A Fixed Parameter Sublinear Time Approximation Scheme for Metric Min-Sum Clustering

Abstract

Metric $k$-minsum clustering is a simple, natural, appealing density-based clustering criterion whose computational complexity remains poorly understood, despite the progress achieved fornother clustering criteria in the past three decades. In this paper we give an efficient approximation scheme for this problem that runs in sublinear time for mildly increasing $k$, and thus is fixed parameter tractable with respect to the parameter $k$. We complement this result by showing that similar algorithms are unlikely to exist for the closely related problems $k$-median and balanced $k$-median, assuming popular fine-grained complexity conjectures. Specifically, we show that no sublinear time algorithm with small approximation factors can exist even for $2$-median and balanced $2$-median. It was previously known that no efficient PTAS can exist for $k$-median and our results show a strong and perhaps surprising separation between the $k$-minsum clustering problem and $k$-median even in the sublinear regime. Our upper bounds improve qualitatively upon decades old results. Our lower bounds establish the first use of the Distributed PCP framework to clustering problems.