Central limit theorems and bootstrap in high dimensions (Q2406563)

Let \(X_1,\ldots,X_n\) be independent, centred random vectors in \(\mathbb{R}^p\), and let \(S_n^X=\frac{1}{\sqrt{n}}\sum_{i=1}^nX_i\). The authors consider Gaussian approximation of \(S_n^X\). That is, letting \(Y_1,\ldots,Y_n\) be independent, centred Gaussian random variables such that \(Y_i\) has the same covariance structure as \(X_i\) for each \(i\), and defining \(S_n^Y=\frac{1}{\sqrt{n}}\sum_{i=1}^nY_i\), the authors provide bounds on \[ \rho_n(\mathcal{A})=\sup_{A\in\mathcal{A}}\left|P(S_n^X\in A)-P(S_n^Y\in A)\right|\,, \] for classes \(\mathcal{A}\) of Borel sets in \(\mathbb{R}^p\). Motivated by modern high-dimensional statistical applications (for example, for statistics of the Kolmogorov-Smirnov or Pearson type), the authors' primary interest is those cases where \(\rho_n(\mathcal{A})\rightarrow0\) even when the dimension \(p=p_n\) is allowed to grow much faster than \(n\). Through explicit bounds (which do not place any restriction on the correlation structure in the underlying random variables), this is shown to be the case if \(\mathcal{A}\) is the class of hyperrectangles in \(\mathbb{R}^p\) or, more generally, the class of simple or sparsely convex sets. In the case of hyperrectangles, the authors show that we have \(\rho_n\rightarrow0\) even when \(p=O(\exp\{an^b\})\), for some constants \(a\) and \(b\). Finally, the authors also derive high-dimensional empirical and multiplier bootstrap theorems, which can be used to provide explicit approximations for the probabilities \(P(S_n^X\in A)\) by data-dependent techniques in cases where the covariance structure of \(S_n^Y\) is unknown and direct computation of \(P(S_n^Y\in A)\) is not feasible.