Curvature and transport inequalities for Markov chains in discrete spaces (Q2405199): Difference between revisions

The paper explores various characteristics of an irreducible Markov chain (MC) with a unique reversible invariant measure \(\pi\), MC being understood as a Markov process in discrete time (for Theorems 1.10 and 1.12) or continuous time (for Theorem 1.13) on a finite or countable state space, in particular on a graph. These characteristics, most of which based on the transition kernel of the MC are: \(\Gamma\) operator, iterated \(\Gamma\) operator (notation \(\Gamma_2\)), two modifications of iterated \(\Gamma_2\), \(L_p\)-Wasserstein distance between two probability measures \(W_p(\mu, \pi)\) (also known as Kantorovich-Rubinstein distance, named after Leonid Vaserstein, see \url{http://personal.psu.edu/lxv1/}), the coarse Ricci curvature (for continuous time) due to \textit{Y. Ollivier} [J. Funct. Anal. 256, No. 3, 810--864 (2009; Zbl 1181.53015)] meaning exponential convergence in the metric \(W_1\) of the corresponding semigroups related to the same generator of the MC and starting with two different measures, Fisher's information of a measure with respect to \(\pi\), relative entropy of a measure with respect to \(\pi\). Two versions of the Ricci curvature for discrete time are the curvature dimension condition due to \textit{D. Bakry} and \textit{M. Émery} [Lect. Notes Math. 1123, 177--206 (1985; Zbl 0561.60080)] and the exponential curvature condition due to \textit{F. Bauer} et al. [J. Differ. Geom. 99, No. 3, 359--405 (2015; Zbl 1323.35189)] defined in terms of \(\Gamma\) and two versions of the iterated \(\Gamma\) operators. It is then defined what it means for the measure \(\pi\) to the satisfy the log-Sobolev inequality, the modified log-Sobolev, the transport-entropy inequality, and the transport-information inequality, and further, what it means for \(\pi\) to satisfy several versions of the weak transport-information inequality. The main results are three theorems which state several transport-information inequalities under one of the curvature conditions: they all mean that one of Wasserstein distances between \(\pi\) and its equivalent, say, \(\nu\) with a density \(f\) (i.e., \(d\nu / d\pi = f\)) does not exceed Fisher's information at this density \(I_\pi(f)\) with some constant multiplier, which constant, in turn, relates to the constants in the corresponding curvature condition. Two more results -- Theorems 2.4 and 2.5 -- state transport-entropy inequalities for the MC satisfying one of transport-information ones. There are applications to concentration and diameter bounds and a long list of references with a substantial review of various links between different conditions. One of the background papers is [\textit{A. Guillin} et al., Probab. Theory Relat. Fields 144, No. 3--4, 669--695 (2009; Zbl 1169.60304)].