Cone-LP ' s and Semidefinite Programs : Geometry and a Simplex-Type Method

A b s t r a c t . We consider optimization problems expressed as a linear program with a cone constraint. Cone~LP's subsume ordinary linear programs, and semidefinlte programs. We study the notions of basic solutions, nondegeneracy, and feasible directions, and propose a generalization of the simplex method for a large class including LP's and SDP's. One key feature of our approach is considering feasible directions as a sum of t~0o directions. In LP, these correspond to variables leaving and entering the basis, respectively. The resulting algorithm for SDP inherits several important properties of the LP-simplex method. In particular, the linesearch can be done in the current face of the cone, similarly to LP, where the linesearch must determine only the variable leaving the basis.