Pith. sign in

REVIEW 2 major objections 4 minor 70 references

Stochastic quantization can be reformulated as a finite-horizon optimal control problem, with an exact path-reweighted terminal ensemble that reaches the Gibbs measure at finite fictitious time.

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

2026-08-01 07:27 UTC pith:ALXUDOT6

load-bearing objection A genuinely new control-based finite-time stochastic quantization scheme with clean formal theory, but the numerical validation does not yet support the headline claim of training-independent exactness because continuous-time Girsanov weights are applied to a discrete Euler–Maruyama chain without a discretization-bias analysis. the 2 major comments →

arxiv 2607.21436 v1 pith:ALXUDOT6 submitted 2026-07-23 hep-lat hep-th

Stochastic Quantization as Optimal Control

Lingxiao Wang This is my paper
classification hep-lat hep-th MSC 81T8060H3093E2082B80 PACS 11.15.Ha05.10.Ln02.50.Ey
keywords stochastic quantizationoptimal controlDoob transformGirsanov reweightingfinite-time samplinglattice phi^4 theorypath reweightingLangevin dynamics
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 tries to establish that stochastic quantization—normally a Langevin dynamics that only reaches the target Gibbs measure as fictitious time goes to infinity—can be recast as a finite-time stochastic optimal control problem. A free-theory Ornstein–Uhlenbeck process serves as a reference, the full interaction enters through a reference-corrected terminal cost, and the optimal residual control is a Doob-transform force. If correct, the terminal ensemble equals the Gibbs measure at an explicitly prescribed time, and the path weights that achieve this are exact, so imperfect neural-network training increases estimator variance but never introduces model bias. The construction is demonstrated on multimodal one-dimensional potentials and on two-dimensional lattice phi^4 theory near criticality, where results agree with hybrid Monte Carlo within errors.

Core claim

The paper claims that stochastic quantization can be implemented as a finite-time control problem rather than an equilibration problem. Splitting the action into free and interacting parts, the free theory supplies an Ornstein–Uhlenbeck reference process, and the interacting terminal cost is corrected by the log-density of the reference marginal, giving S_eff = S + log p_T. The optimal residual drift is v* = g^2 grad log h, where h solves the backward Kolmogorov equation with terminal condition exp(-S_eff); this is a Doob transform of the free reference. At the optimum, the Fokker–Planck equation shows the terminal law becomes (up to a prior-dependent weight) exp(-S_eff) p_T = exp(-S), i.e.,

What carries the argument

The central object is the Doob-transform force v* = g^2 grad log h_t, where h_t is the Feynman–Kac expectation h_t(phi) = E_ref[ exp(-S_eff(phi_T)) | phi_t = phi ]. This h_t converts the free-theory reference into the interacting target at finite time. It is paired with the reference-corrected terminal cost S_eff = S + log p_T, where p_T is the closed-form Ornstein–Uhlenbeck marginal; this correction removes the need for p_T to be flat over the support. The exact path weight, log w = -S_eff - log(dP_theta/dP_ref), obtained via Girsanov, is what preserves unbiasedness even with imperfect training.

Load-bearing premise

The numerical exactness claims assume that the continuous-time Girsanov weight remains the correct importance weight for trajectories generated by a 250-step Euler–Maruyama discretization with time step 0.04, and the paper provides no bound or check on this discretization bias.

What would settle it

Compute the exact Radon–Nikodym derivative for the discretized Euler–Maruyama process (e.g., by including the discrete-time Jacobian or using a discrete Girsanov formula) and compare the resulting reweighted expectation of an observable, such as the susceptibility in the phi^4 model, with the weight used in the paper. A statistically significant difference at fixed dt would indicate that the advertised unbiasedness does not hold for the implemented discretization.

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

If this is right

  • If correct, finite-time exactness of stochastic quantization holds for any noise amplitude g and horizon T, without waiting for equilibration.
  • Imperfect training degrades only the effective sample size; estimators remain unbiased because path weights are exact.
  • The noise amplitude sets a practical diffusion-horizon window: it must be large enough to populate separated modes but small enough to preserve peak structure at T.
  • Independent trajectories replace a single thermalizing Markov chain, enabling parallel generation with fixed per-trajectory cost.
  • The broken phase of lattice phi^4 can be reached through a mixture of Gaussian references, each an Ornstein–Uhlenbeck process, without tunneling.

Where Pith is reading between the lines

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

  • A direct testable extension is to replace the continuous-time Girsanov weight with the exact likelihood ratio of the discretized Euler–Maruyama process; if the two weights differ, the reported unbiasedness may only hold in the continuous-time limit.
  • The reference-corrected terminal cost suggests a connection to nonequilibrium work relations: the path weight resembles a Jarzynski-type exponential, so finite-time exactness may be understood as a fluctuation theorem for the controlled dynamics.
  • Because the free backbone relaxes Fourier modes at rates set by the quadratic kernel, the construction offers a scale-dependent reading of fictitious time that could be compared with renormalization-group inspired diffusion samplers.
  • The prior-dependent optimal weight h_0(phi_0) implies that the effective sample size is bounded by the square of the initial expectation of h_0; choosing a prior aligned with h_0 could further improve sampling efficiency.

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 / 4 minor

Summary. The paper proposes 'optimal stochastic quantization' (OSQ): instead of running Parisi–Wu Langevin dynamics to infinite fictitious time, it formulates finite-horizon stochastic optimal control. A free-theory Ornstein–Uhlenbeck process serves as the reference; the interaction is encoded in a reference-corrected terminal cost S_eff = S + log p_T; the optimal residual control is a Doob-transform force proportional to ∇ log h, where h solves the backward Kolmogorov/Feynman–Kac equation. The authors prove that at the continuous-time optimum the reweighted terminal ensemble equals e^{-S}, with the path weight collapsing to a function of the initial condition. A neural network learns the residual control. Numerical demonstrations are given for one-dimensional multimodal potentials and for two-dimensional lattice scalar φ⁴ theory, comparing OSQ reweighted observables with HMC results at L=16,32,64 near the critical point. The central advertised advantage is that path weights are exact, so imperfect training affects only variance, not bias.

Significance. If the exactness claim survives the discretization gap, this is a conceptually attractive and potentially practical reformulation: it replaces equilibration by a finite-horizon control problem and gives an explicit, closed-form importance weight. The End Matter derivations—Girsanov formula, the KL upper bound, the HJB/Doob solution, and the terminal-law identity—are standard and correct in the continuous-time setting. The paper ships open-source code and gives a clear decomposition of the method. The comparisons at L=16 and L=32, especially near κ=0.2705, show agreement with HMC within errors, and the ESS/N values are honestly reported, including the severe degradation at L=64. The framing in terms of control, rather than merely as another learned sampler, is a useful contribution to the ongoing dialogue between stochastic quantization, diffusion models, and normalizing flows. However, the central 'exactness independent of training quality' statement is only established for continuous SDEs; for the actual Euler–Maruyama implementation the exactness is not demonstrated.

major comments (2)
  1. [End Matter, 'Path-measure objective'; Training paragraph] The exactness claim rests on Eq. (12), the continuous-time Girsanov density for the SDEs (2)–(3). The numerical implementation uses 250-step Euler–Maruyama (T=10, Δt=0.04, g=√2) and states that each trajectory weight is constructed from Eq. (12). For the discretized chain the unbiased importance weight is the ratio of the discrete transition densities, and S_eff must contain the discrete-chain reference terminal marginal p_T^h, not the continuous OU closed-form marginal. The paper does not say which of these is implemented and gives no Δt-refinement study or comparison with the exact discrete-chain weights. Any mismatch introduces O(Δt) (or O(√Δt) pathwise) bias. This is load-bearing: the abstract's claim that 'the path weights are exact' is not demonstrated for the numbers in Table I, and the L=64/κ=0.2705 row (ESS/N=0.002, χ=170(37) vs HMC 213(3)) is far too coarse to detect such a bia
  2. [Lattice φ4, 'stabilized reference' and End Matter] In the broken phase the bare S0 backbone is replaced by a 'stabilized reference' — a mass-shifted OU process with a two-center Gaussian-mixture marginal whose centers are drawn by the prior. Exactness at finite T requires that the P_ref in Eq. (12) and the p_T in Eq. (6) are exactly the path measure and terminal marginal of this stabilized reference. The manuscript does not give the Girsanov factor or the S_eff formula for the mixture/stabilized case, and the sentence 'the residual vθ absorbs … the mismatch between S0 and this stabilized reference, while path reweighting remains exact' is only justified if the weights are computed against that same stabilized reference with its exact transition densities. Without these formulas, the broken-phase entries in Table I (e.g., L=32, κ=0.28; L=64, κ=0.2705) cannot be checked, and the claim of exact reweighting in the broken phase is not fully s
minor comments (4)
  1. [Figures 2 and 3] The top panels compare 'Samples' with 'True'; clarify whether these are unweighted OSQ draws or reweighted samples. The unweighted equivalence to the target requires h0 constant on the prior support, a condition not stated for the toy models.
  2. [Eq. (12)] Please define d w̃_t explicitly: for example, d w̃_t = (dϕ_t − (f0 + vθ)dt)/g is a Pθ-Brownian motion. The sign and normalization of the Girsanov weight are easy to misread.
  3. [Table I caption] The caption states 'reweighted by the exact path weights of Eq. (12)'. In light of the major comment, this wording overstates what is implemented for the discrete chain. Also specify the error estimation method (bootstrap/jackknife) used for the ESS/N column.
  4. [End Matter, 'Optimal control and Doob transform'] The line 'the optimal effective sample size is [Eπ h0]²/Eπ[h0²]' assumes h0 is known exactly; in practice h0 is not computed. It would help to say explicitly that this is a theoretical optimum and that the learned vθ only approximates it.

Circularity Check

1 steps flagged

Finite-time exactness is definitional via S_eff = S + log p_ref_T; broken-phase reference is tuned to the observables it later reports, though the Girsanov/Feynman–Kac core is independent.

specific steps
  1. self definitional [Theory, Eqs. (5)-(6) and End Matter, Eq. (16)]
    "Seff[φ] ≡ S[φ] + log p_ref_T(φ), ... At the optimum, e^{−Seff} p_ref_T = e^{−S} ∝ p_target is recovered at finite T. ... q∗_T ∝ e^{−Seff} pT = e^{−S} (16)."

    The terminal identity q*_T ∝ e^{-S_eff} p_ref_T = e^{-S} is obtained by substituting the definition S_eff = S + log p_ref_T. The target e^{-S} enters through the terminal cost, so the advertised finite-time exactness is an algebraic identity of the construction rather than a consequence of the control dynamics. The Doob h-transform theorem supplies q*_T ∝ e^{-S_eff} p_ref_T, but identifying this with the Gibbs measure is exactly the definition of S_eff. This is the standard importance-sampling identity, and the paper is explicit about the role of the terminal cost, so the circularity is real but minor and acknowledged.

full rationale

The paper's derivation chain is largely self-contained and rests on standard machinery: Girsanov reweighting (Eq. 12), the HJB/Cole–Hopf linearization, and Feynman–Kac (Eqs. 14–15). These steps do not reduce to self-citation or to a fitted input. The central definitional point is the terminal cost S_eff = S + log p_ref_T: once this is chosen, the identity e^{-S_eff} p_ref_T = e^{-S} is immediate, so the finite-time exactness of the reweighted ensemble is an importance-sampling construction rather than a first-principles derivation of the Gibbs measure. The paper is transparent about this, and the nontrivial optimal-control/Doob-transform result is independent of that identity. The numerical comparison against HMC uses Ref. [58], a preprint by the same author; this is a self-citation for the benchmark values but is not load-bearing for the derivation and is reproducible by standard hybrid Monte Carlo. A more serious validation caveat, though not a derivation circularity, is that the broken-phase reference is tuned with 2D Ising finite-size scaling to track the physical magnetization and susceptibility, and the near-critical L=64 comparison has ESS/N ~ 0.002, so the reported agreement for those observables is partly encoded in the proposal and is not a strong independent test. Finally, the exactness proofs are for continuous-time SDEs, while the implementation uses a 250-step Euler–Maruyama discretization without a demonstrated discrete-chain weight; this is a numerical approximation gap, not a circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 1 invented entities

Most free parameters are ordinary numerical choices (g, T, dt, network hyperparameters). The main burden is the broken-phase reference: its well separation and chi_ref are not given in the text, are delegated to code, and appear to be chosen so that the reference tracks target observables. The continuous-to-discrete weight assumption is the most fragile item in the ledger.

free parameters (4)
  • noise amplitude g = g=1, 0.5, 0.1 (toys); g=sqrt(2) (lattice)
    Sets the diffusion-horizon window at finite T. Although g is a fictitious-time reparametrization for the equilibrium measure, it changes finite-time path weights and the learned residual; multiple values are explored, but no principle fixes g.
  • horizon T and time step dt = T=1, dt=0.01 (toys); T=10, dt=0.04 (lattice)
    Chosen by hand; no convergence study in T or dt is reported, although finite-time exactness is central to the claim.
  • broken-phase reference parameters (well separation, chi_ref=1/epsilon) = proportionality constants not given in text; full (L, kappa, lambda) map 'in released code'
    The stabilized two-center Gaussian reference is scaled with volume using Ising exponents beta/nu=1/8, gamma/nu=7/4 'so that the reference tracks the physical magnetization and susceptibility'. If the constants are tuned to HMC values, they are fitted to the very observables later compared.
  • Fourier-feature learnable frequencies (sine-Gordon model) = learned during training
    Network capacity hyperparameters that affect the learned residual; minor for the central claim but part of the method.
axioms (7)
  • standard math Girsanov theorem and mutual absolute continuity of P_theta and P_ref
    Eq. (12) and the KL identity (13) require the controlled and reference path measures to be equivalent; unbounded or learned controls can break absolute continuity.
  • standard math Feynman-Kac/HJB-Cole-Hopf for optimal control
    Used to derive the Doob form v* = g^2 grad log h with h_T = e^{-S_eff} (End Matter, Eq. (15)).
  • domain assumption Reference marginal p_ref_T is exactly known
    S_eff = S + log p_ref_T requires the closed-form marginal of the OU (or two-center mixture) reference; any error in p_ref_T breaks the importance weight.
  • ad hoc to paper Bare S0 is a valid positive quadratic backbone (or is replaced by a hand-stabilized reference)
    In the broken phase the bare free theory has K(0)<0, so the paper abandons the free backbone and constructs a stabilized two-center Gaussian mixture whose parameters are not derived from the action.
  • ad hoc to paper Discrete Euler-Maruyama path weight equals the continuous Girsanov weight
    The 'exact' weight uses Eq. (12) for a continuous SDE, but trajectories are generated with 250 steps; no discretization-bias estimate or correction is supplied.
  • domain assumption The learned residual v_theta can represent the optimal control sufficiently well
    The method relies on neural-network approximation of the Doob force; the paper does not quantify representation error, though exact reweighting mitigates bias.
  • domain assumption For unweighted finite-time exactness, h_0 must be constant on the prior support
    Eq. (16) holds only under this condition; otherwise the paper's general statement of exactness relies on the h_0(phi_0) importance weights.
invented entities (1)
  • Stabilized two-center Gaussian reference (broken-phase OSQ reference) no independent evidence
    purpose: Auxiliary process used instead of the unstable free theory to define the SDE and S_eff in the broken phase
    It is a numerical construction, not a physical entity; it has no falsifiable handle outside the algorithm, and its parameters are set by finite-size scaling/tuning.

pith-pipeline@v1.3.0-alltime-deepseek · 11431 in / 22442 out tokens · 211527 ms · 2026-08-01T07:27:46.862399+00:00 · methodology

0 comments
read the original abstract

Stochastic quantization defines a Euclidean quantum field theory as the equilibrium of a fictitious-time Langevin dynamics, which reaches the Gibbs measure asymptotically. We show that this quantization can be formulated as a finite-time stochastic optimal control problem. A tractable reference process, naturally supplied by the free theory when available, provides an Ornstein--Uhlenbeck dynamics, while the full interaction enters as a reference-corrected terminal cost. The optimal control is a Doob-transform force that steers the path-reweighted terminal ensemble to the target at a prescribed time and for a given noise amplitude. A neural network learns the residual control, realizing this optimal stochastic quantization (OSQ). Because the path weights are exact, imperfect training increases the variance of estimators but does not introduce model bias. On multimodal potentials we recover all modes at finite time and find that the noise amplitude sets a practical diffusion-horizon window. In two-dimensional lattice scalar $\phi^4$ theory we recover observables from hybrid Monte Carlo simulations near the critical point. Quantization is thereby formulated as control rather than equilibration.

Figures

Figures reproduced from arXiv: 2607.21436 by Lingxiao Wang.

Figure 1
Figure 1. Figure 1: FIG. 1. Standard stochastic quantization (SQ) versus op [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Finite- [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Marginal Fisher information [PITH_FULL_IMAGE:figures/full_fig_p003_4.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

70 extracted references · 37 linked inside Pith

  1. [1]

    Parisi, Y

    G. Parisi, Y. S. Wu,et al., Sci. sin24, 483 (1981)

    1981

  2. [2]

    P. H. Damgaard and H. Huffel, Phys. Rept.152, 227 (1987)

    1987

  3. [3]

    Namiki, Prog

    M. Namiki, Prog. Theor. Phys. Suppl.111, 1 (1993)

    1993

  4. [4]

    Parisi, Phys

    G. Parisi, Phys. Lett. B131, 393 (1983)

    1983

  5. [5]

    J. R. Klauder, Acta Phys. Austriaca Suppl.25, 251 (1983)

    1983

  6. [6]

    C. E. Berger, L. Rammelm¨ uller, A. C. Loheac, F. Ehmann, J. Braun, and J. E. Drut, Phys. Rept.892, 1 (2021), arXiv:1907.10183 [cond-mat.quant-gas]

    Pith/arXiv arXiv 2021

  7. [7]

    Aarts and D

    G. Aarts and D. Sexty, Lattice field theories with a sign problem, arXiv preprint (2026), arXiv:2604.24290 [hep- lat]

    Pith/arXiv arXiv 2026

  8. [8]

    G. G. Batrouni, G. R. Katz, A. S. Kronfeld, G. P. Lepage, B. Svetitsky, and K. G. Wilson, Phys. Rev. D32, 2736 (1985)

    1985

  9. [9]

    Wolff, Nucl

    U. Wolff, Nucl. Phys. B Proc. Suppl.17, 93 (1990)

    1990

  10. [10]

    L¨ uscher and S

    M. L¨ uscher and S. Schaefer, JHEP07, 036, arXiv:1105.4749 [hep-lat]

    Pith/arXiv arXiv

  11. [11]

    Luscher, Commun

    M. Luscher, Commun. Math. Phys.293, 899 (2010), arXiv:0907.5491 [hep-lat]

    Pith/arXiv arXiv 2010

  12. [12]

    M. S. Albergo, G. Kanwar, and P. E. Shanahan, Phys. Rev. D100, 034515 (2019), arXiv:1904.12072 [hep-lat]

    Pith/arXiv arXiv 2019

  13. [13]

    Kanwar, M

    G. Kanwar, M. S. Albergo, D. Boyda, K. Cranmer, D. C. Hackett, S. Racani` ere, D. J. Rezende, and P. E. Shanahan, Phys. Rev. Lett.125, 121601 (2020), arXiv:2003.06413 [hep-lat]

    Pith/arXiv arXiv 2020

  14. [14]

    Boyda, G

    D. Boyda, G. Kanwar, S. Racani` ere, D. J. Rezende, M. S. Albergo, K. Cranmer, D. C. Hackett, and P. E. Shana- han, Phys. Rev. D103, 074504 (2021), arXiv:2008.05456 [hep-lat]

    Pith/arXiv arXiv 2021

  15. [15]

    Del Debbio, J

    L. Del Debbio, J. M. Rossney, and M. Wilson, Phys. Rev. D104, 094507 (2021), arXiv:2105.12481 [hep-lat]

    Pith/arXiv arXiv 2021

  16. [16]

    Gerdes, P

    M. Gerdes, P. de Haan, C. Rainone, R. Bondesan, and M. C. N. Cheng, SciPost Phys.15, 238 (2023), arXiv:2207.00283 [hep-lat]. 6

    Pith/arXiv arXiv 2023

  17. [17]

    Caselle, E

    M. Caselle, E. Cellini, and A. Nada, JHEP02, 048, arXiv:2307.01107 [hep-lat]

    Pith/arXiv arXiv

  18. [18]

    Bauer, R

    M. Bauer, R. Kapust, J. M. Pawlowski, and F. L. Tem- men, SciPost Phys.19, 077 (2025), arXiv:2412.12842 [hep-lat]

    Pith/arXiv arXiv 2025

  19. [19]

    Bonanno, A

    C. Bonanno, A. Bulgarelli, E. Cellini, A. Nada, D. Pan- falone, D. Vadacchino, and L. Verzichelli, JHEP04, 051, arXiv:2510.25704 [hep-lat]

    Pith/arXiv arXiv

  20. [20]

    H. Wu, J. K¨ ohler, and F. No´ e, inAdvances in Neural Information Processing Systems, Vol. 33 (2020) pp. 5933– 5944, arXiv:2002.06707 [stat.ML]

    Pith/arXiv arXiv 2020

  21. [21]

    Caselle, E

    M. Caselle, E. Cellini, A. Nada, and M. Panero, JHEP 07(7), 015, arxiv:2201.08862 [hep-lat]

    Pith/arXiv arXiv

  22. [22]

    Bulgarelli, E

    A. Bulgarelli, E. Cellini, and A. Nada, Phys. Rev. D111, 074517 (2025), arXiv:2412.00200 [hep-lat]

    Pith/arXiv arXiv 2025

  23. [23]

    L. Wang, G. Aarts, and K. Zhou, JHEP05, 060, arXiv:2309.17082 [hep-lat]

    Pith/arXiv arXiv

  24. [24]

    Fukushima and S

    K. Fukushima and S. Kamata, J. Phys. Soc. Jap.94, 031010 (2025), arXiv:2411.11297 [hep-lat]

    Pith/arXiv arXiv 2025

  25. [25]

    Hirono, A

    Y. Hirono, A. Tanaka, and K. Fukushima, inPro- ceedings of the 41st International Conference on Ma- chine Learning, PMLR, Vol. 235 (2024) pp. 18324–18351, arXiv:2403.11262 [cs.LG]

    Pith/arXiv arXiv 2024

  26. [26]

    Aarts, D

    G. Aarts, D. E. Habibi, L. Wang, and K. Zhou, Mach. Learn. Sci. Tech.6, 025004 (2025), arXiv:2410.21212 [hep-lat]

    Pith/arXiv arXiv 2025

  27. [27]

    Ranner,Sampling lattice field theories with score- based diffusion models, Diploma thesis, Vienna, Tech

    T. Ranner,Sampling lattice field theories with score- based diffusion models, Diploma thesis, Vienna, Tech. U. (2024)

    2024

  28. [28]

    Q. Zhu, G. Aarts, W. Wang, K. Zhou, and L. Wang, JHEP03, 111, arXiv:2502.05504 [hep-lat]

    arXiv

  29. [29]

    Aarts, D

    G. Aarts, D. E. Habibi, L. Wang, and K. Zhou, JHEP 12, 160, arXiv:2510.01328 [hep-lat]

    Pith/arXiv arXiv

  30. [30]

    Aarts, D

    G. Aarts, D. E. Habibi, A. Ipp, D. I. M¨ uller, T. R. Ranner, L. Wang, W. Wang, and Q. Zhu, Generaliz- able Equivariant Diffusion Models for Non-Abelian Lat- tice Gauge Theory (2026), arXiv:2601.19552 [hep-lat]

    arXiv 2026

  31. [31]

    Komijani, M

    J. Komijani, M. K. Marinkovic, and L. Turgut, Diffusion model for SU(N) gauge theories (2026), arXiv:2605.06134 [hep-lat]

    Pith/arXiv arXiv 2026

  32. [32]

    Tomiya, Lattice Configuration Generation with a Self- Learning Diffusion Model (2026), arXiv:2607.12587 [hep- lat]

    A. Tomiya, Lattice Configuration Generation with a Self- Learning Diffusion Model (2026), arXiv:2607.12587 [hep- lat]

    Pith/arXiv arXiv 2026

  33. [33]

    Cranmer, G

    K. Cranmer, G. Kanwar, S. Racani` ere, D. J. Rezende, and P. E. Shanahan, Nature Rev. Phys.5, 526 (2023), arXiv:2309.01156 [hep-lat]

    Pith/arXiv arXiv 2023

  34. [34]

    Aarts, K

    G. Aarts, K. Fukushima, T. Hatsuda, A. Ipp, S. Shi, L. Wang, and K. Zhou, Nature Rev. Phys.7, 154 (2025), arXiv:2501.05580 [hep-lat]

    Pith/arXiv arXiv 2025

  35. [35]

    Cotler and S

    J. Cotler and S. Rezchikov, Phys. Rev. D108, 025003 (2023), arXiv:2202.11737 [hep-th]

    Pith/arXiv arXiv 2023

  36. [36]

    Sheshmani, Y.-Z

    A. Sheshmani, Y.-Z. You, B. Buyukates, A. Ziashahabi, and S. Avestimehr, Phys. Rev. E111, 015304 (2025), arXiv:2402.17090 [cs.LG]

    Pith/arXiv arXiv 2025

  37. [37]

    Cotler and S

    J. Cotler and S. Rezchikov, Renormalizing Diffusion Models (2023), arXiv:2308.12355 [hep-th]

    Pith/arXiv arXiv 2023

  38. [38]

    L´ eonard, Discrete Contin

    C. L´ eonard, Discrete Contin. Dyn. Syst.34, 1533 (2014)

    2014

  39. [39]

    Vargas, W

    F. Vargas, W. Grathwohl, and A. Doucet, inInterna- tional Conference on Learning Representations(2023) arXiv:2302.13834 [stat.ML]

    Pith/arXiv arXiv 2023

  40. [40]

    Berner, L

    J. Berner, L. Richter, and K. Ullrich, Transactions on Machine Learning Research (2024), arXiv:2211.01364 [cs.LG]

    Pith/arXiv arXiv 2024

  41. [41]

    J. L. Doob,Classical Potential Theory and Its Probabilis- tic Counterpart(Springer, New York, 1984)

    1984

  42. [42]

    Zhang and Y

    Q. Zhang and Y. Chen, inInternational Conference on Learning Representations(2022) arXiv:2111.15141 [stat.ML]

    Pith/arXiv arXiv 2022

  43. [43]

    Richter and J

    L. Richter and J. Berner, inInternational Conference on Learning Representations(2024) arXiv:2307.01198 [cs.LG]

    Pith/arXiv arXiv 2024

  44. [44]

    R. M. Neal, Statistics and Computing11(2001)

    2001

  45. [45]

    M. S. Albergo and E. Vanden-Eijnden, inProceedings of the 42nd International Conference on Machine Learning, PMLR, Vol. 267 (2025) arXiv:2410.02711 [cs.LG]

    Pith/arXiv arXiv 2025

  46. [46]

    S. Chen, M. Qian, G. Aarts, B. Lucini, and K. Zhou, Stochastic Path Sampler For Lattice Field Theory, arXiv preprint (2026), arXiv:2606.13790 [hep-lat]

    arXiv 2026

  47. [47]

    Ihssen, R

    F. Ihssen, R. Kapust, and J. M. Pawlowski, Generative sampling with physics-informed kernels, arXiv preprint (2025), arXiv:2510.26678 [hep-lat]

    arXiv 2025

  48. [48]

    Risken,The Fokker-Planck Equation: Methods of So- lution and Applications, 2nd ed., Springer Series in Syn- ergetics, Vol

    H. Risken,The Fokker-Planck Equation: Methods of So- lution and Applications, 2nd ed., Springer Series in Syn- ergetics, Vol. 18 (Springer, Berlin, 1989)

    1989

  49. [49]

    D. J. E. Callaway and A. Rahman, Phys. Rev. Lett.49, 613 (1982)

    1982

  50. [50]

    Polonyi and H

    J. Polonyi and H. W. Wyld, Phys. Rev. Lett.51, 2257 (1983), [Erratum: Phys.Rev.Lett. 52, 401 (1984)]

    1983

  51. [51]

    Aarts, B

    G. Aarts, B. Lucini, and C. Park, Phys. Rev. E111, 015303 (2025), arXiv:2407.16427 [cond-mat.dis-nn]

    Pith/arXiv arXiv 2025

  52. [52]

    Øksendal,Stochastic Differential Equations: An In- troduction with Applications, 6th ed

    B. Øksendal,Stochastic Differential Equations: An In- troduction with Applications, 6th ed. (Springer, Berlin, 2003)

    2003

  53. [53]

    Karatzas and S

    I. Karatzas and S. E. Shreve,Brownian Motion and Stochastic Calculus, 2nd ed. (Springer, New York, 1991)

    1991

  54. [54]

    Rajaraman,Solitons and Instantons: An Introduc- tion to Solitons and Instantons in Quantum Field Theory (North-Holland, Amsterdam, 1982)

    R. Rajaraman,Solitons and Instantons: An Introduc- tion to Solitons and Instantons in Quantum Field Theory (North-Holland, Amsterdam, 1982)

    1982

  55. [55]

    B. W. Silverman,Density Estimation for Statistics and Data Analysis(Chapman and Hall, London, 1986)

    1986

  56. [56]

    L. F. Richardson, Philos. Trans. R. Soc. Lond. A226, 299 (1927)

    1927

  57. [57]

    T. M. Cover and J. A. Thomas,Elements of Information Theory, 2nd ed. (Wiley-Interscience, Hoboken, 2006)

    2006

  58. [58]

    Y.-y. Tan, G. Aarts, D. E. Habibi, B. Lucini, and L. Wang, Diffusion Models for Sampling Near Criticality in Lattice Field Theories (2026), arXiv:2607.08505 [hep- lat]

    Pith/arXiv arXiv 2026

  59. [59]

    Duane, A

    S. Duane, A. D. Kennedy, B. J. Pendleton, and D. Roweth, Phys. Lett. B195, 216 (1987)

    1987

  60. [60]

    Binder, Phys

    K. Binder, Phys. Rev. Lett.47, 693 (1981)

    1981

  61. [61]

    Bohm, Phys

    D. Bohm, Phys. Rev.85, 166 (1952)

    1952

  62. [62]

    F´ enyes, Z

    I. F´ enyes, Z. Phys.132, 81 (1952)

    1952

  63. [63]

    Nelson, Phys

    E. Nelson, Phys. Rev.150, 1079 (1966)

    1966

  64. [64]

    Ihssen, R

    F. Ihssen, R. Kapust, and J. M. Pawlowski, Solving sign problems with physics-informed kernels, arXiv preprint (2026), arXiv:2603.03159 [hep-lat]

    arXiv 2026

  65. [65]

    Masuki and Y

    K. Masuki and Y. Ashida, Generative diffusion model with inverse renormalization group flows, arXiv preprint (2025), arXiv:2501.09064 [cond-mat.stat-mech]

    Pith/arXiv arXiv 2025

  66. [66]

    Carosso, JHEP01, 172, arXiv:1904.13057 [hep-th]

    A. Carosso, JHEP01, 172, arXiv:1904.13057 [hep-th]

    Pith/arXiv arXiv 1904

  67. [67]

    Oh and D

    J.-H. Oh and D. P. Jatkar, JHEP11, 144, arXiv:1209.2242 [hep-th]

    Pith/arXiv arXiv

  68. [68]

    Gattringer and C

    C. Gattringer and C. B. Lang,Quantum chromodynamics on the lattice, Vol. 788 (Springer, Berlin, 2010). 7 End Matter Path-measure objective LetP θ be the path measure of the controlled SDE (3) andP ref that of the free theory referencedϕ t = f0(ϕt)dt+g dw t started fromp 0 =π, with marginals {pt}t∈[0,T] ; we abbreviatep T ≡p ref T . Both processes share th...

    2010

  69. [69]

    and trained with Adam (learning rate 10 −3, cosine-annealed to 10 −5; batch size

  70. [70]

    for 103 epochs, with identical architecture and budget for all (L, κ)