Marginal parameterizations of discrete models defined by a set of conditional independencies (Q604366)

The subject of the paper is consistency of an arbitrary collection of conditional independencies defined on a finite set of discrete random variables. It is supposed that each random variable has finite range and the joint distribution of all random variables is strictly positive. Consistency of a fixed collection of conditional independencies means that the set of all positive distributions satisfying these independencies has a smooth parametrization of proper dimension (or, in other words, forms a smooth submanifold of proper dimension in the space of distributions). It is well known that for a single conditional independency the log-linear parametrization is of this kind; in the general case of multiple conditional independencies the question is open. The authors suggest to consider log-linear parametrizations for certain collections of hierarchically ordered marginal distributions, which they call hierarchical and complete marginal parametrizations (HCMP). In terms of these HCMP they prove a criterion for a single conditional independency, which provides a class of smooth parametrizations. In the case of multiple independencies they prove a sufficient condition and a necessary condition for existence of smooth parametrizations. On this basis they propose a finite combinatoric algorithm that can determine inconsistency of the model (the absence of smooth parametrization), or determine consistency and provide a smooth HCMP, or, in certain cases, give no answer. The paper is supplied with a number of illustrative examples.