From the Prékopa-Leindler inequality to modified logarithmic Sobolev inequality (Q1000294): Difference between revisions

Let \(\varphi\) be a smooth strictly convex function on \(\mathbb R^n\) and let \(d\mu_\varphi(x)= \exp(-\varphi(x))\,dx\) be a probability measure on \(\mathbb R^n\). The function \(\varphi\) is called the potential of the measure \(\mu_\varphi\). For a function \(g\) define the value \[ \text{Ent}_{\mu_\varphi}(e^g):= \int e^g\log\frac{e^g}{\int e^g d\mu_\varphi} \,d\mu_\varphi. \] The main result of the paper under review is the following inequality \[ \text{Ent}_{\mu_\varphi}(e^g)\leq \int\{x\nabla g(x)-\varphi^*(\nabla\phi(x))+\varphi^*(\nabla\phi(x)-\nabla g(x))\}e^{g(x)}\,d\mu_\varphi(x) \] for all smooth enough functions \(g\) on \(\mathbb R^n\). Here \(\varphi^*\) is the Fenchel-Legendre transform of \(\varphi\). It is assumed that the potential \(\varphi\) is convex, super-linear and satisfy some other conditions. The main application of this result is to extend the modified logarithmic Sobolev inequalities obtained by \textit{I.~Gentil, A.~Guillin} and \textit{L.~Miclo} in [Probab. Theory Relat. Fields 133, No.~3, 409--436 (2005; Zbl 1080.26010) and Rev. Mat. Iberoam. 23, No.~1, 235--258 (2007; Zbl 1123.26022)]. The main result of the paper improves results due to \textit{S. G.~Bobkov} and \textit{M.~Ledoux} [Geom. Funct. Anal. 10, No.~5, 1028--1052 (2000; Zbl 0969.26019)]