The study of random polytopes involves taking a collection of random points in space and looking at the shape formed by their convex hull, which is the smallest convex shape containing all the points. Key measurements of these shapes include quantities like volume, surface area, and more general notions called intrinsic volumes, as well as the number of faces, edges, and vertices. Understanding how much these measurements fluctuate from one random sample to another, captured by their statistical variance, is a central question in geometric probability.
This paper proves lower bounds on that variance, meaning the authors show that these geometric measurements must fluctuate by at least a certain amount, and cannot be nearly constant. The setting is a broad and natural class of probability distributions called rotationally invariant log-concave measures, which includes the standard Gaussian (bell curve) distribution but also many others. Previous results by Barany, Vu, and Thale had established such lower bounds specifically for the Gaussian case. The authors extend those results to this much wider family of distributions, showing that the fluctuation behavior is not a special quirk of the Gaussian but a more universal phenomenon.
The proof strategy builds on a geometric technique originally developed by Barany and Vu, but the authors give it a fresh treatment using barycentric coordinates, which are a way of expressing points in terms of weighted combinations of other reference points. They combine this geometric framework with careful local estimates of variance contributions near the boundary of the polytope. The result is a more flexible and general approach that handles the complexity introduced by moving beyond the Gaussian setting to distributions that may have different tail behaviors and shapes.