Discovering Continuous Symmetries via Lie Algebra Generators

This direction is merged into another direction: automated symmetry discovery. Its last version stays readable and citable here.

Version 1 of 127 papers35 conceptsfrozen 2026-10-04

In modern machine learning, symmetry is often treated as a "cheat code" for generalization. If a model knows that an object remains the same regardless of its rotation or position, it doesn't need to see every possible orientation during training. Traditionally, researchers have manually "baked" these symmetries into architectures—think of the translational invariance in Convolutional Neural Networks.

However, in complex domains like particle physics or climate modeling, the underlying symmetries are often unknown or non-obvious. A growing body of research focuses on symmetry discovery: teaching neural networks to identify the mathematical groups governing a dataset without human intervention.

The Infinitesimal Approach

The core challenge of discovering a continuous symmetry (like rotation) is that the group of possible transformations is infinite. You cannot simply iterate through every possible angle. Instead, researchers like Chris Ick and Vincent Lostanlen suggest looking at the "infinitesimal" level.

In Learning a Lie Algebra from Unlabeled Data Pairs, they demonstrate that a network can recover the infinitesimal generator of a transformation group. Think of this generator as a "nudge" or a derivative at the identity. If you know the direction of a tiny rotation, you can reconstruct the entire circular path by repeatedly applying that nudge. Mathematically, this is done via the matrix exponential, which maps the learned Lie algebra (the local nudges) to the Lie group (the global transformations).

Adversarial Discovery of Physical Laws

How do we find these generators in raw data? One powerful approach is the LieGAN framework, developed by Jianke Yang, Robin Walters, and Rose Yu.

In Generative Adversarial Symmetry Discovery, the authors use a generative adversarial setup. A "symmetry generator" learns a basis of matrices (the Lie algebra) that transform data points. If the transformation is a true symmetry of the data distribution, a discriminator should not be able to distinguish the transformed data from the original. By optimizing this game, the model discovers interpretable physical operations—such as Lorentz boosts or rotations—directly from the data.

From Linear to Nonlinear Symmetries

Early methods focused on linear transformations, but real-world data often moves on curved manifolds. Gyeonghoon Ko and Juho Lee expanded this in Learning Infinitesimal Generators of Continuous Symmetries from Data by modeling generators as learnable vector fields. This allows the discovery of complex, nonlinear symmetries that don't fit into simple matrix multiplications.

Similarly, Jianke Yang introduced Latent LieGAN (LaLiGAN) in Latent Space Symmetry Discovery. This model maps high-dimensional observations into a latent space where complex nonlinear symmetries become linear. By finding the Lie algebra in this simplified latent space, the model can forecast the behavior of dynamical systems with much higher accuracy.

Provable Invariance and Interpretability

Discovery is only half the battle; the other half is using that knowledge to build better models. Pavan Karjol, Vivek Kashyap, and Prathosh AP developed $H_\gamma$-Net, a framework that "canonicalizes" data based on discovered symmetries.

As detailed in Interpretable Discovery of One-parameter Subgroups, $H_\gamma$-Net doesn't just approximate invariance; it provides mathematical guarantees. By parameterizing symmetries through learnable quadratic forms and orientation matrices, the model becomes a "white box." A researcher can look at the learned weights and see, for example, the Minkowski metric of special relativity emerging from raw particle physics data.

Why It Works: The Power of Structure

The success of these methods stems from a shift in perspective: instead of asking a neural network to learn a function $f(x) = y$, we ask it to learn the structure of the space in which $x$ lives.

By focusing on Lie algebras, these models leverage the fact that continuous transformations have a rigid algebraic structure. Whether it's Bispectral Neural Networks extracting Cayley tables or AtlasD discovering local symmetries on a sphere, the goal is the same: to find the fundamental mathematical laws that make the data "predictable." When a model discovers a symmetry, it effectively reduces the dimensionality of the problem, allowing it to generalize to unseen scenarios with far less training data.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_discovering_continuous_symmetries_via_li_248f580d,
  title = {Discovering Continuous Symmetries via Lie Algebra Generators},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/discovering-continuous-symmetries-via-lie-algebra-generators/rdr_248f580d78ba409b93b8fc5daed31c9b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_discovering_continuous_symmetries_via_li_248f580d_v1,
  title = {Discovering Continuous Symmetries via Lie Algebra Generators},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/discovering-continuous-symmetries-via-lie-algebra-generators/rdr_248f580d78ba409b93b8fc5daed31c9b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}