Asymptotic properties of Neyman-Pearson tests for infinite Kullback- Leibler information (Q1095523): Difference between revisions

Classical results of Stein, Chernoff and Rao concerning the rate of convergence of second kind error probabilities \(\beta_{\alpha,n}\) are generalized for the case when the Kullback-Leibler information \(K(P_ 0,P_ 1)\) is infinite. The main result is the following: Let \(q_{\alpha,n}\) denote the logarithm of the critical value of the Neyman-Pearson test of size \(\alpha\) when testing H: P\(=P_ 0\) against K: P\(=P_ 1\) for the sample size n. Moreover let F be the distribution function of \(\log (dP_ 1/dP_ 0)\) with respect to \(P_ 0\). If F is positive and \(\limsup_{x\to -\infty}[F(\lambda_ 0x)/F(x)]<1\) for some \(\lambda_ 0>1\), then \(\lim_{n\to \infty}\beta_{\alpha,n}^{(1/q_{\alpha,n})}=\exp \{-1\}.\) From this follows the interesting fact that in contrast to the classical results the rate of convergence of second kind error probabilities and critical values depends on the level \(\alpha\) if \(K(P_ 0,P_ 1)=\infty.\) Moreover the relation between \(q_{\alpha,n}\) and the local behaviour of the Laplace transform of F is studied. An application to a one-sided test problem is discussed in details.