Diffusion Models Are Statistically Optimal for Learning Low-Dimensional Multi-Modal Distributions

Jingda Wu, Changxiao Cai · 2026

Read Diffusion Models Are Statistically Optimal for Learning Low-Dimensional Multi-Modal Distributions by Jingda Wu, Changxiao Cai (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-30153.txt.

Description

Jingda Wu and Changxiao Cai show score-based diffusion models are near-optimal for learning low-dimensional multi-modal data. Current theory leans on strong regularity assumptions, such as bounded densities or smooth scores, which miss such structure. For distributions on a union of low-dimensional subspaces with subgaussian pieces, diffusion models need sample counts growing only with intrinsic dimension to reach a given error in Wasserstein distance. The rate avoids the curse of dimensionality and needs no smoothness or log-concavity assumptions.

Catalog record

CREATORJingda Wu, Changxiao Cai
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88282
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.30153

Browse the full library · About the Tower of Babel