Deep BSDE Schemes

• Deep BSDE schemes are mesh-free, neural network–based methods that solve high-dimensional backward stochastic differential equations and associated PDEs using time discretization and stochastic gradient descent.
• They leverage stochastic reformulation and neural parameterization to bypass the curse of dimensionality, enabling efficient handling of semilinear parabolic PDEs and stochastic control challenges.
• Applications span option pricing and control problems, with theoretical guarantees on convergence under regularity and Lipschitz conditions.

Deep BSDE schemes constitute a class of mesh-free, neural network–based numerical methods for solving high-dimensional backward stochastic differential equations (BSDEs) and their associated partial differential equations (PDEs). These schemes have demonstrated accuracy and scalability on semilinear parabolic PDEs, fully coupled forward–backward SDEs, stochastic control problems, and nonlocal integro-PDEs. By leveraging stochastic reformulation, neural network parameterization, and stochastic gradient descent optimization, deep BSDE methods overcome the curse of dimensionality that afflicts classical grid-based solvers, and support systematic error control under standard regularity and coupling assumptions.

1. Mathematical Formulation and Probabilistic Representation

The classical formulation begins with the semilinear parabolic PDE

$$\partial_t u(t,x) + \mu(t,x)\cdot\nabla_x u(t,x) + \tfrac12\operatorname{Tr}\bigl[\sigma(t,x)\sigma(t,x)^\top \nabla^2_x u(t,x)\bigr] + f\bigl(t,x,u,\sigma^\top\nabla_x u\bigr) = 0, \quad u(T,x) = g(x),$$

where $(t,x)\in[0,T]\times\mathbb{R}^d$, with coefficients $\mu,\sigma$ and driver $f$. The associated BSDE representation, given a forward SDE

$$dX_t = \mu(t,X_t)dt + \sigma(t,X_t)dW_t,\qquad X_0 = x,$$

is

$$\begin{cases} Y_t = g(X_T) + \int_t^T f(s,X_s,Y_s,Z_s)ds - \int_t^T Z_s dW_s, \ Y_t = u(t,X_t),\quad Z_t = \sigma^\top(t,X_t)\nabla_x u(t,X_t) \end{cases}$$

for $t\in[0,T]$, under appropriate regularity and Lipschitz assumptions ensuring unique adapted solutions (Han et al., 7 May 2025).

This connection underlies the solution of PDEs, stochastic control, and quantitative finance problems—including barrier options, optimal control via HJB, and nonlocal models with jumps.

2. Neural Scheme Design: Time Discretization and Parametrization

Deep BSDE methods employ time discretization—typically via explicit or implicit Euler–Maruyama or Runge–Kutta type schemes—with partitions $0 = t_0 < t_1 < \dots < t_N = T$, and simulation of Brownian increments $\Delta W_n$. For the backward equation, discrete updates take the form

$$\begin{aligned} Y_{n+1} &\approx Y_n - f(t_n, X_n, Y_n, Z_n)\Delta t_n + Z_n^\top \Delta W_n, \ X_{n+1} &= X_n + \mu(t_n,X_n)\Delta t_n + \sigma(t_n,X_n)\Delta W_n. \end{aligned}$$

The unknowns $(t,x)\in[0,T]\times\mathbb{R}^d$0 are parameterized by neural networks:

• Fully-coupled: $(t,x)\in[0,T]\times\mathbb{R}^d$1, $(t,x)\in[0,T]\times\mathbb{R}^d$2, with one network per time-step (Han et al., 7 May 2025).
• Merged/single-network: $(t,x)\in[0,T]\times\mathbb{R}^d$3, using shared weights across steps (Yu et al., 2019).
• Operator scheme: neural operator-valued maps produce $(t,x)\in[0,T]\times\mathbb{R}^d$4 and $(t,x)\in[0,T]\times\mathbb{R}^d$5 as functionals of the terminal data (Nunno et al., 2024).

Automatic differentiation computes the spatial gradient $(t,x)\in[0,T]\times\mathbb{R}^d$6, or, for higher-order sensitivity (e.g., Gamma), additional networks parameterize the Hessian (Kapllani et al., 2024).

3. Training and Optimization Objectives

Training proceeds by minimizing a stochastic loss functional enforcing terminal and discrete-time consistency. Two foundational loss constructs are:

• Terminal condition loss:

$(t,x)\in[0,T]\times\mathbb{R}^d$7

enforcing correct value at the terminal time (Han et al., 7 May 2025).

• Locally additive losses:

$(t,x)\in[0,T]\times\mathbb{R}^d$8

improving training stability in high dimensions (Kapllani et al., 2020).

For problems with boundaries or obstacles (e.g., barrier options), penalties or Brownian-bridge survival factors enforce Dirichlet/absorbing conditions (Yu et al., 2019).

Optimization is typically performed with stochastic gradient descent (Adam), using mini-batches of simulated trajectories, with training terminated by loss plateau or convergence of final loss below the time discretization error (Huang et al., 2024). Regularization (input whitening, batch normalization) can accelerate convergence; batch norm on input alone achieves near-optimal performance at reduced cost (Yu et al., 2019).

4. Advances: High-Order Schemes, Malliavin Sensitivities, and Network Innovations

High-Order Schemes

Deep Runge–Kutta extensions enable higher-order time-discretization (e.g., Crank–Nicolson, two- and three-stage RK)

$(t,x)\in[0,T]\times\mathbb{R}^d$9

with $\mu,\sigma$0 (Euler), $\mu,\sigma$1 (CN/2-stage), $\mu,\sigma$2 (3-stage), contingent on regularity (Chassagneux et al., 2022).

Malliavin Differential Learning

Malliavin-calculus based methods augment the backward scheme to also approximate gradient and Hessian processes, vital for control or risk sensitivities. Both "differential" (forward/backward) and "one-step Malliavin" (OSM) schemes introduce neural networks for the solution triple $\mu,\sigma$3, jointly minimize residuals arising from the discretized system, and demonstrably improve convergence and accuracy for $\mu,\sigma$4 and $\mu,\sigma$5, critical for hedging and HJB (Negyesi et al., 2021, Kapllani et al., 2024, Kapllani et al., 2024).

Network Architectures

• Deep backward schemes deploy separate networks per time step for $\mu,\sigma$6 (Huré et al., 2019).
• Merged time architecture reduces parameter count via weight sharing (Yu et al., 2019).
• XNet/Cauchy-activated networks, and Kolmogorov-Arnold Networks (KANs) with learnable B-spline activations enable significantly faster and smoother gradient estimation, and lower hedging cost, compared to conventional MLPs (Zheng et al., 10 Feb 2025, Handal et al., 16 Jan 2026).

Control Variate and Operator Approaches

Control variate methods decompose the PDE/BSDE into dominant (linear) and residual (nonlinear) parts, reducing variance and accelerating convergence (Takahashi et al., 2021). Operator-based deep BSDE schemes parametrize the entire solution operator as a neural map from terminal data, supporting operator-level generalization (Nunno et al., 2024).

5. Theoretical Guarantees and Convergence

For time-discretized deep BSDE approximations $\mu,\sigma$7, the a posteriori error bounds are

$\mu,\sigma$8

with constants $\mu,\sigma$9 dependent on regularity and coupling (Negyesi et al., 2024, Jiang et al., 2021, Huang et al., 2024).

Convergence holds under Lipschitz continuity, monotonicity/weak coupling, and, in coupled FBSDE, bounds on the cross-coupling parameters. If the forward drift has strong dependence on $f$0, the contraction constant $f$1 may exceed unity, preventing convergence even for perfectly fit losses (see practical guidelines in (Negyesi et al., 2024)).

For non-Lipschitz (e.g., root-mean-square) diffusions, Yamada–Watanabe mollifiers, diagonal structure, and modulus-of-continuity controls ensure stability and theoretical convergence (Jiang et al., 2021).

6. Applications: Option Pricing, Control, Integro-PDEs, and Beyond

Deep BSDE methods are applied in:

• High-dimensional option pricing, including European, American, and barrier derivatives with absorbing boundaries. Survival probabilities via Brownian bridges ensure accuracy in barrier cases (Yu et al., 2019, Handal et al., 16 Jan 2026).
• Stochastic optimal control via FBSDE reformulations. Deep BSDE accommodates both dynamic programming (DP-based, direct value function) and stochastic maximum principle (SMP, adjoint BSDE) (Negyesi et al., 2024, Huang et al., 2024).
• Nonlocal and jump-driven problems, including path-dependent and infinite-activity Lévy models via SDE/BSDE systems with jump correction and neural quadrature (Jakobsen et al., 2024).
• Operator learning for risk measures and dynamic $f$2-expectations, using learned operator-valued neural maps (Nunno et al., 2024).

Benchmarks:

• Allen–Cahn and nonlinear Black–Scholes equations up to $f$3 solved to sub-0.1% accuracy in minutes (Zheng et al., 10 Feb 2025).
• HJB and Hamilton–Jacobi–Bellman equations in $f$4 with robust convergence (Kapllani et al., 2024, Kapllani et al., 2020).
• BSDEs with singular or rough diffusion coefficients (e.g., CIR processes) solved with logarithmic rate error control (Jiang et al., 2021).

7. Extensions, Practical Guidelines, and Outlook

Key advancements include:

• High-order time discretization via deep Runge–Kutta and multistep architectures, balancing approximation and computational cost (Chassagneux et al., 2022, Bussell et al., 2023).
• Backward vs. forward differential learning—backward methods improve training efficiency and sensitivity computation (Kapllani et al., 2024).
• Batch normalization (on input or layer-wise) and feature standardization optimize convergence and runtime (Yu et al., 2019).
• Control variates and asymptotic expansion reduce variance and accelerate fitting in regimes with small parameter coupling (Takahashi et al., 2021).
• Model-free and reinforcement learning connections are realized via “measurability loss” and exploration noise in Deep BSDE-ML frameworks (Wang et al., 2022).

Caveats and limitations are discussed in the convergence literature:

• Full convergence requires careful regularity verification, monotonicity, and weak coupling. In highly coupled SDEs (large $f$5), SMP-FBSDEs circumvent instability by coupling only through $f$6 (Negyesi et al., 2024).
• For high-dimensional $f$7 computation, network width/depth may become a limiting factor; auto-differentiation vs. dedicated networks entails trade-offs in speed and accuracy (Kapllani et al., 2024, Kapllani et al., 2024).
• Theoretical questions persist in the optimization landscape: nonconvex SGD, network expressivity, and generalization bounds remain open problems (Han et al., 7 May 2025).

Ongoing research explores extensions to kinetic (Boltzmann/Vlasov), mean-field game, and fully nonlinear PDE settings, with operator-level learning, adaptive stepping, and global convergence theory driving future methodological developments.