{
  "id": "nyquist-shannon-sampling-theorem",
  "title": "Nyquist–Shannon Sampling Theorem",
  "lang": "en",
  "source_lang": "en",
  "summary": "The Nyquist–Shannon sampling theorem establishes conditions under which a band-limited continuous signal can be reconstructed exactly from uniformly spaced samples.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Shannon sampling theorem",
    "Nyquist sampling theorem",
    "Whittaker–Shannon sampling theorem",
    "Whittaker–Kotelnikov–Shannon sampling theorem"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/nyquist-shannon-sampling-theorem/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/nyquist-shannon-sampling-theorem/",
      "json": "https://aiwiki.page/zh/nyquist-shannon-sampling-theorem.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The Nyquist–Shannon sampling theorem states that a continuous-time signal whose frequency content is confined to a finite band can, under appropriate mathematical conditions, be determined completely by its values at uniformly spaced instants. In its familiar low-pass form, a signal containing no frequencies above \\(B\\) hertz can be reconstructed from samples taken at a rate greater than \\(2B\\) samples per second. The number \\(2B\\) is its Nyquist rate; half a specified sampling rate is the corresponding Nyquist frequency. These related quantities describe different things and should not be confused.\n\nThe theorem supplies a mathematical foundation for converting continuous signals into discrete representations. It explains why digital audio, sampled measurement systems, telecommunications, and digital imaging can preserve the information in an appropriately restricted analog signal. Its central operation is not merely recording points but reconstructing the continuous signal through interpolation, with the Fourier transform connecting the time-domain samples to the signal’s frequency content.\n\nAs a particular sampling theorem, the result depends on assumptions about bandwidth, sampling geometry, and the class of signals under consideration. Exact reconstruction uses idealized infinite data and filtering; practical systems introduce finite records, imperfect filters, noise, timing errors, and amplitude quantization. Extensions address bandpass signals, nonuniform sampling, higher-dimensional data, and signals characterized by constraints other than a single low-pass bandwidth."
    },
    {
      "heading": "Mathematical statement and assumptions",
      "text": "Let \\(x(t)\\) be a continuous-time signal, and adopt the Fourier-transform convention\n\n\\[X(f)=\\int_{-\\infty}^{\\infty}x(t)e^{-i2\\pi ft}\\,dt.\\]\n\nThe signal is band-limited to \\(B\\) hertz if \\(X(f)=0\\) for \\(|f|>B\\). This statement concerns its entire spectrum, including negative frequencies. For a real-valued signal, positive and negative frequencies are related by complex conjugation; the negative-frequency part is not an additional independent physical oscillation. The frequency variable \\(f\\) here is measured in cycles per second, rather than radians per second.\n\nSuppose samples are taken at times \\(t=nT\\), where \\(n\\) ranges over all integers, \\(T>0\\) is the sampling interval, and \\(f_s=1/T\\) is the sampling frequency. A standard version of the theorem states that, for a finite-energy band-limited signal and \\(f_s\\geq 2B\\),\n\n\\[x(t)=\\sum_{n=-\\infty}^{\\infty} x(nT)\\, \\operatorname{sinc}\\!\\left(\\frac{t-nT}{T}\\right),\\]\n\nwhere\n\n\\[\\operatorname{sinc}(u)= \\begin{cases} \\dfrac{\\sin(\\pi u)}{\\pi u},&u\\ne0,\\\\[4pt] 1,&u=0. \\end{cases}\\]\n\nFinite energy means that \\(\\int_{\\mathbb R}|x(t)|^2\\,dt<\\infty\\). Such signals belong to an L^2 space, and the inverse transform supplies a natural continuous representative to which the sampling formula applies.\n\nThe distinction between \\(f_s>2B\\) and \\(f_s=2B\\) requires care. In the finite-energy formulation, values of the spectrum at isolated endpoints do not affect the signal, and critical-rate reconstruction is valid. For broader signal classes, however, spectral lines exactly at the boundary can cause failure. At \\(f_s=2B\\), the sinusoid\n\n\\[x(t)=\\sin(2\\pi Bt)\\]\n\nhas samples \\(x(nT)=\\sin(\\pi n)=0\\), making it indistinguishable from the zero signal. This sinusoid has infinite total energy and lies outside the finite-energy hypothesis. The frequently stated strict inequality avoids this endpoint ambiguity and leaves room for practical filter transitions.\n\nThe assumptions also distinguish uniqueness from physical realizability. The theorem identifies the signal uniquely within a specified band-limited class; without that restriction, infinitely many continuous functions can pass through the same sample values. Conversely, a nonzero finite-energy signal cannot be both strictly band-limited and confined to a finite time interval. Band-limited signals have analytic continuation properties: their inverse Fourier integral extends to an entire analytic function of complex time. Real signals of finite duration therefore satisfy the bandwidth assumption only approximately, although their out-of-band content may be negligibly small for a particular application."
    },
    {
      "heading": "Spectral replication and a Fourier-domain proof",
      "text": "Uniform sampling can be represented mathematically by multiplying the signal by an impulse train:\n\n\\[p_T(t)=\\sum_{n\\in\\mathbb Z}\\delta(t-nT),\\qquad x_s(t)=x(t)p_T(t) =\\sum_{n\\in\\mathbb Z}x(nT)\\delta(t-nT).\\]\n\nThe impulses are Dirac distributions rather than ordinary functions, so these expressions are interpreted through distribution theory. Their role is to encode sample locations and amplitudes exactly; an actual converter measures voltages or charges over nonzero time intervals.\n\nThe Fourier transform of the impulse train is another impulse train:\n\n\\[P_T(f)=\\frac{1}{T} \\sum_{k\\in\\mathbb Z}\\delta(f-kf_s).\\]\n\nMultiplication in time becomes convolution in frequency, giving\n\n\\[X_s(f)=\\frac{1}{T} \\sum_{k\\in\\mathbb Z}X(f-kf_s).\\]\n\nThus sampling produces shifted copies of the original spectrum, centered at every integer multiple of \\(f_s\\). The factor \\(1/T\\) follows from the chosen normalization. This spectral-replication identity is the central mechanism behind both exact reconstruction and its failure when the rate is insufficient.\n\nIf the original spectrum occupies \\([-B,B]\\), adjacent replicas do not overlap in their interiors when \\(f_s\\geq2B\\). An ideal reconstruction filter can select the central copy and restore its amplitude. For example, take a frequency response equal to \\(T\\) on \\((-f_s/2,f_s/2)\\) and zero outside. Its impulse response is\n\n\\[h(t)=\\operatorname{sinc}(t/T).\\]\n\nFiltering the sampled impulse train then yields\n\n\\[(x_s*h)(t) =\\sum_{n\\in\\mathbb Z}x(nT)h(t-nT),\\]\n\nwhich is precisely the reconstruction series. At critical sampling the spectral copies meet at their boundaries; above the critical rate, an unused frequency gap separates them.\n\nA complementary proof uses Fourier series. Regard \\(X(f)\\), extended by zero if necessary, as a square-integrable function on \\([-f_s/2,f_s/2]\\). Its coefficients in the exponential basis are proportional to the samples:\n\n\\[T\\int_{-f_s/2}^{f_s/2} X(f)e^{i2\\pi fnT}\\,df =T\\,x(nT).\\]\n\nConsequently,\n\n\\[X(f)=T\\sum_{n\\in\\mathbb Z} x(nT)e^{-i2\\pi fnT}\\]\n\nin the appropriate mean-square sense on that interval. Inverting each basis term produces a shifted sinc function. This proof shows that the sample sequence is not merely a collection of measurements: it supplies the coordinates of the spectrum in a complete orthonormal basis, after normalization. It also explains why the sampling interval and the spectral interval have reciprocal widths."
    },
    {
      "heading": "Sinc interpolation, convergence, and stability",
      "text": "The reconstruction kernel is often called the cardinal sine, and the resulting expansion the cardinal series. Its interpolation property follows from\n\n\\[\\operatorname{sinc}(m-n)= \\begin{cases} 1,&m=n,\\\\ 0,&m\\ne n. \\end{cases}\\]\n\nAt a sampling instant \\(t=mT\\), every term except the \\(m\\)th disappears. Between sampling instants, however, all samples generally contribute. Exact reconstruction is therefore a global operation, unlike joining adjacent samples by straight lines or selecting the nearest sample.\n\nThe natural setting is the Hilbert space of square-integrable functions with spectra supported in \\([-f_s/2,f_s/2]\\). In that space, the shifted sinc functions are orthogonal, each with squared norm \\(T\\). The sampling expansion converges in the energy norm and satisfies\n\n\\[\\int_{-\\infty}^{\\infty}|x(t)|^2\\,dt = T\\sum_{n\\in\\mathbb Z}|x(nT)|^2.\\]\n\nThis identity states that, with the appropriate scaling, sampling preserves energy. It also makes uniqueness explicit: if every sample of a signal in this space is zero, the signal itself has zero energy.\n\nFor this finite-energy setting, convergence also holds uniformly in time. If \\(e(t)\\) is a band-limited reconstruction error supported in the same frequency interval, the inverse Fourier integral and the Cauchy–Schwarz inequality give\n\n\\[|e(t)|\\leq \\sqrt{f_s}\\,\\|e\\|_2.\\]\n\nThus convergence in energy of the partial sinc expansions implies uniform convergence. Such statements depend on the function space: a formula written for finite-energy signals does not automatically establish identical convergence properties for every bounded waveform, generalized function, or random process.\n\nThe energy identity gives a precise stability result. If sample errors \\(e_n\\) form a square-summable sequence, their sinc interpolation has energy\n\n\\[\\left\\|\\sum_n e_n\\operatorname{sinc}(t/T-n)\\right\\|_2^2 =T\\sum_n|e_n|^2.\\]\n\nNevertheless, stability in an energy norm does not imply every possible form of robustness. The absolute sum of shifted sinc magnitudes is generally unbounded, so uniformly bounded but otherwise arbitrary errors need not produce a uniformly controlled output through the ideal critical-rate interpolator.\n\nOversampling changes the reconstruction possibilities. If \\(f_s>2B\\), the filter need only reproduce the occupied signal band while rejecting the other spectral copies; its response in the intervening gap can be chosen more freely. A smoother frequency response can produce a faster-decaying time-domain kernel than sinc. This freedom improves practical truncation behavior and can strengthen robustness in norms relevant to bounded disturbances. The samples then contain redundancy relative to the smaller signal space, rather than representing independent coordinates at every sampling instant."
    },
    {
      "heading": "Aliasing and the loss of uniqueness",
      "text": "Aliasing occurs when distinct continuous-time frequency components yield the same discrete-time samples. For complex exponentials,\n\n\\[e^{i2\\pi(f+kf_s)nT}=e^{i2\\pi fnT}\\]\n\nfor every integer \\(k\\). Sampling therefore identifies frequencies modulo \\(f_s\\). In a discrete-time spectrum, frequency is periodic; selecting a representative interval such as \\([-f_s/2,f_s/2)\\) does not recover which original continuous-time frequency produced a component.\n\nFor real sinusoids, positive and negative frequencies are also paired. A tone can consequently appear at a frequency obtained by shifting it by integer multiples of \\(f_s\\) and reflecting it into the interval \\([0,f_s/2]\\). With a sampling rate of \\(8\\) kHz, a \\(6\\) kHz cosine produces the same sample sequence as a \\(2\\) kHz cosine of corresponding phase. No processing based only on that sequence can decide which was present unless additional restrictions or measurements are supplied.\n\nFor a general spectrum, undersampling makes the shifted copies in the spectral-replication formula overlap. Their complex amplitudes add, including phase-dependent reinforcement or cancellation. This is more than an error in how a spectrum is displayed: the sampling operation becomes noninjective on the assumed signal class. Different input signals can map to exactly the same data. Once unwanted frequencies have folded into the desired band, an ordinary digital low-pass filter cannot separate them from legitimate components occupying the same frequencies.\n\nAliasing also occurs when a sequence is reduced in sampling rate. Keeping every \\(M\\)th sample changes the rate from \\(f_s\\) to \\(f_s/M\\). Unless the sequence has first been restricted to the new Nyquist band, spectral components that were previously distinguishable overlap after decimation. This is why reducing the number of samples is generally a filtering operation followed by sample selection, not sample selection alone.\n\nThe theorem’s bandwidth condition concerns the input at the instant of sampling, not just the intended signal. Sensor noise, radio-frequency interference, harmonics, and distortion generated upstream can all contribute energy above the Nyquist frequency. A nominally low-frequency experiment can therefore exhibit aliasing even when its desired waveform changes slowly. Conversely, a signal above \\(f_s/2\\) is not necessarily unrecoverable if its occupied band is known in advance and remains separable after sampling. That situation is bandpass sampling, with a different admissible signal class, rather than an exception to the uniqueness requirement."
    },
    {
      "heading": "Analog filtering and physical reconstruction",
      "text": "A practical acquisition chain commonly contains a sensor, an analog anti-aliasing filter, and an analog-to-digital converter. The filter attenuates unwanted high-frequency content before samples are taken. Its placement is essential: filtering after conversion can remove digital frequencies, but cannot undo the mixing of unknown analog components that has already occurred during sampling.\n\nAn ideal low-pass filter would pass one frequency interval perfectly and reject everything beyond a sharp boundary. Its impulse response extends indefinitely in both directions, making it unsuitable as an exactly realizable causal device with finite delay. Real filters have a passband, a transition region, and a stopband with finite attenuation. Consequently, an engineering specification normally separates the highest desired frequency \\(B\\) from \\(f_s/2\\). The available transition width affects filter order, delay, amplitude ripple, and how much unwanted energy remains to alias.\n\nDigital audio illustrates the distinction between theoretical and practical limits. A sampling rate of \\(44.1\\) kHz has a Nyquist frequency of \\(22.05\\) kHz. A desired passband ending at \\(20\\) kHz leaves approximately \\(2.05\\) kHz before that boundary, rather than allowing an ideal sharp cutoff at the edge of the audible passband. Many converter architectures operate internally at a much higher rate, use digital filtering to obtain a narrow transition, and then reduce the rate. The same principle appears in scientific instruments when modest analog filtering is combined with substantial oversampling.\n\nReconstruction hardware introduces a second filtering problem. A digital-to-analog converter may hold each sample value constant for one sampling period, producing a staircase waveform. A zero-order hold has a normalized frequency-response factor\n\n\\[e^{-i\\pi fT}\\operatorname{sinc}(fT).\\]\n\nIt therefore causes both a delay and a frequency-dependent attenuation, often called hold droop. The output also contains spectral images around multiples of the sampling frequency. A reconstruction or anti-imaging filter suppresses those images, while equalization may compensate for passband droop.\n\nThe ideal theorem uses instantaneous point values, whereas actual acquisition circuits have finite aperture times and settling behavior. A measurement that averages over a short interval is a filtered sample, not exactly \\(x(nT)\\). If that aperture response is known and remains nonzero throughout the desired band, compensation may be possible. Zeros in the response, inadequate settling, or clipping can destroy information in ways not described by the sampling-rate condition alone. Correct rate selection is thus one part of a signal-preservation chain, not a guarantee that every converter implementation is lossless."
    },
    {
      "heading": "Finite records, amplitude quantization, and timing errors",
      "text": "The reconstruction formula requires samples from the entire integer time axis. A recorded file supplies only finitely many of them. Truncating the sinc series discards contributions from outside the available interval; the error depends on those missing samples and on the kernel’s slow \\(1/|t|\\) decay. Near record boundaries, the absence of neighboring data is particularly important. No finite record determines an arbitrary finite-energy band-limited signal exactly without additional assumptions about its global form.\n\nFinite-length interpolators commonly multiply sinc by a window and retain a limited number of coefficients. This produces a realizable filter but changes its frequency response. Passband error, transition width, and stopband leakage then depend on the window and filter length. Likewise, observing a signal through a finite time window convolves its spectrum with the window’s transform. Spectral leakage in a finite-record spectrum is distinct from aliasing: leakage comes from time-windowing, whereas aliasing comes from the identification of frequencies under sampling.\n\nThe discrete Fourier transform describes a finite sequence using a finite set of frequency coefficients. Its ordinary interpretation includes periodic continuation of that sequence. Fourier interpolation of such data therefore reconstructs a periodic trigonometric polynomial, not automatically the original nonperiodic waveform. Discontinuities between the end and beginning of the repeated record can introduce strong high-frequency content. The fast Fourier transform accelerates the calculation of these coefficients but does not alter the assumptions or recover information absent from the samples.\n\nAmplitude quantization is separate from sampling in time. The mathematical theorem allows sample values of unlimited precision; a \\(b\\)-bit converter assigns them to finitely many levels. For an ideal uniform quantizer with step size \\(\\Delta\\), a frequently used approximation models the error as uniformly distributed over \\([-\\Delta/2,\\Delta/2]\\), with variance \\(\\Delta^2/12\\). Under the usual assumptions for an ideal full-scale sinusoid, the resulting signal-to-quantization-noise ratio is approximately \\(6.02b+1.76\\) decibels. The model is not universal: quantization error can correlate with the signal, especially for small or simple periodic inputs, and overload creates a different error mechanism.\n\nTiming uncertainty, or clock jitter, perturbs the sample location. For a small error \\(\\delta t_n\\),\n\n\\[x(nT+\\delta t_n)-x(nT)\\approx x'(nT)\\delta t_n.\\]\n\nHigh-frequency signals are more sensitive because their slopes can be larger. For a sinusoidal input of frequency \\(f\\) and small independent timing errors with root-mean-square value \\(\\sigma_t\\), a common approximation is\n\n\\[\\mathrm{SNR}_{\\mathrm{jitter}} \\approx -20\\log_{10}(2\\pi f\\sigma_t).\\]\n\nThese errors persist even when the nominal sampling rate comfortably exceeds the Nyquist rate."
    },
    {
      "heading": "Bandpass sampling, complex signals, and rate conversion",
      "text": "The familiar rule \\(f_s>2B\\) assumes a low-pass spectral support extending from zero to a highest frequency \\(B\\). A real bandpass signal instead occupies two mirrored intervals, \\([f_L,f_H]\\) and \\([-f_H,-f_L]\\), with positive-frequency bandwidth \\(W=f_H-f_L\\). Because the intervening frequencies are absent, it may be possible to sample below \\(2f_H\\) without losing information. The essential requirement remains that the occupied spectral replicas do not overlap.\n\nFor an ideal real bandpass signal with completely known support, standard admissible uniform sampling intervals can be written\n\n\\[\\frac{2f_H}{m}\\leq f_s\\leq \\frac{2f_L}{m-1},\\]\n\nfor integers \\(m\\geq2\\) for which the interval is nonempty. The ordinary high-rate case \\(f_s\\geq2f_H\\) corresponds to \\(m=1\\), with no upper bound. Boundary equalities require the same caution about endpoint components as in low-pass sampling. Although \\(2W\\) is a fundamental bandwidth scale, arbitrary rates above \\(2W\\) are not necessarily valid for a given band location.\n\nFor example, a real signal confined to \\(70\\)–\\(80\\) MHz can ideally be sampled at \\(30\\) million samples per second. Taking \\(m=5\\) gives an admissible interval from \\(32\\) to \\(35\\) MHz and does not establish this example; taking \\(m=6\\) gives\n\n\\[26\\tfrac{2}{3}\\ \\mathrm{MHz}\\leq f_s\\leq28\\ \\mathrm{MHz},\\]\n\nalso excluding \\(30\\) MHz. In fact, \\(30\\) MHz causes the positive and negative band images to overlap and is unsuitable. A rate of \\(34\\) MHz, within the \\(m=5\\) interval, avoids that overlap under the ideal support assumption. This illustrates why bandwidth alone cannot select an undersampling rate. Practical bandpass acquisition also requires an input circuit that can track the original high-frequency waveform and a filter that excludes other bands.\n\nComplex sampling changes how rates are counted. Quadrature receivers produce in-phase and quadrature components, combined as \\(z(t)=I(t)+iQ(t)\\). A complex signal occupying a spectral interval of width \\(W\\) can ideally be sampled at \\(W\\) complex samples per second, with appropriate boundary assumptions. Each complex sample contains two real numbers; this does not contradict the roughly \\(2W\\) real parameters per second associated with a real bandpass waveform.\n\nRate conversion applies the same spectral principles digitally. Increasing a rate by an integer factor inserts zeros and then applies a low-pass interpolation filter to suppress the resulting images. Decreasing it requires low-pass filtering before retaining a subset of samples. A rational conversion factor \\(L/M\\) combines these operations, often through a polyphase filter implementation that avoids calculating outputs that would later be discarded. Oversampling can ease filtering and support noise-shaping architectures, but the extra samples do not create new information about an already perfectly known band-limited waveform."
    },
    {
      "heading": "Nonuniform sampling and higher-dimensional extensions",
      "text": "Uniform spacing makes the classical theorem especially simple, but equal intervals are not the only possible sampling geometry. For nonuniform locations \\(t_n\\), recovery depends on their distribution as well as their average number per unit time. Dense clusters cannot compensate without limit for large gaps. Stable reconstruction is naturally studied using a frame: a possibly redundant system of measurement functions with upper and lower bounds controlling how measurement energy relates to signal energy.\n\nA classical sufficient result is Kadec’s one-quarter theorem. In normalized units, exponential functions with frequencies \\(\\lambda_n\\) form a Riesz basis on \\([-\\pi,\\pi]\\) if\n\n\\[\\sup_{n\\in\\mathbb Z}|\\lambda_n-n|<\\frac14.\\]\n\nThrough Fourier duality, this supplies a stability guarantee for band-limited sampling at locations that deviate from a critical uniform grid by less than one quarter of the sampling interval. The condition is sufficient, not a complete characterization of every stable irregular sampling set. Much more irregular arrangements can work under other hypotheses.\n\nSampling density gives broader necessary conditions. For stable recovery of an entire square-integrable signal class whose spectrum is supported on a set of finite measure, Landau’s density theory relates the required lower sampling density to that spectral measure. A real low-pass class supported on \\([-B,B]\\) has spectral measure \\(2B\\) in hertz, matching the familiar density scale. Density by itself is not generally sufficient: the geometry of the sample set and the required stability still matter.\n\nIn multiple dimensions, spectral copies are translated by the reciprocal lattice of the spatial sampling lattice. For a two-dimensional image supported within\n\n\\[|f_x|<B_x,\\qquad |f_y|<B_y,\\]\n\na rectangular grid with spacings \\(\\Delta_x,\\Delta_y\\) is sufficient when\n\n\\[\\Delta_x<\\frac{1}{2B_x},\\qquad \\Delta_y<\\frac{1}{2B_y}.\\]\n\nThe reconstruction kernel factors into sinc functions along the coordinate directions. For a circular spectral support, other lattice geometries can use samples more efficiently than a rectangular grid. Pixel integration and optical blur further modify the image before its spatial samples are recorded.\n\nOther extensions replace unrestricted band-limited spaces with more structured signal models. Compressed sensing reconstructs certain sparse signals from fewer measurements by exploiting sparsity and suitable measurement operators, often through convex optimization. Its measurements need not be uniform point samples, and its guarantees concern a restricted model rather than all signals within a bandwidth. Similarly, finite-rate-of-innovation methods exploit finite parameter descriptions, such as a limited number of pulses per unit time. These theories change the prior assumptions and measurement design; they do not recover arbitrary out-of-band information from an insufficient classical sample sequence."
    },
    {
      "heading": "Historical development and relation to information theory",
      "text": "The theorem developed through several mathematical and engineering traditions rather than a single discovery. Interpolation formulas involving cardinal sine functions appeared before modern digital electronics. E. T. Whittaker’s 1915 work on interpolation and cardinal functions is an important mathematical antecedent. J. M. Whittaker subsequently developed related interpolation theory. These contributions explain the name “Whittaker–Shannon interpolation,” which emphasizes the reconstruction formula rather than only the minimum sampling rate.\n\nHarry Nyquist’s work at Bell Telephone Laboratories addressed limits on telegraph signaling. His 1924 paper “Certain Factors Affecting Telegraph Speed” and 1928 paper “Certain Topics in Telegraph Transmission Theory” connected signaling speed, waveform structure, and transmission bandwidth. The ideal rate of \\(2B\\) independent real pulse amplitudes per second in a low-pass channel of bandwidth \\(B\\) is closely related to the sampling theorem. Nyquist’s original engineering questions concerned transmission and interference between symbols, however, and should not simply be identified with every part of the later reconstruction theorem.\n\nVladimir Kotelnikov presented a sampling result in 1933 in work on the transmission capacity of communication systems. His formulation established that a band-limited function could be represented through regularly spaced values and reconstructed by an interpolation series. Claude Shannon gave a prominent statement and explanation in his 1949 paper “Communication in the Presence of Noise.” The multiple historical lines account for names including Whittaker–Kotelnikov–Shannon and Nyquist–Shannon. They also reflect differences between priority for the interpolation identity, a communications interpretation, and the theorem’s later dissemination.\n\nThe connection to information theory is substantial but requires a distinction between samples and bits. A band-limited signal has roughly \\(2B\\tau\\) real degrees of freedom over a long interval of duration \\(\\tau\\), in the sense of effective time-frequency concentration. This is not an exact statement that a finite record contains precisely that many independent coordinates: strictly band-limited signals extend indefinitely, and endpoint effects matter. Nor does each ideal sample contain a fixed number of bits. An arbitrary real amplitude requires unbounded precision for exact specification.\n\nNoise and amplitude constraints lead to a different result, the Shannon–Hartley formula for channel capacity:\n\n\\[C=B\\log_2\\!\\left(1+\\frac{S}{N}\\right),\\]\n\nfor the standard band-limited additive white Gaussian noise channel, with signal power \\(S\\), in-band noise power \\(N\\), and capacity \\(C\\) in bits per second. Sampling describes how a restricted waveform can be represented without losing its continuous-time information; channel capacity bounds the rate of reliable communication through a noisy channel. The noisy-channel coding theorem concerns coding for that reliability. Raising the sampling rate alone neither increases the analog channel’s bandwidth nor removes its noise, so samples per second and reliably transmitted bits per second remain distinct quantities."
    }
  ],
  "links": [
    "interpolation",
    "fourier-transform",
    "sampling-theorem",
    "lp-space",
    "analytic-function",
    "convolution",
    "fourier-series",
    "orthonormal-basis",
    "hilbert-space",
    "cauchy-schwarz-inequality",
    "uniform-convergence",
    "aliasing",
    "discrete-fourier-transform",
    "fast-fourier-transform",
    "quantization-signal-processing",
    "variance",
    "convex-optimization",
    "claude-shannon",
    "information-theory",
    "channel-capacity",
    "noisy-channel-coding-theorem"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}