Abstract:We present a multidimensional deep learning implementation of a stochastic branching algorithm for the numerical solution of fully nonlinear PDEs. This approach is designed to tackle functional nonlinearities involving gradient terms of any orders, by combining the use of neural networks with a Monte Carlo branching algorithm. In comparison with other deep learning PDE solvers, it also allows us to check the consistency of the learned neural network function. Numerical experiments presented show that this algorithm can outperform deep learning approaches based on backward stochastic differential equations or the Galerkin method, and provide solution estimates that are not obtained by those methods in fully nonlinear examples.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: to develop a stochastic branching algorithm based on deep learning for solving fully nonlinear partial differential equations (PDEs). Specifically, the authors propose a multi - dimensional deep learning implementation method, combining neural networks with the Monte Carlo branching algorithm to handle functional nonlinear problems involving gradient terms of any order. This method aims to overcome the problem that the computational cost of traditional numerical methods (such as the finite - difference or finite - element methods) increases exponentially with the dimension in high - dimensional cases (i.e., the "curse of dimensionality"), and can provide better performance than methods based on backward stochastic differential equations (BSDEs) or the Galerkin method in some cases.
### Main Objectives:
1. **Solve High - Dimensional Fully Nonlinear PDEs**: By introducing a stochastic branching mechanism, the problem of excessive computational complexity of traditional methods in high - dimensional spaces is avoided.
2. **Handle High - Order Gradient Nonlinearities**: It can handle nonlinear terms containing gradient terms of second - order or higher, which many existing methods cannot do.
3. **Improve the Accuracy and Stability of Numerical Solutions**: Through experiments, this method has shown higher accuracy and better stability in multiple test cases, especially in high - dimensional cases.
4. **Verify the Consistency of Neural Network Learning Results**: Since this method is based on direct Monte Carlo estimation, the consistency between Monte Carlo samples and the learned neural network function can be checked, which is difficult to achieve with other methods.
### Description of the Mathematical Model:
Consider the fully nonlinear partial differential equation in the following form:
\[
\begin{cases}
\partial_t u(t, x)+\frac{1}{2}\Delta u(t, x)+f(\partial^{\lambda_1} u(t, x), \ldots, \partial^{\lambda_n} u(t, x)) = 0, \\
u(T, x)=\phi(x), \quad(t, x)=(t, x_1, \ldots, x_d)\in[0, T]\times\mathbb{R}^d,
\end{cases}
\]
where:
- \( \partial_t u(t, x)=\frac{\partial u(t, x)}{\partial t} \)
- \( \Delta=\sum_{i = 1}^d\frac{\partial^2}{\partial x_i^2} \) is the standard d - dimensional Laplace operator
- \( f \) is a smooth function with respect to the derivatives \( \partial^{\lambda_i} u(t, x) \)
- \( \lambda_i=(\lambda_i^1, \ldots, \lambda_i^d)\in\mathbb{N}^d \)
### Innovation Points of the Method:
- **Stochastic Branching Diffusion Mechanism**: By constructing a random coding tree, the solution of the PDE is represented as a conditional expectation, and then the fully nonlinear Feynman - Kac formula is used for solving.
- **Combination of Deep Learning and Monte Carlo**: Use a neural network to approximate the conditional expectation, and optimize the neural network parameters through the L2 minimization property and the functional approximation theorem.
- **Handle High - Order Nonlinearities**: This method can handle nonlinear terms containing gradient terms of any order, solving the technical problems encountered by existing methods when dealing with high - order gradients.
In conclusion, this paper proposes a new deep - learning framework specifically for the numerical solution of fully nonlinear partial differential equations, especially suitable for high - dimensional and high - order gradient nonlinear problems.