A multivariate Gnedenko law of large numbers (Q378786)

The Gnedenko law of large numbers describes the concentration behavior of the maximum and minimum of i.i.d. real random variables \(x_1,\dots,x_n\), as \(n\to\infty\), under assumptions on the regularity and decay of the distribution function of the \(x_i\). Here, the author investigates multivariate extensions of this classical result. Starting with a log-concave distribution \(\mu\) with density function \(f\) in \({\mathbb R}^d\), i.i.d. random vectors \(x_1,\dots, x_n\) with distribution \(\mu\) are used to build the random polytope \(P_n=\) conv\(\{x_1,\dots ,x_n\}\). It is shown that, with high probability, \(P_n\) approximates the floating body \(H_{1/n}\) of \(\mu\). In fact, explicit bounds for \(n\) and the probability are given which allow the (essentially optimal) upper bound \(1+c \log\log n/\log n\) for the logarithmic Hausdorff distance between \(P_n\) and \(H_{1/n}\). Under a stronger condition on \(\mu\) (\(p\)-log-concavity, with \(p>1\)), the bound \(c \log\log n/(\log n)^{1-1/p}\) is established for the (ordinary) Hausdorff metric \(d_{\mathcal H}(P_n,F_{1/n})\). The results are used to show the existence of a probability measure \(\mu\) such that, with probability 1, the corresponding sequence \(P_{d+1},P_{d+2},\dots\) of random polytopes lies dense with respect to the Banach-Mazur distance in the class \({\mathcal K}_d\) of all convex bodies in \({\mathbb R}^d\). The floating body \(F_\delta\) of \(\mu\) is the intersection of all closed half-spaces \(H\subset{\mathbb R}^d\) such that \(\mu(H)\geq 1-\delta\). If \(\mu\) is the uniform distribution on a convex body \(K\), this notion coincides with the floating body of \(K\) which has been widely used in convex geometry.