PI: Florian Potra DMS-9706894 Interior Point Methods for Semidefinite Programming Abstract Further investigation of interior point algorithms for semidefinite programming (SDP) is proposed with emphasis on the study of global and local convergence, infeasibility detection and software development. While the vast majority of interior point methods for linear programming (LP) have proven polynomial complexity, this is not the case for SDP. The iteration complexity of interior point algorithms for SDP and its dependence on the search direction as well as on the central path neighborhood used by the algorithm will be investigated. Locally superlinearly convergent algorithms will be identified within the class of interior point methods for SDP with proven polynomial complexity. Superlinear convergence is especially important for SDP since no finite termination schemes exist for such problems. The local convergence analysis for interior point algorithms for SDP is much more challenging than those for LP and it will be a major focus point of the project. A C++ package for solving large-scale SDP problems will be developed. The code will efficiently handle different sparsity patterns arising in applications, and will provide refined infeasibility detectors. Semidefinite programming (SDP) represents one of the most important classes of optimization problems with many applications in science and engineering especially related to optimal control (in electrical engineering), optimal allocation of resources (in economics and manufacturing), structural optimization (in civil engineeering), etc. According to a recent survey paper, semidefinite programming is ``the most exciting development in mathematical programming in the 1990's''. Until recently no efficient methods were known for solving general semidefinite programming problems. In the late 80's it has been realized that interior point methods, initially developed for linear programming, can be successfully used for solving semidefinite programming problems. The advent of interior point methods has created new opportunities for the application of SDP in science and engineering. In order for these opportunities to materialize it is important to provide the scientific community with a good theoretical understanding of the behaviour of interior point methods for SDP and with reliable and efficient software capable of solving large-scale SDP problems. The project will investigate some of the most critical issues in the theory of semidefinite programming and will lead to the design of efficient practical algorithms. The software resulting from this project is likely to have a positive impact on several application areas in science and technology.
