Validity of Edgeworth expansions of minimum contrast estimators for Gaussian ARMA processes (Q1088354)

This paper has as its end purpose a careful examination of higher order efficiency aspects of estimates of parameters of scalar Gaussian ARMA processes. In particular adjustments of maximum likelihood (ML) estimates, or of quasi maximum likelihood estimates (QML) to achieve asymptotic median unbiasedness (AMU) are considered. By a quasi maximum likelihood estimator is meant one minimising \[ (*)\quad \int^{\pi}_{-\pi}\{\log f_{\theta}(\lambda)+\frac{I_ T(\lambda)}{f_{\theta}(\lambda)}\}d\lambda,\quad I_ T(\lambda)=\frac{1}{2\pi \tau}| \sum^{T}_{1}X_ te^{it\lambda}|^ 2, \] \(f_{\theta}(\lambda)\) being the spectral density. An estimator \({\tilde \theta}\) is kth order AMU if \[ | P_{\theta}\{T^{1/2}({\tilde \theta}-\theta)\leq 0\}-1/2| =o(T^{- (k-1)/2}). \] The ML and QML procedures are generalised to a class of ''minimum contrast estimators'' satisfying an equation \[ T^{-1}X_ T'A(\theta)^{-1}B(\theta)A(\theta)^{-1}X_ T=b(\theta). \] Here A(\(\theta)\), B(\(\theta)\) are Toeplitz matrices replacing those corresponding to \(f_{\theta}(\lambda)\), \(\partial f_{\theta}(\lambda)/\partial \theta\) and b(\(\theta)\) corresponds to the term obtained by differentiating \((2T)^{-1} \log \det E(X_ TX_ T')\) or log \(f_{\theta}(\lambda)\) in (*). The technique used is the establishment of the validity of an Edgeworth expansion for \({\hat \theta}+T^{-1} m({\hat \theta})\) where \({\hat \theta}\) is the minimum contrast estimator and m(\({\hat \theta}\)) is chosen to achieve AMU, say for \(k=2.\) The paper is difficult to read, with a wealth of special notation, not all of which is well explained.