Pith. sign in

REVIEW 4 major objections 6 minor 27 references

sHGCN: Simplified hyperbolic graph convolutional neural networks

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a simplified hyperbolic graph convolutional network, which cancels the log/exp compositions inside each layer and applies activation and aggregation in Euclidean space, runs faster and matches or beats the original…

desk verdict The simplification is correct and worth having, but the headline accuracy claim rests on an uncontrolled decoder comparison. read the letter →

arxiv 2506.14438 v1 pith:32N6VR3Y submitted 2025-06-17 cs.LG cs.AI

Pol Arévalo , Alexis Molina , Álvaro Ciudad This is my paper
classification cs.LGcs.AI
keywords hyperbolicgraphconvolutionalnetworksPoincaréballlinkpredictionnodeclassificationregressionnumericalstabilitylog-expcancellationEuclideandistancedecoder
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 proposes a simplified hyperbolic graph convolutional network (sHGCN) that removes the redundant $\log(\exp(x))$ compositions inside each layer and carries out activation and aggregation on Euclidean tangent vectors rather than on the Poincaré ball. The authors claim this streamlined model delivers state-of-the-art link prediction, node classification, and graph regression results while running faster than the original HGCN. A sympathetic reading: the paper's core bet is that hyperbolic feature transforms remain useful, but the extra exponential/logarithmic round-trips in the original architecture are both unnecessary and numerically harmful. If that bet holds, hyperbolic GNNs can be adopted at near-Euclidean cost, which is the practical payoff emphasized throughout.

What carries the argument

The load-bearing identity is $\log_0(\exp_0(x)) = x$ in the Poincaré ball, which lets the two maps in each layer cancel. The proposal fixes aggregation weights to one, aggregates in the tangent space at the origin, and applies the nonlinear activation after the log map, so the only remaining hyperbolic operation is the feature transform $\exp^c_0(W x) \oplus^c \exp^c_0(b)$. This machinery turns the HGCN layer into a Euclidean GCN with one log/exp round-trip in the feature transform and also justifies the paper's choice of a Euclidean-distance Fermi-Dirac decoder.

What would settle it

Re-run the link prediction protocol with the original HGCN but with the Euclidean-distance decoder used by sHGCN; if the DISEASE AUC matches the 94.6 reported for sHGCN, then the gain is an artifact of the decoder rather than of removing the log/exp compositions.

Watch

Extended reading notes

Core claim

The central discovery is that the consolidated HGCN update in Eq. (7) contains compositions $\log_0(\exp_0(\cdot))$ that are the identity in exact arithmetic but break down numerically for points far from the origin. Deleting those compositions and performing the nonlinearity directly in Euclidean space yields a model whose message-passing rule is $H^l = \sigma(\tilde{A}\log^{c_{l-1}}_0(\exp^{c_{l-1}}_0(W^l H^{l-1}) \oplus \exp^{c_{l-1}}_0(b)))$. The authors report that this simplified model matches or exceeds HGCN accuracy on link prediction, node classification, and graph regression, with the largest gain on the tree-like DISEASE graph (AUC 94.6 vs 90.8), and that it runs faster per epoch and avoids the out-of-memory failure of the attention-based HGCN.

Load-bearing premise

The load-bearing premise is that scoring edges by ordinary Euclidean distance, as sHGCN's decoder does, is a fair comparison against the hyperbolic-distance scores reported for the original HGCN, and that those published baseline numbers are trustworthy even though the authors could not reproduce them.

Editorial extensions

If this is right

  • On the DISEASE link-prediction task, sHGCN raises test AUC from 90.8 (the reported HGCN value) to 94.6 under the same evaluation protocol.
  • Per-epoch runtime improves by 3 to 96 percent over the origin-aggregated HGCN and by 69 to 146 percent over the attention-aggregated HGCN on the datasets tested, and the attention variant runs out of memory on PubMed while sHGCN completes.
  • Because activations act on Euclidean vectors, the model can use standard tools such as ReLU, batch normalization, and residual connections, which the paper exploits in molecular graph regression.
  • On the ZINC graph regression benchmark, a 4-layer sHGCN with 103k parameters reaches test MAE 0.360, beating comparably sized baselines except 3WL-GNN-E while using far less time per epoch.

Reading between the lines

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

  • If the published HGCN numbers are remeasured with the same Euclidean-distance decoder, the DISEASE gap might shrink or vanish; a direct control would localize whether the gain comes from the simplified architecture or from the decoder.
  • The paper's error analysis predicts that the accuracy advantage should grow under float16 training, where the $\log\circ\exp$ threshold drops to $\|x\|>5$; testing sHGCN in mixed precision would sharpen that prediction.
  • The same cancellation argument may transfer to other hyperbolic architectures whose layers round-trip through the origin, such as hyperbolic attention or hyperbolic recurrent networks.
  • A Euclidean GCN with the same decoder and hyperparameters is the natural next baseline; if it matches sHGCN on non-tree graphs, the hyperbolic feature transform only matters on strongly hyperbolic data.
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 / 6 minor

Summary. The paper proposes sHGCN, a simplified hyperbolic graph convolutional network. The simplification removes the log/exp compositions from the HGCN message-passing update by performing feature transformation in the Poincaré ball but carrying out aggregation and nonlinear activation in the tangent space at the origin, yielding a Euclidean-space embedding at each layer. The authors derive the consolidated matrix update in Eq. (14), argue that the removed compositions are identity in exact arithmetic and numerically lossy in floating point, and report experiments on link prediction, node classification, and graph regression. The headline empirical evidence is Table 1, where sHGCN reaches 94.6 AUC on DISEASE link prediction versus 90.8 for the quoted HGCN baseline, Table 2, which shows substantial speedups over HGCN-AGG0 and HGCN-ATT0, and Tables 6-7, which report graph-regression results on ZINC and AQSOL.

Significance. If the empirical claims hold, the simplification is practically useful: it removes a numerically fragile composition, removes the restriction to hyperbolic-compatible activations, and reduces runtime. The algebraic consolidation in Eqs. (4)-(14) is sound at the symbolic level, and the paper ships code with logged hyperparameters and seeds, which is a strength. However, the central claim of state-of-the-art accuracy is not currently supported by the evidence as presented: the headline link-prediction comparison changes the decoder between sHGCN and the HGCN baseline, the HGCN baseline itself was not reproduced, the numerical-error explanation depends on an unstated precision assumption, and the paper's own node-classification numbers are worse than HGCN on several datasets. The contribution is better characterized as an engineering simplification with promising efficiency gains than as an established accuracy improvement.

major comments (4)
  1. [§5, Table 1; Eq. 30; Appendix C] The headline link-prediction gain on DISEASE (94.6 vs 90.8 AUC) is not a controlled comparison. Eq. 30 scores sHGCN with Euclidean distance in the Fermi-Dirac decoder, whereas the quoted HGCN result from Chami et al. used the hyperbolic-distance decoder from the original paper; the decoder is part of the evaluated model, so the 3.8-point gap could be entirely due to the decoder change. Appendix C further states that the original HGCN model could not be reproduced, so the 90.8 baseline is not verified in this work's setup. To support a state-of-the-art link-prediction claim, the authors should rerun HGCN with the same Euclidean decoder or evaluate sHGCN with the hyperbolic decoder, and report both configurations.
  2. [Appendix B, Table 3; Appendix C] The numerical-error explanation for the accuracy gain assumes that training operated in a precision regime where log(exp(x)) degrades. Table 3 gives thresholds t=5 for float16, t=9 for float32, and t=19 for float64, but Appendix C never states the floating-point precision used in the experiments or reports the norms of intermediate embeddings. If the experiments ran in float32 or higher, the paper's own threshold analysis does not establish that error accumulation was actually present. The authors should state the precision and report the distribution of the relevant norms (e.g., ||y|| in Eq. 28) to show that the threshold is crossed.
  3. [Section 1; Table 1] The abstract and Introduction claim state-of-the-art performance in node classification, but Table 1 shows sHGCN below the quoted HGCN on AIRPORT node classification (85.5 vs 90.6) and CORA node classification (76.5 vs 79.9), and roughly tied on DISEASE and PUBMED. The claim should be narrowed to link prediction and computational efficiency, or the node-classification discussion should explicitly address why the table supports the stated claim.
  4. [Appendix D.3; Tables 6 and 7] The statement that sHGCN is 'significantly faster' than 3WL-GNN-E compares sHGCN epoch times measured on the authors' A30 GPU with benchmark epoch times from Dwivedi et al. measured on different hardware. The authors themselves acknowledge the hardware-dependence when validating GCN timing. This is not a controlled speed comparison. The runtime advantage in the graph-regression section should either be backed by re-running the comparison models on the same hardware or be removed, leaving the speed claim to the controlled comparison in Table 2.
minor comments (6)
  1. [Eq. (14)] The bias index is written as b^{l-1}, but it should be b^l to match Eq. (11) and the layer index of the feature-transform bias.
  2. [Table 1 caption] The caption says 'All settings and results are reported from Chami et al. (2019)', but the table also contains newly run HGCN-AGG0 and HGCN-ATT0 rows and the sHGCN row; clarify which rows are quoted from the literature and which were produced in this work.
  3. [Sections 3-5 and Table 2] The notation 'HGCN-ATT 0' and 'HGCN-ATT0' is used inconsistently; unify it throughout the text and tables.
  4. [Table 2] The header 'Speedup (HGCN-ATT0vs sHGCN)' is missing a space, and the definition of speedup (baseline time divided by sHGCN time) should be stated once in the text.
  5. [Appendix B.1, Lemma B.2] The statement that 'the maximum k is 4' should explicitly say that it applies to float16 arithmetic with unit curvature; this is clear from context but should be stated directly.
  6. [Section 5] The sentence 'one notable case highlights how link prediction can result in an Out-Of-Memory error' does not identify the dataset; refer explicitly to the OOM entry in Table 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the sHGCN simplification is a self-contained algebraic reduction and the empirical comparisons target external baselines.

full rationale

The paper's core derivation is the consolidation of HGCN-AGG0 message passing into Eq. 7 and the removal of log(exp(x)) compositions that are identities in the Poincaré ball (Section 3). This is a mathematical identity, not a fitted parameter or a prediction of data; the resulting sHGCN equations (11)-(14) define a new model, and its accuracy is then measured against the external HGCN results of Chami et al. (2019) and the Dwivedi et al. (2020) benchmarks. The main comparison caveats — the Euclidean Fermi-Dirac decoder in Eq. 30 instead of the hyperbolic distance used in the quoted HGCN numbers, and the admitted inability to reproduce the original HGCN baseline (Appendix C, Table 1 note) — are validity/threat-to-comparison issues, not circularity: the decoder is not fitted from the HGCN outcome, the baseline is not the authors' own prior result, and no load-bearing claim is justified solely by a self-citation. Appendix B's numerical-error analysis uses an external proof (Mishne et al., 2023) and independent floating-point facts. Therefore there is no step in which a prediction is equivalent by construction to its input.

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

The central claim rests on standard definitions of the Poincaré ball and on the identity that exp and log cancel. No new physical entities are introduced. The main fitted inputs are the network parameters and the curvature; the most fragile assumption in the ledger is the numerical-threshold table, because the experimental precision is never stated.

free parameters (3)
  • learnable curvature c^l = not reported (learned)
    A per-layer learnable curvature is introduced in Section 4.4; it affects the log/exp maps and the Möbius addition, and its fitted values are not reported in the paper.
  • Fermi-Dirac decoder hyperparameters r and t = not reported
    Link-prediction probabilities in Eq. (30) depend on r and t; the paper does not state how they are chosen or tuned, though AUC results depend on them.
  • weight matrices W^l and biases b^l = learned
    Standard trainable parameters; the bias is load-bearing because the paper notes that removing it reduces the model to a standard GCN (Appendix B.2).
assumptions (3)
  • standard math Poincaré ball model definitions: Möbius addition, exponential/logarithmic maps at the origin, and the identity log_0(exp_0(x)) = x for points inside the open ball.
    Assumed throughout Sections 3-4 and Appendix A; the simplification is derived from this identity.
  • domain assumption Floating-point thresholds in Table 3 (float16=5, float32=9, float64=19) determine when log(exp(x)) becomes ill-defined.
    The numerical-error argument in Appendix B relies on these thresholds, but the paper does not state the training precision used in the experiments, so the assumption may not apply.
  • domain assumption Dataset hyperbolicity values δ from Table 4 (DISEASE=0, AIRPORT=1, PUBMED=3.5, CORA=11) are taken from prior literature as ground truth.
    The claim that sHGCN is especially suited to hyperbolic graphs depends on these δ values; the paper does not recompute them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of sHGCN: Simplified hyperbolic graph convolutional neural networks." pith.science (2026). https://pith.science/paper/32N6VR3Y

@misc{pith2026250614438,
  author       = {Pith},
  title        = {Pith review of: sHGCN: Simplified hyperbolic graph convolutional neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/32N6VR3Y}},
  note         = {Machine review of arXiv:2506.14438}
}
read the original abstract

Hyperbolic geometry has emerged as a powerful tool for modeling complex, structured data, particularly where hierarchical or tree-like relationships are present. By enabling embeddings with lower distortion, hyperbolic neural networks offer promising alternatives to Euclidean-based models for capturing intricate data structures. Despite these advantages, they often face performance challenges, particularly in computational efficiency and tasks requiring high precision. In this work, we address these limitations by simplifying key operations within hyperbolic neural networks, achieving notable improvements in both runtime and performance. Our findings demonstrate that streamlined hyperbolic operations can lead to substantial gains in computational speed and predictive accuracy, making hyperbolic neural networks a more viable choice for a broader range of applications.

Figures

Figures reproduced from arXiv: 2506.14438 by the authors.

Figure 1
Figure 1. Composition log ◦ exp maps becomes inaccurate as points move away from the origin. 4 PROPOSED METHOD 4.1 FEATURE TRANSFORM A linear transformation involves multiplying the embedding vector x by a weight matrix W and adding a bias term b. For feature transformation, we use matrix-vector multiplication in the Poincare´ ball model along with Mobius addition (see Appendix A for a detailed definition of these operations)… view at source ↗
Figure 2
Figure 2. The proposed sHGCN architecture. 5 RESULTS For details on the experimental setup, please refer to Appendix C. Link Prediction. The performance difference is particularly noticeable when comparing our new method to hyperbolic-based models on specific datasets. This is most evident in the Disease dataset (see [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 17 canonical work pages

  1. [1]

    Adcock, Blair D

    Aaron B. Adcock, Blair D. Sullivan, and Michael W. Mahoney. Tree-like structure in large social and information networks. In 2013 IEEE 13th International Conference on Data Mining, pp.\ 1--10, 2013. doi:10.1109/ICDM.2013.77

    work page doi:10.1109/icdm.2013.77 2013

  2. [2]

    Hyperbolic graph convolutional neural networks

    Ines Chami, Zhitao Ying, Christopher R \'e , and Jure Leskovec. Hyperbolic graph convolutional neural networks. In Advances in Neural Information Processing Systems, pp.\ 4869--4880, 2019

    work page 2019

  3. [3]

    Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson

    Vijay Prakash Dwivedi, Chaitanya K. Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. CoRR, abs/2003.00982, 2020. URL https://arxiv.org/abs/2003.00982

    arXiv 2003

  4. [4]

    Hyperbolic neural networks

    Octavian Ganea, Gary Becigneul, and Thomas Hofmann. Hyperbolic neural networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett (eds.), Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018. URL https://proceedings.neurips.cc/paper_files/paper/2018/file/dbab2adc8f9d078009ee3fa810be...

    work page 2018

  5. [5]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/2017/file/...

    work page 2017

  6. [6]

    Scaled gromov hyperbolic graphs

    Edmond Jonckheere, Poonsuk Lohsoonthorn, and Francis Bonahon. Scaled gromov hyperbolic graphs. J. Graph Theory, 57 0 (2): 0 157–180, February 2008. ISSN 0364-9024

    work page 2008

  7. [7]

    Adam: A method for stochastic optimization

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015

    work page 2015

  8. [8]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR), 2017

    2017

Show all 27 references
  1. [9]

    Hyperbolic geometry of complex networks

    Dmitri Krioukov, Fragkiskos Papadopoulos, Maksim Kitsak, Amin Vahdat, and Mari\'an Bogu\ n\'a. Hyperbolic geometry of complex networks. Phys. Rev. E, 82: 0 036106, Sep 2010. doi:10.1103/PhysRevE.82.036106. URL https://link.aps.org/doi/10.1103/PhysRevE.82.036106

    2010 doi

  2. [10]

    A fully hyperbolic neural model for hierarchical multi-class classification

    Federico L \'o pez and Michael Strube. A fully hyperbolic neural model for hierarchical multi-class classification. In Findings of the Association for Computational Linguistics: EMNLP 2020, pp.\ 460--475, Online, November 2020. Association for Computational Linguistics. URL ht...

    2020

  3. [11]

    Fine-grained entity typing in hyperbolic space

    Federico L \'o pez, Benjamin Heinzerling, and Michael Strube. Fine-grained entity typing in hyperbolic space. In Isabelle Augenstein, Spandana Gella, Sebastian Ruder, Katharina Kann, Burcu Can, Johannes Welbl, Alexis Conneau, Xiang Ren, and Marek Rei (eds.), Proceedings of the...

    2019 doi

  4. [12]

    The numerical stability of hyperbolic representation learning

    Gal Mishne, Zhengchao Wan, Yusu Wang, and Sheng Yang. The numerical stability of hyperbolic representation learning. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.), Proceedings of the 40th International Conferen...

    2023

  5. [13]

    Poincar\' e embeddings for learning hierarchical representations

    Maximillian Nickel and Douwe Kiela. Poincar\' e embeddings for learning hierarchical representations. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems, volume 30. Curran Ass...

    2017

  6. [14]

    PyTorch: an imperative style, high-performance deep learning library, pp.\ 1--12

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K\" o pf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu...

    2019

  7. [15]

    Konstantin Rusch, Michael M

    T. Konstantin Rusch, Michael M. Bronstein, and Siddhartha Mishra. A survey on oversmoothing in graph neural networks, 2023. URL https://arxiv.org/abs/2303.10993

    2023 arXiv

  8. [16]

    Representation tradeoffs for hyperbolic embeddings

    Frederic Sala, Chris De Sa, Albert Gu, and Christopher Re. Representation tradeoffs for hyperbolic embeddings. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research,...

    2018

  9. [17]

    Low distortion delaunay embedding of trees in hyperbolic plane

    Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In Marc van Kreveld and Bettina Speckmann (eds.), Graph Drawing, pp.\ 355--366, Berlin, Heidelberg, 2012. Springer Berlin Heidelberg. ISBN 978-3-642-25878-7

    2012

  10. [18]

    M \"o bius gyrovector spaces in quantum information and computation

    Abraham Albert Ungar. M \"o bius gyrovector spaces in quantum information and computation. In Commentationes Mathematicae Universitatis Carolinae, 2008. URL https://api.semanticscholar.org/CorpusID:55856385

    2008

  11. [19]

    Graph attention networks

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio’, and Yoshua Bengio. Graph attention networks. ArXiv, abs/1710.10903, 2017. URL https://api.semanticscholar.org/CorpusID:3292002

    2017 arXiv

  12. [20]

    Simplifying graph convolutional networks

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. Simplifying graph convolutional networks. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedin...

    2019

  13. [21]

    Hyperbolic graph neural networks: A review of methods and applications

    Menglin Yang, Min Zhou, Zhihao Li, Jiahong Liu, Lujia Pan, Hui Xiong, and Irwin King. Hyperbolic graph neural networks: A review of methods and applications. ArXiv, abs/2202.13852, 2022. URL https://api.semanticscholar.org/CorpusID:247158811

    2022 arXiv

  14. [22]

    H yper T ext: Endowing F ast T ext with hyperbolic geometry

    Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. H yper T ext: Endowing F ast T ext with hyperbolic geometry. In Trevor Cohn, Yulan He, and Yang Liu (eds.), Findings of the Association for Computational Linguistics: EMNLP 2020, pp.\ 1166--1171, Online, Nov...

    2020 doi

  15. [23]

    Zitnik, R

    M. Zitnik, R. Sosič, M. W. Feldman, and J. Leskovec. Evolution of resilience in protein interactomes across the tree of life. In Proceedings of the National Academy of Sciences of the United States of America, 2019

    2019

  16. [24]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  17. [25]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  18. [26]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  19. [27]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.