A central limit theorem for convex sets (Q2457766): Difference between revisions

The author shows a couple of deep theorems, the first of which is as follows. There exist sequences \(\varepsilon_n\downarrow0\), \(\delta_n\downarrow0\) for which the following holds: Let \(n\geq1\), and let \(X\) be a random vector in the Euclidean space \({\mathbb R}^n\) with an isotropic, log-concave density \(f\). Then there exists a subset \(\Theta\) of the unit sphere \(S^{n-1}\) in \({\mathbb R}^n\) with \(\sigma_{n-1}(\Theta)\geq 1-\delta_n\) (\(\sigma_{n-1}\) denoting the unique rotationally-invariant probability measure on \(S^{n-1}\)), such that for all \(\theta\in\Theta\), the total variation distance between the distribution of \(\langle X,\theta\rangle=\sum_{i=1}^n\theta_i X_i\) and a standard normal distribution is bounded by \(\varepsilon_n\). The author also shows that \(\delta_n\leq\exp(-cn^{0.99})\) and \(\varepsilon_n\leq C([\log\log(n+2)]/\log(n+1))^{1/2}\). However in a note added in proofs, he writes that he improved some of the estimates in a subsequent paper \textit{B. Klartag} [J. Funct. Anal. 245, No.~1, 284--310 (2007; Zbl 1140.52004)]. The author notes that a function \(f\) is log-concave when the support of \(f\) is a convex set on which \(\log f\) is concave. In the second theorem, it is shown that, if in addition to the assumptions of the above theorem, the density \(f\) is unconditional, i.e.\ \(f(x_1,\dots,x_n)=f(| x_1| ,\dots,| x_n| )\) for all \((x_1,\dots,x_n)\in{\mathbb R}^n\), then one can choose \(\theta=(1/\sqrt{n},\dots,1/\sqrt{n})\) and \(\varepsilon_n\leq C/(\log(n+1))^{1/5}\). A multivariate extension of the first theorem is given as well. As a tool in the proofs, it is proved that, for an integer \(n\geq1\), a random vector \(X\) with an isotropic and log-concave density in \({\mathbb R}^n\) and all \(0\leq\varepsilon\leq 1\), the inequality \(\text{ Prob}(| n^{-1/2}| X| -1| \geq\varepsilon)\leq Cn^{-c \varepsilon^2}\) holds, where \(c,C>0\) are universal constants.