Stability of extreme quantile function estimation from relatively short records having different parent distributions

Abstract

Stability of the quantile function estimation from the relatively short non-identically distributed data is investigated using large amount of fluctuating pressure measurement records acting on 3-D square section prism, which is exposed in a turbulent boundary layer. For the quantile function estimation, three kinds of frequently used methods, which are General Extreme Value distribution (GEVD), Type I distribution and General Pareto distribution (GPD) of declustered Peak Over Threshold (POT), and newly proposed translation method are considered. The investigation results clearly show that GEVD and a-priori assumed Type I (Gumbel) distribution are not appropriate for the estimation for large variance, and the results from GPD-POT method with declustering procedure show high possibility to underestimate quantiles compared with those derived from the large amount of independent block extremes. Among the considered methods, translation method shows the most stable estimation results with good accuracy.