Asymptotic properties of the maximum likelihood and cross validation estimators for transformed Gaussian processes

Abstract: The asymptotic analysis of covariance parameter estimation of Gaussian processes has been subject to intensive investigation. However, this asymptotic analysis is very scarce for non-Gaussian processes. In this paper, we study a class of non-Gaussian processes obtained by regular non-linear transformations of Gaussian processes. We provide the increasing-domain asymptotic properties of the (Gaussian) maximum likelihood and cross validation estimators of the covariance parameters of a non-Gaussian process of this class. We show that these estimators are consistent and asymptotically normal, although they are defined as if the process was Gaussian. They do not need to model or estimate the non-linear transformation. Our results can thus be interpreted as a robustness of (Gaussian) maximum likelihood and cross validation towards non-Gaussianity. Our proofs rely on two technical results that are of independent interest for the increasing-domain asymptotic literature of spatial processes. First, we show that, under mild assumptions, coefficients of inverses of large covariance matrices decay at an inverse polynomial rate as a function of the corresponding observation location distances. Second, we provide a general central limit theorem for quadratic forms obtained from transformed Gaussian processes. Finally, our asymptotic results are illustrated by numerical simulations.

Review text

In the increasing-domain asymptotic framework, this paper analyses the robustness of the Maximumum Likelihood (ML) and Cross Validation (CV) estimators for covariance function of a stochastic process, defined for Gaussian processes when a non-linear transformation of those processes that alter the supposed normality is assumed. It shows that in those cases such estimators have good asymptotic properties without the need to estimate the transformation of the Gaussian process. The CV estimator was defined and studied for the Gaussian case in a previous paper whose author is co-author of the present paper. Let's see briefly its definition. \((\Omega,F,P)\) is a probability space, and \(Z:\Omega\times\mathbb{R}^{d}\to\mathbb{R}\) is a stationary Gaussian process with a \(k_{Z}\) covariance function that satisfies the following conditions: \begin{itemize} \item[i)] \(k_{Z}\) is subexponential and asymptotically positive, i.e. there are two positive real constants, \(C_{\mathrm{inf}}\) and \(C_{\sup}\) such that \[ |k_{Z}(s)|\leq C_{\sup}\cdot\exp(-C_{\mathrm{inf}}\|s\|),\quad s\in\mathbb{R}^{d}. \] \item[ii)] If \((x_{i})_{i\geq 1}\) is a sequence in \(\mathbb{R}^{d}\) such that \[ \mathrm{inf}(|x_{i}-x_{j}|/i\neq j)>0 \] then \[ \mathrm{inf}_{n\in\mathbb{N}}(\lambda_{n}(\Sigma_{n}))>0 \] where \(\Sigma_{n}\) is the covariance matrix of \(Z_1,\dots,Z_{n}\) and \(\lambda_1(\Sigma_{n})\geq\cdots\geq\lambda_{n}(\Sigma_{n})\) are the eigenvalues of \(\Sigma_{n}\). \end{itemize} Let \(F:\mathbb{R}\to\mathbb{R}\) be a continuously differentiable non-constant, non-decreasing, and subexponential function, i.e. there is a constant \(C>0\) such that for each \(t\in\mathbb{R}\) \[ \max(|F(t)|,|DF(t)|)\leq C\cdot \exp(C\cdot |t|), \] where \(DF\) is the derivative of \(F\). It is assumed that the observed process \(Y:\Omega\times\mathbb{R}^{d}\to\mathbb{R}\) is defined by \[ Y(\omega,s)=F(Z(\omega,s)). \] It is further assumed that \[ E(Y_{s})=0\text{ for all }s\in\mathbb{R}^{d}, \] and that the covariance function of \(Y\) is \[ k_{Y,(\Sigma_0,\Psi_0)}=\Sigma_0^2\cdot c_{Y,\Psi_0} \] where \(c_{Y,\Psi_0}\) is the correlation function of \(Y\), with \(\Sigma_0^2\) and \(\Psi_0\) unknown parameters such that \((\Sigma_0^2,\Psi_0)\in[C_{\mathrm{inf}},C_{\sup}]\times S\), \(-\infty<C_{\mathrm{inf}}<C_{\sup}<+\infty\) and \(S\) a compact set of \(\mathbb{R}^{p-1}\) (\(p\geq 2\)). Either \(\Sigma^2\in[C_{\mathrm{inf}},C_{\sup}]\) for \(\Psi\in S\) and each \(i=1,\dots,n\) be \(Y_{i,\Psi}\) the best linear predictor of \(Y_{i}\) based on \((Y_1,\dots,Y_{i-1},Y_{i+1},\dots,Y_{n})\) (if the true parameter is supposed to be \((\Sigma^2,\Psi)\)), it is defined a cross-validation estimator of \(\Psi\) by \[ \mathrm{CV}_\Psi(Y_1,\dots,Y_{i-1})=(1/n)\sum_{i=1}^n(Y_{i}-Y_{i,\Psi})^2. \] It can be seen that this estimator does not depend on \(\Sigma^2\in[C_{\mathrm{inf}},C_{\sup}]\). It is mentioned in this paper that the proposed CV estimator is for the \(\Psi_0\) parameter and not for \(\Sigma_0\), and references are made to other papers by different authors where CV-like estimators are studied for both parameters but with fixed domain when asymptotic behavior is analyzed. Under the Gaussian model \(\Psi_{\mathrm{ML}}\) is more accurate than \(\Psi_{\mathrm{CV}}\), say the authors based on simulations. The opposite result is observed under the transformed model. Therefore, the authors suggest that a combination of both estimators could lead to new accurate estimators in both models. Therefore, a central limit theorem of the joint asymptotic behavior of both estimators is shown in the paper. The authors explicitly warn that their results on the covariance estimator both the ML and the VC assume a known constant mean. However, they caution that no results are known for the case of unknown constant mean under the transformed Gaussian model and state that it would be useful to address this issue in the future. This paper stands out for its didactic. Through several notes, the authors comment on the main results. These help to understand and possibly extend such results.

Recommendations

Cites work

Cited in

(9)

This page was built for publication: Asymptotic properties of the maximum likelihood and cross validation estimators for transformed Gaussian processes

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2180085)