Source-linked AI summary
Piecewise Linear Equivariant Maps for Compact Groups
Valeriano Aiello
TL;DR
The paper asks which piecewise linear nonlinearities can be equivariant between finite-dimensional real representations of compact groups, motivated by equivariant neural networks. It develops a representation-theoretic structure theorem and shows that nonlinear behavior is confined to identity-component fixed subspaces, yielding an irreducible existence criterion with compact-group rigidity.
Problem
The paper investigates what piecewise linear equivariant maps can occur for compact-group representations, beyond basis-dependent componentwise activations and finite-group criteria.
Method
It analyzes arbitrary piecewise linear equivariant maps using invariant decompositions into identity-component fixed subspaces and orthogonal complements.
Results
For irreducible representations with non-trivial identity-component action, every equivariant piecewise linear map is affine, and a non-zero map exists precisely when L ∼= K or K = Rtriv.
Takeaways & Limitations
The identity component confines genuine piecewise linear nonlinearity to fixed-point subspaces, while the complementary action is rigid and linear.
Takeaways & Limitations
The framework assumes real, finite-dimensional, continuous representations of compact groups, and Haar averaging does not generally preserve piecewise linearity.
Abstract
from arXiv · showhide
Motivated by equivariant neural networks, we study piecewise linear equivariant maps between finite-dimensional real representations of compact groups. We show that all genuinely non-linear piecewise linear behaviour is confined to the subspaces on which the identity component of the group acts trivially, while equivariance forces linearity on the corresponding orthogonal complements. As a consequence, we obtain a compact-group analogue of the finite-group existence criterion of Gibson--Tubbenhauer--Williamson for non-zero equivariant piecewise linear maps between irreducible representations, with the identity component giving rise to a rigidity phenomenon absent from the finite-group case.
Introduction
The paper studies arbitrary piecewise linear equivariant maps between finite-dimensional real representations of compact groups, motivated by equivariant neural networks. It proves that the identity component imposes rigidity: nonlinear behavior is confined to fixed-point subspaces, while complementary components are forced to be linear.
- Motivation: Equivariant neural networks replace ordinary componentwise activations with nonlinear maps compatible with the group action, a compatibility that may fail for general representations.Piecewise linear maps are a natural class to study because they include componentwise ReLU networks.
- Problem: The finite-group kernel criterion does not extend unchanged to compact groups because the identity component introduces a rigidity phenomenon absent from finite groups.The compact-group setting is motivated by rotational symmetries such as SO(2) and SO(3).
- Structure theorem: The equivariant piecewise linear map space decomposes into a piecewise linear part on G◦-fixed subspaces and a linear equivariant part on their orthogonal complements.The group action on the fixed-point spaces has finite image, so the nonlinear part is governed by a finite quotient.
- Rigidity: For connected compact groups, genuinely nonlinear dependence is confined to maximal trivial subrepresentations, while equivariance forces linearity on their orthogonal complements.Geometrically, connectedness prevents a finite orbit of non-affine hyperplanes unless the source dual representation has a nonzero fixed vector.
- Irreducible case: For non-trivial irreducible representations of a connected compact group, a nonzero equivariant piecewise linear map exists exactly when the representations are isomorphic.For arbitrary compact groups, if the identity component acts non-trivially, every such map is affine and nonzero existence occurs precisely when L ∼= K or K = Rtriv.
1. Preliminaries and notation
This section fixes the representation-theoretic and piecewise linear framework used throughout the paper. It also explains how these maps model equivariant neural-network layers while emphasizing that the paper studies basis-independent maps beyond permutation representations.
- Group representations: Throughout, G is a compact group with identity component G◦, and representations are real, finite-dimensional, and continuous.Compactness provides G-invariant inner products, so orthogonal complements of invariant subspaces remain invariant.
- Group representations: Two representations are equivalent when an intertwining linear isomorphism exists, and irreducibility means having no invariant subspaces other than zero and the whole representation.The one-dimensional trivial representation is denoted Rtriv.
- Piecewise linear maps: A map is piecewise linear when a finite polyhedral covering makes its restriction to each piece affine linear.The convention allows lower-dimensional polyhedral pieces, which is useful for restrictions to affine subspaces and faces.
- Piecewise linear maps: For domains equal to the whole ambient vector space, allowing lower-dimensional polyhedral pieces does not change the convention because full-dimensional members already cover the space.Piecewise linear maps are also continuous and are closed under suitable compositions, finite direct sums, and finite sums.
- Equivariant neural networks: In equivariant neural networks, layers are representations, linear maps are equivariant, and nonlinear operations must satisfy the same group-action compatibility.Equivariant affine transformations have equivariant linear parts and translation vectors fixed by the target representation.
- Equivariant neural networks: Componentwise piecewise linear activations are automatically compatible in permutation representations but require additional compatibility between the action, basis, and activation in general representations.The paper therefore studies arbitrary equivariant piecewise linear maps rather than only componentwise activations.
2. Main results
The paper establishes rigidity for piecewise linear equivariant maps under compact-group actions: nonlinearity is confined to identity-component fixed subspaces, while complementary components are forced to be affine or linear. For irreducible representations, this yields compact-group existence criteria that specialize to isomorphism in the connected non-trivial case and recover the finite-group kernel criterion when the identity component acts trivially.
- Rigidity mechanism: Connectedness makes every equivariant piecewise linear map affine linear when the source has no identity-component fixed vectors.The proof uses the connected identity component fixing each hyperplane in a finite orbit, which would produce a forbidden fixed dual vector otherwise.
- Structure theorem: The structure theorem confines genuinely non-linear behaviour to subspaces fixed pointwise by the identity component, while the orthogonal complements admit only linear intertwiners.The non-linear term factors through a finite quotient, whereas the complementary map is necessarily a linear equivariant map.
- Relation to finite groups: When the identity component acts trivially on the domain, the representation factors through a finite quotient and the finite-group kernel criterion is recovered.The resulting criterion is ker(ρL) ⊆ ker(ρK).
- Irreducible representations: For irreducible representations with non-trivial identity-component action on the domain, a non-zero map exists precisely when the representations are isomorphic or the codomain is the trivial representation.Every map has the form of a linear intertwiner plus a group-fixed constant; irreducibility forces either an isomorphism or a one-dimensional trivial codomain.
- Connected groups: For connected groups and non-trivial irreducible domain and codomain, a non-zero equivariant piecewise linear map exists if and only if the representations are isomorphic.The trivial-codomain alternative disappears when both representations are non-trivial.