Estimating Mixture Distributions via Stochastic Mirror Descent

Stochastic mirror descent over M-component mixture distributions yields a family of cross-entropy-minimizing estimators that scale efficiently with the number of candidate components by choice of Bregman divergence, generalizing classical mixture estimators without significant computational overhead as component count grows.