Abstract:The periodization of a stationary Gaussian random field on a sufficiently large torus comprising the spatial domain of interest is the basis of various efficient computational methods, such as the classical circulant embedding technique using the fast Fourier transform for generating samples on uniform grids. For the family of Matérn covariances with smoothness index $\nu$ and correlation length $\lambda$, we analyse the nonsmooth periodization (corresponding to classical circulant embedding) and an alternative procedure using a smooth truncation of the covariance function. We solve two open problems: the first concerning the $\nu$-dependent asymptotic decay of eigenvalues of the resulting circulant in the nonsmooth case, the second concerning the required size in terms of $\nu$, $\lambda$ of the torus when using a smooth periodization. In doing this we arrive at a complete characterisation of the performance of these two approaches. Both our theoretical estimates and the numerical tests provided here show substantial advantages of smooth truncation.

What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper aims to solve two key problems related to the sampling method based on periodization, especially for the Matérn covariance function. Specifically: 1. **Size requirements for periodization extension**: - Problem: To ensure that the periodized covariance matrix $\Sigma_{\text{ext}}$ is positive definite, how much should the original grid size $n$ be extended to $m$? - Impact: This completely determines the efficiency of the sampling scheme. The required $m$ depends on the covariance function $\rho$ and the type of periodization method. 2. **Eigenvalue decay rate**: - Problem: As $n$ (and thus $m$) increases, do the eigenvalues $\{\Lambda_{\text{ext}, j}:j = 1,\ldots,m\}$ maintain a consistent decay rate? - Impact: This determines the efficiency of the numerical method based on decomposition (1.5), because a faster decay of eigenvalues can reduce the number of effective independent variables required to describe the random field $Z$. ### Research background and significance In fields such as uncertainty quantification (UQ), the convergence theory of quasi - Monte Carlo integration methods for high - dimensional problems depends on the decay rate of terms in the discrete Karhunen - Loève (KL) expansion. Moreover, the decay rate of eigenvalues is closely related to the decay rate of the exact KL eigenvalues $\lambda_j$ of the covariance operator of the continuous field $Z$. Therefore, studying these problems is crucial for understanding and improving the performance of these methods. ### Main contributions - **Classical periodization vs. smooth periodization**: - The paper provides a complete quantitative characterization of these two periodization methods. - For smooth periodization, the paper gives an explicit condition regarding the Matérn smoothing parameter $\nu$ and the correlation length $\lambda$ such that the periodized covariance matrix is positive definite. - For classical periodization, the paper proves that the asymptotic decay rate of eigenvalues is the same as the exact KL eigenvalues, but with a slight correction factor related to $\log h$. - **Significant reduction in computational complexity**: - By using a smooth truncation function, the dependence on the grid size $h$ in classical periodization can be avoided, thus greatly reducing the computational complexity. ### Conclusion This paper provides a complete quantitative analysis of the performance of the two periodization methods under the Matérn covariance and shows the significant computational efficiency advantage of smooth periodization. These results not only help in understanding the performance of existing methods but also provide a theoretical basis for developing more efficient sampling algorithms.

Unified Analysis of Periodization-Based Sampling Methods for Matérn Covariances

A Note on the Convergence Rate of the Spectral Distributions of Large Sample Covariance Matrices

Inference for Gaussian Processes with Matérn Covariogram on Compact Riemannian Manifolds

Convergence Rates of Spectral Distributions of Large Sample Covariance Matrices

Hypothesis Testing for the Smoothness Parameter of Matérn Covariance Model on a Regular Grid

Smoothness Estimation for Whittle-Matérn Processes on Closed Riemannian Manifolds

The Circular Matern Covariance Function and its Link to Markov Random Fields on the Circle

A Unifying Framework for Global Gaussianization: Asymptotic Equivalence of Locally Stationary Processes and Bivariate White Noise

Multivariate Matérn Models -- A Spectral Approach

Beyond Matérn: On A Class of Interpretable Confluent Hypergeometric Covariance Functions

Covariate shift in nonparametric regression with Markovian design

Multiscale scattered data analysis in samplet coordinates

An asymptotic study of the joint maximum likelihood estimation of the regularity and the amplitude parameters of a Mat{é}rn model on the circle

Bulk Spectra of Truncated Sample Covariance Matrices

Multilevel approximation of Gaussian random fields: Covariance compression, estimation, and spatial prediction

Learning and Concentration for High Dimensional Linear Gaussians: an Invariant Subspace Approach

Asymptotic covariances for functionals of weakly stationary random fields

Optimal Estimation of Structured Covariance Operators

Smooth Interpolation of Covariance Matrices and Brain Network Estimation

On the empirical spectral distribution for certain models related to sample covariance matrices with different correlations

Linear spectral statistics of sequential sample covariance matrices