Large sample properties of kernel-type score function estimators (Q908630)

Kernel-type score function estimators are proposed and strong uniform consistency and strong approximation results are obtained for these estimators. Let \(X_ 1,...,X_ n\) denote a random sample from a population with distribution function F and differentiable density f. Let \(J(x)=f'(x)/f(x)\) and \(L(t)=J(Q(t))\), where \[ Q(t)=\inf \{x: F(x)\geq t\},\quad 0<t<1, \] is the inverse (quantile) function of F. The kernel estimator for f is defined as \[ f_ n(x)=(nb_ n)^{- 1}\sum^{n}_{j=1}k[(x-X_ j)b_ n^{-1}] \] where k is a chosen density function while the kernel estimator for \(f'\), the derivative of f, is defined as \[ f_ n'(x)=n^{-1}b_ n^{- 2}\sum^{n}_{j=1}k'((x-X_ j)b_ n^{-1}). \] The natural estimator for J is \(J_ n(x)=f_ n'(x)/f(x)\) and for L is \(L_ n(t)=J_ n(Q_ n(t))\), \(0<t<1\), where \[ Q_ n(t)=\inf \{x: F_ n(x)\geq t\},\quad 0<t<1, \] and \(F_ n(x)\) is the empirical distribution function. Under smoothness and boundedness conditions, it is shown that \[ \sup_{\alpha_ n<x<\beta_ n}| J_ n(x)-J(x)| =O(r_{1n}) \] where \(\{\alpha_ n\}\) is nonincreasing, \(\{\beta_ n\}\) is nondecreasing, and \(r_{1n}\) is related to \(b_ n\), \(n^{-1/2}(\log \log n)^{1/2}\), \(\sup_{\alpha_ n<x<\beta_ n}f'(x)[f(x)]^{-2}\) and \(\sup_{\alpha_ n<x<\beta_ n}[f(x)]^{-1}\). Under additional conditions, it is shown that \[ \sup_{n^{-\gamma}\leq t\leq 1-n^{- \gamma}}| L_ n(t)-L(t)| =O(r_{3n})\quad a.s.,\quad where \] \[ r_{3n}=\max \{n^{2\gamma +2\delta -1/2}(\log \log n)^{1/2}, n^{2\gamma -2\delta}\}\quad with\quad 2\gamma +2\delta <1/2\quad and\quad 0<\gamma <\delta. \] Finally, the process \(J_ n(x)-J(x)\), properly normalized, is strongly approximated by a Gaussian process, and the asymptotic distribution of the maximal deviation of this process is deduced. Corresponding results are given for the process \(L_ n(x)- L(x)\).