[Paper Review] A Wigner-Eckart Theorem for Group Equivariant Convolution Kernels

Leon Lang, Maurice Weiler|arXiv (Cornell University)|Oct 21, 2020

Seismic Imaging and Inversion TechniquesEarth and Planetary Sciences31 references21 citations

TL;DR

This paper establishes a general characterization of G-steerable convolution kernels for group equivariant convolutional networks (GCNNs) using a generalized Wigner-Eckart theorem. It proves that the space of G-steerable kernels is fully parameterized by generalized reduced matrix elements, Clebsch-Gordan coefficients, and harmonic basis functions on homogeneous spaces, providing a complete solution for arbitrary compact groups G.

ABSTRACT

Group equivariant convolutional networks (GCNNs) endow classical convolutional networks with additional symmetry priors, which can lead to a considerably improved performance. Recent advances in the theoretical description of GCNNs revealed that such models can generally be understood as performing convolutions with G-steerable kernels, that is, kernels that satisfy an equivariance constraint themselves. While the G-steerability constraint has been derived, it has to date only been solved for specific use cases - a general characterization of G-steerable kernel spaces is still missing. This work provides such a characterization for the practically relevant case of G being any compact group. Our investigation is motivated by a striking analogy between the constraints underlying steerable kernels on the one hand and spherical tensor operators from quantum mechanics on the other hand. By generalizing the famous Wigner-Eckart theorem for spherical tensor operators, we prove that steerable kernel spaces are fully understood and parameterized in terms of 1) generalized reduced matrix elements, 2) Clebsch-Gordan coefficients, and 3) harmonic basis functions on homogeneous spaces.

Motivation & Objective

• To provide a general characterization of G-steerable kernel spaces for group equivariant convolutional networks (GCNNs), which had previously only been solved for specific cases.
• To establish a formal analogy between the constraints on steerable kernels and spherical tensor operators in quantum mechanics, leveraging representation theory.
• To generalize the Wigner-Eckart theorem to the setting of group-equivariant deep learning, enabling a unified framework for kernel parameterization.
• To reduce the problem of finding G-steerable kernels to computing Clebsch-Gordan coefficients, harmonic basis functions, and endomorphisms—key components in representation theory.

Proposed method

• Formalize the G-steerability constraint on convolution kernels as a representation-theoretic condition, analogous to the transformation laws of quantum mechanical operators.
• Apply the Wigner-Eckart theorem in a generalized form to decompose the space of G-steerable kernels into irreducible components using generalized reduced matrix elements.
• Use Clebsch-Gordan coefficients to couple irreducible representations of the group G, enabling the construction of invariant kernels under G-action.
• Parameterize the kernel space using harmonic basis functions on the homogeneous space H/G, which encode the spatial structure of the convolution.
• Reduce the general solution to three fundamental ingredients: reduced matrix elements (degrees of freedom), Clebsch-Gordan coefficients (coupling rules), and harmonic basis functions (spatial modes).
• Establish the role of endomorphisms in the kernel space, particularly in the case of reducible representations, and justify their necessity via representation theory.

Experimental results

Research questions

• RQ1 How can the space of G-steerable convolution kernels be fully characterized for arbitrary compact groups G?
• RQ2 What is the general mathematical structure underlying G-steerable kernels, and how does it relate to quantum mechanical tensor operators?
• RQ3 Can the Wigner-Eckart theorem be generalized to provide a complete parameterization of G-steerable kernels?
• RQ4 What are the minimal ingredients required to construct all possible G-steerable kernels, and how do they relate to representation theory?
• RQ5 How does the solution generalize from irreducible to reducible group representations in the context of GCNNs?

Key findings

• The space of G-steerable kernels for any compact group G is completely characterized and parameterized by generalized reduced matrix elements, Clebsch-Gordan coefficients, and harmonic basis functions on homogeneous spaces.
• The solution generalizes the Wigner-Eckart theorem to the setting of group-equivariant deep learning, providing a rigorous theoretical foundation for kernel design in GCNNs.
• The parameterization reduces the problem of finding G-steerable kernels to computing three well-known objects from representation theory: reduced matrix elements, Clebsch-Gordan coefficients, and harmonic basis functions.
• The framework applies uniformly to all compact groups, including O(3) and SU(N), and enables systematic construction of equivariant layers beyond specific cases.
• The necessity of considering endomorphisms in the kernel space is formally justified, particularly when dealing with reducible representations, ensuring completeness of the solution.
• The work resolves a long-standing open problem in GCNN theory by providing a general solution where only specific cases had been solved previously.