effie
This commit is contained in:
parent
8fb2fd057d
commit
5b3912f21b
2076
LaTeX/main.bib
2076
LaTeX/main.bib
File diff suppressed because it is too large
Load Diff
276
LaTeX/paper.tex
276
LaTeX/paper.tex

@ 13,7 +13,7 @@


\usepackage{environ} % for dynamic TikZ picture scaling


\usepackage{algorithm, algpseudocode} % Pseudo Codes / Algorithms


\usepackage[


style = apa, % citation style


style = authoryear, % citation style


isbn = false, % show isbn?


maxbibnames = 50, % maximal number of names in bibilography


maxcitenames = 2, % maximal number of names in text before et al.



@ 28,8 +28,8 @@


sortcites = false,


natbib = false,


dashed = true,


url = false,


doi = false,


url = true,


doi = true,


bibencoding = utf8


]{biblatex}





@ 272,9 +272,9 @@


\maketitle




\begin{abstract}


We consider supervised learning (regression/classification) problems where the independent variable is tensorvalued. We derive a multilinear sufficient reduction for the regression or classification problem modeling the conditional distribution of the predictors given the response as a member of the quadratic exponential family. Using manifold theory, we prove the consistency and asymptotic normality of the sufficient reduction. We develop estimation procedures of


sufficient reductions for both continuous and binary tensorvalued predictors. For continuous predictors, the algorithm is highly computationally efficient and is also applicable to situations where the dimension of


the reduction exceeds the sample size. We demonstrate the superior performance of our approach in simulations and realworld data examples for both continuous and binary tensorvalued predictors. The \textit{Chess data} analysis results agree with a human player's understanding of the game and confirm the relevance of our approach.


We consider supervised learning (regression/classification) problems where the independent variable is tensor valued. We derive a multilinear sufficient reduction for the regression or classification problem modeling the conditional distribution of the predictors given the response as a member of the quadratic exponential family. Using manifold theory, we prove the consistency and asymptotic normality of the sufficient reduction. We develop estimation procedures of


sufficient reductions for both continuous and binary tensor valued predictors. For continuous predictors, the algorithm is highly computationally efficient and is also applicable to situations where the dimension of


the reduction exceeds the sample size. We demonstrate the superior performance of our approach in simulations and realworld data examples for both continuous and binary tensor valued predictors. The \textit{Chess data} analysis results agree with a human player's understanding of the game and confirm the relevance of our approach.


\end{abstract}




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%



@ 283,61 +283,35 @@




Tensors are a mathematical tool to represent data of complex structure in statistics. \textit{Tensors} are considered as a generalization of matrices to higher dimensions: A tensor is a multidimensional array of numbers. For example, a secondorder tensor can be represented as a matrix, while a thirdorder tensor can be represented as a cube of matrices.




Complex data are collected at different times and/or under several conditions often involving a large number of multiindexed variables represented as tensorvalued data \parencite{KoldaBader2009}. They occur in largescale longitudinal studies \parencite[e.g.][]{Hoff2015}, in agricultural experiments and chemometrics and spectroscopy \parencite[e.g.][]{LeurgansRoss1992,Burdick1995}, in signal and video processing where sensors produce multiindexed data, e.g. over spatial, frequency, and temporal dimensions \parencite[e.g.][]{DeLathauwerCastaing2007,KofidisRegalia2005}, and in telecommunications \parencite[e.g.][]{DeAlmeidaEtAl2007}. Other examples of multiway data include 3D images of the brain, where the modes are the 3 spatial dimensions, and spatiotemporal weather imaging data, a set of image sequences represented as 2 spatial modes and 1 temporal mode.


Complex data are collected at different times and/or under several conditions often involving a large number of multiindexed variables represented as tensor valued data \parencite{KoldaBader2009}. They occur in largescale longitudinal studies \parencite[e.g.][]{Hoff2015}, in agricultural experiments and chemometrics and spectroscopy \parencite[e.g.][]{LeurgansRoss1992,Burdick1995}, in signal and video processing where sensors produce multiindexed data, e.g. over spatial, frequency, and temporal dimensions \parencite[e.g.][]{DeLathauwerCastaing2007,KofidisRegalia2005}, and in telecommunications \parencite[e.g.][]{DeAlmeidaEtAl2007}. Other examples of multiway data include 3D images of the brain, where the modes are the 3 spatial dimensions, and spatiotemporal weather imaging data, a set of image sequences represented as 2 spatial modes and 1 temporal mode.




% \begin{itemize}


% \item Review \cite{ZhouLiZhu2013} and see how you compare with them. They focus on the forward regression model with a scalar response but they claim that "Exploiting the array structure in imaging data, the new method substantially reduces the dimensionality of imaging data, which leads to efficient estimation and prediction."


% \item Read \cite{ZhouEtAl2023} to figure out the distribution they use for the tensorvalued predictors and briefly describe what they do.


% \item Read \cite{RabusseauKadri2016} to figure out what they do. They seem to draw both the response and the predictors from tensornormal with iid N(0,1) entries: "In order to leverage the tensor structure of the output data, we formulate the problem as the minimization of a least squares criterion subject to a multilinear rank constraint on the regression tensor. The rank constraint enforces the model to capture lowrank structure in the outputs and to explain dependencies between inputs and outputs in a lowdimensional multilinear subspace."


% \end{itemize}


Tensor regression models have been proposed to leverage the structure inherent in tensorvalued data. For instance, \textcite{HaoEtAl2021,ZhouLiZhu2013} focus on tensor covariates, while \textcite{RabusseauKadri2016,LiZhang2017,ZhouLiZhu2013} focus on tensor responses, and \textcite{Hoff2015,Lock2018} consider tensor on tensor regression. \textcite{HaoEtAl2021} modeled a scalar response as a flexible nonparametric function of tensor covariates. \textcite{ZhouLiZhu2013} assume the scalar response has a distribution in the exponential family given the tensorvalued predictors with the link modeled as a multilinear function of the predictors. \textcite{RabusseauKadri2016} model the tensorvalued response as a linear model with tensorvalued regression coefficients subject to a multilinear rank constraint. \textcite{LiZhang2017} approach the problem with a similar linear model but instead of a lowrank constraint the error term is assumed to have a separable Kronecker product structure while using a generalization of the envelope model \parencite{CookLiChiaromonte2010}. \textcite{ZhouEtAl2023} focus on partially observed tensor response given vectorvalued predictors with modewise sparsity constraints in the regression coefficients. \textcite{Hoff2015} extends an existing bilinear regression model to a tensor on tensor of conformable modes and dimensions regression model based on a Tucker product. \textcite{Lock2018} uses a tensor contraction to build a penalized least squares model for a tensor with arbitrary number of modes and dimensions.




%  RabusseauKadri2016 Y  x for tensor Y with vector x (HOLRR; Higher Order LowRank Regression)


%  LiZhang2017 Y  x for tensoe Y with vector x (envelope model)


%  ZhouEtAl2023 Y  x for tensor Y with vector x (sparse and partially observed)


Our approach considers the general regression problem of fitting a response of general form (univariate, multivariate, tensorvalued) on a tensorvalue predictor. We operate in the context of sufficient dimension reduction \parencite[e.g.]{Cook1998,Li2018} based on inverse regression, which leads us to regressing the tensor valued predictor on the response. In our setting, this necessitates transforming the response to tensorvalued functions, regardless of whether it is itself tensor valued. Because of the setting, our method shares commonalities with the tensor regression models referred to above, yet the modeling and methodology are novel. Specifically, our tensortotensor regression model is a generalized multilinear model similar to the generalized linear model of \parencite{ZhouLiZhu2013}. To bypass the explosion of the number of parameters to estimate, we assume the inverse regression error covariance has Kronecker product structure as do \textcite{LiZhang2017}. Our maximum likelihoodbased estimation does not require any penalty terms in contrast to the least squares and/or sparse approaches \parencite{ZhouLiZhu2013}. In the case of a tensor (multilinear) normal, given the tensor valued function of the response, our model exhibits similarities to the multilinear modeling of \textcite{Hoff2015}, but we use a generalized multilinear model and estimate the parameters with maximum likelihood instead of least squares. Moreover, a common issue in multilinear tensor regression models is the unidentifiability of the parameters, which we address in a completely different manner. For example, \textcite{LiZhang2017} developed theory that is based on orthogonal projection matrices to uniquely identify a subspace, while our approach is more general as it uses manifold theory.




%  ZhouLiZhu2013 y\in R (GLM) for y  Z, X for tensor X


%  HaoEtAl2021 y\in R for y  X for tensor X (sparse, element wise Bsplines)


In this paper, we present a modelbased \emph{Sufficient Dimension Reduction} (SDR) method for tensorvalued data with distribution in the quadratic exponential family assuming a separable Kronecker product structure of the first and second moment. The quadratic exponential family contains the multilinear normal and the multilinear Ising distributions, for continuous and binary tensorvalued random variables, respectively. By generalizing the parameter space to embedded manifolds we %obtain consistency and asymptotic normality results while


allow great modeling flexibility in the sufficient dimension reduction.




% Tensor regression models have been proposed to exploit the special structure of tensor covariates, e.g. \cite{HaoEtAl2021,ZhouLiZhu2013}, or tensor responses \cite{RabusseauKadri2016,LiZhang2017,ZhouEtAl2023} \cite{HaoEtAl2021} modeled a scalar response as a flexible nonparametric function of tensor covariates. \cite{ZhouLiZhu2013} assume the scalar response has a distribution in the exponential family given the tensorvalued predictors and model the link function as a multilinear function of the predictors. \cite{LiZhang2017} model the tensorvalued response as tensor normal. Rather than using $L_1$ type penalty functions to induce sparsity, they employ the envelope method (Cook, Li, and Chiaromonte Citation2010) to estimate the unknown regression coefficient. Moreover, the envelope method essentially identifies and uses the material information jointly. They develop an estimation algorithm and study the asymptotic properties of the estimator. the scalar response as These models try to utilize the sparse and lowrank structures in the tensors  either in the regression coefficient tensor or the response tensor  to boost performance on the regression task by reducing the number of free parameters.




Tensor regression models have been proposed to leverage the structure inherent in tensor valued data. For instance, \textcite{HaoEtAl2021,ZhouLiZhu2013} focus on tensor covariates, while \textcite{RabusseauKadri2016,LiZhang2017,ZhouLiZhu2013} focus on tensor responses, and \textcite{Hoff2015,Lock2018} consider tensor on tensor regression. \textcite{HaoEtAl2021} modeled a scalar response as a flexible nonparametric function of tensor covariates. \textcite{ZhouLiZhu2013} assume the scalar response has a distribution in the exponential family given the tensorvalued predictors with the link modeled as a multilinear function of the predictors. \textcite{RabusseauKadri2016} model the tensorvalued response as a linear model with tensor valued regression coefficients subject to a multilinear rank constraint. \textcite{LiZhang2017} approach the problem with a similar linear model but instead of a low rank constraint the error term is assumed to have a separable Kronecker product structure while using a generalization of the envelope model \parencite{CookLiChiaromonte2010}. \textcite{ZhouEtAl2023} focus on partially observed tensor response given vectorvalued predictors with modewise sparsity constraints in the regression coefficients. \textcite{Hoff2015} extends an existing bilinear regression model to a tensor on tensor of conformable modes and dimensions regression model based on a Tucker product. \textcite{Lock2018} uses a tensor contraction to build a penalized least squares model for a tensor with arbitrary number of modes and dimensions.




Our approach considers the general regression problem of fitting a response of general form (univariate, multivariate, tensorvalued) on a tensorvalue predictor. We operate in the context of sufficient dimension reduction \parencite[e.g.]{Cook1998,Li2018} based on inverse regression, which leads us to regressing the tensorvalued predictor on the response. In our setting, this necessitates transforming the response to tensorvalued functions, regardless of whether it is itself tensorvalued. Because of the setting, our method shares commonalities with the tensor regression models referred to above, yet the modeling and methodology are novel.


Specifically, our tensortotensor regression model is a generalized multilinear model similar to the generalized linear model of \parencite{ZhouLiZhu2013}. % but with tensor valued response by applying (a known) tensor valued function to the response in an inverse regression setting, reversing the role of response and predictors.


To bypass the explosion of the number of parameters to estimate, we assume the inverse regression error covariance has Kronecker product structure as do \textcite{LiZhang2017}. Our maximum likelihoodbased estimation does not require any penalty terms in contrast to the least squares and/or sparse approaches \parencite{ZhouLiZhu2013}. In the case of a tensor (multilinear) normal, given the tensorvalued function of the response, our model exhibits similarities to the multilinear modeling of \textcite{Hoff2015}, but we use a generalized multilinear model and estimate the parameters with maximum likelihood instead of least squares. Moreover, a common issue in multilinear tensor regression models is the unidentifiability of the parameters, which we address in a completely different manner. For example, \textcite{LiZhang2017} develop theory that is based on orthogonal projection matrices to uniquely identify a subspace, while our approach is more general as it uses manifold theory.






In this paper, we present a modelbased \emph{Sufficient Dimension Reduction} (SDR) method for tensorvalued data with distribution in the quadratic exponential family assuming a separable Kronecker product structure of the first and second moment. By generalizing the parameter space to embedded manifolds we obtain consistency and asymptotic normality results while allowing great modeling flexibility in the linear sufficient dimension reduction.




The quadratic exponential family contains the tensor normal and the tensor Ising distributions, for continuous and binary tensorvalued random variables, respectively.




Multilinear normal models have been used in various applications, including medical imaging \parencite{BasserPajevic2007,DrydenEtAl2009}, spatiotemporal data analysis \parencite{GreenewaldHero2014}, regression analysis for longitudinal relational data \parencite{Hoff2015}. One of the most important uses of the multilinear normal (MLN) distribution, and hence tensor analysis, is perhaps in magnetic resonance imaging (MRI) \parencite{OhlsonEtAl2013}. A recent survey \parencite{WangEtAl2022} and references therein contain more information and potential applications of multilinear tensor normal models.




The Ising\footnote{Also known as the \emph{LenzIsing} model as the physical assumptions of the model where developed by both Lenz and Ising \parencite{Niss2005} where Ising gave a closed form solution for the 1D lattice \parencite{Ising1925}.} model \parencite{Lenz1920,Ising1925,Niss2005} is a mathematical model originating in statistical physics to study ferromagnetism in a thermodynamic setting. It describes magnetic dipoles (atomic ``spins'' with values $\pm 1$) under an external magnetic field (first moments) while allowing twoway interactions (second moments) between direct neighbours on a lattice, a discrete grid. The Ising problem, as known in statistical physics, is to compute observables such as the magnetizations and correlations under the Boltzmann distribution\footnote{The Boltzmann distribution is a probability distribution over the states of a physical system in thermal equilibrium (constant temperature) that assigns higher probabilities to states with lower energy.} while the interaction structure and the magnetic fields are given. The ``reverse'' problem, where the couplings and fields are unknown and to be determined from observations of the spins, as in statistical inference, is known as the \emph{inverse Ising problem} \parencite{NguyenEtAl2017}. From this point of view, the Ising model is a member of a discrete quadratic exponential family \parencite{CoxWermuth1994,JohnsonEtAl1997} for multivariate binary outcomes where the interaction structure (nonzero correlations) is determined by the lattice. Generally, neither the values of couplings nor the interaction structure are known.


The Ising\footnote{Also known as the \emph{LenzIsing} model as the physical assumptions of the model where developed by both Lenz and Ising \parencite{Niss2005} where Ising gave a closed form solution for the 1D lattice \parencite{Ising1925}.} model \parencite{Lenz1920,Ising1925,Niss2005} is a mathematical model originating in statistical physics to study ferromagnetism in a thermodynamic setting. It describes magnetic dipoles (atomic ``spins'' with values $\pm 1$) under an external magnetic field (first moments) while allowing twoway interactions (second moments) between direct neighbors on a lattice, a discrete grid. The Ising problem, as known in statistical physics, is to compute observables such as the magnetizations and correlations under the Boltzmann distribution\footnote{The Boltzmann distribution is a probability distribution over the states of a physical system in thermal equilibrium (constant temperature) that assigns higher probabilities to states with lower energy.} while the interaction structure and the magnetic fields are given. The ``reverse'' problem, where the couplings and fields are unknown and to be determined from observations of the spins, as in statistical inference, is known as the \emph{inverse Ising problem} \parencite{NguyenEtAl2017}. From this point of view, the Ising model is a member of a discrete quadratic exponential family \parencite{CoxWermuth1994,JohnsonEtAl1997} for multivariate binary outcomes where the interaction structure (nonzero correlations) is determined by the lattice. Generally, neither the values of couplings nor the interaction structure are known.




In consequence, the Ising model is mostly used to model multivariate binary data in statistics. The states are ${0, 1}$ instead of $\pm 1$, and full interaction structure. It is related to a multitude of other models, among which the most prominent are: \emph{Graphical Models} and \emph{Markov Random Fields} to describe conditional dependence \parencite{Lauritzen1996,WainwrightJordan2008,LauritzenRichardson2002}, \emph{Potts models} \parencite{Besag1974,ChakrabortyEtAl2022} which generalize the Ising model to multiple states, the \emph{multivariate Bernoulli distribution} \parencite{Whittaker1990,JohnsonEtAl1997,DaiDingWahba2013} that also accommodates higherorder interactions (threeway and higher), \emph{(restricted) Botlzmann machines} \parencite{Smolensky1986,Hinton2002,FischerIgel2012} that introduce additional hidden variables for learning binary distributions. Most of these models can be used both in supervised and unsupervised settings.


Applications of the Ising model (and variations thereof) range from modeling neural firing patterns \parencite{SchneidmanEtAl2006}, gene expression data analysis \parencite{LezonEtAl2006}, and modeling financial markets \parencite{Bury2013}. See also \textcite{NguyenEtAl2017}.


In consequence, the Ising model is mostly used to model multivariate binary data in statistics. The states are ${0, 1}$ instead of $\pm 1$, and full interaction structure. It is related to a multitude of other models, among which the most prominent are: \emph{Graphical Models} and \emph{Markov Random Fields} to describe conditional dependence \parencite{Lauritzen1996,WainwrightJordan2008,LauritzenRichardson2002}, \emph{Potts models} \parencite{Besag1974,ChakrabortyEtAl2022} which generalize the Ising model to multiple states, the \emph{multivariate Bernoulli distribution} \parencite{Whittaker1990,JohnsonEtAl1997,DaiDingWahba2013} that also accommodates higherorder interactions (threeway and higher), \emph{(restricted) Botlzmann machines} \parencite{Smolensky1986,Hinton2002,FischerIgel2012} that introduce additional hidden variables for learning binary distributions. Most of these models can be used both in supervised and unsupervised settings. Applications of the Ising model (and variations thereof) range from modeling neural firing patterns \parencite{SchneidmanEtAl2006}, gene expression data analysis \parencite{LezonEtAl2006}, and modeling financial markets \parencite{Bury2013}. See also \textcite{NguyenEtAl2017}.




The $r$tensor Ising model in statistical physics is a generalization of the Ising model to $r$order interactions. \textcite{MukherjeeEtAl2020} study the oneparameter discrete exponential family for modeling dependent binary data where the interaction structure is given. In \textcite{LiuEtAl2023} the tensor structure itself is to be inferred. These models are fundamentally different from our approach where we rely on properties of the quadratic exponential family which models up to secondorder interactions. Another important difference is that we adopt the multilinear formulation as it is inherently linked to the observable structure of multiway data as opposed to describing the model coefficients with an $r$order tensor structure.


The $r$tensor Ising model in statistical physics is a generalization of the Ising model to $r$order interactions. \textcite{MukherjeeEtAl2022} study the oneparameter discrete exponential family for modeling dependent binary data where the interaction structure is given. In \textcite{LiuEtAl2023} the tensor structure itself is to be inferred. These models are fundamentally different from our approach where we rely on properties of the quadratic exponential family which models up to secondorder interactions. Another important difference is that we adopt the multilinear formulation as it is inherently linked to the observable structure of multiway data as opposed to describing the model coefficients with an $r$order tensor structure.




% \textcite{LiuEtAl2023,MukherjeeEtAl2020,ChengEtAl2014,Habeck2014}


Our main contributions are (a) formulating the dimension reduction problem via the quadratic exponential family, which allows us to derive the sufficient dimension reduction in closed form, (b) defining the parameter space as an embedded manifold, which provides great flexibility in modeling, (c) deriving the maximum likelihood estimator of the sufficient reduction subject to multilinear constraints and overcoming parameter nonidentifiability, (d) developing estimation algorithms which in the case of multilinear normal predictors is fast and efficient, and (e) establishing the consistency and asymptotic normality of the estimators.




% The Ising model for multivariate binary outcomes belongs to the class of discrete exponential families. Its defining feature is that the sufficient statistic involves a quadratic term to capture correlations arising from pairwise interactions.


% The tensor Ising model is a higherorder Ising model for tensorvalued binary outcomes.


% %From \cite{MukherjeeEtAl2020}


% Higherorder Ising models arise naturally in the study of multiatom interactions in lattice gas models, such as the squarelattice eightvertex model, the AshkinTeller model, and Suzuki's pseudo3D anisotropic model (cf. [6, 33, 36, 37, 49, 55, 56, 61, 62] and the references therein). More recently, higherorder spin systems have also been proposed for modeling peergroup effects in social networks [22]. \efi{Daniel: comment on what these guys do and contrast with your setting} In our approach, the parameter is not constrained to be scalar


% We derive maximum likelihood estimates for all first and second order interactions and propose a gradientbased optimization algorithm.


Even though our motivation is rooted in the SDR perspective, our proposal concerns inference on any regression model with a tensor valued response and predictors of any type. Thus, our approach can be used as a standalone model for such data regardless of whether one is interested in deriving sufficient reductions and/or reducing the dimension of the data. Our results in the framework of the quadratic exponential family for tensor valued variables; i.e., consistency and asymptotic normality, apply to both multilinear normal \parencite{KolloVonRosen2005,Hoff2011,OhlsonEtAl2013} and multilinear Ising models, as defined in \cref{sec:ising_estimation}.




As an aside, even though our motivation stems from the SDR perspective, our proposal concerns inference on any regression model with a tensorvalued response and any type of predictors. Thus, our approach can be used as a standalone model for such data regardless of whether one is interested in deriving sufficient reductions and/or reducing the dimension of the data. Our results in the framework of the quadratic exponential family for tensorvalued variables; i.e., consistency and asymptotic normality, apply to both multilinear normal \efi{\ref{?} and multilinear Ising models, as defined in this paper in Sec. ??.}


The structure of this paper is as follows. We introduce our notation in \cref{sec:notation}. \Cref{sec:problemformulation} details the problem we consider and in \cref{sec:gmlmmodel} we introduce our model. Continuing in \cref{sec:mlestimation} we provide the basis for a general maximum likelihood estimation procedure and derive specialized methods for multilinear normal as well as the multilinear Ising distributions. \Cref{sec:manifolds} gives a short introduction into manifolds and provides the basis for applying the consistency and asymtotic normality results from \cref{sec:asymtotics}. Simulations for continuous and binary predictors are the subject of \cref{sec:simulations}. Finally, in \cref{sec:dataanalysis} we apply our model to EEG data and perform a proof of concept data analysis where a chess board is interpreted as a collection of binary $8\times 8$ matrices.








The structure of this paper is as follows. We introduce our notation in \cref{sec:notation}. \Cref{sec:problemformulation} details the problem we consider and in \cref{sec:gmlmmodel} we introduce our model. Continuing in \cref{sec:mlestimation} we provide the basis for a general maximum likelihood estimation procedure and derive specialized methods for tensor normal as well as the tensor Ising distributions. \Cref{sec:manifolds} gives a short introduction into manifolds and provides the basis for applying the consistency and asymtotic normality results from \cref{sec:asymtotics}. Simulations for continuous and binary predictors are the subject of \cref{sec:simulations}. Finally, in \cref{sec:dataanalysis} we apply our model to EEG data and perform a prove of concept data analysis example where a chess board is interpreted as a collection of binary $8\times 8$ matrices.




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\section{Notation}\label{sec:notation}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


Let $\ten{A}\in\mathbb{R}^{q_1\times \ldots\times q_r}$ denotes an order\footnote{Also referred to as rank, therefore the variable name $r$, but this term is \emph{not} used as it leads to confusion with the concept of rank of a matrix.} $r$ tensor, where $r\in\mathbb{N}$ is the number of modes or axes (dimensions) of $\ten{A}$ and $\ten{A}_{i_1,...,i_r} \in \mathbb{R}$ is its $(i_1, \ldots, i_r)$th entry. For example, a $p \times q$ matrix $\mat{B}$ has two modes, the rows and columns. For matrices $\mat{B}_k\in\mathbb{R}^{p_k\times q_k}$, $k\in[r] = \{1, 2, \ldots, r\}$, the \emph{multilinear multiplication}, or \emph{Tucker operator} \parencite{Kolda2006}, is defined element wise as


Let $\ten{A}\in\mathbb{R}^{q_1\times \ldots\times q_r}$ denotes an order\footnote{Also referred to as rank, therefore the variable name $r$, but this term is \emph{not} used as it leads to confusion with the concept of rank of a matrix.} $r$ tensor, where $r\in\mathbb{N}$ is the number of modes or axes (dimensions) of $\ten{A}$ and $\ten{A}_{i_1,...,i_r} \in \mathbb{R}$ is its $(i_1, \ldots, i_r)$th entry. For example, a $p \times q$ matrix $\mat{B}$ has two modes, rows and columns. For matrices $\mat{B}_k\in\mathbb{R}^{p_k\times q_k}$, $k\in[r] = \{1, 2, \ldots, r\}$, the \emph{multilinear multiplication}, or \emph{Tucker operator} \parencite{Kolda2006}, is defined elementwise as


\begin{displaymath}


(\ten{A}\times\{\mat{B}_1, \ldots, \mat{B}_r\})_{j_1, \ldots, j_r} = \sum_{i_1, \ldots, i_r = 1}^{q_1, \ldots, q_r} \ten{A}_{i_1, \ldots, i_r}(\mat{B}_{1})_{j_1, i_1} \cdots (\mat{B}_{r})_{j_r, i_r}


\end{displaymath}



@ 345,25 +319,24 @@ which results in an order $r$ tensor of dimension $p_1\times ...\times p_k$. The


\begin{displaymath}


\ten{A}\times_k\mat{B}_k = \ten{A}\times\{\mat{I}_{q_1}, \ldots, \mat{I}_{q_{k1}}, \mat{B}_{k}, \mat{I}_{q_{k+1}}, \ldots, \mat{I}_{q_r}\}.


\end{displaymath}


The notation $\ten{A}\mlm_{k\in S}\mat{B}_k$ is short hand for the iterative application of the mode product for all indices in $S\subseteq[r]$. For example $\ten{A}\mlm_{k\in\{2, 5\}}\mat{B}_k = \ten{A}\times_2\mat{B}_2\times_5\mat{B}_5$. By only allowing $S$ to be a set, this notation is unambiguous because the mode product commutes for different modes; i.e., $\ten{A}\times_j\mat{B}_j\times_k\mat{B}_k = \ten{A}\times_k\mat{B}_k\times_j\mat{B}_j$ for $j\neq k$. For example, let $\mat{A}, \mat{B}_1, \mat{B}_2$ be matrices (of matching dimensions). The bilinear modeproduct and multilinear multiplication relate to the well known matrixmatrix multiplications as


The notation $\ten{A}\mlm_{k\in S}\mat{B}_k$ is shorthand for the iterative application of the mode product for all indices in $S\subseteq[r]$. For example $\ten{A}\mlm_{k\in\{2, 5\}}\mat{B}_k = \ten{A}\times_2\mat{B}_2\times_5\mat{B}_5$. By only allowing $S$ to be a set, this notation is unambiguous because the mode product commutes for different modes; i.e., $\ten{A}\times_j\mat{B}_j\times_k\mat{B}_k = \ten{A}\times_k\mat{B}_k\times_j\mat{B}_j$ for $j\neq k$. For example, let $\mat{A}, \mat{B}_1, \mat{B}_2$ be matrices (of matching dimensions). The bilinear modeproduct and multilinear multiplication relate to the wellknown matrixmatrix multiplications as


\begin{displaymath}


\mat{A}\times_1\mat{B}_1 = \mat{B}_1\mat{A}, \qquad


\mat{A}\times_2\mat{B}_2 = \mat{A}\t{\mat{B}_2}, \qquad


\mat{A}\mlm_{k = 1}^2\mat{B}_k = \mat{A}\mlm_{k \in \{1, 2\}}\mat{B}_k = \mat{B}_1\mat{A}\t{\mat{B}_2}.


\end{displaymath}




%Matrices and tensors can be \emph{vectorized} by the \emph{vectorization} operator $\vec$.


The operator $\vec$ maps an array to a vector. Specifically, $\vec(\mat{B})$ stands for the $pq \times 1$ vector of the $p \times q$ matrix $\mat{B}$ resulting from stacking the columns of $\mat{B}$ one after the other. For a tensor $\ten{A}$ of order $r$ and dimensions $q_1, \ldots, q_r$, $\vec(\ten{A})$ is the $q_1 q_2 \ldots q_r \times 1$ vector with the elements of $\ten{A}$ stacked one after the other in the order $r$ then $r1$, and so on. For example, if $\ten{A}$ is a 3dimensional array, $\vec(\ten{A})=\t{(\t{\vec(\ten{A}_{:,:,1})},\t{\vec(\ten{A}_{:,:,2})},\ldots,\t{\vec(\ten{A}_{:,:,q_3})})}$. We use the notation $\ten{A}\equiv \ten{B}$ for objects $\ten{A}, \ten{B}$ of any shape if and only if $\vec(\ten{A}) = \vec(\ten{B})$.




The \emph{inner product} between two tensors of the same order and dimensions is


\begin{displaymath}


\langle\ten{A}, \ten{B}\rangle = \sum_{i_1, \ldots, i_r} \ten{A}_{i_1, \ldots, i_r}\ten{B}_{i_1, \ldots, i_r}


\end{displaymath}


This leads to the definition of the \emph{Frobenius norm} for tensors, $\\ten{A}\_F = \sqrt{\langle\ten{A}, \ten{A}\rangle}$ and is the straightforward extension of the Frobenius norm for matrices and vectors. The \emph{outer product} between two tensors $\ten{A}$ of dimensions $q_1, \ldots, q_r$ and $\ten{B}$ of dimensions $p_1, \ldots, p_l$ is a tensor $\ten{A}\circ\ten{B}$ of order $r + l$ and dimensions $q_1, \ldots, q_r, p_1, \ldots, p_l$, such that


This leads to the definition of the \emph{Frobenius norm} for tensors, $\\ten{A}\_F = \sqrt{\langle\ten{A}, \ten{A}\rangle}$ and is the straightforward extension of the Frobenius norm for matrices and vectors. The \emph{outer product} between two tensors $\ten{A}$ of dimensions $q_1, \ldots, q_r$ and $\ten{B}$ of dimensions $p_1, \ldots, p_l$ is a tensor $\ten{A}\circ\ten{B}$ of order $r + l$ and dimensions $q_1, \ldots, q_r, p_1, \ldots, p_l$, such that


\begin{displaymath}


\ten{A}\circ\ten{B} \equiv (\vec\ten{A})\t{(\vec{\ten{B}})}.


\end{displaymath}


Let $\ten{K} : \mathbb{R}^{q_1, ..., q_{2 r}}\to\mathbb{R}^{q_1 q_{r + 1}, ..., q_r q_{2 r}}$ be defined element wise with indices $1\leq i_j + 1\leq q_j q_{r + j}$ for $j = 1, ..., r$ as


Let $\ten{K} : \mathbb{R}^{q_1, ..., q_{2 r}}\to\mathbb{R}^{q_1 q_{r + 1}, ..., q_r q_{2 r}}$ be defined elementwise with indices $1\leq i_j + 1\leq q_j q_{r + j}$ for $j = 1, ..., r$ as


\begin{align*}


\ten{K}(\ten{A})_{i_1 + 1, ..., i_r + 1} &= \ten{A}_{\lfloor i_1 / q_{r + 1}\rfloor + 1, ..., \lfloor i_r / q_{2 r} \rfloor + 1, (i_1\operatorname{mod}q_{r + 1}) + 1, ..., (i_r\operatorname{mod}q_{2 r}) + 1}


\end{align*}



@ 392,7 +365,7 @@ where the vectorized quantities $\vec{\ten{X}}\in\mathbb{R}^p$ and $\vec\ten{F}(


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\section{Problem Formulation}\label{sec:problemformulation}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


Our goal is to infer the cumulative distribution function (cdf) $F$ of $Y\mid \ten{X}$, where $\ten{X}$ is assumed to admit $r$tensor structure of dimension $p_1\times ... \times p_r$ with continuous or discrete entries, and the response $Y$ is unconstrained. The predictor $\ten{X}$ is a complex object; to simplify the problem we assume there exists a tensorvalued function of lower dimension $\ten{R}:\ten{X}\mapsto \ten{R}(\ten{X})$ such that


Our goal is to infer the cumulative distribution function (cdf) $F$ of $Y\mid \ten{X}$, where $\ten{X}$ is assumed to admit $r$tensor structure of dimension $p_1\times ... \times p_r$ with continuous or discrete entries, and the response $Y$ is unconstrained. The predictor $\ten{X}$ is a complex object; to simplify the problem we assume there exists a tensor valued function of lower dimension $\ten{R}:\ten{X}\mapsto \ten{R}(\ten{X})$ such that


\begin{displaymath}


F(Y\mid \ten{X}) = F(Y\mid \ten{R}(\ten{X})).


\end{displaymath}



@ 412,13 +385,13 @@ f_{\mat{\eta}_y}(\ten{X}\mid Y = y)


&= h(\ten{X})\exp(\t{\mat{\eta}_y}\mat{t}(\ten{X})  b(\mat{\eta}_y)) \nonumber \\


&= h(\ten{X})\exp(\langle \mat{t}_1(\ten{X}), \mat{\eta}_{1y} \rangle + \langle \mat{t}_2(\ten{X}), \mat{\eta}_{2y} \rangle  b(\mat{\eta}_{y})) \label{eq:quaddensity}


\end{align}


where $\mat{t}_1(\ten{X})=\vec \ten{X}$ and $\mat{t}_2(\ten{X})$ is linear in $\ten{X}\circ\ten{X}$. The dependence of $\ten{X}$ on $Y$ is fully captured in the natural parameter $\mat{\eta}_y$. The function $h$ is nonnegative realvalued and $b$ is assumed to be at least twice continuously differentiable and strictly convex. An important feature of the \emph{quadratic exponential family} is that the distribution of its members is fully characterized by their first two moments. Distributions within the quadratic exponential family include the \emph{tensor normal} (\cref{sec:tensornormalestimation}) and \emph{tensor Ising model} (\cref{sec:ising_estimation}, a generalization of the (inverse) Ising model which is a multivariate Bernoulli with up to second order interactions) and mixtures of these two.


where $\mat{t}_1(\ten{X})=\vec \ten{X}$ and $\mat{t}_2(\ten{X})$ is linear in $\ten{X}\circ\ten{X}$. The dependence of $\ten{X}$ on $Y$ is fully captured in the natural parameter $\mat{\eta}_y$. The function $h$ is nonnegative realvalued and $b$ is assumed to be at least twice continuously differentiable and strictly convex. An important feature of the \emph{quadratic exponential family} is that the distribution of its members is fully characterized by their first two moments. Distributions within the quadratic exponential family include the \emph{multilinear normal} (\cref{sec:tensornormalestimation}) and \emph{multilinear Ising model} (\cref{sec:ising_estimation}, a generalization of the (inverse) Ising model which is a multivariate Bernoulli with up to second order interactions) and mixtures of these two.




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\section{The Generalized MultiLinear Model}\label{sec:gmlmmodel}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%




In model \eqref{eq:quaddensity}, the dependence of $\ten{X}$ and $Y$ is absorbed in $\mat{\eta}_y$, and $\mat{t}(\ten{X})$ is the minimal sufficient statistic for the \textit{pseudo}parameter\footnote{$\mat{\eta}_y$ is a function of the response $Y$, thus it is not a parameter in the formal statistical sense. It is considered as a parameter when using the equivalence in \eqref{eq:inverseregressionsdr} and view $Y$ as a parameter as a device to derive the sufficient reduction from the inverse regression.} $\mat{\eta}_y = (\mat{\eta}_{1y}, \mat{\eta}_{2y})$ with


In model \eqref{eq:quaddensity}, the dependence of $\ten{X}$ and $Y$ is absorbed in $\mat{\eta}_y$, and $\mat{t}(\ten{X})$ is the minimal sufficient statistic for the \textit{pseudo}parameter\footnote{$\mat{\eta}_y$ is a function of the response $Y$, so it is not a parameter in the formal statistical sense. We view it as a parameter in order to leverage \eqref{eq:inverseregressionsdr} and derive the sufficient reduction from the inverse regression.} $\mat{\eta}_y = (\mat{\eta}_{1y}, \mat{\eta}_{2y})$ with


\begin{align}\label{eq:tstat}


\mat{t}(\ten{X}) &= (\mat{t}_1(\ten{X}),\mat{t}_2(\ten{X}))=(\vec{\ten{X}}, \mat{T}_2\vech((\vec\ten{X})\t{(\vec\ten{X})})),


\end{align}



@ 432,7 +405,7 @@ where $\mat{D}_p$ is the \emph{duplication matrix} from \textcite[Ch.~11]{Abadir


\begin{equation}\label{eq:quadraticexpfam}


f_{\eta_y}(\ten{X}\mid Y = y) = h(\ten{X})\exp\left(\langle \vec \ten{X}, \mat{\eta}_{1y} \rangle + \langle \vec(\ten{X}\circ\ten{X}), \t{(\mat{T}_2\pinv{\mat{D}_p})}\mat{\eta}_{2y} \rangle  b(\mat{\eta}_y)\right)


\end{equation}


The exponential family in \eqref{eq:quadraticexpfam} is easily generalizable to any order. This, though, would result in the number of parameters becoming prohibitive to estimate, which is also the reason why we opted for the second order exponential family in our formulation.


The exponential family in \eqref{eq:quadraticexpfam} is easily generalizable to any order. This, though, would result in the number of parameters becoming prohibitive to estimate, which is also the reason why we opted for the secondorder exponential family in our formulation.




By the equivalence in \eqref{eq:inverseregressionsdr}, in order to find the sufficient reduction $\ten{R}(\ten{X})$ we need to infer $\mat{\eta}_{1y}$, and $\mat{\eta}_{2y}$. This is reminiscent of generalized linear modeling, which we extend to a multilinear formulation next.


Suppose $\ten{F}_y$ is a known mapping of $y$ with zero expectation $\E_Y\ten{F}_Y = 0$. We assume the dependence of $\ten{X}$ and $Y$ is reflected only in the first parameter and let



@ 440,7 +413,7 @@ Suppose $\ten{F}_y$ is a known mapping of $y$ with zero expectation $\E_Y\ten{F}


\mat{\eta}_{1y} &= \vec{\overline{\ten{\eta}}} + \mat{B}\vec\ten{F}_y, \label{eq:eta1manifold} \\


\mat{\eta}_{2} &= \t{(\pinv{(\mat{T}_2\pinv{\mat{D}_p})})}\vec(c\,\mat{\Omega}), \label{eq:eta2manifold}


\end{align}


where $\overline{\ten{\eta}}\in\mathbb{R}^{p_1\times\ldots\times p_r}$, $\mat{\Omega} \in \mathbb{R}^{p \times p}$ is positive definite with $p = \prod_{j = 1}^{r} p_j$, and $c\in\mathbb{R}$ is a known constant determined by the distribution to ease modeling. That is, we assume that only $\mat{\eta}_{1y}$ depends on $Y$ through $\mat{B}$. The second parameter $\mat{\eta}_2$ captures the second order interaction structure of $\ten{X}$, which we assume not to depend on the response $Y$. In order to relate individual modes of $\ten{X}$ to the response, allowing flexibility in modeling, we assume $\ten{F}_y$ takes values in $\mathbb{R}^{q_1\times ...\times q_r}$; that is, $\ten{F}_y$ is a tensor valued independent variable. This, in turn, leads to imposing corresponding tensor structure to the regression parameter $\mat{B}$. Thus, \eqref{eq:eta1manifold} becomes


where $\overline{\ten{\eta}}\in\mathbb{R}^{p_1\times\ldots\times p_r}$, $\mat{\Omega} \in \mathbb{R}^{p \times p}$ is positive definite with $p = \prod_{j = 1}^{r} p_j$, and $c\in\mathbb{R}$ is a known constant determined by the distribution to ease modeling. That is, we assume that only $\mat{\eta}_{1y}$ depends on $Y$ through $\mat{B}$. The second parameter $\mat{\eta}_2$ captures the second order interaction structure of $\ten{X}$, which we assume not to depend on the response $Y$. To relate individual modes of $\ten{X}$ to the response, allowing flexibility in modeling, we assume $\ten{F}_y$ takes values in $\mathbb{R}^{q_1\times ...\times q_r}$; that is, $\ten{F}_y$ is a tensor valued independent variable. This, in turn, leads to imposing a corresponding tensor structure to the regression parameter $\mat{B}$. Thus, \eqref{eq:eta1manifold} becomes


\begin{align}


\mat{\eta}_{1y} &=


\vec\biggl(\overline{\ten{\eta}} + \ten{F}_y\mlm_{j = 1}^{r}\mat{\beta}_j\biggr), \label{eq:eta1}



@ 489,11 +462,9 @@ The reduction in vectorized form is $\vec\ten{R}(\ten{X})=\t{\mat{B}}\vec(\ten{X


\begin{align*}


f_{\theta}(\mat{x}\mid Y = y)


&= h(\mat{x})\exp(\langle\mat{x}, \mat{\eta}_{1y}(\mat{\theta})\rangle + \langle\vec(\mat{x}\circ\mat{x}), \mat{\eta}_2(\mat{\theta})\rangle  b(\mat{\eta}_y(\mat{\theta}))) \\


% &= h(\mat{x})\exp(\t{\mat{\eta}_{1y}(\theta)}\mat{x} + \t{\vec(\mat{x}\circ\mat{x})}\mat{\eta}_2(\mat{\theta})  b(\mat{\eta}_y(\mat{\theta}))) \\


&= h(\mat{x})\exp(\t{(\overline{\mat{\eta}} + \mat{\beta}\mat{f}_y)}\mat{x} + c\,\t{\mat{x}}\mat{\Omega}\,\mat{x}  b(\mat{\eta}_y(\mat{\theta}))).


\end{align*}


using the relation of $\mat{\theta}$ to the natural parameters given by $\mat{\eta}_{1y}(\mat{\theta}) = \overline{\mat{\eta}} + \mat{\beta}\mat{f}_y$ and $\mat{\eta}_2(\theta) = c\,\mat{\Omega}$.


% where the number of unknown parameters is $p + \dim(\StiefelNonCompact{p}{q}) + \dim(\SymPosDefMat{p}) = p\frac{p + 2 q + 3}{2}$.


\end{example}




\begin{example}[Matrix valued $\mat{X}$ ($r = 2$)]



@ 518,7 +489,7 @@ The maximum likelihood estimate of $\mat{\theta}_0$ is the solution to the optim


\end{equation}


with $\hat{\mat{\theta}}_n = (\vec\widehat{\overline{\ten{\eta}}}, \vec\widehat{\mat{B}}, \vech\widetilde{\mat{\Omega}})$ where $\widehat{\mat{B}} = \bigkron_{k = r}^{1}\widehat{\mat{\beta}}_k$ and $\widehat{\mat{\Omega}} = \bigkron_{k = r}^{1}\widehat{\mat{\Omega}}_k$.




A straightforward and general method for parameter estimation is \emph{gradient descent}. To apply gradient based optimization, we compute the gradients of $l_n$ in \cref{thm:grad}.


A straightforward and general method for parameter estimation is \emph{gradient descent}. To apply gradientbased optimization, we compute the gradients of $l_n$ in \cref{thm:grad}.




\begin{theorem}\label{thm:grad}


For $n$ i.i.d. observations $(\ten{X}_i, y_i), i = 1, ..., n$ the loglikelihood is of the form in \eqref{eq:loglikelihood} with $\mat{\theta}$ being the collection of all GMLM parameters $\overline{\ten{\eta}}$, ${\mat{B}} = \bigkron_{k = r}^{1}{\mat{\beta}}_k$ and ${\mat{\Omega}} = \bigkron_{k = r}^{1}{\mat{\Omega}}_k$ for $k = 1, ..., r$. Let $\ten{G}_2(\mat{\eta}_y)$ be a tensor of dimensions $p_1, \ldots, p_r$ such that



@ 535,42 +506,31 @@ A straightforward and general method for parameter estimation is \emph{gradient


If $\mat{T}_2$ is the identity matrix $\mat{I}_{p(p + 1) / 2}$, then $\ten{G}_2(\mat{\eta}_y) = \ten{g}_2(\mat{\eta}_y)$.


\end{theorem}




Although the general case of any GMLM model can be fitted via gradient descent using \cref{thm:grad}, this may be very inefficient. In \cref{thm:grad}, $\mat{T}_2$ can be used to introduce flexible second moment structures. For example, it allows modeling effects differently for predictor components, as described in \cref{sec:ising_estimation} after Eqn. \eqref{eq:isingcondprob}. In the remainder, we focus on $\mat{T}_2$'s that are identity matrices. This approach simplifies the estimation algorithm and the speed of the numerical calculation in the case of tensor normal predictors. % In the case of the tensor normal distribution,


An iterative cyclic updating scheme is derived in \cref{sec:tensornormalestimation}, which has much faster convergence, is stable and does not require hyperparameters, as will be discussed later. On the other hand, the Ising model does not allow such a scheme. There we need to use a gradientbased method, which is the subject of \cref{sec:ising_estimation}.


Although the general case of any GMLM model can be fitted via gradient descent using \cref{thm:grad}, this may be very inefficient. In \cref{thm:grad}, $\mat{T}_2$ can be used to introduce flexible second moment structures. For example, it allows modeling effects differently for predictor components, as described in \cref{sec:ising_estimation} after Eqn. \eqref{eq:isingcondprob}. In the remainder, we focus on $\mat{T}_2$'s that are identity matrices. This approach simplifies the estimation algorithm and the speed of the numerical calculation in the case of multilinear normal predictors. An iterative cyclic updating scheme is derived in \cref{sec:tensornormalestimation}, which has much faster convergence, is stable, and does not require hyperparameters, as will be discussed later. On the other hand, the Ising model does not allow such a scheme. There we need to use a gradientbased method, which is the subject of \cref{sec:ising_estimation}.




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsection{Tensor Normal}\label{sec:tensornormalestimation}


\subsection{MultiLinear Normal}\label{sec:tensornormalestimation}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


The \emph{multilinear normal} is the extension of the matrix normal to tensorvalued random variables and a member of the quadratic exponential family \eqref{eq:quadraticexpfam} under \eqref{eq:eta2}. \textcite{Dawid1981,Arnold1981} introduced the term matrix normal and, in particular, \textcite{Arnold1981} provided several theoretical results, such as its density, moments and conditional distributions of its components. The matrix normal distribution is a bilinear normal distribution; a distribution of a twoway array, each component


representing a vector of observations \parencite{OhlsonEtAl2013}. \textcite{KolloVonRosen2005,Hoff2011,OhlsonEtAl2013} presented the extension of the bilinear to the multilinear normal distribution, what we call tensor normal, using a parallel extension of bilinear matrices to multilinear tensors \parencite{Comon2009}.


The \emph{multilinear normal} is the extension of the matrix normal to tensor valued random variables and a member of the quadratic exponential family \eqref{eq:quadraticexpfam} under \eqref{eq:eta2}. \textcite{Dawid1981,Arnold1981} introduced the term matrix normal and, in particular, \textcite{Arnold1981} provided several theoretical results, such as its density, moments and conditional distributions of its components. The matrix normal distribution is a bilinear normal distribution; a distribution of a twoway array, each component


representing a vector of observations \parencite{OhlsonEtAl2013}. \textcite{KolloVonRosen2005,Hoff2011,OhlsonEtAl2013} presented the extension of the bilinear to the multilinear normal distribution, using a parallel extension of bilinear matrices to multilinear tensors \parencite{Comon2009}.




The defining feature of the matrix normal distribution, and its tensor extension, is the Kronecker product structure of its covariance. This formulation, where the covariates are multivariate normal with multiway covariance structure modeled as a Kronecker product of matrices of much lower dimension, aims to overcome the significant modeling and computational challenges arising from the high computational complexity of manipulating tensor representations \parencite[see, e.g.,][]{HillarLim2013,WangEtAl2022}.


The defining feature of the matrix normal distribution, and its multilinear extension, is the Kronecker product structure of its covariance. This formulation, where the covariates are multivariate normal with multiway covariance structure modeled as a Kronecker product of matrices of much lower dimension, aims to overcome the significant modeling and computational challenges arising from the high computational complexity of manipulating tensor representations \parencite[see, e.g.,][]{HillarLim2013,WangEtAl2022}.




Multilinear tensor normal %Kroneckerseparable covariance


models have been used in various applications, including


medical imaging \parencite{BasserPajevic2007,DrydenEtAl2009}, spatiotemporal data analysis \parencite{GreenewaldHero2014}, regression analysis


for longitudinal relational data \parencite{Hoff2015}.


%, radar [AFJ10], and multipleinputmultipleoutput (MIMO) communications [WJS08].


One of the most important uses of the multilinear normal (MLN) distribution, and hence tensor analysis, is perhaps in magnetic resonance imaging (MRI) \parencite{OhlsonEtAl2013}.


A recent survey \parencite{WangEtAl2022} and references therein contain more information and potential applications of multilinear tensor normal models.


%The first occurrence of the \textit{matrix normal} we found, even though not explicitly called as such, was in \textcite{SrivastavaKhatri1979}.




Suppose $\ten{X}\mid Y = y$ follows a tensor normal distribution with mean $\ten{\mu}_y$ and covariance $\mat{\Sigma} = \bigkron_{k = r}^{1}\mat{\Sigma}_k$. We assume the distribution is nondegenerate which means that the covariances $\mat{\Sigma}_k$ are symmetric positive definite matrices. Its density is given by


Suppose $\ten{X}\mid Y = y$ follows a multilinear normal distribution with mean $\ten{\mu}_y$ and covariance $\mat{\Sigma} = \bigkron_{k = r}^{1}\mat{\Sigma}_k$. We assume the distribution is nondegenerate which means that the covariances $\mat{\Sigma}_k$ are symmetric positive definite matrices. Its density is given by


\begin{displaymath}


f_{\mat{\theta}}(\ten{X}\mid Y = y) = (2\pi)^{p / 2}\prod_{k = 1}^{r}\det(\mat{\Sigma}_k)^{p / 2 p_k}\exp\left( \frac{1}{2}\left\langle\ten{X}  \ten{\mu}_y, (\ten{X}  \ten{\mu}_y)\mlm_{k = 1}^{r}\mat{\Sigma}_k^{1} \right\rangle \right).


\end{displaymath}


For the sake of simplicity and w.l.o.g., we assume $\ten{X}$ has 0 marginal expectation; i.e., $\E\ten{X} = 0$. Rewriting this in the quadratic exponential family form \eqref{eq:quadraticexpfam}, determines the scaling constant $c = 1/2$. The relation to the GMLM parameters $\overline{\ten{\eta}}, \mat{\beta}_k$ and $\mat{\Omega}_k$, for $k = 1, \ldots, r$ is


For the sake of simplicity and w.l.o.g., we assume $\ten{X}$ has 0 marginal expectation; i.e., $\E\ten{X} = 0$. Rewriting this in the quadratic exponential family form \eqref{eq:quadraticexpfam} determines the scaling constant $c = 1/2$. The relation to the GMLM parameters $\overline{\ten{\eta}}, \mat{\beta}_k$ and $\mat{\Omega}_k$, for $k = 1, \ldots, r$ is


\begin{equation}\label{eq:tnormal_cond_params}


\ten{\mu}_y = \ten{F}_y\mlm_{k = 1}^{r}\mat{\Omega}_k^{1}\mat{\beta}_k, \qquad


\mat{\Omega}_k = \mat{\Sigma}_k^{1},


\end{equation}


where we used that $\overline{\ten{\eta}} = 0$ due to $0 = \E\ten{X} = \E\E[\ten{X}\mid Y] = \E\ten{\mu}_Y$ in combination with $\E\ten{F}_Y = 0$. Additionally, all the $\mat{\Omega}_k$'s are symmetric positive definite, because the $\mat{\Sigma}_k$'s are. This lead to another simplification since then $\mat{T}_2$ in \eqref{eq:tstat} equals the identity. This also means that the gradients of the loglikelihood $l_n$ in \cref{thm:grad} are simpler. We obtain


where we used that $\overline{\ten{\eta}} = 0$ due to $0 = \E\ten{X} = \E\E[\ten{X}\mid Y] = \E\ten{\mu}_Y$ in combination with $\E\ten{F}_Y = 0$. Additionally, all the $\mat{\Omega}_k$'s are symmetric positive definite, because the $\mat{\Sigma}_k$'s are. This leads to another simplification since then $\mat{T}_2$ in \eqref{eq:tstat} equals the identity. This also means that the gradients of the loglikelihood $l_n$ in \cref{thm:grad} are simpler. We obtain


\begin{displaymath}


\ten{g}_1(\mat{\eta}_y) = \E[\ten{X}\mid Y = y] = \ten{\mu}_y, \qquad


\ten{G}_2(\mat{\eta}_y) = \ten{g}_2(\mat{\eta}_y) = \E[\ten{X}\circ\ten{X}\mid Y = y] \equiv \bigkron_{k = r}^1\mat{\Sigma}_k + (\vec{\ten{\mu}}_y)\t{(\vec{\ten{\mu}}_y)}.


\end{displaymath}


In practice, we assume we have a random sample of $n$ observations $(\ten{X}_i, \ten{F}_{y_i})$ from the joint distribution. We start the estimation process by demeaning the data. Then, only the reduction matrices $\mat{\beta}_k$ and the scatter matrices $\mat{\Omega}_k$ need to be estimated. To solve the optimization problem \eqref{eq:mle}, with $\overline{\ten{\eta}} = 0$ we initialize the parameters using a simple heuristic approach. % For initial estimates $\hat{\mat{\beta}}_k^{(0)}$ we


First, we compute moment based modewise marginal covariance estimates $\widehat{\mat{\Sigma}}_k(\ten{X})$ and $\widehat{\mat{\Sigma}}_k(\ten{F}_Y)$ as


In practice, we assume we have a random sample of $n$ observations $(\ten{X}_i, \ten{F}_{y_i})$ from the joint distribution. We start the estimation process by demeaning the data. Then, only the reduction matrices $\mat{\beta}_k$ and the scatter matrices $\mat{\Omega}_k$ need to be estimated. To solve the optimization problem \eqref{eq:mle}, with $\overline{\ten{\eta}} = 0$ we initialize the parameters using a simple heuristic approach. First, we compute moment based modewise marginal covariance estimates $\widehat{\mat{\Sigma}}_k(\ten{X})$ and $\widehat{\mat{\Sigma}}_k(\ten{F}_Y)$ as


\begin{displaymath}


\widehat{\mat{\Sigma}}_k(\ten{X}) = \frac{1}{n}\sum_{i = 1}^{n} (\ten{X}_i)_{(k)}\t{(\ten{X}_i)_{(k)}}, \qquad


\widehat{\mat{\Sigma}}_k(\ten{F}_Y) = \frac{1}{n}\sum_{i = 1}^{n} (\ten{F}_{y_i})_{(k)}\t{(\ten{F}_{y_i})_{(k)}}.



@ 601,7 +561,6 @@ Given $\hat{\mat{\beta}}_1, \ldots, \hat{\mat{\beta}}_r, \hat{\mat{\Omega}}_1,


\biggr)


\hat{\mat{\Omega}}_j.


\end{equation}


%For the scatter matrices $\mat{\Omega}_j$, we need to fudge a bit.


Equating the partial gradient of the $j$th scatter matrix $\mat{\Omega}_j$ in \cref{thm:grad} to zero ( $\nabla_{\mat{\Omega}_j}l_n = 0$) gives a quadratic matrix equation due to the dependence of $\ten{\mu}_y$ on $\mat{\Omega}_j$. In practice though, it is faster, more stable, and equally accurate to use modewise covariance estimates via the residuals


\begin{displaymath}


\hat{\ten{R}}_i = \ten{X}_i  \hat{\ten{\mu}}_{y_i} = \ten{X}_i  \ten{F}_{y_i}\mlm_{k = 1}^{r}\hat{\mat{\Omega}}_k^{1}\hat{\mat{\beta}}_k.



@ 619,7 +578,7 @@ so that


\tilde{s} = \biggl(\Bigl(\prod_{k = 1}^{r}\tr{\tilde{\mat{\Sigma}}_k}\Bigr)^{1}\frac{1}{n}\sum_{i = 1}^n \langle \hat{\ten{R}}_i, \hat{\ten{R}}_i \rangle\biggr)^{1 / r}


\end{displaymath}


resulting in the estimates $\hat{\mat{\Omega}}_j = (\tilde{s}\tilde{\mat{\Sigma}}_j)^{1}$.


Estimation is then performed by updating the estimates $\hat{\mat{\beta}}_j$ via \eqref{eq:tensor_normal_beta_solution} for $j = 1, \ldots, r$, and then recompute the $\hat{\mat{\Omega}}_j$ estimates simultaneously keeping the $\hat{\mat{\beta}}_j$'s fixed. This procedure is repeated until convergence. % Convergence is very fast, experiments showed that convergence occures usualy in less than $10$ iterations.


Estimation is then performed by updating the estimates $\hat{\mat{\beta}}_j$ via \eqref{eq:tensor_normal_beta_solution} for $j = 1, \ldots, r$, and then recompute the $\hat{\mat{\Omega}}_j$ estimates simultaneously keeping the $\hat{\mat{\beta}}_j$'s fixed. This procedure is repeated until convergence.




A technical detail for numerical stability is to ensure that the scaled values $\tilde{s}\tilde{\mat{\Sigma}}_j$, assumed to be symmetric and positive definite, are well conditioned. Thus, we estimate the condition number of $\tilde{s}\tilde{\mat{\Sigma}}_j$ before computing the inverse. In case of illconditioning, we use the regularized $\hat{\mat{\Omega}}_j = (\tilde{s}\tilde{\mat{\Sigma}}_j + 0.2 \lambda_{1}(\tilde{s}\tilde{\mat{\Sigma}}_j)\mat{I}_{p_j})^{1}$ instead, where $\lambda_{1}(\tilde{s}\tilde{\mat{\Sigma}}_j)$ is the first (maximum) eigenvalue. Experiments showed that this regularization is usually only required in the first few iterations.





@ 631,11 +590,11 @@ similar “flipflop” approach by iteratively updating the $\mat{\beta}_k$'s a






%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsection{Ising Model}\label{sec:ising_estimation}


\subsection{MultiLinear Ising Model}\label{sec:ising_estimation}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


The Ising\footnote{Also known as the \emph{LenzIsing} model as the physical assumptions of the model where developed by both Lenz and Ising \parencite{Niss2005}. Ising gave a closed form solution for the 1dimensional lattice, that is, a linear chain \parencite{Ising1925}.} model \parencite{Lenz1920,Ising1925,Niss2005} is a mathematical model originating in statistical physics to study ferromagnetism in a thermodynamic setting. It describes magentic dipoles (atomic ``spins'') which can take two states ($\pm 1$) while allowing twoway interactions between direct neighbours on a lattice, a discrete grid. The model assumes all elementary magnets to be the same, which translates to all having the same coupling strength (twoway interactions) governed by a single parameter relating to the temperature of the system. Nowadays, the Ising model, in its general form, allows for different coupling strength for every (symmetric) interaction as well as an external magnetic field acting on every magnetic dipole separately. A review is given by \textcite{NguyenEtAl2017}.


% The Ising\footnote{Also known as the \emph{LenzIsing} model as the physical assumptions of the model where developed by both Lenz and Ising \parencite{Niss2005}. Ising gave a closed form solution for the 1dimensional lattice, that is, a linear chain \parencite{Ising1925}.} model \parencite{Lenz1920,Ising1925,Niss2005} is a mathematical model originating in statistical physics to study ferromagnetism in a thermodynamic setting. It describes magentic dipoles (atomic ``spins'') which can take two states ($\pm 1$) while allowing twoway interactions between direct neighbours on a lattice, a discrete grid. The model assumes all elementary magnets to be the same, which translates to all having the same coupling strength (twoway interactions) governed by a single parameter relating to the temperature of the system. Nowadays, the Ising model, in its general form, allows for different coupling strength for every (symmetric) interaction as well as an external magnetic field acting on every magnetic dipole separately. A review is given by \textcite{NguyenEtAl2017}.




In statistics, the Ising model is used to model multivariate binary data. That is, the states are ${0, 1}$ instead of $\pm 1$. It is related to a multitude of other models; \emph{Graphical Models} and \emph{Markov Random Fields} to describe conditional dependence \parencite{Lauritzen1996,WainwrightJordan2008,LauritzenRichardson2002}, \emph{Potts models} \parencite{Besag1974,ChakrabortyEtAl2022} which generalize the Ising model to multiple states, the \emph{multivariate Bernoulli distribution} \parencite{Whittaker1990,JohnsonEtAl1997,DaiDingWahba2013} considering also interactions (treeway and higher), to give the most prominent.


% In statistics, the Ising model is used to model multivariate binary data. That is, the states are ${0, 1}$ instead of $\pm 1$. It is related to a multitude of other models; \emph{Graphical Models} and \emph{Markov Random Fields} to describe conditional dependence \parencite{Lauritzen1996,WainwrightJordan2008,LauritzenRichardson2002}, \emph{Potts models} \parencite{Besag1974,ChakrabortyEtAl2022} which generalize the Ising model to multiple states, the \emph{multivariate Bernoulli distribution} \parencite{Whittaker1990,JohnsonEtAl1997,DaiDingWahba2013} considering also interactions (treeway and higher), to give the most prominent.




The $p$dimensional Ising model is a discrete probability distribution on the set of $p$dimensional binary vectors $\mat{x}\in\{0, 1\}^p$ with probability mass function (pmf) given by


\begin{displaymath}



@ 652,34 +611,30 @@ Abusing notation, we let $\mat{\gamma}_{j l}$ denote the element of $\mat{\gamma


\end{align}


Conditional Ising models, incorporating the information of covariates $Y$ into the model, were considered by \textcite{ChengEtAl2014,BuraEtAl2022}. The direct way is to parameterize $\mat{\gamma} = \mat{\gamma}_y$ by the covariate $Y = y$ to model a conditional distribution $P_{\mat{\gamma}_y}(\mat{x}\mid Y = y)$.




We extend the conditional pmf by allowing the binary variables to be tensor valued; that is, we set $\mat{x} = \vec{\ten{X}}$, with dimension $p = \prod_{k = 1}^{r}p_k$ for $\ten{X}\in\{ 0, 1 \}^{p_1\times\cdots\times p_r}$. The tensor structure of $\ten{X}$ is accommodated by assuming Kronecker product constraints to the parameter vector $\mat{\gamma}_y$ in a similar fashion as for the tensor normal model. This means that we compare the pmf $P_{\mat{\gamma}_y}(\vec{\ten{X}}  Y = y)$ with the quadratic exponential family \eqref{eq:quadraticexpfam} with the natural parameters modeled by \eqref{eq:eta1} and \eqref{eq:eta2}. A detail to be considered is that the diagonal of $(\vec{\ten{X}})\t{(\vec{\ten{X}})}$ is equal to $\vec{\ten{X}}$, which results in the GMLM being expressed as


We extend the conditional pmf by allowing the binary variables to be tensor valued; that is, we set $\mat{x} = \vec{\ten{X}}$, with dimension $p = \prod_{k = 1}^{r}p_k$ for $\ten{X}\in\{ 0, 1 \}^{p_1\times\cdots\times p_r}$. The tensor structure of $\ten{X}$ is accommodated by assuming Kronecker product constraints to the parameter vector $\mat{\gamma}_y$ in a similar fashion as for the multilinear normal model. This means that we compare the pmf $P_{\mat{\gamma}_y}(\vec{\ten{X}}  Y = y)$ with the quadratic exponential family \eqref{eq:quadraticexpfam} with the natural parameters modeled by \eqref{eq:eta1} and \eqref{eq:eta2}. A detail to be considered is that the diagonal of $(\vec{\ten{X}})\t{(\vec{\ten{X}})}$ is equal to $\vec{\ten{X}}$, which results in the GMLM being expressed as


\begin{align}


P_{\mat{\gamma}_y}(\ten{X} \mid Y = y)


&= p_0(\mat{\gamma}_y)\exp(\t{\vech((\vec{\ten{X}})\t{(\vec{\ten{X}})})}\mat{\gamma}_y) \nonumber \\


&= p_0(\mat{\gamma}_y)\exp\Bigl(\Bigl\langle \ten{X}, \ten{F}_y\mlm_{k = 1}^{r}\mat{\beta}_k \Bigr\rangle + \Bigl\langle\ten{X}\mlm_{k = 1}^{r}\mat{\Omega}_k, \ten{X}\Bigr\rangle\Bigr)\label{eq:isingcondprob}


\end{align}


where we set $\overline{\ten{\eta}} = 0$ and $\mat{T}_2$ to the identity. This imposes an additional constraint to the model, the reason is that the diagonal elements of $\mat{\Omega} = \bigkron_{k = r}^{1}\mat{\Omega}_k$ take the role of $\overline{\ten{\eta}}$, although not fully. Having the diagonal of $\mat{\Omega}$ and $\overline{\ten{\eta}}$ handling the self interaction effects might lead to interference in the optimization routine. Another approach would be to use the $\mat{T}_2$ matrix to set the corresponding diagonal elements of $\mat{\Omega}$ to zero and let $\overline{\ten{\eta}}$ handle the self interaction effect. All of these approaches, namely setting $\overline{\ten{\eta}} = 0$, keeping $\overline{\ten{\eta}}$ or using $\mat{T}_2$, are theoretically solid and compatible with \cref{thm:grad,thm:parammanifold,thm:asymptoticnormalitygmlm}, assuming all axis dimensions $p_k$ are nondegenerate, that is $p_k > 1$ for all $k = 1, \ldots, r$. Regardless, under our modeling choice, the relation between the natural parameters $\mat{\gamma}_y$ of the conditional Ising model and the GMLM parameters $\mat{\beta}_k$ and $\mat{\Omega}_k$ is


where we set $\overline{\ten{\eta}} = 0$ and $\mat{T}_2$ to the identity. This imposes an additional constraint to the model, the reason is that the diagonal elements of $\mat{\Omega} = \bigkron_{k = r}^{1}\mat{\Omega}_k$ take the role of $\overline{\ten{\eta}}$, although not fully. Having the diagonal of $\mat{\Omega}$ and $\overline{\ten{\eta}}$ handling the selfinteraction effects might lead to interference in the optimization routine. Another approach would be to use the $\mat{T}_2$ matrix to set the corresponding diagonal elements of $\mat{\Omega}$ to zero and let $\overline{\ten{\eta}}$ handle the selfinteraction effect. All of these approaches, namely setting $\overline{\ten{\eta}} = 0$, keeping $\overline{\ten{\eta}}$ or using $\mat{T}_2$, are theoretically solid and compatible with \cref{thm:grad,thm:parammanifold,thm:asymptoticnormalitygmlm}, assuming all axis dimensions $p_k$ are nondegenerate, that is $p_k > 1$ for all $k = 1, \ldots, r$. Regardless, under our modeling choice, the relation between the natural parameters $\mat{\gamma}_y$ of the conditional Ising model and the GMLM parameters $\mat{\beta}_k$ and $\mat{\Omega}_k$ is


\begin{equation}\label{eq:isingnaturalparams}


% \t{\pinv{\mat{D}_p}}\mat{\gamma}_y


% = \vec(\mat{\Omega} + \diag(\mat{B}\vec{\ten{F}_y}))


% = \vec\Biggl(\bigkron_{k = r}^{1}\mat{\Omega}_k + \diag\biggl(\vec\Bigl(\ten{F}_y\mlm_{k = 1}^{r}\mat{\beta}_k\Bigr)\biggr)\Biggr).


\mat{\gamma}_y


= \t{\mat{D}_p}\vec(\mat{\Omega} + \diag(\mat{B}\vec{\ten{F}_y}))


= \t{\mat{D}_p}\vec\Biggl(\bigkron_{k = r}^{1}\mat{\Omega}_k + \diag\biggl(\vec\Bigl(\ten{F}_y\mlm_{k = 1}^{r}\mat{\beta}_k\Bigr)\biggr)\Biggr).


\end{equation}


In contract to the tensor normal GMLM, the matrices $\mat{\Omega}_k$ are only required to be symmetric. More specifically, we require $\mat{\Omega}_k$, for $k = 1, \ldots, r$, to be elements of an embedded submanifold of $\SymMat{p_k}$ (see: \cref{sec:kronmanifolds,sec:matrixmanifolds}). The mode wise reduction matrices $\mat{\beta}_k$ are elements of an embedded submanifold of $\mathbb{R}^{p_k\times q_k}$. Common choices are listed in \cref{sec:matrixmanifolds}.


In contrast to the multilinear normal GMLM, the matrices $\mat{\Omega}_k$ are only required to be symmetric. More specifically, we require $\mat{\Omega}_k$, for $k = 1, \ldots, r$, to be elements of an embedded submanifold of $\SymMat{p_k}$ (see: \cref{sec:kronmanifolds,sec:matrixmanifolds}). The mode wise reduction matrices $\mat{\beta}_k$ are elements of an embedded submanifold of $\mathbb{R}^{p_k\times q_k}$. Common choices are listed in \cref{sec:matrixmanifolds}.




To solve the optimization problem \eqref{eq:mle}, given a data set $(\ten{X}_i, y_i)$, $i = 1, \ldots, n$, we use a variation of gradient descent.




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsubsection{Initial Values}




The first step is to get reasonable starting values. Experiments showed that a good starting value of $\mat{\beta}_k$ is to use the tensor normal estimates from \cref{sec:tensornormalestimation} for $k = 1, \ldots, r$, considering $\ten{X}_i$ as continuous. For initial values of $\mat{\Omega}_k$, a different approach is required. Setting everything to the uninformed initial value, that is $\mat{\Omega}_k = \mat{0}$ as this corresponds to the conditional log odds to be $1:1$ for every component and pairwise interaction. This is not possible, since $\mat{0}$ is a stationary point of the loglikelihood. This is directly observed by considering the partial gradients of the loglikelihood in \cref{thm:grad}. Instead, we use a crude heuristic which threads every mode seperately and ignores any relation to the covariates. It is computationaly cheap and better than any of the alternatives we considered. For every $k = 1, \ldots, r$, let the $k$th mode second moment estimate be


The first step is to get reasonable starting values. Experiments showed that a good starting value of $\mat{\beta}_k$ is to use the multilinear normal estimates from \cref{sec:tensornormalestimation} for $k = 1, \ldots, r$, considering $\ten{X}_i$ as continuous. For initial values of $\mat{\Omega}_k$, a different approach is required. Setting everything to the uninformed initial value, that is $\mat{\Omega}_k = \mat{0}$ as this corresponds to the conditional log odds to be $1:1$ for every component and pairwise interaction. This is not possible, since $\mat{0}$ is a stationary point of the loglikelihood. This is directly observed by considering the partial gradients of the loglikelihood in \cref{thm:grad}. Instead, we use a crude heuristic that threads every mode separately and ignores any relation to the covariates. It is computationally cheap and better than any of the alternatives we considered. For every $k = 1, \ldots, r$, let the $k$th mode second moment estimate be


\begin{equation}\label{eq:isingmodemoments}


\hat{\mat{M}}_{2(k)} = \frac{p_k}{n p}\sum_{i = 1}^n (\ten{X}_i)_{(k)}\t{(\ten{X}_i)_{(k)}}


\end{equation}


which contains the $k$th mode first moment estimate in its diagonal $\hat{\mat{M}}_{1(k)} = \diag\hat{\mat{M}}_{2(k)}$. Considering every column of the matricized observation $(\ten{X}_i)_{(k)}$ as a $p_k$ dimensional observation. The number of those artificially generated observations is $n \prod_{j\neq k}p_j$. Let $Z_k$ denote the random variable those artificial observations are realization of. Then, we can interpret the elements $(\hat{\mat{M}}_{1(k)})_{j}$ as the estimates of the probability $P((Z_k)_j = 1)$, that is the marginal probability of the $j$th element of $Z_k$ being $1$. Similar, for $l \neq j$ we have $(\hat{\mat{M}}_{2(k)})_{j l}$ estimating $P((Z_k)_j = 1, (Z_k)_l = 1)$, the marginal probability of twoway interactions. % Without any regard of accuracy ...


Now, we set the diagonal elements of $\mat{\Omega}_k$ to zero. For the off diagonal elements of $\mat{\Omega}_k$, we equate the conditional probabilities $P((Z_k)_j = 1 \mid (Z_k)_{j} = \mat{0})$ and $P((Z_k)_j = 1, (Z_k)_l = 1\mid (Z_k)_{j, l} = \mat{0})$ with the marginal probability estimates $(\hat{\mat{M}}_{1(k)})_{j}$ and $(\hat{\mat{M}}_{2(k)})_{j l}$, respectively. Applying \eqref{eq:isingtwowaylogodds} then gives the initial componentwise estimates $\hat{\mat{\Omega}}_k^{(0)}$,


which contains the $k$th mode first moment estimate in its diagonal $\hat{\mat{M}}_{1(k)} = \diag\hat{\mat{M}}_{2(k)}$. Considering every column of the matricized observation $(\ten{X}_i)_{(k)}$ as a $p_k$ dimensional observation. The number of those artificially generated observations is $n \prod_{j\neq k}p_j$. Let $Z_k$ denote the random variable those artificial observations are realization of. Then, we can interpret the elements $(\hat{\mat{M}}_{1(k)})_{j}$ as the estimates of the probability $P((Z_k)_j = 1)$, that is the marginal probability of the $j$th element of $Z_k$ being $1$. Similar, for $l \neq j$ we have $(\hat{\mat{M}}_{2(k)})_{j l}$ estimating $P((Z_k)_j = 1, (Z_k)_l = 1)$, the marginal probability of twoway interactions. Now, we set the diagonal elements of $\mat{\Omega}_k$ to zero. For the off diagonal elements of $\mat{\Omega}_k$, we equate the conditional probabilities $P((Z_k)_j = 1 \mid (Z_k)_{j} = \mat{0})$ and $P((Z_k)_j = 1, (Z_k)_l = 1\mid (Z_k)_{j, l} = \mat{0})$ with the marginal probability estimates $(\hat{\mat{M}}_{1(k)})_{j}$ and $(\hat{\mat{M}}_{2(k)})_{j l}$, respectively. Applying \eqref{eq:isingtwowaylogodds} then gives the initial componentwise estimates $\hat{\mat{\Omega}}_k^{(0)}$,


\begin{equation}\label{eq:isinginitOmegas}


(\hat{\mat{\Omega}}_k^{(0)})_{j j} = 0,


\qquad



@ 696,36 +651,36 @@ The natural parameter $\mat{\gamma}_y$ is evaluated via \eqref{eq:isingnatural


\begin{displaymath}


\mat{\theta}^{(I + 1)} = \mat{\theta}^{(I)} + \lambda\nabla_{\mat{\theta}} l_n(\mat{\theta})\bigr_{\mat{\theta} = \mat{\theta}^{(I)}},


\end{displaymath}


which is iterated till convergence. In practice, iteration is performed until either a maximum number of iterations is exhausted and/or some break condition is satisfied. A proper choice of the learning rate is needed as a large learning rate $\lambda$ may cause instability, while a very low learning rate requires an enormous amount of iterations. Generically, there are two approaches to avoid the need to determine a proper learning rate. First, \emph{line search methods} determine an appropriate step size for every iteration. This works well if the evaluation of the object function (the loglikelihood) is cheap. This is not the case in our setting, see \cref{sec:isingbiggerdim}. The second approach is an \emph{adaptive learning rate}, where one tracks specific statistics while optimizing and dynamically adapting the learning rate via welltested heuristics using the gathered knowledge from past iterations. We opted to use an adaptive learning rate approach, which not only removes the need to determine an appropriate learning rate, but also accelerates learning.


which is iterated till convergence. In practice, iteration is performed until either a maximum number of iterations is exhausted and/or some break condition is satisfied. A proper choice of the learning rate is needed as a large learning rate $\lambda$ may cause instability, while a very low learning rate requires an enormous amount of iterations. Generically, there are two approaches to avoid the need to determine a proper learning rate. First, \emph{line search methods} determine an appropriate step size for every iteration. This works well if the evaluation of the object function (the loglikelihood) is cheap. This is not the case in our setting, see \cref{sec:isingbiggerdim}. The second approach is an \emph{adaptive learning rate}, where one tracks specific statistics while optimizing and dynamically adapting the learning rate via welltested heuristics using the gathered knowledge from past iterations. We opted to use an adaptive learning rate approach, which not only removes the need to determine an appropriate learning rate but also accelerates learning.




Our method of choice is \emph{root mean squared propagation} (RMSprop) \parencite{Hinton2012}. This is a well known method in machine learning for training neural networks. It is a variation of gradient descent with a per scalar parameter adaptive learning rate. It tracks a moving average of the element wise squared gradient $\mat{g}_2^{(I)}$, which is then used to scale (element wise) the gradient in the update rule. See \textcite{Hinton2012,GoodfellowEtAl2016} among others. The update rule using RMSprop for maximization\footnote{Instead of the more common minimization, therefore $+$ in the update of $\mat{\theta}$.} is


Our method of choice is \emph{root mean squared propagation} (RMSprop) \parencite{Hinton2012}. This is a wellknown method in machine learning for training neural networks. It is a variation of gradient descent with a per scalar parameter adaptive learning rate. It tracks a moving average of the elementwise squared gradient $\mat{g}_2^{(I)}$, which is then used to scale (elementwise) the gradient in the update rule. See \textcite{Hinton2012,GoodfellowEtAl2016} among others. The update rule using RMSprop for maximization\footnote{Instead of the more common minimization, therefore $+$ in the update of $\mat{\theta}$.} is


\begin{align*}


\mat{g}_2^{(I + 1)} &= \nu \mat{g}_2^{(I)} + (1  \nu)\nabla l_n(\mat{\theta}^{(I)})\odot\nabla l_n(\mat{\theta}^{(I)}), \\


\mat{\theta}^{(I + 1)} &= \mat{\theta}^{(I)} + \frac{\lambda}{\sqrt{\mat{g}_2^{(I + 1)}} + \epsilon}\odot\nabla l_n(\mat{\theta}^{(I)}).


\end{align*}


The parameters $\nu = 0.9$, $\lambda = 10^{3}$ and $\epsilon\approx 1.49\cdot 10^{8}$ are fixed. The initial value of $\mat{g}_2^{(0)} = \mat{0}$, the symbol $\odot$ denotes the Hadamard product, or element wise multiplication. The division and square root operation are performed element wise as well. According to our experiments, RMSprop requires iterations in the range of $50$ till $1000$ till convergence while gradient ascent with a learning rate of $10^{3}$ is in the range of $1000$ till $10000$.


The parameters $\nu = 0.9$, $\lambda = 10^{3}$ and $\epsilon\approx 1.49\cdot 10^{8}$ are fixed. The initial value of $\mat{g}_2^{(0)} = \mat{0}$, the symbol $\odot$ denotes the Hadamard product, or element wise multiplication. The division and square root operations are performed elementwise as well. According to our experiments, RMSprop requires iterations in the range of $50$ till $1000$ till convergence while gradient ascent with a learning rate of $10^{3}$ is in the range of $1000$ till $10000$.




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsubsection{Small Data Sets}\label{sec:isingsmalldatasets}


In case of a finite number of observations, specifically in data sets with a small number of observations $n$, the situation where one components is always ether zero or one can occur. Its also possible to observe two exclusive components. This situation of a ``degenerate'' data set needs to be safeguarded against in practice. Working with parameters on a logscale, this gives estimates of $\pm\infty$. This is outside of the parameter space and breaks our optimization algorithm.


In the case of a finite number of observations, specifically in data sets with a small number of observations $n$, the situation where one components is always ether zero or one can occur. It is also possible to observe two exclusive components. This situation of a ``degenerate'' data set needs to be safeguarded against in practice. Working with parameters on a logscale, this gives estimates of $\pm\infty$. This is outside of the parameter space and breaks our optimization algorithm.




The first situation where this needs to be addressed is in \eqref{eq:isinginitOmegas}, where we set initial estimates for $\mat{\Omega}_k$. To avoid divition by zero as well as evaluating the log of zero, we addapt \eqref{eq:isingmodemoments}, the mode wise moment estimates $\hat{\mat{M}}_{2(k)}$. A simple method is to replace the ``degenerate'' components, that are entries with value $0$ or $1$, with the smallest positive estimate of exactly one occurrence $p_k / n p$, or all but one occurrence $1  p_k / n p$, respectively.


The first situation where this needs to be addressed is in \eqref{eq:isinginitOmegas}, where we set initial estimates for $\mat{\Omega}_k$. To avoid division by zero as well as evaluating the log of zero, we adapt \eqref{eq:isingmodemoments}, the modewise moment estimates $\hat{\mat{M}}_{2(k)}$. A simple method is to replace the ``degenerate'' components, that are entries with value $0$ or $1$, with the smallest positive estimate of exactly one occurrence $p_k / n p$, or all but one occurrence $1  p_k / n p$, respectively.




The same problem is present in gradient optimization. Therefore, before starting the optimization, we detect degenerate combinations. We compute upper and lower bounds for the ``degenerate'' element in the Kronecker product $\hat{\mat{\Omega}} = \bigkron_{k = r}^{1}\hat{\mat{\Omega}}_k$. After every gradient update, we check if any of the ``degenerate'' elements fall outside of the bounds. In that case, we adjust all the elements of the Kronecker component estimates $\hat{\mat{\Omega}}_k$, corresponding to the ``degenerate'' element of their Kronecker product, to fall inside the precomputed bounds. While doing so, we try to alter every component as little as possible to ensure that the nondegenerate elements in $\hat{\mat{\Omega}}$, effected by this change due to its Kronecker structure, are altered as little as possible. The exact details are technically cumbersome while providing little insight.


The same problem is present in gradient optimization. Therefore, before starting the optimization, we detect degenerate combinations. We compute upper and lower bounds for the ``degenerate'' element in the Kronecker product $\hat{\mat{\Omega}} = \bigkron_{k = r}^{1}\hat{\mat{\Omega}}_k$. After every gradient update, we check if any of the ``degenerate'' elements fall outside of the bounds. In that case, we adjust all the elements of the Kronecker component estimates $\hat{\mat{\Omega}}_k$, corresponding to the ``degenerate'' element of their Kronecker product, to fall inside the precomputed bounds. While doing so, we try to alter every component as little as possible to ensure that the nondegenerate elements in $\hat{\mat{\Omega}}$, affected by this change due to its Kronecker structure, are altered as little as possible. The exact details are technically cumbersome while providing little insight.








%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsubsection{Slightly Bigger Dimensions}\label{sec:isingbiggerdim}


A big challenge for the Ising model is its high computational complexity as it involves summing over all binary vectors of length $p = \prod_{k = 1}^{r}p_k$ in the partition function \eqref{eq:isingpartitionfunction}. Computing the partition function exactly requires to sum all $2^p$ binary vectors. For small dimensions, say $p\approx 10$, this is easily computed. Increasing the dimension beyond $20$ becomes extremely expensive while it is %absolutely


impossible for dimension bigger than $30$. Trying to avoid the evaluation of the loglikelihood and only computing its partial gradients via \cref{thm:grad} does not resolve the issue. The gradients require the inverse link, in other words the second moment \eqref{eq:isingm2}, where, if dropping the scaling factor $p_0$, still involves to sum $2^p$ summands. Basically, with our model, this means that the optimization of the Ising model using exactly computed gradients is impossible for moderately sized problems.


A big challenge for the Ising model is its high computational complexity as it involves summing over all binary vectors of length $p = \prod_{k = 1}^{r}p_k$ in the partition function \eqref{eq:isingpartitionfunction}. Computing the partition function exactly requires to sum all $2^p$ binary vectors. For small dimensions, say $p\approx 10$, this is easily computed. Increasing the dimension beyond $20$ becomes extremely expensive while it is %absolutely


impossible for a dimension bigger than $30$. Trying to avoid the evaluation of the loglikelihood and only computing its partial gradients via \cref{thm:grad} does not resolve the issue. The gradients require the inverse link, that is the second moment \eqref{eq:isingm2}, where, if dropping the scaling factor $p_0$, still involves to sum $2^p$ summands. Basically, with our model, this means that the optimization of the Ising model using exactly computed gradients is impossible for moderately sized problems.




For estimation of dimensions $p$ bigger than $20$, we use a MonteCarlo method to estimate the second moment \eqref{eq:isingm2}, required to compute the partial gradients of the loglikelihood. Specifically, we use a GibbsSampler to sample from the conditional distribution and approximate the second moment in an importance sampling framework. This can be implemented quite efficiently while the estimation accuracy for the second moment is evaluated experimentally which seems to be very reliable. Simultaneously, we use the same approach to estimate the partition function. This though, is in comparison inaccurate, and may only be used to get a rough idea of the loglikelihood. Regardless, for our method, we only need the gradient for optimization where appropriate break conditions, not based on the likelihood, lead to a working method for MLE estimation.


For the estimation of dimensions $p$ bigger than $20$, we use a MonteCarlo method to estimate the second moment \eqref{eq:isingm2}, required to compute the partial gradients of the loglikelihood. Specifically, we use a GibbsSampler to sample from the conditional distribution and approximate the second moment in an importance sampling framework. This can be implemented quite efficiently while the estimation accuracy for the second moment is evaluated experimentally which seems to be very reliable. Simultaneously, we use the same approach to estimate the partition function. This though, is in comparison inaccurate, and may only be used to get a rough idea of the loglikelihood. Regardless, for our method, we only need the gradient for optimization where appropriate break conditions, not based on the likelihood, lead to a working method for MLE estimation.




\begin{figure}


\centering


\includegraphics[]{plots/simisingperftm2.pdf}


\caption{\label{fig:isingm2perft}Performance test for computing/estimating the second moment of the Ising model of dimension $p$ using ether the exact method or a MonteCarlo (MC) simulation.}


\caption{\label{fig:isingm2perft}Performance test for computing/estimating the second moment of the Ising model of dimension $p$ using either the exact method or a MonteCarlo (MC) simulation.}


\end{figure}







@ 735,16 +690,16 @@ For estimation of dimensions $p$ bigger than $20$, we use a MonteCarlo method t




\cref{thm:sdr} identifies the sufficient reduction for the regression of $Y$ on $\ten{X}$ in the population. Any estimation of the sufficient reduction requires application of some optimality criterion. As we operate within the framework of the exponential family, we opted for maximum likelihood estimation (MLE). For the unconstrained problem, where the parameters are simply $\mat{B}$ and $\mat{\Omega}$ in \eqref{eq:eta1manifold}, maximizing the likelihood of $\ten{X} \mid Y$ is straightforward and yields welldefined MLEs of both parameters. Our setting, though, requires the constrained optimization of the $\ten{X} \mid Y$ likelihood subject to $\mat{B} = \bigotimes_{j = r}^{1}\mat{\beta}_j$ and $\mat{\Omega}=\bigkron_{j = r}^{1}\mat{\Omega}_j$. \Cref{thm:kronmanifolds,thm:parammanifold} provide the setting for which the MLE of the constrained parameter $\mat{\theta}$ is welldefined, which in turn leads to the derivation of its asymptotic normality.




The main problem in obtaining asymptotic results for the MLE of the constrained parameter $\mat{\theta} = (\overline{\ten{\eta}}, \vec\mat{B}, \vech\mat{\Omega})$ stems from the nature of the constraint. We assumed that $\mat{B} = \bigkron_{k = r}^{1}\mat{\beta}_k$, where the parameter $\mat{B}$ is identifiable. This means that different values of $\mat{B}$ lead to different densities $f_{\mat{\theta}}(\ten{X}\mid Y = y)$, a basic property needed to ensure consistency of parameter estimates, which in turn is needed for asymptotic normality. On the other hand, the components $\mat{\beta}_j$, $j = 1, \ldots, r$, are \emph{not} identifiable, which is a direct consequence of the equality $\mat{\beta}_2\otimes\mat{\beta}_1 = (c\mat{\beta}_2)\otimes (c^{1}\mat{\beta}_1)$ for every $c\neq 0$. This is the reason we formulated $\Theta$ as a constrained parameter space instead of parameterizing the densities of $\ten{X}\mid Y$ with respect to the components $\mat{\beta}_1, \ldots, \mat{\beta}_r$. The same is true for $\mat{\Omega} = \bigkron_{k = r}^{1}\mat{\Omega}_k$.


The main problem in obtaining asymptotic results for the MLE of the constrained parameter $\mat{\theta} = (\overline{\ten{\eta}}, \vec\mat{B}, \vech\mat{\Omega})$ stems from the nature of the constraint. We assumed that $\mat{B} = \bigkron_{k = r}^{1}\mat{\beta}_k$, where the parameter $\mat{B}$ is identifiable. This means that different values of $\mat{B}$ lead to different densities $f_{\mat{\theta}}(\ten{X}\mid Y = y)$, a basic property needed to ensure consistency of parameter estimates, which in turn is needed for asymptotic normality. On the other hand, the components $\mat{\beta}_j$, $j = 1, \ldots, r$, are \emph{not} identifiable, which is a direct consequence of the equality $\mat{\beta}_2\otimes\mat{\beta}_1 = (c\mat{\beta}_2)\otimes (c^{1}\mat{\beta}_1)$ for every $c\neq 0$. This is the reason we considered $\Theta$ as a constrained parameter space instead of parameterizing the densities of $\ten{X}\mid Y$ with $\mat{\beta}_1, \ldots, \mat{\beta}_r$. The same is true for $\mat{\Omega} = \bigkron_{k = r}^{1}\mat{\Omega}_k$.




In addition to identifiable parameters, asymptotic normality obtained in \cref{thm:asymptoticnormalitygmlm} requires differentiation. Therefore, the space itself needs to admit defining differentiation, which is usually a vector space. This is too strong an assumption for our purposes. To weaken the vector space assumption we consider \emph{smooth manifolds}. The latter are spaces which look like Euclidean spaces locally and allow the notion of differentiation. The more general \emph{topological} manifolds are too weak for differentiation. To make matters worse, a smooth manifold only allows first derivatives. Without going into details, the solution is a \emph{Riemannian manifold}. Similar to an abstract \emph{smooth manifold}, Riemannian manifolds are detached from our usual intuition as well as complicated to handle in an already complicated setting. This is where an \emph{embedded (sub)manifold} comes to the rescue. Simply speaking, an embedded manifold is a manifold which is a subset of a manifold from which it inherits its properties. If a manifold is embedded in a Euclidean space, almost all the complication of the abstract manifold theory simplifies drastically. Moreover, since a Euclidean space is itself a Riemannian manifold, we inherit the means for higher derivatives. Finally, smooth embedded submanifold structure for the parameter space maintains consistency with existing approaches and results for parameter sets with linear subspace structure. These reasons justify the constraint that the parameter space $\Theta$ be an \emph{smooth embedded submanifold} in an open subset $\Xi$ of a Euclidean space.


In addition to identifiable parameters, the asymptotic normality obtained in \cref{thm:asymptoticnormalitygmlm} requires differentiation. Therefore, the space itself needs to admit defining differentiation, which is usually a vector space. This is too strong an assumption for our purposes. To weaken the vector space assumption we consider \emph{smooth manifolds}. The latter are spaces that look like Euclidean spaces locally and allow the notion of differentiation. The more general \emph{topological} manifolds are too weak for differentiation. To make matters worse, a smooth manifold only allows first derivatives. Without going into details, the solution is a \emph{Riemannian manifold}. Similar to an abstract \emph{smooth manifold}, Riemannian manifolds are detached from our usual intuition as well as complicated to handle in an already complicated setting. This is where an \emph{embedded (sub)manifold} comes to the rescue. Simply speaking, an embedded manifold is a manifold which is a subset of a manifold from which it inherits its properties. If a manifold is embedded in a Euclidean space, almost all the complications of the abstract manifold theory simplify drastically. Moreover, since a Euclidean space is itself a Riemannian manifold, we inherit the means for higher derivatives. Finally, smooth embedded submanifold structure for the parameter space maintains consistency with existing approaches and results for parameter sets with linear subspace structure. These reasons justify the constraint that the parameter space $\Theta$ be an \emph{smooth embedded submanifold} in an open subset $\Xi$ of a Euclidean space.




Now, we directly define a \emph{smooth manifold} embedded in $\mathbb{R}^p$ without any detours to the more generel theory. See for example \textcite{Lee2012,,Lee2018,AbsilEtAl2007,Kaltenbaeck2021} among others.


Now, we define a \emph{smooth manifold} embedded in $\mathbb{R}^p$ without any detours to the more general theory. See for example \textcite{Lee2012,,Lee2018,AbsilEtAl2007,Kaltenbaeck2021} among others.


\begin{definition}[Manifolds]\label{def:manifold}


A set $\manifold{A}\subseteq\mathbb{R}^p$ is an \emph{embedded smooth manifold} of dimension $d$ if for every $\mat{x}\in\manifold{A}$ there exists a smooth\footnote{Here \emph{smooth} means infinitely differentiable or $C^{\infty}$.} bicontinuous map $\varphi:U\cap\manifold{A}\to V$, called a \emph{chart}, with $\mat{x}\in U\subseteq\mathbb{R}^p$ open and $V\subseteq\mathbb{R}^d$ open.


\end{definition}




We also need the concept of a \emph{tangent space} to formulate asymptotic normality in a way which is independent of a particular coordinate representation. Intuitively, the tangent space at a point $\mat{x}\in\manifold{A}$ of the manifold $\manifold{A}$ is the hyperspace of all velocity vectors $\t{\nabla\gamma(0)}$ of any curve $\gamma:(1, 1)\to\manifold{A}$ passing through $\mat{x} = \gamma(0)$, see \cref{fig:torus}. Locally, at $\mat{x} = \gamma(0)$ with a chart $\varphi$ we can written $\gamma(t) = \varphi^{1}(\varphi(\gamma(t)))$ which gives that $\Span\t{\nabla\gamma(0)} \subseteq \Span\t{\nabla\varphi^{1}(\varphi(\mat{x}))}$. Taking the union over all smooth curves through $\mat{x}$ gives equality. The following definition leverages the simplified setup of smooth manifolds in Euclidean space.


We also need the concept of a \emph{tangent space} to formulate asymptotic normality in a way that is independent of a particular coordinate representation. Intuitively, the tangent space at a point $\mat{x}\in\manifold{A}$ of the manifold $\manifold{A}$ is the hyperspace of all velocity vectors $\t{\nabla\gamma(0)}$ of any curve $\gamma:(1, 1)\to\manifold{A}$ passing through $\mat{x} = \gamma(0)$, see \cref{fig:torus}. Locally, at $\mat{x} = \gamma(0)$ with a chart $\varphi$ we can written $\gamma(t) = \varphi^{1}(\varphi(\gamma(t)))$ which gives that $\Span\t{\nabla\gamma(0)} \subseteq \Span\t{\nabla\varphi^{1}(\varphi(\mat{x}))}$. Taking the union over all smooth curves through $\mat{x}$ gives equality. The following definition leverages the simplified setup of smooth manifolds in Euclidean space.




\begin{definition}[Tangent Space]\label{def:tangentspace}


Let $\manifold{A}\subseteq\mathbb{R}^p$ be an embedded smooth manifold and $\mat{x}\in\manifold{A}$. The \emph{tangent space} at $\mat{x}$ of $\manifold{A}$ is defined as



@ 767,7 +722,7 @@ We also need the concept of a \emph{tangent space} to formulate asymptotic norma


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsection{Kronecker Product Manifolds}\label{sec:kronmanifolds}




As a basis to ensure that the constrained parameter space $\Theta$ is a manifold, which is a requirement of \cref{thm:parammanifold}, we need \cref{thm:kronmanifolds}. Therefore, we need the notion of a \emph{spherical} set, which is a set $\manifold{A}$, on which the Frobenius norm is constant. That is, $\\,.\,\_F:\manifold{A}\to\mathbb{R}$ is constant. Forthermore, we call a scale invariant set $\manifold{A}$ a \emph{cone}, that is $\manifold{A} = \{ c \mat{A} : \mat{A}\in\manifold{A} \}$ for all $c > 0$.


As a basis to ensure that the constrained parameter space $\Theta$ is a manifold, which is a requirement of \cref{thm:parammanifold}, we need \cref{thm:kronmanifolds}. Therefore, we need the notion of a \emph{spherical} set, which is a set $\manifold{A}$, on which the Frobenius norm is constant. That is, $\\,.\,\_F:\manifold{A}\to\mathbb{R}$ is constant. Forthermore, we call a scale invariant set $\manifold{A}$ a \emph{cone}, that is $\manifold{A} = \{ c \mat{A} : \mat{A}\in\manifold{A} \}$ for all $c > 0$.




\begin{theorem}[Kronecker Product Manifolds]\label{thm:kronmanifolds}


Let $\manifold{A}\subseteq\mathbb{R}^{p_1\times q_1}\backslash\{\mat{0}\}, \manifold{B}\subseteq\mathbb{R}^{p_2\times q_2}\backslash\{\mat{0}\}$ be smooth embedded submanifolds. Assume one of the following conditions holds.



@ 797,9 +752,9 @@ As a basis to ensure that the constrained parameter space $\Theta$ is a manifold




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsection{Matrix Manifolds}\label{sec:matrixmanifolds}


A powerful side effect of \cref{thm:parammanifold} is the modeling flexibinity it provides. For example, we can perform low rank regression. Or, we may constrain twoway interactions between direct axis neighbors by using band matrices for the $\mat{\Omega}_k$'s, among others.


A powerful feature of \cref{thm:parammanifold} is the modeling flexibility it provides. For example, we can perform lowrank regression. Or, we may constrain twoway interactions between direct axis neighbors by using band matrices for the $\mat{\Omega}_k$'s, among others.




This flexibility derives from many different matrix manifolds that can be used as building blocks $\manifold{B}_k$ and $\manifold{O}_k$ of the parameter space $\Theta$ in \cref{thm:parammanifold}. A list of possible choices, among others, is given in \cref{tab:matrixmanifolds}. As long as parameters in $\Theta$ are valid paramererization of a density (or PMF) of \eqref{eq:quadraticexpfam} subject to \eqref{eq:eta1manifold} and \eqref{eq:eta2manifold}, one may choose any of the manifolds listed in \cref{tab:matrixmanifolds} which are either cones or spherical. We also included an example which is neither a sphere nor a cone. They may also be valid building blocks, but require more work as they are not directly leading to a parameter manifold by \cref{thm:parammanifold}. In case one can show the resulting parameter space $\Theta$ is an embedded manifold, the asymptotic theory of \cref{sec:asymtotics} is applicable.


This flexibility derives from many different matrix manifolds that can be used as building blocks $\manifold{B}_k$ and $\manifold{O}_k$ of the parameter space $\Theta$ in \cref{thm:parammanifold}. A list of possible choices, among others, is given in \cref{tab:matrixmanifolds}. As long as parameters in $\Theta$ are a valid parameterization of a density (or PMF) of \eqref{eq:quadraticexpfam} subject to \eqref{eq:eta1manifold} and \eqref{eq:eta2manifold}, one may choose any of the manifolds listed in \cref{tab:matrixmanifolds} which are either cones or spherical. We also included an example which is neither a sphere nor a cone. They may also be valid building blocks but require more work as they are not directly leading to a parameter manifold by \cref{thm:parammanifold}. In case one can show the resulting parameter space $\Theta$ is an embedded manifold, the asymptotic theory of \cref{sec:asymtotics} is applicable.




\begin{table}


\centering



@ 843,7 +798,7 @@ This flexibility derives from many different matrix manifolds that can be used a


\section{Statistical Properties}


\subsection{Asymptotics}\label{sec:asymtotics}




Let $Z$ be a random variable distributed according to a parameterized probability distribution with density $f_{\mat{\theta_0}}\in\{ f_{\mat{\theta}} : \mat{\theta}\in\Theta \}$ where $\Theta$ is a subset of a Euclidean space. We want to estimate the parameter ${\mat{\theta}}_0$ using $n$ i.i.d. (independent and identically distributed) copies of $Z$. We assume a known, realvalued and measurable function $z\mapsto m_{\mat{\theta}}(z)$ for every $\mat{\theta}\in\Theta$ and that ${\mat{\theta}}_0$ is the unique maximizer of the map $\mat{\theta}\mapsto M(\mat{\theta}) = \E m_{\mat{\theta}}(Z)$. For the estimation we maximize the empirical version


Let $Z$ be a random variable distributed according to a parameterized probability distribution with density $f_{\mat{\theta_0}}\in\{ f_{\mat{\theta}}: \mat{\theta}\in\Theta \}$ where $\Theta$ is a subset of a Euclidean space. We want to estimate the parameter ${\mat{\theta}}_0$ using $n$ i.i.d. (independent and identically distributed) copies of $Z$. We assume a known, realvalued and measurable function $z\mapsto m_{\mat{\theta}}(z)$ for every $\mat{\theta}\in\Theta$ and that ${\mat{\theta}}_0$ is the unique maximizer of the map $\mat{\theta}\mapsto M(\mat{\theta}) = \E m_{\mat{\theta}}(Z)$. For the estimation we maximize the empirical version


\begin{align}\label{eq:Mn}


M_n(\mat{\theta}) &= \frac{1}{n}\sum_{i = 1}^n m_{\mat{\theta}}(Z_i).


\end{align}



@ 915,7 +870,7 @@ for every nonempty compact $K\subset\Xi$. Then, there exists a strong Mest


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\section{Simulations}\label{sec:simulations}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


In this section we report simulation results for the tensor normal and the Ising model where different aspects of the GMLM model are compared against other methods. The comparison methods are Tensor Sliced Inverse Regression (TSIR) \parencite{DingCook2015}, Multiway Generalized Canonical Correlation Analysis (MGCCA) \parencite{ChenEtAl2021,GirkaEtAl2024} and the Tucker decomposition that is a higherorder form of principal component analysis (HOPCA) \textcite{KoldaBader2009}, for both continuous and binary data. For the latter, the binary values are treated as continuous. As part of our baseline analysis, we also incorporate traditional Principal Component Analysis (PCA) on vectorized observations. In the case of the Ising model, we also compare with LPCA (Logistic PCA) and CLPCA (Convex Logistic PCA), both introduced in \textcite{LandgrafLee2020}. All experiments are performed at sample sizes $n = 100, 200, 300, 500$ and $750$. Every experiment is repeated $100$ times.


In this section, we report simulation results for the multilinear normal and the multilinear Ising model where different aspects of the GMLM model are compared against other methods. The comparison methods are Tensor Sliced Inverse Regression (TSIR) \parencite{DingCook2015}, Multiway Generalized Canonical Correlation Analysis (MGCCA) \parencite{ChenEtAl2021,GirkaEtAl2024} and the Tucker decomposition that is a higherorder form of principal component analysis (HOPCA) \textcite{KoldaBader2009}, for both continuous and binary data. For the latter, the binary values are treated as continuous. As part of our baseline analysis, we also incorporate traditional Principal Component Analysis (PCA) on vectorized observations. In the case of the Ising model, we also compare with LPCA (Logistic PCA) and CLPCA (Convex Logistic PCA), both introduced in \textcite{LandgrafLee2020}. All experiments are performed at sample sizes $n = 100, 200, 300, 500$ and $750$. Every experiment is repeated $100$ times.




We are interested in the quality of the estimate of the true sufficient reduction $\ten{R}(\ten{X})$ from \cref{thm:sdr}. Therefore, we compare with the true vectorized reduction matrix $\mat{B} = \bigkron_{k = r}^{1}\mat{\beta}_k$, as it is compatible with any linear reduction method. The distance $d(\mat{B}, \hat{\mat{B}})$ between $\mat{B}\in\mathbb{R}^{p\times q}$ and an estimate $\hat{\mat{B}}\in\mathbb{R}^{p\times \tilde{q}}$ is the \emph{subspace distance} which is proportional to


\begin{displaymath}



@ 924,16 +879,13 @@ We are interested in the quality of the estimate of the true sufficient reductio


the Frobenius norm of the difference between the projections onto the span of $\mat{B}$ and $\hat{\mat{B}}$. The proportionality constant\footnote{Depends on row dimension $p$ and the ranks of $\mat{B}$ and $\hat{\mat{B}}$ given by $(\min(\rank\mat{B} + \rank\hat{\mat{B}}, 2 p  (\rank\mat{B} + \rank\hat{\mat{B}})))^{1/2}$.} of $d(\mat{B}, \hat{\mat{B}})$ ensures that the subspace distance is in the interval $[0, 1]$. A distance of zero implies space overlap, a distance of one means that the subspaces are orthogonal.




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsection{Tensor Normal}\label{sec:simtensornormal}


\subsection{MultiLinear Normal}\label{sec:simtensornormal}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


We generate a random sample $y_i$, $i=1,\ldots, n$, from the standard normal distribution. We then draw i.i.d. samples $\ten{X}_i$ for $i = 1, ..., n$ from the conditional tensor normal distribution of $\ten{X}\mid Y = y_i$. The conditional distribution $\ten{X}\mid Y = y_i$ depends on the choice of the GMLM parameters $\overline{\ten{\eta}}$, $\mat{\beta}_1, ..., \mat{\beta}_r$, $\mat{\Omega}_1, ..., \mat{\Omega}_r$, and the function $\ten{F}_y$ of $y$. In all experiments we set $\overline{\ten{\eta}} = \mat{0}$. The other parameters and $\ten{F}_y$ are described per experiment. With the true GMLM parameters and $\ten{F}_y$ given, we compute the conditional tensor normal mean $\ten{\mu}_y = \ten{F}_y\mlm_{k = 1}^{r}\mat{\Omega}_k^{1}\mat{\beta}_k$ and covariances $\mat{\Sigma}_k = \mat{\Omega}_k^{1}$ as in \eqref{eq:tnormal_cond_params}.


We generate a random sample $y_i$, $i=1,\ldots, n$, from the standard normal distribution. We then draw i.i.d. samples $\ten{X}_i$ for $i = 1, ..., n$ from the conditional multilinear normal distribution of $\ten{X}\mid Y = y_i$. The conditional distribution $\ten{X}\mid Y = y_i$ depends on the choice of the GMLM parameters $\overline{\ten{\eta}}$, $\mat{\beta}_1, ..., \mat{\beta}_r$, $\mat{\Omega}_1, ..., \mat{\Omega}_r$, and the function $\ten{F}_y$ of $y$. In all experiments we set $\overline{\ten{\eta}} = \mat{0}$. The other parameters and $\ten{F}_y$ are described per experiment. With the true GMLM parameters and $\ten{F}_y$ given, we compute the conditional multilinear normal mean $\ten{\mu}_y = \ten{F}_y\mlm_{k = 1}^{r}\mat{\Omega}_k^{1}\mat{\beta}_k$ and covariances $\mat{\Sigma}_k = \mat{\Omega}_k^{1}$ as in \eqref{eq:tnormal_cond_params}.




We consider the following settings:


%We start with a $1$ dimensional linear dependence on $y$ in 1a). Then, the dependence of $y$ is via a cubic polynomial 1bd). In 1b) reduction is full rank, in contrast to 1c) where the $\mat{\beta}_k$'s are of rank $1$, in other words, low rank regression. In 1d) we constrain the inverse covariances $\mat{\Omega}_k$ to be tridiagonal. Both, 1cd) are examples of building the parameter space according to \cref{thm:parammanifold}. The final tensor normal experiment 1e) is a model misspecification. The true model does \emph{not} have a Kronecker structure and the ``known'' function $\ten{F}_y$ of $y$ is misspecified as well.




\begin{itemize}


\item[1a)] $\ten{X}$ is a threeway ($r = 3$) array of dimension $2\times 3\times 5$, %The dependence through the inverse regression model is linear specifically means that


and $\ten{F}_y\equiv y$ is a $1\times 1\times 1$ tensor. The true $\mat{\beta}_k$'s are all equal to $\mat{e}_1\in\mathbb{R}^{p_k}$, the first unit vector, for $k \in \{1, 2, 3\}$. The matrices $\mat{\Omega}_k = \mathrm{AR}(0.5)$ follow an autoregression like structure. That is, the elements are given by $(\mat{\Omega}_k)_{i j} = 0.5^{i  j}$.


\item[1a)] $\ten{X}$ is a threeway ($r = 3$) array of dimension $2\times 3\times 5$, and $\ten{F}_y\equiv y$ is a $1\times 1\times 1$ tensor. The true $\mat{\beta}_k$'s are all equal to $\mat{e}_1\in\mathbb{R}^{p_k}$, the first unit vector, for $k \in \{1, 2, 3\}$. The matrices $\mat{\Omega}_k = \mathrm{AR}(0.5)$ follow an autoregression like structure. That is, the elements are given by $(\mat{\Omega}_k)_{i j} = 0.5^{i  j}$.


\item[1b)] $\ten{X}$ is a threeway ($r = 3$) array of dimension $2\times 3\times 5$, and relates to the response $y$ via a qubic polynomial. This is modeled via $\ten{F}_y$ of dimension $2\times 2\times 2$ by the twice iterated outer product of the vector $(1, y)$. Element wise this reads $(\ten{F}_y)_{i j k} = y^{i + j + k  3}$. All $\mat{\beta}_k$'s are set to $(\mat{e}_1, \mat{e}_2)\in\mathbb{R}^{p_k\times 2}$ with $\mat{e}_i$ the $i$th unit vector and the $\mat{\Omega}_k$'s are $\mathrm{AR}(0.5)$.


\item[1c)] Same as 1b), except that the GMLM parameters $\mat{\beta}_k$ are rank $1$ given by


\begin{displaymath}



@ 942,33 +894,32 @@ We consider the following settings:


\mat{\beta}_3 = \begin{pmatrix} 1 & 1 \\ 1 & 1 \\ 1 & 1 \\ 1 & 1 \\ 1 & 1 \end{pmatrix}.


\end{displaymath}


\item[1d)] Same as 1b), but the true $\mat{\Omega}_k$ is tridiagonal, for $k = 1, 2, 3$. Their elements are given by $(\mat{\Omega}_k)_{i j} = \delta_{0, i  j} + 0.5\delta_{1, i  j}$ with $\delta_{i, j}$ being the Kronecker delta.


\item[1e)] For the misspecification model we let $\ten{X}\mid Y$ be multivariate but \emph{not} tensor normal. Let $\ten{X}$ be a $5\times 5$ random matrix with normal entries, $Y$ univariate standard normal and $\mat{f}_y$ a $4$ dimensional vector given by $\mat{f}_y = (1, \sin(y), \cos(y), \sin(y)\cos(y))$. The true vectorized reduction matrix $\mat{B}$ is $25\times 4$ consisting of the first $4$ columns of the identity; i.e., $\mat{B}_{i j} = \delta_{i j}$. The variancecovariance matrix $\mat{\Sigma}$ has elements $\mat{\Sigma}_{i j} = 0.5^{i  j}$. %is an autoregression like structure with correlation coefficient $0.5$.


Both, $\mat{B}$ and $\mat{\Omega} = \mat{\Sigma}^{1}$ violate the Kronecker product assumptions \eqref{eq:eta1} and \eqref{eq:eta2} of the GMLM model. Then, we set


\item[1e)] For the misspecification model we let $\ten{X}\mid Y$ be multivariate but \emph{not} multilinear normal. Let $\ten{X}$ be a $5\times 5$ random matrix with normal entries, $Y$ univariate standard normal and $\mat{f}_y$ a $4$ dimensional vector given by $\mat{f}_y = (1, \sin(y), \cos(y), \sin(y)\cos(y))$. The true vectorized reduction matrix $\mat{B}$ is $25\times 4$ consisting of the first $4$ columns of the identity; i.e., $\mat{B}_{i j} = \delta_{i j}$. The variancecovariance matrix $\mat{\Sigma}$ has elements $\mat{\Sigma}_{i j} = 0.5^{i  j}$. Both, $\mat{B}$ and $\mat{\Omega} = \mat{\Sigma}^{1}$ violate the Kronecker product assumptions \eqref{eq:eta1} and \eqref{eq:eta2} of the GMLM model. Then, we set


\begin{displaymath}


\vec{\ten{X}}\mid (Y = y) = \mat{B}\mat{f}_y + \mathcal{N}_{25}(\mat{0}, \mat{\Sigma}).


\end{displaymath}


Furthermore, we fit the model with the wrong ``known'' function $\ten{F}_y$. We set $\ten{F}_y$ to be a $2\times 2$ matrix with $(\ten{F}_y)_{i j} = y^{i + j  2}$, $i,j=1,2$.


\end{itemize}




The final tensor normal experiment 1e) is a misspecified model to explore the robustness of our approach. The true model does \emph{not} have a Kronecker structure and the ``known'' function $\ten{F}_y$ of $y$ is misspecified as well.


The final multilinear normal experiment 1e) is a misspecified model to explore the robustness of our approach. The true model does \emph{not} have a Kronecker structure and the ``known'' function $\ten{F}_y$ of $y$ is misspecified as well.






\begin{figure}[hp!]


\centering


\includegraphics[width = \textwidth]{plots/simnormal.pdf}


\caption{\label{fig:simnormal}Visualization of the simulation results for the tensor normal GMLM. Sample size on the $x$axis and the mean of subspace distance $d(\mat{B}, \hat{\mat{B}})$ over $100$ replications on the $y$axis. Described in \cref{sec:simtensornormal}.}


\caption{\label{fig:simnormal}Visualization of the simulation results for the multilinear normal GMLM. Sample size on the $x$axis and the mean of subspace distance $d(\mat{B}, \hat{\mat{B}})$ over $100$ replications on the $y$axis. Described in \cref{sec:simtensornormal}.}


\end{figure}






The results are visualized in \cref{fig:simnormal}. Simulation 1a), given a 1D linear relation between the response $Y$ and $\ten{X}$, TSIR and GMLM are equivalent. This is expected as \textcite{DingCook2015} already established that TSIR gives the MLE estimate under a tensor (matrix) normal distributed setting. For the other methods, MGCCA is only a bit better than PCA which, unexpectedly, beats HOPCA. But none of them are close to the performance of TSIR or GMLM. Continuing with 1b), where we introduced a cubic relation between $Y$ and $\ten{X}$, we observe a bigger deviation in the performance of GMLM and TSIR. This is caused mainly because we are estimating an $8$ dimensional subspace now, which amplifies the small performance boost, in the subspace distance, we gain by avoiding slicing. The GMLM model in 1c) behaves as expected, clearly being the best. The other results are surprising. First, PCA, HOPCA and MGCCA are visually indistinguishable. This is explained by a high signaltonoise ratio in this particular example. But the biggest surprise is the failure of TSIR. Even more surprising is that the conditional distribution $\ten{X}\mid Y$ is tensor normal distributed which, in conjunction with $\cov(\vec\ten{X})$ having a Kronecker structure, should give the MLE estimate. The lowrank assumption is also not an issue, this simply relates to TSIR estimating a 1D linear reduction which fulfills all the requirements. Finally, a common known issue of slicing, used in TSIR, is that conditional multimodal distributions can cause estimation problems due to the different distribution modes leading to vanishing slice means. Again, this is not the case in simulation 1c).


An investigation into this behaviour revealed the problem in the estimation of the mode covariance matrices $\mat{O}_k = \E[(\ten{X}  \E\ten{X})_{(k)}\t{(\ten{X}  \E\ten{X})_{(k)}}]$. The mode wise reductions provided by TSIR are computed as $\hat{\mat{O}}_k^{1}\hat{\mat{\Gamma}}_k$ where the poor estimation of $\hat{\mat{O}}_k$ causes the failure of TSIR. The poor estimate of $\mat{O}_k$ is rooted in the high signal to noise ratio in this particular simulation. GMLM does not have degenerate behaviour for high signal to noise ratios but it is less robust in low signal to noise ratio setting where TSIR performs better in this specific example.


Simulation 1d), incorporating information about the covariance structure behaves similar to 1b), except that GMLM gains a statistically significant lead in estimation performance. The last simulation, 1e), where the model was misspecified for GMLM. GMLM, TSIR as well as MGCCA are on par where GMLM has a sligh lead in the small sample size setting and MGCCA overtakes in higher sample scenarios. The PCA and HOPCA methods both still outperformed. A wrong assumption about the relation to the response is still better than no relation at all.


The results are visualized in \cref{fig:simnormal}. Simulation 1a), given a 1D linear relation between the response $Y$ and $\ten{X}$, TSIR and GMLM are equivalent. This is expected as \textcite{DingCook2015} already established that TSIR gives the MLE estimate under a multilinear (matrix) normal distributed setting. For the other methods, MGCCA is only a bit better than PCA which, unexpectedly, beats HOPCA. But none of them are close to the performance of TSIR or GMLM. Continuing with 1b), where we introduced a cubic relation between $Y$ and $\ten{X}$, we observe a bigger deviation in the performance of GMLM and TSIR. This is caused mainly because we are estimating an $8$ dimensional subspace now, which amplifies the small performance boost, in the subspace distance, we gain by avoiding slicing. The GMLM model in 1c) behaves as expected, clearly being the best. The other results are surprising. First, PCA, HOPCA and MGCCA are visually indistinguishable. This is explained by a high signaltonoise ratio in this particular example. But the biggest surprise is the failure of TSIR. Even more surprising is that the conditional distribution $\ten{X}\mid Y$ is multilinear normal distributed which, in conjunction with $\cov(\vec\ten{X})$ having a Kronecker structure, should give the MLE estimate. The lowrank assumption is also not an issue, this simply relates to TSIR estimating a 1D linear reduction which fulfills all the requirements. Finally, a common known issue of slicing, used in TSIR, is that conditional multimodal distributions can cause estimation problems due to the different distribution modes leading to vanishing slice means. Again, this is not the case in simulation 1c).


An investigation into this behavior revealed the problem in the estimation of the mode covariance matrices $\mat{O}_k = \E[(\ten{X}  \E\ten{X})_{(k)}\t{(\ten{X}  \E\ten{X})_{(k)}}]$. The mode wise reductions provided by TSIR are computed as $\hat{\mat{O}}_k^{1}\hat{\mat{\Gamma}}_k$ where the poor estimation of $\hat{\mat{O}}_k$ causes the failure of TSIR. The poor estimate of $\mat{O}_k$ is rooted in the high signaltonoise ratio in this particular simulation. GMLM does not have degenerate behavior for high signaltonoise ratios but it is less robust in low signaltonoise ratio setting where TSIR performs better in this specific example.


Simulation 1d), incorporating information about the covariance structure behaves similarly to 1b), except that GMLM gains a statistically significant lead in estimation performance. The last simulation, 1e), where the model was misspecified for GMLM. GMLM, TSIR, as well as MGCCA, are on par where GMLM has a slight lead in the small sample size setting and MGCCA overtakes in higher sample scenarios. The PCA and HOPCA methods both still outperformed. A wrong assumption about the relation to the response is still better than no relation at all.






%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsection{Ising Model}\label{sec:simising}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


Assuming for $\ten{X}$ being a $2\times 3$ dimensional binary matrix with conditional matrix (tensor) Ising distribution $\ten{X}\mid Y$ as in \cref{sec:ising_estimation}. We let for $i = 1, \ldots, n$ the response being i.i.d. uniformly distributed in $[1, 1]$ establishing the conditional value in the i.i.d. samples from $\ten{X}_i\mid Y = y_i$ with GMLM parameters $\mat{\beta}_1, \mat{\beta}_2$, $\mat{\Omega}_1, \mat{\Omega}_2$. We let


Assuming for $\ten{X}$ being a $2\times 3$ dimensional binary matrix with conditional matrix (multilinear) Ising distribution $\ten{X}\mid Y$ as in \cref{sec:ising_estimation}. We let for $i = 1, \ldots, n$ the response being i.i.d. uniformly distributed in $[1, 1]$ establishing the conditional value in the i.i.d. samples from $\ten{X}_i\mid Y = y_i$ with GMLM parameters $\mat{\beta}_1, \mat{\beta}_2$, $\mat{\Omega}_1, \mat{\Omega}_2$. We let


\begin{displaymath}


\mat{\beta}_1 = \begin{pmatrix}


1 & 0 \\ 0 & 1



@ 1006,7 +957,7 @@ if not mentioned otherwise in a specific simulation setup given next.


0 & 0 \\ 1 & 1 \\ 0 & 0


\end{pmatrix}.


\end{displaymath}


\item[2d)] We conclude with a simulation relating to the original design of the Ising model. It is a mathematical model to study the behaviour of Ferromagnetism \textcite{Ising1925} in a thermodynamic setting modeling the interaction effects of elementary magnets (spin up/down relating to $0$ and $1$). The model assumes all elementary magnets to be the same, which translates to all having the same coupling strength (twoway interactions) governed by a single parameter relating to the temperature of the system. Assuming the magnets to be arranged in a 2D grid (matrix valued $\ten{X}$), their interactions are constraint to direct neighbours. We can model this by choosing the true $\mat{\Omega}_k$'s to be tridiagonal matrices with zero diagonal entries and all nonzero entries identical. Since this is a 1D matrix manifold, we can enforce the constraint. Setting the true interaction parameters to be


\item[2d)] We conclude with a simulation relating to the original design of the Ising model. It is a mathematical model to study the behavior of Ferromagnetism \textcite{Ising1925} in a thermodynamic setting modeling the interaction effects of elementary magnets (spin up/down relating to $0$ and $1$). The model assumes all elementary magnets to be the same, which translates to all having the same coupling strength (twoway interactions) governed by a single parameter relating to the temperature of the system. Assuming the magnets to be arranged in a 2D grid (matrixvalued $\ten{X}$), their interactions are constrained to direct neighbors. We can model this by choosing the true $\mat{\Omega}_k$'s to be tridiagonal matrices with zero diagonal entries and all nonzero entries identical. Since this is a 1D matrix manifold, we can enforce the constraint. Setting the true interaction parameters to be


\begin{displaymath}


\mat{\Omega}_1 = \frac{1}{2}\begin{pmatrix}


0 & 1 \\ 1 & 0



@ 1025,10 +976,10 @@ if not mentioned otherwise in a specific simulation setup given next.


\caption{\label{fig:simising}Visualization of the simulation results for Ising GMLM. Sample size on the $x$axis and the mean of subspace distance $d(\mat{B}, \hat{\mat{B}})$ over $100$ replications on the $y$axis. Described in \cref{sec:simising}.}


\end{figure}




The simulation results are visualized in \cref{fig:simising}. Regardless of the simulation setting 2ad), the comparative results are similar. We observe that PCA and HOPCA, both treating the response $\ten{X}$ as continuous, perform poorly. Not much better are LPCA and CLPCA. Similar to PCA and HOPCA they do not consider the relation to the response, but they are specifically created for binary predictors. Next we have MGCCA which is the first method considering the relation to the response $y$, clearly outperforming all the PCA variants. Even better is TSIR, regardless of the treatment of the predictors $\ten{X}$ as continuous, achieving very good results. Finally, the Ising GMLM model is the best in all the simulations although TSIR gets very close in some settings.


The simulation results are visualized in \cref{fig:simising}. Regardless of the simulation setting 2ad), the comparative results are similar. We observe that PCA and HOPCA, both treating the response $\ten{X}$ as continuous, perform poorly. Not much better are LPCA and CLPCA. Similar to PCA and HOPCA they do not consider the relation to the response, but they are specifically created for binary predictors. Next, we have MGCCA which is the first method considering the relation to the response $y$, clearly outperforming all the PCA variants. Even better is TSIR, regardless of the treatment of the predictors $\ten{X}$ as continuous, achieving very good results. Finally, the Ising GMLM model is the best in all the simulations although TSIR gets very close in some settings.




% Due to the surprisingly good result of TSIR, we also applied the tensor normal GMLM model to the exact same simulation, simply treating the response $\ten{X}$ as continuous.


% The raw linear reduction estimates of both the Ising GMLM and the tensor normal GMLM are basically indistinguishable, similar to the very similar results of the different PCA variants. The main reason for this specific


% Due to the surprisingly good result of TSIR, we also applied the multilinear normal GMLM model to the exact same simulation, simply treating the response $\ten{X}$ as continuous.


% The raw linear reduction estimates of both the Ising GMLM and the multilinear normal GMLM are basically indistinguishable, similar to the very similar results of the different PCA variants. The main reason for this specific




% \begin{table}


% \begin{tabular}{c  ccc ccc c}



@ 1048,20 +999,20 @@ The simulation results are visualized in \cref{fig:simising}. Regardless of the


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\section{Data Analysis}\label{sec:dataanalysis}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


In this section we perform two applications of the GMLM model on real data. First example is the tensor normal model applied to EEG data. Next, we perform a prove of concept data analysis example for chess.


In this section, we perform two applications of the GMLM model on real data. The first example is the multilinear normal model applied to EEG data. Next, we perform a proof of concept data analysis example for chess.






%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsection{EEG}


The EEG data (\url{http://kdd.ics.uci.edu/databases/eeg/eeg.data.html}) is a small study of $77$ alcoholic and $45$ control subjects. Each data point corresponding to a subject consists of a $p_1\times p_2 = 256\times 64$ matrix, with each row representing a time point and each column a channel. The measurements were obtained by exposing each individual to visual stimuli and measuring voltage values from $64$ electrodes placed on the subjects' scalps sampled at $256$ time points over $1$ second ($256$ Hz). Different stimulus conditions were used, and for each condition, $120$ trials were measured. We used only a single stimulus condition (S1), and for each subject, we took the average of all the trials under that condition. That is, we used $(\ten{X}_i, y_i)$, $i = 1, \ldots, 122$, where $\ten{X}_i$ is a $256\times 64$ matrix, with each entry representing the mean voltage value of subject $i$ at a combination of a time point and a channel, averaged over all trials under the S1 stimulus condition, and $Y$ is a binary outcome variable with $Y_i = 1$ for an alcoholic and $Y_i = 0$ for a control subject.


The EEG data\footnote{\fullcite{eegdataset}} is a small study of $77$ alcoholic and $45$ control subjects. Each data point corresponding to a subject consists of a $p_1\times p_2 = 256\times 64$ matrix, with each row representing a time point and each column a channel. The measurements were obtained by exposing each individual to visual stimuli and measuring voltage values from $64$ electrodes placed on the subjects' scalps sampled at $256$ time points over $1$ second ($256$ Hz). Different stimulus conditions were used, and for each condition, $120$ trials were measured. We used only a single stimulus condition (S1), and for each subject, we took the average of all the trials under that condition. That is, we used $(\ten{X}_i, y_i)$, $i = 1, \ldots, 122$, where $\ten{X}_i$ is a $256\times 64$ matrix, with each entry representing the mean voltage value of subject $i$ at a combination of a time point and a channel, averaged over all trials under the S1 stimulus condition, and $Y$ is a binary outcome variable with $Y_i = 1$ for an alcoholic and $Y_i = 0$ for a control subject.




For a comparison we reproduced the leaveoneout crossvalidation EEG data analysis \textcite[Sec. 7]{PfeifferKaplaBura2021} for the classification task. In this data set, $p= p_1 p_2 = 16384$ is much larger than $n=122$. To deal with this issue, \textcite{PfeifferKaplaBura2021} used two approaches. In the first, prescreening via (2D)$^2$PCA \parencite{ZhangZhou2005} reduced the dimensions to $(p_1, p_2) = (3, 4)$, $(15, 15)$ and $(20, 30)$. In the second, simultaneous dimension reductions and variable selection was carried out using the fast POIC algorithm of \textcite{JungEtAl2019} (due to high computational high burden, only a 10fold crossvalidation was performed for fast POIC).


For a comparison, we reproduced the leaveoneout crossvalidation EEG data analysis \textcite[Sec. 7]{PfeifferKaplaBura2021} for the classification task. In this data set, $p= p_1 p_2 = 16384$ is much larger than $n=122$. To deal with this issue, \textcite{PfeifferKaplaBura2021} used two approaches. In the first, prescreening via (2D)$^2$PCA \parencite{ZhangZhou2005} reduced the dimensions to $(p_1, p_2) = (3, 4)$, $(15, 15)$ and $(20, 30)$. In the second, simultaneous dimension reduction and variable selection were carried out using the fast POIC algorithm of \textcite{JungEtAl2019} (due to high computational burden, only 10fold crossvalidation was performed for fast POIC).




In contrast to \textcite{PfeifferKaplaBura2021}, our GMLM model can be applied directly to the raw data of dimension $(256, 64)$ without prescreening or variable selection. This was not possible for KPIR as the time axis alone was in the large $p$ small $n$ regime with the $p_1 = 256 > n = 122$ leading to a singular time axis covariance. The same issue is present in the GMLM model, but the regularization trick used for numerical stability, as described in \cref{sec:tensornormalestimation}, resolves this without any change to the estimation procedure. In general, the sample size does not need to be large for maximum likelihood estimation in the tensor normal model. In matrix normal models in particular, \cite{DrtonEtAl2020} proved that very small sample sizes, as little as $3$,\footnote{The required minimum sample size depends on a nontrivial algebraic relations between the mode dimensions, while the magnitude of the dimensions has no specific role.} are sufficient to obtain unique MLEs for Kronecker covariance structures.


In contrast to \textcite{PfeifferKaplaBura2021}, our GMLM model can be applied directly to the raw data of dimension $(256, 64)$ without prescreening or variable selection. This was not possible for KPIR as the time axis alone was in the large $p$ small $n$ regime with the $p_1 = 256 > n = 122$ leading to a singular time axis covariance. The same issue is present in the GMLM model, but the regularization trick used for numerical stability, as described in \cref{sec:tensornormalestimation}, resolves this without any change to the estimation procedure. In general, the sample size does not need to be large for maximum likelihood estimation in the multilinear normal model. In matrix normal models in particular, \cite{DrtonEtAl2020} proved that very small sample sizes, as little as $3$,\footnote{The required minimum sample size depends on nontrivial algebraic relations between the mode dimensions, while the magnitude of the dimensions has no specific role.} are sufficient to obtain unique MLEs for Kronecker covariance structures.




We use leaveoneout crossvalidation to obtain unbiased AUC estimates. Then, we compare the GMLM model to the best performing methods from \textcite{PfeifferKaplaBura2021}, namely KPIR (ls) and LSIR from \textcite{PfeifferForzaniBura2012} for $(p_1, p_2) = (3, 4)$, $(15, 15)$ and $(20, 30)$.




In \cref{tab:eeg} we provide the AUC and its standard deviation. For all applied prescreening dimensions, KPIR (ls) has an AUC of $78\%$. LSIR performs better at the price of some instability; it peaked at $85\%$ at $(3, 4)$, then dropped down to $81\%$ at $(15, 15)$ and then increased to $83\%$. In contract, our GMLM method peaked at $(3, 4)$ with $85\%$ and stayed stable at $84\%$, even when no preprocessing was applied. In contrast, fast POIC that carries out simultaneous feature extraction and feature selection resulted in an AUC of $63\%$, clearly outperformed by all other methods.


In \cref{tab:eeg} we provide the AUC and its standard deviation. For all applied prescreening dimensions, KPIR (ls) has an AUC of $78\%$. LSIR performs better at the price of some instability; it peaked at $85\%$ at $(3, 4)$, then dropped down to $81\%$ at $(15, 15)$, and then increased to $83\%$. In contrast, our GMLM method peaked at $(3, 4)$ with $85\%$ and stayed stable at $84\%$, even when no preprocessing was applied. In contrast, fast POIC that carries out simultaneous feature extraction and feature selection resulted in an AUC of $63\%$, clearly outperformed by all other methods.




\begin{table}[!hpt]


\centering



@ 1086,9 +1037,9 @@ In \cref{tab:eeg} we provide the AUC and its standard deviation. For all applied




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\subsection{Chess}\label{sec:chess}


The data set is provided by the \citetitle{lichessdatabase}\footnote{\fullcite{lichessdatabase}}. We randomly selected the November of 2023 data that consist of more than $92$ million games. We removed all games without position evaluations. The evaluations, also denoted as scores, are from Stockfish\footnote{\fullcite{stockfish}}, a free and strong chess engine. The scores take the role of the response $Y$ and correspond to a winning probability from the white pieces point of view. Positive scores are good for white and negative scores indicate an advantage for black pieces. We ignore all highly unbalanced positions, which we set to be positions with absolute score above $5$. We also remove all positions with a mate score (one side can force checkmate). Furthermore, we only consider positions after $10$ halfmoves to avoid oversampling the beginning of the most common openings including the start position which is in every game. Finally, we only consider positions with white to move. This leads to a final data set of roughly $64$ million positions, including duplicates.


The data set is provided by the \citetitle{lichessdatabase}\footnote{\fullcite{lichessdatabase}}. We randomly selected the November of 2023 data that consist of more than $92$ million games. We removed all games without position evaluations. The evaluations, also denoted as scores, are from Stockfish\footnote{\fullcite{stockfish}}, a free and strong chess engine. The scores take the role of the response $Y$ and correspond to a winning probability from the white pieces' point of view. Positive scores are good for white and negative scores indicate an advantage for black pieces. We ignore all highly unbalanced positions, which we set to be positions with absolute score above $5$. We also remove all positions with a mate score (one side can force checkmate). Furthermore, we only consider positions after $10$ halfmoves to avoid oversampling the beginning of the most common openings including the start position which is in every game. Finally, we only consider positions with white to move. This leads to a final data set of roughly $64$ million positions, including duplicates.




A chess position is encoded as a set of $12$ binary matrices $\ten{X}_{\mathrm{piece}}$ of dimensions $8\times 8$. Every binary matrix encodes the positioning of a particular piece by containing a $1$ if the piece is present at the corresponding board position. The $12$ pieces derive from the $6$ types of pieces, namely pawns (\pawn), knights (\knight), bishops (\bishop), queens (\queen) and kings (\king) of two colors, black and white. See \cref{fig:fen2tensor} for a visualization.


A chess position is encoded as a set of $12$ binary matrices $\ten{X}_{\mathrm{piece}}$ of dimensions $8\times 8$. Every binary matrix encodes the positioning of a particular piece by containing a $1$ if the piece is present at the corresponding board position. The $12$ pieces derive from the $6$ types of pieces, namely pawns (\pawn), knights (\knight), bishops (\bishop), queens (\queen), and kings (\king) of two colors, black and white. See \cref{fig:fen2tensor} for a visualization.




\begin{figure}[hp!]


\centering



@ 1096,26 +1047,26 @@ A chess position is encoded as a set of $12$ binary matrices $\ten{X}_{\mathrm{p


\caption{\label{fig:fen2tensor}The chess start position and its 3D binary tensor representation, empty entries are $0$.}


\end{figure}




We assume that $\ten{X}_{\mathrm{piece}}\mid Y = y$ follows an Ising GMLM model \cref{sec:ising_estimation} with different conditional piece predictors being independent. The independence assumption is for the sake of simplicity even though this is clearly not the case in the underlying true distribution. By this simplifying assumption we get a mixture model with the loglikelihood


We assume that $\ten{X}_{\mathrm{piece}}\mid Y = y$ follows an Ising GMLM model \cref{sec:ising_estimation} with different conditional piece predictors being independent. The independence assumption is for the sake of simplicity even though this is not the case in the underlying true distribution. By this simplifying assumption we get a mixture model with the loglikelihood


\begin{displaymath}


l_n(\mat{\theta}) = \frac{1}{12}\sum_{\mathrm{piece}}l_n(\mat{\theta}_{\mathrm{piece}})


\end{displaymath}


where $l_n(\mat{\theta}_{\mathrm{piece}})$ is the Ising GMLM loglikelihood as in \cref{sec:ising_estimation} for $\ten{X}_{\mathrm{piece}}\mid Y = y$. For every component the same relation to the scores $y$ is modeled via a $2\times 2$ dimensional matrix valued function $\ten{F}_y$ consisting of the monomials $1, y, y^2$, specifically $(\ten{F}_y)_{i j} = y^{i + j  2}$.




By the raw scale of the data, millions of observations, it is computationally infeasible to compute the gradients on the entire data set. Simply using a computationally manageable subset is not an option. Due to the high dimension on binary data, which is $12$ times a $8\times 8$ for every observation giving a total dimension of $768$. The main issue is that a manageable subset, say one million observations, still leads to a degenerate data set. In our simplified mixture model, the pawns are a specific issue as there are multiple millions of different combinations of the $8$ pawns per color on the $6\times 8$ sub grid the pawns can be positioned. This alone does not allow to take a reasonable sized subset for estimation. The solution is to switch from a classic gradient based optimization to a stochastic version. This means that every gradient update uses a new random subset of the entire data set. Therefore, we draw independent random samples form the data consisting of $64$ million positions. The independence of samples derived from the independence of games, and every sample is drawn from a different game.


By the raw scale of the data, millions of observations, it is computationally infeasible to compute the gradients on the entire data set. Simply using a computationally manageable subset is not an option. Due to the high dimension on binary data, which is $12$ times a $8\times 8$ for every observation giving a total dimension of $768$. The main issue is that a manageable subset, say one million observations, still leads to a degenerate data set. In our simplified mixture model, the pawns are a specific issue as there are multiple millions of different combinations of the $8$ pawns per color on the $6\times 8$ subgrid where the pawns can be positioned. This alone does not allow us to take a reasonable sized subset for estimation. The solution is to switch from a classic gradientbased optimization to a stochastic version. This means that every gradient update uses a new random subset of the entire data set. Therefore, we draw independent random samples from the data consisting of $64$ million positions. The independence of samples is derived from the independence of games, and every sample is drawn from a different game.




\paragraph{Validation:}


Given the nonlinear nature of the reduction, due to the quadratic matrix valued function $\ten{F}_y$ of the score $y$, we use a \emph{generalized additive model}\footnote{using the function \texttt{gam()} from the \texttt{R} package \texttt{mgcv}.} (GAM) to predict position scores from reduced positions. The reduced positions are $48$ dimensional continuous values by combining the $12$ mixture components from the $2\times 2$ matrix valued reductions per piece. The per piece reduction is


Given the nonlinear nature of the reduction, due to the quadratic matrixvalued function $\ten{F}_y$ of the score $y$, we use a \emph{generalized additive model}\footnote{using the function \texttt{gam()} from the \texttt{R} package \texttt{mgcv}.} (GAM) to predict position scores from reduced positions. The reduced positions are $48$ dimensional continuous values by combining the $12$ mixture components from the $2\times 2$ matrixvalued reductions per piece. The perpiece reduction is


\begin{displaymath}


\ten{R}(\ten{X}_{\mathrm{piece}}) = \mat{\beta}_{1,\mathrm{piece}}(\ten{X}_{\mathrm{piece}}  \E\ten{X}_{\mathrm{piece}})\t{\mat{\beta}_{2, \mathrm{piece}}}


\end{displaymath}


which gives the complete $48$ dimensional vectorized reduction by stacking the piece wise reductions


which gives the complete $48$ dimensional vectorized reduction by stacking the piecewise reductions


\begin{displaymath}


\vec{\ten{R}(\ten{X}})


= (\vec{\ten{R}(\ten{X}_{\text{white pawn}})}, \ldots, \vec{\ten{R}(\ten{X}_{\text{black king}})})


= \t{\mat{B}}\vec(\ten{X}  \E\ten{X}).


\end{displaymath}


The second line encodes all the piece wise reductions in a block diagonal full reduction matrix $\mat{B}$ of dimension $768\times 48$ which is applied to the vectorized 3D tensor $\ten{X}$ combining all the piece components $\ten{X}_{\mathrm{piece}}$ into a single tensor of dimension $8\times 8\times 12$. This is a reduction to $6.25\%$ of the original dimension. The $R^2$ statistic of the GAM fitted on $10^5$ new reduced samples is $R^2_{\mathrm{gam}}\approx 46\%$. A linear model on the reduced data achieves $R^2_{\mathrm{lm}}\approx 26\%$ which clearly shows the nonlinear relation. On the other hand, the static evaluation of the \emph{Schach H\"ornchen}\footnote{Main authors personal chess engine.} engine, given the full position (\emph{not} reduced), achieves an $R^2_{\mathrm{hce}}\approx 52\%$. The $42\%$ are reasonably well compared to $51\%$ of the engine static evaluation which gets the original position and uses chess specific expect knowledge. Features the static evaluation includes, which are expected to be learned by the GMLM mixture model, are; \emph{material} (piece values) and \emph{piece square tables} (PSQT, preferred piece type positions). In addition, the static evaluation includes chess specific features like \emph{king safety}, \emph{pawn structure} or \emph{rooks on open files}. This lets us conclude that the reduction captures most of the relevant features possible, given the oversimplified modeling we performed.


The second line encodes all the piecewise reductions in a block diagonal full reduction matrix $\mat{B}$ of dimension $768\times 48$ which is applied to the vectorized 3D tensor $\ten{X}$ combining all the piece components $\ten{X}_{\mathrm{piece}}$ into a single tensor of dimension $8\times 8\times 12$. This is a reduction to $6.25\%$ of the original dimension. The $R^2$ statistic of the GAM fitted on $10^5$ new reduced samples is $R^2_{\mathrm{gam}}\approx 46\%$. A linear model on the reduced data achieves $R^2_{\mathrm{lm}}\approx 26\%$ which clearly shows the nonlinear relation. On the other hand, the static evaluation of the \emph{Schach H\"ornchen}\footnote{Main author's chess engine.} engine, given the full position (\emph{not} reduced), achieves an $R^2_{\mathrm{hce}}\approx 52\%$. The $42\%$ are reasonably well compared to $51\%$ of the engine static evaluation which gets the original position and uses chess specific expert knowledge. Features the static evaluation includes, which are expected to be learned by the GMLM mixture model, are; \emph{material} (piece values) and \emph{piece square tables} (PSQT, preferred piece type positions). In addition, the static evaluation includes chess specific features like \emph{king safety}, \emph{pawn structure}, or \emph{rooks on open files}. This lets us conclude that the reduction captures most of the relevant features possible, given the oversimplified modeling we performed.




\paragraph{Interpretation:} For a compact interpretation of the estimated reduction we construct PSQTs. To do so we use the linear model from the validation section. Then, we rewrite the combined linear reduction and linear model in terms of PSQTs. Let $\mat{B}$ be the $768\times 48$ full vectorized linear reduction. This is the block diagonal matrix with the $64\times 4$ dimensional per piece reductions $\mat{B}_{\mathrm{piece}} = \mat{\beta}^{\mathrm{piece}}_2\otimes\mat{\beta}^{\mathrm{piece}}_1$. Then, the linear model with coefficients $\mat{b}$ and intercept $a$ on the reduced data is given by


\begin{equation}\label{eq:chesslm}



@ 1123,14 +1074,13 @@ The second line encodes all the piece wise reductions in a block diagonal full r


\end{equation}


with an unknown mean zero error term $\epsilon$ and treating the binary tensor $\ten{X}$ as continuous. Decomposing the linear model coefficients into blocks of $4$ gives per piece coefficients $\mat{b}_{\mathrm{piece}}$ which combine with the diagonal blocks $\mat{B}_{\mathrm{piece}}$ of $\mat{B}$ only. Rewriting \eqref{eq:chesslm} gives


\begin{align*}


y


&= a + \sum_{\mathrm{piece}}\t{(\mat{B}_{\mathrm{piece}}\mat{b}_{\mathrm{piece}})}\vec(\ten{X}_{\mathrm{piece}}  \E\ten{X}_{\mathrm{piece}}) + \epsilon \\


y &= a + \sum_{\mathrm{piece}}\t{(\mat{B}_{\mathrm{piece}}\mat{b}_{\mathrm{piece}})}\vec(\ten{X}_{\mathrm{piece}}  \E\ten{X}_{\mathrm{piece}}) + \epsilon \\


&= \tilde{a} + \sum_{\mathrm{piece}}\langle


\mat{B}_{\mathrm{piece}}\mat{b}_{\mathrm{piece}},


\vec(\ten{X}_{\mathrm{piece}})


\rangle + \epsilon


\end{align*}


with a new intercept term $\tilde{a}$, which is of no interest to us. Finally, we enforce a color symmetry, using known mechanism from chess engines. Specifically, mirroring the position changes the sign of the score $y$. Here, mirroring reverses the rank (row) order, this is the image in a mirror behind a chess board. Let for every $\mat{C}_{\mathrm{piece}}$ be a $8\times 8$ matrix with elements $(\mat{C}_{\mathrm{piece}})_{i j} = (\mat{B}_{\mathrm{piece}}\mat{b}_{\mathrm{piece}})_{i + 8 (j  1)}$. And denote with $\mat{M}(\mat{A})$ the matrix mirror operation which reverses the row order of a matrix. Using this new notation allows to enforcing this symmetry leading to the new approximate linear relation


with a new intercept term $\tilde{a}$, which is of no interest to us. Finally, we enforce color symmetry, using known mechanisms from chess engines. Specifically, mirroring the position changes the sign of the score $y$. Here, mirroring reverses the rank (row) order, this is the image in a mirror behind a chess board. Let for every $\mat{C}_{\mathrm{piece}}$ be a $8\times 8$ matrix with elements $(\mat{C}_{\mathrm{piece}})_{i j} = (\mat{B}_{\mathrm{piece}}\mat{b}_{\mathrm{piece}})_{i + 8 (j  1)}$. And denote with $\mat{M}(\mat{A})$ the matrix mirror operation which reverses the row order of a matrix. Using this new notation allows to enforcement of this symmetry leading to the new approximate linear relation


\begin{align*}


y &= \tilde{a} + \sum_{\mathrm{piece}}\langle


\mat{C}_{\mathrm{piece}},



@ 1141,7 +1091,7 @@ with a new intercept term $\tilde{a}$, which is of no interest to us. Finally, w


\ten{X}_{\text{white piece}}  \mat{M}(\ten{X}_{\text{white piece}})


\rangle + \epsilon


\end{align*}


If for every piece type ($6$ types, \emph{not} distinguishing between color) holds $\mat{C}_{\text{white piece}} = \mat{M}(\mat{C}_{\text{black piece}})$, then we have equality. In our case this is valid given that the estimates $\hat{\mat{C}}_{\mathrm{piece}}$ fulfill this property with a small error. The $6$ matrices $(\mat{C}_{\text{white piece}}  \mat{M}(\mat{C}_{\text{black piece}})) / 2$ are called \emph{piece square tables} (PSQT) which are visualized in \cref{fig:psqt}. The interpretation of those tables is straight forward. A high positive values (blue) means that it is usually good to have a piece of the corresponding type on that square while a high negative value (red) means the opposite. It needs to be considered that the PSQTs are for quiet positions only, that means all pieces are save in the sense that there is no legal capturing moves nore is the king in check.


If for every piece type ($6$ types, \emph{not} distinguishing between color) holds $\mat{C}_{\text{white piece}} = \mat{M}(\mat{C}_{\text{black piece}})$, then we have equality. In our case, this is valid given that the estimates $\hat{\mat{C}}_{\mathrm{piece}}$ fulfill this property with a small error. The $6$ matrices $(\mat{C}_{\text{white piece}}  \mat{M}(\mat{C}_{\text{black piece}})) / 2$ are called \emph{piece square tables} (PSQT) which are visualized in \cref{fig:psqt}. The interpretation of those tables is straightforward. A high positive value (blue) means that it is usually good to have a piece of the corresponding type on that square while a high negative value (red) means the opposite. It needs to be considered that the PSQTs are for quiet positions only, which means all pieces are save in the sense that there is no legal capturing moves nor is the king in check.




\begin{figure}[hp!]


\centering



@ 1156,13 +1106,20 @@ Next, going over the PSQTs one by one, a few words about the preferred positions


The results of our analysis in the previous paragraph agree with the configuration of the chess board most associated with observed chess game outcomes. This arrangement also aligns with the understanding of human chess players of an average configuration at any moment during the game.




\section{Discussion}


We have addressed sufficient dimension reduction for tensor valued predictors for regression or classification problems. Proposing a generalized multilinear model modeling the inverse conditional distribution we provided a multilinear sufficient reduction with consistent and asymptotic normal parameters. Moreover, our ansatz for proving the asymptotic results required by leveraging manifolds as a basis for resolving the issue of unidentifiable parameters lead to an even more flexible modeling framework. This allows to build complex and potentially problem specific parameter spaces incorporating additional domain specific knownledge into the model.


%We have addressed sufficient dimension reduction for tensor valued predictors for regression or classification problems.


We propose a generalized multilinear model formulation for the inverse conditional distribution of a tensorvalued predictor given a response and derive a multilinear sufficient reduction for the corresponding forward regression/classification problem. We also propose estimators for the sufficient reduction and show they are consistent and asymptotically normal. Obtaining the asymptotic results required leveraging manifolds as a basis for resolving the issue of unidentifiable parameters. This in turn led to an even more flexible modeling framework, which allows building complex and potentially problemspecific parameter spaces that incorporate additional domainspecific knowledge into the model.




Our multilinear Ising model can be thought of as the extension of the Ising modelbased approach of \textcite{ChengEtAl2014}, where a $q$dimensional binary vector is regressed on a $p$dimensional continuous vector. Yet, our model does not require penalization or sparsity assumptions, despite the tensor nature of the data, by leveraging the inherent structural information of the tensorvalued covariates assuming separable first and second moments. Moreover, it can accommodate a mixture of continuous and binary tensorvalued predictors, which is a subject of future work.


%A case in point is the popular oneparameter Ising model \parencite[e.g.][]{MukherjeeEtAl2022,NguyenEtAl2017} in Statistical Physics, which is parametrized via a single scaling factor of the covariance. Our model is capable of representing this specific structure via a linear Ising model where the parameters are represented by a one dimensional matrix manifold. Although this is interesting from a theoretical point of view, the design of our approach is \emph{not} intended for this kind of models.




An additional powerful extension of our model involves considering a sum of separable Kronecker predictors. This is motivated by the equivalence of a Kronecker product to a rank 1 tensor. By allowing a sum of a few separable Kronecker predictors, we remove the implicit rank 1 constraint. However, if this extension is to be applied to the SDR setting, as in this paper, it is crucial to ensure that the sum of Kronecker products forms a parameter manifold to apply our theory. While we anticipate that this approach can lead to intriguing and powerful models, there are certain details that need to be resolved first.


We allude to this feature of our approach in \cref{sec:matrixmanifolds}, where we also tabulate different matrix manifolds that can be used as building blocks $\manifold{B}_k$ and $\manifold{O}_k$ of the parameter space in \cref{tab:matrixmanifolds}. For example, our formulation can easily accommodate longitudinal data tabulated in matrix format, where the rows are covariates and the columns are consecutive time points with discrete AR($k$) dependence structure.




\todo{finish!}


Our multilinear Ising model can be thought of as the extension of the Ising modelbased approach of \textcite{ChengEtAl2014}, where a $q$dimensional binary vector is regressed on a $p$dimensional continuous vector. Yet, our model leverages the inherent structural information of the tensorvalued covariates by assuming separable first and second moments. By doing so, it bypasses requiring usual sparsity assumptions or penalization, despite the tensor highdimensional nature of the data. Moreover, it can accommodate a mixture of continuous and binary tensorvalued predictors, which is the subject of future work.






%Another interesting future research is to better understand the surprising behavior of TSIR \parencite{DingCook2015} we discovered in \cref{sec:simtensornormal}, especially as compared with our multilinear normal model for different signaltonoise ratios.




An additional powerful extension of our model involves considering a sum of separable Kronecker predictors. This is motivated by the equivalence of a Kronecker product to a rank $1$ tensor. By allowing a sum of a few separable Kronecker predictors, we remove the implicit rank $1$ constraint. However, if this extension is to be applied to the SDR setting, as in this paper, it is crucial to ensure that the sum of Kronecker products form a parameter manifold.% to apply our theory.


%While we anticipate that this approach can lead to intriguing and powerful models, certain details need to be resolved first.




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\printbibliography[heading=bibintoc, title={References}]



@ 1219,31 +1176,6 @@ as well as for any tensor $\ten{A}$ of even order $2 r$ and matching square matr


\t{(\vec{\ten{A}})}\vec\Bigl(\bigotimes_{k = r}^{1}\t{\mat{B}_k}\Bigr)


\end{displaymath}






% \begin{lemma}\label{thm:kronperm}


% Given $r$ matrices $\mat{A}_k$ of dimension $p_j\times q_j$ for $k = 1, \ldots, r$, then there exists a unique permutation matrix $\mat{S}_{\mat{p}, \mat{q}}$ such that


% \begin{equation}\label{eq:krontoouterperm}


% \vec\bigkron_{k = r}^{1}\mat{A}_j = \mat{S}_{\mat{p}, \mat{q}}\vec\bigouter_{k = 1}^{r}\mat{A}_k.


% \end{equation}


% The permutation $\mat{S}_{\mat{p}, \mat{q}}$ with indices $\mat{p} = (p_1, \ldots, p_r)$ and $\mat{q} = (q_1, \ldots, q_r)$ is the matrixmatrix product of $r  1$ permutation matrices given by


% \begin{multline}\label{eq:S_pq}


% \mat{S}_{\mat{p}, \mat{q}} =


% \Bigl[ \mat{I}_1\otimes \Bigl( \mat{I}_{\prod_{k = r}^{2}q_k}\otimes\mat{K}_{q_1, \prod_{k = r}^{2}p_k}\otimes I_{p_1} \Bigr)\Bigr] \\


% \Bigl[ \mat{I}_{p_1 q_1}\otimes \Bigl( \mat{I}_{\prod_{k = r}^{3}q_k}\otimes\mat{K}_{q_2, \prod_{k = r}^{3}p_k}\otimes I_{p_2} \Bigr) \Bigr]


% \cdots


% \Bigl[ \mat{I}_{\prod_{k = 1}^{r  2}p_k q_k}\otimes \Bigl( \mat{I}_{q_r}\otimes\mat{K}_{q_{r  1}, p_r}\otimes I_{p_{r  1}} \Bigr) \Bigr]


% \end{multline}


% where $\mat{K}_{p, q}$ is the \emph{commutation matrix} from \textcite[Ch.~11]{MatrixAlgebraAbadirMagnus2005}, that is the permutation such that $\vec{\t{\mat{A}}} = \mat{K}_{p, q}\vec{\mat{A}}$ for every $p\times q$ dimensional matrix $\mat{A}$.


% \end{lemma}


% \begin{proof}


% \textcite[Lemma~7]{SymMatandJacobiansMagnusNeudecker1986} states that


% \begin{align*}


% \vec(\mat{A}_2\otimes\mat{A}_1)


% &= (\mat{I}_{q_2}\otimes\mat{K}_{q_1, p_2}\otimes\mat{I}_{p_1})(\vec{\mat{A}_2}\otimes\vec{\mat{A}_1}) \\


% &= (\mat{I}_{q_2}\otimes\mat{K}_{q_1, p_2}\otimes\mat{I}_{p_1})\vec(\mat{A}_1\circ \mat{A}_2).


% \end{align*}


% This proves the statement for $r = 2$. The general statement for $r > 2$ follows via induction using \textcite[Lemma~7]{SymMatandJacobiansMagnusNeudecker1986} in conjunction with $\vec(\mat{C}\mat{a}\t{\mat{b}}) = (\mat{I}_{\dim(\mat{b})}\otimes\mat{C})\vec(\mat{a}\t{\mat{b}})$.


% \end{proof}


\begin{lemma}\label{thm:kronperm}


Given $r \geq 2$ matrices $\mat{A}_k$ of dimension $p_j\times q_j$ for $k = 1, \ldots, r$, then there exists a unique permutation matrix $\mat{S}_{\mat{p}, \mat{q}}$ such that


\begin{equation}\label{eq:krontoouterperm}




Loading…
Reference in New Issue