Generalization Properties of Score-matching Diffusion Models for Intrinsically Low-dimensional Data

Saptarshi Chakraborty, Quentin Berthet, Peter L. Bartlett · 2026

Read Generalization Properties of Score-matching Diffusion Models for Intrinsically Low-dimensional Data by Saptarshi Chakraborty, Quentin Berthet, Peter L. Bartlett (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-2603-03700.txt.

Description

Saptarshi Chakraborty, Quentin Berthet, and Peter Bartlett analyze how well score-based diffusion models generalize when data has low intrinsic dimension. Diffusion models succeed in practice, but theory has lagged, often giving pessimistic rates that ignore the low-dimensional structure of real data such as natural images. The authors study learning an unknown distribution from finitely many samples under mild regularity conditions on the forward process and the data.

They prove finite-sample error bounds in Wasserstein distance that hold for a wide range of cases and need only a finite-moment condition, with no compact support, manifold, or smooth density assumptions. The error scales with a Wasserstein dimension of the data rather than the ambient dimension, showing diffusion models adapt to intrinsic geometry. The results connect diffusion theory with GANs and optimal transport, and the new dimension also covers heavy-tailed distributions.

Catalog record

CREATORSaptarshi Chakraborty, Quentin Berthet, Peter L. Bartlett
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE149867
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03700

Browse the full library · About the Tower of Babel