{"id":"arxiv/2610.35330","url":"https://arxiv.org/abs/2610.35330","tags":["math.ST","stat.ML"],"score":0,"title":"On the asymptotic behaviour of $α$-divergence generalized posteriors","domain":"math.ST","source":"arxiv","authors":["Ioar Casado-Telletxea","Omar Rivasplata"],"reviews":0,"subject":"math.ST","summary":"We study the asymptotic behaviour of the solutions to the unconstrained generalized variational inference problem regularized by an $α$-divergence. Under mild conditions, the minimizer is explicit up to a scalar normalizing constant, and we show that it admits a rescaling under which it converges in $L^1$ to a fixed non-Gaussian law, at a concentration scale around the empirical loss minimizer that departs from the typical $n^{-1/2}$ in the classical Bernstein--von Mises theorem. For $α\\in(0,1)$ and $d 1$ the limiting distribution is a compactly supported Pearson type-II distribution, and…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-26","reactions":{}}