A note on small-sample maximum probability estimation (Q1077830): Difference between revisions

Let X be a random vector with a joint probability density function f(x,\(\theta)\) with \(\theta\in \Omega\), where \(\Omega\) is an open subset of m-dimensional space. A gain is equal to one if the vector D of estimators of \(\theta\) is such that the vector \(D-\theta\) is in a region \(R(\theta)\) specified in m-dimensional space, and the gain is zero otherwise. Let \(\bar W(\delta,\theta)\) be the expected gain with respect to a decision rule \(\delta\) and a parameter \(\theta\). Let C be the class of decision rules defined as follows. A decision rule \(\delta\) is in C if and only if for each \(\theta^*\in \Omega\) and \(\epsilon >0\), each m-dimensional ball \(S(\theta^*,r,\epsilon)\) with center at \(\theta^*\) and positive radius r contains a subset of positive m-dimensional volume such that \[ | \bar W(\delta,\theta)- \bar W(\delta,\theta^*)| \leq \epsilon \text{ for every }\theta \in S(\theta^*,r,\epsilon). \] Then it is shown that, under a mild condition on \(f(x,\theta)\) and for a natural class of regions \(R(\theta)\), every decision rule is in the class C. An example is given where the maximum probability estimator has the expected gain function uniformly arbitrarily close to that of the Bayes decision rule. Very closely related results can be found in the author's paper, Stochastic Processes Appl. 14, 267-277 (1983; Zbl 0499.62013).