Cyclic Correntropy

Cyclic correntropy maps signal samples into a high-dimensional Reproducing Kernel Hilbert Space (RKHS) via a positive-definite kernel function, typically a Gaussian kernel. This nonlinear transformation enables the capture of higher-order statistical relationships that linear correlation methods miss. The kernel bandwidth parameter controls the sensitivity to outliers, with smaller bandwidths providing greater robustness against impulsive noise.

• Defined as the expectation of the kernelized product between time-shifted and frequency-shifted signal versions
• Gaussian kernel: κ(x,y) = exp(-|x-y|²/(2σ²))
• Kernel bandwidth σ acts as a tunable robustness parameter

Unlike conventional correntropy, cyclic correntropy is parameterized by a cyclic frequency α, allowing it to isolate and measure similarity at specific periodicities inherent in modulated signals. This selectivity makes it a powerful tool for blind modulation classification, as different modulation schemes exhibit unique cyclic correntropy signatures at distinct cycle frequencies.

• Extracted at discrete α values corresponding to symbol rate, carrier frequency offsets, and their harmonics
• Zero cyclic frequency reduces to standard correntropy
• Enables separation of overlapping signals with different cyclic frequencies

A defining advantage of cyclic correntropy is its inherent robustness to heavy-tailed, non-Gaussian noise environments such as atmospheric impulsive noise, man-made electromagnetic interference, and co-channel interference. The Gaussian kernel effectively saturates on large-amplitude outliers, bounding their influence on the similarity measure.

• Outperforms cyclic correlation and cyclic cumulants in α-stable noise and Middleton Class A interference
• Kernel bandwidth controls the outlier rejection threshold
• Maintains reliable cyclic feature extraction at low signal-to-noise ratios where second-order methods fail

The Cyclic Correntropy Spectrum (CCES) is the frequency-domain representation obtained by taking the Fourier transform of the cyclic correntropy function with respect to the time lag τ. It reveals the density of spectral correlation in the kernel space, producing a two-dimensional map of spectral frequency f versus cyclic frequency α.

• Analogous to the Spectral Correlation Function but computed in the kernel space
• Provides enhanced resolution of cyclostationary features in impulsive noise
• Can be estimated efficiently using kernelized versions of the FAM or SSCA algorithms

Each digital modulation format produces a distinct pattern of peaks in the cyclic correntropy domain, forming a unique cyclic signature. These signatures are more discriminative than those from linear cyclostationary methods because they capture both the periodic structure and the amplitude distribution of the signal constellation.

• QPSK exhibits strong cyclic correntropy at α = 2fc and α = 2fc ± Rsym
• 16-QAM shows additional features at sub-harmonics due to multi-level amplitude variations
• GMSK and CPM signals produce continuous cyclic features rather than discrete peaks
• Feature vectors extracted from CCES serve as robust inputs to neural network classifiers

Cyclic correntropy enables the design of a kernelized cyclic matched filter for optimal detection of cyclostationary signals in impulsive noise. This nonlinear detector correlates the received signal with a frequency-shifted reference in the RKHS, maximizing the signal-to-noise ratio when noise follows a heavy-tailed distribution.

• Generalizes the classical cyclic matched filter to non-Gaussian environments
• Implemented implicitly via the kernel trick without explicit RKHS computation
• Particularly effective for spread spectrum signal detection and RF fingerprinting in tactical environments

A higher-order statistic (HOS) that extends cumulants to cyclostationary signals. Like cyclic correntropy, cyclic cumulants provide robustness against Gaussian noise (colored or white) because Gaussian cumulants of order greater than two are identically zero. The distinction lies in their response to impulsive noise:

• Cyclic Cumulants: Can be highly sensitive to impulses, as higher-order moments amplify outlier effects.
• Cyclic Correntropy: Uses a kernel function to bound the influence of outliers, offering a fundamentally different and more robust approach to non-Gaussian, heavy-tailed noise environments.

A one-dimensional slice of the cyclic correntropy spectrum at a fixed spectral frequency f. It plots the magnitude of cyclic correntropy across all cyclic frequencies (α). This profile serves as a compact, highly discriminative feature vector for automatic modulation classification. Peaks in the alpha profile directly correspond to the hidden periodicities of a signal, such as its symbol rate and carrier frequency offset, but extracted with a non-linear kernel to suppress impulsive interference.

A Frequency-Shift (FRESH) filtering technique that exploits signal cyclostationarity for interference mitigation. A FRESH filter is a linear periodically time-varying (LPTV) system that processes multiple frequency-shifted versions of the input. By replacing the standard linear correlation in the filter's design with cyclic correntropy, a non-linear FRESH filter can be developed. This allows for the separation of overlapping signals in the frequency domain even when one or both are corrupted by highly impulsive noise, a task where linear FRESH filters fail.

A statistical hypothesis test formulated in the frequency domain to detect the presence of cyclostationarity at a specific cyclic frequency. The test provides a chi-squared statistic to determine if an estimated cyclic feature is statistically significant or just a random fluctuation. A cyclic correntropy-based extension of this test would provide a robust detector for signal presence in the harsh, impulsive noise environments typical of powerline communications and underwater acoustics, where the standard test's Gaussian assumption is violated.