Pith. sign in

REVIEW 3 major objections 6 minor 48 references

Divergence-Kernel method for scores of random systems

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper derives the divergence-kernel formula for scores of random systems: for an SDE with multiplicative noise, the terminal score equals the conditional expectation of a covector process driven by the same noise, yielding a pathwise…

desk verdict A genuinely useful formal algorithm with rigorous discrete-time underpinnings, but the headline continuous-time theorem is a formal limit and the numerics are qualitative. read the letter →

arxiv 2507.04035 v1 pith:SWMTEDBT submitted 2025-07-05 math.PR cs.NAmath.DSmath.NAmath.OC

Angxiu Ni This is my paper · ORCID
classification math.PRcs.NAmath.DSmath.NAmath.OC MSC 60H0760J6065D2565C3065C05
keywords ScorefunctionDivergencemethodKernel-differentiationCameron-Martin-GirsanovDiffusionprocessMultiplicativenoiseMonteCarloestimatorRandomdynamicalsystems
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 tries to establish that the score of the terminal density of a random dynamical system can be represented as the conditional expectation of a single covector process evolved along the same random path. The representation mixes the divergence method with the kernel-differentiation method, and it is designed to work where each alone fails: multiplicative noise, non-hyperbolic dynamics, and long time horizons. The discrete-time version is proved rigorously; the continuous-time SDE version is derived by a formal limit that the paper states as an assumption. If the formula holds, scores can be estimated by pathwise Monte Carlo simulation, which would give diffusion models and linear-response computations a tool they currently lack.

What carries the argument

The load-bearing object is the one-step divergence-kernel identity: for $x_1 = f(x_0)+\sigma(x_0)b_0$, the score is expressed as a weighted combination of the kernel derivative $\nabla\log k(b_0)/\sigma(x_0)$ and the divergence term $g^{-1}_{b_0*}(\nabla\log h_0 - \operatorname{div}g_{b_0*})$, with weights controlled by a matrix $\alpha_1$. Recursing this identity yields the forward covector process in Theorem 4.6, which is rigorous in discrete time. The continuous-time limit replaces $g_{b_*}^{-1}$ and $\operatorname{div}g_{b_*}$ with the approximations in Lemmas 5.2 and 5.3 and uses $\nabla\log k(\Delta B) = -\Delta B/\Delta t$ for Gaussian increments; the backward adapted schedule $\alpha$ is what lets the limit keep the damping term while cancelling the otherwise explosive growth.

What would settle it

Simulate many paths of a one-dimensional multiplicative-noise SDE, for example $dx = -x^3\,dt + (0.5+e^{-x^2})\,dB$, with $T=3$ and small $\Delta t$; compute the conditional average $\mathbb{E}[\nu_T \mid x_T]$ from Theorem 1.1 and compare it with a finite-difference or kernel-density estimate of $\nabla\log h_T$ at several endpoints, using enough samples that Monte Carlo error is negligible. A systematic discrepancy that does not vanish as $\Delta t \to 0$ would disprove the central formula.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1.1. For the SDE $$dx_t = F(x_t)\,dt + \$\sigma$(x_t)\,dB_t,$$ let $h_t$ be the density of $x_t$ and let $\nu_t$ be the forward covector process solving $$d\nu = \big((\nabla\$\sigma$\nabla\$\sigma$^T - \nabla F^T - \$\alpha$)\nu - \nabla\operatorname{div}F + \$nabla^{2}$\$\sigma$\nabla\$\sigma$ + \nabla\$\sigma$\,\$\Delta$\$\sigma$\big)dt - \big(\nabla\$\sigma$\,\nu^T + \$nabla^{2}$\$\sigma$ + \$\alpha$\$sigma^{{-1}}$\big)dB$$ with initial condition $\nu_0 = \nabla\log h_0(x_0)$. Then $$\nabla\log h_T(x_T) = \mathbb{E}[\nu_T \mid x_T].$$ Here $\alpha_t$ is any backward adapted process, typically nonnegative, that damps exponential growth of $\nu$ by shifting part of the divergence onto the probability kernel. The same framework gives pure kernel formulas for additive noise, pure divergence formulas for short time, and a schedule that removes $\nabla\log h_0$ entirely, so singular initial distributions can be handled.

Load-bearing premise

The continuous-time formula rests on the assumption that the small-time-step limit and all interchanges of integrations, averages, and limits are legitimate; the paper states this assumption explicitly rather than proving it.

Editorial extensions

If this is right

  • For any finite horizon, the score of an SDE with multiplicative noise can be estimated by simulating the state and covector SDEs together and averaging over paths conditioned on the same endpoint, with no hyperbolicity assumption.
  • Choosing $\alpha$ large enough, for example constant $\alpha > \frac1T\log|D_T|$, prevents the covector $\nu$ from growing exponentially with $T$, so the estimator does not inherit the divergence method's long-time blow-up.
  • The schedule $\alpha_t = 1/t$ removes $\nabla\log h_0$ from the expression entirely, so the formula applies to singular initial distributions where the initial score is not defined.
  • For additive noise the formula reduces to a pure kernel expression, and for short time to a pure divergence expression, unifying the two classical methods.
  • The numerical demonstration on a 40-dimensional chaotic system with multiplicative noise and a singular initial distribution shows the estimator is implementable in high dimensions.

Reading between the lines

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

  • If the formal limit can be made rigorous under mild regularity, the estimator is unbiased for any finite $T$, making it a drop-in score estimator for score-based diffusion models that currently rely on approximations of the initial score.
  • Because $\alpha_t$ is arbitrary and only affects variance, adaptive schedules that increase $\alpha$ in strongly contracting regions and decrease it where noise is small should improve sample efficiency; the paper tests only constant $\alpha$, so this is a direct testable extension.
  • The same divergence-kernel split should transfer to parameter derivatives, since the divergence term already carries derivatives of terminal densities; the author's stated next step is linear response for diffusions.
  • The rigorous discrete-time Theorem 4.6 already gives a valid gradient estimator for discrete-time generative models, so the method does not have to wait for a fully rigorous continuous-time proof to be useful.
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 / 6 minor

Summary. The paper develops a 'divergence-kernel method' for computing the score function ∇ log h_T of random dynamical systems. For discrete-time systems x_{n+1}=f(x_n)+σ(x_n)b_n with i.i.d. noise b_n, it derives kernel, divergence, and mixed formulas expressing the score as a conditional expectation of a recursively defined covector process ν_n (Theorems 4.2, 4.3, 4.6, 4.7). It then formally passes to the continuous-time limit d x_t = F(x_t) dt + σ(x_t) dB_t and states formal SDE versions, including an arbitrary backward-adapted schedule α_t (Theorem 1.1) and a version avoiding ∇ log h_0 (Theorem 5.5). Numerical experiments on a 1D OU process and a 40-dimensional Lorenz 96 system support the formulas qualitatively, and the code is provided.

Significance. If the continuous-time formula were proven, it would give a pathwise unbiased estimator for scores of SDEs with multiplicative noise, with no hyperbolicity assumption, which is directly relevant to score-based generative modeling and linear response theory. The discrete-time results provide a clean conceptual framework that mixes divergence and kernel differentiation methods. The paper is transparent about the formal nature of the continuous-time limit and posts reproducible code; the zero-mean identity check in Section 6.2 is a useful sanity test. However, the central Theorem 1.1 is explicitly formal, and the discrete-time theorems require an injectivity assumption that is not stated in their hypotheses.

major comments (3)
  1. [Section 5 and Theorem 1.1] The continuous-time result is not established. Section 2 disclaims technical assumptions ('We just assume that all integrations, averages, and change of limits are legit'), and the proof of Theorem 5.1 explicitly says 'Neglect the sum of O(ΔBΔt) terms' without providing L^2 estimates or a convergence argument. Moreover, the ν-SDE in Theorem 1.1 contains the term −(∇σν^T + ∇^2σ + ασ^{−1})dB with backward-adapted α; the stochastic integral is anticipative, and Equation (10) gives only a formal limit, not a definition with existence and uniqueness. Consequently Theorem 1.1 is a conjecture, not a theorem, and the main claim of the paper is unsupported. This must be addressed, either by a rigorous proof of the limit or by clearly reframing the paper's contribution as the rigorous discrete-time results plus a conjectured continuous-time limit.
  2. [Section 3.2 and Theorems 4.3, 4.6, 4.7] The derivation of the divergence formulas assumes injectivity of the map g_b(x) = f(x) + σ(x)b. This assumption is mentioned only parenthetically after Equation (3) ('if it is n-to-1, then we need to further sum over its preimage'), but it is not included in the statements of Lemmas 3.2 and 4.4 or Theorems 4.3, 4.6, and 4.7. For the Euler discretization g_b(x) = x + F(x)Δt + σ(x)b used in Section 5, injectivity can fail with positive probability when σ is non-monotone, e.g., σ(x) = 1 + 0.5 sin x and |b| large. Thus the 'rigorous' discrete-time theorems need either an explicit injectivity assumption, a small-Δt condition, or a version with sums over preimages.
  3. [Section 6] The numerical evidence is too weak to verify the exactness of the continuous-time formula. In Section 6.1 the score is compared with the empirical density only visually and averaged over 10 paths per bin, and in Section 6.2 the check is the zero-mean identity E[Φ(x_T)ν_T·v_T] = −1, which tests the integrated score rather than the pointwise conditional score. A more convincing test would use a system with a known closed-form score (e.g., a linear SDE) and report convergence as Δt → 0 with statistical error bars; the small O(ΔBΔt) terms dropped in Section 5 could produce a bias undetectable by the current checks.
minor comments (6)
  1. [Section 2] The notation σ is used both for the diffusion coefficient and for the σ-algebra generated by (x_0, x_1); this is confusing, particularly in Section 3.3.
  2. [Theorem 4.6] The proof of Theorem 4.6 says 'expand everything and check terms match Lemma 4.5' and does not show the expansion; please include the details.
  3. [Theorem 4.7] In Theorem 4.7, α_n = 1/n is undefined at n = 0; the v_0 term vanishes, but this should be stated explicitly.
  4. [Section 5.1, Eq. (9)] Equation (9) treats ΔB as both a vector and a covector in ∇ log k(ΔB) = −ΔB/Δt; please clarify the convention.
  5. [Throughout] There are several typos: 'fomula' in Theorem 1.1, 'independ' in Section 4.1, and an agreement error in the abstract ('formulas which does not involve').
  6. [Figure 1 caption] The caption states that the score is averaged on 10 paths; this is too few to indicate statistical reliability, and the caption should state that the plot is illustrative.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the divergence-kernel formulas are derived from the density integral representation, and self-citations are contextual rather than load-bearing.

full rationale

The derivation chain is self-contained. Section 3 starts from the density integral representation h1(x1) = integral h0(x0)p(x0,x1)dx0, and Lemmas 3.1 and 3.2 are proved by direct differentiation and change of variables. Lemma 3.3 is an explicit convex combination of those two independently proved one-step identities, as shown in Equation (5). The multi-step formulas in Section 4 are obtained by recursive application of the one-step identities; Lemma 4.5 is proved by a telescoping expansion, and Theorem 4.6 is checked term-by-term against Lemma 4.5 rather than assumed. Theorem 4.7 is derived by substituting the explicit schedule alpha_n = 1/n into Lemma 4.5, with the calculation of v_n shown. The continuous-time results are formally passed to the limit from these discrete formulas; the paper explicitly disclaims rigor in Section 2, noting 'We just assume that all integrations, averages, and change of limits are legit.' A formal unproved limit is a rigor gap, not circularity. The arbitrary schedule alpha_t is a tunable parameter that preserves the identity for any choice, not a fitted input masquerading as a prediction. Self-citations, such as [31], [33], and [35], appear as contextual pointers or proof-style references, while the corresponding lemmas are proved in the text. The numerical verification in Section 6.2 uses the analytically known response -1 and checks that the mean of Phi(x_T) nu_T . v_T + 1 is zero, which is an independent benchmark rather than a restatement of the formula. No load-bearing step reduces to its own input or to a self-citation chain.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central formulas rely on an unstated nondegeneracy of the diffusion coefficient, on the formal interchange of limits in the continuous-time passage, and on the existence of smooth densities for conditional expectations. The only hand-selected values are the numerical schedule parameters, which are variance controls rather than fitted constants.

free parameters (2)
  • Schedule alpha_t = 10 (OU example)
    User-selected process in Theorem 1.1. The theorem holds for any backward adapted alpha_t; the value only affects Monte Carlo variance. The numerical example sets alpha = 10 by hand to control covector growth.
  • Schedule beta_t = t/T (OU example)
    User-selected process in the pure-kernel formula Theorem 5.1. beta_t in [0,1] with beta_T = 1 is arbitrary; the example uses beta_t = t/T to remove the initial-score term.
assumptions (3)
  • domain assumption Diffusion coefficient sigma(x) is invertible for all relevant x, so that sigma^{-1} and sigma^{-M} in the kernel formulas are well defined.
    Lemma 3.1 and Theorem 1.1 use 1/sigma and alpha sigma^{-1}; the paper assumes C1 but never states nondegeneracy.
  • ad hoc to paper All integrals, averages, and changes of limits in the passage to continuous time are legitimate.
    Section 2 states this explicitly as an assumption for the formal SDE limit; it is not proven.
  • domain assumption The SDE has smooth densities h_t and the conditional expectation representation is valid, including for singular initial distributions in Theorem 5.5.
    The score formulas condition on x_T and use h_T; regularity of densities is assumed without proof in the formal limit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Divergence-Kernel method for scores of random systems." pith.science (2026). https://pith.science/paper/SWMTEDBT

@misc{pith2026250704035,
  author       = {Pith},
  title        = {Pith review of: Divergence-Kernel method for scores of random systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWMTEDBT}},
  note         = {Machine review of arXiv:2507.04035}
}
read the original abstract

We derive the divergence-kernel formula for the scores of random dynamical systems, then formally pass to the continuous-time limit of SDEs. Our formula works for multiplicative noise systems over any period of time; it does not require hyperbolicity. We also consider several special cases: (1) for additive noise, we give a pure kernel formula; (2) for short-time, we give a pure divergence formula; (3) we give a formula which does not involve scores of the initial distribution. Based on the new formula, we derive a pathwise Monte-Carlo algorithm for scores, and demonstrate it on the 40-dimensional Lorenz 96 system with multiplicative noise.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 39 canonical work pages

  1. [1]

    Antown, D

    F. Antown, D. Dragi ˇcevi´c, and G. Froyland. Optimal linear responses for Markov chains and stochastically perturbed dynamical systems. Journal of Statistical Physics, 170:1051–1087, 3 2018

    work page 2018

  2. [2]

    Bahsoun, M

    W. Bahsoun, M. Ruziboev, and B. Saussol. Linear response for random dynamical systems. Advances in Mathematics, 364:107011, 4 2020

    work page 2020

  3. [3]

    V . Baladi. Linear response, or else. Proceedings of the International Congress of Mathematicians Seoul 2014, pages 525–545, 2014

    work page 2014

  4. [4]

    V . Baladi. The quest for the ultimate anisotropic Banach space.Journal of Statistical Physics, 166:525–557, 2017

    work page 2017

  5. [5]

    J.-M. Bismut. Large Deviations and the Malliavin Calculus, volume 45. Birkhäuser Boston Inc., Progress in Mathematics, 1984

    work page 1984

  6. [6]

    R. H. Cameron and W. T. Martin. Transformations of weiner integrals under translations. The Annals of Mathematics, 45:386, 4 1944

    work page 1944

  7. [7]

    del Sarto, S

    G. del Sarto, S. Galatolo, and S. Jain. Optimal response for stochastic differential equations by local kernel perturbations. arXiv:2502.09300, 2 2025

    arXiv 2025

  8. [8]

    Dolgopyat

    D. Dolgopyat. On differentiability of SRB states for partially hyperbolic systems. Inventiones Mathemati- cae, 155:389–449, 2004

    work page 2004

Show all 48 references
  1. [9]

    Dragi ˇcevi´c, P

    D. Dragi ˇcevi´c, P. Giulietti, and J. Sedro. Quenched linear response for smooth expanding on average cocy- cles. Communications in Mathematical Physics, 399:423–452, 4 2023

    2023

  2. [10]

    Elworthy and X

    K. Elworthy and X. Li. Formulae for the derivatives of heat semigroups. Journal of Functional Analysis, 125:252–286, 10 1994

    1994

  3. [11]

    G. L. Eyink, T. W. N. Haine, and D. J. Lea. Ruelle’s linear response formula, ensemble adjoint schemes and lévy flights. Nonlinearity, 17:1867–1889, 2004

    2004

  4. [12]

    Fang and G

    L. Fang and G. Papadakis. An augmented shadowing algorithm for calculating the sensitivity of time- average quantities of chaotic systems. Journal of Computational Physics, page 114030, 4 2025

    2025

  5. [13]

    Froyland and S

    G. Froyland and S. Galatolo. Optimal linear response for expanding circle maps. https://arxiv.org/abs/2310.19191, 10 2023

    2023 arXiv

  6. [14]

    Froyland and M

    G. Froyland and M. Phalempin. Optimal linear response for Anosov diffeomorphisms.arXiv:2504.16532v1, 4 2025

    2025 arXiv

  7. [15]

    Galatolo and P

    S. Galatolo and P. Giulietti. A linear response for dynamical systems with additive noise. Nonlinearity, 32:2269–2301, 6 2019

    2019

  8. [16]

    Galatolo and A

    S. Galatolo and A. Ni. Optimal response for hyperbolic systems by the fast adjoint response method. arXiv:2501.02395, 1 2025

    2025 arXiv

  9. [17]

    Galatolo and I

    S. Galatolo and I. Nisoli. An elementary approach to rigorous approximation of invariant measures. SIAM Journal on Applied Dynamical Systems, 13:958–985, 2014

    2014

  10. [18]

    Galatolo and M

    S. Galatolo and M. Pollicott. Controlling the statistical properties of expanding maps. Nonlinearity, 30:2737–2751, 7 2017

    2017

  11. [19]

    L. T. Giorgini, A. N. Souza, D. Lippolis, C. Cvitanovi´c, and P. Schmid. Highlights learning dissipation and instability fields from chaotic dynamics learning dissipation and instability fields from chaotic dynamics. Technical report

  12. [20]

    P. W. Glynn. Likelihood ratio gradient estimation for stochastic systems. Communications of the ACM , 33:75–84, 10 1990

    1990

  13. [21]

    Gouëzel and C

    S. Gouëzel and C. Liverani. Compact locally maximal hyperbolic sets for smooth maps: Fine statistical properties. Journal of Differential Geometry, 79:433–477, 2008

    2008

  14. [22]

    Hairer and A

    M. Hairer and A. J. Majda. A simple framework to justify linear response theory. Nonlinearity, 23:909–922, 4 2010

    2010

  15. [23]

    M. Jiang. Differentiating potential functions of SRB measures on hyperbolic attractors. Ergodic Theory and Dynamical Systems, 32:1350–1369, 2012

    2012

  16. [24]

    C.-H. Lai, Y . Takida, N. Murata, T. Uesaka, Y . Mitsufuji, and S. Ermon. Fp-diffusion: Improving score- based diffusion models by enforcing the underlying score fokker-planck equation. 6 2023

    2023

  17. [25]

    E. N. Lorenz. Predictability – a problem partly solved, pages 40–58. Cambridge University Press, 7 2006

    2006

  18. [26]

    Lucarini, F

    V . Lucarini, F. Ragone, and F. Lunkeit. Predicting climate change using response theory: Global averages and spatial patterns. Journal of Statistical Physics, 166:1036–1064, 2017

    2017

  19. [27]

    Malliavin

    P. Malliavin. Stochastic Analysis, volume 313. Springer Berlin Heidelberg, 1997

    1997

  20. [28]

    Mirafzali, U

    E. Mirafzali, U. Gupta, P. Wyrod, F. Proske, D. Venturi, and R. Marinescu. Malliavin calculus for score- based diffusion models. arXiv:2503.16917, 3 2025

    2025

  21. [29]

    A. Ni. Fast linear response algorithm for differentiating chaos. arXiv:2009.00595, pages 1–28, 2020

    2009

  22. [30]

    A. Ni. Fast adjoint algorithm for linear responses of hyperbolic chaos. SIAM Journal on Applied Dynamical Systems, 22:2792–2824, 12 2023. 22

    2023

  23. [31]

    A. Ni. Backpropagation in hyperbolic chaos via adjoint shadowing. Nonlinearity, 37:035009, 3 2024

    2024

  24. [32]

    A. Ni. Ergodic and foliated kernel-differentiation method for linear responses of random systems. arXiv:2410.10138, 10 2024

    2024 arXiv

  25. [33]

    A. Ni. Differentiating unstable diffusion. arXiv:2503.00718, 3 2025

    2025 arXiv

  26. [34]

    Ni and C

    A. Ni and C. Talnikar. Adjoint sensitivity analysis on chaotic dynamical systems by non-intrusive least squares adjoint shadowing (NILSAS). Journal of Computational Physics, 395:690–709, 2019

    2019

  27. [35]

    Ni and Y

    A. Ni and Y . Tong. Recursive divergence formulas for perturbing unstable transfer operators and physical measures. Journal of Statistical Physics, 190:126, 7 2023

    2023

  28. [36]

    Ni and Y

    A. Ni and Y . Tong. Equivariant divergence formula for hyperbolic chaotic flows. Journal of Statistical Physics, 191:118, 9 2024

    2024

  29. [37]

    Ni and Q

    A. Ni and Q. Wang. Sensitivity analysis on chaotic dynamical systems by non-intrusive least squares shad- owing (NILSS). Journal of Computational Physics, 347:56–77, 2017

    2017

  30. [38]

    Pascanu, T

    R. Pascanu, T. Mikolov, and Y . Bengio. On the difficulty of training recurrent neural networks.International conference on machine learning, pages 1310–1318, 2013

    2013

  31. [39]

    Plechá ˇc, G

    P. Plechá ˇc, G. Stoltz, and T. Wang. Martingale product estimators for sensitivity analysis in computational statistical physics. IMA Journal of Numerical Analysis, 43:3430–3477, 11 2023

    2023

  32. [40]

    M. I. Reiman and A. Weiss. Sensitivity analysis for simulations via likelihood ratios. Operations Research, 37:830–844, 10 1989

    1989

  33. [41]

    Ren and F.-Y

    P. Ren and F.-Y . Wang. Bismut formula for lions derivative of distribution dependent sdes and applications. Journal of Differential Equations, 267:4745–4777, 10 2019

    2019

  34. [42]

    R. Y . Rubinstein. Sensitivity analysis and performance extrapolation for computer simulation models. Op- erations Research, 37:72–81, 2 1989

    1989

  35. [43]

    D. Ruelle. Differentiation of SRB states. Commun. Math. Phys, 187:227–241, 1997

    1997

  36. [44]

    Tang and H

    W. Tang and H. Zhao. Score-based diffusion models via stochastic differential equations – a technical tutorial. 2 2024

    2024

  37. [45]

    C. Wormell. Spectral Galerkin methods for transfer operators in uniformly expanding dynamics. Nu- merische Mathematik, 142:421–463, 2019

    2019

  38. [46]

    C. L. Wormell. Non-hyperbolicity at large scales of a high-dimensional chaotic system. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 478, 5 2022

    2022

  39. [47]

    Zhang, J

    H. Zhang, J. Harlim, and X. Li. Estimating linear response statistics using orthogonal polynomials: An RKHS formulation. Foundations of Data Science, 2:443–485, 2020

    2020

  40. [48]

    Zhang and J

    Q. Zhang and J. Duan. Linear response theory for nonlinear stochastic differential equations with α-stable Levy noises. Journal of Statistical Physics, 182:32, 2 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.