Why are there two frequency axes?

The third frequency is determined as their sum in this convention.

Can the result be imaginary?

Yes; it is generally complex-valued.

Does zero prove a Gaussian process?

No; vanishing third-order structure is insufficient.

AI Glossary / Statistics and probability / Advanced

Bispectrum

Definition

The bispectrum is a third-order frequency-domain description of a signal or random process. For a stationary, zero-mean time series, it is the two-dimensional Fourier transform of the third-order cumulant function, which equals the third-order moment function in this setting. It can reveal relationships among frequency components that an ordinary power spectrum does not describe.

What it means

A power spectrum describes how squared amplitude is distributed over frequencies. A bispectrum also involves combinations of phases, examining two frequencies together with their sum. It is generally complex-valued, so both its magnitude and phase can carry information.
The prefix does not mean that it is simply two power spectra placed together. Two frequency coordinates identify a three-frequency combination.

How it works

One common convention forms the product X(f_1) X(f_2) conjugate[X(f_1+f_2)] from Fourier coefficients and averages across comparable realizations or segments, with the appropriate transform and record-length normalization. Conjugate means reversing the sign of the imaginary part.
Writing each coefficient as amplitude times a phase factor shows that the product's phase is phi_1 + phi_2 - phi_3. Consistent phase combinations can reinforce in the average, while changing combinations can cancel. Opposite conjugation conventions reverse the bispectral phase and must be identified when comparing outputs.
Practical estimation needs choices about centering, windows, segment lengths and averaging. A raw triple product from one segment is not a precise population estimate. Bicoherence is a normalized related measure; its normalization must be specified separately.

Three practical examples

Signal diagnostics: an analyst investigates whether frequency components in a recorded vibration have a consistent sum-phase relationship across segments.
Simulation evaluation: researchers compare signals designed to have identical Fourier magnitudes but different phase combinations.
Feature development: a team tests whether selected bispectral summaries add useful information to a model beyond ordinary spectral features, using separate evaluation data.

Think of it this way

Imagine three rotating hands, with the third associated with the sum of the first two frequencies. Knowing their individual strengths does not tell you whether their angles maintain a relationship. The bispectrum tracks a particular combination of those angles as well as their amplitudes.

Worked example

Consider an original illustration using four segments and normalized Fourier coefficients at three bins, k = 1, 2 and 3. Assume a sufficiently long transform so that these are distinct positive-frequency bins below the Nyquist frequency. For a real signal, the matching negative-frequency coefficients are their complex conjugates. The mean bin is set to zero.
In every segment, let X(1) = 2 and X(2) = 3. First consider dataset A, where X(3) = 4 in all four segments. Each triple product X(1) X(2) conjugate[X(3)] is 2 times 3 times 4 = 24. The four-segment average is also 24.
Now consider dataset B, where X(3) across the four segments is 4, 4i, -4 and -4i, with i squared equal to -1. Its magnitude remains four in every segment. The conjugates are 4, -4i, -4 and 4i, making the triple products 24, -24i, -24 and 24i. Their average is zero because the real and imaginary contributions cancel.
Both datasets have the same squared Fourier magnitudes at these bins in every segment: 4, 9 and 16. Thus these power summaries cannot distinguish A from B. Their averaged triple products do distinguish them: A retains a consistent phase combination, while B's four selected phase combinations cancel exactly.
A third dataset C uses X(3) = 4i in every segment, leaving the other two coefficients unchanged. Its averaged triple product is -24i. This has magnitude 24, just as A does, but phase -pi/2 rather than zero under the stated convention. Looking only at the real part would misleadingly make C appear to have no signal at this frequency pair.
To illustrate scaling, multiply every coefficient in dataset A by two. The squared magnitudes become 16, 36 and 64, four times their previous values. The triple product becomes 4 times 6 times 8 = 192, eight times 24. Raw bispectral magnitude therefore depends on amplitude scale and is not itself a standardized association score.
The conjugate is also essential. In C, omitting it would produce positive 24i instead of negative 24i. More generally, the intended phase difference would become a phase sum. This changes the statistic, rather than merely changing the way its magnitude is displayed.
These are deliberately constructed segment coefficients showing the algebra of an averaged triple-product estimate. A common record-length scaling factor would affect all displayed estimates consistently. The example does not demonstrate statistical significance, prove a physical coupling mechanism or establish that four segments suffice for reliable inference. B's exact cancellation is a property of its chosen phases, not a guarantee about finite samples with random phases.

Illustrative case study

Fictional company Birch Signals compares simulated sensor recordings with similar power summaries. Its analysts compute bispectral estimates using a documented convention and identical preprocessing. They examine magnitude and phase together and repeat the analysis across independently generated records.
When an apparent feature changes with the window or segment length, the team investigates estimation sensitivity before using it as a model input. The scenario illustrates a review process and reports no measured improvement in fault detection.

Common misconceptions and mistakes

A nonzero bispectrum does not by itself prove causal interaction or identify a nonlinear physical mechanism. For example, non-Gaussian innovations passed through a linear filter can produce nonzero third-order structure.
A zero population bispectrum also does not establish Gaussianity: some non-Gaussian processes have vanishing third-order cumulants. Conversely, finite estimates need not be exactly zero when the population value is zero. Confusing raw bispectral magnitude with normalized bicoherence, ignoring phase conventions or comparing differently scaled records without adjustment can produce misleading conclusions.

Business use

Bispectral information can support specialist signal analysis and feature exploration when phase relationships matter. Its value should be tested against simpler summaries and the intended decision. Useful reports state the estimation choices, data scale, uncertainty assessment and whether conclusions concern association or an independently supported mechanism.

Check your understanding

Why did B's estimate vanish?

Answer: Its four complex products canceled.

Why did the power summaries remain identical?

Answer: Changing phase did not change coefficient magnitude.

Why did doubling amplitudes multiply the product by eight?

Answer: The statistic multiplies three coefficients.

Related terms

Power spectrum
Fourier transform
Cumulant
Bicoherence
Biphase
Phase coupling
Trispectrum

Learn next

Vary one segment's phase in dataset B and calculate how the complex average changes while its power summaries stay fixed.

Sources

Oliver T. Schmidt, Bispectral mode decomposition of nonlinear flows (2020), Section 2.1 and equations 9–12: https://arxiv.org/html/2002.04146 . The segment coefficients, scaling calculations and company scenario are original illustrations.

Previous: Biserial correlation Next: Bit-level parallelism