MINIMAX ESTIMATORS FOR A MULTINORMAL PRECISION MATRIX
MINIMAX ESTIMATORS FOR A MULTINORMAL PRECISION MATRIX
复制标题
DOI:
10.1016/0047-259x(77)90079-3
复制
发表时间:
1977-01-01
影响因子:
1.6
通讯作者:
HAFF, LR
中科院分区:
文献类型:
--
作者:
HAFF, LR
Let S p× p∼ Wishart (Σ, k), Σ unknown, k> p+ 1. Minimax estimators of Σ− 1 are given for L 1, an Empirical Bayes loss function; and L 2, a standard loss function (R i≡ E (L i∣ Σ), i= 1, 2). The estimators are Σ ̂− 1= a S− 1+ br (S) I p× p, a, b≥ 0, r (·) a functional on R p (p+ 2) 2. Stein, Efron, and Morris studied the special cases Σ a− 1= a S− 1 (E Σ ̂ k− p− 1− 1= Σ− 1) and Σ ̂ 1− 1= a S− 1+(b/tr S) I, for certain, a, b. From their work R 1 (Σ− 1, Σ ̂ 1− 1; S)≤ R 1 (Σ− 1, Σ ̂ a− 1; S)(∀ Σ), a= k− p− 1, b= p 2+ p− 2; whereas, we prove R 2 (Σ− 1 Σ ̂ a− 1; S)≤ R 2 (Σ− 1, Σ ̂ 1− 1; S)(∀ Σ). The reversal is surprising because L 1 (Σ− 1, Σ ̂ 1− 1; S)→ L 2 (Σ− 1, Σ ̂ 1− 1; S) ae (for a particular L 2). Assume R (compact)⊂ S, S the set of p× p psd matrices. A “divergence theorem” on functions F p× p: R→ S implies identities for R i, i= 1, 2. Then, conditions are given for R i (Σ− 1, Σ ̂− 1; S)≤ R i (Σ− 1, Σ ̂ 1− 1; S)≤ R i (Σ− 1, Σ ̂ a− 1; S)(∀ Σ), i= 1, 2. Most of our results concern estimators with r (S)= t (U)/tr (S), U= p∣ S∣ 1/p/tr (S).