Admissible estimation, Dirichlet principles and recurrence of birth-death chains on \({\mathbb{Z}}^ p_+\) (Q1075707)

The problem considered is that of estimating unknown means \(\lambda_ i\) of each of p independent Poisson observations \(X_ i\), with loss \[ \sum^{p}_{i=1}\lambda_ i^{-1}(d_ i(x)-\lambda_ i)^ 2 \] incured from an estimator \(d(x)=(d_ i(x);\quad 1\leq i\leq p),\) where \(x=(x_ 1,...,x_ p).\) The paper provides probabilistic descriptions and explicit criteria for admissibility of d(x). It has been known for some time that the ML estimator \(d(x)=x\) is inadmissible for \(p\geq 2\), a result analogous to the famous discovery of Stein in the multivariate normal case. The theory developed in this paper is parallel to the one obtained earlier for the normal case [Admissibility from recurrence via Poincaré inequalities in the Gaussian case. Manuscript. (1983)]. Noting that the search for admissible rules may, roughly, be confined to the class of generalized Bayes procedures, the author associates a reversible Markov chain \(\{X_ t\}\) on \(Z^+_ p\) with each generalized Bayes procedure. The main results are that the question of admissibility of such a procedure can be described as a variational problem in probabilistic potential theory associated with \(\{X_ t\}\) and that admissibility is in fact equivalent to recurrence of \(\{X_ t\}\). The variational problem is closely related to the Dirichlet principle for reversible chains, studied recently by \textit{D. Griffeath} and \textit{T. M. Liggett} [Ann. Probab. 10, 881-895 (1982; Zbl 0498.60090)] and \textit{T. Lyons} [ibid. 11, 393-402 (1983; Zbl 0509.60067)].