{"id":"arxiv/2601.18689","url":"https://arxiv.org/abs/2601.18689","tags":["math.ST"],"score":0,"title":"Function estimation in the empirical Bayes setting","domain":"math.ST","source":"arxiv","authors":["Benjamin Kang","Yury Polyanskiy","Anzo Teh"],"reviews":0,"subject":"math.ST","summary":"We study function estimation in the empirical Bayes setting for Poisson and normal means. Specifically, given observations $X_i\\sim f(\\cdot; θ_i)$ with latent parameters $θ_i\\sim π$, the goal is to estimate $\\mathbb{E}_π[\\ell(θ)|X = x]$. This task lies between classical deconvolution (recovering the full prior $π$), and standard empirical Bayes mean estimation. While the minimax risk for estimating $π$ in the Wasserstein distance is known to decay only logarithmically, we show that estimating the corresponding posterior smooth functionals admits dramatically faster rates. In particular, for…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-01-26","reactions":{}}