Wavelet analysis describes a signal at several scales while retaining information about where changes occur. It is useful when an event is brief, when a rhythm changes or when one frequency summary would hide the timing of important behaviour.
The gold standard is not a colourful scalogram or a dramatic coefficient. It is a defined signal, a justified transform, checked boundary conditions and an interpretation that survives comparison with the original observations.
eduKateSG · Wavelet Analysis
Find your next learning step
Start with a four-sample Haar calculation, reconstruct the signal, and learn which time–scale patterns count as defensible evidence.
Full chapter index · Practice and transfer · Education root · How X Works
Full chapter index
Define the signal · Chapters 1–2
Calculate the transform · Chapters 3–4
Check the evidence · Chapters 5–7
Interpret and transfer · Chapters 8–10
Read across the ecosystem · Chapters 11–12
Suppose a sensor records a nearly steady vibration followed by a short burst. A whole-record Fourier spectrum can identify contributing frequencies, but it does not directly identify when the burst occurred. A time–scale representation is useful because timing is part of the research question.
First describe the raw observation: units, sampling interval, observation length, missing values and whether the time axis is regular. A transform of a mistaken time axis gives precisely arranged but misleading coefficients.
Define the intended use before choosing an analysis. Event detection, compression, denoising, periodicity testing and cross-series comparison are different tasks. Each needs its own performance criterion, rather than a claim that one wavelet method is universally best.
An interpretable result identifies a scale, a time interval and a physical phenomenon independently supported by the measurements. A large coefficient alone is a mathematical observation, not a causal explanation.
CHAPTER 2 OF 12 · Define the signal
2. Continuous and discrete wavelets answer related but different questions
A wavelet is a localised analysing function. Shifting it changes the time position being examined; dilating or compressing it changes the scale. The continuous wavelet transform correlates the signal with shifted and scaled versions of a selected wavelet, yielding a map over position and scale.
A discrete wavelet transform typically decomposes a sampled signal through pairs of low-pass and high-pass filters. Under appropriate orthonormal choices, the coefficients permit exact reconstruction and a useful energy-accounting check.
The continuous transform is often redundant: nearby scales and times are not independent observations. A discrete orthonormal transform can be compact, but different boundary rules and wavelet families produce different interpretations.
Scale is related to, but is not automatically identical to, Fourier period or frequency. A scale-to-period conversion depends on the chosen wavelet and the convention used by the implementation. Label an axis honestly instead of relabelling every scale as a measured oscillation period.
CHAPTER 3 OF 12 · Calculate the transform
3. Worked example: a four-sample Haar transform
Consider four equally spaced samples with arbitrary amplitude units: x = [2, 4, 6, 8]. This is an intentionally simple teaching signal, not a record from a real instrument.
Use the orthonormal Haar pair rule: for adjacent samples u and v, average coefficient a = (u + v)/√2 and detail coefficient d = (u − v)/√2. The √2 factor preserves Euclidean energy and is essential to this particular normalization.
At the first level, the two averages are 6/√2 ≈ 4.243 and 14/√2 ≈ 9.899. The two details are −2/√2 ≈ −1.414 and −2/√2 ≈ −1.414. These negative details tell us that the second member of each pair exceeds the first.
Apply the same rule to the two first-level averages. The second-level average is a₂ = 10 and its detail is d₂ = −4. The complete coefficient set is therefore [10; −4; −1.414, −1.414], with its level structure explicitly retained.
The calculation separates a coarser change across the two pairs from finer changes within each pair. It does not yet justify speaking of a dominant physical period: four samples and a Haar basis supply a decomposition, not a validated model of an underlying oscillation.
CHAPTER 4 OF 12 · Calculate the transform
4. Worked example: inverse reconstruction and energy conservation
Reverse the Haar pair with u = (a + d)/√2 and v = (a − d)/√2. Starting from a₂ = 10 and d₂ = −4 reconstructs the two first-level averages, approximately 4.243 and 9.899.
Using the first-level details of approximately −1.414 then reconstructs 2, 4, 6 and 8. The recovered samples match the inputs up to the rounding introduced in the displayed approximations.
An independent check uses energy. The original sum of squares is 2² + 4² + 6² + 8² = 120. The coefficient sum of squares is 10² + (−4)² + 2 × (−√2)² = 100 + 16 + 4 = 120.
This equality is a property of the chosen orthonormal transform and treatment of the complete coefficients; it does not hold unchanged under every redundant continuous transform or every normalization. When a library output differs, inspect convention and boundaries before alleging an error.
Transfer the same check to an eight-sample signal. If coefficients are discarded for compression or denoising, calculate reconstruction error rather than assuming that the omitted detail was merely noise.
CHAPTER 5 OF 12 · Check the evidence
5. A scalogram visualises transform power, not a complete explanation
In a continuous wavelet display, horizontal position commonly represents time, while the vertical axis represents scale or a scale-derived period. Colour may represent the magnitude or power of the transform. The legend must say exactly what has been plotted.
A ridge of high power may indicate a persistent oscillatory contribution under the chosen wavelet. It may also reflect trend, transient forcing, leakage or an amplitude change. Compare the pattern with the original time series and plausible alternate processing choices.
Wavelet coherence and cross-wavelet power are not interchangeable. Coherence is a normalized measure of association across time and scale; it cannot by itself prove that one observed process caused another.
Do not count neighbouring pixels in a scalogram as independent confirmations. Transform redundancy, smoothing and shared input observations create dependencies among the displayed points.
CHAPTER 6 OF 12 · Check the evidence
6. Sampling and resolution constrain what the transform can reveal
An evenly sampled series has a sampling interval Δt and corresponding Nyquist frequency 1/(2Δt) under the usual sampling assumptions. Sampling too slowly can alias high-frequency behaviour into a lower apparent rhythm before any wavelet transform is attempted.
A short analysing wavelet offers sharper time localization but comparatively coarser frequency discrimination. A broader wavelet can resolve frequency more sharply while smearing the time of an event. This trade-off should be tied to the task, not described as a defect.
Record detrending, normalization, interpolation and filtering steps. These choices alter low-frequency content and transient behaviour; they are not cosmetic preparation.
Missing observations and irregular timestamps need a declared policy. Filling a long gap with smooth interpolation can manufacture structure that the instrument never recorded. The NOAA wavelet examples explicitly warn that some workflows require complete regularly spaced data.
Change the sampling cadence or shorten the record and rerun the analysis. A robust interpretation distinguishes stable features from effects created by the chosen representation.
A finite record does not contain observations beyond its beginning or end. A wavelet positioned near either boundary overlaps this unknown region, so its coefficient depends on an assumed extension rule such as reflection, periodic continuation or zero padding.
Padding can make computation convenient, but it does not create genuine measurements. In continuous-wavelet practice, a cone of influence marks regions where edge effects deserve particular caution under an adopted definition.
A feature close to an endpoint can be visually strong and scientifically ambiguous. Check whether its timing and scale persist when you extend the observation, alter the padding rule or exclude boundary-affected regions.
Document how the program handles wavelet length, signal length, filters and residual samples. A library default is part of the method and therefore part of what another reader needs to reproduce the result.
CHAPTER 8 OF 12 · Interpret and transfer
8. Statistical significance requires a justified null model
Coloured regions of a wavelet power plot are not automatically statistically significant. Torrence and Compo’s practical guide explains testing against background noise models and interpreting confidence with attention to scale and edge effects.
For environmental time series, autocorrelation matters. A white-noise null may make slowly varying correlated measurements appear surprising; an autoregressive red-noise model can be a more relevant comparison in some cases, but its parameters must be estimated and checked.
Inspect what is being tested: local pointwise power, averaged power, a region, or an entire map. Testing many pixels raises a multiple-comparisons problem; a nominal local significance threshold does not guarantee a controlled map-wide false-positive rate.
State the null model, confidence rule and correction strategy. Then ask whether independent physical observations support the interpretation. A statistical signal and a causal mechanism are different claims.
CHAPTER 9 OF 12 · Interpret and transfer
9. Denoising and compression make an information trade-off
A discrete wavelet transform can separate fine-scale variation that a user may choose to shrink or discard. Hard thresholding sets some coefficients to zero; soft thresholding reduces their magnitudes. Either can suppress noise, but either can also erase a weak legitimate event.
The choice of threshold, wavelet family, decomposition depth and boundary treatment should follow the scientific objective. A method designed for an audio recording is not automatically suitable for a narrow spectral line or a short meteorological extreme.
Evaluate reconstruction with metrics tied to the use case: retained events, timing error, amplitude bias, false alarms or held-out performance. Global mean-squared error alone can reward a smooth result that loses a rare but important event.
When a denoised series supports a downstream model, validate the whole pipeline on data it did not use for tuning. Otherwise the transform can inadvertently improve a demonstration while damaging real-world generalisation.
For a complete phase test, Why a One-Sample Shift Changes a Decimated Haar Transform computes every Haar level for one short signal and its translated copy. It verifies exact reconstruction and equal energy before showing how the same coefficient threshold can produce different approximations and errors.
CHAPTER 10 OF 12 · Interpret and transfer
10. Practice, FAQs and a changed-context transfer
Practice A: An eight-sample series contains one sharp pulse. Explain why a multiscale plot might help locate the pulse, then show the raw samples alongside the transform. Do not interpret a high coefficient before checking sampling and the edge rule.
Practice B: Repeat the four-sample Haar example after replacing the final sample 8 with 12. Which coefficients change? Reconstruct all samples and compare total energy in the two representations.
FAQ — Does wavelet power prove periodicity? No. It can show localized strength under a chosen basis. A reliable periodicity claim needs a suitable background model, multiple-scale consideration and physical context.
FAQ — Can wavelets recover frequencies that were not sampled? No. A transform cannot undo aliasing or invent observations from gaps. Some modelling methods can make assumptions about missing information, but those assumptions must be stated separately.
FAQ — Is a wavelet transform better than Fourier analysis? Neither dominates every task. A stable global spectrum and a changing local time–scale structure answer different questions; their joint use is often more informative.
The transfer test is to explain an entirely different signal, such as a vibration recording instead of rainfall. Identify what changes in units, noise, sampling and event definition, while keeping the reconstruction and interpretability checks intact.
CHAPTER 11 OF 12 · Read across the ecosystem
11. Connected signal, mathematics and measurement routes
Wavelet analysis belongs to a larger family of methods for representing observations. Move from the broad signal-processing root into mathematics and statistics, and then into measurement disciplines where a transform is only one part of the evidential chain.
A good reading route names the source observation, the mathematical representation and the decision made from it. It also records when a different branch, such as remote sensing or wind profiling, has additional sensor-physics constraints.
Use the following verified eduKateSG reading routes to follow the concepts into adjoining disciplines. Each destination has a different question and should not be treated as a substitute for this chapter.
- The Gold Standard of Signal Processing — signal representations, filters and measurement pipelines
- The Gold Standard of Mathematics — orthogonality, transformations and reconstruction
- The Gold Standard of Statistics — hypothesis tests, uncertainty and multiplicity
- The Gold Standard of Wind Lidar — an example of time-dependent atmospheric sensor measurements
- The Gold Standard of Remote Sensing — acquisition effects and evidence checks before signal interpretation
- The Gold Standard of Wavefront Sensing — different optical measurements that also require inverse reasoning
CHAPTER 12 OF 12 · Read across the ecosystem
12. Primary sources, boundaries and the next learning step
Wavelet calculations in this guide are mathematical teaching examples. They do not establish a real climate cycle, diagnose a rotating machine or validate a detector. For any applied claim, the original samples, metadata, processing code and independent checks are needed.
Continue with a transparent analysis notebook: write down sampling and wavelet parameters, show the untransformed series, mark edge-affected coefficients, declare the statistical null and report how the finding changes under an alternative choice.
Primary references and technical guidance for checking the mechanisms, assumptions, instruments and interpretation discussed in this guide:
- Torrence and Compo (1998), A Practical Guide to Wavelet Analysis — foundational worked treatment of time–scale plots, edge effects and statistical testing
- University of Colorado, Torrence–Compo wavelet research resources — research context, errata and software references
- NOAA Physical Sciences Laboratory, wavelet-analysis interface — sampling, missing data, white/red noise assumptions and practical controls
- UCAR NCL documentation, wavelet function — software interface, coefficient representation and underlying algorithm
All sample amplitudes and calculations are constructed instructional examples. The exact decomposition and energy identity concern the explicitly stated orthonormal Haar convention, not every implementation or type of wavelet transform.
The teaching standard is transferable reasoning: define the quantity, show the working, check the assumptions and apply the method in a changed situation. The immutable Clementi Secondary 1 Mathematics teaching reference is a separate example of that instructional discipline; it is not a technical source for Wavelet Analysis.
Specialist boundary-extension branch
To calculate how zero, repeated-edge, reflected and periodic continuations change exact edge coefficients while keeping observed and invented samples separate, continue with Why Boundary Extension Changes Edge Coefficients.