An Integer Interior Point Method for Min-Cost Flow Using Arc Contractions and Deletions
This work addresses the numerical stability problem in interior point methods for min-cost flow, offering a theoretically efficient algorithm that avoids floating-point errors, which is important for practitioners who prefer combinatorial algorithms.
The paper presents an interior point method for the min-cost flow problem that uses arc contractions and deletions, achieving an expected runtime of O~(m^{3/2}) while using integer arithmetic to avoid numerical issues.
We present an interior point method for the min-cost flow problem that uses arc contractions and deletions to steer clear from the boundary of the polytope when path-following methods come too close. We obtain a randomized algorithm running in expected $\tilde O( m^{3/2} )$ time that only visits integer lattice points in the vicinity of the central path of the polytope. This enables us to use integer arithmetic like classical combinatorial algorithms typically do. We provide explicit bounds on the size of the numbers that appear during all computations. By presenting an integer arithmetic interior point algorithm we avoid the tediousness of floating point error analysis and achieve a method that is guaranteed to be free of any numerical issues. We thereby eliminate one of the drawbacks of numerical methods in contrast to combinatorial min-cost flow algorithms that still yield the most efficient implementations in practice, despite their inferior worst-case time complexity.