arxiv: 2605.15073 · v1 · submitted 2026-05-14 · ⚛️ physics.comp-ph · cond-mat.mtrl-sci· physics.chem-ph

Recognition: 1 theorem link

· Lean Theorem

Fast contracted Clebsch--Gordan tensor products for equivariant graph neural networks

Anton Bochkarev , Yury Lysogorskiy , Ralf Drautz

Authors on Pith no claims yet

Pith reviewed 2026-05-15 02:59 UTC · model grok-4.3

classification ⚛️ physics.comp-ph cond-mat.mtrl-sciphysics.chem-ph

keywords Clebsch-Gordan tensor productsequivariant graph neural networksO(3) equivariancemachine learning potentialsCanonical Polyadic decompositionsurface-curl pairingGauss-Legendre quadratureatomic cluster expansion

0 comments

The pith

An O(L^3) algorithm evaluates contracted Clebsch-Gordan tensor products for O(3)-equivariant machine learning potentials using a structured grid and surface-curl pairing.

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

The paper develops an efficient method to compute contracted Clebsch-Gordan tensor products that are central to O(3)-equivariant graph neural networks and machine learning potentials. It achieves O(L^3) scaling at fixed Canonical Polyadic rank by mapping angular integrals onto a structured Gauss-Legendre and Fourier tensor-product grid. This decouples radial channel contractions from the angular transforms. Antisymmetric parity-odd channels are recovered using the surface-curl pairing, which acts as a spherical Poisson bracket to supply the required L=1 angular momentum while preserving rotational equivariance. The approach supports parity-aware message passing in atomic-cluster-expansion architectures and is confirmed through direct numerical quadrature, with benchmarks indicating practical L^2 scaling that becomes L^3 at large l_max.

Core claim

Contracted Clebsch-Gordan tensor products can be evaluated in O(L^3) operations by decoupling radial and angular parts on a Gauss-Legendre Fourier grid and recovering parity-odd channels via the surface-curl pairing. This extends to parity-aware equivariant message passing while the uncontracted product is limited by an O(L^4) output-size bound.

What carries the argument

The structured Gauss-Legendre and Fourier tensor-product grid combined with the surface-curl pairing, which decouples radial contractions from angular transforms and supplies the L=1 momentum for antisymmetric channels.

Load-bearing premise

The structured Gauss-Legendre and Fourier grid combined with the surface-curl pairing accurately recovers all required Clebsch-Gordan channels, including antisymmetric parity-odd ones, without introducing numerical errors or breaking equivariance for the target L range.

What would settle it

A calculation showing that the surface-curl pairing fails to produce the correct parity-odd Clebsch-Gordan coefficients for some L value, or that rotated inputs lose equivariance in the contracted products, would disprove the algorithm's correctness.

Figures

Figures reproduced from arXiv: 2605.15073 by Anton Bochkarev, Ralf Drautz, Yury Lysogorskiy.

read the original abstract

We present an $\mathcal{O}(L^3)$ algorithm for evaluating contracted Clebsch--Gordan tensor products in $\mathrm{O}(3)$-equivariant machine learning potentials at fixed Canonical Polyadic (CP) rank. Mapping the angular integral to a structured Gauss--Legendre and Fourier tensor-product grid decouples the radial channel contractions from the angular transforms. The antisymmetric parity-odd Clebsch--Gordan channels, unreachable by the symmetric pointwise product on a scalar $S^2$ grid, are recovered through the surface-curl pairing $\hat r \cdot [\nabla_{S^2} A \times \nabla_{S^2} B]$, the spherical Poisson bracket, which supplies the $L=1$ angular momentum on the grid while preserving rotational equivariance. The construction extends to parity-aware equivariant message passing in atomic-cluster-expansion-style architectures and is verified by direct numerical quadrature. The full uncontracted Clebsch--Gordan tensor product remains subject to the $\mathcal{O}(L^4)$ output-size lower bound. A benchmark shows wall-clock scaling empirically as $L^2$ across the practical $l_{\max}$ range. For the on-site contraction this is pre-asymptotic, giving way to $L^3$ at large $l_{\max}$. For message passing it is structural and the runtime is memory-bandwidth bound on $L^2$-sized grid tensors.

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.

Desk Editor's Note private letter to a colleague

The paper gives a grid-plus-surface-curl method that achieves practical O(L^3) contracted CG products for equivariant GNNs, with numerical checks supporting the scaling.

read the letter

The main thing here is a grid-based algorithm that evaluates contracted Clebsch-Gordan tensor products in O(L^3) time for O(3)-equivariant machine learning potentials. It maps the angular part to a Gauss-Legendre and Fourier grid to separate radial contractions, then uses the spherical Poisson bracket (surface curl) to recover the parity-odd antisymmetric channels that a direct pointwise product cannot reach on the sphere. This is the concrete new piece: the specific pairing that keeps rotational equivariance while staying on the grid. The construction is then extended to message passing in atomic-cluster-expansion architectures. The authors verify it against direct quadrature and show wall-clock scaling that tracks L^2 across typical l_max values, which is already better than the O(L^4) output-size bound for the uncontracted tensor. For on-site work the L^3 regime is pre-asymptotic; for message passing the cost is memory-bandwidth limited on the L^2 grid tensors. That distinction is stated clearly. The soft spot is the reliance on quadrature accuracy. Any truncation error in the discrete curl or grid integration would leave a small mismatch with the exact CG coefficients, and the paper does not supply explicit L-dependent error bounds or tests of accumulated equivariance drift in deep stacks. The abstract says the match holds in their checks, so the deviation is small enough for the reported cases, but a referee would want to see how that error behaves at the upper end of practical L. This is aimed at people who code or optimize equivariant potentials for atomic-scale modeling and hit tensor-product cost limits. A reader who needs faster contractions for larger systems or deeper message passing will find usable implementation details and scaling numbers. I would send it for peer review. The algorithmic idea is self-contained, the verification is direct, and the scaling claim is falsifiable, so it merits referee time even if revisions are needed on the error analysis.

Referee Report

2 major / 2 minor

Summary. The paper claims to present an O(L^3) algorithm for evaluating contracted Clebsch-Gordan tensor products in O(3)-equivariant machine learning potentials at fixed CP rank. It maps the angular integral to a structured Gauss-Legendre/Fourier grid that decouples radial contractions from angular transforms, recovers antisymmetric parity-odd channels via the surface-curl pairing (spherical Poisson bracket), verifies the construction by direct numerical quadrature, and reports empirical L^2 wall-clock scaling (pre-asymptotic for on-site contraction, structural for message passing).

Significance. If the grid-based construction recovers all required CG channels to machine precision while preserving exact rotational equivariance, the method would remove a major computational bottleneck for high-order equivariant features in atomic-cluster-expansion and message-passing architectures, enabling practical use of larger l_max without sacrificing the theoretical guarantees of the underlying tensor products.

major comments (2)

[Abstract] Abstract: the claim that the surface-curl pairing 'supplies the L=1 angular momentum on the grid while preserving rotational equivariance' is load-bearing for the central O(L^3) result; the manuscript must supply explicit quadrature-error bounds (or machine-precision agreement metrics) versus l_max and grid density for the parity-odd antisymmetric channels, because any L-dependent truncation would make the output only approximately equivariant.
[Benchmark] Benchmark paragraph: the reported empirical L^2 scaling is stated to be pre-asymptotic for the on-site contraction and to give way to L^3 only at large l_max; the manuscript should include timing data at sufficiently high l_max to demonstrate the transition and confirm that the asymptotic O(L^3) regime is attained within the target range of the method.

minor comments (2)

[Abstract] Abstract: the final sentence notes that the uncontracted CG tensor product remains subject to the O(L^4) output-size lower bound; this distinction should be stated more prominently when the complexity claim is introduced.
Notation: the acronym 'CP rank' appears without expansion on first use; a brief parenthetical definition would improve readability for readers outside the tensor-decomposition literature.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment and the recommendation for minor revision. The two major comments are addressed point-by-point below; both requested additions are feasible and will be incorporated in the revised manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the surface-curl pairing 'supplies the L=1 angular momentum on the grid while preserving rotational equivariance' is load-bearing for the central O(L^3) result; the manuscript must supply explicit quadrature-error bounds (or machine-precision agreement metrics) versus l_max and grid density for the parity-odd antisymmetric channels, because any L-dependent truncation would make the output only approximately equivariant.

Authors: We agree that explicit numerical verification of quadrature accuracy for the parity-odd channels is required to substantiate the exact-equivariance claim. In the revised manuscript we will add a dedicated panel (or table) that reports the maximum absolute error of the recovered L=1 components obtained via the spherical Poisson bracket, plotted against l_max (up to at least 12) and grid density (number of Gauss-Legendre nodes and Fourier modes). The data will demonstrate that the error remains at machine precision (typically < 1e-14) with no visible growth in l_max, confirming that the surface-curl construction introduces no L-dependent truncation within double precision. revision: yes
Referee: [Benchmark] Benchmark paragraph: the reported empirical L^2 scaling is stated to be pre-asymptotic for the on-site contraction and to give way to L^3 only at large l_max; the manuscript should include timing data at sufficiently high l_max to demonstrate the transition and confirm that the asymptotic O(L^3) regime is attained within the target range of the method.

Authors: We acknowledge that the present benchmark only captures the pre-asymptotic L^2 regime. To demonstrate the crossover, we will extend the wall-clock timing experiments to higher l_max values (l_max = 16–20, limited only by available GPU memory) and include the corresponding scaling plot in the revised manuscript. The new data will show the transition from L^2 to the expected L^3 regime for the on-site contraction, while the message-passing timings remain memory-bandwidth bound as stated. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation uses standard quadrature and operators

full rationale

The claimed O(L^3) algorithm is obtained by mapping the angular integral to a fixed Gauss-Legendre/Fourier tensor-product grid whose size scales as L^2, then performing radial contractions separately and recovering parity-odd channels via the surface-curl (spherical Poisson bracket) operator. These steps rest on classical quadrature rules and vector-calculus identities on the sphere, not on any parameter fitted inside the paper or on a self-referential definition of the CG coefficients. The paper states that the construction is verified by direct numerical quadrature, supplying an independent check. No load-bearing equation reduces to a prior result by the same authors, no ansatz is smuggled via citation, and the complexity bound follows directly from the grid cardinality and contraction ordering rather than from any tautological renaming or uniqueness theorem imported from the authors' own work. The result is therefore self-contained against external mathematical benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard properties of spherical harmonics, Clebsch-Gordan coefficients, and numerical quadrature; no free parameters, new entities, or ad-hoc axioms are introduced in the abstract description.

axioms (2)

standard math Standard transformation properties of Clebsch-Gordan coefficients under O(3) rotations and parity
Invoked to guarantee equivariance of the final contraction.
domain assumption Sufficient accuracy of Gauss-Legendre and Fourier quadrature for angular integrals up to given L
Assumed to hold for the chosen grid without further quantification in the abstract.

pith-pipeline@v0.9.0 · 5575 in / 1463 out tokens · 56040 ms · 2026-05-15T02:59:51.677938+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Mapping the angular integral to a structured Gauss–Legendre and Fourier tensor-product grid... surface-curl pairing ˆr·[∇S²A×∇S²B], the spherical Poisson bracket

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

55 extracted references · 55 canonical work pages · 5 internal anchors

[1]

Parity-odd identity recovery For each parity-odd triple (l1, l2, l) we evaluate Eq.(15) on every (m 1, m2, m) entry withm=m 1 +m 2 and ex- tractκ −1(l1, l2, l) = integral/G lm l1m1 l2m2 from the non- vanishing entries. Three properties are checked simul- taneously: (i)κ −1 is purely imaginary in the complex basis, (ii)κ −1 is independent of (m 1, m2, m), ...

work page
[2]

Across all parity-even triples up to (4,4,8), the maximum integral was∼10 −10, set by the quadrature

Parity-even vanishing The same quadrature on parity-even triples (l 1 +l 2 +l even) returns zero to numerical precision, since the parity-even CG channels are symmetric under 1↔2 while the surface-curl integrand is antisymmetric and the two cannot mix. Across all parity-even triples up to (4,4,8), the maximum integral was∼10 −10, set by the quadrature

work page
[3]

Antisymmetry under1↔2exchange The surface-curl integrand ˆr·[∇ S2 Yl1m1 × ∇S2 Yl2m2] flips sign exactly under exchange of the two factors. Di- rect evaluation across all parity-odd triples up to (4,4,7) confirms this to machine precision, with the absolute value of the sum integral 1,2 + integral2,1 consistently be- low 10 −17, i.e. at the level of double...

work page
[4]

End-to-end consistency on random synthetic fields To verify that the surface-curl construction reproduces the contracted CG output not only on plain spheri- cal harmonics but on arbitrary input fields, we gen- erate independent random complex-Gaussian SH coef- ficientsA (1) l1m1 ,A (2) l2m2 , build the corresponding fields A(1)(ˆr), A(2)(ˆr) on the grid, ...

work page
[5]

Φ(sym/anti) lm = X n1,n2 M (sym/anti) n1n2 × X m1m2 Glm l1m1 l2m2 An1l1m1 An2l2m2 14 withA nlm random Gaussian,l 1 =l 2,M sym symmetric, M anti antisymmetric

Self-coupled vanishing under symmetric channel-mixing To validate Eq.(27), we test the contracted parity-odd output with two channel matrices. Φ(sym/anti) lm = X n1,n2 M (sym/anti) n1n2 × X m1m2 Glm l1m1 l2m2 An1l1m1 An2l2m2 14 withA nlm random Gaussian,l 1 =l 2,M sym symmetric, M anti antisymmetric. Across all triples up to (3,3,5), the symmetric-mixing ...

work page
[6]

Parity-indexed message-passing coupling [Eq.(2)] We apply the surface-curl to the parity-indexed cou- pling Eq.(2), comparing (i) the direct CG sum with ex- plicit slot rulep 2 =p(−1) l1 and Category-α/βdispatch, and (ii) the structured-grid pipeline of Sec. VII C. The setup uses one bond at fixed (θ e, ϕe), one radial channel, lmax = 2, randomR nl(rji), ...

work page
[7]

Both are real- valued at every entry, soκ −1 real(l1, l2, l) =Q real/Greal is real with magnitude unitary-invariant,|κ −1 real(l1, l2, l)|= |κ−1 complex(l1, l2, l)|

Real-basisκ −1 extraction and basis consistency We repeat theκ −1-extraction in the real basis Eq.(C1), withG real andQ real defined directly by inte- gration with real spherical harmonics. Both are real- valued at every entry, soκ −1 real(l1, l2, l) =Q real/Greal is real with magnitude unitary-invariant,|κ −1 real(l1, l2, l)|= |κ−1 complex(l1, l2, l)|. A...

work page 1968
[8]

In the Lebedev pipeline, ˜w klm =w k y∗ klm reduces to ˜wklm =w k yklm and all tensors are real

work page
[9]

The rest is un- changed

In the structured-grid pipeline, the forward and in- verse Fourier transforms in Eqs.(37) and (39) be- come real cosine/sine transforms. The rest is un- changed

work page
[10]

The surface-curl construction Eq.(57) uses only point- wise vector operations and is basis-independent, with∂ ϕ acting onq vm as a real derivative ( √ 2 cos(mϕ)↔ −m √ 2 sin(mϕ))

In the parity-aware pipeline, the same applies. The surface-curl construction Eq.(57) uses only point- wise vector operations and is basis-independent, with∂ ϕ acting onq vm as a real derivative ( √ 2 cos(mϕ)↔ −m √ 2 sin(mϕ)). Them↔ −m branching of the real basis requires care in the im- plementation.∂ ϕ maps them >0 cosine slot to the m <0 sine slot (and...

work page
[11]

Drautz, Phys

R. Drautz, Phys. Rev. B99, 014104 (2019), erratum: Phys. Rev. B100, 249901 (2019)

work page 2019
[12]

T. S. Cohen and M. Welling, inProceedings of the 33rd International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, Vol. 48 (2016) pp. 2990–2999, arXiv:1602.07576

work page internal anchor Pith review Pith/arXiv arXiv 2016
[13]

3D Steerable CNNs: Learning Rotationally Equivariant Features in Volumetric Data

M. Weiler, M. Geiger, M. Welling, W. Boomsma, and T. S. Cohen, inAdvances in Neural Information Process- ing Systems, Vol. 31 (2018) arXiv:1807.02547

work page internal anchor Pith review Pith/arXiv arXiv 2018
[14]

Anderson, T.-S

B. Anderson, T.-S. Hy, and R. Kondor, inAdvances in Neural Information Processing Systems, Vol. 32 (2019) pp. 14537–14546, arXiv:1906.04015

work page arXiv 2019
[15]

K. T. Sch¨ utt, O. T. Unke, and M. Gastegger, inProceedings of the 38th International Conference on Machine Learning (ICML)(2021) pp. 9377–9388, arXiv:2102.03150

work page arXiv 2021
[16]

Brandstetter, R

J. Brandstetter, R. Hesselink, E. van der Pol, E. J. Bekkers, and M. Welling, inInternational Con- ference on Learning Representations (ICLR)(2022) arXiv:2110.02905

work page arXiv 2022
[17]

Batzner, A

S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nat. Commun.13, 2453 (2022)

work page 2022
[18]

Batatia, D

I. Batatia, D. P. Kov´ acs, G. N. C. Simm, C. Ort- ner, and G. Cs´ anyi, inAdvances in Neural Informa- tion Processing Systems, Vol. 35 (2022) pp. 11423–11436, arXiv:2206.07697

work page arXiv 2022
[19]

Liao and T

Y.-L. Liao and T. E. Smidt, inInternational Con- ference on Learning Representations (ICLR)(2023) arXiv:2206.11990

work page arXiv 2023
[20]

Musaelian, S

A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen, M. Kornbluth, and B. Kozinsky, Nat. Commun. 14, 579 (2023)

work page 2023
[21]

Bochkarev, Y

A. Bochkarev, Y. Lysogorskiy, and R. Drautz, Phys. Rev. X14, 021036 (2024)

work page 2024
[22]

Y.-L. Liao, B. Wood, A. Das, and T. E. Smidt, inInterna- tional Conference on Learning Representations (ICLR) (2024) arXiv:2306.12059

work page arXiv 2024
[23]

X. Fu, B. M. Wood, L. Barroso-Luque, D. S. Levine, M. Gao, M. Dzamba, and C. L. Zitnick, inProceedings of the 42nd International Conference on Machine Learning (ICML)(2025) arXiv:2502.12147

work page arXiv 2025
[24]

Batatia, S

I. Batatia, S. Batzner, D. P. Kov´ acs, A. Musaelian, G. N. C. Simm, R. Drautz, C. Ortner, B. Kozinsky, and G. Cs´ anyi, Nat. Mach. Intell.7, 56 (2025)

work page 2025
[25]

Y.-L. Liao, A. J. Hoffman, S. C. Shen, A. Duval, S. W. Norwood, and T. Smidt, Equiformerv3: Scaling efficient, expressive, and general se(3)-equivariant graph attention transformers (2026), arXiv:2604.09130 [cs.LG]

work page internal anchor Pith review Pith/arXiv arXiv 2026
[26]

Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

N. Thomas, T. Smidt, S. Kearnes, L. Yang, L. Li, K. Kohlhoff, and P. Riley, Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds (2018), arXiv:1802.08219 [cs.LG]

work page internal anchor Pith review Pith/arXiv arXiv 2018
[27]

Geiger and T

M. Geiger and T. E. Smidt, e3nn: Euclidean neural net- works (2022), arXiv:2207.09453 [cs.LG]

work page arXiv 2022
[28]

O. T. Unke and H. Maennel, E3x: E(3)-equivariant deep learning made easy (2024), arXiv:2401.07595 [cs.LG]

work page arXiv 2024
[29]

Lysogorskiy, C

Y. Lysogorskiy, C. van der Oord, A. Bochkarev, S. Menon, M. Rinaldi, T. Hammerschmidt, M. Mrovec, A. Thompson, G. Cs´ anyi, C. Ortner, and R. Drautz, npj Comput. Mater.7, 97 (2021)

work page 2021
[30]

Bochkarev, Y

A. Bochkarev, Y. Lysogorskiy, S. Menon, M. Qamar, M. Mrovec, and R. Drautz, Phys. Rev. Materials6, 013804 (2022)

work page 2022
[31]

readthedocs.io/(2024)

GRACEmaker documentation,https://gracemaker. readthedocs.io/(2024)

work page 2024
[32]

Lysogorskiy, A

Y. Lysogorskiy, A. Bochkarev, and R. Drautz, npj Com- put. Mater.12, 114 (2026)

work page 2026
[33]

W. C. Witt, C. van der Oord, E. Gelˇ zinyt˙ e, T. J¨ arvinen, A. Ross, J. P. Darby, C. H. Ho, W. J. Baldwin, M. Sachs, J. Kermode, N. Bernstein, G. Cs´ anyi, and C. Ortner, J. Chem. Phys.159, 164101 (2023)

work page 2023
[34]

com/NVIDIA/cuEquivariance(2025)

NVIDIA Corporation, cuEquivariance: CUDA- accelerated equivariant operations,https://github. com/NVIDIA/cuEquivariance(2025)

work page 2025
[35]

J. R. Driscoll and D. M. Healy, Adv. Appl. Math.15, 202 (1994)

work page 1994
[36]

T. S. Cohen, M. Geiger, J. K¨ ohler, and M. Welling, inInternational Conference on Learning Representations (ICLR)(2018) best Paper Award, arXiv:1801.10130

work page internal anchor Pith review Pith/arXiv arXiv 2018
[37]

Passaro and C

S. Passaro and C. L. Zitnick, inProceedings of the 40th International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, Vol. 202 (2023) pp. 27420–27438, arXiv:2302.03655. 16

work page arXiv 2023
[38]

S. Luo, T. Chen, and A. S. Krishnapriyan, inThe Twelfth International Conference on Learning Representations (ICLR)(2024) spotlight, arXiv:2401.10216

work page arXiv 2024
[39]

Liang, Y

Z. Liang, Y. Wang, C. Ding, J. Wang, H.-T. Wang, D. Xing, and J. Sun, Chin. Phys. Lett.43, 020704 (2026)

work page 2026
[40]

Y. Lin, C. Fu, Z. Krueger, H. Yu, M. Nakata, J. Xie, E. Kucukbenli, X. Qian, and S. Ji, inAdvances in Neu- ral Information Processing Systems (NeurIPS)(2025) arXiv:2507.01131

work page arXiv 2025
[41]

Klinteb¨ ack, C

H. Klinteb¨ ack, C. Ortner, and L. Silberman, The high cost of data augmentation for learning equivariant models (2026), arXiv:2602.03118 [math.NA]

work page arXiv 2026
[42]

F. B. Fuchs, D. E. Worrall, V. Fischer, and M. Welling, inAdvances in Neural Information Processing Systems, Vol. 33 (2020) pp. 1970–1981, arXiv:2006.10503

work page arXiv 2020
[43]

Moskalev, M

A. Moskalev, M. Prakash, R. Liao, and T. Mansi, SE(3)-Hyena Operator for Scalable Equivariant Learning (2024), arXiv:2407.01049

work page arXiv 2024
[44]

O. L. Howell, L. Zhao, X. Zhu, Y. Qian, H. Huang, L. Sun, W. Thomason, R. Platt, and R. Walters, Clebsch–Gordan transformer: Fast and global equivari- ant attention (2025), arXiv:2509.24093 [cs.LG]

work page arXiv 2025
[45]

Y. Xie, A. Daigavane, M. Kotak, and T. E. Smidt, Asymptotically fast Clebsch–Gordan tensor products with vector spherical harmonics (2026), arXiv:2602.21466 [cs.LG]

work page arXiv 2026
[46]

Drautz, Phys

R. Drautz, Phys. Rev. B102, 024104 (2020)

work page 2020
[47]

Dusson, M

G. Dusson, M. Bachmayr, G. Cs´ anyi, R. Drautz, S. Etter, C. van der Oord, and C. Ortner, J. Comput. Phys.454, 110946 (2022)

work page 2022
[48]

Bochkarev, Y

A. Bochkarev, Y. Lysogorskiy, C. Ortner, G. Cs´ anyi, and R. Drautz, Phys. Rev. Research4, L042019 (2022)

work page 2022
[49]

J. P. Darby, D. P. Kov´ acs, I. Batatia, M. A. Caro, G. L. W. Hart, C. Ortner, and G. Cs´ anyi, Phys. Rev. Lett.131, 028001 (2023)

work page 2023
[50]

D. A. Varshalovich, A. N. Moskalev, and V. K. Kher- sonskii,Quantum Theory of Angular Momentum(World Scientific, Singapore, 1988)

work page 1988
[51]

V. I. Lebedev, USSR Comput. Math. Math. Phys.16, 10 (1976)

work page 1976
[52]

Abadi, A

M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Is- ard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Lev- enberg, D. Man´ e, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Van...

work page 2015
[53]

E. T. Newman and R. Penrose, J. Math. Phys.7, 863 (1966)

work page 1966
[54]

J. N. Goldberg, A. J. Macfarlane, E. T. Newman, F. Rohrlich, and E. C. G. Sudarshan, J. Math. Phys. 8, 2155 (1967)

work page 1967
[55]

Meurer, C

A. Meurer, C. P. Smith, M. Paprocki, O. ˇCert´ ık, S. B. Kirpichev, M. Rocklin, A. Kumar, S. Ivanov, J. K. Moore, S. Singh, T. Rathnayake, S. Vig, B. E. Granger, R. P. Muller, F. Bonazzi, H. Gupta, S. Vats, F. Johans- son, F. Pedregosa, M. J. Curry, A. R. Terrel, ˇS. Rouˇ cka, A. Saboo, I. Fernando, S. Kulal, R. Cimrman, and A. Scopatz, PeerJ Comput. Sci....

work page 2017