Group Equivariant Convolutions

• Group equivariant convolutions are deep learning layers that generalize classical convolutions by ensuring feature transformations remain consistent under arbitrary symmetry groups.
• They combine rigorous group theory with efficient implementations, including regular, steerable, and PDE-based methods, to achieve significant parameter and data efficiency.
• Their application to images, 3D shapes, graphs, and physical fields illustrates their pivotal role in developing robust, symmetry-aware neural network architectures.

Group equivariant convolutions generalize the classical convolution operation to respect the symmetries of arbitrary groups acting on input data. By enforcing equivariance with respect to a specified group $G$, these layers guarantee that transformations of the input by $G$ induce predictable and consistent transformations in the feature maps throughout the network. This design yields highly parameter-efficient, data-efficient, and generalizable deep learning architectures, especially for structured domains such as images, 3D shapes, graphs, and physical fields, where symmetry plays a critical semantic role. Group equivariant convolutions underpin a significant portion of modern geometric deep learning, enabling neural networks to natively exploit translation, rotation, permutation, scaling, and even more general continuous or discrete symmetry groups (Weijler et al., 11 Feb 2025, Basheer et al., 2024, Cohen et al., 2016, Cohen et al., 2018, Smets et al., 2020).

1. Mathematical Framework of Group Equivariant Convolution

Let $G$ be a (locally compact) group, and let $f: G \to \mathbb{R}^C$ be a feature map defined on $G$, with Haar measure $d\mu$. A group equivariant convolution is defined as

$$(f * k)(g) = \int_{h \in G} f(h)\; k(h^{-1} g)\; d\mu(h)$$

where $$k: G \rightarrow \mathbb{R}^{C_\mathrm{out} \times C_\mathrm{in}}$$ is a learnable kernel. The fundamental property is $G$-equivariance: left translation $L_{g_0} f(h) = f(g_0^{-1} h)$ commutes with this operation, so the output transforms under $G$0 in the same way as the input. For discrete groups, the integral becomes a sum; for standard $G$1 convolutions (i.e., translations), this reduces to the classical convolution formula (Cohen et al., 2016, Cohen et al., 2018, Esteves, 2020).

When extending the input domain to more general homogeneous spaces $G$2 (e.g., spheres $G$3, point clouds, or graphs), feature maps are interpreted as sections of bundles associated to representations of $G$4. The equivariance constraint on kernels formalizes to a bi-equivariance condition: $G$5 where $G$6 and $G$7 are representations for the feature spaces at input and output (Cohen et al., 2018, Lang et al., 2020, Esteves, 2020).

2. Layer Types: Regular, Steerable, and PDE-Based Group Convolutions

Group equivariant convolutional layers are typically classified as follows (Basheer et al., 2024, Azar et al., 2020):

• Regular group convolutions: Feature maps and filters are defined as functions on $G$8 or on $G$9, and symmetry is enforced by convolution over the group or coset structure. Classical examples include G-CNNs equivariant to rotations/reflections/translations (Cohen et al., 2016, Cohen et al., 2018).
• Steerable group convolutions: Feature channels are organized into fields transforming under specific representations. Kernels must satisfy an intertwining constraint relative to the stabilizer subgroup, leading to a solution parameterized in terms of harmonic bases and Clebsch–Gordan coefficients (Weiler et al., 2019, Lang et al., 2020). This includes spherical CNNs, tensor field networks, and SE(3)-equivariant networks.
• PDE-based group convolutions: Layers are constructed as PDE-solvers, where the symmetry is built into a parameterized family of (fractional) diffusion, convection, dilation, and erosion operators on homogeneous spaces. Nonlinearities such as pooling and ReLU can be subsumed into morphological PDE layers, as in the PDE-G-CNN framework (Smets et al., 2020).

The table below summarizes representative architectures and their symmetry structure from the theory of (Cohen et al., 2018):

Model	Group $G$0	Stabilizer $G$1	Base Space $G$2	Feature Type
Classic CNN	$G$3	$G$4	$G$5	Regular
p4-CNN	$G$6	$G$7	$G$8	Regular, irreps
Steerable CNN (E(2))	$G$9	$f: G \to \mathbb{R}^C$0	$f: G \to \mathbb{R}^C$1	Vector, tensor, ...
Spherical CNN	$f: G \to \mathbb{R}^C$2	$f: G \to \mathbb{R}^C$3	$f: G \to \mathbb{R}^C$4	Spherical harmonics
3D Steerable CNN	$f: G \to \mathbb{R}^C$5	$f: G \to \mathbb{R}^C$6	$f: G \to \mathbb{R}^C$7	SO(3) irreps

3. Algorithmic Construction and Practical Implementation

Standard G-convolutions require precomputing filter banks over the action of $f: G \to \mathbb{R}^C$8. For semidirect product groups ($f: G \to \mathbb{R}^C$9), the group structure allows efficient implementations by decomposing a transformation into a spatial translation and an element of the stabilizer (e.g., rotation/reflection). This structure supports weight sharing across group elements and ensures that the parameter count does not increase with the group size (Cohen et al., 2016, Lengyel et al., 2021).

Steerable convolutions enforce constraints on kernels so that $G$0 for all $G$1, thus ensuring compatibility of the angular frequency content of the filter with the representations. The general characterization of such kernel spaces is given by a Wigner–Eckart-type theorem, parameterizing all possible G-steerable filters in terms of harmonic basis functions and Clebsch–Gordan coefficients (Lang et al., 2020, Weiler et al., 2019).

PDE-based approaches encode symmetry by evolving feature maps under symmetry-invariant PDEs on the underlying homogeneous space. The entire structure of nonlinearity and pooling can be realized as solutions to group-invariant PDEs—empirically, this achieves orders-of-magnitude parameter savings (Smets et al., 2020).

Depthwise-separable G-conv variants exploit redundancies induced by group symmetry in learned filters, factorizing spatial and groupwise kernels to reduce computational cost while maintaining exact equivariance (Lengyel et al., 2021).

4. Advanced Methods: Local, Continuous, and Attention-Based Group Convolutions

For groups with high intrinsic dimension (e.g., $G$2 on 3D point clouds), a direct lift to the entire group may be computationally prohibitive. Efficient continuous and local group convolutions have been introduced:

• Efficient Continuous SE(3) Group Convolutions: By constructing a small, per-point, SE(3)-equivariant frame using PCA-based local reference frames, and restricting the group convolution to these local frames, one can achieve exact local SE(3) equivariance with negligible computational overhead. The kernel is parameterized as an MLP over the relative position and rotation, allowing continuous, data-adaptive filtering (Weijler et al., 11 Feb 2025).
• Stochastic Frame Sampling: During training, a random subset of local frames may be used at each step to regularize computation, recovering full equivariance by averaging over frames at test time.
• Attentive Group Equivariant Convolutions: These architectures learn a data-dependent attention kernel over group elements, selectively accentuating plausible symmetry combinations during convolution. The equivariance property is preserved if the attention map satisfies suitable transformation constraints (Romero et al., 2020).
• Variational Partial Group Convolutions: To address the rigidity of full group equivariance, VP G-CNNs learn input-dependent and variationally regularized distributions over output group elements, achieving partial equivariance that adapts to the local symmetry structure of the data (Kim et al., 2024).

5. Applications, Empirical Results, and Complexity Considerations

Group equivariant convolutions are empirically validated across a wide spectrum of domains requiring symmetry-awareness, including:

• Image classification and segmentation: G-CNNs and their variants consistently outperform vanilla CNNs on rotated and reflected data, e.g., reducing rotated MNIST error from 5.03% (CNN) to 2.28% (p4-CNN) (Cohen et al., 2016). Separable GConvs further improve parameter efficiency and performance (Lengyel et al., 2021).
• 3D point cloud and scene understanding: The local SE(3) equivariant convolution achieves mean 86.9% accuracy on ModelNet40 under $G$3 transformation, compared to ~12% for non-equivariant networks, with minimal overhead versus tens of times the compute and memory cost for global or discrete methods (Weijler et al., 11 Feb 2025).
• Pose estimation and part segmentation: Continuous local equivariant methods outperform discrete approaches, yielding orders of magnitude improvements in pose error (mean ~4e-5° vs ~1°) and marked gains in semantic segmentation metrics (Weijler et al., 11 Feb 2025).
• Partial differential equations (PDEs): G-equvariant Fourier Neural Operators (Helwig et al., 2023) and PDE-G-CNNs (Smets et al., 2020) demonstrate superior generalization and parameter efficiency in learning PDE solution operators with built-in physical symmetries.

The following table summarizes memory and speed trade-offs for 3D point cloud group convolution variants, as measured on $G$4, $G$5 (Weijler et al., 11 Feb 2025):

Method	Memory (MB)	FPS
Standard 3D Conv	37	704
Local SE(3), $G$6	37	581
Local SE(3), $G$7	77	433
E2PN (discrete)	1,212	45
EPN (discrete)	1,636	10

Global equivariant and discretized local group convolutions incur prohibitive compute and memory costs when $G$8 is large (e.g., SO(3) discretized at $G$9 orientations yields a $d\mu$0 memory and $d\mu$1 compute penalty).

6. Theoretical Guarantees and Quantum Extensions

A central result is that every $d\mu$2-equivariant linear map between feature spaces (possibly defined on homogeneous spaces) is realized by a group convolution (possibly twisted by bi-equivariant kernels) (Cohen et al., 2018, Esteves, 2020). The exact structure of admissible kernels is characterized by representation-theoretic constraints, and minimal parameterizations are governed by the Wigner–Eckart theorem (Lang et al., 2020).

Quantum algorithms have been developed for group convolutions and cross-correlations on finite groups, providing exponential speedup (logarithmic in $d\mu$3) when quantum oracles are available for the data and filters. These results generalize circulant quantum solvers to arbitrary groups and extend immediately to group-equivariant neural network layers (Castelazo et al., 2021).

7. Limitations, Open Problems, and Future Directions

Outstanding challenges in the development and application of group equivariant convolutions include:

• Scalability to large or continuous groups: Exact global equivariance remains computationally intensive in high dimensions. Local or stochastic sampling alleviates overhead but may introduce locality constraints (Weijler et al., 11 Feb 2025).
• Partial or approximate equivariance: Real-world data often exhibit only partial symmetry. VP G-CNNs and related frameworks address flexibility, but robust, scalable solutions for partial equivariance are nascent (Kim et al., 2024).
• Generalization to non-Euclidean and evolving domains, such as graphs with dynamic topology or manifolds with gauge symmetries, is an open area where new architectures are being developed (Basheer et al., 2024).
• Basis construction for general compact groups and field types: While Wigner–Eckart parameterizations are known for standard groups, efficient basis computation for arbitrary settings and irregular representations remains challenging (Lang et al., 2020).
• Integration of alternative algebraic frameworks, such as Clifford algebras and category-theoretic generalizations, is a promising but largely unexplored direction (Basheer et al., 2024).

Recent work emphasizes the need for richer real-world datasets, improved transferability across domains, and further unification of algebraic, geometric, and machine learning perspectives to realize universally symmetry-aware neural architectures.