{
  "id": "additive-white-gaussian-noise-channel",
  "title": "Additive white Gaussian noise channel",
  "lang": "en",
  "source_lang": "en",
  "summary": "An additive white Gaussian noise channel models communication by adding signal-independent Gaussian noise with a flat power spectrum to the transmitted signal.",
  "category": "computer-science",
  "subcategory": "algorithms-theory",
  "type": "concept",
  "aliases": [
    "AWGN channel",
    "Gaussian noise channel"
  ],
  "generated_at": "2026-11-06",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/additive-white-gaussian-noise-channel/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/additive-white-gaussian-noise-channel/",
      "json": "https://aiwiki.page/zh/additive-white-gaussian-noise-channel.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "An additive white Gaussian noise channel, usually abbreviated AWGN channel, is a mathematical model of a communication channel in which the received signal equals the transmitted signal plus signal-independent noise that has Gaussian statistics and a flat power spectrum. It is a fundamental model in information theory and digital communications, used to analyze reliable transmission, receiver performance, and error-correcting codes. Although it omits many physical impairments, its mathematical tractability makes it an important reference against which more complicated channels are compared."
    },
    {
      "heading": "Meaning of the model",
      "text": "The three qualifiers describe distinct assumptions:\n\n- Additive: noise is added to the signal rather than multiplying it or changing its timing.\n- White: the noise has constant power spectral density across the frequencies represented by the model. Equivalently, ideal continuous-time white noise has an autocorrelation proportional to a Dirac delta.\n- Gaussian: every finite collection of noise coordinates has a joint Gaussian distribution.\n\nThe noise is also assumed to be statistically independent of the transmitted signal. This assumption is essential: Gaussian-looking disturbances whose strength depends on the signal need not constitute an AWGN channel.\n\nWhiteness and Gaussianity are different properties. White noise need not be Gaussian, and Gaussian noise need not be white. For jointly Gaussian noise, however, uncorrelated coordinates are independent. This is why an orthogonal signal-space representation of white Gaussian noise produces independent Gaussian noise components."
    },
    {
      "heading": "Discrete-time formulation",
      "text": "The real-valued, discrete-time channel is written\n\n\\[Y_k=X_k+Z_k,\\]\n\nwhere \\(X_k\\) is the transmitted symbol, \\(Y_k\\) is the received symbol, and the noise variables satisfy\n\n\\[Z_k\\overset{\\mathrm{iid}}{\\sim}\\mathcal N(0,\\sigma^2).\\]\n\nHere, “iid” means independent and identically distributed, and \\(\\sigma^2\\) is the noise variance. The sequence of noise variables is independent of the complete transmitted sequence.\n\nFor one channel use, the conditional probability density is\n\n\\[p(y\\mid x) = \\frac{1}{\\sqrt{2\\pi\\sigma^2}} \\exp\\left[-\\frac{(y-x)^2}{2\\sigma^2}\\right].\\]\n\nIndependence gives the memoryless property:\n\n\\[p(y_1,\\ldots,y_n\\mid x_1,\\ldots,x_n) = \\prod_{k=1}^{n}p(y_k\\mid x_k).\\]\n\nThus, conditioned on the current transmitted symbol, the current output does not depend on earlier inputs or outputs. This does not require the input symbols themselves to be independent: a code generally introduces dependencies among symbols.\n\nInput constraints\n\nA finite capacity requires a constraint on the input. The usual assumption is an average-power constraint,\n\n\\[\\frac{1}{n}\\sum_{k=1}^{n}\\mathbb E[X_k^2]\\leq P,\\]\n\nwhere \\(\\mathbb E\\) denotes expected value. Without a power or amplitude constraint, the AWGN channel has unbounded capacity because distinguishable signal levels can be separated arbitrarily far relative to the fixed noise scale.\n\nA peak-amplitude constraint or a restriction to a finite signal constellation defines a different capacity problem. The familiar Gaussian-input capacity formula does not automatically apply under those restrictions.\n\nComplex-valued formulation\n\nFor quadrature modulation, the channel is commonly represented using complex numbers:\n\n\\[Y_k=X_k+Z_k,\\qquad Z_k\\sim\\mathcal{CN}(0,N).\\]\n\nUnder the usual circularly symmetric convention,\n\n\\[\\mathbb E[|Z_k|^2]=N,\\]\n\nand the real and imaginary components are independent Gaussian variables, each with variance \\(N/2\\). A complex channel use therefore contains two real dimensions. This distinction explains many apparent factors of two in noise and capacity formulas."
    },
    {
      "heading": "Continuous-time noise and bandwidth",
      "text": "The ideal continuous-time model is\n\n\\[y(t)=x(t)+n(t),\\]\n\nwith two-sided noise power spectral density\n\n\\[S_n(f)=\\frac{N_0}{2}.\\]\n\nIts formal autocorrelation is\n\n\\[R_n(\\tau)=\\frac{N_0}{2}\\delta(\\tau).\\]\n\nPerfectly white noise over an unlimited frequency range would have infinite total power. Consequently, ideal white noise is treated mathematically as a generalized random process, rather than an ordinary random function with finite pointwise variance.\n\nPhysical receivers have finite bandwidth. If white noise passes through a filter with frequency response \\(H(f)\\), the output noise power is\n\n\\[P_n=\\frac{N_0}{2} \\int_{-\\infty}^{\\infty}|H(f)|^2\\,df.\\]\n\nThe filtered noise remains Gaussian but is generally correlated in time and is not white over the entire frequency axis. A suitable projection onto orthonormal waveforms, however, produces independent Gaussian noise coordinates. This supplies the connection between the continuous-time model and the discrete-time channel.\n\nFor a real channel with bandwidth \\(B\\) hertz and conventional two-sided noise density \\(N_0/2\\), the in-band noise power is \\(N_0B\\). Statements involving \\(N_0\\), bandwidth, or sample variance must specify whether they use real or complex signals and one-sided or two-sided spectral densities."
    },
    {
      "heading": "Physical motivation",
      "text": "A major motivation is thermal noise, which arises from thermal fluctuations in electrical systems. Over a suitable frequency range and under ordinary operating conditions, it can be approximated as Gaussian with a nearly flat spectrum.\n\nGaussian models are also supported by the central limit theorem: the sum of many sufficiently small, weakly dependent disturbances can have an approximately Gaussian distribution. This argument does not establish whiteness, which is a separate statement about temporal correlation or spectral shape.\n\nAn AWGN channel usually abstracts away propagation loss, distortion, and receiver details. A known constant channel gain may be absorbed into the signal normalization, leaving an equivalent model with a rescaled noise variance. Time-varying attenuation, multipath propagation, impulsive interference, and nonlinear distortion require additional modeling."
    },
    {
      "heading": "Channel capacity",
      "text": "Channel capacity is the supremum of reliably achievable communication rates under a specified input constraint. For the real discrete-time AWGN channel with average input power \\(P\\),\n\n\\[C=\\frac{1}{2}\\log_2\\left(1+\\frac{P}{\\sigma^2}\\right)\\]\n\nin bits per real channel use.\n\nThe ratio \\(P/\\sigma^2\\) is the signal-to-noise ratio under this normalization. Capacity is achieved, in the information-theoretic sense, by a zero-mean Gaussian input with variance \\(P\\).\n\nThe result follows by maximizing mutual information:\n\n\\[I(X;Y)=h(Y)-h(Y\\mid X)=h(Y)-h(Z),\\]\n\nwhere \\(h\\) denotes differential entropy. The noise entropy is fixed, and a Gaussian distribution maximizes differential entropy for a given variance. A Gaussian input makes the output Gaussian and attains this maximum.\n\nFor a circularly symmetric complex AWGN channel with \\(\\mathbb E[|X|^2]\\leq P\\) and \\(\\mathbb E[|Z|^2]=N\\),\n\n\\[C=\\log_2\\left(1+\\frac{P}{N}\\right)\\]\n\nin bits per complex channel use.\n\nBandlimited capacity\n\nThe corresponding continuous-time result is the Shannon–Hartley theorem:\n\n\\[C=B\\log_2\\left(1+\\frac{P}{N_0B}\\right)\\]\n\nin bits per second, for bandwidth \\(B\\), average received signal power \\(P\\), and in-band noise power \\(N_0B\\).\n\nThis is an asymptotic limit. Rates below capacity can be approached with arbitrarily small error probability using sufficiently long, suitably designed codes. The theorem does not imply zero error for a finite codeword or prescribe a particular modulation, code, latency, or decoding complexity.\n\nIf only \\(M\\) distinct input symbols are available, the mutual information cannot exceed \\(\\log_2M\\) bits per symbol. A finite constellation therefore does not achieve unrestricted Gaussian-input capacity at all signal-to-noise ratios.\n\nSpectral and energy efficiency\n\nLet the information rate be \\(R\\), the spectral efficiency be \\(\\eta=R/B\\), and the energy per information bit be \\(E_b=P/R\\). At the capacity boundary,\n\n\\[\\frac{E_b}{N_0} = \\frac{2^\\eta-1}{\\eta}.\\]\n\nAs spectral efficiency approaches zero,\n\n\\[\\lim_{\\eta\\to0}\\frac{E_b}{N_0}=\\ln 2,\\]\n\napproximately \\(-1.59\\) dB. This is the asymptotic minimum energy-per-bit ratio for reliable communication in the standard power-constrained AWGN model. It is not an attainable threshold for every finite-bandwidth, finite-length implementation."
    },
    {
      "heading": "Detection and error probability",
      "text": "For equally likely candidate signal vectors, maximum-likelihood detection selects the vector nearest to the received vector in Euclidean distance:\n\n\\[\\hat{\\mathbf x} = \\operatorname*{arg\\,min}_{\\mathbf x} \\|\\mathbf y-\\mathbf x\\|^2.\\]\n\nThis rule follows directly from the Gaussian likelihood. If candidates have unequal prior probabilities, a minimum-error detector also accounts for those priors.\n\nA matched filter or equivalent correlator extracts relevant signal coordinates for detecting known waveforms in white Gaussian noise. With suitable orthogonal signaling coordinates, the resulting detection problem reduces to a finite-dimensional vector-plus-noise model.\n\nFor coherent, uncoded binary phase-shift keying with ideal synchronization,\n\n\\[P_b=Q\\left(\\sqrt{\\frac{2E_b}{N_0}}\\right),\\]\n\nwhere \\(P_b\\) is the bit error probability and\n\n\\[Q(a)=\\frac{1}{\\sqrt{2\\pi}} \\int_a^\\infty e^{-u^2/2}\\,du.\\]\n\nThis expression depends on its stated assumptions; it is not the error formula for every binary communication system.\n\nAWGN likelihoods also provide soft information for decoding. If \\(X\\in\\{-A,+A\\}\\) has equal priors, then\n\n\\[\\log\\frac{p(Y=y\\mid X=+A)} {p(Y=y\\mid X=-A)} = \\frac{2Ay}{\\sigma^2}.\\]\n\nUnlike a hard binary decision, this likelihood ratio preserves information about confidence in the received symbol."
    },
    {
      "heading": "Simulation and benchmarking",
      "text": "A discrete-time AWGN simulation adds independent Gaussian samples to a transmitted sequence. If the average real-sample signal power is \\(P_s\\) and the desired linear signal-to-noise ratio is \\(\\gamma\\), the required noise variance is\n\n\\[\\sigma^2=\\frac{P_s}{\\gamma}.\\]\n\nFor complex samples, if total noise power is \\(N\\), each quadrature must have variance \\(N/2\\). Independent standard-normal samples can be generated using a pseudorandom number generator and scaled accordingly.\n\nCommon sources of incorrect comparisons include confusing:\n\n- sample signal-to-noise ratio with \\(E_b/N_0\\);\n- information-bit energy with coded-bit energy;\n- real-component variance with total complex noise power;\n- symbol-rate samples with oversampled waveform samples;\n- bandwidth with the noise-equivalent bandwidth of a receiver filter.\n\nAWGN tests isolate the effects of modulation, coding, and detection. They are widely used as reproducible baselines, but their results describe performance under the assumed channel rather than under every deployment condition."
    },
    {
      "heading": "Historical role and limitations",
      "text": "Shannon’s 1948 work on communication theory established the framework for channel capacity and the central importance of Gaussian channels. Subsequent coding and modulation research has used the AWGN channel as a reference model for studying the gap between practical systems and theoretical limits.\n\nIts principal limitations follow directly from its assumptions. It does not by itself represent time-varying multipath fading, frequency-selective distortion, synchronization errors, clipping, quantization, or interference with non-Gaussian statistics. Correlated Gaussian noise requires a colored-noise model; transforming such noise into white coordinates may also transform the signal and its constraints.\n\nAWGN performance is therefore a benchmark, not a universal guarantee or bound for arbitrary physical channels. Its usefulness depends on whether independent additive Gaussian noise is an adequate description of the dominant disturbance at the receiver."
    }
  ],
  "links": [
    "information-theory",
    "error-correcting-code",
    "normal-distribution",
    "variance",
    "expected-value",
    "complex-number",
    "central-limit-theorem",
    "channel-capacity",
    "signal-to-noise-ratio",
    "mutual-information",
    "differential-entropy",
    "shannon-hartley-theorem",
    "pseudorandom-number-generator"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}