An exact form of the local Whittle likelihood is studied with the intent of developing a general-purpose estimation procedure for the memory parameter (d) that does not rely on tapering or differencing prefilters. The resulting exact local Whittle estimator is shown to be consistent and to have the same N(0, 1 4 ) limit distribution for all values of d if the optimization covers an interval of width less than 9 2 and the initial value of the process is known.
Introduction. Semiparametric estimation of the memory parameter (d) in fractionally integrated (I (d))time series is appealing in empirical work because of the general treatment of the short-memory component that it affords. Two common statistical procedures in this class are log-periodogram (LP) regression [1,10] and local Whittle (LW) estimation [5,11]. LW estimation is known to be more efficient than LP regression in the stationary (|d| < 1 2 ) case, although numerical optimization methods are needed in the calculation. Outside the stationary region, it is known that the asymptotic theory for the LW estimator is discontinuous at d = 3 4 and again at d = 1, is awkward to use because of nonnormal limit theory and, worst of all, the estimator is inconsistent when d > 1 [8]. Thus, the LW estimator is not a good general-purpose estimator when the value of d may take on values in the nonstationary zone beyond 3 4 . Similar comments apply in the case of LP estimation [4].To extend the range of application of these semiparametric methods, data differencing and data tapering have been suggested [3,15]. These methods have the advantage that they are easy to implement and they make use of existing algorithms once the data filtering has been carried out. Differencing has the disadvantage that prior information is needed on the appropriate order of differencing. Tapering has the disadvantage that the filter distorts the trajectory of the data and inflates the asymptotic variance. As a consequence, there is presently