Spectral density is a mathematical description of how a quantity—most commonly signal energy, mean-square amplitude, or variance—is distributed over frequency. In time-series analysis and signal processing, the term usually means power spectral density: a nonnegative function whose integral over a frequency band gives that band’s contribution to the total mean-square value. For a centered stationary random process, the integral over all frequencies equals its variance. Spectral density therefore connects fluctuations observed in time with oscillatory structure expressed in frequency.
The central relationship is between spectral density and the Fourier transform of an autocovariance or autocorrelation function. This relationship makes spectral methods useful for detecting periodic components, characterizing noise, analyzing linear systems, and estimating the scales on which a process varies. A spectrum can distinguish slowly varying fluctuations from rapid ones even when the corresponding time records have similar overall variability. Its interpretation nevertheless depends on stationarity, sampling, normalization, and the statistical uncertainty introduced by observing only a finite record.
More generally, spectral distributions need not possess ordinary density functions: periodic components produce discrete spectral lines, and other processes can have singular spectral measures. Related uses of “spectral density” occur in spatial statistics, statistical physics, and operator theory, where the spectral variable may be a wave number, energy, or eigenvalue rather than temporal frequency.
Frequency conventions, normalization, and units
A spectral density is defined relative to a frequency coordinate and an integration convention. For a continuous-time process, frequency , measured in cycles per unit time, and angular frequency , measured in radians per unit time, satisfy . A common Fourier convention for an autocovariance function is
Under this convention, . The area under the density, rather than its height at a single frequency, is the physically relevant contribution to variance.
Two angular-frequency conventions are widespread. One defines
Here , and the integration measure is . Another convention defines a density directly with respect to , namely . These conventions describe the same variance distribution, but their numerical densities differ by . A spectral formula is therefore incomplete unless its transform and integration conventions are specified.
For a real-valued process, the two-sided spectrum is even: . A one-sided spectrum places all continuous spectral power at nonnegative frequencies:
Its integral over positive frequencies equals that of the two-sided spectrum over the whole line. Zero frequency and, in sampled data, the Nyquist frequency require separate treatment when they carry discrete mass or correspond to unpaired transform bins. Doubling every bin indiscriminately gives an incorrect normalization.
If a signal is measured in volts and time in seconds, its power spectral density has units . The term “power” here often means mean-square amplitude, not electrical power in watts. Conversion to dissipated electrical power requires a circuit model, such as division by a resistance. The square root of a power spectral density is an amplitude spectral density, with units . Its integral does not give a root-mean-square amplitude; the density must first be squared and integrated.
Unlike a probability density, a spectral density need not integrate to one. Its integral may equal any finite variance. Dividing a centered process’s spectrum by its variance produces a normalized spectral distribution, but this is an additional normalization rather than part of the definition.
Stationary processes and the spectral representation
The probabilistic foundation begins with a second-order stationary process . Its mean is constant, its second moment is finite, and its covariance depends only on the time difference:
The overbar denotes complex conjugation. The expectation averages over realizations of the process, rather than over an observed time record. The distinction matters because stationarity alone does not guarantee that time averages converge to the corresponding ensemble quantities.
The Wiener–Khinchin theorem identifies the spectrum with the Fourier transform of this covariance, under conditions permitting an ordinary transform. If is absolutely integrable, the formula
gives a continuous, nonnegative spectral density. Absolute integrability is sufficient, not necessary, for a density to exist. Nonnegativity follows from the positive-definite character of covariance: every finite linear combination of process values has a nonnegative variance. Thus covariance functions cannot be chosen arbitrarily.
The more general statement uses measure theory. For a mean-square continuous stationary process, there is a finite nonnegative spectral measure such that
An ordinary spectral density exists when is absolutely continuous with respect to Lebesgue measure, in which case . A spectral measure can instead contain atoms, which represent spectral lines, or a singular continuous component. Treating every spectrum as a smooth function excludes these valid possibilities.
For a discrete-time process , frequency is periodic because . A standard convention is
when the series is appropriately defined. Consequently, . Frequencies outside one interval of length do not provide additional information.
The spectral representation expresses a centered process as an integral of complex oscillations with orthogonal random increments:
The expected squared magnitude of an increment is determined by the corresponding spectral mass. Orthogonal increments are uncorrelated; they need not be independent. For a Gaussian process, however, the mean and covariance determine all finite-dimensional distributions. Its spectrum, together with its mean, therefore supplies a complete second-order and distributional description, whereas a non-Gaussian process can have important higher-order structure invisible to its spectrum.
Energy spectra, power spectra, and spectral lines
For a deterministic signal with finite energy,
its Fourier transform defines an energy spectral density
Parseval’s identity gives . If is measured in volts, this density has units , equivalently , because its frequency integral gives integrated squared voltage rather than mean-square voltage. Calling this quantity a power spectral density obscures the distinction between finite energy and sustained average power.
A persistent sinusoid has infinite energy over the whole real line but finite time-average power. Its ordinary Fourier transform is not a square-integrable function, so simply squaring that transform is not a valid definition of a power spectrum. Instead, one can examine a finite observation:
and study the measures with densities . For a stationary random process, the corresponding expected quantities are related to the spectral measure. The limiting operation may need to be understood weakly, through integrals against test functions, rather than as pointwise convergence of ordinary functions.
For example, let
where is uniform on . This process has zero mean and covariance . Its two-sided spectrum is
where is the Dirac delta distribution. Each line has mass , and the total variance is . A line’s coefficient is an integrated spectral mass, not the finite height of a density peak.
A nonzero mean creates a related distinction. If the spectrum is defined from the uncentered autocorrelation , then
Its spectral measure therefore contains an additional atom of mass at zero frequency. The covariance spectrum excludes this constant component and integrates to variance; the autocorrelation spectrum includes it and integrates to the second moment.
For a periodic signal with Fourier-series coefficients and fundamental frequency , the power spectrum consists of lines with masses at . Spectral magnitudes retain component strengths but discard their relative phases. Different waveforms can consequently share the same power spectrum, just as different random processes can share the same covariance.
Characteristic spectra and stochastic models
White noise provides the reference case of a flat spectrum. A discrete-time zero-mean sequence with covariance for and zero otherwise has under the angular-frequency convention above. Every interval of equal frequency width contributes equally to variance. Whiteness is a second-order property: it specifies lack of correlation, not independence or a Gaussian distribution. Independent identically distributed variables with finite variance are white after centering, but the converse need not hold.
Ideal continuous-time white noise requires a generalized-process interpretation. Its covariance is proportional to a delta distribution and its spectral density is constant over the entire real line. Integrating that constant gives infinite variance, so it cannot describe ordinary pointwise random variables with finite variance. Physical noise is approximately white only over a bounded frequency range. Filtering the idealization through a system with a square-integrable frequency response can nevertheless produce a well-defined finite-variance output.
An elementary colored-noise model is the first-order autoregressive process
where the innovations are white with variance . Its stationary covariance and spectrum are
Positive concentrates power near zero frequency and produces slowly varying records. Negative concentrates power near , producing alternating behavior. As approaches either stationarity boundary, the spectral peak becomes increasingly sharp.
A continuous-time analogue has exponential covariance
where is a correlation time. Its spectrum is the Lorentzian
The characteristic transition occurs near . Short correlation times spread variance over a broad frequency range; long correlation times concentrate it close to zero. The covariance and spectrum provide complementary descriptions of the same temporal persistence.
Many observed spectra follow an approximate power law, , over a finite range. Flicker or pink noise commonly denotes behavior near . Such formulas cannot automatically be extended to all frequencies: a stationary finite-variance process requires an integrable spectrum. In particular, a divergence at zero frequency with is incompatible with finite variance unless modified by a cutoff or another change in behavior.
An ordinary random walk illustrates the danger of formal spectra. It is nonstationary, so the expression proportional to is not an ordinary finite-variance stationary spectral density. By contrast, an integrable low-frequency singularity with can describe stationary long-range dependence. “Large low-frequency power” can therefore indicate either stationary persistence or nonstationarity, depending on the full model.
Linear filtering, sampling, and aliasing
For a stable linear time-invariant system with impulse response , the output is the convolution
If the stationary input has spectral density , and is the Fourier transform of , then
This identity converts the calculation of a filtered covariance into multiplication in frequency. A low-pass filter suppresses rapid fluctuations, a band-pass filter isolates a frequency range, and a resonant response can create a pronounced output peak from an input without a corresponding narrow peak.
The output variance is
For a flat input spectrum, this motivates equivalent noise bandwidth: the integral of the squared frequency response divided by a specified reference gain. Noise bandwidth need not equal the frequency interval between half-power points. The distinction matters when comparing filters that have similar nominal cutoff frequencies but different response shapes.
Sampling at interval produces , with sampling frequency . Under suitable regularity, the sampled spectrum in physical frequency units is
The shifted copies add because continuous-time frequencies separated by integer multiples of produce identical sampled oscillations. This folding of spectral power is aliasing. The integral over the base interval still gives the sampled process’s variance, but the original allocation of power among aliased frequencies cannot generally be recovered.
The sampling theorem gives exact reconstruction conditions for appropriately band-limited signals sampled above twice their highest frequency. Real random noise often has no strict upper frequency limit, so an analog anti-aliasing filter is important before sampling. Filtering the data after acquisition cannot undo an unknown mixture of spectral contributions already folded into the same frequency band.
Changing from physical frequency to angular frequency per sample introduces another scale factor. If is normalized by
then . Confusing these densities causes errors by factors of the sampling rate, even when the plotted spectral shapes are identical.
Operations applied before estimation also act as filters. First differencing has squared gain , suppressing low frequencies. Finite-sample detrending is not simply an ideal high-pass filter, but it likewise alters low-frequency estimates. Differentiation, when defined in the mean-square sense, multiplies a continuous-time spectrum by , requiring sufficient high-frequency decay for the derivative to have finite variance.
Periodograms, windows, and nonparametric estimation
A spectrum is usually inferred from a finite record rather than calculated from a known covariance. For observations at spacing , a two-sided periodogram in physical frequency units is
The discrete Fourier transform evaluates this expression efficiently on a frequency grid, while the fast Fourier transform reduces the computational cost of obtaining those values. Fourier computation, however, does not by itself solve the statistical problem of estimating a population spectrum.
Multiplication by a finite observation window becomes convolution in frequency. Consequently, the expected periodogram is a smoothed version of the true spectrum, with smoothing kernel determined by the observation window. A sinusoid whose frequency falls between transform bins spreads power across nearby bins and often more distant ones. This spectral leakage is not necessarily evidence for multiple physical oscillations; it is partly a consequence of restricting observation to a finite interval.
Tapering reduces the abruptness of the record’s endpoints. Windows such as Hann and Hamming suppress sidelobes relative to a rectangular window, at the cost of a broader main lobe. For a taper , a common density normalization is
The squared-weight denominator calibrates mean-square power. A different normalization is needed when the objective is to estimate the amplitude of a coherent sinusoid. Confusing amplitude calibration with noise-density calibration can produce systematic scale errors.
The raw periodogram is generally not a consistent estimator at an individual frequency. Under standard conditions, at an interior positive frequency its value divided by the true spectrum is approximately exponential with mean one. Increasing record length improves frequency localization but does not make this relative variability disappear. Smoothing nearby frequencies or averaging approximately independent spectral estimates is therefore essential for stable nonparametric estimation.
The Welch method divides a record into possibly overlapping, tapered segments and averages their periodograms. Shorter segments increase the number of averages but reduce frequency resolution. Overlapping segments are correlated, so their nominal count is not the effective number of independent averages. Multitaper estimation instead applies several orthogonal tapers to the same record, commonly discrete prolate spheroidal sequences, and combines their spectral estimates.
Zero-padding evaluates the finite-record Fourier transform on a finer grid but does not create new observations or separate components that the observation duration cannot resolve. If an averaged estimate has an effective degrees of freedom, an approximate uncertainty model is . This supports confidence intervals, subject to assumptions about the process, tapering, averaging, and local smoothness of the spectrum.
Parametric models and frequency-domain inference
Parametric spectral estimation replaces an otherwise unrestricted density with a family indexed by finitely many parameters. For an autoregressive moving-average model,
the backshift operator satisfies . If the innovation variance is , the spectral density is
The notation distinguishes the moving-average polynomial from the backshift operator. Under the usual causal convention, zeros of the autoregressive polynomial lie outside the unit circle for stationarity.
This rational form has a direct interpretation in terms of system dynamics. An autoregressive polynomial zero near the unit circle can produce a narrow spectral peak at the corresponding angle. Complex-conjugate pairs describe damped oscillations in real-valued series. Moving-average factors can create depressions or notches. Thus a small collection of parameters can express behavior that would require substantial smoothing and many frequency bins in a nonparametric estimate.
Parameters can be estimated by covariance equations, prediction-error methods, or maximum likelihood. Autoregressive fitting through the Yule–Walker equations links model coefficients directly to estimated autocovariances. A fitted model can then provide both a spectral estimate and a prediction rule. The apparent smoothness of a parametric spectrum is supplied partly by model restrictions, not solely by information in the observations.
The Whittle likelihood is a frequency-domain approximation to Gaussian time-series likelihood. Up to normalization and constants, its negative log-likelihood has the form
where and use matching conventions. It exploits the approximate decorrelation of Fourier coefficients at distinct Fourier frequencies in long stationary records. The approximation can be inaccurate for short records, strong spectral leakage, or pronounced behavior close to zero frequency.
Model order controls a familiar bias–variance tradeoff. Too few parameters can flatten or merge real peaks; too many can fit random fluctuations and create spurious narrow components. Information criteria, predictive checks, and residual analysis assess different aspects of model adequacy. A residual spectrum that remains strongly structured indicates that the fitted model has not removed all second-order temporal dependence.
Spectral peaks alone do not establish an external periodic forcing. A stochastic system with a damped resonance can generate a peak without any deterministic clock. Likewise, a high fitted peak is not automatically statistically significant: its interpretation depends on the background spectrum, parameter uncertainty, and how many candidate frequencies were examined. Bayesian spectral methods express uncertainty through a posterior over models or densities, but their resolution and conclusions also depend on the chosen prior and likelihood.
Cross-spectra, spatial spectra, and changing frequency structure
For two jointly stationary processes, define the cross-covariance by
Its Fourier transform is the cross-spectral density . Unlike an autospectrum, a cross-spectrum can be complex: its magnitude describes shared frequency-dependent variation, while its phase describes the relative alignment of oscillatory components under the stated lag convention. Reversing the order of the processes conjugates the cross-spectrum.
For a vector process, the auto- and cross-spectra form a spectral density matrix. At each frequency this matrix is Hermitian and positive semidefinite. In particular,
Where the denominator is positive, magnitude-squared coherence is
and lies between zero and one. It measures linear frequency-specific association. A third process can drive both records, so coherence does not by itself establish causation or identify a direction of influence.
For a pure delay , the transfer function is . With the definition above, . The cross-spectral phase therefore changes linearly with frequency, modulo . Phase unwrapping can make that slope visible, but interpreting it as a unique delay requires an appropriate system model. General filters can have frequency-dependent phase responses that do not represent a single transit time.
Spatially stationary random fields use wave-number vectors instead of temporal frequency. For a field on , one convention is
When the field is isotropic, the density depends only on . A radial spectrum formed by integrating over spherical shells contains an additional geometric factor proportional to . Its numerical shape is therefore not identical to the underlying density per unit wave-number volume. This distinction is important when comparing spectra across spatial dimensions or with turbulence energy spectra.
A single global stationary spectrum cannot describe when an oscillation occurs. Short-time Fourier methods estimate spectra in moving windows, while wavelet methods vary temporal resolution with scale. These methods require a compromise between time localization and frequency localization. A process whose oscillation frequency changes over a record can have a broad global spectrum even if its instantaneous oscillation is narrowly concentrated. Time-frequency representations address this limitation, but their definitions and positivity properties differ; not every such representation is a local power spectral density.
Historical development and scientific applications
Spectral analysis developed from harmonic analysis, particularly the representation of functions by sinusoidal components. Joseph Fourier’s 1822 work on heat supplied a major mathematical foundation for Fourier expansions and transforms. Extending these ideas from deterministic functions to irregular observations required a different object: not simply the transform of a waveform, but the frequency distribution of its second-order variation. Arthur Schuster introduced the term “periodogram” in 1898 in work on periodicities in observations.
During the early twentieth century, spectral methods became central to the study of random fluctuations. Norbert Wiener’s 1930 work on generalized harmonic analysis and Aleksandr Khinchin’s 1934 work on stationary random processes established the classical connection between correlation and spectrum. Harald Cramér developed spectral representations of stationary processes, while Andrey Kolmogorov and Wiener connected stationary-process theory with prediction and filtering. These developments made spectra part of probability theory rather than only a computational technique for finding cycles.
Practical estimation advanced alongside electronic instrumentation and digital computation. Maurice Bartlett developed averaging approaches that reduced periodogram variability. John Tukey and collaborators emphasized smoothing and lag-window methods. Peter Welch’s 1967 method combined segment averaging, tapering, and efficient Fourier computation. David Thomson’s 1982 multitaper work provided a systematic framework for controlling leakage and estimating uncertainty using multiple orthogonal data tapers. These methods address different compromises among resolution, statistical stability, and spectral contamination.
In electrical measurements, spectral density quantifies noise contributed by components and instruments. In the classical thermal regime, an ideal resistor’s open-circuit voltage noise has one-sided density
where is Boltzmann’s constant, is absolute temperature, and is resistance. A bandwidth over which the model is flat therefore gives mean-square noise voltage . The expression assumes the classical low-frequency thermal regime and does not extend unchanged to arbitrarily high frequencies.
In mechanical and structural testing, acceleration spectra describe vibration environments and resonances. Integrating an acceleration density over a band gives that band’s mean-square acceleration; converting acceleration to displacement involves frequency-dependent integration and can strongly amplify low-frequency contributions. In geophysics and climate research, spectra characterize variability across temporal and spatial scales, but trends, finite records, and red-noise backgrounds complicate the identification of periodic components.
In quantum mechanics, noise spectra require an additional distinction because observables at different times need not commute. Symmetrized and unsymmetrized correlation functions define different spectra, and positive- and negative-frequency components can encode different transition processes. In operator theory, “spectral density” may instead refer to the density of an operator’s spectral measure; a density of states counts modes or energy levels per interval. These uses share a measure-theoretic language with power spectral density but describe different quantities and require their own normalizations.