{
  "id": "additive-white-gaussian-noise",
  "title": "Additive white Gaussian noise",
  "lang": "en",
  "source_lang": "en",
  "summary": "Additive white Gaussian noise is a mathematical noise model combining additive interference, a flat power spectrum, and jointly Gaussian statistics.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "AWGN",
    "Gaussian white additive noise"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/additive-white-gaussian-noise/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/additive-white-gaussian-noise/",
      "json": "https://aiwiki.page/zh/additive-white-gaussian-noise.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Additive white Gaussian noise (AWGN) is a model of random disturbances in which noise adds to a signal, has constant power per unit bandwidth, and obeys Gaussian statistics. It is widely used in communications, signal processing, and statistical estimation. Its three defining properties describe different aspects of noise: additive specifies how noise enters a system, white specifies its spectral structure, and Gaussian specifies its probability law. The model provides a tractable approximation to many physical disturbances, particularly thermal noise over a restricted frequency range."
    },
    {
      "heading": "Definition and assumptions",
      "text": "A discrete-time additive-noise model is written as\n\n\\[Y_k=X_k+N_k,\\]\n\nwhere \\(X_k\\) is the input signal, \\(Y_k\\) is the observed output, and \\(N_k\\) is the noise. In the standard real-valued AWGN model,\n\n\\[N_k\\overset{\\mathrm{iid}}{\\sim}\\mathcal N(0,\\sigma^2),\\]\n\nmeaning that the noise samples are independent and identically distributed with a zero-mean normal distribution and variance \\(\\sigma^2\\). The noise process is also assumed independent of the input signal.\n\nThe three terms in the name have distinct meanings:\n\n- Additive: noise is added to the signal rather than multiplying it or changing its time scale. Additivity alone does not imply independence from the signal; independence is an additional assumption in the usual AWGN channel.\n- White: the noise has a constant power spectral density. For a zero-mean discrete-time process, this corresponds to zero autocovariance between distinct sample times.\n- Gaussian: every finite collection of noise samples has a joint Gaussian distribution, not merely Gaussian marginal distributions at individual times.\n\nThat last distinction matters. Gaussian one-dimensional marginals do not by themselves establish that a process is Gaussian. For jointly Gaussian samples, however, zero covariance implies independence. Thus discrete-time Gaussian white noise consists of independent Gaussian samples.\n\nThe zero-mean convention separates random fluctuations from a constant offset, which can instead be included in the deterministic part of the model."
    },
    {
      "heading": "Discrete-time mathematical description",
      "text": "For real, zero-mean AWGN,\n\n\\[\\mathbb E[N_k]=0, \\qquad \\mathbb E[N_kN_\\ell]=\\sigma^2\\delta_{k\\ell},\\]\n\nwhere \\(\\mathbb E\\) denotes expected value and \\(\\delta_{k\\ell}\\) is one when \\(k=\\ell\\) and zero otherwise.\n\nFor a vector of \\(m\\) samples,\n\n\\[\\mathbf N=(N_1,\\ldots,N_m)^{\\mathsf T},\\]\n\nthe distribution is\n\n\\[\\mathbf N\\sim\\mathcal N(\\mathbf 0,\\sigma^2 I_m),\\]\n\nwhere \\(I_m\\) is the identity matrix. Its covariance matrix is diagonal, and its probability density function is\n\n\\[p_{\\mathbf N}(\\mathbf n) = \\frac{1}{(2\\pi\\sigma^2)^{m/2}} \\exp\\!\\left( -\\frac{\\|\\mathbf n\\|^2}{2\\sigma^2} \\right).\\]\n\nThe density depends only on distance from the origin. Consequently, an orthogonal change of coordinates leaves this distribution unchanged. In signal processing, this means that projections of white Gaussian noise onto orthogonal unit-energy coordinates have independent Gaussian coefficients with equal variance.\n\nWith angular frequency measured in radians per sample, the discrete-time spectral density is constant:\n\n\\[S_N(e^{j\\omega})=\\sigma^2, \\qquad -\\pi\\leq\\omega<\\pi,\\]\n\nunder the convention\n\n\\[R_N[\\ell] = \\frac{1}{2\\pi} \\int_{-\\pi}^{\\pi} S_N(e^{j\\omega})e^{j\\omega\\ell}\\,d\\omega.\\]\n\nWhiteness concerns the theoretical spectrum. The measured spectrum of a finite noise record generally fluctuates and is not exactly flat."
    },
    {
      "heading": "Continuous-time white noise",
      "text": "Communications theory commonly writes real continuous-time AWGN formally as\n\n\\[Y(t)=X(t)+N(t),\\]\n\nwith two-sided noise spectral density\n\n\\[S_N(f)=\\frac{N_0}{2}, \\qquad -\\infty<f<\\infty.\\]\n\nThe corresponding autocorrelation is\n\n\\[R_N(\\tau)=\\frac{N_0}{2}\\delta(\\tau),\\]\n\nwhere \\(\\delta\\) is the Dirac delta. The spectral and autocorrelation descriptions are related by the Fourier transform.\n\nInfinite bandwidth and generalized processes\n\nAn exactly flat, nonzero spectrum over all frequencies has infinite integrated power:\n\n\\[\\int_{-\\infty}^{\\infty}S_N(f)\\,df=\\infty.\\]\n\nIdeal continuous-time white noise therefore cannot be treated as an ordinary pointwise, finite-variance stochastic process. It is instead understood as a generalized random process, using ideas from distribution theory.\n\nIts action on a real square-integrable test function \\(\\phi\\) is well defined:\n\n\\[N[\\phi]=\\int_{-\\infty}^{\\infty}N(t)\\phi(t)\\,dt,\\]\n\nwith\n\n\\[\\mathbb E[N[\\phi]]=0, \\qquad \\operatorname{Cov}(N[\\phi],N[\\psi]) = \\frac{N_0}{2} \\int_{-\\infty}^{\\infty}\\phi(t)\\psi(t)\\,dt.\\]\n\nThus filtered or integrated noise has finite variance even though an ideal instantaneous noise value does not. In particular, projections onto real orthonormal functions have independent Gaussian coefficients of variance \\(N_0/2\\).\n\nGaussian white noise is also described formally as the derivative of a Wiener process. This is a generalized derivative: Wiener paths do not have ordinary time derivatives.\n\nBand limitation and filtering\n\nReal receivers have finite bandwidth. If ideal white noise passes through a linear time-invariant filter with frequency response \\(H(f)\\), its output spectrum is\n\n\\[S_{\\mathrm{out}}(f) = \\frac{N_0}{2}|H(f)|^2,\\]\n\nand its variance, when the integral is finite, is\n\n\\[\\sigma_{\\mathrm{out}}^2 = \\frac{N_0}{2} \\int_{-\\infty}^{\\infty}|H(f)|^2\\,df.\\]\n\nFiltering preserves Gaussianity but generally destroys whiteness.\n\nFor an ideal real low-pass filter passing \\(|f|\\leq B\\) with unit gain,\n\n\\[\\sigma_{\\mathrm{out}}^2=N_0B.\\]\n\nIts output autocorrelation is\n\n\\[R_{\\mathrm{out}}(\\tau) = N_0B\\,\\operatorname{sinc}(2B\\tau),\\]\n\nwhere \\(\\operatorname{sinc}(u)=\\sin(\\pi u)/(\\pi u)\\). Samples spaced \\(1/(2B)\\) apart are therefore uncorrelated and, because they are jointly Gaussian, independent. Samples at arbitrary spacings need not be independent."
    },
    {
      "heading": "Complex baseband representation",
      "text": "Digital communications often represent signals using complex baseband coordinates. A common noise model is\n\n\\[N_k=N_{I,k}+jN_{Q,k},\\]\n\nwhere the in-phase and quadrature components are independent and satisfy\n\n\\[N_{I,k},N_{Q,k} \\sim\\mathcal N\\!\\left(0,\\frac{\\sigma_c^2}{2}\\right).\\]\n\nThen\n\n\\[\\mathbb E[|N_k|^2]=\\sigma_c^2,\\]\n\nand the noise is written\n\n\\[N_k\\sim\\mathcal{CN}(0,\\sigma_c^2).\\]\n\nIts density is\n\n\\[p_N(n) = \\frac{1}{\\pi\\sigma_c^2} \\exp\\!\\left(-\\frac{|n|^2}{\\sigma_c^2}\\right).\\]\n\nThis distribution is circularly symmetric: multiplication by a complex number of unit magnitude does not change its distribution.\n\nNoise-variance conventions require care. A variance quoted per real component is half the total complex variance under this convention. Likewise, one-sided and two-sided spectral densities differ by a factor of two for real noise. Equations involving \\(N_0\\) are meaningful only with the accompanying bandwidth and coordinate conventions."
    },
    {
      "heading": "Physical interpretation",
      "text": "A major physical motivation for AWGN is thermal noise, produced by thermal fluctuations in electrical systems. Within a suitable operating frequency range, its spectrum can be approximately flat, while its amplitude statistics are approximately Gaussian.\n\nThe Gaussian approximation is often motivated by the central limit theorem: many small, sufficiently weakly dependent contributions can combine into an approximately normal disturbance. This argument concerns amplitude statistics; it does not establish whiteness. Temporal correlations depend on the physical source and on the system through which the noise passes.\n\nThe model is therefore an approximation over specified time scales, bandwidths, and operating conditions. A disturbance can be approximately Gaussian without being white, or approximately white without being Gaussian."
    },
    {
      "heading": "Detection and estimation",
      "text": "AWGN makes many inference problems analytically tractable. Suppose\n\n\\[\\mathbf Y=\\mathbf s(\\theta)+\\mathbf N, \\qquad \\mathbf N\\sim\\mathcal N(\\mathbf 0,\\sigma^2I),\\]\n\nwhere \\(\\theta\\) is an unknown parameter and \\(\\sigma^2\\) is fixed. The conditional density is proportional to\n\n\\[p(\\mathbf y\\mid\\theta) \\propto \\exp\\!\\left( -\\frac{\\|\\mathbf y-\\mathbf s(\\theta)\\|^2}{2\\sigma^2} \\right).\\]\n\nConsequently, maximum-likelihood estimation is equivalent to minimizing squared Euclidean error:\n\n\\[\\widehat{\\theta}_{\\mathrm{ML}} = \\operatorname*{arg\\,min}_{\\theta} \\|\\mathbf y-\\mathbf s(\\theta)\\|^2.\\]\n\nThis gives a probabilistic basis for least-squares methods under independent Gaussian errors of equal variance.\n\nFor equally likely transmitted signal vectors, optimal minimum-error detection likewise selects the candidate nearest to the received vector in Euclidean distance. With unequal prior probabilities, the decision rule also includes the priors.\n\nIf Gaussian noise instead has a known positive-definite covariance matrix \\(C\\), the relevant quadratic distance becomes\n\n\\[(\\mathbf y-\\mathbf s)^{\\mathsf T} C^{-1} (\\mathbf y-\\mathbf s).\\]\n\nSuch noise is not white unless \\(C\\) is proportional to the identity. A suitable linear transformation can whiten it, but that transformation must also be applied to the signal model."
    },
    {
      "heading": "The AWGN channel and capacity",
      "text": "The AWGN channel is a fundamental model in information theory. For the real discrete-time channel\n\n\\[Y=X+N, \\qquad N\\sim\\mathcal N(0,\\sigma^2),\\]\n\nwith independent noise and average input-power constraint\n\n\\[\\mathbb E[X^2]\\leq P,\\]\n\nthe capacity is\n\n\\[C=\\frac{1}{2} \\log_2\\!\\left(1+\\frac{P}{\\sigma^2}\\right)\\]\n\nbits per real channel use. A zero-mean Gaussian input of variance \\(P\\) achieves the maximum mutual information under these assumptions.\n\nFor a real band-limited AWGN channel of bandwidth \\(B\\) hertz, total average signal power \\(P\\), and noise power \\(N_0B\\), the corresponding formula is\n\n\\[C = B\\log_2\\!\\left(1+\\frac{P}{N_0B}\\right)\\]\n\nbits per second. This is the Shannon–Hartley formula, associated with Claude Shannon’s development of mathematical communication theory.\n\nThe ratio of signal power to noise power is the signal-to-noise ratio. Capacity describes an asymptotic limit on reliable communication, not the throughput of a particular modulation or code. Finite block lengths, restricted signal constellations, latency constraints, and imperfect channel knowledge can reduce achievable rates."
    },
    {
      "heading": "Simulation and limitations",
      "text": "A real discrete-time AWGN sequence can be simulated by generating independent standard normal samples \\(Z_k\\) and setting\n\n\\[N_k=\\sigma Z_k.\\]\n\nFor circular complex noise with total variance \\(\\sigma_c^2\\),\n\n\\[N_k = \\sqrt{\\frac{\\sigma_c^2}{2}} \\left(Z_{1,k}+jZ_{2,k}\\right),\\]\n\nwhere both real sequences are independent standard normal sequences. Such simulations are widely used in Monte Carlo evaluations of estimation error and communication performance.\n\nA sampled simulation has finite bandwidth. Its sample variance must be related to the intended continuous-time spectral density through the sampling and filtering conventions; variance and spectral density are not interchangeable quantities.\n\nImportant departures from AWGN include:\n\n- Colored noise: frequency-dependent spectral power and correlated samples.\n- Impulsive noise: occasional large disturbances that Gaussian tails may substantially underestimate.\n- Signal-dependent noise: disturbance strength or distribution varies with the signal.\n- Multiplicative disturbances: random gain or phase changes, rather than purely additive errors.\n- Nonstationary noise: statistics change with time.\n- Structured interference: unwanted signals may have periodicity, modulation, or dependence not captured by independent Gaussian samples.\n\nAWGN can coexist with other channel effects. A model such as\n\n\\[\\mathbf Y=H\\mathbf X+\\mathbf N\\]\n\ncan include a nontrivial linear channel \\(H\\) while retaining an AWGN disturbance. The noise assumption alone does not imply that the complete system is distortion-free or memoryless."
    }
  ],
  "links": [
    "normal-distribution",
    "variance",
    "multivariate-normal-distribution",
    "statistical-independence",
    "expected-value",
    "identity-matrix",
    "covariance-matrix",
    "probability-density-function",
    "fourier-transform",
    "stochastic-process",
    "wiener-process",
    "complex-number",
    "central-limit-theorem",
    "maximum-likelihood-estimation",
    "least-squares",
    "information-theory",
    "channel-capacity",
    "mutual-information",
    "claude-shannon",
    "signal-to-noise-ratio",
    "monte-carlo-method"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}