Harmonic Analysis: Fourier Series, Transforms, and Singular Integrals

1. Fourier Series on the Circle: L^2 Theory

The circle group T is identified with the interval (-pi, pi] with endpoints identified, or equivalently with the unit circle in the complex plane. Every function in L^2(T) has a Fourier series expansion in terms of the orthonormal basis of complex exponentials.

Fourier Coefficients and Series

For f in L^2(T), the n-th Fourier coefficient is defined by the inner product of f with the character e^(inx). The resulting series is the Fourier series of f.

Parseval's Identity and L^2 Completeness

The system (e^(inx)) forms a complete orthonormal basis for L^2(T). Completeness means that the partial sums S_N(f) = sum from n = -N to N of c_n e^(inx) converge to f in the L^2 norm. Parseval's identity, which says the L^2 norm of f equals the l^2 norm of its coefficient sequence, is the content of this completeness statement. The map f to (c_n) is an isometric isomorphism from L^2(T) to l^2(Z).

Dirichlet Kernel and Pointwise Convergence

The N-th partial sum S_N(f)(x) equals the convolution of f with the Dirichlet kernel D_N(x) = sum from n = -N to N of e^(inx). By summing the geometric series, D_N(x) = sin((N + 1/2)x) divided by sin(x/2). The Dirichlet kernel has unit integral but its L^1 norm grows like log(N), which is why pointwise convergence fails in general. The Riemann-Lebesgue lemma says c_n(f) tends to 0 as n tends to infinity for any f in L^1(T).

Fejer's Theorem and Cesaro Summability

While partial sums may diverge pointwise, the Cesaro means sigma_N(f) = (1/(N+1)) times sum from k = 0 to N of S_k(f) behave much better. These averages equal the convolution of f with the Fejer kernel F_N(x) = (1/(N+1)) times |D_N(x)|^2, which is non-negative and has unit integral. Fejer's theorem states:

Fejer's theorem has a remarkable corollary: the trigonometric polynomials are dense in C(T) in the uniform norm (by Fejer), and hence dense in L^p(T) for all 1 less than or equal to p less than infinity. This is the harmonic analysis analog of the Weierstrass approximation theorem.

Pointwise Convergence: Dini and Jordan Conditions

The Fourier series of f converges to f(x) pointwise at x under various local regularity conditions. The Dini condition requires that the integral from 0 to delta of |f(x+t) + f(x-t) - 2f(x)| / t dt be finite. The Jordan condition applies when f is of bounded variation in a neighborhood of x: the partial sums converge to (f(x+) + f(x-))/2. However, Kolmogorov constructed an L^1 function whose Fourier series diverges everywhere — pointwise convergence for L^1 functions is not automatic.

2. Fourier Transform on R

The Fourier transform extends the idea of frequency decomposition from periodic functions to functions on the entire real line. It converts a function of time (or position) into a function of frequency, and underpins virtually all of modern signal processing and PDE theory.

Definition and Basic Properties

The Fourier transform is well defined for f in L^1(R) and the result f-hat is continuous and vanishes at infinity (Riemann-Lebesgue lemma). The inversion formula holds when both f and f-hat are in L^1. Key algebraic properties: the Fourier transform converts differentiation to multiplication by 2pi i xi, and converts convolution to pointwise multiplication.

Convolution Theorem

The convolution of two functions f and g is defined as the integral of f(y) times g(x-y) over all y. The convolution theorem states that the Fourier transform of a convolution is the pointwise product of the transforms:

The convolution theorem is fundamental in signal processing: passing a signal through a linear time-invariant filter corresponds to multiplying in the frequency domain. The transfer function of the filter is the Fourier transform of its impulse response.

Plancherel Theorem

For f in L^1 intersect L^2, the Fourier transform satisfies the isometry property: the L^2 norm of f-hat equals the L^2 norm of f. The Plancherel theorem asserts that the Fourier transform extends uniquely to a unitary automorphism of L^2(R).

Heisenberg Uncertainty Principle

A function and its Fourier transform cannot both be highly concentrated. The mathematical uncertainty principle quantifies this trade-off precisely.

The proof uses integration by parts and the Cauchy-Schwarz inequality. The key step is that x times d/dx of |f|^2 integrates to -||f||_2^2, which follows from the divergence theorem (or integration by parts).

3. Tempered Distributions and Schwartz Space

To extend the Fourier transform to objects like the Dirac delta function or to functions that grow at infinity, we work in the framework of distributions introduced by Laurent Schwartz.

The Schwartz Space S(R^n)

The Schwartz space consists of smooth functions that decay rapidly together with all their derivatives. Formally, phi is in S(R^n) if phi is infinitely differentiable and for every pair of multi-indices alpha and beta, the seminorm sup over x of |x^alpha times D^beta phi(x)| is finite. Informally, Schwartz functions decrease faster than any polynomial, as do all their derivatives.

The Fourier transform is an automorphism of S(R^n): if phi is in S then phi-hat is in S. This is the key advantage of Schwartz space for Fourier analysis. The inversion formula holds unconditionally in S.

Tempered Distributions S'(R^n)

A tempered distribution is a continuous linear functional on S(R^n). The space of tempered distributions is denoted S'(R^n) and is the topological dual of S. Every function of polynomial growth defines a tempered distribution by integration. More singular objects like the Dirac delta are also tempered distributions.

Distributional Fourier Transform

For u in S'(R^n), the Fourier transform u-hat is defined by duality: u-hat(phi) = u(phi-hat) for all phi in S. This extends the Fourier transform to all tempered distributions. Derivatives of distributions are also defined by duality: (D^alpha u)(phi) = (-1)^|alpha| u(D^alpha phi). Key examples include: the derivative of delta is -delta', and the principal value distribution p.v.(1/x) (which is the distributional kernel of the Hilbert transform) has Fourier transform -i pi sign(xi).

4. Hardy-Littlewood Maximal Function

The Hardy-Littlewood maximal function is a fundamental tool for controlling pointwise behavior of functions through averages. It appears in the proof of Lebesgue's differentiation theorem and in the theory of singular integrals.

Definition

Weak-Type (1,1) Bound

The maximal function is not L^1-bounded (taking the constant function 1 on R shows Mf = infinity). However, it satisfies a weaker substitute: the weak-type (1,1) inequality.

The proof of the weak-type bound uses the Vitali covering lemma: from any collection of balls, extract a subcollection of disjoint balls whose 3-fold dilations cover the original collection. This geometric lemma is fundamental to all of real-variable harmonic analysis.

Vitali Covering Lemma

Marcinkiewicz Interpolation and Strong-Type Bounds

The Marcinkiewicz interpolation theorem provides a soft method to upgrade weak-type estimates to strong-type estimates. Given a sublinear operator T that is weak-type (1,1) and bounded on L^infinity, Marcinkiewicz interpolation yields that T is bounded on L^p for all 1 less than p less than infinity. Applied to the maximal function (which is trivially L^infinity bounded by the L^infinity norm of f), this gives the strong-type (p,p) bound for p greater than 1.

5. Hilbert Transform and Riesz Transforms

The Hilbert transform is the prototypical example of a singular integral operator: its kernel 1/(pi x) is not integrable, yet the operator is bounded on L^2 and L^p. It plays a central role in complex analysis, signal processing, and the theory of Calderon-Zygmund operators.

Definition of the Hilbert Transform

The Fourier multiplier formula shows immediately that H is bounded on L^2 with norm 1 (since |sign(xi)| = 1), and that H composed with H is -I (the negative identity on functions with zero mean). The operator H is its own inverse up to sign: H^2 = -I.

L^p Boundedness and the M. Riesz Theorem

Marcel Riesz proved that the Hilbert transform extends to a bounded operator on L^p(R) for all 1 less than p less than infinity. The norm ||H||_(L^p to L^p) grows like p as p tends to infinity and like 1/(p-1) as p tends to 1. The Hilbert transform does not extend to a bounded operator on L^1 or L^infinity, but it satisfies a weak-type (1,1) bound.

Riesz Transforms in Higher Dimensions

The Riesz transforms R_j, for j = 1, ..., n, are the natural n-dimensional generalizations of the Hilbert transform. They are defined by the Fourier multiplier -i xi_j / |xi|.

Connection to Complex Analysis

The Hilbert transform on R arises naturally from the Cauchy integral formula. If F = u + iv is an analytic function in the upper half-plane with boundary values u on R, then v = Hu. The pair (u, v) is a conjugate pair satisfying the Cauchy-Riemann equations. This connection explains why the Hilbert transform is bounded on L^p: it follows from the L^p theory of harmonic conjugates.

6. Calderon-Zygmund Theory

Calderon-Zygmund theory provides a unified framework for studying singular integral operators far beyond the Hilbert and Riesz transforms. The theory originated in the 1950s with Calderon and Zygmund's study of second-order elliptic PDEs, and was vastly generalized by subsequent work including the T(1) theorem.

Calderon-Zygmund Kernels

A Calderon-Zygmund (CZ) kernel K on R^n is a function defined away from the origin satisfying size and smoothness conditions:

The Calderon-Zygmund Decomposition

The fundamental tool in the theory is the Calderon-Zygmund decomposition: given f in L^1 and a height lambda, decompose f = g + b where g (the "good" part) satisfies ||g||_1 is less than or equal to ||f||_1 and ||g||_infinity is less than or equal to C lambda, and b (the "bad" part) is supported on a union of disjoint cubes Q_j with total measure at most C||f||_1/lambda, with each b_j = b restricted to Q_j having mean zero.

BMO Space and the Sharp Maximal Function

The space BMO (functions of Bounded Mean Oscillation) consists of locally integrable functions f such that the "sharp maximal function" is bounded. The John-Nirenberg theorem gives precise exponential integrability for BMO functions.

T(1) Theorem of David-Journe

The T(1) theorem gives necessary and sufficient conditions for a singular integral operator T to be bounded on L^2. The conditions are: (1) T(1) is in BMO, (2) T*(1) is in BMO, and (3) T satisfies the weak boundedness property (a testing condition on indicator functions of balls). This theorem unified many results in the theory and led to the T(b) theorem and broader developments.

7. Littlewood-Paley Theory

Littlewood-Paley theory provides a powerful method for decomposing functions into frequency bands and measuring their size. It is the key tool for proving Fourier multiplier theorems and for defining modern function spaces like Sobolev, Besov, and Triebel-Lizorkin spaces.

Dyadic Decomposition and the Partition of Unity

Choose a smooth function psi on R^n supported in the annulus (1/2) less than |xi| less than 2, with sum over j in Z of psi(2^(-j) xi) = 1 for all xi not equal to 0. Define the j-th Littlewood-Paley piece of f by P_j f, whose Fourier transform is psi(2^(-j) xi) times f-hat(xi). This localizes f to frequencies of size approximately 2^j.

Multiplier Theorems: Mikhlin and Hormander

A key application of Littlewood-Paley theory is proving that certain Fourier multiplier operators are bounded on L^p. A Fourier multiplier operator T_m is defined by (T_m f)-hat(xi) = m(xi) f-hat(xi).

Besov and Triebel-Lizorkin Spaces

The Littlewood-Paley pieces P_j f measure the frequency content of f at scale 2^j. By controlling these pieces in different ways, one obtains different function spaces.

8. Wavelets and Multiresolution Analysis

Wavelets combine the time-frequency localization of Fourier analysis with a recursive structure that makes them especially suited for analyzing signals at multiple scales. The theory was developed in the 1980s by Daubechies, Mallat, Meyer, and others, and immediately found applications in image compression, numerical analysis, and signal processing.

The Haar Wavelet

The simplest wavelet, introduced by Haar in 1910, is defined as: psi(x) = 1 on [0, 1/2), -1 on [1/2, 1), and 0 otherwise. The Haar system consists of all dilates and translates psi_(j,k)(x) = 2^(j/2) psi(2^j x - k) for j, k in Z. These form an orthonormal basis for L^2(R). The Haar wavelet is discontinuous; constructing smooth wavelets requires the framework of multiresolution analysis.

Multiresolution Analysis (MRA)

A multiresolution analysis is a sequence of closed subspaces V_j of L^2(R) satisfying:

Given an MRA, the wavelet psi is constructed as an element of W_0 = V_1 ominus V_0 (the orthogonal complement of V_0 in V_1). Since V_1 is generated by phi(2x - k), and the scaling function satisfies phi(x) = sum_k h_k phi(2x - k) for some filter coefficients h_k (the refinement equation), the wavelet is psi(x) = sum_k (-1)^k h_(1-k) phi(2x - k).

Daubechies Wavelets and Regularity

Ingrid Daubechies constructed a family of wavelets with compact support and increasing regularity. The Daubechies-N wavelet has N vanishing moments (integral of x^k psi(x) dx = 0 for k = 0, ..., N-1), which gives approximation order N and implies that smooth function coefficients decay rapidly. The Daubechies-1 wavelet is the Haar wavelet; Daubechies-2 and higher are smoother and better suited for compression applications.

Applications to Compression and Signal Processing

9. Pontryagin Duality and Abstract Fourier Analysis

Classical Fourier analysis on R and on the circle T can both be understood as special cases of a general theory of harmonic analysis on locally compact abelian (LCA) groups. Pontryagin duality places these examples in a unified framework and reveals the deep symmetry between a group and its dual.

Characters and the Dual Group

Let G be a locally compact abelian group. A character of G is a continuous group homomorphism from G to the circle group T = (z in C : |z| = 1). The set of all characters of G, denoted G-hat, forms a group under pointwise multiplication and carries a natural topology (the compact-open topology) that makes it locally compact abelian as well. G-hat is called the Pontryagin dual of G.

Pontryagin Duality Theorem

The Pontryagin duality theorem asserts that the canonical map from G to its double dual G-hat-hat is an isomorphism of topological groups. This is the exact analog of the finite-dimensional statement that a vector space is canonically isomorphic to its double dual, but it holds in full generality for all LCA groups.

Haar Measure

Every locally compact group G admits a left-invariant Borel measure, unique up to positive scalar multiple, called the Haar measure. For R^n, Haar measure is Lebesgue measure. For the circle T, it is arc length measure normalized to have total mass 1. For a finite group, it is counting measure divided by the order of the group. The existence of Haar measure (proved by Haar in 1933, with uniqueness by von Neumann) is what makes integration and Fourier analysis possible on general LCA groups.

10. Applications: PDEs, Signal Processing, and Number Theory

Harmonic analysis is not merely an abstract theory — it provides the computational and conceptual machinery for solving PDEs, compressing and transmitting data, and even proving results in analytic number theory.

Heat Equation via Fourier Transform

Consider the heat equation on R: partial_t u = Delta u, with initial condition u(x, 0) = f(x). Taking the Fourier transform in the space variable x converts the PDE to an ODE in t for each frequency xi:

Wave Equation and Huygens' Principle

For the wave equation partial_tt u = Delta u on R^n, the Fourier transform gives u-hat(xi, t) = f-hat(xi) cos(2 pi |xi| t) + g-hat(xi) sin(2 pi |xi| t) / (2 pi |xi|), where f and g are initial position and velocity. In odd dimensions at least 3, the solution at time t depends only on data on the sphere of radius t (Huygens' principle); in even dimensions, it depends on data in the entire ball.

Signal Processing: Sampling and Reconstruction

The Shannon-Nyquist sampling theorem, a fundamental result of signal processing, is a direct application of Fourier analysis. A band-limited signal — one whose Fourier transform vanishes for |xi| greater than B — is completely determined by its values at the sample points n/(2B) for n in Z. Reconstruction is given by the Whittaker-Shannon interpolation formula using sinc functions.

Number Theory: Hardy-Ramanujan and the Circle Method

Fourier analysis on the circle (equivalently, the theory of exponential sums) is a central tool in analytic number theory. The Hardy-Ramanujan circle method, developed in 1918, uses Fourier series on the circle to extract asymptotic formulas for arithmetic quantities from generating functions.

The circle method was later applied by Hardy, Littlewood, and Vinogradov to Goldbach's problem (every even integer is a sum of two primes) and Waring's problem (every positive integer is a sum of at most g(k) perfect k-th powers). The key analytical tool is the estimation of exponential sums sum of e^(2 pi i n alpha) over primes or other arithmetic sequences, which requires deep estimates from both harmonic analysis and analytic number theory.

Practice Problems with Solutions

Exam Tips and Common Mistakes

Key Theorems at a Glance

Related Topics

Further Reading