\(\alpha\)-variational inference with statistical guarantees (Q2196198)

The authors research issues on point estimation where data are coming from ``mean-field or other variational approximations to a Bayesian posterior''. They derive that there is no reason for denying or preferring it with respect to using the true posterior in terms of accuracy. The affirmation seems to be true for a broad class of statistical models, particularly for the class of variational objective functions indexed by a temperature parameter. For the particular case of $\alpha = 1$ bounds for the risk characterize the minimax optimality of variational of the point estimates. The authors analyze the case $\alpha < 1$. Dealing with a prior mass condition supports the derivation of the risk bounds of Rényi's divergence. They consider some particular problems for illustrating the behavior of their proposal, and claim that under the fulfillment of all the hypotheses the rates of convergence are the same whenever $\alpha\in (0,1]$.