No label defined (Q4239176)

Let \(X=\{X_n\}\) be a regular real-valued stationary Gaussian process with positive continuous spectral density \(g(\lambda)\). Without loss of generality, we assume that \(EX_n=0\). Let \(u_k(\lambda)\in L^2[0,1\pi]\) \((k\geq 1)\) be linearly independent continuous functions such that \(u_k(-\lambda)= \overline{u_k(\lambda)}\). Define the random vector \(Z_n^{(K)}= (Z_{n,1},\dots, Z_{n,k})\) by \(Z_{n,k}= \int^\pi_{-\pi} u_k(\lambda) I_n(\lambda) d\lambda\), where \(I_n(\lambda)\) denotes the periodogram of the data \(\{X_1,\dots,X_n\}\). Using the large deviation theorem [cf. \textit{J. Gärtner}, Theory Probab. Appl. 22, 24-39 (1977); translation from Teor. Veroyatn. Primen. 22, 27-42 (1977; Zbl 0375.60033) and \textit{R. S. Ellis}, Ann. Probab. 12, 1-12 (1984; Zbl 0534.60026)], the author proves the following: If for some \(\widehat \gamma^{(K)}= (\widehat\gamma_1, \dots,\widehat \gamma_K)\in R^K\) there exists a spectral density function \(f^*=f_{\theta^*}\) \((\theta^*\in R^K)\) such that \(\int^\pi_{-\pi} u_k(\lambda)f^*(\lambda)d\lambda=\widehat\gamma_k\), \(k=1,\dots,K\), then for any \(L\) \((>K)\) and \(\delta\) \((>0)\) there is an \(\varepsilon_0\) \((>0)\) such that for \(\varepsilon\) \((\in(0,\varepsilon_0])\) \[ \lim_{n\to\infty} P(Z_n^{(L)} \in B_L(\widehat\gamma^{(L)}, \delta)\mid Z_n^{(K)}\in B_K(\widehat \gamma^{(K)}, \varepsilon)| Z_n^{K})=1, \] where \(B_n(\widehat\gamma^{(m)}, \delta)\) \((\subset R^m)\) denotes the open ball with center \(\widehat\gamma^{(m)}\) and radius \(\delta\) and \(\widehat\gamma_k\) \((k>K)\) are given by \(\widehat\gamma_k= \int^\pi_{-\pi} u_k(\lambda)f^*(\lambda) d\lambda\). The result extends the known results for both the i.i.d. case and the Markov chain case. Some applications are also considered.