This function estimates the bispectrum by first estimating the third-order cumulants of the random process, x(t), which is formally
E\{x*(t)x(t+k)x(t+l)\}
The asterisk denotes complex conjugation. k and l are lags.
'biased' (the default) means that the average in the expectation is obtained by dividing by the number of samples 'unbiased' means the average in the expectation is obtained by dividing by the number of samples-1
Look at biased vs. unbiased estimates in statistics and you'll see where that comes from. I would recommend using 'biased' in this case for the same reasons you should use biased in the autocorrelation sequence estimates. This is well documented in the literature.
Without the application of a smoothing window, the bispectrum estimate is an inconsistent estimator of the true bispectrum. The nlags parameter determines how many samples are used in the smoothing window. I would start with approximately 1/4 of your input time series length as a first pass.
