{"id":"arxiv/2601.06715","url":"https://arxiv.org/abs/2601.06715","tags":["math.ST","cs.LG","math.PR","stat.ML"],"score":0,"title":"Diffusion Models with Heavy-Tailed Targets: Score Estimation and Sampling Guarantees","domain":"math.ST","source":"arxiv","authors":["Yifeng Yu","Lu Yu"],"reviews":0,"subject":"math.ST","summary":"Score-based diffusion models have become a powerful framework for generative modeling, with score estimation as a central statistical bottleneck. Existing guarantees for score estimation largely focus on light-tailed targets or rely on restrictive assumptions such as compact support, which are often violated by heavy-tailed data in practice. In this work, we study conventional (Gaussian) score-based diffusion models when the target distribution is heavy-tailed and belongs to a Sobolev class with smoothness parameter $β>0$. We consider both exponential and polynomial tail decay, indexed by a…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-01-10","reactions":{}}