The Fast Fourier Transform & Spectral Methods

the Nyquist frequency

/ NYE-kwist /

If you check a clock only once every hour, you can tell the hour hand is moving but you have no hope of following the second hand — it spins many times between your glances. Sampling has the same ceiling: the rate at which you look sets a hard limit on the fastest wiggle you can possibly track. The Nyquist frequency is that ceiling — the highest frequency a given sampling rate can honestly represent.

If you sample at f_s samples per second, the Nyquist frequency is exactly f_s / 2. The reason is intuitive: to capture an oscillation you need at least two samples per cycle — one to catch a peak and one a trough — so the fastest cycle you can pin down repeats every two samples, which is f_s / 2 cycles per second. In a length-N DFT this corresponds to the middle bin, index N/2, the fastest pattern the data can represent (alternating +1, -1, +1, -1...). Every real frequency from 0 up to f_s / 2 maps to its own DFT bin; anything faster has nowhere distinct to go and instead folds back below f_s / 2 as an alias, which is why the Nyquist frequency is also called the folding frequency.

The Nyquist frequency is the practical design number behind every digitizer. Audio sampled at 44.1 kHz has a Nyquist limit of 22.05 kHz, just above human hearing; a medical sensor sampling at 1000 Hz can faithfully report rhythms only up to 500 Hz. The companion rule is the Nyquist RATE — the minimum SAMPLING rate (twice the highest frequency present) you must use to avoid aliasing — which is the same number seen from the other side. Engineers usually sample comfortably above it and add an anti-aliasing filter, because a signal with content right at Nyquist is borderline and easily corrupted.

A CD samples at 44,100 Hz, so its Nyquist frequency is 22,050 Hz — chosen to sit just above the roughly 20,000 Hz upper limit of human hearing. Frequencies below 22,050 Hz are captured cleanly; anything above is removed by an anti-aliasing filter before sampling so it cannot fold down into the audible band.

Half the sampling rate is the hard ceiling on representable frequency.

Sampling at exactly twice a frequency is not enough in practice: a sine sampled precisely at its zero crossings reads as all zeros. The sampling theorem requires the rate to be STRICTLY greater than twice the highest frequency, with margin to spare.

Also called
folding frequencyhalf the sampling rate奈奎斯特極限折疊頻率