A new preprint by Thomas Bonis, Thanh Mai Pham Ngoc, and Viet Chi Tran provides a statistical theory for why score-based diffusion models can generalize well when data lies on a low-dimensional manifold. The authors analyze empirical risk minimization over Sobolev classes with derivative constraints and derive convergence rates for the estimated score function.

The key result is that these score estimators yield minimax optimal Wasserstein-1 rates for the generated distribution, up to logarithmic factors. The rates depend on the intrinsic dimension of the data manifold rather than the ambient dimension, offering a theoretical explanation for why such models avoid overfitting in practice.

The paper gives sufficient conditions for score-based generative modeling to generalize from finite samples, bridging the gap between empirical success and statistical guarantees. As a single source, the claims reflect the authors' own analysis and have not yet been independently verified.