Concentration of tempered posteriors and of their variational approximations (Q2196229)

The authors consider variational Bayes (VB) fast Bayesian inference to compute variational approximations of the posterior for massive datasets where sampling schemes such as Markov chain Monte Carlo are computationally impractical. Variational Bayesian methods aim at approximating the posterior by a distribution in a tractable family \(F\). The authors propose a general oracle inequality that relates the quality of the VB approximation to the prior \(\pi\) and to the structure of \(F\). They provide a simple condition that allows to derive rates of convergence from this oracle inequality. Various examples are presented. It is shown that for parametric models with log-Lipschitz likelihood, Gaussian VB leads to efficient algorithms and consistent estimators. A matrix completion is demonstrated as a high-dimensional example, and a density estimation is presented as a nonparametric example.