Wigner-Ville Distribution (WVD)

The WVD is fundamentally computed by taking the Fourier transform of the instantaneous autocorrelation function. Unlike the STFT, which uses a fixed window, the WVD's kernel is derived directly from the signal itself.

• Kernel Definition: The signal is correlated with a time-reversed copy of itself: R(t, τ) = s(t + τ/2) s(t - τ/2)*.
• No Window Trade-off: Because no external windowing function is applied, the WVD bypasses the Heisenberg-Gabor uncertainty principle inherent in linear transforms.
• Perfect Localization: For a linear chirp signal, the WVD produces a delta function concentrated exactly on the signal's instantaneous frequency law.

The most significant drawback of the WVD is the generation of cross-terms when analyzing multi-component signals. These are spurious oscillatory artifacts that appear midway between genuine signal components.

• Bilinear Origin: Cross-terms arise because the quadratic transform creates interaction terms between every pair of signal components.
• Oscillatory Nature: These artifacts oscillate rapidly in the time-frequency plane, often taking negative values, which violates the intuitive notion of an energy density.
• Mitigation Strategies: Techniques like the Choi-Williams Distribution or smoothed pseudo-WVD apply kernel filters in the ambiguity domain to suppress cross-terms at the cost of reduced auto-term resolution.

The WVD satisfies the fundamental marginal properties expected of a proper time-frequency energy distribution, making it a member of Cohen's class.

• Time Marginal: Integrating the WVD over frequency yields the instantaneous power of the signal: |s(t)|².
• Frequency Marginal: Integrating over time yields the energy spectral density: |S(f)|².
• Total Energy: The double integral over time and frequency equals the total signal energy, ensuring physical consistency.

The WVD preserves the temporal and spectral boundaries of the original signal, a property known as finite support.

• Time Support: If a signal is zero outside a specific time interval, its WVD is also zero outside that interval.
• Frequency Support: If the signal's spectrum is band-limited, the WVD is zero outside that frequency band.
• Strong Support: This contrasts with linear transforms like the STFT, where windowing can smear energy into regions where the signal has no content.

Despite being derived from a complex-valued analytic signal, the Wigner-Ville Distribution is always real-valued. This is a critical property for interpreting the result as a quasi-probability density.

• Hermitian Symmetry: The instantaneous autocorrelation function possesses Hermitian symmetry, guaranteeing that its Fourier transform is real.
• Negative Values: While real, the WVD can take negative values in regions of cross-term interference. This prevents it from being a true probability density but allows it to capture phase information and interference patterns.

Computing the WVD on discrete sampled data requires careful handling to avoid aliasing in the frequency domain.

• Nyquist Constraint: The instantaneous autocorrelation effectively doubles the signal's bandwidth. To prevent aliasing, the analytic signal must be sampled at twice the Nyquist rate or the original signal must be oversampled by a factor of two.
• Discrete Form: The discrete WVD is computed using a symmetric lag window and FFT, often implemented with zero-padding to improve visual interpolation.
• Practical Libraries: Implementations are available in scipy.signal (Python) and MATLAB's Signal Processing Toolbox under time-frequency analysis functions.

The WVD's superior resolution makes it exceptionally well-suited for analyzing transient signals—the brief turn-on and turn-off periods of a transmitter. These moments are rich in unique hardware-specific artifacts.

• Captures the instantaneous frequency trajectory of power amplifier ramp-up
• Reveals subtle, device-specific ringing and overshoot patterns
• Provides a high-resolution view of the transition from noise to steady-state transmission

Beyond communications, WVD is a gold standard for extracting micro-Doppler signatures from radar returns. These signatures are generated by the mechanical vibrations or rotations of a target, such as a drone's propellers or a vehicle's engine.

• Resolves closely spaced, time-varying frequency components
• Enables classification of drone type and payload based on propeller blade flash
• Used in gait analysis for through-wall human identification

The primary drawback of the WVD is cross-term interference. For multi-component signals, the distribution's quadratic nature generates spurious oscillatory artifacts at the midpoint between every pair of true signal components.

• Cross-terms can obscure weaker signal features and lead to misinterpretation
• They appear midway in time and frequency between actual components
• This limitation directly motivates the use of Cohen's class distributions with smoothing kernels

The WVD is a first-order estimator of instantaneous frequency for linear frequency-modulated (LFM) signals. The peak of the WVD in the time-frequency plane directly corresponds to the signal's instantaneous frequency law.

• Provides an unbiased estimate for chirp signals
• Ideal for analyzing frequency-hopping patterns in cognitive radio
• Used to characterize voltage-controlled oscillator (VCO) tuning non-linearity as a unique device fingerprint

To mitigate cross-terms while retaining high resolution, the WVD is often filtered using a 2D kernel in the ambiguity domain, leading to distributions like the Choi-Williams Distribution (CWD).

• The kernel acts as a low-pass filter to smooth cross-terms
• A trade-off exists between cross-term suppression and auto-term resolution
• This kernel design is a central theme in Cohen's class of time-frequency distributions

WVD provides a direct visual map of hardware impairments that manifest as non-linear frequency modulations. For example, a power amplifier's AM/PM distortion creates a characteristic curvature in the instantaneous frequency trajectory.

• Visualizes I/Q modulator memory effects as time-varying frequency shifts
• Reveals phase noise as a broadening of the spectral line over time
• Enables the isolation of specific impairment fingerprints for deep learning classifiers