Nonclassical Berry-Esseen inequalities and accuracy of the bootstrap (Q2215718)

Let \(X_1,\ldots,X_n\) be independent random vectors with zero mean, taking values in \(\mathbb{R}^p\), and let \(S_n=n^{-1/2}\sum_{i=1}^nX_i\). The main results of this paper are uniform error bounds on bootstrap approximations to \(\mathbf{P}(S_n\in B)\) which hold with high probability, where \(B\) is a Euclidean ball in \(\mathbb{R}^p\). Two different bootstrap estimators are considered. The first is Efron's bootstrap, where data are sampled uniformly at random with replacement from \(X_1,\ldots,X_n\), which in this case are assumed to be IID. The second is a weighted bootstrap estimator of the form \(n^{-1/2}\sum_{i=1}^nX_i\varepsilon_i\), where \(\varepsilon_1,\varepsilon_2,\ldots\) are IID random variables, independent of the \(X_i\), taking values in \(\mathbb{R}\), each with zero mean, unit second and third moments and bounded fourth moment. The error bounds established here have explicit dependence on \(n\) and \(p\), and show that if the underlying random variables are sub-Gaussian and \(p\) is small enough compared to \(n\), then the bootstrap approximations are valid. Important ingredients of the proofs of these bootstrap error bounds are multivariate generalized Berry-Esseen inequalities which give explicit comparisons between two sums of zero-mean random vectors in which the summands have identical moments up to some given order. The paper concludes with numerical investigation of both these generalized Berry-Esseen bounds and the bootstrap approximations.