Diffusion Models for High-Dimensional Clustered Data: Intrinsic-Dimension Adaptivity via Bayesian Classification

What the math gives to ML

The paper identifies a concrete phase structure in diffusion denoising for clustered, heterogeneous low-dimensional data: an early mixing phase followed by a cluster-commitment phase. Its transferable asset is the exact Bayesian decomposition of a mixture score into posterior-weighted cluster scores, together with the result that posterior weights concentrate once the signal-to-noise ratio is sufficiently large. This suggests a routed diffusion architecture in which a lightweight noisy-sample classifier selects a cluster-specific low-rank score expert, replacing an ambient-dimensional score computation after commitment. The most direct validation is on synthetic Gaussian mixtures with different intrinsic ranks, measuring whether routed low-rank denoising preserves sample quality while reducing score-network FLOPs.