Conditional empirical processes (Q1077107)

Let \((X_ 1,Y_ 1),(X_ 2,Y_ 2),..\). be a sequence of independent identically distributed random vectors such that \(X_ n\) and \(Y_ n\) take values in R and \(R^ d\), \(d\geq 1\), respectively. In the case where \(d=1\), the author [ibid. 12, 917-926 (1984; Zbl 0539.62026)] introduced a nearest-neighbour-type estimate for the regression function \(E(Y| X=x)\) and established its asymptotic normality. To extend these results, a similar kernel-type estimate \(m_ n(y| x_ 0)\) is defined for the conditional distribution function (d.f.) \(m(y| x_ 0)=P(Y\leq y| X=x_ 0)\), for \(x_ 0\in R\), \(y\in R^ d.\) The main result gives conditions under which \(m_ n(\cdot | x_ 0)- m(\cdot | x_ 0)\), suitably normed, converges in distribution for Lebesgue-almost all \(x_ 0\in (0,1)\) to a centred Gaussian process. It is also shown that \(m_ n(\cdot | x_ 0)\) can be modified to produce a bona fide d.f. \(m^*\!_ n(\cdot | x_ 0)\) which obeys the same Donsker-type invariance principle. These results are applied to prove the asymptotic normality of conditional quantile functions and to derive distribution-free techniques for conditional d.f.'s.