English

When Scores Learn Geometry: Rate Separations under the Manifold Hypothesis

Machine Learning 2026-03-17 v2 Machine Learning

Abstract

Score-based methods, such as diffusion models and Bayesian inverse problems, are often interpreted as learning the data distribution in the low-noise limit (σ0\sigma \to 0). In this work, we propose an alternative perspective: their success arises from implicitly learning the data manifold rather than the full distribution. Our claim is based on a novel analysis of scores in the small-σ\sigma regime that reveals a sharp separation of scales: information about the data manifold is Θ(σ2)\Theta(\sigma^{-2}) stronger than information about the distribution. We argue that this insight suggests a paradigm shift from the less practical goal of distributional learning to the more attainable task of geometric learning, which provably tolerates O(σ2)O(\sigma^{-2}) larger errors in score approximation. We illustrate this perspective through three consequences: i) in diffusion models, concentration on data support can be achieved with a score error of o(σ2)o(\sigma^{-2}), whereas recovering the specific data distribution requires a much stricter o(1)o(1) error; ii) more surprisingly, learning the uniform distribution on the manifold-an especially structured and useful object-is also O(σ2)O(\sigma^{-2}) easier; and iii) in Bayesian inverse problems, the maximum entropy prior is O(σ2)O(\sigma^{-2}) more robust to score errors than generic priors. Finally, we validate our theoretical findings with preliminary experiments on large-scale models, including Stable Diffusion.

Keywords

Cite

@article{arxiv.2509.24912,
  title  = {When Scores Learn Geometry: Rate Separations under the Manifold Hypothesis},
  author = {Xiang Li and Zebang Shen and Ya-Ping Hsieh and Niao He},
  journal= {arXiv preprint arXiv:2509.24912},
  year   = {2026}
}

Comments

Accepted at ICLR 2026

R2 v1 2026-07-01T06:04:48.920Z