Exactness of belief propagation for some graphical models with loops

Abstract. It is well known that an arbitrary graphical model of statistical inference defined on a tree, i.e. on a graph without loops, is solved exactly and efficiently by an iterative Belief Propagation (BP) algorithm convergent to unique minimum of the so-called Bethe free energy functional. For a general graphical model on a loopy graph the functional may show multiple minima, the iterative BP algorithm may converge to one of the minima or may not converge at all, and the global minimum of the Bethe free energy functional is not guaranteed to correspond to the optimal Maximum-Likelihood (ML) solution in the zero-temperature limit. However, there are exceptions to this general rule, discussed in 12 and 2 in two different contexts, where zerotemperature version of the BP algorithm finds ML solution for special models on graphs with loops. These two models share a key feature: their ML solutions can be found by an efficient Linear Programming (LP) algorithm with a Totally-Uni-Modular (TUM) matrix of constraints. Generalizing the two models we consider a class of graphical models reducible in the zero temperature limit to LP with TUM constraints. Assuming that a gedanken algorithm, g-BP, finding the global minimum of the Bethe free energy is available we show that in the limit of zero temperature g-BP outputs the ML solution. Our consideration is based on equivalence established between gapless Linear Programming (LP) relaxation of the graphical model in the T → 0 limit and respective LP version of the Bethe-Free energy minimization.