Research Radar — 2026-10-30

Methodology Review: Wavelet shrinkage: removing small details without flattening every scale

English · About 10 minutes

Primary texts support the algorithm and cited applications; the four-point calculation is an original illustration. No current omics benchmark or universal denoising guarantee is claimed.

The intuition

Imagine a measured curve with a broad plateau, one abrupt jump and many small zigzags. Averaging every neighboring pair may remove the zigzags, but it can also blur the jump. A useful question is whether we can first describe variation by its location and scale, and then decide which details deserve to survive. Wavelets provide such a representation: coarse components describe broad structure, while detail coefficients describe localized contrasts. The transform itself does not remove noise. It changes the coordinates in which we make that decision. [R2]

Soft-thresholding then moves each selected detail coefficient toward zero by a chosen amount. A small coefficient disappears; a large one remains, but with reduced magnitude. The hope is that the underlying signal has relatively few important wavelet coefficients while measurement noise is spread more widely. This is a modeling assumption, not a rule that small biological features must be unimportant. Donoho studies the statistical consequences under a Gaussian-noise model and specified function classes. The lesson here is the algorithm and its tradeoff, not a universal guarantee for every assay. [R1]

A small worked example

Consider an invented four-position measurement \(y=(1,3,5,7)\). These numbers are a teaching example, not an experimental dataset. Use an orthonormal Haar transform with four coefficients: \(a=(y_1+y_2+y_3+y_4)/2\) for the overall level, \(b=(y_1+y_2-y_3-y_4)/2\) for the contrast between halves, and \(c=(y_1-y_2)/\sqrt{2}\) and \(d=(y_3-y_4)/\sqrt{2}\) for the two within-pair contrasts. The symbols \(y_i\) denote the four measurements, with position index \(i\) running from one to four. The square-root factors make the transform preserve squared length.

The calculation gives \(a=8\), \(b=-4\) and \(c=d=-\sqrt{2}\). Choose the illustrative threshold \(\lambda=\sqrt{2}\), keep the overall-level coefficient unchanged, and soft-threshold the other three. The new coefficients are \(\widehat{a}=8\), \(\widehat{b}=-4+\sqrt{2}\) and \(\widehat{c}=\widehat{d}=0\). A hat denotes an estimated or modified quantity. This threshold was selected to make the arithmetic transparent; it was not estimated from these four measurements and is not a recommended automatic choice.

To invert, the first two positions are \(\widehat{a}/2+\widehat{b}/2\) plus and minus \(\widehat{c}/\sqrt{2}\). The last two are \(\widehat{a}/2-\widehat{b}/2\) plus and minus \(\widehat{d}/\sqrt{2}\). Thus the reconstructed curve is approximately \((2.707,2.707,5.293,5.293)\). The within-pair zigzags have vanished and the overall mean remains four. However, the difference between the two plateaus has also become smaller. Removing noise and shrinking a genuine contrast happened through exactly the same numerical operation.

For this exercise only, suppose we reveal the hidden target \(f=(2,2,6,6)\). The original sum of squared errors is four; the reconstructed sum is two, because each reconstructed value differs from its target by \(1/\sqrt{2}\). That improvement is one deliberately chosen example, not an estimate of average performance. If the hidden target had instead been the observed ramp itself, thresholding would introduce error. Without replicated measurements or a defensible noise model, the curve alone does not tell us which of these interpretations is scientifically correct.

The necessary mathematics

For the general setup, let \(y=f+\varepsilon\) be a column vector of \(n\) ordered measurements, with unknown signal vector \(f\) and noise vector \(\varepsilon\). Let \(W\) be an orthonormal wavelet-transform matrix, so \(W^{\mathsf{T}}W=I\), where the superscript denotes transpose and \(I\) is the identity matrix. The observed coefficients are \(z=Wy\) and the signal coefficients are \(\theta=Wf\). Under independent Gaussian noise of variance \(\sigma^2\) at every position, the transformed noise has that same covariance. The positive quantity \(\sigma\) is the noise standard deviation. Orthonormality alone preserves covariance; Gaussianity supplies independence after rotation. [R2] [R3]

For a scalar coefficient \(z_j\), indexed by \(j\), soft-thresholding is defined below. The function \(\operatorname{sign}\) returns its argument's sign, \(\max\) takes the larger argument, and \(\lvert z_j\rvert\) is absolute value. The nonnegative threshold is \(\lambda\). After treating the chosen detail coefficients, invert using \(\widehat{f}=W^{\mathsf{T}}\widehat{\theta}\). Coarse coefficients may be retained as in our example.

[R1]

Normalization matters when implementing the paper. Donoho writes an empirical-function convention involving observations divided by \(\sqrt{n}\). The example above transforms the unscaled data vector, so its coefficient noise standard deviation remains \(\sigma\) rather than \(\sigma/\sqrt{n}\). Moving a threshold between these conventions without rescaling changes the estimator. Independently of the probabilistic model, orthonormality also gives \(\lVert\widehat{f}-f\rVert_2^2=\lVert\widehat{\theta}-\theta\rVert_2^2\); the subscript two denotes the Euclidean norm. This explains why coefficient errors are relevant to reconstruction error. [R1] [R2]

Assumptions and failure modes

A basis is a scientific choice. The Haar example favors flat regions separated by jumps. For a smoothly varying target it may leave block-like features. Position and boundary handling also matter: pairing positions one and two is different from starting the pairing one position later. As a diagnostic exercise, shift the same toy signal, repeat the calculation and compare the restored curves after shifting back. Agreement is something to check, not assume from the word multiresolution. This is a proposed sensitivity exercise rather than a new performance claim.

Noise assumptions require equal care. Raw RNA counts are not automatically independent Gaussian measurements with constant variance. Nearby measurements may share technical errors, and a sample may contain real narrow peaks whose coefficients are small. Applying this particular rule to an arbitrary ordering of genes also gives adjacency a meaning it may not have. A smooth result cannot establish that a batch effect has been removed or that a biological interaction exists. If repeated curves are analyzed together, their dependence and experimental units belong in a statistical model; the functional mixed-model reading makes that distinction explicit. [R3]

Verified applications

Mallat demonstrates the representation in image coding, texture discrimination and fractal analysis. Morris and Carroll develop a Bayesian functional mixed model and illustrate it using spatial biomarker measurements along rat colonic crypts. Their procedure models replicated curves and regularizes wavelet coefficients; it is not simply the single-curve soft-threshold rule used in our example. These are verified applications of wavelet representations and regularization. They do not validate a new single-cell denoising pipeline, and this review does not claim comparative performance on current omics benchmarks. [R2] [R3]

Possible transfers — speculative

A possible omics experiment would begin with an already aligned one-dimensional profile, such as a predefined genomic signal track, rather than an unordered cell-by-gene matrix. This is speculative. Preserve raw measurements, define the coordinate axis and decide which scales correspond to features the assay can actually resolve. Use independent biological samples for evaluation. Compare a wavelet reconstruction with an unsmoothed baseline and a simple local average, testing both preservation of known sharp features and behavior in negative-control regions. The published applications above do not establish that this proposed transfer will work.

What to try and what to read next

Try the four-point example before choosing a library or a large dataset. Reconstruct the original vector with no threshold, reproduce the two-unit squared error after thresholding, and then replace the hidden target with a ramp. Write down which result would count as failure for your scientific question. Read Donoho for the statistical estimator, Mallat for the representation, and Morris–Carroll for the extra modeling needed when curves come from multiple experimental units. The practical question is which details you can justify shrinking, with which noise assumptions and which independent evidence.

Reading list

• [R1] Primary reading: De-noising by soft-thresholding. David L. Donoho (1995). Main reading: start with the introduction and three-step algorithm. Note that the paper normalizes observations by the square root of sample size; the worked example here uses the ordinary orthonormal discrete transform instead.

• [R2] Related reading: A theory for multiresolution signal decomposition: the wavelet representation. Stéphane G. Mallat (1989). Representation background: understand why a signal can be split into a coarse component and localized details before any denoising decision is made.

• [R3] Related reading: Wavelet-based functional mixed models. Jeffrey S. Morris and Raymond J. Carroll (2006). Modeling extension and biological application: see how replicated curves require fixed/random effects and a covariance model, rather than treating each curve as an isolated denoising problem.