acceptodds
Under review as a conference paper at ICLR 2027

A Stability Theory for Generator-Based Continuous Symmetry Learning

Abstract

Symmetry discovery learns transformations that preserve functions, data distributions, or the solution sets of differential equations, revealing structure that can guide model design and scientific understanding. Existing methods address these target objects through different formulations, making it difficult to identify common conditions for success. We develop a stability theory for generator-based symmetry discovery by formulating it as a subspace recovery problem through a task-specific symmetry-measuring operator. This provides a common framework for analyzing how surrogate models and numerical approximations affect recovered symmetries. We establish sufficient conditions for accurate recovery and derive an error bound governed by the surrogate operator error relative to the separation between symmetry and non-symmetry directions. A taxonomy clarifies how existing methods relate and which settings our theory covers. We validate the theoretical predictions in controlled environments across three settings, where the true symmetry subgroup, perturbation strength, and failure modes can be explicitly controlled.

Primary area:
learning theory
Keywords:
symmetry discovery, geometric deep learning, Lie group

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RuiaOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a bound on subspace error between estimated and ground truth symmetries in the context of generator-based discovery of global subgroup symmetries which use a vector field formulation. The authors provide a taxonomy of symmetry discovery based on many recently-published methods of symmetry discovery.

Critical strengths

Strength 1: this paper begins to address an important aspect of symmetry discovery, which is the tendency of existing methods to consider different metrics for success.

Strength 2: this paper shows potential, but requires revision before publication should be considered.

Strength 3: this paper seems to be sufficiently comprehensive in terms of the most recent and salient symmetry discovery methods considered.

Critical weaknesses

Weakness 1: The paper is not sufficiently rigorous from a mathematical point of view. The field of symmetry discovery sits at the intersection of machine learning and the rich and rigorous field of differential geometry (and adjacent mathematical fields). As such, advances in symmetry discovery must be sufficiently rigorous, not merely as a service to mathematics, but rather that the reader can precisely understand the claimed contribution. My opinion is that the paper in its current form is imprecise so as to obscure the claimed contributions, and that publication in its current form would enable incorrect notions of symmetry discovery to persist. To support this argument, I present the following examples as a subset.

(1) In section 2, it seems that the principle of symmetry is summarized by the so-called symmetry relation Rω(x,g)=0.R_{\omega}(x,g)=0. This is objectionable in part because it is not defined: moreover, efforts by the authors to justify their claim that each symmetry method studied can be reduced to the task of producing SxS_x are absent, or so apparently. This is not fully resolved in Appendix E.

(2) The proposed taxonomy may ultimately be suitable, but the authors do not discuss or acknowledge the simplicity thereof. In the target object "dimension," for example, lines between so-called functional symmetry and distributional symmetry are blurred when the underlying distribution can be described by a probability density function. Locality is another case where lines can be blurred, as vector fields can define local or global flows.

(3) Section 3.1 (specifically lines 167-176) begins with "Let X be the data domain..." which seems quite general. But subsequent statements imply much stronger assumptions about the structure of the so-called "data domain;" e.g., that it has a tangent space with a norm. I think the authors should be more clear/specific about what precise mathematical construction is needed/assumed, including about the probability measure q. The statement "the Hilbert space in which infinitesimal generators are represented as vector fields" moves too quickly in the sense that the authors should more formally introduce symmetry and infinitesimal generators in the context of their focus, which is generator-based discovery of global subgroup symmetries using vector field formulation (and sometimes matrix generators). Additionally, lines 172-177 are simply not sufficiently explained: in fact, it is not clear to me how the operator A\mathcal{A} is defined.

(4) In line 191, gϵ=exp(ϵv)g_{\epsilon} = \text{exp}(\epsilon v) is not defined.

(5) I don't believe the authors have actually accurately captured symmetry discovery in the case of distributional symmetry: the authors should connect their provided infinitesimal condition to their cited sources, such as Yang et al., 2023.

Weakness 2: the paper attempts to introduce a taxonomy in a way that only a survey can do. Many claims about the methods considered are simply not justified because there is not room in a 9-page conference paper to do so. However, this leaves critical gaps in the justification of the proposed taxonomy (some of which tie in to weakness 1). This is not resolved in Appendix E.

Weakness 3: some contributions do not seem warranted. In lines 239-240, the authors claim to have reformulated generator-based symmetry discovery as the problem of recovering the kernel of a task-specific symmetry-measuring operator. But it's not clear that this is a novel reformulation, particularly given Otto et al., 2024 and Shaw et al., 2024, in which symmetry discovery is explicitly described as recovering such a kernel or nullspace. Additionally, the subspace distance question seems to have already been taken up in Hu et al. in their paper introducing LieNLSD: at the very least, an empirical comparison should be attempted.

Weakness 4: The proposed taxonomy distinguishes linear and nonlinear approaches, but the experiments seem lacking in the nonlinear category. Furthermore, despite the assurance of a vector field focus, any recovered generators appear to be matrices.

Weakness 5: in terms of writing-specific weaknesses, my impression was that terms/phrases were emphasized excessively, and that much of the paper was too conversational/informal.

Questions for authors

Question 1. What is the purpose of encouraging orthonormality (section 3.1, lead-up to equation 3)? Some existing methods already attempt to impose orthonormality, whereas the writing does not address this, seemingly claiming to introduce orthonormality as a novelty.

Question 2. Is the reference to Yang et al., 2023 in both the functional and distributional target object sections in Appendix E intentional?

1 comment

I appreciate the authors' dilligence in supplying explanations and revisions, which have changed my perception of the paper. I believe conference attendees would find this work to be of interest, and I have raised the overall score accordingly.

Reviewer_wYmMOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Symmetry discovery has been studied across various domains, such as functions, probability distributions, and partial differential equations. However, a unified framework for analyzing the stability of discovered symmetries across these domains remains lacking. The authors propose an operator-based perspective that unifies different symmetry discovery settings. In this framework, symmetry generators correspond to the kernel of a symmetry-measuring operator. Furthermore, the authors establish a stability bound for the recovered symmetry subspace under surrogate operator perturbations, showing that the recovery error depends on the surrogate operator error relative to the spectral gap of the ground-truth operator.

Critical strengths
  1. The problem setting is timely, and the novelty is sufficient. Previous works have studied symmetry discovery in different domains, often using different formulations and definitions. This work provides a unified operator-based perspective and a common theoretical framework for analyzing the stability of symmetry discovery across these domains.

  2. The theoretical formulation and results are convincing under the stated assumptions. I did not find any obvious errors in the main theoretical arguments.

Critical weaknesses
  1. The theoretical guarantees rely on relatively strong assumptions, including a finite-dimensional symmetry subspace, an isolated spectral gap, and sufficiently small surrogate operator error. These conditions may be difficult to satisfy or verify in practical settings, potentially limiting the applicability of the proposed framework beyond theoretical analysis.
Questions for authors
  1. Could the authors clarify the scope of applicability of Theorem 3.1? In particular, the theorem assumes a finite-dimensional kernel and an isolated spectral gap, while the generator space is defined as an infinite-dimensional Hilbert space. Does the theorem apply to the full function space under these assumptions, or should the analysis be restricted to a finite-dimensional candidate generator space, as in the experiments?
Reviewer_rFpJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes the stability of recovering symmetry learned from data. They reformulate symmetry discovery as finding the null space of a linear operator. Consequently, they conclude that the accuracy of the discovered symmetry is bounded by the approximation error relative to the spectral gap of the symmetry-measuring operator, without directly analyzing the prediction error of the surrogate model.

Critical strengths
  1. Good idea that reformulates different formulations of functional, distributional, and PDE symmetry discovery into a null-space recovery of a linear operator.
  2. Clear results that support the theoretical findings such as measuring subspace recovery error as delta/lambda varies.
Critical weaknesses
  1. The accuracy of the discovered symmetry is bounded by the approximation error of the symmetry-measuring operator relative to its spectral gap. However, the theory does not establish how the surrogate model's prediction error affects this operator approximation error, which might be the most critical factor in practical symmetry discovery. In particular, a small prediction error does not necessarily imply a small gradient error, which directly affects the operator approximation error. Therefore, it is unclear whether the theoretical guarantee is meaningful for surrogate models learned from finite data.

  2. The experiments are all conducted on synthetic data with controlled operator perturbations. Although these experiments may be sufficient to validate the relationship between symmetry accuracy and the critical factors identified in the theory, they do not demonstrate whether the theory applies to surrogate models learned from real data. Real-world data with known physical symmetries could provide an opportunity to verify the practical relevance of the theory.

Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.