From Stein's unbiased risk estimates to the method of generalized cross- validation (Q1083798)

Let \(y_ i=\mu_ i+\epsilon_ i\), \(i=1,...,n\), where \(y_ 1,...,y_ n\) are n independent observations with unknown means \(\mu_ 1,...,\mu_ n\) and \(\epsilon_ i\) has mean 0 and common variance \(\sigma^ 2\). To estimate \({\underset \tilde{} \mu}=(\mu_ 1,...,\mu_ n)'\), a class of linear estimators \({\hat \mu}\)(h)\(=M(h)\underset \tilde{} y\) is considered, where h is an indexing number in an index set H, M(h) is an \(n\times n\) matrix and \(\underset \tilde{} y=(y_ 1,...,y_ n)'\). The generalized cross-validation (GCV) method chooses h by minimizing \[ \| \underset \tilde{} y-{\hat \mu}(h)\|^ 2/(1-n^{-1}tr M(h))^ 2 \] where ''tr'' means trace and \(\| \cdot \|\) denotes the Euclidean norm. The GCV method may be applied to problems in model selection, ridge regression, smoothing splines, for example. In this paper a new approach is given to the GVC, based on Stein estimates and the associated unbiased risk estimates. Consistency results are obtained for the cross-validated (Steinized) estimates. A variant of the GCV is proposed for the case in which the dimension of the raw data is known to be greater than that of their expected values.