Bispectrum Analysis

The bispectrum's fundamental advantage is its theoretical immunity to additive white Gaussian noise (AWGN). For any zero-mean Gaussian process, the bispectrum is identically zero. This means the technique extracts signal features from below the noise floor, isolating the deterministic hardware fingerprint from random thermal noise that plagues power-spectrum-based methods. This property is critical for low-SNR environments where traditional fingerprinting fails.

Bispectrum analysis uniquely identifies quadratic phase coupling (QPC) — the phenomenon where two frequency components interact non-linearly to generate a third component whose phase is the sum of the original two. This coupling is a direct fingerprint of power amplifier non-linearity and mixer imperfections. The bispectrum detects when φ(f1) + φ(f2) = φ(f1+f2), revealing the specific harmonic generation patterns that distinguish one transmitter from another.

Unlike the 1D power spectrum, the bispectrum B(f1, f2) is a 2D complex-valued function defined over a bifrequency plane. This representation captures inter-frequency correlations that are invisible to second-order statistics. Key regions include:

• Principal domain: The triangular region 0 ≤ f2 ≤ f1 ≤ f1+f2 ≤ 1 containing all non-redundant information
• Diagonal slices: B(f, f) reveals self-coupling from harmonic generation
• Off-diagonal regions: B(f1, f2) for f1 ≠ f2 captures intermodulation products

To reduce the high dimensionality of the full bispectrum for classifier input, integrated bispectrum techniques project the 2D function onto 1D representations while preserving discriminative information:

• Radially Integrated Bispectrum (RIB): Integrates along radial lines in the bifrequency plane, capturing phase coupling at specific frequency ratios
• Axially Integrated Bispectrum (AIB): Integrates along lines parallel to one frequency axis
• Circularly Integrated Bispectrum (CIB): Integrates over concentric circles, providing rotation-invariant features These projections make bispectrum analysis computationally feasible for real-time SEI systems.

The bispectrum is estimated from a finite sequence of complex I/Q samples using either:

• Direct method (FFT-based): Segment the data, compute the Fourier transform of each segment, and average the triple product X(f1) · X(f2) · X*(f1+f2) across segments
• Indirect method: First estimate the third-order cumulant sequence, then apply a 2D Fourier transform Both methods require variance reduction through averaging and careful windowing to minimize spectral leakage. The direct method is preferred for computational efficiency in real-time systems.

The bicepstrum is the inverse Fourier transform of the logarithm of the bispectrum. This transformation converts the non-linear convolution of the signal with the channel impulse response into an additive operation in the cepstral domain. By applying a lifter (cepstral filter), channel effects can be separated from the intrinsic device fingerprint, making the extracted features robust to multipath propagation and varying channel conditions.

The bispectrum is the gold standard feature space for SEI because it captures the unintentional non-linear phase couplings introduced by a transmitter's unique power amplifier and oscillator chain. These couplings are invariant to Gaussian noise and form a robust, device-specific fingerprint.

• Extracts features from power amplifier non-linearity and phase noise
• Enables identification of same-model radios from different production batches
• Used in military IFF systems and civilian spectrum enforcement

Bispectrum analysis excels at detecting signals buried in additive white Gaussian noise (AWGN) because the bispectrum of any Gaussian process is identically zero. This property allows operators to detect the presence of non-Gaussian communication signals at very low signal-to-noise ratios (SNR).

• Detects spread spectrum signals below the noise floor
• Identifies transient turn-on signatures for emitter classification
• Applied in spectrum surveillance and cognitive radio sensing

When a signal passes through a non-linear system, harmonics interact to produce quadratic phase coupling (QPC) — a phenomenon where the phase of a frequency component at f1+f2 equals the sum of the phases at f1 and f2. The bispectrum is the definitive tool for detecting and quantifying QPC.

• Identifies harmonic distortion products from saturated amplifiers
• Diagnoses non-linear mechanical faults in rotating machinery via vibration analysis
• Used in EEG analysis to study non-linear brain wave interactions during epileptic seizures

Bispectrum-derived features serve as robust inputs to deep learning classifiers for automatic modulation classification. Unlike raw IQ samples or basic spectral features, bispectrum representations are translation-invariant and suppress Gaussian noise, making them ideal for CNNs and few-shot learning architectures.

• Generates 2D bispectrum images for ResNet-style classifiers
• Distinguishes modulation families with identical power spectra (e.g., QPSK vs. OQPSK)
• Enables open set recognition by clustering unknown signal types in bispectral space

Malicious modifications to integrated circuits introduce subtle, unintended non-linearities in the chip's electromagnetic emissions. Bispectrum analysis of unintentional radiated emissions reveals these anomalies by detecting statistically significant deviations from the expected bispectral signature of a genuine chip.

• Compares bispectral templates of golden reference chips against units under test
• Detects dormant Trojans that evade functional testing
• Applied in supply chain authentication for critical infrastructure and defense systems

Bispectrum analysis quantifies non-linear interactions in physiological signals that linear methods like the power spectrum miss entirely. It reveals phase coupling between different frequency bands of the EEG, which is a hallmark of specific neurological states and pathologies.

• Identifies ictal EEG patterns for seizure localization
• Analyzes heart rate variability (HRV) to assess autonomic nervous system coupling
• Detects sleep spindles and K-complexes via their non-linear phase relationships