Density-Difference Estimation

摘要: We address the problem of estimating difference between two probability densities. A naive approach is a two-step procedure first densities separately and then computing their difference. However, this does not necessarily work well because step performed without regard to second step, thus small estimation error incurred in stage can cause big stage. In letter, we propose single-shot for directly density derive nonparametric finite-sample bound proposed density-difference estimator show that it achieves optimal convergence rate. how be used L2-distance approximation. Finally, experimentally demonstrate usefulness method robust distribution comparison such as class-prior change-point detection.