Decomposing Conditional Independence Ideals with Hidden Variables: A Matroid-Theoretic Approach
Emiliano Liwski
TL;DR
The paper develops a matroid-theoretic framework for decomposing CI ideals with hidden variables by linking determinantal CI conditions to circuit and matroid varieties. Introducing quasi-paving matroids, it proves that irreducible components of relevant varieties are realizable as matroid varieties $V_M$, and it provides a decomposition along nice partitions of dependent hyperplanes. Specializations to grid-like matroids $G_{k,l}$ and line-configuration matroids $L_n$ yield explicit decompositions and, for certain parameters, complete descriptions, together with generating functions that count irreducible components. The work integrates algebraic statistics with combinatorial geometry, delivering concrete counting tools and structural insights for CI models and their algebraic invariants.
Abstract
We study a class of determinantal ideals arising from conditional independence (CI) statements with hidden variables. Such CI statements translate into determinantal conditions on a matrix whose entries represent the probabilities of events involving the observed random variables. Our main objective is to determine the irreducible components of the corresponding varieties and to provide a combinatorial or geometric interpretation of each. We achieve this by introducing a new approach rooted in matroid theory. In particular, we introduce a new class of matroids, which we term quasi-paving matroids, and show that the components of these determinantal varieties are precisely matroid varieties of quasi-paving matroids. Moreover, we derive generating functions that encode the number of irreducible components of these CI ideals.
