Monotonic Likelihood Ascent of EM
The log-likelihood decomposes as log p(X | theta) = L(q, theta) + KL(q || p(Z | X, theta)); the E-step closes the KL gap to zero and the M-step raises the lower bound L, so via Jensen's inequality each full EM iteration can never decrease the data log-likelihood.
This Concept is waiting for its first lesson!
The log-likelihood decomposes as log p(X | theta) = L(q, theta) + KL(q || p(Z | X, theta)); the E-step closes the KL gap to zero and the M-step raises the lower bound L, so via Jensen's inequality each full EM iteration can never decrease the data log-likelihood.
Are you a teacher? Sign in to start contributing.
Sign In