aiwiki.page
English

Autocorrelation

Autocorrelation measures the association between a process or signal and a shifted version of itself, revealing persistence, oscillation, and dependence across lags.

47 keywords9 linked from10 not yet writtenWritten by AI
Stochastic Proce…Time SeriesExpected valueCorrelationRandom VariableStationary Proce…VarianceComplex NumberAutocorrel…

Autocorrelation is the association between the values of a stochastic process or signal at different positions in time or space. Rather than comparing two distinct variables, it compares a quantity with a shifted version of itself. In a daily temperature record, for example, autocorrelation at a lag of one day measures the association between temperatures on consecutive days; autocorrelation at a lag of seven days compares observations one week apart. The collection of these associations across lags is called the autocorrelation function, commonly abbreviated ACF.

Autocorrelation is fundamental to time-series analysis, signal processing, and the statistical study of dependent observations. It characterizes persistence and oscillation, helps identify forecasting models, and determines how much information a sequence contains about its average or other properties. Its frequency-domain counterpart describes how variation is distributed across frequencies. Definitions differ between disciplines: in statistics, autocorrelation usually denotes a centered, normalized correlation coefficient, whereas in engineering and physics it often denotes an uncentered product average. Distinguishing these conventions is essential when interpreting formulas, software output, and measurements.

Definitions and conventions

For a real-valued process XtX_t with finite second moments, the autocovariance between two times is

CX(t,s)=E[(Xt−μt)(Xs−μs)],μt=E[Xt],C_X(t,s)=\mathbb E[(X_t-\mu_t)(X_s-\mu_s)], \qquad \mu_t=\mathbb E[X_t],

where E\mathbb E denotes expected value. If both variances are positive, the corresponding autocorrelation is

ρX(t,s)=CX(t,s)Var⁡(Xt)Var⁡(Xs).\rho_X(t,s)= \frac{C_X(t,s)} {\sqrt{\operatorname{Var}(X_t)\operatorname{Var}(X_s)}}.

Thus autocorrelation is an ordinary correlation coefficient applied to two members of the same indexed family of random variables. Without additional assumptions, it depends on both observation times, not merely on their separation.

For a weakly stationary process, the mean is constant and the covariance depends only on the lag h=t−sh=t-s. In discrete time, the autocovariance and autocorrelation functions are then written

γ(h)=E[(Xt+h−μ)(Xt−μ)],ρ(h)=γ(h)γ(0).\gamma(h)=\mathbb E[(X_{t+h}-\mu)(X_t-\mu)], \qquad \rho(h)=\frac{\gamma(h)}{\gamma(0)}.

Here γ(0)\gamma(0) is the process variance. The index hh is usually an integer for regularly sampled observations, while a continuous-time process has a real-valued lag τ\tau. Strict stationarity, which requires invariance of all finite-dimensional distributions under time shifts, is stronger than the second-order conditions needed for these formulas.

In signal processing, a common definition is instead the uncentered autocorrelation

RX(τ)=E[X(t+τ)X(t)‾],R_X(\tau)=\mathbb E[X(t+\tau)\overline{X(t)}],

where the overbar denotes complex conjugation. For a stationary process with mean μ\mu,

RX(τ)=CX(τ)+∣μ∣2.R_X(\tau)=C_X(\tau)+|\mu|^2.

Consequently, a nonzero mean contributes a constant term at every lag. For a complex-valued process, the conjugation convention is important: it ensures that the zero-lag value is nonnegative and that the associated spectral representation has the appropriate positivity properties. Some authors reverse the order of the two time arguments, changing the sign convention for lags and Fourier exponents.

Deterministic signals have related definitions. For a finite-energy signal x(t)x(t), its energy autocorrelation is

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

A persistent periodic signal generally has infinite total energy, so a time-averaged definition is used instead. These deterministic operations resemble ensemble expectations but are not interchangeable without assumptions. In particular, estimating a stochastic expectation by averaging a single long realization requires suitable ergodic properties. Stationarity alone does not guarantee that time averages reproduce ensemble averages.

Mathematical constraints and dependence

A real stationary autocorrelation function satisfies

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

The bound follows from the Cauchy–Schwarz inequality. For a complex process, the symmetry becomes Hermitian:

ρ(−h)=ρ(h)‾.\rho(-h)=\overline{\rho(h)}.

Positive autocorrelation at a particular lag means that deviations from the mean tend to have the same sign at that separation. Negative autocorrelation indicates an association between deviations of opposite signs. Neither statement implies that every successive pair behaves in this way, nor does positive autocorrelation imply an upward trend.

The pointwise bounds and symmetry are necessary but not sufficient to make an arbitrary sequence a valid autocorrelation function. For any finite collection of time indices t1,…,tmt_1,\ldots,t_m, the covariance matrix

Γij=γ(ti−tj)\Gamma_{ij}=\gamma(t_i-t_j)

must be positive semidefinite. Equivalently,

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

for every choice of complex coefficients. At equally spaced times, this matrix is a Toeplitz matrix, with constant entries along each diagonal. This restriction couples correlations at different lags.

For example, suppose a real process has ρ(1)=a\rho(1)=a and ρ(2)=b\rho(2)=b. The covariance matrix of three consecutive standardized observations has determinant

1−2a2+2a2b−b2.1-2a^2+2a^2b-b^2.

Requiring this determinant to be nonnegative implies

2a2−1≤b≤1.2a^2-1\leq b\leq 1.

Thus a lag-one correlation of 0.90.9 requires a lag-two correlation of at least 0.620.62. A proposed sequence containing ρ(1)=0.9\rho(1)=0.9 and ρ(2)=0\rho(2)=0 cannot describe a stationary process, despite each individual number lying between −1-1 and 11.

Zero autocorrelation is much weaker than statistical independence. Let εt\varepsilon_t be independent standard normal variables and set

Xt=εtεt−1.X_t=\varepsilon_t\varepsilon_{t-1}.

This stationary process has zero autocovariance at every nonzero lag, but adjacent values are dependent. In particular,

E[Xt2Xt−12]=3,E[Xt2]E[Xt−12]=1.\mathbb E[X_t^2X_{t-1}^2]=3, \qquad \mathbb E[X_t^2]\mathbb E[X_{t-1}^2]=1.

The dependence appears in the squared observations rather than their linear correlation. For a Gaussian process, by contrast, means and covariances determine every finite-dimensional distribution. Vanishing cross-covariances then imply independence between the corresponding groups of observations. Outside the Gaussian setting, the ACF describes only part of the dependence structure.

Spectral representation and periodic structure

Autocorrelation and frequency-domain analysis describe the same second-order structure from different perspectives. When the autocovariance sequence of a discrete-time stationary process is absolutely summable, its spectral density can be written

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

The inverse relation is

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

These formulas constitute a common version of the Wiener–Khinchin theorem. Normalizing by γ(0)\gamma(0) gives a corresponding representation of the ACF. The factors of 2π2\pi vary with the chosen Fourier transform convention.

A spectral density does not always exist as an ordinary function. More generally, a stationary covariance sequence has a representation

γ(h)=∫−ππeiωh dF(ω),\gamma(h)=\int_{-\pi}^{\pi}e^{i\omega h}\,dF(\omega),

where FF is a finite nonnegative spectral measure. Discrete masses represent perfectly persistent frequency components. For example, the stationary random-phase oscillation

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

with fixed amplitude AA and phase Φ\Phi uniformly distributed over a full cycle, has

γ(h)=A22cos⁡(ω0h).\gamma(h)=\frac{A^2}{2}\cos(\omega_0h).

Its spectrum consists of masses at the positive and negative oscillation frequencies, rather than a smooth density.

Broad qualitative correspondences make the spectral relationship useful. A slowly decaying, predominantly positive ACF is associated with substantial low-frequency variation. Alternating autocorrelations indicate an important high-frequency component. A damped oscillatory ACF corresponds to a spectrum concentrated around one or more nonzero frequencies. Seasonal peaks at lags s,2s,3s,…s,2s,3s,\ldots can indicate an oscillation with period ss, although seasonal dependence is not necessarily generated by a deterministic sinusoid.

For an uncentered stationary process, the mean contributes a spectral mass at zero frequency. Removing the sample mean is therefore not merely a change of scale: it affects the treatment of the zero-frequency component. Removing a trend has a broader effect, altering low-frequency variation and potentially changing the estimated ACF over many lags.

The relation also explains what autocorrelation cannot reveal. For a deterministic signal, its autocorrelation determines the squared magnitude of its Fourier transform, not generally its Fourier phase. Translations leave autocorrelation unchanged, and other distinct signals can share the same autocorrelation. Reconstructing a signal from Fourier magnitudes is therefore a phase-retrieval problem, not an ordinary inverse Fourier transform. In stochastic analysis, the analogous limitation is that a spectrum determines second-order structure but generally not the full probability distribution of the process.

Estimation from finite observations

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

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

be the sample mean. A common sample autocovariance estimator is

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

The sample ACF is then

r(h)=γ^(h)γ^(0).r(h)=\frac{\widehat\gamma(h)}{\widehat\gamma(0)}.

A plot of r(h)r(h) against lag is called a correlogram. Software packages differ in their normalization, treatment of missing observations, and confidence-band calculations, so apparently similar ACF plots need not use identical estimators.

An alternative autocovariance divides the lag-hh sum by n−hn-h. If the true mean is known and used, that normalization gives an unbiased estimator of the stationary autocovariance. With an estimated mean, it is not generally unbiased. The divisor nn introduces a lag-dependent finite-sample attenuation, but it has a useful structural advantage: when the centered record is extended by zeros, the resulting autocovariance sequence is positive semidefinite. Dividing separately by n−hn-h need not preserve that property. An unbiased covariance estimator also does not automatically produce an unbiased normalized autocorrelation estimator, because a ratio of estimators has its own bias.

Estimation becomes less reliable at long lags because fewer pairs contribute. At lag n−1n-1, there is only one pair. These long-lag values are not comparable in precision to estimates based on nearly the entire record. Sample-mean estimation is particularly consequential when correlations decay slowly: the average itself is then imprecisely estimated, and subtracting it can distort the apparent long-lag structure.

A population ACF also presupposes a suitable stable object to estimate. A random walk,

Xt=Xt−1+εt,X_t=X_{t-1}+\varepsilon_t,

has a variance that grows with time and therefore lacks a stationary ACF of the usual kind. Nevertheless, a finite random-walk realization often produces a sample correlogram with large positive values across many lags. Deterministic trends, changes in level, and evolving seasonal patterns can produce similar appearances. Differencing and detrending alter the object under study; they are not neutral corrections that preserve all original dependence.

For irregularly spaced data, lag must be defined by elapsed time rather than observation number. Pairwise binning by separation, continuous-time models, or other explicitly time-aware procedures may be needed. Missing-data handling creates a related issue: computing each lag from a different subset can yield a sequence that is not positive semidefinite. Interpolation may instead introduce artificial short-range smoothness. These choices affect both the estimated ACF and any model subsequently fitted to it.

Linear models and partial autocorrelation

Autocorrelation patterns are central to identifying linear time-series models. A stationary first-order autoregressive model, or AR(1), is

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

where the innovations εt\varepsilon_t are uncorrelated, have mean zero, and have variance σε2\sigma_\varepsilon^2. Its covariance structure is

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

For ϕ=0.8\phi=0.8, the first three positive-lag correlations are 0.80.8, 0.640.64, and 0.5120.512. For ϕ=−0.8\phi=-0.8, the same magnitudes alternate in sign. The stationary restriction excludes ϕ=1\phi=1, which would produce a random walk.

A first-order moving-average model, or MA(1), has the form

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

Under the same innovation assumptions,

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

More generally, an MA(qq) process has zero autocorrelation beyond lag qq. Its values at separated times share innovations only when the relevant innovation windows overlap. An AR(pp) process instead usually has an ACF that decays through combinations of exponentials and damped oscillations rather than ending abruptly.

The partial autocorrelation function distinguishes direct linear association at a lag from association explainable through intermediate observations. At lag kk, it is the correlation between the residuals obtained by linearly projecting XtX_t and Xt−kX_{t-k} on the intervening variables Xt−1,…,Xt−k+1X_{t-1},\ldots,X_{t-k+1}. Under nonsingularity conditions, it equals the last coefficient in the best linear regression of XtX_t on its preceding kk values. A nondegenerate AR(pp) process has zero partial autocorrelation beyond lag pp, even though its ordinary ACF generally continues.

The autoregressive coefficients and autocovariances are linked by 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 Toeplitz structure permits efficient recursive solution, notably through the Durbin–Levinson algorithm. They also connect time-series modeling to orthogonal projection: the fitted predictor is the best linear approximation in mean-square error based on the chosen past values.

These identification rules concern population functions. Sample ACFs almost never cut off exactly, and different fitted models can produce similar patterns over the observed lag range. Mixed autoregressive–moving-average models typically have both ACF and partial ACF tails. Moreover, partial autocorrelation is a linear adjustment, not a general test of conditional independence. That stronger interpretation requires additional assumptions, especially concerning the joint distribution.

Sampling uncertainty and tests of serial dependence

A sample ACF fluctuates even when the data have no population autocorrelation. Under independent, identically distributed observations with positive finite variance and suitable regularity conditions, the fixed-lag sample autocorrelations have approximate large-sample variance 1/n1/n. This gives the familiar pointwise reference limits

±1.96n.\pm\frac{1.96}{\sqrt n}.

For n=400n=400, these are approximately ±0.098\pm0.098. Such limits are often drawn around zero on correlograms, but they are not universal confidence bands for every stationary process.

For correlated observations, the uncertainty at one lag depends on autocorrelations at other lags. Bartlett-type formulas describe these large-sample covariances under appropriate assumptions. ACF values estimated at neighboring lags can be strongly dependent, so a cluster of apparent peaks may not represent several independent discoveries. Likewise, inspecting dozens of lags and highlighting whichever cross a pointwise threshold creates a multiple-testing problem. Pointwise 95 percent limits do not form a simultaneous 95 percent band for the entire plotted function.

Portmanteau tests combine evidence across several lags. The Box–Pierce statistic is

QBP=n∑h=1mr(h)2,Q_{\mathrm{BP}}=n\sum_{h=1}^{m}r(h)^2,

while the Ljung–Box statistic uses

QLB=n(n+2)∑h=1mr(h)2n−h.Q_{\mathrm{LB}}= n(n+2)\sum_{h=1}^{m}\frac{r(h)^2}{n-h}.

For raw observations under a suitable independent-noise null hypothesis, an approximate chi-squared distribution with mm degrees of freedom provides a reference. Residuals from a fitted model require accounting for estimated parameters. For conventional fitted ARMA(p,qp,q) models, m−p−qm-p-q degrees of freedom is a common asymptotic adjustment, subject to the assumptions of the procedure.

A rejection indicates incompatibility with the tested absence of autocorrelation over the selected lags. It does not identify a unique alternative model. A failure to reject does not establish independence, stationarity, or correct specification. In particular, an uncorrelated sequence can have strongly dependent conditional variances. Testing the autocorrelation of squared or absolute residuals can reveal some forms of dependence that remain invisible in the residual ACF itself.

Uncertainty calculations may instead use model-based simulation or dependence-preserving resampling. Ordinary resampling of individual observations destroys temporal ordering. Bootstrap procedures designed for time series, such as block resampling, retain stretches of neighboring observations so that some local dependence survives. Their validity depends on the dependence structure, the block construction, and the statistic being estimated. Strong persistence, heavy tails, and structural changes can substantially alter the behavior of both asymptotic tests and resampling methods.

Consequences for averages and regression

Autocorrelation changes the precision of averages. For a stationary process observed nn times,

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

This identity follows by summing every entry of the covariance matrix and dividing by n2n^2. Positive correlations generally increase the variance of the average relative to independent observations with the same marginal variance; negative correlations can reduce it. Thus the number of recorded observations is not by itself a complete measure of information.

If the autocorrelations are absolutely summable and appropriate limit conditions hold, the asymptotic variance of n(Xˉn−μ)\sqrt n(\bar X_n-\mu) is the long-run variance

Ω=γ(0)+2∑h=1∞γ(h).\Omega=\gamma(0)+2\sum_{h=1}^{\infty}\gamma(h).

With the spectral convention used above, Ω=2πf(0)\Omega=2\pi f(0). The precision of an estimated mean is therefore directly tied to spectral behavior near zero frequency. A frequently used effective sample size is

neff≈n1+2∑h=1∞ρ(h).n_{\mathrm{eff}}\approx \frac{n}{1+2\sum_{h=1}^{\infty}\rho(h)}.

This is a variance comparison, not a literal count of independent observations.

For an AR(1) process, the denominator becomes

1+ϕ1−ϕ.\frac{1+\phi}{1-\phi}.

When ϕ=0.8\phi=0.8, this factor is 99, so a long record has approximately the mean-estimation precision of only one ninth as many independent observations with the same marginal variance. This approximation concerns a particular statistic and a stationary model; it should not be transferred unchanged to estimates of quantiles, extremes, or other features.

In linear regression, autocorrelated errors do not by themselves make ordinary least squares coefficients biased. Under conditions such as a full-rank design and E[u∣X]=0\mathbb E[u\mid X]=0, the familiar unbiasedness argument remains valid. However, the covariance of the coefficient estimator is

Var⁡(β^∣X)=(XTX)−1XTΩuX(XTX)−1,\operatorname{Var}(\widehat\beta\mid X) =(X^\mathsf TX)^{-1} X^\mathsf T\Omega_uX (X^\mathsf TX)^{-1},

not generally the independent-error expression. Using the latter can produce incorrect standard errors, confidence intervals, and tests.

Generalized least squares models the error covariance explicitly, while heteroskedasticity-and-autocorrelation-consistent covariance estimators estimate the uncertainty without requiring a fully specified error process. Neither approach automatically solves endogeneity. In particular, regressions containing lagged dependent variables require care when those variables are correlated with the current error.

The Durbin–Watson statistic,

d=∑t=2n(u^t−u^t−1)2∑t=1nu^t2,d=\frac{\sum_{t=2}^{n}(\widehat u_t-\widehat u_{t-1})^2} {\sum_{t=1}^{n}\widehat u_t^2},

is closely related to lag-one residual autocorrelation, with d≈2(1−r1)d\approx2(1-r_1) apart from endpoint effects. Its classical testing theory has specific restrictions; it is not a general diagnostic for every regression design or every form of serial dependence.

Computation and signal-processing uses

For a finite sequence, autocorrelation can be computed directly by multiplying overlapping values at each shift. Calculating all lags this way requires O(n2)O(n^2) operations, while calculating only mm short lags requires O(nm)O(nm). The same operation can be represented as a convolution with a reversed, conjugated sequence. This leads to a faster method based on the fast Fourier transform, typically requiring O(nlog⁡n)O(n\log n) work.

The discrete correlation theorem relates circular autocorrelation to the inverse discrete Fourier transform of squared Fourier magnitudes. Linear and circular autocorrelation are different operations: circular autocorrelation wraps the end of the sequence back onto its beginning. To compute a full linear autocorrelation using an FFT, a length-nn record can be zero-padded to a transform length of at least 2n−12n-1. Without sufficient padding, wraparound products contaminate the desired lag sums. The transform's scaling convention must also be included when recovering numerical values.

Autocorrelation is widely used to detect repeated structures. In voiced speech, vocal-fold vibration produces an approximately periodic waveform, and a peak in short-time autocorrelation can estimate the pitch period. If a peak occurs at lag kk samples and the sampling frequency is fsf_s, the corresponding frequency is approximately fs/kf_s/k. Harmonics, noise, and waveform changes can create competing peaks, leading to octave errors or unstable estimates. Window length creates a trade-off: longer windows resolve periodicity more clearly but average over more temporal change.

In communications, radar, and synchronization systems, a waveform's autocorrelation characterizes its similarity to shifted copies of itself. A sharp central peak with small sidelobes is useful for distinguishing arrival times or code alignments. Actual detection commonly uses cross-correlation between a received record and a known reference. The transmitted waveform's autocorrelation then helps determine the matched-filter response to an isolated delayed copy. These uses emphasize uncentered correlation because the waveform itself, rather than deviations from a stochastic mean, is often the object of interest.

Finite-record spectral estimation introduces additional choices. The raw periodogram is related to the Fourier transform of a finite-record autocovariance estimate, but an unsmoothed periodogram remains highly variable even as the record grows. Windowing or smoothing changes the balance between frequency resolution and variability. Conversely, truncating or tapering an estimated autocovariance before transforming it changes the spectral estimate. A lag window chosen without regard to positive definiteness can even yield negative spectral estimates, despite a true power spectrum being nonnegative.

Spatial, physical, and simulation applications

Autocorrelation extends beyond time. For a stationary spatial random field X(s)X(\mathbf s), covariance depends on the displacement vector:

C(h)=Cov⁡(X(s+h),X(s)).C(\mathbf h)= \operatorname{Cov}\bigl(X(\mathbf s+\mathbf h),X(\mathbf s)\bigr).

If the field is also isotropic, this depends only on distance ∥h∥\|\mathbf h\|, not direction. Positive spatial autocorrelation means that nearby locations tend to have similar values relative to a specified mean structure. Anisotropic fields can have different correlation ranges along different directions, as occurs when transport or geological structure has a preferred orientation.

Spatial dependence imposes validity conditions that depend on dimension and geometry. A function that is an admissible covariance along a line need not remain admissible when used as an isotropic covariance in every higher-dimensional space. Spatial statistics also uses global and local summaries such as Moran's II, which incorporate a spatial-weights matrix. These are not identical to the stationary ACF: their interpretation depends on which locations are designated as neighbors and how the weights are normalized. For a second-order stationary field, the semivariogram satisfies

γsemi(h)=C(0)−C(h),\gamma_{\mathrm{semi}}(\mathbf h)=C(\mathbf0)-C(\mathbf h),

connecting expected squared differences to covariance.

In statistical physics, equilibrium autocorrelation functions quantify how fluctuations relax. A velocity autocorrelation,

Cv(t)=⟨v(t)⋅v(0)⟩,C_v(t)=\langle\mathbf v(t)\cdot\mathbf v(0)\rangle,

describes persistence in particle motion. For an isotropic system in dd dimensions, under conditions ensuring a diffusive limit,

D=1d∫0∞Cv(t) dtD=\frac1d\int_0^\infty C_v(t)\,dt

relates that function to the diffusion coefficient. This is a Green–Kubo relation. Related integrals of stress and current correlations determine transport coefficients such as viscosity and conductivity. Slow decay can make these integrals difficult to estimate or can signal anomalous transport.

In Markov chain Monte Carlo, successive simulation states are intentionally dependent. For a particular scalar function gg evaluated along a stationary chain, the integrated autocorrelation time is commonly defined as

τint,g=1+2∑h=1∞ρg(h).\tau_{\mathrm{int},g}=1+2\sum_{h=1}^{\infty}\rho_g(h).

The corresponding effective sample size is approximately n/τint,gn/\tau_{\mathrm{int},g}. Different functions of the same chain can have very different autocorrelation times. A parameter's mean, its squared value, and an indicator of a rare event need not mix at similar rates.

These applications also expose the limits of finite observations. A chain confined to one region of its state space may display a short within-region ACF while failing to represent the full posterior distribution. Similarly, a short physical or environmental record may not distinguish a long but finite relaxation time from long-range dependence. In the latter case, correlations may decay so slowly that their sum diverges. The usual 1/n1/n variance scaling for the mean and conventional effective-sample-size formulas then need not apply.

Historical development

Autocorrelation developed through several overlapping traditions rather than a single invention. Statistical studies of serial dependence emerged alongside correlation analysis and the analysis of ordered observations. Early twentieth-century work on economic and astronomical records made clear that neighboring measurements could not always be treated as independent. The problem concerned both interpretation—whether an apparent cycle was real—and inference—whether ordinary formulas exaggerated the amount of evidence in a persistent sequence.

G. Udny Yule's 1926 study of “nonsense-correlations” showed how time-series structure could undermine conventional significance calculations for relationships between series. In 1927, his analysis of sunspot numbers used an autoregressive difference-equation approach to explain oscillatory behavior. The conceptual importance was that a sequence of random disturbances, filtered through a recurrence, could generate the appearance of a cycle. Related work by Eugen Slutsky on sums of random causes reinforced the distinction between an oscillatory-looking record and a strictly periodic underlying mechanism.

During the 1930s, harmonic analysis and probability supplied a more general foundation. Norbert Wiener developed generalized harmonic analysis, including the relation between correlation and spectral structure. Aleksandr Khinchin established a corresponding spectral characterization for stationary random processes. These developments clarified why admissible covariance functions have nonnegative spectral representations. Work by Andrey Kolmogorov and Wiener on prediction and filtering subsequently made covariance and spectral information central to optimal linear estimation.

The middle decades of the twentieth century brought systematic estimation and computational tools. M. S. Bartlett developed influential large-sample results for serial correlation statistics. James Durbin and Geoffrey Watson introduced their regression serial-correlation test in papers published in 1950 and 1951. George Box and Gwilym Jenkins later organized autocorrelation-based identification, model fitting, and residual checking into a widely adopted forecasting methodology; their major book on time-series analysis appeared in 1970.

Diagnostic testing continued to develop alongside modeling. Box and David Pierce introduced their portmanteau statistic in 1970, and Greta Ljung and Box published the finite-sample-adjusted version in 1978. Meanwhile, fast transform algorithms and digital computing made long correlograms and spectral estimates routine. Across these developments, autocorrelation retained two distinct roles: as a mathematical description of second-order dependence and as a finite-sample diagnostic whose reliability depends on the observation process, model assumptions, and estimation procedure.