Rotationally Equivariant Representations

Updated 13 January 2026

Rotationally equivariant representations are structured feature spaces that ensure predictable, linear responses to rotations.
They employ kernel constraints, moment kernels, and irreducible decompositions to capture inherent geometric symmetries.
Applications in molecular modeling, computer vision, and physics yield improved efficiency, accuracy, and sample complexity.

Rotationally equivariant representations are structured feature spaces, operators, or architectures in which the action of a group of rotations on the input induces a predictable, typically linear action on the output. Such representations are foundational in geometric deep learning, enabling models to respect the intrinsic symmetries of physical, biological, and mathematical systems, dramatically improving sample efficiency, parameterization, and generalization, especially for systems where rotation is a nuisance or a fundamental symmetry.

1. Mathematical Foundations of Rotational Equivariance

Let $G = {\rm SO}(d)$ or ${\rm O}(d)$ denote the relevant group of rotations (or rotations plus reflections) in dimension $d$ . For a function or field $u:\mathbb{R}^d \to V$ or $X \mapsto Y$ , rotational equivariance requires that for all $R \in G$ ,

$\Phi(R \cdot x) = \rho_{\text{out}}(R) \, \Phi(x)$

where $\rho_{\text{out}}$ is the representation on the output (identity for scalars, standard rotation for vectors, tensor products for higher-order tensors) and $R \cdot x$ denotes the rotated input.

In the setting of linear operators on fields, rotational equivariance, together with translation invariance (for Euclidean domains), fully characterizes the operator as a convolution with a radially symmetric tensor kernel,

$L(u)(x) = \int_{\mathbb{R}^d} K(\|x-y\|) \otimes u(y) \, dy$

with the critical property that the kernel satisfies

$K(R x) = \rho_{\text{out}}(R) K(x) \rho_{\text{in}}(R)^{-1}$

for all $R \in G$ (Shen et al., 2021, Schlamowitz et al., 27 May 2025, Weiler et al., 2018).

Analysis via group representation theory, especially the use of irreducible representations (irreps) and Schur’s Lemma, establishes that all such covariant kernels must decompose as sums of products of radial functions and angular harmonics (circular or spherical), or as moment monomials times radial profiles for general O(d)-equivariant kernels. In harmonic language, the only nontrivial equivariant directions correspond to irreducible representations of the underlying symmetry group (Shen et al., 2021, Schlamowitz et al., 27 May 2025).

2. Core Architectural Mechanisms and Kernel Parameterizations

There are three main generic frameworks for constructing rotationally equivariant neural operators:

Kernel constraint approach: All convolutional kernels are parameterized as linear combinations of invariant combinations of radial functions and angular (spherical/circular) harmonics (classic steerable/filter basis) (Weiler et al., 2018, Shen et al., 2021, Schlamowitz et al., 27 May 2025).
Moment kernels: Any equivariant linear map between tensor fields is, up to a change of basis, a finite sum of terms of the form $R_p(\|x\|) \cdot x^{\otimes n}$ , possibly contracted with identity tensors (Kronecker deltas), where $R_p$ are scalar learnable radial profiles (Schlamowitz et al., 27 May 2025).
Irreducible decomposition and channel organization: Feature spaces are explicitly organized by irreducible representations (channels of type-ℓ features), and convolutions, nonlinearities, and learnable maps respect the block structure dictated by group representation theory. This approach is essential for higher-order SO(3)-equivariant networks, such as in molecular modeling and protein structure prediction (Visani et al., 2023, Dang et al., 8 May 2025, Schütt et al., 2021, Miller et al., 2020).

The constraint on the convolutional kernel $K$ is expressed as: $K(Rx) = \rho_{\text{out}}(R) K(x) \rho_{\text{in}}(R)^{-1}$ which, via harmonic or moment expansion, ensures the linear operator's equivariance. This principle is fully constructive: by enforcing the weight-sharing and channel-mixing structure prescribed above, standard convolution layers can be extended to arbitrary high-dimensional and high-rank equivariant architectures (Schlamowitz et al., 27 May 2025, Weiler et al., 2018).

3. Nonlinearities, Deep Architectures, and Expressivity

A central result is that pointwise scalar nonlinearities (e.g., ReLU) cannot yield nontrivial rotation-equivariant networks for connected compact groups such as SO(d); any such architecture automatically collapses to a rotation-invariant mapping (Pacini et al., 2024). To maintain genuine equivariance at depth, nonlinearities must be non-pointwise—gated, tensor-product, or norm nonlinearities applied to group-theoretically structured channels. In the context of SO(3)-equivariant architectures, the standard mechanisms are:

Gated nonlinearities: Scalar-by-vector (or higher-order) gating, where each non-scalar channel is modulated by a learned scalar gate (Weiler et al., 2018, Esteves, 2020).
Tensor-product/Clebsch–Gordan nonlinearities: Channels of order ℓ are mixed according to Clebsch–Gordan rules, ensuring the block-structure resulting from tensor products of irreps is preserved (Visani et al., 2023, Wiersma et al., 2020, Dang et al., 8 May 2025).
Norm-based nonlinearities: Nonlinearity is applied to the Euclidean norm of a vector-valued channel, then broadcast as a scaling (Weiler et al., 2018, Shutty et al., 2020).

In spherical or surface CNNs, complex-valued features organized according to angular momentum order, and pointwise magnitude-based nonlinearities (e.g., complex ReLU), guarantee phase-correct equivariance (rotation induces a predictable exponential phase shift in each stream) (Wiersma et al., 2020).

4. Application Domains and Empirical Impact

Rotationally equivariant representations yield substantial gains in multiple domains with innate or imposed rotational symmetry:

Molecular property prediction and quantum chemistry: Equivariant graph neural networks drastically reduce sample complexity and parameter count compared to invariant architectures. For example, explicitly including vector (ℓ=1) channels in message passing yields ≈23% lower mean absolute error across the QM9 chemical property suite, with improvements up to 50% for dipole prediction (Miller et al., 2020, Schütt et al., 2021). EquiHGNN demonstrates further advances by extending E(3)-equivariance to hypergraphs, improving accuracy for large, conformationally rich molecular systems (Dang et al., 8 May 2025).
Geometric vision and surface analysis: Surface CNNs with vector-valued rotation-equivariant features resolve the local gauge ambiguity on curved manifolds, achieving state-of-the-art results in shape classification, segmentation, and correspondence with full discrete rotational equivariance (Wiersma et al., 2020).
3D volumetric perception: Steerable 3D CNNs, constructed using analytic SO(3) (or SE(3)) kernel bases, robustly solve classification and regression tasks under arbitrary rotations with much smaller models and without requiring augmentation (Weiler et al., 2018, Shen et al., 2021). Biomedical imaging and 3D image registration tasks benefit from moment kernel approaches, which are scalable and highly interpretable (Schlamowitz et al., 27 May 2025).
Physics-informed machine learning: Operators such as Green's functions and differential operators on fields are naturally cast as rotationally equivariant convolutions with radially symmetric kernels, forming a bridge between classical scientific modeling and neural operator learning (Shen et al., 2021).
Self-supervised and contrastive learning: Explicitly enforcing that augmentations act as rotations in embedding space leads to more informative, structured representations, as seen in CARE (Gupta et al., 2023).
Quantum machine learning: Equivariant circuit ansätze avoid barren plateaus in optimization and substantially improve accuracy and trainability over generic architectures, with theoretical guarantees via commutant algebra constraints and representation theory (West et al., 2023).

5. Implementation Paradigms, Algorithms, and Software

Practical deployment of rotationally equivariant representations requires:

Automated kernel construction: Tools such as e3nn, EquivariantOperators.jl, and Lie Algebraic Networks (LAN) automate finding irreducible representations, kernel parameterizations, and channel structures from either explicit group generators or structure constants of the Lie algebra (Shutty et al., 2020, Shen et al., 2021).
Radial profile parameterization: Nearly all methods reduce the high-dimensional kernel search to a set of learnable scalar radial functions, typically via MLPs or radial basis expansions. The group action is implemented by wrapping these profiles with appropriate angular harmonics or by enforcing weight-sharing patterns (Shen et al., 2021, Weiler et al., 2018, Schlamowitz et al., 27 May 2025).
Equivariant message passing on graphs/hypergraphs: Tensor-valued messages and node features, together with message aggregation and update mechanisms that respect equivariant transformations, form the backbone of molecular graph neural networks (e.g., PaiNN, Equiformer, EquiHGNN) (Schütt et al., 2021, Dang et al., 8 May 2025).
Invariant projections from equivariant features: For downstream tasks requiring invariance, explicit random invariant projections (assembled from the basis of group-invariant linear maps) yield high-fidelity, rotation-invariant summaries of equivariant features without destroying information (Hansen et al., 2024).
Attention and spatio-temporal encoding: For dynamic or time-series imaging data, spatially equivariant features are pooled and concatenated with time encodings and processed via attention blocks to yield approximately equivariant spatio-temporal codes (Wang et al., 2024).
Clebsch–Gordan solver and representation learning: For arbitrary Lie algebras, algorithms exist that compute numeric representations and CG coefficients from structure constants, extending equivariant architectures beyond just SO(2)/SO(3) (Shutty et al., 2020).

6. Theoretical Limits and Open Challenges

A central theoretical finding is that there is no nontrivial pointwise nonlinear activation that is strictly equivariant under SO(d) for finite-dimensional representations; exact equivariance at depth necessitates non-pointwise/fiberwise nonlinearities structured according to group irreps (Pacini et al., 2024). This restricts the design space for deep rotation-equivariant architectures, except in trivial (rotation-invariant) cases.

For group actions with nontrivial stabilizers (i.e., when the data admits symmetries that fix points, such as images of objects with rotational symmetry), minimal equivariant representations must encode both cosets of the stabilizer and orbit labels. Failure to account for stabilizers leads to loss of information ("collapsing orbits"), which is remedied by the orbit-stabilizer theory encoding (e.g., in the EquIN framework) (Rey et al., 2023).

Moment-kernel characterizations demonstrate that the only equivariant linear operators under O(d) are linear combinations of monomials in the coordinates multiplied by radially symmetric functions (and contraction with deltas), placing strong algebraic constraints on what is learnable via equivariant convnets (Schlamowitz et al., 27 May 2025).

Finally, while these principles extend naturally to translations and finite/matrix Lie groups, extensions to noncompact groups, continuous stabilizer subgroups, or more general data manifolds remain active areas of research.

7. Summary Table: Selected Architectural Blueprints

Domain/Task	Representation Stack	Kernel/Nonlinearity Constraint
Planar/surface CNNs	Circular harmonics; complex phase–order streams	Steerable kernels, equivariant parallel-transport
3D volumetric data	Spherical harmonics; tensor bundles; type-ℓ channels	SO(3)-steerable basis, tensor-product nonlinearities
Molecular property graphs	Scalar/vector/tensor node features (irreps); message passing	Equivariant message update and aggregation
Physics-inspired convolutions (fields)	Scalar/vector/tensor fields	Radially symmetric moment kernels
Spherical/omni-valued data	Spherical harmonics (band-limited); scattering/wavelet layers	Modulus pointwise; all convolution harmonic-structured
Hypergraph equivariant networks	Scalar/vector features on nodes/hyperedges; AllSet bipartite GNN	EGNN/SE(3)-Transformer block, basis expansion