Minimaxity of Pitman estimators (Q1093271)

It is known that the Pitman estimator is minimax for translation invariant experiments with a Euclidean parameter space. This result is extended to experiments whose parameter space is homogeneous under the operation of a topological group. Let \(\mu\) be G-invariant and E be \(\mu\)-integrable of order \(p+1\) (p\(\geq 0)\). Let F be a posterior distribution, given \(\mu\), and let the loss function W be separating, level compact of order p, and G-invariant. Then the following results hold: (i) there exist non-randomized Pitman estimators, (ii) if E is G- invariant then every Pitman-estimator is equivariant under a given condition, and (iii) if E is G-invariant and G is locally compact then every equivariant Pitman estimator is minimax, under a set of given conditions.