🤖 machine learning

Bases of Steerable Kernels for Equivariant CNNs: From 2D Rotations to the Lorentz Group

This paper presents a general and accessible method for constructing explicit real and complex bases of steerable kernels for equivariant CNNs across various symmetry groups and tensor types by leveraging a point-wise invariance condition and the steerability equation, thereby eliminating the need to compute Clebsch-Gordan coefficients.

Original authors: Alan Garbarz

Published 2026-03-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alan Garbarz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to recognize objects in a room. If you show the robot a picture of a cat, and then you rotate the picture 90 degrees, the robot should still know it's a cat. But if the robot is just a "vanilla" neural network, it has to relearn what a cat looks like from scratch every time you turn the picture. That's inefficient.

Steerable Equivariant CNNs are like training that robot with a built-in rulebook: "If I rotate the input, the output must rotate in the exact same way." This makes the robot much smarter and faster to train.

However, there's a catch. To make this work, the "brain" of the robot (the mathematical kernel) has to follow a very strict, complex set of rules. Usually, figuring out these rules is like trying to solve a giant, messy puzzle using a complicated instruction manual called Clebsch-Gordan coefficients. It's hard, computationally expensive, and requires deep knowledge of advanced math.

This paper presents a much simpler, more direct way to solve that puzzle.

Here is the breakdown of the author's idea using everyday analogies:

1. The Problem: The "Rotating Room"

Imagine you are in a room with a specific object on a table (the input). You want to describe how that object looks from every possible angle in the room.

The Old Way: You try to calculate the view from every single angle simultaneously by solving a massive equation that connects all angles at once. It's like trying to predict the weather for the whole planet by solving one giant equation for every atom in the atmosphere.
The Paper's Way: The author says, "Let's just look at one specific spot first."

2. The Strategy: The "Anchor Point"

The author's method is like this:

Pick a "Home Base": Choose one specific point in the room (let's call it $x_0$ ) and one specific angle (say, facing North).
Solve the Easy Part: Ask, "What does the rule look like if I am standing right here, facing North?" Because you are in a fixed spot, the rules are much simpler. You just need to find a few basic shapes or patterns that stay the same (invariant) when you wiggle slightly.
The "Steering" Move: Once you have the solution for that one spot, you don't need to solve the math again for the rest of the room. You simply rotate (or "steer") your solution to match any other angle.
- Analogy: Imagine you have a stencil of a star. If you want to know what the star looks like rotated 45 degrees, you don't need to draw a new star from scratch. You just take your original stencil and rotate the paper. The paper does the work for you.

3. Why This is a Big Deal

No More "Magic Numbers": The old methods required calculating Clebsch-Gordan coefficients. Think of these as a massive, pre-calculated dictionary of how different shapes combine when rotated. Looking them up is slow and hard to generate for new shapes. The new method bypasses this dictionary entirely.
Direct Translation: Instead of translating "Shape A" into "Shape B" via a complex middleman, the author's method translates "Shape A" directly into "Shape B" using the rotation rules themselves.
Ready-to-Use: The paper doesn't just give a theory; it provides the actual "stencils" (bases) for many different types of symmetry, including:
- 2D Rotations: Like spinning a photo on a table.
- 3D Rotations: Like spinning a globe.
- The Lorentz Group: This is the tricky one. It's the math behind Einstein's Special Relativity. It describes how space and time mix when you move at near-light speed. The paper shows how to build these "steerable" rules for particles moving at relativistic speeds, which is huge for physics simulations.

4. The "Massive" vs. "Massless" Particles

The paper also explains how to handle two types of particles in physics:

Massive Particles (like electrons): They have a "rest frame" (you can sit still next to them). The math here is like rotating a ball; it's stable.
Massless Particles (like light): They always move at the speed of light. They don't have a "rest frame." The math here is trickier because the "direction" of the light is the only thing that matters. The author shows how to build the rules for these "light-speed" particles using a special kind of projector that filters out the noise, ensuring the rules work no matter how fast the observer is moving.

Summary

In short, this paper is a shortcut.

Instead of forcing a computer to solve a complex, global puzzle to understand how to rotate data, the author says: "Solve it for one easy spot, then just rotate the answer."

This makes it possible to build AI that understands 3D space, 4D spacetime, and relativistic physics much more efficiently, without needing to be a PhD in group theory to set it up. It turns a mountain of complex math into a simple, reusable toolkit.

1. Problem Statement

The paper addresses the design of Steerable Equivariant Convolutional Neural Networks (CNNs). In these networks, feature maps transform according to specific representations of a symmetry group $G$ (e.g., rotations, Lorentz boosts). For a convolutional layer to be equivariant, its kernel $K$ must satisfy a specific steerability constraint:
$K(g \cdot x) = \frac{1}{|\det g|} \rho_{\text{out}}(g) K(x) \rho_{\text{in}}(g)^{-1}$
where $\rho_{\text{in}}$ and $\rho_{\text{out}}$ are the representations of the input and output feature maps.

The Challenge:
Finding a basis of solutions for this constraint for arbitrary groups and tensor types is mathematically complex. Existing methods typically rely on:

Decomposing tensor products into irreducible representations (coupled basis).
Computing Clebsch-Gordan (CG) coefficients to transform between uncoupled and coupled bases.
Using harmonic basis functions on homogeneous spaces.

These approaches are computationally expensive (especially for non-compact groups like the Lorentz group) and require deep knowledge of representation theory, creating a barrier for practical implementation in deep learning.

2. Methodology

The author proposes a direct, geometric strategy to solve the steerability constraint without pre-computing CG coefficients or changing to a coupled basis. The core insight is to reduce the problem to finding intertwiners (homomorphisms) at a single point $x_0$ and then "steering" them to the rest of the orbit.

The Algorithm:

Select a Reference Point: Choose a point $x_0$ on a specific orbit of the group $G$ .
Identify the Stabilizer: Determine the isotropy subgroup $H = \text{Stab}(x_0)$ , which leaves $x_0$ invariant.
Solve the Local Constraint: Find the space of linear maps $K(x_0): V_{\text{in}} \to V_{\text{out}}$ $K (x_{0}) : V_{in} \to V_{out}$ that commute with the action of $H$ $H$ .
- The condition becomes: $K(x_0) = \rho_{\text{out}}(h) K(x_0) \rho_{\text{in}}(h)^{-1}$ for all $h \in H$ .
- Since $H$ is often a simpler group (e.g., $SO(2)$ inside $SO(3)$), the representations $\rho|_H$ decompose into direct sums of irreps.
- By Schur's Lemma, the solution space consists of block-diagonal matrices where non-zero blocks only exist between isomorphic irreducible subspaces of the input and output.
Steer to the Orbit: Once the basis of solutions at $x_0$ is found, the kernel at any arbitrary point $x = g \cdot x_0$ is generated by applying the group action:
$K(x) = \rho_{\text{out}}(g) K(x_0) \rho_{\text{in}}(g)^{-1}$
This step naturally generates the required functional dependence (e.g., spherical harmonics or Lorentz boost factors) without explicit CG coefficient calculation.

3. Key Contributions

Elimination of Clebsch-Gordan Coefficients: The method bypasses the need to compute or store CG coefficients, which are difficult to derive for non-compact groups and computationally heavy for large dimensions.
Unified Framework: It provides a single, consistent strategy applicable to both compact groups (SO(2), O(2), SO(3), O(3)) and non-compact groups (Lorentz group $SO^+(1,3)$ ).
Explicit Real and Complex Bases: The paper derives ready-to-use bases for kernels acting on feature maps of arbitrary tensor types (scalars, vectors, tensors, spinors) for the groups listed above.
Geometric Interpretation: It reframes steerability as a projection problem. For integer spins, the intertwiners at $x_0$ act as projectors onto invariant subspaces (e.g., projecting a 4-vector onto its time-like or space-like components).

4. Key Results and Examples

The paper details the derivation of bases for several groups:

SO(2) & O(2):
- For 2D rotations, the stabilizer is trivial (identity), making the local constraint trivial. The steering operation naturally generates trigonometric functions ( $\cos, \sin$ ) corresponding to the rotation angles.
- For O(2) (including reflections), the stabilizer includes a reflection matrix, which imposes parity constraints on the kernel components, reducing the solution space.
SO(3) & O(3):
- The stabilizer of a point on a sphere is $SO(2)$. The method decomposes the 3D representations into $SO(2)$ irreps (labeled by magnetic quantum numbers $m$ ).
- The solution space is determined by matching $m$ values in the input and output. The resulting kernels are expressed directly in terms of Wigner D-matrices (or real spherical harmonics), avoiding the explicit use of 3j/6j symbols.
Lorentz Group $SO^+(1,3)$ :
- Massive Particles (Timelike orbits): The stabilizer is $SO(3)$. The method decomposes Lorentz representations $(j_L, j_R)$ $(j_{L}, j_{R})$ into $SO(3)$ spins. The kernels are constructed by steering $SO(3)$ projectors (e.g., projecting onto the subspace orthogonal to the 4-velocity $u^\mu$ $u^{μ}$ ).
  - Result: Explicit formulas for kernels acting on 4-vectors, rank-2 tensors, and spinors, expressed using the projection tensor $\Delta^{\mu}_{\nu} = \delta^{\mu}_{\nu} - u^\mu u_\nu$ .
- Massless Particles (Lightlike orbits): The stabilizer is the Euclidean group $ISO(2)$. The paper simplifies this by considering only the $SO(2)$ subgroup (ignoring the non-semisimple translation part, which is physically irrelevant for standard particles).
  - Result: Kernels are constructed using a transverse projector $\Delta^{\mu\nu} = \eta^{\mu\nu} - n^\mu \bar{n}^\nu - n^\nu \bar{n}^\mu$ , where $n$ is the null direction.

5. Significance and Impact

Computational Efficiency: By avoiding the coupled basis transformation, the method reduces the computational overhead associated with generating equivariant layers, particularly for high-rank tensors and non-compact groups.
Accessibility: The approach relies on basic linear algebra (Schur's Lemma) and group actions rather than advanced representation theory, making it accessible to a broader audience of machine learning practitioners.
Flexibility: The method allows for independent cutoffs on input and output feature map irreps, potentially leading to more expressive and flexible network architectures compared to fixed coupled-basis approaches.
Relativistic Physics Applications: The explicit construction of Lorentz-steerable kernels enables the design of CNNs for relativistic systems (e.g., particle physics simulations, cosmological data analysis) where standard Euclidean equivariant CNNs fail.

In conclusion, Garbarz presents a robust, generalizable, and computationally efficient framework for constructing steerable kernels, bridging the gap between abstract group theory and practical deep learning implementation for both Euclidean and relativistic symmetries.