Probability Read this week

Sharp and explicit Berry--Esseen inequalities over convex sets

Zhilova

Finite-sample normal approximation bounds over all convex sets in \(\mathbb{R}^d\) for sums of independent random vectors, with numerical constants, optimal dependence on the sample size, dimension and moments, and matching lower bounds.

Filed under Probability

The paper bounds the distance between the law of a normalised sum of \(n\) independent random vectors and the Gaussian, uniformly over all convex subsets of \(\mathbb{R}^d\). The upper bounds carry numerical constants, allow arbitrary dependence among coordinates and cover every moment order from three upwards; for identically distributed summands the dependence on \(n\), \(d\) and the moment bound is shown to be optimal, with matching lower bounds from bounded distributions.

The headline finding is a fourth-moment threshold. Four is the first moment order for which uniform bounds on coordinate moments give a worst-case error linear in the dimension, and bounds on higher moments give the same rate. The proof adds independent Gaussian noise to exploit convexity, proves a normal approximation for Rademacher sums on Parseval frames, and then uses random swaps with independent copies to apply that bound conditionally to general sums.