Local empirical spectral measure of multivariate processes with long range dependence. (Q2574622)

Let \((X_t)_{t\in \mathbb Z}\) be a \(p\)-dimensional stochastic process with a spectral density matrix \(f(\lambda )\) such that \(f(\lambda )\sim \Lambda ^{-1}G\Lambda ^{-1}\) as \(\lambda \to 0^+\) where \(\Lambda = \operatorname {diag}(\lambda ^{H_1-1/2}, \dots , \lambda ^{H_p-1/2})\) and \(G\) is a \(p\times p\) real, symmetric, positive definite matrix. For a sample \(X_1,\dots ,X_n\) define \(w(\lambda ) = (2\pi n)^{- 1/2} \sum _{t=1}^n X_t\text e^{it\lambda }\), \(I(\lambda )=w(\lambda )w^*(\lambda )\), and \(\hat F(\lambda _m) =(2\pi /n) \sum _{j=1}^m\text{Re}(I(\lambda _j))\) for \(\lambda _j=2\pi j/n\). Let \(r\in (0,1)\). The author deals with \[ F_n(r)= \sqrt m \lambda _m^{-1} \Lambda _m(\hat F(\lambda _{[mr]}) - F(\lambda _{mr})) \Lambda _m \] where \(\Lambda _m=\operatorname {diag}(\lambda _m^{H_1-1/2}, \dots , \lambda _m^{H_p-1/2})\). If \(H_i\in [1/2,3/4)\), then it is proved that \(F_n(r)\) is asymptotically normal and the process \((F_n(r))_{0\leq r\leq 1}\) converges weakly to a Gaussian process with zero mean and independent increments. The results are applied to narrow-band least squares estimation in time series regression with long range dependent regressors and errors. The second application is to goodness-of-fit testing in the frequency domain based on the empirical spectral measure.