A general framework for Bayes structured linear models (Q2215762)

The paper under review provides a unified methodology and theory for both Bayes high dimensional statistics and Bayes nonparametric statistics in a general framework of structured linear models. The authors first introduce a unified view of various high dimensional and nonparametric models, and then propose a single prior distribution for all models in the considered framework. Optimal rates of convergence of the posterior distributions are established under appropriate conditions. The results directly lead to exact minimax posterior contraction rates in stochastic block model, biclustering, sparse linear regression, regression with group sparsity, multitask learning and dictionary learning. Moreover, a general posterior oracle inequality, which allows arbitrary model misspecification, is also derived. The main results are illustrated by examples ranging from nonparametric estimation to high dimensional statistics.