Asymptotics of the quantization problem on metric measure spaces

Abstract

The problem of quantization of measures looks for best approximations of probability measures on a metric space by discrete measures supported on $N$ points, where the error of approximation is measured with respect to the Wasserstein distance. Zador's theorem states that, for measures on ${\mathbb{R}}^d$ or $d$-dimensional Riemannian manifolds satisfying appropriate integrability conditions, the quantization error decays to zero as $N\to \infty$ at the rate $N^{-1/d}$. In this paper, we provide a general treatment of the asymptotics of quantization on metric measure spaces $(X,\nu )$. We show that a weaker version of Zador's theorem involving the Hausdorff densities of $\nu$ holds also in this general setting. We also prove Zador's theorem in full for appropriate $m$-rectifiable measures on Euclidean space, answering a conjecture by Graf and Luschgy in the affirmative. For both results, the higher integrability conditions of Zador's theorem are replaced with a general notion of $(p,s)$-quantizability, which follows from Pierce-type (non-asymptotic) upper bounds on the quantization error, and we also prove multiple such bounds at the level of metric measure spaces.

47 pages + 12 page appendix