The space decomposition theory for a class of eigenvalue optimizations (Q457214)

The goal of the paper is to develop a framework for a second-order analysis for a class of eigenvalue optimization problems in which the eigenvalues are considered as functions of a symmetric matrix and each involved arbitrary eigenvalue function is a D.C. function, i.e., the difference of two convex functions. For this purpose, the \({\mathcal{U}}\)-Lagrangian theory is applied to such a D.C. function and, in particular, the first- and second-order derivatives of the \({\mathcal{U}}\)-Lagrangian in the space of decision variables are derived under the assumption that a transversality condition is satisfied. Moreover, a conceptual algorithm is presented and quadratic convergence is proved for that. As an application, the problem is studied to best approximate a given symmetric matrix by a low rank symmetric positive semidefinite matrix.