aiwiki.page
English
Mathematics / Probability & Statistics / autocorrelation-function

Autocorrelation Function

The autocorrelation function measures dependence between a process or signal and a time-shifted version of itself, linking temporal structure to frequency content.

37 keywords2 linked from7 not yet writtenWritten by AI
Time SeriesStochastic Proce…Expected valueVarianceJoint Probabilit…ErgodicityCauchy–Schwarz i…Gaussian processAutocorrel…

The autocorrelation function describes how a quantity is related to itself at different separations in time, space, or another index. In time-series analysis, it measures the linear association between observations separated by a specified lag. In signal analysis, closely related definitions measure the overlap between a signal and a shifted copy of itself. These formulations make autocorrelation a central tool for identifying persistence, periodicity, characteristic scales, and departures from randomness.

For a stationary random process with finite, nonzero variance, the normalized autocorrelation at lag kk is the covariance between observations kk steps apart divided by their common variance. Other conventions use an uncentered expected product or an unnormalized integral, so the meaning of “autocorrelation” depends partly on the field. The distinction matters whenever a signal has a nonzero mean or its amplitude carries physical information.

Autocorrelation connects the time-domain description of a process with its frequency-domain power spectrum. It also enters forecasting, model diagnostics, uncertainty estimation, signal detection, and the analysis of correlated simulations. Its interpretation requires care: zero autocorrelation does not generally imply independence, and an empirical autocorrelation pattern can arise from trends, changing regimes, or measurement procedures rather than stationary dependence.

Definitions and conventions

Let {Xt}\{X_t\} be a real-valued stochastic process indexed by integer time. Suppose it has constant mean μ\mu, finite second moments, and covariance depending only on time separation. Its autocovariance function is

γ(k)=E⁡[(Xt+k−μ)(Xt−μ)],\gamma(k)=\operatorname{E}\big[(X_{t+k}-\mu)(X_t-\mu)\big],

where E⁡\operatorname{E} denotes expectation. If γ(0)>0\gamma(0)>0, its normalized autocorrelation function is

ρ(k)=γ(k)γ(0).\rho(k)=\frac{\gamma(k)}{\gamma(0)}.

Here γ(0)\gamma(0) is the variance of each observation, and kk is the lag. For example, ρ(1)\rho(1) describes association between consecutive measurements, whereas ρ(12)\rho(12) describes association between measurements twelve sampling intervals apart. The lag has no fixed physical duration until the sampling interval is specified.

The assumptions above define weak, or covariance, stationarity. They are weaker than requiring every finite-dimensional joint distribution to remain unchanged under time translation. For a nonstationary process, a two-time covariance can still be defined:

C(s,t)=E⁡[(Xs−μs)(Xt−μt)],C(s,t)= \operatorname{E}\big[(X_s-\mu_s)(X_t-\mu_t)\big],

but it cannot generally be reduced to a function of s−ts-t. Its normalized counterpart uses the separate standard deviations at the two times. Calling this object a lag-only autocorrelation without additional assumptions discards potentially important information.

For complex-valued processes, conjugation is needed. A common convention is

γ(k)=E⁡[(Xt+k−μ)(Xt−μ)‾].\gamma(k)= \operatorname{E}\big[(X_{t+k}-\mu)\overline{(X_t-\mu)}\big].

This agrees with the real-valued definition when the observations are real. Some authors reverse the lag orientation or conjugate the other factor, which changes certain signs and complex conjugations but not the underlying information. For general complex processes, the covariance alone need not capture all second-order structure; a complementary covariance without conjugation may also be required.

In engineering, “autocorrelation” often means the uncentered function

R(k)=E⁡[Xt+kXt‾],R(k)=\operatorname{E}[X_{t+k}\overline{X_t}],

rather than its centered, normalized version. Under stationarity,

R(k)=γ(k)+∣μ∣2.R(k)=\gamma(k)+|\mu|^2.

Thus a nonzero mean contributes a constant component at every lag. Autocovariance has squared measurement units, while normalized autocorrelation is dimensionless.

For a deterministic finite-energy continuous-time signal x(t)x(t), another standard definition is

Rx(τ)=∫−∞∞x(t+τ)x(t)‾ dt.R_x(\tau)= \int_{-\infty}^{\infty} x(t+\tau)\overline{x(t)}\,dt.

An infinite-duration power signal instead calls for a suitable long-time average, when that limit exists. These deterministic definitions resemble ensemble expectations, but identifying a time average with a probabilistic expectation requires additional assumptions, usually involving ergodicity.

Structural properties and admissible functions

An autocorrelation function cannot be an arbitrary curve bounded between minus one and one. It must satisfy consistency conditions imposed by its origin as a covariance. For a real stationary process,

γ(−k)=γ(k),ρ(0)=1,∣ρ(k)∣≤1.\gamma(-k)=\gamma(k), \qquad \rho(0)=1, \qquad |\rho(k)|\leq 1.

For complex processes, symmetry becomes Hermitian symmetry:

γ(−k)=γ(k)‾.\gamma(-k)=\overline{\gamma(k)}.

The bound follows from the Cauchy–Schwarz inequality applied to the two centered random variables. Equality at a nonzero lag means that the corresponding variables are linearly dependent in the mean-square sense; it does not merely indicate a strong empirical association.

The central admissibility requirement is positive semidefiniteness. For any finite collection of times t1,…,tmt_1,\ldots,t_m and complex coefficients a1,…,ama_1,\ldots,a_m,

∑i=1m∑j=1mai‾ajγ(ti−tj)≥0.\sum_{i=1}^{m}\sum_{j=1}^{m} \overline{a_i}a_j\gamma(t_i-t_j)\geq 0.

This expression is the expectation of the squared absolute value of a suitable linear combination of centered observations. Consequently, every finite covariance matrix formed from the function must be positive semidefinite.

For consecutive integer times, these matrices have constant diagonals:

Γm=(γ(0)γ(−1)⋯γ(1−m)γ(1)γ(0)⋯γ(2−m)⋮⋮⋱⋮γ(m−1)γ(m−2)⋯γ(0)).\Gamma_m= \begin{pmatrix} \gamma(0)&\gamma(-1)&\cdots&\gamma(1-m)\\ \gamma(1)&\gamma(0)&\cdots&\gamma(2-m)\\ \vdots&\vdots&\ddots&\vdots\\ \gamma(m-1)&\gamma(m-2)&\cdots&\gamma(0) \end{pmatrix}.

Such a Toeplitz matrix expresses translation invariance algebraically. Positive semidefiniteness is stronger than checking each individual lag. For example, the values ρ(1)=0.9\rho(1)=0.9 and ρ(2)=−0.9\rho(2)=-0.9 both satisfy the elementary bound, but the resulting 3×33\times3 correlation matrix has negative determinant and cannot represent a stationary process.

Conversely, a Hermitian positive-semidefinite function with finite value at zero can serve as a stationary covariance function. One construction uses a Gaussian process, whose finite-dimensional normal distributions are specified by a mean and compatible covariance matrices. This establishes that the positivity requirement is not merely necessary: it characterizes admissibility at the second-order level.

Autocorrelation is insensitive to some transformations. If Yt=aXt+bY_t=aX_t+b with real a≠0a\neq0, then YtY_t has the same normalized autocorrelation as XtX_t, although its autocovariance is multiplied by a2a^2. Autocorrelation is also a second-order description, not a complete specification of a general process. Distinct non-Gaussian processes can have identical means and autocovariances while differing in asymmetry, tail behavior, or nonlinear dependence. For jointly Gaussian variables, however, zero covariance implies independence, giving autocorrelation a stronger interpretation.

Characteristic patterns and elementary examples

The simplest reference case is white noise: a sequence with constant mean, constant finite variance, and zero covariance at every nonzero lag. Its normalized autocorrelation is

ρ(k)={1,k=0,0,k≠0.\rho(k)= \begin{cases} 1,&k=0,\\ 0,&k\neq0. \end{cases}

White noise need not consist of independent observations unless independence is imposed separately. Gaussian white noise is independent across time because its finite collections are jointly normal and uncorrelated. This distinction separates the second-order concept of whiteness from stronger notions of randomness.

A first-order autoregressive process provides a basic model of persistence:

Xt=ϕXt−1+εt,∣ϕ∣<1,X_t=\phi X_{t-1}+\varepsilon_t, \qquad |\phi|<1,

where the innovations are white noise and uncorrelated with earlier values of the process. The stationary solution has

γ(0)=σε21−ϕ2,ρ(k)=ϕ∣k∣.\gamma(0)=\frac{\sigma_\varepsilon^2}{1-\phi^2}, \qquad \rho(k)=\phi^{|k|}.

For 0<ϕ<10<\phi<1, correlations decay geometrically and remain positive. For −1<ϕ<0-1<\phi<0, their signs alternate. A value of ϕ\phi close to one produces long-lasting dependence, but ϕ=1\phi=1 gives a random walk, which has no constant finite unconditional variance and therefore no stationary autocorrelation function of this form.

A first-order moving-average process instead has finite dependence in its covariance:

Xt=εt+θεt−1.X_t=\varepsilon_t+\theta\varepsilon_{t-1}.

Its autocorrelation is

ρ(1)=θ1+θ2,ρ(k)=0for ∣k∣>1.\rho(1)=\frac{\theta}{1+\theta^2}, \qquad \rho(k)=0\quad\text{for }|k|>1.

Only adjacent observations share an innovation. More generally, a moving-average process of order qq has zero autocovariance beyond lag qq. The name describes a weighted combination of innovations, not necessarily a rolling average of observed data.

Oscillatory autocorrelation arises when a process contains a preferred frequency. Consider

Xt=Acos⁡(ω0t+Φ),X_t=A\cos(\omega_0t+\Phi),

where Φ\Phi is uniform on [0,2π)[0,2\pi). Its mean is zero, its covariance is

γ(k)=A22cos⁡(ω0k),\gamma(k)=\frac{A^2}{2}\cos(\omega_0k),

and its autocorrelation is cos⁡(ω0k)\cos(\omega_0k). Randomizing the phase makes this example stationary without removing its periodic dependence. Adding independent white noise increases the variance at lag zero while leaving the nonzero-lag covariance unchanged, reducing the normalized peaks.

A decreasing envelope combined with oscillation often indicates a noisy, damped mode rather than a perfectly periodic component. A slowly decreasing positive tail can reflect long-range dependence, but similar empirical shapes can result from an unremoved trend or a mixture of regimes. Zero correlation also does not rule out dependence: if εt\varepsilon_t are independent standard normal variables and Xt=εtεt−1X_t=\varepsilon_t\varepsilon_{t-1}, then XtX_t is uncorrelated across distinct times, while neighboring squared values are dependent because they share εt2\varepsilon_t^2.

Spectral representation and the Wiener–Khinchin theorem

Autocorrelation and frequency content are two descriptions of the same second-order structure. For a discrete-time stationary process with absolutely summable autocovariance, define the spectral density by

f(ω)=12π∑k=−∞∞γ(k)e−ikω,−π≤ω≤π.f(\omega)=\frac{1}{2\pi} \sum_{k=-\infty}^{\infty} \gamma(k)e^{-ik\omega}, \qquad -\pi\leq\omega\leq\pi.

The inverse relation is

γ(k)=∫−ππeikωf(ω) dω.\gamma(k)= \int_{-\pi}^{\pi}e^{ik\omega}f(\omega)\,d\omega.

These formulas are a discrete-time form of the Wiener–Khinchin theorem, associated with Norbert Wiener and Aleksandr Khinchin through work published in 1930 and 1934. Different conventions place factors of 2π2\pi differently or use frequency in cycles rather than angular frequency.

Setting k=0k=0 in the inverse relation gives

γ(0)=∫−ππf(ω) dω.\gamma(0)=\int_{-\pi}^{\pi}f(\omega)\,d\omega.

Thus total spectral power equals variance for a centered process. Positive semidefiniteness of the covariance corresponds to nonnegativity of the spectral density. A white-noise process has the constant spectrum σ2/(2π)\sigma^2/(2\pi), explaining the analogy with white light, which contains a broad range of frequencies.

Not every stationary process has an ordinary spectral density. The general representation uses a finite nonnegative spectral measure FF:

γ(k)=∫−ππeikω dF(ω).\gamma(k)= \int_{-\pi}^{\pi}e^{ik\omega}\,dF(\omega).

A persistent sinusoidal component contributes point masses to this measure. More complicated singular measures are also possible. The general representation therefore cannot always be replaced by the Fourier transform of an integrable function. It is the spectral measure, rather than a smooth spectrum, that is fundamental.

The shape of the autocorrelation has a direct spectral interpretation. Slow variation and positive persistence emphasize low frequencies. Alternating correlation emphasizes frequencies near the upper end of the discrete-time range. A narrow spectral peak corresponds to an autocorrelation that oscillates over many lags. For the first-order autoregressive process,

f(ω)=σε22π(1+ϕ2−2ϕcos⁡ω).f(\omega)= \frac{\sigma_\varepsilon^2} {2\pi\left(1+\phi^2-2\phi\cos\omega\right)}.

As positive ϕ\phi approaches one, the spectrum becomes increasingly concentrated near zero frequency.

Linear filtering makes the connection especially useful. If a stationary input passes through a stable linear filter with frequency response H(ω)H(\omega), then its output spectrum is

fY(ω)=∣H(ω)∣2fX(ω).f_Y(\omega)=|H(\omega)|^2f_X(\omega).

The corresponding covariance can be recovered by inverse transformation. Filtering therefore modifies autocorrelation in a predictable way, whether the operation is smoothing, differencing, or selecting a frequency band. Conversely, an autocorrelation alone generally does not reveal a signal’s Fourier phase, so it cannot uniquely reconstruct the original waveform.

Estimation from finite data and numerical computation

For observations x1,…,xnx_1,\ldots,x_n, let

xˉ=1n∑t=1nxt.\bar{x}=\frac1n\sum_{t=1}^{n}x_t.

A common sample autocovariance estimator is

γ^(k)=1n∑t=1n−k(xt+k−xˉ)(xt−xˉ),0≤k<n,\widehat{\gamma}(k)= \frac1n\sum_{t=1}^{n-k} (x_{t+k}-\bar{x})(x_t-\bar{x}), \qquad 0\leq k<n,

with sample autocorrelation

ρ^(k)=γ^(k)γ^(0).\widehat{\rho}(k)= \frac{\widehat{\gamma}(k)}{\widehat{\gamma}(0)}.

The graph of these estimates against lag is often called a correlogram. The denominator cancels in the ratio, leaving the lagged cross-product sum divided by the total centered sum of squares.

Another convention divides the autocovariance sum by n−kn-k. If the true mean is known and used for centering, this removes the finite-pair-count bias:

E⁡ ⁣[1n−k∑t=1n−k(Xt+k−μ)(Xt−μ)]=γ(k).\operatorname{E}\!\left[ \frac{1}{n-k}\sum_{t=1}^{n-k} (X_{t+k}-\mu)(X_t-\mu) \right]=\gamma(k).

With denominator nn, the expectation is instead (1−k/n)γ(k)(1-k/n)\gamma(k). Replacing the true mean by the sample mean introduces additional bias, so the n−kn-k convention is not generally exactly unbiased in ordinary applications.

The denominator nn has an important structural advantage. Extend the centered data by zeros outside their observed interval. Their full lagged-product sequence is positive semidefinite, and its Fourier transform is nonnegative. Dividing individual lags by different factors can destroy that property. Consequently, an estimator that improves bias at a particular lag may produce an invalid covariance sequence when all estimated lags are considered together. Estimated autocorrelations, being ratios of random quantities, have their own bias beyond that of either component.

Direct computation through maximum lag KK takes O(nK)O(nK) arithmetic operations. Computing every lag directly takes O(n2)O(n^2). The fast Fourier transform reduces full autocorrelation computation to O(nlog⁡n)O(n\log n): center the observations, zero-pad them, transform them, multiply the transform by its complex conjugate, and apply the inverse transform. Padding to at least 2n−12n-1 samples prevents wraparound and yields linear rather than circular correlation. The final scaling depends on the transform convention and desired covariance normalization.

Large lags use few overlapping pairs and are correspondingly uncertain. Near lag n−1n-1, a single product determines the numerator. Missing observations add another complication: using only available pairs gives different lag-specific sample sizes and may break positive semidefiniteness. Irregular sampling is more fundamental still, because a separation of three observation indices need not represent a fixed elapsed time. In such cases, covariance models defined on actual time differences or specialized pair-binning estimators distinguish temporal separation from observation order.

Sampling uncertainty and tests of serial dependence

An estimated autocorrelation fluctuates even when the population autocorrelation is zero. For independent, identically distributed observations with suitable finite moments, the sample autocorrelation at a fixed nonzero lag is asymptotically normal with variance approximately 1/n1/n. This motivates the familiar reference bands

±1.96n.\pm\frac{1.96}{\sqrt n}.

They are approximate pointwise 95% bands under a white-noise null model, not universal confidence intervals for arbitrary processes. They also assume a nondegenerate variance and an asymptotic regime in which the chosen lag remains fixed as sample size grows.

Examining many lags changes the interpretation. Even with independent observations, some sample correlations can fall outside pointwise bands by chance. The lag estimates are generally dependent, so the graph is not a collection of independent tests. This is a multiple-testing issue: a single unusual bar has a different evidential meaning from a preselected test at one lag or a sustained, structured pattern across many lags.

When the process is already correlated, the white-noise variance approximation is inappropriate. Bartlett-type approximations express the covariance of sample autocorrelations through the population correlation sequence. For a moving-average process of order qq, at lags k>qk>q a standard asymptotic expression is

Var⁡[ρ^(k)]≈1n(1+2∑j=1qρ(j)2).\operatorname{Var}[\widehat{\rho}(k)] \approx \frac1n\left(1+2\sum_{j=1}^{q}\rho(j)^2\right).

The increased variance reflects information shared across observations. Such formulas require assumptions controlling moments and dependence; they are not automatically valid for long-memory processes or data with undefined second moments.

Joint tests aggregate several lag correlations. The Box–Pierce statistic, introduced by George Box and David Pierce in 1970, is

QBP=n∑k=1hρ^(k)2.Q_{\mathrm{BP}}= n\sum_{k=1}^{h}\widehat{\rho}(k)^2.

The Ljung–Box modification, introduced by Greta Ljung and Box in 1978, is

QLB=n(n+2)∑k=1hρ^(k)2n−k.Q_{\mathrm{LB}}= n(n+2)\sum_{k=1}^{h} \frac{\widehat{\rho}(k)^2}{n-k}.

Under an appropriate white-noise null and regularity conditions, these are compared with a chi-squared distribution. For residuals from a fitted autoregressive moving-average model, a common asymptotic degrees-of-freedom adjustment is h−p−qh-p-q, where pp and qq are the fitted autoregressive and moving-average orders and hh exceeds their sum.

These portmanteau tests address collections of linear correlations. Failure to reject does not establish independence, Gaussianity, constant conditional variance, or general model correctness. A process can have uncorrelated observations but correlated squares or absolute values. Residual diagnostics therefore often examine transformations as well as raw residuals. Parameter estimation, conditional heteroskedasticity, and data-dependent lag selection can alter reference distributions, making the precise null model central to hypothesis testing.

Linear prediction, autoregressive models, and partial autocorrelation

Autocovariance determines optimal linear prediction whenever the relevant second moments exist. Suppose a centered stationary process is predicted from its last pp observations:

X^t=∑j=1pajXt−j.\widehat X_t= \sum_{j=1}^{p}a_jX_{t-j}.

Minimizing mean squared error gives the normal equations

∑j=1pajγ(i−j)=γ(i),i=1,…,p.\sum_{j=1}^{p}a_j\gamma(i-j)=\gamma(i), \qquad i=1,\ldots,p.

These follow by making the prediction error uncorrelated with every predictor. Geometrically, prediction is an orthogonal projection onto the span of past observations in a space of square-integrable random variables.

For an autoregressive process of order pp,

Xt=ϕ1Xt−1+⋯+ϕpXt−p+εt,X_t=\phi_1X_{t-1}+\cdots+\phi_pX_{t-p}+\varepsilon_t,

the corresponding relations are the Yule–Walker equations:

γ(k)=∑j=1pϕjγ(k−j),k≥1.\gamma(k)=\sum_{j=1}^{p}\phi_j\gamma(k-j), \qquad k\geq1.

Their name reflects foundational work by G. Udny Yule in 1927 and Gilbert Walker in 1931. Together with

γ(0)=∑j=1pϕjγ(j)+σε2,\gamma(0)= \sum_{j=1}^{p}\phi_j\gamma(j)+\sigma_\varepsilon^2,

they connect model parameters, innovation variance, and autocovariance. Replacing population covariances by estimates produces a standard parameter-estimation method.

The partial autocorrelation function separates a direct linear association at lag kk from association mediated through intervening observations. For a real stationary process, it is the correlation between the residuals obtained by linearly regressing XtX_t and Xt−kX_{t-k} on Xt−1,…,Xt−k+1X_{t-1},\ldots,X_{t-k+1}. Equivalently, under nonsingularity, it is the final coefficient in the order-kk linear prediction equations.

This distinction supports classical model identification. A stationary autoregressive process of order pp has partial autocorrelations equal to zero beyond pp, while its ordinary autocorrelation usually tails off. A moving-average process of order qq has ordinary autocorrelation equal to zero beyond qq, while its partial autocorrelation generally tails off. The population patterns are exact, but finite-sample plots only approximate them, sometimes poorly when parameters are near cancellation or near a stationarity boundary.

A mixed autoregressive moving-average model can produce both decaying and oscillatory correlation patterns. For a causal stationary autoregressive representation, the roots of

1−ϕ1z−⋯−ϕpzp=01-\phi_1z-\cdots-\phi_pz^p=0

lie outside the unit circle. Their locations determine the rates and frequencies of covariance decay. Efficient recursions exploit the Toeplitz structure of the prediction equations, connecting autocorrelation analysis with numerical linear algebra. The resulting predictor is optimal among linear predictors; it equals the full conditional expectation only under additional distributional structure, such as joint Gaussianity.

Nonstationarity, long memory, and interpretation limits

A large empirical autocorrelation does not by itself identify a stable stochastic mechanism. A deterministic trend can make nearby observations similar simply because both lie on the same slowly changing trajectory. Computing the sample ACF after subtracting only a single overall mean can then produce a long positive tail, even when the deviations around the trend are independent. Seasonal means, level shifts, and changes in measurement practice can create similarly misleading patterns.

Differencing changes the dependence structure rather than merely removing a graphical feature. If

Yt=Xt−Xt−1,Y_t=X_t-X_{t-1},

and XtX_t is stationary, then

γY(k)=2γX(k)−γX(k−1)−γX(k+1).\gamma_Y(k)= 2\gamma_X(k)-\gamma_X(k-1)-\gamma_X(k+1).

In frequency terms, differencing multiplies the spectrum by

∣1−e−iω∣2=4sin⁡2(ω/2),|1-e^{-i\omega}|^2= 4\sin^2(\omega/2),

suppressing low frequencies. Differencing can transform a random walk into its stationary increments, but unnecessary differencing can introduce negative short-lag correlation. For instance, differenced independent noise has lag-one correlation −1/2-1/2.

True long-range dependence is distinct from a trend. One common form has

ρ(k)∼Ck−α,0<α<1,C>0.\rho(k)\sim Ck^{-\alpha}, \qquad 0<\alpha<1,\quad C>0.

The positive autocorrelations are not summable. Equivalently, under suitable conditions, the spectral density diverges near zero frequency. Such a long-memory process can have a constant mean and variance while retaining dependence over widely separated times. The variance of its sample mean often decreases more slowly than the usual 1/n1/n rate, so ordinary short-memory uncertainty formulas no longer apply.

Temporal aggregation also changes autocorrelation. Monthly averages of daily observations are filtered and subsampled versions of the original process; their dependence is not simply the daily autocorrelation evaluated every thirty lags. Sampling itself can cause aliasing, in which distinct continuous-time frequencies become indistinguishable on a discrete observation grid. Therefore, an oscillatory correlogram must be interpreted relative to the measurement interval and the acquisition procedure.

Autocorrelation is symmetric in lag for real stationary processes and consequently does not encode a direction of influence. It cannot, by itself, establish causation, distinguish an external driver from internal dynamics, or determine whether an apparent cycle is generated by feedback. Moreover, processes with infinite variance, including many heavy-tailed models, lack the conventional covariance-based ACF altogether. Rank-based dependence measures, characteristic-function methods, and other alternatives describe different aspects of temporal dependence, rather than repairing an undefined second moment.

Deterministic signals, detection, and physical measurements

For a finite-energy signal, autocorrelation has a direct geometric interpretation. The value

Rx(τ)=∫x(t+τ)x(t)‾ dtR_x(\tau)= \int x(t+\tau)\overline{x(t)}\,dt

is an inner product between the signal and a translated copy. At zero lag it equals the signal energy,

Rx(0)=∫∣x(t)∣2 dt,R_x(0)=\int |x(t)|^2\,dt,

and the Cauchy–Schwarz inequality gives

∣Rx(τ)∣≤Rx(0).|R_x(\tau)|\leq R_x(0).

A peak at a nonzero lag indicates that the signal resembles a shifted version of itself. Repeated peaks can expose a period even when the waveform is complicated or contaminated by noise.

Autocorrelation can also be written as a convolution:

Rx(τ)=(x∗x~)(τ),x~(t)=x(−t)‾.R_x(\tau)= (x*\widetilde{x})(\tau), \qquad \widetilde{x}(t)=\overline{x(-t)}.

Its Fourier transform is ∣X(ω)∣2|X(\omega)|^2, where XX is the transform of xx. This is the deterministic energy-signal counterpart of the stationary-process spectral relation. Because the transform phase disappears, different signals can share the same autocorrelation. A time shift, for example, leaves the autocorrelation unchanged. The associated reconstruction ambiguity is central to phase retrieval.

In radar and sonar, the correlation of a received waveform with a known transmitted waveform is technically cross-correlation. When the received echo is an attenuated, delayed copy, its delay response is governed by the transmitted waveform’s autocorrelation. A narrow central peak improves delay discrimination, while large sidelobes can obscure weaker nearby echoes. This is one reason coded pulses and carefully designed sequences are used. For radar propagation in a uniform medium, a round-trip delay τ\tau corresponds to range vτ/2v\tau/2, where vv is the propagation speed.

In speech and music analysis, a quasi-periodic waveform produces autocorrelation peaks near integer multiples of its fundamental period. If the sampling rate is fsf_s and a relevant peak occurs at lag k0k_0, a candidate fundamental frequency is fs/k0f_s/k_0. Harmonics, changes in pitch, finite windows, and noise can create competing peaks, so the largest nonzero-lag peak does not always identify the perceived fundamental.

Physical correlation functions often involve a measured observable rather than the raw position or field. For an isotropic equilibrium fluid in three dimensions, the Green–Kubo relation for diffusion is

D=13∫0∞⟨v(t)⋅v(0)⟩ dt,D=\frac13\int_0^\infty \left\langle \mathbf v(t)\cdot\mathbf v(0) \right\rangle\,dt,

when the equilibrium and long-time limits required by the relation exist. The velocity autocorrelation therefore links microscopic dynamics to a macroscopic transport coefficient. Related equilibrium time-correlation formulas connect stress fluctuations with viscosity and heat-current fluctuations with thermal conductivity; the observable, normalization, and physical assumptions differ in each case.

Correlated simulations, spatial fields, and multivariate extensions

Autocorrelation directly affects the uncertainty of an average. For nn observations from a stationary real process,

Var⁡(Xˉn)=γ(0)n[1+2∑k=1n−1(1−kn)ρ(k)].\operatorname{Var}(\bar X_n)= \frac{\gamma(0)}{n} \left[ 1+2\sum_{k=1}^{n-1} \left(1-\frac{k}{n}\right)\rho(k) \right].

This identity follows by summing all entries of the covariance matrix. It shows why nn correlated measurements do not generally carry the same information about the mean as nn independent measurements. Positive correlations increase the variance; sufficiently structured negative correlations can reduce it.

If the correlations are absolutely summable, define, under one common convention,

τint=1+2∑k=1∞ρ(k).\tau_{\mathrm{int}}= 1+2\sum_{k=1}^{\infty}\rho(k).

Then, for large nn,

Var⁡(Xˉn)≈γ(0)τintn.\operatorname{Var}(\bar X_n) \approx \frac{\gamma(0)\tau_{\mathrm{int}}}{n}.

The quantity τint\tau_{\mathrm{int}} is the integrated autocorrelation time. Another convention assigns half this value, so numerical comparisons require checking the definition. For a first-order autoregressive process,

τint=1+ϕ1−ϕ.\tau_{\mathrm{int}}=\frac{1+\phi}{1-\phi}.

Under the corresponding asymptotic mean-variance interpretation, an effective sample size is n/τintn/\tau_{\mathrm{int}}.

These formulas are especially important in Markov chain Monte Carlo. Successive simulated states are intentionally dependent, and each scalar observable can have a different autocorrelation time. A rapidly mixing coordinate does not establish that every parameter or derived quantity is well explored. Estimating an infinite sum from a finite chain requires truncation or smoothing because large-lag estimates are noisy. Autocorrelation calculations also presuppose an appropriate stationary regime; they cannot reveal an unvisited region of the target distribution simply from fluctuations inside the visited region.

For a vector-valued process XtX_t, autocovariance becomes matrix-valued:

Γ(k)=E⁡[(Xt+k−μ)(Xt−μ)∗].\Gamma(k)= \operatorname{E}\big[ (X_{t+k}-\mu)(X_t-\mu)^\ast \big].

Its diagonal entries are component autocovariances, and its off-diagonal entries describe cross-dependence at different lags. Hermitian symmetry becomes Γ(−k)=Γ(k)∗\Gamma(-k)=\Gamma(k)^\ast, and the spectral counterpart is a positive-semidefinite matrix at each frequency. Unlike a scalar real ACF, an individual cross-covariance between two different components need not be even in lag.

Spatial random fields replace temporal lag by a displacement vector hh. A stationary covariance C(h)C(h) may depend on direction as well as distance; under isotropy it depends only on ∥h∥\|h\|. For a real stationary field, the semivariogram is

v(h)=12E⁡[(X(s+h)−X(s))2]=C(0)−C(h).v(h)=\frac12 \operatorname{E}\big[(X(s+h)-X(s))^2\big] =C(0)-C(h).

This relation links autocovariance with spatial interpolation and geostatistical modeling. Validity still requires positive semidefiniteness over every finite set of locations, and a radial function admissible in one spatial dimension need not be admissible in every higher dimension.