Inapproximability of shortest paths on perfect matching polytopes

TL;DR

This paper proves that finding short paths in the perfect matching polytope's skeleton is computationally hard, disproving a previous conjecture and showing implications for simplex algorithm pivot rules.

Contribution

It establishes the inapproximability of shortest paths in perfect matching polytopes, disproving a conjecture and linking complexity to simplex pivot rule performance.

Findings

No polynomial-time algorithm for constant-length paths unless P=NP.

Under ETH, no algorithm can find paths shorter than a logarithmic bound.

Results hold even for bipartite graphs with maximum degree three.

Abstract

We consider the computational problem of finding short paths in the skeleton of the perfect matching polytope of a bipartite graph. We prove that unless $P=NP$, there is no polynomial-time algorithm that computes a path of constant length between two vertices at distance two of the perfect matching polytope of a bipartite graph. Conditioned on $P\neq NP$, this disproves a conjecture by Ito, Kakimura, Kamiyama, Kobayashi and Okamoto [SIAM Journal on Discrete Mathematics, 36(2), pp. 1102-1123 (2022)]. Assuming the Exponential Time Hypothesis we prove the stronger result that there exists no polynomial-time algorithm computing a path of length at most $\left(\frac{1}{4}-o(1)\right)\frac{\log N}{\log \log N}$ between two vertices at distance two of the perfect matching polytope of an $N$-vertex bipartite graph. These results remain true if the bipartite graph is restricted to be of maximum degree three. The above has the following interesting implication for the performance of pivot rules for the simplex algorithm on simply-structured combinatorial polytopes: If $P\neq NP$, then for every simplex pivot rule executable in polynomial time and every constant $k \in \mathbb{N}$ there exists a linear program on a perfect matching polytope and a starting vertex of the polytope such that the optimal solution can be reached in two monotone steps from the starting vertex, yet the pivot rule will require at least $k$ steps to reach the optimal solution. This result remains true in the more general setting of pivot rules for so-called circuit-augmentation algorithms.