Toric invariant theory for maximum likelihood estimation in log-linear models (Q2076293): Difference between revisions

In this interesting paper, the authors study connections between the \emph{stability} of vectors under the action of an algebraic torus and maximum likelihood estimation (MLE) for log-linear models in algebraic statistics. A log-linear model corresponding to a matrix \(A\in \mathbb Z^{d\times m}\) is the set \(\mathcal M_A = \lbrace p\in \Delta_{m-1} \ | \ \log(p)\in \mathrm{rowspan}(A) \rbrace\), where \(\Delta_{m-1}\) is the \((m-1)\)-dimensional standard simplex. A vector of counts \(u\in \mathbb Z^m\), whose \(j\)-th coordinate is the number of times the \(j\)-th state is observed, gives an empirical distribution \(\overline{u}\in \Delta_{m-1}\). The maximum likelihood estimation (MLE) in the model \(\mathcal M_A\) finds a point most likely to give rise to the observed data \(u\). More precisely, given \(u\), an MLE is a maximizer \(\hat{p}\) of the likelihood function \(L_u(p) = p_1^{u_1}\cdots p_m^{u_m}\) over \(\mathcal M_A\). Consider a vector of counts \(u\in \mathbb Z^m\) with sample size \(n = \sum_{j=1}^m u_j\). Let \(A\in \mathbb Z^{d\times m}\) be a matrix with \(\mathbf{1}=(1,\ldots, 1)\) in its rowspan and \(b=Au\). The main theorem (Theorem 4.3) relates the stability of \(\mathbf{1}\) under the action of \((\mathbb C^{*})^d\), given by the matrix \(nA\) with linearization \(b\), to the MLE in \(\mathcal M_A\) as follows: \begin{itemize} \item \(\mathbf{1}\) is semistable if and only if extended MLE exists and is unique (here ``extended'' means that the estimation is allowed to be in the model euclidean closure \(\overline{\mathcal M_A}\)); \item \(\mathbf{1}\) is polystable if and only if MLE exists and is unique. \end{itemize} The authors also obtain a nice description of the MLE in terms of the \emph{moment map} in Theorem 4.7.