Signal Processing

Cyclostationary Detection

Q: How much sensitivity advantage does cyclostationary detection have over energy detection?

The decisive advantage is not raw sensitivity at a single observation length but immunity to noise-power uncertainty. Energy detection suffers from an SNR wall: when the noise variance is known to within only a few tenths of a dB, no energy detector can reliably distinguish signal from noise below roughly -3 to 0 dB SNR regardless of how long it integrates. Because thermal noise is wide-sense stationary it exhibits no spectral correlation at any nonzero cyclic frequency, so a cyclostationary detector evaluated at a known cyclic frequency alpha is theoretically immune to that wall and can keep gaining about 1.5 dB of detectable SNR for every doubling of the observation interval. In practice cyclostationary detectors operate down to -15 to -20 dB SNR for known signals, at the cost of roughly 10 to 100 times more computation than energy detection.

Q: What cyclic frequencies should I search for a given modulation?

The cyclic frequencies are set by the deterministic periodicities the modulation imposes. Every linearly modulated signal (BPSK, QPSK, QAM) shows spectral correlation peaks at cyclic frequencies equal to integer multiples of the symbol rate, 1/T. Real-valued single-rail formats such as BPSK also show carrier features at twice the carrier frequency, 2 f_c, plus combinations such as 2 f_c plus or minus 1/T, but balanced complex QPSK and QAM suppress that 2 f_c feature because their in-phase and quadrature symbol streams are independent. The symbol-rate features survive even after the carrier is unknown, which is why symbol-rate cyclic features are used for blind detection. OFDM produces peaks at the inverse of the total symbol-plus-cyclic-prefix duration, while a square-wave-clocked or pulsed radar shows features at the pulse repetition frequency and its harmonics. Computing the full spectral correlation over a grid of cyclic frequencies with the FFT accumulation method or the strip spectral correlation analyzer reveals all of these features at once.

Q: Why is cyclostationary detection so computationally expensive?

Energy detection collapses the whole observation into a single power number, but cyclostationary detection must estimate a two-dimensional surface, the spectral correlation density over both spectral frequency f and cyclic frequency alpha. The FFT accumulation method (FAM) and the strip spectral correlation analyzer (SSCA) are the standard efficient algorithms; for an N-point record the FAM costs on the order of N log N per cyclic-frequency strip and the full bifrequency plane scales with the number of strips. A 1 ms record at 20 MHz sampling (20,000 samples) resolving a few hundred cyclic frequencies is well within a modern FPGA or DSP budget, but searching a wide unknown cyclic-frequency range at fine resolution can require billions of operations, which is why detectors usually target a short list of known candidate cyclic frequencies.

/sy-kloh-STAY-shuh-ner-ee dih-TEK-shuhn/

Most man-made communication and radar waveforms carry hidden periodicities, in their carrier, symbol clock, or pulse repetition rate, that make their statistics repeat in time even though the data they convey looks random. Cyclostationary detection exploits exactly that property: rather than measuring raw power like an energy detector, it estimates the cyclic spectral density and looks for spectral correlation at the signal's known cyclic frequencies. Because thermal noise is wide-sense stationary and produces no such correlation, the method extends usable spectrum sensing down to roughly -15 to -20 dB SNR, far below where energy detection's noise-uncertainty wall stops it. The same cyclic features double as a modulation fingerprint, which is why the technique is central to cognitive radio and signal-classification receivers.

Category: Signal Processing

Typical Sensitivity: −15 to −20 dB SNR

Key Statistic: Spectral Correlation S_x^α(f)

Exploiting Hidden Periodicity in Modulated Signals

A signal is second-order cyclostationary when its autocorrelation function is periodic in time rather than constant. Formally, R_x(t, τ) = E[x(t + τ/2) x*(t − τ/2)] repeats with one or more periods, so it can be expanded in a Fourier series whose coefficients are the cyclic autocorrelation function R_x^α(τ). Each nonzero cyclic frequency α corresponds to a periodicity the modulation introduces: the symbol clock contributes features at multiples of the symbol rate 1/T, while the suppressed carrier of a double-sideband signal contributes a feature at twice the carrier, 2f_c. Stationary thermal noise has energy only at α = 0, so any measurable correlation at α ≠ 0 is a near-certain signature of a man-made signal.

The frequency-domain counterpart of R_x^α(τ) is the spectral correlation density S_x^α(f), obtained by Fourier transforming the cyclic autocorrelation over τ. It measures the correlation between spectral components separated by α Hz, and it forms a two-dimensional surface over the (f, α) bifrequency plane. The ordinary power spectral density is simply the α = 0 slice of this surface; everything off that axis is invisible to conventional spectral analysis. Detection reduces to testing whether the magnitude of S_x^α(f) at a target cyclic frequency exceeds a threshold, typically after normalizing by the noise floor to form the spectral coherence.

Cyclic Spectral Density and Cyclic Frequencies

Cyclic Autocorrelation Function:
R_x^α(τ) = lim_T→∞ (1/T) ∫_−T/2^T/2 x(t + τ/2) x*(t − τ/2) e^−j2παt dt

Spectral Correlation Density (Cyclic Spectrum):
S_x^α(f) = ∫_−∞^∞ R_x^α(τ) e^−j2πfτ dτ

Spectral Coherence (test statistic):
C_x^α(f) = S_x^α(f) / √[ S_x⁰(f + α/2) × S_x⁰(f − α/2) ]

Where α = cyclic frequency, τ = lag, f = spectral frequency. Noise is wide-sense stationary so S_n^α(f) ≈ 0 for all α ≠ 0. Linear modulations place symbol-clock features at α = k/T; real-valued (single-rail) formats such as BPSK and DSB also show carrier features at α = 2f_c ± k/T, whereas balanced complex QPSK/QAM suppress the 2f_c feature.

Detector Comparison for Spectrum Sensing

Detector	Prior knowledge	Usable SNR (typical)	Noise-uncertainty robust?	Relative compute	Best use
Energy detection	Noise variance only	−3 to +3 dB	No (hits SNR wall)	1× (baseline)	Fast, signal-agnostic scan
Cyclostationary	Cyclic frequency α	−15 to −20 dB	Yes	10 to 100×	Weak-signal sensing, classification
Matched filter	Full waveform	−20 dB and below	Yes	~5×	Known pilots, sync words
Eigenvalue / covariance	None (blind)	−8 to −12 dB	Partially	~20×	Correlated multi-antenna sensing

Common Questions

Frequently Asked Questions

How much sensitivity advantage does cyclostationary detection have over energy detection?

The decisive advantage is immunity to noise-power uncertainty, not raw sensitivity at one observation length. Energy detection hits an SNR wall: with noise variance known only to within a few tenths of a dB, no energy detector works reliably below roughly −3 to 0 dB SNR regardless of integration time. Because wide-sense stationary thermal noise shows no spectral correlation at any α ≠ 0, a cyclostationary detector at a known cyclic frequency dodges that wall and keeps gaining about 1.5 dB per doubling of the observation interval, reaching −15 to −20 dB SNR at 10 to 100× the compute cost.

What cyclic frequencies should I search for a given modulation?

The cyclic frequencies follow the periodicities the modulation imposes. Every linear format (BPSK, QPSK, QAM) shows peaks at integer multiples of the symbol rate 1/T. Real-valued single-rail formats such as BPSK additionally show carrier features at twice the carrier 2f_c and combinations like 2f_c ± 1/T, but balanced complex QPSK and QAM suppress that 2f_c feature because their in-phase and quadrature symbol streams are independent. The symbol-rate features survive an unknown carrier, so they drive blind detection. OFDM peaks at the inverse of the symbol-plus-cyclic-prefix duration, and a pulsed radar shows features at the PRF and its harmonics. Computing the full spectral correlation with the FFT accumulation method exposes all of them at once.

Why is cyclostationary detection so computationally expensive?

Energy detection collapses a record into one power number; cyclostationary detection estimates a two-dimensional surface, S_x^α(f) over both f and α. The FFT accumulation method (FAM) and strip spectral correlation analyzer (SSCA) are the efficient standards: for an N-point record the FAM costs on the order of N log N per cyclic-frequency strip. A 1 ms record at 20 MHz sampling resolving a few hundred cyclic frequencies fits a modern FPGA budget, but a fine wide-range blind search can reach billions of operations, so detectors target a short list of known candidate α values.

Related Terms

← Cyclist Detection Cylinder Function →