Estimates for the rate of strong approximation in Hilbert space (Q642076)

The paper deals with an accuracy of a strong Gaussian approximation of sums of i.i.d. Hilbert space-valued random vectors. Let \(Z= (Z_1, Z_2,\dots)\) be a mean zero random vector with values in the space \(\ell_2\) such that its coordinates are uncorrelated and \(\sigma^2_k= EZ^2_k\) form a nonincreasing sequence. One of the main results states that if \(E\| Z\|^\gamma<\infty\) for some \(\gamma> 2\) then for every fixed \(n\) and \(d\geq 1\) satisfying some condition (roughly \(d\) should not be too big with respect to \(n\)) there exist independent copies \(X_1,\dots, X_n\) of \(Z\) and independent copies \(Y_1,\dots, Y_n\) of a mean zero Gaussian vector with the same covariance operator \(\text{Cov\,}Z\) and defined on the same probability space such that \[ E\Biggl(\max_{1\leq k\leq n}\,\| S_k- T_k\|\Biggr)^\gamma\leq C(\gamma)(\sigma_1/\sigma_d)^\gamma nd^{11\gamma}E\| Z\|^\gamma+ \Biggl(n\sum_{k> d} \sigma^2_k\Biggr)^{\gamma/2}, \] where \(S_k=\sum_{i\leq k}X_i\), \(T_k= \sum_{i\leq k}Y_i\) and \(C(\gamma)\) depends only on \(\gamma\). The theorems are obtained as corollaries of recent finite-dimensional results of the authors.