Saddlepoint approximations to p-values for comparison of density estimates (Q2488377)

A nonparametric two-sample test for homogeneity is considered. The test is based on comparison of kernel density estimates \(\widehat f_n(x)\) for the first and \(\widehat g_m(x)\) for the second sample. The test statistic is \(\Delta=\sup_x| \widehat f_n(x)-\widehat g_m(x)| \). Gaussian and saddlepoint approximations are developed for the distribution of \(\Delta\) for large sample sizes. Results of simulations are presented which demonstrate that the saddlepoint approximations are superior to the Gaussian ones when the bandwidth is narrow and/or sample size is small.