We prove a new simple and explicit bound on the total variation distance between a measure $\pi\propto e^{-nf}$ on $\mathbb R^d$ and its Laplace approximation. The bound is proportional to $d/\sqrt n$, which has recently been shown to be the tight rate in terms of dimension dependence. Our bound holds under weak regularity conditions on $f$ and at least linear growth of $f$ at infinity. We then apply this bound to prove the first ever Bernstein-von Mises (BvM) theorems on the asymptotic normality of posterior distributions in the regime $n\gg d^2$. This improves on the tightest previously known condition, $n\gg d^3$. We establish the BvM for the following data-generating models: 1) exponential families, 2) arbitrary probability mass functions on $d+1$ states, and 3) logistic regression with Gaussian design. Our statements of the BvM are nonasymptotic, taking the form of explicit high-probability bounds. We also show in a general setting that the prior can have a stronger regularizing effect than previously known while still vanishing in the large sample limit.
翻译:暂无翻译