CHAPTER 2
A Primal-Dual Interior Point
Algorithm for Linear
Programming
Masakazu Kojima, Shinji Mizuno, and Akiko Yoshise
Abstract. This chapter presents an algorithm that works simultaneously on
primal and dual linear programming problems and generates a sequence of pairs
of their interior feasible solutions. Along the sequence generated, the duality gap
converges to zero at least linearly with a global convergence ratio (1 - Yf/n); each
iteration reduces the duality gap by at least Yf/n. Here n denotes the size of the
problems and Yf a positive number depending on initial interior feasible solutions
of the problems. The algorithm is based on an application of the classical
logarithmic barrier function method to primal and dual linear programs, which
has recently been proposed and studied by Megiddo.
§1. Introduction
This chapter deals with the pair of the standard form linear program and its
dual.
(P)
(D)
Minimize c T x
subject to XES = {x E R
n
: Ax = b, x O}.
Maximize b T Y
subject to (y,z) E T = {(y,z) E Rm+n: ATy + Z = c, Z
Here R
n
denotes the n-dimensional Euclidean space, cERn and bERm are
constant column vectors, and A is an m x n constant matrix. Throughout the
chapter we impose the following assumptions on these problems:
(a) The set Sf = {x E S : x> O} of the strictly positive feasible solutions of
the primal problem (P) is nonempty.
N. Megiddo (ed.), Progress in Mathematical Programming
© Springer-Verlag New York Inc. 1989