PaperlayerView exact revision on arXiv

Source-linked AI summary

Rotational Equivariance in Machine Learning: A Comprehensive Tutorial

Peter Lippmann, Fred A. Hamprecht

arXiv:2608.31045v1cs.LG

TL;DR

The tutorial addresses how machine-learning models for 3D data can respect coordinate independence through rotational equivariance. It builds the required geometric and representation-theoretic foundations, surveys architectural strategies, and emphasizes their design trade-offs and computational limitations.

  • Problem

    3D machine-learning predictions must remain consistent across arbitrary coordinate frames, but raw-coordinate models do not automatically preserve the required transformation behavior.

  • Method

    The tutorial develops rotational equivariance from geometric deep learning, group representations, spherical harmonics, Wigner matrices, tensor products, and Clebsch-Gordan decomposition, then surveys group convolutions, tensorial representations, and canonicalization.

  • Results

    The tutorial connects these mathematical formalisms to modern equivariant architectures and clarifies their practical strengths, limitations, and design trade-offs.

  • Takeaways & Limitations

    Practitioners can use the tutorial to navigate competing equivariant model choices in settings where high-quality 3D or simulation data are costly.

Abstract

from arXiv · show

Rotational symmetry is one of the most important structural principles in machine learning on 3D data. In applications ranging from physics and materials science to 3D computer vision, predictions should not depend on an arbitrary choice of coordinate frame. Rotational equivariance captures this requirement mathematically by enforcing that a rotation of the input induces a corresponding transformation of the model output. This tutorial provides a comprehensive introduction to rotational equivariance, starting from the physical and geometric intuition behind coordinate independence and building up the necessary machinery from geometric deep learning, group theory, and representation theory. We introduce message passing on Euclidean graphs, group actions and representations, spherical harmonics, Wigner matrices, tensor products, and Clebsch-Gordan decomposition, and explain how these ingredients give rise to modern equivariant architectures. We then survey the principal strategies for incorporating rotational equivariance in deep learning, including group convolutions, internal tensorial representations, and canonicalization-based methods, and discuss their practical strengths and limitations. The tutorial aims to lower the barrier to the subject by connecting the underlying mathematics to practical model design, by unifying ideas that are often expressed in different formal languages, and by helping practitioners choose among competing approaches through a clear discussion of their trade-offs.

I. INTRODUCTION

Rotational equivariance encodes coordinate independence by requiring outputs to transform consistently when inputs are rotated. The tutorial motivates this principle physically and practically, then connects it to mathematical tools and model-design choices.

  • Physical laws are formulated so predictions do not depend on arbitrary coordinate systems or reference frames.
  • Molecules require invariant scalar observables, such as energy, and consistently transforming vector or tensor quantities, such as forces and dipole moments.
  • Built-in symmetry can improve generalization and data efficiency by avoiding the need to learn symmetries from finite data while guaranteeing exact behavior out of distribution.
  • The tutorial develops equivariance from symmetry and message passing through group representations, spherical harmonics, Wigner matrices, tensor products, and Clebsch-Gordan decomposition.
  • Naive dense layers and component-wise nonlinearities can mix geometric components incorrectly, so equivariance generally requires carefully designed architectures.

III. GEOMETRIC DEEP LEARNING AND MESSAGE PASSING

Geometric data can be represented as Euclidean graphs whose nodes carry features and positions, enabling message passing to build increasingly abstract representations. The same framework also encompasses fully connected attention-based models and spatial symmetry groups.

  • Euclidean graphs represent positioned nodes with geometric features and edges constructed, for example, by k-nearest-neighbor or radius rules.
  • Message passing neural networks update each node by constructing messages from neighboring node features and aggregating them permutation-invariantly.
  • Stacking message-passing layers propagates information over longer graph distances and combines local geometric information into higher-level representations.
  • Fully connected graphs are a special MPNN case, and Transformer self-attention can be viewed as content-dependent message aggregation.
  • Euclidean symmetry handling focuses on translations, rotations, and reflections, with O(d) including reflections and SO(d) containing orientation-preserving rotations.

V. GROUP REPRESENTATIONS AND EQUIVARIANCE

Group representations specify how geometric quantities transform, allowing equivariance to be stated as commutation between input and output actions. Examples include vectors, higher-order tensors, and the reflection-sensitive cross product.

  • A velocity transforms as a vector under rotations, while its length remains invariant under orthogonal transformations.
  • A group representation maps symmetry transformations to invertible linear operators and specifies how those transformations act on vectors in a representation space.
  • Cartesian tensors transform by applying the rotation to every index, producing a dn-dimensional representation space for rank-n tensors.
  • Equivariance means transforming an input before applying a function is equivalent to applying the function first and transforming its output afterward; invariance is the identity-output special case.
  • The cross product is equivariant under rotations but becomes an axial-vector transformation under reflections, acquiring a determinant-dependent sign.
  • Representations provide the basis for transforming features between different local reference frames in tensorial message passing.

VI. SPHERICAL HARMONICS AND THE IRREDUCIBLE REPRESENTATIONS OF SO(3)

Spherical harmonics provide rotation-adapted bases whose fixed-degree subspaces realize the irreducible representations of SO(3). This connects Cartesian tensor decomposition to spherical tensors and practical equivariant features.

  • Cartesian tensor representations can be decomposed into irreducible SO(3) subrepresentations realized by Wigner-D matrices, while spherical harmonics transform under the same irreducibles.
  • A rank-2 Cartesian tensor is reducible because it contains the invariant identity-matrix subspace, corresponding to its trace component.
  • Spherical harmonics form a complete orthonormal basis for square-integrable functions on the sphere and may be represented in complex or real bases.
  • Under rotations, harmonics of degree l mix only among themselves through a (2l+1)-dimensional Wigner-D matrix.
  • Degree-0 harmonics are invariant, while degree-1 harmonics transform like ordinary vectors in a suitable real basis.
  • SO(3) irreducible representations are indexed by l and have dimension 2l+1, motivating spherical tensors with optional even or odd parity under reflections.

A. Computing Wigner-D matrices

Wigner-D matrices represent how degree-l spherical harmonics transform under rotations. The Pinchon-Hoggan factorization computes them efficiently by reducing y-rotations to z-rotations using Euler decomposition and a coordinate-swap operator.

  • Wigner-D representation: Wigner-D matrices encode the rotation behavior of degree-l spherical harmonics in a real basis.The construction derives representation matrices by studying spherical-harmonic transformations under rotations.
  • Euler-angle construction: A z-rotation mixes each (cos(mϕ), sin(mϕ)) pair through a planar rotation by angle mα.This simple z-rotation structure is the basis for the factorization strategy.
  • Euler-angle construction: Any R ∈ SO(3) is decomposed into z-y-z Euler rotations whose representation matrices combine by the homomorphism property.The factors correspond to rotations Rz(·) and Ry(·) with Euler angles α, β, and γ.
  • Pinchon-Hoggan reduction: The coordinate swap J maps (x,y,z) to (x,z,y) and conjugates the harder y-rotation into a z-rotation.Its induced degree-l operator J^(l) is well defined on spherical tensors.
  • Computational cost: The resulting real-basis algorithm computes D^(l)(R) in O(l^3) time using cheap 2×2 blocks and reusable, degree-dependent J^(l) matrices.The matrices J^(l) can be precomputed once for each degree l and reused across rotations.

VII. DECOMPOSITION OF CARTESIAN TENSORS AND THE TENSOR PRODUCT

Cartesian tensor products preserve equivariance but generally produce reducible representations. Clebsch-Gordan decomposition changes basis to compact irreducible SO(3) components, while exposing computational costs that motivate efficient implementations.

  • Tensor products: Cartesian tensor products transform equivariantly under the product representation but can cause combinatorial growth in feature dimensionality.Decomposition into irreducible representations provides smaller building blocks with well-defined transformation behavior.
  • Cartesian tensor decomposition: A rank-2 Cartesian tensor decomposes into antisymmetric, trace, and symmetric-traceless components with dimensions 3, 1, and 5.The antisymmetric component is a pseudovector, the trace is the l = 0 irrep, and the symmetric-traceless part is the l = 2 irrep.
  • Cartesian tensor decomposition: A rank-3 Cartesian tensor decomposes as 1 ⊕ 3 ⊕ 3 ⊕ 3 ⊕ 5 ⊕ 5 ⊕ 7.This corresponds to one scalar, three vector-type components, two l = 2 components, and one l = 3 component.
  • Clebsch-Gordan decomposition: Clebsch-Gordan coefficients convert the product basis of two irreps into coupled irreducible components labeled by L.Allowed degrees satisfy |l1−l2| ≤ L ≤ l1+l2, and the coupled product transforms as an L-type spherical tensor.
  • Clebsch-Gordan decomposition: Equivariance requires that transforming tensors before coupling equals coupling before transforming.This condition determines the Clebsch-Gordan coefficients up to normalization and phase convention.
  • Computational scaling: Naively evaluating all Clebsch-Gordan couplings up to degree L scales as O(L^6), creating a bottleneck in architectures that use them at every layer.An equivalent SO(2)-convolution formulation reduces the total complexity from O(L^6) to O(L^3).

VIII. ROTATIONAL EQUIVARIANCE IN DEEP LEARNING

Rotational equivariance in neural networks requires geometric features to transform consistently under SO(3), unlike permutation and translation handling that often follows directly from message-passing design. The tutorial surveys data augmentation and tensorial internal representations as approaches with different guarantees and information flow.

  • Architectural requirement: Rotational equivariance requires every geometric feature to have a well-defined transformation behavior under SO(3).Permutation equivariance commonly follows from set aggregation, while translation invariance commonly uses relative coordinates and distances.
  • Data augmentation: Data augmentation works with any architecture and group but learns approximate equivariance without guaranteeing behavior out of distribution.Randomly transformed training samples provide the symmetry examples from which the model must learn.
  • Tensorial representations: Invariant-input methods use distances, angles, or dihedral angles with standard network layers to achieve rotational invariance.These methods are described as the historically first approach to exact rotational invariance.

A. Equivariance via Group Convolutions

Group convolutions achieve equivariance by lifting signals to group-indexed feature maps and then correlating over group coordinates. Their explicit group index supports standard channel-wise nonlinearities but introduces resolution-dependent cost and 3D discretization trade-offs.

  • Lifting: A lifting layer maps an input signal on X to a group-indexed feature map by evaluating transformed frames across g ∈ G.The output transforms under the left regular action on the group index.
  • Lifting equivariance: Transforming the input before lifting is equivalent to lifting first and re-indexing the group coordinate afterward.The re-indexing takes the form [LuF](g) = F(u^-1g).
  • Group convolution: Group convolution generalizes translation correlation by using kernels that depend on the relative group element g^-1h.The operation is equivariant under the left regular action.
  • Spatial-group convolution: Joint spatial-group convolutions use relative spatial displacement x−y and relative group displacement g^-1h.This construction is used in roto-translation equivariant SE(2) CNNs.
  • Nonlinearities and readout: Because the left regular action only re-indexes group coordinates, channel-wise nonlinearities commute with it and can alternate with group correlations.The group index is typically removed at the end by pooling such as averaging.
  • Practical limitations: Feature-map cost scales with |G|, while finer 3D discretizations improve full-group approximation but increase runtime and memory.Exact finite closed 3D rotation subgroups beyond planar dihedral groups are limited to Platonic-solid symmetry groups.

B. Equivariance via Internal Tensorial Representations

Tensor field networks enforce exact SO(3) equivariance by representing internal features as irreducible tensor channels and composing equivariant operations. Their message passing uses spherical harmonics, Wigner-D transformations, and Clebsch-Gordan products, while specialized nonlinearities and normalization preserve equivariance.

  • Tensor field networks represent internal features as direct sums of SO(3) irreducible representations and compose equivariant building blocks.
  • Messages couple tensorial node features with tensorial convolution filters through Clebsch-Gordan products.Filters combine learnable scalar radial functions with angular spherical harmonics evaluated from relative directions.
  • Equivariant linear layers mix channels within each degree l but cannot mix different representation degrees.The same weight matrix is applied to every irrep component, preserving the SO(3) action on the component index.
  • Gated nonlinearities use scalar features to modulate higher-degree tensor features, while invariant-statistics normalization avoids direction-dependent biases.Scalar nonlinearities are standard for l = 0; for l > 0, scalar gates preserve equivariance when scaling tensor components.
  • Steerable CNNs apply the same representation-constrained principle to regular Cartesian grids, differing mainly in their kernel implementation and data domain.
  • Alternative architectures use projective or Clifford algebra, or Cartesian tensors that avoid costly Clebsch-Gordan tensor products.

C. Canonicalization-Based Approaches

Canonicalization achieves exact equivariance by predicting reference frames, transforming inputs into those frames, and applying standard backbones. Global frames remove overall orientation, whereas local frames additionally capture local symmetries but introduce communication and continuity challenges.

  • Canonicalization provides an alternative route to exact equivariance that can use non-equivariant backbones without specialized equivariant building blocks.
  • Group averaging applies shared backbones to transformed inputs and averages outputs, but becomes expensive for large or continuous groups.Continuous-group averages require approximation by a finite transformation set.
  • Local canonicalization can capture local symmetries that global canonicalization cannot, potentially improving learning and generalization.
  • Global canonicalization predicts one equivariant orientation, transforms the input into that pose, and then processes it with a standard backbone.
  • Local canonicalization predicts an equivariant frame for each node, making local coordinates independent of global orientation and similar local motifs similarly represented.
  • Different local frames create a communication barrier because identical geometric objects acquire different coordinates, limiting faithful message passing between nodes.Exact frame-to-frame transitions have been proposed to enable directed tensor-valued communication.
  • Canonicalization frames can be discontinuous or ill-conditioned near symmetric geometries because small perturbations may cause finite orientation changes.
  • The five approaches differ in approximation, architectural flexibility, computational cost, and ability to capture local symmetries.Data augmentation is approximate; group convolutions and tensorial representations are exact but specialized; global and local canonicalization trade architectural constraints for frame-prediction issues.

D. Conceptual Links between Different Approaches

Canonicalization and tensorial equivariant architectures can be understood through a shared scalarization–tensorization lens. Both transform geometric information into invariant quantities for processing and then reconstruct equivariant outputs, but they use different bases and internal pathways.

  • Global canonicalization becomes a special case of local canonicalization when all local frames are identical.
  • Geometry-adapted local frames can improve performance when explicitly exploited, creating a trade-off between robustness to frame choice and extracted geometric information.
  • Canonicalization scalarizes features by projecting them into equivariantly predicted local frames, processes invariant coordinates, and tensorizes outputs back into the global frame.
  • For vector outputs, predicted basis vectors combine scalar coefficients to produce outputs that transform equivariantly under global rotations.
  • The same scalarization and tensorization construction extends from vectors to rank-2 and higher-order Cartesian tensors.Higher-order tensors are formed by contracting indices with local basis vectors and expanding back with the same basis.
  • In e3nn-style architectures, scalar Clebsch-Gordan paths project tensor features onto spherical-harmonic bases, while later products reconstruct spherical tensors.
  • Canonicalization uses a complete orthonormal local basis, whereas e3nn-style projection is defined by the relative direction and spherical-harmonic basis.

IX. PROBLEMS WITH ROTATIONAL EQUIVARIANCE IN PRACTICE

Rotationally equivariant architectures offer physical guarantees but often incur computational and optimization costs. Evidence on data efficiency and scaling is mixed, so the tutorial argues against universal prescriptions and favors problem-dependent evaluation.

  • Equivariant architectures can impose substantial runtime overhead because their specialized operations align poorly with optimized dense-matrix hardware primitives.
  • Exact-equivariance constraints complicate optimization, with less standardized training recipes and reported difficulty scaling models successfully.The added compute, engineering complexity, and tuning burden can limit both industrial development and academic projects.
  • Relaxing or removing built-in symmetries has re-emerged in practice because standard components offer greater architectural freedom and efficiency.The tutorial cites AlphaFold 3 as a prominent example of omitting global rotational and translational symmetry constraints.
  • Some recent works suggest exact symmetry enforcement may hinder optimal scaling as more data becomes available, but this remains a hypothesis.
  • Built-in symmetry has been reported to improve data efficiency, while other evidence shows data augmentation can achieve better low-data accuracy than local canonicalization.Local canonicalization may instead produce a steeper performance-versus-data curve, so scaling slope and small-data accuracy can diverge.
  • The tutorial concludes that the value of exact symmetry depends on the problem, dataset size, target quantity, and optimization landscape.

X. CONCLUSION

Rotational equivariance offers a principled encoding of coordinate independence for 3D machine learning, connecting foundational mathematics to practical architectures. The tutorial also frames equivariant model design as a choice among approaches with distinct trade-offs.

  • The tutorial connects symmetry, geometric deep learning, group representations, spherical harmonics, Wigner matrices, tensor products, and Clebsch-Gordan decomposition to modern equivariant learning.It relates these ingredients across mathematical and architectural formalisms.
  • Understanding rotational equivariance informs practical model design by clarifying how architectural choices encode geometric structure.
  • Different ways of incorporating geometric structure trade off expressivity, computational efficiency, implementation complexity, and numerical stability.The discussion supports choosing between exact symmetry and approximate methods according to application needs.
Loading 2608.31045v1…