On the relative approximation error of the generalized Pareto approximation for a high quantile (Q2488436)

Let \(F\) be a PDF in the domain of max-attraction of the extreme value distribution with tail index \(\gamma\geq 0\). Then the high quantile \(q_\alpha\): \(\overline F(q_\alpha)=1-F(q_\alpha)=\alpha\) (\(\alpha\approx 0\)) can be approximated by using the generalized Pareto approximation of the distribution of excess over a high threshold \(u\). This approximation reads: \[ \begin{aligned} \tilde q_\alpha & =u-{\sigma(u)\over\gamma} (1-(\alpha/ \overline F(u))^{-\gamma})\quad \text{if } \gamma>0, \\ \tilde q_\alpha & =u-\sigma(u)\ln (\alpha/\overline F(u))\quad \text{if } \gamma=0, \end{aligned} \] where \(\sigma(u)=V'(V^{-1}(u))\), \(V(t)=\overline F^{-1}(e^{-t})\). The authors analyse the relative error \(\varepsilon_\alpha=(q_\alpha-\tilde q_\alpha)/q_\alpha\) of such approximation when \(\alpha\to 0\). The general asymptotic expansion is applied to Weibull, normal, gamma and lognormal quantiles.