Discrete Wavelet Transform (DWT)

The DWT decomposes a signal using a pair of quadrature mirror filters—a low-pass scaling filter and a high-pass wavelet filter—followed by dyadic downsampling. This process projects the signal onto mutually orthogonal subspaces, ensuring that the approximation and detail coefficients are linearly independent. The orthogonality property guarantees that the transform preserves energy (Parseval's theorem) and enables perfect reconstruction without redundancy, making it ideal for compression applications like JPEG 2000.

The DWT implements Mallat's multiresolution framework by iteratively decomposing only the low-frequency approximation coefficients. This creates a hierarchical representation where:

• Coarse scales capture long-duration, low-frequency trends
• Fine scales capture short-duration, high-frequency transients This variable time-frequency tiling overcomes the fixed resolution limitation of the Short-Time Fourier Transform, providing high temporal resolution at high frequencies and high frequency resolution at low frequencies.

A wavelet has N vanishing moments if it is orthogonal to all polynomials up to degree N-1. This property causes the detail coefficients to be zero or negligible in regions where the signal is smooth and polynomial-like, resulting in a sparse representation. The Daubechies family of wavelets is explicitly constructed to maximize vanishing moments for a given support width. Higher vanishing moments produce more concentrated signal energy in fewer coefficients, which is critical for denoising and compression.

A direct extension of the DWT, the Wavelet Packet Decomposition (WPD) generalizes the binary tree structure by decomposing both the approximation and detail coefficients at each level. This yields a complete binary tree of subspaces, providing a richer library of orthonormal bases. An entropy-based best-basis selection algorithm can then adaptively choose the optimal tiling of the time-frequency plane for a specific signal, offering superior flexibility for non-stationary signal analysis compared to the fixed octave-band partitioning of the standard DWT.

The DWT is mathematically equivalent to a two-channel sub-band coding system. The analysis filter bank (H0 for low-pass, H1 for high-pass) splits the signal into sub-bands, which are then critically decimated. The synthesis filter bank (G0, G1) upsamples and interpolates these sub-bands. The filters are designed to satisfy the perfect reconstruction condition:

• Alias cancellation: G0(z)H0(-z) + G1(z)H1(-z) = 0
• No distortion: G0(z)H0(z) + G1(z)H1(z) = 2z^(-l) This ensures the original signal is recovered exactly from its wavelet coefficients.

A translation-invariant variant of the DWT, the Stationary Wavelet Transform (also called 'à trous' algorithm) omits the downsampling step and instead upsamples the filters by inserting zeros between coefficients at each level. This produces a redundant, overcomplete representation where the number of coefficients at each scale equals the original signal length. The SWT is particularly valuable for RF fingerprinting because it preserves precise temporal localization of transient events, making it robust to small time shifts that would cause aliasing in the critically-sampled DWT.

Provides an overcomplete representation by allowing scale and translation parameters to vary continuously. Unlike DWT's dyadic sampling, CWT maps a 1D time series into a 2D time-frequency representation with arbitrary precision. - Key advantage: Finer scale resolution for detecting subtle hardware impairments - Trade-off: Computationally redundant compared to DWT's orthogonal decomposition - Common wavelet: Morlet wavelet for optimal joint time-frequency localization

A generalization of DWT that decomposes both approximation and detail coefficients at each level. This creates a complete binary tree structure, partitioning the frequency spectrum into equal-width sub-bands. - DWT limitation: Only the low-frequency branch is recursively decomposed - WPD advantage: Adaptive selection of the best basis for a given signal - Application: Isolating narrowband impairments in specific frequency regions of an RF emission

The mathematical framework underlying DWT that decomposes a function space into a sequence of nested subspaces. MRA provides coarse approximations and fine details simultaneously. - Scaling function (father wavelet): Captures the approximation coefficients - Wavelet function (mother wavelet): Captures the detail coefficients - RF context: Separates slow-varying oscillator drift from fast transient ringing artifacts in transmitter signatures

A time-frequency reassignment technique that sharpens wavelet-based representations by reallocating coefficients along the frequency axis. SST concentrates energy onto the true instantaneous frequency ridges. - Benefit over DWT: Eliminates spectral smearing inherent in fixed filter banks - Mechanism: Uses phase information to squeeze diffuse energy into precise frequency bins - Use case: Extracting clean, high-resolution signatures from noisy RF captures before feeding into a classifier

A data-driven, adaptive decomposition that breaks a signal into Intrinsic Mode Functions (IMFs) without predefined basis functions. Unlike DWT's fixed filter banks, EMD is fully signal-dependent. - Process: Iterative sifting extracts local mean envelopes - Output: Finite set of oscillatory components ordered from high to low frequency - RF fingerprinting: Isolates non-stationary transient events that fixed wavelet bases might miss or smear across multiple scales

The predecessor to wavelet methods, computing the Fourier transform on windowed signal segments. STFT provides a fixed time-frequency resolution across all frequencies. - Heisenberg limitation: Time resolution × frequency resolution = constant - DWT advantage: Multi-resolution—good time resolution at high frequencies, good frequency resolution at low frequencies - Comparison: STFT is a tiling of the time-frequency plane with fixed rectangles; DWT uses a variable tiling suited for transient detection