Jordan algebras and Bayesian quadratic estimation of variance components (Q1189634)

In a variance-component model there are in general no uniformly best (quadratic) unbiased estimators for linear functions of variance components. \textit{J. Seely} [Ann. Math. Statistics 42, 710-721 (1971; Zbl 0249.62067)] has found a necessary and sufficient condition for their existence. He found that a certain class of square matrices should form a Jordan algebra, or, as he called it, a quadratic subspace. This condition is in particular met if the variance-component model under investigation is a balanced mixed linear model. It is the purpose of this paper to give some properties of Jordan algebras: simple formulae for \(A\circ B\), \(A^ +\), \(A^ -\); solutions of matrix equations \(AXB=C\), and so on. Applications are the determination of locally optimal estimators and Bayesian quadratic estimators of variance components.