Coresets for $k$-Means: When Does Uniform Sampling Suffice?

Amir Carmel; Robert Krauthgamer

Home/Archive/arXiv

arXiv · 2610.31491

Coresets for $k$-Means: When Does Uniform Sampling Suffice?

Amir Carmel·Robert Krauthgamer

View original source ↗Read source document ↗Export metadata ↓

Abstract

Data summarization techniques, such as coresets, are essential for scaling clustering algorithms to massive datasets. We study uniform sampling, which is an extremely simple and ultra-efficient method to construct coresets, in the context of weak, stable, and strong coresets for the $k$-means problem in Hilbert space. We begin with the fundamental case $k=1$, and establish matching upper and lower bounds on the sample complexity of stable and weak coresets, finding that the two notions require different sample complexities. As a consequence, we obtain tight bounds for stable $1$-median coresets in $\ell_1$ and $\ell_2$ (and thus also Kendall--tau and Jaccard metrics), removing the dimension dependence of Carmel and Krauthgamer (ICLR 2026). Our analysis is substantially simpler than prior work and reduces to bounding the deviation of the sample mean. For the general $k$-means problem on $ρ$-separated instances, a well-studied structural assumption on the input, we show that uniform sampling yields a stable coreset for all $O(1)$-approximate solutions, and a strong coreset if the instance additionally has bounded kurtosis. In all these results, the sample complexity is independent of the dimension and the instance size.

Explore related subjects

cs.DS cs.CG

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Amir Carmel, Robert Krauthgamer. 2026-10-22. Coresets for $k$-Means: When Does Uniform Sampling Suffice?. https://arxiv.org/abs/2610.31491

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Compressed Set Representations based on Set Difference

We introduce a compressed representation of sets of sets that exploits how similar they are to each other. Our representation supports access, membership, predecessor and successor queries on the sets within logarithmic time. It also computes the support of any element (i.e., the sets containing it) in optimal time. In addition, we give a new MST-based construction algorithm for the representation that outperforms standard ones.

cs.DS↗

arXiv2026

Fast Approximation Algorithms for Balanced Edge and Vertex Separators

We present a family of fast pseudo-approximation algorithms for the minimum balanced element cut problem in a graph, which includes vertex and edge separators. Given a graph $G=(V,E)$ with polynomially-bounded finite integer element weights with $n$ vertices and $m$ edges, and a (constant) balance parameter $c\in(0,1/2)$, where $G$ has some (unknown) $c$-balanced element cut of size $OPT_c$, we give a (Monte-Carlo randomized) algorithm running in $O(n^{\varepsilon}m^{1+o(1)})$ time that produces a $Θ(1)$-balanced element cut of size $O(OPT_c\cdot\sqrt{(\log n)/\varepsilon})$ for any value $\varepsilon\in[Θ(1/\log(n)),Θ(1)]$. Our algorithms are based on a semidefinite programming (SDP) relaxation of the problem. A standard approach to solving an SDP relaxation for this kind of problem would use the primal-dual Matrix Multiplicative Weights Update framework of Arora and Kale, but this framework does not seem to apply, since we cannot express all of our SDPs primal variables explicitly. Instead, we present a new MMWU framework for our problem that allows us to represent some of our variables implicitly. We also extend our framework to a general "Partial-variable MMWU framework", which we expect will be useful for future work. Our oracle for the balanced element cut problem within our framework uses $O(n^{\varepsilon}\text{polylog}(n))$ almost-linear time maximum-flow computations, and would be sped up if the time complexity of maximum-flow improves. For graphs with padding bound $β$, another oracle gives an $O(\sqrtβ/\varepsilon)$ approximation in $m^{1+o(1)}n^{\varepsilon}\operatorname{polylog}(n)$ time. It is known that $β\le O(\log h)$ for $K_h$-minor-free graphs (and in particular $β\le O(\log n)$).

cs.DS↗

arXiv2026

Optimal-Time Move Structure Construction

The move structure represents a permutation $π$ of $[0,n)$ by partitioning the domain into $O(r)$ disjoint, contiguously permuted intervals, with $r$ being the minimum number of such intervals. This data structure occupies $O(r)$ words of space and enables $O(1)$-time computation of $π(i)$ given the interval that contains $i$. For permutations where $r \ll n$, this provides an efficient, compressed representation for navigation. While the existing best $O(r)$-space construction approaches require $O(r\log r)$-time, we present an optimal $O(r)$-time and space construction algorithm. This is achieved by replacing balanced search trees with pointer-based lists through introducing a bidirectional strategy that synchronizes construction of the structures for $π$ and its inverse $π^{-1}$ in a single, unified pass. By applying this algorithm to an existing result, we achieve the first optimal $O(n)$-time construction of the longest common prefix (LCP) array from a run-length-encoded Burrows-Wheeler transform (RLBWT) of $r$ runs in $O(r)$ working space. Empirical evaluation on pangenome-scale data confirms that our move structure construction algorithm is consistently faster than the previous best, achieving speedups of up to $\sim 2\times$ while using less memory.

cs.DS↗