When does Edgeworth beat Berry and Esséen? Numerical evaluations of Edgeworth expansions (Q1372343)

Edgeworth expansions are now common practice in some areas of statistics such as the bootstrap, risk theory in actuarial sciences, econometrics, quantile, regression curve and density estimation, etc. Since Edgeworth expansions are expected to improve on the traditional \(O(n^{-1/2})\) Berry-Esséen bounds, the authors of the present paper try to answer to the following natural question: how large should \(n\) be for Edgeworth beating Berry-Esséen. The complexity of the task is quite tough since, in the special case of the Wilcoxon signed rank statistic, the Edgeworth expansions do not improve upon Berry-Esséen bounds unil the sample size reaches the astronomically huge value of \(e^{5000}\) [\textit{W. Albers, P.J. Bickel} and \textit{W.R. van Zwet}, Ann. of Statist. 4, 108-156 (1976; Zbl 0321.62049)]. The aim of the paper is to revise this result, considering instead of the universal constant (on the Berry-Esséen model) a sequence of constants depending on the sample size, and allowing for non-monotonic bounds on the remainder term. The obtained numerical bounds turn out to be much sharper: the minimal sample size at which Edgeworth beats Berry-Esséen (with 0.7915 constant) drops down considerably from the huge number \(e^{5000}\) to the moderate (but still very large) number 128098. Unfortunately, a sample of size 128098 can hardly be considered as a small nor medium size sample. The conclusion is that the users of Edgeworth expansions in statistical applications should remain extremely cautious, and keep in mind the warning that higher-order expansions do not necessarily imply better approximations, even for fairly large sample sizes.