Fourier Series: Rigorous Foundations

Can a vibrating string, a temperature distribution, or any reasonably well-behaved periodic signal be decomposed into pure sine and cosine waves? This question was first attacked seriously by Joseph Fourier while studying heat conduction. It forces us to confront a deep issue in analysis. Given a $2\pi$ -periodic function $f$ , can we find coefficients $a_n, b_n$ so that

f(x) = \frac{a_0}{2} + \sum_{n=1}^{\infty} \left(a_n \cos nx + b_n \sin nx\right),

and in what sense does this series actually converge to $f$ ? The answer turns out to be subtle. The naive hope that every periodic function equals its Fourier series pointwise is false in general. Making the theory rigorous requires the full strength of the Riemann Integral. It also needs careful limit arguments from the theory of Uniform Convergence and Integration. This matters because of Fourier's idea. He proposed that "arbitrary" functions decompose into elementary oscillations. This became one of the most productive ideas in mathematics and physics. It underlies the analysis of heat, sound, light, and (much later) digital signal processing and quantum mechanics. The trigonometric system $\{1, \cos x, \sin x, \cos 2x, \sin 2x, \dots\}$ has a remarkable orthogonality structure on any interval of length $2\pi$ . This is exactly the tool needed to extract the coefficients $a_n, b_n$ from $f$ by integration. This works regardless of whether the resulting series converges back to $f$ . The big picture is this: we first define the Fourier coefficients of any Riemann-integrable periodic function purely formally. This is done via explicit integral formulas, with no convergence assumed. We then form the partial sums $S_N(f)$ of the resulting series. We study them as ordinary functions built from finitely many trigonometric terms. Only after this scaffolding is in place does the harder question make sense. Coefficients must be well-defined, partial sums well-defined, and orthogonality proven. Then we can ask when and how $S_N(f) \to f$ . This question belongs to the theory of convergence of Fourier series. It is treated separately. This article delivers the rigorous foundational layer. It gives the precise definition of Fourier coefficients for Riemann-integrable periodic functions. It covers the algebraic and integral structure of the trigonometric system. It also presents the Riemann–Lebesgue-type integrability facts needed to make sense of the coefficients. Finally, it gives a fully proved orthogonality theorem that underlies everything downstream. This includes Bessel's Inequality and the eventual convergence theorems.

Key takeaways

Fourier coefficients $a_n, b_n$ of a $2\pi$ -periodic, Riemann-integrable $f$ are defined by explicit integral formulas — no convergence is assumed a priori.
The trigonometric system $\{1,\cos nx,\sin nx\}_{n\ge1}$ is orthogonal on any interval of length $2\pi$ , which is the algebraic engine behind the coefficient formulas.
Partial sums $S_N(f)(x) = \tfrac{a_0}{2} + \sum_{n=1}^N (a_n\cos nx + b_n \sin nx)$ are trigonometric polynomials of degree $\le N$ , always well-defined and continuous.
Riemann integrability of $f$ guarantees the coefficients exist and are finite, via boundedness plus integrability of $f$ times bounded continuous functions.
The orthogonality theorem is proved directly from elementary trigonometric product-to-sum identities and definite integrals — no circularity, no appeal to convergence results.
Key idea: separate the algebra of extracting coefficients (rigorous, always works) from the analysis of whether the series reconstructs $f$ (subtle, treated afterward).

Where this lives today

Fourier's coefficient formulas and the orthogonality of the trigonometric system underlie an enormous range of modern mathematics and applied science. In pure analysis, they are the launching point for Bessel's Inequality and Parseval's theorem, which show the Fourier map is an isometry between periodic $L^2$ functions and square-summable coefficient sequences — a cornerstone of Hilbert space theory. In partial differential equations, separation of variables for the heat and wave equations on bounded domains produces exactly the trigonometric eigenfunctions Fourier used, making this the standard tool for solving boundary value problems. In signal processing and electrical engineering, the discrete and fast Fourier transforms (algorithmic descendants of the same coefficient formulas) underlie audio compression (MP3), image compression (JPEG's use of the related discrete cosine transform), and digital filtering. In cryptography and number theory, Fourier analysis on finite groups underlies primality testing algorithms and the Fast Fourier Transform speeds polynomial multiplication used in RSA and other cryptosystems. In statistics and time-series analysis, spectral density estimation decomposes a random process into frequency components using the same orthogonality principle. In quantum mechanics, Fourier series and transforms relate position and momentum representations of a wavefunction, a direct descendant of the position/frequency duality Fourier first isolated for heat.

:::faq[Why use $[-\pi,\pi]$ instead of $[0,2\pi]$ for the defining integrals?] Any interval of length $2\pi$ gives the same coefficients, because $f$ , and hence $f(x)\cos(nx)$ and $f(x)\sin(nx)$ , are $2\pi$ -periodic, and the integral of a periodic function over any interval of length equal to one full period is the same regardless of where that interval starts. This is a direct consequence of periodicity and additivity of the Riemann integral over adjacent subintervals. The interval $[-\pi,\pi]$ is often chosen for convenience because it is symmetric about $0$ , which simplifies computations for even or odd functions (as seen in the worked examples), but $[0,2\pi]$ or any shifted interval works equally well and yields identical numbers.

Ask the tutor: How does the best-approximation property proved here get upgraded into a full convergence theorem for Fourier series?

f:\mathbb{R}\to\mathbb{R}

Intuition

Think of the trigonometric functions $\cos nx$ and $\sin nx$ as an infinite family of "pure tones," analogous to the way a musical chord decomposes into individual notes. The claim behind Fourier's construction is that these tones form a kind of coordinate system for periodic functions, and that the coefficients $a_n, b_n$ measure "how much" of the frequency- $n$ tone is present in $f$ . This is directly analogous to finding the components of a vector along orthogonal axes in $\mathbb{R}^n$ : if $\{e_1,\dots,e_n\}$ is an orthonormal basis, the $k$ -th component of a vector $v$ is $\langle v, e_k\rangle$ . The integral

\langle f, g\rangle := \int_{-\pi}^{\pi} f(x) g(x)\, dx

plays the role of an inner product on the (infinite-dimensional) space of Riemann-integrable periodic functions, and the trigonometric system is orthogonal with respect to it — meaning distinct basis functions have inner product zero. The coefficient formulas for $a_n$ and $b_n$ are then nothing but "take the inner product of $f$ with the $n$ -th basis function, and divide by the squared length of that basis function" — exactly the Gram–Schmidt/orthogonal-projection formula from linear algebra, transplanted into an infinite-dimensional integral setting. The crucial subtlety, which the rest of the Fourier theory is built to resolve, is that in infinite dimensions "expressing $f$ in the basis" is not automatic: unlike in $\mathbb{R}^n$ , a function need not equal the sum of its projections onto the basis vectors, even when there are infinitely many of them. That is precisely why this foundational article stops short of claiming $S_N(f) \to f$ ; it only builds the machinery — coefficients, orthogonality, well-defined partial sums — on which such convergence theorems can later rest.

Figure

The trigonometric system as an orthogonal "coordinate grid" for periodic functions

m, n \geq 0

x

Worked examples

Example 1: the sawtooth-type function $f(x) = x^2$ on $(-\pi,\pi]$ , extended $2\pi$ -periodically. Let $f$ be the $2\pi$ -periodic extension of $f(x) = x^2$ for $x \in (-\pi,\pi]$ . This $f$ is continuous (check: at $x=\pi$ , the left-hand value from the formula is $\pi^2$ ; periodicity forces $f(\pi) = f(-\pi)$ , and indeed the formula gives $f(-\pi) = (-\pi)^2 = \pi^2$ too, so there is no jump), hence certainly Riemann integrable on $[-\pi,\pi]$ , so its Fourier coefficients are well-defined. Since $f$ is even ( $f(-x) = (-x)^2 = x^2 = f(x)$ ), and $\sin(nx)$ is odd, the product $f(x)\sin(nx)$ is odd, so $b_n = \frac{1}{\pi}\int_{-\pi}^\pi f(x)\sin(nx)\,dx = 0$ for all $n \geq 1$ (using the odd-function-vanishing fact established in the proof above). Compute $a_0$ :

a_0 = \frac{1}{\pi}\int_{-\pi}^\pi x^2\,dx = \frac{1}{\pi}\left[\frac{x^3}{3}\right]_{-\pi}^\pi = \frac{1}{\pi}\cdot \frac{2\pi^3}{3} = \frac{2\pi^2}{3}.

Compute $a_n$ for $n \geq 1$ via integration by parts twice. Let $u = x^2$ , $dv = \cos(nx)\,dx$ , so $du = 2x\,dx$ , $v = \sin(nx)/n$ :

\int_{-\pi}^\pi x^2\cos(nx)\,dx = \left[\frac{x^2\sin(nx)}{n}\right]_{-\pi}^\pi - \frac{2}{n}\int_{-\pi}^\pi x\sin(nx)\,dx.

The boundary term vanishes since $\sin(n\pi) = 0$ and $\sin(-n\pi)=0$ for integer $n$ . For the remaining integral, let $u=x$ , $dv = \sin(nx)\,dx$ , $du=dx$ , $v = -\cos(nx)/n$ :

\int_{-\pi}^\pi x \sin(nx)\,dx = \left[-\frac{x\cos(nx)}{n}\right]_{-\pi}^{\pi} + \frac{1}{n}\int_{-\pi}^\pi \cos(nx)\,dx.

The remaining integral is $0$ (shown earlier, $n \neq 0$ ). The boundary term:

-\frac{\pi\cos(n\pi)}{n} - \left(-\frac{(-\pi)\cos(-n\pi)}{n}\right) = -\frac{\pi\cos(n\pi)}{n} - \frac{\pi\cos(n\pi)}{n} = -\frac{2\pi\cos(n\pi)}{n},

using $\cos(-n\pi)=\cos(n\pi)$ . Since $\cos(n\pi) = (-1)^n$ , this equals $-\dfrac{2\pi(-1)^n}{n}$ . So $\int_{-\pi}^\pi x\sin(nx)\,dx = -\dfrac{2\pi(-1)^n}{n}$ , and therefore

\int_{-\pi}^\pi x^2\cos(nx)\,dx = -\frac{2}{n}\cdot\left(-\frac{2\pi(-1)^n}{n}\right) = \frac{4\pi(-1)^n}{n^2}.

Hence

a_n = \frac{1}{\pi}\cdot \frac{4\pi(-1)^n}{n^2} = \frac{4(-1)^n}{n^2}, \qquad n \geq 1.

So the Fourier series of this $f$ is

f(x) \sim \frac{\pi^2}{3} + \sum_{n=1}^\infty \frac{4(-1)^n}{n^2}\cos(nx).

As a numerical check, the first coefficient is $a_1 = -4$ and $a_2 = 1$ ; note $a_0/2 = \pi^2/3 \approx 3.2899$ , matching the average value of $x^2$ over the interval, a sensible sanity check since $a_0/2$ is always the mean of $f$ over one period.

Example 2: an asymmetric, discontinuous function — a shifted, scaled step function. Let $f$ be $2\pi$ -periodic with

f(x) = \begin{cases} 3, & -\pi < x < 0 \\ -1, & 0 \le x < \pi\end{cases}.

This $f$ is bounded and has only one discontinuity per period (a jump at $x=0$ , and correspondingly at $x=\pi \equiv -\pi$ ), so it is Riemann integrable on $[-\pi,\pi]$ by the standard criterion that a bounded function with finitely many discontinuities on a compact interval is Riemann integrable. Compute $a_0$ :

a_0 = \frac{1}{\pi}\left(\int_{-\pi}^0 3\,dx + \int_0^\pi (-1)\,dx\right) = \frac{1}{\pi}\left(3\pi - \pi\right) = 2.

Compute $a_n$ , $n\geq 1$ :

a_n = \frac{1}{\pi}\left(3\int_{-\pi}^0 \cos(nx)\,dx - \int_0^\pi \cos(nx)\,dx\right) = \frac{1}{\pi}\left(3\left[\frac{\sin nx}{n}\right]_{-\pi}^0 - \left[\frac{\sin nx}{n}\right]_0^\pi\right).

Since $\sin(n\cdot 0)=0$ and $\sin(\pm n\pi)=0$ for integer $n$ , every bracket evaluates to $0$ . So $a_n = 0$ for all $n \geq 1$ . Compute $b_n$ , $n \geq 1$ :

b_n = \frac{1}{\pi}\left(3\int_{-\pi}^0 \sin(nx)\,dx - \int_0^\pi \sin(nx)\,dx\right).

Now $\int_{-\pi}^0 \sin(nx)\,dx = \left[-\frac{\cos nx}{n}\right]_{-\pi}^0 = -\frac{1}{n} + \frac{\cos(n\pi)}{n}\cdot(-1)\cdot(-1)$ ; let's be careful:

\int_{-\pi}^0 \sin(nx)\,dx = \left[-\frac{\cos(nx)}{n}\right]_{-\pi}^{0} = -\frac{\cos 0}{n} - \left(-\frac{\cos(-n\pi)}{n}\right) = -\frac{1}{n} + \frac{(-1)^n}{n} = \frac{(-1)^n - 1}{n}.

And

\int_0^\pi \sin(nx)\,dx = \left[-\frac{\cos(nx)}{n}\right]_0^\pi = -\frac{(-1)^n}{n} + \frac{1}{n} = \frac{1-(-1)^n}{n}.

b_n = \frac{1}{\pi}\left(3\cdot\frac{(-1)^n-1}{n} - \frac{1-(-1)^n}{n}\right) = \frac{1}{\pi n}\Big(3(-1)^n - 3 - 1 + (-1)^n\Big) = \frac{4(-1)^n - 4}{\pi n}.

When $n$ is even, $(-1)^n=1$ , giving $b_n = 0$ . When $n$ is odd, $(-1)^n=-1$ , giving $b_n = \dfrac{-8}{\pi n}$ . So

b_n = \begin{cases} 0 & n \text{ even} \\ -\dfrac{8}{\pi n} & n \text{ odd}\end{cases}.

As a check, $b_1 = -8/\pi \approx -2.546$ , and the series so far reads

f(x) \sim 1 - \frac{8}{\pi}\sin x - \frac{8}{3\pi}\sin 3x - \frac{8}{5\pi}\sin 5x - \cdots,

consistent with $a_0/2 = 1$ being the mean of $f$ (mean of $3$ on half the interval and $-1$ on the other half is indeed $1$ ), a useful numerical sanity check on the computation.

f

Scope, edge cases & common traps

Situation	What happens	Why
$f$ continuous and $2\pi$ -periodic	Coefficients always well-defined; $S_N(f)$ well-defined for every $N$	Continuous $\Rightarrow$ Riemann integrable on compact intervals
$f$ bounded with finitely many jump discontinuities per period	Coefficients well-defined	Bounded + finitely many discontinuities $\Rightarrow$ Riemann integrable
$f$ unbounded (even if "integrable" in an improper sense)	Formulas for $a_n, b_n$ may not make sense as Riemann integrals	Riemann integration requires boundedness on a compact interval
$f$ bounded but wildly discontinuous (e.g. Dirichlet-type)	Coefficients may fail to exist	Riemann integrability can fail even for bounded functions
Asking whether $S_N(f) \to f$ pointwise	Not addressed by this theorem	Convergence requires separate hypotheses (e.g. piecewise smoothness) — a distinct theorem
Using $S_N(f) = f$ as if automatic	A common error	Orthogonality only shows $S_N(f)$ is the best degree- $N$ approximation in mean square, not that the limit recovers $f$ pointwise or uniformly

Common trapThe best-approximation property is not a convergence theorem

It is tempting to read the minimization result — $S_N(f)$ minimizes $\int (f-T)^2$ among degree- $\le N$ trigonometric polynomials — as already proving $S_N(f) \to f$ as $N \to \infty$ . This is a non sequitur: minimality at each fixed $N$ says nothing by itself about the limiting behavior of the mean-square error as $N\to\infty$ , still less about pointwise or uniform convergence. Establishing $\int_{-\pi}^\pi (f - S_N(f))^2\,dx \to 0$ requires an additional density argument (completeness of the trigonometric system), and establishing pointwise convergence requires yet further hypotheses on $f$ (such as piecewise continuous differentiability). Conflating " $S_N(f)$ is the best approximant" with " $S_N(f)$ converges to $f$ " is one of the most common logical errors when first learning this material — the correct quantifier order is: for each fixed $N$ , $S_N(f)$ is optimal among degree- $\le N$ polynomials; nothing is claimed yet about the limit as $N \to \infty$ .

The mathematician behind it

Origins · 1 minJean-Baptiste Joseph Fourier and how this was discovered

Jean-Baptiste Joseph Fourier

1768–1830 · Napoleonic France, mathematical physics of heat

Fourier was a French mathematician and physicist whose 1822 treatise Théorie analytique de la chaleur ("The Analytic Theory of Heat") introduced the claim that arbitrary periodic functions could be represented as infinite sums of sines and cosines, in order to solve the heat equation on physical domains. His argument was not rigorous by modern standards — the notion of function, the concept of convergence, and the theory of the integral were all still being developed — but the mathematical content it forced into existence (orthogonality, convergence of infinite series of functions, the need for a precise theory of integration) reshaped 19th-century analysis. Fourier also served as a scientific advisor on Napoleon's Egyptian expedition and later as prefect of the Isère department, only returning fully to mathematical physics work in his final years.

History

- 1807: Fourier submits a memoir to the French Academy claiming arbitrary functions decompose into trigonometric series; the claim is met with skepticism (Lagrange objects).

- 1822: Fourier publishes Théorie analytique de la chaleur, cementing the trigonometric series method for solving the heat equation.

- 1829: Dirichlet gives the first rigorous convergence conditions for Fourier series (piecewise monotonic functions), putting Fourier's claims on partially solid footing.

- 1854: Riemann's habilitation work develops the Riemann integral partly in order to make precise sense of Fourier coefficients for more general functions.

- 1873: du Bois-Reymond constructs a continuous function whose Fourier series diverges at a point, showing pointwise convergence is genuinely delicate and not automatic from continuity alone.

Why do we require

f

to be Riemann integrable rather than continuous?

Continuity is sufficient but far stronger than necessary. Many functions of genuine interest — square waves, sawtooth signals, piecewise-defined physical models — have jump discontinuities, yet remain perfectly well-behaved for integration purposes: bounded with finitely many discontinuities on any bounded interval. The Riemann integrability criterion (a bounded function is Riemann integrable iff its set of discontinuities has measure zero, or more simply, iff it has at most finitely many discontinuities on a compact interval in the elementary case) is exactly the right level of generality to make the coefficient integrals $\int f\cos(nx)\,dx$ and $\int f \sin(nx)\,dx$ well-defined, without requiring the unnecessary luxury of continuity. This is why the foundational definition uses Riemann integrability rather than continuity as its hypothesis.

Does this article prove that Fourier series converge to the original function?

No, and deliberately so. This article establishes only the foundational layer: that the Fourier coefficients are well-defined real numbers whenever $f$ is Riemann integrable and periodic, that the trigonometric system is orthogonal, and that the partial sum $S_N(f)$ is the best mean-square approximation to $f$ among trigonometric polynomials of degree at most $N$ . Whether $S_N(f)$ actually converges to $f$ — pointwise, uniformly, or in mean square — is a separate and considerably subtler question, requiring additional hypotheses (such as piecewise continuous differentiability for pointwise convergence, or general density/completeness arguments for mean-square convergence). That material belongs to the theory of convergence of Fourier series.

What is the geometric meaning of the orthogonality relations?

They say that distinct members of the trigonometric system $\{1, \cos x, \sin x, \cos 2x, \sin 2x, \dots\}$ have zero "inner product" with respect to the pairing $\langle f,g\rangle = \int_{-\pi}^\pi fg\,dx$ , exactly as orthogonal vectors in Euclidean space have zero dot product. This makes the trigonometric system behave like an orthogonal basis for an infinite-dimensional inner product space of periodic functions. The Fourier coefficient formulas are then the exact analogue of computing components of a vector along orthogonal axes: divide the inner product with a basis vector by that basis vector's squared length. This geometric viewpoint is what makes the minimization theorem (best mean-square approximation) essentially a Pythagorean-theorem argument in disguise.

Are the coefficients

a_n, b_n

unique given

f

Yes, trivially, because they are defined by an explicit formula (a specific Riemann integral) applied to $f$ ; there is no ambiguity or choice involved. Uniqueness of representation is a different and subtler question: could two different Riemann-integrable functions $f \ne g$ (differing on a "small" set) have exactly the same Fourier coefficients? The answer is yes in general for arbitrary Riemann-integrable functions differing on a measure-zero set (e.g. at finitely many points), since altering finitely many values does not change any of the defining integrals. However, if $f$ and $g$ are both continuous and have identical Fourier coefficients, then $f = g$ everywhere — this stronger uniqueness statement follows from later convergence theorems, not from the coefficient definitions alone.

What role does linearity of the integral play in defining Fourier coefficients?

Linearity of the Riemann integral is used constantly and silently throughout: it justifies splitting $\int_{-\pi}^\pi fT\,dx$ into a sum of simpler integrals against $\cos(nx)$ and $\sin(nx)$ in the proof of the minimization theorem, it justifies computing $a_n$ and $b_n$ term-by-term for piecewise-defined $f$ , and it underlies the very fact that the map sending $f$ to its sequence of Fourier coefficients is a linear map. Without linearity of integration — which itself rests on the additive and scaling properties proved when the Riemann integral is first constructed — none of the coefficient extraction machinery would be algebraically well-behaved.

How does this connect to Bessel's Inequality mentioned as a "leads to" topic?

The minimization theorem proved here shows $\int_{-\pi}^\pi (f-S_N(f))^2\,dx \le \int_{-\pi}^\pi f^2\,dx$ once the completed-square computation in the proof is examined closely: since the "error" term after minimizing is $\int f^2\,dx - \frac{\pi}{2}a_0^2 - \pi\sum_{n=1}^N(a_n^2+b_n^2)$ , and this quantity must remain non-negative (it is a sum of squares, the minimum mean-square error), rearranging gives $\frac{a_0^2}{2} + \sum_{n=1}^N (a_n^2+b_n^2) \le \frac{1}{\pi}\int_{-\pi}^\pi f^2\,dx$ for every $N$ . Letting $N \to \infty$ yields exactly Bessel's Inequality, showing the foundational orthogonality result proved in this article is the direct algebraic ancestor of that later theorem.