Kurtosis

Excess kurtosis (kurtosis minus 3) directly measures deviation from a Gaussian distribution. In RF fingerprinting, a non-zero excess kurtosis indicates the presence of transmitter-specific non-linearities.

• Leptokurtic (excess > 0): Heavy tails caused by amplifier saturation spikes or rare high-amplitude transients.
• Platykurtic (excess < 0): Thin tails typical of clipped or compressed signals from power amplifier compression.
• Mesokurtic (excess = 0): Gaussian distribution, suggesting no exploitable hardware signature is present in the amplitude domain.

Kurtosis is the fourth standardized moment of a signal's amplitude distribution. For a zero-mean random variable X, it is defined as:

Kurt[X] = E[X⁴] / (E[X²])²

• E[X⁴] is the fourth moment, highly sensitive to outlier amplitudes.
• E[X²] is the variance, used for normalization.
• The subtraction of 3 yields excess kurtosis, normalizing the Gaussian distribution to zero.
• This normalization makes kurtosis a dimensionless, scale-invariant feature robust to varying signal power levels.

Power amplifiers (PAs) are a primary source of kurtosis-based fingerprints. When driven near saturation, PAs exhibit non-linear gain compression.

• 1 dB Compression Point (P1dB): The output power where gain drops by 1 dB. Operation beyond this point generates heavy-tailed amplitude distributions.
• AM/AM Distortion: Amplitude-dependent amplitude distortion creates specific kurtosis signatures unique to each PA's transfer curve.
• AM/PM Distortion: Amplitude-dependent phase shifts also manifest in the complex signal's kurtosis, providing a multi-dimensional fingerprint.

A key advantage of kurtosis in RF environments is its theoretical insensitivity to additive white Gaussian noise (AWGN).

• The kurtosis of a Gaussian process is exactly 3 (excess = 0).
• When a non-Gaussian signal is combined with AWGN, the mixture's kurtosis is a weighted combination. The signal's non-Gaussian signature persists.
• This property allows kurtosis-based features to survive below the noise floor, extracting hardware fingerprints even at low signal-to-noise ratios (SNR) where power spectrum analysis fails.

While both are higher-order moments, they capture distinct hardware impairments.

• Skewness (3rd moment): Measures asymmetry. Detects directional biases like even-order harmonic distortion or DC offset in I/Q modulators.
• Kurtosis (4th moment): Measures tailedness. Detects symmetric non-linearities like odd-order harmonic distortion and PA compression.
• Together, they form a complementary pair for characterizing the full non-Gaussian signature of a transmitter's analog front-end.

Kurtosis is rarely used in isolation. It is integrated into higher-order statistical feature vectors for machine learning classifiers.

• Per-band kurtosis: Compute kurtosis on individual sub-bands after filter bank analysis to capture frequency-dependent non-linearities.
• Time-slice kurtosis: Track kurtosis over short time windows to capture transient behavior during PA turn-on sequences.
• Joint cumulant vectors: Combine kurtosis with skewness, variance, and higher-order cumulants (5th, 6th order) to create a comprehensive statistical fingerprint for support vector machines (SVMs) or neural network classifiers.

The third standardized moment quantifying the asymmetry of a signal's amplitude distribution. While kurtosis measures tail weight, skewness detects directional bias.

• Positive skew: Right tail extends further (e.g., amplifier saturation in positive quadrant)
• Negative skew: Left tail extends further (e.g., clipping in negative quadrant)
• Zero skew: Symmetric distribution (ideal Gaussian)

Together, skewness and kurtosis form a compact non-Gaussian signature that reveals specific hardware impairment types.

A statistical hypothesis test that determines whether a signal's amplitude distribution deviates from the Gaussian (normal) model. Kurtosis is a primary test statistic.

• Jarque-Bera test: Combines skewness and kurtosis into a single omnibus statistic
• D'Agostino's K-squared: Decomposes into separate skewness and kurtosis components
• Excess kurtosis ≠ 0: Rejects the Gaussian null hypothesis

These tests validate that exploitable non-Gaussian fingerprints exist in a captured signal before committing to higher-order feature extraction pipelines.

Statistical measures beyond second-order variance that quantify deviations from Gaussianity. The fourth-order cumulant is directly related to kurtosis.

• 2nd cumulant: Variance (σ²)
• 3rd cumulant: Related to skewness
• 4th cumulant: κ₄ = μ₄ − 3σ⁴ (zero for Gaussian)

Cumulants are theoretically blind to Gaussian noise, making them ideal for extracting weak hardware fingerprints buried below the noise floor in low-SNR environments.

The systematic examination of signal components that violate the Central Limit Theorem assumption. Kurtosis is the entry point to this analytical framework.

• Central Limit Theorem: Sums of independent random variables tend toward Gaussian distribution
• Hardware impairments introduce deterministic non-linearities that break this assumption
• Excess kurtosis indicates the presence of these exploitable deterministic components

This framework exploits the fact that thermal noise is Gaussian, while transmitter-specific distortion is not, enabling physical-layer device authentication.

A compact statistical fingerprint constructed from estimated higher-order cumulants, including kurtosis-derived measures, that serves as input to machine learning classifiers.

• Feature components: 2nd, 3rd, and 4th-order cumulant estimates per signal segment
• Dimensionality: Typically 4-10 features per observation window
• Classifier compatibility: Works with SVMs, neural networks, and ensemble methods

These vectors provide computationally efficient yet highly discriminative representations for real-time emitter identification on edge hardware.

A higher-order statistical function that captures both cyclostationary periodicity and non-Gaussian distribution simultaneously. Extends kurtosis into the cyclic domain.

• Cyclic frequency (α): Related to symbol rate, carrier frequency, and guard intervals
• Order (n): 4th-order cyclic cumulant incorporates kurtosis-like tail sensitivity
• Dual robustness: Features persist through both stationary noise and time-varying channels

Cyclic cumulants provide modulation-specific fingerprints that remain stable even when simple kurtosis estimates are corrupted by multipath fading.