Beyond Gaussian approximation: bootstrap for maxima of sums of independent random vectors (Q1996788)

Let \(\mathbf{X} = (X_1, \ldots, X_n)^\top \in \mathbb{R}^{n \times p}\) denote a random matrix with stochastically independent rows \(X_i = (X_{i, 1}, \ldots, X_{i, p}) \in \mathbb{R}^p\) for \(1 \leq i \leq n\), where \(p \equiv p_n\) may depend on \(n\). For \(1 \leq j \leq p\), let \(\bar{X}_{n, j} = n^{-1} \sum_{i=1}^n X_{i, j}\). The authors are concerned with the consistency of several bootstrap methods for approximating the distribution of \[ \max_{1 \leq j \leq p} \sqrt{n} \left(\bar{X}_{n, j} - \mathbb{E}\left[\bar{X}_{n, j}\right]\right) \] in the case of large \(p = p_n\), which may grow exponentially as \(n \to \infty\). In particular, the authors improve convergence rates published in this context in previous literature. Their proof methodology relies on new comparison and anticoncentration theorems. Applications in large-scale simultaneous statistical inference are also discussed. Besides these theoretical results, some computer simulations are presented in order to compare different bootstrap schemes.