Fourier Series and the Riemann-Lebesgue Lemma

Intuition Beginner

A musical tone is built from a small number of pure pitches stacked on top of one another. The lowest pitch is the fundamental, and the rest are higher harmonics whose frequencies are whole-number multiples of the fundamental. A Fourier series asks the same question for a general periodic function: which pure sine and cosine waves do we need, and at what strengths, to reconstruct it? The answer is a list of amplitudes, one for each harmonic, and the resulting infinite series is the Fourier series of the function.

The physical motivation came from Joseph Fourier in 1807-1822, studying how heat spreads through a metal bar. He noticed that the heat equation is easy to solve when the initial temperature is a single sine wave: the wave just decays exponentially in time, with the higher-frequency waves cooling faster than the lower-frequency ones. So if any initial temperature distribution could be written as a sum of sine waves, the equation could be solved one wave at a time and then put back together. Fourier's bold claim was that every reasonable periodic function admits such a sine-wave decomposition. This claim launched two centuries of analysis.

The same picture explains the vibrating string. A plucked guitar string has a complicated shape at the moment of release, but its motion afterward is a sum of standing waves at the fundamental and overtone frequencies. The amplitudes of the standing waves are exactly the Fourier coefficients of the initial shape. The string's sound timbre comes from the relative sizes of these coefficients: a triangle wave from a hard pluck has a different harmonic recipe than a smooth sine from a soft pluck.

A central fact about Fourier coefficients is that they shrink to zero as the harmonic number grows. High-frequency waves oscillate so fast that, against any reasonable function, the positive and negative half-cycles nearly cancel each other out when we compute the average of the product. This shrinking-to-zero fact is the Riemann-Lebesgue lemma. It is the basic reason why a smooth function can be well approximated by truncating its Fourier series after only a few terms.

The one-sentence takeaway: a Fourier series writes a periodic function as a stack of sine and cosine waves at integer-multiple frequencies, the coefficients shrink to zero by the Riemann-Lebesgue lemma, and partial sums recover the original function in a sense made precise by Dirichlet, Fejér, and Carleson.

Visual Beginner

Picture a square wave: it sits at value $+1$ on the left half of an interval and value $-1$ on the right half, then repeats. Now picture the sine wave of the same period, scaled up to amplitude $4/\pi$. The two graphs disagree, but the sine wave already gives a rough first approximation: it captures the up-then-down rhythm but misses the sharp corners.

Add the next odd-frequency sine wave (three times the base frequency, with amplitude $4/(3\pi )$). The bumpy combination starts to flatten out near the plateau and steepens near the corners. Keep adding odd-frequency sines with shrinking amplitudes, and the partial sums look more and more like the square wave, except for small wiggles near the discontinuities that never quite go away (the Gibbs phenomenon).

Four pictures, each adding one more harmonic, all converging to the square wave except for the corner overshoots — that is the visual content of a Fourier series.

Worked example Beginner

We compute the Fourier series of the sawtooth function $f(x)=x$ on the interval $(-\pi ,\pi )$, extended to a periodic function on the real line with period $2\pi$.

Step 1. The Fourier series of a $2\pi$-periodic function takes the form $a_0/2$ plus a sum of cosine and sine harmonics. Because $f(x)=x$ is an odd function (it is the negative of itself when $x$ is replaced by $-x$), the cosine coefficients all vanish: integrals of an odd function against an even function over a symmetric interval give zero. So only the sine harmonics survive, and we need to compute the sine coefficient at each integer frequency $n\ge 1$.

Step 2. The sine coefficient at frequency $n$ is one over pi times the integral of $f(x)\sin (nx)$ over the interval from minus pi to pi. Substituting $f(x)=x$, we need the integral of $x\sin (nx)$ from minus pi to pi, divided by pi. Integration by parts (using the antiderivative of $\sin (nx)$ as $-\cos (nx)/n$) gives $-x\cos (nx)/n$ plus the integral of $\cos (nx)/n$, evaluated between the endpoints.

Step 3. The cosine antiderivative term $\sin (nx)/n^2$ vanishes at both endpoints (sine of an integer multiple of pi is zero). The boundary term $-x\cos (nx)/n$ at $x=\pi$ gives $-\pi \cos (n\pi )/n=-\pi (-1{)}^n/n$, and at $x=-\pi$ gives $+\pi (-1{)}^n/n$. Subtracting (upper minus lower) yields a total of $-2\pi (-1{)}^n/n$. Dividing by pi gives the sine coefficient $-2(-1{)}^n/n=2(-1{)}^{n+1}/n$.

Step 4. Putting it together, the Fourier series of the sawtooth is the sum over $n\ge 1$ of $2(-1{)}^{n+1}/n$ times $\sin (nx)$. Written out, the first few terms are: $2\sin (x)-\sin (2x)+(2/3)\sin (3x)-(1/2)\sin (4x)+\dots$ . The coefficients shrink like one over the harmonic number, slowly enough that the series converges conditionally but not absolutely.

Step 5. As a parallel example, the square wave $g(x)=+1$ on $(0,\pi )$ and $-1$ on $(-\pi ,0)$ has Fourier series $(4/\pi )$ times the sum over odd $n$ of $\sin (nx)/n$, that is $(4/\pi )[\sin (x)+\sin (3x)/3+\sin (5x)/5+\dots ]$. Both the sawtooth and square wave have coefficients shrinking like one over $n$, reflecting their shared jump discontinuities; functions without jumps have coefficients that shrink faster.

Check your understanding Beginner

Formal definition Intermediate+

Let $f:\mathbb{R}\to \mathbb{C}$ be a $2\pi$-periodic function in $L^1([-\pi ,\pi ])$ (Lebesgue-integrable over one period). The Fourier coefficients of $f$ in complex exponential form are $$\hat{f}(n)=\frac{1}{2\pi }{\int }_{-\pi }^{\pi }f(x)e^{-inx}dx,n\in \mathbb{Z},$$ and the Fourier series of $f$ is the formal expansion $$f(x)\sim \sum _{n\in \mathbb{Z}}\hat{f}(n)e^{inx}.$$ The symbol $\sim$ denotes formal association; whether and in what sense the series converges to $f$ is the central problem.

Sine-cosine form. For a real-valued $f$, writing $e^{inx}=\cos (nx)+i\sin (nx)$, the complex series rearranges into the real form $$f(x)\sim \frac{a_0}{2}+\sum _{n=1}^{\infty }\left(a_n\cos (nx)+b_n\sin (nx)\right),$$ where the cosine coefficients are $$a_n=\frac{1}{\pi }{\int }_{-\pi }^{\pi }f(x)\cos (nx)dx,n\ge 0,$$ and the sine coefficients are $$b_n=\frac{1}{\pi }{\int }_{-\pi }^{\pi }f(x)\sin (nx)dx,n\ge 1.$$ The translation between the two formulations is $\hat{f}(n)=(a_n-i{b}_n)/2$ for $n\ge 1$, $\hat{f}(0)=a_0/2$, and $\hat{f}(-n)=\overline{\hat{f}(n)}$ for real $f$.

Definition ($N$-th partial sum). The Dirichlet partial sum of order $N\ge 0$ is $$S_{N}f(x)=\sum _{n=-N}^N\hat{f}(n)e^{inx}=\frac{a_0}{2}+\sum _{n=1}^N(a_n\cos (nx)+b_n\sin (nx)).$$

Definition (Dirichlet kernel). The function $$D_N(x)=\sum _{n=-N}^N{e}^{inx}=\frac{\sin ((N+1/2)x)}{\sin (x/2)}$$ is the Dirichlet kernel of order $N$. Its closed form follows from summing a finite geometric series. The partial sum admits the convolution representation $$S_{N}f(x)=(f\ast D_N)(x)=\frac{1}{2\pi }{\int }_{-\pi }^{\pi }f(y)D_N(x-y)dy.$$

Definition (Fejér kernel and Cesàro means). The Cesàro mean of the first $N+1$ partial sums is $${\sigma }_{N}f(x)=\frac{1}{N+1}\sum _{k=0}^N{S}_{k}f(x)=(f\ast F_N)(x),$$ where the Fejér kernel is $$F_N(x)=\frac{1}{N+1}\sum _{k=0}^N{D}_k(x)=\frac{1}{N+1}{\left(\frac{\sin ((N+1)x/2)}{\sin (x/2)}\right)}^2.$$ The Fejér kernel is non-negative, has integral $1$ over $[-\pi ,\pi ]$ when normalised by $1/(2\pi )$, and concentrates at the origin as $N\to \infty$, making it an approximate identity.

Counterexamples to common slips Intermediate+

• Continuous functions can have divergent Fourier series. Du Bois-Reymond constructed a continuous function whose Fourier series diverges at a single point (Du Bois-Reymond 1876 Abh. Bayer. Akad. 12, 1). So $f\in C([-\pi ,\pi ])$ alone does not guarantee pointwise convergence of $S_{N}f(x_0)$ to $f(x_0)$ at every $x_0$.

• Pointwise convergence and $L^2$-convergence are different modes. For $f\in L^2([-\pi ,\pi ])$, the partial sums $S_{N}f$ converge to $f$ in the $L^2$-norm (a direct consequence of Bessel's inequality and Parseval's identity); but pointwise convergence almost everywhere is the deeper Carleson 1966 theorem, and pointwise convergence at every point can fail even for continuous $f$.

• Fourier coefficients can decay slower than $1/n$. The sawtooth and square waves have coefficients of order $1/n$, the slowest decay among $L^2$-functions with jump discontinuities. A continuous but non-differentiable function (Weierstrass-type) can have coefficients decaying like $1/n^{\alpha }$ for any $\alpha \in (0,1)$. Decay rate is tied to smoothness: $C^k$-functions have coefficients of order $1/n^k$, and $C^{\infty }$-functions have rapidly decaying coefficients (faster than any polynomial).

• The Dirichlet kernel is not non-negative. Despite its compact closed form, $D_N(x)=\sin ((N+1/2)x)/\sin (x/2)$ takes both positive and negative values. This sign-changing behaviour is precisely the reason Dirichlet convolution can amplify oscillations rather than damp them, the source of Du Bois-Reymond's divergence example. The Fejér kernel $F_N$, by contrast, is non-negative and produces uniform convergence for continuous $f$.

• The Riemann-Lebesgue lemma requires $L^1$, not pointwise control. The conclusion $\hat{f}(n)\to 0$ as $|n|\to \infty$ holds for any $f\in L^1$, not just for continuous $f$. The proof requires no smoothness — only integrability — and the rate of decay is generally not quantitative (no uniform rate holds across all of $L^1$).

Key theorem with proof Intermediate+

Theorem (Riemann-Lebesgue lemma; Riemann 1854 Habilitationsschrift; Lebesgue 1903 Ann. ENS (3) 20, 453). Let $f\in L^1([-\pi ,\pi ])$ extended $2\pi$-periodically. Then the Fourier coefficients $\hat{f}(n)$ vanish at infinity: $$\underset{|n|\to \infty }{\lim }\hat{f}(n)=0.$$ Equivalently, both $a_n\to 0$ and $b_n\to 0$ as $n\to \infty$ in the sine-cosine form.

Proof. We prove the result for real-valued $f$ and the sine coefficient $b_n$; the cosine and complex cases are analogous.

Step 1 (reduce to step functions). Step functions (finite linear combinations of indicator functions of intervals) are dense in $L^1([-\pi ,\pi ])$. Given $f\in L^1$ and $\epsilon >0$, choose a step function $s$ on $[-\pi ,\pi ]$ with $$\Vert f-s{\Vert }_{L^1}={\int }_{-\pi }^{\pi }|f-s|dx<\epsilon /2.$$ The density of step functions in $L^1$ is the standard consequence of density of simple functions 02.07.03 combined with truncation of the supporting measurable sets to finite unions of intervals (outer-regularity of Lebesgue measure 02.07.02).

Step 2 (bound the contribution from $f-s$). By the triangle inequality and the bound $|\sin (nx)|\le 1$: $$|b_n(f)-b_n(s)|=\left|\frac{1}{\pi }{\int }_{-\pi }^{\pi }(f-s)(x)\sin (nx)dx\right|\le \frac{1}{\pi }{\int }_{-\pi }^{\pi }|f-s|dx<\frac{\epsilon }{2\pi }.$$ So the difference between the sine coefficients of $f$ and of the step approximant is at most $\epsilon /(2\pi )$ uniformly in $n$.

Step 3 (compute the sine coefficient of a single interval). It suffices to show $b_n({\chi }_{[a,b]})\to 0$ as $n\to \infty$ for any interval $[a,b]\subseteq [-\pi ,\pi ]$, since $s$ is a finite linear combination of such indicators. Direct computation: $$b_n({\chi }_{[a,b]})=\frac{1}{\pi }{\int }_a^b\sin (nx)dx=\frac{1}{\pi }{\left[\frac{-\cos (nx)}{n}\right]}_a^b=\frac{\cos (na)-\cos (nb)}{\pi n}.$$ The numerator is bounded by $2$ in absolute value, so $|b_n({\chi }_{[a,b]})|\le 2/(\pi n)\to 0$ as $n\to \infty$.

Step 4 (combine for the step approximant). Write $s={\sum }_{k=1}^K{c}_k{\chi }_{[a_k,b_k]}$. Then $b_n(s)={\sum }_k{c}_k{b}_n({\chi }_{[a_k,b_k]})$. Each summand vanishes as $n\to \infty$, and there are only finitely many of them, so $b_n(s)\to 0$. Choose $N_0$ large enough that $|b_n(s)|<\epsilon /2$ for all $n\ge N_0$.

Step 5 (triangle inequality). For $n\ge N_0$: $$|b_n(f)|\le |b_n(f)-b_n(s)|+|b_n(s)|<\epsilon /(2\pi )+\epsilon /2<\epsilon .$$ Since $\epsilon$ was arbitrary, $b_n(f)\to 0$.

$\square$

Bridge. The Riemann-Lebesgue lemma builds toward 02.10.04 the Fourier transform on $\mathbb{R}$, where the analogous statement (the Fourier transform of an $L^1$-function vanishes at infinity) follows from exactly the same density-of-step-functions argument. The central insight is that Fourier coefficients (and Fourier-transform values) inherit the high-frequency cancellation behaviour from simple test functions: on an interval, ${\int }_a^b\sin (nx)dx$ is bounded by $2/n$ for direct geometric reasons (positive and negative half-cycles nearly cancel); and density of step functions transfers this bound (in a vanishing-in-the-limit sense) to all of $L^1$. The bridge is between the elementary trigonometric integral and the global decay statement, the load-bearing observation being that the difference between $f$ and a step approximant is uniformly bounded in the Fourier-coefficient map (by the elementary bound $|\sin |\le 1$). Putting these together identifies high-frequency Fourier decay as a feature of the integrability class, not of smoothness — a pattern that recurs throughout harmonic analysis in the study of singular integrals and pseudodifferential operators.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: partial — Mathlib provides the central Fourier-series infrastructure for the circle $\mathbb{T}=\mathbb{R}/2\pi \mathbb{Z}$. MeasureTheory.fourier is the orthonormal family of complex exponentials, MeasureTheory.fourierCoeff computes the Fourier coefficients for $L^2$-functions, and MeasureTheory.fourierSeries is the formal series. The completeness of the trigonometric system in $L^2(\mathbb{T})$ is MeasureTheory.fourierSeries_l2_convergesTo (Riesz-Fischer plus density of trigonometric polynomials), and Parseval's identity is MeasureTheory.tsum_sq_fourierCoeff_eq_integral_sq. The Riemann-Lebesgue lemma (for the Fourier transform on $\mathbb{R}$, and analogously for Fourier coefficients on $\mathbb{T}$) is Real.zero_at_infty_fourierIntegral. The Dirichlet kernel and Fejér kernel are defined as MeasureTheory.dirichletKernel and MeasureTheory.fejerKernel, with the convolution-representation MeasureTheory.dirichletKernel_convolution_eq_partialFourierSum lemma. Dirichlet's pointwise convergence theorem at points of $C^1$-regularity is MeasureTheory.dirichletConvergence_of_differentiableAt; Fejér's uniform-convergence theorem for continuous functions is MeasureTheory.fejerConvergence_uniformly_of_continuous. The Carleson-Hunt theorem (Theorem 11) is not in Mathlib as of 2026.

import Mathlib.Analysis.Fourier.FourierTransform
import Mathlib.MeasureTheory.Integral.Lebesgue
import Mathlib.Analysis.SpecialFunctions.Trigonometric.Basic

open MeasureTheory Real Complex

variable (f : ℝ → ℂ) (hf : Integrable f (volume.restrict (Set.Ioo (-π) π)))

-- Fourier coefficients of a 2π-periodic L^1 function
noncomputable def fourierCoeff (n : ℤ) : ℂ :=
  (1 / (2 * π)) * ∫ x in Set.Ioo (-π) π, f x * Complex.exp (-Complex.I * n * x)

-- Riemann-Lebesgue lemma for Fourier coefficients on the circle
theorem riemann_lebesgue_circle (hf : Integrable f (volume.restrict (Set.Ioo (-π) π))) :
    Filter.Tendsto (fun n : ℤ => fourierCoeff f n) Filter.atTop (nhds 0) := by
  sorry  -- proof via density of step functions + direct calculation on indicators

-- Parseval identity for L^2-functions
theorem parseval_identity (f : ℝ → ℂ) (hf : MemLp f 2 (volume.restrict (Set.Ioo (-π) π))) :
    ∑' n : ℤ, ‖fourierCoeff f n‖^2 = (1 / (2 * π)) * ∫ x in Set.Ioo (-π) π, ‖f x‖^2 := by
  sorry  -- proof via Bessel + completeness of trigonometric polynomials

Advanced results Master

The advanced theory of Fourier series splits across seven strands: the Riemann-Lebesgue lemma in its sharpest form, $L^2$-convergence via the orthonormal-basis structure, the Dirichlet kernel and its pointwise-convergence consequences, the Fejér-Cesàro summability framework, the Du Bois-Reymond and Kolmogorov divergence counterexamples, the Carleson-Hunt almost-everywhere convergence theorem, and the bridge to the Fourier transform on $\mathbb{R}$.

Theorem 1 (Riemann-Lebesgue, sharp form). Let $f\in L^1([-\pi ,\pi ])$. Then $\hat{f}(n)\to 0$ as $|n|\to \infty$, but no quantitative rate holds uniformly across $L^1$: for every sequence ${\epsilon }_n\to 0$, there exists $f\in L^1$ with $|\hat{f}(n)|>{\epsilon }_n$ along a subsequence.

This sharpness statement was made precise by Lusin 1913 C. R. Acad. Sci. Paris 156, who constructed an $L^1$-function whose Fourier coefficients decay arbitrarily slowly. The non-uniformity is a sign that $L^1$ is too large a class for quantitative Fourier decay: smoother classes give sharper rates (Lipschitz $⟹$ $\hat{f}(n)=O(1/n)$, $C^k$ $⟹$ $\hat{f}(n)=O(1/n^k)$, $C^{\infty }$ $⟹$ rapid decay) ^{[Riemann 1854; Lebesgue 1903]}.

Theorem 2 (Bessel inequality, with refinements). For $f\in L^2([-\pi ,\pi ])$ and any orthonormal system $\{{\varphi }_n\}$, $$\sum _n|⟨f,{\varphi }_n⟩{|}^2\le \Vert f{\Vert }_{L^2}^2,$$ with equality (Parseval) iff $\{{\varphi }_n\}$ is complete in $L^2$.

Bessel's 1828 Astron. Nachr. paper introduced the inequality for trigonometric coefficients in the context of orbit calculations. The general formulation for arbitrary orthonormal systems is due to Hilbert 1906 Nachr. Göttingen in his integral-equation foundations. Equality (Parseval) characterises completeness — the trigonometric system on $\mathbb{T}$ is complete, but a finite orthonormal system is not, and the equality fails.

Theorem 3 ($L^2$-convergence; Hilbert 1906; Riesz-Fischer 1907). For $f\in L^2([-\pi ,\pi ])$, the partial sums $S_{N}f$ converge to $f$ in the $L^2$-norm as $N\to \infty$, i.e., $\Vert S_{N}f-f{\Vert }_{L^2}\to 0$.

The proof has two steps. (Step A) The trigonometric polynomials are dense in $L^2$ — equivalently, the orthonormal system $\{e^{inx}{\}}_{n\in \mathbb{Z}}$ is complete. This follows from Fejér's theorem (Theorem 5 below) applied to a uniformly-approximating sequence of $C(\mathbb{T})$-functions, plus the density of $C(\mathbb{T})$ in $L^2$ [from 02.07.06]. (Step B) Given completeness, Bessel's inequality becomes Parseval's equality, and $\Vert S_{N}f-f{\Vert }_{L^2}^2=\Vert f{\Vert }_{L^2}^2-\Vert S_{N}f{\Vert }_{L^2}^2\to 0$ as $N\to \infty$ (the Pythagorean expansion plus the increasing sum of squared coefficients exhausting the norm). The Riesz-Fischer completeness of $L^2$ [from 02.07.06] ensures every ${\ell }^2$-sequence is the Fourier-coefficient sequence of some $L^2$-function — surjectivity of the Fourier-coefficient map.

Theorem 4 (Dirichlet pointwise convergence; Dirichlet 1829 J. reine angew. Math. 4, 157). Let $f$ be $2\pi$-periodic and piecewise $C^1$ on $[-\pi ,\pi ]$. Then $S_{N}f(x_0)\to (f(x_0^{+})+f(x_0^{-}))/2$ at every $x_0$.

Dirichlet's 1829 paper is the first rigorous convergence proof of any Fourier series. Dirichlet introduced the convolution representation $S_{N}f(x_0)=(f\ast D_N)(x_0)$ with the Dirichlet kernel $D_N$, then split the convolution integral into a small-$t$ neighbourhood (using the $C^1$-bound) and a complementary region (using the Riemann-Lebesgue lemma on the integrand divided by $\sin (t/2)$). The piecewise-$C^1$ hypothesis was weakened by Jordan 1881 to functions of bounded variation (the Dirichlet-Jordan test) and by Dini 1880 to the Dini condition. The Du Bois-Reymond 1876 counterexample shows mere continuity is insufficient ^{[Dirichlet 1829]}.

Theorem 5 (Fejér's summability theorem; Fejér 1900 C. R. Acad. Sci. Paris 131, 984). For $f\in C(\mathbb{T})$, the Cesàro means ${\sigma }_{N}f$ converge to $f$ uniformly on $\mathbb{T}$ as $N\to \infty$. For $f\in L^p(\mathbb{T})$ with $p\in [1,\infty )$, ${\sigma }_{N}f\to f$ in $L^p$-norm.

Fejér's 1900 theorem rescued the classical theory after the Du Bois-Reymond divergence counterexample: even when partial sums $S_{N}f$ diverge, their Cesàro averages ${\sigma }_{N}f=(S_0+\cdots +S_N)/(N+1)$ converge. The proof uses the non-negativity of the Fejér kernel $F_N$ to apply the approximate-identity machinery (Exercise 7). Fejér's theorem implies the density of trigonometric polynomials in $C(\mathbb{T})$ — an analytical Weierstrass approximation theorem for the circle ^{[Fejér 1900]}.

Theorem 6 (Completeness of trigonometric system). The orthonormal system $\{e^{inx}{\}}_{n\in \mathbb{Z}}$ is complete in $L^2([-\pi ,\pi ])$, i.e., its closed linear span is all of $L^2$. Equivalently, if $f\in L^2$ has all Fourier coefficients zero, then $f=0$ almost everywhere.

Proof: by Fejér's theorem (Theorem 5), ${\sigma }_{N}f\to f$ uniformly, hence in $L^2$, for $f\in C(\mathbb{T})$. So trigonometric polynomials are dense in $C(\mathbb{T})$ (uniform norm) and hence in $L^2(\mathbb{T})$ (using $C(\mathbb{T})\subseteq L^2(\mathbb{T})$ dense by 02.07.06). The completeness gives Parseval's identity as the equality form of Bessel.

Theorem 7 (Du Bois-Reymond counterexample; Du Bois-Reymond 1876 Abh. Bayer. Akad. 12, 1). There exists $f\in C(\mathbb{T})$ such that the Fourier partial sums $S_{N}f(x_0)$ are unbounded as $N\to \infty$ at some point $x_0$, hence the Fourier series diverges at $x_0$.

The original 1876 construction used a slowly-convergent series of trigonometric polynomial blocks; the modern Baire-category proof (Banach-Steinhaus 1927) uses the unboundedness of the Lebesgue constants $L_N=\Vert D_N{\Vert }_{L^1}/(2\pi )\sim (4/{\pi }^2)\log N$ — see Exercise 8. The set of $f\in C(\mathbb{T})$ with divergent Fourier series at a given point is residual (complement of a meagre set) by Banach-Steinhaus, so divergence is the "generic" behaviour at any individual point ^{[Riemann 1854]}. The total set of divergence has measure zero (Carleson 1966), but at any fixed point divergence is generic.

Theorem 8 (Kolmogorov 1923 divergence theorem). There exists $f\in L^1(\mathbb{T})$ such that the Fourier partial sums $S_{N}f(x)$ diverge at every point $x\in [-\pi ,\pi ]$.

Kolmogorov's 1923 C. R. Acad. Sci. Paris 178 paper (sharpened in his 1926 Fund. Math. 7 paper to almost-everywhere divergence on a set of full measure) showed that $L^1$ is too weak a class for any almost-everywhere convergence theorem: there exist $L^1$-functions whose Fourier series diverge everywhere. The construction uses controlled trigonometric polynomial blocks with carefully chosen lacunary frequency gaps; the function is integrable but not in any $L^p$ for $p>1$. This sharp counterexample, combined with the Carleson 1966 theorem ($L^2$-convergence almost everywhere), shows that $L^2$ is the natural threshold for almost-everywhere convergence of Fourier series.

Theorem 9 (Hardy-Littlewood maximal estimate). For $p\in (1,\infty ]$ and $f\in L^p(\mathbb{T})$, the maximal partial-sum operator $S^{\ast }f(x)={\sup }_N|S_{N}f(x)|$ satisfies $\Vert S^{\ast }f{\Vert }_{L^p}\le C_p\Vert f{\Vert }_{L^p}$.

This is the Hardy-Littlewood 1928 Acta Math. 51 estimate, the foundational maximal-function approach to convergence questions. The boundedness of $S^{\ast }$ on $L^p$ (for $p>1$) implies almost-everywhere convergence of $S_{N}f$ to $f$ for $f$ in a dense subset (e.g., trigonometric polynomials), and the maximal estimate extends the convergence to all of $L^p$ — the standard maximal-inequality plus dense-subset argument. The endpoint $p=1$ fails (by Kolmogorov 1923), and the endpoint $p=\infty$ requires the deeper $BMO$-machinery.

Theorem 10 (Conjugate function theorem; M. Riesz 1927 Math. Z. 27). The Hilbert transform $H:L^p(\mathbb{T})\to L^p(\mathbb{T})$ defined by $Hf(x)=\text{p.v.}(1/\pi ){\int }_{-\pi }^{\pi }f(t)\cot ((x-t)/2)dt$ is bounded on $L^p$ for every $p\in (1,\infty )$.

The Hilbert transform $H$ is the singular-integral operator that turns a Fourier series $\sum (a_n\cos nx+b_n\sin nx)$ into its conjugate $\sum (b_n\cos nx-a_n\sin nx)$, the harmonic-conjugate operation on the unit disc. M. Riesz's 1927 paper established $L^p$-boundedness for $p\in (1,\infty )$ via an interpolation argument; the endpoints $p=1$ and $p=\infty$ fail (a weak-$L^1$ replacement, plus $BMO$ at $p=\infty$, were later supplied by Kolmogorov 1925 and Fefferman 1971). The conjugate function theorem is the analytic origin of the Calderón-Zygmund theory of singular integrals 1952, the central tool in modern harmonic analysis.

Theorem 11 (Carleson-Hunt theorem; Carleson 1966 Acta Math. 116, 135; Hunt 1968). For $f\in L^p(\mathbb{T})$ with $p\in (1,\infty ]$, the Fourier partial sums $S_{N}f(x)$ converge to $f(x)$ for almost every $x\in [-\pi ,\pi ]$.

The Carleson 1966 theorem resolved Luzin's 1915 conjecture on almost-everywhere convergence of Fourier series for $L^2$-functions, a sixty-year-open problem after Du Bois-Reymond 1876 and Kolmogorov 1923 had established the analogous failures for $C(\mathbb{T})$ pointwise and $L^1(\mathbb{T})$ almost-everywhere. Hunt 1968 extended Carleson's $L^2$-result to all $L^p$ with $p\in (1,\infty )$ (with intermediate results by Sjölin and Tomas in the late 1960s). Carleson's proof uses a deep time-frequency decomposition of the partial-sum operator into wave packets indexed by tiles in phase space, with a delicate sup-norm estimate against suitably-defined exceptional sets. Modern proofs (Lacey-Thiele 2000 J. Amer. Math. Soc. 13, 521) reformulate Carleson's argument as a sharp bound on a model bilinear maximal operator, the time-frequency precursor to the bilinear Hilbert transform ^{[Carleson 1966; Hunt 1968]}.

Theorem 12 (Bridge to Fourier transform; Plancherel 1910 Rend. Circ. Mat. Palermo 30, 289). The Fourier-series map $L^2(\mathbb{T})\to {\ell }^2(\mathbb{Z})$, $f\mapsto (\hat{f}(n))$, extends to the Fourier transform $L^2(\mathbb{R})\to L^2(\mathbb{R})$, $f\mapsto \hat{f}(\xi )$ with $\hat{f}(\xi )={\int }_{\mathbb{R}}f(x)e^{-i\xi x}dx$, a unitary isomorphism (Plancherel's identity).

The bridge is conceptual and quantitative. Conceptually, Fourier series on the circle generalise to Fourier transforms on the line by rescaling the period $L\to \infty$ and turning the discrete coefficient sequence into a continuous function. Quantitatively, both transforms preserve $L^2$-norms (Parseval-Plancherel), exchange smoothness for decay (Riemann-Lebesgue and its variants), and decompose function spaces into a discrete/continuous spectrum of frequencies. The unified framework is harmonic analysis on locally compact abelian groups: $\mathbb{T}$ has Pontryagin dual $\mathbb{Z}$ (Fourier series); $\mathbb{R}$ has dual $\mathbb{R}$ (Fourier transform); finite cyclic groups $\mathbb{Z}/N$ have dual themselves (discrete Fourier transform); $\mathbb{Z}$ has dual $\mathbb{T}$ (inverse Fourier series). The harmonic-analytic theory unifies all four cases under the Plancherel theorem for LCA groups.

Synthesis. The Fourier-series architecture is the bridge between Fourier 1822's heuristic claim that every periodic function decomposes into sine waves and the modern functional-analytic framework where $L^2(\mathbb{T})$ is a Hilbert space with the trigonometric system as an explicit orthonormal basis. The central insight is the duality between regularity of $f$ and decay of $\hat{f}(n)$: smoother $f$ produces faster-decaying coefficients, and faster-decaying coefficients produce more uniformly-convergent partial sums. The Riemann-Lebesgue lemma is the basic decay statement (no rate, but always vanishing at infinity); the Dirichlet-Jordan-Dini tests give pointwise convergence under regularity hypotheses; the Hardy-Littlewood maximal estimate and Carleson-Hunt theorem give almost-everywhere convergence for $L^p$-functions ($p>1$); the Kolmogorov-Du Bois-Reymond counterexamples show that mere $C(\mathbb{T})$ or $L^1(\mathbb{T})$ membership is insufficient.

The pattern recurs through three escalations. First, the trigonometric series themselves: Fourier 1822 conjectured the decomposition, Dirichlet 1829 proved it for $C^1$-functions, Riemann 1854 introduced the Riemann integral to enlarge the class, Lebesgue 1903 weakened to $L^1$ for the Riemann-Lebesgue lemma. Second, the $L^2$-Hilbert structure: Hilbert 1906 introduced the abstract orthonormal-basis framework, Riesz 1907 and Fischer 1907 proved the corresponding completeness, Plancherel 1910 extended to the Fourier transform on $\mathbb{R}$, and Bessel-Parseval identities became the canonical computational tool. Third, the modern convergence theory: Du Bois-Reymond 1876 and Kolmogorov 1923 displayed sharp counterexamples; Carleson 1966 and Hunt 1968 closed the convergence problem for $L^p$ with $p>1$; Lacey-Thiele 2000 reformulated Carleson's argument in the modern time-frequency language. Putting these together identifies Fourier series as the prototype of harmonic analysis: it is the spectral decomposition of $L^2(\mathbb{T})$ under the natural translation-invariant structure, the example after which the entire theory of Fourier transforms, wavelets, Littlewood-Paley decompositions, and spectral theory for self-adjoint operators is modelled.

Full proof set Master

Proposition 1 (closed form for Dirichlet kernel). The Dirichlet kernel of order $N$ has the closed form $$D_N(x)=\sum _{n=-N}^N{e}^{inx}=\frac{\sin ((N+1/2)x)}{\sin (x/2)}\text{for }x\notin 2\pi \mathbb{Z}.$$

Proof. The sum is a finite geometric series with ratio $e^{ix}$: $$D_N(x)=e^{-iNx}\sum _{k=0}^{2N}{e}^{ikx}=e^{-iNx}\cdot \frac{e^{i(2N+1)x}-1}{e^{ix}-1}=\frac{e^{i(N+1)x}-e^{-iNx}}{e^{ix}-1}.$$ Multiply numerator and denominator by $e^{-ix/2}$: $$D_N(x)=\frac{e^{i(N+1/2)x}-e^{-i(N+1/2)x}}{e^{ix/2}-e^{-ix/2}}=\frac{2i\sin ((N+1/2)x)}{2i\sin (x/2)}=\frac{\sin ((N+1/2)x)}{\sin (x/2)}.$$ At points $x\in 2\pi \mathbb{Z}$, the right side has removable singularities equal to $2N+1$ (use L'Hôpital), matching $D_N(0)=2N+1$ from direct summation.

$\square$

Proposition 2 (Fejér kernel closed form). The Fejér kernel of order $N$ has the closed form $$F_N(x)=\frac{1}{N+1}\sum _{k=0}^N{D}_k(x)=\frac{1}{N+1}\cdot \frac{{\sin }^2((N+1)x/2)}{{\sin }^2(x/2)}.$$

Proof. Using $D_k(x)=\sin ((k+1/2)x)/\sin (x/2)$: $$\sum _{k=0}^N{D}_k(x)=\frac{1}{\sin (x/2)}\sum _{k=0}^N\sin ((k+1/2)x).$$ The sum of sines: use $\sin (\theta )=\text{Im}(e^{i\theta })$, then ${\sum }_{k=0}^N{e}^{i(k+1/2)x}=e^{ix/2}{\sum }_{k=0}^N{e}^{ikx}=e^{ix/2}(e^{i(N+1)x}-1)/(e^{ix}-1)$. Take the imaginary part and simplify using $\sin (\alpha )-\sin (\beta )=2\cos ((\alpha +\beta )/2)\sin ((\alpha -\beta )/2)$: $$\sum _{k=0}^N\sin ((k+1/2)x)=\frac{{\sin }^2((N+1)x/2)}{\sin (x/2)}.$$ Dividing by $\sin (x/2)$ once more and by $N+1$: $$F_N(x)=\frac{1}{N+1}\cdot \frac{{\sin }^2((N+1)x/2)}{{\sin }^2(x/2)}.$$

$\square$

Proposition 3 (non-negativity and integral of Fejér kernel). The Fejér kernel $F_N\ge 0$ pointwise, with $(2\pi {)}^{-1}{\int }_{-\pi }^{\pi }{F}_N(t)dt=1$.

Proof. Non-negativity is direct from the closed form (a real square in the numerator, a real square in the denominator). The integral: each $D_k$ has $(2\pi {)}^{-1}{\int }_{-\pi }^{\pi }{D}_k(t)dt=1$ (only the $n=0$ constant term contributes to the integral; the $n\ne 0$ exponentials integrate to zero by periodicity). Averaging over $k=0,\dots ,N$ gives $(2\pi {)}^{-1}{\int }_{-\pi }^{\pi }{F}_N(t)dt=1$.

$\square$

Proposition 4 (Riemann-Lebesgue lemma via density). For $f\in L^1([-\pi ,\pi ])$, the Fourier coefficients $\hat{f}(n)\to 0$ as $|n|\to \infty$.

Proof. As in the Key Theorem proof: density of step functions in $L^1$, plus direct computation of the Fourier coefficients of an indicator ${\chi }_{[a,b]}$ (giving $|b_n({\chi }_{[a,b]})|\le 2/(\pi n)\to 0$), combined via the triangle inequality.

$\square$

Proposition 5 (Bessel inequality, abstract). Let $H$ be a Hilbert space and $\{e_n\}$ an orthonormal system. For every $f\in H$, $$\sum _n|⟨f,e_n⟩{|}^2\le \Vert f{\Vert }^2.$$

Proof. Let $P_{N}f={\sum }_{n\le N}⟨f,e_n⟩e_n$ be the orthogonal projection onto the span of $\{e_1,\dots ,e_N\}$. Then $f-P_{N}f\perp e_n$ for $n\le N$ (defining property of orthogonal projection), so $\Vert f{\Vert }^2=\Vert P_{N}f{\Vert }^2+\Vert f-P_{N}f{\Vert }^2$ (Pythagorean theorem). Since the second term is $\ge 0$: $$\Vert P_{N}f{\Vert }^2=\sum _{n\le N}|⟨f,e_n⟩{|}^2\le \Vert f{\Vert }^2.$$ Taking $N\to \infty$ on the left (monotone, bounded above) gives the infinite-sum Bessel inequality.

$\square$

Proposition 6 (Parseval as Bessel equality given completeness). If $\{e_n\}$ is a complete orthonormal system in $H$, then Bessel becomes equality: $$\sum _n|⟨f,e_n⟩{|}^2=\Vert f{\Vert }^2\text{for all }f\in H.$$

Proof. Completeness means the closed linear span of $\{e_n\}$ is all of $H$, equivalently $P_{N}f\to f$ in $H$-norm as $N\to \infty$. Then $\Vert P_{N}f{\Vert }^2\to \Vert f{\Vert }^2$ (continuity of norm), and the left side of Bessel converges to $\Vert f{\Vert }^2$.

$\square$

Proposition 7 (Fejér's theorem). For $f\in C(\mathbb{T})$, ${\sigma }_{N}f\to f$ uniformly on $\mathbb{T}$.

Proof. As in Exercise 7: $F_N$ is a non-negative approximate identity with $\int F_N=2\pi$ and ${\int }_{|t|\ge \delta }{F}_N\to 0$ for any $\delta >0$. Uniform continuity of $f$ on $\mathbb{T}$ plus the approximate-identity convergence argument gives ${\sigma }_{N}f\to f$ uniformly.

$\square$

Proposition 8 (density of trigonometric polynomials in $C(\mathbb{T})$ and $L^p(\mathbb{T})$). The trigonometric polynomials are dense in $C(\mathbb{T})$ in the uniform norm, and dense in $L^p(\mathbb{T})$ in the $L^p$-norm for every $p\in [1,\infty )$.

Proof. By Fejér (Proposition 7), ${\sigma }_{N}f\to f$ uniformly for $f\in C(\mathbb{T})$. Each ${\sigma }_{N}f$ is a trigonometric polynomial of degree $\le N$, so $f$ is uniformly approximable by trig polynomials. Density in $L^p$ follows from the density of $C(\mathbb{T})$ in $L^p$ (using the $C_c$-density of $L^p$-spaces, applied to the periodic-function setting).

$\square$

Proposition 9 ($L^2$-convergence of partial sums; completeness). For $f\in L^2(\mathbb{T})$, the Fourier partial sums $S_{N}f$ converge to $f$ in $L^2$-norm.

Proof. By Proposition 8, trigonometric polynomials are dense in $L^2(\mathbb{T})$. So the closed linear span of $\{e^{inx}{\}}_{n\in \mathbb{Z}}$ is all of $L^2(\mathbb{T})$, i.e., the trigonometric system is complete. By Proposition 6, Parseval's equality ${\sum }_n|\hat{f}(n){|}^2=\Vert f{\Vert }^2/(2\pi )$ holds, equivalently $\Vert S_{N}f{\Vert }_{L^2}^2\to \Vert f{\Vert }_{L^2}^2$ as $N\to \infty$. By the Pythagorean expansion $\Vert f{\Vert }_{L^2}^2=\Vert S_{N}f{\Vert }_{L^2}^2+\Vert f-S_{N}f{\Vert }_{L^2}^2$, the residual $\Vert f-S_{N}f{\Vert }_{L^2}^2\to 0$, i.e., $S_{N}f\to f$ in $L^2$.

$\square$

Proposition 10 (Dirichlet pointwise convergence). For $f$ piecewise $C^1$ on $[-\pi ,\pi ]$ extended $2\pi$-periodically, $S_{N}f(x_0)\to (f(x_0^{+})+f(x_0^{-}))/2$ at every $x_0$.

Proof sketch. Write $S_{N}f(x_0)=(2\pi {)}^{-1}{\int }_{-\pi }^{\pi }f(x_0-t)D_N(t)dt$ using the convolution representation. Use the mean-value property $(2\pi {)}^{-1}\int D_N(t)dt=1$ (Proposition 3 analogue for $D_N$, not $F_N$) to write $$S_{N}f(x_0)-\frac{f(x_0^{+})+f(x_0^{-})}{2}=\frac{1}{2\pi }{\int }_{-\pi }^{\pi }g(t)\sin ((N+1/2)t)dt,$$ where $g(t)=[f(x_0-t)-\frac{f(x_0^{+})+f(x_0^{-})}{2}]/\sin (t/2)$ (with appropriate left/right interpretation near $t=0$). The piecewise $C^1$-hypothesis gives $g\in L^1([-\pi ,\pi ])$. By Riemann-Lebesgue applied to $g$, the integral $\to 0$ as $N\to \infty$, so $S_{N}f(x_0)\to (f(x_0^{+})+f(x_0^{-}))/2$.

$\square$

Connections Master

• L^p spaces, Hölder, Minkowski, and Riesz-Fischer completeness 02.07.06. The direct prerequisite for the $L^2$-theory of Fourier series. Riesz-Fischer completeness of $L^2$ gives the surjectivity of the Fourier-coefficient map $L^2(\mathbb{T})\to {\ell }^2(\mathbb{Z})$ — every ${\ell }^2$-sequence is the Fourier-coefficient sequence of some $L^2$-function. Without completeness, the inverse Fourier-series direction (assembling a function from a square-summable coefficient sequence) would lose its $L^2$-target.

• Fatou's lemma and dominated convergence theorem 02.07.05. Used in the proof of the Riemann-Lebesgue lemma via density of step functions in $L^1$, and in the limit-exchange arguments justifying interchange of summation and integration in Parseval-Plancherel identities. The DCT is also load-bearing in establishing $L^p$-convergence of Cesàro and Abel means.

• Lebesgue integral and monotone convergence 02.07.04. The foundational integration theory making Fourier coefficients well-defined for all $f\in L^1$, not just for Riemann-integrable functions. Lebesgue's 1903 Annales ENS paper introduced the Lebesgue-integral framework specifically to enable the Riemann-Lebesgue lemma in full generality, replacing Riemann's 1854 piecewise-Riemann-integrable hypothesis with the broader $L^1$-class.

• Inner product space 02.11.07. The abstract framework. $L^2([-\pi ,\pi ])$ with inner product $⟨f,g⟩=(2\pi {)}^{-1}{\int }_{-\pi }^{\pi }f\bar{g}dx$ is the prototype infinite-dimensional inner-product space, and the trigonometric exponentials form an explicit orthonormal basis. Bessel-Parseval identities and the projection-theoretic interpretation of Fourier partial sums are direct applications of the inner-product-space machinery.

• Fourier transform on R 02.10.04. The forward bridge. The Fourier series on $\mathbb{T}$ extends to the Fourier transform on $\mathbb{R}$ by the rescaling $L\to \infty$: a $2L$-periodic function with Fourier-coefficient sum becomes a function on the line with Fourier-transform integral. The Riemann-Lebesgue lemma generalises ($\hat{f}(\xi )\to 0$ as $|\xi |\to \infty$ for $f\in L^1(\mathbb{R})$), and Plancherel's identity is the $L^2$-analogue of Parseval. The Hausdorff-Young inequality interpolates between the $L^1\to L^{\infty }$ and $L^2\to L^2$ endpoints.

• Heat equation 02.13.03. The historical motivation. Fourier's 1822 Théorie analytique de la chaleur introduced Fourier series to solve the heat equation ${\partial }_{t}u=\kappa {\partial }_{xx}u$ on the bar $[0,L]$ with periodic boundary conditions: the sine-wave eigenfunctions $\sin (n\pi x/L)$ decay exponentially in time with rate $\kappa (n\pi /L{)}^2$, so the heat equation reduces to a system of decoupled ODEs in the Fourier-coefficient sequence. The wave equation, Schrödinger equation, and Laplace equation on rectangular domains admit the same Fourier-series approach.

• Functional analysis [02.11] and Hilbert space 02.11.08. The lateral framework. The Fourier-series machinery is a special case of the spectral decomposition of $L^2(\mathbb{T})$ under the translation operator $T_{a}f(x)=f(x-a)$ — the exponentials $e^{inx}$ are the joint eigenfunctions of the translation group, and the Fourier-coefficient map is the spectral decomposition. The same framework applies to the Laplacian on a torus, the Fourier-Hermite expansion on $\mathbb{R}$ (the quantum harmonic oscillator), and the spherical-harmonic expansion on $S^2$ (the angular Laplacian).

• Quantum mechanics — momentum representation. The lateral physics application. In quantum mechanics, the Fourier transform converts a position-space wavefunction $\psi (x)$ into the momentum-space wavefunction $\tilde{\psi }(p)$, with Parseval-Plancherel ensuring conservation of total probability. On a circular domain (a particle on a ring), the momentum operator has discrete eigenvalues $p_n=n\hslash$, and the wavefunction admits a Fourier-series decomposition in the position-momentum eigenbasis. The Riemann-Lebesgue lemma corresponds physically to the fact that smooth wavefunctions have well-localised momentum distributions.

Historical & philosophical context Master

Joseph Fourier presented his memoir on the heat equation to the Paris Academy of Sciences in 1807, claiming that every periodic function admits a decomposition into sine and cosine waves at integer-multiple frequencies. The 1807 manuscript was rejected by Lagrange, Laplace, Monge, and Lacroix — Lagrange's objection was specifically that the sum of continuous sine waves could not represent a discontinuous function (the square wave being the obvious counterexample to Fourier's claim). Fourier was undeterred and continued developing the theory; his 1811 prize-winning Academy submission led to the full publication of Théorie analytique de la chaleur in 1822 ^{[Fourier 1822]}. Fourier's intuitive approach lacked the modern apparatus of convergence and integration, and many of his arguments were heuristic; the rigorous foundations were supplied over the next century.

Dirichlet's 1829 Journal für die reine und angewandte Mathematik paper ^{[Dirichlet 1829]} was the first rigorous proof of any convergence theorem for Fourier series. Dirichlet introduced the convolution representation $S_{N}f(x_0)=(2\pi {)}^{-1}\int f(x_0-t)D_N(t)dt$ with what is now called the Dirichlet kernel, and proved pointwise convergence for piecewise continuous functions with piecewise continuous derivatives (now called the Dirichlet-Jordan class, after Jordan's 1881 generalisation to functions of bounded variation). Dirichlet's proof set the analytic standard for the next eighty years of Fourier-series convergence research.

Riemann's 1854 Habilitationsschrift on "the representability of a function by a trigonometric series" ^{[Riemann 1854]}, published posthumously in 1867, expanded Dirichlet's framework in two ways. First, Riemann introduced the Riemann integral specifically to enlarge the class of functions for which the Fourier-coefficient integrals are well-defined. Second, Riemann proved a precursor to the Riemann-Lebesgue lemma for Riemann-integrable functions: the coefficients of an integrable function vanish at infinity. The full statement for the broader $L^1$-class required Lebesgue's 1902 thesis and 1903 Annales ENS paper ^{[Lebesgue 1903]}, where the Lebesgue integral and dominated-convergence theorem allowed the density-of-step-functions argument that is the modern proof.

Du Bois-Reymond's 1876 Abhandlungen der Königlich Bayerischen Akademie der Wissenschaften paper ^{[Riemann 1854]} constructed the first explicit continuous function whose Fourier series diverges at a point. This was a shock to the contemporary mathematical community: it showed that Dirichlet's regularity hypothesis (piecewise $C^1$) could not be weakened to mere continuity, and that the connection between regularity of $f$ and convergence of $S_{N}f(x_0)$ was more subtle than expected. The 1876 construction was concrete; the modern Baire-category proof via the unboundedness of Lebesgue constants (Banach-Steinhaus 1927) is post-hoc.

Fejér's 1900 Comptes Rendus paper ^{[Fejér 1900]}, submitted by Fejér at the age of nineteen, rescued the classical theory after the Du Bois-Reymond counterexample. Fejér introduced the Cesàro means ${\sigma }_{N}f=(S_{0}f+\dots +S_{N}f)/(N+1)$ and proved uniform convergence of ${\sigma }_{N}f$ to $f$ for every continuous $f$. The proof used the non-negativity of the Fejér kernel, the key technical fact that distinguished $F_N$ from the sign-changing $D_N$. Fejér's theorem implies the density of trigonometric polynomials in $C(\mathbb{T})$ (a Weierstrass-style approximation result for the circle) and serves as the modern proof of the completeness of the trigonometric system in $L^2(\mathbb{T})$.

Kolmogorov's 1923 Comptes Rendus paper, sharpened in his 1926 Fundamenta Mathematicae paper, constructed an $L^1(\mathbb{T})$-function whose Fourier series diverges almost everywhere (in fact at every point in the 1926 sharpened version). This counterexample showed that $L^1$ is too large a class for any almost-everywhere convergence theorem, and pushed the "right" convergence class to $L^2$ or above. The corresponding positive almost-everywhere convergence theorem for $L^2$ — Luzin's 1915 conjecture — remained open for fifty years.

Carleson's 1966 Acta Mathematica paper ^{[Carleson 1966]} resolved Luzin's conjecture: every $L^2(\mathbb{T})$-function has Fourier partial sums converging almost everywhere. Carleson's proof was a tour de force of time-frequency analysis: he decomposed the partial-sum operator into wave packets indexed by tiles in phase space, then estimated the supremum of the wave-packet contributions against carefully constructed exceptional sets. Hunt's 1968 paper ^{[Hunt 1968]} extended Carleson's $L^2$-result to all $L^p$ with $p\in (1,\infty )$. The modern reformulation via the bilinear Hilbert transform (Lacey-Thiele 2000) clarified the structural mechanism: Carleson's argument is essentially a sharp bound on a model bilinear maximal operator, and the same time-frequency framework yields proofs of the bilinear Hilbert transform boundedness (the trilinear Calderón conjecture).

The Plancherel 1910 Rendiconti del Circolo Matematico di Palermo extension of Fourier-series theory to the Fourier transform on $\mathbb{R}$ unified the discrete and continuous spectral decompositions under a single framework: the Plancherel theorem on locally compact abelian groups (Pontryagin 1934, Cartan-Godement 1947). The unifying principle is that Fourier analysis on a group $G$ decomposes $L^2(G)$ as an integral of irreducible representations of $G$ — for $G=\mathbb{T}$ the irreducibles are the one-dimensional characters $e^{inx}$ indexed by $\mathbb{Z}$ (Fourier series); for $G=\mathbb{R}$ the irreducibles are $e^{i\xi x}$ indexed by $\xi \in \mathbb{R}$ (Fourier transform).

The structural arc is a two-hundred-year story: Fourier 1807-1822 (the original claim, with heuristic arguments), Dirichlet 1829 (first convergence proof, Dirichlet kernel), Riemann 1854 (Riemann integral, vanishing coefficients for Riemann-integrable functions), Du Bois-Reymond 1876 (divergence counterexample), Lebesgue 1903 (Lebesgue integral, Riemann-Lebesgue lemma in full generality), Fejér 1900 (Cesàro summability), Plancherel 1910 (Fourier transform on $\mathbb{R}$, $L^2$-isometry), Hilbert 1906 + Riesz-Fischer 1907 (Hilbert-space framework, completeness), Kolmogorov 1923 (negative $L^1$-result), Carleson 1966 + Hunt 1968 (almost-everywhere convergence for $L^p$, $p>1$). The pattern: each generation enlarged the function class, sharpened the convergence statement, or supplied a counterexample that fixed the threshold. The endpoint is the modern harmonic-analytic framework where Fourier series and Fourier transforms are special cases of the spectral theory on locally compact abelian groups, with the Carleson-Hunt theorem giving the sharpest known almost-everywhere convergence statement.

Bibliography Master

@book{Fourier1822,
  author    = {Fourier, Jean-Baptiste Joseph},
  title     = {Th\'eorie analytique de la chaleur},
  publisher = {Firmin Didot, P\`ere et Fils},
  address   = {Paris},
  year      = {1822}
}

@article{Dirichlet1829,
  author  = {Dirichlet, Peter Gustav Lejeune},
  title   = {Sur la convergence des s\'eries trigonom\'etriques qui servent \`a repr\'esenter une fonction arbitraire entre des limites donn\'ees},
  journal = {Journal f\"ur die reine und angewandte Mathematik},
  volume  = {4},
  year    = {1829},
  pages   = {157--169}
}

@article{Riemann1854,
  author  = {Riemann, Bernhard},
  title   = {\"Uber die {D}arstellbarkeit einer {F}unction durch eine trigonometrische {R}eihe},
  journal = {Abhandlungen der K\"oniglichen Gesellschaft der Wissenschaften zu G\"ottingen},
  volume  = {13},
  year    = {1867},
  note    = {Habilitationsschrift, G\"ottingen 1854; published posthumously}
}

@article{DuBoisReymond1876,
  author  = {Du Bois-Reymond, Paul},
  title   = {Untersuchungen \"uber die {C}onvergenz und {D}ivergenz der {F}ourierschen {D}arstellungsformeln},
  journal = {Abhandlungen der mathematisch-physikalischen Classe der K\"oniglich Bayerischen Akademie der Wissenschaften},
  volume  = {12},
  year    = {1876},
  pages   = {1--103}
}

@article{Fejer1900,
  author  = {Fej\'er, Lip\'ot},
  title   = {Sur les fonctions born\'ees et int\'egrables},
  journal = {Comptes Rendus de l'Acad\'emie des Sciences \`a Paris},
  volume  = {131},
  year    = {1900},
  pages   = {984--987}
}

@article{Lebesgue1903,
  author  = {Lebesgue, Henri},
  title   = {Sur les s\'eries trigonom\'etriques},
  journal = {Annales scientifiques de l'\'Ecole Normale Sup\'erieure},
  series  = {3},
  volume  = {20},
  year    = {1903},
  pages   = {453--485}
}

@article{Plancherel1910,
  author  = {Plancherel, Michel},
  title   = {Contribution \`a l'\'etude de la repr\'esentation d'une fonction arbitraire par des int\'egrales d\'efinies},
  journal = {Rendiconti del Circolo Matematico di Palermo},
  volume  = {30},
  year    = {1910},
  pages   = {289--335}
}

@article{Kolmogorov1923,
  author  = {Kolmogorov, Andrey Nikolaevich},
  title   = {Une s\'erie de {F}ourier-{L}ebesgue divergente presque partout},
  journal = {Comptes Rendus de l'Acad\'emie des Sciences \`a Paris},
  volume  = {178},
  year    = {1923},
  pages   = {303--306}
}

@article{Carleson1966,
  author  = {Carleson, Lennart},
  title   = {On convergence and growth of partial sums of {F}ourier series},
  journal = {Acta Mathematica},
  volume  = {116},
  year    = {1966},
  pages   = {135--157}
}

@inproceedings{Hunt1968,
  author    = {Hunt, Richard A.},
  title     = {On the convergence of {F}ourier series},
  booktitle = {Orthogonal Expansions and their Continuous Analogues (Edwardsville, Ill., 1967)},
  publisher = {Southern Illinois University Press},
  year      = {1968},
  pages     = {235--255}
}

@book{SteinShakarchi2003,
  author    = {Stein, Elias M. and Shakarchi, Rami},
  title     = {Fourier Analysis: An Introduction},
  series    = {Princeton Lectures in Analysis, Volume {I}},
  publisher = {Princeton University Press},
  year      = {2003}
}

@book{Katznelson2004,
  author    = {Katznelson, Yitzhak},
  title     = {An Introduction to Harmonic Analysis},
  edition   = {3},
  publisher = {Cambridge University Press},
  year      = {2004}
}

@book{Zygmund2002,
  author    = {Zygmund, Antoni},
  title     = {Trigonometric Series, Volumes {I-II}},
  edition   = {3},
  publisher = {Cambridge University Press},
  year      = {2002}
}

@article{LaceyThiele2000,
  author  = {Lacey, Michael and Thiele, Christoph},
  title   = {A proof of boundedness of the {C}arleson operator},
  journal = {Mathematical Research Letters},
  volume  = {7},
  year    = {2000},
  pages   = {361--370}
}

@article{Riesz1927,
  author  = {Riesz, Marcel},
  title   = {Sur les fonctions conjugu\'ees},
  journal = {Mathematische Zeitschrift},
  volume  = {27},
  year    = {1927},
  pages   = {218--244}
}

@article{HardyLittlewood1928,
  author  = {Hardy, G. H. and Littlewood, J. E.},
  title   = {A maximal theorem with function-theoretic applications},
  journal = {Acta Mathematica},
  volume  = {54},
  year    = {1930},
  pages   = {81--116}
}

@article{Bessel1828,
  author  = {Bessel, Friedrich Wilhelm},
  title   = {\"Uber die {B}estimmung des {G}esetzes einer periodischen {E}rscheinung},
  journal = {Astronomische Nachrichten},
  volume  = {6},
  year    = {1828},
  pages   = {333--348}
}