A Fourier series describes a periodic function — a wave that repeats itself forever.
But most signals worth listening to or looking at are not periodic at all. A spoken sentence
never repeats exactly. A single flash of lightning happens once and is gone. A heartbeat drifts
and changes from beat to beat. None of these have a period 2L to plug
into the series formula — so does Fourier's idea simply stop working for them?
It does not — it generalises beautifully. Imagine taking a periodic signal and slowly stretching
its period 2L longer and longer, out towards infinity. In the limit,
a single non-repeating pulse is just a "periodic" signal whose neighbouring copies have been
pushed infinitely far away, so they never come back to interfere. Let
L \to \infty in the complex
Fourier series,
and the spacing \pi/L between allowed frequencies shrinks to zero. The
sum over discrete n melts into an integral over a
continuous frequency k. That limit is the Fourier
transform — the same idea as the series, generalised to work on absolutely any signal,
periodic or not.
\hat{f}(k) = \int_{-\infty}^{\infty} f(x)\, e^{-ikx}\,dx, \qquad f(x) = \frac{1}{2\pi}\int_{-\infty}^{\infty} \hat{f}(k)\, e^{ikx}\,dk.
The first integral analyses f into its frequency
content \hat{f}(k); the second synthesises
f back out of it, exactly the way the series rebuilds a periodic wave
from its coefficients. The discrete coefficients c_n have become a
continuous spectrum — instead of a strength attached to each whole-number
frequency n, every real frequency k gets
its own (typically infinitesimal) share.
It is tempting to picture the transform as a Fourier series with infinitely many terms, but
that is not quite right — a series with infinitely many discrete frequencies still only "sees"
those whole-number multiples of the fundamental. What actually happens as
L \to \infty is that the fundamental frequency
\pi/L itself shrinks to zero, so the gaps between the allowed
frequencies close up completely and every real number becomes a frequency the signal can use.
A sum over a comb of dots becomes an integral over a solid line — that is the real
difference between a spectrum with isolated lines and a genuinely continuous
spectrum.
The hallmark: narrow in one domain, wide in the other
The Fourier transform trades localisation between the two domains, and this trade-off is the
single most important fact about it. Think of a sustained, nearly pure musical note — a flute
holding one steady pitch. Stretched out broadly in time, that signal needs barely any
frequencies to build: its transform is a sharp, narrow spike sitting right at
that one pitch. Now think of the opposite: a single sharp click, like a
snapped finger or a drumstick tapping a rim. It is concentrated into a tiny sliver of time —
and to build something that abrupt out of smooth waves, you need a huge range of frequencies all
added together, cancelling everywhere except that one instant. Its transform is spread thin
across almost every frequency at once.
A broad, gentle pulse (like the note) needs only a few low frequencies, so its transform is
narrow; a sharply concentrated pulse (like the click) needs many, so its transform is spread out.
This reciprocity — the foundation of the Heisenberg uncertainty principle in quantum mechanics —
is cleanest to see for the Gaussian bump, which happens to be its own transform
shape:
Push the "sustained note" idea to its absolute limit: a tone that has been playing forever,
f(x) = e^{ik_0 x}, one single frequency k_0
and nothing else, stretching infinitely far in both directions of time. Its transform is, quite
literally, zero everywhere except at the one point k = k_0, where it
is infinitely tall and infinitesimally thin — a genuine spike, formally written using the
Dirac delta, \hat f(k) = 2\pi\,\delta(k-k_0). That is the
idealised endpoint the Gaussian below is edging towards as its pulse widens: an infinitely
patient note collapses onto one exact frequency, while a signal squeezed into an instant does
the opposite and spreads across every frequency there is. Real tones and clicks sit somewhere
between these two extremes, but the two limits show you exactly which way each knob turns.
f(x) = e^{-a x^2} \;\Longrightarrow\; \hat{f}(k) = \sqrt{\tfrac{\pi}{a}}\; e^{-k^2 / (4a)}.
A wide bump in x (small a, like the
sustained note) becomes a narrow bump in k; a narrow bump in
x (large a, like the click) becomes a wide
one in k.
Try it: from a "note" to a "click"
The pulse is on the left, its transform on the right. Slide the width parameter
a down towards its minimum and the pulse spreads out like a long,
held musical note — watch its transform collapse into a tall, narrow spike at
k=0, a single dominant frequency. Now slide
a up towards its maximum: the pulse squeezes into a sharp, click-like
spike — and its transform spreads out flat and wide, using a broad range of frequencies almost
equally. The two widths always move in opposite directions: you can never make both
narrow at once, no matter how you choose a.
Where this actually shows up
This is not an abstract curiosity — reaching for the frequency content of a signal, instead of
its raw up-and-down wiggle over time, is what makes a whole family of everyday devices work:
-
Audio equalizers. The bass, mid and treble sliders on a stereo (or the graphic
equalizer in a music app) work by Fourier-transforming the sound, boosting or cutting the
bands of frequency you choose, and transforming back — something that is completely
impractical to do directly on the raw waveform.
-
MRI scanners. A medical MRI machine never photographs the body directly. It
collects raw data that lives in frequency space (radiologists literally call it "k-space",
after the same k you've just met) and reconstructs the final image
with an inverse Fourier transform.
-
WiFi and radio. A WiFi signal splits your data across a whole comb of separate
carrier frequencies packed close together; the receiver's antenna picks up one messy combined
wave, and only by Fourier-transforming it can the chip separate that jumble back into the
individual frequency channels carrying the data.
It is easy to think of the time/frequency trade-off you just saw with the Gaussian as some
practical limitation — a fuzziness caused by cheap equipment, a sloppy algorithm, or noisy
measurements that a better microphone or a smarter piece of software could someday fix around.
It cannot be fixed, by any device or any algorithm, ever. A signal that is sharply localised in
time is, as a matter of pure mathematics, necessarily spread out in
frequency — and a signal concentrated in one narrow band of frequency is
necessarily spread out over time. You cannot design a signal, however cleverly, that is
perfectly narrow in both domains at once. This is exactly the same mathematics behind the
Heisenberg uncertainty principle in physics, where position and momentum play the roles of time
and frequency — it is a genuine property of waves themselves, not a shortcoming of any
particular technology.
Computing a Fourier transform directly, term by term, is expensive: doubling how finely you
sample a signal roughly quadruples the work. In 1965, James Cooley and John Tukey published
the Fast Fourier Transform (FFT), a way to compute exactly the same answer
using dramatically fewer operations — turning an N^2 calculation
into one that scales like N\log N. That change of scaling is why
your phone can equalise, compress, or analyse audio in real time at all, and it sits quietly
underneath the MP3 and JPEG formats, digital television, and pretty much every piece of audio
or image software you've ever used. It is routinely ranked among the most important algorithms
of the twentieth century — a purely mathematical idea about sines and cosines that ended up
reshaping consumer electronics.