Nested covariance determinants and restricted trek separation in Gaussian graphical models

DOI10.3150/19-BEJ1179MaRDI QIDQ2203613zbMATH OpenFDO

Authors Mathias Drton, Elina Robeva, Luca Weihs

Publication date 7 October 2020

Published in Bernoulli (Search for Journal in Brave)

Full work available at URL https://arxiv.org/abs/1807.07561, https://projecteuclid.org/euclid.bj/1598493622

conditional independence covariance matrix graphical model open problems trek separation Verma constraint

Lists of open problems (00A27) Estimation in multivariate analysis (62H12) Probabilistic graphical models (62H22) Analysis of variance and covariance (ANOVA) (62J10) Algebraic statistics (62R01)

Abstract: Directed graphical models specify noisy functional relationships among a collection of random variables. In the Gaussian case, each such model corresponds to a semi-algebraic set of positive definite covariance matrices. The set is given via parametrization, and much work has gone into obtaining an implicit description in terms of polynomial (in-)equalities. Implicit descriptions shed light on problems such as parameter identification, model equivalence, and constraint-based statistical inference. For models given by directed acyclic graphs, which represent settings where all relevant variables are observed, there is a complete theory: All conditional independence relations can be found via graphical

d

-separation and are sufficient for an implicit description. The situation is far more complicated, however, when some of the variables are hidden (or in other words, unobserved or latent). We consider models associated to mixed graphs that capture the effects of hidden variables through correlated error terms. The notion of trek separation explains when the covariance matrix in such a model has submatrices of low rank and generalizes

d

-separation. However, in many cases, such as the infamous Verma graph, the polynomials defining the graphical model are not determinantal, and hence cannot be explained by

d

-separation or trek-separation. In this paper, we show that these constraints often correspond to the vanishing of nested determinants and can be graphically explained by a notion of restricted trek separation.

Review text

This paper is placed in the context of Gaussian graphical models [\textit{S. L. Lauritzen}, Graphical models. Oxford: Oxford Univ. Press (1996; Zbl 0907.62001)], and its focus is on \textit{acyclic directed mixed graphs} (ADMGs), which are an extension of the better known directed acylic graphs (DAGs). Specifically, the authors are interested in explaining the relations among the entries of a covariance matrix that belongs to such linear structural equation models. It is well known in the graphical models literature that probabilistic conditional independence (CI) statements among the entries of Gaussian random vectors correspond to the vanishing of specific minors of the covariance matrix, i.e., a CI statement holds if and only if the determinant of certain submatrix of the covariance matrix is zero. These are enough to completely describe the constraints in the covariance matrix of a DAG model, and correspond graphically to \textit{d-separation} [\textit{M. Studený}, Probabilistic conditional independence structures. London: Springer (2005; Zbl 1070.62001)]. However, in the presence of hidden variables (represented in the mixed graph by bidirected edges between observed nodes) this is no longer the case. In particular, there may be minors that vanish on the model that do not correspond to CI statements. The work of [\textit{S. Sullivant} et al., Ann. Stat. 38, No. 3, 1665--1685 (2010; Zbl 1189.62091)] introduced the concept of \textit{trek separation} and managed to characterize all vanishing minors of the covariance matrix. However, there may exist other polynomial constraints that hold for ADMG models that are not of this form, such as the (in)famous \textit{Verma constraint} [\textit{T. van Ommen} and \textit{J. M. Mooij}, ``Algebraic equivalence of linear structural equation models, in: Proceedings of the 33rd Annual Conference on Uncertainty in Artificial Intelligence (UAI-17) (2017), \url{arXiv:1807.03527}]. These kinds of relations appear to be harder to understand, and the main contribution of the paper under review is that while they cannot be expressed as minors of the covariance matrix, they can actually be expressed as determinants of matrices whose entries are determinants themselves: \textit{nested determinants}. Furthermore, in their main theorem the authors are able to explain the structure of many of these instances via the concept of \textit{restricted trek-separation}. Through several illuminating and well-chosen examples, the authors illustrate not only how the theory they develop applies to different kinds of ADMGs, but also how there are other graphical models, such as ones that now allow cycles, which have constraints that also admit a nested determinant representation (which is in general not unique). Indeed, these observations reveal that the story of nested determinants and graphical model relations is not fully understood yet, and the paper suggests many interesting open problems for follow-up work in this direction.

Recommendations

Cites work

Cited in

(10)

Describes a project that uses

Uses Software

TETRAD

This page was built for publication: Nested covariance determinants and restricted trek separation in Gaussian graphical models

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2203613)