Persistent Clustering and a Theorem of J. Kleinberg

We construct a framework for studying clustering algorithms, which includes two key ideas: persistence and functoriality. The first encodes the idea that the output of a clustering scheme should carry a multiresolution structure, the second the idea that one should be able to compare the results of clustering algorithms as one varies the data set, for example by adding points or by applying functions to it. We show that within this framework, one can prove a theorem analogous to one of J. Kleinberg, in which one obtains an existence and uniqueness theorem instead of a non-existence result. We explore further properties of this unique scheme, stability and convergence are established.

💡 Research Summary

The paper proposes a novel theoretical framework for clustering algorithms built around two central ideas: persistence and functoriality. Persistence captures the intuition that a clustering output should not be a single static partition but a multi‑resolution structure—a filtration of partitions that reflects the data at various scales. Functoriality encodes the requirement that clustering results should behave consistently under transformations of the data set, such as adding or removing points, applying Lipschitz maps, or more general continuous functions. By formalizing these concepts within category theory, the authors are able to revisit Kleinberg’s celebrated impossibility theorem, which states that no clustering function can simultaneously satisfy consistency, scale‑invariance, and richness.

The authors first critique Kleinberg’s result, arguing that the impossibility stems from insisting on a single, scale‑agnostic partition. They then introduce a categorical setting where objects are finite metric spaces and morphisms are non‑expansive (1‑Lipschitz) maps. Within this category, a “filtered object” represents a hierarchy of partitions indexed by a resolution parameter ε. A clustering algorithm is modeled as a functor from the metric‑space category to the category of filtered objects. This functorial viewpoint forces the algorithm to respect data transformations in a mathematically precise way.

Within this framework the authors construct a unique clustering scheme that satisfies three axioms: (1) Persistence – the output is a filtration of partitions; (2) Functoriality – the scheme is a functor with respect to 1‑Lipschitz maps; and (3) Non‑triviality – there exist data sets for which the scheme yields more than one cluster. Remarkably, the unique scheme turns out to be the classic single‑linkage (or “minimum‑spacing”) clustering, but reinterpreted as a persistent filtration: for each ε the clusters are the connected components of the ε‑neighbourhood graph, and as ε grows clusters merge in a nested fashion.

The paper proves an existence‑and‑uniqueness theorem: under the three axioms there is exactly one clustering functor, and it is precisely the persistent single‑linkage construction. The authors then analyze several important properties of this scheme. First, stability: small perturbations of the underlying metric (e.g., adding bounded noise) induce only small changes in the filtration, measured by the bottleneck distance between persistence diagrams. Second, convergence: when the data are sampled increasingly densely from an underlying compact metric space, the empirical filtration converges (in the Gromov‑Hausdorff sense) to the true filtration of the underlying space. Third, computational tractability: because the construction coincides with single‑linkage, existing O(n²) algorithms or fast approximate MST‑based methods can be employed without additional overhead.

Beyond these technical results, the authors discuss broader implications. By embedding clustering in a functorial, persistent setting, the work bridges classical clustering theory with topological data analysis (TDA). The filtration viewpoint naturally accommodates multi‑scale analysis, making the approach suitable for data where scale matters (e.g., hierarchical community detection, image segmentation, or time‑varying networks). Moreover, functoriality provides a principled way to compare clusterings across different data sets or across successive snapshots of a dynamic data set, something that traditional static clustering lacks.

In conclusion, the paper shows that Kleinberg’s impossibility theorem is not a universal barrier but a consequence of restricting attention to static partitions. When clustering is reconceived as a persistent, functorial process, one obtains a mathematically well‑posed problem with a single, canonical solution. This solution inherits the desirable algorithmic properties of single‑linkage while gaining robustness, multi‑scale interpretability, and a solid categorical foundation, opening new avenues for both theoretical investigation and practical applications.