We study conditional independence (CI) models in statistical theory, in the case of discrete random variables, from the point of view of algebraic geometry and matroid theory. Any CI model with hidden random variables corresponds to a variety defined by certain determinantal conditions on a matrix whose entries are probabilities of events involving the observed random variables. We show that any CI variety, and more generally any hypergraph variety, admits a matroid stratification. Our main motivation for studying decompositions of CI varieties is the realizability problem: given a collection of CI relations, the goal is to determine the existence of random variables that satisfy these constraints and violates the rest. We show that the realization spaces of CI models and the matroid varieties in their decompositions are closely related. We use ideas from incidence geometry, in particular point and line configurations, to find minimal decompositions of general hypergraph varieties in terms of matroid varieties, which are not necessarily irreducible by Mnev--Sturmfels universality theorem, and may have arbitrary singularities. We focus on various families of hypergraph varieties for which we explicitly compute an irredundant irreducible decomposition. Our main findings in this direction are threefold: (1) we describe minimal matroids of such hypergraphs; (2) we prove that the varieties of these matroids are irreducible and their union is the hypergraph variety; and (3) we show that every such matroid is realizable over real numbers. Our decomposition strategy gives immediate matroid interpretations of the irreducible components of many families of CI varieties in algebraic statistics, and unravels the symmetric structures in CI varieties which hugely simplifies the computations.
Download