Bispectrum

The bispectrum of a Gaussian process is identically zero. This property provides a powerful theoretical foundation for emitter identification, as it allows the analysis to completely reject additive white Gaussian noise in expectation. In practice, this means the bispectrum isolates the non-Gaussian, deterministic signal components generated by transmitter hardware impairments, dramatically improving the signal-to-noise ratio for feature extraction in low-SNR environments.

Unlike the power spectrum, which discards all phase information, the bispectrum is a complex-valued statistic that preserves both magnitude and phase relationships. This is critical for RF fingerprinting because many hardware impairments—such as amplifier non-linearities and I/Q modulator imbalances—manifest as specific phase couplings between harmonically related frequency components. The bispectrum captures these couplings, providing a richer, more discriminative feature set than second-order statistics alone.

The bispectrum is uniquely sensitive to quadratic phase coupling (QPC), a phenomenon where two frequency components, f1 and f2, interact non-linearly to generate a third component at f1+f2 with a phase equal to the sum of the original phases. This is a direct signature of non-linear hardware behavior such as power amplifier compression and mixer intermodulation. Detecting QPC allows a fingerprinting system to identify the specific non-linear transfer function of a transmitter's analog chain.

The bispectrum exhibits 12 distinct symmetry regions in the bifrequency plane, a consequence of its mathematical definition and the assumption of real-valued signals. These symmetries mean that all non-redundant information is contained within a principal triangular domain. Understanding these regions is essential for efficient computation and feature extraction, as it reduces the dimensionality of the bispectral representation without losing any discriminative information about the emitter's hardware signature.

While the full bispectrum is a two-dimensional function, it can be computationally prohibitive for real-time classification. The integrated bispectrum reduces this to a one-dimensional function by integrating along radial slices in the bifrequency plane. This preserves the phase-preserving and noise-suppressing properties while creating a compact, rotation-invariant feature vector suitable as direct input to classifiers like support vector machines or neural networks.

The bicoherence is a normalized version of the bispectrum, constrained to values between 0 and 1. It measures the fraction of signal power at a given bifrequency that is due to quadratic phase coupling. This normalization makes the metric independent of signal amplitude, providing a scale-invariant feature that is robust to variations in transmission power. A bicoherence value near 1 indicates strong, deterministic non-linear coupling—a highly reliable fingerprint indicator.

The bispectrum directly captures quadratic phase coupling generated by non-linear components in the transmitter chain, such as power amplifiers operating near saturation. Unlike the power spectrum, which discards phase, the bispectrum reveals frequency-domain correlations that serve as a unique, unclonable hardware signature.

• Detects harmonics generated by amplifier non-linearity
• Identifies coupled frequency pairs unique to specific DACs
• Suppresses Gaussian thermal noise that obscures subtle impairments

A defining property of the bispectrum is that it is theoretically zero for any Gaussian process. This makes it exceptionally valuable in low-SNR environments where additive white Gaussian noise dominates the received signal. By computing the bispectrum, the noise floor is mathematically suppressed while the non-Gaussian signal of interest—containing the transmitter's fingerprint—is preserved.

• Ideal for long-range or low-power signal interception
• Extracts features buried below the noise floor
• Enables fingerprinting at SNR levels where spectral methods fail

The power spectrum discards all phase information, losing critical identifying characteristics. The bispectrum, as a complex-valued statistic, retains both magnitude and phase relationships between harmonically related frequency components. This phase information encodes manufacturing variances in analog filters, mixers, and oscillators.

• Captures phase distortion from I/Q imbalance
• Preserves oscillator phase noise signatures
• Provides a richer feature space than spectral methods alone

Bispectral features are largely invariant to the modulation format being transmitted. Because the bispectrum characterizes the underlying hardware non-linearity rather than the intentional modulation, the same device can be identified whether it is transmitting QPSK, 16-QAM, or OFDM waveforms. This enables cross-protocol emitter tracking.

• Identifies hardware regardless of waveform type
• Enables tracking across frequency-hopping patterns
• Reduces need for modulation-specific models

Bispectral estimates can be formatted as 2D images and fed directly into Convolutional Neural Networks for automatic feature learning. Alternatively, integrated bispectral features—such as radially integrated bispectrum or axially integrated bispectrum—provide compact, rotation-invariant representations suitable for lightweight classifiers deployed on edge hardware.

• 2D bispectral images as CNN inputs
• Radially integrated bispectrum for rotation invariance
• Compatible with Siamese networks for one-shot verification

Because the bispectrum captures inherent physical non-linearities that are extremely difficult to model or replicate, it provides strong resistance against adversarial spoofing attacks. An attacker would need to precisely replicate not just the linear spectral signature but the specific higher-order phase couplings of the target device—a computationally intractable problem.

• Higher-order couplings are physically unclonable
• Resistant to replay attacks with modified hardware
• Complements cryptographic authentication at the physical layer

A fourth-order spectral cumulant that extends the bispectrum to analyze cubic non-linear coupling between frequency components. While the bispectrum captures quadratic phase coupling, the trispectrum reveals more complex interactions in signals with non-Gaussian distributions. It is particularly effective for characterizing transmitters with strong amplifier non-linearities but requires significantly more computational resources and longer signal observation windows than third-order methods.

The mathematical framework underlying bispectrum estimation. Cumulants of order n > 2 are zero for Gaussian processes, making them ideal for separating a transmitter's unique non-linear signature from thermal noise. Key properties include:

• Third-order cumulant: Measures skewness and quadratic phase coupling
• Fourth-order cumulant: Measures kurtosis and cubic interactions
• Blind source separation: Cumulants enable the isolation of individual emitter signatures from co-channel interference

The general class of higher-order spectra that includes the power spectrum (2nd order), bispectrum (3rd order), and trispectrum (4th order). Polyspectra preserve both magnitude and phase information, unlike the power spectrum which discards phase. This phase preservation is critical for SEI because the non-linear phase coupling patterns are direct manifestations of hardware-specific impairments in mixers and amplifiers that cannot be cloned or spoofed.

A phenomenon where two frequency components, f1 and f2, interact through a non-linear system to generate energy at their sum and difference frequencies with a consistent phase relationship. The bispectrum is the definitive tool for detecting and quantifying this coupling. In RF fingerprinting, quadratic phase coupling arises from mixer intermodulation and amplifier memory effects, creating a unique, device-specific signature that persists across different modulation schemes.

A widely-used MATLAB toolbox implementing algorithms for bispectrum estimation, bicoherence analysis, and cumulant computation. It provides both direct (FFT-based) and indirect (cumulant-based) methods for polyspectral estimation. Key functions include:

• bispecd: Direct bispectrum estimation via the Brillinger-Rosenblatt method
• bispeci: Indirect estimation using windowed cumulants
• bicoher: Bicoherence calculation for normalized coupling measurement

A normalized form of the bispectrum that quantifies the strength of quadratic phase coupling on a scale from 0 to 1, independent of signal amplitude. Bicoherence values near 1 indicate strong, consistent non-linear coupling, while values near 0 suggest random phase relationships. This normalization makes bicoherence features robust to varying transmit power levels, a critical advantage for operational SEI systems where signal strength fluctuates with distance and channel conditions.