Some properties of \([tr(Q^{2p})]^{1/2p}\) with application to linear minimax estimation (Q917459): Difference between revisions

This paper tackles the solution of the nondifferentiable problem to minimize f(C), where \(C\in R^{r\times n}\). Such problems arise in regression theory. The function f is convex and the level set \(W(C_ 0)=\{C|\) \(f(C)\leq f(C_ 0)\}\) is compact for each \(C_ 0\). Therefore the problem has a solution \(C'\). The obtained results determine that the use of nondifferentiable optimization is avoidable. Auxiliary problems \(f_ p(C)\to f(C)\) as \(p\to \infty\) are considered. As a result it is obtained that the solution \(C_ p\) is unique and tends to \(C'\). Therefore the solution of \(f_{p_ j}(C)\to Min!\) gives a starting point. This result is supported by six propositions and a corollary. An iteration procedure is proposed of the form \(C_{t+1}=C_ t+\lambda_ tS_ t\), where \(S_ t\in R^{r\times n}\) and \(\lambda_ t\) is a stepsize that satisfies a set of inequalities. A theorem states that \(C_ t\to C_ p\) and bounds the error estimates. Another theorem presents similar results for the computed solution. The related regression problem is solved for \(y\sim (XB,V)\). The solution is an approximate minimax inhomogeneous linear estimator.