A nearest neighbour-estimator for the score function (Q2266537): Difference between revisions

Let f(\(\cdot)\) be a strictly positive density function defined on \((a,b)\subseteq R^ 1\) with a continuous derivative f'(\(\cdot)\) and let \(F(x)=\int^{x}_{a}f(t)dt\), \(-\infty \leq a<x<+\infty\), be the corresponding distribution function. Define the quantile function Q of F by \(Q(y)=F^{-1}(y)=\inf \{x:\) F(x)\(\geq y\}\), \(0<y<1\); the score function (-1)J of the density function f by \(J(y)=f'(Q(y))/f(Q(y))\), and the Fisher information I(f) of f by \(I(f)=\int^{1}_{0}(J(y))^ 2dy\), assumed to be finite. Given some regularity conditions on F, we propose a sequence of nearest neighbour type estimators \(J_ n\) for J and prove that for all \(\epsilon\in (0,1/5)\) there exists an estimator \(J_{n,\epsilon}\) of J such that for all \(\delta \in (0,(5\epsilon /18)\wedge (\epsilon /12+1/40))\) we have \[ \sup_{n^{-\delta}\leq y\leq 1-n^{- \delta}}| J_{n,\epsilon}(y)-J(y)| =^{a.s.}O(n^{- 1/5+\epsilon}(\log n)^{1/2}),\quad and\quad I_ n(f)\to^{a.s.}I(f), \] \[ where\quad I_ n(f)=\int^{1}_{0}(\bar J_ n(y))^ 2dy,\quad with\quad \bar J_ n(y)=J_{n,\epsilon}(y)\quad if\quad n^{- \delta}\leq y\leq 1-n^{-\delta}\quad and\quad zero\quad otherwise. \]