New proof: diffusion models beat curse of dimensionality, converge faster

Denoising Diffusion Probabilistic Models (DDPMs) have become the state-of-the-art for generating synthetic data, powering image, audio, and video generation. Yet their remarkable success in high-dimensional spaces has outpaced theoretical understanding. The manifold hypothesis suggests that real-world data lies on lower-dimensional manifolds embedded in a high-dimensional ambient space, and this latest paper provides the strongest evidence yet that diffusion models exploit this structure. Iskander Azangulov, George Deligiannidis, and Judith Rousseau (University of Oxford) have proven that DDPMs achieve score-learning rates that are independent of the ambient dimension, meaning the curse of dimensionality does not apply when data lies on a manifold.

The authors achieved this by developing a novel framework that connects diffusion models to the theory of extrema of Gaussian Processes, a well-studied area in probability and statistics. This connection allowed them to derive sampling complexity bounds in Wasserstein distance that also do not depend on the ambient dimension—only on the intrinsic dimension of the data manifold. Prior work had shown some adaptation to manifold structure, but not to this extent, leaving a gap between theory and the empirical performance of diffusion models. The new results close that gap significantly, offering rigorous justification for why DDPMs work so well in practice on complex, high-dimensional datasets.

This proof has immediate implications for researchers and engineers working on generative models. It provides a theoretical foundation for scaling diffusion models to even higher dimensions without fear of exponential sample complexity. It also suggests that future algorithmic improvements should focus on manifold learning and intrinsic dimensionality rather than the raw input space. For the broader AI community, this work is a step toward making generative AI more predictable and efficient, with potential impact on scientific applications such as molecular generation and medical imaging, where high-dimensional data is the norm.