Efficient construction and explicit dimensionality of Lie group-equivariant and permutation-invariant spaces
math.NA, cs.NA, quant-ph
Submitted: 2026-04-02
Updated: 2026-11-05
Comments: 42 pages, 9 figures, 5 tables
Code: https://github.com/ACEsuit/EquivariantTensors
License: http://creativecommons.org/licenses/by/4.0/
The gist: We give a practical construction of functions of N variables that are simultaneously equivariant under a group action and invariant under permutations, or simply group-equivariant…
Terminology
Abstract
We give a practical construction of functions of N variables that are simultaneously equivariant under a group action and invariant under permutations, or simply group-equivariant, valid for any finite-dimensional space stable under that action. The method applies to any connected linear Lie group: it uses the Lie algebra to build a matrix M whose kernel corresponds exactly to the subspace with the desired symmetry properties, removing the need for Clebsch-Gordan coefficients and explicit symmetrization over permutations. For the groups SO(3) and SU(2), the special structure of the matrix M yields a linear scaling method which significantly outperforms the typically exponential cost of existing methods in the literature. As a byproduct, we obtain the exact dimension of group-equivariant and permutation-invariant spaces. This allows us to compare dimensionalities and show that, for large N, this dimension is of comparable order to that of the permutation-invariant space alone, while pre-asymptotically it is orders of magnitude lower. Enforcing group-equivariance explicitly, on top of permutation-invariance, therefore yields a substantial computational gain when approximating such functions.
Sources
- e3nn: Euclidean Neural Networks
- Equivariant Electronic Hamiltonian Prediction with Many-Body Message Passing
- A symmetry-preserving and transferable representation for learning the Kohn-Sham density matrix
Related papers
- Do physics-informed neural networks (PINNs) need to be deep? Shallow PINNs using the Levenberg-Marquardt algorithm
- A Neural-preconditioned Poisson Solver for Mixed Dirichlet and Neumann Boundary Conditions
- Advanced Linear Algebra with Applications - Part I (Numerical linear algebra for PDEs, machine learning, and data assimilation)
- Operator learning meets inverse problems: A probabilistic perspective
- Second-order consistency for learning chaotic dynamics via randomized Jacobian matching
- Windowed thinning and query complexity for the bouncy particle and Zigzag samplers