Adaptive Fully Dynamic $k$-Center Clustering with (Near-)Optimal Worst-Case Guarantees

Abstract: Given a sequence of adversarial point insertions and point deletions, is it possible to simultaneously optimize the approximation ratio, update time, and recourse for a $k$-clustering problem? If so, can this be achieved with worst-case guarantees against an adaptive adversary? These questions have garnered significant attention in recent years. Prior works by Bhattacharya, Costa, Garg, Lattanzi, and Parotsidis [FOCS '24] and by Bhattacharya, Costa, and Farokhnejad [STOC '25] have taken significant steps toward this direction for the $k$-median clustering problem and its generalization, the $(k, z)$-clustering problem. In this paper, we study the $k$-center clustering problem, which is one of the most classical and well-studied $k$-clustering problems. Recently, Bhattacharya, Costa, Farokhnejad, Lattanzi, and Parotsidis [ICML '25] provided an affirmative answer to the first question for the $k$-center clustering problem. However, their work did not resolve the second question, as their result provides only expected amortized guarantees against an oblivious adversary. In this work, we make significant progress and close the gap by answering both questions in the affirmative. Specifically, we show that the fully dynamic $k$-center clustering problem admits a constant-factor approximation, near-optimal worst-case update time, and constant worst-case recourse, even against an adaptive adversary. This is achieved by first developing a fully dynamic bicriteria approximation algorithm with (near-)optimal worst-case bounds, and then designing a suitable fully dynamic $k$-center algorithm with near-linear update time. For the fully dynamic bicriteria approximation algorithm, we establish the worst-case recourse and worst-case update time guarantees separately, and then merge them into a single algorithm through a simple yet elegant process.

• The paper presents a novel dynamic algorithm that achieves an O(1) approximation for k-center clustering with worst-case guarantees in adversarial settings.
• It utilizes lazy bicriteria maintenance and efficient data structures to ensure near-linear update time and constant recourse during insertions and deletions.
• The approach is supported by formal proofs and lower bounds, highlighting its optimality and necessity of randomization over deterministic methods.

Adaptive Fully Dynamic $k$-Center Clustering: Near-Optimal Worst-Case Guarantees

Introduction and Context

This work addresses the central question of whether it is possible to maintain a $k$-center clustering solution dynamically with provable, worst-case guarantees on approximation, update time, and solution recourse, even against a strong adaptive adversary. The $k$-center problem, a canonical hard optimization over metric spaces, asks for $k$ representative points (centers) that minimize the maximum distance from any point to its closest center. While static and amortized versions are extensively studied, maintaining high-quality solutions with worst-case guarantees in fully dynamic settings (insertions and deletions) with low recourse (the number of changes to the solution after each update) remains a challenging open problem, particularly under adversarial adaptivity.

The authors definitively answer this challenge with new algorithmic frameworks that, for the first time, provide fully dynamic $k$-center clustering with all of the following:

• $O(1)$ approximation ratio,
• $\tilde{O}(k)$ worst-case update time,
• $O(1)$ worst-case recourse,
• and resilience against a fully adaptive adversary.

This is achieved via a combination of novel bicriteria sparsification, careful lazy updating strategies, and the introduction of new near-linear time dynamic cluster maintenance.

Problem Setting and Adversarial Model

The goal is to maintain an approximate $k$-center clustering for a point set $P \subseteq \mathcal{X}$ under adversarial updates (insert/delete), with no restriction on the update sequence or knowledge of future updates. The adversary is adaptive, observing the maintained solution (and in the strong model, internal randomness). The desiderata are:

• Accuracy: Constant-factor approximation to the optimum $k$0-center cost.
• Efficiency: Polylogarithmic (in $k$1 and aspect ratio) worst-case update time, preferably near-linear in $k$2.
• Stability: Constant worst-case recourse. These collectively ensure practical and theoretical robustness for evolving large-scale datasets and real-time systems.

Technical Components

1. Fully Dynamic Bicriteria Approximation with Worst-Case Guarantees

The foundational step is constructing and maintaining a bicriteria approximation: instead of insisting on exactly $k$3 centers, the algorithm permits $k$4 centers, achieving better structural properties. The dynamic core exploits a generalization of the Mettu-Plaxton (MP-bi) static algorithm, which operates in "lazy" layers. These layers are only rebuilt upon substantial change (quantified precisely via thresholded counters) rather than every update, which is crucial to both efficiency and worst-case recourse.

A central innovation lies in making the maintenance of these layers lazier—each is permitted to deviate from the ground truth proportionally to its own size, and the bicriteria solution is only gradually synchronized to the target set across update phases. This is orchestrated so that (1) each update only affects $k$5 centers, and (2) the global solution is always within a constant factor of the current optimum, despite not tracking every update instantaneously.

The main technical lemma establishes that the lazy properties and the structure of the update sequence guarantee that the solution cost is at most 8 times the current optimal $k$6-center cost (with high probability), and the entire clustering adapts with $k$7 recourse per update (details are elaborated in the phase-based analysis of target set synchronization).

Figure 1: Representation of lazy set maintenance across epochs, highlighting the staggered execution set construction and lazy update propagation for adaptation under adversarial changes.

2. Near-Optimal Worst-Case Update Time

To ensure near-optimal worst-case update time, the authors design an independent framework where recoloring/rebuilding work for each layer is amortized and then de-amortized by spreading work uniformly over the allowed period of laziness. Each hierarchical layer (subproblem) is handled by an independent instance, and updates are scheduled in a round-robin/lazy fashion without ever having to process the entire dataset in a single update step.

Care is taken to ensure that layers are only rebuilt when their own laziness budget is depleted, with formal proofs that their optimality certificates (via carefully defined balls and coverage thresholds) remain approximately valid despite adversarial perturbations in the underlying pointset.

3. Efficient $k$8-Center Solution Extraction

The final component is the conversion of any lazily-maintained bicriteria solution into a true $k$9-center clustering with optimal recourse and near-linear time. Adapting and refining a recent low-recourse deterministic dynamic algorithm (Forster--Skarlatos), the authors remove its $k$0 time bottleneck via a refined reassignment cascade. Instead of exhaustively searching all inter-cluster assignments upon center deletions, an efficient $k$1 procedure is implemented using heap-based data structures and selective reassignment along canonical paths in an implicit assignment graph. This ensures each dynamic event is handled efficiently even when complex cascades are required by adversarial deletions or insertions.

Figure 2: Sequence of cluster responsibility reassignments after a center deletion: the algorithm maintains regular cluster invariants and minimal recourse via fast path selection.

Figure 3: Extended illustration showing step-by-step transitions between algorithmic states during adaptive adversarial insertions and deletions.

4. Impossibility Results for Deterministic Algorithms

A novel implication of the work is that, under established lower bounds, the guarantees of the presented (randomized) fully dynamic algorithms are optimal up to polylogarithmic factors. In particular, any deterministic fully dynamic constant-approximation $k$2-center algorithm with poly($k$3) update time would contradict known lower bounds. Thus, the randomization (even against adaptive adversaries) is not a technical artifact but a provable necessity at this regime.

Main Theorem and Guarantees

The main result proves the existence of a randomized, fully dynamic algorithm for $k$4-center clustering in arbitrary metric spaces such that, with high probability and against an adaptive adversary:

• The maintained solution is an $k$5-approximation to the optimum;
• Worst-case update time is $k$6;
• Worst-case recourse is $k$7.

This is the first such result in the literature for any non-trivial hard $k$8-clustering objective (including $k$9-median, $k$0-means, and generalizations), and it holds both for the usual metric setting and for graph metrics.

Practical and Theoretical Implications

The developed framework advances the state-of-the-art for consistency and adaptivity in dynamic clustering:

• Fully dynamic clustering under real-time constraints becomes practical: high-throughput, low-latency, and stability are achieved simultaneously.
• No compromise in solution quality for robustness to adaptive attacks, a pivotal requirement for use as a subroutine or in adversarial/online environments.
• The explicit lower bounds for deterministic dynamic clustering highlight that randomization is essential, indicating fundamental limits for current algorithmic design.

These algorithms are directly applicable in active databases, online learning and streaming settings, adversarial ML, and real-time recommender or summarization systems, especially where stable yet responsive cluster maintenance is required.

Future Directions

Several extensions are evident:

• The presented techniques generalize to $k$1-clustering objectives, and it would be compelling to realize full worst-case guarantees for other costs (e.g., $k$2-median, $k$3-means).
• Tightening dependence on $k$4, $k$5, and the metric aspect ratio remains of interest.
• The practical implementation of the efficient reassignment and hierarchical lazy maintenance strategies deserves empirical study in large-scale, high-throughput environments.
• Understanding fine-grained tradeoffs between recourse, update time, and approximation in related dynamic combinatorial optimization problems may further extend the algorithmic toolkit for robust adaptive data analysis.

Conclusion

This paper establishes, for the first time, that robust, efficient, and fully dynamic $k$6-center clustering with worst-case guarantees is possible against the strongest adversaries. The combination of lazy bicriteria maintenance, efficient sparsification and conversion, and formal lower bounds draws new boundaries for what is algorithmically attainable under adaptivity, and opens avenues for robust, scalable unsupervised learning and combinatorial optimization in dynamic, hostile environments.

Reference: "Adaptive Fully Dynamic $k$7-Center Clustering with (Near-)Optimal Worst-Case Guarantees" (2604.01726)