Fourth-Order Cumulant (C40/C42)

C40 is the fourth-order cumulant at zero lag:

• C40 = cum(x, x, x, x) = E[x⁴] - 3(E[x²])²
• C42 is the two-lag variant:
• C42 = cum(x, x, x*, x*) = E[|x|⁴] - |E[x²]|² - 2(E[|x|²])²

These capture the kurtosis of the signal distribution, measuring the weight of the tails relative to a Gaussian.

Raw cumulants depend on signal power. To achieve amplitude independence, normalize by the squared variance:

• C40_norm = C40 / (σ²)²
• C42_norm = C42 / (σ²)²

This ensures the feature is robust to gain variations in the receiver chain and varying path loss, making it ideal for non-cooperative classification.

Normalized C40 takes distinct theoretical values:

• BPSK: C40 = -2.0
• QPSK: C40 = -1.0
• 8-PSK: C40 = 0.0
• 16-QAM: C40 = -0.68
• 64-QAM: C40 = -0.619
• Gaussian Noise: C40 = 0.0

These values form a one-dimensional decision space for hierarchical classification.

By the Central Limit Theorem, the sum of many independent signals tends toward a Gaussian distribution. C40/C42 exploit this:

• Gaussian signals (OFDM, noise): C40 ≈ 0
• Sub-Gaussian signals (PSK, QAM): C40 < 0
• Super-Gaussian signals: C40 > 0

This property enables blind separation of OFDM from single-carrier modulations without prior knowledge.

C40 and C42 are phase-invariant cumulants. A constant phase rotation φ multiplies x by e^(jφ), but:

• C40 involves x·x·x·x, accumulating e^(j4φ)
• C42 involves x·x·x*·x*, canceling to e^(j0φ) = 1

This makes C42 fully invariant to carrier phase offset, while C40 requires absolute value |C40| for rotation robustness.

In practice, cumulants are estimated from N received IQ samples:

• C40_hat = (1/N)Σ x[n]⁴ - 3[(1/N)Σ x[n]²]²
• C42_hat = (1/N)Σ |x[n]|⁴ - |(1/N)Σ x[n]²|² - 2[(1/N)Σ |x[n]|²]²

Estimation variance decreases with √N. For reliable classification at low SNR, N must exceed the cumulant SNR wall threshold.

A scale-invariant cumulant value obtained by dividing a higher-order cumulant by a power of the signal variance.

• Ensures classification features are independent of received signal amplitude
• Critical for real-world deployment where automatic gain control varies
• C40 and C42 are typically normalized by σ⁴ (variance squared)

A discriminative feature formed by dividing two different cumulant orders, such as |C40|/|C42|.

• Creates a modulation fingerprint robust to phase and frequency offsets
• Eliminates dependence on absolute signal power
• Enables hierarchical separation of QAM, PSK, and ASK families

A statistical hypothesis test using sample cumulants to determine if a signal's distribution deviates from Gaussian.

• C40 and C42 are zero for Gaussian processes, enabling threshold-based detection
• Separates linear modulations from Gaussian noise or OFDM signals
• Forms the first decision node in hierarchical cumulant classifiers

An empirical estimate of the theoretical cumulant computed from a finite block of received IQ samples.

• Practical input feature for real-time modulation classifiers
• Estimation variance decreases with observation length
• Subject to the cumulant SNR wall—a theoretical threshold below which classification becomes unreliable

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

• Starts with coarse PSK/QAM separation using C42
• Refines to specific orders using C40 and higher-order cumulants
• Computationally efficient compared to exhaustive multi-class approaches