A statistical diptych: Admissible inferences -- recurrence of symmetric Markov chains (Q1206701)

Given a statistical model \(P(dx\mid\theta)\) on a sample space \(X\) and an improper prior distribution \(\nu\) on the parameter space \(\Theta\), the marginal measure on \(X\) is given by \(M(dx)=\int P(dx\mid\theta) \nu(d\theta)\). It is assumed that \(\nu\) and \(M\) are \(\sigma\)-finite measures, so a formal posterior distribution exists. Consider a statistical decision problem with a risk function \(R(\delta,\theta)\), where \(\delta\) is a decision rule and \(\theta\in\Theta\) is a parameter. A decision rule \(\delta_ 0\) is almost-\(\nu\)-admissible (a-\(\nu\)-a) if for any decision rule \(\delta\) which satisfies \(R(\delta,\theta)\leq R(\delta_ 0,\theta)\) for all \(\theta\), the set \(\{\theta\mid R(\delta,\theta)<R(\delta_ 0,\theta)\}\) has \(\nu\) measure zero. A class of decision problems is introduced which are called quadratically regular problems. The quadratically regular problems include the problem of estimating a bounded measurable function of \(\theta\) and \(x\) when the loss is quadratic and a prediction problem when the loss is a quadratic measure of distance between distributions. A key result of the paper is that if for each measurable subset \(C\in\Theta\) satisfying \(0<\nu(C)<\infty\), \[ \inf_{h\in V(C)}\Delta(h)=0, \quad\text{where}\quad \Delta(h)=\iint (h(\theta)- h(\eta))^ 2 R(d\theta\mid\nu)\nu(d\eta) \] and \(V\) is a given class of real valued functions \(h\) on \(\Theta\), then for all quadratically regular decision problems, the formal Bayes rules induced by \(\nu\) are a-\(\nu\)-a.