{
  "id": "one-over-f-noise",
  "title": "1/f Noise",
  "lang": "en",
  "source_lang": "en",
  "summary": "1/f noise is a fluctuation pattern whose spectral power decreases approximately inversely with frequency, appearing in electronic devices and many complex systems.",
  "category": "science",
  "subcategory": "physics",
  "type": "concept",
  "aliases": [
    "Flicker noise",
    "Pink noise",
    "One-over-f noise"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/one-over-f-noise/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/one-over-f-noise/",
      "json": "https://aiwiki.page/zh/one-over-f-noise.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "1/f noise is a form of fluctuation in which the power spectral density varies approximately as the inverse of frequency over a specified range. Its defining relation is \\(S(f)\\propto 1/f\\): slower fluctuations contribute more spectral power per unit frequency than faster ones. In broader usage, the name includes spectra proportional to \\(1/f^\\alpha\\), where the exponent \\(\\alpha\\) is close to 1. The terms flicker noise and pink noise overlap with 1/f noise, although flicker noise usually refers to low-frequency device fluctuations, whereas pink noise often denotes a deliberately generated signal.\n\nThe phenomenon matters both as a practical limitation and as a problem in statistical physics. It restricts the stability of amplifiers, sensors, oscillators, and quantum devices, while similar spectral patterns occur in biological recordings, geophysical measurements, and human activity. A common spectral shape does not establish a common physical cause: broad distributions of relaxation times, trapping processes, interacting defects, and collective dynamics can all produce approximately inverse-frequency spectra. Understanding a particular case therefore requires both statistical characterization and evidence about its underlying mechanism."
    },
    {
      "heading": "Spectral definition and physical meaning",
      "text": "For a measured quantity \\(x(t)\\), such as voltage, the power spectral density \\(S_x(f)\\) describes how its mean-square fluctuations are distributed across frequency. With a one-sided spectrum, defined for \\(f\\geq 0\\), the contribution to the variance from a frequency interval is\n\n\\[\\sigma_x^2(f_1,f_2)=\\int_{f_1}^{f_2}S_x(f)\\,df.\\]\n\nIf \\(x\\) is a voltage, the spectral density has units of \\(\\mathrm{V^2/Hz}\\). Its square root, the amplitude spectral density, has units of \\(\\mathrm{V/\\sqrt{Hz}}\\). This distinction is essential: a power spectrum proportional to \\(1/f\\) has an amplitude spectrum proportional to \\(1/\\sqrt f\\), not \\(1/f\\).\n\nAn exact inverse-frequency spectrum over a finite band can be written\n\n\\[S_x(f)=\\frac{A}{f}, \\qquad f_{\\mathrm L}<f<f_{\\mathrm H}.\\]\n\nIts integrated contribution is\n\n\\[\\sigma_x^2=A\\ln\\!\\left(\\frac{f_{\\mathrm H}}{f_{\\mathrm L}}\\right).\\]\n\nConsequently, every frequency interval with the same ratio of upper to lower limits contributes the same variance. The bands 1–10 Hz and 10–100 Hz contribute equally, as do successive octaves. This equal contribution per octave explains the analogy with pink light: compared with a spectrally flat signal, more weight lies toward the low-frequency end. For example, a voltage spectrum with \\(A=10^{-12}\\,\\mathrm{V^2}\\) contributes approximately \\(2.30\\times10^{-12}\\,\\mathrm{V^2}\\) in each decade, corresponding to about \\(1.52\\,\\mathrm{\\mu V}\\) root-mean-square voltage.\n\nThe more general empirical model is\n\n\\[S_x(f)=C f^{-\\alpha}.\\]\n\nOn logarithmic axes, its slope is \\(-\\alpha\\). White noise has \\(\\alpha=0\\); ideal pink noise has \\(\\alpha=1\\); and an ideal random walk has a spectrum proportional to \\(f^{-2}\\), subject to qualifications about stationarity and finite observation. These labels describe spectral behavior rather than complete physical models. Different processes can have the same exponent but different amplitude distributions, temporal asymmetries, or responses to external perturbations.\n\nA strict \\(1/f\\) law cannot extend from zero frequency to infinite frequency in an ordinary finite-variance stationary process, because its integral diverges at both ends. Real measurements always concern a bounded interval. At sufficiently high frequency, microscopic response times or instrument bandwidth change the spectrum; at low frequency, the spectrum may flatten, change exponent, or remain unresolved because the observation is too short. The phrase “1/f noise” therefore normally means an approximate scaling regime, not an exact law at all frequencies."
    },
    {
      "heading": "Temporal correlations, stationarity, and averaging",
      "text": "The temporal interpretation of a noise spectrum depends on the properties of the underlying stochastic process. A process is weakly stationary when its mean is constant and its covariance depends on time separation rather than absolute time. Under appropriate conditions, the Wiener–Khinchin theorem relates the spectrum to the autocorrelation function through a Fourier transform. This connects enhanced low-frequency power with fluctuations that persist over relatively long times.\n\nFor a one-sided, band-limited inverse-frequency spectrum, the covariance is\n\n\\[C_x(\\tau)= A\\int_{f_{\\mathrm L}}^{f_{\\mathrm H}} \\frac{\\cos(2\\pi f\\tau)}{f}\\,df.\\]\n\nBetween the timescales set by the two cutoffs, this expression can vary approximately logarithmically with lag. Such behavior differs from a single exponential relaxation, which has one characteristic correlation time. A broad 1/f region instead indicates that many timescales contribute, or that no single timescale dominates the observed range. The precise covariance still depends on the cutoffs and on the spectrum outside that range.\n\nLarge low-frequency power complicates averaging. For independent measurements, the variance of their mean decreases inversely with the number of observations. That familiar rule does not apply unchanged when successive observations remain correlated. If a continuous record is averaged for duration \\(T\\), the averaging operation has squared frequency response\n\n\\[|H_T(f)|^2= \\left[\\frac{\\sin(\\pi fT)}{\\pi fT}\\right]^2.\\]\n\nThe variance of that average is obtained by integrating \\(S_x(f)|H_T(f)|^2\\). Averaging strongly suppresses fast fluctuations but passes fluctuations much slower than \\(T\\). A measurement can therefore become insensitive to additional short-term samples while remaining affected by slow changes in its baseline.\n\nStationarity must not be inferred from a straight line on a log–log plot. A drifting mean, intermittent events, or a random walk can generate strong low-frequency power without being stationary. Conversely, a stationary process with a very long but finite correlation time may resemble 1/f noise throughout an accessible experiment. Whether extending the record reveals a stable low-frequency cutoff is an empirical question.\n\nMathematical models also require care. For fractional Brownian motion, the position-like process is nonstationary, whereas its increments form fractional Gaussian noise. In the standard model, the increment spectrum behaves at low frequency as \\(f^{1-2H}\\), with \\(H\\) the Hurst parameter. Its relation to 1/f-like behavior differs from the spectral scaling assigned to the integrated process. Thus a spectral exponent, a Hurst exponent, and a statement about long-range dependence are not interchangeable without specifying which process is being analyzed."
    },
    {
      "heading": "Discovery and development of the problem",
      "text": "The experimental history of 1/f noise began in electronic devices. In 1925, John B. Johnson reported low-frequency fluctuations in vacuum-tube currents whose intensity increased toward lower frequencies. These fluctuations were distinct from the noise expected from the discrete passage of charge. Walter Schottky subsequently developed an early explanation involving fluctuating surface conditions associated with the emitting cathode. The term flicker effect became associated with this excess low-frequency noise.\n\nThis history overlaps with, but is separate from, the discovery of thermal noise. Johnson’s 1928 measurements and Harry Nyquist’s theoretical treatment established the equilibrium voltage fluctuations of an electrical resistor. In the classical regime, their one-sided voltage spectrum is \\(4k_{\\mathrm B}TR\\), where \\(k_{\\mathrm B}\\) is the Boltzmann constant, \\(T\\) is absolute temperature, and \\(R\\) is electrical resistance. Thermal noise is approximately white within the frequency range where this expression applies. Flicker noise commonly appears as an additional low-frequency component, especially when a device is biased.\n\nThe growth of semiconductor electronics made the problem technologically urgent. Transistors and integrated circuits brought active interfaces, insulating layers, and small conducting volumes into increasingly sensitive applications. In 1957, A. L. McWhorter developed a trapping-based account of low-frequency noise associated with semiconductor surfaces. The approach connected a wide range of switching times to the distance-dependent probability of quantum tunneling between a conducting region and nearby traps.\n\nIn 1969, F. N. Hooge introduced a widely used empirical relation connecting normalized resistance or current noise to carrier number. It encouraged systematic comparison among materials and devices, while also prompting debate about whether mobility fluctuations or carrier-number fluctuations were dominant. These explanations are not mutually exclusive in all devices, and the interpretation of a fitted noise parameter depends on device geometry, transport, and operating conditions.\n\nTheoretical work broadened beyond electronics during the following decades. In 1981, P. Dutta and P. M. Horn reviewed how distributions of thermally activated processes could account for many observed low-frequency spectra and their temperature dependence. In 1987, Per Bak, Chao Tang, and Kurt Wiesenfeld introduced self-organized criticality, linking scale-free collective dynamics with the search for explanations of 1/f behavior. These developments established several useful mechanisms rather than a single universal theory. The historical problem evolved from explaining one device imperfection into determining which physical and statistical models apply to each observed scaling regime."
    },
    {
      "heading": "Microscopic mechanisms and broad relaxation spectra",
      "text": "One of the most direct routes to 1/f noise is the superposition of many fluctuators with widely distributed relaxation times. A single variable that relaxes exponentially has a Lorentzian spectrum. If its covariance is \\(\\sigma^2e^{-|\\tau|/\\tau_c}\\), its one-sided spectrum is\n\n\\[S(f)= \\frac{4\\sigma^2\\tau_c} {1+(2\\pi f\\tau_c)^2}.\\]\n\nIt is approximately white below the corner frequency \\(1/(2\\pi\\tau_c)\\) and decreases as \\(f^{-2}\\) above it. A single such process therefore does not produce a broad 1/f spectrum.\n\nNow suppose statistically independent fluctuators contribute comparable variance per logarithmic interval of relaxation time. Their combined spectrum has the form\n\n\\[S_{\\mathrm{tot}}(f)\\propto \\int_{\\tau_{\\min}}^{\\tau_{\\max}} \\frac{d\\tau}{1+(2\\pi f\\tau)^2}.\\]\n\nFor frequencies well between the inverse limiting times, the integral is approximately proportional to \\(1/f\\). Physically, different fluctuators dominate at different frequencies, and no narrow group supplies a unique characteristic time. Unequal amplitudes, finite limits, or a nonuniform distribution shift the exponent or introduce curvature.\n\nThermally activated defects provide a concrete way to obtain this distribution. A typical relaxation time follows\n\n\\[\\tau=\\tau_0 \\exp\\!\\left(\\frac{E_a}{k_{\\mathrm B}T}\\right),\\]\n\nwhere \\(E_a\\) is an activation energy. A moderately broad range of barriers can therefore generate an enormous span of times. If the appropriately weighted barrier distribution changes slowly over the energies sampled by an experiment, the spectrum can approximate 1/f. Changing temperature moves different defects into the measurement band, giving a test of the mechanism rather than merely another fitted exponent.\n\nCharge trapping provides another route. A defect that alternately captures and releases an electron may produce random telegraph noise, visible as switching between two signal levels. One dominant trap gives a Lorentzian spectrum; many traps can give an approximate inverse-frequency continuum. If tunneling rates vary exponentially with distance from an interface, even a relatively simple spatial trap distribution can span many relaxation times.\n\nOther mechanisms involve interacting defects, magnetic domain motion, fluctuating conducting paths, or collective avalanches. A material near a phase transition may exhibit broad relaxation associated with critical dynamics. Disordered conductors may amplify the influence of a few unstable links in a current-carrying network. These mechanisms cannot be identified solely from \\(\\alpha\\approx1\\). Higher-order statistics, responses to temperature and bias, sample-size dependence, and direct observations of switching are needed to distinguish a sum of nearly independent fluctuators from correlated collective motion."
    },
    {
      "heading": "Electronic devices and low-frequency transport",
      "text": "In electrical conductors, low-frequency noise is often measured as voltage fluctuations under fixed current or current fluctuations under fixed voltage. For a device obeying Ohm’s law and measured at constant current \\(I\\), a small resistance fluctuation gives \\(\\delta V=I\\delta R\\). Consequently,\n\n\\[S_V(f)=I^2S_R(f).\\]\n\nThis explains why resistance noise often scales with the square of the applied bias in a sufficiently weakly driven regime. It also distinguishes resistance fluctuations from a bias-independent additive voltage source. At stronger bias, Joule heating, nonlinear transport, and changing trap populations can invalidate the simple scaling.\n\nHooge’s empirical relation is commonly written\n\n\\[\\frac{S_R(f)}{R^2} = \\frac{S_I(f)}{I^2} = \\frac{\\alpha_{\\mathrm H}}{Nf},\\]\n\nwhere \\(N\\) is the number of mobile carriers and \\(\\alpha_{\\mathrm H}\\) is the Hooge parameter. The equality between the normalized resistance and current spectra assumes the appropriate fixed-voltage measurement and small fluctuations. The relation captures an often-observed reduction of relative noise with increasing carrier population, but \\(\\alpha_{\\mathrm H}\\) is not a universal constant. Its value depends on material quality, transport mechanism, and how the active volume and carrier number are defined.\n\nThe metal–oxide–semiconductor field-effect transistor is a particularly important example. Traps in or near the gate dielectric exchange charge with the channel, changing the effective threshold voltage. For small fluctuations,\n\n\\[\\delta I_D\\approx g_m\\,\\delta V_{\\mathrm{th}},\\]\n\nwhere \\(g_m\\) is the transconductance. The corresponding drain-current spectrum contains a contribution proportional to \\(g_m^2S_{V_{\\mathrm{th}}}\\). Trapped charge can also change carrier mobility through scattering, so number and mobility fluctuations may be correlated rather than independent.\n\nDevice scaling changes the character as well as the amplitude of noise. Large transistors average over many microscopic defects, often yielding relatively smooth spectra. In a very small transistor, one trap can noticeably alter the channel conductance, producing discrete switching and strongly sample-dependent behavior. Increasing gate area commonly reduces input-referred flicker noise in otherwise comparable devices, but exact scaling depends on trap distribution, electrostatics, fabrication, and bias.\n\nA practical spectrum often contains a white background plus a flicker contribution,\n\n\\[S(f)=S_{\\mathrm w}+\\frac{K}{f^\\alpha}.\\]\n\nTheir intersection defines a flicker corner frequency. Below that corner, low-frequency noise dominates; above it, thermal noise or shot noise may dominate. Contact resistance, surface adsorption, mechanical stress, and fluctuating magnetization can add further contributions. Noise measurements thus serve both as performance tests and as sensitive probes of defects that ordinary current–voltage measurements may not reveal."
    },
    {
      "heading": "Occurrence outside conventional electronics",
      "text": "Inverse-frequency spectra occur in many nonelectronic records, but the measured variable and frequency range are crucial. A spectrum of event counts is not equivalent to a spectrum of intervals between events, and neither is automatically equivalent to a spectrum of a continuously measured field. Sampling and preprocessing can change apparent scaling. The observation that unrelated systems have similar slopes is therefore a starting point for comparison, not evidence that they share an underlying microscopic mechanism.\n\nIn neuroscience, electrical recordings often contain an aperiodic background whose power decreases with frequency. This background can coexist with narrow oscillatory peaks. Depending on the signal, frequency range, and recording conditions, a power law or a power law with a bend can describe portions of the spectrum. Synaptic filtering, population activity, and the electrical properties of tissue can all shape the observed slope. A fitted exponent alone does not uniquely measure a particular cellular process or establish a specific balance of excitation and inhibition.\n\nBiological timing records also show broad correlations. Heartbeat-interval series have displayed approximately inverse-frequency scaling over selected ranges, although rhythms, activity, sleep, age, and recording procedures affect the result. Fluctuations in ion-channel currents may instead be connected more directly to molecular transitions among conducting and nonconducting states. A small number of transition rates generates recognizable spectral components; a broad distribution of rates or interacting channels can produce a more extended background. These are distinct modeling situations despite the shared language of colored noise.\n\nGeophysical and environmental records contain low-frequency variability from coupled reservoirs and transport processes. Atmospheric and oceanic responses span times from short-lived disturbances to slow heat storage and circulation. River discharge combines precipitation, soil storage, groundwater exchange, and routing through drainage networks. Such combinations can generate broad or approximately power-law spectra, but seasonal cycles, trends, and changes in the observing system must be separated from stochastic variability. Long records do not by themselves guarantee a single scaling regime.\n\nIn acoustics, synthesized pink noise is useful because each octave contains equal mean-square amplitude. It is used in audio testing and sound-system measurements, with the generated spectrum limited by playback and measurement bandwidth. Statistical analyses of some musical features have also reported inverse-frequency-like behavior, but results depend on whether the variable is loudness, pitch, timing, or another feature. In astronomy, broad red-noise continua appear in variable sources, sometimes with slopes near 1 over part of the spectrum and breaks elsewhere. Across these applications, specifying the variable, band limits, and competing spectral models is more informative than assigning the label “1/f” alone."
    },
    {
      "heading": "Measurement, estimation, and statistical pitfalls",
      "text": "Measuring 1/f noise requires control of both the instrument and the statistical procedure. A typical electronic experiment biases a sample with a low-noise source, amplifies its fluctuations, applies an anti-alias filter, and records the signal digitally. The sample spectrum must be distinguished from amplifier voltage and current noise, reference fluctuations, environmental pickup, and temperature drift. Shielding and grounding reduce interference, while thermal and mechanical control help prevent slow external disturbances from masquerading as intrinsic device noise.\n\nRecord length sets a fundamental scale. An observation lasting \\(T_{\\mathrm{obs}}\\) has Fourier-frequency spacing \\(1/T_{\\mathrm{obs}}\\), and frequencies near that limit are represented by very few independent fluctuations. Extending a claim from 1 Hz to 1 mHz changes the relevant period from one second to roughly 17 minutes; reliable characterization generally requires many such periods. At these timescales, stability of the sample, bias, and laboratory environment becomes part of the experiment.\n\nA periodogram estimates spectral power from squared Fourier amplitudes, but its values fluctuate strongly between records. Averaging independent records or overlapping, windowed segments reduces this statistical scatter at the cost of frequency resolution. Windowing reduces leakage from strong components, although it also modifies spectral resolution and normalization. For steep spectra, leakage from unresolved low frequencies can distort a substantial portion of the measured band. Aliasing causes fluctuations above the Nyquist frequency to enter the sampled band unless sufficiently attenuated before digitization.\n\nFitting a straight line to a logarithmic spectrum is convenient but not automatically statistically sound. Spectral estimates have nonuniform uncertainty, neighboring bins may be correlated, and logarithmic binning changes weighting. Plausible models include a power law plus a white floor, one or more Lorentzians, or a broken power law. A fitted exponent should be accompanied by its frequency interval, treatment of peaks and drift, and uncertainty. Removing a polynomial trend or applying a high-pass filter can suppress genuine low-frequency fluctuations as well as unwanted drift.\n\nCross-correlation methods can suppress independent noise from two readout channels, but common interference and correlated back-action remain. In frequency metrology, Allan variance complements ordinary spectra by comparing adjacent time averages. Under the usual idealized power-law model, flicker noise in fractional frequency produces an approximately averaging-time-independent Allan variance over its scaling range. For other measured quantities, the correspondence must be derived rather than assumed. Reporting whether a result describes voltage, phase, frequency, or fractional frequency prevents otherwise serious ambiguity in noise classification and measurement uncertainty."
    },
    {
      "heading": "Mitigation, simulation, and limits of interpretation",
      "text": "The most effective suppression strategy depends on where the fluctuations originate. In semiconductor devices, improved interfaces, lower defect densities, suitable geometry, and control of contacts can reduce flicker noise. Increasing active area often improves averaging over defects, but can increase capacitance, power requirements, or settling time. Temperature changes may shift a thermally activated process out of one measurement band and into another; cooling does not guarantee that all low-frequency noise decreases.\n\nModulation techniques move information away from the noisiest frequencies. A lock-in amplifier detects a signal near a known reference frequency and translates it back to a slowly varying output. Chopper-stabilized amplifiers similarly modulate a low-frequency input before amplification and demodulate it afterward, reducing the effect of amplifier offset and low-frequency noise generated between the modulation stages. They do not remove fluctuations already inseparable from the input signal. Switching artifacts, residual ripple, and the noise of other circuit elements remain design constraints.\n\nAuto-zeroing and correlated double sampling compare measurements made at different times so that slowly varying offsets partially cancel. Their effectiveness depends on the correlation between those measurements and on the timing sequence; sampling can also fold broadband noise into the signal band. In oscillators, flicker noise in active components may be converted into phase fluctuations. A spectrum proportional to \\(1/f\\) in fractional frequency corresponds to a phase spectrum proportional to \\(1/f^3\\), apart from convention-dependent constants, because phase integrates frequency. The noise variable must therefore be stated explicitly.\n\nSynthetic 1/f signals are used to test estimators, control systems, and signal-processing algorithms. One method assigns random Fourier coefficients with mean-square magnitude proportional to \\(1/f^\\alpha\\), imposes conjugate symmetry, and transforms back to the time domain. Another sums relaxation processes over logarithmically spaced timescales. Both methods require finite cutoffs. Fourier synthesis also creates a finite periodic representation, which can matter when the record is repeated or when long-lag correlations are analyzed.\n\nMatching a spectrum is not equivalent to reproducing the complete noise process. A Gaussian synthetic signal can share its spectrum with a signal dominated by rare switching events while having very different extreme-value statistics. Likewise, a many-defect device and an interacting system may exhibit similar second-order spectra but different higher-order correlations and responses to perturbation. Mechanistic interpretation therefore relies on additional observables: temperature dependence, bias dependence, spatial correlations, distributions of switching times, and changes caused by controlled alterations to the material or system. These measurements distinguish a useful empirical 1/f description from a tested explanation of its origin."
    }
  ],
  "links": [
    "variance",
    "stochastic-process",
    "wiener-khinchin-theorem",
    "autocorrelation",
    "fourier-transform",
    "boltzmann-constant",
    "temperature",
    "electrical-resistance",
    "quantum-tunneling",
    "activation-energy",
    "electron",
    "phase-transition",
    "ohms-law",
    "joule-heating",
    "mosfet",
    "scattering",
    "adsorption",
    "magnetization",
    "neuroscience",
    "ion-channel",
    "river-discharge",
    "aliasing",
    "allan-variance",
    "measurement-uncertainty",
    "capacitance",
    "signal-processing"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}