Training Fourier Analysis Wavelets & Time-Frequency Analysis
7 / 7

Wavelets & Time-Frequency Analysis

35 min Fourier Analysis

Wavelets & Time-Frequency Analysis

The Fourier transform gives perfect frequency resolution but zero time resolution — a single sinusoid extends over all time. The Short-Time Fourier Transform (STFT) localizes in time with a sliding window, but is limited by the uncertainty principle. Wavelets provide simultaneous time-frequency localization by dilating and translating a mother wavelet $\psi$. The Haar, Daubechies, and Mexican hat wavelets are used in signal compression, edge detection, and multi-resolution analysis.

Wavelet Transform

The continuous wavelet transform (CWT) of $f\in L^2(\mathbb{R})$ is $W_f(a,b)=\frac{1}{\sqrt{|a|}}\int f(t)\overline{\psi\left(\frac{t-b}{a}\right)}\,dt$ where $a\neq 0$ is the scale (dilation) and $b$ is the translation. The mother wavelet $\psi$ must satisfy the admissibility condition $C_\psi=\int\frac{|\hat{\psi}(\xi)|^2}{|\xi|}\,d\xi<\infty$, implying $\hat{\psi}(0)=0$ (zero mean). The inverse CWT: $f(t)=\frac{1}{C_\psi}\int\int W_f(a,b)\psi\left(\frac{t-b}{a}\right)\frac{da\,db}{a^2}$.

Multi-Resolution Analysis (MRA)

An MRA is a sequence of subspaces $\cdots V_{-1}\subset V_0\subset V_1\cdots$ of $L^2(\mathbb{R})$ with: $\overline{\bigcup V_j}=L^2$; $\bigcap V_j=\{0\}$; $f\in V_j\iff f(2\cdot)\in V_{j+1}$ (scaling); there exists a scaling function $\phi$ with $\{\phi(x-k)\}$ an orthonormal basis for $V_0$. The wavelet $\psi$ spans the orthogonal complement $W_j=V_{j+1}\ominus V_j$. Discrete wavelet transform (DWT): apply low-pass (scaling) and high-pass (wavelet) filter banks, then downsample — iterated to get multi-scale decomposition.

Example 1

Compute the Haar wavelet transform of $x=[4,6,10,2]$.

Solution: Haar mother wavelet: $\psi=[1,-1]/\sqrt{2}$, scaling: $\phi=[1,1]/\sqrt{2}$. Level 1: averages $[(4+6)/2,(10+2)/2]=[5,6]$; differences $[(4-6)/2,(10-2)/2]=[-1,4]$. Level 2: averages $[(5+6)/2]=[5.5]$; differences $[(5-6)/2]=[-0.5]$. Wavelet coefficients: $[-0.5,-1,4]$; scaling $[5.5]$. Perfect reconstruction: invert the filters. Large coefficients indicate edges/discontinuities; small coefficients indicate smooth regions — basis of wavelet compression.

Example 2

Explain why the Mexican hat wavelet $\psi(x)=(1-x^2)e^{-x^2/2}$ is used for edge detection.

Solution: $\psi(x)$ is the negative second derivative of a Gaussian: $\psi=-\frac{d^2}{dx^2}e^{-x^2/2}$. At large scales $a$, $W_f(a,b)\approx a^2\cdot f''(b)$ (up to constants). Thus zero crossings of $W_f$ at scale $a$ correspond to inflection points of $f$ smoothed at scale $a$ — these are edges. The Mexican hat wavelet detects blobs and ridges; at multiple scales, it implements a scale-space edge detector (Marr-Hildreth algorithm).

Practice

  1. Prove the admissibility condition implies $\int\psi(t)\,dt=0$ (zero mean), giving wavelets their oscillatory character.
  2. Show the Haar basis $\{\psi_{j,k}(t)=2^{j/2}\psi(2^j t-k):j,k\in\mathbb{Z}\}$ is orthonormal in $L^2(\mathbb{R})$.
  3. Compare the time-frequency resolution of the STFT and CWT using the Heisenberg uncertainty principle.
  4. Explain why JPEG 2000 uses wavelets instead of the DCT used in JPEG, in terms of ringing artifacts and compression efficiency.
Show Answer Key

1. Admissibility: $C_\psi=\int_0^\infty\frac{|\hat{\psi}(\xi)|^2}{\xi}d\xi<\infty$ requires $\hat{\psi}(0)=0$, i.e., $\int\psi(t)dt=0$. So wavelets must oscillate (have zero mean).

2. Orthonormality: $\langle\psi_{j,k},\psi_{j',k'}\rangle=\delta_{jj'}\delta_{kk'}$. For Haar: $\psi(t)=\mathbf{1}_{[0,1/2)}-\mathbf{1}_{[1/2,1)}$. Translates are supported on disjoint intervals (orthogonal), and dilates are orthogonal by the multiresolution structure.

3. STFT: fixed window $\Rightarrow$ fixed time-frequency resolution $\Delta t\cdot\Delta f\ge1/(4\pi)$. CWT: variable window (dilated wavelet) $\Rightarrow$ good time resolution at high frequency, good frequency resolution at low frequency. CWT adapts the resolution to the signal's local frequency content.

4. DCT produces block artifacts at $8\times8$ boundaries. Wavelets use overlapping basis functions, eliminating blocking artifacts. Wavelets also provide multiresolution decomposition, better preserving edges. JPEG 2000 achieves 10-20% better compression at the same quality compared to JPEG.