On quadratic transportation cost inequalities (Q864183)

Define on a metric space \((E,d)\) the \(L^p\)-Wasserstein distance between two probability measures \(\mu,\nu\) by \[ W_p(\mu,\nu) := \left[\inf_\pi\int d^p(x,y)\,\pi(dx,dy)\right]^{1/p}, \] where the infimum is taken over all probability measures \(\pi\) on \(E\times E\) with marginals \(\mu\) and \(\nu\). In the sequel, \(E\) is a complete, finite-dimensional smooth Riemannian manifold, \(d\) is the geodesic distance, \(dx\) is the volume measure and \(\mu\) will be of the form \(\exp\{-V(x)\}\,dx\). It is assumed that \(V\) is \(C^3\) and that the diffusion process on \(E\) with generator \(L=\frac 12 \text{div}\nabla - \frac 12 \nabla V\cdot\nabla\) is nonexplosive. The authors introduce the following condition: \[ \sup_{1\leq p<2} \frac{\int f^2\,d\mu - \big(\int| f| ^p\,d\mu\big)^{2/p}}{(2-p)^\alpha} \leq C(\alpha)\int | \nabla f| ^2\,d\mu, \tag{I\((\alpha)\)} \] which `interpolates' between the Poincaré inequality \((I(0))\) and the log-Sobolev inequality \((I(1))\). The main results of the paper are as follows. If \((I(\alpha))\) holds for all \(\nu\) such that \(\| d\nu/d\mu\| _\infty\leq K\) the the following transportation inequality holds: \[ W_2(\nu,\mu)\leq D(\alpha, K)\sqrt{C(\alpha) H(\nu,\mu)}, \] where \(H\) is the Kullback-Leibler (relative entropy) information and \(D\) is a suitable, explicitly given constant. Moreover, it is shown that the following statements are equivalent: (a) \(I(0)\); (b) the above transportation inequality for \(\alpha=0\); (c) a certain log-Sobolev inequality. Consequently, the authors can give some examples of measures satisfying the \(T_2\)-inequality but not the log-Sobolev inequality.