Pith. sign in

REVIEW 4 major objections 6 minor 38 references

Continuous symmetries of a stochastic differential equation can be recovered directly from trajectory data, with the learned span aligning with the analytic symmetry Lie algebra to within a few degrees on every benchmark.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-05 00:39 UTC pith:BI3IQLMK

load-bearing objection First credible end-to-end pipeline for SDE Lie symmetry discovery; the theory and losses are genuine, but the empirical validation leans on an unverified surrogate-derivative assumption and some large principal angles. the 4 major comments →

arxiv 2608.01582 v1 pith:BI3IQLMK submitted 2026-08-03 cond-mat.stat-mech cond-mat.dis-nncs.LGmath-phmath.MP

LieStoNet: Learning Lie Symmetries from Spatiotemporal Data for Stochastic Dynamical Systems

Shida Liu , Abhishek Gupta , Sumit Sinha , L. Mahadevan This is my paper
classification cond-mat.stat-mech cond-mat.dis-nncs.LGmath-phmath.MP MSC 60H1035A3068T07 PACS 05.10.Gg
keywords Lie-point symmetriesstochastic differential equationssymmetry discoveryneural surrogatesLie algebrasdetermining equationsFokker-Planck equationItô calculus
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Symmetries of a stochastic dynamical system, if known, would let models exploit invariance for better sample efficiency, robustness, and out-of-distribution generalization; but unlike the deterministic case, there has been no data-driven way to discover them. LieStoNet claims to close that gap: given only discretely sampled trajectories, it fits differentiable neural surrogates for drift and diffusion, then learns infinitesimal symmetry generators by enforcing the SDE determining equations together with Lie-algebra consistency (bracket closure, Jacobi identity, antisymmetry, bilinearity, independence). The paper reports correct symmetry-algebra dimensions on four canonical SDEs and a ten-dimensional Brownian system, spans aligned with the analytic ground truth, and a stable two-dimensional algebra on high-frequency BTC/USDT data validated by finite-flow preservation. If the claim holds, stochastic symmetries become things a model can discover rather than assume, opening the door to symmetry-constrained surrogates, data augmentation, and interpretable structure for noisy dynamics.

Core claim

The central claim: the projectable Lie-point symmetry algebra of an unknown Itô SDE is learnable end-to-end from trajectory data. A neural surrogate (f_θ, σ_θ) is fit by local maximum likelihood over Euler increments; then m generators X_i = τ_i(t)∂_t + ξ_i(t,x)∂_x are learned as neural fields under two loss families — SDE-validity losses enforcing the Gaeta–Quintero determining equations (two PDEs linking generator components to drift, diffusion, and their derivatives) plus a finite-ε pushforward check, and Lie-algebra losses enforcing closure, Jacobi, antisymmetry, bilinearity, independence. Dimension is selected as the first minimum of the post-training closure loss L1 over a bounded swee

What carries the argument

The carrying object is the projectable Lie-point generator X = τ(t)∂_t + ξ(t,x)·∂_x, an infinitesimal transformation whose time component depends only on time, validated by the Gaeta–Quintero SDE determining equations — a pair of PDEs linking generator components to drift, diffusion, and their derivatives. Enforced through a learned neural surrogate (f_θ, σ_θ), these PDEs become an optimization loss (L6) with all derivatives from autodiff. Around this core, structural losses (bracket closure with constant coefficients, Jacobi identity, skew-symmetry, bilinearity, functional independence) organize generators into a genuine Lie algebra, and the closure loss L1(m) doubles as the dimension-selec

Load-bearing premise

The load-bearing premise is that the derivatives of the learned neural surrogate — ∂_t f, ∂_x f, ∂_xx σ — match the true drift and diffusion derivatives on the evaluation region (assumed in Section 3.1 and in loss L6 of Appendix C.2, and acknowledged in the paper's own limitations as 'fidelity and coverage' of the surrogate); if they drift, L6 enforces symmetries of the wrong equation and the recovered generators are fitting artifacts.

What would settle it

Train the full pipeline on data from an SDE with a known symmetry algebra (for instance dx = x dt + dW), but sabotage the surrogate's derivatives rather than its values — add a high-frequency component to f_θ that leaves pointwise agreement intact while spoiling ∂_x f or ∂_xx σ — and check the recovered span. If principal angles stay small despite large surrogate-derivative error, loss L6 is not the operative constraint; if they blow up, surrogate-derivative fidelity is confirmed as the load-bearing link. A cleaner variant: report relative errors between the trained surrogate's autodiff deriva

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Discovered generators can serve as structural priors: the paper's stated motivation is that known invariances improve sample efficiency, robustness, and out-of-distribution generalization, so recovered SDE symmetries are direct inputs for symmetry-constrained stochastic surrogates and equivariant models of noisy dynamics.
  • Symmetry dimension becomes a data-driven output, not an input: the closure-loss minimum selects the correct algebra dimension across all benchmarks, and theoretical bounds (at most 3 for scalar SDEs, at most n+2 for full-rank n-dimensional systems) make the sweep finite rather than open-ended.
  • The learned SDE generators embed into the learned Fokker–Planck generator span (Example 1 principal angles 2.00°–18.81°), empirically confirming the theoretical nesting of the Itô symmetry algebra inside the FP symmetry algebra.
  • The unmodified pipeline transfers to real-world data: on high-frequency BTC/USDT ticks it yields a stable two-dimensional symmetry algebra whose finite pushes preserve drift, diffusion, and the Euler–Maruyama residual distribution far better than random-push controls.
  • Recovery scales beyond low-dimensional benchmarks: the ten-dimensional Brownian SDE produces a 12-dimensional projectable algebra with principal angles mostly below 2°, consistent with the full-rank dimension bound.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the pipeline's reliance on surrogate derivatives is its sharpest untested seam; a decisive stress test would compare autodiff derivatives (∂_x f, ∂_xx σ) of the fitted surrogate against the true coefficients' analytic derivatives on the evaluation grid, since the paper reports recovery quality but not surrogate-derivative fidelity.
  • Editorial inference: the projectable restriction (τ = τ(t)) is the method's structural boundary; removing it heads directly toward random Lie-point symmetries, which the paper cites as a broader class, and would require new determining equations and new validation metrics — a natural follow-up.
  • Editorial inference: nothing in the pipeline is specific to forward simulation; applied to partially observed, coarse-grained, or nonlinearly transformed observations, the same two-stage loop could discover symmetries of effective or reduced stochastic models, though the surrogates would need to fit those coordinates first.
  • Editorial inference: the dimension-selection criterion was tested only on SDEs with small algebras; running the m-sweep on systems engineered to have nearly redundant or accidentally closing generator sets would probe how sharply the L1 landscape separates true dimension from spurious minima.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. LieStoNet proposes an end-to-end pipeline for discovering projectable Lie-point symmetries of Itô SDEs from raw trajectory data. In Stage 1, a neural drift/diffusion surrogate is fit to increments via local Gaussian likelihood. In Stage 2, a set of m generator networks (τ_i(t), ξ_i(t,x)) is trained by minimizing a weighted sum of Lie-algebra structural losses (closure, Jacobi, skew-symmetry, bilinearity, independence) and SDE-validity losses: the Gaeta–Quintero determining equations (L6) and a finite-step pushforward residual (L7). The symmetry dimension is selected as the m that minimizes the post-training closure loss L1. The method is evaluated on four analytic SDEs, a 10-dimensional Brownian benchmark, and BTC/USDT high-frequency data, with principal-angle span alignment and after-push residuals as the main quantitative checks. The appendices contain full analytic symmetry derivations, loss definitions, implementation details, ablations, and weight-sensitivity studies.

Significance. If the central claim holds, LieStoNet would be a meaningful first step in a genuinely unexplored area: data-driven discovery of continuous symmetries of stochastic dynamics, grounded in the rigorous Gaeta–Quintero framework rather than ad hoc invariance heuristics. The paper is careful to separate the SDE-level and Fokker–Planck-level notions of symmetry, provides basis-invariant evaluation via principal angles, includes an independent finite-step after-push check, reports a higher-dimensional 12-generator example, and ships code and complete analytic derivations for all benchmarks. These are real strengths. However, the load-bearing validation is incomplete: the determining-equation loss is computed on a neural surrogate whose derivatives are never checked against the true coefficients, and the independent after-push check lacks null baselines on synthetic data. The reported span alignments also vary widely (6.26°–19.37° at the selected dimension), and the method is highly sensitive to loss weights. These issues do not invalidate the framework but they do preclude accepting the abstract's strong recovery claim as established.

major comments (4)
  1. [§3.1, Appendix C.2 (L6)] The central claim is that LieStoNet recovers generators of the data-generating SDE, but the only term that ties the generators to that SDE is L6, which evaluates Theorem 2.1 on the learned surrogate fθ, σθ and its autodiff derivatives. The manuscript never validates that ∂t fθ, ∂x fθ, ∂xx σθ, ∂t σθ approximate the true coefficient derivatives on the evaluation region. If they do not, minimizing L6 yields symmetries of the surrogate, not of the original SDE. Appendix H gives a perturbation bound, but it is conditional on the C1 error of the surrogate and is never instantiated for the actually trained surrogates; Table 9 perturbs an already-trained surrogate and does not measure this error. Please report the empirical C1 deviation of the fitted fθ, σθ from the true coefficients on the evaluation box, and recompute the L6 residual using the true derivatives for the learned generators. Witho
  2. [§5.1, Table 6] The after-push check is presented as independent validation that the learned generators map the analytic SDE to itself. This is a good idea, but for the synthetic benchmarks there is no random-push control. The residuals in Table 6 are small in absolute terms, but smallness is only meaningful relative to a null distribution: a random vector field of comparable magnitude might produce similarly small residuals given 500 trajectories and a finite evaluation box. The BTC/USDT experiments do include random controls (Tables 7–8), so the same protocol should be applied to Examples 1–4. Without synthetic baselines, the claim that 'the learned generators induce finite transformations that approximately map the SDE to itself' is not calibrated.
  3. [Table 4, §4.2] The headline numbers are not uniformly small: the maximum principal angles at the selected dimension are 6.26°, 19.37°, 7.64°, and 14.29°. For Example 2, a 19.37° maximal angle means one learned direction is substantially misaligned with the analytic span; calling this 'consistent with the ground-truth symmetry algebra' needs a quantitative threshold and ideally confidence intervals over the independent runs whose distributions are shown in Figure 3. Please clarify whether Table 4 reports a representative run, the mean over seeds, or the median, and discuss what angular threshold would be considered successful recovery. Without this, the abstract's wording is stronger than the reported evidence.
  4. [Appendix Q] The sensitivity of the maximum principal angle to the two main loss weights is large and non-monotone: across the sweep in Table 16 the angle ranges from 14.71° to 55.31°. This is a free-parameter dependence on the reported results. The main-text ablation in Appendix O uses a single fixed weight vector, and the weights differ across Examples 1–4 (§P.0.2). Please report how the selected results in Table 4 vary under the weight grid, or justify a principled weight-selection rule; otherwise the reported angles may reflect tuned losses rather than the method's intrinsic accuracy.
minor comments (6)
  1. [§5.1, Table 6] The units switch from 10^-4 for ε=1,3 to 10^-3 for ε=5 without explanation; please state this in the caption so the reader can compare across columns.
  2. [Appendix P, Table 11] Example 2 uses an i.i.d. increment dataset rather than simulated trajectories, which is a different observational setting from the other examples. This should be disclosed in the main text, since the method is advertised as learning from spatiotemporal trajectories.
  3. [Figure 2–3] Please specify the number of random seeds represented in the box plots and whether Table 4 values are from a single seed or averaged; the current text is ambiguous.
  4. [§5.2] The dimension-selection criterion for BTC/USDT is described as 'the first nontrivial minimum' of L1. The word 'nontrivial' is not defined; please give an operational threshold (e.g., relative drop, significance against repeated runs) so that the reported m*=2 is reproducible.
  5. [Appendix C.2, L7] The prolonged pushforward system is stated without derivation. Since L7 is a novel finite-step loss and is shown in the ablation to be important, a short derivation linking the propagated (µ, ς) to the Itô transformation formulas of Gaeta–Quintero would improve rigor.
  6. [§3.3, L1] The label 'fixed' for L1 is confusing: the coefficients c^k_ij(t,x) are fitted and penalized for non-constancy. Consider renaming to 'closure with slowly varying structure coefficients' to avoid implying they are fixed exogenously.

Circularity Check

0 steps flagged

No significant circularity: learned symmetries are validated against independent analytic generators; surrogate-derivative mismatch is a correctness risk, not a circular reduction.

full rationale

The derivation chain is self-contained against external benchmarks and does not reduce to its inputs. Stage 1 fits a neural surrogate (f_theta, sigma_theta) from trajectory increments; Stage 2 trains projectable generators by minimizing the Gaeta-Quintero determining-equation residuals (L6) on that surrogate together with Lie-algebraic structure losses and a finite-flow pushforward loss (L7). The reported evaluation is independent: principal angles compare the learned generator span to analytic generators derived in Appendices J-M by solving the same determining equations for the known coefficients, not from the data, and the ground-truth dimension is not fed into the pipeline (it is selected as the L1 minimum). No fitted parameter is renamed as a prediction, and no load-bearing self-citation appears; the theoretical citations are to Gaeta & Quintero (1999) and Kozlov, not to the present authors. The skepticism about unvalidated surrogate derivatives is a correctness/robustness gap, not circularity: Appendix H provides a C1 perturbation bound, Section 5.1 performs an after-push check against analytic coefficients on data pushed from the analytic SDE, and the paper's 'Limitations and outlook' section acknowledges that LieStoNet depends on surrogate fidelity. Even if that assumption were violated, the failure mode would be inaccurate recovery, not the method's outputs being equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The method's central claim rests on the Gaeta-Quintero determining equations (taken as definition of SDE symmetry), the Kozlov dimension bounds, the Euler-Maruyama approximation for surrogate fitting, and the accuracy of autodiff derivatives of the surrogate. The main tuned quantities are the loss weights, which the paper itself shows are influential, the finite-step size in the pushforward loss, and the generator count. No invented physical entities are introduced.

free parameters (4)
  • loss weights w1..w7 = Examples 1-3: (1,0.1,0.1,0.1,0.1 or 0.5,1,0.1); Example 4: scheduled w6=10, w7=2, w5=2; grid search over (w6,w7)
    These weights balance algebraic structure vs. SDE validity. Appendix Q shows the maximum principal angle varies from 14.71 to 55.31 degrees over the (w6,w7) grid, so the chosen operating point materially affects the results.
  • finite-step size epsilon in L7 = 1e-2 with steps=1
    The pushforward consistency loss uses a hand-chosen finite step size and a single integration step; this choice affects the non-local validation strength.
  • generator count m = swept 1..3 (and up to 12 for 10D Brownian), selected by minimum L1
    The dimension is an output of the method, but the sweep range and the L1-minimization selection rule are choices. The L1 criterion is also part of the training objective, so the selection is not independent of training dynamics.
  • surrogate architecture = MLPs of widths 64 and 32, tanh/swish activations, softplus diffusion floors
    Model capacity and activation functions are manually selected per example; the surrogate's derivative fidelity is not directly validated.
axioms (5)
  • domain assumption Theorem 2.1 SDE determining equations define the symmetry condition (Gaeta and Quintero, 1999)
    The central loss L6 enforces these equations as the definition of SDE symmetry. The theorem is cited from prior literature, not proved in the paper.
  • domain assumption Kozlov dimension bounds: scalar Ito SDEs have Lie algebras of dimension at most 3; n-dimensional full-rank diffusion has maximal dimension n+2
    Used in Section 3.4 to justify the finite sweep over generator count m. These bounds come from cited literature, not established here.
  • standard math Euler-Maruyama increment Gaussianity for surrogate fitting
    The surrogate is trained by assuming Δx | (t,x) ~ N(f Δt, σ^2 Δt). This is a standard weak approximation valid in the Δt to 0 limit; the paper uses Δt=0.01.
  • domain assumption Autodiff derivatives of the neural surrogate approximate derivatives of the true drift and diffusion
    L6 and L7 compute loss terms using partial derivatives of f_θ and σ_θ. The paper does not directly measure derivative error, only overall surrogate perturbation effects.
  • domain assumption Restriction to projectable generators X = τ(t)∂t + ξ(t,x)∂x
    The method deliberately excludes state-dependent time reparameterizations. This is an explicit modeling restriction, stated in Section 2.2.

pith-pipeline@v1.3.0-daily-deepseek · 27619 in / 14401 out tokens · 156134 ms · 2026-08-05T00:39:08.528221+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of LieStoNet: Learning Lie Symmetries from Spatiotemporal Data for Stochastic Dynamical Systems." pith.science (2026). https://pith.science/paper/BI3IQLMK

@misc{pith2026260801582,
  author       = {Pith},
  title        = {Pith review of: LieStoNet: Learning Lie Symmetries from Spatiotemporal Data for Stochastic Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BI3IQLMK}},
  note         = {Machine review of arXiv:2608.01582}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Symmetry is central to modern machine learning and physics: invariances and equivariances improve sample efficiency, robustness, and out-of-distribution generalization, while symmetry principles guide scientific modeling. Yet for stochastic dynamical systems the relevant continuous symmetries are rarely known, and symmetry discovery for SDEs has remained essentially unexplored. We introduce \textit{LieStoNet}, an end-to-end, \emph{template-free} framework for discovering Lie-point symmetries of SDEs directly from spatiotemporal trajectories, without prespecifying symmetry groups, templates, or canonical coordinates. Building on the seminal SDE Lie-symmetry theory of Gaeta and Quintero (1999), which formalizes Lie-point SDE symmetries and their relation to Fokker-Planck symmetries, LieStoNet learns neural surrogates for drift and diffusion from increments, then learns projectable generators by enforcing the SDE determining equations, separately regularizing for closure under Lie brackets, adherence to the Lie algebra axioms (bilinearity, antisymmetry, Jacobi), and a non-redundant independent basis. The surrogate also defines an associated Fokker-Planck equation, enabling optional discovery of its Lie-point symmetries in parallel. Across multiple canonical SDEs with known analytic symmetries, LieStoNet recovers generators consistent with the ground-truth symmetry algebra, providing interpretable symmetry discovery for noisy dynamics. Code is available at \href{https://github.com/sumit-sinha-seas/LieStoNet_Final.git}{this link}.

Figures

Figures reproduced from arXiv: 2608.01582 by Abhishek Gupta, L. Mahadevan, Shida Liu, Sumit Sinha.

Figure 1
Figure 1. Figure 1: Pipeline for LieStoNet ∆t = tn+1 − tn, and define increments ∆x (k) n := x (k) (tn+1) − x (k) (tn). We fit a differentiable surrogate SDE dxt = fθ(t, xt) dt + σθ(t, xt) dWt, (7) where fθ and σθ > 0 are parameterized by neural networks (details in the appendix P.0.1). The role of this surrogate is to provide smooth coefficient functions so that all derivatives required by the symmetry constraints in Theorem… view at source ↗
Figure 2
Figure 2. Figure 2: Post-training Lie bracket closure loss L1 vs. number of generators m across four experiments. Colors denote independent runs with identical settings (different random seeds), showing the minimum is consistently attained at the ground-truth m rather than a one-off. • L3 (Skew-symmetry). Enforces antisymmetry of the bracket: swapping generator order flips the commuta￾tor sign. • L4 (Bilinearity). Enforces li… view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of combination-wise maximum principal angles. Number of learned generators (m) vs. principal angles. The blue boxes denote the interquartile range; the orange horizontal lines inside the box denote the median; the green up triangles and the blue down triangles denote the maximum and minimum of maximum principal angles, respectively; the orange dots denote the mean. 7 [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 4
Figure 4. Figure 4: Learned generator components for Example 1. Part (a): temporal component τi(t). Part (b): spatial component ξi(t, x) visualized as heat maps. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Learned generator components for Example 3. Part (a): τi(t). Part (b): ξi(t, x) heat maps. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Learned generator components for Example 2. Part (a): τi(t). Part (b): ξi(t, x) heat maps. (a) τi(t) (Example 3). (b) ξi(t, x) (Example 3) [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Learned generator components for Example 4. Part (a): τi(t). Part (b): ξi(t, x) heat maps. H. Sensitivity to Surrogate Approximation Error LieStoNet computes the symmetry losses using the learned neural SDE surrogate ( ˆf, σˆ), so surrogate misspecification can affect the recovered generators. Let (f, σ) denote the true drift and diffusion, G⋆ a true symmetry generator, Gˆ a learned generator, and Df,σ(·) … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages

  1. [1]

    Proceedings of the 41st International Conference on Machine Learning , series=

    Latent Space Symmetry Discovery , author=. Proceedings of the 41st International Conference on Machine Learning , series=. 2024 , publisher=

    work page 2024

  2. [2]

    Machine Learning: Science and Technology , volume=

    Deep learning symmetries and their Lie groups, algebras, and subalgebras from first principles , author=. Machine Learning: Science and Technology , volume=. 2023 , publisher=

    work page 2023

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    Learning infinitesimal generators of continuous symmetries from data , author=. Advances in Neural Information Processing Systems , volume=

    work page

  4. [4]

    Advances in neural information processing systems , volume=

    Learning invariances in neural networks from training data , author=. Advances in neural information processing systems , volume=

  5. [5]

    Physical Review Letters , volume=

    Machine learning hidden symmetries , author=. Physical Review Letters , volume=. 2022 , publisher=

    work page 2022

  6. [6]

    Proceedings of the 42nd International Conference on Machine Learning , series=

    Explicit Discovery of Nonlinear Symmetries from Dynamic Data , author=. Proceedings of the 42nd International Conference on Machine Learning , series=. 2025 , publisher=

    work page 2025

  7. [7]

    Physics Today , volume=

    Symmetry in physics: Wigner's legacy , author=. Physics Today , volume=. 1995 , publisher=

    work page 1995

  8. [8]

    I , author=

    Asymptotically free gauge theories. I , author=. Physical Review D , volume=. 1973 , publisher=

    work page 1973

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Clebsch--gordan nets: a fully fourier space spherical convolutional neural network , author=. Advances in Neural Information Processing Systems , volume=

    work page

  10. [10]

    International conference on machine learning , pages=

    E (n) equivariant graph neural networks , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  11. [11]

    Physical Review Letters , volume=

    Equivariant flow-based sampling for lattice gauge theory , author=. Physical Review Letters , volume=. 2020 , publisher=

    work page 2020

  12. [12]

    International conference on machine learning , pages=

    Group equivariant convolutional networks , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  13. [13]

    Physical Review D , volume=

    Sampling using SU (N) gauge equivariant flows , author=. Physical Review D , volume=. 2021 , publisher=

    work page 2021

  14. [14]

    Advances in neural information processing systems , volume=

    Se (3)-transformers: 3d roto-translation equivariant attention networks , author=. Advances in neural information processing systems , volume=

  15. [15]

    arXiv preprint arXiv:1802.08219 , year=

    Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds , author=. arXiv preprint arXiv:1802.08219 , year=

    Pith/arXiv arXiv

  16. [16]

    1993 , publisher=

    Applications of Lie groups to differential equations , author=. 1993 , publisher=

    work page 1993

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Symmetry discovery beyond affine transformations , author=. Advances in Neural Information Processing Systems , volume=

    work page

  18. [18]

    Advances in Neural Information Processing Systems , volume=

    Noether networks: meta-learning useful conserved quantities , author=. Advances in Neural Information Processing Systems , volume=

  19. [19]

    Advances in neural information processing systems , volume=

    A general theory of equivariant cnns on homogeneous spaces , author=. Advances in neural information processing systems , volume=

    work page

  20. [20]

    Advances in neural information processing systems , volume=

    Hamiltonian neural networks , author=. Advances in neural information processing systems , volume=

  21. [21]

    arXiv preprint arXiv:2002.03061 , year=

    Incorporating symmetry into deep dynamics models for improved generalization , author=. arXiv preprint arXiv:2002.03061 , year=

    Pith/arXiv arXiv 2002

  22. [22]

    arXiv preprint arXiv:2107.01272 , year=

    Physics-guided deep learning for dynamical systems: A survey , author=. arXiv preprint arXiv:2107.01272 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2310.02299 , year=

    Discovering symmetry breaking in physical systems with relaxed group convolution , author=. arXiv preprint arXiv:2310.02299 , year=

    Pith/arXiv arXiv

  24. [24]

    Proceedings of the 42nd International Conference on Machine Learning , series=

    Symmetry-Driven Discovery of Dynamical Variables in Molecular Simulations , author=. Proceedings of the 42nd International Conference on Machine Learning , series=. 2025 , publisher=

    work page 2025

  25. [25]

    Advances in Neural Information Processing Systems , volume=

    Symmetry-informed governing equation discovery , author=. Advances in Neural Information Processing Systems , volume=

    work page

  26. [26]

    Proceedings of the 40th International Conference on Machine Learning , series=

    Generative Adversarial Symmetry Discovery , author=. Proceedings of the 40th International Conference on Machine Learning , series=. 2023 , publisher=

    work page 2023

  27. [27]

    Physical Review E , volume=

    Finding discrete symmetry groups via machine learning , author=. Physical Review E , volume=. 2024 , publisher=

    work page 2024

  28. [28]

    Proceedings of the 42nd International Conference on Machine Learning , series=

    AtlasD: Automatic Local Symmetry Discovery , author=. Proceedings of the 42nd International Conference on Machine Learning , series=. 2025 , publisher=

    work page 2025

  29. [29]

    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=

    work page 2025

  30. [30]

    arXiv preprint arXiv:2505.08219 , year=

    Lie Group Symmetry Discovery and Enforcement Using Vector Fields , author=. arXiv preprint arXiv:2505.08219 , year=

    work page arXiv

  31. [31]

    Journal of Physics A: Mathematical and General , volume=

    Lie-point symmetries and stochastic differential equations , author=. Journal of Physics A: Mathematical and General , volume=. 1999 , publisher=

    work page 1999

  32. [32]

    Journal of Mathematical Physics , volume=

    Random Lie-point symmetries of stochastic differential equations , author=. Journal of Mathematical Physics , volume=. 2017 , publisher=

    work page 2017

  33. [33]

    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=

    work page 2024

  34. [34]

    Proceedings of the 39th International Conference on Machine Learning (ICML) , series=

    Lie Point Symmetry Data Augmentation for Neural PDE Solvers , author=. Proceedings of the 39th International Conference on Machine Learning (ICML) , series=. 2022 , publisher=

    work page 2022

  35. [35]

    Journal of Physics A: Mathematical and Theoretical , volume=

    Symmetries of Systems of Stochastic Differential Equations with Diffusion Matrices of Full Rank , author=. Journal of Physics A: Mathematical and Theoretical , volume=. 2010 , publisher=

    work page 2010

  36. [36]

    Journal of Physics A: Mathematical and Theoretical , volume=

    The Group Classification of a Scalar Stochastic Differential Equation , author=. Journal of Physics A: Mathematical and Theoretical , volume=. 2010 , publisher=

    work page 2010

  37. [37]

    Journal of Physics A: Mathematical and Theoretical , volume=

    On Maximal Lie Point Symmetry Groups Admitted by Scalar Stochastic Differential Equations , author=. Journal of Physics A: Mathematical and Theoretical , volume=. 2011 , publisher=

    work page 2011

  38. [38]

    Journal of Mathematical Physics , volume=

    Symmetries of linear systems of second-order ordinary differential equations , author=. Journal of Mathematical Physics , volume=. 1988 , publisher=

    work page 1988