Marginal Likelihood / Evidence

Definition

The marginal likelihood, also called the evidence, is

$$p(x) = \int p(x, z)\,dz.$$

It normalizes the posterior and is often intractable in latent variable models.

Why it matters

It is the reason direct posterior computation is difficult, and why VI introduces a surrogate objective.