Feed a 25 kHz tone into a digital recorder running at 44.1 kHz with the input filter bypassed, and the recording will play back a tone at 19.1 kHz. The number is not a glitch and not a rounding error; it is exactly what the arithmetic predicts, since 44.1 minus 25 equals 19.1. A frequency the system cannot represent has not been lost or attenuated. It has been relabelled as a different frequency, one the system can represent, and it is now indistinguishable from a tone that was genuinely there.
The Nyquist-Shannon sampling theorem is the result that tells you exactly when this cannot happen. In its precise form: if a signal x(t) contains no frequency components at or above B hertz, then it is completely determined by its samples taken at a uniform rate f_s strictly greater than 2B samples per second — and “completely determined” means the continuous signal can be reconstructed from the discrete samples with zero error, not approximately. Everything below is an unpacking of that sentence: why the factor of two, why the inequality is strict, what the reconstruction actually is, and what happens in the enormous gap between the theorem’s hypothesis and any real recording.
The statement, and why the inequality is strict
Write for the Fourier transform of . Say that is band-limited to if for all with . Then:
If is band-limited to and , the samples for determine uniquely, and there is an explicit formula recovering for every real .
The quantity is called the Nyquist rate of the signal; the quantity is called the Nyquist frequency of the sampler. These are different things, and the literature confuses them constantly.
Many textbooks write the condition as . The strict inequality is the correct form, and the reason is a one-line counterexample. Take
whose spectrum consists of two impulses, at and . Sample it at exactly , so the sample instants are . Then
Every single sample is zero. The sampled sequence is identical to the sampled sequence of the zero signal, so no reconstruction procedure of any kind can distinguish from silence. At the theorem is false.
The fix is in the hypothesis, not the rate. Define band-limited so that the band edge itself is excluded — require to vanish on a neighbourhood of — and no longer qualifies, so one can then prove the theorem with . The two conventions are consistent, but only if you track which one you are using. Textbooks that write while allowing spectral content at are stating something false, and the sine above is the witness. For practical purposes this is a technicality — real signals have no isolated spectral lines sitting precisely at the band edge — but it is the kind of technicality that distinguishes a theorem from a slogan.
The reconstruction formula
The theorem is not merely an existence claim. The reconstruction is explicit:
This is the Whittaker-Shannon interpolation formula, or in the older mathematical literature the cardinal series. Three things about it deserve emphasis.
First, it is exact. This is not a curve fitted through the sample points in the spirit of a spline or a least-squares approximation; under the hypothesis, the right-hand side equals the left-hand side at every real , not just at the sample instants.
Second, it is a specific interpolation — not linear, not cubic, not polynomial. The kernel equals 1 at and exactly 0 at every other integer, so each term contributes its sample value at its own instant and nothing at the others; and, crucially, its Fourier transform is the indicator function of the band. Sinc interpolation is the unique interpolation that commits to the band-limitation hypothesis.
Third, it needs infinitely many samples, and the kernel is badly behaved in two ways that matter. It is non-causal: is nonzero for well before , so reconstructing the signal at any instant requires samples from the future. And it decays only as . Both properties are consequences of the ideal rectangular spectrum: a spectrum with a perfectly sharp edge has an impulse response with a heavy tail.
The decay has a quantifiable cost. Truncate the series to the nearest samples. Since , the discarded terms are bounded by roughly — and diverges. Summing numerically for a representative between samples, the tail out to comes to about , matching : the bound grows logarithmically without limit, so truncation error cannot be controlled by peak amplitude alone. For finite-energy signals it is better but still slow. Cauchy-Schwarz bounds the tail by the signal energy times , numerically at , matching the asymptotic . Truncation error decays only like — which is why practical reconstruction filters do not truncate the ideal sinc at all, but design a different, faster-decaying kernel with a controlled transition band. Every digital-to-analogue converter approximates a formula that cannot be implemented.
Why twice: the argument in the frequency domain
The factor of two looks arbitrary until you look at sampling in the Fourier domain, where it becomes almost visual. (This section assumes the Fourier transform; for the groundwork see our articles on Fourier series and the fast Fourier transform.)
Model ideal sampling as multiplication by a Dirac comb of spacing :
The comb has the remarkable property of being its own transform type: the Fourier transform of a comb of spacing is a comb of spacing , scaled by . And multiplication in the time domain is convolution in the frequency domain. Therefore
Convolving with a shifted delta is just translation, so the spectrum of the sampled signal is an infinite sum of copies of the original spectrum, one centred at every integer multiple of the sampling rate. This single identity is the whole theorem.
Now count widths. If is supported in the open interval , each copy occupies a window of width centred at . The copies are spaced apart. Adjacent copies are disjoint precisely when the spacing exceeds the width:
That is the factor of two, and it is not a safety margin. It is the ratio between the two-sided bandwidth , which is what the spectrum actually occupies once you account for negative frequencies, and the one-sided bandwidth , which is what engineers quote. A real-valued signal has a Hermitian spectrum, , so the negative-frequency half is redundant as information but entirely real as spectral occupancy, and it is the occupancy that collides.
When the copies are disjoint, the baseband copy can be isolated by an ideal low-pass filter with cutoff — and multiplying by a rectangle in frequency is convolving with a sinc in time, which is exactly the interpolation formula above. The two arguments are one argument in two coordinate systems.
When , the copies overlap. Content from the copy centred at lands on top of baseband content and the sum is recorded as a single number per frequency. There is no inverse: two different inputs have produced the same output. That overlap is aliasing, and the lost distinction is gone for good.
Aliasing, computed
The bookkeeping of where a frequency lands is simple. A pure component at frequency , sampled at , is indistinguishable from a component at
the distance from to the nearest multiple of . Some worked values:
| True frequency | Sampling rate | Nearest multiple | Apparent frequency |
|---|---|---|---|
| 5 Hz | 4 Hz | 4 Hz | 1 Hz |
| 25 kHz | 44.1 kHz | 44.1 kHz | 19.1 kHz |
| 50 kHz | 44.1 kHz | 44.1 kHz | 5.9 kHz |
| 30 kHz | 48 kHz | 48 kHz | 18 kHz |
| 1 kHz | 44.1 kHz | 0 | 1 kHz (no aliasing) |
The 50 kHz case is the alarming one: an ultrasonic component far above anything audible, well outside the band the system claims to handle, reappears at 5.9 kHz, squarely in the most sensitive region of human hearing. This is why the anti-alias filter is not optional.
The first row is the clearest case to see directly, because the two functions do not merely produce similar samples — they produce identical ones. Take Hz, so the sample instants are . Then
since , an exact multiple of the period. The two sinusoids agree at every sampling instant, for every integer , exactly and not approximately. Evaluated numerically at , both give , with the largest discrepancy across all nine instants equal to — floating-point noise, nothing more.
Figure 1 — A 1 Hz sinusoid (blue) and a 5 Hz sinusoid (red) take exactly the same value at every instant t = n/4. The nine black dots lie on both curves simultaneously. From the samples alone the two signals are not merely hard to tell apart; they are mathematically identical data.
The wagon wheel. The same arithmetic explains the stagecoach wheels that spin backwards in old films. Cinema runs at 24 frames per second, and a wheel with 12 identical spokes looks the same after a rotation of one twelfth of a turn. So the visual signal has a period of revolution, and the frame rate samples it 24 times a second. The wheel appears stationary whenever the rotation between frames is an exact whole number of spoke-spacings — that is, when the rotation rate is
So a 12-spoke wheel turning at exactly 2 rev/s, which is 120 rpm, appears frozen. For a wheel 0.9 m in diameter that corresponds to a ground speed of m/s, about 20.4 km/h — a plausible trotting pace, which is why the effect is so common in westerns.
Just below that speed the wheel reverses. At 1.9 rev/s each frame advances the wheel by revolutions, which is spoke-spacings: the nearest spoke falls of a spacing short of where the previous spoke was, so the eye reads a backward step, an apparent rev/s. At 2.1 rev/s the same calculation gives rev/s. The wheel is a sampled signal, the frame rate is , and the backwards motion is the folded frequency.
Nothing real is both time-limited and band-limited
Here is the uncomfortable fact underneath all applications: the theorem’s hypothesis is never exactly satisfied by any real recording.
A non-zero function compactly supported in time cannot be band-limited, and vice versa. If vanishes outside , then
is an integral over a bounded interval, so the integrand is entire in the complex variable and the integral converges locally uniformly. Hence extends to an entire function on — in fact one of exponential type at most , which is the content of the Paley-Wiener theorem (Paley and Wiener’s 1934 American Mathematical Society colloquium volume Fourier Transforms in the Complex Domain; for the version, Rudin’s Real and Complex Analysis, Theorem 19.3). Now apply the identity theorem: an entire function vanishing on a set with an accumulation point vanishes identically. A time-limited signal vanishes on an entire half-line. So .
Contrapositive: any signal that starts and stops — every recording ever made — has a spectrum nonzero at arbitrarily high frequencies. There is no for which the hypothesis holds exactly. For the analytic machinery here, see complex analysis.
The practical response is engineering rather than mathematics. Before the converter, a real system applies an anti-alias filter, with stopband attenuation chosen so that whatever energy survives above folds back below the converter’s noise floor. For a 16-bit converter that floor sits around dB relative to full scale, so roughly 100 dB of stopband rejection makes the aliased residue irrelevant in practice even though it is not zero. The theorem’s exact hypothesis is traded for a quantified approximation to it, and the whole of digital audio rests on that trade.
Why 44.1 kHz, concretely
Human hearing extends to roughly 20 kHz in young listeners, which sets a Nyquist rate of 40 kHz. Any rate above that will do; 44.1 kHz leaves kHz of transition band between the top of the audio band and the Nyquist frequency, which is what the anti-alias filter needs in order to fall from passband to deep stopband without an impractically steep slope.
But 44.1 is a strange-looking number, and the reason for the specific figure is historical rather than psychoacoustic. In the late 1970s there was no mass storage medium capable of holding a stereo digital audio stream, and the pragmatic solution was the PCM adaptor: a device that encoded audio samples as black-and-white pixel patterns and recorded them onto ordinary video tape, which was already a mature, high-bandwidth consumer and professional medium. Sony’s PCM-1 (1977) and the professional PCM-1600 (1979) worked this way.
Once you are storing samples as video lines, the sampling rate is dictated by the video format. In the 60 Hz (NTSC-style) format, blanking leaves 490 usable lines per frame, or 245 per field, at 60 fields per second; storing 3 samples per line per channel gives
In the 50 Hz (PAL-style) format, 294 usable lines per field at 50 fields per second with the same 3 samples per line gives
the same number. That coincidence is the point: 44.1 kHz was the highest rate comfortably compatible with both video standards at no more than three samples per line, so a single figure served studios on both sides of the standards divide. When the Compact Disc specification was written, it inherited the rate from the mastering equipment. (A caveat often omitted: colour NTSC actually runs at 59.94 fields per second, which gives 44,056 Hz, and both figures circulated in early equipment.)
The broadcast and film world standardised instead on 48 kHz, which divides cleanly against video frame rates and gives a more generous transition band. Higher rates — 96 kHz and 192 kHz — remain contested: advocates cite gentler filters, reduced pre-ringing and headroom for processing, sceptics that the extra bandwidth is inaudible by construction. The argument is about filter design and workflow, not about the sampling theorem, which has nothing to say in either direction.
Who actually proved it
The attribution is genuinely tangled, and the usual potted version is wrong in a specific way.
E. T. Whittaker (1915) introduced the cardinal series — the sinc interpolation formula — in a paper on interpolation theory, asking which function through a given set of equally spaced ordinates is freest of rapid oscillation. He was not doing signal processing and made no reference to communications. Butzer and co-authors have argued that attributing the sampling theorem to him, as opposed to the interpolation formula, is therefore misleading. The name “cardinal series” predates every engineering version of the result.
Harry Nyquist (1928), in Certain Topics in Telegraph Transmission Theory (Transactions of the AIEE, vol. 47, pp. 617–644), showed that a channel of bandwidth can carry at most independently resolvable pulses per second. This is the most commonly misattributed item in the whole story: Nyquist did not state the sampling theorem. He did not discuss sampling a continuous signal, and he did not give a reconstruction formula. What he established was a limit on signalling rate, which is a closely related but logically distinct statement. The factor of two matches, which is why his name became attached.
Vladimir Kotelnikov (1933) gave a rigorous statement and proof in a paper submitted to a Soviet conference on communications — in the modern form, including reconstruction. His work was essentially unknown in the West for decades.
Herbert Raabe (1939) derived the theorem in his doctoral thesis, with the practically important refinement of treating sampling pulses of finite duration rather than ideal impulses. The sampling condition is sometimes called the Raabe condition in German-language sources.
Isao Someya (1949) published the theorem in the Japanese book Signal Transmission, independently; it is occasionally called Someya’s theorem in Japanese literature.
Claude Shannon (1949), in Communication in the Presence of Noise, stated the theorem in the form everyone now uses, placed it inside a general theory of communication, and explicitly cited Whittaker. Shannon himself was careful, remarking that the result was “common knowledge in the communication art” but had not been stated explicitly in the communication-theory literature. The standard historical account is Hans Dieter Lüke’s “The Origins of the Sampling Theorem” (IEEE Communications Magazine, April 1999), which traces how practitioners, theoreticians and mathematicians arrived at it largely independently.
The honest summary: Whittaker wrote the formula without the application, Nyquist found the factor of two in a different problem, Kotelnikov and Raabe and Someya proved the theorem independently in three countries, and Shannon gave it the statement and context that made it canonical. For Shannon’s broader framework see information theory and entropy.
Sampling below the Nyquist rate, legitimately
Several techniques sample at rates below twice the highest frequency present. None of them breaks the theorem; each replaces the band-limitation hypothesis with a different assumption, and the new assumption does the work.
Bandpass sampling (undersampling). Suppose the signal occupies a band from to of width , positioned far from zero. The spectrum is mostly empty, so the copies produced by sampling can be made to interleave in the gaps rather than collide. The condition is that for some positive integer ,
with recovering ordinary baseband sampling. A valid zone exists only when the lower bound does not exceed the upper, which bounds . Worked example: for a band from 90 Hz to 100 Hz, gives the zone Hz, and the largest admissible is 10, where the inequalities collapse to the single point Hz — exactly . One may sample a 100 Hz signal at 20 Hz. For the FM broadcast band, 88 to 108 MHz, the condition admits with MHz, against a naive requirement of 216 MHz: a fivefold reduction. Note that rates above can still fail if they fall between zones; the condition is not a simple threshold.
Compressed sensing. Replace bandwidth with sparsity: assume the signal has few significant coefficients in some basis, measure it with projections incoherent with that basis, and reconstruct by minimisation rather than linear interpolation. The number of measurements scales with the sparsity rather than the bandwidth. This is a large subject with its own guarantees; see our full treatment in compressed sensing.
Finite rate of innovation. Vetterli, Marziliano and Blu (“Sampling Signals with Finite Rate of Innovation”, IEEE Transactions on Signal Processing, vol. 50, no. 6, 2002, pp. 1417–1428) considered signals with a limited number of free parameters per unit time — streams of Diracs, piecewise polynomials, non-uniform splines. These are emphatically not band-limited, yet sampling with a suitable kernel at or above the rate of innovation, then recovering the parameters with an annihilating-filter technique borrowed from spectral estimation, gives perfect reconstruction. The hypothesis has shifted from “narrow band” to “few degrees of freedom”.
The pattern is always the same: a stronger assumption about the signal class, paired with a more elaborate reconstruction, buys a lower rate. The classical theorem remains exactly true on its own terms, which assume nothing except bandwidth.
Beyond audio
Images. A camera sensor samples a two-dimensional field, and the same folding occurs in two dimensions. Fine periodic detail — a striped shirt, a brick wall, a halftone print — aliases into coarse low-frequency patterns, the familiar moiré. Cameras have historically included an optical low-pass filter, a thin birefringent layer that deliberately blurs the image by about one pixel pitch; it is an analogue anti-alias filter, trading a little sharpness for the absence of artefacts no post-processing can remove.
Magnetic resonance imaging. An MRI scanner samples the Fourier transform of the body — k-space — directly, so the sampling grid lives in the frequency domain and aliasing shows up in the image as wrap-around: anatomy outside the field of view appears folded onto the opposite side. The field of view is set by the k-space sample spacing, which is the sampling theorem read in reverse.
Time series generally. Monthly economic data cannot resolve weekly cycles; a sensor logging once an hour cannot see a twenty-minute oscillation and will report it at some spurious slow period. Whenever a quantity is recorded at a fixed interval, the folding formula applies. For related ideas about resolution at multiple scales, see wavelets and multiresolution analysis.
What the hypothesis buys you
It is tempting to read the sampling theorem as a budget: this much data is required. That reading is wrong in both directions. It is not a lower bound on data — bandpass sampling, compressed sensing and finite-rate-of-innovation methods all do better under other assumptions — and it is not really about data volume at all.
What the theorem measures is the value of a hypothesis. Band-limitation is extraordinarily strong: it says the signal lies in a space of entire functions of exponential type, a space so rigid that a member is determined by its values on a discrete set, in the same spirit that an analytic function is determined by its germ at a point. A signal band-limited to and observed over duration has about degrees of freedom — the time-bandwidth product — and the sampling theorem says that uniform samples at a rate just above read off precisely those and no more.
Seen this way, the strict inequality, the infinite sinc tails, and the impossibility of simultaneous time- and band-limitation stop being footnotes and become the same fact viewed from three sides. The assumption is strong enough to make exact reconstruction possible, and tight enough that it fails the instant you weaken it by a single frequency at the band edge. That is an unusually precise bargain, and it is why a result that fits in one sentence has survived a century of engineering practice essentially unamended.
Related topics: Compressed Sensing, Fourier Series, Fast Fourier Transform, Wavelets and Multiresolution Analysis, Information Theory and Entropy, Complex Analysis, Hamming Codes, The Mathematics of Music
Frequently asked
What is the Nyquist-Shannon sampling theorem?
It says that if a signal contains no frequency components at or above B hertz, then it is completely and exactly determined by its values at evenly spaced instants, provided the sampling rate is strictly greater than 2B samples per second. The original continuous signal can be rebuilt from those samples by sinc interpolation, with no error at all. The theorem is a statement about what the band-limitation hypothesis buys you, not a rule of thumb about data volume.
What is aliasing?
Aliasing is what happens when a frequency component too high for the sampling rate is misread as a lower one. A component at frequency f sampled at rate fs appears in the samples at the folded frequency given by the absolute value of f minus fs times the nearest integer to f divided by fs. A 5 Hz sinusoid sampled at 4 Hz is indistinguishable from a 1 Hz sinusoid, because the two functions take exactly the same value at every sampling instant. Once aliasing has occurred no amount of later processing can undo it, which is why anti-alias filtering happens before the converter.
Why must the sampling rate be more than twice the highest frequency?
Sampling a signal is equivalent to multiplying it by a Dirac comb, and multiplication in time is convolution in frequency. The spectrum is therefore replicated at every integer multiple of the sampling rate fs. A spectrum occupying the interval from minus B to B has width 2B, so the replicas are spaced fs apart and each is 2B wide. They fail to overlap precisely when fs is greater than 2B. The factor of two is the ratio of the full two-sided bandwidth to the one-sided bandwidth, not an engineering safety margin.
Why is CD audio sampled at 44.1 kHz?
Human hearing extends to roughly 20 kHz, which sets a Nyquist rate near 40 kHz, and the remaining 4.1 kHz is transition band for the anti-alias filter. The specific figure 44,100 comes from the practice of storing early digital audio on video tape using PCM adaptors. In the 60 Hz video format, 245 usable lines per field times 60 fields per second times 3 samples per line gives exactly 44,100. In the 50 Hz format, 294 lines times 50 fields times 3 samples gives the same number, so one rate served both.
Can you sample below the Nyquist rate?
You can, if you replace the band-limitation hypothesis with a different one. Bandpass sampling exploits a signal confined to a narrow band far from zero and can work at rates near twice the bandwidth rather than twice the highest frequency, subject to a condition on where the band sits. Compressed sensing replaces bandwidth with sparsity in some basis and uses nonlinear reconstruction. Finite rate of innovation sampling counts degrees of freedom per unit time. None of these contradicts the theorem; each assumes something the theorem does not.