MIT 公开课 - 高等算法 lec14.pdfVIP

MIT OpenCourseWare 6.854J / 18.415J Advanced Algorithms Fall 2008 �� For information about citing these materials or our Terms of Use, visit: /terms. 18.415/6.854 Advanced Algorithms October 27, 2008 Lecture 14 Lecturer: Michel X. Goemans 1 Introduction For this lecture we’ll look at using interior point algorithms for solving linear programs, and more generally convex programs . Developing originally in 1984 by Narendra Karmarkar, there have been many variants (with some of the keywords ’path following’, ’primal-dual’, ’potential reduction’, etc.) on interior point algorithms, especially through the late 80s and early 90s. In the late 90s, people began to realize that interior point algorithms could also be used to solve semidefinite programs (or, even more generally, convex programs). As much as possible, we will discuss linear programming, semidefinite programming, and even a larger class called conic programming in a unified way. 2 Linear Programming We will start with linear programming. Remember that in linear programming, we have: Primal: Given A ∈ Rm ×n , c ∈ Rn and b ∈ Rm , find x ∈ Rn : Min cT x s.t. Ax = b, x ≥ 0. Its dual linear program is: Dual: Find y ∈ Rm : Max bT y s.t. AT y ≤ c. We can introduce non-negative slack variables and rewrite this as: Dual: Find y ∈ Rm , s ∈ Rn : Max bT y s.t. AT y + s = c, s ≥ 0. We know that, for a feasible solution, x in the primal, and a feasible solution (y , s) in the dual, we know by complementary slackne