On estimation of a regression model with long-memory stationary errors (Q1113248)

Let \(Y_ t=X'_ t \beta +\epsilon_ t\) where \(X_ t\) has k components and \(\epsilon_ t\) is stationary with zero mean, finite variance and absolutely continuous spectrum. The author considers the case where the spectral density is \[ f(\omega)=f^*(\omega)/| 1- e^{i\omega}|^{2d},\quad 0<d<, \] and \(f^*(\omega)\) is bounded and nonnegative. He also considers some other related cases. He shows in particular that the LSE for \(\beta\) is strongly consistent using only certain conditions essentially on the rate of increase, with T, of the smallest eigenvalue of \(\sum^{T}_{1}X_ tX'_ t\). For \(X'_ t \beta\) a polynomial of degree \(\ell\) in t he evaluates the asymptotic efficiency of the LSE relative to the BLUE, which is 1 for \(d=0\). For \(\ell =2\) this decreases to 8/9 ad \(d=0\cdot 5\). For both LSE and BLUE the variances of the estimates of \(\beta\) increase at a rate increased by the factor \(T^{2d}\) as compared to \(d=0.\) Estimates of d and the innovation variance, \(\sigma^ 2\), for \(\epsilon_ t\) are defined and shown to be strongly consistent when the asymptotic correlation matrix of the \(X_ t\) exists and is non singular and \(\epsilon_ t\) is ergodic. For the polynomial case a central limit theorem is established for the estimates of d and \(\sigma^ 2\), under, also, martingale difference conditions on the innovations.