Sources and References

Primary textbook — Katznelson, Y. An Introduction to Harmonic Analysis, 3rd ed. Cambridge University Press, 2004. Supplementary texts — Rudin, W. Functional Analysis, 2nd ed. McGraw-Hill, 1991. Folland, G.B. Real Analysis: Modern Techniques and Their Applications, 2nd ed. Wiley, 1999. Stein, E.M. and Shakarchi, R. Fourier Analysis: An Introduction, Princeton Lectures in Analysis, Vol. 1, 2003. Online resources — MIT OCW 18.103 Fourier Analysis; Tao, T. lecture notes on harmonic analysis; Stein & Shakarchi Fourier Analysis (Princeton Lectures).

---

Chapter 1: Fourier Series — Foundations

Historical Origins: Heat, Controversy, and a Revolution in Mathematics

The story of Fourier analysis begins in 1807, when Joseph Fourier submitted a memoir to the French Academy of Sciences titled Mémoire sur la propagation de la chaleur dans les corps solides (“Memoir on the propagation of heat in solid bodies”). The manuscript contained a claim that would ignite decades of mathematical controversy: any function defined on an interval could be represented as an infinite sum of sines and cosines.

This was not merely a technical assertion. It was a philosophical bombshell. The eighteenth century had been dominated by the view, championed by Euler and d’Alembert, that “arbitrary” functions and “analytic” (formula-defined) functions were categorically different objects. Fourier’s claim — that a step function, a function with corners, even a function defined piecewise by completely unrelated formulas, could be expressed as a sum of smooth sinusoidal waves — seemed to violate the very nature of these functions. Lagrange, one of the greatest mathematicians alive, was unconvinced. He had worked on similar problems years earlier and was skeptical that the series could converge to discontinuous functions. The Academy declined to publish the memoir.

Fourier persisted. In 1822, he published the full theory as Théorie analytique de la chaleur (“The Analytic Theory of Heat”). He was right about the physics — and mostly right about the mathematics. The precise question of when and in what sense the Fourier series of a function converges to that function would take another century to fully resolve, driving the development of:

• Measure theory (Lebesgue, 1901) — because “almost everywhere” convergence requires a precise notion of sets of measure zero;
• Functional analysis — because the convergence question in \(L^2\) required the theory of Hilbert spaces;
• Set theory and descriptive set theory — Cantor invented transfinite ordinals while working on sets of uniqueness for Fourier series;
• The Riemann integral itself — Riemann’s 1854 habilitation thesis defined the Riemann integral specifically to study Fourier series.

In short, Fourier’s 1807 memoir did not merely solve a problem in heat conduction. It created modern analysis. We study Fourier analysis not just for its applications but because it is the lens through which the structure of \(L^p\) spaces, Hilbert spaces, distribution theory, and harmonic analysis came into focus.

The Circle Group and Periodic Functions

The natural domain for Fourier series is the circle group \(\mathbb{T} = \mathbb{R}/2\pi\mathbb{Z}\), which we identify with the interval \([-\pi, \pi)\) equipped with addition modulo \(2\pi\). A function \(f : \mathbb{R} \to \mathbb{C}\) is \(2\pi\)-periodic if \(f(x + 2\pi) = f(x)\) for all \(x \in \mathbb{R}\). The natural function spaces on \(\mathbb{T}\) are the Lebesgue spaces \(L^p(\mathbb{T})\) for \(1 \leq p \leq \infty\), defined with respect to normalized Lebesgue measure \(\frac{dx}{2\pi}\).

The circle group \(\mathbb{T}\) is simultaneously a group, a compact topological space, and a smooth manifold. This triple structure — algebraic, topological, and geometric — is what makes Fourier analysis on \(\mathbb{T}\) so rich. The “correct” measure on \(\mathbb{T}\) for group-theoretic purposes is the one that is translation-invariant: the Haar measure, which here is simply normalized arc length \(\frac{dx}{2\pi}\). The normalization ensures \(\int_\mathbb{T} 1\, d\mu = 1\), making \(\mathbb{T}\) a probability space and the Fourier transform a unitary operation.

\[ \langle e_n, e_m \rangle = \frac{1}{2\pi}\int_{-\pi}^{\pi} e^{inx}\overline{e^{imx}}\,dx = \frac{1}{2\pi}\int_{-\pi}^{\pi} e^{i(n-m)x}\,dx = \delta_{nm}. \]

This orthonormality is the computational backbone of Fourier analysis: it means the exponentials are “independent” in an inner product sense, and the Fourier coefficients can be extracted by projecting onto each character. The fact that these characters form a complete orthonormal set — equivalently, that their finite linear combinations are dense in \(L^2(\mathbb{T})\) — is the content of the completeness theorem, which we prove later via Fejér’s theorem.

Fourier Coefficients and Fourier Series

The tilde notation \(\sim\) is deliberate: the series need not converge to \(f\) pointwise. The central question of Fourier analysis is precisely when and in what sense the partial sums recover \(f\).

\[ \frac{1}{2\pi}\int_{-\pi}^\pi f(x) e^{-imx}\,dx = \sum_n c_n \underbrace{\frac{1}{2\pi}\int_{-\pi}^\pi e^{i(n-m)x}\,dx}_{= \delta_{nm}} = c_m. \]

This orthogonality computation shows that if \(f\) has a Fourier series, then its coefficients must be given by the formula above. The subtle issue is going the other direction: given the coefficients, does the series converge back to \(f\)?

Worked Examples: Computing Fourier Coefficients

Let us compute the Fourier coefficients of several fundamental functions. These are not just exercises — each computation illuminates a different aspect of the theory, and several lead to famous classical identities.

Example 1: The Sawtooth Wave \(f(x) = x\)

Consider \(f(x) = x\) on \([-\pi, \pi)\), extended to be \(2\pi\)-periodic. This function is odd (\(f(-x) = -f(x)\)), which immediately tells us that its Fourier series involves only sines (the imaginary parts of \(e^{inx}\)).

Computation of \(\hat{f}(0)\): Since \(f\) is odd, \(\hat{f}(0) = \frac{1}{2\pi}\int_{-\pi}^\pi x\,dx = 0\).

\[ \hat{f}(n) = \frac{1}{2\pi}\int_{-\pi}^{\pi} x e^{-inx}\,dx. \]\[ \hat{f}(n) = \frac{1}{2\pi}\left[\frac{x e^{-inx}}{-in}\right]_{-\pi}^{\pi} - \frac{1}{2\pi}\int_{-\pi}^{\pi}\frac{e^{-inx}}{-in}\,dx. \]

The boundary term: \(\frac{1}{2\pi}\cdot\frac{\pi e^{-in\pi} - (-\pi)e^{in\pi}}{-in} = \frac{1}{2\pi}\cdot\frac{\pi(e^{-in\pi} + e^{in\pi})}{-in} = \frac{1}{2\pi}\cdot\frac{2\pi\cos(n\pi)}{-in} = \frac{\cos(n\pi)}{-in} = \frac{(-1)^n}{-in}.\)

The remaining integral: \(-\frac{1}{2\pi}\cdot\frac{1}{-in}\int_{-\pi}^\pi e^{-inx}\,dx = 0\) since \(\int_{-\pi}^\pi e^{-inx}\,dx = 0\) for \(n \neq 0\).

\[ \hat{f}(n) = \frac{(-1)^n}{-in} = \frac{(-1)^{n+1}}{in} = \frac{(-1)^{n+1}\cdot (-i)}{n} = \frac{i(-1)^n \cdot (-1)}{n} \]\[ \hat{f}(n) = \frac{(-1)^n}{-in} = \frac{(-1)^n}{-in}\cdot\frac{i}{i} = \frac{i(-1)^n}{n}. \]

But we know \(f\) is real and odd, so \(\hat{f}(n)\) should be purely imaginary and odd in \(n\). Indeed: \(\hat{f}(n) = \frac{i(-1)^n}{n}\). Since \(\hat{f}(-n) = \frac{i(-1)^{-n}}{-n} = \frac{-i(-1)^n}{n} = -\hat{f}(n)\) (noting \((-1)^{-n} = (-1)^n\)), this is consistent.

\[ x \sim \sum_{n \neq 0} \frac{i(-1)^n}{n} e^{inx} = 2\sum_{n=1}^\infty \frac{(-1)^{n+1}}{n}\sin(nx), \]

where we combined the \(n\) and \(-n\) terms: \(\frac{i(-1)^n e^{inx}}{n} + \frac{-i(-1)^n e^{-inx}}{n} = \frac{2(-1)^n \sin(nx)}{-n} = \frac{2(-1)^{n+1}\sin(nx)}{n}\).

\[ x = 2\left(\sin x - \frac{\sin 2x}{2} + \frac{\sin 3x}{3} - \frac{\sin 4x}{4} + \cdots\right), \quad x \in (-\pi, \pi). \]\[ \frac{\pi}{4} = 1 - \frac{1}{3} + \frac{1}{5} - \frac{1}{7} + \cdots \]

Example 2: The Square Wave

Let \(f(x) = \text{sgn}(x) = \begin{cases} 1 & 0 < x < \pi \\ -1 & -\pi < x < 0 \end{cases}\), extended \(2\pi\)-periodically. This is an odd function, so again only sine terms appear.

\[ \hat{f}(n) = \frac{1}{2\pi}\int_{-\pi}^{\pi} f(x) e^{-inx}\,dx = \frac{1}{2\pi}\left(\int_0^\pi e^{-inx}\,dx - \int_{-\pi}^0 e^{-inx}\,dx\right). \]

Computing: \(\int_0^\pi e^{-inx}\,dx = \frac{1-e^{-in\pi}}{in} = \frac{1-(-1)^n}{in}\). By oddness, \(\int_{-\pi}^0 e^{-inx}\,dx = -\frac{1-(-1)^n}{in}\).

So \(\hat{f}(n) = \frac{1}{2\pi}\cdot\frac{2(1-(-1)^n)}{in} = \frac{1-(-1)^n}{\pi i n}\).

\[ \text{sgn}(x) \sim \frac{4}{\pi}\left(\sin x + \frac{\sin 3x}{3} + \frac{\sin 5x}{5} + \cdots\right). \]

This is the Fourier series of the square wave. It only has odd harmonics — a consequence of the half-wave symmetry \(f(x+\pi) = -f(x)\). The series converges to \(f(x)\) at all points where \(f\) is continuous (which is everywhere except \(0, \pm\pi\)), converges to \(0 = \frac{f(0^+)+f(0^-)}{2}\) at the jump discontinuities, but exhibits the Gibbs phenomenon near the jumps (see Section 1.6 below).

Example 3: The Triangular Wave \(f(x) = |x|\)

The function \(f(x) = |x|\) on \([-\pi, \pi]\) is even (\(f(-x) = f(x)\)), so its Fourier series involves only cosines.

Computation of \(\hat{f}(0)\): \(\hat{f}(0) = \frac{1}{2\pi}\int_{-\pi}^\pi |x|\,dx = \frac{1}{2\pi}\cdot 2\int_0^\pi x\,dx = \frac{1}{2\pi}\cdot\pi^2 = \frac{\pi}{2}\).

\[ \hat{f}(n) = \frac{1}{2\pi}\int_{-\pi}^\pi |x|e^{-inx}\,dx = \frac{1}{\pi}\int_0^\pi x\cos(nx)\,dx, \]\[ \int_0^\pi x\cos(nx)\,dx = \left[\frac{x\sin(nx)}{n}\right]_0^\pi - \int_0^\pi \frac{\sin(nx)}{n}\,dx = 0 + \frac{1}{n}\left[\frac{\cos(nx)}{n}\right]_0^\pi = \frac{\cos(n\pi) - 1}{n^2} = \frac{(-1)^n - 1}{n^2}. \]

Therefore \(\hat{f}(n) = \frac{(-1)^n - 1}{\pi n^2}\). When \(n\) is even, \(\hat{f}(n) = 0\). When \(n\) is odd, \(\hat{f}(n) = \frac{-2}{\pi n^2}\).

\[ |x| = \frac{\pi}{2} - \frac{4}{\pi}\left(\cos x + \frac{\cos 3x}{9} + \frac{\cos 5x}{25} + \cdots\right) = \frac{\pi}{2} - \frac{4}{\pi}\sum_{k=0}^\infty \frac{\cos((2k+1)x)}{(2k+1)^2}. \]\[ |\hat{f}(0)|^2 + 2\sum_{n=1}^\infty |\hat{f}(n)|^2 = \frac{\pi^2}{4} + 2\sum_{k=0}^\infty \frac{4}{\pi^2(2k+1)^4}. \]

Setting these equal: \(\frac{\pi^2}{3} = \frac{\pi^2}{4} + \frac{8}{\pi^2}\sum_{k=0}^\infty \frac{1}{(2k+1)^4}\), giving \(\sum_{k=0}^\infty \frac{1}{(2k+1)^4} = \frac{\pi^4}{96}\), and from this one can recover \(\sum_{n=1}^\infty \frac{1}{n^4} = \frac{\pi^4}{90}\).

Example 4: The Parabola \(f(x) = x^2\)

Consider \(f(x) = x^2\) on \([-\pi, \pi]\), extended periodically. This is an even function.

Computation of \(\hat{f}(0)\): \(\hat{f}(0) = \frac{1}{2\pi}\int_{-\pi}^\pi x^2\,dx = \frac{1}{2\pi}\cdot\frac{2\pi^3}{3} = \frac{\pi^2}{3}\).

\[ \hat{f}(n) = \frac{1}{\pi}\int_0^\pi x^2\cos(nx)\,dx = \frac{1}{\pi}\left[\frac{x^2\sin(nx)}{n}\right]_0^\pi - \frac{2}{\pi n}\int_0^\pi x\sin(nx)\,dx. \]\[ \frac{2}{\pi n}\int_0^\pi x\sin(nx)\,dx = \frac{2}{\pi n}\left(\left[-\frac{x\cos(nx)}{n}\right]_0^\pi + \frac{1}{n}\int_0^\pi\cos(nx)\,dx\right) = \frac{2}{\pi n}\cdot\frac{-\pi\cos(n\pi)}{n} = \frac{-2(-1)^n}{n^2}. \]\[ x^2 = \frac{\pi^2}{3} + 4\sum_{n=1}^\infty \frac{(-1)^n}{n^2}\cos(nx) = \frac{\pi^2}{3} - 4\cos x + \cos 2x - \frac{4\cos 3x}{9} + \frac{4\cos 4x}{16} - \cdots \]

Wait — let us recheck: \(\hat{f}(n) = \frac{2(-1)^n}{n^2}\), so the series is \(\frac{\pi^2}{3} + \sum_{n\neq 0}\frac{2(-1)^n}{n^2}e^{inx} = \frac{\pi^2}{3} + 4\sum_{n=1}^\infty \frac{(-1)^n}{n^2}\cos(nx)\).

\[ 0 = \frac{\pi^2}{3} + 4\sum_{n=1}^\infty \frac{(-1)^n}{n^2} = \frac{\pi^2}{3} - \frac{4\pi^2}{12} = \frac{\pi^2}{3} - \frac{\pi^2}{3}. \]\[ \pi^2 = \frac{\pi^2}{3} + 4\sum_{n=1}^\infty \frac{(-1)^n \cdot (-1)^n}{n^2} = \frac{\pi^2}{3} + 4\sum_{n=1}^\infty \frac{1}{n^2}. \]

Rearranging: \(\sum_{n=1}^\infty \frac{1}{n^2} = \frac{\pi^2}{6}\).

This is the famous Basel problem, solved by Euler in 1734 (he used a different method — factoring \(\sin(\pi x)/(\pi x)\) as a product over its zeros). Fourier’s method gives it as a byproduct of knowing the Fourier series of \(x^2\). One can also recover it from Parseval: \(\|x^2\|_{L^2}^2 = \frac{\pi^4}{5}\) and \(\sum|\hat{f}(n)|^2 = \frac{\pi^4}{9} + 2\sum_{n=1}^\infty \frac{4}{n^4}\), giving \(\sum \frac{1}{n^4} = \frac{\pi^4}{90}\) — Euler’s quartic formula as well.

Motivation: The Heat Equation on the Circle

\[ \frac{\partial u}{\partial t} = \frac{\partial^2 u}{\partial x^2}, \quad u(x,0) = f(x), \]

where \(f\) is the initial temperature distribution. This was Fourier’s original problem. He considered a thin metal ring (the circle \(\mathbb{T}\)) with an initial non-uniform temperature, and asked: how does the temperature evolve over time?

\[ X(x)T'(t) = X''(x)T(t) \implies \frac{T'(t)}{T(t)} = \frac{X''(x)}{X(x)} = -\lambda, \]

where \(\lambda\) is a constant (since the left side depends only on \(t\) and the right only on \(x\)). For periodic boundary conditions, \(X(x) = e^{inx}\) with \(\lambda = n^2\) solves \(X'' = -\lambda X\). Then \(T'(t) = -n^2 T(t)\), so \(T(t) = e^{-n^2 t}\). The separated solution is \(u_n(x,t) = e^{inx-n^2 t}\).

\[ u(x,t) = \sum_{n=-\infty}^{\infty} \hat{f}(n)e^{-n^2 t}e^{inx}. \]

For \(t > 0\), the factor \(e^{-n^2 t}\) provides rapid decay, making the series absolutely and uniformly convergent. This is the Poisson–Weierstrass kernel approach and illustrates how Fourier series decompose the evolution operator into independent modes.

What does this tell us physically? High-frequency oscillations (large \(|n|\)) are damped out extremely rapidly by the factor \(e^{-n^2 t}\), which decays much faster than the low-frequency ones. The “roughness” of the initial temperature distribution disappears instantaneously — for any \(t > 0\), no matter how rough \(f\) is, the solution \(u(\cdot, t)\) is a \(C^\infty\) function (indeed, real-analytic). The heat equation is a smoothing operator. This is in stark contrast to the wave equation, which preserves singularities.

The Weierstrass approximation theorem follows as an elegant corollary: since \(u(\cdot,t)\) is a trigonometric polynomial (a finite sum of exponentials, to any desired precision for small \(t\)) and converges uniformly to the continuous initial data \(f\), continuous periodic functions can be uniformly approximated by trigonometric polynomials.

The Riemann–Lebesgue Lemma

The Riemann–Lebesgue lemma tells us that Fourier coefficients must lie in \(c_0(\mathbb{Z})\) (sequences vanishing at infinity). This is a necessary condition, but far from sufficient: characterizing the image of \(L^1(\mathbb{T})\) under the Fourier transform is a difficult open problem. The image is strictly contained in \(c_0(\mathbb{Z})\) — for example, one can construct sequences in \(c_0(\mathbb{Z})\) that are not Fourier coefficients of any \(L^1\) function.

Interpretation: The Riemann–Lebesgue lemma says that rapid oscillations (high frequencies) are “averaged out” in the integral \(\int f(x) e^{-inx}\,dx\) as \(n \to \infty\). The factor \(e^{-inx}\) oscillates so rapidly for large \(n\) that the contributions from neighboring intervals nearly cancel. This is a form of the “principle of stationary phase” in its simplest incarnation.

---

Chapter 2: Convolutions and Summability Kernels

Convolution on the Torus

Convolution is commutative, associative, and bilinear. By Young’s convolution inequality, \(\|f*g\|_p \leq \|f\|_1 \|g\|_p\) for any \(1 \leq p \leq \infty\). The crucial algebraic property is that the Fourier transform converts convolution to pointwise multiplication: \(\widehat{f*g}(n) = \hat{f}(n)\hat{g}(n)\). This makes the Fourier transform a homomorphism of algebras from \((L^1(\mathbb{T}), *)\) into \((c_0(\mathbb{Z}), \cdot)\).

Why does convolution arise? In the heat equation, the solution at time \(t\) is \(u(\cdot, t) = f * K_t\) where \(K_t\) is a kernel depending only on time. This is because the heat equation is translation-invariant: shifting the initial data \(f\) just shifts the solution. Any translation-invariant linear operation on periodic functions must be a convolution (by a version of the representation theorem for such operators). Thus convolution is not an arbitrary algebraic construction — it is the fundamental structure of any linear, translation-invariant operation.

\[ \|f*g\|_p \leq \|f\|_1\|g\|_p. \]

Proof: by Minkowski’s inequality for integrals, \(\|f*g\|_p = \|\int f(t)g(\cdot - t)\frac{dt}{2\pi}\|_p \leq \int |f(t)|\|g(\cdot-t)\|_p\frac{dt}{2\pi} = \|g\|_p \|f\|_1\).

The Dirichlet Kernel

\[ S_N f(x) = \frac{1}{2\pi}\int_{-\pi}^{\pi} f(t) \sum_{n=-N}^{N} e^{in(x-t)}\,dt = (f * D_N)(x), \]

where the Dirichlet kernel is:

\[ D_N(x) = e^{-iNx}\sum_{k=0}^{2N} e^{ikx} = e^{-iNx}\cdot\frac{e^{i(2N+1)x}-1}{e^{ix}-1}. \]\[ D_N(x) = e^{-iNx}\cdot\frac{e^{i(2N+1/2)x} - e^{-ix/2}}{e^{ix/2} - e^{-ix/2}} = \frac{e^{i(N+1/2)x} - e^{-i(N+1/2)x}}{e^{ix/2} - e^{-ix/2}} = \frac{\sin((N+\frac{1}{2})x)}{\sin(x/2)}. \]

The Dirichlet kernel has the properties:

• \(\frac{1}{2\pi}\int_{-\pi}^\pi D_N(x)\,dx = 1\) (the constant term of \(D_N\) is 1).
• \(D_N\) is even.
• \(D_N(x) = 2N+1\) at \(x = 0\).
• \(D_N\) oscillates and can be positive or negative.

\[ \|D_N\|_{L^1} = \frac{1}{2\pi}\int_{-\pi}^{\pi} |D_N(x)|\,dx \sim \frac{4}{\pi^2}\log N \to \infty. \]\[ \|D_N\|_1 \geq C\int_0^\pi \frac{|\sin((N+\frac12)x)|}{x}\,dx \geq C\sum_{k=1}^N \int_{(k-1)\pi/(N+1/2)}^{k\pi/(N+1/2)} \frac{|\sin((N+\frac12)x)|}{x}\,dx \geq C\sum_{k=1}^N \frac{1}{k} \sim C\log N. \]

This divergence of the Lebesgue constants \(L_N = \|D_N\|_{L^1}\) is the root cause of all the pathological behavior of Fourier partial sums. By the Banach–Steinhaus theorem, since the operators \(f \mapsto S_N f(0)\) have operator norms \(\|S_N(\cdot)(0)\|_{L^1 \to \mathbb{C}} = L_N \to \infty\), there must exist continuous functions \(f\) for which \(\sup_N |S_N f(0)| = \infty\) — i.e., the Fourier series diverges at \(x = 0\).

The Gibbs Phenomenon

The Gibbs phenomenon refers to the persistent overshoot of Fourier partial sums near a jump discontinuity. It was observed by Henry Wilbraham in 1848, rediscovered and reported by J. Willard Gibbs in 1898, and fully explained by Maxime Bôcher in 1906.

Consider the partial sums \(S_N f\) of the sawtooth wave \(f(x) = x\) (which has a jump discontinuity of size \(2\pi\) at \(x = \pm\pi\)). The maximum of \(S_N f\) near the discontinuity exceeds the value \(\pi\) of the function itself by approximately a fixed fraction, even as (N \to \infty$.

\[ (S_N f)'(x) = 2\sum_{k=1}^N (-1)^{k+1}\cos(kx). \]

Using the formula for the sum of cosines, one finds \((S_N f)'(x) = D_N(x+\pi) - D_N(x)\) (up to factors). Setting this to zero to find critical points near \(x = 0^+\): the first zero of \((S_N f)'\) after 0 is at \(x_N = \frac{\pi}{N+1/2} \approx \frac{\pi}{N}\).

\[ S_N f(x_N) = 2\sum_{k=1}^N \frac{(-1)^{k+1}}{k}\sin\!\left(\frac{k\pi}{N+1/2}\right). \]\[ \lim_{N\to\infty} S_N f(x_N) = 2\int_0^\pi \frac{\sin t}{t}\,dt = 2 \cdot \text{Si}(\pi) \approx 2 \times 1.8519 \approx 3.7038 \cdots \]\[ \lim_{N\to\infty} S_N f\!\left(\pi - \frac{\pi}{N}\right) - \pi = 2\int_0^\pi \frac{\sin t}{t}\,dt - \pi \approx 3.7038 - \pi \approx 0.5619 \cdots \]

As a fraction of the jump (of size \(2\pi\)): \(\frac{0.5619}{2\pi} \approx 8.9\% \approx 9\%\).

This 9% overshoot is universal: it does not depend on the function, only on the presence of a jump discontinuity. The Gibbs phenomenon says that Fourier partial sums always overshoot jump discontinuities by about 9% of the jump size, and this overshoot does not diminish — only the width of the overshoot region (around \(\pi/N\)) shrinks.

Why does Gibbs happen? The Dirichlet kernel \(D_N\) has oscillatory lobes. When we compute \(S_N f = f * D_N\), the lobes of \(D_N\) near the discontinuity create artificial oscillations. The Cesàro means \(\sigma_N f = f * K_N\) avoid this because the Fejér kernel \(K_N \geq 0\) has no oscillatory lobes — it is non-negative, so it cannot create overshoots. This is the key advantage of Cesàro summation.

The Fejér Kernel and Cesàro Means

\[ \sigma_N f(x) = \frac{1}{N}\sum_{k=0}^{N-1} S_k f(x) = (f * K_N)(x), \]

where the Fejér kernel is:

\[ K_N(x) = \sum_{|n| \leq N-1}\left(1 - \frac{|n|}{N}\right)e^{inx}. \]

The Cesàro mean thus applies triangular weights \(1 - |n|/N\) to the Fourier coefficients, tapering them to zero as \(|n| \to N\). This is called a Fejér kernel or Bartlett window in signal processing.

For the closed form \(\frac{1}{N}\left(\frac{\sin(Nx/2)}{\sin(x/2)}\right)^2\): this follows from the identity \(\left(\sum_{k=0}^{N-1}e^{ikx}\right)^2 = N\sum_{k=0}^{N-1}D_k(x)\) after careful manipulation.

The Fejér kernel has three crucial properties:

1. Non-negativity: \(K_N(x) = \frac{1}{N}\left(\frac{\sin(Nx/2)}{\sin(x/2)}\right)^2 \geq 0\) for all \(x\). This is the key property that the Dirichlet kernel lacks.
2. Normalization: \(\frac{1}{2\pi}\int_{-\pi}^{\pi} K_N(x)\,dx = 1\).
3. Concentration: For any \(\delta > 0\), \(\int_{\delta \leq |x| \leq \pi} K_N(x)\,dx \to 0\) as \(N \to \infty\).

Proof of concentration: For \(|x| \geq \delta > 0\), \(|\sin(x/2)| \geq |\sin(\delta/2)| > 0\), so \(K_N(x) \leq \frac{1}{N\sin^2(\delta/2)} \to 0\) uniformly. Therefore \(\int_{|x|\geq\delta}K_N(x)\,dx \leq \frac{2\pi}{N\sin^2(\delta/2)} \to 0\).

These three properties define an approximate identity (or summability kernel).

Intuition: The Fejér kernel \(K_N\) looks like a bump of height \(\approx N\) and width \(\approx 2\pi/N\), with total area 1 (normalized), concentrated near 0, and non-negative. Convolution with \(K_N\) averages \(f\) over a window of width \(\approx 2\pi/N\). As \(N \to \infty\), the window shrinks to a point, so the average converges to the function value (at points of continuity).

Summability Kernels and the General Convergence Theorem

Fejér’s theorem is a cornerstone result. As an immediate corollary, the trigonometric polynomials are dense in \(C(\mathbb{T})\), which in turn implies that \(\{e^{inx}\}_{n \in \mathbb{Z}}\) is a complete orthonormal set in \(L^2(\mathbb{T})\). Another corollary: if \(\hat{f}(n) = 0\) for all \(n \in \mathbb{Z}\), then \(f = 0\) a.e. (the uniqueness theorem for Fourier series).

Historical context: Lipót Fejér proved this theorem in 1900, at the age of 19, as his dissertation work. It was a complete surprise — after du Bois-Reymond had shown in 1873 that there exist continuous functions whose Fourier partial sums diverge at a point, it seemed that Fourier series might be fundamentally flawed. Fejér showed that one need only average the partial sums to restore convergence. The beauty of the idea is that the Fejér kernel is non-negative — a fact that makes it behave much better than the oscillatory Dirichlet kernel.

The Poisson Kernel

Another important summability kernel is the Poisson kernel, which arises from the solution to Laplace’s equation on the disk.

\[ P_r(x) = \text{Re}\left(\frac{1+re^{ix}}{1-re^{ix}}\right) = \frac{1-r^2}{|1-re^{ix}|^2} = \frac{1-r^2}{1-2r\cos x + r^2}. \]

The Poisson kernel is non-negative, normalized (\(\frac{1}{2\pi}\int_{-\pi}^\pi P_r(x)\,dx = 1\)), and concentrates near \(x=0\) as \(r \to 1^-\). Thus \(\{P_r\}_{r\to 1^-}\) is a summability kernel (in the continuous parameter version).

Connection to harmonic analysis: If \(u(r,x) = (f * P_r)(x) = \sum_n \hat{f}(n) r^{|n|} e^{inx}\), then \(u\) is the harmonic extension of \(f\) to the unit disk (in polar coordinates). The Poisson kernel solves the Dirichlet problem for the disk: given continuous boundary data \(f\) on \(\partial\mathbb{D} \cong \mathbb{T}\), the harmonic extension to the interior is \(u(re^{ix}) = (f*P_r)(x)\), and \(u(re^{ix}) \to f(e^{ix})\) uniformly as \(r \to 1^-\).

The Weierstrass Kernel

\[ W_t(x) = \sum_{n=-\infty}^\infty e^{-n^2 t}e^{inx}, \quad t > 0. \]

This arises directly from the heat equation on \(\mathbb{T}\) (the Fourier multiplier \(e^{-n^2 t}\) represents the heat flow). As \(t \to 0^+\), \(\{W_t\}\) is a summability kernel. The Weierstrass approximation theorem — trigonometric polynomials are dense in \(C(\mathbb{T})\) — follows from the fact that for small \(t\), \(W_t * f\) is a trigonometric polynomial (to any desired precision) converging uniformly to \(f\).

---

Chapter 3: Convergence of Fourier Series

Why \(L^2\) is the Right Space

Before studying convergence, it is worth asking: why do we focus so much on \(L^2\)? There are several compelling reasons.

Hilbert space completeness: \(L^2(\mathbb{T})\) is a Hilbert space — it has an inner product, and it is complete (every Cauchy sequence converges). This makes orthonormal expansions work perfectly: given an orthonormal system, the Fourier series always converges in norm. For \(L^1\) or other \(L^p\), we lack this inner product structure, and convergence is much harder to guarantee.

Physical motivation: In quantum mechanics, states are vectors in a Hilbert space, and the “square-integrability” of wavefunctions is a physical requirement (the norm-squared represents probability). In signal processing, energy is proportional to the square of amplitude, so \(L^2\) is the natural energy space.

The Riesz–Fischer theorem: In \(L^2\), there is a perfect bijection between functions and their Fourier coefficient sequences. This bijectivity fails in \(L^1\) (the image of the Fourier transform is not all of \(c_0\)) and in \(L^\infty\) (a bounded function’s Fourier series may not converge anywhere). In \(L^2\), the Fourier transform is an isometric isomorphism \(L^2(\mathbb{T}) \cong \ell^2(\mathbb{Z})\) — a perfect duality.

Parseval’s theorem: In \(L^2\), the energy of a function equals the sum of the squares of its Fourier coefficients. This is the mathematical content of the fact that frequencies are “independent” — they carry non-overlapping energy. This has no analog in \(L^1\).

Failure in \(L^1\): There exist \(f \in L^1(\mathbb{T})\) whose Fourier series diverges everywhere (Kolmogorov’s 1923 example). There exist \(f \in C(\mathbb{T})\) whose Fourier series diverges at a point (du Bois-Reymond, 1873). In \(L^2\), by Carleson’s theorem (1966), the Fourier series converges almost everywhere — but this required 60 years to prove after the convergence in \(L^2\)-norm was established.

Uniform Convergence

While Fejér’s theorem gives uniform convergence of Cesàro means for continuous \(f\), the partial sums themselves may diverge. To guarantee \(S_N f \to f\) uniformly, one needs additional smoothness.

A cleaner sufficient condition is the Dini-type condition: if \(f\) is \(\alpha\)-Hölder continuous (i.e., \(|f(x) - f(y)| \leq C|x-y|^\alpha\) for some \(\alpha > 0\)), then \(S_N f \to f\) uniformly. More generally, if \(f \in C^1(\mathbb{T})\), then \(|\hat{f}(n)| = O(1/n)\) and the Fourier series converges absolutely and uniformly.

Why smoothness helps: Integration by parts relates \(\hat{f}'(n)\) and \(n\hat{f}(n)\). If \(f\) is \(C^k\), then \(|\hat{f}(n)| = O(|n|^{-k})\), and the series \(\sum_n |\hat{f}(n)|\) converges absolutely for \(k \geq 2\), giving uniform convergence. For \(C^\infty\) functions, the Fourier series converges faster than any polynomial — it is an analytic “frequency-domain” representation of the function.

\(L^p\) Convergence

This follows from Hilbert space theory: the exponentials \(\{e^{inx}\}\) form a complete orthonormal basis of \(L^2(\mathbb{T})\), and \(S_N f\) is the orthogonal projection of \(f\) onto the span of \(\{e^{inx} : |n| \leq N\}\). In any Hilbert space, the partial sums of a Fourier expansion (with respect to a complete orthonormal basis) converge in norm — this is the content of Bessel’s inequality and completeness.

The proof of M. Riesz’s theorem requires deeper methods — specifically, the boundedness of the conjugate function operator (Hilbert transform) on \(L^p\) for \(1 < p < \infty\). This fails at the endpoints: the Hilbert transform is not bounded on \(L^1$ or \(L^\infty\).

Failure of \(L^1\) convergence: Du Bois-Reymond constructed a continuous function whose Fourier series diverges at a point. By a Baire category argument (Banach–Steinhaus theorem applied to the diverging Lebesgue constants), one can show that for a generic \(f \in C(\mathbb{T})\), the Fourier series diverges on a dense set.

Kolmogorov’s Example

In 1923, Andrei Kolmogorov (age 19, like Fejér before him) constructed an \(L^1\) function whose Fourier series diverges everywhere — not just at a point, but at every point. This was a shocking demonstration that \(L^1\) is genuinely the wrong space for pointwise convergence of Fourier series.

The construction is intricate but the idea is clear: one builds \(f\) as a sum \(f = \sum_k f_k\) where each \(f_k\) is a trigonometric polynomial chosen so that the partial sums of \(f\) are large at a dense set of points. The \(L^1\) norm of \(f\) is controlled because the \(f_k\) are spread out, but the partial sums blow up everywhere because the \(f_k\) conspire to create large values at different points at different scales.

Kolmogorov later (1926) strengthened his result to show an \(L^1\) function whose Fourier series diverges everywhere in a much stronger sense (the partial sums are unbounded almost everywhere). The complete picture for \(L^p\) spaces is:

• \(p = 1\): Fourier series can diverge everywhere (Kolmogorov).
• \(1 < p < \infty\): Fourier series converges in \(L^p\)-norm (M. Riesz), and even converges a.e. (Carleson–Hunt theorem, 1968).
• \(p = \infty\): \(L^\infty\) convergence fails (same as for \(L^1\) by duality arguments).
• \(C(\mathbb{T})\) (continuous functions): Fourier series converges a.e. (Carleson, 1966), but can diverge at individual points (du Bois-Reymond).

---

Chapter 4: Pointwise Convergence — Dini’s Theorem and Localization

Dini’s Test

Interpreting Dini’s condition: The condition asks that the “average oscillation” \(\frac{f(x_0+t)+f(x_0-t)-2s}{t}\) be integrable near 0. This is automatic if:

• \(f\) has a one-sided derivative at \(x_0\) from both sides (then the numerator is \(O(t)\), so the ratio is bounded);
• \(f\) is Hölder continuous at \(x_0\) (then the numerator is \(O(t^\alpha)\), so the ratio is \(O(t^{\alpha-1})\), integrable for \(\alpha > 0\));
• \(f\) satisfies a Lipschitz condition at \(x_0\).

In all these cases, \(S_N f(x_0) \to \frac{f(x_0^+)+f(x_0^-)}{2}\).

Dini’s test generalizes Dirichlet’s classical theorem: if \(f\) is Hölder continuous at \(x_0\) (say with exponent \(\alpha > 0\)) or has one-sided limits with finite one-sided derivatives, the Dini condition is satisfied with \(s = \frac{f(x_0^+) + f(x_0^-)}{2}\).

Riemann’s Localization Principle

A striking feature of pointwise convergence is its local nature.

The localization principle has the striking consequence that the convergence of \(S_N f(x_0)\) depends only on the behavior of \(f\) near \(x_0\) — a global operation (summing all Fourier coefficients) yields a result determined by local data. This is remarkable: the Fourier coefficients involve integrating \(f\) over the entire circle, yet their sum at a point depends only on what \(f\) does in an arbitrarily small neighborhood of that point.

Contrast with global phenomena: The failure of pointwise convergence (Kolmogorov’s example) does require global properties of \(f\), but any information about convergence at a specific point \(x_0\) is purely local. This is why one can make local modifications to \(f\) (in a small neighborhood of a point) without affecting convergence elsewhere.

Hardy’s Tauberian Theorem

Hardy’s theorem provides a striking bridge between summability and ordinary convergence.

The condition \(\hat{f}(n) = O(1/n)\) is a Tauberian condition — it provides the extra regularity that converts Cesàro convergence back to ordinary convergence. The proof uses summation by parts (Abel summation) and the estimate on the size of Fourier coefficients to control the difference \(S_N f - \sigma_N f\).

This theorem is a model for the general Tauberian theory developed by Hardy–Littlewood and Wiener, which plays a central role in analytic number theory (Wiener’s Tauberian theorem and the prime number theorem).

---

Chapter 5: Hilbert Space Theory and L² Analysis

\(L^2(\mathbb{T})\) as a Hilbert Space

The space \(L^2(\mathbb{T})\) with inner product \(\langle f, g \rangle = \frac{1}{2\pi}\int_{-\pi}^{\pi} f(x)\overline{g(x)}\,dx\) is a complete inner product space (Hilbert space). The norm is \(\|f\|_2 = \langle f, f\rangle^{1/2}\).

The completeness of \(L^2\) — the Riesz–Fischer theorem — was proved in 1907 simultaneously and independently by Frigyes Riesz and Ernst Fischer. This was a foundational result: it showed that \(L^2\) is a complete metric space (hence a Banach space, and with the inner product structure, a Hilbert space). The proof used the Lebesgue integral crucially — the Riemann integral does not give a complete space, because limits of Riemann-integrable functions need not be Riemann-integrable.

Completeness here means that the only element orthogonal to all \(e^{inx}\) is the zero function — equivalently, if \(\hat{f}(n) = 0\) for all \(n\) then \(f = 0\) in \(L^2\). This follows from Fejér’s theorem: if all Fourier coefficients vanish, then all Cesàro means vanish, and since they converge to \(f\) in \(L^2\), we have \(f = 0\).

General Hilbert space context: In any Hilbert space \(\mathcal{H}\) with complete orthonormal set \(\{e_n\}\), every element \(f \in \mathcal{H}\) has a Fourier expansion \(f = \sum_n \langle f, e_n\rangle e_n\) converging in norm, and the map \(f \mapsto \{\langle f, e_n\rangle\}_n\) is an isometric isomorphism from \(\mathcal{H}\) to \(\ell^2\). Applied to \(L^2(\mathbb{T})\) with \(e_n = e^{inx}\), this gives the fundamental theorem of Fourier series in \(L^2\).

Parseval’s and Plancherel’s Theorems

\[ \frac{\pi^2}{3} = \sum_{n \neq 0} \frac{1}{n^2} = 2\sum_{n=1}^\infty \frac{1}{n^2}, \]

so \(\sum_{n=1}^\infty \frac{1}{n^2} = \frac{\pi^2}{6}\). This is the most elegant proof of the Basel problem — arguably more natural than Euler’s original argument. It took Euler enormous effort (and a non-rigorous passage through infinite products) to establish this in 1734; Parseval’s identity (once the theory is set up) makes it a one-line calculation.

\[ \frac{\pi^4}{5} = \frac{\pi^4}{9} + 2\sum_{n=1}^\infty \frac{4}{n^4} \implies \sum_{n=1}^\infty\frac{1}{n^4} = \frac{\pi^4}{90}. \]\[ 1 = \sum_{k=-\infty}^\infty |\hat{f}(2k+1)|^2 = 2\sum_{k=0}^\infty \frac{4}{\pi^2(2k+1)^2} \implies \sum_{k=0}^\infty\frac{1}{(2k+1)^2} = \frac{\pi^2}{8}. \]

These computations illustrate the power of Parseval’s identity: the \(L^2\) theory of Fourier series transforms analytic number theory questions (sums of series) into geometric questions (computing norms in Hilbert spaces).

The Riesz–Fischer Theorem

The Riesz–Fischer theorem is a completeness statement — it says that every square-summable sequence of “frequencies” corresponds to an actual \(L^2\) function. This was a major result in the early 20th century, establishing a precise correspondence between function theory and sequence spaces.

Construction: Given \(\{c_n\} \in \ell^2(\mathbb{Z})\), define \(f_N = \sum_{|n|\leq N} c_n e^{inx}\). These are Cauchy in \(L^2\): \(\|f_M - f_N\|_{L^2}^2 = \sum_{N<|n|\leq M}|c_n|^2 \to 0\) since \(\{c_n\} \in \ell^2\). By completeness of \(L^2\), the limit \(f = \lim f_N\) exists in \(L^2\), and one verifies \(\hat{f}(n) = c_n\).

Convergence for Smooth Functions

Conversely, if \(\hat{f}(n)\) decays rapidly, then \(f\) is smooth. For \(C^\infty\) functions the Fourier series converges absolutely and uniformly: \(\sum_n |\hat{f}(n)| < \infty\).

The regularity ladder: The smooth-function theorem gives us a precise dictionary between function regularity and Fourier coefficient decay:

| Regularity of \(f\) | Decay of \(|\hat{f}(n)|\) | Convergence | |—|—|—| | \(f \in L^1(\mathbb{T})\) | \(o(1)\) (R-L lemma) | No pointwise guarantee | | \(f\) Hölder \(C^{0,\alpha}\) | \(O(n^{-\alpha})\) | Uniform | | \(f \in C^1(\mathbb{T})\) | \(O(1/n)\) | Absolute + uniform | | \(f \in C^k(\mathbb{T})\) | \(O(n^{-k})\) | Absolutely \(\sum n^{k-1}|\hat{f}(n)| < \infty\) | | \(f \in C^\infty(\mathbb{T})\) | Faster than any \(n^{-k}\) | Absolutely and rapidly | | \(f\) real-analytic | Exponential \(O(r^{-|n|})\), \(r > 1\) | Absolutely and exponentially |

This table encodes the fundamental insight: smoothness in the spatial domain corresponds to decay in the frequency domain. The Fourier transform is a “microscope” that reveals the regularity of functions through their frequency content.

---

Chapter 6: Interpolation Theorems

The Riesz–Thorin Interpolation Theorem

Many operators in Fourier analysis are bounded on two endpoint spaces but one wants to deduce boundedness on intermediate spaces. The Riesz–Thorin theorem makes this precise.

Application: Fourier transform on \(L^p(\mathbb{T})\): The Fourier transform (mapping \(f\) to its sequence of Fourier coefficients) is bounded from \(L^1 \to \ell^\infty\) (trivially: \(|\hat{f}(n)| \leq \|f\|_1\)) and from \(L^2 \to \ell^2\) (isometrically, by Plancherel). By Riesz–Thorin with \(\theta = 2/p' - 1\), it is bounded from \(L^p \to \ell^{p'}\) for \(1 \leq p \leq 2\). This is the Hausdorff–Young inequality.

The Hausdorff–Young Inequality

The Hausdorff–Young inequality is sharp: equality holds for Gaussian functions. The reverse inequality (Young’s inequality) states that if \(\{\hat{f}(n)\} \in \ell^p\) for \(1 \leq p \leq 2\), then \(f \in L^{p'}(\mathbb{T})\). Together these establish a duality between the integrability of a function and the summability of its Fourier coefficients.

---

Chapter 7: Fourier Transform on the Real Line

The Fourier Transform on \(L^1(\mathbb{R})\)

Moving from the compact group \(\mathbb{T}\) to the locally compact group \(\mathbb{R}\), the Fourier transform becomes an integral over all of \(\mathbb{R}\).

This convention (with the \(2\pi\) in the exponent rather than the measure) makes the inversion formula symmetric: \(f(x) = \int_{\mathbb{R}} \hat{f}(\xi)e^{2\pi ix\xi}\,d\xi\). The Fourier transform of an \(L^1\) function is continuous, bounded (\(|\hat{f}(\xi)| \leq \|f\|_1\)), and vanishes at infinity (Riemann–Lebesgue).

Alternative conventions: In physics and engineering, common conventions are \(\hat{f}(\xi) = \int f(x)e^{-i\xi x}\,dx\) (with the \(2\pi\) factor appearing in the inversion formula), or \(\hat{f}(\xi) = \frac{1}{\sqrt{2\pi}}\int f(x)e^{-i\xi x}\,dx\) (the symmetric convention). All are equivalent; the \(2\pi\)-in-exponent convention is the “analyst’s convention” that makes Plancherel an isometry without extra constants.

Key properties of the Fourier transform:

• Linearity: \(\widehat{af+bg} = a\hat{f} + b\hat{g}\).
• Translation: \(\widehat{f(\cdot - a)}(\xi) = e^{-2\pi ia\xi}\hat{f}(\xi)\).
• Modulation: \(\widehat{e^{2\pi ib\cdot}f}(\xi) = \hat{f}(\xi - b)\).
• Scaling: \(\widehat{f(a\cdot)}(\xi) = \frac{1}{|a|}\hat{f}(\xi/a)\) for \(a \neq 0\).
• Convolution: \(\widehat{f*g}(\xi) = \hat{f}(\xi)\hat{g}(\xi)\).
• Differentiation: \(\widehat{f'}(\xi) = 2\pi i\xi\hat{f}(\xi)\) (if \(f' \in L^1\)).
• Riemann–Lebesgue: \(\hat{f}(\xi) \to 0\) as \(|\xi| \to \infty\).

The Fundamental Example: The Gaussian

The most important example in all of Fourier analysis is the Gaussian:

This is the statement that the Gaussian function is an eigenfunction of the Fourier transform with eigenvalue 1. Let us prove this in full detail.

Proof: We compute \(\hat{f}(\xi) = \int_{-\infty}^\infty e^{-\pi x^2} e^{-2\pi ix\xi}\,dx\).

\[ I'(\xi) = \int_{-\infty}^\infty e^{-\pi x^2}(-2\pi ix)e^{-2\pi ix\xi}\,dx = -i\int_{-\infty}^\infty x e^{-\pi x^2}e^{-2\pi ix\xi}\,dx. \]\[ I'(\xi) = -i\left[-\frac{e^{-\pi x^2}e^{-2\pi ix\xi}}{2\pi}\Bigg|_{-\infty}^\infty + \frac{1}{2\pi}\int_{-\infty}^\infty e^{-\pi x^2}(-2\pi i\xi)e^{-2\pi ix\xi}\,dx\right] = -i \cdot \frac{-2\pi i\xi}{2\pi}\cdot I(\xi) = -2\pi\xi I(\xi). \]

Step 2: Solve the ODE. We have \(I'(\xi) = -2\pi\xi I(\xi)\), which is separable: \(\frac{dI}{I} = -2\pi\xi\,d\xi\), giving \(I(\xi) = Ce^{-\pi\xi^2}\).

Step 3: Determine the constant. \(I(0) = \int_{-\infty}^\infty e^{-\pi x^2}\,dx\). The standard Gaussian integral gives \(\int_{-\infty}^\infty e^{-\pi x^2}\,dx = 1\) (by the substitution \(u = \sqrt{\pi}x\) and the well-known \(\int e^{-u^2}\,du = \sqrt{\pi}\)). Thus \(C = I(0) = 1\), and \(\hat{f}(\xi) = e^{-\pi\xi^2}\). \(\square\)

Why does this matter? The Gaussian arises everywhere because:

1. It is the unique (up to scaling) function that is its own Fourier transform — a “fixed point” of \(\mathcal{F}\).
2. It saturates the Heisenberg uncertainty principle (achieves equality).
3. It is the fundamental solution of the heat equation: \(H_t(x) = \frac{1}{\sqrt{4\pi t}}e^{-x^2/(4t)}\) is a Gaussian.
4. It is the limit in the Central Limit Theorem: the distribution of \((X_1+\cdots+X_n)/\sqrt{n}\) converges to a Gaussian.

More generally, \(\hat{f}(\xi) = e^{-a\pi\xi^2}\) when \(f(x) = a^{-1/2}e^{-\pi x^2/a}\). The scaling \(a \leftrightarrow 1/a\) in the uncertainty principle is reflected in the Fourier transform: a wide Gaussian (\(a\) large) transforms to a narrow Gaussian (\(1/a\) small).

Eigenfunctions of the Fourier transform: More generally, the Hermite functions \(h_n(x) = H_n(\sqrt{2\pi}x)e^{-\pi x^2}\) (where \(H_n\) are Hermite polynomials) are eigenfunctions of \(\mathcal{F}\) with eigenvalues \(e^{-\pi i n/2} = i^{-n}\). They form a complete orthonormal basis of \(L^2(\mathbb{R})\). The Fourier transform acts on this basis by \(\mathcal{F}h_n = i^{-n}h_n\), showing \(\mathcal{F}^4 = \text{Id}\) (the Fourier transform has order 4 as an operator on \(L^2(\mathbb{R})\)).

Test Functions and the Schwartz Class

To develop a robust theory including inversion, one works with spaces of rapidly decaying smooth functions.

Examples of Schwartz functions include: Gaussians \(e^{-ax^2}\) (for \(a > 0\)), functions with compact support that are smooth (compactly supported smooth functions \(C_c^\infty(\mathbb{R}) \subset \mathcal{S}(\mathbb{R})\)), and products of polynomials with Gaussians.

Why Schwartz? The Schwartz class is specifically designed so that the Fourier transform preserves it. The Fourier transform swaps differentiation and multiplication by \(x\): \(\widehat{f'}(\xi) = 2\pi i\xi\hat{f}(\xi)\) and \(\widehat{xf}(\xi) = \frac{1}{-2\pi i}\hat{f}'(\xi)\). If \(f\) decays rapidly and is smooth, then \(\hat{f}\) is smooth (because \(x^m f \in L^1\) for all \(m\)) and decays rapidly (because \(f\) is smooth). So \(\mathcal{S}\) is closed under \(\mathcal{F}\). The key point is that both conditions — rapid decay AND infinite smoothness — are needed simultaneously, and Schwartz functions have both.

The Fourier Transform as an Isomorphism on \(\mathcal{S}\)

The key insight is that the Fourier transform exchanges differentiation and multiplication by polynomials, so the Schwartz conditions (rapid decay times smooth) are preserved.

---

Chapter 8: Distributions and Tempered Distributions

Distributions

The theory of distributions extends classical analysis to include objects like the Dirac delta that arise naturally in physics and PDE.

Every locally integrable function \(f\) defines a regular distribution \(T_f(\phi) = \int f(x)\phi(x)\,dx\). The Dirac delta \(\delta_a(\phi) = \phi(a)\) is a singular distribution not given by a function.

Why distributions? Classical analysis runs into walls when working with PDE: the derivative of a discontinuous function does not exist classically, yet it arises naturally in physics (the derivative of the Heaviside step function is the Dirac delta). Distributions (also called “generalized functions”) extend the notion of function so that every distribution has a derivative, every \(L^1_\text{loc}\) function can be differentiated infinitely often, and solutions to PDE can be found in spaces much larger than classical function spaces.

Laurent Schwartz introduced distributions rigorously in 1945, for which he was awarded the Fields Medal in 1950.

Operations on distributions are defined by duality:

• Differentiation: \(T'(\phi) = -T(\phi')\) (justified by integration by parts for regular distributions).
• Multiplication by smooth functions: \((gT)(\phi) = T(g\phi)\) for \(g \in C^\infty(\mathbb{R})\).

The principal value distribution \(\text{p.v.}\frac{1}{x}(\phi) = \lim_{\varepsilon \to 0^+}\int_{|x|>\varepsilon} \frac{\phi(x)}{x}\,dx\) is a fundamental singular distribution arising in the Hilbert transform.

Distributional derivative examples:

• The Heaviside step function \(H(x) = \mathbf{1}_{x>0}\) has distributional derivative \(H' = \delta_0\).
• The absolute value \(|x|\) has distributional derivative \(\text{sgn}(x)\), and second derivative \(2\delta_0\).
• The function \(\log|x|\) has distributional derivative \(\text{p.v.}\frac{1}{x}\).

Tempered Distributions

Since \(\mathcal{D}(\mathbb{R}) \subset \mathcal{S}(\mathbb{R})\), restriction gives \(\mathcal{S}'(\mathbb{R}) \subset \mathcal{D}'(\mathbb{R})\). Tempered distributions are precisely those distributions that can be extended continuously from \(\mathcal{D}\) to \(\mathcal{S}\).

Functions of polynomial growth define tempered distributions: if \(f\) is measurable and \(|f(x)| \leq C(1 + |x|)^N\) for some \(N\), then \(T_f \in \mathcal{S}'(\mathbb{R})\).

Why tempered? “Tempered” means “not growing too fast.” A distribution is tempered if it does not grow faster than a polynomial — fast enough for the Fourier transform to be defined (since Schwartz functions decay faster than any polynomial, tempered distributions can be “tested” against them). The Fourier transform is well-defined on \(\mathcal{S}'\), but not on all of \(\mathcal{D}'\) (which can include things like \(e^{x^2}\) that grow too fast for the Fourier transform to make sense).

Fourier Transform of Tempered Distributions

This is consistent with the classical definition for \(L^1\) functions (by Parseval’s identity) and extends the Fourier transform to a continuous automorphism of \(\mathcal{S}'(\mathbb{R})\).

Key examples:

Differentiation: \(\widehat{T'} = 2\pi i\xi \hat{T}\) and \(\widehat{xT} = \frac{i}{2\pi}\hat{T}'\), extending the classical rules to distributions.

---

Chapter 9: Inversion and Plancherel on ℝ

Fourier Inversion on the Schwartz Class

Inversion on \(L^1(\mathbb{R})\)

For \(f \in L^1(\mathbb{R})\), the Fourier transform \(\hat{f}\) is continuous and bounded but need not be in \(L^1\), so the inversion integral may not converge absolutely.

The difficulty: Unlike the circle \(\mathbb{T}\), on \(\mathbb{R}\) the Fourier transform of an \(L^1\) function need not be in \(L^1\). Consider the function \(f = \mathbf{1}_{[-1,1]}\): its Fourier transform is \(\hat{f}(\xi) = \frac{\sin(2\pi\xi)}{\pi\xi}\), which is in \(L^2(\mathbb{R})\) but not in \(L^1(\mathbb{R})\) (since \(\int |\sin(2\pi\xi)/(\pi\xi)|\,d\xi = \infty\)). The inversion formula \(f = \check{\hat{f}}\) holds but requires principal value interpretation.

Plancherel’s Theorem on \(L^2(\mathbb{R})\)

The \(L^2\) Fourier transform is defined as an \(L^2\) limit: for \(f \in L^2(\mathbb{R})\), choose \(f_n \in L^1 \cap L^2\) with \(f_n \to f\) in \(L^2\) (e.g., \(f_n = f \cdot \mathbf{1}_{[-n,n]}\)), and then \(\hat{f} = \lim_{n\to\infty} \hat{f}_n\) in \(L^2\). The pointwise formula \(\hat{f}(\xi) = \int f(x)e^{-2\pi ix\xi}\,dx\) holds only in a principal value sense.

Plancherel as a symmetry: The statement \(\|\hat{f}\|_2 = \|f\|_2\) says that the Fourier transform preserves the \(L^2\) norm — it is a unitary operator on the Hilbert space \(L^2(\mathbb{R})\). This is a kind of “energy conservation” principle: the total energy of a signal is the same whether measured in the time domain or the frequency domain.

---

Chapter 10: Applications — PDE, Probability, Number Theory

Heat Equation on \(\mathbb{R}\)

\[ \hat{u}_t(\xi, t) = -(2\pi\xi)^2 \hat{u}(\xi,t), \quad \hat{u}(\xi,0) = \hat{f}(\xi). \]\[ H_t(x) = \frac{1}{\sqrt{4\pi t}}e^{-x^2/(4t)}. \]

For \(t > 0\), \(H_t \in \mathcal{S}(\mathbb{R})\), and convolution with \(H_t\) smooths \(f\) instantly — even if \(f\) is only in \(L^2\) or \(L^1\), the solution is \(C^\infty\) for \(t > 0\) (hypoellipticity / infinite speed of propagation of smoothness).

\[ H_t(x) = \frac{1}{\sqrt{4\pi t}}e^{-x^2/(4t)} = \mathcal{F}^{-1}(e^{-4\pi^2\xi^2 t})(x). \]

This is a beautiful circle: the heat equation’s fundamental solution is a Gaussian, which is the function that is its own Fourier transform (up to scaling). The deep reason is that the Gaussian is the unique fixed point of the Fourier transform — and the heat equation’s smoothing action in the spatial domain corresponds to Gaussian damping in the frequency domain.

Wave Equation

The wave equation \(u_{tt} = u_{xx}\) transforms to \(\hat{u}_{tt} = -(2\pi\xi)^2\hat{u}\), giving \(\hat{u}(\xi,t) = \hat{f}(\xi)\cos(2\pi\xi t) + \hat{g}(\xi)\frac{\sin(2\pi\xi t)}{2\pi\xi}\) for initial data \(u(\cdot,0) = f\), \(u_t(\cdot,0) = g\). The d’Alembert formula \(u(x,t) = \frac{f(x+t)+f(x-t)}{2} + \frac{1}{2}\int_{x-t}^{x+t} g(s)\,ds\) follows by inverting. Unlike the heat equation, the wave equation propagates at finite speed (the solution at \((x,t)\) depends only on the initial data in \([x-t,x+t]\)).

Contrast with heat: The heat equation has infinite propagation speed (any perturbation of the initial data affects the solution everywhere immediately), while the wave equation has finite propagation speed (the “cone of dependence” is \([x-t, x+t]\)). This is visible in the Fourier transform: the heat damping factor \(e^{-4\pi^2\xi^2 t}\) kills all high frequencies exponentially, while the wave factor \(e^{\pm 2\pi i\xi t}\) is a pure phase rotation — no damping, just propagation.

Probability: Characteristic Functions

Characteristic functions are the Fourier–Stieltjes transforms of probability measures. They enjoy several key properties: \(\varphi_X(0) = 1\), \(|\varphi_X(t)| \leq 1\), and \(\varphi_X\) is uniformly continuous. The characteristic function of the standard normal \(N(0,1)\) is \(e^{-t^2/2}\) — again a Gaussian! This reflects the fact that the Gaussian is stable under convolution (sums of Gaussians are Gaussian) and is the unique distribution with this property (up to scaling).

Central Limit Theorem sketch: If \(X_1, X_2, \ldots\) are i.i.d. with mean 0 and variance 1, then \(\varphi_{S_n/\sqrt{n}}(t) = \left(\varphi_X(t/\sqrt{n})\right)^n\). Taylor expanding \(\varphi_X(t/\sqrt{n}) \approx 1 - t^2/(2n) + o(1/n)\) gives \(\varphi_{S_n/\sqrt{n}}(t) \to e^{-t^2/2}\), the characteristic function of \(N(0,1)\). By Lévy’s theorem, \(S_n/\sqrt{n} \xrightarrow{d} N(0,1)\).

Poisson Summation Formula

\[ \theta(t) := \sum_{n \in \mathbb{Z}} e^{-\pi n^2 t} = \frac{1}{\sqrt{t}}\sum_{n \in \mathbb{Z}} e^{-\pi n^2/t} = \frac{1}{\sqrt{t}}\theta(1/t). \]

This functional equation for \(\theta(t)\) is central to the proof of the functional equation of the Riemann zeta function \(\zeta(s)\).

Application — Number theory: Via the Poisson summation formula, one can relate sums over lattice points to sums over frequencies. For example, the number of ways to represent an integer \(n\) as a sum of four squares is related to the Fourier coefficients of \(\theta(t)^4\) — a deep connection between automorphic forms and arithmetic (Jacobi’s four-square theorem).

Application — Sampling: The Poisson summation formula is the mathematical foundation of the Shannon–Nyquist sampling theorem. Sampling a function \(f\) at integer points and then summing over \(\mathbb{Z}\) creates a periodic “aliased” version \(F(x) = \sum_n f(x+n)\) whose Fourier coefficients are the values \(\hat{f}(k)\) at integer frequencies. If \(\hat{f}\) is supported in \([-1/2, 1/2]\) (band-limited), there is no aliasing and the samples perfectly determine \(f\).

---

Chapter 11: Further Applications

The Sampling Theorem

The Shannon–Nyquist sampling theorem answers the question: when can a continuous function be perfectly reconstructed from its values at discrete sample points?

The sampling rate \(2B\) (the Nyquist rate) is optimal: sampling below this rate leads to aliasing, where different frequency components become indistinguishable. The sinc function \(\frac{\sin(\pi y)}{\pi y}\) is the inverse Fourier transform of the indicator function of \([-1/2, 1/2]\), so the reconstruction formula is exactly the statement that convolution with an ideal low-pass filter (the indicator of \([-B,B]\) in frequency space) reconstructs the band-limited function from its samples.

The Uncertainty Principle

The Heisenberg uncertainty principle in quantum mechanics has a precise mathematical formulation in terms of Fourier analysis.

Equality condition: Equality holds in Cauchy–Schwarz when \(f'\) and \(xf\) are proportional: \(f'(x) = \lambda x f(x)\) for some \(\lambda \in \mathbb{C}\). Since we also need \(\text{Re}(\lambda) < 0\) for \(f \in L^2\), writing \(\lambda = -2\pi a\) for \(a > 0\) gives \(f'(x) = -2\pi ax f(x)\), solved by \(f(x) = Ce^{-\pi a x^2}\). The equality is thus achieved uniquely (up to normalization, translation, and modulation) by Gaussian functions!

Physical interpretation: In quantum mechanics, \(|f(x)|^2\,dx\) is the probability of finding a particle in \([x, x+dx]\), and \(|\hat{f}(\xi)|^2\,d\xi\) is the probability of the particle having momentum in \([\xi, \xi+d\xi]\). The uncertainty principle says: you cannot simultaneously pin down both position and momentum with arbitrary precision. The Gaussian wavepacket (coherent state) is the unique state that saturates this bound — it is the most “quantum mechanically localized” state possible.

Different formulation: In signal processing, \(\Delta t \cdot \Delta\omega \geq \frac{1}{2}\) where \(\Delta t\) is the “time spread” and \(\Delta\omega\) is the “frequency spread” (standard deviations). This is why speech compression algorithms and spectrogram analysis face fundamental resolution limits: you cannot simultaneously achieve fine time resolution and fine frequency resolution.

Sobolev Spaces

For integer \(s = k \geq 0\), \(H^k(\mathbb{R})\) coincides with the classical Sobolev space of functions with \(k\) derivatives in \(L^2\): \(H^k = \{f : f^{(j)} \in L^2 \text{ for } 0 \leq j \leq k\}\). The Fourier definition extends this to fractional and negative \(s\).

Why Sobolev spaces? In PDE, solutions often exist in “weak” form — they satisfy an equation only in a distributional sense, with limited regularity. Sobolev spaces provide the correct framework: they are Hilbert spaces with norms measuring both function values and derivative sizes, and they have good compactness and embedding properties.

Sobolev embedding theorem: if \(s > 1/2\), then \(H^s(\mathbb{R}) \hookrightarrow C_0(\mathbb{R})\) (functions in \(H^s\) are continuous and vanish at infinity). More generally, if \(s > k + 1/2\), then \(H^s(\mathbb{R}) \hookrightarrow C^k(\mathbb{R})\).

Proof of embedding: If \(f \in H^s\) with \(s > 1/2\), then \(\hat{f} \in L^1(\mathbb{R})\) (by Cauchy–Schwarz: \(\int |\hat{f}(\xi)|\,d\xi \leq (\int(1+|\xi|^2)^s|\hat{f}|^2)^{1/2}(\int(1+|\xi|^2)^{-s})^{1/2} < \infty\) when \(2s > 1\)), so \(f = \check{\hat{f}}\) is a continuous function vanishing at infinity.

\[ \hat{u}(\xi)(1+|\xi|^2)^{(s+2)/2} = \hat{f}(\xi)(1+|\xi|^2)^{s/2} \cdot \frac{(1+|\xi|^2)^{1}}{4\pi^2|\xi|^2+1} \in L^2, \]

so \(u \in H^{s+2}\). The elliptic operator gains two derivatives in Sobolev regularity — a manifestation of elliptic regularity.

---

Chapter 12: Fourier–Stieltjes Transform and Bochner’s Theorem

Measures and the Fourier–Stieltjes Transform

Let \(M(\mathbb{R})\) denote the space of finite (complex) Borel measures on \(\mathbb{R}\), equipped with total variation norm \(\|\mu\| = |\mu|(\mathbb{R})\). This is a Banach algebra under convolution \((\mu * \nu)(E) = \int \mu(E-x)\,d\nu(x)\).

Properties of \(\hat{\mu}\):

• \(\hat{\mu}\) is bounded: \(|\hat{\mu}(\xi)| \leq \|\mu\|\).
• \(\hat{\mu}\) is uniformly continuous (but need not vanish at infinity — e.g., \(\mu = \delta_a\) gives \(\hat{\mu}(\xi) = e^{-2\pi ia\xi}\)).
• \(\widehat{\mu * \nu} = \hat{\mu} \cdot \hat{\nu}\) (the Fourier–Stieltjes transform is a homomorphism).

For absolutely continuous measures \(d\mu = f\,dx\) with \(f \in L^1(\mathbb{R})\), this reduces to the ordinary Fourier transform. The Riemann–Lebesgue lemma no longer applies: \(\hat{\mu}\) need not tend to zero at infinity.

Positive-Definite Functions

\[ \sum_{j,k} c_j\overline{c_k}\hat{\mu}(x_j - x_k) = \int_{\mathbb{R}} \left|\sum_j c_j e^{-2\pi ix_j\xi}\right|^2 d\mu(\xi) \geq 0. \]

Bochner’s theorem is the profound converse.

Bochner’s Theorem

The Wiener algebra: The image of \(M(\mathbb{R})\) under the Fourier–Stieltjes transform is the algebra of Fourier–Stieltjes transforms, a subalgebra of \(C_b(\mathbb{R})\) (bounded continuous functions). Understanding this image — which continuous functions are Fourier–Stieltjes transforms — is one of the central questions in abstract harmonic analysis, motivating the study of the Fourier algebra and Fourier–Stieltjes algebra \(B(G)\) for locally compact groups \(G\), a topic of ongoing research including work by Prof. Spronk.

---

Chapter 13: Abstract Harmonic Analysis and Pontryagin Duality

Locally Compact Abelian Groups

The framework of this course — Fourier analysis on \(\mathbb{T}\) and \(\mathbb{R}\) — extends to a vast generalization: the theory of locally compact abelian groups (LCA groups). This generalization, developed primarily by André Weil and Lev Pontryagin in the 1930s–40s, unifies Fourier series (on \(\mathbb{T}\)), the Fourier transform (on \(\mathbb{R}\)), and discrete Fourier analysis (on \(\mathbb{Z}/n\mathbb{Z}\)) into a single framework.

Examples:

• \(\mathbb{R}^n\) with addition (non-compact).
• \(\mathbb{T} = \mathbb{R}/\mathbb{Z}\) (compact).
• \(\mathbb{Z}\) with the discrete topology (discrete).
• \(\mathbb{Z}/n\mathbb{Z}\) (finite, hence compact and discrete).
• \(\mathbb{Q}_p\) (the \(p\)-adic numbers — compact open subgroups).
• Any product \(G_1 \times G_2\) of LCA groups.

Haar Measure

The Haar measure is the natural notion of “size” on \(G\), analogous to Lebesgue measure on \(\mathbb{R}\). It allows the definition of \(L^p(G)\) spaces and convolution on \(G\).

Examples of Haar measures:

• \(\mathbb{R}^n\): Lebesgue measure \(dx\).
• \(\mathbb{T}\): normalized arc length \(\frac{d\theta}{2\pi}\).
• \(\mathbb{Z}\): counting measure.
• \(\mathbb{Z}/n\mathbb{Z}\): \(\frac{1}{n}\) times counting measure (normalized).
• Compact group: the Haar measure is a probability measure (total mass 1).
• Discrete group: counting measure.

The Dual Group and Pontryagin Duality

The characters \(\chi : G \to \mathbb{T}\) are the “pure frequencies” on (G$. They generalize:

• On \(\mathbb{T}\): characters are \(\chi_n(\theta) = e^{in\theta}\) for \(n \in \mathbb{Z}\), so \(\hat{\mathbb{T}} = \mathbb{Z}\).
• On \(\mathbb{R}\): characters are \(\chi_\xi(x) = e^{2\pi ix\xi}\) for \(\xi \in \mathbb{R}\), so \(\hat{\mathbb{R}} = \mathbb{R}\).
• On \(\mathbb{Z}\): characters are \(\chi_\theta(n) = e^{in\theta}\) for \(\theta \in \mathbb{T}\), so \(\hat{\mathbb{Z}} = \mathbb{T}\).
• On \(\mathbb{Z}/n\mathbb{Z}\): characters are \(\chi_k(j) = e^{2\pi ijk/n}\) for \(k \in \mathbb{Z}/n\mathbb{Z}\), so \(\widehat{\mathbb{Z}/n\mathbb{Z}} = \mathbb{Z}/n\mathbb{Z}\).

Pontryagin duality says that “taking characters twice” returns to the original group. This is a profound symmetry between a group and its dual:

• \(\mathbb{T}\) and \(\mathbb{Z}\) are Pontryagin duals of each other.
• \(\mathbb{R}\) is self-dual: \(\hat{\mathbb{R}} \cong \mathbb{R}\).
• \(\mathbb{Z}/n\mathbb{Z}\) is self-dual.
• Compact groups have discrete duals; discrete groups have compact duals.

The duality table:

Group \(G\)	Dual \(\hat{G}\)	Fourier analysis
\(\mathbb{T}\) (circle)	\(\mathbb{Z}\) (integers)	Fourier series
\(\mathbb{Z}\) (integers)	\(\mathbb{T}\) (circle)	Discrete Fourier series
\(\mathbb{R}\) (reals)	\(\mathbb{R}\) (reals)	Fourier transform
\(\mathbb{Z}/n\mathbb{Z}\) (cyclic)	\(\mathbb{Z}/n\mathbb{Z}\) (cyclic)	DFT
Compact \(\Leftrightarrow\)	Discrete	Fourier series \(\leftrightarrow\) Fourier transform

The Abstract Fourier Transform

\[ \hat{f}(\chi) = \int_G f(g)\overline{\chi(g)}\,d\mu_G(g), \quad \chi \in \hat{G}. \]

This is the Fourier transform on the dual group \(\hat{G}\). The abstract theory gives:

• Riemann–Lebesgue: \(\hat{f} \in C_0(\hat{G})\) for \(f \in L^1(G)\).
• Convolution: \(\widehat{f*g} = \hat{f}\cdot\hat{g}\).
• Plancherel: The Fourier transform extends to a unitary isomorphism \(L^2(G) \cong L^2(\hat{G})\) (with appropriate Haar measures on \(\hat{G}\)).
• Inversion: For \(f \in L^1(G)\) with \(\hat{f} \in L^1(\hat{G})\), \(f(g) = \int_{\hat{G}}\hat{f}(\chi)\chi(g)\,d\mu_{\hat{G}}(\chi)\).

Abstract Fourier Analysis on the Circle: Connection to Representation Theory

The characters \(e^{inx}\) on \(\mathbb{T}\) are precisely the irreducible unitary representations of \(\mathbb{T}\). Since \(\mathbb{T}\) is abelian, every irreducible representation is one-dimensional (a character). The Fourier series expansion \(f = \sum_n \hat{f}(n)e_n\) is the decomposition of \(f\) (viewed as acting by multiplication in the regular representation) into irreducibles.

For non-abelian groups \(G\), irreducible representations are higher-dimensional, and the Fourier transform \(\hat{f}(\pi) = \int_G f(g)\pi(g)\,d\mu_G(g)\) takes values in operators (matrices). The Peter–Weyl theorem for compact groups and the Plancherel theorem for locally compact groups give the non-abelian analogues of Parseval’s identity. This is the starting point of the representation-theoretic approach to Fourier analysis, which is central to Prof. Spronk’s research.

The Wiener Tauberian Theorem

A crowning achievement of abstract harmonic analysis is Wiener’s Tauberian theorem, which generalizes Hardy’s theorem and has deep applications in number theory.

The condition \(\hat{f} \neq 0\) everywhere is called “non-vanishing spectrum.” Wiener’s theorem says: if \(f\) has non-vanishing spectrum, then the translates of \(f\) are “dense” in \(L^1\). As a corollary, if \(\mu\) is a measure on \(\mathbb{R}\) and \(\int f(x-t)\,d\mu(t) \to L\int f(x)\,dx\) for some \(f\) with non-vanishing spectrum, then the same holds for all \(g \in L^1\). This is the Tauberian content: knowing the “limit” for one test function with non-vanishing spectrum forces the limit for all test functions.

Application to the prime number theorem: The prime number theorem (\(\pi(x) \sim x/\log x\) where \(\pi(x)\) counts primes up to \(x\)) can be proved using Wiener’s theorem. The key step is showing that the Riemann zeta function \(\zeta(s)\) has no zeros on the line \(\text{Re}(s) = 1\), which corresponds (via an analytic Tauberian argument) to the non-vanishing of a certain Fourier transform.

---

Chapter 14: The Hilbert Transform

Definition and Motivation

The Hilbert transform is the most important singular integral operator in harmonic analysis. It arises naturally in several contexts: the relationship between real and imaginary parts of analytic functions, the conjugate Fourier series, and the theory of \(L^p\) convergence of Fourier series.

The Hilbert transform is a convolution with the kernel \(k(x) = \frac{1}{\pi x}\), which is not integrable at 0 — hence the principal value interpretation.

\[ \widehat{Hf}(\xi) = -i\,\text{sgn}(\xi)\hat{f}(\xi), \]

where \(\text{sgn}(\xi) = \pm 1\) for \(\xi \gtrless 0\). The Hilbert transform is thus the Fourier multiplier with symbol \(-i\,\text{sgn}(\xi)\) — it rotates the phase of each frequency component by \(\pm\pi/2\).

Derivation: The Fourier transform of the principal value distribution \(\text{p.v.}\frac{1}{\pi x}\) is \(-i\,\text{sgn}(\xi)\). This can be seen by regularization: \(\text{p.v.}\frac{1}{\pi x} = \lim_{\varepsilon\to 0}(\frac{1}{\pi x}\mathbf{1}_{|x|>\varepsilon})\), and the Fourier transform of \(\frac{1}{\pi x}\mathbf{1}_{|x|>\varepsilon}\) converges to \(-i\,\text{sgn}(\xi)\) as \(\varepsilon\to 0\).

Properties of the Hilbert Transform

1. \(L^2\) isometry: Since \(|\text{sgn}(\xi)| = 1\), Plancherel gives \(\|Hf\|_2 = \|f\|_2\) — the Hilbert transform is an isometry on \(L^2(\mathbb{R})\).

2. Involution: \(H^2 = -I\) (the Hilbert transform squares to minus the identity), since \((-i\,\text{sgn}(\xi))^2 = -1\). So \(H\) is a unitary operator with \(H^{-1} = -H\).

3. \(L^p\) boundedness (M. Riesz, 1927): For \(1 < p < \infty\), \(H : L^p(\mathbb{R}) \to L^p(\mathbb{R})\) is bounded. This is the central tool in proving \(L^p\) convergence of Fourier series.

4. Weak \(L^1\) bound: \(H\) is not bounded on \(L^1\) (the integral of \(1/x\) diverges), but it satisfies the weak-type estimate \(\mu(\{|Hf| > \lambda\}) \leq C\|f\|_1/\lambda\) for all \(\lambda > 0\).

5. Analytic functions: If \(f\) is the real part of an analytic function \(F = f + ig\) in the upper half-plane, then \(g\) is (up to constants) the Hilbert transform of \(f|_{\mathbb{R}}\). The Hilbert transform thus converts between the real and imaginary parts of boundary values of analytic functions.

Connection to Fourier Series Convergence

\[ \tilde{f}(x) = -i\sum_{n\neq 0} \text{sgn}(n)\hat{f}(n)e^{inx}. \]

This is the Fourier series with the Fourier multiplier \(-i\,\text{sgn}(n)\) — the discrete analogue of the Hilbert transform. The conjugate Fourier series converges in \(L^p\) for \(1 < p < \infty\) because the discrete Hilbert transform (convolution with \(\text{p.v.}\cot(x/2)/(2\pi)\) on \(\mathbb{T}\)) is bounded on \(L^p(\mathbb{T})\). M. Riesz’s theorem on \(L^p\) convergence of Fourier series follows directly.

---

Chapter 15: Harmonic Analysis Beyond the Course

Wavelets: An Alternative to Fourier

The Fourier transform decomposes a function into pure frequencies — infinite sine waves. This is powerful for stationary signals (whose frequency content does not change over time), but poor for non-stationary signals (like speech, music, or images with sharp edges).

The wavelet transform decomposes a function into “wavelets” — localized oscillatory pieces supported in small intervals, dilated and translated to cover all scales and locations. The wavelet transform \(Wf(a,b) = \frac{1}{\sqrt{|a|}}\int f(x)\overline{\psi\!\left(\frac{x-b}{a}\right)}\,dx\) (where \(\psi\) is a “mother wavelet”) provides simultaneous time-frequency localization, unlike the Fourier transform.

Wavelets inherit the uncertainty principle: a wavelet cannot be perfectly localized in both time and frequency, but wavelets are specifically designed to balance this tradeoff better than pure Fourier modes for practical signals.

\(L^p\) Harmonic Analysis and Calderón–Zygmund Theory

Beyond the Hilbert transform, there is a large family of singular integral operators of the form \(Tf(x) = \int K(x,y)f(y)\,dy\) where \(K\) is a singular kernel. The Calderón–Zygmund theory (1952) provides a unified framework for proving \(L^p\) boundedness of such operators:

1. Show \(T\) is bounded on \(L^2\) (often via Fourier multipliers).
2. Show \(T\) satisfies a “weak type (1,1)” estimate (using the Calderón–Zygmund decomposition).
3. Interpolate to get boundedness on \(L^p\) for \(1 < p < 2\).
4. Use duality to extend to \(2 < p < \infty\).

This program, combining interpolation, decomposition, and duality, is the backbone of modern harmonic analysis and PDE.

Hardy Spaces and BMO

The Hardy space \(H^1(\mathbb{R})\) is the “right” replacement for \(L^1\) in harmonic analysis: it consists of those \(L^1\) functions whose Hilbert transform is also in \(L^1\). The space \(\text{BMO}(\mathbb{R})\) (functions of bounded mean oscillation) is the dual of \(H^1\) and the “right” replacement for \(L^\infty\). The Fefferman–Stein theorem (\(H^1\)–\(\text{BMO}\) duality) is a deep result that shows \(\text{BMO}\) is the correct endpoint for many singular integral estimates.

The Fourier multiplier characterization of singular integrals, the role of the Hilbert transform as the canonical singular integral, and the Calderón–Zygmund decomposition together form the foundation of modern harmonic analysis and its applications to PDE, number theory, and geometry.

---

Appendix: Key Identities and Summary Formulas

Fourier Series Formulas

For \(f \in L^1(\mathbb{T})\) with Fourier coefficients \(\hat{f}(n) = \frac{1}{2\pi}\int_{-\pi}^\pi f(x)e^{-inx}\,dx\):

Identity	Formula
Parseval	(|f|_{L^2}^2 = \sum_n
Convolution	\(\widehat{f*g}(n) = \hat{f}(n)\hat{g}(n)\)
Differentiation	\(\widehat{f'}(n) = in\hat{f}(n)\)
Translation	\(\widehat{f(\cdot-a)}(n) = e^{-ina}\hat{f}(n)\)
Modulation	\(\widehat{e^{iax}f}(n) = \hat{f}(n-a)\)
Conjugation	\(\widehat{\bar{f}}(n) = \overline{\hat{f}(-n)}\)

Key Fourier Coefficient Computations

Function \(f(x)\)	Fourier coefficients \(\hat{f}(n)\)
\(f(x) = x\) (sawtooth)	\(\hat{f}(0) = 0\), \(\hat{f}(n) = \frac{i(-1)^n}{n}\) for \(n\neq 0\)
\(f(x) = x^2\) (parabola)	\(\hat{f}(0) = \frac{\pi^2}{3}\), \(\hat{f}(n) = \frac{2(-1)^n}{n^2}\)
(f(x) =	x
\(f(x) = \text{sgn}(x)\) (square)	\(\hat{f}(n) = \frac{1-(-1)^n}{\pi in}\) (odd \(n\): \(\frac{2}{\pi in}\))
\(f(x) = e^{ax}\), \(a \notin i\mathbb{Z}\)	\(\hat{f}(n) = \frac{e^{a\pi}-e^{-a\pi}}{2\pi(a-in)}\)

Series Identities Derived from Parseval

Identity	Source
\(\sum_{n=1}^\infty \frac{1}{n^2} = \frac{\pi^2}{6}\)	Parseval for \(f(x)=x\) or evaluation of \(f(x)=x^2\)
\(\sum_{n=1}^\infty \frac{1}{n^4} = \frac{\pi^4}{90}\)	Parseval for \(f(x)=x^2\)
\(\sum_{n=0}^\infty \frac{1}{(2n+1)^2} = \frac{\pi^2}{8}\)	Parseval for square wave
\(\sum_{n=0}^\infty \frac{(-1)^n}{2n+1} = \frac{\pi}{4}\)	Leibniz formula, from \(f(x)=x\) at \(x=\pi/2\)
\(\sum_{n=1}^\infty \frac{1}{(2n-1)^4} = \frac{\pi^4}{96}\)	Parseval for (f(x)=

Fourier Transform Formulas on \(\mathbb{R}\)

Convention: \(\hat{f}(\xi) = \int_\mathbb{R} f(x)e^{-2\pi ix\xi}\,dx\).

Function \(f(x)\)	Fourier transform \(\hat{f}(\xi)\)
\(e^{-\pi x^2}\) (Gaussian)	\(e^{-\pi\xi^2}\) (self-dual!)
(e^{-a	x
\(\mathbf{1}_{[-a,a]}(x)\) (box)	\(\frac{\sin(2\pi a\xi)}{\pi\xi} = 2a\,\text{sinc}(2a\xi)\)
\(\text{sinc}(x) = \frac{\sin(\pi x)}{\pi x}\)	\(\mathbf{1}_{[-1/2,1/2]}(\xi)\)
\(H_t(x) = \frac{1}{\sqrt{4\pi t}}e^{-x^2/4t}\)	\(e^{-4\pi^2\xi^2 t}\)
\(\delta_a(x)\)	\(e^{-2\pi ia\xi}\)
\(e^{2\pi iax}\)	\(\delta_a(\xi)\)
\(\text{p.v.}\frac{1}{\pi x}\)	\(-i\,\text{sgn}(\xi)\)

Kernels and Their Properties

Kernel	Formula	Non-negative?	\(\|\cdot\|_{L^1}\)
Dirichlet \(D_N\)	\(\frac{\sin((N+\frac12)x)}{\sin(x/2)}\)	No	\(\sim \frac{4}{\pi^2}\log N\)
Fejér \(K_N\)	\(\frac{1}{N}\left(\frac{\sin(Nx/2)}{\sin(x/2)}\right)^2\)	Yes	1
Poisson \(P_r\)	\(\frac{1-r^2}{1-2r\cos x+r^2}\)	Yes	1
Weierstrass \(W_t\)	\(\sum_n e^{-n^2t}e^{inx}\)	Yes	1
Heat \(H_t\)	\(\frac{1}{\sqrt{4\pi t}}e^{-x^2/(4t)}\)	Yes	1

Fundamental Theorems Summary

---

Chapter 16: Why Exponentials? The Deep Reasons

Eigenfunctions of Translation and Differentiation

One of the most important conceptual questions in Fourier analysis is: why do we use \(e^{inx}\) as the basis functions? It is not an arbitrary choice. There are at least three independent, compelling reasons.

Reason 1: Eigenfunctions of the differentiation operator. Consider the ordinary differential equation \(Df = \lambda f\) where \(D = \frac{d}{dx}\). The solutions are \(f(x) = Ce^{\lambda x}\). For periodic solutions on \(\mathbb{T} = \mathbb{R}/2\pi\mathbb{Z}\), we need \(e^{\lambda(x+2\pi)} = e^{\lambda x}\), which forces \(e^{2\pi\lambda} = 1\), so \(\lambda = in\) for some \(n \in \mathbb{Z}\). Therefore, the functions \(e^{inx}\) are precisely the eigenfunctions of \(d/dx\) on \(\mathbb{T}\), with eigenvalues \(in\). Since \(d/dx\) is a fundamental operator in analysis and physics (it appears in every linear PDE), it is natural to decompose functions into eigenfunctions of \(d/dx\). The Fourier series is this eigenfunction expansion.

More concretely: any constant-coefficient linear differential operator \(L = \sum_{k=0}^N a_k \frac{d^k}{dx^k}\) is diagonalized by the Fourier basis. We have \(Le^{inx} = \left(\sum_k a_k (in)^k\right)e^{inx} = p(in)e^{inx}\) where \(p\) is a polynomial. So \(L\) acts on each Fourier mode independently, multiplying it by the scalar \(p(in)\). This is the fundamental reason the Fourier transform converts differential equations to algebraic ones: it simultaneously diagonalizes all constant-coefficient differential operators.

Reason 2: Characters of the group \(\mathbb{T}\). The circle \(\mathbb{T} = \mathbb{R}/2\pi\mathbb{Z}\) is an abelian group. A character of \(\mathbb{T}\) is a continuous group homomorphism \(\chi : \mathbb{T} \to \mathbb{C}^\times\) of modulus 1 (i.e., mapping into the unit circle \(\mathbb{T} \subset \mathbb{C}^\times\)). The homomorphism property means \(\chi(x+y) = \chi(x)\chi(y)\) for all \(x, y \in \mathbb{T}\).

If \(\chi\) is smooth and satisfies this, differentiate: \(\chi'(x+y) = \chi'(x)\chi(y)\). Setting \(x = 0\): \(\chi'(y) = \chi'(0)\chi(y)\). This is the same ODE as before, giving \(\chi(y) = e^{\chi'(0)y}\). The condition \(|\chi| = 1\) forces \(\chi'(0) = in\) for \(n \in \mathbb{Z}\). So the characters of \(\mathbb{T}\) are exactly \(e_n(x) = e^{inx}\), \(n \in \mathbb{Z}\).

This is the group-theoretic explanation: characters are the simplest functions on a group, encoding its symmetry. Any group-invariant analysis on \(\mathbb{T}\) must decompose into characters — this is the Peter–Weyl theorem for compact abelian groups.

Reason 3: Eigenfunctions of convolution (translation-invariant operators). A translation-invariant linear operator \(T : L^2(\mathbb{T}) \to L^2(\mathbb{T})\) satisfies \(T(\tau_a f) = \tau_a(Tf)\) where \(\tau_a f(x) = f(x-a)\). Such operators include: convolution with a fixed kernel \(Tf = f * g\), the Hilbert transform, the Fourier partial sums \(S_N\), and all differential operators with constant coefficients.

Claim: Every bounded translation-invariant operator on \(L^2(\mathbb{T})\) is a Fourier multiplier: \(\widehat{Tf}(n) = m(n)\hat{f}(n)\) for some bounded sequence \(m(n)\), called the multiplier sequence. The action of \(T\) on each Fourier mode \(e^{inx}\) is multiplication by \(m(n)\), and the functions \(e^{inx}\) are simultaneous eigenfunctions of all such operators.

\[ (T e_n)(x) = T(\tau_{-y}e_n)(x+y) = \tau_{-y}(Te_n)(x+y) \cdot e^{iny}/ e^{iny} \]

More cleanly: let \(m(n) = \widehat{Te_n}(n)/1\). Translation-invariance gives \(\tau_a(Te_n) = T(\tau_a e_n) = T(e^{-ina}e_n) = e^{-ina}Te_n\). So \(Te_n\) is an eigenfunction of every translation with eigenvalue \(e^{-ina}\) — which forces \(Te_n = c_n e_n\) for some constant \(c_n = m(n)\). For general \(f = \sum \hat{f}(n)e_n\), linearity and continuity give \(Tf = \sum m(n)\hat{f}(n)e_n\). Thus \(T\) is a Fourier multiplier. \(\square\)

This is the spectral theorem for translation-invariant operators: the Fourier basis simultaneously diagonalizes all such operators. In engineering terms, sinusoids are the “eigenmodes” of any linear time-invariant (LTI) system, and the transfer function \(m(n)\) captures the system’s response at each frequency.

\[ \widehat{f * g}(n) = \hat{f}(n)\hat{g}(n). \]

In algebraic language: the Fourier transform is an algebra homomorphism, and the characters \(e^{inx}\) are the “maximal ideals” (Gelfand theory) of the convolution algebra. The Fourier series is the Gelfand transform — the most natural representation of the algebra.

Why Not Other Bases?

Could we use other orthonormal bases for \(L^2(\mathbb{T})\)? Yes — for example, wavelets, or the Haar system, or Legendre polynomials adapted to the circle. But none of these have the four properties above simultaneously. Exponentials are special because they are:

1. Smooth (unlike the Haar wavelet which has jumps).
2. Eigenfunctions of differentiation (unlike wavelets).
3. Characters of the group (a group-theoretic property no other system has).
4. Eigenfunctions of all convolution operators simultaneously.

The price we pay is that exponentials are non-localized: a single \(e^{inx}\) is spread over all of \(\mathbb{T}\), contributing to every point. This is why the Fourier transform is poor for analyzing localized events (like a click in audio, or an edge in an image) — and why wavelets were invented as a complementary tool.

---

Chapter 17: Bessel’s Inequality, Completeness, and the Proof of Parseval

Bessel’s Inequality

Let \(\mathcal{H}\) be a Hilbert space with inner product \(\langle \cdot, \cdot \rangle\), and let \(\{e_n\}_{n \in \mathbb{Z}}\) be an orthonormal system (not necessarily complete). For any \(f \in \mathcal{H}\), the Fourier coefficients with respect to this system are \(c_n = \langle f, e_n \rangle\).

Geometric interpretation: Bessel’s inequality says that the energy in the Fourier components cannot exceed the total energy of \(f\). The “missing energy” \(\|f\|^2 - \sum_n |c_n|^2 = \|f - S_\infty\|^2 \geq 0\) is the energy in the “orthogonal complement” of the system — it is zero if and only if the system is complete (spans a dense subspace).

Completeness: Parseval’s Identity

The orthonormal system \(\{e^{inx}\}_{n \in \mathbb{Z}}\) is complete in \(L^2(\mathbb{T})\) if and only if no non-zero function is orthogonal to all \(e^{inx}\) — equivalently, if \(\hat{f}(n) = 0\) for all \(n \in \mathbb{Z}\), then \(f = 0\) in \(L^2(\mathbb{T})\).

Polarization: The inner product version \(\langle f, g\rangle = \sum_n \hat{f}(n)\overline{\hat{g}(n)}\) follows by polarizing: \(4\langle f,g\rangle = \|f+g\|^2 - \|f-g\|^2 + i\|f+ig\|^2 - i\|f-ig\|^2\).

---

Chapter 18: The Banach–Steinhaus Argument for Divergence

The Operator Norm of Partial Sums

We have seen that the Dirichlet kernel has \(L^1\) norm \(\|D_N\|_1 \sim \frac{4}{\pi^2}\log N \to \infty\). Let us make precise why this implies the existence of continuous functions with divergent Fourier series.

\[ \|L_{N,x_0}\| = \sup_{\|f\|_\infty \leq 1}|S_N f(x_0)| = \frac{1}{2\pi}\int_{-\pi}^\pi |D_N(x_0 - t)|\,dt = \|D_N\|_{L^1}. \]

(The equality of the operator norm to \(\|D_N\|_{L^1}\) follows because the supremum is attained by \(f(t) = \text{sgn}(D_N(x_0 - t))\), which one approximates by continuous functions.)

Since \(\|L_{N,x_0}\| = \|D_N\|_1 \to \infty\, \text{ as } N \to \infty\), the Banach–Steinhaus theorem (uniform boundedness principle) applies:

The contrapositive: if \(\sup_\alpha \|T_\alpha\| = \infty\), then there exists \(x \in X\) with \(\sup_\alpha \|T_\alpha x\| = \infty\). Applying this to \(T_N = L_{N,0}: C(\mathbb{T}) \to \mathbb{C}\) with \(\|T_N\| = \|D_N\|_1 \to \infty\), we conclude:

Refinement: By using all points \(x_0 \in \mathbb{T}\) simultaneously, one obtains: there exists \(f \in C(\mathbb{T})\) whose Fourier series diverges on a set of full measure (though du Bois-Reymond only showed divergence at a single point). Carleson’s 1966 theorem shows that for \(f \in L^2(\mathbb{T})\), the Fourier series converges almost everywhere — a far deeper result.

\[ \|S_N f\|_\infty \leq L_N \|f\|_\infty, \quad \text{but this bound is tight.} \]

The logarithmic growth \(L_N \sim \frac{4}{\pi^2}\log N\) is the reason partial sums of Fourier series can diverge — but only logarithmically slowly. In practice, for smooth functions, the divergence is overwhelmed by the rapid decay of Fourier coefficients.

---

Chapter 19: The Gibbs Phenomenon — Detailed Analysis

Precise Statement and Computation

\[ f(x) = \frac{\pi}{2}\text{sgn}(x) = \begin{cases} \pi/2 & 0 < x < \pi \\ -\pi/2 & -\pi < x < 0 \end{cases}, \quad \text{extended } 2\pi\text{-periodically}. \]\[ f(x) = \sum_{\substack{n=1 \\ n \text{ odd}}}^\infty \frac{2}{n}\sin(nx) = 2\sin x + \frac{2\sin 3x}{3} + \frac{2\sin 5x}{5} + \cdots \]

The partial sums are \(S_N f(x) = \sum_{\substack{n=1 \\ n \text{ odd}, n \leq 2N-1}} \frac{2}{n}\sin(nx) = 2\sum_{k=0}^{N-1}\frac{\sin((2k+1)x)}{2k+1}\).

\[ \cos\theta + \cos 3\theta + \cdots + \cos(2N-1)\theta = \frac{\sin 2N\theta}{2\sin\theta}. \]

(Proof: multiply by \(2\sin\theta\) and telescope.) So \((S_N f)'(x) = \frac{\sin 2Nx}{\sin x}\). The first critical point in \((0, \pi)\) is at \(x_N = \frac{\pi}{2N}\) (the first zero of \(\sin 2Nx\) after 0).

\[ S_N f(x_N) = 2\sum_{k=0}^{N-1}\frac{\sin((2k+1)\pi/(2N))}{2k+1}. \]\[ \frac{1}{\pi}\cdot 2\cdot\int_0^\pi \frac{\sin t}{t}\,dt \cdot \pi = 2\int_0^\pi \frac{\sin t}{t}\,dt = 2\,\text{Si}(\pi), \]\[ \lim_{N\to\infty} S_N f(x_N) = 2\,\text{Si}(\pi) \approx 2 \times 1.8519 \approx 3.7038\cdots \]\[ \text{Overshoot} = \lim_{N\to\infty}S_N f(x_N) - f(0^+) = 2\,\text{Si}(\pi) - \frac{\pi}{2} \approx 3.7038 - 1.5708 \approx 2.133. \]\[ \lim_{N\to\infty} S_N f(x_N) = \frac{4}{\pi}\int_0^\pi \frac{\sin t}{t}\,dt = \frac{4}{\pi}\text{Si}(\pi) \approx \frac{4}{\pi}\times 1.8519 \approx 2.3591. \]\[ G = \frac{2}{\pi}\text{Si}(\pi) - 1 = \frac{2}{\pi}\int_0^\pi\frac{\sin t}{t}\,dt - 1 \approx \frac{2\times 1.8519}{\pi} - 1 \approx 1.1790 - 1 = 0.1790. \]

So the overshoot is approximately \(17.9\%\) of the half-jump (or equivalently, \(8.9\%\) of the full jump), universally for any function with a jump discontinuity. This constant does not depend on the shape of the function near the discontinuity — only on the magnitude of the jump.

Why the overshoot persists: The key insight is that the overshoot does not come from “misadventure” or numerical error. It is an exact consequence of the Riemann sum approximation to \(\int_0^\pi \frac{\sin t}{t}\,dt\), which is a fixed positive number regardless of \(N\). The region of overshoot shrinks (width \(\approx \pi/N \to 0\)) but the height remains constant. The Gibbs phenomenon is therefore invisible in \(L^p\) for any \(p < \infty\) (because the overshoot region has vanishing measure), but it is visible in \(L^\infty\) (the max norm).

Gibbs for general jump discontinuities: If \(f\) has a jump of size \([f](x_0) = f(x_0^+) - f(x_0^-)\) at \(x_0\), then by the localization principle, \(S_N f\) near \(x_0\) looks like \(S_N g\) where \(g\) is the square wave with the same jump size. Therefore \(S_N f\) overshoots the jump by approximately \(G \cdot [f](x_0) \approx 8.9\%\) of the jump, universally.

Remedies for Gibbs: The three main methods to avoid Gibbs are:

1. Cesàro/Fejér summation: Averaging partial sums gives \(\sigma_N f\), which converges uniformly on \(\mathbb{T}\) for continuous \(f\). For discontinuous \(f\), \(\sigma_N f\) converges to \([f(x^+)+f(x^-)]/2\) pointwise. The Fejér kernel is non-negative, which prevents overshoot.
2. Lanczos \(\sigma\)-factors: Multiply Fourier coefficients by \(\sigma(n/N) = \text{sinc}(n/N)\) before summing. This smooths the cutoff and reduces overshoot by introducing a taper.
3. Spectral resampling: In computational methods, use a different basis (wavelets, polynomials) near discontinuities.

---

Chapter 20: Abel Summability and the Poisson Integral

Abel Summability: Definition and Basic Properties

\[ A\text{-lim}_{r\to 1^-}\sum_{n=0}^\infty a_n r^n = L. \]\[ A\text{-lim}_{r\to 1^-}\sum_{n=-\infty}^\infty a_n r^{|n|} = L. \]

Abel summability is strictly weaker than convergence. The classic example: \(\sum_{n=0}^\infty (-1)^n\) diverges (partial sums alternate 1, 0, 1, 0, …) but is Abel summable to \(1/2\): \(\sum (-1)^n r^n = \frac{1}{1+r} \to \frac{1}{2}\) as \(r\to 1^-\).

Fourier Series is Always Abel Summable

Back to top