Ollivier curvature, isoperimetry, concentration, and log-Sobolev inequality

Abstract

We introduce a Laplacian separation principle for the eikonal equation on Markov chains. As application, we prove an isoperimetric concentration inequality for reversible Markov chains with non-negative Ollivier curvature. That is, every single point from the concentration profile yields an estimate for every point of the isoperimetric profile. Applying this to exponential and Gaussian concentration, we obtain affirmative answers to two open questions by Erbar and Fathi. Moreover, we prove that the modified log-Sobolev constant is at least the minimal Ollivier Ricci curvature, assuming non-negative Ollivier sectional curvature, i.e., the Ollivier Ricci curvature when replacing the ${\ell }_1$ by the ${\ell }_{\infty }$ Wasserstein distance. This settles a recent open problem by Pedrotti. We give a simple example showing that non-negative Ollivier sectional curvature is necessary to obtain a modified log-Sobolev inequality via a positive Ollivier Ricci bound. This provides a counterexample to a conjecture by Peres and Tetali.