Pith. sign in

REVIEW 4 major objections 6 minor 37 references

This paper claims that the continuous martingale condition for linear, semilinear, and fully nonlinear PDEs can be reformulated as a sequence of least-squares regression problems, so that the solution is found by iteratively fitting a neura

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

2026-08-03 06:27 UTC pith:G7LBBBIP

load-bearing objection Useful iterative sampling–regression PDE solver with strong numerics; convergence theory covers only the exact-conditional-expectation iteration, and the empirical error check is diagonal-slice only — worth serious review, but the gap and missing artifacts should be addressed. the 4 major comments →

arxiv 2607.29470 v1 pith:G7LBBBIP submitted 2026-07-31 math.NA cs.NA

iSMART: An Iterative Sampling-and-Regression Technique for Solving Martingale-Based PDEs

Tiejun Li , Xiaoguang Li , Fugui Ma This is my paper
classification math.NA cs.NA MSC 65M7565C3068T07
keywords martingale-based PDEsiterative sampling and regressionstop-gradient learningleast-squares regressionHamilton-Jacobi-Bellman equationshigh-dimensional numerical PDEsdeep neural networks
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 sets out to show that a large class of high-dimensional partial differential equations with a martingale structure — linear, semilinear, and fully nonlinear — can be solved by a simple iterative loop: simulate stochastic paths under the frozen current solution estimate, build pathwise target values, and fit a neural network to those targets by ordinary squared-error regression. The key mathematical fact is that a conditional expectation is exactly the L^2 minimiser over functions of the conditioning variable, so the expensive conditional expectation inside the martingale representation can be replaced by a regression against raw samples. The paper proves that in the idealised setting where each regression is exact, the iteration is a contraction in a weighted Sobolev space and its fixed point is the PDE solution. On numerical benchmarks in 200 to 1000 dimensions it reports relative L2 errors between roughly 1e-3 and 3e-2 while avoiding adversarial training and nested Monte Carlo estimation.

Core claim

The central claim is that the PDE solution u is the fixed point of the iteration u_{n+1}(t,x)=E^{t,x}[ u_n(t+s,X_{t+s}) - ∫_t^{t+s} f(τ,X_τ,u_n,∇u_n) dτ ], and that this fixed point can be reached not by computing the expectation but by least-squares regression of a neural network u_θ(t,x) against single-path targets with the stop-gradient operator. The theorem states that if the terminal condition u_n(T,x)=g(x) holds and the regression subproblems are solved exactly, the iterates converge to the unique classical solution. The numerical evidence shows the procedure is stable and accurate for reaction-diffusion, Burgers-type, and Hamilton-Jacobi-Bellman equations in high dimensions, including

What carries the argument

Proposition 3.1 (the L^2-projection property of conditional expectation): for a square-integrable random variable ξ, the minimiser over F_t-measurable functions of E[|φ−ξ|^2] is exactly E[ξ|F_t]. This identity lets the algorithm replace conditional expectations by regression losses. The algorithm iterates (3.7)/(3.9) with target ξ^{m,i}=u_θ(t_i+s_i,X^{(m,i)}_{t_i+s_i}) - ∫ f(τ,·) dτ and loss L(θ)=1/(NM) Σ|u_θ(t_i,x_i)−Sg[ξ^{m,i}]|^2. Method (III), freezing-and-compensating, freezes the gradient-dependent drift and moves part of the nonlinearity into the source term to stabilise HJB iterations.

Load-bearing premise

The algorithm's convergence in practice depends on the neural network solving each least-squares regression well enough that the error stays small and does not accumulate; the proven theorem assumes exact conditional expectations and exact terminal matching, not the finite-network, minibatch optimisation actually used.

What would settle it

On a benchmark with a known exact solution, compute a highly accurate Monte Carlo estimate of the conditional expectation at a handful of points after each iSMART iteration and compare it with the network's regression fit; if the gap does not shrink toward zero or if it grows over iterations, the contractive fixed-point argument is not being realised. Alternatively, deliberately break the terminal ansatz by using a network that violates u(T,x)=g(x) and check whether convergence to the PDE solution still occurs.

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

If this is right

  • A single algorithmic template now covers linear, semilinear, and fully nonlinear martingale-based PDEs, eliminating the need to switch between Feynman-Kac, BSDE, or adversarial solvers.
  • Because the regression target is detached via stop-gradient, the training loop is standard least squares, so the method inherits the stability and parallelisation properties of ordinary deep learning optimizers.
  • For equations whose drift and source do not depend on ∇u, the method is derivative-free: path generation needs only DNN forward evaluations, making it cheaper than gradient-based alternatives.
  • If the contraction theorem holds in the idealised setting, the only remaining algorithmic question is the accuracy of each regression fit; this identifies a clear target for future capacity and sampling analysis.
  • The freezing-and-compensating strategy for HJB equations suggests that other problems with quadratic gradient nonlinearities can be handled by the same drift/source split.

Where Pith is reading between the lines

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

  • A natural testable extension: use the same L^2-projection trick to replace conditional expectations in other contexts, such as backward SDEs or filtering, where nested simulation is the bottleneck.
  • The paper's own remark that deep-martingale solvers produce wavy errors where trajectories are sparse suggests adaptive sampling of initial points (t,x) could further reduce error without changing the regression core.
  • Because the convergence theorem assumes exact terminal matching, applying iSMART to problems with boundary conditions rather than terminal conditions would require an additional boundary treatment; this is a limitation, not addressed here.
  • The reported GPU time advantage over prior martingale methods may be partly due to the fixed terminal ansatz (5.1), which is a modelling choice as much as an algorithmic one.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes iSMART, an iterative sampling-and-regression method for high-dimensional PDEs admitting a martingale representation. Starting from the conditional-expectation form (2.2) of the martingale property, the method uses the classical L2-projection characterization of conditional expectation (Proposition 3.1) to replace nested-expectation evaluation by regression against single-path targets. Given a current approximation u_n, the next approximation u_{n+1} is obtained by least-squares fitting a neural network u_theta to the stop-gradient targets u_n(t+s, X_{t+s}) - ∫ f(..., u_n, ∇u_n), with trajectories generated by one of three sampling strategies (Methods I, II, III), the last using a freezing-and-compensating drift for HJB equations. Section 4 proves a contraction result for the idealized fixed-point iteration (3.7)/(3.9) in a weighted sup-norm space. Numerical experiments on linear reaction-diffusion, semilinear Burgers-type, and fully nonlinear HJB equations in dimensions up to 1000 report relative L2 errors of order 1e-3 to 3e-2 and favorable comparison with DeepMartNet.

Significance. If the numerical claims are taken at face value, iSMART is a simple, derivative-free-in-some-regimes, and scalable alternative to nested-expectation and adversarial formulations, and the freezing-and-compensating trick for HJB equations is a useful contribution. The paper is also transparent about the fact that the convergence theorem covers only the exact conditional-expectation iteration, not the implemented minibatch/DNN regression, which is a significant gap. The benchmarks are systematic across PDE classes and dimensions and include comparisons with DeepMartNet, which is a strength. However, the reported high-dimensional errors are computed on one-dimensional diagonal slices, so the full-domain accuracy claim is not empirically established.

major comments (4)
  1. [Section 4 vs. Algorithm 1] The convergence result Theorem 4.1 analyzes the exact iteration (3.7)/(3.9) in the weighted sup-norm space X_β (4.2), and the section explicitly disclaims the minibatch optimization step. The implemented Algorithm 1 (lines 8-11) performs only K=1 gradient update on the empirical L2 loss (3.16) with M=4096 trajectories and a finite-capacity DNN. No bound is given for the projection/regression error of the data-dependent map actually iterated, nor for its accumulation over outer iterations. Since the loss minimized is empirical L2 while the contraction metric is a sup-norm, small empirical loss does not control the theorem's norm. This is the central gap between theory and practice.
  2. [Section 5, Eq. (5.2), Figs. 2-9] For the high-dimensional tests, the reported relative L2 errors are evaluated on the diagonal slice {s 1_d}, as stated in Section 5.1 ('To visualize the high-dimensional results, we compare ... along a one dimensional curve') and used in Figures 2, 6, 8, and 9. The errors in Tables 1-4 for d=200/1000 therefore do not test accuracy over the full spatial domain; off-diagonal errors are invisible. To support the claimed high-dimensional accuracy, full-dimensional test errors on samples from the training distribution (and preferably multiple seeds) should be reported.
  3. [Section 3.3, Method (III), Eqs. (3.25)-(3.26)] The freezing-and-compensating strategy for HJB equations is specified at the PDE level (3.25) and the SDE level (3.26), but the actual regression target minimized in Algorithm 1 is not written. The generic pathwise target (3.15) contains only f(τ, X, u, ∇u), whereas the consistency of (3.25) requires a modified pseudo-source including the compensation term α_n φ_n^T ∇u_n (and the drift compensation). Since the HJB experiments are a main contribution, this missing explicit loss makes the implementation non-reproducible and prevents verification of the claimed mathematical equivalence.
  4. [Theorem 4.1 proof, Eqs. (4.4)-(4.8)] The proof of Theorem 4.1 uses two different iterations: Case 1 (t+h ≥ T) uses the terminal representation with g(X_T), corresponding to lookahead s = T−t, while Case 2 uses lookahead s = h. The iteration (3.7) is written with a fixed lookahead s. The proof therefore does not analyze a single map for all t unless the lookahead is adaptive, s_t = min(h, T−t). The statement 'for any h>0' needs clarification or the proof needs to be adjusted to a consistently defined iteration.
minor comments (6)
  1. [Eq. (3.24) and following text] For the HJB source term f = (δ/2)|∇u|^2, the integral identity gives φ = (δ/2)∇u, not φ = (1/2)∇u; the displayed formula omits the factor δ.
  2. [Section 5.2.2] The heading contains a typo: 'High-dimensitional' should be 'High-dimensional'.
  3. [Figures 3-5] Several axis labels read 't=1:0'; these should be 't=1.0'. The same issue appears in some other figure captions.
  4. [Eq. (5.2)] The definition of the error set D is given generically, but for high-dimensional results it is never stated that D is the one-dimensional diagonal. This should be made explicit in the text and table captions.
  5. [Algorithm 1] The algorithm lists max outer iterations ITER_max but no stopping criterion or convergence check is described; the text only says to iterate until convergence. A precise stopping rule would improve reproducibility.
  6. [Code availability] The code is 'available upon request' rather than publicly deposited. Given the algorithmic complexity (three path-generation methods, freezing compensation), public code would substantially aid reproducibility.

Circularity Check

0 steps flagged

No significant circularity: iSMART's regression-based iteration is derived from the martingale representation; the main limitations are acknowledged idealization gaps, not definitional circularity.

full rationale

No circular step can be exhibited from the paper's equations. The derivation chain is: PDE (1.1) -> Itô formula -> martingale identity (2.2) -> fixed-point iteration (3.7)/(3.9) -> L2 projection Proposition 3.1 -> regression loss (3.16). Each arrow is derived in the text; in particular, Proposition 3.1 is proved in the paper and the citation [11] is only a pointer to a standard fact, not a load-bearing self-citation. The regression target (3.15) uses the previous iterate under stop-gradient, so the fitted minimum is not equal to the target by construction; in the ideal limit it equals the conditional expectation, and the paper's convergence theorem covers that ideal iteration while explicitly disclaiming the finite-sample optimization step. The terminal ansatz (5.1) enforces the known boundary condition g, hence zero error at T is input matching, not a prediction. Numerical validation uses exact or Monte Carlo reference solutions. The acknowledged idealization gap (exact regression vs. K=1 empirical gradient steps, noted in Section 4) is a missing error-propagation analysis, not a circular definition. Self-citations ([11], [32], [33]) are background and are not used to justify the central equivalence or to rule out alternatives.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

All substantive content is derived from classical martingale theory and L2 projection; the paper introduces no new physical entities. The hand-tuned hyperparameters listed above are training and algorithm choices; the main unfunded assumption is that finite-network regression tracks the exact conditional-expectation iteration.

free parameters (5)
  • Terminal-matching decay gamma = 5.0
    Appears in ansatz (5.1); chosen from [3,6] and set to 5.0; controls boundary transition and can affect accuracy near the terminal time.
  • Freezing-compensation coefficient alpha_n = not specified (constant in (0,1))
    Method III, Eq (3.26), shifts gradient nonlinearity into drift; the value is not reported in experiments, yet affects stability and convergence of the iteration.
  • Spatial sampling mixture weights and radii = [0.01,0.05,0.1,0.5,1.5,2.5] for sigma_k
    Eq (3.19) multi-layer Gaussian shells plus uniform boundary; chosen by hand to balance coverage in high dimensions.
  • Lookahead time s and quadrature intervals = s* = min(s0,T-t), N_Int up to 20
    Method I uses fixed lookahead and composite trapezoidal steps; choices affect bias and variance of the regression target.
  • Network and training hyperparameters = N=20, M=4096, K=1, lr=1e-3, hidden widths 128-410, SiLU
    Listed in Section 5; chosen empirically, with no sensitivity study reported.
axioms (7)
  • standard math Ito formula and martingale representation: a smooth solution u of (1.1) yields the compensated process (1.5) as a zero-mean martingale.
    Used to set up iteration (3.7) and all losses; standard under the stated integrability conditions.
  • standard math Classical L2-projection property of conditional expectation (Proposition 3.1).
    Core equivalence replacing conditional expectation with a regression loss; proof is given in the paper.
  • domain assumption Assumptions A1-A4: bounded Lipschitz drift, uniformly elliptic Lipschitz diffusion, Lipschitz source, and Lipschitz terminal data g.
    Needed for Theorem 4.1 and for the parabolic estimates in Lemma A.1; restricts the problem class.
  • domain assumption Unique smooth solution of (1.1) exists (cf. ref [3]).
    Invoked in Section 1 and relied on by the convergence/identification argument.
  • ad hoc to paper DNN capacity and optimization: the least-squares minimizer of (3.16) approximates the conditional expectation sufficiently well at each iteration.
    Algorithm 1 substitutes the regression-trained network for the exact conditional expectation; no capacity or sample-complexity analysis is given, and the convergence theorem explicitly excludes the minibatch step.
  • ad hoc to paper Terminal condition is enforced exactly by ansatz (5.1).
    Theorem requires u_n(T)=g; the ansatz imposes this with gamma>0. If the ansatz cannot represent the solution class near the terminal time, the fixed-point argument is altered.
  • ad hoc to paper Freezing-and-compensating reformulation (3.25)-(3.26) is exactly equivalent to (1.1) for HJB equations.
    Requires the phi identity (3.24), the chosen alpha_n, and that frozen drift paths remain controlled; the algorithmic stability argument is heuristic.

pith-pipeline@v1.3.0-daily-deepseek · 36805 in / 14634 out tokens · 154876 ms · 2026-08-03T06:27:42.876222+00:00 · methodology

0 comments
read the original abstract

We propose the {\bf i}terative {\bf S}a{\bf M}pling-{\bf A}nd-{\bf R}egression {\bf T}echnique (iSMART) for high-dimensional martingale-based partial differential equations (PDEs) in this paper. By leveraging the $L^2$-projection property of conditional expectation and adopting the stop-gradient technique, iSMART reformulates the continuous martingale condition derived from PDEs into a sequence of tractable sampling-regression problems within an iterative framework. This approach relies solely on standard SDE path simulation and plain squared-error loss minimization, completely bypassing the need for adversarial optimization or nested expectation estimation in previous methods. iSMART accommodates linear, semi-linear, and fully nonlinear martingale-based PDEs within a unified iterative procedure. In particular, for fully nonlinear Hamilton-Jacobi-Bellman (HJB) equations, a freezing-and-compensating technique is introduced to strategically shift a portion of the nonlinearity into the SDE drift, thereby improving the convergence behavior of the iterations. Numerous numerical experiments on linear reaction-diffusion equations with sharp gradients, semilinear Burgers-type equations, and fully nonlinear HJB equations demonstrate the accuracy, efficiency, and robustness of the proposed approach in various high dimensions.

Figures

Figures reproduced from arXiv: 2607.29470 by Fugui Ma, Tiejun Li, Xiaoguang Li.

Figure 1
Figure 1. Figure 1: Visualization of the spatiotemporal sampling scheme. Subfigures (a) and (b) illustrate the temporal sampling profile at epoch 0 and the [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Validation and quantitative assessment of the iSMART algorithm for the 200-dimensional Problem ( [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Validation and quantitative assessment of the iSMART algorithm for the one-dimensional Problem ( [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Validation and quantitative performance evaluation of the iSMART algorithm for the 1D Problem ( [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Numerical performance comparison between DeepMartNet and iSMART for the 1D Problem ( [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance of iSMART on the 200-dimensional semi-linear problem ( [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Numerical performance comparison between DeepMartNet and iSMART for the 200-dimensional semi-linear problem ( [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison between the iSMART numerical solutions and the exact solutions for the HJB equation ( [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison between the iSMART numerical solutions and the exact solutions for the HJB equation ( [PITH_FULL_IMAGE:figures/full_fig_p022_9.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

37 extracted references · 8 canonical work pages

  1. [1]

    Bellman, Dynamic programming, 1957th Edition, Princeton Landmarks in Mathematics, Princeton University Press, Princeton, NJ, 2010, with a new introduction by Stuart Dreyfus

    R. Bellman, Dynamic programming, 1957th Edition, Princeton Landmarks in Mathematics, Princeton University Press, Princeton, NJ, 2010, with a new introduction by Stuart Dreyfus

    2010

  2. [2]

    Bellman, Adaptive control processes: A guided tour, Princeton University Press, Princeton, NJ, 1961

    R. Bellman, Adaptive control processes: A guided tour, Princeton University Press, Princeton, NJ, 1961

    1961

  3. [3]

    Friedman, Partial differential equations of parabolic type, Prentice-Hall, Inc., Englewood Cliffs, NJ, 1964

    A. Friedman, Partial differential equations of parabolic type, Prentice-Hall, Inc., Englewood Cliffs, NJ, 1964

    1964

  4. [4]

    H. Chen, A. Didisheim, S. Scheidegger, Deep surrogates for finance: With an application to option pricing, J. Financ. Econ. 177 (2026) 104222.doi:10.1016/j.jfineco.2025.104222

    arXiv 2026

  5. [5]

    Y . Zhao, W. Zhang, T. Li, Epr-net: constructing a non-equilibrium potential landscape via a variational force projection formulation, National Science Review 11 (7) (2024) nwae052

    2024

  6. [6]

    J. Han, R. Hu, J. Long, Learning high-dimensional McKean-Vlasov forward-backward stochastic differential equations with general distribution dependence, SIAM J. Numer. Anal. 62 (1) (2024) 1–24.doi:10.1137/ 22M151861X. 30

    2024

  7. [7]

    J. Han, A. Jentzen, W. E, Solving high-dimensional partial differential equations using deep learning, Proc. Natl. Acad. Sci. USA 115 (34) (2018) 8505–8510.doi:10.1073/pnas.1718942115

    doi:10.1073/pnas.1718942115 2018

  8. [8]

    Capponi, C.-A

    A. Capponi, C.-A. Lehalle (Eds.), Machine learning and data sciences for financial markets—a guide to contem- porary practices, Cambridge University Press, Cambridge, 2023

    2023

  9. [9]

    Pardoux, S

    E. Pardoux, S. Peng, Backward stochastic differential equations and quasilinear parabolic partial differential equations, in: Stochastic Partial Differential Equations And Their Applications, V ol. 176 of Lect. Notes Control Inf. Sci., Springer, Berlin, 1992, pp. 200–217.doi:10.1007/BFb0007334

    work page doi:10.1007/bfb0007334 1992

  10. [10]

    J. Ma, P. Protter, J. M. Yong, Solving forward-backward stochastic differential equations explicitly—a four step scheme, Probab. Theory Related Fields 98 (3) (1994) 339–359.doi:10.1007/BF01192258

    doi:10.1007/bf01192258 1994

  11. [11]

    W. E, T. Li, E. Vanden-Eijnden, Applied Stochastic Analysis, V ol. 199 of Graduate Studies in Mathematics, American Mathematical Society, Providence, Rhode Island, 2019

    2019

  12. [12]

    W. E, B. Yu, The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems, Commun. Math. Stat. 6 (1) (2018) 1–12.doi:10.1007/s40304-018-0127-z

    doi:10.1007/s40304-018-0127-z 2018

  13. [13]

    Sirignano, K

    J. Sirignano, K. Spiliopoulos, DGM: a deep learning algorithm for solving partial differential equations, J. Comput. Phys. 375 (2018) 1339–1364.doi:10.1016/j.jcp.2018.08.029

    doi:10.1016/j.jcp.2018.08.029 2018

  14. [14]

    Raissi, P

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, J. Comput. Phys. 378 (2019) 686–707.doi:10.1016/j.jcp.2018.10.045

    doi:10.1016/j.jcp.2018.10.045 2019

  15. [15]

    Y . Zang, G. Bao, X. Ye, H. Zhou, Weak adversarial networks for high-dimensional partial differential equations, J. Comput. Phys. 411 (2020) 109409, 14.doi:10.1016/j.jcp.2020.109409

    arXiv 2020

  16. [16]

    Z. Li, N. B. Kovachki, K. Azizzadenesheli, B. liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Fourier neural operator for parametric partial differential equations, in: International Conference on Learning Representations, 2021

    2021

  17. [17]

    Z. You, Z. Xu, W. Cai, MscaleFNO: multi-scale Fourier neural operator learning for oscillatory functions and wave scattering problems, J. Comput. Phys. 547 (2026) Paper No. 114530, 18.doi:10.1016/j.jcp.2025. 114530

    doi:10.1016/j.jcp.2025 2026

  18. [18]

    Oksendal, Stochastic differential equations: an introduction with applications, Springer Science & Business Media, 2013

    B. Oksendal, Stochastic differential equations: an introduction with applications, Springer Science & Business Media, 2013

    2013

  19. [19]

    Pardoux, S

    E. Pardoux, S. G. Peng, Adapted solution of a backward stochastic differential equation, Systems Control Lett. 14 (1) (1990) 55–61.doi:10.1016/0167-6911(90)90082-6

    doi:10.1016/0167-6911(90)90082-6 1990

  20. [20]

    C. Huré, H. Pham, X. Warin, Deep backward schemes for high-dimensional nonlinear PDEs, Math. Comp. 89 (324) (2020) 1547–1579.doi:10.1090/mcom/3514

    work page doi:10.1090/mcom/3514 2020

  21. [21]

    C. Beck, S. Becker, P. Cheridito, A. Jentzen, A. Neufeld, Deep splitting method for parabolic PDEs, SIAM J. Sci. Comput. 43 (5) (2021) A3135–A3154.doi:10.1137/19M1297919

    work page doi:10.1137/19m1297919 2021

  22. [22]

    Zhang, W

    W. Zhang, W. Cai, Fbsde based neural network algorithms for high-dimensional quasilinear parabolic pdes, Journal of Computational Physics 470 (2022) 111557

    2022

  23. [23]

    J. Seo, S. R. Lee, J. Y . Lee, Unbiased and second-order-free training for high-dimensional pdes, in: International Conference on Machine Learning, 2026

    2026

  24. [24]

    J. Han, J. Long, Convergence of the deep BSDE method for coupled FBSDEs, Probab. Uncertain. Quant. Risk 5 (2020) Paper No. 5, 33.doi:10.1186/s41546-020-00047-w. 31

    work page doi:10.1186/s41546-020-00047-w 2020

  25. [25]

    Jentzen, D

    A. Jentzen, D. Salimova, T. Welti, A proof that deep artificial neural networks overcome the curse of dimension- ality in the numerical approximation of Kolmogorov partial differential equations with constant diffusion and nonlinear drift coefficients, Commun. Math. Sci. 19 (5) (2021) 1167–1205.doi:10.4310/CMS.2021.v19. n5.a1

    work page doi:10.4310/cms.2021.v19 2021

  26. [26]

    Grohs, F

    P. Grohs, F. Hornung, A. Jentzen, P. von Wurstemberger, A proof that artificial neural networks overcome the curse of dimensionality in the numerical approximation of Black-Scholes partial differential equations, Mem. Amer. Math. Soc. 284 (1410) (2023) v+93.doi:10.1090/memo/1410

    work page doi:10.1090/memo/1410 2023

  27. [27]

    W. Cai, A. He, D. Margolis, DeepMartNet: a martingale-based deep neural network learning method for Dirich- let BVPs and eigenvalue problems of elliptic PDEs inR d∗, SIAM J. Sci. Comput. 48 (1) (2026) C25–C50. doi:10.1137/24M1630773

    work page doi:10.1137/24m1630773 2026

  28. [28]

    W. Cai, S. Fang, T. Zhou, Deep random difference method for high-dimensional quasilinear parabolic partial dif- ferential equations, J. Comput. Phys. 555 (2026) Paper No. 114767, 31.doi:10.1016/j.jcp.2026.114767

    arXiv 2026

  29. [29]

    W. Cai, DeepMartNet – A martingale based deep neural network learning algorithm for eigenvalue/BVP prob- lems and optimal stochastic controls (2023).arXiv:2307.11942,doi:10.48550/arXiv.2307.11942

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.11942 2023

  30. [30]

    Y . Song, P. Dhariwal, M. Chen, I. Sutskever, Consistency models, in: A. Krause, E. Brunskill, K. Cho, B. En- gelhardt, S. Sabato, J. Scarlett (Eds.), Proceedings of the 40th International Conference on Machine Learning, V ol. 202 of Proceedings of Machine Learning Research, PMLR, 2023, pp. 32211–32252

    2023

  31. [31]

    Z. Geng, M. Deng, X. Bai, Z. Kolter, K. He, Mean flows for one-step generative modeling, in: D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, N. Chen (Eds.), Advances in Neural Information Processing Systems, V ol. 38, Curran Associates, Inc., 2025, pp. 75460–75482

    2025

  32. [32]

    M. Deng, H. Li, T. Li, Y . D. Du, K. He, Generative modeling via drifting (2026).arXiv:2602.04770v1

    Pith/arXiv arXiv 2026

  33. [33]

    Z. Liu, T. Li, Free energy surface sampling via reduced flow matching (2026).arXiv:2605.00337

    Pith/arXiv arXiv 2026

  34. [34]

    Andersson, A

    K. Andersson, A. Gnoatto, M. Patacca, A. Picarelli, A deep solver for BSDEs with jumps, SIAM J. Financial Math. 16 (3) (2025) 875–911.doi:10.1137/23M1615048

    work page doi:10.1137/23m1615048 2025

  35. [35]

    E. H. Georgoulis, A. Papapantoleon, C. Smaragdakis, A deep implicit-explicit minimizing movement method for partial integro-differential equations, with application to option pricing in jump-diffusion models, Commun. Nonlinear Sci. Numer. Simul. 156 (2026) Paper No. 109709, 16.doi:10.1016/j.cnsns.2026.109709

    arXiv 2026

  36. [36]

    Karatzas, S

    I. Karatzas, S. E. Shreve, Brownian motion and stochastic calculus, 2nd Edition, V ol. 113 of Graduate Texts in Mathematics, Springer-Verlag, New York, 1991.doi:10.1007/978-1-4612-0949-2

    doi:10.1007/978-1-4612-0949-2 1991

  37. [37]

    D. L. Burkholder, B. J. Davis, R. F. Gundy, Integral inequalities for convex functions of operators on martingales, in: Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability theory, Univ. California Press, Berkeley, CA, 1972, pp. 223– 240. 32

    1970