Pith. sign in

REVIEW 2 major objections 5 minor 54 references

Neural feedback approximation for stochastic control with degenerate diffusions: error estimates and numerical analysis

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Pith's one-line read Direct neural feedback learning for stochastic control has an explicit averaged error bound that still works for degenerate and deterministic dynamics.

desk verdict Solid averaged error bound for direct neural feedback learning without density assumptions; the only real soft spot is the non-quantitative Lipschitz approximation the authors already flag. read the letter →

arxiv 2607.10254 v1 pith:JGHU72Z5 submitted 2026-07-11 math.OC cs.NAmath.NA

classification math.OCcs.NAmath.NA MSC 93E2049M2565C3068T0760H10
keywords stochasticoptimalcontroldegeneratediffusionsneuralnetworksfeedbackapproximationerrorestimatestime-discretedirectpolicylearning
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 studies finite-horizon stochastic optimal control and replaces the usual grid or PDE solvers with a direct learning problem: parametrize the feedback policy by a neural network, simulate the controlled trajectories, and minimize the expected cost. The main theoretical claim is a quantitative averaged error bound between the value of a fixed time-discrete control problem and the value induced by an approximately optimized neural policy. That bound splits cleanly into three pieces: how well a near-optimal feedback can be approximated by the network class on a compact set, the probability that stochastic trajectories leave that set, and the training optimization tolerance. Because the argument uses only moment stability and localization, it does not need transition densities and therefore covers non-degenerate, degenerate, and fully deterministic controlled dynamics in one framework. Numerical examples on a radial target problem, an HJB benchmark, and a gas-storage problem show that the same decomposition appears in practice, so time-step error, piecewise-constant restriction, network approximation, and Monte Carlo noise can be isolated separately.

What carries the argument

Theorem 4.6: an averaged error estimate that decomposes the neural-policy value error into feedback approximation on compact sets, probabilistic localization of trajectories, and training tolerance, proved via discrete Gronwall stability for controlled Euler schemes together with a density argument that replaces measurable feedbacks by Lipschitz ones in value.

What would settle it

On a problem whose optimal feedback is discontinuous (e.g., bang-bang gas storage), keep the network architecture and training fixed while refining the time grid: if the observed value error fails to track the policy-approximation-plus-localization bound of Theorem 4.6 once time-discretization error is negligible, the central estimate is false.

Watch

Extended reading notes

Core claim

For a fixed time grid, the averaged L1 error between the discrete value V0 and the value induced by an η-suboptimal neural feedback is controlled by the uniform approximation of an ε-optimal feedback on a localization ball, plus a localization term of order 1/(M−R) from trajectories leaving that ball, plus ε and η. The estimate does not require transition densities and therefore remains valid for degenerate diffusions and deterministic dynamics.

Load-bearing premise

Near-optimal measurable feedbacks can be replaced in value by Lipschitz feedbacks, but the paper gives no bound on how large that Lipschitz constant must grow when the true optimum is discontinuous.

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 / 5 minor

Summary. The paper studies finite-horizon stochastic optimal control and approximates the time-discrete value V0 by a direct policy-learning problem over neural-network feedback maps. The main result (Theorem 4.6) gives an averaged L1 error bound between V0 and the value induced by an approximately optimized neural policy, for compactly supported initial laws. The bound separates (i) uniform approximation of an ε-optimal feedback on a localization ball BM, (ii) a localization term controlled by trajectory moments, and (iii) the training tolerance η. The analysis relies on discrete Gronwall stability (Lemma 4.1), localization (Lemmas 4.3–4.5), and density of Lipschitz feedbacks in value (Proposition 4.11), without transition-density or non-degeneracy assumptions, and thus covers degenerate and deterministic dynamics. Convergence of the neural approximation is stated in Theorem 4.10. Three numerical examples (degenerate radial target, HJB benchmark, gas storage) illustrate the method and separate time-discretization, piecewise-constant policy restriction, neural approximation/optimization, and Monte Carlo errors.

Significance. If the estimates hold as stated, the paper supplies a useful quantitative framework for global direct neural feedback learning that is not restricted to non-degenerate diffusions. The explicit separation of approximation, localization, and optimization errors, and the avoidance of density assumptions, distinguish the analysis from related DPP-based neural control work (e.g., Huré et al.). The numerical section is carefully designed to isolate error sources and recovers expected scheme orders when neural/MC errors are subdominant. The authors correctly flag that the Lipschitz approximation of measurable feedbacks is non-quantitative (Remarks 4.8, 4.13), so no rate in network size is claimed when the optimum is discontinuous. Overall this is a solid contribution to the theory of neural methods for stochastic control with degenerate dynamics.

major comments (2)
  1. The central bound (Theorem 4.6) and the convergence theorem (Theorem 4.10) are carefully derived under the stated assumptions. The only structural limitation that is load-bearing for interpretability is already acknowledged by the authors: Proposition 4.11 and Remarks 4.8/4.13 show that ε-optimal Lipschitz feedbacks exist, but L=L(ε) is not quantified and may blow up for discontinuous (e.g., bang-bang) optima, so CT,L may grow and no rate in network size follows. This does not invalidate the theorems as written, but the abstract and introduction should state more prominently that the quantitative estimate does not yield a rate when the true feedback is discontinuous, and that the numerical examples partly operate in that regime (gas storage is bang-bang; radial target has non-unique controls inside the reachable set).
  2. Section 5 and Remark 4.15: the theory is stated for Lipschitz-constrained classes ĐGL (and cites GroupSort universality), yet the experiments use unconstrained feedforward nets with tanh/SiLU. The paper notes that finite-parameter nets are Lipschitz a posteriori, but does not discuss how the trained Lipschitz constants relate to the L appearing in CT,L, nor whether the approximation property (16) is effectively realized. A short discussion or diagnostic (e.g., estimated Lip constants of trained policies) would strengthen the link between theory and numerics without changing the main claims.
minor comments (5)
  1. Notation: the same symbol α is used for continuous controls, discrete feedbacks, and neural policies; a short notational table or consistent superscripts would help.
  2. Figure 3 and Table 1: the empirical rates for the Platen scheme are close to 2, but the classical weak order-2 theory requires smooth coefficients including the feedback; the paper correctly flags this as empirical (p. 18). A one-sentence reminder in the figure caption would avoid over-reading the rates.
  3. Example 3: the reference N=320 and the piecewise-constant futures curve on N=5 windows are free experimental choices; stating that the observed rate ~N^{-1.32} is relative to this numerical reference (already done) is good; a brief sensitivity check to the reference grid would be welcome if space allows.
  4. Typos/clarity: 'analy sis' in the title line; occasional spacing issues in displayed equations; 'recomputed' misspelled in a figure caption in the gas-storage section.
  5. Related work: the comparison with [28,4] on density assumptions is clear and fair; a short pointer to recent actor-critic HJB work already cited could note that those methods target the PDE residual rather than the global policy cost.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the averaged error bound is derived from SDE stability, localization, and density of Lipschitz feedbacks, not by construction from the neural class or fitted targets.

full rationale

Theorem 4.6 bounds E0[|V̂0(ξ)−V0(ξ)|] by uniform approximation of an ε-optimal feedback on BM, a moment-based localization term, and training tolerance η. The proof chain (Lemmas 4.1–4.5, Cauchy–Schwarz on cost differences, Markov localization) is standard discrete SDE analysis and does not define V0 or the target error in terms of the neural class ĜL. Proposition 4.11 shows Lipschitz policies are dense in value via a measure-theoretic density argument (Lemma A.1); it is openly non-quantitative (Remarks 4.8, 4.13) and does not smuggle the conclusion. Numerical benchmarks (exact radial formula, Hopf–Cole semi-analytics, N=320 storage reference) are external to the training loss. Self-citations to related deterministic neural-HJB/game work supply context only; they are not load-bearing uniqueness or ansatz imports. No fitted-input-as-prediction, self-definitional, or renaming circularity is present.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The theory rests on standard controlled-SDE Lipschitz assumptions, compact convex action sets, and classical discrete-time control facts, plus the paper-specific choice to measure policy approximation in the uniform norm on localization balls and to leave the Lipschitz constant of near-optimal feedbacks unquantified. No new physical entities are introduced; free parameters appear only as numerical training/architecture choices that affect experiments, not the abstract bound’s form.

free parameters (3)
  • Neural architecture and training hyperparameters (depth/width, activation, Adam rates, batch sizes, epochs)
    Chosen by hand per example; they determine the practical approximation class and η, and therefore the observed residual errors, but are not derived from the theory.
  • Localization radius M and reference Lipschitz bound L
    Free analysis parameters balancing approximation on BM against the localization term and the growth of CT,L; not fixed by data but chosen to make the bound informative.
  • Storage reference grid N=320 and piecewise-constant futures curve on N=5 windows
    Numerical reference and input-curve construction choices that define the empirical convergence baseline in Example 3.
assumptions (6)
  • domain assumption Assumptions 2.1 and 2.3: compact convex A; Lipschitz-in-(t,x,a) coefficients b,σ,f and Lipschitz g with linear growth.
    Used throughout well-posedness, value regularity, and all stability estimates.
  • standard math Existence of optimal measurable Markov feedback for the discrete problem under compactness/continuity (Bertsekas–Shreve-type selection).
    Invoked for the feedback formulation and randomization formula (Proposition 3.1).
  • standard math Discrete Euler–Maruyama moment bounds and Gronwall stability under Lipschitz feedbacks.
    Lemmas 4.1 and 4.4; standard SDE numerics adapted to controlled schemes.
  • standard math Lipschitz maps are dense in L2(μ;A) for finite Radon μ when A is compact convex (Lemma A.1).
    Key tool for Proposition 4.11 approximating measurable feedbacks in value.
  • domain assumption Approximation classes satisfy uniform approximation of L-Lipschitz maps on compact balls (Eq. 16), realizable e.g. by GroupSort networks.
    Needed for Theorem 4.10; experiments use unconstrained smooth nets whose Lipschitz constants are not a priori controlled.
  • ad hoc to paper Policy approximation measured by pointwise sup-norm on BM rather than Lebesgue ess-sup or Lp.
    Remark 4.9: required because degenerate laws may be singular; this strengthens the approximation demand relative to density-based analyses.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural feedback approximation for stochastic control with degenerate diffusions: error estimates and numerical analysis." pith.science (2026). https://pith.science/paper/JGHU72Z5

@misc{pith2026260710254,
  author       = {Pith},
  title        = {Pith review of: Neural feedback approximation for stochastic control with degenerate diffusions: error estimates and numerical analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGHU72Z5}},
  note         = {Machine review of arXiv:2607.10254}
}
read the original abstract

We study finite-horizon stochastic optimal control problems and approximate the resulting time-discrete formulation by a direct policy-learning problem over neural-network feedback maps. We prove a quantitative convergence estimate, in an averaged sense, for the error between the time-discrete value and the value induced by an approximately optimized neural policy. The bound separates the approximation of near-optimal feedback policies, the localization of stochastic trajectories on compact sets, and the optimization tolerance in training. The analysis does not require transition-density assumptions and covers possibly degenerate diffusions and deterministic controlled dynamics in a unified framework. Numerical experiments are provided for a degenerate stochastic radial target problem, a Hamilton--Jacobi--Bellman benchmark, and a gas storage problem, illustrating the approach and separating the main error sources: time discretization, restriction to piecewise-constant policies, neural-network approximation, and Monte Carlo evaluation.

Figures

Figures reproduced from arXiv: 2607.10254 by the authors.

Figure 1
Figure 1. (Radial target problem) Left: Simulated independent trajectories over the time hori￾zon [0, T = 0.5] under the trained neural-network policy, with control constraint M = 1. Dots and crosses represent the initial and terminal states, respectively. The blue circle is the target of radius r0 = 1. The pink annular region represents the backward reachable set at time t = 0, that is, Z0. Right: Contour plot of the Monte C… view at source ↗
Figure 2
Figure 2. (Radial target problem) Comparison between the ex [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗
Figure 3
Figure 3. (Radial target problem) Log–log error for the valu [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: (HJB benchmark problem) Error tables in discrete [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]
Figure 5
Figure 5. Figure 5: (HJB benchmark problem) Comparison between the no [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: (Gas storage problem) Simulated trajectories of t [PITH_FULL_IMAGE:figures/full_fig_p028_6.png]
Figure 7
Figure 7. Figure 7: (Gas storage problem) Left: Monte Carlo estimates of the optimized gain for different time grids, with 95% confidence intervals, in log–log scale. Right: Oriented relative error (v320 −vN )/v320 in log–log scale, with confidence intervals computed by the linear delta m…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 6 linked inside Pith

  1. [1]

    L. A. Abbas-Turki, J.-F. Chassagneux, J.-P. Lemor, G. Lo eper, and S. Sananes. Stochastic Policy Gradient Methods in the Uncertain Volati lity Model. arXiv preprint arXiv:2605.06670, 2026

  2. [2]

    C. Anil, J. Lucas, and R. Grosse. Sorting out Lipschitz fu nction approximation. In Proceed- ings of the 36th International Conference on Machine Learni ng, volume 97 of Proceedings of Machine Learning Research , pages 291–301. PMLR, 2019

  3. [3]

    Assellaou, O

    M. Assellaou, O. Bokanowski, and H. Zidani. Error estima tes for second order Hamilton- Jacobi-Bellman equations. Approximation of probabilisti c reachable sets. Discrete Contin. Dyn. Syst. , 35(9):3933–3964, 2015

  4. [4]

    Bachouch, C

    A. Bachouch, C. Huré, N. Langrené, and H. Pham. Deep neura l networks algorithms for stochastic control problems on finite horizon: Numerical ap plications. Methodology and Computing in Applied Probability , 24(1):143–178, 2022

  5. [5]

    Barrera-Esteve, F

    C. Barrera-Esteve, F. Bergeret, C. Dossal, E. Gobet, A. M eziou, R. Munos, and D. Reboul- Salze. Numerical methods for the pricing of swing options: a stochastic control approach. Methodology and Computing in Applied Probability , 8(4):517–540, 2006

  6. [6]

    C. Beck, W. E, and A. Jentzen. Machine learning approxima tion algorithms for high- dimensional fully nonlinear partial differential equation s and second-order backward stochastic differential equations. J. Nonlinear Sci. , 29:1563–1619, 2019

  7. [7]

    D. P. Bertsekas and S. E. Shreve. Stochastic Optimal Control: The Discrete-Time Case . Athena Scientific, 1996

  8. [8]

    Bokanowski, J.-F

    O. Bokanowski, J.-F. Chassagneux, X. Li, and C. Reisinge r. Numerical approximation for path-dependent McKean–Vlasov control with non-asympt otic error estimates. arXiv preprint arXiv:2606.27181, 2026

Show all 54 references
  1. [9]

    Bokanowski, A

    O. Bokanowski, A. Prost, and X. Warin. Neural networks fo r first order HJB equations and application to front propagation with obstacle terms. Partial Differ. Equ. Appl. , 4(5):Paper No. 45, 36, 2023

  2. [10]

    Bokanowski and X

    O. Bokanowski and X. Warin. Representation results and error estimates for differen- tial games with applications using neural networks. Dynamic Games and Applications , 15(2):417–453, 2025

  3. [11]

    Carmona and M

    R. Carmona and M. Ludkovski. Valuation of energy storag e: An optimal switching ap- proach. Quantitative Finance, 10(4):359–374, 2010

  4. [12]

    Chassagneux and A

    J.-F. Chassagneux and A. Richou. Numerical simulation of quadratic BSDEs. The Annals of Applied Probability , 26(1):262–304, 2016. 30

  5. [13]

    S. N. Cohen, J. Hebner, D. Jiang, and J. Sirignano. Neura l actor-critic methods for Hamilton–Jacobi–Bellman PDEs: Asymptotic analysis and nu merical studies. arXiv preprint arXiv:2507.06428, 2025

  6. [14]

    C. Cox. Fieller’s theorem, the likelihood and the delta method. Biometrics, 46(3):709–718, 1990

  7. [15]

    W. E, J. Han, and A. Jentzen. Deep learning-based numeri cal methods for high-dimensional parabolic partial differential equations and backward stoc hastic differential equations. Com- munications in Mathematics and Statistics , 5(4):349–380, 2017

  8. [16]

    El Karoui, S

    N. El Karoui, S. Peng, and M.-C. Quenez. Backward stocha stic differential equations in finance. Mathematical Finance, 7(1):1–71, 1997

  9. [17]

    W. H. Fleming and H. M. Soner. Controlled Markov Processes and Viscosity Solutions , volume 25 of Stochastic Modelling and Applied Probability. Springer, New York, 2nd edition, 2006

  10. [18]

    G. B. Folland. Real Analysis: Modern Techniques and Their Applications. Pure and Applied Mathematics. John Wiley & Sons, New York, 2nd edition, 1999

  11. [19]

    Frikha, M

    N. Frikha, M. Germain, M. Laurière, H. Pham, and X. Song. Actor-critic learning for mean-field control in continuous time. Journal of Machine Learning Research , 26(127):1– 42, 2025

  12. [20]

    Germain, H

    M. Germain, H. Pham, and X. Warin. Neural networks-base d algorithms for stochastic control and PDEs in finance. arXiv preprint arXiv:2101.08068 , 2021

  13. [21]

    X. Guo, A. Hu, and Y. Zhang. Reinforcement learning for l inear-convex models with jumps via stability analysis of feedback controls. SIAM Journal on Control and Optimization , 61(2):755–787, 2023

  14. [22]

    Han and W

    J. Han and W. E. Deep learning approximation for stochas tic control problems. In Ad- vances in Neural Information Processing Systems (NeurIPS) , Deep Reinforcement Learning Workshop, 2016

  15. [23]

    J. Han, A. Jentzen, and W. E. Solving high-dimensional p artial differential equations using deep learning. Proc. Natl. Acad. Sci. USA , 115(34):8505–8510, 2018

  16. [24]

    Han and J

    J. Han and J. Long. Convergence of the deep BSDE method fo r coupled FBSDEs. Proba- bility, Uncertainty and Quantitative Risk , 5(1):1–33, 2020

  17. [25]

    Hu and M

    R. Hu and M. Laurière. Recent developments in machine le arning methods for stochastic control and games. Numerical Algebra, Control and Optimization , 14(3):435–525, 2024

  18. [26]

    M. Hua, M. Laurière, and E. Vanden-Eijnden. A simulatio n-free deep learning approach to stochastic optimal control. arXiv preprint arXiv:2410.05163 , 2024

  19. [27]

    Huang, B

    Z. Huang, B. Negyesi, and C. W. Oosterlee. Convergence o f the deep BSDE method for stochastic control problems formulated through the stocha stic maximum principle. Math- ematics and Computers in Simulation , 227:553–568, 2025

  20. [28]

    C. Huré, H. Pham, A. Bachouch, and N. Langrené. Deep neur al networks algorithms for stochastic control problems on finite horizon: convergence analysis. SIAM Journal on Numerical Analysis, 59(1):525–557, 2021

  21. [29]

    C. Huré, H. Pham, and X. Warin. Deep backward schemes for high-dimensional nonlinear PDEs. Mathematics of Computation , 89(324):1547–1579, 2020. 31

  22. [30]

    E. R. Jakobsen, A. Picarelli, and C. Reisinger. Improve d order 1/4 convergence for piecewise constant policy approximation of stochastic control probl ems. Electronic Communications in Probability, 24:Paper No. 59, 10, 2019

  23. [31]

    P. E. Kloeden and E. Platen. Numerical Solution of Stochastic Differential Equations . Springer, Berlin, Heidelberg, 1992

  24. [32]

    N. V. Krylov. Approximating value functions for contro lled degenerate diffusion processes by using piece-wise constant policies. Electronic Journal of Probability , 4(2):1–19, 1999

  25. [33]

    Lefebvre, G

    W. Lefebvre, G. Loeper, and H. Pham. Differential learni ng methods for solving fully nonlinear PDEs. Digital Finance, 5(1):183–229, 2023

  26. [34]

    G. N. Milstein and M. V. Tretyakov. Stochastic Numerics for Mathematical Physics . Sci- entific Computation. Springer, Berlin, Heidelberg, 2004

  27. [35]

    S. Peng. A general stochastic maximum principle for opt imal control problems. SIAM Journal on Control and Optimization , 28(4):966–979, 1990

  28. [36]

    H. Pham. Continuous-Time Stochastic Control and Optimization with Financial Applica- tions, volume 61 of Stochastic Modelling and Applied Probability . Springer, Berlin, 2009

  29. [37]

    Pham and X

    H. Pham and X. Warin. Actor-critic learning algorithms for mean-field control with moment neural networks. Methodology and Computing in Applied Probability , 27:13, 2025

  30. [38]

    H. Pham, X. Warin, and M. Germain. Neural networks-base d backward scheme for fully nonlinear PDEs. SN Partial Differential Equations and Applications , 2(1):16, 2021

  31. [39]

    Picarelli and C

    A. Picarelli and C. Reisinger. Probabilistic error ana lysis for some approximation schemes to optimal control problems. Systems & Control Letters , 137:104619, 2020

  32. [40]

    Picarelli, M

    A. Picarelli, M. Scaratti, and J. Tam. Extended mean fiel d control: A global numerical solution via finite-dimensional approximation. arXiv preprint arXiv:2503.20510 , 2025

  33. [41]

    ReduceLROnPlateau, 2026

    PyTorch Contributors. ReduceLROnPlateau, 2026. PyTo rch 2.12 documentation

  34. [42]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physic s-informed neural networks: A deep learning framework for solving forward and inverse problem s involving nonlinear partial differential equations. Journal of Computational Physics , 378:686–707, 2019

  35. [43]

    Sirignano and K

    J. Sirignano and K. Spiliopoulos. DGM: A deep learning a lgorithm for solving partial differential equations. Journal of Computational Physics , 375:1339–1364, 2018

  36. [44]

    H. M. Soner, J. Teichmann, and Q. Yan. Learning algorith ms for mean field optimal control. arXiv preprint arXiv:2503.17869 , 2025

  37. [45]

    Stannat and A

    W. Stannat and A. Vogler. Approximation of optimal feed back controls for stochastic reaction–diffusion equations. ESAIM: Control, Optimisation and Calculus of Variations , 31:6, 2025

  38. [46]

    Stannat, A

    W. Stannat, A. Vogler, and L. Wessels. Neural network ap proximation of optimal con- trols for stochastic reaction–diffusion equations. Chaos: An Interdisciplinary Journal of Nonlinear Science, 33(9):093118, 2023

  39. [47]

    Talay and L

    D. Talay and L. Tubaro. Expansion of the global error for numerical schemes solving stochastic differential equations. Stochastic Analysis and Applications , 8(4):483–509, 1990. 32

  40. [48]

    Tanielian and G

    U. Tanielian and G. Biau. Approximating Lipschitz cont inuous functions with GroupSort neural networks. In Proceedings of The 24th International Conference on Artific ial In- telligence and Statistics , volume 130 of Proceedings of Machine Learning Research , pages 442–450. PMLR, 2021

  41. [49]

    Wang and X

    H. Wang and X. Y. Zhou. Continuous-time mean–variance p ortfolio selection: A reinforce- ment learning framework. Mathematical Finance, 30(4):1273–1308, 2020

  42. [50]

    X. Warin. Gas storage hedging. In Numerical Methods in Finance: Bordeaux, June 2010 , pages 421–445. Springer, 2012

  43. [51]

    X. Warin. Reservoir optimization and machine learning methods. EURO Journal on Computational Optimization , 11:100068, 2023

  44. [52]

    Yong and X

    J. Yong and X. Y. Zhou. Stochastic Controls: Hamiltonian Systems and HJB Equation s, volume 43 of Applications of Mathematics . Springer, New York, 1999

  45. [53]

    M. Zhou, J. Han, and J. Lu. Actor-critic method for high d imensional static Hamilton– Jacobi–Bellman partial differential equations based on neu ral networks. SIAM Journal on Scientific Computing , 43(6):A4043–A4066, 2021. A Technical results A.1 Proof of Lemma 4.4 Proof. We sh...

  46. [54]

    Finally, by a discrete Gronwall’s lemma, we get Γ n≤ (C1 +C2|x|2)τn ·eC3τn, whereC1 := 4[b]2 3T + 16[σ ]2 3,C2 := 8[b]2 1 + 4[σ ]2 1 andC3 := 8([b]2 1T + 4[σ ]2 1)

    + (8[b]2 1 + 4[σ ]2 1)|x|2) τn + 8([b]2 1T + 4[σ ]2 1) n− 1∑ i=0 Γ iτ. Finally, by a discrete Gronwall’s lemma, we get Γ n≤ (C1 +C2|x|2)τn ·eC3τn, whereC1 := 4[b]2 3T + 16[σ ]2 3,C2 := 8[b]2 1 + 4[σ ]2 1 andC3 := 8([b]2 1T + 4[σ ]2 1). We conclude, taking C := max{C1,C 2,C 3} ...

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.