Paper: arXiv 2401.17472

Authors: Zhipeng Huang, Balint Negyesi, Cornelis W. Oosterlee

Abstract

It is well-known that decision-making problems from stochastic control can be formulated by means of a forward-backward stochastic differential equation (FBSDE). Recently, the authors of Ji et al. 2022 proposed an efficient deep learning algorithm based on the stochastic maximum principle (SMP). In this paper, we provide a convergence result for this deep SMP-BSDE algorithm and compare its performance with other existing methods. In particular, by adopting a strategy as in Han and Long 2020, we derive a-posteriori estimate, and show that the total approximation error can be bounded by the value of the loss functional and the discretization error. We present numerical examples for high-dimensional stochastic control problems, both in case of drift- and diffusion control, which showcase superior performance compared to existing algorithms.

Complexity vs Empirical Score

  • Math Complexity: 9.0/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper is highly mathematical, featuring advanced stochastic calculus, FBSDE theory, and detailed convergence proofs with a-posteriori error estimates. It demonstrates strong empirical rigor by presenting numerical experiments on high-dimensional stochastic control problems, comparing performance against existing algorithms.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Analyze convergence of<br>deep SMP-BSDE algorithm for<br>high-dimensional stochastic control"] --> B["Key Inputs<br>• FBSDE formulations (Drift/Diffusion control)<br>• Neural network approximations<br>• SMP-inspired loss functions"]
  B --> C["Methodology Steps<br>1. SMP-based FBSDE formulation<br>2. A-posteriori error estimation<br>3. Discretization error analysis"]
  C --> D["Computational Process<br>• Deep learning optimizer<br>• Monte Carlo simulations<br>• High-dimensional PDE approximation"]
  D --> E["Key Outcomes<br>• Convergence proof with error bounds<br>• Loss functional < A-posteriori error < Discretization error<br>• Superior performance vs. existing methods<br>• Numerical validation on 100+ dimensions"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper provides a convergence result for the deep SMP‑BSDE algorithm and derives a‑posteriori error estimates that bound the total approximation error by the loss functional value and the discretization error.

Source sentence
In this paper, we provide a convergence result for this deep SMP-BSDE algorithm and compare its performance with other existing methods. In particular, by adopting a strategy as in Han and Long 2020, we derive a-posteriori estimate, and show that the total approximation error can be bounded by the value of the loss functional and the discretization error.

Assumes 2

  • The functions \(\bar\phi = \bar b, \bar\sigma, \bar f\) and \(g\) are twice continuously differentiable in \(x\), bounded at zero, and their first and second derivatives are uniformly Lipschitz in \(x\) and \(u\). model
    Source sentence
    The map ¯φ is C2 in x, and ¯φ(t, 0, u) is bounded for any (t, u) ∈[0, T]×U. Moreover, ¯φ, ¯φx and ¯φxx are uniformly Lipschitz in x and u. Remark 2.
  • The terminal cost \(g\) is convex and the Hamiltonian \(\bar H\) is concave in \((x,u)\) for the optimality theorem. model
    Source sentence
    Suppose that g(·) is convex, ¯H (t, ·, ·, P ∗ t , Q∗ t) defined by (2) is concave for all t ∈[0, T] P almost surely, and the maximum condition ¯H (t, X∗ t , u∗ t, P ∗ t , Q∗ t) = max u∈U ¯H (t, X∗ t , u, P ∗ t , Q∗ t) , a.e. t ∈[0, T], P-a.s. (4) holds.

Data

  • Linear‑quadratic synthetic examples synthetic public
    Source sentence
    Our numerical examples are given for so-called linear quadratic (LQ) type problems, a special case of (1), where ¯b(t, x, u) = Ax + Bu + β, ¯σj(t, x, u) = Cjx + Dju + Σj, (41) ¯f(t, x, u) = 1 2(x⊤Rxx + u⊤Rxux + u⊤Ruu), g(x) = 1 2x⊤Gx, with A ∈Rd×d, B ∈Rd×ℓ, β ∈Rd, Rx ∈Rd×d, Rxu ∈Rl×d, Ru ∈Rℓ×ℓand G ∈Rd×d, Rx, Ru and G are symmetric, and, for j = 1, . . . , m, Cj ∈Rd×d, Dj ∈Rd×ℓ, Σj ∈Rd.

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • The relative approximation errors of both state and adjoint processes are on the order of 10⁻³ for over 95 % of sampling paths. theoretical
    Source sentence
    The relative approximation errors of both processes are O(10−3) for over 95% of the sampling paths.
  • Approximation errors of all processes decay with at least an O(h) rate as the number of discretization points N increases. theoretical
    Source sentence
    Indeed, Figure 2a confirms our theoretical findings, where the approximation errors of all processes decay with a rate of at least O(h).
  • The total approximation error can be bounded by the sum of the time step size h and the value of the objective functional, confirming the a‑posteriori estimate. theoretical
    Source sentence
    In particular, by adopting a strategy as in Han and Long 2020, we derive a-posteriori estimate, and show that the total approximation error can be bounded by the value of the loss functional and the discretization error. The convergence of the proposed algorithm is collected in Figure 2a, where the absolute errors of all corresponding processes are plotted for different val-ues of N. We recall that the insights of

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 11 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (38)

  1. Solving high-dimensional partial differential equations using deep learning (2018)
  2. Convergence of a robust deep FBSDE method for stochastic control (2023)
  3. Solving stochastic optimal control problem via stochastic maximum principle with deep learning method (2022)
  4. Dynamic programming and stochastic control processes (1958)
  5. Mathematical theory of optimal processes (2018)
  6. An introductory approach to duality in optimal stochastic control (1978)
  7. Numerical methods for stochastic control problems in continuous time (1990)
  8. The rate of convergence of finite-difference approximations for parabolic Bellman equations with Lipschitz coefficients in cylindrical domains (2007)
and 30 more
  1. On the rate of convergence of approximation schemes for Bellman equations associated with optimal stopping time problems (2003)
  2. The Rate of Convergence of Finite - Difference Approximations for Bellman Equations with Lipschitz Coefficients (2005) — link
  3. Deep learning approximation for stochastic control problems (2016) — link
  4. Neural networks for control systems—a survey (1992)
  5. A multilayered neural network controller (1988) — link
  6. Piecewise affine neural networks and nonlinear control (1998)
  7. Deep learning methods for mean field control problems with delay (2020)
  8. Deep backward schemes for high-dimensional nonlinear PDEs (2020)
  9. Convergence of the backward deep BSDE method with applications to optimal stopping problems (2023)
  10. Convergence of the deep BSDE method for coupled FBSDEs (2020)
  11. Recurrent neural networks for stochastic control problems with delay (2021)
  12. Deep neural networks algorithms for stochastic control problems on finite horizon: numerical applications (2022)
  13. Deep neural networks algorithms for stochastic control problems on finite horizon: convergence analysis (2021)
  14. Feynman-Kac neural network architectures for stochastic control using second-order FBSDE theory (2020)
  15. Time discretization and Markovian iteration for coupled FBSDEs (2008) — link
  16. The One Step Malliavin scheme: new discretization of BSDEs implemented with deep learning regressions (2024) — link
  17. Convergence analysis of machine learning algorithms for the numerical solution of mean field control and games: II —the finite horizon case (2022)
  18. Neural networks-based algorithms for stochastic control and PDEs in finance (2021) — link
  19. Recent developments in machine learning methods for stochastic control and games (2023) — link
  20. Continuous-time stochastic control and optimization with financial applications (2009)
  21. Stochastic controls: Hamiltonian systems and HJB equations (1999)
  22. Backward stochastic differential equations (2017)
  23. A posteriori error estimates for fully coupled McKean–Vlasov forward-backward SDEs (2023) — link
  24. A Posteriori Estimates for Backward SDEs (2013) — link
  25. Forward-backward stochastic differential equations and quasilinear parabolic PDEs (1999)
  26. Backward forward stochastic differential equations (1993)
  27. A numerical scheme for BSDEs (2004)
  28. Convergence of the Deep BSDE method for FBSDEs with non-Lipschitz coefficients (2021) — link
  29. Well-posedness of mean-field type forward–backward stochastic differential equations (2015) — link
  30. Fully coupled forward-backward stochastic differential equations and applications to optimal control (1999)

0 of 38 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.