Numerical Considerations for the Construction of Karhunen-Loève Expansions

Karhunen-Loève expansions (KLE) are mathematical tools used to represent second-order stochastic processes, widely applied in uncertainty quantification, material science, and subsurface flows. KLE provides an optimal mean-square-convergent series spectral representation of a stochastic process in terms of deterministic orthogonal basis functions and uncorrelated scalar random variables. This property makes KLE valuable for dimensionality reduction and high-fidelity simulations. Since Ghanem and Spanos introduced KLE in the stochastic finite element method, its applications in computational science and engineering have increased. Le Maître and Knio provide a comprehensive overview of spectral methods for uncertainty quantification, including KLE methods. Despite its theoretical advantages, the numerical implementation of KLE faces challenges, especially in high-dimensional complex domains. Existing methods typically rely on the spectral decomposition of the covariance operator, requiring the solution of Fredholm integral equations. However, these methods face limitations in computational complexity and numerical stability.

Constructing Karhunen-Loève expansions (KLE) numerically poses challenges. The core problem is how to effectively represent, discretize, and sample infinite-dimensional stochastic processes while preserving the statistical structure of the underlying field. Discretizing the Fredholm integral equation converts the continuous eigenproblem into a discrete one, but computational complexity and numerical stability become major bottlenecks in high-dimensional cases. Additionally, the choice of covariance kernel affects the smoothness of random field realizations and the rate of eigenvalue decay, determining the number of KLE terms needed. How to effectively choose distance metrics to evaluate the covariance function in complex domains is also a crucial issue.

The core innovation of this paper is combining Fredholm integral equation eigensolutions with singular value decomposition (SVD) to provide a unified numerical method for Karhunen-Loève expansions (KLE) of second-order stochastic processes. Specifically, the paper derives the algebraic equivalence between Fredholm-based eigensolutions and the SVD of the weight-scaled sample matrix, achieving consistency in model-driven and data-driven KLE construction. This approach not only enhances numerical solution accuracy but also provides new insights for handling high-dimensional complex domains in stochastic processes. Additionally, the paper validates the accuracy of numerical methods in 1D cases using analytical eigensolutions for exponential and squared-exponential covariance kernels as benchmarks.

• �� Construct Karhunen-Loève expansions (KLE) through the spectral decomposition of the covariance operator via the Fredholm integral equation.
• �� Discretize the Fredholm integral equation, forming an eigendecomposition task.
• �� Derive the algebraic equivalence between Fredholm-based eigensolutions and the singular value decomposition (SVD) of the weight-scaled sample matrix.
• �� Validate the accuracy of numerical methods in 1D cases using analytical eigensolutions for exponential and squared-exponential covariance kernels as benchmarks.
• �� In 2D and 3D cases, test using unstructured triangular meshes and a 3D toroidal domain, examining the impact of discretization strategy, quadrature rule, and sample count on KLE results.
• �� Compare Euclidean and shortest interior path distances between grid points, highlighting their effects on eigensolutions.

The experimental design includes 1D, 2D, and 3D test cases. In 1D cases, analytical eigensolutions for exponential and squared-exponential covariance kernels serve as benchmarks to validate numerical methods' accuracy. In 2D cases, tests using unstructured triangular meshes examine the impact of discretization strategy, quadrature rule, and sample count on KLE results. In 3D cases, tests using a 3D toroidal domain compare Euclidean and shortest interior path distances between grid points, highlighting their effects on eigensolutions. Experiments also include convergence analysis of SVD-based eigenvalue estimates and empirical distributions of KL coefficients with varying sample counts.

Experimental results show that SVD-based eigenvalue estimates and empirical distributions of KL coefficients converge to the theoretical N(0,1) target as sample count increases. In 2D and 3D cases, the impact of discretization strategy, quadrature rule, and sample count on KLE results is significant. Particularly in the 3D toroidal domain, Euclidean and shortest interior path distances between grid points significantly affect eigensolutions. Experiments also indicate that covariance kernels with short correlation lengths require more KLE terms to achieve a given accuracy.

The method has significant implications for uncertainty quantification and high-fidelity simulations. By enhancing the accuracy and consistency of KLE numerical solutions, the method can be applied to handle high-dimensional complex domains in stochastic processes. This approach is particularly suitable for applications requiring high-precision random field representations, such as microstructure simulations in material science and uncertainty quantification in subsurface flows.

The method may have high computational complexity when dealing with very high-dimensional stochastic processes, especially with large sample sizes. The choice of distance metric in non-simply-connected complex topologies still requires further research to understand its impact on eigensolutions. For certain covariance kernels, analytical eigensolutions may be difficult to obtain, affecting the validation of numerical methods. Future research will focus on optimizing the algorithm to reduce computational complexity and validating the method's effectiveness in more complex non-simply-connected domains.