Harmonic analysis is a branch of mathematical analysis concerned with the representation of functions or signals as superpositions of basic waves, such as sines, cosines, or more general eigenfunctions of differential operators. It originates from the study of Fourier series and Fourier transforms, and has expanded to include abstract harmonic analysis on topological groups, the analysis of singular integrals, and modern time-frequency analysis. The field is central to many areas of mathematics, physics, and engineering.
1 Foundations
The foundations of harmonic analysis lie in decomposing functions into elementary oscillatory components. This section covers the classical theories of Fourier series, Fourier transforms on Euclidean space, and discrete Fourier analysis.
1.1 Fourier series
Fourier series represent periodic functions as infinite sums of sines and cosines. For a function \(f\) of period \(2\pi\), its Fourier series is given by \(f(x) \sim \frac{a_0}{2} + \sum_{n=1}^{\infty} (a_n \cos(nx) + b_n \sin(nx))\), where the coefficients are computed via integrals. The theory addresses when and how such series converge to the original function.
1.1.1 Convergence and summability
The convergence of Fourier series is not guaranteed pointwise for all continuous functions. Methods of summability, such as Cesàro and Abel means, provide ways to recover the function even when the series diverges.
1.1.1.1 Pointwise convergence theorems
Pointwise convergence of Fourier series was a major historical problem. The Dirichlet–Jordan theorem states that if a function is of bounded variation, its Fourier series converges pointwise to the average of left and right limits. Carleson's theorem (1966) showed that for square-integrable functions, the Fourier series converges almost everywhere.
1.1.1.2 Cesàro and Abel summability
Cesàro summation averages partial sums of the Fourier series; the Fejér theorem states that the Cesàro means of a continuous periodic function converge uniformly to the function. Abel summation uses the Poisson kernel to recover the function from its Fourier coefficients, providing a continuous analog.
1.2 Fourier transform on ℝⁿ
The Fourier transform extends Fourier series to non-periodic functions on Euclidean space. It converts a function \(f\) on \(\mathbb{R}^n\) into a function \(\hat{f}\) on the dual space, representing \(f\) as a superposition of complex exponentials.
1.2.1 L¹ theory and inversion
For integrable functions, the Fourier transform is well-defined and continuous. The inversion formula recovers the original function from its transform under appropriate conditions.
1.2.1.1 The Riemann–Lebesgue lemma
| The Riemann–Lebesgue lemma states that the Fourier transform of an integrable function vanishes at infinity: \(\lim_{ | \xi | \to\infty} \hat{f}(\xi) = 0\). This is a fundamental property indicating that high-frequency components decay. |
|---|
1.2.1.2 Inversion formula and Fourier multipliers
The inversion formula \(f(x) = \frac{1}{(2\pi)^n} \int_{\mathbb{R}^n} \hat{f}(\xi) e^{i x\cdot\xi} \, d\xi\) holds almost everywhere if \(\hat{f}\) is integrable. Fourier multipliers are operators defined by multiplying \(\hat{f}\) by a function \(m(\xi)\); boundedness of such operators is a rich area of study.
1.2.2 L² theory and Plancherel's theorem
| Plancherel's theorem establishes that the Fourier transform is an isometry on \(L^2(\mathbb{R}^n)\): \(\|f\|_2 = (2\pi)^{-n/2} \|\hat{f}\|_2\). This allows extension of the transform from the dense subspace \(L^1 \cap L^2\) to all of \(L^2\), preserving inner products. |
|---|
1.3 Discrete Fourier analysis
Discrete Fourier analysis deals with functions defined on finite or countable sets, essential for digital computation.
1.3.1 Discrete Fourier transform (DFT)
The DFT maps a finite sequence of \(N\) complex numbers to another sequence of the same length, defined by \(X_k = \sum_{n=0}^{N-1} x_n e^{-2\pi i k n / N}\). It is widely used in signal processing and numerical analysis.
1.3.2 Fast Fourier transform (FFT)
The FFT is an efficient algorithm for computing the DFT, reducing the complexity from \(O(N^2)\) to \(O(N \log N)\). Discovered in its modern form by Cooley and Tukey (1965), it revolutionized digital signal processing.
2 Abstract harmonic analysis
Abstract harmonic generalizes Fourier analysis to functions on groups, particularly locally compact groups, using representation theory.
2.1 Locally compact abelian groups
On locally compact abelian (LCA) groups, the Fourier transform is defined via characters—continuous homomorphisms from the group to the circle.
2.1.1 Pontryagin duality
Pontryagin duality states that the dual group of an LCA group is itself LCA, and the bidual is naturally isomorphic to the original group. This provides a symmetric framework for Fourier analysis on abelian groups.
2.1.2 The dual group and characters
The dual group \(\widehat{G}\) consists of all continuous characters of \(G\). For \(\mathbb{R}\), the characters are \(x \mapsto e^{i\xi x}\); for the circle \(\mathbb{T}\), they are \(n \mapsto e^{in\theta}\). The Fourier transform on an LCA group maps functions on \(G\) to functions on \(\widehat{G}\).
2.1.3 Plancherel and Peter–Weyl theorems
The Plancherel theorem for LCA groups extends the \(L^2\) isometry property. The Peter–Weyl theorem applies to compact groups (not necessarily abelian), stating that the matrix coefficients of irreducible representations form an orthonormal basis for \(L^2\) of the group.
2.2 Non-abelian groups
Harmonic analysis on non-abelian groups uses representation theory to decompose functions.
2.2.1 Representation theory and character formulas
Representations of groups are homomorphisms into linear operators. Characters, the traces of these operators, play the role that exponentials play in the abelian case.
2.2.1.1 Finite groups
For finite groups, the Fourier transform is a matrix-valued transform defined by irreducible representations. Orthogonality relations for matrix coefficients lead to inversion formulas and convolution theorems.
2.2.1.2 Compact groups
Compact groups have a complete theory similar to finite groups, with integration via Haar measure. The Peter–Weyl theorem ensures that irreducible representations are finite-dimensional and span \(L^2\).
2.3 Gelfand pairs and spherical functions
A Gelfand pair \((G,K)\) is a group and a subgroup such that the convolution algebra of bi-\(K\)-invariant functions is commutative. Spherical functions are eigenfunctions of invariant differential operators and generalize the exponential functions on symmetric spaces.
3 Modern topics
Modern harmonic analysis addresses singular integral operators, time-frequency representations, and non-linear extensions.
3.1 Singular integral operators
Singular integrals are operators that involve kernels with singularities, yet are bounded on \(L^p\) spaces.
3.1.1 Calderón–Zygmund theory
Calderón–Zygmund operators are defined by kernels satisfying certain size and smoothness conditions. The theory provides \(L^p\) boundedness for \(1<p<\infty\) and weak-type (1,1) bounds.
3.1.1.1 The Hilbert transform
The Hilbert transform on the real line is given by the principal value integral \(Hf(x) = \frac{1}{\pi} \int_{-\infty}^{\infty} \frac{f(y)}{x-y} dy\). It is a model Calderón–Zygmund operator and is bounded on \(L^p\) for \(1<p<\infty\).
3.1.1.2 Riesz transforms
| Riesz transforms are higher-dimensional analogs of the Hilbert transform. They appear in studying harmonic functions and have kernel \(K_j(x) = c_n x_j / | x | ^{n+1}\). Together they decompose the gradient into a vector of singular integrals. |
|---|
3.1.2 Boundedness in \(L^p\) and weak-type estimates
The Calderón–Zygmund decomposition splits a function into "good" and "bad" parts, leading to weak-type (1,1) estimates. Interpolation then yields \(L^p\) boundedness for all \(1<p<\infty\). The endpoint \(p=\infty\) is replaced by BMO (bounded mean oscillation) spaces.
3.2 Time-frequency analysis
Time-frequency analysis studies signals simultaneously in time and frequency, capturing non-stationary behavior.
3.2.1 Short-time Fourier transform
The short-time Fourier transform (STFT) multiplies a signal by a sliding window function before taking the Fourier transform. It yields a two-dimensional representation that shows how frequency content evolves over time, but its resolution is limited by the uncertainty principle.
3.2.2 Wavelet analysis
Wavelet analysis uses basis functions that are localised in both time and frequency, allowing multiresolution decomposition.
3.2.2.1 Continuous wavelet transform
The continuous wavelet transform (CWT) convolves a signal with scaled and translated versions of a mother wavelet. It provides a redundant representation that can be used for singularity detection and scale analysis.
3.2.2.2 Multiresolution analysis
Multiresolution analysis (MRA) constructs orthonormal wavelet bases through a hierarchy of approximation spaces. The scaling function and wavelet satisfy two-scale equations; the fast wavelet transform enables efficient computation.
3.3 Non-linear harmonic analysis
Non-linear harmonic analysis adapts linear techniques to study non-linear differential equations.
3.3.1 Littlewood–Paley theory
Littlewood–Paley theory decomposes a function into frequency blocks (dyadic pieces). It provides powerful tools for studying function spaces such as Sobolev and Besov spaces, and is essential for treating non-linear PDEs.
3.3.2 Para-differential operators
Para-differential operators are a calculus that linearizes non-linear operators by separating high and low frequencies. Developed by Bony and others, they are used to handle quasi-linear and fully non-linear PDEs.
4 Applications
Harmonic analysis finds applications across science and engineering, wherever data or physical processes involve oscillations or waves.
4.1 Signal processing
Fourier transforms and wavelets are fundamental tools for filtering, compression, and feature extraction in audio, image, and video processing. The DFT and FFT enable efficient spectral analysis.
4.2 Partial differential equations
Harmonic analysis underpins the study of PDEs through Fourier methods: solving the heat equation by convolution with the Gauss kernel, analyzing dispersive equations such as Schrödinger, and estimating solutions via Littlewood–Paley decompositions.
4.3 Quantum mechanics
In quantum mechanics, the Fourier transform connects position and momentum representations. The Heisenberg uncertainty principle is a consequence of the time-frequency uncertainty law. Spectral theory of operators relies heavily on harmonic analysis.
4.4 Data science and machine learning
Harmonic analysis methods are used in data science for dimensionality reduction (e.g., graph Laplacians), manifold learning, and analyzing functions on graphs. Wavelets and scattering transforms provide stable representations for classification and regression tasks.