Uniform in bandwidth consistency of kernel-type function estimators (Q2569245)

The authors consider kernel estimators based on iid data in density and regression estimation. Their aim is to prove uniform consistency with rates, where uniformity is meant w.r. to the argument and the smoothing parameter. E.g., let \((X_i)_1^\infty\) be an iid sequence of random vectors following a d-dimensional bounded density \(f\,.\) Under some assumptions on the kernel it is shown that the estimator \(f_{n,h}(x)=(nh)^{-1} \sum_{k=1}^n K((x-X_k)/h^{1/d})\) satisfies with probability one \[ \limsup_{n \to \infty} \sup_{c \log n\,/n\leq h\leq 1}\frac{\sqrt{nh}\| \widehat{f}_{n,h}- E(\widehat{f}_{n,h})\| _\infty}{\sqrt{\log(1/h)\vee \log \log n}}= \kappa(c)<\infty. \] From this result one can, e.g., conclude consistency of estimators with data-driven bandwidths. Related results are shown for the Nadaraya-Watson estimator in regression and for an estimator for conditional distributions. One important tool for the proof is empirical process theory applied to the family of functions \({\mathcal K}=\{K(x-.)/h^{1/d})\), \(x \in {\mathbb R}^d\), \(h>0\}.\)