arXiv · 2611.18738
Support recovery of diffusion models under the manifold hypothesis
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.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Ayya Alieva, Gautam Iyer, Alexei Novikov, Lenya Ryzhik. 2026-11-11. Support recovery of diffusion models under the manifold hypothesis. https://arxiv.science/abs/2611.18738
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.