Title:Nonparametric Bernstein-von Mises Phenomenon: A Tuning Prior Perspective

Abstract:Statistical inference on infinite-dimensional parameters in a Bayesian framework is investigated. The main contribution of this paper is to demonstrate that nonparametric Bernstein-von Mises theorem can be established in a {\em general} class of nonparametric regression models under a novel tuning prior (indexed by a non-random hyper-parameter). Surprisingly, this type of prior connects two important classes of statistical methods: nonparametric Bayes and smoothing spline. The intrinsic connection with smoothing spline greatly facilitates both theoretical analysis and practical implementation for nonparametric Bayesian inference. For example, we can employ generalized cross validation to select a proper tuning prior, under which the constructed credible regions/intervals are frequentist valid. This new methodology is supported by our simulations. The posterior contraction rate (under Sobolev norm) is derived as a by-product. A collection of technical tools such as RKHS theory, Cameron-Martin theorem and Gaussian correlation inequality are employed in this paper.