A Fast Approximation Scheme for Low-Dimensional $k$-Means

TL;DR

This paper presents a faster approximation algorithm for low-dimensional $k$-means, achieving near state-of-the-art practical running times while maintaining theoretical guarantees, by innovatively combining local search with randomized dissections.

Contribution

It introduces a $(1+ ext{epsilon})$-approximation algorithm for low-dimensional $k$-means with improved running time matching practical heuristics, using randomized dissections to speed up local search.

Findings

Achieves $(1+\epsilon)$-approximation in near-linear time for low-dimensional $k$-means.

Uses randomized dissections to efficiently identify optimal local moves.

Matches the practical running time of $k$-means++ and $D^2$-sampling heuristics.

Abstract

We consider the popular $k$-means problem in $d$-dimensional Euclidean space. Recently Friggstad, Rezapour, Salavatipour [FOCS'16] and Cohen-Addad, Klein, Mathieu [FOCS'16] showed that the standard local search algorithm yields a $(1+\epsilon)$-approximation in time $(n \cdot k)^{1/\epsilon^{O(d)}}$, giving the first polynomial-time approximation scheme for the problem in low-dimensional Euclidean space. While local search achieves optimal approximation guarantees, it is not competitive with the state-of-the-art heuristics such as the famous $k$-means++ and $D^2$-sampling algorithms. In this paper, we aim at bridging the gap between theory and practice by giving a $(1+\epsilon)$-approximation algorithm for low-dimensional $k$-means running in time $n \cdot k \cdot (\log n)^{(d\epsilon^{-1})^{O(d)}}$, and so matching the running time of the $k$-means++ and $D^2$-sampling heuristics up to polylogarithmic factors. We speed-up the local search approach by making a non-standard use of randomized dissections that allows to find the best local move efficiently using a quite simple dynamic program. We hope that our techniques could help design better local search heuristics for geometric problems. We note that the doubly exponential dependency on $d$ is necessary as $k$-means is APX-hard in dimension $d = \omega(\log n)$.