Interactive demo · Probability and Statistical Inference · Chapter 18

Cliques and potentials: a product over the maximal cliques has no interactions outside them

Every positive distribution of binary variables can be written uniquely as log f(x) = log f(0) + ΣS λS ∏v∈S xv, one interaction term λS per set of variables. The Hammersley-Clifford theorem says that f is Markov to a graph exactly when λS = 0 for every S that is not a clique. Here f is built as in (18.1) from random positive potentials on the maximal cliques (panel a). Panel (b) shows all the λS on a log scale: the non-cliques (red) are zero up to rounding error. Add an interaction between two non-adjacent variables and its term appears, and their pairwise independence breaks. Panel (c) shows that a clique potential is not the clique's marginal distribution.

Graph

Strength of the potentials s

Extra interaction on a non-edge

Its size δ

Try these

For each maximal clique C and each configuration of its variables, log ψC = s(2u − 1) with u from the seeded generator (seed in the readout). The extra factor is exp(δ xaxb). Interaction terms: λS = ΣU⊆S (−1)|S|−|U| log f(1U, 0rest). Bars below 10−16 are drawn at the floor. Conditional mutual informations as in the companion demo on Markov properties.

In the lecture notes