Pith. sign in

REVIEW 4 major objections 5 minor 65 references

Heat Kernel Goes Topological

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Heat-kernel signatures from a combinatorial-complex Laplacian distinguish every non-isomorphic pair.

desk verdict A plausible spectral descriptor for combinatorial complexes saddled with an ill-defined Laplacian and an expressivity theorem that does not hold. read the letter →

arxiv 2507.12380 v1 pith:D75WX27J submitted 2025-07-16 cs.LG

Maximilian Krahn , Vikas Garg This is my paper
classification cs.LG
keywords heatkernelsignaturescombinatorialcomplexestopologicaldeeplearninggraphLaplacianHodgespectralmethodsexpressivitymolecularpropertyprediction
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

This paper sets out to show that a Laplacian defined directly on combinatorial complexes gives heat-kernel descriptors that are both maximally expressive and inexpensive to use in deep learning. The authors define $L = \sum_{i=1}^{R} b_i \delta_i \delta_i^\top$ using incidence matrices between rank-0 cells and each higher rank, then use the diagonal of the matrix exponential $e^{-tL}$ at several diffusion times as a multiscale node signature. The payoff, if the theory is correct, is that no two non-isomorphic complexes produce the same descriptor, so a sufficiently expressive network can separate them by design, without running higher-order message passing. That combination of expressive power and speed is what would make the method useful for molecular and other structured-data benchmarks.

What carries the argument

The load-bearing object is the combinatorial-complex Laplacian of Definition 3.1: a weighted sum $L = \sum_{i=1}^{R} b_i \delta_i \delta_i^\top$ of Gram matrices of incidence matrices from rank-0 cells to rank-$i$ cells, with coefficients $b_i$ chosen so that all subset sums are distinct, for example powers of $1/2$. This single symmetric operator turns every higher-rank cell into a contribution to the connectivity of rank-0 cells, which is why one Laplacian can see rank-3 or rank-4 structure that the Hodge Laplacian misses. The heat kernel $K_t = \exp(-tL)$ and its diagonal entries at selected times $t_1,\dots,t_d$ define the HKS descriptor of each rank-0 cell. The argument's chain is: incidence structure fixes the Laplacian, the Laplacian fixes the spectrum, the spectrum fixes the diffusion kernel, distinct kernels give distinct descriptors, and a universal function approximator separates distinct descriptors.

What would settle it

Take a known pair of non-isomorphic graphs with identical Laplacian spectra (such pairs exist in the standard graph-theory literature) and compute the paper's HKS descriptors for them; since the proposed Laplacian reduces to the graph Laplacian on graphs, identical descriptors would contradict the uniqueness theorem, while distinct descriptors would support it.

Watch

Extended reading notes

Core claim

The central claim is that heat kernel signatures computed from the proposed combinatorial-complex Laplacian are complete: any two non-isomorphic combinatorial complexes have distinct descriptors, and a Universal Function Approximator can therefore learn to tell them apart. The proof route is Theorem 4.1, which asserts that the Laplacian is uniquely determined by the complex up to orthogonal similarity, and Theorem 4.3, which asserts that Laplacians not related by an orthogonal transformation give distinct diffusion kernels and hence distinct HKS descriptors. The paper also establishes supporting properties: on graphs the operator reduces exactly to the graph Laplacian, the quadratic form $f^\top L f$ is a Dirichlet-energy smoothness measure, and on combinatorial complexes the new Laplacian is strictly more expressive than the Hodge Laplacian, while the two agree on simplicial complexes. Empirically, the descriptors are reported to separate torus pairs that confound the SMCN and CIN baselines, run roughly twelve times faster than SMCN, and match or beat comparisons on MolHIV, PROTEIN, Glycose, and Immunogenicity.

Load-bearing premise

The argument's load-bearing premise is that the Laplacian built from a complex's incidence pattern is a complete fingerprint: any two complexes whose Laplacians are the same up to relabelling are already the same complex.

Editorial extensions

If this is right

  • If the uniqueness theorem holds, the method separates every pair of non-isomorphic combinatorial complexes, including the pairs that the paper shows are invisible to the Hodge Laplacian and to the WL-style message-passing baselines.
  • Because the descriptor is read off from one eigendecomposition instead of being built by message passing, the per-iteration cost stays roughly flat as the number of cells grows; the paper reports about 12x faster inference than SMCN on its torus scalability test.
  • On ordinary graphs the operator is exactly the graph Laplacian, so the framework hands graph neural networks a ready-made multiscale spectral feature channel without changing their architecture.
  • On the four benchmarks tested, the descriptors with a transformer or MLP-Mixer backbone match or beat the compared topological and graph baselines, so the claimed expressiveness does not appear to trade away benchmark accuracy.

Reading between the lines

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

  • A strict reading of the uniqueness proof assumes that orthogonally similar Laplacians force isomorphism; because the operator reduces to the graph Laplacian on graphs, standard non-isomorphic cospectral graph pairs are a direct stress test of that assumption, and the paper does not discuss them.
  • The theoretical separation guarantee is stated for the diffusion process, while the descriptor samples it at finitely many times; a natural extension is adaptive or learned selection of the time grid so the finite descriptor realizes the guarantee for every pair.
  • Since the descriptor is a per-node spectral feature, it could be combined with other inductive biases rather than replacing them; feeding HKS into existing message-passing or substructure-counting models is a cheap testable extension of the paper's idea.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes TopoHKS, a framework that defines a Laplacian on combinatorial complexes as a weighted sum of incidence Gram matrices, uses the diagonal of the resulting heat kernel at multiple diffusion times as node descriptors, and feeds these descriptors into a transformer or MLP-Mixer backbone. The authors claim that the proposed Laplacian is uniquely determined by a combinatorial complex, that spectrally equivalent Laplacians force isomorphism, and that the resulting HKS descriptors distinguish arbitrary non-isomorphic combinatorial complexes, which would make the method maximally expressive. Experiments on torus datasets, MolHIV, PROTEIN, Glycose, and Immunogenicity report competitive or superior accuracy against topological baselines, with substantially higher throughput. The appendix contains proofs of the main theoretical statements, a smoothness interpretation, and the claimed Hodge-Laplacian comparison.

Significance. If the theoretical claims were correct, the paper would make a strong contribution: a spectral descriptor that is a complete invariant for combinatorial complexes, combined with an efficient transformer-free architecture, would be of considerable interest to the topological deep learning community. The empirical observation that a precomputed spectral descriptor can replace higher-order message passing while retaining accuracy is also potentially valuable. However, the central theoretical machinery is unsound. The sign ambiguity in the incidence matrix definition makes the Laplacian ill-defined; the graph reduction is off by a constant factor; and the uniqueness/isomorphism theorem is directly contradicted by well-known cospectral graphs. Because these failures are load-bearing for the abstract's central claim of maximal expressiveness, the paper's main contribution is not established. The empirical efficiency results are positive but secondary, and the paper does not provide code or sufficient experimental detail to independently verify them.

major comments (4)
  1. [Definition 2.3 and Definition 3.1] Definition 2.3 does not determine the incidence matrix: for a cell y containing n rank-0 cells, the definition states that exactly one entry is +1 and the remaining entries are -1, but it does not specify which cell receives the +1. For a rank-1 cell y={a,b,c}, choosing +1 at a yields δδ^T with off-diagonal entry +1 between b and c, while choosing +1 at b yields a different matrix with off-diagonal entry +1 between a and c and -1 between b and c. These matrices are not equal, so the Laplacian in Definition 3.1, and hence the HKS descriptors in Definition 3.2, are not well-defined functions of the combinatorial complex. This ambiguity invalidates the uniqueness claimed in Theorem 4.1 and the expressiveness chain built on it.
  2. [Corollary 4.1 and Appendix Corollary 8.1] The claimed reduction to the graph Laplacian is false as stated. For a graph (a CC of rank 1), Definition 3.1 gives L = b_1 δ_1 δ_1^T, whereas the standard graph Laplacian is δ_1 δ_1^T. The coefficient b_1 is an element of the set B and is not required to be 1; for example, the paper suggests B={2^-1, 2^-2, ...}, which can give b_1=1/2. Thus the CC Laplacian is only proportional to the graph Laplacian in the rank-1 case, not identical. This is a load-bearing error because the paper uses the graph case both as a sanity check and as a bridge to the uniqueness theorem.
  3. [Theorem 4.1 and Appendix Theorem 8.1] Theorem 4.1 asserts that if L' = Π L Π^T for an orthogonal matrix Π, then the underlying combinatorial complexes are isomorphic. This is false even for graphs, which are CCs of rank 1. Non-isomorphic cospectral graphs have Laplacians that are orthogonally similar: if L = U Λ U^T and L' = V Λ V^T are eigendecompositions of cospectral Laplacians, then L' = Π L Π^T with Π = V U^T. Since non-isomorphic cospectral graphs exist, the claimed bijective isomorphic mapping cannot exist. The appendix proof only shows that all Gram factorizations of a given matrix are related by an orthogonal transformation; it does not establish that the transformed matrix corresponds to an isomorphic combinatorial complex. Consequently, Theorem 4.1 is false.
  4. [Theorem 4.3 and Corollary 4.3] The HKS uniqueness theorem and its expressiveness corollary do not follow. Step 1 of the proof of Theorem 4.3 relies on the false claim that non-isomorphic complexes have non-similar Laplacians, which is contradicted by cospectral graphs as noted above. Step 2, that non-similar Laplacians yield distinct HKS descriptors, is asserted without proof and is also false in a relevant regime: for any vertex-transitive graph, K_t(v,v) equals (1/n) tr(e^{-tL}) for every vertex v, so the HKS descriptor depends only on the Laplacian spectrum; two non-isomorphic vertex-transitive cospectral graphs therefore have identical HKS descriptors for all vertices and all times. Such pairs are well documented in the spectral graph theory literature. Corollary 4.3 then assumes distinct input descriptors and invokes universal approximation, but the unconditional claim that the method distinguishes arbitrary non-isomorphic CCs is exactly the part that fails.
minor comments (5)
  1. [Definition 3.1] The subset-sum condition on the set B is garbled: the phrase 'iff ... than B'=B''' is not a complete mathematical statement, and the intended distinctness condition should be written as ∀B',B''⊆B, if Σ_{b∈B'} b = Σ_{b∈B''} b then B'=B''.
  2. [Section 4.1 and Appendix Section 8.2] The numbering and cross-references are inconsistent: Lemma 4.1 is described as 'This corollary can be demonstrated', and the same figure appears as both Fig. 4 and Fig. 7; the reader cannot tell which result corresponds to which statement.
  3. [Theorem 4.2] The smoothness theorem cites 'Eq. (10)' and 'Eq. (11)' but the displayed equations in Section 4.1 are numbered (7) and (8); moreover, the claimed reformulation as Σ w_ij(f_i-f_j)^2 requires nonnegative off-diagonal weights, whereas Definition 3.1 produces positive off-diagonal entries for cells containing more than two rank-0 cells, so the expansion as written is not generally valid.
  4. [Section 5.1] The text alternates among 'rank 3', 'rank 4', and 'cell of rank 4, which covers two cells of rank 2' when describing the new torus dataset; these descriptions should be reconciled with the formal rank definitions.
  5. [Table 3 and Section 5.4] The dataset name 'Glycose' appears to be a typo, and the table is missing dataset statistics and standard deviations for some baselines; also, the paper states that code will be released only after acceptance, which currently prevents verification of the reported efficiency and accuracy numbers.

Circularity Check

1 steps flagged · score 6.0 of 10

Maximal expressiveness claim reduces to the assumption that HKS descriptors are distinctive.

  1. self definitional [Section 4.2, Corollary 4.3; Appendix Corollary 8.3 proof]
    "Corollary 4.3 (Expressiveness) states: "Given two combinatorial complexes with distinct input descriptors, it is possible to learn a function using a Universal Function Approximator (UFA) approach that effectively distinguishes between them." Appendix Corollary 8.3 proof: "Assume that their node-level input features (e.g., heat kernel signatures) are such that ... the inputs are distinctive up to isomorphism.""

    The paper's headline expressiveness claim is that the method distinguishes arbitrary non-isomorphic CCs. Corollary 4.3 takes as its premise 'distinct input descriptors' — precisely the property that must be proved for all non-isomorphic CCs — and then applies UFA. The Appendix proof makes the reduction explicit by assuming inputs are 'distinctive up to isomorphism.' A UFA separating already-distinct vectors is analytically true and contributes no content; it is a restatement of the antecedent. Thus the central claim that the method is maximally expressive is, by construction, equivalent to the assumption that the HKS descriptors are a complete invariant for CCs. That assumption is delegated to Theorem 4.1/8.1, which is asserted rather than derived.

full rationale

Score is 6 rather than 0 because the final step of the expressiveness derivation is circular: Corollary 4.3/8.3 reduces 'maximal expressiveness' to the assumption of distinctive input descriptors. There is no fitted-parameter circularity: the HKS descriptor is a deterministic function of the CC Laplacian, and the model is trained on those descriptors; no fitted value is renamed as a prediction. The self-citation to [5] (Garg is a co-author) is used only for dataset splits and as a baseline, so it is not load-bearing. Separately, the Laplacian-uniqueness theorem that would supply the missing premise is not established and is in fact doubtful: Definition 2.3 leaves the +1 sign in delta_i unspecified, so L is not a well-defined function of the CC, and graph Laplacians are known to have non-isomorphic cospectral pairs. These correctness failures make the circular step more damaging, but they are not themselves instances of circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central result depends on the Laplacian being a well-defined complete invariant. That depends on an unspecified sign convention in Definition 2.3 and on a false spectral-determination premise. The descriptor also depends on hand-chosen diffusion times and weights. No new physical entities are proposed.

free parameters (2)
  • Diffusion times t_1,...,t_d = d=10, equal spacing between 0 and 3
    Hand-chosen hyperparameters controlling the multiscale descriptor; no selection criterion or sensitivity analysis is given.
  • Laplacian weights b_i = powers of 1/2 (suggested)
    Weights in the weighted incidence Gram sum; required to have distinct subset sums but otherwise arbitrary, and the results likely depend on their values.
assumptions (4)
  • ad hoc to paper A fixed combinatorial complex determines a unique incidence matrix delta_i.
    Definition 2.3 and Definition 3.1 require a sign convention; the paper never specifies which rank-0 cell gets +1, so the Laplacian is not uniquely determined by the complex.
  • ad hoc to paper Spectrally equivalent Laplacians imply isomorphic combinatorial complexes.
    Theorem 4.1 and Corollary 4.2 rely on this; it is contradicted by non-isomorphic cospectral graphs.
  • standard math A symmetric positive-semidefinite Laplacian can be exponentiated via eigendecomposition.
    Used to define the heat kernel K_t = exp(-tL); this is standard linear algebra.
  • standard math A universal function approximator can realize any separation function on finite descriptor sets.
    Cited universal approximation results are used in Corollary 4.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Heat Kernel Goes Topological." pith.science (2026). https://pith.science/paper/D75WX27J

@misc{pith2026250712380,
  author       = {Pith},
  title        = {Pith review of: Heat Kernel Goes Topological},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D75WX27J}},
  note         = {Machine review of arXiv:2507.12380}
}
read the original abstract

Topological neural networks have emerged as powerful successors of graph neural networks. However, they typically involve higher-order message passing, which incurs significant computational expense. We circumvent this issue with a novel topological framework that introduces a Laplacian operator on combinatorial complexes (CCs), enabling efficient computation of heat kernels that serve as node descriptors. Our approach captures multiscale information and enables permutation-equivariant representations, allowing easy integration into modern transformer-based architectures. Theoretically, the proposed method is maximally expressive because it can distinguish arbitrary non-isomorphic CCs. Empirically, it significantly outperforms existing topological methods in terms of computational efficiency. Besides demonstrating competitive performance with the state-of-the-art descriptors on standard molecular datasets, it exhibits superior capability in distinguishing complex topological structures and avoiding blind spots on topological benchmarks. Overall, this work advances topological deep learning by providing expressive yet scalable representations, thereby opening up exciting avenues for molecular classification and property prediction tasks.

Figures

Figures reproduced from arXiv: 2507.12380 by the authors.

Figure 1
Figure 1. Qualitative example of how the Heat Kernel captures local and global information. Let 𝑡 denote the diffusion time and 𝐿 the Laplacian of the combinatorial complex. For ef￾ficient computation, we use the spectral decompo￾sition of 𝐿, given by 𝐿 = Φ⊤diag(𝜆1, . . . , 𝜆𝑛)Φ, which allows us to express the heat kernel as 𝐾𝑡 = Φ⊤diag(𝑒 −𝑡𝜆1 , . . . , 𝑒−𝑡𝜆𝑛 )Φ. Let 𝑐 be a rank-0 cell in the combinatorial complex. For notati… view at source ↗
Figure 2
Figure 2. Qualitative example of how the HKS descriptor differs for non-isometric cells of rank 0. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Showing a model setup for learning features. Input: Concatenation of cell [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Presenting two Combina￾torial Complexes with their CC and Hodge Laplacian. While the CC Laplacian differs, the Hodge Lapla￾cian is the same for both complexes. This corollary can be demonstrated by constructing a simple counterexample, as illustrated in [PITH_FULL_IMA…
Figure 5
Figure 5. Figure 5: In Fig. 5a we show that only our method can differentiate between tori [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Inference tim￾ing for differently sized combinatorial complexes. Mean over five runs We evaluate the scalability of our method against the Scalable-MCN (SMCN) model from [6], using a mod￾ified Torus dataset. Instead of focusing on classification accuracy, we concentrat…
Figure 7
Figure 7. Figure 7: The figure depicts two non-isomorphic combinatorial complexes. In both [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 42 canonical work pages

  1. [1]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolu- tional networks. InInternational Conference on Learning Representations, 2017

    work page 2017

  2. [2]

    Provably powerful graph networks.Advances in neural information processing systems, 32, 2019

    Haggai Maron, Heli Ben-Hamu, Hadar Serviansky, and Yaron Lipman. Provably powerful graph networks.Advances in neural information processing systems, 32, 2019. 15

    work page 2019

  3. [3]

    Demystifying higher-order graph neural networks

    Maciej Besta, Florian Scheidl, Lukas Gianinazzi, Grzegorz Kwasniewski, Shachar Klaiman, Jürgen Müller, and Torsten Hoefler. Demystifying higher-order graph neural networks. arXiv preprint arXiv:2406.12841, 2024

    arXiv 2024

  4. [5]

    Topological neural networks go persistent, equivariant, and continuous.International Conference on Machine Learning, 2024

    Yogesh Verma, Amauri H Souza, and Vikas Garg. Topological neural networks go persistent, equivariant, and continuous.International Conference on Machine Learning, 2024

    work page 2024

  5. [6]

    Bronstein, and Haggai Maron

    Yam Eitan, Yoav Gelberg, Guy Bar-Shalom, Fabrizio Frasca, Michael M. Bronstein, and Haggai Maron. Topological blindspots: Understanding and extending topological deep learning through the lens of expressivity. InThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  6. [7]

    Cell complex neural networks.arXiv preprint arXiv:2010.00743, 2020

    Mustafa Hajij, Kyle Istvan, and Ghada Zamzmi. Cell complex neural networks.arXiv preprint arXiv:2010.00743, 2020

    arXiv 2010

  7. [8]

    Convergence of invariant graph networks

    Chen Cai and Yusu Wang. Convergence of invariant graph networks. InInternational Conference on Machine Learning, pages 2457–2484. PMLR, 2022

    2022

  8. [9]

    Invariant and equivariantgraphnetworks

    Haggai Maron, Heli Ben-Hamu, Nadav Shamir, and Yaron Lipman. Invariant and equivariantgraphnetworks. In InternationalConferenceonLearningRepresentations , 2019

    work page 2019

Show all 65 references
  1. [10]

    Aconciseandprovablyinformative multi-scalesignaturebasedonheatdiffusion

    JianSun, MaksOvsjanikov, andLeonidasGuibas. Aconciseandprovablyinformative multi-scalesignaturebasedonheatdiffusion. In Computergraphicsforum ,volume28, pages 1383–1392. Wiley Online Library, 2009

    2009

  2. [11]

    Representing edge flows on graphs via sparse cell complexes

    Josef Hoppe and Michael T Schaub. Representing edge flows on graphs via sparse cell complexes. InLearning on Graphs Conference, pages 1–1. PMLR, 2024

    2024

  3. [12]

    Graph laplacians and their convergence on random neighborhood graphs.Journal of Machine Learning Research, 8(6), 2007

    Matthias Hein, Jean-Yves Audibert, and Ulrike von Luxburg. Graph laplacians and their convergence on random neighborhood graphs.Journal of Machine Learning Research, 8(6), 2007

    2007

  4. [13]

    Global intrinsic symmetries of shapes

    Maks Ovsjanikov, Jian Sun, and Leonidas Guibas. Global intrinsic symmetries of shapes. InComputer graphics forum, volume 27, pages 1341–1348. Wiley Online Library, 2008

    2008

  5. [14]

    Witten–morse theory for cell complexes.Topology, 37(5):945–979, 1998

    Robin Forman. Witten–morse theory for cell complexes.Topology, 37(5):945–979, 1998

    1998

  6. [15]

    Multilayer feedforward networks are universal approximators.Neural networks, 2(5):359–366, 1989

    Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators.Neural networks, 2(5):359–366, 1989

    1989

  7. [16]

    Are transformers universal approximators of sequence-to-sequence functions? International Conference on Learning Representations, 2020

    Chulhee Yun, Srinadh Bhojanapalli, Ankit Singh Rawat, Sashank J Reddi, and Sanjiv Kumar. Are transformers universal approximators of sequence-to-sequence functions? International Conference on Learning Representations, 2020. 16

    2020

  8. [17]

    How powerful are graph neural networks?arXiv preprint arXiv:1810.00826, 2018

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks?arXiv preprint arXiv:1810.00826, 2018

    2018 arXiv

  9. [18]

    arXiv preprint arXiv:2409.13467, 2024

    RomanJoeresandDanielBojar.Higher-ordermessagepassingforglycanrepresentation learning. arXiv preprint arXiv:2409.13467, 2024

    2024 arXiv

  10. [19]

    Graph laplacian for semi-supervised learning

    Or Streicher and Guy Gilboa. Graph laplacian for semi-supervised learning. In InternationalConferenceonScaleSpaceandVariationalMethodsinComputerVision , pages 250–262. Springer, 2023

    2023

  11. [20]

    American Mathematical Soc., 1997

    Fan RK Chung.Spectral graph theory, volume 92. American Mathematical Soc., 1997

    1997

  12. [21]

    On spectral clustering: Analysis and an algorithm

    Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001

    2001

  13. [22]

    Laplacian eigenmaps for dimensionality reduction and data representation.Neural computation, 15(6):1373–1396, 2003

    Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps for dimensionality reduction and data representation.Neural computation, 15(6):1373–1396, 2003

    2003

  14. [23]

    Diffusion maps.Applied and computational harmonic analysis, 21(1):5–30, 2006

    Ronald R Coifman and Stéphane Lafon. Diffusion maps.Applied and computational harmonic analysis, 21(1):5–30, 2006

    2006

  15. [24]

    In International conference on machine learning, pages 6878–6917

    Guoji Fu, Peilin Zhao, and Yatao Bian.𝑝-laplacian based graph neural networks. In International conference on machine learning, pages 6878–6917. PMLR, 2022

    2022

  16. [25]

    Robust graph neural networks using weighted graph laplacian.arXiv preprint arXiv:2208.01853, 2022

    Bharat Runwal, Sandeep Kumar, et al. Robust graph neural networks using weighted graph laplacian.arXiv preprint arXiv:2208.01853, 2022

    2022 arXiv

  17. [26]

    Weisfeiler and lehman go topological: Message passing simplicial networks

    Cristian Bodnar, Fabrizio Frasca, Yuguang Wang, Nina Otter, Guido F Montufar, Pietro Lio, and Michael Bronstein. Weisfeiler and lehman go topological: Message passing simplicial networks. InInternational conference on machine learning, pages 1026–1037. PMLR, 2021

    2021

  18. [27]

    E(n) equivariant topological neural networks

    Claudio Battiloro, Ege Karaismailoglu, Mauricio Tec, George Dasoulas, Michelle Audirac, and Francesca Dominici. E(n) equivariant topological neural networks. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  19. [28]

    Principled simplicial neural networks for trajectory prediction

    T Mitchell Roddenberry, Nicholas Glaze, and Santiago Segarra. Principled simplicial neural networks for trajectory prediction. InInternational Conference on Machine Learning, pages 9020–9029. PMLR, 2021

    2021

  20. [29]

    Convolutional learning on simplicial complexes

    Maosheng Yang and Elvin Isufi. Convolutional learning on simplicial complexes. arXiv preprint arXiv:2301.11163, 2023

    2023 arXiv

  21. [30]

    Simplicial neural networks

    Stefania Ebli, Michaël Defferrard, and Gard Spreemann. Simplicial neural networks. arXiv preprint arXiv:2010.03633, 2020

    2010 arXiv

  22. [31]

    Simplicial attention networks

    Christopher Wei Jin Goh, Cristian Bodnar, and Pietro Lio. Simplicial attention networks. arXiv preprint arXiv:2204.09455, 2022

    2022 arXiv

  23. [32]

    Generalized simplicial attention neural networks.IEEE Transactions on Signal and Information Processing over Networks, 2024

    Claudio Battiloro, Lucia Testa, Lorenzo Giusti, Stefania Sardellitti, Paolo Di Lorenzo, and Sergio Barbarossa. Generalized simplicial attention neural networks.IEEE Transactions on Signal and Information Processing over Networks, 2024. 17

    2024

  24. [33]

    Simplicial representation learning with neural𝑘-forms

    Kelly Maggs, Celia Hacker, and Bastian Rieck. Simplicial representation learning with neural𝑘-forms. arXiv preprint arXiv:2312.08515, 2023

    2023 arXiv

  25. [34]

    Higher-order topological directionality and directed simplicial neural networks

    ManuelLecha,AndreaCavallo,FrancescaDominici,ElvinIsufi,andClaudioBattiloro. Higher-order topological directionality and directed simplicial neural networks. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025

    2025

  26. [35]

    Weisfeiler and lehman go cellular: Cw networks

    Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yuguang Wang, Pietro Lio, Guido F Montufar, and Michael Bronstein. Weisfeiler and lehman go cellular: Cw networks. Advances in neural information processing systems, 34:2625–2640, 2021

    2021

  27. [36]

    Higher-order attention networks.arXiv preprint arXiv:2206.00606, 2(3):4, 2022

    Mustafa Hajij, Ghada Zamzmi, Theodore Papamarkou, Nina Miolane, Aldo Guzmán- Sáenz, and Karthikeyan Natesan Ramamurthy. Higher-order attention networks.arXiv preprint arXiv:2206.00606, 2(3):4, 2022

    2022 arXiv

  28. [37]

    Architecturesof topological deep learning: A survey of message-passing topological neural networks

    MathildePapillon,SophiaSanborn,MustafaHajij,andNinaMiolane. Architecturesof topological deep learning: A survey of message-passing topological neural networks. arXiv preprint arXiv:2304.10031, 2023

    2023 arXiv

  29. [38]

    Topological graph neural networks

    Max Horn, Edward De Brouwer, Michael Moor, Yves Moreau, Bastian Rieck, and Karsten Borgwardt. Topological graph neural networks. arXiv preprint arXiv:2102.07835, 2021

    2021 arXiv

  30. [39]

    Topological relational learning on graphs

    Yuzhou Chen, Baris Coskunuzer, and Yulia Gel. Topological relational learning on graphs. Advances in neural information processing systems, 34:27029–27042, 2021

    2021

  31. [40]

    On the expressivity of persistent homology in graph learning.arXiv preprint arXiv:2302.09826, 2023

    Rubén Ballester and Bastian Rieck. On the expressivity of persistent homology in graph learning.arXiv preprint arXiv:2302.09826, 2023

    2023 arXiv

  32. [41]

    Cliqueph: Higher-order information for graph neural networks through persistent homology on clique graphs

    Davide Buffelli, Farzin Soleymani, and Bastian Rieck. Cliqueph: Higher-order information for graph neural networks through persistent homology on clique graphs. arXiv preprint arXiv:2409.08217, 2024

    2024 arXiv

  33. [42]

    Schaub, Nina Miolane, Simone Scardapane, and TheodorePapamarkou

    Lev Telyatnikov, Guillermo Bernárdez, Marco Montagna, Pavlo Vasylenko, Ghada Zamzmi, Mustafa Hajij, Michael T. Schaub, Nina Miolane, Simone Scardapane, and TheodorePapamarkou. Topobenchmarkx: Aframeworkforbenchmarkingtopological deep learning.CoRR, abs/2406.06642, 2024

    2024 arXiv

  34. [43]

    Topotune: Aframeworkforgeneralizedcombinatorialcomplexneuralnetworks

    Mathilde Papillon, Guillermo Bernárdez, Claudio Battiloro, and Nina Miolane. Topotune: Aframeworkforgeneralizedcombinatorialcomplexneuralnetworks. arXiv preprint arXiv:2410.06530, 2024

    2024

  35. [44]

    Weisfeiler and leman go neural: Higher- order graph neural networks

    Christopher Morris, Martin Ritzert, Matthias Fey, William L Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and leman go neural: Higher- order graph neural networks. InProceedings of the AAAI conference on artificial intelligence, volume 33, pages 4602–...

    2019

  36. [45]

    The reduction of a graph to canonical form and the algebra which appears therein.nti, Series, 2(9):12–16, 1968

    Boris Weisfeiler and Andrei Leman. The reduction of a graph to canonical form and the algebra which appears therein.nti, Series, 2(9):12–16, 1968. 18

    1968

  37. [46]

    The surprising power of graph neural networks with random node initialization.arXiv preprint arXiv:2010.01179, 2020

    Ralph Abboud, Ismail Ilkan Ceylan, Martin Grohe, and Thomas Lukasiewicz. The surprising power of graph neural networks with random node initialization.arXiv preprint arXiv:2010.01179, 2020

    2010 arXiv

  38. [47]

    Improving graph neural network expressivity via subgraph isomorphism counting

    Giorgos Bouritsas, Fabrizio Frasca, Stefanos Zafeiriou, and Michael M Bronstein. Improving graph neural network expressivity via subgraph isomorphism counting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):657–668, 2022

    2022

  39. [48]

    Equivariant polynomials for graph neural networks

    Omri Puny, Derek Lim, Bobak Kiani, Haggai Maron, and Yaron Lipman. Equivariant polynomials for graph neural networks. InInternational Conference on Machine Learning, pages 28191–28222. PMLR, 2023

    2023

  40. [49]

    Equivariant subgraphaggregationnetworks

    Beatrice Bevilacqua, Fabrizio Frasca, Derek Lim, Balasubramaniam Srinivasan, Chen Cai, Gopinath Balamurugan, Michael M Bronstein, and Haggai Maron. Equivariant subgraphaggregationnetworks. InternationalConferenceonLearningRepresentations , 2022

    2022

  41. [50]

    Under- standing and extending subgraph gnns by rethinking their symmetries.Advances in Neural Information Processing Systems, 35:31376–31390, 2022

    Fabrizio Frasca, Beatrice Bevilacqua, Michael Bronstein, and Haggai Maron. Under- standing and extending subgraph gnns by rethinking their symmetries.Advances in Neural Information Processing Systems, 35:31376–31390, 2022

    2022

  42. [51]

    A complete expressiveness hierarchy for subgraph gnns via subgraph weisfeiler-lehman tests

    Bohang Zhang, Guhao Feng, Yiheng Du, Di He, and Liwei Wang. A complete expressiveness hierarchy for subgraph gnns via subgraph weisfeiler-lehman tests. In International Conference on Machine Learning, pages 41019–41077. PMLR, 2023

    2023

  43. [52]

    Nested graph neural networks.Advances in Neural Information Processing Systems, 34:15734–15747, 2021

    Muhan Zhang and Pan Li. Nested graph neural networks.Advances in Neural Information Processing Systems, 34:15734–15747, 2021

    2021

  44. [53]

    Reconstruction for powerful graph representations.Advances in Neural Information Processing Systems, 34:1713– 1726, 2021

    Leonardo Cotta, Christopher Morris, and Bruno Ribeiro. Reconstruction for powerful graph representations.Advances in Neural Information Processing Systems, 34:1713– 1726, 2021

    2021

  45. [54]

    A flexible, equivariant framework for subgraph GNNs via graph products and graph coarsening

    Guy Bar-Shalom, Yam Eitan, Fabrizio Frasca, and Haggai Maron. A flexible, equivariant framework for subgraph GNNs via graph products and graph coarsening. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  46. [55]

    A topological characterisation of weisfeiler-leman equivalence classes

    Jacob Bamberger. A topological characterisation of weisfeiler-leman equivalence classes. InTopological, Algebraic and Geometric Learning Workshops 2022, pages 17–27. PMLR, 2022

    2022

  47. [56]

    Theory of graph neural networks: Representation and learning

    Stefanie Jegelka. Theory of graph neural networks: Representation and learning. In The International Congress of Mathematicians, pages 1–23, 2022

    2022

  48. [57]

    Weisfeilerandlemangomachine learning: The story so far.Journal of Machine Learning Research, 24(333):1–59, 2023

    Christopher Morris, Yaron Lipman, Haggai Maron, Bastian Rieck, Nils M Kriege, MartinGrohe,MatthiasFey,andKarstenBorgwardt. Weisfeilerandlemangomachine learning: The story so far.Journal of Machine Learning Research, 24(333):1–59, 2023. 19

    2023

  49. [58]

    The expressive power of graph neural networks: A survey.IEEE Transactions on Knowledge and Data Engineering, 2024

    BingxuZhang,ChangjunFan,ShixuanLiu,KuihuaHuang,XiangZhao,JincaiHuang, and Zhong Liu. The expressive power of graph neural networks: A survey.IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  50. [59]

    Scale-invariant heat kernel signatures for non-rigid shape recognition

    Michael M Bronstein and Iasonas Kokkinos. Scale-invariant heat kernel signatures for non-rigid shape recognition. In2010 IEEE computer society conference on computer vision and pattern recognition, pages 1704–1711. IEEE, 2010

    2010

  51. [60]

    Volu- metric heat kernel signatures

    Dan Raviv, Michael M Bronstein, Alexander M Bronstein, and Ron Kimmel. Volu- metric heat kernel signatures. InProceedings of the ACM workshop on 3D object retrieval, pages 39–44, 2010

    2010

  52. [61]

    Onepoint isometric matching with the heat kernel

    MaksOvsjanikov,QuentinMérigot,FacundoMémoli,andLeonidasGuibas. Onepoint isometric matching with the heat kernel. InComputer Graphics Forum, volume 29, pages 1555–1564. Wiley Online Library, 2010

    2010

  53. [62]

    A novel graph kernel based on the wasserstein distance and spectral signatures

    Yantao Liu, Luca Rossi, and Andrea Torsello. A novel graph kernel based on the wasserstein distance and spectral signatures. InJoint IAPR International Workshops on Statistical Techniques in Pattern Recognition (SPR) and Structural and Syntactic Pattern Recognition (SSPR), pag...

    2022

  54. [63]

    The wave kernel signature: A quantum mechanical approach to shape analysis

    Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The wave kernel signature: A quantum mechanical approach to shape analysis. In2011 IEEE international conferenceoncomputervisionworkshops(ICCVworkshops) ,pages1626–1633.IEEE, 2011

    2011

  55. [64]

    Learning structural node embeddings via diffusion wavelets

    Claire Donnat, Marinka Zitnik, David Hallac, and Jure Leskovec. Learning structural node embeddings via diffusion wavelets. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1320–1329, 2018

    2018

  56. [65]

    Spectral coarsening with hodge laplacians

    Alexandros Keros and Kartic Subr. Spectral coarsening with hodge laplacians. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023

    2023

  57. [66]

    Laplacian matrices of graphs: a survey.Linear algebra and its applications, 197:143–176, 1994

    Russell Merris. Laplacian matrices of graphs: a survey.Linear algebra and its applications, 197:143–176, 1994. 20 8 Appendix In this part of the Appendix, we fully describe the proofs and provide further definitions if needed. We also include the full text for completeness and...

    1994

Pith tools

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