Harmonic analysis is the mathematics of decomposing complicated functions and signals into simpler oscillatory components. It begins with Fourier series and transforms, then expands into convolution, orthogonal decompositions, group representations, wavelets and the study of how structure appears across frequency and scale.
A sound can be viewed as a waveform in time or as a combination of frequencies. A heat distribution can be viewed in physical space or as a sum of diffusion modes. A periodic function can be analysed point by point or decomposed into sines and cosines. Harmonic analysis studies these changes of representation and the mathematics that makes them reliable.
Series route: Mathematics Learning Hub → How Mathematics Works → Harmonic Analysis. Important neighbours are Functional Analysis, Real Analysis, Partial Differential Equations and Representation Theory.
1. The central move is decomposition
A complicated object may become easier to understand when expressed as a sum or integral of simpler modes.
Harmonic analysis chooses basis functions or characters suited to translation, oscillation or symmetry, then studies how much of each mode is present.
2. Periodic functions lead to Fourier series
A sufficiently regular periodic function can be represented by a series of sine and cosine terms or complex exponentials.
Each frequency contributes an amplitude and phase. The time-domain shape becomes a frequency-domain sequence.
3. Orthogonality separates independent modes
Sine and cosine functions of different integer frequencies are orthogonal over a full period under the usual L² inner product.
This lets coefficients be extracted by projection, just as coordinates are extracted along perpendicular axes in Euclidean space.
4. Fourier coefficients are inner products
In complex notation, the coefficient at integer frequency n is obtained by integrating the function against e^{-inx} over the period with the appropriate normalisation.
The coefficient measures alignment with that oscillatory mode.
5. Convergence is not one single concept
A Fourier series can converge pointwise, uniformly, in L² norm or in other senses.
These are different statements. A series converging in mean square need not converge uniformly everywhere.
Harmonic analysis therefore requires precise control of the function space and convergence mode.
6. Parseval and Plancherel conserve energy across representations
Parseval-type identities equate the L² energy of a function with the squared magnitudes of its Fourier coefficients.
Plancherel’s theorem extends this principle to the Fourier transform on L².
The representation changes, but a key quadratic quantity remains invariant.
7. Nonperiodic functions lead to the Fourier transform
For functions on the real line, discrete Fourier coefficients become a continuous frequency function.
The Fourier transform expresses a function as a superposition of complex exponentials across a continuum of frequencies under appropriate conditions.
8. Inversion reconstructs the original function
Under suitable assumptions, the inverse Fourier transform recovers the original function from its frequency representation.
This establishes the transform as a reversible change of coordinates rather than a lossy summary.
9. Translation becomes phase
Shifting a function in physical space multiplies its Fourier transform by a complex phase factor.
The magnitude spectrum can therefore remain unchanged while phase records location information.
10. Differentiation becomes multiplication by frequency
Under standard hypotheses, differentiating a function corresponds to multiplying its Fourier transform by a factor proportional to iξ.
Differential equations can therefore become algebraic equations in frequency space.
11. Convolution becomes multiplication
The convolution f*g combines one function with translated copies of another.
Fourier transformation turns convolution into pointwise multiplication: the transform of f*g is the product of the transforms under standard conventions.
This is why filtering becomes especially simple in frequency space.
12. Multiplication becomes convolution
The dual statement also matters: multiplying functions in one domain corresponds to convolution in the transformed domain.
Windowing a signal therefore spreads frequency content instead of leaving frequencies perfectly isolated.
13. The uncertainty principle limits simultaneous localisation
A function cannot be arbitrarily concentrated in both physical space and frequency space.
Narrow localisation in one domain necessarily spreads the representation in the other under precise quantitative inequalities.
This is a mathematical structure, not merely a physical slogan.
14. Smoothness controls Fourier decay
Smoother functions tend to have Fourier transforms or coefficients that decay faster at high frequencies.
Repeated integration by parts makes this relationship visible for sufficiently regular functions.
Frequency decay therefore carries information about regularity.
15. Singularities create slow decay and broad frequency content
A sharp edge or discontinuity requires many high-frequency components for accurate reconstruction.
This explains why smooth signals compress differently from signals with abrupt changes.
16. Gibbs oscillations expose limits of truncated series
Near a jump discontinuity, partial Fourier sums overshoot and oscillate.
Adding more terms narrows the oscillatory region but does not remove the characteristic relative overshoot in the naive partial-sum process.
Approximation quality therefore depends on what feature is being measured.
17. Summability can improve convergence behaviour
Cesàro and Fejér summation average partial sums instead of taking one raw partial sum.
This can produce better convergence for continuous periodic functions and reduce oscillatory artefacts.
18. Fourier analysis solves the heat equation mode by mode
On a bounded interval with suitable boundary conditions, the initial temperature can be decomposed into sine or cosine modes.
Each mode decays exponentially at a rate related to its frequency.
High-frequency spatial variation decays faster, explaining mathematically why diffusion smooths temperature profiles.
19. Fourier analysis solves wave equations through oscillatory modes
Wave equations decompose into normal modes with characteristic frequencies.
Each mode oscillates simply, while the full solution is their superposition.
Complex motion becomes many simple motions added together.
20. Fourier multipliers represent operators
An operator that becomes multiplication by a function m(ξ) in Fourier space is called a Fourier multiplier.
Boundedness properties of the multiplier determine whether the operator behaves well on function spaces.
This connects harmonic analysis directly to functional analysis.
21. Singular integrals generalise differentiation-like behaviour
Operators such as the Hilbert transform involve kernels singular at the origin but interpreted through principal values.
Their boundedness on Lp spaces is a central theme of modern harmonic analysis.
22. Maximal functions control local averaging
The Hardy–Littlewood maximal operator measures the largest local average of |f| over balls or intervals containing a point.
Maximal inequalities help prove differentiation theorems and convergence results.
23. Harmonic analysis extends beyond Euclidean space
On locally compact abelian groups, characters generalise complex exponentials.
Fourier analysis becomes representation theory of the group, with dual groups indexing frequency-like modes.
24. Nonabelian harmonic analysis uses matrix-valued representations
When the symmetry group is noncommutative, irreducible representations may have dimension greater than one.
Frequency components then become matrix-valued or operator-valued rather than scalar.
This is the natural continuation of Representation Theory.
25. Spherical harmonics analyse functions on spheres
Spherical harmonics are eigenfunctions of the spherical Laplacian and provide an orthogonal basis for functions on the sphere.
They appear in geophysics, quantum angular momentum, acoustics and global field modelling.
26. Wavelets localise in both position and scale
Fourier modes are globally spread across the domain. Wavelets use translated and rescaled localised functions.
This makes wavelets effective for signals whose frequency content changes with position or time.
27. Multiresolution analysis organises scales
A multiresolution framework nests approximation spaces from coarse to fine scales.
Wavelet detail spaces capture what is gained when resolution increases.
This creates a mathematically controlled zoom system.
28. The discrete Fourier transform makes the theory computable
For finite sampled data, the discrete Fourier transform maps N samples to N frequency coefficients.
The fast Fourier transform computes this efficiently by exploiting recursive symmetry.
The algorithmic gain changes what becomes practical without changing the underlying transform.
29. Sampling introduces aliasing
Sampling a continuous signal too coarsely can make high frequencies appear as lower frequencies.
Under band-limited assumptions, sampling above the Nyquist rate permits exact reconstruction in the ideal Shannon setting.
The bandwidth assumption is part of the theorem, not an optional detail.
30. Windowing trades spectral leakage against localisation
Finite observations multiply an underlying signal by a time window.
Because multiplication becomes convolution in frequency, sharp truncation spreads spectral energy.
Smoother windows reduce some leakage at the cost of broader frequency peaks.
31. A worked mechanism: decomposing one simple wave
Take f(x)=3cos(2x)+2sin(5x).
- The signal contains a cosine mode at frequency 2 with amplitude 3.
- It contains a sine mode at frequency 5 with amplitude 2.
- Orthogonality means these coefficients can be recovered independently over a full period.
- Differentiation multiplies each mode by its frequency and shifts phase.
A seemingly complicated waveform is already a sum of two independent harmonic modes.
32. Common harmonic-analysis failure modes
- Convergence collapse: saying “the Fourier series converges” without specifying the sense.
- Magnitude-only thinking: ignoring phase and therefore location information.
- Sampling blindness: interpreting aliased frequencies as genuine source frequencies.
- Window blindness: treating finite observation as if it did not alter the spectrum.
- Basis universality: assuming Fourier modes are always the best representation.
- Regularity overclaim: reading rapid coefficient decay without checking function-space assumptions.
- Transform=reality confusion: forgetting that frequency representation is one mathematical view of the same object.
33. Harmonic analysis as a mathematical machine
Function/Signal → Symmetry or Translation Structure → Orthogonal/Representation Modes → Coefficients/Transform → Operator Simplification → Reconstruction → Physical or Analytical Interpretation.
34. What mastery looks like
- interpret Fourier coefficients as projections;
- distinguish periodic series from transforms on the line;
- track pointwise, uniform and norm convergence separately;
- use convolution and multiplier relationships correctly;
- connect smoothness with frequency decay;
- recognise sampling and window effects;
- use wavelets when localisation across scale matters;
- see Fourier analysis as representation theory of translation symmetry.
35. Conclusion
Harmonic analysis works by changing representation. Oscillatory modes expose frequency. Orthogonality separates components. Fourier transformation turns differentiation into multiplication and convolution into products. Wavelets add localisation across scale. Group representations generalise the idea beyond ordinary Euclidean translation.
The point is not that frequency is more real than time or space. The point is that some structures become visible only after the right change of coordinates.
How Mathematics Works | Batch 10
- How Mathematics Works | Lie Groups & Lie Algebras
- Harmonic Analysis — this article
- How Mathematics Works | Calculus of Variations
- How Mathematics Works | Tensor Analysis
Return to the Mathematics Learning Hub or the How Mathematics Works root.