Solving stochastic differential equations and Kolmogorov equations by means of deep learning

Christian Beck, Sebastian Becker, Philipp Grohs, Nor Jaafari, Arnulf Jentzen

TL;DR

This paper introduces a novel deep learning-based numerical method to solve high-dimensional Kolmogorov partial differential equations (PDEs) across an entire spatial region. By combining stochastic optimization with deep neural networks, the method successfully bypasses the "curse of dimensionality" while providing a global solution function rather than just a single-point approximation.

What Problem Does This Paper Address?

Kolmogorov PDEs and their closely related counterparts, Stochastic Differential Equations (SDEs), are foundational tools used to model complex systems in engineering, physics, and financial mathematics (such as derivative pricing). However, solving these equations in high dimensions (e.g., 100 dimensions) is notoriously difficult.

Traditional deterministic approaches, like finite difference or finite element methods, suffer from the curse of dimensionality—their computational cost grows exponentially as the number of dimensions increases. Conversely, random approximation methods like Monte Carlo simulations can handle high dimensions but only provide the solution at a single fixed space-time point. If you need to know the solution across a whole region, Monte Carlo becomes prohibitively expensive. This paper tackles this gap by proposing a method that is both scalable to high dimensions and capable of delivering a continuous solution over an entire domain $[a, b]^d$.

Methods

The authors bridge the gap between local stochastic solutions and global deterministic solutions using a deep learning framework:

  1. Stochastic Reformulation: Using the classical Feynman-Kac formula, the authors rewrite the Kolmogorov PDE as an infinite-dimensional stochastic minimization problem.
  2. Temporal Discretization: The underlying SDE is discretized in time using the standard Euler-Maruyama scheme, which simulates the random paths of the system.
  3. Spatial Parameterization via ANNs: Instead of using a rigid grid, the solution to the PDE is approximated across the entire spatial domain using a fully connected deep artificial neural network (ANN).
  4. Stochastic Gradient Descent: The resulting finite-dimensional optimization problem is solved using stochastic gradient descent-type algorithms—specifically, the Adam optimizer. The network learns the global solution function by minimizing the error along the simulated random trajectories.

Datasets

The paper does not use real-world empirical datasets. Instead, the algorithm is rigorously evaluated on several high-dimensional synthetic benchmark problems derived from physics and mathematical finance:

  • The Heat Equation ($d = 100$)
  • Geometric Brownian Motions ($d = 100$)
  • The Black-Scholes Model with Correlated Noise ($d = 100$)
  • Stochastic Lorenz Equations ($d = 3$)
  • The Heston Model ($d = 50$, representing 25 asset prices and 25 corresponding volatilities)

Key Insights and Ideas

  • The "Aha!" Marriage: The core innovation lies in using neural networks as mesh-free function approximators trained via SDE path sampling. This marries the dimension-agnostic scaling of Monte Carlo path sampling with the global generalization power of deep neural networks.
  • Overcoming the Grid: Because neural networks can generalize from sparse data points across a continuous space, the method completely avoids the need for space-discretization grids, effectively breaking the curse of dimensionality.
  • Global vs. Local: Unlike prior deep learning methods that only solved PDEs at a single starting point, this approach trains the network to be accurate across an entire specified region $[a, b]^d$ simultaneously.

Key Results

The numerical results demonstrate that the algorithm achieves remarkable accuracy and speed in high dimensions:

  • High-Dimensional Accuracy: For a 100-dimensional Heat Equation, the relative $L^1$ approximation error dropped to 0.000822 ($0.08\%$) after 750,000 training steps.
  • Financial Models: In the 100-dimensional Black-Scholes model with correlated noise, the relative $L^1$ error reached 0.000471 after 750,000 steps, taking roughly 57 minutes (3466.8 seconds) on a single NVIDIA GeForce GTX 1080 GPU.
  • Complex Dynamics: For the 50-dimensional Heston model, the method achieved a relative $L^1$ error of 0.00174 in about 5.7 hours.

Limitations and Future Work

While the paper demonstrates excellent empirical performance, some limitations and areas for future exploration remain:

  • Computational Intensity for Complex SDEs: For highly nonlinear drift coefficients, such as in the Stochastic Lorenz equations, the runtime increases significantly (taking over 6 hours for a 3D problem due to the complexity of the simulated paths).
  • Hyperparameter Sensitivity: The choice of network architecture, learning rate schedules, and batch sizes (set to 8192 in these examples) heavily influences convergence speed and stability.
  • Theoretical Convergence Bounds: Although empirical results prove the method works well on the chosen examples, a comprehensive mathematical proof establishing general convergence rates for all classes of linear and non-linear Kolmogorov PDEs remains an open area for research.

Summary updated

Abstract

Stochastic differential equations (SDEs) and the Kolmogorov partial differential equations (PDEs) associated to them have been widely used in models from engineering, finance, and the natural sciences. In particular, SDEs and Kolmogorov PDEs, respectively, are highly employed in models for the approximative pricing of financial derivatives. Kolmogorov PDEs and SDEs, respectively, can typically not be solved explicitly and it has been and still is an active topic of research to design and analyze numerical methods which are able to approximately solve Kolmogorov PDEs and SDEs, respectively. Nearly all approximation methods for Kolmogorov PDEs in the literature suffer under the curse of dimensionality or only provide approximations of the solution of the PDE at a single fixed space-time point. In this paper we derive and propose a numerical approximation method which aims to overcome both of the above mentioned drawbacks and intends to deliver a numerical approximation of the Kolmogorov PDE on an entire region $[a,b]^d$ without suffering from the curse of dimensionality. Numerical results on examples including the heat equation, the Black-Scholes model, the stochastic Lorenz equation, and the Heston model suggest that the proposed approximation algorithm is quite effective in high dimensions in terms of both accuracy and speed.
BibTeX
@article{beck2021solving,
  author = {Christian Beck and Sebastian Becker and Philipp Grohs and Nor Jaafari and Arnulf Jentzen},
  title = {{Solving stochastic differential equations and Kolmogorov equations by means of deep learning}},
  journal = {J. Sci. Comput.},
  year = {2021},
  doi = {10.1007/s10915-021-01590-0},
}
Figure