Hierarchical community detection via maximum entropy partitions and the renormalization group

Jorge Martinez Armas

Hierarchical community detection via maximum entropy partitions and the renormalization group

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A dendrogram-only entropy criterion selects informative community scales without edge statistics.

desk verdict Abstract promises a clean, size-only dendrogram-level criterion for choosing informative scales; the idea is worth a look, but the full text is unreadable in this pipeline and the reported evidence boils down to an unquantified claim of ground-truth agreement, with real circularity risk. read the letter →

arxiv 2508.04034 v1 pith:3TAIE62Q submitted 2025-08-06 cs.SI physics.data-anphysics.soc-ph

Jorge Martinez Armas This is my paper · ORCID

classification cs.SIphysics.data-anphysics.soc-ph

keywords hierarchicalcommunitydetectionmaximumentropydendrogramrenormalizationgroupmultiscalenetworksLFRbenchmarkmodularitynetworkscience

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces Hierarchical Clustering Entropy (HCE), a model-agnostic rule for choosing which level of a hierarchical community-detection tree actually carries meaning. Instead of looking at edges, HCE scores each level by a trade-off between the entropy of the community-size distribution and the number of communities, and selects the level where that trade-off is maximized. The paper claims this identifies scales of high structural heterogeneity in any dendrogram, regardless of how the dendrogram was produced. A sympathetic reader should care because this turns scale selection in community detection into a single, parameter-free criterion that could apply broadly across networks and other hierarchical data.

What carries the argument

Hierarchical Clustering Entropy (HCE) is a dendrogram-level score: at each level of a tree of nested partitions, it combines the entropy of the community-size distribution with the number of communities, and the selected scale is the level maximizing that combined objective. The renormalization-group framing in the title ties the chosen levels to successive coarse-grainings of the network, but the operative mechanism is the HCE score applied to whatever dendrogram the user already has.

What would settle it

Construct an LFR-style network with a planted two-level hierarchy, record HCE's selected level, then rewire edges so the planted top-level communities are equal-sized while keeping connectivity otherwise unchanged; if HCE stops selecting the planted top level, the criterion is tracking size heterogeneity rather than the underlying structure.

Watch

Extended reading notes

Core claim

Hierarchical Clustering Entropy is a criterion that works directly on the dendrogram, not on edge-level statistics. For each candidate partition level, it forms a score from two ingredients: how spread out the community sizes are, measured by entropy, and how many communities exist. The informative resolution is claimed to be the level where these two terms balance in a way that corresponds to high structural heterogeneity. The paper demonstrates on synthetic benchmarks, including LFR and symmetric and asymmetric multiscale models, that HCE's chosen partitions closely match planted ground truth across varying hierarchy, size imbalance, and noise. On real social and neuroscience networks, the

Load-bearing premise

The argument stands or falls on the premise that the most heterogeneous distribution of community sizes, rather than any edge-level statistic, is what makes a resolution level informative.

Editorial extensions

If this is right

Any existing hierarchical clustering method can be paired with HCE without modifying the underlying algorithm; the criterion only reads the dendrogram.
Because HCE avoids edge statistics, it applies equally to modularity-based and correlation-based dendrograms, and to any distance metric that produces a hierarchy.
HCE returns multiple informative levels, not just one partition, so it yields a hierarchy of meaningful scales rather than a single cut.
On LFR and multiscale benchmarks, the selected levels align with ground truth under varying hierarchy depth, community-size imbalance, and noise.
Applied to social and brain networks, the selected levels give interpretable modular hierarchies matching known structural and functional organization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because HCE consumes only a dendrogram, it should transfer to non-network hierarchical clusterings, such as gene-expression or document trees, where edge statistics are meaningless; the paper does not demonstrate this transfer.
If community-size heterogeneity is the signal, HCE may systematically discount scales with balanced but genuinely distinct communities; a targeted benchmark with planted equal-size communities would test whether the criterion tracks structure or merely size skew.
The entropy-versus-count objective resembles an information-theoretic model-selection score, so one could build a null distribution of HCE values from randomized dendrograms and attach significance to selected levels; the paper does not provide such a statistical test.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

I could only read the abstract; the full-text rendering in our pipeline is garbage, so everything below the abstract is an inference. That alone forces a provisional verdict.

What the paper does, as stated, is neat: it defines Hierarchical Clustering Entropy (HCE), a criterion that scores each cut of a dendrogram using only the multiset of community sizes. The level you pick maximizes a trade-off between size entropy and community count. That is genuinely different from the usual edge-statistic or modularity-based level selection, and if it works, it is a fast, model-agnostic tool for anyone using hierarchical clustering on networks. The abstract is well written and the framing is honest about the method's generality.

The soft spot is the one the stress-test flags, and I think it is real. The criterion ignores everything except the sizes of the communities at a given cut: no edge weights, no merge heights, no node attributes, no internal structure. So HCE is blind to whether a partition aligns with true modules; it only sees how uneven the sizes are. A high entropy of sizes could come from a spurious cut in noise, and a balanced, meaningful subdivision could be penalized simply for being balanced. The abstract gives no first-principles argument for why maximal size heterogeneity should track ground truth. Then there is the benchmark circularity: the synthetic models deliberately vary size imbalance, which is exactly what HCE rewards. If the planted communities have size distributions that look like the ones HCE likes, agreement might be baked in. The abstract reports no NMI/ARI numbers, no baselines, no error bars, so I cannot tell whether the agreement is impressive or trivial.

Now, fair credit: the idea is not crazy. Entropy-based selection has precedents, but working directly on dendrogram size distributions without edge information is a real angle, and the claimed scalability and domain independence are nice if substantiated. The paper is not incoherent; the concern is under-justification, not nonsense. But with only the abstract in hand, the central claim is unverified and the load-bearing assumption is unexamined.

Who is this for? People building practical community-detection pipelines who want an automatic, parameter-light way to pick a level. That is a useful audience. Whether the paper deserves a referee depends on the full text: if it includes the derivation, the benchmark protocol, and at least one null model showing HCE rejects noise, then yes, it is a legitimate submission. Based on the abstract alone, I would not desk-reject it; the idea is specific enough to warrant a careful reviewer. I just would not cite it or put it in our reading group until I can see the actual numbers.

Referee Report

3 major / 4 minor

Summary. The paper proposes Hierarchical Clustering Entropy (HCE), a criterion for choosing informative resolution levels from any hierarchical clustering dendrogram. HCE scores each cut by a trade-off between the entropy of the community-size distribution and the number of communities, and the authors claim it is model-agnostic, scalable, and selects partitions closely aligned with ground truth on LFR and multiscale synthetic benchmarks, while yielding interpretable hierarchies on real social and neuroscience networks.

Significance. If fully substantiated, HCE would be a useful and computationally light addition to the hierarchical community-detection toolbox, since it operates directly on dendrograms and does not require edge-level statistics. The core idea is appealing and falsifiable, and the paper's emphasis on a size-distribution criterion is novel relative to modularity- or likelihood-based approaches. However, the submission as available does not establish the central premise or empirical alignment: no quantitative metrics, baselines, or statistical tests are reported, and the abstract alone cannot support the strength of the claim.

major comments (3)

[Abstract, empirical evaluation] The central claim of 'consistently identifies partitions closely aligned with ground truth' is unsupported by any reported quantity. The abstract gives no NMI/ARI values, no error bars, no baseline comparisons, and no significance testing. This makes the primary empirical assertion unfalsifiable as written and prevents the reader from judging the method's practical value. At minimum, the paper must report standard agreement metrics and compare against a simple size-based baseline.
[Abstract, criterion rationale] The load-bearing premise is that the cut maximizing a trade-off between community-size entropy and community count marks the informative scale. The abstract asserts this without an independent argument. Because the criterion ignores edge weights, merge heights, and node labels, it cannot distinguish a partition aligned with true modules from any other partition with the same size multiset. A derivation, consistency argument, or a null-model test showing why this specific objective identifies signal rather than artifact is needed.
[Synthetic benchmark design] The abstract states that the benchmarks vary 'size imbalance' and include both symmetric and asymmetric multiscale models. If the planted ground-truth partitions are themselves size-heterogeneous, then a criterion that rewards size heterogeneity may succeed in part by construction. The paper must demonstrate that HCE does not also select spurious levels on balanced ground-truth communities or on null models with no community structure, e.g., by reporting performance as a function of planted size imbalance.

minor comments (4)

[Abstract, terminology] The phrase 'scales of high structural heterogeneity' is used without a precise definition. This should be formalized or replaced with a quantitative statement.
[Abstract, parameters] The trade-off between entropy and community count is not specified as fixed, tuned, or free. State whether a single default weight is used across all experiments and how sensitive the results are to that choice.
[Full text accessibility] The supplied manuscript text is heavily corrupted, with most characters replaced by placeholders. If this reflects the submitted PDF, a clean and readable version is essential; equations, figures, and benchmark tables cannot currently be verified.
[References and positioning] The abstract does not cite prior work on entropy-based clustering or the LFR benchmark. The body should clearly state the incremental contribution relative to existing scale-selection methods such as modularity-based or persistence-based approaches.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrable circularity; HCE's entropy objective is a modeling assumption rather than a fitted or self-referential prediction.

full rationale

The readable portion of the manuscript (the abstract) defines HCE as a fixed criterion: select resolution levels that maximize a trade-off between the Shannon entropy of the community-size distribution and the number of communities. It then evaluates the selected partitions against independent ground truth on synthetic benchmarks and real networks. No equation is visible that makes the selected partition equivalent to the criterion's own inputs; no parameter is fitted to a portion of the data and then reported as a prediction; no self-citation or imported uniqueness theorem is used as load-bearing support. The assumption that informative scales correspond to high size heterogeneity could be empirically wrong or biased, but that is a correctness/validity concern, not circularity. Without quotable evidence that the trade-off weight was tuned on the same benchmarks or that ground-truth partitions were constructed to maximize the HCE objective, the benchmark agreement remains an external empirical check rather than a construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

Review based on abstract only. The method introduces one implied free parameter (the trade-off weight) and one postulate (heterogeneity equals informativeness). No physical or data entity is invented; HCE is a new selection rule, not a new object in the network. The full text could reveal additional fitted constants or ad hoc choices that this ledger cannot capture.

free parameters (1)

trade-off weight between entropy and community count
The abstract defines the objective as a trade-off between entropy of community sizes and number of communities, but does not state the functional form or the relative weight. This is a free choice of the method that the benchmarks would be used to justify.

assumptions (3)

ad hoc to paper The informative scale in a hierarchy is the one maximizing the entropy-versus-count trade-off
The criterion is introduced as a postulate in the abstract, with no independent theoretical justification (such as an information-theoretic derivation) visible at abstract level. It is the core assumption the method rests on.
domain assumption Dendrograms produced by modularity- and correlation-based methods contain a level whose communities correspond to ground truth
Benchmark evaluation presupposes that the underlying clustering algorithms produce a dendrogram that contains the true partition at some level. This is standard practice in community detection but is an empirical assumption inherited from the clustering methods.
standard math Shannon entropy is the correct measure of community-size heterogeneity
The method uses the standard entropy definition. Its use as the heterogeneity measure is conventional, though entropy encodes a specific notion of uniformity that may not match all notions of structural meaningfulness.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hierarchical community detection via maximum entropy partitions and the renormalization group." pith.science (2026). https://pith.science/paper/3TAIE62Q

@misc{pith2026250804034,
  author       = {Pith},
  title        = {Pith review of: Hierarchical community detection via maximum entropy partitions and the renormalization group},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TAIE62Q}},
  note         = {Machine review of arXiv:2508.04034}
}

read the original abstract

Identifying meaningful structure across multiple scales remains a central challenge in network science. We introduce Hierarchical Clustering Entropy (HCE), a general and model-agnostic framework for detecting informative levels in hierarchical community structures. Unlike existing approaches, HCE operates directly on dendrograms without relying on edge-level statistics. It selects resolution levels that maximize a principled trade-off between the entropy of the community size distribution and the number of communities, corresponding to scales of high structural heterogeneity. This criterion applies to dendrograms produced by a wide range of clustering algorithms and distance metrics, including modularity-based and correlation-based methods. We evaluate HCE on synthetic benchmarks with varying degrees of hierarchy, size imbalance, and noise, including LFR and both symmetric and asymmetric multiscale models, and show that it consistently identifies partitions closely aligned with ground truth. Applied to real-world networks in social and neuroscience systems, HCE reveals interpretable modular hierarchies that align with known structural and functional organizations. As a scalable and principled method, HCE offers a general, domain-independent approach to hierarchical community detection with potential applications across biological, social, and technological systems.

Discussion (0). Continue with ORCID to comment.

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.