Exact Symmetry as Algebra: A Machine-Verified Tensor Calculus that Enforces Physical Selection Rules

Paulina Hoyos, Shashanka Ubaru, Dongsung Huh, Vasileios Kalantzis, Kenneth L. Clarkson, Misha Kilmer, Haim Avron, Lior HoresharXiv 2026 · arXiv

TL;DR

This paper introduces $\star_G$, a novel tensor algebra that embeds physical symmetries directly into the mathematical multiplication rules of data rather than forcing neural network architectures to learn them. By doing so, it eliminates compounding approximation errors over deep networks, driving symmetry violations down to machine zero ($10^{-32}$) and providing a machine-verified, mathematically optimal framework for physical data processing.

What Problem Does This Paper Address?

Symmetry is fundamental to physics and chemistry—think of a molecule's properties remaining unchanged when rotated. Modern Equivariant Neural Networks (ENNs) attempt to preserve these symmetries by carefully engineering specialized network layers.

However, the authors identify two major flaws in this architectural approach:

  1. Compounding Errors: ENNs only satisfy symmetry approximately due to floating-point and optimization limits. A tiny per-layer equivariance error ($\varepsilon$) accumulates linearly with network depth ($M$), leading to a total drift of $O(M\varepsilon$). In deep networks or iterative simulations, this causes physical predictions to drift into impossible states.
  2. The Combinatorial Wall: Engineering networks that respect multiple coexisting symmetries (like simultaneous rotations and permutations) requires redesigning architectures from scratch for every new combination.

Additionally, standard machine learning pipelines often flatten multidimensional tensors into vectors, destroying the geometric meaning and structural relationships inherent in the data.

Methods

Instead of changing the network architecture to fit the symmetry, the authors change the underlying mathematics to fit the geometry of the data. They construct the $\star_G$ tensor algebra, where any finite or compact group $G$ defines the tensor multiplication rule via group convolution.

Key components of their approach include:

  • Intrinsic Equivariance: Equivariance becomes a built-in property of multiplication. Composing multiple symmetries is as simple as specifying a direct product of groups ($G_1 \times G_2 \times \dots \times G_d$), requiring zero architectural redesign.
  • Group-Fourier Transform ($F_G$): This transform block-diagonalizes every $\star_G$ tensor into independent irreducible representation (irrep) blocks. This allows standard matrix operations (SVD, QR, least-squares) to be applied blockwise while remaining exactly equivariant.
  • $\star_G$-SVD & Eckart–Young Optimality: They prove that truncating tensors in this Fourier domain provides the mathematically optimal low-rank approximation that preserves group structure.
  • Formal Verification: The core algebraic laws and optimality proofs were formally machine-checked using the Lean 4 proof assistant to guarantee absolute mathematical correctness under an explicit axiom budget.

Datasets

The framework was evaluated on several established physical benchmarks and synthetic datasets:

  • QM9: A dataset of 130,831 small organic molecules used to evaluate scalar property predictions (e.g., HOMO–LUMO gap, dipole magnitude).
  • QM7-X: Used specifically for its full molecular polarizability tensors to test rank-2 tensorial selection rules.
  • de Jong Elastic-Tensor Dataset: A collection of 1,181 inorganic crystals spanning various crystal systems (cubic, hexagonal, etc.) used to validate point-group selection rules.
  • Synthetic Datasets: Custom-generated molecular configurations exhibiting exact $Z_{12}$ rotational symmetry and $Z_6 \times Z_4$ product-group symmetries to isolate and verify exactness guarantees.

Key Insights and Ideas

  • The Exactness Dichotomy: Approximate symmetry drifts, but exact symmetry is stable at unbounded depth. The authors demonstrate a 14-order-of-magnitude gap in symmetry preservation between approximate methods and the $\star_G$ algebra over deep iterations.
  • Symmetry as a Spectroscope: Because the group-Fourier transform maps data directly onto angular-momentum channels, the algebra acts as a physical spectroscope. It can read off selection rules (like the Wigner–Eckart theorem) directly from raw geometric data without any quantum-mechanical inputs.
  • Canonicity: Requiring a tensor product to be equivariant uniquely forces the transform to be the group-Fourier transform. It is a mathematical necessity, not an arbitrary modeling choice.

Key Results

Eliminating Forbidden Leakage

When applied as a post-hoc "physicality layer" on top of standard crystal graph neural networks predicting elastic tensors, the $\star_G$ projection drove symmetry-forbidden channel leakage from $10^{-2}$ down to machine zero ($\sim 1.5 \times 10^{-32}$).

Screening Recovery

In material discovery pipelines, unconstrained models often predict mechanically unstable tensors that violate Born stability criteria, causing valid materials to be falsely rejected. The physicality layer completely eliminated these false rejections, recovering thousands of viable material candidates.

Optimal Compression

The $\star_G$-SVD achieved exact data reconstruction of molecular tensors at rank 3, whereas standard matrix SVD retained over 20% reconstruction error at the same rank because it could not exploit the underlying group structure.

Symmetry-preserving compression performance compared to standard matrix SVD and Tucker decomposition.

Data Efficiency

In data-scarce regimes (e.g., training on only 100 molecules), the closed-form $\star_G$-SVD ridge regressor significantly outperformed unconstrained neural networks while using orders of magnitude fewer parameters and requiring zero training time.

Limitations and Future Work

  • Pooled Accuracy vs. Expressivity: The authors openly state that the $\star_G$ algebra is a structural and diagnostic tool, not a replacement for massive deep models. When data is abundant, graph neural networks still lead in aggregate accuracy because they capture complex bond topologies that a fixed local descriptor cannot.
  • Topological Blindness: The $\star_G$-SVD certifies approximation quality and irrep content based on singular values, but it is "blind" to the winding of singular vectors. It cannot certify topological properties like Chern numbers or Wannier localizability.
  • Non-Compact Groups: While extended to band-limited compact groups and crystallographic space groups, extending full covariance to non-compact groups (like the full Poincaré group without fiber reduction) remains an open theoretical frontier.

Summary updated

Abstract

Symmetry is central to the physical sciences, yet machine learning usually captures it only approximately, leaving a residual per-step equivariance error $\varepsilon$ that compounds with depth $M$ as $M\varepsilon$, whereas exact equivariance holds at unbounded depth; we demonstrate this divergence at fourteen orders of magnitude. We show that a symmetry can be made exact by construction, as the multiplication rule of a tensor algebra. In the resulting $\starG$ algebra, defined by any finite group $G$, the group-Fourier transform block-diagonalizes every tensor into irreducible-representation blocks, making equivariance intrinsic; requiring equivariance conversely \emph{forces} this suitably normalized transform, so the algebra is determined by $G$ rather than chosen. The standard matrix toolbox, including a Frobenius-optimal low-rank factorization, transfers blockwise, machine-checked in Lean~4 under an explicit axiom budget, and extends unchanged to band-limited compact groups and, under periodic boundary conditions, to all 230 crystallographic space groups and the compact little-group fibers of Euclidean and Poincaré symmetry. This exactness is an applied capability: on inorganic-crystal elastic tensors the algebra enforces point-group selection rules exactly on the output of \emph{any} predictor, driving a trained graph network's forbidden-channel leakage from $10^{-2}$ to machine zero, eliminating mechanically unstable predictions, and recovering viable materials that an unconstrained screen discards; on molecular data, with no quantum-mechanical input, it exposes octahedral selection-rule signatures consistent with the Wigner--Eckart theorem. Matched networks lead on pooled molecular accuracy, which we report plainly: the contribution is a complementary algebraic calculus, structural and diagnostic, exact at any depth.
BibTeX
@misc{hoyos2026exact,
  author = {Paulina Hoyos and Shashanka Ubaru and Dongsung Huh and Vasileios Kalantzis and Kenneth L. Clarkson and Misha Kilmer and Haim Avron and Lior Horesh},
  title = {{Exact Symmetry as Algebra: A Machine-Verified Tensor Calculus that Enforces Physical Selection Rules}},
  year = {2026},
  eprint = {2605.20440},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.20440},
}
Figure