Displacement convexity of entropy and related inequalities on graphs (Q466893): Difference between revisions

In recent years, there has been a great interest in developing a notion of Ricci curvature in the setting of discrete spaces, i.e., finite graphs or finite Markov chains. A fruitful direction has been one where Ricci curvature is defined via convexity properties of the relative entropy regarded as a function on the space of probability measures equipped with a \(L_p\)-Wasserstein metric or a related metric. The study of convexity properties of relative entropy is important because of its connection to several geometric and function inequalities, and several approaches to extend this theory to discrete settings have been proposed. One which uses a pseudo Wasserstein metric was developed in [\textit{J. Maas}, J. Funct. Anal. 261, No. 8, 2250--2292 (2011; Zbl 1237.60058)] and [\textit{M. Erbar} and \textit{J. Maas}, Arch. Ration. Mech. Anal. 206, No. 3, 997--1038 (2012; Zbl 1256.53028)]. Consider the discrete hypercube \(X = \{0,1\}^n\), and let \(\nu\) be the uniform probability measure on \(X\). \textit{Y. Ollivier} and \textit{C. Villani} [SIAM J. Discrete Math. 26, No. 3, 983--996 (2012; Zbl 1267.52010)] proved that, for any probability measures \(\mu_0\) and \(\mu_1\) in \(X\), one can construct a probability measure \(\mu_{1/2}\) on the set of midpoints of the supports of \(\mu_0\) and \(\mu_1\) such that \[ H(\mu_{1/2}|\nu) \leq \frac{1}{2}\, H(\mu_0|\nu) + \frac{1}{2}\, H(\mu_1|\nu) - \frac{1}{16n}\, W_1(\mu_0,\mu_1)^2, \] where \(H(\cdot|\nu)\) denotes relative entropy, and \(W_1\) denotes the \(L_1\)-Wasserstein distance. In this article, the authors propose another interpolation between probability measures and derive a similar inequality which holds in a more general setting. More precisely, let \(G\) be a finite graph with vertex set \(V\). For any two vertices \(x\) and \(y\), the authors define a path \(\mu_t^{x,y}\) between the corresponding delta measures \(\delta_x\) and \(\delta_y\) based on a binomial interpolation. The definition of \(\mu_t^{x,y}\) includes the contribution of all the geodesics from \(x\) to \(y\). From this, they construct a path between any two probability measures on \(V\). That is, let \(\mu_0\) and \(\mu_1\) be any two probability measures on \(V\). For any coupling \(\pi\) of \(\mu_0\) and \(\mu_1\), the authors define a path \(\mu_t^{\pi}\) from \(\mu_0\) to \(\mu_1\) by \[ \mu_t^{\pi} (\cdot) = \sum_{(x,y)\in V^2} \pi (x,y) \mu_t^{x,y}(\cdot) ,\quad \text{for all } t \in [0,1]. \] Using this new notion the authors derive a displacement convexity of entropy inequality with respect to the \(L_1\)-Wasserstein-metric on the complete graph and products of complete graphs, and from this result they derive several important inequalities, i.e., a discrete version of the Prékopa-Leindler type inequalities, and HWI type inequalities on graphs. In particular, for the discrete hypercube, they prove the following result: Theorem. Let \(\mu\) be a probability on \(\{0,1\}\), and denote by \(\mu^{\otimes n}\) its \(n\)-fold product on \(X\). Then, for any two probability measures \(\mu_0\) and \(\mu_1\) on \(X\), there exists a coupling \(\pi\) of \(\mu_0\) and \(\mu_1\) such that, for all \(t\in [0,1]\), \[ H(\mu_t^{\pi}|\mu^{\otimes n}) \leq (1-t) H (\mu_0|\mu^{\otimes n}) + t H(\mu_1| \mu^{\otimes n}) - \frac{2t(1-t)}{n} W_1(\mu_0,\mu_1)^2. \] In the introduction, the authors provide an overview of known results with extensive references for both continuous and discrete settings. A related work which uses another type of binomial interpolation can be found in [\textit{E. Hillion}, Potential Anal. 41, No. 3, 679--698 (2014; Zbl 1305.60014)].