Pith. sign in

REVIEW 2 major objections 4 minor 28 references

Moment Estimates and DeepRitz Methods on Learning Diffusion Systems with Non-gradient Drifts

T0 review · 2 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A two-phase Moment-DeepRitz method learns the full decomposition of a non-gradient drift from density snapshots, covering cases where the gradient and rotational parts are not pointwise orthogonal.

desk verdict A clean, incremental two-phase extension of the authors' prior moment method; the main unresolved issue is the unflagged harmonic gauge ambiguity in the DeepRitz stage, which the reported experiments do not rule out. read the letter →

arxiv 2509.10495 v1 pith:WKP7QBNC submitted 2025-08-31 cs.LG physics.comp-ph

classification cs.LGphysics.comp-ph MSC 60H3565N3068T07
keywords Fokker-Planckequationnon-gradientdriftdecompositionmomentestimatesDeepRitzmethodpseudo-potentialquasi-potentialgradient-divergence-free
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

This paper proposes a data-driven method, the Moment-DeepRitz method, for learning the drift of a stochastic diffusion from snapshots of its probability density. The drift is assumed to split into a gradient part and a divergence-free rotational part, b = -∇ψ + R; the method recovers all three objects—b, ψ, and R—in two phases. Phase 1 uses the exact evolution of first moments to fit the full drift b from pairs of density snapshots; Phase 2 uses a variational (DeepRitz) formulation of the Poisson equation −Δψ = ∇·b to extract the potential part, with R obtained as the remainder. This removes the pointwise orthogonality restriction of an earlier two-stage approach, and the reported two-dimensional experiments cover double- and quadruple-well potentials, a rough oscillatory potential, and oscillatory rotations under both noisy and vanishing-noise conditions.

What carries the argument

The two load-bearing pieces are the first-moment identity d/dt ∫ f x_i dx = ∫ b_i f dx, which converts density snapshots into linear constraints on the drift, and the variational principle I(ψ) = ½ ∫ |∇ψ|² dx + ∫ ∇ψ·b dx, whose minimizer satisfies −Δψ = ∇·b. Phase 1 minimizes the squared error between finite-difference centroid velocities and a neural-network drift averaged against the observed densities; Phase 2 applies the DeepRitz variational method to I with the learned b fixed. The divergence-free rotation is then recovered as R = b + ∇ψ.

What would settle it

Run the same two-phase scheme on a double-well system with the computational box halved to [-2,2]² while keeping identical density data near the origin. If the reconstructed ψ and R change noticeably with box size, especially R near the boundary, then the truncation of the domain, not the underlying dynamics, is controlling the error.

Watch

Extended reading notes

Core claim

The paper's central claim is that for Fokker-Planck dynamics, the first-moment identity d/dt ∫ f x_i dx = ∫ b_i f dx turns paired density snapshots into a regression problem for the drift b. With b in hand, the potential ψ in the decomposition b = −∇ψ + R, ∇·R = 0, is characterized as the minimizer of the convex variational problem I(ψ) = ½ ∫ |∇ψ|² dx + ∫ ∇ψ·b dx, whose Euler–Lagrange equation is −Δψ = ∇·b. The Moment-DeepRitz method implements both steps with neural networks and reports small relative errors for b, ψ, and R in two-dimensional tests. Because the divergence-free condition is enforced through the decomposition itself rather than by requiring pointwise orthogonality of ∇ψ and R

Load-bearing premise

The method assumes that cutting the plane down to the box [-4,4]² and imposing no condition at its edges does not change the potential being learned; the paper does not quantify the boundary error this truncation introduces.

Editorial extensions

If this is right

  • Given two nearby density snapshots for a set of Gaussian initial packets, the full drift b, the potential ψ, and the rotation R can be reconstructed without any trajectory or velocity data.
  • The pointwise orthogonality restriction of the earlier two-stage framework is removed, so the method applies to general gradient-plus-divergence-free drift decompositions.
  • The method tolerates noise added to the densities and remains accurate as the noise intensity goes to zero, where the learned scalar field approaches the quasi-potential.
  • The recovered R carries the non-gradient, circulatory part of the dynamics, giving a route from data to non-equilibrium quantities such as entropy production.

Reading between the lines

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

  • Because Phase 1 uses only centroids, a natural extension is to feed the method particle-based density estimates instead of perfectly known snapshots and to measure how centroid estimation error propagates into b and R.
  • Since R is computed as b + ∇ψ, errors from both phases concentrate in the rotation; checking ∇·R ≈ 0 on the learned field would be a cheap, output-only consistency test the paper does not report.
  • The same two-phase structure should extend to time-dependent drifts by windowing the moment equation and letting both networks depend on time, matching the paper's stated open problem.
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

2 major / 4 minor

Summary. The paper proposes a two-phase data-driven method, Moment-DeepRitz, to learn the drift decomposition b = -∇ψ + R with ∇·R = 0 for generalized diffusion processes. Phase 1 fits the drift b from first-moment dynamics, and Phase 2 obtains ψ by minimizing a DeepRitz-style variational loss given the learned b. The method is applied to four two-dimensional examples (double-well, quadruple-well, oscillatory rotation, rough potential) with reported rRMSE below 0.25 for b, ψ, and R. The authors claim the method extends prior work to non-pointwise orthogonal drifts and is robust to noisy data and rough/oscillatory components.

Significance. The contribution is potentially significant: the idea of combining moment dynamics with a variational principle is natural and the test problems are well chosen. If the method indeed recovers the non-unique decomposition b = -∇ψ + R in general settings, it would be a useful addition to the data-driven SDE/PDE learning literature. The paper also demonstrates the approach on challenging cases (rough potentials, oscillatory rotations) that go beyond the pointwise-orthogonal setting of [15]. However, the validation is undermined by a boundary-condition/gauge issue in the DeepRitz stage and by the absence of baselines, error bars, and out-of-sample checks. The central claim is therefore not yet established, but the issues appear fixable within the manuscript's scope.

major comments (2)
  1. [Section 2, Eq. (6)–(9)] The variational problem (6) is minimized over H without any boundary condition. Its first variation forces the natural Neumann condition (∇ψ+b)·n = 0 on ∂Ω. For the ground-truth examples in Section 3, R = b + ∇ψ has non-zero normal components on the box boundary (e.g., R = (y, -x) gives R·n = ±y on x = ±4). Consequently, the minimizer of (6) is not the stated ψ but ψ + h for a harmonic h selected by this boundary condition; equivalently the decomposition b = -∇ψ + R is not unique up to an additive constant, as the paper claims, but up to an arbitrary harmonic gauge on the bounded domain. The reported rRMSE for ψ and R in Table 1 therefore do not directly validate recovery of the intended decomposition. The paper should either add a boundary term to the loss that enforces the correct boundary condition, evaluate the error only in the interior where the gauge drift is small, or explicitly
  2. [Section 3, Table 1 and paragraph on noisy data] The abstract and Section 3 claim robustness to noisy data, but the only observation-noise experiment uses a single perturbation level (Gaussian noise with standard deviation 0.1, convolved with training snapshots). The three rows of Table 1a vary the SDE noise intensity σ², not the observation noise, so they do not support the 'robust to noisy data' claim. In addition, no error bars or multiple random seeds are reported; the reader cannot assess whether the rRMSE values in Table 1 are representative or the result of a favorable initialization. Since robustness is advertised as a key advantage, the experimental setup needs to be expanded (e.g., varying noise level, several seeds, perhaps with mean ± std) before the claim is supported.
minor comments (4)
  1. [Section 2, Eq. (7)–(9)] The rRMSE definitions do not state whether the error is computed on the full grid [-4,4]² or only on the plotting window (e.g., [-1.5,1.5]²). This matters precisely because of the boundary-condition ambiguity discussed above. Please specify the evaluation grid.
  2. [Section 3, Figures 1 and 2] The figures are small and difficult to read, especially the quiver and difference plots. Please use larger panels and consistent color scales so the reader can compare learned vs. ground truth across rows.
  3. [Section 4] The claim that 'in the vanishing noise limit (σ → 0), the pseudo-potential converges to the quasi-potential' is not substantiated by any experiment or reference to a computed quasi-potential. It is merely a qualitative remark and should be reworded or supported.
  4. [Introduction] There is a typo: 'Kong at al.' should be 'Kong et al.' Also, the relationship to [15] and [19] is only stated qualitatively; a short comparison table on the same examples would help the reader judge the claimed advantage.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; the derivation is self-contained apart from a minor non-load-bearing self-citation.

full rationale

The Moment-DeepRitz derivation is self-contained and does not reduce to its own inputs. Phase 1 uses the exact first-moment identity (3), restated explicitly in the loss (8), so the fit of b is not a relabeled prediction. Phase 2 minimizes the variational form (6), whose Euler-Lagrange equation is the Poisson equation (5) derived from the Helmholtz-type decomposition (4); the ground-truth ψ and R are not used in either loss. The only self-citation is [15], which is noted as providing an identical Stage 1, but because the loss is restated in the present paper and follows from standard moment dynamics, this citation is not load-bearing. A separate correctness caveat, distinct from circularity, is that the truncated-domain variational problem (6) has no boundary term, so its minimizer satisfies the natural condition (∇ψ+b)·n=0 on ∂Ω, which the reported true solutions do not satisfy; this may bias ψ and R near the boundary. This is a model misspecification concern, not a circular reduction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method contains no fitted physical constants. Its free parameters are training-protocol choices (initial variance, snapshot times, number of initial conditions, noise level). The derivation itself is parameter-free, relying on standard Helmholtz decomposition, the moment equation, and the DeepRitz variational principle. The only questionable input is the unstated boundary treatment on the truncated domain.

free parameters (4)
  • initial variance sigma0^2 = 0.01
    Chosen to make initial Gaussians localized; controls the sampling of drift values in Phase 1.
  • observation times t1, t2 = t1=0.015, t2=0.016
    Finite-difference approximation of the moment derivative in (8); small window avoids nonlinearity but increases discretization error.
  • number of initial conditions M = 40 (80 for QW and Osc. R)
    Density of coverage of the state space; directly affects identifiability of b in Phase 1.
  • noise perturbation level = 0.1
    The single noise level used to claim robustness; not swept.
assumptions (6)
  • standard math Any smooth drift b admits a unique decomposition b = -∇ψ + R with ∇·R = 0 (Helmholtz decomposition)
    Invoked in Section 2, eqs. (4)-(5), to define the pseudo-potential and rotation as targets of the learning method.
  • standard math First-moment evolution d/dt ∫ f x_i dx = ∫ b_i f dx for the Fokker-Planck density
    Eq. (3) in Section 2; it is the basis of the Phase 1 loss (8).
  • standard math Minimizing I(ψ) in (6) is equivalent to solving -Δψ = ∇·b
    Section 2, eq. (6); calculus of variations with first variation ∫∇δψ·(∇ψ+b)=0.
  • ad hoc to paper Truncating the domain to [-4,4]^2 and minimizing (6) without boundary terms approximates the R^2 problem
    Section 3 truncates the FP problem to [-4,4]^2; the implicit natural Neumann condition (∇ψ+b)·n=0 is not satisfied by the true potential and is never discussed.
  • domain assumption The numerical FP solver produces density snapshots accurate enough to serve as training data
    Section 3 generates training data by solving (2) with Δx=0.1, Δt=1e-4; solver error is neither reported nor propagated.
  • domain assumption Two-layer Tanh networks with 50 hidden units can represent b and ψ at the required accuracy
    Section 3 chooses the architecture without approximation-error analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Moment Estimates and DeepRitz Methods on Learning Diffusion Systems with Non-gradient Drifts." pith.science (2026). https://pith.science/paper/WKP7QBNC

@misc{pith2026250910495,
  author       = {Pith},
  title        = {Pith review of: Moment Estimates and DeepRitz Methods on Learning Diffusion Systems with Non-gradient Drifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WKP7QBNC}},
  note         = {Machine review of arXiv:2509.10495}
}
read the original abstract

Conservative-dissipative dynamics are ubiquitous across a variety of complex open systems. We propose a data-driven two-phase method, the Moment-DeepRitz Method, for learning drift decompositions in generalized diffusion systems involving conservative-dissipative dynamics. The method is robust to noisy data, adaptable to rough potentials and oscillatory rotations. We demonstrate its effectiveness through several numerical experiments.

Figures

Figures reproduced from arXiv: 2509.10495 by the authors.

Figure 1
Figure 1. From left to right: bNN/b (quiver plots), ψNN/ψ/ψNN − ψ (heatmaps), and RNN/R (quiver plots). Rows: (1) Quadruple-well potential, (2) Oscillatory rotation. Rough Potential. To demonstrate the robustness of our method to rough, oscillatory potentials, we consider ψ(x) = 1 4 (x 2 − 1)2 + 1 2 y 2 + ε 4 sin( 2πx ε ) sin( 2πy ε ) with a parameter ε = 0.2. Here ε controls the oscillations: smaller values produce stronger,… view at source ↗
Figure 2
Figure 2. From left to right: bNN/b (quiver plots), ψNN/ψ/ψNN − ψ (heatmaps), and RNN/R (quiver plots). Rows: (1) σ 2 = 2, (2) σ 2 = 0.001, (3) σ 2 = 0, (4) σ 2 = 2 with the rough potential. 6 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 26 canonical work pages

  1. [15]

    Moment Estimate and Variational Approach for Learning Generalized Diffusion with Non-gradient Structures

    Kong, F., Lai, C.-C. and Lu, Y . (2025) Moment Estimate and Variational Approach for Learning Generalized Diffusion with Non-gradient Structures. arXiv preprint arXiv:2508.01854

  2. [1]

    Chen, X., Yang, L., Duan, J., Karniadakis, G. Em. (2021) Solving Inverse Stochastic Problems from Discrete Particle Observations Using the Fokker–Planck Equation and Physics-Informed Neural Networks. SIAM J. Sci. Comput. 43(3):B811-B830

  3. [2]

    and Kutz, J

    Brunton, S., Proctor, J. and Kutz, J. (2016) Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proc. Natl. Acad. Sci. 113(15):3932–3937

  4. [3]

    (1999) Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences

    Crooks, G. (1999) Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences. Phys. Rev. E 60(3):2721

  5. [4]

    (1999) An introduction to chaos in nonequilibrium statistical mechanics

    Dorfman, J. (1999) An introduction to chaos in nonequilibrium statistical mechanics. No. 14, Cambridge university press

  6. [5]

    and Yu, B

    E, W. and Yu, B. (2018) The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems. Commun. Math. Stat. 6(1):1-12

  7. [6]

    and Van den Broeck, C

    Esposito, M. and Van den Broeck, C. (2010) Three detailed fluctuation theorems. Phys. Rev. Lett. 104(9):090601

  8. [7]

    (2012) Stochastic thermodynamics under coarse graining

    Esposito, M. (2012) Stochastic thermodynamics under coarse graining. Phys. Rev. E. 84(4):041125

Show all 28 references
  1. [8]

    (2012) An introduction to stochastic differential equations

    Evans, L. (2012) An introduction to stochastic differential equations. V ol. 82, American Mathematical Society

  2. [9]

    Gao, Lang, Q

    Y . Gao, Lang, Q. and Lu, F. (2024) Self-test loss functions for learning weak-form operators and gradient flows. arXiv preprint arXiv:2412.03506

  3. [10]

    and Sasa, S.-I

    Hatano, T. and Sasa, S.-I. (2001) Steady-state thermodynamics of Langevin systems. Phys. Rev. Lett. 86(16):3463

  4. [11]

    and Reina, C

    Huang, S., He, Z. and Reina, C. (2022) Variational Onsager Neural Networks (VONNs): A thermodynamics-based variational learning strategy for non-equilibrium PDEs. J. Mech. Phys. Solids 163:104856

  5. [12]

    and Reina, C

    Huang, S., He, Z., Dirr N., Zimmer, J. and Reina, C. (2024) Statistical-Physics-Informed Neural Networks (Stat-PINNs): A machine learning strategy for coarse-graining dissipative dynamics. J. Mech. Phys. Solids:105908

  6. [13]

    and Jiang, D.-H

    Jiang, D.-Q. and Jiang, D.-H. (2004) Mathematical theory of nonequilibrium steady states: on the frontier of probability and dynamical systems. Springer Science & Business Media

  7. [14]

    (1998) Langevin equation and thermodynamics

    Ken, S. (1998) Langevin equation and thermodynamics. Prog. Theor . Phys. Suppl.130:17–27

  8. [16]

    and Karniadakis, G

    Kharazmi, E., Zhang, Z. and Karniadakis, G. Em. (2021) hp-VPINNs: Variational physics-informed neural networks with domain decomposition. Comput. Methods Appl. Mech. Eng. 374:113547

  9. [17]

    (1998) Fluctuation theorem for stochastic dynamics

    Kurchan, J. (1998) Fluctuation theorem for stochastic dynamics. J. Phys. A: Math. Gen. 31(16):3719

  10. [18]

    and Trask, N

    Lee, K. and Trask, N. and Stinis, P. (2021) Machine learning structure preserving brackets for forecasting irreversible processes. Advances in Neural Information Processing Systems

  11. [19]

    and Ren W

    Lin, B., Li, Q. and Ren W. (2022) A data driven method for computing quasipotentials. Math. Sci. Mach. Learn.:652–670

  12. [20]

    and Wang Y

    Lu, Y ., Li, X., Liu, C., Tang, Q. and Wang Y . (2024) Learning Generalized Diffusions using an Energetic Variational Approach. arXiv preprint arXiv:2412.04480

  13. [21]

    and Vulpiani, A

    Marconi, U., Puglisi, A., Rondoni, L. and Vulpiani, A. (2008) Fluctuation–dissipation: response theory in statistical physics. Phys. Rep. 461(4–6):111–195

  14. [22]

    Messenger, D.A

    D.A. Messenger, D.A. and D.M. Bortz, D.M. (2021) Weak SINDy for partial differential equations. J. Comput. Phys. 443:110525. 5

  15. [23]

    and Gong, G

    Qian, M.-P., Qian, M. and Gong, G. (1991) The reversibility and the entropy production of Markov processes. Contemp. Math 118:255–261

  16. [24]

    and Wang, Z

    Qian, M. and Wang, Z. (1999) The entropy production of diffusion processes on manifolds and its circulation decompositions. Commun. Math. Phys. 206:429–445

  17. [25]

    and Probert, M

    Quigley, D. and Probert, M. (2004) Langevin dynamics in constant pressure extended systems. J. Chem. Phys. 120(24):11432–11441

  18. [26]

    and Karniadakis, G

    Raissi, M., Perdikaris, P. and Karniadakis, G. Em. (2019) Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378 :686–707

  19. [27]

    and Solin, A

    Särkkä, S. and Solin, A. (2019) Applied stochastic differential equations. V ol. 10, Cambridge University Press

  20. [28]

    best” parameters of the neural networks to reconstruct bNN for learning the drift b. • Phase 2: Minimize the loss function I dyn ψ given by (9) and find the “best

    Zang, Y ., Bao, G., Ye, X. and Zhou, H. (2020) Weak adversarial networks for high-dimensional partial differential equations. J. Comput. Phys. 411:109409. A Algorithm for the Moment-DeepRitz Method Here we present the algorithm for the Moment-DeepRitz method described in Secti...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.