Skip to content

Wavelet Multiresolution Analysis

Multiresolution method — instantiates Model-Guided Signal Separation

Re-expresses the signal across a ladder of scales at once, so structure living at one scale can be separated from nuisance living at another — then reconstructs the target from the scales that hold it.

Wavelet Multiresolution Analysis separates a signal by rewriting it simultaneously at many scales and doing the cleanup scale by scale. Where a single-window transform forces one time-frequency resolution on the whole record, a wavelet decomposition tiles the plane adaptively — fine time resolution for fast, localized features, fine frequency resolution for slow ones. Its defining move is that target and nuisance often live at different scales: baseline wander is coarse, a spike is fine, white noise is spread thin across all scales. Once the signal is spread across the scale ladder, separation becomes a matter of keeping the coefficients that hold the target and shrinking the ones that hold noise, then reconstructing. That scale-adaptive tiling is what distinguishes it from Short-Time Fourier Transform Window Selection, which commits to a single fixed window for the entire record.

Example

A wearable records an ECG corrupted two ways at once: a slow baseline wander as the patient breathes and the skin moves, and a fast fuzz of muscle and electrical noise. A plain band-pass filter struggles because the sharp QRS spike shares frequencies with the muscle noise. Wavelet Multiresolution Analysis decomposes the trace across scales: the baseline wander collects almost entirely in the coarsest approximation coefficients, the muscle fuzz in the finest detail coefficients, and the diagnostic QRS morphology sits in the middle scales, concentrated in a few large coefficients.

Zero out the coarsest band, shrink the small high-scale coefficients toward zero, keep the rest, and reconstruct — the heartbeat returns with its sharp edges intact, because the transient was represented compactly instead of being smeared across a fixed window.

How it works

What distinguishes it from the transforms and smoothers around it is the adaptive, self-contained pipeline:

  • Decompose across scales — successive halving of resolution builds the multiresolution representation.
  • Threshold the coefficients — hard or soft shrinkage kills small (noise-dominated) coefficients while sparing the few large ones that carry a transient.[1]
  • Reconstruct — invert from the kept coefficients to recover the target in the original domain.

Unlike a fixed-window spectrogram, which only sets up a frame for a later filter, the wavelet method performs the separation itself — the threshold is the separator.

Tuning parameters

  • Wavelet basis — the shape matched to the target; a wavelet resembling a QRS complex represents it in fewer coefficients and survives thresholding better.
  • Decomposition depth — how many scales to peel; deeper separates slower nuisance but accrues boundary artifacts.
  • Threshold rule and level — hard versus soft shrinkage and how aggressive; higher thresholds kill more noise but start eroding the target.
  • Per-scale vs. global threshold — allow different aggressiveness at each scale.
  • Boundary handling — how the edges are extended; a wrong choice injects artifacts at the ends of the record.

When it helps, and when it misleads

Its strength shows wherever target and nuisance separate by scale and transients matter: sharp edges survive that a smoother would blur away.

Over-thresholding erases real low-amplitude structure and can leave tell-tale ripples near sharp features, while a poorly matched basis spreads the target across many small coefficients that thresholding then eats. The classic misuse is to crank the threshold until the output looks clean, mistaking smooth for true. The discipline is to judge by whether the removed part — the residual — looks like noise and nothing else, rather than by whether the kept part looks tidy.

How it implements the components

Wavelet Multiresolution Analysis fills the transform-and-separate side of the archetype:

  • separation_operator_specification — the transform → threshold → reconstruct chain is the separator, defined by basis and shrinkage rule.
  • observation_frame — the multiresolution (scale × position) representation is the frame in which separation happens.
  • information_loss_metric — every shrunk or dropped coefficient is a chosen loss; the metric accounts for how much target energy went out with the noise.

It does not model the sources generatively — State-Space Model does that — and it is not a validation of the result; a Held-Out Sample Test or residual check judges whether the cleanup was honest.

References

[1] Wavelet shrinkage denoises by transforming to the wavelet domain, shrinking small coefficients toward zero (soft thresholding) on the assumption that noise spreads into many small coefficients while signal concentrates into a few large ones, then reconstructing from what remains. ↩