acceptodds
Under review as a conference paper at ICLR 2027

CLEARMIND: Closed-Form Laplacian Embeddings of Arakelov-Green Resistance for Morphological Intrinsic Neuronal Distances

Abstract

Tropical geometry turns a metric graph into a flat torus, its tropical Jacobian, and measures distances between the images of its points by a closest vector problem on the period lattice. We prove that for points of the graph this problem is solved by geodesics: the squared tropical polarization distance equals the path metric minus the Arakelov–Green distance, an effective-resistance kernel given in closed form by the Laplacian pseudoinverse. Equivalently, the Arakelov–Green distance and the path metric are the minimal energies of real and of integral unit flows, and the squared tropical distance is exactly their integrality gap. The full tropical distance matrix is thus computable in cubic time. On this identity we build CLEARMIND, a training-free descriptor of 3D neuronal morphology. A reconstruction is reduced to its branching skeleton, neurite tips near the soma are joined to it, short bridges are contracted, and the Arakelov–Green matrix of the result is summarized by its spectrum. Every step has an exact algebraic description; for instance, the period matrix records the shared soma-to-tip path lengths, and the spectrum has exactly one positive eigenvalue, so its absolute values determine it. The signature needs no lattice search, is invariant to vertex order, rigid motions and subdivision, and is Lipschitz in the edge lengths. On ACT-4, JML-4 and BIL-6, appending the 64-dimensional spectrum improves a point-cloud GNN and MorphVAE on every dataset, by up to 20.3 points; a Tree-LSTM with the spectrum outperforms every reimplemented baseline; and an MLP on the spectrum alone is a competitive morphology classifier that surpasses every reimplemented deep baseline on ACT-4. On BREC, a GIN with Arakelov–Green eigenvector encodings distinguishes 70.0% of the pairs beyond the 1-WL limit, including 33% of the CFI pairs, more than PPGN (23%) and I-GNN (21%), and the spectrum alone separates 217 of the 400 pairs without training.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neuronal morphology, tropical Jacobian, Arakelov-Green, metric graphs, graph representation learning, spectral descriptors, connectomics, Weisfeiler-Lehman

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hz8hOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLEARMIND, a learning-free morphological descriptor for neurons. It simplifies the neuronal tree structure and augments it with cycles to capture morphological information via the Arakelov-Green metric. CLEARMIND can be plugged into existing neuron classification models to improve their performance.

Critical strengths

CLEARMIND has a clear geometric definition and computational pipeline, exhibits good invariance properties, and does not rely on labeled data. These features make it easy to integrate with existing models and suitable for analyzing large-scale reconstructed neurons.

Critical weaknesses

The discussion of limitations in CLEARMIND is insufficient. Both Kron reduction and augmentation inevitably lose information and depend on two empirical hyperparameters. While simplifying the neuronal morphology is necessary, the paper lacks an analysis of how these steps and parameter choices affect the range of applicable neuron types. Such analysis is critical for assessing the performance boundaries of CLEARMIND.

Questions for authors
  1. For neurons with dense dendrites, could Kron reduction compress them into identical graphs, making them difficult to distinguish?

  2. Does the augmentation step artificially introduce cycles and potentially inject spurious topological information?

  3. What are the impacts of the two empirical hyperparameters discussed in Table 4 on the scope of applicability of CLEARMIND?

  4. Message-passing GNNs can enlarge their receptive fields by stacking more layers. In this setting, what new morphological information does CLEARMIND provide that deeper GNNs cannot capture?

  5. Table 2 reports results when AG is incorporated as an additional feature into existing models. How would the classification performance of AG alone compare directly with those baselines? This comparison would help evaluate the descriptive power of the proposed method.

  6. CLEARMIND lacks comparisons with traditional morphological descriptors such as Sholl analysis, the Topological Morphology Descriptor, and persistent homology.

Reviewer_eTYiOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies tropical polarization distances between points of a metric graph and proposes CLEARMIND, a training-free spectral descriptor for neuronal morphology. Its main theoretical claim is that the squared tropical distance equals the graph path distance minus the Arakelov–Green/effective-resistance distance. For neurons, the method reduces a reconstruction to a branching skeleton, adds edges from the soma to spatially nearby tips, contracts short bridges, and summarizes the resulting AG matrix by its spectrum. The paper reports neuron-classification results on ACT-4, JML-4, and BIL-6, plus BREC graph-pair discrimination.

Critical strengths
  1. Theorem 3.3 gives a closed-form relation between tropical polarization distance, path distance, and effective resistance, and interprets the result through integral versus real flow energies. If correct under the stated assumptions, this is a substantive mathematical contribution, not merely a new feature engineering recipe.

  2. CLEARMIND combines the graph metric with spatial information through cycle augmentation, and the spectrum improves the reported MorphoGNN and MorphVAE results on all three datasets. The paper also gives a transparent, training-free BREC result, while appropriately exposing that the spectrum alone does not separate every pair.

Critical weaknesses
  1. The morphology evaluation appears to use the evaluation folds for model selection. Appendix B.2 says folds 8–9 are evaluated after every epoch and the best epoch is reported. As written, there is no clearly untouched final test set, making the reported classification results optimistic as estimates of generalization. This affects the main empirical evidence for the descriptor’s utility.

  2. Some superiority and attribution claims exceed the comparisons shown. The paper does not report a Tree-LSTM without AG, so the strong Tree-LSTM + AG result does not isolate the spectrum’s benefit in that model. Several comparisons also use different protocols or previously reported baselines. On BREC, AG-GIN’s 70.0% overall is below I²-GNN’s 70.2%, although its CFI result is higher; it also reports the best of ten seeds per pair. These results support promising, task-specific performance, but not broad superiority over competing methods.

Questions for authors
  1. Are folds 8–9 used both for epoch selection and for the final reported evaluation? If so, can the author report results from a separate untouched test set, with model selection performed using only training/validation data?

  2. Can the author provide a Tree-LSTM baseline with the same architecture and training protocol but without AG features, and clarify which Table 2 baselines were rerun under the same splits and preprocessing? For BREC, can the author clarify how the “best of ten seeds per pair” result compares with the seed aggregation used for the cited baselines?

Reviewer_QRn3OpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
2(−1)
Presentation
1
Soundness
2
Read review
Paper summary

The paper describes a novel spectral descriptor for neuronal graphs, based on the usage of electrical equivalence, simplification through Kron reduction, and metric computation based on tropical geometry, and leading to a Laplacian eigenvalue decomposition. The theoretical construction is fascinating, and very well designed, and results on neuronal reconstructions showcase promising results in terms of expressiveness, that can support available graph learning paradigms. Authors compares with respect to last available baselines, even if the GraphFormer was partially compared.

Critical strengths
  1. Novel spectral signature for graphs with strong theoretical contribution. The application of tropical geometry and spectral graph theory appears novel to me.
  2. Estensive assessment and comparison with SOTA methods, highlighting the potential of the proposed descriptor, and with convincing and interesting results.
Critical weaknesses
  1. Manuscript presentation. I had hard time to try to understand the concepts, and I think many parts can be improved to ensure proper readability:
  • abstract too long, and with many concepts out of context, especially in the first part.
  • the whole theoretical construction can be greatly simplified with a toy example (a graph with few edges, to showcase how Kron reduction works, how metrics are computed, etc.)
  • explain in a concise style what are the connections between differential geometry, spectral graph theory, and electrical circuit theory in the proposed network (this needs to be concise but clear)
  • theoretical appendix is unusually long ( most part of it looks to be AI-generated) and various mathematical parts are repeated from previous papers.
  1. Analysis of the recovered spectral signature; it seems to me that experiments were performed with a signature SΓ∈R64S_\Gamma \in \mathcal{R}^{64}, I would have expected examples of histograms, and also visual representation of embeddings to showcase the potential for phenotyping; in spectral theory, the first eigenvalues have physical significance, what about the proposed one?
  2. Complexity and scalability. It seems to me that bottleneck is in the Kron reduction; how the whole computation is depending on the size of the spectral domain? 16,32,64, and 128, do they make sense?
  3. Assessment: given that the paper proposes a novel spectral signature, I would expect comparison with established descriptors, like Laplacian Spectrum, especially for what concerns the expressiveness (again, visualization of embeddings is important to assess that)
  4. For assessment of the graph signature alone, authors considered an MLP, Given that the signature is compact, what about using KNNs the same way reported baselines do?
Questions for authors
  1. showcase embeddings (with comparisons)
  2. improve the description of the theoretical formulation, highlighting the connections with spectral graph theory and with simplified toy examples
  3. extend the analysis of the capabilities of the proposed descriptor with visual examples (histogram of the spectrum), and perform KNN experiments for comparing with respect to baselines

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.