Probability Papers

Sharp and explicit Berry--Esseen inequalities over convex sets

Mayya Zhilova

Finite-sample Berry-Esseen bounds over all convex subsets of \(\mathbb{R}^d\) for sums of independent random vectors, with explicit constants and optimal dependence on sample size, dimension and moments.

Filed under Probability

The question is how fast the law of a normalised sum of \(n\) independent random vectors in \(\mathbb{R}^d\) approaches a Gaussian, measured uniformly over all convex sets, and how the error depends on \(n\), on \(d\) and on the moment assumptions. The paper gives upper bounds with numerical constants that allow arbitrary dependence between coordinates and cover every moment order from three upwards.

For identically distributed summands it determines the optimal worst-case error and proves matching lower bounds with bounded distributions. The answer has a threshold: four is the first moment order at which uniform bounds on coordinate moments give an error linear in \(d\), and higher moments do not improve the rate. The proof adds independent Gaussian noise to obtain a bound for Rademacher sums on Parseval frames, then uses random swaps with independent copies to reach general sums. Explicit high-dimensional normal approximation of this kind underlies bootstrap and simultaneous-inference guarantees.