🤖 machine learning

Group-Algebraic Tensors: Provably-optimal Equivariant Learning and Physical Symmetry Discovery

This paper introduces the $\star_G$ tensor algebra, a framework that transforms equivariance from an architectural constraint into an intrinsic algebraic property, enabling provably optimal symmetry-preserving tensor approximation, closed-form per-irreducible-representation predictions, and data-driven discovery of physical symmetries without requiring prior quantum mechanical knowledge.

Original authors: Paulina Hoyos, Shashanka Ubaru, Dongsung Huh, Vasileios Kalantzis, Kenneth L. Clarkson, Misha Kilmer, Haim Avron, Lior Horesh

Published 2026-05-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Paulina Hoyos, Shashanka Ubaru, Dongsung Huh, Vasileios Kalantzis, Kenneth L. Clarkson, Misha Kilmer, Haim Avron, Lior Horesh

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

The Core Idea: Changing the Rules of the Game

Imagine you have a complex origami crane. Traditional machine learning tries to study this crane by flattening it into a single sheet of paper. You can technically unfold it without tearing it, but you lose the 3D shape that gives it meaning. To understand the crane again, the computer has to painfully try to "fold" the paper back into a crane, guessing the structure every time.

Furthermore, if you rotate the crane, it's still the same crane. But standard computers don't "know" this; they see the rotated paper as a completely different, confusing object.

The authors of this paper propose a different approach. Instead of forcing the crane into a flat sheet and then trying to guess the folds, they invent a new kind of math (a new "language") where the rules of folding and rotating are built-in from the start.

They call this the $\star_G$ Tensor Algebra.

The Three Magic Pillars

The paper claims this new math rests on three solid, machine-verified foundations:

1. The Perfect Compression (The "Best Possible Fold")
In standard math, trying to compress a complex 3D object into a smaller version often involves guessing or settling for "good enough."

The Claim: The authors prove that their new math allows for a "perfect" compression. If you want to shrink the data while keeping its essential shape, their method finds the absolute best possible version, guaranteed by math.
The Analogy: Imagine trying to fit a suitcase into a tiny box. Standard methods might leave gaps or crush the clothes. This new method is like a magical vacuum seal that fits the clothes perfectly with zero wasted space, and they have a mathematical certificate proving it's the best fit possible.

2. The Lego Block of Symmetries (Mixing Rules)
Usually, if you want a computer to understand both rotation (spinning) and permutation (swapping parts), you have to build a custom, complicated machine for that specific mix. If you add a third rule, you have to rebuild the whole machine.

The Claim: Their new math is like a universal Lego set. You can snap different symmetry rules together (like rotation + swapping) just by saying "Group A times Group B." You don't need to redesign the machine; the math handles the combination automatically.
The Analogy: Instead of building a new car engine every time you want to add a radio or AC, you just plug in a standard module. The engine (the math) knows exactly how to run with the new part attached.

3. The Machine-Verified Blueprint

The Claim: They didn't just write code; they wrote a formal proof in a computer language called Lean 4. The computer checked every single step of their logic and found zero errors.
The Analogy: It's like an architect not just drawing a bridge, but having a super-computer simulate every wind, weight, and earthquake to prove the bridge cannot collapse before a single brick is laid.

What Can This New Math Do?

The paper demonstrates three specific superpowers that standard AI (like Neural Networks) cannot do:

1. Reading the "Symmetry" of Nature
The authors tested this on a dataset of molecules (QM9). They asked the math to look at the shapes of molecules and figure out the rules of physics governing them, without being told the rules.

The Result: The math successfully "discovered" the Wigner-Eckart selection rules. These are complex physics laws that say, for example: "To predict a molecule's size, you only need to look at its spherical shape. But to predict its magnetic direction, you must look at its directional shape."
The Analogy: Imagine showing a child a pile of different toys and asking them to guess the rules of how they move. The child (the AI) correctly guesses, "Round things roll, but things with handles need to be held." The paper claims their math did this for molecules, discovering deep physics laws just by looking at geometry.

2. Explaining Its Own Answers
Standard AI is often a "black box"—it gives an answer, but you don't know why.

The Result: Because this new math breaks data down into specific "channels" (like separating a song into bass, drums, and vocals), it can tell you exactly which part of the symmetry contributed to the answer.
The Analogy: If a standard AI says, "This molecule is toxic," it's a guess. This new math says, "This molecule is toxic because its 'bass' (scalar part) is high, but its 'drums' (vector part) are low." It gives a clear, mathematical recipe for the prediction.

3. Doing More with Less

The Result: On the molecular dataset, their method achieved high accuracy using 50 to 90 times fewer parameters (computer memory/settings) than standard neural networks.
The Analogy: Standard AI is like trying to solve a puzzle by buying a million random pieces and hoping they fit. This method is like having the picture on the box and only buying the exact pieces you need. It works better with less data and less computing power.

The Big Takeaway

The paper argues that we shouldn't force nature to fit into our computer's flat, rigid boxes. Instead, we should change our math to fit the natural, symmetrical shape of the data.

Old Way: Flatten the data, build a giant, complex neural network to guess the patterns, and hope it learns the rules of rotation and symmetry.
New Way ( $\star_G$ ): Build the rules of rotation and symmetry directly into the math. This makes the computer naturally understand the data, leading to perfect compression, guaranteed accuracy, and the ability to discover hidden physical laws on its own.

The authors conclude that this isn't just a "faster" version of existing AI; it is a different kind of tool that offers mathematical guarantees and interpretability that current AI simply cannot provide.

Technical Summary: Group-Algebraic Tensors for Provably-Optimal Equivariant Learning

Problem Statement

Scientific data, such as molecular configurations and quantum states, is inherently multidimensional and governed by physical symmetries (e.g., rotation, permutation). Traditional machine learning approaches typically vectorize this tensor data, collapsing its natural geometric structure into flat feature vectors. This process destroys the structural information that gives the data its meaning, forcing subsequent models to implicitly recover this structure at a high computational cost.

While Equivariant Neural Networks (ENNs) have successfully incorporated symmetry by engineering specific architectural constraints (e.g., rotation-equivariant layers), they face significant limitations:

Combinatorial Complexity: When multiple symmetries coexist, the network architecture must be redesigned from scratch for each new combination.
Lack of Optimality: Existing tensor decomposition methods (Tucker, CP, Tensor-Train) used for compression or approximation lack global optimality guarantees or are NP-hard to compute.
Interpretability: ENNs operate as black boxes; they do not provide a closed-form decomposition of predictions into specific symmetry channels (irreducible representations).
Discovery: Current frameworks cannot discover the underlying symmetry group of a dataset from data alone; the group must be specified manually.

Methodology: The $\star_G$ Tensor Algebra

The authors propose a paradigm shift: instead of constraining the architecture to fit the symmetry, they change the mathematics to fit the geometry of the data. They introduce the $\star_G$ tensor algebra, where a finite group $G$ defines the multiplication rule, making equivariance an intrinsic algebraic property rather than an architectural constraint.

Core Components

Algebraic Definition:
- For a finite group $G$ of order $n$ , a convolution tensor $T \in \mathbb{R}^{n \times n \times n}$ is defined based on the group multiplication table ( $T(a,b,c) = 1$ if $ab=c$, else 0).
- The $\star_G$ product of two tensors is defined via group convolution along the group dimensions. This product generalizes classical matrix algebra, embedding the symmetry group directly into the multiplicative structure.
- The algebra supports standard matrix-mimetic operations (inverses, transposes, norms) that inherently preserve equivariance.
Spectral Decomposition (Peter-Weyl Theorem):
- The convolution tensor admits a spectral decomposition $T(a,b,c) = \sum C(i,j,k) F_G(a,i) F_G(b,j) F_G^{-1}(c,k)$ , where $F_G$ is a generalized Group Fourier transform matrix constructed from the irreducible unitary representations (irreps) of $G$ .
- In the Fourier domain, group convolution becomes block-diagonal matrix multiplication, allowing for efficient computation: transform to Fourier domain $\to$ parallel matrix products per irrep $\to$ inverse transform.
$\star_G$ -SVD and Optimality:
- Every tensor in this algebra admits a Singular Value Decomposition (SVD): $A = U \star_G S \star_G V^H$ .
- Theorem 2.1 (Eckart–Young for $\star_G$ ): The rank- $k$ truncation $A_k$ minimizes the Frobenius norm error $\|A - B\|_F$ over all tensors $B$ of $\star_G$ -rank at most $k$ . This provides the first provably optimal symmetry-preserving tensor approximation.
- Unlike Tucker (quasi-optimal with $\sqrt{d}$ factor), CP (NP-hard), or Tensor-Train (no global optimality), $\star_G$ -SVD is exact and polynomial-time.
Composition of Symmetries:
- Theorem 2.2 (Product Groups): For a product group $G = G_1 \times \dots \times G_d$ , the convolution tensor and Fourier matrix factorize as Kronecker products ( $T_G = T_{G_1} \otimes \dots \otimes T_{G_d}$ ). This allows multiple symmetries to be composed without architectural redesign, simply by updating the Fourier transform matrix.

Key Contributions

Provably Optimal Compression: The paper establishes the first Eckart–Young optimality guarantee for symmetry-preserving tensor approximation, enabling principled rank selection with closed-form error bounds.
Machine-Verified Equivariance: The core algebraic results (associativity, distributivity, SVD optimality, product group composition) are formalized in Lean 4 (600 lines of code) with zero unresolved proof obligations, providing a certificate of mathematical correctness.
Data-Driven Symmetry Discovery: The framework enables the discovery of the optimal symmetry group and its factorization structure directly from data, without prior knowledge.
Per-Irrep Interpretability: The algebraic decomposition allows for a closed-form breakdown of predictive power by irreducible representation channels, revealing the symmetry content of physical observables.

Experimental Results

1. Synthetic Validation

On synthetic data with exact $Z_{12}$ rotational symmetry:

$\star_G$ -SVD + Ridge Regression achieved perfect prediction ( $R^2 = 1.000$ ) and exact invariance (rotation variance $\approx 10^{-31}$ ) using only 101 parameters.
Neural baselines (Standard MLP, Augmented MLP, Invariant MLP) required 5,000–14,000 parameters, achieved lower accuracy, and exhibited significantly higher rotation variance (up to $10^{-5}$ for Augmented MLP).

2. QM9 Molecular Property Prediction

On the QM9 benchmark (130,831 molecules) predicting the HOMO–LUMO gap:

$\star_G$ -SVD + Ridge achieved $R^2 = 0.556$ with only 107 parameters, outperforming neural baselines in the parameter-efficient regime.
Standard and Invariant MLPs collapsed ( $R^2 < 0$ ) on small sample sizes (1,000 molecules), while $\star_G$ maintained positive $R^2$ from as few as 100 molecules.
Within-Isomer Analysis: When controlling for molecular size (a major confounder in QM9), the performance gap between $\star_G$ and MLPs vanished, indicating that larger models' pooled $R^2$ advantages were largely due to size prediction rather than superior chemistry learning.

3. Product Group Composition

In a task with two commuting symmetries ( $Z_6 \times Z_4$ ):

The product group model achieved perfect prediction ( $R^2 = 1.000$ ) with 186 parameters.
Single-factor models ( $Z_6$ or $Z_4$ alone) captured only $\le 23\%$ of the variance.
A naive cyclic approximation ( $Z_{24}$ ) achieved $R^2 = 0.986$ , failing to resolve the coupled frequency modes that the product structure uniquely captures.

4. Symmetry and Factorization Discovery

Group Discovery: Scanning candidate groups on QM9 data identified $Z_4$ as the best fit, consistent with the $C_4$ rotational symmetry of many organic molecules, without any crystallographic input.
Factorization Discovery: For data on a 24-element group, the algorithm identified $Z_3 \times Z_8$ as the optimal decomposition ( $R^2 = 1.000$ ), outperforming the cyclic $Z_{24}$ ( $R^2 = 0.985$ ) and revealing latent product structures.

5. Empirical Recovery of Wigner–Eckart Selection Rules

Decomposing molecular geometry over the chiral octahedral group ( $O$ ):

Scalar properties (e.g., HOMO-LUMO gap) were dominated by the $A_1$ ( $l=0$ ) channel.
Vector components (dipole moments) were dominated by the $T_1$ ( $l=1$ ) channel, with a $T_1/A_1$ predictive power ratio of $\sim 2.8$ (vs. $\sim 0.5$ for scalars).
Rank-2 tensors (isotropic polarizability) showed near-zero dependence on the $l=1$ channel, consistent with the theoretical decomposition $l=0 \oplus l=2$ .
These physical selection rules were recovered solely from data, without any quantum-mechanical input.

6. Comparison with Equivariant Neural Networks (ENNs)

Pooled $R^2$ : State-of-the-art ENNs (MACE, SchNet) achieve higher pooled $R^2$ on QM9 but require $\sim 10^6$ parameters and consume full atomic graphs.
Input Contract: $\star_G$ operates on molecule-level summaries. When compared to MLPs using the same molecule-level input, $\star_G$ is competitive or superior at orders of magnitude fewer parameters.
Cross-Selectivity: On polarizability components, $\star_G$ models restricted to single irreps ( $E$ or $T_2$ ) achieved $\sim 96\%$ cross-selectivity (predicting one component while ignoring the other), whereas ENNs treated both components symmetrically ( $< 1\%$ selectivity), demonstrating $\star_G$ 's unique ability to disentangle representation-theoretic structures.

Significance and Claims

The paper claims that the $\star_G$ framework offers a different mathematical affordance rather than simply a "faster-better-cheaper" alternative to ENNs. Its significance lies in:

Provably Optimal Symmetry-Preserving Compression: Providing the first global optimality guarantee for symmetry-aware tensor approximation.
Machine-Verified Equivariance: Ensuring mathematical correctness through formal verification.
Per-Irrep Interpretability: Offering a closed-form decomposition of predictions that reveals the underlying symmetry structure of physical observables.
Data-Driven Physical Discovery: Enabling the recovery of physical laws (like Wigner–Eckart selection rules) and symmetry groups directly from empirical data.

The authors conclude that in data-scarce scientific settings, algebraic structure serves as a stronger inductive bias than architectural expressivity, allowing deep algebra to outperform big data approaches.