27
7

Distributed Maximal Matching and Maximal Independent Set on Hypergraphs

Abstract

We investigate the distributed complexity of maximal matching and maximal independent set (MIS) in hypergraphs in the LOCAL model. A maximal matching of a hypergraph H=(VH,EH)H=(V_H,E_H) is a maximal disjoint set MEHM\subseteq E_H of hyperedges and an MIS SVHS\subseteq V_H is a maximal set of nodes such that no hyperedge is fully contained in SS. Both problems can be solved by a simple sequential greedy algorithm, which can be implemented naively in O(Δr+logn)O(\Delta r + \log^* n) rounds, where Δ\Delta is the maximum degree, rr is the rank, and nn is the number of nodes. We show that for maximal matching, this naive algorithm is optimal in the following sense. Any deterministic algorithm for solving the problem requires Ω(min{Δr,logΔrn})\Omega(\min\{\Delta r, \log_{\Delta r} n\}) rounds, and any randomized one requires Ω(min{Δr,logΔrlogn})\Omega(\min\{\Delta r, \log_{\Delta r} \log n\}) rounds. Hence, for any algorithm with a complexity of the form O(f(Δ,r)+g(n))O(f(\Delta, r) + g(n)), we have f(Δ,r)Ω(Δr)f(\Delta, r) \in \Omega(\Delta r) if g(n)g(n) is not too large, and in particular if g(n)=logng(n) = \log^* n (which is the optimal asymptotic dependency on nn due to Linial's lower bound [FOCS'87]). Our lower bound proof is based on the round elimination framework, and its structure is inspired by a new round elimination fixed point that we give for the Δ\Delta-vertex coloring problem in hypergraphs. For the MIS problem on hypergraphs, we show that for Δr\Delta\ll r, there are significant improvements over the naive O(Δr+logn)O(\Delta r + \log^* n)-round algorithm. We give two deterministic algorithms for the problem. We show that a hypergraph MIS can be computed in O(Δ2logr+Δlogrlogr+logn)O(\Delta^2\cdot\log r + \Delta\cdot\log r\cdot \log^* r + \log^* n) rounds. We further show that at the cost of a worse dependency on Δ\Delta, the dependency on rr can be removed almost entirely, by giving an algorithm with complexity ΔO(Δ)logr+O(logn)\Delta^{O(\Delta)}\cdot\log^* r + O(\log^* n).

View on arXiv
Comments on this paper