Fourier Series: Building Any Wave from Sines and Cosines

In 1807, the French mathematician Joseph Fourier submitted a paper to the Paris Academy of Sciences containing a claim that the established mathematicians on the review panel — Lagrange, Laplace, Legendre, Lacroix — found impossible to accept. Fourier was claiming that any periodic function, no matter how jagged or discontinuous, could be exactly represented as a sum of sines and cosines.

A sine wave is smooth and rounded. A square wave is flat with vertical jumps. To these mathematicians, the idea that you could build a square wave by adding smooth sine waves was absurd. Sines are smooth; sums of smooth things are smooth; therefore the sum cannot have jumps. Fourier’s claim contradicted their intuition.

Fourier was right. The paper was rejected (his presentation lacked rigor by the standards of the time), but he kept refining the argument. By 1822 he had published Théorie analytique de la chaleur — the foundational text on what we now call Fourier series — and the mathematical world had begun to accept that any reasonable periodic function really could be decomposed into sine and cosine components.

This article is about what Fourier series are, why they work despite the apparent contradiction, and how they came to power most of modern technology.

The basic idea

A periodic function $f(x)$ with period $2\pi$ can be written as an infinite sum:

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

The coefficients $a_n$ and $b_n$ are determined by the function:

$a_n = \frac{1}{\pi} \int_{-\pi}^{\pi} f(x) \cos(nx) \, dx, \quad b_n = \frac{1}{\pi} \int_{-\pi}^{\pi} f(x) \sin(nx) \, dx.$

These integrals project $f$ onto the cosine and sine basis. Each $a_n$ measures “how much” of $\cos(nx)$ is in $f$; each $b_n$ measures “how much” of $\sin(nx)$.

The remarkable fact is that this series, in principle, reconstructs $f$ exactly (at points where $f$ is continuous) using only the values of these coefficients. The frequency-domain coefficients $\{a_n, b_n\}$ contain all the information needed to recover the time-domain function $f$.

Building a square wave

The cleanest illustration of Fourier’s claim is the square wave — a function that takes value +1 for half the period and -1 for the other half:

This function has no continuous derivative. It jumps abruptly. How can adding smooth sine waves produce these jumps?

The Fourier series of the square wave (period $2\pi$, antisymmetric) is:

$f(x) = \frac{4}{\pi} \left[ \sin(x) + \frac{\sin(3x)}{3} + \frac{\sin(5x)}{5} + \frac{\sin(7x)}{7} + \cdots \right].$

Only odd harmonics, with amplitudes decreasing as $1/n$. Watch what happens as you add more terms:

The blue curve is just $\sin(x)$ — a single sine. The green curve adds $\sin(3x)/3$ — two terms. The red curve sums the first eight odd terms ($n = 1, 3, 5, 7, 9, 11, 13, 15$). As more terms are added, the result hugs the square wave more tightly. With infinite terms, the agreement becomes exact (except exactly at the jumps, where Fourier series converge to the average of the left and right limits).

The Gibbs phenomenon

You may notice something in the red curve: near the jumps, the partial sum overshoots the target by a small amount. This overshoot doesn’t go away as you add more terms — it gets narrower but stays about 9% above the target.

This is the Gibbs phenomenon. It’s a fact about how Fourier series converge near discontinuities: the partial sums always overshoot by approximately $1.0894...$ times the jump amount (a 9% overshoot above the jump). The overshoot gets compressed into a narrower region as you add more terms, but never disappears.

The Gibbs phenomenon was first noticed by Henry Wilbraham in 1848 and named after Josiah Willard Gibbs, who analyzed it in the 1890s. It has practical consequences: when you reconstruct a signal with a finite number of Fourier terms, you get characteristic “ringing” near sharp edges. This is visible in JPEG-compressed images near sharp boundaries — the small ringing artifacts you sometimes see are essentially Gibbs phenomena.

Why Fourier series work

The mathematical reason Fourier series can represent any reasonable periodic function comes from orthogonality.

In the space of periodic functions, the functions $\{1, \cos(x), \sin(x), \cos(2x), \sin(2x), \ldots\}$ form an orthogonal basis: the inner product $\int_{-\pi}^{\pi} f(x) g(x) \, dx$ is zero for any two distinct functions in this collection.

This is exactly analogous to how $\{1, x, y, z\}$ are orthogonal in 3D space. Once you have an orthogonal basis, any vector can be decomposed into its components along each basis direction. The “components” are computed by projection, which is exactly the formula for Fourier coefficients.

So Fourier series are nothing more than decomposing a function into its components along an orthogonal basis — just an infinite-dimensional version of expressing a vector in $\mathbb{R}^3$ as a sum of $x$, $y$, and $z$ components. (See our vector spaces piece for the broader picture.)

This perspective also explains why the series exists for any reasonable function: orthogonal bases are general mathematical objects, and you can decompose along them. The question of convergence — whether the infinite series actually equals the function — is more subtle and depends on the function’s smoothness.

Convergence subtleties

A function with discontinuities has a Fourier series that converges to it almost everywhere, but not in the strongest possible sense:

• Pointwise convergence: at every point where $f$ is continuous, the Fourier series converges to $f$. At a jump, it converges to the average of the left and right limits.
• Uniform convergence: requires $f$ to be continuous. Discontinuities break uniform convergence — that’s what the Gibbs phenomenon reveals.
• $L^2$ convergence: the series converges in the mean-square sense for any square-integrable function. This is the cleanest statement and the foundation of much modern analysis.

The history of getting these subtleties right is dramatic. Fourier’s original 1822 work was correct in spirit but lacked the rigor demanded by his contemporaries. Dirichlet (1829) gave the first rigorous conditions guaranteeing convergence. Riemann’s habilitation lecture (1854) developed integration theory partly to handle Fourier series. The 20th century saw measure theory emerge in part as the right framework for these convergence questions — particularly through Lebesgue’s work.

In 1966 Lennart Carleson proved that the Fourier series of any $L^2$ function converges almost everywhere (i.e., on a set of measure 1). This was a major outstanding question; Carleson’s proof is widely considered one of the most difficult theorems of 20th-century analysis.

Real-world applications

Fourier series and their continuous cousin, the Fourier transform, are everywhere in modern technology.

Audio compression (MP3, AAC, Opus): convert audio to its frequency representation, discard frequencies humans can’t hear, transmit only the kept ones. The whole MP3 industry rests on Fourier analysis.

Image compression (JPEG, JPEG 2000): the same idea in 2D. The DCT (Discrete Cosine Transform) — a close relative of Fourier — converts each 8×8 image block into frequency components. Low-information frequencies are quantized aggressively, producing the compressed file.

Wireless communication (Wi-Fi, 5G, LTE): OFDM (Orthogonal Frequency-Division Multiplexing) sends data on many parallel Fourier sub-carriers. Modern wireless protocols are essentially elaborate uses of Fourier-based modulation.

Music synthesis: additive synthesizers literally add sinusoids at different frequencies to produce sounds. Subtractive synthesizers do the opposite — start with rich harmonic content and filter out frequencies. Both rely on Fourier thinking.

Speech recognition: convert speech to a “spectrogram” showing frequency content over time. The patterns are how phonemes are identified.

Medical imaging: MRI scans reconstruct images via Fourier inversion — the scanner measures spatial frequencies of the body, and a 2D inverse Fourier transform produces the image you see.

Astronomy: spectroscopy decomposes starlight into its frequency components. The Hubble Space Telescope and many ground-based observatories use Fourier-based analysis to extract chemical signatures.

Vibration analysis: structural engineers analyze building vibrations via Fourier methods. The natural frequencies of a structure (Fourier eigenvalues) determine its response to earthquakes, wind, and other forces.

The reach is enormous. Almost any technology that processes signals does Fourier analysis internally.

The Fast Fourier Transform

A footnote on practical computation. Computing a Fourier series naively requires $O(n^2)$ operations for $n$ frequency components. In 1965, James Cooley and John Tukey rediscovered (after Gauss had it in 1805) an algorithm that does it in $O(n \log n)$. This is the Fast Fourier Transform (FFT).

The FFT made Fourier analysis computationally tractable at scale. Before FFT, computing the spectrum of a long signal was infeasible. After FFT, it became a basic operation that runs in real time even on embedded hardware. Modern signal processing is unimaginable without FFT.

For most of computational mathematics, FFT is a fundamental subroutine — used billions of times per second worldwide in everything from phone calls to scientific simulations.

What Fourier series teach

The deepest lesson of Fourier series is that decomposition into frequency components is a fundamental operation on functions. Whatever signal you’re working with — audio, image, video, biomedical, financial — the frequency representation often reveals structure that the time-domain representation hides.

This perspective also generalizes far beyond classical Fourier series:

• Wavelets: alternative basis functions that capture local features better than global sinusoids. Used in image compression (JPEG 2000), denoising, and feature detection.
• Spectrograms: short-time Fourier transforms showing how frequency content changes over time. Foundational in speech and music processing.
• Generalized Fourier analysis on groups, manifolds, and abstract spaces — central to modern harmonic analysis, representation theory, and number theory.

For students, Fourier series are one of the more rewarding subjects to learn — accessible (just sums of trig functions), visually compelling (you can plot partial sums and watch them converge), and astonishingly applicable. The cognitive shift from “this signal is a function of time” to “this signal is a sum of frequencies” is one of the most useful conceptual moves in applied mathematics.

For working scientists and engineers, Fourier analysis is the lingua franca of signal processing. When you see something labeled “DSP” (digital signal processing), Fourier is at the core. When you compress audio or video, Fourier is in the codec. When you transmit data wirelessly, Fourier is in the protocol.

Joseph Fourier, in 1807, made a claim that his contemporaries thought was impossible: smooth sinusoids can build any periodic function. He was right. Two hundred years later, that single insight is doing more computational work, in more places, than perhaps any other piece of 19th-century mathematics. The square wave really does come from sines, if you have enough of them. The reviewers who rejected the original paper would, with the benefit of hindsight, probably reconsider.

Frequently asked

What's the difference between Fourier series and Fourier transform?

Fourier series decompose periodic functions into discrete sines and cosines at integer multiples of a fundamental frequency. Fourier transforms decompose non-periodic functions into a continuous spectrum of frequencies. Series is the discrete case (sum); transform is the continuous case (integral). They're two faces of the same mathematical idea — frequency-domain representation of functions.

What is the Gibbs phenomenon?

When a Fourier series approximates a function with a jump discontinuity (like a square wave), the partial sums overshoot near the jump by about 9% of the jump height — no matter how many terms you add. The overshoots get narrower but not shorter. This is the Gibbs phenomenon, discovered by Henry Wilbraham in 1848 and named for Josiah Willard Gibbs who studied it later.

Are Fourier series actually used in everyday technology?

Constantly. JPEG image compression decomposes images into Fourier-like frequency components and discards high-frequency information that humans don't notice. MP3 audio compression does the analogous thing in time. Wi-Fi, 5G, and most modern wireless protocols use Fourier methods to multiplex multiple signals on the same channel. Every electronic device with audio or video processing uses some form of Fourier analysis.

Continue reading

Formula

The Fourier Transform

The integral transform that decomposes a function into its frequency spectrum — foundation of harmonic analysis, quantum mechanics, and modern signal processing.

Read →Person

Leonhard Euler

Swiss mathematician of extraordinary range and productivity. Formulated Euler's identity, founded graph theory, and produced more mathematical output than any mathematician before or since.

Read →Blog

Stokes' Theorem: The One Equation That Unifies Vector Calculus

The fundamental theorem of calculus, Green's theorem, the divergence theorem, and the classical Stokes' theorem are not four separate results — they are one short equation written four times.

Read →