Cumulant Invariant

Higher-order cumulants of order k naturally absorb or cancel carrier phase offsets (θ) that plague likelihood-based classifiers. For a complex signal x, the transformation x → x·e^(jθ) leaves the magnitude of many cumulants unchanged.

• C40/C42 Invariance: The fourth-order cumulant ratio |C40|/|C42| is inherently phase-blind because the phase term e^(j4θ) cancels in the magnitude operation
• Conjugate Pairing: Cumulants defined with p conjugated and q non-conjugated terms (Cpq) can be tuned so that the net phase exponent sums to zero
• Practical Impact: Eliminates the need for a phase-locked loop or carrier recovery circuit before the classifier, enabling true blind identification on raw IQ streams

Normalized cumulants are constructed to be independent of the received signal power, a critical property when automatic gain control is absent or the receiver operates over a wide dynamic range.

• Variance Normalization: Dividing a k-th order cumulant by the k-th power of the signal variance (σ²) yields a scale-invariant statistic, e.g., C40 / (σ²)²
• Ratio Invariance: Forming ratios like |C63| / |C42|^(3/2) cancels amplitude dependence entirely, as both numerator and denominator scale identically with signal power
• Practical Impact: The same classifier threshold works whether the signal is at -80 dBm or -20 dBm, eliminating the need for continuous recalibration in dynamic spectrum environments

Cumulants of stationary processes are, by definition, independent of the time origin. A shift t → t + τ does not alter the cumulant value, making them ideal for asynchronous interception scenarios.

• Stationarity Property: For a wide-sense stationary signal, the joint cumulant depends only on relative time lags, not absolute time, so a fixed observation window yields consistent estimates regardless of when sampling begins
• Symbol Timing Robustness: Unlike constellation-based methods that require precise symbol synchronization, cumulant estimates computed on oversampled IQ data remain stable even with an arbitrary sampling offset
• Practical Impact: Eliminates the need for a symbol timing recovery loop before feature extraction, significantly reducing the preprocessing pipeline latency in real-time electronic warfare receivers

All cumulants of order k ≥ 3 are identically zero for Gaussian processes. This property provides a built-in theoretical shield against additive white Gaussian noise (AWGN), the dominant impairment in wireless channels.

• Theoretical Insensitivity: The k-th order cumulant of a signal-plus-noise mixture equals the cumulant of the signal alone when the noise is Gaussian, because the Gaussian contribution vanishes for k ≥ 3
• Finite-Sample Robustness: While sample estimates exhibit variance due to finite data, the mean of the estimator converges to the noise-free theoretical value as the observation length increases
• Practical Impact: Cumulant-based classifiers maintain high accuracy at low SNR where raw IQ or constellation-based methods degrade rapidly, making them preferred for SIGINT and spectrum monitoring applications operating near the noise floor

Each ideal modulation format maps to a unique, deterministic set of cumulant values, creating a lookup table for classification without requiring training data.

• BPSK: C40 = -2.0, C42 = -2.0, |C40|/|C42| = 1.0
• QPSK: C40 = 1.0, C42 = -1.0, |C40|/|C42| = 1.0
• 16-QAM: C40 = -0.68, C42 = -0.68, |C40|/|C42| = 1.0
• 64-QAM: C40 = -0.619, C42 = -0.619, |C40|/|C42| = 1.0
• Practical Impact: A hierarchical decision tree can separate BPSK-like (C42 < 0) from QPSK-like (C42 > 0) modulations, then use C40 magnitude to distinguish QAM orders—all from analytically derived thresholds with zero training overhead

The cumulant of a sum of statistically independent signals equals the sum of their individual cumulants. This linear decomposition property enables blind source separation in co-channel interference scenarios.

• Source Separation: In a multi-antenna array receiving M mixed signals, the joint fourth-order cumulant tensor can be jointly diagonalized to recover each source's modulation fingerprint independently
• JADE Algorithm: The Joint Approximate Diagonalization of Eigenmatrices exploits this additivity to separate mixed communication signals without training data or channel state information
• Practical Impact: Enables modulation classification of individual spatial streams in MIMO systems or identification of overlapping emitters in dense electromagnetic environments where traditional single-signal classifiers fail

The most direct form of a scale-invariant cumulant, obtained by dividing a higher-order cumulant by a power of the signal's variance. This normalization ensures the feature is independent of the received signal amplitude, making it a fundamental building block for robust classification. For example, normalizing the fourth-order cumulant by the square of the variance yields a constant value for a given modulation type regardless of signal power.

A discriminative feature formed by dividing two different cumulant orders, such as |C40|/|C42|. This ratio creates a modulation fingerprint that is inherently robust to both phase and frequency offsets, as these nuisance parameters cancel out in the division. It is a primary invariant used in hierarchical classifiers to separate QAM, PSK, and ASK constellations.

A cumulant function that exploits the periodicity in a signal's statistics to provide modulation features robust to stationary noise and interference. By analyzing the cycle frequency domain, cyclic cumulants can separate overlapping signals and identify modulation types even in highly congested spectrum environments where traditional cumulants fail.

A structured set of estimated cumulants and their ratios concatenated into a single input vector for a machine learning classifier. This vector typically includes:

• Normalized fourth-order cumulants (C40, C42)
• Cumulant ratios for phase invariance
• Higher-order statistics (sixth-order) for fine-grained discrimination This physics-informed input dramatically reduces the dimensionality required by deep learning models.

A decision tree architecture that uses specific cumulant thresholds at each node to partition the modulation candidate set. The process is:

• Root Node: Separate QAM vs. PSK using Gaussianity tests
• Intermediate Nodes: Refine by order (e.g., QPSK vs. 8PSK) using higher-order invariants
• Leaf Nodes: Final modulation decision This approach minimizes computational complexity by avoiding exhaustive multi-class comparisons.

The inherent resistance of cumulant features to small, adversarial perturbations. Unlike raw IQ samples, which can be easily manipulated by subtle waveform distortions, higher-order statistics are less sensitive to minor changes. This property makes cumulant invariants a preferred feature set for security-critical applications like electronic warfare and spectrum monitoring, where evasion attacks are a concern.