Dimension reduction for semidefinite programs via Jordan algebras (Q2188241)

A new approach for simplifying semidefinite optimization problems inspired by the technique of symmetry reduction is proposed, where if an orthogonal projection map satisfies certain invariance conditions, an equivalent primal-dual pair of optimization problems over a lower-dimensional symmetric cone is derived. An algorithm for minimizing the rank of this projection and hence the dimension of the involved subalgebra is proposed and implemented as well. By means of Jordan algebras, the proposed method can be easily extended to linear, second-order cone and even symmetric cone optimization. Moreover, it is proven that rank minimization has a beneficial effect on the direct-sum decomposition of the algebra into simple ideals, leading to an optimal ``block-diagonalization'' of the initial semidefinite optimization problem. Combinatorial versions of the proposed algorithm illustrate its effectiveness on concrete examples.