Empirical smoothing parameter selection in adaptive estimation (Q1208647): Difference between revisions

The paper deals with selecting the smoothing parameter involved in the construction of adaptive estimates for the symmetric location model. In \textit{P. J. Bickel's} [Ann. Stat. 10, 647-671 (1982; Zbl 0489.62033)] original paper on adaptivity a necessary condition for adaption was given. Moreover, some adaptive estimators were explicitly constructed. The idea of this construction was to replace the score function \(\phi=f'/f\), involved in the one-step maximum-likelihood estimation of the parameter, by an estimator obtained from a kernel estimate \(\widehat f_ \lambda\) of the density and its derivative \(\widehat f_ \lambda'\). Here \(\lambda\) is a smoothing parameter (the bandwidth) and its choice is crucial for the behavior of the estimator. The large sample theory for the choice of a sequence \(\lambda_ n\) depending on the sample size \(n\) is well established. The author offers alternatives for small to moderate sample sizes. He develops a new empirical smoothing parameter selection method. First, for fixed \(k\), the score function is estimated directly by linear \(B\)-splines \(\widehat \phi_ k\) where \(k\), the number of knots, is the smoothing parameter. Second, new empirical methods to select \(\widehat k_ n\), the smoothing parameter, and to locate the knots, are introduced. Moreover, asymptotic bounds on \(\widehat k_ n\) are derived. Adaptivity of the estimator is also proved. The results are also generalized for the general linear regression model with errors having a density symmetric about zero.