Kurtosis

Kurtosis is the standardized fourth central moment of a probability distribution, defined as:

K = E[(X - μ)⁴] / (E[(X - μ)²])²

• For a real-valued signal, the excess kurtosis is K - 3, where 3 is the kurtosis of a Gaussian distribution
• For complex baseband signals, the normalized fourth-order cumulant C40 is often used as a kurtosis analog
• Measures the tailedness and peakedness of the amplitude distribution relative to a Gaussian
• A mesokurtic distribution (K = 3) matches Gaussian behavior, typical of OFDM signals with many subcarriers
• The theoretical kurtosis for BPSK is 1, for QPSK is 1, and for 16-QAM is approximately 1.32

Kurtosis provides a binary classification boundary for modulation recognition by partitioning schemes based on their distribution shape:

• Sub-Gaussian (Platykurtic, K < 3): Modulations with bounded, uniform-like amplitude distributions
  • PSK family (BPSK, QPSK, 8-PSK) exhibits constant envelope, yielding kurtosis values below 3
  • The flat phase distribution contributes to thinner tails than Gaussian
• Super-Gaussian (Leptokurtic, K > 3): Modulations with heavy-tailed amplitude distributions
  • Higher-order QAM (16-QAM, 64-QAM) shows amplitude variability producing heavier tails
  • The multi-level amplitude structure creates infrequent but large deviations from the mean
• This hierarchical partitioning is often the first node in a cumulant-based decision tree classifier
• The separation is robust to phase and frequency offsets since kurtosis operates on magnitude information

In practical modulation classifiers, kurtosis is estimated from a finite block of N received IQ samples:

• The sample kurtosis estimator is: K̂ = (1/N) Σ|x[n]|⁴ / ((1/N) Σ|x[n]|²)²
• This is a biased estimator for small sample sizes; bias correction factors are applied for N < 1000
• The variance of the estimator decreases with √N, requiring sufficient samples for reliable classification
• At low SNR, the estimator variance increases, defining a cumulant SNR wall below which classification becomes unreliable
• Recursive/online estimators enable streaming kurtosis computation without storing the entire sample buffer
• The complex baseband kurtosis uses |x[n]| rather than x[n] to achieve phase invariance
• Typical implementation requires 1000-5000 samples for stable estimates at moderate SNR (>10 dB)

Kurtosis serves as the root node feature in hierarchical cumulant classifiers, enabling coarse modulation set partitioning:

• Step 1: Compute sample kurtosis from received IQ block
• Step 2: Compare against Gaussian threshold (K = 3)
  • If K < 3: Signal is sub-Gaussian → candidate set = {BPSK, QPSK, 8-PSK, CPM}
  • If K > 3: Signal is super-Gaussian → candidate set = {16-QAM, 64-QAM, 256-QAM, APSK}
  • If K ≈ 3: Signal may be Gaussian → candidate set = {OFDM, noise, spread spectrum}
• Step 3: Apply higher-order cumulants (C40, C42, C63) within each branch for fine-grained identification
• This divide-and-conquer strategy reduces computational complexity from O(M) to O(log M) for M modulation candidates
• The kurtosis threshold can be adaptive, adjusting based on estimated SNR to maintain constant false-alarm rate

Kurtosis exhibits inherent robustness to several common channel impairments, making it a preferred feature for non-cooperative classification:

• Phase offset invariance: Since kurtosis operates on |x[n]|, arbitrary carrier phase rotations do not affect the estimate
• Frequency offset tolerance: Slow frequency drift does not alter the amplitude distribution shape over short observation windows
• Gaussian noise resilience: The fourth-order cumulant formulation theoretically suppresses Gaussian noise, as Gaussian moments above second order are zero
  • In practice, noise increases estimator variance but does not bias the mean at moderate SNR
• Scale invariance: Normalization by signal power makes kurtosis independent of received signal amplitude
• Limitations:
  • Multipath fading distorts the amplitude distribution, requiring equalization before kurtosis estimation
  • Impulsive noise (non-Gaussian) introduces bias, as it directly affects higher-order moments
  • Co-channel interference mixes distributions, potentially shifting kurtosis across the Gaussian boundary

Beyond modulation classification, kurtosis functions as a statistical hypothesis test for non-Gaussianity:

• Null Hypothesis H₀: The signal samples are drawn from a Gaussian distribution (K = 3)
• Alternative Hypothesis H₁: The signal exhibits non-Gaussian behavior (K ≠ 3)
• The test statistic: Z = (K̂ - 3) / σ_K̂ where σ_K̂ is the standard error of the kurtosis estimate under Gaussianity
• For large N, Z follows a standard normal distribution under H₀
• Applications in modulation recognition:
  • OFDM detection: OFDM signals with many subcarriers approach Gaussianity by the Central Limit Theorem, yielding K ≈ 3
  • Noise-only channel detection: Distinguishes between idle spectrum (Gaussian noise) and active transmissions
  • Spread spectrum detection: Direct-sequence signals with long spreading codes approximate Gaussian statistics
• The p-value from this test provides a confidence metric for the sub/super-Gaussian classification decision