Simultaneous estimation and prediction using the expected coverage measure criterion (Q1120216)

Let X and Y be independent random vectors with \[ X\sim N_ n(C_ 1\beta,\sigma^ 2I_ n)\quad and\quad Y\sim N_ m(C_ 2\beta,C_ 0\sigma^ 2I_ m) \] where \(\beta \in {\mathbb{R}}^ q\), \(\sigma >0\) are unknown and \(C_ 1\), \(C_ 2\) and \(C_ 0\) are known. We observe X but not Y and we want intervals for \(a'Y\) for some \(a\in R^ m\). It is desired that the intervals R(a,X) cover \(a'Y\) in most cases and they should be narrow. Following \textit{D. Q. Naiman} [ibid. 12, 702-715 (1984; Zbl 0546.62046)], define a probability measure \(\mu\) on \({\mathbb{R}}^ m\), a positive weight function w(a,\(\sigma)\) and a bound 1-\(\alpha\in (0,1)\) and choose R to minimize \[ (1)\quad E\int w(a,\sigma)\ell (R(a,x))\mu (da) \] \[ subject\quad to\quad (2)\quad E\int I_{R(a,x)}(a'y)\mu (da)\geq 1-\alpha \quad for\quad all\quad \beta \in {\mathbb{R}}^ q,\quad \sigma >0. \] Here \(\ell\) is the Lebesgue measure on \({\mathbb{R}}\), \(I_ R(Z)=1\) if \(Z\in R\) and \(I_ R(Z)=0\) if \(Z\not\in R\). Naiman minimizes (1) uniformly in (\(\beta\),\(\sigma)\) subject to (2) with restrictions on R. This paper considers general invariance theory and proves that the optimal intervals are minimax. Applications to multivariate generalizations are also given.