A general frequency domain method for assessing spatial covariance structures

DOI10.3150/19-BEJ1160MaRDI QIDQ2203611zbMATH OpenOpenAlexFDO

Authors Matthew van Hala, Soutir Bandyopadhyay, Daniel J. Nordman, Soumendra N. Lahiri

Publication date 7 October 2020

Published in Bernoulli (Search for Journal in Brave)

Full work available at URL https://projecteuclid.org/euclid.bj/1598493620

confidence sets random fields sampling spatial periodogram spatial testing spectral moment conditions variogram

Nonparametric tolerance and confidence regions (62G15) Inference from stochastic processes and spectral analysis (62M15) Inference from spatial processes (62M30) Random fields; image analysis (62M40)

Review text

The authors deal with the mean-value zero second-order homogeneous random fields \(Z(s)\), \(s\in\mathbb{R}^d\), which are observed at \(n\) locations \(s_1,\dots, s_n\) in a sampling region \(\mathcal{D}_n\subset\mathbb{R}^d\), \(\mathcal{D}_n=\lambda_n\mathcal{D}_0\), \(\mathcal{D}_0\subset(-1/2,1/2]^d\), \(\lambda_n\to\infty\) as \(n\to\infty\). The sample locations \(s_1,\dots, s_n\) are generated in the following way \(s_k = \lambda_nX_k\), \(k=1,\dots,n\), where \(\{X_k:X_k\subset \mathcal{D}_0, k=1,2\dots\}\) is a sequence of i.i.d. \(\mathbb{R}^d\)-valued random vectors, with probability density function \(f (x)\) with support on the closure of \(\mathcal{D}_0\). This formulation allows the number \(n\) of sampling sites to grow at a different rate than the volume \(O(\lambda^d_n)\) of the sampling region \(\mathcal{D}_n=\lambda_n\mathcal{D}_0\), and get different asymptotic structures depending on the limit \(c=\lim_{n\to\infty}n/\lambda^d_n\in(0,\infty]\). The case where \(c\) is finite corresponds to pure increasing domain (PID) asymptotics and the case \(c=\infty\) corresponds to sampling referred to as mixed increasing domain (MID). As a complication, limit laws of even simple statistics, such as sample means, typically change with the type of spatial asymptotic structure. Note that neither the type of spatial sampling structure (PID or MID) nor the density \(f (x)\) of locations need to be known or estimated in the proposed testing approach. The proposed spatial testing method uses a periodogram \(I_n(\omega)=|d_n(\omega)|^2\), where \[ d_n(\omega)= \lambda_n^{d/2}n^{-1}\sum_{k=1}^{n}Z(s_k)\exp\{i\omega' s_k\}. \] The spatial periodogram here can have a nontrivial bias depending on the spatial asymptotic structure and the spatial sampling density \(f(x)\). Really (see [\textit{Y. Matsuda} and \textit{Y. Yajima}, J. R. Stat. Soc., Ser. B, Stat. Methodol. 71, No. 1, 191--217 (2009; Zbl 1231.62169)]), \[ \lim_{n\to\infty} EI_n(\omega)=c^{-1}\sigma(0)+K\varphi(\omega), \quad\omega\in\mathbb R^d, \] where \(\varphi(\omega)\) is the spectral density of the field \(Z(s)\), \(\sigma(s)\) denotes the field covariance function, \(K=(2\pi)^d\int _{\mathbb R^d} f^2(x)dx\). In the PID case, where \(c\) is finite, there exists a nontrivial bias component, which vanishes asymptotically in the MID case. The bias-corrected periodogram is \[ \tilde{I_n}(\omega)=I_n(\omega)-n^{-1}\lambda_n^{d}\hat{\sigma}(0), \] where \(\hat{\sigma}(0)= n^{-1}\sum_{k=1}^{n}(Z(s_k)-\bar{Z}_n)^2\), \(\bar{Z}_n= n^{-1}\sum_{k=1}^{n}Z(s_k)\), is the sampling variance. We next describe a framework for setting spectral estimating functions and the device for assessing these with empirical likelihood (EL) for the spatial testing. Suppose that a spatial parameter \(\theta\in\Theta\subset\mathbb R^p\) is connected to the spectral density \(\varphi(\omega)\) of the random field \(Z(s)\) by a system of estimating equations with the estimating functions \(G: \mathbb R^d \times \Theta\to\mathbb R^r\) such that the spectral moment condition \[ \int_{\mathbb R^d}G(\omega,\theta) \varphi(\omega)d\omega=0_r \] at a true parameter \(\theta_0\in\Theta\), where \(0_r\in\mathbb R^r\) denotes the zero vector. To construct the estimating functions satisfying the indicated spectral moment condition, the authors propose to compute the (bias-corrected) spatial periodogram on a frequency grid with appropriate frequency spacing. To estimate the value of a parameter \(\theta\), using an estimating function \(G(\omega,\theta)\), the corresponding empirical likelihood (which is a multinomial likelihood) function is defined \[ \mathcal{R}_n(\theta) = \sup \left\{ \prod_{k=1}^{n}Np_k: \sum_{k=1}^{n}p_k=1,p_k\geq 0, \sum_{k=1}^{n}p_k G(\omega_{kn},\theta)\tilde{I_n}(\omega_{kn})=0_r \right\}, \] based on spatial periodogram on a frequency grid. Maximizing the function \(\mathcal{R}_n(\theta)\) over the parameter space \(\Theta\subset\mathbb R^p\) we get a point estimator \(\hat{\theta}_n\in\Theta\) (or the maximum EL estimator (MELE) of \(\theta\)). The authors propose a detailed construction and numerically illustrate the frequency domain methodology, respectively, for tests of isotropy, separability, and variogram parameters. A wide range of results is provided to validate the general methodology for testing spatial covariance structure over broad classes of potential problems.

Recommendations

Cites work

Cited in

(9)

This page was built for publication: A general frequency domain method for assessing spatial covariance structures

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2203611)