1997 年 27 巻 1 号 p. 19-35
This paper discusses the problem of classifying an observed stretch X=(x1, …, xr) into Π1 or Π2, where Πi is a Gaussian stationary process with zero mean and spectral density fθi(λ). We propose a new discriminant statistic based on some estimator θ=θ(X) of a spectral parameter. The statistic D[θ, W] is motivated by a spectral measure with divergence function W. Most of the work presented is devoted to higher order asymptotic theory when θ2 is contiguous to θ1, in order to study the asymptotic difference between different D[θ, W]. In particular, it is shown that for any choice of W, D[θ, W] has the same second order averaged risk as the optimal likelihood ratio (LR) if θ belongs to an appropriate class of asymptotically efficient estimators, and the third order term of the averaged risk is minimized by the (bias-adjusted) maximum likelihood estimator (MLE). We also examine the case of the rule based on the MLE without bias adjustment.