Fourier Series — Advanced Treatment

Fourier series represent a periodic function

f

as an infinite sum of sines and cosines (equivalently, complex exponentials). Key result: under mild conditions,

f (x) = \sum_{n} c_{n} e^{in x}

where the coefficients

c_{n}

are integrals of

f

against

e^{- in x}

, and convergence holds in the

L^{2}

sense (Parseval) and pointwise where

f

is nice. Applies to signal processing, PDEs (heat/wave), harmonic analysis, number theory. Originated with Joseph Fourier (1807–1822) studying heat conduction; rigorous foundations by Dirichlet, Riemann, Lebesgue, Fejér, Carleson.

A Fourier series is the expansion of a $2 π$ -periodic (or $L$ -periodic) function as a superposition of harmonics $e^{in x}$ , with coefficients determined by orthogonality. It matters because it diagonalizes translation-invariant linear operators (differentiation, convolution), turning differential equations into algebra. The deepest surprise is that the naive expectation — "smooth function $\Rightarrow$ pointwise convergence everywhere" — is both more delicate and more robust than it looks: there exist continuous functions whose Fourier series diverge at a point (du Bois-Reymond, 1873), yet every $L^{2}$ function's series converges almost everywhere (Carleson, 1966). The theory sits at the crossroads of analysis, geometry (Hilbert spaces), and physics.

The Core Insight

The complex exponentials ${e_{n} (x) = e^{in x}}_{n \in Z}$ form a complete orthonormal basis of the Hilbert space $L^{2} [- π, π]$ (normalized by $\frac{1}{2 π} \int$ ). Expanding $f$ in this basis is exactly projecting a vector onto orthonormal axes — except the space is infinite-dimensional. Everything else (decay of coefficients $\leftrightarrow$ smoothness, Parseval $\leftrightarrow$ Pythagoras, convolution $\leftrightarrow$ multiplication) follows from this single geometric fact.

Formal Statement

Fourier Coefficients and Series

Let $f : R \to C$ be $2 π$ -periodic and integrable on $[- π, π]$ . Its complex Fourier coefficients are

\hat{f} (n) = c_{n} = \frac{1}{2 π} \int_{- π}^{π} f (x) e^{- in x} d x, n \in Z .

⟳project

f

onto

e_{n}

using the inner product

⟨ f, g ⟩ = \frac{1}{2 π} \int_{- π}^{π} f \overset{g}{ˉ} d x

; orthonormality

⟨ e_{m}, e_{n} ⟩ = δ_{mn}

isolates the

n

-th coefficient.

The Fourier series of $f$ is the formal sum

f (x) \sim n = - \infty \sum \infty c_{n} e^{in x} .

⟳the symbol

\sim

("is associated with") avoids asserting convergence until proven; it becomes

=

under conditions specified below.

The real form uses $a_{n}, b_{n}$ :

f (x) \sim \frac{a _{0}}{2} + n = 1 \sum \infty (a_{n} cos n x + b_{n} sin n x),

a_{n} = \frac{1}{π} \int_{- π}^{π} f (x) cos n x d x, b_{n} = \frac{1}{π} \int_{- π}^{π} f (x) sin n x d x .

⟳substitute

e^{\pm in x} = cos n x \pm i sin n x

into [2]; collect terms to get

a_{n} = c_{n} + c_{- n}

b_{n} = i (c_{n} - c_{- n})

Symbols:

c_{n}, \hat{f} (n)

= complex coefficient;

a_{n}, b_{n}

= real cosine/sine coefficients;

n

= harmonic index (frequency);

e^{in x}

= the

n

-th harmonic of angular frequency

n

Partial Sums and Dirichlet Kernel

The $N$ -th partial sum and its integral representation are

S_{N} f (x) = n = - N \sum N c_{n} e^{in x} = \frac{1}{2 π} \int_{- π}^{π} f (x - t) D_{N} (t) d t,

where the Dirichlet kernel is

D_{N} (t) = n = - N \sum N e^{in t} = \frac{sin ( ( N + \frac{1}{2} ) t )}{sin ( t /2 )} .

⟳insert [1] into the sum [5], swap sum and integral, recognize

\sum e^{in (x - t)} = D_{N} (x - t)

; the closed form is a geometric series

\sum_{- N}^{N} r^{n}

with

r = e^{i t}

Parseval / Plancherel Identity

For $f \in L^{2} [- π, π]$ ,

\frac{1}{2 π} \int_{- π}^{π} ∣ f (x) ∣^{2} d x = n = - \infty \sum \infty ∣ c_{n} ∣^{2} .

⟳this is the Pythagorean theorem in

L^{2}

; completeness of

{e_{n}}

upgrades Bessel's inequality

\sum ∣ c_{n} ∣^{2} \leq ∥ f ∥^{2}

to equality.

In particular the map $f \mapsto (c_{n})$ is an isometry $L^{2} [- π, π] \to ℓ^{2} (Z)$ , and $S_{N} f \to f$ in $L^{2}$ -norm for every square-integrable $f$ .

Dirichlet Pointwise Convergence

If $f$ is $2 π$ -periodic, piecewise continuous, and has left/right derivatives at $x_{0}$ , then

N \to \infty lim S_{N} f (x_{0}) = \frac{f ( x _{0}^{+} ) + f ( x _{0}^{-} )}{2} .

⟳write

S_{N} f (x_{0}) - \frac{f ( x _{0}^{+} ) + f ( x _{0}^{-} )}{2}

as an integral against

D_{N}

; split into

[- δ, δ]

and the rest; apply the Riemann–Lebesgue lemma after dividing by

sin (t /2)

(the local-integrability "Dini condition" makes the quotient integrable).

At points of continuity this gives $S_{N} f (x_{0}) \to f (x_{0})$ .

Why It Holds / Derivation

Orthogonality and the Coefficient Formula

Step 1 — Inner product and orthonormality. Define

⟨ f, g ⟩ = \frac{1}{2 π} \int_{- π}^{π} f \overset{g}{ˉ} d x

. Compute for

m \neq = n

⟨ e_{m}, e_{n} ⟩ = \frac{1}{2 π} \int_{- π}^{π} e^{i (m - n) x} d x = \frac{1}{2 π} \cdot \frac{e ^{i (m - n) x}}{i ( m - n )}_{- π}^{π} = 0,

since

e^{i (m - n) π} = e^{- i (m - n) π}

. For

m = n

the integrand is

1

, giving

⟨ e_{n}, e_{n} ⟩ = 1

. Thus

⟨ e_{m}, e_{n} ⟩ = δ_{mn}

. Step 2 — Extract coefficients. Assuming

f = \sum_{k} c_{k} e_{k}

converges nicely, take inner product with

e_{n}

⟨ f, e_{n} ⟩ = k \sum c_{k} ⟨ e_{k}, e_{n} ⟩ = c_{n} .

Expanding

⟨ f, e_{n} ⟩ = \frac{1}{2 π} \int f e^{- in x} d x

recovers [1]. Step 3 — Bessel's inequality. For any

N

, the projection

S_{N} f

is the closest point in

span {e_{- N}, \dots, e_{N}}

f

. Since

f - S_{N} f ⊥ e_{n}

for

∣ n ∣ \leq N

∥ f ∥^{2} = ∥ S_{N} f ∥^{2} + ∥ f - S_{N} f ∥^{2} \geq ∥ S_{N} f ∥^{2} = ∣ n ∣ \leq N \sum ∣ c_{n} ∣^{2} .

Letting

N \to \infty

\sum ∣ c_{n} ∣^{2} \leq ∥ f ∥^{2} < \infty

. Hence

c_{n} \to 0

(Riemann–Lebesgue). Step 4 — Completeness $\Rightarrow$ Parseval. The key analytic input is that trigonometric polynomials are dense in

L^{2}

(proved via Fejér's theorem, below). Density forces

∥ f - S_{N} f ∥ \to 0

, so Bessel becomes equality [7].

■

Fejér's Theorem — Cesàro Sums Always Work

The Dirichlet kernel is not positive and not an approximate identity, which is why $S_{N} f$ can misbehave. The fix: average.

Step 1 — Cesàro means. Define

σ_{N} f = \frac{1}{N + 1} \sum_{k = 0}^{N} S_{k} f = \frac{1}{2 π} \int_{- π}^{π} f (x - t) K_{N} (t) d t

, where the Fejér kernel is

K_{N} (t) = \frac{1}{N + 1} k = 0 \sum N D_{k} (t) = \frac{1}{N + 1} (\frac{sin (( N + 1 ) t /2 )}{sin ( t /2 )})^{2} \geq 0.

⟳sum the Dirichlet closed forms using the telescoping identity

\sum_{k} sin ((k + \frac{1}{2}) t) = \frac{s i n ^{2} (( N + 1 ) t /2 )}{s i n ( t /2 )}

; positivity is the crucial gain over

D_{N}

Step 2 — Approximate identity properties.

K_{N} \geq 0

\frac{1}{2 π} \int K_{N} = 1

, and

K_{N}

concentrates at

0

. For continuous

f

σ_{N} f \to f

uniformly. Step 3 — Consequence. Trigonometric polynomials (the

σ_{N} f

) approximate every continuous function uniformly, hence are dense in

L^{2}

, completing Step 4 of the previous proof.

■

Joseph Fourier | 1768–1830

A French mathematician and physicist who accompanied Napoleon to Egypt and later served as prefect of Isère. In his 1807 memoir and 1822 treatise Théorie analytique de la chaleur, Fourier made the audacious claim that any function — even discontinuous ones with corners — could be written as a sum of sines and cosines, in order to solve the heat equation. His contemporaries (Lagrange, Laplace) objected strenuously; the claim was too broad and his arguments non-rigorous. The 19th century vindicated his physical instinct while sharpening exactly which functions and what kind of convergence. Fourier also discovered the greenhouse effect.

Building Intuition

Think of a periodic signal as a musical chord. The Fourier coefficients are the recipe: how much of each pure tone (pitch $n$ ) is present, and with what phase. Three guiding pictures:

Geometry: $L^{2}$ is an infinite-dimensional space with perpendicular axes $e_{n}$ . Computing $c_{n}$ is dropping a perpendicular onto axis $n$ ; Parseval is "length² = sum of squared components."
Smoothness $\leftrightarrow$ decay: A smooth function is built mostly from low frequencies, so $c_{n}$ decays fast. Each derivative multiplies $c_{n}$ by $in$ , so $f \in C^{k} \Rightarrow c_{n} = o (n^{- k})$ . Conversely, jumps force slow $\sim 1/ n$ decay.
Why averaging (Fejér) helps: $S_{N}$ can overshoot near jumps (Gibbs) because $D_{N}$ has large negative lobes that "ring." The Fejér kernel $K_{N}$ is a positive bump — a blunt, honest averaging instrument — so it never overshoots and always converges for continuous data.

Dirichlet kernel $D_5$ vs. Fejér kernel $K_5$: the negative lobes of $D_N$ cause Gibbs ringing; $K_N\ge 0$ does not.

Worked Examples

Square Wave — Coefficients, Sum, and Gibbs

Given: the odd

2 π

-periodic square wave

f (x) = sgn (sin x)

, i.e.

f (x) = + 1

(0, π)

- 1

(- π, 0)

. Step 1 — Symmetry.

f

is odd

\Rightarrow a_{n} = 0

for all

n

, and we need only

b_{n}

. Step 2 — Compute $b_{n}$ .

b_{n} = \frac{1}{π} \int_{- π}^{π} f sin n x d x = \frac{2}{π} \int_{0}^{π} sin n x d x = \frac{2}{π} \cdot \frac{1 - cos nπ}{n} = \frac{2}{π} \cdot \frac{1 - ( - 1 ) ^{n}}{n} .

b_{n} = 0

for even

n

, and

b_{n} = \frac{4}{π n}

for odd

n

. Step 3 — Series.

f (x) = \frac{4}{π} k = 0 \sum \infty \frac{sin ( ( 2 k + 1 ) x )}{2 k + 1} = \frac{4}{π} (sin x + \frac{1}{3} sin 3 x + \frac{1}{5} sin 5 x + \dots)

Step 4 — Check at $x = π /2$ : the series gives

\frac{4}{π} (1 - \frac{1}{3} + \frac{1}{5} - \dots) = \frac{4}{π} \cdot \frac{π}{4} = 1 = f (π /2) .

✓ (Leibniz series for

π /4

.) Step 5 — Parseval check:

\frac{1}{2 π} \int ∣ f ∣^{2} = 1

and

\sum ∣ c_{n} ∣^{2} = \frac{1}{2} \sum b_{n}^{2} = \frac{1}{2} \cdot \frac{16}{π ^{2}} \sum_{odd} \frac{1}{n ^{2}} = \frac{8}{π ^{2}} \cdot \frac{π ^{2}}{8} = 1.

✓ Gibbs: near the jump,

S_{N}

overshoots

f

\approx 8.95%

of the jump height (a fixed

\approx 0.0895 \times 2

) regardless of

N

Parabola — Solving for $\zeta(2)$

Given:

f (x) = x^{2}

[- π, π]

, extended

2 π

-periodically (even function). Step 1 — $a_{0}$ :

a_{0} = \frac{1}{π} \int_{- π}^{π} x^{2} d x = \frac{1}{π} \cdot \frac{2 π ^{3}}{3} = \frac{2 π ^{2}}{3} .

Step 2 — $a_{n}$ ( $n \geq 1$ ): integrate by parts twice,

a_{n} = \frac{2}{π} \int_{0}^{π} x^{2} cos n x d x = \frac{2}{π} \cdot \frac{2 π ( - 1 ) ^{n}}{n ^{2}} = \frac{4 ( - 1 ) ^{n}}{n ^{2}} .

(

b_{n} = 0

by evenness.) Step 3 — Series:

x^{2} = \frac{π ^{2}}{3} + 4 n = 1 \sum \infty \frac{( - 1 ) ^{n}}{n ^{2}} cos n x

Step 4 — Evaluate at $x = π$ (a continuity point of the periodic extension, value

π^{2}

π^{2} = \frac{π ^{2}}{3} + 4 n = 1 \sum \infty \frac{( - 1 ) ^{n} ( - 1 ) ^{n}}{n ^{2}} = \frac{π ^{2}}{3} + 4 n = 1 \sum \infty \frac{1}{n ^{2}} .

Solving:

\sum_{n = 1}^{\infty} \frac{1}{n ^{2}} = \frac{π ^{2}}{6}

— the Basel problem, units dimensionless. ✓ Step 5 — Evaluate at $x = 0$ :

0 = \frac{π ^{2}}{3} + 4 \sum \frac{( - 1 ) ^{n}}{n ^{2}}

gives

\sum \frac{( - 1 ) ^{n - 1}}{n ^{2}} = \frac{π ^{2}}{12}

Common Misconceptions

"Fourier series of a continuous function converges to it everywhere." False. Du Bois-Reymond (1873) constructed a continuous periodic function whose Fourier series diverges at a point. Continuity alone is not enough for pointwise convergence of

S_{N}

. What is guaranteed: (a) Cesàro sums

σ_{N} f

converge uniformly (Fejér); (b)

S_{N} f \to f

L^{2}

; (c)

S_{N} f \to f

almost everywhere (Carleson). Pointwise everywhere needs extra regularity (Dini/Dirichlet conditions or bounded variation). "You can differentiate a Fourier series term by term." Not always. Term-by-term differentiation multiplies

c_{n}

in

, which may destroy convergence. The square wave's series converges, but differentiating gives

\frac{4}{π} \sum cos ((2 k + 1) x)

which diverges (its "sum" is a Dirac comb, a distribution). Term-by-term integration is much safer — it divides by

n

and improves convergence.

A continuous function whose series diverges at 0

One classical construction (lacunary, Fejér–Lebesgue style): build

g (x) = k = 1 \sum \infty \frac{1}{k ^{2}} T_{n_{k}} (x), T_{n} (x) = j = 1 \sum n \frac{sin j x}{j} \cdot (suitable shift),

with rapidly increasing

n_{k}

. Each block

T_{n_{k}}

is a trigonometric polynomial whose partial sums grow like

lo g n_{k}

x = 0

. Choosing

n_{k}

so that

\frac{l o g n _{k}}{k ^{2}} \to \infty

along a subsequence makes

S_{N} g (0)

unbounded, while the

\frac{1}{k ^{2}}

weights keep

g

uniformly convergent, hence continuous. Moral: uniform convergence of the function's defining series does not control the Dirichlet partial sums.

Connections

Fourier series is the periodic/discrete-spectrum member of a family:

Fourier Transform — the non-periodic limit ( $L \to \infty$ ): sums over $n$ become integrals over continuous frequency $ξ$ .
Hilbert Space — Parseval is the abstract Pythagorean theorem; ${e_{n}}$ is the canonical complete orthonormal basis, prototype for general orthonormal expansions.
Sturm–Liouville Theory — sines/cosines are eigenfunctions of $- d^{2} / d x^{2}$ with periodic boundary conditions; general SL problems give other orthogonal bases (Legendre, Bessel).
Convolution Theorem — convolution on the circle becomes pointwise multiplication of coefficients: $f * g (n) = \hat{f} (n) \overset{g}{^} (n)$ .
Distribution Theory — the Dirac comb and termwise-differentiated series live here; $\sum e^{in x} = 2 π \sum_{k} δ (x - 2 π k)$ .
Discrete Fourier Transform — sampling Fourier series at $N$ points gives the DFT, computed by the FFT.

Real-World Applications

Audio & music (MP3, AAC): signals are decomposed into frequency components; perceptual coding discards inaudible coefficients. Equalizers directly scale $∣ c_{n} ∣$ .
Image compression (JPEG): uses the discrete cosine transform (a real Fourier cousin) on $8 \times 8$ blocks; quantizing high-frequency coefficients yields huge compression.
Solving PDEs (heat & wave equations): separation of variables expands initial data in Fourier modes; each mode evolves independently ( $e^{- n^{2} t}$ for heat). This is exactly Fourier's original 1822 application.
Electrical engineering / AC circuits: any periodic voltage is a sum of harmonics; impedance analysis treats each harmonic separately (phasors).
Crystallography (X-ray diffraction): the diffraction pattern is the Fourier transform of the electron density; the periodic lattice gives a Fourier-series structure (structure factors $F_{hk l}$ ).
Vibration & structural analysis: modal decomposition of vibrating strings, membranes, and beams uses Fourier/eigenfunction series.
Number theory: Fourier expansions of periodic arithmetic functions and theta functions underlie analytic number theory and the functional equation of $ζ$ .

Historical Development

1748 — Euler & d'Alembert: the vibrating string controversy; trigonometric sums appear but their generality is fiercely disputed.
1753 — Daniel Bernoulli: claims all string motions are superpositions of harmonics; Euler and d'Alembert reject the universality of the claim.
1807 — Fourier: submits his memoir on heat to the French Academy; the committee (Lagrange, Laplace) is skeptical of the convergence claims.
1822 — Fourier: publishes Théorie analytique de la chaleur, the foundational text.
1829 — Dirichlet: gives the first rigorous convergence theorem (Dirichlet conditions), founding modern analysis of the question.
1854 — Riemann: his Habilitationsschrift on trigonometric series introduces the Riemann integral to handle Fourier coefficients rigorously.
1873 — du Bois-Reymond: continuous function with divergent Fourier series — convergence is genuinely subtle.
1900–1906 — Fejér & Lebesgue: Fejér's Cesàro-summability theorem (1900) and Lebesgue integration (1902) place the theory on firm modern footing; Bessel/Parseval completed.
1966 — Carleson: proves $L^{2}$ functions converge a.e. (extended to $L^{p}$ , $p > 1$ , by Hunt 1968), resolving the Lusin conjecture.

Subtleties & Limits

Convergence at the boundary of $L^{1}$ . Carleson's theorem fails for

L^{1}

: Kolmogorov (1923) built an

L^{1}

function whose Fourier series diverges everywhere. The precise frontier — exactly which integrability/regularity conditions force a.e. convergence — involves subtle maximal-function estimates (the Carleson operator) whose sharp constants and multidimensional analogues remain active research. In several variables, the order of summation matters enormously: spherical vs. rectangular partial sums behave very differently, and sharp convergence results for Bochner–Riesz means in dimension

\geq 2

are connected to deep open problems in harmonic analysis (the Bochner–Riesz and restriction conjectures), which intertwine with the Kakeya problem and remain unsolved.