Spherical & Cartesian Tensor Equivariant Models

• Spherical and Cartesian tensor-based equivariant models are deep learning architectures that enforce symmetry by transforming features under E(3) and O(3) groups.
• They utilize spherical harmonics with Clebsch–Gordan products or Cartesian tensor decompositions to rigorously encode physical invariance for properties such as energies, forces, and high-order tensor attributes.
• Comparative analyses highlight trade-offs in computational efficiency, memory scaling, and applicability, guiding model selection in atomistic simulations and materials modeling.

Spherical and Cartesian Tensor-Based Equivariant Models

Equivariant models are fundamental in geometric deep learning, atomistic simulations, and materials modeling due to the physical requirement that predicted quantities must transform covariantly or remain invariant under rigid motions, specifically under the E(3) or O(3) symmetry groups. Two primary frameworks for encoding and enforcing equivariance in neural architectures have emerged: one based on spherical harmonics and irreducible representations of SO(3) ("spherical-tensor" or "spherical-harmonic" models), and the other constructed from (possibly irreducible) Cartesian tensors ("Cartesian-tensor" models). Both paradigms admit rigorous mathematical formulations, efficient computational realizations, and are grounded in classical invariant theory and modern representation theory, but they differ significantly in practical construction, computational scaling, and applicability domains.

1. Mathematical and Representation-Theoretic Foundations

Equivariant neural layers exploit the principle that many molecular and condensed-matter properties—energies, forces, dipoles, stress tensors, polarizabilities, etc.—transform beneath the rotation (and, for E(3), translation) groups according to well-defined group representations. The choice of tensor-type directly affects both the theoretical rigor and computational tractability of enforcing O(3) or E(3) equivariance within model architectures.

• Spherical tensors: Features are decomposed into irreducible representations of SO(3), indexed by angular momentum quantum number ℓ. A feature of type ℓ ("rank-ℓ spherical tensor") is a (2ℓ+1)-component vector transforming via the Wigner D-matrix, $D^{(\ell)}(R)$, under rotation $R$:

$$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$

Clebsch–Gordan (CG) coefficients prescribe bilinear and multilinear equivariant interactions, i.e., $\ell_1\otimes\ell_2\rightarrow\ell_3$.

• Cartesian tensors: Features are represented as rank-$r$ arrays $T_{i_1...i_r}\in\mathbb{R}^{3^r}$ transforming by

$$T'_{i_1...i_r} = R_{i_1j_1}\cdots R_{i_rj_r}T_{j_1...j_r}$$

under rotations. The decomposition into irreducible (traceless-symmetric) tensors of weight ℓ is essential for minimality and efficiency, but general Cartesian models can operate directly, only subsequently projecting onto irreducible sectors as needed.

The block-diagonalization of Cartesian tensors and the correspondences between spherical and Cartesian irreps are mathematically formalized and operationalized in frameworks such as e3nn, TensorNet, TACE, and CartNN (Geiger et al., 2022, Xu et al., 18 Dec 2025, Xu et al., 18 Sep 2025, Simeon et al., 2023, Shao et al., 2024).

2. Core Equivariant Operations and Tensor Bases

Both classes of models build all equivariant operations—message passing, convolutions, attention—using a small set of group-theoretic building blocks.

Spherical Harmonic-Based Equivariant Layers

• Spherical harmonics $Y_{\ell}^m(\hat{r})$ serve as angular bases for relative directions, forming the backbone for rotationally equivariant message-passing kernels. Convolutions and filters are expanded on $Y_\ell^m$; filter construction employs learned radial functions and fixed angular bases (Geiger et al., 2022, Tang, 15 Dec 2025).
• Clebsch–Gordan tensor product: Linear maps that couple two features of degree $\ell_1$ and $R$0 into all allowable output degrees $R$1, implemented through precomputed CG coefficients.
• SO(3)-equivariant kernels are constructed as sums of radial basis functions times angular projectors, $R$2.

Cartesian Tensor-Based Equivariant Layers

• Tensor product and contraction: Expressed as explicit index manipulations—outer products, contractions, and symmetrizations—which are naturally equivariant owing to the transformation law for each index (Wang et al., 2024, Shao et al., 2024).
• Irreducible Cartesian tensor decomposition (ICTD): Symmetric-traceless projections extract the minimal "active" irreps from a full Cartesian tensor, enabling parameter and memory compression, with explicit constructions available up to $R$3 (Shao et al., 2024).
• Cartesian-3j and -nj symbols: Generalize the CG rules to the Cartesian basis, allowing the construction of high-order equivariant couplings without reference to spherical harmonics (Xu et al., 18 Dec 2025).
• Channelized bases: Models such as CEITNet aggregate local environments into multi-channel Cartesian tensors, and all interaction is performed in channel space followed by basis assembly (Jin et al., 4 Feb 2026).

3. Model Architectures and Computational Scaling

The implementation details, architectural constraints, and computational scaling of equivariant models are governed by the nature of their tensor bases.

Paradigm	Basis Dim. (rank ℓ)	Coupling Mechanism	Computational Scaling	Memory Scaling
Spherical	$R$4	Clebsch–Gordan tensor prod.	$R$5 (full), $R$6 Gaunt	$R$7
Cartesian	$R$8 (full rep)	Outer prod./contraction + sym.	$R$9 naïve; with ICTD $$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$0	$$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$1

• Spherical: Models such as Tensor Field Networks, SE(3)-Transformer, and Allegro employ layers of spherical-harmonic features, typically with $$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$2 to avoid prohibitive cost. CG products are parameter- and memory-efficient up to moderate degree but become impractical for high-rank interactions or high-body order (Tang, 15 Dec 2025).
• Cartesian: TensorNet, TACE, HotPP, and CEITNet can handle high-rank tensors, arbitrary-order contractions, and direct tensor property prediction, often with lower floating-point operation counts and fewer learnable parameters at low rank, but incur exponentially increasing storage at large $$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$3 (Xu et al., 18 Sep 2025, Simeon et al., 2023, Zaverkin et al., 2024, Jin et al., 4 Feb 2026).

Practical implementations benefit from mapping between bases: e3nn supports seamless conversion between Cartesian and spherical representations, and frameworks such as CartNN provide generalized ICTP/ICTC support (Xu et al., 18 Dec 2025, Geiger et al., 2022).

4. Empirical Performance, Expressivity, and Applicability

Quantitative head-to-head comparisons in the literature demonstrate near-equivalence in achievable accuracy for molecular energies, force fields, dipole/polarizability, and high-order tensor properties for $$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$4–$$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$5.

• Energy/force benchmarks: On liquid water, carbon allotropes, and diverse QM9/rMD17 molecules, pure Cartesian and pure spherical models (TensorNet, TACE, HotPP, cNequIP, MACE, Allegro) all achieve sub-meV/atom and sub-10 meV/Å force MAEs, with no systematic advantage across benchmarks (Xu et al., 18 Sep 2025, Simeon et al., 2023, Wang et al., 2024, Xu et al., 18 Dec 2025).
• High-order tensor prediction: Tasks such as crystal elastic, dielectric, and piezoelectric tensor prediction reveal that Cartesian models such as CEITNet or TACE consistently match or exceed the accuracy of CG/spherical-based models, while being 4–13× faster at rank 3–4, and reducing model size by up to 46% (Jin et al., 4 Feb 2026).
• Expressivity: Spherical transform-based architectures, such as the Equivariant Spherical Transformer (EST), can, in principle, subsume all CG-based function spaces and break degree-bounded limitations via spatial attention, distinguishing high-fold symmetries unreachable to truncated CG product models (2505.23086).
• Parameter vs. memory scaling: Spherical models are more stable at high angular resolution (large $$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$6), whereas pure Cartesian approaches experience superlinear growth in memory and computational complexity—hybrid designs or irreducible decomposition are necessary for practical scalability (Xu et al., 18 Dec 2025).

5. Theoretical Advances: Decomposition, Coupling Algebra, and Universal Approximation

Recent research has made major advances in the analytic understanding and efficient implementation of both frameworks.

• Explicit ICT decomposition: Path-matrix constructions, exploiting chain contractions of Clebsch–Gordan tensors, have enabled explicit, orthonormal bases for irreducible Cartesian tensors up to rank 9, reducing factorial complexity to exponential (Shao et al., 2024).
• Cartesian-3j/nj algebra: The algebra of tensor products, contractions, and irreducible projections (ICTP, ICTC) in Cartesian models is now fully compatible with the spherical Clebsch–Gordan scheme, enabling analytic interoperability and hybrid designs (Xu et al., 18 Dec 2025).
• Universal function spaces: For spherical tensors, any proper equivariant map can be written as a sum of maximally coupled CG-basis tensors weighted by scalar functions. Efficient approximations can reduce to a minimal set of 2λ+1 frame vectors plus a correction, with negligible loss in practical accuracy (Domina et al., 8 May 2025).
• Symmetric tensor networks: Both Cartesian and spherical equivariant functions (invariant polynomials and higher-order messages) can be generated programmatically from graphical tensor networks of Kronecker delta, Levi–Civita, and irreducible projectors, yielding systematic model families and automatable basis construction (Zhang et al., 18 Aug 2025).

6. Practical Guidelines, Limitations, and Future Directions

The regime of application, implementation resources, and scientific objectives influence the optimal choice of tensor paradigm.

• Scenarios favoring Cartesian models:
  • Direct prediction of arbitrary-rank, structure-determined tensors (e.g. polarizabilities, high-rank crystal response) (Xu et al., 18 Sep 2025, Simeon et al., 2023, Zaverkin et al., 2024, Jin et al., 4 Feb 2026).
  • Unified scalar/tensorial modeling and integration of invariants (charge, basis, field) at low cost.
  • Platform independence (works in general E(n)), implementation simplicity via built-in PyTorch/NumPy contractions.
• Scenarios favoring spherical models:
  • Efficiency for high angular momentum ($$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$7) and full irreducibility.
  • Minimal number of feature channels at high rank; legacy infrastructure for Clebsch–Gordan algebra.
• Known bottlenecks and directions:
  • Memory/compute scaling for high-order Cartesian tensors remains challenging; hybrid architectures (low-rank Cartesian blocks + spherical harmonics at higher $$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$8) may offer the best trade-offs (Xu et al., 18 Dec 2025).
  • Fast numerical or analytic routines for generic ICT contractions (Cartesian-k-j) are an open area.
  • Extensions to space groups, enforcement of crystal symmetries beyond O(3), and generative tasks exploiting full equivariant expressivity are under active development (Heilman et al., 2024, Zhang et al., 18 Aug 2025, 2505.23086).
  • As hardware and software support for high-order tensor computation matures, pure Cartesian models are likely to become standard for applications requiring up to $$x'_{m} = \sum_{m'=-\ell}^\ell D_{mm'}^{(\ell)}(R) x_{m'}$$9 precision (Xu et al., 18 Sep 2025, Jin et al., 4 Feb 2026).

7. Representative Models and Benchmarks

The following table summarizes leading models representative of each paradigm:

Model	Tensor Paradigm	Notable Features	Key References
TFN/SE(3)-Transformer	Spherical	CG-based convolution/attention	(Tang, 15 Dec 2025, Geiger et al., 2022)
TACE/TMP	Cartesian/Irreducible	Universal property prediction, field embeddings, LES	(Xu et al., 18 Sep 2025)
TensorNet	Cartesian	O(3) matrix-embedding, low param/compute budget	(Simeon et al., 2023)
CEITNet	Cartesian, Channelized	Efficient high-rank tensor property prediction in crystals	(Jin et al., 4 Feb 2026)
EST	Spherical (spatial domain)	Transformer; breaks degree-bound expressivity	(2505.23086)
CartNN/cNequIP/cAllegro/cMACE	Cartesian, ICT	Systematic port of e3nn-based spherical architectures	(Xu et al., 18 Dec 2025)

The landscape of equivariant neural models now includes well-established spherical-tensor designs and a family of Cartesian-tensor models with efficient, theoretically principled couplings and proven performance parity. Continued advances in algebraic decomposition, numerical optimization, and software integration are driving both paradigms toward broader applicability and increasing specialty for high-precision scientific learning tasks.