Self-normalized Cramér-type large deviations for independent random variables. (Q1433897)

The authors consider Cramér-type tail approximations to the distribution of a sum \(S_n\) of independent random variables \(X_1,\dots, X_n\) with zero means and finite variances, normalized by \(V_n:= \{\sum^n_{i=1} X_i 2\}^{1/2}\). The relative error in the normal approximation to \(P[S_n/V_n> x]\) is shown to be small as long as \(x = o(d_{n,\delta})\) if all the \(X_i\) have finite \((2+\delta)\)th moments, for some \(0<\delta\leq 1\),where \[ d_{n,\delta}= \{EV_n 2\}^{1/2}\Biggl/\Biggl(\sum^n_{i=1} E| X_i|^{2+\delta}\Biggr)^{1/(2+\delta)}\asymp n^{1/2(2+\delta)} \] in the i.i.d. case. This result is in stark contrast to the corresponding approximation when \(\sqrt{EV_n2}\) is used as normalization, when a finite exponential moment is needed. An application to the studentized bootstrap is given, extending a corresponding result of \textit{G. P. Chistyakov} and \textit{F. Götze} [Preprint, 1999], which was proved for symmetric random variables.