Continuous Wavelet Transform
Multi-Resolution Analysis of Non-Stationary RF Signals
The continuous wavelet transform was formalized by Grossmann and Morlet in the early 1980s as a remedy for a fundamental limitation of windowed Fourier analysis: a single window length cannot simultaneously resolve a slow carrier and a fast transient. The CWT replaces the fixed window with a wavelet that is stretched or compressed by a scale parameter. Large scales dilate the wavelet to capture low-frequency, long-duration behavior, while small scales compress it to capture high-frequency, short-duration events. This scale-dependent windowing satisfies the Heisenberg uncertainty bound everywhere but distributes the time-frequency resolution where each portion of the signal needs it, which is precisely the property a radar or EMC engineer wants when a signal's bandwidth changes within a few microseconds.
For RF work the signal is typically a digitized baseband or IF record. After downconversion and sampling, the CWT slides each scaled wavelet across the record and integrates the product, building a coefficient surface indexed by scale (related to frequency) and translation (time). The magnitude squared of those coefficients is the scalogram, the wavelet analog of the spectrogram. Engineers read ridges in the scalogram to extract instantaneous frequency from a linear FM chirp, to time-stamp the leading edge of a pulse to sub-sample accuracy, or to separate overlapping emitters whose tones drift. Because the complex Morlet wavelet carries phase, the same transform also yields instantaneous-phase and group-delay estimates useful for pulse-compression diagnostics.
The price of this fidelity is computation and redundancy. Evaluating many scales at every sample produces far more coefficients than input points, so the CWT is an analysis and visualization tool rather than a coding or compression scheme. When an invertible, memory-efficient representation is required, designers move to the dyadic discrete wavelet transform; when smooth, interpretable time-frequency maps matter more than efficiency, the CWT remains the preferred instrument.
Governing Equations
W(a,b) = (1/√a) × ∫ x(t) · ψ*((t − b)/a) dt
Admissibility condition (required for invertibility):
Cψ = ∫ |Ψ(ω)|2 / |ω| dω < ∞, so Ψ(0) = 0 (zero mean wavelet)
Scale-to-frequency mapping:
f ≈ Fc / (a × Ts)
Where a = scale (dilation), b = translation (time shift), ψ = mother wavelet, Ψ = its spectrum, Fc = wavelet center frequency (≈ 0.81 for Morlet), Ts = sampling period. Example: Fc = 0.81, Ts = 10 ns (100 MHz sampler), a = 8 → f ≈ 10 MHz.
Time-Frequency Method Comparison
| Method | Resolution | Sampling | Redundancy | Invertible | Best RF Use |
|---|---|---|---|---|---|
| Continuous wavelet (CWT) | Multi-resolution (scale dependent) | Dense scale & shift grid | High (oversampled) | Yes, with Cψ | Chirp / transient analysis, scalograms |
| Discrete wavelet (DWT) | Multi-resolution (dyadic) | Power-of-two grid | None (critical) | Yes | Denoising, compression, embedded |
| Short-time Fourier (STFT) | Fixed (single window) | Uniform time-frequency | Tunable by hop | Yes | Stationary / slowly varying tones |
| Wigner-Ville | Highest, no window | Quadratic kernel | High (plus cross-terms) | Recoverable to a constant phase | Single-component IF tracking |
| FFT (plain) | Frequency only | Uniform bins | None | Yes | Stationary spectrum, no time info |
Frequently Asked Questions
How does the continuous wavelet transform differ from the short-time Fourier transform?
The STFT uses one fixed-width window, so resolution is constant across the band. The CWT dilates a mother wavelet, giving short windows at high frequencies (sharp time localization for fast transients) and long windows at low frequencies (sharp frequency localization). This multi-resolution behavior suits chirped pulses and EMI bursts whose bandwidth changes quickly, where a fixed STFT window forces a single resolution compromise.
How do you convert wavelet scale to an equivalent RF frequency?
Each scale a maps to a pseudo-frequency f ≈ Fc / (a × Ts), where Fc is the mother wavelet center frequency (≈ 0.81 for the Morlet) and Ts is the sampling period. With a 100 MHz sampler (Ts = 10 ns) and the Morlet wavelet, scale a = 8 corresponds to roughly 10 MHz. The hyperbolic mapping spaces frequency bins logarithmically, matching octave-band and EMI reporting.
Which mother wavelet should be used for RF transient analysis?
The complex Morlet is the usual choice for RF and radar because it is well localized in time and frequency and carries analytic phase for instantaneous-frequency estimation of LFM chirps. Real wavelets like the Mexican hat (Ricker) are used for edge detection, such as locating a pulse rising edge. The complex Morlet bandwidth and center-frequency parameters (cmor B-C) trade time resolution against frequency resolution to match the expected pulse width.
Why is the continuous wavelet transform redundant compared with the discrete wavelet transform?
The CWT evaluates every scale and shift on a fine grid, so its coefficient surface contains many more samples than the input. That oversampling yields a smooth, interpretable scalogram but is computationally heavy and unsuited to coding. The DWT samples scale and translation on a dyadic grid, giving a critically sampled, invertible, non-redundant representation that is efficient for denoising, compression, and real-time embedded processing.