Cumulant Contrast Function

The contrast function measures non-Gaussianity as a proxy for independence. By the Central Limit Theorem, a sum of independent random variables is more Gaussian than any individual source. Maximizing a higher-order cumulant (e.g., kurtosis or negentropy) on the extracted component drives the system toward a source signal.

• Kurtosis-based contrast: Fast to compute, sensitive to outliers
• Negentropy-based contrast: Statistically optimal, requires approximation
• The contrast function achieves its maximum when the output equals a single source

Cumulant contrast functions enable blind source separation—no training data, channel estimates, or modulation labels are required. The algorithm exploits only the statistical structure of the received mixture.

• Works on co-channel signals overlapping in time and frequency
• No need for pilot tones or preambles
• Applicable to unknown modulation types in electronic warfare scenarios
• Recovers both the sources and the mixing matrix simultaneously

All cumulants of order greater than two are identically zero for Gaussian processes. This property makes cumulant contrast functions theoretically immune to additive white Gaussian noise.

• Noise does not shift the location of the contrast maximum
• Enables separation at negative SNR conditions where power-based methods fail
• Fourth-order cumulants suppress Gaussian interference while preserving modulated signal structure

The contrast function landscape is non-convex with multiple local maxima, each corresponding to a different source signal. Optimization algorithms must navigate this terrain carefully.

• Fixed-point iteration (FastICA): Converges quadratically near the solution
• Gradient ascent: Slower but more stable with step-size control
• Deflationary approach: Extract sources one by one, removing each before the next
• Symmetric orthogonalization prevents convergence to the same source twice

Different modulation schemes exhibit distinct cumulant signatures. The contrast function naturally separates signals by their modulation type when sources have different higher-order statistics.

• Sub-Gaussian signals (PSK): Negative kurtosis, contrast minimized
• Super-Gaussian signals (QAM with high-order constellations): Positive kurtosis
• Platykurtic vs. leptokurtic distributions separate cleanly in the contrast landscape
• Enables simultaneous separation and modulation identification

Sample cumulant estimation scales as O(N) with the number of samples, making contrast function evaluation efficient for streaming applications. The full ICA iteration adds matrix operations.

• Batch estimation: Compute cumulants over a fixed observation window
• Recursive updates: Online algorithms update cumulant estimates per sample
• FPGA implementations achieve microsecond latency for tactical SIGINT
• Complexity grows with the square of the number of sources for joint diagonalization