Correlation Length · aiwiki.page

Correlation length, usually denoted by $\xi$ , is a characteristic distance describing how rapidly spatial correlations diminish in a system. It is central to statistical mechanics, especially the study of fluctuations near a phase transition. A small correlation length indicates relatively localized fluctuations; a large one indicates coordinated variations across widely separated regions. It describes the spatial organization of fluctuations rather than necessarily the range of the underlying interactions. (damtp.cam.ac.uk)

Definition through correlation functions

For a local observable $A(\mathbf{x})$ , such as density or magnetization, the connected correlation function is

C(\mathbf{r})= \langle A(\mathbf{x})A(\mathbf{x}+\mathbf{r})\rangle -\langle A(\mathbf{x})\rangle \langle A(\mathbf{x}+\mathbf{r})\rangle .

Angle brackets denote an ensemble expectation value. This expression is the spatial covariance of the observable. Subtracting the mean values isolates fluctuations, which is particularly important when the system has a nonzero order parameter. In a homogeneous system, $C$ depends on separation rather than absolute position. (damtp.cam.ac.uk)

When correlations decay exponentially at large separation, their asymptotic form may be written

C(r)\sim B r^{-p}e^{-r/\xi}.

The exponential sets the characteristic length, while the algebraic prefactor modifies the detailed decay. For a pure exponential, increasing separation by $\xi$ reduces the correlation by a factor $e^{-1}$ . Correlations do not abruptly disappear at this distance: $\xi$ is a decay scale, not a sharp boundary. (damtp.cam.ac.uk)

Different definitions and directional dependence

The exponential correlation length can be defined, when the limit exists, by

\xi_{\mathrm{exp}}^{-1} =-\lim_{r\to\infty}\frac{1}{r} \ln\left|\frac{C(r)}{C_*}\right|,

where $C_*$ is a fixed nonzero normalization with the same units as $C$ . This definition extracts the long-distance decay rate. On a lattice, the result can depend on the direction of separation, so a single isotropic value may not capture all spatial behavior. (arxiv.org)

A frequently used alternative is the second-moment correlation length. For an isotropic system in $d$ spatial dimensions,

\xi_{\mathrm{2nd}}^2= \frac{\int d^d r\,r^2 C(\mathbf{r})} {2d\int d^d r\,C(\mathbf{r})},

provided the integrals converge and define a positive ratio. Lattice definitions replace integrals with sums. Unlike the exponential definition, this measure weights correlations across all separations. The two lengths can be close, but they are not generally identical; their distinction matters in interpreting numerical results and finite-size scaling. (arxiv.org)

Critical behavior

Near an ordinary continuous critical point, the bulk correlation length commonly diverges as

\xi\simeq \xi_0^\pm |t|^{-\nu}, \qquad t=\frac{T-T_c}{T_c}.

Here $T$ is temperature, $T_c$ is the critical temperature, and $\nu$ is a critical exponent. The amplitudes $\xi_0^\pm$ describe approach from above or below the transition and need not be equal. The divergence concerns the ideal thermodynamic limit, rather than an unrestricted length measurable in a finite sample. (damtp.cam.ac.uk)

At criticality, correlations often follow a power law,

C(r)\propto r^{-(d-2+\eta)},

rather than possessing a finite exponential cutoff. The exponent $\eta$ describes the anomalous spatial decay. The renormalization group explains this behavior through changes of observation scale and identifies universality classes: microscopically different systems can share critical exponents because their long-distance behavior is governed by the same critical structure. (damtp.cam.ac.uk)

The divergence is not a defining feature of every transition. Second-moment correlation lengths can remain finite at first-order transitions, as illustrated by calculations for the two-dimensional Potts model. (arxiv.org)

Example: the Ising chain

An exactly solvable example is the infinite, one-dimensional Ising model with nearest-neighbor ferromagnetic coupling $J>0$ , zero external field, and lattice spacing $a$ . At positive temperature,

\langle s_0s_n\rangle=[\tanh(\beta J)]^{|n|}, \qquad \xi=-\frac{a}{\ln[\tanh(\beta J)]},

where $s_n=\pm1$ and $\beta=1/(k_BT)$ , with $k_B$ the Boltzmann constant. Thus the correlation length is finite at every positive temperature but grows without bound as temperature approaches zero. This example distinguishes a growing correlation scale from a finite-temperature phase transition. (cpt.univ-mrs.fr)

Measurement and physical consequences

Correlation lengths can be inferred from spatial measurements or scattering experiments. The structure factor expresses correlations in reciprocal space and is related to their Fourier transform. For fluctuations described by a simple Ornstein–Zernike form,

S(q)\simeq\frac{S(0)}{1+q^2\xi^2},

where $q$ is the magnitude of the scattering wavevector. A larger $\xi$ produces a narrower low- $q$ feature. Experiments on polymer mixtures, for example, extract concentration-fluctuation correlation lengths from this dependence. The inferred value depends on the correlation model and the measured range of wavevectors. (damtp.cam.ac.uk)

Growing density fluctuations near fluid critical points can strongly scatter visible light, producing critical opalescence. Correlations also connect microscopic fluctuations to macroscopic response: the spatial integral of connected magnetic correlations determines magnetic susceptibility, with appropriate temperature and normalization factors. (damtp.cam.ac.uk)

In Ginzburg–Landau theory, the simplest Gaussian treatment gives $\xi\propto |T-T_c|^{-1/2}$ . This is a mean-field prediction, not a universal exponent: fluctuations beyond that approximation can change the critical scaling. (damtp.cam.ac.uk)