Caught on camera. Lying on tape.
Both wheels are the same wheel, always turning forward at the speed you pick. Left is what your eyes see. Right is what a camera records: one photo per frame, nothing in between. Right now the camera takes only 6 photos per second, and its wheel rolls backwards. Nobody touched the brakes. Press the button to fix it, then play with the sliders to break it again.
The wheel turns forward the whole time. The camera footage shows it frozen, or rolling backwards.
A wheel with 4 identical spokes. Between two photos it turns 80° forward, a little less than the 90° from one spoke to the next. Here are the camera's photos:
The wheel in See it has 8 spokes: at 0.5 turns/s its pattern repeats 8 × 0.5 = 4 times a second, so the camera needs more than 8 fps. That's the Nyquist rule, applied to spokes.
A tone sweeps steadily up from 200 Hz to 11 kHz. Played correctly, it just rises. "Recorded" at a cheap 8 kHz with no anti-aliasing filter, it rises, then falls, then rises again, even though the real tone only ever went up.
Wheels that roll backwards, tones that sing the wrong note: same crime, same culprit. The rest of this page explains how it works.
A tone rises steadily from 200 Hz to 11 kHz. Recorded at 8 kHz, it is heard rising, then falling, then rising again.
8,000 samples per second can only describe tones up to 4 kHz (half of 8). A higher tone doesn't disappear: it gets folded back below 4 kHz, like paper folded along the 4 kHz line.
| real tone | what you hear |
|---|---|
| 3 kHz | 3 kHz |
| 4 kHz | 4 kHz |
| 5 kHz | 3 kHz |
| 6 kHz | 2 kHz |
| 7 kHz | 1 kHz |
| 8 kHz | 0 kHz |
| 9 kHz | 1 kHz |
| 11 kHz | 3 kHz |
Real recorders put an anti-aliasing filter in front of the sampler. It removes everything above 4 kHz before it can sneak in folded.
Translation: more than 2 snapshots per wiggle.
Same rule, the other way round: fs > 2 · fmax
*Terms & conditions apply. Signal must be band-limited (no frequencies above fmax, not even a little bit). Reconstruction requires an ideal low-pass filter, which, like a perfectly tidy desk, exists only in theory. Note the strict ">": right on the limit is a trap. Offer not valid for square waves.
The faint line is the real signal. The dots are your samples. The dashed line is what anyone holding only the dots would reconstruct. When the dashed line doesn't match: impostor detected.
A 9 Hz wave, 10 samples per second. The dots fit a 1 Hz wave perfectly.
Between two samples the 9 Hz wave goes through 0.9 of a cycle. Going 0.9 forward lands on exactly the same spot as going 0.1 backward, just like the wheel. So the dots also fit a wave that does only 0.1 cycles per sample: 1 Hz.
Same dots, and the phase comes out mirrored (φ becomes −φ). That's why the playground shows "mirrored!" above the limit.
At exactly fs/2 you take exactly 2 snapshots per wiggle. Catch the peaks and it looks fine. Catch the zero crossings and you record nothing at all.
Green is the real wave. The dots are your snapshots. Connect the dots and you get the dashed wave: smaller than the real one, or even flat. With exactly 2 snapshots per wiggle, how much you see is pure luck: it depends on how the wave lines up with your snapshots. Same wiggle, tiny shift, different result.
A full-size wave, sampled exactly twice per wiggle, can produce samples that are all zero. Where did it go?
Picture the wave as a clock hand going round. A lamp above throws its shadow on the floor. The shadow sliding left and right is the wave you record. Each sample is one photo.
4 photos per turn. 12 → 3 → 6 → 9 o'clock: clearly turning clockwise, and the shadow swings left and right. ✓ That's a wave.
The fix: anything more than 2 photos per turn. Even 3 is enough: The shortest way between photos is clockwise again, and the shadows trace the wave whatever moment you start. ✓
The punchline: the shadow only shows left-right. The up-down position of the hand never reaches the floor; mathematicians call that part "imaginary". Normally you can still work it out from how the shadow moves between photos. At exactly 2 photos per turn you can't, not even the direction. 🔍 Im: missing.
For the mathematicians: the clock hand is a complex number eiθ. Its shadow is the real part, its up-down position the imaginary part. A real wave is two such hands turning in opposite directions:
At f = fs/2 the angle per sample is π, so eiπn = (−1)n, so the twins now spin at the same frequency and merge into one: ½A·eiφ + ½A·e−iφ = A·cos φ. Real, and the sine part, A·sin φ, which would have told you the phase, is gone for good. That's why the safe rule says > and not ≥. In the FFT the same thing appears as the lonely middle bin:
The camera's frame rate matches the rotor, so the blades stand still and the chopper "flies" anyway.
Fine stripes plus a TV camera's pixel grid make swirly moiré ghosts. Presenters are told not to wear them.
Humans hear up to about 20 kHz. Times 2 is 40 kHz, plus room for the anti-aliasing filter. Why exactly 44.1? Early digital audio was stored on video tape: 3 samples × 245 lines × 60 fields = 44,100.
In Europe: 3 × 294 lines × 50 fields, the same 44,100.
Pixels are spatial samples. A slanted line becomes a staircase. Games smooth edges by taking extra samples per pixel (anti-aliasing).