Fourier Analysis

⚡ Key Takeaways

Fourier Analysis decomposes functions into sums or integrals of sines and cosines (or complex exponentials). Key result: any "reasonable" periodic function equals a sum of harmonics whose amplitudes are computed by integration; non-periodic functions get a continuous analogue, the Fourier transform. Applies to: signal processing, PDEs, quantum mechanics, number theory, data compression. Who/when: Joseph Fourier (1807–1822), formalized by Dirichlet, Riemann, Lebesgue, and others through the 19th–20th centuries.

💡 The Core Insight

The sines and cosines form an orthogonal basis for a space of functions. Just as a vector in $\mathbb{R}^3$ is reconstructed from its components along orthogonal axes, a function is reconstructed from its "components" along each frequency. Each Fourier coefficient is an inner product — a projection of the function onto a single pure tone. The entire theory is linear algebra in an infinite-dimensional inner-product space.

DefinitionFourier Series — Trigonometric Form

Let $f$ be a real function of period $2L$ that is (absolutely) integrable over one period. Its Fourier series is

1

\quad f(x) \sim \frac{a_0}{2} + \sum_{n=1}^{\infty}\left[a_n\cos\!\frac{n\pi x}{L} + b_n\sin\!\frac{n\pi x}{L}\right]

Derivation hint: Posit that $f$ is a superposition of harmonics with period dividing $2L$ ; the constant term is written $a_0/2$ so that the coefficient formula below is uniform across all $n$ . with coefficients

2

\quad a_n = \frac{1}{L}\int_{-L}^{L} f(x)\cos\!\frac{n\pi x}{L}\,dx,\qquad b_n = \frac{1}{L}\int_{-L}^{L} f(x)\sin\!\frac{n\pi x}{L}\,dx

Derivation hint: Multiply [1] by $\cos(m\pi x/L)$ or $\sin(m\pi x/L)$ , integrate over a period, and use the orthogonality relations $\int_{-L}^L \cos\frac{n\pi x}{L}\cos\frac{m\pi x}{L}\,dx = L\delta_{nm}$ (for $n,m\ge1$ ), with sines analogous and mixed products vanishing. Here:

$f$ — the function being analyzed;
$L$ — half-period (period is $2L$ );
$a_n,b_n$ — Fourier coefficients (amplitudes of each harmonic);
$n$ — harmonic number (integer $\ge 0$ );
the symbol $\sim$ means "has the Fourier series," reserved because equality requires convergence hypotheses.

DefinitionFourier Series — Complex Form

Using Euler's formula $e^{i\theta}=\cos\theta+i\sin\theta$ , the series collapses to

3

\quad f(x) \sim \sum_{n=-\infty}^{\infty} c_n\, e^{i n\pi x/L},\qquad c_n = \frac{1}{2L}\int_{-L}^{L} f(x)\, e^{-i n\pi x/L}\,dx

Derivation hint: Substitute $\cos\theta = \tfrac12(e^{i\theta}+e^{-i\theta})$ , $\sin\theta = \tfrac1{2i}(e^{i\theta}-e^{-i\theta})$ into [1]; matching coefficients gives $c_n=\tfrac12(a_n-ib_n)$ for $n>0$ , $c_{-n}=\overline{c_n}$ for real $f$ , and $c_0=a_0/2$ . The exponentials $\{e^{in\pi x/L}\}_{n\in\mathbb{Z}}$ satisfy the orthogonality relation $\int_{-L}^{L} e^{in\pi x/L}\,\overline{e^{im\pi x/L}}\,dx = 2L\,\delta_{nm}$ .

DefinitionFourier Transform

For a function $f:\mathbb{R}\to\mathbb{C}$ that is absolutely integrable ( $f\in L^1(\mathbb{R})$ ), the Fourier transform and its inverse are

4

\quad \hat f(\xi) = \int_{-\infty}^{\infty} f(x)\, e^{-2\pi i \xi x}\,dx,\qquad f(x) = \int_{-\infty}^{\infty} \hat f(\xi)\, e^{2\pi i \xi x}\,d\xi

Derivation hint: Take the period $2L\to\infty$ in [3]; the discrete frequencies $n/(2L)$ merge into a continuous variable $\xi$ , the sum becomes a Riemann integral, and the factor $1/(2L)$ becomes the differential $d\xi$ . Here $\xi$ is frequency. (Conventions vary: some authors use angular frequency $\omega=2\pi\xi$ and place factors of $1/\sqrt{2\pi}$ or $1/(2\pi)$ ; the $e^{-2\pi i\xi x}$ convention keeps the transform pair perfectly symmetric.)

TheoremDirichlet's Pointwise Convergence Theorem

If $f$ is periodic, piecewise continuous, and has a left and right derivative at a point $x_0$ (e.g. $f$ is piecewise smooth), then the Fourier series converges at $x_0$ to the average of the one-sided limits:

5

\quad \frac{a_0}{2}+\sum_{n=1}^\infty\left[a_n\cos\tfrac{n\pi x_0}{L}+b_n\sin\tfrac{n\pi x_0}{L}\right] = \frac{f(x_0^+)+f(x_0^-)}{2}

Derivation hint: Express the $N$ -th partial sum as the convolution of $f$ with the Dirichlet kernel $D_N(t)=\frac{\sin((N+\tfrac12)t)}{\sin(t/2)}$ , then apply the Riemann–Lebesgue lemma to the difference, which kills all but the value at the jump's midpoint. At points of continuity the right side equals $f(x_0)$ .

TheoremPlancherel / Parseval

The Fourier transform preserves the $L^2$ norm (energy):

6

\quad \int_{-\infty}^{\infty}|f(x)|^2\,dx = \int_{-\infty}^{\infty}|\hat f(\xi)|^2\,d\xi

and for Fourier series, $\displaystyle \frac{1}{2L}\int_{-L}^{L}|f|^2\,dx = \sum_{n=-\infty}^{\infty}|c_n|^2$ . Derivation hint: Expand $\int |f|^2 = \int f\bar f$ , substitute the inverse transform for one factor, swap order of integration, and use $\int e^{2\pi i(\xi-\xi')x}dx = \delta(\xi-\xi')$ .

ProofDeriving the Coefficient Formulas via Orthogonality

1. Assume a representation. Suppose $f(x)=\frac{a_0}{2}+\sum_{n\ge1}[a_n\cos\frac{n\pi x}{L}+b_n\sin\frac{n\pi x}{L}]$ with convergence good enough to integrate term by term. 2. Establish orthogonality. For integers $m,n\ge 1$ :

7

\quad \int_{-L}^{L}\cos\tfrac{n\pi x}{L}\cos\tfrac{m\pi x}{L}\,dx = L\,\delta_{nm},\quad \int_{-L}^{L}\sin\tfrac{n\pi x}{L}\sin\tfrac{m\pi x}{L}\,dx = L\,\delta_{nm}

Derivation hint: Use product-to-sum identities $\cos A\cos B=\tfrac12[\cos(A-B)+\cos(A+B)]$ ; the integral of any nonzero-frequency cosine over a full period is zero, leaving only the $n=m$ term. 3. Mixed and constant terms. $\int_{-L}^{L}\cos\frac{n\pi x}{L}\sin\frac{m\pi x}{L}\,dx=0$ for all $n,m$ (odd integrand symmetry), and $\int_{-L}^{L}1\cdot\cos\frac{n\pi x}{L}\,dx=0$ for $n\ge1$ . 4. Project. To extract $a_m$ ( $m\ge1$ ), multiply both sides by $\cos\frac{m\pi x}{L}$ and integrate over $[-L,L]$ . By steps 2–3 every term vanishes except $a_m\int\cos^2 = a_m L$ , giving $a_m=\frac1L\int_{-L}^{L}f\cos\frac{m\pi x}{L}\,dx$ , matching [2]. 5. The constant. Integrating [1] directly gives $\int_{-L}^L f\,dx = a_0 L$ , i.e. $a_0=\frac1L\int_{-L}^L f\,dx$ — which is exactly [2] with $n=0$ , justifying the $a_0/2$ convention. 6. Best approximation property. Among all trigonometric polynomials of degree $\le N$ , these coefficients minimize the mean-square error $\int_{-L}^L|f-P_N|^2$ . This is the projection theorem in the Hilbert space $L^2[-L,L]$ : the Fourier partial sum is the orthogonal projection of $f$ onto the span of the first $2N+1$ harmonics. $\blacksquare$

∎

JF

Joseph Fourier

1768–1830

Jean-Baptiste Joseph Fourier was a French mathematician and physicist, orphaned young, who trained at the École Normale, accompanied Napoleon to Egypt, and later served as prefect of Isère. While studying heat conduction he made the audacious claim (in his 1807 memoir, expanded in Théorie analytique de la chaleur, 1822) that any function — even one with corners or jumps — could be written as a trigonometric series. The leading mathematicians of the day, including Lagrange and Laplace, were skeptical, and rigorous justification took another century. Fourier is also credited with an early recognition of the greenhouse effect.

🧠 Intuition

Think of a function as a musical chord and Fourier analysis as the ear (or a prism). The chord is a single complicated pressure wave; your ear decomposes it into pure tones with definite pitches and loudnesses. The Fourier coefficient $c_n$ is "how much of pitch $n$ " is present and "with what phase." Two complementary viewpoints:

Geometric: the harmonics are perpendicular coordinate axes in an infinite-dimensional space. $c_n$ is the shadow of $f$ on the $n$ -th axis. Parseval [6] is the Pythagorean theorem: total energy = sum of squared components.
Dynamical: $e^{in\omega t}$ are the eigenfunctions of $\frac{d}{dt}$ . Decomposing $f$ into them diagonalizes differentiation, which is why Fourier methods crush linear differential equations: each frequency evolves independently.

Figure

Square wave (gold) and its Fourier partial sums — note the Gibbs overshoot near the jumps

ExampleFourier Series of a Square Wave

Given: the odd square wave of period $2\pi$ ( $L=\pi$ ): $f(x)=+1$ for $0<x<\pi$ , $f(x)=-1$ for $-\pi<x<0$ .

Steps: 1. Since $f$ is odd, all $a_n=0$ (cosines are even). Only $b_n$ survive. 2. Compute $b_n=\frac1\pi\int_{-\pi}^{\pi}f(x)\sin(nx)\,dx = \frac2\pi\int_0^\pi \sin(nx)\,dx$ (even integrand). 3. $\int_0^\pi \sin(nx)\,dx = \frac{1-\cos(n\pi)}{n}=\frac{1-(-1)^n}{n}$ . 4. So $b_n=\frac{2}{\pi}\cdot\frac{1-(-1)^n}{n}$ : this is $0$ for even $n$ and $\frac{4}{\pi n}$ for odd $n$ .

Boxed answer:

8

\quad f(x)=\frac{4}{\pi}\sum_{k=0}^{\infty}\frac{\sin\big((2k+1)x\big)}{2k+1}=\frac{4}{\pi}\left(\sin x+\frac{\sin 3x}{3}+\frac{\sin 5x}{5}+\cdots\right)

Derivation hint: only odd harmonics appear because the half-period symmetry $f(x+\pi)=-f(x)$ annihilates even harmonics.

Bonus: setting $x=\pi/2$ (where $f=1$ ) gives the Leibniz series $\frac{\pi}{4}=1-\frac13+\frac15-\frac17+\cdots$ .

ExampleFourier Transform of a Gaussian

Given: $f(x)=e^{-\pi x^2}$ (a normalized Gaussian).

Steps: 1. $\hat f(\xi)=\int_{-\infty}^\infty e^{-\pi x^2}e^{-2\pi i\xi x}\,dx$ . 2. Complete the square in the exponent: $-\pi x^2-2\pi i\xi x = -\pi(x+i\xi)^2 - \pi\xi^2$ . 3. Factor out $e^{-\pi\xi^2}$ : $\hat f(\xi)=e^{-\pi\xi^2}\int_{-\infty}^\infty e^{-\pi(x+i\xi)^2}\,dx$ . 4. The remaining integral equals $\int_{-\infty}^\infty e^{-\pi u^2}\,du = 1$ (contour shift / Gaussian integral).

Boxed answer:

9

\quad \widehat{e^{-\pi x^2}}(\xi)=e^{-\pi \xi^2}

Derivation hint: the Gaussian with this normalization is a fixed point ("eigenfunction") of the Fourier transform.

Interpretation: A Gaussian transforms to a Gaussian — and a narrow one transforms to a wide one (try $e^{-\pi a x^2}\mapsto a^{-1/2}e^{-\pi\xi^2/a}$ ). This trade-off is the analytic heart of the Uncertainty Principle.

⚠️ Common mistake

"A Fourier series always equals the function." Not pointwise. Convergence depends on the function's smoothness and the mode of convergence. There exist continuous functions (du Bois-Reymond, 1873) whose Fourier series diverges at a point. Convergence in $L^2$ (mean-square) is guaranteed for any square-integrable function (Riesz–Fischer), but pointwise convergence needs extra hypotheses like piecewise smoothness. The deep result of Carleson (1966) is that for $f\in L^2$ the series converges almost everywhere — a far harder fact than it sounds.

⚠️ Common mistake

"More terms always give a better approximation everywhere." Near a jump discontinuity the partial sums overshoot by a fixed fraction (~8.95% of the jump) regardless of how many terms you take — the Gibbs phenomenon. The overshoot does not shrink; it merely narrows and slides toward the discontinuity. The series still converges in the $L^2$ sense, just not uniformly.

Counterexample

Define $f$ on $[-\pi,\pi]$ by $f(x)=\sum_{k=1}^\infty \frac{\sin(2^{k}x)}{k^2}$ -type lacunary constructions, or more simply consider the Weierstrass-style nowhere-differentiable function. Its Fourier coefficients decay like $1/k^2$ along sparse frequencies, the series converges uniformly (so $f$ is continuous), yet $f$ is nowhere differentiable. Conclusion: a uniformly convergent, perfectly continuous Fourier series can represent a function with no derivative anywhere — slow decay of coefficients ⟺ rough function.

connection

Fourier analysis sits at a crossroads:

It is infinite-dimensional <span class="wiki-link-missing">Linear Algebra</span>: coefficients are inner products onto an orthonormal basis, and Parseval is the Pythagorean theorem in Hilbert Space.
It is the natural tool for Partial Differential Equations — separation of variables in the heat, wave, and Laplace equations produces exactly these trigonometric eigenfunctions.
The Fourier transform turns <span class="wiki-link-missing">Convolution</span> into pointwise multiplication, the foundation of Linear Time-Invariant Systems.
It generalizes to the Laplace Transform (complex frequencies), the Discrete Fourier Transform / Fast Fourier Transform (finite data), and abstract Harmonic Analysis on groups.
In Quantum Mechanics the position and momentum representations of a state are a Fourier transform pair, making [9]'s Gaussian trade-off literally Heisenberg's uncertainty.

🔬 Where this appears

⏳ History

- 1748 — Euler relates trigonometric series to the vibrating string problem; debate with d'Alembert and Daniel Bernoulli over which functions a string's shape can take.

- 1807 — Fourier submits his memoir on heat to the French Academy, claiming arbitrary functions expand in trigonometric series; the committee (Lagrange, Laplace, Legendre) withholds full approval.

- 1822 — Fourier publishes Théorie analytique de la chaleur, establishing the method despite incomplete rigor.

- 1829 — Dirichlet gives the first rigorous convergence theorem for piecewise-monotone functions ([5]).

- 1854 — Riemann, in his habilitation, develops the Riemann integral partly to clarify which functions have Fourier series, and constructs the Riemann–Lebesgue ideas.

- 1873 — du Bois-Reymond exhibits a continuous function whose Fourier series diverges at a point, ending hopes for universal pointwise convergence.

- 1899 — Gibbs (after Wilbraham, 1848) explains the persistent overshoot at jumps — the Gibbs phenomenon.

- 1902–1910 — Lebesgue's integration theory and the Riesz–Fischer theorem put

L^2

convergence on solid ground; Plancherel proves [6] for the transform.

- 1965 / 1966 — Cooley & Tukey publish the Fast Fourier Transform (FFT), reducing the DFT cost from

O(N^2)

to

O(N\log N)

; Carleson proves a.e. convergence of

L^2

Fourier series.

open-problem

Several boundaries of the classical theory remain delicate or active:

Multidimensional convergence: in $\mathbb{R}^d$ the order of summation matters; the "ball multiplier" problem (Fefferman, 1971) showed that spherical partial sums need not converge in $L^p$ for $p\ne 2$ when $d\ge2$ — a counterintuitive failure absent in one dimension.
Pointwise convergence at the edge of integrability: Kolmogorov (1923) built an $L^1$ function whose Fourier series diverges everywhere. The exact boundary between Carleson's a.e.-convergence ( $L^2$ ) and total divergence ( $L^1$ ) is governed by subtle conditions (e.g. $L\log L$ classes).
Restriction & Kakeya: modern harmonic analysis ties Fourier restriction estimates to the Kakeya conjecture, with open problems linking geometry, number theory, and PDE that drive contemporary research (work of Bourgain, Tao, and others).
Uncertainty limits: a function and its transform cannot both be sharply localized; the precise extremal trade-offs (Beurling–Malliavin, Hardy's theorem) remain a refined area.

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also