Cyclic Correntropy

Unlike the cyclic autocorrelation function (CAF), which relies on second-order statistics and is highly sensitive to outliers, cyclic correntropy applies a Gaussian kernel function to suppress the influence of large-amplitude impulsive samples. This makes it exceptionally robust in environments with alpha-stable or Laplace-distributed noise, such as atmospheric interference or man-made electromagnetic pulses.

• Mechanism: Maps input samples into a high-dimensional kernel space where inner products correspond to a non-linear similarity metric.
• Advantage: Maintains reliable cyclostationary feature extraction even when conventional methods fail due to heavy-tailed noise distributions.

Cyclic correntropy generalizes the concept of cyclostationarity by measuring the kernelized similarity between a signal and its frequency-shifted version. It reduces to the traditional cyclic autocorrelation when the kernel function is linear, but provides a richer, higher-order statistical characterization in the non-linear case.

• Cyclic Correntropy Function: Defined as the Fourier coefficient of the time-varying correntropy function, revealing periodicities in the signal's probability density structure.
• Information-Theoretic Link: Directly relates to the Renyi entropy of the signal, capturing information content beyond simple power distribution.

The kernel size parameter in cyclic correntropy acts as a tunable filter that can be optimized to maximize the detection of specific cyclic frequencies while suppressing broadband impulsive noise. This enables blind symbol rate estimation and carrier offset recovery in highly degraded signal environments.

• Application: Automatically detecting the baud rate of an unknown emitter in electronic warfare or spectrum monitoring scenarios.
• Method: Sweeping the cyclic frequency domain and computing the correntropy spectral density to identify peaks corresponding to the symbol rate and its harmonics.

Different digital modulation schemes exhibit unique higher-order cyclostationary signatures that are amplified by the non-linear kernel mapping of cyclic correntropy. This property enables robust automatic modulation classification (AMC) without requiring prior synchronization or timing recovery.

• BPSK/QPSK: Strong cyclic correntropy peaks at twice the carrier frequency plus the symbol rate.
• 16-QAM/64-QAM: Distinct cyclic patterns emerge at multiples of the symbol rate due to the multi-level amplitude structure.
• FSK Signals: Cyclic correntropy reveals periodicities at the frequency deviation spacing.

Practical estimation of cyclic correntropy is performed using a sliding window kernel estimator that computes the sample mean of the kernelized product between the original and frequency-shifted signal segments. Efficient implementations leverage the FFT accumulation method (FAM) adapted for non-linear kernel operations.

• Kernel Size Selection: The Gaussian kernel bandwidth controls the trade-off between outlier suppression and signal fidelity.
• Complexity: Comparable to cyclic autocorrelation estimation but with an added kernel evaluation step per sample pair.

Cyclic correntropy features serve as highly discriminative inputs for deep neural network classifiers in RF fingerprinting applications. The kernelized representation naturally suppresses noise while preserving device-specific hardware impairment signatures that manifest as subtle cyclostationary distortions.

• Feature Vector Construction: Sampling the cyclic correntropy function at key cyclic frequencies (symbol rate, carrier offset) creates a compact, robust fingerprint.
• Adversarial Robustness: The non-linear kernel mapping provides inherent resistance to low-power jamming or spoofing attempts that target linear statistical features.

The fundamental building block of cyclic correntropy. It is a localized similarity measure that maps the input signal into a high-dimensional kernel space. Unlike standard correlation, which is a second-order moment, correntropy incorporates higher-order moments of the signal's probability density function. This allows it to capture non-linear relationships and provides inherent robustness to large outliers or impulsive noise that would otherwise dominate a conventional autocorrelation calculation.

The mathematical mechanism that enables correntropy's non-linearity. A kernel function (typically a Gaussian kernel) computes the inner product between data points in a transformed feature space without explicitly calculating the coordinates. Key properties include:

• Mercer's Theorem: Guarantees the kernel corresponds to a valid inner product space.
• Bandwidth Selection: The kernel size controls the outlier rejection threshold; a smaller bandwidth makes the measure more robust to impulsive noise but can discard useful signal structure.

The conventional linear method for analyzing cyclostationary signals. It measures the density of correlation between frequency-shifted versions of a signal. While optimal for Gaussian noise, its performance degrades catastrophically in environments with heavy-tailed, non-Gaussian interference. Cyclic correntropy directly addresses this limitation by extending the analysis into a kernel space, providing a non-linear alternative that suppresses impulsive noise while preserving the cyclic periodicities of the signal of interest.

The primary engineering advantage of cyclic correntropy. In real-world environments like power lines, industrial machinery, or atmospheric discharges, noise often exhibits an impulsive alpha-stable distribution. Standard second-order statistics diverge in this context. Cyclic correntropy acts as an adaptive filter that effectively zeros out large amplitude spikes. This ensures that the extracted cyclostationary features—such as symbol rate or carrier offset—remain stable and detectable even when the signal-to-noise ratio is dominated by rare, high-energy bursts.

A representation that displays the cyclic frequency content of a signal's envelope or instantaneous frequency. When computed using correntropy instead of linear correlation, the resulting non-linear cyclic modulation spectrum reveals modulation-specific periodicities that are invisible to conventional methods in impulsive noise. This is critical for automatic modulation classification tasks where the goal is to identify the transmission scheme of an unknown emitter without prior knowledge of the noise distribution.

The practical output of a cyclic correntropy analysis. By sampling the cyclic correntropy function at key cyclic frequencies (alpha)—such as the symbol rate and twice the carrier offset—engineers construct a compact feature vector. This vector serves as a highly robust input to machine learning classifiers for RF fingerprinting. Its insensitivity to impulsive noise makes it a superior alternative to features derived from the spectral correlation function for authenticating devices in harsh electromagnetic environments.