Bispectral estimation: A multitaper approach

We have compared the bias and variance properties of several multitaper based bispectral estimators, with those of frequency smoothed biperiodograms, and Hanning-tapered/frequency-smoothed biperiodograms. The test processes are a white Gaussian process and a non-Gaussian ARMA(2,2) process. We found that the bias and variance properties of the multitaper bispectral estimators are governed mainly by a quantity we call the total bispectral window. We conclude that the multitaper based bispectral estimator are superior to the conventional non-parametric bispectral estimators.