Spectral Kurtosis

Spectral Kurtosis is fundamentally a fourth-order cumulant computed in the frequency domain. While variance (second-order) measures average power, SK measures the temporal variability of that power. A purely Gaussian signal has an SK of zero. A positive SK indicates a 'peaky' spectrum with impulsive components, while a negative SK suggests a more uniform or 'flat' spectral distribution than Gaussian noise. This sensitivity to the shape of the probability distribution of frequency bins allows SK to detect structured signals buried in noise.

For a pure Gaussian process, the expected value of Spectral Kurtosis is exactly zero. The variance of the SK estimate depends on the number of averaged power spectral density estimates. This allows for the calculation of a precise statistical threshold to separate noise from signal. A common rule of thumb: if the absolute value of SK exceeds 2 / sqrt(N), where N is the number of spectral averages, the sample is considered statistically non-Gaussian. This provides a mathematically rigorous, blind detection criterion that does not require a priori knowledge of the signal.

SK excels at detecting rare, high-energy events that are invisible to standard power spectral density analysis. In environments plagued by lightning, switching transients, or radar pulses, these impulsive events create heavy tails in the statistical distribution of spectral power. SK acts as a matched filter for non-Gaussianity, producing a sharp peak at the frequency and time where the impulse occurs. This makes it invaluable for Radio Frequency Interference (RFI) mitigation in radio astronomy and spectrum enforcement, where identifying intermittent bursty interference is critical.

Because different signal types exhibit distinct kurtosis signatures, SK can be used as a contrast function for source separation. In a mixed signal environment, a constant-modulus signal like FM has a negative SK, while a pulsed radar has a highly positive SK. Algorithms like FastICA can use kurtosis maximization or minimization to isolate individual components from a composite signal without any prior training. This 'blind' capability is essential for cognitive radios that must autonomously characterize unknown electromagnetic environments.

SK provides a direct metric for classifying emitters based on their temporal duty cycle. A continuous wave (CW) signal with constant amplitude will exhibit a strongly negative SK. A bursty or intermittent signal, such as a frequency-hopping transmitter or a packet-based IoT device, will show a SK value near zero or slightly positive during its active periods. A highly impulsive signal, like a spark gap or ultra-wideband radar, will show a large positive SK. This creates a one-dimensional feature space for rapid emitter triage.

Unlike time-domain kurtosis, Spectral Kurtosis is computed on individual frequency bins across multiple spectral estimates. This produces a kurtosis spectrum that identifies exactly which frequencies are contaminated by non-Gaussian interference. A narrowband jammer will produce a high SK value only in its occupied bandwidth, while the rest of the spectrum remains at zero. This frequency-selective property allows for surgical notching of interference without discarding the entire wideband signal, preserving maximum communication throughput.