Efficient Matroid Intersection via a Batch-Update Auction Algorithm

TL;DR

This paper introduces a simple auction-based algorithm for matroid intersection that achieves near-linear time approximation and sublinear parallel algorithms, significantly improving efficiency over previous methods.

Contribution

It presents a novel auction algorithm reducing approximate matroid intersection to a simpler problem, leading to faster algorithms in both sequential and parallel models.

Findings

First near-linear time approximation algorithm for matroid intersection.

First sublinear parallel algorithms for weighted matroid intersection.

Improved round complexity over previous algorithms.

Abstract

Given two matroids $\mathcal{M}_1$ and $\mathcal{M}_2$ over the same $n$-element ground set, the matroid intersection problem is to find a largest common independent set, whose size we denote by $r$. We present a simple and generic auction algorithm that reduces $(1-\varepsilon)$-approximate matroid intersection to roughly $1/\varepsilon^2$ rounds of the easier problem of finding a maximum-weight basis of a single matroid. Plugging in known primitives for this subproblem, we obtain both simpler and improved algorithms in two models of computation, including: * The first near-linear time/independence-query $(1-\varepsilon)$-approximation algorithm for matroid intersection. Our randomized algorithm uses $\tilde{O}(n/\varepsilon + r/\varepsilon^5)$ independence queries, improving upon the previous $\tilde{O}(n/\varepsilon + r\sqrt{r}/{\varepsilon^3})$ bound of Quanrud (2024). * The first sublinear exact parallel algorithms for weighted matroid intersection, using $O(n^{2/3})$ rounds of rank queries or $O(n^{5/6})$ rounds of independence queries. For the unweighted case, our results improve upon the previous $O(n^{3/4})$-round rank-query and $O(n^{7/8})$-round independence-query algorithms of Blikstad (2022).