pith. sign in

arxiv: 2607.00556 · v1 · pith:5OWRBKVNnew · submitted 2026-07-01 · 💻 cs.LG · cs.AI

Group-Equivariant Poincar\'e Convolutional Networks

Aiden Durrant , Rahul Baburajan , Georgios Leontidis This is my paper

Pith reviewed 2026-07-02 15:59 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords hyperbolic neural networksgroup equivariancePoincaré ballequivariant convolutionsResNethyperbolic geometrydiscrete symmetry groups
0
0 comments X

The pith

Equivariant Poincaré ResNets embed discrete symmetry groups into hyperbolic space to reduce optimization space and accelerate convergence while respecting manifold boundaries.

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

The paper proposes Equivariant Poincaré ResNets that combine hyperbolic geometry with discrete symmetry groups C4 and D4. Standard hyperbolic networks treat spatial transformations as distinct concepts, causing redundant parameters and slow optimization due to Riemannian gradients. By introducing geometrically safe tensor reshaping, left-regular permutations for group convolutions, and joint-orientation batch normalisation, the approach embeds equivariance directly. This reduces the search space during training and speeds convergence without violating the Poincaré ball's strict boundaries or breaking equivariance properties.

Core claim

Embedding equivariance drastically reduces the optimisation space, accelerating convergence while respecting the boundary constraints of the Poincaré ball and preserving spatial-group equivariance, achieved through adaptations that maintain hyperbolic geometry.

What carries the argument

Geometrically safe tensor reshaping combined with left-regular permutations for hyperbolic group convolutions and joint-orientation Poincaré Midpoint Batch normalisation, which allow discrete group actions to be applied in hyperbolic space.

If this is right

  • Equivariance reduces redundant parameter usage by treating transformed versions of objects as the same hierarchical concept.
  • Convergence accelerates because the optimization space is smaller while still respecting manifold constraints.
  • Spatial-group equivariance is preserved, meaning the network respects both spatial transformations and the group symmetries.
  • The methods ensure no hidden distortions are introduced to the hyperbolic manifold.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • These techniques could extend to other manifolds where standard Euclidean equivariance methods fail due to curvature.
  • Applications in hierarchical data like trees or graphs might benefit from combined hyperbolic and equivariant properties.
  • Testing on datasets with rotational symmetries could show improved generalization.

Load-bearing premise

The proposed geometrically safe tensor reshaping, left-regular permutations, and joint-orientation Poincaré Midpoint Batch normalisation actually maintain both hyperbolic geometry and group equivariance without introducing hidden distortions or breaking the manifold constraints.

What would settle it

An experiment measuring whether the proposed batch normalisation keeps all points strictly inside the Poincaré ball after transformations, or if group convolutions produce outputs that violate the manifold geometry.

Figures

Figures reproduced from arXiv: 2607.00556 by Aiden Durrant, Georgios Leontidis, Rahul Baburajan.

Figure 1
Figure 1. Figure 1: Data-Efficiency Curve. Top￾1 test accuracy evaluated across discrete fractions of the training dataset. 0 20 40 60 80 100 Epoch 30 40 50 60 70 80 90 Accuracy (%) Poincaré C4 Poincaré D4 Poincaré [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualizing C4 equivariance. Standard convolutions (a) fail to preserve the geometric relationship of the rotated inputs, whereas C4-equivariant lifting layers (b) predictably rotate and permute the feature maps. Note, visualisations are projected onto the Euclidean plane. Features near the boundary appear distorted due to the Poincaré metric, yet the equivariant channel routing remains invariant to the di… view at source ↗
Figure 4
Figure 4. Figure 4: Effect of varying curvature mag￾nitudes (c ∈ {0.001, 0.01, 0.1, 0.5, 1.0}) on CIFAR-10 performance within the Poincaré manifold [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
read the original abstract

While recent advancements like the Poincar\'e ResNet have demonstrated the potential of learning visual representations directly in hyperbolic space, their optimisation remains hampered by the computationally intensive nature of Riemannian gradients and the strict boundaries of the manifold. Furthermore, standard hyperbolic networks treat spatial transformations of the same object as distinct hierarchical concepts, leading to redundant parameter usage and vanishing signals. We propose Equivariant Poincar\'e ResNets, combining hyperbolic geometry with discrete symmetry groups ($C_4$ and $D_4$). We identify critical roadblocks in applying Euclidean equivariance to hyperbolic space and propose geometrically safe tensor reshaping, left-regular permutations for hyperbolic group convolutions, and joint-orientation Poincar\'e Midpoint Batch normalisation. Empirically, embedding equivariance drastically reduces the optimisation space, accelerating convergence while accelerating convergence while respecting the boundary constraints of the Poincar\'e ball and preserving spatial-group equivariance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes Equivariant Poincaré ResNets that integrate discrete group equivariance (C4 and D4) into hyperbolic Poincaré ball networks. It identifies roadblocks in applying Euclidean equivariance to hyperbolic space and introduces three adaptations: geometrically safe tensor reshaping, left-regular permutations for hyperbolic group convolutions, and joint-orientation Poincaré Midpoint Batch normalisation. The central claim is that embedding equivariance drastically reduces the optimisation space, accelerates convergence, respects the boundary constraints of the Poincaré ball, and preserves spatial-group equivariance.

Significance. If the proposed adaptations are shown to maintain both hyperbolic geometry and group equivariance without distortions, the work could advance efficient representation learning in hyperbolic spaces for symmetric visual data by reducing redundant parameters and mitigating vanishing signals, building on prior Poincaré ResNet results.

major comments (2)
  1. [Abstract] Abstract: The empirical claim that the adaptations 'drastically reduce the optimisation space, accelerating convergence while respecting the boundary constraints' supplies no quantitative results, error bars, baselines, or experiment details, rendering the central claim unverifiable from the provided text.
  2. [Abstract] Abstract: The assertion that geometrically safe tensor reshaping, left-regular permutations, and joint-orientation Poincaré Midpoint Batch normalisation 'maintain both hyperbolic geometry and group equivariance without introducing hidden distortions or breaking the manifold constraints' is stated without derivation, proof, or verification that these operations commute with C4/D4 actions and preserve the Poincaré ball radius, which is load-bearing for the strongest claim.
minor comments (1)
  1. [Abstract] Abstract: The phrase 'accelerating convergence while accelerating convergence' contains an obvious repetition that should be removed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting issues in the abstract. We agree the claims require better support and will revise the abstract in the next version to address both points directly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The empirical claim that the adaptations 'drastically reduce the optimisation space, accelerating convergence while respecting the boundary constraints' supplies no quantitative results, error bars, baselines, or experiment details, rendering the central claim unverifiable from the provided text.

    Authors: We acknowledge the abstract presents the empirical benefits at a high level without numbers. The full manuscript reports experiments with baselines, convergence metrics, and parameter counts demonstrating the claimed reductions. We will revise the abstract to include specific quantitative results (e.g., convergence speed-ups and optimization-space reductions with error bars) drawn from those experiments. revision: yes

  2. Referee: [Abstract] Abstract: The assertion that geometrically safe tensor reshaping, left-regular permutations, and joint-orientation Poincaré Midpoint Batch normalisation 'maintain both hyperbolic geometry and group equivariance without introducing hidden distortions or breaking the manifold constraints' is stated without derivation, proof, or verification that these operations commute with C4/D4 actions and preserve the Poincaré ball radius, which is load-bearing for the strongest claim.

    Authors: The body of the manuscript contains the geometric arguments, permutation definitions, and batch-norm derivations showing preservation of the Poincaré ball and commutation with C4/D4 actions. The abstract is intended as a summary. We will revise it to note that these properties are verified in Sections 3–4 and, if space permits, add a one-sentence indication of the key invariance arguments. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on proposed architectural adaptations without reduction to inputs by construction.

full rationale

The provided abstract and context describe identification of roadblocks in applying Euclidean equivariance to hyperbolic space, followed by proposals for geometrically safe tensor reshaping, left-regular permutations, and joint-orientation Poincaré Midpoint Batch normalisation. These are presented as novel solutions whose empirical effects (reduced optimisation space, faster convergence, preserved constraints) are asserted as outcomes rather than derived quantities. No equations, fitted parameters, or self-citations are quoted that would make any prediction equivalent to its inputs by definition. The derivation chain is self-contained against external benchmarks, with no load-bearing self-referential steps or renamings of known results.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; all technical content is at the level of named operations without derivation details.

pith-pipeline@v0.9.1-grok · 5686 in / 1042 out tokens · 19064 ms · 2026-07-02T15:59:34.790583+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 9 canonical work pages · 3 internal anchors

  1. [1]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Atigh, M.G., Schoep, J., Acar, E., Van Noord, N., Mettes, P.: Hyperbolic image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4453–4462 (2022)

    2022

  2. [2]

    VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning

    Bardes, A., Ponce, J., LeCun, Y.: Vicreg: Variance-invariance-covariance regular- ization for self-supervised learning. arXiv preprint arXiv:2105.04906 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  3. [3]

    Advances in neural information processing systems34, 18673– 18684 (2021)

    Bietti, A., Venturi, L., Bruna, J.: On the sample complexity of learning under geometric stability. Advances in neural information processing systems34, 18673– 18684 (2021)

    2021

  4. [4]

    Journal of High Energy Physics2024(3), 1–66 (2024)

    Bogatskiy, A., Hoffman, T., Miller, D.W., Offermann, J.T., Liu, X.: Explainable equivariant neural networks for particle physics: Pelican. Journal of High Energy Physics2024(3), 1–66 (2024)

    2024

  5. [5]

    Flavors of geometry31(59-115), 2 (1997)

    Cannon, J.W., Floyd, W.J., Kenyon, R., Parry, W.R., et al.: Hyperbolic geometry. Flavors of geometry31(59-115), 2 (1997)

    1997

  6. [6]

    Advances in neural information processing systems32(2019)

    Chami, I., Ying, Z., Ré, C., Leskovec, J.: Hyperbolic graph convolutional neural networks. Advances in neural information processing systems32(2019)

    2019

  7. [7]

    In: International conference on machine learning

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: International conference on machine learning. pp. 1597–1607. PMLR (2020)

    2020

  8. [8]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606–3613 (2014)

    2014

  9. [9]

    In: Interna- tional conference on machine learning

    Cohen, T., Welling, M.: Group equivariant convolutional networks. In: Interna- tional conference on machine learning. pp. 2990–2999. PMLR (2016)

    2016

  10. [10]

    In: International conference on machine learning

    Elesedy, B., Zaidi, S.: Provably strict generalisation benefit for equivariant models. In: International conference on machine learning. pp. 2959–2969. PMLR (2021)

    2021

  11. [11]

    In: International Conference on Machine Learning

    Elsayed, G., Ramachandran, P., Shlens, J., Kornblith, S.: Revisiting spatial invari- ance with low-rank local connectivity. In: International Conference on Machine Learning. pp. 2868–2879. PMLR (2020)

    2020

  12. [12]

    In: International conference on machine learning

    Finzi, M., Welling, M., Wilson, A.G.: A practical method for constructing equivari- ant multilayer perceptrons for arbitrary matrix groups. In: International conference on machine learning. pp. 3318–3328. PMLR (2021)

    2021

  13. [13]

    Advances in neural information processing systems31(2018)

    Ganea, O., Bécigneul, G., Hofmann, T.: Hyperbolic neural networks. Advances in neural information processing systems31(2018)

    2018

  14. [14]

    Knowledge-Based Systems291, 111590 (2024) 16 A

    Huang, B., Yu, W., Xie, R., Luo, J., Xiao, J., Huang, J.: Lorentz equivariant model for knowledge-enhanced hyperbolic collaborative filtering. Knowledge-Based Systems291, 111590 (2024) 16 A. Durrant et al

    2024

  15. [15]

    In: International Conference on Machine Learning

    Kaba, S.O., Mondal, A.K., Zhang, Y., Bengio, Y., Ravanbakhsh, S.: Equivariance with learned canonicalization functions. In: International Conference on Machine Learning. pp. 15546–15566. PMLR (2023)

    2023

  16. [16]

    arXiv preprint arXiv:2312.09016 (2023)

    Kaba, S.O., Ravanbakhsh, S.: Symmetry breaking and equivariant neural networks. arXiv preprint arXiv:2312.09016 (2023)

    work page arXiv 2023

  17. [17]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Khrulkov, V., Mirvakhabova, L., Ustinova, E., Oseledets, I., Lempitsky, V.: Hyper- bolic image embeddings. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 6418–6428 (2020)

    2020

  18. [18]

    Advances in Neural Infor- mation Processing Systems36, 18582–18612 (2023)

    Kim, J., Nguyen, D., Suleymanzade, A., An, H., Hong, S.: Learning probabilistic symmetrization for architecture agnostic equivariance. Advances in Neural Infor- mation Processing Systems36, 18582–18612 (2023)

    2023

  19. [19]

    In: International conference on machine learning

    Kondor, R., Trivedi, S.: On the generalization of equivariance and convolution in neural networks to the action of compact groups. In: International conference on machine learning. pp. 2747–2755. PMLR (2018)

    2018

  20. [20]

    Krioukov, D., Papadopoulos, F., Kitsak, M., Vahdat, A., Boguná, M.: Hyperbolic geometryofcomplexnetworks.PhysicalReviewE—Statistical,Nonlinear,andSoft Matter Physics82(3), 036106 (2010)

    2010

  21. [21]

    Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. . (2009)

    2009

  22. [22]

    In: Proceedings of the 57th annual meeting of the association for computational linguistics

    Le, M., Roller, S., Papaxanthos, L., Kiela, D., Nickel, M.: Inferring concept hier- archies from text corpora via hyperbolic embeddings. In: Proceedings of the 57th annual meeting of the association for computational linguistics. pp. 3231–3241 (2019)

    2019

  23. [23]

    Advances in neural information processing systems33, 21464–21475 (2020)

    Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems33, 21464–21475 (2020)

    2020

  24. [24]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Long, T., Mettes, P., Shen, H.T., Snoek, C.G.: Searching for actions on the hyper- bole. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1141–1150 (2020)

    2020

  25. [25]

    In: International conference on machine learning

    Lou, A., Katsman, I., Jiang, Q., Belongie, S., Lim, S.N., De Sa, C.: Differentiating through the fréchet mean. In: International conference on machine learning. pp. 6393–6403. PMLR (2020)

    2020

  26. [26]

    In: Medical Imaging with Deep Learning (2026)

    Mishra, S.S., van Spengler, M., Berkhout, E., Mettes, P.: Hyperbolic u-net for ro- bust medical image segmentation. In: Medical Imaging with Deep Learning (2026)

    2026

  27. [27]

    Hyperbolic Graph Neural Networks Under the Microscope: The Role of Geometry-Task Alignment

    Naddeo, D., Linkerhägner, J., Toschi, N., Skenderi, G., Lachi, V.: Hyperbolic graph neural networks under the microscope: The role of geometry-task alignment. arXiv preprint arXiv:2602.01828 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  28. [28]

    In: NIPS workshop on deep learning and unsupervised feature learning

    Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y., et al.: Reading digits in natural images with unsupervised feature learning. In: NIPS workshop on deep learning and unsupervised feature learning. vol. 2011, p. 4. Granada (2011)

    2011

  29. [29]

    Advances in neural information processing systems30(2017)

    Nickel, M., Kiela, D.: Poincaré embeddings for learning hierarchical representa- tions. Advances in neural information processing systems30(2017)

    2017

  30. [30]

    Advances in Neural Information Processing Systems36, 61936–61959 (2023)

    Petrache, M., Trivedi, S.: Approximation-generalization trade-offs under (approxi- mate) group equivariance. Advances in Neural Information Processing Systems36, 61936–61959 (2023)

    2023

  31. [31]

    arXiv preprint arXiv:2110.03336 , year=

    Puny, O., Atzmon, M., Ben-Hamu, H., Misra, I., Grover, A., Smith, E.J., Lipman, Y.: Frame averaging for invariant and equivariant network design. arXiv preprint arXiv:2110.03336 (2021)

    work page arXiv 2021

  32. [32]

    In: International conference on machine learning

    Ravanbakhsh, S., Schneider, J., Poczos, B.: Equivariance through parameter- sharing. In: International conference on machine learning. pp. 2892–2901. PMLR (2017) Group-Equivariant Poincaré Convolutional Networks 17

    2017

  33. [33]

    arXiv preprint arXiv:2010.10241 (2020)

    Richemond, P.H., Grill, J.B., Altché, F., Tallec, C., Strub, F., Brock, A., Smith, S., De, S., Pascanu, R., Piot, B., et al.: Byol works even without batch statistics. arXiv preprint arXiv:2010.10241 (2020)

    work page arXiv 2010

  34. [34]

    In: International Symposium on Graph Drawing

    Sarkar, R.: Low distortion delaunay embedding of trees in hyperbolic plane. In: International Symposium on Graph Drawing. pp. 355–366. Springer (2011)

    2011

  35. [35]

    In: International Conference on Artificial Intelligence and Statistics

    Shakerinava, M., Sohrabi, M.M., Ravanbakhsh, S., Lacoste-Julien, S.: Weight- sharing regularization. In: International Conference on Artificial Intelligence and Statistics. pp. 4204–4212. PMLR (2024)

    2024

  36. [36]

    arXiv preprint arXiv:2006.08210 (2020)

    Shimizu, R., Mukuta, Y., Harada, T.: Hyperbolic neural networks++. arXiv preprint arXiv:2006.08210 (2020)

    work page arXiv 2006

  37. [37]

    arXiv preprint arXiv:2303.14027 (2023)

    van Spengler, M., Berkhout, E., Mettes, P.: Poincar\’e resnet. arXiv preprint arXiv:2303.14027 (2023)

    work page arXiv 2023

  38. [38]

    In: Proceedings of the 31st ACM International Conference on Multimedia

    van Spengler, M., Wirth, P., Mettes, P.: Hypll: The hyperbolic learning library. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 9676–9679 (2023)

    2023

  39. [39]

    Poincar\'e GloVe: Hyperbolic Word Embeddings

    Tifrea, A., Bécigneul, G., Ganea, O.E.: Poincar\’e glove: Hyperbolic word embed- dings. arXiv preprint arXiv:1810.06546 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  40. [40]

    Springer Nature (2022)

    Ungar, A.: A gyrovector space approach to hyperbolic geometry. Springer Nature (2022)

    2022

  41. [41]

    Synthesis Lec- tures on Mathematics and Statistics1(1), 1–194 (2008)

    Ungar, A.A.: A gyrovector space approach to hyperbolic geometry. Synthesis Lec- tures on Mathematics and Statistics1(1), 1–194 (2008)

    2008

  42. [42]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Van Spengler, M., Berkhout, E., Mettes, P.: Poincare resnet. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5419–5428 (2023)

    2023

  43. [43]

    CoRRabs/1809.01703(2018),http://arxiv.org/abs/1809.01703

    Vinh, T.D.Q., Tay, Y., Zhang, S., Cong, G., Li, X.: Hyperbolic recommender sys- tems. CoRRabs/1809.01703(2018),http://arxiv.org/abs/1809.01703

    work page arXiv 2018

  44. [44]

    In: The 29th International Conference on Artificial Intelligence and Statistics

    Vlasenko, M., Atigh, M.G., Mettes, P.: Hyperbolic part-whole image segmentation. In: The 29th International Conference on Artificial Intelligence and Statistics

  45. [45]

    In: International Conference on Machine Learning

    Wang, R., Walters, R., Yu, R.: Approximately equivariant networks for imper- fectly symmetric dynamics. In: International Conference on Machine Learning. pp. 23078–23091. PMLR (2022)

    2022

  46. [46]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Weiler, M., Hamprecht, F.A., Storath, M.: Learning steerable filters for rotation equivariant cnns. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 849–858 (2018)

    2018

  47. [47]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yan, J., Luo, L., Deng, C., Huang, H.: Unsupervised hyperbolic metric learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12465–12474 (2021)

    2021

  48. [48]

    ACM Transactions on Infor- mation Systems44(5), 1–27 (2026)

    Zhang, Q., Wen, H., Yuan, W., Chen, C., Yang, M., Yiu, S.M., Yin, H.: Hmamba: Hyperbolic mamba for sequential recommendation. ACM Transactions on Infor- mation Systems44(5), 1–27 (2026)

    2026

  49. [49]

    IEEE transactions on pattern analysis and machine intelligence40(6), 1452–1464 (2017) 18 A

    Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence40(6), 1452–1464 (2017) 18 A. Durrant et al. A Prerequisites A.1 Hyperbolic Learning: The Poincaré Ball Model Hyperbolic spaceB n is the unique simply connectedn-dimensional ...

    2017