The Nyquist rate is the minimum sampling frequency required to represent a bandlimited signal without aliasing: it equals twice the highest frequency component present in the signal. Sampling below this rate causes higher-frequency content to fold back and corrupt lower-frequency content in an unrecoverable way.
In practice
In ADC-based data acquisition, the Nyquist rate sets the lower bound on your sample rate. If your signal contains energy up to 10 kHz, you must sample at more than 20 kHz to avoid aliasing. In practice, real anti-aliasing filters have finite roll-off slopes, so most designs sample at 2.5x to 10x the maximum signal frequency of interest and rely on the filter to attenuate content above the Nyquist frequency (half the sample rate) before the signal reaches the ADC.
The term "Nyquist frequency" is sometimes used interchangeably with "Nyquist rate," but many engineers reserve "Nyquist frequency" to mean half the sample rate (fs/2), i.e., the highest frequency that can be represented at a given sample rate, while "Nyquist rate" refers to the minimum sample rate required for a given signal bandwidth. The distinction matters when reading datasheets and DSP literature.
Bandpass sampling is an important exception to the naive rule. A bandlimited signal centered at a high carrier frequency does not always need to be sampled at twice the carrier frequency; it only needs to be sampled at twice the signal's bandwidth, provided aliasing is managed deliberately. This technique is used in software-defined radio and IF digitization on resource-constrained embedded platforms. The EmbeddedRelated post "Sampling bandpass signals" covers the conditions and constraints in detail.
On embedded targets running fixed-point DSP (Cortex-M4/M7 with DSP extensions, C2000, Blackfin, etc.), the Nyquist rate governs filter design choices as well as sample rate. Violating it in firmware -- for example, by skipping ADC samples under interrupt load or using a loop-based delay instead of a hardware timer -- introduces effective sample rate jitter that can create spectral artifacts even when the nominal rate appears sufficient.
Learn this in DSP Foundations
Discussed on DSPRelated
Frequently asked
What is the difference between the Nyquist rate and the Nyquist frequency?
The Nyquist rate is a property of the signal: it is twice the signal's highest frequency component, and it is the minimum
sample rate you need. The Nyquist frequency (also called the folding frequency) is a property of your sampling system: it is half your actual sample rate (fs/2), and it is the highest frequency your system can represent without
aliasing. For example, sampling at 48 kHz gives a Nyquist frequency of 24 kHz; a signal with content up to 20 kHz has a Nyquist rate of 40 kHz. The two values coincide only when you are sampling at exactly the minimum rate, which is rarely done in practice.
Why do real systems sample well above the Nyquist rate instead of right at it?
The Nyquist theorem assumes a perfectly bandlimited signal and an ideal brick-wall anti-aliasing filter. Neither exists in hardware. Real analog filters have gradual roll-off, so some energy exists above the intended cutoff. Sampling at 2x to 10x the signal
bandwidth (
oversampling) moves the
aliasing problem to higher frequencies where a gentler, cheaper analog filter can suppress the out-of-band content before it reaches the
ADC. Oversampling also improves effective resolution in sigma-delta ADCs and simplifies reconstruction filter requirements in
DAC systems.
Does the Nyquist rate apply to bandpass signals in the same way as baseband signals?
Not always. A bandpass signal occupying a
bandwidth B centered at a high carrier frequency theoretically requires a
sample rate of only 2B -- not twice the carrier frequency -- if the sampling aliases are managed carefully. This is bandpass or IF sampling. The conditions are specific: the band must fit within one Nyquist zone without straddling a zone boundary. In practice this technique is used in IF digitization and
SDR front ends. The EmbeddedRelated post 'Sampling bandpass signals' covers the zone conditions and practical constraints.
What happens in firmware if I accidentally sample below the Nyquist rate?
Frequencies above
fs/2 fold back (alias) into the
baseband spectrum and are indistinguishable from legitimate lower-frequency content. The effect is permanent -- you cannot remove aliases in post-processing because the information needed to separate them is gone. Common firmware causes include using a software delay loop instead of a hardware timer for sample pacing, losing samples due to interrupt latency or DMA misconfiguration, or setting a peripheral clock divider incorrectly. Always use a hardware timer or a peripheral with a built-in sample rate clock to guarantee consistent sample intervals.
Is the Nyquist rate the same as the Nyquist theorem or the Shannon sampling theorem?
These names refer to the same foundational result. The theorem is variously attributed to Nyquist (1928, related work on telegraph signaling) and Shannon (1949, formal proof for communication systems), and is sometimes called the Nyquist-Shannon sampling theorem. In DSP and embedded literature you will see all three names. The core statement is always the same: a bandlimited signal with maximum frequency fmax can be exactly reconstructed from samples taken at a rate of at least 2*fmax. The EmbeddedRelated post 'The Sampling Theorem - An Intuitive Approach' provides an accessible derivation.
Differentiators vs similar concepts
Nyquist rate (minimum required
sample rate for a given signal, = 2 * fmax) is frequently confused with Nyquist frequency (the maximum representable frequency in a given sampled system, = fs/2). The two are reciprocal perspectives on the same relationship but are properties of different things: the signal versus the sampling system. Additionally, "Nyquist rate" is sometimes conflated with "Nyquist
bandwidth" in communications contexts, where it refers to the maximum symbol rate (in baud) that a channel of bandwidth B can carry without intersymbol interference (= 2B). That communications usage is distinct from the
ADC/DSP sampling usage described here.
