Wavelets

Listen to a piece of music. There is a deep bass line rolling slowly beneath rapid guitar riffs and a crash of cymbal that vanishes in a millisecond. To describe all of it you need very different time scales at the same moment — broad strokes for the bass, fine strokes for the cymbal.

The Fourier transform, invented in the early nineteenth century, tells you which frequencies are present in a signal, but it smears them across all time. A single cymbal crash looks the same as a continuous tone of the same pitch — because Fourier has no notion of when.

Wavelets solve this. A wavelet is a small oscillating wave, like a ripple, that is localized in both time and frequency. By scaling and shifting that ripple across the signal, the wavelet transform produces a map of what is happening where and at what scale. You see the slow bass and the fast cymbal simultaneously, each at the resolution it deserves.

This multi-resolution view is not just elegant — it is extremely efficient. Most signals spend most of their energy at coarse scales; fine detail is sparse. Drop the smallest, weakest fine-scale coefficients and almost nothing perceptible changes. That observation is the engine behind JPEG 2000, the FBI fingerprint database, and seismic signal processing.

The demo below lets you build a simple signal, decompose it with the Haar wavelet transform, and then throw away the smallest detail coefficients to see how much you can compress it before quality suffers.

<p class="hint">{{hint}}</p>
<div class="controls-row">
  <div class="ctrl-group">
    <label>{{label_preset}}</label>
    <div id="presets" class="btn-group">
      <button class="preset active" data-id="smooth">{{btn_smooth}}</button>
      <button class="preset" data-id="step">{{btn_step}}</button>
      <button class="preset" data-id="spiky">{{btn_spiky}}</button>
      <button class="preset" data-id="mixed">{{btn_mixed}}</button>
    </div>
  </div>
  <div class="ctrl-group">
    <label>{{label_threshold}} <span id="tval">0.00</span></label>
    <input type="range" id="thresh" min="0" max="100" value="0" step="1">
  </div>
</div>
<canvas id="cv" width="560" height="180"></canvas>
<div id="coeff-wrap">
  <canvas id="cc" width="560" height="90"></canvas>
</div>
<div id="stats">{{stats_kept_prefix}} <span id="kept">32</span> / 32 {{stats_kept_suffix}} &nbsp;|&nbsp; {{stats_ratio_label}} <span id="ratio">1.0x</span></div>

* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; }
.hint { font-size: .88rem; color: #444; margin: 0 0 .7rem; line-height: 1.45; }
.controls-row { display: flex; gap: 1.2rem; flex-wrap: wrap; align-items: flex-end; margin-bottom: .6rem; }
.ctrl-group { display: flex; flex-direction: column; gap: .25rem; }
.ctrl-group label { font-size: .8rem; font-weight: 600; color: #555; }
.btn-group { display: flex; gap: .3rem; flex-wrap: wrap; }
.preset { font: 600 13px system-ui; padding: .3rem .65rem; border-radius: 6px;
          border: 1px solid #1d3557; background: #fff; color: #1d3557; cursor: pointer; }
.preset.active { background: #1d3557; color: #fff; }
input[type=range] { width: 180px; accent-color: #1d3557; cursor: pointer; }
canvas { display: block; border: 1px solid #dde3eb; border-radius: 8px; background: #f8fafc;
         max-width: 100%; }
#coeff-wrap { margin-top: .5rem; }
#stats { font-size: .85rem; color: #444; margin-top: .4rem; }
#stats span { font-weight: 700; color: #1d3557; }

// Code not found

Threshold controls how aggressively fine detail is dropped. At threshold 0 you keep every coefficient and reconstruct perfectly. As you raise the threshold, coefficients smaller than that value are zeroed out — the reconstructed signal (in red) loses fine wiggles but keeps the overall shape. Notice how many coefficients you can discard before the reconstruction looks bad. That is the heart of wavelet compression.

The simplest wavelet is the Haar wavelet, introduced by Alfréd Haar in 1909. Take a signal of n samples. Repeatedly pair adjacent values, record their average (the coarse trend) and their difference (the fine detail). After one pass you have n/2 averages and n/2 differences. Apply the same step to the averages only, then again, until you reach a single number. This is the fast wavelet transform (FWT) and it runs in $O(n)$ — linear time, faster than the $O(n \log n)$ FFT.

The averages encode the broad shape of the signal; the differences encode detail at each scale. To reconstruct, undo the steps: replace each average–difference pair with the two original values. The process is lossless — perfect reconstruction is guaranteed by design.

Daubechies wavelets (Ingrid Daubechies, 1988) generalize this idea. By carefully choosing the averaging and differencing filter coefficients, Daubechies constructed wavelets that are smooth, compactly supported (zero outside a finite window), and orthogonal. The "Daubechies 4" wavelet underpins JPEG 2000; the "Daubechies 8" is standard in the FBI fingerprint compression format (WSQ).

The key insight behind all these wavelets is multiresolution analysis: the idea (formalized by Stéphane Mallat, 1989) that a signal can be approximated at a hierarchy of resolutions, with each finer level adding detail. This framework makes the transform modular, fast, and invertible — properties no ad-hoc method can guarantee.

Unlike Fourier, wavelets are not a fixed basis. You can design a wavelet for your problem: smooth or spiky, wide or narrow, real or complex. This flexibility is why there are dozens of wavelet families (Morlet, Meyer, coiflets, symlets) each suited to different tasks.

Learn about compression to see how discarding coefficients leads to practical file-size savings, and how wavelet ideas connect to the broader theory of information and entropy.

Wavelets show up wherever signals must be compressed, denoised, or analyzed across multiple scales:

• Image compression — JPEG 2000: the successor to JPEG replaces the block-based DCT with a wavelet transform. The result supports lossless and lossy compression in the same format, scales smoothly from thumbnail to full resolution, and handles arbitrarily large images without block artifacts.
• Law enforcement — FBI WSQ: the FBI stores 70 million fingerprint cards. Since 1993 they have used the Wavelet Scalar Quantization (WSQ) standard, based on Daubechies wavelets, to compress each card to about 10:1 without losing ridge detail needed for matching.
• Medical imaging — MRI and CT denoising: scan acquisition is noisy. Wavelet thresholding (proposed by David Donoho and Iain Johnstone, 1994) removes noise while preserving sharp tissue boundaries far better than Gaussian blurring.
• Audio — MP3 and beyond: while MP3 uses modified DCT, many newer codecs (Vorbis, parts of AAC) use wavelet-like filter banks. The perceptual principle is the same: concentrate bits on the frequency bands and time frames where energy is high.
• Geophysics and finance: seismologists use wavelets to isolate earthquake signatures buried in noise; quantitative traders analyze price series at multiple time scales to separate trend from micro-volatility.
• Numerical PDEs: wavelet bases can represent the solution to a partial differential equation adaptively — fine grid where the solution is rough, coarse grid elsewhere — reducing computation dramatically.

The wavelet transform is a mathematical microscope with a built-in zoom wheel. Turn it to the coarsest setting and you see the slow sweep of a signal; turn it fine and the sharpest transients snap into focus — all without losing track of when each feature occurs. That simultaneity is the thing Fourier cannot do, and it changes what compression, denoising, and analysis can achieve.

From a 1909 curiosity by Haar to a rigorous theory by Mallat and Daubechies in the late 1980s, wavelets went from pure mathematics to the backbone of image and fingerprint databases in under a decade. The lesson is a recurring one in computation: a well-chosen mathematical structure can make a hard problem easy, a large file small, and a noisy signal clean — all at the same time.