The double kernel method in density estimation (Q914283): Difference between revisions

The author considers kernel estimators \[ \hat f_ n(x,h)=(nh)^{- 1}\sum^{n}_{j=1}K((x-X_ j)/h), \] based on i.i.d. samples \((X_ 1,...,X_ n)\) with density f on \({\mathbb{R}}\). The major problem in this context is to find a good smoothing parameter h. Various data-driven bandwidth choices have been suggested, mainly based on \(L_ 2\)-error criteria. This paper is concerned with \(L_ 1\)-convergence and suitable bandwidth choices H. The aim is to find asymptotically optimal procedures, which means \[ E(\int | \hat f_ n(x,H)-f(x)| dx)/\inf_{h>0}E(\int | \hat f_ n(x,h)-f(x)| dx)\to 1,\quad as\quad n\to \infty, \] in analogy to the result of \textit{C. J. Stone}, Ann. Stat. 12, 1285-1297 (1984; Zbl 0599.62052) for the \(L_ 2\)-error based on a CV-bandwidth. Cross validation is not useful in the \(L_ 1\)- case, so the author suggests the following method: Use a second kernel estimator \(\tilde f\) based on a different kernel L. The Fourier transforms of K and L should be different near 0 (e.g. K and L are kernels of different order). Then, choose H as the minimizer of \(\int | \hat f_ n(x,h)-\tilde f_ n(x,h)| dx.\) The following results give a flavour of the theorems obtained: 1.) E(\(\int | \hat f_ n(x,H)-f(x)| dx)\to 0\), \(n\to \infty\), for all densities f on \({\mathbb{R}}.\) 2.) Assume that \(f^{(s-1)}\) is absolutely continuous on \({\mathbb{R}}\) and \(\int \sqrt{f}dx<\infty\), then \[ \limsup_{n\to \infty}\{E \int | \hat f_ n(x,H)-f(x)| dx/\inf_{h>0} E \int | \hat f_ n(x,h)- f(x)| dx\}\leq c(L,K) \] (c-optimality), provided K is a kernel of order s, so that f is in the saturation class of K. 3.) \(H\to 0\), nH\(\to \infty\) and \(\int | \hat f_ n(x,H)-f(x)| dx\to 0\), completely, for an arbitrary density f on \({\mathbb{R}}\).