Sampling does not merely make a signal discrete. It makes frequencies separated by integer multiples of the sampling rate observationally identical.
The ambiguity in the dots
For a sinusoid sampled at rate fₛ, the samples of frequencies f and f + m fₛ coincide for every integer m. Restricting a real sinusoid to the interval from zero to fₛ/2 chooses one representative from each equivalence class. That boundary is the Nyquist frequency.
Interactive experiment
Frequency folding
When does a sampled sinusoid become indistinguishable from a slower one?
Blue is the source, dashed amber is the apparent sinusoid, and circles are samples shared by both. The display covers two seconds.
Why folding appears
Increase a tone towards Nyquist and its samples trace a faster oscillation. Cross the boundary and the same dots can be joined by a waveform whose frequency decreases. The sampled data cannot decide which continuous path was present. Reconstruction requires an assumption that the input was band-limited before sampling.
Anti-aliasing happens before decimation
A low-pass filter removes content that the lower rate cannot represent; then decimation discards samples. Reversing those operations is too late because high and low frequencies have already become inseparable. Real filters have transition bands, so engineering a sample-rate conversion also means choosing acceptable pass-band loss and stop-band leakage.
The machine-learning connection
Image resize, temporal subsampling, pooling and strided convolution all reduce a discrete grid. If appreciable content lies above the new grid's Nyquist limits, it folds. A learned model may exploit or partially compensate for these artefacts, and nonlinearities can create new frequencies, so a neural network is not simply a textbook linear sampler. Still, inspectable low-pass filtering before stride is a useful design hypothesis—not a guarantee of better task performance.
Two dimensions complicate the picture
An image has a two-dimensional spectrum. Fine stripes can become coarse moiré patterns when pixel centres sample them; diagonal structure can fold in both axes. Camera optics, sensor pixels, demosaicing and resizing all participate in the effective prefilter. “The image resolution” is therefore not one isolated sampling event.
Practical implications
- Filter before reducing rate, not afterwards.
- Record the exact resize and resampling conventions in data pipelines.
- Test model behaviour under sub-pixel translations; phase-sensitive aliases can change sharply.
- Do not mistake a smooth plotted interpolation for recovered information.
A 7 Hz signal sampled at 10 Hz
The source advances by 0.7 cycles between samples. A −3 Hz sinusoid advances by −0.3 cycles, which reaches the same phase after adding one complete turn. They therefore produce the same samples. The apparent frequency magnitude is 3 Hz, but the signed frequency preserves the phase direction: a zero-phase positive 3 Hz sine would have the opposite sign.
In the explorer, start at 3 Hz with a 10 Hz sampling rate, then move to 7 Hz. Compare the sample values, not just the smooth curves. At exactly 5 Hz, a zero-phase sine lands on zero at every sample; shifting the phase changes what is observed. The Nyquist boundary is a delicate endpoint, not a safe target for arbitrary signal recovery.
Why two samples per cycle is not an engineering recipe
The ideal reconstruction result assumes an infinite sequence, an exactly band-limited source and an ideal reconstruction filter. Acquisition systems have finite records, noise and filters with finite transition bands. Leave room between the highest frequency you need to preserve and the Nyquist boundary so the analogue filter can attenuate unwanted content before the converter.
For an already sampled signal, reducing rate by a factor of two halves the representable band. A moving average may reduce some high-frequency content, but its stop band is not an ideal brick wall. Select and verify the filter against the tolerated alias level.
A useful diagnostic for images and models
Imagine alternating bright and dark pixels. Keeping every second pixel can produce an entirely bright or entirely dark image depending on the starting offset. Averaging each adjacent pair first produces a uniform middle value. The coarse grid cannot retain the alternating pattern; filtering makes the discarded detail less dependent on sampling phase.
For a vision model, repeat an input with small translations and compare outputs before assuming semantic change. For a time-series model, record the sensor rate and every decimation stage: a later Fourier transform cannot identify which of two aliased physical frequencies was originally present.
Explore the frequency representation to distinguish folding caused by sampling from leakage caused by a finite observation. Both spread or relocate apparent spectral structure, but they arise at different stages.