Simultaneous concentration of order statistics

Abstract

Let $\mu$ be a probability measure on $\mathbb{R}$ with cumulative distribution function $F$, $(x_{i})_{1}^{n}$ a large i.i.d. sample from $\mu$, and $F_{n}$ the associated empirical distribution function. The Glivenko-Cantelli theorem states that with probability 1, $F_{n}$ converges uniformly to $F$. In so doing it describes the macroscopic structure of $\{x_{i}\}_{1}^{n}$, however it is insensitive to the position of individual points. Indeed any subset of $o(n)$ points can be perturbed at will without disturbing the convergence. We provide several refinements of the Glivenko-Cantelli theorem which are sensitive not only to the global structure of the sample but also to individual points. Our main result provides conditions that guarantee simultaneous concentration of all order statistics. The example of main interest is the normal distribution.