Poincaré's inequalities and Talagrand's concentration phenomenon for the exponential distribution (Q679155): Difference between revisions

The authors prove a logarithmic Sobolev inequality for the exponential measure \(\nu_1:\text{Ent}_{\nu_1}(e^f)\leq C\int f^{\prime 2}e^fd\nu_1\), where \(f\) is Lipschitz function on \(\mathbb{R}\) such that \(|f'|\leq c<1\), \(C= 2/(1- c)\), \(\nu_1(dx)= (1/2)e^{-|x|}dx\), \(\text{Ent}_\mu(g)= \int g\log gd\mu-(\int gd\mu)(\log\int gd\mu)\), \(g\geq 0\), \(\mu\) being a probability measure. As a consequence a Talagrand's concentration phenomenon in \(\mathbb{R}^n\) is derived: \(\nu_n(A+\sqrt rB_2+ rB_1)\geq 1-e^{-r/K}\), where \(K\) is a positive constant, \(r\geq 0\), \(B_k=\{(x_1,\dots,x_n); \sum^n_{i=0}|x_i|^k\leq 1\}\), \(\nu_n=\nu_1\otimes\cdots \otimes\nu_1\) and \(A\) is a Borel set such that \(\nu_n(A)\geq 1/2\). Suppose more generally that \(\mu\) is a probability measure on a metric space \((E,d)\). It is said that \(\mu\) satisfies a Poincaré inequality with constant \(\lambda\), if for any \(f\) such that \(\int f^2d\mu<\infty\), then \(\lambda_1\text{Var}_\mu(f)\leq \int|\nabla f|^2d\mu<\infty\), where \(\text{Var}_\mu(f)\) denotes the variance of \(f\) with respect to \(\mu\) and \(|\nabla f|(x)=\limsup_{y\to x} |f(x)- f(y)|/d(x,y)\). As a main result, the authors prove that any measure satisfying such a Poincaré inequality will satisfy an inequality on entropy (i.e. \(\text{Ent}_\mu(e^f)\leq C\int|\nabla f|^2 e^fd\mu)\), and therefore a concentration result similar to the one for the exponential distribution. The authors also discuss exponential integrability under Poincaré inequalities and its consequence to sharp diameter upper bounds on spectral gaps.