Maximum Likelihood Estimation of the Parameters of Discrete Fractionally Differenced Gaussian Noise Process

A maximum likelihood estimation procedure is constructed for estimating the parameters of discrete fractionally differenced Gaussian noise from an observation set of finite size N. The procedure does not involve the computation of any matrix inverse or determinant. It requires N²/ 2 + 0(N) op-erations. The expected value of the loglikelihood function for estimating the parameter d of fractionally differenced Gaussian noise (which corresponds to a parameter of the equivalent continuous-time fractional Brownian motion related to its fractal dimension) is shown to have a unique maximum with the range of allowable values of d, The maximum occurs at the true value of d. A Cramer-Rao bound on the variance of any unbiased estimate of d obtained from a finite size observation set is derived. It is shown experimentally that the maximum likelihood estimate of d is unbiased and efficient when finite size data sets are used in the estimation procedure. The proposed procedure is also extended to deal with noisy observations of discrete fractionally differenced Gaussian noise.