Invariance principles for non-isotropic long memory random fields (Q2475288): Difference between revisions

Following the scheme of \textit{G. Lang} and \textit{P. Soulier} [Stat. Inference Stoch. Process. 3, No. 1--2, 41--51 (2000; Zbl 0978.60023)] proposed for stochastic processes, the author extends the spectral convergence theorem to random fields. Let \((\xi_k)_{k\in\mathbb{Z}^d}\) be a centered random field having a spectral density \(f_{\xi}\) bounded by \(M>0\). Assume that finite dimensional distributions of the partial sums constructed in a standard manner for a field \(\xi\) converge to the corresponding distributions of a field \(B\) defined on \(]0,\infty[^d\). Let \(W_n(A)=n^{d/2}W(n^{-1}A)\), where \(W\) arises in a spectral representation for \(\xi\), \(A\) be a Borel subset of \([-n\pi,n\pi]^d\), \(n\in\mathbb{N}\). The main result (Theorem 1) states that there exists a linear mapping \(I: L^2(\mathbb{R}^d) \to L^2(\Omega,\mathcal{A},{ P})\) such that if \(\Phi_n \to \Phi\) in \(L^2(\mathbb{R}^2)\) then \(\int \Phi_n(x)W_n(dx) \to I(\Phi)\) in law as \(n\to \infty\). Moreover, this \(I\) need not be an isometry, thus it cannot be viewed as a stochastic integral. This result is applied to the study of appropriately normalized partial sums generated by a field \(X\) constructed by filtering the noise \(\xi\). The author obtained different results when \(d \leq 2\) and \(d \geq 3\). The convergence of partial sums is established for strongly dependent random fields whose memory does not satisfy the regularity conditions previously met in the literature.