A0372
Title: Bandwidth selection method for estimating difference between two densities with kernel density estimation
Authors: Sixiao Zhu - Paris 1 University (France) [presenting]
Alain Celisse - Paris 1 University (France)
Abstract: Measuring the difference between two probability distributions is the key issue of many statistical applications such as change point detection and clustering. Calculating the L2 distance between two kernel estimations of densities serves a straight forward measure of this kind. This brings forward the classical issue of choosing the bandwidth of kernel used in this procedure. In this work, we address this issue by borrowing idea from the recent work of penalized comparison to overfitting (PCO) method aiming originally the density estimation problem, and propose a new bandwidth selection method for our two-sample setting. We justify the performance of the proposed method by establishing theoretical results such as oracle inequality, together with numerical simulation results.
