On Kullback-Leibler loss and density estimation (Q1124241)

The paper deals with optimal estimation of Kullback-Leibler loss and density function. The estimate of Kullback-Leibler loss is based on the non-parametric kernel density estimate. Its consistency and rate are studied and the smoothing parameter is chosen in the cross-validation approach. It is observed that the asymptotic properties of the estimate are profoundly influenced by the tails of both the kernel and the unknown density function. Standard consistency conditions on the smoothing sequence in density estimation such as \(h_ n\to 0\) and \(nh_ n\to \infty\) are not sufficient for expected Kullback-Leibler loss convergence.