From dimension free concentration to the Poincaré inequality (Q2017824)

The main result of this paper is to show the equivalence between probability measures satisfying a certain concentration property and those satisfying a Poincaré inequality. More precisely, if \((\mathcal{X},d)\) is a Polish metric space and \(p\geq1\), define the distance \(d_p\) by \[ d_p(x,y)=\left[\sum_{i=1}^nd^p(x_i,y_i)\right]^{1/p} \] on the product space \(\mathcal{X}^n\), and for a Borel subset \(A\) of \(\mathcal{X}^n\) let \[ A_{r,p}=\left\{x\in\mathcal{X}^n:d_p(x,A)\leq r\right\}\,. \] A Borel probability measure \(\mu\) on \(\mathcal{X}\) is said to satisfy the dimension free concentration property (with a non-increasing concentration profile \(\alpha:\mathbb{R}^+\mapsto\mathbb{R}^+\)) if \[ \mu^n(A_{r,p})\geq1-\alpha(r)\,, \] for all \(r\geq0\) and all \(A\subseteq\mathcal{X}^n\) with \(\mu^n(A)\geq1/2\). The main result of the present paper is that if \(\mu\) satisfies the dimension free concentration property with \(p=2\), then \(\mu\) satisfies the Poincaré inequality \[ \lambda\text{Var}_\mu(f)\leq\int|\nabla^-f|^2\,d\mu\,, \] for all Lipschitz functions \(f:\mathcal{X}\mapsto\mathbb{R}\) and with an explicitly defined constant \(\lambda\), where \[ |\nabla^-f|(x)=\limsup_{y\rightarrow x}\frac{\max\{f(x)-f(y),0\}}{d(x,y)}\; \] when \(x\) is not isolated in \(\mathcal{X}\), and \(|\nabla^-f|(x)=0\) if \(x\) is isolated in \(\mathcal{X}\). Moreover, the converse also holds when \(\alpha\) has the form \(\alpha(r)=be^{-ar}\) for some \(a,b>0\). The proof of the main result uses a formulation of the concentration property in terms of deviation inequalities for inf-convolution operators, and employs a central limit theorem for triangular arrays of independent random variables. The present paper also discusses this main result in the context of other notions of concentration appearing in the literature, and gives further results on the links between restrictions of the concentration property and Poincaré inequality to convex sets in the case where \((\mathcal{X},d)\) is geodesic.