Saramsha

No. 022  ·  8 October 2026

The Wheel That Runs Backward

The wheel on the left never changes speed — it is always turning exactly as fast as the slider says. The one on the right is the same wheel, but you are only allowed to glimpse it at fixed moments, like a camera's shutter. Tune the two speeds against each other and watch the right-hand wheel lie to you.

the true motion a sample — a single glimpse the ghost frequency fooling you

About this piece

A camera does not watch. It glances.

A movie camera does not record motion continuously. It opens its shutter, takes a still picture, closes it, waits, and takes another — traditionally twenty-four times a second. Everything that happened to the wheel between two of those glances is simply gone; the film only ever contains the twenty-four frozen moments, never the journey between them. So when you watch the film, your eye is not shown the wheel's motion at all. It is shown a sequence of still positions, and your visual system — which is very good at assuming the simplest explanation for what it sees — reconstructs some motion that would produce exactly that sequence of positions. Usually that reconstruction matches reality. Sometimes, when the wheel turns fast enough relative to the shutter, it does not, and the simplest explanation your eye finds is a lie.

This is exactly what the right-hand wheel above is doing to you. It is not slowed down, blurred, or faked — its angle really is recalculated fresh at every sample tick, from the same true rotation as the left-hand wheel, and held steady in between. The illusion is produced entirely by throwing away everything that happened between ticks, which is precisely what a real shutter, a real strobe light, and a real digital sample-and-hold circuit all do as a matter of course.

Why a slower-looking wheel can be exactly as true as a fast one

Here is the uncomfortable fact underneath all of this: a set of still positions, sampled at a fixed rate, cannot tell you which of several different speeds produced them. A point that moves through one full turn between two samples looks, in those two samples, identical to a point that stood still, or one that went all the way around twice, or one that crept a little short of a full turn backward. Any of those motions puts the point in the same place at both sampling instants. Given just the samples, there is no way to prefer one story over the others — except that your eye (and most reconstruction algorithms) always picks the slowest, simplest motion that fits. That slowest-fitting motion is called the alias, and it is what you perceive, correctly computed and entirely wrong about reality.

The graph beneath the wheels makes this concrete for a single point going round. The light curve is the true motion — a sine wave at the true frequency. The solid dots are where it was actually sampled. The dashed curve is the alias: the slowest sine wave that passes through those exact same dots. Both curves are completely honest about where the dots are. Only one of them is honest about what happened between the dots.

The rule, and the line you must not cross

Harry Nyquist worked out the boundary at Bell Labs in 1928, and Claude Shannon turned it into the theorem that carries both names in 1949: if you sample at rate fs, you can faithfully recover any motion whose own frequency is below fs/2 — the Nyquist limit — and nothing faster. Above that line, the true frequency and some slower impostor produce identical samples, and the samples alone can never tell you which one actually happened. The exact impostor frequency is the true frequency reflected around the nearest multiple of fs, which is why pushing the true speed up past the limit does not make the apparent motion faster — it makes it fold back and come down again, then stop, then reverse, then speed back up the wrong way, over and over, as you keep turning the "really turning" slider. Try it slowly and watch the readout track each fold.

Why spokes make it so much worse

A single marked point has nowhere to hide: it aliases only once it genuinely passes the Nyquist limit of its own rotation. A wheel with N identical, unmarked spokes is a different problem, because the wheel does not have to complete a full turn to look unchanged — it only has to turn 1/N of a turn, since every spoke is standing in for every other one. For the purposes of this illusion, an N-spoked wheel behaves exactly like a single point rotating N times faster, so its effective frequency is the true rotation rate multiplied by N. That is why the default above — three turns a second, eight identical spokes, a camera at twenty-four frames a second — lands exactly on a multiple of the frame rate (3 × 8 = 24) and freezes stone still, despite the wheel genuinely spinning three times a second the whole time. This is the real reason stagecoach wheels famously "stopped" in old Westerns, and the same reason helicopter rotor blades on the news sometimes appear to hang motionless in mid-air.

Where else this shows up

Anywhere a continuous signal is cut into discrete snapshots, the same trap is waiting. A digital camera's rolling shutter does it to spinning propellers. Standard video, sampled at its frame rate, does it to any rotating or oscillating thing fast enough — fan blades, car wheels under the strobing of fluorescent light, the stepper motor in disassembled hardware. Audio works identically: a microphone's signal is sampled at a fixed rate by a digital recorder, and any true pitch above half that sample rate folds down into a false, lower pitch that was never played — which is precisely why every digital audio converter has an anti-aliasing filter wired in front of it, quietly removing anything above the Nyquist limit before sampling, so it has no chance to masquerade as something lower. CD audio samples at 44.1 kHz for exactly this reason: twice the roughly 20 kHz edge of human hearing, with headroom for the filter to work.

Everything here runs from one formula: fold a frequency x into the band ± fs/2 by subtracting the nearest multiple of fs. The wheels use that fold on the true rate times the spoke count; the graph below uses it directly on the true rate. The right-hand wheel's displayed angle is not computed from this formula at all — it is the literal true angle at the most recent sample instant, held fixed until the next one, exactly as a real shutter would produce it; the formula only predicts, and then explains, what you end up seeing. ?f=, ?fs= and ?n= set the three controls directly. Space pauses.