Expectation Maximization: Coordinate Ascent on the Lower Bound

The expectation-maximization algorithm is a classical foundation for learning latent variable models, famously used to fit Gaussian mixture models to clustered data.

This algorithm can be viewed as coordinate ascent on the evidence lower bound, a green surface that never rises above the purple surface of the marginal log-likelihood.

The rest of the narration is in the book's Companion.