What the FFT Actually Shows You — Fourier Transform Explained | PinePaper

The Question

Play a chord on a piano — say, C and E together. Your ear hears one sound. But that sound is two frequencies superimposed: 261.6 Hz and 329.6 Hz. Your cochlea physically separates them — different hair cells resonate at different frequencies, sending distinct signals to your brain.

The Fast Fourier Transform does the same thing, but with numbers instead of hair cells. Give it a signal (a sequence of amplitude samples over time) and it returns a list of frequencies and their strengths. It answers: what frequencies are present, and how much of each?

What's Actually Happening

A signal sampled over time is a list of numbers: the amplitude at each sample point. A 1-second recording at 44,100 Hz is 44,100 numbers. These numbers describe the signal in the time domain — amplitude as a function of time.

The FFT converts this to the frequency domain — amplitude as a function of frequency. Same information, different representation. Like switching between Cartesian and polar coordinates: nothing is created or destroyed, only re-expressed.

The mathematical core: every periodic signal can be written as a sum of sine and cosine waves at different frequencies. This is Fourier's theorem (1807). The FFT computes the coefficients of that sum — how much of each frequency is in the signal.

Why "Fast"

The naive way to compute a Fourier transform requires N² operations for N samples. For 1024 samples, that's about 1 million operations. The Cooley-Tukey algorithm (1965) reduces this to N·log₂(N) — about 10,000 operations for the same input. A 100x speedup. For a million samples, the speedup is 50,000x.

The trick: split the N-point transform into two N/2-point transforms, recursively. This requires N to be a power of 2 (or you pad with zeros). Each split halves the problem. The "butterfly" operation combines the halves:

X[k]     = Even[k] + W · Odd[k]
X[k+N/2] = Even[k] - W · Odd[k]

Where W is a complex exponential (a rotation in the complex plane). The same two sub-results give you two output points. This is why the algorithm is "fast" — it reuses every computation twice.

PinePaper's implementation is a textbook Cooley-Tukey radix-2 DIT (decimation in time). 40 lines of JavaScript. We wrote it from scratch rather than importing a library because we wanted students to be able to read the source and understand every line.

What Those Bars Mean

When you see a spectrum analyzer — bars jumping to music — each bar represents a frequency bin. The height is the magnitude (strength) of that frequency in the current signal.

A pure sine wave produces one tall bar at its frequency and nothing else.
A square wave produces bars at the fundamental and every odd harmonic (3rd, 5th, 7th...), decreasing as 1/n. This is why square waves sound "buzzy" — they contain high-frequency energy that pure sines don't.
White noise produces bars of roughly equal height everywhere. Every frequency is present with equal probability.
A human voice produces a fundamental (the pitch you hear) plus formants — resonant peaks from the shape of your vocal tract that distinguish vowels.

Windowing: Why the Edges Matter

There's a catch. The FFT assumes the signal repeats forever. But our sample is finite — it starts and stops. If the signal doesn't happen to be at zero at both endpoints, the abrupt cutoff creates artificial high-frequency content. This is called spectral leakage.

The fix: multiply the signal by a window function that tapers smoothly to zero at the edges. Common windows:

Hann (cosine bell): good general purpose, loses some frequency resolution
Hamming: similar to Hann but doesn't reach zero at edges, slightly better sidelobe suppression
Blackman: narrower main lobe, better sidelobe suppression, loses more frequency resolution

The choice is always a tradeoff between frequency resolution (how precisely you can identify a frequency) and spectral leakage (how much energy bleeds into neighboring bins). There is no perfect window. This is a consequence of the uncertainty principle — you cannot have arbitrarily precise knowledge of both time and frequency simultaneously.

Where the FFT Lives

You interact with FFT results constantly:

MP3 and AAC compression: transform audio to frequency domain, discard frequencies below the hearing threshold, compress what remains. The transform is the entire basis of lossy audio compression.
JPEG compression: the 2D version (DCT) transforms 8×8 pixel blocks to frequency domain, quantizes high-frequency components. That's why JPEG artifacts appear as blocks.
WiFi and 5G: OFDM encoding splits data across many frequency sub-carriers. The FFT converts between time-domain transmission and frequency-domain data symbols.
MRI imaging: the raw signal from an MRI scanner is in frequency space. The inverse FFT reconstructs the spatial image. Literally: every MRI you've ever seen is an inverse Fourier transform.
Shazam: computes the spectrogram (FFT over sliding windows), extracts peaks, matches the pattern against a database. The FFT is the first step in recognizing every song.

A 60-year-old algorithm, in your pocket, running billions of times per day.

Try It

Open PinePaper, select the Spectrum Analyzer generator. Generate a square wave. Look at the bars — you'll see the odd harmonics falling off as 1/n. Switch to a sawtooth — now all harmonics are present, falling off as 1/n. Switch to noise — flat spectrum, every frequency equally likely.

Change the window function. Watch how Hann smooths the spectrum at the cost of wider peaks. Switch to Blackman — narrower peaks but lower sidelobes.

You're not reading about the FFT. You're measuring signals and observing what the transform reveals. That's the difference between knowing and understanding.

References

Brigham, E.O. (1988). The Fast Fourier Transform and Its Applications. Prentice Hall.
Cooley, J.W. & Tukey, J.W. (1965). An Algorithm for the Machine Calculation of Complex Fourier Series. Mathematics of Computation, 19(90), 297-301.
Fourier, J. (1822). Théorie analytique de la chaleur. Paris: Firmin Didot.
Harris, F.J. (1978). On the Use of Windows for Harmonic Analysis with the Discrete Fourier Transform. Proceedings of the IEEE, 66(1), 51-83.
Oppenheim, A.V. & Schafer, R.W. (2009). Discrete-Time Signal Processing (3rd ed.). Prentice Hall.
Shannon, C.E. (1949). Communication in the Presence of Noise. Proceedings of the IRE, 37(1), 10-21.
Smith, S.W. (1997). The Scientist and Engineer's Guide to Digital Signal Processing. California Technical Publishing.
Wang, A., et al. (2003). An Industrial-Strength Audio Search Algorithm. Proceedings of ISMIR 2003. (Shazam's audio fingerprinting algorithm.)
Wallace, G.K. (1991). The JPEG Still Picture Compression Standard. Communications of the ACM, 34(4), 30-44.

PinePaper's FFT is a Cooley-Tukey radix-2 implementation with Hann, Hamming, and Blackman windowing, plus low-pass and high-pass filters. Try it free at pinepaper.studio/editor.