No label defined (Q4025599)

Consider the linear model \(Y=X\beta+\varepsilon\), where the random variable \(\varepsilon\) is normally distributed \(N_ n(0,\sigma^ 2 V)\), \(X\) and \(V\) are known \(n\times p\) and \(n\times n\) matrices, respectively, \(\text{rank}(X)=p<n\), \(V\) is a positive definite matrix, \(\beta\in\mathbb{R}^ p\) is a parametric vector, and \(\sigma>0\) is known or unknown. Let \[ \widehat\beta =(X'V^{-1}X)^{-1}X'V^{-1}Y, \qquad S^ 2=Y'[V^{-1}-V^{-1}X(X'V^{-1}X)^{-1}X' V^{-1}]Y. \] Consider the loss function \((*)\) \([a-(\widehat\beta-\beta) (\widehat\beta- \beta)]^ 2\). The author proved the following results: (1) If \(\sigma\) is known, and \(A=(X'V^{-1}X)\), then for \(p\leq4\), \(\sigma^ 2\text{tr}(A)\) is a consistent minimum risk unbiased and admissible estimator of \((\widehat\beta-\beta) (\widehat\beta-\beta)\), and for \(p\geq 5\), \(\sigma^ 2\text{tr}(A)\) is inadmissible, for the loss function \((*)\). (2) If \(\sigma\) is unknown, \((n-p)^{-1}S^ 2\text{tr}(A)\) is a consistent minimum risk unbiased estimator of \((\widehat\beta-\beta) (\widehat\beta-\beta)\) for \((*)\). For the loss function \([a-\sigma^{-2}(\widehat\beta-\beta) (\widehat\beta-\beta)]^ 2\), and \(p\leq 4\), \(\text{tr}(A)\) is a consistent minimum risk unbiased and admissible estimator of \(\sigma^{- 2}(\widehat\beta-\beta) (\widehat\beta-\beta)\), but for \(p\geq 5\), it is inadmissible.