Paperlane

no. 16 · Tuesday 17 November 2026

Support recovery of diffusion models under the manifold hypothesis

Ayya Alieva, Gautam Iyer, Alexei Novikov and Lenya Ryzhik

math.ST · 57 pp

Abstract

We study diffusion models under the manifold hypothesis with possibly large score estimation and discretization errors, and show that while they may not exactly recover the data distribution, they will exactly recover the support (or data manifold). In continuous time, we use a Lyapunov function to show that even if the relative score estimation error is fixed (and not necessarily vanishing), then the rate at which the process approaches the support is $O(\sqrt{(T-t)|\log(T-t)|})$ as $t \to T$, where $T$ is the final time. In discrete time, we use geometric time steps with a fixed (non-vanishing) ratio and show that with a fixed (also not small) relative score estimation error, the process approaches the data manifold exponentially with the number of iterations. The proof relies on controlling a Gaussian measure of a neighborhood of the cut-locus of the data manifold, using a polynomial approximation argument combined with quantitative semi-algebraic estimates.

comments
57 pages, 1 figure
fields
Statistics Theory, Probability
pdf ↗ page arXiv:2611.18738 ↗