Coreset Strikes Back: Improved Parameterized Approximation Schemes for (Constrained) k-Median/Means
Sujoy Bhore, Ameet Gadekar, Tanmay Inamdar
TL;DR
The paper addresses constrained $(k,z)$-clustering in metric spaces with bounded algorithmic scatter dimension, proposing a unified EPAS framework that leverages coresets to reduce problem size and a ball-intersection-based refinement to enforce feasibility under capacities and fairness. The framework combines a coreset construction, a constrained assignment oracle, and a ball-intersection oracle to produce a $(1+\varepsilon)$-approximation with running time $2^{O\left(\frac{k}{\varepsilon}\log (k\ell) \cdot \λ\left(\frac{\varepsilon}{60z}\right) \log(1/\varepsilon)\right)}$ times poly$(|\mathcal{I}|)$, with deterministic variants when coresets are deterministic and improved exponent behavior for vanilla clustering. It yields the first EPAS for capacitated and fair $(k,z)$-clustering in broad metric families, including those with bounded highway dimension, and extends prior Euclidean results by Cohen-Addad and Li and others. The approach improves upon previous EPAS for uncapacitated clustering by achieving tighter dependence on $k$ and $\varepsilon$ via a more refined radii bounding argument in the request-sets, enabled by the scatter-dimension framework. These results provide practical, scalable algorithms for constrained clustering in diverse metric spaces and connect coreset theory with parameterized approximation in new settings.
Abstract
Algorithmic scatter dimension is a notion of metric spaces introduced recently by Abbasi et al. (FOCS 2023), which unifies many well-known metric spaces, including continuous Euclidean space, bounded doubling space, planar and bounded treewidth metrics. Recently, Bourneuf and Pilipczuk (SODA 2025) showed that metrics induced by graphs from any fixed proper minor closed graph class have bounded scatter dimension. Abbasi et al. presented a unified approach to obtain EPASes (i.e., $(1+ε)$-approximations running in time FPT in $k$ and $ε$) for $k$-Clustering in metrics of bounded scatter dimension. However, a seemingly inherent limitation of their approach was that it could only handle clustering objectives where each point was assigned to the closest chosen center. They explicitly asked, if there exist EPASes for constrained $k$-Clustering in metrics of bounded scatter dimension. We present a unified framework which yields EPASes capacitated and fair $k$-Median/Means in metrics of bounded algorithmic scatter dimension. Our framework exploits coresets for such constrained clustering problems in a novel manner, and notably requires only coresets of size $(k\log n/ε)^{O(1)}$, which are usually constuctible even in general metrics. Note that due to existing lower bounds it is impossible to obtain such an EPAS for Capacitated $k$-Center, thus essentially answering the complete spectrum of the question. Our results on capacitated and fair $k$-Median/Means provide the first EPASes for these problems in broad families of metric spaces. Earlier such results were only known in continuous Euclidean spaces due to Cohen-Addad & Li, (ICALP 2019), and Bandyapadhyay, Fomin & Simonov, (ICALP 2021; JCSS 2024), respectively. Along the way, we obtain faster EPASes for uncapacitated $k$-Median/Means, improving upon the running time of the algorithm by Abbasi et al.