LP-rounding algorithms for facility-location problems

We study LP-rounding approximation algorithms for metric uncapacitated facility-location problems. We first give a new analysis for the algorithm of Chudak and Shmoys, which differs from the analysis of Byrka and Aardal in that now we do not need any bound based on the solution to the dual LP program. Besides obtaining the optimal bifactor approximation as do Byrka and Aardal, we can now also show that the algorithm with scaling parameter equaling 1.58 is, in fact, an 1.58-approximation algorithm. More importantly, we suggest an approach based on additional randomization and analyses such as ours, which could achieve or approach the conjectured optimal 1.46 · · ·–approximation for this basic problem. Next, using essentially the same techniques, we obtain improved approximation algorithms in the 2-stage stochastic variant of the problem, where we must open a subset of facilities having only stochastic information about the future demand from the clients. For this problem we obtain a 2.2975approximation algorithm in the standard setting, and a 2.4957-approximation in the more restricted, per-scenario setting. We then study robust fault-tolerant facility location, introduced by Chechik and Peleg: solutions here are designed to provide low connection cost in case of failure of up to k facilities. Chechik and Peleg gave a 6.5-approximation algorithm for k = 1 and a (7.5k + 1.5)-approximation algorithm for general k. We improve this to an LP-rounding (k+ 5+ 4/k)-approximation algorithm. We also observe that in case of oblivious failures the expected approximation ratio can be reduced to k + 1.5, and that the integrality gap of the natural LP-relaxation of the problem is at least k + 1. Institute of Computer Science, University of Wroclaw, Poland. [email protected]. Work of this author was partially conducted at CWI Amsterdam, TU Eindhoven, EPFL Lausanne, and while visiting the University of Maryland. Partially supported by FNP HOMING PLUS/2010-1/3 grant and MNiSW grant number N N206 368839, 2010-2013. Dept. of Computer Science, University of Maryland, College Park, MD 20742, USA. [email protected] Dept. of Computer Science and Institute for Advanced Computer Studies, University of Maryland, College Park, MD 20742, USA. [email protected]. Supported by NSF Award CNS-0626636.