A minimax bivariate normal model selection (Q1114247)

This paper considers the problem of choosing one among all possible submodels \(N((0,0),I_ 2)\), \(N((\theta_ 1,0),I_ 2)\), \(N((0,\theta_ 2),I_ 2)\), \(N((\theta_ 1,\theta_ 2),I_ 2)\) based on the observation X from \(N((\theta_ 1,\theta_ 2),I_ 2)\) where \((\theta_ 1,\theta_ 2)\in R^ 2\) and \(I_ 2\) is the \(2\times 2\) identity matrix. The general form of a model selection rule is given by \(I_{\underline f}(x)=(I_{f_ 1}(x),I_{f_ 2}(x))\) where \(f_ j\) is a Borel function with \(0\leq f_ j(x)\leq 1\) for \(j=1,2\) and \(I_{f_ j}(x)=1\) with probability \(f_ j(x)\) and 0 with probability \((1-f_ j(x))\). It is shown that there exists a minimax model selection rule relative to a loss function which takes into account both inaccuracy and complexity.