How Fourier Transforms Work and Why They Matter

Here's a question that sounds simple and isn't: if you hand someone a single squiggly line, a microphone recording, a photo's brightness values, a stock price, can they tell you what "notes" are hiding inside it?

Not metaphorical notes. Literal ones. Every signal, no matter how messy, turns out to be a stack of plain sine waves added together. Find those waves and you can compress the signal, clean it up, or redesign it piece by piece. Miss them and you're stuck staring at a jagged line with no idea what's driving it.

The Fourier transform is the tool that finds them. And the way it works is stranger, and more physical, than the formula makes it look.

A wave hidden inside another wave is still there

Picture two speakers. One plays a steady 3 Hz hum, the other a faint 11 Hz whine on top. Record both at once and you don't get two waves. You get one wave: a single messy line that is the two originals added together, sample by sample.

Nothing about that hum and whine gets deleted. They're both still in there, just tangled up. The question is how you untangle them without already knowing what to look for.

Drag the winding speed slider. At most speeds, the signal loops around the circle and spreads out evenly, so its average position sits right near the center. But set the speed to 3 or 11, the actual frequencies hiding in the signal, and the loop stops being even. It bunches up on one side, and the center of mass jumps away from zero.

That jump is the whole trick. A frequency is "in" a signal exactly when winding the signal at that speed makes it lopsided.

Why winding it up works

Think of the signal as a rope, and you're wrapping it around a pole. If you wrap at some random, unrelated speed, each loop lands in a slightly different spot than the last. Over many loops, the wraps smear out evenly around the pole. Average their positions and you get roughly the center, because the smearing cancels itself out.

But if you wrap at the exact speed the rope is already oscillating at, something different happens. Every loop lands in the same place relative to the last one. The high points of the rope keep landing on the same side of the pole, over and over. Nothing cancels. The wraps pile up, and the average position gets dragged away from the center, toward wherever the rope's high points are.

That's not a metaphor for the math. It is the math. Winding a signal $g(t)$ around a circle at frequency $f$ means plotting the point

for every moment $t$, then averaging all those points. Written with complex numbers, that average is exactly the Fourier transform:

The $e^{-2\pi i f t}$ term is just shorthand for "wind at frequency $f$." The integral is just "average all the wound-up points." When $f$ matches a real frequency in the signal, the average lands far from zero. When it doesn't, the average collapses to nothing. There's no extra magic beyond winding and averaging.

Building the whole picture, one speed at a time

Winding at one frequency tells you whether that one frequency is present. To get the full picture, the natural move is to just try every speed and record where the loop bunches up.

Press play and watch the scan sweep from slow to fast. Most of the time, the bar stays low, because most speeds don't match anything in the signal. Then the scan crosses 3 Hz, and a spike appears. It crosses 7 Hz, another spike. Cross 11 Hz, a third. Those three bars are the whole answer: the signal is exactly those three frequencies, stacked, nothing else.

That bar chart has a name: the frequency spectrum. It's the same information as the original wiggly line, just reorganized. Instead of "here's the height at each moment in time," it says "here's how much of each frequency is present." Two languages, one signal. And here's the part that makes it genuinely useful:

you can edit a signal by editing its spectrum, in ways that would be nearly impossible in the time domain.

Why this actually matters: cutting out noise

Say you've got a clean signal, but it got recorded alongside a bunch of high-pitched noise, static, hiss, interference. In the raw waveform, the noise and the real signal are mixed into one line. There's no obvious place to grab a pair of scissors.

But in the frequency domain, the real signal usually lives at a few low frequencies, and the noise is scattered across a wide, high-frequency range. Once you can see that separation, cleaning up the signal is just deleting the high frequencies and keeping the low ones.

Drag the cutoff down. Watch the reconstructed line, the accent-colored one, smooth out and start tracking the true signal underneath (click the button to reveal it). This is exactly what a noise-canceling headphone, a static filter on an old radio recording, or a denoising step in a camera sensor is doing under the hood. None of them touch the waveform directly. They all edit the spectrum and transform back.

Why this actually matters: compressing files

Here's a related trick, and it's the one behind nearly every music and image file on your phone.

A real signal, like a second of someone's voice, usually contains dozens of frequency components. But they're not all equally loud. A few carry almost all the energy. The rest are quiet enough that dropping them barely changes what you hear or see.

Pull K down and compare the two strategies. Keeping the lowest frequencies seems like the safe, conservative choice, but watch the error: it stays high, because some of the loudest, most important components in this signal sit at higher frequencies. Keeping the strongest components, regardless of where they sit, gets you a near-perfect reconstruction with a fraction of the data.

That's the entire idea behind MP3 and JPEG compression. Transform the signal into frequencies, throw away the ones your ears or eyes barely notice, and keep the rest. The file gets smaller because most of a real signal's information is concentrated in surprisingly few frequencies.

The short version

Any signal, sound, image, radio wave, voltage, can be written as a sum of plain sine waves at different frequencies. To find them, wind the signal around a circle at a candidate speed and check whether the result balances evenly or bunches up. Bunching up means that frequency is really there. Doing this across every possible speed builds the full frequency spectrum, a second, equally complete description of the same signal.

Once a signal lives in the frequency domain, editing it gets easy. Delete the noisy high frequencies and the signal gets cleaner. Keep only the loudest frequencies and the file gets smaller. Every time your phone plays music, compresses a photo, or filters a call, some version of this winding-and-averaging trick is running in the background, quietly taking a messy signal apart into the simple waves it was made of.