Sharp High-Dimensional Gaussian Approximations over Rectangles without Covariance Restrictions
Victor Chernozhukov, Denis Chetverikov, Kengo Kato, Yuta Koike
econ.EM
Nov 16, 2026 · v1
math.ST
TL;DR
The paper's 31 proved statements and their substantive proof arguments were formalized and verified in Lean 4, with the formalization performed using an AI system.
Abstract
We study how well a sum of $n$ independent random vectors with $p$ coordinates can be approximated by a Gaussian vector with the same mean and covariance. We measure the error over rectangles, so the bounds apply to maxima and simultaneous confidence regions. The coordinates may be arbitrarily dependent, even exactly linearly dependent, and the bounds do not deteriorate as the covariance matrix becomes singular. After standardizing the coordinate variances, bounded or sub-exponential coordinates give the $n^{-1/2}$ sample-size rate up to logarithmic factors, with the size or tail bounds held fixed. The main upper bounds retain the Gaussian expected maximum, improving the rates when the coordinates are strongly dependent. Fourth-moment refinements give further improvements when large observations contribute little to individual coordinate variances. If we instead impose a fixed moment bound of order $q>4$ on the largest absolute coordinate, together with a coordinate fourth-moment bound, the rate becomes $n^{-1/2+1/q}$ up to logarithmic factors. At $q=4$, the slower $n^{-1/4}$ rate is unavoidable, confirming our earlier conjecture. We also allow the bounds on observation size to grow with $n$. Lower-bound examples show that the dependence on sample size, observation size, and Gaussian width is sharp up to logarithmic factors in the regimes studied, and in several regimes the upper and lower bounds match including the logarithmic terms.
Problem
The paper studies how well a sum of n independent random p-dimensional vectors can be approximated by a Gaussian vector with the same mean and covariance. Error is measured in Kolmogorov distance over rectangles. The coordinates may be arbitrarily dependent, and the covariance matrix may be singular.
Approach
The upper bounds come from smoothing rectangle indicators and interpolating between the sum and its Gaussian counterpart, after changing to support coordinates so that singular covariances can be handled. The bounds retain the Gaussian expected maximum (Gaussian width). Unbounded observations are handled by conditioning on large observations, or alternatively through fourth-moment and Bennett-inequality arguments. Lower-bound constructions show sharpness, and the archived version's statements and proofs were verified in Lean 4.
Results
Bounded or sub-exponential coordinates give the n^{-1/2} rate up to logarithmic factors, and q-th moment bounds with q>4 give n^{-1/2+1/q}. At q=4 the n^{-1/4} rate is unavoidable, and the lower bounds match the upper bounds up to logarithmic factors, sometimes exactly. Lean 4 verification covers all 31 proved statements, with the derivative-envelope estimates restricted to integer orders q≥1.