Moment measures and stability for Gaussian inequalities

Alexander V. Kolesnikov ; Egor D. Kosov

Teoriâ slučajnyh processov, Tome 22 (2017) no. 2, pp. 47-61

Résumé

Let $\gamma$ be the standard Gaussian measure on $\mathbb{R}^n$ and let $\mathcal{P}_{\gamma}$ be the space of probability measures that are absolutely continuous with respect to $\gamma$. We study lower bounds for the functional $\mathcal{F}_{\gamma}(\mu) = {\rm Ent}(\mu) - \frac{1}{2} W^2_2(\mu, \nu)$, where $\mu \in \mathcal{P}_{\gamma}, \nu \in \mathcal{P}_{\gamma}$, ${\rm Ent}(\mu) = \int \log\bigl( \frac{\mu}{\gamma}\bigr) d \mu$ is the relative Gaussian entropy, and $W_2$ is the quadratic Kantorovich distance. The minimizers of $\mathcal{F}_{\gamma}$ are solutions to a dimension-free Gaussian analog of the (real) Kähler–Einstein equation. We show that $\mathcal{F}_{\gamma}(\mu) $ is bounded from below under the assumption that the Gaussian Fisher information of $\nu$ is finite and prove a priori estimates for the minimizers. Our approach relies on certain stability estimates for the Gaussian log-Sobolev and Talagrand transportation inequalities.

Détail

Keywords: Gaussian inequalities, Kähler-Einstein equation, moment measure.
Mots-clés : optimal transportation