Does the rule apply to images?

Yes, through spatial sampling and spatial frequency. Two-dimensional image sampling requires attention to direction and the sampling grid.

Can software always remove aliasing later?

No. Distinct inputs can share the same samples. Extra assumptions or information may help, but the ambiguity is not generally reversible.

Is a rate just above twice the limit always practical?

Not necessarily. Filter transitions, timing errors and application requirements can justify additional margin.

AI Glossary / Computer vision and multimodal AI / Intermediate

Nyquist sampling rate

Definition

The Nyquist sampling rate is twice the highest frequency present in a band-limited signal under the usual baseband sampling model. Sampling above this rate allows ideal reconstruction from uniformly spaced samples, subject to the theorem's assumptions. It is a threshold derived from the signal, not the same quantity as a sampler's Nyquist frequency.

What it means

Sampling records values at discrete positions or times. If those samples are too widely spaced, different continuous patterns can produce the same recorded values. This ambiguity is aliasing: a fast variation can appear to be a slower one.

For a signal with highest frequency B, the Nyquist rate is 2B. For a device sampling at rate fs, its Nyquist frequency is fs/2. These related terms are easily confused. The first asks how rapidly a particular signal needs to be sampled; the second identifies a frequency boundary for a chosen sampling rate.

The familiar rule assumes an appropriately band-limited signal and uniform sampling. Special bandpass-sampling arrangements use additional knowledge and conditions; they should not be treated as permission to undersample an arbitrary signal.

How it works

Under ideal conditions, keeping the signal's frequency content strictly below half the sampling frequency prevents overlap between the repeated spectra created by sampling. The original continuous signal can then be reconstructed using the appropriate ideal interpolation.

In practice, acquisition systems use filtering before sampling to reduce unwanted high-frequency content. Digital downsampling similarly needs appropriate filtering before samples are discarded. Real filters have transition regions, so engineers normally allow margin instead of relying on a perfect cutoff at the boundary.

Exactly two samples per cycle can be ambiguous for a sinusoid at the boundary: an unfortunate phase can place every sample at a zero crossing. “Above twice the highest frequency” avoids treating this boundary as a universal guarantee.

The theorem does not promise a noise-free recording, unlimited numerical precision or an accurate instrument. Sampling rate is one part of a complete acquisition design.

Three practical examples

An audio acquisition system limits high-frequency content before conversion so that it does not fold into the recorded band.
An image-resizing pipeline smooths appropriately before reducing pixel density, helping avoid false patterns in fine textures.
A vibration-monitoring experiment selects a sampling rate and input filter together, using the highest relevant frequency content rather than the average speed of a machine.

Think of it this way

Think of observing a rotating patterned wheel through regularly timed glimpses. Several different rotation rates may place the pattern at the same observed positions. Suitable sampling can reduce ambiguity.

The analogy illustrates aliasing, but a wheel is not the complete sampling theorem. Reconstructing an arbitrary signal requires the stated bandwidth and sampling assumptions; simply taking “a few more pictures” is not a mathematical guarantee.

Worked example

Finding an aliased frequency

Suppose an ideal cosine signal oscillates at 120 Hz and is sampled at 200 samples per second. The sampler's Nyquist frequency is 100 Hz, so the 120 Hz component lies above it. At integer sample index n, the samples are cos(2π × 120n/200).

These are identical to samples from an 80 Hz cosine because 120/200 = 1 − 80/200, and cosine has the relevant periodicity and symmetry at integer n. A record interpreted only within the lower frequency band can therefore show 80 Hz instead of the original 120 Hz.

If the sampling rate becomes 300 samples per second, the Nyquist frequency becomes 150 Hz. The 120 Hz component is now below that boundary. This satisfies the condition for that component, but other higher-frequency inputs may still need attention.

For another boundary check, a sine wave at 100 Hz sampled at 200 samples per second can give sin(πn) = 0 at every sample. The disappearing signal explains why an exact two-samples-per-cycle slogan needs qualification. These calculations assume ideal sinusoids and timing.

Illustrative case study

Illustrative case study - fictional

A fictional equipment team records a periodic vibration and sees a prominent 80 Hz component. Before changing its maintenance plan, an engineer checks the acquisition settings. A controlled test shows that a 120 Hz input produces the same apparent frequency at the selected 200-sample-per-second rate.

The team repeats the measurement with suitable acquisition bandwidth, filtering and a higher sampling rate. It documents the changed setup so later measurements remain comparable.

This illustrates an acquisition error, not a machinery diagnosis. Interpret frequencies in light of how the data was collected.

Common misconceptions and mistakes

Nyquist rate and Nyquist frequency are not interchangeable: one is twice the signal limit, while the other is half the sampling rate.
The relevant limit is not merely the frequency of the feature you hope to measure. Other input components can alias into that band.
Filtering after aliasing has occurred cannot generally recover which original frequency produced the samples without additional information.
More samples do not automatically improve sensor accuracy, bit depth or noise performance.
The ideal theorem concerns reconstruction, not drawing straight lines between a sparse set of samples and expecting a visually smooth waveform.

Business use

Businesses encounter this principle in audio, cameras, industrial sensing and data reduction. It helps explain why a fast-looking digital record can still misrepresent the physical input.

When evaluating a system, ask for the acquisition rate, usable input bandwidth and anti-alias filtering behavior together. For images, test fine patterns at the intended output resolution. For sensors, include plausible unwanted frequency content in validation. Decisions based on apparent trends should not ignore the possibility that the sampling process created those trends.

Check your understanding

A device samples at 1,000 samples per second. Is its Nyquist frequency 2,000 Hz?

Answer: No. It is 500 Hz. A signal limited to 200 Hz has a Nyquist rate of 400 samples per second under the baseband model.

Related terms

Aliasing (factorial experiments)
Spatial frequency
aliasing
anti-aliasing filter
sampling frequency
downsampling
signal bandwidth

Learn next

Calculate sampled values for the 120 Hz and 80 Hz cosine examples at 200 samples per second. Compare them, then change the rate. Keep the signal's frequency limit and the sampler's half-rate boundary labeled separately.

Sources

Sources: D. Rowell, MIT 2.161 Lecture 10, Sampling Theorem: https://ocw.mit.edu/courses/2-161-signal-processing-continuous-and-discrete-fall-2008/2934569026a399e9062597924c730a78_lecture_10.pdf ; NI, Anti-Aliasing Filters and Their Usage Explained: https://www.ni.com/en/shop/data-acquisition/measurement-fundamentals/analog-fundamentals/anti-aliasing-filters-and-their-usage-explained..html . Relevant theorem and filtering explanations were reviewed.

Previous: NYQUIST PLOT Next: Nyquist - Shannon sampling theorem