Interactive demo · Probability and Statistical Inference · Chapter 18

Pairwise versus global Markov properties: missing edges imply separation statements, but only for positive distributions

A graph's missing edges say "independent given all the other variables" (pairwise). For a positive distribution these statements imply that every separating set gives a conditional independence (global, Theorem 18.1), with the local statements in between. Panel (a) shows the graph, panel (b) the probability of every configuration of the binary variables (zeros in red), and panel (c) the conditional mutual information of every statement the graph makes: a statement holds exactly when its bar is empty (value 0). With identical copies of one coin the pairwise statements hold but the global ones fail.

Graph

Distribution

Coupling J on every edge

Field h on every vertex

Noise η (each copy flipped with probability η)

Which variables are copies

Sample size n for the tests

Try these

Ising: f(x) ∝ exp(J Σedges sisj + h Σ si) with s = 2x − 1, positive and factorised over the edges, hence Markov to the graph. Copies: a fair coin U; each copy equals U except with probability η; the other variables are independent fair coins. A statement A ⊥⊥ B | C holds exactly when I(A; B | C) = 0. G² = 2nI is the likelihood ratio statistic that a sample of size n with exactly these frequencies would give (for the data it is the actual statistic), with df = (2|A| − 1)(2|B| − 1)2|C|.

In the lecture notes