Bicoherence

The following applies to the analysis of data or signals

$X_i(t)$

The standard cross spectrum is the Fourier transform of the correlation

$C_1(t_1)=⟨X_1(t)X_2(t+t_1)⟩$

where the square brackets imply averaging over t. Similarly, the bispectrum is the Fourier transform of the bicorrelation

$C_2(t_1,t_2)=⟨X_1(t)X_2(t+t_1)X_2(t+t_2)⟩$

The signals X_i can either be different or identical. In the latter case, one speaks of the autocorrelation, autospectrum, auto-bicorrelation or auto-bispectrum.

Bispectrum

Denoting the Fourier transforms of the signals X_i(t) by

${\hat{X}}_i(\omega )$

the bispectrum is defined as

$B({\omega }_1,{\omega }_2)={\hat{X}}_1^{*}(\omega ){\hat{X}}_2({\omega }_1){\hat{X}}_2({\omega }_2)$

where

$\omega ={\omega }_1+{\omega }_2$

Bicoherence

The bicoherence is obtained by averaging the bispectrum over statistically equivalent realizations, and normalizing the result:

$b^2({\omega }_1,{\omega }_2)=\frac{|⟨B({\omega }_1,{\omega }_2)⟩{|}^2}{⟨|{\hat{X}}_1(\omega ){|}^2⟩⟨|{\hat{X}}_2({\omega }_1){\hat{X}}_2({\omega }_2){|}^2⟩}$

The normalization is such that 0 ≤ b² ≤ 1.

The bicoherence is symmetric under the transformations (ω₁,ω₂) → (ω₂,ω₁) and (ω₁,ω₂) → (-ω₁,-ω₂), so that only one quarter of the plane (ω₁,ω₂) contains independent information. Additionally, for discretely sampled data all frequencies must be less than the Nyquist frequency: |ω₁|,|ω₂|,|ω| ≤ ω_Nyq. These restrictions define a polygonal subspace of the plane, which is how the bicoherence is usually represented.

The summed bicoherence is defined by

$\frac{1}{N(\omega )}\sum _{{\omega }_1+{\omega }_2=\omega }{b}^2({\omega }_1,{\omega }_2)$

where N is the number of terms in the sum. Similarly, the total mean bicoherence is

$\frac{1}{N_{tot}}\sum _{{\omega }_1,{\omega }_2}{b}^2({\omega }_1,{\omega }_2)$

where N_tot is the number of terms in the sum.

Interpretation

The bicoherence measures three-wave coupling and is only large when the phase between the wave at ω and the sum wave ω₁+ω₂ is nearly constant over a significant number of realizations.

Notes

• The bicoherence can of course be defined in wavenumber space instead of frequency space by applying the replacements x → t and ω → k.
• The bicoherence can be computed using the (continuous) wavelet transform instead of the Fourier transform, in order to improve statistics.