Convergence results for maximum likelihood type estimators in multivariable ARMA models (Q580858): Difference between revisions

This paper seems in part to have been motivated by the results in \textit{W. Dunsmuir} and the reviewer, Advances Appl. Probab. 8, 339-364 (1976; Zbl 0327.62055), which we refer to below as DH. The situation basically being studied is one where data are generated by a stationary ergodic vector ARMA process belonging to some family, for example the manifold of all such systems of given McMillan degree or with given Kronecker indices. The objective is to prove the strong law of large numbers for the maximum likelihood (ML) estimate constructed as if the data were Gaussian, but assuming only ergodicity and the existence of second moments. If the observed data vectors are y(t), \(t=1,...,T\), and \(L_ T(f)\) is \(2T^{- 1}\) by the likelihood, f(\(\omega)\) being the spectral density, then in DH it is shown that \[ \limsup_{T\to \infty}L_ T(\hat f)\leq \inf_{\theta}\limsup_{T\to \infty}L_ T(f_{\theta})=\inf_{\theta} L(f_{\theta}) \] when \(\hat f\) is the ML estimate, inf is over \(\theta\) such that \(\det f_{\theta}(\omega)\neq 0\), \(\omega\in [-\pi,\pi]\) and \[ L(f_{\theta})=\log \det \Sigma +(1/2\pi)tr.\int f_{\theta}^{-1}(\omega)f_ 0(\omega)d\omega, \] \(f_ 0\) being the true spectral density at \(\theta =\theta_ 0\). Condition B6 of DH asserts that \(\inf_{\theta}L(f_{\theta})=\log \det \Sigma +\dim y(t)\). While this is obviously true if det \(f_ 0\neq 0\), \(\omega\in [-\pi,\pi]\), the proof that B6 holds more generally is incomplete in DH, as the author of the present paper points out. He gives a very detailed, very precise and very general treatment of the overall problem. Consideration is given to the situation where convergence may be not to a point but to a set, as in the situation where the true point lies at the boundary of the model set and may therefore be lower dimensional. Failing B6 the author points out, in particular, that, since the essential part of \(L_ T(f_ 0)\) is \(T^{-1}y_ T'\Gamma_ T(f_ 0)y_ T\), where \(y_ T\) is the vector of all data, \(\Gamma_ T(f_ 0)=E(y_ Ty_ T')\), the strong law will hold if \(T^{-1}y_ T'\Gamma_ T(f_ 0)y_ T\to \dim y_ T\) a.s. This is clearly true if y(t) is Gaussian. It seems probable, incidentally, that B6 does hold for important cases, e.g. systems for given McMillan degree or given Kronecker indices.