Probability Papers

High-Dimensional Gaussian and Bootstrap Approximation: Near 1/√n Rates with General Covariance

Lopes and Masini

Berry–Esseen bounds over rectangles of order \(1/\sqrt{n}\), up to factors of \(\log(pn)\), hold for sums of i.i.d. vectors in \(\mathbb{R}^p\) with sub-Gaussian entries and an unrestricted covariance matrix.

Filed under Probability, Learning theory

For sums of \(n\) i.i.d. random vectors in \(\mathbb{R}^p\), Gaussian approximation over hyperrectangles underpins simultaneous inference when \(p\) far exceeds \(n\). Whether the rate \(1/\sqrt{n}\) is attainable up to factors of \(\log(pn)\) without restrictions on the population covariance had been a long-standing open question.

The authors answer it affirmatively for vectors with sub-Gaussian entries of non-zero variance, and prove matching results for bootstrap approximation and Gaussian comparison using a discrepancy measure between covariance matrices that improves on earlier work. The proofs use interpolation and bound the interpolation derivatives in a way that does not depend on the covariance structure.