Some inequalities on Riemannian manifolds linking entropy, Fisher information, Stein discrepancy and Wasserstein distance (Q6156061)

Let \(M\) be a complete connected Riemannian manifold, equipped with a probability measure \(\mu\) defined by \(\mu(\mathrm{d}x)=e^{-V(x)}\mathrm{vol}(\mathrm{d}x)\) for a suitable function \(V\) and where \(\mathrm{vol}\) is the Riemannian volume measure. For another probability measure \(\nu\) on \(M\), the authors establish inequalities linking the relative entropy \(H(\nu|\mu)\), Fisher information \(I(\nu|\mu)\), Stein discrepancy \(S(\nu|\mu)\) and Wasserstein distance \(\mathbb{W}_2(\nu,\mu)\). For example, in the setting where \(V\) satisfies \(\mathrm{Hess}_V=K>0\) is constant, the authors show that \begin{align*} H(\nu|\mu)&\leq\frac{1}{2}S^2(\nu|\mu)\log\left(1+\frac{I(\nu|\mu)}{KS^2(\nu|\mu)}\right)\,,\\ \mathbb{W}_2(\nu,\mu)&\leq\frac{S(\nu|\mu)}{K^{1/2}}\arccos\left(\exp\left\{-\frac{H(\nu|\mu)}{S^2(\nu|\mu)}\right\}\right). \end{align*} Inequalities established in the present paper strengthen the well-known log-Sobolev and transportation cost inequalities, and generalise the HSI inequality for a Gaussian measure on a Euclidean space to Riemannian manifolds.