Fourier & the frequency domain · 03 · Frequency for everything · 9 min
HardThe Fourier transform
Fourier series handle signals that repeat. For everything else — a spoken word, a door slam, a stock chart — let the repeat period grow to infinity: harmonics crowd together into a continuum, and the series becomes the Fourier transform. Every signal gets a spectrum.
Build the intuition
From discrete recipe to continuous spectrum
A repeating signal has energy only at harmonic spikes. Stretch the period: spikes pack tighter; in the limit of “never repeats,” the spectrum X(ω) becomes a smooth curve — how much of every frequency the signal contains. Same projection idea, now against e^{iωt} probes at all frequencies.
Two complete views of one object
x(t) and X(ω) carry identical information — the inverse transform rebuilds the signal perfectly. Some questions are easy in time (when did the door slam?), others in frequency (what pitch is the hum?). Professionals hop between views constantly; the transform is the turnstile.
The crown jewel: convolution becomes multiplication
Filtering — convolution in time — becomes ordinary multiplication in frequency: Y(ω) = H(ω)X(ω). The output spectrum is input spectrum times frequency response, frequency by frequency. This is why filter design happens in the frequency domain, and a major reason transforms run the world's signal chains.
The uncertainty trade
Squeeze a signal short in time and its spectrum spreads wide; a long pure tone has a razor spectrum. Time-width × bandwidth has a floor — the same mathematics as quantum's uncertainty principle. A click can't have a pitch; a pitch can't be instantaneous. This trade haunts windowing (lesson 7).
See it move
Time domain — what the microphone sees
Frequency domain — what the ear hears
The transform's deliverable: the tangled top view and the legible bottom view are the same signal. The spectrum reads what the waveform hides.
A worked example
Diagnose a hum
A podcast recording sounds dirty. The waveform shows… wiggles. Useless.
Transform it: a clean speech spectrum plus one needle at exactly 60 Hz — mains interference from a ground loop.
Fix: multiply the spectrum by a notch H(ω) that zeroes 60 Hz, transform back. The waveform view never even saw the problem; the frequency view solved it.
Out in the world
MRI and the chemistry lab
MRI machines excite hydrogen nuclei and record a tangled echo; Fourier transforming it (in 2D) reveals the spatial image. NMR spectroscopy identifies molecules by the frequencies in their response. Whole industries are built on “measure mess in time, read truth in frequency.”
Common confusion, cleared
“The spectrum is an approximation of the signal.”
It's a lossless re-description — perfectly invertible. Nothing is discarded; you've only rotated to better axes (lesson 1's basis change, taken to a continuum).
“Negative frequencies are a glitch in the math.”
They're the second half of each phasor pair: a real cosine is two counter-rotating arrows (e^{iωt} + e^{−iωt})/2. Negative ω is the counter-rotation — bookkeeping for realness, not nonsense.
Recap
- The transform extends Fourier's recipe to non-repeating signals: a spectrum for everything.
- Time and frequency views are equivalent and invertible — pick whichever answers your question.
- Convolution ↔ multiplication; short-in-time ↔ wide-in-frequency.
Progress saves in this browser.