High-Rank Irreducible Cartesian Tensor Decomposition and Bases of Equivariant Spaces

Qinghua Cui; Shihao Shao; Yikang Li; Zhouchen Lin

High-Rank Irreducible Cartesian Tensor Decomposition and Bases of Equivariant Spaces

T0 review · 0 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Path-matrix contractions yield orthogonal rank-9 tensor decompositions in minutes.

desk verdict A solid, genuinely useful construction of rank 6–9 ICT decompositions and orthogonal equivariant bases; the stress-test worry about Lemma 15 dissolves on close reading, though a few lemmas lean on unproved standard CG orthogonality. read the letter →

arxiv 2412.18263 v7 pith:JA4Y7OYU submitted 2024-12-24 cs.LG math-phmath.MPphysics.chem-phphysics.comp-phquant-ph

Shihao Shao , Yikang Li , Zhouchen Lin , Qinghua Cui This is my paper · ORCID

classification cs.LGmath-phmath.MPphysics.chem-phphysics.comp-phquant-ph MSC 15A6922E70

keywords irreducibleCartesiantensordecompositionClebsch-GordancoefficientsparentageschemeequivariantgraphneuralnetworksorthogonalbasisO(3)equivariancechangeofhigh-ranktensors

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that orthogonal decompositions of rank-n Cartesian tensors into irreducible components, previously known only through rank 5 with factorial cost, can be built for ranks 6 through 9 by chaining Clebsch-Gordan contractions along the parentage scheme. The construction produces path matrices whose concatenation is an orthonormal change of basis between the Cartesian tensor product space and a direct sum of spherical spaces, which makes the decomposition orthogonal automatically and removes the need for row-reduction algorithms. The paper further claims these same path matrices give an orthogonal basis of the equivariant linear design space, not just a spanning set, and extends the construction to arbitrary tensor product and direct-sum spaces. A reader should care because high-rank equivariant layers are what let equivariant graph neural networks capture many-body interactions without switching to spherical Clebsch-Gordan transforms at every step.

What carries the argument

The central object is the path matrix, built by contracting a growing tensor with Clebsch-Gordan matrices according to the parentage scheme: $P^{(0\to l_1\to\cdots\to l_k)}$ is O(3)-equivariant from $(l{=}1)^{\otimes k}$ to weight $l_k$, has mutually orthogonal columns, and has columns of equal L2 norm. Normalizing each path matrix and concatenating them yields the orthogonal change-of-basis matrix $D$ between $(\mathbb{R}^3)^{\otimes n}$ and the spherical direct sum; orthogonality makes the inverse equal to the transpose, which is what produces orthogonal ICT decomposition matrices without RREF.

What would settle it

Compute the concatenated matrix $D$ for rank 9 from the released code and evaluate $\|D^\top D - I\|$; alternatively, directly test the dot product of one column from each of two same-length path matrices that diverge at an earlier step. A single nonzero off-diagonal entry in $D^\top D$ would falsify the orthogonality claim on which both the ICT decomposition and the equivariant basis construction rest.

Watch

Extended reading notes

Core claim

Algorithm 1 constructs normalized path matrices by sequential contraction of a starting (0,0,0) Clebsch-Gordan tensor with (1,l_i,l_{i+1}) CG tensors following the parentage scheme. Theorem 17 states that each matrix $\hat{P}(p)(\hat{P}(p))^\top$ has rank $2l+1$, is O(3)-equivariant, and the full set over all paths of length $n+1$ forms an orthogonal ICT decomposition, because the concatenation $D$ of all normalized path matrices is orthogonal. Theorem 21 states that $\hat{P}(p_i)(\hat{P}(p_j))^\top$ over pairs of paths with the same final weight form a Frobenius-orthogonal basis of the equivariant endomorphism space, and Proposition 25 extends this to bases of equivariant maps between arbitrary input and output spaces. The paper reports obtaining the full decomposition matrices for n = 6, 7, 8, 9 in 1s, 3s, 11s, and 4m32s respectively.

Load-bearing premise

The whole construction collapses if the Clebsch-Gordan tensors used at any intermediate contraction stop having orthogonal columns of equal norm; the paper imports this as a standard property rather than proving it.

Editorial extensions

If this is right

ICT decomposition matrices for ranks 6 through 9 become available in seconds to minutes, extending the previous explicit rank-5 limit and enabling high-rank Cartesian equivariant layers.
Equivariant linear layers can use an orthogonal basis rather than a redundant spanning set, reducing the number of free parameters and removing the need for post-processing to extract a basis.
Equivariant maps between arbitrary tensor product and direct-sum spaces, including spherical tensor product spaces, can be built directly, allowing steerable layers beyond pure Cartesian inputs.
The basis dimension formula gives a direct count of the equivariant design space dimension for rank n, which can be used to size equivariant networks without constructing the full layer.
Because path matrices are derived analytically from CG coefficients, the same construction extends in principle to O(n), SO(n), SU(n), and U(n) wherever CG coefficients are available.

Reading between the lines

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

The reported timings suggest the practical bottleneck for n = 10 shifts from computation to storage of the full $3^{10}\times 3^{10}$ matrix set, so a natural extension is to generate only the subset of decomposition matrices for selected weights on demand.
The equal-column-norm property of path matrices implies a closed-form value for each column norm; using it explicitly could remove even the normalization pass and yield fully symbolic decomposition matrices.
The Frobenius-orthogonal basis could be used to derive principled parameter-sharing schemes or to measure the effective rank of equivariant layers by projecting weight matrices onto the basis.
The parity-matching condition in Proposition 25 suggests that equivariant bias and normalization layers for arbitrary spaces should keep a dedicated $l=0$ irreducible representation available, as the paper notes for practical EGNN design.

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.

Desk Editor's Note

read the letter

You should know this paper is the real thing: it gives the first explicit ICT decomposition matrices for ranks 6–9, via a path-matrix construction that avoids the factorial blow-up of Bonvicini and the RREF bottleneck. It also produces orthogonal bases for the equivariant design space, not just spanning sets, and extends the machinery to arbitrary tensor product spaces. That is an actual advance, not an incremental repackaging, and it is well positioned against Bonvicini (2024), Pearce-Crump (2023b), and Finzi et al. (2021). The core proofs, especially Theorem 17 and Theorem 21, are detailed and rest on standard CG orthogonality; the algorithm is simple and reproducible, with code released and timing benchmarks that match the claims.

Where are the soft spots? The stress-test concern about Lemma 15 does not hold up. The induction hypothesis is exactly that columns of different path matrices of the same length are orthogonal, so in the displayed sum over j2, j2', the factor with the path matrices is zero termwise, and the CG prefactors do not matter. The proof is fine there. The more legitimate gap is that Lemma 6, the 'general' orthogonality of CG tensors, is asserted without proof. It is standard (it follows from unitarity of CG couplings), but the paper does not cite or derive it, and a skeptical reader has to take it as an external fact. That is a minor issue, not a load-bearing flaw. Corollary 23's proof is terse, and the paper asserts squareness of D by dimension counting rather than showing the multiplicity count, but those are cosmetic. The paper also says it verified orthogonality of D only through visualizations, not a numerical residual; given Theorem 17 proves D^T D = I, that is a presentation nicety rather than a missing argument.

The citation pattern is honest and complete, and the paper's self-positioning against predecessors matches what I know. I did not run the code myself, but the repository is linked and the runtime numbers are concrete.

Verdict: this paper deserves a serious referee and is essentially acceptable. For revision, add a proof or a direct citation for Lemma 6, and include a line reporting the max |D^T D - I| for n = 6..9. Both are cheap. I would bring this to reading group and would cite it if I were working on high-rank Cartesian equivariant layers.

Referee Report

0 major / 7 minor

Summary. The paper develops a method for constructing orthogonal irreducible Cartesian tensor (ICT) decomposition matrices and orthogonal bases of equivariant linear spaces for O(3). The central construction is a family of "path matrices" formed by chaining Clebsch-Gordan contractions along the parentage scheme. The main theorems (Theorem 17 and Theorem 21) prove that the concatenation of normalized path matrices is an orthogonal change-of-basis matrix between (R^3)^⊗n and a direct sum of spherical irreducible representation spaces, and that products of the form P(pi)P(pj)^T with matching final weights form an orthogonal basis of the equivariant endomorphism space. The method is extended to arbitrary tensor product spaces through a generalized parentage scheme. Empirically, the authors report obtaining decomposition matrices for ranks n=6,...,9 in 1s, 3s, 11s, and 4m32s, respectively, with code provided.

Significance. If the results hold, this is a substantial advance for equivariant neural networks and theoretical chemistry: it extends explicit ICT decompositions from rank 5 to rank 9, avoids factorial-complexity RREF computations, delivers orthogonal decompositions (unlike several previous constructions), and gives orthogonal bases for equivariant linear layers rather than merely spanning sets. The paper's strengths include a fully analytical construction with reproducible code, explicit timing benchmarks, and a clear extension to general tensor product spaces. The central proofs are grounded in standard Clebsch-Gordan orthogonality and appear sound; the remaining issues are local gaps in exposition rather than flaws in the core derivation.

minor comments (7)

[Theorem 17, Eq. (10)] The proof asserts that the concatenated matrix D is 3^n × 3^n, but this dimension equality is not justified. The number of columns is the sum over all paths of (2l+1), and the equality with 3^n follows from summing the multiplicities in Proposition 10 (or from the standard irreducible decomposition of (R^3)^⊗n). Please add a short counting argument, because the claim DD^T = I depends on squareness.
[Lemma 6 and Appendix A, Eq. (40)] The proof of Lemma 15 relies on the 'general' orthogonality statement of Lemma 6 for two different CG tensors with the same first two weights and differing third weight. This is a standard property, but the lemma as stated is vague about which index pair is contracted. Please state the precise orthogonality relation used in Eq. (40) and either prove it or cite a standard reference, so the proof of Lemma 15 is self-contained.
[Appendix A, Lemma 15, Eq. (41)] The induction step for distinct path matrices is valid: in Eq. (41) the inner sum over j1 vanishes for every pair (j2,j'2) by the induction hypothesis, so the CG prefactors do not affect the argument. This concern, if raised, does not constitute a gap. The only external input is the standard CG orthogonality referenced above.
[Corollary 23] The proof of Corollary 23 is compressed to 'mostly identical to that of Lemma 13.' Since the general parentage scheme allows arbitrary bridge numbers, please spell out that Lemmas 6–8 hold for general CG tensors (l1,l2,lo) and that the same induction applies, or at least state the exact claim being invoked.
[Proposition 25] The block-matrix orthogonality argument in the proof of Proposition 25 is terse. The sentence 'If they appear in the same M, then the question is equivalent to Theorem 21' should be expanded to make the reduction explicit, particularly how the generalized path matrices inherit the column orthogonality and normalization properties.
[Algorithm 2, Step 2] Step 2 states the selection rule as '|lj − lt(j+1)| ≤ lj+1 ≤ lt(j+1) + 1,' which appears to be a typo; it should read '≤ lj + lt(j+1)' to match the CG selection rule stated in Section 6.1.
[Section 2.1] The text says l can take values from {1/2, 1, 3/2, ...}, but later restricts attention to integer l for O(3). Please clarify that half-integer l correspond to SU(2)/spin representations, not to the O(3) spherical harmonics used in the main construction.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is self-contained and rests on standard, externally supplied Clebsch–Gordan coefficient properties rather than on fitted parameters or self-citations.

full rationale

The paper's central claims—that concatenated normalized path matrices form an orthogonal rank-n ICT decomposition (Theorem 17) and that the products \hat P^{(p_i)}(\hat P^{(p_j)})^\top form an orthogonal equivariant basis (Theorem 21)—are derived from CG coefficient orthogonality, norm equality, and equivariance (Lemmas 6-8), which the paper imports as standard external facts (identifiable with Wigner-3j symbols; see Section 2.2 and the reproducibility statement citing e3nn and qutip for CG coefficients). These are not the target decomposition matrices and are not fitted to any subset of the paper's outputs. Algorithm 1 contains no trainable or fitted parameters; the timing results (1s, 3s, 11s, 4m32s) are reported benchmarks, not predictions made from a fitted model. No load-bearing self-citation appears: the references include standard external works (Coope, Andrews, Bonvicini, Pearce-Crump, Alex et al.) but no argument reduces to a prior paper by the present authors. The relation DI^{(j)}D^\top = \hat P^{(p_j)}(\hat P^{(p_j)})^\top in equation (11) is a straightforward algebraic identity given D's block structure, not a circular definition. The skeptical concern about a possible gap in the induction proof of Lemma 15 (the handling of distinct path matrices in equation (41)) is a correctness or rigor issue, not a circularity issue: even if that proof step were incomplete, the claim would depend on an unproved intermediate lemma rather than on the target conclusion being assumed as an input. Because there is no exhibited reduction of any claimed prediction to its own inputs, the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

No free parameters are fitted. The method assumes standard Clebsch-Gordan coefficient properties and the usual representation-theoretic facts. The only unproved structural input is the dimension-counting that D is square, which is standard but not derived.

assumptions (6)

standard math Clebsch-Gordan tensors satisfy column orthogonality, equal column norms, and equivariance (Lemmas 6-8)
Imported from quantum angular momentum theory (Racah 1942; Wigner-3j), used in Section 2.4 and throughout Section 4.
standard math The weight-l spherical harmonics form irreducible O(3) representations and the l=1 spherical space is O(3)-isomorphic to R^3 (Lemma 4)
Standard result; used to identify Cartesian and spherical spaces.
standard math Schur's lemma (Lemma 16)
Used in Theorems 17 and 21 to prove irreducibility of image spaces and basis structure.
standard math Multiplicity formula N(n;l) (Proposition 10, from Mihailov 1977 and Andrews 2023)
Used for dimension counting in Theorem 21; cited from prior literature.
domain assumption The total number of columns in all path matrices equals 3^n, so D is square (dimension counting in Theorem 17)
Not explicitly proved in the paper; relies on the standard fact that the sum of dimensions of irreps in the decomposition of (l=1)^⊗n equals the dimension of the space.
standard math Sylvester's rank inequality (Lemma 20)
Used in Theorem 21 to compute the rank of basis elements.

how reviews work

0 comments

Cite this review

Pith. "Pith review of High-Rank Irreducible Cartesian Tensor Decomposition and Bases of Equivariant Spaces." pith.science (2026). https://pith.science/paper/JA4Y7OYU

@misc{pith2026241218263,
  author       = {Pith},
  title        = {Pith review of: High-Rank Irreducible Cartesian Tensor Decomposition and Bases of Equivariant Spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JA4Y7OYU}},
  note         = {Machine review of arXiv:2412.18263}
}

abstract

Irreducible Cartesian tensors (ICTs) play a crucial role in the design of equivariant graph neural networks, as well as in theoretical chemistry and chemical physics. Meanwhile, the design space of available linear operations on tensors that preserve symmetry presents a significant challenge. The ICT decomposition and a basis of this equivariant space are difficult to obtain for high-rank tensors. After decades of research, Bonvicini (2024) has recently achieved an explicit ICT decomposition for $n=5$ with factorial time/space complexity. In this work we, for the first time, obtain decomposition matrices for ICTs up to rank $n=9$ with reduced and affordable complexity, by constructing what we call path matrices. The path matrices are obtained via performing chain-like contractions with Clebsch-Gordan matrices following the parentage scheme. We prove and leverage that the concatenation of path matrices is an orthonormal change-of-basis matrix between the Cartesian tensor product space and the spherical direct sum spaces. Furthermore, we identify a complete orthogonal basis for the equivariant space, rather than a spanning set (Pearce-Crump, 2023b), through this path matrices technique. Our method avoids the RREF algorithm and maintains a fully analytical derivation of each ICT decomposition matrix, thereby significantly improving the algorithm's speed to obtain arbitrary rank orthogonal ICT decomposition matrices and orthogonal equivariant bases. We further extend our result to the arbitrary tensor product and direct sum spaces, enabling free design between different spaces while keeping symmetry. The Python code is available at https://github.com/ShihaoShao-GH/ICT-decomposition-and-equivariant-bases, where the $n=6,\dots,9$ ICT decomposition matrices are obtained in 1s, 3s, 11s, and 4m32s on 28-core Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz, respectively.

Figures

Figures reproduced from arXiv: 2412.18263 by the authors.

**Figure 1.** The mind map of our method. The blocks in orange color are aligned with the contributions of this work. The ICT decomposition algorithm (Theorem 17), together with Schur’s lemma (Lemma 16), shows that path matrices can also be used to construct an equivariant basis. Building on the fact that our ICT decomposition relies on the parentage scheme, we propose a generalized parentage scheme that allows decomposition in a… view at source ↗

**Figure 2.** The parentage scheme. For ICTs Tˆ(n;l;q) , the x-axis corresponds to the rank n, the y-axis corresponds to the weight l, and the numbers in the circles represent the multiplicity N(n;l) of index q in equation (4). The line connection indicates that the irreducible representation with the weight in the left column can be decomposed into the irreducible representation with the weight in the right column, when coupling… view at source ↗

**Figure 3.** Construction of decomposition matrix. Top: The path matrices are formed via sequential contractions according to the paths in the parentage scheme. Bottom: The decomposition matrix is obtained as the product of the path matrix and its transpose. Ti1i2i3 = X j1 Cˆ lxlyl2 i2j1 C l1l2l3 i1j1i3 (7) and we flatten the first two indices of T, converting it into a (2l1+1)(2lx+1)(2ly+1)×(2l3+1) matrix. Then the following st… view at source ↗

Figures from the paper (13 more)

**Figure 4.** Figure 4: The general parentage scheme with the example of finding basis for ((2 ⊗ 2 ⊗ 2), p = −1) ⊕ ((1 ⊗ 3), p = −1) → ((3 ⊗ 4), p = −1). Compared to the classic parentage scheme, we need to additionally mark the bridge number in the scheme. outputs the possible weights accord…

**Figure 4.** Figure 4: This approach follows from the previous proofs and aligns with the core idea of [PITH_FULL_IMAGE:figures/full_fig_p032_4.png]

**Figure 5.** Figure 5: The parentage scheme of decomposition of (l = 0.5)⊗n for SU(2) and U(2), where n is the rank. The numbers in the circle represent multiplicity. 8.1 Relation to Other Works The remarkable work done by Pearce-Crump (2023b) has successfully found spanning sets for the equ…

**Figure 6.** Figure 6: Decomposition matrix for rank-2 ICT with path (0 [PITH_FULL_IMAGE:figures/full_fig_p044_6.png]

**Figure 7.** Figure 7: Decomposition matrix for rank-3 ICT with path (0 [PITH_FULL_IMAGE:figures/full_fig_p045_7.png]

**Figure 8.** Figure 8: Decomposition matrix for rank-4 ICT with path (0 [PITH_FULL_IMAGE:figures/full_fig_p045_8.png]

**Figure 9.** Figure 9: Decomposition matrix for rank-5 ICT with path (0 [PITH_FULL_IMAGE:figures/full_fig_p046_9.png]

**Figure 10.** Figure 10: Decomposition matrix for rank-6 ICT with path (0 → 1 → 1 → 0 → 1 → 2 → 1). 46 [PITH_FULL_IMAGE:figures/full_fig_p046_10.png]

**Figure 11.** Figure 11: Decomposition matrix for rank-7 ICT with path (0 → 1 → 2 → 3 → 2 → 2 → 1 → 1). 0.020 0.015 0.010 0.005 0.000 0.005 0.010 0.015 0.020 [PITH_FULL_IMAGE:figures/full_fig_p047_11.png]

**Figure 12.** Figure 12: Decomposition matrix for rank-7 ICT with path (0 → 1 → 1 → 1 → 2 → 1 → 2 → 2). 47 [PITH_FULL_IMAGE:figures/full_fig_p047_12.png]

**Figure 13.** Figure 13: Equivariant basis matrix of EndO(3) R3 ⊗6 with paths (1 → 2 → 2 → 3 → 3 → 3) and (1 → 1 → 2 → 3 → 4 → 3). 48 [PITH_FULL_IMAGE:figures/full_fig_p048_13.png]

**Figure 14.** Figure 14 [PITH_FULL_IMAGE:figures/full_fig_p049_14.png]

**Figure 15.** Figure 15: Equivariant basis matrix of HomSU(2) ((0.5 ⊗ 0.5 ⊗ 1.5),(0.5 ⊗ 0.5 ⊗ 0.5)) with paths (0.5 0.5 → 0 1.5 → 1.5) and (0.5 0.5 → 1 0.5 → 1.5). 50 [PITH_FULL_IMAGE:figures/full_fig_p050_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 39 canonical work pages

[1]

A numerical algorithm for the explicit calculation of SU(N) and SL(N,C) Clebsch--Gordan coefficients

Arne Alex, Matthias Kalus, Alan Huckleberry, and Jan von Delft. A numerical algorithm for the explicit calculation of SU(N) and SL(N,C) Clebsch--Gordan coefficients. Journal of Mathematical Physics, 52 0 (2): 0 023507, 2011

work page 2011
[2]

David L. Andrews. Symmetry-based identification and enumeration of independent tensor properties in nonlinear and chiral optics. The Journal of Chemical Physics, 158 0 (3): 0 034101, 2023

work page 2023
[3]

Irreducible fourth-rank Cartesian tensors

DL Andrews and WA Ghoul. Irreducible fourth-rank Cartesian tensors. Physical Review A, 25 0 (5): 0 2647, 1982

work page 1982
[4]

MACE : Higher order equivariant message passing neural networks for fast and accurate force fields

Ilyes Batatia, David P Kovacs, Gregor Simm, Christoph Ortner, and G \'a bor Cs \'a nyi. MACE : Higher order equivariant message passing neural networks for fast and accurate force fields. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 11423--11436, 2022

work page 2022
[5]

E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials

Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications, 13 0 (1): 0 2453, 2022

work page 2022
[6]

Irreducible Cartesian tensor decomposition: A computational approach

Andrea Bonvicini. Irreducible Cartesian tensor decomposition: A computational approach. The Journal of Chemical Physics, 160 0 (22): 0 224105, 2024

work page 2024
[7]

On algebras which are connected with the semisimple continuous groups

Richard Brauer. On algebras which are connected with the semisimple continuous groups. Annals of Mathematics, 38 0 (4): 0 857--872, 1937

work page 1937
[8]

Representations of compact Lie groups, volume 98

Theodor Br \"o cker and Tammo Tom Dieck. Representations of compact Lie groups, volume 98. Springer Science & Business Media, 2013

work page 2013

Show all 41 references

[9]

Cartesian atomic cluster expansion for machine learning interatomic potentials

Bingqing Cheng. Cartesian atomic cluster expansion for machine learning interatomic potentials. npj Computational Materials, 10 0 (1): 0 157, 2024

2024
[10]

u tt, and Klaus-Robert M \

Stefan Chmiela, Alexandre Tkatchenko, Huziel E Sauceda, Igor Poltavsky, Kristof T Sch \"u tt, and Klaus-Robert M \"u ller. Machine learning of accurate energy-conserving molecular force fields. Science Advances, 3 0 (5): 0 e1603015, 2017

2017
[11]

Group equivariant convolutional networks

Taco Cohen and Max Welling. Group equivariant convolutional networks. In International Conference on Machine Learning (ICML), pages 2990--2999, 2016

2016
[12]

Irreducible Cartesian tensors

JAR Coope, RF Snider, and FR McCourt. Irreducible Cartesian tensors. The Journal of Chemical Physics, 43 0 (7): 0 2269--2275, 1965

1965
[13]

Orthonormal decomposition of third rank tensors and applications

C i g dem Din c kal. Orthonormal decomposition of third rank tensors and applications. In Proceedings of the World Congress on Engineering, volume 1, 2013

2013
[14]

Atomic cluster expansion for accurate and transferable interatomic potentials

Ralf Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials. Physical Review B, 99 0 (1): 0 014104, 2019

2019
[15]

A practical method for constructing equivariant multilayer perceptrons for arbitrary matrix groups

Marc Finzi, Max Welling, and Andrew Gordon Wilson. A practical method for constructing equivariant multilayer perceptrons for arbitrary matrix groups. In International Conference on Machine Learning (ICML), pages 3318--3328, 2021

2021
[16]

Forces are not enough: Benchmark and critical evaluation for machine learning force fields with molecular simulations

Xiang Fu, Zhenghao Wu, Wujie Wang, Tian Xie, Sinan Keten, Rafael Gomez-Bombarelli, and Tommi Jaakkola. Forces are not enough: Benchmark and critical evaluation for machine learning force fields with molecular simulations. Transactions on Machine Learning Research (TMLR), 2023

2023
[17]

Directional message passing for molecular graphs

Johannes Gasteiger, Janek Gro , and Stephan G \"u nnemann. Directional message passing for molecular graphs. In International Conference on Learning Representations (ICLR), 2020

2020
[18]

Gemnet: Universal directional graph neural networks for molecules

Johannes Gasteiger, Florian Becker, and Stephan G \"u nnemann. Gemnet: Universal directional graph neural networks for molecules. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, pages 6790--6802, 2021

2021
[19]

Neural message passing for quantum chemistry

Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International Conference on Machine Learning (ICML), pages 1263--1272, 2017

2017
[20]

A new model for learning in graph domains

Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In International Joint Conference on Neural Networks (IJCNN), volume 2, pages 729--734, 2005

2005
[21]

Group theory and its applications in physics, volume 78

Teturo Inui, Yukito Tanabe, and Yositaka Onodera. Group theory and its applications in physics, volume 78. Springer Science & Business Media, 2012

2012
[22]

Clebsch-- G ordan nets: a fully F ourier space spherical convolutional neural network

Risi Kondor, Zhen Lin, and Shubhendu Trivedi. Clebsch-- G ordan nets: a fully F ourier space spherical convolutional neural network. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018

2018
[23]

Backpropagation applied to handwritten zip code recognition

Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel. Backpropagation applied to handwritten zip code recognition. Neural Computation, 1 0 (4): 0 541--551, 1989

1989
[24]

Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations

Yi-Lun Liao, Brandon M Wood, Abhishek Das, and Tess Smidt. Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations. In International Conference on Learning Representations (ICLR), 2023

2023
[25]

Equalities and inequalities for ranks of matrices

George Marsaglia and George PH Styan. Equalities and inequalities for ranks of matrices. Linear and Multilinear Algebra, 2 0 (3): 0 269--292, 1974

1974
[26]

Addition or arbitrary number of identical angular momenta

V V Mihailov. Addition or arbitrary number of identical angular momenta. Journal of Physics A: Mathematical and General, 10 0 (2): 0 147, 1977

1977
[27]

Learning local equivariant representations for large-scale atomistic dynamics

Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J Owen, Mordechai Kornbluth, and Boris Kozinsky. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications, 14 0 (1): 0 579, 2023

2023
[28]

An algorithm for computing with B rauer's group equivariant neural network layers

Edward Pearce-Crump. An algorithm for computing with B rauer's group equivariant neural network layers. arXiv preprint arXiv:2304.14165, 2023 a

2023 arXiv
[29]

Brauer’s group equivariant neural networks

Edward Pearce-Crump. Brauer’s group equivariant neural networks. In International Conference on Machine Learning (ICML), pages 27461--27482, 2023 b

2023
[30]

Theory of complex spectra

Giulio Racah. Theory of complex spectra. II . Physical Review, 62: 0 438--462, 1942

1942
[31]

Enumeration of 166 billion organic small molecules in the chemical universe database gdb-17

Lars Ruddigkeit, Ruud Van Deursen, Lorenz C Blum, and Jean-Louis Reymond. Enumeration of 166 billion organic small molecules in the chemical universe database gdb-17. Journal of Chemical Information and Modeling, 52 0 (11): 0 2864--2875, 2012

2012
[32]

u tt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert M \

Kristof Sch \"u tt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert M \"u ller. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. In Advances in Neural Information Processing Sys...

2017
[33]

Equivariant message passing for the prediction of tensorial properties and molecular spectra

Kristof Sch \"u tt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning (ICML), pages 9377--9388, 2021

2021
[34]

Tensornet: Cartesian tensor representations for efficient learning of molecular potentials

Guillem Simeon and Gianni De Fabritiis. Tensornet: Cartesian tensor representations for efficient learning of molecular potentials. In Advances in Neural Information Processing Systems (NeurIPS), volume 36, 2024

2024
[35]

On the inclusion of charge and spin states in C artesian tensor neural network potentials

Guillem Simeon, Antonio Mirarchi, Raul P Pelaez, Raimondas Galvelis, and Gianni De Fabritiis. On the inclusion of charge and spin states in C artesian tensor neural network potentials. arXiv preprint arXiv:2403.15073, 2024

2024 arXiv
[36]

Irreducible Cartesian Tensors , volume 43

Robert F Snider. Irreducible Cartesian Tensors , volume 43. Walter de Gruyter GmbH & Co KG, 2017

2017
[37]

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

Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3 D point clouds. arXiv preprint arXiv:1802.08219, 2018

2018 arXiv
[38]

Equivariant diffusion policy

Dian Wang, Stephen Hart, David Surovik, Tarik Kelestemur, Haojie Huang, Haibo Zhao, Mark Yeatman, Jiuguang Wang, Robin Walters, and Robert Platt. Equivariant diffusion policy. In The Conference on Robot Learning (CoRL), 2024

2024
[39]

3 D steerable CNN s: Learning rotationally equivariant features in volumetric data

Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco S Cohen. 3 D steerable CNN s: Learning rotationally equivariant features in volumetric data. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018

2018
[40]

Higher-rank irreducible cartesian tensors for equivariant message passing

Viktor Zaverkin, Francesco Alesiani, Takashi Maruyama, Federico Errica, Henrik Christiansen, Makoto Takamoto, Nicolas Weber, and Mathias Niepert. Higher-rank irreducible cartesian tensors for equivariant message passing. arXiv preprint arXiv:2405.14253, 2024

2024 arXiv
[41]

Orthogonal irreducible decompositions of tensors of high orders

W-N Zou, Q-S Zheng, D-X Du, and J Rychlewski. Orthogonal irreducible decompositions of tensors of high orders. Mathematics and Mechanics of Solids, 6 0 (3): 0 249--267, 2001

2001

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

A numerical algorithm for the explicit calculation of SU(N) and SL(N,C) Clebsch--Gordan coefficients

Arne Alex, Matthias Kalus, Alan Huckleberry, and Jan von Delft. A numerical algorithm for the explicit calculation of SU(N) and SL(N,C) Clebsch--Gordan coefficients. Journal of Mathematical Physics, 52 0 (2): 0 023507, 2011

work page 2011

[2] [2]

David L. Andrews. Symmetry-based identification and enumeration of independent tensor properties in nonlinear and chiral optics. The Journal of Chemical Physics, 158 0 (3): 0 034101, 2023

work page 2023

[3] [3]

Irreducible fourth-rank Cartesian tensors

DL Andrews and WA Ghoul. Irreducible fourth-rank Cartesian tensors. Physical Review A, 25 0 (5): 0 2647, 1982

work page 1982

[4] [4]

MACE : Higher order equivariant message passing neural networks for fast and accurate force fields

Ilyes Batatia, David P Kovacs, Gregor Simm, Christoph Ortner, and G \'a bor Cs \'a nyi. MACE : Higher order equivariant message passing neural networks for fast and accurate force fields. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 11423--11436, 2022

work page 2022

[5] [5]

E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials

Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications, 13 0 (1): 0 2453, 2022

work page 2022

[6] [6]

Irreducible Cartesian tensor decomposition: A computational approach

Andrea Bonvicini. Irreducible Cartesian tensor decomposition: A computational approach. The Journal of Chemical Physics, 160 0 (22): 0 224105, 2024

work page 2024

[7] [7]

On algebras which are connected with the semisimple continuous groups

Richard Brauer. On algebras which are connected with the semisimple continuous groups. Annals of Mathematics, 38 0 (4): 0 857--872, 1937

work page 1937

[8] [8]

Representations of compact Lie groups, volume 98

Theodor Br \"o cker and Tammo Tom Dieck. Representations of compact Lie groups, volume 98. Springer Science & Business Media, 2013

work page 2013