A theorem on large-deviation probabilities for decomposable statistics and its statistical application (Q1061425)

Let \(X_ 1,...,X_ n\) be a sample of multinomial trials with outcomes from the set \(\{\) 1,...,N\(\}\) under probabilities \(p_ 1,...,p_ N\). In order to test a simple hypothesis \(H_ 0:p_ k=q_ k,\) \(k=1,...,N\), against a composite alternative \(H:p_ k\neq q_ k\) for some k, the author considers test statistics of the form \(L_ n=\sum^{N}_{k=1}f(h_ k^{(n)}),\) where f is a function growing not faster than x ln x (x\(\to \infty)\), and \(h_ k^{(n)}\) is the frequency of outcome k in the first n trials. Using moment-generating function techniques the limiting behaviour of \(N^{-1}\ln P(L_ n>c_ n)\) is studied which serves as a measure of efficiency of the test rejecting \(H_ 0\) for large values of \(L_ n\). Here \(c_ n\) is a certain sequence of critical values, and the author obtains results under equiprobable probabilities \(p_ 1=...p_ N=1/N\) as well as under alternative distributions. Some particular cases of f are discussed in more detail.