Properties of a Cutting Plane Method for Semidefinite Programming1
نویسندگان
چکیده
We analyze the properties of an interior point cutting plane algorithm that is based on a semi-infinite linear formulation of the dual semidefinite program. The cutting plane algorithm approximately solves a linear relaxation of the dual semidefinite program in every iteration and relies on a separation oracle that returns linear cutting planes. We show that the complexity of a variant of the interior point cutting plane algorithm is slightly smaller than that of a direct interior point solver for semidefinite programs where the number of constraints is approximately equal to the dimension of the matrix. Our primary focus in this paper is the design of good separation oracles that return cutting planes that support the feasible region of the dual semidefinite program. The intersection of a cutting plane at a boundary point y with the tangent cone at that point gives a face of the tangent cone; we propose to use the dimension of this face to measure the strength of the cutting plane, we characterize the supporting hyperplanes that give higher dimensional faces of the tangent cone, and show how such cutting planes can be found efficiently. Our procedures are analogous to finding facets of an integer polytope in cutting plane methods for integer programming. We illustrate these concepts with two examples in the paper. We present computational results that highlight the strength of these cutting planes in a practical setting. Our technique of finding higher dimensional cutting planes can conceivably be used to improve the convergence of the spectral bundle method of Helmberg et al. [10, 11], and the non-polyhedral cutting surface algorithms of Sivaramakrishnan et al. [38] and Oskoorouchi et al. [27, 28].
منابع مشابه
Properties of a Cutting Plane Method for Semidefinite Programming
We analyze the properties of an interior point cutting plane algorithm that is based on a semi-infinite linear formulation of the dual semidefinite program. The cutting plane algorithm approximately solves a linear relaxation of the dual semidefinite program in every iteration and relies on a separation oracle that returns linear cutting planes. We show that the complexity of a variant of the i...
متن کاملA unifying framework for several cutting plane methods for semidefinite programming
Cutting plane methods provide the means to solve large scale semidefinite programs (SDP) cheaply and quickly. We give a survey of various cutting plane approaches for SDP in this paper. These cutting plane approaches arise from various perspectives, and include techniques based on interior point cutting plane approaches, and an approach which mimics the simplex method for linear programming. We...
متن کاملImplementation of a Cutting Plane Method for Semidefinite Programming
Semidefinite programming refers to a class of problems that optimizes a linear cost function while insuring that a linear combination of symmetric matrices is positive definite. Currently, interior point methods for linear programming have been extended to semidefinite programming, but they possess some drawbacks. So, researchers have been investigating alternatives to interior point schemes. K...
متن کاملA Toolbox for a Cutting { Plane Approachbased on Semidefinite
CUTSDP is a package of C programs containing an implementation of a cutting plane approach based on semideenite programming. This manual describes the package itself and also three applications: max-cut, graph bisection, and graph equipartition.
متن کاملLP and SDP branch-and-cut algorithms for the minimum graph bisection problem: a computational comparison
While semidefinite relaxations are known to deliver good approximations for combinatorial optimization problems like graph bisection, their practical scope is mostly associated with small dense instances. For large sparse instances, cutting plane techniques are considered the method of choice. These are also applicable for semidefinite relaxations via the spectral bundle method, which allows to...
متن کامل