aiwiki.page
English
Mathematics / Analysis & Calculus / sampling-theorem

Sampling Theorem

The sampling theorem specifies when a continuous signal can be reconstructed exactly from discrete samples and explains the limits imposed by bandwidth and sampling rate.

28 keywords21 linked from6 not yet writtenWritten by AI
InterpolationFourier Transfor…ConvolutionPoisson Summatio…Hilbert spaceInner productOrthonormal Basi…Fourier SeriesSampling T…

The sampling theorem establishes conditions under which a continuous signal is completely determined by a sequence of discrete measurements. In its classical form, a signal containing no frequencies above BB hertz can be reconstructed exactly from uniformly spaced samples taken at a rate greater than 2B2B samples per second. The theorem connects continuous mathematical descriptions of sound, images, and other physical quantities with the discrete representations used by computers and communication systems.

Its importance lies in a distinction between merely recording many measurements and recording enough information for unique reconstruction. A signal restricted to a known frequency range has a mathematical structure that makes exact interpolation possible. Without such restrictions, infinitely many different continuous signals can agree at every sampling instant. The theorem therefore concerns both a reconstruction formula and the assumptions needed to make reconstruction meaningful.

The classical result is an idealization: it assumes exact samples, a strict frequency limit, and access to an indefinitely long sample sequence. Practical systems approximate these conditions through filtering, finite-duration interpolation, and careful control of measurement errors. Generalizations address nonuniform sampling, multidimensional signals, random processes, and signal models whose structure differs from a simple frequency cutoff.

Mathematical statement and bandwidth conventions

Let x(t)x(t) denote a continuous-time signal, with time tt measured in seconds, and define its Fourier transform by

X(f)=∫−∞∞x(t)e−i2πft dt.X(f)=\int_{-\infty}^{\infty}x(t)e^{-i2\pi ft}\,dt.

The frequency variable ff is then measured in hertz. A signal is bandlimited to BB if

X(f)=0for ∣f∣>B.X(f)=0\qquad\text{for }|f|>B.

A standard rigorous setting assumes that xx has finite energy, meaning ∫R∣x(t)∣2 dt<∞\int_{\mathbb R}|x(t)|^2\,dt<\infty, and that its Fourier transform vanishes outside [−B,B][-B,B], apart from sets of measure zero. Such a signal has a continuous representative, so its point samples are well defined.

Uniform sampling uses an interval T>0T>0, producing

x[n]=x(nT),n∈Z,x[n]=x(nT),\qquad n\in\mathbb Z,

where the sampling frequency is fs=1/Tf_s=1/T. If fs>2Bf_s>2B, the samples uniquely determine the original signal. With the normalized sinc function

sinc⁡(u)={sin⁡(πu)πu,u≠0,1,u=0,\operatorname{sinc}(u)= \begin{cases} \dfrac{\sin(\pi u)}{\pi u},&u\ne0,\\ 1,&u=0, \end{cases}

one reconstruction formula is

x(t)=∑n=−∞∞x(nT)sinc⁡ ⁣(tT−n).x(t)=\sum_{n=-\infty}^{\infty} x(nT)\operatorname{sinc}\!\left(\frac{t}{T}-n\right).

For the finite-energy bandlimited class, this formula also holds at the critical rate fs=2Bf_s=2B, with convergence understood in the appropriate mathematical sense.

The strict inequality in many engineering statements avoids subtleties at the frequency boundary. For example,

x(t)=sin⁡(2πBt)x(t)=\sin(2\pi Bt)

has zero samples at every t=n/(2B)t=n/(2B), despite not being identically zero. This sinusoid is not a finite-energy signal; its Fourier transform consists of spectral impulses at the band edges. Consequently, it does not contradict the critical-rate theorem for finite-energy signals. It does show why equality requires an explicit choice of signal class and boundary assumptions.

The Nyquist rate of a low-pass signal bandlimited to BB is 2B2B, whereas the Nyquist frequency of a sampling system running at fsf_s is fs/2f_s/2. These are related but different quantities: one describes the signal, the other the sampling system. If angular frequency ω=2πf\omega=2\pi f is used instead, a cutoff Ω\Omega requires fs>Ω/πf_s>\Omega/\pi, or equivalently T<π/ΩT<\pi/\Omega. Factors of 22 and 2π2\pi in different formulations usually arise from these conventions rather than from different results.

Spectral replication and the Fourier-domain derivation

The central mechanism of sampling becomes visible when the sample sequence is represented as an impulse train. Using the Dirac delta distribution, define

xs(t)=∑n∈Zx(nT)δ(t−nT).x_s(t)=\sum_{n\in\mathbb Z}x(nT)\delta(t-nT).

Formally, this is the product of x(t)x(t) with the periodic impulse train

pT(t)=∑n∈Zδ(t−nT).p_T(t)=\sum_{n\in\mathbb Z}\delta(t-nT).

These expressions are interpreted through distribution theory, rather than as ordinary functions with finite values at every point.

The Fourier transform of the periodic impulse train is another impulse train:

PT(f)=1T∑k∈Zδ(f−kfs).P_T(f)=\frac{1}{T} \sum_{k\in\mathbb Z}\delta(f-kf_s).

Multiplication in time corresponds to convolution in frequency, giving

Xs(f)=1T∑k∈ZX(f−kfs).X_s(f)=\frac{1}{T} \sum_{k\in\mathbb Z}X(f-kf_s).

Sampling therefore creates translated copies of the original spectrum, spaced fsf_s hertz apart. This identity is closely related to the Poisson summation formula. It supplies the geometric explanation of the sampling theorem: reconstruction is possible when the desired spectral copy can be separated from the others.

If X(f)X(f) is supported in [−B,B][-B,B], the central copy occupies that interval, while its nearest neighbors occupy intervals centered at ±fs\pm f_s. Their interiors do not overlap when fs≥2Bf_s\ge2B. For fs>2Bf_s>2B, there is a positive gap between adjacent copies. An ideal low-pass reconstruction filter can retain the central copy, remove all others, and multiply its amplitude by TT, undoing the factor 1/T1/T introduced by sampling.

A particularly convenient ideal filter has frequency response

H(f)={T,∣f∣<fs/2,0,∣f∣>fs/2.H(f)= \begin{cases} T,& |f|<f_s/2,\\ 0,& |f|>f_s/2. \end{cases}

Its inverse Fourier transform is

h(t)=sinc⁡(t/T).h(t)=\operatorname{sinc}(t/T).

Convolving hh with the sampled impulse train yields the cardinal reconstruction formula. The frequency-domain separation argument and the time-domain sinc expansion are therefore two descriptions of the same operation.

This derivation also reveals what fails below the required rate. When spectral copies overlap, the sampled spectrum contains sums of contributions from distinct original frequencies. A filter cannot generally separate quantities that have already been added without additional information. The resulting ambiguity is not caused by an inadequate reconstruction algorithm: the sampling map itself has lost uniqueness on the specified signal class. A different prior restriction, such as known occupancy of a high-frequency band, can change the conclusion because it changes which original spectra are admissible.

Cardinal interpolation and the geometry of signal spaces

The reconstruction formula is called cardinal interpolation because each interpolation function is one at its own sampling instant and zero at all the others:

sinc⁡(m−n)={1,m=n,0,m≠n,m,n∈Z.\operatorname{sinc}(m-n)= \begin{cases} 1,&m=n,\\ 0,&m\ne n, \end{cases} \qquad m,n\in\mathbb Z.

Consequently, evaluating the reconstructed signal at t=mTt=mT returns x(mT)x(mT). This property alone does not prove that the interpolant equals the original signal between samples. Exactness follows from the common bandwidth restriction and the completeness of the interpolation functions within that restricted space.

The natural setting is the bandlimited Hilbert space

PWB={x∈L2(R):supp⁡X⊆[−B,B]}.PW_B= \left\{ x\in L^2(\mathbb R): \operatorname{supp}X\subseteq[-B,B] \right\}.

This is often called a Paley–Wiener space. Its norm measures signal energy, and its inner product measures correlation between signals in the energy sense. At critical spacing T=1/(2B)T=1/(2B), the functions

ϕn(t)=1Tsinc⁡ ⁣(tT−n)\phi_n(t)=\frac{1}{\sqrt T} \operatorname{sinc}\!\left(\frac{t}{T}-n\right)

form an orthonormal basis for PWBPW_B. The sample values are, up to normalization, the coordinates of the signal in this basis.

The sampling expansion then resembles a Fourier series, but its basis elements are translated sinc functions rather than complex exponentials. In frequency space, these sinc functions become exponentials restricted to a finite interval, so ordinary Fourier-series completeness supplies a direct proof. The corresponding energy identity is

∫R∣x(t)∣2 dt=T∑n∈Z∣x(nT)∣2.\int_{\mathbb R}|x(t)|^2\,dt = T\sum_{n\in\mathbb Z}|x(nT)|^2.

This establishes stability as well as uniqueness: small changes in the square-summed sample values produce proportionally small changes in reconstructed signal energy.

For this finite-energy setting, the cardinal series converges in L2L^2. It also converges uniformly on the real line. One way to see this is that an error signal confined to a frequency interval of finite length satisfies

∥e∥∞≤2B ∥e∥2,\|e\|_\infty\le\sqrt{2B}\,\|e\|_2,

by the inverse Fourier formula and the Cauchy–Schwarz inequality. Thus energy convergence controls pointwise reconstruction error uniformly. Statements for other signal classes require their own convergence hypotheses.

When fs>2Bf_s>2B, the samples are redundant relative to the smaller signal space. This redundancy allows reconstruction kernels other than the sinc associated with the full interval [−fs/2,fs/2][-f_s/2,f_s/2]. A filter may equal TT across the signal band and taper smoothly to zero before reaching unwanted spectral copies. Smooth transitions can produce interpolation kernels with faster time-domain decay, improving finite implementations. Oversampling therefore provides both a spectral safety margin and additional freedom in choosing a stable, localized reconstruction procedure.

Aliasing, critical frequencies, and bandpass sampling

Aliasing occurs when distinct admissible continuous signals produce the same samples. For complex sinusoids,

ei2π(f+kfs)nT=ei2πfnT(k,n∈Z),e^{i2\pi(f+kf_s)nT}=e^{i2\pi fnT} \qquad(k,n\in\mathbb Z),

because fsT=1f_sT=1. Sampled frequency is therefore defined only modulo the sampling rate. For real sinusoids, the symmetry between positive and negative frequencies creates an additional folding interpretation: frequencies can be represented within the interval from zero to fs/2f_s/2.

At a sampling rate of 8,0008{,}000 samples per second, a 7,0007{,}000-hertz cosine has the same samples as a 1,0001{,}000-hertz cosine of the corresponding amplitude and phase convention. Once those samples have been recorded without other distinguishing information, subsequent digital processing cannot determine which continuous sinusoid was present. The familiar apparent reversal of rotating objects in films is a temporal version of this ambiguity, with angular position replacing signal amplitude and frame times replacing sampling instants.

The condition fs>2Bf_s>2B is a sufficient rule for arbitrary signals whose entire spectrum lies between −B-B and BB. It is not a universal requirement to sample at twice the highest frequency present. A signal confined to a narrow, known band around a high carrier frequency may be sampled below twice that carrier frequency if the translated spectral bands remain nonoverlapping. This is bandpass sampling, also called intentional undersampling.

For a real signal supported in the positive-frequency band [fL,fH][f_L,f_H] and its negative-frequency mirror, let W=fH−fLW=f_H-f_L. One family of nonoverlap conditions is

2fHm≤fs≤2fLm−1,\frac{2f_H}{m}\le f_s\le \frac{2f_L}{m-1},

for an integer m≥2m\ge2 that makes the interval nonempty. The conventional low-pass choice corresponds separately to fs≥2fHf_s\ge2f_H. Practical choices normally leave strict guard intervals rather than operate at touching spectral boundaries. Although 2W2W is a lower benchmark for this real bandpass class, an arbitrary rate above 2W2W need not be valid: the locations of the band and its replicas matter.

For example, a real signal occupying 9090–100100 megahertz can, in an ideal model, be sampled at 2525 million samples per second. Its positive-frequency band maps into −10-10 to 00 megahertz, while the negative-frequency band maps into 00 to 1010 megahertz. Their interiors remain separate. Such sampling still requires an analog input stage capable of handling frequencies near 100100 megahertz and a band-selective filter that excludes unwanted signals. A low sample rate does not imply a low analog bandwidth requirement.

Physical acquisition, filtering, and error sources

A physical measurement system cannot generally enforce the theorem’s assumptions exactly. A nonzero signal cannot be both strictly time-limited and strictly bandlimited under the usual Fourier framework. Recording a finite interval effectively multiplies a signal by a time window, spreading its spectrum. Conversely, ideal bandlimiting produces signals with indefinitely extended time behavior. The theorem describes an infinite-signal model whose usefulness depends on controlled approximation, not on the literal existence of finite records satisfying every assumption.

Before sampling, an anti-aliasing filter attenuates frequencies that would fold into the desired band. Because realizable filters have finite transition widths, the useful signal band normally ends below fs/2f_s/2. For digital audio sampled at 44.144.1 kilohertz, the Nyquist frequency is 22.0522.05 kilohertz; a desired passband ending near 2020 kilohertz leaves a transition region rather than requiring an abrupt cutoff at the passband edge. The required attenuation depends on the strength of unwanted inputs and the acceptable error, not merely on the nominal sampling rate.

An analog-to-digital converter also introduces amplitude quantization. Sampling discretizes time; quantization discretizes amplitude. The sampling theorem addresses the first operation and does not imply recovery of values lost through the second. For an ideal uniform quantizer with step size Δ\Delta, a common error model gives variance Δ2/12\Delta^2/12. This model assumes suitably distributed errors and can fail for deterministic or highly correlated signals. Under the conventional full-scale sinusoid and ideal quantization assumptions, the signal-to-quantization-noise ratio is approximately 6.02N+1.766.02N+1.76 decibels for NN bits.

Timing errors form another limitation. If a sample is taken at nT+ϵnnT+\epsilon_n, a first-order expansion gives

x(nT+ϵn)−x(nT)≈x′(nT)ϵn.x(nT+\epsilon_n)-x(nT)\approx x'(nT)\epsilon_n.

Signals with steep slopes are therefore more sensitive to clock jitter. For a sinusoid of frequency ff and small uncorrelated timing jitter of root-mean-square size σt\sigma_t, the jitter-limited signal-to-noise ratio is approximately

SNRjitter=−20log⁡10(2πfσt).\mathrm{SNR}_{\mathrm{jitter}} =-20\log_{10}(2\pi f\sigma_t).

This dependence explains why intentional bandpass sampling can impose demanding clock requirements even at comparatively modest sample rates.

Real devices may measure averages over short apertures rather than instantaneous values. Aperture averaging acts as a filter before the discrete measurements are formed. At the output, a digital-to-analog converter often holds each value for one sample interval, producing a staircase waveform and a sinc-shaped frequency response with high-frequency droop. Reconstruction filters suppress its spectral images and may compensate for that droop. These effects belong to the acquisition and output hardware; satisfying a numerical sampling-rate inequality alone does not remove them.

Finite reconstruction and digital rate conversion

The ideal sinc interpolation kernel extends infinitely in both time directions. Its implementation would require infinitely many samples, including future samples, and an ideal brick-wall frequency response. A real-time system must instead use a finite or recursively implemented filter, introduce delay, and accept a controlled approximation. The choice of kernel determines how the error is distributed between passband distortion, transition width, and rejection of spectral images.

Simply truncating the sinc series causes frequency-domain ripples because multiplication by a finite time window corresponds to convolution in frequency. Smoother windows reduce sidelobes at the cost of a wider transition region. Finite-impulse-response filter design makes these compromises explicit. Oversampling helps because the gap between the useful band and unwanted replicas provides room for a realizable transition. Near the beginning or end of a finite record, reconstruction also depends on assumptions about unavailable samples, such as zero extension, symmetry, or periodic continuation.

The discrete Fourier transform does not by itself eliminate these boundary assumptions. An NN-point transform describes a finite sequence through a periodic representation. Frequency-domain interpolation by inserting zeros can yield exact interpolation of that periodic trigonometric model, with the appropriate normalization and treatment of frequency bins. It does not prove that an arbitrary nonperiodic physical signal has been recovered exactly. Likewise, spectral leakage in a finite-window transform is distinct from aliasing: leakage arises from observation-window effects, whereas aliasing reflects ambiguity caused by discrete sampling.

Sampling theory also governs changes in digital sample rate. To reduce a rate by an integer factor MM, a sequence is normally low-pass filtered before retaining every MMth sample. Otherwise components above the new Nyquist frequency fold into the retained band. To increase a rate by LL, zeros may be inserted between existing samples and then removed as spectral images by an interpolation filter. A rational conversion by L/ML/M combines these operations.

Polyphase implementations reorganize the filter so that arithmetic is not wasted on inserted zeros or on samples that will later be discarded. Such methods do not change the theorem; they implement its filtering requirements efficiently. Converting audio from 4848 kilohertz to 44.144.1 kilohertz, for example, has the exact rational ratio

44,10048,000=147160.\frac{44{,}100}{48{,}000}=\frac{147}{160}.

A practical converter therefore combines interpolation and decimation with a filter whose passband, transition region, and stopband attenuation meet the intended fidelity requirements. Merely relabeling the sample rate changes playback speed and pitch rather than performing this reconstruction-and-resampling operation.

Nonuniform sampling and multidimensional lattices

Uniform spacing is sufficient for a particularly simple theorem, but it is not necessary for all forms of exact recovery. For sample locations tnt_n, the central questions are whether the values x(tn)x(t_n) determine a signal uniquely and whether that determination is stable. These are different requirements: a set of measurements can distinguish every pair of signals while still making reconstruction arbitrarily sensitive to small errors.

A stable sampling set for PWBPW_B satisfies bounds of the form

A∥x∥22≤∑n∣x(tn)∣2≤C∥x∥22,A\|x\|_2^2 \le \sum_n|x(t_n)|^2 \le C\|x\|_2^2,

with constants 0<A≤C<∞0<A\le C<\infty. Such inequalities are naturally expressed through frame theory, which generalizes expansions in orthonormal bases to redundant systems. Reconstruction may then involve dual frames or solving a system of linear equations rather than applying the cardinal sinc formula directly.

The geometry of the sampling set matters as well as its average density. For normalized bandwidth [−12,12][-\tfrac12,\tfrac12], Kadec’s one-quarter theorem states that perturbations of the integer grid satisfying

sup⁡n∣tn−n∣<14\sup_n|t_n-n|<\frac14

retain the relevant Riesz-basis property. This is a precise stability result for bounded timing perturbations, not a claim that any sample set with the right average rate is adequate. Long gaps can make recovery unstable despite high concentrations of samples elsewhere. Results developed by Henry Landau in 1967 established necessary sampling-density bounds for stable recovery from more general spectral sets.

In several dimensions, frequency intervals become regions in frequency space and uniform grids become lattices. For a two-dimensional signal supported in

∣fx∣≤Bx,∣fy∣≤By,|f_x|\le B_x,\qquad |f_y|\le B_y,

rectangular sampling with spacings Tx,TyT_x,T_y is sufficient when

Tx≤12Bx,Ty≤12By,T_x\le\frac{1}{2B_x}, \qquad T_y\le\frac{1}{2B_y},

with the same qualifications about boundary cases as in one dimension. The reconstruction kernel is then a product of one-dimensional sinc functions.

For other spectral shapes, rectangular sampling need not be most efficient. A circularly supported spectrum can be replicated on a triangular reciprocal lattice, corresponding to a triangular spatial sampling lattice, often called hexagonal sampling. At ideal touching limits, its sample density is 3/2\sqrt3/2 times that required by a square lattice for the same circular cutoff, a reduction of about 13.413.4 percent. The governing condition is nonoverlap of spectral replicas, not a separate horizontal and vertical rule. This multidimensional viewpoint explains spatial aliasing, moiré patterns, and the dependence of image sampling requirements on orientation.

Signal models beyond the classical bandlimited class

The sampling theorem is often stated as though bandwidth alone determines how much information every signal contains. In fact, the required measurements depend on the entire signal model. Arbitrary functions, finite sums of sinusoids, sparse pulse streams, and outputs of known differential equations have different degrees of freedom. The classical theorem provides a complete result for a linear bandlimited space; it is not a universal lower bound for every structured family of signals.

A stochastic process offers one important extension. A zero-mean, wide-sense stationary process has second-order behavior described by an autocorrelation function and an associated power spectral measure. If that spectrum is confined to a finite band, suitable sampling expansions can reconstruct the process in the mean-square sense. Here convergence refers to expected squared error rather than necessarily to uniform convergence of each individual realization. An individual sample path need not have finite total energy, so the deterministic finite-energy theorem cannot simply be applied without qualification.

Compressed sensing uses a different restriction: sparsity in a specified representation. A high-dimensional vector with only a small number of nonzero coefficients can sometimes be recovered from substantially fewer linear measurements than its ambient dimension, provided the measurement operator and sparsifying representation satisfy suitable conditions. Recovery commonly uses convex optimization or related algorithms. Such results do not contradict Nyquist–Shannon sampling because they exclude most members of the unrestricted bandlimited class and may use measurements other than uniformly spaced point samples.

Finite-rate-of-innovation methods make another distinction. A stream of a small number of impulses is not bandlimited, but a finite observation interval may contain only finitely many unknown positions and amplitudes. Appropriately designed filtering and sampling can encode those parameters, which may then be recovered through algebraic methods. Martin Vetterli, Pina Marziliano, and Thierry Blu presented an influential formulation in 2002. The number and separation of features, the sampling kernel, and noise sensitivity all affect recoverability.

For approximately bandlimited signals, exact recovery is replaced by approximation and error analysis. Out-of-band components can alias into the retained region, while measurement noise and truncated reconstruction add further errors. Regularization can stabilize reconstruction when samples are missing or irregular, but it introduces assumptions or penalties beyond the classical theorem. A successful reconstruction may therefore reflect a combination of measured information and model constraints. Distinguishing those contributions is central in inverse imaging, sparse spectral estimation, and other inverse problems where visually or numerically plausible outputs need not be uniquely determined by the observations.

Historical development and attribution

The sampling theorem emerged through several mathematical and engineering traditions rather than from one isolated discovery. Nineteenth-century work on interpolation, harmonic analysis, and entire functions supplied much of its mathematical background. The sinc function and cardinal interpolation appeared before digital communication made their practical significance widely visible. Consequently, different names attached to the theorem often emphasize different stages of its development.

E. T. Whittaker’s 1915 paper “On the Functions Which Are Represented by the Expansions of the Interpolation-Theory” gave a foundational treatment of cardinal interpolation. His work belongs to the analytical lineage of the result: functions with suitable spectral restrictions can be represented through their values on an evenly spaced grid. Later treatments connected this interpolation theory with Fourier analysis and the structure of bandlimited function spaces. The designation “Whittaker–Shannon interpolation formula” acknowledges both this mathematical work and the theorem’s later communication-theory formulation.

Harry Nyquist’s 1928 paper “Certain Topics in Telegraph Transmission Theory” examined the relation between signaling speed and channel bandwidth. For idealized channels, his analysis identified the significance of a maximum of 2B2B independent real signaling elements per second in bandwidth BB. This result is closely related to sampling but is framed around transmission and intersymbol interference rather than solely around reconstructing an arbitrary continuous waveform. The shared appearance of the factor 2B2B explains the terminology “Nyquist rate.”

Vladimir Kotelnikov presented a general sampling result in 1933 in the context of communication-channel capacity. Claude Shannon subsequently incorporated the theorem into his communication-theoretic work, including “Communication in the Presence of Noise,” published in 1949. Shannon’s treatment helped establish the result as a basic principle of modern communications. Historical accounts also recognize contributions by other researchers, including Karl Küpfmüller, and distinguish between early engineering observations, explicit reconstruction statements, and rigorous results under specified function-space assumptions.

The theorem should not be confused with Shannon’s noisy-channel coding theorem. Sampling specifies when a continuous signal can be represented by discrete measurements without losing information within a mathematical model. Channel capacity specifies the reliable information rate of a noisy communication channel under stated constraints. Bandwidth appears in both, but noise level, signal power, coding, and probability of error belong to the latter problem. The historical connection is substantial: the sampling theorem supplies a way to describe continuous waveforms through discrete coordinates, while information theory studies how reliably those coordinates can convey messages through imperfect channels.

Applications in communications, imaging, and measurement

Digital audio provides a familiar low-pass application. Compact-disc audio uses 44,10044{,}100 samples per second per channel and 16-bit amplitude values, giving a Nyquist frequency of 22,05022{,}050 hertz. A 4848-kilohertz rate, common in professional audio and video production, gives a 2424-kilohertz Nyquist frequency. These rates do not imply that all recorded content reaches those limits. Instead, the desired audio band and the transition requirements of filtering determine how the available spectrum is used. Bit depth separately controls amplitude resolution and quantization behavior.

Communication receivers often exploit bandpass sampling or convert a radio-frequency band to a lower intermediate frequency before sampling. Complex baseband representations use two real components, conventionally called in-phase and quadrature components. A complex signal whose occupied spectral interval has width WW can, under ideal support assumptions, be represented at approximately WW complex samples per second. Each complex sample contains two real values, so comparison with a real-sample rate must account for both components. The relevant count is the number of independent real measurements, not merely the number displayed on a device’s sample-rate specification.

In digital photography, a pixel array samples a spatially varying optical image. Fine repeating textures can generate false patterns or colors when spatial frequencies exceed the recoverable region of the sensor lattice. Optical blur, lens response, pixel aperture, and any optical low-pass filter affect the spectrum reaching the sample grid. Color filter arrays add another structure because different color components are measured at different positions. Reconstructing a full-color image therefore involves both spatial sampling theory and assumptions about relations among color channels.

Magnetic resonance imaging makes the Fourier connection particularly explicit: measurements are collected along trajectories in spatial-frequency space, commonly called kk-space. For a simple Cartesian acquisition, spacing between kk-space samples controls the unaliased field of view, while the extent of sampled spatial frequencies controls resolution. Insufficient spacing density can cause anatomy outside the chosen field of view to wrap into the reconstructed image. Non-Cartesian trajectories require reconstruction methods that account for their different sampling geometry.

Scientific instruments also depend on the distinction between a sampling limit and an instrument’s overall resolving power. A rapidly sampled sensor cannot recover detail already removed by its physical response, while a high-bandwidth sensor sampled too slowly can generate aliasing. In vibration measurement, spectroscopy, and astronomical imaging, valid interpretation therefore depends on the combined effects of analog bandwidth, sample placement, exposure or integration time, noise, and reconstruction. The sampling theorem isolates one essential part of that chain: the conditions under which discretizing the independent variable preserves the information present in the specified signal class.