Asymptotic performance bounds for the kernel estimate (Q1119308)

The author considered an arbitrary sequence of kernel density estimates \(f_ n\) with kernels \(K_ n\) possibly depending on the number of observations (which are independent and identically distributed). Under a mild restriction on the sequence \(K_ n\), the author obtained the lower bounds for the expected \(L_ 1\)-error. As an example, he considered densities with finite and infinite characteristic functions and also densities with a large tail. In the second half of the paper, the author restricted the class of kernels to \(K_ n\in K_ s\), where \(K_ s\) is a saturation class such as the class of kernels of the form \(h^{-1}L(x/h)\) for symmetric L where all the moments of L up to and not including the s th moment are zero.