Integration by parts with empirical measure

152 Views Asked by At

I'm currently reading through the paper Asymptotic normality of nearest neighbor regression function estimates and am struggling to understand the asymptotic equalities that were shown in the proof of Lemma 4 (p. 923) where the last one was proven by using integration by parts:

$\begin{align*}&\qquad \frac{m(x_0)}{\sqrt{a_n^3}}\int \left[\alpha_n(x_0)-\alpha_n(x)\right]K'\left(\frac{F(x_0)-F(x)}{a_n}\right)F(d x)\\ &= -\frac{m(x_0)}{\sqrt{a_n^3}}\int \alpha_n(x)K'\left(\frac{F(x_0)-F(x)}{a_n}\right)F(d x)\\ &= -\frac{m(x_0)}{\sqrt{a_n}}\int K\left(\frac{F(x_0)-F(x)}{a_n}\right)\alpha_n(dx). \end{align*}$

Here $m(x) = E[Y|X=x]$, $\alpha_n(x) = \sqrt{n} [F_n(x)-F(x)]$ denotes the empirical process of the random variable $X$ (having continuous distribution function $F$), $K$ being a twice continuously differentiable probability kernel with bounded support and $na_n^3 \to \infty$, $a_n \to 0$ for $n \to \infty$.

I know that I could expand the second to last term like so: $\begin{align} &\quad-\frac{m(x_0)}{\sqrt{a_n^3}}\int \alpha_n(x)K'\left(\frac{F(x_0)-F(x)}{a_n}\right)F(d x)\\ &= -\frac{m(x_0)\sqrt{n}}{\sqrt{a_n^3}}\int F_n(x)K'\left(\frac{F(x_0)-F(x)}{a_n}\right)F(d x) + \frac{m(x_0)\sqrt{n}}{\sqrt{a_n^3}}\int F(x)K'\left(\frac{F(x_0)-F(x)}{a_n}\right)F(d x)\\ \end{align}$

where the second integral contains transformations with $F$, so a transformation of variables and the integral a la $u = F(x)$ is possible and then application of the "classic" integration by parts yields the first of the following terms and a "constant" term.

$\begin{align} &= \frac{m(x_0)\sqrt{n}}{\sqrt{a_n}}\int K\left(\frac{F(x_0)-F(x)}{a_n}\right) F(d x) -\frac{m(x_0)\sqrt{n}}{\sqrt{a_n}}\int K\left(\frac{F(x_0)-F(x)}{a_n}\right) F_n(d x)\\ &= -\frac{1}{\sqrt{a_n}}m(x_0)\int K\left(\frac{F(x_0)-F(x)}{a_n}\right) \alpha_n(d x)\end{align}$

It would be nice if someone could provide me an explanation on the upper equalities since I'm not even sure if I'm on the right track. I only consider the classic integration by parts, maybe the Lebesgue-Stieltjes variant is more helpful.