Admissible linear estimators in the general Gauss-Markov model (Q1110221)

The authors consider the Gauss-Markov linear model, \(y=X\beta +\sigma \epsilon\), \(E(\epsilon)=0\), \(E(\epsilon \epsilon ')=V\). The problem is to characterize linear admissible estimators \(Fy+f\) of a parametric function \(K\beta\) under quadratic risk. After a review of some of the literature available on this topic the authors first give the solution in the Zyskind-Martin model. Then the solution of the general problem for estimable \(K\beta\), \(K=CX\), is given as follows: The NASC is given by the five properties \[ (i)\quad im(VF')\subseteq im(X);\quad (ii)\quad (FVC)'=FVC;\quad (iii)\quad FVF'\leq FVC; \] \[ (iv)\quad im((F-C)X)=im((F-C)W),\quad im(W)=im(X)\cap im(V), \] \[ (v)\quad f\in im((F-C)X). \] This general result is then compared with the results known from literature.