Pith. sign in

REVIEW 3 major objections 5 minor 108 references

Adaptive Symmetry Discovery for Dynamical System Identification

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Unknown finite symmetries can be discovered from a single trajectory at no additional observation cost, and the dynamics can be identified within the same trajectory length as if the symmetry were known, provided the family of candidate…

desk verdict A strong representation-theoretic core with a load-bearing overclaim: the formal theorem only guarantees the worst-case candidate threshold, not the advertised known-symmetry trajectory length. read the letter →

arxiv 2608.08091 v1 pith:BOPXBCCE submitted 2026-08-08 cs.LG cs.AImath.DS

Behrooz Tahmasebi , Melanie Weber This is my paper · ORCID
classification cs.LGcs.AImath.DS MSC 37M1068Q3220C1505C25
keywords systemidentificationequivariancesymmetrydiscoveryfinitegroupssamplecomplexityrepresentationtheoryCayleygraphssingletrajectory
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper addresses a practical challenge in scientific machine learning: when a dynamical system's only symmetry is an unknown finite group, can one still identify the system from a short single trajectory? For the class of feature-lifted linear systems $f(x)=W\Phi(x)$, it first gives a representation-theoretic formula for the minimal trajectory length $T_\Phi(G)$ needed when $G$ is known, showing that symmetries shorten the trajectory from the generic length $m$ to a quantity determined by multiplicities of irreducible representations. Its central claim is that the same length $T_\Phi(\mathcal G)=\max_{H\in\mathcal G}T_\Phi(H)$ suffices when the symmetry is unknown, under the assumption of generic candidate separation: no candidate group outside the true one's subgroups can fit the data. The paper proposes a concrete algorithm that tests random generating sets of candidate groups by linear feasibility and selects the largest feasible group, so the saving in data is real while the extra cost is computational. If right, symmetry discovery is observationally free for such systems, and the theory quantifies exactly when and how.

What carries the argument

Three pieces carry the argument. The first is the isotypic decomposition of the state and feature representations of a finite group: after a change of basis, every equivariant map takes the block form $\bigoplus_\pi C_\pi\otimes I_{V_\pi}$, so identification reduces to checking whether the per-block design matrices $\Phi_{\pi,T}$ have full row rank. The second is the random-generation property of finite groups: $O(\log|G|+\log(1/\delta))$ i.i.d. uniform samples generate $G$ with probability $1-\delta$, equivalently making the Cayley graph $\mathrm{Cay}(G,S)$ connected. The third is the feasibility formulation that combines the trajectory equation $W\Phi(x_t)=x_{t+1}$ with the sampled intertwining constraints $\rho(g)W=W\rho_\Phi(g)$, which are linear in the entries of $W$ and, by the generating-set property, enforce full equivariance. Generic candidate separation is the condition that lets the algorithm pick the true group as the unique largest feasible candidate.

What would settle it

For a fixed feature map such as $\Phi_{\le 2}$, enumerate all pairs of incomparable subgroups $H_1,H_2$ of $S_4$ or $S_5$; for a generic system with full symmetry group $H_1$, solve the Algorithm 1 feasibility program for $H_2$ at trajectory length $T=\max(T_\Phi(H_1),T_\Phi(H_2))$. If any $H_2$ that is not a subgroup of $H_1$ is feasible, generic candidate separation fails for that family, and the no-extra-trajectory guarantee does not hold unconditionally.

Watch

Extended reading notes

Core claim

The central claim is that, under a separation condition on the candidate family, discovering the symmetry group is observationally free: Algorithm 1 recovers both the parameter matrix $W$ and a generating set of the full symmetry group $G$ from a single generic trajectory of length $T_\Phi(\mathcal G)$, the same worst-case trajectory length required when $G$ is known. The proof builds on a precise characterization of the known-symmetry threshold: $T_\Phi(G)$ is the smallest $T$ such that, in every isotypic component $\pi$ that appears in the state representation, the generic rank $h_{\pi,\Phi}(T)$ of the stacked feature matrix $\Phi_{\pi,T}$ reaches the feature multiplicity $m_\pi$. Because equivariance on a generating set implies full equivariance, logarithmically many random samples per candidate suffice; the largest feasible candidate is then the true group, since every feasible candidate must be a subgroup of it. The paper also proves a bounded-index variant in which samples are drawn from a known ambient group and rejected until they generate the unknown subgroup, avoiding enumeration of candidate subgroups.

Load-bearing premise

The load-bearing premise is that the candidate family is generically separating: for the true group $G$, no group $H$ outside the subgroups of $G$ can admit an equivariant system consistent with the observed trajectory; the paper proves this automatically only when the candidates are totally ordered by inclusion, and otherwise it is assumed rather than derived.

Editorial extensions

If this is right

  • For candidate families totally ordered by inclusion, the separation condition is automatic, so adaptive discovery provably matches the known-symmetry trajectory length without any extra observations.
  • For fixed-degree permutation-equivariant polynomial systems, the known-symmetry threshold is $O_k(1)$, independent of the state dimension $d$, so the adaptive result would make unknown permutation symmetries equally cheap when separation holds.
  • When the true symmetry is a bounded-index subgroup of a known ambient group, the algorithm discovers and generates it without enumerating the candidate subgroups, with expected sampling overhead at most a factor $B$.
  • The computational cost scales polynomially in $d,m,T,|\mathcal G|,\log|G|_{\max},\log(1/\delta)$, so groups of size exponential in the state dimension are handled with polylogarithmic random samples.
  • If a generating set for each candidate is supplied, randomization is unnecessary and the algorithm becomes a deterministic feasibility search.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 'same length as known-symmetry' promise is conditional on the threshold not being dominated by the least symmetric candidate: if the trivial group belongs to $\mathcal G$, then $T_\Phi(\mathcal G)=m$ and the adaptive guarantee collapses to the generic no-symmetry length, so the meaningful regime is a family that excludes near-trivial candidates or a guarantee at $T_\Phi(G)$ itself.
  • Generic candidate separation is proved automatically only for nested families; for incomparable candidates it is an extra hypothesis that should be checked. A concrete computational test on incomparable subgroups of a symmetric group, such as $C_4$ and $V_4$ inside $S_4$, would reveal how often separation holds for natural feature maps.
  • The feasibility tests assume exact data; under noise the true system would generically fail the equality constraints, so a practical variant would need tolerant feasibility, likely least-squares residual thresholds, and the question of how the trajectory-length gain degrades with noise remains open.
  • The bounded-index sampler requires elementwise separation, a stronger per-element condition than group-level separation; if spurious elements pass the feasibility test, rejection sampling no longer yields uniform samples from the hidden subgroup and the generating-set guarantee fails.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies identification of discrete-time feature-lifted linear dynamical systems f(x)=WΦ(x) from a single noiseless trajectory, under finite group equivariance. It first gives a representation-theoretic characterization T_Φ(G) of the generic identification threshold for a known symmetry group, with several examples (linear, affine, finite Abelian, permutation-equivariant polynomial) showing large reductions relative to the generic threshold. It then proposes adaptive symmetry discovery: Algorithm 1 samples random generators for each candidate group, tests feasibility of the trajectory and equivariance constraints, and returns the largest feasible group. Theorem 4.5 guarantees success for T at least the worst-case threshold over the candidate family, under a 'generic candidate separation' assumption and a random-generation condition. Algorithm 2 handles bounded-index subgroups by rejection-sampling from an ambient group under a 'generic elementwise separation' assumption. The appendix gives detailed proofs and a small matched-symmetry experiment.

Significance. If the results were as advertised, the paper would be a useful contribution: a clean sample-complexity theory for equivariant system identification, an elegant use of random generating sets and Cayley graphs to avoid enumerating finite groups, and a striking example where permutation equivariance reduces the required trajectory length from Θ(d^2) to a constant. The known-symmetry part (Theorem 4.1 and its corollaries) appears coherent, and Proposition A.2 on random generation is elementary and correct. However, the central adaptive claim in the abstract is not supported by Theorem 4.5, and the paper's own Remark 4.10 concedes that general candidate families require a larger horizon. This overclaim materially reduces the significance of the paper in its current form.

major comments (3)
  1. [Abstract; §4.2, Theorem 4.5] The central claim that adaptive discovery "achieves the same optimal trajectory length as in the known-symmetry case" is not what Theorem 4.5 proves. Theorem 4.5 requires T ≥ max_{H∈G} T_Φ(H), whereas the known-symmetry length for the true group G is T_Φ(G). If G contains the trivial group, max_{H∈G} T_Φ(H) = m, the generic no-symmetry threshold, so the theorem's guarantee degenerates to the baseline. Definition 4.4 defines generic separation only for T ≥ max_{H∈G} T_Φ(H), so it cannot be invoked at T = T_Φ(G). Remark 4.10 explicitly concedes that for incomparable candidate families the worst-case threshold need not suffice and a larger horizon is required, which contradicts the abstract and Remark 4.9. The authors must either weaken the abstract and main text to a worst-case threshold or to nested families, or add a new theorem establishing success at T = T_Φ(G) under an explicit and verified separation condition.
  2. [Definition 4.4; Proposition B.5] Generic candidate separation is the load-bearing premise of Algorithm 1, but it is proved only for families totally ordered by inclusion (Proposition B.5). For incomparable candidates H that are neither subgroups nor supergroups of the true group G, no argument rules out an H-equivariant system that fits the trajectory when T is below T_Φ(H); in that regime the feasible set can have many degrees of freedom. The manuscript offers no example of a non-nested family satisfying Definition 4.4 and no structural condition beyond inclusion that implies it. Since Theorem 4.5's conclusion is conditional on this assumption, the adaptive result is much narrower than the introduction and abstract suggest. The assumption should be either derived for a substantive class of examples or explicitly represented as a restrictive condition in the statement of the main theorem.
  3. [Theorem 4.11] Theorem 4.11 relies on "generic elementwise separation," which is simply assumed. The acceptance test in Algorithm 2 is used to decide membership in the unknown subgroup H, and the proof requires that a sampled g ∈ Γ passes the feasibility test iff g ∈ H for a generic system. This is a strong correctness condition: for trajectory lengths below the known-symmetry threshold of a supergroup or an incomparable subgroup, spurious feasible elements may pass. No conditions on Γ, Φ, or T are given that imply this equivalence. As with Definition 4.4, the theorem is a conditional statement rather than an unconditional near-zero-overhead guarantee. The main text should state the assumption clearly and provide at least one nontrivial family where it is verified.
minor comments (5)
  1. [§3.3 and Definition 4.4] The symbol T_Φ(G) is used both for the known-symmetry threshold of a group G and for the worst-case threshold max_{G∈G} T_Φ(G) of a candidate family. This notation collision makes the gap between the theorem and the abstract easy to miss; use a distinct symbol such as T_Φ(𝒢) for the family threshold.
  2. [§4.1] In the paragraph on representation stability, the phrase "stabilize stabilize" contains a duplicated word.
  3. [Appendix A.1] The examples paragraph contains grammatical slips: "Here is a few examples" and "permutation group (also known as symmetric group) of group of all permutations" should be rewritten.
  4. [Appendix C] The experiments only illustrate the known-symmetry thresholds for matched symmetry classes; they do not run Algorithm 1 or Algorithm 2. A small experiment executing the adaptive discovery procedure, even on the linear systems in Table 1, would directly test the paper's main algorithmic claim.
  5. [Proof of Theorem 4.5] In the proof, the line "Since T ≥ T_Φ(G) ≥ T_Φ(G)" prints the same symbol for the family threshold and the group threshold; this is a consequence of the notation collision noted above and should be corrected for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is self-contained and the adaptive guarantee is a genuine conditional theorem.

full rationale

I walked the derivation chain from Theorem 4.1 to Theorem 4.5 and found no step in which a claimed prediction is equivalent to an input by construction. Theorem 4.1 characterizes the known-symmetry threshold T_Phi(G) by a generic-rank condition on isotypic feature blocks; the proof is a direct linear-algebra identifiability argument, not a fit to the adaptive algorithm's outputs. Algorithm 1 uses the trajectory only through linear feasibility constraints and uses randomness only to draw generating sets; no parameter is fitted to a subset of data and then reported as a prediction. Definition 4.4 (generic candidate separation) is an external assumption, not a restatement of the theorem: the theorem's proof applies it to conclude that every feasible candidate is a subgroup of the true group, and the nested-family case is genuinely proved in Proposition B.5. Remark 4.10 explicitly acknowledges that for incomparable families the worst-case threshold may be insufficient, which further shows the authors do not hide the assumption behind the conclusion. The self-citations (e.g., Tahmasebi and Weber 2025) appear only in related-work context and are not load-bearing for the main theorems. I did note that the abstract's phrase 'same optimal trajectory length as in the known-symmetry case' is stronger than Theorem 4.5, which guarantees success at the worst-case threshold max_{H in G} T_Phi(H); unless that maximum equals T_Phi(G) for the true group, the advertised equality is not established. That is a correctness/scope gap, not circularity, because the theorem does not define the known-symmetry length to equal the worst-case threshold—it simply states a sufficient condition.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central theorem rests on a new, unverified-for-general-families condition (generic candidate separation), and the headline claim requires an additional unstated equality between the worst-case-over-family threshold and the true group's threshold. No free parameters are fitted to data; the theory is parametric in the feature map and group representations.

assumptions (6)
  • domain assumption The candidate family mathcal G is known a priori and contains the true full symmetry group G of the unknown dynamics.
    Invoked in Section 3.3 and Theorem 4.5; without this, the search space is undefined.
  • ad hoc to paper Generic candidate separation (Definition 4.4): for every G in mathcal G, a generic G-equivariant system with a generic initial state has the property that, for T >= T_Phi(mathcal G), any feasible candidate H is a subgroup of G.
    This is the central structural assumption that makes Algorithm 1 correct. It is proven automatic only for nested families (Proposition B.5), and is otherwise assumed, not derived.
  • ad hoc to paper For the bounded-index setting, generic elementwise separation: a sampled g in Gamma satisfies g in H iff the feasibility test succeeds for a generic system with full symmetry H.
    Stated just before Theorem 4.11; this converts rejection sampling into uniform sampling from the unknown subgroup H.
  • domain assumption The feature map Phi is analytic with linearly independent coordinates and is compatible with the group action.
    Theorem 4.1 and the genericity arguments rely on analyticity and linear independence of the feature coordinates.
  • domain assumption Genericity is measured in the Lebesgue measure-zero sense in the space of initial states and equivariant parameters.
    Definition 3.7 and Appendix A; the paper assumes the observed trajectory avoids a measure-zero exceptional set.
  • ad hoc to paper To support the abstract's 'same trajectory length as known-symmetry' claim, the paper requires the unstated assumption that max_{H in mathcal G} T_Phi(H) = T_Phi(G) for the true G, or that separation holds already at T_Phi(G).
    Theorem 4.5 only guarantees T >= max_{H in mathcal G} T_Phi(H); the advertised 'no overhead' claim needs this additional equality, which is not stated or proved anywhere in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Symmetry Discovery for Dynamical System Identification." pith.science (2026). https://pith.science/paper/BOPXBCCE

@misc{pith2026260808091,
  author       = {Pith},
  title        = {Pith review of: Adaptive Symmetry Discovery for Dynamical System Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BOPXBCCE}},
  note         = {Machine review of arXiv:2608.08091}
}
read the original abstract

Dynamical systems model trajectory data generated by fixed underlying dynamics, with applications ranging from biology to physics. Especially in scientific settings, dynamical systems are not generic but often exhibit symmetries imposed by physical laws, formalized through equivariance with respect to group actions. The identification problem concerns recovering the parameters of a system from observed trajectories. In this work, we study adaptive symmetry discovery for dynamical system identification and address how a system can be identified from a single trajectory when it is equivariant with respect to an unknown symmetry group. To this end, we first show that for known symmetries, the system can be identified from a significantly shorter single trajectory than in the generic setting, and we precisely characterize this improvement. We then consider the automatic symmetry discovery setting, proposing a method to learn the symmetry group directly from a single trajectory and incorporate it into the identification procedure, achieving the same optimal trajectory length as in the known-symmetry case. Our analysis relies on tools from group representation theory and the expander properties of Cayley graphs, and may be of independent interest for the study of symmetries in dynamical systems.

Figures

Figures reproduced from arXiv: 2608.08091 by the authors.

Figure 1
Figure 1. Dimension of the feasible solution set ST (H) as a function of trajectory length T for the correctly matched symmetry classes. The dimension reaches zero at the predicted generic identification thresholds T = 1, 9, and 10 for Sd, C2, and the trivial group, respectively. Computing the feasible-set dimension. Let r = dim(C(H)), and fix a basis A1, . . . , Ar of C(H). Writing A = Pr i=1 θiAi , the constraint Y = AX bec… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

108 extracted references · 60 canonical work pages

  1. [1]

    Approximate Equivariance in Reinforcement Learning , author =

  2. [2]

    Group equivariant convolutional networks , author=

  3. [3]

    Lorentz group equivariant neural network for particle physics , author=

  4. [4]

    E(n) equivariant graph neural networks , author=

  5. [5]

    Wang, Dian and Walters, Robin and Platt, Robert , booktitle=iclr, year=

  6. [6]

    Steerable cnns , author=

  7. [7]

    Incorporating symmetry into deep dynamics models for improved generalization , author=

  8. [8]

    The exact sample complexity gain from invariances for kernel regression , author=

Show all 108 references
  1. [9]

    Provably strict generalisation benefit for invariance in kernel methods , author=

  2. [10]

    Relaxing equivariance constraints with non-stationary continuous filters , author=

  3. [11]

    Regularizing towards soft equivariance under mixed symmetries , author=

  4. [12]

    Residual pathway priors for soft equivariance constraints , author=

  5. [13]

    Learning partial equivariances from data , author=

  6. [14]

    Approximation-generalization trade-offs under (approximate) group equivariance , author=

  7. [15]

    Symmetry and Geometry in Neural Representations (NeurReps) Workshop , year=

    Almost equivariance via lie algebra convolutions , author=. Symmetry and Geometry in Neural Representations (NeurReps) Workshop , year=

  8. [16]

    Symmetry-based structured matrices for efficient approximately equivariant networks , author=

  9. [17]

    Approximately equivariant networks for imperfectly symmetric dynamics , author=

  10. [18]

    Sample complexity bounds for estimating probability divergences under invariances , author=

  11. [19]

    Sample complexity of probability divergences under group symmetry , author=

  12. [20]

    arXiv preprint arXiv:2104.13478 , year=

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges , author=. arXiv preprint arXiv:2104.13478 , year=

    arXiv

  13. [21]

    Imagenet classification with deep convolutional neural networks , author=

  14. [22]

    Invariant and Equivariant Graph Networks , author=

  15. [23]

    PRX Quantum , volume=

    Theory for equivariant quantum neural networks , author=. PRX Quantum , volume=. 2024 , publisher=

    2024

  16. [24]

    Generative adversarial symmetry discovery , author=

  17. [25]

    Improving equivariant model training via constraint relaxation , author=

  18. [26]

    Learning layer-wise equivariances automatically using gradients , author=

  19. [27]

    Equivariance via Minimal Frame Averaging for More Symmetries and Efficiency , author=

  20. [28]

    A canonicalization perspective on invariant and equivariant learning , author=

  21. [29]

    Regularity in Canonicalized Models: A Theoretical Perspective , author=

  22. [30]

    Generalization Bounds for Canonicalization: A Comparative Study with Group Averaging , author=

  23. [31]

    Equivariant frames and the impossibility of continuous canonicalization , author=

  24. [32]

    Lie Algebra Canonicalization: Equivariant Neural Operators under arbitrary Lie Groups , author=

  25. [33]

    Frame Averaging for Invariant and Equivariant Network Design , author=

  26. [34]

    Equivariance with learned canonicalization functions , author=

  27. [35]

    Approximately equivariant neural processes , author=

  28. [36]

    Discovering Symmetry Breaking in Physical Systems with Relaxed Group Convolution , author=

  29. [37]

    Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs , author=

  30. [38]

    Physical Review Letters , volume=

    Approximately Symmetric Neural Networks for Quantum Spin Liquids , author=. Physical Review Letters , volume=. 2025 , publisher=

    2025

  31. [39]

    arXiv preprint arXiv:2506.02269 , year=

    A Tale of Two Symmetries: Exploring the Loss Landscape of Equivariant Models , author=. arXiv preprint arXiv:2506.02269 , year=

  32. [40]

    NeurIPS AI for Science Workshop , year=

    Relaxed octahedral group convolution for learning symmetry breaking in 3d physical systems , author=. NeurIPS AI for Science Workshop , year=

  33. [41]

    Physical Review D , volume=

    Symmetry discovery with deep learning , author=. Physical Review D , volume=. 2022 , publisher=

    2022

  34. [42]

    Automatic symmetry discovery with lie algebra convolutional network , author=

  35. [43]

    Symmetry discovery beyond affine transformations , author=

  36. [44]

    Latent Space Symmetry Discovery , author=

  37. [45]

    2025 , journal=

    Geometric machine learning , author=. 2025 , journal=

    2025

  38. [46]

    Approximately equivariant graph networks , author=

  39. [47]

    Relaxed rotational equivariance via g-biases in vision , author=

  40. [48]

    Learning with Exact Invariances in Polynomial Time , author=

  41. [49]

    Frame averaging for equivariant shape space learning , author=

  42. [50]

    Designs, codes and cryptography , volume=

    Unitary designs and codes , author=. Designs, codes and cryptography , volume=. 2009 , publisher=

    2009

  43. [51]

    Physical Review A—Atomic, Molecular, and Optical Physics , volume=

    Exact and approximate unitary 2-designs and their application to fidelity estimation , author=. Physical Review A—Atomic, Molecular, and Optical Physics , volume=. 2009 , publisher=

    2009

  44. [52]

    Uniform expansion bounds for Cayley graphs of

    Bourgain, Jean and Gamburd, Alex , journal=. Uniform expansion bounds for Cayley graphs of. 2008 , publisher=

    2008

  45. [53]

    Journal of Algebra , volume=

    Tensor product Markov chains , author=. Journal of Algebra , volume=. 2020 , publisher=

    2020

  46. [54]

    Random Structures & Algorithms , volume=

    Random Cayley graphs and expanders , author=. Random Structures & Algorithms , volume=. 1994 , publisher=

    1994

  47. [55]

    Bulletin of the American Mathematical Society , volume=

    Expander graphs and their applications , author=. Bulletin of the American Mathematical Society , volume=

  48. [56]

    Discovering Group Structures via Unitary Representation Learning , author=

  49. [57]

    On the hardness of learning under symmetries , author=

  50. [58]

    The 28th International Conference on Artificial Intelligence and Statistics , year=

    Symmetry-Based Structured Matrices for Efficient Approximately Equivariant Networks , author=. The 28th International Conference on Artificial Intelligence and Statistics , year=

  51. [59]

    A Robust Kernel Statistical Test of Invariance: Detecting Subtle Asymmetries , author=

  52. [60]

    2013 , publisher=

    Representation theory: a first course , author=. 2013 , publisher=

    2013

  53. [61]

    2018 , publisher=

    High-dimensional probability: An introduction with applications in data science , author=. 2018 , publisher=

    2018

  54. [62]

    1994 , publisher=

    Character theory of finite groups , author=. 1994 , publisher=

    1994

  55. [63]

    1977 , publisher=

    Linear representations of finite groups , author=. 1977 , publisher=

    1977

  56. [64]

    Foundations of computational mathematics , volume=

    User-friendly tail bounds for sums of random matrices , author=. Foundations of computational mathematics , volume=. 2012 , publisher=

    2012

  57. [65]

    2012 , publisher=

    Theory of groups of finite order , author=. 2012 , publisher=

    2012

  58. [66]

    Proceedings of the American Mathematical Society , volume=

    Complete sets of representations of algebras , author=. Proceedings of the American Mathematical Society , volume=

  59. [67]

    Proceedings of the American Mathematical Society , volume=

    A note on theorems of Burnside and Blichfeldt , author=. Proceedings of the American Mathematical Society , volume=. 1964 , publisher=

    1964

  60. [68]

    Theory of Computing , volume =

    Alon, Noga and Lovett, Shachar , title =. Theory of Computing , volume =. 2013 , pages =

    2013

  61. [69]

    arXiv preprint arXiv:1409.7632 , year=

    On the Burnside-Brauer-Steinberg theorem , author=. arXiv preprint arXiv:1409.7632 , year=

    arXiv

  62. [70]

    arXiv preprint arXiv:2303.15596 , year=

    Symmetric powers , author=. arXiv preprint arXiv:2303.15596 , year=

    arXiv

  63. [71]

    Data Augmentation: A

    Behrooz Tahmasebi and Melanie Weber and Stefanie Jegelka , booktitle=. Data Augmentation: A

  64. [72]

    NeurIPS 2025 Workshop on Symmetry and Geometry in Neural Representations , year=

    From Finite to Infinite Groups: A Polynomial-Time Algorithm for Learning with Exact Invariances , author=. NeurIPS 2025 Workshop on Symmetry and Geometry in Neural Representations , year=

    2025

  65. [73]

    arXiv preprint arXiv:2512.11855 , year=

    Achieving Approximate Symmetry Is Exponentially Easier than Exact Symmetry , author=. arXiv preprint arXiv:2512.11855 , year=

    arXiv

  66. [74]

    Finite time

    Sarkar, Tuhin and Rakhlin, Alexander and Dahleh, Munther A , journal=jmlr, volume=. Finite time

  67. [75]

    Learning for Dynamics and Control , year=

    Learning nonlinear dynamical systems from a single trajectory , author=. Learning for Dynamics and Control , year=

  68. [76]

    Near optimal finite time identification of arbitrary linear dynamical systems , author=

  69. [77]

    Learning from many trajectories , author=

  70. [78]

    Learning without mixing: Towards a sharp analysis of linear system identification , author=

  71. [79]

    Single trajectory nonparametric learning of nonlinear dynamics , author=

  72. [80]

    arXiv preprint arXiv:2503.03014 , year=

    Learning finite symmetry groups of dynamical systems via equivariance detection , author=. arXiv preprint arXiv:2503.03014 , year=

    arXiv

  73. [81]

    Learning infinitesimal generators of continuous symmetries from data , author=

  74. [82]

    arXiv preprint arXiv:2510.01855 , year=

    Explicit discovery of nonlinear symmetries from dynamic data , author=. arXiv preprint arXiv:2510.01855 , year=

  75. [83]

    arXiv preprint arXiv:2511.09779 , year=

    A model-free method for discovering symmetry in differential equations , author=. arXiv preprint arXiv:2511.09779 , year=

  76. [84]

    Explicit Discovery of Nonlinear Symmetries from Dynamic Data , author =

  77. [85]

    Machine Learning: Science and Technology , volume=

    Machine learning for detection of equivariant finite symmetry groups in dynamical systems , author=. Machine Learning: Science and Technology , volume=. 2025 , publisher=

    2025

  78. [86]

    arXiv preprint arXiv:2505.08219 , year=

    Continuous Symmetry Discovery and Enforcement Using Infinitesimal Generators of Multi-parameter Group Actions , author=. arXiv preprint arXiv:2505.08219 , year=

  79. [87]

    arXiv preprint arXiv:2505.18798 , year=

    Governing Equation Discovery from Data Based on Differential Invariants , author=. arXiv preprint arXiv:2505.18798 , year=

    arXiv

  80. [88]

    arXiv preprint arXiv:2509.22184 , year=

    Learning Equivariant Functions via Quadratic Forms , author=. arXiv preprint arXiv:2509.22184 , year=

  81. [89]

    arXiv preprint arXiv:2512.20043 , year=

    Discovering Lie Groups with Flow Matching , author=. arXiv preprint arXiv:2512.20043 , year=

  82. [90]

    arXiv preprint arXiv:2505.12083 , year=

    Discovering Symbolic Differential Equations with Symmetry Invariants , author=. arXiv preprint arXiv:2505.12083 , year=

  83. [91]

    arXiv preprint arXiv:2506.03914 , year=

    Learning equivariant models by discovering symmetries with learnable augmentations , author=. arXiv preprint arXiv:2506.03914 , year=

  84. [92]

    Neural Networks , pages=

    Symmetry discovery for different data types , author=. Neural Networks , pages=. 2025 , publisher=

    2025

  85. [93]

    2025 , volume =

    Andrea Perin and Stephane Deny , title =. 2025 , volume =

    2025

  86. [94]

    AtlasD: Automatic Local Symmetry Discovery , author=

  87. [95]

    Latent Mixture of Symmetries for Sample-Efficient Dynamic Learning , author=

  88. [96]

    A unified framework to enforce, discover, and promote symmetry in machine learning , author=

  89. [97]

    Symmetry-informed governing equation discovery , author=

  90. [98]

    Proceedings of the National Academy of Sciences , volume=

    Learning dynamical systems from data: An introduction to physics-guided deep learning , author=. Proceedings of the National Academy of Sciences , volume=. 2024 , publisher=

    2024

  91. [99]

    arXiv preprint arXiv:2512.06315 , year=

    Control-Oriented System Identification: Classical, Learning, and Physics-Informed Approaches , author=. arXiv preprint arXiv:2512.06315 , year=

    arXiv

  92. [100]

    ACM Computing Surveys , volume=

    Physics-guided deep learning for dynamical systems: A survey , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    2025

  93. [101]

    Machine Learning: Science and Technology , volume=

    Data-driven Lie point symmetry detection for continuous dynamical systems , author=. Machine Learning: Science and Technology , volume=. 2024 , publisher=

    2024

  94. [102]

    2012 , publisher=

    Subspace identification for linear systems: Theory—Implementation—Applications , author=. 2012 , publisher=

    2012

  95. [103]

    2011 , publisher=

    Identification of dynamic systems: an introduction with applications , author=. 2011 , publisher=

    2011

  96. [104]

    Proceedings of the national academy of sciences , volume=

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems , author=. Proceedings of the national academy of sciences , volume=. 2016 , publisher=

    2016

  97. [105]

    arXiv preprint arXiv:2512.19410 , year=

    Research Program: Theory of Learning in Dynamical Systems , author=. arXiv preprint arXiv:2512.19410 , year=

  98. [106]

    Revista Matematica Iberoamericana , volume=

    Optimal control with learning on the fly: a toy problem , author=. Revista Matematica Iberoamericana , volume=

  99. [107]

    Revista Matematica Iberoamericana , volume=

    Controlling Unknown Linear Dynamics with Bounded Multiplicative Regret , author=. Revista Matematica Iberoamericana , volume=

  100. [108]

    arXiv preprint arXiv:2403.06320 , year=

    Almost Optimal Agnostic Control of Unknown Linear Dynamics , author=. arXiv preprint arXiv:2403.06320 , year=

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.