An Edgeworth expansion for U-statistics based on samples from finite populations (Q916197)

Let \(\{w_{ij}\}\) be an \(N\times N\) array of real numbers satisfying \(w_{ij}=w_{ji}\) and \(w_{ii}=0\). Let \((R_ 1,...,R_ N)\) be a vector selected at random from the set of N! of permutations of (1,2,...,N). For \(1<n<N\) we define \[ U=\sum^{n}_{i=2}\sum^{i- 1}_{j=1}w_{ij},\text{ where } w_{ij}=w_{R_ i}w_{R_ j}. \] Put \(p=n/N\), \(q=1-p\), \(w_ i=\sum^{N}_{j=1}w_{ij}\), \(\nu^ 2=p^ 3q\sum^{N}_{i=1}w^ 2_ i.\) The following Edgeworth expansion for \(\nu^{-1}U\) is derived: \[ F(x)=\Phi (x)-H_ 2(x)\phi (x)\{((q- p)/6(pq)^{1/2})(\sum^{N}_{k=1}w^ 3_ k)(\sum^{N}_{k=1}w^ 2_ i)^{-3/2}+ \] \[ +(q/p)^{1/2}(\sum^{n}_{i=2}\sum^{i- 1}_{j=1}w_{ij}w_ iw_ j)(\sum^{N}_{k=1}w^ 2_ i)^{-3/2}\} \] where \(\Phi (x)=\int^{\infty}_{-\infty}\phi (t)dt\), \(\phi (x)=(2\pi)^{-1/2}e^{-x^ 2/2}\), \(h_ 2(x)=(x^ 2-1)\).