Why Higher-Order Spectra?
Because two time series can look identical through covariance and the ordinary spectrum while hiding very different nonlinear and non-Gaussian dependence.
The ordinary spectrum sees only second-order structure
That is powerful—but incomplete. If the process is nonlinear or non-Gaussian, important dependence can live beyond pairs of observations.
The bispectrum
The bispectrum looks at interactions among triples of frequencies. It can reveal phase coupling and third-order dependence that an ordinary periodogram cannot show.
Polyspectra
Polyspectra extend the same idea to cumulants of order four and beyond. The price is complexity: the object quickly becomes high-dimensional.
Same spectrum, different dynamics
Series A
Linear, symmetric, and fully described by second-order behavior.
Series B
Same covariance pattern, but with nonlinear interactions or asymmetric higher-order dependence.
An ordinary spectrum may struggle to distinguish them. Higher-order spectra are designed precisely for the structure that remains.
Polyspectral means turn a huge spectral object into a targeted statistic
Rather than estimating and interpreting an entire high-dimensional polyspectrum, a weighted mean can focus attention on the frequency interactions relevant to the scientific question.
Statistical advantage
Lower-dimensional summaries are easier to estimate, compare, test, and use as features.
Scientific advantage
The weight function can emphasize particular regions or interaction patterns rather than treating every frequency combination equally.
Use higher-order spectra when second-order structure is not the whole story
They are especially useful when the question involves nonlinear dynamics, asymmetry, non-Gaussianity, frequency coupling, or feature construction for clustering and classification.