A complete class for linear estimation in a general linear model (Q582736)

Consider a random vector Y in \(R^ n\) with expectation u and variance- covariance matrix V, where u and V are unknown. The author considers estimation of a parameter c'u by estimators of the form d'Y, \(c\in R^ n\), with mean square error risk. The author shows that any unique (up to equivalence with respect to risk) locally best estimator is admissible. In particular, any linear estimator being locally best in the relative interior of the canonical parameter set is admissible. Also, basing on a well-known necessary condition for admissibility, he shows that any admissible linear estimator may be presented as a limit of estimators being locally best in the relative interior. And then, he shows that any unique (up to equivalence with respect to risk) locally best estimators and their limits constitute a complete class, but not a minimal complete class.