Correlation Length and Its Divergence

A many-body system in equilibrium is never truly still; it fluctuates, and those fluctuations are spatially correlated. The correlation length $\xi$ is the single number that says how far the influence of a local fluctuation reaches: two regions separated by much more than $\xi$ behave as statistically independent. Away from any phase transition $\xi$ is microscopic — comparable to the lattice spacing or interaction range — and the system looks featureless on macroscopic scales. The remarkable physics of critical phenomena is precisely that $\xi$ can be made arbitrarily large by tuning a control parameter (temperature, field, density) toward a critical point. Quantitatively, $\xi$ is read off from the connected correlation function $G_c(r)=\langle s(\mathbf 0)s(\mathbf r)\rangle-\langle s\rangle^2$ . In a disordered or generic ordered phase this decays exponentially, $G_c(r)\propto e^{-r/\xi}$ , defining $\xi$ as the inverse decay rate. As $T\to T_c$ the decay slows, $\xi\to\infty$ , and exactly at $T_c$ the correlation function loses its scale and decays only algebraically. The divergence is governed by the critical exponent $\nu$ through $\xi\sim|t|^{-\nu}$ , with $t=(T-T_c)/T_c$ the reduced distance to the critical temperature. The diverging correlation length is the conceptual cornerstone of the modern theory of phase transitions. Because $\xi$ dwarfs every microscopic length near $T_c$ , the long-wavelength physics cannot depend on microscopic detail — this is the origin of universality. Because $\xi$ is the only relevant length, observables become homogeneous functions of it, giving the scaling hypothesis and the scaling relations among exponents. And because fluctuations span every scale between the lattice spacing and $\xi$ , naive perturbation theory diverges and one must coarse-grain scale by scale — the renormalization group. Historically, $\xi$ entered through the Ornstein–Zernike theory of critical opalescence (1914), which captured the diverging range of density fluctuations responsible for the milky scattering of light near a liquid–gas critical point, while predicting the mean-field value $\nu=\tfrac12$ . The measured deviation of the exponents from these values — especially in three dimensions — is the experimental signature that fluctuations on the scale $\xi$ dominate the physics.

Key takeaways

The correlation length $\xi$ is the distance over which order-parameter fluctuations stay statistically coupled; beyond it $G_c(r)\sim e^{-r/\xi}$ decays exponentially.
Approaching a continuous transition, $\xi$ diverges as $\xi\sim|t|^{-\nu}$ with $t=(T-T_c)/T_c$ ; the universal exponent $\nu$ controls every other critical singularity.
At $T_c$ the exponential is replaced by a power law $G_c(r)\sim r^{-(d-2+\eta)}$ because the only length scale has become infinite — the system is scale invariant.
Diverging $\xi$ underlies the breakdown of mean-field theory below the upper critical dimension, the failure of Ornstein–Zernike, and the success of the renormalization group.
Scaling relations such as Fisher's $\gamma=\nu(2-\eta)$ and hyperscaling $2-\alpha=d\nu$ follow from treating $\xi$ as the single relevant length.
Key idea: near a continuous transition there is one emergent length $\xi$ diverging as $|t|^{-\nu}$ ; every critical singularity is a power law fixed by how the observable scales with $\xi$ .

The core insight

Near a continuous transition there is a single emergent length, the correlation length $\xi$ , which diverges as $|t|^{-\nu}$ ; every critical singularity — susceptibility, specific heat, opalescence, slowing down — is a power law fixed by how that quantity scales with $\xi$ .

Where this lives today

The correlation length is the quantity every precision test of universality actually measures. Neutron scattering on magnets fits the Ornstein–Zernike lineshape to extract $\xi(T)$ and hence $\nu$ ; light scattering on near-critical fluids tracks $\xi$ growing toward optical wavelengths; and large-scale Monte Carlo studies use finite-size scaling — the cutoff of $\xi$ at the box size $L$ — to pin $T_c$ and $\nu$ to many digits by demanding data at different $L$ collapse under the variable $L^{1/\nu}t$ . The same idea reaches far beyond equilibrium magnets: diverging correlation lengths signal continuous transitions in cold-atom gases, in the QCD phase diagram probed by heavy-ion collisions, and in dynamical phase transitions of driven systems.

The physical picture

Picture the Ising model as a sea of up and down spins. Far from $T_c$ a spin influences only its near neighbours: flip one and the perturbation dies within a few lattice spacings, so $\xi$ is microscopic. Heat toward $T_c$ and ordered patches grow; cool toward $T_c$ from below and disordered bubbles grow. At $T_c$ patches of every size — from the lattice spacing up to the system — coexist in a self-similar froth, and $\xi$ , the typical patch size, diverges. The diverging $\xi$ is also why critical points fluctuate violently and relax slowly. The susceptibility integrates the whole correlated volume, $\chi\sim\int\mathrm{d}^d r\,G_c(r)\sim\xi^{2-\eta}$ , so it diverges. Light scattering off density fluctuations correlated over a length $\xi$ produces critical opalescence. Relaxation times grow as a power of $\xi$ — critical slowing down — because reorganizing a correlated region of size $\xi$ takes ever longer. Scale invariance is the deep meaning of $\xi\to\infty$ . The length $\xi$ appears in $G_c(r)\sim e^{-r/\xi}$ ; setting $\xi=\infty$ removes it, leaving a function with no scale, which must be a pure power law $r^{-(d-2+\eta)}$ . The system then looks the same under magnification — the fixed point of the renormalization group. Because $\xi$ dwarfs the lattice spacing, the precise interaction shape and chemical identity wash out; only $d$ , the number of order-parameter components $n$ , and the range and symmetry of interactions survive. That is universality, a direct corollary of a single diverging length.

Figure

Correlation length diverges as T approaches T_c — patch size grows without bound

Figure

Divergence of ξ versus t on a log-log plot — slope minus nu differs between MF and Ising

The core equations

The correlation length is defined operationally through the decay of the connected correlation function, which subtracts the square of the mean so that it measures genuine fluctuations only.

G_c(\mathbf r)=\langle s(\mathbf 0)\,s(\mathbf r)\rangle-\langle s\rangle^2 \qquad (1)

where: $s(\mathbf r)$ — local order-parameter field; $\langle\cdot\rangle$ — thermal average; $G_c$ — connected two-point function. Away from criticality the decay is exponential, with a power-law prefactor that comes from inverting the Ornstein–Zernike form in $d$ dimensions.

G_c(r)\;\xrightarrow{\,r\to\infty\,}\;A\,\frac{e^{-r/\xi}}{r^{(d-1)/2}} \qquad (2)

where: $\xi$ — correlation length; $A$ — non-universal amplitude; $d$ — spatial dimension. For $d=3$ this is the familiar $e^{-r/\xi}/r$ . The temperature dependence of $\xi$ near a continuous transition is a power law with the universal exponent $\nu$ .

\xi(t)\sim|t|^{-\nu},\qquad t\equiv\frac{T-T_c}{T_c} \qquad (3)

where: $\nu$ — correlation-length exponent; $t$ — reduced temperature. Mean field gives $\nu=\tfrac12$ , 2D Ising $\nu=1$ , 3D Ising $\nu\approx0.630$ . Exactly at $T_c$ the exponential is replaced by a power law, whose deviation from the Gaussian $d-2$ defines the anomalous dimension.

G_c(r)\big|_{T=T_c}\sim\frac{1}{r^{\,d-2+\eta}} \qquad (4)

where: $\eta$ — anomalous dimension; $d-2$ — mean-field power. Integrating $G_c$ over space gives the susceptibility, tying its divergence directly to that of $\xi$ .

\chi=\beta\!\int\mathrm{d}^d r\,G_c(r)=\beta\,\tilde G(0)\sim\xi^{2-\eta} \qquad (5)

where: $\chi$ — order-parameter susceptibility; $\tilde G(0)$ — zero-momentum correlation. Combined with $\xi\sim|t|^{-\nu}$ this yields Fisher's relation.

\gamma=\nu(2-\eta) \qquad (6)

where: $\gamma$ — susceptibility exponent, $\chi\sim|t|^{-\gamma}$ . Because the singular free energy carries one factor of $k_BT_c$ per correlation volume $\xi^d$ , the specific-heat exponent is fixed by $\nu$ and the dimension — the hyperscaling relation, the only one that contains $d$ explicitly.

2-\alpha=d\,\nu\qquad\text{(Josephson hyperscaling)} \qquad (7)

where: $\alpha$ — specific-heat exponent. It holds only below the upper critical dimension $d_c=4$ .

Derivation AOrnstein–Zernike correlation length from Gaussian Landau theory

The goal is to derive the correlation length, the mean-field exponents $\nu=\tfrac12,\eta=0$ , and the screened-Coulomb correlation function from a Gaussian truncation of the Landau–Ginzburg free energy, then to expose exactly where the approximation breaks down. This is the calculation that both establishes the mean-field baseline and, through the Ginzburg criterion at the end, predicts its own failure below four dimensions.

Step 1. Write the Landau–Ginzburg functional $\beta\mathcal H[\phi]=\int\mathrm{d}^d r\big[\tfrac{K}{2}(\nabla\phi)^2+\tfrac{a}{2}\phi^2+\tfrac{u}{4}\phi^4-h\phi\big]$ with $a=a_0t$ changing sign at $T_c$ . Why this form: it is the most general local, analytic, symmetric free energy; the gradient term penalizes rapid variation and sets the stiffness.

Step 2. For $T>T_c$ fluctuate about $\phi=0$ and keep only quadratic terms, so in Fourier space $\beta\mathcal H_0=\tfrac12\int\frac{\mathrm{d}^d k}{(2\pi)^d}(Kk^2+a)|\tilde\phi(k)|^2$ . Why drop $u\phi^4$ : near $\phi=0$ the quartic term is subleading — its validity is exactly what the Ginzburg criterion will test.

Step 3. Equipartition gives $\langle|\tilde\phi(k)|^2\rangle=\dfrac{1}{Kk^2+a}=\dfrac{1}{K}\dfrac{1}{k^2+\xi^{-2}}$ , so reading off $\xi^{-2}=a/K=a_0t/K$ gives $\xi\propto t^{-1/2}$ , i.e. $\nu=\tfrac12$ . Why $\nu=\tfrac12$ : the mass term is linear in $t$ , and $\xi$ is its inverse square root.

Step 4. Inverse-transform in $d=3$ to get $G_c(r)=\dfrac{1}{4\pi K r}e^{-r/\xi}$ ; at $T_c$ , $\xi=\infty$ leaves $G_c\sim1/r=r^{-(d-2)}$ , so $\eta=0$ . Why $\eta=0$ : the Gaussian propagator inverts to the pure Coulomb power with no anomalous correction.

Step 5. Test self-consistency: the fluctuation inside one correlation volume must be smaller than the mean, $\langle\delta\phi^2\rangle_{\xi^d}\ll\langle\phi\rangle^2$ , which reduces to $|t|^{(d-4)/2}\ll\text{const}$ . Why this bounds the theory: the inequality holds as $t\to0$ only for $d>4$ ; below $d_c=4$ fluctuations dominate and the mean-field exponents fail. *Verified: the MF values $\nu=\tfrac12,\eta=0$ , the 3D OZ form, and the $d_c=4$ Ginzburg boundary are checked in verify_aq-sm-correlation-length-divergence.py; they reproduce Pathria & Beale §13.*

Go deeper · Derivation B · 1 minScaling relations from a single diverging length

Derivation BScaling relations from a single diverging length

The goal is to show that postulating $\xi$ as the only relevant length forces the singular free energy to be homogeneous and recovers Fisher, Josephson (hyperscaling), and Rushbrooke without any microscopic calculation. The argument is the seed of the whole scaling hypothesis: one physical input — one diverging length carrying one unit of singular free energy per correlation volume — pins the relations among six exponents.

Step 1. Assume $f_{\rm sing}(t,h)=\xi^{-d}\,\Phi(h\,\xi^{y_h})=|t|^{d\nu}\,\Phi(h\,|t|^{-\nu y_h})$ . Why $\xi^{-d}$ : one factor of $k_BT_c$ of singular energy lives in each correlation volume $\xi^d$ , so the singular free-energy density scales as its inverse.

Step 2. Differentiate twice in $t$ : $c\sim\partial_t^2 f_{\rm sing}\sim|t|^{d\nu-2}\equiv|t|^{-\alpha}$ , giving $2-\alpha=d\nu$ . Why this is "hyper": $d$ appears explicitly, so the relation is dimension-dependent and holds only for $d\le d_c=4$ .

Step 3. Use the susceptibility sum rule $\chi\sim\xi^{2-\eta}\sim|t|^{-\nu(2-\eta)}\equiv|t|^{-\gamma}$ , yielding Fisher's $\gamma=\nu(2-\eta)$ . Why the sum rule enters: $\chi$ is the integrated correlator, and integrating the critical scaling form produces $\xi^{2-\eta}$ .

Step 4. Combine the homogeneity in $t$ and $h$ with $\langle\phi\rangle\sim|t|^\beta$ to get Rushbrooke $\alpha+2\beta+\gamma=2$ ; Widom's $\gamma=\beta(\delta-1)$ follows the same way. Why only two are free: all six exponents come from the two scaling powers $y_t=1/\nu$ and $y_h$ , so four algebraic relations must hold. *Verified: hyperscaling $2-\alpha=d\nu$ , Fisher, and Rushbrooke are checked against MF and both the 2D ( $\nu=1,\eta=\tfrac14,\gamma=\tfrac74$ ) and 3D Ising values in verify_aq-sm-correlation-length-divergence.py.*

Go deeper · Derivation C · 1 minCorrelation length and critical slowing down

Derivation CCorrelation length and critical slowing down

The goal is to link the relaxation time to the correlation length through the dynamic exponent $z$ , explaining why timescales diverge near $T_c$ (critical slowing down). Static criticality tells us how large the correlated region is; dynamics asks how long it takes to reorganize such a region, and the answer is a power of $\xi$ .

Step 1. Adopt relaxational (model A) dynamics $\partial_t\tilde\phi(k)=-\Gamma(Kk^2+\xi^{-2})\tilde\phi(k)+\text{noise}$ . Why this equation: it is the simplest overdamped relaxation of a non-conserved order parameter toward equilibrium, with kinetic coefficient $\Gamma$ .

Step 2. Each mode relaxes at rate $\tau(k)^{-1}=\Gamma(Kk^2+\xi^{-2})$ , so the slowest ( $k\to0$ ) mode has $\tau=\xi^2/(\Gamma K)$ . Why the longest wavelength is slowest: its restoring "stiffness" $\xi^{-2}$ is the smallest, so it drifts back most sluggishly.

Step 3. Hence $\tau\sim\xi^2\sim|t|^{-\nu z}$ with the van Hove value $z=2$ ; including fluctuations renormalizes $z$ slightly upward (e.g. $z\approx2.04$ for 3D model A) but the divergence persists. Why it matters: this is critical slowing down — the reason simulations and experiments near $T_c$ equilibrate painfully slowly. *Verified: the $\tau\sim\xi^z$ form and van Hove $z=2$ are recorded in verify_aq-sm-correlation-length-divergence.py; consistent with Hohenberg–Halperin RMP 1977.*

Limiting cases

How $\xi$ behaves depends on where one sits relative to $T_c$ and on dimensionality.

Case	Result	Meaning
Disordered, $T>T_c$	$\xi\simeq\xi_0^{+}t^{-\nu}$	Exponential decay $G_c\sim e^{-r/\xi}$ ; $\xi_0^{+}$ microscopic, $\nu$ universal.
Ordered, $T<T_c$	$\xi\simeq\xi_0^{-}\|t\|^{-\nu}$	Same exponent both sides; the amplitude ratio $\xi_0^{+}/\xi_0^{-}$ is universal though the amplitudes are not.
At $T_c$	$G_c(r)\sim r^{-(d-2+\eta)}$ , $\xi=\infty$	Scale invariance; $\eta$ measures deviation from the mean-field power.
$d>4$ (above $d_c$ )	$\nu=\tfrac12,\eta=0$	Ornstein–Zernike exact; fluctuations irrelevant (Ginzburg).
2D Ising (exact)	$\nu=1,\eta=\tfrac14$	Onsager class; strongly non-mean-field.
3D Ising (best)	$\nu\approx0.630,\eta\approx0.036$	Conformal bootstrap / high-order RG; realized in fluids and uniaxial magnets.

Worked examples

Example 1 — How much faster ξ grows in 3D Ising than in mean field. For the 3D Ising class $\nu=0.630$ ; in mean field $\nu=\tfrac12$ . Compare the growth of $\xi$ when $t$ drops from $0.02$ to $0.002$ (a tenfold approach toward $T_c$ ).

Step 1. Ising: $\dfrac{\xi(0.002)}{\xi(0.02)}=\left(\dfrac{0.02}{0.002}\right)^{0.630}=10^{0.630}=\mathbf{4.27}$ .

Step 2. Mean field: the same step gives $10^{0.5}=3.16$ . The correlation length in the 3D Ising system grows by a factor 4.27, about 35 % more than the mean-field 3.16 — the excess is the fluctuation enhancement that makes real critical points more singular than Landau theory predicts.

Example 2 — When ξ reaches the wavelength of light. Take a simple fluid with microscopic scale $\xi_0=0.3$ nm and 3D Ising $\nu=0.63$ . How large is $\xi$ at $t=10^{-4}$ ?

Step 1. $\xi=\xi_0\,|t|^{-\nu}=0.3\times(10^{-4})^{-0.63}=0.3\times10^{2.52}$ nm.

Step 2. $10^{2.52}=331$ , so $\xi=0.3\times331\approx\mathbf{99\ nm}$ — already a fifth of the wavelength of visible light. A mere $0.01\%$ approach toward $T_c$ swells the correlation length from a molecular $0.3$ nm to about $100$ nm, and a factor ten closer would push $\xi$ past $400$ nm, into the optical band where the fluid turns milky — critical opalescence.

Origins · 1 minmichael fisher and how this was discovered

michael fisher

1931–2021 · Michael Ellis Fisher was, with Leo Kadanoff and Kenneth Wilson, one of the architects of the modern theory of critical phenomena. He sharpened the definition of the correlation length, clarified the anomalous dimension, and derived the scaling relation γ = ν(2 − η) that now bears his name — the bridge between the correlation function and the susceptibility. Fisher's work on series expansions, scaling, and the ε-expansion (with Wilson) helped turn critical exponents from empirical numbers into consequences of the renormalization group. He shared the 1980 Wolf Prize and the 1995 Lars Onsager Prize for this body of work.

History

1914

— Ornstein and Zernike introduce the correlation length to explain critical opalescence

1944

— Onsager solves the 2D Ising model exactly, giving ν = 1

1964

— Fisher relates the correlation-length and susceptibility exponents, γ = ν(2 − η)

1966

— Kadanoff's block-spin picture makes the diverging ξ the basis of scaling

1971

— Wilson's renormalization group derives ν as the inverse relevant eigenvalue

1977

— Hohenberg and Halperin systematize dynamic critical exponents and critical slowing down

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 12–14 · Kardar, Statistical Physics of Fields, Ch. 2–4 · Goldenfeld, Lectures on Phase Transitions and the Renormalization Group, Ch. 5–7 · Sethna, Statistical Mechanics, Ch. 12 · Stanley, Introduction to Phase Transitions and Critical Phenomena · Landau & Lifshitz, Statistical Physics Part 1, §116, §146.

Provenance. All exponent values are standard: mean-field $\nu=\tfrac12,\eta=0$ (Pathria & Beale §13); 2D Ising $\nu=1,\eta=\tfrac14$ (Onsager); 3D Ising $\nu\approx0.630,\eta\approx0.036$ (conformal bootstrap, Kos et al.). Scaling relations $\gamma=\nu(2-\eta)$ , $2-\alpha=d\nu$ , $\alpha+2\beta+\gamma=2$ are derived in Stanley and Kardar. A verifier should recompute the OZ Fourier inversion, the sum rule, the scaling algebra, and the van Hove dynamic exponent $z=2$ (Hohenberg–Halperin, RMP 49, 435 (1977)).

Builds on: Correlation Function · Continuous Phase Transitions · Order Parameter · Landau Theory of Phase Transitions. Leads to: Critical Exponents · Scaling Hypothesis and Scaling Relations · Universality and Universality Classes · Goldstone Modes.

What exactly is the correlation length — the decay length, or something else?

It is the decay length of the connected correlation function: for large separations $G_c(r)\sim e^{-r/\xi}$ , so $\xi$ is the distance over which correlations fall by a factor $e$ . More precisely, one can define $\xi$ as the second moment of $G_c$ (the "second-moment correlation length"), which is what scattering experiments extract from the width of the Ornstein–Zernike Lorentzian. All reasonable definitions agree on the exponent $\nu$ governing the divergence, even though they differ in the non-universal amplitude $\xi_0$ .

Why does a single diverging length produce power laws in every observable?

Once $\xi$ is the only relevant length, dimensional analysis forces every singular quantity to be a power of $\xi$ : the susceptibility is an inverse length squared times fluctuation ( $\xi^{2-\eta}$ ), the singular free-energy density is an inverse volume ( $\xi^{-d}$ ), and so on. Since $\xi\sim|t|^{-\nu}$ , each of these becomes a power of $|t|$ — a critical exponent. This is why the exponents are not independent: they are all built from how the same $\xi$ enters each observable, which is the content of the scaling relations.

How does mean-field theory get ν wrong, and why only below four dimensions?

Mean-field (Gaussian) theory computes $\xi$ while ignoring interactions between fluctuations, giving $\nu=\tfrac12$ . The Ginzburg criterion asks whether the neglected fluctuations inside one correlation volume are actually small: the ratio scales as $|t|^{(d-4)/2}$ . For $d>4$ this vanishes as $T\to T_c$ , so mean field is self-consistent and exact. For $d<4$ it grows, so fluctuations dominate near $T_c$ and shift $\nu$ away from $\tfrac12$ (to $0.630$ in 3D Ising, to $1$ in 2D). The dimension $d_c=4$ where the crossover happens is the upper critical dimension.

What is critical slowing down and how does it relate to ξ?

Near $T_c$ the system takes longer and longer to relax to equilibrium: the characteristic time diverges as $\tau\sim\xi^z$ , where $z$ is the dynamic critical exponent (van Hove $z=2$ in mean field). The physical reason is that reorganizing a correlated region of size $\xi$ requires coordinated motion over that whole region, and larger regions take proportionally longer. This is why experiments and Monte Carlo simulations near a critical point are notoriously slow to equilibrate, and why measuring $z$ requires time-resolved probes like dynamic light scattering.

Can the correlation length ever actually reach infinity in a real sample?

No — a real sample has a finite size $L$ , so $\xi$ can grow only until it saturates near $L$ , at which point finite-size effects round off the would-be singularity. The true thermodynamic divergence $\xi\to\infty$ requires the thermodynamic limit $L\to\infty$ . Far from being a limitation, this saturation is exploited: finite-size scaling theory predicts exactly how observables depend on $L^{1/\nu}t$ , so measuring the same quantity in several box sizes and collapsing the data yields $T_c$ and $\nu$ with high precision.

Why is the amplitude ratio universal but the amplitudes are not?

The prefactors $\xi_0^{+}$ (above $T_c$ ) and $\xi_0^{-}$ (below) depend on microscopic details — lattice spacing, coupling strength — so they differ from material to material and are not universal. But their ratio $\xi_0^{+}/\xi_0^{-}$ is a pure number fixed by the universality class, because the microscopic scale cancels. The same is true of amplitude ratios for the specific heat and susceptibility. Universal amplitude ratios are a stronger test of universality than the exponents alone, since two systems can share exponents yet be distinguished by their amplitude ratios if they belong to different classes.