Improved ARMA spectral estimation using the canonical variate method

Abstract

The canonical variate method of rational system identification is investigated here for its performance in a high-resolution spectral analysis environment. A computationally simple version of the method, due to White [9], is briefly reviewed and applied to several standard examples. It is found that, at the cost of some computational complexity, the method is capable of yielding significantly improved resolution and SNR performance for multiple sinusoids in noise as compared to the computationally efficient high-performance method of Cadzow [8], which can also yield a negative power spectral density estimate under certain conditions. Another interesting feature of this method is that it directly applies to the problem of spectral matrix estimation of a multidimensional time series.