Maximum-likelihood estimation of the Matérn covariance structure of isotropic spatial random fields on finite, sampled grids

Frederik J. Simons 1,2 , Olivia L. Walbert 1 , Arthur P. Guillaumin 3 , Gabriel L. Eggers 4 , Kevin W. Lewis 5 , and Sofia C. Olhede 6 1 Department of Geosciences, Princeton University, Princeton, NJ 08544, USA. E-mail: [email protected] 2 Program in Applied & Computational Mathematics, Princeto · 2025

Read Maximum-likelihood estimation of the Matérn covariance structure of isotropic spatial random fields on finite, sampled grids by Frederik J. Simons 1,2 , Olivia L. Walbert 1 , Arthur P. Guillaumin 3 , Gabriel L. Eggers 4 , Kevin W. Lewis 5 , and Sofia C. Olhede 6 1 Department of Geosciences, Princeton University, Princeton, NJ 08544, USA. E-mail: [email protected] 2 Program in Applied & Computational Mathematics, Princeto (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-06223.txt.

Introduction

Fitting the Matérn covariance, the workhorse model of spatial randomness, to data on a finite grid gets a careful maximum-likelihood treatment in a 2025 paper led by Frederik J. Simons with five co-authors. Variance, smoothness, and range govern the field, but discretization and edge effects distort naive estimates. Their spectral procedure uses a debiased Whittle likelihood that treats those effects explicitly.

A specification test accompanies the estimates, because even the best fit may deserve rejection. Worked examples on real geophysical fields show the workflow, aimed at interpolation, kriging, machine learning, and feature detection in geological data. The authors stress honest uncertainty quantification for sampled fields.

Catalog record

CREATORFrederik J. Simons 1,2 , Olivia L. Walbert 1 , Arthur P. Guillaumin 3 , Gabriel L. Eggers 4 , Kevin W. Lewis 5 , and Sofia C. Olhede 6 1 Department of Geosciences, Princeton University, Princeton, NJ 08544, USA. E-mail: [email protected] 2 Program in Applied & Computational Mathematics, Princeto
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE101 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-06223

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Maximum-likelihood estimation of the Matérn covariance structure of isotropic spatial random fields on finite, sampled grids Frederik J. Simons 1,2 , Olivia L. Walbert 1 , Arthur P. Guillaumin 3 , Gabriel L. Eggers 4 , Kevin W. Lewis 5 , and Sofia C. Olhede 6 1 Department of Geosciences, Princeton University, Princeton, NJ 08544, USA. E-mail: [email protected] 2 Program in Applied & Computational Mathematics, Princeton University, Princeton, NJ 08544, USA 3 Queen Mary University of London, London, UK 4 Wesleyan University, Middletown, CT 06459, USA 5 Earth and Planetary Sciences, Johns Hopkins University, Baltimore, MD 21218, USA 6 Ecole Polytechnique Fédérale de Lausanne, Switzerland Abstract keywords Fourier transforms. Geostatistics. Spectral analysis. Statistical methods. {summary} We present a statistically and computationally efficient spectral-domain maximum-likelihood procedure to solve for the structure of Gaussian spatial random fields within the Matérn covariance hyperclass. For univariate, stationary, and isotropic fields, the three controlling parameters are the process variance, smoothness, and range. The debiased Whittle likelihood maximization explicitly treats discretization and edge effects for finite sampled regions in parameter estimation and uncertainty quantification. As even the ‘best’ parameter estimate may not be ‘good enough’, we provide a test for whether the model specification itself warrants rejection. Our results are practical and relevant for the study of a variety of geophysical fields, and for spatial interpolation, out-of-sample extension, kriging, machine learning, and feature detection of geological data. We present procedural details and high-level results on real-world examples. 1 I n t r o d u c t i o n What numbers, which statistical notions capture the “essence” of a spatial patch of geophysical data? If it were a stationary, white, Gaussian process, simply, its population mean and variance would be sufficient. However, computed over differing window sizes, or at varying resolution, sample means and variances of geophysical data sets fluctuate non-erratically, hence “whiteness” immediately proves to be an untenable assumption. Reporting summary statistics over changing baselines [ [ , e.g.,] for Earth and planetary topography]Sharpton+85,Aharonson+98,Grohmann+2010,Rosenburg+2011 effectively subscribes to the data as realizations from a spatially correlated (locally stationary, Gaussian) process. Estimating even just the variance of said field (that is, at a point), with little bias (that is, accurately) and with a reasonable estimation variance (that is, precisely), requires knowledge of the covariance (between pairs of points). We are led to the estimation of (the parameters of) a spatial covariance function, or alternatively and equivalently, of a spectral density function, from sampled data. The first (spatial estimation) is notoriously noisy since it requires finding data pairs at increasing offsets, and computationally time consuming when it entails inverting a covariance matrix [ [ , see, e.g.,] ]Mardia+84,Kitanidis+85,Vecchia88,Grainger+2021. The second (spectral estimation), while typically faster and less noisy, is famously affected by aliasing, finite-field, and edge effects [ [ , e.g.,] ]Stein95,Appourchaux+98a,Appourchaux+98b,Hamilton2009a,Hamilton2009b. Fourier-domain artifacts lead to estimation bias especially for multidimensional data sets, where the nefarious influence of boundary terms dominates the mean-squared error [ Dahlhaus & Künsch(1987) ] . Common choices for parameterized covariance functions of Gaussian random fields are exponential or squared-exponential forms [ [ , e.g.,] ]Tarantola+84,Montagner86,Gudmundsson+90,Baig+2003,Baig+2004a, defined solely by a variance and a correlation range. Both are special cases of the [ Matérn(1960) ] class of covariance functions [ Guttorp & Gneiting(2006) ] , with a specific smoothness or mean-squared differentiability [ Adler(1981) , Christakos(1992) ] , itself an inversion parameter of interest in a multitude of geophysical application domains such as seismology [ [ , e.g.,] ]Wu+85b,Wu+90,Carpentier+2007,Becker+2007b, seafloor bathymetry and oceanography [ Goff & Jordan(1989a) , Goff & Jordan(1989b) , Goff & Arbic(2010) , Sandwell et al.(2022)Sandwell, Goff, Gevorgian, Harper, Kim, Yu, Tozer, Wessel, & Smith ] , helioseismology [ [ , e.g.] ]Gizon+2004, hydrology [ [ , e.g.,] ]Rodriguez+74,Kitanidis+85, meteorology, and climate science [ [ , e.g.,] ]Handcock+94,Paciorek+2006,Lindgren+2011,North+2011,Sun+2015. In this paper we present a univariate spectral-domain debiased “Whittle” maximum-likelihood procedure [ Simons & Olhede(2013) , Guillaumin et al.(2017)Guillaumin, Sykulski, Olhede, Early, & Lilly , Sykulski et al.(2019)Sykulski, Olhede, Guillaumin, Lilly, & Early ] that estimates the variance, smoothness, and range of an isotropic Matérn Gaussian process, from sampled spatial data. We show how to obtain unbiased estimates for these size, scale, and shape parameters when the region studied is neither rectangular nor circular, nor completely sampled. We exactly calculate their estimation covariance, correctly blurred for finite field effects and without neglecting commonly omitted wave vector correlation effects, such that the results from differently sized and sampled patches can be compared robustly. We fully develop and comprehensively illustrate our algorithms for fully sampled rectangular data sets. Our results have widespread implications for the study of geophysical fields, and should be interpreted in the light of our trying to derive “process” from “parameters”, the end goal being to be able to assign likely formation mechanisms and histories for the patches under consideration. Our results should also be relevant for whomever needs to perform spatial interpolation or out-of-sample extension (e.g. via kriging) on geological data [ Journel & Huijbregts(1978) , Christakos(1992) , Cressie(1993) , Stein(1999) ] . They also carry consequences for machine learning and feature detection [ Rasmussen & Williams(2006) , Porcu et al.(2024)Porcu, Bevilacqua, Schaback, & Oates ] . We present procedural details but also focus on high-level results that have real-world applications. We illustrate our methodology on four geologically and geophysically relevant data sets, assuming stationarity within patches that have been selected via user interpretation. 2 P R E L I M I N A R I E S Readers wishing to come to terms with the geological, geophysical, and geodetic definitions of ‘relief’, ‘topography’, or ‘elevation’ are directed to [ Lambeck(1988) ] , [ Hofmann-Wellenhof & Moritz(2006) ] , and [ Wieczorek(2015) ] . To make the jump from geology and geophysics to statistics, particularly in this context, we first and foremost recommend (re)reading [ Goff & Jordan(1988) , Goff & Jordan(1989a) ] , who also discuss anisotropic processes. The material in this section is both an extension and a specialization of the multivariate results of [ Simons & Olhede(2013) ] , which is to be consulted for further details. Here, we use a more explicit notation, adapt some of the normalizations, and make a number of modifications—but most importantly, we restrict our analysis to univariate two-dimensional Cartesian isotropic Gaussian fields. 2.1 Continuous framework Here we draw most heavily on Sections 2.1 and 4.1 of [ Simons & Olhede(2013) ] . Referring furthermore to [ Percival & Walden(1993) ] , [ Stein(1999) ] , and [ Vanmarcke(2010) ] for additional considerations and terminology, to [ Abramowitz & Stegun(1965) ] and [ Gradshteyn & Ryzhik(2000) ] for properties of special functions, we begin by defining the particular quantities of interest in the spatial and the spectral domains. 2.1.1 Stationarity A geophysical field {\mathcal{H}}(\mathbf{x}) is considered to be a zero-mean, finite-variance, stationary, two-dimensional random field with finite second moments. Under what is known as the [ Cramér(1942) ] representation, there exists a spectral orthogonal-increment process, d{\mathcal{H}}(\mathbf{k}) , according to which the spatial field {\mathcal{H}}(\mathbf{x})=\int\!\!\!\!\int e^{i\mathbf{k}\cdot\mathbf{x}}\,d{\mathcal{H}}(\mathbf{k}),\qquad\mathbf{x}\in\mathbb{R}^{2}. (1) The integration is over the space containing all wave vectors \mathbf{k} . In the case of strict band-limitation or very fast spectral decay we may restrict the computations to the Nyquist plane [-\pi,\pi]\times[-\pi,\pi] . The expectation

Read the complete text