A marginal probability assigns probability to one variable after accounting for other possible values. In Bayes’ theorem, P(D) is the evidence: the total probability of observing D across competing hypotheses, used to normalize P(H|D). section 2.2.1
In probabilistic graphical models, belief propagation passes local probability summaries to estimate marginals. It is exact on loop-free graphs; on graphs with loops, updates are generally approximate and may not converge. section 10.3.2
Bayes’ theorem places the marginal probability of the data in the denominator, normalizing the posterior. section 2.2.1




