}(y)\|^2}{|x-y|^{2(\beta-\lfloor\beta\rfloor)+u}}dxdy,\] where $\lfloor \beta\rfloor$ denotes the integer part of $\beta$, and by $\sum_{k:|k|=1}\|f^{(k)}\|^2_{L^2}$ otherwise, where $k\in\{0,1\}^u$ denotes a multiindex with $|k|=\sum^u_{i=1}k_i$. For $u=1$ we often write $f':=f^{(1)}$. Within $H^{\beta}$ the ball of radius $L>0$ is defined via \[H^{\beta}_L:=\{f\in H^{\beta}:\|f\|_{H^{\beta}}\leq L\}.\] For $\gamma\in(0,1]$ and $N>0$ H\"older balls are given by \[C^{\gamma}_N:=\{f:\Omega\to\mathbb R:\sup\nolimits_{s,t\in\Omega}|f(s)-f(t)|^{\gamma}/|s-t|\leq N\}.\] Symmetric co-domains $\mathbb R^{d\times d}_{\text{sym}}:=\{A\in\mathbb R^{d\times d}:A=A^{\top}\}$ are highlighted by the notation $L^2_{\text{sym}}:=L^2(\Omega^u,\mathbb R^{d\times d}_{\text{sym}})$ and $H^{\beta}_{\text{sym}}:=H^{\beta}(\Omega^u,\mathbb R^{d\times d}_{\text{sym}})$. It is a basic fact that if $\beta>u/2$ for any $f\in H^{\beta}(\Omega^u,\mathbb R^{v\times w})$ a continuous version can be obtained after possibly modifying $f$ on a zero-subset of $\Omega^u$. An overview over Sobolev spaces and their embedding properties with respect to H\"older spaces can be found in \citet{Triebel[2010]}. For $Z\sim\mathcal N(0,I_d)$ the matrix $\mathcal Z=\text{Cov}(\text{vec}(ZZ^{\top}))$ is twice the so-called symmetriser matrix, i.e., it has the property $\mathcal Z\text{vec}(A)=\text{vec}(A+A^{\top})$, $A\in\mathbb R^{d\times d}$, see e.g. \citet{Abadir[2005]}. Any $d^2\times d^2$-matrix $A\otimes A$ commutes with $\mathcal Z$. Moreover, $\mathcal Z$ is positive semi-definite and therefore not invertible. For $(A_n)_{n\geq1}$ and $(B_n)_{n\geq1}$ in $\mathbb R^{d\times d}$ the expression $A_n\lesssim B_n$ means $\|A_n\|=\mathcal O(\|B_n\|)$ and $A_n\sim B_n$ means $A_n=\mathcal O(B_n)$ as well as $B_n=\mathcal O(A_n)$. Finally, for a set of parameters $\Theta$ the Le Cam distance between two statistical experiments $\mathcal E=\{(X,\mathcal X,P_{\theta}):\theta\in\Theta\}$ and $\mathcal F=\{(Y,\mathcal Y, Q_{\theta}):\theta\in\Theta\}$ on Polish spaces is given by $\Delta(\mathcal E,\mathcal F):=\max\{\delta(\mathcal E,\mathcal F),\delta(\mathcal F,\mathcal E)\}$. Here $\delta$ denotes the one-sided deficiency \[\delta(\mathcal E,\mathcal F):=\inf_K\sup_{\theta\in\Theta}\|K\cdot P_{\theta}-Q_{\theta}\|_{\text{TV}},\] where the infimum is taken over all Markov kernels from $(X,\mathcal X)$ to $(Y,\mathcal Y)$ and $\|\cdot\|_{\text{TV}}$ denotes the total variation norm. Sequences $(\mathcal E_n)_{n\geq1}$ and $(\mathcal F_n)_{n\geq1}$ of experiments are called asymptotically equivalent if $\Delta(\mathcal E_n,\mathcal F_n)=o(1)$. The latter implies that asymptotic properties transfer from one model to the other, and vice versa. Properties of $\Delta$ can be found in Appendix~\ref{SsecLeCamDistance} and \ref{SsecWeakLanReg}, see also \cite{LeCam[2000]} for a thorough introduction. \subsection{\textbf{Fundamental parametric model}} Consider the $d$-dimensional discrete observation model generated by the observations \begin{equation}\label{DiscPar} \tilde Y_i=\Sigma^{1/2}G_{i/n}+\varepsilon_i,\quad i=1,\ldots,n, \end{equation} where $G=(G_t)_{t\in[0,1]}$ is such that $G\sim\mathcal N^{\otimes d}_{0,\Gamma}$, for a centred Gaussian measure $\mathcal N_{0,\Gamma}$ on $L^2([0,1],\mathbb R)$ with covariance operator $\Gamma$. Assume that $G$ is independent of the i.i.d. errors $\varepsilon_1,\ldots,\varepsilon_n\sim\mathcal N(0,\eta^2I_d)$. The noise level $\eta>0$ is a nuisance parameter, whereas $\Sigma$ is the parameter of interest subject to \begin{equation}\label{Parspace} \Theta_0:=\{\Sigma\in\mathbb R^{d\times d}_+:0<\Sigma<SI_d\}, \end{equation} where $S>0$. Here $\mathbb R^{d\times d}_+$ denotes all positive-definite $\mathbb R^{d\times d}$-matrices and the ordering $\Sigma<SI_d$ is meant with respect to positive definiteness. An important tool paving the way to asymptotic lower bounds in the present work are several asymptotic equivalences in Le Cam's sense. In order to obtain a mathematically more convenient working basis, consider the spectral analogue of \eqref{DiscPar} given by \begin{equation}\label{SeqPar} Y_p\sim\mathcal N(0,C_p),\quad C_p:=\Sigma\lambda_p+\frac{\eta^2}{n}I_d,\quad p\geq1. \end{equation} The sequence $\lambda=(\lambda_p)_{p\geq1}$ denotes the eigenvalue sequence of the covariance operator of $\Gamma$. The approximation error between the models \eqref{DiscPar} and \eqref{SeqPar} is quantifiable by the Le Cam $\Delta$-distance, which is negligible under the following regularity assumption, cf. Proposition~\ref{PropFDiscSeq} below. \begin{assumption}\label{AssSoboCov}\textbf{-}$\pmb{G(\beta).}$ The function $(s,t)\mapsto\text{Cov}(G_s,G_t),\ s,t\in[0,1]$, lies in $H^{\beta}$ for some $\beta\in(1,2)$. \end{assumption} As an important consequence of asymptotic equivalence, LAN-expansions and convolution theorems in \eqref{DiscPar} and \eqref{SeqPar} coincide. However, as there are infinitely many non-identically distributed vectors $Y_p$ in \eqref{SeqPar} it is not clear at all whether a LAN-expansion holds since the sum of infinitely many remainder terms needs to be controlled. For the latter it will be crucial that the behaviour of certain subsequences $(\lambda_{p_n})_{n\geq1}$ carries over to the entire sequence $(\lambda_p)_{p\geq1}$ which can be done under the following. \begin{assumption}\label{regvarass}\textbf{-}$\pmb{\lambda(\delta).}$ The eigenvalues $\lambda=(\lambda_p)_{p\geq1}$ of $\Gamma$ are strictly-positive and regularly varying at infinity with index $-\delta,\ \delta>1$, i.e., \begin{equation}\label{regvar} \lim_{p\to\infty}\frac{\lambda_{\lfloor ap\rfloor}}{\lambda_p}=a^{-\delta},\ \forall a>0. \end{equation} \end{assumption} If $P^n_{\Sigma}$ denotes the measure induced by \eqref{SeqPar} then Assumption~\ref{regvarass}-$\lambda(\delta)$ ensures that a certain LAN-expansion holds, i.e., for $H\in\R^{d\times d}_{\text{sym}}$ one has \[\log\frac{dP^n_{\Sigma+r_nH}}{dP^n_{\Sigma}}\overset{P^n_{\Sigma}}{\to}\Delta_H-\frac{1}{2}\|H\|^2_{\mathcal I(\Sigma)\mathcal Z},\] where $\Delta_H\sim\mathcal N(0,\|H\|^2_{\mathcal I(\Sigma)\mathcal Z})$ and $\mathcal I(\Sigma)\mathcal Z\in\mathbb R^{d^2\times d^2}$ is the asymptotic Fisher information matrix, cf. Proposition~\ref{LAN_gen}. The rate $r_n\to0$ is obtained by \[\lim_{n\to\infty}n\lambda_{\lfloor r^{-2}_n\rfloor}=c,\] where $c>0$ is chosen such that $r_n$ is normalised with respect to multiplicative scalars, e.g. $r_n=n^{-1/4}$ but not $r_n=2n^{-1/4}$. Thus a slow decay of $\lambda$ implies a fast decay of $r_n$, and vice versa. Since the Fisher information $\mathcal I(\Sigma)\mathcal Z$ is singular it is not obvious how classical implications from LAN-theory, e.g. a convolution theorem, can be obtained. This problem is overcome by symmetrising properties of $\mathcal Z$ which allow for certain isometries, cf. Remark~\ref{RmkZZZ} below. In a non-noisy set-up \citet{Brouste[2018]} recently derived asymptotic lower bounds despite singularity by usage of certain rate matrices. For a further discussion of $r_n$ and $\mathcal I(\Sigma)$ see Section~\ref{efficiency}. \subsection{\textbf{Parametric main result}} Let $\psi(\Sigma)\in\mathbb R^k$ be a differentiable target of estimation in the sense that there is some $\nabla\psi_{\Sigma}\in\mathbb R^{k\times d^2}$ such that \begin{equation}\label{reg_esti} r^{-1}_n(\psi(\Sigma+r_nH)-\psi(\Sigma))\to\nabla\psi_{\Sigma}\text{vec}(H),\quad H\in\R^{d\times d}_{\text{sym}}, \end{equation} as $n\to\infty$. In the following, sequences of so-called regular estimators $\hat\vartheta_n$ of $\psi(\Sigma)$ are regarded, cf. Appendix~\ref{SsecWeakLanReg} for a definition. \begin{theorem}\label{ThmConvPara} Let $\hat\vartheta_n$ be a sequence of regular estimators of $\psi(\Sigma)\in\mathbb R^k$ with \eqref{reg_esti} and suppose that Assumptions~\ref{AssSoboCov}-$G(\beta)$ and \ref{regvarass}-$\lambda(\delta)$ are met. Then under $P^n_{\Sigma+r_nH},\ H\in\R^{d\times d}_{\text{sym}}$, and as $n\to\infty$ it holds that \[r^{-1}_n(\hat\vartheta_n-\psi(\Sigma+r_nH))\overset{d}{\to}\mathcal N\left(0,\tfrac{1}{4}\nabla\psi^{\top}_{\Sigma}\mathcal I(\Sigma)^{-1}\mathcal Z\nabla\psi_{\Sigma}\right)\ast R,\] for some distribution $R$. \end{theorem} The deduction of the above result offers a comprehensive understanding of how efficient estimation, particularly the optimal estimation rate $r_n$ and the geometry of the Fisher information matrix, depends on the spectral properties of the signal. Moreover, Theorem~\ref{ThmConvPara} extends the knowledge of asymptotic lower bounds in a few one-dimensional models to a general class of underlying multidimensional Gaussian processes. It is noted that only the leading term of $(\lambda_p)_{p\geq1}$ has to be known for the derivation of lower bounds. As mentioned before, several estimators have been designed for particular Gaussian models. In this work a universal estimation approach is given by \[\hat\vartheta^{\text{ad}}_n:=\sum_{p\in\pi_n}W_p\lambda^{-1}_p\text{vec}(Y_pY^{\top}_p-\eta^2/n I_d),\] where $\pi_n\subsetneq\mathbb N$ and $W_p\in\mathbb R^{d^2\times d^2}$ are adaptive weights. A spectral approach has been already used, e.g. by \citet{Bibinger[2014]}, for a covariation estimator, where martingale properties inherited from the Brownian motion are a key argument. In contrary, constructing $W_p$ independently of $(Y_p)_{p\in\pi_n}$ is the crucial idea in this work, which yields generality and gives \[r^{-1}_n(\hat\vartheta^{\text{ad}}_n-\psi(\Sigma+r_nH))\overset{d}{\to}\mathcal N\left(0,\tfrac{1}{4}\nabla\psi^{\top}_{\Sigma}\mathcal I(\Sigma)^{-1}\mathcal Z\nabla\psi_{\Sigma}\right),\] under $P^n_{\Sigma+r_nH}$, for any $H\in\R^{d\times d}_{\text{sym}}$, cf. Theorem~\ref{ThmOrAd}. The matching upper bounds imply that the derived lower bounds from Theorem~\ref{ThmConvPara} are sharp. \begin{remark}\label{RmkDepNoise} If the model is generalised to non-diagonal noise $\varepsilon_1,\ldots,\varepsilon_n\sim\mathcal N(0,H)$ with $H\in\mathbb R^{d\times d}_+$ known, then lower and upper bounds can be derived in the same way if the transformations $\tilde Y'_i:=H^{-1/2}Y_i$, $i=1,\ldots,n$ are used. In particular, $\Sigma$ in $\mathcal I(\Sigma)$ has to be replaced by $H^{-1/2}\Sigma H^{-1/2}$ and $\eta^2$ is set to the value $1$. \end{remark} \begin{remark}\label{RmkWeakDep} Another possible extension is given by weakly dependent noise. Let us consider stationary $m$-dependent noise, i.e., $\mathbb E[\varepsilon_i\varepsilon_{i+j}]=\eta_j$ with $\eta_j=0,\ j>m$, which is used in high-frequency statistics, e.g. by \citet{Hautsch[2013]}. With $\eta'_n:=\text{Var}(n^{-1/2}\sum^n_{i=1}\varepsilon_i)=\eta_0+2\sum^m_{j=1}\frac{n-j}{n}\eta_j$ a `big-block-small-block' argument gives rise to the desired connection between discrete and sequence space model in the sense that $\eta^2$ in \eqref{SeqPar} should be replaced with $\lim_{n\to\infty}\eta'_n$ and the theory provided by this work can be applied. However, this results in more assumptions on $\beta$, $\gamma$ and $m$ and is therefore omitted. \end{remark} \begin{remark}\label{RmkSigmaRandom} The techniques of this work can also be carried out if $\Sigma$ is random but $G$ given $\Sigma$ is still Gaussian. The derivation of a conditional convolution theorem is then obtained if Assumption H0 (which replaces the usage of Le Cam's third Lemma) of the general result by \citet{Clement[2013]} is met. Again, precise derivations are omitted. \end{remark} \begin{example}\label{ExBM} If $G$ denotes a $d$-dimensional Brownian motion, then $\lambda^{\text{BM}}_p=(\pi(p-1/2))^{-2}$, i.e., Assumption~\ref{regvarass} holds with $\delta=2$. Then efficient regular estimators $\hat\vartheta_n$ of $\vartheta=\text{vec}(\Sigma)$ satisfy (cf. Theorem~\ref{ThmRateFisher} below) \begin{equation}\label{clt_bm} n^{1/4}(\hat\vartheta_n-\vartheta)\overset{d}{\to}\mathcal N(0,2\eta(\Sigma\otimes\Sigma^{1/2}+\Sigma^{1/2}\otimes\Sigma)\mathcal Z). \end{equation} For $d=1$ this result coincides with \citet{Gloter[2001]} and for $d\geq 1$, \eqref{clt_bm} extends asymptotically the Cram\'er-Rao bound of \citet{Bibinger[2014]}. \end{example} \begin{example}\label{ExfBM} If $G$ is a fractional Brownian motion with Hurst exponent $H\in(0,1)$, then, by \citet{Chigansky[2018]}, the corresponding eigenvalues satisfy \eqref{regvar} with $\delta=2H+1$: \[\lambda^{\text{fBM}}_p=\frac{\sin(H\pi)\Gamma(2H+1)}{(\pi p)^{2H+1}}+o(p^{-(2H+1)}),\quad p\geq1.\] Precise asymptotic lower bounds have only been known for $d=1$ in a non-noisy setting, cf. \citet{Brouste[2018]}. In the multivariate noisy set-up Theorem~\ref{ThmConvPara} implies for $H>1/4$ that the rate of of efficient estimators is $r_n=n^{-1/(4H+2)}$, where the restriction $H>1/4$ ensures Assumption~\ref{AssSoboCov}-$G(\beta)$. The optimal asymptotic covariance can be easily calculated by Theorem~\ref{ThmRateFisher} below. Note that the Cram\'er-Rao bound in \citet{Sabel[2014]} holds for any $H\in(0,1)$. Whether the models \eqref{DiscPar} and \eqref{SeqPar} can be separated for $H\leq 1/4$ lies beyond the scope of this paper. \end{example} \begin{example} The eigenvalues $\lambda^{\text{BB}}_p=(\pi p)^{-2}$ corresponding to a Brownian bridge have the same leading term as $\lambda^{\text{BM}}_p$ in Example~\ref{ExBM}, hence \eqref{clt_bm} holds as well. Similarly, regard the (stationary) Ornstein-Uhlenbeck process \[\Sigma^{1/2}G_t=\Sigma^{1/2}G_0e^{-\beta t}+\Sigma^{1/2}\int^t_0e^{-\beta(t-s)}dB_s,\quad t\in[0,1],\] where $G_0\sim\mathcal N(0,(2\beta)^{-1}I_d),\ \beta>0$ and $B$ is a standard Brownian motion. Under the normalisation $\beta=1/2$ the eigenvalues $\lambda^{\text{OU}}_p=\frac{2\beta}{p^2\pi^2}+o(p^{-2})$ imply \eqref{clt_bm} as well. This means that mean-reversion or the behaviour of bridges have no impact on estimation of $\Sigma$. In fact, the three models corresponding to $\lambda^{\text{BM}}_p$, $\lambda^{\text{BB}}_p$ and $\lambda^{\text{OU}}_p$ are even asymptotically equivalent, cf. Proposition~\ref{PropEigLeCamEqui}. Similarly a fractional Brownian bridge and a fractional Ornstein-Uhlenbeck process seem to offer the same asymptotics as $\lambda^{\text{fBM}}_p$, cf. the (yet unpublished) drafts by \citet{Chigansky[2017]} and \citet{Chigansky[2018b]}. \end{example} \begin{example} For the $m$-fold integrated Brownian motion the eigenvalues satisfy $\lambda^{m\text{BM}}_p=(\pi p)^{-(2m+2)}+o(p^{-(2m+2)})$, cf. \citet{Wang[2008]}. This implies $r_n=n^{-1/(4m+4)}$, which reveals the interesting phenomenon that very smooth signal paths lead to rather poor estimation rates, also cf. Example~\ref{ExfBM}, where regularity is increasing in $H$ whereas $r_n$ is decreasing. \end{example} \subsection{\textbf{Semi-parametric asynchronous model}} On the basis of the parametric results asymptotic lower bounds in the more sophisticated asynchronous observation model \begin{equation}\label{DiscSemi} Y_{i,j}=(X_{t_{i,j}})_j+\varepsilon_{i,j},\quad 1\leq i\leq n_j,\ 1\leq j\leq d, \end{equation} are derived, where $X_t=X_0+\int^t_0\Sigma^{1/2}(s)dB_s$ denotes a continuous martingale in terms of a $d$-dimensional standard Brownian motion $B=(B_t)_{t\in[0,1]}$. The noise variables $\varepsilon_{i,j}\sim\mathcal N(0,\eta^2_j),\ 1\leq i\leq n_j$, with $\eta_j>0$ known, $1\leq j\leq d$, are mutually independent and independent of the signal $X=(X_t)_{t\in[0,1]}$. Moreover, suppose for the asymptotics $n_{\min}:=\min_{1\leq j\leq d}n_j\to\infty$ that $n_{\min}/n_j\to\nu_j$ for some $\nu_j\in(0,1],\ j=1,\ldots,d$. \begin{assumption}\label{AssSigma}\textbf{-}$\pmb{\Sigma(\beta,M,S).}$ For some $\beta>1/2$, $M>0$, and $S>1$ we assume that $\Sigma$ belongs to the parameter set \[\Theta_1:=\left\{A:[0,1]\to\R^{d\times d}_{\text{sym}}\Big|A\in H^{\beta}_M:S^{-1}I_d<A(t)<SI_d,\forall t\in[0,1]\right\}.\] \end{assumption} \begin{assumption}\label{AssRegF}\textbf{-}$\pmb{F(\gamma,N,\beta).}$ The observation times obey $t_{i,j}=F^{-1}_j(i/n_j)$ for a distribution function $F_j:[0,1]\to[0,1]$ with derivative $F'_j$ and \begin{itemize} \item[(i)] $F_j(0)=0$ and $F_j(1)=1$, \item[(ii)] $F'_j\in C^{\gamma}_N$ and $F'_j>0$, \end{itemize} for $j=1,\ldots,d$, and some $\gamma\in(\beta,1],\ N>0$. \end{assumption} As in the parametric set-up, \eqref{DiscSemi} is approximated by a spectral representation for which the conditions $\gamma>\beta>1/2$ and $\Sigma>S^{-1}I_d$ are needed. The latter one is slightly restrictive but not uncommon, cf. \citet{Reiss[2011]}. The spectral representation is given by the mutually independent random vectors \begin{equation}\label{SeqSemi} Y_{pk}\sim\mathcal N(0,C_{pk}),\quad k=0,\ldots,m-1,\ p\geq 1, \end{equation} where $C_{pk}:=\Sigma(k/m)\lambda_{mp}+n^{-1}_{\min}\Xi^2(k/m),\ \lambda_{mp}:=(\pi pm)^{-2}$ and \[\Xi^2(t):=\text{diag}(\eta^2_j\nu_j/(F'_j(t)))_{1\leq j\leq d}.\] However, the approximation of \eqref{DiscSemi} by \eqref{SeqSemi} holds only for localisations $\Sigma+n^{-1/4}_{\min}H,\ H\inH^{\beta}_{\text{sym}}$, which nevertheless is the right ingredient to ensure that LAN-expansions in the sequence space carry over to \eqref{DiscSemi}, cf Proposition~\ref{PropLANequi}. \subsection{\textbf{Semi-parametric main result}} For each $k$ the sequence $(Y_{pk})_{p\geq1}$ in \eqref{SeqSemi} is of the same type as the fundamental sequence space model in \eqref{SeqPar}. Indeed the parametric results can be applied simultaneously (over $k$) to the setting \eqref{SeqSemi}, for which we consider targets of estimation given by \begin{equation}\label{targetinf} \psi(\Sigma):=\int^1_0(W(\Sigma))(t)dt \end{equation} with a differentiable weight $W:\Theta_1\to L^2([0,1],\mathbb R^{d^2})$ in the sense that \begin{equation}\label{targetreg} n^{1/4}_{\min}(W(\Sigma+n^{-1/4}_{\min}H)-W(\Sigma))\to\nabla W_{\Sigma}\cdot\text{vec}(H),\quad H\inH^{\beta}_{\text{sym}}, \end{equation} as $n_{\min}\to\infty$, for some $\nabla W_{\cdot}\in L^2([0,1],\mathbb R^{d^2\times d^2})$. An example is given by the choice $W(\Sigma)=\text{vec}(\Sigma)$ with $\nabla W_{\cdot}=I_{d^2}$. \begin{theorem}\label{ThmConvNonpara} Let $\hat\vartheta_n$ be a sequence of regular estimators of $\psi(\Sigma)$ as in \eqref{targetinf} with \eqref{targetreg} and suppose that Assumptions~\ref{AssSigma}-$\Sigma(\beta,M,S)$ and \ref{AssRegF}-$F(\gamma,N,\beta)$ are met. Then under $Q^n_{\Sigma+n^{-1/4}_{\min}H},\ H\inH^{\beta}_{\text{sym}}$, it holds that \[n^{1/4}_{\min}(\hat\vartheta_n-\psi(\Sigma+n^{-1/4}_{\min}H))\overset{d}{\to}\mathcal N\Big(0,\frac{1}{4}\int^1_0(\nabla W_{\Sigma}\mathcal I^{-1}_{\Sigma}\mathcal Z\nabla W_{\Sigma}^{\top})(t)dt\Big)\ast R,\] as $n_{\min}\to\infty$, for some $R$, where $Q^n_{\Sigma}$ is the measure induced by \eqref{SeqSemi} and \[\mathcal I^{-1}_{\Sigma}(t)=8(\Sigma^{1/2}_{\Xi}(t)\otimes\Sigma(t)+\Sigma(t)\otimes \Sigma^{1/2}_{\Xi}(t)),\ t\in[0,1],\] with $\Sigma^{1/2}_{\Xi}:=\Xi(\Xi^{-1}\Sigma\Xi^{-1})^{1/2}\Xi$. \end{theorem} The above statement extends the one-dimensional asymptotic efficiency results of \citet{Reiss[2011]} in various ways. Firstly, the needed H\"older-regularity $(1+\sqrt{5})/4\approx0.81$ in \citet{Reiss[2011]} can be relaxed to Sobolev regularity $\beta>1/2$. This relaxation is achieved by focussing on asymptotically equivalent experiments that share the same semi-parametric lower bounds for targets as in \eqref{targetinf}, whereas Rei\ss\ even considers experiments with common asymptotic non-parametric lower bounds. Moreover, Theorem~\ref{ThmConvNonpara} allows for multidimensionality of $\Sigma$ as well as for asynchronicity and therefore extends asymptotically the basic case Cram\'er-Rao bound for continuously differentiable $\Sigma$ by \citet{Bibinger[2014]}. Since the local method of moments estimator provided by \citet{Bibinger[2014]} attains the Gaussian part of the limit distribution of Theorem~\ref{ThmConvNonpara}, the derived bounds are sharp. \begin{remark} The steps that are taken to establish Theorem~\ref{ThmConvNonpara} can be developed analogously if $\Sigma=(\Sigma_t)_{t\in[0,1]}$ is assumed to be random with realisations in $\Theta_1$ and if $X$ conditioned on $\Sigma$ is still Gaussian. Again the result by \citet{Clement[2013]} gives a conditional convolution theorem, cf. Remark~\ref{RmkSigmaRandom}. The estimator provided by \citet{Altmeyer[2015]} attains the corresponding asymptotic stochastic lower bounds. Similarly, extensions for the noise can be obtained as illustrated in Remark~\ref{RmkDepNoise} and \ref{RmkWeakDep}. \end{remark} \section{Analysis of the fundamental parametric model}\label{SecParametric} Throughout this section we assume that $\Sigma\in\Theta_0$ for some $S>0$, cf. \eqref{Parspace}, and that Assumption~\ref{AssSoboCov}-$G(\beta)$ and Assumption~\ref{regvarass}-$\lambda(\delta)$ are satisfied. \subsection{\textbf{Connection between discrete and sequence space model}}\label{secCtsSeq} Consider the discrete observation model \eqref{DiscPar} and its continuous analogue \begin{equation}\label{ContPar} dY_t=\Sigma^{1/2}G_tdt+\frac{\eta}{\sqrt{n}}dW_t,\quad t\in[0,1], \end{equation} where $W$ is a Wiener process independent of $G$. The model \eqref{ContPar} is consistent with observing the stochastic bilinear forms \begin{equation}\label{cylmeas} Y_f:=(f,dY):=\sum^d_{j=1}\int^1_0(f(t))_jd(Y_t)_j,\quad f\in L^2([0,1],\mathbb R^d). \end{equation} $Y_f$ is Gaussian with $\mathbb E[Y_f]=0$ and $\text{Cov}(Y_f,Y_g)=\langle K_{\Sigma,n}f,g\rangle_{L^2}$. The underlying covariance operator $K_{\Sigma,n}$ is given by \[K_{\Sigma,n}:=T_{\Sigma^{1/2}}\text{diag}(\Gamma)_{1\leq j\leq d}T_{\Sigma^{1/2}}+\frac{\eta^2}{n}\text{Id},\] with $T_{\Sigma^{1/2}}:f\mapsto\Sigma^{1/2} f,\ \text{Id}:f\mapsto f$ and $\text{diag}(\Gamma)_{1\leq j\leq d}:f\mapsto(\Gamma f_j)_{1\leq j\leq d}$ being the covariance operator of $G$. For the orthonormal eigenbasis $(\varphi_p)_{p\geq1}$ of $\Gamma$ and $e_{pi}:=(\mathbbm{1}_{\{i=j\}}\varphi_p)_{1\leq j\leq d}$ the vectors $(Y_{e_{p1}},\ldots,Y_{e_{pd}})^{\top},\ p\geq1,$ follow the same distribution as the sequence $(Y_p)_{p\geq1}$ in \eqref{SeqPar}. \begin{definition} Denote by $\mathcal F_n$ and $\mathcal F^s_n$ the statistical experiments that are generated by the observations \eqref{DiscPar} and \eqref{SeqPar}, respectively. \end{definition} Since $(\varphi_p)_{p\geq 1}$ is a basis, observing the sequence $(Y_p)_{p\geq1}$ in \eqref{SeqPar} is equivalent to observe \eqref{ContPar}. Moreover, the following is just a consequence of the more general Theorem~\ref{ThmLeCamDiscCont} given in the Appendix. \begin{proposition}\label{PropFDiscSeq} Under Assumption~\ref{AssSoboCov}-$G(\beta)$ the experiments $\mathcal F_n$ and $\mathcal F^s_n$ are asymptotically equivalent. More precisely, the Le Cam distance obeys \[\Delta(\mathcal F_n,\mathcal F^s_n)=\mathcal O(Sn^{1-\beta}).\] \end{proposition} \subsection{\textbf{Local asymptotic normality}}\label{efficiency} Denote the score in $\mathcal F^s_n$ by $\nabla\ell_n(\Sigma):=\sum_{p\geq1}\ell_{np}(\Sigma)$ and set $\mathcal I_n(\Sigma)\mathcal Z:=\text{Cov}(\nabla\ell_n(\Sigma))$, where \begin{equation}\label{SeqScore} \nabla\ell_{np}(\Sigma):=\frac{1}{2}\lambda_p\text{vec}(C^{-1}_pY_{p}Y^{\top}_{p}C^{-1}_p-C^{-1}_p). \end{equation} The Fisher information $\mathcal I_n(\Sigma)\mathcal Z=\sum_{p\geq1}\mathcal I_{np}(\Sigma)\mathcal Z\in\mathbb R^{d^2\times d^2}$ is driven by \[\mathcal I_{np}(\Sigma):=\frac{1}{4}\lambda^2_p(C^{-1}_p\otimes C^{-1}_p)\quad p\geq1.\] In the derivation of $\ell_n$ and $\mathcal I_n$ the following well-known identity was used: \[\text{vec}(ABC)=(C^{\top}\otimes A)\text{vec}(B),\quad A,B,C\in\mathbb R^{d\times d}.\] As a consequence of Assumption~\ref{regvarass}, $\mathcal I_n(\Sigma)\mathcal Z$ is well-defined. A crucial quantity is the rate $r_n\to0$ such that the asymptotic Fisher information \[\mathcal I(\Sigma)\mathcal Z:=\lim_{n\to\infty}r^2_n\mathcal I_n(\Sigma)\mathcal Z\] is well-defined, where $r_n$ is assumed to be normalised with respect to scalars, e.g. $r_n=n^{-1/4}$. The key to finding this rate $r_n$ lies in the interplay between the operators $\text{diag}(\Gamma)_{1\leq j\leq d}$ and $\tfrac{1}{n}\text{Id}$ along with the regular variation of $\lambda$. More precisely, in the covariance matrices $C_p=\Sigma\lambda_p+\tfrac{\eta^2}{n}I_d$, the impact of signal and noise is (nearly) balanced at the index $p_n$ with $\lambda(p_n)=n^{-1}$, where we identify the sequence $\lambda$ with some continuously interpolated non-increasing analogue $\lambda:\mathbb R_+\to\mathbb R_+$. It is well-known, that the representation \begin{equation}\label{regvarrepr} \lambda(p)=p^{-\delta}L(p) \end{equation} is valid, for some slowly varying $L:\mathbb R_+\to\mathbb R_+$, cf. \citet{Bingham[1989]}. \begin{theorem}\label{ThmRateFisher} Grant Assumption~\ref{regvarass}-$\lambda(\delta)$ on $\Gamma$. Then the Fisher information satisfies for any $\Sigma\in\R^{d\times d}_{\text{sym}}$ with $\Sigma>0$ \begin{equation}\label{FisherRate} p^{-1}_n\mathcal I_n(\Sigma)\mathcal Z\to\mathcal I(\Sigma)\mathcal Z,\quad\text{as }n\to\infty, \end{equation} where $p_n$ is given by $\lambda(p_n)=n^{-1}$. If $Q$ is an orthogonal matrix such that $\Sigma=Q^{\top}\text{diag}(s_1,\ldots,s_d)Q$ then \[\mathcal I(\Sigma)=(Q\otimes Q)^{\top}\text{diag}(v_{11},\ldots,v_{1d},v_{21},\ldots,v_{2d},v_{31},\ldots,v_{dd})(Q\otimes Q)\] with eigenvalues \[v_{i,j}=\frac{\zeta}{4\eta^{2/\delta}}\int^1_0(s_i+x^{\delta})^{-1}(s_j+x^{\delta})^{-1}dx,\quad i,j=1,\ldots,d,\] where $\zeta=\lim_{n\to\infty}r^2_np_n$ for $r_n\sim p^{-1/2}_n$ standardised. Moreover, the convergence in \eqref{FisherRate} already holds for $\mathcal I_{\pi_n}(\Sigma):=\sum_{p\in\pi_n}\mathcal I_{np}(\Sigma)$, whenever $\pi_n=[\underline{\pi_n},\overline{\pi_n}]\cap\mathbb N,$ with $\underline{\pi_n}/p_n\to0$ and $(\underline\pi_n\wedge\overline{\pi_n}/p_n)\to\infty$. \end{theorem} By the above statement the rate $r_n$ satisfies the relation \[r_n L(r^{-2}_n)^{1/(2\delta)}\sim n^{-1/(2\delta)},\] with $L$ as in \eqref{regvarrepr}. Thus the rate $r_n$ is completely determined by the decay of $\lambda$. The slower $\lambda$ decreases the more observations $Y_p$ carry significant information about $\Sigma$ and the faster $\Sigma$ can be estimated. Moreover, solely the limiting behaviour of $L$ determines the constant $\zeta$. For instance, in the Brownian motion case $\lambda^{\text{BM}}_p=(p-1/2)^{-2}\pi^{-2}$ one has $\delta=2,p_n=\sqrt{n}/\pi+1/2$ and $L(p)=(\pi(2-1/(2p)))^{-2}$, which gives $r_n=n^{-1/4}$ and $\zeta=1/\pi$. A simple calculation, cf. Remark~\ref{RmkFisherExp}, shows, that the eigenvalues obey \[v_{i,j}=\frac{\zeta\pi}{4\delta\sin(\pi/\delta)\eta^{2/\delta}}\cdot\frac{s^{1/\delta-1}_j-s^{1/\delta-1}_i}{s_i-s_j}\] and that they are driven by the slope of $x\mapsto-x^{1/\delta-1}$ between all pairs $(s_i,s_j)$. Whenever $s_i=s_j$ the slope equals the derivative at $s_i$. In particular, for the case $\Sigma=\sigma^2\in\mathbb R_+$ the Fisher information becomes \[\mathcal I(\sigma^2)=\frac{\zeta\pi(1-1/\delta)}{4\delta\sin(\pi/\delta)\eta^{2/\delta}}\sigma^{2/\delta-4}.\] Sufficient information to estimate $\Sigma$ efficiently in asymptotics is already provided by those observations $Y_p$ in $\mathcal F^s_n$, such that $p$ is subject to an interval $\pi_n$ as in Theorem~\ref{ThmRateFisher}. This means that maximal information about $\Sigma$ is asymptotically contained in (arbitrarily slowly) increasing neighbourhoods of $p_n$ within the spectrum of $Y=(Y_t)_{t\in[0,1]}$ in $\mathcal F_n$. This gives canonical choices of truncation indices for spectral estimators of $\Sigma$, cf. Section~\ref{estimation}. For $\Sigma\in\Theta_0$ consider local alternatives of the form $\Sigma+r_nH,\ H\in\R^{d\times d}_{\text{sym}}$, where $r_n$ is chosen according to Theorem~\ref{ThmRateFisher}. Note that $\Sigma+r_nH\in\Theta_0$ for $n$ sufficiently large, hence $P^n_{\Sigma+r_nH}$ might be defined arbitrarily, whenever $\Sigma+r_nH\notin\Theta_0$. Denote by $\Delta_H$ the centred Gaussian process with \[\text{Cov}(\Delta_{H_1},\Delta_{H_2})=\langle H_1,H_2\rangle_{\mathcal I(\Sigma)\mathcal Z},\quad H_1,H_2\in\R^{d\times d}_{\text{sym}},\] where it is noted that $\mathcal I(\Sigma)\mathcal Z$ is positive definite on $\{\text{vec}(H):H\in\R^{d\times d}_{\text{sym}}\}$. \begin{proposition}\label{LAN_gen} Under Assumption~\ref{regvarass}-$\lambda(\delta)$, for any $\Sigma\in\Theta_0$, the following asymptotic expansion is satisfied in $\mathcal F^s_n$ as $n\to\infty$: \begin{equation}\label{LAN} \log\frac{dP^n_{\Sigma+r_nH}}{dP^n_{\Sigma}}=\Delta_{n,H}-\frac{r^2_n}{2}\|H\|^2_{\mathcal I_n(\Sigma)\mathcal Z}+\rho_n,\quad H\in\R^{d\times d}_{\text{sym}}, \end{equation} where $\Delta_{n,H}\overset{d}{\to}\Delta_H$, under $P^n_{\Sigma},\ r^2_n\|H\|^2_{\mathcal I_n(\Sigma)\mathcal Z}\to\| H\|^2_{\mathcal I(\Sigma)\mathcal Z}$ and $\rho_n=o_{P^n_{\Sigma}}(1)$. \end{proposition} Note that $\Delta_{n,H}=r_n\text{vec}(H)^{\top}\nabla\ell_n(\Sigma)$, where $\nabla\ell_n$ denotes the score in $\mathcal F^s_n$. Moreover, the remainder obeys $\rho_n=\rho^{(1)}_n+\rho^{(2)}_n$ with $\mathbb E[\rho^{(1)}_n]=0$ and \begin{align} \label{LANrem1}\text{Var}(\rho^{(1)}_n)\leq&r^2_n\|H\|^2\|\Sigma^{-1}\|^2r^2_n\|H\|^2_{\mathcal I_n(\Sigma)\mathcal Z}=\mathcal O(r^2_n),\\ \label{LANrem2}|\rho^{(2)}_n|\leq&2r_n\|H\|\|\Sigma^{-1}\|r^2_n\|H\|^2_{\mathcal I_n(\Sigma)\mathcal Z}=\mathcal O(r_n), \end{align} hence \eqref{LAN} holds uniformly in $H$ over balls within $\R^{d\times d}_{\text{sym}}$.\newpage An implication of the LAN-property~\eqref{LAN} is weak convergence of the localisations $\{P^n_{\Sigma+r_nH}:H\in\R^{d\times d}_{\text{sym}}\}$ to the Gaussian shift experiment $\mathcal G:=\{\mathcal N(\mathcal I(\Sigma)\mathcal Z\text{vec}(H),\mathcal I(\Sigma)\mathcal Z):H\in\R^{d\times d}_{\text{sym}}\}$. Given an observation $Y$ in $\mathcal G$ the property $\mathcal Z\text{vec}(H)=2\text{vec}(H)$ implies that the best unbiased estimator of $\text{vec}(H)$ is given by $\frac{1}{2}\mathcal I(\Sigma)^{-1}Y\sim\mathcal N(\text{vec}(H),\frac{1}{4}\mathcal I(\Sigma)^{-1}\mathcal Z)$. This determines the asymptotic distribution of regular estimators, which is made precise in the following. \subsection{\textbf{Verification of Theorem~\ref{ThmConvPara}}} \begin{proof} If one closely follows the steps as in the verification of the general (convolution) Theorem 3.11.2 in \citet{VanDVaart[2013]} then the only peculiarity to be taken into account is the matrix $\mathcal Z$. More precisely, for an orthonormal basis $h_1,\ldots,h_{d^*},\ d^*:=d(d+1)/2$, of $\text{vec}(\R^{d\times d}_{\text{sym}}):=\{\text{vec}(A):A\in\R^{d\times d}_{\text{sym}}\}$ with respect to the inner product $\langle \cdot,\cdot\rangle_{\mathcal I(\Sigma)\mathcal Z}$, Proposition~\ref{LAN_gen} and Le Cam's Third Lemma yield \begin{equation}\label{LimitDecomp} r^{-1}_n(\hat\vartheta_n-\psi(\Sigma+r_nH))\overset{d}{\to}\mathcal N\Big(0,\sum^{d^*}_{k=1}\nabla\psi_{\Sigma}h_kh^{\top}_k\nabla\psi^{\top}_{\Sigma}\Big)\ast R, \end{equation} under $P^n_{\Sigma+r_nH}$, for some $R$. The independence of $h$ now follows by \begin{align*} &\Big(\sum^{d^*}_{k=1}\nabla\psi_{\Sigma}h_kh^{\top}_k\nabla\psi^{\top}_{\Sigma}\Big)_{i,j}=\sum^{d^*}_{k=1}\langle\nabla\psi_{\Sigma}^{(i)},h_k\rangle\langle\nabla\psi_{\Sigma}^{(j)},h_k\rangle\\ =&\frac{1}{4}\sum^{d^*}_{k=1}\langle\mathcal I^{-1}_{\Sigma}\nabla\psi_{\Sigma}^{(i)},h_k\rangle_{\mathcal I(\Sigma)\mathcal Z}\langle\mathcal I^{-1}_{\Sigma}\nabla\psi_{\Sigma}^{(j)},h_k\rangle_{\mathcal I(\Sigma)\mathcal Z}=\frac{1}{4}\langle\nabla\psi^{(i)}_{\Sigma},\mathcal I^{-1}_{\Sigma}\nabla\psi^{(j)}_{\Sigma}\rangle_{\mathcal Z}, \end{align*} where $\nabla\psi^{(i)}_{\Sigma}$ denotes the $i$-th column of $\nabla\psi^{\top}_{\Sigma}$ and $1\leq i,j\leq d$. \end{proof} \begin{remark}\label{RmkZZZ} Note that the singularity of $\mathcal I(\Sigma)\mathcal Z$ has no critical impact as $\langle\cdot,h_k\rangle_{\mathcal I(\Sigma)\mathcal Z}=2\langle\cdot,h_k\rangle_{\mathcal I(\Sigma)}$ is the essential isometry-type ingredient used. \end{remark} \subsection{\textbf{Estimation}}\label{estimation} For each observation $Y_p$ in \eqref{SeqPar} an unbiased estimator of $\psi(\Sigma)=\text{vec}(\Sigma)$ can be obtained via \[\hat\vartheta_p:=\lambda_p^{-1}\text{vec}\Big(Y_{np}Y^{\top}_{np}-\frac{\eta^2}{n}I_d\Big).\] Since $\hat\vartheta_p,\ p\geq1,$ are independent it is reasonable to consider a weighted average to reduce variability. Let $\pi_n=[\underline \pi_n,\overline\pi_n]\cap\mathbb N$ be as in Theorem~\ref{ThmRateFisher} and set $\mathcal I_J(\Sigma):=\sum_{p\in J}\mathcal I_{np}(\Sigma)$, for $J\subseteq\mathbb N$. Then, by a Lagrange approach, the choice of weights \[W_p(\Sigma):=\mathcal I_{\pi_n}(\Sigma)^{-1}\mathcal I_{np}(\Sigma)\] ensures unbiasedness and minimal covariance of the oracle estimator \[\hat\vartheta^{\text{or}}_n:=\sum_{p\in\pi_n}W_p(\Sigma)\hat\vartheta_p.\] Let $\pi'_n\subsetneq\mathbb N$ be with $\pi_n\cap\pi'_n=\emptyset$, $n\geq1$, and $|\pi'_n|\to\infty$, as $n\to\infty$. Set $\hat\vartheta^{\text{pre}}_n:=\sum_{p\in\pi'_n}W^{\pi'_n}_p(SI_d)\hat\vartheta_p$, where $W^{\pi'_n}_p(\Sigma):=\mathcal I_{\pi'_n}(\Sigma)^{-1}\mathcal I_{np}(\Sigma)$, and set $\hat\Sigma^{\text{pre}}_n:=\text{mat}(\hat\vartheta^{\text{pre}}_n)$, where $\text{mat}:\mathbb R^{d^2}\to\mathbb R^{d\times d}$ is the inverse of $\text{vec}$. Then an adaptive version of $\hat\vartheta^{\text{or}}_n$ is obtained by \begin{equation}\label{adaptest} \hat\vartheta^{\text{ad}}_n:=\sum_{p\in\pi_n}W_p(\hat\Sigma^{\text{pre}}_n)\hat\vartheta_p. \end{equation} Note that it is crucial that $(W_p(\hat\Sigma^{\text{pre}}_n))_{p\in\pi'_n}$ is independent of $(Y_p)_{p\in\pi_n}$. \begin{theorem}\label{ThmOrAd} The estimators $\hat\vartheta^{\text{or}}_n$ and $\hat\vartheta^{\text{ad}}_n$ of $\psi(\Sigma)=\text{vec}(\Sigma)$ are regular and efficient in the sense of Theorem~\ref{ThmConvPara}. In particular, it holds that \[r^{-1}_n(\hat\vartheta^{ad}_n-\psi(\Sigma+r_nH))\overset{d}{\to}\mathcal N(0,\tfrac{1}{4}\mathcal I(\Sigma)^{-1}\mathcal Z),\quad\text{as }n\to\infty,\] under $P^n_{\Sigma+r_nH}$, for any $H\in\mathbb R^{d\times d}_{\text{sym}}$. \end{theorem} \begin{remark} The estimator $\hat\vartheta^{\text{ad}}_n=\hat\vartheta^{\text{ad}}_n((Y_p)_{p\geq1})$ in $\mathcal F^s_n$ can be obtained in the initial model $\mathcal F_n$ by the explicit construction via interpolations given in the proof of Theorem~\ref{ThmLeCamDiscCont}. In particular, for an interpolated version $(\bar Y_t)_{t\in[0,1]}$ of \eqref{DiscGen}, cf. \eqref{contmodel_zwischen}, the estimator $\hat\vartheta^{\text{ad}}_n=\hat\vartheta^{\text{ad}}_n((\bar Y_p)_{p\geq1})$ in $\mathcal F_n$ can be built as in \eqref{adaptest} from \[\bar Y_p:=((e_{pj},\bar Y))_{1\leq j\leq d},\quad p\geq1,\] where $e_{pi}=(\mathbbm{1}\{i=j\}\varphi_p)_{1\leq j\leq d}$ and $\varphi_p$ is the eigenfunction corresponding to $\lambda_p$, cf. Section~\ref{secCtsSeq}. For the limit distribution of $\hat\vartheta^{\text{ad}}_n((\bar Y_p)_{p\geq1})$ note that for $\bar P^n_{\Sigma}:=\mathcal L((\bar Y_p)_{p\geq1})$ and $f$ continuous and bounded it easily can be seen that \[\mathbb E_{\bar P^n_{\Sigma}}[f(\hat\vartheta^{\text{ad}}_n)]=\mathbb E_{P^n_{\Sigma}}[f(\hat\vartheta^{\text{ad}}_n)]+\mathcal O(\|f\|_{\infty}\|P^n_{\Sigma}-\bar P^n_{\Sigma}\|_{\text{TV}})\] where the total variation norm satisfies $\|P^n_{\Sigma}-\bar P^n_{\Sigma}\|_{\text{TV}}\to0$, by the proof of Theorem~\ref{ThmLeCamDiscCont}. In particular, the estimator $\hat\vartheta^{\text{ad}}_n((\bar Y_p)_{p\geq1})$ has the same asymptotic properties as its counterpart constructed in $\mathcal F^s_n$ and it satisfies the statement of Theorem~\ref{ThmOrAd}. \end{remark} \subsection{\textbf{Further asymptotic equivalences}}\label{morecam} The adaptive estimator $\hat\vartheta^{\text{ad}}_n$ in \eqref{adaptest} allows for further asymptotic equivalence statements that completes the asymptotic analysis of the fundamental parametric model $\mathcal F_n$. By Theorem~\ref{ThmRateFisher} the asymptotically significant information for estimating $\Sigma$ efficiently in $\mathcal F^s_n$ is already contained in the subexperiment $\mathcal F^s_{n,\pi_n}$ that is generated by the observations $(Y_p)_{p\in\pi_n}$, where $\pi_n$ is as in Theorem~\ref{ThmRateFisher}, i.e., \[\pi_n=[a_np_n,b_np_n]\cap\mathbb N,\quad a_n\downarrow0,\quad b_n\to\infty.\] Clearly, $\mathcal F^s_n$ is at least as informative as $\mathcal F^s_{n,\pi_n}$, but even the reverse can be shown, at least asymptotically, given that the parameter set $\Theta_0$ is replaced by the more restrictive set (with $S>1$) \begin{equation}\label{sparprime} \Theta_0'=\{\Sigma\in\R^{d\times d}_{\text{sym}}:S^{-1}I_d<\Sigma<SI_d\}. \end{equation} \begin{proposition}\label{PropLeCamProj} For parameter set $\Theta_0'$ in \eqref{sparprime} the experiments $\mathcal F^s_n$ and $\mathcal F^s_{n,\pi_n}$ are asymptotically equivalent in Le Cam's sense. More precisely, \[\Delta(\mathcal F^s_n,\mathcal F^s_{n,\pi_n})=\mathcal O(S\log n)\mathcal O(a^{\delta-1/2}_n\vee b^{1/2-\delta}_n).\] \end{proposition} Proposition~\ref{PropLeCamProj} gives a further intuition on smoothing choices for several known estimation methods such as pre-averaging, where the frequencies of order $\sqrt{n}$ play a central role for models driven by a Brownian motion, cf. \citet{Jacod[2009]}. Next the impact of deviations in the underlying eigenvalue sequence $\lambda$ is investigated. As we have seen in Theorem~\ref{ThmRateFisher}, the leading term of $(\lambda_p)_{p\geq1}$ completely determines the asymptotic lower bounds. As an example consider the cases in which $G$ in \eqref{ContPar} is a Brownian bridge or a Brownian motion. The respective underlying eigenvalue sequences read as \[\lambda^{\text{BB}}_p=(\pi p)^{-2}\quad\text{and}\quad\lambda^{\text{BM}}_p=\pi^{-2}(p-1/2)^{-2},\] respectively, and thus the bounds obtained by Theorem~\ref{ThmConvPara} coincide. In fact, even a general characterisation of asymptotic equivalence on the basis of the underlying eigenvalue sequence can be given. \begin{proposition}\label{PropEigLeCamEqui} For sequences $\lambda$ and $\lambda'$ satisfying Assumption~\ref{regvarass}-$\lambda(\delta)$ (with possibly different $\delta$) let $\mathcal F^s_n$ and $\mathcal F^{s'}_n$, respectively, be sequence space models of type \eqref{SeqPar} on $\Theta_0'$ as in \eqref{sparprime}. Then the following are equivalent: \begin{enumerate} \item $\lambda_p/\lambda'_p\to1$, as $p\to\infty$. \item $r_n/r'_n\to1$, as $n\to\infty$, and $\mathcal I(\Sigma)=\mathcal I'(\Sigma)$, for all $\Sigma\in\Theta_0$. \item $\Delta(\mathcal F^s_n,\mathcal F^{s'}_n)\to0$, as $n\to\infty$, \end{enumerate} where $r'_n$ and $\mathcal I'(\Sigma)\mathcal Z$ are the rate and asymptotic Fisher information in $\mathcal F^{s'}_n$. \end{proposition} \newpage The impact of the leading term of $\lambda$ yields an interesting finding in the particular scenario, in which the signal process is a mixture \[G_t=Z_{1,t}+Z_{2,t},\] of two independent Gaussian processes $Z_i=(Z_{i,t})_{t\in[0,1]},\ i=1,2$. If the covariance operators of $Z_1$ and $Z_2$ are diagonalisable by the same basis then the process with more slowly decaying eigenvalues completely determines the asymptotic properties of the estimation problem. Therefore one might conjecture for $G$ being a so-called mixed fractional Brownian motion of Hurst index $H>1/2$, cf. \citet{Cheridito[2001]}, that solely the Brownian motion part contributes to the underlying asymptotics. \section{Semiparametric efficiency under asynchronicity}\label{SecSemiPara} In the following we suppose that Assumptions~\ref{AssSigma}-$\Sigma(\beta,M,S)$ and \ref{AssRegF}-$F(\gamma,N,\beta)$ hold. \subsection{\textbf{Locally parametric approximation}}\label{SsecPwConAppr} As in the parametric set-up, observing \eqref{DiscSemi} is approximated by its continuous analogue. However, in order to use the parametric results, locally constant approximations of $\Sigma$ and $F=(F_j)_{1\leq j\leq d}$ are considered. More precisely, for $m$ disjoint blocks $\text{I}_{mk}:=[k/m,(k+1)/m),k=0,\ldots,m-1$, and $\Sigma_{m,k}:=\Sigma(k/m)$ introduce \[\Sigma_m:=\sum^{m-1}_{k=0}\Sigma_{m,k}\mathbbm{1}_{\text{I}_{mk}}(\cdot),\quad F'_{j,m}:=\sum^{m-1}_{k=0}F'_j\Big(\frac{k}{m}\Big)\mathbbm{1}_{\text{I}_{mk}}(\cdot),\quad j=1,\ldots,d.\] and the corresponding continuous observation model \begin{equation}\label{ContSemi} dY^m_t=\Big(\int^t_0\Sigma^{1/2}_m(s)dB_s\Big)dt+\Xi_m(t)dW_t,\quad t\in[0,1], \end{equation} where \[\Xi^2_m:=\text{diag}(\eta^2_j/(n_jF'_{j,m}))_{1\leq j\leq d}.\] \begin{definition}\label{DefExpMCm} For $n:=(n_1,\ldots,n_d)$ let $\mathcal M_n$ and $\mathcal M^c_n$ be the statistical experiments that are generated by the discrete and continuous observations \eqref{DiscSemi} and \eqref{ContSemi}, respectively. \end{definition} Let $\lambda_{mp}:=(\pi pm)^{-2}$, $p\geq1$, and set $n_{\max}:=\max_{1\leq j\leq d}n_j$. The Le Cam distance between $\mathcal M_n$ and $\mathcal M^c_n$ is bounded by the approximation errors of $\Sigma_m$ and $F'_{j,m}$. As $m$ will have to be chosen later in this section such that $m=o(\sqrt{n_{\min}})$, the restriction $\beta>1/2$ is evident in view of the following. \begin{proposition}\label{PropPwc} For any $\kappa\in(0,1/2)$ and $n_{\min}\to\infty$ it holds that \[\Delta(\mathcal M_n,\mathcal M^c_n)=\mathcal O(MSn_{\max}n^{-3/2+\kappa}_{\min})+\mathcal O(MSn^{1/4}_{\max}m^{-\beta}).\] In particular, asymptotic equivalence holds, given that $m=o(\sqrt{n_{\min}})$. \end{proposition} \subsection{\textbf{LAN for correlated and uncorrelated sequence space models}} As described in Section~\ref{secCtsSeq} a continuous experiment can be represented in the sequence space. To this end, consider the (normalised) $L^2([0,1],\mathbb R)$-basis \begin{align*} \varphi_{0,0}(t)&:=\sqrt{m}\mathbbm{1}_{\text{I}_{m,0}}(t),\\ \varphi_{0,k+1}(t)&:=\frac{\sqrt{m}}{\sqrt{2}}(\mathbbm{1}_{\text{I}_{mk}}(t)-\mathbbm{1}_{\text{I}_{m,k+1}}(t)),\quad k=0,\ldots,m-2,\\ \varphi_{pk}(t)&:=\sqrt{2m}\cos(p\pi(tm-k))\mathbbm{1}_{\text{I}_{mk}}(t),\quad p\geq1,\quad k=0,\ldots,m-1. \end{align*} Via $e_{pki}=(\mathbbm{1}_{\{i=j\}}\varphi_{pk})_{1\leq j\leq d}$ Gaussian random vectors \begin{equation}\label{SeqSemiTwo} S_{pk}:=((e_{pki},dY^m))_{1\leq i\leq d},\quad p\geq0,\ k=0,\ldots,m-1 \end{equation} are obtained, cf. \eqref{cylmeas}. Clearly observing the correlated vectors $(S_{pk})_{p\geq 0,k=0,\ldots,m-1}$ is equivalent to observing \eqref{ContSemi} and more informative than observing $(S_{pk})_{p\geq 1,k=0,\ldots,m-1}$. However, the latter sequence is independent and close to observing \eqref{SeqSemi}, hence it is similar to experiment $\mathcal F^s_n$ which has been intensively studied in Section~\ref{SecParametric}. \begin{proposition}\label{PropLANequi} Let $m=o(\sqrt{n_{\min}})$ be satisfied. Then any LAN-expansion with respect to $\Sigma+n^{-1/4}_{\min}H,\ H\inH^{\beta}_{\text{sym}}$ for the model \eqref{SeqSemi} is also valid in $\mathcal M^c_n$ and $\mathcal M_n$. \end{proposition} \subsection{\textbf{Verification of Theorem~\ref{ThmConvNonpara}}} \begin{proof} The score induced by \eqref{SeqSemi} equals $\nabla\ell_{n}(\Sigma):=\text{vec}(\nabla\ell^{(0)}_n(\Sigma),\ldots,\nabla\ell^{(m-1)}_n(\Sigma))$, where $\nabla\ell^{(k)}_n(\Sigma)$ is of the exact same shape as the parametric score in \eqref{SeqScore} with $\lambda_{mp},\ C_{pk}$ and $Y_{pk}$ replacing $\lambda_p,\ C_p$ and $Y_p$, respectively. Therefore the (not $\mathcal Z$-normalised) Fisher information in $\mathcal M^s_n$ is given by the block diagonal matrix \[\mathcal I_{n,m}(\Sigma):=\begin{pmatrix}\mathcal I_n^{(0)}(\Sigma)&0&\cdots&0\\0&\mathcal I_n^{(1)}(\Sigma)&\cdots&0\\\vdots&&\ddots&\vdots\\0&\cdots&\cdots&\mathcal I_n^{(m-1)}(\Sigma)\end{pmatrix},\] with blocks \[\mathcal I_n^{(k)}(\Sigma):=\frac{1}{4}\sum^{\infty}_{p=1}\lambda^2_{mp}(C^{-1}_{pk}\otimes C^{-1}_{pk}),\ k=0,\ldots,m-1.\] As in Theorem~\ref{ThmRateFisher}, regular variation of the eigenvalues $\lambda$ yields that on each block $\text{I}_{mk}$ the Fisher information grows with rate $\sqrt{n_{\min}}/m$ such that \begin{equation}\label{FisherRiemann} n^{-1/2}_{\min}\sum^{m-1}_{k=0}\mathcal I_n^{(k)}(\Sigma)=\frac{1}{m}\sum^{m-1}_{k=0}\mathcal I_{\Sigma}(k/m)+o(1)\to\int^1_0\mathcal I_{\Sigma}(t)dt, \end{equation} i.e., the rate is $n^{-1/4}_{\min}$, where (cf. proof of Theorem~\ref{ThmRateFisher} and Remark~\ref{RmkFisherExp}) \[\mathcal I_{\Sigma}(t)=\frac{1}{8}(\Sigma^{1/2}_{\Xi}(t)\otimes\Sigma(t)+\Sigma(t)\otimes \Sigma^{1/2}_{\Xi}(t))^{-1},\ t\in[0,1].\] For $H\inH^{\beta}_{\text{sym}}$ - (as before) in the sense that $\Sigma+n^{-1/4}_{\min}H\in\Theta_1$, for $n$ sufficiently large - note that \eqref{LANrem1} and \eqref{LANrem2} hold uniformly in $H_{m,k}:=H(k/m),\ k=0,\ldots,m-1$. Thus applying Proposition~\ref{LAN_gen} simultaneously leads to (denoting by $Q^n_{\Sigma}$ the measure induced by \eqref{SeqSemi}) \begin{align*} \log\frac{dQ^n_{\Sigma+n^{-1/4}_{\min}H}}{dQ^n_{\Sigma}}=\sum^{m-1}_{k=0}\Big(&\text{vec}(H_{m,k})^{\top}\nabla\ell^{(k)}_n(\Sigma)-\frac{1}{2\sqrt{n_{\min}}}\|H_{m,k}\|^2_{\mathcal I_n^{(k)}(\Sigma)\mathcal Z,L^2}\\ &+\rho^{(1,k)}_n+\rho^{(2,k)}_n\Big), \end{align*} where \eqref{FisherRiemann} implies $n^{-1/2}_{\min}\sum^{m-1}_{k=0}\|H_{m,k}\|^2_{\mathcal I_n^{(k)}(\Sigma)\mathcal Z,L^2}\to\|H\|^2_{\mathcal I_{\Sigma}\mathcal Z,L^2}$ with $\langle H,H\rangle_{\mathcal I_{\Sigma}\mathcal Z,L^2}:=\int^1_0\langle H(t),H(t)\rangle_{\mathcal I_{\Sigma}(t)\mathcal Z}dt$ (similarly for $\|\cdot\|_{\mathcal I_n^{(k)}(\Sigma)\mathcal Z,L^2}$). Moreover, for $k=0,\ldots,m-1$, \eqref{LANrem1} and \eqref{LANrem2} imply $\mathbb E[\rho^{(1,k)}_n]=0$ as well as \[\text{Var}\Big(\sum^{m-1}_{k=0}\rho^{(1,k)}_n\Big)=\mathcal O(n^{-1/2}_{\min}),\quad\sum^{m-1}_{k=0}|\rho^{(2)}_{n,k}|=\mathcal O(n^{-1/4}_{\min}).\] Since a central limit theorem applies for $n^{-1/4}_{\min}\sum^m_{k=1}\text{vec}(H_k)\nabla\ell^{(k)}_n(\Sigma)$ analogously as in Theorem~\ref{LAN} the sequence of experiments $\mathcal M^s_n$ satisfies \begin{equation}\label{LANNonpara} \log\frac{dP^{n,m}_{\Sigma+n^{-1/4}_{\min}H}}{dP^{n,m}_{\Sigma}}=\Delta_{n,\Sigma,H}+\frac{1}{2}\|H\|^2_{\mathcal I_{\Sigma}\mathcal Z,L^2},\quad H\inH^{\beta}_{\text{sym}}, \end{equation} where $\Delta_{n,\Sigma,H}\overset{d}{\to}\Delta_{\Sigma,H}$, under $Q^n_{\Sigma}$, with $\Delta_{\Sigma,H}$ being the centred Gaussian process with $\text{Cov}(\Delta_{\Sigma,H_1},\Delta_{\Sigma,H_2})=\langle H_1,H_2\rangle_{\mathcal I_{\Sigma}\mathcal Z,L^2}$. In order to establish a convolution theorem, the verification of Theorem 3.11.2 in \citet{VanDVaart[2013]} is once more closely followed. First denote for the asymptotic perturbation error by \[\dot\kappa(H):=\int^1_0(\nabla W_{\Sigma}\text{vec}(H))(t)dt=\lim_{n_{\min}\to\infty}n^{1/4}_{\min}(\psi(\Sigma+n^{-1/4}_{\min}H)-\psi(\Sigma)).\] For $U\geq1$ let $L_U$ be a $U$-dimensional subspace of $H^{\beta}_{\text{sym}}$ and let $H_1,\ldots,H_U$ be an orthonormal basis of $L_U$ with respect to $\langle\cdot,\cdot\rangle_{\mathcal I_{\Sigma}\mathcal Z,L^2}$. Denote by $\dot W^{(i)}_{\Sigma}$ the $i$-th column of $(\nabla W_{\Sigma})^{\top}$ and let $h_u:=\text{vec}(H_u),\ u=1\ldots,U$. Then \eqref{LANNonpara} and Le Cam's third Lemma yield that the limit distribution of regular estimators under $Q^n_{\Sigma+n^{-1/4}_{\min}H}$, $H\in L_U$, is a convolution of some $R$ with $\mathcal N(0,\sum^U_{u=1}\dot\kappa(H_u)\dot\kappa(H_u)^{\top})$, cf. \eqref{LimitDecomp}. Thus the $(i,j)$-entry of the optimal asymptotic covariance of estimating $\psi(\Sigma+n^{-1/4}_{\min}H)$ is obtained by a limiting argument and (once more) by the properties of $\mathcal Z$ via \begin{align*} &\lim_{U\to\infty}\sum^U_{u=1}(\dot\kappa(H_u)\dot\kappa(H_u)^{\top})_{i,j}=\lim_{U\to\infty}\sum^U_{u=1}\langle\dot W_{\Sigma}^{(i)},h_u\rangle_{L^2}\langle\dot W_{\Sigma}^{(j)},h_u\rangle_{L^2}\\ =&\lim_{U\to\infty}\frac{1}{4}\sum^U_{u=1}\langle\mathcal I^{-1}_{\Sigma}\dot W_{\Sigma}^{(i)},h_u\rangle_{\mathcal I_{\Sigma}\mathcal Z,L^2}\langle\mathcal I^{-1}_{\Sigma}\dot W_{\Sigma}^{(j)},h_u\rangle_{\mathcal I_{\Sigma}\mathcal Z,L^2}\\ =&\frac{1}{4}\langle\mathcal I^{-1}_{\Sigma}\dot W_{\Sigma}^{(i)},\mathcal I^{-1}_{\Sigma}\dot W_{\Sigma}^{(j)}\rangle_{\mathcal I_{\Sigma}\mathcal Z,L^2}. \end{align*} \end{proof}

a78c16d14272e6cb785c677e7e4c9ca127628517

\section{Introduction} One of the most fundamental topics in natural language processing is how best to derive high-level representations from constituent parts, as natural language meanings are a function of their constituent parts. How best to construct a sentence representation from distributed word embeddings is an example domain of this larger issue. Even though sequential neural models such as recurrent neural networks (RNN) \cite{elman1990finding} and their variants including Long Short-Term Memory (LSTM) \cite{hochreiter1997long} and Gated Recurrent Unit (GRU) \cite{cho2014learning} have become the de-facto standard for condensing sentence-level information from a sequence of words into a fixed vector, there have been many lines of research towards better sentence representation using other neural architectures, e.g. convolutional neural networks (CNN) \cite{kim2014convolutional} or self-attention based models \cite{shen2018reinforced}. From a linguistic point of view, the underlying tree structure---as expressed by its constituency and dependency trees---of a sentence is an integral part of its meaning. Inspired by this fact, some recursive neural network (RvNN\footnote{To avoid confusion, we call recursive neural networks (or tree-structured NNs) RvNNs to distinguish them from recurrent neural networks RNNs, following the convention of some previous works.}) models are designed to reflect the syntactic tree structure, achieving impressive results on several sentence-level tasks such as sentiment analysis \cite{socher2012semantic,socher2013recursive}, machine translation \cite{yang2017towards}, natural language inference \cite{bowman2016fast}, and discourse relation classification \cite{wang2017tag}. However, some recent works have \cite{yogatama2017learning,choi2018learning} proposed latent tree models, which learn to construct task-specific tree structures without explicit supervision, bringing into question the value of linguistically-motivated recursive neural models. Witnessing the surprising performance of the latent tree models on some sentence-level tasks, there arises a natural question: \textit{Are linguistic tree structures the optimal way of composing sentence representations for NLP tasks?} In this paper, we demonstrate that linguistic priors are in fact useful for devising effective neural models for sentence representations, showing that our novel architecture based on constituency trees and their tag\footnote{In this work, we refer to both part-of-speech (POS) tags (e.g. DT-determiner, JJ-adjective) for words and phrase-level tags (e.g. NP-noun phrase, VP-verb phrase) simply as `tags'.} information obtains superior performance on several sentence-level tasks, including sentiment analysis and natural language inference. A chief novelty of our approach is that we introduce a small separate tag-level tree-LSTM to control the composition function of the existing word-level tree-LSTM, which is in charge of extracting helpful syntactic signals for meaningful semantic composition of constituents by considering both the structures and linguistic tags of constituency trees simultaneously. In addition, we demonstrate that applying a typical LSTM to preprocess the leaf nodes of a tree-LSTM greatly improves the performance of the tree models. Moreover, we propose a clustered tag set to replace the existing tags on the assumption that the original syntactic tags are too fined-grained to be useful in neural models. In short, our contributions in this work are as follows: \begin{itemize} \item We propose a new linguistically-motivated neural model which generates high-quality sentence representations by considering all the information extracted from constituency parse trees. \item In addition, we demonstrate the superiority of the proposed models achieving new state-of-the-art performance within the similar model class on 4 out of 5 sentence classification benchmarks, as well as showing competitive results compared to other types of neural models. \item We empirically show that another key point to the success of tree-structured models is to contextualize input word embeddings so that the corresponding input for each word in a sentence can better reflect the meaning of the whole sentence. \end{itemize} \section{Related Work} Recursive neural networks (RvNN) are a kind of neural architecture which model sentences by exploiting syntactic structure. While earlier RvNN models proposed utilizing diverse composition functions, including feed-forward neural networks \cite{socher2011parsing}, matrix-vector multiplication \cite{socher2012semantic}, and tensor computation \cite{socher2013recursive}, tree-LSTMs \cite{tai2015improved} remain the standard for several sentence-level tasks. Even though classic RvNNs have demonstrated superior performance on a variety of tasks, their inflexibility, i.e. their inability to handle \textit{dynamic compositionality} for different syntactic configurations, is a considerable weakness. For instance, it would be desirable if our model could distinguish e.g. adjective-noun composition from that of verb-noun or preposition-noun composition, as models failing to make such a distinction ignore real-world syntactic considerations such as `-arity' of function words (i.e. types), and the adjunct/argument distinction. To enable dynamic compositionality in recursive neural networks, many previous works \cite{hashimoto2013simple,dong2014adaptive,qian2015learning,wang2017tag,liu2017dynamic,huang2017encoding,teng2017head} have proposed various methods. One main direction of research leverages tag information, which is produced as a by-product of parsing. In detail, \citeauthor{qian2015learning} (\citeyear{qian2015learning}) suggested TG-RNN, a model employing different composition functions according to POS tags, and TE-RNN/TE-RNTN, models which leverage tag embeddings as additional inputs for the existing tree-structured models. Despite the novelty of utilizing tag information, the explosion of the number of parameters (in case of the TG-RNN) and the limited performance of the original models (in case of the TE-RNN/TE-RNTN) have prevented these models from being widely adopted. Meanwhile, \citeauthor{wang2017tag} (\citeyear{wang2017tag}) and \citeauthor{huang2017encoding} (\citeyear{huang2017encoding}) proposed models based on a tree-LSTM which also uses the tag vectors to control the gate functions of the tree-LSTM. In spite of their impressive results, there is a limitation that the trained tag embeddings are too simple to reflect the rich information which tags provide in different syntactic structures. To alleviate this problem, we introduce structure-aware tag representations in the next section. Another way of building dynamic compositionality into RvNNs is to take advantage of a meta-network (or hyper-network). Inspired by recent works on dynamic parameter prediction, DC-TreeLSTMs \cite{liu2017dynamic} dynamically create the parameters for compositional functions in a tree-LSTM. Specifically, the model has two separate tree-LSTM networks whose architectures are similar, but the smaller of the two is utilized to calculate the weights of the bigger one. A possible problem for this model is that it may be easy to be trained such that the role of each tree-LSTM is ambiguous, as they share the same input, i.e. word information. Therefore, we design two disentangled tree-LSTMs in our model so that one focuses on extracting useful features from only syntactic information while the other composes semantic units with the aid of the features. Furthermore, our model reduces the complexity of computation by utilizing typical tree-LSTM frameworks instead of computing the weights for each example. Finally, some recent works \cite{yogatama2017learning,choi2018learning} have proposed latent tree-structured models that learn how to formulate tree structures from only sequences of tokens, without the aid of syntactic trees or linguistic information. The latent tree models have the advantage of being able to find the optimized task-specific order of composition rather than a sequential or syntactic one. In experiments, we compare our model with not only syntactic tree-based models but also latent tree models, demonstrating that modeling with explicit linguistic knowledge can be an attractive option. \section{Model} In this section, we introduce a novel RvNN architecture, called \textbf{SATA Tree-LSTM}\footnote{The implementation of our model and supplemental materials are available at https://github.com/galsang/SATA-Tree-LSTM.} (\textbf{S}tructure-\textbf{A}ware \textbf{T}ag \textbf{A}ugmented \textbf{Tree-LSTM}). This model is similar to typical Tree-LSTMs, but provides dynamic compositionality by augmenting a separate tag-level tree-LSTM which produces structure-aware tag representations for each node in a tree. In other words, our model has two independent tree-structured modules based on the same constituency tree, one of which (word-level tree-LSTM) is responsible for constructing sentence representations given a sequence of words as usual, while the other (tag-level tree-LSTM) provides supplementary syntactic information to the former. In section 3.1, we first review tree-LSTM architectures. Then in section 3.2, we introduce a tag-level tree-LSTM and structure-aware tag representations. In section 3.3, we discuss an additional technique to boost the performance of tree-structured models, and in section 3.4, we describe the entire architecture of our model in detail. \subsection{Tree-LSTM} The LSTM \cite{hochreiter1997long} architecture was first introduced as an extension of the RNN architecture to mitigate the vanishing and exploding gradient problems. In addition, several works have discovered that applying the LSTM cell into tree structures can be an effective means of modeling sentence representations. To be formal, the composition function of the cell in a tree-LSTM can be formulated as follows: \begin{equation} \label{eq:1} \begin{bmatrix} \mathbf{i} \\ \mathbf{f}_l \\ \mathbf{f}_r \\ \mathbf{o} \\ \mathbf{g} \end{bmatrix} = \begin{bmatrix} \sigma \\ \sigma \\ \sigma \\ \sigma \\ \tanh \end{bmatrix} \Bigg( \mathbf{W} \begin{bmatrix} \mathbf{h}_l \\ \mathbf{h}_r \\ \end{bmatrix} + \mathbf{b} \Bigg) \end{equation} \begin{equation} \label{eq:2} \mathbf{c} = \mathbf{f}_l \odot \mathbf{c}_l + \mathbf{f}_r \odot \mathbf{c}_r + \mathbf{i} \odot \mathbf{g}\\ \end{equation} \begin{equation} \label{eq:3} \mathbf{h} = \mathbf{o} \odot \tanh{\left(\mathbf{c}\right)} \end{equation} \noindent where $\mathbf{h}, \mathbf{c} \in\mathbb{R}^{d}$ indicate the hidden state and cell state of the LSTM cell, and $\mathbf{h}_l, \mathbf{h}_r, \mathbf{c}_l, \mathbf{c}_r \in\mathbb{R}^{d}$ the hidden states and cell states of a left and right child. $\mathbf{g} \in\mathbb{R}^{d}$ is the newly composed input for the cell and $\mathbf{i}, \mathbf{f}_{l}, \mathbf{f}_{r}, \mathbf{o} \in\mathbb{R}^{d}$ represent an input gate, two forget gates (left, right), and an output gate respectively. $\mathbf{W} \in\mathbb{R}^{5d\times2d}$ and $\mathbf{b} \in\mathbb{R}^{5d}$ are trainable parameters. $\sigma$ corresponds to the sigmoid function, $\tanh$ to the hyperbolic tangent, and $\odot$ to element-wise multiplication. Note the equations assume that there are only two children for each node, i.e. binary or binarized trees, following the standard in the literature. While RvNN models can be constructed on any tree structure, in this work we only consider constituency trees as inputs. In spite of the obvious upside that recursive models have in being so flexible, they are known for being difficult to fully utilize with batch computations as compared to other neural architectures because of the diversity of structure found across sentences. To alleviate this problem, \citeauthor{bowman2016fast} (\citeyear{bowman2016fast}) proposed the SPINN model, which brings a shift-reduce algorithm to the tree-LSTM. As SPINN simplifies the process of constructing a tree into only two operations, i.e. shift and reduce, it can support more effective parallel computations while enjoying the advantages of tree structures. For efficiency, our model also starts from our own SPINN re-implementation, whose function is exactly the same as that of the tree-LSTM. \subsection{Structure-aware Tag Representation} In most previous works using linguistic tag information \cite{qian2015learning,wang2017tag,huang2017encoding}, tags are usually represented as simple low-dimensional dense vectors, similar to word embeddings. This approach seems reasonable in the case of POS tags that are attached to the corresponding words, but phrase-level constituent tags (e.g. NP, VP, ADJP) vary greatly in size and shape, making them less amenable to uniform treatment. For instance, even the same phrase tags within different syntactic contexts can vary greatly in size and internal structure, as the case of NP tags in Figure \ref{fig:figure1} shows. Here, the NP consisting of DT[the]-NN[stories] has a different internal structure than the NP consisting of NP[the film 's]-NNS[shortcomings]. One way of deriving \textit{structure-aware} tag representations from the original tag embeddings is to introduce a separate tag-level tree-LSTM which accepts the typical tag embeddings at each node of a tree and outputs the computed structure-aware tag representations for the nodes. Note that the module concentrates on extracting useful syntactic features by considering only the tags and structures of the trees, excluding word information. \begin{figure}[!t] \centering \includegraphics[width=0.9\columnwidth]{figure1.pdf} \caption{A constituency tree example from Stanford Sentiment Treebank.} \label{fig:figure1} \end{figure} Formally, we denote a tag embedding for the tag attached to each node in a tree as $\textbf{e} \in\mathbb{R}^{d_\text{T}}$. Then, the function of each cell in the tag tree-LSTM is defined in the following way. Leaf nodes are defined by the following: \begin{equation} \label{eq:4} \begin{bmatrix} \hat{\mathbf{c}} \\ \hat{\mathbf{h}} \\ \end{bmatrix} = \tanh{\left(\mathbf{U}_\text{T} \mathbf{e} + \mathbf{a}_\text{T}\right)} \end{equation} \noindent while non-leaf nodes are defined by the following: \begin{equation} \label{eq:5} \begin{bmatrix} \hat{\mathbf{i}} \\ \hat{\mathbf{f}}_l \\ \hat{\mathbf{f}}_r \\ \hat{\mathbf{o}} \\ \hat{\mathbf{g}} \end{bmatrix} = \begin{bmatrix} \sigma \\ \sigma \\ \sigma \\ \sigma \\ \tanh \end{bmatrix} \Bigg( \mathbf{W_\text{T}} \begin{bmatrix} \hat{\mathbf{h}}_l \\ \hat{\mathbf{h}}_r \\ \mathbf{e} \\ \end{bmatrix} + \mathbf{b}_\text{T} \Bigg) \end{equation} \begin{equation} \label{eq:6} \hat{\mathbf{c}} = \hat{\mathbf{f}}_l \odot \hat{\mathbf{c}}_l + \hat{\mathbf{f}}_r \odot \hat{\mathbf{c}}_r + \hat{\mathbf{i}} \odot \hat{\mathbf{g}}\\ \end{equation} \begin{equation} \label{eq:7} \hat{\mathbf{h}} = \hat{\mathbf{o}} \odot \tanh{\left(\hat{\mathbf{c}}\right)} \end{equation} \noindent where $\hat{\mathbf{h}}, \hat{\mathbf{c}} \in\mathbb{R}^{d_\text{T}}$ represent the hidden state and cell state of each node in the tag tree-LSTM. We regard the hidden state ($\hat{\mathbf{h}}$) as a structure-aware tag representation for the node. $ \mathbf{U}_\text{T} \in\mathbb{R}^{2d_\text{T} \times d_\text{T}}, \textbf{a}_\text{T} \in\mathbb{R}^{2d_\text{T}}, \mathbf{W}_\text{T} \in\mathbb{R}^{5d_\text{T} \times 3d_\text{T}}$, and $\mathbf{b}_\text{T} \in\mathbb{R}^{5d_\text{T}}$ are trainable parameters. The rest of the notation follows equations \ref{eq:1}, \ref{eq:2}, and \ref{eq:3}. In case of leaf nodes, the states are computed by a simple non-linear transformation. Meanwhile, the composition function in a non-leaf node absorbs the tag embedding ($\mathbf{e}$) as an additional input as well as the hidden states of the two children nodes. The benefit of revising tag representations according to the internal structure is that the derived embedding is a function of the corresponding makeup of the node, rather than a monolithic, categorical tag. With regard to the tags themselves, we conjecture that the taxonomy of the tags currently in use in many NLP systems is too complex to be utilized effectively in deep neural models, considering the specificity of many tag sets and the limited amount of data with which to train. Thus, we cluster POS (word-level) tags into 12 groups following the universal POS tagset \cite{petrov2012universal} and phrase-level tags into 11 groups according to criteria analogous to the case of words, resulting in 23 tag categories in total. In this work, we use the revised coarse-grained tags instead of the original ones. For more details, we refer readers to the supplemental materials. \subsection{Leaf-LSTM} An inherent shortcoming of RvNNs relative to sequential models is that each intermediate representation in a tree is unaware of its external context until all the information is gathered together at the root node. In other words, each composition process is prone to be locally optimized rather than globally optimized. To mitigate this problem, we propose using a leaf-LSTM following the convention of some previous works \cite{eriguchi2016tree,yang2017towards,choi2018learning}, which is a typical LSTM that accepts a sequence of words in order. Instead of leveraging word embeddings directly, we can use each hidden state and cell state of the leaf-LSTM as input tokens for leaf nodes in a tree-LSTM, anticipating the proper contextualization of the input sequence. Formally, we denote a sequence of words in an input sentence as $w_{1:n}$ ($n$: the length of the sentence), and the corresponding word embeddings as $\mathbf{x}_{1:n}$. Then, the operation of the leaf-LSTM at time $t$ can be formulated as, \begin{equation} \label{eq:8} \begin{bmatrix} \tilde{\mathbf{i}} \\ \tilde{\mathbf{f}} \\ \tilde{\mathbf{o}} \\ \tilde{\mathbf{g}} \end{bmatrix} = \begin{bmatrix} \sigma \\ \sigma \\ \sigma \\ \tanh \end{bmatrix} \Bigg( \mathbf{W}_\text{L} \begin{bmatrix} \tilde{\mathbf{h}}_{t-1} \\ \mathbf{x}_t \\ \end{bmatrix} + \mathbf{b}_\text{L} \Bigg) \end{equation} \begin{equation} \label{eq:9} \tilde{\mathbf{c}}_t = \tilde{\mathbf{f}} \odot \tilde{\mathbf{c}}_{t-1} + \tilde{\mathbf{i}} \odot \tilde{\mathbf{g}}\\ \end{equation} \begin{equation} \label{eq:10} \tilde{\mathbf{h}}_t = \tilde{\mathbf{o}} \odot \tanh{\left(\tilde{\mathbf{c}}_t\right)} \end{equation} \noindent where $\mathbf{x}_t \in\mathbb{R}^{d_w}$ indicates an input word vector and $\tilde{\mathbf{h}}_t$, $\tilde{\mathbf{c}}_t \in\mathbb{R}^{d_h}$ represent the hidden and cell state of the LSTM at time $t$ ($\tilde{\mathbf{h}}_{t-1}$ corresponds to the hidden state at time $t$-1). $\mathbf{W}_\text{L}$ and $\mathbf{b}_\text{L} $ are learnable parameters. The remaining notation follows that of the tree-LSTM above. In experiments, we demonstrate that introducing a leaf-LSTM fares better at processing the input words of a tree-LSTM compared to using a feed-forward neural network. We also explore the possibility of its bidirectional setting in ablation study. \subsection{SATA Tree-LSTM} \begin{figure}[!t] \centering \includegraphics[width=1\columnwidth]{figure2.pdf} \caption{A diagram of SATA Tree-LSTM. The model has two separate tree-LSTM modules, the right of which (tag tree-LSTM) extracts a structure-aware tag representation to control the composition function of the remaining tree-LSTM (word tree-LSTM). Fully-connected: one-layered non-linear transformation.} \label{fig:figure2} \end{figure} In this section, we define \textbf{SATA Tree-LSTM} (\textbf{S}tructure-\textbf{A}ware \textbf{T}ag \textbf{A}ugmented \textbf{Tree-LSTM}, see Figure \ref{fig:figure2}) which joins a tag-level tree-LSTM (section 3.2), a leaf-LSTM (section 3.3), and the original word tree-LSTM together. As above we denote a sequence of words in an input sentence as $w_{1:n}$ and the corresponding word embeddings as $\mathbf{x}_{1:n}$. In addition, a tag embedding for the tag attached to each node in a tree is denoted by $\textbf{e} \in\mathbb{R}^{d_\text{T}}$. Then, we derive the final sentence representation for the input sentence with our model in two steps. First, we compute structure-aware tag representations ($\hat{\mathbf{h}}$) for each node of a tree using the tag tree-LSTM (the right side of Figure \ref{fig:figure2}) as follows: \begin{equation} \label{eq:11} \begin{bmatrix} \hat{\mathbf{c}} \\ \hat{\mathbf{h}} \\ \end{bmatrix} = \begin{cases} \text{Tag-Tree-LSTM}(\mathbf{e}) & \text{if a leaf node} \\ \text{Tag-Tree-LSTM}(\hat{\mathbf{h}}_l, \hat{\mathbf{h}}_r, \mathbf{e}) & \text{otherwise} \end{cases} \end{equation} \noindent where Tag-Tree-LSTM indicates the module we described in section 3.2. Second, we combine semantic units recursively on the word tree-LSTM in a bottom-up fashion. For leaf nodes, we leverage the Leaf-LSTM (the bottom-left of Figure \ref{fig:figure2}, explained in section 3.3) to compute $\tilde{\mathbf{c}}_{t}$ and $\tilde{\mathbf{h}}_{t}$ in sequential order, with the corresponding input $\mathbf{x}_t$. \begin{equation} \label{eq:12} \begin{bmatrix} \tilde{\mathbf{c}}_{t} \\ \tilde{\mathbf{h}}_{t} \\ \end{bmatrix} = \text{Leaf-LSTM}(\tilde{\textbf{h}}_{t-1}, \textbf{x}_t) \end{equation} \noindent Then, the $\tilde{\mathbf{c}}_{t}$ and $\tilde{\mathbf{h}}_{t}$ can be utilized as input tokens to the word tree-LSTM, with the left (right) child of the target node corresponding to the $t$th word in the input sentence. \begin{equation} \label{eq:13} \begin{bmatrix} \check{\textbf{c}}_{\{l, r\}} \\ \check{\textbf{h}}_{\{l, r\}} \end{bmatrix} = \begin{bmatrix} \tilde{\textbf{c}}_{t} \\ \tilde{\textbf{h}}_{t} \end{bmatrix} \end{equation} In the non-leaf node case, we calculate phrase representations for each node in the word tree-LSTM (the upper-left of Figure \ref{fig:figure2}) recursively as follows: \begin{equation} \label{eq:14} \check{\mathbf{g}} = \tanh{\left( \mathbf{U}_\text{w} \begin{bmatrix} \check{\mathbf{h}}_l \\ \check{\mathbf{h}}_r \\ \end{bmatrix} + \mathbf{a}_\text{w} \right)} \end{equation} \begin{equation} \label{eq:15} \begin{bmatrix} \check{\mathbf{i}} \\ \check{\mathbf{f}}_l \\ \check{\mathbf{f}}_r \\ \check{\mathbf{o}} \end{bmatrix} = \begin{bmatrix} \sigma \\ \sigma \\ \sigma \\ \sigma \end{bmatrix} \Bigg( \mathbf{W_\text{w}} \begin{bmatrix} \check{\mathbf{h}}_l \\ \check{\mathbf{h}}_r \\ \hat{\mathbf{h}} \\ \end{bmatrix} + \mathbf{b}_\text{w} \Bigg) \end{equation} \begin{equation} \label{eq:16} \check{\mathbf{c}} = \check{\mathbf{f}}_l \odot \check{\mathbf{c}}_l + \check{\mathbf{f}}_r \odot \check{\mathbf{c}}_r + \check{\mathbf{i}} \odot \check{\mathbf{g}} \end{equation} \begin{equation} \label{eq:17} \check{\mathbf{h}} = \check{\mathbf{o}} \odot \tanh{\left(\check{\mathbf{c}}\right)} \end{equation} \noindent where $\check{\mathbf{h}}$, $\check{\mathbf{c}} \in \mathbb{R}^{d_h}$ represent the hidden and cell state of each node in the word tree-LSTM. $\mathbf{U}_\text{w} \in \mathbb{R}^{d_h \times 2d_h}$, $\mathbf{W}_\text{w} \in \mathbb{R}^{4d_h \times \left(2d_h+d_\text{T}\right)}$, $\mathbf{a}_\text{w} \in \mathbb{R}^{d_h}$, $\mathbf{b}_\text{w} \in \mathbb{R}^{4d_h}$ are learned parameters. The remaining notation follows those of the previous sections. Note that the structure-aware tag representations ($\hat{\mathbf{h}}$) are only utilized to control the gate functions of the word tree-LSTM in the form of additional inputs, and are not involved in the semantic composition ($\check{\mathbf{g}}$) directly. Finally, the hidden state of the root node ($\check{\mathbf{h}}_\text{root}$) in the word-level tree-LSTM becomes the final sentence representation of the input sentence. \section{Experiment and Discussion} \subsection{Quantitative Analysis} \subsubsection{Sentence classification tasks} One of the most basic approaches to evaluate a sentence encoder is to measure the classification performance with the sentence representations made by the encoder. Thus, we conduct experiments on the following five datasets. (Summary statistics for the datasets are reported in the supplemental materials.) \begin{itemize} \item \textbf{MR}: A group of movie reviews with binary (positive / negative) classes. \cite{pang2005seeing} \item \textbf{SST-2}: Stanford Sentiment Treebank \cite{socher2013recursive}. Similar to MR, but each review is provided in the form of a binary parse tree whose nodes are annotated with numeric sentiment values. For SST-2, we only consider binary (positive / negative) classes. \item \textbf{SST-5}: Identical to SST-2, but the reviews are grouped into fine-grained (very negative, negative, neutral, positive, very positive) classes. \item \textbf{SUBJ}: Sentences grouped as being either subjective or objective (binary classes). \cite{pang2004sentimental} \item \textbf{TREC}: A dataset which groups questions into six different question types (classes). \cite{li2002learning} \end{itemize} As a preprocessing step, we construct parse trees for the sentences in the datasets using the Stanford PCFG parser \cite{klein2003accurate}. Because syntactic tags are by-products of constituency parsing, we do not need further preprocessing. To classify the sentence given our sentence representation ($\check{\mathbf{h}}_\text{root}$), we use one fully-connected layer with a ReLU activation, followed by a softmax classifier. The final predicted probability distribution of the class $y$ given the sentence $w_{1:n}$ is defined as follows, \begin{equation} \mathbf{s} = \text{ReLU}(\mathbf{W}_\text{s} \check{\mathbf{h}}_\text{root}+ \mathbf{b}_\text{s}) \end{equation} \begin{equation} p(y|w_{1:n}) = \text{softmax}(\mathbf{W}_\text{c}\mathbf{s} + \mathbf{b}_\text{c}) \end{equation} \noindent where $\textbf{s} \in \mathbb{R}^{d_\text{s}}$ is the computed task-specific sentence representation for the classifier, and $\textbf{W}_\text{s} \in \mathbb{R}^{d_\text{s} \times d_h}$, $\textbf{W}_\text{c} \in \mathbb{R}^{d_\text{c} \times d_s}$, $\textbf{b}_\text{s} \in \mathbb{R}^{d_s}$, $\textbf{b}_\text{c} \in \mathbb{R}^{d_c}$ are trainable parameters. As an objective function, we use the cross entropy of the predicted and true class distributions. The results of the experiments on the five datasets are shown in table \ref{table1}. In this table, we report the test accuracy of our model and various other models on each dataset in terms of percentage. To consider the effects of random initialization, we report the best numbers obtained from each several runs with hyper-parameters fixed. Compared with the previous syntactic tree-based models as well as other neural models, our SATA Tree-LSTM shows superior or competitive performance on all tasks. Specifically, our model achieves new state-of-the-art results within the tree-structured model class on 4 out of 5 sentence classification tasks---SST-2, SST-5, MR, and TREC. The model shows its strength, in particular, when the datasets provide phrase-level supervision to facilitate tree structure learning (i.e. SST-2, SST-5). Moreover, the numbers we report for SST-5 and TREC are competitive to the existing state-of-the-art results including ones from structurally pre-trained models such as ELMo \cite{peters2018deep}, proving our model's superiority. Note that the SATA Tree-LSTM also outperforms the recent latent tree-based model, indicating that modeling a neural model with explicit linguistic knowledge can be an attractive option. On the other hand, a remaining concern is that our SATA Tree-LSTM is not robust to random seeds when the size of a dataset is relatively small, as tag embeddings are randomly initialized rather than relying on pre-trained ones in contrast with the case of words. From this observation, we could find out there needs a direction of research towards pre-trained tag embeddings. \begin{table*}[t!] \centering \resizebox{0.7\textwidth}{!}{ \begin{tabular}{|l|c|c|c|c|c|} \hline \multicolumn{1}{|c|}{\textbf{Models}} & \textbf{SST-2} & \textbf{SST-5} & \textbf{MR} & \textbf{SUBJ} & \textbf{TREC} \\ \hline \hline \multicolumn{6}{|l|}{\textbf{Tree-structured models}} \\ \hline RNTN \cite{socher2013recursive} & 85.4 & 45.7 & - & - & - \\ AdaMC-RNTN \cite{dong2014adaptive} & 88.5 & 46.7 & - & - & - \\ TE-RNTN \cite{qian2015learning} & 87.7 & 49.8 & - & - & - \\ TBCNN \cite{mou2015discriminative} & 87.9 & 51.4 & - & - & 96.0 \\ Tree-LSTM \cite{tai2015improved} & 88.0 & 51.0 & - & - & - \\ AdaHT-LSTM-CM \cite{liu2017adaptive} & 87.8 & 50.2 & 81.9 & 94.1 & - \\ DC-TreeLSTM \cite{liu2017dynamic} & 87.8 & - & 81.7 & 93.7 & 93.8 \\ TE-LSTM \cite{huang2017encoding} & 89.6 & 52.6 & 82.2 & - & - \\ BiConTree \cite{teng2017head} & 90.3 & 53.5 & - & - & 94.8 \\ Gumbel Tree-LSTM$^\star$ \cite{choi2018learning} & 90.7 & 53.7 & - & - & - \\ TreeNet \cite{cheng2018treenet} & - & - & 83.6 & \underline{95.9} & 96.1 \\ \textbf{SATA Tree-LSTM (Ours)} & \textbf{91.3} & \underline{\textbf{54.4}} & \textbf{83.8} & \textbf{95.4} & \underline{\textbf{96.2}} \\ \hline \hline \multicolumn{6}{|l|}{\textbf{Other neural models}} \\ \hline CNN \cite{kim2014convolutional} & 88.1 & 48.0 & 81.5 & 93.4 & 93.6 \\ AdaSent \cite{zhao2015self} & - & - & 83.1 & 95.5 & 92.4 \\ LSTM-CNN \cite{zhou2016text} & 89.5 & 52.4 & 82.3 & 94.0 & 96.1 \\ byte-mLSTM$^\dagger$ \cite{radford2017learning} & \underline{91.8} & 52.9 & \underline{86.9} & 94.6 & - \\ BCN + Char + CoVe$^\dagger$ \cite{mccann2017learned} & 90.3 & 53.7 & - & - & 95.8 \\ BCN + Char + ELMo$^\dagger$ \cite{peters2018deep} & - & \underline{54.7$\pm$0.5} & - & - & - \\ \hline \end{tabular} } \caption{The comparison of various models on different sentence classification tasks. We report the test accuracy of each model in percentage. Our SATA Tree-LSTM shows superior or competitive performance on all tasks, compared to previous tree-structured models as well as other sophisticated models. $\star$: Latent tree-structured models. $\dagger$: Models which are pre-trained with large external corpora.} \label{table1} \end{table*} \subsubsection{Natural language inference} To estimate the performance of our model beyond the tasks requiring only one sentence at a time, we conduct an experiment on the Stanford Natural Language Inference \cite{snli} dataset, each example of which consists of two sentences, the premise and the hypothesis. Our objective given the data is to predict the correct relationship between the two sentences among three options--- contradiction, neutral, or entailment. We use the siamese architecture to encode both the premise ($p_{1:m}$) and hypothesis ($h_{1:n}$) following the standard of sentence-encoding models in the literature. (Specifically, $p_{1:m}$ is encoded as $\check{\mathbf{h}}_\text{root}^p \in \mathbb{R}^{d_h}$ and $h_{1:n}$ is encoded as $\check{\mathbf{h}}_\text{root}^h \in \mathbb{R}^{d_h}$ with the same encoder.) Then, we leverage some heuristics \cite{mou2016natural}, followed by one fully-connected layer with a ReLU activation and a softmax classifier. Specifically, \begin{equation} \mathbf{z} = \left[ \check{\mathbf{h}}_\text{root}^p; \check{\mathbf{h}}_\text{root}^h; | \check{\mathbf{h}}_\text{root}^p - \check{\mathbf{h}}_\text{root}^h |; \check{\mathbf{h}}_\text{root}^p \odot \check{\mathbf{h}}_\text{root}^h \right] \end{equation} \begin{equation} \mathbf{s} = \text{ReLU}(\mathbf{W}_\text{s} \mathbf{z} + \mathbf{b}_\text{s}) \end{equation} \begin{equation} p(y|p_{1:m}, h_{1:n}) = \text{softmax}(\mathbf{W}_\text{c}\textbf{s} + \mathbf{b}_\text{c}) \end{equation} \noindent where $\textbf{z} \in \mathbb{R}^{4d_h}$, $\textbf{s} \in \mathbb{R}^{d_s}$ are intermediate features for the classifier and $\textbf{W}_\text{s} \in \mathbb{R}^{d_\text{s} \times 4d_h}$, $\textbf{W}_\text{c} \in \mathbb{R}^{d_\text{c} \times d_s}$, $\textbf{b}_\text{s} \in \mathbb{R}^{d_s}$, $\textbf{b}_\text{c} \in \mathbb{R}^{d_c}$ are again trainable parameters. Our experimental results on the SNLI dataset are shown in table \ref{table2}. In this table, we report the test accuracy and number of trainable parameters for each model. Our SATA-LSTM again demonstrates its decent performance compared against the neural models built on both syntactic trees and latent trees, as well as the non-tree models. (Latent Syntax Tree-LSTM: \citeauthor{yogatama2017learning} (\citeyear{yogatama2017learning}), Tree-based CNN: \citeauthor{mou2016natural} (\citeyear{mou2016natural}), Gumbel Tree-LSTM: \citeauthor{choi2018learning} (\citeyear{choi2018learning}), NSE: \citeauthor{munkhdalai2017neural} (\citeyear{munkhdalai2017neural}), Reinforced Self-Attention Network: \citeauthor{shen2018reinforced} (\citeyear{shen2018reinforced}), Residual stacked encoders: \citeauthor{nie2017shortcut} (\citeyear{nie2017shortcut}), BiLSTM with generalized pooling: \citeauthor{chen2018enhancing} (\citeyear{chen2018enhancing}).) Note that the number of learned parameters in our model is also comparable to other sophisticated models, showing the efficiency of our model. Even though our model has proven its mettle, the effect of tag information seems relatively weak in the case of SNLI, which contains a large amount of data compared to the others. One possible explanation is that neural models may learn some syntactic rules from large amounts of text when the text size is large enough, reducing the necessity of external linguistic knowledge. We leave the exploration of the effectiveness of tags relative to data size for future work. \subsubsection{Experimental details} Here we go over the settings common across our models during experimentation. For more task-specific details, refer to the supplemental materials. For our input embeddings, we used 300 dimensional 840B GloVe \cite{pennington2014glove} as pre-trained word embeddings, and tag representations were randomly sampled from the uniform distribution [-0.005, 0.005]. Tag vectors are revised during training while the fine-tuning of the word embedding depends on the task. Our models were trained using the Adam \cite{kingma2014adam} or Adadelta \cite{zeiler2012adadelta} optimizer, depending on task. For regularization, weight decay is added to the loss function except for SNLI following \citeauthor{loshchilov2017fixing} (\citeyear{loshchilov2017fixing}) and Dropout \cite{srivastava2014dropout} is also applied for the word embeddings and task-specific classifiers. Moreover, batch normalization \cite{ioffe2015batch} is adopted for the classifiers. As a default, all the weights in the model are initialized following \citeauthor{he2015delving} (\citeyear{he2015delving}) and the biases are set to 0. The total norm of the gradients of the parameters is clipped not to be over 5 during training. Our best models for each dataset were chosen by validation accuracy in cases where a validation set was provided as a part of the dataset. Otherwise, we perform a grid search on probable hyper-parameter settings, or run 10-fold cross-validation in cases where even a test set does not exist. \begin{table}[t!] \centering \small \resizebox{0.9\columnwidth}{!}{ \begin{tabular}{|l|c|c|} \hline \multicolumn{1}{|c|}{\textbf{Models}} & \textbf{Acc.} & \textbf{\# Params} \\ \hline \hline \multicolumn{3}{|l|}{\textbf{Tree-structured models}} \\ \hline 100D Latent Syntax Tree-LSTM$^\star$ & 80.5 & 500K \\ 300D Tree-based CNN & 82.1 & 3.5M \\ 300D SPINN-PI & 83.2 & 3.7M \\ 300D Gumbel Tree-LSTM$^\star$ & 85.6 & 2.9M \\ \textbf{300D SATA Tree-LSTM (Ours)} & \textbf{85.9} & \textbf{3.3M} \\ \hline \hline \multicolumn{3}{|l|}{\textbf{Other neural models}} \\ \hline 300D NSE & 84.6 & 3.0M \\ 300D Reinforced Self-Attention Network & 86.3 & 3.1M \\ 600D Residual stacked encoders & 86.0 & 29M \\ 600D BiLSTM with generalized pooling & \underline{86.6} & 65M \\ \hline \end{tabular}} \caption{The accuracy of diverse models on Stanford Natural Language Inference. For fair comparison, we only consider sentence-encoding based models. Our model achieves a comparable result with a moderate number of parameters. $\star$: Latent tree models.} \label{table2} \end{table} \subsection{Ablation Study} In this section, we design an ablation study on the core modules of our model to explore their effectiveness. The dataset used in this experiment is SST-2. To conduct the experiment, we only replace the target module with other candidates while maintaining the other settings. To be specific, we focus on two modules, the leaf-LSTM and structure-aware tag embeddings (tag-level tree-LSTM). In the first case, the leaf-LSTM is replaced with a fully-connected layer with a $\tanh$ activation or Bi-LSTM. In the second case, we replace the structure-aware tag embeddings with naive tag embeddings or do not employ them at all. The experimental results are depicted in Figure \ref{fig:figure3}. As the chart shows, our model outperforms all the other options we have considered. In detail, the left part of the chart shows that the leaf-LSTM is the most effective option compared to its competitors. Note that the sequential leaf-LSTM is somewhat superior or competitive than the bidirectional leaf-LSTM when both have a comparable number of parameters. We conjecture this may because a backward LSTM does not add additional useful knowledge when the structure of a sentence is already known. In conclusion, we use the uni-directional LSTM as a leaf module because of its simplicity and remarkable performance. Meanwhile, the right part of the figure demonstrates that our newly introduced structure-aware embeddings have a real impact on improving the model performance. Interestingly, employing the naive tag embeddings made no difference in terms of the test accuracy, even though the absolute validation accuracy increased (not reported in the figure). This result supports our assumption that tag information should be considered in the structure. \subsection{Qualitative Analysis} \begin{figure}[!t] \centering \includegraphics[width=0.9\columnwidth]{figure3.pdf} \caption{An ablation study on the core modules of our model. The test accuracy of each model on SST-2 is reported. The results demonstrate that the modules play an important role for achieving the superior performance of our model. FC: A fully connected-layer with a $\tanh$ function. w/o tags: Tag embeddings are not used. w/ tags: The naive tag embeddings are directly inserted into each node of a tree.} \label{fig:figure3} \end{figure} In previous sections, we have numerically demonstrated that our model is effective in encouraging useful composition of semantic units. Here, we directly investigate the computed representations for each node of a tree, showing that the remarkable performance of our model is mainly due to the gradual and recursive composition of the intermediate representations on the syntactic structure. To observe the phrase-level embeddings at a glance, we draw a scatter plot in which a point represents the corresponding intermediate representation. We utilize PCA (Principal Component Analysis) to project the representations into a two-dimensional vector space. As a target parse tree, we reuse the one seen in Figure \ref{fig:figure1}. The result is shown in Figure \ref{fig:figure4}. From this figure, we confirm that the intermediate representations have a hierarchy in the semantic space, which is very similar to that of the parse tree. In other words, as many tree-structured models pursue, we can see the tendency of constructing the representations from the low-level (the bottom of the figure) to the high-level (the top-left and top-right of the figure), integrating the meaning of the constituents recursively. An interesting thing to note is that the final sentence representation is near that of the phrase \textit{`, the stories are quietly moving.'} rather than that of \textit{`Despite the film's shortcomings'}, catching the main meaning of the sentence. \begin{figure}[!t] \centering \includegraphics[width=0.95\columnwidth]{figure4.pdf} \caption{A scatter plot whose points represent the intermediate representations for each node of the tree in Figure 1. From this figure, we can see the tendency of constructing the representations recursively from the low to the high level.} \label{fig:figure4} \end{figure} \section{Conclusion} We have proposed a novel RvNN architecture to fully utilize linguistic priors. A newly introduced tag-level tree-LSTM demonstrates that it can effectively control the composition function of the corresponding word-level tree-LSTM. In addition, the proper contextualization of the input word vectors results in significant performance improvements on several sentence-level tasks. For future work, we plan to explore a new way of exploiting dependency trees effectively, similar to the case of constituency trees. \section*{Acknowledgments} We thank anonymous reviewers for their constructive and fruitful comments. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2016M3C4A7952587). \fontsize{9pt}{10pt} \selectfont

e7fb01aca8163d63cc09b76ab3620b5606bd35e6

\section{Introduction} Let $R$ be a Noetherian standard graded algebra over a field $k=R_0$, $\mathfrak{m}=\oplus_{n\in \mathbb Z_{> 0}}R_n$, and $I$ a homogeneous $R$-ideal. In this paper we study asymptotic behavior of the lowest degree of the local cohomology modules $\{\HH{i}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$, provided that they are finite. As we make clear below, such behavior can be viewed as an ``asymptotic Kodaira vanishing for thickenings" phenomenon, and have recently appeared in various works such as \cite{BBLSZ, DM, Claudiu}. To describe our motivations and questions precisely, let us recall some notations. For a graded $R$-module $M=\oplus_{i\in \mathbb Z}M_i$ one defines $$\indeg{M}=\min\{i\mid M_i\neq0\},\qquad \topdeg{M}=\max\{i\mid M_i\neq 0\}.$$ If $M=0$, we set $\indeg M=\infty$ and $\topdeg M=-\infty$. We also set $$\beta(M)=\max\{i\mid (M/\mathfrak{m} M)_i\neq0\},$$ i.e., $\beta(M)$ is the maximal degree of an element in a minimal set of homogeneous generators of $M$. The {\it Castelnuovo-Mumford regularity} of $M$ is defined as $$\reg(M)=\max\{\topdeg{\HH{i}{\mathfrak{m}}{M}}+i\}.$$ It is known that $\reg(R/I^n)$ agrees with a linear function for $n\gg 0$, this fact was proved independently in \cite{CHT} and \cite{Kod} when $R$ is a polynomial ring over a field, and extended in \cite{TW} for arbitrary standard graded rings. If $\HH{i}{\mathfrak{m}}{M}\neq 0$, the $\topdeg$ of $\HH{i}{\mathfrak{m}}{M}$ is always finite, however this is not the case for $\indeg$. In fact, since $\HH{i}{\mathfrak{m}}{M}$ is an Artinian module, we have that $\indeg\HH{i}{\mathfrak{m}}{M}>-\infty$ if and only if $\HH{i}{\mathfrak{m}}{M}$ is Noetherian. Our work is guided by the following questions raised in \cite{DM}. \begin{question}\label{motivQ} Assume $\HH{i}{\mathfrak{m}}{R/I^n}$ is Noetherian for $n\gg 0$. \begin{enumerate} \item Does there exist $\alpha \in \mathbb Z$ such that $\indeg \HH{i}{\mathfrak{m}}{R/I^n} >\alpha n$ for every $n\gg 0$? In other words, is $\displaystyle\liminf_{n\rightarrow \infty} \frac{ \indeg\HH{i}{\mathfrak{m}}{R/I^n}}{n}$ finite? \item If so, does the limit $\displaystyle\lim_{n\rightarrow \infty} \frac{ \indeg\HH{i}{\mathfrak{m}}{R/I^n}}{n}$ exist? \end{enumerate} \end{question} It follows from \cite[1.4]{BBLSZ} that when $R$ is a polynomial ring over a field of characteristic $0$, $I$ is a prime ideal, $X:=\Proj(R/I)$ is locally a complete intersection (lci), and $i$ is at most the codimension of the singular locus of $X$, then $ \indeg\HH{i}{\mathfrak{m}}{R/I^n} \geqslant 0$ for all $n>0$. As explained there, this can be viewed as a Kodaira Vanishing Theorem for thickenings of $X$. When $I$ is a determinantal ideal, more precise behavior of vanishing results, and other homological invariants of thickenings of $I$ are available, see for instance \cite{Claudiu} and \cite{Raicu2}. Our initial interest in the question came from \cite{DM} where we need an affirmative answer to part (1) of Question \ref{motivQ} to obtain efficient bounds on the lengths of local cohomology modules of powers. Robert Lazarsfeld pointed out to us that this is indeed the case when $X=\Proj(R/I)$ is a l.c.i variety, $k$ is of characteristic zero, and $i$ is at most the dimension of $X$. Thus Kodaira vanishing may not hold, but the lowest degrees of $\HH{i}{\mathfrak{m}}{R/I^n}$ are still bounded below by a linear function. In this work we provide further answers to Question \ref{motivQ} above, in the case when $R$ is not necessarily a polynomial ring and $I$ may not be prime, or even reduced. The first main general result of this article (Theorem \ref{mainVB}), provides a linear lower bound for the initial degrees of local cohomology of the symmetric powers $\{S^n(E)\}_{n\in\mathbb N}$, where $E$ is a graded module that is locally free on $\Spec R \setminus \{\mathfrak{m}\}$. Our proof relies on a duality statement and the result on regularity by Trung and Wang in \cite{TW}. Here $(-)^*$ denotes the $R$-dual $\Hom_R(-,R)$. \begin{thm} \label{MainT1} Let $(R,\mathfrak{m})$ be a standard Cohen-Macaulay graded algebra over a field $k$ of characteristic zero. Set $d=\dim R\geqslant 2$. Let $E$ be a graded $R$-module which is free locally on $\Spec R\setminus \{\mathfrak{m}\}$. Then there exists an integer $\varepsilon$ such that $$\indeg \HH{i}{\mathfrak{m}}{S^n(E)} \geqslant -\beta(E^*) n+\varepsilon$$ for every $n\geqslant 1$ and $ 1\leqslant i< d$. \end{thm} We apply this Theorem to answer Question \ref{motivQ} (1) affirmatively and effectively when $R$ is any standard graded algebra over a field of characteristic $0$, $R/I$ is Cohen-Macaulay, and $I$ is a complete intersection locally on $\Spec R\setminus \{\mathfrak{m}\}$. In this case, $\liminf_{n\rightarrow \infty} \frac{ \indeg\HH{i}{\mathfrak{m}}{R/I^n}}{n}$ is bounded below by $-\max\{\beta(E^*),0\}$ where $E$ is the conormal module $I/I^2$ (see Corollary \ref{conormal}). This result can be seen as an algebraic version of \cite[1.4]{BBLSZ} and \cite[5.6]{DM}, our proof via local cohomology of symmetric powers of conormal modules is inspired by the proofs of these results. Theorem \ref{MainT1} also allows us to show a result on stabilization of maps of local cohomology of powers of ideals. We show that, if $I$ is as in the previous paragraph, and if $\beta(E^*)<0$, then the maps between local cohomology and Ext modules of consecutive powers of $I$ eventually stabilize on each graded degree. This result is closely related to \cite[1.1]{BBLSZ} and provide a partial answer to a question of Eisenbud, Musta\c{t}\u{a}, and Stillman (\cite[6.1]{EMS}), see Corollary \ref{stabExt} and its preceding paragraph for more details. In general, if one only assumes that $\HH{i}{\mathfrak{m}}{R/I^n}$ is Noetherian for $n\gg 0$, it is complicated to find bounds on its lowest degrees. However, we are able to prove that there is a polynomial lower bound, regardless of the characteristic of $k$. The proof rests on a result by Chardin, Ha, and Hoa (\cite{CHH}), and is provided in Section \ref{polySec}. In Section \ref{monoSec} we focus on the case where $I$ is a monomial ideal in a polynomial ring $R$. As expected, the extra combinatorial structure allows for better results. Assuming that $\HH{i}{\mathfrak{m}}{R/I^n}$ is Noetherian for $n\gg 0$, one can show that either $\indeg\HH{i}{\mathfrak{m}}{R/I^n}=0$ for $n\gg 0$ or $\displaystyle\liminf_{n\rightarrow \infty} \frac{ \indeg\HH{i}{\mathfrak{m}}{R/I^n}}{n}\geqslant 1$, and the latter holds precisely when $\tilde H_{i-1}(\Delta(I))= 0$, where $\Delta(I)$ is the simplicial complex whose Stanley-Reisner ideal is $\sqrt{I}$. \section{Symmetric Powers of Locally Free Modules and Linear Lower Bound}\label{linearSec} Let $E$ be a Noetherian graded module and set $u=\mu(E):=\dim_k E/\mathfrak{m} E$. Let $$F_1\xrightarrow{\phi} F_0\to E\to 0$$ be a minimal presentation of $E$, where $F_0$ and $F_1$ are graded free $R$-modules, and $\phi$ is an $u\times s$ matrix with entries in $\mathfrak{m}$. Let $T_1,\ldots, T_u$ be a set of variables and $\ell_1,\ldots, \ell_s$ the linear forms determined by $$\begin{bmatrix}\ell_1,&\cdots,&\ell_s \end{bmatrix}= \begin{bmatrix}T_1,&\cdots,&T_u \end{bmatrix}\phi. $$ The ring $\Sym(E):=R[T_1,\ldots, T_u]/(\ell_1,\cdots, \ell_s)$ is the {\it symmetric algebra} of $E$. Let $d_1,\ldots, d_u$ be the degrees of a homogeneous minimal generating set of $E$. We can assign to $\Sym(E)$ a bi-graded structure where $T_i$ has bi-degree $(d_i,1)$ for every $i=1,\ldots, u$. The $n$th-graded component of $\Sym(E)$, $S^n(E)=\oplus_{a\in \mathbb Z}\Sym(E)_{(a,n)}$, is the {\it $n$th-symmetric power} of $E$. Let $M$ be any Noetherian graded $R$-module and $U\subseteq E$ a graded submodule. We say $U$ is an {\it $M$-reduction} of $E$, if $S^n(E)\otimes_R M=S^1(U)S^{n-1}(E)\otimes_RM$ for $n\gg 0$, where $S^1(U)$ is seen as a submodule of $S^1(E)$. Following \cite{TW}, we define $$\rho_M(E):=\min\{\beta(U)\mid U \text{ is an $M$-reduction of } E \}.$$ We note that $\rho_M(E)\leqslant \beta(E)$ for every $R$-module $M$. The following theorem is the module version of \cite[3.2]{TW} and the proofs of these results are identical, however we include some relevant details for the reader's convenience. We remark that even though the algebras in \cite{TW} are positively graded, the proof of this result does not use this assumption. \begin{thm} \label{regModules} Let $R$ be a standard graded algebra over a Noetherian ring $A$. Let $E$ and $M$ be finitely generated graded $R$-modules. Then $$\reg(S^n(E)\otimes_R M)=\rho_M(E)n+e$$ for some integer $e\geqslant \indeg M$ and every $n\gg 0$. \end{thm} \begin{proof} Let $U$ be an $M$-reduction of $E$ such that $d(U)=\rho_M(E)$. Let $\mathcal M=\Sym(E)\otimes_R M = \oplus_{n\in \mathbb N} S^n(E)\otimes_R M$ and notice $\mathcal M$ is a finitely generated graded $\Sym(U)$-module. Let $s=\mu_A(R_1)$, $v=\mu(U)$, and $u_1,\ldots, u_v$ the degrees of a homogeneous minimal generating set of $U$, then $\Sym(U)$ is a quotient ring of the bi-graded polynomial ring $A[x_1,\ldots, x_s, y_1,\ldots, y_v]$ where $x_i$ has degree $(1,0)$ for each $i$ and $y_j$ has degree $(u_i,1)$ for each $j$. Therefore, \cite[2.2]{TW} implies $\reg(S^n(E)\otimes_R M)$ is a linear function $\rho n+e$ for some $e$ and $\rho\leqslant \rho_M(E)$. Finally, proceeding as in \cite[3.1]{TW} we obtain $\rho \geqslant \rho_M(E)$ and $e\geqslant \indeg M$, finishing the proof. \end{proof} For the next result, we assume $R$ is a local ring or standard graded over a field. Let $\mathfrak{m}$ be the (irrelevant) maximal ideal of $R$, $k=R/\mathfrak{m}$, and $E_R(k)$ the {\it (graded) injective hull} of $k$. For a (graded) $R$-module $M$ we set $$M^\vee:=\Hom_R(M,\, E_R(k)).$$ The following is a generalization of a duality result of Horrocks (\cite{Hor}). \begin{prop}\label{duality} Let $(R,\mathfrak{m},k)$ be a Cohen-Macaulay local ring $($or positively graded $k$-algebra$)$. Set $d=\dim R\geqslant 2$ and $\omega$ a (graded) canonical module of $R$. Fix $1\leqslant i\leqslant d-1$, then for a $($graded$)$ $R$-module $M$ of dimension $d$ that is $S_{i+1}$ locally on $\Spec R\setminus \{\mathfrak{m}\}$ we have $($graded$)$ isomorphisms $$\HH{i}{\mathfrak{m}}{M}^\vee\cong \HH{d-i+1}{\mathfrak{m}}{\Hom_R(M,\omega)}\quad\text{if } i\geqslant 2,$$ and, $$\HH{1}{\mathfrak{m}}{M}^\vee\cong \ker\big(\HH{d}{\mathfrak{m}}{\Hom_R(M,\omega)}\rightarrow \HH{d}{\mathfrak{m}}{\Hom_R(F_0,\omega)}\big),$$ where $F_0\twoheadrightarrow M$ is a free module. \end{prop} \begin{proof} We begin the proof with the following claim. \ \noindent {\it {\bf Claim:} Let $N$ be a $($graded$)$ $R$-module that is Maximal Cohen-Macaulay $($MCM$)$ locally on $\Spec R\setminus \{\mathfrak{m}\}$ and $\cdots\to F_{ 0} \rightarrow N\to 0$ a $($graded$)$ free resolution of $N$. Then, $\Ext^1_R(N, \omega)\cong \ker\big(\HH{2}{\mathfrak{m}}{\Hom_R(N,\omega)}\rightarrow \HH{2}{\mathfrak{m}}{\Hom_R(F_0,\omega)}\big)$ if $d=2$, and $\Ext^1_R(N, \omega)\cong \HH{2}{\mathfrak{m}}{\Hom_R(N, \omega)}$ if $d\geqslant 3$.} In order to prove this claim, we consider the $R$-modules $K$ and $C$ that fit in the following two exact sequences \begin{equation}\label{qq1} 0\rightarrow\Hom_R(N,\omega)\rightarrow \Hom_R(F_0,\omega)\rightarrow C\rightarrow 0 \quad \quad\text{and} \end{equation} \begin{equation}\label{qq2} 0\rightarrow K\rightarrow \Hom_R(F_1,\omega)\rightarrow\Hom_R(F_2,\omega). \end{equation} By applying the depth lemma to \eqref{qq2} we obtain $\depth K\geqslant 2$, therefore $$\Ext^1_R(N,\omega)=\HH{0}{\mathfrak{m}}{\Ext^1_R(N,\omega)}=\HH{0}{\mathfrak{m}}{K/C}\cong \HH{1}{\mathfrak{m}}{C}$$ where the first equality follows by the assumption on $N$. Hence, the conclusion of the claim follows from \eqref{qq1}. \ Now, back to the original statement, we note that the result follows by the claim and local duality \cite[3.6.19]{BH} if $d=2$, then we may assume $d\geqslant 3$. Let $\Omega^nM$ be the $n$th-syzygy module of $M$. Again by local duality and the claim we have \begin{equation}\label{rr1} \HH{i}{\mathfrak{m}}{M}^\vee\cong \Ext_R^{d-i}(M,\omega)\cong \Ext^1_R(\Omega^{d-i-1}M,\omega)\cong \HH{2}{\mathfrak{m}}{\Hom_R(\Omega^{d-i-1}M,\omega)}. \end{equation} Let $0\leqslant t\leqslant d-i-2$, by assumption $\Omega^tM$ is MCM in codimension $i+t+1$, which implies that $\dim \Ext^1_R(\Omega^{t}M,\omega)<d-i-t-1$. Let $\cdots\to F_{ 0} \rightarrow M\to 0$ be a (graded) free resolution of $M$. From the exact sequence $$0\rightarrow\Hom_R(\Omega^{t}M,\omega)\rightarrow\Hom_R(F_t,\omega)\rightarrow\Hom_R(\Omega^{t+1}M,\omega)\rightarrow \Ext^1_R(\Omega^{t}M,\omega)\rightarrow 0$$ we obtain \begin{equation}\label{rr2} \HH{d-i-t}{\mathfrak{m}}{\Hom_R(\Omega^{t+1}M,\omega)}\cong \HH{d-i-t+1}{\mathfrak{m}}{\Hom_R(\Omega^{t}M,\omega)},\,\,\,\,\,\text{ if } i+t\geqslant 2, \end{equation} and \begin{equation}\label{rr3} \HH{d-1}{\mathfrak{m}}{\Hom_R(\Omega^{1}M,\omega)}\cong \ker\big(\HH{d}{\mathfrak{m}}{\Hom_R(M,\omega)}\rightarrow \HH{d}{\mathfrak{m}}{\Hom_R(F_0,\omega)}\big). \end{equation} The statement now follows from \eqref{rr1}, \eqref{rr2}, and \eqref{rr3}. \end{proof} Given an $R$-module $M$, we denote by $\Gamma(M)$ the {\it divided powers algebra} of $M$ \cite[Appendix 2]{E}. We set $M^*:=\Hom_R(M,R)$ and for a graded $R$-algebra $S=\oplus_{n\in \mathbb N}S_n$, we denote by $S^*:=\oplus_{n\in \mathbb N}S_n^*$ the {\it graded dual} of $S$. We need the following technical lemma for the proof of our main result. \begin{lemma}\label{symm} Let $R$ be a commutative $($graded$)$ ring and $M$ a $($graded$)$ $R$-module. Then there exist natural $($graded$)$ maps $$\Sym(M^*)\xrightarrow{\alpha} \Gamma(M^*)\xrightarrow{\beta} \Sym(M)^*.$$ Moreover, $\alpha$ is an isomorphism if $R$ contains the field of rational numbers and $\beta$ is an isomorphism if $M$ is free. \end{lemma} \begin{proof} For the construction and results on $\alpha$, see \cite[Proposition III.3., page 256]{Roby}. See \cite[A2.6 and A2.7(c)]{E} for the corresponding information for $\beta$. \end{proof} The following is the main theorem of this section. \begin{thm}\label{mainVB} Let $(R,\mathfrak{m})$ be a Cohen-Macaulay standard graded algebra over a field $k$ of characteristic zero. Set $d=\dim R\geqslant 2$. Let $E$ be a graded $R$-module which is free locally on $\Spec R\setminus \{\mathfrak{m}\}$. Then there exists an integer $\varepsilon$ such that $$\indeg \HH{i}{\mathfrak{m}}{S^n(E)} \geqslant -\beta(E^*) n+\varepsilon$$ for every $n\geqslant 1$ and $ 1\leqslant i< d$. \end{thm} \begin{proof} First, assume $i\geqslant 2$. Let $\omega$ be the canonical module of $R$. By Hom-Tensor adjointness and the isomorphism $R\cong \Hom_R(\omega,\, \omega)$, we have $S^n(E^*)^*\cong \Hom_R(S^n(E^*)\otimes_R \omega, \omega).$ By the assumption we have $S^n(E^*)\otimes_R \omega$ is MCM locally on $\Spec R\setminus \{\mathfrak{m}\}$, therefore the natural map $$S^n(E^*)\otimes_R \omega\rightarrow \Hom_R(\Hom_R( S^n(E^*)\otimes_R \omega,\, \omega),\, \omega)$$ is an isomorphism locally on $\Spec R\setminus \{\mathfrak{m}\}$. Hence, by Proposition \ref{duality} $$ \HH{i}{\mathfrak{m}}{S^n( E^*)^*}\cong \HH{i}{\mathfrak{m}}{\Hom_R( S^n(E^*)\otimes_R \omega,\, \omega)}\cong \HH{d-i+1}{\mathfrak{m}}{S^n(E^*)\otimes_R \omega }^\vee, $$ By Theorem \ref{regModules}, we have $$\topdeg \HH{d-i+1}{\mathfrak{m}}{S^n(E^*) \otimes_R \omega }\leqslant \beta(E^*) n-\varepsilon$$ for some $\varepsilon \in \mathbb Z$ and every $n\geqslant 1$. Therefore $ \indeg \HH{i}{\mathfrak{m}}{S^n( E^*)^*}\geqslant -\beta(E^*) n+\varepsilon$ for every $n\geqslant 1$ and $i\geqslant 2$. The map $\Sym(E^*)\xrightarrow{\beta\circ\alpha} \Sym(E)^*$ in Lemma \ref{symm} (with $M=E$) is an isomorphism locally on $\Spec R\setminus \{\mathfrak{m}\}$, hence \begin{equation}\label{isomSym} S^n( E^*)^*\cong S^n( E)^{**}. \end{equation} The result now follows for $i\geqslant 2$ by observing that $\HH{i}{\mathfrak{m}}{S^n( E)}\cong \HH{i}{\mathfrak{m}}{S^n( E)^{**}} $ as $S^n(E)$ is free, hence reflexive, locally on $\Spec R\setminus \{\mathfrak{m}\}$. Now, we show the statement for $i=1$. Fix $n\gg 0$ and consider the short exact sequence $$0\to \Ker(\varphi) \to S^n(E)\xrightarrow{\varphi} S^n(E)^{**}\to \mathcal C\to 0.$$ Since $\varphi$ is an isomorphism locally on $\Spec R\setminus \{\mathfrak{m}\}$, we have $\dim \Ker (\varphi) = \dim \mathcal C = 0$. Then $\HH{1}{\mathfrak{m}}{S^n(E)}\cong\mathcal C$, as $\depth S^n( E)^{**}\geqslant 2$. Therefore, $$\indeg \HH{1}{\mathfrak{m}}{S^n(E)} = \indeg \mathcal C \geqslant \indeg S^n(E)^{**}=\indeg S^n( E^*)^*,$$ where the last equality follows from \eqref{isomSym}. Let $\oplus_{i=1}^u R(-a_i)\rightarrow S^n( E^*)\rightarrow 0$ be the first map of a minimal homogeneous resolution of $S^n( E^*)$, where $u=\mu(S^n( E^*))$. Then, $S^n( E^*)^*\hookrightarrow \oplus_{i=1}^u R(a_i)$. We conclude $$\indeg S^n( E^*)^*\geqslant -\max_i\{a_i\}\geqslant -\reg (S^n( E^*))\geqslant -\beta(E^*)n+\varepsilon,$$ for some $\varepsilon\in \mathbb Z$ and $n\geqslant 1$ by Theorem \ref{regModules}. \end{proof} Assume $E$ is a graded submodule of a free graded $R$-module $F=\oplus_{i=1}^\gamma R(-d_i)$. We have the natural map of symmetric algebras $$\Sym(E)\rightarrow \Sym(F)=R[T_1,\ldots,T_\gamma],$$ where each $T_i$ has bidegree $(d_i, 1)$. The image of this map is the bi-graded algebra$$\mathcal R[E]:=\oplus_{n\in \mathbb N}E^n\subset R[T_1,\ldots,T_\gamma].$$ The ring $\mathcal R[E]$ is called the {\it Rees algebra} of $E$ with respect to the embedding $E\subset F$. It is known that if $E$ has a {\it rank}, i.e., $E_P$ is free of constant rank for every $P\in \Ass(R)$, then $\mathcal R[E]$ is isomorphic fo $\Sym(E)/(\text{$R$-torsion})$ and hence it is independent of the graded embedding of $E$ into a free module (\cite{EHU}). \begin{cor} Let $(R,\mathfrak{m},k)$ and $E$ be as in Theorem \ref{mainVB}. Assume that $E$ has a rank, then $$\indeg \HH{i}{\mathfrak{m}}{E^n} \geqslant -\beta(E^*)n+\varepsilon$$ for some $\varepsilon\in \mathbb Z$ and every $n\geqslant 1$, $1\leqslant i< d$. \end{cor} \begin{proof} Since $E$ has a rank, and is free locally on $\Spec R \setminus \{\mathfrak{m}\}$, for every $n\geqslant 1$ we have $E^n\cong S^n(E)/(R\text{-torsion})$ and this $R$-torsion submodule is supported on $\{\mathfrak{m}\}$. Therefore, $\HH{i}{\mathfrak{m}}{E^n}=\HH{i}{\mathfrak{m}}{S^n(E)}$ for every $i\geqslant 1$ and the statement follows from Theorem \ref{mainVB}. \end{proof} \begin{cor}\label{conormal} Let $(R,\mathfrak{m})$ be a standard graded algebra over a field $k$ of characteristic zero. Let $I$ be a homogeneous $R$-ideal such that $S=R/I$ is Cohen-Macaulay. Assume $I_\mathfrak{p}$ is generated by a regular sequence in $R_\mathfrak{p}$ for every $\mathfrak{p}\in \Spec(R)\setminus \{\mathfrak{m}\}$ and that $\dim S\geqslant 2$. Let $E=I/I^2$ be the {\it conormal module} of $I$ and $E^*=\Hom_{S}(E, S)$, then \begin{enumerate} \item if $\beta(E^*)\ls0$, there exists $C\in \mathbb Z$ such that $$\indeg \HH{i}{\mathfrak{m}}{R/I^n} \geqslant C$$ for every $n\geqslant 1$ and $1\leqslant i< \dim R/I$; \item if $\beta(E^*)>0$, then there exists $\varepsilon\in \mathbb Z$ such that $$\indeg \HH{i}{\mathfrak{m}}{R/I^n} \geqslant -\beta(E^*)n+\varepsilon$$ for every $n\geqslant 1$ and $1\leqslant i< \dim R/I$. \end{enumerate} \end{cor} \begin{proof} By assumption $ E$ is a $R/I$-module that is free locally on $\Spec R\setminus \{\mathfrak{m}\}$. Since the natural epimorphism $S^n(E)\twoheadrightarrow I^n/I^{n+1}$ is an isomorphism locally on $\Spec R\setminus \{\mathfrak{m}\}$, we have $\HH{i}{\mathfrak{m}}{S^n(E)}\cong \HH{i}{\mathfrak{m}}{I^n/I^{n+1}}$ for $i\geqslant 1$. The conclusion now follows from Theorem \ref{mainVB} and by induction on $n$ via the inequality $$\indeg \HH{i}{\mathfrak{m}}{R/I^{n+1}} \geqslant \min\{\indeg \HH{i}{\mathfrak{m}}{I^n/I^{n+1}},\,\indeg \HH{i}{\mathfrak{m}}{R/I^{n}}\}$$ for $n\geqslant 1$. \end{proof} \begin{example}\label{limKnown} Let $R=k[x,y,u,v]/(xu^t-yv^t)$ for some $t\geqslant 1$ and $\chara k = 0$. Let $I=(x,y)$ and notice that $I$ and the ring $S=R/I$ satisfy the assumptions of Corollary \ref{conormal}. The graded free resolution of $I/I^2$ is $$0\rightarrow S(-1-t) \xrightarrow{ \begin{bmatrix} u^t\\-v^t\end{bmatrix} } S^2(-1) \rightarrow I/I^2\rightarrow 0.$$ Therefore, $(I/I^2)^*$ is isomorphic to the kernel of the map $S^2(1) \xrightarrow{ \begin{bmatrix} u^t&-v^t\end{bmatrix} } S(1+t) $, which is generated by $\begin{bmatrix} v^t\\u^t\end{bmatrix}$. Hence $\beta((I/I^2)^*)=t-1$. If $t\geqslant 2$, by Corollary \ref{conormal} we have $\indeg \HH{1}{\mathfrak{m}}{R/I^n} \geqslant -(t-1)n+\varepsilon$ for some $\varepsilon\in\mathbb Z$ and every $n\geqslant 1$. On the other hand, computing $ \HH{1}{\mathfrak{m}}{R/I^n}$ via the \u{C}ech complex of the system of parameters $\{u,v\}$ of the ring $R/I^n$, we obtain that the class of $[ \frac{x^{n-1}}{v^{tn-t}} , \, \frac{y^{n-1}}{u^{tn-t}} ]$ is nonzero and has degree $-(t-1)n+(t-1)$. We conclude that $$-(t-1)n+(t-1)\geqslant \indeg \HH{1}{\mathfrak{m}}{R/I^n} \geqslant -(t-1)n+\varepsilon$$ for every $n\geqslant 1$. Therefore, $$\lim_{n\rightarrow \infty}\frac{ \indeg \HH{1}{\mathfrak{m}}{R/I^n}}{n}=-(t-1).$$ Now, if $t=1$ we have $\beta((I/I^2)^*)=0$ and hence $\{\indeg \HH{1}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$ is bounded below by a constant. In fact, computations with Macaulay2 \cite{GS} suggests that the sequence $\{\indeg \HH{1}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$ in Example \ref{limKnown} agrees with the linear function $-(t-1)(n-1)$ for each $t\geqslant 1$ and $n\geqslant 2$. We record some of these values below. \begin{table}[ht] \centering \begin{tabular}{c| c c c c c c c c c c c c c c c c} \backslashbox{$t$}{$n$} & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & 16 \\ \hline 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 2 & -1 & -2 & -3 & -4 & -5 & -6 & -7 & -8 & -9 & -10 & -11 & -12 & -13 & -14 & -15 \\ 3 & -2 & -4 & -6 & -8 & -10 & -12 & -14 & -16 & -18 & -20 & -22 & -24 & -26 & -28 & -30 \\ 4 & -3 & -6 & -9 & -12 & -15 & -18 & -21 & -24 & -27 & -30 & -33 & -36 & -39 & -42 & -45 \\ 5 & -4 & -8 & -12 & -16 & -20 & -24 & -28 & -32 & -36 & -40 & -44 & -48 & -52 & -56 & -60 \\ \end{tabular} \label{sequences} \end{table} \end{example} \begin{example}\label{maxMinors} Let $X$ be a $2\times 3$ generic matrix and $ R=k[X]$ with $\chara k = 0$. Let $I=I_2(X)$ the ideal generated by the $2\times 2$ minors of $X$, then $R/I$ is Cohen-Macaulay of dimension 4, and $\Proj R/I$ is lci. Using Macaulay2 \cite{GS} we obtain $\beta((I/I^2)^*) =-1$, therefore by Corollary \ref{conormal} the sequence $\{\indeg \HH{3}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$ is bounded below by a constant. Indeed, $\indeg \HH{3}{\mathfrak{m}}{R/I^n} = 0 $ for every $n\geqslant 2$ \cite[5.1]{BBLSZ}. \end{example} In the following example we demonstrate that the lower bound $C$ in Corollary \ref{conormal} (1) may be negative. \begin{example} Let $R=k[x,y,z,u,v,w]/(x^2u^2+y^2v^2+z^2w^2)$ with $\chara k = 0$ and let $I=(x,y,z)$. Computations with Macaulay2 \cite{GS} show that $\beta((I/I^2)^*) =-1$ and suggest that $\indeg \HH{i}{\mathfrak{m}}{R/I^n}=-2$ for every $n \geqslant 3$. \end{example} In the following example we observe that, even when the ring $R$ is regular, the sequence $\{\indeg \HH{i}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$ may have linear behavior with negative slope. \begin{example} Let $R=k[x,y,u,v]$ with $\chara k = 0$ and $I = (x^2u-y^2v,u^2,uv,v^2)$. Computations with Macaulay2 \cite{GS} show that $\beta((I/I^2)^*) =2$ and suggest that $\indeg \HH{i}{\mathfrak{m}}{R/I^n}=-2n+1$ for every $n \geqslant 1$. \end{example} In \cite[6.1]{EMS}, Eisenbud, Musta\c{t}\u{a}, and Stillman asked for which ideals $I$ in a polynomial ring $R$ over a field $k$, there exists a decreasing chain of ideals $\{I_n\}_{n\in \mathbb N}$, cofinal with the regular powers $\{I^n\}_{n\in \mathbb N}$, such that the natural map $$\Ext_R^i(R/I_n,R)\rightarrow \lim_{\longrightarrow}\Ext_R^i(R/I_n,R)=\HH{i}{I}{R}$$ is injective for every $i$ and $n$. In \cite[1.2]{BBLSZ} the authors provide a partial answer to this question, showing that if $I$ is a homogeneous prime ideal such that $\Proj R/I$ is smooth, and $\chara k = 0$, then for each $i\in \mathbb N$ and $j\in \mathbb Z$ the map $$\Ext_R^i(R/I^n,R)_j\rightarrow \HH{i}{I}{R}_j$$ is injective for $n\gg 0$. Part (3) of the following corollary provides another partial answer to the question above for a class of ideals in Gorenstein rings. Part (1) and (2) are closely related to \cite[1.1]{BBLSZ}. \begin{cor}\label{stabExt} Let $R$, $I$, and $E$ be as in Corollary \ref{conormal}. Assume $\dim R/I\geqslant 3$ and $\beta(E^*)<0$, then for any $D\in \mathbb Z$, we have \begin{enumerate} \item for each $1\leqslant i\leqslant \dim R/I-2$ the natural map $$\HH{i}{\mathfrak{m}}{R/I^{n+1}}_{\leqslant D}\rightarrow\HH{i}{\mathfrak{m}}{R/I^{n}}_{\leqslant D}$$ is an isomorphism for $n\gg 0$; this map is injective for $i=\dim R/I -1$ and $n\gg 0$. \item If $R$ is Cohen-Macaulay and $\omega$ is the canonical module of $R$, then for each $\height I +2\leqslant i<\dim R$ the natural map $$\Ext_R^i(R/I^n,\omega)_{\geqslant D}\rightarrow \Ext_R^i(R/I^{n+1},\omega)_{\geqslant D}$$ is an isomorphism for $n\gg 0$. Furthermore, this map is injective for $i=\height I$ and $n\geqslant 1$. \item If $R$ is Cohen-Macaulay and $\omega$ is the canonical module of $R$, then for every $i<\dim R$ such that $i\neq \height I +1 $ the natural map $$\Ext_R^i(R/I^n,\omega)_{\geqslant D}\rightarrow \HH{i}{I}{\omega}_{\geqslant D}$$ is injective for $n\gg 0$. In fact the map is injective for $i=\height I$ and $n\geqslant 1$, and it is an isomorphism for $i\geqslant \height I+2$ and $n\gg 0$. \end{enumerate} \end{cor} \begin{proof} Part (1) follows by Theorem \ref{mainVB}, as the assumptions imply that $\indeg \HH{i}{\mathfrak{m}}{I^n/I^{n+1}} = \indeg \HH{i}{\mathfrak{m}}{S^n(E)}>D$ for $1\leqslant i\leqslant \dim R/I-1$ and $n\gg 0$. Part (2) follows from (1) and local duality for the case $i\geqslant\height I+2$. The injectivity for $i=\height I$ follows by local duality and the epimorphism $$\HH{\dim R/I}{\mathfrak{m}}{R/I^{n+1}}\twoheadrightarrow \HH{\dim R/I}{\mathfrak{m}}{R/I^{n}}$$ for $n\geqslant 1$. Now, Part (3) follows from (2) as $\HH{i}{I}{\omega}=\displaystyle\lim_{\longrightarrow}\Ext_R^i(R/I^n,\omega)$ for every $i$. \end{proof} A local ring $(S,\mathfrak{n})$ is said to be {\it cohomologically full} if for every surjection $T\twoheadrightarrow S$ from a local ring $(T,\mathfrak{q})$, such that $T_{\text{red}} = S_{\text{red}}$ and $T$ and $S$ have the same characteristic, the natural map $\HH{i}{\mathfrak{q}}{T}\rightarrow \HH{i}{\mathfrak{n}}{S}$ is surjective for every $i$. If $R$ is a standard graded algebra over a field $k$ and irrelevant maximal ideal $\mathfrak{m}$, then we say $R$ is cohomologically full if the local ring $R_\mathfrak{m}$ is. For more information and examples of cohomologically full rings see \cite{DDM}. \vspace{1mm} The following result answers Question \ref{motivQ}, (2), in a particular case. \begin{cor}\label{cohFull} Let $R$, $I$, and $E$ be as in Corollary \ref{stabExt}, and fix an integer $1\leqslant i\leqslant \dim R/I-2$. Assume $R/J$ is cohomologically full for some $R$-ideal $J$ such that $\sqrt{J}=\sqrt{I}$ and $\HH{i}{\mathfrak{m}}{R/J}\neq 0$. Then there exists an integer $C\leqslant 0$ such that $\indeg \HH{i}{\mathfrak{m}}{R/I^n}=C$ for every $n\gg 0$. \end{cor} \begin{proof} By assumption we have that the map $\HH{i}{\mathfrak{m}}{R/I^n}\rightarrow \HH{i}{\mathfrak{m}}{R/J}$ is surjective for $n\gg 0$. Therefore, $ \HH{i}{\mathfrak{m}}{R/J}$ has finite length and $$\indeg \HH{i}{\mathfrak{m}}{R/I^n}\leqslant \indeg \HH{i}{\mathfrak{m}}{R/J} =0,$$ where the last equality follows from \cite[4.9]{DDM}. Now, by Corollary \ref{stabExt}, (1) we have that $\HH{i}{\mathfrak{m}}{R/I^{n+1}}_{\leqslant 0}\rightarrow\HH{i}{\mathfrak{m}}{R/I^{n}}_{\leqslant 0}$ is an isomorphism for $n\gg 0$. The conclusion follows. \end{proof} \begin{remark} In the setting of Corollary \ref{cohFull}, let $X = \Proj R/I$. If $i\leqslant \codim \Sing X$ it was proved in \cite[3.1]{BBLSZ} that $\HH{i}{\mathfrak{m}}{R/I^n}_{<0}=0$ for every $n\geqslant 1$. Hence, if $I^n$ is cohomologically full for every $n\gg 0$, \cite[4.9]{DDM} shows $\indeg \HH{i}{\mathfrak{m}}{R/I^n}=0$ for $n\gg 0$. \end{remark} The following example shows that the assumption on the characteristic is necessary in Corollary \ref{conormal} and hence in Theorem \ref{mainVB}. \begin{example} Let $R$ and $I$ be as in Example \ref{maxMinors} but assume instead that $R$ has characteristic $p>0$. Moreover, assume the conclusion of Theorem \ref{mainVB} holds in positive characteristic. Computations by Macaulay2 \cite{GS} show $\beta((I/I^2)^*) =-1$ and then by Corollary \ref{stabExt}, (1) we have $$\HH{3}{\mathfrak{m}}{R/I^{n+1}}_0\rightarrow \HH{3}{\mathfrak{m}}{R/I^{n}}_0$$ is injective for $n\gg 0$. Furthermore, by \cite[5.5]{BBLSZ}, $\HH{3}{\mathfrak{m}}{R/I^{n}}_0\neq 0$ for every $n\geqslant 2$. However, if $n'\gg n\gg0$, there exists $e\in \mathbb N$ such that $I^{n'}\subseteq I^{[p^e]}\subseteq I^n$ and hence $\HH{3}{\mathfrak{m}}{R/I^{n'}}\rightarrow \HH{3}{\mathfrak{m}}{R/I^{n}}$ is the zero map as $R/I^{[p^e]}$ is Cohen-Macaulay, which is a contradiction. \end{example} \section{Polynomial Bound for Homogeneous Ideals}\label{polySec} Let $I$ be a homogeneous ideal in a standard graded ring over a field. The purpose of this section is to prove that whenever the modules $ \HH{i}{\mathfrak{m}}{R/I^n}$ are Noetherian for $n\gg 0$, the rate of growth of the sequence $\{\indeg \HH{i}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$ is at most polynomial. The results of this section apply in wide generality and without assumptions on the characteristic of the base field. Let $M$ be a Noetherian $R$-module of dimension $d$. We denote by $e_0(M), \ldots, e_d(M)$ the {\it Hilbert coefficients} of $M$, i.e., $$\lambda(M/\mathfrak{m}^nM)=\sum_{i=0}^r (-1)^ie_i(M){n+d-i\choose d-i},\qquad \text{ for }n\gg 0,$$ where $\lambda(N)$ denotes the {\it length} of the $R$-module $N$. \vspace{1mm} We now present the main theorem of this section. \begin{thm}\label{polBound} Let $R$ be a standard graded algebra over a field $k$ and with irrelevant maximal ideal $\mathfrak{m}$. Let $I$ be a homogeneous $R$-ideal and set $d=\dim R/I\geqslant 2$. Assume $\HH{i}{\mathfrak{m}}{R/I^n}$ is Noetherian for some $1\leqslant i< d$ and $n\gg 0$. Then there exists $s\in \mathbb N$ such that for every $n\gg 0$ we have $$|\indeg \HH{i}{\mathfrak{m}}{R/I^n}|<n^s.$$ \end{thm} \begin{proof} Let $e=\dim_k R_1$. Consider an epimorphism $S := k[x_1,\ldots, x_e]\xtwoheadrightarrow{\varphi} R$ from a polynomial ring $S$ and let $\mathfrak{n} = (x_1,\ldots,x_e)\subseteq S$. By graded local duality \cite[3.6.19]{BH} we have a graded isomorphism $$ \HH{i}{\mathfrak{m}}{R/I^n}\cong \HH{i}{\mathfrak{n}}{R/I^n} \cong \Ext^{e-i}_S(R/I^n, S)^{\vee}(e) $$ for every $n\in \mathbb N$. Therefore, by the assumption the module $\Ext^{e-i}_S(R/I^n, S)$ has finite length for $n\gg 0$, and then \begin{equation}\label{upsDo} \indeg \HH{i}{\mathfrak{m}}{R/I^n} = -\topdeg \Ext^{e-i}_S(R/I^n, S)-e= -\reg( \Ext^{e-i}_S(R/I^n, S))-e. \end{equation} For an $R$-module $M$ of dimension $r$, we set $Q_M(n)= \sum_{i=0}^r |e_i(M)|{n+r-i\choose r-i}. $ For an $R$-ideal $J$, set $\tilde{J}= (0:_{R}\mathfrak{m}^\infty)$. \vspace{1mm} Let $r_n=\reg(R/I^n) $ and notice that from the regular sequence $$0\to \tilde{I^n}/I^n\to R/I^n \to R/\tilde{I^n}\to 0$$ we obtain $\reg(R/\tilde{I^n})\leqslant r_n $. Therefore, by \cite[3.5]{CHH}, there exists $C\in \mathbb N$ such that $$\reg( \Ext^{e-i}_S(R/I^n, S))<C\big(Q_{R/I^n}(r_n)\big)^{2^d-2}.$$ By \cite[1.1]{HPV}, the functions $e_i(R/I^n)$ agree with a polynomial of degree $\leqslant e-d-i$ for every $i$ and $n\gg 0$, therefore there exists a polynomial in two variables, $q(n,t)\in \mathbb Z[n,t]$ of degree at most $e$ in $n$ and of degree $d$ in $t$, such that $Q_{R/I^n}(t)\leqslant q(t,n)$ for $t, n\gg 0$. Since $r_n$ eventually agrees with a linear function by \cite[3.2]{TW}, it follows that $$\big(Q_{R/I^n}(r_n)\big)^{2^d-2} < D n^{(e+d)(2^d-2)}$$ for some $D\in \mathbb N$ and $n\gg 0$. The conclusion now follows from \eqref{upsDo} and the fact that the sequence $\{\indeg \HH{i}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$ is bounded above by the linear function $\reg(R/I^n)$ for $n\gg 0$. \end{proof} The previous result can be used to show a polynomial bound for the lengths of local cohomology modules of powers of homogeneous ideals. The next result can be seen as a partial answer to Question \cite[7.1]{DM}. \begin{cor}\label{polynomialLength} Let $R=k[x_1,\ldots, x_d]$, $\mathfrak{m}=(x_1,\ldots, x_d)$, and $I$ a homogeneous $R$-ideal. Assume $\HH{i}{\mathfrak{m}}{R/I^n}$ is Noetherian for $n\gg 0$, then there exists $t\in \mathbb N$ such that $$\limsup_{n\rightarrow \infty} \frac{\lambda(\HH{i}{\mathfrak{m}}{R/I^n})}{n^t}<\infty.$$ \end{cor} \begin{proof} In this proof we follow Notation \ref{takNot}. By extending the field $k$ we can assume it is infinite. As in the proof of \cite[5.3]{DM}, using Gr\"obner deformation and \cite[2.4]{Sba} we can construct a sequence of monomial ideals $J_n$ such that $\reg(J_n)=\reg(I^n)$ and $\dim_k \HH{i}{\mathfrak{m}}{R/I^n}_{j}\leqslant \dim_k \HH{i}{\mathfrak{m}}{R/J_n}_{j}$ for every $j\in\mathbb Z$. Let $\beta\in \mathbb N$ be such that $\reg(I^n)\leqslant\beta n$ for every $n\geqslant 1$. By Theorem \ref{polBound} there exists $s\in \mathbb Z_{>0}$ such that $\HH{i}{\mathfrak{m}}{R/I^n}_{<-n^s} = 0$, then \begin{equation}\label{beqn1} \lambda(\HH{i}{\mathfrak{m}}{R/I^n})=\lambda(\HH{i}{\mathfrak{m}}{R/I^n}_{\geqslant -n^s})\leqslant\lambda(\HH{i}{\mathfrak{m}}{R/J_n}_{\geqslant -n^s}) \leqslant \sum_{-n^s \leqslant |\mathbf{a}| \leqslant\beta n} \dim_k \HH{i}{\mathfrak{m}}{R/J_n}_{\mathbf{a}}. \end{equation} By \cite[5.2]{DM} if $|\mathbf{a}^+|>\beta n$, then $\HH{i}{\mathfrak{m}}{R/J_n}_{\mathbf{a}}=0$. Moreover, if $\mathbf{a}=(a_1,\ldots, a_d)\in \mathbb Z^d$ satisfies that $-n^s \leqslant |\mathbf{a} | \leqslant\beta n$ and $|\mathbf{a}^+|\leqslant \beta n$, then for every $j\in G_\mathbf{a}$ we must have $a_j\geqslant -n^s-\beta n$. Set $S=G_\mathbf{a}$, then by \cite[3.3, 5.2, 5.1]{DM} there exists $C\in \mathbb N$ such that \begin{equation}\label{beqn2} \sum_{-n^s \leqslant |\mathbf{a}| \leqslant\beta n,\, G_\mathbf{a} =S} \dim_k \HH{i}{\mathfrak{m}}{R/J_n}_{\mathbf{a}}\leqslant (n^s+\beta n)^{|S|}Cn^{d-|S|}\leqslant C(n^{s+1}+\beta n^2)^{d}. \end{equation} The conclusion now follows from \eqref{beqn1} and Inequality \eqref{beqn2}, by adding the latter over all possible $S$ and setting $t=d(s+1)$. \end{proof} We finish the section with the following remark. \begin{remark} Assume $I$ is generated in a single degree $\gamma$. In \cite[3.4]{BCH} the authors showed that for every $i\geqslant 0$, there exists a set $\Lambda_i\subseteq \mathbb Z$ and a function $\eta_i:\Lambda_i\to \mathbb N$, such that $\HH{i}{\mathfrak{m}}{R/I^n}_{l+n\gamma }\neq 0$ for every $l\in \Lambda_i$ and $n\geqslant \eta_i(l)$. We note that if one is able to show that the for certain $i$ the image of the function $\eta_i$ is bounded, then the Noetherian assumption on $ \HH{i}{\mathfrak{m}}{R/I^n}$ for $n\gg 0$ would imply $\Lambda_i$ is finite, and then by \cite[3.4]{BCH} the sequence $ \{\indeg \HH{i}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$ would agree with a linear function for $n\gg 0$. \end{remark} \section{Monomial ideals}\label{monoSec} The purpose of this section is to analyze the asymptotic behavior of $\{\indeg \HH{i}{\mathfrak{m}}{R/I^n}\}_{n\in \mathbb N}$ for monomial ideals. From now on we assume $R=k[x_1,\ldots, x_d]$, $\mathfrak{m}=(x_1,\ldots, x_d)$, and $I$ is a monomial ideal. \begin{Notation}\label{takNot} Let $F$ be a subset of $[d]= \{1,\ldots, d\}$. We consider the map $\pi_F: R\longrightarrow R,$ defined by $\pi_F(x_i)=1$ if $i\in F$, and $\pi_F(x_i)=x_i$ otherwise. We set $I_F:=\pi_F(I)$. For $\mathbf{a}=(a_1,\ldots, a_d)\in \mathbb N^d$, we use the notation $\mathbf{x}^\mathbf{a} := x_1^{a_1}\cdots x_d^{a_d}$. We also consider $G_\mathbf{a}=\{i\mid a_i<0\}$ and define $\mathbf{a}^+=(a_1^+,\ldots, a_d^+)$, where $a_i^+=a_i$ if $i\not\in G_\mathbf{a} $ and $a_i^+=0$ otherwise. We set $\Delta_{\mathbf{a}}(I)$ to be the simplicial complex of all subsets $F$ of $[d]\setminus G_\mathbf{a}$ such that $\mathbf{x}^{\mathbf{a}^+}\not\in I_{F\cup G_\mathbf{a}}$. We note that $\Delta_{\mathbf{a}}(I)$ is a subcomplex of $\Delta(I)$, the simplicial complex whose Stanley-Reisner ideal is $\sqrt{I}$ (\cite[1.3]{MT}). \end{Notation} We now state Takayama's formula which expresses the graded components of local cohomology of monomial ideals in terms of reduced homology of some associated simplicial complexes. \begin{thm}[ {\cite[Theorem 1]{Tak}}]\label{Takayama} For every $\mathbf{a}\in \mathbb Z^d$ and $i\geqslant 0$ we have $$\dim_k \HH{i}{\mathfrak{m}}{R/I}_{\mathbf{a}} = \dim_k \tilde{H}_{i-|G_\mathbf{a} |-1}(\Delta_{\mathbf{a}}(I), k)$$ \end{thm} The following is the main theorem of this section. \begin{thm}\label{monomials} Let $I$ be a monomial ideal and assume $\HH{i}{\mathfrak{m}}{R/I^n}$ is Noetherian for $n\gg 0$. Then one of the following holds \begin{enumerate} \item If $\tilde{H}_{i-1}(\Delta(I),k)\neq 0$, then $\indeg \HH{i}{\mathfrak{m}}{R/I^n} =0$ for $n\gg 0$. \item If $\tilde{H}_{i-1}(\Delta(I),k) = 0$ then $\liminf_{n\rightarrow \infty}\frac{\indeg \HH{i}{\mathfrak{m}}{R/I^n} }{n}\geqslant 1$. \end{enumerate} \end{thm} \begin{proof} By \cite[Proposition 1]{Tak} and the assumption we have $\HH{i}{\mathfrak{m}}{R/I^n}_{\mathbf{a}}=0$ for $n\gg 0$, and every $\mathbf{a}\in \mathbb Z^d$ such that $G_\mathbf{a}\neq \emptyset$. In particular, $\indeg \HH{i}{\mathfrak{m}}{R/I^n} \geqslant 0$ for every $n\gg 0$. Set $\mathbf{0}=(0,\ldots,0)\in \mathbb N^d$. We note that $\Delta_{\mathbf{0}}(I^n) = \Delta(I)$ for every $n$ (\cite[1.4]{MT}), therefore if $\tilde{H}_{i-1}(\Delta(I),k)\neq 0$, Theorem \ref{Takayama} implies $\indeg \HH{i}{\mathfrak{m}}{R/I^n} =0$ for every $n\gg 0$. Now, assume $\tilde{H}_{i-1}(\Delta(I),k) = 0$. Fix $n\in \mathbb N$ and $\mathbf{a}\in \mathbb N^d$ such that $|\mathbf{a} |<n$. For every facet $F$ of $\Delta(I)$ we have $I_F^n\neq 1$, hence by degree reasons $\mathbf{x}^\mathbf{a}\not\in I_F^n$. It follows $\Delta_\mathbf{a}(I^n)=\Delta(I)$ and then $\HH{i}{\mathfrak{m}}{R/I^n}_{\mathbf{a}}=0$ by Theorem \ref{Takayama}. We conclude $\indeg \HH{i}{\mathfrak{m}}{R/I^n}\geqslant n$, finishing the proof. \end{proof} \begin{remark} The condition $\tilde{H}_{i-1}(\Delta(I),k)\neq 0$ in Theorem \ref{monomials} (1) is automatically satisfied if $\HH{i}{\mathfrak{m}}{R/I^n}$ is Noetherian for some $n\in \mathbb N $ and $\HH{i}{\mathfrak{m}}{R/\sqrt{I}}\neq 0$ (\cite[4.9]{DDM}). \end{remark} The following example answers Question \ref{motivQ}, (2) in the particular case that $\Delta(I)$ is a cycle graph $\mathcal C_d$ for $d\geqslant 5$. \begin{example} Let $d\geqslant 5$ and $\mathcal C_d$ be the cycle graph of length $d$, i.e., the edges of $\mathcal C_d$ are indexed by $\{i,i+1\}$ for $1\leqslant i\leqslant d$ where $\{d,d+1\}:=\{d,1\}$. Let $I$ be the Stanley-Reisner ideal of $R$ associated to the complex $\Delta(I)=\mathcal C_d$. Then \begin{equation}\label{limitExMon} \lim_{n\rightarrow \infty}\frac{\indeg\HH{1}{\mathfrak{m}}{R/I^n}}{n}=1. \end{equation} To show this, notice that since $\Delta(I)$ is connected, we have $\tilde{H}_0(\Delta(I))=0$. Then by Theorem \ref{monomials} it suffices to show $\displaystyle\limsup_{n\rightarrow \infty}\frac{\indeg\HH{1}{\mathfrak{m}}{R/I^n}}{n}\leqslant 1$. For each $1\leqslant i\leqslant d$, set $\mathfrak{p}_i = (\{x_j\mid j\neq i,i+1\})$. Hence, $\Ass(I)=\{\mathfrak{p}_1,\ldots, \mathfrak{p}_d\}$. Note that $$I_{\{1\}}=(x_3,\ldots,x_n)\cap(x_2,\ldots, x_{n-1})=(x_2x_n, x_3,x_4,\ldots, x_{n-1})$$ which is a complete intersection. Likewise, $I_{\{i\}}$ is a complete intersection for every $1\leqslant i\leqslant d$, therefore $\Proj R/I$ is lci. Hence, $\HH{1}{\mathfrak{m}}{R/I^n}$ is Noetherian for every $n\in \mathbb N$ and we have $$\tilde{I^n}=(I^n:_R \mathfrak{m}^\infty ) = \cap_{i=1}^d \mathfrak{p}_i^n .$$ Fix $n\gg 0$ and let $\mathbf{a}_n = (n-d+4, 0, 1, \ldots, 1, 0)\in \mathbb N^d $, then one readily verifies $\Delta_{\mathbf{a}_n}(\tilde{I^n})$ is the subcomplex of $\Delta(I)$ whose facets are $\{i,i+1\}$ for $i\neq 2, d-1$. Since $\Delta_{\mathbf{a}_n}(\tilde{I^n})$ is disconnected, we have $\dim_k \HH{1}{\mathfrak{m}}{R/\tilde{I^n}}_{\mathbf{a}_n} = \dim_k \tilde{H}_0(\Delta_{\mathbf{a}_n}, k)\neq 0$. We conclude $$\indeg \HH{1}{\mathfrak{m}}{R/\tilde{I^n}}\leqslant |\mathbf{a}_n|=n+1.$$ Finally, the limit \eqref{limitExMon} follows by noticing $\HH{1}{\mathfrak{m}}{R/I^n} \cong \HH{1}{\mathfrak{m}}{R/\tilde{I^n}}$. \end{example} \section*{Acknowledgments} The authors are grateful to David Eisenbud, Jack Jeffries, Robert Lazarsfeld, Luis N\'u\~nez-Betan\-court, and Claudiu Raicu for very helpful discussions. They also thank Tai H\`a for bringing the reference \cite{BCH} to their attention. Part of the research included in this article was developed in the Mathematisches Forschungsinstitut Oberwolfach (MFO) while the authors were in residence at the institute under the program {\it Oberwolfach Leibniz Fellows}. The authors thank MFO for their hospitality and excellent conditions for conducting research. The authors would also like to thank the referee for her or his helpful comments and suggestions that improved this paper.

64bd2c4a066a33b97c7278e3e43bf25cf63a86ad

\section{Introduction}\label{intro} Hamiltonian systems of differential equations are of widespread importance in physics and mathematics. There is particular interest in systems that are superintegrable in the sense of Liouville by possessing the maximal number of globally well-defined first integrals which are functionally independent. Noether's theorem provides an obvious connection between first integrals and local symmetries of a given Hamiltonian system, whether or not the system is (super)integrable. Specifically, using the Lagrangian formulation of the system, each first integral corresponds to a variational local symmetry, which can be either a point symmetry or a dynamical symmetry \cite{BA-book,Olver-book}. Point symmetries are distinguished by involving only the canonical variables and the time variable; most importantly, all point symmetries can be obtained systematically for any Hamiltonian system through the use of Lie's method \cite{BA-book,Olver-book}. A natural first question is: \emph{To what extent can (super)integrability of a Hamiltonian system be detected just by looking at its point symmetries?} The answer is, in general, that variational point symmetries do not always provide a sufficient number of first integrals. A widely studied example is central force motion in the Newtonian case of an inverse-square force law (see e.g.~\cite{GolPooSaf}). The variational point symmetries of this Hamiltonian system consist of rotations and time-translation, which yield the components of the angular momentum vector and the energy as first integrals. There are additional first integrals given by the components of the well-known Laplace-Runge-Lenz (LRL) vector. Recall that this vector lies in the plane orthogonal to the angular momentum vector and is oriented in the direction of the apsis line from the center of mass to the apsis point on any non-circular orbit. The first integrals corresponding to the angle determined by the LRL vector arise from ``hidden'' dynamical symmetries \cite{BacRueSou,Fra,AncMeaPas} rather than point symmetries of the \eom/. A second question then is: \emph{Can (super)integrability of a Hamiltonian system be detected by knowing its dynamical symmetries?} The answer, surprisingly, is neither simple nor universal. Because variational dynamical symmetries correspond to first integrals by Noether's theorem, these local symmetries do contain some information about the first integrals. In some situations, the global form of the symmetry group transformations acting on solutions may indicate if the first integrals are globally single-valued and non-singular. But, in general, this global question about first integrals is distinct from the properties of the variational symmetry group, because the existence and nature of local symmetries depends solely on the local structure of the \eom/. The example of central force motion with a general radial potential is a nice illustration of the subtleties. It has been known for several decades that an analog of the LRL vector exists for any radial potential \cite{BacRueSou,Fra,AncMeaPas}, but the resulting first integrals given by the components of this generalized vector are globally single-valued and non-singular only for the Kepler-Coulomb potential and the isotropic oscillator potential \cite{BacRueSou}. Namely, those are the only two central force systems that are superintegrable in Euclidean space. (For the situation in curved spaces with radial symmetry, see~\cite{commun}.) When any other central force system is considered in Euclidean space, the generalized LRL vector instead is multi-valued and jumps each time the apsis point on a non-circular orbit is reached \cite{SerSha,BucDen}. Moreover, the ``hidden'' dynamical symmetries that correspond to these first integrals have the same symmetry algebra \cite{BacRueSou,Fra} for all radial potentials, whether the central force system is superintegrable or not. As a consequence, superintegrability is not related to either the size or the algebra structure of the variational dynamical symmetries. Of course, if all variational dynamical symmetries of a given Hamiltonian system are known, then the first integrals can be obtained in an explicit form through Noether's theorem, so that their global properties then can be studied. But a priori it is not possible to find all variational dynamical symmetries without in essence integrating the \eom/ of the Hamiltonian system, and this task involves the same level of difficulty as directly finding all first integrals \cite{BA-book}. More specifically, it is no easier to find all variational dynamical symmetries than it is to find all first integrals. Nevertheless, there is an extended symmetry method developed in Ref.~\cite{AncMeaPas} that can be used to obtain all first integrals systematically for many Hamiltonian systems. The purpose of the present paper is to get a deeper insight into the connection between first integrals and local symmetries by studying a superintegrable Hamiltonian system introduced in Ref.~\cite{DarbouxPhysD}. This system describes a radially symmetric nonlinear oscillator which is physically interesting both from dynamical and geometrical viewpoints, since it can be identified both with the motion of a particle on a space having non-constant curvature, and also with an oscillator whose mass is position dependent. The main results in the paper will be to show how to go systematically from the Hamiltonian system to local symmetries and then to first integrals, and reciprocally, from the Hamiltonian system directly to first integrals and then to local symmetries. This will be accomplished without the need for ansatzes or guess-work by adapting the extended symmetry method from \cite{AncMeaPas}. As will be seen, superintegrability of the system is not related in any straightforward or universal way to either its point symmetries or its variational symmetry algebra. These results will reinforce the preceding discussion. In particular, on one hand, point symmetries are generally insufficient to characterize when a Hamiltonian system is superintegrable, and on the other hand, the size and structure of the variational symmetry algebra of a given Hamiltonian system is not enough to detect if the system is superintegrable. \section{Lie point symmetries and (super)integrability}\label{toymodels} In this section, the basic Hamiltonian systems for two uncoupled oscillators and for central force planar motion are studied as benchmark models to discuss and clarify the questions raised in section~\ref{intro}. Recall that the first system is superintegrable provided that the two oscillator frequencies are commensurate, and that the second system is superintegrable only for the Kepler-Coulomb potential and the isotropic oscillator potential. This situation is ideal for analyzing the difficulties in detecting superintegrability through point symmetries, since the superintegrable systems are specific cases belonging to the respective family of oscillator and central force potentials. \subsection{Uncoupled oscillators}\label{uncoupledoscil} The Hamiltonian system describing two uncoupled oscillators is given by the \eom/ \begin{equation*} \ddot q_1 +\omega_1{}^2 q_1=0, \qquad \ddot q_2 +\omega_2{}^2 q_2=0, \end{equation*} for $(q_1(t),q_2(t))$, where $\omega_1$ and $\omega_2$ are the frequencies. The Lagrangian is ${\mathcal L} = \tfrac{1}{2}(\dot q_1^2 +\dot q_2^2 - \omega_1{}^2 q_1^2 - \omega_2{}^2 q_2^2)$. \\ \emph{First integrals}: All constants of motion (\com/) $I(q_1,q_2,\dot q_1,\dot q_2)$ arise from the determining equation \begin{equation*} 0=\dot I =I_{q_1} \dot q_1 +I_{q_2} \dot q_2 -\omega_1{}^2 q_1 I_{\dot q_1} -\omega_2{}^2 q_2 I_{\dot q_2}, \end{equation*} which can be solved easily by the method of characteristics. A maximal set of three functionally-independent \com/ is given by the energies of the two oscillators \begin{equation}\label{uncoupledoscil-com-E} E_1= \tfrac{1}{2}( \dot q_1^2 +\omega_1{}^2 q_1^2 ), \quad E_2= \tfrac{1}{2}( \dot q_2^2 +\omega_2{}^2 q_2^2 ) , \end{equation} and a phase quantity \begin{equation}\label{uncoupledoscil-com-Phi} \Phi= (1+\tfrac{\omega_2}{\omega_1})\arctan(\omega_1 q_1/\dot q_1) -(1+\tfrac{\omega_1}{\omega_2})\arctan(\omega_2 q_2/\dot q_2) \ \mod \pi. \end{equation} This quantity can be shown to describe the difference in the relative phase shifts $\Delta\phi_1$ and $\Delta\phi_2$ between the two oscillators measured at the times $t_1$ and $t_2$ when each one of oscillators passes through zero, namely $\Phi=\Delta\phi_2-\Delta\phi_1$. It undergoes a jump each time one of the oscillators changes direction. Thus, in general the \com/~$\Phi$ is multi-valued. In contrast, the \com/~$E_1$, $E_2$, and $E=E_1+E_2$ (total energy of the oscillators) are single-valued. The oscillator system is superintegrable iff the frequencies of two oscillators are commensurate: $\omega_1/\omega_2\in\mathbb Q$. In the superintegrable case, the set of values of the \com/ $\Phi$ is finite, whereas for incommensurate frequencies, this set of values is infinite. In both cases, $\Phi$ is always non-singular. In addition to the three functionally-independent \com/ $E_1$, $E_2$, $\Phi$, there is a first integral that depends explicitly on $t$: \begin{equation} T= t- \Big(\tfrac{1}{2\omega_1}\arctan(\omega_1 q_1/\dot q_1) +\tfrac{1}{2\omega_2}\arctan(\omega_2 q_2/\dot q_2)\Big) \end{equation} This first integral yields the average of the times at which the respective oscillators pass through zero. \\ \emph{Variational symmetries}: According to Noether's theorem, each of the first integrals $E_1$, $E_2$, $\Phi$, $T$ corresponds to a variational symmetry. The evolutionary form of these symmetries, acting on $(q_1(t),q_2(t))$, is given by the generators \begin{gather*} \hat\mathbf{X}_{E_1}= \dot q_1\partial/\partial q_1, \quad \hat\mathbf{X}_{E_2}= \dot q_2\partial/\partial q_2, \\ \hat\mathbf{X}_{\Phi}=\tfrac{\omega_1+\omega_2}{2}( \tfrac{1}{E_2} q_2\partial/\partial q_2 -\tfrac{1}{E_1} q_1\partial/\partial q_1 ), \quad \hat\mathbf{X}_{T}=\tfrac{1}{2E_1} q_1\partial/\partial q_1 +\tfrac{1}{2E_2} q_2\partial/\partial q_2 . \end{gather*} By comparison, an infinitesimal point symmetry $t\to t + \epsilon \tau(t,q_1,q_2) +O(\epsilon^2)$, $q_1\to q_1+\epsilon \eta_1(t,q_1,q_2) +O(\epsilon^2)$, $q_2\to q_2+\epsilon \eta_2(t,q_1,q_2) +O(\epsilon^2)$ in evolutionary form has the generator $\hat\mathbf{X}=(\eta_1-\tau\dot q_1)\partial/\partial_{q_1}+ (\eta_2-\tau\dot q_2)\partial/\partial_{q_2}$. It is straightforward to see that none of the four symmetries $\hat\mathbf{X}_{E_1}$, $\hat\mathbf{X}_{E_2}$, $\hat\mathbf{X}_{\Phi}$, $\hat\mathbf{X}_{T}$ represent point symmetries, due to the form of their dependence on $\dot q_1$ and $\dot q_2$ through the expressions \eqref{uncoupledoscil-com-E}--\eqref{uncoupledoscil-com-Phi} for $E_1,E_2,\Phi$. Therefore, they are dynamical symmetries. On solutions of the \eom/, the symmetry generators are mutually commuting, namely the symmetry algebra is abelian. Moreover, none of the symmetries contain information about superintegrability of the \eom/. In particular, the components of each symmetry generator are single-valued and non-singular for arbitrary frequencies $\omega_1\neq0$, $\omega_2\neq0$.\\ \emph{Point symmetries}: The point symmetries of the uncoupled oscillator system for general frequencies $\omega_1$ and $\omega_2$ are generated by a time translation $\mathbf{X}_{\rm trans}= \partial/\partial t$ and two scalings $\mathbf{X}_{{\rm scal}_1}= q_1\partial/\partial q_1$ and $\mathbf{X}_{{\rm scal}_2}= q_2\partial/\partial q_2$, along with elementary symmetries $\mathbf{X} = f_1(t)\partial/\partial_{q_1} + f_2(t)\partial/\partial_{q_2}$, where $f_j=a_j\cos(\omega_jt +\phi_j)$, $j=1,2$, are arbitrary solutions of the \eom/. Note that time translation is a variational symmetry whose evolutionary form is given by $\hat\mathbf{X}_{\rm trans} =\hat\mathbf{X}_{E_1} +\hat\mathbf{X}_{E_2} = \hat\mathbf{X}_{E}$. Additional point symmetries arise only when the two frequencies are equal, $\omega_1=\omega_2=\omega$. In this special case, there are four additional point symmetries, which are generated by \begin{align*} & \mathbf{X}_{\rm rot}= q_2\partial/\partial q_1 - q_1\partial/\partial q_2, \\ & \mathbf{X}_1 = e^{\pm i\omega t}( q_1\partial/\partial q_1 + q_2\partial/\partial q_2 \mp\tfrac{i}{\omega} \partial/\partial_t ), \\ & \mathbf{X}_2 = e^{\pm i\omega t}q_1(q_1\partial/\partial q_1 + q_2\partial/\partial q_2 \mp\tfrac{i}{\omega} \partial/\partial_t ), \\ & \mathbf{X}_3 = e^{\pm i\omega t}q_2(q_1\partial/\partial q_1 + q_2\partial/\partial q_2 \mp\tfrac{i}{\omega} \partial/\partial_t ). \end{align*} Hence, because $\omega_1=\omega_2=\omega$ belongs to the case of commensurate frequencies, this special superintegrable case has a larger point symmetry group. However, in all other superintegrable cases, for which $\omega_1/\omega_2\in \mathbb Q$ with $\omega_1\neq\omega_2$, the point symmetry group of the system has the same size as in the general non-superintegrable case. \subsection{Central force motion}\label{centralforce} For any central force, motion in the plane orthogonal to the conserved angular momentum vector is given by the Hamiltonian system \begin{equation*} \ddot r=\dot \theta^2 r -U'(r), \quad \ddot \theta=-2\dot \theta \dot r/r \end{equation*} in polar coordinates $(r(t),\theta(t))$, where $U(r)$ is the potential and ${\mathcal L} = \tfrac{1}{2}( \dot r^2 + \dot\theta^2 r^2 ) -U(r)$ is the Lagrangian.\\ \emph{First integrals}: All \com/ $I(r,\theta,\dot r,\dot \theta)$ arise from the determining equation \begin{equation*} 0=\dot I =I_{r} \dot r +I_{\theta} \dot \theta +(\dot\theta^2 r -U'(r)) I_{\dot r} -2\dot\theta r^{-1}\dot r I_{\dot \theta} . \end{equation*} This is a first-order linear PDE for $I$, which can be explicitly solved by the method of characteristics and yields a maximal set of three functionally-independent \com/: \begin{gather} L=\dot\theta r^2, \quad E= \tfrac{1}{2}( \dot r^2 +L^2/r^2) +U(r), \label{centralforce-LE} \\ \Theta= \theta-L\int_{r_0}^r \frac{{\rm sgn}(\dot r)}{r\sqrt{2(E+U(r_{\rm equil}) - U(r))r^2 -L^2}}dr \ \mod 2\pi, \label{centralforce-Theta} \end{gather} where $r_{\rm equil}$ is any equilibrium point, $U'(r_{\rm equil})=0$. For any solution $(r(t),\theta(t))$ of the \eom/, $L$ is the planar angular momentum; $E$ is the energy (Hamiltonian); and $\Theta$ is the angle reached at some point $r=r_0$. As shown in \cite{AncMeaPas}, a natural intrinsic choice of $r_0$ is any turning point $r^*$ or any inertial point $r_*$, which are given by $U_{\rm eff}(r^*)=E$ or $U_{\rm eff}'(r_*)=0$ in terms of the effective potential $U_{\rm eff}(r)=U(r) +\tfrac{1}{2}L^2/r^2 - U(r_{\rm equil})$. In addition to the three functionally-independent \com/~ $L$, $E$, $\Theta$, there is a first integral that depends explicitly on $t$: \begin{equation}\label{centralforce-T} T = t- \int^r_{r_0} \frac{{\rm sgn}(\dot r)}{\sqrt{2(E+U(r_{\rm equil})-U(r)) -L^2/r^2}}\,dr . \end{equation} It is well known that the central force system is superintegrable iff $U(r)=-k/r$ is the Coulomb potential or $U(r)=k r^2$ is the isotropic oscillator potential. Thus, $\Theta$ is single-valued and non-singular in these two cases. In particular, if $r_0=r^*$ is a turning point at which $r(t)$ reaches a local maximum, then as shown in \cite{AncMeaPas}, $\Theta$ is the angle of the LRL vector, which is a \com/ for the Coulomb potential and the isotropic oscillator potential. In these two cases, $T$ is the time at which $\theta(t)$ coincides with the LRL angle, modulo the period of the solution $(r(t),\theta(t))$. As a consequence, all bounded orbits for both of these potentials do not precess. For any other central force system, we can infer that $\Theta$ is multi-valued and possibly singular. In particular, the angle $\Theta$ defining the generalized LRL vector undergoes a jump each time $t=T$ when $r$ reaches a turning point. An example is the perturbed Coulomb potential $U(r)=-k/r-K/r^2$, where bounded non-circular orbits exhibit precession \cite{GolPooSaf}, and thus this central force system is not superintegrable. \\ \emph{Variational symmetries}: By Noether's theorem, each of the \com/~$L$, $E$, $\Theta$ corresponds to a variational symmetry. In evolutionary form, acting on $(r(t),\theta(t))$, these symmetries are given by the generators \cite{AncMeaPas} \begin{gather*} \hat\mathbf{X}_{L}= -\partial/\partial \theta, \quad \hat\mathbf{X}_{E}= -\dot r\partial/\partial r -\dot\theta \partial/\partial \theta, \\ \hat\mathbf{X}_{\Theta}=-(\dot r \partial_E\Theta) \partial/\partial r -(\partial_L\Theta +\dot\theta\partial_E \Theta)\partial/\partial \theta, \quad \hat\mathbf{X}_{T}=-(\dot r \partial_E T) \partial/\partial r -(\partial_L T +\dot\theta\partial_E T)\partial/\partial \theta . \end{gather*} Both $\hat\mathbf{X}_{L}$ and $\hat\mathbf{X}_{E}$ represent infinitesimal point symmetries, as can be easily seen by comparison with $\hat\mathbf{X}=(\eta^r -\tau\dot r)\partial/\partial{r}+ (\eta^\theta-\tau\dot \theta)\partial/\partial{\theta}$ which is general evolutionary form for an infinitesimal point symmetry $t\to t + \epsilon \tau(t,r,\theta) +O(\epsilon^2)$, $r\to r+\epsilon \eta^r(t,r,\theta) +O(\epsilon^2)$, $\theta\to \theta+\epsilon \eta^\theta(t,r,\theta) +O(\epsilon^2)$. In contrast, $\hat\mathbf{X}_{\Phi}$ and $\hat\mathbf{X}_{T}$ do not represent point symmetries but instead are dynamical symmetries, because of their nonlinear dependence on $\dot r$ and $\dot \theta$ through the expressions \eqref{centralforce-LE} for $L,E$. A straightforward computation shows that, on solutions of the \eom/, the symmetry algebra is abelian, namely, the four generators $\hat\mathbf{X}_{L}$, $\hat\mathbf{X}_{E}$, $\hat\mathbf{X}_{\Phi}$, $\hat\mathbf{X}_{T}$ are mutually commuting. Clearly, the point symmetries $\mathbf{X}_{L}$ and $\mathbf{X}_{E}$ contain no information about superintegrability of the \eom/, since they are admitted for an arbitrary central force potential $U(r)$. An interesting question is whether the dynamical symmetry $\mathbf{X}_{\Theta}$ contains any information about superintegrability. To answer this, we need to examine the components \begin{equation}\label{centralforce_Theta_components} \begin{aligned} \partial_E \Theta & = L\int_{r_0}^r \frac{{\rm sgn}(\dot r)r}{\sqrt{2(E+U(r_{\rm equil}) - U(r))r^2 -L^2}^3}dr, \\ \partial_L\Theta & = \int_{r_0}^r \frac{{\rm sgn}(\dot r)2(U(r)-U(r_{\rm equil})-E)r}{\sqrt{2(E+U(r_{\rm equil}) - U(r))r^2 -L^2}^3}dr. \end{aligned} \end{equation} In the cases of the Coulomb potential $U(r)=-k/r$ and the isotropic oscillator potential $U(r)=k r^2$, we find that both components \eqref{centralforce_Theta_components} are single-valued but become singular at turning points. For the case of the perturbed Coulomb potential $U(r)=-k/r-K/r^2$, we find that the component $\partial_L\Theta$ is not single-valued. Consequently, in these three examples, the form of the dynamical symmetry $\mathbf{X}_{\Theta}$ detects if the central force system is superintegrable. \\ \emph{Point symmetries}: The point symmetries of the central force \eom/ are generated by $\mathbf{X}_{L}= -\partial/\partial \theta$ and $\mathbf{X}_{E}= \partial/\partial t$ for a general potential $U(r)$. Additional point symmetries are admitted only for \cite{AncMeaPas} two special potentials: $U(r)=kr^p$, which admits $\mathbf{X}_1 = t\partial/\partial t -\tfrac{2}{p}r\partial/\partial r$; $U(r)=kr+K/r^3$, which admits $\mathbf{X}_2 = e^{2\sqrt{k}t}(\partial/\partial t +\sqrt{k}r\partial/\partial r)$. Notice that the point symmetry group is \emph{not larger} in the superintegrable cases. \section{Connections among first integrals, symmetries, and superintegrability} We will study the $N=2$ version of the dynamical system given by the Hamiltonian \begin{equation}\label{ND-Ham} H(\mathbf{q},\mathbf{p}) =\tfrac{1}{2}(1+\lambda \mathbf{q}^2)^{-1} ( \mathbf{p}^2 + \omega^2 \mathbf{q}^2) \end{equation} where $\lambda>0$ and $\omega> 0$ are real parameters, and $(\mathbf{q},\mathbf{p})$ are $2N$ canonical coordinates. This system was proven in~Ref.~\cite{DarbouxPhysD} to be maximally superintegrable, namely it possesses the maximum number $(2N-1)$ of \com/, which are functionally independent and globally well-defined for a general solution $(\mathbf{q}(t),\mathbf{p}(t))$. These \com/ are explicitly given by \begin{equation}\label{ND-com} \begin{gathered} C^{(m)}=\!\! \sum_{1\leq i<j\leq m} \!\!\!\! (q_ip_j-q_jp_i)^2 , \quad C_{(m)}=\!\!\! \sum_{N-m<i<j\leq N}\!\!\!\!\!\! (q_ip_j-q_jp_i)^2 , \quad m=2,\dots,N, \\ E_i=p_i^2-\bigl(2\lambda H(\mathbf{q},\mathbf{p})-\omega^2\bigr) q_i^2 ,\quad i=1,\dots,N. \end{gathered} \end{equation} Some non-local symmetries for the $N=1$ case were found in~\cite{GandariasPLA}. When $\lambda\to 0$, this system~\eqref{ND-Ham} reduces to the $N$-dimensional Euclidean isotropic oscillator with frequency $\omega$, which is indeed a maximally superintegrable system. Thus, $\lambda$ can be viewed as a deformation parameter, and the system~\eqref{ND-Ham} can be thought of as a maximally superintegrable deformation of the Euclidean isotropic oscillator. Geometrically, the term $\tfrac{1}{2}(1+\lambda \mathbf{q}^2)^{-1}\mathbf{p}^2$ in the Hamiltonian $H(\mathbf{q},\mathbf{p})$ can be interpreted as the kinetic energy defined by the geodesic motion of a particle with unit mass on a conformally flat space whose metric is given by ${\rm d} s^2= (1+\lambda \mathbf{q}^2){\rm d} \mathbf{q}^2$ (see also~\cite{KKMW02}--\cite{Gonera}). The scalar curvature of this space is negative and asymptotically vanishes for large $|\mathbf{q}|$. Further discussion on the geometrical interpretation of the system and its quantization can be found in~\cite{commun} and~\cite{NDAnnals}--\cite{BGSNpla}. From a physical viewpoint, the system describes a particle with position-dependent mass of the form $m(\mathbf{q})=1+\lambda \mathbf{q}^2$. We recall that the quantum version of such systems (see for instance~\cite{Roos}--\cite{MR} and references therein) is relevant for the description of semiconductor heterostructures and nanostructures and, in particular, models constructed in terms of quadratic mass functions have been considered in~\cite{Koc, Schd}. Our main result will be to show systematically how to derive the local symmetry group underlying the \com/ ~\eqref{ND-com} of this system in the case $N=2$. We do this derivation in two different ways. First, we directly integrate the determining equation for first integrals, and then we apply Noether's theorem (in reverse) to obtain the corresponding variational symmetries. This process can be summarized as \emph{Hamiltonian system $\Rightarrow$ first integrals $\Rightarrow$ local symmetries}. Next, and most importantly, we show how to use the symmetry method outlined in Ref.\cite{AncMeaPas} to do the reverse process: \emph{Hamiltonian system $\Rightarrow$ local symmetries $\Rightarrow$ first integrals}. This method is systematic and explicit, and no ansatzes are needed. \subsection{From symmetries to first integrals} The Hamiltonian \eqref{ND-Ham} in the planar case $N=2$ is given by \begin{equation}\label{Ham} H = \frac{p_1^2 +p_2^2 +\omega^2(q_1^2+q_2^2)}{2(1+\lambda(q_1^2+q_2^2))}. \end{equation} Note, when the deformation parameter $\lambda$ is taken to be $\lambda=0$, this Hamiltonian reduces to the one for the planar isotropic oscillator, which is given by oscillator system discussed in Section~\ref{uncoupledoscil} in the case $\omega_1=\omega_2=\omega$. Hereafter it will be useful to change from planar coordinates to polar coordinates \begin{equation} q_1=r\sin\theta, \quad q_2=r\sin\theta, \quad p_1 = p^r\cos\theta +p^\theta r^{-1}\sin\theta, \quad p_2 = p^r\sin\theta-p^\theta r^{-1}\cos\theta. \end{equation} The Hamiltonian \eqref{Ham} becomes \begin{equation} H= \frac{(p^r)^2 + (p^\theta/r)^2 +\omega^2 r^2}{2(1+\lambda r^2)} , \end{equation} which yields the second-order ODE system \begin{equation}\label{eom} \ddot r = f^r(r,\theta,\dot r,\dot\theta) = \frac{((2\lambda r^2 +1)\dot \theta^2 +\lambda \dot r^2)r}{\lambda r^2+1} -\frac{\omega^2 r}{(\lambda r^2+1)^3}, \quad \ddot\theta = f^\theta(r,\theta,\dot r,\dot\theta) =-\frac{2\dot \theta \dot r (2\lambda r^2 +1)}{(\lambda r^2 +1)r}. \end{equation} This system is superintegrable. Hereafter, the set of solutions $(r(t),\theta(t))$ of the \eom/ will be denoted ${\mathcal E}$. \subsection{From first integrals to symmetries} A first integral is a function $I$ of $t,r,\theta,\dot r,\dot \theta$ that is time-independent, $\dot I=0$, when it is evaluated on the solution space ${\mathcal E}$ of the \eom/. If $I$ does not depend explicitly on $t$, then it is a \com/. All first integrals can be found by solving the determining equation \begin{equation}\label{I-deteqn} 0=\dot I(t,r,\theta,\dot r,\dot \theta)\big|_{\mathcal E} =I_{t} + I_{r} \dot r +I_{\theta} \dot\theta +f^r I_{\dot r} +f^\theta I_{\dot \theta}, \end{equation} which is a linear first-order PDE for $I(t,r,\theta,\dot r,\dot \theta)$. Solving this PDE amounts to integrating the \eom/ \eqref{eom}. This can be done by applying the method of characteristics, yielding the ODE system $dt/1=dr/\dot r=d\theta/\dot\theta=d\dot r/f^r=d\dot\theta/f^\theta$. Integration of the system gives four functionally-independent first integrals: \begin{align} & L= r^2(1+\lambda r^2) \dot\theta =p^\theta, \label{L} \\ & E= \tfrac{1}{2}\big( (1+\lambda r^2) \dot r^2 + (\omega^2 r^2 +L^2/r^2)(1+\lambda r^2) ^{-1} \big) =H, \label{E} \\ & \begin{aligned} \Theta & = \theta - \tfrac{1}{2}\arctan\Big(\frac{{\rm sgn}(\dot r)(Er^2-L^2)}{L\sqrt{2Er^2(1+\lambda r^2)-L^2-\omega^2r^4}}\Big)\Big|^{r}_{r_0} \mod 2\pi, \end{aligned} \label{Theta} \\ & \begin{aligned} T & = t -\tfrac{1}{2}\frac{w^2-\lambda E}{\sqrt{w^2-2\lambda E}^3}\bigg(\arctan\Big(\frac{{\rm sgn}(\dot r)\big( (2\lambda r^2+1)E-\omega^2 r^2\big)}{\sqrt{w^2-2\lambda E}\sqrt{2Er^2(1+\lambda r^2)-L^2-\omega^2r^4}}\Big) \Big|^{r}_{r_0} \bigg) \\&\qquad +\tfrac{1}{2}\frac{\lambda}{w^2-2\lambda E}\Big( {\rm sgn}(\dot r)\sqrt{2Er^2(1+\lambda r^2)-L^2-\omega^2r^4}\Big)\Big|^{r}_{r_0} . \end{aligned} \label{T} \end{align} Here $L$ is the planar angular momentum and $E$ is the energy, which respectively arise from solving $dr/\dot r=d\dot\theta/f^\theta$ and $dr/\dot r=d\dot r/f^r$; $\Theta$ is an angular quantity given by solving $dr/\dot r=d\theta/\dot\theta$ and involves an arbitrary constant of integration $r_0=r(t_0)$. These quantities are \com/, while $T$ is a temporal first integral arising from $dt/1=dr/\dot r$. The physical meaning of $\Theta$ and $T$ is similar to the analogous quantities that appear in the superintegrable cases of central force motion discussed in Section~\ref{centralforce}. Notice that $L$, $E$, $\Theta$, $T$ are functionally independent because they each have different physical units. A natural physical choice of $r_0$ is any turning point $r^*$ or any inertial point $r_*$, which are respectively given by $U_{\rm eff}(r^*)=E$ or $U_{\rm eff}'(r_*)=0$ in terms of the effective potential \begin{equation} U_{\rm eff}(r)=\frac{\omega^2 r^2 +L^2/r^2}{1+\lambda r^2}. \end{equation} On the orbit of a solution $(r(t),\theta(t))$, a turning point is thus a point $r=r^*$ at which the radial velocity $\dot r=0$, and an inertial point is a point $r=r_*$ at which the radial acceleration $\ddot r=0$. These points are determined intrinsically by the dynamics of each solution $(r(t),\theta(t))$. With such a choice of $r_0$, the angular quantity $\Theta$ physically represents the angle $\theta$ on the orbit of a solution $(r(t),\theta(t))$ at the point $r=r_0$ given by either a turning point $r_0=r^*$ or an inertial point $r_0=r_*$. Likewise, the temporal quantity $T$ physically represents the time $t$ at which this point is reached on the orbit. Superintegrability of the system \eqref{eom} is distinguished by the feature that both $\Theta$ and $T$ are single-valued and non-singular for a general solution $(r(t),\theta(t))$. Consequently, $\Theta$ provides an analog of the LRL angle. All first integrals $I(t,r,\theta,\dot r,\dot\theta)$ are associated to multiplier pairs $(Q^r(r,\theta,\dot r,\dot\theta),Q^\theta(r,\theta,\dot r,\dot\theta))$ given by expressing the conservation property $\dot I|_{\mathcal E}=0$ as an identity \begin{equation*} \dot I = (\ddot r - f^r)Q^r + (\ddot \theta - f^\theta)Q^\theta, \qquad Q^r=\partial_{\dot r}I, \quad Q^\theta=\partial_{\dot\theta}I \end{equation*} holding off of the solution space ${\mathcal E}$. Multiplier pairs are directly related to variational symmetries through Noether's theorem using the Lagrangian formulation of the system \eqref{eom} as follows. The Lagrangian is given by ${\mathcal L} = \tfrac{1}{2}(1+\lambda r^2)( \dot r^2 + \dot\theta^2 r^2 ) -\tfrac{1}{2}w^2r^2/(1+\lambda r^2)$, which yields \begin{equation*} \ddot r -f^r =-(1+\lambda r^2)^{-1}\frac{\delta{\mathcal L}}{\delta r}, \qquad \ddot \theta -f^\theta =-(r^2 (1+\lambda r^2))^{-1}\frac{\delta{\mathcal L}}{\delta \theta} . \end{equation*} Now consider any vector field $\hat\mathbf{X} = P^r(t,r,\theta,\dot r,\dot\theta)\partial_r + P^\theta(t,r,\theta,\dot r,\dot\theta)\partial_\theta$ in evolutionary form which acts only on the coordinates $(r,\theta)$. This vector field induces a variation of the Lagrangian, yielding Noether's identity \begin{equation*} {\rm pr}^{(1)}\hat\mathbf{X} ({\mathcal L}) = \frac{\delta{\mathcal L}}{\delta r} P^r + \frac{\delta{\mathcal L}}{\delta\theta} P^\theta + \frac{d}{dt}( P^r\partial_{\dot r}{\mathcal L} + P^\theta \partial_{\dot\theta}{\mathcal L}), \end{equation*} where ${\rm pr}^{(1)}\hat\mathbf{X} = P^r\partial_{r} + P^\theta\partial_{\theta} + \dot P^r\partial_{\dot r} + \dot P^\theta\partial_{\dot\theta}$ is the prolongation of $\hat\mathbf{X}$ to the coordinate space $(r,\theta,\dot r,\dot\theta)$. The condition for the vector field to be a variational symmetry is that the induced variation of the Lagrangian is a total time derivative, ${\rm pr}^{(1)}\hat\mathbf{X}({\mathcal L}) = \dot R$, for some function $R(t,r,\theta,\dot r,\dot\theta)$. This implies \begin{equation*} (\ddot r -f^r)\big( (1+\lambda r^2)P^r \big) + (\ddot \theta -f^\theta)\big( r^2(1+\lambda r^2)P^\theta \big) = \frac{d}{dt}\big( P^r\partial_{\dot r}{\mathcal L} +P^\theta \partial_{\dot\theta}{\mathcal L} -R \big) . \end{equation*} When this equation is evaluated on solutions $(r(t),\theta(t))$ of the \eom/, it yields a first integral \begin{equation}\label{I} I=R - P^r\partial_{\dot r}{\mathcal L} -P^\theta \partial_{\dot\theta}{\mathcal L} \end{equation} which has the multiplier pair \begin{equation}\label{QfromP} Q^r = -(1+\lambda r^2)P^r , \qquad Q^\theta = -r^2(1+\lambda r^2)P^\theta . \end{equation} This relation establishes a one-to-one correspondence between multiplier pairs $(Q^r,Q^\theta)$ and components $P^r$, $P^\theta$ of variation symmetries. Specifically, any variational symmetry yields a first integral whose corresponding multiplier is determined by equation \eqref{QfromP} in terms of the components of the symmetry generator; conversely, any first integral yields a variational symmetry whose generator has components determined in terms of the multiplier through inverting equation \eqref{QfromP} to get \begin{equation}\label{PfromQ} P^r = \frac{-Q^r}{1+\lambda r^2}, \qquad P^\theta = \frac{-Q^\theta}{r^2(1+\lambda r^2)} . \end{equation} These correspondences \eqref{QfromP}--\eqref{PfromQ}, along with the first integral expression \eqref{I}, constitute the statement of Noether's theorem. The multiplier pairs for the four first integrals \eqref{L}--\eqref{T} are given by \begin{align} & (Q^r,Q^\theta)_L = \big(0,r^2(\lambda r^2+1)\big), \qquad (Q^r,Q^\theta)_E = \big((\lambda r^2+1)\dot r,r^2(\lambda r^2+1)\dot \theta\big), \label{Q-L-E} \\ & (Q^r,Q^\theta)_\Theta = \big((\lambda r^2+1)\dot r\partial_E\Theta,r^2(\lambda r^2+1)(\dot \theta \partial_E\Theta+ \partial_L\Theta)\big), \label{Q-Theta} \\ & (Q^r,Q^\theta)_T = \big((\lambda r^2+1)\dot r\partial_E T,r^2(\lambda r^2+1)(\dot \theta \partial_E T+ \partial_L T)\big), \label{Q-T} \end{align} where \begin{align} &\begin{aligned} \partial_L\Theta & = \frac{1}{E^2+(2E\lambda-\omega^2)L^2} \bigg( {\rm sgn}(\dot r)\frac{(2(\lambda r^2 +1)E -\omega^2 r^2)E+(2\lambda E-\omega^2)L^2 }{2\sqrt{2Er^2(\lambda r^2 +1) - L^2 - \omega^2r^4}} \bigg)\bigg|^{r}_{r_0}, \end{aligned} \\ &\begin{aligned} \partial_E\Theta = - \partial_L T &= \frac{L}{E^2+(2E\lambda-\omega^2)L^2} \bigg( {\rm sgn}(\dot r)\frac{\omega^2r^2 -(\lambda r^2 +1)E -\lambda L^2}{2\sqrt{2Er^2(\lambda r^2 +1) - L^2 - \omega^2r^4}} \bigg)\bigg|^{r}_{r_0}, \end{aligned} \\ &\begin{aligned} \partial_ET & =\frac{\lambda(2w^2-\lambda E)}{2\sqrt{w^2-2\lambda E}^5} \arctan\Big(\frac{{\rm sgn}(\dot r)\big((2\lambda r^2+1)E -\omega^2 r^2\big)}{\sqrt{2Er^2(\lambda r^2 +1)-L^2-\omega^2r^4}\sqrt{w^2-2\lambda E}}\Big)\Big|^{r}_{r_0} \\&\qquad +\frac{1}{2(w^2-2\lambda E)} \Big( (\lambda r^2+1)\Big( \lambda r^2 +\frac{(\omega^2-\lambda E)(Er^2-L^2)}{E^2+(2E\lambda-\omega^2)L^2}\Big)\Big)\Big|^{r}_{r_0}, \\&\qquad +\frac{\lambda}{2(w^2-2\lambda E)^2} \Big(2\lambda +\frac{E(\omega^2-\lambda E)}{E^2+(2E\lambda-\omega^2)L^2}\Big)\Big({\rm sgn}(\dot r)\sqrt{2Er^2(\lambda r^2 +1)-L^2-\omega^2r^4}\Big)\Big|^{r}_{r_0}. \end{aligned} \end{align} Applying the Noether correspondence \eqref{PfromQ} to each multiplier pair, we obtain the corresponding variational symmetries \begin{gather} \hat\mathbf{X}_{L}= -\partial/\partial \theta, \quad \hat\mathbf{X}_{E}= -\dot r\partial/\partial r -\dot\theta \partial/\partial \theta, \label{hatX-L-E} \\ \hat\mathbf{X}_{\Theta} =-(\dot r \partial_E\Theta) \partial/\partial r -(\partial_L\Theta +\dot\theta\partial_E \Theta)\partial/\partial \theta, \label{hatX-Theta} \\ \hat\mathbf{X}_{T} =-(\dot r \partial_E T) \partial/\partial r -(\partial_L T +\dot\theta\partial_E T)\partial/\partial \theta. \label{hatX-T} \end{gather} These symmetries can be understood as acting on solutions $(r(t),\theta(t))$ of the \eom/. Equivalently, the symmetries can be formulated as acting on the variables $(t,r,\theta)$ by means of a standard transformation \cite{BA-book,Olver-book} which has the general form $\eta^r=P^r +\tau \dot r$, $\eta^\theta=P^\theta +\tau \dot\theta$, yielding \begin{equation*} \mathbf{X} = \tau\partial/\partial_t + \eta^r\partial/\partial_r + \eta^\theta\partial/\partial_\theta \end{equation*} where $\tau$ can be chosen freely as a function of $t,r,\theta,\dot r,\dot\theta$. We take $\tau=0$ for $\hat\mathbf{X}_{L}$, $\tau=1$ for $\hat\mathbf{X}_{E}$, giving \begin{equation}\label{X-L-E} \mathbf{X}_{L}= -\partial/\partial \theta, \qquad \mathbf{X}_{E}= \partial/\partial t . \end{equation} In this form, these generators represent point symmetries, consisting of rotations and time-translations. For $\hat\mathbf{X}_{\Theta}$ we take $\tau = \partial_E\Theta$, which yields \begin{equation}\label{X-Theta} \begin{aligned} \mathbf{X}_{\Theta} & = \partial_E\Theta\partial/\partial_t -\partial_L\Theta\partial/\partial \theta . \end{aligned} \end{equation} Similarly, for $\hat\mathbf{X}_{T}$ we take $\tau = \partial_E T$, giving \begin{equation}\label{X-T} \begin{aligned} \mathbf{X}_{T} & = \partial_E T \partial/\partial_t -\partial_L T\partial/\partial \theta . \end{aligned} \end{equation} Both of these generators $\mathbf{X}_{\Theta}$ and $\mathbf{X}_{T}$ represent dynamical symmetries. Moreover, there is no choice of $\tau$ that can transform them into point symmetries, because the components of the generators $\hat\mathbf{X}_{L}$ and $\hat\mathbf{X}_{T}$ have a nonlinear dependence on $\dot r$ and $\dot\theta$ through the expressions for $L,E$. The commutators of the variational symmetries \eqref{X-L-E}, \eqref{X-Theta}, \eqref{X-T} turn out to vanish, as shown later, whereby the variational symmetries comprise a four-dimensional abelian algebra. In summary, we have shown how to go \emph{Hamiltonian system $\Rightarrow$ first integrals $\Rightarrow$ local symmetries} in a systematic way by using the determining equation \eqref{I-deteqn} for first integrals and the reverse version \eqref{PfromQ} of Noether's theorem. \subsection{From symmetries to first integrals} We will now show how to go \emph{Hamiltonian system $\Rightarrow$ local symmetries $\Rightarrow$ first integrals} in a systematic way without having to use any ansatzes or guess-work, by following the extended symmetry method outlined in Ref.~\cite{AncMeaPas}. \emph{Step 1}: Compute all variational point symmetries of the Hamiltonian system \eqref{eom}. Point symmetries are given by generators (vector fields) of the form $\mathbf{X} = \tau(t,r,\theta)\partial/\partial{t} + \eta^r(t,r,\theta)\partial/\partial{r} + \eta^\theta(t,r,\theta)\partial/\partial{\theta}$ under which the system \eqref{eom} is infinitesimally invariant, \begin{equation}\label{symmcond} {\rm pr}^{(2)}\mathbf{X} (\ddot r - f^r)|_{\mathcal E} = 0, \qquad {\rm pr}^{(2)}\mathbf{X} (\ddot\theta - f^\theta)|_{\mathcal E} = 0 . \end{equation} Here ${\rm pr}^{(2)}\mathbf{X}$ is the second prolongation of $\mathbf{X}$, acting on the coordinate space $(t,r,\theta,\dot r,\dot\theta,\ddot r,\ddot\theta)$. The prolongation formula is somewhat complicated. It can be avoided by working with the generator $\mathbf{X}$ in evolutionary form \begin{equation}\label{pointsymm} \hat\mathbf{X} = P^r \partial/\partial{r} + P^\theta \partial/\partial{\theta}, \qquad P^r = \eta^r -\tau \dot r, \quad P^\theta = \eta^\theta-\tau \dot\theta, \end{equation} acting on solutions $(r(t),\theta(t))$. Then the invariance condition \eqref{symmcond} becomes simply \begin{equation} \begin{aligned} 0 & ={\rm pr}^{(1)}\hat\mathbf{X}(\ddot r - f^r)|_{\mathcal E} = \big( \ddot P^r - P^r \partial_r f^r - P^\theta \partial_\theta f^r - \dot P^r \partial_{\dot r} f^r - \dot P^\theta \partial_{\dot\theta} f^r \big)\big|_{\mathcal E}, \\ 0 & ={\rm pr}^{(1)}\hat\mathbf{X}(\ddot\theta -f^\theta)|_{\mathcal E} = \big( \ddot P^\theta - P^r \partial_r f^\theta - P^\theta \partial_\theta f^\theta - \dot P^r \partial_{\dot r} f^\theta - \dot P^\theta \partial_{\dot\theta} f^\theta \big)\big|_{\mathcal E}. \end{aligned} \end{equation} This pair of determining equations splits with respect to $\dot r$, $\dot \theta$, and thereby yields an overdetermined linear system of equations for $\tau$, $\eta^r$, $\eta^\theta$. After simplification, the linear system reduces to $\partial_t\tau=\partial_r\tau=\partial_\theta\tau=0$, $\eta^r=0$, $\partial_t\eta^\theta=\partial_r\eta^\theta=\partial_\theta\eta^\theta=0$, whose solution is given by $\tau=C_1$, $\eta^r=0$, $\eta^\theta=C_2$, where $C_1,C_2$ are constants. Hence, we obtain the two point symmetries \eqref{X-L-E}, consisting of rotations ($C_1=0$, $C_2=-1$) and time-translations ($C_1=1$, $C_2=0$). Both of these point symmetries are variational. In particular, in evolutionary form \eqref{hatX-L-E}, their action on the Lagrangian is a total time derivative given by \begin{equation}\label{XactionLagr} {\rm pr}^{(1)}\hat\mathbf{X}_{L}({\mathcal L}) = \dot R =0, \qquad {\rm pr}^{(1)}\hat\mathbf{X}_{E}({\mathcal L}) = \dot R = -\dot{\mathcal L} . \end{equation} \emph{Step 2}: Use Noether's theorem to obtain first integrals from the variational point symmetries. The action \eqref{XactionLagr} of the variational point symmetries \eqref{hatX-L-E} on ${\mathcal L}$ gives $R=0$ and $R=-{\mathcal L}$ respectively. Then, through the Noether correspondence \eqref{I}, this yields the first integrals $I=L$ and $I=E$, which are given by the angular momentum \eqref{L} and the energy \eqref{E}. \emph{Step 3}: Re-write the Hamiltonian system \eqref{eom} in first-order form using the previous first integrals. First, expressions \eqref{L} and \eqref{E} for the first integrals directly yield \begin{equation}\label{1stordsys1} \dot r = \frac{{\rm sgn}(\dot r)\sqrt{2E(\lambda r^2 +1) - L^2/r^2 - \omega^2r^2}}{\lambda r^2 +1} = F^r, \qquad \dot \theta = \frac{L}{r^2(\lambda r^2 +1)} = F^\theta. \end{equation} Next, since the first integrals are time-independent, they satisfy \begin{equation}\label{1stordsys2} \dot L =0, \qquad \dot E =0. \end{equation} These four equations constitute a first-order form for the Hamiltonian system \eqref{eom}. \emph{Step 4}: Find all point symmetries of the first-order system such that every joint invariant of the variational point symmetries is preserved. An invariant of a point symmetry $\mathbf{X}$ is a function $\chi$ of $t,r,\theta$ that is annihilated by the symmetry generator, $\mathbf{X}(\chi)=0$. The joint invariants of the two variational point symmetries \eqref{X-L-E} clearly consist of only $\chi=r$. To preserve this invariant, we search for point symmetries on $(t,r,\theta,E,L)$ of the infinitesimal form \begin{equation}\label{Y} \mathbf{Y} = \tau(r,L,E)\partial/\partial t + \eta^\theta(r,L,E)\partial/\partial\theta + \eta^L(r,L,E)\partial/\partial L + \eta^E(r,L,E)\partial/\partial E. \end{equation} The condition for the first-order system \eqref{1stordsys1}--\eqref{1stordsys2} to be infinitesimally invariant is given by \begin{gather*} 0 ={\rm pr}^{(1)}\hat\mathbf{Y}(\dot L)|_{\mathcal E} = F^r\partial_r\eta^L, \qquad 0 ={\rm pr}^{(1)}\hat\mathbf{Y}(\dot E)|_{\mathcal E} = F^r\partial_r\eta^E, \\ 0 ={\rm pr}^{(1)}\hat\mathbf{Y}(\dot \theta - F^\theta)|_{\mathcal E} = \big( F^r(\partial_r\eta^\theta -F^\theta\partial_r\tau) -\eta^E\partial_E F^\theta - \eta^L\partial_L F^\theta \big), \\ 0 ={\rm pr}^{(1)}\hat\mathbf{Y}(\dot r - F^r)|_{\mathcal E} = -\big( (F^r)^2\partial_r\tau + \eta^E\partial_E F^r + \eta^L\partial_L F^r \big), \end{gather*} using the symmetry generator in evolutionary form $\hat\mathbf{Y}= (\eta^\theta -\tau F^\theta)\partial/\partial{\theta} -\tau F^r\partial/\partial{r} + \eta^L\partial/\partial{L} + \eta^E\partial/\partial{E}$. These determining equations yield the linear system \begin{gather} \partial_r\eta^L=0, \quad \partial_r\eta^E=0, \\ \partial_r\tau = -(\eta^E\partial_E F^r + \eta^L\partial_L F^r)/(F^r)^2, \quad \partial_r\eta^\theta = -(\eta^E\partial_E(F^r F^\theta) +\eta^L\partial_L(F^r F^\theta))/(F^r)^2, \end{gather} which can be straightforwardly solved. Up to arbitrary functions of $L$ and $E$, the general solution is given by $\tau = C_1\partial_E\Theta + C_2\partial_E T +C_3$, $\eta^\theta = -C_1\partial_L\Theta - C_2\partial_LT -C_4$, $\eta^L = C_1$, $\eta^E=C_2$, where $C_1,C_2,C_3,C_4$ are constants. Hence we obtain four point symmetries \begin{gather} \mathbf{Y}_L = -\partial/\partial{\theta} = \mathbf{X}_L, \qquad \mathbf{Y}_E = \partial/\partial{t} = \mathbf{X}_E, \label{Y-L-E} \\ \mathbf{Y}_\Theta =\partial_E\Theta\partial/\partial{t} -\partial_L\Theta \partial/\partial{\theta} + \partial/\partial{L}, \qquad \mathbf{Y}_T =-\partial_E T\partial/\partial{t} +\partial_L T \partial/\partial{\theta} - \partial/\partial{E}. \label{Y-Theta-T} \end{gather} \emph{Step 5}: Convert the additional point symmetries into dynamical symmetries of the Hamiltonian system \eqref{eom}. The two symmetries $\mathbf{Y}_L$ and $\mathbf{Y}_E$ are clearly inherited from the rotation and time-translation symmetries \eqref{X-L-E} of the Hamiltonian system \eqref{eom}. Since these symmetries are the only point symmetries admitted by this system, the two additional symmetries $\mathbf{Y}_\Theta$ and $\mathbf{Y}_T$ must therefore yield dynamical symmetries when they are transformed to act on solutions $(r(t),\theta(t))$. Their action is obtained simply by first expressing the symmetries in evolutionary form and then projecting the generators onto the coordinate space $(r,\theta)$. This yields $\hat\mathbf{Y}_\Theta = -(\partial_L \Theta +\dot\theta\partial_E \Theta)\partial/\partial{\theta} -\dot r\partial_E\Theta\partial/\partial{r} =\hat\mathbf{X}_\Theta$ and $\hat\mathbf{Y}_T = (\partial_L T+\dot\theta\partial_E T )\partial/\partial{\theta} +\dot r\partial_E T\partial/\partial{r} =-\hat\mathbf{X}_T$, which are the two dynamical symmetries \eqref{hatX-Theta} and \eqref{hatX-T} obtained previously. \emph{Step 6}: Apply Noether's theorem to obtain first integrals by using the dynamical symmetries that are variational. There are two methods to verify a priori that the two dynamical symmetries $\hat\mathbf{Y}_\Theta$ and $\hat\mathbf{Y}_T$ are variational, and to derive the corresponding first integrals. A direct method \cite{BA-book,Olver-book} consists of showing that the action of the symmetries on the Lagrangian is a total time derivative, ${\rm pr}^{(1)}\hat\mathbf{Y}({\mathcal L})=\dot R$. This also yields $R$ so that the first integral \eqref{I} can be obtained. However, it is somewhat complicated to find $R$ explicitly. An alternative method which by-passes this complication uses only the dynamical symmetries themselves \cite{BA-book,AncBlu98}. First, a dynamical symmetry $\hat\mathbf{Y} = P^r\partial/\partial{r} + P^\theta\partial/\partial{\theta}$ is variational iff the pair $(Q^r,Q^\theta)$ given by the Noether correspondence \eqref{QfromP} is a multiplier, satisfying \begin{equation}\label{Q-deteqns} \frac{\delta}{\delta r}\Big((\ddot r -f^r)Q^r+(\ddot \theta -f^\theta)Q^\theta\Big)=0, \quad \frac{\delta}{\delta \theta}\Big((\ddot r -f^r)Q^r+(\ddot \theta -f^\theta)Q^\theta\Big)=0. \end{equation} Then, the resulting first integral \eqref{I} can be obtained by a line integral formula \begin{equation}\label{I-lineintegral} I = \int_{\mathcal C} \Big( (1+\lambda r^2)^{-1}\dot P^r\,dr + (r^2(1+\lambda r^2))^{-1}\dot P^\theta\,d\theta -(1+\lambda r^2)P^r\,d\dot r -r^2(1+\lambda r^2)P^\theta\,d\dot \theta \Big)\big|_{\mathcal E} \end{equation} where $\mathcal C$ is any curve from $(t_0,r_0,\theta_0,\dot r_0,\dot\theta_0)$ to $(t,r,\theta,\dot r,\dot\theta)$. It is straightforward to verify that the multiplier equations \eqref{Q-deteqns} hold for both of the dynamical symmetries $\hat\mathbf{Y}_\Theta$ and $\hat\mathbf{Y}_T$, using the pairs \eqref{Q-Theta} and \eqref{Q-T}. The corresponding first integrals from formula \eqref{I-lineintegral} (up to an additive constant) are $I=\Theta$ and $I=T$, given by expressions \eqref{Theta} and \eqref{T}, respectively. Altogether, the two first integrals arising from the dynamical symmetries, plus the two first integrals arising from the point symmetries, comprise the complete set of four functionally-independent first integrals \eqref{L}--\eqref{T} for the Hamiltonian system \eqref{eom}. \subsection{Variational symmetry algebra} The commutator structure of the four variational symmetries \eqref{hatX-L-E}--\eqref{hatX-T} can be derived in several different ways. A direct method is to express the commutators in terms of the action of the prolonged symmetries on the components of the symmetry generators. However, the prolongations can be avoided by instead using the representation of the symmetries in the form \eqref{Y-L-E}--\eqref{Y-Theta-T} given by point symmetries acting on the coordinate space $(t,r,\theta,E,L)$. The commutators are then simple to compute, and we find that every commutator vanishes. Hence the four point symmetries \eqref{Y-L-E}--\eqref{Y-Theta-T} comprise an abelian algebra. The same algebra structure then holds for the four variational symmetries \eqref{hatX-L-E}--\eqref{hatX-T}. We remark that this result provides an alternative way to obtain the corresponding first integrals, by utilizing the canonical coordinates of the four symmetries \eqref{Y-L-E}--\eqref{Y-Theta-T} as follows. The canonical form of a symmetry \eqref{Y} is given by $\mathbf{Y} = \partial/\partial{\zeta}$ where $\zeta(t,r,\theta,L,E)$ is a canonical coordinate satisfying $\mathbf{Y}(\zeta)=\tau\partial_t\zeta + \eta^\theta\partial_\theta\zeta + \eta^L\partial_L\zeta + \eta^E\partial_E\zeta=1$. Since the four symmetries \eqref{Y-L-E}--\eqref{Y-Theta-T} are mutually commuting, there exists a point transformation from $(t,r,\theta,E,L)$ to the coordinate space $(r,\zeta^L,\zeta^E,\zeta^\Theta,\zeta^T)$ consisting of the four canonical coordinates and the joint invariant of all four symmetries. It is straightforward to see that $\zeta^L = -\Theta$, $\zeta^E = T$, $\zeta^\Theta=L$, $\zeta^T=-E$. \section{Concluding remarks} Finally, we stress that the approach presented in this paper is widely applicable to many other Hamiltonian systems and will be helpful for unveiling the connections between local and global aspects of their integrability and symmetry properties. In particular, we plan to study first integrals and local symmetries of more general central force systems with Hamiltonians of the form $H=m(r)^{-1}((p^r)^2 + (p^\theta/r)^2)+ U(r)$ having a position-dependent mass, as well as generic (for instance, H\'enon-Heiles type) nonlinearly coupled systems of oscillators. Indeed, from a more general perspective, an important open problem is how to detect global regularity of first integrals. \section*{Acknowledgements} S.C.A. is supported by an NSERC research grant. A.B. has been partially supported by Ministerio de Ciencia, Innovaci\'on y Universidades (Spain) under grant MTM2016-79639-P (AEI/FEDER, UE), by Junta de Castilla y Le\'on (Spain) under grant BU229P18. M.L.G. gratefully acknowledges support by Junta de Andaluc\'{i}a (Spain) for the research group grant FQM-201.

acc9cf73a0b305a073859a93791bcb652dd4e64a

\section{Introduction}\label{s1} The problem of constructing a consistent theory of interacting fields with higher spin is a long-standing one. Mainly, it was an interesting theoretical curiosity, but nowadays, thanks to advances in string field theory and attempts to construct a quantum theory of gravity, we see a renaissance of interest in higher spins, which may illuminate some fundamental aspects behind those theories. Before embarking on a journey into the yet uncharted territory of higher spin theory, in this section we review what we already know about spin. After a basic introduction, in \textbf{Section \ref{s2}} we discuss some of the main motivating factors that keep driving new research in this area. As history frequently teaches us important lessons, in \textbf{Section \ref{s3}} we go through a "folk" history of higher spin theory, having the obvious advantage of hindsight. One should keep in mind that this is far from a complete history, it is a story consciously biased towards the topics we will be dealing with here. This section also serves as an appropriate place to expose some of the most important ideas and developments in higher spin theory.\newline We review several theorems whose implications seem to render the investigation of higher spins pointless. In \textbf{Section \ref{s5}}, as a warm-up exercise, we review the well-known theories of fields with spin $0, 1$ and $2$, carefully examining the role of spin and trying to catch a glimpse of the underlying pattern common to all spins. Finally, in \textbf{Section \ref{s6}}, we present the higher spin theory of massless bosons in flat spacetime using an elegant mathematical formalism that works for all spins in all dimensions. We discuss both free and interacting theory (with a generic external current) in their constrained and unconstrained forms. We also examine the issue of non-locality and we investigate the geometric formulation of higher spin theories. In \textbf{Appendix \ref{a1}}, we provide some explicit calculational results in higher spin theory. These results were obtained and checked using a simple C\texttt{++} code written specifically for this purpose. The core snippets of the code are provided in \textbf{Appendix \ref{a2}}. \newpage \subsection{What is spin?} Spin is an intrinsic property of relativistic fields, which (after quantization) give rise to particles. It can be viewed as an additional degree of freedom unrelated to spatial degrees of freedom specified by position and momentum. Its name comes from the fact that, mathematically, spin behaves like quantized angular momentum. Unlike orbital angular momentum, spin quantum numbers may have half-integer values and fundamental particles cannot be made to stop "spinning" or to spin faster or slower, spin can only change its orientation. Particles (fields) with integral spin are called \textbf{bosons}, and those with half-integral spin are known as \textbf{fermions}. \textit{Spin-statistics theorem}, one of the rigorous results of axiomatic quantum field theory, tells us that bosons and fermions behave in a drastically different manner. While the former respect the \textbf{Bose-Einstein} statistics, the latter respect the \textbf{Fermi-Dirac} statistics, and as a consequence are subject to the \textit{Pauli exclusion principle}. Fermions form particles of matter, while bosons mediate the interactions between them. \subsection{Where does spin come from?} The existence of spin is a direct consequence of the most fundamental mathematical properties of our universe, \textbf{spacetime symmetries}. This is why the true meaning of spin has to be discussed in the context of a fully Lorentz-invariant theory. Quantum field theory, which is the underlying formalism describing the Standard Model of particle physics, is such a theory. We introduce a field for each fundamental particle species, which transforms \textit{nicely} under Lorentz transformations. Once we pick a particular representation of the Lorentz transformations, it specifies the spin. After quantizing the field, one finds that the field operator creates or annihilates particles of definite spin, which was of course, the spin associated with the classical field to begin with. \subsubsection{Spin from irreducible representations in four-dimensional spacetime}\label{IRREPs} Instead of simply looking at Lorentz transformations, we have to look at the full class of spacetime isometries (i.e. isometries of \textit{Minkowski space} $\mathcal{M}^4$). They are locally described by the \textit{Poincaré group} \begin{equation} \mathcal{P} = \mathbb{R}^{(1,3)} \rtimes \mathrm{SO} (1,3) \, , \end{equation} a ten-dimensional noncompact Lie group, corresponding to ten independent symmetries (3 spatial translations, time translation, 3 spatial rotations and 3 Lorentz boosts).\newline A closer look at continuous local symmetries leads us to analyze the identity component of $\mathcal{P}$, which acts as a stabilizer group of the origin, the \textit{proper orthochronous Lorentz group}, \begin{equation} \mathrm{SO}^+ (1,3) = \left\{ \Lambda \in \mathrm{GL} (4, \mathbb{R}) \, \big| \, \Lambda^T \eta \Lambda = \eta, \, \eta = \mathbf{diag} (-1,1,1,1) \right\} \, . \end{equation} The fundamental physical fields must carry the irreducible representations of its Lie algebra, \begin{equation} \mathfrak{so}(1,3) \hookrightarrow \mathfrak{so}(1, 3)_\mathbb{C} \cong \mathfrak{su}(2)_\mathbb{C} \oplus \mathfrak{su}(2)_\mathbb{C} \cong \mathfrak{sl} (2, \mathbb{C}) \, , \end{equation} which generates the covering Lie group of $\mathcal{P}$, \begin{equation} \mathbb{R}^{(1,3)} \rtimes \mathrm{SL} (2,\mathbb{C}) \hookleftarrow \mathcal{P} \, . \end{equation} Elements of this group are generated by terms of the form \begin{equation} \exp(i a_\mu P^\mu) \exp\left(\frac{i}{2} \omega_{\mu\nu} M^{\mu\nu }\right) \, , \end{equation} where $a_\mu$ parametrizes translations generated by $P^\mu$, and $\omega_{\mu\nu}$ parametrizes Lorentz transformations (rotations and boosts) generated by $M^{\mu\nu}$. This Lie algebra is defined by \begin{align} \left[ P_\mu, P_\nu \right] &= 0 \, , \\ \left[ M_{\mu\nu}, P_\rho \right] &= i \, \eta_{\rho [ \mu} P_{\nu ]} \, , \\ \left[ M_{\mu\nu}, M_{\rho\sigma} \right] &= i \, \eta_{[ \mu \nu} M_{\rho \sigma]} \, , \end{align} where $[\dots]$ stands for unweighted anti-symmetrization of indices with the minimal number of terms and $M_{\mu\nu}$ is defined in terms of the rotation generator $J_i$ and boost generator $K_\mu$ as \begin{align} J_i &= \frac{1}{2} \varepsilon_{ijk} M^{jk} \, , \\ K_i &= M_{0i}. \end{align} The \textit{Casimir invariants} of the Poincaré group are \begin{equation} P_\mu P^\mu := m^2 \, , \end{equation} where $m$ stands for mass and\footnotemark \footnotetext{This is true if $m \neq 0$, but $m=0$ does not imply $W^2=0$.} \begin{equation} W_\mu W^\mu = m^2 s (s+1), \end{equation} where $s$ stands for spin\footnotemark.\footnotetext{At this level of analysis, which is purely mathematical, \textit{mass} and \textit{spin} are simply names we give to these quantities. Their physical properties only become apparent after introducing the actual physics, i.e. equations of motion.} $W_\mu$ is the \textit{Pauli-Lubanski} pseudovector, defined as the Hodge dual of $\mathbf{J} \wedge \mathbf{P}$, i.e. \begin{equation} W_\mu := \frac{1}{2} \varepsilon_{\mu\nu\rho\sigma} J^{\nu\rho} P^\sigma. \end{equation} \newpage Next, we look at \textit{Wigner's little groups}, stabilizer subgroups of various mass states. \begin{itemize} \item $m>0$, stabilizer of $P=(m,0,0,0)$ \newline$\implies$ massive states with mass $m$ and spin $s \in \mathbb{N}_0/2$ \item $m=0$ and $P_0 > 0$, stabilizer of $P=(k,0,0,k)$ \newline$\implies$ $s \in \mathbb{N}_0/2$ IRREPs and \textit{continuous spin} representation \item $m^2 < 0$, stabilizer of $P=(0,0,0,m)$ \newline$\implies$ \textit{tachyons}\footnotemark \footnotetext{Fields with an imaginary mass, which propagate faster-than-light excitations and lead to theories with instabilities and violation of causality.} \item $m=0$ and $P^\mu=0$ \newline$\implies$ trivial representation, the \textit{vacuum} state \end{itemize} We can now classify the physically relevant\footnotemark\, finite-dimensional irreducible representations of the double cover of the Poincaré group by two numbers, $m \in \mathbb{R}^+$ and $s \in \frac{1}{2} \mathbb{Z}$. \footnotetext{We follow the standard prescription of ignoring tachyons and \textit{continuous spin} representations.\newline The latter seem to give rise to fields whose excitations cannot be compactly localized, but are instead localized on semi-infinite spacelike strings (see \cite{infspin}) that we do not seem to find in nature.} These irreducible representations are further classified by two numbers, $j_1$ and $j_2$ such that $j_1 + j_2 = s$ and are labeled as $(j_1, j_2)$ representations, summarized in the table below.\\ \begin{tabular}{ r | c | c | c | c} \hline \textbf{spin} & \textbf{representation} & \textbf{field} & \textbf{eq. of motion} & \textbf{example} \\ \hline $0$ & $(0,0)$ & scalar & \textit{Klein-Gordon} & Higgs \\ $1/2$ & $(\frac{1}{2},0) \oplus (0, \frac{1}{2})$ & spinor & \textit{Dirac} & electron \\ $1$ & $(\frac{1}{2}, \frac{1}{2})$ & vector & \textit{Proca} & photon \\ $3/2$ & $(\frac{1}{2},1) \oplus (1, \frac{1}{2})$ & spinor-vector & \textit{Rarita-Schwinger} & gravitino\footnotemark \\ $2$ & $(1,1)$ & 2-tensor\footnotemark & \textit{linearized Einstein} & graviton\footnotemark \\ $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ \\ \hline \end{tabular}\newline \footnotetext[5]{Graviton's fermionic superpartner in theories with Bose-Fermi symmetry (supersymmetry), specifically in supergravity.} \footnotetext[6]{Symmetric tensor of order two.} \footnotetext[7]{Hypothetical quantized excitation of the gravitational field. } The dots at the end of the table represent the fact that, from a mathematical point of view, nothing prevents the existence of higher spin fields at this level of analysis, nor does anything directly imply that we should expect them to behave differently from their lower spin counterparts. For a detailed group-theoretical analysis of spin, see, for example, \cite{weinbergQFT}, \cite{sternberg}, \cite{BBIRREPs} or \cite{schwichtenberg}. \newpage \subsubsection{Spin from an action principle} It has long been thought that spin cannot be formulated with an action principle within the framework of Lagrangian mechanics with particles. This issue was particularly relevant with the introduction of Feynman's sum-over-histories approach to quantization.\newline Even Feynman himself wrote in his 1965 book \cite{feynmanpath}:\newline \q{With regards to quantum mechanics, path integrals suffer most grievously from a serious defect. They do not permit a discussion of spin operators or other such operators in a simple and lucid way. ... It is a serious limitation that the half-integral spin of the electron does not find a simple and ready representation.}\newline However, this is actually possible and relatively straightforward, but the appropriate phase space formulation was not fully obvious until several years later \cite{pathspin}. One begins by examining how spin behaves classically. It can be thought of as a spinning top, or a little arrow with fixed length, sticking out from the particle and pointing in a particular direction in three-dimensional physical space. Therefore, it is reasonable to assume that its phase space is a $2$-sphere $\mathcal{S}^2$, and its dynamical variables can be taken to be the polar and azimuthal angles $\theta$ and $\phi$. This establishes the point particle as an entity described not only by its position and momentum, but also by the orientation of its spin "arrow". The rest is simply a matter of mathematical construction. It is precisely this that caused so much confusion, because in order to write the action, one has to find the proper local geometric invariant on $\mathcal{S}^2$. The volume form on a $2$-sphere is \begin{equation} \mathbf{\omega} = \sin{\theta} \, d \phi \, \wedge \, d\theta \, , \end{equation} and it is invariant under rotations. Since $\mathbf{\omega}$ is a closed form, we can write it locally as an exact form, i.e. \begin{equation} \mathbf{\omega} = d \mathbf{\chi} = d (\cos{\theta} \, d \phi) \, , \end{equation} so the action for spin $J$ can be written as \begin{equation} S = J \int \chi = J \int d\phi \, \cos{\theta} = J \int dt \, \dot{\phi} \cos{\theta} \, . \label{int} \end{equation} A distinctive property of this system is that its phase space is compact (i.e. closed and bounded), which implies a finite-dimensional Hilbert space. Furthermore, invariance of $\mathcal{S}^2$ (as a manifold embedded in $\mathbb{R}^3$) under $\mathrm{SO}(3)$ guarantees the operators with correct commutation relations\footnotemark. \footnotetext{We will not derive this here. The interested reader is referred to \cite{condmatter}.} Let us demonstrate how quantized spin arises when we plug \eqref{int} into the Feynman integral. The spin term in the Feynman integral, with $J=j\hbar$ reads \footnotemark \footnotetext{Here and only here, we write $\hbar$ explicitly instead of working with natural units, where $\hbar=c=1$} \begin{equation} e^{i S / \hbar} = \exp\left(ij\int dt \, \dot{\phi} \cos{\theta}\right) \, . \end{equation}\newpage We proceed by using the Stokes' theorem\footnotemark\, on the integral, \footnotetext{$\int_{\partial \Omega} \omega = \int_{\Omega} d\omega$} \begin{equation} \int dt \, \dot{\phi} \cos{\theta} = \oint_C d \phi \cos{\theta} = \int_M d \phi d \theta \sin{\theta} \end{equation} where $C$ denotes a closed path on $\mathcal{S}^2$, bounding a $2$-surface $M$, i.e. $\partial M = C$. Since $\mathcal{S}^2$ is compact, the choice of $M$ is not unique, but the difference between two possible choices is simply the integral over entire $\mathcal{S}^2$. In other words, the difference between two possible choices for the action is \begin{equation} \Delta S = j \hbar \int_{\mathcal{S}^2} d \phi d \theta \sin{\theta} = 4 \pi \hbar j \, . \end{equation} The path integral cannot be multivalued, which in turn means that $e^{iS/\hbar}$ has to be single-valued. Therefore, \begin{equation} e^{i \Delta S / \hbar} = 1 \quad \implies \quad 4 \pi j = 2 \pi N \quad ( \forall N \in \mathbb{Z}) \, , \end{equation} i.e. spin $j$ can take only integer and half-integer values, the same conclusion we arrived at using group representation theory in \textbf{Section \ref{IRREPs}}. Once again, nothing at the mathematical level of analysis prevents the existence of arbitrarily high spins nor does it indicate any sort of inconsistency. \newpage \subsection{What do we mean by higher spin?} Historically, the term \textit{higher spin}\footnotemark\, was used to refer to several different domains of theoretical constructs. \footnotetext{Often abbreviated as HS.}One of the reasons for including only spin $0$, $1/2$ and $1$ fields in the domain of \textit{lower spin} was the fact that only those result in renormalizable quantum field theories. Today, by \textit{higher spin}, we mean spin greater than \textbf{two}, i.e. spin-$5/2$ and higher for fermions, spin-$3$ and higher for bosons. This seems more appropriate since we \textit{do} have consistent \textit{classical} field theories for $s\leq2$, but all higher spins yield problematic constructions even before quantization. Constructing a consistent interacting theory of HS fields (sometimes referred to as \textit{higher-spin gravity} in the case of massless interacting fields) has been a long-standing problem in theoretical physics. So far, we only have a fully consistent interacting HS theory in $\mathrm{(A)dS}$ spacetimes, which has become known as \textit{Vasiliev's theory}. Similar attempts at constructing such theories in flat space have not been successful. Unfortunately, taking the flat-space limit of Vasiliev's theory in $\mathrm{(A)dS}$ in hope of recovering a theory in flat spacetime is by no means trivial and possibly not even well-defined. Interestingly, $\mathrm{(A)dS}$ spacetimes are highly symmetrical, and their symmetry group $\mathrm{SO}(1,4) \cong \mathbf{Sp}(2,2)$ reduces\footnotemark\, to the Poincaré group in the limiting case of infinite \newline(anti-)de Sitter radius, which may point to the $\mathrm{(A)dS}$ group as being more fundamental\cite{missed}. \footnotetext{This can be accomplished rigorously using the İnönü-Wigner group contraction.} \newpage \section{Why study HS theory?}\label{s2} \subsection{String theory}\label{ST} String theory is a promising candidate for a consistent theory of quantum gravity. Its perturbative spectrum consists of states with arbitrarily high spins and masses. One could say that higher spin gravity lies between supergravity and string theory, which makes it particularly interesting. In string theory, one finds an \textbf{infinite tower of massive string excitations} with increasing spin. The existence of this infinite tower of higher-spin fields is crucial for the absence of ultraviolet divergences, an extremely important feature of string theory. The only free parameter in string theory is the \textit{string constant}, denoted by $\alpha'$, which determines the characteristic length and mass scale of strings. In the $\alpha' \to 0$ limit, the theory reduces to supergravity, i.e. a theory with massless modes. On the other hand, in the $\alpha' \to \infty$ limit, all excitations become massless, and the theory resembles higher spin gravity. Furthermore, we know that it is possible to have a theory with only massless fields in its formal construction, which nevertheless produces no massless excitations after quantization. In other words, it is possible to begin with a Lagrangian with massless fields, which describes a quantum field theory without massless propagating degrees of freedom. There are at least two mechanisms, familiar from the Standard Model, that exhibit such behaviour. One is \textit{spontaneous symmetry breaking}, i.e. the \textit{Higgs mechanism} that gives mass to massive fundamental\footnotemark\footnotetext{Spontaneous symmetry breaking also occurs in non-fundamental descriptions, for example in the theory of superconductivity and superfluidity.} particles. The other one is \textit{color confinement}, the mechanism responsible for clumping of gluons and quarks into colorless hadrons. It is possible that a similar mechanism underlies the generation of massive states in string theory, in which case we would have to know how to construct a massless higher spin theory. There are strong indications that symmetries of string theory form a very large group, much larger than what can be seen using the perturbative approach, which spontaneously breaks down to a smaller group, giving mass to higher-spin excitations. Therefore, it is plausible that a firm understanding of HS theory could shed some light on the underlying mathematical structure of string theory. In particular, we would like to know what symmetries the theory possesses and what is the notion of spacetime geometry in string theory and HS gravity. \newpage \subsection{AdS/CFT correspondence} The $\mathrm{AdS/CFT}$ (\textit{anti-de Sitter/conformal field theory}) correspondence\footnotemark \footnotetext{Also known as \textit{Maldacena duality} or \textit{gauge/gravity duality}.}, in its strictest formulation, posits an equivalence between the theory of quantum gravity in anti-de Sitter spacetimes, as formulated in string theory, and conformal field theory on its boundary. However, although it seems to be valid generally, it is still technically a conjecture and its rigorous construction has not yet been completed. There are reasons to believe that studying HS theory could help us not only to understand string theory but also to elucidate this conjectured correspondence, particularly in the prominent example of type IIB string theory (a theory on $AdS_5 \times \mathcal{S}^5$ with five spacetime and five compact dimensions) and $\mathcal{N}=4$ supersymmetric Yang-Mills theory on its four-dimensional boundary. \subsection{Why not?} Of course, from the viewpoint of pure mathematics, one needs no justification for studying anything. But in the case of HS theory, there is more to it than just curiosity. We know that a rich mathematical structure emerges from the fully consistent (necessarily non-linear!) theory of spin-$2$ fields, i.e. \textbf{pseudo-Riemannian geometry}. Some say that Einstein's general theory of relativity was, in a sense, discovered prematurely, and it was only Einstein's deep geometric intuition that allowed him to make the leap to a fully geometrized description of gravity. Had we persisted on building the theory in a bottom-up way from a linear spin-$2$ theory, we might have not ended up with such an elegant theory years before the dawn of quantum field theory. It is not known at the time of writing whether fully consistent spin-$3$ and higher spin theories give rise to some exciting new connections between physics and mathematics, perhaps even hitherto unknown mathematical structures, but it doesn't seem so unlikely that they might. \newpage \section{Folk history of HS theory}\label{s3} One cannot fully appreciate the struggle to understand higher spins without its history. For that purpose, we review here the most important steps forward\footnotemark\footnotetext{Like in every area of research, some steps that did not quite lead forward have been made, which was not understood at the time. Today, we understand more, so we can only pick those results that lead somewhere, hence the title \textit{folk} history.} in understanding higher spins, and we use this opportunity to expose the very basics of the theory. We restrict our attention to fields with integral spin, since those are the focus of this thesis. \subsection{Fierz-Pauli equations (1939)} \label{fpe} Fierz and Pauli constructed a consistent set of equations describing free massive fields of arbitrary spin \cite{fierzpauli}. They start with the Klein-Gordon equation, \begin{equation} (\Box - M^2) \phi = 0 \, , \end{equation} which describes spin-$0$ fields, and generalize it directly to higher spins, imposing additional consistency constraints. The Fierz-Pauli equations describing a spin-$s$ field are \begin{align} \phi_{\mu_1 \cdots \mu_s} &= \phi_{(\mu_1 \cdots \mu_s)}\, , \label{FP1} \\ (\Box - M^2) \phi_{\mu_1 \cdots \mu_s} &= 0 \, , \label{FP2} \\ \partial^{\mu_1} \phi_{\mu_1 \cdots \mu_s} &= 0 \, , \label{FP3} \\ \eta^{\mu_1 \mu_2} \phi_{\mu_1 \cdots \mu_s} &= 0 \, , \label{FP4} \end{align} where $(\dots)$ denotes the normalized symmetrization of indices. Equation \eqref{FP1} establishes the field as a fully symmetric tensor of order $s$. This condition ensures that the field transforms in accordance with the desired spin representation. To ensure that it is an irreducible representation, it must be traceless, which is guaranteed by \eqref{FP4}. Finally, \eqref{FP3} imposes the transversality condition, needed for the field to propagate the correct number of degrees of freedom, which we calculate in the following segment. From group-theoretical considerations, we expect all\footnotemark\, massless bosons in four-dimensional Minkowski spacetime to propagate exactly \textbf{two} independent degrees of freedom.\footnotetext{Except scalar bosons, which always have a single degree of freedom. More precisely, \textit{all} massless bosons have a single degree of freedom, but it gets doubled due to parity transformations, except in the case of scalars, whose irreducible representations are one-dimensional.} In general, a spin-$s$ field in $D$-dimensional spacetime should propagate\footnotemark \begin{equation} \#(D-2,s) - \#(D-2, s-2) = {{D+s-3} \choose {s}} - {{D+s-5} \choose {s-2}} \end{equation} independent degrees of freedom, as can be seen for example, using Wigner's classification. $\#(D,s)$ denotes the number of independent components of a fully symmetric tensor of order $s$ in $D$-dimensional spacetime. It is a simple exercise in combinatorics to check that $\#(D,s)={{D+s-1} \choose {s}}$. A spin-$s$ field is described by a totally symmetric doubly traceless tensor of order $s$, which contains \begin{equation} \#(D, s) - \#(D, s-4) = \underbrace{{{D+s-1} \choose {s}}}_{\parbox{2cm}{\tiny{components of a\\ symmetric tensor}}} - \underbrace{{{D+s-5} \choose {s-4}}}_{\parbox{1.8cm}{\tiny{components of \\ its second trace}}} \end{equation} independent components. Gauge invariance eliminates the propagation of spurious degrees of freedom ($2 s^2$ components in $D=4$), leaving the correct number of remaining degrees of freedom. For a detailed calculation, see, for example, \cite{weinbergQFT} or \cite{BBIRREPs}. \subsection{Singh-Hagen Lagrangian (1974) } It was not until 35 years later that the proper Lagrangian formulation of Fierz-Pauli equations was constructed, by Singh and Hagen \cite{singhhagen}. The fundamental obstacle lay in the need for auxiliary non-dynamical fields of spins $s-2, s-3, \dots$, along with the spin-$s$ field. Let us motivate their construction by starting with the trivial example of $s=1$, and then proceeding to the first non-trivial case of $s=2$. \subsubsection{Spin-1: no auxiliary fields} A spin-$1$ field is described by $\phi_\mu(x)$, a Lorentz-tensor of order one (i.e. a 4-vector). Since it only has a single index, we do not have to worry about equation \eqref{FP4}, nor do we have to worry about the symmetry condition \eqref{FP1}. The Lagrangian for $s=1$ is the Proca Lagrangian \begin{equation} \mathcal{L}_{(1)} = -\frac{1}{2} (\partial_\mu \phi_\nu)^2 + \frac{1}{2} (\partial \cdot \phi )^2 - \frac{M^2}{2} \phi^2 \label{SH1} \end{equation} which produces the Proca equation of motion, \begin{equation} \Box \phi_\mu - \partial_\mu (\partial \cdot \phi) - M^2 \phi_\mu = 0 \, . \label{proca} \end{equation} At first glance, this is not equal to \eqref{FP2} for $s=1$, but a single divergence of \eqref{proca} gives \begin{equation} \partial^\mu \phi_\mu = 0 \, , \end{equation} which gives the transversality condition \eqref{FP3}. Putting this back into \eqref{proca}, we are indeed left with the spin-$1$ version of equation \eqref{FP2}, \begin{equation} (\Box - M^2) \phi_\mu = 0 \, . \end{equation} \newpage \subsubsection{Spin-2: scalar auxiliary field} Following \eqref{FP1} and \eqref{FP4}, a spin-$2$ field is described by $\phi_{\mu \nu}(x)$, a symmetric traceless Lorentz-tensor of order two. Instead of directly generalizing \eqref{SH1} to the spin-$2$ case by using a tensor of order two, we write the Lagrangian with an undetermined real parameter $\alpha$ in place of $1$, \begin{equation} \mathcal{L}_{(2)} = -\frac{1}{2} (\partial_\mu \phi_{\nu\rho})^2 + \frac{\alpha}{2} (\partial \cdot \phi_\mu )^2 - \frac{M^2}{2} \phi^2 \, . \label{SH2} \end{equation} The reason for doing so will become apparent soon. The corresponding equation of motion\footnotemark\, is found to be \footnotetext{$\mathcal{L}_{(2)}$ is varied taking into consideration the symmetry and the tracelessness of $\phi_{\mu \nu}$} \begin{equation} \Box \phi_{\mu \nu} - \frac{\alpha}{2} \left( \partial_\mu \partial \cdot \phi_\nu + \partial_\nu \partial \cdot \phi_\mu - \frac{2}{D} \eta_{\mu \nu} \partial^2 \cdot \phi \right) - M^2 \phi_{\mu \nu} = 0 \, , \label{sheom} \end{equation} where $D$ is the dimension of spacetime. A single divergence of \eqref{sheom} gives \begin{equation} \left( 1 - \frac{\alpha}{2} \right) \Box \partial \cdot \phi_\mu + \alpha \left( \frac{1}{D} - \frac{1}{2} \right) \partial_\mu \partial^2 \cdot \phi - M^2 \partial \cdot \phi_\mu = 0 \, . \label{shtrouble} \end{equation} We seem to be in trouble, because (assuming $D>2$) we can only partially restore the transversality condition $\eqref{FP3}$ by setting $\alpha=2$, which eliminates the first term in \eqref{shtrouble}. Had we generalized \eqref{SH1} directly, instead of leaving $\alpha$ undetermined, we would not have been able to eliminate it. To eliminate the second term in \eqref{shtrouble}, we introduce the auxiliary scalar field $\pi(x)$ by adding to $\mathcal{L}_{(2)}$ (with $\alpha=2$) additional terms with two undetermined real parameters, $c_1$ and $c_2$, \begin{equation} \mathcal{L}_\pi = \pi \partial^2 \cdot \phi + c_1 (\partial_\mu \pi)^2 + c_2 \pi^2 \, . \end{equation} The corresponding equations of motion for $\mathcal{L} = \mathcal{L}_{(2)}\big\rvert_{\alpha=2} + \mathcal{L}_\pi$ are found to be \begin{align} \phi :& \quad \Box \phi_{\mu \nu} - \partial_\mu \partial \cdot \phi_\nu - \partial_\nu \partial \cdot \phi_\mu + \frac{2}{D} \eta_{\mu \nu} \partial^2 \cdot \phi - M^2 \phi_{\mu\nu} + \partial_\mu \partial_\nu \pi - \frac{1}{D} \eta_{\mu \nu} \Box \pi = 0 \, , \label{phieq} \\ \pi :& \quad \partial^2 \cdot \phi + 2(c_2 - c_1 \Box) \pi = 0 \, . \label{sheom2pi} \end{align} Taking twice the divergence of \eqref{phieq}, i.e. contracting it by $\partial_\mu \partial_\nu$, and multiplying it by $D$, yields \begin{equation} \left( (2-D) \Box - D M^2 \right) \partial^2 \cdot \phi + (D-1) \Box^2 \pi = 0 \, .\label{sheom2phi} \end{equation} The two equations, \eqref{sheom2pi} and \eqref{sheom2phi}, can be seen as a linear homogeneous system in $\partial^2 \cdot \phi$ and $\pi$. The system is solved by requiring that its determinant be non-vanishing and purely algebraic (without $\Box$ operators). Fortunately, this is possible if we choose \begin{align} c_1 &= \frac{D-1}{2(D-2)} \, , \\ c_2 &= \frac{D(D-1)M^2}{2(D-2)^2} \, . \end{align} This way, the only solution of the linear system \eqref{sheom2pi}-\eqref{sheom2phi} is $\pi = 0$ and $\partial^2 \cdot \phi = 0$, which we plug into \eqref{shtrouble} with $\alpha=2$ to obtain the Fierz-Pauli transversality condition \eqref{FP3} for $s=2$, \begin{equation} \partial^\mu \phi_{\mu\nu} = 0 \, . \end{equation} Finally, plugging the transversality condition and the solution $\partial^2 \cdot \phi = 0$ into \eqref{sheom}, we get the Fierz-Pauli equation of motion \eqref{FP2} for $s=2$, \begin{equation} (\Box - M^2) \phi_{\mu\nu} = 0 \, . \end{equation} A similar procedure with $s-1$ auxiliary fields was shown to yield the correct Lagrangian for spin-$s$ fields, equivalent to the Fierz-Pauli equations \cite{singhhagen}. \subsection{Fronsdal equation (1978)}\label{fronsdaleq} Soon after Singh and Hagen, Fronsdal investigated the massless case, taking the $M \to 0$ limit of their Lagrangian formulation \cite{fronsdal}. In this limit, only the spin-$s$ and the first auxiliary spin-$s-2$ field survive, while all the lower spin auxiliary fields decouple. Furthermore, the remaining two fields can be neatly packed into a single field, with additional consistency constraints. Let us demonstrate here what happens in the spin-$2$ case. \subsubsection{Spin-$2$ Fronsdal equation} We start from the $M \to 0$ limit of the Singh-Hagen Lagrangian for $s=2$, \begin{equation} \mathcal{L} = -\frac{1}{2} (\partial_\mu \phi_{\nu\rho})^2 + (\partial \cdot \phi_\mu)^2 + \pi \partial^2 \cdot \phi + \frac{D-1}{2(D-2)} (\partial_\mu \pi)^2 \, . \end{equation} Next, we redefine $\pi$ and $\phi_{\mu \nu}$ into a new field $\varphi_{\mu\nu}$, \begin{equation} \varphi_{\mu\nu} := \phi_{\mu \nu} + \frac{1}{D-2} \eta_{\mu \nu} \pi \, , \end{equation} which is no longer traceless. The resulting Lagrangian is \begin{equation} \mathcal{L} = -\frac{1}{2} (\partial_\mu \varphi_{\nu\rho})^2 + (\partial \cdot \varphi_\mu)^2 + \frac{1}{2} (\partial_\mu \varphi)^2 + \varphi \partial^2 \cdot \varphi \, . \end{equation} Note that this is exactly the linearized Einstein-Hilbert Lagrangian, an important fact to which we will return later. The equation of motion that follows from this Lagrangian is \begin{equation} \Box \varphi_{\mu\nu} - (\partial_\mu \partial \cdot \varphi_\nu + \partial_\nu \partial \cdot \varphi_\mu) + \partial_\mu \partial_\nu \varphi + \eta_{\mu \nu} \left( \partial^2 \cdot \varphi - \Box \varphi \right) = 0 \, , \label{f2eom} \end{equation} which is precisely the free linearized Einstein equation, \begin{equation} G_{\mu\nu}^{(lin)} = R_{\mu\nu}^{(lin)} - \frac{1}{2} \eta_{\mu \nu} R^{(lin)} = 0 \, , \end{equation} but more on this in \textbf{Section \ref{spin2}}. We define $\mathcal{F}_{\mu\nu}$ as the \textit{Fronsdal tensor} of order two, \begin{equation} \mathcal{F}_{\mu\nu} = \Box \varphi_{\mu\nu} - (\partial_\mu \partial \cdot \varphi_\nu + \partial_\nu \partial \cdot \varphi_\mu) + \partial_\mu \partial_\nu \varphi \, , \end{equation} which is equal to $R_{\mu\nu}^{(lin)}$. Note that we can now write \eqref{f2eom} as \begin{equation} \mathcal{F}_{\mu\nu} - \frac{1}{2} \eta_{\mu \nu} \mathcal{F} = 0 \, , \label{fgeom1} \end{equation} which simply reduces to \begin{equation} \mathcal{F}_{\mu \nu} = \Box \varphi_{\mu\nu} - (\partial_\mu \partial \cdot \varphi_\nu + \partial_\nu \partial \cdot \varphi_\mu) + \partial_\mu \partial_\nu \varphi = 0 \, . \label{fgeom2} \end{equation} This is the \textbf{Fronsdal equation} for spin-$2$ fields. Note that \eqref{fgeom1} can reduce to \eqref{fgeom2} only because the theory is \textit{free}, analogous to the reduction of Einstein field equations to the vanishing of $R_{\mu\nu}$ in vacuum. Fronsdal equation is invariant under the gauge transformation \begin{equation} \delta \varphi_{\mu \nu} = \partial_\mu \Lambda_\nu + \partial_\nu \Lambda_\mu \, . \end{equation} This fact will be particularly important when we begin investigating the theory in detail. \subsubsection{Spin-$3$ Fronsdal equation} We can try to generalize the spin-$2$ case directly on a totally symmetric (but not traceless!) tensor of order three\footnotemark, \footnotetext{Here, we begin to use a prime to denote a trace, e.g. $\varphi'_\nu := \varphi^\mu{}_{\mu\nu}$.} \begin{align} \mathcal{F}_{\mu\nu\sigma} &= \Box \varphi_{\mu\nu\sigma} - (\partial_\mu \partial \cdot \varphi_{\nu \sigma} + \partial_\nu \partial \cdot \varphi_{\sigma \mu} + \partial_\sigma \partial \cdot \varphi_{\mu \nu}) \\ \nonumber &+ \partial_\mu \partial_\nu \varphi'_\sigma + \partial_\nu \partial_\sigma \varphi'_\mu + \partial_\nu \partial_\sigma \varphi'_\mu = 0\, . \end{align} The corresponding generalized gauge transformations reads \begin{equation} \delta \varphi_{\mu\nu\sigma} = \partial_\mu \Lambda_{\nu\sigma} + \partial_\nu \Lambda_{\sigma\mu} + \partial_\sigma \Lambda_{\mu\nu} \, , \end{equation} but unlike in the case of $s=2$, now we do not have a fully gauge-invariant Fronsdal tensor. Instead, \begin{equation} \delta \mathcal{F}_{\mu\nu\sigma} = 3 \partial_\mu \partial_\nu \partial_\sigma \Lambda' \, . \end{equation} In his original formulation\cite{fronsdal}, Fronsdal circumvents this problem by simply restricting the space of gauge parameters to ones that are traceless, i.e. by imposing the unusual constraint \begin{equation} \Lambda' = 0 \, . \end{equation} This amounts to restricting ourselves to a subclass of gauge transformations, instead of having fully unrestricted gauge invariance. \subsubsection{Spin-$s$ Fronsdal equation} The traceless $\Lambda$ constraint leaves us with a fully consistent gauge-invariant theory of free higher spin fields obeying the spin-$s$ \textit{Fronsdal equation}\footnotemark, \begin{equation} \mathcal{F}_{\mu_1 \cdots \mu_s} = \Box \varphi_{\mu_1 \cdots \mu_s} - (\partial_{\underline{\mu_1}} \partial \cdot \varphi_{\underline{\mu_2 \cdots \mu_s}}) + \partial_{\underline{\mu_1}} \partial_{\underline{\mu_2}} \varphi'_{\underline{\mu_3 \cdots \mu_{s} }} = 0 \, . \end{equation}\footnotetext{Underlined indices stand for unweighted symmetrization with the minimal number of terms.} \subsubsection{Fronsdal Lagrangian} Fronsdal started with the Singh-Hagen Lagrangian formulation and naturally, he wanted to describe his theory using an action principle. The Lagrangian that makes this possible is \begin{equation} \mathcal{L}_{\mathcal{F}} = \frac{1}{2} \varphi^{\mu_1 \cdots \mu_s} \left( \mathcal{F}_{\mu_1 \cdots \mu_s} - \frac{1}{2} \eta_{\underline{\mu_1 \mu_2}} \mathcal{F}'_{\underline{\mu_3 \cdots \mu_s}} \right) \, . \label{flang} \end{equation} As we show through explicit calculation in \textbf{Section \ref{freefronsdal}}, where we switch to a simpler notation, \eqref{flang} indeed yields the Fronsdal equation for $s<4$. For spins higher than four, we have to impose another unusual constraint, \begin{equation} \varphi'' = 0 \end{equation} if we are to arrive at the Fronsdal equation of motion $\mathcal{F}=0$. \subsection{Vasiliev's equations (1990) } M.A.Vasiliev successfully constructed a fully consistent non-linear theory of interacting higher spin fields in (anti-)de Sitter spacetimes\cite{vasiliev90}. The equations are notoriously complicated and since we will be dealing with massless bosonic fields in flat spacetime, we will not reproduce them here. It suffices to quote \cite{vasiliev}:\newline \q{The shortest route to Vasiliev equations covers 40 pages.}\newline \q{It is a sort of conventional wisdom that Vasiliev equations cannot be derived...} Similarly to string theory, Vasiliev's theory in spacetime dimensions four and higher can be consistent only if it contains an infinite tower of higher-spin fields. Only in dimensions three and lower can it be consistent with an upper limit on spin. \subsection{No-go theorems}\label{s4} Throughout the history of HS theory, several important results have been obtained that severely constrain the properties of would-be interacting theories of higher spin fields. Vasiliev's theory\cite{vasiliev90}\cite{vasiliev} shows that the class of such theories is not empty, but we have yet to arrive at other theories of this kind. We list here some of the most important \textit{no-go} theorems. For a more detailed discussion, see \cite{nogo},\cite{nogo2} and references therein. \paragraph{No long-range HS interactions\newline} Using the \textit{S-matrix} approach, Weinberg proved in 1964 that there are no consistent \textbf{long-range} interactions by massless bosons with spin greater than two\cite{weinberg}. \paragraph{No local Lagrangians in HS theories\newline} Using the local \textit{Lagrangian formalism} and working in the \textit{soft limit}, Aragone and Deser proved in 1979 \cite{AragoneDeserNoGo1} (see also \cite{AragoneDeserNoGo2}) that HS fields cannot consistently interact with gravity. Since gravitational interaction is universal, this implies that there can be no consistent interacting HS fields. \paragraph{No massless HS interactions in flat spacetime\newline} The Weinberg-Witten theorem\cite{WeinbergWitten} from 1980 states that no massless HS field can consistently interact with gravity in \textbf{flat spacetime}.\newline It is important to keep in mind that all no-go theorems start with some underlying assumptions that are not obviously satisfied in all physically possible cases. Therefore, the effort to construct consistent interacting HS theories might not be a fool's errand after all. \newpage \section{Review of lower spin theories}\label{s5} Instead of jumping head-first into some deeper problems of higher spin theory, let us review the familiar territory of lower spin bosonic theories. \subsection{Spin-0 theory} Fields of spin $0$ are described by Lorentz scalars. The general Lagrangian for these fields is \begin{equation} \mathcal{L}_{0} [\phi]= \frac{1}{2} ( \partial_\mu \phi )^2 - \frac{m^2}{2} \phi^2 \, , \end{equation} and it produces the equation of motion for scalar fields, the \textbf{Klein-Gordon equation}, \begin{equation} (\Box + m^2) \phi = 0 \, . \end{equation} \subsubsection{Example: Higgs boson} \textbf{Higgs field} is a well-known example of a scalar field, and it is the only scalar fundamental field in the Standard Model. It is a \textit{complex} scalar field, described by the Lagrangian \begin{equation} \mathcal{L}_H = \left| \partial_\mu \phi \right|^2 - V(\phi) \, . \end{equation} \subsection{Spin-1 theory} Fields of spin $1$ are described by Lorentz vectors. The general Lagrangian for these fields is \begin{equation} \mathcal{L}_1 [A^\mu]= -\frac{1}{2} F_{\mu \nu} F^{\mu \nu} + \frac{m^2}{2} A_\mu A^\mu \, , \end{equation} with $F_{\mu \nu} = \partial_\mu A_\nu - \partial_\nu A_\mu$, and it produces the \textbf{Proca equation}, \begin{equation} \Box A^\nu - \partial^\nu ( \partial_\mu A^\mu ) + m^2 A^\nu = 0 \, . \end{equation} \subsubsection{Example: Maxwell's electrodynamics} In the Standard Model, four vector bosons take part in the \textbf{electroweak} interaction, the \textbf{photon} and three intermediate bosons, $W^\pm$ and $Z^0$. The free massive intermediate boson fields satisfy the Proca equation while the massless photon field satisfies \textbf{Maxwell's equations}, \begin{equation} \partial_\mu F^{\mu \nu} = j^\nu. \end{equation} The strong force is also mediated by vector bosons, described by the massless \textbf{gluon} field. Before addressing the spin-2 theory, let us briefly discuss the issue of gauge invariance. \paragraph{Spin-1 gauge invariance:} A massless spin-$1$ field $A^\nu (x)$ has 4 components, and it satisfies the equation of motion \begin{equation} \Box A^\nu - \partial^\nu (\partial_\lambda A^\lambda ) = 0. \end{equation} This theory is invariant under the Abelian gauge transformation \begin{equation} \delta A_\mu (x) = \partial_\mu \Lambda (x) \, . \end{equation} The equation of motion can be cast into a simple wave equation form, \begin{equation} \Box A_\mu (x) = 0 \, , \end{equation} by choosing the Lorentz-invariant \textit{Lorenz gauge}, \begin{equation} \partial_\mu A^\mu (x) = 0. \end{equation} This choice is a scalar constraint, which eliminates one of two spurious degrees of freedom, but there is a degree of gauge freedom left, i.e. \begin{equation} \delta (\partial_\mu A^\mu) = \Box \Lambda = 0. \end{equation} This is also a scalar constraint, so we are indeed left with two propagating degrees of freedom. \subsection{Spin-2 theory}\label{spin2} Fields of spin $2$ are described by symmetric Lorentz tensors of order two. The general Lagrangian for these fields is \begin{equation} \mathcal{L}_2 [h^{\mu\nu}] = -\frac{1}{2} (\partial_\sigma h_{\mu\nu})^2 + \partial_\sigma h_{\mu\nu} \partial^\mu h^{\nu\sigma} - \partial \cdot h_\nu \partial^\nu h + \frac{1}{2} (\partial_\mu h)^2 \, , \label{Ls2} \end{equation} and it produces the equation of motion \begin{equation} \Box h_{\mu\nu} - \partial_\mu \partial \cdot h_\nu - \partial_\nu \partial \cdot h_\mu + \partial_\mu \partial_\nu h + \eta_{\mu\nu} \partial^2 \cdot h - \eta_{\mu\nu} \Box h = 0 \, . \label{Es2} \end{equation} \subsubsection{Example: General Relativity} \textbf{Einstein's General Relativity} is the archetypal example of a spin-2 theory. It describes gravitation, and it is the only spin-$2$ theory found in nature. In its full form, general relativity is highly nonlinear, and it is described by \textbf{Einstein field equations}, \begin{equation} G_{\mu \nu} \equiv R_{\mu \nu} - \frac{1}{2} g_{\mu \nu} R = 8 \pi T_{\mu \nu} , \end{equation} where $G_{\mu \nu}$ is the \textbf{Einstein tensor}, $R_{\mu \nu}$ is the \textbf{Ricci tensor}, $R$ is the \textbf{Ricci scalar} and $T_{\mu \nu}$ is the matter \textbf{energy-momentum} tensor. The \textbf{Riemann curvature tensor} can be defined as \begin{equation} R^\rho{}_{\sigma\mu\nu} = \partial_\mu \Gamma^\rho{}_{\nu \sigma} - \partial_\nu \Gamma^\rho{}_{\mu \sigma} + \Gamma^\rho{}_{\mu \lambda} \Gamma^\lambda{}_{\nu \sigma} - \Gamma^\rho{}_{\nu \lambda} \Gamma^\lambda{}_{\mu \sigma} \, , \end{equation} using the \textit{torsionless connection} $\Gamma^\rho{}_{\mu\nu} = \Gamma^\rho{}_{\nu\mu}$, \begin{equation} \Gamma^\rho{}_{\mu\nu} = \frac{1}{2} g^{\rho\lambda} (\partial_\mu g_{\nu \lambda} + \partial_\nu g_{\mu \lambda} - \partial_\lambda g_{\mu \nu}) \, . \end{equation} Ricci tensor and scalar are simply given by \begin{align} R_{\mu \nu} &= R^\rho{}_{\mu \rho \nu} \, ,\\ R &= R^\lambda{}_\lambda \, . \end{align} \subsubsection{Example: Linearized Gravity} By considering small metric perturbations from the flat Minkowski spacetime, we can construct a linear theory of a dynamical spin-2 field in a static flat background. Explicitly, we decompose the metric so that \begin{equation} g_{\mu \nu} (x) = \eta_{\mu \nu} + h_{\mu \nu} (x) + \mathcal{O}(h^2) \, , \end{equation} and we truncate the expansion to first order in $h_{\mu\nu}$, assuming $\| h(x) \| \ll 1$. The resulting theory is what we call \textbf{linearized gravity}, and it is described by \textit{linearized} Einstein field equations, \begin{equation} G^{(lin)}_{\mu \nu} = R^{(lin)}_{\mu \nu} - \frac{1}{2} \eta_{\mu \nu} R^{(lin)} = T^{(lin)}_{\mu\nu}. \end{equation} Or, using the metric perturbation field explicitly, \begin{align} R^{(lin)}_{\mu \nu} &= \Box h_{\mu \nu} - \partial_\mu (\partial^\lambda h_{\lambda \nu}) - \partial_\nu (\partial^\lambda h_{\lambda \mu}) + \partial_\mu \partial_\nu h^\lambda{}_\lambda \, , \\ R^{(lin)} &= 2 \Box h^\lambda{}_\lambda -2 \partial^\lambda \partial^\sigma h_{\lambda \sigma} \, , \\ G^{(lin)}_{\mu \nu} &= \Box h_{\mu \nu} - \partial_\mu (\partial^\lambda h_{\lambda \nu}) - \partial_\nu (\partial^\lambda h_{\lambda \mu}) + \partial_\mu \partial_\nu h^\lambda{}_\lambda + \eta_{\mu \nu} \partial^\lambda \partial^\sigma h_{\lambda \sigma} - \eta_{\mu\nu}\Box h^\lambda{}_\lambda \, . \end{align} Note that $G^{(lin)}_{\mu \nu} = 0$ corresponds to \eqref{Es2}. This is no coincidence, since \eqref{Ls2} precisely describes the Lagrangian for linearized gravity in the absence of sources, i.e. with \newline$T^{(lin)}_{\mu\nu}=0$. \paragraph{Spin-2 gauge invariance:} The spin-$2$ field is described by a doubly traceless tensor $h_{\mu\nu} (x)$ of rank two and therefore has $\mathbf{10}$ independent components. In free theory, it satisfies the equation of motion \begin{equation} R^{(lin)}_{\mu\nu} = \Box h_{\mu\nu} - \partial_\mu (\partial^\lambda h_{\lambda\nu}) - \partial_\nu (\partial^\lambda h_{\mu\lambda}) + \partial_\mu \partial_\nu h^\lambda{}_\lambda = 0 \, . \end{equation} This theory is invariant under the Abelian gauge transformation \begin{equation} \delta h_{\mu \nu} (x) = \partial_\mu \xi_\nu (x) + \partial_\nu \xi_\mu (x) \, , \end{equation} which allows us to cast the above equation into a simple wave equation form, \begin{equation} \Box h_{\mu \nu} (x) = 0 \, , \end{equation} by choosing the Lorentz-invariant \textit{de Donder Gauge}\footnotemark, \begin{equation} \mathcal{D}_\mu (x) \equiv \partial^\lambda h_{\lambda \mu} - \frac{1}{2} \partial_\mu h^\lambda{}_\lambda = 0. \end{equation} \textit{De Donder tensor} $\mathcal{D}_\mu$ is a $4$-vector, so we are left with $10 - 4 = \mathbf{6}$ degrees of freedom. Fixing the gauge in this way does not eliminate the gauge freedom completely. This can be seen from the \textit{de Donder} gauge condition, since \footnotetext{Also known as the \textit{harmonic gauge}, \textit{Lorentz gauge}, \textit{Einstein gauge}, \textit{Hilbert gauge} or \textit{Fock gauge}.} \begin{equation} \delta \mathcal{D}_\mu (x) = \Box \xi_\mu (x) = 0. \end{equation} This too is a $4$-vector constraint, which eliminates the remaining $4$ spurious degrees of freedom, leaving us with $6 - 4 = \mathbf{2}$ propagating degrees of freedom, as expected. \newpage \section{Higher spin theory of massless bosons}\label{s6} \subsection{Francia-Sagnotti formalism} There exists an elegant formalism\footnotemark\, developed by D.Francia and A.Sagnotti\cite{fs1} \cite{fs2} \cite{st} \cite{introfree} \cite{fs3} \cite{fms} \cite{dariomass} \cite{dariopropm} \cite{dariocrete} that makes it easy to express and manipulate most of mathematical objects of HS theory in the linear approximation.\footnotetext{To be fair, it would be more precise to call it \textit{notation}, but as Feynman said\cite{feynmannotation}: \q{We could, of course, use any notation we want; do not laugh at notations; invent them, they are powerful. In fact, mathematics is, to a large extent, invention of better notations.}} This formalism is suitable for higher spin theory since the tensorial indices and spin are left implicit, but are easily recovered. A spin-$s$ field is simply written as \begin{equation} \phi_{\mu_1 \cdots \mu_s} \equiv \phi \, . \end{equation} The $n$-th gradient of $\phi$ is written as $\partial^n \phi$, the $n$-th divergence\footnotemark\, as $\partial^n \cdot \phi$ and the $n$-th trace as $\phi^{[n]}$. Lower traces are simply written with a prime, e.g. $\phi''$ for the second trace. \footnotetext{Where Francia and Sagnotti would use (for example) $\partial \cdot \partial \cdot \partial \cdot \varphi$, here we use $\partial^3 \cdot \varphi $ instead. This simplification seems to produce no ambiguities, as the reader is welcome to check.} All indices are implicitly symmetrized, without weight factors, using the minimal number of terms. For example, if $s=2$, \begin{align} \partial^2 \phi &\equiv \partial_\mu \partial_\nu \phi_{\sigma \rho} + \partial_\mu \partial_\sigma \phi_{\nu \rho} + \partial_\mu \partial_\rho \phi_{\sigma \nu} + \partial_\nu \partial_\sigma \phi_{\mu \rho} + \partial_\nu \partial_\rho \phi_{\sigma \mu} + \partial_\sigma \partial_\rho \phi_{\mu \nu} \, , \label{ex1} \\ \partial (\partial \cdot \phi) &\equiv \partial_\nu (\partial^\lambda \phi_{\mu \lambda}) + \partial_\mu (\partial^\lambda \phi_{\lambda \nu}) \, , \label{ex2} \\ \eta \partial^2 \cdot \phi &\equiv \eta_{\mu\nu} \partial^\rho \partial^\sigma \phi_{\rho \sigma} \, . \label{ex3} \end{align} The formalism implies the following set of rules: \begin{align} ( \partial^p \phi )' &= \Box \partial^{p-2} \phi + 2 \partial^{p-1} \left( \partial \cdot \phi \right) + \partial^p \phi' \label{fs1} \\ \partial \cdot (\partial^p \phi) &= \Box \partial^{p-1} \phi + \partial^p \left( \partial \cdot \phi \right) \label{fs2} \\ \left( \eta^k T_{(s)} \right)' &= [D + 2(s+k-1)] \eta^{k-1} T_{(s)} + \eta^k T_{(s)}' \label{traces} \\ \partial^p \partial^q &= {{p+q}\choose{q}} \partial^{p+q} \\ \eta^p \eta^q &= {{p+q}\choose{q}} \eta^{p+q} \\ \partial \cdot \eta^{p} &= \eta^{p-1} \partial \end{align} \eqref{fs1} and \eqref{fs2} can further be generalized to: \begin{align} ( \partial^n \phi )^{[p]} &= \sum_{k=0}^{p} \sum_{l=0}^{k} {p \choose k} {k \choose l} 2^l \Box^{p-k} \partial^{n - 2p + 2k -l} \left( \partial^l \cdot \phi^{[k-l]} \right) \\ \partial^n \cdot ( \partial^p \phi ) &= \sum_{k=0}^{n} {n \choose k} \Box^{n-k} \partial^{p-n+k} \left( \partial^k \cdot \phi \right) \label{gen2} \end{align} The relations \eqref{fs1}-\eqref{gen2} will prove to be useful in simplifying our calculations. Note that this formalism is also implicit in the dimension of spacetime, as long as the relevant expressionts do not include traces of terms containing the metric tensor, as implied by \eqref{traces}. We introduce "$ \scalerel*{\cdot}{\bigodot} $" to denote maximal contraction between two tensors\footnotemark.\footnotetext{Francia and Sagnotti do not use this notation. Instead, such contractions are left implicit, which may look confusing to the untrained eye.} For tensors $\varphi$ of order $s$ and $\chi$ of order $r$, with $s>r$, the contraction is defined as \begin{equation} \varphi \scalerel*{\cdot}{\bigodot} \chi \equiv \varphi_{\mu_1 \cdots \mu_r \mu_{r+1} \cdots \mu_s} \chi^{\mu_1 \cdots \mu_r} \, , \end{equation} where both tensors are assumed to be symmetrized with the minimal number of unweighted terms, before contraction. For example, if $\varphi$ is a tensor of order three and $\chi$ is a tensor of order two, \begin{align} \varphi \scalerel*{\cdot}{\bigodot} \chi &\equiv \varphi_{\mu\nu\sigma} \chi^{\mu \nu} \, , \\ \varphi \scalerel*{\cdot}{\bigodot} \partial \chi &\equiv \varphi_{\mu\nu\sigma} \left( \partial^\mu \chi^{\nu\sigma} + \partial^\nu \chi^{\sigma\mu} + \partial^\sigma \chi^{\mu\nu} \right) = 3 \varphi_{\mu\nu\sigma} \partial^\mu \chi^{\nu\sigma} \, , \\ \partial \varphi \scalerel*{\cdot}{\bigodot} \eta \chi &\equiv \left( \partial_\mu \varphi_{\nu\sigma\rho} + \partial_\nu \varphi_{\sigma\rho\mu} + \partial_\sigma \varphi_{\rho\mu\nu} + \partial_\rho \varphi_{\mu\nu\sigma} \right) \\ \nonumber &\left( \eta^{\mu\nu} \chi^{\sigma\rho} + \eta^{\mu\sigma} \chi^{\nu\rho} + \eta^{\mu\rho} \chi^{\sigma\nu} + \eta^{\nu\sigma} \chi^{\mu\rho} + \eta^{\nu\rho} \chi^{\mu\sigma} + \eta^{\sigma\rho} \chi^{\mu\nu} \right) \\ \nonumber &= 12 \partial^\sigma \varphi_{\sigma \mu \nu} \chi^{\mu\nu} + 12 \chi^{\mu\nu} \eta^{\sigma\rho} \partial_\mu \varphi_{\nu\sigma\rho} \, . \end{align} In other words, \textit{first} we symmetrize the tensors as in examples \eqref{ex1}-\eqref{ex3}, and \textit{then} we contract them. This notation will prove to be particularly useful in the analysis of actions and their variations. In the following segments, when we vary a Lagrangian, we will always vary it under the integral sign, as a variation of the action, i.e. \begin{equation} \delta \mathcal{S} [\varphi(x)] = \delta \int d^D x \, \mathcal{L}[\varphi(x)] = \int d^D x \, \delta \mathcal{L}[\varphi(x)] \, . \end{equation} When calculating such variations, we will often encounter terms of the form \begin{equation} \int d^D x \, A(x) \delta (\partial_\mu B(x)) \, , \end{equation} where we perform partial integration to obtain \begin{equation} - \int d^D x \, \partial_\mu A(x) \delta B(x) + \text{(boundary terms)} \, . \end{equation} The boundary terms vanish due to the standard assumption that all fields vanish at infinity and that there are no non-trivial topological features of spacetime. This allows us to use the following relation: \begin{equation} \int d^D x \, A(x) \delta (\partial_\mu B(x)) = - \int d^D x \, \partial_\mu A(x) \delta B(x) \, . \end{equation} In the Francia-Sagnotti formalism, one should be careful when performing partial integration, since this operation might produce additional symmetry factors. For example, if $\varphi$ is a symmetric tensor of order $s$ and $\Lambda$ is a symmetric tensor of order $s-1$, \begin{align} \int d^D x \, \partial \Lambda \scalerel*{\cdot}{\bigodot} \varphi &\equiv \int d^D x \, \left( \underbrace{\partial_{\mu_1} \Lambda_{\mu_2 \cdots \mu_{s}} + \dots + \partial_{\mu_s} \Lambda_{\mu_1 \cdots \mu_{s-1}}}_{s \, \text{terms}} \right) \varphi^{\mu_1 \cdots \mu_s} \\ &= s \int d^D x \, \partial_{\mu_1} \Lambda_{\mu_2 \cdots \mu_{s}} \varphi^{\mu_1 \cdots \mu_s} \\ &= -s \int d^D x \, \Lambda_{\mu_1 \cdots \mu_{s-1}} \partial_{\mu_s} \varphi^{\mu_1 \cdots \mu_s} \\ &\equiv -s \int d^D x \, \Lambda \scalerel*{\cdot}{\bigodot} \partial \cdot \varphi \, . \end{align} Similarly, one should be careful when writing terms of the form $\varphi \scalerel*{\cdot}{\bigodot} \eta \varphi'$ as terms of the form $\varphi' \scalerel*{\cdot}{\bigodot} \varphi'$, because \begin{align} \varphi \scalerel*{\cdot}{\bigodot} \eta \varphi' &\equiv \varphi_{\mu_1 \cdots \mu_s} \eta_{\nu \sigma} \left( \underbrace{ \eta^{\mu_1 \mu_2} \varphi^{\mu_3 \cdots \mu_s \nu \sigma} + \dots + \eta^{\mu_{s-1} \mu_s} \varphi^{\mu_1 \cdots \mu_{s-2} \nu \sigma}}_{{s \choose 2} \, \text{terms}} \right) \\ &= {s \choose 2} \eta^{\mu_1 \mu_2} \varphi_{\mu_3 \cdots \mu_s} \eta_{\mu_1 \mu_2} \varphi^{\mu_3 \cdots \mu_s} \\ &\equiv {s \choose 2} \varphi' \scalerel*{\cdot}{\bigodot} \varphi' \, . \end{align} To drive the point home, we provide two additional examples that we will encounter in our calculations: \begin{align} \int d^D x \, \Lambda \scalerel*{\cdot}{\bigodot} \partial^3 \varphi'' &= - {{s-1} \choose 3} \int d^D x \, \partial^3 \cdot \Lambda \scalerel*{\cdot}{\bigodot} \varphi'' \\ \Lambda \scalerel*{\cdot}{\bigodot} \eta \partial \cdot \varphi' &= {s-1 \choose 2} \Lambda' \scalerel*{\cdot}{\bigodot} \partial \cdot \varphi' \end{align} \subsection{Fronsdal's constrained theory}\label{fronsdalconstrained} \subsubsection{Free theory}\label{freefronsdal} As we have already seen in \textbf{Section \ref{fronsdaleq}}, Fronsdal's HS theory, in the absence of sources, consists of the Fronsdal equation along with two unusual constraints, i.e. \begin{align} \mathcal{F} &= \Box \varphi - \partial (\partial \cdot \varphi) + \partial^2 \varphi' = 0 \, , \label{freq} \\ \mathcal{L}_{\mathcal{F}} &= \frac{1}{2} \varphi \scalerel*{\cdot}{\bigodot} \left( \mathcal{F} - \frac{1}{2} \eta \mathcal{F}' \right) \, , \label{lagrangianF} \\ \delta \varphi &= \partial \Lambda \, , \label{gauge} \\ \Lambda' &= 0 \, \label{constr1},\\ \varphi'' &= 0 \, \label{constr2}. \end{align} Let us show how \eqref{lagrangianF} produces \eqref{freq} as the equation of motion in the absence of sources. Varying the action gives \begin{align} \delta \mathcal{S}_\mathcal{F} = \int d^D x \, \delta \mathcal{L}_\mathcal{F} &= \frac{1}{2} \int d^D x \, \left[ \delta \varphi \scalerel*{\cdot}{\bigodot} \left( \mathcal{F} - \frac{1}{2} \eta \mathcal{F}' \right) + \varphi \scalerel*{\cdot}{\bigodot} \left( \delta (\mathcal{F}) - \frac{1}{2} \eta (\delta \mathcal{F}') \right) \right] \\ &= \frac{1}{2} \int d^D x \, \Big\{ \delta \varphi \scalerel*{\cdot}{\bigodot} \left( \Box \varphi - \partial (\partial \cdot \varphi) + \partial^2 \varphi' + \eta \partial^2 \cdot \varphi - \eta \Box \varphi \right) \\ &+ \varphi \scalerel*{\cdot}{\bigodot} \left[ \delta (\Box \varphi) - \delta (\partial (\partial \cdot \varphi)) + \delta (\partial^2 \varphi') + \delta (\eta \partial^2 \cdot \varphi) - \delta (\eta \Box \varphi ) \right] \Big\} \nonumber \\ &= \int d^D x \, \left( \Box \varphi - \partial (\partial \cdot \varphi) + \partial^2 \varphi' + \eta \partial^2 \cdot \varphi - \eta \Box \varphi \right) \scalerel*{\cdot}{\bigodot} \delta \varphi \\ &= \int d^D x \, \left( \mathcal{F} - \frac{1}{2} \eta \mathcal{F}' \right) \scalerel*{\cdot}{\bigodot} \delta \varphi \, , \end{align} so the equation of motion reads \begin{equation} \mathcal{F} - \frac{1}{2} \eta \mathcal{F}' = 0 \, , \label{ffeq} \end{equation} which indeed reduces to \begin{equation} \mathcal{F} = 0 \, , \end{equation} since there are no sources on the right-hand side of \eqref{ffeq}. Armed with the powerful formalism, let us now take a closer look at the two constraints \eqref{constr1} and \eqref{constr2}. We would like to find where exactly they come from so that we can construct an equivalent \textit{unconstrained} theory. \paragraph{Why traceless $\Lambda$?\newline} The Fronsdal equation \eqref{freq} transforms under the gauge variation \eqref{gauge} as \begin{equation} \delta \mathcal{F} = 3 \partial^3 \Lambda' \label{FVar} \, , \end{equation} which is why we demand that the gauge parameter be traceless. If we could find an appropriate linear combination of fully gauge-invariant terms, we could formulate a theory without imposing this constraint. One way of getting around this would be through a differential constraint, \begin{equation} \partial^3 \Lambda' \, (x) = 0 \, , \end{equation} without directly constraining $\Lambda'$. If the gauge parameter $\Lambda(x)$ vanishes at infinity, the only solution would indeed be $\Lambda'=0$. Another way to dispense with this constraint is to introduce a non-dynamical spin-$(s-3)$ \textit{compensator} field $\alpha(x)$, which transforms under the gauge variation as \begin{equation} \delta \alpha = \Lambda' \, , \end{equation} and modify the equation of motion to \begin{equation} \mathcal{F} - 3 \partial^3 \alpha = 0 \, . \end{equation} If we introduce a second non-dynamical spin-$(s-4)$ field, this theory can be described by a Lagrangian, as we will explain in \textbf{Section \ref{localunconstrained}}. The third way to avoid the traceless $\Lambda$ is to work within a manifestly gauge-invariant geometric framework. Unfortunately, as we will see in \textbf{Section \ref{nonlocal}}, this forces us to abandon locality and instead work with non-local or higher-order (in derivatives) terms. \paragraph{Why doubly-traceless $\varphi$?\newline} The relation that lies at the heart of this constraint is the so-called \textit{anomalous\footnotemark Bianchi identity},\footnotetext{It is called \textit{anomalous} because it does not vanish. If the right-hand side vanishes, it is simply the \textit{Bianchi identity}.} \begin{equation} \partial \cdot \mathcal{F} - \frac{1}{2} \partial \mathcal{F}' = - \frac{3}{2} \partial^3 \varphi'' \, . \label{bianchi} \end{equation} We would like the Fronsdal action to be gauge invariant, so let us see what its gauge variation\footnotemark\, produces. Using \eqref{bianchi}, one obtains \footnotetext{We use $\delta_{\Lambda}$ to avoid confusing this variation with the usual functional variation $\delta$.} \begin{equation} \delta_{\Lambda} \mathcal{S}_\mathcal{F} = \int d^D x \, \delta_{\Lambda} \mathcal{L}_{\mathcal{F}} = \frac{1}{2} \int d^D x \, \left[ \partial \Lambda \scalerel*{\cdot}{\bigodot} \left( \mathcal{F} - \frac{1}{2} \eta \mathcal{F}' \right) + \varphi \scalerel*{\cdot}{\bigodot} \delta_{\Lambda} \left( - \frac{3}{2} \partial^3 \varphi'' \right) \right] \, . \end{equation} The second term under the integral vanishes if we impose $\Lambda' = 0$, so we have \begin{align} \delta_{\Lambda} \mathcal{S}_\mathcal{F} = \int d^D x \, \delta_{\Lambda} \mathcal{L}_{\mathcal{F}} &= \frac{1}{2} \int d^D x \, \partial \Lambda \scalerel*{\cdot}{\bigodot} \left( \mathcal{F} - \frac{1}{2} \eta \mathcal{F}' \right) \\ &= -\frac{s}{2} \int d^D x \, \Lambda \scalerel*{\cdot}{\bigodot} \partial \cdot \left( \mathcal{F} - \frac{1}{2} \eta \mathcal{F}' \right) \\ &= -\frac{s}{2} \int d^D x \, \Lambda \scalerel*{\cdot}{\bigodot} \left( \underbrace{\partial \cdot \mathcal{F} - \frac{1}{2} \partial \mathcal{F}'}_{\eqref{bianchi}} - \frac{1}{2} \eta \partial \cdot \mathcal{F}' \right) \\ &= -\frac{s}{2} \int d^D x \, \left[ \Lambda \scalerel*{\cdot}{\bigodot} \left( -\frac{3}{2} \partial^3 \varphi'' \right) - \frac{1}{2} \Lambda \scalerel*{\cdot}{\bigodot} \eta \partial \cdot \mathcal{F}' \right] \\ &= -3 \int d^D x \, \left[ {s \choose 4} \partial^3 \cdot \Lambda \scalerel*{\cdot}{\bigodot} \varphi'' - \frac{1}{4} {s \choose 3} \Lambda' \scalerel*{\cdot}{\bigodot} \partial \cdot \mathcal{F}' \right] \, . \end{align} Once again, the second term under the integral vanishes if we impose $\Lambda' = 0$. It follows that it is necessary to impose the additional constraint $\varphi'' = 0$ for the action to be gauge-invariant. An alternative way to get around the double-tracelessness constraint is to work within a geometric framework, where we generalize the Fronsdal tensor into an equivalent object satisfying generalized Bianchi identities. As previously mentioned, the price to pay for the elegant geometric theory is higher-order terms or non-locality. \paragraph{Counting degrees of freedom$\newline$} Let us show that the constrained Fronsdal equation propagates the correct number of degrees of freedom. In case of a massless spin-$s$ bosonic field, arguments from representation theory (as discussed in \textbf{Section \ref{fpe}} and specifically in relation to the Fronsdal equation in \cite{verybasics}) show that the correct number is \begin{equation} \#(D-2,s) - \#(D-2, s-2) = {{D+s-3} \choose {s}} - {{D+s-5} \choose {s-2}} \, . \label{dof} \end{equation} We begin by counting the number of independent components of $\varphi$. It is a fully symmetric doubly-traceless $D-$dimensional tensor of order $s$, so that number is \begin{equation} \#(D,s) - \#(D,s-4) = {{D+s-1} \choose {s}} - {{D+s-5} \choose {s-4}} \, . \end{equation} We proceed by partially fixing the gauge, imposing the de Donder gauge condition, \begin{equation} \mathcal{D} = \partial \cdot \varphi - \frac{1}{2} \partial \varphi' = 0 \, , \end{equation} which reduces the Fronsdal equation to a wave equation, \begin{equation} \Box \varphi = 0 \, . \end{equation} Since $\mathcal{D}$ is traceless and of order $s-1$, fixing the de Donder tensor corresponds to eliminating \begin{equation} \#(D,s-1) - \#(D,s-3) = {{D+s-2} \choose {s-1}} - {{D+s-4} \choose {s-3}} \end{equation} independent components. However, fixing $\mathcal{D}$ does not fully fix the gauge, since \begin{equation} \delta \mathcal{D} = \Box \Lambda \, . \end{equation} Fixing this residual gauge freedom also corresponds to eliminating $\#(D,s-1) - \#(D,s-3)$ independent components. In total, this leaves us with \begin{align} &\#(D,s) - \#(D,s-4) - 2\left\{\#(D,s-1) - \#(D,s-3)\right\} = \nonumber \\ &\#(D-2,s) - \#(D-2, s-2) \, , \end{align} which is the same as \eqref{dof}. Note that we used the double tracelessness of $\varphi$ to count the propagating degrees of freedom. However, this is merely a \textit{sufficient} condition for the correct number, not a \textit{necessary} one. \subsubsection{Interacting theory with an external current}\label{interacting constrained} Let us begin the analysis of the HS gauge field coupled to an external current within the framework of Fronsdal's constrained theory by defining the \textit{Fronsdal-Einstein} tensor $\mathcal{G}$, \begin{equation} \mathcal{G} := \mathcal{F} - \frac{1}{2} \eta \mathcal{F}' \, . \end{equation} We showed in \textbf{Section \ref{freefronsdal}} that this is precisely the left-hand side of the equation of motion, as obtained from \eqref{lagrangianF}. An interaction term in the Lagrangian, for some \textit{generic} totally symmetric external current $J$ can be written as \begin{equation} \mathcal{L}_{int} = -\frac{1}{2} \varphi \scalerel*{\cdot}{\bigodot} J \, , \label{current} \end{equation} so the total action reads \begin{equation} \mathcal{S} [\varphi, J] = \int d^D x \, \mathcal{L} = \int d^D x \, \left( \mathcal{L}_{\mathcal{F}} + \mathcal{L}_{int} \right) = \mathcal{S}_\mathcal{F} - \frac{1}{2} \int d^D x \, \varphi \scalerel*{\cdot}{\bigodot} J \, . \label{constrainedS} \end{equation} As demonstrated in \textbf{Section \ref{freefronsdal}}, \begin{equation} \delta_{\Lambda} \mathcal{S}_{\mathcal{F}} = 0 \, , \end{equation} so we need to investigate the effect of the interaction term $\mathcal{L}_{int}$, since it need not be gauge-invariant. The equation of motion obtained by varying \eqref{constrainedS} reads \begin{equation} \mathcal{G} = J \, . \label{el} \end{equation} Taking the trace of \eqref{el} yields \begin{equation} \mathcal{F}' = \frac{-2}{D+2(s-3)} J' \, , \label{fj} \end{equation} which in turn implies \begin{equation} J'' = 0 \, , \label{jpp0} \end{equation} since $\mathcal{F}''=0$ when $\varphi''=0$. We can now rewrite \eqref{el} as \begin{equation} \mathcal{F} = J - \frac{1}{D+2(s-3)} \eta J' \, . \label{fronsINT} \end{equation} Taking the divergence of \eqref{el} and using \eqref{bianchi}, we get \begin{align} \partial \cdot J = -\frac{1}{2} \eta \partial \cdot \mathcal{F}' \label{Whencefore It Cometh?} \end{align} Substituting \eqref{fj} into \eqref{Whencefore It Cometh?} yields \begin{equation} \partial \cdot J - \frac{1}{D+2(s-3)} \eta \partial \cdot \ J' = 0 \, . \label{divtracelessJ} \end{equation} The left-hand side of \eqref{divtracelessJ} is actually the traceless part of $\partial \cdot J$. In general, the traceless part of a fully symmetric tensor $\chi$ of order $s$ in $D$-dimensional spactime is\footnotemark\footnotetext{To compactify the notation, here we begin to use the \textbf{falling factorial} function, defined as \newline $n^{\underline{k}} = \frac{n!}{(n-k)!}$, and we define the \textbf{falling double factorial} function $n^{\uuline{k}} = \frac{n!!}{(n-k)!!}$.} \begin{equation} \mathcal{T}_D[\chi] = \sum_{k=0}^{[s/2]} \frac{(-1)^k}{[D+2(s-2)]^{\uuline{k}}} \eta^k \chi^{[k]} := \sum_{k=0}^{[s/2]} \rho_k (D,s) \eta^k \chi^{[k]} \, , \label{traceless} \end{equation} which is easily checked by direct computation. In \eqref{traceless}, we define coefficients $\rho_k (D,s)$ for later convenience. Using \eqref{traceless} and \eqref{jpp0}, we see that indeed \begin{equation} \mathcal{T}_D[\partial \cdot J] = \partial \cdot J - \frac{1}{D+2(s-3)} \eta \partial \cdot \ J' \, . \label{jtraceless} \end{equation} Also, \begin{equation} \mathcal{F} = \mathcal{T}_{D-2} [J] \, . \end{equation} Therefore, in general, only the traceless part of the divergence of $J$ vanishes. \bigskip\par\centerline{*\,*\,*}\medskip\par To understand the physical meaning of \eqref{jtraceless}, we need to introduce the concept of \textit{current exchange}. Let us motivate the idea on a familiar case of spin-$1$ fields, i.e. Maxwell's theory of electrodynamics. In the manifestly Lorentz-covariant formalism, Maxwell's equations coupled to an external current $J^\mu$ read \begin{equation} \Box A^\mu - \partial^\mu (\partial \cdot A) = J^\mu \, , \end{equation} where consistency demands that the current be conserved, i.e. \begin{equation} \partial_\mu j^\mu = 0 \, . \end{equation} In the momentum space, this translates to \begin{align} (p^2 \eta_{\mu\nu} - p_\mu p_\nu) A^\nu &= J_\mu \, , \\ p^\mu J_\mu &= 0 \, . \end{align} It follows that, for a current-current interaction, \begin{equation} p^2 A_\mu J^\mu = J_\mu J^\mu \, . \end{equation} By current exchange, we mean the exchange between the degrees of freedom that take part in this interaction. As we know from electrodynamics, interactions mediated by photons only respond to the transverse part of the current, since the photon has no longitudinal degrees of freedom. Therefore, instead of considering the full Lorentzian product $J_\mu J^\mu$, we can project an on-shell current (i.e. a current satisfying the equation of motion) $J_\mu (p)$ to its transverse part using the projection operator $\Pi$ \begin{equation} \Pi_{\mu\nu} = \eta_{\mu\nu} - p_\mu \bar{p}_\nu - p_\nu \bar{p}_\mu \, , \label{projektor} \end{equation} where $p$ is the exchanged on-shell momentum, satisfying $p^2 = 0$, and $\bar{p}$ is a vector that satisfies $\bar{p}^2 = 0$ and $p_\mu \bar{p}^\mu = 1$. One can check by direct computation that, indeed, \begin{equation} p^\mu \Pi_{\mu\nu} J^\nu = 0 \, , \end{equation} and \begin{equation} J_\mu J^\mu = J^\mu \Pi_{\mu\nu} J^\nu \, . \label{jpj} \end{equation} Now, since \begin{equation} \eta_{\mu \nu} \Pi^{\mu\nu} = D - 2 \, , \end{equation} it follows\footnotemark\footnotetext{Equation \eqref{jpj} is basically an eigenvalue problem. The trace of a linear operator equals the sum of its eigenvalues. Since the eigenvalues of a projection operator equal $0$ or $1$, its trace equals the dimension of the subspace to which it projects.} that the number of degrees of freedom taking part in the interaction is $D-2$. Instead of working in the manifestly Lorentz-covariant formulation, we can repeat the procedure in the light-cone formulation, where we have two null coordinates (i.e. coordinates on the light cone), \begin{align} x^+ &= \frac{t+x}{\sqrt{2}} \, , \\ x^- &= \frac{t-x}{\sqrt{2}} \, , \end{align} and the remaining $D-2$ coordinates are spatial. This allows us to work in the light-cone gauge, \begin{equation} A^+ = 0 \, , \end{equation} which eliminates all unphysical degrees of freedom. In this formulation, Maxwell's equations in the momentum space take a simple form that involves only the spatial coordinates, \begin{equation} p^2 A_i = j_i \, , \end{equation} where latin indices denote the components of ($D-2$)-dimensional Euclidean vectors. The current-current interaction becomes \begin{equation} p^2 j_i A^i = j_i j^i \, . \end{equation} Since all components are physical\footnotemark\footnotetext{Because the light-cone formulation corresponds to working in the "reference frame" of a massless particle, where all degrees of freedom are particle's proper degrees of freedom. This is why the light-cone frame is sometimes referred to as the \textit{infinite momentum frame.}}, we simply count the number of components of $j_i$, which is $D-2$, in agreement with our previous conclusion. The general idea is to check whether \begin{equation} J_{\mu_1 \cdots \mu_s} \mathcal{P}^{\mu_1 \cdots \mu_s \nu_1 \cdots \nu_s} J_{\nu_1 \cdots \nu_s} = j_{a_1 \cdots a_s} j^{a_1 \cdots a_s} \label{currx} \end{equation} holds for spin-$s$ current exchanges, where $\mathcal{P}$ denotes the proper analogue of the projection operator \eqref{projektor}. The right-hand side of \eqref{currx} implies that the proper number of degrees of freedom in the current exchange is equal to the number of independent components of the current in the light-cone gauge. As we saw in \textbf{Section \ref{fpe}}, this number is equal to the number of independent components of a traceless fully symmetric tensor of order $s$ in $D-2$ dimensions. This means that $\mathcal{P}$ should be an operator that projects the current to its transverse part and then extracts its traceless part. Using \eqref{traceless} and \eqref{projektor}, we see that $\mathcal{P}$ has to be \begin{equation} \mathcal{P}^{(\mu)(\nu)} J_{(\nu)} = \mathcal{T}_{D-2} [\Pi \cdot J] \, , \label{conserved} \end{equation} where we write $(\mu)$ and $(\nu)$ to indicate a totally symmetric set of $s$ indices. \bigskip\par\centerline{*\,*\,*}\medskip\par Coming back to the interacting theory with an external current in the constrained formulation, we see that \eqref{jtraceless} determines the currrent exchange. It implies that \begin{align} J_{(\mu)} \mathcal{P}^{(\mu)(\nu)} J_{(\nu)} &= J \scalerel*{\cdot}{\bigodot} \left( J - \frac{1}{D+2(s-3)} \eta J'\right) \, , \\ &= J \scalerel*{\cdot}{\bigodot} J - \frac{1}{D+2(s-3)} J \scalerel*{\cdot}{\bigodot} \eta J' \, , \\ &= J \scalerel*{\cdot}{\bigodot} J - \frac{s(s-1)}{2[D+2(s-3)]} J' \scalerel*{\cdot}{\bigodot} J' \, \\ &= J \scalerel*{\cdot}{\bigodot} J + \rho_1 (D-2,s) {s \choose 2} J' \scalerel*{\cdot}{\bigodot} J' . \label{constrained exchange} \end{align} We will return to this result to compare it with the analogous result in the unconstrained formulation. \subsection{Local unconstrained theory}\label{localunconstrained} Let us demonstrate how we can rewrite Fronsdal's theory without the usual \begin{equation} \Lambda' = 0 \quad \& \quad \varphi'' = 0 \end{equation} constraints. This is accomplished here by introducing two compensator fields. \subsubsection{Free theory} We begin by considering the Fronsdal tensor $\mathcal{F}$ and its gauge transformation \eqref{FVar}. From $\mathcal{F}$, one can build a fully gauge-invariant tensor, \begin{equation} \mathcal{A} := \mathcal{F} - 3 \partial^3 \alpha \, , \label{Acomp} \end{equation} where we introduce the field $\alpha(x)$ as a spin-$(s-3)$ \textit{compensator}, which transforms as \begin{equation} \delta_\Lambda \alpha = \Lambda' \end{equation} under the gauge transformation \eqref{FVar}. The Bianchi identity for $\mathcal{A}$ reads \begin{equation} \partial \cdot \mathcal{A} - \frac{1}{2} \partial \mathcal{A}' = - \frac{3}{2} \partial^3 \left( \varphi'' - 4 \partial \cdot \alpha - \partial \alpha' \right) =: -\frac{3}{2} \partial^3 \mathcal{C} \, , \label{ABianchi} \end{equation} where we have identified a gauge-invariant tensor, which we denote by $\mathcal{C}$, i.e. \begin{equation} \mathcal{C} = \varphi'' - 4 \partial \cdot \alpha - \partial \alpha' \, . \end{equation} In analogy with \eqref{lagrangianF}, we write the Lagrangian \begin{equation} \mathcal{L}_0 = \frac{1}{2} \varphi \scalerel*{\cdot}{\bigodot} \left( \mathcal{A} - \frac{1}{2} \eta \mathcal{A}' \right) \, . \end{equation} Varying the action, we get \begin{align} \delta_{\Lambda} \mathcal{S}_0 = \int d^D x \, \delta_{\Lambda} \mathcal{L}_0 &= \frac{1}{2} \int d^D x \, \partial \Lambda \scalerel*{\cdot}{\bigodot} \left( \mathcal{A} - \frac{1}{2} \eta \mathcal{A}' \right) \\ &= - \frac{s}{2} \int d^D x \, \Lambda \scalerel*{\cdot}{\bigodot} \left( \underbrace{\partial \cdot \mathcal{A} - \frac{1}{2} \partial \mathcal{A}'}_{\eqref{ABianchi}} - \frac{1}{2} \eta \partial \cdot \mathcal{A}' \right) \\ &= -3 \int d^D x \, \left[ {s \choose 4} \partial^3 \cdot \Lambda \scalerel*{\cdot}{\bigodot} \mathcal{C} - \frac{1}{4} {s \choose 3} \Lambda' \scalerel*{\cdot}{\bigodot} \partial \cdot \mathcal{A}' \right] \, . \end{align} We can make all the terms under the integral vanish by adding to $\mathcal{L}_0$ \begin{equation} \mathcal{L}_1 = -\frac{3}{4} {s \choose 3} \alpha \scalerel*{\cdot}{\bigodot} \partial \cdot \mathcal{A}' + 3 {s \choose 4} \beta \scalerel*{\cdot}{\bigodot} \mathcal{C} \, , \end{equation} where we introduce the second compensator\footnotemark \footnotetext{Technically, it is just a Lagrange multiplier.}, a spin-$(s-4)$ field denoted by $\beta$ that transforms as \begin{equation} \delta_\Lambda \beta = \partial^3 \cdot \Lambda \end{equation} under the gauge transformation \eqref{FVar}. Finally, we can write the fully gauge-invariant Lagrangian for the unconstrained local theory as \begin{equation} \mathcal{L} = \frac{1}{2} \varphi \scalerel*{\cdot}{\bigodot} \left( \mathcal{A} - \frac{1}{2} \eta \mathcal{A}' \right) -\frac{3}{4} {s \choose 3} \alpha \scalerel*{\cdot}{\bigodot} \partial \cdot \mathcal{A}' + 3 {s \choose 4} \beta \scalerel*{\cdot}{\bigodot} \mathcal{C} \, . \label{abplagrangian} \end{equation} We can introduce the third gauge-invariant tensor $\mathcal{B}$, \begin{equation} \mathcal{B} := \beta + \Box \partial \cdot \alpha + \frac{1}{2} \partial (\partial^2 \cdot \alpha) - \frac{1}{2} \partial^2 \cdot \varphi' \, \label{Bcomp} \end{equation} and note that \eqref{abplagrangian} may be generalized to \begin{equation} \mathcal{L}_k = \frac{1}{2} \varphi \scalerel*{\cdot}{\bigodot} \left( \mathcal{A} - \frac{1}{2} \eta \mathcal{A}' \right) -\frac{3}{4} {s \choose 3} \alpha \scalerel*{\cdot}{\bigodot} \partial \cdot \mathcal{A}' + 3 {s \choose 4} \left( \beta - k \mathcal{B} \right) \scalerel*{\cdot}{\bigodot} \mathcal{C} \, , \label{abplagrangianK} \end{equation} without affecting the equations of motion, so that \eqref{abplagrangian} corresponds to $k=0$. As shown in \cite{fms}, a more general analysis reveals that adding quadratic terms in $\mathcal{A}$, $\mathcal{B}$ and $\mathcal{C}$ to \eqref{abplagrangianK} does not produce any terms that would lead to different equations of motion. Therefore, the free unconstrained local theory is parametrized by a real parameter $k$ and the gauge-invariant field equations read \begin{align} E_\varphi (k) &:= \mathcal{A} - \frac{1}{2} \eta \mathcal{A}' + \frac{1+k}{4} \eta \partial^2 \mathcal{C} + (1-k) \eta^2 \mathcal{B} = 0 \, , \label{Evarphi} \\ E_\alpha (k) &:= -\frac{3}{2} {s \choose 3} \left[ \partial \cdot \mathcal{A}' -\frac{1+k}{2} \left( \partial \Box + \partial^2 \partial \cdot \right) \mathcal{C} + (k-1) \left( 2 \partial + \eta \partial \cdot \right) \mathcal{B} \right] = 0 \, , \label{Ealpha} \\ E_\beta (k) &:= 3 {s \choose 4} (1-k) \mathcal{C} = 0 \, . \label{Ebeta} \end{align} We can use these three tensors to write the final Lagrangian in a particularly elegant form, \begin{equation} \mathcal{L}_k = \frac{1}{2} \varphi \scalerel*{\cdot}{\bigodot} E_\varphi (k) + \frac{1}{2} \alpha \scalerel*{\cdot}{\bigodot} E_\alpha (k) + \frac{1}{2} \beta \scalerel*{\cdot}{\bigodot} E_\beta (k) \, . \end{equation} From equations \eqref{Evarphi}-\eqref{Ebeta}, if $k\neq1$, it follows that \begin{align} \mathcal{A} &\equiv \mathcal{F} - 3 \partial^3 \alpha = 0 \, , \\ \mathcal{C} &\equiv \varphi'' - 4 \partial \cdot \alpha - \partial \alpha' = 0 \, . \end{align} After fixing the gauge to $\Lambda' = 0$ we are left with \begin{align} \mathcal{F} &= 0 \, , \\ \varphi'' &= 0 \, , \end{align} which is exactly equivalent to Fronsdal's constrained formulation. \subsubsection{Interacting theory with an external current} In the unconstrained formulation described in the previous subsection, setting $k=0$, coupling to an external source \eqref{current} is described by \begin{equation} \mathcal{A} - \frac{1}{2} \eta \mathcal{A}' + \eta^2 \mathcal{B} = J \label{abj} \end{equation} We can define a quantity $\mathcal{K}$, \begin{equation} \mathcal{K} := J - \eta^2 \mathcal{B} \, , \label{kdef} \end{equation} and write the equation of motion as \begin{equation} \mathcal{A} - \frac{1}{2} \eta \mathcal{A}' = \mathcal{K} \, , \label{unconseom} \end{equation} so that, formally, $\mathcal{A}$ and $\mathcal{K}$ play the same role as $\mathcal{F}$ and $J$ play in the constrained formalism of \textbf{Section \ref{interacting constrained}}. Note that \begin{equation} \mathcal{A}'' = 3 \Box \mathcal{C} + 3 \partial (\partial \cdot \mathcal{C}) + \partial^2 \mathcal{C}' = 0 \, , \label{app} \end{equation} since $\mathcal{C}$ vanishes as a result of \eqref{Ebeta}. Since $\mathcal{A}''$ vanishes when the equations of motion are satisfied, we can write \begin{equation} \mathcal{B} = \sum_{k=2}^{n+1} \sigma_k \eta^k J^{[k]} \, , \end{equation} where $n=\left[ \frac{s-1}{2} \right]$ and we can determine the coefficients $\sigma_k$ from the condition $\mathcal{K}'' = 0$. A direct computations yields \begin{equation} \mathcal{B} = \sum_{k=2}^{n+1} (1-n) \, \rho_n(D-2,s) \eta^k J^{[k]} \, , \label{bcons} \end{equation} which allows us to rewrite \eqref{abj} as \begin{equation} \mathcal{A} - \frac{1}{2} \eta \mathcal{A}' = J - \sum_{k=2}^{n+1} (1-n) \, \rho_n(D-2,s) \eta^k J^{[k]} \end{equation} We can use the formal correspondence between $\mathcal{A}$ and $\mathcal{K}$, and $\mathcal{F}$ and $J$, to skip the explicit calculation and immediately write \begin{equation} \mathcal{A} = \mathcal{K} - \frac{1}{D + 2(s-3)} \eta \mathcal{K}' \, , \label{acons} \end{equation} in analogy with \eqref{fronsINT}. Using \eqref{bcons} and \eqref{kdef}, we arrive at \begin{equation} \mathcal{A} = \sum_{k=0}^{n+1} \rho_k (D-2,s) \eta^k J^{[k]} \end{equation} The current exchange is thus \begin{align} J_{(\mu)} \mathcal{P}^{(\mu)(\nu)} J_{(\nu)} &= \sum_{k=0}^{n+1} \rho_k (D-2,s) J \scalerel*{\cdot}{\bigodot} \eta^k J^{[k]} \\ &= \sum_{k=0}^{n+1} \rho_k (D-2,s) {{s-2k} \choose k} J^{[k]} \scalerel*{\cdot}{\bigodot} J^{[k]} \, , \label{unconstrained exchange} \end{align} which agrees with \eqref{constrained exchange}, as we can see by expanding the first two terms, \begin{align} J_{(\mu)} \mathcal{P}^{(\mu)(\nu)} J_{(\nu)} & = J \scalerel*{\cdot}{\bigodot} J + \rho_1 (D-2,s) {s \choose 2} J' \scalerel*{\cdot}{\bigodot} J' \\ \nonumber &\quad + \sum_{k=2}^{n+1} \rho_k (D-2,s) J \scalerel*{\cdot}{\bigodot} \eta^k J^{[k]} \, . \end{align} \newpage \subsection{Non-local unconstrained theory}\label{nonlocal} Instead of introducing compensator fields $\alpha$ and $\beta$ and formulating the theory in terms of $\mathcal{A}$, $\mathcal{C}$ and $\mathcal{B}$ tensors, we can construct it using only the gauge field $\varphi$ if we allow non-local operators, i.e. powers of $\frac{1}{\Box}$ \footnotemark.\footnotetext{Alternatively, we could multiply the equations with the appropriate power of $\Box$ and have a higher-order derivative theory instead. However, it is not clear if the higher-order formulation of the theory is equivalent to the non-local formulation.} Let us show here how to construct the theory in this manner. \subsubsection{Free theory} One begins by building a non-local tensor $\mathcal{H}$ that satisfies \begin{equation} \delta_{\Lambda} \mathcal{H} = 3 \Lambda' \, , \end{equation} so that $\mathcal{F} - \partial^3 \mathcal{H}$ becomes gauge-invariant without any additional constraints or compensator fields. As shown in \cite{fs1}, inspired by HS generalizations of metric connections from general relativity (developed in \cite{dwf} and later explained in more detail in \textbf{Section \ref{geometric}}), we can try to construct a generalized Fronsdal tensor $\mathcal{F}_n$ that transforms as \begin{equation} \delta_\Lambda \mathcal{F}_{n} = (2n+1) \frac{\partial^{2n+1}}{\Box^{n-1}} \Lambda^{[n]} \, \label{gaugenonlocal} \end{equation} under the gauge variation \eqref{gauge}. This way, for high enough $n$, $\mathcal{F}_n$ becomes gauge-invariant. Since we also want the action to be gauge invariant, we require that $\mathcal{F}_n$ satisfies a generalization of the Bianchi identity, \begin{equation} \partial \cdot \mathcal{F}_{n} - \frac{1}{2n} \partial \mathcal{F}_{n}{}' = - \left( 1 + \frac{1}{2n} \right) \frac{\partial^{2n+1}}{\Box^{n-1}} \varphi^{[n+1]} \, , \label{bianchinonlocal} \end{equation} which also vanishes for high enough $n$. The generalized Fronsdal tensor $\mathcal{F}_n$ that satisfies all these requirements reads \begin{equation} \mathcal{F}_{n+1} = \mathcal{F}_{n} - \frac{1}{n+1} \frac{\partial}{\Box} \left( \partial \cdot \mathcal{F}_{n} \right) + \frac{1}{(n+1)(2n+1)} \frac{\partial^2}{\Box} \mathcal{F}_{n}{}' \, , \end{equation} where $\mathcal{F}_{1} = \mathcal{F}$ (or equivalently, $\mathcal{F}_0 = \Box \varphi$), as one can easily check through direct computation and using simple inductive arguments. To construct a spin-$s$ theory, we use $\mathcal{F}_{n+1}$ with $n=\left[ \frac{s-1}{2} \right]$, the minimal value for which the gauge variation and the Bianchi identity \eqref{bianchinonlocal} both vanish. The corresponding generalized Einstein-like tensor reads \begin{equation} \mathcal{G}_{n} = \sum_{k=0}^{n+1} \frac{(-1)^k}{2^k (n+1)^{\underline{k}}} \eta^k \mathcal{F}_{n+1}^{[k]} \, . \end{equation} One can check that $\mathcal{G}_{n}$ is indeed divergenceless, as required by the gauge-invariance of the action, using the traces of \eqref{bianchinonlocal}, which satisfy \begin{equation} \partial \cdot \mathcal{F}_{n+1}^{[k]} - \frac{1}{2(n-k+1)} \partial \mathcal{F}_{n+1}^{[k+1]} = 0 \, , \quad (k \leq n) \label{pbianchinonlocal} \end{equation} and applying it successively to terms in $\partial \cdot \mathcal{G}_{n}$. For clarity, let us show how all the pieces fit together to make the action gauge-invariant. \begin{align} \delta_\Lambda \mathcal{S}_n &= \int d^D x \, \delta_\Lambda \mathcal{L}_n \\ &= \frac{1}{2} \int d^D x \, \left( \partial \Lambda \scalerel*{\cdot}{\bigodot} \mathcal{G}_n + \varphi \scalerel*{\cdot}{\bigodot} \delta_\Lambda \mathcal{G}_n \right) \\ &= \frac{1}{2} \sum_{k=0}^{n+1} \frac{(-1)^k}{2^k (n+1)^{\underline{k}}} \int d^D x \, \left( \partial \Lambda \scalerel*{\cdot}{\bigodot} \eta^k \mathcal{F}_{n+1}^{[k]} + \varphi \scalerel*{\cdot}{\bigodot} \eta^k \underbrace{\delta_\Lambda \mathcal{F}_{n+1}^{[k]}}_{=0, \eqref{gaugenonlocal}} \right) \\ &= -\frac{s}{2} \sum_{k=0}^{n+1} \frac{(-1)^k}{2^k (n+1)^{\underline{k}}} \int d^D x \, \left[ \Lambda \scalerel*{\cdot}{\bigodot} \partial \cdot \left( \eta^k \mathcal{F}_{n+1}^{[k]} \right) \right] \\ &= -\frac{s}{2} \int d^D x \, \Lambda \scalerel*{\cdot}{\bigodot} \partial \cdot \mathcal{F}_{n+1} -\frac{s}{2} \sum_{k=1}^{n+1} \frac{(-1)^k}{2^k (n+1)^{\underline{k}}} \int d^D x \, \Lambda \scalerel*{\cdot}{\bigodot} \eta^k \partial \cdot \mathcal{F}_{n+1}^{[k]} \\ \nonumber &\quad+s \sum_{k=0}^{n} \frac{(-1)^k}{2^k (n+1)^{\underline{k+1}}} \int d^D x \, \Lambda \scalerel*{\cdot}{\bigodot} \eta^{k} \partial \mathcal{F}_{n+1}^{[k+1]} \\ &= -\frac{s}{2} \sum_{k=0}^{n} \frac{(-1)^k}{2^k (n+1)^{\underline{k}}} \int d^D x \, \eta^k \left( \underbrace{ \partial \cdot \mathcal{F}_{n+1}^{[k]} - \frac{1}{2(n-k+1)} \partial \mathcal{F}_{n+1}^{[k+1]} }_{=0, \eqref{pbianchinonlocal}} \right) \\ &= 0 \end{align} \subsubsection{Interacting theory with an external current} If the system is coupled to a generic totally symmetric external current $\mathcal{J}$, a natural starting point would be to write the Lagrangian as \begin{equation} \mathcal{L} = \frac{1}{2} \varphi \scalerel*{\cdot}{\bigodot} \left( \mathcal{G}_{n} - \mathcal{J} \right) \end{equation} and the field equations read \begin{equation} \mathcal{G}_{n} \equiv \sum_{k=0}^{n+1} \frac{(-1)^k}{2^k (n+1)^{\underline{k}}} \eta^k \mathcal{F}_{n+1}^{[k]} = \mathcal{J} \, . \label{nonlocal current} \end{equation} We proceed like in all previous cases, inverting \eqref{nonlocal current} to extract the current exchange. Taking successive traces of \eqref{nonlocal current} and multiplying both sides with metric tensors to obtain a tensor of order $s$, one finds a general relation, \begin{equation} \rho_k (D-2n,s-1) \eta^k \mathcal{J}^{[k]} = (-1)^k \sum_{p=k}^{n+1} \frac{(-1)^{p}}{2^{p} (n+1)^{\underline{p}}} {p \choose k} \eta^p \mathcal{F}_{n+1}^{[p]} \, . \label{JFrelation} \end{equation} Summing both sides of \eqref{JFrelation} over $k$ ($0 \leq k \leq n+1$), one finds that the factor $(-1)^k {p \choose k}$ cancels all the terms over $p$ on the right-hand side, except $\mathcal{F}_{n+1}$, i.e. \begin{equation} \mathcal{F}_{n+1} = \sum_{k=0}^{n+1} \rho_k (D-2n,s-1) \eta^k \mathcal{J}^{[k]} \, . \end{equation} Therefore, the current exchange is described by \begin{align} \mathcal{J}_{(\mu)} \mathcal{P}^{(\mu)(\nu)} \mathcal{J}_{(\nu)} &= \sum_{k=0}^{n+1} \rho_k (D-2n,s-1) \mathcal{J} \scalerel*{\cdot}{\bigodot} \eta^k \mathcal{J}^{[k]} \\ &= \sum_{k=0}^{n+1} \rho_k (D-2n,s-1) {{s-2k} \choose 2} \mathcal{J}^{[k]} \scalerel*{\cdot}{\bigodot} \mathcal{J}^{[k]} \label{thexchangesum} \end{align} Expanding the first two terms of the current exchange, we see that \begin{align} \mathcal{J}_{(\mu)} \mathcal{P}^{(\mu)(\nu)} \mathcal{J}_{(\nu)} &= \mathcal{J} \scalerel*{\cdot}{\bigodot} \mathcal{J} + \frac{1}{2} \rho_1 (D-2n,s-1) \mathcal{J}' \scalerel*{\cdot}{\bigodot} \mathcal{J}'\\ \nonumber &\quad + \sum_{k=2}^{n+1} \rho_k (D-2n,s-1) {{s-2k} \choose 2} \mathcal{J}^{[k]} \scalerel*{\cdot}{\bigodot} \mathcal{J}^{[k]} \, , \end{align} which, due to the presence of an additional $-2n$ in the denominator, clearly \textit{disagrees} with the constrained case \eqref{constrained exchange} and the unconstrained case \eqref{unconstrained exchange}, except in the case of lower spins, i.e. $s\leq2$. We explore the implications of this disagreement in the following segment. \subsubsection{HS theory with proper current exchange} So far, we have seen three different ways to formulate a higher spin theory of massless bosons in flat spacetime. One is \textit{Fronsdal's constrained formulation}, explored in \textbf{Section \ref{freefronsdal}}, in which the gauge-invariance of the action is enforced by restricting the gauge parameter $\Lambda$ to a traceless tensor and restricting the gauge field $\varphi$ to a doubly-traceless tensor. The second one is the \textit{local unconstrained formulation}, explored in \textbf{Section \ref{localunconstrained}}, which requires additional non-dynamical fields $\alpha$ and $\beta$ to ensure a fully gauge-invariant action. The third one is the \textit{non-local unconstrained formulation}, explored in this section, which allows for a fully gauge-invariant theory without any additional fields, at the cost of having to use non-local operators $\frac{1}{\Box}$. As we concluded in the previous segment, current exchanges in the non-local unconstrained theory seem to disagree with other two formulations. Since the non-local formulation is based on simple geometric arguments, without imposing \textit{ad-hoc} constraints or adding additional fields to the theory, it is natural to take it as a starting point, try to understand the disagreement and try to formulate it in a way that naturally reduces to other two formulations. As we shall see, this leads us to a \textit{unique} form of the theory for each spin. Equation \eqref{unconstrained exchange} suggests that, in the constrained formulation, the operator $\mathcal{P}_c$, as defined in \eqref{currx}, is \begin{equation} \mathcal{P}_c \scalerel*{\cdot}{\bigodot} J= \sum_{k=0}^{n+1} \rho_k (D-2,s) \eta^k J^{[k]} \, . \end{equation} A direct computation shows that \begin{equation} (\mathcal{P}_c \scalerel*{\cdot}{\bigodot} J)' = 2 \sum_{k=0}^{n+1} \rho_{k+1} (D-2,s) \eta^k J^{[k]} \, , \end{equation} and \begin{equation} (\mathcal{P}_c \scalerel*{\cdot}{\bigodot} J)'' = 0 \, . \label{p''} \end{equation} Note that $\mathcal{P}_c$ precisely corresponds to \eqref{currx} if the current is conserved, since $\Pi$ effectively gets replaced by $\eta$. Thus, if we want to build an unconstrained theory with proper current exchanges, in analogy with \eqref{abj}, we postulate the non-local Einstein tensor $\mathcal{E}$ of form\footnotemark\footnotetext{We put $\varphi$ in the subscript to stress the fact that these quantities are to be built using only the gauge field $\varphi$.} \begin{equation} \mathcal{E} = \mathcal{A}_\varphi - \frac{1}{2} \eta \mathcal{A}_\varphi ' + \eta^2 \mathcal{B}_\varphi \, , \end{equation} requiring that $\mathcal{A}_\varphi'' = 0$, reflecting \eqref{p''}, and $\partial \cdot \mathcal{E} = 0$, reflecting the fact that $\mathcal{P}_c$ corresponds to the generalized projection operator for a \textit{conserved} external current. We construct $\mathcal{A}_\varphi$ using gauge-invariant building blocks, which can all be expressed in terms of $\mathcal{F}_{n+1}$\footnotemark\footnotetext{Alternatively, we could have used $\mathcal{G}_n$ as the main building block.}, where $n=\left[ \frac{s-1}{2} \right]$. Since \begin{equation} \partial \cdot \mathcal{F}^{[k]}_{n+1} = \frac{1}{2(n-k+1)} \partial \mathcal{F}^{[k+1]}_{n+1} \, , \end{equation} all divergences can be expressed in terms of traces, which means that our gauge-invariant building blocks are \begin{equation} \mathcal{F}_{n+1}, \, \mathcal{F}_{n+1}' , \, \dots \, , \mathcal{F}^{[N]}_{n+1} \end{equation} where $N=\left[ \frac{s}{2} \right]$. The general linear combination thus reads \begin{equation} \mathcal{A}_\varphi = \sum_{k=0}^{N} a_k \frac{\partial^{2k}}{\Box^k} \mathcal{F}^{[k]}_{n+1} \, . \end{equation} Demanding that $\mathcal{A}_\varphi$ satisfies the Bianchi identity (because we want to write the Lagrangian as $\mathcal{L}_\varphi = \frac{1}{2} \varphi \scalerel*{\cdot}{\bigodot} \mathcal{E}$), \begin{equation} \partial \cdot \mathcal{A}_\varphi - \frac{1}{2} \partial \mathcal{A}_\varphi ' = 0 \, \label{bianchnl} \end{equation} implies \begin{equation} a_k = (-1)^{k+1} (2k-1) \prod_{j=0}^{k-1} \frac{n+j}{n-j+1} \, . \end{equation} We can write $\mathcal{B}_\varphi$ as \begin{equation} \mathcal{B}_\varphi = \sum_{k=0}^{N-2} \eta^k \mathcal{B}_k \, , \end{equation} where $\mathcal{B}_k$ terms contain no metric tensors. We solve for $\mathcal{B}_k$ by demanding that $\mathcal{E}$ be divergenceless, i.e. \begin{align} \partial \cdot & \left\{ \mathcal{A}_\varphi - \frac{1}{2} \eta \mathcal{A}_\varphi ' + \eta^2 \mathcal{B}_\varphi \right\} = 0 \, \\ \implies & \partial \cdot \mathcal{A}_\varphi ' = 2 \partial \mathcal{B}_\varphi + \eta \partial \cdot \mathcal{B}_\varphi \, . \end{align} This in turn implies that $\mathcal{B}_0$ is pure gradient \begin{equation} \partial \cdot \mathcal{A}_\varphi' = 2 \partial \mathcal{B}_0 \, , \label{AB} \end{equation} and $\mathcal{B}_k$ tensors can therefore be expressed as traces of $\mathcal{B}_0$, i.e. \begin{equation} \mathcal{B}_\varphi = \sum_{k=0}^{N-2} \frac{1}{2^{k-1} (k+2)!} \eta^k \mathcal{B}^{[k]}_0 \, . \end{equation} Solving \eqref{AB} for $\mathcal{B}_0$ gives \begin{equation} \mathcal{B}_0 = \sum_{k=0}^{N} b_k \frac{\partial^{2k}}{\Box^k} \mathcal{F}^{[k+2]}_{n+1} \, , \end{equation} where\footnotemark \footnotetext{This calculation was worked out with an error in \cite{fms} and later corrected in \cite{dariomass}, which unfortunately also seems to contain an error.} \begin{equation} b_k = \frac{a_k}{4(n-k)(n-k+1)} \frac{1-4n^2}{1-4k^2} \, . \end{equation} Note that the denominator of $b_k$ can be equal to zero, but that is not a problem, since $\mathcal{F}^{[k+2]}_{n+1}$ vanishes for those values. \subsection{Geometric theory}\label{geometric} It is instructive to formulate the theory purely in terms of generalized geometric objects. We start by defining \textit{higher-spin curvatures}, which leads us to the construction of generalized Riemann and Einstein tensors. \subsubsection{Higher-spin curvature} As explained in \cite{dwf}, HS curvatures are essentially a hierarchy of generalized (linearized!) \textit{Christoffel connections} $\Gamma$ built from derivatives of the gauge field $\varphi$. The $m$-th connection reads \begin{equation} \Gamma^{(m)}_{\mu_1 \cdots \mu_m ; \nu_1 \cdots \nu_s} \equiv \Gamma^{(m)} = \sum_{k=0}^{m} \frac{(-1)^k}{{m \choose k}} \partial_{(\nu)}^{m-k} \partial_{(\mu)}^k \varphi \, , \label{gamma} \end{equation} where we write $(\mu)$ and $(\nu)$ in the subscript of $\partial$ to denote that $m-k$ derivatives carry one set of symmetric indices $(\nu_1 \cdots \nu_s)$, whereas $k$ derivatives carry the other set of symmetric indices $(\mu_1 \cdots \mu_m)$. For example, if $s=2$ the first connection is \begin{align} \Gamma^{(1)} &= \partial_{(\nu)} \varphi - \partial_{(\mu)} \varphi \\ & \equiv \partial_{\nu_1} \varphi_{\nu_2 \mu} + \partial_{\nu_2} \varphi_{\nu_1 \mu} - \partial_\mu \varphi_{\nu_1 \nu_2} \end{align} which is exactly the linearized first connection as we know it from linearized general relativity, up to a multiplicative constant. The gauge variation of $\Gamma^{(m)}$ is \begin{align} \delta \Gamma^{(m)} &= \sum_{k=0}^{m} \frac{(-1)^k}{{m \choose k}} \partial_{(\nu)}^{m-k} \partial_{(\mu)}^k (\partial_{(\nu)} \Lambda + \partial_{(\mu)} \Lambda) \\ &= \sum_{k=0}^{m} \frac{(-1)^k}{{m \choose k}} \left[ (m-k+1) \partial_{(\nu)}^{m-k+1} \partial_{(\mu)}^k \Lambda + (k+1) \partial_{(\nu)}^{m-k} \partial_{(\mu)}^{k+1} \Lambda \right] \\ & = (m+1) \partial_{(\nu)}^{m+1} \Lambda \,. \end{align} Since $\partial$ should carry $m+1$ $\nu$-indices, $\Lambda$ should carry $s-1$ $\mu$-indices and there are $m+s$ indices in total, the gauge variation vanishes for $m \geq s$. For this reason, we define the generalized Riemann tensor as \begin{align} \mathcal{R} := \Gamma^{(s)} \, . \end{align} \subsubsection{Relationship between $\mathcal{R}$ and $\mathcal{F}$} Following \cite{fms}, we can relate the Fronsdal tensor $\mathcal{F}$ to the generalized Riemann tensor $\mathcal{R}$ using the generalized Fronsdal tensor $\mathcal{F}_n$, \begin{equation} \mathcal{F}_{n+1} = \frac{1}{\Box^n} \partial^{s - 2N} \cdot \mathcal{R}^{[N]} \, , \label{geometry} \end{equation} where, as before, $n=\left[ \frac{s-1}{2} \right]$, $N=\left[ \frac{s}{2} \right]$ and the contraction is performed on $\mu$-indices, i.e. the first set of indices in the $s$-th connection. The correspondence \eqref{geometry} thus allows us to reformulate the theory using only (linearized!) geometric objects, which was one of the motivating factors that led to the construction of $\mathcal{F}_{n+1}$, as mentioned in \textbf{Section \ref{nonlocal}} \newpage \section{Discussion}\label{s7} Having gone through the analysis of massless bosonic massless fields, we are now in a position to concisely define the proper theory and show how it reduces to some interesting special cases. The full set of equations reads \begin{empheq}[box=\widefbox]{align} \mathcal{L} &= \frac{1}{2} \varphi \left( \mathcal{E} - \mathcal{J} \right) \\ \mathcal{E} &= \mathcal{A}_\varphi - \frac{1}{2} \eta \mathcal{A}_\varphi ' + \eta^2 \mathcal{B}_\varphi\\ \mathcal{A}_\varphi &= \sum_{k=0}^{N} a_k \frac{\partial^{2k}}{\Box^k} \mathcal{F}^{[k]}_{n+1} \\ \mathcal{B}_\varphi &= \sum_{k=0}^{N-2} \frac{1}{2^{k-1} (k+2)!} \eta^k \mathcal{B}_0^{[k]} \label{B} \\ \mathcal{B}_0 &= \sum_{k=0}^{N} b_k \frac{\partial^{2k}}{\Box^k} \mathcal{F}^{[k+2]}_{n+1} \label{D} \\ a_k &= (-1)^{k+1} (2k-1) \prod_{j=0}^{k-1} \frac{n+j}{n-j+1} \label{ak} \\ b_k &= \frac{a_k}{4(n-k)(n-k+1)} \frac{1-4n^2}{1-4k^2} \label{bk} \\ \mathcal{F}_{n+1} &= \mathcal{F}_n - \frac{1}{n+1} \frac{\partial}{\Box} \partial \cdot \mathcal{F}_n + \frac{1}{(n+1)(2n+1)} \frac{\partial^2}{\Box} \mathcal{F}_n' \\ \mathcal{F}_0 &= \Box \varphi \\ n &= \left[ \frac{s-1}{2} \right], \quad N = \left[ \frac{s}{2} \right] \end{empheq} If the theory is free, $\mathcal{E} = 0$ reduces to $\mathcal{A}_\varphi = 0$ which in turn reduces to $\mathcal{F}_{n+1} = 0$. If we want to cast the theory in a local unconstrained form, we express $\mathcal{A}$ in terms of $\varphi$ and $\alpha$ as described in \eqref{Acomp}, and we express $\mathcal{B}$ in terms of $\varphi$ and $\beta$ as described in \eqref{Bcomp}. Finally, to go full circle and arrive at Fronsdal's constrained theory, we simply dispense with compensator fields $\alpha$ and $\beta$ and we impose $\Lambda'=0$ and $\varphi''=0$. \newpage \subsection{A single equation?}\label{disc} The Einstein-like tensor $\mathcal{E}$ reveals an interesting peculiarity when expressed explicitly in terms of $\varphi$. As you can see in \textbf{Appendix \ref{E}}, its general form for the spin-$s$ tensor $\mathcal{E}_s$ seems to be \begin{equation} \mathcal{E}_{s} [\varphi_s]= \mathcal{E}_{s-1}[\varphi_s] + \Delta_{s}[\varphi_s] \, , \end{equation} where $\Delta_s$ contains only those terms that become non-vanishing for spin $s$, i.e. \begin{equation} \Delta_s [\varphi_{s'}] = 0 \quad \text{for} \quad s' < s \, . \end{equation} In other words, for spin $s$, all tensors $\mathcal{E}_k$ where $k\geq s$ are equally valid, since they trivially reduce to $\mathcal{E}_s$. What this seems to imply is not only that there is a unique Einstein-like tensor that leads to a valid theory \textbf{for each spin}, but that there is a \textbf{single tensor} $\mathcal{E}_\infty$ that leads to a valid theory \textbf{for all spins}. Note that we have only evaluated $\mathcal{E}_s$ up to $s=15$ using computer-assisted methods described in \textbf{Appendix \ref{program}}, but it certainly seems natural that this pattern holds for general spin $s$. This conjecture remains to be proved, and the possibility of an explicit construction of $\mathcal{E}_\infty$ also remains an open question. \section{Conclusion}\label{s8} We have shown the proper form of equations for a theory of massless higher-spin bosons interacting with a generic external current. As it turns out, to construct a consistent unconstrained local theory, we either have to introduce non-local operators or high derivatives. This construction leads to a \textit{unique} theory for each spin, perhaps even a unique theory for \textit{all spins}, as discussed in \textbf{Section \ref{disc}}. Putting the $\mathrm{(A)dS}$ and the fermionic theory aside, an interesting step forward would perhaps be to find the proper HS gauge corresponding to the \textit{de Donder} gauge for the spin-$2$ theory. Some interesting results regarding generalized \textit{de Donder} gauges can be found in \cite{fs2}, but they certainly deserve further investigation. Note that in spacetimes with more than four dimensions, fully symmetric tensors do not exhaust all the available possibilities and one should also consider mixed-symmetry tensors. This was considered, for example, in \cite{mix1}, \cite{mix2} or \cite{mix3}.

d7bb7f9746110e6682bfc47ed2a045b05745d9f4

\section{Introduction} \section{Introduction} In recent years, there has been considerable interest in investigating the physical properties related to quantum systems with long-range interactions \cite{Lahaye2009, Peter2012, Gong2016, Zeiher2017}. This is primarily in response to the significant developments made in experimental atomic, molecular and optical (AMO) physics \cite{Bloch2008,Saffman2010,Douglas2015}, where such interactions can be implemented in a well controlled setting \cite{Schauss2012,Britton2012,Yan2013,Islam2013}. These studies have led to a flurry of exciting new physical phenomena \cite{Hauke2013,Sciolla2011,Zunkovic2018, Richerme2014, Jurcevic2014, Schachenmayer2013, Buyskikh2016, Bruno2001, Laflorencie2005, Lobos2013, Maghrebi2017, Smith2016, Jaschke2017, Neyenhuis2017, Eldredge2017}, for instance, propagation of correlations faster than the Lieb-Robinson bound \cite{Hauke2013,Richerme2014,Jurcevic2014}, emergence of exotic long-range order \cite{Laflorencie2005,Lobos2013,Bruno2001,Maghrebi2017}, and dynamical phase transitions \cite{Sciolla2011,Zunkovic2018}. {Most of these phenomena arise in the presence of long-range interactions due to the breakdown of ``quasi-locality"~\cite{Eisert2013,Gong2017} (cf.~\cite{Luitz2019}). In this context, quasi-locality affirms the existence of a non-relativistic spatial light-cone within which most of the causal information travels with the finite Lieb-Robinson velocity~\cite{Lieb1972}. Any correlations or response to local fluctuations appear to be strongly suppressed at small distances away from this light-cone boundary \cite{Hastings2006,Nachtergaele2006}, which may not be the case when long-range interactions are present.} Importantly, the loss of quasi-locality can lead to nontrivial distribution of quantum entanglement \cite{Horodecki2009}, which over the years has been established as an important resource in implementation of various quantum information and computation protocols \cite{Ekert1991,Bennett1993,Bennett1992} (also see \cite{Nielsen2000}). While recent studies have { focused} on the growth of entanglement between two parties in a variable-range interacting system \cite{Schachenmayer2013, Buyskikh2016}, the effect of emergent quasi-nonlocality due to long-range interactions on the global entanglement of these systems remains elusive \cite{Pappalardi2018}. Here, we address this void by investigating the multipartite entanglement in the ground and quenched states of quantum many-body systems with long-range interactions. It is known that entanglement jointly distributed among many parties has richer features \cite{Vidal2000,Verstraete2002}, which has allowed for the design of sophisticated protocols such as cryptographic conference \cite{Horne1992, Bose1998} and multiparty quantum communication \cite{Karlsson1998,Bandyopadhyay2000,Rigolin2005,Yeo2006,Agrawal2006,Ghose2016}. Multiparty entangled states are also intrinsic resources in implementation of novel quantum computation models such as measurement-based quantum computation \cite{Raussendorf2001}. In the past decade, notable progress in experimental physics has allowed for the efficient creation and manipulation of multiparty entanglement \cite{Eibl2004,Prevedel2009,Gao2010,Pan2012,Yao2012}. This opens up the exciting potential for harnessing systems with tunable range of interactions for physical realization of these quantum protocols. Moreover, multiparty entanglement is also an important characteristic quantity in the study of critical phenomena in many-body systems \cite{Wei2005,Cui2008,Orus2008,Giampaolo2013,Stasinska2014,Hofmann2014,Roy2017,Pezze2017}. In this work, we consider a spin-1/2 Heisenberg chain with spin interactions that follow a power-law decay (${1}/{r^{\alpha}}$). Efficient implementation of such variable interactions has been possible with recent developments in AMO physics, in particular with cold atoms \cite{Douglas2015} where the parameter $\alpha$ can be tuned. Other systems include Rydberg atoms \cite{Schauss2012}, trapped ions \cite{Britton2012} and polar molecules \cite{Yan2013}. Incidentally, it is known that such power-law decay in Heisenberg chain can lead to breakdown of quasi-locality \cite{Maghrebi2017,Luitz2019}. An important ramification of this is that the area law no longer bounds the entanglement entropy \cite{Gong2017} (cf.~\cite{Hastings2007,Eisert2010}), especially for $\alpha \leq 1$. In the same vein, intuitively, one would expect that the spatial nonlocal effects induced by the long-range interactions will result in quantum phases with enhanced global entanglement. To explore this further, we characterize the multiparty entanglement in both the ground and quenched states of the considered Hamiltonian. We observe a clear dichotomy between two different regimes, depending on whether the interactions in the $x$--$y$ spin plane are antiferromagnetic (AFM) or ferromagnetic (FM). {We note that while the ground states} in the FM regime have enhanced multiparty entanglement for increased range of interactions in the system, counterintuitively, for the AFM regime the global entanglement weakly diminishes. {Interestingly, we note that this is no longer the case when quantum states are quenched with such long-range interactions. Here, we start from a completely separable or product spin state and switch on the interactions. The subsequent growth of multipartite entanglement in the time-evolved system is then numerically analyzed.} {Here, we observe that the} AFM interactions with long-range are more favorable towards the growth of multiparty entanglement, in contrast to the FM interactions, where the growth appears almost independent of the range of interactions. Thus, our findings clearly demonstrate that long-range interaction selectively enhances quantum resources, such as global entanglement in the system, and this is important for experimental efforts to generate entanglement and implement {quantum information and computation} protocols using systems with variable-range interactions. The paper is arranged as follows. We introduce the spin-1/2 Heisenberg chain with long-range interactions in Sec.~\ref{model}. Our measure of genuine multipartite entanglement and its computaion is discussed in Sec.~\ref{sec:GGM_def}. In Sec.~\ref{sec:results}, we analyze the genuine multiparty entanglement in the ground states of the long-range model. The growth of entanglement under quantum quench is then investigated in Sec.~\ref{sec:quench}, before we end with a final discussion on the results in Sec.~\ref{sec:discussions}. \section{\label{model} Model} We start by introducing the physical system of our interest, the one dimensional (1D) quantum spin lattice, consisting of spin-1/2 particles, coupled via long-range interactions with power-law decay. The Heisenberg Hamiltonian governing such a system can be written as \begin{eqnarray} \mathcal{H}=\sum_{i<j}\frac{1}{|i-j|^{\alpha}}(J_x\sigma^x_i \sigma^x_{j}+J_y\sigma^y_i\sigma^y_{j}+\Delta\sigma^z_i \sigma^z_{j}), \label{Long_Ham} \end{eqnarray} where $\alpha\geq0$ is the continuous exponent that controls the long-range interaction. $J_x$ and $J_y$ are the coupling constants along the $x$ and $y$ spin axes, respectively, and $\Delta$ is the anisotropy along the $z$-direction. Here, $\sigma^m$'s are the Pauli spin matrices ($m\in \{x,y,z\}$). For $J_x, J_y < 0$, the interaction in the $x$--$y$ plane is ferromagnetic, while for $J_x, J_y > 0$, we obtain the antiferromagnetic coupling. The above Hamiltonian, in the presence of long-range interactions, has a rich phase diagram. For $J_x = J_y = -1$, an exotic continuous symmetry breaking (CSB) phase emerges \cite{Maghrebi2017} for low values of $\alpha$, apart from the known \emph{XY} and AFM phases observed in the short-range Hamiltonian. This is a true hallmark of the quasi-nonlocal effect and {the change in effective dimensionality of the system induced by long-range interactions. This is due to the fact that spontaneous breaking of continuous symmetry typically appears only in higher dimensional spin lattices and is otherwise forbidden in low-dimensional systems by the Mermin-Wagner theorem~\cite{Mermin1966} (also see Ref.~\cite{Maghrebi2017}).} For $\alpha=\infty$, the model reduces to the short-range Heisenberg model with nearest-neighbor (NN) interactions. A schematic of a 1D long-range quantum spin system of arbitrary size and with power-law decay of interactions is provided in Fig.~\ref{fig1}. In our study, we consider periodic boundary conditions for the spin chain. { In order to do that we always choose the interaction corresponding to the shortest distance, $|i-j|$, from one site and another in the periodic spin chain.} We note that at $\Delta=0$, for $J_x = J_y$, the above Hamiltonian gives us the long-range \emph{XX} model, which we analyze in our study. Moreover, for $\Delta = J_x = J_y$ and $\alpha=2$, the model reduces to the exactly solvable Haldane-Shastry model \cite{Gaudiano2008,Shastry1988,Haldane1988}. \begin{figure}[t] \includegraphics[width=3.4in,angle=00]{longrange2.pdf} \caption{Schematic of a quantum spin chain with interactions that follow a power-law decay, ${1}/{r^{\alpha}}$. The black (bold) and red (dotted) lines show the short- and long-range coupling between the $j^{th}$ and the other spins in the quantum system. } \label{fig1} \end{figure} \section{\label{sec:GGM_def} Measure of genuine multiparty entanglement} Before going into the detailed analysis of the entanglement properties of the ground and quenched states of the long-ranged Heisenberg model, we begin by defining the genuine multiparty entanglement of a quantum state. We note that there exists several equivalent definitions and measures of multiparty entanglement in the literature \cite{Horodecki2009}. In our work, we are mainly { focused} on the genuine multiparty entanglement of a quantum system \cite{Chiara2018}, which is defined as follows: {\it An $N$-party pure quantum state, $|\psi\rangle_N$, is said to be genuinely multiparty entangled if it cannot be written as a product in any bipartition}. In other words, a genuine multiparty entangled state is entangled across all bipartitions of the system \cite{Goldbart2003, Blasone2008, Sen2010}. In order to estimate this quantity in $|\psi\rangle_N$, we consider the generalized geometric measure (GGM) \cite{Sen2010}, which is a computable measure of genuine multiparty entanglement of a state. It is defined as an optimized distance of the given quantum state, $|\psi\rangle_N$, from the set of all states that are not genuinely multiparty entangled. This can be mathematically expressed as $ \mathcal{G}(|\psi\rangle_N)=1-\Lambda_{\max}^2(|\psi\rangle_N), $ where $\Lambda_{\max} (|\psi\rangle_N ) = \max | \langle \chi|\psi\rangle_N |$, with the maximization being over all such pure quantum state $|\chi\rangle$ that are not genuinely multiparty entangled. Following some simplifications, one can derive an equivalent expression for the above equation, given by \cite{Sen2010} \begin{equation} \mathcal{G} (|\psi \rangle_N ) = 1 - \max \{\lambda^2_{ A : B} | A \cup B = \{s_1,\ldots, s_N\}, A \cap B = \emptyset\}, \label{GGM} \end{equation} where \(\lambda_{A:B}\) is the maximal Schmidt coefficient of $|\psi\rangle_N$, in the bipartition $A:B$. The measure is then optimized over all possible bipartitions of the state, $|\psi\rangle_N$, and for spin-1/2 or qubit systems, takes values in the range: $ 0 \leq \mathcal{G} (|\psi\rangle_N ) \leq 1/2$. In recent years, GGM has been used to characterize genuine multiparty entanglement in strongly-correlated systems, including quantum spin liquids \cite{Dhar2013,Roy2016}, doped spin lattices \cite{Roy2017b,Das2018}, and other many-body systems \cite{Prabhu2011, Jindal2014, Mishra2016, Biswas2014, Sadhukhan2017}. We note that the computation of GGM in many-body quantum systems requires access to the complete state of the system and all its reduced density matrices. In general, for a quantum system with $N$ number of sites, the number of such reduced density matrices is given by $\sum_{i=1}^{N/2} {N \choose i} $, which increases exponentially with the size of the system. Moreover, in the presence of long-range interactions, there are no known analytical or approximate methods, such as tensor-networks or matrix product states, which can be used to compute GGM, as is the case in several short-range models (see Refs.~\cite{Dhar2013,Roy2016,Roy2017b,Roy2018}). Therefore, in our case, we are restricted to exact numerical solutions for small, finite spin chains. In our work, we have considered systems with up to $N$ = 20 spins, and use diagonalization and propagation methods based on the Krylov subspace and Lanczos algorithm. {To mitigate the effect of unstable finite-size effects in the presence of long-range interactions, we have also checked the qualitative consistency of our main results against smaller system-sizes.} \section{\label{sec:results} Multiparty entanglement in the ground state} We now study the variation of genuine multiparty entanglement ($\mathcal{G}$) in the {ground states} of the spin-1/2 Heisenberg chain, with long-range interactions, given by Eq.~(\ref{Long_Ham}). Towards that aim, we consider two distinct regimes emanating from the Hamiltonian, i) The ferromagnetic regime, with interactions in the $x$--$y$ plane given by $J_x = J_y = -1$, and ii) the antiferromagnetic, with $J_x=J_y=1$. Here, we consider only the antiferromagnetic interactions along the $z$-axis, i.e.,$0 \leq \Delta/J \leq 2$ ($J$ = $|J_x| = |J_y|$), where there always exists a distinct gap between the lowest energy values. The long-range interaction in the system is controlled through the exponent $\alpha$, which is varied in the integer range, $1 \leq \alpha \leq 10$. We exclude the extreme points corresponding to systems with infinite interactions ($\alpha = 0$) or strictly NN interactions ($\alpha = \infty$). We start with the FM regime, and consider the case where the interaction is defined by $\alpha = 10$, with variable anisotropy between the $x$--$y$ and $z$ directions. We note that the system is already short-range for $\alpha = 10$. In Fig.~\ref{fig2}, we note that the ground state is genuinely multiparty entangled for all values of the anisotropy parameter, $\Delta/J$ ($0\leq\Delta/J\leq 2$), with the minimum $\mathcal{G}$ at the point, $\Delta/J = 1$. As the range of interaction is increased, by decreasing $\alpha$, the genuine multipartite entanglement increases monotonically, for $\Delta/J < 1$. Subsequently, at higher values of anisotropy ($\Delta/J > 1$) the local minimum shifts to larger values of $\Delta/J$ for decreasing values of $\alpha$. Moreover, there is a cross-over between $\mathcal{G}$ values of different $\alpha$, which gives rise to an interesting regime where the shortest-range ($\alpha = 10$) and the longest-range of interactions ($\alpha = 1$) generate ground states with higher global entanglement than the intermediate range of interactions. More importantly, $\mathcal{G}$ remains the highest at $\alpha = 1$, and gradually decreases with increasing $\Delta$. Therefore, as long-range interactions in the system {increase} there is an expected hike in the genuine multipartite entanglement {in the ground state} of the ferromagnetic spin-1/2 Heisenberg Hamiltonian. \begin{figure}[t] \includegraphics[width=3in,angle=00]{FM_new.pdf} \caption{(Color online.) Variation of genuine multipartite entanglement. Here, we consider a Heisenberg chain with $N = 20$ spins, and FM interactions ($J_x=J_y=-1$) in the $x$--$y$ plane. The plot shows the variation in $\mathcal{G}$ with the parameter $\Delta/J$, where $J$ = $|J_x| = |J_y|$, for ten different integer values of the exponent $\alpha$, ranging from $\alpha = 1$ (red-squares) to $\alpha = 10$ (green-circles). Here, the dashed lines are fits to the plotted data points. The regime $\Delta=0$ corresponds to the \emph{XX} model which also mimics the result obtained for Heisenberg chain at low $\Delta$. We note that the plots for higher $\alpha$ values are very close together, which shows that the short-range character is reached fairly quickly. Moreover, the solid-red vertical arrows highlight specific parameter regimes where $\mathcal{G}$ increases with decreasing $\alpha$ (or increasing long-range interactions), whereas the dashed-green vertical arrows show regions where $\mathcal{G}$ increases but now for increasing $\alpha$ (or decreasing long-range interactions). Both the axes are dimensionless. } \label{fig2} \end{figure} In the AFM regime, the situation is drastically different. For the short-range interaction ($\alpha = 10$), the genuine multiparty entanglement of the ground state is minimum at $\Delta/J = 1$, with a distinct symmetry around the point, as shown in Fig.~\ref{fig3}. In contrast to the FM regime, $\Delta/J = 1$ is the local minima of $\mathcal{G}$, for all values of $\alpha$, {although in the vicinity of this point the multipartite entanglement increases for more long-range interactions, which is similar to the FM regime.} However, away from this point, $\mathcal{G}$ decreases as the long-range interaction in the system is increased. This intriguing behavior of genuine multipartite entanglement is in direct contrast to the behavior of the system in the FM regime, and implies a negative interdependence between global entanglement and long-range induced nonlocal effects in the system. This is significant from the perspective of physical implementation of quantum protocols where multiparty entanglement is an important resource. In the AFM regime, long-range interactions appear to be detrimental to generating large entangled states, as compared to the FM regime. \begin{figure}[t] \includegraphics[width=3in,angle=00]{AFM_new.pdf} \caption{(Color online.) Variation of genuine multipartite entanglement. Here, we consider a Heisenberg chain with $N = 20$ spins, and AFM interactions ($J_x=J_y=1$) in the $x$--$y$ plane. The plot shows the variation in $\mathcal{G}$ with the parameter $\Delta/J$ (where, $J$ = $|J_x| = |J_y|$), for ten different integer values of the exponent $\alpha$, ranging from $\alpha = 1$ (red-squares) to $\alpha = 10$ (green-circles). Once more, the dashed lines are just fits to the plotted data points and both the axes here are dimensionless. The red and green vertical arrows here imply the same behavior as outlined in Fig.~\ref{fig2}. } \label{fig3} \end{figure} {The difference in the behavior of the genuine multipartite entanglement between the FM and AFM regimes, can be partly explained using a heurtistic description of the ground state in these regimes based on our numerical simulations. Here, the competition between different ground state configurations in interacting many-body systems gives rise to the phenomena of entanglement-frustration \cite{Dawson2019} (also see Refs.~\cite{Roy2017,Jindal2014}), which can potentially define the complex behavior of entanglement in our model. In the long-range interaction model that we consider, the ground state can be written as a superposition between two stable, but competing configurations, such that $|\psi_g\rangle = a~|\psi_\mathrm{N}\rangle + b~|\phi_\mathrm{\bar{N}}\rangle$. Here, $|\psi_\mathrm{N}\rangle$ is the state arising due to the N{\'e}el order at $\Delta > 1$ (for $J$ = 1). It is expected that at large $\Delta$, the ground state will be closer to the N{\'e}el state for both the AFM and FM model. For large $\alpha$, this is simply given by $|\psi\rangle$ = $|\uparrow\downarrow\uparrow\cdots\downarrow\rangle$, where $\{\uparrow, \downarrow\}$ are the eigenstates of $\sigma_z$. However, the complementary configuration $|\psi'\rangle$ = $|\downarrow\uparrow\downarrow\cdots\uparrow\rangle$ is also a likely ground state at large $\Delta$. Hence, at dominant $\Delta$ values, frustration ensures that, $|\psi_\mathrm{N}\rangle = \beta_1|\psi\rangle + \beta_2|\psi'\rangle$. The parity symmetry of $\mathcal{H}$ results in $\beta_1 = \pm \beta_2 = 1/\sqrt{2}$, which ensures $|\psi_\mathrm{N}\rangle$ is maximally multiparty entangled. On the other hand, $|\phi_\mathrm{\bar{N}}\rangle$ refers primarily to the non-N{\'e}el configurations in the ground state, which are orthogonal to both $|\psi\rangle$ and $|\psi'\rangle$. These states are significant in regimes where $\Delta$ is not large and seem to arise from the \emph{XY} terms in the Hamiltonian. However, unlike $|\psi_\mathrm{N}\rangle$, the entanglement properties of $|\phi_\mathrm{N}\rangle$ are a priori not known. While the above description is intuitively appealing for regimes that correspond to either large or small values of the anisotropy parameter, numerical analysis suggests that it can also provide a broad picture of the ground state for intermediate values of $\Delta$. By investigating the quantum fidelity of the ground state to $|\psi\rangle$ and $|\psi'\rangle$, one can deduce that the overall entanglement of the ground state is dependent on the trade-off between the states $|\psi_\mathrm{N}\rangle$ and $|\phi_\mathrm{\bar{N}}\rangle$, i.e., the ratio $a/b$. In the AFM case, two distinct regimes emerge for all $\alpha$, symmetric around the point $\Delta = 1$, viz. the region with $a >b$ (for $\Delta > 1$) and the one with $b > a$ (for $\Delta < 1$). We call these the N{\'e}el and non-N{\'e}el regimes. Later, we discuss how these regimes closely correspond to the AFM and \textit{XY} phases respectively. In the N{\'e}el regime, we observe that the ratio $a/b$ not only increases with $\Delta$ but also with $\alpha$, resulting in higher entanglement for shorter-range interactions. Interestingly, in the non-N{\'e}el regime the opposite behavior is observed. Here, it can be numerically shown that the ratio $b/a$ increases for decreasing $\Delta$ values, resulting in more entanglement close to $\Delta = 0$. However, $b/a$ also increases with increasing $\alpha$, which again leads to higher entanglement in short-range systems. This allows a distinct symmetry in multiparty entanglement to emerge around the vicinity of $\Delta = 1$ (in the region, $0\leq\Delta\leq2$), for all values of $\alpha$ in the AFM model, but with short-range interactions leading to more entanglement, as shown in Fig.~\ref{fig3}. Things look more interesting in the FM case, where the effects of long-range interactions become more prominent. Firstly, for $\alpha > 1$, the transition from the N{\'e}el to the non-N{\'e}el regime is no longer centered at $\Delta = 1$, apart from the short-range cases ($\alpha > 6$). The different points of transition on $\Delta$ increases for decreasing $\alpha$. This allows for cross-over between the multiparty entanglement corresponding to different values of $\alpha$. Secondly, and more importantly, there is no N{\'e}el to non-N{\'e}el transition for $\alpha = 1$ in the FM case, at least within the considered parameter regime. Therefore, the ground state always corresponds to high values of $b/a$ (in the non-N{\'e}el regime) and has high multiparty entanglement compared to other values of $\alpha$ (see Fig.~\ref{fig2}).} {Incidentally, we note that in the short-range limit (i.e. $\alpha =10$), due to the $SU(2)$ invariance of the Hamiltonian, the AFM and FM model turns out to be the same at $\Delta=\pm 1$. Moreover, in the short-range limit, the FM and AFM models here are also connected via local unitary operations (spin-flip operations at alternate sites in the spin chain) that keep the global entanglement unchanged. This is reflected in Figs.~\ref{fig2} and \ref{fig3}, where the plots for multiparty entanglement ($\mathcal{G}$) at $\alpha = 10$ are almost same for both the FM and AFM cases.} \begin{figure}[t] \includegraphics[width=3.5in,angle=00]{3dphase.pdf} \caption{(Color online.) Variation of genuine multiparty entanglement ($\mathcal{G}$) in the ground states of the long-range Heisenberg chain consisting of $N = 20$ spins in the $\Delta/J-\alpha$ plane for (a) AFM and (b) FM interactions in the $x$--$y$ plane. Here, $J$ = $|J_x| = |J_y|$. We note here that the dashed yellow lines represent the extremal (minima) points of $\mathcal{G}$. However, they only serve as a visual aid to deconstruct the known quantum phases of the model (see Ref.~\cite{Maghrebi2017}). Both the axes and the color bar in the figures are dimensionless. } \label{fig4} \end{figure} {The dichotomy} in the behavior of genuine multipartite entanglement in the ground state of FM and AFM regimes of the Heisenberg Hamiltonian is closely related to their respective phase structures. In Fig.~\ref{fig4}, we show that the genuine multipartite entanglement is able to deconstruct the different phases in these regimes, as has been established in earlier work \cite{Maghrebi2017}. For large $\alpha$, the phases are similar to their counterparts corresponding to NN spin-1/2 Heisenberg chain, with two distinct phases: the \emph{XY} spin liquid phase and the AFM Ising-like phase. Figures~\ref{fig4}(a)-\ref{fig4}(b), shows how $\mathcal{G}$ distinctly highlights these phases in both the AFM and FM regimes, respectively. We note that the ferromagnetic phase corresponding to $\Delta < -1$ is not shown in the diagram, as $\mathcal{G}$ cannot be uniquely computed for degenerate ground states. The anomalous behavior arises as $\alpha$ is decreased and one enters the quasi-nonlocal regime. For the AFM case, a regime of relatively weak entangled phase appears, with lower values of $\mathcal{G}$. In contrast, in the FM regime, the continuous symmetry breaking phase emerges with decreasing $\alpha$ ($\alpha \lesssim 2$) \cite{Maghrebi2017}, which is marked by a region of high genuine multiparty entanglement. Therefore, in terms of the phase diagram, the increase in genuine multipartite entanglement with increasing long-range interactions is related to the \emph{XY}--CSB phase transition in the FM regime. In addition to this, at the truly long-range interaction limit ($\alpha\sim1$), the ground state mostly remains in the CSB phase, which apparently does not decrease quickly even when the anisotropy and N{\'e}el order increase with $\Delta$. \section{\label{sec:quench}Generating entanglement through quantum quench} We now look at how genuine multipartite entanglement can be generated through a quantum quench mediated {by the variable-range interactions in either the FM or AFM regimes of the spin Hamiltonian}. In particular, we start with a product or completely separable initial state of the system, given by $|\psi\rangle_{in} = \prod_{i}^N|\phi\rangle_i$, where $|\phi\rangle_i = \frac{1}{\sqrt{2}}(|0\rangle_i+|1\rangle_i)$. Here, $|0\rangle_i$ and $|1\rangle_i$ are the eigen states of $\sigma_i^z$. The initial states here can be thought to be ground states of some local Hamiltonian acting identically on all the spins. For the quantum quench, the long-range interactions are instantaneously switched on in the spin system. Subsequently, the initially separable quantum state rapidly evolves in time leading to potential growth of multiparty entanglement in the system. We note that the quench performed in our study is motivated from the perspective of various quantum information and computation protocols, where entanglement is necessary for successful implementation of the protocol. To this end, in our quench process we begin with a completely separable product state, which is a resource-less state, and wish to generate useful resource (entanglement) in the system. Our main aim here is to see whether the presence of long-range interactions in the Heisenberg Hamiltonian can generate higher entanglement or quantum resource in these quenched states {as compared to process that only invoke short-range interactions during the quench. \begin{figure}[h] \includegraphics[width=3.5in,angle=00]{Evolution.pdf} \caption{(Color online.) Genuine multiparty entanglement after a quantum quench. The growth of $\mathcal{G}$ in a system consisting of $N = 12$ spins after a quantum quench of the initially product state, $|\psi\rangle_{in}$, for (a)-(b) AFM and (c)-(d) FM interactions in the $x$--$y$ plane. Here, $J$ = $|J_x| = |J_y|$ and the plots correspond to $\alpha =$ 1 (red-squares), 2 (green-circles), 5 (yellow-triangles), 10 (orange-circles). The red vertical arrow here implies the same behavior as outlined in Fig.~\ref{fig2}. The axes in the above figures are all dimensionless. } \label{fig5} \end{figure} The initial state is subjected to a quantum quench and coherently evolves to $|\psi(t)\rangle = \exp(-i\mathcal{H}t)|\psi\rangle_{in}$. Subsequently, we measure how much GGM is generated in the quenched state, i.e., we calculate $\mathcal{G}(|\psi(t)\rangle)$. We are interested in the parameter regimes away from $\Delta = 1$, where the dichotomy between {the ground states in the FM and AFM cases} appears to be the most distinct. Figure~\ref{fig5}, shows the evolution of the state after the quench. Surprisingly, for the quenched dynamics, long-range interactions ($\alpha = 1$) appears to play a strong role in the growth of multipartite entanglement when $|\psi\rangle_{in}$ is quenched in the AFM regime. In contrast, the generation of multipartite entanglement in the FM regime is almost independent of the range of interactions in the system. This implies that highly entangled quantum states can be generated through quenching in the FM regime even in the absence of any significant long-range interactions. Therefore, in quenched dynamics long-range interactions seem to affect the multiparty entanglement favorably in the AFM regime, while remaining ambivalent in the FM regime. This is converse to the outcome that was observed in the ground state phases of the system. \section{\label{sec:discussions}Discussion} In this work, we have demonstrated how the quasi-nonlocal effect induced by long-range interactions in many-body systems, selectively affects the multipartite entanglement of the system. By investigating different ground state phases of the spin-1/2 Heisenberg Hamiltonian we observed that multiparty entanglement can be enhanced or counterintuitively, can reduce as the range of interactions are increased. In particular, these opposing effects were observed for two distinct ground states phases depending on whether the interaction in the $x$--$y$ plane was ferromagnetic or antiferromagnetic. While the global entanglement is expectedly boosted with more quasi-nonlocal effects for the FM regime, in contrast, long-range interactions {appear to} act detrimentally in the AFM case. {A possible reason for the unexpected behavior in the AFM regime, as we observed in our ground state analysis, is the entanglement-frustration arising from N{\'e}el and non-N{\'e}el terms, which appears to favor more global entanglement in the short-range limit. In the FM case, no transition occur between the N{\'e}el and non-N{\'e}el regimes for long-range interactions, leading to higher entanglement.} Interestingly, the observed dichotomy in these regimes was intriguingly different while considering the generation of multiparty entanglement through quenched dynamics of initially separable states. Here, long-range interactions allow for robust growth of global entanglement in the AFM regime, in contrast to the FM regime, where there is no perceptible advantage in using longer interactions in the quenched dynamics. Overall, our results clearly demonstrate that the system in the ferromagnetic interaction regime is more susceptible to allow significant global entanglement for both short- and long-range interactions. Our findings provide significant insights for physical implementation of quantum protocols where multiparty entanglement is the necessary resource, such as measurement based computation or secure multiparty communication. With recent technological breakthroughs in experimental atomic, molecular and optical physics, where the systems often contain tunable long-range interactions, it is essential to determine the optimal range of interactions that will allow for maximal global entanglement in these system, which can then be harnessed in the quantum protocol. \acknowledgments The authors thank R. Fazio for useful discussions at ICTP, Trieste. HSD acknowledges funding by the Austrian Science Fund (FWF), project no. M 2022-N27, under the Lise Meitner programme of the FWF.

7a5a52a166a99c262975619640b95a80fc7e0c2e

\section{Introduction} We investigate the $\tay$-time-periodic Stokes problem \begin{align}\label{SH} \begin{pdeq} \partial_tu - \Deltau + \nabla p &= f && \text{in }\mathbb{R}\times\R^n_+, \\ \Divu &= g && \text{in }\mathbb{R}\times\R^n_+, \\ u &= h && \text{on }\mathbb{R}\times\partial\R^n_+, \\ u(\tay+t, \cdot) &= u(t, \cdot), \end{pdeq} \end{align} in a half-space $\R^n_+$ of dimension $n\geq 2$. Here, $u\colon\mathbb{R}\times\R^n_+\to{\R^n}$ denotes the velocity field and $p\colon\mathbb{R}\times\R^n_+\to\mathbb{R}$ the pressure term. As customary in the formulation of time-periodic problems, the time-axis is taken to be the whole of $\mathbb{R}$. Variables in the time-space domain $\mathbb{R}\times\R^n_+$ are denoted by $(t,x)$. The time-period $\tay>0$ shall remain fixed. Data $f\colon\mathbb{R}\times\R^n_+\to{\R^n}$, $g\colon\mathbb{R}\times\R^n_+\to\mathbb{R}$ and $h\colon\mathbb{R}\times\partial\R^n_+\to{\R^n}$ that are also $\tay$-time-periodic are considered. The time-periodic Stokes equations play a fundamental role in a wide range of problems in fluid mechanics. Although comprehensive $\LR{p}$ estimates of maximal regularity type are available in the whole-space case \cite{KyedMaxReg14}, similar estimates in the more complicated half-space case were only established recently by \textsc{Maekawa} and \textsc{Sauer} \cite{MaekawaSauer17}. The analysis in \cite{MaekawaSauer17}, however, does not include inhomogeneous boundary data $h\neq 0$. In the following, we shall establish maximal regularity $\LR{p}$ estimates that include the case $h\neq 0$. Such estimates are crucial in a number of applications. For example, the classical approach to free boundary problems in fluid-structure interaction relies heavily on maximal regularity frameworks that include inhomogeneous boundary data. When time-periodic driving forces are studied in such settings, time-periodic Stokes equations appear in the linearization. The nature of the Stokes problem does not allow the treatment of inhomogeneous boundary data by a simple ``lifting'' argument. Consequently, an extension of the results in \cite{MaekawaSauer17} to include the case $h\neq 0$ is by no means trivial. In the following, we employ a different approach than the reflection type argument used in \cite{MaekawaSauer17}. Instead, we use the Fourier-transform $\mathscr{F}_{{\mathbb T}\times\mathbb{R}^{n-1}}$ to reduce \eqref{SH} to an ordinary differential equation in the variable $x_n$. Here, ${\mathbb T}$ denotes the torus $\mathbb{R}/\tay\mathbb{Z}$. The $\LR{p}$ estimates are then established with arguments based on Fourier-multipliers and interpolation techniques. Although the main idea behind this approach is not new, indeed it has been applied successfully by various authors to investigate the initial-value Stokes problem, a number of non-trivial modifications are needed to adapt the arguments to the time-periodic case. Notably, the system \eqref{SH} has to be decomposed into a steady-state part and a so-called purely oscillatory part. Without this decomposition, it seems impossible to establish optimal $\LR{p}$ estimates. Whereas the estimates for the resulting steady-state problem are well-known and can be found in contemporary literature, the estimates for the purely oscillatory part in the following are new. It is convenient to formulate $\tay$-time-periodic problems in a setting of function spaces where the torus ${\mathbb T}\coloneqq\mathbb{R}/\tay\mathbb{Z}$ is used as a time-axis. Indeed, via lifting with the quotient map $\pi\colon\mathbb{R}\to{\mathbb T}$, $\tay$-time-periodic functions are canonically identified as functions defined on ${\mathbb T}$ and vice versa. For such functions, we introduce the simple decomposition \begin{align}\label{intro_projections} \calpu(x) \coloneqq \int_{\mathbb T} u(t,x)\,{\mathrm d}t = \frac{1}{\tay}\int_0^\tay u(t,x)\,{\mathrm d}t,\qquad \calp_\botu(t,x) \coloneqq u(t,x)-\calpu(x) \end{align} into a time-independent part $\calpu$, and a part $\calp_\botu$ with vanishing time-average over the period. We shall refer to $\calpu$ as the \emph{steady-state} part, and to $\calp_\botu$ as the \emph{purely oscillatory} part of $u$. Equipped with the quotient topology, the time-space domain ${\mathbb T}\times\mathbb{R}^n$ is a locally compact abelian group and therefore has a Fourier transform $\mathscr{F}_{{\mathbb T}\times\mathbb{R}^n}$ associated to it. Moreover, we may introduce the Schwartz-Bruhat space $\mathscr{S}({\mathbb T}\times\mathbb{R}^n)$ and its dual space $\mathscr{S^\prime}({\mathbb T}\times\mathbb{R}^n)$ of tempered distributions. Consequently, Bessel potential spaces with underlying time-space domain ${\mathbb T}\times\mathbb{R}^n$ can be defined as subspaces of $\mathscr{S^\prime}({\mathbb T}\times\mathbb{R}^n)$ in a completely standard manner. Sobolev spaces are introduced as Bessel potential spaces with integer exponents, and Sobolev-Slobodecki\u{\i} spaces, \textit{i.e} Sobolev spaces with non-integer exponents, via real interpolation. The same scale of function spaces with respect to the half-space ${\halfspace}$ is obtained by restriction. In a setting of these function spaces (see Section \ref{pre} for the precise definitions and Remark \ref{PerVsTorusFunctionSpacesRemark} below), the main theorem of this article can be formulated as follows: \begin{thm}[Main Theorem]\label{MainThm} Let $q\in(1,\infty)$ and $n\geq 2$. For all \begin{align} \begin{aligned}\label{MainThm_Data} &f\in\LR{q}\bp{{\mathbb T}; \LR{q}(\R^n_+)}^n,\\ &g\in\LR{q}\bp{{\mathbb T}; \WSR{1}{q}(\R^n_+)}\cap\WSR{1}{q}\bp{{\mathbb T}; \WSRD{-1}{q}(\R^n_+)},\\ &h\in\WSR{1-\frac{1}{2q}}{q}\bp{{\mathbb T};\LR{q}(\wholespace)}^n\cap\LR{q}\bp{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}^n \end{aligned} \end{align} with \begin{align}\label{MainThm_DataCompCond} \begin{aligned} &h_{n}\in\WSR{1}{q}({\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)) \end{aligned} \end{align} there is a solution $(u,p)$ to \eqref{SH} with \begin{align} \begin{aligned}\label{MainThm_SolReg} &\calpu\in\WSRD{2}{q}({\halfspace})^n,\\ &\calp_\botu\in\calp_\bot\WSR{1}{q}\bp{{\mathbb T};\LR{q}({\halfspace})}^n\cap\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{2}{q}({\halfspace})}^n,\\ &p\in\LR{q}\bp{{\mathbb T};\WSRD{1}{q}({\halfspace})}, \end{aligned} \end{align} which satisfies \begin{multline}\label{MainThm_ProjEst} \norm{\nabla^2\calpu}_{\LR{q}({\halfspace})} + \norm{\nabla\calpp}_{\LR{q}({\halfspace})}\\ \leq \Ccn{C} \bp{\norm{\calp f}_{\LR{q}({\halfspace})} + \norm{\calpg}_{\WSR{1}{q}({\halfspace})}+\norm{\calph}_{\WSR{2-\frac{1}{q}}{q}(\wholespace)}} \end{multline} and \begin{align}\label{MainThm_ProjComplEst} \begin{aligned} &\norm{\calp_\botu}_{\WSR{1}{q}\np{{\mathbb T};\LR{q}({\halfspace})}\cap\LR{q}\np{{\mathbb T};\WSR{2}{q}({\halfspace})}} + \norm{\nabla\calp_\botp}_{\LR{q}\np{{\mathbb T}; \LR{q}({\halfspace})}} \\ &\qquad \leq \Ccn{C}\, \bp{\norm{\calp_\botf}_{\LR{q}\np{{\mathbb T};\LR{q}(\R^n_+)}}+\norm{\calp_\botg}_{\LR{q}\np{{\mathbb T}; \WSR{1}{q}(\R^n_+)}\cap\WSR{1}{q}\np{{\mathbb T}; \WSRD{-1}{q}(\R^n_+)}}\\ &\qquad\qquad +\norm{\calp_\both}_{\WSR{1-\frac{1}{2q}}{q}\np{{\mathbb T};\LR{q}(\wholespace)}\cap\LR{q}\np{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}} + \norm{\calp_\both_n}_{\WSR{1}{q}\np{{\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)}}\, } \end{aligned} \end{align} with $\Ccn{C}=\Ccn{C}\np{n,q,\tay}$. If $(\tilde{u},\tilde{p})$ is another solution to \eqref{SH} in the class \eqref{MainThm_SolReg}, then $\calp_\botu=\calp_\bot\tilde{u}$, $\calpu=\calp\tilde{u}+(a_1x_n,\ldots,a_{n-1}x_{n},0)$ for some vector $a\in\mathbb{R}^{n-1}$, and $p=\tilde{p} + d(t)$ for some function $d$ that depends only on time. \end{thm} The two separated estimates \eqref{MainThm_ProjEst} and \eqref{MainThm_ProjComplEst} of different regularity type for the steady-state $\calpu$ and the {purely oscillatory} part $\calp_\botu$ of the solution, respectively, demonstrate the necessity of the decomposition. Observe that the purely oscillatory part $\calp_\botu$ of the solution is unique, whereas the steady-state part $\calpu$ is not. The projections $\calp$ and $\calp_\bot$ decompose the time-periodic Stokes problem \eqref{SH} into a classical steady-state Stokes problem with respect to data $(\calpf,\calpg,\calph)$ and a time-periodic Stokes problem with respect to purely oscillatory data $(\calp_\botf,\calp_\botg,\calp_\both)$, respectively. The first estimate \eqref{MainThm_ProjEst} is well-known for the former problem, whence the main objective in the following will be to establish existence of a unique solution to the latter that satisfies \eqref{MainThm_ProjComplEst}. \begin{rem}\label{PerVsTorusFunctionSpacesRemark} It is possible to avoid the analysis on the torus group ${\mathbb T}\coloneqq\mathbb{R}/\tay\mathbb{Z}$ and instead define the function spaces appearing in Theorem \ref{MainThm} as spaces of $\tay$-periodic functions on $\mathbb{R}$. For a Banach space $E$, let \begin{align*} \CR{\infty}_{\mathrm{per}}\left(\mathbb{R}; E\right)\coloneqq \setc{f\in\CR \infty\left(\mathbb{R}; E\right)}{f(t+\tay, x) = f(t, x) } \end{align*} denote the space of smooth $\tay$-time-periodic $E$-valued functions. The Bochner-Lebesgue and Bochner-Sobolev spaces of time-periodic functions can then be introduced as \begin{align*} &\LRper{q}\left(\mathbb{R}; E\right)\coloneqq \closure{\CR{\infty}_{\mathrm{per}}\left(\mathbb{R}; E\right)}{\norm{\cdot}_{\LR{q}\left((0,\tay); E\right)}},\\ &\WSRper{k}{q}\left(\mathbb{R}; E\right) \coloneqq \closure{\CR{\infty}_{\mathrm{per}}\left(\mathbb{R}; E\right)}{\norm{\cdot}_{\WSR{k}{q}\left((0,\tay ); E\right)}}. \end{align*} Observe that the closures above are taken with respect to a time interval of period $\tay$. Time-periodic Sobolev-Slobodecki\u{\i} spaces can then be defined as real interpolation spaces in the usual way. Via the canonical quotient map $\pi\colon\mathbb{R}\to{\mathbb T}$, the spaces $\CR{\infty}_{\mathrm{per}}\bp{\mathbb{R}; E}$ and $\CR \infty\bp{{\mathbb T}; E}$ are isometrically isomorphic with respect to the norms $\norm{\cdot}_{\WSR{k}{q}\left((0,\tay ); E\right)}$ and $\norm{\cdot}_{\WSR{k}{q}\left({\mathbb T}; E\right)}$, respectively, provided the Haar measure on ${\mathbb T}$ is normalized appropriately. It follows that $\WSRper{s}{q}\left(\mathbb{R}; E\right)$ and $\WSR{s}{q}\left({\mathbb T}; E\right)$ are also isometrically isomorphic for all $s$. In this manner, all the function spaces appearing in Theorem \ref{MainThm} have interpretations as $\tay$-time-periodic Bochner spaces. \end{rem} \section{Preliminaries}\label{pre} The objective of this section is to formalize the reformulation of \eqref{SH} in a setting where the time axis is replaced with the torus group ${\mathbb T}\coloneqq\mathbb{R}/\tay\mathbb{Z}$. This includes definitions of the function spaces appearing in Theorem \ref{MainThm}. \subsection{Topology, differentiable structure and Fourier transform} We utilize ${{\mathbb T}\times{\R^n}}$ as a time-space domain. Equipped with the quotient topology induced by the quotient mapping \begin{align*} \pi :\mathbb{R}\times\mathbb{R}^n\to{{\mathbb T}\times{\R^n}}, \ \pi\left(t, x\right) \coloneqq \left(\left[t\right], x\right), \end{align*} ${{\mathbb T}\times{\R^n}}$ becomes a locally compact abelian group. We can identify ${{\mathbb T}\times{\R^n}}$ with the domain $\left[0,\tay\right)\times\mathbb{R}^n$ via the restriction $\pi\big|_{\left[0, \tay\right)\times{\R^n}}$. The Haar measure ${\mathrm d}g$ on ${{\mathbb T}\times{\R^n}}$ is the product of the Lebesgue measure on $\mathbb{R}^n$ and the Lebesgue measure on $\left[0,\tay\right)$. We normalize ${\mathrm d}g$ so that \begin{align*} \int_{{\mathbb T}\times{\R^n}} u(g)\,{\mathrm d}g = \frac{1}{\tay}\int_0^\tay\int_{{\R^n}} u(t, x)\,{\mathrm d}x{\mathrm d}t. \end{align*} There is a bijective correspondence between points $(k, \xi)\in{\frac{2\pi}{\tay}\Z\times{\R^n}}$ and characters $\chi: {{\mathbb T}\times{\R^n}}\to\mathbb{C}$, $\chi\left(t, x\right)\coloneqq e^{ix\cdot\xi + ikt}$ on ${{\mathbb T}\times{\R^n}}$. Consequently, we can identify the dual group of ${{\mathbb T}\times{\R^n}}$ with ${\frac{2\pi}{\tay}\Z\times{\R^n}}$. The compact-open topology on ${\frac{2\pi}{\tay}\Z\times{\R^n}}$ reduces to the product of the Euclidean topology on ${\R^n}$ and the discrete topology on ${\frac{2\pi}{\tay}\Z}$. The Haar measure on ${\frac{2\pi}{\tay}\Z\times{\R^n}}$ is therefore the product of the counting measure on ${\frac{2\pi}{\tay}\Z}$ and the Lebesgue measure on ${\R^n}$. The spaces of smooth functions on ${{\mathbb T}\times{\R^n}}$ and ${\frac{2\pi}{\tay}\Z\times{\R^n}}$ are defined as \begin{align}\label{DefOfSmoothFunctionsOnGrp} \CR \infty({{\mathbb T}\times{\R^n}}) \coloneqq \setc{u:{{\mathbb T}\times{\R^n}}\rightarrow\mathbb{R}}{\exists U\in\CR \infty\bp{\mathbb{R}\times{\R^n}}:\ U=u\circ\pi} \end{align} and \begin{align*} \CR \infty\Bp{{\frac{2\pi}{\tay}\Z\times{\R^n}}} \coloneqq \setcL{u\in\CR{}\Bp{{\frac{2\pi}{\tay}\Z\times{\R^n}}}}{\forall k\in{\frac{2\pi}{\tay}\Z}: u(k, \cdot)\in\CR \infty({\R^n})}, \end{align*} respectively. Derivatives of a function $u\in\CR \infty({{\mathbb T}\times{\R^n}})$ are defined by \begin{align*} \partial_t^\beta\partial_x^\alphau \coloneqq \left[\partial_t^\beta\partial_x^\alpha\left(u\circ\pi\right)\right]\circ\Pi^{-1}, \end{align*} with $\Pi \coloneqq \pi\big|_{\left[0,\tay\right)\times{\R^n}}$. The notion of Schwartz spaces can be extended to locally compact abelian groups (see \cite{Bruhat} and \cite{kyedeiter_PIFBook}). The so-called Schwartz-Bruhat space on ${{\mathbb T}\times{\R^n}}$ is given by \begin{align*} \mathscr{S}\np{{{\mathbb T}\times{\R^n}}} \coloneqq \setc{u\in\CR \infty({{\mathbb T}\times{\R^n}})}{\forall (\alpha, \beta, \gamma)\in\mathbb{N}_0^n\times\mathbb{N}_0\times\mathbb{N}_0^n: \ \rho_{\alpha, \beta, \gamma}(u)<\infty}, \end{align*} where \begin{align*} \rho_{\alpha, \beta, \gamma}(u) \coloneqq \sup_{(t,x)\in{{\mathbb T}\times{\R^n}}}{\left|x^\gamma\partial_t^\beta\partial_x^\alphau(t,x)\right|}. \end{align*} Equipped with the semi-norm topology of the family ${\{\rho_{\alpha, \beta, \gamma} | \left(\alpha, \beta, \gamma\right)\in\mathbb{N}_0^n\times\mathbb{N}_0\times\mathbb{N}_0^n \}}$, $\mathscr{S}({{\mathbb T}\times{\R^n}})$ becomes a topological vector space. The corresponding topological dual space $\mathscr{S^\prime}({{\mathbb T}\times{\R^n}})$ equipped with the weak* topology is referred to as the space of tempered distributions on ${{\mathbb T}\times{\R^n}}$. Distributional derivatives for a tempered distribution $u$ are defined by duality as in the classical case. The Schwartz-Bruhat space on ${\frac{2\pi}{\tay}\Z\times{\R^n}}$ is \begin{multline*} \mathscr{S}\Bp{{\frac{2\pi}{\tay}\Z\times{\R^n}}} \\\coloneqq \setcL{u\in\CR \infty\Bp{{\frac{2\pi}{\tay}\Z\times{\R^n}}}}{\forall (\alpha, \beta, \gamma)\in\mathbb{N}_0^n\times\mathbb{N}_0^n\times\mathbb{N}_0: \ \hat{\rho}_{\alpha, \beta, \gamma}(u)<\infty}, \end{multline*} with the generic semi-norms \begin{align*} \hat{\rho}_{\alpha, \beta, \gamma}(u) \coloneqq \sup_{(k,\xi)\in{\frac{2\pi}{\tay}\Z\times{\R^n}}}{\left|\xi^\alpha\partial_\xi^\beta k^\gammau(k,\xi)\right|} \end{align*} inducing the topology. By $\mathscr{F}_{{\mathbb T}\times{\R^n}}$ we denote the Fourier transform associated to the locally compact abelian group ${{\mathbb T}\times{\R^n}}$ equipped with the Haar measure introduced above: \begin{align*} &\mathscr{F}_{{\mathbb T}\times{\R^n}}:\mathscr{S}\np{{{\mathbb T}\times{\R^n}}}\rightarrow\mathscr{S}\bp{{\frac{2\pi}{\tay}\Z\times{\R^n}}},\\ &\mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{u}(k,\xi)\coloneqq \ft{u}(k,\xi) \coloneqq \frac{1}{\tay}\int_0^\tay\int_{{\R^n}} u(t,x)\,\e^{-ix\cdot\xi-ik t}\,{\mathrm d}x{\mathrm d}t. \end{align*} Recall that $\mathscr{F}_{{\mathbb T}\times{\R^n}}:\mathscr{S}({{\mathbb T}\times{\R^n}})\rightarrow\mathscr{S}({\frac{2\pi}{\tay}\Z\times{\R^n}})$ is a homeomorphism with inverse given by \begin{align*} &\mathscr{F}^{-1}_{{\mathbb T}\times{\R^n}}:\mathscr{S}\bp{{\frac{2\pi}{\tay}\Z\times{\R^n}}}\rightarrow\mathscr{S}\np{{{\mathbb T}\times{\R^n}}},\\ &\mathscr{F}^{-1}_{{\mathbb T}\times{\R^n}}\nb{w}(t,x)\coloneqq \sum_{k\in{\frac{2\pi}{\tay}\Z}}\,\int_{{\R^n}} w(k,\xi)\,\e^{ix\cdot\xi+ik t}\,{\mathrm d}\xi, \end{align*} provided the Lebesgue measure ${\mathrm d}\xi$ is normalized appropriately. By duality, $\mathscr{F}_{{\mathbb T}\times{\R^n}}$ extends to a homeomorphism $\mathscr{F}_{{\mathbb T}\times{\R^n}}:\mathscr{S^\prime}({{\mathbb T}\times{\R^n}})\rightarrow\mathscr{S^\prime}({\frac{2\pi}{\tay}\Z\times{\R^n}})$. The Fourier symbol, with respect to $\mathscr{F}_{{\mathbb T}\times{\R^n}}$, of the projection $\calp$ introduced in \eqref{intro_projections} is the delta distribution $\delta_{\frac{2\pi}{\tay}\Z}$, \textit{i.e.}, the function $\delta_{\frac{2\pi}{\tay}\Z}:{\frac{2\pi}{\tay}\Z}\rightarrow\mathbb{C}$ with $\delta_{\frac{2\pi}{\tay}\Z}(0):=1$ and $\delta_{\frac{2\pi}{\tay}\Z}(k):=0$ for $k\neq 0$. Via the symbol, the projections $\calp$ and $\calp_\bot$ extend to projections on $\mathscr{S^\prime}\np{{{\mathbb T}\times{\R^n}}}$: \begin{align}\label{SymbolsOfProjections} \begin{aligned} &\calp:\mathscr{S^\prime}\np{{{\mathbb T}\times{\R^n}}}\rightarrow\mathscr{S^\prime}\np{{{\mathbb T}\times{\R^n}}},\quad \calpu := \mathscr{F}^{-1}_{{\mathbb T}\times{\R^n}}\bb{\delta_{\frac{2\pi}{\tay}\Z}\,\mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{u}},\\ &\calp_\bot:\mathscr{S^\prime}\np{{{\mathbb T}\times{\R^n}}}\rightarrow\mathscr{S^\prime}\np{{{\mathbb T}\times{\R^n}}},\quad \calp_\botu := \mathscr{F}^{-1}_{{\mathbb T}\times{\R^n}}\bb{\np{1-\delta_{\frac{2\pi}{\tay}\Z}}\,\mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{u}}. \end{aligned} \end{align} At this point, we have introduced ample formalism to reformulate \eqref{SH} equivalently as a system of partial differential equations in the time-space domain ${{\mathbb T}\times{\R^n}}$. Moreover, the Fourier transform $\mathscr{F}_{{\mathbb T}\times{\R^n}}$ enables us the investigate the systems in terms of Fourier-multipliers. Due to the lack of a comprehensive $\LR{q}$-multiplier theory in the general group setting, we shall utilize a so-called Transference Principle for this purpose. The Transference Principle goes back to \textsc{De Leeuw} \cite{Leeuw1965}. The lemma below is due to \textsc{Edwards} and \textsc{Gaudry} \cite{EdwardsGaudry}. \begin{thm}[\textsc{De Leeuw}, \textsc{Edwards} and \textsc{Gaudry}]\label{transference} Let $G$ and $H$ be locally compact abelian groups. Moreover, let $\Phi:\widehat{G}\rightarrow\widehat{H}$ be a continuous homomorphism and $q\in [1,\infty]$. Assume that $m\in\LR{\infty}(\widehat{H};\mathbb{C})$ is a continuous $\LR{q}$-multiplier, i.e., there is a constant $C$ such that \begin{align*} \forall f\in\LR{2}\left(H\right)\cap\LR{q}\left(H\right):\quad \norm{\mathscr{F}^{-1}_H\nb{m\, \mathscr{F}_H\nb{f}}}_q\leq C\norm{f}_q. \end{align*} Then $m\circ\Phi\in\LR{\infty}(\widehat{G}; \mathbb{C})$ is also an $\LR{q}$-multiplier with \begin{align*} \forall f\in\LR{2}\left(G\right)\cap\LR{q}\left(G\right):\quad \norm{\mathscr{F}^{-1}_G\nb{m\circ\Phi\, \mathscr{F}_G\nb{f}}}_q\leq C\norm{f}_q. \end{align*} \end{thm} \begin{proof} See \cite[Theorem B.2.1]{EdwardsGaudry}. \end{proof} \subsection{Function spaces}\label{FunktionSpacesSection} Since our proof of the main theorem is based on Fourier-multipliers and interpolation theory, we find it convenient to introduce all the relevant Sobolev spaces via Bessel-Potential spaces. Classical (inhomogeneous) Bessel-Potential spaces are defined for $s\in\mathbb{R}$ and $q\in[1,\infty)$ by \begin{align*} &\HSR{s}{q}\np{\mathbb{R}^n} \coloneqq \setcl{u\in\mathscr{S^\prime}\np{{\R^n}}}{\mathscr{F}^{-1}_{{\R^n}}\bb{(1+\snorm{\xi}^2)^\frac{s}{2}\mathscr{F}_{{\R^n}}\nb{u}}\in\LR{q}\np{{\R^n}}},\\ &\norm{u}_{\HSR{s}{q}\np{\mathbb{R}^n}} \coloneqq \norm{u}_{s,q}\coloneqq\norm{\mathscr{F}^{-1}_{{\R^n}}\bb{(1+\snorm{\xi}^2)^\frac{s}{2}\mathscr{F}_{{\R^n}}\nb{u}}}_q. \end{align*} Classical Sobolev spaces on ${\R^n}$ are defined as Bessel-Potential spaces of integer order $k\in\mathbb{Z}$, and Sobolev spaces on the half-space $\R^n_+$ via restriction: \begin{align*} \WSR{k}{q}\np{\mathbb{R}^n}\coloneqq\HSR{k}{q}\np{{\R^n}},\qquad\WSR{k}{q}\np{\R^n_+}\coloneqq\setcl{u_{|\R^n_+}}{u\in\WSR{k}{q}\np{{\R^n}}}. \end{align*} Observe that for negative-order spaces, \textit{i.e} when $k<0$, the Sobolev space $\WSR{k}{q}\np{\R^n_+}$ coincides with the dual space $\bp{\WSR{-k}{q'}\np{\R^n_+}}'$ and \emph{not} with the dual space $\np{\WSRN{-k}{q'}\np{\R^n_+}}'$. In the following, it is essential that the former meaning of $\WSR{k}{q}\np{\R^n_+}$ is used. Homogeneous Bessel-Potential spaces are defined in accordance with \cite{TriebelTheoryFunctionSpaces} by introducing the subspace \begin{align*} Z\np{{\R^n}}\coloneqq\setcl{\phi\in\mathscr{S}\np{{\R^n}}}{\forall\alpha\in\mathbb{N}_0^n:\ \partial_x^\alpha\ft{\phi}\np{0}=0} \end{align*} of $\mathscr{S}\np{{\R^n}}$, and for $s\in\mathbb{R}$ and $q\in[1,\infty)$ letting \begin{align*} &\HSRD{s}{q}\np{{\R^n}}\coloneqq\setcl{u\in\SRh^\prime\np{{\R^n}}}{\mathscr{F}^{-1}_{{\R^n}}\bb{\snorm{\xi}^s \mathscr{F}_{{\R^n}}\nb{u}}\in\LR{q}\np{{\R^n}}},\\ &\norm{u}_{\HSRD{s}{q}\np{{\R^n}}}\coloneqq\norm{\mathscr{F}^{-1}_{{\R^n}}\bb{\snorm{\xi}^s \mathscr{F}_{{\R^n}}\nb{u}}}_q. \end{align*} Due to the lack of regularity of $\snorm{\xi}^s$ at the origin, the above definition of $\HSRD{s}{q}\np{{\R^n}}$ is not meaningful as a subspace of $\mathscr{S^\prime}\np{{\R^n}}$. Instead, $\HSRD{s}{q}\np{{\R^n}}$ is defined as a subspace of $\SRh^\prime\np{{\R^n}}$. As such, $\HSRD{s}{q}\np{{\R^n}}$ is clearly a Banach space. As above, we define homogeneous Sobolev spaces on ${\R^n}$ as homogeneous Bessel-Potential spaces of integer order $k\in\mathbb{Z}$, and introduce homogeneous Sobolev spaces on the half-space $\R^n_+$ via restriction: \begin{align*} \WSRD{k}{q}\np{\mathbb{R}^n}\coloneqq\HSRD{k}{q}\np{{\R^n}},\qquad\WSRD{k}{q}\np{\R^n_+}\coloneqq\setcl{u_{|\R^n_+}}{u\in\WSRD{k}{q}\np{{\R^n}}}. \end{align*} By the Hahn-Banach Theorem, any functional in $\SRh^\prime\np{{\R^n}}$ can be extended to a tempered distribution in $\mathscr{S^\prime}\np{{\R^n}}$. If $s\leq 0$, the extension of an element in $\HSRD{s}{q}\np{{\R^n}}$ to $\mathscr{S^\prime}\np{{\R^n}}$ is unique. In the case $s>0$, one may verify that two extensions of an element in $\HSRD{s}{q}\np{{\R^n}}$ differ at most by addition of a polynomial of order strictly less than $s$. With this ambiguity in mind, one may consider $\HSRD{s}{q}\np{{\R^n}}$ as a normed ($s\leq 0$) and semi-normed ($s> 0$) subspace of $\mathscr{S^\prime}\np{{\R^n}}$. Sobolev-Slobodecki\u{\i} spaces of both homogeneous and inhomogeneous type are defined via real interpolation in the usual way. For example, the spaces appearing in Theorem \ref{MainThm} are defined by \begin{align*} &\WSR{2-\frac{1}{q}}{q}(\wholespace)\coloneqq\bp{\LR{q}\np{\wholespace},\WSR{2}{q}\np\wholespace}_{1-\frac{1}{2q},q},\\ &\WSRD{-\frac{1}{q}}{q}(\wholespace)\coloneqq\bp{\WSRD{-1}{q}\np{\wholespace},\LR{q}\np\wholespace}_{1-\frac{1}{q},q} \end{align*} and equipped with the associated interpolation norms. Analogously, Bessel-Potential spaces with underlying time-space domain ${{\mathbb T}\times{\R^n}}$ are introduced via the Fourier transform $\mathscr{F}_{{{\mathbb T}\times{\R^n}}}$ as \begin{align*} &\HSR{r}{q}\bp{{\mathbb T};\HSR{s}{q}\np{\R^n}}\coloneqq\\ &\qquad\setcl{u\in\mathscr{S^\prime}\np{{{\mathbb T}\times{\R^n}}}}{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\np{1+\snorm{k}^2}^{\frac{r}{2}}(1+\snorm{\xi}^2)^\frac{s}{2}\mathscr{F}_{{{\mathbb T}\times{\R^n}}}\nb{u}}\in\LR{q}\bp{{\mathbb T}; \LR{q}\np{\R^n}}} \end{align*} equipped with the canonical norm. Again, we refer to Bessel-Potential spaces of integer order $k,l\in\mathbb{N}$ as Sobolev spaces: \begin{align*} \WSR{k}{q}\bp{{\mathbb T};\WSR{l}{q}\np{\R^n}}\coloneqq\HSR{k}{q}\bp{{\mathbb T};\HSR{l}{q}\np{\R^n}}. \end{align*} Sobolev spaces on the time-space domain ${\mathbb T}\times{\halfspace}$ are defined via restriction of the elements in the spaces above. In order to introduce homogeneous spaces, we let \begin{align*} Z\np{{{\mathbb T}\times{\R^n}}}\coloneqq\setcl{\phi\in\mathscr{S}\np{{{\mathbb T}\times{\R^n}}}}{\forall\alpha\in\mathbb{N}_0^n:\ \partial_x^\alpha\mathscr{F}_{\mathbb{R}^n}\nb{\phi}\np{t,0}=0} \end{align*} and put \begin{align*} &\HSR{r}{q}\bp{{\mathbb T};\HSRD{s}{q}\np{\R^n}}\coloneqq\\ &\qquad\setcl{u\in\SRh^\prime\np{{{\mathbb T}\times{\R^n}}}}{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\np{1+\snorm{k}^2}^{\frac{r}{2}}\snorm{\xi}^s\mathscr{F}_{{{\mathbb T}\times{\R^n}}}\nb{u}}\in\LR{q}\bp{{\mathbb T}; \LR{q}\np{\R^n}}}. \end{align*} As above, we may consider $\HSR{r}{q}\bp{{\mathbb T};\HSRD{s}{q}\np{\R^n}}$ as a subspace of $\mathscr{S^\prime}\np{{{\mathbb T}\times{\R^n}}}$ by extension. Finally, Sobolev-Slobodecki\u{\i} spaces on the domain ${{\mathbb T}\times{\R^n}}$ are defined via real interpolation. For example, \begin{align*} \WSR{1-\frac{1}{2q}}{q}\bp{{\mathbb T};\LR{q}(\wholespace)} \coloneqq \bp{\LR{q}\bp{{\mathbb T};\LR{q}\np\wholespace},\WSR{1}{q}\bp{{\mathbb T};\LR{q}\np\wholespace}}_{1-\frac{1}{2q},q}. \end{align*} In this way, all the function spaces appearing in Theorem \ref{MainThm} attain rigorous definitions. It is easy to verify that these definitions coincide with a classical interpretation as Bochner spaces of vector-valued functions defined on the torus ${\mathbb T}$. \subsection{Interpolation} Although the function spaces appearing in Theorem \ref{MainThm} can all be defined in terms of classical interpolation, our proof of the theorem relies on a somewhat more refined scale of interpolation spaces. More specifically, it is based on anisotropic Besov spaces with underlying time-space domain ${{\mathbb T}\times{\R^n}}$, which we shall show coincide with the function spaces obtained by real interpolation of the Bessel-Potential spaces introduced above. Although this task is mainly technical and does not require any significantly new ideas, indeed we shall mimic the proofs of similar results for classical isotropic Besov spaces, these spaces and their interpolation properties are not part of contemporary literature and we shall therefore carry out the identification here (even in slightly more generality than actually needed for the proof of the main theorem). To this end, we fix an $m\in\mathbb{N}$ and introduce the parabolic length scale \begin{align}\label{BS_LengthScale} \parnorm{\eta}{\xi}\coloneqq(|\eta|^2+|\xi|^{4m})^{\frac{1}{4m}}\quad {\text{for } (\eta,\xi)\in\mathbb{R}\times\mathbb{R}^n}. \end{align} The anisotropic Besov spaces defined below pertain to time-periodic parabolic problems of order $2m$. In our analysis of the Stokes problem, we thus put $m=1$. For simplicity, we omit $m$ in the notation for the function spaces below. The anisotropic Besov spaces shall be based on the following anisotropic partition of unit: \begin{lem}\label{BS_PartitionOfUnity} Let $m\in\mathbb{N}$ and $\parnorm{\eta}{\xi}$ be given by \eqref{BS_LengthScale}. There is a $\phi\in\CR \infty_0\np{\mathbb{R}\times\mathbb{R}^n}$ satisfying \begin{align} &\supp\phi = \setc{(\eta,\xi)}{2^{-1}\leq \parnorm{\eta}{\xi}\leq 2} \label{BS_PartitionOfUnity_Support},\\ &\phi(\eta,\xi)>0 \quad\text{for}\quad 2^{-1}<\parnorm{\eta}{\xi}<2,\label{BS_PartitionOfUnity_Positivity}\\ &\sum_{l=-\infty}^\infty \phi(2^{-2ml}\eta,2^{-l}\xi)=1\quad \text{for}\quad \parnorm{\eta}{\xi}\neq 0.\label{BS_PartitionOfUnity_PartUnity} \end{align} \end{lem} \begin{proof} Let $h\in\CR \infty(\mathbb{R})$ with $\supph=\setc{y\in\mathbb{R}}{2^{-1}\leq\snorm{y}\leq2}$ and $h(y)>0$ for $2^{-1}<\snorm{y}<2$. Then $f:\mathbb{R}\times\mathbb{R}^n\rightarrow\mathbb{R},\ f(\eta,\xi)\coloneqqh\bp{\parnorm{\eta}{\xi}}$ satisfies \eqref{BS_PartitionOfUnity_Support} and \eqref{BS_PartitionOfUnity_Positivity}. Moreover, $f(2^{-2ml}\eta,2^{-l}\xi)\neq 0$ iff $2^{l-1}<\parnorm{\eta}{\xi}<2^{l+1}$. Thus $f(2^{-2ml}\eta,2^{-l}\xi)\neq 0$ for at least one and at most two $l\in\mathbb{Z}$. Consequently, \begin{align*} \phi:\mathbb{R}\times\mathbb{R}^n\rightarrow\mathbb{R},\quad \phi(\eta,\xi)\coloneqq \begin{pdeq} &\frac{f(\eta,\xi)}{\sum_{l=-\infty}^\infty f(2^{-2ml}\eta,2^{-l}\xi)} &&\text{if }\parnorm{\eta}{\xi}\neq 0,\\ &0 && \text{if }\parnorm{\eta}{\xi}=0 \end{pdeq} \end{align*} is well-defined. It is easy to verify that $\phi$ satisfies \eqref{BS_PartitionOfUnity_Support}--\eqref{BS_PartitionOfUnity_PartUnity}. \end{proof} \begin{defn}[Anisotropic Besov and Bessel-Potential Spaces]\label{BS_DefOfBesovSpace} Let $\phi\in\CR \infty_0\np{\mathbb{R}\times\mathbb{R}^n}$ be as in Lemma \ref{BS_PartitionOfUnity}, $s\in\mathbb{R}$ and $p,q\in[1,\infty)$. We define anisotropic Besov spaces \begin{align}\label{BS_DefOfBesovSpace_Defn} \begin{aligned} &\BSRcompl{s}{pq}\np{{{\mathbb T}\times{\R^n}}}\coloneqq\setc{f\in\calp_\bot\mathscr{S^\prime}({{\mathbb T}\times{\R^n}})}{\norm{f}_{\BSRcompl{s}{pq}}<\infty},\\ &\norm{f}_{\BSRcompl{s}{pq}} \coloneqq \Bp{\sum_{l=0}^\infty \bp{2^{sl} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_p}^q}^{\frac{1}{q}}, \end{aligned} \end{align} and anisotropic Bessel-Potential spaces \begin{align}\label{BS_DefOfBesovSpace_DefnBessel} \begin{aligned} &\ABPSRcompl{s}{p}\np{{{\mathbb T}\times{\R^n}}}\coloneqq\setc{f\in\calp_\bot\mathscr{S^\prime}({{\mathbb T}\times{\R^n}})}{\norm{f}_{\ABPSRcompl{s}{p}}<\infty},\\ &\norm{f}_{\ABPSRcompl{s}{p}} \coloneqq \norm {\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\parnorm{k}{\xi}^s\mathscr{F}_{{{\mathbb T}\times{\R^n}}}\nb{f}}}_p. \end{aligned} \end{align} \end{defn} Observe that $\BSRcompl{s}{pq}\np{{{\mathbb T}\times{\R^n}}}$ and $\ABPSRcompl{s}{p}\np{{{\mathbb T}\times{\R^n}}}$ are defined as subspaces of the purely oscillatory distributions $\calp_\bot\mathscr{S^\prime}({{\mathbb T}\times{\R^n}})$ rather than $\mathscr{S^\prime}({{\mathbb T}\times{\R^n}})$. Recalling \eqref{SymbolsOfProjections}, it is easy to verify that $\norm{\cdot}_{\BSRcompl{s}{pq}}$ and $\norm{\cdot}_{\ABPSRcompl{s}{p}}$ are therefore norms (rather than mere semi-norms), and $\BSRcompl{s}{pq}\np{{{\mathbb T}\times{\R^n}}}$ and $\ABPSRcompl{s}{p}\np{{{\mathbb T}\times{\R^n}}}$ Banach spaces. As in the case of classical (isotropic) spaces, real interpolation of anisotropic Bessel-potential spaces yields anisotropic Besov spaces: \begin{lem}\label{BS_InterpolationLem} Let $p,q\in(1,\infty)$, $\theta\in(0,1)$, $s_0,s_1\in\mathbb{R}$ and $s\coloneqq(1-\theta)s_0+\theta s_1$. If $s_0\neq s_1$, then $\bp{\ABPSRcompl{s_0}{p}\np{{\mathbb T}\times{\R^n}},\ABPSRcompl{s_1}{p}\np{{\mathbb T}\times{\R^n}}}_{\theta,q}=\BSRcompl{s}{pq}\np{{\mathbb T}\times{\R^n}}$ with equivalent norms. \end{lem} \begin{proof} For $l\in\mathbb{N}_0$ and $r\in\mathbb{R}$ let \begin{align*} \mathfrak{m}^r_l:\mathbb{R}\times\mathbb{R}^n\rightarrow\mathbb{C},\quad\mathfrak{m}^r_l(\eta,\xi)\coloneqq \phi\bp{2^{-2ml}\eta,2^{-l}\xi}\parnorm{\eta}{\xi}^{-r}. \end{align*} We claim that $\mathfrak{m}^r_l$ is an $\LR{p}\np{\mathbb{R}; \LR{p}\np{\R^n}}$-multiplier, which we verify by showing that $\mathfrak{m}^r_l$ meets the condition of Marcinkiewicz's multiplier theorem (see for example \cite[Corollary 6.2.5]{Grafakos}). For this purpose, we utilize only \begin{align} &\supp\phi\bp{2^{-2ml}\cdot,2^{-l}\cdot}=\setc{(\eta,\xi)\in\mathbb{R}\times\mathbb{R}^n}{2^{l-1}\leq\parnorm{\eta}{\xi}\leq2^{l+1}},\label{BS_InterpolationThm_SuppProperty} \end{align} and that $g(\eta,\xi)\coloneqq\parnorm{\eta}{\xi}^{-r}$ is parabolically $\np{-r}$-homogeneous, that is, \begin{align} &\forall\lambda>0:\quad g(\eta,\xi)=\lambda^{-r} g(\lambda^{-2m}\eta,\lambda^{-1}\xi).\label{BS_InterpolationThm_HomoProperty} \end{align} From \eqref{BS_InterpolationThm_SuppProperty} we immediately obtain $\norm{\mathfrak{m}^r_l}_\infty \leq \Ccn{C} \norm{\phi}_\infty 2^{-lr}$, with $\Ccnlast{C}$ independent on $l$. By \eqref{BS_InterpolationThm_HomoProperty}, we further observe that \begin{align*} \eta\,\partial_\eta\mathfrak{m}^r_l(\eta,\xi) &= 2^{-2ml}\eta\, \partial_\eta\phi\bp{2^{-2ml}\eta,2^{-l}\xi}\, g(\eta,\xi) \\ &\quad + \phi\bp{2^{-2ml}\eta,2^{-l}\xi}\,\lambda^{-r}\, \partial_\eta g(\lambda^{-2m}\eta,\lambda^{-1}\xi)\,\lambda^{-2m}\eta. \end{align*} Choosing $\lambda\coloneqq\parnorm{\eta}{\xi}$ and recalling \eqref{BS_InterpolationThm_SuppProperty}, we thus deduce $\norm{\eta\,\partial_\eta\mathfrak{m}^r_l}_\infty \leq \Ccn{C} \norm{\phi}_\infty 2^{-lr}$, with $\Ccnlast{C}$ independent on $l$. Similarly, we obtain \begin{align*} \sum_{\alpha\in\set{0,1}^{n+1}} \norm{\xi_1^{\alpha_1}\cdots\xi_n^{\alpha_n}\eta^{\alpha_{n+1}} \partial_{\xi_1}^{\alpha_1}\cdots\partial_{\xi_n}^{\alpha_n}\partial_{\eta}^{\alpha_{n+1}} \mathfrak{m}^r_l}_\infty\leq \Ccn{C} \norm{\phi}_\infty 2^{-lr} \end{align*} with $\Ccnlast{C}$ independent on $l$. It follows from Marcinkiewicz's multiplier theorem that $\mathfrak{m}^r_l$ is an $\LR{p}(\mathbb{R}; \LR{p}(\mathbb{R}^n))$-multiplier. Consequently, the Transference Principle (Theorem \ref{transference}) implies that $\multrestriction{{\mathfrak{m}^r_l}}{{\frac{2\pi}{\tay}\Z}\times\mathbb{R}^n}$ is an $\LR{p}({\mathbb T}; \LR{p}({\R^n}))$-multiplier with \begin{align*} \normL{\phi\mapsto\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\mathfrak{m}^r_l(k,\xi)\mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{\phi}}}_{\mathscr{L}\np{\LR{p}\np{{\mathbb T}; \LR{p}\np{\R^n}},\LR{p}\np{{\mathbb T}; \LR{p}\np{\R^n}}}}<\Ccn{C} \norm{\phi}_\infty 2^{-lr}. \end{align*} Let $f\in\bp{\ABPSRcompl{s_0}{p}\np{{\mathbb T}\times{\R^n}},\ABPSRcompl{s_1}{p}\np{{\mathbb T}\times{\R^n}}}_{\theta,q}$. Consider a decomposition $f=f_0+f_1$ with $f_0\in\ABPSRcompl{s_0}{p}\np{{\mathbb T}\times{\R^n}}$ and $f_1\in\ABPSRcompl{s_1}{p}\np{{\mathbb T}\times{\R^n}}$. We deduce \begin{align*} &\norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_p \\ &\quad\leq \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\mathfrak{m}^{s_0}_l\mathscr{F}_{{\mathbb T}\times{\R^n}}\bb{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\parnorm{k}{\xi}^{s_0}\mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f_0}}}}}_p\\ &\quad\quad + \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\mathfrak{m}^{s_1}_l\mathscr{F}_{{\mathbb T}\times{\R^n}}\bb{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\parnorm{k}{\xi}^{s_1}\mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f_1}}}}}_p\\ &\quad\leq \Ccn{C}\bp{2^{-ls_0}\norm{f_0}_{\ABPSRcompl{s_0}{p}} + 2^{-ls_1}\norm{f_1}_{\ABPSRcompl{s_1}{p}} }\\ &\quad\leq \Ccn{C}2^{-ls_0}\bp{\norm{f_0}_{\ABPSRcompl{s_0}{p}} + 2^{l(s_0-s_1)}\norm{f_1}_{\ABPSRcompl{s_1}{p}} }. \end{align*} We now employ the $K$-method (see for example \cite[Chapter 3.1]{BL76}) to characterize the interpolation space $\bp{\ABPSRcompl{s_0}{p}\np{{\mathbb T}\times{\R^n}},\ABPSRcompl{s_1}{p}\np{{\mathbb T}\times{\R^n}}}_{\theta,q}$. Taking infimum over all decompositions $f_0,f_1$ in the inequality above, we find that \begin{align*} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_p \leq \Ccn{C}\, 2^{-ls_0}\,K\bp{2^{l(s_0-s_1)},f,\ABPSRcompl{s_0}{p},\ABPSRcompl{s_1}{p}}, \end{align*} which implies \begin{align*} \norm{f}_{\BSRcompl{s}{pq}} \leq \Ccn{C} \Bp{\sum_{l=0}^\infty\bp{2^{\theta l(s_1-s_0)}\,K\np{2^{l(s_0-s_1)},f,\ABPSRcompl{s_0}{p},\ABPSRcompl{s_1}{p}}}^q}^{\frac{1}{q}} \leq \Ccn{C}\,\norm{f}_{\bp{\ABPSRcompl{s_0}{p},\ABPSRcompl{s_1}{p}}_{\theta,q}}, \end{align*} where the last inequality above is valid since $s_0\neq s_1$. Now consider $f\in\BSRcompl{s}{pq}\np{{\mathbb T}\times{\R^n}}$. Let $l\in\mathbb{N}_0$. Choose $\psi\in\CR \infty_0(\mathbb{R}\times\mathbb{R}^n)$ with $\psi(\eta,\xi)=1$ for $2^{-1}\leq\parnorm{\eta}{\xi}\leq2$ and $\supp\psi=\setc{(\eta,\xi)\in\mathbb{R}\times\mathbb{R}^n}{4^{-1}\leq\parnorm{\eta}{\xi}\leq 4}$. Using the same technique as above, this time utilizing the multiplier \begin{align*} \mathfrak{\widetilde{m}}^r_l:\mathbb{R}\times\mathbb{R}^n\rightarrow\mathbb{C},\quad\mathfrak{\widetilde{m}}^r_l(\eta,\xi)\coloneqq \psi\bp{2^{-2ml}\eta,2^{-l}\xi}\parnorm{\eta}{\xi}^{-r}, \end{align*} we can estimate \begin{align*} &\norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{\ABPSRcompl{s_1}{p}}\\ &\qquad=\norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\psi(2^{-2ml}k,2^{-l}\xi)\, \phi(2^{-2ml}k,2^{-l}\xi)\, \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{\ABPSRcompl{s_1}{p}}\\ &\qquad \leq \Ccn{C}\,2^{ls_1}\norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{p}, \end{align*} and similarly \begin{align*} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{\ABPSRcompl{s_0}{p}} \leq \Ccn{C}\,2^{ls_0}\norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{p}. \end{align*} We thus obtain \begin{align*} &2^{-l\theta(s_0-s_1)}{2^{l(s_0-s_1)} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{\ABPSRcompl{s_1}{p}}} \\ &\qquad=2^{ls} 2^{-ls_1} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{\ABPSRcompl{s_1}{p}} \\ &\qquad\leq \Ccn{C} 2^{ls} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{p} \end{align*} and \begin{align*} &2^{-l\theta(s_0-s_1)}{ \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{\ABPSRcompl{s_0}{p}}} \\ &\qquad=2^{ls} 2^{-ls_0} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{\ABPSRcompl{s_0}{p}} \\ &\qquad\leq \Ccn{C} 2^{ls} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}_{p}. \end{align*} We now employ the $J$-method (see for example \cite[Chapter 3.2]{BL76}) to characterize the interpolation space $\bp{\ABPSRcompl{s_0}{p}\np{{\mathbb T}\times{\R^n}},\ABPSRcompl{s_1}{p}\np{{\mathbb T}\times{\R^n}}}_{\theta,q}$. By the last two estimates above, we see that \begin{align*} &2^{-l\theta(s_0-s_1)}\,J\bp{2^{l(s_0-s_1)},\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f} }}\\ &\qquad\leq \Ccn{C}\,2^{ls} \norm{\mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f}}}_p. \end{align*} Since $\calp_\botf=f$, we find that $f=\sum_{l=0}^\infty \mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\bb{\phi(2^{-2ml}k,2^{-l}\xi) \mathscr{F}_{{\mathbb T}\times{\R^n}}\nb{f}}$ with convergence in the space $\ABPSRcompl{s_0}{p}\np{{\mathbb T}\times{\R^n}}+\ABPSRcompl{s_1}{p}\np{{\mathbb T}\times{\R^n}}$. Recalling that $s_0\neq s_1$, we thus conclude that \begin{align*} \norm{f}_{\bp{\ABPSRcompl{s_0}{p},\ABPSRcompl{s_1}{p}}_{\theta,q}} \leq \Ccn{C}\,\norm{f}_{\BSRcompl{s}{pq}}, \end{align*} and thereby the lemma. \end{proof} \section{Proof of Main Theorem} Utilizing the formalism introduced in Section \ref{pre}, we can equivalently reformulate \eqref{SH} in a setting where the time axis $\mathbb{R}$ is replaced with the torus ${\mathbb T}\coloneqq\mathbb{R}/\tay\mathbb{Z}$. In this setting, the periodicity condition is no longer needed and we obtain the equivalent problem \begin{align}\label{SHTorus} \begin{pdeq} \partial_tu - \Deltau + \nabla p &= f && \text{in }{\mathbb T}\times\R^n_+, \\ \Divu &= g && \text{in }{\mathbb T}\times\R^n_+, \\ u &= h && \text{on }{\mathbb T}\times\partial\R^n_+. \end{pdeq} \end{align} In order to investigate \eqref{SHTorus}, we employ the projections $\calp$ and $\calp_\bot$ to decompose the problem into a steady-state and a so-called \emph{purely oscillatory} problem. More specifically, we observe that $\np{u,p}$ is a solution to \eqref{SHTorus} if and only if $(v,\Pi)\coloneqq(\calpu,\calpp)$ is a solution to the steady-state problem \begin{align}\label{SHGS} \begin{pdeq} - \Deltav + \nabla\Pi &= \calpf && \text{in }\R^n_+, \\ \Divv &= \calpg && \text{in }\R^n_+, \\ v &= \calph && \text{on }\partial\R^n_+ \end{pdeq} \end{align} and $\np{w,\pi}\coloneqq\np{\calp_\botu,\calp_\botp}$ is a solution to \begin{align}\label{SHGP} \begin{pdeq} \partial_tw - \Deltaw + \nabla\pi &= \calp_\botf && \text{in }{\mathbb T}\times\R^n_+, \\ \Divw &= \calp_\botg && \text{in }{\mathbb T}\times\R^n_+, \\ w &= \calp_\both && \text{on }{\mathbb T}\times\partial\R^n_+. \end{pdeq} \end{align} The steady-state problem \eqref{SHGS} is a classical Stokes problem, for which a comprehensive theory is available. We therefore focus on the purely oscillatory problem \eqref{SHGP}, which only differs from \eqref{SH} by having purely oscillatory data. We start with the case of non-homogeneous boundary values. \begin{prop}\label{PurelyOscProblem_HomoBrdData} Let $q\in (1, \infty)$ and $n\geq 2$. For any vector field $H$ with \begin{align}\label{PurelyOscProblem_HomoBrdData_DataReg} \begin{aligned} &H\in\calp_\bot\WSR{1-\frac{1}{2q}}{q}\bp{{\mathbb T};\LR{q}(\wholespace)}^n\cap\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}^n,\\ &H_{n}\in\calp_\bot\WSR{1}{q}({\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)) \end{aligned} \end{align} there is a solution \begin{align}\label{PurelyOscProblem_HomoBrdData_SolReg} \begin{aligned} &u\in\calp_\bot\WSR{1}{q}\bp{{\mathbb T};\LR{q}({\halfspace})}^n\cap\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{2}{q}({\halfspace})}^n,\\ &p\in\calp_\bot\LR{q}\bp{{\mathbb T};\WSRD{1}{q}({\halfspace})} \end{aligned} \end{align} to \begin{align}\label{PurelyOscProblem_HomoBrdData_Eq} \begin{pdeq} \partial_tu - \Deltau + \nabla p &= 0 && \text{in }{\mathbb T}\times\R^n_+, \\ \Divu &= 0 && \text{in }{\mathbb T}\times\R^n_+, \\ u &= H && \text{on }{\mathbb T}\times\partial\R^n_+ \end{pdeq} \end{align} that satisfies \begin{align} \begin{aligned}\label{PurelyOscProblem_HomoBrdData_Est} &\norm{u}_{\WSR{1}{q}\np{{\mathbb T};\LR{q}({\halfspace})}\cap\LR{q}\np{{\mathbb T};\WSR{2}{q}({\halfspace})}} + \norm{\nablap}_{\LR{q}\np{{\mathbb T};\LR{q}({\halfspace})}}\\ &\qquad \leq \Ccn{C}\, \bp{\norm{H}_{\WSR{1-\frac{1}{2q}}{q}\np{{\mathbb T};\LR{q}(\wholespace)}\cap\LR{q}\np{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}} +\norm{H_n}_{\WSR{1}{q}\np{{\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)}} } \end{aligned} \end{align} with $\Ccn{C}=\Ccn{C}(n,q,\tay)$. \end{prop} \begin{proof} We shall employ the Fourier transform $\mathscr{F}_{{\mathbb T}\times\wholespace}$ to transform \eqref{PurelyOscProblem_HomoBrdData_Eq} into a system of ODEs. For this purpose, we denote by $(k,\xi)\in {\frac{2\pi}{\tay}\Z}\times\wholespace$ the coordinates in the dual group of ${\mathbb T}\times\wholespace$. Letting $v\coloneqqu^\prime \coloneqq (u_1,\ldots,u_{n-1})$, $w\coloneqqu_n$, $\ft{v}\coloneqq\mathscr{F}_{{\mathbb T}\times\wholespace}\bb{v}$, and $\ft{w}\coloneqq\mathscr{F}_{{\mathbb T}\times\wholespace}\bb{u_n}$ in \eqref{PurelyOscProblem_HomoBrdData_Eq}, we obtain an equivalent formulation of the system as a family of ODEs. More precisely, \eqref{PurelyOscProblem_HomoBrdData_Eq} is equivalent to the following ODE being satisfied for each (fixed) $(k,\xi)\in{\frac{2\pi}{\tay}\Z}\times\wholespace$: \begin{align}\label{ODE} \begin{pdeq} ik\ft{v}(x_n) + \snorm{\xi}^2\ft{v}(x_n) - \partial_{x_n}^2\ft{v}(x_n) + i\xi\ft{p}(x_n) &= 0 && \text{in }\mathbb{R}_+, \\ ik\ft{w}(x_n) + \snorm{\xi}^2\ft{w}(x_n) - \partial_{x_n}^2\ft{w}(x_n) + \partial_{x_n}\ft{p}(x_n) &= 0 && \text{in }\mathbb{R}_+, \\ i\xi\cdot\ft{v}(x_n) + \partial_{x_n}\ft{w}(x_n) &= 0 && \text{in }\mathbb{R}_+, \\ (\ft{v}(0), \ft{w}(0)) &= (\ft{\rhsh^\prime}, \ft{\rhsh_n}). \end{pdeq} \end{align} To solve the ODE, we first consider the case $k\neq 0$. Taking divergence on both sides in $\eqrefsub{PurelyOscProblem_HomoBrdData_Eq}{1}$ and utilizing that $\Divu=0$, we find that $\Deltap = 0$ and thus $-\snorm{\xi}^2\ft{p}(x_n) + \partial_{x_n}^2\ft{p}(x_n) = 0$ in $\mathbb{R}_+$. Consequently, \begin{align}\label{PurelyOscProblem_tfupresFormula} \ft{p}(x_n) = q_0(k,\xi) e^{-\snorm{\xi}x_n}\quad\text{in }\mathbb{R}_+ \end{align} for some function $q_0:{\frac{2\pi}{\tay}\Z}\times\wholespace\rightarrow\mathbb{C}$. Inserting \eqref{PurelyOscProblem_tfupresFormula} into \eqref{ODE}, we find that \begin{align*} \partial_{x_n}^2\ft{v} &= (ik + \snorm{\xi}^2)\ft{v} + i\xi q_0 e^{-\snorm{\xi}x_n}, \\ \partial_{x_n}^2\ft{w} &= (ik + \snorm{\xi}^2)\ft{w} - \snorm{\xi} q_0 e^{-\snorm{\xi}x_n}. \end{align*} Since $k\neq 0$, the resolution hereof yields \begin{align} \ft{v}(x_n) &= -\frac{\xiq_0(k, \xi)}{k} e^{-\snorm{\xi}x_n} + \alpha(k, \xi) e^{-\sqrt{\snorm{\xi}^2 + ik} \, x_n}, \\ \ft{w}(x_n) &= \frac{\snorm{\xi}q_0(k, \xi)}{ik} e^{-\snorm{\xi}x_n} + \beta(k, \xi) e^{-\sqrt{\snorm{\xi}^2 + ik} \, x_n}, \end{align} for some functions $\alpha,\beta:{\frac{2\pi}{\tay}\Z}\times\wholespace\rightarrow\mathbb{C}$. Utilizing $\eqrefsub{ODE}{3}$ and the boundary conditions $\eqrefsub{ODE}{5}$, we deduce \begin{align*} &\alpha = \ft{\rhsh^\prime} + \frac{\xiq_0}{k}, \qquad \beta = \ft{\rhsh_n} - \frac{\snorm{\xi}q_0}{ik} \end{align*} and \begin{align}\label{PurelyOscProblem_HomoBrdData_SolFormulaq0} &q_0 = -i\left(\snorm{\xi} + \sqrt{\snorm{\xi}^2 + ik}\right)\frac{\xi}{\snorm{\xi}}\cdot\ft{\rhsh^\prime} + \sqrt{\snorm{\xi}^2+ik} \, \ft{\rhsh_n} + \snorm{\xi}\ft{\rhsh_n} + \frac{ik}{\snorm{\xi}}\ft{\rhsh_n}. \end{align} By \eqref{PurelyOscProblem_tfupresFormula}--\eqref{PurelyOscProblem_HomoBrdData_SolFormulaq0} a solution to \eqref{ODE} is identified in the case $k\neq 0$. Since $H$ is purely oscillatory, we have $\ft{\rhsh^\prime}(0,\xi)=\ft{\rhsh_n}(0,\xi)=0$, whence $(v,w,p)\coloneqq(0,0,0)$ solves \eqref{ODE} in the case $k=0$. We thus obtain a formula for the solution to \eqref{PurelyOscProblem_HomoBrdData_Eq}: \begin{align}\label{PurelyOscProblem_HomoBrdData_SolFormula} \begin{aligned} v &= \mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\Bb{-\frac{\xiq_0}{k} e^{-\snorm{\xi}x_n} + \bp{\ft{\rhsh^\prime} + \frac{\xiq_0}{k}} \e^{-\sqrt{\snorm{\xi}^2 + ik} \, x_n}}, \\ w &= \mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\Bb{\frac{\snorm{\xi}q_0}{ik} e^{-\snorm{\xi}x_n} + \bp{\ft{\rhsh_n} - \frac{\snorm{\xi}q_0}{ik}} e^{-\sqrt{\snorm{\xi}^2 + ik} \, x_n}}, \\ p &= \mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\bb{q_0 e^{-\snorm{\xi}x_n}}. \end{aligned} \end{align} Formally at least, $(v,w,p)$ as defined above is a solution to \eqref{PurelyOscProblem_HomoBrdData_Eq}. It remains to show that this solution is well-defined in the class \eqref{PurelyOscProblem_HomoBrdData_SolReg} for data in the class \eqref{PurelyOscProblem_HomoBrdData_DataReg}. We start by considering data $H\in\calp_\botZ\np{{\mathbb T}\times\wholespace}^n$. The space $Z\np{{\mathbb T}\times\wholespace}$ is dense in \begin{align*} \WSR{1-\frac{1}{2q}}{q}\bp{{\mathbb T};\LR{q}(\wholespace)}\cap\LR{q}\bp{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)} \cap \WSR{1}{q}\bp{{\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)}, \end{align*} which is not a trivial assertion since it entails the construction of an approximating sequence that converges simultaneously in Sobolev spaces of positive order and in homogeneous Sobolev spaces of negative order. Nevertheless, it can be shown by a standard ``cut-off'' and mollifier technique; see \cite[proof of Theorem 2.3.3 and Theorem 5.1.5]{TriebelTheoryFunctionSpaces}. Consequently, $\calp_\botZ\np{{\mathbb T}\times\wholespace}^n$ is dense in the class \eqref{PurelyOscProblem_HomoBrdData_DataReg}. Clearly, for purely oscillatory data $H$ the solution given by \eqref{PurelyOscProblem_HomoBrdData_SolFormula} is also purely oscillatory. If we can therefore show \eqref{PurelyOscProblem_HomoBrdData_Est} for arbitrary $H\in\calp_\botZ\np{{\mathbb T}\times\wholespace}^n$, the claim of the proposition will follow by a density argument. We first examine the pressure term $p$ (more specifically $\nablap$). The terms in \eqref{PurelyOscProblem_HomoBrdData_SolFormulaq0} have different order of regularity, so we decompose $q_0=q_1+q_2$ by \begin{align}\label{PurelyOscProblem_HomoBrdData_DefOfqdfs} \begin{aligned} &q_1(k,\xi) \coloneqq -i\left(\snorm{\xi} + \sqrt{\snorm{\xi}^2 + ik}\right)\frac{\xi}{\snorm{\xi}}\cdot\ft{\rhsh^\prime} + \sqrt{\snorm{\xi}^2+ik} \, \ft{\rhsh_n} + \snorm{\xi}\ft{\rhsh_n},\\ &q_2(k,\xi) \coloneqq \frac{ik}{\snorm{\xi}}\ft{\rhsh_n}, \end{aligned} \end{align} and introduce the operators \begin{align}\label{PurelyOscProblem_HomoBrdData_DefOfGoodOpr} \begin{aligned} &\mathscr G: Z\np{{\mathbb T}\times\wholespace}^n \rightarrow \mathscr{S}\np{{\mathbb T}\times{\halfspace}}^n, \\ &\mathscr G(H)\coloneqq \mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\Bb{\xiq_1(k,\xi)e^{-\snorm{\xi}x_n}} \end{aligned} \end{align} and \begin{align}\label{PurelyOscProblem_HomoBrdData_DefOfBadOpr} \begin{aligned} &\mathscr B: Z\np{{\mathbb T}\times\wholespace} \rightarrow \mathscr{S}\np{{\mathbb T}\times{\halfspace}}^n, \\ &\mathscr B(\rhsh_n)\coloneqq \mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\Bb{\xiq_2(k,\xi)e^{-\snorm{\xi}x_n}}. \end{aligned} \end{align} For $m\in\mathbb{N}_0$, we observe for any $x_n>0$ that the symbol $\xi\rightarrow\np{\snorm{\xi}x_n}^m\e^{-\snorm{\xi}x_n}$ is an $\LR{q}\np{\wholespace}$-multiplier. Specifically, one verifies that \begin{align*} \sup_{x_n>0}\sup_{\varepsilon\in\{0, 1\}^{n-1}}\sup_{\xi\in\wholespace}\left|\xi_1^{\varepsilon_1}\cdots\xi_{n-1}^{\varepsilon_{n-1}}\partial_{\xi_1}^{\varepsilon_1}\cdots\partial_{\xi_{n-1}}^{\varepsilon_{n-1}}\bb{\np{\snorm{\xi}x_n}^m\e^{-\snorm{\xi}x_n}} \right|<\infty, \end{align*} whence it follows from the Marcinkiewicz Multiplier Theorem (see for example \cite[Corollary 6.2.5]{Grafakos}) that the Fourier-multiplier operator with symbol $\xi\rightarrow\np{\snorm{\xi}x_n}^m\e^{-\snorm{\xi}x_n}$ is a bounded operator on $\LR{q}\np\wholespace$ with operator norm independent on $x_n$, that is, \begin{align}\label{PurelyOscProblem_HomoBrdData_MultiplierOprNormIndep} \sup_{x_n>0}\,\normL{\phi\mapsto\mathscr{F}_{\wholespace}\Bb{\np{\snorm{\xi}x_n}^m\e^{-\snorm{\xi}x_n}\mathscr{F}_{\wholespace}\nb{\phi}}}_{\mathscr{L}\np{\LR{q}\np{\wholespace},\LR{q}\np{\wholespace}}}<\infty. \end{align} We can thus estimate \begin{align*} &\norm{\mathscr G(H)}_{\LR{\infty}\np{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}} \leq \Ccn{C}\, \norm{\mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\bb{\xiq_1(k,\xi)}}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}\\ &\qquad\leq \Ccn{C}\, \Bp{\norm{H}_{\LR{q}\np{{\mathbb T};\HSR{2}{q}\np\wholespace}} + \normL{\mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\bb{M(k,\xi)\,{\np{\snorm{\xi}^2+ik}\, \frac{\xi\otimes\xi}{\snorm{\xi}^2}\ft{\rhsh^\prime}}}}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}\\ &\qquad\qquad\ +\normL{\mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\bb{M(k,\xi)\,{\np{\snorm{\xi}^2+ik}\, \frac{\xi}{\snorm{\xi}}\ft{\rhsh_n}}}}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}} }, \end{align*} where \begin{align*} M:\mathbb{R}\times\wholespace\rightarrow\mathbb{C},\quad M(\eta,\xi)\coloneqq \frac{\snorm{\xi}}{\sqrt{\snorm{\xi}^2+i\eta}}. \end{align*} Employing again the Marcinkiewicz Multiplier Theorem, we find that the symbol $M$ is an $\LR{q}\np{\mathbb{R};\LR{q}(\wholespace)}$-multiplier. An application of the Transference Principle (Theorem \ref{transference}) therefore implies that the restriction $M_{|\Zgrp\times\ws}$ is an $\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}$-multiplier. We thus conclude \begin{align}\label{PurelyOscProblem_HomoBrdData_GoodOprInterpolationEst1} \begin{aligned} &\norm{\mathscr G(H)}_{\LR{\infty}\np{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}} \leq \Ccn{C}\, {\norm{H}_{\LR{q}\np{{\mathbb T};\HSR{2}{q}\np\wholespace}\cap\HSR{1}{q}\np{{\mathbb T};\LR{q}\np{\wholespace}}}}. \end{aligned} \end{align} This estimate shall serve as an interpolation endpoint. To obtain the opposite endpoint, we again employ \eqref{PurelyOscProblem_HomoBrdData_MultiplierOprNormIndep} to estimate \begin{align*} \sup_{x_n>0}\,\norm{x_n \mathscr G\np{H}}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}\leq \Ccn{C}\,\norm{q_1}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}, \end{align*} which implies \begin{align*} \norm{\mathscr G\np{H}}_{\LR{1,\infty}\np{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}} &= \normL{\frac{1}{x_n} \norm{x_n \mathscr G\np{H}}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}}_{\LR{1,\infty}\np{\mathbb{R}_+}} \\ &\leq \Ccn{C}\,\norm{q_1}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}. \end{align*} Recalling \eqref{PurelyOscProblem_HomoBrdData_DefOfqdfs}, we estimate \begin{align*} &\norm{q_1}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}\\ &\qquad \leq \Ccn{C}\,\Bp{\norm{H}_{\LR{q}\np{{\mathbb T};\HSR{1}{q}\np\wholespace}} +\normL{\mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\bb{\MmultiplierNr{1}(k,\xi)\cdot{\np{\snorm{\xi}+\snorm{k}^\frac{1}{2}}\ft{\rhsh^\prime}}}}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}\\ &\qquad\qquad\quad +\normL{\mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\bb{\MmultiplierNr{2}(k,\xi)\,{\np{\snorm{\xi}+\snorm{k}^\frac{1}{2}}\ft{\rhsh_n}}}}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}} \end{align*} with \begin{align*} &\MmultiplierNr{1}:\mathbb{R}\times\wholespace\rightarrow\mathbb{C}^{n-1},\quad \MmultiplierNr{1}(\eta,\xi)\coloneqq \frac{\sqrt{\snorm{\xi}^2+i\eta}}{\snorm{\xi}+\snorm{\eta}^{\frac{1}{2}}}\frac{\xi}{\snorm{\xi}},\\ &\MmultiplierNr{2}:\mathbb{R}\times\wholespace\rightarrow\mathbb{C},\quad \MmultiplierNr{2}(\eta,\xi)\coloneqq \frac{\sqrt{\snorm{\xi}^2+i\eta}}{{\snorm{\xi}+\snorm{\eta}^{\frac{1}{2}}}}. \end{align*} Again, one can utilize the Marcinkiewicz Multiplier Theorem to show that both $\MmultiplierNr{1}$ and $\MmultiplierNr{2}$ are $\LR{q}\np{\mathbb{R};\LR{q}(\wholespace)}$-multipliers, and subsequently obtain from the Transference Principle (Theorem \ref{transference}) that their restrictions to ${\frac{2\pi}{\tay}\Z}\times\wholespace$ are $\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}$-multipliers. Consequently, we find that \begin{align}\label{PurelyOscProblem_HomoBrdData_GoodOprInterpolationEst2} &\norm{\mathscr G\np{H}}_{\LR{1,\infty}\np{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}} \leq \Ccn{C}\, {\norm{H}_{\LR{q}\np{{\mathbb T};\HSR{1}{q}\np\wholespace}\cap\HSR{\frac{1}{2}}{q}\np{{\mathbb T};\LR{q}\np{\wholespace}}}}. \end{align} By \eqref{PurelyOscProblem_HomoBrdData_GoodOprInterpolationEst1} and \eqref{PurelyOscProblem_HomoBrdData_GoodOprInterpolationEst2}, the operator $\mathscr G$ extends uniquely to a bounded operator \begin{align}\label{PurelyOscProblem_HomoBrdData_GoodOprInterpolationPoles1} \begin{aligned} &\mathscr G:\LR{q}\bp{{\mathbb T};\HSR{2}{q}\np\wholespace}^n\cap\HSR{1}{q}\bp{{\mathbb T};\LR{q}\np{\wholespace}}^n\rightarrow\LR{\infty}\bp{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}^n,\\ &\mathscr G:\LR{q}\bp{{\mathbb T};\HSR{1}{q}\np\wholespace}^n\cap\HSR{\frac{1}{2}}{q}\bp{{\mathbb T};\LR{q}\np{\wholespace}}^n\rightarrow\LR{1,\infty}\bp{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}^n. \end{aligned} \end{align} These extensions rely on the fact that $Z\np{{\mathbb T}\times\wholespace}$ is dense in the function spaces on the left-hand side above. We once more refer to \cite[proof of Theorem 2.3.3 and Theorem 5.1.5]{TriebelTheoryFunctionSpaces} for a verification of this fact. Using the projection $\calp_\bot$ on the left-hand side in \eqref{PurelyOscProblem_HomoBrdData_GoodOprInterpolationPoles1}, we obtain scales of the anisotropic Bessel-Potential spaces introduced in \eqref{BS_DefOfBesovSpace_DefnBessel}. Consequently, $\mathscr G$ is a bounded operator: \begin{align}\label{PurelyOscProblem_HomoBrdData_GoodOprInterpolationPoles2} \begin{aligned} &\mathscr G:\ABPSRcompl{2}{q}\np{{\mathbb T}\times\wholespace}^n \rightarrow\LR{\infty}\bp{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}^n,\\ &\mathscr G:\ABPSRcompl{1}{q}\np{{\mathbb T}\times\wholespace}^n \rightarrow\LR{1,\infty}\bp{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}^n. \end{aligned} \end{align} Utilizing Lemma \ref{BS_InterpolationLem}, we find that \begin{align*} &\Bp{\ABPSRcompl{1}{q}\np{{\mathbb T}\times\wholespace},\ABPSRcompl{2}{q}\np{{\mathbb T}\times\wholespace}}_{1-\frac{1}{q},q}=\BSRcompl{2-\frac{1}{q}}{qq}\np{{\mathbb T}\times\wholespace}\\ &\quad=\Bp{\calp_\bot\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)},\ABPSRcompl{2}{q}\np{{\mathbb T}\times\wholespace}}_{1-\frac{1}{2q},q}\\ &\quad=\Bp{\calp_\bot\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}, \calp_\bot\LR{q}\bp{{\mathbb T};\HSR{2}{q}\np\wholespace}\cap\calp_\bot\HSR{1}{q}\bp{{\mathbb T};\LR{q}\np{\wholespace}}}_{1-\frac{1}{2q},q}\\ &\quad=\calp_\bot\Bp{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}, \LR{q}\bp{{\mathbb T};\HSR{2}{q}\np\wholespace}}_{1-\frac{1}{2q},q}\\ &\qquad\qquad \cap \calp_\bot\Bp{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)},\HSR{1}{q}\bp{{\mathbb T};\LR{q}\np{\wholespace}}}_{1-\frac{1}{2q},q}\\ &\quad=\calp_\bot\WSR{1-\frac{1}{2q}}{q}\bp{{\mathbb T};\LR{q}(\wholespace)}\cap\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}. \end{align*} One can employ \cite[Theorem 1.12.1]{TriebelInterpolation} to verify the interpolation of the intersection space in the fourth equality above. Moreover, real interpolation yields \begin{align*} \Bp{ \LR{1, \infty}\bp{\mathbb{R}_+; \LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}, &\, \LR{\infty}\bp{\mathbb{R}_+; \LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}}_{1-\frac{1}{q}, q} \\ &\qquad\qquad\qquad\qquad\qquad = \LR{q}\bp{\mathbb{R}_+; \LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}. \end{align*} Recalling \eqref{PurelyOscProblem_HomoBrdData_GoodOprInterpolationPoles2}, we conclude that $\mathscr G$ extends uniquely to a bounded operator \begin{align}\label{PurelyOscProblem_HomoBrdData_GoodOprFinalMappingProperty} \mathscr G: \calp_\bot\WSR{1-\frac{1}{2q}}{q}\bp{{\mathbb T};\LR{q}(\wholespace)}^n\cap\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}^n \rightarrow \LR{q}\bp{{\mathbb T};\LR{q}({\halfspace})}^n. \end{align} We now recall \eqref{PurelyOscProblem_HomoBrdData_DefOfBadOpr} and examine the operator $\mathscr B$. Utilizing \eqref{PurelyOscProblem_HomoBrdData_MultiplierOprNormIndep} with $m=0$, we obtain \begin{align}\label{PurelyOscProblem_HomoBrdData_BadOprInterpolationPoles1} \begin{aligned} \norm{\mathscr B(\rhsh_n)}_{\LR{\infty}\np{\mathbb{R}_+;\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}} &\leq \Ccn{C}\, \norm{\mathscr{F}^{-1}_{{\mathbb T}\times\wholespace}\bb{\xiq_2\np{k,\xi}}}_{\LR{q}\np{{\mathbb T};\LR{q}(\wholespace)}}\\ &\leq \Ccn{C}\, \norm{\rhsh_n}_{\HSR{1}{q}\np{{\mathbb T};\LR{q}\np{\wholespace}}}. \end{aligned} \end{align} We again employ \eqref{PurelyOscProblem_HomoBrdData_MultiplierOprNormIndep} to estimate \begin{align*} \sup_{x_n>0}\,\norm{x_n \mathscr B\np{\rhsh_n}}_{\LR{q}\np{{\mathbb T}; \LR{q}(\wholespace)}}\leq \Ccn{C}\,\norm{\rhsh_n}_{\HSR{1}{q}\np{{\mathbb T};\HSRD{-1}{q}\np{\wholespace}}}, \end{align*} which implies \begin{align*} \norm{\mathscr B\np{\rhsh_n}}_{\LR{1,\infty}\np{\mathbb{R}_+;\LR{q}\np{{\mathbb T}; \LR{q}(\wholespace)}}} &= \normL{\frac{1}{x_n} \norm{x_n \mathscr B\np{\rhsh_n}}_{\LR{q}\np{{\mathbb T}; \LR{q}(\wholespace)}}}_{\LR{1,\infty}\np{\mathbb{R}_+}}\\ &\leq \Ccn{C}\,\norm{\rhsh_n}_{\HSR{1}{q}\np{{\mathbb T};\HSRD{-1}{q}\np{\wholespace}}}. \end{align*} It follows that $\mathscr B$ extends to a bounded operator \begin{align*} &\mathscr B: \calp_\bot\HSR{1}{q}\bp{{\mathbb T};\LR{q}\np{\wholespace}} \rightarrow\LR{\infty}\bp{\mathbb{R}_+;\LR{q}\np{{\mathbb T}; \LR{q}(\wholespace)}}^n,\\ &\mathscr B: \calp_\bot\HSR{1}{q}\bp{{\mathbb T};\HSRD{-1}{q}\np{\wholespace}}\rightarrow\LR{1,\infty}\bp{\mathbb{R}_+;\LR{q}\np{{\mathbb T}; \LR{q}(\wholespace)}}.^n \end{align*} Real interpolation thus implies that $\mathscr B$ extends to a bounded operator \begin{align}\label{PurelyOscProblem_HomoBrdData_BadOprFinalMappingProperty} \mathscr B: \calp_\bot\WSR{1}{q}\bp{{\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)}\rightarrow \LR{q}\bp{{\mathbb T}; \LR{q}({\halfspace})}^n. \end{align} We now return to the solution formulas \eqref{PurelyOscProblem_HomoBrdData_SolFormula} and consider $H\in\calp_\botZ\np{{\mathbb T}\times\wholespace}^n$. In this case, an application of \eqref{PurelyOscProblem_HomoBrdData_MultiplierOprNormIndep} ensures that $p$ is well-defined as an element in the function space $\LR{q}\bp{{\mathbb T};\HSRD{1}{q}\np{\halfspace}}$. By \eqref{PurelyOscProblem_HomoBrdData_GoodOprFinalMappingProperty} and \eqref{PurelyOscProblem_HomoBrdData_BadOprFinalMappingProperty}, we obtain \begin{align}\label{PurelyOscProblem_HomoBrdData_EstimateForGradPresure} \begin{aligned} &\norm{\nablap}_{\LR{q}\np{{\mathbb T}; \LR{q}({\halfspace})}} = \norm{\mathscr G\np{H}+\mathscr B\np{\rhsh_n}}_{\LR{q}\np{{\mathbb T}; \LR{q}({\halfspace})}}\\ &\quad \leq \Ccn{C} \bp{\norm{H}_{\WSR{1-\frac{1}{2q}}{q}\np{{\mathbb T};\LR{q}(\wholespace)}\cap\LR{q}\np{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}} +\norm{H_n}_{\WSR{1}{q}({\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)}} \end{aligned} \end{align} In a similar manner, it can be shown that $(v,w)$ is well-defined as an element in the space $\LR{q}\bp{{\mathbb T};\HSR{2}{q}\np{{\halfspace}}}\cap\HSR{1}{q}\bp{{\mathbb T};\LR{q}\np{{\halfspace}}}$. To this end, one may consider the symbol \begin{align*} m:\mathbb{R}\times\wholespace\rightarrow\mathbb{C},\quad m\np{\eta,\xi}\coloneqq \bp{\sqrt{\snorm{\xi}^2+i\eta}\,x_n}^m \e^{-\sqrt{\snorm{\xi}^2 + i\eta} \, x_n} \end{align*} and verify that \begin{align*} \sup_{x_n>0}\sup_{\varepsilon\in\{0, 1\}^n}\sup_{(\eta,\xi)\in\mathbb{R}\times\wholespace}\left|\eta^{\varepsilon_0}\xi_1^{\varepsilon_1}\cdots\xi_{n-1}^{\varepsilon_{n-1}}\partial_{\eta}^{\varepsilon_0}\partial_{\xi_1}^{\varepsilon_1}\cdots\partial_{\xi_{n-1}}^{\varepsilon_{n-1}} m(\eta,\xi) \right|<\infty. \end{align*} It follows that the Fourier-multiplier operator corresponding to the symbol $m$ is a bounded operator on $\LR{q}\np{\mathbb{R}; \LR{q}(\wholespace)}$ with operator norm independent on $x_n$. An application of the Transference Principle (Theorem \ref{transference}) therefore implies that the operator corresponding to the symbol $m_{|{\frac{2\pi}{\tay}\Z}\times\wholespace}$ is a bounded operator on $\LR{q}\np{{\mathbb T}; \LR{q}(\wholespace)}$ with operator norm independent on $x_n$, that is, \begin{align}\label{PurelyOscProblem_HomoBrdData_MultiplierOprNormIndep2} \sup_{x_n>0}\,\normL{\phi\mapsto\mathscr{F}_{{\mathbb T}\times\wholespace}\bb{m(k,\xi)\mathscr{F}_{{\mathbb T}\times\wholespace}\nb{\phi}}}_{\mathscr{L}\np{\LR{q}\np{{\mathbb T}; \LR{q}(\wholespace)},\LR{q}\np{{\mathbb T}; \LR{q}(\wholespace)}}}<\infty. \end{align} With both \eqref{PurelyOscProblem_HomoBrdData_MultiplierOprNormIndep} and \eqref{PurelyOscProblem_HomoBrdData_MultiplierOprNormIndep2} at our disposal, it is now straightforward to verify that $u\coloneqq(v,w)$ is well-defined as element in the space $\LR{q}\bp{{\mathbb T};\HSR{2}{q}\np{{\halfspace}}}\cap\HSR{1}{q}\bp{{\mathbb T};\LR{q}\np{{\halfspace}}}$. By construction, this choice of $(u,p)$ is a solution to \eqref{PurelyOscProblem_HomoBrdData_Eq}. Moreover, since $\calp_\botH=H$, also $\calp_\botu=u$. This means that $u$ is a purely oscillatory solution in the aforementioned function space to the time-periodic heat equation in the half-space \begin{align*} \begin{pdeq} \partial_tu - \Deltau &= -\nablap && \text{in }{\mathbb T}\times\R^n_+, \\ u &= H && \text{on }{\mathbb T}\times\partial{\halfspace}. \end{pdeq} \end{align*} By \cite[Theorem 2.1]{KyedSauer_Heat} (see also \cite[Theorem 1.3]{KyedSauer_ADN1}), it is known that this problem has a unique purely oscillatory solution in the space $\LR{q}\bp{{\mathbb T};\HSR{2}{q}\np{{\halfspace}}}\cap\HSR{1}{q}\bp{{\mathbb T};\LR{q}\np{{\halfspace}}}$, which satisfies \begin{align*} &\norm{u}_{\HSR{1}{q}\np{{\mathbb T};\LR{q}({\halfspace})}\cap\LR{q}\np{{\mathbb T};\HSR{2}{q}({\halfspace})}}\\ &\qquad \leq \Ccn{C} \bp{\norm{\nablap}_{\LR{q}\np{{\mathbb T}; \LR{q}({\halfspace})}} + \norm{H}_{\WSR{1-\frac{1}{2q}}{q}\np{{\mathbb T};\LR{q}(\wholespace)}\cap\LR{q}\np{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}}}. \end{align*} Combining this estimate with \eqref{PurelyOscProblem_HomoBrdData_EstimateForGradPresure}, we conclude \eqref{PurelyOscProblem_HomoBrdData_Est}. \end{proof} In the next step, we consider the resolution of the fully non-homogeneous system \eqref{SHGP}, that is, \eqref{SH} with purely oscillatory data, and establish $\LR{q}$ estimates. This step concludes the main result of the article. \begin{thm}\label{PurelyOscProblem_HomoDataThm} Let $q\in(1,\infty)$ and $n\geq 2$. For all \begin{align} \begin{aligned} &f\in\calp_\bot\LR{q}\bp{{\mathbb T}; \LR{q}({\halfspace})}^n,\\ &g\in\calp_\bot\LR{q}\bp{{\mathbb T}; \WSR{1}{q}(\R^n_+)}\cap\calp_\bot\WSR{1}{q}\bp{{\mathbb T}; \WSRD{-1}{q}(\R^n_+)},\\ &h\in\calp_\bot\WSR{1-\frac{1}{2q}}{q}\bp{{\mathbb T};\LR{q}(\wholespace)}^n\cap\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}^n \end{aligned} \end{align} with \begin{align} &h_{n}\in\calp_\bot\WSR{1}{q}\bp{{\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)} \end{align} there is a solution $(u,p)$ to \eqref{SH} with \begin{align} \begin{aligned}\label{PurelyOscProblem_HomoDataThm_SolReg} &u\in\calp_\bot\WSR{1}{q}\bp{{\mathbb T};\LR{q}({\halfspace})}^n\cap\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{2}{q}({\halfspace})}^n,\\ &p\in\calp_\bot\LR{q}\bp{{\mathbb T};\WSRD{1}{q}({\halfspace})}, \end{aligned} \end{align} which satisfies \begin{align}\label{PurelyOscProblem_HomoDataThm_ProjComplEst} \begin{aligned} &\norm{u}_{\WSR{1}{q}\np{{\mathbb T};\LR{q}({\halfspace})}\cap\LR{q}\np{{\mathbb T};\WSR{2}{q}({\halfspace})}} + \norm{\nablap}_{\LR{q}\np{{\mathbb T}; \LR{q}({\halfspace})}} \\ &\qquad \leq \Ccn{C}\, \bp{\norm{f}_{\LR{q}\np{{\mathbb T}; \LR{q}({\halfspace})}}+\norm{g}_{\LR{q}\np{{\mathbb T}; \WSR{1}{q}(\R^n_+)}\cap\WSR{1}{q}\np{{\mathbb T}; \WSRD{-1}{q}(\R^n_+)}}\\ &\qquad\qquad +\norm{h}_{\WSR{1-\frac{1}{2q}}{q}\np{{\mathbb T};\LR{q}(\wholespace)}\cap\LR{q}\np{{\mathbb T};\WSR{2-\frac{1}{q}}{q}(\wholespace)}} + \norm{h_n}_{\WSR{1}{q}\np{{\mathbb T}; \WSRD{-\frac{1}{q}}{q}(\wholespace)}}\, } \end{aligned} \end{align} with $\Ccn{C}=\Ccn{C}(n,q,\tay)$. Moreover, if $(\tilde{u},\tilde{p})$ is another solution to \eqref{SH} in the class \eqref{PurelyOscProblem_HomoDataThm_SolReg}, then $u=\tilde{u}$ and $p=\tilde{p} + d(t)$ for some function $d$ that depends only on time. \end{thm} \begin{proof} Let $v\in\calp_\bot\WSR{1}{q}\bp{{\mathbb T};\LR{q}({\halfspace})}^n\cap\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{2}{q}({\halfspace})}^n$ be the solution to the purely oscillatory time-periodic n-dimensional heat equation \begin{align*} \begin{pdeq} \partial_tv - \Deltav &= f && \text{in }{\mathbb T}\times\R^n_+, \\ v &= 0 && \text{on }{\mathbb T}\times\partial{\halfspace}. \end{pdeq} \end{align*} The existence of such a solution $v$ that satisfies \begin{align*} \norm{v}_{\WSR{1}{q}\np{{\mathbb T};\LR{q}({\halfspace})}\cap\LR{q}\np{{\mathbb T};\WSR{2}{q}({\halfspace})}} \leq \Ccn{C} \norm{f}_{\LR{q}\np{{\mathbb T}; \LR{q}({\halfspace})}} \end{align*} follows from \cite[Theorem 2.1]{KyedSauer_Heat}. Denote by $G$ the extension of $g-\Divv$ to ${{\mathbb T}\times{\R^n}}$ by even reflection in the $x_n$ variable. Then $G\in\calp_\bot\LR{q}\bp{{\mathbb T};\WSR{1}{q}\np{{\R^n}}}$. Moreover, identifying $\WSRD{-1}{q}\np{{\R^n}}$ as the dual of $\WSRD{1}{q'}\np{{\R^n}}$ and recalling that $g\in\calp_\bot\WSR{1}{q}\bp{{\mathbb T}; \WSRD{-1}{q}(\R^n_+)}$, one directly verifies that $G\in\calp_\bot\WSR{1}{q}\bp{{\mathbb T};\WSRD{-1}{q}\np{{\R^n}}}$ with \begin{multline*} \norm{G}_{\WSR{1}{q}\np{{\mathbb T};\WSRD{-1}{q}\np{{\R^n}}}\cap\LR{q}\np{{\mathbb T};\WSR{1}{q}\np{{\R^n}}}}\\ \leq \Ccn{C}\bp{\norm{g}_{\WSR{1}{q}\np{{\mathbb T};\WSRD{-1}{q}\np{{\halfspace}}}\cap\LR{q}\np{{\mathbb T};\WSR{1}{q}\np{{\halfspace}}}}+ \norm{v}_{\WSR{1}{q}\np{{\mathbb T};\LR{q}({\halfspace})}\cap\LR{q}\np{{\mathbb T};\WSR{2}{q}({\halfspace})}}}. \end{multline*} A solution to the purely oscillatory Stokes system \begin{align}\label{PurelyOscProblem_HomoDataThm_StokesRnReduction} \begin{pdeq} \partial_tw-\Deltaw + \nabla\pi &=0 && \text{in }{{\mathbb T}\times{\R^n}},\\ \Divw &= G &&\text{in }{{\mathbb T}\times{\R^n}} \end{pdeq} \end{align} is obtained via the solution formulas \begin{align*} w \coloneqq \mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\Bb{\frac{-i\xi}{\snorm{\xi}^2}\,\mathscr{F}_{{{\mathbb T}\times{\R^n}}}\nb{G}},\quad \pi \coloneqq \mathscr{F}^{-1}_{{{\mathbb T}\times{\R^n}}}\Bb{\frac{ik+\snorm{\xi}^2}{\snorm{\xi}^2}\,\mathscr{F}_{{{\mathbb T}\times{\R^n}}}\nb{G}}. \end{align*} From these formulas, we immediately obtain the estimate \begin{align*} \norm{w}_{\WSR{1}{q}\np{{\mathbb T};\LR{q}({\R^n})}\cap\LR{q}\np{{\mathbb T};\WSR{2}{q}({\R^n})}} &+ \norm{\nabla\pi}_{\LR{q}\np{{\mathbb T}; \LR{q}(\mathbb{R}^n)}} \\ &\qquad\qquad \leq \Ccn{C} \norm{G}_{\LR{q}\np{{\mathbb T};\WSR{1}{q}\np{{\R^n}}}\cap\WSR{1}{q}\np{{\mathbb T};\WSRD{-1}{q}\np{{\R^n}}}}. \end{align*} By the symmetry of $G$, the vector field $\widetilde{w}$ obtained by odd reflection with respect to $x_n$ of the n'th component of $w$, that is, \begin{align*} \widetilde{w}\np{t,x',x_n}\coloneqq \bp{w_1\np{t,x',x_n},\ldots,w_{n-1}\np{t,x',x_n},-w_n\np{t,x',-x_n}}, \end{align*} is also a solution to \eqref{PurelyOscProblem_HomoDataThm_StokesRnReduction} corresponding to the same pressure term $\pi$. This means that $w$ and $\widetilde{w}$ both solve the same time-periodic heat equation in the whole-space ${{\mathbb T}\times{\R^n}}$. By \cite[Theorem 2.1]{KyedSauer_Heat}, $w=\widetilde{w}$. It follows that $\trace_{{\mathbb T}\times\wholespace}\nb{w_n}=0$. Consequently, $H\coloneqqh-\trace_{{\mathbb T}\times\wholespace}\nb{w}$ belongs to the space \eqref{PurelyOscProblem_HomoBrdData_DataReg} (see for example \cite{KyedSauer_ADN1} for a rigorous definition of the trace operator in this setting). Let $(U,\mathfrak{P})$ be the corresponding solution from Proposition \ref{PurelyOscProblem_HomoBrdData}. It follows that $\np{u,p}\coloneqq\np{U+w+v,\mathfrak{P}+\pi}$ is a solution to \eqref{SH} in the class \eqref{PurelyOscProblem_HomoDataThm_SolReg} satisfying \eqref{PurelyOscProblem_HomoDataThm_ProjComplEst}. It remains to show uniqueness, which follows from a standard duality argument. To this end, assume that $(\tilde{u},\tilde{p})$ is a solution in the class \eqref{PurelyOscProblem_HomoDataThm_SolReg} to the homogeneous Stokes problem \begin{align*} \begin{pdeq} \partial_t\tilde{u} - \Delta\tilde{u} + \nabla \tilde{p} &= 0 && \text{in }{\mathbb T}\times\R^n_+, \\ \Div\tilde{u} &= 0 && \text{in }{\mathbb T}\times\R^n_+, \\ \tilde{u} &= 0 && \text{on }{\mathbb T}\times\partial\R^n_+. \end{pdeq} \end{align*} Let $\phi\in\CR \infty_0\np{{\mathbb T}\times{\halfspace}}^n$. With exactly the same arguments as above, one can establish existence of a solution \begin{align*} &\psi\in\calp_\bot\WSR{1}{q\prime}\bp{{\mathbb T};\LR{q\prime}({\halfspace})}^n\cap\calp_\bot\LR{q\prime}\bp{{\mathbb T};\WSR{2}{q\prime}({\halfspace})}^n,\\ &\eta\in\calp_\bot\LR{q\prime}\bp{{\mathbb T};\WSRD{1}{q\prime}({\halfspace})}, \end{align*} to the adjoint Stokes problem \begin{align*} \begin{pdeq} \partial_t\psi + \Delta\psi + \nabla \eta &= \phi && \text{in }{\mathbb T}\times\R^n_+, \\ \Div\psi &= 0 && \text{in }{\mathbb T}\times\R^n_+, \\ \psi &= 0 && \text{on }{\mathbb T}\times\partial\R^n_+, \end{pdeq} \end{align*} where $q\prime$ denotes the H\"older conjugate of $q$. Integration by parts yields \begin{align*} \int_{\mathbb T}\int_{\halfspace} \tilde{u}\cdot\phi\,{\mathrm d}x{\mathrm d}t = \int_{\mathbb T}\int_{\halfspace} \tilde{u}\cdot\bp{\partial_t\psi + \Delta\psi + \nabla \eta}\,{\mathrm d}x{\mathrm d}t= 0. \end{align*} Since this identity holds for all $\phi\in\CR \infty_0\np{{\mathbb T}\times{\halfspace}}^n$, it follows that $\tilde{u}=0$. In turn, we deduce $\nabla\tilde{p}=0$, whence $\tilde{p}\in\calp_\bot\LR{q}\np{{\mathbb T}}$, that is, $\tilde{p}$ depends only on time. \end{proof} \begin{proof}[Proof of Theorem \ref{MainThm}] Let $f,g,h$ be vector fields in the class \eqref{MainThm_Data}, with $h$ satisfying \eqref{MainThm_DataCompCond}. By \cite[Theorem IV.3.2]{Galdi}, the steady-state Stokes problem \eqref{SHGS} admits a solution $\np{v,\Pi}\in\WSRD{2}{q}\np{{\halfspace}}^n\times\WSRD{1}{q}\np{{\halfspace}}$ that satisfies \begin{multline*} \norm{\nabla^2v}_{\LR{q}({\halfspace})} + \norm{\nabla\Pi}_{\LR{q}({\halfspace})}\leq \Ccn{C} \bp{\norm{\calp f}_{\LR{q}({\halfspace})} + \norm{\calpg}_{\WSR{1}{q}({\halfspace})}+\norm{\calph}_{\WSR{2-\frac{1}{q}}{q}(\wholespace)}}. \end{multline*} By Theorem \ref{PurelyOscProblem_HomoDataThm}, the purely oscillatory Stokes problem \eqref{SHGP} admits a solution $(w,\pi)$ in the class \eqref{PurelyOscProblem_HomoDataThm_SolReg} satisfying \eqref{PurelyOscProblem_HomoDataThm_ProjComplEst}. Putting $(u,p)\coloneqq(v+w,\Pi+\pi)$, we obtain a solution to \eqref{SH} that satisfies \eqref{MainThm_ProjEst} and \eqref{MainThm_ProjComplEst}. Finally, if $(\tilde{u},\tilde{p})$ is another solution to \eqref{SH} in the class \eqref{MainThm_SolReg}, then $\calp_\botu=\calp_\bot\tilde{u}$ by Theorem \ref{PurelyOscProblem_HomoDataThm}, and $\calpu=\calp\tilde{u}+(a_1x_n,\ldots,a_{n-1}x_{n},0)$ for some vector $a\in\mathbb{R}^{n-1}$ by \cite[Theorem IV.3.2]{Galdi}. It follows that $\nablap=\nabla\tilde{p}$, and thus $p=\tilde{p} + d(t)$ for some function $d$ that depends only on time. \end{proof} \bibliographystyle{plainurl}

9424e051d75f9627d593c9881f6e8af61c427f43

\section{Introduction} Metastability is a dynamical phenomenon observed in many different contexts, such as physics, chemistry, biology, climatology, economics. Despite the variety of scientific areas, the common feature of all these situations is the existence of multiple, well-separated \emph{time scales}. On short time scales the system is in a quasi-equilibrium within a single region, while on long time scales it undergoes rapid transitions between quasi-equilibria in different regions. A rigorous description of metastability in the setting of stochastic dynamics is relatively recent, dating back to the pioneering paper \cite{CGOV}, and has experienced substantial progress in the last decades. See \cite{BL,Bo, BdH,OV} for reviews and for a list of the most important papers on this subject. One of the big challenges in rigorous study of metastability is understanding the dependence of the metastable behaviour and of the nucleation process of the stable phase on the dynamics. The nucleation process of the critical droplet, i.e. the configuration triggering the crossover, has been indeed studied in different dynamical regimes: serial (\cite{BM, CO}) vs. parallel dynamics (\cite{BCLS,CN,CNS01}); non-conservative (\cite{BM, CO}) vs. conservative dynamics (\cite{HNT,HNT1,HOS}); finite (\cite{BHN}) vs. infinite volumes (\cite{BHS}); competition (\cite{CNS02,CNS03,I,SLF}) vs. non-competition of metastable phases (\cite{CN2013,CNS2015}). All previous studies assumed that the microscopic interaction is of short-range type. The present paper pushes further this investigation, studying the dependence of the metastability scenario on the \emph{range} of the interaction of the model. Long-range Ising models in low dimensions are known to behave like higher-dimensional short-range models. For instance in \cite{Dys, Cas} (and later generalized by \cite{Picco, WBruno}) it was shown that long-range Ising models undergo a phase transition already in one dimension, and this transition persists in fast enough decaying fields. Furthermore, Dobrushin interfaces are rigid already in two dimensions for anisotropic long-range Ising models, see \cite{Loren}. We consider the question: does indeed a \emph{long-range} interaction change substantially the nucleation process? Are we able to define in this framework a critical configuration triggering the crossover towards the stable phase? In (\cite{MCAW}) the author already considered the \emph{Dyson-like} long-range models, i.e. the one-dimensional lattice model of Ising spins with interaction decaying with a power $\alpha$, in a external magnetic field. Despite the long-range potential, the author showed, by \emph{instanton} arguments, that the system has a finite-sized critical droplet. In this manuscript we want to make rigorous this claim for a general long-range interaction, showing as well that the long-range interaction completely changes the metastability scenario: in the short--range one-dimensional Ising model a droplet of size one, already nucleates the stable phase. We show instead that for a given external field $h$, and pair long range potential $J(n)$, we can define a nucleation droplet which gets larger for smaller $h$. For $d=1$ finite range interactions, inserting a minus interval of size $\ell$ in the plus phase costs a finite energy, which is uniform in the length of the interval, the same is almost true for a fast decaying interaction, as there is a uniform bound on the energy an interval costs. Thus, for low temperature, there is a diverging timescale and we will talk in case (maybe by abuse of terminology) of metastability. The spatial scale of a nucleating interval, however, defined as an interval which lowers its energy when growing, is finite for finite range interactions, but diverges as $h\to0$ for infinite range. The Dyson model has energy and spatial scale of nucleating droplet diverging as $h$ goes to zero. We will show that, depending on the value of $h$, the critical droplet can be \emph{macroscopic} or \emph{mesoscopic}. Roughly speaking, an interval of minuses of length $\ell$ which grows to $\ell+1$ gains energy $2h$, but loses $E_\ell= \sum_{ n=\ell }^{\infty} J(n)$. $E_\ell$ converges to zero as $\ell\rightarrow \infty$, but the smaller $h$ is, the larger the size of the critical droplet. Moreover, by taking $h$ volume-dependent, going to zero with $N$ as $ N^{-\delta}$, one can make the nucleation interval mesoscopic (e.g. $O(N^\delta)$, with $\delta\in(0,1)$) or macroscopic (i.e. $O(N)$). The paper is organised as follows. In Section~2 we describe the lattice model and we give the main definitions; in Section~3 the main results of the paper are stated, while in Section~4 and 5 the proofs of the model-dependent results are given. \section{The model and main definitions} Let $\Lambda$ be a finite interval of $\mathbb{Z}$, and let us denote by $h$ a positive external field. Given a configuration $\sigma$ in $\Omega_{\Lambda} = \{-1,1\}^{\Lambda}$, we define the \textit{Hamiltonian} with respect to free boundary condition by \begin{equation}\label{ham1} H_{\Lambda, h}(\sigma) = -\sum_{\{i,j\} \subseteq \Lambda} J(|i-j|)\sigma_{i}\sigma_{j} - \sum_{i \in \Lambda}h \sigma_{i}, \end{equation} where $J: \mathbb{N} \rightarrow \mathbb{R}$, the \textit{pair interaction}, is assumed to be positive and decreasing. The class of interactions that we want to include in the present analysis are of \textit{long-range type}, for instance, \begin{enumerate} \item exponential decay: $J(|i-j|)= J \cdot \lambda^{-|i-j|}$ with constants $J>0$ and $\lambda >1$; \item polynomial decay: $J(|i-j|)= J \cdot |i-j|^{-\alpha}$, where $\alpha>0$ is a parameter. \end{enumerate} The {\it finite-volume Gibbs measure} will be denoted by \begin{equation} \label{e:gibbs} \mu_{\Lambda}(\sigma) = \frac{1}{Z_{\Lambda}} \exp \left(-\beta H_{\Lambda, h}(\sigma)\right ), \end{equation} where $\beta>0$ is proportional to the inverse temperature and $Z_{\Lambda}$ is a normalizing constant. The set of \emph{ground states} $\mathscr{X}^{s}$ is defined as $\mathscr{X}^{s}: =\textnormal{argmin}_{\sigma\in \Omega_\Lambda} H_{\Lambda, h}(\sigma)$. Note that for the class of interactions considered $\mathscr{X}^{s} = \{\mathbf{+1}\}$, where $\mathbf{+1}$ {stands for} the configuration with all spins equal to $+1$. \\ Given an integer $k \in \{0, \dots, \#\Lambda\}$, we consider $\mathcal{M}_k:=\{\sigma\in \Omega_{\Lambda}: \# \{i: \sigma_{i} = 1\} = k\}$ consisting of configurations in $\Omega_{\Lambda}$ with $k$ positive spins, and we define the configurations $L^{(k)}$ and $R^{(k)}$ as follows. Let \begin{equation} L_{i}^{(k)} = \begin{cases} +1 & \text{if $1 \leq i \leq k $, and}\\ -1 & \text{otherwise,} \end{cases} \end{equation} and \begin{equation} R_{i}^{(k)} = \begin{cases} -1 & \text{if $1 \leq i \leq \# \Lambda-k$, and}\\ +1 & \text{otherwise,} \end{cases} \end{equation} i.e., the configurations respectively with $k$ positive spins on {\it left} side of the interval and on the {\it right} one. We will show that $L^{(k)}$ and $R^{(k)}$ are the {minimizers} of the energy function $H_{\Lambda,h}$ on $\mathcal{M}_k$ (see {Proposition}~\ref{csgo2}). Let us denote {by $\mathcal{P}^{(k)}$ the set $\mathcal{P}^{(k)}:= \{L^{(k)}, R^{(k)}\}$ consisting} of the {minimizers} of the energy on $\mathcal{M}_k$. With abuse of notation we will indicate with $H_{\Lambda,h}(\mathcal{P}^{(k)})$ the energy of the elements of the set, {that is, $H_{\Lambda,h}(\mathcal{P}^{(k)}):=H_{\Lambda,h}({L}^{(k)})=H_{\Lambda, h}({R}^{(k)})$.} We {choose} the evolution of the system {to be described} by a discrete-time Markov chain {$X=(X(t))_{t \geq 0}$}, in particular, we consider the discrete-time serial Glauber dynamics given by the Metropolis weights, i.e., {the transition matrix of such dynamics is given by} $$ p(\sigma,\eta):= c(\sigma,\eta)e^{-\beta[H_{\Lambda,h}(\eta)-H_{\Lambda,h}(\sigma)]_+}, $$ where $[\cdot]_+$ denotes the positive part, and $c(\cdot,\cdot)$ is its connectivity matrix {that is equal to $1/|\Lambda|$ in case the two configurations $\sigma$ and $\eta$ coincide up to the value of a single spin, and zero otherwise.} Notice that such dynamics is reversible with respect to the Gibbs measure defined in (\ref{e:gibbs}). Let us define the \emph{hitting time} {$\tau_{\eta}^{\sigma}$} of a configuration $\eta$ of the chain $X$ started at $\sigma$ as \begin{equation} \label{e:hit} {\tau_{\eta}^{\sigma}:=\inf\{t>0: X(t)=\eta\}.} \end{equation} For any positive integer $n$, {a sequence $\gamma = (\sigma^{(1)}, \dots, \sigma^{(n)})$ such that $\sigma^{(i)}\in \Omega_\Lambda $ and $c(\sigma^{(i)},\sigma^{(i+1)})>0$} for all $i=1,\dots,n-1$ is called a \emph{path} joining {$\sigma^{(1)}$ to $\sigma^{(n)}$}; we also say that $n$ is the length of the path. For any path {$\gamma$} of length $n$, we let \begin{equation} \label{height} {\Phi_\gamma :=\max_{i=1,\dots,n} H_{\Lambda, h}(\sigma^{(i)})} \end{equation} be the \emph{height} of the path. {We also define the \emph{communication height} between $\sigma$ and $\eta$ by \begin{equation} \label{communication} \Phi(\sigma,\eta) := \min_{\gamma\in\Omega(\sigma,\eta)} \Phi_\gamma, \end{equation} where the minimum is restricted to the set $\Omega(\sigma,\eta)$ of all paths joining $\sigma$ to $\eta$.} By reversibility, it easily follows that \begin{equation} \label{rev02} \Phi(\sigma,\eta)=\Phi(\eta,\sigma) \end{equation} for all $\sigma,\eta\in \Omega_\Lambda$. {We extend the previous definition for sets $\mathcal{A},\mathcal{B}\subseteq \Omega_\Lambda$ by letting} \begin{equation} \label{communication-set} \Phi(\mathcal{A},\mathcal{B}) := {\min_{\gamma\in\Omega(\mathcal{A},\mathcal{B})}\Phi_\gamma} = \min_{\sigma\in \mathcal{A},\eta\in \mathcal{B}}\Phi(\sigma,\eta), \end{equation} where {$\Omega(\mathcal{A},\mathcal{B})$ denotes} the set of paths joining a state in $\mathcal{A}$ to a state in $\mathcal{B}$. The {\it communication cost} of passing from $\sigma$ to $\eta$ is given by the quantity $\Phi(\sigma,\eta)-H_{\Lambda, h}(\sigma)$. {Moreover, if we define $\mathscr{I}_\sigma$ as the set of all states $\eta$ in $\Omega_\Lambda$ such that $H_{\Lambda, h}(\eta)< H_{\Lambda, h}(\sigma)$, then the \emph{stability level} of any $\sigma\in \Omega_\Lambda \setminus \mathscr{X}^{s}$ is given by \begin{equation} \label{stability} V_\sigma:=\Phi(\sigma,\mathscr{I}_\sigma)-H_{\Lambda, h}(\sigma) \ge 0. \end{equation} } Following \cite{MNOS}, we now introduce the notion of \emph{maximal stability level}. {Assuming that} $\Omega_\Lambda\setminus \mathscr{X}^{s}\neq\emptyset$, we let the \emph{maximal stability level} be \begin{equation} \label{gamma} \Gamma_\textnormal{m}:=\sup_{\sigma\in \Omega_\Lambda\setminus \mathscr{X}^{s}}V_\sigma. \end{equation} We give the following definition. \begin{definition} \label{def1} We call metastable set $\mathscr{X}^{m}$, the set \begin{equation} \label{metastabile} \mathscr{X}^{m} := \{\sigma\in \Omega_\Lambda\setminus \mathscr{X}^{s}:\,V_\sigma=\Gamma_{\textnormal{m}}\}. \end{equation} \end{definition} Following \cite{MNOS}, we shall call $\mathscr{X}^{m}$ the set of \emph{metastable} states of the system {and refer to each of its elements as \emph{metastable}.} We denote {by} $\Gamma$ the quantity \begin{equation} \label{e:gamma} {\Gamma:=\max_{k=0, \dots, \#\Lambda} H_{\Lambda,h}(\mathcal{P}^{(k)})- H_{\Lambda,h}(\mathbf{-1}).} \end{equation} We will show in Corollary~\ref{t:meta} that {under certain assumptions} $\Gamma=\Gamma_\textnormal{m}$. \section{Main Results} \subsection{Mean exit time} In this section we will study the first hitting time of the configuration $\mathbf{+1}$ when the system is prepared in $\mathbf{-1}$, in the limit $\beta\to \infty$. We will restrict our analysis to the case given by the following condition. \begin{condition} \label{c:condition} {Let $N$ be an integer such that $N \geq 2$. We consider $\Lambda = \{1, \dots, N\}$ and $h$ such that \begin{equation}\label{field} 0 < h < \sum_{n = 1}^{N-1}J(n). \end{equation} } \end{condition} By using the general theory developed in \cite{MNOS}, we need first to solve two \emph{model-dependent} problems: the calculation of the \emph{minimax} between {$\mathbf{-1}$} and $\mathbf{+1}$ ({item} \ref{minmax1} of Theorem~\ref{t:minmax}) and the proof of a \emph{recurrence} property in the energy landscape ({item} \ref{minmax2} of Theorem~\ref{t:minmax}). \begin{teo} \label{t:minmax} Assume that Condition~\ref{c:condition} is satisfied.{Then, we have} \begin{enumerate} \item $\Phi(-\mathbf{1},\mathbf{+1})=\Gamma+H_{\Lambda,h}(\mathbf{-1})$, \label{minmax1} \item {$V_{\mathbf{-1}}= \Gamma > 0$}, and \label{minmax3} \item $V_\sigma<\Gamma$ for any {$\sigma\in \Omega_\Lambda\setminus \{\mathbf{-1}, \mathbf{+1}\}$.\label{minmax2}} \end{enumerate} \end{teo} As a corollary we have that $-\mathbf{1}$ is the only metastable state for this model. \begin{cor} \label{t:meta} Assume that Condition~\ref{c:condition} is satisfied. {It follows that \begin{equation} \Gamma = \Gamma_m , \end{equation} and \begin{equation} \mathscr{X}^{m}=\{-\mathbf{1}\}. \end{equation} } \end{cor} Therefore, the asymptotic of the exit time for the system started at the metastable states {is given by the following theorem.} \begin{teo} \label{t:meantime} Assume that Condition~\ref{c:condition} is satisfied. {It follows that} \begin{enumerate} \item for any $\epsilon>0$ $$ {\lim_{\beta\to\infty} \mathbb{P}\left(e^{\beta(\Gamma-\epsilon)}<\tau_{\mathbf{+1}}^{\mathbf{-1}}<e^{\beta(\Gamma+\epsilon)}\right)=1,} $$ \item the limit $$ {\lim_{\beta\to\infty}\frac{1}{\beta}\log\left(\mathbb{E}\left(\tau_{\mathbf{+1}}^{\mathbf{-1}}\right)\right)=\Gamma} $$ holds. \end{enumerate} \end{teo} Once the model-dependent results in Theorem~\ref{t:minmax} have been proven, the proof of Theorem~\ref{t:meantime} easily follows from the general theory present in \cite{MNOS}: {item} 1 follows from Theorem~4.1 in \cite{MNOS} and {item} 2 from Theorem~4.9 in \cite{MNOS}. \subsection{Nucleation of the metastable phase} We are going to show that for small enough external magnetic field, the size of the critical droplet is a macroscopic fraction of the system, while for $h$ sufficiently large, the critical configuration will be a {mesoscopic} fraction of the system. Let us define $L := \left\lfloor \frac{N}{2} \right\rfloor$, and {let $h_{k}^{(N)}$ be} \begin{equation} {h_{k}^{(N)} := \sum_{n=1}^{N-k-1} J(n) - \sum_{n=1}^{k} J(n)} \end{equation} for each $k = 0,\dots, L-1$. {One can easily verify that \begin{equation} 0 < h_{L-1}^{(N)} < \dots < h_{1}^{(N)} < h_{0}^{(N)} = \sum_{n = 1}^{N-1}J(n) \end{equation} } \begin{prop}\label{critdrop} {Under the assumption that Condition (\ref{c:condition}) is satisfied, one of the following conditions holds. \begin{enumerate} \item Case $h < h_{L-1}^{(N)}$, we have \[H_{\Lambda,h}(\mathcal{P}^{(L)}) > \max_{\substack{0 \leq k \leq N \\ k \neq L}} H_{\Lambda,h}(\mathcal{P}^{(k)}).\] \item Case $h_{k}^{(N)} < h < h_{k-1}^{(N)}$ for some $k \in \{1,\dots, L-1\}$, we have \[H_{\Lambda,h}(\mathcal{P}^{(k)}) > \max_{\substack{0 \leq i \leq N \\ i \neq \bar{k}}} H_{\Lambda,h}(\mathcal{P}^{(i)}).\] \item Case $h = h_{k}^{(N)}$ for some $k \in \{1,\dots, L-1\}$, we have \[H_{\Lambda,h}(\mathcal{P}^{(k)}) = H_{\Lambda,h}(\mathcal{P}^{(k+1)}) > \max_{\substack{0 \leq i \leq N \\ i \neq k, i \neq k+1}} H_{\Lambda,h}(\mathcal{P}^{(i)}).\] \end{enumerate} } \end{prop} The first point of Proposition~\ref{critdrop} describes the less interesting and, in a way, artificial, situation of very low external magnetic fields: in this regime the \emph{bulk} term is negligible so that the energy of the droplet increases until the positive spins are the majority (i.e. {$k=L$}, see Figure~\ref{fig:test2}). Therefore, {the second point} contains the most interesting situation, where there is an interplay between the bulk and the \emph{surface} term. The following Corollary is a consequence of Proposition~\ref{critdrop} when $N$ is large enough and gives a characterisation of the critical size $k_c$ of the critical droplet. \begin{cor}\label{corcrit} {If we assume that $\sum_{n = 1}^{\infty}J(n)$ converges and \begin{equation} 0 < h < \sum_{n = 1}^{\infty}J(n), \end{equation} then, the size of the critical droplet will be given by \begin{equation}\label{crit} k_{c} = \min\left\{k \in \mathbb{N}: \sum_{n = k+1}^{\infty}J(n) \leq h\right\} \end{equation} whenever $N$ is sufficiently large.} \end{cor} As a consequence {of Corollary \ref{corcrit}}, the set of \emph{critical configurations} $\mathcal{P}_c$ is given by { \begin{equation} \label{critical_conf} \mathcal{P}_c:= \{L^{(k_c)}, R^{(k_c)}\} \end{equation} for $N$ large enough.} {The following result shows the reason why configurations in $\mathcal{P}_c$ are referred to as \emph{critical} configurations: they indeed trigger the transition towards the stable phase.} \begin{lemma} \label{nucleation_1} Under the conditions stated above, we have \begin{enumerate} \item any path {$\gamma\in \Omega(\mathbf{-1},\mathbf{+1})$ such that $\Phi_\gamma- H_{\Lambda,h}(-\mathbf{1})=\Gamma$} visits $\mathcal{P}_c$, and \item {the limit} { $$ \lim_{\beta\to\infty} \mathbb{P}(\tau_{\mathcal{P}_c}^{-\mathbf{1}}<\tau_{+\mathbf{1}}^{-\mathbf{1}})=1 $$ holds.} \end{enumerate} \end{lemma} The proof of the previous Theorem is a straightforward consequence of Theorem~5.4 in \cite{MNOS}. \subsection{Examples} Let us give two interesting examples of the general theory so far developed. \subsubsection{Example 1: exponentially decaying coupling} We consider {$$J(n) = \frac{J}{\lambda^{n-1}},$$} where $J$ and $\lambda$ are positive real numbers with $\lambda > 1$. \begin{prop} \label{expo} {Under the same hypotheses as Corollary~\ref{corcrit}, we have that the critical droplet length $k_c$ is equal to \begin{equation} k_{c} = \left\lceil \log_{\lambda}\left(\frac{J}{h(1-\lambda^{-1})}\right) \right\rceil \end{equation} whenever N is sufficiently large.} \end{prop} \begin{proof} {By Corollary~\ref{corcrit}, we have \[ J \sum_{n = k_{c}+1}^{\infty}\lambda^{-(n-1)} \leq h < J \sum_{n = k_{c}}^{\infty}\lambda^{-(n-1)}\] that implies \[\frac{\lambda^{-k_{c}}}{1-\lambda^{-1}} \leq \frac{h}{J} < \frac{\lambda^{-(k_{c}-1)}}{1-\lambda^{-1}}\] Thus \begin{equation} k_{c} - 1 < - \frac{\log\left(\frac{h(1-\lambda^{-1})}{J}\right)}{\log \lambda} \leq k_{c}. \end{equation} } \end{proof} {As a remark we notice that in case of exponential decay of the interaction, the system behaves essentially as the nearest neighbours one-dimensional Ising model. Note that \begin{equation} \lim_{\lambda \to \infty} J(n) = \begin{cases} J &\text{if $n =1$, and}\\ 0 &\text{otherwise;} \end{cases} \end{equation} moreover, if $h < J = \lim_{\lambda \to \infty} \sum_{n =1}^{\infty}J(n)$, then $k_{c} = 1$ whenever $\lambda$ is large enough. So, we conclude that typically a single plus spin in the lattice will trigger the nucleation of the stable phase. As you can see in Figure~\ref{expo22} the energy exitations $H_{\Lambda,h}(\mathcal{P}^{(k)})-H_{\Lambda,h}(\mathbf{-1})$ are strictly descreasing in $k$, as expected.} \begin{figure} \centering \includegraphics[height=7cm]{N1000_exponential} \caption{Blue line is the excitation energy $H_{\Lambda,h}(\mathcal{P}^{(k)})-H_{\Lambda,h}(\mathbf{-1})$ for $N=1000$, $\lambda=2, h=0.21,J=1$; red line is the critical droplet.} \label{expo22} \end{figure} \subsubsection{Example 2: polynomially decaying coupling} Let the coupling constants be given by $$J(n) = J\cdot n^{-\alpha},$$ where $J$ and $\alpha$ are positive real numbers with $\alpha > 1$. As it is shown in Figures~\ref{fig:test1} and \ref{fig:test2}, for the polynomially decaying coupling model, we have that, for $h$ small enough the critical droplet is essentially the half interval, while for large enough magnetic external magnetic field, the critical droplet is the configuration with $k_c$ plus spins at the sides, with $k_c\approx \left(\frac{J}{h(\alpha -1)}\right)^{\frac{1}{\alpha -1}} $. \begin{figure} \centering \begin{minipage}{.5\textwidth} \centering \includegraphics[width=.9\linewidth]{N10000} \captionof{figure}{Blue line is the excitation energy \\ $H_{\Lambda,h}(\mathcal{P}^{(k)})-H_{\Lambda,h}(\mathbf{-1})$ for $N=10000$,\\ $\alpha=3/2, h=0.21,J=1$; {the red line represents the \\ critical length $k_c\approx 91$.}} \label{fig:test1} \end{minipage}% \begin{minipage}{.5\textwidth} \centering \includegraphics[width=.9\linewidth]{N500} \captionof{figure}{ Blue line is the excitation energy \\ $H_{\Lambda,h}(\mathcal{P}^{(k)})-H_{\Lambda,h}(\mathbf{-1})$ for $N=500$,\\ $\alpha=3/2, h=0.0001,J=1$; {the red line represents the \\critical length $k_c= 250$.}} \label{fig:test2} \end{minipage} \end{figure} We can prove indeed the following proposition. \begin{prop}\label{dyson} {Under the same hypotheses as Corollary~\ref{corcrit}, we have that $k_c$ satisfies \begin{equation} \left | k_{c} - \left(\frac{J}{h(\alpha -1)}\right)^{\frac{1}{\alpha -1}} \right |< 1 \end{equation} whenever $N$ is large enough.} \end{prop} \begin{proof} {By Corollary~\ref{corcrit}}, it follows that $$ J \sum_{n = k_{c}+1}^{\infty}n^{-\alpha} \leq h < J \sum_{n = k_{c}}^{\infty}n^{-\alpha}. $$ Moreover, note that $$ \int_{k_{c}+1}^{\infty}\frac{1}{x^{\alpha}} dx < \sum_{n = k_{c}+1}^{\infty}n^{-\alpha} $$ and \[ \sum_{n = k_{c}}^{\infty}n^{-\alpha} < \int_{k_{c}-1}^{\infty}\frac{1}{x^{\alpha}} dx \] so that \[\frac{(k_{c}+1)^{1-\alpha}}{\alpha - 1} < \frac{h}{J} < \frac{(k_{c}-1)^{1-\alpha}}{\alpha - 1}. \] Hence, \begin{equation} (k_{c} - 1)^{\alpha - 1} < \frac{J}{h(\alpha -1)} < (k_{c} + 1)^{\alpha - 1} . \end{equation} \end{proof} \section{Proof Theorem~\ref{t:minmax}} We start the proof of the main theorem giving some general results about the control of the energy of a general configuration. First of all we note that equation (\ref{ham1}) can be written as \begin{eqnarray*} H_{\Lambda, h}(\sigma) &=& -\frac{1}{2}\sum_{i \in \Lambda}\sum_{j \in \Lambda} J(|i-j|)\sigma_{i}\sigma_{j} - h\sum_{i \in \Lambda} \sigma_{i}\\ &=& \sum_{i \in \Lambda}\sum_{j \in \Lambda} J(|i-j|)\left(\frac{1 - \sigma_{i}\sigma_{j}}{2}\right) - h\sum_{i \in \Lambda} \sigma_{i} - \frac{1}{2}\sum_{i \in \Lambda}\sum_{j \in \Lambda} J(|i-j|)\\ &=& \sum_{i \in \Lambda}\sum_{j \in \Lambda} J(|i-j|) \mathds{1}_{\{\sigma_{i} \neq \sigma_{j}\}} - h\sum_{i \in \Lambda}\sigma_{i} - \frac{1}{2}\sum_{i \in \Lambda}\sum_{j \in \Lambda} J(|i-j|). \end{eqnarray*} Moreover, given an integer $k \in \{0, \dots, N\}$, if $\sigma\in\mathcal{M}_k$, then \begin{equation}\label{main} H_{\Lambda, h}(\sigma) = \sum_{i \in \Lambda}\sum_{j \in \Lambda} J(|i-j|) \mathds{1}_{\{\sigma_{i} \neq \sigma_{j}\}} + h (N - 2k) - \frac{1}{2}\sum_{i \in \Lambda}\sum_{j \in \Lambda} J(|i-j|). \end{equation} Therefore, restricting ourselves to configurations that contains only $k$ spins with the value $1$, in order to find such configurations with minimal energy, it is sufficient to minimize the first term of the right-hand side of equation (\ref{main}). \begin{prop}\label{csgo} Let $N$ be a positive integer and $k \in \{0, \dots, N\}$, if we restrict to all $\sigma\in\mathcal{M}_k$, then \begin{equation}\label{csgo1} \sum_{i=1}^{N}\sum_{j=1}^{N} J(|i-j|) \mathds{1}_{\{\sigma_{i} \neq \sigma_{j}\}} \geq 2 \sum_{i=1}^{k}\sum_{j=k+1}^{N} J(|i-j|). \end{equation} Under this restriction, the equality in the equation above holds if and only if $\sigma = L^{(k)}$ or $\sigma = {R}^{(k)}$. \end{prop} \begin{proof} Let us prove the result by induction. Let $\mathcal{H}_{N}$ be defined by \begin{equation} \mathcal{H}_{N}(\sigma_{1}, \dots,\sigma_{N}) = \sum_{i=1}^{N}\sum_{j=1}^{N} J(|i-j|) \mathds{1}_{\{\sigma_{i} \neq \sigma_{j}\}} = 2 \sum_{i :\,\sigma_{i} = 1}\sum_{j :\,\sigma_{j} = -1} J(|i-j|). \end{equation} Note that the result is trivial if $N=1$. Assuming that it holds for $N \geq 1$, let us prove that it also holds for $N+1$. In case $\sigma_{1} = 1$, applying our induction hypothesis and Lemma \ref{lero}, we have \begin{eqnarray} \mathcal{H}_{N+1}(1,\sigma_{2}, \dots,\sigma_{N+1}) &=& 2 \sum_{j = 1}^{N} J(j) \mathds{1}_{\{\sigma_{j+1} = -1\}} + \mathcal{H}_{N}(\sigma_{2}, \dots,\sigma_{N+1}) \\ \label{awp1}&\geq& 2 \sum_{j = k}^{N} J(j) + 2 \sum_{i=1}^{k-1}\sum_{j=k}^{N} J(|i-j|)\\ &=& 2 \sum_{i=1}^{k}\sum_{j=k+1}^{N+1} J(|i-j|). \end{eqnarray} Replacing the inequality sign in equation (\ref{awp1}) by an equality, it follows that \begin{equation} 0 \leq \mathcal{H}_{N}(\sigma_{2}, \dots,\sigma_{N+1}) - 2 \sum_{i=1}^{k-1}\sum_{j=k}^{N} J(|i-j|) = 2 \sum_{j = k}^{N} J(j) - 2 \sum_{j = 1}^{N} J(j) \mathds{1}_{\{\sigma_{j+1} = -1\}} \leq 0, \end{equation} hence, \begin{equation} \sum_{j = 1}^{k-1} J(j) - \sum_{j = 1}^{N} J(j) \mathds{1}_{\{\sigma_{j+1} = 1\}} = 0. \end{equation} Using Lemma \ref{lero} again, we conclude that $\sigma_{j} = 1$ whenever $1 \leq j \leq k$, and $\sigma_{j} = -1$ whenever $k+1 \leq j \leq N+1$. Now, in case $\sigma_{1} = -1$, we write $\mathcal{H}_{N+1}(-1,\sigma_{2}, \dots,\sigma_{N+1})$ as \begin{eqnarray} \mathcal{H}_{N+1}(-1,\sigma_{2}, \dots,\sigma_{N+1}) = \mathcal{H}_{N+1}(1,-\sigma_{2}, \dots,-\sigma_{N+1}) \end{eqnarray} and apply our previous result in order to obtain \begin{equation} \mathcal{H}_{N+1}(-1,\sigma_{2}, \dots,\sigma_{N+1}) \geq 2 \sum_{i=1}^{N+1-k}\sum_{j=N+2-k}^{N+1} J(|i-j|) = 2 \sum_{i=1}^{k}\sum_{j=k+1}^{N+1} J(|i-j|), \end{equation} where the equality holds only if $\sigma_{j} = -1$ whenever $1 \leq j \leq N+1-k$, and $\sigma_{j} = 1$ whenever $N+2-k \leq j \leq N+1$. \end{proof} As an immediate consequence of Proposition \ref{csgo} the next results follows. \begin{teo} \label{csgo2} Given an integer $k \in \{0, \dots, N\}$, if we restrict to all $\sigma\in\mathcal{M}_k$, then \begin{equation} H_{\Lambda, h}(\sigma) \geq 2 \sum_{i=1}^{k}\sum_{j=k+1}^{N} J(|i-j|) + h (N - 2k) - \frac{1}{2}\sum_{i=1}^{N}\sum_{j =1}^{N} J(|i-j|). \end{equation} Under this restriction, the equality in the equation above holds if and only if $\sigma = R^{(k)}$ or $\sigma = {L}^{(k)}$ \end{teo} \subsection{Proof of Theorem \ref{t:minmax}.\ref{minmax1}(minimax)} \begin{proof}[Proof of Theorem \ref{t:minmax}.\ref{minmax1}] Define $f: \{0,\dots, N\} \rightarrow \mathbb{R}$ as \begin{equation} f(k) = H_{\Lambda, h} (\mathscr{P}{^{(k)}}). \end{equation} It follows that \begin{eqnarray*} \Delta f (k) &=& f(k+1) - f(k) \\ &=& 2\left( \sum_{i=1}^{k+1} \sum_{j=k+2}^{N} J(|i-j|) - \sum_{i=1}^{k} \sum_{j=k+1}^{N} J(|i-j|) - h \right) \\ &=& 2\left( \sum_{j=k+2}^{N} J(|k+1-j|) + \sum_{i=1}^{k} \sum_{j=k+2}^{N} J(|i-j|) - \sum_{i=1}^{k} \sum_{j=k+1}^{N} J(|i-j|) - h \right) \\ &=& 2\left( \sum_{j=k+2}^{N} J(|k+1-j|) - \sum_{i=1}^{k} J(|i-(k+1)|) - h \right) \\ &=& 2\left( \sum_{i=1}^{N-k-1} J(i) - \sum_{i=1}^{k} J(i) - h \right) \end{eqnarray*} holds for all $k$ such that $0 \leq k \leq N-1$, and \begin{eqnarray*} \Delta^{2} f (k) &=& \Delta f(k+1) - \Delta f(k) \\ &=& 2\left( \sum_{i=1}^{N-k-2} J(i) - \sum_{i=1}^{N-k-1} J(i) - \sum_{i=1}^{k+1} J(i) + \sum_{i=1}^{k} J(i) \right) \\ &=& -2(J(N-k-1) + J(k+1)) \end{eqnarray*} holds whenever $0 \leq k \leq N-2$. Note that \begin{equation}\label{omg1} \Delta f (0) = 2\left( \sum_{i=1}^{N-1} J(i) - h \right) > 0, \end{equation} $1 \leq \left\lfloor \frac{N}{2} \right\rfloor\leq N-1$, and \begin{equation}\label{omg2} \Delta f \left(\left\lfloor \frac{N}{2} \right\rfloor\right) < 0. \end{equation} It follows from $\Delta^{2} f < 0 $ and equations (\ref{omg1}) and (\ref{omg2}) that $f$ satisfies \begin{equation}\label{gammapos} f(0) < f(1) \end{equation} and \begin{equation} f\left(\left\lfloor \frac{N}{2} \right\rfloor\right) > \dots > f(N), \end{equation} therefore, $f(k_{0}) = \max_{0 \leq k \leq N} f(k)$ for some $k_{0} \in \{1,\dots, \left\lfloor \frac{N}{2} \right\rfloor\}$. Defining the path $\gamma: \mathbf{-1} \rightarrow \mathbf{+1}$ by $\gamma = (L^{(0)}, L^{(1)}, \dots, L^{(N)})$, it is easy to see that \begin{equation} \Phi(\mathbf{-1}, \mathbf{+1}) = \max_{\sigma \in \gamma}H_{\Lambda,h}(\sigma) = \max_{0 \leq k \leq N}H_{\Lambda, h} (\mathscr{P}{^{(k)}}) =\Gamma+H_{\Lambda,h}(\mathbf{-1}). \end{equation} \end{proof} \subsection{Proof of Theorem \ref{t:minmax}.\ref{minmax3} and \ref{t:minmax}.\ref{minmax2}} Before giving the proof of the second point of the main theorem, we give some results about the control of the energy of a spin-flipped configuration. Given a configuration $\sigma$ and $k \in \Lambda$, the \emph{spin-flipped} configuration $\theta_{k}\sigma$ is defined as: \begin{equation} (\theta_{k}\sigma)_{i} = \begin{cases} -\sigma_{k} &\text{if $i = k$, and}\\ \sigma_{i} &\text{otherwise.} \end{cases} \end{equation} Note that the energetic cost to flip the spin at position $k$ from the configuration $\sigma$ is given by \begin{eqnarray*} H_{\Lambda,h}(\theta_{k}\sigma) - H_{\Lambda,h}(\sigma) &=& \sum_{\{i,j\} \subseteq \Lambda} J(|i-j|)(\sigma_{i}\sigma_{j} - (\theta_{k}\sigma)_{i}(\theta_{k}\sigma)_{j}) + h \sum_{i \in \Lambda} (\sigma_{i} - (\theta_{k}\sigma)_{i}) \\ &=& \left(\sum_{j \in \Lambda} J(|k-j|)2\sigma_{k}\sigma_{j} + 2h \sigma_{k} \right) \\ &=& 2\sigma_{k}\left(\sum_{j \in \Lambda} J(|k-j|)\sigma_{j} + h \right). \end{eqnarray*} \begin{prop} Under Condition~\ref{c:condition}, given a configuration $\sigma$ such that \begin{equation} \label{property} H_{\Lambda,h}(\theta_{k}\sigma) - H_{\Lambda,h}(\sigma) \geq 0 \end{equation} for every $k \in \{1,\dots,N\}$, then either $\sigma = \mathbf{-1}$ or $\sigma = \mathbf{+1}$. \end{prop} \begin{proof} Let $k \in \{1,\dots,N-1\}$, and let $\sigma$ be a configuration such that $\sigma_{i}= +1$ whenever $1 \leq i \leq k$ and $\sigma_{k+1} = -1$. In the following, we show that every such $\sigma$ cannot satisfy property (\ref{property}). If property (\ref{property}) is satisfied, then \begin{equation} \begin{cases} H_{\Lambda,h}(\theta_{k}\sigma) - H_{\Lambda,h}(\sigma) \geq 0 \\ H_{\Lambda,h}(\theta_{k+1}\sigma) - H_{\Lambda,h}(\sigma) \geq 0 \end{cases} \end{equation} that is, \begin{equation} \begin{cases} \sum_{i=1}^{k-1}J(|k-i|) - J(1) + \sum_{i=k+2}^{N}J(|k-i|)\sigma_{i} + h \geq 0 \\ -\left(\sum_{i=1}^{k}J(|k+1-i|) + \sum_{i=k+2}^{N}J(|k+1-i|)\sigma_{i} + h \right)\geq 0. \end{cases} \end{equation} Summing both equations above, we have \begin{eqnarray*} 0 &\leq& -J(k) - J(1) + \sum_{i=k+2}^{N}(J(i-k) - J(i-k-1))\sigma_{i} \\ &\leq& -J(k) - J(1) + \sum_{i=k+2}^{N}(J(i-k-1) - J(i-k))\\ &=& -J(k) - J(1) + \sum_{i=1}^{N-k-1}(J(i) - J(i+1))\\ &=& -J(k)-J(N-k) \end{eqnarray*} that is a contradiction. Analogously, every configuration $\sigma$ such that such that $\sigma_{i}= -1$ whenever $1 \leq i \leq k$ and $\sigma_{k+1} = 1$ for some $k \in \{1,\dots,N-1\}$, property (\ref{property}) cannot be satisfied. Therefore, we conclude that for every $\sigma$ different from $\mathbf{-1}$ and $\mathbf{+1}$, property (\ref{property}) does not hold. The proof of the converse statement is straightforward. \end{proof} As an immediate consequence of the result above, the next result follows. \begin{cor}\label{path} Under Condition~\ref{c:condition}, for every configuration $\sigma$ different from $\mathbf{-1}$ and $\mathbf{+1}$, there is a path $\gamma = (\sigma^{(1)}, \dots, \sigma^{(n)})$, where $\sigma^{(1)} = \sigma$ and $\sigma^{(n)} \in \{\mathbf{-1},\mathbf{+1}\}$, such that $H_{\Lambda,h}(\sigma^{(i+1)}) < H_{\Lambda,h}(\sigma^{(i)})$. \end{cor} We have now all the element for proving item \ref{minmax3} and \ref{minmax2} of Theorem ~\ref{t:minmax}. \begin{proof}[Proof of Theorem \ref{t:minmax}.\ref{minmax3}] First, note that it follows from inequality (\ref{gammapos}) that $\Gamma > 0$. Now, let us show that $V_{\mathbf{-1}}$ satisfies \begin{equation} V_{\mathbf{-1}} = \Phi(\mathbf{-1},\mathbf{+1}) - H_{\Lambda,h}(\mathbf{-1}). \end{equation} Since $\mathbf{+1}\in \mathscr{I}_{\mathbf{-1}}$, we have \begin{equation} \label{v1} V_{\mathbf{-1}} \leq \Phi(\mathbf{-1},\mathbf{+1}) - H_{\Lambda,h}(\mathbf{-1}). \end{equation} So, we conclude the proof if we show that \begin{equation} \label{v2} \Phi(\mathbf{-1},\mathbf{+1}) \leq \Phi(\mathbf{-1},\eta) \end{equation} holds for every $\eta\in \mathscr{I}_{\mathbf{-1}}$. Let $\gamma_{1}: \mathbf{-1} \rightarrow \eta$ be a path from $\mathbf{-1}$ to $\eta$ given by $\gamma_{1} = (\sigma^{(1)}, \dots, \sigma^{(n)})$, then, according to Corollary \ref{path}, there is a path $\gamma_{2} : \eta \rightarrow \mathbf{+1}$, say $\gamma_{2} = (\eta^{(1)},\dots, \eta^{(m)})$, along which the energy decreases. Hence, the path $\gamma: \mathbf{-1} \rightarrow \mathbf{+1}$ given by \begin{equation} \gamma = (\sigma^{(1)},\dots, \sigma^{(n-1)}, \eta^{(1)},\dots, \eta^{(m)}) \end{equation} satisfies \begin{equation} \Phi_{\gamma} (-\mathbf{1},+\mathbf{1}) = \Phi_{\gamma_1} (-\mathbf{1},\eta) \vee \Phi_{\gamma_2} (\eta,+\mathbf{1})) =\Phi_{\gamma_1} (-\mathbf{1},\eta). \end{equation} Hence, the inequality \begin{equation} \Phi(\mathbf{-1},\mathbf{+1}) \leq \Phi_{\gamma_{1}}(\mathbf{-1},\eta) \end{equation} holds for every path $\gamma_{1}: \mathbf{-1} \rightarrow \eta$, and equation (\ref{v2}) follows. \end{proof} \begin{proof}[Proof of Theorem \ref{t:minmax}.\ref{minmax2}] Given $\sigma \notin \{\mathbf{-1}, \mathbf{+1}\}$, let us show now that \begin{equation} \Phi(\sigma,\eta) - H_{\Lambda,h}(\sigma) < V_{\mathbf{-1}} \end{equation} holds for any $\eta\in \mathscr{I}_{\sigma}$. Let us consider the following cases. \begin{enumerate} \item Case $\eta = \mathbf{+1}$. According to Corollary (\ref{path}), there is a path $\gamma = (\sigma^{(1)}, \dots, \sigma^{(n)})$ from $\sigma^{(1)}= \sigma$ to $\sigma^{(n)} \in \{\mathbf{-1},\mathbf{+1}\}$ along which the energy decreases. \begin{itemize} \item[(a)] If $\sigma^{(n)} = \mathbf{-1}$, then the path $\gamma_{0}: \sigma \rightarrow \mathbb{\eta}$ given by $\gamma_{0} = (\sigma^{(1)}, \dots, \sigma^{(n-1)}, L^{(0)}, \dots, L^{(N)})$ satisfies \begin{eqnarray*} \Phi(\sigma,\eta) - H_{\Lambda,h}(\sigma) &\leq& \max_{\zeta \in \gamma_{0}}H_{\Lambda,h}(\zeta) - H_{\Lambda,h}(\sigma) \\ &\leq& \left(\max_{\zeta \in \gamma}H_{\Lambda,h}(\zeta)\right) \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)})\right) - H_{\Lambda,h}(\sigma) \\ &=& 0 \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)}) - H_{\Lambda,h}(\sigma)\right) \\ &<& \max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)}) - H_{\Lambda,h}(\mathbf{-1}) \\ &=& V_{\mathbf{-1}}. \end{eqnarray*} \item[(b)] Otherwise, if $\sigma^{(n)} = \mathbf{+1}$, then \begin{eqnarray*} \Phi(\sigma,\eta) - H_{\Lambda,h}(\sigma) &\leq& \max_{\zeta \in \gamma}H_{\Lambda,h}(\zeta) - H_{\Lambda,h}(\sigma) \\ &=& 0 \\ &<& V_{\mathbf{-1}}. \end{eqnarray*} \end{itemize} \item Case $\eta = \mathbf{-1}$. According to Corollary (\ref{path}), there is a path $\gamma = (\sigma^{(1)}, \dots, \sigma^{(n)})$ from $\sigma^{(1)}= \sigma$ to $\sigma^{(n)} \in \{\mathbf{-1},\mathbf{+1}\}$ along which the energy decreases. \begin{itemize} \item[(a)] If $\sigma^{(n)} = \mathbf{+1}$, then the path $\gamma_{0}: \sigma \rightarrow \mathbb{\eta}$ given by $\gamma_{0} = (\sigma^{(1)}, \dots, \sigma^{(n-1)}, L^{(N)}, \dots, L^{(0)})$ satisfies \begin{eqnarray*} \Phi(\sigma,\eta) - H_{\Lambda,h}(\sigma) &\leq& \max_{\zeta \in \gamma_{0}}H_{\Lambda,h}(\zeta) - H_{\Lambda,h}(\sigma) \\ &\leq& \left(\max_{\zeta \in \gamma}H_{\Lambda,h}(\zeta)\right) \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)})\right) - H_{\Lambda,h}(\sigma) \\ &=& 0 \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)}) - H_{\Lambda,h}(\sigma)\right) \\ &<& \max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)}) - H_{\Lambda,h}(\mathbf{-1}) \\ &=& V_{\mathbf{-1}}. \end{eqnarray*} \item[(b)] Otherwise, if $\sigma^{(n)} = \mathbf{-1}$, then \begin{eqnarray*} \Phi(\sigma,\eta) - H_{\Lambda,h}(\sigma) &\leq& \max_{\zeta \in \gamma}H_{\Lambda,h}(\zeta) - H_{\Lambda,h}(\sigma) \\ &=& 0 \\ &<& V_{\mathbf{-1}}. \end{eqnarray*} \end{itemize} \item Case $\eta \notin \{\mathbf{-1}, \mathbf{+1}\}$. Let $\gamma_{1} = (\sigma^{(1)}, \dots, \sigma^{(n)})$ and $\gamma_{2} = (\eta^{(1)}, \dots, \eta^{(m)})$ be paths from $\sigma^{(1)} = \sigma$ to $\sigma^{(n)} \in \{\mathbf{-1},\mathbf{+1}\}$ and from $\eta^{(1)} = \eta$ to $\eta^{(m)} \in \{\mathbf{-1},\mathbf{+1}\}$, respectively, along which the energy decreases. \begin{itemize} \item[(a)] If $\sigma^{(n)} = \eta^{(m)}$, define the path $\gamma: \sigma \rightarrow \eta$ given by $\gamma_{0} = (\sigma^{(1)}, \dots, \sigma^{(n-1)}, \eta^{(m)},\dots, \eta^{(1)})$ in order to obtain \begin{eqnarray*} \Phi(\sigma,\eta) - H_{\Lambda,h}(\sigma) &\leq& \max_{\zeta \in \gamma_{0}}H_{\Lambda,h}(\zeta) - H_{\Lambda,h}(\sigma) \\ &=& \left(\max_{\zeta \in \gamma_{1}}H_{\Lambda,h}(\zeta)\right) \vee \left(\max_{\zeta \in \gamma_{2}}H_{\Lambda,h}(\zeta)\right) - H_{\Lambda,h}(\sigma) \\ &=& H_{\Lambda,h}(\sigma) \vee H_{\Lambda,h}(\eta) - H_{\Lambda,h}(\sigma) \\ &=& 0 \\ &<& V_{\mathbf{-1}}. \end{eqnarray*} \item[(b)] If $\sigma^{(n)} = \mathbf{-1}$ and $\eta^{(m)} = \mathbf{+1}$, let us define the path $\gamma_{0}: \sigma \to \eta$ given by \begin{equation} \gamma_{0} = (\sigma^{(1)}, \dots, \sigma^{(n-1)}, L^{(0)}, \dots, L^{(N)}, \eta^{(m-1)}, \dots, \eta^{(1)}) \end{equation} satisfies \begin{eqnarray*} \Phi(\sigma,\eta) - H_{\Lambda,h}(\sigma) &\leq& \max_{\zeta \in \gamma_{0}}H_{\Lambda,h}(\zeta) - H_{\Lambda,h}(\sigma) \\ &=& \left(\max_{\zeta \in \gamma_{1}}H_{\Lambda,h}(\zeta)\right) \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)})\right) \vee \left(\max_{\zeta \in \gamma_{2}}H_{\Lambda,h}(\zeta)\right) - H_{\Lambda,h}(\sigma) \\ &=& H_{\Lambda,h}(\sigma) \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)})\right) \vee H_{\Lambda,h}(\eta) - H_{\Lambda,h}(\sigma) \\ &=& 0 \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)}) - H_{\Lambda,h}(\sigma)\right) \\ &<& \max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)}) - H_{\Lambda,h}(\mathbf{-1}) \\ &=& V_{\mathbf{-1}}. \end{eqnarray*} \item[(c)] If $\sigma^{(n)} = \mathbf{+1}$ and $\eta^{(m)} = \mathbf{-1}$, let us define the path $\gamma_{0}: \sigma \to \eta$ given by \begin{equation} \gamma_{0} = (\sigma^{(1)}, \dots, \sigma^{(n-1)}, L^{(N)}, \dots, L^{(0)}, \eta^{(m-1)}, \dots, \eta^{(1)}) \end{equation} satisfies \begin{eqnarray*} \Phi(\sigma,\eta) - H_{\Lambda,h}(\sigma) &\leq& \max_{\zeta \in \gamma_{0}}H_{\Lambda,h}(\zeta) - H_{\Lambda,h}(\sigma) \\ &=& \left(\max_{\zeta \in \gamma_{1}}H_{\Lambda,h}(\zeta)\right) \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)})\right) \vee \left(\max_{\zeta \in \gamma_{2}}H_{\Lambda,h}(\zeta)\right) - H_{\Lambda,h}(\sigma) \\ &=& H_{\Lambda,h}(\sigma) \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)})\right) \vee H_{\Lambda,h}(\eta) - H_{\Lambda,h}(\sigma) \\ &=& 0 \vee \left(\max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)}) - H_{\Lambda,h}(\sigma)\right) \\ &<& \max_{0 \leq k \leq N}H_{\Lambda,h}(L^{(k)}) - H_{\Lambda,h}(\mathbf{-1}) \\ &=& V_{\mathbf{-1}}. \end{eqnarray*} \end{itemize} \end{enumerate} We conclude that for every $\sigma \notin \{\mathbf{-1},\mathbf{+1}\}$, we have $V_{\sigma} < V_{\mathbf{-1}}$. \end{proof} \section{Proofs of the critical droplets results} \begin{proof}[Proof of Proposition~\ref{critdrop}] As in the proof of Theorem~\ref{t:minmax}, let us define $f: \{0,\dots, N\} \rightarrow \mathbb{R}$ as \begin{equation} f(i) = H_{\Lambda, h} (L^{(i)}), \end{equation} and recall that \begin{equation} \Delta f (i) = 2\left( \sum_{n=1}^{N-i-1} J(n) - \sum_{n=1}^{i} J(n) - h \right). \end{equation} In the first case, we have $\Delta f(L-1) = 2(h_{L-1}^{(N)} - h) > 0$, thus, since $f$ decreases for all $i$ greater than $L$, and since $\Delta^2 f<0$, we conclude that $f$ attains a unique strict global maximum at $L$. In the second case, we have $\Delta f(k-1) = 2(h_{k-1}^{(N)} - h) > 0$ and $\Delta f(k) = 2(h_{k}^{(N)} - h) < 0$, so, $f$ attains a unique strict global maximum at $k$. Finally, in the third case, we have $\Delta f(k) = 0$, that is, $f(k) = f(k+1)$. Using the fact that $\Delta f(k+1) < 0 < \Delta f(k-1)$, we conclude that the global maximum of $f$ can we only be reached at $k$ and $k+1$. \end{proof} \begin{proof}[Proof of Corollary~\ref{corcrit}] Since $\sum_{n = 1}^{\infty}J(n)$ converges, it follows that the set in equation (\ref{crit}) is nonempty, thus $k_{c}$ is well defined. Then, we have \begin{equation} \sum_{n = k_{c}+1}^{\infty}J(n) \leq h < \sum_{n = k_{c}}^{\infty}J(n). \end{equation} For all $N$ sufficiently large such that $\left\lfloor \frac{N}{2} \right\rfloor > k_{c}$ and \begin{equation} \sum_{n = N - k_{c}+1}^{\infty}J(n) < \sum_{n = k_{c}}^{\infty}J(n) - h, \end{equation} we have \begin{equation} h < \sum_{n = k_{c}}^{\infty}J(n) - \sum_{n = N - k_{c}+1}^{\infty}J(n) = h_{k_{c}-1}^{(N)} \end{equation} and \begin{equation} h_{k_{c}}^{(N)} = \sum_{n = k_{c}+1}^{\infty}J(n) - \sum_{n = N - k_{c}}^{\infty}J(n) < h. \end{equation} Therefore, by means of Proposition \ref{critdrop}, we conclude that for $N$ large enough, $k_{c}$ satisfies \begin{equation} H_{\Lambda,h}(\mathscr{P}^{(k_{c})}) > \max_{\substack{0 \leq i \leq N \\ i \neq k_{c}}} H_{\Lambda,h}(\mathscr{P}^{(i)}). \end{equation} \end{proof}

632acd5af32069040ee465b92bc6fa1af55f861a

\section{INTRODUCTION} \label{sec:intro} The {\it Chandra X-ray Observatory} (CXO) was launched on 23 July 1999 on the Space Shuttle {\it Columbia}. An overview of the mission and its instruments are presented in Weisskopf~{\em et al.}~(2000)~\cite{weisskopf2000} and an update on the mission was provided in Weisskopf~{\em et al.}~(2012)\cite{weisskopf2012}. The CXO carries two imaging instruments, the {\it Advanced CCD Imaging Spectrometer} (ACIS) discussed in Garmire~{\em et al.}~(1992)~\cite{garmire92} and Garmire~{\em et al.}~(2003)~\cite{garmire03} and the {\it High Resolution Camera} (HRC) discussed in Murray~{\em et al.}~(1997)~\cite{murray1997}. In addition, the CXO carries two gratings instruments known as the {\it High Energy Transmission Grating} (HETG) described in Canizares~{\em et al.}~(2000)~\cite{canizares2000} and the {\it Low Energy Transmissions Grating} (LETG) Brinkman~{\em et al.}~(2000)~\cite{brinkman2000}. ACIS is the primary instrument on the CXO with a nominal bandpass of 0.3--10.0~keV, conducting over 90\% of the observations. ACIS contains 10 CCDs arranged into two arrays. One array, the ACIS Imaging array (ACIS-I), consists of four frontside illuminated (FI) CCDs arranged in a $2\times2$ array, and the other array, the ACIS Spectroscopy array (ACIS-S), consists of four FI CCDs and two backside illuminated (BI) CCDs arranged in a $1\times6$ array. The ACIS-I array is used primarily for imaging spectroscopy and the ACIS-S array is used primarily as the readout detector for the HETG and LETG, although the ACIS-S is also used for imaging spectroscopy. The BI CCDs have higher quantum efficiency at low energies than the FI CCDs and are therefore preferred over the FI CCDs for some imaging observations. In order to suppress optical to infrared photons but to transmit the X-ray photons of interest, both ACIS arrays have an {\em Optical Blocking Filter} (OBF) inserted in the optical path. The filters were produced by ${\tt Luxel^{TM}}$ and are made of polyimide with Al deposited on both sides of the polyimide. The two filters are of slightly different thicknesses, the ACIS-S OBF (OBF-S) is 100/200/30 nm of Al/polyimide/Al and the ACIS-I OBF is 130/200/30 nm of Al/polyimide/Al. The OBFs sit about 12~mm in front of the CCDs facing the mirrors on the CXO. The volume around the CCDs is effectively isolated from the interior of the spacecraft, while the surface of the OBFs facing the mirrors is exposed to the interior of the spacecraft. The CCD focal plane is regulated at a temperature of -120~C. The OBFs are positioned at the top of the ACIS Camera Body (CB) which was regulated at -60~C early in the mission, but has been unregulated from April 2008 fluctuatng between between -72~C and -60~C. The CB was regulated at -60~C from August 2015 until July 2016 but has since been unregulated (see Plucinsky~{\em et al.}~2016~\cite{plucinsky2016} for details). In normal operations, the centers of the filters are warmer by $\sim2-4$ degrees due to the radiative heat load of the warm mirrors (+20~C) and the optical bench assembly. It was noticed early in the mission\cite{plucinsky2003} that the low energy sensitivity of the ACIS instrument was decreasing with time. It was quickly determined that this loss of detection efficiency was the result of a contamination layer building up on the surface of the OBFs facing the spacecraft interior. The contamination layer continues to accumulate even after 18 years on orbit. The accumulation rate, the chemical composition, and the spatial distribution of the contaminant have all varied with time over the mission. The accumulation rate exhibited a steep rise at the beginning of the mission, a flattening from 2003 until 2010, and then another steep rise from 2010 onwards. We reported in 2016\cite{plucinsky2016} on our efforts to reduce the accumulation rate by turning on the ACIS Detector Housing (DH) heater which regulates the CB and hence OBF edges at -60~C. There was no measurable effect on the accumulation rate due to the DH heater regulating the CB at -60~C. In this paper we report that the accumulation rate has decreased significantly starting in 2017 and we discuss our current understanding of the time-variable accumulation rate and chemical composition. \section{ACIS Contamination Layer} \label{sec:contam} \subsection{Discovery and Initial Characterization} \label{sec:discovery} The existence of the contamination layer was discovered in 2002\cite{plucinsky2003} as a gradual decrease in the low energy detection efficiency of all of the CCDs. The growth of the contamination layer was tracked by repeated observations of the {\em external calibration source}~(ECS) which has lines of Al-K~(1.5 keV), Ti-K~(4.5~keV), and Mn-K~(5.9~keV) from an ${\rm Fe^{55}}$ radioactive source with a half-life of 2.7~yr. The ECS also produced a line complex from Mn-L around 0.67 keV. The ratio of the Mn-L/Mn-K count rates on the S3 BI CCD became the most useful measure of the declining sensitivity at low energies. Unfortunately, the observed flux from the Mn-L line complex decreased with time due to the decay of the radioactive source and the increasing thickness of the contamination layer. Eventually the uncertainties on the measurements became so large that they were no longer useful to track the growth of the contamination layer. As the mission progressed, we transitioned to using celestial sources to monitor the growth of the contamination layer. We used celestial sources that are believed to be constant (or nearly constant on human time scales), such as clusters of galaxies and supernova remnants (SNRs), to monitor the change in low energy detection efficiency. We also used bright, variable sources with the HETG and LETG to constrain the absorption as a function of energy produced by the contaminant. Early efforts to determine the chemical composition of the contaminant\cite{marshall2004} identified absorption edges of C, O, and F that were in excess of the edges in the ACIS OBF. The ACIS OBF has absorption edges of C and O, but no edge due to F. The ACIS detection efficiency as a function of energy was carefully calibrated before launch\cite{bautz1998,nousek1998} including the transmission and absorption edges of the OBFs. The flight measurements used a bright continuum source dispersed with the HETG and/or LETG to achieve the highest spectral resolution possible with the CXO. In these high resolution spectra, it became obvious that some absorption edges were deeper than in the pre-flight measurements or only appeared (in the case of F) after launch. The newly-detected absorption edges were also found to be increasing in time. C was by far the dominant species in the contaminant while the O and F were approximately equal in concentration. We believe the contaminant started accumulating as soon as the ACIS door was opened and the OBFs were exposed to the interior of the spacecraft. The contaminant has continued to accumulate for the entire 19 year mission of the CXO, see Section~\ref{sec:accumulation} for a detailed time history. \subsection{OBF and Camera Body Temperatures} \label{sec:heater} ACIS has two separate filters, one for the Imaging array, OBF-I, and one for the spectroscopy array, OBF-S. For diagrams and pictures of the flight hardware, see the figures in Plucinsky~{\em et al.}~(2004)\cite{plucinsky2004}. Both OBFs are secured to the top surface of the ACIS Camera Body (CB). The OBFs have no active thermal control but respond to the environment around them. The edges of the filter are in good thermal contact with the CB and are therefore at the same temperature as the CB. At the beginning of the mission, the CB was held at -60~C. The centers of the filter are warmer than the edges due to the radiative heat load from the warm mirrors and optical bench cavity. The center of the OBF-I is modeled to achieve a temperature of $\sim-56$~C while the center of the OBF-S is at about $\sim-58$~C. There is no temperature sensor on the OBFs themselves. In April 2008, it was decided to turn off the ACIS Detector Housing (DH) heater which kept the CB temperature at -60 C. With the DH heater off, the CB temperature fluctuated between -72~C and -62.5~C depending on the orientation of the CXO spacecraft. The cooler CB temperature provided more margin for keeping the CCDs in the focal plane at -120~C. From launch in 1999 until April 2008, the CB temperature regulated at -59.9~C except for a few excursions during special activities. After April 2008, the CB temperature was unregulated and varied with the orientation of the spacecraft. In August 2015, it was decided to turn the ACIS DH heater back on with the hope that the accumulation rate of the contaminant would decrease. But as we reported in 2016\cite{plucinsky2016}, the warmer CB temperatures has no effect on the accumulation rate of the contaminant. Therefore, it was decided in July 2016 to turn the DH heater back off and leave the CB temperatures unregulated. \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics{fig1.eps} \end{tabular} \end{center} \caption[] { \label{fig:akos_aciss} Optical depth at 0.67 keV for the ACIS-S aimpoint as determined by fits to the E0102(blue) and A1795(red) data. The black line is the model for the optical depth in the N0010 contamination model. } \end{figure} \subsection{Time Dependence of the Accumulation Rate} \label{sec:accumulation} As mentioned in Section~\ref{sec:discovery}, the accumulation rate of the contamination layer was monitored with the ECS until the radioactive source became too faint to produce reliable results. At this point, we switched to using the brightest SNR in the Small Magellanic Cloud, 1E~0102.2-7219 (hereafter E0102), a bright cluster of galaxies known as Abell~1795 (hereafter A1795), and a bright Blazar called Markarian 421 (hereafter Mkn~421). E0102 has a soft, line-dominated spectrum and we have used it throughout the mission to characterize the contamination layer\cite{plucinsky2008,plucinsky2012}. The development of the standard IACHEC model for E0102 and its application to the current generation of X-ray instruments is presented in our 2017 paper\cite{plucinsky2017}. A1795 has a harder thermal spectrum with some significant line emission. Mkn~421 has a continuum spectrum described by a curved power-law model. \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics{fig2.eps} \end{tabular} \end{center} \caption[] { \label{fig:akos_acisi} Optical depth at 0.67 keV for the ACIS-I aimpoint as determined by fits to the E0102(blue) and A1795(red) data. The black line is the model for the optical depth in the N0010 contamination model. } \end{figure} We have used the A1795 and E0102 data on ACIS-S and ACIS-I to measure the optical depth of the contaminant at 0.67~keV (the energy of the Mn-L complex in the ECS). The results for the ACIS-S aimpoint are plotted in Figure~\ref{fig:akos_aciss}. The blue data points are derived from the E0102 data, the red data points are derived from the A1795 data and the black curve shows the expected increase in the contamination layer that is contained in the current release of the CXC contamination file ``{\tt acisD1999-08-13contamN0010.fits}'', called ``N0010'' for short. The measured optical depths from the E0102 and A1795 data are consistent within the uncertainties. The accumulation history of the contaminant is shown in this figure, a steep rise early in the mission, a reduction in the rate from 2003 to 2010, another sharp increase after 2010, and an apparent decrease starting in 2017. The data in 2017 begin to deviate from the expected accumulation rate and the trend continues into 2018. The decrease in the accumulation rate is not correlated with the DH heater which was on from 11 August 2015 until 20 July 2016. The behavior at the aimpoint on ACIS-I is even more dramatic and shown in Figure~\ref{fig:akos_acisi}. The first data point to deviate from the expectation is in late 2017. Perhaps more interesting, the last data point in 2018 is consistent with no accumulation over the last 6 months. The uncertainties are relatively large so future measurements will be necessary to confirm this result. Note that the maximum optical depth is about 3.0 on the OBF-I and is about 2.5 on the OBF-S. The contaminant has apparently accumulated more rapidly at the center of the OBF-I than at the center of the OBF-S. This can be seen more clearly in Figure~\ref{fig:od_diff} which shows the difference in the optical depth at the aimpoints on ACIS-I and ACIS-S as a function of time. For most of the mission, the optical depths were within 0.2 of each other. But starting in 2015, the contaminant grew more rapidly near the aimpoint on ACIS-I reaching a maximum difference of 0.6 optical depths. Curiously, the most recent data point in 2018 shows the difference between OBF-I and OBF-S is decreasing. This suggests that the accumulation rate on OBF-I is close to zero while the accumulation rate is still positive and small on OBF-S. Future observations will be necessary to determine if this trend will continue. \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics{fig3.eps} \end{tabular} \end{center} \caption[] { \label{fig:od_diff} The difference in the optical depth at 0.67 keV at the ACIS-I and ACIS-S aimpoints. The contaminant grew more quickly on OBF-I than OBF-S from 2015 until 2017. } \end{figure} \subsection{Time Dependence of the Chemical Composition} \label{sec:composition} The high resolution spectra provided by the HETG have been used\cite{marshall2004} to constrain the chemical composition of the contaminant and how it has changed with time. The contaminant is composed mostly of C, with some O and F. One limitation of the HETG data is that they only provide information on the contaminant for the OBF-S filter. The optical depth of the contaminant for each element (C, O, \& F) is modeled as a functions of time and position with a two component model: \begin{center} $\tau(t;x,y) = \tau_0(t) + [\tau_1(t) \times f(x,y)]$ \end{center} \noindent where $\tau_0(t)$ represents a time-variable, spatially uniform component, $\tau_1(t)$ represents a time-variable, spatially variable component, and $ f(x,y)$ is the spatial distribution for the spatially variable component. Figure~\ref{fig:ck_tau0} shows the time dependence of the $\tau_0(t)$ and $\tau_1(t)$ components for C near the aimpoint on the ACIS-S detector derived from HETG observations of Mkn~421. The time dependence of $\tau_0(t)$ for C matches that of the N0010 model until the last few data points which are significantly below the line. This is similar to the behavior seen for A1795 and E0102 shown in Figure~\ref{fig:akos_aciss}. The time dependence of $\tau_1(t)$ for C matches that of the N0010 model in shape, but the N0010 model might be slightly under-predicting at late times. The $\tau_1(t)$ component has been mostly flat with time from 2015 onwards, while the $\tau_0(t)$ continues to accumulate, albeit at a lower rate than predicted by the N0010 model. One interpretation of this behavior is that the spatially uniform component and the spatially variable component correspond to separate materials and the spatially variable component has ceased to accumulate. Figure~\ref{fig:ok_tau0} shows the time dependence $\tau_0(t)$ and $\tau_1(t)$ components for O, again near the aimpoint on the ACIS-S detector. The time dependence of $\tau_0(t)$ for O matches that of the N0010 model until the last few data points which are significantly above the line. The time dependence of $\tau_1(t)$ for O matches that of the N0010 model in shape and amplitude. However, the data since 2015 are consistent with no growth in this component so the N0010 model may be over-predicting the contaminant at late times but the uncertainties are still large enough that the case is not definitive. The $\tau_0(t)$ result indicates that the N0010 model has less O than it should. But note that the total optical depth of O is significantly less than that of C, $\sim2.0$ versus $\sim15$, so that any error in the O optical depth has less effect on the observed spectra and is therefore more difficult to discern. \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics[width=2.4in,angle=90]{fig4l.ps} \includegraphics[width=2.4in,angle=90]{fig4r.ps} \end{tabular} \end{center} \caption[] { \label{fig:ck_tau0} LEFT: The optical depth at C-K of the spatially uniform component near the center of the ACIS-S array. RIGHT: The optical depth at C-K of the spatially variable component near the center of the ACIS-S array. The solid line for both is the prediction from the N0010 contamination model. } \end{figure} \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics[width=2.4in,angle=90]{fig5l.ps} \includegraphics[width=2.4in,angle=90]{fig5r.ps} \end{tabular} \end{center} \caption[] { \label{fig:ok_tau0} LEFT: The optical depth at O-K of the spatially uniform component near the center of the ACIS-S array. RIGHT: The optical depth at O-K of the spatially variable component near the center of the ACIS-S array. The solid line for both is the prediction from the N0010 contamination model. } \end{figure} \section{Performance of the Current Contamination File} \label{sec:verification} \subsection{CXC Calibration Files} \label{sec:calibration_files} The CXC calibration group is responsible for providing calibration files that accurately model the additional absorption produced by the contamination layer. As mentioned above the characterization of the contamination layer is complicated by the temporal variation of the thickness, the chemical composition and the spatial distribution. The CXC regularly acquires calibration data of standard targets such as E0102, A1795, and Mkn~421 to verify the current contamination calibration file. If deficiencies are found, a new calibration file is created to address those deficiencies. The ACIS contamination file has been updated 7 times over the course of the CXO mission. For the analysis that follows, we use version {\tt N0010} of the model, which is called {\tt acisD1999-08-13contamN0010.fits} in the CXC {\em Calibration Database}~(CALDB). We used {\em Chandra Interactive Analysis of Observations}~(CIAO) version~4.9 and CALDB version~4.7.8. \subsection{E0102 Model} \label{sec:e0102_model} We have defined a standard model for E0102 as part of the activities of the IACHEC. We have used this model extensively~\cite{plucinsky2008,plucinsky2012,plucinsky2017} to test and improve the ACIS response model earlier in the mission. The model is intended for calibration analyses and is not intended to provide any insight into E0102 as a SNR. The model is empirical in that it uses 52 Gaussians to model the line emission. It uses a two component absorption model, one component for the Galactic contribution and one for the Small Magellanic Cloud (SMC) contribution. We modeled the continuum using a modified version of the {\em Astrophysical Plasma Emission Code}~({\tt APEC})~\cite{smith2001} called the {\tt ``No-Line''} model. This model excludes all line emission, while retaining all continuum processes including bremsstrahlung, radiative recombination continua (RRC), and the two-photon continuum from hydrogenic and helium-like ions (from the strictly forbidden ${}^2S_{1/2} 2s \rightarrow$\ gnd and ${}^1S_0 1s2s \rightarrow$\ gnd transitions, respectively). We included two continuum components of this type in the E0102 model. For details of the model and the parameters assumed see Plucinsky~{\em et al.}~(2017)\cite{plucinsky2017}. \input{s3tab.tex} Although the standard IACHEC model has many parameters, most of them are held fixed when we fit the data for calibration purposes. The continuum components are fixed and the interstellar absorption components are held fixed. All the line energies and widths are also held fixed. Typically, we freeze all line normalizations except for the four normalizations of the brightest lines/line complexes. We allow the normalizations for the \ion{O}{vii}~He$\alpha$~{\em r} line, the \ion{O}{viii}~Ly$\alpha$ line, the \ion{Ne}{ix}~He$\alpha$~{\em r} line, and \ion{Ne}{X}~Ly$\alpha$ line to vary in the fit. For the \ion{O}{vii}~He$\alpha$ and \ion{Ne}{ix}~He$\alpha$ triplets, we link the normalizations of the {\em f}, {\em i}, and {\em r} lines to each other and only allow one of them to vary during the fitting process. In this way, the triplet can increase or decrease its normalization as a group but the normalizations of the individual lines in the triplet can not vary independently of each other. There is also a global constant that multiplies the entire spectrum that is allowed to vary. In this manner, we allow only 5 of the 208 parameters in the IACHEC model to vary when we fit. We are essentially allowing the normalizations of the four brightest line/line complexes to vary while freezing the weaker lines and the continuum. We assume that E0102 is not changing significantly over the 19 year lifetime of the CXO mission such that the flux from the source in 1999 is not significantly different from the flux in 2018. And therefore we assume the total flux in a given line is not changing or changing very little over the 19 year mission. \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics[width=4.0in,angle=270]{fig6.ps} \end{tabular} \end{center} \caption[] { \label{fig:s3_e0102_sp} ACIS-S3 spectra of E0102 from OBSIDs 18418(2016), 19850(2017) and 20639(2018). The 2016 data are fit with the standard IACHEC model and that model is overplotted on the 2017 and 2018 data. } \end{figure} \subsection{ACIS-S Results} \label{sec:acis_s} E0102 has been observed many times with ACIS-S since the beginning of the mission. The mirrors on the CXO mission produce such sharp X-ray images that observations of a point source can be affected by ``pileup''. ``Pileup'' is defined as two photons interacting with the CCD within one detection cell (typically a $3\times3$ pixel region) within one readout frame of the CCD. Even though E0102 is an extended source, some of the bright filaments in E0102 are bright enough to have significant pileup. Most of the observations of E0102 early in the mission were executed in ``full-frame'' mode with an integration time of 3.2~s. We have excluded the ``full-frame'' observations from our analysis and selected only the ``subarray'' observations with shorter frametimes of 1.1~s, 0.8~s, and 0.4~s in order to minimize the effects of pileup on our data. There are 32 subarray observations of E0102 on S3 included in our analysis listed in Table~\ref{tab:s3obs}. Most of these observations are near the center of the CCD with {\tt chipy} values around 512, but 13 of the observations are at different {\tt chipy} positions. \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics[width=5.5in]{fig7.ps} \end{tabular} \end{center} \caption[] { \label{fig:e0102S3} Line normalizations from E0102 on S3 as a function of time. The solid black line is the average of the data points near the on-axis point aimpoint. The red dashed lines are $+/-10\%$ above and below the average. The points away from the nominal aimpoint are indicated in green and blue. } \end{figure} We have fit all of the S3 observations with the standard IACHEC model allowing only the global normalization and the normalizations for the \ion{O}{vii}~He$\alpha$~{\em r} line, the \ion{O}{viii}~Ly$\alpha$ line, the \ion{Ne}{ix}~He$\alpha$~{\em r} line, and \ion{Ne}{X}~Ly$\alpha$ line to vary. Figure~\ref{fig:s3_e0102_sp} shows an example of these fits for the three most recent observations near the S3 aimpoint from 2016, 2017 and 2018. The model was fit to the 2016 data and then frozen for the 2017 and 2018 observations to demonstrate deficiencies in the time-dependent contamination model. The large residuals in the 2018 spectrum at the \ion{O}{viii}~Ly$\alpha$ and \ion{Ne}{X}~Ly$\alpha$ lines indicate that the contaminant is over-estimated. Note that the difference between the 2017 and 2018 observations is not as large as the model predicts. The residuals indicate that the \ion{O}{viii}~Ly$\alpha$ line and the \ion{Ne}{X}~Ly$\alpha$ are not well fitted in 2018 but the \ion{Ne}{ix}~He$\alpha$~{\em r} line is well fitted. This will be challenging to correct with a revised contamination model. We compared the fitted line normalizations for the \ion{O}{vii}~He$\alpha$~{\em r} line, the \ion{O}{viii}~Ly$\alpha$ line, the \ion{Ne}{ix}~He$\alpha$~{\em r} line, and the \ion{Ne}{X}~Ly$\alpha$ line as a function of time. The results are plotted in Figure~\ref{fig:e0102S3}. The solid black line is the average of the on-axis data and the black dashed lines are +/-10\% from the average. Figure~\ref{fig:e0102S3} shows that the line normalizations are mostly consistent to within $\pm10\%$ from 2003 through 2016 for the on-axis data points with the exception of the \ion{O}{vii}~He$\alpha$~{\em r} line. After 2016, the \ion{O}{vii}~He$\alpha$~{\em r} line and \ion{O}{viii}~Ly$\alpha$ deviate dramatically from the previous values. The 2018 normalizations on-axis for the \ion{O}{viii}~Ly$\alpha$ line and \ion{O}{vii}~He$\alpha$~{\em r} line are $\sim28\%$ and $\sim49\%$ higher than the average value. The \ion{Ne}{ix}~He$\alpha$~{\em r} line, and \ion{Ne}{X}~Ly$\alpha$ line normalizations are consistent with the average within 10\% so the problem on S3 appears to effect energies below 0.9~keV. \subsection{ACIS-I Results} \label{sec:acis_i} There are 16 subarray observations of E0102 on the I3 CCD in the ACIS-I array. Table~\ref{tab:i3obs} lists the observations with their locations on the CCD and exposure times and count rates. Unlike the S3 CCD where the aimpoint is near the middle of the CCD, the aimpoint on the I3 CCD is near the top, right corner (high {\tt chipx} and {\tt chipy}). Hence most of these observations have {\tt chipx} of $\sim875$ and {\tt chipy} values of $\sim930$. This position is close to the center of the OBF-I filter, so the contamination layer is thinner at this position than near the edges. There are only 3 of the 14 observations that are at positions other than the nominal aimpoint. \input{i3tab.tex} Figure~\ref{fig:i3_e0102_sp} shows an example of these fits for the three most recent observations near the I3 aimpoint from 2016, 2017 and 2018. Again, the model was fit to the 2016 data and then frozen for the 2017 and 2018 data to demonstrate deficiencies in the time-dependent contamination model. Note the dramatic difference in the expected model spectrum for the 2018 data. The N0010 contamination model is over-estimating the contamination by a large amount at the aimpoint on I3. This is partly due to the fact that the accumulation rate has decreased but it is also due to the fact that the N0010 contamination model predicted significantly more contamination at the aimpoint on I3 than S3 (see Figures~\ref{fig:akos_aciss} and ~\ref{fig:akos_acisi}). \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics[width=4.0in,angle=270]{fig8.ps} \end{tabular} \end{center} \caption[] { \label{fig:i3_e0102_sp} ACIS-I3 spectra of E0102 from OBSIDs 18417(2016), 19849 (2017) and 20638(2018). The 2016 data are fit with the standard IACHEC model and that model if overplotted on the 2017 and 2018 data. } \end{figure} We compared the fitted line normalizations for the \ion{O}{vii}~He$\alpha$~{\em r} line, the \ion{O}{viii}~Ly$\alpha$ line, the \ion{Ne}{ix}~He$\alpha$~{\em r} line, and the \ion{Ne}{X}~Ly$\alpha$ line as a function of time. The results are plotted in Figure~\ref{fig:e0102I3}. The over-correction for the contamination layer in 2018 is large. The normalizations for the \ion{O}{vii}~He$\alpha$~{\em r} line, the \ion{O}{viii}~Ly$\alpha$ line, the \ion{Ne}{ix}~He$\alpha$~{\em r} line, and the \ion{Ne}{X}~Ly$\alpha$ line are over-estimated by $\sim98\%$, $\sim125\%$, $\sim32\%$, and $\sim25\%$ respectively. The data from 2016 and earlier are mostly consistent with each other to within 10\%. The discrepancy begins in 2017 and dramatically worsens in 2018. The revised contamination file soon to be released by the CXC should address most of this dicrepancy. \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics[width=5.5in]{fig9.ps} \end{tabular} \end{center} \caption[] { \label{fig:e0102I3} Line normalizations from E0102 on I3 as a function of time. The solid black line is the average of the data points near the on-axis point aimpoint. The red dashed lines are $+/-10\%$ above and below the average. The points away from the nominal aimpoint are indicated in red and blue. } \end{figure} \section{Possible Explanations For the Reduction in the Accumulation Rate} The analyses presented up to this point measure the accumulation rate of the contaminant which is the difference between the deposition rate and the vaporization rate. If the accumulation changes, we do not know if the deposition rate changed or the vaporization rate changed or both. Over the course of the mission, many components on the CXO spacecraft have increased in temperature, reaching mission high values within the last few years. It is conceiveable that a component on the spacecraft was not out-gassing significantly early in the mission, but as its temperature increased it began to out-gas at a higher rate. Perhaps the out-gassing from this component has now started to decrease, as the source of the contaminant has diminished. Another possibility is that the temperature distributions on the filters have changed with time. Figure~\ref{fig:tice_emittance} shows the expected temperature distributions on the filters in the presence of no contamination when the emittance is expected to be 0.05. In this case, the center of the OBF-I is at -55.8~C and the center of the OBF-S is at $\sim -58.0$~C. As contaminant accumulates on the filters and the surrounding surfaces the temperature distribution will change, with the centers of the filters becoming warmer. For an emittance of 0.20, the center of the OBF-I increases to -41.7~C and the center of the OBF-S increases to $\sim -46.0$~C. The temperatures of the OBFs increase as the emittance increases because the OBFs are more coupled to the temperature of the warm optical bench assembly (+20~C). But as the emittance continues increasing the OBF temperatures start to decrease again because in this model, the surfaces around the OBF are also accumulating a contamination layer and those surfaces have a higher emittance which results in better coupling between those relatively cold surfaces and the OBFs. As shown previously\cite{odell2013}, the vaporization rate of materials is a steep function of temperature with the vaporization rate increasing by roughly one order of magnitude for every 5~C increase in temperature. Therefore, it is possible that the vaporization rate has increased by about two to three orders of magnitude in the centers of the filters as the temperatures have increased from $\sim -56$~C to $\sim-42$~C. . This could be part of the explanation for the reduction in the accumulation rate that has been observed. This would be consistent with the center of the OBF-I showing a larger reduction in the accumulation rate than the center of the OBF-S since the center of the OBF-I is warmer than the center of the OBF-S. \begin{figure} \begin{center} \begin{tabular}{c} \includegraphics[trim={0.7cm 0.5cm 1.3cm 0.5cm},clip,width=6.5in]{fig10.ps} \end{tabular} \end{center} \caption[] { \label{fig:tice_emittance} The expected temperature distributions on the OBF-S and OBF-I as a function of emittance. } \end{figure} \section{Future Work} \label{sec:future} The temporal model of the contamination correction in the {\tt N0010} file {\tt acisD1999-08-13contamN0010.fits} contained in CALDB~4.7.8 needs modification to predict less absorption near the center and edges of the OBFs. This is clear from the E0102 line normalizations presented in Figures~\ref{fig:e0102S3} and~\ref{fig:e0102I3}. The CXC calibration team is working on a revision to the {\tt N0010} model that will change the time dependence of the spatial distribution and will update the chemical composition as a function of time. We expect this revised contamination model to be released in two stages (both in 2018), one release for the OBF-I and one for OBF-S. The accumulation rate of the contaminant will need to be monitored more frequently in the coming years. The accumulation rate at the centers and edges of the OBFs for both ACIS-S and ACIS-I have all changed in unexpected ways over the past two years. The continued characterization of these accumulation rates with time may provide constraints on the deposition and vaporization rates. The CXO project considered a ``Bakeout'' of the ACIS instrument\cite{plucinsky2004} soon after the contamination layer was discovered in 2004. The project decided at that time that a Bakeout was not worth the risk. There have been several papers written describing models of an ACIS Bakeout, see O'Dell~{\em et al.}~(2005)\cite{odell2005}, O'Dell~{\em et al.}~(2013)\cite{odell2013}, and O'Dell~{\em et al.}~(2015)\cite{odell2015}. These papers predict a range of outcomes from successful to unsuccessful depending on the assumed volatilities for the contaminants. The recently discovered reduction in the accumulation rate makes it less likely the project will consider a Bakeout worth the risk. Nevertheless, we will continue to monitor the accumulation rate and spatial distribution of the contaminant to constrain the volatilities of the possible contaminants to hopefully to constrain the range of possible outcomes for a Bakeout. If the contaminant were observed to decrease in the center of the OBF-I and OBF-S, we would know that the vaporization rate is larger than the deposition rate at the current temperatures. Such a result would indicate that a Bakeout is likely to be successful, even at temperatures not much higher than the current range of -70~C to -42~C. \section{Conclusions} \label{sec:conclusions} We have presented the accumulation rate of the ACIS contamination layer as a function of time. The accumulation rate decreased from launch until 2005, was fairly linear from 2005 to 2010, increased after 2010 but has sharply decreased since 2016. The chemical composition of the contamination has changed with time, possibly indicating that multiple sources are responsible for the contamination. The C, O, and F all exhibit different time dependencies again indicating that multiple materials have accumulated at different rates over the course of the mission. Nevertheless, all three have shown a dramatic decrease over the past year. The explanation for this sudden decrease is not clear. It could be that the deposition rate has decreased or the vaporization rate has increased, or both. The CXC will need to monitor the contamination layer frequently with dedicated calibration observations over the coming years to accurately model the contamination layer We tested the current contamination model {\tt N0010} with the SNR E0102. We find that the fitted values for the normalizations of the \ion{O}{vii}~He$\alpha$~{\em r} line, the \ion{O}{viii}~Ly$\alpha$ line, the \ion{Ne}{ix}~He$\alpha$~{\em r} line, and \ion{Ne}{X}~Ly$\alpha$ line are mostly consistent to within $\pm10\%$ for both ACIS-S and ACIS-I near the aimpoint from 2003 through 2016. After 2016, the line normalizations begin to deviate from the average value, with deviations as large as 49\% at the aimpoint on ACIS-S and 125\% at the aimpoint on ACIS-I for the \ion{O}{viii}~Ly$\alpha$ line. The CXC is preparing a revised contamination file that will significantly improve the agreement from 2016 onwards for release this year. \acknowledgments We acknowledge support under NASA contract NAS8-03060. The {\em Chandra X-ray Observatory} is operated by the Smithsonian Astrophysical Observatory under contract to the NASA Marshall Space Flight Center (MSFC). The Advanced CCD Imaging Spectrometer (ACIS) was developed by the Massachusetts Institute of Technology and the Pennsylvania State University.\\ We sincerely thank all of our colleagues in the IACHEC that contributed to the development of the highly successful spectral model for E0102.

b5598063541aa0319956c33e4a79750d2dad7b54

\section{Introduction}\label{secintro} {Based on over three decades of large-scale investigations \cite{Sta80,Gut81,Lin82,Alb,Lin83,Lin86a,Lin86b}, obviously the inflationary paradigm is going to be the corner stones of modern cosmology. The inflation theory can describe the early Universe evolutions successfully and also could be considered as a remedy for three vital problems which old big bang theory was faced i.e. the flatness, horizon and heavy monopoles problems \cite{Lid00,Bas,Lem,Kin,Bau09,Bau14}. Besides, it seems to obtain a correct ratio for tensor-to-scalar ratio and, in general, a correct behavior of primordial perturbations this model is requisite \cite{Liddle0,Langl,Lyth,Guth00,Lidsey97,Bas,Mukhanov-etal,Haidar,Haidar2}. In the standard inflation model, the potential term of Lagrangian is dominated comparing to the canonical kinetic term, i.e. the potential term dominated during inflation \cite{Lid00,Bas,Bau09,Bau14}. However there exist inflationary models in which the kinetic term has different form from the canonical one, namely non-canonical models, such as Dirac-Born-Infeld (DBI) action where the non-canonical kinetic term is attributed to the scalar field. It could be realized that DBI scalar field model can be assumed as a subset of k-inflation scenario \cite{Arm,Gar,Li,Hwa,Fra10a,Fra10b,Unn12,Unn13,Zha14a,Gol,Nazavari}. The observational constraints on $k$-inflation and its perturbations have been considered in literature \cite{Gar} and \cite{Li}. Here we should emphasis despite a huge number of inflationary models, the precise data originated from inspecting Cosmic Microwave Background, CMB, have been reduced dramatically the number of allowable inflationary models \cite{Mar13,Mar14,Hos14a}. Additionally, some other noticeable researches have been worked out in the context of non-canonical inflationary scenario and we refer the reader to \cite{Hwa,Fra10a,Fra10b,Unn12,Unn13,Zha14a,Gol}. In \cite{Unn12} it was shown that one can reduce the values of slow-roll parameters and accordingly the slow-roll regime condition can be obtained very easily by using of a non-canonical Lagrangian than the canonical case. In addition, it has been shown that the steep potentials connected to dark energy in the canonical setting can drive inflation in the non-canonical framework precisely \cite{Unn12}. In the non-canonical setup, the power law inflation is consistent with the observational results and one can obtain a way to end the inflation and it causes to get rid of changing the form of the power law of the scale factor surrounding the horizon exit \cite{Unn13}.\\ Mostly there are exist various ways to, in the slow-rolling inflationary scenarios, obtain expression of observables such as the tensor and scalar spectral indices, the running of them and tensor-to-scalar ratio. Amongst them one can refer to the introducing different types for scale factor to run inflation and then examining the results comparing to observations. Here we are interesting to the intermediate type, which is the most generic and well-known, to investigate its role to derive inflation and for the first time has been presented by \cite{Barrow11}. In this procedure, the scale factor introduced as an exponential function based on the cosmic time, i.e. $a(t) = \exp\big( \kappa t^f \big)$, $\kappa>0$ where usually $0<f<1$ \cite{Vallinotto,Starobinsky}. It leads to an asymptotically negative power-law potential, we can refer the steep potentials for instance \cite{Rendall}. Here we can supply the reason why people named these set of scale factors the intermediate ones. For them usually the expansion of the Universe is faster than the case which made by power-law one, i.e. ($a(t)=t^p, p>1$), and slower than de-Sitter inflation ($a(t)=\exp(Ht), H=constant$)}. It is interesting also we mention here that in Einstein gravity intermediate inflation for $\alpha=2/3$ creates scale invariant perturbations \cite{Barrow11,Barrow-etal,Vallinotto,Starobinsky}. One important reason to consider the intermediate inflation is its results for tensor-to-scalar ratio and scalar spectrum index which are in a good agreement comparing to the CMB data \cite{BarrowNunes}. Due to advantages of the intermediate proposal in solving problems of inflation this scenario preserve an appropriate place in the community and for more details we can refer the reader to the literature \cite{Muslimov,BarowLiddle02,BarrowLiddle,kk,mohammadi} and references there.\\ The majority of investigations to find out dynamical evolution of inflationary models have been done in a homogenous and isotropic background for instance Friedmann-Lemaitre-Robertson-Walker, FLRW, metric. However, a bit little deviation of isotropy at the level of $10^{-5}$, has also been proposed by Bennett \textit{et al}, and subsequently this suggestion was approved by high resolution Wilkinson Microwave Anisotropy Probe, WMAP \cite{ch1:9,WM2}. We should emphasise here although according to recent studies the anisotropy should be small its imprints in large scale structure formation are considerable. To justify this claim the the effects of anisotropy on the early Universe evolutions and especially the primary seeds of structure formation in the frame work of Bianchi type I (BI) exactly have been investigated \cite{ko,ku,Y1,17,18,aghaohamadi}. Amongst the Bianchi different types we can refer to the Kasner-type as a specific one; in which cosmological scale factors evolve by a power-law function of time. In General Relativity, GR, the vacuum Kasner solutions \cite{r7} and their fluid filled counterparts, the BI models, were verified effectively as a starting point for the investigation of the structure of anisotropic models. Barrow and Clifton \cite{r8,r9} have shown that it is possible finding out the solutions of the Kasner type for $R^n$-gravity. Newly, the authors in \cite{18a} have discussed the effects of low anisotropy on the interacting Dark Energy, DE, models and have shown the advantages of their model comparing to the standard FLRW, $\Lambda$ Cold Dark Matter, $\Lambda$CDM, and $w$CDM model results. Additionally, they showed that the anisotropy should get a non-zero value at the present time. Let's again turn our attention to the BI Universe; in fact the BI model is a straightforward extension of the flat FLRW metric so we can emphasize that it is the simplest model of anisotropic but homogenous Universe with spatial flatness. Against the FLRW Universe, which has a same scale factor for its three spatial directions, in the BI Universe the scale factor could change in different independent directions. Hence the study of inflation in an anisotropic Universe has much more advantages than isotropic one. Therefore, based on aforementioned reasons, in this work we are going to consider the anisotropic model to investigate the effects of the intermediate inflation with a non-canonical scalar field \cite{BarowLiddle02,BarrowLiddle}. By the way, here we want to answer this question why we need to deal with non-canonical model instead of pure canonical one. An immediately and temporary answer may be is intermediate inflation in the standard canonical inflation has some drawbacks and faced to failure comparing with observations. In Refs.\cite{BarowLiddle02,BarrowLiddle} it was showed that in light of the observations risen from the Cosmic Background Explorer ,COBE, the scalar and tensor power spectral expressed by intermediate inflation has no any valuable results. Besides, the aforementioned drawbacks without adding some extra processes to the model the intermediate inflation never cease, behaves as same as the eternal inflation \cite{BarrowLiddle}. At present work, we want to seek a probable remedy for these problems in which the canonical, more even non-canonical, inflation with intermediate inflation faced but in an anisotropic framework.\\ This work is organized as follows: In Sec.\,\ref{secnon-can}, we will express the main dynamical equations for non-canonical Lagrangian in an anisotropic metric. And in Sec.\,\ref{secInfnon-can} by virtue of an intermediate scale factor we will evaluate the inflationary observables and will compare their results with the Planck $2015$ data as a well-known criterion. Also, for the asymptotical regimes, i.e. canonical intermediate inflation and isotropic background, we will show that their results are not in a good agreement compared to general at hand proposal and the planck data. At last, Sec.\,\ref{seccon} is devoted to conclusion and discussions. \section{ non-canonical model in an anisotropic metric}\label{secnon-can} Usually non-canonical inflation could be expressed by the following action \begin{equation} \label{action} S = \int {{{\rm{d}}^4}} x\;\sqrt{-g}~\mathcal{L}(X,\phi), \end{equation} where Lagrangian ${\cal L}$ is a function of scalar field $\phi$ and its derivatives, i.e. the kinetic term $X \equiv {g^{\mu\nu}\nabla _\mu }\phi {\nabla_\nu }\phi /2$. By varying the action and after some algebra the equations of energy density $\rho_{\phi}$ and pressure $p_{\phi}$ are obtained as follows: \cite{Arm,Gar,Li,Hwa,Fra10a,Fra10b,Unn12,Unn13,Zha14a} \begin{eqnarray} \label{rhodef} {\rho _\phi } &=& 2X\left( {\frac{{\partial {\cal L}}}{{\partial X}}} \right) - {\cal L}~, \\ \label{pdef} {p_\phi } &=& {\cal L}~. \end{eqnarray} \label{sec2} As mentioned in introduction the BI cosmology refers to a spatially homogeneous background but not necessarily isotropic one. As a remembrance please note that we will consider BI cosmology in entire of this work expect when it has mentioned obviously. The metric of the BI model could be given by \begin{equation}\label{1} ds^2=dt^{2}-A^{2}(t)dx^{2}-B^{2}(t)dy^{2}-C^{2}(t)dz^{2}, \end{equation} where the metric components $A, B$ and $C$ are merely functions of time, for more details about Lie algebra and isometry group of the BI metric we refer the reader to \cite{35a} . From the literature we know the energy momentum tensor for perfect fluid is expressed by \begin{eqnarray}\label{4} T^{\mu}_{\nu}=diag[\rho,-p,-p,-p], \end{eqnarray} where $\rho$ and $p$ represent the energy density and pressure respectively. Additionally, the field equations in the axial symmetry BI metric are obtained as \cite{36,37,38} \begin{eqnarray} 3H^{2}-\sigma^{2}&=&\frac{1}{M_p^2}(\rho_{\phi}), \label{Fri} \\ 3H^2+2\dot{H}+\sigma^{2}&=&-\frac{1}{M_p^2}\left(p_{\phi}\right), \label{Fri2} \end{eqnarray} where $M_p^2=1/(8\pi G)$ is the reduced Planck mass, and $\sigma^2=\sigma_{ij}\sigma^{ij}/2 $ in which $\sigma_{ij}=u_{i,j}+\frac{1}{2}(u_{i;k}u^{k}u_{j}+u_{j;k}u^{k}u_{i})+\frac{1}{3}\theta(g_{ij}+u_{i}u_{j})$ is the shear tensor. By virtue of this tensor we can write down $(\sigma_{ij}u^j=0, \sigma^i_{~i}=0)$ that describes the rate of distortion of the matter flow. The scalar expansion introduced by $3H=u^{j}_{;j}$ where $u^j$ is 4-velocity and in a comoving coordinate it is given by $(u^i=\delta^i_0)$. Also the components of the Hubble parameter and the shear tensor based on the Eq.(\ref{1}) are expressed as \cite{36} \begin{eqnarray} H&=&\frac{1}{3}(\frac{\dot{A}}{A}+\frac{\dot{B}}{B}+\frac{\dot{C}}{C}), \label{14} \\ \sigma^{2}&=&3H^2-(\frac{\dot{A}\dot{B}}{AB}+\frac{\dot{B}\dot{C}}{BC}+\frac{\dot{A}\dot{C}}{AC}). \label{15} \end{eqnarray} If one takes $A=B^{\lambda}$ with $B=C$ the scale factor is appeared as $a = (ABC)^{1/3}=(B)^{(\lambda+2)/3} $ where $\lambda$ is a real constant. Then by assuming $H_2 = \frac{\dot{B} }{B}$ the Hubble parameter and the shear are reduced to the following simplified equations \begin{eqnarray} H = \frac{2 + \lambda }{3}H_2,\label{1an}\\ \sigma^2 = \frac{(\lambda-1)^2H_2^2}{3}.\label{2an} \end{eqnarray} By combining the Friedmann equations and embedded them into the Klein-Gordon equation the conservation equation is resulted as \begin{equation} \label{rhophidot} {\dot \rho _\phi } + 3H\left( {{\rho _\phi } + {p_\phi }} \right) = 0. \end{equation} \section{ Intermediate inflation for an anisotropic Universe and non-canonical Lagrangian}\label{secInfnon-can} In this section the inflationary behaviour for an intermediate scale factor by means of extended canonical Lagrangian inside a BI Universe is studied. Now let us turn our attention to investigate the inflationary evolution in the aforementioned framework. To do so we want to begin with introducing the namely first and second slow-roll parameters, viz. \begin{eqnarray} \label{eps} \varepsilon &=& - \frac{{\dot H}}{{{H^2}}}, \\ \label{eta} \eta &=& \frac{\dot{\varepsilon}}{H \varepsilon}. \end{eqnarray} To receive an accelerating phase as a necessary part of initial Universe, i.e. $\ddot a>0$, from Eq.(\ref{eps}) one immediately realize that the first slow-roll parameter should behave like $\varepsilon<1$. Also as mentioned above one of the big triumphes of inflation paradigm was the finding a remedy to cope with the horizon problem; inflation should drag on in order to persist the relation of $\varepsilon<1$ but the acceleration gets much smaller amounts than unity to run inflation. Hence, inflation occurs and persists if and only if both $\varepsilon $ and $\left| \eta \right|$ being much less than unity and so these assumptions in the literature named usually the slow-roll approximations. Another critical parameter to drive inflation in expected way is the number of e-fold which is defined as \begin{equation} \label{N} N = \int_{t_i}^t H dt = \int_{\phi_i}^{\phi} \frac{H}{{\dot \phi }}d\phi. \end{equation} In order to solve the horizon problem the number of e-fold should at least become more than 60 \cite{kk}. Now after introducing the necessary instruments of running the inflation we can go back to the Lagrangian again. The Lagrangian density which we shall consider can be considered as the following \cite{Unn12,Unn13} \begin{equation} \label{Lag} \mathcal{L}(X,\phi ) = X{\left( {\frac{X}{{{M^4}}}} \right)^{\alpha - 1}} - \;V(\phi ), \end{equation} where $M$ has the dimension of mass and $\alpha$ is a dimensionless parameter introduced to afford turning about to canonical case, i.e. ${\cal L}(X,\phi ) = X - V(\phi )$. Additionally, the Lagrangian \ref{Lag} satisfies the requirements $\partial {\cal L}/\partial X \ge 0 $ and ${\partial ^2}{\cal L}/\partial {X^2} > 0$ to cope with both the null-energy condition and physical propagations of perturbations respectively \cite{Fra10a}. This type of Lagrangian has been taken in account in vast number of prior literature to investigate some steep potentials for chaotic or other inflationary scenarios \cite{Unn12}. To refine the power law inflation in light of Planck $2013$ this Lagrangian has been considered as well \cite{Unn13}.\\ Now let's start the calculations based on the Lagrangian introduced in Eq.(\ref{Lag}). To do so we want to substitute the Lagrangian (\ref{Lag}) into the Eqs. (\ref{rhodef}) and (\ref{pdef}) and thence the energy density and pressure of the scalar field $\phi$ are given by \begin{eqnarray} \label{rhophi} {\rho _\phi } &=& \left( {2\alpha - 1} \right)X{\left( {\frac{X}{{{M^4}}}} \right)^{\alpha - 1}} + V(\phi), \\ \label{p} {p_\phi } &=& X{\left( {\frac{X}{{{M^4}}}} \right)^{\alpha - 1}} - V(\phi ). \end{eqnarray} In addition the dynamical equation of the scalar field, i.e. Klein-Gordon equation, by embedding the Eqs.(\ref{rhophi}) and (\ref{p}) into the conservation equation (\ref{rhophidot}) will be expressed as follows \begin{equation} \label{phiddot} \ddot \phi + \frac{{3H\dot \phi }}{{2\alpha - 1}} + \left( {\frac{{V'(\phi )}}{{\alpha (2\alpha - 1)}}} \right){\left( {\frac{{2{M^4}}}{{{{\dot \phi }^2}}}} \right)^{\alpha - 1}} = 0. \end{equation} Whereas we have no any interaction of type of non-minimally coupled chameleonic mechanism here, so by varying the lagrangian with respect to the scalar field we can obtain the above equation equally to the procedure which have been used in the papers \cite{Kho2,MOF1,ch19,ch20,ch1:20a}. Now by substituting Eqs. (\ref{14},\ref{15},\ref{1an},\ref{2an}) into the Eqs.(\ref{rhophi}) and (\ref{p}), the slow-roll parameters, i.e. Eqs.(\ref{eps}) and (\ref{eta}), based on the potential $V(\phi)$ are expressed as follows \begin{eqnarray} \label{epsV} {\varepsilon _V} &=&\frac{\sqrt{3(2\lambda +1)}}{2+\lambda} {\left[ {\frac{1}{\alpha }{{\left( {\frac{{3{M^4}}}{V(\phi)}} \right)}^{\alpha - 1}}{{\left( {\frac{{{M_P}V'(\phi)}}{{\sqrt 2 \;V(\phi)}}} \right)}^{2\alpha }}} \right]^{\frac{1}{{2\alpha - 1}}}}, \\ \label{etaV} {\eta _V} &=& \frac{\sqrt{3(2\lambda +1)}}{2+\lambda}\left( {\frac{{\alpha {\varepsilon _V}}}{{2\alpha - 1}}} \right)\left( {\frac{{2V(\phi )V''(\phi )}}{{V'{{(\phi )}^2}}} - 1} \right), \end{eqnarray} The Eqs.(\ref{epsV}) and (\ref{etaV}), so called the first and second potential based slow-roll parameters respectively. In addition, the slow-roll approximation implies the potential energy should be more larger than the kinetic one and therefore the Friedmann equation (\ref{Fri}) is reduced to \begin{equation} \label{Frisr} H^2\left(\phi\right) =\frac{(2+\lambda)^2}{9(2\lambda+1)} \frac{1}{{M_P^2}}V(\phi ). \end{equation} Meanwhile, under the slow-roll condition the dynamical equation of the scalar field, (\ref{phiddot}), is took the form \begin{equation} \label{phidot} \dot \phi = - \theta {\left \{\frac{\sqrt{3(2\lambda+1)}}{2+\lambda} {\left( {\frac{{{M_P}}}{{\sqrt 3 \alpha }}} \right)\left( {\frac{{\theta V'(\phi )}}{{\sqrt {V(\phi )} }}} \right){{\left( {2{M^4}} \right)}^{\alpha - 1}}} \right\}^{\frac{1}{{2\alpha - 1}}}}, \end{equation} where $\theta = \pm 1$ when the sign of $V'(\phi ) $ is $\pm$ \cite{Unn12, kk}. As mentioned in the aforementioned sections, the main aim of this study goes back to investigate the intermediate inflation in an anisotropic Universe, i.e. BI Universe. The scale factor expressed as $a = (ABC)^{1/3}=(B)^{(\lambda+2)/3} $ in which parameter $\lambda$ introduced to indicate the deviations of the isotropic background and could be considered a little bit larger or smaller than unity. Hence the appellation of low anisotropy implies these small deviations; and the component $B$ of the metric in intermediate inflation is expressed as \begin{equation} \label{at} B(t) = a_i \exp \left[ {{\kappa }{{\left( {{M_P}t} \right)}^f}} \right], \end{equation} where $a_i$ is the scale factor in $y$ axis direction, i.e. the $g_{22}$ component of the metric tensor at the initial time of the inflation. Thereupon, one will be able to obtain the main scale factor as $a=(a_i \exp [ {{\kappa ^2}{{\left( {{M_P}t} \right)}^f}} ])^{(\lambda+2)/3}$. Signally by virtue of this definition, the parameters of Hubble and shear could be obtained as follows \begin{equation} \label{Hubble} H^2=\frac{{{\kappa}^2{f^2}{{({M_p}t)}^{2f}}{{(2 + \lambda )}^2}}}{{9{t^2}}}, \end{equation} and \begin{equation} \label{shear} \sigma^2=\frac{{{\kappa}^2{f^2}{{({M_p}t)}^{2f}}{{( - 1 + \lambda )}^2}}}{{3{t^2}}}. \end{equation} In the above expressions we have the constraints ${\kappa }>0$ and $0<f<1$ \cite{BarowLiddle02,BarrowLiddle}. For more convenient the scale factor is normalized to the present time values as $a_0=1$. By using the Eqs.(\ref{Fri}) and (\ref{Fri2}) with the intermediate scale factor (\ref{at}) one receives \begin{eqnarray} \label{rhot} \rho _\phi &=& \frac{{\kappa }^{2f}M_p^2(M_pt)^{2f}(1 + 2\lambda )}{t^2} , \\ \label{pt} {p_\phi } &=& - \frac{{{\kappa}fM_p^2{{({M_p}t)}^f}[2( - 1 + f)(2 + \lambda ) + {\kappa }f{{({M_p}t)}^f}(5 + 2\lambda (1 + \lambda ))]}}{{3{t^2}}}. \end{eqnarray} Considering the slow-roll condition, i.e. ${\rho _\phi } = V(\phi )$, and Eq.(\ref{rhot}) we obtain \begin{equation} \label{Vt1} V(\phi )=\frac{{{\kappa ^2}{f^2}{M_P^2}{{\left( {M_P t} \right)}^{2f}}\left( {1 + 2\lambda } \right)}}{{{t^2}}}. \end{equation} Substituting Eq.(\ref{Vt1}) into (\ref{phidot}) we receive a first order differential equation to the scalar field as follows \begin{equation} \label{phit} \dot \phi(t) = {\left( { - \frac{{{2^\alpha }\left( { - 1 + f} \right){{\left( {{M^4}} \right)}^{ - 1 + \alpha }}{M_P}\sqrt {1 + 2\lambda } \sqrt {\frac{{{\kappa ^2}{f^2}M_P^2{{\left( {{M_P}t} \right)}^{2f}}\left( {1 + 2\lambda } \right)}}{{{t^2}}}} }}{{\alpha \left( {2 + \lambda } \right)t}}} \right)^{\frac{1}{2}/\alpha }}. \end{equation} Now, by integrating Eq.(\ref{phit}) and after some manipulations time $t$ could be obtained as a function of $\phi$, \begin{eqnarray} \label{tphi} \nonumber t(\phi ) &= &{2^{ - \frac{{2\alpha }}{{ - 2 + f + 2\alpha }}}}\\ &\times&{\left( { - \frac{{\left( {2 - f - 2\alpha } \right)\phi }}{{{\left( { - \frac{{{2^\alpha }\left( { - 1 + f} \right){{\left( {{M^4}} \right)}^{ - 1 + \alpha }}{M_P}\sqrt {1 + 2\lambda } \sqrt {{\kappa ^2}{f^2}M_P^{2 + 2f}\left( {1 + 2\lambda } \right)} }}{{\alpha \left( {2 + \lambda } \right)}}} \right)}^{ \frac{1}{2}/\alpha }}\alpha }} \right)^{\frac{{2\alpha }}{{ - 2 + f + 2\alpha }}}}. \end{eqnarray} Then to find the form of the potential, we can substitute the above solution in Eq.(\ref{Vt1}) and it gives \begin{equation} \label{Vt2} V(\phi ) = {V_0}{\phi ^s}, \end{equation} where \[\begin{array}{l} {V_0} = {\kappa ^2}{f^2}M_P^4\left( {1 + 2\lambda } \right)\\ \times \left( {{M_P}{{\left( { \frac{{-\left( {2 - f - 2\alpha } \right)}}{{2\alpha {{\left( { \frac{{{2^\alpha }\left( { 1 - f} \right){{\left( {{M^4}} \right)}^{ - 1 + \alpha }}\sqrt {{\kappa ^2}{f^2}M_P^{4 + 2f}{{\left( {1 + 2\lambda } \right)}^2}} }}{{\alpha \left( {2 + \lambda } \right)}}} \right)}^{\frac{1}{2}/\alpha }}}}} \right)}^{\frac{{2\alpha }}{{ - 2 + f + 2\alpha }}}}} \right)^{-2+2f}\equiv { V_0}^\ast{\kappa ^\frac{4\alpha-2}{-2+2\alpha+f}} \end{array}\] is a constant and \begin{equation} \label{s} s = \frac{{2\alpha ( - 2 + 2f)}}{{ - 2 + f + 2\alpha }}. \end{equation} It is obvious that the achieved potential behaves like the power law potentials \cite{BarowLiddle02,BarrowLiddle}. Whereas the value of the parameter $f$ for the intermediate scale factor (\ref{at}) gets the values of betwixt $0$ and $1$ \cite{Muslimov,BarowLiddle02,BarrowLiddle,kk,mohammadi}. Thence, from Eq.(\ref{s}) one can conclude that the parameter $s$ in (\ref{Vt2}) must be in the range $0 < s <-2\alpha/(\alpha - 1)$ to authorize the existence of intermediate inflation which $\alpha>1$ according to equation (\ref{Lag}). Since in the standard canonical setting ($\alpha =1$), so the $s$ parameter is varying between $-\infty< s <0$. Now, given the inverse power law potential form as source of inflation in the slow roll condition, we can obtain the necessary relations for determining the inflationary observables. The expression of the scalar and tensor power spectrum in the slow roll regime are given as \cite{Gar} \begin{eqnarray} \label{psk} {{\cal P}_s}&=&(\frac{H^2}{2\pi(c_s(\rho_{\phi}+p_{}\phi)^{1/2}})^2_{aH_{iso}=c_sk} , \\ \label{ptk} {{\cal P}_t} &=&\frac{8}{M_p^2}(\frac{H}{2\pi})^2_{aH_{iso}=k}. \end{eqnarray} By considering the Lagrangian (\ref{Lag}) and also the Eqs.(\ref{1an}, \ref{epsV}) in anisotropic metric, the above equations are expressed as \cite{Unn12,Unn13} \begin{eqnarray} \label{Ps} {{\cal P}_s} &=&\frac{(2+\lambda)^3}{(3(2\lambda+1))^3/2} \frac{1}{{72{\pi ^2}{c_s}}}\left( {\frac{{{6^\alpha }\alpha V{{(\phi )}^{5\alpha - 2}}}}{{M_P^{14\alpha - 8}{{ M}^{4(\alpha - 1)}}V'{{(\phi )}^{2\alpha }}}}} \right)_{a{ani} = {c_s}k}^{\frac{1}{{2\alpha - 1}}}.\\ {{\cal P}_t}&=&\frac{(2+\lambda)^2}{3(2\lambda+1)}\Big(\frac{2V(\phi)}{3{\pi}^2 M_p^4}\Big)_{aH_{ani}= k}.\label{PTV} \end{eqnarray} To receive equations \ref{Ps} and \ref{PTV} we used $H_{ani}=\frac{2+\lambda}{\sqrt{3(2\lambda+1)}}H_{iso}$ where subscribes ${ani} $ and ${iso}$ refer to the anisotropic and isotropic respectively. Now let's explain a little bit more about the constraint $aH_{ani} = {c_s}k$ in above equations. In fact, based on leading literature and textbooks the scalar power spectrum should be assessed at the sound horizon exit that specified by $aH_{ani} = {c_s}k$ where $k$ is the comoving wave number and $c_{s}$ refers to the sound speed \cite{Arm,Gar,Li,Hwa,Fra10a,Fra10b,Unn12,Unn13,Zha14a,kk}. Additionally, the sound speed has a definition as follows \begin{equation} \label{csdef} c_s^2 \equiv \frac{{\partial {p_\phi }/\partial X}}{{\partial {\rho _\phi }/\partial X}} = \frac{{\partial {\cal L}(X,\phi )/\partial X}}{{\left( {2X} \right){\partial ^2}{\cal L}(X,\phi )/\partial {X^2} + \partial {\cal L}(X,\phi )/\partial X}}. \end{equation} And for our investigation here it takes the following form \begin{equation} \label{cs} {c_s} = \frac{1}{{\sqrt{2\alpha - 1} }}, \end{equation} where behaves just as a constant. Replacing the potential (\ref{Vt2}) onto Eqs.(\ref{Ps}) and (\ref{PTV}) after some algebra gives \begin{eqnarray}\label{Psphi} {P_s} &= &{(\frac{{\left( {2 + \lambda } \right)}}{{\sqrt {3\left( {2\lambda + 1} \right)} }})^3}\\ \nonumber &\times &\frac{{{{\left( {{6^\alpha }M_P^{8 - 14\alpha }{\mkern 1mu} \alpha {\mkern 1mu} {\mu ^{4 - 4\alpha }}{{\left( {\frac{{2\alpha \left( { - 2 + 2f} \right)}}{{ - 2 + f + 2\alpha }}} \right)}^{ - 2\alpha }}{{\left( {{V_0}} \right)}^{ - 2 + 3\alpha }}} \right)}^{\frac{1}{{ - 1 + 2\alpha }}}}}}{{72{\pi ^2}{c_s}}}(\phi )_{aH_{ani} = {c_s}k}^{\frac{{\alpha \left( {6f - 4} \right)}}{{2\alpha + f - 2}}}, \end{eqnarray} and \begin{equation} \label{PTphi} {{\cal P}_t} = {(\frac{{\left( {2 + \lambda } \right)}}{{\sqrt {3\left( {2\lambda + 1} \right)} }})^2}\frac{{2{V_0}}}{{3M_P^4{\pi ^2}}}(\phi )_{aH_{ani} = k}^{\frac{{4\left( { - 1 + f} \right)\alpha }}{{ - 2 + f + 2\alpha }}}, \end{equation} where $\mu= M/M_p$. From Eq.(\ref{Psphi}), it obviously could be seen that for the value of $f=2/3$ the scalar power spectrum is appeared as an independent parameter of $\phi$ and so it makes sense like the scale-invariant Harrison-Zel$'$dovich spectrum. Now, in order to calculate the evolution of power spectrum based on $N$, we need the scalar field in terms of the number of e-folds. Hence, we might need to calculate the values of scalar field at the initiation of the inflation namely $\phi_{i}$. To do this end according to the slow-roll parameter definition i.e. equation (\ref{epsV}) we have \begin{equation} \label{epsilonphi} {\varepsilon _V}=\sqrt {3\left( {2\lambda + 1} \right)} \frac{{{{\left( {{\alpha ^{ - 1}}{2^{ - \alpha }}{3^{ - 1 + \alpha }}M_P^{2\alpha }{{\left( {\frac{{2\alpha \left( { - 2 + 2f} \right)}}{{ - 2 + f + 2\alpha }}} \right)}^{2\alpha }}{{\left( {\frac{{{M^4}}}{{{V_0}}}} \right)}^{ - 1 + \alpha }}} \right)}^{\frac{1}{{ - 1 + 2\alpha }}}}}}{{2 + \lambda }}{\phi ^{\frac{{s - \alpha s - 2\alpha }}{{ - 1 + 2\alpha }}}}. \end{equation} Now we can rely on these facts that at the beginning of the inflation $\varepsilon _V=1$ therefore easily we obtain the related value of scalar field as \begin{equation} \label{phibegin} \phi_i = {{\Big[}{{\Big(}{\alpha ^{ - 1}}{2^{ - \alpha }}{3^{ - 1 + \alpha }}{M^{4\alpha - 4}}{V_o}^{1 - \alpha }M_p^{2\alpha }{\Big)}^{ - \frac{1}{{ - 1 + 2\alpha }}}}\chi{\Big]}^{\frac{{1 - 2\alpha }}{{ - s + 2\alpha + s\alpha }}}}, \end{equation} where $\chi = \frac{{(2 + \lambda )}}{{\sqrt {3(2\lambda + 1)} }}$. By bringing in account the Eq.(\ref{N}) we get \begin{equation} \label{NN} \phi = {{\Big(}\phi_i^{\frac{{2 - f}}{{ - 2 + f + 2\alpha } + s}} + \frac{N}{\Lambda }{\Big)}^{\frac{{ - 2 + f + 2\alpha }}{{(2 - f)(1 - s) + 2\alpha s}}}}, \end{equation} where $$\Lambda = \frac{{2\alpha (2 + \lambda )\sqrt {{V_0}} {{({\gamma ^{ - \frac{{2\alpha }}{{ - 2 + f + 2\alpha }}}})}^{ - \frac{{ - 2 + f}}{{2\alpha }}}}}}{{3{M_p}( - 2 + f + 2\alpha )(1 + \frac{s}{2} - \frac{{ - 2 + f}}{{ - 2 + f + 2\alpha }})\gamma \sqrt {1 + 2\lambda } }},$$ and $$\gamma = \frac{{2\alpha {{( - \frac{{{2^\alpha }( - 1 + f){M^{ - 4 + 4\alpha }}{M_p}\sqrt {1 + 2\lambda } \sqrt {{\kappa ^2}{f^2}{M_p}^{2 + 2f}(1 + 2\lambda )} }}{{\alpha (2 + \lambda )}})}^{1/2\alpha }}}}{{ - 2 + f + 2\alpha }}.$$ Now, by virtue of Eqs.(\ref{Psphi}) and (\ref{NN}), the scalar power spectrum in terms of the number of e-folds is given by \begin{eqnarray} \label{Psk} {{\cal P}_s}={(\frac{{\left( {2 + \lambda } \right)}}{{\sqrt {3\left( {2\lambda + 1} \right)} }})^3}\frac{{{{\left( {{6^\alpha }M_P^{8 - 14\alpha }\,\alpha \,{\mu ^{4 - 4\alpha }}{{\left( {\frac{{2\alpha \left( { - 2 + 2f} \right)}}{{ - 2 + f + 2\alpha }}} \right)}^{ - 2\alpha }}{{\left( {{V_0}} \right)}^{ - 2 + 3\alpha }}} \right)}^{\frac{1}{{ - 1 + 2\alpha }}}}}}{{72{\pi ^2}{c_s}}}\times\cr {{\Big(}{\Big(}\phi_i^{\frac{{2 - f}}{{ - 2 + f + 2\alpha } + s}} + \frac{N}{\Lambda }{\Big)}^{\frac{{ - 2 + f + 2\alpha }}{{(2 - f)(1 - s) + 2\alpha s}}}}{\Big)}^{\frac{{\alpha \left( {6f - 4} \right)}}{{2\alpha + f - 2}}}. \end{eqnarray} Whereas both $H$, in the slow roll inflation, and $c_s$, in this work are constant. Consequently at the sound horizon exit, $aH = c_s k$, we have \cite{Unn13} \begin{equation} \label{exithorizon} \frac{{\rm{d}}}{{{\rm{d ln k}}}} \simeq -\frac{{\rm{d}}}{{{\rm{dN}}}}. \end{equation} The importance of this relation goes back to calculation the especially spectral indices. For the scalar spectral index we can write \begin{equation} \label{nsdef} {n_s} - 1 \equiv \frac{{d\ln {{\cal P}_s}}}{{d\ln k}}, \end{equation} in which by using Eq. (\ref{Psk}) we get \begin{eqnarray} \label{nsk} n_s&=&1 - \frac{{( - 4 + 6f)\alpha }}{{(2 - f)(1 - s) + 2\alpha s)\Lambda }}\cr &\times&\frac{1}{{\frac{N}{\Lambda } + {{\Big{(}{{({{(\frac{{{2^{ - \alpha }}{3^{ - 1 + \alpha }}{M^{4( - 1 + \alpha )}}V_0^{1 - \alpha }{{({M_p}s)}^{2\alpha }}}}{\alpha })}^{\frac{1}{{1 - 2\alpha }}}}\chi )}^{\frac{{1 - 2\alpha }}{{s( - 1 + \alpha ) + 2\alpha }}}}\Big{)}}^{s + \frac{{2 - f}}{{ - 2 + f + 2\alpha }}}}}}. \end{eqnarray} Another important parameter one can refer, to investigate the behavior and evolution of the initial cosmos, is the running parameter. We can consider the running of the scalar spectral index as \begin{eqnarray} \label{dnsk} \alpha_s&=&\frac{{d{n_s}}}{{d\ln k}} =\frac{{( - 4 + 6f)\alpha }}{{{\Lambda ^2}(2 + f( - 1 + s) + 2s( - 1 + \alpha ))}}\cr &\times& {{\Bigg[\frac{N}{\Lambda } + {{\Big{(}{{({{(\frac{{{2^{ - \alpha }}{3^{ - 1 + \alpha }}{M^{4( - 1 + \alpha )}}V_0^{1 - \alpha }{{({M_p}s)}^{2\alpha }}}}{\alpha })}^{\frac{1}{{1 - 2\alpha }}}}\chi )}^{\frac{{1 - 2\alpha }}{{s( - 1 + \alpha ) + 2\alpha }}}}\Big{)}}^{s + \frac{{2 - f}}{{ - 2 + f + 2\alpha }}}}\Bigg]^{-2}}} \end{eqnarray} After the completion of the scalar part, now we can go through the tensor part. In terms of the number of e-folds the tensor power spectrum, by means of Eqs.(\ref{Lag}) and (\ref{PTphi}), is given by \begin{equation} \label{Pt} {{\cal P}_t}=\frac{{2V_0{\chi ^2}{{3{M_p}^{-4}{\pi^{-2}}}}}}{\Bigg[{{\Big{(}\frac{N}{\Lambda } + {{({{({{(\frac{{{2^{ - \alpha }}{3^{ - 1 + \alpha }}{M^{4( - 1 + \alpha )}}V_0^{1 - \alpha }{{({M_p}s)}^{2\alpha }}}}{\alpha })}^{\frac{1}{{1 - 2\alpha }}}}\chi )}^{\frac{{1 - 2\alpha }}{{s( - 1 + \alpha ) + 2\alpha }}}})}^{s + \frac{{2 - f}}{{ - 2 + f + 2\alpha }}}}\Big{)}}^{\frac{1}{{s + \frac{{2 - f}}{{ - 2 + f + 2\alpha }}}}}}\Bigg]^{-s}}. \end{equation} The tensor spectral index is defined as \begin{equation} \label{ntdef} {n_t} \equiv \frac{{d\ln {{\cal P}_t}}}{{d\ln k}}. \end{equation} Now by using Eqs.(\ref{exithorizon}), (\ref{Pt}), and the above equation one can obtain \begin{eqnarray} \label{ntn} {n_t} &=& \frac{{-s( - 2 + f + 2\alpha )}}{{((2 - f)(1 - s) + 2\alpha s)\Lambda }}\cr &\times&\frac{1}{{\frac{{N}}{\Lambda } + {{\Big{(}{{({{(\frac{{{2^{ - \alpha }}{3^{ - 1 + \alpha }}{M^{4( - 1 + \alpha )}}V_0^{1 - \alpha }{{({M_p}s)}^{2\alpha }}}}{\alpha })}^{\frac{1}{{1 - 2\alpha }}}}\chi )}^{\frac{{1 - 2\alpha }}{{s( - 1 + \alpha ) + 2\alpha }}}}\Big{)}}^{s + \frac{{2 - f}}{{ - 2 + f + 2\alpha }}}}}}. \end{eqnarray} To measure the amplitude of the primordial fluctuations we need to calculate the tensor-to-scalar ratio which is defined as \begin{equation} \label{rdef} r \equiv \frac{{{{\cal P}_t}}}{{{{\cal P}_s}}}, \end{equation} where by using Eqs. (\ref{Psk}), (\ref{Pt}) and (\ref{rdef}), it can be expressed by \begin{eqnarray} \label{rn} r& =& {\Bigg[{\Big[\frac{N}{\Lambda } + {\Big{(}{({(\frac{{{2^{ - \alpha }}{3^{ - 1 + \alpha }}{M^{4( - 1 + \alpha )}}V_0^{1 - \alpha }{{({M_p}s)}^{2\alpha }}}}{\alpha })^{\frac{1}{{1 - 2\alpha }}}}\chi )^{\frac{{1 - 2\alpha }}{{s( - 1 + \alpha ) + 2\alpha }}}}\Big{)}^{s + \frac{{2 - f}}{{ - 2 + f + 2\alpha }}}}\Big]^{\frac{1}{{s + \frac{{2 - f}}{{ - 2 + f + 2\alpha }}}}}}\Bigg]^{s + \frac{{(4 - 6f)\alpha }}{{ - 2 + f + 2\alpha }}}}\cr &\times&\frac{{48{c_{s}}{V_0}{{({6^\alpha }M_{_p}^{^{8 - 14\alpha }}V_0^{^{ - 2 + 3\alpha }}{{(s)}^{ - 2\alpha }}\alpha {\mu ^{4 - 4\alpha }})}^{\frac{1}{{1 - 2\alpha }}}}}}{{M_p^4\chi }}. \end{eqnarray} The consistency relation between observable $r$ and $n_t$ in the non-canonical inflation is as follows \begin{equation} \label{rnt3} r \approx -8 c_s n_t, \end{equation} which has an extra $c_s$ coefficient comparing to the canonical case, i.e. ($r=-8n_t$) \cite{Unn12,Unn13}. Replacing Eq.(\ref{ntn}) into Eq(\ref{rnt3}) gives \begin{eqnarray} \label{rnt2} r &\approx & \frac{{8 c_s s( - 2 + f + 2\alpha )}}{{((2 - f)(1 - s) + 2\alpha s)\Lambda }}\cr &\times&\frac{1}{{\frac{{N}}{\Lambda } + {{\Big{(}{{({{(\frac{{{2^{ - \alpha }}{3^{ - 1 + \alpha }}{M^{4( - 1 + \alpha )}}V_0^{1 - \alpha }{{({M_p}s)}^{2\alpha }}}}{\alpha })}^{\frac{1}{{1 - 2\alpha }}}}\chi )}^{\frac{{1 - 2\alpha }}{{s( - 1 + \alpha ) + 2\alpha }}}}\Big{)}}^{s + \frac{{2 - f}}{{ - 2 + f + 2\alpha }}}}}}. \end{eqnarray} Subsequently we are going to check the accuracy and consistency of our theoretical results. To do so we have to make a comparison with observation. One of the best criterions for our aim could be considered is the data risen by Planck $2013$ and $2015$ \cite{Planck2015}. It is distinct that one of the most important results of Planck data is the $r-n_s$ diagram and the validity of different models relies on their compatibility to this observation. Therefore, by virtue of Eqs.(\ref{nsk}) and (\ref{rnt2}) we excited to depict the $r-n_s$ diagram for our scenario. This diagram is shown in figure \ref{fignsr}. Besides, the marginalized likelihoods based on Confidence Levels (CLs) 68\% and 95\% are allowed by Planck $2015$ , TT, TE, EE+lowP data \cite{Planck2015} and we illustrated them in the figure \ref{fignsr}. Predictions of our model are specified by solid black line for the values of $\alpha=3$, $\kappa=3.02\times 10^{-12}$, $f=10^{-4}$ and $\lambda=3.5$. From figure \ref{fignsr} it could be visualized that our results can be considered in acceptable ranges compared to the observations. Then, it could be concluded this scenario is able to be regarded as a valid case for explaining the inflationary scenario. \begin{figure}[ht] \centering \includegraphics[scale=.50]{wnsas} \caption{{\it{The $r-n_s$ diagram indicates prediction of the non-canonical intermediate inflationary model in anisotropy background for the specified values of $\alpha=3$, $\kappa=3.02\times 10^{-12}$, $f=10^{-4}$ ,$M_p=10^{18}$, $M=10^{12}$ and $\mu=10^{-6}$, in comparison with the observational results of Planck 2015. Where the likelihood of Planck 2013 (grey contours), Planck TT+lowP(red contours), Planck TT,TE,EE+lowP(blue contours) and the thick black line indicate the predictions of our model in which small and large dots are the value of $n_s$ at the number of e-fold $N=55,~N=66$.}}} \label{fignsr} \end{figure} It is clear that the grey, red and blue CLs are correspond to Planck $2013$, Planck $2015$ TT+lowP and Planck $2015$ TT, TE, EE+lowP data respectively \cite{Planck2015}. Also in Table (\ref{FT}) we want to study the behaviour of the parameter $\lambda$, i.e. the effect of low anisotropy; one can obviously observe that the prediction of the model for perturbation parameters are relied upon the specific values of free parameters $f,~\alpha, ~\kappa $ but different values for parameter $\lambda$. Given the Table (\ref{FT}), it is clear that the $r-n_s$ diagram for $\alpha=3$ and $\lambda=3.5$ is in more consistency with the Planck $2015$ TT, TE, EE+lowP results \cite{Planck2015}. Meanwhile, at $\lambda=1$ and $\alpha=3$ the results lead to the non-canonical but isotropic Universe, that the $\mathcal{P}_s$has a little bit deviation of its observed value by Planck $\mathcal{P}^{\ast}_s=2.17\times 10^{-9}$ \begin{table}[h] \centering {\footnotesize \begin{tabular}{p{1.2cm}p{1.5cm}p{0.8cm}p{0.8cm}p{1.2cm}p{2.5cm}p{2.5cm}} \hline $ \alpha$ & $\kappa*10^{-12}$ & $\ \ f$ & $\lambda$ & $\ \ n_s$ &\ \ $ r$ & $\qquad \mathcal{P}_s$ \\[0.1mm] $3$ & $3.02$ & $10^{-4}$ & $3.5$ & $0.978$ & $0.078$ & $2.17\times 10^{-9}$ \\[2mm] $3$ & $3.02$ & $10^{-4}$ & $1$ & $0.978$ & $0.078$ & $6.199\times 10^{-10}$ \\[2mm] $1$ & $3.02$ & $10^{-4}$ & $3.5$ & $0.964$ & $0.167$ & $2.7\times 10^{590950}$ \\[0.1mm] $2$ & $3.02$ & $10^{-4}$ & $2.5$ & $0.976$ & $0.290$ & $7293.78$ \\[0.1mm] $1$ & $3.02$ & $10^{-4}$ & $1$ & $0.964$ & $0.130$ & $9.07\times 10^{587695}$ \\[0.1mm] \end{tabular} } \caption{\footnotesize The prediction of the model for the perturbation parameters $n_s$, $r$ and $\mathcal{P}_s$ are prepared for different values of the free parameters $\lambda$ and $ \alpha$ besides the specified values of other free parameters. Also we used $M_p=10^{18}$, $M=10^{12}$ and $N=55$. This analyze shows the best behaviour for the first row of the table. On the second row we can see the behaviour of non-canonical model in isotropic background in which there is some deviations almost around one order for $\mathcal{P}_s$. We also examine the canonical case with anisotropic condition on third row and the result was very far from the observed results, especially the value for $\mathcal{P}_s$. The amounts for free parameters on the fourth row are supplied for more clarity in comparison. And finally we consider the canonical case in an isotropic background and the results again were not according to excepted results originated from observations}\label{FT} \end{table} Now we can turn into the running spectral index, $\alpha_s=d{n_s}/dN -{n_s}$, behaviour in comparison to the observational results originated of Planck data. So at first we regard $\alpha=3$, $\kappa=3.02\times 10^{-12}$, $f=10^{-4}$ ,$M_p=10^{18}$, $M=10^{12}$, $\mu=10^{-6}$, and $\lambda=3.5$. Then, by using Eqs. (\ref{nsk}) and (\ref{dnsk}) we will plot $d{n_s}/dN$ versus $n_s$. The plot \ref{fignsdns1} shows the prediction of the model could lie insides the joint 68\% CLs region of Planck $2015$ TT, TE, EE+lowP data, and satisfies the agreement with observations \cite{Planck2015}. \begin{figure}[ht] \centering \includegraphics[scale=.50]{nsdns} \caption{{\it{ The $d{n_s}/dN - {n_s}$ diagram show Prediction of the non-canonical intermediate inflationary model in anisotropy background for the specified values of $\alpha=3$, $\kappa=3.02\times 10^{-12}$, $f=10^{-4}$ ,$M_p=10^{18}$, $M=10^{12}$ and $\lambda=3.5$, in comparison with the observational results of Planck 2015. Where the likelihood of Planck $2013$ (grey contours), Planck $2015$ TT+lowP(red contours), Planck $2015$ TT,TE,EE+lowP(blue contours) and the thick black line indicate the predictions of our model in which small and large dots are the amount of $n_s$ at the e-folding value $N=55,~N=65$.}}} \label{fignsdns1} \end{figure} \newpage \section{Conclusions}\label{seccon} A well-known class of scale factors namely the intermediate ones for a non-canonical Lagrangian in the anisotropic background has been investigated. The main motivation of doing such investigation goes back to cope with drawbacks of canonical and isotropic version of inflationary scenarios. Despite some complications in formulas and calculations, have been raised because of extension in the model, fortunately and without any hand made conditions it has bee shown that the obtained potential automatically takes a steep form, i.e. $V = V_0 \: \phi^{s}$ with $s<0$. This class of potentials, as it has been shown, can be considered as a suitable candidate to run the inflation in an acceptable process, compared to observational constraints. To examine our proposal we have been followed the slow-roll method and all necessary parameters have been estimated based on a powerful criterion such as Planck $2015$. Amongst those aforementioned observables we have been focused on the amplitudes of scalar and tensor perturbations, their ratio, scalar and tensor spectral indices and their running as well. So by combining resulted potential and slow-rolling approach we have been tried to examine the accuracy of our estimations and also the claims about the succusses of Non-canonical anisotropic model. It has been clear one of the most important results of Planck data is the $r-n_s$ diagram, and the validity of theoretical models rely upon their acceptable compatibility with this criterion. Whereas we have been obtained all the necessary instruments to examine the validity of our results we could plot them based on the original figures originated from the Planck collaboration papers, e.g. \cite{Planck2015}. Therefore based on our investigations for $r-n_s$ analyzes at first we have been supplied a diagram in Fig.\ref{fignsr} and it has been observed that non-canonical anisotropic inflation with an intermediate scalar field could be considered as a suitable candidate to drive inflation. And then, consequently in Table. \ref{FT} different asymptotical behaviour based on definitions for the Lagrangian and also BI metric and the comparability with $n_s$ and $\mathcal{P}_s$ appeared in Planck have been evaluated. The best free parameters have been obtained as $\alpha=3$, $\lambda=3.5$ and $f=10^{-4}$ in which we have been used some specific values for other parameters like the $\kappa=3.02\times 10^{-12}$ ,$M_p=10^{18}$, and $M=10^{12}$. To visualize the aforementioned asymptotical behaviour at first we have been considered $\lambda=1$ to go back to isotropic background. From this point of view it has been concluded that even for the well accepted non-canonical Lagrangian the results in the isotropic universe have some deviations compared to data. Even more the situation could be absolutely teerible for canonical Lagrangian even in anisotropic background. Besides, in Fig.\ref{fignsdns1}, the predictions for the running spectral index have been appeared also in acceptable ranges comparing with observational data, that has been relied insides the joint 68\% CLs region of Planck 2015 TT, TE, EE+lowP data \cite{Planck2015}. \section{Acknowledgement} The authors thank the anonymous referee for his/her useful comments and suggestions have resulted in an improved version of their manuscript. HS would like to appreciate IPM, and specially H. Firouzjahi, for their hospitality and constructive discussions during his visit of there. Also he is grateful ICTP, during Summer School 2018, to give him constructive ideas about inflation and primordial fluctuations. He appreciates G. Ellis, A. Weltamn and UCT to arrange his short visit of there and good discussions about primordial universe. He is also grateful his wife E. Avirdi for her valuable notes and being patience during our stay in South Africa.

ab9e5c41f075def0f880fa10341ed70df26613c0

\section{\label{Sec: Introduction} Introduction} The cochlea is a highly sensitive device that is capable of sensing sound waves across a broad spectrum of frequencies ($20-20000 \si{Hz}$) and across a wide range of sound intensities ranging from $0 \si{dB}$ (threshold of hearing) up to $120 \si{dB}$ (sound of a jet engine). The cochlea was believed to be a passive device that acts like a Fourier analyzer: each frequency causes a vibration at a particular location on the basilar membrane (BM). This mechanism was discovered by the Nobel Prize winner George von B\'ek\'esy who carried out his experiments on cochleae of human cadavers. However, in 1948, Thomas Gold hypothesized that the ear is rather an active device that has a component termed the cochlear amplifier. Although Gold's hypothesis was rejected by von B\'ek\'esy, David Kemp validated it thirty years later by measuring emissions from the ear. These emissions, termed otoacoustic emissions (OAEs) are sound waves that are produced by the cochlea and can be measured in the ear canal. It is widely accepted that the outer hair cells, anchored on the cochlear partition, are responsible for the active gain in the cochlea that produces these emissions. However, the underlying mechanism is still not well understood. For example, spontaneous otoacoustic emissions (SOAEs) -- emissions generated in the absence of any stimulus -- are studied in \cite{fruth2014active} and \cite{ku2008statistics}. The remarkable high sensitivity of the cochlea makes it vulnerable to stochastic perturbations that are believed to be the cause of these emissions. Particularly, in \cite{ku2008statistics}, the authors studied the instabilities that arise in a linear biomechanical cochlear model with spatially random active gain profiles that are static in time. In \cite{fruth2014active}, similar analysis was carried out on simplified cochlear models comprised of coupled active nonlinear oscillators. The randomness, or disorder, was introduced via static variations of a bifurcation parameter. In these previous works, the analysis was carried out through Monte Carlo simulations by studying the stability of different randomly generated active gain (or bifurcation) profiles. In this paper, we carry out a \textit{simulation-free} stability analysis of the linearized dynamics of a nonlinear model of the cochlea. Our analysis employs structured stochastic uncertainty theory (\cite{bamieh2018structured}, \cite{filo2018structured}, \cite{lu2002mean}, \cite{elia2005remote}) rather than Monte Carlo simulations, where the active gain is stochastic in space and time and may have a spatially-varying expectation and/or covariance. It turns out that letting the active gain be a stochastic process puts the model in a standard setting of linear time-invariant (LTI) systems in feedback with a diagonal stochastic process that enters the dynamics multiplicatively (see Figure~\ref{Fig: Feedback Block Diagram}). This analysis allows us to predict the locations on the BM where the dynamics are more likely to destabilize due to the underlying uncertainties. It also provides a bound on the variance of the perturbations allowed such that stability is maintained. The rest of the paper is organized as follows: we start by providing a brief description of a class of biomechanical models of the cochlea in section \ref{Section: Brief Model Description}. Then, in section \ref{Section: DSS Formulation}, we recast this class of models in a descriptor state space (DSS) form using operator language (i.e. in continuous space-time). In section \ref{Section: Active Gain Uncertainties}, we reformulate the DSS form in a standard setting that is particularly useful to carry out our stochastic uncertainty analysis. We also provide the conditions for mean-square stability (MSS). In section \ref{Section: Cochlear Instabilities}, we present the numerical results of the possible instabilities caused by stochastic gain profiles with different statistic properties. To validate our analysis, we show a stochastic simulation for the full nonlinear model in section~\ref{Section: Simulations}. Finally, before we conclude, we give a discussion in section~\ref{Section: Discussion} to give a physical interpretation of our results and provide some comments on previous works. \section{Biomechanical Model of the Cochlea} Throughout the literature, cochlear modeling attempts varied depending on two main factors. The first is concerned with the degree of biological realism of the mathematical model. This is realized by the incorporation of various biological structures (\cite{geisler1995cochlear}, \cite{lamar2006signal}, \cite{neely1986model}) and the dimensionality of the fluid filling the cochlear chambers (\cite{steele1979comparison}, \cite{givelberg2003comprehensive}). The second factor is concerned with the computational aspect of the models. Different numerical methods were devised to approach the spatio-temporal nature of the cochlea (\cite{neely1981finite}, \cite{elliott2013wave}). Particularly, \cite{elliott2007state} used a finite difference method developed in \cite{neely1981finite} to discretize space and formulate the model in state space form. Moreover, computationally efficient methods and model reduction techniques were developed for fast simulations of cochlear response (\cite{bertaccini2011fast}, \cite{filo2016order}). This section starts by describing the mathematical model adopted in this paper. Then, we reformulate the latter in a continuous space-time descriptor state space form, using operator language. This form has two advantages: (a) it encompasses a wider class of cochlear models and (b) it makes the dynamics more transparent by treating the exact model and its finite dimensional approximation (i.e. discretizing space by some numerical method) separately \cite{filo2017topics}. \subsection{Mathematical Model Description} \label{Section: Brief Model Description} The mathematical model can be divided into two main blocks as illustrated in Figure~\ref{Fig: Block Diagram of the Ear}(a). For a detailed derivation of the governing mechanics, refer to \cite{elliott2007state} and \cite{filo2016order} for a one and two dimensional modeling of the fluid stage, respectively. \begin{figure}[h] \centering \begin{tabular}{l} \includegraphics[scale = 0.7]{Figures/EarBlockDiagram.pdf} \\ \footnotesize{(a) Block Diagram of the Cochlea} \\ \includegraphics[scale = 0.7]{Figures/MicroMechanics.pdf} \\ \footnotesize{(b) Detailed Schematic Representing the Membranes Block } \end{tabular} \caption{\footnotesize{(a) The cochlea processes the acceleration of the stapes $\ddot s(t)$, in two stages, to produce the vibrations at every location of the BM, $u(x,t)$. The first stage is governed by the fluid that is stimulated by both the stapes and BM accelerations to yield a pressure $p(x,t)$ acting on every location of the BM. The second stage is governed by the dynamics of the membranes. The two stages are in feedback through the BM acceleration. (b) This figure is a schematic of a cross section (at a location $x$) of the cochlear partition showing the membranes governing the dynamics of the micro-mechanical stage. The spatially varying parameters $m_i$, $c_i(x)$ and $k_i(x)$ are the mass, damping coefficient and stiffness of the BM and TM for $i = 1$ and $2$, respectively. Furthermore, $c_3(x)$ and $k_3(x)$ are the mutual damping coefficient and stiffness, respectively; while $c_4(x)$ and $k_4(x)$ are the damping coefficient and stiffness associated with the active feedback gain from the outer hair cells (OHC) to the BM. The spring and damper between the BM and the OHC have variable negative values to capture the effect of the active force acting only on the BM without any direct effect on the TM. Their values depend on the the BM displacement $u$ via the nonlinear gain $\mathcal G(u)$. Equation(\ref{Eqn: Micromechanics}) describes the underlying dynamics.}} \label{Fig: Block Diagram of the Ear} \end{figure} The fluid block, commonly referred to as the macro-mechanical stage, is linear and memoryless under the appropriate assumptions and approximations (refer to Appendix-\ref{Section: Mass Operators}). This block introduces spatial coupling along the different locations on the BM. Its output is the pressure $p(x,t)$ acting on each location of the BM. The governing equation can be written as a general expression, regardless of the dimensionality of the fluid and the numerical method used, as \begin{equation} \label{Eqn: BM Pressure} p(x,t) = -[\mathcal M_f \ddot u](x,t) - [\mathcal M_s \ddot s](t), \end{equation} where $\ddot{ }$ represents the second time derivative operation, and $\mathcal M_f$ and $\mathcal M_s$ are linear spatial operators associated with the fluid and stapes mass, respectively. Refer to Appendix-\ref{Section: Mass Operators} for a more detailed discussion of these mass operators and their finite dimensional approximations as matrices $M_f$ and $M_s$, respectively. The second block, commonly referred to as the micro-mechanical stage, takes the distributed pressure $p(x,t)$ as an input to produce the BM vibrations $u(x,t)$ at every location according to the following differential equations \begin{equation} \label{Eqn: Micromechanics} \begin{aligned} \begin{bmatrix} \frac{g}{b} m_1 & 0 \\ 0 & m_2 \end{bmatrix} \begin{bmatrix} \ddot u \\ \ddot v \end{bmatrix} + \begin{bmatrix} \frac{g}{b}(c_1 + c_3 - \mathcal G(u)c_4) & \mathcal G(u) c_4 - c_3 \\ -\frac{g}{b}c_3 & c_2 + c_3 \end{bmatrix} \begin{bmatrix} \dot u \\ \dot v \end{bmatrix} \\ + \begin{bmatrix} \frac{g}{b}(k_1 + k_3 - \mathcal G(u)k_4) & \mathcal G(u) k_4 - k_3 \\ -\frac{g}{b}k_3 & k_2 + k_3 \end{bmatrix} \begin{bmatrix} u \\ v \end{bmatrix} = \begin{bmatrix} p \\ 0 \end{bmatrix}, \end{aligned} \end{equation} where $v(x,t)$ is the tectorial membrane (TM) vibration (refer to Figure~\ref{Fig: Block Diagram of the Ear}(b)). Note that the space and time variables $(x,t)$ are dropped where necessary for notational compactness. The constant $b$ is the ratio of the average to maximum vibration along the width of the BM, and $g$ is the BM to outer hair cells lever gain. Refer to \cite{neely1986model} for a detailed explanation of the parameters. Finally, $\mathcal G$ is the nonlinear active gain operator that captures the active nature of the outer hair cells, commonly referred to as the cochlear amplifier. In the spirit of \cite{lamar2006signal}, the action of $\mathcal G$ on a distributed BM displacement profile $u$ is given by \begin{equation} \label{Eqn: Nonlinear Gain} \left[\mathcal G(u)\right](x,t) = \frac{\gamma(x)}{1 + \theta \left[\Phi_\eta\left(\frac{u^2}{R^2}\right)\right](x,t)}, \end{equation} where the gain coefficient $\gamma(x)$ represents the gain at a location $x$, in the absence of any stimulus ($u(x,t) = 0$). The constants $\theta$ and $R$ are the nonlinear coupling coefficient and BM displacement normalization factor, respectively. The operator $\Phi_{\eta}$ is a normalized Gaussian operator such that its action on $u$ is defined as \begin{align} [\Phi_\eta(u)](x,t) &:= \frac{\int_0^L \phi_{\eta}(x-\xi) u(\xi,t) d\xi}{\int_0^L \phi_{\eta}(x-\xi)d\xi}; \label{Eqn: Guassian Weighing Operator}\\ \phi_{\eta}(x) &:= \frac{1}{\eta\sqrt{2\pi}}e^{\frac{-x^2}{2\eta^2}}\label{Eqn: Gaussian Kernel}, \end{align} where $L$ is the length of the BM and $\phi_{\eta}$ is the Gaussian kernel with a width $\eta$. Note that $\eta = 0.5345 \si{mm}$ corresponds to the equivalent rectangular bandwidth on the BM (refer to Appendix-\ref{Section: ERB} for a detailed explanation). Observe that the spatial coupling in the micro-mechanical stage appears only in the nonlinear active gain (\ref{Eqn: Nonlinear Gain}). \subsection{Deterministic Descriptor State Space Formulation of the Linearized Dynamics in Continuous Space-Time} \label{Section: DSS Formulation} This section gives a Descriptor State Space (DSS) formulation of the cochlear model described in (\ref{Eqn: BM Pressure}) and (\ref{Eqn: Micromechanics}). The DSS form is given for the linearized dynamics around the only fixed point which is the origin. It can be shown (Appendix-\ref{Section: System Linearization}) that the linearized dynamics can be achieved by simply replacing the nonlinear active gain $[\mathcal G(u)](x,t)$ in (\ref{Eqn: Micromechanics}) by its gain coefficient $\gamma(x)$. First, define the state space variable $\psi(x,t)$ in continuous space-time as \begin{equation} \label{Eqn: State Space Variable} \psi(x,t) := \begin{bmatrix} u(x,t) & v(x,t) & \dot u(x,t) & \dot v(x,t) \end{bmatrix}^T. \end{equation} Then the DSS form of the linearized dynamics is \begin{equation} \label{Eqn: Descriptor State Space Operator Form} \begin{aligned} \mathcal E \frac{\partial}{\partial t} \psi(x,t) &= \mathcal A_{\gamma} \psi(x,t) + \mathcal B \ddot s(t) \\ u(x,t) &= \mathcal C \psi(x,t), \end{aligned} \end{equation} where $\mathcal E$, $\mathcal A_{\gamma}$ and $\mathcal B$ are matrices of linear spatial operators defined as follows \begin{align*} &\mathcal E := \begin{bmatrix} \mathcal I & 0 & 0 & 0 \\ 0 & \mathcal I & 0 & 0 \\ 0 & 0 & \frac{g}{b}m_1 \mathcal I + \mathcal M_f & 0 \\ 0 & 0 & 0 & m_2 \mathcal I \\ \end{bmatrix}; \quad \mathcal B := \begin{bmatrix} 0 \\ 0 \\ -\mathcal M_s \\ 0 \end{bmatrix}; \\ &\mathcal A_{\gamma} := \mathcal A_0 + \mathcal B_0 \gamma \mathcal C_0; \qquad \mathcal C := \begin{bmatrix} \mathcal I & 0 & 0 & 0 \end{bmatrix}; \\ &\Scale[0.97]{\mathcal A_0 := \resizebox{0.92\hsize}{!}{$ \begin{bmatrix} 0 & 0 & \mathcal I & 0\\ 0 & 0 & 0 & \mathcal I\\ -\frac{g}{b}(k_1+k_3) & k_3 & -\frac{g}{b}(c_1+c_3) & c_3\\ \frac{g}{b}k_3 & -(k_2+k_3) & \frac{g}{b}c_3 &-(c_2+c_3) \end{bmatrix}$};} \\ &\mathcal B_0 := \begin{bmatrix} 0 & 0 & \mathcal I & 0 \end{bmatrix}^T; \qquad \mathcal C_0 := \begin{bmatrix} \frac{g}{b}k_4 & -k_4 & \frac{g}{b}c_4 & -c_4 \end{bmatrix}; \end{align*} and $\mathcal I $ is the identity operator. The equations in (\ref{Eqn: Descriptor State Space Operator Form}) represent a deterministic evolution differential equation and an output equation that provides the distributed displacement of the BM $u(x,t)$. Other outputs can be selected, such as the TM displacement, by appropriately constructing the $\mathcal C$ operator. In the subsequent section, we slightly modify the dynamical equations to account for stochastic perturbations in the gain coefficient $\gamma(x)$. \section{Stochastic Uncertainties in the Active Gain} \label{Section: Active Gain Uncertainties} This section investigates the Mean Square Stability (MSS, which we will formally define in section \ref{Section: Stochastic Feedback Interconnection}) of the linearized cochlear dynamics when the gain coefficient is a \textit{spatio-temporal stochastic process}. The stochastic gain coefficient, now denoted by $\gamma(x,t)$ to account for spatio-temporal perturbations, enters the dynamics (\ref{Eqn: Descriptor State Space Operator Form}) multiplicatively. We first reformulate the dynamics as an LTI system in feedback with a diagonal stochastic gain which is a standard setting in robust control theory\cite[Section 10.3]{zhou1996robust}. Then we carry out our MSS analysis based on \cite{filo2018structured}. By tracking the evolution of the instantaneous spatial covariances, MSS analysis allows us to predict the locations on the BM that are more likely to become unstable due to the underlying stochastic uncertainty. We conclude this section by defining and analyzing a linear operator, whose spectral radius provides a condition for MSS. \subsection{Stochastic Feedback Interconnection} \label{Section: Stochastic Feedback Interconnection} The purpose of this section is to separate the stochastic portion of the gain coefficient in a feedback interconnection. We assume that $\gamma(x,t)$ is a \textit{spatio-temporal stochastic process} that is white in time (but may be colored in space), and whose expectation and covariance are independent of time. More precisely, let $\bar \gamma(x)$ be the expectation of $\gamma(x,t)$ and $\tilde \gamma(x,t)$ be a temporally independent, zero mean stochastic perturbation, such that \begin{equation} \label{Eqn: Stochastic Gain} \begin{aligned} \gamma(x,t) &= \bar \gamma(x) + \epsilon \tilde \gamma(x,t), \\ \text{with } & \left\{ \begin{aligned} &\mathbb E[\gamma(x,t)] = \bar \gamma(x) \\ &\mathbb E[\tilde \gamma(x,t) \tilde \gamma(\xi,\tau)] = \mathbf \Gamma(x,\xi) \delta(t-\tau) \end{aligned}\right. ~~ \forall t\geq 0, \end{aligned} \end{equation} where $\mathbb E[.]$ denotes the expectation, $\epsilon$ is a perturbation parameter, $\delta(t)$ is the Dirac Delta function, and $\mathbf \Gamma(x,\xi)$ is a positive semi-definite covariance kernel. Substituting (\ref{Eqn: Stochastic Gain}) in (\ref{Eqn: Descriptor State Space Operator Form}) yields \begin{equation} \label{Eqn: Stochastic Feedback} \begin{aligned} \mathcal E \frac{\partial}{\partial t}\psi(x,t) &= (\mathcal A_{\bar \gamma} + \epsilon \mathcal B_0 \tilde \gamma \mathcal C_0) \psi(x,t) + \mathcal B \ddot s(t)\\ u(x,t) &= \mathcal C \psi(x,t). \end{aligned} \end{equation} The evolution equation in (\ref{Eqn: Stochastic Feedback}) is a Stochastic Partial Differential Equation (SPDE) that is given an It\=o interpretation in the time variable. For more details on It\=o calculus, refer to \cite{oksendal2003stochastic}. Define a secondary output related to the difference in BM and TM displacements and velocities as \begin{equation} \label{Eqn: Secondary Output} \begin{aligned} y(x,t) &:= \epsilon \mathcal C_0 \psi(x,t). \end{aligned} \end{equation} Furthermore, define the active feedback pressure resulting from the stochastic perturbations to be \begin{equation} \label{Eqn: Feedback Pressure} p_a(x,t) := \tilde \gamma(x,t) y(x,t). \end{equation} Therefore, using (\ref{Eqn: Stochastic Feedback}), (\ref{Eqn: Secondary Output}) and (\ref{Eqn: Feedback Pressure}), construct the feedback block diagram depicted in Figure \ref{Fig: Feedback Block Diagram}. \begin{figure}[h] \centering \includegraphics[scale = 0.9]{Figures/FeedbackDiagram.pdf} \caption{\footnotesize{The linearized cochlear model in feedback with multiplicative stochastic gain. The block to the top represents the deterministic portion of the linearized cochlear dynamics casted in a descriptor state space form. The feedback block is a diagonal spatial operator that represents the multiplicative stochastic gain. $y(x,t)$ is the differential vibration and velocity between the BM and TM as given by (\ref{Eqn: Secondary Output}). $p_a(x,t)$ is the active pressure that results from the stochastic component of the active gain.}} \label{Fig: Feedback Block Diagram} \end{figure} This is a standard setting(\cite{filo2018structured} for structured stochastic uncertainty analysis, where the feedback gain is a diagonal spatial operator. This configuration is used to investigate the MSS of the cochlea which is formally defined next. \textit{Definition}: The feedback system in Figure~\ref{Fig: Feedback Block Diagram} is MSS if, in the absence of an input (i.e. $\ddot s(t) = 0$), the state $\psi(x,t)$ and the active feedback pressure $p_a(x,t)$ have bounded variances for all time. Therefore, to study MSS, we need to track the temporal evolution of the variances and look at their steady state limits as $t$ goes to $+\infty$. This is the topic of the next subsection. \subsection{Temporal Evolution of the Covariance Operators} \label{Section: Covariance Evolution} This section tracks the time evolution of the covariance operators in the absence of any input (i.e. we set $\ddot s(t) = 0$ for the rest of the paper). We use the term covariance ``operators" rather than covariance matrices because the spatial variables $x$ and $\xi$ are continuous. After using some numerical method to discretize space, the covariance operators can be approximated by covariance matrices. With slight abuse of notation, we use the same symbol to denote the covariance operator and its associated kernel. Define the following instantaneous spatial covariance kernels \begin{equation} \label{Eqn: Covariance Definitions} \begin{aligned} \mathcal X(x,\xi;t) &:= \mathbb E[\psi(x,t) \psi(\xi,t)] \\ \mathcal Y(x,\xi;t) &:= \mathbb E[y(x,t) y(\xi,t)] \\ \mathcal P(x,\xi;t) &:= \mathbb E[p_a(x,t) p_a(\xi,t)] \\ \mathcal U(x,\xi;t) &:= \mathbb E[u(x,t) u(\xi,t)] \\ \mathbf \Gamma(x,\xi) &:= \mathbb E[\tilde \gamma(x,t) \tilde \gamma(\xi,t)] \qquad \forall t\geq 0. \end{aligned} \end{equation} Given that the stochastic perturbations $\tilde \gamma$ are temporally independent, it can be shown \cite[Section V]{filo2018structured} that the time evolution of the covariance operators are governed by the following operator-valued, \textit{differential algebraic} equations \begin{equation} \label{Eqn: Covariance Evolution Equations} \begin{aligned} \mathcal E\dot {\mathcal X} \mathcal E^* &= \mathcal A_{\bar \gamma} \mathcal X \mathcal E^* + \mathcal E \mathcal X \mathcal A_{\bar \gamma}^* + \mathcal B_0 \mathcal P \mathcal B_0^* \\ \mathcal Y &= \epsilon ^2 \mathcal C_0 \mathcal X \mathcal C_0^*\\ \mathcal P &= \mathbf \Gamma \circ \mathcal Y, \end{aligned} \end{equation} where $*$ is the adjoint operation and $\circ$ is the Hadamard product; i.e. the element-by-element multiplication of the kernels $\mathcal P(x,\xi;t) = \mathbf \Gamma(x,\xi) \mathcal Y(x,\xi;t)$. In order to study the MSS, we need to look at the steady state limit of the covariances. We denote by the asymptotic limit of a covariance operator, when it exists, by an overbar. That is \begin{equation} \label{Eqn: Covariance Limits} \begin{aligned} \bar {\mathcal X} := \lim_{t\to\infty} \mathcal X(t); ~~ \bar {\mathcal Y} := \lim_{t\to\infty} \mathcal Y(t); ~~ \bar {\mathcal P} := \lim_{t\to\infty} \mathcal P(t) . \end{aligned} \end{equation} At the steady state, the covariances become constant in time and thus their time derivatives go to zero. Hence, the steady state covariances, if they exist, are governed by the following operator-valued \textit{algebraic} equations: \begin{equation} \label{Eqn: State State Covariance Equations} \begin{aligned} &\mathcal A_{\bar \gamma} \bar{\mathcal X} \mathcal E^* + \mathcal E \bar{\mathcal X} \mathcal A_{\bar \gamma}^* + \mathcal B_0 \bar{\mathcal P} \mathcal B_0^* = 0 \\ &\bar{\mathcal Y} = \epsilon^2 \mathcal C_0 \bar{\mathcal X} \mathcal C_0^*\\ &\bar{\mathcal P} = \mathbf \Gamma \circ \bar{\mathcal Y}. \end{aligned} \end{equation} In the next section, we will use (\ref{Eqn: State State Covariance Equations}) to define a new operator as a tool to check the boundedness of the steady state covariances. \subsection{Loop Gain Operator \& MSS} Using (\ref{Eqn: State State Covariance Equations}), define the loop gain operator $\mathbb{L}_{\mathbf \Gamma}$, parametrized by the perturbation covariance $\mathbf \Gamma$, as \begin{equation} \label{Eqn: Loop Gain Operator} \begin{aligned} \mathbb{L}_{\mathbf \Gamma}(\bar{\mathcal P}_{\text{in}}) = \bar{\mathcal P}_{\text{out}} \Longleftrightarrow \left\{ \begin{aligned} &\bar{\mathcal P}_{\text{out}} = \mathbf \Gamma \circ (\mathcal C_0 \bar{\mathcal X} \mathcal C_0^*) \\ &\mathcal A_{\bar \gamma} \bar{\mathcal X} \mathcal E^* + \mathcal E \bar{\mathcal X} \mathcal A_{\bar \gamma}^* + \mathcal B_0 \bar{\mathcal P}_{\text{in}} \mathcal B_0^* = 0. \end{aligned} \right. \end{aligned} \end{equation} The MSS condition is given in terms of the spectral radius of the loop gain operator as explained next. \textit{\textbf{Theorem}}: \textit{Consider the system in Figure~\ref{Fig: Feedback Block Diagram} where $\tilde \gamma$ is a temporally independent multiplicative noise, interpreted in the sense of It\=o, with instantaneous spatial covariance $\mathbf \Gamma$, and $\mathcal M$ is a stable causal LTI system. The feedback system is MSS if and only if the spectral radius of the loop gain operator is strictly less than one, i.e. \begin{equation} \label{Eqn: MSS Condition} \epsilon^2 \rho(\mathbb L_{\mathbf \Gamma}) < 1, \end{equation} where $\mathbb L_{\mathbf \Gamma}$ is defined in (\ref{Eqn: Loop Gain Operator}) and $\rho(\mathbb L_{\mathbf \Gamma})$ is its spectral radius.} The proof of this theorem is given in \cite{filo2018structured}. This theorem will be used to find an upper bound on the perturbation constant $\epsilon$ above which MSS is violated. \subsection{Worst-Case Covariances} The loop gain operator maps a covariance operator $\bar{\mathcal P}_{\text{in}}$ into another covariance operator $\bar{\mathcal P}_{\text{out}}$. Hence, the eigenvectors of $\mathbb L_{\mathbf \Gamma}$ are themselves operators. When a finite dimensional approximation of $\mathbb L_{\mathbf \Gamma}$ is carried out using some numerical method, these eigenvectors can be approximated as matrices. We are particularly interested in the eigenvector (or eigen-operator) of $\mathbb L_{\mathbf \Gamma}$ associated with the largest eigenvalue because it has a significant meaning explained in this subsection. First, since the loop gain operator is a monotone operator \cite{bamieh2018structured}, it is guaranteed to have a real largest eigenvalue equal to $\rho(\mathbb L_{\mathbf \Gamma})$. It is also guaranteed that the eigen-operator associated with the largest eigenvalue is positive semidefinite, i.e. there exists a positive semidefinite covariance operator $\textbf{P}$ such that \begin{equation} \label{Eqn: Worst Case Covariance} \mathbb L_{\mathbf \Gamma}(\textbf{P}) = \rho(\mathbb L_{\mathbf \Gamma}) \textbf{P}. \end{equation} Note that $\textbf{P}$ is the operator counterpart of the Perron-Frobenius eigenvector for matrices with non-negative entries. Refer to \cite[Thm 2.3]{bamieh2018structured} for a proof of the aforementioned guarantees. If the stability condition (\ref{Eqn: MSS Condition}) is violated, $\textbf{P}$ will be the covariance mode that has the highest growth rate, hence the name ``worst-case" covariance. This provides information about the locations on the BM that are more likely to destabilize due to the stochastic perturbations of the gain. Particularly, since we are interested in the instabilities at the BM, the worst-case covariance of the BM vibrations, denoted by $\mathbf U$, can be computed by propagating the worst-case pressure covariance $\textbf{P}$ through the cochlear dynamics (at steady state) as follows \begin{equation} \label{Eqn: BM Worst Case Covariance} \begin{aligned} &\mathcal A_{\bar \gamma} \textbf{X} \mathcal E^* + \mathcal E \textbf{X} \mathcal A_{\bar \gamma}^* + \mathcal B_0 \textbf{P} \mathcal B_0^* = 0 \\ &\textbf{U} = \mathcal C \textbf{X} \mathcal C^*, \end{aligned} \end{equation} where $\textbf{X}$ denotes the worst-case covariance operator corresponding to the state space variable $\psi$. \section{Instabilities in Linearized Cochlear Dynamics} \label{Section: Cochlear Instabilities} This section contains the main results on the effects of stochastic uncertainties on cochlear instabilities. The analysis is carried out for three different scenarios of the perturbation covariance $\mathbf \Gamma(x,\xi)$: \begin{itemize} \item $\textbf{S}_1$: spatially uncorrelated uncertainties, i.e. ${\mathbf \Gamma(x,\xi) = \delta(x-\xi)}$ \item $\textbf{S}_2$: spatially correlated uncertainties with a correlation length $\lambda$, i.e. $\mathbf \Gamma(x,\xi) = \phi_{\lambda}(x-\xi)$ \item $\textbf{S}_3$: spatially localized and uncorrelated uncertainties, i.e. $\mathbf \Gamma(x,\xi) = \phi_{\sigma}(x-\mu) \delta(x-\xi)$, \end{itemize} where $\phi_\lambda$ and $\phi_{\sigma}$ are the Gaussian kernels defined in (\ref{Eqn: Gaussian Kernel}) such that $\lambda$ is the spatial correlation length and $\sigma$ is the spatial localization length. In the subsequent analysis, scenarios $\textbf{S}_1$ and $\textbf{S}_2$ are treated simultaneously because, in both cases, the perturbation covariance is a Toeplitz operator since $\mathbf \Gamma(x,\xi)$ depends solely on the difference $x-\xi$ rather than the absolute locations $x$ and $\xi$. However, in scenario $\textbf{S}_3$, the perturbation covariance is spatially localized and $\mathbf \Gamma(x,\xi)$ depends on the absolute locations, and thus it is treated separately in subsection \ref{Section: Spatially Variant Covariance}. Recall that the linearized cochlear dynamics excludes micro-mechanical spatial coupling along different locations of the BM; whereas, scenario $\textbf{S}_2$ sort of reintroduces spatial coupling via the spatial correlations of the stochastic active gain. The condition of MSS (\ref{Eqn: MSS Condition}) can be rewritten as \begin{equation} \label{Eqn: MSS Condition for S1, S2 and S3} \begin{aligned} \epsilon &< \frac{1}{\sqrt{\rho(\mathbb L_{\mathbf{\Gamma}})}}, \end{aligned} \end{equation} for scenarios $\mathbf S_1, \mathbf S_2$ and $\mathbf S_3$. This bound is the maximum allowed perturbation in (\ref{Eqn: Stochastic Feedback}) such that MSS is maintained. In this section, we compute the upper bound on $\epsilon$ and the ``worst-case" covariance $\textbf{U}$ for the linearized cochlear dynamics. \subsection{Numerical Considerations} This section describes the numerical considerations of the model and the numerical method used to compute the spectral radius and worst-case covariance of $\mathbb L_{\mathbf \Gamma}$. The numerical values of the parameters in this paper are taken from Table~I in \cite{ku2008statistics} for the linear cochlea. However, the expectation of the gain coefficient, $\bar \gamma(x)$, (which was considered to be spatially constant in \cite{ku2008statistics}) is left as a spatially distributed parameter to be tuned. The fluids block in Figure~\ref{Fig: Block Diagram of the Ear}(a) considered here is the one dimensional traveling wave as described in Appendix-\ref{Section: Mass Operators}. A spatial discretization grid of step size $\Delta_x := L/N_x$, where $N_x = 400$, is used to give a finite dimensional approximation of the operators (as matrices) describing the dynamics in Figure~\ref{Fig: Feedback Block Diagram} (refer to Appendix-\ref{Section: Finite Realizations}). Special care has to be taken when dealing with spatially white continuous processes (Scenario $\textbf{S}_1$). Let $\Gamma$ denote a matrix approximation of the uncertainty covariance operator $\mathbf \Gamma$ and approximate the Dirac delta function as \begin{equation} \label{Eqn: Delta Function} \begin{aligned} \delta(x) &\approx \frac{1}{\Delta_x} \text{rect}_{\Delta_x}(x) \\ \text{such that}, \quad &\text{rect}_{\Delta_x} := \begin{cases} 1, & \text{if}\ -\frac{\Delta_x}{2} \leq x \leq \frac{\Delta_x}{2} \\ 0, & \text{otherwise} \end{cases}. \end{aligned} \end{equation} Hence, the finite dimensional approximation of the perturbation covariance needs to be scaled with the discretization step $\Delta_x$ as follows \begin{equation} \label{Eqn: Discretized Covariance} \Gamma = \frac{1}{\Delta_x}I, \end{equation} where $I$ is the identity matrix. Furthermore, our analysis requires the computation of the largest eigenvalue of the loop gain operator and its associated eigenvector (or eigen-operator). The matrices that approximate the spatial operators have a size of ($4(N_x+1) = 1604$), and keeping track of the underlying sparsity of all the approximated operators is essential for carrying out the computations efficiently. Note that to maintain the sparsity of (\ref{Eqn: Loop Gain Operator}) for scenario $\textbf{S}_2$, we use a truncated Gaussian kernel to approximate $\phi_{\lambda}$ given in (\ref{Eqn: Gaussian Kernel}), i.e. $ \phi_\lambda(x-\xi) \approx 0, \text{ for } |x-\xi| > d$, where $d$ is a pre-specified constant that represents a compromise between computational accuracy and sparsity. Finally, the power iteration method is employed for eigenvalue and eigenmatrix computations as recommended by \cite{parrilo2000cone}. This requires solving the Lyapunov-like equation in (\ref{Eqn: Loop Gain Operator}) at each iteration. \subsection{Stochastic Gain Coefficient with a Spatially Constant Expectation} \label{Section: Constant Expectation} In this section, we set the expectation of the gain coefficient to one everywhere along the BM, i.e. $\bar \gamma(x) = 1$. To study the effects of the spatial correlations in the gain coefficient, we compare scenarios $\mathbf S_1$ and $\mathbf S_2$ by keeping in mind that $\mathbf S_1$ can be seen as a special case of $\mathbf S_2$ at the limit when $\lambda$ goes to zero. First, we compute the upper bounds on $\epsilon$ in (\ref{Eqn: MSS Condition for S1, S2 and S3}) such that MSS is maintained. Then we compute the worst-case covariance $\textbf{U}$ in (\ref{Eqn: BM Worst Case Covariance}). By applying the power iteration method on (\ref{Eqn: Worst Case Covariance}), we compute the spectral radii $\rho(\mathbb L_{\mathbf \Gamma})$ and their associated eigen-operators $\mathbf P$ for scenarios $\textbf{S}_1$ and $\textbf{S}_2$ with different correlation lengths $\lambda$. Then, (\ref{Eqn: MSS Condition for S1, S2 and S3}) yields the upper bounds on $\epsilon$. The results are illustrated in Figure~\ref{Fig: Sweeping lambda} showing the small upper bounds on $\epsilon$. This reflects the high sensitivity of the model to such stochastic perturbations. As one would expect, a larger correlation length $\lambda$ requires a larger perturbation to destabilize the linearized cochlea. \begin{figure}[h] \centering \includegraphics[scale = .25]{Figures/Sweeping_lambda_ConstantGain.pdf} \caption{\footnotesize{Mean Square Stability Curve: Upper bound on the perturbation parameter, $\epsilon$, of the stochastic gain (\ref{Eqn: Stochastic Gain}) whose expectation is $\bar \gamma(x) = 1$. The black dot corresponds to scenario $\textbf{S}_1$ (uncorrelated gain perturbations) and the solid black line corresponds to scenario $\textbf{S}_2$ (correlated gain perturbations) for different spatial correlation lengths $\lambda$. The figure shows that larger correlation lengths make the model more immune to stochastic perturbations.}} \label{Fig: Sweeping lambda} \end{figure} The eigen-operator $\textbf{P}$ computed by the power iteration method is the worst-case pressure covariance. The corresponding worst-case covariance of the BM displacement $\textbf{U}$ is then computed using (\ref{Eqn: BM Worst Case Covariance}). Figure~\ref{Fig: Sweeping lambda Covariance}(a) shows $\textbf{U}$ for scenario $\textbf{S}_1$, zoomed in for $0 \leq x,\xi \leq L/10$. The intensity plot shows two sets of axes. The first axis represents the location on the BM and the second represents the corresponding characteristic frequency at each location, calculated using the Greenwood location-to-frequency mapping \cite{greenwood1990cochlear}. Observe that the covariance is band limited and the diagonal entries are dominant near the stapes ($x = 0$). This shows that instabilities essentially occur at high frequencies. Figure~\ref{Fig: Sweeping lambda Covariance}(b) plots the diagonal entries of $\textbf{U}$ for scenarios $\textbf{S}_1$ and $\textbf{S}_2$ for different correlation lengths $\lambda$. \begin{figure}[h!] \centering \begin{tabular}{c} \includegraphics[scale = .25]{Figures/CovarianceUncorrelated_ConstantGain.pdf} \\ \footnotesize{(a) Worst-Case Covariance of BM Displacement $\textbf{U}(x,\xi)$} \\ \\ \includegraphics[scale = .25]{Figures/DiagonalEntries.pdf} \\ \footnotesize{(b) Diagonal Entries of $\textbf{U}$}\\ \\ \includegraphics[scale = .25]{Figures/Eigenfunctions_ConstantGain.pdf} \\ \footnotesize{(c) Dominant Eigenfunction of $\textbf{U}$} \end{tabular} \caption{\footnotesize{Figure (a) shows an intensity plot of the worst-case covariance $\textbf{U}$ for scenario $\textbf{S}_1$ (uncorrelated gain perturbation) zoomed in for $0 \leq x,\xi \leq 3.5\si{mm}$. The axes correspond to the physical location $x$ in mm on the BM and the corresponding characteristic frequency $f$ in kHz. Figure (b) shows the diagonal entries of $\textbf{U}$ for scenarios $\textbf{S}_1$ and $\textbf{S}_2$ for different correlation lengths $\lambda$. Figure (c) depicts the dominant eigenfunction of $\textbf{U}$ for the different cases indicating the insignificant effect of $\lambda$ on the shape of the dominant eigenfunctions.}} \label{Fig: Sweeping lambda Covariance} \end{figure} A smaller correlation length gives a slightly broader spectrum of unstable frequencies. However, for small $\epsilon$, the effect of the correlation length on the shape of the unstable BM modes is negligible. This is illustrated in Figure~\ref{Fig: Sweeping lambda Covariance}(c), where the dominant eigenfunction of $\textbf{U}$ is plotted for different cases. \subsection{Stochastic Gain Coefficient with a Spatially Varying Expectation} This section shows that the frequencies of instabilities (or, equivalently, the locations on the BM) can shift depending on the shape of the expectation of the gain coefficient $\bar \gamma(x)$. For illustration purposes, four different profiles of $\bar{\gamma}_0(x)$ are generated as \begin{equation} \label{Eqn: Gain Mean Profiles} \bar{\gamma}_0(x) = \frac{\tanh(x/10) + \beta}{\tanh(L/10) + \beta}, \end{equation} where $x$ and $L$ are expressed in mm and $\beta = 0,2,4$ and $6$. First, we show the MSS curves, similar to Figure~\ref{Fig: Sweeping lambda} for the four different profiles generated using (\ref{Eqn: Gain Mean Profiles}). Figure~\ref{Fig: Sweeping lambda beta}(b) clearly shows that the shape of $\bar \gamma(x)$ affects the margin of MSS. Particularly, the larger the dip in the gain coefficient, the higher $\epsilon$ needs to be to destabilize the linearized dynamics in the MSS sense. \begin{figure}[h] \centering \begin{tabular}{l} \includegraphics[scale = .25]{Figures/GainProfiles.pdf} \\ \footnotesize{(a) Gain Coefficient Expectation Profiles} \\ \includegraphics[scale = .25]{Figures/Sweeping_lambda_VariableGain.pdf}\\ \footnotesize{(b) Corresponding MSS Curves} \\ \includegraphics[scale = .25]{Figures/Sweeping_beta_Eigenfunctions.pdf}\\ \footnotesize{(c) Eigenfunctions for scenario $\mathbf S_1$} \\ \end{tabular} \caption{\footnotesize{Mean Square Stability Curves for different gain coefficient expectation profiles: Figure(a) shows four different profiles of $\bar{\gamma}(x)$ generated as examples of spatially varying gain coefficients using (\ref{Eqn: Gain Mean Profiles}). The same values of $\beta$ are used in figures (b) and (c). Particularly, Figure(b) shows the upper bound on the perturbation parameter $\epsilon$ for the corresponding profiles of $\bar \gamma(x)$ in Figure(a). The circles correspond to scenario $\textbf{S}_1$ (uncorrelated gain perturbations) and the solid lines correspond to scenario $\textbf{S}_2$ (correlated gain perturbations) for different spatial correlation lengths $\lambda$. Figure (c) shows the eigenfunctions of the worst-case covariance operator $\textbf{U}$ corresponding to the different profiles of $\bar \gamma(x)$. The peaks of the eigenfunctions shift consistently with the shape of the gain profiles.}} \label{Fig: Sweeping lambda beta} \end{figure} Since the correlation length for small values of $\epsilon$ has a negligible effect on the shape of the unstable modes as shown in Figure~\ref{Fig: Sweeping lambda Covariance}(c), we only present the worst-case covariances for scenario $\textbf{S}_1$. In fact, the correlation length only affects the margin of stability as illustrated in Figure~\ref{Fig: Sweeping lambda beta}(b). Figure~\ref{Fig: Sweeping lambda beta}(c) depicts the dominant eigenfunctions of $\textbf{U}$ for the four different profiles of $\bar \gamma(x)$. Clearly, the peaks of the unstable modes of the BM shift depending on the shape of $\bar \gamma(x)$. In fact, as the dip in $\bar \gamma(x)$ is increased, the peaks shift farther from the stapes resulting in instabilities of lower frequencies. \subsection{Stochastic Gain Coefficient with a Spatially Localized Covariance} \label{Section: Spatially Variant Covariance} We now treat the case where the gain coefficient $\gamma(x,t)$ in (\ref{Eqn: Stochastic Feedback}) has a spatially constant expectation, but spatially localized covariance given in scenario $\textbf{S}_3$, i.e. $$ \bar \gamma(x) = 1 \qquad \text{and} \qquad \mathbf \Gamma(x,\xi) = \phi_\sigma(x-\mu) \delta(x-\xi),$$ for different values of $\sigma$ and $\mu$. Observe that for this form of $\mathbf \Gamma(x,\xi)$, the covariance is localized around $\mu$. Hence, this section investigates the cochlear instabilities that emerge as a result of stochastic perturbations localized around a particular location on the BM. In particular, we are interested in tracking the unstable BM modes for different values of $\mu$ and $\sigma$, where $\mu$ is the location of the perturbation and $\sigma$ represents the local spread of the perturbation in the neighborhood of $\mu$. Following the same calculations of the previous sections, we compute the dominant eigenfunction of the worst-case covariance of the BM displacement $\mathbf U$ for different values of $\mu$ and $\sigma$. The results are depicted in Figure~\ref{Fig: Spatially Localized Covariance}. \begin{figure}[h] \centering \begin{tabular}{c} \includegraphics[scale = .25]{Figures/Sweeping_x0_sigma1.pdf} \\ (a) \\ \includegraphics[scale = .25]{Figures/Sweeping_x0_sigma2.pdf} \\ (b) \\ \includegraphics[scale = .25]{Figures/Sweeping_x0_sigma3.pdf} \\ (c) \end{tabular} \caption{\footnotesize{Eigenfunctions of the worst-case covariance operator $\textbf{U}$ for different localized gain coefficient perturbations. These figures show the dominant eigenfunctions of the worst-case covariance operators for three different values of $\mu$ and $\sigma$. Particularly, in each figure, we fix $\sigma$ and vary $\mu$. Each thin curve represents a particular uncertainty spread function $\phi_{\sigma}(x-\mu)$ (not drawn to scale in the vertical axis) and each thick curve (with the same color) represents the corresponding dominant eigenfunction of the worst-case covariance operator. This figure illustrates the ``basal shifting" observation that resembles the phenomenon of detuning. }} \label{Fig: Spatially Localized Covariance} \end{figure} Observe that localized perturbations of the active gain coefficient at some location $\mu$ of the BM causes instabilities in that neighborhood. Particularly, for relatively small spread $\sigma = L/100$, the instabilities emerge at the same locations of the perturbations as shown in Figure~\ref{Fig: Spatially Localized Covariance}(a). However, as the spread of the uncertainty is increased up to $\sigma = L/30$ and $ L/10$, the location of the instability shifts towards the stapes. In fact, the wider the spread the larger the shift is as illustrated in Figures~\ref{Fig: Spatially Localized Covariance}(b) and (c). This ``basal shifting" resembles the phenomenon of detuning observed in the cochlea. Acting as a frequency analyzer (or ``inverse-piano"), each location on the BM vibrates in response to a sound stimulus at a particular frequency. Thus, the BM has a frequency-to-location map such that every stimulus frequency has a preferred place on the BM called Characteristic Place (CP). The detuning phenomenon is observed as the shifting of the CP towards the stapes as the intensity of the stimulus (in dB) is increased. In this section, we showed that increasing the spread of the stochastic perturbations also shifts the BM vibrations towards the stapes. Nonlinear dynamics are necessary to model the detuning phenomenon. However, modeling this ``detuning-like" phenomenon doesn't require nonlinearities, instead a locally perturbed active gain is sufficient to explain it. It is believed that these instabilities in the BM reflect back to the middle ear causing SOAEs \cite{nuttall2004spontaneous}. It is also believed that if these BM vibrations are intense enough, they can be perceived as tinnitus. Our results suggest a mechanism that explains the frequencies that can be detected in the ear canal due to SOAEs and/or perceived as tinnitus. As a matter of fact, the shape of the statistics (expectation and covariance) of the gain coefficient is a factor that controls the bands of the frequencies that are emitted as SOAEs. These emissions arise due to (a) \textit{spatially variant inhomogeneities} along the cochlear partition and (b) \textit{temporal stochastic perturbations} that give rise to structured stochastic uncertainties. \section{Nonlinear Stochastic Simulations} \label{Section: Simulations} So far, the MSS analysis is carried out on the linearized dynamics. In this section, we carry out stochastic simulations of the nonlinear model to validate the predictions of our analysis of the linearized dynamics. \subsection{Nonlinear Descriptor State Space Formulation in Continuous Space-Time} We first start by formulating the nonlinear dynamics in a DSS form similar to that given in section \ref{Section: DSS Formulation}. Recall that, the nonlinear deterministic active gain is given by (\ref{Eqn: Nonlinear Gain}) with $\gamma(x)$ representing the gain coefficient. To include stochastic perturbations, we substitute (\ref{Eqn: Stochastic Gain}) in (\ref{Eqn: Nonlinear Gain}) so that the nonlinear stochastic active gain can be written as \begin{equation} \label{Eqn: Stochastic Nonlinear Active Gain} \begin{aligned} \left[\mathcal G(u)\right](x,t) &=\frac{\bar\gamma(x) + \epsilon \tilde \gamma(x,t)}{1 + \theta \left[\Phi_\eta\left(\frac{u^2}{R^2}\right)\right](x,t)} \\ & =: \bigg(\bar\gamma(x) + \epsilon \tilde \gamma(x,t)\bigg) \left[\tilde{\mathcal G}(u)\right](x,t). \end{aligned} \end{equation} Recall that $\Phi_{\eta}$ is the Gaussian spatial operator given by (\ref{Eqn: Guassian Weighing Operator}), $\theta = 0.5$, $R = 1\si{nm}$ and $\eta = 0.5345\si{mm}$. By substituting (\ref{Eqn: Stochastic Nonlinear Active Gain}) in (\ref{Eqn: Micromechanics}), we can rewrite the nonlinear model in a nonlinear DSS form as \begin{equation} \label{Eqn: Nonlinear DSS} \mathcal E \frac{\partial}{\partial t} \psi(x,t) = \big(\mathcal A_{\bar \gamma}(u) + \epsilon \tilde{\mathcal{B}}_0(u)\tilde \gamma \mathcal C_0\big) \psi(x,t), \end{equation} where $\mathcal A_{\bar{\gamma}}(u) := \mathcal A_0 + \tilde{\mathcal B_0}(u) \bar \gamma \mathcal C_0$ and $\tilde{\mathcal{B}}_0(u)\tilde \gamma \mathcal C_0 $ are nonlinear spatial operators that represent the deterministic and stochastic portions of the dynamics, respectively. Note that $\mathcal E, \mathcal A_0$, and $\mathcal C_0$ are all defined in (\ref{Eqn: Descriptor State Space Operator Form}), and $\tilde{\mathcal B}_0(u) = \begin{bmatrix} 0 & 0 & \tilde{\mathcal G}(u) & 0 \end{bmatrix}^T$. Therefore, (\ref{Eqn: Nonlinear DSS}) represents the nonlinear stochastic dynamics given in a DSS operator form, where the spatial variable is continuous. This is really a Stochastic Partial Differential Equation (SPDE) that needs to be discretized in space and time in order to carry out our simulations. \subsection{Description of the Numerical Method for Simulations} In this section, we discretize (\ref{Eqn: Nonlinear DSS}) in space and time so that numerical simulations become fairly straightforward to implement. On a side note, if the stochastic perturbation $\tilde \gamma = 0$, (\ref{Eqn: Nonlinear DSS}) becomes a deterministic Partial Differential Equation (PDE). This can be easily integrated by discretizing space using a spatial grid, and then employ a time marching solver such as ODE45 in MATLAB. However, for an SPDE, one has to carefully treat the scaling of the covariances with the discretization steps. Space and time are discretized as $x_i = i\Delta_x$ and $t_n = n\Delta_t$ with discretization steps $\Delta_x = L/N_x$ and $ \Delta_t = t_f/N_t$ for $i = 0, 1, ..., N_x$ and $ n = 0, 1, ..., N_t$, where $t_f$ is the final time. Let the BM and TM displacements on the discretized space-time grid be denoted by the vectors $u_n$ and $v_n \in \mathbb R^{N_x+1}$, respectively such that \begin{align*} u_n &:= \begin{bmatrix} u(x_0,t_n) & \cdots & u(x_{N_x},t_n) \end{bmatrix}^T \\ v_n &:= \begin{bmatrix} v(x_0,t_n) & \cdots & v(x_{N_x},t_n) \end{bmatrix}^T. \end{align*} Then the discretized state space variable can be expressed by $\psi_n \in \mathbb R^{4(N_x+1)}$ as $$ \psi_n := \begin{bmatrix} u_n^T & v_n^T & \dot u_n^T & \dot v_n^T \end{bmatrix}^T.$$ For scenarios $\textbf{S}_1$ and $\textbf{S}_3$, $\tilde \gamma(x,t)$ is a zero-mean white process in space and time. It can be approximated at the spatial grid points $\{x_i\}_{i=0,1,...,N_x}$ and at time $t_n$ as follows \begin{equation*} \begin{bmatrix} \tilde \gamma(x_0,t_n) & \tilde \gamma(x_1,t_n) & \cdots & \tilde \gamma(x_{N_x},t_n) \end{bmatrix}^T\approx \frac{1}{\sqrt{\Delta_x \Delta_t}} w_n, \end{equation*} where $w_n \in \mathbb R^{N_x+1}$ is a zero-mean Gaussian random vector with a covariance matrix $\mathbb E\left[w_n w_n^T\right] = I$ for $\mathbf S_1$ and $\mathbb E\left[w_n w_n^T\right] = \mathcal D\left( \begin{bmatrix} \phi_{\sigma}(x_0-\mu) & \cdots & \phi_{\sigma}(x_{N_x}-\mu) \end{bmatrix}\right)$ for $\mathbf S_3$, where $\mathcal D$ is the diagonal operator such that $\mathcal D(w_n)$ is a diagonal matrix with $w_n$ arranged on its diagonal entries. For scenario $\textbf{S}_2$, $\tilde \gamma(x,t)$ is a stochastic process that is white in time but ``colored" in space with a spatial covariance $\mathbf \Gamma(x,\xi) = \epsilon^2 \phi_\lambda(x-\xi)$. In this scenario, the noise is smooth in space and there is no need to scale the covariance by the spatial discretization step. More precisely, $\tilde \gamma(x,t)$ can be approximated as \begin{equation*} \begin{bmatrix} \tilde \gamma(x_0,t_n) & \tilde \gamma(x_1,t_n) & \cdots & \tilde \gamma(x_{N_x},t_n) \end{bmatrix}^T\approx \frac{1}{\sqrt{\Delta_t}} w_n, \end{equation*} where $\mathbb E\left[w_n w_n^T\right]$ is now a symmetric matrix whose $(i,j)^{\text{th}}$ entry is given by $\phi_{\lambda}(x_i-x_j)$. Therefore, a first order approximation of (\ref{Eqn: Nonlinear DSS}) can be carried out in the spirit of the Euler-Maruyama method \cite{maruyama1955continuous} to obtain \begin{equation}\label{Eqn: Nonlinear Numerics} E \psi_{n+1} = E\psi_n + \Delta_t A_{\bar \gamma}(u_n) \psi_n + \alpha \tilde B_0(u_n) \mathcal D(w_n) C_0\psi_n \end{equation} where $\alpha = \epsilon \sqrt{\Delta_t/\Delta_x}$ for $\mathbf S_1$ and $\mathbf S_3$; and $\alpha = \epsilon \sqrt{\Delta_t}$ for $\mathbf S_2$. The matrices $E , A_{\bar{\gamma}}(u_n) , \tilde B_0(u_n)$ and $C_0 $ are all finite dimensional approximations of the operators $\mathcal E, \mathcal A_{\bar \gamma}(u), \tilde{\mathcal B}_0(u)$ and $\mathcal C_0$, respectively (Appendix-\ref{Section: Finite Realizations}). Equation (\ref{Eqn: Nonlinear Numerics}) represents the recursive numerical methods to solve (\ref{Eqn: Nonlinear DSS}) for all three scenarios with the right choice of $\alpha$ and $\mathbb E[w_nw_n^T]$. \subsection{Simulation of the Nonlinear Stochastic Model} To validate our MSS analysis of the linearized dynamics and evaluate how well it copes with the nonlinear dynamics, we carry out a simulation of (\ref{Eqn: Nonlinear DSS}). This section considers scenario $\textbf{S}_1$. Hence, the numerical method used here is that given in (\ref{Eqn: Nonlinear Numerics}) for $\alpha = \epsilon^2 \sqrt{\Delta_t/\Delta_x}$ and $\mathbb E[w_nw_n^T] = I$. The nonlinear stochastic simulation shown here is for $\bar \gamma(x)$ given in (\ref{Eqn: Gain Mean Profiles}) with $\beta = 2$. All other scenarios are in agreement with our MSS analysis; however, this particular case study ($\beta = 2$) is chosen here to illustrate the effectiveness of our analysis. Observe using Figure~\ref{Fig: Sweeping lambda beta}(b) that for $\beta = 2$, the MSS condition is violated if $\epsilon \geq 9.1\times 10^{-6}$. We choose $\epsilon = 1.1 \times 10^{-5}$ which slightly violates the MSS condition for the linearized dynamics and allows the nonlinearity to kick in and saturate the response. The spatio-temporal response of the BM is depicted in Figure~\ref{Fig: Nonlinear Stochastic Simulation}(a) for $t \in [0, t_f]$ with $t_f = 200\si{ms}$. The response is maximal in a band limited region $10\si{mm} < x < 20\si{mm}$ which corresponds to a frequency range of roughly between $1\si{kHz}$ and $5\si{kHz}$. \begin{figure}[h!] \centering \begin{tabular}{c} \includegraphics[scale = .22]{Figures/Stochastic_Simulation.pdf} \\ \footnotesize{(a) Spatio-Temporal Stochastic Evolution of the BM} \\ \includegraphics[scale = .22]{Figures/Empirical_Covariance.pdf} \\ \footnotesize{(b) Empirical Covariance $\mathbf U_{\text{Emp}}(x,\xi)$} \\ \includegraphics[scale = .22]{Figures/Theoretical_Covariance.pdf} \\ \footnotesize{(c) Predicted Worst-Case Covariance $\mathbf U(x,\xi)$} \\ \begin{tabular}{ll} \includegraphics[scale =.119]{Figures/Eigenvalues.pdf} & \includegraphics[scale =.119]{Figures/EmpiricalTheoretical_Comparison.pdf} \\ \includegraphics[scale=.119]{Figures/Second_EV.pdf} & \includegraphics[scale=.119]{Figures/Third_EV.pdf} \end{tabular}\\ \footnotesize{(d) Empirical and Theoretical Dominant Eigenvalues/functions} \end{tabular} \caption{\footnotesize{Nonlinear Stochastic Simulation. Figure (a) shows the BM response to spatially uncorrelated stochastic active gain (scenario $\mathbf S_1$) with an expectation given by (\ref{Eqn: Gain Mean Profiles}) where $\beta = 2$ and a perturbation of $\epsilon = 1.1\times 10^{-5}$. Figures (b) and (c) show a comparison between the empirical and predicted covariances. The predicted covariance is computed for the linearized dynamics via the power iteration method applied on the loop gain operator (\ref{Eqn: Loop Gain Operator}). The empirical covariance is computed using the data obtained from one nonlinear stochastic simulation using (\ref{Eqn: Nonlinear Numerics}) and integrated in time using (\ref{Eqn: Ergodic Integration}) assuming ergodicity. Figure (d) shows a comparison between the dominant eigenvalues/functions of the empirical and predicted covariances shown in Figures (b) and (c), respectively. This eigen-decomposition is referred to as the \textit{Karhunen{–}Lo\`ve} decomposition. Clearly the theoretical predictions match the empirical data, thus suggesting that the nonlinearities only saturate the response without significantly deforming the waveforms.}} \label{Fig: Nonlinear Stochastic Simulation} \end{figure} To be more precise, we compute the empirical covariance $\textbf{U}_{\text{Emp}}(x,\xi)$ as follows \begin{equation} \label{Eqn: Ergodic Integration} \textbf{U}_{\text{Emp}}(x,\xi) = \frac{1}{t_f} \int_0^{t_f} u(x,\tau) u(\xi, \tau) d\tau. \end{equation} The time averaging replaces the expectation assuming ergodicity. Figures~\ref{Fig: Nonlinear Stochastic Simulation}(b) and (c) compare the empirical covariance to the predicted worst-case covariance. By visual inspection, we observe that the empirical results are in good agreement with our theoretical predictions. For a more precise comparison, we plot the first twenty dominant eigenvalues and first three dominant eigenfunctions of both the predicted and empirical covariances in Figure~\ref{Fig: Nonlinear Stochastic Simulation}(d). This eigen-decomposition is referred to as the \textit{Karhunen{-}Lo\`eve decomposition}. The eigenfunctions are the modes of BM vibrations that have the highest growth rate and are more likely to destabilize for small perturbations of the active gain. The plots doesn't show any significant difference between the empirical and theoretical results. In fact, although the nonlinear active gain slightly deforms the response, but its fundamental role (in the absence of a stimulus) is to saturate the linearized instabilities to form oscillations that remain bounded in time. \section{Discussion} \label{Section: Discussion} The mechanisms underlying cochlear instabilities such as SOAEs and tinnitus are still controversial and not well understood. This paper suggests a new possible source of cochlear instabilities: spatio-temporal stochastic perturbations of the active gain. It is widely accepted that Outer Hair Cells (OHC) are responsible for the active gain in the cochlea. This work proposes a \textit{simulation-free}, system theoretic framework to analyze the effects of small stochastic perturbations that may occur on the level of the OHCs. These perturbations can have several physical origins such as noisy nearby neuronal activities, cellular activities, blood flow, etc... Studying the effects of randomness in the active gain is not new \cite{fruth2014active}, \cite{ku2008statistics}. However, the previous studies on this matter considered random spatial perturbations that are time-invariant. This type of randomness is referred to as ``frozen" or \textit{quenched disorder} in the statistical physics community. In fact, \cite{ku2008statistics} investigated the effects of the frozen spatial randomness by carrying out Monte Carlo simulations to study the statistics of the instabilities. However, to achieve a broad spectrum of unstable frequencies, the authors allowed severe perturbations of the active gain which is not realistic. Without these severe perturbations, the unstable frequencies would be limited to a band of high frequencies only (Section~\ref{Section: Constant Expectation}). This doesn't agree with the experimental observations where, for example, SOAEs are mainly found between $0.5$ and $4.5\si{kHz}$. A more realistic case is to treat the active gain as a stochastic process, where the randomness may occur in space and time, simultaneously. In addition to that, only small perturbations of the active gain are considered (three to four orders of magnitude less than \cite{ku2008statistics}). A major advantage of our analysis is that it is \textit{simulation-free} and no Monte Carlo simulations are required to study the statistics of the emerging instabilities. In our analysis, we also show that the band of unstable frequencies can be controlled by the tuning of the structural parameters of the cochlea such as the active gain coefficient. Hence, we show that even for very small perturbations, the unstable frequencies can be shifted dramatically. Furthermore, examining localized stochastic perturbations in the active gain allowed us to observe local instabilities that shift toward the stapes as the localization length or spread is larger. This observation resembles the detuning phenomenon present in the cochlea. \section{Conclusion and Future Work} This paper examines the instabilities that occur in the linearized dynamics due to spatio-temporal stochastic perturbations in the distributed structure of the cochlear partition. The simulation-free analysis is carried out through a structured stochastic uncertainty framework. It is shown that the spatial shape of the expectation and covariance of the gain coefficient affect the locations of the instabilities on the basilar membrane. These instabilities eventually saturate to form bounded oscillations due to the saturation nonlinearity of the active gain (\ref{Eqn: Nonlinear Gain}) producing spontaneous basilar membrane vibrations. It is believed that these instabilities are reflected to the middle ear as spontaneous otoacoustic emissions (SOAEs) \cite{nuttall2004spontaneous} with frequencies corresponding to the location of the instability on the basilar membrane. This analysis also suggests an explanation of one possible source of tinnitus, which is less addressed in the literature. Particularly, if the spontaneous BM vibrations were intense enough, they may be perceived as tinnitus. Future work will address instabilities that may occur due to stochastic uncertainties in structural parameters other than the active gain coefficient, such as the cochlear fluid density. \bibliographystyle{ieeetr}

9e397df8ad288fd9e94c09ddd82178778ca0d308

\section{Introduction} Consider $n$ possible treatments, say, drugs in a clinical trial, where each treatment either has a positive expected effect relative to a baseline (actual positive), or no difference (null), with a goal of identifying as many actual positive treatments as possible. If evaluating the $i$th trial results in a noisy outcome (e.g. due to variance in the actual measurement or just diversity in the population) then given a total measurement budget of $B$, it is standard practice to execute and average $B/n$ measurements of each treatment, and then output a set of predicted actual positives based on the measured effect sizes. False alarms (i.e. nulls predicted as actual positives) are controlled by either controlling \emph{family-wise error rate (FWER)}, where one bounds the probability that at least one of the predictions is null, or \emph{false discovery rate (FDR)}, where one bounds the expected proportion of the number of predicted nulls to the number of predictions. FDR is a weaker condition than FWER but is often used in favor of FWER because of its higher \emph{statistical power}: more actual positives are output as predictions using the same measurements. In the pursuit of even greater statistical power, there has recently been increased interest in the biological sciences to reject the uniform allocation strategy of $B/n$ trials to the $n$ treatments in favor of an \emph{adaptive} allocation. Adaptive allocations partition the budget $B$ into sequential rounds of measurements in which the measurements taken at one round inform which measurements are taken in the next \cite{hao2008drosophila,rocklin}. Intuitively, if the effect size is relatively large for some treatment, fewer trials will be necessary to identify that treatment as an actual positive relative to the others, and that savings of measurements can be allocated towards treatments with smaller effect sizes to boost the signal. However, both \cite{hao2008drosophila,rocklin} employed ad-hoc heuristics which may not only have sub-optimal statistical power, but also may even result in more false alarms than expected. As another example, in the domain of A/B/n testing in online environments, the desire to understand and maximize click-through-rate across treatments (e.g., web-layouts, campaigns, etc.) has become ubiquitous across retail, social media, and headline optimization for the news. And in this domain, the desire for statistically rigorous adaptive sampling methods with high statistical power are explicit \cite{johari2015always}. In this paper we propose an adaptive measurement allocation scheme that achieves near-optimal statistical power subject to FWER or FDR false alarm control. Perhaps surprisingly, we show that even if the treatment effect sizes of the actual positives are identical, adaptive measurement allocation can still substantially improve statistical power. That is, more actual positives can be predicted using an adaptive allocation relative to the uniform allocation under the same false alarm control. \subsection{Problem Statement}\label{sec:problem_statement} Consider $n$ distributions (or arms) and a game where at each time $t$, the player chooses an arm $i \in [n] := \{1,\dots,n\}$ and immediately observes a reward $X_{i,t} \overset{iid}{\sim} \nu_i$ where $X_{i,t} \in [0,1]$\footnote{All results without modification apply to unbounded, sub-Gaussian random variables.} and $\mathbb{E}_{\nu_i}[X_{i,t}] = \mu_i$. For a \emph{known} threshold $\mu_0$, define the sets\footnote{All results generalize to the case when $\mathcal{H}_0 = \{ i : \mu_i \leq \mu_0 \}$.} \begin{align*} \mathcal{H}_1 = \{ i \in [n]: \mu_i > \mu_0 \} \quad \text{ and } \quad \mathcal{H}_0 = \{ i \in [n] : \mu_i = \mu_0 \} = [n] \setminus \mathcal{H}_1 . \end{align*} The value of the means $\mu_i$ for $i\in[n]$ and the cardinality of $\mathcal{H}_1$ are \emph{unknown}. The arms (treatments) in $\mathcal{H}_1$ have means greater than $\mu_0$ (positive effect) while those in $\mathcal{H}_0$ have means equal to $\mu_0$ (no effect over baseline). At each time $t$, after the player plays an arm, she also outputs a set of indices $\mathcal{S}_t \subseteq [n]$ that are interpreted as \emph{discoveries} or rejections of the null-hypothesis (that is, if $i \in \mathcal{S}_t$ then the player believes $i \in \mathcal{H}_1$). For as small a $\tau \in \mathbb{N}$ as possible, the goal is to have the number of true detections $|\mathcal{S}_t \cap \mathcal{H}_1|$ be approximately $|\mathcal{H}_1|$ for all $t \geq \tau$, subject to the number of false alarms $|\mathcal{S}_t \cap \mathcal{H}_0|$ being small uniformly over all times $t \in \mathbb{N}$. We now formally define our notions of false alarm control and true discoveries. \begin{definition}[False Discovery Rate, FDR-$\delta$] Fix some $\delta \in (0,1)$. We say an algorithm is FDR-$\delta$ if for all possible problem instances $(\{\nu_i\}_{i=1}^n, \mu_0)$ it satisfies $\displaystyle\mathbb{E}[\tfrac{|\mathcal{S}_t \cap \mathcal{H}_0|}{|\mathcal{S}_t|\vee 1}] \leq \delta$ for all $t \in \mathbb{N}$ simultaneously. \end{definition} \begin{definition}[Family-wise Error Rate, FWER-$\delta$] Fix some $\delta \in (0,1)$. We say an algorithm is FWER-$\delta$ if for all possible problem instances $(\{\nu_i\}_{i=1}^n, \mu_0)$ it satisfies $\mathbb{P}(\bigcup_{t=1}^\infty \{ \mathcal{S}_t \cap \mathcal{H}_0 \neq \emptyset\} ) \leq \delta$. \end{definition} \noindent Note FWER-$\delta$ implies FDR-$\delta$, the former being a stronger condition than the latter. Allowing a relatively small number of false discoveries is natural, especially if $|\mathcal{H}_1|$ is relatively large. Because $\mu_0$ is known, there exist schemes that guarantee FDR-$\delta$ or FWER-$\delta$ even if the arm means $\mu_i$ and the cardinality of $\mathcal{H}_1$ are unknown (see Section~\ref{sec:false_alarm_control}). It is also natural to relax the goal of identifying \emph{all} arms in $\mathcal{H}_1$ to simply identifying a \emph{large proportion} of them. \begin{definition}[True Positive Rate, TPR-$\delta,\tau$] Fix some $\delta \in (0,1)$. We say an algorithm is TPR-$\delta,\tau$ on an instance $(\{\nu_i\}_{i=1}^n, \mu_0)$ if $\mathbb{E}[\frac{|\mathcal{S}_t \cap \mathcal{H}_1|}{|\mathcal{H}_1|}] \geq 1-\delta$ for all $t \geq \tau$. \end{definition} \begin{definition}[Family-wise Probability of Detection, FWPD-$\delta,\tau$] Fix some $\delta \in (0,1)$. We say an algorithm is FWPD-$\delta,\tau$ on an instance $(\{\nu_i\}_{i=1}^n, \mu_0)$ if $\mathbb{P}(\mathcal{H}_1 \subseteq \mathcal{S}_t ) \geq 1- \delta$ for all $t \geq \tau$. \end{definition} Note that FWPD-$\delta,\tau$ implies TPR-$\delta,\tau$, the former being a stronger condition than the latter. Also note $\mathbb{P}( \bigcup_{t=1}^\infty \{ \mathcal{S}_t \cap \mathcal{H}_0 \neq \emptyset\} ) \leq \delta$ and $\mathbb{P}( \mathcal{H}_1 \subseteq \mathcal{S}_\tau) \geq 1-\delta$ together imply $\mathbb{P}(\mathcal{H}_1 = \mathcal{S}_\tau) \geq 1-2\delta$. We will see that it is possible to control the number of false discoveries $|\mathcal{S}_t \cap \mathcal{H}_0|$ regardless of how the player selects arms to play. It is the rate at which $\mathcal{S}_t$ includes $\mathcal{H}_1$ that can be thought of as the statistical power of the algorithm, which we formalize as its \emph{sample complexity}: \begin{definition}[Sample Complexity] Fix some $\delta \in (0,1)$ and an algorithm $\mathcal{A}$ that is FDR-$\delta$ (or FWER-$\delta$) over all possible problem instances. Fix a particular problem instance $(\{\nu_i\}_{i=1}^n, \mu_0)$. At each time $t \in \mathbb{N}$, $\mathcal{A}$ chooses an arm $i \in [n]$ to obtain an observation from, and before proceeding to the next round outputs a set $\mathcal{S}_t \subseteq [n]$. The \emph{sample complexity} of $\mathcal{A}$ on this instance is the smallest time $\tau \in \mathbb{N}$ such that $\mathcal{A}$ is TPR-$\delta,\tau$ (or FWPD-$\delta,\tau$). \end{definition} The sample complexity and value of $\tau$ of an algorithm will depend on the particular instance $(\{\nu_i\}_{i=1}^n, \mu_0)$. For example, if $\mathcal{H}_1 = \{i \in [n] : \mu_i =\mu_0 + \Delta\}$ and $\mathcal{H}_0 = [n] \setminus \mathcal{H}_1$, then we expect the sample complexity to increase as $\Delta$ decreases since at least $\Delta^{-2}$ samples are necessary to determine whether an arm has mean $\mu_0$ versus $\mu_0 + \Delta$. The next section will give explicit cases. \begin{remark}[Impossibility of stopping time] We emphasize that just as in the non-adaptive setting, at no time can an algorithm \emph{stop} and declare that it is TPR-$\delta,\tau$ or FWPD-$\delta,\tau$ for any finite $\tau \in \mathbb{N}$. This is because there may be an arm in $\mathcal{H}_1$ with a mean infinitesimally close to $\mu_0$ but distinct such that no algorithm can determine whether it is in $\mathcal{H}_0$ or $\mathcal{H}_1$. Thus, the algorithm must run indefinitely or until it is stopped externally. However, using an anytime confidence bound (see Section~\ref{sec:algorithm}) one can always make statements like ``either $\mathcal{H}_1 \subseteq \mathcal{S}_t$, or $\max_{i \in \mathcal{H}_1 \setminus \mathcal{S}_t} \mu_i-\mu_0 \leq \epsilon$'' where the $\epsilon$ will depend on the width of the confidence interval. \end{remark} \subsection{Contributions and Informal Summary of Main Results}\label{sec:contributions} \begin{table} \begin{center} \begin{tabular}{c | c c } & \multicolumn{2}{c}{\textbf{False alarm control}}\\[2pt] & \makecell{FDR-$\delta$ \\ $\max_t \mathbb{E}[\frac{|\mathcal{S}_t \cap \mathcal{H}_0|}{|\mathcal{S}_t| \vee 1}] \leq \delta$} & \makecell{FWER-$\delta$ \\ $\mathbb{P}( \bigcup_{t=1}^\infty \{ \mathcal{S}_t \cap \mathcal{H}_0 \neq \emptyset\} ) \leq \delta$} \\ \hline \\[-8pt] \makecell{\textbf{Detection Probability}\\[4pt] TPR-$\delta,\tau$ \\ $\mathbb{E}[\frac{|\mathcal{S}_\tau \cap \mathcal{H}_1|}{|\mathcal{H}_1|}] \geq 1-\delta$\\[8pt] } & \makecell{ {\small Theorem~\ref{thm:FDR_TPR}}\\[-0pt] $n\Delta^{-2}$} & \makecell{ {\small Theorem~\ref{thm:FWER_TPR}}\\[-0pt] $(n-k)\Delta^{-2} + k \Delta^{-2} \log(n-k)$} \\[16pt] \makecell{FWPD-$\delta,\tau$ \\ $\mathbb{P}( \mathcal{H}_1 \subseteq \mathcal{S}_\tau) \geq 1-\delta$} & \makecell{ {\small Theorem~\ref{thm:FDR_FWPD}}\\[-0pt] $\quad (n-k) \Delta^{-2} \log(k) + k \Delta^{-2} \quad$} & \makecell{ {\small Theorem~\ref{thm:FWER_FWPD}}\\[-0pt] $(n-k) \Delta^{-2}\log(k) + k \Delta^{-2}\log(n-k)$} \end{tabular} \end{center} \caption{\small Informal summary of sample complexity results proved in this paper for $|\mathcal{H}_1|=k$, constant $\delta$ (e.g., $\delta=.05$) and $\Delta=\min_{i \in \mathcal{H}_1} \mu_i - \mu_0$. Uniform sampling across all settings requires at least $n \Delta^{-2} \log(n/k)$ samples, and in the FWER+FWPD setting requires $n \Delta^{-2}\log(n)$. Constants and $\log\log$ factors are ignored.\label{tab:complexity_table}} \vspace{-.25in} \end{table} In Section~\ref{sec:algorithm} we propose an algorithm that handles all four combinations of \{FDR-$\delta$, FWER-$\delta$\} and \{TPR-$\delta,\tau$, FWPD-$\delta,\tau$\}. A reader familiar with the multi-armed bandit literature would expect an adaptive sampling algorithm to have a large advantage over uniform sampling when there is a large diversity in the means of $\mathcal{H}_1$ since larger means can be distinguished from $\mu_0$ with fewer samples. However, one should note that to declare all of $\mathcal{H}_1$ as discoveries, one must sample every arm in $\mathcal{H}_0$ \emph{at least} as many times as the \emph{most sampled} arm in $\mathcal{H}_1$, otherwise they are statistically indistinguishable. As discoveries are typically uncovering rare phenomenon, it is common to assume $|\mathcal{H}_1| = n^\beta$ for $\beta \in (0,1)$ \cite{castro2014adaptive,rabinovich2017optimal}, or $|\mathcal{H}_1| = o(n)$, but this implies that the number of samples taken from the arms in $\mathcal{H}_1$, regardless of how samples are allocated to those arms, will almost always be dwarfed by the number of samples allocated to those arms in $\mathcal{H}_0$ since there are $\Omega(n)$ of them. This line of reasoning, in part, is what motivates us to give our sample complexity results in terms of the quantities that best describe the contributions from those arms in $\mathcal{H}_0$, namely, the cardinality $|\mathcal{H}_1| = n-|\mathcal{H}_0|$, the confidence parameter $\delta$ (e.g., $\delta=.05$), and the gap $\Delta := \min_{i \in \mathcal{H}_1} \mu_i - \mu_0$ between the means of the arms in $\mathcal{H}_0$ and the smallest mean in $\mathcal{H}_1$. Reporting sample complexity results in terms of $\Delta$ also allows us to compare to known lower bounds in the literature \cite{2017arXiv170306222R,castro2014adaptive,malloy2014sequential,simchowitz2017simulator}. Nevertheless, we do address the case where the means of $\mathcal{H}_1$ are varied in Theorem~\ref{thm:FDR_TPR}. An informal summary of the sample complexity results proven in this work are found in Table~\ref{tab:complexity_table} for $|\mathcal{H}_1|=k$. For the least strict setting of FDR+TPR, the upper-left quadrant of Table~\ref{tab:complexity_table} matches the lower bound of \cite{castro2014adaptive}, a sample complexity of just $\Delta^{-2}n$. In this FDR+TPR setting (which requires the fewest samples of the four settings), uniform sampling which pulls each arm an equal number of times has a sample complexity of at least $n \Delta^{-2} \log(n/|\mathcal{H}_1|)$ (see Theorem \ref{thm:uniform_fdr_lower_bound} in Appendix~\ref{sec:succ-elim}), which exceeds all results in Table~\ref{tab:complexity_table} demonstrating the statistical power gained by adaptive sampling. For the most strict setting of FWER+FWPD, the lower-right quadrant of Table~\ref{tab:complexity_table} matches the lower bounds of \cite{malloy2014sequential,kalyanakrishnan2012pac,simchowitz2017simulator}, a sample complexity of $(n-k) \Delta^{-2}\log(k) + k \Delta^{-2}\log(n-k)$. Uniform sampling in the FWER+FWPD setting has a sample complexity lower bounded by $n \Delta^{-2} \log(n)$ (see Theorem~\ref{thm:uniform_fwer_lower_bound} in Appendix~\ref{sec:succ-elim}). The settings of FDR+FWPD and FWER+TPR are sandwiched between these results, and we are unaware of existing lower bounds for these settings. All the results in Table~\ref{tab:complexity_table} are novel, and to the best of our knowledge are the first non-trivial sample complexity results for an adaptive algorithm in the \emph{fixed confidence} setting where a desired confidence $\delta$ is set, and the algorithm attempts to minimize the number of samples taken to meet the desired conditions. We also derive tools that we believe may be useful outside this work: for always valid $p$-values (c.f. \cite{johari2015always,yang2017framework}) we show that FDR is controlled for all times using the Benjamini-Hochberg procedure \cite{benjamini1995controlling} (see Lemma~\ref{lem:expected_FDR}), and also provide an anytime high probability bound on the false discovery proportion (see Lemma~\ref{lem:fdr_high_prob}). Finally, as a direct consequence of the theoretical guarantees proven in this work and the empirical performance of the FDR+TPR variant of the algorithm on real data, an algorithm faithful to the theory was implemented and is in use in production at a leading A/B testing platform \cite{optimizely}. \subsection{Related work}\label{sec:related_work} Identifying arms with means above a threshold, or equivalently, multiple testing via rejecting null-hypotheses with small $p$-values, is an ubiquitous problem in the biological sciences. In the standard setup, each arm is given an equal number of measurements (i.e., a uniform sampling strategy), a $p$-value $P_i$ is produced for each arm where $\mathbb{P}(P_i \leq x) \leq x$ for all $x \in (0,1]$ and $i \in \mathcal{H}_0$, and a procedure is then run on these $p$-values to declare small $p$-values as rejections of the null-hypothesis, or discoveries. For a set of $p$-values $P_1 \leq P_2 \leq \dots \leq P_n$, the so-called Bonferroni selection rule selects $\mathcal{S}_{BF} = \{ i : P_{i} \leq \delta/n \}$. The fact that FWER control implies FDR control, $\mathbb{E}[ | \mathcal{S}_{BF} \cap \mathcal{H}_0| ] \leq \mathbb{P}(\bigcup_{i \in \mathcal{H}_0} \{ P_{i} \leq \delta/n\} ) \leq \delta \frac{|\mathcal{H}_0|}{n} \leq \delta$, suggests that greater statistical power (i.e. more discoveries) could be achieved with procedures designed specifically for FDR. The BH procedure \cite{benjamini1995controlling} is one such procedure to control FDR and is widely used in practice (with its many extensions \cite{2017arXiv170306222R} and performance investigations \cite{rabinovich2017optimal}). Recall that a uniform measurement strategy where every arm is sampled the same number of times requires $n \Delta^{-2} \log(n/k)$ samples in the FDR+TPR setting, and $n \Delta^{-2}\log(n)$ samples in the FWER+FWPD setting (Theorems~\ref{thm:uniform_fdr_lower_bound} and \ref{thm:uniform_fwer_lower_bound} in Appendix~\ref{sec:succ-elim}), which can be substantially worse than our adaptive procedure (see Table~\ref{tab:complexity_table}). Adaptive sequential testing has been previously addressed in the \emph{fixed budget} setting: the procedure takes a sampling budget as input, and the guarantee states that if the given budget is larger than a problem dependent constant, the procedure drives the error probability to zero and the detection probability to one. One of the first methods called \emph{distilled sensing} \cite{haupt2011distilled} assumed that arms from $\mathcal{H}_0$ were Gaussian with mean at most $\mu_0$, and successively discarded arms after repeated sampling by thresholding at $\mu_0$--at most the median of the null distribution--thereby discarding about half the nulls at each round. The procedure made guarantees about FDR and TPR, which were later shown to be nearly optimal \cite{castro2014adaptive}. Specifically, \cite[Corollary 4.2]{castro2014adaptive} implies that any procedure with $\max\{FDR+(1-TPR)\} \leq \delta$ requires a budget of at least $\Delta^{-2} n \log(1/\delta)$, which is consistent with our work. Later, another thresholding algorithm for the fixed budget setting addressed the FWER and FWPD metrics \cite{malloy2014sequential}. In particular, if their procedure is given a budget exceeding $(n-|\mathcal{H}_1|) \Delta^{-2}\log(|\mathcal{H}_1|) + |\mathcal{H}_1| \Delta^{-2} \log(n-|\mathcal{H}_1|)$ then the FWER is driven to zero, and the FWPD is driven to one. By appealing to the optimality properties of the SPRT (which knows the distributions precisely) it was argued that this is optimal. These previous works mostly focused on the asymptotic regime as $n \rightarrow \infty$ and $|\mathcal{H}_1| = o(n)$. Our paper, in contrast to these previous works considers the \emph{fixed confidence} setting: the procedure takes a desired FDR (or FWER) and TPR (or FWPD) and aims to minimize the number of samples taken before these constraints are met. To the best of our knowledge, our paper is the first to propose a scheme for this problem in the fixed confidence regime with near-optimal sample complexity guarantees. A related line of work is the threshold bandit problem, where all the means of $\mathcal{H}_1$ are assumed to be strictly above a given threshold, and the means of $\mathcal{H}_0$ are assumed to be strictly below the threshold \cite{locatelli2016optimal,kano2017good}. To identify this partition, each arm must be pulled a number of times inversely proportional to the square of its deviation from the threshold. This contrasts with our work, where the majority of arms may have means \emph{equal} to the threshold and the goal is to identify arms with means greater than the threshold subject to discovery constraints. If the arms in $\mathcal{H}_0$ are assumed to be strictly below the threshold it is possible to declare arms as in $\mathcal{H}_0$. In our setting we can only ever determine that an arm is in $\mathcal{H}_1$ and not $\mathcal{H}_0$, but it is impossible to detect that an arm is in $\mathcal{H}_0$ and not in $\mathcal{H}_1$. Note that the problem considered in this paper is very related to the top-$k$ identification problem where the objective is to identify the unique $k$ arms with the highest means with high probability \cite{chen2017nearly,kalyanakrishnan2012pac,simchowitz2017simulator}. Indeed, if we knew $|\mathcal{H}_1|$, then our FWER+FWPD setting is equivalent to the top-$k$ problem with $k=|\mathcal{H}_1|$. Lower bounds derived for the top-$k$ problem assume the algorithm has knowledge of the values of the means, just not their indices \cite{chen2017nearly, simchowitz2017simulator}. Thus, these lower bounds also apply to our setting and are what are referenced in Section \ref{sec:contributions}. As pointed out by \cite{locatelli2016optimal}, both our setting and the threshold bandit problem can be posed as a combinatorial bandits problem as studied in \cite{chen2014combinatorial,cao2017disagreement}, but such generality leads to unnecessary $\log$ factors. The techniques used in this work aim to reduce extraneous $\log$ factors, a topic of recent interest in the top-$1$ and top-$k$ arm identification problem \cite{even2006action,karnin2013almost,jamieson2014lil,chen2017towards,chen2017nearly,simchowitz2017simulator}. While these works are most similar to exact identification (FWER+FWPD), there also exist examples of \emph{approximate} top-$k$ where the objective is to find any $k$ means that are each within $\epsilon$ of the best $k$ means \cite{kalyanakrishnan2012pac}. Approximate recovery is also studied in a ranking context with a symmetric difference metric \cite{heckel2018approximaterank} which is more similar to the FDR and TPR setting, but neither this nor that work subsumes one another. Finally, maximizing the number of discoveries subject to a FDR constraint has been studied in a sequential setting in the context of A/B testing with uniform sampling \cite{johari2015always}. This work popularized the concept of an always valid $p$-value that we employ here (see Section~\ref{sec:algorithm}). The work of \cite{yang2017framework} controls FDR over a \emph{sequence} of independent bandit problems that each outputs at most one discovery. While \cite{yang2017framework} shares much of the same vocabulary as this paper, the problem settings are very different. \section{Algorithm and Discussion}\label{sec:algorithm} \begin{algorithm}[t] \textbf{Input:} Threshold $\mu_0$, confidence $\delta\in (0, e^{-1}]$, confidence interval $\phi(\cdot,\cdot)$\\ \textbf{Initialize:} Pull each arm $i \in [n]$ once and let $T_i(t)$ denote the number of times arm $i$ has been pulled up to time $t$. Set $\mathcal{S}_{n+1} = \emptyset$, $\mathcal{R}_{n+1} = \emptyset$, and\\ \textbf{If } TPR\\ \hspace*{.25in}$\xi_t=1$, \hspace{.2in} and \hspace{.2in} $\nu_t=1 \quad \forall t$\\ \textbf{Else if } FWPD\\ \hspace*{.25in}$\xi_t=\max\{ 2|\mathcal{S}_t|, \tfrac{5}{3(1-4\delta)} \log(1/\delta) \}$, \hspace{.2in} and \hspace{.2in} $\nu_t=\max\{ |\mathcal{S}_t|, 1\} \quad \forall t$ \\[4pt] \textbf{For} $t = n+1,n+2,\dots$ \\ \hspace*{.25in}\textbf{Pull arm} $\displaystyle I_t = \arg\max_{i \in [n] \setminus \mathcal{S}_t } \widehat{\mu}_{i,T_i(t)} + \phi(T_i(t),\tfrac{\delta}{\xi_t})$, \\ \hspace*{.25in} \textbf{Apply} Benjamini-Hochberg \cite{benjamini1995controlling} selection at level $\delta' = \tfrac{\delta}{6.4\log(36/\delta)}$ to obtain $\delta$ FDR-controlled set $\mathcal{S}_t$:\\[2pt] \hspace*{.5in}$s(k) = \{ i \in [n] : \widehat{\mu}_{i,T_i(t)} - \phi(T_i(t),\delta' \tfrac{k}{n}) \geq \mu_0 \}$, $\forall k \in [n]$ \\ \hspace*{.5in}$\displaystyle\mathcal{S}_{t+1} = s(\widehat{k}) \text{ where }\widehat{k} = \max \{ k \in [n]: |s(k)| \geq k \}$ (if $\not\exists \widehat{k}$ set $\mathcal{S}_{t+1} = \mathcal{S}_t$)\\[4pt] \hspace*{.25in}\textbf{If } FWER and $\mathcal{S}_t \neq \emptyset$:\\ \hspace*{.5in}\textbf{Pull arm} $\displaystyle J_t = \arg\max_{i \in \mathcal{S}_t\setminus \mathcal{R}_t} \widehat{\mu}_{i,T_i(t)} + \phi(T_i(t),\tfrac{\delta}{\nu_t})$ \\ \hspace*{.5in} \textbf{Apply} Bonferroni-like selection to obtain FWER-controlled set $\mathcal{R}_t$:\\[2pt] \hspace*{.75in}$\chi_t = n - (1-2\delta'(1+4\delta')) |\mathcal{S}_t| + \tfrac{4(1+4\delta')}{3}\log(5\log_2(n/\delta')/\delta')$\\ \hspace*{.75in}$\displaystyle\mathcal{R}_{t+1} = \mathcal{R}_t \cup \{ i \in \mathcal{S}_t : \widehat{\mu}_{i,T_i(t)} - \phi(T_i(t), \tfrac{\delta}{\chi_t}) \geq \mu_0 \}$ \\ \caption{\small An algorithm for identifying arms with means above a threshold $\mu_0$ using as few samples as possible subject to false alarm and true discovery conditions. The set $\mathcal{S}_t$ is designed to control FDR at level $\delta$. The set $\mathcal{R}_t$ is designed to control FWER at level $\delta$. \label{alg:FDRMAB}} \end{algorithm} Throughout, we will assume the existence of an \emph{anytime confidence interval}. Namely, if $\widehat{\mu}_{i,t}$ denotes the empirical mean of the first $t$ bounded i.i.d. rewards in $[0,1]$ from arm $i$, then for any $\delta \in (0,1)$ we assume the existence of a function $\phi$ such that for any $\delta$ we have $\mathbb{P}\left( \bigcap_{t=1}^\infty \{ |\widehat{\mu}_{i,t} - \mu_i| \leq \phi(t,\delta) \} \right) \geq 1-\delta$. We assume that $\phi( t , \delta )$ is non-increasing in its second argument and that there exists an absolute constant $c_\phi$ such that $\phi( t , \delta ) \leq \sqrt{\frac{ c_\phi \log( \log_2(2 t)/\delta)}{t}}$. It suffices to define $\phi$ with this upper bound with $c_\phi=4$ but there are much sharper known bounds that should be used in practice (e.g., they may take empirical variance into account), see \cite{jamieson2014lil,Balsubramani2014,kaufmann2016complexity,tanczosnowak2017}. Anytime bounds constructed with such a $\phi(t,\delta)$ are known to be tight in the sense that $\mathbb{P}(\bigcup_{t=1}^\infty \{|\widehat{\mu}_{i,t} - \mu_i| \geq \phi(t,\delta)\}) \leq \delta$ and that there exists an absolute constant $h \in (0,1)$ such that $\mathbb{P}(\{|\widehat{\mu}_{i,t} - \mu_i| \geq h \, \phi(t,\delta)\text{ for infinitely many $t \in \mathbb{N}$}\}) = 1$ by the Law of the Iterated Logarithm \cite{HartmanWintnerLIL}. Consider Algorithm~\ref{alg:FDRMAB}. Before entering the for loop, time-dependent variables $\xi_t$ and $\nu_t$ are defined that should be updated at each time $t$ for different settings. If just FDR control is desired, the algorithm merely loops over the three lines following the for loop, pulling the arm $I_t$ not in $\mathcal{S}_t$ that has the highest upper confidence bound; such strategies are common for pure-exploration problems \cite{jamieson2014lil,yang2017framework}. But if FWER control is desired then at most one additional arm $J_t$ is pulled per round to provide an extra layer of filtering and evidence before an arm is added to $\mathcal{R}_t$. Below we describe the main elements of the algorithm and along the way sketch out the main arguments of the analysis, shedding light on the constants $\xi_t$ and $\nu_t$. \subsection{False alarm control}\label{sec:false_alarm_control} \textbf{$\mathcal{S}_t$ is FDR-controlled.} In addition to its use as a confident bound, we can also use $\phi(t,\delta)$ to construct: \begin{align}\label{eqn:anytime_pvalue} P_{i,t} := \sup \{ \alpha \in (0,1] : \widehat{\mu}_{i,t}-\mu_0 \leq \phi(t,\alpha) \} \leq \log_2(2t) \exp(-t (\widehat{\mu}_{i,t}-\mu_0)^2/c_\phi). \end{align} Proposition~1 of \cite{yang2017framework} (and the proof of our Lemma~\ref{lem:expected_FDR}) shows that if $i \in \mathcal{H}_0$ so that $\mu_i = \mu_0$ then $P_{i,t}$ is an \emph{anytime, sub-uniformly distributed $p$-value} in the sense that $\mathbb{P}( \bigcup_{t=1}^\infty \{ P_{i,t} \leq x \} ) \leq x$. Sequences that have this property are sometimes referred to as \emph{always-valid} $p$-values \cite{johari2015always}. Note that if $i \in \mathcal{H}_1$ so that $\mu_i > \mu_0$, we would intuitively expect the sequence $\{P_{i,t}\}_{t=1}^\infty$ to be point-wise smaller than if $\mu_i = \mu_0$ by the property that $\phi(\cdot,\cdot)$ is non-increasing in its second argument. This leads to the intuitive rule to reject the null-hypothesis (i.e., declare $i \notin \mathcal{H}_0$) for those arms $i \in [n]$ where $P_{i,t}$ is very small. The Benjamini-Hochberg (BH) procedure introduced in \cite{benjamini1995controlling} proceeds by first sorting the $p$-values so that $P_{(1),T_{(1)}(t)} \leq P_{(2),T_{(2)}(t)} \leq \dots \leq P_{(n),T_{(n)}(t)}$, then defines $\widehat{k} = \max\{ k : P_{(k),T_{(k)}(t)} \leq \delta \tfrac{k}{n} \}$, and sets $\mathcal{S}_{BH} = \{ i : P_{i,T_i(t)} \leq \delta \tfrac{\widehat{k}}{n} \}$. Note that this procedure is identical to defining sets \begin{align*} s(k) = \{ i : P_{i,T_i(t)} \leq \delta \tfrac{k}{n} \} = \{ i : \widehat{\mu}_{i,T_i(t)} - \phi(T_i(t),\delta\tfrac{k}{n}) \geq \mu_0 \}, \end{align*} setting $\widehat{k} = \max\{ k : |s(k)| \geq k \}$, and $\mathcal{S}_{BH} = s(\widehat{k})$, which is exactly the set $\mathcal{S}_t = \mathcal{S}_{BH}$ in Algorithm~\ref{alg:FDRMAB}. Thus, $\mathcal{S}_t$ in Algorithm~\ref{alg:FDRMAB} is equivalent to applying the BH procedure at a level $O(\delta/\log(1/\delta))$ to the anytime $p$-values of \eqref{eqn:anytime_pvalue}. We now discuss the extra logarithmic factor. Because the algorithm is pulling arms sequentially, some dependence between the $p$-values may be introduced. Because the anytime $p$-values are not independent, the BH procedure at level $\delta$ does not directly guarantee FDR-control at level $\delta$. However, it has been shown \cite{benjamini2001control} that for even arbitrarily dependent $p$-values the BH procedure at level $\delta$ controls FDR at level $\delta \log(n)$ (and that it is nearly tight). Similarly, the following theorem, which may be of independent interest, is a significant improvement when applied to our setting. \begin{theorem}\label{thm:newFDRcontrol} Fix $\delta \in (0,e^{-1})$. Let $p_1,\dots,p_n$ be random variables such that $\{ p_i \}_{i\in \mathcal{H}_0}$ are independent and sub-uniformly distributed so that $\max_{i \in \mathcal{H}_0} \mathbb{P}(p_i \leq x ) \leq x$. For any $k \in \{0,1,\dots,n\}$, let $R_k := \{ i : p_i \leq \delta \frac{k}{n} \}$ and $\widehat{FDP}(R_k) := \tfrac{\max_{p_i \in R_k} p_i }{|R_k| \vee 1}$. \begin{align*} \mathbb{E}\left[ \max_{k : \widehat{FDP}(R_k) \leq \delta} FDP(R_k) \right] &\leq \frac{|\mathcal{H}_0|\delta}{n} \left( 2 \log( \tfrac{2n}{|\mathcal{H}_0|\delta }) + \log( 8 e^5\log( \tfrac{8n}{|\mathcal{H}_0| \delta} ) ) \right) \\ &\leq 4 \delta \log( 9/\delta) \end{align*} In other words, any procedure that chooses a set $\{ i : p_i \leq \frac{\delta k}{n} \}$ satisfying $|\{ i : p_i \leq \frac{\delta k}{n} \}| \geq k$ is FDR controlled at level $O(\delta \log(1/\delta))$. \end{theorem} Recall, if $\widehat{k} = \max\{ k : \widehat{FDP}(R_k) \leq \delta \}$ then $\mathbb{E}[ FDP(R_{\widehat{k}})] \leq \delta$ by the standard BH result. When running the algorithm we recommend using BH at level $\delta$, not level $O(\delta/\log(1/\delta))$. As $T_i$ gets very large, $P_{i,T_i(t)} \rightarrow P_{i,*}$ and we know that if BH is run on $P_{i,*}$ at level $\delta$ then FDR would be controlled at level $\delta$. We believe this inflation to be somewhat of an artifact of our proofs. \textbf{$\mathcal{R}_t$ is FWER-controlled.} A core obstacle in our analysis is the fact that we don't know the cardinality of $\mathcal{H}_1$. If we did know $|\mathcal{H}_1|$ (and equivalently know $|\mathcal{H}_0| = n - |\mathcal{H}_1|$) then a FWER+FWPD algorithm is equivalent to the so-called top-$k$ multi-armed bandit problem \cite{kalyanakrishnan2012pac,simchowitz2017simulator} and controlling FWER would be relatively simple {using a Bonferroni correction: \begin{align*} \mathbb{P}\Big( \bigcup_{i \in \mathcal{H}_0} \cup_{t=1}^\infty \{ \widehat{\mu}_{i,t} - \phi(t, \tfrac{\delta}{n-|\mathcal{H}_1|}) \geq \mu_0\}\Big) \leq \sum_{i \in \mathcal{H}_0} \mathbb{P}\left(\cup_{t=1}^\infty \{ \widehat{\mu}_{i,t} - \phi(t, \tfrac{\delta}{|\mathcal{H}_0|}) \geq \mu_0\} \right)\leq |\mathcal{H}_0| \tfrac{\delta}{|\mathcal{H}_0|} \end{align*} which implies FWER-$\delta$. Comparing the first expression immediately above to the definition of $\mathcal{R}_t$ in the algorithm, it is clear our strategy is to use $|\mathcal{S}_t|$ as a surrogate for $|\mathcal{H}_1|$. Note that we could use the bound $|\mathcal{H}_0| = n - |\mathcal{H}_1| \leq n$ to guarantee FWER-$\delta$, but this could be very loose and induce an $n\log(n)$ sample complexity. Using $|\mathcal{S}_t|$ as a surrogate for $|\mathcal{H}_1|$ in $\mathcal{R}_t$ is intuitive because by the FDR guarantee, we know $|\mathcal{H}_1| \geq \mathbb{E}[|\mathcal{S}_t \cap \mathcal{H}_1|] = \mathbb{E}[|\mathcal{S}_t|] - \mathbb{E}[|\mathcal{S}_t \cap \mathcal{H}_0|] \geq (1-\delta) \mathbb{E}[ |\mathcal{S}_t|]$, implying that $|\mathcal{H}_0| = n - |\mathcal{H}_1| \leq n - (1-\delta) \mathbb{E}[|\mathcal{S}_t|]$ which may be much tighter than $n$ if $\mathbb{E}[|\mathcal{S}_t|] \rightarrow |\mathcal{H}_1|$. Because we only know $|\mathcal{S}_t|$ and not its expectation, the extra factors in the surrogate expression used in $\mathcal{R}_t$ are used to ensure correctness with high-probability (see Lemma~\ref{lem:fwer}). \subsection{Sampling strategies to boost statistical power} The above discussion about controlling false alarms for $\mathcal{S}_t$ and $\mathcal{R}_t$ holds for \emph{any} choice of arms $I_t$ and $J_t$ that may be pulled at time $t$. Thus, $I_t$ and $J_t$ are chosen in order to minimize the amount of time necessary to add arms into $\mathcal{S}_t$ and $\mathcal{R}_t$, respectively, and optimize the sample complexity. \textbf{TPR-$\delta,\tau$ setting} implies $\xi_t=\nu_t=1$. Define the random set $\mathcal{I} = \{ i \in \mathcal{H}_1 : \widehat{\mu}_{i,T_i(t)} + \phi(T_i(t),\delta) \geq \mu_i \ \ \forall t \in \mathbb{N} \}$. Because $\phi$ is an anytime confidence bound, $\mathbb{E}\left[ \left| \mathcal{I} \right| \right] \geq (1-\delta) |\mathcal{H}_1|$. If $\Delta = \min_{i \in \mathcal{H}_1} \mu_i - \mu_0$, then $\min_{i \in \mathcal{I}} \mu_i \geq \mu_0 + \Delta$ and we claim that with probability at least $1-O(\delta)$ (Section~\ref{sec:FDR_TPR_proof}) \begin{align*} \textstyle\sum_{t=1}^\infty \mathbf{1}\{I_t \in \mathcal{H}_0, \mathcal{I} \not\subseteq \mathcal{S}_t \} &\leq \textstyle\sum_{t=1}^\infty \mathbf{1}\{I_t \in \mathcal{H}_0, \widehat{\mu}_{I_t,T_{I_t}(t)} + \phi(T_{I_t}(t),\delta) \geq \mu_0 + \Delta \}\\ &\leq c|\mathcal{H}_0| \Delta^{-2}\log(\log(\Delta^{-2}/\delta). \end{align*} Thus once this number of samples has been taken, either $\mathcal{I} \subseteq \mathcal{S}_t$, or arms in $\mathcal{I}$ will be repeatedly sampled until they are added to $\mathcal{S}_t$ since each arm $i \in \mathcal{I}$ has its upper confidence bound larger than those arms in $\mathcal{H}_0$ by definition. It is clear that an arm in $\mathcal{H}_1$ that is repeatedly sampled will eventually be added to $\mathcal{S}_t$ since its anytime $p$-value of \eqref{eqn:anytime_pvalue} approaches $0$ at an exponential rate as it is pulled, and BH selects for low $p$-values. A similar argument holds for $J_t$ and adding arms to $\mathcal{R}_t$. \begin{remark} While the main objective of Algorithm~\ref{alg:FDRMAB} is to identify all arms with means above a given threshold, we note that prior to adding an arm to $\mathcal{S}_t$ in the TPR setting (i.e., when $\xi_t=1$) Algorithm~\ref{alg:FDRMAB} behaves identically to the nearly optimal best-arm identification algorithm lil'UCB of \cite{jamieson2014lil}. Thus, whether the goal is best-arm identification or to identify all arms with means above a certain threshold, Algorithm~\ref{alg:FDRMAB} is applicable. \end{remark} \textbf{FWPD-$\delta,\tau$ setting} is more delicate and uses inflated values of $\xi_t$ and $\nu_t$. This time, we must ensure that $\{\mathcal{H}_1 \not\subseteq \mathcal{S}_t\} \implies \max_{i \in \mathcal{H}_1 \cap S_t^c} \widehat{\mu}_{i,T_i(t)} + \phi(T_i(t),\delta) \geq \min_{i \in \mathcal{H}_1 \cap \mathcal{S}_t^c} \mu_i \geq \mu_0 + \Delta$. Because then we could argue that either $\mathcal{H}_1 \subset \mathcal{S}_t$, or only arms in $\mathcal{H}_1$ are sampled until they are added to $\mathcal{S}_t$ (mirroring the TPR argument). As in the FWER setting above, if we knew the value of $|\mathcal{H}_1|$ the we could set $\xi_t \geq |\mathcal{H}_1|$ to observe that \begin{align*} \textstyle\mathbb{P}( \bigcup_{i \in \mathcal{H}_1} \cup_{t=1}^\infty \{ \widehat{\mu}_{i,t} + \phi(t, \tfrac{\delta}{\xi_t}) < \mu_i\} ) \leq \sum_{i \in \mathcal{H}_1} \mathbb{P}\left(\cup_{t=1}^\infty \{ \widehat{\mu}_{i,t} + \phi(t, \tfrac{\delta}{\xi_t}) < \mu_i\} \right)\leq |\mathcal{H}_1| \tfrac{\delta}{\xi_t} \end{align*} which is less than $\delta$, to guarantee such a condition. But we don't know $|\mathcal{H}_1|$ so we use $|\mathcal{S}_t|$ as a surrogate, resulting in the inflated definitions of $\xi_t$ and $\nu_t$ relative to the TPR setting. The key argument is that either $\mathcal{I} \not\subseteq \mathcal{S}_t$ so that $\max_{i \in \mathcal{I} \cap \mathcal{S}_t^c} \widehat{\mu}_{i,T_i(t)} + \phi(T_i(t),\tfrac{\delta}{\xi_t}) \geq \mu_0+\Delta$ by the definition of $\mathcal{I}$ (since $\xi_t \geq 1$), or $\mathcal{I} \subset \mathcal{S}_t$ and $|\mathcal{S}_t| \geq \frac{1}{2} |\mathcal{H}_1|$ with high probability which implies $\xi_t = \max\{ 2|\mathcal{S}_t|, \tfrac{5}{3(1-4\delta)} \log(1/\delta) \} \geq |\mathcal{H}_1|$ and the union bound of the display above holds. \section{Main Results}\label{sec:main_results} In what follows, we say $f \lesssim g$ if there exists a $c >0$ that is independent of all problem parameters and $f \leq c g$. The theorems provide an upper bound on the sample complexity $\tau \in \mathbb{N}$ as defined in Section~\ref{sec:problem_statement} for TPR-$\delta,\tau$ or FWER-$\delta,\tau$ that holds with probability at least $1-c \delta$ for different values of $c$\footnote{ Each theorem relies on different events holding with high probability, and consequently a different $c$ for each. To have $c=1$ for each of the four settings, we would have had to define different constants in the algorithm for each setting. We hope the reader forgives us for this attempt at minimizing clutter.}. We begin with the least restrictive setting, resulting in the smallest sample complexity of all the results presented in this work. Note the slight generalization in the below theorem where the means of $\mathcal{H}_0$ are assumed to be no greater than $\mu_0$. \begin{theorem}[FDR, TPR]\label{thm:FDR_TPR} Let $\mathcal{H}_1 = \{ i \in [n]: \mu_i > \mu_0\}$, $\mathcal{H}_0 = \{ i \in [n]: \mu_i \leq \mu_0 \}$. Define $\Delta_i = \mu_i - \mu_0$ for $i \in \mathcal{H}_1$, $\Delta = \min_{i \in \mathcal{H}_1} \Delta_i$, and $\Delta_i=\min_{j \in \mathcal{H}_1} \mu_j - \mu_i = \Delta + (\mu_0 - \mu_i)$ for $i\in\mathcal{H}_0$. For all $t \in \mathbb{N}$ we have $\mathbb{E}[\frac{|\mathcal{S}_t \cap \mathcal{H}_0|}{|\mathcal{S}_t| \vee 1}] \leq \delta$. Moreover, with probability at least $1-2\delta$ there exists a $T$ such that \begin{align*} \textstyle T \lesssim \min\big\{&n \Delta^{-2} \log( \log(\Delta^{-2})/\delta), \\ &\textstyle\sum_{i \in \mathcal{H}_0} \Delta_i^{-2} \log( \log(\Delta_i^{-2})/\delta) + \textstyle\sum_{i \in \mathcal{H}_1} \Delta_i^{-2} \log( n \log(\Delta_i^{-2})/\delta)\big\} \end{align*} and $\mathbb{E}[\frac{|\mathcal{S}_t \cap \mathcal{H}_1|}{|\mathcal{H}_1|}] \geq 1-\delta$ for all $t \geq T$. Neither argument of the minimum follows from the other. \end{theorem} If the means of $\mathcal{H}_1$ are very diverse so that $\max_{i \in \mathcal{H}_1} \mu_i-\mu_0 \gg \min_{i \in \mathcal{H}_1} \mu_i-\mu_0$ then the second argument of the min in Theorem~\ref{thm:FDR_TPR} can be tighter than the first. But as discussed above, this advantage is inconsequential if $|\mathcal{H}_1| = o(n)$. The remaining theorems are given in terms of just $\Delta$. The $\log\log(\Delta^{-2})$ dependence is due to inverting the $\phi$ confidence interval and is unavoidable on at least one arm when $\Delta$ is unknown a priori due to the law of the iterated logarithm \cite{HartmanWintnerLIL,jamieson2014lil,chen2017towards}. Informally, Theorem~\ref{thm:FDR_TPR} states that if just most true detections suffice while not making too many mistakes, then $O(n)$ samples suffice. The first argument of the min is known to be tight in a minimax sense up to doubly logarithmic factors due to the lower bound of \cite{castro2014adaptive}. As a consequence of this work, an algorithm inspired by Algorithm~\ref{alg:FDRMAB} in this setting is now in production at one of the largest A/B testing platforms on the web. The full proof of Theorem~\ref{thm:FDR_TPR} (and all others) is given in the Appendix due to space. \begin{theorem}[FDR, FWPD]\label{thm:FDR_FWPD} For all $t \in \mathbb{N}$ we have $\mathbb{E}[\frac{|\mathcal{S}_t \cap \mathcal{H}_0|}{|\mathcal{S}_t| \vee 1}] \leq \delta$. Moreover, with probability at least $1-5\delta$, there exists a $T$ such that \begin{align*} T \lesssim (n-|\mathcal{H}_1|) \Delta^{-2} \log ( \max\{|\mathcal{H}_1|,& \log\log(n/\delta)\} \log(\Delta^{-2})/ \delta) + |\mathcal{H}_1| \Delta^{-2} \log ( \log(\Delta^{-2})/ \delta) \end{align*} and $\mathcal{H}_1 \subseteq \mathcal{S}_t$ for all $t \geq T$. \end{theorem} Here $T$ roughly scales like $(n-|\mathcal{H}_1|) \max\{\log(|\mathcal{H}_1|), \log\log\log(n/\delta)\} + |\mathcal{H}_1|$ where the $\log\log\log(n/\delta)$ term comes from a high probability bound on the false discovery proportion for anytime $p$-values (in contrast to just expectation) in Lemma~\ref{lem:fdr_high_prob} that may be of independent interest. While negligible for all practical purposes, it appears unnatural and we suspect that this is an artifact of our analysis. We note that if $|\mathcal{H}_1| = \Omega(\log(n))$ then the sample complexity sheds this awkwardness\footnote{In the asymptotic $n$ regime, it is common to study the case when $|\mathcal{H}_1| = n^\beta$ for $\beta \in (0,1)$ \cite{castro2014adaptive,haupt2011distilled}.}. The next two theorems are concerned with controlling FWER on the set $\mathcal{R}_t$ and determining how long it takes before the claimed detection conditions are satisfied on the set $\mathcal{R}_t$. Note we still have that FDR is controlled on the set $\mathcal{S}_t$ but now this set feeds into $\mathcal{R}_t$. \begin{theorem}[FWER, FWPD]\label{thm:FWER_FWPD} For all $t$ we have $\mathbb{E}[\frac{|\mathcal{S}_t \cap \mathcal{H}_0|}{|\mathcal{S}_t| \vee 1}] \leq \delta$. Moreover, with probability at least $1-6\delta$, we have $\mathcal{H}_0 \cap \mathcal{R}_t = \emptyset$ for all $t \in \mathbb{N}$ and there exists a $T$ such that \begin{align*} T \lesssim& (n-|\mathcal{H}_1|) \Delta^{-2} \log ( \max\{|\mathcal{H}_1|, \log\log(n/\delta)\} \log(\Delta^{-2}) / \delta) \\ &+ |\mathcal{H}_1| \Delta^{-2} \log ( \max\{n-(1-2\delta(1+4\delta))|\mathcal{H}_1|, \log\log(n/\delta)\}\log(\Delta^{-2}) / \delta) \end{align*} and $\mathcal{H}_1 \subseteq \mathcal{R}_t$ for all $t \geq T$. Note, together this implies $\mathcal{H}_1 = \mathcal{R}_t$ for all $t \geq T$. \end{theorem} Theorem~\ref{thm:FWER_FWPD} has the strongest conditions, and therefore the largest sample complexity. Ignoring $\log\log\log(n)$ factors, $T$ roughly scales as $(n-|\mathcal{H}_1|) \log(|\mathcal{H}_1|) + |\mathcal{H}_1| \log(n-(1-2\delta(1+4\delta))|\mathcal{H}_1|)$. Inspecting the top-k lower bound of \cite{simchowitz2017simulator} where the arms' means in $\mathcal{H}_1$ are equal to $\mu_0 + \Delta$, the arms' means in $\mathcal{H}_0$ are equal to $\mu_0$, and the algorithm has knowledge of the cardinality of $\mathcal{H}_1$, a necessary sample complexity of $(n-|\mathcal{H}_1|)\log(|\mathcal{H}_1|) + |\mathcal{H}_1| \log(n-|\mathcal{H}_1|)$ is given. It is not clear whether this small difference of $\log(n-(1-2\delta(1+4\delta)) |\mathcal{H}_1|)$ versus $\log(n-|\mathcal{H}_1|)$ is an artifact of our analysis, or a fundamental limitation when the cardinality $|\mathcal{H}_1|$ is unknown. We now state our final theorem. \begin{theorem}[FWER, TPR]\label{thm:FWER_TPR} For all $t$ we have $\mathbb{E}[\frac{|\mathcal{S}_t \cap \mathcal{H}_0|}{|\mathcal{S}_t| \vee 1}] \leq \delta$. Moreover, with probability at least $1-7\delta$ we have $\mathcal{H}_0 \cap \mathcal{R}_t = \emptyset$ for all $t \in \mathbb{N}$ and there exists a $T$ such that \begin{align*} T \lesssim& (n-|\mathcal{H}_1|) \Delta^{-2} \log ( \log(\Delta^{-2})/ \delta) \\ &+ |\mathcal{H}_1| \Delta^{-2} \log ( \max\{n-(1-\eta)|\mathcal{H}_1|, \log\log(n\log(1/\delta)/\delta)\}\log(\Delta^{-2}) / \delta) \end{align*} and $\mathbb{E}[\frac{|\mathcal{R}_t \cap \mathcal{H}_1|}{|\mathcal{H}_1|}] \geq 1-\delta$ for all $t \geq T$, where $\eta=( 1 - 3\delta- \sqrt{2\delta \log(1/\delta)/|\mathcal{H}_1|})$. \end{theorem} \section{Experiments} The distribution of each arm equals $\nu_i = \mathcal{N}(\mu_i,1)$ where $\mu_i = \mu_0 = 0$ if $i \in \mathcal{H}_0$, and $\mu_i>0$ if $i \in \mathcal{H}_1$. We consider three algorithms: $i$) uniform allocation with anytime BH selection as done in Algorithm~1, $ii$) successive elimination (SE) (see Appendix~\ref{sec:succ-elim})\footnote{Inspired by the best-arm identification literature \cite{even2006action}.} that performs uniform allocation on only those arms that have not yet been selected by BH, and $iii$) Algorithm 1 (UCB). Algorithm 1 and the BH selection rule for all algorithms use $\phi(t,\delta) = \sqrt{\tfrac{2\log(1/\delta)+6 \log\log(1/\delta) + 3 \log(\log(e t/2))}{t}}$ from \cite[Theorem 8]{kaufmann2016complexity}. In addition, we ran BH at level $\delta$ instead of $\delta/(6.4\log(36/\delta))$ as discussed in section \ref{sec:main_results}. Here we present the sample complexity for TPR+FDR with $\delta=0.05$ and different parameterizations of $\mu$, $n$, $|\mathcal{H}_1|$.\\[6pt] \begin{tabular}{l l l l l} \includegraphics[width=.252\textwidth,trim={.6cm .7cm .7cm .7cm},clip]{plot_TPR_time.png} & \includegraphics[width=.201\textwidth,trim={.3cm .7cm .7cm .4cm},clip]{plot_H1eq2.png} & \includegraphics[width=.18\textwidth,trim={1.3cm .7cm .7cm .4cm},clip]{plot_H1sqrtN.png} & \includegraphics[width=.18\textwidth,trim={1.3cm .7cm .7cm .4cm},clip]{plot_H1eqN5.png} & \includegraphics[width=.18\textwidth,trim={1.3cm .7cm .7cm .4cm},clip]{plot_H1var.png} \end{tabular} The first panel shows an empirical estimate of $\mathbb{E}[ \frac{|\mathcal{S}_t \cap \mathcal{H}_1|}{|\mathcal{H}_1|} ]$ at each time $t$ for each algorithm, averaged over 1000 trials. The black dashed line on the first panel denotes the level $\mathbb{E}[ \frac{|\mathcal{S}_t \cap \mathcal{H}_1|}{|\mathcal{H}_1|} ]=1-\delta = .95$, and corresponds to the dashed black line on the second panel. The right four panels show the number of samples each algorithm takes before the true positive rate exceeds $1-\delta=.95$, relative to the number of samples taken by UCB, for various parameterizations. Panels two, three, and four have $\Delta_i=\Delta$ for $i \in \mathcal{H}_1$ while panel five is a case where the $\Delta_i$'s are linear for $i \in \mathcal{H}_1$. While the differences are most clear on the second panel when $|\mathcal{H}_1| =2= o(n)$, over all cases UCB uses at least $\approx3$ times fewer samples than uniform and SE. For FDR+TPR, Appendix~\ref{sec:succ-elim} shows uniform sampling roughly has a sample complexity that scales like $n \Delta^{-2} \log(\tfrac{n}{|\mathcal{H}_1|})$ while SE's is upper bounded by $\min\{ n \Delta^{-2} \log(\tfrac{n}{|\mathcal{H}_1|}), (n-|\mathcal{H}_1|)\Delta^{-2} \log(\tfrac{n}{|\mathcal{H}_1|}) + \sum_{i \in \mathcal{H}_1} \Delta_i^{-2} \log(n )\}$. Comparing with Theorem~\ref{thm:FDR_TPR} for the difference cases (i.e., $|\mathcal{H}_1| =2, \sqrt{n}, n/5$) provides insight into the relative difference between UCB, uniform, and SE on the different panels. \subsection*{Acknowledgments} This work was informed and inspired by early discussions with Aaditya Ramdas on methods for controlling the false discovery rate (FDR) in multiple testing; we are grateful to have learned from a leader in the field. We also thank him for his careful reading and feedback. We'd also like to thank Martin J. Zhang for his input. We also thank the leading experimentation and A/B testing platform on the web, \emph{Optimizely}, for its support, insight into its customers' needs, and for committing engineering time to implementing this research into their platform \cite{optimizely}. In particular, we thank Whelan Boyd, Jimmy Jin, Pete Koomen, Sammy Lee, Ajith Mascarenhas, Sonesh Surana, and Hao Xia at Optimizely for their efforts. \clearpage \bibliographystyle{unsrt}

cc2f26dac2ea215f5dc4bd1129a069d90f61e262

\section{Introduction} \label{sec:intro} As we approach the end of the semiconductor roadmap~\cite{semiconductor2015}, we are entering a regime in which fundamental thermodynamic considerations limit the sub-threshold slope, practical switching voltages, and gate energies---implying that further downscaling of device sizes and gate capacitances will soon no longer yield improvements in energy efficiency for conventional logic. Industry's shift towards 3D geometries~\cite{semiconductor2015} will somewhat reduce parasitic energy losses in circuit structures, but once that line of improvements is played out, the only remaining approach to further increase energy efficiency will be to begin applying techniques of energy recovery. In this regard, resonant circuit techniques to recycle and reuse logic signal energies, rather than dissipating the entire $\frac{1}{2}CV^2$ circuit node energy on each logic-level transition, are promising. Unlike all other options, no fundamental theoretical limits on the ultimate energy efficiency of energy recovery are currently known for this direction---offering a path towards future growth of computing performance within any given energy dissipation constraints. But apparently the ideal of 100\% energy recovery implies that all switching activity of a device must be carried out in a manner that is asymptotically adiabatic---avoiding any abrupt loss of signal energy to heat. This motivated the consideration of \emph{adiabatic circuits} which allow for computations with an asymptotically close to zero energy dissipation (at the expense of a slower execution). Due to Landauer's limit~\cite{Landauer61}, this in turn implies that the computational function of the switching circuit must be \emph{logically reversible}, in the appropriately generalized sense discussed in~\cite{DBLP:conf/rc/Frank17}. Otherwise, the information lost from a conventional circuit leads to an entropy increase and, therefore, to an irreducible energy dissipation. This was recently also advocated to a larger community in~\cite{frank2017reversible} stating that the future of computing depends on reversible computations. While these concepts have already been around for a while---general techniques for designing fully-adiabatic and reversible circuits have been introduced in the 1990's and resulted in a large body of literature (see e.g.~\cite{koller1992adiabatic,hall1992electroid,merkle1992towards,younis1993practical})---most of the adiabatic design families that have been proposed contain flaws preventing them from being truly adiabatic~\cite{DBLP:conf/csreaESA/Frank03}. In this regard, \emph{two-level adiabatic logic} (2LAL as proposed in~\cite{anantharam2004driving}) represents a very promising, \mbox{fully-adiabatic} transmission-gate logic family that relies on simple but rather efficient building blocks. However, to realize correct adiabatic and reversible circuit designs that could truly approach arbitrarily low levels of energy dissipation requires to satisfy certain \emph{switching rules} which differ from the design of conventional circuitry---crying out for automated approaches for the design of such adiabatic circuits. Heading into this direction recently also gained relevance in industry---triggered e.g.~by investments of funding agencies and national departments~\cite{frank2017reversible}. Accordingly, researchers started to work towards such solutions. However, previously proposed approaches either focus on their electrical realization (see e.g.~\cite{younis1993practical,anantharam2004driving}) or on designing purely reversible building blocks like Toffoli gates (see e.g.~\cite{DBLP:journals/tcad/MorrisonR14,rauchenecker2017exploiting} in combination with corresponding synthesis approaches such as~\cite{zulehner2017one,ZulehnerW18Exploiting}). While the former approaches are restricted to small and \mbox{hand-crafted} circuits only, relying on purely reversible building blocks results in an unnecessarily large overhead. Instead, recent findings (summarized in~\cite{DBLP:conf/rc/Frank17}) show that conditional reversibility is sufficient for adiabatic circuits. But thus far, no design automation approach for adiabatic circuits exists which exploits that in an automatic fashion. In this work, we overcome this issue by combining expertise from both adiabatic circuits and design automation. More precisely, we review the theoretical and technical background of adiabatic circuits and, based on that, propose an automatic \emph{and} dedicated design flow for this promising technology. Two complementary design styles (namely retractile and fully-pipelined) are thereby considered which allow for the generation of adiabatic circuits either focusing on reducing the number of gates or keeping the number of so-called power clocks small. Furthermore, optimizations for both design styles are proposed which utilize application-specific properties and, by this, allow e.g.~for a reduction in the number of gates by approx. 37\% and 30\% on average for the retractile and fully-pipelined design styles, respectively. Evaluations confirm the benefits and applicability of the proposed solution. The remainder of this work is structured as follows: Section~\ref{sec:background} provides a review of the theoretical and technical background of adiabatic circuits. Based on that, the proposed design flow is introduced in Section~\ref{sec:flow} followed by the descriptions of the respective mapping methods following the retractile and fully-pipelined design style in Section~\ref{sec:retractile} and Section~\ref{sec:pipelined}, respectively. Finally, a summary of the results from our evaluations is given in Section~\ref{sec:exp} and the paper is concluded in Section~\ref{sec:conclusions}. \vfill \section{Adiabatic Circuits} \label{sec:background} In this work, we consider design automation for adiabatic circuits according to the \emph{two-level adiabatic logic} (2LAL,~\cite{anantharam2004driving}) circuit family. This type of adiabatic circuit uses only two different voltage levels and heavily relies on transmission gates. Furthermore, a dual-rail encoding is used for the signals of the circuit, i.e.~each signal occurs in uncomplemented as well as in complemented form.\footnote{The uncomplemented form of a signal is labeled with a subscript $N$ and the complemented form is labeled with subscript $P$.} Fig.~\ref{fig:transmission_gate} provides the notation for transmission gates: If the signal $P$ is 1 (the gate is \emph{turned on}), $A$ and $B$ are connected.\footnote{Note that logic 1 (i.e.~$X=1$) is realized by $X_N = 1$ and $X_P=0$ since a dual-rail encoding is employed.} Otherwise (the gate is \emph{turned off}), $A$ and $B$ are not connected. Since $A$ and~$B$ are both encoded in a dual-rail fashion and, thus, have an uncomplemented as well as complemented form, two transmission gates as shown in the right-hand side of Fig.~\ref{fig:transmission_gate} are required.\footnote{For sake of simplicity, we abstract the two transmission gates in the following illustrations and use the more compact form as shown in the left-hand side of Fig.~\ref{fig:transmission_gate} instead.} The general \emph{switching rules} for transistors in adiabatic circuits (e.g.~outlined in~\cite{anantharam2004driving,DBLP:conf/csreaESA/Frank03}) imply that a transmission gate shall never be turned on if $A$ and~$B$ have different values. \begin{figure}\centering \scalebox{0.9}{ \begin{tikzpicture} \draw[line width=0.75pt] (0,0) rectangle ++(1,0.5); \draw (-0.25,0.25) node[left] {A} -- (0, 0.25); \draw (1,0.25) -- (1.25,0.25) node[right] {$B$}; \draw (0.5,0.5) -- (0.5, 0.75) node[above] {$P$} ; \draw (3.25,0.25) node[left] {$A_P$} -| (3.5,0.4) -| (4,0.25) -- (4.25,0.25) node[right] {$B_P$}; \draw (3.5,0.25) |- (4,0.1) -- (4, 0.25); \draw (3.5,0.05) -- (4, 0.05); \draw (3.5,0.45) -- (4, 0.45); \draw[fill=white] (3.75,0.5) circle(0.05); \draw (3.75,0.55) -- (3.75,0.7); \draw (3.75,0.05) -- (3.75,-0.1); \draw[line width = 1pt] (2,0.125) -- (2.25,0.125); \draw[line width = 1pt] (2,0.25) -- (2.25,0.25); \draw[line width = 1pt] (2,0.375) -- (2.25,0.375); \draw (5.5,0.25) node[left] {$A_N$} -| (5.75,0.4) -| (6.25,0.25) -- (6.5,0.25) node[right] {$B_N$}; \draw (5.75,0.25) |- (6.25,0.1) -- (6.25, 0.25); \draw (5.75,0.05) -- (6.25, 0.05); \draw (5.75,0.45) -- (6.25, 0.45); \draw[fill=white] (6,0.5) circle(0.05); \draw (6,0.55) -- (6,0.7); \draw (6,0.05) -- (6,-0.1); \draw (6,-0.1) -- (3.75,-0.1); \draw (4.875,-0.1) -- (4.875,-0.25) node[below] {$P_N$}; \draw (6,0.7) -- (3.75,0.7); \draw (4.875,0.7) -- (4.875,0.85) node[above] {$P_P$}; \end{tikzpicture}} \vspace*{-3mm} \caption{Transmission gate for dual-rail signals} \label{fig:transmission_gate} \vspace*{-3mm} \end{figure} Besides that, so-called \emph{power clocks} (denoted $\phi_i$) are additionally utilized to realize typical functions such as OR or AND. More precisely, the inputs of the gate control a network of transmission gates which connect the output $Y$ of the gate to one of the power clocks~$\phi_i$ in case the function to be realized evaluates to 1. To obey the switching rules, the output $Y$ of the gate as well as the power clock $\phi_i$ are assumed to be 0 initially. By transitioning the power clock to 1, the output of the gate is set to the desired value. Moreover, when resetting all inputs of a gate to 0 (and, thus, disconnecting $\phi_i$ and $Y$) while $\phi_i$ is still 1, the output preserves its value (even if resetting $\phi_i$ to 0 afterwards). This allows for an inherent \emph{latching} of an output value to be used by following gates. An example illustrates the idea: \begin{myexample} Fig.~\ref{fig:gates} shows the 2LAL realization of an OR gate and an AND gate. The OR gate is composed of two parallel transmission gates whose outputs are connected. In case $A=1$ ($B=1$), the upper (the lower) transmission gate is turned on and connects the power clock~$\phi_i$ to the output $Y$. Consequently, $Y$ is connected to $\phi_i$ if \mbox{$A+B=1$}. Transitioning $\phi_i$ to 1 sets $Y$ to the desired value. If we now reset the inputs $A$ and $B$ to 0, the output $Y$ is latched---its value is preserved even when setting $\phi_i$ back to 0 afterwards. The AND gate is realized similarly as a sequence of two transmission gates. Note that a second output $Y_2=A$ is required in this case to operate the gate in an adiabatic fashion in case $A=1$ and $B=0$ when used in a \mbox{fully-pipelined} circuit (cf.~Section~\ref{sec:pipelined}). \end{myexample} \begin{figure} \begin{subfigure}[b]{0.44\linewidth} \centering \scalebox{0.7}{ \begin{tikzpicture} \draw[line width=0.75pt] (0,0) rectangle ++(1,0.5); \draw[line width=0.75pt] (0,-0.75) rectangle ++(1,0.5); \draw (0, 0.25) -- (-0.25,0.25) |- (0, -0.5) ; \draw (1,0.25) -- (1.25,0.25) |- (1,-0.5); \draw (0.5,0.5) -- (0.5, 0.75) node[above] {$A$}; \draw (0.5,-0.75) -- (0.5, -1) node[below] {$B$}; \draw (-0.25,-0.125) -- (-0.5,-0.125) node[left] {$\phi_i$}; \draw (1.25,-0.125) -- (1.5,-0.125) node[right] {$Y=A+B$}; \end{tikzpicture}} \vspace*{-1mm} \caption{OR gate}\label{fig:gates_or} \end{subfigure} \begin{subfigure}[b]{0.55\linewidth} \centering \scalebox{0.7}{ \begin{tikzpicture} \draw[line width=0.75pt] (0,0) rectangle ++(1,0.5); \draw[line width=0.75pt] (1.5,0) rectangle ++(1,0.5); \draw (0, 0.25) -- (-0.25,0.25) node[left] {$\phi_i$}; \draw (1,0.25) -- (1.5,0.25); \draw (0.5,0.5) -- (0.5, 0.75) node[above] {$A$}; \draw (2,0.5) -- (2, 0.75) node[above] {$B$}; \draw (2.5,0.25) -- (2.75,0.25) node[right] {$Y_1=A\cdot B$}; \draw (1.25,0.25) |- (2.75,-0.25) node[right]{$Y_2=A$}; \draw[draw=none] (0.5, -0.625) node[below] {\phantom{$B$}}; \end{tikzpicture}} \vspace*{-1mm} \caption{AND gate} \end{subfigure} \vspace*{-7mm} \caption{Adiabatic gates} \label{fig:gates} \vspace*{-3mm} \end{figure} Once the output of a gate is not needed anymore (e.g.~by a following gate), an essential step for adiabatic circuits is the ability to decompute it---feeding charge back to the power clocks. In case that the output was not latched (i.e.~the output is still connected to the power clock), it is decomputable by simply resetting the power clock to 0 (as discussed above). If the output is latched (i.e.~it was disconnected from the power clock by setting the inputs back to~0), the power clock has to be transitioned to 1 as well, before the inputs are applied in order to obey the switching rules. Then, the output is decomputable by transitioning the power clock back to 0. \begin{myexample}\label{ex:decompute} Consider again the 2LAL realization of an OR gate (cf.~Fig.~\ref{fig:gates_or}). Assume that the output $Y=A+B$ of the gate is latched and that all other signals are set to~0. To unlatch the output $Y$, we first have to set the power clock $\phi_i$ to 1. By this, $\phi_i$ and $Y$ have the same value if they get connected by resetting the inputs to their original value. Then, $Y$ is decomputable by changing the power clock $\phi_i$ back to 0---the charge representing $Y=1$ is fed back to the power supply. \end{myexample} Following this main principle allows for conducting operations with an asymptotically close to zero energy dissipation (at the expense of a slower execution since more steps have to be conducted). In fact, in contrast to conventional circuits in which energy is frequently ``grounded'', adiabatic circuits allow for feeding energy back to the clocks providing the power supply. However, this concept of feeding back charge to the power clocks by decomputing signals demands for a logical reversibility of the underlying computations. This is because, in order to not violate the switching rules, the original input assignments have to be applied so that signals with different values are never connected (cf.~Example~\ref{ex:decompute}). While in the past, a pure reversible scheme has been assumed (see e.g.~\cite{DBLP:journals/tcad/MorrisonR14,rauchenecker2017exploiting}), findings recently summarized in~\cite{DBLP:conf/rc/Frank17} showed that conditional reversibility is actually sufficient for adiabatic circuits. Again, this is illustrated by means of an example: \begin{myexample} Consider again the OR gate shown in Fig.~\ref{fig:gates_or}. Considering the state of the signals $A$, $B$, and $Y$, the gate describes a function $f:\mathbb{B}^3 \rightarrow \mathbb{B}^3 = (A,B,Y) \rightarrow (A,B,A+B)$. This function is not reversible in general, since the initial value of $Y$ can not be computed from the output values. However, the function is conditionally reversible under the precondition that the value of $Y$ is initially set to 0, i.e.~an input combination like e.g.~$(1,0,1)$ can never occur. Conditional reversibility is a much weaker constraint than unconditional reversibility (as e.g.~considered in~\cite{DBLP:journals/tcad/MorrisonR14,rauchenecker2017exploiting})---allowing to realize adiabatic gates as e.g.~shown in Fig.~\ref{fig:gates}. \end{myexample} Obviously, conducting computations in such a fashion requires the corresponding circuits to be designed in a significantly different fashion than conventional circuitry. Besides the generation of a proper netlist composed of transmission gates, this additionally requires dedicated power clocks which correspondingly trigger the required operations at the correct point in time. Moreover, also the design objectives change. While the number of required (transmission) gates is still a factor (e.g.~to approximate the required area), their impact on energy consumption is smaller than for conventional circuits. This is because energy is never grounded in adiabatic circuits but frequently fed back to the power supply as described above. In contrast, the number of power clocks is much more crucial as they are the entities which actually require energy and whose waveform might be hard to generate. Besides that, more clocks usually also require longer execution times. \section{Proposed Design Flow}\label{sec:flow} As discussed above, previous design methods for designing adiabatic circuits (e.g.~\cite{DBLP:journals/tcad/MorrisonR14,rauchenecker2017exploiting}) assumed the requirement of full reversibility. As recently discussed in~\cite{DBLP:conf/rc/Frank17}, this leads to a significant overhead and is not necessarily needed. In fact, conditional reversibility as reviewed above is sufficient and constitutes a much weaker constraint. However, thus far, no design automation for this kind of adiabatic circuits exists. Also, solely employing conventional design solutions is not an option since, despite the pure functionality, a dedicated mapping and clocking scheme is required. In this work, we present different methods which address these issues. All of them employ thereby a two-stage process. The first step is similar to the design of conventional circuits: We realize the function to be synthesized with respect to a certain logic gate library. Afterwards, the resulting netlist is mapped to an adiabatic circuit which respectively satisfies and optimizes the rules and objectives reviewed in Section~\ref{sec:background}. For the first part, we utilize a solution based on \emph{AND-Inverter Graphs} (AIGs~\cite{KPKG:2002}) which realize the function to be synthesized in terms of NAND gates.\footnote{Note that the design methods proposed in this work can correspondingly be adjusted to any other synthesis solution and, hence, logic gate library as well.} AIGs allow for a graph-based representation of Boolean functions. The graph has one root node for each output of the function. The inputs of the function are provided as terminals. The intermediate nodes of an AIG represent an AND operation and, thus, have two successors each. To gain universality, the inputs of the AND operation can be inverted. This is denoted by black circles on the respective edges. Equal nodes occur frequently and can be shared---allowing for a compact representation of the function to be realized. \begin{myexample} Fig.~\ref{fig:aig} shows the AIG of a 3-input 2-output Boolean function with inputs $x_2$, $x_1$, and $x_0$ as well as outputs $y_1$ and $y_0$ which represent $y_1 = \overline{x}_2\overline{x}_1 + \overline{x}_2x_0 + \overline{x}_1x_0$ and $y_0 = \overline{x}_2x_1+x_1x_0+x_2\overline{x}_1\overline{x}_0$ in terms of an AIG and, hence, NAND operations. \end{myexample} How to determine and optimize an AIG (e.g.~minimizing its number of nodes/gates) has intensely been considered in the literature (see e.g.~\cite{DBLP:conf/dac/MishchenkoCB06}) and, hence, is not covered further in the following. Instead, we focus on the second step, i.e.~how to map the resulting NAND netlist to an adiabatic circuit, i.e.~a network of transmission gates and the corresponding power clocks. To this end, we translate the AIG into an \emph{OR-Inverter graph} (OIG) so that a NOR gate netlist results. An OIG can easily be derived from an AIG by simply applying De Morgan's laws, i.e.~by relabeling the inner nodes from AND to OR and inverting the polarity of the edges to the terminals and the edges to the root nodes (cf.~\ref{fig:oig}). \begin{figure} \begin{subfigure}[b]{0.49\linewidth} \centering \scalebox{0.75}{ \centering \begin{tikzpicture}[terminal/.style={draw,rectangle,inner sep=2pt}] \matrix[matrix of nodes,ampersand replacement=\&,every node/.style={vertex},column sep={1.5cm,between origins},row sep={0.9cm,between origins}] (qmdd) { \node[regular polygon,regular polygon sides=4, inner sep=1pt] (y1) {$y_1$}; \& \node[regular polygon,regular polygon sides=4, inner sep=1pt] (y0) {$y_0$}; \& \\ \node(n1a) {$\wedge_6$}; \& \node(n1b) {$\wedge_7$}; \& \\ \& \node (n2a) {$\wedge_4$}; \& \node (n2b) {$\wedge_5$}; \\ \node(n3a) {$\wedge_2$}; \& \node(n3b) {$\wedge_3$}; \& \\ \& \node(n4) {$\wedge_1$}; \& \& \\ \node[regular polygon,regular polygon sides=3,draw,inner sep=0pt] (x2) {$x_2$}; \& \node[regular polygon,regular polygon sides=3,draw,inner sep=0pt] (x1) {$x_1$}; \& \node[regular polygon,regular polygon sides=3,draw,inner sep=0pt] (x0) {$x_0$}; \\ }; \draw (x2.north) -- (n3a); \draw (x2.north) -- (n4) node[midway, circle, inner sep=1pt, fill] {}; \draw (x1.north) -- (n3a); \draw (x1.north) -- (n4) node[midway, circle, inner sep=1pt, fill] {}; \draw (x1.north) -- (n2b); \draw (x0.north) -- (n3b) node[midway, circle, inner sep=1pt, fill] {}; \draw (x0.north) -- (n2b); \draw (n4) -- (n3b) node[midway, circle, inner sep=1pt, fill] {}; \draw (n3a) -- (n1a) node[midway, circle, inner sep=1pt, fill] {}; \draw (n3a) -- (n2a) node[midway, circle, inner sep=1pt, fill] {}; \draw (n3b) -- (n1a) node[midway, circle, inner sep=1pt, fill] {}; \draw (n3b) -- (n2a); \draw (n2a) -- (n1b) node[midway, circle, inner sep=1pt, fill] {}; \draw (n2b) -- (n1b) node[midway, circle, inner sep=1pt, fill] {}; \draw (n1a) -- (y1); \draw (n1b) -- (y0) node[midway, circle, inner sep=1pt, fill] {}; \end{tikzpicture}} \caption{AND-Inverter graph} \label{fig:aig} \end{subfigure} \begin{subfigure}[b]{0.49\linewidth} \centering \scalebox{0.75}{ \centering \begin{tikzpicture}[terminal/.style={draw,rectangle,inner sep=2pt}] \matrix[matrix of nodes,ampersand replacement=\&,every node/.style={vertex},column sep={1.5cm,between origins},row sep={0.9cm,between origins}] (qmdd) { \node[regular polygon,regular polygon sides=4, inner sep=1pt] (y1) {$y_1$}; \& \node[regular polygon,regular polygon sides=4, inner sep=1pt] (y0) {$y_0$}; \& \\ \node(n1a) {$\vee_6$}; \& \node(n1b) {$\vee_7$}; \& \\ \& \node (n2a) {$\vee_4$}; \& \node (n2b) {$\vee_5$}; \\ \node(n3a) {$\vee_2$}; \& \node(n3b) {$\vee_3$}; \& \\ \& \node(n4) {$\vee_1$}; \& \& \\ \node[regular polygon,regular polygon sides=3,draw,inner sep=0pt] (x2) {$x_2$}; \& \node[regular polygon,regular polygon sides=3,draw,inner sep=0pt] (x1) {$x_1$}; \& \node[regular polygon,regular polygon sides=3,draw,inner sep=0pt] (x0) {$x_0$}; \\ }; \draw (x2.north) -- (n3a) node[midway, circle, inner sep=1pt, fill] {}; \draw (x2.north) -- (n4); \draw (x1.north) -- (n3a) node[midway, circle, inner sep=1pt, fill] {}; \draw (x1.north) -- (n4); \draw (x1.north) -- (n2b) node[midway, circle, inner sep=1pt, fill] {}; \draw (x0.north) -- (n3b); \draw (x0.north) -- (n2b) node[midway, circle, inner sep=1pt, fill] {}; \draw (n4) -- (n3b) node[midway, circle, inner sep=1pt, fill] {}; \draw (n3a) -- (n1a) node[midway, circle, inner sep=1pt, fill] {}; \draw (n3a) -- (n2a) node[midway, circle, inner sep=1pt, fill] {}; \draw (n3b) -- (n1a) node[midway, circle, inner sep=1pt, fill] {}; \draw (n3b) -- (n2a); \draw (n2a) -- (n1b) node[midway, circle, inner sep=1pt, fill] {}; \draw (n2b) -- (n1b) node[midway, circle, inner sep=1pt, fill] {}; \draw (n1a) -- (y1) node[midway, circle, inner sep=1pt, fill] {}; \draw (n1b) -- (y0); \end{tikzpicture}} \caption{OR-Inverter graph} \label{fig:oig} \end{subfigure} \vspace*{-2mm} \caption{Graph representations for Boolean functions} \vspace*{-2mm} \end{figure} Now, the nodes of an OIG can directly be mapped to the adiabatic OR gates introduced before in Fig~\ref{fig:gates_or}. However, it remains open and non-trivial how to connect these gates to the power clocks and how to generate a corresponding waveform of these clocks (again, following the switching rules and optimization objectives reviewed in Section~\ref{sec:background}). To this end, two (complementary) design styles are considered: \emph{retractile} circuits (cf.~\cite{hall1992electroid}) as well as \mbox{\emph{fully-pipelined}} circuits (cf.~\cite{younis1993practical,anantharam2004driving,DBLP:conf/rc/Frank17}). Note that for both design styles the conditional reversibility is inherently satisfied by preserving the inputs of the signals throughout the whole computation and by assuming that all additional (intermediate) signals are initially set to 0. In the following sections, we discuss advantages and disadvantages of both design styles and present according (automatic) mapping schemes. More precisely, for each design style we first describe a straightforward mapping scheme (conveying the main idea of the design style) followed by an advanced mapping scheme (which results in a significantly smaller number of gates as well as, in case of retractile circuits, to a smaller number of power clocks). These considerations eventually motivate the implementation of different methods for design automation of adiabatic circuits whose performance is eventually discussed in Section~\ref{sec:exp}. \section{Retractile Circuits} \label{sec:retractile} \subsection{Straightforward Solution} \label{sec:retractile_sf} The straightforward mapping for retractile circuits is similar to conventional circuitry, where an AIG or OIG is directly mapped to the target technology. In fact, we can realize each node of the OIG with an OR gate and negations with inverters. Moreover, in case of adiabatic circuits, the inverters come ``for free'' since we are operating on dual-rail signals and, hence, an inverted input can easily be realized with no further hardware by swapping the rails of the signal. \begin{myexample} Consider again the OIG depicted in Fig.~\ref{fig:oig}. Mapping the OIG to conventional gates results in the circuit shown in Fig.~\ref{fig:circuit_retractile_gates}. Doing this mapping for adiabatic circuits following the retractile design style, each OR-gate is realized with two transmission gates as discussed in Section~\ref{sec:background}. \end{myexample} To operate the circuit in an adiabatic fashion, all intermediate signals are first initialized with~0. Furthermore, each stage $s_i$ ($0 \le i < N$) of the circuit with depth $N$ has an associated dual-rail encoded clock $\phi_i$---allowing to compute the individual stages sequentially. Then, the computations are started by transitioning the $0^{th}$ clock from 0 to 1---triggering the desired operations of the first stage. Once stable, the operations of the next stages are sequentially triggered. To allow for decomputing the intermediate results, the clocks transition back to 0 in reverse order, i.e.~first the $N-1^{th}$ clock is set back to~0, then the other ones. This way, all intermediate results are decomputed and restored back to 0. Overall, this requires $2N+1$ time steps for a single computation (assuming one additional time step is required to process the outputs of the circuit). During these time steps, the inputs have to remain constant---yielding a rather low throughput. \addtocounter{myexample}{-1} \begin{myexample}[continued] Since the resulting circuit has four stages (the OIG has a depth of 4), we need four different clocks (eight if we take the dual-rail encoding into account). The waveforms of these clocks are shown in Fig.~\ref{fig:clocks}. Overall, this causes that a single computation of this circuit requires 9 timesteps. \end{myexample} \begin{figure} \begin{subfigure}{0.58\linewidth} \centering \scalebox{0.7}{ \begin{tikzpicture} \node (x2) at (0, 2) {$x_2$}; \node (x1) at (0, 1) {$x_1$}; \node (x0) at (0, 0) {$x_0$}; \node[or gate US, draw, anchor=input 1] at ($(x2) + (2.5, 0)$) (or2) {$g_2$}; \node[or gate US, draw, anchor=input 2] at ($(x1) + (1, 0)$) (or1) {$g_1$}; \node[or gate US, draw, anchor=input 1] at (2.5,0 |- or1.output) (or3) {$g_3$}; \node[or gate US, draw, anchor=input 2] at (4,0 |- or3.output) (or4) {$g_4$}; \node[or gate US, draw, anchor=input 2] at ($(x0) + (4, 0)$) (or5) {$g_5$}; \node[or gate US, draw, anchor=input 1] at (5.5,0 |- or2.output) (or6) {$g_6$}; \coordinate (pos7) at ($(or4.output)!.5!(or5.output)$); \node[or gate US, draw, anchor=output] at (or6.output |- pos7) (or7) {$g_7$}; \draw (x2) -- (or2.input 1); \draw (x1) -- (or1.input 2); \draw (or1.output) -- (or3.input 1); \draw (x0) -- (or5.input 2); \draw (or3.output) -- (or4.input 2); \draw (or2.output) -- (or6.input 1); \draw ($(x2)+(0.75,0)$) node[branch] {} |- ($(or1.input 1)$); \draw ($(x1)+(0.5,0)$) node[branch] {} |- ($(or2.input 2)$); \draw ($(x0)+(2,0)$) node[branch] {} |- ($(or3.input 2)$); \draw ($(or2.output)+(0.5,0)$) node[branch] {} |- ($(or4.input 1)$); \draw ($(or3.output)+(0.25,0)$) node[branch] {} |- ($(or6.input 2)$); \draw ($(x1)+(0.5,0)$) |- ($(or5.input 1)$); \draw (or5.output) -- ++ (0.5,0) |- ($(or7.input 2)$); \draw (or4.output) -- ++ (0.5,0) |- ($(or7.input 1)$); \draw (or6.output) -- ++ (0.25,0) node[anchor=west] {$y_1$}; \draw (or7.output) -- ++ (0.25,0) node[anchor=west] {$y_0$}; \draw[line width=0.4, fill=white] ($(or2.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or2.input 2)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or3.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or4.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or5.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or5.input 2)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or6.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or6.input 2)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or7.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or7.input 2)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or5.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or6.output)+(0.07,0)$) circle (0.07); \draw[dashed] (1.75,-0.25) -- (1.75,2.5); \node[above] at (1,2.25) {$s_0$}; \draw[dashed] (3.25,-0.25) -- (3.25,2.5); \node[above] at (2.5,2.25) {$s_1$}; \draw[dashed] (4.75,-0.25) -- (4.75,2.5); \node[above] at (4,2.25) {$s_2$}; \node[above] at (5.5,2.25) {$s_3$}; \end{tikzpicture}} \caption{Circuit} \label{fig:circuit_retractile_gates} \end{subfigure} \begin{subfigure}{0.39\linewidth} \centering \scalebox{0.6}{ \begin{tikzpicture} \draw[->](0,0) -- (5.0, 0) node[right] {$t$}; \draw[-](0, 0) -- (0, 2.5) node[above] {}; \draw[domain=0:.5, smooth, variable=\x, red] plot ({\x}, {(\x)}); \draw[domain=.5:4, smooth, variable=\x, red] plot ({\x}, {.5}); \draw[domain=4:4.5, smooth, variable=\x, red] plot ({\x}, {-\x+4.5}); \draw[domain=0:0.5, smooth, variable=\x, red] plot ({\x}, {0.6}); \draw[domain=0.5:1, smooth, variable=\x, red] plot ({\x}, {0.1+(\x)}); \draw[domain=1:3.5, smooth, variable=\x, red] plot ({\x}, {1.1}); \draw[domain=3.5:4, smooth, variable=\x, red] plot ({\x}, {-\x+4.6}); \draw[domain=4:4.5, smooth, variable=\x, red] plot ({\x}, {0.6}); \draw[domain=0:1, smooth, variable=\x, red] plot ({\x}, {1.2}); \draw[domain=1:1.5, smooth, variable=\x, red] plot ({\x}, {0.2+(\x)}); \draw[domain=1.5:3, smooth, variable=\x, red] plot ({\x}, {1.7}); \draw[domain=3:3.5, smooth, variable=\x, red] plot ({\x}, {-\x+4.7}); \draw[domain=3.5:4.5, smooth, variable=\x, red] plot ({\x}, {1.2}); \draw[domain=0:1.5, smooth, variable=\x, red] plot ({\x}, {1.8}); \draw[domain=1.5:2, smooth, variable=\x, red] plot ({\x}, {0.3+(\x)}); \draw[domain=2:2.5, smooth, variable=\x, red] plot ({\x}, {2.3}); \draw[domain=2.5:3, smooth, variable=\x, red] plot ({\x}, {-\x+4.8}); \draw[domain=3:4.5, smooth, variable=\x, red] plot ({\x}, {1.8}); \node[anchor = center] at (-0.25,0.25) {$\phi_0$}; \node[anchor = center] at (-0.25,0.85) {$\phi_1$}; \node[anchor = center] at (-0.25,1.45) {$\phi_2$}; \node[anchor = center] at (-0.25,2.05) {$\phi_3$}; \draw (0,-0.1) node[below] {0}; \foreach \x in {1,...,9} { \draw (0.5*\x,0.1) -- (0.5*\x,-0.1) node[below]{\x}; } \end{tikzpicture} } \caption{Power clocks} \label{fig:clocks} \end{subfigure} \vspace*{-2mm} \caption{Synthesized retractile circuit} \label{fig:circuit_retractile} \vspace*{-2mm} \end{figure} \subsection{Advanced Solution} \label{sec:retractile_opt} The straightforward mapping described above can significantly be optimized to reduce the number of required transmission gates and power clocks. The optimized mapping scheme is motivated by an analysis of the realization of an OR gate, which is composed of two parallel buffers (i.e.~a transmission gate), whose outputs are connected (cf.~Fig.~\ref{fig:gates_or}). Consequently, an OR gate with multiple inputs can be generated by adding further buffers in parallel. This way, each OIG node, whose children both have a fanout of 1 (and, thus, represents a 4-inputs OR gate) can be realized in a single stage of the circuit composed of two 2-input OR gates whose outputs are connected. A similar optimization can be performed for OIG nodes, where only one of the children has a fanout of 1. Here, one buffer is required for the child which has a fanout larger than one (in order to avoid sneak-paths). Additionally, the gate representing the child with fanout 1 has to be lifted to the next stage of the circuit since both, the buffer as well as the gate, have to be operated by the same power clock to allow for an adiabatic computation. The optimization rules are shown in Fig.~\ref{fig:retractile_opt_rules} and denoted \emph{Rule 1} and \emph{Rule 2} in the following. \begin{figure} \centering \scalebox{0.7}{ \begin{tikzpicture} \node[or gate US, draw] (or1) {}; \node[or gate US, draw, below = 0.25cm of or1] (or2) {}; \coordinate (pos7) at ($(or1.output)!.5!(or2.output)$); \node[or gate US, draw] at (1.15,0 |- pos7) (or3) {}; \node[above=-0.1cm of or1,xshift=0.4cm] {fanout $=$ 1}; \node[below=-0.1cm of or2,xshift=0.4cm] {fanout $=$ 1}; \draw (or1.output) -- ++ (0.25,0) |- ($(or3.input 1)$); \draw (or2.output) -- ++ (0.25,0) |- ($(or3.input 2)$); \draw (or1.input 1) -- ++ (-0.25,0); \draw (or1.input 2) -- ++ (-0.25,0); \draw (or2.input 1) -- ++ (-0.25,0); \draw (or2.input 2) -- ++ (-0.25,0); \draw (or3.output) -- ++ (0.25,0); \draw[dashed] ($(or1.input 1)+(-0.15,0.5)$) |- ($(or2.input 2)+(1.95,-0.5)$) |- ($(or1.input 1)+(-0.15,0.5)$); \node[or gate US, draw, right=2.9cm of or1] (or4) {}; \node[or gate US, draw, below = 0.25cm of or4] (or5) {}; \coordinate (pos8) at ($(or4.output)!.5!(or5.output)$); \draw (or4.output) -- ++ (0.25,0) |- ($(pos8)+(0.55,0)$); \draw (or5.output) -- ++ (0.25,0) |- ($(pos8)+(0.55,0)$) node[midway, branch] {}; \draw (or4.input 1) -- ++ (-0.25,0); \draw (or4.input 2) -- ++ (-0.25,0); \draw (or5.input 1) -- ++ (-0.25,0); \draw (or5.input 2) -- ++ (-0.25,0); \draw[dashed] ($(or4.input 1)+(-0.15,0.5)$) |- ($(or5.input 2)+(1.1,-0.5)$) |- ($(or4.input 1)+(-0.15,0.5)$); \path[draw=black,solid,line width=2mm,fill=black, preaction={-triangle 90,thin,draw,shorten >=-1mm} ]($(or3.output)+(0.4,0)$)-- ++ (0.8,0) node[above, midway]{Rule 1}; \node[or gate US, draw, right = 5.5 cm of or1] (or6) {}; \node[or gate US, draw, below = 0.25cm of or6] (or7) {}; \coordinate (pos1) at ($(or6.output)!.5!(or7.output)$); \node[or gate US, draw] at (7.3,0 |- pos1) (or8) {}; \node[above=-0.1cm of or6,xshift=0.4cm] {fanout $\neq$ 1}; \node[below=-0.1cm of or7,xshift=0.4cm] {fanout $=$ 1}; \draw (or6.output) -- ++ (0.25,0) |- ($(or8.input 1)$); \draw (or7.output) -- ++ (0.25,0) |- ($(or8.input 2)$); \draw (or6.input 1) -- ++ (-0.25,0); \draw (or6.input 2) -- ++ (-0.25,0); \draw (or7.input 1) -- ++ (-0.25,0); \draw (or7.input 2) -- ++ (-0.25,0); \draw (or8.output) -- ++ (0.25,0); \draw ($(or6.output)+(0.25,0)$) node[branch] {} |- ($(or6.output)+(0.7,0.75)$); \draw[dashed] ($(or6.input 1)+(-0.15,0.5)$) |- ($(or7.input 2)+(1.95,-0.5)$) |- ($(or6.input 1)+(-0.15,0.5)$); \node[or gate US, draw, right = 2.9cm of or6] (or9) {}; \node[or gate US, draw=none, below = 0.25cm of or9] (or10) {}; \coordinate (pos11) at ($(or9.output)!.5!(or10.output)$); \node[or gate US, draw] at ($(or10)+(0.9,0)$) (or11) {}; \node[buffer gate US, draw, anchor=output] at (or11.output |- or9.output) (buf) {}; \draw (or9.output) -- (buf.input); \draw ($(or10.input 1)+(-0.25,0)$) -- (or11.input 1); \draw ($(or10.input 2)+(-0.25,0)$) -- (or11.input 2); \draw (or9.input 1) -- ++ (-0.25,0); \draw (or9.input 2) -- ++ (-0.25,0); \coordinate (pos2) at ($(or11.output)!.5!(buf.output)$); \draw (buf.output) -- ++ (0.2,0) |- ($(pos2)+(0.55,0)$) node[midway, branch] {}; \draw (or11.output) -- ++ (0.2,0) |- ($(pos2)+(0.55,0)$); \draw ($(or9.output)+(0.2,0)$) node[branch] {} |- ($(or9.output)+(0.7,0.75)$); \draw[dashed] ($(or9.input 1)+(-0.15,0.5)$) |- ($(or10.input 2)+(1.95,-0.5)$) |- ($(or9.input 1)+(-0.15,0.5)$); \path[draw=black,solid,line width=2mm,fill=black, preaction={-triangle 90,thin,draw,shorten >=-1mm} ]($(or8.output)+(0.4,0)$)-- ++ (0.8,0) node[above, midway]{Rule 2}; \end{tikzpicture}} \caption{Rules for optimization} \label{fig:retractile_opt_rules} \end{figure} Note that one has to be careful when applying the rules if the corresponding input of the gate is inverted. In this case, the inversion has to be pushed towards the inputs. This is possible by applying De Morgan's law ($\overline{a+b} = \overline{a}\cdot\overline{b}$). Consequently, we have to invert the inputs on this level and exchange the OR gate with an AND gate.\footnote{Note that this is also possible if there are several subsequent nodes for which the rules can be applied.} \begin{myexample} Consider again the circuit shown in Fig.~\ref{fig:circuit_retractile}. The children of gate $g_7$ (i.e.~$g_4$ and $g_5$) both have a fanout of 1. Consequently, we can apply \emph{Rule 1} to remove $g_7$. Furthermore, one child of gate $g_3$ has a fanout of 1 (i.e.~the input $x_0$). Consequently, we can apply \emph{Rule 2} for gate $g_3$. The resulting (optimized) circuit is shown in Fig.~\ref{fig:circuit_retractile_opt}. Since both inputs of $g_7$ and one input of $g_3$ are inverted, we have to apply De Morgan's law. Consequently, the gates $g_1$, $g_4$, and $g_5$ are transformed into an AND-gate. The resulting circuit only requires 11 transmission gates and has only two stages (and, thus, suddenly requires only two different dual-rail encoded clocks). \end{myexample} \begin{figure}[t] \begin{minipage}[c]{0.5\linewidth} \centering \scalebox{0.7}{\begin{tikzpicture} \node (x2) at (0, 2) {$x_2$}; \node (x1) at (0, 1) {$x_1$}; \node (x0) at (0, 0) {$x_0$}; \node[or gate US, draw, anchor=input 1] at ($(x2) + (1, 0)$) (or2) {$g_2$}; \node[and gate US, draw, anchor=input 2] at ($(x1) + (1, 0)$) (or1) {$g_1$}; \node[buffer gate US, draw, anchor=output] at (or1.output |- 0,0.5) (buf) {$g_3$}; \node[and gate US, draw, anchor=input 2] at (2.5,0 |- or1.output) (or4) {$g_4$}; \node[and gate US, draw, anchor=input 2] at ($(x0) + (2.5, 0)$) (or5) {$g_5$}; \node[or gate US, draw, anchor=input 1] at (2.5,0 |- or2.output) (or6) {$g_6$}; \coordinate (pos7) at ($(or4.output)!.5!(or5.output)$); \draw (x2) -- (or2.input 1); \draw (x1) -- (or1.input 2); \draw (x0) -- (or5.input 2); \draw (or1.output) -- (or4.input 2); \draw (or2.output) -- (or6.input 1); \draw ($(x2)+(0.75,0)$) node[branch] {} |- ($(or1.input 1)$); \draw ($(x1)+(0.5,0)$) node[branch] {} |- ($(or2.input 2)$); \draw ($(x0)+(0.75,0)$) node[branch] {} |- ($(buf.input)$); \draw ($(buf.output)$) -| ($(or1.output)+(0.25,0)$) node[branch] {}; \draw ($(or2.output)+(0.5,0)$) node[branch] {} |- ($(or4.input 1)$); \draw ($(or1.output)+(0.25,0)$) |- ($(or6.input 2)$); \draw ($(x1)+(0.5,0)$) |- ($(or5.input 1)$); \draw (or5.output) -| ($(or4.output)+(0.25,0)$) node[branch] {}; \draw (or4.output) -- ++ (0.5,0) node[anchor=west] {$y_0$}; \draw (or6.output) -- ++ (0.5,0) node[anchor=west] {$y_1$}; \draw[line width=0.4, fill=white] ($(or1.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or1.input 2)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or2.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or2.input 2)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or4.input 2)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or6.input 1)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or6.input 2)-(0.07,0)$) circle (0.07); \draw[line width=0.4, fill=white] ($(or6.output)+(0.07,0)$) circle (0.07); \end{tikzpicture}} \end{minipage} \begin{minipage}[c]{0.49\linewidth} \centering \scalebox{0.75}{ \begin{tikzpicture} \draw[->](0,0) -- (3.0, 0) node[right] {$t$}; \draw[-](0, 0) -- (0, 1.5) node[above] {}; \draw[domain=0:.5, smooth, variable=\x, red] plot ({\x}, {(\x)}); \draw[domain=.5:2, smooth, variable=\x, red] plot ({\x}, {.5}); \draw[domain=2:2.5, smooth, variable=\x, red] plot ({\x}, {-\x+2.5}); \draw[domain=0:0.5, smooth, variable=\x, red] plot ({\x}, {0.6}); \draw[domain=0.5:1, smooth, variable=\x, red] plot ({\x}, {0.1+(\x)}); \draw[domain=1:1.5, smooth, variable=\x, red] plot ({\x}, {1.1}); \draw[domain=1.5:2, smooth, variable=\x, red] plot ({\x}, {-\x+2.6}); \draw[domain=2:2.5, smooth, variable=\x, red] plot ({\x}, {0.6}); \node[anchor = center] at (-0.25,0.25) {$\phi_0$}; \node[anchor = center] at (-0.25,0.85) {$\phi_1$}; \draw (0,-0.1) node[below] {0}; \foreach \x in {1,...,5} { \draw (0.5*\x,0.1) -- (0.5*\x,-0.1) node[below]{\x}; } \end{tikzpicture}} \end{minipage} \vspace*{-2mm} \caption{Optimized retractile circuit} \label{fig:circuit_retractile_opt} \vspace*{-2mm} \end{figure} \section{Fully-Pipelined Circuits} \label{sec:pipelined} The main disadvantages of the retractile circuits considered in Section~\ref{sec:retractile} are that many different power clocks are required (one for each stage) and that a computation can be conducted only every $2N+1$ time steps---resulting in a rather low throughput. These issues can be avoided by using fully-pipelined circuits. In conventional design, this would require a register after each stage of the circuit. For the adiabatic circuits considered here, however, this is not necessary, because the gates inherently allow for latching their output (cf.~Section~\ref{sec:background}). In fact, we only have to compute the outputs of a stage $s_i$ while decomputing the signals of stage $s_{i-1}$ (i.e.~resetting them back to 0). This way, only two different power clocks (four if we take the dual-rail encoding into account) are required (independent from the circuit depth) and computations can be conducted in a pipelined fashion (leading to a much higher throughput). To realize this, however, the functions computed in the individual stages have to be (conditionally) reversible. This can easily be achieved by forwarding all the input signals of stage $s_{i-1}$ to the stage $s_i$ by using buffers. The following example illustrates the idea of such buffers. \begin{myexample} Fig.~\ref{fig:buffer} shows the structure of a buffer that sets \mbox{$x_{t} = x_{t-1}$} while decomputing $x_{t-1}$ (i.e.~while resetting~$x_{t-1}$ back to 0). Initially, both clocks $\phi_0$ and $\phi_1$ as well as $x_{t}$ are set to~0. If $x_{t-1}=1$, the transmission gate on the right connects $\phi_1$ with $x_{t}$. In the first time step, $\phi_0$ transitions to 1 (c.f.~Fig.~\ref{fig:clocks_fully_pipelined}). Afterwards, $\phi_1$ transitions to 1, setting $x_{t} = x_{t-1}$. If $x_{t}=x_{t-1}=1$, the transmission gate on the left hand side in Fig.~\ref{fig:buffer} connects $\phi_0$ with $x_{t-1}$. This does not violate the switching rules discussed in Section~\ref{sec:background} since $\phi_0$ is also 1. In the next time step, $\phi_0$ transitions back to 0---decomputing $x_{t-1}$ and, thus, disconnecting $\phi_i$ and $x_{t}$. Consequently, the output $x_{t}$ remains at its voltage level when eventually transitioning $\phi_1$ back to 0---the output is latched. \end{myexample} \begin{figure} \centering \begin{subfigure}[b]{.43\linewidth} \centering \scalebox{0.7}{ \begin{tikzpicture} \draw (0.25,-1.25) node[below] {$\phi_0$} |- (0.75, 0.5); \draw (0.25,0.5) -- (-0.25,0.5) node[left] {$x_{t-1}$}; \draw (1.,1.25) node[above] {$\phi_1$} |- (0.5, -0.5); \draw (1,-0.5) -- (1.5,-0.5) node[right] {$x_{t}$}; \draw[line width=0.75pt, fill=white] (0,0) rectangle ++(0.5,-1); \draw[line width=0.75pt, fill=white] (0.75,0) rectangle ++(0.5,1); \end{tikzpicture}} \captionof{figure}{Transmission gates} \label{fig:buffer} \end{subfigure} \begin{subfigure}[b]{0.54\linewidth} \centering \scalebox{0.75}{ \begin{tikzpicture} \draw[->](0,0) -- (2.5, 0) node[right] {$t$}; \draw[-](0, 0) -- (0, 1.5) node[above] {}; \draw (0,-0.1) node[below] {0}; \foreach \x in {1,...,4} { \draw (0.5*\x,0.1) -- (0.5*\x,-0.1) node[below]{\x}; } \draw[domain=0:0.5, smooth, variable=\x, red] plot ({\x}, {0.6}); \draw[domain=0.5:1, smooth, variable=\x, red] plot ({\x}, {.1+\x)}); \draw[domain=1:1.5, smooth, variable=\x, red] plot ({\x}, {1.1}); \draw[domain=1.5:2, smooth, variable=\x, red] plot ({\x}, {-\x+2.6}); \draw[domain=0:0.5, smooth, variable=\x, red] plot ({\x}, {\x}); \draw[domain=0.5:1, smooth, variable=\x, red] plot ({\x}, {.5}); \draw[domain=1:1.5, smooth, variable=\x, red] plot ({\x}, {1.5-\x}); \draw[domain=1.5:2, smooth, variable=\x, red] plot ({\x}, {0}); \node[anchor = east] at (-0.0,0.85) {$\phi_1$}; \node[anchor = east] at (-0.0,0.25) {$\phi_0$}; \end{tikzpicture}} \captionof{figure}{Clocks} \label{fig:clocks_fully_pipelined} \end{subfigure} \vspace*{-3mm} \caption{Buffer element for fully-pipelined circuits} \vspace*{-3mm} \end{figure} To allow for inverted inputs of gates, a quad-rail encoding is required for the signals to properly decompute the inputs~\cite{anantharam2004driving}. Here, each signal $X$ is represented by two dual-rail signals (one for $X=1$ and one for $X=0$). Initially, both dual-rail signals are set to 0. This again allows to realize inverters without any transmission gates---just swapping the two dual-rail signals $X=1$ and $X=0$. In the following we again abstract this fact when illustrating the required transmission gates. \subsection{Straightforward Solution} \label{sec:pipelined_sf} As for retractile circuits, we again map the OIG nodes to an adiabatic realizations of an OR gate. As mentioned above, this requires to realize each OR gate as shown in Fig.~\ref{fig:or_pipelined}.\footnote{Signals with fanout do not have to be buffered multiple times.} This way, the signals from stage $s_{t-1}$ (e.g.~$A_{t-1}$ and $B_{t-1}$) serve as input to compute $(A+B)_t = A_{t-1} + B_{t-1}$. Since $(A+B)_{t}$ is driven by clock $\phi_1$, its value is inherently latched. In fact, the input signals $A_{t-1}$ and $B_{t-1}$ are reset to 0 by the according buffers (disconnecting $\phi_1$ and $(A+B)_t$), before the clock $\phi_1$ is transitioned back to 0. \begin{figure} \centering \begin{subfigure}[b]{0.49\linewidth} \centering \begin{tikzpicture} \draw[line width=0.75pt] (0.1,0) rectangle ++(0.25,0.5); \draw[line width=0.75pt] (0.85,0) rectangle ++(0.25,0.5); \draw[line width=0.75pt] (1.35,0.5) rectangle ++(0.25,0.5); \draw[line width=0.75pt] (-0.1,0) rectangle ++(-0.25,0.5); \draw[line width=0.75pt] (-0.85,0) rectangle ++(-0.25,0.5); \draw[line width=0.75pt] (-1.35,0.5) rectangle ++(-0.25,0.5); \draw (0.35,0.25) -| (0.45,1.25) node[above] {$B_{t-1}$}; \draw (0.45,0.75) -- (1.35,0.75); \draw (0.975,0.5) -- (0.975,0.75); \draw (1.1,0.25) -| (1.475,0.5); \draw (0.975,0) -- (0.975,-0.25) node[below] {$\phi_0$}; \draw (1.475,0.25) -| (1.475,-0.5) node[below] {$B_t$}; \draw (1.475,1) -| (1.475,1.25) node[above] {$\phi_1$}; \draw (0.225,0) |- (-0.225,-0.1) -- ++(0,0.1); \draw (0,-0.1) -- (0,-0.5)node[below]{$(A+B)_{t}$}; \draw (0.225,0.5) |- (-0.225,0.6) -- ++(0,-0.1); \draw (0,0.6) -- (0,.85)node[above]{$\phi_1$}; \draw (-0.35,0.25) -| (-0.45,1.25) node[above] {$A_{t-1}$}; \draw (-0.45,0.75) -- (-1.35,0.75); \draw (-0.975,0.5) -- (-0.975,0.75); \draw (-1.1,0.25) -| (-1.475,0.5); \draw (-0.975,0) -- (-0.975,-0.25) node[below] {$\phi_0$}; \draw (-1.475,0.25) -| (-1.475,-0.5) node[below] {$A_t$}; \draw (-1.475,1) -| (-1.475,1.25) node[above] {$\phi_1$}; \end{tikzpicture} \caption{Computing OR} \label{fig:or_pipelined} \end{subfigure} \begin{subfigure}[b]{0.49\linewidth} \centering \begin{tikzpicture} \draw[line width=0.75pt] (0.1,0) rectangle ++(0.25,-0.5); \draw[line width=0.75pt] (0.85,0) rectangle ++(0.25,-0.5); \draw[line width=0.75pt] (1.35,-0.5) rectangle ++(0.25,-0.5); \draw[line width=0.75pt] (-0.1,0) rectangle ++(-0.25,-0.5); \draw[line width=0.75pt] (-0.85,0) rectangle ++(-0.25,-0.5); \draw[line width=0.75pt] (-1.35,-0.5) rectangle ++(-0.25,-0.5); \draw (0.35,-0.25) -| (0.45,-1.25) node[below] {$B_{t}$}; \draw (0.45,-0.75) -- (1.35,-0.75); \draw (0.975,-0.5) -- (0.975,-0.75); \draw (1.1,-0.25) -| (1.475,-0.5); \draw (0.975,0) -- (0.975,0.25) node[above] {$\phi_1$}; \draw (1.475,-0.25) -| (1.475,0.5) node[above] {$B_{t-1}$}; \draw (1.475,-1) -| (1.475,-1.25) node[below] {$\phi_0$}; \draw (0.225,0) |- (-0.225,0.1) -- ++(0,-0.1); \draw (0,0.1) -- (0,0.5)node[above]{$(A+B)_{t-1}$}; \draw (0.225,-0.5) |- (-0.225,-0.6) -- ++(0,0.1); \draw (0,-0.6) -- (0,-.85)node[below]{$\phi_0$}; \draw (-0.35,-0.25) -| (-0.45,-1.25) node[below] {$A_{t}$}; \draw (-0.45,-0.75) -- (-1.35,-0.75); \draw (-0.975,-0.5) -- (-0.975,-0.75); \draw (-1.1,-0.25) -| (-1.475,-0.5); \draw (-0.975,0) -- (-0.975,0.25) node[above] {$\phi_1$}; \draw (-1.475,-0.25) -| (-1.475,0.5) node[above] {$A_{t-1}$}; \draw (-1.475,-1) -| (-1.475,-1.25) node[below] {$\phi_0$}; \end{tikzpicture} \caption{Decomputing OR} \label{fig:or_pipelined_decompute} \end{subfigure} \vspace*{-3mm} \caption{OR gate for fully-pipelined circuits} \vspace*{-3mm} \end{figure} Now, in contrast to retractile circuits, new hardware is required to decompute the result (after e.g.~copying it elsewhere) since the stages of the pipeline already contain the values of the next computation. The (conditionally) reversible function calculated by the pipeline is $F = f_{N-1}\circ f_{N-2} \circ \cdots \circ f_0$, where $f_i$ is the conditionally reversible function computed by stage $s_i$.\footnote{Note that $\circ$ denotes functional composition, i.e.~$g(x)\circ f(x) = g(f(x))$.} Since the function $f_i$ computed by each stage is conditionally reversible, the inverse of $F$ (i.e.~$F^{-1}$) exists and is determined by $F^{-1} = f_0^{-1} \circ f_1^{-1} \circ \cdots \circ f_{N-1}^{-1}$. The inverse $f_i^{-1}$ of the function $f_i$ computed by stage $s_i$ can be easily realized by duplicating the hardware for stage $s_i$ and connecting the power clocks $\phi_0$ and $\phi_1$ in opposite fashion (as shown for an OR gate in~Fig.~\ref{fig:or_pipelined_decompute}). Consequently, decomputing the results requires to double the depth of the pipeline and, thus, doubles the number of required transmission gates. \begin{myexample} Consider again the circuit shown in Fig.~\ref{fig:circuit_retractile}. The first stage contains a single OR gate. Additionally, three buffers are required to forward the inputs $x_2$, $x_1$, and $x_0$ to stage $s_1$ (while decomputing them in stage $s_0$). Consequently, $(1+3)\cdot 2=8$ transmission gates are required. The second stage has four input signals and requires two OR gates. Therefore, $(4+2)\cdot 2 = 12$ transmission gates are required to realize stage $s_1$. The third stage has then 6 inputs and requires $16$ transmission gates. Finally, the last stage has 8 inputs and requires $20$ transmission gates. Overall, this sums up to $56$ transmission gates. The reverse cascade of the stages again requires 56 transmission gates. Consequently, a total of 112 transmission gates are required (448 if we take the quad-rail encoding into account) to realize the function in a fully-pipelined fashion---a huge overhead compared to the retractile design methodology. However, the circuit has a higher throughput and only requires two different clocks to be operated (four if we take into account that their complement is also needed due to a dual-rail encoding). \end{myexample} \subsection{Advanced Solution} \label{sec:pipelined_opt} The mapping scheme discussed above yields circuits with a huge overhead since many signals are pushed through the whole pipeline---even though they are not required as outputs, nor to obtain reversibility of a stage. Hence, we propose to decompute such unnecessary signals as soon as possible. As shown in Fig.~\ref{fig:or_pipelined_decompute}, the inputs of a gate have to be present until its output is decomputed. This means, the signals resulting from the gates in the next-to-last stage can be decomputed while computing the outputs of the function to be realized. Afterwards, the signals generated in the stage before can be decomputed---eventually resulting in the mapping scheme discussed in the previous subsection---hence, no signal can be decomputed before the final outputs of the function to be realized are determined. However, we can easily circumvent this problem by choosing some signals that shall not be decomputed.\footnote{Note that, in the end, all signals are decomputed since each stage is duplicated as discussed in Section~\ref{sec:pipelined_sf}.} To this end, we mark the corresponding OIG nodes that generate these signals. This allows to decompute several other signals earlier---while continuing to compute the outputs of the function. Consequently, fewer signals are pushed through the pipeline---reducing the number of required transmission gates. Recall, that each node $v$ of the OIG is translated to an OR gate on a certain stage of the circuit. To determine when the signal resulting from $v$ can be decomputed we traverse all parents (denoted $p_j$ in the following). For each parent node $p_j$ we determine the stage in which the signal generated by $v$ can be decomputed at the earliest. Then, we take the stage with the largest index, since the constraints for all parents have to be satisfied. If $p_j$ is a node that is marked, we can immediately decompute the signal generated by $v$ in the same stage (since the signal computed by $p_j$ is not decomputed afterwards). If $p_j$ is not marked, we can decompute the signal generated by $v$ at the earliest one stage after the signal generated by $p_j$ can be decomputed (because the signal generated by $v$ is required to decompute the signal generated by $p_j$). \begin{myexample} \label{ex:pipelined_opt} Consider again the OIG shown in Fig.~\ref{fig:oig} (as well as the corresponding circuit shown in Fig.~\ref{fig:circuit_retractile}). Assume that we marked the nodes labeled $\vee_2$ and $\vee_3$ (the nodes labeled $\vee_6$ and $\vee_7$ are inherently marked since they are directly connected to an output). Consequently, we want to decompute the signals generated by the OIG nodes labeled $\vee_1$, $\vee_4$, and $\vee_5$ as soon as possible. In the second stage (i.e.~$s_1$) of the circuit, we compute the result of the nodes labeled $\vee_2$ and $\vee_3$. Since the signal generated by node $\vee_1$ is not required anymore (its single parent labeled $\vee_3$ is marked), it can be decomputed in the second stage as well. Consequently, we can save the buffers for this signal in the third and fourth stage of the circuit. Furthermore, the signals generated by nodes labeled $\vee_4$ and $\vee_5$ can be decomputed while computing the outputs of the function (in stage $s_3$). Since this is the last stage of the circuit, no buffers can be saved. However, fewer output signals result. Considering the fact that each pipeline stage has to be duplicated, a reduction of four buffers (i.e.~8 transmission gates) can be obtained. \end{myexample} This leads to the question how to determine a suitable marking scheme for the nodes, i.e.~a marking scheme that results in a circuit with a smaller number of transmission gates. A very simple but also effective marking scheme is to mark all nodes of the OIG with a depth that is a multiple of a constant $k\in \mathbb{N}$. For $k=2$, this means to mark all nodes with an even depth (as done in Example~\ref{ex:pipelined_opt}). The experimental evaluations summarized in Section~\ref{sec:exp} show that significant improvements can be obtained by using this marking scheme. \section{Evaluation} \label{sec:exp} In this section, we summarize and discuss the results obtained by our evaluations of the proposed design methods for adiabatic circuits. To this end, we implemented the approaches discussed in Section~\ref{sec:retractile} and Section~\ref{sec:pipelined} in C++ and used the tool ABC~\cite{DBLP:conf/cav/BraytonM10} to generate the initially required AIGs/OIGs (to reduce the number of AIG nodes, we used the synthesis command \emph{dc2}). Afterwards, we evaluated the resulting methods using benchmarks taken from the ISCAS~\cite{ISCAS:89} and the IWLS benchmark suite~\cite{McE:93}. Table~\ref{tab:results} summarizes the obtained results. The first columns show the name of the benchmark as well as the number of primary inputs~\emph{PI} and primary outputs \emph{PO}. Then, we list the results obtained for retractile and fully-pipelined adiabatic circuits. For each design style, we list the number of required transmission gates (denoted~$\left|tg\right|$) and the number of required power clocks (denoted~$\left|\phi\right|$) of the straightforward solution as well as the advanced solution (columns denoted \emph{Straight-forward} and \emph{Advanced}, respectively). Having a dual-rail (for retractile circuits) or quad-rail encoding (for fully-pipelined circuits) is taken into account in the numbers listed for the required transmission gates, as well as the fact that each power clock has to be supplied in two polarities (i.e.~a power clock is dual-rail encoded for both types of circuits). For sake of completeness, we also list the parameter $k$ used in the solution discussed in Section~\ref{sec:pipelined_opt}. The runtime is not listed in Table~\ref{tab:results} since all methods are capable to produce these results in negligible runtime (i.e.~a fraction of a second). \begin{table}[t] \caption{Evaluation} \label{tab:results} \vspace*{-2mm} \centering \scriptsize \setlength{\tabcolsep}{3pt} \renewcommand{\arraystretch}{1.1} \begin{tabular}{lrr||rr|rr||rr|rrr} \multicolumn{3}{c||}{} & \multicolumn{4}{c||}{Retractile (Section~\ref{sec:retractile})} & \multicolumn{5}{c}{Fully-pipelined (Section~\ref{sec:pipelined})} \\ \multicolumn{3}{c||}{} & \multicolumn{2}{c|}{Straight-forward} & \multicolumn{2}{c||}{Advanced} & \multicolumn{2}{c|}{Straight-forward} & \multicolumn{3}{c}{Advanced} \\ \multicolumn{3}{c||}{} & \multicolumn{2}{c|}{(Section~\ref{sec:retractile_sf})} & \multicolumn{2}{c||}{(Section~\ref{sec:retractile_opt})} & \multicolumn{2}{c|}{(Section~\ref{sec:pipelined_sf})} & \multicolumn{3}{c}{(Section~\ref{sec:pipelined_opt})} \\ Name & $PI$ & $PO$ & $\left|\phi\right|$ & $\left|tg\right|$ & $\left|\phi\right|$ & $\left|tg\right|$ & $\left|\phi\right|$ & $\left|tg\right|$ & $k$ & $\left|\phi\right|$ & $\left|tg\right|$ \\ \hline \csvreader[ late after line=\\, late after last line=\\, ]{results.csv} {1=\Name,2=\In, 3=\Out, 4=\Dsf, 5=\Dopt, 6=\retractileSF, 7=\retractileOPT, 8=\pipelinedSF, 9=\kopt, 10=\pipelinedOPT} {\Name & \In & \Out & \Dsf & \optnum{\retractileSF} & \Dopt & \optnum{\retractileOPT} & 4 & \optnum{\pipelinedSF} & \kopt & 4 & \optnum{\pipelinedOPT}} \end{tabular}\\ \raggedright{$\left|\phi\right|$: \#required clocks \hspace*{0.4cm} $\left|tg\right|$: \#transmission gates \hspace*{0.4cm} $k$: parameter discussed in Section~\ref{sec:pipelined_opt}} \vspace*{-5mm} \end{table} First, the results nicely show the impact of the respectively chosen design style. Retractile circuits are clearly the better choice when it comes to reducing the number of gates, while pipelined circuits are efficient with respect to the number of power clocks and, following that, also the throughput. At a first glance, it might look that the costs of having fewer power clocks in pipelined circuits is not acceptable (in fact, magnitudes more gates are required). However, if area is not an issue, this might still acceptable since, as discussed in Section~\ref{sec:background}, gates in adiabatic circuits do not affect the energy consumption as much as they do in conventional circuits. Hence, each design style has its own advantages and disadvantages and, eventually, the user is presented with complementary solutions out of which the best suitable can be chosen. Besides that, the results clearly show the improvement of the advanced schemes. On average an improvement of approx.~42\% in the number of required power clocks, as well as an average improvement of approx.~37\% with respect to the number of required transmission gates is obtained for retractile circuits. For the fully-pipelined circuits, we observe a reduction in the number of transmission gates of approx.~30\% on average. Overall, these results clearly confirm the benefit and applicability of the proposed design automation techniques for this kind of circuits. While previously considered circuits were either handcrafted (following approaches e.g.~proposed in~\cite{younis1993practical,anantharam2004driving}) or relied on fully reversible realizations which led to an unnecessarily large overhead (as conducted in~\cite{DBLP:journals/tcad/MorrisonR14,rauchenecker2017exploiting} and discussed in~\cite{DBLP:conf/rc/Frank17}), the proposed design flow allows for generating the desired adiabatic circuits in an automatic fashion while, at the same time, satisfying the switching rules by conditional reversibility only. The improvements obtained by the advanced schemes additionally show the further potential that can be exploited following this direction. \section{Conclusions} \label{sec:conclusions} In this work, we proposed an automatic and dedicated design flow for adiabatic circuits which explicitly takes recent findings in this domain (namely that conditional reversibility is sufficient for adiabatic circuits) into account. The proposed flow first realizes the desired functionality in terms of an AIG/OIG and, afterwards, dedicatedly maps the resulting structure to an adiabatic description. For the latter step, two complementary schemes (namely retractile or fully-pipelined) are considered which allow the designer to either focus on reducing the number of gates or keeping the number of power clocks small. Furthermore, optimizations are proposed which allow for a reduction in the number of gates by approx.~37\% and 30\%, respectively, for both design styles on average. By this, expertise from both, adiabatic circuits and design automation, is combined yielding an automatic \emph{and} dedicated design scheme for this promising technology. This eventually provides the basis for further studies including, besides others, more sophisticated optimizations, the design and use of larger building blocks, as well as the application of the proposed design flow in the physical implementation of adiabatic circuits. \begin{acks} This work has partially been supported by the European Union through the COST Action IC1405. M. Frank was supported by the Laboratory Directed Research and Development program at Sandia National Laboratories and by the Advanced Simulation and Computing program under the U.S. Department of Energy’s National Nuclear Security Administration (NNSA). Sandia National Laboratories is a multimission laboratory managed and operated by National Technology and Engineering Solutions of Sandia, LLC., a wholly owned subsidiary of Honeywell International, Inc., for NNSA under contract DE-NA0003525. Approved for public release, SAND2018-9936 O. \end{acks} \bibliographystyle{ACM-Reference-Format} {

8573689e6f248b756e3a1e502abd3252b5fb4190

\section{Introduction} The Schr\"odinger equation for a free particle has attracted the search for wave functions that evolve without distortion. Berry and Balasz have shown that an Airy wave function keeps its form under evolution, just showing some acceleration \cite{Berry}. However, Airy wave functions are not square integrable functions and therefore are not proper wave functions. If one wants to use them, they need to be apodized, either by cutting them or by super-imposing a Gaussian function; i.e., instead considering a Gauss-Airy beam. In such case, it is too much to say that they loose their shape as they evolve, and therefore, their beauty. Effects such as focusing of waves may occur when particles go through a single slit \cite{Schleich2}, as it has been shown by studying the time dependent wave function in position space and its Wigner function \cite{Schleich1}.\\ In this contribution, we want to show that by adding a positive quadratic phase to an initial arbitrary wavefunction, its free evolution maintains an invariant structure, while it spreads by the action of an squeeze operator. That means, that the effect of passing a beam of particles (for instance electrons \cite{elec}, neutrons \cite{neut} or atoms \cite{atom}) through a negative lens, provides the wave function with the property of evolution invariance, while it diffracts by the application of a squeeze operator to the initial state \cite{Yuen,Caves,Satya,Vidiella,Knight,Schleich}.\\ In the following, we will revisit Airy beams and Airy-Gauss beams in order to show that the later ones deform as they evolve. In Section III, we show that the acquisition of a quadratic phase helps any field to become invariant under free evolution; in Section IV, we give some examples, namely initial Sinc and Bessel functions, while Section V is left for conclusions. \section{Revisiting Airy beams} Berry and Balasz \cite{Berry} have shown that an initial wave function of the form (for simplicity we set $\hbar=1$) \begin{equation}\label{Airy0} \psi(x,0)=\mathrm{Ai}(\epsilon x), \end{equation} where $\epsilon$ is an arbitrary real constant, evolves according to the Schr\"odinger equation for a free particle of mass $m=1$ \begin{equation}\label{schr0} i\frac{\partial \psi(x,t)}{\partial t}=\frac{\hat{p}^2}{2}\psi(x,t), \end{equation} as \begin{equation} \psi(x,t)=\mathrm{Ai}\left[\epsilon \left(x-\frac{\epsilon^3t^2}{4}\right)\right] \exp\left[ i \frac{\epsilon^3t}{2}\left( x-\frac{\epsilon^3t^2}{6}\right)\right], \end{equation} as can be verified by substitution into (\ref{schr0}). It is clear from this solution that the Airy wave packet is conserved, meaning that it evolves without spreading. Besides, the evolution shows an acceleration which may be obtained also in some other initial distributions of wave packets, like half Bessel functions \cite{Aleahmad}. Propagation of Airy wavelet-related patterns has also been considered in \cite{Torre0} and it has been shown they provide \textquotedblleft source functions\textquotedblright \; for freely propagating paraxial fields. The acceleration may be corrected by propagating the Airy function in a linear potential \cite{Chavez}. Unfortunately, the Airy wave packet is not a proper wave function as it is not square integrable. A possibility for making it normalizable would be to cut it (have a window) or to {\it apodize} it by multiplying it by a Gauss function, and effectively cutting it. If instead of the initial state (\ref{Airy0}), we consider as initial condition the normalizable wave function \begin{equation}\label{Airy1} \psi(x,0)=\textrm{Ai}(\epsilon x)\exp\left( -\beta x^2\right), \end{equation} with $\beta$ another arbitrary real constant, the solution then reads \begin{equation}\label{Airy1Sol} \psi(x,t)=\frac{1}{\sqrt{1-2i\beta t}}\textrm{Ai}\left[ \zeta(x,t)\right] \exp\left( \frac{\beta x^2}{2i\beta t -1}\right) \exp\left[i\gamma(x,t)\right] \end{equation} with \begin{equation} \zeta(x,t)=\frac{\epsilon^4t^2+\epsilon x(2i\beta t -1)}{(2\beta t+i)^2}, \qquad \gamma(x,t)=\frac{3\epsilon^3xt(2\beta t+i)-2i\epsilon^6t^3}{3(2\beta t+i)^3}; \end{equation} again, this can be proved by direct substitution into Eq.(\ref{schr0}). In Figure 1, we plot the probability density $|\psi(x,t)|^2$ for Eq.(\ref{Airy1Sol}) for different times. We can see that for $\beta=0.01$, the Airy-Gauss beam still accelerates, but it looses its shape. \begin{figure}[H] \centering{} \includegraphics[width=12cm]{victor1} \caption{Plot of the probability density $|\psi(x,t)|^2$ of the wavefuntion in equation (\ref{Airy1Sol}) for the parameters $\epsilon=1$ and $\beta=0.01$ at (a) $t=0$, (b) $t=1$ and (c) $t=2$.} \label{fig1} \end{figure} \section{Evolution invariant beams} Now consider an initial condition of the form \begin{equation} \psi(x,0)=\exp\left( i\alpha x^2\right) \phi(x,0), \end{equation} where $\alpha$ is a real parameter which must be set in each specific case \cite{victor18}. The solution of the Schr\"odinger equation then reads \begin{equation}\label{sol} \psi(x,t)=\exp\left( -i\frac{t}{2}\hat{p}^2\right) \exp\left( i\alpha x^2\right) \phi(x,0). \end{equation} Writing the identity operator as $\hat{I}=\left( i\frac{t}{2}\hat{p}^2\right) \exp\left( -i\frac{t}{2}\hat{p}^2\right)$, the previous equation can be cast as \begin{equation}\label{ec090} \psi(x,t)=\exp\left( -i\frac{t}{2}\hat{p}^2\right) \exp\left( i\alpha x^2\right) \exp \left( i\frac{t}{2}\hat{p}^2\right) \exp\left( -i\frac{t}{2}\hat{p}^2\right) \phi(x,0). \end{equation} As is well known, $\exp\left( -i\frac{t}{2}\hat{p}^2\right) x \exp\left( i\frac{t}{2}\hat{p}^2\right) =x-t\hat{p}$, and this implies that \begin{eqnarray} \exp\left( -i\frac{t}{2}\hat{p}^2\right) \exp\left( i\alpha x^2\right) \exp \left( i\frac{t}{2}\hat{p}^2\right)&=&\exp \left[ i \alpha \left( x-t\hat{p}\right) ^2 \right] \\ \nonumber &=&\exp\left\lbrace i\alpha [x^2-t(x\hat{p}+\hat{p}x)+t^2\hat{p}^2]\right\rbrace, \end{eqnarray} which substituted in equation (\ref{ec090}) gives us \begin{equation} \psi(x,t)=\exp\left\lbrace i\alpha [x^2-t(x\hat{p}+\hat{p}x)+t^2\hat{p}^2]\right\rbrace \exp\left( -i\frac{t}{2}\hat{p}^2\right) \phi(x,0). \end{equation} It is not difficult to show that the first exponential above may be factorized as \cite{metop} \begin{equation} \exp \left[ if_1(t)x^2\right] \exp\left[ i f_2(t)(x\hat{p}+\hat{p}x)\right] \exp\left[ i f_3(t)\hat{p}^2\right] , \end{equation} with \begin{equation} f_1(t)= \frac{\alpha}{1+2\alpha t}, \qquad f_2(t)=-\frac{1}{2} \ln(1+2\alpha t), \qquad f_3(t)= \frac{\alpha t^2}{1+2\alpha t}. \end{equation} This allows us to give a final form for equation (\ref{sol}) as \begin{equation} \psi(x,t)=\exp\left[ if_1(t)x^2\right] \exp\left[ if_2(t)(x\hat{p}+\hat{p}x)\right] \exp\left[ i f_4(t)\hat{p}^2\right] \phi(x,0) \end{equation} with $f_4(t)=f_3(t)-t/2$.\\ We now examine the behaviour of $f_4(t)$ as a function of the parameter $\alpha$. The Taylor series of $f_4(t)$ for $\alpha \approx 0$ is \begin{equation}\label{0150} f_4(t) = -\frac{t}{2}+t^2 \alpha-2t^3 \alpha^2+\textrm{O}(\alpha)^3 \end{equation} and for $\alpha \approx \infty$ is \begin{equation}\label{0160} f_4(t) = -\frac{1}{4\alpha}+\frac{1}{8t\alpha^2}+\textrm{O}\left( \frac{1}{\alpha}\right) ^3. \end{equation} In Figure 2, we plot $f_4(t)$ as a function of time for different values of the $\alpha$ parameter. It may be seen that for small values of $\alpha$ it remains close to zero, and for large values of $\alpha$ it becomes very small, as expected from the approximation in Equation (16). \begin{figure}[H] \centering{} \includegraphics[width=10cm]{figura2.jpg} \caption{Plot of the function $f_4(t)$ for $\alpha=10$ (dotted line), $\alpha=5.0$ (dashed line) and $\alpha=0.5$ (continuous line).} \label{fig2} \end{figure} Thus, for small values of $\alpha$, we take the first two terms in the Taylor development of the operator $\exp\left[ i f_4(t)\hat{p}^2\right]$ and we get \begin{equation} \label{appsol} \psi_1(x,t)\approx \exp\left[ if_1(t)x^2\right] \exp\left[ if_2(t)(x\hat{p}+\hat{p}x)\right] \left[1+if_4(t)\hat{p}^2\right]\phi(x,0). \end{equation} For $\alpha$ large enough, we completely disregard the term $\exp\left[ i f_4(t)p^2\right] $ and, to a very good approximation (as will be see below), we write simply the zeroth order solution \begin{equation}\label{appsol2} \psi_0(x,t)\approx \exp\left[ if_1(t)x^2\right] \exp\left[ i f_2(t)(x\hat{p}+\hat{p}x)\right] \phi(x,0). \end{equation} The operator $\exp\left[i f_2(t)(x\hat{p}+\hat{p}x)\right]$ is the squeeze operator, and by its application to the initial function, the equation above may be cast into \begin{equation}\label{appsol3} \psi_0(x,t)=\frac{1}{\sqrt{1+2\alpha t}} \exp\left[ if_1(t)x^2\right] \phi\left(\frac{x}{1+2\alpha t},0\right). \end{equation} It is clear that the above wave function gives a probability density that remains invariant during evolution \begin{equation}\label{invariant} |\psi_0(x,t)|^2=\frac{1}{{1+2\alpha t}}\left\vert \phi\left(\frac{x}{1+2\alpha t},0\right)\right\vert^2. \end{equation} The choice of the $\alpha$ parameter depends of the problem that is being studied and on the propagation distance that must be considered, as will be shown in the examples below. From Eqs. (\ref{0150}) and (\ref{0160}), it is also clear that different values of $\alpha$ must be considered if the zeroth order or the first order solutions are going to be used. In \cite{victor18} we present a discussion on the election of this parameter in the realm of classical optics. \section{Some examples} In this section, we study some examples where we apply our approximation and compare it with the exact solution. \subsection{Sinc function} We start with an initial (unnormalized, but normalizable) wave packet of the form \begin{equation} \psi(x,0)= \exp\left( i\alpha x^2\right) \mathrm{Sinc}(bx), \end{equation} where $b$ is an arbitrary real constant and where we define the Sinc function as \begin{equation} \mathrm{Sinc}(bx)=\frac{1}{b}\int_{-b}^b \exp\left( i u x\right) du. \label{sinc} \end{equation} We write the approximations to zeroth and first order as \begin{equation} \psi_0(x,t)=\frac{\exp\left[ if_1(t)x^2\right]} {b\sqrt{1+2\alpha t}} \int_{-b}^b \exp\left( iu\frac{x}{1+2\alpha t}\right) du, \end{equation} and \begin{equation} \psi_1(x,t)=\psi_0(x,t)+i\frac{f_4(t) \exp\left[ if_1(t)x^2\right]} {b\sqrt{1+2\alpha t}} \int_{-b}^b u^2 \exp\left( iu\frac{x}{1+2\alpha t}\right) du, \end{equation} respectively. For the sake of comparison, we can also write the exact solution as \begin{equation} \psi(x,t)=\frac{\exp\left[ if_1(t)x^2\right] }{b\sqrt{1+2\alpha t}} \int_{-b}^b \exp\left[ if_4(t)u^2\right] \exp\left( iu\frac{x}{1+2\alpha t}\right) du. \end{equation} We plot in Figure \ref{fig3} (a) and (c) the probability densities for the zeroth order and exact solutions, showing that they match very well for a value of $\alpha=0.3$ and have an excellent agreement for a greater value ($\alpha=3$). In Figure \ref{fig3} (b) and (d), the quantities $|\psi_0(x,t)|^2$ (dashed line) and $|\psi_0(x,t)-\psi_0(x,t)|^2$ (solid line) are plotted in order to show that their contributions to the first order approximation are negligible, already for such small values of $\alpha$. \begin{figure}[H] \centering{} \includegraphics[width=11cm]{victorsinc} \caption{\label{fig3} Plot of the probability densities for an initial Sinc function as given in equation (\ref{sinc}) with $b=1$ and the parameters (a) $\alpha=0.3$ and (c) $\alpha=3$ for $t=5$ for the exact (solid line) and approximate solutions (dashed line). In (b) and (d) are the comparison between the zero (dashed line) and first order (solid line) contributions.} \end{figure} \subsection{Bessel function} We consider now the initial wave function given by a Bessel function \cite{Leija,Optica} \begin{equation} \psi(x,0)=\exp \left( i\alpha x^2\right) J_n(x), \end{equation} with $J_n(x)$ a Bessel function of order $n$, defined as \cite{Arfken} \begin{equation} J_n(x)=\frac{1}{2\pi}\int_{-\pi}^{\pi} \exp\left( in\theta\right) \exp\left( -ix\sin\theta\right) d\theta. \end{equation} It is not difficult to show that the zeroth order solution is given by \begin{equation} \psi_0(x,t)=\frac{\exp\left[ if_1(t)x^2\right]} {\sqrt{1+2\alpha t}} J_n\left(\frac{x}{1+2\alpha t}\right), \end{equation} while the solution to first order reads \begin{eqnarray} & & \psi_1 \left( x,t\right)= \frac{\exp\left[ i f_1\left( t \right) x^2\right]}{\sqrt{1+2\alpha t}} \times \nonumber \\ & & \left\lbrace \left[1+\frac{f_4\left( t \right) }{2}\right] J_n\left( \frac{x}{1+2\alpha t} \right) -i \frac{f_4\left( t \right) }{4} \left[ J_{n+2}\left( \frac{x}{1+2\alpha t} \right)+ J_{n-2}\left( \frac{x}{1+2\alpha t} \right) \right] \right\rbrace. \end{eqnarray} In order to show that the approximation is good, we write also the exact solution as \begin{eqnarray} \psi(x,t)=\frac{ \exp\left[ if_1(t)x^2\right] }{2\pi\sqrt{1+2\alpha t}}\int_{-\pi}^{\pi} \exp\left( in\theta\right) \exp\left( -ix\sin\theta\right) \exp\left[ if_4(t)\sin^2\theta\right] d\theta, \end{eqnarray} which is a so-called generalized Bessel function \cite{Leija,Dattoli,Torre}. In Figure \ref{fig4}, we plot the probability densities for the exact (solid lines) and zeroth order solutions (dashed lines) which again show an excellent agreement. \begin{figure}[H] \centering{}\includegraphics[width=12cm]{VictorBessel} \caption{\label{fig4} Plot of the probability densities for a initial wave function given by a Bessel function $J_0(x)$ with parameters (a) $\alpha=10$, (b) $\alpha=5$ and (c) $\alpha=0.5$ for $t=5$ for the exact (solid lines) and first order approximate solutions (dashed lines). } \end{figure} \section{Conclusions} We have shown that by adding a quadratic phase to an initial wave packet, its structure may be kept invariant through free evolution. The main result of this contribution is equation (\ref{invariant}), which shows clearly this fact. Although the invariance is an approximation, it was shown that it perfectly matches the exact evolution. The price that has to be paid is the usual spread of the wave function due to free evolution, which is given here by the application of the squeeze operator to the initial wave function.\\

0d48485ac6b278796dfb40ebefd88a66fc4accbb

\section{Introduction} Ultra-faint dwarf galaxies (UFDs) are the least luminous galaxies known. They have only been discovered relatively recently, after the advent of deep, wide-area photometric surveys like the Sloan Digital Sky Survey, Pan-STARRS, and the Dark Energy Survey found several low surface-brightness satellites of the Milky Way \citep[e.g.,][]{Willman05a,Belokurov07,Laevens15a,Bechtol15,Koposov15a}. Though at first it was unclear if such objects were dwarf galaxies or globular clusters \citep{Willman05a}, subsequent spectroscopic followup found most of them displayed velocity dispersions implying mass-to-light ratios $>100$ and large metallicity spreads \citep[e.g.,][]{Simon07}. These properties contrast with globular clusters, which display no evidence for dark matter or large metallicity spreads \citep{Willman12}. UFDs are now understood to be the natural result of galaxy formation in small dark matter halos in standard $\Lambda$CDM cosmology. Theoretically, these galaxies begin forming at $z \sim 10$ in small $\sim 10^8~M_\odot$ dark matter halos \citep{Bromm11}. Supernova feedback is especially effective in these small galaxies \citep[e.g.,][]{BlandHaw15}, so they form stars inefficiently for $1-2$ Gyr before their star formation is quenched by reionization \citep{Bullock00,Benson02}. All observed properties of UFDs are also consistent with this picture. Color-magnitude diagrams show they contain uniformly old stellar populations \citep{Brown14,Weisz14a}. Spectroscopy shows their stars have low metallicities that extend the mass-metallicity relation all the way to $M_\star \sim 1000 M_\odot$ \citep{Kirby08,Kirby13b}. At such tiny stellar masses, the chemical abundances of individual UFDs will not even sample a full initial mass function's worth of supernovae \citep[e.g.,][]{Koch08,Simon10,Lee13}, let alone rarer nucleosynthesis events like neutron star mergers \citep{Ji16b}. Given the likely association between UFDs and small scale dark matter substructure, it is extremely important to distinguish between UFDs and globular clusters. Currently, the largest telescopes can perform spectroscopy to establish velocity and metallicity dispersions from a reasonable number of stars in the closest and/or most luminous UFDs \citep[e.g.,][]{Simon07}. However, many of the most recently discovered UFDs are very faint and/or far away. In such cases, only a handful of stars are accessible for followup spectroscopy, so it is difficult to clearly establish a velocity or metallicity dispersion for these galaxy candidates \citep[e.g.,][]{Koch09,Koposov15b,Kirby15,Kirby15b,Martin16,Martin16b}. Exacerbating this concern is the presence of unresolved binary stars, which can inflate velocity dispersions and can therefore lead to premature UFD classifications \citep{McConnachie10,Ji16a,Kirby17}. As a result, many UFD candidates still do not have clear velocity and/or metallicity dispersions \citep{Kirby15,Kirby17,Martin16,Martin16b,Walker16,Simon17}. For some UFDs, an alternative is to examine the detailed chemical abundances of the brightest stars. The first high-resolution spectroscopic abundances of stars in UFDs revealed that most elemental abundances in UFDs follow the average trends defined by metal-poor Milky Way halo stars, with the obvious exception of neutron-capture elements (e.g. Sr, Ba, Eu) that were extremely \emph{low} \citep{Koch08,Koch13,Frebel10b,Frebel14,Simon10}. This view was recently revised by the discovery that some UFDs (Reticulum~II and Tucana~III) have extremely high abundances of neutron-capture elements synthesized in the $r$-process \citep{Ji16b,Roederer16b,Hansen17}. In stark contrast, neutron-capture elements in globular clusters closely follow the abundance trends of the Milky Way halo \citep[e.g.,][]{Gratton04,Gratton12,Pritzl05}, including the globular clusters that display some internal neutron-capture abundance scatter \citep{Roederer11b}. Extreme neutron-capture element abundances have thus been suggested to be a distinguishing factor between UFDs and globular clusters \citep{Frebel15}. Here we study the detailed chemical abundances of the dwarf galaxy candidates Grus~I (Gru~I) and Triangulum~II (Tri~II). Gru~I was discovered in Dark Energy Survey data by \citet{Koposov15a}. \citet{Walker16} identified seven likely members of this galaxy, but did not resolve a metallicity or velocity dispersion. Tri~II was discovered by \citet{Laevens15a} in Pan-STARRS. As one of the closest but also least luminous galaxy candidates ($d_\odot = 28.4$ kpc, $M_V = -1.2$; \citealt{Carlin17}), Tri~II has already been the subject of numerous spectroscopic studies \citep{Kirby15,Kirby17,Martin16,Venn17}. We report the first detailed chemical abundances of two stars in Gru~I and a reanalysis of two stars in Tri~II with additional data. We describe our observations and abundance analysis in Sections~\ref{s:obs} and \ref{s:analysis}. Section~\ref{s:abunds} details the results for individual elements. We consider the classification of Gru~I and Tri~II in Section~\ref{s:discussion}, with an extended discussion of the origin and interpretation of neutron-capture elements in UFDs, larger dSph satellites, and globular clusters. We conclude in Section~\ref{s:conclusion}. \vspace{1cm} \section{Observations and Data Reduction} \label{s:obs} Our program stars were observed from two telescopes with two different echelle spectrographs. Details of the observations can be found in Table~\ref{tbl:obs}. Selected spectral regions of these four stars are shown in Figure~\ref{f:spec}. The Gru~I stars were selected as the two brightest probable members of Gru~I from \citet{Walker16}. We observed these stars with the Magellan Inamori Kyocera Echelle (MIKE) spectrograph \citep{Bernstein03} on the Magellan-Clay telescope in Aug 2017 with the 1\farcs0 slit, providing resolution $R \sim 28,000$ from ${\sim}3900-5000${\AA} on the blue arm and $R \sim 22,000$ from ${\sim}5000-9000${\AA} on the red arm. Individual exposures were 50-55 minutes long. The data were reduced with CarPy \citep{Kelson03}. Heliocentric corrections were determined with \texttt{rvcor} in IRAF\footnote{IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.}. The two stars in Tri~II were observed with the Gemini Remote Access to CFHT ESPaDOnS Spectrograph (GRACES) \citep{Donati03, Chene14}\footnote{See \url{http://www.gemini.edu/sciops/instruments/visiting/graces} for more details}. These stars were selected as the brightest probable members of Tri~II from \citet{Kirby15} and \citet{Martin16}. We combined data from two programs\footnote{GN-2015B-DD-2 (PI Venn) and GN-2016B-Q-44 (PI Ji)} that both used the 2-fiber object+sky GRACES mode providing $R \sim 40,000$ from ${\sim}5000-10,000${\AA}. The GRACES throughput for these faint stars was worse than predicted by the integration time calculator, especially at wavelengths $<6000${\AA} where the signal-to-noise ratio (S/N) was less than half that expected. The data were reduced with the OPERA pipeline for ESPaDOnS that was adapted for GRACES \citep{Martioli12}. This pipeline automatically includes a heliocentric velocity correction. \begin{deluxetable*}{lccrcrlrrrl} \tablecolumns{11} \tablewidth{0pt} \tabletypesize{\footnotesize} \tablecaption{Observing Details\label{tbl:obs}} \tablehead{ \colhead{Star} & \colhead{$\alpha$} & \colhead{$\delta$} & \colhead{$V$} & \colhead{Observation Date} & \colhead{$t_{\rm exp}$} & \colhead{$v_{\rm hel}$} & \colhead{S/N} & \colhead{S/N} &\colhead{S/N} & \colhead{Instrument}\\ \colhead{} & \colhead{(J2000)} & \colhead{(J2000)} & \colhead{(mag)} & \colhead{} & \colhead{(min)} & \colhead{(km s$^{-1}$)} & \colhead{(4500\AA)} & \colhead{(5300\AA)} & \colhead{(6500\AA)} & \colhead{} } \startdata GruI-032 & 22 56 58.1 & $-$50 13 57.9 & 18.1 & 2017 Aug 16,25 & 165 & $-139.8 \pm 0.7$ & 22 & 25 & 60 & MIKE 1\farcs0 slit\\ GruI-038 & 22 56 29.9 & $-$50 04 33.3 & 18.7 & 2017 Aug 15,16,25 & 430 & $-143.9 \pm 0.4$ & 20 & 22 & 55 & MIKE 1\farcs0 slit\\ TriII-40 & 02 13 16.5 & $+$36 10 45.9 & 17.3 & 2015 Dec 15 & 60 & $-381.5 \pm 1.3$ & 5 & 15 & 35 & GRACES 2-fiber\\ & & & & 2016 Sep 8 & 80 & $-381.5$ & & & & GRACES 2-fiber\\ TriII-46 & 02 13 21.5 & $+$36 09 57.6 & 18.8 & 2015 Dec 16,17 & 160 & $-396.5 \pm 3.2$ & 1 & 7 & 17 & GRACES 2-fiber\\ & & & & 2016 Sep 7 & 120 & $-381.5 \pm 5.0$ & & & & GRACES 2-fiber\\ \enddata \tablecomments{S/N values are per pixel. S/N values for Tri~II stars were determined after coadding. Velocity precision is computed with coadded spectra except for TriII-46, where each visit is measured separately because of the binary orbital motion.} \end{deluxetable*} We used IRAF and SMH \citep{Casey14} to coadd, normalize, stitch orders, and Doppler correct the reduced spectra. We estimated the S/N per pixel on coadded spectra by running a median absolute deviation filter across the normalized spectra in a ${\approx}$5{\AA} window. The signal-to-noise at the order center closest to rest wavelengths of 4500{\AA}, 5300{\AA}, and 6500{\AA} is given in Table~\ref{tbl:obs}. Radial velocities were determined by cross correlating the Mg b triplet against a MIKE spectrum of HD122563. \citet{Venn17} found one of the stars in Tri~II (TriII-46) to be a binary, so we Doppler shifted spectra from each visit to rest frame before coadding. The implications of this binary star were previously considered in \citet{Venn17} and \citet{Kirby17}. Our added velocity measurement does not affect their conclusions. Other than TriII-46, the velocities are consistent with constant heliocentric velocity in our data and with previous velocity measurements \citep{Kirby15,Kirby17,Martin16,Walker16,Venn17}. Velocity precision was estimated using the coadded spectra by cross-correlating all orders from $5000-6500${\AA} for MIKE and $4500-6500${\AA} for GRACES against HD122563. We excluded orders where the velocity was not within 10~km~s$^{-1}$ of the Mg b velocity, then took the standard deviation of the remaining order velocities. This value was added in quadrature to the combined statistical velocity uncertainty to obtain the velocity uncertainties listed in Table~\ref{tbl:obs}. The most discrepant velocity other than TriII-46 is for GruI-032, which is ${\approx}1$~km~s$^{-1}$ away from the measurement in \citet{Walker16} ($-138.4 \pm 0.4$~km~s$^{-1}$), but not large enough that we would consider this a clear binary candidate. Note that the two Gru~I stars differ by ${\approx}4$ km s$^{-1}$, which could be consistent with a significant velocity dispersion. \begin{figure*} \centering \includegraphics[height=8cm]{spec_mgb.pdf} \includegraphics[height=8cm]{spec_ba.pdf} \caption{ \emph{Left panels}: Spectra of the target stars around the Mg b triplet. Mg b, Ti II, and Fe I lines are labeled in black, blue, and red, respectively. Notice the large drop in S/N in TriII-46 at the red end due to reaching the order edge. \emph{Right panels}: our four stars near the Ba line at 6497{\AA}. For Gru~I stars, the solid red curve indicates our best-fit synthesis, while the dotted red curves indicate $\pm 0.15$ dex. For Tri~II stars, the dashed red curves indicate upper limits. In all panels, the dashed blue line indicates $\mbox{[Ba/Fe]} = 0$ for comparison. \label{f:spec}} \end{figure*} \section{Abundance Analysis}\label{s:analysis} We analyzed all four stars using the 2011 version of the 1D LTE radiative transfer code MOOG \citep{Sneden73, Sobeck11} with the \citet{Castelli04} (ATLAS) model atmospheres. We measured equivalent widths and ran MOOG with SMH \citep{Casey14}. The abundance of most elements was determined from equivalent widths. We used spectral synthesis to account for blends, molecules, and hyperfine structure for the species CH, Sc, Mn, Sr, Ba, and Eu. Atomic data references can be found in table~3 of \citet{Roederer10}. Measurements and uncertainties of individual features are in Table~\ref{tbl:eqw}. Stellar parameters and uncertainties for this work and previous measurements are in Table~\ref{tbl:sp}. Final abundances and uncertainties are in Table~\ref{tbl:abunds}. Detailed abundance uncertainties due to stellar parameter variations are in Table~\ref{tbl:spabunderr}. \input{ew_short} \subsection{Standard analysis for brighter stars} For three of our stars (GruI-032, GruI-038, and TriII-40), our spectra are of sufficient quality for a standard equivalent width analysis. We first fit Gaussian profiles to the line list in \citet{Roederer10}. We applied the formula from \citet{Battaglia08} to determine equivalent width uncertainties. The S/N per pixel was calculated with median absolute deviation in a running 5{\AA} window. Varying the window size affected the S/N estimates by only 2-3\%, but we conservatively add an additional 10\% uncertainty to each equivalent width. Using this estimate, we rejected most lines with equivalent width uncertainties larger than 30\%. The exceptions were lines of Al, Si, Cr, Co, and Zn that otherwise would have had all lines of that element rejected; and some clean lines near regions of large true variation (e.g., near CH bands) where the S/N was clearly underestimated. We propagate these to a $1\sigma$ abundance uncertainty for each line (Table~\ref{tbl:eqw}). Synthesis uncertainties are calculated by varying abundances until the entire synthesized profile encompasses the spectrum noise around the feature, corresponding to $1\sigma$ uncertainties. We derived the effective temperature, surface gravity, and microturbulence ($T_{\rm eff}$, $\log g$, $\nu_t$) with excitation, ionization, and line strength balance of Fe lines. We then applied the $T_{\rm eff}$ correction from \citet{Frebel13} and redetermined $\log g$ and $\nu_t$. Statistical uncertainties for $T_{\rm eff}$ and $\nu_t$ correspond to the $1\sigma$ error on the fitted slopes of abundance with respect to excitation potential and reduced equivalent width, respectively. The statistical uncertainty for $\log g$ was derived by varying the parameter to match the combined standard error of the Fe\,I and Fe\,II abundances. We then further adopt systematic uncertainties of 150 K for $T_{\rm eff}$ from scatter in the \citet{Frebel13} calibration; and 0.3 dex for $\log g$, and 0.3 km s$^{-1}$ for $\nu_t$ to reflect this systematic temperature uncertainty. We use the standard deviation of Fe\,I lines as the statistical uncertainty in the stellar atmosphere's model metallicity. We add the statistical and systematic uncertainties in quadrature to obtain the stellar parameter uncertainties in Table~\ref{tbl:sp}. These three stars are all $\alpha$-enhanced, so we used the $\mbox{[$\alpha$/Fe]} = +0.4$ \citet{Castelli04} model atmospheres. \subsection{Analysis of TriII-46} The data for star TriII-46 has very low signal-to-noise (Table~\ref{tbl:obs}) and thus requires special care. We rebin the spectra by a factor of 2 to improve the signal-to-noise. This allowed us to measure equivalent widths for lines at the center of echelle orders with wavelengths $>5000${\AA}. After keeping only lines with equivalent width uncertainty less than 30\%, we have 18 Fe\,I lines and only one Fe\,II line. For this small number of lines, spectroscopic determination of stellar parameters is subject to many degeneracies based on line selection. Still, we examine here what parameters would be derived with the information from Fe lines. If we apply the same procedure as for the other three stars (i.e., excitation, ionization, and line strength balance with the \citealt{Frebel13} correction but using only these 19 lines), we obtain $T_{\rm eff}=5260$\,K, $\log g=2.1$\,dex, $\nu_t=2.60$\,km/s, and $\mbox{[Fe/H]} = -2.01$. However, the ionization equilibrium is set by a single Fe\,II line with equivalent width $164 \pm 50$ m{\AA}, so this is extremely unreliable. Ignoring the Fe\,II line and using a $\mbox{[Fe/H]}=-2$ Yonsei-Yale isochrone to set $\log g$ as a function of $T_{\rm eff}$ \citep{Kim02}, we obtain $T_{\rm eff}=5260$\,K, $\log g=2.7$\,dex, $\nu_t=2.50$\,km/s, and $\mbox{[Fe/H]} = -2.01$. The statistical errors are large: 240K, 0.6 dex, 0.5 km/s, and 0.3 dex respectively. We summarize this and other derived stellar parameters for this star in Table~\ref{tbl:sp}. For comparison, \citet{Venn17} derived $T_{\rm eff}=5050$\,K, $\log g=2.6$\,dex, and $\nu_t=2.5$\,km/s for TriII-46 using photometry, distance, and a modified scaling relation for $\nu_t$. An updated distance modulus \citep{Carlin17} would slightly increase $\log g$ to 2.7\,dex. \citet{Kirby17} derived $T_{\rm eff}=5282$\,K, $\log g=2.74$\,dex, and $\nu_t=1.5$\,km/s using photometry and distance to set $\log g$ and $\nu_t$ but allowing $T_{\rm eff}$ to vary to fit their spectrum. Our stellar parameters are somewhat in between their values, preferring the higher temperature from \citet{Kirby17} but with the higher microturbulence from \citet{Venn17}. Our data for this star are insufficient to make any further refinements, so we decided to adopt intermediate values with large uncertainties that encompass other stellar parameter determinations: $T_{\rm eff}=5150 \pm 200$\,K, $\log g=2.7 \pm 0.5$\,dex, and $\nu_t=2.0 \pm 0.5$\,km/s. Regardless of the stellar parameters, this star is not $\alpha$-enhanced so we use the \citet{Castelli04} model atmospheres with $\mbox{[$\alpha$/Fe]} = 0$. We propagate these uncertainties through to the final abundance uncertainties. \subsection{Final Abundances and Uncertainties} Table~\ref{tbl:abunds} contains the final abundance results for our stars. For each element, $N$ is the number of lines measured. $\log \epsilon(X)$ is the average abundance of those lines weighted by the abundance uncertainty. Letting $\log\epsilon_i$ and $\sigma_i$ be the abundance and uncertainty of line~$i$, we define $w_i = 1/\sigma_i^2$ and $\log\epsilon(X) = \sum_i (w_i \log\epsilon_i) / \sum_i w_i$. $\sigma$ is the standard deviation of those lines. $\sigma_{\rm w}$ is the standard error from propagating individual line uncertainties, i.e., $1/\sigma_{\rm w}^2 = \sum_i w_i$ \citep{McWilliam95}. $\mbox{[X/H]}$ is the abundance relative to solar abundances from \citet{Asplund09}. $\mbox{[X/Fe]}$ is calculated using either [Fe\,I/H] or [Fe\,II/H], depending on whether X is neutral or ionized; except for TriII-46, where all [X/Fe] are calculated relative to [Fe\,I/H] because of an unreliable Fe\,II abundance. $\sigma_{\rm [X/H]}$ is the quadrature sum of $\sigma/\sqrt{N}$, $\sigma_{\rm w}$, and abundance uncertainties due to $1\sigma$ stellar parameter variations. Detailed abundance variations from changing each stellar parameter are given in Table~\ref{tbl:spabunderr}. $\sigma_{\rm [X/Fe]}$ is similar to $\sigma_{\rm [X/H]}$, but when calculating the stellar parameter uncertainties we include variations in Fe. We use the difference in Fe\,I abundance for neutral species and the difference in Fe\,II abundance for ionized species to calculate this error. The [X/Fe] error is usually smaller than the [X/H] error, since abundance differences from changing $T_{\rm eff}$ and $\log g$ usually (but not always) affect X and Fe in the same direction when using the same ionization state. Since most of our elements have very few lines, we adopt the standard deviation of the Fe\,I lines as the minimum $\sigma$ when calculating $\sigma_{\rm [X/H]}$ and $\sigma_{\rm [X/Fe]}$. Upper limits were derived by spectrum synthesis. Using several features of each element (Table~\ref{tbl:eqw}), we found the best-fit synthesis to the observed spectrum to determine a reference $\chi^2$ and smoothing for the synthetic spectrum. The minimum smoothing was calculated using $\rm{FWHM} = \lambda/R$ where $\lambda$ is the line wavelength. Holding the continuum, smoothing, and radial velocity fixed, we increased the abundance until $\Delta\chi^2 = 25$. This is formally a $5\sigma$ upper limit, though it does not include uncertain continuum placement. \begin{deluxetable}{lcrrllllll} \tablecolumns{12} \tablewidth{\linewidth} \tabletypesize{\footnotesize} \tablecaption{Stellar Parameters\label{tbl:sp}} \tablehead{ \colhead{Star} & \colhead{Ref} & \colhead{$T_{\rm eff}$} & \colhead{$\sigma$} & \colhead{$\log g$} & \colhead{$\sigma$} & \colhead{$\nu_t$}\tablenotemark{a} & \colhead{$\sigma$} & \colhead{[Fe/H]} & \colhead{$\sigma$}} \startdata GruI-032 & TW & 4495 & 155 & 0.85 & 0.37 & 2.60 & 0.32 & $-$2.57 & 0.19 \\ GruI-032 &W16\tablenotemark{b} & 4270 & 69 & 0.72 & 0.22 & 2.0\ & \nodata & $-$2.69 & 0.10 \\ \hline GruI-038 & TW & 4660 & 158 & 1.45 & 0.39 & 2.40 & 0.32 & $-$2.50 & 0.24 \\ GruI-038 &W16\tablenotemark{b} & 4532 & 100 & 0.87 & 0.31 & 2.0 & \nodata & $-$2.42 & 0.15 \\ \hline TriII-40 & TW & 4720 & 175 & 1.35 & 0.42 & 2.48 & 0.34 & $-$2.95 & 0.21 \\ TriII-40 &V17 & 4800 & 50 & 1.80 & 0.06 & 2.7 & 0.2 & $-$2.87 & 0.19 \\ TriII-40 &K17\tablenotemark{c} & 4816 & \nodata & 1.64 & \nodata & 2.51 & \nodata & $-$2.92 & 0.21 \\ TriII-40 &K17\tablenotemark{d} & 4917 & \nodata & 1.89 & \nodata & 1.70 & \nodata & $-$2.78 & 0.11 \\ \hline TriII-46 & TW & 5150 & 200 & 2.7 & 0.5 & 2.00 & 0.5 & $-$1.96 & 0.28 \\ TriII-46 &V17 & 5050 & 50 & 2.60 & 0.06 & 2.5 & \nodata & $-$2.5 & 0.2 \\ TriII-46 &K17\tablenotemark{d} & 5282 & \nodata & 2.74 & \nodata & 1.50 & \nodata & $-$1.91 & 0.11 \\ TriII-46 &Spec\tablenotemark{e} & 5260 & 240 & 2.7 & 0.6 & 2.5 & 0.5 & $-$2.01 & 0.26 \\ \enddata \tablerefs{TW = this work; W16 = \citealt{Walker16}; V17 = \citealt{Venn17}; K17 = \citealt{Kirby17}} \tablenotetext{a}{$\nu_t$ for W16 is always 2 km/s \citep{Lee08a}. $\nu_t$ for DEIMOS data in K17 according to the equation $\nu_t = 2.13 - 0.23 \log g$ \citep{Kirby09}} \tablenotetext{b}{[Fe/H] for W16 stars have a 0.32 dex offset removed; see text} \tablenotetext{c}{HIRES data} \tablenotetext{d}{DEIMOS data} \tablenotetext{e}{Spectroscopic balances in this work using isochrones to determine $\log g$} \end{deluxetable} \subsection{Comparison to literature measurements}\label{s:litcomp} For the two Gru~I stars, \citet{Walker16} determined stellar parameters and metallicities from high-resolution M2FS spectra near the Mg b triplet using a large synthesized grid. The grid fixes $\nu_t = 2.0$ \citep{Lee08a}. \citet{Walker16} increased all their [Fe/H] measurements by $0.32$ dex, which is the offset they obtained from fitting twilight spectra of the Sun. It is not clear that the same offset should be applied for both dwarf stars (like the Sun) and giants. If we remove the offset, our stellar parameters and metallicities are in good agreement (also see \citealt{Ji16d}). \citet{Venn17} analyzed both stars in Tri~II, and we have combined their previous GRACES data with additional observations\footnote{\citet{Venn17} labeled the stars as Star~40 and Star~46 instead of TriII-40 and TriII-46. We have retained the number but changed the label to TriII for clarity.}. For TriII-40, we find good agreement for all stellar parameters except $\log g$. This is because we determined our $\log g$ spectroscopically, while \citet{Venn17} did so photometrically using the distance to Tri~II. Adjusting for the different $\log g$, our abundances for this star agree within $1\sigma$. For TriII-46, \citet{Venn17} fixed stellar parameters with photometry and used spectral synthesis to measure all abundances. We measured $\mbox{[Fe/H]} = -2.01 \pm 0.37$, while \citet{Venn17} obtained $\mbox{[Fe/H]} = -2.5 \pm 0.2$. Our large abundance uncertainty means these are only $1.2 \sigma$ discrepant, but we might expect better agreement given that so much of the data overlaps. Detailed investigation of the discrepancy shows that 0.3 dex of the difference is due to differences in stellar parameters (mostly $T_{\rm eff}$ and $\nu_t$). The remaining 0.2 dex is attributable to systematic differences in continuum placement that are individually within $1\sigma$ uncertainties. Finally, we note that the stellar parameter uncertainties in \citet{Venn17} reflect statistical photometric errors, but could be larger due to systematic uncertainties in photometric calibrations, filter conversions, and reddening maps. \citet{Kirby17} determined abundances of TriII-40 with a high-resolution, high signal-to-noise Keck/HIRES spectrum. Our abundances agree within 0.15 dex, except for Cr which is still within $1\sigma$. \citet{Kirby17} also analyzed the Mg, Ca, Ti, and Fe abundance of TriII-46 by matching a synthetic grid to an $R \sim 7000$ Keck/DEIMOS spectrum. They measured $\mbox{[Mg/Fe]}=+0.21 \pm 0.28$, $\mbox{[Ca/Fe]}=-0.39 \pm 0.15$, and $\mbox{[Ti/Fe]}=-0.79 \pm 0.76$. There are some significant discrepancies, especially for Mg. One possible reason for these differences is that we used stronger blue lines with lower excitation potentials for Mg and Ti, while the synthetic grid is driven by combining multiple higher excitation potential lines that we could not individually measure in our spectrum. This explanation is supported by the fact that our Ca abundances agree better because they are derived from similar spectral features. \section{Abundance Results}\label{s:abunds} In Gru~I we measured the abundances of C, Na, Mg, Al, Si, K, Ca, Sc, Ti, Cr, Mn, Fe, Co, Ni, Sr, and Ba. In Tri~II we were only able to measure Mg, K, Ca, Ti, Cr, Fe, Co, Ni, and Ba due to a combination of lower S/N and the fact that the strongest features for other elements are found $\lambda<5000${\AA}. Figures~\ref{f:grid1}, \ref{f:kmg}, and \ref{f:grid2} show the abundances of our four stars compared to other UFDs and a literature sample of halo stars (\citealt{Frebel10}, and \citealt{Roederer14c} for K). The UFDs are Bootes~I \citep{Feltzing09,Norris10a,Gilmore13,Ishigaki14,Frebel16}, Bootes~II \citep{Ji16b}, CVn~II \citep{Francois16}, Coma Berenices \citep{Frebel10b}, Hercules \citep{Koch08, Koch13}, Hor~I \citep{Nagasawa18}, Leo~IV \citep{Simon10,Francois16}, Reticulum~II \citep{Ji16c,Roederer16b}, Segue~1 \citep{Frebel14}, Segue~2 \citep{Roederer14a}, Tuc~II \citep{Ji16d,Chiti18}, Tuc~III \citep{Hansen17}, and UMa~II \citep{Frebel10b}. Overall, the two Gru~I stars have the same [Fe/H] to within our abundance uncertainties, and all [X/Fe] ratios are very similar except for Ba. The metallicities of the Tri~II stars differ by more than $2\sigma$ and display different abundance ratios. We now discuss each element in more detail. \startlongtable \input{abund_final} \begin{figure*} \centering \includegraphics[width=18cm]{lightgrid.pdf} \caption{Abundance of light elements in Gru~I (red squares) and Tri~II (red triangles) compared to halo stars (gray points) and other UFDs (colored points). Upper limits are indicated as open symbols with arrows. The element $X$ is indicated in the top-left corner of each panel. Tri~II stars are not plotted for C. Limits on Tri~II abundances are above the top axis for Sc and Mn. Essentially all [X/Fe] ratios in these two galaxies follow trends defined by the Milky Way halo stars and other UFDs. The notable exceptions are the Na and Ni in TriII-40, and the low Mg and Ca in TriII-46. \label{f:grid1}} \end{figure*} \subsection{Carbon} Spectral synthesis of the $G$-band features at 4313{\AA} and 4323{\AA} was used to measure carbon in the Gru~I stars (using a list from B. Plez 2007, private communication). The oxygen abundance can affect molecular equilibrium, but since oxygen cannot be measured in these stars we assume $\mbox{[O/Fe]}=0.4$. Since they are red giant branch stars, some C has been converted to N. The corrections from \citet{Placco14} were applied to estimate the natal abundance, which are $\mbox{[C/Fe]} = +0.21$ and $+0.57$ for GruI-032 and GruI-038, respectively. Varying $\log g$ by the uncertainty in Table~\ref{tbl:sp} causes the correction to change by $\pm 0.1$ dex. Both stars are carbon-normal ($\mbox{[C/Fe]} < 0.7$) even after this carbon correction. Note that the uncorrected carbon abundances are used in Figure~\ref{f:grid1} and Table~\ref{tbl:abunds}. We were unable to place any constraints on carbon in Tri~II. The GRACES spectra are not usable below 4800{\AA}, so the $G$-band cannot be measured. The CH lists from \citet{Masseron14,Kurucz11} do suggest strong CH features should exist at 5893{\AA} and 8400{\AA} that were used to place a [C/Fe] upper limit by \citet{Venn17}, but we could not find these features in several carbon-enhanced metal-poor stars or atlas spectra of the Sun and Arcturus \citep{Hinkle03}\footnote{\url{ftp://ftp.noao.edu/catalogs/arcturusatlas}}. No other C features are available. \citet{Kirby17} were able to measure $\mbox{[C/Fe]} \sim -0.1$ for TriII-40 from their HIRES spectrum, so this star is not carbon enhanced. \subsection{$\alpha$-elements: Mg, Si, Ca, Ti} The abundances of these four $\alpha$-elements are determined from equivalent widths. The magnesium abundance is determined from $3-6$ lines, but always using two of the Mg b lines. Silicon can only be measured in the Gru~I stars, using the 4102{\AA} line that is in the wing of H$\delta$. The abundance uncertainty from only this single line is quite large, ${\approx}0.6$ dex. Neutral calcium is well-determined by a large number of lines, and it should be considered the most reliable $\alpha$-element. Titanium has several strong lines in both the neutral and singly ionized state, though only a handful ($1-5$) Ti lines can be measured in the Tri~II stars. The abundance of Ti\,I is affected by NLTE effects \citep[e.g.,][]{Mashonkina17}, so we only plot Ti\,II abundances in Figure~\ref{f:grid1} both to avoid NLTE effects and because a Ti\,II line can be measured in all four of our stars. The literature sample also uses Ti\,II whenever possible. \subsection{Odd-Z elements: Na, Al, K, Sc} Sodium is measured from the Na D lines for GruI-032, GruI-038, and TriII-40. While we can identify the presence of Na D lines in TriII-46, the lines are too noisy for a reliable abundance measurement. An upper limit $\mbox{[Na/Fe]} < 1.04$ is found from the subordinate Na lines near 8190\,{\AA}, and for completeness we include the best estimate of equivalent widths for the Na D lines in Table~\ref{tbl:eqw}. NLTE corrections are not applied since most stars in the literature comparison sample do not have these corrections, but the grid from \citet{Lind11} gives corrections of $-0.28$ for GruI-032, $-0.32$ for GruI-038, and $-0.06$ for TriII-40. The two Gru~I stars have solar ratios of Na, following the usual halo trend. In contrast, TriII-40 has significantly subsolar $\mbox{[Na/Fe]} =-0.79 \pm 0.22$ that is an outlier from the halo trend, as first reported by \citet{Venn17}. A similarly low [Na/Fe] ratio has previously been seen in one of three stars in the UFD Coma Berenices \citep{Frebel10b}. The primordial (first generation) population of stars in globular clusters also have low Na, but all with $\mbox{[Na/Fe]} > -0.5$, \citep{Gratton12}. Aluminum and scandium are only measured in the Gru~I stars. Al is determined from a single line at 3961{\AA}. Given the low S/N in this region, Al is the least certain abundance of all elements measured here. The measurement is consistent with that of other halo stars at $\mbox{[Fe/H]} \approx -2.5$, but it is not a meaningful constraint. Sc lines in Gru~I are synthesized due to hyperfine structure \citep{Kurucz95}, and the abundances are also similar to other halo stars. For completeness, we place Sc upper limits in the Tri~II stars with some weak red lines that provide no interesting constraint. \begin{figure*} \centering \includegraphics[width=18cm]{spec_k.pdf} \caption{Spectrum around K lines for our four stars. Black lines are the data, solid red lines indicate synthesis fit, dotted red lines indicate uncertainty. In the third row (TriII-40), dark blue lines indicate data from Dec 2015, while cyan lines indicate data from Sep 2016, showing how the location of telluric absorption shifts relative to the K line. The 7699 line is cleanly detected in Sep 2016 (telluric lines are at 7699.5 {\AA} and 7701.8 {\AA}). The same abundance is synthesized at the expected strength of the 7665 line, but we do not use that line because it is too blended with telluric absorption. In the fourth row (TriII-46), the dashed red line is the K upper limit, and the dashed blue line indicates [K/Fe]=1. \label{f:kspec}} \end{figure*} Potassium has two strong lines at 7665\,{\AA} and 7699\,{\AA}. These lines are located near several telluric absorption features. Figure~\ref{f:kspec} shows these two lines and the best-fit synthetic spectrum or upper limits. The top two spectra are Gru~I observations, where observations were conducted within the span of one month, so the telluric features do not move much due to heliocentric corrections. Both Gru~I stars have K lines that are easily distinguished from the telluric features. The bottom two spectra of Figure~\ref{f:kspec} are Tri~II observations, which were conducted in Dec 2015 and Sep 2016. The heliocentric correction is different between these epochs by ${\sim}40$ km/s, so the telluric features shift by ${\approx}1${\AA} between 2015 and 2016. We emphasize this for TriII-40 by showing individual frames from Dec 2015 (thin blue lines) and Sep 2016 (thin cyan lines). Note that we used \texttt{scombine} in IRAF with \texttt{avsigclip} rejection to obtain the coadded black spectra, so it tends to follow the telluric lines from Sep 2016 (four exposures) rather than Dec 2015 (two exposures). The 7699\,{\AA} line is detected in TriII-40. It is significantly blended with a telluric line in the Dec 2015 observations (see \citealt{Venn17} figure 4, dark blue lines here), but cleanly separated in the Sep 2016 observations. We find [K/Fe] = 0.8 in TriII-40, in agreement with the measurement by \citet{Venn17}. The 7665\,{\AA} line is severely blended with telluric lines in both epochs, so we do not use it but just highlight its position in Figure~\ref{f:kspec} with a synthesized K line. Neither K line is detected for TriII-46, and an upper limit $\mbox{[K/Fe]} < 0.77$ is set with the 7699\,{\AA} line. We could not account for the telluric lines when setting this upper limit, but this makes the limit more conservative. NLTE corrections have not been applied, but they can be large (as high as $-0.4$ dex for the most K-enhanced stars in LTE, \citealt{Andrievsky10}). \begin{figure} \centering \includegraphics[width=8cm]{KMg.pdf} \caption{K and Mg abundances of stars in the Tri~II and Gru~I from this work, compared to K and Mg in the stellar halo \citep{Roederer14c}, NGC2419 \citep{Mucciarelli12} and other UFDs (Boo~II, \citealt{Ji16a}; Ret~II, \citealt{Ji16c}; Tuc~II, \citealt{Ji16d}; Segue~2, \citealt{Roederer14a}; and Tuc~III, \citealt{Hansen17}). The K abundance of TriII-46 is not enhanced, so Tri~II does not follow the strange K-Mg anticorrelation in NGC2419. Note that the halo sample here is different than in Figures~\ref{f:grid1} and \ref{f:grid2} because our usual halo compilation does not have K abundances \citep{Frebel10}. Adapted from \citet{Venn17}. \label{f:kmg}} \end{figure} \subsection{Iron-peak elements: Cr, Mn, Co, Ni, Zn} The Fe-peak abundances were determined with equivalent widths, except for Mn, which is synthesized due to hyperfine structure \citep{Kurucz95}. In Gru~I we can constrain Cr, Mn, Co, and Ni, finding that both stars have essentially identical abundances of these elements. Though the Cr and Mn abundances are similar to those in metal-poor stars in other UFDs or in the Milky Way halo, the Co and Ni abundances are somewhat higher. However, this difference is not very significant, especially for Co which is derived from only a few bluer lines. One Zn line is marginally detected in GruI-032 with an abundance consistent to the halo trend, though with large uncertainty. In Tri~II, we can detect Cr and Ni in TriII-40 and provide upper limits in TriII-46. Mn, Co, and Zn are unconstrained as they only have strong lines blueward of 5000{\AA}. The upper limits for Cr and Ni in TriII-46 are uninteresting. For TriII-40, we detect a normal [Cr/Fe] ratio, but Ni appears significantly enhanced ($\mbox{[Ni/Fe]} = 0.57 \pm 0.16$), in agreement with \citet{Kirby17} and \citet{Venn17}. \subsection{Neutron-capture elements: Sr, Ba, Eu} Strontium is detected only in Gru~I, as the strong Sr\,II lines at 4077{\AA} and 4215{\AA} are out of the range of the Tri~II (GRACES) spectra. The abundance of both lines is determined with spectrum synthesis. The Sr abundances in these two stars are very similar, $\mbox{[Sr/Fe]} \approx -2$, which is much lower than what is found in most halo stars but similar to most UFDs (Figure~\ref{f:grid2}). Barium is measured with four different lines in the Gru~I stars including hyperfine structure and isotope splitting \citep{McWilliam98}. We use solar isotope ratios \citep{Sneden08}, but given the low overall abundance, changing this to $r$- or $s$-process ratios does not significantly affect our abundances. GruI-032 has a low $\mbox{[Ba/Fe]} \approx -1.6$, but GruI-038 has a much higher Ba abundance $\mbox{[Ba/Fe]} \approx -1.0$. This is formally only 1.6$\sigma$ different, but differential comparison of the line strengths (e.g., the 6497{\AA} line in Figure~\ref{f:spec}) suggests that the difference is real. We discuss this more in Section~\ref{s:discoutlier}, but both Ba abundances are low and similar to those in most UFDs. Ba is not detected in either Tri~II star, so instead we place $5\sigma$ upper limits. The Ba limit for TriII-40 is $\mbox{[Ba/Fe]} < -1.25$, suggesting a low Ba abundance similar to other UFDs. \citet{Kirby17} determined $\mbox{[Sr/Fe]}=-1.5$ and $\mbox{[Ba/Fe]}=-2.4$ from their HIRES spectrum of this star, consistent with our upper limit and showing TriII-40 clearly has very low neutron-capture element abundances. The Ba limit for TriII-46 is only $\mbox{[Ba/Fe]} \lesssim -0.2$, but this is still at the lower envelope of the halo trend (Figure~\ref{f:grid2}). Eu is not detected in any of these four stars, as expected given the low Sr and Ba abundances. Upper limits are placed from the 4129{\AA} line for the Gru~I stars (MIKE data) and from the 6645{\AA} line for the Tri~II stars (GRACES data). \begin{figure*} \centering \includegraphics[width=18cm]{ncap_1x3_v2.pdf} \caption{Abundances of Sr and Ba in UFDs compared to halo stars. Symbols are as in Figure~\ref{f:grid1}. The left two panels show the abundance trend with respect to [Fe/H]. Note that there is no constraint on Sr for Tri~II stars. The rightmost panel shows that most halo stars cluster near $\mbox{[Sr,Ba/Fe]} \approx 0$, but most UFDs are clearly offset to lower Sr and Ba. \label{f:grid2}} \end{figure*} \section{Discussion} \label{s:discussion} \subsection{Abundance Anomalies}\label{s:discoutlier} The abundance ratios of the two stars in Gru~I are nearly identical to each other, and similar to typical UFD stars at $\mbox{[Fe/H]} \approx -2.5$. The most notable exception is the Ba abundance, where GruI-038 has 0.6 dex higher [Ba/Fe] than GruI-032. After applying corrections from \citet{Placco14}, GruI-038 also has a higher carbon abundance than GruI-032 ($\mbox{[C/Fe]} = +0.57$ vs $+0.21$, respectively). The differences are both somewhat low significance, and it is reasonable to consider these two stars chemically identical. However if the differences are real, one possible explanation is that GruI-038 formed from gas that had been polluted by more AGB stars compared to GruI-032. A lower mass ($1-4 M_\odot$) AGB star could add significant Ba and C without changing the Sr abundance too much \citep[e.g.,][]{Lugaro12}. Since AGB winds are low velocity, their C and Ba production would be more inhomogeneously distributed in the star-forming gas of a UFD progenitor \citep[e.g.,][]{Emerick18}. However, many more stars in Gru~I would be needed to test this scenario. We confirm the result from \citet{Venn17} that TriII-46, the more Fe-rich star in Tri~II, has very low [Mg/Fe] and [Ca/Fe] ratios. The standard interpretation is that TriII-46 must have formed after significant enrichment by Type~Ia supernovae, and this star does follow the decreasing [$\alpha$/Fe] trend of other stars in Tri~II \citep{Kirby17}. Indeed, most other UFDs show a similar downturn in [$\alpha$/Fe] ratios as [Fe/H] increases \citep{Vargas13}, though Horologium~I is unique in that all known stars in the system have low [$\alpha$/Fe] \citep{Nagasawa18}, and Segue~1 is unique in that it shows no downturn in $\alpha$-elements at high [Fe/H] \citep{Frebel14}. It is actually somewhat surprising that the very low-luminosity Tri~II appears to have formed stars long enough to be enriched by Type~Ia supernovae, since its luminosity is very similar to Segue~1. If Tri~II were significantly tidally stripped by now \citep{Kirby15,Kirby17,Martin16} this would help reconcile enrichment by Type~Ia supernovae with the small present-day luminosity. However, the orbital pericenter of Tri~II is 20 kpc, where tidal effects are not too strong \citep{Simon18}; and there are no visible signs of tidal disruption in deep imaging \citep{Carlin17}. An alternate explanation could be the presence of very prompt Type~Ia supernovae \citep[e.g.,][]{Mannucci06}. If this is the case, it may have implications for the single-degenerate vs. double-degenerate debate of Type~Ia supernova progenitors. Short detonation delay times (${\sim}100s$ of Myr) are a common feature of double-degenerate models, and less common (though still possible) for single-degenerate models \citep[e.g.,][]{Maoz14}. One way to distinguish these models in Tri~II would be to examine Fe-peak elements like Mn, Co, and Ni \citep[see][]{McWilliam18}; but these elements are unavailable in our GRACES spectra. \citet{Venn17} first noticed that the K and Mg abundances in Tri~II could match the unusual globular cluster NGC2419, which displays a K-Mg anticorrelation of unknown origin \citep{Cohen12,Mucciarelli12}. If so, then TriII-46 should have very high $1 < \mbox{[K/Fe]} < 2$ (Figure~\ref{f:kmg}). Our new limit of $\mbox{[K/Fe]} \lesssim 0.8$ in TriII-46 suggests that Tri~II probably does not display the same K-Mg anticorrelation as NGC2419. [K/Fe] is often enhanced in LTE, both for UFD stars and halo stars \citep{Roederer14c}. NLTE effects tend to amplify the strengths of the resonance lines for K-enhanced stars, so they likely contribute to the apparent overabundance of K in these stars \citep{Andrievsky10}. We also confirm results from \citet{Kirby17} and \citet{Venn17} that TriII-40 has very low $\mbox{[Na/Fe]} = -0.79 \pm 0.22$ and somewhat high $\mbox{[Ni/Fe]} = 0.57 \pm 0.16$. This star has $\mbox{[Fe/H]} \sim -3$ and enhanced $\alpha$-elements, so we would nominally expect its abundance ratios to predominantly reflect the yields of metal-poor core-collapse supernovae (CCSNe). It is somewhat counterintuitive to find enhanced Ni and depressed Na in a CCSN, as the production of both elements is positively correlated with the neutron excess in a supernova \citep[e.g.,][]{Venn04,Nomoto13}. However, this appears to break down at the lowest metallicities, and the online \emph{Starfit} tool\footnote{\url{http://starfit.org/}} finds that a Pop\,III supernova progenitor (11.3 $M_\odot$, $E=3 \times 10^{51}$\,erg, from the supernova yield grid of \citealt{Heger10}) provides a decent fit to the Mg, Ca, Ti, Fe, and Ni abundances ($\mbox{[Na/Fe]} \approx -1.0$, $\mbox{[Ni/Fe]} \approx +0.2$). An alternate possibility is that this $\mbox{[Fe/H]} \sim -3$ star formed from gas already affected by Type~Ia supernovae, as Chandrasekhar mass explosions can produce high [Ni/Fe] \citep[e.g.,][]{Fink14} while reducing [Na/Fe] by adding iron. It seems very unlikely to form and explode a white dwarf so early in this galaxy's history, but age and metallicity may be decoupled at early times due to inhomogeneous metal mixing \citep[e.g.,][]{Frebel12,Leaman12,Nomoto13}. A very prompt population of Type Ia's with merging delay times as low as 30 Myr could also exist \citep{Mannucci06}. We note that the Na and Ni lines in our spectrum of TriII-46 are very noisy and cannot provide a reliable abundance, but the best-fit abundance estimates (Table~\ref{tbl:eqw}) do suggest this star also has low [Na/Fe] and enhanced [Ni/Fe]. \subsection{Classification as dwarf galaxy or globular cluster} \label{s:discclassify} In this paper, we consider three criteria that can be used to classify Tri~II and Gru~I as either ultra-faint dwarf galaxies or globular clusters. \noindent (1) a velocity dispersion indicating the presence of dark matter, \\ \noindent (2) an [Fe/H] spread implying the ability to form multiple generations of stars despite supernova feedback, or significant internal mixing, and \\ \noindent (3) unusually low neutron-capture element abundances compared to halo stars. The first two criteria were codified by \citet{Willman12} and imply that the stellar system is the result of extended star formation in a dark matter halo. The third criterion is based on previous studies of UFDs confirmed by the other two criteria \citep[e.g.,][]{Frebel10b, Frebel14, Frebel15, Simon10, Koch13}, and it has recently been used as a way to distinguish UFD stars from other stars \citep[e.g.,][]{Kirby17,Casey17,Roederer17}. Unlike the first two criteria, this is a criterion specifically for the lowest mass galaxies, rather than defining galaxies in general. Note that violating the criterion also does not preclude an object from being a UFD, as is evident from the $r$-process outliers Ret~II and Tuc~III that experienced rare $r$-process enrichment events. However, when multiple stars are observed in the same UFD, the majority of stars do tend to have similar neutron-capture element abundances. We discuss possible explanations for criterion (3) in Section~\ref{s:discwhy}, but first accept it as an empirical criterion. \subsubsection{Triangulum~II} The case of Tri~II was already extensively discussed by \citet{Kirby15,Kirby17,Martin16,Venn17,Carlin17}, generally finding that it is most likely a UFD rather than a star cluster. Our high-resolution abundance results are consistent with the discussion in \citet{Venn17} and \citet{Kirby17}, namely that we find a difference in [Fe/H] between these two stars at about $2\sigma$ significance, and TriII-46 has lower [$\alpha$/Fe] ratios compared to TriII-40. \citet{Kirby17} previously found very low Sr and Ba abundances in TriII-40, and our Ba limit on TriII-46 is consistent with overall low neutron-capture element abundances in Tri~II (though additional data is needed to confirm that TriII-46 is well below the halo scatter). Tri~II thus likely satisfies criteria (2) and (3), though it is unclear if it satisfies criterion (1) (see \citealt{Kirby17}, figure~2). Our main additional contribution here is a more stringent upper limit on K in TriII-46 (Figure~\ref{f:kmg}) as discussed above in Section~\ref{s:discoutlier}, which shows Tri~II does not have the abundance signature found in the globular cluster NGC2419. \subsubsection{Grus~I} \citet{Walker16} identified seven probable members in Gru~I. This sample was insufficient to resolve either a velocity dispersion or metallicity dispersion. Our high-resolution followup of two stars has found that those stars have indistinguishable [Fe/H]. Thus, Gru~I does not currently satisfy criteria (1) or (2) to be considered a galaxy. However, we have found that the neutron-capture element abundances in Gru~I are both low and similar to UFDs, satisfying criterion (3). Gru~I thus most likely appears to be a UFD, and we expect that further spectroscopic study of Gru~I will reveal both metallicity and velocity dispersions. We note that the velocity difference in our two Gru~I stars alone does already suggest a potentially significant velocity dispersion. The mean metallicity determined by \citet{Walker16} for Gru~I is $\mbox{[Fe/H]} \sim -1.4 \pm 0.4$, which placed it far from the luminosity-metallicity trend of other dSph galaxies, while globular clusters do not have such a relationship. However, the two brightest stars, analyzed here, both have $\mbox{[Fe/H]} \sim -2.5$ that would be consistent with the mean trend. Only ${\sim}0.3$\,dex of the difference can be attributed to their metallicity zero-point offset (see Section~\ref{s:litcomp}). The rest of the discrepancy is due to the fact that \citet{Walker16} found their other five members of Gru~I to have a much higher [Fe/H] than these two stars, ranging from [Fe/H] = $-2$ to $-1$. Those five fainter stars are over 1 mag fainter than our stars, currently out of reach for high-resolution spectroscopic abundances so we cannot test the true metallicity of Gru~I with our data. However, those stars also have very low S/N, and inferred effective temperatures that are much higher than expected based on photometry alone. We thus suggest the metallicity of Gru~I is probably closer to the value measured from our two stars. Recently, \citet{Jerjen18} published deep photometry of Gru~I, with isochrone-based metallicities of $\mbox{[Fe/H]} = -2.5 \pm 0.3$. \subsection{Why do most UFDs have low neutron-capture element abundances?}\label{s:discwhy} \begin{figure*} \includegraphics[width=18cm]{cldw_ncapfe_allsame_2x2.pdf} \caption{Neutron-capture element abundances for UFDs (yellow diamonds; separating Boo~I as red diamonds; and Ret~II and Tuc~III as large dark red stars), classical dSphs (blue and orange symbols), globular clusters (large purple circles), and halo stars (grey points). Classical dSph stars come from \citealt{Aoki09,Cohen09,Cohen10,Frebel10a,Fulbright04,Geisler05,JHansen18,Jablonka15,Kirby12,Norris17b,Shetrone01,Shetrone03,Simon15Scl,Skuladottir15,Tafelmeyer10,Tsujimoto15a,Tsujimoto17,Ural15,Venn12}. \label{f:ncapcomp}} \end{figure*} \begin{figure} \includegraphics[width=8.5cm]{dsph_Mstar_Mdyn_ncap.pdf} \caption{Absolute $V$ magnitude vs dynamical mass within half light radius for dSphs with neutron-capture element constraints. Galaxies are color-coded according to their [Sr/Fe] and [Ba/Fe] abundance at $-3.5 \lesssim \mbox{[Fe/H]} \lesssim -2.5$. Yellow points have both low Sr and Ba, orange points have low Sr but regular Ba, blue points have regular Sr and Ba. For comparison, we also show globular clusters in the \citet{Pritzl05} sample with $\mbox{[Fe/H]} \lesssim -2$. The dynamical data and luminosity for dwarf galaxies come from \citet{Munoz18}, supplemented by \citealt{Majewski03,Bechtol15}. Velocity dispersions are from \citealt{Bellazzini08,Kirby13a,Kirby17,Koch09,Koposov11,Simon07,Simon11,Simon15,Simon17,Simon19,Walker09c,Walker09,Walker16}. $M_{\rm dyn}$ is computed with the equation in \citet{Walker09}. Globular cluster data are from \citet{Harris10}. \label{f:dsphbyncap}} \end{figure} Figure~\ref{f:ncapcomp} shows the neutron-capture element abundances of UFD stars relative to halo stars and classical dSph stars. Excluding Ret~II and Tuc~III, it is clear that UFDs have low neutron-capture element abundances relative to these other populations in both Sr and Ba, and most apparent in Sr. The astrophysical origin (or origins) of these low-but-nonzero neutron-capture element abundances is still an open question (see Section~\ref{s:discncap}). However, the abundance signature of this (or these) low-yield site(s) is usually hidden in more metal-rich stars. This is clearly seen by examining the classical dSph galaxies, which are somewhat more evolved than UFDs due to their higher mass. In Sculptor, we can see a ${>}1$ dex rise in [Sr/Fe] from UFD levels to typical halo star levels occurring at very low metallicity $-4 < \mbox{[Fe/H]} < -3.3$; while a rise in [Ba/Fe] occurs later, at $\mbox{[Fe/H]} \sim -2.5$ \citep[also see][]{Jablonka15,Mashonkina17}. Similar trends exist for Sagittarius, Sextans, and Ursa Minor. We highlight Draco and Carina separately, as their stars' [Sr/Fe] ratios stay similarly low to UFDs until $\mbox{[Fe/H]} \gtrsim -2.5$, but unlike UFDs their [Ba/Fe] ratios rise with [Fe/H]. The UFD Boo~I is similar to Draco and Carina and unlike most UFDs in this sense, as well. The rise in Sr and Ba suggests the delayed onset of different, more prolific, sources of neutron-capture elements, presumably some combination of AGB stars and neutron star mergers. These higher-yield later-onset sources of Sr and Ba will eventually dominate total Sr and Ba production. Overall, it seems that larger galaxies manage to reach a ``normal'' halo-like neutron-capture element abundance at lower [Fe/H] than smaller galaxies, implying that they can be enriched by those dominant sources of Sr and Ba \citep[also see][]{Tafelmeyer10,Jablonka15}. The question of why UFDs have low neutron-capture element abundances thus boils down to why these high-yield sources of neutron-capture elements do not contribute metals to most UFDs while they are forming stars. We can imagine three possible reasons: \begin{enumerate} \item UFDs do not form enough stars to fully sample all metal yields from a stellar population. If the dominant sources of Sr and Ba are produced rarely or stochastically, they will only occasionally enrich a given UFD; so most UFDs would have low [Sr,Ba/Fe] \citep[e.g.,][]{Koch08,Koch13,Simon10,Venn12,Venn17,Ji16b}. \item UFDs form in small potential wells, so they do not retain metals very well \citep[e.g.,][]{Kirby11outflow,Venn12}. If the dominant sources of Sr and Ba are lost with higher efficiency in UFDs (relative to iron), this would result in low [Sr,Ba/Fe]. \item UFDs form stars for only a short time. If the dominant neutron-capture element sources have long delay times (e.g., neutron star mergers or AGBs), these sources may only produce metals after UFDs have finished forming stars. Then, surviving UFD stars would not preserve the metals from those sources. \end{enumerate} We note that Sr and Ba appear to have differing trends, so the explanations for Sr and Ba may differ as well. As one attempt to distinguish between these possibilities, we consider whether there are correlations with stellar mass or current dynamical mass. Figure~\ref{f:dsphbyncap} shows the absolute magnitude and inferred dynamical mass within the half light radius for several classical dSphs and UFDs. The yellow points are UFDs that have low [Sr/Fe] and [Ba/Fe]. Blue points are classical dSphs (UMi, Sex, Scl, Sgr) that have regular Sr and Ba trends. In orange we highlight Boo~I, Carina, and Draco, which have low Sr at $\mbox{[Fe/H]} \sim -3$, but Ba behavior similar to the more massive dSphs. We also note that Draco, UMi, and all UFDs have CMDs indicating purely old stellar populations ($>10-12$\,Gyr old), while more luminous dSphs (Carina and above) show evidence for some late time star formation \citep{Weisz14a,Brown14}. There is a broad transition in neutron-capture element content occurring somewhere between $-6 > M_V > -10$ and $10^6 < M_{\rm dyn}/M_\odot < 10^7$, also roughly corresponding to the purely old dSphs. Unfortunately, given the strong correlations between luminosity, dynamical mass, and overall age in this sample, it is hard to distinguish between the three reasons listed above for low neutron-capture elements in UFDs. Explanation (2) is somewhat disfavored if one accepts two stronger assumptions. First, $M_{\rm dyn}(<r_{1/2})$ is not a good measure of the total halo mass, because the half light radius is only a tiny fraction of the overall halo size. Correcting for this requires extrapolating an assumed density profile to larger radii, but such extrapolations imply that UFDs and even some of the larger dSphs may all reside in dark halos of similar mass \citep{Strigari08}. A similar dark halo mass is also expected from a stellar-mass-to-halo-mass relation with large intrinsic scatter \citep[e.g.,][]{Jethwa18}. Second, one must assume that $z=0$ halo masses are highly correlated with halo masses at the time of star formation. This is true on average in $\Lambda$CDM, but it breaks down in specific cases due to scatter in halo growth histories \citep[e.g.,][]{Torrey15} and tidal stripping from different subhalo infall times \citep[e.g.,][]{Dooley14}. Together, these two assumptions would imply that neutron-capture element behavior is uncorrelated with halo mass, disfavoring explanation (2). Furthermore, comparison to classical dSphs suggests the short star formation timescale (3) is unlikely for Sr: more massive dSphs like Scl and Sgr are much more efficient at forming stars, but they are already Sr-enriched at $\mbox{[Fe/H]} \sim -3$. It may thus be the case that explanation (1) is the most likely one for Sr, i.e. that the dominant source of Sr is stochastically produced. However, explanations (1) and (3) both remain viable for Ba; and explanation (2) remains for both Sr and Ba as well if the two stronger assumptions do not hold. \subsection{Comparison to globular clusters}\label{s:gc} Globular clusters (GCs) have very different neutron-capture element abundances than UFDs. Figure~\ref{f:ncapcomp} shows the mean abundances of GCs as purple circles (compiled in \citealt{Pritzl05}\footnote{We have removed NGC 5897, NGC 6352, and NGC 6362 from this compilation, which were outliers in [Ba/Fe]. These three GCs were all observed by \citet{Gratton87} and scaled to a common $\log gf$ scale by \citet{Pritzl05}. However, the abundances derived by \citet{Gratton87} appear to conflict with the $\log gf$, and we suspect a typographical error for $\log gf$. We confirm this in NGC 5897 with more recent measurements by \citet{Koch14c}.}). Sr is usually not measured in GCs, so we also show Y (which has similar nucleosynthetic origins as Sr). It is immediately obvious that all neutron-capture elements in globular clusters closely trace the overall halo trend, as well as more metal-rich stars in classical dSphs. In contrast, UFDs tend to lie at the extremes of the halo trend. The origin of globular clusters is unknown, but one class of theories posits that metal-poor GCs form as the dominant stellar component of a small dark matter halo, rather than as a part of a larger galaxy \citep[e.g.,][and references therein]{Forbes18}. Such theories usually have GCs form in the \emph{same} dark matter halos as UFDs (i.e., ${\sim}10^8 M_\odot$ dark matter halos that experience atomic line cooling), but something (e.g., a gas-rich merger) triggers them to become GCs instead of UFDs \citep[e.g.,][]{Griffen10,Trenti15,Ricotti16,Creasey18}. However, if GCs do form in these small atomic cooling halos, their neutron-capture element enrichment should match that of UFDs, i.e. be very low, or at least show significant GC-to-GC scatter\footnote{At least one metal-poor globular cluster, M15, does show a significant \emph{internal} dispersion in neutron-capture element abundances ($>0.6$\,dex; \citealt{Sneden97}). Some other GCs might also display such a dispersion, though it is much smaller ($0.3$\,dex) and could be due to systematic effects \citep{Roederer11b,Cohen11b,Roederer15}. Either way, this dispersion is not enough to match the neutron-capture element deficiency seen in most UFD stars with $\mbox{[Fe/H]} \gtrsim -2.5$.}. The difference in neutron-capture element abundances thus seems to imply that the known metal-poor GCs in the Milky Way formed as part of larger galaxies (e.g., \citealt{BoylanKolchin17}), rather than in their own dark matter halos. Note that the neutron-capture element abundances are not affected by the multiple abundance populations usually discussed in globular clusters \citep[e.g.,][]{Gratton04,Roederer11b}. Those variations in lighter abundances are due to an internal mechanism, rather than tracing the natal abundance of the gas the GCs formed from \citep[see e.g.,][and references therein.]{Gratton12,Bastian18}. \subsection{On the origin of the ubiquitous neutron-capture element floor} \label{s:discncap} We briefly discuss the most viable candidates for this ubiquitous low-yield neutron-capture element source occurring at low metallicity. This is important not just for understanding UFD enrichment, but also for the most metal-poor halo stars, where Sr and/or Ba appear to be ubiquitously present at the level $\mbox{[Sr,Ba/H]} \sim -6$ \citep{Roederer13}\footnote{To our knowledge, the only star with limits below this threshold is a star with no detected Fe, SMSS 0313$-$6708, with extremely low limits $\mbox{[Sr/H]} < -6.7$ and $\mbox{[Ba/H]} < -6.1$ \citep{Keller14}.}. The sources must explain the ubiquitous presence of both Sr and Ba, the overall low but nonzero yield of both Sr and Ba, and the fact that the [Sr/Ba] ratio in UFDs varies over ${\sim}2$ dex. \emph{Neutrino-driven wind.} The high-entropy neutrino-driven wind in CCSNe was initially thought to be a promising site for Sr and Ba production in the $r$-process \citep[e.g.,][]{Woosley92}, but contemporary simulations suggest wind entropies an order of magnitude too low to produce the full set of $r$-process elements up to uranium \citep[e.g.,][]{Arcones07}. It still seems that this mechanism robustly produces a limited form of the $r$-process that always synthesizes Sr, but a little bit of Ba only under extreme conditions (e.g., neutron star mass $>2 M_\odot$, \citealt{Wanajo13}). Supporting this, \citet{Mashonkina17} recently argued for two types of Sr production, one of which was highly correlated with Mg, implying CCSNe could produce Sr alone. However, current models suggest that even extreme neutrino-driven winds cannot produce $\mbox{[Sr/Ba]} \sim 0$ \citep{Arcones11,Wanajo13}, so while they may be an important factor they probably are not the only source of neutron-capture elements in most UFDs. \emph{Magnetorotationally driven jets.} A dying massive star with extremely strong magnetic fields and fast rotation speeds can launch a neutron-rich jet that synthesizes copious Sr and Ba in the r-process \citep[e.g.,][]{Winteler12,Nishimura15}. It is still debated whether such extreme conditions can be physically achieved in massive star evolution \citep[e.g.,][]{Rembiasz16a,Rembiasz16b, Mosta17}. However, if the conditions are less extreme, such supernovae can actually produce both Sr and Ba without synthesizing the heaviest r-process elements in a delayed jet \citep{Nishimura15,Nishimura17,Mosta17}. These more moderate rotation speeds and magnetic fields may be more plausible results of massive and metal-poor stellar evolution, and so the rate of these moderate jet explosions could occur much more often than is invoked to explain prolific $r$-process yields. If so, then we propose that delayed magnetorotationally driven jets are a viable source of the low Sr and Ba abundances in UFDs. Additional modeling focusing on the frequency of less-extreme jets is needed for a more detailed evaluation, and zinc abundances may help as well \citep{Ji18}. \emph{Spinstars.} Spinstars are rapidly rotating massive stars that can produce Sr and Ba in the $s$-process \citep[e.g.,][]{Meynet06}. The amount of rotation changes the amount of internal mixing in the star, allowing these models to produce a wide range of [Sr/Ba] ratios, though the amount of Ba is still subject to nuclear reaction rate uncertainties \citep{Cescutti13, Frischknecht16, Choplin18}. The fiducial spinstar models in \citet{Frischknecht16} underproduce Sr and Ba by a factor of ${>}100$ to explain the observed values in UFDs \citep[e.g.,][]{Ji16d}, and having hundreds of spinstars in each UFD is unlikely given there are only hundreds of massive stars to begin with in each galaxy. However, extreme spinstar models with particularly fast rotation velocities and a modified nuclear reaction rate increase the abundance yields by a factor $>10$ \citep{Cescutti13,Frischknecht12,Frischknecht16}\footnote{Yields from \url{http://www.astro.keele.ac.uk/shyne/datasets/s-process-yields-from-frischknecht-et-al-12-15}}. These models also produce [C/Sr] and [C/Ba] $\sim +2.0$, consistent or somewhat lower than the C abundances in UFDs like Gru~I. The [C/Fe] ratios are very high ($>3.0$), but the spinstar yields do not include any carbon or iron generated in a supernova explosion, which would reduce this extreme abundance ratio. Thus, the extreme spinstar models are also a viable source for the neutron-capture elements found in UFDs. Note that rotation is not the only way that neutron-capture processes can occur in metal-poor or metal-free stars, as it is just one of many possible mechanisms that can induce internal mixing and thus create free neutrons. Recently, \citet{Banerjee18a} and \citet{Clarkson18} have shown that proton ingestion into convective He shells can result in a low level of $s$-, $i$-, and $r$-processes in even in metal-free stars. Some of the metal-poor models by \citet{Banerjee18a} are able to produce explain the low but nonzero amounts of Sr and Ba found in UFDs, as well as the diversity of [Sr/Ba] ratios. \emph{An unknown low-yield r-process source.} As of now, binary neutron star mergers are the only confirmed source of the full $r$-process (i.e., produces all elements from the 1st through 3rd $r$-process peaks). However, there is evidence from halo stars with low Sr and Ba that UFDs are enriched by a low-yield (or heavily diluted) version of the same abundance pattern. \citet{Roederer17} found three halo stars with low Sr and Ba as well as Eu detections consistent with the $r$-process ($-4 < \mbox{[Eu/H]} < -3.5$). \citet{Casey17} found a halo star with $\mbox{[Sr,Ba/H]} \approx -6$, with $\mbox{[Sr/Ba]} \sim 0$ consistent with the full $r$-process. Assuming that these halo stars originated in now-tidally-disrupted UFDs, that might imply that a low-yield but robust $r$-process does occur. This has long been assumed to take place in some subset of core-collapse supernovae, but as mentioned above, current models cannot achieve this reliably. However, UFDs display variations in [Sr/Ba] that cannot be explained by just a single $r$-process. Disentangling these different sites will require determining abundances of neutron-capture elements other than Sr and Ba in UFD stars. Given the distance to known UFDs, this will require significant time investments with echelle spectrographs on 30m class telescopes. In the meantime, progress can be made by study of bright, nearby halo stars with low Sr and Ba abundances \citep[e.g.,][]{Roederer17}. For this purpose, the best stars are the relatively Fe-rich but Sr- and Ba-poor stars, as these are the ones most clearly associated with UFDs (Figure~\ref{f:ncapcomp}). Such stars are expected to comprise $1-3$\% of halo stars at $-2.5 < \mbox{[Fe/H]} < -2.0$ \citep{Brauer18}. \section{Conclusion} \label{s:conclusion} We present detailed chemical abundances from high-resolution spectroscopy of two stars in Gru~I and two stars in Tri~II. Overall, the abundance ratios of these stars are generally similar to those found in other ultra-faint dwarf galaxies, including extremely low neutron-capture element abundances. The Gru~I stars are nearly chemically identical, except for possibly a different Ba abundance. A possible similarity between Tri~II and the cluster NGC~2419 is probably ruled out by a new K upper limit, and there may also be an anomaly in Na and Ni (Section~\ref{s:discoutlier}). The velocity and metallicity dispersions of Gru~I and Tri~II have not been decisive about whether they are ultra-faint dwarf galaxies or globular clusters, but we conclude they are both likely UFDs rather than GCs because both systems have extremely low neutron-capture element abundances (Section~\ref{s:discclassify}). We thus expect future observations of these systems to confirm metallicity spreads, as well as significant velocity dispersions or signs of tidal disruption. The low neutron-capture element abundances in UFDs reflect chemical enrichment at the the extreme low-mass end of galaxy formation in $\Lambda$CDM (Section~\ref{s:discwhy}): stochastic enrichment, metal loss in winds, and short star formation durations. The dissimilarity in neutron-capture elements also suggests that globular clusters and UFDs do not form in the same environments, and thus that globular clusters probably did not form in their own dark matter halos (Section~\ref{s:gc}). However, the nucleosynthetic origin of the low neutron-capture element abundances in UFDs like Gru~I and Tri~II is still an open question (Section~\ref{s:discncap}). \acknowledgments We thank Nidia Morrell for assisting with MIKE observations of Gru~I; Kristin Chiboucas and Lison Malo for assistance with GRACES and data reduction; Vini Placco for computing carbon corrections; and Projjwal Banerjee, Gabriele Cescutti, Anirudh Chiti, Brendan Griffen, Evan Kirby, Andrew McWilliam, Tony Piro, and \'Asa Sk\'ulad\'ottir for useful discussions. A.P.J. is supported by NASA through Hubble Fellowship grant HST-HF2-51393.001 awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS5-26555. J.D.S. and T.T.H. acknowledge support from the National Science Foundation under grant AST-1714873. A.F. acknowledges support from NSF grants AST-1255160 and AST-1716251. K.A.V. acknowledges funding from the National Science and Engineering Research Council of Canada (NSERC), funding reference number 327292-2006. Based on observations obtained with ESPaDOnS, located at the Canada-France-Hawaii Telescope (CFHT). CFHT is operated by the National Research Council of Canada, the Institut National des Sciences de l'Univers of the Centre National de la Recherche Scientique of France, and the University of Hawai'i. ESPaDOnS is a collaborative project funded by France (CNRS, MENESR, OMP, LATT), Canada (NSERC), CFHT and ESA. ESPaDOnS was remotely controlled from the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the National Research Council (Canada), CONICYT (Chile), Ministerio de Ciencia, Tecnología e Innovación Productiva (Argentina) and Ministério da Ciência, Tecnologia e Inovação (Brazil). This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France \citep{Simbad}, and NASA's Astrophysics Data System Bibliographic Services. \facilities{Magellan-Clay (MIKE, \citealt{Bernstein03}), Gemini-N (GRACES, \citealt{Chene14,Donati03})} \software{CarPy \citep{Kelson03}, OPERA \citep{Martioli12}, IRAF, MOOG \citep{Sneden73,Sobeck11}, SMH \citep{Casey14}, \texttt{numpy} \citep{numpy}, \texttt{scipy} \citep{scipy}, \texttt{matplotlib} \citep{matplotlib}, \texttt{pandas} \citep{pandas}, \texttt{seaborn}, \citep{seaborn}, \texttt{astropy} \citep{astropy}} \input{sperr}

2bf8fa12aa70d03088909d4e3fb45f98c74e3906

\section{Introduction} \label{sec:intro} As a longstanding, fundamental and challenging problem in computer vision, object detection (illustrated in Fig.~\ref{Fig:conferencekeywords}) has been an active area of research for several decades \cite{Fischler1973}. The goal of object detection is to determine whether there are any instances of objects from given categories (such as humans, cars, bicycles, dogs or cats) in an image and, if present, to return the spatial location and extent of each object instance (\emph{e.g.,} via a bounding box \cite{Everingham2010,Russakovsky2015}). As the cornerstone of image understanding and computer vision, object detection forms the basis for solving complex or high level vision tasks such as segmentation, scene understanding, object tracking, image captioning, event detection, and activity recognition. Object detection supports a wide range of applications, including robot vision, consumer electronics, security, autonomous driving, human computer interaction, content based image retrieval, intelligent video surveillance, and augmented reality. Recently, deep learning techniques \cite{Hinton2006Reducing,LeCun15} have emerged as powerful methods for learning feature representations automatically from data. In particular, these techniques have provided major improvements in object detection, as illustrated in Fig.~\ref{fig:GODResultsStatistics}. As illustrated in Fig.~\ref{fig:ObjectInstancevsCategory}, object detection can be grouped into one of two types \cite{Grauman2011Visual,Zhang13}: detection of specific instances versus the detection of broad categories. The first type aims to detect instances of a particular object (such as Donald Trump's face, the Eiffel Tower, or a neighbor's dog), essentially a matching problem. The goal of the second type is to detect (usually previously unseen) instances of some predefined object categories (for example humans, cars, bicycles, and dogs). Historically, much of the effort in the field of object detection has focused on the detection of a single category (typically faces and pedestrians) or a few specific categories. In contrast, over the past several years, the research community has started moving towards the more challenging goal of building general purpose object detection systems where the breadth of object detection ability rivals that of humans. \begin {figure}[!t] \centering \includegraphics[width=0.48\textwidth]{conferencekeywords.pdf} \caption{Most frequent keywords in ICCV and CVPR conference papers from 2016 to 2018. The size of each word is proportional to the frequency of that keyword. We can see that object detection has received significant attention in recent years.} \label{Fig:conferencekeywords} \end {figure} \begin {figure}[!t] \centering \includegraphics[width=0.45\textwidth]{ObjectInstancevsCategory.pdf} \caption{Object detection includes localizing instances of a {\em particular} object (top), as well as generalizing to detecting object {\em categories} in general (bottom). This survey focuses on recent advances for the latter problem of generic object detection.} \label{fig:ObjectInstancevsCategory} \end {figure} \begin {figure}[!t] \centering \includegraphics[width=0.5\textwidth]{GODResultsStatistics.pdf} \caption{An overview of recent object detection performance: We can observe a significant improvement in performance (measured as mean average precision) since the arrival of deep learning in 2012. (a) Detection results of winning entries in the VOC2007-2012 competitions, and (b) Top object detection competition results in ILSVRC2013-2017 (results in both panels use only the provided training data).} \label{fig:GODResultsStatistics} \end {figure} \begin {figure*}[!t] \centering \includegraphics[width=0.95\textwidth]{ADecadeOfObjectDetection.pdf} \caption{Milestones of object detection and recognition, including feature representations \cite{Csurka2004,Dalal2005HOG,He2016ResNet,Krizhevsky2012,Lazebnik2006SPM,Lowe1999Object,Lowe2004, Perronnin2010,Simonyan2014VGG,Sivic2003,GoogLeNet2015,Viola2001,HOGLBP2009}, detection frameworks \cite{Felzenszwalb2010b,Girshick2014RCNN,OverFeat2014,Uijlings2013b,Viola2001}, and datasets \cite{Everingham2010,Lin2014,Russakovsky2015}. The time period up to 2012 is dominated by handcrafted features, a transition took place in 2012 with the development of DCNNs for image classification by Krizhevsky \emph{et al.} \cite{Krizhevsky2012}, with methods after 2012 dominated by related deep networks. Mostof the listed methods are highly cited and won a major ICCV or CVPR prize. See Section~\ref{Sec:Progress} for details.} \label{fig:milestones} \end {figure*} In 2012, Krizhevsky \emph{et al.} \cite{Krizhevsky2012} proposed a Deep Convolutional Neural Network (DCNN) called AlexNet which achieved record breaking image classification accuracy in the Large Scale Visual Recognition Challenge (ILSVRC) \cite{Russakovsky2015}. Since that time, the research focus in most aspects of computer vision has been specifically on deep learning methods, indeed including the domain of generic object detection \cite{Girshick2014RCNN,He2014SPP,Girshick2015FRCNN,OverFeat2014,Ren2016a}. Although tremendous progress has been achieved, illustrated in Fig.~\ref{fig:GODResultsStatistics}, we are unaware of comprehensive surveys of this subject over the past five years. Given the exceptionally rapid rate of progress, this article attempts to track recent advances and summarize their achievements in order to gain a clearer picture of the current panorama in generic object detection. \begin{table*}[!t] \caption {Summary of related object detection surveys since 2000.}\label{Tab:Surveys} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{2pt} \resizebox*{16cm}{!}{ \begin{tabular}{!{\vrule width1.2bp}c|p{6cm}<{\centering}|c|c|c|p{9cm}<{\centering}!{\vrule width1.2bp}} \Xhline{1pt} \footnotesize No. & \footnotesize Survey Title & \footnotesize Ref. & \footnotesize Year & \footnotesize Venue & \footnotesize Content \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 1 }& \footnotesize Monocular Pedestrian Detection: Survey and Experiments &\raisebox{-1.5ex}[0pt]{ \footnotesize \cite{Enzweiler2009Monocular} } &\raisebox{-1.5ex}[0pt]{ \footnotesize 2009} & \raisebox{-1.5ex}[0pt]{\footnotesize PAMI} & \footnotesize An evaluation of three pedestrian detectors \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 2 }& \footnotesize Survey of Pedestrian Detection for Advanced Driver Assistance Systems & \raisebox{-1.5ex}[0pt]{\footnotesize \cite{Geronimo2010Survey}} &\raisebox{-1.5ex}[0pt]{ \footnotesize 2010 } & \raisebox{-1.5ex}[0pt]{\footnotesize PAMI }& \footnotesize \raisebox{-1.5ex}[0pt]{ A survey of pedestrian detection for advanced driver assistance systems} \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 3} & \footnotesize Pedestrian Detection: An Evaluation of the State of The Art & \raisebox{-1.5ex}[0pt]{\footnotesize \cite{Dollar2012Pedestrian} } & \raisebox{-1.5ex}[0pt]{\footnotesize 2012 }& \raisebox{-1.5ex}[0pt]{\footnotesize PAMI } & \footnotesize A thorough and detailed evaluation of detectors in monocular images \\ \hline \footnotesize 4 & \footnotesize Detecting Faces in Images: A Survey & \footnotesize \cite{Yang2002b} & \footnotesize 2002 & \footnotesize PAMI & \footnotesize First survey of face detection from a single image \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 5 } & \footnotesize A Survey on Face Detection in the Wild: Past, Present and Future & \raisebox{-1.5ex}[0pt]{\footnotesize \cite{Zafeiriou2015}} & \raisebox{-1.5ex}[0pt]{\footnotesize 2015 }& \raisebox{-1.5ex}[0pt]{ \footnotesize CVIU}& \raisebox{-1.5ex}[0pt]{ \footnotesize A survey of face detection in the wild since 2000} \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 6 }& \raisebox{-1.5ex}[0pt]{ \footnotesize On Road Vehicle Detection: A Review} &\raisebox{-1.5ex}[0pt]{ \footnotesize \cite{Sun2006Road} }&\raisebox{-1.5ex}[0pt]{ \footnotesize 2006 }& \raisebox{-1.5ex}[0pt]{ \footnotesize PAMI} & \footnotesize A review of vision based on-road vehicle detection systems \\ \hline \footnotesize 7 & \footnotesize Text Detection and Recognition in Imagery: A Survey & \footnotesize \cite{Ye2015Text} & \footnotesize 2015 & \footnotesize PAMI & \footnotesize A survey of text detection and recognition in color imagery \\ \hline \raisebox{-1.5ex}[0pt]{ \footnotesize 8 }&\raisebox{-1.5ex}[0pt]{ \footnotesize Toward Category Level Object Recognition} & \raisebox{-1.5ex}[0pt]{\footnotesize \cite{Ponce2007Toward}} &\raisebox{-1.5ex}[0pt]{ \footnotesize 2007} &\raisebox{-1.5ex}[0pt]{ \footnotesize Book } & \footnotesize Representative papers on object categorization, detection, and segmentation \\ \hline \raisebox{-1.5ex}[0pt]{ \footnotesize 9 }& \footnotesize The Evolution of Object Categorization and the Challenge of Image Abstraction & \raisebox{-1.5ex}[0pt]{\footnotesize \cite{Dickinson2009}} &\raisebox{-1.5ex}[0pt]{ \footnotesize 2009}&\raisebox{-1.5ex}[0pt]{ \footnotesize Book } & \raisebox{-1.5ex}[0pt]{ \footnotesize A trace of the evolution of object categorization over four decades} \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 10} & \footnotesize Context based Object Categorization: A Critical Survey &\raisebox{-1.5ex}[0pt]{ \footnotesize \cite{Galleguillos2010}} &\raisebox{-1.5ex}[0pt]{ \footnotesize 2010} & \raisebox{-1.5ex}[0pt]{\footnotesize CVIU} & \footnotesize A review of contextual information for object categorization \\ \hline \footnotesize 11 & \footnotesize 50 Years of Object Recognition: Directions Forward & \footnotesize \cite{Andreopoulos13} & \footnotesize 2013& \footnotesize CVIU & \footnotesize A review of the evolution of object recognition systems over five decades \\ \hline \raisebox{-1.5ex}[0pt]{ \footnotesize 12} & \raisebox{-1.5ex}[0pt]{ \footnotesize Visual Object Recognition} &\raisebox{-1.5ex}[0pt]{ \footnotesize \cite{Grauman2011Visual} } &\raisebox{-1.5ex}[0pt]{ \footnotesize2011}&\raisebox{-1.5ex}[0pt]{ \footnotesize Tutorial} & \footnotesize Instance and category object recognition techniques \\ \hline \footnotesize 13 & \footnotesize Object Class Detection: A Survey & \footnotesize \cite{Zhang13} & \footnotesize 2013 & \footnotesize ACM CS & \footnotesize Survey of generic object detection methods before 2011 \\ \hline \raisebox{-1.5ex}[0pt]{ \footnotesize 14} & \footnotesize Feature Representation for Statistical Learning based Object Detection: A Review &\raisebox{-1.5ex}[0pt]{ \footnotesize \cite{Li2015Feature}} &\raisebox{-1.5ex}[0pt]{ \footnotesize 2015} & \raisebox{-1.5ex}[0pt]{ \footnotesize PR} & \footnotesize Feature representation methods in statistical learning based object detection, including handcrafted and deep learning based features \\ \hline \footnotesize 15 & \footnotesize Salient Object Detection: A Survey & \footnotesize \cite{Borji14} & \footnotesize2014 & \footnotesize arXiv & \footnotesize A survey for salient object detection \\ \hline \raisebox{-1.5ex}[0pt]{ \footnotesize 16}& \footnotesize Representation Learning: A Review and New Perspectives & \raisebox{-1.5ex}[0pt]{ \footnotesize \cite{Bengio13Feature}} &\raisebox{-1.5ex}[0pt]{ \footnotesize 2013} &\raisebox{-1.5ex}[0pt]{ \footnotesize PAMI} & \footnotesize Unsupervised feature learning and deep learning, probabilistic models, autoencoders, manifold learning, and deep networks \\ \hline \footnotesize 17 & \footnotesize Deep Learning & \footnotesize \cite{LeCun15} & \footnotesize2015& \footnotesize Nature & \footnotesize An introduction to deep learning and applications \\ \hline \raisebox{-1.5ex}[0pt]{ \footnotesize 18 }& \footnotesize A Survey on Deep Learning in Medical Image Analysis& \raisebox{-1.5ex}[0pt]{\footnotesize \cite{Litjens2017} }&\raisebox{-1.5ex}[0pt]{ \footnotesize 2017} &\raisebox{-1.5ex}[0pt]{ \footnotesize MIA} & \footnotesize A survey of deep learning for image classification, object detection, segmentation and registration in medical image analysis \\ \hline \raisebox{-1.5ex}[0pt]{ \footnotesize 19} & \footnotesize Recent Advances in Convolutional Neural Networks&\raisebox{-1.5ex}[0pt]{ \footnotesize \cite{Gu2015Recent}} &\raisebox{-1.5ex}[0pt]{ \footnotesize 2017}& \raisebox{-1.5ex}[0pt]{\footnotesize PR}& \footnotesize A broad survey of the recent advances in CNN and its applications in computer vision, speech and natural language processing \\ \hline \footnotesize 20 & \footnotesize Tutorial: Tools for Efficient Object Detection & \footnotesize $-$ & \footnotesize 2015& \footnotesize ICCV15& \footnotesize A short course for object detection only covering recent milestones \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 21 }&\raisebox{-1.5ex}[0pt]{ \footnotesize Tutorial: Deep Learning for Objects and Scenes} &\raisebox{-1.5ex}[0pt]{ \footnotesize $-$} & \raisebox{-1.5ex}[0pt]{ \footnotesize 2017}&\raisebox{-1.5ex}[0pt]{ \footnotesize CVPR17} & \footnotesize A high level summary of recent work on deep learning for visual recognition of objects and scenes \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 22} &\raisebox{-1.5ex}[0pt]{ \footnotesize Tutorial: Instance Level Recognition}&\raisebox{-1.5ex}[0pt]{ \footnotesize $-$ }& \raisebox{-1.5ex}[0pt]{ \footnotesize 2017} &\raisebox{-1.5ex}[0pt]{ \footnotesize ICCV17} & \footnotesize A short course of recent advances on instance level recognition, including object detection, instance segmentation and human pose prediction \\ \hline \raisebox{-1.5ex}[0pt]{\footnotesize 23}&\raisebox{-1.5ex}[0pt]{ \footnotesize Tutorial: Visual Recognition and Beyond} &\raisebox{-1.5ex}[0pt]{ \footnotesize $-$} & \raisebox{-1.5ex}[0pt]{\footnotesize 2018}&\raisebox{-1.5ex}[0pt]{ \footnotesize CVPR18 }& \footnotesize A tutorial on methods and principles behind image classification, object detection, instance segmentation, and semantic segmentation. \\ \hline \footnotesize \textbf{24}& \footnotesize \textbf{Deep Learning for Generic Object Detection} & \footnotesize \textbf{Ours} & \footnotesize \textbf{2019} & \footnotesize \textbf{VISI} & \footnotesize \textbf{A comprehensive survey of deep learning for generic object detection} \\ \Xhline{1pt} \end{tabular} } \end{table*} \subsection{Comparison with Previous Reviews} Many notable object detection surveys have been published, as summarized in Table~\ref{Tab:Surveys}. These include many excellent surveys on the problem of {\em specific} object detection, such as pedestrian detection \cite{Enzweiler2009Monocular,Geronimo2010Survey,Dollar2012Pedestrian}, face detection \cite{Yang2002b,Zafeiriou2015}, vehicle detection \cite{Sun2006Road} and text detection \cite{Ye2015Text}. There are comparatively few recent surveys focusing directly on the problem of generic object detection, except for the work by Zhang \emph{et al.} \cite{Zhang13} who conducted a survey on the topic of object class detection. However, the research reviewed in \cite{Grauman2011Visual}, \cite{Andreopoulos13} and \cite{Zhang13} is mostly pre-2012, and therefore prior to the recent striking success and dominance of deep learning and related methods. Deep learning allows computational models to learn fantastically complex, subtle, and abstract representations, driving significant progress in a broad range of problems such as visual recognition, object detection, speech recognition, natural language processing, medical image analysis, drug discovery and genomics. Among different types of deep neural networks, DCNNs \cite{LeCun1998Gradient,Krizhevsky2012,LeCun15} have brought about breakthroughs in processing images, video, speech and audio. To be sure, there have been many published surveys on deep learning, including that of Bengio \emph{et al.} \cite{Bengio13Feature}, LeCun \emph{et al.} \cite{LeCun15}, Litjens \emph{et al.} \cite{Litjens2017}, Gu \emph{et al.} \cite{Gu2015Recent}, and more recently in tutorials at ICCV and CVPR. In contrast, although many deep learning based methods have been proposed for object detection, we are unaware of any comprehensive recent survey. A thorough review and summary of existing work is essential for further progress in object detection, particularly for researchers wishing to enter the field. Since our focus is on {\em generic} object detection, the extensive work on DCNNs for {\em specific} object detection, such as face detection \cite{Li2015CasecadeCNN,Zhang2016Joint,Hu2017Finding}, pedestrian detection \cite{Zhang2016faster,Hosang2015taking}, vehicle detection \cite{Zhou2016dave} and traffic sign detection \cite{Zhu2016traffic} will not be considered. \begin {figure}[!t] \centering \includegraphics[width=0.45\textwidth]{TheProblem.pdf} \caption{Recognition problems related to generic object detection: (a) Image level object classification, (b) Bounding box level generic object detection, (c) Pixel-wise semantic segmentation, (d) Instance level semantic segmentation.} \label{Fig:TheProblem} \end {figure} \subsection{Scope} The number of papers on generic object detection based on deep learning is breathtaking. There are so many, in fact, that compiling any comprehensive review of the state of the art is beyond the scope of any reasonable length paper. As a result, it is necessary to establish selection criteria, in such a way that we have limited our focus to top journal and conference papers. Due to these limitations, we sincerely apologize to those authors whose works are not included in this paper. For surveys of work on related topics, readers are referred to the articles in Table~\ref{Tab:Surveys}. This survey focuses on major progress of the last five years, and we restrict our attention to still pictures, leaving the important subject of video object detection as a topic for separate consideration in the future. The main goal of this paper is to offer a comprehensive survey of deep learning based generic object detection techniques, and to present some degree of taxonomy, a high level perspective and organization, primarily on the basis of popular datasets, evaluation metrics, context modeling, and detection proposal methods. The intention is that our categorization be helpful for readers to have an accessible understanding of similarities and differences between a wide variety of strategies. The proposed taxonomy gives researchers a framework to understand current research and to identify open challenges for future research. The remainder of this paper is organized as follows. Related background and the progress made during the last two decades are summarized in Section~\ref{Sec:Background}. A brief introduction to deep learning is given in Section \ref{Sec:CNNintro}. Popular datasets and evaluation criteria are summarized in Section \ref{Sec:DataEval}. We describe the milestone object detection frameworks in Section~\ref{Sec:Frameworks}. From Section \ref{Sec:DCNNFeatures} to Section \ref{sec:otherissue}, fundamental sub-problems and the relevant issues involved in designing object detectors are discussed. Finally, in Section \ref{Sec:Conclusions}, we conclude the paper with an overall discussion of object detection, state-of-the- art performance, and future research directions. \section{Generic Object Detection} \label{Sec:Background} \subsection{The Problem} \label{Sec:TheProblem} \emph{Generic object detection}, also called generic object category detection, object class detection, or object category detection \cite{Zhang13}, is defined as follows. Given an image, determine whether or not there are instances of objects from predefined categories (usually \emph{many} categories, \emph{e.g.,} 200 categories in the ILSVRC object detection challenge) and, if present, to return the spatial location and extent of each instance. A greater emphasis is placed on detecting a broad range of natural categories, as opposed to specific object category detection where only a narrower predefined category of interest (\emph{e.g.,} faces, pedestrians, or cars) may be present. Although thousands of objects occupy the visual world in which we live, currently the research community is primarily interested in the localization of highly structured objects (\emph{e.g.,} cars, faces, bicycles and airplanes) and articulated objects (\emph{e.g.,} humans, cows and horses) rather than unstructured scenes (such as sky, grass and cloud). The spatial location and extent of an object can be defined coarsely using a bounding box (an axis-aligned rectangle tightly bounding the object) \cite{Everingham2010,Russakovsky2015}, a precise pixelwise segmentation mask \cite{Zhang13}, or a closed boundary \cite{Lin2014,Russell2008}, as illustrated in Fig.~\ref{Fig:TheProblem}. To the best of our knowledge, for the evaluation of generic object detection algorithms, it is bounding boxes which are most widely used in the current literature \cite{Everingham2010,Russakovsky2015}, and therefore this is also the approach we adopt in this survey. However, as the research community moves towards deeper scene understanding (from image level object classification to single object localization, to generic object detection, and to pixelwise object segmentation), it is anticipated that future challenges will be at the pixel level \cite{Lin2014}. There are many problems closely related to that of generic object detection\footnote{To the best of our knowledge, there is no universal agreement in the literature on the definitions of various vision subtasks. Terms such as detection, localization, recognition, classification, categorization, verification, identification, annotation, labeling, and understanding are often differently defined \cite{Andreopoulos13}.}. The goal of \emph{object classification} or \emph{object categorization} (Fig.~\ref{Fig:TheProblem} (a)) is to assess the presence of objects from a given set of object classes in an image; \emph{i.e.,} assigning one or more object class labels to a given image, determining the presence without the need of location. The additional requirement to locate the instances in an image makes detection a more challenging task than classification. The \emph{object recognition} problem denotes the more general problem of identifying/localizing all the objects present in an image, subsuming the problems of object detection and classification \cite{Everingham2010,Russakovsky2015,Opelt2006generic,Andreopoulos13}. Generic object detection is closely related to \emph{semantic image segmentation} (Fig.~\ref{Fig:TheProblem} (c)), which aims to assign each pixel in an image to a semantic class label. \emph{Object instance segmentation} (Fig.~\ref{Fig:TheProblem} (d)) aims to distinguish different instances of the same object class, as opposed to semantic segmentation which does not. \begin {figure}[!t] \centering \includegraphics[width=0.5\textwidth]{challenges.pdf} \caption{Taxonomy of challenges in generic object detection.} \label{Fig:challenges} \end {figure} \begin {figure}[!t] \centering \includegraphics[width=0.5\textwidth]{IntraInterClass.pdf} \caption{\footnotesize{Changes in appearance of the same class with variations in imaging conditions (a-h). There is an astonishing variation in what is meant to be a single object class (i). In contrast, the four images in (j) appear very similar, but in fact are from four {\emph different} object classes. Most images are from ImageNet \cite{Russakovsky2015} and MS COCO \cite{Lin2014}.}} \label{Fig:IntraInterClass} \end {figure} \subsection{Main Challenges} \label{Sec:MainChallenges} The ideal of generic object detection is to develop a general-purpose algorithm that achieves two competing goals of \emph{high quality/accuracy} and \emph{high efficiency} (Fig.~\ref{Fig:challenges}). As illustrated in Fig.~\ref{Fig:IntraInterClass}, high quality detection must accurately localize and recognize objects in images or video frames, such that the large variety of object categories in the real world can be distinguished (\emph{i.e.,} high distinctiveness), and that object instances from the same category, subject to intra-class appearance variations, can be localized and recognized (\emph{i.e.,} high robustness). High efficiency requires that the entire detection task runs in real time with acceptable memory and storage demands. \subsubsection{Accuracy related challenges} Challenges in detection accuracy stem from 1) the vast range of intra-class variations and 2) the huge number of object categories. Intra-class variations can be divided into two types: intrinsic factors and imaging conditions. In terms of intrinsic factors, each object category can have many different object instances, possibly varying in one or more of color, texture, material, shape, and size, such as the ``chair'' category shown in Fig.~\ref{Fig:IntraInterClass} (\emph{i}). Even in a more narrowly defined class, such as human or horse, object instances can appear in different poses, subject to nonrigid deformations or with the addition of clothing. Imaging condition variations are caused by the dramatic impacts unconstrained environments can have on object appearance, such as lighting (dawn, day, dusk, indoors), physical location, weather conditions, cameras, backgrounds, illuminations, occlusion, and viewing distances. All of these conditions produce significant variations in object appearance, such as illumination, pose, scale, occlusion, clutter, shading, blur and motion, with examples illustrated in Fig.~\ref{Fig:IntraInterClass} (\emph{a}-\emph{h}). Further challenges may be added by digitization artifacts, noise corruption, poor resolution, and filtering distortions. In addition to \emph{intra}class variations, the large number of object categories, on the order of $10^4-10^5$, demands great discrimination power from the detector to distinguish between subtly different \emph{inter}class variations, as illustrated in Fig.~\ref{Fig:IntraInterClass} (j). In practice, current detectors focus mainly on structured object categories, such as the 20, 200 and 91 object classes in PASCAL VOC \cite{Everingham2010}, ILSVRC \cite{Russakovsky2015} and MS COCO \cite{Lin2014} respectively. Clearly, the number of object categories under consideration in existing benchmark datasets is much smaller than can be recognized by humans. \subsubsection{Efficiency and scalability related challenges} The prevalence of social media networks and mobile/wearable devices has led to increasing demands for analyzing visual data. However, mobile/wearable devices have limited computational capabilities and storage space, making efficient object detection critical. The efficiency challenges stem from the need to localize and recognize, computational complexity growing with the (possibly large) number of object categories, and with the (possibly very large) number of locations and scales within a single image, such as the examples in Fig.~\ref{Fig:IntraInterClass} (c, d). A further challenge is that of scalability: A detector should be able to handle previously unseen objects, unknown situations, and high data rates. As the number of images and the number of categories continue to grow, it may become impossible to annotate them manually, forcing a reliance on weakly supervised strategies. \begin {figure*}[!t] \centering \includegraphics[width=0.45\textwidth]{ConvReLuMax.pdf} \includegraphics[width=0.45\textwidth]{VGGNet.pdf} \caption{\textcolor{black}{(a) Illustration of three operations that are repeatedly applied by a typical CNN: Convolution with a number of linear filters; Nonlinearities (\emph{e.g.} ReLU); and Local pooling (\emph{e.g.} Max Pooling). The $M$ feature maps from a previous layer are convolved with $N$ different filters (here shown as size $3\times3\times M$), using a stride of 1. The resulting $N$ feature maps are then passed through a nonlinear function (\emph{e.g.} ReLU), and pooled (\emph{e.g.} taking a maximum over $2\times2$ regions) to give $N$ feature maps at a reduced resolution. (b) Illustration of the architecture of VGGNet \cite{Simonyan2014VGG}, a typical CNN with 11 weight layers. An image with 3 color channels is presented as the input. The network has 8 convolutional layers, 3 fully connected layers, 5 max pooling layers and a softmax classification layer. The last three fully connected layers take features from the top convolutional layer as input in vector form. The final layer is a $C$-way softmax function, $C$ being the number of classes. The whole network can be learned from labeled training data by optimizing an objective function (\emph{e.g.} mean squared error or cross entropy loss) via Stochastic Gradient Descent.}} \label{fig:ConvReLuMax} \end {figure*} \subsection{Progress in the Past Two Decades} \label{Sec:Progress} Early research on object recognition was based on template matching techniques and simple part-based models \cite{Fischler1973}, focusing on specific objects whose spatial layouts are roughly rigid, such as faces. Before 1990 the leading paradigm of object recognition was based on geometric representations~\cite{Mundy2006Object,Ponce2007Toward}, with the focus later moving away from geometry and prior models towards the use of statistical classifiers (such as Neural Networks \cite{Rowley1998}, SVM \cite{Osuna1997Train} and Adaboost \cite{Viola2001,Xiao2003Boosting}) based on appearance features \cite{Murase1995,Schmid1997Local}. This successful family of object detectors set the stage for most subsequent research in this field. The milestones of object detection in more recent years are presented in Fig.~\ref{fig:milestones}, in which two main eras (SIFT\emph{ vs.} DCNN) are highlighted. The appearance features moved from global representations \cite{Murase1995Visual,Swain1991Color,Turk1991Face} to local representations that are designed to be invariant to changes in translation, scale, rotation, illumination, viewpoint and occlusion. Handcrafted local invariant features gained tremendous popularity, starting from the Scale Invariant Feature Transform (SIFT) feature \cite{Lowe1999Object}, and the progress on various visual recognition tasks was based substantially on the use of local descriptors \cite{Mikolajczyk2005} such as Haar-like features \cite{Viola2001}, SIFT \cite{Lowe2004}, Shape Contexts \cite{Belongie2002shape}, Histogram of Gradients (HOG) \cite{Dalal2005HOG} Local Binary Patterns (LBP) \cite{Ojala02}, and region covariances \cite{Tuzel2006Region}. These local features are usually aggregated by simple concatenation or feature pooling encoders such as the Bag of Visual Words approach, introduced by Sivic and Zisserman \cite{Sivic2003} and Csurka \emph{et al.} \cite{Csurka2004}, Spatial Pyramid Matching (SPM) of BoW models \cite{Lazebnik2006SPM}, and Fisher Vectors \cite{Perronnin2010}. For years, the multistage hand tuned pipelines of handcrafted local descriptors and discriminative classifiers dominated a variety of domains in computer vision, including object detection, until the significant turning point in 2012 when DCNNs \cite{Krizhevsky2012} achieved their record-breaking results in image classification. The use of CNNs for detection and localization \cite{Rowley1998} can be traced back to the 1990s, with a modest number of hidden layers used for object detection \cite{Vaillant1994,Rowley1998,Sermanet2013c}, successful in restricted domains such as face detection. However, more recently, deeper CNNs have led to record-breaking improvements in the detection of more general object categories, a shift which came about when the successful application of DCNNs in image classification \cite{Krizhevsky2012} was transferred to object detection, resulting in the milestone Region-based CNN (RCNN) detector of Girshick \emph{et al.} \cite{Girshick2014RCNN}. The successes of deep detectors rely heavily on vast training data and large networks with millions or even billions of parameters. The availability of GPUs with very high computational capability and large-scale detection datasets (such as ImageNet \cite{ImageNet2009,Russakovsky2015} and MS COCO \cite{Lin2014}) play a key role in their success. Large datasets have allowed researchers to target more realistic and complex problems from images with large intra-class variations and inter-class similarities \cite{Lin2014,Russakovsky2015}. However, accurate annotations are labor intensive to obtain, so detectors must consider methods that can relieve annotation difficulties or can learn with smaller training datasets. The research community has started moving towards the challenging goal of building general purpose object detection systems whose ability to detect many object categories matches that of humans. This is a major challenge: according to cognitive scientists, human beings can identify around 3,000 entry level categories and 30,000 visual categories overall, and the number of categories distinguishable with domain expertise may be to the order of $10^5$ \cite{Biederman1987}. Despite the remarkable progress of the past years, designing an accurate, robust, efficient detection and recognition system that approaches human-level performance on $10^4-10^5$ categories is undoubtedly an unresolved problem. \section{A Brief Introduction to Deep Learning} \label{Sec:CNNintro} \textcolor{black}{Deep learning has revolutionized a wide range of machine learning tasks, from image classification and video processing to speech recognition and natural language understanding. Given this tremendously rapid evolution, there exist many recent survey papers on deep learning \cite{Bengio13Feature,Goodfellow2016Deep,Gu2015Recent,LeCun15, Litjens2017,Pouyanfar2018Survey, Wu2019Comprehensive,Young2018Recent, Zhang2018Deep,Zhou2018Graph,Zhu2017Deep}. These surveys have reviewed deep learning techniques from different perspectives \cite{Bengio13Feature,Goodfellow2016Deep,Gu2015Recent, LeCun15,Pouyanfar2018Survey,Wu2019Comprehensive,Zhou2018Graph}, or with applications to medical image analysis \cite{Litjens2017}, natural language processing \cite{Young2018Recent}, speech recognition systems \cite{Zhang2018Deep}, and remote sensing \cite{Zhu2017Deep}.} Convolutional Neural Networks (CNNs), the most representative models of deep learning, are able to exploit the basic properties underlying natural signals: translation invariance, local connectivity, and compositional hierarchies \cite{LeCun15}. A typical CNN, illustrated in Fig.~\ref{fig:ConvReLuMax}, has a hierarchical structure and is composed of a number of layers to learn representations of data with multiple levels of abstraction \cite{LeCun15}. We begin with a convolution \begin{equation} \textbf{\emph{x}}^{l-1} * \textbf{\emph{w}}^{l} \end{equation} between an input feature map $\textbf{\emph{x}}^{l-1}$ at a feature map from previous layer $l-1$, convolved with a 2D convolutional kernel (or filter or weights) $\textbf{\emph{w}}^{l}$. This convolution appears over a sequence of layers, subject to a nonlinear operation $\sigma$, such that \begin{equation} \textbf{\emph{x}}^l_j = \sigma(\sum_{i=1}^{N^{l-1}} \textbf{\emph{x}}^{l-1}_i * \textbf{\emph{w}}^{l}_{i, j} +b^{l}_j), \label{eq:conv} \end{equation} with a convolution now between the $N^{l-1}$ input feature maps $\textbf{\emph{x}}^{l-1}_i$ and the corresponding kernel $\textbf{\emph{w}}^{l}_{i, j}$, plus a bias term $b^{l}_j$. The elementwise nonlinear function $\sigma(\cdot)$ is typically a rectified linear unit (ReLU) for each element, \begin{equation} \sigma(x) = \max\{x, 0\}. \end{equation} Finally, pooling corresponds to the downsampling/upsampling of feature maps. These three operations (convolution, nonlinearity, pooling) are illustrated in Fig. \ref{fig:ConvReLuMax} (a); CNNs having a large number of layers, a ``deep'' network, are referred to as Deep CNNs (DCNNs), with a typical DCNN architecture illustrated in Fig.~\ref{fig:ConvReLuMax} (b). Most layers of a CNN consist of a number of feature maps, within which each pixel acts like a neuron. Each neuron in a convolutional layer is connected to feature maps of the previous layer through a set of weights $\textbf{\emph{w}}_{i,j}$ (essentially a set of 2D filters). As can be seen in Fig.~\ref{fig:ConvReLuMax} (b), where the early CNN layers are typically composed of convolutional and pooling layers, the later layers are normally fully connected. From earlier to later layers, the input image is repeatedly convolved, and with each layer, the receptive field or region of support increases. In general, the initial CNN layers extract low-level features (\emph{e.g.,} edges), with later layers extracting more general features of increasing complexity \cite{ZeilerFergus2014,Bengio13Feature,LeCun15,Oquab2014Learning}. DCNNs have a number of outstanding advantages: a hierarchical structure to learn representations of data with multiple levels of abstraction, the capacity to learn very complex functions, and learning feature representations directly and automatically from data with minimal domain knowledge. What has particularly made DCNNs successful has been the availability of large scale labeled datasets and of GPUs with very high computational capability. Despite the great successes, known deficiencies remain. In particular, there is an extreme need for labeled training data and a requirement of expensive computing resources, and considerable skill and experience are still needed to select appropriate learning parameters and network architectures. Trained networks are poorly interpretable, there is a lack of robustness to degradations, and many DCNNs have shown serious vulnerability to attacks \cite{Goodfellow2015Explaining}, all of which currently limit the use of DCNNs in real-world applications. \begin{table}[!t] \caption{Most frequent object classes for each detection challenge. The size of each word is proportional to the frequency of that class in the training dataset.} \centering \setlength\tabcolsep{2pt} \begin{tabular}{!{\vrule width1.2bp}c!{\vrule width1.2bp}c!{\vrule width1.2bp}} \Xhline{1.2pt} \includegraphics[width=0.24\textwidth]{ClassFreqpascal2012_trainval_first.pdf}& \includegraphics[width=0.24\textwidth]{ClassFreqcoco2017_trainval_first.pdf}\\ \Xhline{1.2pt} \multicolumn{1}{c}{(a) PASCAL VOC (20 Classes) }& \multicolumn{1}{c}{(b) MS COCO (80 Classes)}\\ \multicolumn{1}{c}{} \\ \Xhline{1.2pt} \multicolumn{2}{!{\vrule width1.2bp}c!{\vrule width1.2bp}}{\includegraphics[width=0.48\textwidth]{ClassFreqImageNet_trainval.pdf}}\\ \Xhline{1.2pt} \multicolumn{2}{c}{(c) ILSVRC (200 Classes)}\\ \multicolumn{1}{c}{} \\ \Xhline{1.2pt} \multicolumn{2}{!{\vrule width1.2bp}c!{\vrule width1.2bp}}{\includegraphics[width=0.48\textwidth]{ClassFreqopen2018_train_first.pdf}}\\ \Xhline{1.2pt} \multicolumn{2}{c}{(d) Open Images Detection Challenge (500 Classes)} \end{tabular} \label{fig:ClassFrequency} \end {table} \section{Datasets and Performance Evaluation} \label{Sec:DataEval} \subsection{Datasets} \label{sec:datasets} Datasets have played a key role throughout the history of object recognition research, not only as a common ground for measuring and comparing the performance of competing algorithms, but also pushing the field towards increasingly complex and challenging problems. In particular, recently, deep learning techniques have brought tremendous success to many visual recognition problems, and it is the large amounts of annotated data which play a key role in their success. Access to large numbers of images on the Internet makes it possible to build comprehensive datasets in order to capture a vast richness and diversity of objects, enabling unprecedented performance in object recognition. For generic object detection, there are four famous datasets: PASCAL VOC \cite{Everingham2010,Everingham2015}, ImageNet \cite{ImageNet2009}, MS COCO \cite{Lin2014} and Open Images \cite{Kuznetsova2018Open}. The attributes of these datasets are summarized in Table~\ref{Tab:maindatasets}, and selected sample images are shown in Fig.~\ref{fig:ObjectImages}. There are three steps to creating large-scale annotated datasets: determining the set of target object categories, collecting a diverse set of candidate images to represent the selected categories on the Internet, and annotating the collected images, typically by designing crowdsourcing strategies. Recognizing space limitations, we refer interested readers to the original papers \cite{Everingham2010,Everingham2015,Lin2014,Russakovsky2015,Kuznetsova2018Open} for detailed descriptions of these datasets in terms of construction and properties. \begin{table*}[!t] \caption {Popular databases for object recognition. Example images from PASCAL VOC, ImageNet, MS COCO and Open Images are shown in Fig.~\ref{fig:ObjectImages}.}\label{Tab:maindatasets} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{1pt} \resizebox*{18cm}{!}{ \begin{tabular}{!{\vrule width1.2bp}c|c|c|c|c|c|c|p{8cm}!{\vrule width1.2bp}} \Xhline{1pt} \scriptsize \shortstack [c] {\textbf{Dataset}\\ \textbf{Name}} & \scriptsize \shortstack [c] {\textbf{Total} \\ \textbf{Images}} & \scriptsize \shortstack [c] {\textbf{Categories}} & \scriptsize \shortstack [c] {\textbf{Images Per} \\ \textbf{Category}} & \scriptsize \shortstack [c] {\textbf{Objects Per }\\ \textbf{Image}} & \scriptsize \shortstack [c] {\textbf{Image} \\ \textbf{Size} } & \scriptsize \shortstack [c] {\textbf{Started} \\ \textbf{Year}} &\raisebox{1.3ex}[0pt]{ \scriptsize \shortstack [c] {$\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad$\textbf{Highlights}}} \\ \hline \raisebox{-6.3ex}[0pt]{\scriptsize \shortstack [c] {PASCAL \\VOC \\ (2012) \cite{Everingham2015} } } &\raisebox{-5ex}[0pt]{\scriptsize $11,540$ }& \raisebox{-5ex}[0pt]{\scriptsize $20$} & \raisebox{-5ex}[0pt]{ \scriptsize $303\sim4087$}& \raisebox{-5ex}[0pt]{\scriptsize $2.4$} & \raisebox{-5ex}[0pt]{\scriptsize $470\times380$}& \raisebox{-5ex}[0pt]{ \scriptsize $2005$} & \scriptsize Covers only 20 categories that are common in everyday life; Large number of training images; Close to real-world applications; Significantly larger intraclass variations; Objects in scene context; Multiple objects in one image; Contains many difficult samples. \\ \hline \raisebox{-3.3ex}[0pt]{\scriptsize ImageNet \cite{Russakovsky2015}} &\raisebox{-4.3ex}[0pt]{\shortstack [c] {14 \\ millions+} } &\raisebox{-3.3ex}[0pt]{ \scriptsize $21,841$} &\raisebox{-3.3ex}[0pt]{ \scriptsize $-$} &\raisebox{-3.3ex}[0pt]{ \scriptsize $1.5$} & \raisebox{-3.3ex}[0pt]{\scriptsize $500\times400$}&\raisebox{-3.3ex}[0pt]{ \scriptsize $2009$} & \scriptsize Large number of object categories; More instances and more categories of objects per image; More challenging than PASCAL VOC; Backbone of the ILSVRC challenge; Images are object-centric. \\ \hline \raisebox{-3.3ex}[0pt]{\scriptsize MS COCO \cite{Lin2014}} & \raisebox{-3.3ex}[0pt]{\scriptsize $328,000+$ } & \raisebox{-3.3ex}[0pt]{\scriptsize $91$} & \raisebox{-3.3ex}[0pt]{ \scriptsize $-$} & \raisebox{-3.3ex}[0pt]{ \scriptsize $7.3$} & \raisebox{-3.3ex}[0pt]{ \scriptsize $640\times480$}& \raisebox{-3.3ex}[0pt]{\scriptsize $2014$} & \scriptsize Even closer to real world scenarios; Each image contains more instances of objects and richer object annotation information; Contains object segmentation notation data that is not available in the ImageNet dataset. \\ \hline \raisebox{-1.3ex}[0pt]{\scriptsize Places \cite{Zhou2017Places}} &\raisebox{-2.3ex}[0pt]{\shortstack [c] {10 \\ millions+} } &\raisebox{-1.3ex}[0pt]{ \scriptsize $434$ }&\raisebox{-1.3ex}[0pt]{ \scriptsize $-$}&\raisebox{-1.3ex}[0pt]{ \scriptsize $-$} & \raisebox{-1.3ex}[0pt]{ \scriptsize $256\times256$}&\raisebox{-1.3ex}[0pt]{ \scriptsize $2014$} & \scriptsize The largest labeled dataset for scene recognition; Four subsets Places365 Standard, Places365 Challenge, Places 205 and Places88 as benchmarks. \\ \hline \raisebox{-3.3ex}[0pt]{\scriptsize Open Images \cite{Kuznetsova2018Open}} &\raisebox{-4.3ex}[0pt]{\shortstack [c] {9 \\ millions+} } &\raisebox{-3.3ex}[0pt]{ \scriptsize $6000$+ }&\raisebox{-3.3ex}[0pt]{ \scriptsize $-$}&\raisebox{-3.3ex}[0pt]{ \scriptsize $8.3$} & \raisebox{-3.3ex}[0pt]{ \scriptsize varied}&\raisebox{-3.3ex}[0pt]{ \scriptsize $2017$} & \scriptsize Annotated with image level labels, object bounding boxes and visual relationships; Open Images V5 supports large scale object detection, object instance segmentation and visual relationship detection. \\ \Xhline{1pt} \end{tabular} } \end{table*} \begin {figure*}[!t] \centering \includegraphics[width=0.98\textwidth]{ObjectImages.pdf} \caption{Some example images with object annotations from PASCAL VOC, ILSVRC, MS COCO and Open Images. See Table \ref{Tab:maindatasets} for a summary of these datasets.} \label{fig:ObjectImages} \end {figure*} The four datasets form the backbone of their respective detection challenges. Each challenge consists of a publicly available dataset of images together with ground truth annotation and standardized evaluation software, and an annual competition and corresponding workshop. Statistics for the number of images and object instances in the training, validation and testing datasets\footnote{The annotations on the test set are not publicly released, except for PASCAL VOC2007.} for the detection challenges are given in Table~\ref{Tab:detdatasets}. The most frequent object classes in VOC, COCO, ILSVRC and Open Images detection datasets are visualized in Table \ref{fig:ClassFrequency}. \textbf{PASCAL VOC} \cite{Everingham2010,Everingham2015} is a multi-year effort devoted to the creation and maintenance of a series of benchmark datasets for classification and object detection, creating the precedent for standardized evaluation of recognition algorithms in the form of annual competitions. Starting from only four categories in 2005, the dataset has increased to 20 categories that are common in everyday life. Since 2009, the number of images has grown every year, but with all previous images retained to allow test results to be compared from year to year. Due the availability of larger datasets like ImageNet, MS COCO and Open Images, PASCAL VOC has gradually fallen out of fashion. \textbf{ILSVRC}, the ImageNet Large Scale Visual Recognition Challenge \cite{Russakovsky2015}, is derived from ImageNet \cite{ImageNet2009}, scaling up PASCAL VOC's goal of standardized training and evaluation of detection algorithms by more than an order of magnitude in the number of object classes and images. ImageNet1000, a subset of ImageNet images with 1000 different object categories and a total of 1.2 million images, has been fixed to provide a standardized benchmark for the ILSVRC image classification challenge. \textbf{MS COCO} is a response to the criticism of ImageNet that objects in its dataset tend to be large and well centered, making the ImageNet dataset atypical of real-world scenarios. To push for richer image understanding, researchers created the MS COCO database \cite{Lin2014} containing complex everyday scenes with common objects in their natural context, closer to real life, where objects are labeled using fully-segmented instances to provide more accurate detector evaluation. The COCO object detection challenge \cite{Lin2014} features two object detection tasks: using either bounding box output or object instance segmentation output. COCO introduced three new challenges: \begin{enumerate} \item It contains objects at a wide range of scales, including a high percentage of small objects \cite{Singh2018SNIP}; \item Objects are less iconic and amid clutter or heavy occlusion; \item The evaluation metric (see Table~\ref{Tab:Metrics}) encourages more accurate object localization. \end{enumerate} Just like ImageNet in its time, MS COCO has become the standard for object detection today. \textbf{OICOD} (the Open Image Challenge Object Detection) is derived from Open Images V4 (now V5 in 2019) \cite{Kuznetsova2018Open}, currently the largest publicly available object detection dataset. OICOD is different from previous large scale object detection datasets like ILSVRC and MS COCO, not merely in terms of the significantly increased number of classes, images, bounding box annotations and instance segmentation mask annotations, but also regarding the annotation process. In ILSVRC and MS COCO, instances of all classes in the dataset are exhaustively annotated, whereas for Open Images V4 a classifier was applied to each image and only those labels with sufficiently high scores were sent for human verification. Therefore in OICOD only the object instances of human-confirmed positive labels are annotated. \begin{table*}[!t] \caption {Statistics of commonly used object detection datasets. Object statistics for VOC challenges list the non-difficult objects used in the evaluation (all annotated objects). For the COCO challenge, prior to 2017, the test set had four splits (\emph{Dev}, \emph{Standard}, \emph{Reserve}, and \emph{Challenge}), with each having about 20K images. Starting in 2017, the test set has only the \emph{Dev} and \emph{Challenge} splits, with the other two splits removed. Starting in 2017, the train and val sets are arranged differently, and the test set is divided into two roughly equally sized splits of about $20,000$ images each: Test Dev and Test Challenge. Note that the 2017 Test Dev/Challenge splits contain the same images as the 2015 Test Dev/Challenge splits, so results across the years are directly comparable.}\label{Tab:detdatasets} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{2pt} \resizebox*{14cm}{!}{ \begin{tabular}{!{\vrule width1.2bp}c|c|r|r|c!{\vrule width1.2bp}r|c!{\vrule width1.2bp}r|r|c!{\vrule width1.2bp}} \Xhline{1pt} \multirow{2}{*}{\footnotesize Challenge} & \multirow{2}{*}{\footnotesize \shortstack [c] {Object \\Classes}}& \multicolumn{3}{c!{\vrule width1.2bp}}{\footnotesize Number of Images} & \multicolumn{2}{c!{\vrule width1.2bp}}{\footnotesize \shortstack [c] {Number of Annotated Objects}} & \multicolumn{3}{c!{\vrule width1.2bp}}{\footnotesize Summary (Train$+$Val)} \\ \cline{3-10} \footnotesize & \footnotesize & \footnotesize Train & \footnotesize Val & \footnotesize Test & \footnotesize \shortstack [c] { Train} & \footnotesize Val & \footnotesize Images & \footnotesize Boxes & \footnotesize Boxes/Image\\ \Xhline{1pt} \multicolumn{10}{!{\vrule width1.2bp}c!{\vrule width1.2bp}}{\footnotesize PASCAL VOC Object Detection Challenge}\\ \hline \footnotesize VOC07 & \footnotesize $20$ & \footnotesize$ 2,501$ & \footnotesize $2,510 $ & \footnotesize$4,952 $ & \footnotesize $6,301(7,844)$ & \footnotesize $6,307(7,818)$ & \footnotesize $5,011$ & \footnotesize$12,608$& \footnotesize $2.5$\\ \hline \footnotesize VOC08 & \footnotesize $20$ & \footnotesize$2,111$ & \footnotesize$2,221$ & \footnotesize$4,133$ & \footnotesize $5,082(6,337) $ & \footnotesize$5,281(6,347) $ & \footnotesize $4,332$& \footnotesize$10,364$& \footnotesize $2.4$ \\ \hline \footnotesize VOC09 & \footnotesize $20 $& \footnotesize$3,473$& \footnotesize $3,581$ & \footnotesize $6,650 $& \footnotesize$8,505(9,760) $ & \footnotesize$ 8,713(9,779)$& \footnotesize$7,054$& \footnotesize $17,218$& \footnotesize $2.3$\\ \hline \footnotesize VOC10 & \footnotesize $20$ & \footnotesize$4,998 $& \footnotesize$5,105$ & \footnotesize$9,637$ & \footnotesize $11,577(13,339)$ & \footnotesize $11,797(13,352)$ & \footnotesize$10,103$& \footnotesize $23,374$& \footnotesize$2.4$ \\ \hline \footnotesize VOC11 & \footnotesize $20$ & \footnotesize $5,717$ & \footnotesize $5,823$ & \footnotesize$10,994$ & \footnotesize$13,609 (15,774) $& \footnotesize $13,841(15,787) $ & \footnotesize$11,540$& \footnotesize$27,450$& \footnotesize $2.4$ \\ \hline \footnotesize VOC12 & \footnotesize $20$ & \footnotesize$5,717$& \footnotesize $5,823$ & \footnotesize$10,991$& \footnotesize $13,609 (15,774) $& \footnotesize $13,841(15,787) $& \footnotesize $11,540$& \footnotesize $27,450$& \footnotesize$2.4$\\ \Xhline{1pt} \multicolumn{10}{!{\vrule width1.2bp}c!{\vrule width1.2bp}}{\footnotesize ILSVRC Object Detection Challenge}\\ \hline \footnotesize ILSVRC13& \footnotesize $200$ & \footnotesize$395,909$ & \footnotesize $20,121 $& \footnotesize$40,152$ & \footnotesize$345,854 $& \footnotesize$55,502$& \footnotesize$416,030$& \footnotesize $401,356$& \footnotesize $1.0$\\ \hline \footnotesize ILSVRC14 & \footnotesize$200 $& \footnotesize$456,567$ & \footnotesize$20,121$ & \footnotesize$40,152$& \footnotesize$ 478,807 $& \footnotesize$55,502$& \footnotesize $476,668$& \footnotesize$534,309$& \footnotesize$1.1$\\ \hline \footnotesize ILSVRC15& \footnotesize $200$& \footnotesize $456,567$ & \footnotesize$20,121$& \footnotesize $51,294$ & \footnotesize$478,807$& \footnotesize $55,502$& \footnotesize$476,668$& \footnotesize $534,309$ & \footnotesize$1.1$\\ \hline \footnotesize ILSVRC16 & \footnotesize$200$& \footnotesize$ 456,567$& \footnotesize$ 20,121$& \footnotesize $60,000$& \footnotesize$ 478,807$& \footnotesize $55,502$& \footnotesize$476,668$& \footnotesize $534,309$ & \footnotesize$1.1$\\ \hline \footnotesize ILSVRC17 & \footnotesize$ 200$& \footnotesize $456,567 $& \footnotesize$ 20,121$& \footnotesize$ 65,500 $& \footnotesize$478,807$ & \footnotesize$55,502$& \footnotesize$476,668$& \footnotesize $534,309$& \footnotesize$1.1$\\ \Xhline{1pt} \multicolumn{10}{!{\vrule width1.2bp}c!{\vrule width1.2bp}}{\footnotesize MS COCO Object Detection Challenge}\\ \hline \footnotesize MS COCO15 & \footnotesize $80 $& \footnotesize $82,783$ & \footnotesize $40,504 $ & \footnotesize$81,434$ & \footnotesize $604,907$ & \footnotesize$ 291,875$& \footnotesize$123,287$& \footnotesize $896,782$& \footnotesize$7.3$\\ \hline \footnotesize MS COCO16 & \footnotesize $80$ & \footnotesize$ 82,783$ & \footnotesize $40,504$ & \footnotesize $81,434$ & \footnotesize$ 604,907 $& \footnotesize $291,875$& \footnotesize$123,287$& \footnotesize$896,782$& \footnotesize $7.3$ \\ \hline \footnotesize MS COCO17 & \footnotesize$ 80$ & \footnotesize$118,287$ & \footnotesize$ 5,000$& \footnotesize$40,670$& \footnotesize$860,001$& \footnotesize$36,781$ & \footnotesize$123,287$& \footnotesize $896,782$& \footnotesize $7.3$ \\ \hline \footnotesize MS COCO18 & \footnotesize$ 80$ & \footnotesize$118,287$ & \footnotesize$ 5,000$& \footnotesize$40,670$& \footnotesize$860,001$& \footnotesize$36,781$& \footnotesize$123,287$& \footnotesize $896,782$& \footnotesize $7.3$ \\ \Xhline{1pt} \multicolumn{10}{!{\vrule width1.2bp}c!{\vrule width1.2bp}}{\footnotesize Open Images Challenge Object Detection (OICOD) (Based on Open Images V4 \cite{Kuznetsova2018Open})}\\ \hline \footnotesize OICOD18 & \footnotesize $500$& \footnotesize $1,643,042$ & \footnotesize $100,000$ & \footnotesize$99,999$ & \footnotesize $11,498,734$ & \footnotesize $696,410$ & \footnotesize $1,743,042$& \footnotesize $12,195,144$& \footnotesize $7.0$\\ \Xhline{1pt} \end{tabular} } \end{table*} \subsection{Evaluation Criteria} \label{sec:EvaluationCriteria} There are three criteria for evaluating the performance of detection algorithms: detection speed in Frames Per Second (FPS), precision, and recall. The most commonly used metric is \emph{Average Precision} (AP), derived from precision and recall. AP is usually evaluated in a category specific manner, \emph{i.e.}, computed for each object category separately. To compare performance over all object categories, the \emph{mean AP} (mAP) averaged over all object categories is adopted as the final measure of performance\footnote{In object detection challenges, such as PASCAL VOC and ILSVRC, the winning entry of each object category is that with the highest AP score, and the winner of the challenge is the team that wins on the most object categories. The mAP is also used as the measure of a team's performance, and is justified since the ranking of teams by mAP was always the same as the ranking by the number of object categories won \cite{Russakovsky2015}.}. More details on these metrics can be found in \cite{Everingham2010,Everingham2015,Russakovsky2015,Hoiem2012}. \begin {figure}[!t] \centering \includegraphics[width=0.49\textwidth]{Algorithm.pdf} \caption{The algorithm for determining TPs and FPs by greedily matching object detection results to ground truth boxes.} \label{fig:Algorithm1} \end {figure} The standard outputs of a detector applied to a testing image $\textbf{I}$ are the predicted detections $\{(b_j,c_j,p_j)\}_j$, indexed by object $j$, of Bounding Box (BB) $b_j$, predicted category $c_j$, and confidence $p_j$. A predicted detection $(b,c,p)$ is regarded as a True Positive (TP) if \begin{itemize} \renewcommand{\labelitemi}{$\bullet$} \item The predicted category $c$ equals the ground truth label $c_g$. \item The overlap ratio IOU (Intersection Over Union) \cite{Everingham2010,Russakovsky2015} \begin{equation}\label{eqn:IOU} \textrm{IOU}(b,b^g)=\frac{area(b\cap b^g)}{area(b\cup b^g)}, \end{equation} between the predicted BB $b$ and the ground truth $b^g$ is not smaller than a predefined threshold $\varepsilon$, where $\cap$ and $cup$ denote intersection and union, respectively. A typical value of $\varepsilon$ is 0.5. \end{itemize} Otherwise, it is considered as a False Positive (FP). The confidence level $p$ is usually compared with some threshold $\beta$ to determine whether the predicted class label $c$ is accepted. AP is computed separately for each of the object classes, based on \emph{Precision} and \emph{Recall}. For a given object class $c$ and a testing image $\textbf{I}_i$, let $\{(b_{ij},p_{ij})\}_{j=1}^M$ denote the detections returned by a detector, ranked by confidence $p_{ij}$ in decreasing order. Each detection $(b_{ij},p_{ij})$ is either a TP or an FP, which can be determined via the algorithm\footnote{It is worth noting that for a given threshold $\beta$, multiple detections of the same object in an image are not considered as all correct detections, and only the detection with the highest confidence level is considered as a TP and the rest as FPs.} in Fig.~\ref{fig:Algorithm1}. Based on the TP and FP detections, the precision $P(\beta)$ and recall $R(\beta)$ \cite{Everingham2010} can be computed as a function of the confidence threshold $\beta$, so by varying the confidence threshold different pairs $(P,R)$ can be obtained, in principle allowing precision to be regarded as a function of recall, \emph{i.e.} $P(R)$, from which the Average Precision (AP) \cite{Everingham2010,Russakovsky2015} can be found. Since the introduction of MS COCO, more attention has been placed on the accuracy of the bounding box location. Instead of using a fixed IOU threshold, MS COCO introduces a few metrics (summarized in Table~\ref{Tab:Metrics}) for characterizing the performance of an object detector. For instance, in contrast to the traditional mAP computed at a single IoU of $0.5$, $AP_{coco}$ is averaged across all object categories and multiple IOU values from $0.5$ to $0.95$ in steps of $0.05$. Because $41\%$ of the objects in MS COCO are small and $24\%$ are large, metrics $AP_{coco}^{small}$, $AP_{coco}^{medium}$ and $AP_{coco}^{large}$ are also introduced. Finally, Table~\ref{Tab:Metrics} summarizes the main metrics used in the PASCAL, ILSVRC and MS COCO object detection challenges, with metric modifications for the Open Images challenges proposed in \cite{Kuznetsova2018Open}. \begin{table}[!t] \caption {Summary of commonly used metrics for evaluating object detectors.}\label{Tab:Metrics} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{1pt} \resizebox*{9cm}{!}{ \begin{tabular}{!{\vrule width1.2bp}c|c|l|l!{\vrule width1.2bp}} \Xhline{1pt} \footnotesize Metric & \footnotesize Meaning & \multicolumn{2}{c!{\vrule width1.2bp}}{\footnotesize Definition and Description} \\ \hline \raisebox{1ex}[0pt]{\footnotesize TP} & \footnotesize \shortstack [c] {True \\ Positive} & \multicolumn{2}{l!{\vrule width1.2bp}}{\raisebox{1ex}[0pt]{\footnotesize A true positive detection, per Fig.~\ref{fig:Algorithm1}.}} \\ \hline \raisebox{1ex}[0pt]{\footnotesize FP} & \footnotesize \shortstack [c] {False \\ Positive} & \multicolumn{2}{l!{\vrule width1.2bp}}{\raisebox{1ex}[0pt]{\footnotesize A false positive detection, per Fig.~\ref{fig:Algorithm1}.}} \\ \hline \raisebox{1ex}[0pt]{\footnotesize $\beta$ } & \footnotesize \shortstack [c] {Confidence \\ Threshold} & \multicolumn{2}{l!{\vrule width1.2bp}}{\raisebox{1ex}[0pt]{\footnotesize A confidence threshold for computing $P(\beta)$ and $R(\beta)$.}} \\ \hline \multirow{3}{*}{\footnotesize $\varepsilon$} & \multirow{3}{*}{\footnotesize \shortstack [c] {IOU \\ Threshold}} & \footnotesize VOC & \footnotesize Typically around $0.5$ \\ \cline{3-4} & \footnotesize & \footnotesize ILSVRC & \footnotesize $\min(0.5,\frac{wh}{(w+10)(h+10)})$; $w\times h$ is the size of a GT box. \\ \cline{3-4} & \footnotesize & \footnotesize MS COCO & \footnotesize Ten IOU thresholds $\varepsilon\in\{0.5:0.05:0.95\}$ \\ \hline \footnotesize $P(\beta)$ & \raisebox{1.3ex}[0pt]{\footnotesize \shortstack [c] {Precision}} & \multicolumn{2}{l!{\vrule width1.2bp}}{\footnotesize \shortstack [l] {The fraction of correct detections out of the total detections returned \\ by the detector with confidence of at least $\beta$.}} \\ \hline \footnotesize $R(\beta)$ & \raisebox{1.3ex}[0pt]{\footnotesize \shortstack [c] {Recall}}& \multicolumn{2}{l!{\vrule width1.2bp}}{\footnotesize \shortstack [l] {The fraction of all $N_c$ objects detected by the detector having a \\ confidence of at least $\beta$.}} \\ \hline \raisebox{1ex}[0pt]{\footnotesize AP } & \footnotesize \shortstack [c] {Average \\ Precision} & \multicolumn{2}{l!{\vrule width1.2bp}}{\footnotesize \shortstack [l] {Computed over the different levels of recall achieved by varying \\ the confidence $\beta$.}} \\ \hline \multirow{8}{*}{\footnotesize mAP} & \multirow{8}{*}{\footnotesize \shortstack [c] {mean \\Average\\Precision}} & \footnotesize VOC & \footnotesize AP at a single IOU and averaged over all classes. \\ \cline{3-4} & \footnotesize & \footnotesize ILSVRC & \footnotesize AP at a modified IOU and averaged over all classes. \\ \cline{3-4} & \footnotesize & \multirow{6}{*}{\footnotesize MS COCO}& \footnotesize $\bullet AP_{coco}$: mAP averaged over ten IOUs: $\{0.5:0.05:0.95\}$;\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AP^{\textrm{IOU}=0.5}_{coco}$: mAP at IOU=0.50 (PASCAL VOC metric);\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AP^{\textrm{IOU}=0.75}_{coco}$: mAP at IOU=0.75 (strict metric);\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AP^{\textrm{small}}_{coco}$: mAP for small objects of area smaller than $32^2$;\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AP^{\textrm{medium}}_{coco}$: mAP for objects of area between $32^2$ and $96^2$;\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AP^{\textrm{large}}_{coco}$: mAP for large objects of area bigger than $96^2$; \\ \hline \raisebox{1ex}[0pt]{\footnotesize AR} & \footnotesize \shortstack [c] {Average \\ Recall} & \multicolumn{2}{c!{\vrule width1.2bp}}{\footnotesize \shortstack [l] {The maximum recall given a fixed number of detections per image, \\ averaged over all categories and IOU thresholds.}} \\ \hline \multirow{6}{*}{\footnotesize AR} & \multirow{6}{*}{\footnotesize \shortstack [c] {Average\\Recall}} & \multirow{6}{*}{\footnotesize MS COCO}& \footnotesize $\bullet AR^{\textrm{max}=1}_{coco}$: AR given 1 detection per image;\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AR^{\textrm{max}=10}_{coco}$: AR given 10 detection per image;\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AR^{\textrm{max}=100}_{coco}$: AR given 100 detection per image;\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AR^{\textrm{small}}_{coco}$: AR for small objects of area smaller than $32^2$;\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AR^{\textrm{medium}}_{coco}$: AR for objects of area between $32^2$ and $96^2$;\\ & \footnotesize& \footnotesize& \footnotesize$\bullet$ $AR^{\textrm{large}}_{coco}$: AR for large objects of area bigger than $96^2$; \\ \Xhline{1pt} \end{tabular} } \end{table} \section{Detection Frameworks} \label{Sec:Frameworks} There has been steady progress in object feature representations and classifiers for recognition, as evidenced by the dramatic change from handcrafted features \cite{Viola2001,Dalal2005HOG,Felzenszwalb08CVPR, Harzallah2009Combining,Vedaldi09Multiple} to learned DCNN features \cite{Girshick2014RCNN,Ouyang2015deepid,Girshick2015FRCNN, Ren2015NIPS,Dai2016RFCN}. In contrast, in terms of localization, the basic ``sliding window'' strategy \cite{Dalal2005HOG,Felzenszwalb2010b,Felzenszwalb08CVPR} remains mainstream, although with some efforts to avoid exhaustive search \cite{lampert2008beyond,Uijlings2013b}. However, the number of windows is large and grows quadratically with the number of image pixels, and the need to search over multiple scales and aspect ratios further increases the search space. Therefore, the design of efficient and effective detection frameworks plays a key role in reducing this computational cost. Commonly adopted strategies include cascading, sharing feature computation, and reducing per-window computation. This section reviews detection frameworks, listed in Fig.~\ref{fig:MilestonesAfter2014} and Table~\ref{Tab:Detectors}, the milestone approaches appearing since deep learning entered the field, organized into two main categories: \begin{enumerate} \item [a.] Two stage detection frameworks, which include a preprocessing step for generating object proposals; \item [b.] One stage detection frameworks, or region proposal free frameworks, having a single proposed method which does not separate the process of the detection proposal. \end{enumerate} Sections~\ref{Sec:DCNNFeatures} through~\ref{sec:otherissue} will discuss fundamental sub-problems involved in detection frameworks in greater detail, including DCNN features, detection proposals, and context modeling. \begin {figure*}[!t] \centering \includegraphics[width=0.9\textwidth]{MilestonesPastSeveralYears.pdf} \caption{Milestones in generic object detection.} \label{fig:MilestonesAfter2014} \end {figure*} \subsection{Region Based (Two Stage) Frameworks} \label{Sec:RegionBased} In a region-based framework, category-independent region proposals\footnote{Object proposals, also called region proposals or detection proposals, are a set of candidate regions or bounding boxes in an image that may potentially contain an object. \cite{Chavali2016,Hosang2016}} are generated from an image, CNN \cite{Krizhevsky2012} features are extracted from these regions, and then category-specific classifiers are used to determine the category labels of the proposals. As can be observed from Fig.~\ref{fig:MilestonesAfter2014}, DetectorNet \cite{Szegedy2013Deep}, OverFeat \cite{OverFeat2014}, MultiBox \cite{MultiBox1} and RCNN \cite{Girshick2014RCNN} independently and almost simultaneously proposed using CNNs for generic object detection. \begin {figure}[!t] \centering \includegraphics[width=0.5\textwidth]{RegionBased.pdf} \caption{Illustration of the RCNN detection framework \cite{Girshick2014RCNN,Girshick2016TPAMI}.} \label{fig:RegionBased} \end {figure} \begin {figure}[!t] \centering \includegraphics[width=0.46\textwidth]{RegionVsUnified1.pdf} \includegraphics[width=0.46\textwidth]{RegionVsUnified2.pdf} \caption{High level diagrams of the leading frameworks for generic object detection. The properties of these methods are summarized in Table \ref{Tab:Detectors}.} \label{Fig:RegionVsUnified} \end {figure} \textbf{RCNN} \cite{Girshick2014RCNN}: Inspired by the breakthrough image classification results obtained by CNNs and the success of the selective search in region proposal for handcrafted features \cite{Uijlings2013b}, Girshick \emph{et al.} were among the first to explore CNNs for generic object detection and developed RCNN \cite{Girshick2014RCNN,Girshick2016TPAMI}, which integrates AlexNet \cite{Krizhevsky2012} with a region proposal selective search \cite{Uijlings2013b}. As illustrated in detail in Fig.~\ref{fig:RegionBased}, training an RCNN framework consists of multistage pipelines: \begin{enumerate} \item \emph{Region proposal computation:} Class agnostic region proposals, which are candidate regions that might contain objects, are obtained via a selective search \cite{Uijlings2013b}. \item \emph{CNN model finetuning:} Region proposals, which are cropped from the image and warped into the same size, are used as the input for fine-tuning a CNN model pre-trained using a large-scale dataset such as ImageNet. At this stage, all region proposals with $\geqslant0.5$ IOU \footnote{Please refer to Section \ref{sec:EvaluationCriteria} for the definition of IOU.} overlap with a ground truth box are defined as positives for that ground truth box's class and the rest as negatives. \item \emph{Class specific SVM classifiers training:} A set of class-specific linear SVM classifiers are trained using fixed length features extracted with CNN, replacing the softmax classifier learned by fine-tuning. For training SVM classifiers, positive examples are defined to be the ground truth boxes for each class. A region proposal with less than 0.3 IOU overlap with all ground truth instances of a class is negative for that class. Note that the positive and negative examples defined for training the SVM classifiers are different from those for fine-tuning the CNN. \item \emph{Class specific bounding box regressor training:} Bounding box regression is learned for each object class with CNN features. \end{enumerate} In spite of achieving high object detection quality, RCNN has notable drawbacks \cite{Girshick2015FRCNN}: \begin{enumerate} \item Training is a multistage pipeline, slow and hard to optimize because each individual stage must be trained separately. \item For SVM classifier and bounding box regressor training, it is expensive in both disk space and time, because CNN features need to be extracted from each object proposal in each image, posing great challenges for large scale detection, particularly with very deep networks, such as VGG16 \cite{Simonyan2014VGG}. \item Testing is slow, since CNN features are extracted per object proposal in each test image, without shared computation. \end{enumerate} All of these drawbacks have motivated successive innovations, leading to a number of improved detection frameworks such as SPPNet, Fast RCNN, Faster RCNN \emph{etc}., as follows. \textbf{SPPNet} \cite{He2014SPP}: During testing, CNN feature extraction is the main bottleneck of the RCNN detection pipeline, which requires the extraction of CNN features from thousands of warped region proposals per image. As a result, He \emph{et al.} \cite{He2014SPP} introduced traditional spatial pyramid pooling (SPP) \cite{Grauman2005pyramid,Lazebnik2006SPM} into CNN architectures. Since convolutional layers accept inputs of arbitrary sizes, the requirement of fixed-sized images in CNNs is due only to the Fully Connected (FC) layers, therefore He \emph{et al.} added an SPP layer on top of the last convolutional (CONV) layer to obtain features of fixed length for the FC layers. With this SPPNet, RCNN obtains a significant speedup without sacrificing any detection quality, because it only needs to run the convolutional layers {\emph once} on the entire test image to generate fixed-length features for region proposals of arbitrary size. While SPPNet accelerates RCNN evaluation by orders of magnitude, it does not result in a comparable speedup of the detector training. Moreover, fine-tuning in SPPNet \cite{He2014SPP} is unable to update the convolutional layers before the SPP layer, which limits the accuracy of very deep networks. \textbf{Fast RCNN} \cite{Girshick2015FRCNN}: Girshick proposed Fast RCNN \cite{Girshick2015FRCNN} that addresses some of the disadvantages of RCNN and SPPNet, while improving on their detection speed and quality. As illustrated in Fig.~\ref{Fig:RegionVsUnified}, Fast RCNN enables end-to-end detector training by developing a streamlined training process that simultaneously learns a softmax classifier and class-specific bounding box regression, rather than separately training a softmax classifier, SVMs, and Bounding Box Regressors (BBRs) as in RCNN/SPPNet. Fast RCNN employs the idea of sharing the computation of convolution across region proposals, and adds a Region of Interest (RoI) pooling layer between the last CONV layer and the first FC layer to extract a fixed-length feature for each region proposal. Essentially, RoI pooling uses warping at the feature level to approximate warping at the image level. The features after the RoI pooling layer are fed into a sequence of FC layers that finally branch into two sibling output layers: softmax probabilities for object category prediction, and class-specific bounding box regression offsets for proposal refinement. Compared to RCNN/SPPNet, Fast RCNN improves the efficiency considerably -- typically 3 times faster in training and 10 times faster in testing. Thus there is higher detection quality, a single training process that updates all network layers, and no storage required for feature caching. \textbf{Faster RCNN} \cite{Ren2015NIPS,Ren2016a}: Although Fast RCNN significantly sped up the detection process, it still relies on external region proposals, whose computation is exposed as the new speed bottleneck in Fast RCNN. Recent work has shown that CNNs have a remarkable ability to localize objects in CONV layers \cite{Zhoubolei2014,Zhou2016learning, Cinbis2017,Oquab2015object,Hariharan2016}, an ability which is weakened in the FC layers. Therefore, the selective search can be replaced by a CNN in producing region proposals. The Faster RCNN framework proposed by Ren \emph{et al.} \cite{Ren2015NIPS,Ren2016a} offered an efficient and accurate Region Proposal Network (RPN) for generating region proposals. They utilize the same backbone network, using features from the last shared convolutional layer to accomplish the task of RPN for region proposal and Fast RCNN for region classification, as shown in Fig.~\ref{Fig:RegionVsUnified}. RPN first initializes $k$ reference boxes (\emph{i.e.} the so called \emph{anchors}) of different scales and aspect ratios at each CONV feature map location. The anchor {\emph positions} are image content independent, but the feature vectors themselves, extracted from anchors, are image content dependent. Each anchor is mapped to a lower dimensional vector, which is fed into two sibling FC layers --- an object category classification layer and a box regression layer. In contrast to detection in Fast RCNN, the features used for regression in RPN are of the same shape as the anchor box, thus $k$ anchors lead to $k$ regressors. RPN shares CONV features with Fast RCNN, thus enabling highly efficient region proposal computation. RPN is, in fact, a kind of Fully Convolutional Network (FCN) \cite{FCNCVPR2015,FCNTPAMI}; Faster RCNN is thus a purely CNN based framework without using handcrafted features. For the VGG16 model \cite{Simonyan2014VGG}, Faster RCNN can test at 5 FPS (including all stages) on a GPU, while achieving state-of-the-art object detection accuracy on PASCAL VOC 2007 using 300 proposals per image. The initial Faster RCNN in \cite{Ren2015NIPS} contains several alternating training stages, later simplified in \cite{Ren2016a}. Concurrent with the development of Faster RCNN, Lenc and Vedaldi \cite{Lenc2015} challenged the role of region proposal generation methods such as selective search, studied the role of region proposal generation in CNN based detectors, and found that CNNs contain sufficient geometric information for accurate object detection in the CONV rather than FC layers. They showed the possibility of building integrated, simpler, and faster object detectors that rely exclusively on CNNs, removing region proposal generation methods such as selective search. {\textbf{RFCN (Region based Fully Convolutional Network)}}: While Faster RCNN is an order of magnitude faster than Fast RCNN, the fact that the region-wise sub-network still needs to be applied per RoI (several hundred RoIs per image) led Dai \emph{et al.} \cite{Dai2016RFCN} to propose the RFCN detector which is \emph{fully convolutional} (no hidden FC layers) with almost all computations shared over the entire image. As shown in Fig.~\ref{Fig:RegionVsUnified}, RFCN differs from Faster RCNN only in the RoI sub-network. In Faster RCNN, the computation after the RoI pooling layer cannot be shared, so Dai \emph{et al.} \cite{Dai2016RFCN} proposed using all CONV layers to construct a shared RoI sub-network, and RoI crops are taken from the last layer of CONV features prior to prediction. However, Dai \emph{et al.} \cite{Dai2016RFCN} found that this naive design turns out to have considerably inferior detection accuracy, conjectured to be that deeper CONV layers are more sensitive to category semantics, and less sensitive to translation, whereas object detection needs localization representations that respect translation invariance. Based on this observation, Dai \emph{et al.} \cite{Dai2016RFCN} constructed a set of position-sensitive score maps by using a bank of specialized CONV layers as the FCN output, on top of which a position-sensitive RoI pooling layer is added. They showed that RFCN with ResNet101 \cite{He2016ResNet} could achieve comparable accuracy to Faster RCNN, often at faster running times. \textbf{Mask RCNN}: He \emph{et al.} \cite{MaskRCNN2017} proposed Mask RCNN to tackle pixelwise object instance segmentation by extending Faster RCNN. Mask RCNN adopts the same two stage pipeline, with an identical first stage (RPN), but in the second stage, in parallel to predicting the class and box offset, Mask RCNN adds a branch which outputs a binary mask for each RoI. The new branch is a Fully Convolutional Network (FCN) \cite{FCNCVPR2015,FCNTPAMI} on top of a CNN feature map. In order to avoid the misalignments caused by the original RoI pooling (RoIPool) layer, a RoIAlign layer was proposed to preserve the pixel level spatial correspondence. With a backbone network ResNeXt101-FPN \cite{Xie2016Aggregated,FPN2016}, Mask RCNN achieved top results for the COCO object instance segmentation and bounding box object detection. It is simple to train, generalizes well, and adds only a small overhead to Faster RCNN, running at 5 FPS \cite{MaskRCNN2017}. \textbf{Chained Cascade Network and Cascade RCNN}: The essence of cascade \cite{Felzenszwalb2010Cascade,Bourdev2005Robust,Li2004Floatboost} is to learn more discriminative classifiers by using multistage classifiers, such that early stages discard a large number of easy negative samples so that later stages can focus on handling more difficult examples. Two-stage object detection can be considered as a cascade, the first detector removing large amounts of background, and the second stage classifying the remaining regions. Recently, end-to-end learning of more than two cascaded classifiers and DCNNs for generic object detection were proposed in the Chained Cascade Network \cite{Ouyang2017Chained}, extended in Cascade RCNN \cite{CascadeRCNN2018}, and more recently applied for simultaneous object detection and instance segmentation \cite{Chen2019Hybrid}, winning the COCO 2018 Detection Challenge. \textbf{Light Head RCNN}: In order to further increase the detection speed of RFCN \cite{Dai2016RFCN}, Li \emph{et al.} \cite{Li2018Light} proposed Light Head RCNN, making the head of the detection network as light as possible to reduce the RoI computation. In particular, Li \emph{et al.} \cite{Li2018Light} applied a convolution to produce thin feature maps with small channel numbers (\emph{e.g.,} 490 channels for COCO) and a cheap RCNN sub-network, leading to an excellent trade-off of speed and accuracy. \subsection{Unified (One Stage) Frameworks} \label{Sec:Unified} The region-based pipeline strategies of Section~\ref{Sec:RegionBased} have dominated since RCNN \cite{Girshick2014RCNN}, such that the leading results on popular benchmark datasets are all based on Faster RCNN \cite{Ren2015NIPS}. Nevertheless, region-based approaches are computationally expensive for current mobile/wearable devices, which have limited storage and computational capability, therefore instead of trying to optimize the individual components of a complex region-based pipeline, researchers have begun to develop \emph{unified} detection strategies. Unified pipelines refer to architectures that directly predict class probabilities and bounding box offsets from full images with a single feed-forward CNN in a monolithic setting that does not involve region proposal generation or post classification / feature resampling, encapsulating all computation in a single network. Since the whole pipeline is a single network, it can be optimized end-to-end directly on detection performance. \textbf{DetectorNet}: Szegedy \emph{et al.} \cite{Szegedy2013Deep} were among the first to explore CNNs for object detection. DetectorNet formulated object detection a regression problem to object bounding box masks. They use AlexNet \cite{Krizhevsky2012} and replace the final softmax classifier layer with a regression layer. Given an image window, they use one network to predict foreground pixels over a coarse grid, as well as four additional networks to predict the object's top, bottom, left and right halves. A grouping process then converts the predicted masks into detected bounding boxes. The network needs to be trained per object type and mask type, and does not scale to multiple classes. DetectorNet must take many crops of the image, and run multiple networks for each part on every crop, thus making it slow. \begin {figure}[!t] \centering \includegraphics[width=0.5\textwidth]{OverFeat.pdf} \caption{Illustration of the OverFeat \cite{OverFeat2014} detection framework.} \label{fig:OverFeat} \end {figure} \textbf{OverFeat}, proposed by Sermanet \emph{et al.} \cite{OverFeat2014} and illustrated in Fig.~\ref{fig:OverFeat}, can be considered as one of the first single-stage object detectors based on fully convolutional deep networks. It is one of the most influential object detection frameworks, winning the ILSVRC2013 localization and detection competition. OverFeat performs object detection via a single forward pass through the fully convolutional layers in the network (\emph{i.e.} the ``Feature Extractor", shown in Fig. \ref{fig:OverFeat} (a)). The key steps of object detection at test time can be summarized as follows: \begin{enumerate} \item \emph{Generate object candidates by performing object classification via a sliding window fashion on multiscale images.} OverFeat uses a CNN like AlexNet \cite{Krizhevsky2012}, which would require input images ofa fixed size due to its fully connected layers, in order to make the sliding window approach computationally efficient, OverFeat casts the network (as shown in Fig. \ref{fig:OverFeat} (a)) into a fully convolutional network, taking inputs of any size, by viewing fully connected layers as convolutions with kernels of size $1\times1$. OverFeat leverages multiscale features to improve the overall performance by passing up to six enlarged scales of the original image through the network (as shown in Fig. \ref{fig:OverFeat} (b)), resulting in a significantly increased number of evaluated context views. For each of the multiscale inputs, the classifier outputs a grid of predictions (class and confidence). \item \emph{Increase the number of predictions by offset max pooling}. In order to increase resolution, OverFeat applies offset max pooling after the last CONV layer, \emph{i.e.} performing a subsampling operation at every offset, yielding many more views for voting, increasing robustness while remaining efficient. \item \emph{Bounding box regression.} Once an object is identified, a single bounding box regressor is applied. The classifier and the regressor share the same feature extraction (CONV) layers, only the FC layers need to be recomputed after computing the classification network. \item \emph{Combine predictions.} OverFeat uses a greedy merge strategy to combine the individual bounding box predictions across all locations and scales. \end{enumerate} OverFeat has a significant speed advantage, but is less accurate than RCNN \cite{Girshick2014RCNN}, because it was difficult to train fully convolutional networks at the time. The speed advantage derives from sharing the computation of convolution between overlapping windows in the fully convolutional network. OverFeat is similar to later frameworks such as YOLO \cite{YoLo2016} and SSD \cite{Liu2016SSD}, except that the classifier and the regressors in OverFeat are trained sequentially. \textbf{YOLO}: Redmon \emph{et al.} \cite{YoLo2016} proposed YOLO (You Only Look Once), a unified detector casting object detection as a regression problem from image pixels to spatially separated bounding boxes and associated class probabilities, illustrated in Fig.~\ref{Fig:RegionVsUnified}. Since the region proposal generation stage is completely dropped, YOLO directly predicts detections using a small set of candidate regions\footnote{YOLO uses far fewer bounding boxes, only 98 per image, compared to about 2000 from Selective Search.}. Unlike region based approaches (\emph{e.g.} Faster RCNN) that predict detections based on features from a local region, YOLO uses features from an entire image globally. In particular, YOLO divides an image into an $S\times S$ grid, each predicting $C$ class probabilities, $B$ bounding box locations, and confidence scores. By throwing out the region proposal generation step entirely, YOLO is fast by design, running in real time at 45 FPS and Fast YOLO \cite{YoLo2016} at 155 FPS. Since YOLO sees the entire image when making predictions, it implicitly encodes contextual information about object classes, and is less likely to predict false positives in the background. YOLO makes more localization errors than Fast RCNN, resulting from the coarse division of bounding box location, scale and aspect ratio. As discussed in \cite{YoLo2016}, YOLO may fail to localize some objects, especially small ones, possibly because of the coarse grid division, and because each grid cell can only contain one object. It is unclear to what extent YOLO can translate to good performance on datasets with many objects per image, such as MS COCO. \textbf{YOLOv2 and YOLO9000}: Redmon and Farhadi \cite{YOLO9000} proposed YOLOv2, an improved version of YOLO, in which the custom GoogLeNet \cite{GoogLeNet2015} network is replaced with the simpler DarkNet19, plus batch normalization \cite{He2015delving}, removing the fully connected layers, and using good anchor boxes\footnote{Boxes of various sizes and aspect ratios that serve as object candidates.} learned via \emph{k}means and multiscale training. YOLOv2 achieved state-of-the-art on standard detection tasks. Redmon and Farhadi \cite{YOLO9000} also introduced YOLO9000, which can detect over 9000 object categories in real time by proposing a joint optimization method to train simultaneously on an ImageNet classification dataset and a COCO detection dataset with WordTree to combine data from multiple sources. Such joint training allows YOLO9000 to perform weakly supervised detection, \emph{i.e.} detecting object classes that do not have bounding box annotations. \textbf{SSD}: In order to preserve real-time speed without sacrificing too much detection accuracy, Liu \emph{et al.} \cite{Liu2016SSD} proposed SSD (Single Shot Detector), faster than YOLO \cite{YoLo2016} and with an accuracy competitive with region-based detectors such as Faster RCNN \cite{Ren2015NIPS}. SSD effectively combines ideas from RPN in Faster RCNN \cite{Ren2015NIPS}, YOLO \cite{YoLo2016} and multiscale CONV features \cite{Hariharan2016} to achieve fast detection speed, while still retaining high detection quality. Like YOLO, SSD predicts a fixed number of bounding boxes and scores, followed by an NMS step to produce the final detection. The CNN network in SSD is fully convolutional, whose early layers are based on a standard architecture, such as VGG \cite{Simonyan2014VGG}, followed by several auxiliary CONV layers, progressively decreasing in size. The information in the last layer may be too coarse spatially to allow precise localization, so SSD performs detection over multiple scales by operating on multiple CONV feature maps, each of which predicts category scores and box offsets for bounding boxes of appropriate sizes. For a $300\times300$ input, SSD achieves $74.3\%$ mAP on the VOC2007 test at 59 FPS versus Faster RCNN 7 FPS / mAP $73.2\%$ or YOLO 45 FPS / mAP $63.4\%$. \textbf{CornerNet:} Recently, Law \emph{et al.} \cite{Law2018CornerNet} questioned the dominant role that anchor boxes have come to play in SoA object detection frameworks \cite{Girshick2015FRCNN,MaskRCNN2017,YoLo2016,Liu2016SSD}. Law \emph{et al.} \cite{Law2018CornerNet} argue that the use of anchor boxes, especially in one stage detectors \cite{DSSD2016,LinICCV2017,Liu2016SSD,YoLo2016}, has drawbacks \cite{Law2018CornerNet,LinICCV2017} such as causing a huge imbalance between positive and negative examples, slowing down training and introducing extra hyperparameters. Borrowing ideas from the work on Associative Embedding in multiperson pose estimation \cite{Newell2017Associative}, Law \emph{et al.} \cite{Law2018CornerNet} proposed CornerNet by formulating bounding box object detection as detecting paired top-left and bottom-right keypoints\footnote{The idea of using keypoints for object detection appeared previously in DeNet \cite{SmithICCV2017}. }. In CornerNet, the backbone network consists of two stacked Hourglass networks \cite{Newell2016Stacked}, with a simple corner pooling approach to better localize corners. CornerNet achieved a $42.1\%$ AP on MS COCO, outperforming all previous one stage detectors; however, the average inference time is about 4FPS on a Titan X GPU, significantly slower than SSD \cite{Liu2016SSD} and YOLO \cite{YoLo2016}. CornerNet generates incorrect bounding boxes because it is challenging to decide which pairs of keypoints should be grouped into the same objects. To further improve on CornerNet, Duan \emph{et al.} \cite{Duan2019CenterNet} proposed CenterNet to detect each object as a triplet of keypoints, by introducing one extra keypoint at the centre of a proposal, raising the MS COCO AP to $47.0\%$, but with an inference speed slower than CornerNet. \section{Object Representation} \label{Sec:DCNNFeatures} As one of the main components in any detector, good feature representations are of primary importance in object detection \cite{Dickinson2009,Girshick2014RCNN,Gidaris2015,Zhu2016Do}. In the past, a great deal of effort was devoted to designing local descriptors (\emph{e.g.,} SIFT \cite{Lowe1999Object} and HOG \cite{Dalal2005HOG}) and to explore approaches (\emph{e.g.,} Bag of Words \cite{Sivic2003} and Fisher Vector \cite{Perronnin2010}) to group and abstract descriptors into higher level representations in order to allow the discriminative parts to emerge; however, these feature representation methods required careful engineering and considerable domain expertise. In contrast, deep learning methods (especially {\em deep} CNNs) can learn powerful feature representations with multiple levels of abstraction directly from raw images \cite{Bengio13Feature,LeCun15}. As the learning procedure reduces the dependency of specific domain knowledge and complex procedures needed in traditional feature engineering \cite{Bengio13Feature,LeCun15}, the burden for feature representation has been transferred to the design of better network architectures and training procedures. The leading frameworks reviewed in Section \ref{Sec:Frameworks} (RCNN \cite{Girshick2014RCNN}, Fast RCNN \cite{Girshick2015FRCNN}, Faster RCNN \cite{Ren2015NIPS}, YOLO \cite{YoLo2016}, SSD \cite{Liu2016SSD}) have persistently promoted detection accuracy and speed, in which it is generally accepted that the CNN architecture (Section \ref{Sec:PopularNetworks} and Table \ref{fig:ILSVRCclassificationResults}) plays a crucial role. As a result, most of the recent improvements in detection accuracy have been via research into the development of novel networks. Therefore we begin by reviewing popular CNN architectures used in Generic Object Detection, followed by a review of the effort devoted to improving object feature representations, such as developing invariant features to accommodate geometric variations in object scale, pose, viewpoint, part deformation and performing multiscale analysis to improve object detection over a wide range of scales. \begin {figure}[!t] \centering \includegraphics[width=0.45\textwidth]{ILSVRCclassificationResults.pdf} \caption{\footnotesize{Performance of winning entries in the ILSVRC competitions from 2011 to 2017 in the image classification task.}} \label{fig:ILSVRCclassificationResults} \end {figure} \begin{table*}[!t] \caption {DCNN architectures that were commonly used for generic object detection. Regarding the statistics for ``\#Paras'' and ``\#Layers'', the final FC prediction layer is not taken into consideration. ``Test Error'' column indicates the Top 5 classification test error on ImageNet1000. When ambiguous, the ``\#Paras'', ``\#Layers'', and ``Test Error'' refer to: OverFeat (accurate model), VGGNet16, ResNet101 DenseNet201 (Growth Rate 32, DenseNet-BC), ResNeXt50 (32*4d), and SE ResNet50. }\label{Tab:dcnnarchitectures} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{2pt} \resizebox*{18cm}{!}{ \begin{tabular}{!{\vrule width1.2bp}c|c|c|c|c|c|p{9cm}!{\vrule width1.2bp}} \Xhline{1pt} \footnotesize No. & \footnotesize \shortstack [c] {DCNN \\ Architecture} & \footnotesize \shortstack [c] {\#Paras \\ ($\times10^6$)} & \footnotesize \shortstack [c] {\#Layers \\ (CONV+FC)} & \footnotesize \shortstack [c] {Test Error \\ (Top 5)} & \footnotesize \shortstack [c] { \shortstack [c] {First \\ Used In}} & \footnotesize Highlights \\ \Xhline{1pt} \raisebox{-3.3ex}[0pt]{$1$} & \raisebox{-3.3ex}[0pt]{\footnotesize AlexNet \cite{AlexNet2012}} & \raisebox{-3.3ex}[0pt]{ \footnotesize $57$}& \raisebox{-3.3ex}[0pt]{\footnotesize $5+2 $} & \raisebox{-3.3ex}[0pt]{\footnotesize $15.3\%$} & \raisebox{-3.3ex}[0pt]{\footnotesize \cite{Girshick2014RCNN}} & \footnotesize The first DCNN found effective for ImageNet classification; the historical turning point from hand-crafted features to CNN; Winning the ILSVRC2012 Image classification competition. \\ \hline $2$ & \footnotesize ZFNet (fast) \cite{ZeilerFergus2014} & \footnotesize $58$ & \footnotesize$ 5+2 $ &\footnotesize $14.8\%$ & \footnotesize \cite{He2014SPP}& \footnotesize Similar to AlexNet, different in stride for convolution, filter size, and number of filters for some layers. \\ \hline $3 $& \footnotesize OverFeat \cite{OverFeat2014} & \footnotesize $140 $& \footnotesize$ 6+2$ &\footnotesize $13.6\%$ & \footnotesize \cite{OverFeat2014} & \footnotesize Similar to AlexNet, different in stride for convolution, filter size, and number of filters for some layers. \\ \hline \raisebox{-1.3ex}[0pt]{$4$}&\raisebox{-1.3ex}[0pt]{ \footnotesize VGGNet \cite{Simonyan2014VGG}} & \raisebox{-1.3ex}[0pt]{\footnotesize$ 134$} & \raisebox{-1.3ex}[0pt]{\footnotesize $13+2 $} & \raisebox{-1.3ex}[0pt]{\footnotesize $6.8\%$}&\raisebox{-1.3ex}[0pt]{ \footnotesize \cite{Girshick2015FRCNN}}& \footnotesize Increasing network depth significantly by stacking $3\times3$ convolution filters and increasing the network depth step by step. \\ \hline \raisebox{-2.3ex}[0pt]{$5$}&\raisebox{-2.3ex}[0pt]{ \footnotesize GoogLeNet \cite{GoogLeNet2015} }&\raisebox{-2.3ex}[0pt]{ \footnotesize$ 6$} & \raisebox{-2.3ex}[0pt]{\footnotesize $ 22$ } & \raisebox{-2.3ex}[0pt]{\footnotesize $6.7\%$} &\raisebox{-2.3ex}[0pt]{ \footnotesize \cite{GoogLeNet2015} }& \footnotesize Use Inception module, which uses multiple branches of convolutional layers with different filter sizes and then concatenates feature maps produced by these branches. The first inclusion of bottleneck structure and global average pooling. \\ \hline $6$ & \footnotesize Inception v2 \cite{Ioffe2015} & \footnotesize $12$ & \footnotesize $31$ & \footnotesize $4.8\%$ & \footnotesize \cite{Howard2017MobileNets} & \footnotesize Faster training with the introduce of Batch Normalization.\\ \hline $7$ & \footnotesize Inception v3 \cite{ Szegedy2016a} & \footnotesize $22$ & \footnotesize $47$ & \footnotesize $3.6\%$ & \footnotesize & \footnotesize Inclusion of separable convolution and spatial resolution reduction. \\ \hline $8$ & \footnotesize YOLONet \cite{YoLo2016} & \footnotesize $64 $& \footnotesize $24+1$ & \footnotesize$-$ & \footnotesize \cite{YoLo2016} & \footnotesize A network inspired by GoogLeNet used in YOLO detector. \\ \hline $9$& \footnotesize ResNet50 \cite{He2016ResNet} & \footnotesize$ 23.4 $& \footnotesize $49$ & \footnotesize $3.6\%$ & \footnotesize \cite{He2016ResNet} & \footnotesize With identity mapping, substantially deeper networks can be learned. \\ \cline{1-4}\cline{6-6} $10$ & \footnotesize ResNet101 \cite{He2016ResNet} & \footnotesize $42$ & \footnotesize $100 $ & \footnotesize (ResNets) & \footnotesize \cite{He2016ResNet} & \footnotesize Requires fewer parameters than VGG by using the global average pooling and bottleneck introduced in GoogLeNet. \\ \hline \raisebox{-1.3ex}[0pt]{$11$} & \raisebox{-1.3ex}[0pt]{ \footnotesize InceptionResNet v1 \cite{InceptionV4} } & \raisebox{-1.3ex}[0pt]{\footnotesize $21$ }& \raisebox{-1.3ex}[0pt]{\footnotesize $87$} &\multirow{3}{*}{ \footnotesize $3.1\%$} &\raisebox{-1.3ex}[0pt]{ \footnotesize} & \footnotesize Combination of identity mapping and Inception module, with similar computational cost of Inception v3, but faster training process. \\ \cline{1-4}\cline{6-7} \raisebox{-1.3ex}[0pt]{$12$} &\raisebox{-1.3ex}[0pt]{ \footnotesize InceptionResNet v2 \cite{InceptionV4}} &\raisebox{-1.3ex}[0pt]{ \footnotesize $30$ }& \raisebox{-1.3ex}[0pt]{\footnotesize $95$ }& \raisebox{-0.3ex}[0pt]{ \footnotesize (Ensemble) } &\raisebox{-1.3ex}[0pt]{ \footnotesize \cite{Huang2016Speed}} & \footnotesize A costlier residual version of Inception, with significantly improved recognition performance. \\ \cline{1-4}\cline{6-7} \raisebox{-1.3ex}[0pt]{$13 $}&\raisebox{-1.3ex}[0pt]{ \footnotesize Inception v4 \cite{InceptionV4}} & \raisebox{-1.3ex}[0pt]{\footnotesize $41$} &\raisebox{-1.3ex}[0pt]{ \footnotesize $75$ } & \raisebox{-1.3ex}[0pt]{\footnotesize} & \footnotesize & \footnotesize An Inception variant without residual connections, with roughly the same recognition performance as InceptionResNet v2, but significantly slower. \\ \hline \raisebox{-1.3ex}[0pt]{$14$} &\raisebox{-1.3ex}[0pt]{ \footnotesize ResNeXt \cite{ Xie2016Aggregated}} &\raisebox{-1.3ex}[0pt]{ \footnotesize $23 $}&\raisebox{-1.3ex}[0pt]{ \footnotesize $49 $} &\raisebox{-1.3ex}[0pt]{ \footnotesize $3.0\%$} &\raisebox{-1.3ex}[0pt]{ \footnotesize \cite{Xie2016Aggregated}}& \footnotesize Repeating a building block that aggregates a set of transformations with the same topology. \\ \hline \raisebox{-3.3ex}[0pt]{$15$} & \raisebox{-3.3ex}[0pt]{\footnotesize DenseNet201 \cite{Huang2016Densely}} &\raisebox{-3.3ex}[0pt]{ \footnotesize $18$} & \raisebox{-3.3ex}[0pt]{ \footnotesize$ 200$ } &\raisebox{-3.3ex}[0pt]{ \footnotesize $-$} & \raisebox{-3.3ex}[0pt]{\footnotesize \cite{Zhou2018Scale}} & \footnotesize Concatenate each layer with every other layer in a feed forward fashion. Alleviate the vanishing gradient problem, encourage feature reuse, reduction in number of parameters.\\ \hline $16$ & \footnotesize DarkNet \cite{YOLO9000} & \footnotesize $20$ & \footnotesize $19$ & \footnotesize $-$& \footnotesize \cite{YOLO9000} & \footnotesize Similar to VGGNet, but with significantly fewer parameters. \\ \hline \raisebox{-1.3ex}[0pt]{$17$} &\raisebox{-1.3ex}[0pt]{ \footnotesize MobileNet \cite{Howard2017MobileNets}} &\raisebox{-1.3ex}[0pt]{ \footnotesize $3.2$ }& \raisebox{-1.3ex}[0pt]{\footnotesize $27+1$ }& \raisebox{-1.3ex}[0pt]{\footnotesize $-$ } & \raisebox{-1.3ex}[0pt]{ \footnotesize \cite{Howard2017MobileNets}} & \footnotesize Light weight deep CNNs using depth-wise separable convolutions. \\ \hline \raisebox{-3.3ex}[0pt]{$18$}&\raisebox{-3.3ex}[0pt]{ \footnotesize SE ResNet \cite{ Hu2018Squeeze} }& \raisebox{-3.3ex}[0pt]{\footnotesize $26$} &\raisebox{-3.3ex}[0pt]{ \footnotesize $50$} &\raisebox{-3.3ex}[0pt]{ \shortstack [c] {$2.3\%$ \\ (SENets)} }& \raisebox{-3.3ex}[0pt]{\footnotesize \cite{ Hu2018Squeeze} }& \footnotesize Channel-wise attention by a novel block called \emph{Squeeze and Excitation}. Complementary to existing backbone CNNs. \\ \Xhline{1pt} \end{tabular} } \end{table*} \subsection{Popular CNN Architectures} \label{Sec:PopularNetworks} CNN architectures (Section \ref{Sec:CNNintro}) serve as network backbones used in the detection frameworks of Section~\ref{Sec:Frameworks}. Representative frameworks include AlexNet \cite{AlexNet2012}, ZFNet \cite{ZeilerFergus2014} VGGNet \cite{Simonyan2014VGG}, GoogLeNet \cite{GoogLeNet2015}, Inception series \cite{Ioffe2015,Szegedy2016a,InceptionV4}, ResNet \cite{He2016ResNet}, DenseNet \cite{Huang2016Densely} and SENet \cite{ Hu2018Squeeze}, summarized in Table \ref{Tab:dcnnarchitectures}, and where the improvement over time is seen in Fig.~\ref{fig:ILSVRCclassificationResults}. A further review of recent CNN advances can be found in \cite{Gu2015Recent}. The trend in architecture evolution is for greater depth: AlexNet has 8 layers, VGGNet 16 layers, more recently ResNet and DenseNet both surpassed the 100 layer mark, and it was VGGNet \cite{Simonyan2014VGG} and GoogLeNet \cite{GoogLeNet2015} which showed that increasing depth can improve the representational power. As can be observed from Table~\ref{Tab:dcnnarchitectures}, networks such as AlexNet, OverFeat, ZFNet and VGGNet have an enormous number of parameters, despite being only a few layers deep, since a large fraction of the parameters come from the FC layers. Newer networks like Inception, ResNet, and DenseNet, although having a great depth, actually have far fewer parameters by avoiding the use of FC layers. With the use of Inception modules \cite{GoogLeNet2015} in carefully designed topologies, the number of parameters of GoogLeNet is dramatically reduced, compared to AlexNet, ZFNet or VGGNet. Similarly, ResNet demonstrated the effectiveness of skip connections for learning extremely deep networks with hundreds of layers, winning the ILSVRC 2015 classification task. Inspired by ResNet \cite{He2016ResNet}, InceptionResNets \cite{InceptionV4} combined the Inception networks with shortcut connections, on the basis that shortcut connections can significantly accelerate network training. Extending ResNets, Huang \emph{et al.} \cite{Huang2016Densely} proposed DenseNets, which are built from dense blocksconnecting each layer to every other layer in a feedforward fashion, leading to compelling advantages such as parameter efficiency, implicit deep supervision\footnote{DenseNets perform deep supervision in an implicit way, \emph{i.e.} individual layers receive additional supervision from other layers through the shorter connections. The benefits of deep supervision have previously been demonstrated in Deeply Supervised Nets (DSN) \cite{Lee2015Deeply}.}, and feature reuse. Recently, Hu \emph{et al.} \cite{He2016ResNet} proposed Squeeze and Excitation (SE) blocks, which can be combined with existing deep architectures to boost their performance at minimal additional computational cost, adaptively recalibrating channel-wise feature responses by explicitly modeling the interdependencies between convolutional feature channels, and which led to winning the ILSVRC 2017 classification task. Research on CNN architectures remains active, with emerging networks such as Hourglass \cite{Law2018CornerNet}, Dilated Residual Networks \cite{Yu2017Dilated}, Xception \cite{Chollet2017Xception}, DetNet \cite{Li2018DetNet}, Dual Path Networks (DPN) \cite{Chen2017Dual}, FishNet \cite{Sun2018Fishnet}, and GLoRe \cite{Chen2019Graph}. The training of a CNN requires a large-scale labeled dataset with intraclass diversity. Unlike image classification, detection requires localizing (possibly many) objects from an image. It has been shown \cite{Ouyang2016} that pretraining a deep model with a large scale dataset having object level annotations (such as ImageNet), instead of only the image level annotations, improves the detection performance. However, collecting bounding box labels is expensive, especially for hundreds of thousands of categories. A common scenario is for a CNN to be pretrained on a large dataset (usually with a large number of visual categories) with image-level labels; the pretrained CNN can then be applied to a small dataset, directly, as a generic feature extractor \cite{Razavian2014,Azizpour2016,Donahue2014DeCAF,Yosinski2014Transferable}, which can support a wider range of visual recognition tasks. For detection, the pre-trained network is typically fine-tuned\footnote{Fine-tuning is done by initializing a network with weights optimized for a large labeled dataset like ImageNet. and then updating the network's weights using the target-task training set.} on a given detection dataset \cite{Donahue2014DeCAF,Girshick2014RCNN,Girshick2016TPAMI}. Several large scale image classification datasets are used for CNN pre-training, among them ImageNet1000 \cite{ImageNet2009,Russakovsky2015} with 1.2 million images of 1000 object categories, Places \cite{Zhou2017Places}, which is much larger than ImageNet1000 but with fewer classes, a recent Places-Imagenet hybrid \cite{Zhou2017Places}, or JFT300M \cite{Hinton2015Distilling,Sun2017Revisiting}. Pretrained CNNs without fine-tuning were explored for object classification and detection in \cite{Donahue2014DeCAF,Girshick2016TPAMI,Agrawal2014}, where it was shown that detection accuracies are different for features extracted from different layers; for example, for AlexNet pre-trained on ImageNet, FC6 / FC7 / Pool5 are in descending order of detection accuracy \cite{Donahue2014DeCAF,Girshick2016TPAMI}. Fine-tuning a pre-trained network can increase detection performance significantly \cite{Girshick2014RCNN,Girshick2016TPAMI}, although in the case of AlexNet, the fine-tuning performance boost was shown to be much larger for FC6 / FC7 than for Pool5, suggesting that Pool5 features are more general. Furthermore, the relationship between the source and target datasets plays a critical role, for example that ImageNet based CNN features show better performance for object detection than for human action \cite{Zhoubolei2014,Azizpour2016}. \begin{table*}[!t] \caption {Summary of properties of representative methods in improving DCNN feature representations for generic object detection. Details for Groups (1), (2), and (3) are provided in Section \ref{Sec:EnhanceFeatures}. Abbreviations: Selective Search (SS), EdgeBoxes (EB), InceptionResNet (IRN). \emph{Conv-Deconv} denotes the use of upsampling and convolutional layers with lateral connections to supplement the standard backbone network. Detection results on VOC07, VOC12 and COCO were reported with mAP@IoU=0.5, and the additional COCO results are computed as the average of mAP for IoU thresholds from 0.5 to 0.95. Training data: ``07''$\leftarrow$VOC2007 trainval; ``07T''$\leftarrow$VOC2007 trainval and test; ``12''$\leftarrow$VOC2012 trainval; CO$\leftarrow$ COCO trainval. The COCO detection results were reported with COCO2015 Test-Dev, except for MPN \cite{Zagoruyko2016} which reported with COCO2015 Test-Standard.}\label{Tab:EnhanceFeatures} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{1pt} \resizebox*{18.5cm}{!}{ \begin{tabular}{!{\vrule width1.2bp}c|c|c|c|c|c|c|c|c|c|p{8cm}<{\centering}!{\vrule width1.2bp}} \Xhline{1.5pt} \footnotesize & \footnotesize Detector & \footnotesize Region & \footnotesize Backbone & \footnotesize Pipelined & \multicolumn{3}{c|}{mAP@IoU=0.5} & \footnotesize mAP & \footnotesize Published & \footnotesize \\ \cline{6-9} \footnotesize Group & \footnotesize Name & \footnotesize Proposal & \footnotesize DCNN & \footnotesize Used & \footnotesize VOC07 & \footnotesize VOC12 & \footnotesize COCO & \footnotesize COCO & \footnotesize In & \footnotesize Highlights \\ \Xhline{1.5pt} \footnotesize \multirow{3}{*}{\rotatebox{90}{\scriptsize \shortstack [c] {\textbf{(1) Single detection }\\ \textbf{with multilayer features}$\quad$} }} &\footnotesize \raisebox{-3.5ex}[0pt]{ION \cite{Bell2016ION}} & \footnotesize \raisebox{-4.5ex}[0pt]{ \shortstack [c] {SS+EB\\MCG+RPN} }& \footnotesize \raisebox{-3.5ex}[0pt]{VGG16} & \footnotesize \raisebox{-4.5ex}[0pt]{ \shortstack [c] {Fast \\ RCNN}} & \footnotesize \raisebox{-4.5ex}[0pt]{ \shortstack [c] {$79.4$\\(07+12)}} & \footnotesize \raisebox{-4.5ex}[0pt]{\shortstack [c] {$76.4$\\(07+12)}} & \footnotesize \raisebox{-3.5ex}[0pt]{$55.7$}& \footnotesize \raisebox{-3.5ex}[0pt]{$33.1$} & \footnotesize \raisebox{-3.5ex}[0pt]{CVPR16}& \footnotesize Use features from multiple layers; use spatial recurrent neural networks for modeling contextual information; the Best Student Entry and the $3^{\textrm{rd}}$ overall in the COCO detection challenge 2015. \\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{HyperNet \cite{HyperNet2016}} & \footnotesize\raisebox{-2ex}[0pt]{ RPN }& \footnotesize \raisebox{-2ex}[0pt]{ VGG16} & \footnotesize \raisebox{-2.5ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \footnotesize \raisebox{-2.5ex}[0pt]{ \shortstack [c] {$76.3$\\(07+12)}} & \footnotesize \raisebox{-2.5ex}[0pt]{ \shortstack [c] {$71.4$\\(07T+12)}} & \footnotesize \raisebox{-2ex}[0pt]{$-$}& \footnotesize \raisebox{-2ex}[0pt]{$-$} & \footnotesize \raisebox{-2ex}[0pt]{CVPR16 } & \footnotesize Use features from multiple layers for both region proposal and region classification. \\ \cline{2-11} & \footnotesize \raisebox{-3ex}[0pt]{ PVANet \cite{PVANET2016}} & \footnotesize \raisebox{-3ex}[0pt]{RPN} & \footnotesize \raisebox{-3ex}[0pt]{PVANet} & \footnotesize \raisebox{-3.5ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \footnotesize \raisebox{-3.5ex}[0pt]{ \shortstack [c] {$\textbf{84.9}$\\(07+12+CO)}}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] { $\textbf{84.2}$\\(07T+12+CO)}}& \footnotesize \raisebox{-3ex}[0pt]{ $-$}& \footnotesize \raisebox{-3ex}[0pt]{ $-$} & \footnotesize \raisebox{-3ex}[0pt]{NIPSW16} & \footnotesize Deep but lightweight; Combine ideas from concatenated ReLU \cite{Shang2016Understanding}, Inception \cite{GoogLeNet2015}, and HyperNet \cite{HyperNet2016}. \\ \Xhline{1.5pt} \footnotesize \multirow{5}{*}{\rotatebox{90}{\scriptsize \shortstack [c] {\textbf{(2) Detection at multiple layers}$\quad\quad\quad\quad\quad\quad$} }} & \footnotesize \raisebox{-5ex}[0pt]{SDP+CRC \cite{Yang2016Exploit}} & \footnotesize \raisebox{-5ex}[0pt]{EB} & \footnotesize \raisebox{-5ex}[0pt]{VGG16 }& \footnotesize \raisebox{-6ex}[0pt]{ \shortstack [c] {Fast \\ RCNN}} & \footnotesize \raisebox{-6ex}[0pt]{ \shortstack [c] { $69.4$\\(07)}}& \footnotesize \raisebox{-5ex}[0pt]{$-$} & \footnotesize\raisebox{-5ex}[0pt]{ $-$}& \footnotesize\raisebox{-5ex}[0pt]{ $-$} & \footnotesize \raisebox{-5ex}[0pt]{CVPR16} & \footnotesize Use features in multiple layers to reject easy negatives via CRC, and then classify remaining proposals using SDP. \\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{MSCNN \cite{MSCNN2016}} & \footnotesize \raisebox{-2ex}[0pt]{RPN} & \footnotesize \raisebox{-2ex}[0pt]{VGG }& \footnotesize \raisebox{-2.5ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \multicolumn{4}{c|}{\raisebox{-2ex}[0pt]{\footnotesize Only Tested on KITTI}} & \footnotesize\raisebox{-2ex}[0pt]{ ECCV16 } & \footnotesize Region proposal and classification are performed at multiple layers; includes feature upsampling; end to end learning. \\ \cline{2-11} & \footnotesize \raisebox{-5ex}[0pt]{MPN \cite{Zagoruyko2016}} & \footnotesize \raisebox{-5ex}[0pt]{SharpMask \cite{Pinheiro2016}} & \footnotesize \raisebox{-5ex}[0pt]{VGG16} & \footnotesize \raisebox{-6ex}[0pt]{ \shortstack [c] {Fast \\ RCNN}} & \footnotesize \raisebox{-5ex}[0pt]{$-$} & \footnotesize \raisebox{-5ex}[0pt]{$-$}& \footnotesize \raisebox{-5ex}[0pt]{$51.9$} & \footnotesize \raisebox{-5ex}[0pt]{$33.2$} & \footnotesize \raisebox{-5ex}[0pt]{BMVC16} & \footnotesize Concatenate features from different convolutional layers and features of different contextual regions; loss function for multiple overlap thresholds; ranked $2^{\textrm{nd}}$ in both the COCO15 detection and segmentation challenges. \\ \cline{2-11} & \footnotesize\raisebox{-2ex}[0pt]{ DSOD \cite{ShenICCV2017}} & \footnotesize \raisebox{-2ex}[0pt]{ Free} & \footnotesize \raisebox{-2ex}[0pt]{DenseNet} & \footnotesize \raisebox{-2ex}[0pt]{SSD} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $77.7$\\(07+12)}} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $72.2$\\(07T+12)}}& \footnotesize \raisebox{-2ex}[0pt]{ $47.3$ } & \footnotesize \raisebox{-2ex}[0pt]{ $29.3$ } & \footnotesize \raisebox{-2ex}[0pt]{ICCV17} & \footnotesize Concatenate feature sequentially, like DenseNet. Train from scratch on the target dataset without pre-training. \\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{RFBNet \cite{Liu2017Receptive}} & \footnotesize \raisebox{-2ex}[0pt]{Free} & \footnotesize\raisebox{-2ex}[0pt]{ VGG16 }& \footnotesize \raisebox{-2ex}[0pt]{SSD} & \footnotesize\raisebox{-3ex}[0pt]{ \shortstack [c] { $82.2$\\(07+12)}}& \footnotesize \raisebox{-3ex}[0pt]{\shortstack [c] { $81.2$\\(07T+12)}} & \footnotesize \raisebox{-2ex}[0pt]{ $55.7$} & \footnotesize \raisebox{-2ex}[0pt]{ $34.4$} & \footnotesize \raisebox{-2ex}[0pt]{ ECCV18} & \footnotesize Propose a multi-branch convolutional block similar to Inception \cite{GoogLeNet2015}, but using dilated convolution. \\ \Xhline{1.5pt} \footnotesize \multirow{8}{*}{\rotatebox{90}{\scriptsize \textbf{(3) Combination of (1) and (2) $\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad$} }} & \footnotesize \raisebox{-2ex}[0pt]{DSSD \cite{DSSD2016}} & \footnotesize\raisebox{-2ex}[0pt]{ Free} & \footnotesize \raisebox{-2ex}[0pt]{ResNet101} & \footnotesize \raisebox{-2ex}[0pt]{SSD }& \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $81.5$\\(07+12)}} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $80.0$\\(07T+12)}}& \footnotesize \raisebox{-2ex}[0pt]{$53.3$} & \footnotesize \raisebox{-2ex}[0pt]{$33.2$} & \footnotesize \raisebox{-2ex}[0pt]{2017 } & \footnotesize \raisebox{-2ex}[0pt]{Use Conv-Deconv, as shown in Fig. \ref{fig:MultiLayerCombine} (c1, c2).} \\ \cline{2-11} & \footnotesize\raisebox{-2ex}[0pt]{ FPN \cite{FPN2016}} & \footnotesize \raisebox{-2ex}[0pt]{RPN} & \footnotesize \raisebox{-2ex}[0pt]{ResNet101} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \footnotesize\raisebox{-2ex}[0pt]{ $-$}& \footnotesize \raisebox{-2ex}[0pt]{$-$}& \footnotesize\raisebox{-2ex}[0pt]{ $59.1$} & \footnotesize \raisebox{-2ex}[0pt]{$36.2$}& \footnotesize \raisebox{-2ex}[0pt]{CVPR17} & \footnotesize Use Conv-Deconv, as shown in Fig. \ref{fig:MultiLayerCombine} (a1, a2); Widely used in detectors. \\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{ TDM \cite{Shrivastava2017} }& \footnotesize \raisebox{-2ex}[0pt]{RPN} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] {ResNet101\\ VGG16} }& \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \footnotesize \raisebox{-2ex}[0pt]{$-$} & \footnotesize \raisebox{-2ex}[0pt]{$-$} & \footnotesize \raisebox{-2ex}[0pt]{ $57.7$} & \footnotesize \raisebox{-2ex}[0pt]{$36.8$} & \footnotesize \raisebox{-2ex}[0pt]{CVPR17} & \footnotesize Use Conv-Deconv, as shown in Fig. \ref{fig:MultiLayerCombine} (b2). \\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{RON \cite{Kong2017ron}} & \footnotesize \raisebox{-2ex}[0pt]{RPN} & \footnotesize \raisebox{-2ex}[0pt]{VGG16} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $81.3$\\(07+12+CO)}} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $80.7$\\(07T+12+CO)}} & \footnotesize \raisebox{-2ex}[0pt]{$49.5$} & \footnotesize \raisebox{-2ex}[0pt]{$27.4$} & \footnotesize \raisebox{-2ex}[0pt]{CVPR17} & \footnotesize Use Conv-deconv, as shown in Fig. \ref{fig:MultiLayerCombine} (d2); Add the objectness prior to significantly reduce object search space. \\ \cline{2-11} & \footnotesize \raisebox{-3ex}[0pt]{ZIP \cite{Hongyang2018Zoom}} & \footnotesize \raisebox{-3ex}[0pt]{RPN} & \footnotesize \raisebox{-3ex}[0pt]{Inceptionv2} & \footnotesize \raisebox{-4ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \footnotesize \raisebox{-4ex}[0pt]{ \shortstack [c] {$79.8$\\ (07+12)}} & \footnotesize \raisebox{-3ex}[0pt]{$-$} & \footnotesize \raisebox{-3ex}[0pt]{$-$} & \footnotesize\raisebox{-3ex}[0pt]{ $-$ }& \footnotesize\raisebox{-3ex}[0pt]{ IJCV18 } & \footnotesize Use Conv-Deconv, as shown in Fig. \ref{fig:MultiLayerCombine} (f1). Propose a map attention decision (MAD) unit for features from different layers.\\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{STDN \cite{Zhou2018Scale} } & \footnotesize \raisebox{-2ex}[0pt]{Free} & \footnotesize \raisebox{-2ex}[0pt]{DenseNet169} & \footnotesize \raisebox{-2ex}[0pt]{SSD} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] {$80.9$\\(07+12)}} & \footnotesize\raisebox{-2ex}[0pt]{ $-$ }& \footnotesize \raisebox{-2ex}[0pt]{$51.0$ }& \footnotesize \raisebox{-2ex}[0pt]{$31.8$ }& \footnotesize \raisebox{-2ex}[0pt]{ CVPR18} & \footnotesize A new scale transfer module, which resizes features of different scales to the same scale in parallel. \\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{RefineDet \cite{Zhang2018Single}} & \footnotesize \raisebox{-2ex}[0pt]{RPN} & \footnotesize \raisebox{-3ex}[0pt]{\shortstack [c] {VGG16\\ResNet101}} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $83.8$\\(07+12)}} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] {$83.5$\\(07T+12)}} & \footnotesize\raisebox{-2ex}[0pt]{ $ 62.9$} & \footnotesize \raisebox{-2ex}[0pt]{$41.8$} & \footnotesize \raisebox{-2ex}[0pt]{CVPR18} & \footnotesize Use cascade to obtain better and less anchors. Use Conv-deconv, as shown in Fig. \ref{fig:MultiLayerCombine} (e2) to improve features. \\ \cline{2-11} & \footnotesize \raisebox{-5ex}[0pt]{PANet \cite{Liu2018Path} }& \footnotesize \raisebox{-5ex}[0pt]{RPN} & \footnotesize \raisebox{-6.5ex}[0pt]{\shortstack [c] {ResNeXt101\\+FPN}} & \footnotesize \raisebox{-5ex}[0pt]{Mask RCNN} & \footnotesize \raisebox{-5ex}[0pt]{ $-$ } & \footnotesize \raisebox{-5ex}[0pt]{ $-$ }& \footnotesize\raisebox{-5ex}[0pt]{ $\textbf{67.2}$ } & \footnotesize \raisebox{-5ex}[0pt]{$\textbf{47.4}$} & \footnotesize\raisebox{-5ex}[0pt]{ CVPR18}& \footnotesize Shown in Fig. \ref{fig:MultiLayerCombine} (g). Based on FPN, add another bottom-up path to pass information between lower and topmost layers; adaptive feature pooling. Ranked $1^{st}$ and $2^{nd}$ in COCO 2017 tasks. \\ \cline{2-11} & \footnotesize \raisebox{-1ex}[0pt]{DetNet \cite{Li2018DetNet}}& \footnotesize \raisebox{-1ex}[0pt]{RPN} & \footnotesize \raisebox{-1ex}[0pt]{DetNet59+FPN} & \footnotesize \raisebox{-1ex}[0pt]{Faster RCNN} & \footnotesize \raisebox{-1ex}[0pt]{ $-$ } & \footnotesize \raisebox{-1ex}[0pt]{ $-$ }& \footnotesize\raisebox{-1ex}[0pt]{ $61.7$ } & \footnotesize \raisebox{-1ex}[0pt]{$40.2$} & \footnotesize\raisebox{-1ex}[0pt]{ ECCV18}& \footnotesize Introduces dilated convolution into the ResNet backbone to maintain high resolution in deeper layers; Shown in Fig. \ref{fig:MultiLayerCombine} (i). \\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{FPR \cite{Kong2018Deep} }& \footnotesize \raisebox{-2ex}[0pt]{$-$} & \footnotesize \raisebox{-3ex}[0pt]{\shortstack [c] {VGG16\\ ResNet101}} & \footnotesize \raisebox{-2ex}[0pt]{SSD} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $82.4$\\(07+12)}} & \footnotesize \raisebox{-3ex}[0pt]{ \shortstack [c] { $81.1$\\(07T+12)}}& \footnotesize\raisebox{-2ex}[0pt]{$54.3$} & \footnotesize \raisebox{-2ex}[0pt]{ $34.6$} & \footnotesize\raisebox{-2ex}[0pt]{ ECCV18}& \footnotesize Fuse task oriented features across different spatial locations and scales, globally and locally; Shown in Fig. \ref{fig:MultiLayerCombine} (h).\\ \cline{2-11} & \footnotesize \raisebox{-5ex}[0pt]{M2Det \cite{Zhao2019M2Det}}& \footnotesize \raisebox{-5ex}[0pt]{$-$} & \footnotesize \raisebox{-5ex}[0pt]{SSD} & \footnotesize \raisebox{-6.5ex}[0pt]{\shortstack [c] {VGG16\\ ResNet101}} & \footnotesize \raisebox{-5ex}[0pt]{ $-$ } & \footnotesize \raisebox{-5ex}[0pt]{ $-$ } & \footnotesize\raisebox{-5ex}[0pt]{ $64.6$} & \footnotesize \raisebox{-5ex}[0pt]{$44.2$} & \footnotesize\raisebox{-5ex}[0pt]{ AAAI19}& \footnotesize Shown in Fig. \ref{fig:MultiLayerCombine} (j), newly designed top down path to learn a set of multilevel features, recombined to construct a feature pyramid for object detection. \\ \Xhline{1.5pt} \scriptsize \multirow{6}{*}{\rotatebox{90}{ \textbf{(4) Model Geometric Transforms$\quad$} }} & \footnotesize \raisebox{-8ex}[0pt]{DeepIDNet \cite{Ouyang2015deepid} }& \footnotesize \raisebox{-8ex}[0pt]{SS+ EB} & \footnotesize \raisebox{-12ex}[0pt]{\shortstack [c] {AlexNet \\ZFNet \\OverFeat \\GoogLeNet}} & \footnotesize \raisebox{-8ex}[0pt]{ RCNN}& \footnotesize \raisebox{-9ex}[0pt]{\shortstack [c] {$69.0$ \\(07)}} & \footnotesize \raisebox{-8ex}[0pt]{$-$}& \footnotesize\raisebox{-8ex}[0pt]{$-$}& \footnotesize \raisebox{-8ex}[0pt]{$25.6$}& \footnotesize \raisebox{-8ex}[0pt]{CVPR15} & \footnotesize Introduce a deformation constrained pooling layer, jointly learned with convolutional layers in existing DCNNs. Utilize the following modules that are not trained end to end: cascade, context modeling, model averaging, and bounding box location refinement in the multistage detection pipeline. \\ \cline{2-11} & \footnotesize \raisebox{-2ex}[0pt]{DCN \cite{Dai17Deformable}} & \footnotesize \raisebox{-2ex}[0pt]{RPN} & \footnotesize \raisebox{-2.5ex}[0pt]{\shortstack [c] {ResNet101\\IRN}} & \footnotesize\raisebox{-2ex}[0pt]{ RFCN} & \footnotesize \raisebox{-2.5ex}[0pt]{ \shortstack [c] {$82.6$ \\(07+12)}}& \footnotesize\raisebox{-2ex}[0pt]{ $-$ } & \footnotesize \raisebox{-2ex}[0pt]{ $58.0$} & \footnotesize\raisebox{-2ex}[0pt]{ $37.5$} & \footnotesize \raisebox{-2ex}[0pt]{CVPR17} & \footnotesize Design deformable convolution and deformable RoI pooling modules that can replace plain convolution in existing DCNNs. \\ \cline{2-11} & \footnotesize \raisebox{-2.5ex}[0pt]{DPFCN \cite{Mordan2018End}} & \footnotesize \raisebox{-2.5ex}[0pt]{AttractioNet \cite{Gidaris2016Attend} }& \footnotesize \raisebox{-2.5ex}[0pt]{ ResNet} & \footnotesize \raisebox{-2.5ex}[0pt]{ RFCN} & \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$83.3$ \\(07+12)}}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$81.2$\\(07T+12)}} & \footnotesize \raisebox{-2.5ex}[0pt]{$59.1$}& \footnotesize\raisebox{-2.5ex}[0pt]{$39.1$} & \footnotesize \raisebox{-2.5ex}[0pt]{IJCV18}& \footnotesize Design a deformable part based RoI pooling layer to explicitly select discriminative regions around object proposals. \\ \Xhline{1.5pt} \end{tabular} } \end{table*} \subsection{Methods For Improving Object Representation} \label{Sec:EnhanceFeatures} Deep CNN based detectors such as RCNN \cite{Girshick2014RCNN}, Fast RCNN \cite{Girshick2015FRCNN}, Faster RCNN \cite{Ren2015NIPS} and YOLO \cite{YoLo2016}, typically use the deep CNN architectures listed in Table \ref{Tab:dcnnarchitectures} as the backbone network and use features from the top layer of the CNN as object representations; however, detecting objects across a large {\em range} of scales is a fundamental challenge. A classical strategy to address this issue is to run the detector over a number of scaled input images (\emph{e.g.,} an image pyramid) \cite{Felzenszwalb2010b,Girshick2014RCNN,He2014SPP}, which typically produces more accurate detection, with, however, obvious limitations of inference time and memory. \subsubsection{Handling of Object Scale Variations} \label{sec:objectscale} Since a CNN computes its feature hierarchy layer by layer, the sub-sampling layers in the feature hierarchy already lead to an inherent multiscale pyramid, producing feature maps at different spatial resolutions, but subject to challenges \cite{Hariharan2016,FCNCVPR2015,Shrivastava2017}. In particular, the higher layers have a large receptive field and strong semantics, and are the most robust to variations such as object pose, illumination and part deformation, but the resolution is low and the geometric details are lost. In contrast, lower layers have a small receptive field and rich geometric details, but the resolution is high and much less sensitive to semantics. Intuitively, semantic concepts of objects can emerge in different layers, depending on the size of the objects. So if a target object is small it requires fine detail information in earlier layers and may very well disappear at later layers, in principle making small object detection very challenging, for which tricks such as dilated or ``atrous'' convolution \cite{Yu2015Multiscale,Dai2016RFCN,Chen2016deeplab} have been proposed, increasing feature resolution, but increasing computational complexity. On the other hand, if the target object is large, then the semantic concept will emerge in much later layers. A number of methods \cite{Shrivastava2017,Zhang2018Object,FPN2016,Kong2017ron} have been proposed to improve detection accuracy by exploiting multiple CNN layers, broadly falling into three types of \textbf{multiscale object detection}: \begin{enumerate} \item Detecting with combined features of multiple layers; \item Detecting at multiple layers; \item Combinations of the above two methods. \end{enumerate} \begin {figure}[!t] \centering \includegraphics[width=0.48\textwidth]{HyperFeature.pdf} \caption{Comparison of HyperNet and ION. LRN is Local Response Normalization, which performs a kind of ``lateral inhibition'' by normalizing over local input regions \cite{Jia2014Caffe}.} \label{fig:HyperFeature} \end {figure} \textbf{(1) Detecting with combined features of multiple CNN layers:} Many approaches, including Hypercolumns \cite{Hariharan2016}, HyperNet \cite{HyperNet2016}, and ION \cite{Bell2016ION}, combine features from multiple layers before making a prediction. Such feature combination is commonly accomplished via concatenation, a classic neural network idea that concatenates features from different layers, architectures which have recently become popular for semantic segmentation \cite{FCNCVPR2015,FCNTPAMI,Hariharan2016}. As shown in Fig.~\ref{fig:HyperFeature} (a), ION \cite{Bell2016ION} uses RoI pooling to extract RoI features from multiple layers, and then the object proposals generated by selective search and edgeboxes are classified by using the concatenated features. HyperNet \cite{HyperNet2016}, shown in Fig.~\ref{fig:HyperFeature} (b), follows a similar idea, and integrates deep, intermediate and shallow features to generate object proposals and to predict objects via an end to end joint training strategy. The combined feature is more descriptive, and is more beneficial for localization and classification, but at increased computational complexity. \begin {figure*}[!t] \centering \includegraphics[height=0.9\textheight]{MultiLayerCombine.pdf} \caption{Hourglass architectures: Conv1 to Conv5 are the main Conv blocks in backbone networks such as VGG or ResNet. The figure compares a number of Feature Fusion Blocks (FFB) commonly used in recent approaches: FPN \cite{FPN2016}, TDM \cite{Shrivastava2017}, DSSD \cite{DSSD2016}, RON \cite{Kong2017ron}, RefineDet \cite{Zhang2018Single}, ZIP \cite{Hongyang2018Zoom}, PANet \cite{Liu2018Path}, FPR \cite{Kong2018Deep}, DetNet \cite{Li2018DetNet} and M2Det \cite{Zhao2019M2Det}. FFM: Feature Fusion Module, TUM: Thinned U-shaped Module} \label{fig:MultiLayerCombine} \end {figure*} \textbf{(2) Detecting at multiple CNN layers:} A number of recent approaches improve detection by predicting objects of different resolutions at different layers and then combining these predictions: SSD \cite{Liu2016SSD} and MSCNN \cite{MSCNN2016}, RBFNet \cite{Liu2017Receptive}, and DSOD \cite{ShenICCV2017}. SSD \cite{Liu2016SSD} spreads out default boxes of different scales to multiple layers within a CNN, and forces each layer to focus on predicting objects of a certain scale. RFBNet \cite{Liu2017Receptive} replaces the later convolution layers of SSD with a Receptive Field Block (RFB) to enhance the discriminability and robustness of features. The RFB is a multibranch convolutional block, similar to the Inception block \cite{GoogLeNet2015}, but combining multiple branches with different kernels and convolution layers \cite{Chen2016deeplab}. MSCNN \cite{MSCNN2016} applies deconvolution on multiple layers of a CNN to increase feature map resolution before using the layers to learn region proposals and pool features. Similar to RFBNet \cite{Liu2017Receptive}, TridentNet \cite{Li2019Scale} constructs a parallel multibranch architecture where each branch shares the same transformation parameters but with different receptive fields; dilated convolution with different dilation rates are used to adapt the receptive fields for objects of different scales. \textbf{(3) Combinations of the above two methods:} Features from different layers are complementary to each other and can improve detection accuracy, as shown by Hypercolumns \cite{Hariharan2016}, HyperNet \cite{HyperNet2016} and ION \cite{Bell2016ION}. On the other hand, however, it is natural to detect objects of different scales using features of approximately the same size, which can be achieved by detecting large objects from downscaled feature maps while detecting small objects from upscaled feature maps. Therefore, in order to combine the best of both worlds, some recent works propose to detect objects at multiple layers, and the resulting features obtained by combining features from different layers. This approach has been found to be effective for segmentation \cite{FCNCVPR2015,FCNTPAMI} and human pose estimation \cite{Newell2016Stacked}, has been widely exploited by both one-stage and two-stage detectors to alleviate problems of scale variation across object instances. Representative methods include SharpMask \cite{Pinheiro2016}, Deconvolutional Single Shot Detector (DSSD) \cite{DSSD2016}, Feature Pyramid Network (FPN) \cite{FPN2016}, Top Down Modulation (TDM)\cite{Shrivastava2017}, Reverse connection with Objectness prior Network (RON) \cite{Kong2017ron}, ZIP \cite{Hongyang2018Zoom}, Scale Transfer Detection Network (STDN) \cite{Zhou2018Scale}, RefineDet \cite{Zhang2018Single}, StairNet \cite{Woo18StairNet}, Path Aggregation Network (PANet) \cite{Liu2018Path}, Feature Pyramid Reconfiguration (FPR) \cite{Kong2018Deep}, DetNet \cite{Li2018DetNet}, Scale Aware Network (SAN) \cite{Kim2018San}, Multiscale Location aware Kernel Representation (MLKP) \cite{Wang2018Multiscale} and M2Det \cite{Zhao2019M2Det}, as shown in Table~\ref{Tab:EnhanceFeatures} and contrasted in Fig.~\ref{fig:MultiLayerCombine}. Early works like FPN \cite{FPN2016}, DSSD \cite{DSSD2016}, TDM \cite{Shrivastava2017}, ZIP \cite{Hongyang2018Zoom}, RON \cite{Kong2017ron} and RefineDet \cite{Zhang2018Single} construct the feature pyramid according to the inherent multiscale, pyramidal architecture of the backbone, and achieved encouraging results. As can be observed from Fig. \ref{fig:MultiLayerCombine} (a1) to (f1), these methods have very similar detection architectures which incorporate a top-down network with lateral connections to supplement the standard bottom-up, feed-forward network. Specifically, after a bottom-up pass the final high level semantic features are transmitted back by the top-down network to combine with the bottom-up features from intermediate layers after lateral processing, and the combined features are then used for detection. As can be seen from Fig.~\ref{fig:MultiLayerCombine} (a2) to (e2), the main differences lie in the design of the simple Feature Fusion Block (FFB), which handles the selection of features from different layers and the combination of multilayer features. FPN \cite{FPN2016} shows significant improvement as a generic feature extractor in several applications including object detection \cite{FPN2016,LinICCV2017} and instance segmentation \cite{MaskRCNN2017}. Using FPN in a basic Faster RCNN system achieved state-of-the-art results on the COCO detection dataset. STDN \cite{Zhou2018Scale} used DenseNet \cite{Huang2016Densely} to combine features of different layers and designed a scale transfer module to obtain feature maps with different resolutions. The scale transfer module can be directly embedded into DenseNet with little additional cost. More recent work, such as PANet \cite{Liu2018Path}, FPR \cite{Kong2018Deep}, DetNet \cite{Li2018DetNet}, and M2Det \cite{Zhao2019M2Det}, as shown in Fig. \ref{fig:MultiLayerCombine} (g-j), propose to further improve on the pyramid architectures like FPN in different ways. Based on FPN, Liu \emph{et al.} designed PANet \cite{Liu2018Path} (Fig.~\ref{fig:MultiLayerCombine} (g1)) by adding another bottom-up path with clean lateral connections from low to top levels, in order to shorten the information path and to enhance the feature pyramid. Then, an adaptive feature pooling was proposed to aggregate features from all feature levels for each proposal. In addition, in the proposal sub-network, a complementary branch capturing different views for each proposal is created to further improve mask prediction. These additional steps bring only slightly extra computational overhead, but are effective and allowed PANet to reach 1st place in the COCO 2017 Challenge Instance Segmentation task and 2nd place in the Object Detection task. Kong \emph{et al.} proposed FPR \cite{Kong2018Deep} by explicitly reformulating the feature pyramid construction process (\emph{e.g.} FPN \cite{FPN2016}) as feature reconfiguration functions in a highly nonlinear but efficient way. As shown in Fig.~\ref{fig:MultiLayerCombine} (h1), instead of using a top-down path to propagate strong semantic features from the topmost layer down as in FPN, FPR first extracts features from multiple layers in the backbone network by adaptive concatenation, and then designs a more complex FFB module (Fig. \ref{fig:MultiLayerCombine} (h2)) to spread strong semantics to all scales. Li \emph{et al.} proposed DetNet \cite{Li2018DetNet} (Fig. \ref{fig:MultiLayerCombine} (i1)) by introducing dilated convolutions to the later layers of the backbone network in order to maintain high spatial resolution in deeper layers. Zhao \emph{et al.} \cite{Zhao2019M2Det} proposed a MultiLevel Feature Pyramid Network (MLFPN) to build more effective feature pyramids for detecting objects of different scales. As can be seen from Fig. \ref{fig:MultiLayerCombine} (j1), features from two different layers of the backbone are first fused as the base feature, after which a top-down path with lateral connections from the base feature is created to build the feature pyramid. As shown in Fig.~\ref{fig:MultiLayerCombine} (j2) and (j5), the FFB module is much more complex than those like FPN, in that FFB involves a Thinned U-shaped Module (TUM) to generate a second pyramid structure, after which the feature maps with equivalent sizes from multiple TUMs are combined for object detection. The authors proposed M2Det by integrating MLFPN into SSD, and achieved better detection performance than other one-stage detectors. \subsection{Handling of Other Intraclass Variations} \label{sec:Otherchanges} Powerful object representations should combine distinctiveness and robustness. A large amount of recent work has been devoted to handling changes in object scale, as reviewed in Section~\ref{sec:objectscale}. As discussed in Section~\ref{Sec:MainChallenges} and summarized in Fig.~\ref{Fig:challenges}, object detection still requires robustness to real-world variations other than just scale, which we group into three categories: \begin{itemize} \renewcommand{\labelitemi}{$\bullet$} \item Geometric transformations, \item Occlusions, and \item Image degradations. \end{itemize} To handle these intra-class variations, the most straightforward approach is to augment the training datasets with a sufficient amount of variations; for example, robustness to rotation could be achieved by adding rotated objects at many orientations to the training data. Robustness can frequently be learned this way, but usually at the cost of expensive training and complex model parameters. Therefore, researchers have proposed alternative solutions to these problems. \textbf{Handling of geometric transformations:} DCNNs are inherently limited by the lack of ability to be spatially invariant to geometric transformations of the input data \cite{Lenc2018Understanding,Liu2017Local,Chellappa2016}. The introduction of local max pooling layers has allowed DCNNs to enjoy some translation invariance, however the intermediate feature maps are not actually invariant to large geometric transformations of the input data \cite{Lenc2018Understanding}. Therefore, many approaches have been presented to enhance robustness, aiming at learning invariant CNN representations with respect to different types of transformations such as scale \cite{Kim2014Locally,Bruna13Invariant}, rotation \cite{Bruna13Invariant,RIFDCNN2016,Worrall2017Harmonic,Zhou2017Oriented}, or both \cite{Jaderberg2015Spatial}. One representative work is Spatial Transformer Network (STN) \cite{Jaderberg2015Spatial}, which introduces a new learnable module to handle scaling, cropping, rotations, as well as nonrigid deformations via a global parametric transformation. STN has now been used in rotated text detection \cite{Jaderberg2015Spatial}, rotated face detection and generic object detection \cite{Wang2017}. Although rotation invariance may be attractive in certain applications, such as scene text detection \cite{He2018End,Ma2018Arbitrary}, face detection \cite{Shi2018Real}, and aerial imagery \cite{Ding2018Learning,Xia2018DOTA}, there is limited generic object detection work focusing on rotation invariance because popular benchmark detection datasets (\emph{e.g.} PASCAL VOC, ImageNet, COCO) do not actually present rotated images. Before deep learning, Deformable Part based Models (DPMs) \cite{Felzenszwalb2010b} were successful for generic object detection, representing objects by component parts arranged in a deformable configuration. Although DPMs have been significantly outperformed by more recent object detectors, their spirit still deeply influences many recent detectors. DPM modeling is less sensitive to transformations in object pose, viewpoint and nonrigid deformations, motivating researchers \cite{Dai17Deformable,Girshick2015DPMCNN,Mordan2018End,Ouyang2015deepid,Wan2015end} to explicitly model object composition to improve CNN based detection. The first attempts \cite{Girshick2015DPMCNN,Wan2015end} combined DPMs with CNNs by using deep features learned by AlexNet in DPM based detection, but without region proposals. To enable a CNN to benefit from the built-in capability of modeling the deformations of object parts, a number of approaches were proposed, including DeepIDNet \cite{Ouyang2015deepid}, DCN \cite{Dai17Deformable} and DPFCN \cite{Mordan2018End} (shown in Table~\ref{Tab:EnhanceFeatures}). Although similar in spirit, deformations are computed in different ways: DeepIDNet \cite{Ouyang2016} designed a deformation constrained pooling layer to replace regular max pooling, to learn the shared visual patterns and their deformation properties across different object classes; DCN \cite{Dai17Deformable} designed a deformable convolution layer and a deformable RoI pooling layer, both of which are based on the idea of augmenting regular grid sampling locations in feature maps; and DPFCN \cite{Mordan2018End} proposed a deformable part-based RoI pooling layer which selects discriminative parts of objects around object proposals by simultaneously optimizing latent displacements of all parts. \textbf{Handling of occlusions:} In real-world images, occlusions are common, resulting in information loss from object instances. A deformable parts idea can be useful for occlusion handling, so deformable RoI Pooling \cite{Dai17Deformable,Mordan2018End,Ouyang2013Joint} and deformable convolution \cite{Dai17Deformable} have been proposed to alleviate occlusion by giving more flexibility to the typically fixed geometric structures. Wang \emph{et al.} \cite{Wang2017} propose to learn an adversarial network that generates examples with occlusions and deformations, and context may be helpful in dealing with occlusions \cite{Zhang2018Occluded}. Despite these efforts, the occlusion problem is far from being solved; applying GANs to this problem may be a promising research direction. \textbf{Handling of image degradations:} Image noise is a common problem in many real-world applications. It is frequently caused by insufficient lighting, low quality cameras, image compression, or the intentional low-cost sensors on edge devices and wearable devices. While low image quality may be expected to degrade the performance of visual recognition, most current methods are evaluated in a degradation free and clean environment, evidenced by the fact that PASCAL VOC, ImageNet, MS COCO and Open Images all focus on relatively high quality images. To the best of our knowledge, there is so far very limited work to address this problem. \section{Context Modeling} \label{sec:ContextInfo} \begin{table*}[!t] \caption {Summary of detectors that exploit context information, with labelling details as in Table \ref{Tab:EnhanceFeatures}.}\label{Tab:ContextMethods} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{1pt} \resizebox*{18.5cm}{!}{ \begin{tabular}{!{\vrule width1.5bp}c|c|c|c|c|c|c|c|c|p{8cm}<{\centering}!{\vrule width1.5bp}} \Xhline{1.5pt} \footnotesize & \footnotesize Detector & \footnotesize Region & \footnotesize Backbone & \footnotesize Pipelined & \multicolumn{2}{c|}{mAP@IoU=0.5} & \footnotesize mAP & \footnotesize Published & \footnotesize \\ \cline{6-8} \footnotesize Group & \footnotesize Name & \footnotesize Proposal & \footnotesize DCNN & \footnotesize Used & \footnotesize VOC07 & \footnotesize VOC12 & \footnotesize COCO & \footnotesize In & \footnotesize Highlights \\ \Xhline{1.5pt} \footnotesize \multirow{4}{*}{\rotatebox{90}{\scriptsize \textbf{Global Context} $\quad\quad\quad\quad\quad$}} & \raisebox{-1.5ex}[0pt]{\footnotesize SegDeepM \cite{SegDeepM2015} } &\raisebox{-1.5ex}[0pt]{ \footnotesize SS+CMPC} & \raisebox{-1.5ex}[0pt]{\footnotesize VGG16} & \raisebox{-1.5ex}[0pt]{\footnotesize RCNN} & \raisebox{-1.5ex}[0pt]{\footnotesize VOC10} &\raisebox{-1.5ex}[0pt]{ \footnotesize VOC12} & \raisebox{-1.5ex}[0pt]{\footnotesize $-$}&\raisebox{-1.5ex}[0pt]{ \footnotesize CVPR15 } & \footnotesize Additional features extracted from an enlarged object proposal as context information. \\ \cline{2-10} & \footnotesize \raisebox{-1.5ex}[0pt]{DeepIDNet \cite{Ouyang2015deepid}} & \footnotesize \raisebox{-1.5ex}[0pt]{SS+EB} &\raisebox{-2.5ex}[0pt]{ \footnotesize \shortstack [c] {AlexNet\\ZFNet}} & \footnotesize \raisebox{-1.5ex}[0pt]{ RCNN } & \footnotesize \raisebox{-2.5ex}[0pt]{\shortstack [c] {$69.0$ \\(07)}} & \footnotesize \raisebox{-1.5ex}[0pt]{$-$ } & \footnotesize \raisebox{-1.5ex}[0pt]{$-$} & \footnotesize \raisebox{-1.5ex}[0pt]{CVPR15} & \footnotesize Use image classification scores as global contextual information to refine the detection scores of each object proposal.\\ \cline{2-10} & \raisebox{-1.5ex}[0pt]{\footnotesize ION \cite{Bell2016ION} }& \raisebox{-1.5ex}[0pt]{ \footnotesize SS+EB} &\raisebox{-1.5ex}[0pt]{ \footnotesize VGG16} &\raisebox{-2.5ex}[0pt]{ \footnotesize\shortstack [c] { Fast \\RCNN}} & \raisebox{-1.5ex}[0pt]{\footnotesize $80.1$} & \raisebox{-1.5ex}[0pt]{\footnotesize $77.9$ } &\raisebox{-1.5ex}[0pt]{ \footnotesize $33.1$} &\raisebox{-1.5ex}[0pt]{ \footnotesize CVPR16 } & \footnotesize The contextual information outside the region of interest is integrated using spatial recurrent neural networks. \\ \cline{2-10} & \raisebox{-1ex}[0pt]{\footnotesize CPF \cite{ Shrivastava2016}} & \raisebox{-1ex}[0pt]{\footnotesize RPN} & \raisebox{-1ex}[0pt]{\footnotesize VGG16} &\raisebox{-2.5ex}[0pt]{ \footnotesize \shortstack [c] { Faster \\ RCNN} }&\raisebox{-2.5ex}[0pt]{ \footnotesize \shortstack [c] {$76.4$\\ (07+12)}} & \raisebox{-2ex}[0pt]{\footnotesize \shortstack [c] { $72.6$\\ (07T+12) }}& \raisebox{-1ex}[0pt]{\footnotesize $-$ }&\raisebox{-1ex}[0pt]{ \footnotesize ECCV16} & \footnotesize \raisebox{-1ex}[0pt]{Use semantic segmentation to provide top-down feedback. } \\ \Xhline{1.5pt} \footnotesize \multirow{7}{*}{\rotatebox{90}{\scriptsize \textbf{Local Context}$\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad$}} & \footnotesize \raisebox{-3ex}[0pt]{MRCNN \cite{Gidaris2015}} & \footnotesize \raisebox{-3ex}[0pt]{SS} & \footnotesize \raisebox{-3ex}[0pt]{VGG16} & \footnotesize \raisebox{-3ex}[0pt]{SPPNet} & \footnotesize \raisebox{-4ex}[0pt]{\shortstack [c] {$78.2$ \\(07+12)}}& \footnotesize \raisebox{-4ex}[0pt]{\shortstack [c] { $73.9$\\(07+12)} }& \footnotesize \raisebox{-3ex}[0pt]{$-$} & \footnotesize\raisebox{-3ex}[0pt]{ ICCV15 }& \footnotesize Extract features from multiple regions surrounding or inside the object proposals. Integrate the semantic segmentation-aware features. \\ \cline{2-10} & \raisebox{-6ex}[0pt]{\footnotesize GBDNet \cite{ GBDCNN2016, Zeng2017Crafting}} & \raisebox{-6ex}[0pt]{\footnotesize CRAFT \cite{ CRAFT2016}} &\raisebox{-7ex}[0pt]{ \footnotesize \shortstack [c] {Inception v2\\ResNet269\\PolyNet \cite{Zhang2017PolyNet} } }&\raisebox{-7ex}[0pt]{ \footnotesize \shortstack [c] {Fast \\ RCNN} }&\raisebox{-7ex}[0pt]{ \footnotesize\shortstack [c] { $77.2$\\ (07+12)} } &\raisebox{-6ex}[0pt]{ \footnotesize $-$} &\raisebox{-6ex}[0pt]{ \footnotesize $27.0$} &\raisebox{-6ex}[0pt]{ \footnotesize \shortstack [c] {ECCV16\\ TPAMI18} } & \footnotesize A GBDNet module to learn the relations of multiscale contextualized regions surrounding an object proposal; GBDNet passes messages among features from different context regions through convolution between neighboring support regions in two directions. \\ \cline{2-10} & \raisebox{-2.5ex}[0pt]{\footnotesize ACCNN\cite{Li2017Attentive}} & \raisebox{-2.5ex}[0pt]{\footnotesize SS }& \raisebox{-2.5ex}[0pt]{\footnotesize VGG16 }& \raisebox{-2.5ex}[0pt]{\footnotesize \shortstack [c] {Fast \\ RCNN} }&\raisebox{-2.5ex}[0pt]{ \footnotesize \shortstack [c] { $72.0$\\ (07+12) }}& \raisebox{-2.5ex}[0pt]{\footnotesize \shortstack [c] { $70.6$\\ (07T+12)} }&\raisebox{-5.5ex}[0pt]{ \footnotesize $-$ }&\raisebox{-2.5ex}[0pt]{ \footnotesize TMM17 }& \footnotesize Use LSTM to capture global context. Concatenate features from multi-scale contextual regions surrounding an object proposal. The global and local context features are concatenated for recognition. \\ \cline{2-10} & \raisebox{-2.5ex}[0pt]{\footnotesize CoupleNet\cite{ ZhuICCV2017}} &\raisebox{-2.5ex}[0pt]{ \footnotesize RPN} &\raisebox{-2.5ex}[0pt]{ \footnotesize ResNet101} & \raisebox{-2.5ex}[0pt]{\footnotesize RFCN }&\raisebox{-2.5ex}[0pt]{ \footnotesize \shortstack [c] { $\textbf{82.7}$\\(07+12)}} &\raisebox{-2.5ex}[0pt]{ \footnotesize\shortstack [c] { $\textbf{80.4}$ \\(07T+12)} }& \raisebox{-2.5ex}[0pt]{\footnotesize $34.4$}&\raisebox{-2.5ex}[0pt]{ \footnotesize ICCV17} & \footnotesize Concatenate features from multiscale contextual regions surrounding an object proposal. Features of different contextual regions are then combined by convolution and element-wise sum. \\ \cline{2-10} & \raisebox{-2.5ex}[0pt]{\footnotesize SMN \cite{ChenSpatial2017}} & \raisebox{-2.5ex}[0pt]{\footnotesize RPN} & \raisebox{-2.5ex}[0pt]{\footnotesize VGG16 } & \raisebox{-3.5ex}[0pt]{\footnotesize\shortstack [c] { Faster \\ RCNN} } &\raisebox{-3.5ex}[0pt]{ \footnotesize \shortstack [c] {$70.0$ \\(07)}} & \raisebox{-2.5ex}[0pt]{\footnotesize $-$} & \raisebox{-2.5ex}[0pt]{\footnotesize $-$}& \raisebox{-2.5ex}[0pt]{\footnotesize ICCV17} & \footnotesize Model object-object relationships efficiently through a spatial memory network. Learn the functionality of NMS automatically. \\ \cline{2-10} &\raisebox{-3.5ex}[0pt]{ \footnotesize ORN \cite{Hu2018Relation}} & \raisebox{-3.5ex}[0pt]{\footnotesize RPN} &\raisebox{-5ex}[0pt]{ \footnotesize \shortstack [c] {ResNet101\\+DCN}}& \raisebox{-5ex}[0pt]{\footnotesize \shortstack [c] { Faster \\ RCNN}}&\raisebox{-3.5ex}[0pt]{ \footnotesize $-$} & \raisebox{-3.5ex}[0pt]{\footnotesize $-$} & \raisebox{-3.5ex}[0pt]{\footnotesize $\textbf{39.0}$}&\raisebox{-3.5ex}[0pt]{ \footnotesize CVPR18} & \footnotesize Model the relations of a set of object proposals through the interactions between their appearance features and geometry. Learn the functionality of NMS automatically. \\ \cline{2-10} &\raisebox{-3.5ex}[0pt]{ \footnotesize SIN \cite{Liu2018Structure}} & \raisebox{-3.5ex}[0pt]{\footnotesize RPN} &\raisebox{-3.5ex}[0pt]{ \footnotesize VGG16}& \raisebox{-4.5ex}[0pt]{\footnotesize \shortstack [c] { Faster \\ RCNN}}&\raisebox{-4.5ex}[0pt]{ \footnotesize \shortstack [c] {$76.0$\\(07+12)}} & \raisebox{-4.5ex}[0pt]{\footnotesize\shortstack [c] { $73.1$\\(07T+12)}} & \raisebox{-3.5ex}[0pt]{\footnotesize \shortstack [c] {$23.2$}}&\raisebox{-3.5ex}[0pt]{ \footnotesize CVPR18} & \footnotesize Formulate object detection as graph-structured inference, where objects are graph nodes and relationships the edges.\\ \Xhline{1.5pt} \end{tabular} } \end{table*} \begin {figure*}[!t] \centering \includegraphics[width=0.9\textwidth]{LocalContext.pdf} \caption{Representative approaches that explore local surrounding contextual features: MRCNN \cite{Gidaris2015}, GBDNet \cite{ GBDCNN2016,Zeng2017Crafting}, ACCNN \cite{Li2017Attentive} and CoupleNet \cite{ZhuICCV2017}; also see Table~\ref{Tab:ContextMethods}.} \label{Fig:LocalContext} \end {figure*} In the physical world, visual objects occur in particular environments and usually coexist with other related objects. There is strong psychological evidence \cite{Biederman1972Contextual,Bar2004Visual} that context plays an essential role in human object recognition, and it is recognized that a proper modeling of context helps object detection and recognition \cite{Torralba2003,Oliva2007Role,Chen2016deeplab, Chen2015Semantic,Divvala2009,Galleguillos2010}, especially when object appearance features are insufficient because of small object size, object occlusion, or poor image quality. Many different types of context have been discussed \cite{Divvala2009,Galleguillos2010}, and can broadly be grouped into one of three categories: \begin{enumerate} \item Semantic context: The likelihood of an object to be found in some scenes, but not in others; \item Spatial context: The likelihood of finding an object in some position and not others with respect to other objects in the scene; \item Scale context: Objects have a limited set of sizes relative to other objects in the scene. \end{enumerate} A great deal of work \cite{Chen2015c,Divvala2009,Galleguillos2010,Malisiewicz09Beyond, Murphy03Using,Rabinovich2007Objects,Parikh2012} preceded the prevalence of deep learning, and much of this work has yet to be explored in DCNN-based object detectors \cite{ChenSpatial2017,Hu2018Relation}. The current state of the art in object detection \cite{Ren2015NIPS,Liu2016SSD,MaskRCNN2017} detects objects without explicitly exploiting any contextual information. It is broadly agreed that DCNNs make use of contextual information implicitly \cite{ZeilerFergus2014,Zheng15Conditional} since they learn hierarchical representations with multiple levels of abstraction. Nevertheless, there is value in exploring contextual information explicitly in DCNN based detectors \cite{Hu2018Relation,ChenSpatial2017,Zeng2017Crafting}, so the following reviews recent work in exploiting contextual cues in DCNN- based object detectors, organized into categories of {\em global} and {\em local} contexts, motivated by earlier work in \cite{Zhang13,Galleguillos2010}. Representative approaches are summarized in Table~\ref{Tab:ContextMethods}. \subsection{Global Context} Global context \cite{Zhang13,Galleguillos2010} refers to image or scene level contexts, which can serve as cues for object detection (\emph{e.g.,} a bedroom will predict the presence of a bed). In DeepIDNet \cite{Ouyang2015deepid}, the image classification scores were used as contextual features, and concatenated with the object detection scores to improve detection results. In ION \cite{Bell2016ION}, Bell \emph{et al.} proposed to use spatial Recurrent Neural Networks (RNNs) to explore contextual information across the entire image. In SegDeepM \cite{SegDeepM2015}, Zhu \emph{et al.} proposed a Markov random field model that scores appearance as well as context for each detection, and allows each candidate box to select a segment out of a large pool of object segmentation proposals and score the agreement between them. In \cite{Shrivastava2016}, semantic segmentation was used as a form of contextual priming. \subsection{Local Context} Local context \cite{Zhang13,Galleguillos2010,Rabinovich2007Objects} considers the relationship among locally nearby objects, as well as the interactions between an object and its surrounding area. In general, modeling object relations is challenging, requiring reasoning about bounding boxes of different classes, locations, scales \emph{etc}. Deep learning research that explicitly models object relations is quite limited, with representative ones being Spatial Memory Network (SMN) \cite{ChenSpatial2017}, Object Relation Network \cite{Hu2018Relation}, and Structure Inference Network (SIN) \cite{Liu2018Structure}. In SMN, spatial memory essentially assembles object instances back into a pseudo image representation that is easy to be fed into another CNN for object relations reasoning, leading to a new sequential reasoning architecture where image and memory are processed in parallel to obtain detections which further update memory. Inspired by the recent success of attention modules in natural language processing \cite{Vaswani2017Attention}, ORN processes a set of objects simultaneously through the interaction between their appearance feature and geometry. It does not require additional supervision, and it is easy to embed into existing networks, effective in improving object recognition and duplicate removal steps in modern object detection pipelines, giving rise to the first fully end-to-end object detector. SIN \cite{Liu2018Structure} considered two kinds of context: scene contextual information and object relationships within a single image. It formulates object detection as a problem of graph inference, where the objects are treated as nodes in a graph and relationships between objects are modeled as edges. A wider range of methods has approached the context challenge with a simpler idea: enlarging the detection window size to extract some form of local context. Representative approaches include MRCNN \cite{Gidaris2015}, Gated BiDirectional CNN (GBDNet) \cite{ GBDCNN2016,Zeng2017Crafting}, Attention to Context CNN (ACCNN) \cite{Li2017Attentive}, CoupleNet \cite{ZhuICCV2017}, and Sermanet \emph{et al.} \cite{Sermanet2013c}. In MRCNN \cite{Gidaris2015} (Fig.~\ref{Fig:LocalContext} (a)), in addition to the features extracted from the original object proposal at the last CONV layer of the backbone, Gidaris and Komodakis proposed to extract features from a number of different regions of an object proposal (half regions, border regions, central regions, contextual region and semantically segmented regions), in order to obtain a richer and more robust object representation. All of these features are combined by concatenation. Quite a number of methods, all closely related to MRCNN, have been proposed since then. The method in \cite{Zagoruyko2016} used only four contextual regions, organized in a foveal structure, where the classifiers along multiple paths are trained jointly end-to-end. Zeng \emph{et al.} proposed GBDNet \cite{GBDCNN2016,Zeng2017Crafting} (Fig.~\ref{Fig:LocalContext} (b)) to extract features from multiscale contextualized regions surrounding an object proposal to improve detection performance. In contrast to the somewhat naive approach of learning CNN features for each region separately and then concatenating them, GBDNet passes messages among features from different contextual regions. Noting that message passing is not always helpful, but dependent on individual samples, Zeng \emph{et al.} \cite{GBDCNN2016} used gated functions to control message transmission. Li \emph{et al.} \cite{Li2017Attentive} presented ACCNN (Fig.~\ref{Fig:LocalContext} (c)) to utilize both global and local contextual information: the global context was captured using a Multiscale Local Contextualized (MLC) subnetwork, which recurrently generates an attention map for an input image to highlight promising contextual locations; local context adopted a method similar to that of MRCNN \cite{Gidaris2015}. As shown in Fig.~\ref{Fig:LocalContext} (d), CoupleNet \cite{ZhuICCV2017} is conceptually similar to ACCNN \cite{Li2017Attentive}, but built upon RFCN \cite{Dai2016RFCN}, which captures object information with position sensitive RoI pooling, CoupleNet added a branch to encode the global context with RoI pooling. \section{Detection Proposal Methods} \label{Sec:DetectionProposal} An object can be located at any position and scale in an image. During the heyday of handcrafted feature descriptors (SIFT \cite{Lowe2004}, HOG \cite{Dalal2005HOG} and LBP \cite{Ojala02}), the most successful methods for object detection (\emph{e.g.} DPM \cite{Felzenszwalb08CVPR}) used \emph{sliding window} techniques \cite{Viola2001,Dalal2005HOG,Felzenszwalb08CVPR,Harzallah2009Combining,Vedaldi09Multiple}. However, the number of windows is huge, growing with the number of pixels in an image, and the need to search at multiple scales and aspect ratios further increases the search space\footnote{Sliding window based detection requires classifying around $10^4$-$10^5$ windows per image. The number of windows grows significantly to $10^6$-$10^7$ windows per image when considering multiple scales and aspect ratios.}. Therefore, it is computationally too expensive to apply sophisticated classifiers. Around 2011, researchers proposed to relieve the tension between computational tractability and high detection quality by using \emph{detection proposals}\footnote{We use the terminology \emph{detection proposals}, \emph{object proposals} and \emph{region proposals} interchangeably.} \cite{Van2011SS,Uijlings2013b}. Originating in the idea of \emph{objectness} proposed by \cite{Alexe2010Object}, object proposals are a set of candidate regions in an image that are likely to contain objects, and if high object recall can be achieved with a modest number of object proposals (like one hundred), significant speed-ups over the sliding window approach can be gained, allowing the use of more sophisticated classifiers. Detection proposals are usually used as a pre-processing step, limiting the number of regions that need to be evaluated by the detector, and should have the following characteristics: \begin{enumerate} \item High recall, which can be achieved with only a few proposals; \item Accurate localization, such that the proposals match the object bounding boxes as accurately as possible; and \item Low computational cost. \end{enumerate} The success of object detection based on detection proposals \cite{Van2011SS,Uijlings2013b} has attracted broad interest \cite{Carreira2012,Arbelaez2014,Alexe2012,Cheng2014bing,EdgeBox2014,Endres14Category, Philipp14Geodesic,Manen2013Prime}. A comprehensive review of object proposal algorithms is beyond the scope of this paper, because object proposals have applications beyond object detection \cite{Arbelaez2012Semantic,Guillaumin2014ImageNet,Xhu2017Soft}. We refer interested readers to the recent surveys \cite{Hosang2016,Chavali2016} which provide in-depth analysis of many classical object proposal algorithms and their impact on detection performance. Our interest here is to review object proposal methods that are based on DCNNs, output class agnostic proposals, and are related to generic object detection. In 2014, the integration of object proposals \cite{Van2011SS,Uijlings2013b} and DCNN features \cite{Krizhevsky2012} led to the milestone RCNN \cite{Girshick2014RCNN} in generic object detection. Since then, detection proposal has quickly become a standard preprocessing step, based on the fact that all winning entries in the PASCAL VOC \cite{Everingham2010}, ILSVRC \cite{Russakovsky2015} and MS COCO \cite{Lin2014} object detection challenges since 2014 used detection proposals \cite{Girshick2014RCNN,Ouyang2015deepid,Girshick2015FRCNN, Ren2015NIPS,Zeng2017Crafting,MaskRCNN2017}. Among object proposal approaches based on traditional low-level cues (\emph{e.g.,} color, texture, edge and gradients), Selective Search \cite{Uijlings2013b}, MCG \cite{Arbelaez2014} and EdgeBoxes \cite{EdgeBox2014} are among the more popular. As the domain rapidly progressed, traditional object proposal approaches \cite{Uijlings2013b,Hosang2016,EdgeBox2014}, which were adopted as external modules independent of the detectors, became the speed bottleneck of the detection pipeline \cite{Ren2015NIPS}. An emerging class of object proposal algorithms \cite{MultiBox1,Ren2015NIPS,DeepBox2015,Deepproposal2015,DeepMask2015,CRAFT2016} using DCNNs has attracted broad attention. \begin{table*}[!t] \caption {Summary of object proposal methods using DCNN. Blue indicates the number of object proposals. The detection results on COCO are based on mAP@IoU[0.5, 0.95], unless stated otherwise.}\label{Tab:ObjectProposals} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{1pt} \resizebox*{18cm}{!}{ \begin{tabular}{!{\vrule width1.5bp}c|c|p{1.5cm}<{\centering}|c|c|c|c|c|c|c|c|p{9cm}<{\centering}!{\vrule width1.5bp}} \Xhline{1.5pt} \footnotesize $\quad\quad$ & \footnotesize Proposer & \footnotesize Backbone & \footnotesize Detector & \multicolumn{3}{c|}{Recall@IoU (VOC07)} &\multicolumn{3}{c|}{Detection Results (mAP)} & \footnotesize Published & \footnotesize \\ \cline{5-7}\cline{8-10} \footnotesize & \footnotesize Name & \footnotesize Network & \footnotesize Tested & \footnotesize $0.5$ & \footnotesize $0.7$ & \footnotesize $0.9$ & \footnotesize VOC07 & \footnotesize VOC12 & \footnotesize COCO & \footnotesize In & \footnotesize Highlights \\ \cline{2-12} \footnotesize \multirow{8}{*}{\rotatebox{90}{Bounding Box Object Proposal Methods$\quad\quad\quad\quad\quad\quad\quad\quad$ }} & \footnotesize \raisebox{-3ex}[0pt]{MultiBox1\cite{ MultiBox1}} & \footnotesize \raisebox{-3.5ex}[0pt]{AlexNet} & \footnotesize \raisebox{-3ex}[0pt]{RCNN}& \footnotesize \raisebox{-3ex}[0pt]{$-$}& \footnotesize \raisebox{-3ex}[0pt]{$-$} & \footnotesize \raisebox{-3ex}[0pt]{$-$} & \footnotesize \raisebox{-5.5ex}[0pt]{ \shortstack [c] {$29.0$ \\(\textcolor{blue}{10})\\ (12)}} & \footnotesize \raisebox{-3ex}[0pt]{ $-$} & \footnotesize \raisebox{-3ex}[0pt]{$-$} & \footnotesize \raisebox{-3ex}[0pt]{CVPR14} & \footnotesize Learns a class agnostic regressor on a small set of 800 predefined anchor boxes. Do not share features for detection. \\ \cline{2-12} & \footnotesize \raisebox{-3.5ex}[0pt]{ DeepBox \cite{ DeepBox2015}} & \footnotesize \raisebox{-3.5ex}[0pt]{ VGG16 } & \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {Fast \\ RCNN}} & \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.96$ \\ (\textcolor{blue}{1000})}}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.84$ \\ (\textcolor{blue}{1000})}}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.15$ \\ (\textcolor{blue}{1000})}}& \footnotesize \raisebox{-3.5ex}[0pt]{$-$} & \footnotesize \raisebox{-3.5ex}[0pt]{ $-$} & \footnotesize \raisebox{-5.5ex}[0pt]{\shortstack [c] {$37.8$\\(\textcolor{blue}{500})\\([email protected])} } & \footnotesize \raisebox{-3.5ex}[0pt]{ICCV15} & \footnotesize Use a lightweight CNN to learn to rerank proposals generated by EdgeBox. Can run at 0.26s per image. Do not share features for detection. \\ \cline{2-12} & \footnotesize\raisebox{-4.5ex}[0pt]{ RPN\cite{Ren2015NIPS,Ren2016a}} & \footnotesize \raisebox{-4.5ex}[0pt]{ \shortstack [c] { VGG16 } }& \footnotesize \raisebox{-4.5ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}}& \footnotesize \raisebox{-7.5ex}[0pt]{\shortstack [c] {$0.97$ \\ (\textcolor{blue}{300})\\0.98\\(\textcolor{blue}{1000})}}& \footnotesize \raisebox{-7.5ex}[0pt]{\shortstack [c] {$0.79$ \\ (\textcolor{blue}{300})\\0.84\\(\textcolor{blue}{1000})}}& \footnotesize \raisebox{-7.5ex}[0pt]{\shortstack [c] {$0.04$ \\ (\textcolor{blue}{300})\\0.04\\(\textcolor{blue}{1000})}}& \footnotesize \raisebox{-6.5ex}[0pt]{\shortstack [c] { $73.2$ \\ (\textcolor{blue}{300})\\(07+12)}} & \footnotesize \raisebox{-6.5ex}[0pt]{\shortstack [c] { $70.4$ \\ (\textcolor{blue}{300})\\(07++12)} }& \footnotesize \raisebox{-4.5ex}[0pt]{ \shortstack [c] { $21.9$ \\ (\textcolor{blue}{300})} }& \footnotesize\raisebox{-4.5ex}[0pt]{ NIPS15} & \footnotesize The first to generate object proposals by sharing full image convolutional features with detection. Most widely used object proposal method. Significant improvements in detection speed. \\ \cline{2-12} & \footnotesize \raisebox{-3.5ex}[0pt]{DeepProposal\cite{ Deepproposal2015} }& \footnotesize \raisebox{-3.5ex}[0pt]{ VGG16 } & \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {Fast \\ RCNN}}& \footnotesize \raisebox{-7.5ex}[0pt]{\shortstack [c] {$0.74$ \\ (\textcolor{blue}{100})\\0.92\\(\textcolor{blue}{1000})}}& \footnotesize \raisebox{-7.5ex}[0pt]{\shortstack [c] {$0.58$ \\ (\textcolor{blue}{100})\\0.80\\(\textcolor{blue}{1000})}}& \footnotesize\raisebox{-7.5ex}[0pt]{\shortstack [c] {$0.12$ \\ (\textcolor{blue}{100})\\0.16\\(\textcolor{blue}{1000})}}& \footnotesize \raisebox{-6.5ex}[0pt]{ \shortstack [c] {$53.2$ \\ (\textcolor{blue}{100})\\(07)}} & \footnotesize \raisebox{-3.5ex}[0pt]{ $-$}& \footnotesize \raisebox{-3.5ex}[0pt]{$-$}& \footnotesize\raisebox{-3.5ex}[0pt]{ ICCV15} & \footnotesize Generate proposals inside a DCNN in a multiscale manner. Share features with the detection network.\\ \cline{2-12} & \footnotesize \raisebox{-3.5ex}[0pt]{CRAFT \cite{ CRAFT2016}} & \footnotesize \raisebox{-3.5ex}[0pt]{ VGG16} & \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {Faster \\ RCNN}} & \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.98$ \\ (\textcolor{blue}{300}) }}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.90$ \\ (\textcolor{blue}{300}) }}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.13$ \\ (\textcolor{blue}{300}) }}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$75.7$ \\ (07+12) }}& \footnotesize \raisebox{-3.5ex}[0pt]{ \shortstack [c] {71.3\\ (12)}} & \footnotesize \raisebox{-3.5ex}[0pt]{ $-$} & \footnotesize\raisebox{-3.5ex}[0pt]{ CVPR16} & \footnotesize Introduced a classification network (\emph{i.e.} two class Fast RCNN) cascade that comes after the RPN. Not sharing features extracted for detection.\\ \cline{2-12} & \footnotesize\raisebox{-3ex}[0pt]{ AZNet \cite{ Lu2016Adaptive} }& \footnotesize \raisebox{-3ex}[0pt]{VGG16 }& \footnotesize \raisebox{-4ex}[0pt]{\shortstack [c] {Fast \\ RCNN}} & \footnotesize\raisebox{-4ex}[0pt]{ \shortstack [c] {$0.91$ \\ (\textcolor{blue}{300}) }} & \footnotesize \raisebox{-4ex}[0pt]{\shortstack [c] {$0.71$ \\ (\textcolor{blue}{300}) }}& \footnotesize \raisebox{-4ex}[0pt]{\shortstack [c] {$0.11$ \\ (\textcolor{blue}{300}) }}& \footnotesize \raisebox{-4ex}[0pt]{\shortstack [c] {$70.4$\\ (07)} }& \footnotesize \raisebox{-3.5ex}[0pt]{$-$}& \footnotesize \raisebox{-3.5ex}[0pt]{$22.3$} & \footnotesize \raisebox{-3.5ex}[0pt]{CVPR16 }& \footnotesize Use coarse-to-fine search: start from large regions, then recursively search for subregions that may contain objects. Adaptively guide computational resources to focus on likely subregions. \\ \cline{2-12} & \footnotesize \raisebox{-3.5ex}[0pt]{ ZIP \cite{Hongyang2018Zoom}} & \footnotesize \raisebox{-3.5ex}[0pt]{Inception v2} & \footnotesize \raisebox{-3.5ex}[0pt]{ \shortstack [c] {Faster \\ RCNN}} & \footnotesize \raisebox{-5.5ex}[0pt]{\shortstack [c] {$0.85$ \\ (\textcolor{blue}{300})\\COCO }} & \footnotesize \raisebox{-5.5ex}[0pt]{\shortstack [c] {$0.74$ \\ (\textcolor{blue}{300})\\COCO }} & \footnotesize\raisebox{-5.5ex}[0pt]{\shortstack [c] {$0.35$ \\ (\textcolor{blue}{300})\\COCO }} & \footnotesize\raisebox{-3ex}[0pt]{ \shortstack [c] { $79.8$\\ (07+12)}} & \footnotesize \raisebox{-3ex}[0pt]{ $-$ }& \footnotesize \raisebox{-3ex}[0pt]{$-$} & \footnotesize\raisebox{-3ex}[0pt]{ IJCV18 } & \footnotesize Generate proposals using conv-deconv network with multilayers; Proposed a map attention decision (MAD) unit to assign the weights for features from different layers. \\ \cline{2-12} & \footnotesize \raisebox{-3.5ex}[0pt]{DeNet\cite{SmithICCV2017}} & \footnotesize \raisebox{-3.5ex}[0pt]{ResNet101 } & \footnotesize\raisebox{-4.5ex}[0pt]{ \shortstack [c] {Fast \\ RCNN}} & \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.82$ \\ (\textcolor{blue}{300})}}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.74$ \\ (\textcolor{blue}{300})}}& \footnotesize\raisebox{-3.5ex}[0pt]{\shortstack [c] {$0.48$ \\ (\textcolor{blue}{300})}} & \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {$77.1$ \\ (07+12)}}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] { $73.9$ \\ (07++12)} }& \footnotesize \raisebox{-3.5ex}[0pt]{$33.8$} & \footnotesize \raisebox{-3.5ex}[0pt]{ ICCV17 }& \footnotesize A lot faster than Faster RCNN; Introduces a bounding box corner estimation for predicting object proposals efficiently to replace RPN; Does not require predefined anchors.\\ \Xhline{1.5pt} \footnotesize & \footnotesize \shortstack [c] {Proposer\\Name} & \footnotesize \shortstack [c] {Backbone\\Network} & \footnotesize \shortstack [c] { Detector \\ Tested}& \multicolumn{3}{c|}{\shortstack [c] {Box Proposals \\ (AR, COCO)}} & \multicolumn{3}{c|}{\shortstack [c] {Segment Proposals\\ (AR, COCO)}} & \footnotesize \shortstack [c] { \shortstack [c] {Published \\ In}} & \footnotesize Highlights \\ \cline{2-12} \footnotesize \multirow{5}{*}{\rotatebox{90}{Segment Proposal Methods $\quad\quad$ }} & \footnotesize \raisebox{-3.5ex}[0pt]{DeepMask \cite{DeepMask2015}}& \footnotesize \raisebox{-3.5ex}[0pt]{VGG16 }& \footnotesize \raisebox{-3.5ex}[0pt]{ \shortstack [c] {Fast \\ RCNN}} &\multicolumn{3}{c|}{\raisebox{-3.5ex}[0pt]{$0.33$ (\textcolor{blue}{100}), $0.48(\textcolor{blue}{1000})$}} &\multicolumn{3}{c|}{\raisebox{-3.5ex}[0pt]{$0.26$ (\textcolor{blue}{100}), $0.37(\textcolor{blue}{1000})$}} & \footnotesize \raisebox{-3.5ex}[0pt]{ NIPS15 }& \footnotesize First to generate object mask proposals with DCNN; Slow inference time; Need segmentation annotations for training; Not sharing features with detection network; Achieved mAP of $69.9\%$ (\textcolor{blue}{500}) with Fast RCNN. \\ \cline{2-12} & \footnotesize \raisebox{-3ex}[0pt]{InstanceFCN \cite{Dai2016Instance}}& \footnotesize \raisebox{-3ex}[0pt]{ VGG16}& \footnotesize \raisebox{-3ex}[0pt]{$-$}&\multicolumn{3}{c|}{\raisebox{-3ex}[0pt]{$-$}} &\multicolumn{3}{c|}{\raisebox{-3ex}[0pt]{$0.32$ (\textcolor{blue}{100}), $0.39(\textcolor{blue}{1000})$}} & \footnotesize\raisebox{-3ex}[0pt]{ECCV16} & \footnotesize Combines ideas of FCN \cite{FCNCVPR2015} and DeepMask \cite{DeepMask2015}. Introduces instance sensitive score maps. Needs segmentation annotations to train the network. \\ \cline{2-12} & \footnotesize \raisebox{-3.5ex}[0pt]{SharpMask \cite{Pinheiro2016} }& \footnotesize \raisebox{-3.5ex}[0pt]{ MPN \cite{Zagoruyko2016}}& \footnotesize \raisebox{-3.5ex}[0pt]{\shortstack [c] {Fast \\ RCNN}}&\multicolumn{3}{c|}{\raisebox{-3.5ex}[0pt]{$0.39$ (\textcolor{blue}{100}), $0.53(\textcolor{blue}{1000})$}} &\multicolumn{3}{c|}{\raisebox{-3.5ex}[0pt]{$0.30$ (\textcolor{blue}{100}), $0.39(\textcolor{blue}{1000})$}}& \footnotesize\raisebox{-3.5ex}[0pt]{ ECCV16} & \footnotesize Leverages features at multiple convolutional layers by introducing a top-down refinement module. Does not share features with detection network. Needs segmentation annotations for training. \\ \cline{2-12} & \footnotesize \raisebox{-3.5ex}[0pt]{FastMask\cite{ Hu2017FastMask} }& \footnotesize \raisebox{-3.5ex}[0pt]{ResNet39} & \footnotesize \raisebox{-3.5ex}[0pt]{$-$} &\multicolumn{3}{c|}{\raisebox{-3.5ex}[0pt]{$0.43$ (\textcolor{blue}{100}), $0.57(\textcolor{blue}{1000})$}} & \multicolumn{3}{c|}{\raisebox{-3.5ex}[0pt]{$0.32$ (\textcolor{blue}{100}), $0.41(\textcolor{blue}{1000})$}} & \footnotesize \raisebox{-3.5ex}[0pt]{ CVPR17} & \footnotesize Generates instance segment proposals efficiently in one-shot manner similar to SSD \cite{Liu2016SSD}. Uses multiscale convolutional features. Uses segmentation annotations for training. \\ \Xhline{1.5pt} \end{tabular} } \end{table*} Recent DCNN based object proposal methods generally fall into two categories: {\em bounding box} based and {\em object segment} based, with representative methods summarized in Table \ref{Tab:ObjectProposals}. \begin {figure}[!t] \centering \includegraphics[width=0.4\textwidth]{Anchor.pdf} \caption{Illustration of the Region Proposal Network (RPN) introduced in \cite{Ren2015NIPS}.} \label{fig:anchor} \end {figure} \textbf{Bounding Box Proposal Methods} are best exemplified by the RPC method \cite{Ren2015NIPS} of Ren \emph{et al.}, illustrated in Fig.~\ref{fig:anchor}. RPN predicts object proposals by sliding a small network over the feature map of the last shared CONV layer. At each sliding window location, $k$ proposals are predicted by using $k$ anchor boxes, where each anchor box\footnote{The concept of ``anchor'' first appeared in \cite{Ren2015NIPS}.} is centered at some location in the image, and is associated with a particular scale and aspect ratio. Ren \emph{et al.} \cite{Ren2015NIPS} proposed integrating RPN and Fast RCNN into a single network by sharing their convolutional layers, leading to Faster RCNN, the first end-to-end detection pipeline. RPN has been broadly selected as the proposal method by many state-of-the-art object detectors, as can be observed from Tables \ref{Tab:EnhanceFeatures} and \ref{Tab:ContextMethods}. Instead of fixing {\em a priori} a set of anchors as MultiBox \cite{MultiBox1,MultiBox2} and RPN \cite{Ren2015NIPS}, Lu \emph{et al.} \cite{Lu2016Adaptive} proposed generating anchor locations by using a recursive search strategy which can adaptively guide computational resources to focus on sub-regions likely to contain objects. Starting with the whole image, all regions visited during the search process serve as anchors. For any anchor region encountered during the search procedure, a scalar zoom indicator is used to decide whether to further partition the region, and a set of bounding boxes with objectness scores are computed by an Adjacency and Zoom Network (AZNet), which extends RPN by adding a branch to compute the scalar zoom indicator in parallel with the existing branch. Further work attempts to generate object proposals by exploiting multilayer convolutional features. Concurrent with RPN \cite{Ren2015NIPS}, Ghodrati \emph{et al.} \cite{Deepproposal2015} proposed DeepProposal, which generates object proposals by using a cascade of multiple convolutional features, building an inverse cascade to select the most promising object locations and to refine their boxes in a coarse-to-fine manner. An improved variant of RPN, HyperNet \cite{HyperNet2016} designs Hyper Features which aggregate multilayer convolutional features and shares them both in generating proposals and detecting objects via an end-to-end joint training strategy. Yang \emph{et al.} proposed CRAFT \cite{ CRAFT2016} which also used a cascade strategy, first training an RPN network to generate object proposals and then using them to train another binary Fast RCNN network to further distinguish objects from background. Li \emph{et al.} \cite{Hongyang2018Zoom} proposed ZIP to improve RPN by predicting object proposals with multiple convolutional feature maps at different network depths to integrate both low level details and high level semantics. The backbone used in ZIP is a ``zoom out and in'' network inspired by the conv and deconv structure \cite{FCNCVPR2015}. Finally, recent work which deserves mention includes Deepbox \cite{DeepBox2015}, which proposed a lightweight CNN to learn to rerank proposals generated by EdgeBox, and DeNet \cite{SmithICCV2017} which introduces bounding box corner estimation to predict object proposals efficiently to replace RPN in a Faster RCNN style detector. \textbf{Object Segment Proposal Methods} \cite{DeepMask2015,Pinheiro2016} aim to generate segment proposals that are likely to correspond to objects. Segment proposals are more informative than bounding box proposals, and take a step further towards object instance segmentation \cite{Hariharan2014,Dai2016Aware,Li2017Fully}. In addition, using instance segmentation supervision can improve the performance of bounding box object detection. The pioneering work of DeepMask, proposed by Pinheiro \emph{et al.} \cite{DeepMask2015}, segments proposals learnt directly from raw image data with a deep network. Similarly to RPN, after a number of shared convolutional layers DeepMask splits the network into two branches in order to predict a class agnostic mask and an associated objectness score. Also similar to the efficient sliding window strategy in OverFeat \cite{OverFeat2014}, the trained DeepMask network is applied in a sliding window manner to an image (and its rescaled versions) during inference. More recently, Pinheiro \emph{et al.} \cite{Pinheiro2016} proposed SharpMask by augmenting the DeepMask architecture with a refinement module, similar to the architectures shown in Fig.~\ref{fig:MultiLayerCombine} (b1) and (b2), augmenting the feed-forward network with a top-down refinement process. SharpMask can efficiently integrate spatially rich information from early features with strong semantic information encoded in later layers to generate high fidelity object masks. Motivated by Fully Convolutional Networks (FCN) for semantic segmentation \cite{FCNCVPR2015} and DeepMask \cite{DeepMask2015}, Dai \emph{et al.} proposed InstanceFCN \cite{Dai2016Instance} to generate instance segment proposals. Similar to DeepMask, the InstanceFCN network is split into two fully convolutional branches, one to generate instance sensitive score maps, the other to predict the objectness score. Hu \emph{et al.} proposed FastMask \cite{Hu2017FastMask} to efficiently generate instance segment proposals in a one-shot manner, similar to SSD \cite{Liu2016SSD}, in order to make use of multiscale convolutional features. Sliding windows extracted densely from multiscale convolutional feature maps were input to a scale-tolerant attentional head module in order to predict segmentation masks and objectness scores. FastMask is claimed to run at 13 FPS on $800\times600$ images. \begin{table*}[!t] \caption {Representative methods for training strategies and class imbalance handling. Results on COCO are reported with Test Dev. The detection results on COCO are based on mAP@IoU[0.5, 0.95].}\label{Tab:ClassImbalance} \centering \renewcommand{\arraystretch}{1.2} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{1pt} \resizebox*{18.5cm}{!}{ \begin{tabular}{!{\vrule width1.5bp}c|c|c|c|c|c|c|c|p{8cm}<{\centering}!{\vrule width1.5bp}} \Xhline{1.5pt} \footnotesize \shortstack [c] {Detector \\ Name} & \footnotesize \shortstack [c] {Region \\ Proposal} & \footnotesize \shortstack [c] {Backbone \\ DCNN} & \footnotesize \shortstack [c] {Pipelined \\ Used} & \footnotesize \shortstack [c] {VOC07 \\ Results} & \footnotesize \shortstack [c] {VOC12 \\ Results} & \footnotesize \shortstack [c] {COCO \\ Results} & \footnotesize \shortstack [c] {Published \\ In} & \footnotesize Highlights \\ \Xhline{1.5pt} \raisebox{-4ex}[0pt]{MegDet \cite{Peng2018MegDet}} & \footnotesize \raisebox{-4ex}[0pt]{ RPN} & \footnotesize \raisebox{-5ex}[0pt]{ \shortstack [c] { ResNet50\\+FPN }}& \footnotesize \raisebox{-5ex}[0pt]{ \shortstack [c] { Faster\\RCNN}} & \footnotesize \raisebox{-4ex}[0pt]{$-$}& \footnotesize \raisebox{-4ex}[0pt]{$-$} & \footnotesize \raisebox{-4ex}[0pt]{$52.5$} & \footnotesize \raisebox{-4ex}[0pt]{CVPR18} & \footnotesize Allow training with much larger minibatch size than before by introducing cross GPU batch normalization; Can finish the COCO training in 4 hours on 128 GPUs and achieved improved accuracy; Won COCO2017 detection challenge. \\ \hline \raisebox{-2.5ex}[0pt]{SNIP \cite{Singh2018sniper} }& \footnotesize \raisebox{-2.5ex}[0pt]{RPN }& \footnotesize \raisebox{-3.5ex}[0pt]{ \shortstack [c] {DPN \cite{Chen2017Dual}\\+DCN \cite{Dai17Deformable}}} & \footnotesize \raisebox{-2.5ex}[0pt]{RFCN} & \footnotesize \raisebox{-2.5ex}[0pt]{$-$ }& \footnotesize \raisebox{-2.5ex}[0pt]{$-$}& \footnotesize \raisebox{-2.5ex}[0pt]{$48.3$} & \footnotesize \raisebox{-2.5ex}[0pt]{CVPR18} & \footnotesize A new multiscale training scheme. Empirically examined the effect of up-sampling for small object detection. During training, only select objects that fit the scale of features as positive samples. \\ \hline \raisebox{-1ex}[0pt]{SNIPER \cite{Singh2018sniper}} & \footnotesize \raisebox{-1ex}[0pt]{RPN} & \footnotesize \raisebox{-2.5ex}[0pt]{\shortstack [c] {ResNet101\\+DCN}} & \footnotesize \raisebox{-2.5ex}[0pt]{\shortstack [c] {Faster \\ RCNN}} & \footnotesize\raisebox{-1ex}[0pt]{ $-$} & \footnotesize \raisebox{-1ex}[0pt]{$-$} & \footnotesize \raisebox{-1ex}[0pt]{$47.6$} & \footnotesize\raisebox{-1ex}[0pt]{ 2018 }& An efficient multiscale training strategy. Process context regions around ground-truth instances at the appropriate scale. \\ \hline \raisebox{-1.5ex}[0pt]{OHEM \cite{Shrivastava2016OHEM} }& \footnotesize \raisebox{-1.5ex}[0pt]{SS} & \footnotesize \raisebox{-1.5ex}[0pt]{VGG16} & \footnotesize \raisebox{-2.5ex}[0pt]{ \shortstack [c] { Fast \\ RCNN} }& \footnotesize\raisebox{-2.5ex}[0pt]{ \shortstack [c] {$78.9$\\ (07+12)}}& \footnotesize \raisebox{-2.5ex}[0pt]{\shortstack [c] { $76.3$\\ (07++12)}}& \footnotesize \raisebox{-1.5ex}[0pt]{$22.4$}& \footnotesize \raisebox{-1.5ex}[0pt]{CVPR16}& \footnotesize A simple and effective Online Hard Example Mining algorithm to improve training of region based detectors. \\ \hline \raisebox{-1.5ex}[0pt]{ FactorNet \cite{Ouyang2016Factors} }& \footnotesize \raisebox{-1.5ex}[0pt]{SS} & \footnotesize \raisebox{-1.5ex}[0pt]{GooglNet} & \footnotesize \raisebox{-2.5ex}[0pt]{ \shortstack [c] {RCNN} }& \footnotesize\raisebox{-2.5ex}[0pt]{$-$ } & \footnotesize \raisebox{-2.5ex}[0pt]{$-$ } & \raisebox{-2.5ex}[0pt]{$-$ } & \footnotesize \raisebox{-1.5ex}[0pt]{CVPR16}& \footnotesize Identify the imbalance in the number of samples for different object categories; propose a divide-and-conquer feature learning scheme. \\ \hline \raisebox{-5.5ex}[0pt]{Chained Cascade \cite{CascadeRCNN2018} }& \footnotesize \raisebox{-7ex}[0pt]{\shortstack [c] {SS \\ CRAFT} } & \footnotesize \raisebox{-7ex}[0pt]{\shortstack [c] {VGG \\ Inceptionv2}} & \footnotesize \raisebox{-7.5ex}[0pt]{ \shortstack [c] {Fast RCNN, \\ Faster RCNN} }& \footnotesize\raisebox{-8.5ex}[0pt]{ \shortstack [c] {$80.4$\\ (07+12) \\ (SS+VGG)}}& \footnotesize \raisebox{-5.5ex}[0pt]{$-$}& \footnotesize \raisebox{-5.5ex}[0pt]{$-$}& \footnotesize \raisebox{-5.5ex}[0pt]{ICCV17}& \footnotesize Jointly learn DCNN and multiple stages of cascaded classifiers. Boost detection accuracy on PASCAL VOC 2007 and ImageNet for both fast RCNN and Faster RCNN using different region proposal methods. \\ \hline \raisebox{-5.5ex}[0pt]{Cascade RCNN \cite{CascadeRCNN2018} }& \footnotesize \raisebox{-5.5ex}[0pt]{RPN} & \footnotesize \raisebox{-8.5ex}[0pt]{\shortstack [c] {VGG\\ResNet101\\+FPN}} & \footnotesize \raisebox{-6.5ex}[0pt]{ \shortstack [c] {Faster RCNN}}& \footnotesize\raisebox{-6.5ex}[0pt]{ \shortstack [c] {$-$}}& \footnotesize \raisebox{-6.5ex}[0pt]{\shortstack [c] {$-$}}& \footnotesize \raisebox{-5.5ex}[0pt]{$42.8$}& \footnotesize \raisebox{-5.5ex}[0pt]{CVPR18}& \footnotesize Jointly learn DCNN and multiple stages of cascaded classifiers, which are learned using different localization accuracy for selecting positive samples. Stack bounding box regression at multiple stages. \\ \hline \raisebox{-3ex}[0pt]{RetinaNet \cite{LinICCV2017}} & \footnotesize \raisebox{-4ex}[0pt]{$-$} & \footnotesize \raisebox{-4ex}[0pt]{\shortstack [c] {ResNet101\\+FPN }}& \footnotesize \raisebox{-3ex}[0pt]{RetinaNet }& \footnotesize \raisebox{-3ex}[0pt]{$-$}& \footnotesize \raisebox{-3ex}[0pt]{$-$} & \footnotesize \raisebox{-3ex}[0pt]{$39.1$} & \footnotesize \raisebox{-3ex}[0pt]{ICCV17} & \footnotesize Propose a novel Focal Loss which focuses training on hard examples. Handles well the problem of imbalance of positive and negative samples when training a one-stage detector. \\ \Xhline{1.5pt} \end{tabular} } \end{table*} \section{Other Issues} \label{sec:otherissue} \textbf{Data Augmentation.} Performing data augmentation for learning DCNNs \cite{Chatfield2014,Girshick2015FRCNN,Girshick2014RCNN} is generally recognized to be important for visual recognition. Trivial data augmentation refers to perturbing an image by transformations that leave the underlying category unchanged, such as cropping, flipping, rotating, scaling, translating, color perturbations, and adding noise. By artificially enlarging the number of samples, data augmentation helps in reducing overfitting and improving generalization. It can be used at training time, at test time, or both. Nevertheless, it has the obvious limitation that the time required for training increases significantly. Data augmentation may synthesize completely new training images \cite{Peng2015Learning,Wang2017}, however it is hard to guarantee that the synthetic images generalize well to real ones. Some researchers \cite{Dwibedi2017Cut,Gupta2016Synthetic} proposed augmenting datasets by pasting real segmented objects into natural images; indeed, Dvornik \emph{et al.} \cite{Dvornik2018Modeling} showed that appropriately modeling the visual context surrounding objects is crucial to place them in the right environment, and proposed a context model to automatically find appropriate locations on images to place new objects for data augmentation. \textbf{Novel Training Strategies.} Detecting objects under a wide range of scale variations, especially the detection of very small objects, stands out as a key challenge. It has been shown \cite{Huang2016Speed,Liu2016SSD} that image resolution has a considerable impact on detection accuracy, therefore scaling is particularly commonly used in data augmentation, since higher resolutions increase the possibility of detecting small objects \cite{Huang2016Speed}. Recently, Singh \emph{et al.} proposed advanced and efficient data argumentation methods SNIP \cite{Singh2018SNIP} and SNIPER \cite{Singh2018sniper} to illustrate the scale invariance problem, as summarized in Table \ref{Tab:ClassImbalance}. Motivated by the intuitive understanding that small and large objects are difficult to detect at smaller and larger scales, respectively, SNIP introduces a novel training scheme that can reduce scale variations during training, but without reducing training samples; SNIPER allows for efficient multiscale training, only processing context regions around ground truth objects at the appropriate scale, instead of processing a whole image pyramid. Peng \emph{et al.} \cite{Peng2018MegDet} studied a key factor in training, the minibatch size, and proposed MegDet, a Large MiniBatch Object Detector, to enable the training with a much larger minibatch size than before (from 16 to 256). To avoid the failure of convergence and significantly speed up the training process, Peng \emph{et al.} \cite{Peng2018MegDet} proposed a learning rate policy and Cross GPU Batch Normalization, and effectively utilized 128 GPUs, allowing MegDet to finish COCO training in 4 hours on 128 GPUs, and winning the COCO 2017 Detection Challenge. \begin {figure}[!t] \centering \includegraphics[width=0.4\textwidth]{IOU.pdf} \caption{Localization error could stem from insufficient overlap or duplicate detections. Localization error is a frequent cause of false positives.} \label{fig:iou} \end {figure} \textbf{Reducing Localization Error.} In object detection, the Intersection Over Union\footnote{Please refer to Section \ref{sec:EvaluationCriteria} for more details on the definition of IOU.} (IOU) between a detected bounding box and its ground truth box is the most popular evaluation metric, and an IOU threshold (\emph{e.g.} typical value of $0.5$) is required to define positives and negatives. From Fig.~\ref{Fig:RegionVsUnified}, in most state of the art detectors \cite{Girshick2015FRCNN,Liu2016SSD,MaskRCNN2017,Ren2015NIPS,YoLo2016} object detection is formulated as a multitask learning problem, \emph{i.e.,} jointly optimizing a softmax classifier which assigns object proposals with class labels and bounding box regressors, localizing objects by maximizing IOU or other metrics between detection results and ground truth. Bounding boxes are only a crude approximation for articulated objects, consequently background pixels are almost invariably included in a bounding box, which affects the accuracy of classification and localization. The study in \cite{Hoiem2012} shows that object localization error is one of the most influential forms of error, in addition to confusion between similar objects. Localization error could stem from insufficient overlap (smaller than the required IOU threshold, such as the green box in Fig.~\ref{fig:iou}) or duplicate detections (\emph{i.e.,} multiple overlapping detections for an object instance). Usually, some post-processing step like NonMaximum Suppression (NMS) \cite{Bodla2017Soft,Hosang2017Learning} is used for eliminating duplicate detections. However, due to misalignments the bounding box with better localization could be suppressed during NMS, leading to poorer localization quality (such as the purple box shown in Fig.~\ref{fig:iou}). Therefore, there are quite a few methods aiming at improving detection performance by reducing localization error. MRCNN \cite{Gidaris2015} introduces iterative bounding box regression, where an RCNN is applied several times. CRAFT \cite{CRAFT2016} and AttractioNet \cite{Gidaris2016Attend} use a multi-stage detection sub-network to generate accurate proposals, to forward to Fast RCNN. Cai and Vasconcelos proposed Cascade RCNN \cite{CascadeRCNN2018}, a multistage extension of RCNN, in which a sequence of detectors is trained sequentially with increasing IOU thresholds, based on the observation that the output of a detector trained with a certain IOU is a good distribution to train the detector of the next higher IOU threshold, in order to be sequentially more selective against close false positives. This approach can be built with any RCNN-based detector, and is demonstrated to achieve consistent gains (about 2 to 4 points) independent of the baseline detector strength, at a marginal increase in computation. There is also recent work \cite{Jiang2018Acquisition,Rezatofighi2019Generalized,Huang2019Mask} formulating IOU directly as the optimization objective, and in proposing improved NMS results \cite{Bodla2017Soft,He2019Bounding,Hosang2017Learning,Smith2018Improving}, such as Soft NMS \cite{Bodla2017Soft} and learning NMS \cite{Hosang2017Learning}. \textbf{Class Imbalance Handling.} Unlike image classification, object detection has another unique problem: the serious imbalance between the number of labeled object instances and the number of background examples (image regions not belonging to any object class of interest). Most background examples are easy negatives, however this imbalance can make the training very inefficient, and the large number of easy negatives tends to overwhelm the training. In the past, this issue has typically been addressed via techniques such as bootstrapping \cite{Sung1996Learning}. More recently, this problem has also seen some attention \cite{Li2019Gradient,LinICCV2017,Shrivastava2016OHEM}. Because the region proposal stage rapidly filters out most background regions and proposes a small number of object candidates, this class imbalance issue is mitigated to some extent in two-stage detectors \cite{Girshick2014RCNN,Girshick2015FRCNN,Ren2015NIPS,MaskRCNN2017}, although example mining approaches, such as Online Hard Example Mining (OHEM) \cite{Shrivastava2016OHEM}, may be used to maintain a reasonable balance between foreground and background. In the case of one-stage object detectors \cite{YoLo2016,Liu2016SSD}, this imbalance is extremely serious (\emph{e.g.} 100,000 background examples to every object). Lin \emph{et al.} \cite{LinICCV2017} proposed Focal Loss to address this by rectifying the Cross Entropy loss, such that it down-weights the loss assigned to correctly classified examples. Li \emph{et al.} \cite{Li2019Gradient} studied this issue from the perspective of gradient norm distribution, and proposed a Gradient Harmonizing Mechanism (GHM) to handle it. \section{Discussion and Conclusion} \label{Sec:Conclusions} Generic object detection is an important and challenging problem in computer vision and has received considerable attention. Thanks to remarkable developments in deep learning techniques, the field of object detection has dramatically evolved. As a comprehensive survey on deep learning for generic object detection, this paper has highlighted the recent achievements, provided a structural taxonomy for methods according to their roles in detection, summarized existing popular datasets and evaluation criteria, and discussed performance for the most representative methods. We conclude this review with a discussion of the state of the art in Section~\ref{Sec:Performance}, an overall discussion of key issues in Section~\ref{Sec:Discussion}, and finally suggested future research directions in Section~\ref{Sec:Directions}. \subsection{State of the Art Performance} \label{Sec:Performance} A large variety of detectors has appeared in the last few years, and the introduction of standard benchmarks, such as PASCAL VOC \cite{Everingham2010,Everingham2015}, ImageNet \cite{Russakovsky2015} and COCO \cite{Lin2014}, has made it easier to compare detectors. As can be seen from our earlier discussion in Sections~\ref{Sec:Frameworks} through~\ref{sec:otherissue}, it may be misleading to compare detectors in terms of their originally reported performance (\emph{e.g.} accuracy, speed), as they can differ in fundamental / contextual respects, including the following choices: \begin{itemize} \renewcommand{\labelitemi}{$\bullet$} \item Meta detection frameworks, such as RCNN \cite{Girshick2014RCNN}, Fast RCNN \cite{Girshick2015FRCNN}, Faster RCNN \cite{Ren2015NIPS}, RFCN \cite{Dai2016RFCN}, Mask RCNN \cite{MaskRCNN2017}, YOLO \cite{YoLo2016} and SSD \cite{Liu2016SSD}; \item Backbone networks such as VGG \cite{Simonyan2014VGG}, Inception \cite{GoogLeNet2015,Ioffe2015,Szegedy2016a}, ResNet \cite{He2016ResNet}, ResNeXt \cite{Xie2016Aggregated}, and Xception \cite{Chollet2017Xception} \emph{etc.} listed in Table \ref{Tab:dcnnarchitectures}; \item Innovations such as multilayer feature combination \cite{FPN2016,Shrivastava2017,DSSD2016}, deformable convolutional networks \cite{Dai17Deformable}, deformable RoI pooling \cite{Ouyang2015deepid,Dai17Deformable}, heavier heads \cite{Ren2016NOC,Peng2018MegDet}, and lighter heads \cite{Li2018Light}; \item Pretraining with datasets such as ImageNet \cite{Russakovsky2015}, COCO \cite{Lin2014}, Places \cite{Zhou2017Places}, JFT \cite{Hinton2015Distilling} and Open Images \cite{OpenImages2017}; \item Different detection proposal methods and different numbers of object proposals; \item Train/test data augmentation, novel multiscale training strategies \cite{Singh2018SNIP,Singh2018sniper} \emph{etc}, and model ensembling. \end{itemize} Although it may be impractical to compare every recently proposed detector, it is nevertheless valuable to integrate representative and publicly available detectors into a common platform and to compare them in a unified manner. There has been very limited work in this regard, except for Huang's study \cite{Huang2016Speed} of the three main families of detectors (Faster RCNN \cite{Ren2015NIPS}, RFCN \cite{Dai2016RFCN} and SSD \cite{Liu2016SSD}) by varying the backbone network, image resolution, and the number of box proposals. \begin {figure}[!t] \centering \includegraphics[width=0.49\textwidth]{cocoresults.pdf} \caption{Evolution of object detection performance on COCO (Test-Dev results). Results are quoted from \cite{Girshick2015FRCNN,MaskRCNN2017,Ren2016a}. The backbone network, the design of detection framework and the availability of good and large scale datasets are the three most important factors in detection accuracy.} \label{fig:cocoresults} \end {figure} As can be seen from Tables~\ref{Tab:EnhanceFeatures}, \ref{Tab:ContextMethods}, \ref{Tab:ObjectProposals}, \ref{Tab:ClassImbalance}, \ref{Tab:Detectors}, we have summarized the best reported performance of many methods on three widely used standard benchmarks. The results of these methods were reported on the same test benchmark, despite their differing in one or more of the aspects listed above. Figs.~\ref{fig:GODResultsStatistics} and~\ref{fig:cocoresults} present a very brief overview of the state of the art, summarizing the best detection results of the PASCAL VOC, ILSVRC and MSCOCO challenges; more results can be found at detection challenge websites \cite{ILSVRCResults,COCOResults,VOCResults}. The competition winner of the open image challenge object detection task achieved $61.71\%$ mAP in the public leader board and $58.66\%$ mAP on the private leader board, obtained by combining the detection results of several two-stage detectors including Fast RCNN \cite{Girshick2015FRCNN}, Faster RCNN \cite{Ren2015NIPS}, FPN \cite{FPN2016}, Deformable RCNN \cite{Dai17Deformable}, and Cascade RCNN \cite{CascadeRCNN2018}. In summary, the backbone network, the detection framework, and the availability of large scale datasets are the three most important factors in detection accuracy. Ensembles of multiple models, the incorporation of context features, and data augmentation all help to achieve better accuracy. In less than five years, since AlexNet \cite{Krizhevsky2012} was proposed, the Top5 error on ImageNet classification \cite{Russakovsky2015} with 1000 classes has dropped from 16\% to 2\%, as shown in Fig.~\ref{fig:ILSVRCclassificationResults}. However, the mAP of the best performing detector \cite{Peng2018MegDet} on COCO \cite{Lin2014}, trained to detect only 80 classes, is only at $73\%$, even at 0.5 IoU, illustrating how object detection is much harder than image classification. The accuracy and robustness achieved by the state-of-the-art detectors far from satisfies the requirements of real world applications, so there remains significant room for future improvement. \subsection{Summary and Discussion} \label{Sec:Discussion} With hundreds of references and many dozens of methods discussed throughout this paper, we would now like to focus on the key factors which have emerged in generic object detection based on deep learning. \textbf{(1) Detection Frameworks: Two Stage vs. One Stage} \\ In Section~\ref{Sec:Frameworks} we identified two major categories of detection frameworks: region based (two stage) and unified (one stage): \begin{itemize} \renewcommand{\labelitemi}{$\bullet$} \item When large computational cost is allowed, two-stage detectors generally produce higher detection accuracies than one-stage, evidenced by the fact that most winning approaches used in famous detection challenges like are predominantly based on two-stage frameworks, because their structure is more flexible and better suited for region based classification. The most widely used frameworks are Faster RCNN \cite{Ren2015NIPS}, RFCN \cite{Dai2016RFCN} and Mask RCNN \cite{MaskRCNN2017}. \item It has been shown in \cite{Huang2016Speed} that the detection accuracy of one-stage SSD \cite{Liu2016SSD} is less sensitive to the quality of the backbone network than representative two-stage frameworks. \item One-stage detectors like YOLO \cite{YoLo2016} and SSD \cite{Liu2016SSD} are generally faster than two-stage ones, because of avoiding preprocessing algorithms, using lightweight backbone networks, performing prediction with fewer candidate regions, and making the classification subnetwork fully convolutional. However, two-stage detectors can run in real time with the introduction of similar techniques. In any event, whether one stage or two, the most time consuming step is the feature extractor (backbone network) \cite{Law2018CornerNet,Ren2015NIPS}. \item It has been shown \cite{Huang2016Speed,YoLo2016,Liu2016SSD} that one-stage frameworks like YOLO and SSD typically have much poorer performance when detecting small objects than two-stage architectures like Faster RCNN and RFCN, but are competitive in detecting large objects. \end{itemize} There have been many attempts to build better (faster, more accurate, or more robust) detectors by attacking each stage of the detection framework. No matter whether one, two or multiple stages, the design of the detection framework has converged towards a number of crucial design choices: \begin{itemize} \renewcommand{\labelitemi}{$\bullet$} \item A fully convolutional pipeline \item Exploring complementary information from other correlated tasks, \emph{e.g.}, Mask RCNN \cite{MaskRCNN2017} \item Sliding windows \cite{Ren2015NIPS} \item Fusing information from different layers of the backbone. \end{itemize} The evidence from recent success of cascade for object detection \cite{CascadeRCNN2018,Cheng2018Decoupled,Cheng18Revisiting} and instance segmentation on COCO~\cite{Chen2019Hybrid} and other challenges has shown that multistage object detection could be a future framework for a speed-accuracy trade-off. A teaser investigation is being done in the 2019 WIDER Challenge~\cite{Loy2019Wider}. \textbf{(2) Backbone Networks} \\ As discussed in Section~\ref{Sec:PopularNetworks}, backbone networks are one of the main driving forces behind the rapid improvement of detection performance, because of the key role played by discriminative object feature representation. Generally, deeper backbones such as ResNet \cite{He2016ResNet}, ResNeXt \cite{Xie2016Aggregated}, InceptionResNet \cite{InceptionV4} perform better; however, they are computationally more expensive and require much more data and massive computing for training. Some backbones \cite{Howard2017MobileNets,SqueezeNet2016,Zhang18ShuffleNet} were proposed for focusing on speed instead, such as MobileNet \cite{Howard2017MobileNets} which has been shown to achieve VGGNet16 accuracy on ImageNet with only $\frac{1}{30}$ the computational cost and model size. Backbone training from scratch may become possible as more training data and better training strategies are available \cite{Wu2018Group,Luo2019Switchable,Luo2018Towards}. \textbf{(3) Improving the Robustness of Object Representation}\\ The variation of real world images is a key challenge in object recognition. The variations include lighting, pose, deformations, background clutter, occlusions, blur, resolution, noise, and camera distortions. \textbf{(3.1) Object Scale and Small Object Size} \\ Large variations of object scale, particularly those of small objects, pose a great challenge. Here a summary and discussion on the main strategies identified in Section~\ref{Sec:EnhanceFeatures}: \begin{itemize} \renewcommand{\labelitemi}{$\bullet$} \item Using image pyramids: They are simple and effective, helping to enlarge small objects and to shrink large ones. They are computationally expensive, but are nevertheless commonly used during inference for better accuracy. \item Using features from convolutional layers of different resolutions: In early work like SSD \cite{Liu2016SSD}, predictions are performed independently, and no information from other layers is combined or merged. Now it is quite standard to combine features from different layers, e.g. in FPN \cite{FPN2016}. \item Using dilated convolutions \cite{Li2018DetNet,Li2019Scale}: A simple and effective method to incorporate broader context and maintain high resolution feature maps. \item Using anchor boxes of different scales and aspect ratios: Drawbacks of having many parameters, and scales and aspect ratios of anchor boxes are usually heuristically determined. \item Up-scaling: Particularly for the detection of small objects, high-resolution networks \cite{Sun2019Deep,Sun2019High} can be developed. It remains unclear whether super-resolution techniques improve detection accuracy or not. \end{itemize} Despite recent advances, the detection accuracy for small objects is still much lower than that of larger ones. Therefore, the detection of small objects remains one of the key challenges in object detection. Perhaps localization requirements need to be generalized as a function of scale, since certain applications, e.g. autonomous driving, only require the identification of the existence of small objects within a larger region, and exact localization is not necessary. \textbf{(3.2) Deformation, Occlusion, and other factors} \\ As discussed in Section~\ref{Sec:MainChallenges}, there are approaches to handling geometric transformation, occlusions, and deformation mainly based on two paradigms. The first is a spatial transformer network, which uses regression to obtain a deformation field and then warp features according to the deformation field \cite{Dai17Deformable}. The second is based on a deformable part-based model \cite{Felzenszwalb2010b}, which finds the maximum response to a part filter with spatial constraints taken into consideration \cite{Ouyang2015deepid, Girshick2015DPMCNN,Wan2015end}. Rotation invariance may be attractive in certain applications, but there are limited generic object detection work focusing on rotation invariance, because popular benchmark detection datasets (PASCAL VOC, ImageNet, COCO) do not have large variations in rotation. Occlusion handling is intensively studied in face detection and pedestrian detection, but very little work has been devoted to occlusion handling for generic object detection. In general, despite recent advances, deep networks are still limited by the lack of robustness to a number of variations, which significantly constrains their real-world applications. \textbf{(4) Context Reasoning} \\ As introduced in Section \ref{sec:ContextInfo}, objects in the wild typically coexist with other objects and environments. It has been recognized that contextual information (object relations, global scene statistics) helps object detection and recognition \cite{Oliva2007Role}, especially for small objects, occluded objects, and with poor image quality. There was extensive work preceding deep learning \cite{Malisiewicz09Beyond,Murphy03Using,Rabinovich2007Objects,Divvala2009,Galleguillos2010}, and also quite a few works in the era of deep learning \cite{Gidaris2015,GBDCNN2016,Zeng2017Crafting,ChenSpatial2017,Hu2018Relation}. How to efficiently and effectively incorporate contextual information remains to be explored, possibly guided by how human vision uses context, based on scene graphs \cite{Li2017Scene}, or via the full segmentation of objects and scenes using panoptic segmentation \cite{Kirillov2018Panoptic}. \textbf{(5) Detection Proposals} \\ Detection proposals significantly reduce search spaces. As recommended in \cite{Hosang2016}, future detection proposals will surely have to improve in repeatability, recall, localization accuracy, and speed. Since the success of RPN \cite{Ren2015NIPS}, which integrated proposal generation and detection into a common framework, CNN based detection proposal generation methods have dominated region proposal. It is recommended that new detection proposals should be assessed for object detection, instead of evaluating detection proposals alone. \textbf{(6) Other Factors} \\ As discussed in Section~\ref{sec:otherissue}, there are many other factors affecting object detection quality: data augmentation, novel training strategies, combinations of backbone models, multiple detection frameworks, incorporating information from other related tasks, methods for reducing localization error, handling the huge imbalance between positive and negative samples, mining of hard negative samples, and improving loss functions. \subsection{Research Directions} \label{Sec:Directions} Despite the recent tremendous progress in the field of object detection, the technology remains significantly more primitive than human vision and cannot yet satisfactorily address real-world challenges like those of Section~\ref{Sec:MainChallenges}. We see a number of long-standing challenges: \begin{itemize} \renewcommand{\labelitemi}{$\bullet$} \item Working in an open world: being robust to any number of environmental changes, being able to evolve or adapt. \item Object detection under constrained conditions: learning from weakly labeled data or few bounding box annotations, wearable devices, unseen object categories etc. \item Object detection in other modalities: video, RGBD images, 3D point clouds, lidar, remotely sensed imagery \emph{etc}. \end{itemize} Based on these challenges, we see the following directions of future research: \textbf{(1) Open World Learning:} The ultimate goal is to develop object detection capable of accurately and efficiently recognizing and localizing instances in thousands or more object categories in open-world scenes, at a level competitive with the human visual system. Object detection algorithms are unable, in general, to recognize object categories outside of their training dataset, although ideally there should be the ability to recognize novel object categories \cite{Lake2015Human,Hariharan2017Low}. Current detection datasets \cite{Everingham2010,Russakovsky2015,Lin2014} contain only a few dozen to hundreds of categories, significantly fewer than those which can be recognized by humans. New larger-scale datasets \cite{Hoffman2014lsda,Singh2018RFCN,YOLO9000} with significantly more categories will need to be developed. \textbf{(2) Better and More Efficient Detection Frameworks:} One of the reasons for the success in generic object detection has been the development of superior detection frameworks, both region-based (RCNN \cite{Girshick2014RCNN}, Fast RCNN \cite{Girshick2015FRCNN}, Faster RCNN \cite{Ren2015NIPS}, Mask RCNN \cite{MaskRCNN2017}) and one-stage detectors (YOLO \cite{YoLo2016}, SSD \cite{Liu2016SSD}). Region-based detectors have higher accuracy, one-stage detectors are generally faster and simpler. Object detectors depend heavily on the underlying backbone networks, which have been optimized for image classification, possibly causing a learning bias; learning object detectors from scratch could be helpful for new detection frameworks. \textbf{(3) Compact and Efficient CNN Features:} CNNs have increased remarkably in depth, from several layers (AlexNet \cite{AlexNet2012}) to hundreds of layers (ResNet \cite{He2016ResNet}, DenseNet \cite{Huang2016Densely}). These networks have millions to hundreds of millions of parameters, requiring massive data and GPUs for training. In order reduce or remove network redundancy, there has been growing research interest in designing compact and lightweight networks \cite{Chen2017Learning,Alvarez2016Learning,CondenseNet18,Howard2017MobileNets, Lin2017Towards,Yu2017NISP} and network acceleration \cite{Cheng2018Model,Hubara2016Binarized,Han2016Deep,Li2017Pruning,Li2017Mimicking,Wei2018Quantization}. \textbf{(4) Automatic Neural Architecture Search:} Deep learning bypasses manual feature engineering which requires human experts with strong domain knowledge, however DCNNs require similarly significant expertise. It is natural to consider automated design of detection backbone architectures, such as the recent Automated Machine Learning (AutoML) \cite{Quanming2018Taking}, which has been applied to image classification and object detection \cite{Cai2018Path,Chen2019DetNAS,Ghiasi2019NASFPN,Liu2018Progressive, Zoph2016Neural,Zoph2018Learning}. \textbf{(5) Object Instance Segmentation:} For a richer and more detailed understanding of image content, there is a need to tackle pixel-level object instance segmentation \cite{Lin2014,MaskRCNN2017,Ronghang2018}, which can play an important role in potential applications that require the precise boundaries of individual objects. \textbf{(6) Weakly Supervised Detection:} Current state-of-the-art detectors employ fully supervised models learned from labeled data with object bounding boxes or segmentation masks \cite{Everingham2015,Lin2014,Russakovsky2015,Lin2014}. However, fully supervised learning has serious limitations, particularly where the collection of bounding box annotations is labor intensive and where the number of images is large. Fully supervised learning is not scalable in the absence of fully labeled training data, so it is essential to understand how the power of CNNs can be leveraged where only weakly / partially annotated data are provided \cite{Bilen2016Weakly,Diba2017Weakly,Shi2017PAMI}. \textbf{(7) Few / Zero Shot Object Detection:} The success of deep detectors relies heavily on gargantuan amounts of annotated training data. When the labeled data are scarce, the performance of deep detectors frequently deteriorates and fails to generalize well. In contrast, humans (even children) can learn a visual concept quickly from very few given examples and can often generalize well \cite{Biederman1987Recognition,Lake2015Human,Fei2006One}. Therefore, the ability to learn from only few examples, \emph{few} shot detection, is very appealing \cite{Chen2018LSTD,Dong2018Few,Finn2017Model, Kang2018Few,Lake2015Human,Ren2018Meta, Schwartz2019RepMet}. Even more constrained, \emph{zero} shot object detection localizes and recognizes object classes that have never been seen\footnote{Although side information may be provided, such as a wikipedia page or an attributes vector.} before \cite{Bansa2018Zero,Demirel2018Zero,Rahman2018Zero,Rahman2018Polarity}, essential for life-long learning machines that need to intelligently and incrementally discover new object categories. \textbf{(8) Object Detection in Other Modalities:} Most detectors are based on still 2D images; object detection in other modalities can be highly relevant in domains such as autonomous vehicles, unmanned aerial vehicles, and robotics. These modalities raise new challenges in effectively using depth \cite{Chen20153D,Pepik2015GCPR,Xiang2014Beyond,Wu20153D}, video \cite{Feichtenhofer17Detect,Kang2016Object}, and point clouds \cite{Qi2017PointNet,Qi2018Frustum}. \textbf{(9) Universal Object Detection:} Recently, there has been increasing effort in learning \emph{universal representations}, those which are effective in multiple image domains, such as natural images, videos, aerial images, and medical CT images \cite{Rebuffi2017Learning,Rebuffi2018Efficient}. Most such research focuses on image classification, rarely targeting object detection \cite{Wang2019Towards}, and developed detectors are usually domain specific. Object detection independent of image domain and cross-domain object detection represent important future directions. The research field of generic object detection is still far from complete. However given the breakthroughs over the past five years we are optimistic of future developments and opportunities. \begin{table*}[!t] \begin{sideways} \begin{minipage}{\textheight} \centering \caption {Summary of properties and performance of milestone detection frameworks for generic object detection. See Section~\ref{Sec:Frameworks} for a detailed discussion. Some architectures are illustrated in Fig.~\ref{Fig:RegionVsUnified}. The properties of the backbone DCNNs can be found in Table~\ref{Tab:dcnnarchitectures}. }\label{Tab:Detectors} \renewcommand{\arraystretch}{1} \setlength\arrayrulewidth{0.2mm} \setlength\tabcolsep{1pt} \resizebox*{!}{16cm}{ \begin{tabular}{!{\vrule width1.5bp}c|c|c|c|c|c|c|c|c|c|p{10cm}!{\vrule width1.5bp}} \Xhline{1.5pt} &\scriptsize \shortstack [c] { Detector\\Name} & \scriptsize RP & \scriptsize \shortstack [c] { Backbone\\DCNN} & \scriptsize \shortstack [c] {Input \\ ImgSize} & \scriptsize \shortstack [c] {VOC07 \\Results}& \scriptsize \shortstack [c] { VOC12\\Results }& \scriptsize \shortstack [c] { Speed \\(FPS) } & \scriptsize \shortstack [c] {Published \\ In}& \scriptsize \shortstack [c] { Source\\Code } & \scriptsize Highlights and Disadvantages \\ \Xhline{1.5pt} \multirow{7}*{\hfil \rotatebox{90}{\footnotesize \textbf{Region based} (Section \ref{Sec:RegionBased})$\quad\quad\quad\quad\quad\quad\quad\quad$ }} &\raisebox{-4.3ex}[0pt]{ \scriptsize RCNN \cite{Girshick2014RCNN} }&\raisebox{-4.3ex}[0pt]{\scriptsize SS} & \raisebox{-4.3ex}[0pt]{\scriptsize AlexNet } & \raisebox{-4.3ex}[0pt]{\scriptsize Fixed } &\raisebox{-5.3ex}[0pt]{ \scriptsize \shortstack [c] { $58.5$ \\ (07)}} & \scriptsize \raisebox{-6.3ex}[0pt]{ \scriptsize \shortstack [c] { $53.3$ \\ (12)}} &\raisebox{-4.3ex}[0pt]{ \scriptsize $<0.1$} &\raisebox{-4.3ex}[0pt]{ \scriptsize CVPR14} &\raisebox{-4.3ex}[0pt]{\scriptsize \shortstack [c] {Caffe \\ Matlab}} & \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} First to integrate CNN with RP methods; Dramatic performance improvement over previous state of the artP. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Multistage pipeline of sequentially-trained (External RP computation, CNN finetuning, each warped RP passing through CNN, SVM and BBR training); Training is expensive in space and time; Testing is slow. \\ \cline{2-11} &\raisebox{-5ex}[0pt]{ \scriptsize SPPNet \cite{He2014SPP} }&\raisebox{-5ex}[0pt]{\scriptsize SS }&\raisebox{-5ex}[0pt]{ \scriptsize ZFNet } &\raisebox{-5ex}[0pt]{ \scriptsize Arbitrary } & \scriptsize \raisebox{-7ex}[0pt]{ \scriptsize \shortstack [c] { $60.9$ \\ (07)}} & \scriptsize \raisebox{-7ex}[0pt]{ \scriptsize $-$ }&\raisebox{-5ex}[0pt]{ \scriptsize $<1$}&\raisebox{-5ex}[0pt]{ \scriptsize ECCV14} &\raisebox{-5ex}[0pt]{\scriptsize \shortstack [c] {Caffe \\ Matlab}} & \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} First to introduce SPP into CNN architecture; Enable convolutional feature sharing; Accelerate RCNN evaluation by orders of magnitude without sacrificing performance; Faster than OverFeat. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Inherit disadvantages of RCNN; Does not result in much training speedup; Fine-tuning not able to update the CONV layers before SPP layer. \\ \cline{2-11} & \raisebox{-4.5ex}[0pt]{\scriptsize Fast RCNN \cite{Girshick2015FRCNN}} & \raisebox{-4.5ex}[0pt]{\scriptsize SS} & \raisebox{-7ex}[0pt]{ \scriptsize \shortstack [c] { AlexNet\\VGGM\\VGG16 }}&\raisebox{-4.5ex}[0pt]{ \scriptsize Arbitrary } & \raisebox{-7ex}[0pt]{ \scriptsize \shortstack [c] { $70.0$ \\ (VGG)\\(07+12)}} &\raisebox{-7ex}[0pt]{ \scriptsize \shortstack [c] { $68.4$ \\ (VGG)\\(07++12)}}& \raisebox{-4.5ex}[0pt]{\scriptsize $<1$ }& \raisebox{-4.5ex}[0pt]{\scriptsize ICCV15 } &\raisebox{-4.5ex}[0pt]{\scriptsize \shortstack [c] {Caffe \\ Python}} & \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} First to enable end-to-end detector training (ignoring RP generation); Design a RoI pooling layer; Much faster and more accurate than SPPNet; No disk storage required for feature caching. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} External RP computation is exposed as the new bottleneck; Still too slow for real time applications. \\ \cline{2-11} &\raisebox{-6.5ex}[0pt]{ \scriptsize Faster RCNN \cite{Ren2015NIPS} }&\raisebox{-6.5ex}[0pt]{\scriptsize RPN } & \raisebox{-7ex}[0pt]{ \scriptsize \shortstack [c] { ZFnet\\VGG }}&\raisebox{-6.5ex}[0pt]{ \scriptsize Arbitrary } & \raisebox{-9ex}[0pt]{ \scriptsize \shortstack [c] { $73.2$ \\ (VGG)\\(07+12)}} &\raisebox{-9ex}[0pt]{ \scriptsize \shortstack [c] { $70.4$ \\ (VGG)\\(07++12)}}& \raisebox{-6.5ex}[0pt]{\scriptsize $<5$}& \raisebox{-6.5ex}[0pt]{\scriptsize NIPS15} &\raisebox{-6.5ex}[0pt]{\scriptsize \shortstack [c] {Caffe \\ Matlab\\Python}}& \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} Propose RPN for generating nearly cost-free and high quality RPs instead of selective search; Introduce translation invariant and multiscale anchor boxes as references in RPN; Unify RPN and Fast RCNN into a single network by sharing CONV layers; An order of magnitude faster than Fast RCNN without performance loss; Can run testing at 5 FPS with VGG16. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Training is complex, not a streamlined process; Still falls short of real time. \\ \cline{2-11} & \raisebox{-2ex}[0pt]{\scriptsize RCNN$\ominus$R \cite{Lenc2015} }& \raisebox{-2ex}[0pt]{\scriptsize New} & \raisebox{-4ex}[0pt]{ \scriptsize \shortstack [c] { ZFNet\\+SPP}}&\raisebox{-2ex}[0pt]{ \scriptsize Arbitrary } & \raisebox{-4ex}[0pt]{ \scriptsize \shortstack [c] { $59.7$ \\(07)}} & \raisebox{-2ex}[0pt]{ \scriptsize $-$ }& \raisebox{-2ex}[0pt]{ \scriptsize $<5$ }& \raisebox{-2ex}[0pt]{ \scriptsize BMVC15} &\raisebox{-2ex}[0pt]{\scriptsize $-$} & \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} Replace selective search with static RPs; Prove the possibility of building integrated, simpler and faster detectors that rely exclusively on CNN. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Falls short of real time; Decreased accuracy from poor RPs. \\ \cline{2-11} &\raisebox{-3.5ex}[0pt]{ \scriptsize RFCN \cite{Dai2016RFCN} }&\raisebox{-3.5ex}[0pt]{ \scriptsize RPN} & \raisebox{-3.5ex}[0pt]{\scriptsize ResNet101 } &\raisebox{-3.5ex}[0pt]{ \scriptsize Arbitrary } & \raisebox{-6.5ex}[0pt]{ \scriptsize \shortstack [c] { $80.5$\\(07+12)\\$83.6$\\(07+12+CO)}} &\raisebox{-6.5ex}[0pt]{ \scriptsize \shortstack [c] { $77.6$ \\(07++12)\\$82.0$\\(07++12+CO)}}&\raisebox{-3.5ex}[0pt]{ \scriptsize $<10$}&\raisebox{-3.5ex}[0pt]{ \scriptsize NIPS16} &\raisebox{-3.5ex}[0pt]{\scriptsize \shortstack [c] {Caffe \\ Matlab}}& \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} Fully convolutional detection network; Design a set of position sensitive score maps using a bank of specialized CONV layers; Faster than Faster RCNN without sacrificing much accuracy. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Training is not a streamlined process; Still falls short of real time. \\ \cline{2-11} &\raisebox{-4.5ex}[0pt]{ \scriptsize Mask RCNN \cite{MaskRCNN2017}} & \raisebox{-4.5ex}[0pt]{\scriptsize RPN }& \raisebox{-5.5ex}[0pt]{\scriptsize \shortstack [c] { ResNet101 \\ ResNeXt101 } } &\raisebox{-4.5ex}[0pt]{ \scriptsize Arbitrary }& \multicolumn{2}{|c|}{\raisebox{-7ex}[0pt]{\scriptsize \shortstack [c] { $50.3$\\(ResNeXt101)\\ (COCO Result)}}} & \raisebox{-4.5ex}[0pt]{\scriptsize $<5$} & \raisebox{-4.5ex}[0pt]{\scriptsize ICCV17 } &\raisebox{-4.5ex}[0pt]{\scriptsize \shortstack [c] {Caffe \\ Matlab\\Python}}& \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} A simple, flexible, and effective framework for object instance segmentation; Extends Faster RCNN by adding another branch for predicting an object mask in parallel with the existing branch for BB prediction; Feature Pyramid Network (FPN) is utilized; Outstanding performance. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Falls short of real time applications. \\ \Xhline{1.5pt} \multirow{4}{*}{\rotatebox{90}{\footnotesize \textbf{Unified } (Section \ref{Sec:Unified}) $\quad\quad\quad\quad\quad\quad$ }} & \raisebox{-5.5ex}[0pt]{\scriptsize OverFeat \cite{OverFeat2014} } & \raisebox{-5.5ex}[0pt]{ \scriptsize $-$} & \raisebox{-5.5ex}[0pt]{\scriptsize AlexNet like }&\raisebox{-5.5ex}[0pt]{ \scriptsize Arbitrary }&\raisebox{-5.5ex}[0pt]{$-$}&\raisebox{-5.5ex}[0pt]{$-$}& \raisebox{-5.5ex}[0pt]{\scriptsize $<0.1$ }& \raisebox{-5.5ex}[0pt]{\scriptsize ICLR14 } &\raisebox{-5.5ex}[0pt]{\scriptsize c++}& \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} Convolutional feature sharing; Multiscale image pyramid CNN feature extraction; Won the ISLVRC2013 localization competition; Significantly faster than RCNN. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Multi-stage pipeline sequentially trained; Single bounding box regressor; Cannot handle multiple object instances of the same class; Too slow for real time applications. \\ \cline{2-11} & \raisebox{-4.5ex}[0pt]{\scriptsize YOLO \cite{YoLo2016} } & \raisebox{-4.5ex}[0pt]{\scriptsize $-$ }& \raisebox{-4.5ex}[0pt]{ \scriptsize \shortstack [c] { GoogLeNet\\like }} &\raisebox{-4.5ex}[0pt]{ \scriptsize Fixed }& \raisebox{-5ex}[0pt]{ \scriptsize \shortstack [c] { $66.4$ \\(07+12)}} &\raisebox{-5ex}[0pt]{ \scriptsize \shortstack [c] { $57.9$ \\(07++12)}} & \raisebox{-4.5ex}[0pt]{\scriptsize \shortstack [c] { $<25$\\(VGG)}} & \raisebox{-4.5ex}[0pt]{\scriptsize CVPR16} &\raisebox{-4.5ex}[0pt]{\scriptsize DarkNet} & \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} First efficient unified detector; Drop RP process completely; Elegant and efficient detection framework; Significantly faster than previous detectors; YOLO runs at 45 FPS, Fast YOLO at 155 FPS; \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Accuracy falls far behind state of the art detectors; Struggle to localize small objects. \\ \cline{2-11} & \raisebox{-3.5ex}[0pt]{ \scriptsize YOLOv2\cite{YOLO9000}} & \raisebox{-3.5ex}[0pt]{\scriptsize $-$ } & \raisebox{-3.5ex}[0pt]{\scriptsize DarkNet} &\raisebox{-3.5ex}[0pt]{ \scriptsize Fixed } & \raisebox{-4ex}[0pt]{ \scriptsize \shortstack [c] { $78.6$ \\(07+12)}} &\raisebox{-4ex}[0pt]{ \scriptsize \shortstack [c] { $73.5$ \\(07++12)}} & \scriptsize $<50$ & \raisebox{-3.5ex}[0pt]{ \scriptsize CVPR17} &\raisebox{-3.5ex}[0pt]{\scriptsize DarkNet} & \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} Propose a faster DarkNet19; Use a number of existing strategies to improve both speed and accuracy; Achieve high accuracy and high speed; YOLO9000 can detect over 9000 object categories in real time. \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Not good at detecting small objects. \\ \cline{2-11} & \raisebox{-3.5ex}[0pt]{\scriptsize SSD \cite{Liu2016SSD} }& \raisebox{-3.5ex}[0pt]{ \scriptsize $-$ } & \raisebox{-3.5ex}[0pt]{ \scriptsize VGG16} &\raisebox{-3.5ex}[0pt]{ \scriptsize Fixed } & \raisebox{-6.5ex}[0pt]{ \scriptsize \shortstack [c] { $76.8$\\(07+12)\\$81.5$\\(07+12+CO)}} &\raisebox{-6.5ex}[0pt]{ \scriptsize \shortstack [c] { $74.9$ \\(07++12)\\$80.0$\\(07++12+CO)}}& \raisebox{-3.5ex}[0pt]{ \scriptsize $<60$}& \raisebox{-3.5ex}[0pt]{ \scriptsize ECCV16} &\raisebox{-3.5ex}[0pt]{\scriptsize \shortstack [c] {Caffe \\ Python}} & \scriptsize \textcolor{DarkGreen}{\textbf{Highlights:}} First accurate and efficient unified detector; Effectively combine ideas from RPN and YOLO to perform detection at multi-scale CONV layers; Faster and significantly more accurate than YOLO; Can run at 59 FPS; \par \textcolor{DarkRed}{\textbf{Disadvantages:}} Not good at detecting small objects. \\ \Xhline{1.5pt} \multicolumn{9}{c}{$\quad$}\\ \end{tabular} } \par \raggedright \small{\emph{Abbreviations in this table: Region Proposal (RP), Selective Search (SS), Region Proposal Network (RPN), RCNN$\ominus$R represents ``RCNN minus R'' and used a trivial RP method. Training data: ``07''$\leftarrow$VOC2007 trainval; ``07T''$\leftarrow$VOC2007 trainval and test; ``12''$\leftarrow$VOC2012 trainval; ``CO''$\leftarrow$COCO trainval. The ``Speed'' column roughly estimates the detection speed with a single Nvidia Titan X GPU.}} \end{minipage} \end{sideways} \end{table*} \section{Acknowledgments} The authors would like to thank the pioneering researchers in generic object detection and other related fields. The authors would also like to express their sincere appreciation to Professor Ji\v{r}\'{\i} Matas, the associate editor and the anonymous reviewers for their comments and suggestions. This work has been supported by the Center for Machine Vision and Signal Analysis at the University of Oulu (Finland) and the National Natural Science Foundation of China under Grant 61872379. \bibliographystyle{spbasic} \footnotesize

e711aba737b3141567227a825af5fa5d57d00756

\section*{Introduction} For a compact complex space $X$, Kobayashi hyperbolicity is equivalent to the fact that every holomorphic map $\mathbb{C}\to X$ is constant, thanks to a classical result of Brody. When $X$ is moreover projective (or, more generally, compact K\"ahler), hyperbolicity is further expected to be completely characterized by (algebraic) positivity properties of $X$ and of its subvarieties. More precisely, we have the following conjecture, due to S.~Lang. \begin{conj}\cite[Conjecture 5.6]{Lan86} A projective variety $X$ is hyperbolic if and only if every subvariety (including $X$ itself) is of general type. \end{conj} Recall that a projective variety $X$ is of general type if the canonical bundle of any smooth projective birational model of $X$ is big, \textsl{i.e.}~has maximal Kodaira dimension. This is for instance the case when $X$ is smooth and \emph{canonically polarized}, \textsl{i.e.} with an ample canonical bundle $K_X$. Note that Lang's conjecture in fact implies that every smooth hyperbolic projective manifold $X$ is canonically polarized, as conjectured in 1970 by S.~Kobayashi. It is indeed a well-known consequence of the Minimal Model Program that any projective manifold of general type without rational curves is canonically polarized (see for instance~\cite[Theorem A]{BBP}). Besides the trivial case of curves and partial results for surfaces~\cite{MM83,DES79,GG80,McQ98}, Lang's conjecture is still almost completely open in higher dimension as of this writing. General projective hypersurfaces of high degree in projective space form a remarkable exception: they are known to be hyperbolic~\cite{Bro17} (see also~\cite{McQ99,DEG00,DT10,Siu04,Siu15,RY18}), and they satisfy Lang's conjecture~\cite{Cle86,Ein88,Xu94,Voi96,Pac04}. \medskip It is natural to test Lang's conjecture for the following two basic classes of manifolds, known to be hyperbolic since the very beginning of the theory: \begin{itemize} \item[(N)] compact K\"ahler manifolds $X$ with negative holomorphic sectional curvature; \item[(B)] compact, free quotients $X$ of bounded domains $\Omega\Subset\mathbb{C}^n$. \end{itemize} In case (N), ampleness of $K_X$ was established in~\cite{WY16a,WY16b,TY17} (see also~\cite{DT16}). By curvature monotonicity, this implies that every smooth subvariety of $X$ also has ample canonical bundle. More generally, Guenancia recently showed~\cite{Gue18} that each (possibly singular) subvariety of $X$ is of general type, thereby verifying Lang's conjecture in that case. One might even more generally consider the case where $X$ carries an arbitrary Hermitian metric of negative holomorphic sectional curvature, which seems to be still open. \medskip In this note, we confirm Lang's conjecture in case (B). While the case of quotients of bounded \emph{symmetric} domains has been widely studied (see, just to cite a few,~\cite{Nad89,BKT13,Bru16,Cad16,Rou16,RT18}), the general case seems to have somehow passed unnoticed. Instead of bounded domains, we consider more generally the following class of manifolds, which comprises relatively compact domains in Stein manifolds, and has the virtue of being stable under passing to an \'etale cover or a submanifold. \begin{defiint} We say that a complex manifold $M$ is \emph{of bounded type} if it carries a bounded, strictly plurisubharmonic function $\varphi$. \end{defiint} By a well-known result of Richberg, any \emph{continuous} bounded strictly psh function on a complex manifold $M$ can be written as a decreasing limit of smooth strictly psh functions, but this fails in general for discontinuous functions~\cite[p.66]{For}, and it is thus unclear to us whether every manifold of bounded type should carry also a \emph{smooth} bounded strictly psh function. \begin{thmA}\label{thm:main} Let $X$ be a compact K\"ahler manifold admitting an \'etale (Galois) cover $\tilde X\to X$ of bounded type. Then: \begin{itemize} \item[(i)] $X$ is Kobayashi hyperbolic; \item[(ii)] $X$ has large fundamental group; \item[(iii)] $X$ is projective and canonically polarized; \item[(iv)] every subvariety of $X$ is of general type. \end{itemize} \end{thmA} Note that $\tilde X$ can always be replaced with the universal cover of $X$, and hence can be assumed to be Galois. \medskip By~\cite[3.2.8]{Kob98}, (i) holds iff $\tilde X$ is hyperbolic, which follows from the fact that manifolds of bounded type are Kobayashi hyperbolic~\cite[Theorem 3]{Sib81}. Alternatively, any entire curve $f:\mathbb{C}\to X$ lifts to $\tilde X$, and the pull-back to $\mathbb{C}$ of the bounded, strictly psh function carried by $\tilde X$ has to be constant, showing that $f$ itself is constant. \smallskip By definition, (ii) means that the image in $\pi_1(X)$ of the fundamental group of any subvariety $Z\subseteq X$ is infinite~\cite[\S 4.1]{Kol}, and is a direct consequence of the fact that manifolds of bounded type do not contain nontrivial compact subvarieties. According to the Shafarevich conjecture, $\tilde X$ should in fact be Stein; in case $\tilde X$ is a bounded domain of $\mathbb{C}^n$, this is indeed a classical result of Siegel~\cite{Sie50} (see also ~\cite[Theorem 6.2]{Kob59}). \smallskip By another classical result, this time due to Kodaira~\cite{Kod}, any compact complex manifold $X$ admitting a Galois \'etale cover $\tilde X\to X$ biholomorphic to a bounded domain in $\mathbb{C}^n$ is projective, with $K_X$ ample. Indeed, the Bergman metric of $\tilde X$ is non-degenerate, and it descends to a positively curved metric on $K_X$. Our proof of (iii) and (iv) is a simple variant of this idea, inspired by~\cite{CZ02}. For each subvariety $Y\subseteq X$ with desingularization $Z\to Y$ and induced Galois \'etale cover $\tilde Z\to Z$, we use basic H\"ormander--Andreotti--Vesentini--Demailly $L^2$-estimates for $\overline{\partial}$ to show that the Bergman metric of $\tilde Z$ is generically non-degenerate. It then descends to a psh metric on $K_Z$, smooth and strictly psh on a nonempty Zariski open set, which is enough to conclude that $K_Z$ is big, by~\cite{Bou02}. \medskip As a final comment, note that K\"ahler hyperbolic manifolds, \textsl{i.e.} compact K\"ahler manifolds $X$ carrying a K\"ahler metric $\omega$ whose pull-back to the universal cover $\pi:\tilde X\to X$ satisfies $\pi^*\omega=d\a$ with $\a$ bounded, also satisfy (i)--(iii) in Theorem A \cite{Gro}. It would be interesting to check Lang's conjecture for such manifolss as well. \begin{ackn} This work was started during the first-named author's stay at SAPIENZA Universit\`a di Roma. He is very grateful to the mathematics department for its hospitality, and to INdAM for financial support. Both authors would also like to thank Stefano Trapani for helpful discussions, in particular for pointing out the reference~\cite{For}. \end{ackn} \section{The Bergman metric and manifolds of general type} \subsection{Non-degeneration of the Bergman metric} Recall that the \emph{Bergman space} of a complex manifold $M$ is the separable Hibert space $\mathcal{H}=\mathcal{H}(M)$ of holomorphic forms $\eta\in H^0(M,K_M)$ such that $$ \|\eta\|_\mathcal{H}^2:=i^{n^2}\int_{\tilde X}\eta\wedge\bar\eta<\infty, $$ with $n=\dim M$. Assuming $\mathcal{H}\ne\{0\}$, we get an induced (possibly singular) psh metric $h_M$ on $K_M$, invariant under $\Aut(M)$, characterized pointwise by $$ h/h_M=\sup_{\eta\in\mathcal{H}\setminus\{0\}}\frac{|\eta|^2_h}{\|\eta\|_\mathcal{H}^2}=\sum_j |\eta_j|^2_h, $$ for any choice of smooth metric $h$ on $K_M$ and orthonormal basis $(\eta_j)$ for $\mathcal{H}$ (see for instance~\cite[\S 4.10]{Kob98}). The curvature current of $h_M$ is classically called the \lq\lq Bergman metric\rq\rq{} of $M$; it is a \textsl{bona fide} K\"ahler form precisely on the Zariski open subset of $M$ consisting of points at which $\mathcal{H}$ generates $1$-jets~\cite[Proposition 4.10.11]{Kob98}. \begin{defi} We shall say that a complex manifold $M$ has a \emph{non-degenerate (resp.~generically non-degenerate) Bergman metric} if its Bergman space $\mathcal{H}$ generates $1$-jets at each (resp.~some) point of $M$. \end{defi} We next recall the following standard consequence of $L^2$-estimates for $\overline{\partial}$. \begin{lem}\label{lem:1jet} Let $M$ be a complete K\"ahler manifold with a bounded psh function $\varphi$. If $\varphi$ is strictly psh on $M$ (resp.~at some point of $M$), then the Bergman metric of $M$ is non-degenerate (resp.~generically non-degenerate). \end{lem} \begin{proof} Pick a complete K\"ahler metric $\omega$ on $M$. Assume $\varphi$ strictly psh at $p\in M$, and fix a coordinate ball $(U,z)$ centered at $p$ with $\varphi$ strictly psh near $\overline U$. Pick also $\chi\in C^\infty_c(U)$ with $\chi\equiv 1$ near $p$. Since $\chi\log|z|$ is strictly psh in an open neighbourhood $V$ of $p$, smooth on $U\setminus\overline V$, and compactly supported in $U$, we can then choose $A\gg 1$ such that $$ \psi:=(n+1)\chi\log|z|+A\varphi $$ is psh on $M$, with $dd^c\psi\ge\omega$ on $U$. Note that $\psi$ is also bounded above on $M$, $\varphi$ being assumed to be bounded. For an appropriate choice of holomorphic function $f$ on $U$, the smooth $(n,0)$-form $\eta:=\chi f\,dz_1\wedge\dots\wedge dz_n$, which is compactly supported in $U$ and holomorphic in a neighborhood of $x$, will have any prescribed jet at $p$. The $(n,1)$-form $\bar\partial\eta$ is compactly supported in $U$, and identically zero in a neighborhood of $p$, so that $|\overline{\partial}\eta|_\omega e^{-\psi}\in L^2(U)$. Since $dd^c\psi\ge\omega$ on $U$,~\cite[Th\'eor\`eme 5.1]{Dem82} yields an $L^2_{\mathrm{loc}}$ $(n,0)$-form $u$ on $M$ such that $\overline{\partial} u=\overline{\partial}\eta$ and \begin{equation}\label{equ:l2} i^{n^2} \int_M u\wedge\bar u\,e^{-2\psi}\le\int_U|\overline{\partial}\eta|^2_\omega e^{-2\psi}dV_\omega. \end{equation} As a result, $v:=\eta-u$ is a holomorphic $n$-form on $X$. Since $u=\eta-v$ is holomorphic at $x$ and $\psi$ has an isolated singularity of type $(n+1)\log|z|$ at $x$, (\ref{equ:l2}) forces $u$ to vanish to order $2$ at $p$, so that $v$ and $\eta$ have the same $1$-jet at $p$. Finally, (\ref{equ:l2}) and the fact that $\psi$ is bounded above on $M$ shows that $u$ is $L^2$. Since $\eta$ is clearly $L^2$ as well, $v$ belongs to the Bergman space $\mathcal{H}$, with given $1$-jet at $p$, and we are done. \end{proof} \subsection{Manifolds of general type} Let $X$ be a compact complex manifold, $\tilde X\to X$ a Galois \'etale cover, and assume that the Bergman metric of $\tilde X$ is non-degenerate, so that the canonical metric $h_{\tilde X}$ on $K_{\tilde X}$ defined by $\mathcal{H}(\tilde X)$ is smooth, strictly psh. Being invariant under automorphisms, this metric descends to a smooth, strictly psh metric on $K_X$, and the latter is thus ample by \cite{Kod}. This argument, which goes back to the same paper by Kodaira, admits the following variant. \begin{lem}\label{lem:big} Let $X$ be a compact K\"ahler manifold admitting a Galois \'etale cover $\tilde X\to X$ with generically non-degenerate Bergman metric. Then $X$ is projective and of general type. \end{lem} \begin{proof} The assumption now means that the psh metric $h_{\tilde X}$ on $K_{\tilde X}$ is smooth and strictly psh on a non-empty Zariski open subset. It descends again to a psh metric on $K_X$, smooth and strictly psh on a non-empty Zariski open subset, and we conclude that $K_X$ is big by~\cite[\S 2.3]{Bou02} (see also~\cite[\S 1.5]{BEGZ10}). Being both Moishezon and K\"ahler, $X$ is then projective. \end{proof} \section{Proof of Theorem A} Let $X$ be a compact K\"ahler manifold with an \'etale cover $\pi:\tilde X\to X$ of bounded type, which may be assumed to be Galois after replacing $\tilde X$ by the universal cover of $X$. Since $\tilde X$ is also complete K\"ahler, its Bergman metric is non-degenerate by Lemma~\ref{lem:1jet}, and $X$ is thus projective and canonically polarized by \cite{Kod}. Now let $Y\subseteq X$ be an irreducible subvariety. On the one hand, pick any connected component $\tilde Y$ of the preimage $\pi^{-1}(Y)\subset\tilde X$, so that $\pi$ induces a Galois \'etale cover $\pi|_{\tilde Y}\colon\tilde Y\to Y$. On the other hand, let $\mu\colon Z\to Y$ be a projective modification with $Z$ smooth and $\mu$ isomorphic over $Y_{\reg}$, whose existence is guaranteed by Hironaka. Since $Y$ is K\"ahler and $\mu$ is projective, $Z$ is then a compact K\"ahler manifold. The fiber product $\tilde Z=Z\times_{Y }\tilde Y$ sits in the following diagram $$ \xymatrix{ \tilde Z \ar[dr]^{\tilde\mu}\ar[dd]_\nu & &\\ & \tilde Y \ar@{^{(}->}[r] \ar[dd]^{\pi|_{\tilde Y}}& \tilde X\ar[dd]^\pi \\ Z \ar[dr]_\mu & &\\ & Y \ar@{^{(}->}[r] & X. } $$ Being a base change of a Galois \'etale cover, $\nu$ is a Galois \'etale cover, and $\tilde\mu$ is a resolution of singularities of $\tilde Y$. Since $\pi$ is \'etale, we have $\tilde Y_{\reg}=\pi^{-1}(Y_{\reg})$, and $\tilde\mu$ is an isomorphism over $\tilde Y_{\reg}$. The pull-back of $\varphi$ to $\tilde Z$ is thus a bounded psh function, strictly psh at any point $p\in\tilde\mu^{-1}(\tilde Y_{\reg})$. Since $Z$ is compact K\"ahler, $\tilde Z$ is complete K\"ahler. By Lemma~\ref{lem:1jet}, the Bergman metric of $Z$ is generically non-degenerate, and $Z$ is thus of general type, by Lemma~\ref{lem:big}.

6a9da3a672d8aa8a5100c05f72460498a86d21cb

\section{Introduction} \label{sec:intro} In the field of natural language processing (NLP), one of the most prevalent neural approaches to obtaining sentence representations is to use recurrent neural networks (RNNs), where words in a sentence are processed in a sequential and recurrent manner. Along with their intuitive design, RNNs have shown outstanding performance across various NLP tasks e.g. language modeling \citep{mikolov2010rnnlm,graves2013generating}, machine translation \citep{cho2014nmt,sutskever2014sequence,bahdanau2015nmt}, text classification \citep{zhou2015c,tang2015document}, and parsing \citep{kiperwasser2016parsing,dyer2016rnng}. Among several variants of the original RNN \citep{elman1990finding}, gated recurrent architectures such as long short-term memory (LSTM) \citep{hochreiter1997long} and gated recurrent unit (GRU) \citep{cho2014nmt} have been accepted as de-facto standard choices for RNNs due to their capability of addressing the vanishing and exploding gradient problem and considering long-term dependencies. Gated RNNs achieve these properties by introducing additional gating units that learn to control the amount of information to be transferred or forgotten \citep{goodfellow2016deeplearningboook}, and are proven to work well without relying on complex optimization algorithms or careful initialization \citep{sutskever2013training}. Meanwhile, the common practice for further enhancing the expressiveness of RNNs is to stack multiple RNN layers, each of which has distinct parameter sets (stacked RNN) \citep{schmidhuber1992learning,el1996hierarchical}. In stacked RNNs, the hidden states of a layer are fed as input to the subsequent layer, and they are shown to work well due to increased depth \citep{pascanu2014construct} or their ability to capture hierarchical time series \citep{hermans2013training} which are inherent to the nature of the problem being modeled. \begin{figure}[tb] \centering \subfigure{% \includegraphics[width=0.36\textwidth]{plain_stacked_simple.pdf} } \quad \subfigure{% \includegraphics[width=0.36\textwidth]{ca_stacked_simple.pdf} } \caption{ Visualization of (\textit{a}) plain stacked LSTM and (\textit{b}) CAS-LSTM. The red nodes indicate the blocks whose cell states directly affect the cell state $\mathbf{c}_t^l$. } \label{fig:comparison} \end{figure} However this setting of stacking RNNs might hinder the possibility of more sophisticated structures since the information from lower layers is simply treated as input to the next layer, rather than as another class of state that participates in core RNN computations. Especially for gated RNNs such as LSTMs and GRUs, this means that the vertical layer-to-layer connections cannot fully benefit from the carefully constructed gating mechanism used in temporal transitions. In this paper, we study a method of constructing multi-layer LSTMs where memory cell states from the previous layer are used in controlling the vertical information flow. This system utilizes states from the left and the lower context equally in computation of the new state, thus the information from lower layers is elaborately filtered and reflected through a soft gating mechanism. % Our method is easy-to-implement, effective, and can replace conventional stacked LSTMs without much modification of the overall architecture. We call this architecture Cell-aware Stacked LSTM, or CAS-LSTM, and evaluate our method on multiple benchmark tasks: natural language inference, paraphrase identification, sentiment classification, and machine translation. From experiments we show that the CAS-LSTMs consistently outperform typical stacked LSTMs, opening the possibility of performance improvement of architectures based on stacked LSTMs. Our contribution is summarized as follows. Firstly, we bring the idea of utilizing states coming from multiple directions to construction of stacked LSTM and apply the idea to the research of sentence representation learning. There is some prior work addressing the idea of incorporating more than one type of state \citep{graves2007mdrnn,kalchbrenner2016grid,zhang2016highway}, however to the best of our knowledge there is little work on applying the idea to modeling sentences for better understanding of natural language text. Secondly, we conduct extensive evaluation of the proposed method and empirically prove its effectiveness. The CAS-LSTM architecture provides consistent performance gains over the stacked LSTM in all benchmark tasks: natural language inference, paraphrase identification, sentiment classification, and machine translation. Especially in SNLI, SST-2, and Quora Question Pairs datasets, our models outperform or at least are on par with the state-of-the-art models. We also conduct thorough qualitative analysis to understand the dynamics of the suggested approach. This paper is organized in the following way. We study prior work related to our objective in \S\ref{sec:related}, and \S\ref{sec:caslstm} gives a detailed description about the proposed method. Experimental results are given in \S\ref{sec:experiments}, and \S\ref{sec:conclusion} concludes this paper. \section{Related Work} \label{sec:related} In this section, we summarize prior work related to the proposed method. We group the previous work that motivated our work into three classes: i) enhancing interaction between vertical layers, ii) RNN architectures that accepts latticed data, and iii) tree-structured RNNs. \paragraph{Stacked RNNs.} There is some prior work on methods of stacking RNNs beyond the plain stacked RNNs \citep{schmidhuber1992learning,el1996hierarchical}. Residual LSTMs \citep{kim2017residual,tran2017stack} add residual connections between the hidden states computed at each LSTM layer, and shortcut-stacked LSTMs \citep{nie2017shortcut} concatenate hidden states from all previous layers to make the backpropagation path short. In our method, the lower context is aggregated via a gating mechanism, and we believe it modulates the amount of information to be transmitted in a more efficient and effective way than vector addition or concatenation. Also, compared to concatenation, our method does not significantly increase the number of parameters.\footnote{The $l$-th layer of a typical stacked LSTM requires $(d_{l-1} + d_l + 1) \times 4d_l$ parameters, and the $l$-th layer of a shortcut-stacked LSTM requires $(\sum_{k=0}^{l-1} {d_k} + d_l + 1) \times 4d_l$ parameters. CAS-LSTM uses $(d_{l-1} + d_l + 1) \times 5d_l$ parameters at the $l$-th ($l>1$) layer.} Highway LSTMs \citep{zhang2016highway} and depth-gated LSTMs \citep{yao2015depth} are similar to our proposed models in that they use cell states from the previous layer, and they are successfully applied to the field of automatic speech recognition and language modeling. However in contrast to CAS-LSTM, where the additional forget gate aggregates the previous layer states and thus contexts from the left and below participate in computation equitably, in Highway LSTMs and depth-gated LSTMs the states from the previous time step are not considered in computing vertical gates. The comparison of our method and this architecture is presented in \S\ref{exp:variations}. \paragraph{Multidimensional RNNs.} There is another line of research that aims to extend RNNs to operate with multidimensional inputs. Grid LSTMs \citep{kalchbrenner2016grid} are a general $n$-dimensional LSTM architecture that accepts $n$ sets of hidden and cell states as input and yields $n$ sets of states as output, in contrast to our architecture, which emits a single set of states. In their work, the authors utilize 2D and 3D Grid LSTMs in character-level language modeling and machine translation respectively and achieve performance improvement. Multidimensional RNNs \citep{graves2007mdrnn,graves2009offline} have similar formulation to ours, except that they reflect cell states via simple summation and weights for all columns (vertical layers in our case) are tied. However they are only employed to model multidimensional data such as images of handwritten text with RNNs, rather than stacking RNN layers for modeling sequential data. From this view, CAS-LSTM could be interpreted as an extension of two-dimensional LSTM architecture that accepts a 2D input $\{\mathbf{h}_t^l\}_{t=1,l=0}^{T,L}$ where $\mathbf{h}_t^l$ represents the hidden state at time $t$ and layer $l$. \paragraph{Tree-structured RNNs.} The idea of having multiple states is also related to tree-structured RNNs \citep{goller1996learning,socher2011parsing}. Among them, tree-structured LSTMs (tree-LSTMs) \citep{tai2015treelstm,zhu2015treelstm,le2015treelstm} are similar to ours in that they use both hidden and cell states of children nodes. In tree-LSTMs, states of children nodes are regarded as input, and they participate in computing the states of a parent node equally through weight-shared or weight-unshared projection. From this perspective, each CAS-LSTM layer can be seen as a binary tree-LSTM where the structures it operates on are fixed to right-branching trees. Indeed, our work is motivated by the recent analysis \citep{williams2018do,shi2018ontree} on latent tree learning models \citep{yogatama2017learning,choi2018learning} which has shown that tree-LSTM models outperform the sequential LSTM models even when the resulting parsing strategy generates strictly left- or right-branching parses, where a tree-LSTM model should read words in the manner identical to a sequential LSTM model. We argue that the active use of cell state in computation could be one reason of these counter-intuitive results and empirically prove the hypothesis in this work. \section{Model Description} \label{sec:caslstm} \begin{figure}[tb] \centering \includegraphics[width=0.5\textwidth]{ca_block.pdf} \caption{Schematic diagram of a CAS-LSTM block.} \label{fig:diagram} \end{figure} In this section, we give the detailed formulation of architectures used in experiments. \subsection{Stacked LSTMs} \label{ssec:stacked-lstm} While there exist various versions of LSTM formulation, in this work we use the following, the most common variant: \begin{align} \mathbf{i}_t^l &= \sigma(\mathbf{W}_i^l \mathbf{h}_t^{l-1} + \mathbf{U}_i^l \mathbf{h}_{t-1}^l + \mathbf{b}_i^l) \\ \mathbf{f}_t^l &= \sigma(\mathbf{W}_f^l \mathbf{h}_t^{l-1} + \mathbf{U}_f^l \mathbf{h}_{t-1}^l + \mathbf{b}_f^l) \\ \tilde{\mathbf{c}}_t^l &= \tanh(\mathbf{W}_c^l \mathbf{h}_t^{l-1} + \mathbf{U}_c^l \mathbf{h}_{t-1}^l + \mathbf{b}_c^l) \\ \mathbf{o}_t^l &= \sigma(\mathbf{W}_o^l \mathbf{h}_t^{l-1} + \mathbf{U}_o^l \mathbf{h}_{t-1}^l + \mathbf{b}_o^l) \\ \mathbf{c}_t^l &= \mathbf{i}_t^l \odot \tilde{\mathbf{c}}_t^l + \mathbf{f}_t^l \odot \mathbf{c}_{t-1}^l \label{eq:conventional-cell}\\ \mathbf{h}_t^l &= \mathbf{o}_t^l \odot \tanh(\mathbf{c}_t^l), \end{align} where $t \in \{1,\cdots,T\}$ and $l \in \{1,\cdots,L\}$. $\mathbf{W}_{\cdot}^{l} \in \mathbb{R}^{d_l \times d_{l-1}}$, $\mathbf{U}_{\cdot}^{l} \in \mathbb{R}^{d_l \times d_l}$, $\mathbf{b}_{\cdot}^{l} \in \mathbb{R}^{d_l}$ are trainable parameters, and $\sigma(\cdot)$ and $\tanh(\cdot)$ are the sigmoid and the hyperbolic tangent function respectively. Also we assume that $\mathbf{h}_t^0=\mathbf{x}_t \in \mathbb{R}^{d_0}$ where $\mathbf{x}_t$ is the $t$-th element of an input sequence. The input gate $\mathbf{i}_t^l$ and the forget gate $\mathbf{f}_t^l$ control the amount of information transmitted from $\tilde{\mathbf{c}}_t^l$ and $\mathbf{c}_{t-1}^l$, the candidate cell state and the previous cell state, to the new cell state $\mathbf{c}_t^l$. Similarly the output gate $\mathbf{o}_t^l$ soft-selects which portion of the cell state $\mathbf{c}_t^l$ is to be used in the final hidden state. We can clearly see that the cell states $\mathbf{c}_{t-1}^l$, $\tilde{\mathbf{c}}_t^l$, $\mathbf{c}_t^l$ play a crucial role in forming horizontal recurrence. However the current formulation does not consider the cell state from $(l-1)$-th layer ($\mathbf{c}_t^{l-1}$) in computation and thus the lower context is reflected only through the rudimentary way, hindering the possibility of controlling vertical information flow. \subsection{Cell-aware Stacked LSTMs} Now we extend the stacked LSTM formulation defined above to address the problem noted in the previous subsection. To enhance the interaction between layers in a way similar to how LSTMs keep and forget the information from the previous time step, we introduce the \textit{additional forget gate} $\mathbf{g}_t^l$ that determines whether to accept or ignore the signals coming from the previous layer. The proposed Cell-aware Stacked LSTM (CAS-LSTM) architecture is defined as follows: \begin{align} \mathbf{i}_t^l &= \sigma(\mathbf{W}_i^l \mathbf{h}_t^{l-1} + \mathbf{U}_i^l \mathbf{h}_{t-1}^l + \mathbf{b}_i^l) \\ \mathbf{f}_t^l &= \sigma(\mathbf{W}_f^l \mathbf{h}_t^{l-1} + \mathbf{U}_f^l \mathbf{h}_{t-1}^l + \mathbf{b}_f^l) \\ \mathbf{g}_t^l &= \sigma(\mathbf{W}_g^l \mathbf{h}_t^{l-1} + \mathbf{U}_g^l \mathbf{h}_{t-1}^l + \mathbf{b}_g^l) \label{eq:new-forget-gate}\\ \tilde{\mathbf{c}}_t^l &= \tanh(\mathbf{W}_c^l \mathbf{h}_t^{l-1} + \mathbf{U}_c^l \mathbf{h}_{t-1}^l + \mathbf{b}_c^l) \\ \mathbf{o}_t^l &= \sigma(\mathbf{W}_o^l \mathbf{h}_t^{l-1} + \mathbf{U}_o^l \mathbf{h}_{t-1}^l + \mathbf{b}_o^l) \\ \mathbf{c}_t^l &= \mathbf{i}_t^l \odot {\tilde{\mathbf{c}}}_t^l + (\bm{1 - \lambda})\odot\mathbf{f}_t^l \odot \mathbf{c}_{t-1}^l + \bm{\lambda}\odot\mathbf{g}_t^l \odot \mathbf{c}_t^{l-1} \\ \mathbf{h}_t^l &= \mathbf{o}_t^l \odot \tanh(\mathbf{c}_t^l), \end{align} where $l > 1$ and $d_l=d_{l-1}$. $\bm\lambda$ can either be a vector of constants or parameters. When $l=1$, the equations defined in the previous subsection are used. Therefore, it can be said that each non-bottom layer of CAS-LSTM accepts two sets of hidden and cell states---one from the left context and the other from the below context. The left and the below context participate in computation with the equivalent procedure so that the information from lower layers can be efficiently propagated. Fig. \ref{fig:comparison} compares CAS-LSTM to the conventional stacked LSTM architecture, and Fig. \ref{fig:diagram} depicts the computation flow of the CAS-LSTM. We argue that considering $\mathbf{c}_t^{l-1}$ in computation is beneficial for the following reasons. First, contrary to $\mathbf{h}_t^{l-1}$, $\mathbf{c}_t^{l-1}$ contains information which is not filtered by $\mathbf{o}_t^{l-1}$. Thus a model that directly uses $\mathbf{c}_t^{l-1}$ does not rely solely on $\mathbf{o}_t^{l-1}$ for extracting information, due to the fact that it has access to the raw information $\mathbf{c}_t^{l-1}$, as in temporal connections. In other words, $\mathbf{o}_t^{l-1}$ no longer has to take all responsibility for selecting useful features for both horizontal and vertical transitions, and the burden of selecting information is shared with $\mathbf{g}_t^l$. \begin{figure} \centering \includegraphics[width=0.42\textwidth]{ca_path.pdf} \caption{ Visualization of paths between $\mathbf{c}_t^{l-1}$ and $\mathbf{c}_t^l$. In CAS-LSTM, the direct connection between $\mathbf{c}_t^{l-1}$ and $\mathbf{c}_t^l$ exists (denoted as red dashed lines). } \label{fig:ca_path} \end{figure} Another advantage of using the $\mathbf{c}_t^{l-1}$ lies in the fact that it directly connects $\mathbf{c}_t^{l-1}$ and $\mathbf{c}_t^l$. This direct connection could help and stabilize training, since the terminal error signals can be easily backpropagated to the model parameters by the shortened propagation path. Fig. \ref{fig:ca_path} illustrates paths between the two cell states. Regarding $\bm\lambda$, we find experimentally that there is little difference between having it be a constant and a trainable vector bounded in $(0, 1)$, and we practically find that setting $\lambda_i=0.5$ works well across multiple experiments. We also experimented with the architecture without $\bm\lambda$ i.e. two cell states are combined by unweighted summation similar to multidimensional RNNs \citep{graves2009offline}, and found that it leads to performance degradation and unstable convergence, likely due to mismatch in the range of cell state values between layers ($(-2, 2)$ for the first layer and $(-3, 3)$ for the others). Experimental results on various $\bm\lambda$ are presented in \S\ref{exp:variations}. \subsection{Sentence Encoders} For text classification tasks, a variable-length sentence should be represented as a fixed-length vector. We describe the sentence encoder architectures used in experiments in this subsection. First, we assume that a sequence of $T$ one-hot word vectors is given as input: $(\mathbf{w}_1, \cdots, \mathbf{w}_T)$, $\mathbf{w}_t \in \mathbb{R}^{|V|}$ where $V$ is the vocabulary set. The words are projected to corresponding word representations: $\mathbf{X}=(\mathbf{x}_1, \cdots, \mathbf{x}_T)$ where $\mathbf{x}_{t} = \mathbf{E}^\top \mathbf{w}_t \in \mathbb{R}^{d_0}$, $\mathbf{E}\in \mathbf{R}^{|V| \times d_0}$. Then $\mathbf{X}$ is fed to a $L$-layer CAS-LSTM model, resulting in the representations $\mathbf{H}=(\mathbf{h}_1^L, \cdots, \mathbf{h}_T^L)\in \mathbb{R}^{T\times d_L}$. The encoded sentence representation $\mathbf{s} \in \mathbb{R}^{d_L}$ is computed by max-pooling $\mathbf{H}$ over time as in the work of \citet{conneau2017infersent}. Similar to their results, from preliminary experiments we found that the max-pooling performs consistently better than the mean-pooling and the last-pooling. For better modeling of semantics, a bidirectional CAS-LSTM network may also be used. In the bidirectional case, the representations obtained by left-to-right reading $\mathbf{H}=(\mathbf{h}_1^L, \cdots, \mathbf{h}_T^L) \in \mathbb{R}^{T\times d_L}$ and those by right-to-left reading $\widehat{\mathbf{H}}=(\widehat{\mathbf{h}}_1^L, \cdots, \widehat{\mathbf{h}}_T^L) \in \mathbb{R}^{T \times d_L}$ are concatenated and max-pooled to yield the sentence representation $\mathbf{s} \in \mathbb{R}^{2d_L}$. We call this bidirectional architecture Bi-CAS-LSTM in experiments. To predict the final task-specific label, we apply a task-specific feature extraction function $\phi$ to the sentence representation(s) and feed the extracted features to a classifier network. For the classifier network, a multi-layer perceptron (MLP) with the ReLU activation followed by the linear projection and the softmax function is used: \begin{equation} P(\mathbf{y}|\mathbf{X}) = \text{softmax}(\mathbf{W}_c\text{MLP}(\phi(\cdot))), \end{equation} where $\mathbf{W}_c \in \mathbb{R}^{|L|\times d_h}$, $|L|$ is the number of label classes, and $d_h$ the dimension of the MLP output. \section{Experiments} \label{sec:experiments} We evaluate our method on three benchmark tasks on sentence encoding: natural language inference (NLI), paraphrase identification (PI), and sentiment classification. To further demonstrate the general applicability of our method on text generation, we also evaluate the proposed method on machine translation. In addition, we conduct analysis on gate values model variations for the understanding of the architecture. We refer readers to the supplemental material for detailed experimental settings. The code will be made public for reproduction. For the NLI and PI tasks, there exists architectures specializing in sentence pair classification. However in this work we confine our model to the architecture that encodes each sentence using a shared encoder without any inter-sentence interaction, in order to focus on the effectiveness of the architectures in extracting semantics. But note that the applicability of CAS-LSTM is not limited to sentence encoder--based approaches. \subsection{Natural Language Inference} \begin{table*}[t] \centering \begin{tabular}{l r r} \hline \bf{Model} & \bf{Acc. (\%)} & \bf {\# Params} \\ \hline 300D LSTM \citep{bowman2016spinn} & 80.6 & 3.0M \\ 300D TBCNN \citep{mou2016snli} & 82.1 & 3.5M \\ 300D SPINN-PI \citep{bowman2016spinn} & 83.2 & 3.7M \\ 600D BiLSTM with intra-attention \citep{liu2016learning} & 84.2 & 2.8M \\ 4096D BiLSTM with max-pooling \citep{conneau2017infersent} & 84.5 & 40M \\ 300D BiLSTM with gated pooling \citep{chen2017gated} & 85.5 & 12M \\ 300D Gumbel Tree-LSTM \citep{choi2018learning} & 85.6 & 2.9M \\ 600D Shortcut stacked BiLSTM \citep{nie2017shortcut} & 86.1 & 140M \\ 300D Reinforced self-attention network \citep{shen2018reinforced} & 86.3 & 3.1M \\ 600D BiLSTM with generalized pooling \citep{chen2018generalized} & 86.6 & 65M \\ \hline 300D 2-layer CAS-LSTM (ours) & 86.4 & 2.9M \\ 300D 2-layer Bi-CAS-LSTM (ours) & {86.8} & 6.8M \\ 300D 3-layer CAS-LSTM (ours) & 86.4 & 4.8M \\ 300D 3-layer Bi-CAS-LSTM (ours) & \bf{87.0} & 8.6M \\ \hline \end{tabular} \caption{Results of the models on the SNLI dataset.} \label{table:snli} \end{table*} \begin{table*}[t] \centering \begin{tabular}{l r r r} \hline \bf{Model} & \bf{In (\%)} & \bf{Cross (\%)} & \bf{\# Params} \\ \hline CBOW \citep{williams2018mnli} & 64.8 & 64.5 & - \\ BiLSTM \citep{williams2018mnli} & 66.9 & 66.9 & - \\ Shortcut stacked BiLSTM \citep{nie2017shortcut}$^\ast$ & \bf{74.6} & 73.6 & 140M \\ BiLSTM with gated pooling \citep{chen2017gated} & 73.5 & 73.6 & 12M \\ BiLSTM with generalized pooling \citep{chen2018generalized} & 73.8 & \bf{74.0} & 18M$^{\ast\ast}$ \\ \hline 2-layer CAS-LSTM (ours) & 74.0 & 73.3 & 2.9M \\ 2-layer Bi-CAS-LSTM (ours) & \bf{74.6} & 73.7 & 6.8M \\ 3-layer CAS-LSTM (ours) & 73.8 & 73.1 & 4.8M \\ 3-layer Bi-CAS-LSTM (ours) & 74.2 & 73.4 & 8.6M \\ \hline \end{tabular} \caption{Results of the models on the MultiNLI dataset. `In' and `Cross' represent accuracy calculated from the matched and mismatched test set respectively. $^\ast$: SNLI dataset is used as additional training data. $^{\ast\ast}$: computed from hyperparameters provided by the authors.} \label{table:mnli} \end{table*} For the evaluation of performance of the proposed method on the NLI task, SNLI \citep{bowman2015snli} and MultiNLI \citep{williams2018mnli} datasets are used. The objective of both datasets is to predict the relationship between a premise and a hypothesis sentence: \textit{entailment}, \textit{contradiction}, and \textit{neutral}. SNLI and MultiNLI datasets are composed of about 570k and 430k premise-hypothesis pairs respectively. GloVe pretrained word embeddings\footnote{\url{https://nlp.stanford.edu/projects/glove/}} \citep{pennington2014glove} are used and remain fixed during training. The dimension of encoder states ($d_l$) is set to 300 and a 1024D MLP with one or two hidden layers is used. We apply dropout \citep{srivastava2014dropout} to the word embeddings and the MLP layers. The features used as input to the MLP classifier are extracted by the following equation: \begin{equation} \label{eq:nli-matching} \phi(\mathbf{s}_1, \mathbf{s}_2) = \mathbf{s}_1 \oplus \mathbf{s}_2 \oplus |\mathbf{s}_1 - \mathbf{s}_2| \oplus (\mathbf{s}_1 \odot \mathbf{s}_2), \end{equation} where $\oplus$ is the vector concatenation operator. Table \ref{table:snli} and \ref{table:mnli} contain results of the models on SNLI and MultiNLI datasets. Along with other state-of-the-art models, the tables include several stacked LSTM--based models to facilitate comparison of our work with prior related work. \citet{liu2016learning,chen2017gated,chen2018generalized} adopt advanced pooling algorithms motivated by the attention mechanism to obtain a fixed-length sentence vector. \citet{nie2017shortcut} use the concatenation of all outputs from previous layers as input to the next layer. In SNLI, our best model achieves the accuracy of 87.0\%, which is the new state-of-the-art among the sentence encoder--based models, with relatively fewer parameters. Similarly in MultiNLI, our models match the accuracy of state-of-the-art models in both in-domain (matched) and cross-domain (mismatched) test sets. Note that only the GloVe word vectors are used as word representations, as opposed to some models that introduce character-level features. It is also notable that our proposed architecture does not restrict the selection of pooling method; the performance could further be improved by replacing max-pooling with other advanced algorithms e.g. intra-sentence attention \citep{liu2016learning} and generalized pooling \citep{chen2018generalized}. \subsection{Paraphrase Identification} \begin{table}[t] \centering \begin{tabular}{l r} \hline \bf{Model} & \bf{Acc. (\%)} \\ \hline CNN \citep{wang2017bilateral} & 79.6 \\ LSTM \citep{wang2017bilateral} & 82.6 \\ Multi-Perspective LSTM \citep{wang2017bilateral} & 83.2 \\ LSTM + ElBiS \citep{choi2018elbis} & 87.3 \\ REGMAPR (BASE+REG) \citep{brahma2018regmapr} & 88.0 \\ \hline CAS-LSTM (ours) & 88.4 \\ Bi-CAS-LSTM (ours) & \bf{88.6} \\ \hline \end{tabular} \caption{Results of the models on the Quora Question Pairs dataset.} \label{table:quora} \end{table} We use Quora Question Pairs dataset \citep{wang2017bilateral} in evaluating the performance of our method on the PI task. The dataset consists of over 400k question pairs, and each pair is annotated with whether the two sentences are paraphrase of each other or not. Similarly to the NLI experiments, GloVe pretrained vectors, 300D encoders, and 1024D MLP are used. The number of CAS-LSTM layers is fixed to 2 in PI experiments. Two sentence vectors are aggregated using the following equation and fed as input to the classifier. \begin{equation} \label{eq:pi-matching} \phi(\mathbf{s}_1, \mathbf{s}_2) = |\mathbf{s}_1 - \mathbf{s}_2| \oplus (\mathbf{s}_1 \odot \mathbf{s}_2) \end{equation} The results on the Quora Question Pairs dataset are summarized in Table \ref{table:quora}. Again we can see that our models outperform other models, especially compared to conventional LSTM--based models. Also note that Multi-Perspective LSTM \citep{wang2017bilateral}, LSTM + ElBiS \citep{choi2018elbis}, and REGMAPR (BASE+REG) \citep{brahma2018regmapr} in Table \ref{table:quora} are approaches that focus on designing a more sophisticated function for aggregating two sentence vectors, and their aggregation functions could be also applied to our work for further improvement. \subsection{Sentiment Classification} \begin{table*}[t] \centering \begin{tabular}{l r r} \hline \bf{Model} & \bf{SST-2 (\%)} & \bf{SST-5 (\%)} \\ \hline Recursive Neural Tensor Network \citep{socher2013recursive} & 85.4 & 45.7 \\ 2-layer LSTM \citep{tai2015treelstm} & 86.3 & 46.0 \\ 2-layer BiLSTM \citep{tai2015treelstm} & 87.2 & 48.5 \\ Constituency Tree-LSTM \citep{tai2015treelstm} & 88.0 & 51.0 \\ Constituency Tree-LSTM with recurrent dropout \citep{looks2017deep} & 89.4 & 52.3 \\ byte mLSTM \citep{radford2017learning}$^\ast$ & \underline{91.8} & 52.9 \\ Gumbel Tree-LSTM \citep{choi2018learning} & 90.7 & \bf{53.7} \\ BCN + Char + ELMo \citep{peters2018elmo}$^\ast$ & - & \underline{54.7} \\ \hline 2-layer CAS-LSTM (ours) & 91.1 & 53.0 \\ 2-layer Bi-CAS-LSTM (ours) & \bf{91.3} & 53.6 \\ \hline \end{tabular} \caption{Results of the models on the SST dataset. $^\ast$: models pretrained on large external corpora are used.} \label{table:sst} \end{table*} In evaluating sentiment classification performance, the Stanford Sentiment Treebank (SST) \citep{socher2013recursive} is used. It consists of about 12,000 binary-parsed sentences where constituents (phrases) of each parse tree are annotated with a sentiment label (\textit{very positive}, \textit{positive}, \textit{neutral}, \textit{negative}, \textit{very negative}). Following the convention of prior work, all phrases and their labels are used in training but only the sentence-level data are used in evaluation. In evaluation we consider two settings, namely SST-2 and SST-5, the two differing only in their level of granularity with regard to labels. In SST-2, data samples annotated with `neutral' are ignored from training and evaluation. The two positive labels (very positive, positive) are considered as the same label, and similarly for the two negative labels. As a result 98,794/872/1,821 data samples are used in training/validation/test, and the task is considered as a binary classification problem. In SST-5, all 318,582/1,101/2,210 data samples are used and the task is a 5-class classification problem. Since the task is a single-sentence classification problem, we use the sentence representation itself as input to the classifier. We use 300D GloVe vectors, 2-layer 150D or 300D encoders, and a 300D MLP classifier for the models, however unlike previous experiments we tune the word embeddings during training. The results on SST are listed in Table \ref{table:sst}. Our models clearly outperform plain LSTM- and BiLSTM-based models, and are competitive to other state-of-the-art models, without utilizing parse tree information. \subsection{Machine Translation} \begin{table}[tb] \centering \begin{tabular}{l l} \hline \bf{Model} & \bf{BLEU} \\ \hline 256D LSTM & 28.1 $\pm$ 0.22 \\ 256D CAS-LSTM & 28.8 $\pm$ 0.04$^\ast$ \\ 247D CAS-LSTM & 28.7 $\pm$ 0.07$^\ast$ \\ \hline \end{tabular} \caption{ Results of the models on the IWSLT 2014 de-en dataset. $^\ast$: $p < 0.0005$ (one-tailed paired t-test). } \label{table:mt} \end{table} We use the IWSLT 2014 machine evaluation campaign dataset \citep{cettolo2014report} in machine translation experiments. We used the fairseq library\footnote{\url{https://github.com/pytorch/fairseq}} \citep{gehring2017fairseq} for experiments. Moses tokenizer\footnote{\url{https://github.com/moses-smt/mosesdecoder/blob/master/scripts/tokenizer/tokenizer.perl}} is used for word tokenization and the byte pair encoding \citep{sennrich2016bpe} is applied to confine the size of the vocabulary set up to 10,000. Similar to \citet{wiseman2016nmt}, a 2-layer 256D sequence-to-sequence LSTM model with the attentional decoder is used as baseline, and we replace the encoder and the decoder network with the proposed architecture for the evaluation of performance improvement. For decoding, beam search with $B=10$ is used. For fair comparison, we tune hyperparameters for all models based on the performance on the validation dataset and train the same model for five times with different random seeds. Also, to cancel out the increased number of parameters, we experiment with the 247D CAS-LSTM model which has the roughly same number of parameters as the baseline model (8.2M). From Table \ref{table:mt}, we can see that the CAS-LSTM models bring significant performance gains over the baseline model. \subsection{Forget Gate Analysis} \begin{figure}[t] \centering \subfigure{% \label{fig:gate-a} \includegraphics[width=0.3\textwidth]{g2.pdf} } \quad \subfigure{% \label{fig:gate-b} \includegraphics[width=0.3\textwidth]{g3.pdf} } \par\bigskip \subfigure{% \label{fig:gate-c} \includegraphics[width=0.3\textwidth]{g2_range.pdf} } \quad \subfigure{% \label{fig:gate-d} \includegraphics[width=0.3\textwidth]{g3_range.pdf} } \par\bigskip \subfigure{% \label{fig:gate-e} \includegraphics[width=0.3\textwidth]{g2o1.pdf} } \quad \subfigure{% \label{fig:gate-f} \includegraphics[width=0.3\textwidth]{g3o2.pdf} } \caption{ (\textit{a}) $g^2_i$, (\textit{b}) $g^3_i$, (\textit{c}) $R(\mathbf{g}^2_\cdot)$, (\textit{d}) $R(\mathbf{g}^3_\cdot)$, (\textit{e}) $\vert g^2_i - o^1_i \vert$, (\textit{f}) $\vert g^3_i - o^2_i \vert$. (\textit{a}), (\textit{b}): Histograms of vertical forget gate values. (\textit{c}), (\textit{d}): Histograms of the ranges of vertical forget gate per time step. (\textit{e}), (\textit{f}): Histograms of the absolute difference between the previous output gate and the current vertical forget gate values. } \label{fig:gate-analysis} \end{figure} To inspect the effect of the additional forget gate, we investigate how the values of vertical forget gates are distributed. We sample 1,000 random sentences from the development set of the SNLI dataset, and use the 3-layer CAS-LSTM model trained on the SNLI dataset to compute gate values. If all values from a vertical forget gate $\mathbf{g}_t^l$ were to be 0, this would mean that the introduction of the additional forget gate is meaningless and the model would reduce to a plain stacked LSTM. On the contrary if all values were 1, meaning that the vertical forget gates were always \textit{open}, it would be impossible to say that the information is modulated effectively. Fig. \ref{fig:gate-a} and \ref{fig:gate-b} represent histograms of the vertical forget gate values from the second and the third layer. From the figures we can validate that the trained model does not fall into the degenerate case where vertical forget gates are ignored. Also the figures show that the values are right-skewed, which we conjecture to be a result of focusing more on a strong interaction between adjacent layers. To further verify that the gate values are diverse enough within each time step, we compute the distribution of the range of values per time step, $R(\mathbf{g}_t^l)=\max_i{g_{t,i}^l} - \min_i{{g}_{t,i}^l}$, where $\mathbf{g}_t^l=[g_{t,1}^l, \cdots, g_{t,d_l}^l]^\top$. We plot the histograms in Fig. \ref{fig:gate-c} and \ref{fig:gate-d}. From the figures we see that the vertical forget gate controls the amount of information flow effectively, making diverse decisions of retaining or discarding signals across dimensions. Finally, to investigate the argument presented in \S\ref{sec:caslstm} that the additional forget gate helps the previous output gate with reducing the burden of extracting all needed information, we inspect the distribution of the values from $\vert \mathbf{g}_t^l - \mathbf{o}_t^{l-1} \vert$. This distribution indicates how differently the vertical forget gate and the previous output gate select information from $\mathbf{c}_t^{l-1}$. From Fig. \ref{fig:gate-e} and \ref{fig:gate-f} we can see that the two gates make fairly different decisions, from which we demonstrate that the direct path between $\mathbf{c}_t^{l-1}$ and $\mathbf{c}_t^l$ enables a model to utilize signals overlooked by $\mathbf{o}_t^{l-1}$. \subsection{Model Variations} \label{exp:variations} In this subsection, we see the influence of each component of a model on performance by removing or replacing its components. the SNLI dataset is used for experiments, and the best performing configuration is used as a baseline for modifications. We consider the following variants: (\textit{i}) models with different $\bm\lambda$, (\textit{ii}) models without $\bm\lambda$, and (\textit{iii}) models that integrate lower contexts via peephole connections. Variant (\textit{iii}) calculates and applies the forget gate $\mathbf{g}_t^l$ which takes charge of integrating lower contexts via the equations below, following the work of \citet{zhang2016highway}: \begin{align} \begin{split} \mathbf{g}_t^l &= \sigma(\mathbf{W}_g^l \mathbf{h}_t^{l-1} + \mathbf{p}_{g_1}^l \odot \mathbf{c}_{t-1}^l + \mathbf{p}_{g_2}^l \odot \mathbf{c}_{t}^{l-1} + \mathbf{b}_g^l) \label{eq:peephole1} \end{split} \\ \mathbf{c}_t^l &= \mathbf{i}_t^l \odot \tilde{\mathbf{c}}_t^l + \mathbf{f}_t^l \odot \mathbf{c}_{t-1}^l + \mathbf{g}_t^l \odot \mathbf{c}_t^{l-1}, \label{eq:peephole2} \end{align} where $\mathbf{p}_\cdot^l \in \mathbb{R}^{d_l}$ represent peephole weight vectors that take cell states into account. We can see that the computation formulae of $\mathbf{f}_t^l$ and $\mathbf{g}_t^l$ are not consistent, in that $\mathbf{h}_{t-1}^l$ does not participate in computing $\mathbf{g}_{t-1}^l$, and that the left and the below context are reflected in $\mathbf{g}_{t-1}^l$ only via element-wise multiplications which do not consider the interaction among dimensions. By contrast, ours uses the analogous formulae in calculating $\mathbf{f}_t^l$ and $\mathbf{g}_t^l$, considers $\mathbf{h}_{t-1}^{l}$ in calculating $\mathbf{g}_t^l$, and introduces the scaling factor $\bm{\lambda}$. Table \ref{table:variations} summarizes the results of model variants. From the results of \textit{baseline} and \textit{(i)}, we validate that the selection of $\bm\lambda$ does not significantly affect performance but introducing $\bm\lambda$ is beneficial (\textit{baseline vs. (ii)}) possibly due to its effect on normalizing information from multiple sources, as mentioned in \S\ref{sec:caslstm}. Also, from the comparison between \textit{baseline} and \textit{(iii)}, we show that the proposed way of combining the left and the lower contexts leads to better modeling of sentence representations than that of \citet{zhang2016highway}. \begin{table}[t] \centering \begin{tabular}{l r r} \hline \bf{Model} & \bf{Acc. (\%)} & \bf{$\Delta$} \\ \hline Bi-CAS-LSTM (\textit{baseline}) & 87.0 & \\ \quad \textit{(i) Diverse $\bm\lambda$} & & \\ \qquad \textit{(a) $\lambda_i=0.25$} & 86.8 & -0.2 \\ \qquad \textit{(b) $\lambda_i=0.75$} & 86.8 & -0.2 \\ \qquad \textit{(c) Trainable $\bm\lambda$} & 86.9 & -0.1 \\ \quad \textit{(ii) No $\bm\lambda$} & 86.6 & -0.4 \\ \quad \textit{(iii) Integration through peepholes} & 86.5 & -0.5 \\ \hline \end{tabular} \caption{Results of model variants.} \label{table:variations} \end{table} \section{Conclusion} \label{sec:conclusion} In this paper, we proposed a method of stacking multiple LSTM layers for modeling sentences, dubbed CAS-LSTM. It uses not only hidden states but also cell states from the previous layer, for the purpose of controlling the vertical information flow in a more elaborate way. We evaluated the proposed method on various benchmark tasks: natural language inference, paraphrase identification, and sentiment classification. Our models outperformed plain LSTM-based models in all experiments and were competitive other state-of-the-art models. The proposed architecture can replace any stacked LSTM only under one weak restriction---the size of states should be identical across all layers. For future work we plan to apply the CAS-LSTM architecture beyond sentence modeling tasks. Various problems such as sequence labeling and language modeling might benefit from sophisticated modulation on context integration. Aggregating diverse contexts from sequential data, e.g. those from forward and backward reading of text, could also be an intriguing research direction. \section*{Acknowledgments} This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea Government (MSIT) (NRF2016M3C4A7952587).

1d89691e8178fdde06141bebe30bf7fc291b0f60

\section{Introduction} An ever increasing interest in the field of automatic summarization has led to a plethora of extractive techniques, with a lack of clarity about which is the best technique. Unclear implementation details and variation in evaluation setups only make the problem worse. However, there is no doubt that none of these techniques would always work. Not only will there be a difference in performance across different datasets, there is also a good amount of variation across documents in the same dataset. In several cases this difference is also attributed to the fact that different ROUGE setups are used for evaluation, which can result substantial variation in the scores. \cite{hong2014repository} propose using a fixed set of parameters for ROUGE and report comparable results for several summarization algorithms. Even with this normalization, the system performance still varies a lot, and there is a possibility of exploiting this variation to generate better performing systems. To give an example, we show a simple comparison of two extractive summarization systems from those used by \cite{hong2014repository} in their experiments. We pick two extreme systems, in terms of performance, from the those reported in the work. The FreqSum system\cite{nenkova2006compositional}, which has the weakest performance, was compared to the DPP system\cite{kulesza2012determinantal} which was the best performing system amongst those compared. On DUC 2004 dataset, FreqSum performed better on more than 10$\%$ of the document clusters. There are documents for which a system which is overall very weak, outperforms the system that has a very good performance on an average. Going a step further, an oracle of just the five baseline systems, outperforms DPP in a little over fifty percent of the document clusters. The argument is clear: ensemble of several systems can definitely improve the performance as compared to individual systems. The ideal way of forming an ensemble summary would be to select the relevant information from each candidate while discarding the rest.\\ In this work we propose a new method for estimating the authority of a particular system for a given document and at the same time also estimating the importance of each sentence within the summary generated by that system. The ensemble summary is then a function of the authority of each candidate system as well as the relative importance of each sentence in the candidate summaries. The fact, that informative or \emph{summary worthy} sentences in a document cluster are much less in number compared to the non-informative ones, forms the basis of our hypothesis. We argue that since this content is much less, any substantial overlap between two summaries will likely be due to the \emph{important} content rather than the redundant one. Simply because there is less content to choose from when it comes to the important sentences two good summaries will have a good amount of overlap in content. At the same time it is highly unlikely that two summaries will also chose the same not-important content. Keeping this argument in mind we associate higher similarity in content between two summaries with the summaries having more informative content. \\ We use graph based ranking that takes into account the similarity of a candidate summary with other candidates to generate its local (or document specific) ranking. We also determine the overall global ranking of a system from its ROUGE score on a development dataset. In the same way \emph{informativeness} of a sentence is linked to its overlap with sentences of other summaries. The HybridRank model proposed here, combines these three factors to generate a new aggregate ranking of sentences. \section{Related Work} In contrast to the amount of attention automatic summarization, and especially the extractive techniques, has achieved from researchers, aggregation techniques have been explored little. This is counter intuitive given several studies which show that even in cases where two systems achieve a comparable ROUGE score, the actual content can be quite different \cite{hong2014repository}. Existing aggregation techniques can be broadly classified into two categories: \emph{rank aggregation} and \emph{summary aggregation}. These techniques are used post-summarization, i.e. each candidate system is first used without any modification. The output, whether ranked lists or summaries, are then used for aggregation. The former method solely relies on candidate summaries, without any information or assumption about the original sentence rankings. In contrast the latter combines existing ranked lists to generate a new aggregate ranking. Apart from these two, there is another type of aggregation which combines various aspects of candidates and incorporates them into the algorithm itself.\\ \cite{pei2012supervised} and \cite{hongsystem} are two instances of the summary aggregation techniques. \cite{pei2012supervised} use SVM-Rank to learn the optimum ranking of sentences from candidate systems. Each sentence is labeled as $-1,0,1$ depending on its \emph{summary worthiness} and then it is used to learn pairwise ranking for each sentence. As opposed to this, \cite{hongsystem} attempt to generate rankings of entire summaries to find out the combination of sentences that maximizes ROUGE-1 or ROUGE-2. They use summaries from four different systems to begin with. Next they combine these summaries and list out all possible candidate summaries by selecting a fixed number of sentences from the combined set. Several word and summary level features are then used to train a SVM-Rank algorithm that can learn to rank candidate summaries to maximize ROUGE scores. \\ Excluding the common techniques like \emph{Round Robin}, \emph{Borda Count} or \emph{Reciprocal Rank} only notable attempt at using rank aggregation was made by \cite{wang2012weighted}. The \emph{weighted consensus summarization} system proposed by them treats rank aggregation as a optimization problem. This approach also introduced the concept of consensus between candidate summaries. They create a weighted combination of ranked lists, under the constraint that the aggregate ranking be as close to original rankings as possible. Unlike the approach proposed in this paper, \cite{wang2012weighted} do not differentiate candidates based on their \emph{trustworthiness}, and instead try to maintain as much information from each candidate as possible. Apart from not taking into account the content of candidate summaries, there are two major limitations with this approach. One, it uses $L_1$ norm for computing similarity (or distance) between candidate rankings, which can be a sub-optimal choice when compared to traditional metrics like \emph{Kendall's Tau}. The other major problem is, this technique tries to optimize rankings over all sentences in the document cluster. So even if the systems agree on ranking of top-k sentence, which actually go into the summary, the aggregation will try to optimize over the entire rankings. This is not only unnecessary, but can also affect the performance adversely.\\ Neither \emph{summary aggregation} nor \emph{rank aggregation} take into consideration the original summarization algorithms or in any way modify them. There have been a few attempts at modifying summarization algorithms to incorporate features from several candidates into a single algorithm. But such attempts are limited by the non-triviality of being able to meaningfully combine unique aspects of the candidates. One popular approach of this kind is combining several sentence similarity scores to generate an aggregate similarity which can then be used by any of the existing algorithms. The \emph{MultSum} technique proposed by \cite{mogrenextractive} builds upon the submodular optimization technique\cite{lin2012learning}. They replace the cosine similarity used in the original approach with multiplicative combination of several sentence similarity measures. This can be extended to several techniques which rely on a sentence similarity metric. But in general it is difficult to define an \emph{aggregate} for other components of a summarization algortihm. As an alternate, \cite{mehta2018effective} propose combining several aspects of candidate systems like \emph{sentence similarity}, \emph{ranking algorithm} and \emph{text representation scheme} in a post-summarization setup. This approach falls within the rank aggregation technique, except that the candidate systems are created by varying one of these three components of the original systems. The \emph{LexRank} algorithm can be used with Kullback Leibler Divergence or word overlap, instead of cosine similarity. Such variations are then combined using existing rank aggregation techniques. \section{Proposed Approach} In this paper we propose a new summary aggregation technique which takes into account the content of each candidate systems, rather than only ranked lists as done in \cite{wang2012weighted}. In a multi document summarization setup, especially in case of newswire, it is quite common to have duplicate or near duplicate content getting repeated across multiple documents. To put things into perspective, DUC 2003 dataset has, on an average, 34 sentences per cluster which have an exact match within the document cluster. More than 50 sentences have a 80\% match with another sentence. Most existing summarization techniques do not handle this redundancy explicitly, neither do most existing aggregation techniques. This has a huge impact on a class of aggregation techniques which rely only on ranked lists aggregation without taking into account the actual content of individual summaries. For instance, consider the example shown below where $S_1$ and $S_2$ are individual sentence rankings and $S_A$ is the aggregate ranking. This would be fine if each sentence is different and equally important. But consider a case where $sim(s_{1}, s_4) = 1$. The fact that $s_1$ and $s_5$ are repeated across documents makes them more important. But the rank aggregation techniques fail to take into account their actual content and treat these separately, which may result in lowering of their aggregate scores. $S_A^*$ indicates the ideal rank aggregate. \begin{figure}[H] \centering \includegraphics[scale=0.6]{aclproblem.png} \caption{Issue with rank aggregation} \end{figure} Instead, in the proposed approach we take into account content of the summaries being aggregated to assign weights to candidate systems. The proposed system takes into account the overlap in content between summaries of the candidate systems to compute their reliability, and in turn use it to assign relative importance to individual sentences in those summaries. \\ The proposed method estimates importance of a given sentence in a particular candidate summary, and then uses this information to generate a meta-summary. In contrast to the solution proposed by \cite{hongsystem}, where they use word and summary level features to estimate the relative importance of each candidate, we propose using similarity between candidates as a measure for their importance. We argue that given two good summaries there is bound to be a good amount of overlap in their content, simply because the relevant content in a document cluster is much less compared to non-relevant content. Arguing the other way round, if two summaries have a good amount of overlap, it is likely because both have more \emph{informative} content. Due to a large number of \emph{non-informative} sentences, selecting the same set of \emph{non-informative} sentences in two distinct summaries is much less likely. The only possible exception where two bad summaries might have higher overlap, will be when both have very similar sentence rankings. As the number of candidate systems increase, this will automatically be taken care of. We start with a simple method to estimate informativeness of sentences in candidate summaries. \subsection{SentRank} This system takes into account similarity of each sentence in a candidate summary, with that of other candidate summaries, and uses it to assign a relative importance to the sentence. The argument presented above, in favor of using similarity as a measure for reliability of a summary can easily be extended at sentence level. A sentence that is \emph{summary worthy} will share more information with another \emph{summary worthy sentences}. For $i^{th}$ sentence in the $j^{th}$ candidate summary($s_{ij}$), we find the best matching sentence in the remaining candidate summaries. The score of that sentence can then be computed as shown in equation \ref{eq1} below. Score of the sentence is the sum of its similarity with the best matching sentences from remaining candidates. Here $j$,$k$ are the candidate systems. $i$ and $l$ are the sentences in candidate $J$ and $K$ respectively. \begin{eqnarray} \boldsymbol{R}(s_{ij}) = \sum_{k, k\neq j} \max_{l}(Sim(s_{ij}, s_{lk})) \label{eq1} \end{eqnarray} Each sentence in the candidate summary is then ranked according to their score $\boldsymbol{R}$ and top k sentences are selected in the summary. We experimented with n-gram overlap, cosine and KL Divergence for computing the similarity between sentences and empirically select cosine similarity, which is also used in the subsequent systems. \subsection{GlobalRank} One limitation of the SentRank approach proposed above is that does not take into account the reliability of candidate systems into account and treats each candidate equally. A sentence that comes from a well performing candidate system is more likely to be \emph{informative} compared to a sentence from a poor summary. The proposed \emph{GlobalRank} system does exactly that. It builds over the SentRank system by incorporating a candidates \emph{global reputation} score into the sentence ranking scheme. The new scoring mechanism is shown in equation \ref{eq2} below. $G(k)$ refers to the global reputation of candidate system $k$. \begin{eqnarray} \boldsymbol{R}(s_{ij}) = \sum_{k, k\neq j} G(k)*\max_{l}(Sim(s_{ij}, s_{lk})) \label{eq2} \end{eqnarray} $G(k)$ is estimated using the average ROUGE-1 recall of each candidate systems as shown in equation 4 and 5 below. $R1_k$ is the rouge-1 recall of the $k^{th}$ candidate. $R1^{'}$ is the normalized version of $R1$. Here we do not subtract mean, to avoid negative values in scoring, and instead subtract the minimum of $R1^{'}$. Additionally we scale it using a scaling factor $a$, which is dependent on the total number of candidate systems. We empirically set $a$ to $0.1$. We used the results on DUC2002 dataset for estimating the ROUGE-1 recall, and in turn the GlobalRank of a candidate. \begin{eqnarray} G(k) = aR1'(k) \\ R1^{'}_{k} = \frac{R1_k}{\sigma(R1_k)} - \min_{k}\Big[\frac{R1_k}{\sigma(R1_k)}\Big] \label{eq3} \end{eqnarray} As compared to SentRank, which can be overwhelmed by too many poor performing systems, GlobalRank provides a smoothing effect, by giving more importance to the systems that are known to perform well generally. \subsection{LocalRank} One major limitation of the existing aggregation systems, which we highlighted in section 2, is their inability to predict which candidate system will perform better for a given document cluster. Neither of the systems suggested above, \emph{SentRank} and \emph{GlobalRank}, address this problem. The next system, \emph{LocalRank} tries to mitigate this problem. We do not rely on any lexical or corpus specific features, simply because the training data is not sufficient to estimate these features reliably. Instead we continue on our line of argument, using the similarity between summaries as a measure of reliability. For a give document cluster, we estimate reliability of a candidate $k$ from the content it shares with other candidates, and also the reliability of those candidates. We first create a graph with the nodes as the candidate summaries and edge as the similarity between nodes. Each candidate starts with the same reputation score or LocalRank ($L$). The local rank is then updated iteratively using the pagerank algorithm \cite{page1999pagerank}. The Local rank for a given node is estimated as shown in equation \ref{eq4} below: \begin{equation} L(k) = \sum_{j}L(j)*Sim(S_j, S_k) \label{eq4} \end{equation} $L(k)$ indicates local rank of $k^{th}$ candidate, $S_k$ indicate summary generated by the $k^{th}$ candidate. We use cosine similarity as the similarity score. The overall sentence scores are then computed just like in the GlobalRank algorithm. \begin{eqnarray} \boldsymbol{R}(s_{ij}) = \sum_{k, k\neq j} L(k)*\max_{l}(Sim(s_{ij}, s_{lk})) \label{eq5} \end{eqnarray} \subsection{HybridRank} While LocalRank is useful for estimating how well a given candidate might perform for a given document cluster, it does not make use of the actual system performance. HybridRank overcomes that limitation. As the name suggests, HybridRank combines strengths of both GlobalRank as well as LocalRank by taking a weighted combination of both. The HybridRank is defined in the equation \ref{eq6} below \begin{equation} H(k) = \alpha L(k) + (1-\alpha)G(k) \label{eq6} \end{equation} Here the value of $\alpha$ determines the balance between Local and GlobalRank. Higher value of Alpha gives more importance to the estimate of how good a system will perform on a particular cluster, while ignoring the overall aggregate performance of candidate. $\alpha = 0$ leads to the original GlobalRank, without any local information. We empirically set the value of $\alpha$ to 0.3. Once the systems are ranked, the sentence rankings are computed in the same manner as LocalRank or GlobalRank (equation \ref{eq2} and \ref{eq5}). \section{Experimental Setup and Results} We report the experimental results on the DUC 2003 and DUC 2004 datasets. We report standard ROUGE scores\cite{lin2004rouge} that are well accepted across the community, ROUGE-1, ROUGE-2 and ROUGE-4 recall. To be consistent and in order to make our work reproducible, we use the same set of ROUGE parameters as that used by \cite{hong2014repository}\footnote{ROUGE-1.5.5.pl -n 4 -m -a -x -l 100 -c 95 -r 1000 -f A -p 0.5 -t 0}. We use eleven candidate systems which are a mix of several state of art extractive techniques and other well known baseline systems. The complete list is shown below, with a brief description of each system. For the DUC 2004 dataset, \cite{hong2014repository} provide summaries for all these systems. We directly use these pre-generated summaries provided for that particular year, to provide a fair comparison. We generated results for the DUC 2003 dataset ourselves. We do post-processing on the generated summaries, by dropping sentences that have a cosine similarity of more than 0.5 with the already selected sentences. Apart from that no other pre/post-processing was done. We use the following systems as candidate systems for our experiments: \begin{itemize} \item[] \textbf{LexRank} This method proposed by \cite{erkan2004lexrank} treats each document as a undirected graph and each sentence in the document constitutes a node. The edges represent cosine similarity between nodes. The importance of each node is then iteratively determined by the number of other nodes to which it is connected and the importance of those nodes. \item[] \textbf{FreqSum} A simple approach\cite{nenkova2006compositional} that ranks each sentence based on the frequency of its constituent words. Higher the frequency more the informativeness of the word. \item[] \textbf{TsSum} This method defines \emph{topic signatures} as the words which are more frequent in a given document compared to a background corpus\cite{lin2000automated}. Sentenced with more \emph{topic signatures} are considered more informative \item[] \textbf{Greedy-KL} This method follows a greedy algorithm\cite{haghighi2009exploring} to minimize the KL Divergence between the original document and resultant summary. Sentences are sequentially added to the summary so as to minimize the KL-Divergence between word distributions of the set of sentences selected so far and the overall document \item[] \textbf{CLASSY04} Judged best among the submissions at DUC 2004\cite{conroy2004left}, uses a hidden markov model with topic signatures as the features. It links the usefulness of a sentence to that of its neighboring sentences. \item[] \textbf{CLASSY11} This method builds over the CLASSY04 technique, and uses topic signatures as features while estimating the probability that a bigram will occur in human generated summary. It employs non negative matrix factorization to select a subset of non-redundant sentences with highest scores. \item[] \textbf{Submodular} \cite{lin2012learning} treat summarization as a submodular maximization problem. It incrementally computes the \emph{informativeness} of a summary and also provides a confidence score as to how close the approximation is to globally optimum summary. \item[] \textbf{DPP} Detrimental point processing\cite{kulesza2012determinantal} is the best performing state-of-art system amongst all the candidate systems. DPP scores each sentence individually, while at the same time trying to maintain a global diversity to reduce redundancy in the content selected. \item[] \textbf{RegSum} uses diverse features like parts of speech tags, name entity tags, locations and categories for supervised prediction of word importance\cite{hong2014improving}. The sentence with most number of \emph{important} words are then included in the summary. \item[] \textbf{OCCAMS\_V} The system by \cite{davis2012occams}, employs LSA to estimate word importance and then use the budgeted maximal coverageand the knapsack problem to generate sentence rankings. \item[] \textbf{ICSISumm} treats summarization as a global linear optimization problem\cite{gillick2009icsi}, to find globally best summary instead of selecting sentences greedily. The final summary includes most important concepts in the documents. \end{itemize} As benchmark ensemble techniques, we use two existing techniques: \emph{Borda count} and \emph{Weighted consensus summarization}\cite{wang2012weighted} described in section 2. For the GlobalRank system, we used DUC 2002 dataset as a development dataset to estimate the overall performance of candidate systems. We ranked the systems based on ROUGE-1 recall scores for this purpose. The results are shown in table 1 below: \vspace{3mm} \begin{table*}[t] \centering \begin{tabular}{|c|c|c|c| c |c | c|} \hline & \multicolumn{3}{c|}{DUC2003} & \multicolumn{3}{c|}{DUC2004} \\ \hline System & R-1 & R-2 & R-4 & R-1 & R-2 & R-3 \\ \hline LexRank & 0.3572 & 0.0742 & 0.0079 & 0.3595 & 0.0747 & 0.0082 \\ FreqSum & 0.3542 & 0.0815 & 0.0101 & 0.3530 & 0.0811 & 0.0099 \\ TsSum & 0.3589 & 0.0863 & 0.0103 & 0.3588 & 0.0815 & 0.0103 \\ Greedy-KL & 0.3692 & 0.0880 & 0.0129 & 0.3780 & 0.0853 & 0.0126 \\ CLASSY04 & 0.3744 & 0.0902 & 0.0148 & 0.3762 & 0.0895 & 0.0150 \\ CLASSY11 & 0.3730 & 0.0925 & 0.0142 & 0.3722 & 0.0920 & 0.0148 \\ Submodular & 0.3888 & 0.0930 & 0.0141 & 0.3918 & 0.0935 & 0.0139 \\ DPP & 0.3992 & 0.0958 & 0.0159 & 0.3979 & 0.0962 & 0.0157 \\ RegSum & 0.3840 & 0.0980 & 0.0165 & 0.3857 & 0.0975 & 0.0160 \\ OCCAMS\_V & 0.3852 & 0.0976 & 0.0142 & 0.3850 & 0.0976 & 0.0133 \\ ICSISumm & 0.3855 & 0.0977 & 0.0185 & 0.3840 & 0.0978 & 0.0173 \\ \hline Borda Count & 0.3700 & 0.0738 & 0.0115 & 0.3772 & 0.0734 & 0.0110 \\ WCS & 0.3815 & 0.0907 & 0.0120 & 0.3800 & 0.0923 & 0.0125 \\ SentRank & 0.3880 & 0.1010 & 0.0163 & 0.3870 & 0.1008 & 0.0159 \\ GlobalRank & 0.3562 & 0.1045 & 0.0185 & 0.3955 & 0.1039 & \textbf{0.0191}$^\dagger$ \\ LocalRank & 0.3992 & 0.1058 & 0.0192 & 0.3998 & 0.1050 & 0.0187 \\ HybridRank & \textbf{0.4082}$^\dagger$ & \textbf{0.1102}$^\dagger$ & \textbf{0.0195}$^\dagger$ & \textbf{0.4127}$^\dagger$ & \textbf{0.1098}$^\dagger$ & 0.0180 \\ \hline \end{tabular} \caption{Results on DUC 2003 dataset} \end{table*} \vspace{3mm} As shown in the table, the proposed systems outperform most existing systems on all three ROUGE scores. Both Borda and WCS failed to outperform the best state of art results. Even the simplistic SentRank algorithm outperforming most candidate systems and achieves a performance at par with the state of art systems in terms of ROUGE-2. While HybridRank achieves the best performance for ROUGE-1 and ROUGE-2 on both DUC 2003 and DUC 2004 datasets, GlobalRank performs best in terms of ROUGE-4 on DUC 2004. We performed a two sided sign test for determining whether the results were significantly different. The results clearly show that a rank aggregation technique that takes into account content of the summaries achieve a much higher ROUGE score vis-a-vis the systems that use only the ranked lists of sentences. \section{Conclusion} In this work we propose a new technique to estimate \emph{reliability} of a summarization system for a given document cluster. We define three systems, \emph{SentRank}, \emph{LocalRank} and \emph{GlobalRank} which take into account \emph{informativeness} of individual sentences, performance of candidates on a given document cluster, and overall performance of candidates on a held out development set, respectively. We use content overlap between summaries generated from several systems to estimate the relative importance of each system in case of LocalRank and SentRank. We combine the information from all these three systems to generate the final hybrid ranking (HybridRank) system. Summaries generated from such an aggregate system outperforms all the baseline and state of art systems as well as the baseline aggregation techniques by a good margin.

32a3f52bc8a9a50460e42fcfe10c612687939086