Asymptotic behavior of the empiric distribution of M-estimated residuals from a regression model with many parameters (Q1088338)

The author considers the general linear model \(Y_ i=x_ i'\beta +R_ i\), \(i=1,...,n\), where \(\{R_ 1,...,R_ n\}\) are i.i.d. with c.d.f. F, \(\{x_ 1,...,x_ n\}\) are (fixed) vectors in \({\mathbb{R}}^ p\) and \(\beta \in {\mathbb{R}}^ p\). Let \({\hat\beta}\) be an M-estimator defined by using the kernel \(\psi\) ; let \(\hat F_ n(x)\) denote the empirical distribution of the residuals \(Y_ i-x_ i'{\hat\beta}\), and let \(\hat F^*_ n\) be the empirical c.d.f. of the errors \(\{R_ i\}\). Under suitable smoothness conditions on \(\psi\), F, and the density \(F'=f\) it is shown that, for fixed x \[ \sqrt{n}(\hat F_ n(x)-\hat F^*_ n(x)- H_ n(x))-(p/\sqrt{n})g(x)\to^{p}0 \] where \(g(x)=af(x)\psi(x)+bf'(x)\) and \(H_ n(x)=(1/nd)f(x)\sum^{n}_{i=1}\psi (R_ i)\) if the design has a constant term (and \(H_ n(x)\) vanishes otherwise). The author also proves that if \(p/\sqrt{n}\to c\), \(\sqrt{n}(\hat F_ n(x)-F(x))\) converges weakly to a Gaussian process with drift given by the bias term cg(x).