Pith. sign in

REVIEW 2 major objections 1 minor 52 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · grok-4.3

Nonreversible perturbations of Fokker-Planck dynamics can be cast as gauge fields that preserve the stationary density while changing the relaxation spectrum.

2026-06-27 22:40 UTC pith:XBWL6X5E

load-bearing objection The paper casts nonreversible Fokker-Planck perturbations as gauge fields via antisymmetric tensors and adds an actor-critic step to pick finite gauge strength, but the claimed generality of that representation is asserted without a completeness proof. the 2 major comments →

arxiv 2606.06412 v1 pith:XBWL6X5E submitted 2026-06-04 cond-mat.dis-nn quant-phstat.ML

Nonreversible Gauge Fields in Fokker--Planck Dynamics: Supersymmetric Hamiltonians and Learned Finite Forces

Masayuki Ohzeki This is my paper
classification cond-mat.dis-nn quant-phstat.ML
keywords Fokker-Planck dynamicsnonreversible perturbationsgauge fieldssupersymmetric Hamiltoniansactor-critic learningOrnstein-Uhlenbeck processstochastic gradient descent
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.

The paper establishes that nonreversible changes to Fokker-Planck evolution can be written as gauge fields generated by antisymmetric tensors. These fields leave the long-time invariant probability unchanged but alter how the system approaches it. When detailed balance is present, the reversible operator maps to a supersymmetric Hamiltonian via similarity transformation, and the gauges enter as non-Hermitian corrections that keep the zero eigenvalue fixed. The same language covers circulating currents, relaxation gaps, and finite control costs. Finite gauge strengths are then chosen by minimizing a regularized finite-time objective learned with an actor-critic method, which recovers known optima on solvable benchmarks.

Core claim

Admissible nonreversible perturbations that preserve the stationary density are represented by antisymmetric tensor fields acting as gauge currents; these appear as non-Hermitian perturbations to the supersymmetric Hamiltonian obtained from the reversible case, leaving the zero mode intact while shifting the rest of the spectrum. The Ohzeki-Ichiki force provides a constant symplectic example whose infinite-strength limit recovers Hamiltonian dynamics. Because the continuous-time spectral gap alone does not fix a finite strength, a finite-time objective is introduced and optimized by actor-critic learning, which exactly recovers the Lyapunov-equation optimum on an anisotropic Gaussian Ornstei

What carries the argument

Gauge fields realized by antisymmetric tensor fields that generate probability currents while leaving the stationary density invariant.

Load-bearing premise

Any admissible nonreversible gauge current can be written exactly as an antisymmetric tensor field that leaves the stationary density unchanged.

What would settle it

In the exactly solvable anisotropic Gaussian Ornstein-Uhlenbeck process, check whether the gauge strength returned by the actor-critic procedure coincides with the value that minimizes the finite-time cost derived from the Lyapunov equation.

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

If this is right

  • The continuous-time spectral gap supplies no unique finite gauge strength, so a finite-time regularized objective is required.
  • An actor-critic procedure can learn the gauge that recovers the Lyapunov optimum on the Ornstein-Uhlenbeck benchmark.
  • The same constrained selection appears in a nonconvex double-well metastable landscape.
  • Mini-batch stochastic gradient descent corresponds to a Fokker-Planck system whose diffusion tensor encodes the noise and whose metric may carry nonequilibrium currents.
  • Adaptive methods such as Adam arise as particular metric choices that can include such currents.

Where Pith is reading between the lines

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

  • Viewing training dynamics as tunable gauge fields suggests a route to accelerate convergence by deliberately adding learned nonreversible terms.
  • The framework supplies a common operator language that could link hypocoercive estimates in kinetic theory to practical optimization schedules.
  • Nonequilibrium currents already latent in common adaptive optimizers might be diagnosed and adjusted for faster mixing without changing the target distribution.

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

2 major / 1 minor

Summary. The paper formulates stationary-density-preserving nonreversible perturbations of Fokker-Planck dynamics as gauge fields represented by antisymmetric tensor fields. These deform relaxation spectra while leaving the invariant measure fixed. When detailed balance holds, a similarity transform maps the reversible operator to a supersymmetric Witten-Laplacian Hamiltonian; nonreversible gauges appear as non-Hermitian perturbations. The Ohzeki-Ichiki force is identified as a constant symplectic example. A finite-time regularized objective plus actor-critic learning selects finite gauge strength, recovering the Lyapunov optimum on an exactly solvable anisotropic Gaussian OU process and illustrated on a double-well potential. Stochastic gradient methods are interpreted as Fokker-Planck systems with effective diffusion and possible nonequilibrium currents.

Significance. If the representation of all admissible currents is shown to be exhaustive and the finite-time selection procedure is shown to be general, the work supplies a common operator language linking relaxation gaps, circulating currents, hypocoercive acceleration, and optimization dynamics. The machine-learning interpretation of SGD/Adam as nonequilibrium Fokker-Planck systems is a potentially useful bridge between sampling theory and training algorithms.

major comments (2)
  1. [Abstract] Abstract (paragraph on representation of gauge currents): the claim that every stationary-density-preserving nonreversible perturbation admits an antisymmetric-tensor gauge representation is asserted without a derivation of completeness. In domains with boundaries or nontrivial topology, divergence-free currents may exist that cannot be expressed as the divergence of an antisymmetric tensor while remaining admissible under the Fokker-Planck operator; if such currents are excluded, the spectral-deformation statements apply only to a subclass.
  2. [Benchmarks section] The finite-time objective is introduced to select gauge strength, yet the manuscript provides no general proof that the learned gauge recovers the Lyapunov optimum outside the two benchmark cases; the central claim that the procedure yields the optimal finite gauge therefore rests on the specific examples rather than a general argument.
minor comments (1)
  1. Notation for the antisymmetric tensor field and its relation to the current should be introduced with an explicit equation in the main text rather than only in the abstract.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments. We respond point by point to the major comments below.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraph on representation of gauge currents): the claim that every stationary-density-preserving nonreversible perturbation admits an antisymmetric-tensor gauge representation is asserted without a derivation of completeness. In domains with boundaries or nontrivial topology, divergence-free currents may exist that cannot be expressed as the divergence of an antisymmetric tensor while remaining admissible under the Fokker-Planck operator; if such currents are excluded, the spectral-deformation statements apply only to a subclass.

    Authors: We agree that the manuscript asserts the representation without an explicit derivation of completeness. The formulation defines admissible gauge currents via antisymmetric tensor fields in the settings of the paper (typically unbounded Euclidean domains or periodic tori). To address the concern, we will add a concise derivation and domain statement in a revised section or appendix, making clear that the spectral results apply to this class of perturbations. revision: yes

  2. Referee: [Benchmarks section] The finite-time objective is introduced to select gauge strength, yet the manuscript provides no general proof that the learned gauge recovers the Lyapunov optimum outside the two benchmark cases; the central claim that the procedure yields the optimal finite gauge therefore rests on the specific examples rather than a general argument.

    Authors: The referee correctly observes that a general proof is absent. The actor-critic procedure is constructed to apply broadly, with the Gaussian benchmark recovering the Lyapunov optimum exactly and the double-well providing a nonconvex illustration. We will revise the text to state explicitly that optimality is demonstrated on these cases and that a universal proof for arbitrary potentials is not provided. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's central formulation represents nonreversible perturbations as gauge fields via antisymmetric tensors and introduces a finite-time objective whose optimum is validated against an independent Lyapunov-equation solution in the Gaussian benchmark. The Ohzeki-Ichiki force appears only as an illustrative constant-symplectic example, not as a load-bearing premise whose validity is assumed without external support. No derivation step equates a claimed prediction to a fitted parameter or self-citation by construction, and the benchmarks remain falsifiable against external optima.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The central claim rests on the existence of a similarity transformation that maps the reversible Fokker-Planck operator to a Witten-Laplacian supersymmetric Hamiltonian, on the representation of gauge currents by antisymmetric tensors that leave the stationary measure invariant, and on the introduction of a finite-time regularized objective whose optimum is learned by actor-critic. These are domain assumptions in nonequilibrium statistical mechanics plus one new modeling choice (the finite-time objective).

free parameters (1)
  • gauge strength
    The continuous-time spectral gap does not select a finite value, so a finite-time objective is introduced whose optimum must be learned or fitted.
axioms (2)
  • domain assumption A similarity transformation exists that maps the reversible Fokker-Planck operator to a Witten-Laplacian-type supersymmetric Hamiltonian when detailed balance holds.
    Invoked in the abstract when nonreversible gauges are described as non-Hermitian perturbations preserving the zero mode.
  • domain assumption Admissible nonreversible gauge currents can be represented exactly by antisymmetric tensor fields that preserve the stationary density.
    Stated directly in the abstract as the representation of gauge currents.
invented entities (1)
  • nonreversible gauge field no independent evidence
    purpose: To deform relaxation spectra of Fokker-Planck dynamics while keeping the invariant state fixed.
    New postulated object introduced to unify nonreversible perturbations; no independent experimental handle given in the abstract.

pith-pipeline@v0.9.1-grok · 5787 in / 1751 out tokens · 22132 ms · 2026-06-27T22:40:19.548074+00:00 · methodology

0 comments
read the original abstract

We formulate stationary-density-preserving nonreversible perturbations of Fokker--Planck dynamics as gauge fields that deform relaxation spectra while leaving the invariant state fixed. When detailed balance holds, a similarity transformation maps the reversible Fokker--Planck operator to a Witten-Laplacian-type supersymmetric Hamiltonian; nonreversible gauges then appear as non-Hermitian perturbations that preserve the zero mode but modify the excited spectrum. This operator viewpoint gives a common language for relaxation gaps, circulating probability currents, hypocoercive acceleration, and finite control costs. We represent admissible gauge currents by antisymmetric tensor fields and identify the detailed-balance-violating Ohzeki--Ichiki force as a constant symplectic example whose infinite-strength limit is Hamiltonian dynamics. The continuous-time spectral gap alone does not select a finite gauge strength, so we introduce a finite-time regularized objective and an actor--critic procedure for learning the gauge. An exactly solvable anisotropic Gaussian Ornstein--Uhlenbeck benchmark separates the spectral transition from the finite-time optimum and shows that the learned gauge recovers the Lyapunov-equation optimum. A double-well benchmark then illustrates the same constrained selection in a nonconvex metastable landscape. Stochastic gradient methods enter this framework as physically relevant Fokker--Planck systems: mini-batch noise acts as an effective diffusion tensor, and adaptive methods such as Adam correspond to metric choices with possible nonequilibrium currents.

Figures

Figures reproduced from arXiv: 2606.06412 by Masayuki Ohzeki.

Figure 1
Figure 1. Figure 1: FIG. 1. Finite-gauge learning in the anisotropic Gaussian OU benchmark (color online). In the left [PITH_FULL_IMAGE:figures/full_fig_p025_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Mean trajectory in the anisotropic Gaussian OU benchmark from [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Relaxation of the mean position in the double-well model (color online). The black, blue, [PITH_FULL_IMAGE:figures/full_fig_p028_3.png] 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

52 extracted references · 8 canonical work pages · 4 internal anchors

  1. [1]

    Ichiki and M

    A. Ichiki and M. Ohzeki, Violation of detailed balance accelerates relaxation, Physical Review E88, 020101(R) (2013)

    2013

  2. [2]

    Ohzeki and A

    M. Ohzeki and A. Ichiki, Langevin dynamics neglecting detailed balance condition, Physical Review E92, 012105 (2015)

    2015

  3. [3]

    Ohzeki and A

    M. Ohzeki and A. Ichiki, Mathematical understanding of detailed balance condition violation 29 and its application to langevin dynamics, Journal of Physics: Conference Series638, 012003 (2015)

    2015

  4. [4]

    Rey-Bellet and K

    L. Rey-Bellet and K. Spiliopoulos, Improving the convergence of reversible samplers, Journal of Statistical Physics164, 472 (2016)

    2016

  5. [5]

    A. B. Duncan, T. Lelièvre, and G. A. Pavliotis, Variance reduction using nonreversible langevin samplers, Journal of Statistical Physics163, 457 (2016)

    2016

  6. [6]

    Ichiki and M

    A. Ichiki and M. Ohzeki, Stochastic dynamics without detailed balance condition con- necting simple gradient method and hamiltonian monte carlo, arXiv:2107.08633 (2021), arXiv:2107.08633

    work page arXiv 2021

  7. [7]

    Risken,The Fokker–Planck Equation: Methods of Solution and Applications(Springer, 1984)

    H. Risken,The Fokker–Planck Equation: Methods of Solution and Applications(Springer, 1984)

    1984

  8. [8]

    Witten, Supersymmetry and morse theory, Journal of Differential Geometry17, 661 (1982)

    E. Witten, Supersymmetry and morse theory, Journal of Differential Geometry17, 661 (1982)

    1982

  9. [9]

    Welling and Y

    M. Welling and Y. W. Teh, Bayesian learning via stochastic gradient langevin dynamics, in Proceedings of the 28th International Conference on Machine Learning(2011)

    2011

  10. [10]

    Sato and H

    I. Sato and H. Nakagawa, Approximation analysis of stochastic gradient langevin dynamics by using fokker–planck equation and itô process, inProceedings of the 31st International Confer- ence on Machine Learning, Proceedings of Machine Learning Research, Vol. 32 (2014)

    2014

  11. [11]

    Mandt, M

    S. Mandt, M. D. Hoffman, and D. M. Blei, Stochastic gradient descent as approximate bayesian inference, Journal of Machine Learning Research18, 1 (2017)

    2017

  12. [12]

    Three Factors Influencing Minima in SGD

    S. Jastrzębski, Z. Kenton, D. Arpit, N. Ballas, A. Fischer, Y. Bengio, and A. Storkey, Three factors influencing minima in SGD, arXiv:1711.04623 (2017), arXiv:1711.04623

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  13. [13]

    S. L. Smith and Q. V. Le, A bayesian perspective on generalization and stochastic gradient descent, inInternational Conference on Learning Representations(2018)

    2018

  14. [14]

    J.-F. Bony, D. Le Peutrec, and L. Michel, Eyring–kramers law for fokker–planck type differ- ential operators, arXiv:2201.01660 (2022), arXiv:2201.01660

    work page arXiv 2022

  15. [15]

    Yaida, Fluctuation–dissipation relations for stochastic gradient descent, inInternational Conference on Learning Representations(2019)

    S. Yaida, Fluctuation–dissipation relations for stochastic gradient descent, inInternational Conference on Learning Representations(2019)

    2019

  16. [16]

    Mignacco and P

    F. Mignacco and P. Urbani, The effective noise of stochastic gradient descent, Journal of Statistical Mechanics: Theory and Experiment2022, 083405 (2022)

    2022

  17. [17]

    Z. Zhu, J. Wu, B. Yu, L. Wu, and J. Ma, The anisotropic noise in stochastic gradient descent: Its behaviorof escaping from sharp minima and regularization effects, inProceedings of the 36th 30 International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 97 (2019) pp. 7654–7663

    2019

  18. [18]

    Dai and Y

    X. Dai and Y. Zhu, On large batch training and sharp minima: A fokker–planck perspective, arXiv:2112.00987 (2021), arXiv:2112.00987

    work page arXiv 2021

  19. [19]

    N. Yang, C. Tang, and Y. Tu, Stochastic gradient descent introduces an effective landscape- dependent regularization favoring flat solutions, Physical Review Letters130, 237101 (2023)

    2023

  20. [20]

    Geman and D

    S. Geman and D. Geman, Stochastic relaxation, gibbs distributions, and the bayesian restora- tion of images, IEEE Transactions on Pattern Analysis and Machine Intelligence6, 721 (1984)

    1984

  21. [21]

    Hajek, Cooling schedules for optimal annealing, Mathematics of Operations Research13, 311 (1988)

    B. Hajek, Cooling schedules for optimal annealing, Mathematics of Operations Research13, 311 (1988)

    1988

  22. [22]

    S. L. Smith, P.-J. Kindermans, C. Ying, and Q. V. Le, Don’t decay the learning rate, increase the batch size, inInternational Conference on Learning Representations(2018)

    2018

  23. [23]

    S. L. Smith, E. Elsen, and S. De, On the generalization benefit of noise in stochastic gradient descent, inProceedings of the 37th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 119 (2020) pp. 9058–9067

    2020

  24. [24]

    D. P. Kingma and J. Ba, Adam: A method for stochastic optimization, inInternational Con- ference on Learning Representations(2015)

    2015

  25. [25]

    Girolami and B

    M. Girolami and B. Calderhead, Riemann manifold langevin and hamiltonian monte carlo methods, Journal of the Royal Statistical Society: Series B73, 123 (2011)

    2011

  26. [26]

    Patterson and Y

    S. Patterson and Y. W. Teh, Stochastic gradient riemannian langevin dynamics on the prob- ability simplex, inAdvances in Neural Information Processing Systems, Vol. 26 (2013)

    2013

  27. [27]

    Y.-A. Ma, T. Chen, and E. B. Fox, A complete recipe for stochastic gradient MCMC, in Advances in Neural Information Processing Systems, Vol. 28 (2015)

    2015

  28. [28]

    Ohzeki, Stochastic gradient method with accelerated stochastic dynamics, Journal of Physics: Conference Series699, 012019 (2016)

    M. Ohzeki, Stochastic gradient method with accelerated stochastic dynamics, Journal of Physics: Conference Series699, 012019 (2016)

    2016

  29. [29]

    Loshchilov and F

    I. Loshchilov and F. Hutter, Decoupled weight decay regularization, inInternational Confer- ence on Learning Representations(2019)

    2019

  30. [30]

    Shazeer and M

    N. Shazeer and M. Stern, Adafactor: Adaptive learning rates with sublinear memory cost, in Proceedings of the 35th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 80 (2018)

    2018

  31. [31]

    H. Liu, Z. Li, D. Hall, P. Liang, and T. Ma, Sophia: A scalable stochastic second-order 31 optimizer for language model pre-training, arXiv:2305.14342 (2023), arXiv:2305.14342

    work page Pith review arXiv 2023

  32. [32]

    X. Chen, C. Liang, D. Huang, E. Real, K. Wang, Y. Liu, H. Pham, X. Dong, T. Luong, C.-J. Hsieh, Y. Lu, and Q. V. Le, Symbolic discovery of optimization algorithms, inAdvances in Neural Information Processing Systems, Vol. 36 (2023)

    2023

  33. [33]

    N. S. Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. P. Tang, On large-batch training for deep learning: Generalization gap and sharp minima, inInternational Conference on Learning Representations(2017)

    2017

  34. [34]

    An Empirical Model of Large-Batch Training

    S. McCandlish, J. Kaplan, D. Amodei, and O. D. Team, An empirical model of large-batch training, arXiv:1812.06162 (2018), arXiv:1812.06162

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  35. [35]

    N. S. Keskar and R. Socher, Improving generalization performance by switching from adam to SGD, arXiv:1712.07628 (2017), arXiv:1712.07628

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  36. [36]

    L. Luo, Y. Xiong, Y. Liu, and X. Sun, Adaptive gradient methods with dynamic bound of learning rate, inInternational Conference on Learning Representations(2019)

    2019

  37. [37]

    A. C. Wilson, R. Roelofs, M. Stern, N. Srebro, and B. Recht, The marginal value of adaptive gradient methods in machine learning, inAdvances in Neural Information Processing Systems, Vol. 30 (2017)

    2017

  38. [38]

    S. J. Reddi, S. Kale, and S. Kumar, On the convergence of adam and beyond, inInternational Conference on Learning Representations(2018)

    2018

  39. [39]

    Cohen, S

    J. Cohen, S. Kaur, Y. Li, J. Z. Kolter, and A. Talwalkar, Gradient descent on neural networks typically occurs at the edge of stability, inInternational Conference on Learning Representa- tions(2021)

    2021

  40. [40]

    F.Liao, A.Kolomvaki,andA.Kyrillidis,SGDattheedgeofstability: Thestochasticsharpness gap, arXiv:2604.21016 (2026), arXiv:2604.21016

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  41. [41]

    Foret, A

    P. Foret, A. Kleiner, H. Mobahi, and B. Neyshabur, Sharpness-aware minimization for ef- ficiently improving generalization, inInternational Conference on Learning Representations (2021)

    2021

  42. [42]

    Benamou and Y

    J.-D. Benamou and Y. Brenier, A computational fluid mechanics solution to the monge– kantorovich mass transfer problem, Numerische Mathematik84, 375 (2000)

    2000

  43. [43]

    Jordan, D

    R. Jordan, D. Kinderlehrer, and F. Otto, The variational formulation of the fokker–planck equation, SIAM Journal on Mathematical Analysis29, 1 (1998)

    1998

  44. [44]

    R. S. Sutton, D. McAllester, S. Singh, and Y. Mansour, Policy gradient methods for reinforce- 32 ment learning with function approximation, inAdvances in Neural Information Processing Systems, Vol. 12 (2000)

    2000

  45. [45]

    V. R. Konda and J. N. Tsitsiklis, Actor-critic algorithms, inAdvances in Neural Information Processing Systems, Vol. 12 (2000)

    2000

  46. [46]

    Metropolis, A

    N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, Equation of state calculations by fast computing machines, Journal of Chemical Physics21, 1087 (1953)

    1953

  47. [47]

    W. K. Hastings, Monte carlo sampling methods using markov chains and their applications, Biometrika57, 97 (1970)

    1970

  48. [48]

    Duane, A

    S. Duane, A. D. Kennedy, B. J. Pendleton, and D. Roweth, Hybrid monte carlo, Physics Letters B195, 216 (1987)

    1987

  49. [49]

    M. D. Hoffman and A. Gelman, The no-u-turn sampler: Adaptively setting path lengths in hamiltonian monte carlo, Journal of Machine Learning Research15, 1593 (2014)

    2014

  50. [50]

    Bakry and M

    D. Bakry and M. Émery, Diffusions hypercontractives, inSéminaire de Probabilités XIX 1983/84, Lecture Notes in Mathematics, Vol. 1123 (Springer, 1985) pp. 177–206

    1983

  51. [51]

    J. Song, S. Zhao, and S. Ermon, A-nice-mc: Adversarial training for MCMC, inAdvances in Neural Information Processing Systems, Vol. 30 (2017)

    2017

  52. [52]

    D. Levy, M. D. Hoffman, and J. Sohl-Dickstein, Generalizing hamiltonian monte carlo with neural networks, inInternational Conference on Learning Representations(2018). 33

    2018