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

Correlation Function

A correlation function describes how fluctuations or measured quantities at different times, positions, or other indexed arguments are statistically related.

46 keywords2 linked from14 not yet writtenWritten by AI
CorrelationStochastic Proce…Expected valueCovarianceRandom VariableVarianceStatistical Inde…Covariance matri…Correlatio…

A correlation function is a function that describes statistical relationships between quantities evaluated at different times, positions, or other indexed arguments. In its most common form, it assigns to each pair of arguments the expected product of two quantities, sometimes after their means have been subtracted or their magnitudes normalized. It extends the idea of correlation from a pair of variables to an entire family of measurements, making it possible to describe temporal memory, spatial organization, and collective fluctuations.

Correlation functions are fundamental in probability, signal analysis, statistical mechanics, and quantum theory. A temporal correlation function can quantify how rapidly a fluctuating signal loses memory of its past; a spatial correlation function can reveal characteristic distances in a material; and a quantum correlation function can encode the probabilities and dynamics of measurement outcomes. Their mathematical properties connect them to positive semidefinite kernels, spectral measures, and Fourier analysis.

Terminology varies across disciplines. “Correlation function” may mean an uncentered product average, a covariance function, a normalized correlation coefficient, or a higher-order expectation involving several quantities. Consequently, the definition, ordering of arguments, and normalization must be specified before formulas from different fields can be compared.

Definitions and conventions

Let X(t)X(t) be a real- or complex-valued stochastic process, where tt belongs to an index set TT. The index may represent time, a point in space, or a more general parameter. Assuming finite second moments, the uncentered autocorrelation function is commonly defined as

RX(s,t)=E[X(s)X(t)‾].R_X(s,t)=\mathbb E[X(s)\overline{X(t)}].

Here E\mathbb E denotes expectation, and the overbar denotes complex conjugation. For real-valued processes, conjugation has no effect. The prefix auto- indicates that both quantities come from the same process. For two processes XX and YY, a cross-correlation function can instead be defined by

RXY(s,t)=E[X(s)Y(t)‾].R_{XY}(s,t)=\mathbb E[X(s)\overline{Y(t)}].

Different engineering and mathematical texts reverse the arguments or place the conjugate on the first factor. These conventions change the sign of a lag or frequency but not the underlying statistical information.

Writing mX(t)=E[X(t)]m_X(t)=\mathbb E[X(t)], the covariance function is

CX(s,t)=E ⁣[(X(s)−mX(s))(X(t)−mX(t))‾].C_X(s,t) = \mathbb E\!\left[ \bigl(X(s)-m_X(s)\bigr) \overline{\bigl(X(t)-m_X(t)\bigr)} \right].

Thus CX(s,t)=RX(s,t)−mX(s)mX(t)‾C_X(s,t)=R_X(s,t)-m_X(s)\overline{m_X(t)}. This function extends covariance to an indexed family of random variables. Its diagonal value CX(t,t)C_X(t,t) is the variance at tt. Subtracting the mean matters physically and statistically: a nonzero constant background can produce a nonzero uncentered correlation even when the fluctuations at distinct arguments are independent.

When both variances are positive, a normalized covariance is

ρX(s,t)=CX(s,t)CX(s,s)CX(t,t).\rho_X(s,t)= \frac{C_X(s,t)} {\sqrt{C_X(s,s)C_X(t,t)}}.

For real-valued processes, this lies between −1-1 and 11; for complex-valued processes, its absolute value is at most 11. Normalization removes physical units and separates the strength of linear association from fluctuation amplitude. If a variable has zero variance, this normalized expression is undefined, although its covariance remains well defined.

The relationship between correlation and dependence requires care. Zero covariance means that the centered variables are orthogonal in a second-moment sense, not that they are generally independent. For example, if ZZ has a standard normal distribution, then ZZ and Z2Z^2 have zero covariance but are dependent. Statistical independence implies zero covariance whenever the required moments exist; the converse holds for jointly Gaussian variables but not for arbitrary distributions.

Some deterministic signals also have correlation functions defined through time integrals rather than probability expectations. For finite-energy functions, a common definition is Rxy(τ)=∫x(t+τ)y(t)‾ dtR_{xy}(\tau)=\int x(t+\tau)\overline{y(t)}\,dt. This is mathematically related to stochastic correlation, but the existence assumptions, normalization, and interpretation differ.

Positive semidefiniteness and admissible kernels

A covariance function cannot be an arbitrary function of two arguments. Its defining restriction is positive semidefiniteness: for every finite set t1,…,tnt_1,\ldots,t_n and every collection of complex coefficients a1,…,ana_1,\ldots,a_n,

∑i,j=1nai‾ CX(ti,tj) aj≥0.\sum_{i,j=1}^{n} \overline{a_i}\,C_X(t_i,t_j)\,a_j\geq 0.

The inequality follows because the left-hand side is the expected squared absolute value of a linear combination of centered variables. Accordingly, the matrix with entries CX(ti,tj)C_X(t_i,t_j) is a positive semidefinite covariance matrix. In kernel terminology, “positive definite” often includes this non-strict condition; in matrix terminology, positive definite usually means strictly positive for every nonzero coefficient vector.

Covariance functions also satisfy Hermitian symmetry,

CX(s,t)=CX(t,s)‾,C_X(s,t)=\overline{C_X(t,s)},

and the Cauchy–Schwarz inequality gives

∣CX(s,t)∣2≤CX(s,s)CX(t,t).|C_X(s,t)|^2\leq C_X(s,s)C_X(t,t).

These restrictions explain why a normalized real autocorrelation cannot exceed one in magnitude. They also show that zero variance at one argument forces its covariance with every other argument to vanish. For real processes, Hermitian symmetry reduces to ordinary symmetry.

Conversely, a real symmetric positive semidefinite kernel, together with a prescribed real mean function, specifies the finite-dimensional distributions of a Gaussian process. For each finite collection of arguments, the corresponding vector is assigned a multivariate normal distribution with that mean vector and covariance matrix. The consistency of these distributions permits construction of a process through the Kolmogorov extension theorem. This existence result does not by itself guarantee continuous sample paths; path regularity requires further conditions.

The kernel can also be interpreted geometrically. Centered random variables with finite second moments form a Hilbert space under an expectation-based inner product. Their covariance matrices are therefore Gram matrices. This viewpoint explains why sums of covariance kernels with nonnegative coefficients remain valid, why products of valid scalar kernels remain valid, and why a change of coordinates can pull a valid kernel back to a new index set. Such closure properties are widely used to construct models with several characteristic scales.

Regularity of a kernel controls certain mean-square properties. For a real process,

E ⁣[((X(t)−mX(t))−(X(s)−mX(s)))2]=CX(t,t)+CX(s,s)−2CX(s,t).\mathbb E\!\left[ \bigl((X(t)-m_X(t))-(X(s)-m_X(s))\bigr)^2 \right] = C_X(t,t)+C_X(s,s)-2C_X(s,t).

If this expression tends to zero as ss approaches tt, the centered process is mean-square continuous. Analogous difference quotients relate derivatives of the covariance to mean-square derivatives of the process. Mean-square continuity is weaker than almost-sure continuity of sample paths, so these notions cannot be used interchangeably.

Stationarity, temporal memory, and basic models

A process is weakly stationary, or second-order stationary, when its mean is constant and its covariance depends only on the difference between arguments. For a continuous-time process, this permits the notation

C(τ)=E[(X(t+τ)−m)(X(t)−m)‾].C(\tau)= \mathbb E[(X(t+\tau)-m)\overline{(X(t)-m)}].

The lag τ\tau, rather than the absolute time tt, becomes the relevant variable. A strictly stationary process has joint distributions invariant under time shifts; when second moments exist, strict stationarity implies weak stationarity. The reverse implication does not generally hold, although it does for Gaussian processes.

For a real stationary process, C(−τ)=C(τ)C(-\tau)=C(\tau). Positive values indicate that deviations from the mean tend to have the same sign at the specified separation, whereas negative values indicate a tendency toward opposite signs. An oscillating covariance can reveal recurrent dynamics, while monotone decay often reflects relaxation. Neither shape uniquely identifies a mechanism: different stochastic models can have identical second-order correlation functions.

Discrete white noise provides the simplest example. If εn\varepsilon_n has constant variance σ2\sigma^2 and is uncorrelated at distinct indices, then

Cε(k)={σ2,k=0,0,k≠0.C_\varepsilon(k)= \begin{cases} \sigma^2,&k=0,\\ 0,&k\neq0. \end{cases}

Independence is a stronger assumption than the zero-lag covariance condition alone. In continuous time, ideal white noise is represented through a delta covariance, C(τ)=qδ(τ)C(\tau)=q\delta(\tau), in distribution theory. It is not an ordinary pointwise process with finite variance.

A first-order autoregressive model,

Xn=ϕXn−1+εn,∣ϕ∣<1,X_n=\phi X_{n-1}+\varepsilon_n,\qquad |\phi|<1,

with zero-mean independent innovations of variance σε2\sigma_\varepsilon^2, has stationary covariance

C(k)=σε21−ϕ2ϕ∣k∣.C(k)= \frac{\sigma_\varepsilon^2}{1-\phi^2}\phi^{|k|}.

Positive ϕ\phi produces persistent correlations; negative ϕ\phi produces alternating signs. The continuous-time Ornstein–Uhlenbeck process, governed by dXt=−λXt dt+σ dWtdX_t=-\lambda X_t\,dt+\sigma\,dW_t with λ>0\lambda>0, has stationary covariance

C(τ)=σ22λe−λ∣τ∣,C(\tau)=\frac{\sigma^2}{2\lambda}e^{-\lambda|\tau|},

where WtW_t is a Wiener process. Its relaxation time is 1/λ1/\lambda.

Not every process with a simple covariance is stationary. For standard Brownian motion starting at zero, C(s,t)=min⁡(s,t)C(s,t)=\min(s,t) for nonnegative times. Its variance grows with time, although its increments are stationary. This distinction is important when interpreting empirical correlation functions: a changing mean, a trend, or a time-dependent variance can imitate persistence that is not described by a stationary covariance.

Long memory refers to correlations that decay sufficiently slowly to remain important across many lags. For example, a positive asymptotic behavior C(k)∼ck−αC(k)\sim c k^{-\alpha}, with 0<α<10<\alpha<1, produces a divergent sum of covariances. This changes the scaling of averages and can invalidate uncertainty formulas based on rapidly decaying dependence.

Spectral representation and Fourier duality

The connection between correlation and frequency content is expressed by the Wiener–Khinchin theorem. Under an integrability assumption, the covariance of a continuous-time stationary process and its power spectral density form a Fourier transform pair. One angular-frequency convention is

S(ω)=∫−∞∞C(τ)e−iωτ dτ,C(τ)=12π∫−∞∞S(ω)eiωτ dω.S(\omega)= \int_{-\infty}^{\infty} C(\tau)e^{-i\omega\tau}\,d\tau, \qquad C(\tau)= \frac{1}{2\pi} \int_{-\infty}^{\infty} S(\omega)e^{i\omega\tau}\,d\omega.

The spectral density is nonnegative. In particular,

C(0)=12π∫−∞∞S(ω) dω,C(0)=\frac{1}{2\pi}\int_{-\infty}^{\infty}S(\omega)\,d\omega,

so the total spectral weight equals the variance. Frequency conventions using cycles per unit time replace ω\omega by 2πf2\pi f and redistribute the factors of 2π2\pi.

A more general statement uses a spectral measure rather than an ordinary density. Bochner’s theorem states that a continuous positive-definite function on the real line is the Fourier transform of a finite positive measure:

C(τ)=∫Reiωτ μ(dω).C(\tau)=\int_{\mathbb R}e^{i\omega\tau}\,\mu(d\omega).

When μ(dω)=S(ω) dω/(2π)\mu(d\omega)=S(\omega)\,d\omega/(2\pi), the density formula follows. Discrete spectral components are represented by atoms of the measure. For example, a sinusoid with a uniformly random phase has a covariance proportional to cos⁡(ω0τ)\cos(\omega_0\tau), and its spectral measure is concentrated at ω0\omega_0 and −ω0-\omega_0. Thus a stationary covariance need not decay with lag or possess an integrable spectral density.

For a discrete-time stationary process, the spectrum is periodic in angular frequency. When the covariance sequence is absolutely summable,

S(ω)=∑k=−∞∞C(k)e−iωk,C(k)=12π∫−ππS(ω)eiωk dω.S(\omega)=\sum_{k=-\infty}^{\infty}C(k)e^{-i\omega k}, \qquad C(k)=\frac{1}{2\pi}\int_{-\pi}^{\pi} S(\omega)e^{i\omega k}\,d\omega.

This representation underlies frequency-domain analysis of time series. It also links rapid covariance decay to broad spectral content, while slowly decaying or oscillatory covariance can produce sharply concentrated spectral features.

Linear filtering makes the relationship operational. If a stationary input passes through a suitable linear time-invariant filter with impulse response hh, the output is a convolution, and its spectrum satisfies SY(ω)=∣H(ω)∣2SX(ω)S_Y(\omega)=|H(\omega)|^2S_X(\omega). The corresponding output covariance is obtained by filtering the input covariance in both arguments. An exponential covariance, for example, transforms into a Lorentzian spectrum proportional to (λ2+ω2)−1(\lambda^2+\omega^2)^{-1}.

The historical foundations were established through work on stochastic and harmonic analysis in the early twentieth century. Norbert Wiener developed generalized harmonic analysis in 1930, and Aleksandr Khinchin established a central stationary-correlation result in 1934. The spectral-measure formulation clarifies the distinction between broadband fluctuations, narrow spectral peaks, and exact spectral lines.

Higher-order and connected correlation functions

A two-point correlation function captures only part of a general probability law. Higher-order functions examine products at three or more arguments, such as

G(n)(t1,…,tn)=E[X(t1)⋯X(tn)]G^{(n)}(t_1,\ldots,t_n) = \mathbb E[X(t_1)\cdots X(t_n)]

for a real-valued process. Such quantities can detect asymmetry, non-Gaussian dependence, and interactions among several observations that cannot be inferred from pairwise covariance alone. For complex quantities or operator-valued fields, conjugation and ordering must be specified separately.

The distinction between moments and connected correlations is formalized by cumulants. For a finite random vector, derivatives of the logarithm of its moment-generating function produce joint cumulants when that function exists near the origin. The first cumulant is the mean, the second is the covariance, and the third measures a form of joint asymmetry. Cumulants can also be defined directly from moments through sums over partitions, avoiding the need to assume a moment-generating function when the relevant moments exist.

For four zero-mean variables, the fourth connected correlation is

⟨X1X2X3X4⟩c=⟨X1X2X3X4⟩−⟨X1X2⟩⟨X3X4⟩−⟨X1X3⟩⟨X2X4⟩−⟨X1X4⟩⟨X2X3⟩.\begin{aligned} \langle X_1X_2X_3X_4\rangle_c ={}&\langle X_1X_2X_3X_4\rangle\\ &-\langle X_1X_2\rangle\langle X_3X_4\rangle\\ &-\langle X_1X_3\rangle\langle X_2X_4\rangle\\ &-\langle X_1X_4\rangle\langle X_2X_3\rangle. \end{aligned}

The subtraction removes contributions that factor into products of two-point functions. For nonzero means, the general cumulant expression contains additional terms. Connected correlations vanish whenever their arguments divide into two nonempty statistically independent groups. Vanishing of one connected function, however, does not establish independence.

Gaussian processes occupy a special position because their mean and covariance determine all finite-dimensional distributions. For a centered jointly Gaussian family, every odd-order moment vanishes, while each even-order moment is a sum over pairings. At fourth order,

E[X1X2X3X4]=C12C34+C13C24+C14C23.\mathbb E[X_1X_2X_3X_4] = C_{12}C_{34}+C_{13}C_{24}+C_{14}C_{23}.

Consequently, all Gaussian cumulants above order two vanish. This identity is known in probability as Isserlis’ theorem, published in 1918, and is closely related to Wick’s theorem in field theory. Nonzero higher connected correlations therefore provide evidence against an exactly Gaussian description.

Pairwise information can nevertheless miss strong constraints. Let UU and VV be independent variables taking values +1+1 and −1-1 with equal probabilities, and define W=UVW=UV. Each pair among U,V,WU,V,W is independent, and all three pairwise covariances vanish. Yet UVW=1UVW=1 always, so the three variables are not jointly independent. Their third-order product exposes a dependence invisible to every two-point covariance.

Fourier transforms of higher-order stationary correlations yield polyspectra, including the bispectrum and trispectrum. These describe relations among several frequencies rather than merely the power at each frequency. They are useful in detecting non-Gaussian signals and phase relationships, although estimating them generally requires substantially more data than estimating a two-point function.

Estimation from finite observations

A theoretical correlation function is an ensemble expectation, whereas observations usually supply one finite realization. Replacing ensemble averages with averages along a record requires assumptions about stationarity and ergodicity. Stationarity alone is insufficient: a process that equals a single random constant at every time is stationary, but its temporal average does not reveal the ensemble distribution of that constant. A finite record therefore contains information about a correlation function only through an assumed relationship between repeated observations and the underlying ensemble.

For observations x1,…,xNx_1,\ldots,x_N with sample mean xˉ\bar x, a common estimator at lag k≥0k\geq0 is

C^N(k)=1N∑t=1N−k(xt+k−xˉ)(xt−xˉ).\widehat C_N(k)= \frac{1}{N} \sum_{t=1}^{N-k} (x_{t+k}-\bar x)(x_t-\bar x).

Another convention divides by N−kN-k. With a known mean, the latter removes the loss-of-pairs bias at each lag. With an estimated mean, it is not generally unbiased. The denominator NN has a different advantage: its full lag sequence can be viewed as the autocorrelation of a zero-padded finite sequence and has a nonnegative spectral transform. Choosing a denominator therefore involves more than a cosmetic adjustment.

Large lags use fewer overlapping pairs and are consequently difficult to estimate. Moreover, estimated correlations at neighboring lags are themselves statistically dependent. For fixed positive lags under an independent white-noise null model with suitable moment conditions, sample autocorrelations are asymptotically normal with variance approximately 1/N1/N. This motivates familiar approximate bands of width 1.96/N1.96/\sqrt N, but those bands are not universal confidence intervals. They require adjustment when model parameters have been estimated, dependence is present, or many lags are examined simultaneously.

The impact of serial correlation on averages follows from an exact identity. For a real stationary sequence with variance σ2\sigma^2 and normalized autocorrelation ρ(k)\rho(k),

Var⁡(XˉN)=σ2N[1+2∑k=1N−1(1−kN)ρ(k)].\operatorname{Var}(\bar X_N) = \frac{\sigma^2}{N} \left[ 1+2\sum_{k=1}^{N-1} \left(1-\frac{k}{N}\right)\rho(k) \right].

For sufficiently fast decay, the bracket approaches 1+2∑k≥1ρ(k)1+2\sum_{k\geq1}\rho(k). Its reciprocal times NN defines a commonly used effective sample size. Positive persistence usually reduces effective information; negative correlations can increase it relative to independent observations. This calculation is central to uncertainty assessment in Markov chain Monte Carlo.

Computationally, all lags can be obtained through the fast Fourier transform by transforming the data, forming the squared magnitude, and applying an inverse transform. Adequate zero-padding is necessary to avoid replacing ordinary correlation by circular correlation. Missing observations, irregular sampling, measurement noise, and nonstationarity require additional modeling. Independent additive measurement noise increases the zero-lag covariance without changing nonzero-lag covariance, potentially making normalized correlations appear weaker. Detrending and smoothing also alter the estimated function and can create or remove apparent long-range structure.

Spatial correlation and point patterns

For a spatially indexed random field X(x)X(\mathbf x), the covariance is initially a function of two positions. Homogeneity reduces it to a function of their displacement, C(h)C(\mathbf h); isotropy further reduces it to a function of distance, C(r)C(r), where r=∥h∥r=\|\mathbf h\|. These assumptions are distinct. A material may be homogeneous but directionally anisotropic, as in a layered medium whose fluctuations have different correlation ranges parallel and perpendicular to the layers.

In geostatistics, spatial dependence is frequently described by the variogram. For a real field with stationary increments, the semivariogram is

γ(h)=12E[(X(x+h)−X(x))2].\gamma(\mathbf h) = \frac12\mathbb E[ (X(\mathbf x+\mathbf h)-X(\mathbf x))^2].

For a constant-mean, second-order stationary field,

γ(h)=C(0)−C(h).\gamma(\mathbf h)=C(\mathbf 0)-C(\mathbf h).

The variogram can exist even when a stationary covariance does not. It also obeys different admissibility conditions: a valid variogram is conditionally negative semidefinite rather than a covariance kernel. Variograms enter kriging, where spatial observations are combined to form linear predictions with minimum prediction-error variance under a specified model.

Common stationary covariance models include the exponential, squared-exponential, and Matérn covariance functions. The squared-exponential kernel is exceptionally smooth and yields mean-square derivatives of every order. The Matérn family includes a parameter that controls smoothness, allowing rougher and often more realistic fields. A function that is positive definite for locations on a line need not be valid for every spatial dimension; the index space is part of the admissibility problem.

For point patterns, the relevant quantity is often a pair correlation function rather than a covariance of ordinary pointwise field values. In a homogeneous isotropic point process of number density nn, the pair correlation g(r)g(r) is defined so that, conditional on a typical point, the expected number of other points in a three-dimensional shell of radius rr and thickness drdr is

4πr2n g(r) dr.4\pi r^2n\,g(r)\,dr.

For a homogeneous Poisson point process, g(r)=1g(r)=1. Values below one indicate fewer pairs than the Poisson benchmark at that separation; values above one indicate an excess. Hard-core exclusion produces g(r)=0g(r)=0 below an exclusion distance.

The total correlation h(r)=g(r)−1h(r)=g(r)-1 connects real-space pair structure with the structure factor:

S(k)=1+n∫Rdh(r)e−ik⋅r ddr,S(\mathbf k) = 1+n\int_{\mathbb R^d} h(\mathbf r)e^{-i\mathbf k\cdot\mathbf r}\,d^d r,

under appropriate homogeneity and convergence assumptions. The constant term represents the self contribution. This relation links measured scattering intensity to spatial organization in fluids and disordered materials. Estimation requires corrections for finite observation windows, inaccessible separations, and inhomogeneous sampling; otherwise boundary effects can be mistaken for physical correlations.

Equilibrium fluctuations, critical behavior, and transport

In statistical mechanics, correlation functions quantify how microscopic fluctuations combine into macroscopic behavior. For observables AA and BB, the connected equilibrium correlation is

CAB(t)=⟨δA(t) δB(0)⟩,δA=A−⟨A⟩.C_{AB}(t) = \langle\delta A(t)\,\delta B(0)\rangle, \qquad \delta A=A-\langle A\rangle.

The brackets denote an equilibrium ensemble average. The subtraction distinguishes genuine fluctuation correlations from products of equilibrium means. Depending on the system, the observables may be particle densities, spin components, velocities, or currents.

A basic example is the Ising model, whose lattice spins sis_i take values +1+1 or −1-1. Its connected two-point function is

Gij=⟨sisj⟩−⟨si⟩⟨sj⟩.G_{ij}=\langle s_i s_j\rangle-\langle s_i\rangle\langle s_j\rangle.

Above a ferromagnetic transition, correlations usually decay at large separation with a characteristic correlation length ξ\xi. Near a continuous transition, a scaling form is

G(r)∼r−(d−2+η)F(r/ξ),ξ∼∣t∣−ν,G(r)\sim r^{-(d-2+\eta)}F(r/\xi), \qquad \xi\sim |t|^{-\nu},

where t=(T−Tc)/Tct=(T-T_c)/T_c, dd is dimension, and η\eta and ν\nu are critical exponents. At criticality, the divergence of ξ\xi permits scale-free correlations over increasing distances. For the two-dimensional square-lattice Ising model at zero field, ν=1\nu=1 and η=1/4\eta=1/4, so the critical spin correlation decays as r−1/4r^{-1/4}.

Correlation sums determine response coefficients. If a magnetic field enters the energy as −h∑isi-h\sum_i s_i, the susceptibility per spin obeys

χ=βN∑i,jGij,β=1kBT.\chi = \frac{\beta}{N} \sum_{i,j}G_{ij}, \qquad \beta=\frac{1}{k_{\mathrm B}T}.

Thus enhanced long-distance correlation can produce a large macroscopic response. This is an equilibrium instance of the fluctuation–dissipation theorem, which connects spontaneous fluctuations to linear response. The precise factors depend on whether the observable and applied field include magnetic moments or other dimensional constants.

Time correlations also determine transport. For an isotropic classical fluid in dd dimensions, the self-diffusion coefficient is

D=1d∫0∞⟨v(t)⋅v(0)⟩ dt,D=\frac1d\int_0^\infty \langle\mathbf v(t)\cdot\mathbf v(0)\rangle\,dt,

provided the integral converges and the long-time motion is diffusive. This is a Green–Kubo relation. Similar expressions connect shear-stress correlations to viscosity and heat-current correlations to thermal conductivity. The framework was developed through the work of Melville Green and Ryogo Kubo in the 1950s.

Decay need not be purely exponential. Conservation laws and hydrodynamic modes can generate algebraic long-time tails; critical systems can show growing relaxation times, known as critical slowing down. These features affect both physical transport coefficients and numerical simulations. A truncation of a correlation integral can miss a substantial tail contribution, while a finite simulation box limits the largest accessible wavelengths and correlation lengths.

Quantum correlation functions and operator ordering

In quantum mechanics, observables are operators, and their products generally depend on order. Given a density operator ρ^\hat\rho, a two-time correlation may be written

GAB(t,t′)=Tr⁡ ⁣[ρ^ A^(t)B^(t′)],G_{AB}(t,t') = \operatorname{Tr}\!\left[ \hat\rho\,\hat A(t)\hat B(t') \right],

with Heisenberg-picture operators

A^(t)=eiH^t/ℏA^e−iH^t/ℏ.\hat A(t)= e^{i\hat Ht/\hbar}\hat A e^{-i\hat Ht/\hbar}.

Because A^(t)B^(t′)\hat A(t)\hat B(t') need not equal B^(t′)A^(t)\hat B(t')\hat A(t), several inequivalent correlation functions arise. A quantum correlation can also be complex even when both observables are Hermitian.

The symmetrized connected correlation is

CABsym(t,t′)=12⟨δA^(t)δB^(t′)+δB^(t′)δA^(t)⟩.C_{AB}^{\mathrm{sym}}(t,t') = \frac12 \left\langle \delta\hat A(t)\delta\hat B(t') + \delta\hat B(t')\delta\hat A(t) \right\rangle.

By contrast, the commutator determines linear response. If a perturbation is introduced as H^′(t)=−f(t)B^\hat H'(t)=-f(t)\hat B, the response of ⟨A^(t)⟩\langle\hat A(t)\rangle is governed by

χAB(t−t′)=iℏθ(t−t′)⟨[A^(t),B^(t′)]⟩\chi_{AB}(t-t') = \frac{i}{\hbar}\theta(t-t') \langle[\hat A(t),\hat B(t')]\rangle

in equilibrium, with this choice of sign convention. The step function enforces causality: the response vanishes before the perturbation. Symmetrized fluctuations and dissipative response are related by a quantum fluctuation–dissipation relation containing temperature- and frequency-dependent factors.

In quantum field theory, nn-point functions are expectations of products of field operators at spacetime points. Time-ordered two-point functions place the later operator to the left and define the Feynman propagator, up to conventional factors. Such objects are often called Green’s functions, although not every correlation function is a Green’s function for a differential operator, and not every Green’s function is a probabilistic covariance. Relativistic fields are operator-valued distributions, so their correlators may be singular at coincident points and must be interpreted through smearing or regularization.

Quantum optics distinguishes correlation functions by how field operators are ordered. For a single mode with annihilation operator aa, the normalized equal-time intensity correlation is

g(2)(0)=⟨a†a†aa⟩⟨a†a⟩2,g^{(2)}(0)= \frac{\langle a^\dagger a^\dagger aa\rangle} {\langle a^\dagger a\rangle^2},

when the denominator is nonzero. A coherent state has g(2)(0)=1g^{(2)}(0)=1, an ideal single-mode thermal state has g(2)(0)=2g^{(2)}(0)=2, and a single-photon number state has g(2)(0)=0g^{(2)}(0)=0. These values distinguish Poissonian photon statistics, thermal bunching, and suppression of simultaneous detections.

The Hanbury Brown–Twiss effect, demonstrated in landmark optical experiments in 1956, established intensity correlations as a practical measurement tool. Modern correlation measurements appear in photon counting, neutron scattering, magnetic spectroscopy, and many-body quantum experiments. The measured object depends on detector coupling and operator ordering, so a reported “correlation function” must identify the observable, state, ordering convention, and normalization.