pith. sign in

arxiv: 2605.07270 · v1 · submitted 2026-05-08 · 💻 cs.LG

bispectrum: Selective G-Bispectra Made Practical

Johan Mathe , Adele Myers , Simon Mataigne , Nina Miolane This is my paper

Pith reviewed 2026-05-11 02:19 UTC · model grok-4.3

classification 💻 cs.LG
keywords G-bispectrumgroup invariancepooling layersmachine learningdeep learningPyTorch librarycomputational efficiencyinvariant features
0
0 comments X

The pith

Selective G-bispectra can be computed in linear time and used as pooling layers to improve neural network performance when training data is scarce.

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

The paper develops practical versions of the G-bispectrum, a mathematical object that encodes all information in a signal while remaining unchanged under group transformations such as rotations. By introducing selectivity for finite groups and augmentation for continuous rotation groups, the authors reduce the computational burden enough to embed these invariants directly into deep learning models as differentiable layers. Experiments across three benchmark datasets demonstrate that networks using these selective bispectra as pooling operations achieve higher accuracy than norm pooling, max pooling, or data-augmentation baselines, especially when the number of training samples is small and model capacity is moderate. The work is accompanied by a fully tested PyTorch library covering seven group actions, with reported runtimes in the sub-millisecond range on GPU for typical problem sizes.

Core claim

Selective G-bispectra preserve the completeness and invariance properties of the full bispectrum while lowering the cost from quadratic in group order to linear for finite groups and from cubic to quadratic in band-limit for spherical rotations; when inserted as pooling layers, they yield higher classification accuracy than norm, gated, Fourier-ELU, max, or data-augmented convolutional baselines on standard benchmarks in the low-data regime.

What carries the argument

The selective G-bispectrum, a reduced set of coefficients that retains invariance and completeness while enabling direct differentiation and GPU execution inside neural networks.

If this is right

  • Deep networks gain a drop-in mechanism for exact group invariance without requiring custom group-equivariant layers or heavy data augmentation.
  • Training data requirements for rotation-invariant tasks can be lowered while maintaining or improving accuracy.
  • The same selective construction extends across translations, planar rotations, and spherical rotations inside one unified library.
  • Sub-millisecond per-sample computation makes the method viable for real-time or large-batch pipelines at common band-limits.

Where Pith is reading between the lines

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

  • The same selectivity pattern could be applied to other complete invariants such as higher-order moments or learned descriptors.
  • In very high-data regimes the relative benefit over data augmentation may shrink, suggesting a regime-dependent choice of pooling.
  • Combining selective bispectra with equivariant convolutions might further reduce sample complexity beyond what either technique achieves alone.

Load-bearing premise

The reduced selective and augmented bispectra still capture enough signal information to produce the observed accuracy gains over simpler pooling methods.

What would settle it

On a rotation-invariant image classification task with a few hundred training examples, a network using selective bispectrum pooling would need to show no accuracy improvement over max pooling or data-augmented baselines under identical architecture and training conditions.

Figures

Figures reproduced from arXiv: 2605.07270 by Adele Myers, Johan Mathe, Nina Miolane, Simon Mataigne.

Figure 1
Figure 1. Figure 1: Φsel preserves enough information to recover signals up to rotation (L=12). Top: targets. Bottom: reconstructions from Φsel alone, Procrustes-aligned. Median image residual ≈0.20; bispec￾trum residuals ∼10−3 . 10 5 10 6 Parameters 0.750 0.775 0.800 0.825 0.850 0.875 0.900 0.925 0.950 Test AUC (a) AUC vs. parameter count (10% data) Bispectrum Standard (augmented) NormReLU Gated Fourier-ELU 0.75 0.80 0.85 0.… view at source ↗
Figure 2
Figure 2. Figure 2: Bispectral pooling on PatchCamelyon (C8, 10% data). (a) All equivariant methods dominate the augmented CNN; the bispectrum is most stable across capacities. (b) At ∼100K params, methods cluster within 0.5pp. (c) At 1% data the bispectrum leads (0.912 AUC), consistent with complete invariants compensating for scarce data. Multi-seed statistics in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Data efficiency on OrganMNIST3D. At 5% data, the bispectrum reaches [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Completeness, not capacity, drives the accuracy gap. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Bispectral coefficient count vs. group order [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Forward-pass wall-clock time (batch = 16, H100 GPU). [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a) GPU throughput (samples/s) vs. batch size for the selective forward pass. Cn and T 2 exceed 107 samples/s; the octahedral module reaches 5.7 × 106 samples/s. All modules scale linearly with batch size. (b) Inversion wall-clock time. Abelian groups (Cn, Dn, SO(2)/disk) scale roughly linearly. SO(3) on S 2 inversion is not yet implemented. C PCam: Extended Results [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: OrganMNIST3D test accuracy on the original and octahedrally rotated test sets. Error [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
read the original abstract

Many machine learning tasks are invariant under the action of a group $G$ of transformations: signal classification can be invariant under translations, image classification under 2D rotations, and spherical-image classification under 3D rotations. The $G$-bispectrum is a principled complete invariant of a signal (retaining all all signal's information up to the group action) with proven benefits in machine learning and as a pooling layer in deep networks. However, its deployment has been hampered by high computational cost and a patchwork of group-specific implementations. We present bispectrum, an open-source, fully unit-tested PyTorch library that implements selective $G$-bispectra for seven different group actions, as differentiable modules that can be directly incorporated into machine learning pipelines and deep learning architectures. For finite groups $G$, selectivity reduces the computational cost from $O(|G|^2)$ to $O(|G|)$. For planar rotations, we leverage the disk bispectrum. For spherical 3D rotations, we introduce an augmented selective bispectrum at band-limit $L$ which reduces the cost from $O(L^3)$ to $\Theta(L^2)$ coefficients. We profile the entire library (for which we implemented various compute optimizations), showing that it delivers near-exact $G$-invariance with its selective $G$-bispectra computed in sub-millisecond time on GPU (up to commonly used bandlimits). We evaluate the benefits of incorporating $G$-bispectra as pooling layers into deep learning architectures on three classical benchmark datasets --comparing against norm pooling, gated pooling, Fourier-ELU pooling, max pooling, and (non-equivariant) data-augmented convolutional baselines. Results show that $G$-bispectra consistently outperform alternatives in the low-data, moderate-capacity regime.

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 / 2 minor

Summary. The manuscript introduces the 'bispectrum' open-source PyTorch library implementing selective G-bispectra as differentiable modules for seven group actions. For finite groups it reduces cost from O(|G|^2) to O(|G|), for spherical rotations it introduces an augmented selective version reducing from O(L^3) to Θ(L^2); the library is profiled for sub-millisecond GPU times with near-exact invariance and evaluated as pooling layers, showing consistent outperformance versus norm/gated/Fourier-ELU/max pooling and data-augmented CNN baselines specifically in the low-data, moderate-capacity regime on three benchmarks.

Significance. If the selective and augmented constructions preserve the completeness and invariance of the full G-bispectrum, the library would make principled invariant representations practically usable in ML pipelines, with particular value in data-scarce settings. The unit-tested multi-group implementation with compute optimizations is a concrete strength for reproducibility.

major comments (2)
  1. [Abstract and selective construction] Abstract and selective-bispectrum sections: the claim that the O(|G|) selective coefficients (finite groups) and Θ(L^2) augmented spherical version remain complete invariants is not supported by a dimension count, orbit-separation argument, or proof; only empirical near-exact invariance after profiling is reported. This is load-bearing for attributing benchmark gains to the invariant construction rather than implementation details.
  2. [Experiments] Benchmark evaluation: the reported outperformance in the low-data regime lacks visible error bars, number of runs, statistical tests, or precise dataset sizes, weakening the strength of the central empirical claim.
minor comments (2)
  1. [Spherical implementation] Clarify in the spherical case whether the augmented selective bispectrum is defined via an explicit equation or only described procedurally.
  2. [Introduction] The abstract states 'proven benefits' for the full G-bispectrum; add a brief reference to the relevant prior completeness theorem for context.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful and constructive review. We appreciate the positive assessment of the library's practical contributions and reproducibility. Below we respond point-by-point to the major comments and indicate the revisions made to address them.

read point-by-point responses

  1. Referee: [Abstract and selective construction] Abstract and selective-bispectrum sections: the claim that the O(|G|) selective coefficients (finite groups) and Θ(L^2) augmented spherical version remain complete invariants is not supported by a dimension count, orbit-separation argument, or proof; only empirical near-exact invariance after profiling is reported. This is load-bearing for attributing benchmark gains to the invariant construction rather than implementation details.

    Authors: We thank the referee for highlighting this important point. The completeness of the standard (non-selective) G-bispectrum follows from classical results in bispectral analysis. For the selective finite-group case, the O(|G|) coefficients are obtained by retaining only the interactions with a fixed reference element, which still separates orbits because the selection corresponds to a generating set of the group algebra that preserves the necessary phase information. For the augmented spherical case, the Θ(L^2) coefficients combine the disk bispectrum with an additional radial augmentation that restores the missing degrees of freedom. We acknowledge that an explicit dimension count and orbit-separation argument were not present in the submitted manuscript. In the revised version we have added a dedicated subsection (Section 3.3) that supplies (i) a dimension count for finite groups showing that the number of retained coefficients equals the dimension of the space of orbit-separating functions, and (ii) a short orbit-separation argument for the augmented spherical construction based on the injectivity of the augmented map. These additions allow the benchmark gains to be attributed to the invariant construction with greater rigor. revision: yes

  2. Referee: [Experiments] Benchmark evaluation: the reported outperformance in the low-data regime lacks visible error bars, number of runs, statistical tests, or precise dataset sizes, weakening the strength of the central empirical claim.

    Authors: We agree that the experimental presentation would be strengthened by explicit statistical reporting. The original runs were performed with multiple random seeds, but these details were omitted from the text and figures. In the revised manuscript we now report: (a) exact low-data regime sizes (500 samples for the first benchmark, 1 000 for the second, and 2 000 for the third), (b) error bars showing mean ± one standard deviation over five independent runs with different seeds, and (c) results of a Wilcoxon signed-rank test confirming that the bispectrum pooling layer outperforms each baseline with p < 0.05 in the low-data, moderate-capacity regime. The updated figures and experimental section incorporate these changes. revision: yes

Circularity Check

0 steps flagged

No circularity: implementation library with external empirical comparisons.

full rationale

The paper describes a PyTorch library implementing selective G-bispectra for various groups, with profiling for speed and invariance, plus benchmark evaluations against norm/gated/Fourier-ELU/max pooling and data-augmented CNNs. No derivation chain, equations, or predictions appear; claims rest on code implementation and direct comparisons to independent baselines. No self-citations, fitted parameters renamed as predictions, or completeness proofs that reduce to inputs are present. The selective completeness question is a correctness/verification issue, not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper introduces no new free parameters, axioms beyond standard group representation theory, or invented entities; it implements and optimizes prior mathematical concepts with selectivity and engineering choices.

axioms (1)
  • domain assumption G-bispectrum is a complete invariant under the group action
    Invoked to justify that selective versions retain all signal information up to the group action.

pith-pipeline@v0.9.0 · 5628 in / 1265 out tokens · 47946 ms · 2026-05-11T02:19:22.926951+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

  1. [1]

    arXiv preprint arXiv:2511.19706 , year=

    The Selective Disk Bispectrum and Its Inversion, with Application to Multi-Reference Alignment , author=. arXiv preprint arXiv:2511.19706 , year=

    work page arXiv

  2. [2]

    The Selective

    Mataigne, Simon and Mathe, Johan and Sanborn, Sophia and Hillar, Christopher and Miolane, Nina , booktitle=. The Selective. 2024 , volume=. doi:10.52202/079017-3674 , url=

    work page doi:10.52202/079017-3674 2024

  3. [3]

    Triple Correlation on Groups , author=

    work page

  4. [4]

    arXiv preprint arXiv:1712.10163 , year=

    Estimation under group actions: recovering orbits from invariants , author=. arXiv preprint arXiv:1712.10163 , year=

    work page arXiv

  5. [5]

    The Bispectrum as a Source of Phase-Sensitive Invariants for

    Kakarala, Ramakrishna , journal=. The Bispectrum as a Source of Phase-Sensitive Invariants for. 2012 , publisher=. doi:10.1007/s10851-012-0330-6 , url=

    work page doi:10.1007/s10851-012-0330-6 2012

  6. [6]

    International Conference on Machine Learning (ICML) , pages=

    Group Equivariant Convolutional Networks , author=. International Conference on Machine Learning (ICML) , pages=

    work page

  7. [7]

    Weiler, Maurice and Cesa, Gabriele , booktitle=. General

    work page

  8. [8]

    International Conference on Learning Representations (ICLR) , year=

    Bispectral Neural Networks , author=. International Conference on Learning Representations (ICLR) , year=

    work page

  9. [9]

    and Depeursinge, Adrien , booktitle=

    Oreiller, Valentin and Fageot, Julien and Andrearczyk, Vincent and Prior, John O. and Depeursinge, Adrien , booktitle=. Robust Multi-Organ Nucleus Segmentation Using a Locally Rotation Invariant Bispectral. 2022 , volume=

    work page 2022

  10. [10]

    and Linmans, Jasper and Winkens, Jim and Cohen, Taco and Welling, Max , booktitle=

    Veeling, Bastiaan S. and Linmans, Jasper and Winkens, Jim and Cohen, Taco and Welling, Max , booktitle=. Rotation Equivariant. 2018 , publisher=

    work page 2018

  11. [11]

    Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for

    Thomas, Nathaniel and Smidt, Tess and Kearnes, Steven and Yang, Lusann and Li, Li and Kohlhoff, Kai and Riley, Patrick , journal=. Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for. 2018 , url=

    work page 2018

  12. [12]

    Advances in Neural Information Processing Systems (NeurIPS) , pages=

    Batatia, Ilyes and Cs. Advances in Neural Information Processing Systems (NeurIPS) , pages=. 2022 , volume=. doi:10.52202/068431-0830 , url=

    work page doi:10.52202/068431-0830 2022

  13. [13]

    Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and Killeen, Trevor and Lin, Zeming and Gimelshein, Natalia and Antiga, Luca and Desmaison, Alban and K. Py. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  14. [14]

    and Andrearczyk, Vincent and Depeursinge, Adrien , booktitle=

    Chevalley, Arthur and Oreiller, Valentin and Fageot, Julien and Prior, John O. and Andrearczyk, Vincent and Depeursinge, Adrien , booktitle=. A Bispectral. 2024 , publisher=. doi:10.1007/978-3-031-73967-5_5 , url=

    work page doi:10.1007/978-3-031-73967-5_5 2024

  15. [15]

    Proceedings of the 42nd International Conference on Machine Learning (ICML) , pages=

    The Price of Freedom: Exploring Expressivity and Runtime Tradeoffs in Equivariant Tensor Products , author=. Proceedings of the 42nd International Conference on Machine Learning (ICML) , pages=. 2025 , volume=

    work page 2025

  16. [16]

    and Geiger, Mario and K

    Cohen, Taco S. and Geiger, Mario and K. Spherical. International Conference on Learning Representations (ICLR) , year=

    work page

  17. [17]

    SIAM Journal on Applied Algebra and Geometry , volume=

    Orbit Recovery for Band-Limited Functions , author=. SIAM Journal on Applied Algebra and Geometry , volume=. 2024 , publisher=. doi:10.1137/23M1577808 , url=

    work page doi:10.1137/23m1577808 2024

  18. [18]

    2023 , publisher=

    Yang, Jiancheng and Shi, Rui and Wei, Donglai and Liu, Zequan and Zhao, Lin and Ke, Bilian and Pfister, Hanspeter and Ni, Bingbing , journal=. 2023 , publisher=

    work page 2023

  19. [19]

    and Bekkers, Erik J

    Kuipers, Thijs P. and Bekkers, Erik J. , booktitle=. Regular. 2023 , publisher=. doi:10.1007/978-3-031-43898-1_25 , url=

    work page doi:10.1007/978-3-031-43898-1_25 2023

  20. [20]

    Moving Frame Net:

    Sangalli, Mateus and Blusseau, Samy and Velasco-Forero, Santiago and Angulo, Jes. Moving Frame Net:. Proceedings of the 1st NeurIPS Workshop on Symmetry and Geometry in Neural Representations , pages=. 2023 , volume=

    work page 2023

  21. [21]

    2024 , publisher=

    Zhemchuzhnikov, Dmitrii and Grudinin, Sergei , booktitle=. 2024 , publisher=. doi:10.1007/978-3-031-70359-1_21 , url=

    work page doi:10.1007/978-3-031-70359-1_21 2024

  22. [22]

    Zhemchuzhnikov, Dmitrii and Grudinin, Sergei , booktitle=. On the. 2025 , url=

    work page 2025

  23. [23]

    JAMA , volume=

    Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer , author=. JAMA , volume=. 2017 , doi=

    work page 2017

  24. [24]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Densely Connected Convolutional Networks , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    work page

  25. [25]

    International Conference on Learning Representations (ICLR) , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations (ICLR) , year=

    work page

  26. [26]

    A Program to Build

    Cesa, Gabriele and Lang, Leon and Weiler, Maurice , booktitle=. A Program to Build. 2022 , url=

    work page 2022

  27. [27]

    Geiger, Mario and Smidt, Tess , journal=. e3nn:. 2022 , url=

    work page 2022

  28. [28]

    Learning

    Esteves, Carlos and Allen-Blanchette, Christine and Makadia, Ameesh and Daniilidis, Kostas , booktitle=. Learning. 2018 , publisher=

    work page 2018

  29. [29]

    Communications on Pure and Applied Mathematics , volume=

    Group Invariant Scattering , author=. Communications on Pure and Applied Mathematics , volume=. 2012 , publisher=

    work page 2012

  30. [30]

    Orbit recovery for spherical functions

    Orbit Recovery for Spherical Functions , author=. arXiv preprint arXiv:2508.02674 , year=

    work page arXiv

  31. [31]

    and Millman, K

    Harris, Charles R. and Millman, K. Jarrod and van der Walt, St. Array programming with. Nature , volume=. 2020 , publisher=

    work page 2020

  32. [32]

    Spherical

    Bonev, Boris and Kurth, Thorsten and Hundt, Christian and Pathak, Jaideep and Balu, Maximilian and Kashinath, Karthik and Anandkumar, Animashree , booktitle=. Spherical

    work page

  33. [33]

    and Pellegrini, Franziska and Jurhar, Tin and Nguyen, Tien D

    Binns, Thomas S. and Pellegrini, Franziska and Jurhar, Tin and Nguyen, Tien D. and K. Journal of Open Source Software , volume=. 2025 , publisher=. doi:10.21105/joss.08504 , url=

    work page doi:10.21105/joss.08504 2025

  34. [34]

    A General Framework for Equivariant Neural Networks on Reductive

    Batatia, Ilyes and Geiger, Mario and Munoz, Jose and Smidt, Tess and Silberman, Lior and Ortner, Christoph , year=. A General Framework for Equivariant Neural Networks on Reductive. 2306.00091 , archivePrefix=

    work page arXiv

  35. [35]

    Journal of the Optical Society of America A , volume=

    Signal reconstruction from multiple correlations: frequency- and time-domain approaches , author=. Journal of the Optical Society of America A , volume=. 1989 , month=. doi:10.1364/JOSAA.6.000682 , url=

    work page doi:10.1364/josaa.6.000682 1989

  36. [36]

    Efficient computation of the

    Diaconis, Persi and Rockmore, Daniel , journal=. Efficient computation of the. 1990 , month=. doi:10.1090/S0894-0347-1990-1030655-4 , url=

    work page doi:10.1090/s0894-0347-1990-1030655-4 1990

  37. [37]

    Kondor, A novel set of rotationally and transla- tionally invariant features for images based on the non-commutative bispectrum (2007), arXiv:cs/0701127 [cs.CV]

    A novel set of rotationally and translationally invariant features for images based on the non-commutative bispectrum , author=. arXiv preprint cs/0701127 , year=

    work page arXiv

  38. [38]

    Rotationally invariant image representation for viewing direction classification in cryo-

    Zhao, Zhizhen and Singer, Amit , journal=. Rotationally invariant image representation for viewing direction classification in cryo-. 2014 , publisher=

    work page 2014