“…If u t is, for example, an instantaneous nonlinear function of a Gaussian long memory process v t , then by extension of results for the sample mean of Rosenblatt (1961) or Taqqu (1975), one expects the limit distributions of OLS and GLS to be non-normal in case u t has Hermite rank greater than 1 (that is, its expansion in Hermite polynomials of v t contains no linear component). Even when OLS and GLS are asymptotically normal, they have rate of convergence which is not only affected by trending behaviour in deterministic (such as polynomial-in-t) x t but, unlike our GLS estimates with stochastic x t , is also adversely affected by long memory u t when the limiting spectral distribution function of the normalized x t has a jump at frequency zero (as in the polynomial-in-t case); see Yajima (1988Yajima ( , 1991. Dahlhaus (1995) studied a class of weighted estimates deriving from Adenstedt's (1974) treatment of the simple location model, studying to what extent they achieve the asymptotic Gauss-Markov bound in the presence of long (and negative) memory u t .…”