Gaussian stationary processes: Adaptive wavelet decompositions, discrete approximations, and their convergence (Q2483007): Difference between revisions

This article studies wavelet decompositions for stationary Gaussian processes. Let \(X(t)\) a stationary Gaussian process having the representation \[ X(t)=\int_{\mathbb{R}}g(t-u)dB(u)=\int_{\mathbb{R}}e^{itx}\hat g(x) d\hat B(x), \] where \(g\) is a \(L^2\) function and \(\hat g\) its Fourier Transform. Such a process has the following wavelet decomposition \[ X(t)=\sum_{n=-\infty}^{\infty}a_{J,n}\theta^J(t-2^{-J}n)+\sum_{j=J}^{\infty} \sum_{n=-\infty}^{\infty}d_{j,n}\Psi^J(t-2^{-j}n). \] Where the sequences \(a_J=\{a_{J,n}\}_{n\in\mathbb{Z}}\) and \(d_j=\{d_{j,n}\}_{j\geq J\,,n\in\mathbb{Z}}\) are standard independent Gaussian random variables and \(\hat\theta^j(x)=\hat g (x)2^{-j/2}\hat\phi(2^{-j}x)\), \(\hat\Psi^j(x)=\hat g (x)2^{-j/2}\hat\psi(2^{-j}x)\), where \(\phi\) and \(\psi\) are scaling and wavelets functions, associated with a Multiresolution Analysis. The functions \(\theta^j\) and \(\Psi^j\) are the wavelet's basis. In this work the authors modify the first term in the above decomposition to enhance the approximation. The modification is \[ \sum_{n=-\infty}^{\infty} a_{J,n}\theta^J(t-2^{-J}n) =\sum_{n=-\infty}^{\infty} X_{J,n}\Phi^J(t-2^{-J}n), \] such that \(2^{J/2}X_{J,[2^Jt]}\approx X(t)\) in a certain sense whenever \(t\to\infty\). In the above relation we have \[ \hat\Phi^J(x)=\frac{\hat g(x)}{\hat g_J(2^{-J}x)}2^{-J/2}\hat\phi(2^{-J}x)=\frac{\hat\theta^J(x)}{\hat g_J(2^{-J}x)}, \] with the discrete transform \(\hat g_J(y)\) of a sequence \(g_J=\{g_{J,n}\}\). The random sequence \(X_J=\{X_{J,n}\}\) is defined as \(\hat X_J(x)=\hat g(x)\hat a_J(x)\) in the frequency domain. In section 3 some assumptions that allow built the sequence \(g_J=\{g_{J,n}\}\) are given. Then the procedure is applied to some specific processes, for instance the Ornstein-Uhlenbeck stationary process and also a process having a rational spectral function.