A note on Talagrand's transportation inequality and logarithmic Sobolev inequality (Q5961965): Difference between revisions

Let \((E,d)\) be a metric space -- in the sequel, a complete connected Riemannian manifold with geodesic distance -- and let \(p\in[1,\infty)\).\ Let, for \(\mu, \nu \in M^1(E)\), \(W_p(\mu,\nu)\) denote the \(L^p\)-Wasserstein distance.\ \(\mu\) satisfies the transport-entropy inequality \(W_pH(C)\), for some constant \(C>0\), if for all probabilities \(\nu\), \(W_p(\mu,\nu)\leq \left(2CH(\mu|\nu)\right)^{1/2}\), where \(H(\mu|\nu)\) denotes the Kullback-Leibler information.\ It is known that \(W_1H\) implies Gaussian concentration of measures [cf.\ \textit{K. Marton}, Ann. Probab. 24, No. 2, 857--866 (1996; Zbl 0865.60017)].\ Gaussian laws of any dimension satisfy \(W_2H(C)\) for \(C=1\) [cf. \textit{M. Talagrand}, Geom. Funct. Anal. 6, 587--600 (1996; Zbl 0859.46030)]. The authors' intention is to obtain practical criteria for \(W_2H(C)\) in terms of other prominent inequalities, in particular logarithmic Sobolev inequalities \(LSI\) and Lyapunov inequalities. Let \(d x\) denote the volume measure and, for \(V\in C^2\) with \(Z = \int e^{-V} dx <\infty\), let \(\mu\) denote the Boltzmann measure, \(\mu(d x) = e^{-V(x)} d x / Z\). Then \(LSI(C)\) holds if, for all locally Lipschitz functions \(g\), we have \(\int g^2 \log\left(g^2/(\int g^2 d \mu )\right) \,d\mu \leq 2 C \int |\nabla g|^2 \,d\mu\). It is known that \(LSI(C)\) implies \(W_2H(C)\) and that \(W_2H\) is strictly weaker than \(LSI\) [cf.\ \textit{P. Cattiaux} and \textit{A. Guillin}, J.~Math. Pures Appl. (9) 86, No. 4, 342--361 (2006; Zbl 1118.58017), and see the literature mentioned there]. Consider the \(\mu\)-symmetric operator \(\mathcal{L} = \Delta - \nabla V\cdot \nabla\) on \(C^2\). \(W\) satisfies a Lyapunov condition if \(\mathcal{L}W \leq -\phi W + b 1_{B(x_0,r)}\) for some \(\phi\geq 0\), \(r,b>0\) and \(x_0\in E\). In the main theorem (Theorem 1.2) the authors show that Lyapunov conditions of the form \(\mathcal{L}W\leq (-c d^2(x,x_0) +b)\cdot W\) (for \(x_0\in E\), \(b,c>0\), \(W\geq 1\)) imply \(W_2H(C)\) for some \(C>0\), and, moreover, that such a Lyapunov condition together with \(\text{Hess}(V) + \text{Ric}\geq K\cdot \text{Id}\) (for some \(K\leq 0\)) implies \(LSI\). Section 2 compares the criteria of Theorem 1.2 with other practical criteria; Section 3 contains the proof of Theorem~1.2, while Section 4 contains some more criteria related to \(W_2H\) and \(LSI\).