Skewness

Skewness (γ₁) is defined as the expected value of the cubed standardized deviation: γ₁ = E[((X - μ)/σ)³]. For a set of N complex IQ samples, the sample skewness is computed as the normalized sum of cubed deviations from the mean.

• Zero skewness: Indicates a perfectly symmetric distribution (e.g., QPSK, 16QAM)
• Positive skewness: The right tail is longer; mass concentrates on the left
• Negative skewness: The left tail is longer; mass concentrates on the right
• Complex-valued extension: Skewness is computed separately on I and Q rails or as a joint complex statistic

Pulse Amplitude Modulation (PAM) schemes exhibit non-zero skewness because their one-dimensional constellation points are not symmetrically distributed around the origin. This property enables rapid separation from symmetric modulations.

• 4-PAM: Exhibits measurable skewness due to asymmetric amplitude levels
• 8-PAM: Stronger skewness signature with more distinct amplitude states
• QPSK/16QAM: Skewness approaches zero due to quadrant symmetry
• Practical threshold: |γ₁| > 0.15 often indicates a PAM signal in moderate SNR conditions

Hardware imperfections in direct-conversion receivers create IQ imbalance, which manifests as a skewed circular constellation. Skewness serves as a sensitive diagnostic metric for this impairment.

• Gain imbalance: Differential amplification between I and Q rails shifts the distribution asymmetrically
• Phase imbalance: Non-orthogonal mixing introduces correlation that alters skewness
• Pre-classification correction: Skewness thresholds trigger compensation algorithms before modulation identification
• Monitoring application: Drift in skewness over time indicates degrading receiver hardware

As a higher-order statistic, skewness estimation is more sensitive to noise than variance. The variance of the skewness estimator grows with noise power, requiring sufficient samples for reliable estimation.

• Sample complexity: O(10³) to O(10⁴) samples needed for stable estimates at low SNR
• SNR wall: Below approximately 5-8 dB, skewness becomes unreliable for PAM detection
• Robustness advantage: Skewness is theoretically zero for Gaussian noise, providing inherent noise rejection
• Complementary use: Often paired with kurtosis for hierarchical classification decisions

Skewness is typically deployed at the first decision node of a hierarchical cumulant classifier to partition the modulation candidate set into symmetric and asymmetric subsets.

• Node 1: If |skewness| > threshold → candidate set = {PAM, ASK variants}
• Node 1 (else): Candidate set = {PSK, QAM, circular modulations}
• Computational efficiency: Reduces downstream processing by 50% immediately
• Feature vector inclusion: Skewness is concatenated with C40, C42, and kurtosis for neural network classifiers

Skewness estimation can be implemented in streaming hardware using recursive moment accumulators, avoiding batch processing latency.

• Online algorithm: Update running sums of (x - μ)² and (x - μ)³ with each new sample
• FPGA resource usage: Requires only a few multiply-accumulate units and dividers
• Latency: Sub-microsecond feature extraction on modern FPGAs
• Sliding window: Exponentially weighted moving average adapts to non-stationary channels

Pulse Amplitude Modulation (PAM) constellations exhibit inherent amplitude asymmetry along a single axis, producing non-zero skewness values. In contrast, square QAM constellations are symmetric about both I and Q axes, yielding skewness near zero. This property enables a binary decision boundary in hierarchical classifiers: compute the sample skewness of the received IQ samples; if it exceeds a threshold, classify as PAM, otherwise proceed to QAM/PSK sub-classification. This approach is computationally lightweight and robust to phase offsets.

Hardware imperfections in direct-conversion receivers create gain and phase mismatches between the I and Q branches. This imbalance distorts the received constellation, introducing measurable skewness in the signal distribution even when the transmitted modulation is perfectly symmetric. By monitoring the skewness of the baseband IQ samples, an operational system can:

• Detect receiver front-end degradation before it causes bit errors
• Trigger automatic digital compensation algorithms
• Flag hardware for preventive maintenance in SIGINT collection platforms

Power amplifier non-linearities near saturation compress signal peaks asymmetrically, introducing skewness into the transmitted waveform. This effect is particularly pronounced in OFDM signals with high peak-to-average power ratios. By analyzing the skewness of the received signal's amplitude distribution, an electronic warfare system can:

• Identify specific transmitter hardware through its unique non-linear fingerprint
• Estimate the operating point of an adversary's power amplifier
• Distinguish between intentional modulation and unintentional distortion artifacts

Skewness is rarely used in isolation. In modern cumulant-based feature vectors, it serves as the third-order component alongside variance (2nd order) and kurtosis (4th order). This multi-order statistical profile creates a robust fingerprint for modulation identification:

• Skewness separates asymmetric constellations (PAM, APSK)
• Kurtosis separates sub-Gaussian (PSK) from super-Gaussian modulations
• Variance normalizes against signal power variations Together, these features feed into classifiers like Support Vector Machines or lightweight neural networks for real-time AMC.

In contested electromagnetic environments, adversaries may inject spoofed signals that mimic legitimate modulation formats. However, generating a signal with precisely controlled higher-order statistics is computationally difficult. By establishing a baseline skewness profile for authorized transmitters, a monitoring system can:

• Detect anomalous skewness deviations in real-time
• Reject spoofed signals before they enter the decision chain
• Trigger alerts for electronic attack attempts This technique exploits the fact that skewness is a non-linear statistic that is hard for an adversary to precisely replicate without access to the original transmitter hardware.

Before modulation classification can occur, the received constellation must be recovered from the raw IQ stream. Skewness-based algorithms assist in blind equalization by:

• Identifying the correct sampling phase by maximizing distribution asymmetry for PAM signals
• Correcting phase rotations by aligning the skewness vector along the expected axis
• Validating convergence of Constant Modulus Algorithm (CMA) equalizers by monitoring the stabilization of the output skewness value This enables non-cooperative signal processing without pilot tones or training sequences.