Abstract:This study introduces an uncertainty-aware, mesh-free numerical method for solving Kolmogorov PDEs. In the proposed method, we use Gaussian process regression (GPR) to smoothly interpolate pointwise solutions that are obtained by Monte Carlo methods based on the Feynman-Kac formula. The proposed method has two main advantages: 1. uncertainty assessment, which is facilitated by the probabilistic nature of GPR, and 2. mesh-free computation, which allows efficient handling of high-dimensional PDEs. The quality of the solution is improved by adjusting the kernel function and incorporating noise information from the Monte Carlo samples into the GPR noise model. The performance of the method is rigorously analyzed based on a theoretical lower bound on the posterior variance, which serves as a measure of the error between the numerical and true solutions. Extensive tests on three representative PDEs demonstrate the high accuracy and robustness of the method compared to existing methods.
Numerical Analysis,Computational Engineering, Finance, and Science,Optimization and Control
What problem does this paper attempt to address?
### What problems does this paper attempt to solve?
This paper aims to solve two main challenges in the numerical solution of Kolmogorov partial differential equations (PDEs):
1. **Difficulty in evaluating the validity of numerical results**:
- Existing numerical methods such as the finite - difference method and the finite - element method require complex mathematical derivations for error analysis, and different forms of PDEs require different analysis methods.
- The error bounds provided by these methods are sometimes not tight enough to evaluate the errors of computational results in detail.
2. **High computational complexity of high - dimensional problems**:
- Using grid discretization for PDEs will cause the solution complexity of the linear equation system to increase cubically with the grid size. This complexity becomes extremely large in high - dimensional spaces, making it impractical to solve high - dimensional PDEs.
To solve the above problems, the paper proposes a new numerical method, a meshless numerical method based on Gaussian Process Regression (GPR). The main features of this method include:
- **Uncertainty evaluation**: Through the probabilistic nature of GPR, the uncertainty of the numerical solution can be evaluated.
- **Meshless computation**: It avoids the computational complexity problems caused by relying on grids in traditional methods, and thus can efficiently handle high - dimensional PDEs.
Specifically, this method uses the Monte Carlo method combined with the Feynman - Kac formula to obtain point - like solutions and performs smooth interpolation on them through GPR. GPR not only provides the mean function of the regression result but also allows the uncertainty of the regression result to be evaluated through the posterior variance. In addition, by adjusting the kernel function and combining the noise information in the Monte Carlo samples, the quality of the solution can be further improved.
### Main contributions
1. **Uncertainty awareness**: Through the probabilistic characteristics of GPR, it provides a quantitative evaluation of the uncertainty of the numerical solution.
2. **Meshless computation**: Significantly reduces the computational complexity of solving high - dimensional PDEs.
3. **Theoretical analysis**: Through strict analysis of the lower bound of the posterior variance, it provides a measurement standard for the error between the numerical solution and the true solution.
4. **Experimental verification**: Through extensive testing of three representative PDEs, it is proved that this method has higher accuracy and robustness compared with existing methods.
### Formula summary
- The form of Kolmogorov PDE:
\[
\frac{\partial v(t, x)}{\partial t}+\frac{1}{2} \text{tr}\left(a(t, x) a(t, x)^{T} \frac{\partial^{2} v(t, x)}{\partial x^{2}}\right)+b(t, x)^{T} \frac{\partial v(t, x)}{\partial x}+c(t, x) v(t, x)+h(t, x)=0
\]
Boundary condition:
\[
v(T, x)=g(x)
\]
- Mean and variance of the posterior distribution:
\[
p(v(0, x) \mid X, U, x)=\mathcal{N}(\hat{u}(x), \hat{k}(x, x))
\]
where,
\[
\hat{u}(x)=k_{x}^{T}\left(K_{XX}+\frac{R_{XX}}{M}\right)^{-1} U
\]
\[
\hat{k}(x, x^{\prime})=k(x, x^{\prime})-k_{x}^{T}\left(K_{XX}+\frac{R_{XX}}{M}\right)^{-1} k_{x}^{\prime}
\]
Through these improvements, this method performs excellently in solving high - dimensional PDEs and at the same time provides an effective evaluation of the uncertainty of the numerical solution.