Smoothing technique and its applications in semidefinite optimization (Q879964)

The author extends the smoothing technique he has developed in two previous papers to the problems of semidefinite optimization. First (in Section 2) a simple technique is proposed for estimating the Lipschitz constant for the gradient of some symmetric function of eigenvalues of symmetric matrices. Then (in Section 3) an optimal method is described for solving a smooth convex optimization problem. This method will be used in the next sections for treating two applications, namely, the minimizations of the maximal eigenvalue and of the spectral radius of a symmetric matrix depending linearly on the design variables. For each of these problems an upper bound is given on the number of iterations. For the first application this number is bounded by O(\(1/\varepsilon\)) where \(\varepsilon\) is the required absolute accuracy of the problem while for the second one, this number is bounded by \((4/\delta)\sqrt{(1+\delta)r \log{r}}\) where \(\delta\) is the required relative accuracy and \(r\) is the maximal rank of the corresponding linear matrix. In particular, the latter method is a fully polynomial approximation scheme.