Aliasing
A Socratic walk-through of aliasing — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why can a filmed wagon wheel appear to turn backwards, and why does the same trap corrupt digital measurements?
A stagecoach accelerates across the screen and its wheels, at some point, begin to rotate backwards. Nobody filmed a wheel turning backwards. The camera recorded still frames of a wheel turning forwards, and the backwards motion appeared between frames — where nothing was recorded at all.
The companion explanation here on apparent motion ends by naming exactly this gap: the visual system must decide which element of one frame corresponds to which of the next, and it can guess wrong. Let us take up that thread, because the wrong guess is not a quirk of eyes. It sits inside every digital measurement ever taken, and there it is not an illusion but a fabrication no processing can undo.
Reasoning it through
REASONING #Give the wheel twelve spokes and film it at twenty-four frames per second. Twelve spokes means it looks identical after every thirty degrees — one spoke has taken the previous one's place.
In one frame interval a wheel turning at f revolutions per second advances 360*f*/24 degrees, which is 15*f*. At four revolutions per second that is sixty degrees: exactly two spoke spacings, so the wheel sits in a visually identical position in every frame. The film contains a wheel that is not moving — not as an error, but as a faithful record of something that looked the same at every exposure.
Now nudge it to 3.9 revolutions per second. Per frame it advances 58.5 degrees, one and a half short of two spoke spacings, so every spoke sits one and a half degrees behind where a spoke sat before. Nothing in the film says the wheel went forward 58.5 degrees rather than backward 1.5; both fit every frame. And the visual system, built to link each element to its nearest candidate, takes the short reading: creeping backwards.
Notice the rule falling out. Apparent rotation is true rotation minus the nearest multiple of the spoke-repeat rate — here twenty-four frames over twelve spokes, two revolutions per second. Speed the wheel up steadily and it reads true, slow, stopped, reversed, true again, over and over. That is no defect in the camera: there is genuinely no more information in the frames.
Now strip the wheel away, because the structure is the point. Sample any signal at rate fs, so measurements fall at times n/fs, and take a cosine at frequency f: its sampled values are cos(2*pi f n*/fs). Now take a cosine at f plus any whole multiple of fs. Its values are cos(2*pi f n*/fs + 2*pi k n*), and since k and n are whole numbers that extra term is a whole number of turns. The two produce identical numbers — not similar ones, the same ones.
So sampling does not approximate; it collapses an infinite family of frequencies onto one record that cannot distinguish its members. The escape is Nyquist's condition: guarantee that nothing lies at or above half the sampling rate, and exactly one member of each family survives.
What if you cannot guarantee it? Then a frequency above half the rate does not disappear — it arrives below half the rate wearing a false name. Sample at 1000 hertz and a genuine 1200-hertz vibration is recorded as 200 hertz, indistinguishable from a real shaft harmonic. And here is the sting: the corruption happens in the converter, at the instant of sampling. Averaging cannot remove it, nor a longer record, nor a cleverer algorithm, because the distinguishing information was never digitised. The only cure is an analogue filter before the converter.
The analogy
THE ANALOGY #Think of a clock face and a friend who glances at it once an hour, on the hour. The hour hand tells them the time. The second hand tells them nothing — it reads the same at every glance — and if it ran very slightly slow it would appear to creep backwards at a minute or so per hour. Their record of it is not noisy — it is a confident, repeatable measurement of a motion that never happened.
the friend knows a second hand exists and moves faster than they are looking, whereas a data logger does not — the false frequency arrives as an ordinary number with nothing marking it suspect, which is why the filter must be physical rather than a matter of interpretation.
Clarifying the model
THE MODEL #Three refinements, one of them a genuine dispute.
The first is that the folk explanation — "the frame rate is too slow to keep up" — gets the failure wrong. It sounds like blur or missing detail. But every frame is sharp and correct, and the false motion is entirely an artefact of joining correct frames by the nearest-match rule. The samples are not degraded, they are ambiguous, and ambiguity is a different failure mode from imprecision.
The second is that the visual and electronic cases share the mathematics but differ in the deciding rule: a logger picks the lowest member of the family by convention, the eye picks the nearest correspondence between frames. That is why the effect needs a repeating pattern of spokes at all — a featureless disc films perfectly well at any speed.
The third is contested. The wagon-wheel effect also gets reported under continuous daylight, with no camera, no strobe and no sampling anywhere in the physical chain. Purves and colleagues took that as evidence that visual processing is itself discrete; Kline, Holcombe and Eagleman argued it arises from competition between motion detectors tuned to opposite directions, with no sampling involved. That is not settled and I would not lean on either side. The filmed and strobed cases, by contrast, are arithmetic.
On numbers, softest first. The identity between f and f plus k times fs is exact and derived above; it is the load-bearing claim: sampling makes a whole family of signals produce identical records, so the information distinguishing them is destroyed at the sampling instant and must be removed before it, not after. The wagon-wheel figures follow exactly from the twelve spokes and twenty-four frames I chose. Film shutters conventionally expose for about half the frame interval, a figure I am recalling rather than deriving, and that exposure blurs fast spokes — the weakest link here, since whether an audience notices the reversal depends on blur and spoke contrast rather than on the arithmetic.
A picture of it
THE PICTURE #How to readThe only thing changing is the wheel's true speed, rising left to right, and each box is what an observer of the film reports at that speed. Start at the left, where the motion between frames is unambiguous and the film tells the truth. Then follow the transitions as speed rises. The loop from the fourth box back to the second is the essential feature: the cycle repeats at every multiple of the spoke-repeat rate, so there is no speed beyond which the wheel finally reads true again.
What became clearer
WHAT CLEARED #Sampling is not a fine-grained approximation to watching. It is a lossy identification: a family of possible motions is mapped onto one record, and whatever chooses among them is inventing, honestly and inevitably. The wagon wheel is that invention made visible, because we happen to know what a wheel does. In a measurement we do not — which is why a sampling rate is not a quality setting but a promise about what the signal may contain, enforced in hardware before the first number is written down.
Where to go next
ONWARD #- Why an anti-alias filter's imperfect roll-off forces sampling well above twice the highest frequency of interest.
- Why deliberate undersampling is useful when the band of interest is known, as in bandpass sampling.
Key terms
TERMS #| Term | What it means |
|---|---|
| Aliasing | the appearance of a frequency above half the sampling rate as a different, lower frequency in the record. |
| Nyquist frequency | half the sampling rate, the upper limit on signal content for unambiguous reconstruction. |
| Anti-alias filter | an analogue filter before a converter, removing content that would otherwise fold. |
| Correspondence problem | which element in one frame is the same as in the next; wrong answers produce apparent reversal. |
Every term the collection defines is gathered in the glossary.