Paperlane

no. 6 · Monday 2 November 2026

Simple and Sharp Generalization Bounds via Lifting

Jingbo Liu

math.ST · v5

Abstract

We bound the supremum of stochastic processes using a rate-distortion functional analogous to classical entropy integrals. A lifting argument embeds the process into a higher-dimensional space with permutation symmetry, reducing the analysis to a stationary process. This gives a two-sided characterization of the supremum of Gaussian processes, including cases where Dudley's entropy integral is not sharp. Combining the sharpness of Dudley's integral for stationary processes with information-theoretic arguments gives a proof of the majorizing measure theorem. The method also yields localized generalization bounds without a slicing argument. As an application, we derive risk bounds for empirical risk minimization over Lipschitz images of Sobolev ellipsoids. Our proof does not use the majorizing measure theorem.

comments
1 figure; Annals of Applied Probability, to appear
fields
Statistics Theory
pdf ↗ page arXiv:2508.18682 ↗