\documentclass[11pt]{amsart} \usepackage{amssymb,amsmath,amsthm,mathtools,booktabs,xcolor} \usepackage[margin=1.15in]{geometry} \usepackage{hyperref} \hypersetup{colorlinks=true,linkcolor=blue!50!black,citecolor=blue!50!black,urlcolor=blue!50!black} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{example}[theorem]{Example} \newtheorem{remark}[theorem]{Remark} \newcommand{\diag}{\operatorname{diag}} \newcommand{\link}{\mathcal{L}} \newcommand{\C}{\mathbb{C}} \newcommand{\Q}{\mathbb{Q}} \newcommand{\R}{\mathbb{R}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\PP}{\mathbb{P}} \newcommand{\ones}{\mathbf 1} \newcommand{\Res}{\operatorname{Res}} \newcommand{\Disc}{\operatorname{Disc}} \newcommand{\one}{\mathbf{1}} \title[Determinantal formula for Sinkhorn limits]{An explicit determinantal formula for Sinkhorn limits of matrices and tensors} \author{Shankar Sivarajan} \makeatletter \g@addto@macro\@setauthors{% \begin{center}\footnotesize Department of Computational Medicine and Bioinformatics\\ University of Michigan, Ann Arbor, MI, USA\\ \href{mailto:sivarajn@umich.edu}{\texttt{sivarajn@umich.edu}} \end{center}} \makeatother \begin{document} \begin{abstract} We give an explicit formula, as a single determinant, for the polynomial satisfied by any entry of the limit obtained by scaling a positive tensor of format $n_1\times\dots\times n_r$ to arbitrary prescribed one-dimensional margins (Sinkhorn--Knopp scaling, Kruithof's method, iterative proportional fitting). The polynomial is the multilinear resultant, on the Segre variety $\PP^{n_1-1}\times\dots\times\PP^{n_r-1}$, of the linear forms expressing the margin conditions. Its degree is the multinomial coefficient $N=\binom{d}{n_1-1,\dots,n_r-1}$ with $d=\sum_k(n_k-1)$, and for generic tensors its roots are the corresponding entries of all complex solutions of the scaling equations. A Sylvester matrix turns it into an explicit $N\times N$ generalized eigenvalue problem, whose null vector recovers the whole limit. The leading coefficient is the principal $A$-determinant of the product of simplices, that is, the product of the hyperdeterminants of all faces of the tensor. The constant term is likewise a product of face hyperdeterminants, and the faces avoiding the chosen entry split off as a factor independent of that entry. For matrices we also give a short proof, in Pl\"ucker coordinates, of a diagonal form of the eliminant. For every positive tensor, the limit is the unique root for which the Sylvester matrix has a positive null vector of monomials. Generically, the other roots are the corner entries of the solutions of the same scaling equations with complex scaling factors. \end{abstract} \maketitle \section{Introduction} Let $T$ be a positive tensor of format $n_1\times\dots\times n_r$, and for each mode $k$ let $\mu^{(k)}\in\R^{n_k}_{>0}$ be target margins with a common total $\tau$. Repeatedly rescaling the slices of $T$ so that its mode-$k$ margins equal $\mu^{(k)}$ converges to a tensor \[ P=T\circ\bigl(u^{(1)}\otimes\dots\otimes u^{(r)}\bigr),\qquad \textstyle\sum_{c:\,c_k=i}P_c=\mu^{(k)}_i\quad(1\le k\le r,\ 1\le i\le n_k), \] which is the unique tensor of this form with these margins~\cite{Kruithof,Sinkhorn,FranklinLorenz}. For $r=2$ this is matrix scaling, known as the Sinkhorn--Knopp algorithm, as Kruithof's double-factor method, and, in statistics, as iterative proportional fitting; it also arises in entropically regularized optimal transport, where the marginals are arbitrary. Everything below allows arbitrary positive margins; the special targets $\mu^{(1)}=\ones$, $\mu^{(2)}=\frac mn\ones$ give what Rowland and Wu call the Sinkhorn limit. The entries of $P$ are algebraic over the entries of $T$ and the margins. For $2\times2$ matrices they involve square roots~\cite{Nathanson,Fang25}. Rowland and Wu~\cite{RW} conjectured, on the basis of extensive computation, that the top-left entry of the limit of an $m\times n$ matrix is a root of an explicit polynomial of degree $\binom{m+n-2}{m-1}$; Fang~\cite{Fang25} proved the degree bound by viewing the scaling factors as a point of a Segre variety, and later proved the full conjecture~\cite{Fang26}. We extend this to tensors of any order with arbitrary margins. Write $y=u^{(1)}\otimes\dots\otimes u^{(r)}$. The margin conditions are \emph{linear} in $y$, and $y$ ranges over the cone over the Segre variety $\Sigma=\PP^{n_1-1}\times\dots\times\PP^{n_r-1}$, of dimension $d=\sum_k(n_k-1)$. After normalization the conditions are $d$ multilinear forms $\ell_1,\dots,\ell_d$; adding the form $\ell_x$ that fixes the corner entry to $x$ gives $d+1$ forms on a $d$-dimensional variety. For generic $T$ they have a common zero exactly when $x$ is the corner entry of a solution. The number of complex solutions, $\deg\Sigma$, is known: it is the maximum likelihood degree of the scaled Segre variety~\cite{ABBGHHNRS}. Our main result (Theorem~\ref{thm:main}) is that the multilinear resultant $\Res_\Sigma(\ell_1,\dots,\ell_d,\ell_x)$ is, up to a factor independent of $x$, the product of $x-x_i$ over all complex solutions; by a theorem of Sturmfels and Zelevinsky~\cite{SZ} it is a single Sylvester determinant, of size $(d+1)N$ with $x$ in exactly $N=\binom{d}{n_1-1,\dots,n_r-1}=\deg\Sigma$ columns, so that the corner entry is an eigenvalue of an explicit $N\times N$ pencil (Corollary~\ref{cor:pencil}). Section~\ref{sec:structure} identifies the leading coefficient with the principal $A$-determinant (a product of hyperdeterminants of faces), shows that the constant term is also such a product, and determines the content. Section~\ref{sec:matrix} treats matrices, where the pencil has a diagonal form; this form was conjectured by Rowland and Wu~\cite{RW} and proved by Fang~\cite{Fang26}, and we give a short proof in Pl\"ucker coordinates. Theorem~\ref{thm:select} characterizes the limit among the roots for every positive $T$: it is the only root for which $\Phi$ has a left null vector of monomials at a positive point. Section~\ref{sec:other} accounts for the other roots, which come from the same scaling equations with complex factors (equivalently, from the other critical points of the dual potential); the real ones are positive scalings of sign-twisted copies of $T$. Appendix~\ref{app:examples} gives numerical examples, including cases in which several roots lie in the range of possible values of the corner entry and a degenerate tensor of reduced degree. \section{The Segre formulation}\label{sec:segre} Fix the format $\mathbf n=(n_1,\dots,n_r)$ with all $n_j\geq2$, write $k_j=n_j-1$, $d=k_1+\dots+k_r$, and \[ N=\binom{d}{k_1,\dots,k_r}=\deg\Sigma. \] Index cells by $c=(c_1,\dots,c_r)$, write $\one=(1,\dots,1)$ for the corner, $t_c$ for the entries of $T$, and $y_c=u^{(1)}_{c_1}\cdots u^{(r)}_{c_r}$ for the Segre coordinates. Let $\sigma(y)=\sum_ct_cy_c$ and define the multilinear forms \[ \ell_{k,i}(y)=\sum_c\Bigl(\delta_{c_k,i}-\frac{\mu^{(k)}_i}{\tau}\Bigr)t_c\,y_c\quad(1\le k\le r,\ 2\le i\le n_k), \qquad \ell_x(y)=t_\one y_\one-\frac{x}{\tau}\,\sigma(y). \] There are $d$ forms $\ell_{k,i}$; we list them as $\ell_1,\dots,\ell_d$. \begin{lemma}\label{lem:segre} A point $y\in\Sigma$ with $\sigma(y)\neq0$ is a common zero of $\ell_1,\dots,\ell_d$ if and only if $P=\frac{\tau}{\sigma(y)}\,T\circ y$ has margins $\mu^{(1)},\dots,\mu^{(r)}$. In that case $\ell_x(y)=0$ exactly when $x=P_\one$. \end{lemma} \begin{proof} The mode-$k$ margin of $P$ at $i$ is $\frac\tau{\sigma(y)}\sum_{c_k=i}t_cy_c$, and $\ell_{k,i}(y)=\frac{\sigma(y)}{\tau}\bigl(\text{margin}-\mu^{(k)}_i\bigr)$. The forms with $i=1$ are omitted because the margins in each mode sum to $\tau$ automatically. Finally, $\ell_x(y)=\frac{\sigma(y)}{\tau}(P_\one-x)$. \end{proof} Thus the complex solutions of the scaling problem are the points of $Z=\Sigma\cap\{\ell_1=\dots=\ell_d=0\}$ with $\sigma\neq0$, and the scaling limit is the unique positive one. \begin{lemma}\label{lem:generic} Fix margins with $\tau\neq0$ and all $\mu^{(k)}_i\neq0$ (for instance positive). For $T$ outside a proper Zariski-closed subset, $Z$ lies in the dense torus $\Sigma^\circ$ of $\Sigma$ and consists of $N$ simple points, at each of which $\sigma\neq0$. \end{lemma} \begin{proof} Each $y\in\Sigma$ lies in the torus orbit $O_F$ of a unique face $F=F_1\times\dots\times F_r$ ($\emptyset\neq F_j\subseteq[n_j]$), and $y_c\neq0$ exactly for $c\in F$. For fixed $y\in O_F$ the $\ell_{k,i}(y)$ are linear forms in $(t_c)_{c\in F}$ with coefficient vectors $y_c\bigl(\chi_{k,i}-\frac{\mu^{(k)}_i}{\tau}\one\bigr)|_F$, where $\chi_{k,i}$ is the indicator of the slice $c_k=i$. The vectors $\chi_{k,i}|_F$ ($i\in F_k$, all $k$) span a space of dimension $\dim F+1$ containing $\one_F$. If $F\neq[n_1]\times\dots\times[n_r]$, choose $k$ with $F_k\neq[n_k]$: either some $i\geq2$ is not in $F_k$, and then the $(k,i)$ vector is $-\frac{\mu^{(k)}_i}\tau\one_F\neq0$, or $F_k=\{2,\dots,n_k\}$, and then the $(k,i)$ vectors for $i\geq2$ sum to $\frac{\mu^{(k)}_1}\tau\one_F\neq0$. In both cases $\one_F$, and hence every $\chi_{k,i}|_F$, lies in the span of our vectors, so the forms impose $\dim F+1$ conditions on $T$. The incidence variety $\{(T,y):y\in O_F,\ \ell_1(y)=\dots=\ell_d(y)=0\}$ therefore has dimension less than the number of cells, and for generic $T$ we get $Z\cap O_F=\emptyset$. For $F$ the whole box the same computation gives $d$ independent conditions, so $I=\{(T,y):y\in\Sigma^\circ,\ \ell_1=\dots=\ell_d=0\}$ is a vector bundle over $\Sigma^\circ$, smooth and irreducible, of dimension equal to the number of cells. Its projection to $T$ is dominant, since for the rank-one tensor $T_c=\tau^{1-r}\prod_k\mu^{(k)}_{c_k}$ the point $y=\one$ lies in $Z$ and is a simple zero (for positive $\mu$, every positive $T$ has a positive solution), so it is generically finite and, by generic smoothness, has reduced generic fibers. With no zeros on the boundary of $\Sigma$, Bernstein's theorem~\cite{Bernstein} counts the torus zeros, with multiplicity, as the normalized volume of $\Delta_{n_1-1}\times\dots\times\Delta_{n_r-1}$, which is $\deg\Sigma=N$. Finally, a point of $Z\cap\Sigma^\circ$ with $\sigma=0$ annihilates $f=\sum_ct_c\prod_ju^{(j)}_{c_j}$ and all of its logarithmic derivatives $u^{(k)}_i\partial f/\partial u^{(k)}_i$, so it is a singular point of $\{f=0\}$ in the torus; that happens only for $T$ in a proper closed subset (the closure of this set is the dual variety). \end{proof} \section{The main theorem}\label{sec:main} For a multidegree $a=(a_1,\dots,a_r)$ let $S(a)$ be the space of polynomials of multidegree $a$ in the variables $u^{(1)},\dots,u^{(r)}$, so $\dim S(a)=\prod_j\binom{a_j+k_j}{k_j}$. Put $K_0=0$, $K_j=k_1+\dots+k_j$ ($j\ge1$), and \[ m=(m_1,\dots,m_r),\qquad m_j=1+K_{j-1}. \] Each $\ell_s$ has multidegree $\one$. The \emph{Sylvester map} is \[ \Phi(x):\ S(m-\one)^{\oplus(d+1)}\longrightarrow S(m),\qquad (g_1,\dots,g_d,g_x)\longmapsto \sum_{s=1}^dg_s\ell_s+g_x\ell_x. \] Explicitly, in monomial bases (rows indexed by monomials $u^a$ of multidegree $m$, columns by pairs $(s,u^b)$ with $u^b$ of multidegree $m-\one$), the entry in row $u^a$ and column $(s,u^b)$ is the coefficient of $y_c$ in $\ell_s$ if $u^a=u^b\,y_c$ for a cell $c$, and $0$ otherwise. Thus every entry is $0$, $(\delta_{c_k,i}-\mu^{(k)}_i/\tau)\,t_c$, or $\delta_{c,\one}t_\one-(x/\tau)\,t_c$. \begin{lemma}\label{lem:square} $\Phi(x)$ is square, of size $(d+1)N$, and $x$ enters only the $N=\dim S(m-\one)$ columns of the last summand, linearly. \end{lemma} \begin{proof} $\dim S(m-\one)=\prod_j\binom{K_j}{k_j}=N$. Since $\binom{1+K_j}{k_j}=\binom{K_j}{k_j}\frac{1+K_j}{1+K_{j-1}}$, the product telescopes: $\dim S(m)=N\prod_j\frac{1+K_j}{1+K_{j-1}}=(d+1)N$. \end{proof} \begin{theorem}\label{thm:main} For every $T$ and $\mu$, $\det\Phi(x)=\pm\Res_\Sigma(\ell_1,\dots,\ell_d,\ell_x)$, and when $T$ and $\mu$ are positive this vanishes at $x=P_\one$. Suppose moreover that $Z$ is contained in the dense torus of $\Sigma$ (by Lemma~\ref{lem:generic} this holds for generic $T$). Then \[ \det\Phi(x)=\pm\Res_\Sigma(\ell_1,\dots,\ell_d,\ell_x) =C\prod_{\xi\in Z}\ell_x(\xi)^{\operatorname{mult}_\xi} =C'\prod_{\xi\in Z,\ \sigma(\xi)\neq0}\bigl(x-x_\xi\bigr)^{\operatorname{mult}_\xi}, \] where $\operatorname{mult}_\xi$ is the intersection multiplicity, $\sum_\xi\operatorname{mult}_\xi=N$, $x_\xi=\tau\,t_\one\xi_\one/\sigma(\xi)$ is the corner entry of the solution $\xi$, and $C\neq0$ and $C'$ do not depend on $x$. For generic $T$ the polynomial $\det\Phi(x)$ has degree exactly $N$, with the corner entries of the $N$ complex solutions as its roots. \end{theorem} \begin{proof} $\Res_\Sigma$ is the resultant of type $(k_1,\dots,k_r;1,\dots,1)$: an irreducible polynomial in the coefficients of $d+1$ multilinear forms that vanishes exactly when they have a common zero on $\Sigma$~\cite[Ch.~13, \S2A]{GKZ}. Since $\min(k_j,1)=1$ for all $j$, by Sturmfels and Zelevinsky~\cite{SZ} (see~\cite[Ch.~13, Prop.~2.2 and (2.3)]{GKZ}) the resultant equals the determinant of the Sylvester map $S(m^\pi-\one)\otimes\C^{d+1}\to S(m^\pi)$ for each of the $r!$ vectors $m^\pi_j=1+\sum_{\pi(i)<\pi(j)}k_i$, one for each permutation $\pi$ of the factors. For the identity permutation this is our $m$, and $\det\Phi=\pm\Res_\Sigma$ identically in the coefficients of the forms; being a polynomial identity, it holds for our specialized coefficients. At $x=P_\one$ the forms have the common zero $y^*$ given by the limit, so the resultant vanishes. When $Z$ lies in the torus (it is then finite, being closed in $\Sigma$), the Poisson formula for sparse resultants~\cite[Thm.~1.1]{DS} gives $\Res_\Sigma(\ell_1,\dots,\ell_d,f)=C\prod_{\xi\in Z}f(\xi)^{\operatorname{mult}_\xi}$ for every $f$ of multidegree $\one$. Here $C$ is a product of powers of directional resultants of $\ell_1,\dots,\ell_d$; these are nonzero because $Z$ meets no boundary orbit, and $C$ does not depend on $f$. By Bernstein's theorem~\cite{Bernstein} the total multiplicity is the normalized volume of $\Delta_{n_1-1}\times\dots\times\Delta_{n_r-1}$, that is, $\deg\Sigma=N$. For $f=\ell_x$ each factor is $\ell_x(\xi)=t_\one\xi_\one-\frac x\tau\sigma(\xi)$, which is $-\frac{\sigma(\xi)}\tau(x-x_\xi)$ if $\sigma(\xi)\neq0$ and independent of $x$ otherwise. By Lemma~\ref{lem:segre} the $x_\xi$ are the corner entries of the complex solutions. The last sentence follows from Lemma~\ref{lem:generic}. \end{proof} \begin{corollary}[Eigenvalue form]\label{cor:pencil} Write $\Phi(x)=[\,M_0\mid K-xJ\,]$, where $M_0$ collects the first $dN$ columns and $K-xJ$ the last $N$. Whenever $\det\Phi\not\equiv0$ (in particular for generic $T$), $M_0$ has full column rank; let the columns of $Z_0$, with entries in $\Q(T,\mu)$, span its left kernel. Then \[ \det\Phi(x)=c\,\det\bigl(Z_0^{\!\top}K-x\,Z_0^{\!\top}J\bigr), \] so $P_\one$ is a generalized eigenvalue of an explicit $N\times N$ pencil. More generally, let $\xi\in Z$ with $\sigma(\xi)\neq0$, and suppose $x_\xi$ is a simple root of $\det\Phi$. Then the left null vector of $\Phi(x_\xi)$ is, up to scale, the evaluation $h\mapsto h(\xi)$ on $S(m)$. So it determines every ratio $u^{(k)}_i/u^{(k)}_{i'}$ of $\xi$, and hence the whole solution $P_\xi=\frac{\tau}{\sigma(\xi)}T\circ\xi$, with entries in $\Q(T,\mu)(x_\xi)$. For $\xi=y^*$ this recovers the limit from $P_\one$. \end{corollary} \begin{proof} Multiply $\Phi$ on the left by an invertible matrix whose last $N$ rows are $Z_0^{\!\top}$; the result is block triangular. The functional $h\mapsto h(\xi)$ is nonzero, and it kills every $\sum g_s\ell_s+g_x\ell_{x_\xi}$ because $\ell_1,\dots,\ell_d$ and $\ell_{x_\xi}$ all vanish at $\xi$. When the root is simple the left kernel is one-dimensional. Since $m_k\geq1$ for every $k$, the monomials of $S(m)$ include pairs differing by one variable of each mode, and the corresponding ratios of coordinates of the null vector are the ratios $u^{(k)}_i/u^{(k)}_{i'}$. \end{proof} \begin{theorem}[The root of the limit]\label{thm:select} Let $T$ and the margins be positive, let $P^*$ be the limit, and let $y^*\in\Sigma$ be its Segre point. For $y\in\Sigma$ write $\operatorname{ev}_y=(y^a)_a$ for the vector of monomials of multidegree $m$ evaluated at $y$, a row vector indexed like the rows of $\Phi$. \begin{enumerate} \item $P^*$ is the only solution with all entries positive, and $y^*$ is a simple point of $Z$. \item For $x\in\C$, the left kernel of $\Phi(x)$ contains $\operatorname{ev}_y$ for some $y\in\Sigma$ with all coordinates positive if and only if $x=P^*_\one$, and then $y=y^*$. In particular $\det\Phi(P^*_\one)=0$. \item If $P^*_\one$ is a simple root of $\det\Phi$, its left null vector is, up to scale, $\operatorname{ev}_{y^*}$, which has all entries positive. If moreover $Z$ lies in the torus and all roots are simple, then $P^*_\one$ is the only root of $\det\Phi$ whose left null vector can be scaled to have all entries positive. \end{enumerate} \end{theorem} \begin{proof} (1) Existence and uniqueness of the positive scaling are classical~\cite{Sinkhorn,FranklinLorenz}. To show that $y^*$ is a simple zero, scale $y^*$ so that $T\circ y^*=P^*$, and use the coordinates $a=\log u$, writing tangent vectors as $\alpha=(\alpha^{(1)},\dots,\alpha^{(r)})$. The directions with $\alpha^{(k)}=\lambda_k\one$ for every $k$ are tangent to the fibers of the cone over $\Sigma$, so $T_{y^*}\Sigma$ is the quotient by them. The derivative of the map $a\mapsto(\text{all margins of }T\circ e^a)$ at $y^*$ is the symmetric matrix $A$ of the form $q(\alpha)=\sum_cP^*_c\bigl(\sum_k\alpha^{(k)}_{c_k}\bigr)^2$. Since $P^*>0$, $A$ is positive semidefinite, and $A\alpha=0$ exactly when $\sum_k\alpha^{(k)}_{c_k}=0$ for all $c$, that is, when $\alpha^{(k)}=\lambda_k\one$ with $\sum_k\lambda_k=0$. Now let $\alpha$ satisfy $d\ell_{k,i}(\alpha)=0$ for $i\geq2$. Summing over $i$ gives $\sum_i\ell_{k,i}=0$ identically when $\ell_{k,1}$ is defined by the same formula, so also $d\ell_{k,1}(\alpha)=0$. As $\ell_{k,i}=M_{k,i}-\frac{\mu^{(k)}_i}{\tau}\sigma$ with $M_{k,i}=\sum_{c_k=i}t_cy_c$, this reads $(A\alpha)^{(k)}_i=\frac{\mu^{(k)}_i}{\tau}\,d\sigma(\alpha)$ for all $k,i$. Let $e$ be the vector with $e^{(1)}=\one$ and $e^{(k)}=0$ for $k\geq2$, so that $Ae=(\mu^{(1)},\dots,\mu^{(r)})$. Then $A\bigl(\alpha-\frac{d\sigma(\alpha)}{\tau}e\bigr)=0$, so every $\alpha^{(k)}$ is a multiple of $\one$, and $\alpha$ is zero in $T_{y^*}\Sigma$. Thus $\ell_1,\dots,\ell_d$ have independent differentials at $y^*$, and $y^*$ is a simple zero. (2) The entry of $\operatorname{ev}_y\Phi(x)$ in the column $(s,u^b)$ is $y^b\ell_s(y)$, with $\ell_{d+1}=\ell_x$. If $y$ is positive then $y^b>0$, so $\operatorname{ev}_y$ lies in the left kernel exactly when $\ell_1(y)=\dots=\ell_d(y)=\ell_x(y)=0$. Since $\sigma(y)>0$, Lemma~\ref{lem:segre} shows that this means $\frac\tau{\sigma(y)}T\circ y$ is a positive solution with corner entry $x$, hence equals $P^*$ by (1). (3) A simple root has a one-dimensional left kernel, which by (2) is spanned by $\operatorname{ev}_{y^*}$. For another simple root $x_\xi$, the left null vector is $\operatorname{ev}_\xi$ by Corollary~\ref{cor:pencil}. If it can be scaled to be positive, then all ratios of monomials at $\xi$ are positive. These include all ratios $u^{(k)}_i/u^{(k)}_{i'}$, so $\xi$ is positive and $\xi=y^*$. \end{proof} Thus the limit is distinguished from the other roots by positivity of the whole solution, read off from the null vector of $\Phi$. When $Z$ lies in the torus, (1) shows that the factor $x-P^*_\one$ occurs in $\det\Phi$ with multiplicity one unless another complex solution has the same corner entry. Since Theorem~\ref{thm:main} applies to any entry after permuting the indices, every entry of the limit is a root of a determinant of this kind; when $P^*_\one$ is a simple root of $\det\Phi$, a single corner determinant and its left null vector already recover all entries of $P^*$ via Corollary~\ref{cor:pencil}. \begin{remark} Theorem~\ref{thm:main} needs no genericity beyond the assumption on $Z$; there are $r!$ admissible multidegrees, one for each ordering of the factors, and all give the same determinant up to sign. \end{remark} \section{Leading and constant coefficients}\label{sec:structure} Write $R(x)=\det\Phi(x)$. We treat the entries $t_c$ as indeterminates and the margins as coordinates on the linear space of margin vectors with a common total $\tau$, so the coefficients of $R$ lie in $\mathcal R=\Q[T,\mu][\tau^{-1}]$; ``up to a unit'' means up to a factor $q\tau^e$ with $q\in\Q^\times$, $e\in\Z$. Let $A=\{e_{c}\}\subset\Z^{n_1}\oplus\dots\oplus\Z^{n_r}$ be the vertex set of $\Delta_{n_1-1}\times\dots\times\Delta_{n_r-1}$. For a face $F=F_1\times\dots\times F_r$ ($\emptyset\neq F_j\subseteq[n_j]$) let $T_F$ be the corresponding subtensor and $\Disc_F(T)$ its $A$-discriminant $\Delta_{A\cap F}$: $t_c$ for a cell, the $k\times k$ minor for a square face of a matrix, Cayley's hyperdeterminant for a $2\times2\times2$ face, and in general the hyperdeterminant of $T_F$ when it is nontrivial (and $1$ otherwise)~\cite{GKZ}. The principal $A$-determinant is $E_A(T)=\prod_F\Disc_F(T)$, with all multiplicities equal to one because the toric variety $X_A=\Sigma$ is smooth~\cite[Ch.~10, Thm.~1.2]{GKZ}. \begin{theorem}\label{thm:lead} Up to a unit, the leading coefficient of $R(x)$ is $E_A(T)$; in particular it does not depend on the margins except through a power of $\tau$. \end{theorem} \begin{proof} The coefficient of $x^N$ is $(-1/\tau)^N\Res_\Sigma(\ell_1,\dots,\ell_d,\sigma)$. A resultant is unchanged by adding to one form a multiple of another of the same multidegree, so we may replace $\ell_{k,i}$ by $M_{k,i}(y)=\sum_{c_k=i}t_cy_c$, which removes $\mu$. Write $f(u)=\sum_ct_c\prod_ju^{(j)}_{c_j}$, the polynomial with support $A$ and coefficients $T$. Then $\sigma=f$ and $M_{k,i}=u^{(k)}_i\,\partial f/\partial u^{(k)}_i$, and together these span the space of all logarithmic derivatives of $f$ and $f$ itself (the omitted $i=1$ derivatives are determined by the Euler relations $\sum_iu^{(k)}_i\partial f/\partial u^{(k)}_i=f$). The $A$-resultant is the resultant on the toric variety $X_A=\Sigma$, so $\Res_\Sigma(M,\sigma)$ is, up to a nonzero constant, the principal $A$-determinant $E_A(f)=R_A\bigl(u\,\partial f/\partial u,f\bigr)$ as defined in~\cite[Ch.~10, (1.1)]{GKZ}. By~\cite[Ch.~10, Thm.~1.2]{GKZ}, since $X_A=\Sigma$ is smooth, it is the product of the discriminants of all faces, each to the first power. \end{proof} \begin{corollary} The degree of $R$ is less than $N$ exactly on the hypersurface $E_A(T)=0$, that is, when some face of $T$ has vanishing hyperdeterminant (or minor, or entry). \end{corollary} The faces through the corner and the others play different roles. \begin{proposition}\label{prop:content} If $F$ does not contain the corner, then $\Disc_F(T)$ divides every coefficient of $R(x)$. \end{proposition} \begin{proof} For generic $T$ with $\Disc_F(T)=0$, the restriction of $f$ to the face $F$ has a singular point in its torus, since by definition $\{\Disc_F=0\}$ is the closure of the set of such coefficient vectors~\cite[Ch.~9, (1.1) and Def.~1.2]{GKZ}. This gives a point $\xi\in\Sigma$ in the torus orbit of $F$ at which $f$ and all its logarithmic derivatives vanish, that is, all margins of $T\circ\xi$ are zero. So $\sigma(\xi)=0$ and every $\ell_{k,i}(\xi)=0$. Since $\one\notin F$, also $\xi_\one=0$, hence $\ell_x(\xi)=0$ for every $x$, and $R\equiv0$. As this holds on a dense subset of the hypersurface $\Disc_F=0$, the irreducible polynomial $\Disc_F$ divides $R$ as a polynomial in $T$ and $x$, and being free of $x$ it divides each coefficient. \end{proof} \begin{proposition}\label{prop:const} For $1\le k\le r$ let $T^{(k)}$ be the slab $\{t_c:c_k\geq2\}$, of format $n_1\times\dots\times(n_k-1)\times\dots\times n_r$, let $A_k$ be its vertex configuration, and let $N_k=\binom{d-1}{k_1,\dots,k_k-1,\dots,k_r}$ be the degree of $\Sigma_k=\Sigma\cap\{u^{(k)}_1=0\}$. Up to a unit, \[ R(0)=t_\one^N\prod_{k=1}^r\bigl(\mu^{(k)}_1\bigr)^{N_k}E_{A_k}\bigl(T^{(k)}\bigr). \] \end{proposition} \begin{proof} At $x=0$, $\ell_x=t_\one\,u^{(1)}_1\cdots u^{(r)}_1$. The resultant is multiplicative in its last argument, and $\Res_\Sigma(\ell_1,\dots,\ell_d,u^{(k)}_1)=\pm\Res_{\Sigma_k}(\ell_1|_{\Sigma_k},\dots,\ell_d|_{\Sigma_k})$, so $R(0)=\pm t_\one^N\prod_kR_k$ with $R_k$ the latter resultant. On $\Sigma_k$ only cells of $T^{(k)}$ survive. Write $\sigma_k$ and $M^{(k)}_{j,i}=\sum_{c\in T^{(k)},\,c_j=i}t_cy_c$ for the restricted total and margin forms. For mode $k$ the restricted forms $\ell_{k,i}|_{\Sigma_k}=M^{(k)}_{k,i}-\frac{\mu^{(k)}_i}\tau\sigma_k$ ($2\le i\le n_k$) sum to $\frac{\mu^{(k)}_1}{\tau}\sigma_k$, and the matrix taking $(M^{(k)}_{k,i})_{i\geq2}$ to them has determinant $\mu^{(k)}_1/\tau$. So, after a triangular change of basis of determinant $\mu^{(k)}_1/\tau$, the $d$ restricted forms become $\sigma_k$, $M^{(k)}_{k,i}$ ($i\geq3$) and $M^{(k)}_{j,i}$ ($j\neq k$, $i\geq2$); the other modes are handled by subtracting multiples of $\sigma_k$. Since $\Res_{\Sigma_k}$ has degree $N_k$ in each argument, $R_k=\pm(\mu^{(k)}_1/\tau)^{N_k}\Res_{\Sigma_k}(M^{(k)},\sigma_k)$, and the last resultant is, up to a nonzero constant, $E_{A_k}(T^{(k)})$, exactly as in the proof of Theorem~\ref{thm:lead}. \end{proof} Thus both extreme coefficients of $R$ are products of face discriminants. In particular $R(0)\neq0$ for generic $T$ and any margins with $\mu^{(k)}_1\neq0$. \begin{theorem}\label{thm:content} Fix margins with $\tau\neq0$ and all $\mu^{(k)}_i\neq0$. As a polynomial in $x$, $R(x)$ has content, up to a unit, \[ c(T)=t_\one\prod_{F\not\ni\one}\Disc_F(T), \] the product over all faces not containing the corner (including the vertices $t_c$, $c\neq\one$). Consequently $R=c(T)\,\tilde E_T(x)$, where $\tilde E_T$ is primitive of degree $N$; up to a unit, its leading coefficient and constant term are \[ \prod_{F\ni\one,\ F\neq\{\one\}}\Disc_F(T) \qquad\text{and}\qquad t_\one^{N-1}\prod_k\bigl(\mu^{(k)}_1\bigr)^{N_k}\prod_{F\not\ni\one}\Disc_F(T)^{\,e_F-1}, \] where $e_F=\#\{k:1\notin F_k\}$. Moreover $\tilde E_T$ is a power of an irreducible polynomial in $\C[T,x]$. \end{theorem} \begin{proof} The content divides the leading coefficient $E_A(T)$ (Theorem~\ref{thm:lead}). This is squarefree and free of $\mu$, and it is a product of irreducible face discriminants that are pairwise non-associate, since each $\Disc_F$ involves exactly the variables $t_c$, $c\in F$ (see below). So $c=\prod_{F\in S}\Disc_F$ for some set $S$ of faces. Proposition~\ref{prop:content} gives $F\in S$ for every $F\not\ni\one$. At $t_\one=0$ we have $\ell_x=-\frac x\tau\sigma$, so $R=(-x/\tau)^NE_A(T)|_{t_\one=0}=0$, since $t_\one$ is a factor of $E_A$; hence $\{\one\}\in S$. Now let $F\ni\one$ be a face of positive dimension with nontrivial discriminant. $\Disc_F$ is irreducible~\cite[Ch.~9, Def.~1.2]{GKZ} and invariant up to sign under permutations of the indices within each mode of $F$. These permutations act transitively on the cells of $F$, so the nonconstant polynomial $\Disc_F$ involves every $t_c$ with $c\in F$, in particular $t_\one$. By Proposition~\ref{prop:const}, $R(0)$ is a nonzero polynomial whose only dependence on $t_\one$ is the factor $t_\one^N$, since the slabs do not contain the corner. An irreducible polynomial involving $t_\one$ and not associate to $t_\one$ cannot divide it. Hence $F\notin S$. The formula for $\tilde E_T(0)=R(0)/c(T)$ follows from Proposition~\ref{prop:const}. Every face $F\not\ni\one$ lies in exactly the $e_F\geq1$ slabs $T^{(k)}$ with $1\notin F_k$, and $E_{A_k}(T^{(k)})$ is the product of the discriminants of the faces of $T^{(k)}$. For the last claim, let $I$ be the incidence variety of Lemma~\ref{lem:generic}. It is irreducible, and its projection to $T$ has degree $N$ with reduced generic fibers. By Theorem~\ref{thm:main}, $\tilde E_T$ is, up to its leading coefficient, the norm of $X-x$ from the function field of $I$ to $\C(T)$, which is a power of the minimal polynomial of $x$. \end{proof} \begin{remark}[Degenerate tensors] For special positive $T$ the full determinant $R$ can vanish identically. For instance, by Proposition~\ref{prop:content} this happens for a $3\times3$ matrix with a vanishing $2\times2$ minor away from the $(1,1)$ entry. The right object is therefore $\tilde E_T$. Since the limit depends continuously on $T$, the identity $\tilde E_T(P_\one)=0$, valid for generic $T$, holds for \emph{every} positive $T$. The coefficients of $\tilde E_T$ have no common factor, so $\tilde E_T\not\equiv0$ outside a subset of codimension at least two. \end{remark} \section{Matrices}\label{sec:matrix} For $r=2$, Theorem~\ref{thm:main} gives an $N\times N$ pencil with $N=\binom{m+n-2}{m-1}$ whose determinant is the eliminant of the top-left entry. The faces with nontrivial discriminant are then exactly the square submatrices, so the coefficients in Section~\ref{sec:structure} are products of minors: the leading coefficient of $\tilde E_T$ is the product of the minors of size at least two that contain $t_{11}$, and the minors avoiding both the first row and the first column enter its constant term. Rowland and Wu~\cite{RW} found, and Fang~\cite{Fang26} proved, a diagonal form of this eliminant, in which the matrix enters only through minors on a diagonal and the margins only through an integer-coefficient matrix. Two regular pencils with the same simple eigenvalues are equivalent, so for generic $T$ the two forms differ by a change of basis; we give a short proof of this formula. \medskip \noindent\emph{Pl\"ucker coordinates and derivations.} Let $e_1,\dots,e_M$ be the standard basis of $\C^M$, totally ordered. For a $k$-subset $J=\{j_1<\dots0}$ with row targets $\rho=\mu^{(1)}$ and column targets $\kappa=\mu^{(2)}$. Index $\C^{m+n}$ by $1<\dots