Correlation Functions and Correlation Length in Lattice Models

Thermodynamic averages tell you the mean magnetization or energy, but they hide the most interesting question about a phase transition: how does a fluctuation at one point of a lattice reach out and influence a distant point? That is what a correlation function measures. For a spin model the connected two-point function $G(\mathbf{r})=\langle s_0 s_{\mathbf{r}}\rangle_c$ reports the typical size and shape of the ordered clusters that form spontaneously as the system is cooled. And because thermodynamic response functions are integrals of correlation functions, the microscopic and macroscopic descriptions are rigidly locked together: a diverging susceptibility is nothing but the thermodynamic shadow of a diverging correlation length. The qualitative story is universal. Far above $T_c$ spins are nearly independent and $G$ decays over a microscopic distance — the correlation length $\xi$ is a lattice spacing or two. As $T_c$ is approached from either side, $\xi$ grows without bound, $\xi\sim|t|^{-\nu}$ , where $t=(T-T_c)/T_c$ . At criticality the only length scale has run off to infinity, scale invariance sets in, and the exponential decay is replaced by a pure power law characterized by the anomalous dimension $\eta$ . Below $T_c$ the full correlation saturates at $\langle s\rangle^2$ , and only the connected part decays. Two exactly solvable settings anchor the picture. The one-dimensional Ising chain, solved by the transfer matrix, gives $G(r)=(\tanh K)^r=e^{-r/\xi}$ with $\xi^{-1}=\ln\coth(\beta J)$ , finite for every $T>0$ — the analytic statement that the 1D chain never orders at finite temperature. The two-dimensional Ising model shows the genuine critical divergence with $\nu=1$ and $\eta=1/4$ , the canonical test of scaling. Throughout, the correlation length is the organizing idea: the length that controls the crossover between microscopic and macroscopic behavior, and whose divergence defines a second-order phase transition.

Key takeaways

The two-point correlation function $G(\mathbf{r})=\langle s_0 s_{\mathbf{r}}\rangle-\langle s_0\rangle\langle s_{\mathbf{r}}\rangle$ measures how strongly the spin at the origin is correlated with the spin a distance $\mathbf{r}$ away.
Away from criticality $G(\mathbf{r})$ decays exponentially, $G\sim e^{-r/\xi}$ , defining the correlation length $\xi$ — the size of correlated domains.
The fluctuation–response theorem ties susceptibility to integrated correlation, $\chi=\beta\sum_{\mathbf{r}}G(\mathbf{r})$ , so $\chi$ diverges precisely when $\xi$ does.
At a continuous transition $\xi\to\infty$ as $\xi\sim|t|^{-\nu}$ , and exactly at $T_c$ the exponential is replaced by a power law $G(r)\sim r^{-(d-2+\eta)}$ .
In 1D the transfer matrix gives $\xi^{-1}=\ln(\lambda_1/\lambda_2)=\ln\coth(\beta J)$ , finite at all $T>0$ — no order in 1D.
The structure factor $S(\mathbf{k})$ , the Fourier transform of $G$ , is what neutron and light scattering measure; its small- $k$ Lorentzian peak sharpens as $\xi$ grows (critical opalescence).
Key idea: the correlation length $\xi$ is the single emergent length of a lattice model — correlations decay as $e^{-r/\xi}$ off criticality, $\xi$ diverges as $|t|^{-\nu}$ at a continuous transition, and $\chi=\beta\sum_{\mathbf{r}}G(\mathbf{r})$ diverges with it because response is the spatial integral of correlation.

The core insight

The correlation length $\xi$ is the single emergent length scale of a lattice model: correlations decay as $e^{-r/\xi}$ away from criticality, $\xi$ diverges as $|t|^{-\nu}$ at a continuous transition, and the susceptibility $\chi=\beta\sum_{\mathbf{r}}G(\mathbf{r})$ diverges with it because response is the spatial integral of correlation.

Where this lives today

The correlation length is the workhorse of computational statistical mechanics. Because $\xi$ diverges at $T_c$ , Monte Carlo studies extract it from the measured $G(r)$ or structure factor $S(\mathbf{k})$ and locate criticality by its divergence — in practice through finite-size scaling, where $\xi$ is cut off by the system size $L$ , and the crossings of scaled observables pin $T_c$ and $\nu$ with remarkable precision. High-temperature series for $\chi=\beta\sum_r G(r)$ and for the second moment $\sum_r r^2 G(r)$ yield $\xi^2$ and locate $T_c$ from their radius of convergence. Scattering experiments read the structure factor directly: neutron scattering off magnets and light scattering off binary fluids both measure a small- $k$ Lorentzian of width $1/\xi$ , and watching that peak sharpen as $T\to T_c$ is a live measurement of a diverging correlation length — the modern descendant of Ornstein and Zernike's 1914 explanation of critical opalescence. In the renormalization group the correlation length is the deepest concept of all: under a coarse-graining step that rescales lengths by $b$ , $\xi\to\xi/b$ , so the fixed point $\xi=\infty$ is exactly $T_c$ , the eigenvalues of the linearized RG give $\nu$ , and the fixed-point value of $\eta$ fixes the critical decay — making $\xi$ the bridge from microscopic lattice models to the universal exponents that classify all continuous transitions.

The physical picture

Picture cooling a magnet. Far above $T_c$ each spin flips almost independently; a snapshot looks like television static, and knowing the spin at the origin tells you nothing about a spin a few sites away. As $T\to T_c$ , patches of aligned spins grow — and $\xi$ is the typical patch size. At $T_c$ patches of all sizes coexist; the system is scale-invariant, like a fractal, and $\xi$ has diverged to infinity. The correlation length is also a response range. Poke the system at the origin with a tiny field and the magnetization responds out to a distance $\xi$ ; integrating that response over all sites gives the bulk susceptibility. That is precisely why $\chi$ and $\xi$ diverge together. Scattering experiments see this directly: the small-angle peak in the structure factor $S(\mathbf{k})$ is a Lorentzian of width $1/\xi$ , so as $\xi\to\infty$ the peak becomes a sharp spike — the phenomenon of critical opalescence, where a transparent fluid near its critical point turns milky.

Figure

Correlated patches grow as T falls toward Tc — ξ is the typical patch size, diverging at criticality

The punchline is the exactly solvable 1D Ising chain, whose correlation $G(r)=(\tanh K)^r$ is a perfect exponential $e^{-r/\xi}$ on a semilog plot — a straight line whose slope is $-1/\xi$ . Lower the temperature (raise $K=\beta J$ ) and the line flattens: $\xi=-1/\ln(\tanh K)$ grows, diverging only as $T\to0$ . The three straight lines of different slope, one per temperature, show the correlation length being born as a measurable number — and show, by never becoming horizontal at any finite $T$ , exactly why the 1D chain has no phase transition.

Figure

1D Ising G(r) = (tanh K)^r on a semilog plot: exponential decay, slope −1/ξ, ξ grows as T falls

The core equations

The connected two-point function subtracts the disconnected piece, so it decays to zero even in the ordered phase where the full correlation saturates at $m^2$ . This subtraction is essential below $T_c$ .

G(\mathbf{r})=\langle s_{\mathbf{0}}\,s_{\mathbf{r}}\rangle-\langle s_{\mathbf{0}}\rangle\langle s_{\mathbf{r}}\rangle \qquad (1)

where: $s_{\mathbf{r}}$ — spin at site $\mathbf{r}$ (e.g. $\pm1$ for Ising); $\langle\cdot\rangle$ — thermal average at temperature $T$ ; $G$ — the connected correlation, the cumulant that decays to zero at large $r$ . Physically it isolates genuine correlation from the trivial "both spins point along the magnetization." The correlation length is defined by the exponential decay rate of $G$ at large distance — the single length scale controlling how far correlations reach.

\xi^{-1}=-\lim_{r\to\infty}\frac{1}{r}\ln G(\mathbf{r}) \qquad (2)

where: $\xi$ — correlation length; $r=|\mathbf{r}|$ along a chosen direction. When $\xi$ is a few lattice spacings the system is "boring"; when it diverges, criticality has arrived. The fluctuation–response sum rule ties the susceptibility to the integral of the correlation — the exact link between microscopic correlation and macroscopic response.

\chi=\beta\sum_{\mathbf{r}}G(\mathbf{r})=\frac{\beta}{N}\big(\langle M^2\rangle-\langle M\rangle^2\big) \qquad (3)

where: $\chi$ — isothermal susceptibility per spin; $\beta=1/k_BT$ ; $M=\sum_i s_i$ — total magnetization; $N$ — number of sites. Read physically: $\chi$ is nothing but the summed correlation, which is why it diverges exactly when $\xi$ (and hence the sum) diverges. The structure factor is the Fourier transform of $G$ — the quantity that scattering experiments measure directly.

S(\mathbf{k})=\sum_{\mathbf{r}}e^{-i\mathbf{k}\cdot\mathbf{r}}\,G(\mathbf{r}) \qquad (4)

where: $S(\mathbf{k})$ — static structure factor; $\mathbf{k}$ — momentum transfer. Its small- $k$ form is a Lorentzian, and the forward value $S(0)=\chi/\beta$ is the susceptibility. For a short-range $G$ the small- $k$ structure factor is a Lorentzian whose half-width is $1/\xi$ — the Ornstein–Zernike form seen in scattering.

S(\mathbf{k})\simeq\frac{S(0)}{1+k^2\xi^2},\qquad \chi=\beta\,S(0) \qquad (5)

where: $S(0)\propto\chi$ — forward-scattering intensity; $k^2\xi^2$ — dimensionless combination setting the peak width. As $\xi\to\infty$ the Lorentzian narrows to a spike: critical opalescence. At a continuous transition the correlation length diverges as a power law, and exactly at $T_c$ the exponential decay becomes a pure power law with anomalous dimension $\eta$ .

G(r)\sim r^{-(d-2+\eta)}\,e^{-r/\xi},\qquad \xi\sim|t|^{-\nu} \qquad (6)

where: $\eta$ — anomalous (Fisher) dimension; $\nu$ — correlation-length exponent; $t=(T-T_c)/T_c$ . At $T_c$ the length $\xi=\infty$ removes the exponential, leaving pure $r^{-(d-2+\eta)}$ scale invariance.

Correlation length and response

Derivation ACorrelation function and ξ from the transfer matrix

We want to compute the exact spin–spin correlation of the 1D Ising chain and read off its correlation length, showing that $\xi$ is finite for every $T>0$ — the precise statement that the chain has no finite-temperature transition. The strategy: write the correlation as a trace of spin operators sandwiched between transfer matrices, diagonalize the transfer matrix, and take the thermodynamic limit so that the ratio of its two largest eigenvalues controls the decay. Keep an eye on that ratio $\lambda_2/\lambda_1<1$ — its logarithm is the inverse correlation length.

Step 1. Write the $2\times2$ transfer matrix and the partition function as its trace,

T=\begin{pmatrix} e^{\beta(J+h)} & e^{-\beta J}\\ e^{-\beta J} & e^{\beta(J-h)}\end{pmatrix},\qquad Z=\mathrm{Tr}\,T^N.

Why: the 1D chain factorizes bond by bond, so the partition function is a matrix product whose trace closes the periodic chain.

Step 2. Express the correlation as a trace with spin operators $\sigma_z=\mathrm{diag}(+1,-1)$ inserted at sites $0$ and $r$ ,

\langle s_0 s_r\rangle=\frac{\mathrm{Tr}\,(\sigma_z T^{r}\sigma_z T^{N-r})}{\mathrm{Tr}\,T^{N}}.

Why: inserting the diagonal spin operator picks out the spin value at that site; the transfer matrices propagate the chain between insertions.

Step 3. Diagonalize $T=\sum_a\lambda_a|a\rangle\langle a|$ and expand spectrally,

\langle s_0 s_r\rangle=\frac{\sum_{a,b}\lambda_a^{\,r}\lambda_b^{\,N-r}|\langle a|\sigma_z|b\rangle|^2}{\sum_a \lambda_a^{\,N}}.

Why: in the eigenbasis every power of $T$ becomes a power of an eigenvalue, turning the trace into a sum of geometric terms.

Step 4. Take $N\to\infty$ , where $\lambda_1$ (the largest eigenvalue) dominates,

\langle s_0 s_r\rangle\to\langle1|\sigma_z|1\rangle^2+\Big(\tfrac{\lambda_2}{\lambda_1}\Big)^{r}|\langle1|\sigma_z|2\rangle|^2.

Why: the first term is the disconnected piece $m^2$ ; the second, decaying because $\lambda_2<\lambda_1$ , is the connected correlation.

Step 5. Identify the correlation length,

G(r)=\Big(\tfrac{\lambda_2}{\lambda_1}\Big)^{r}|\langle1|\sigma_z|2\rangle|^2=e^{-r/\xi}|\langle1|\sigma_z|2\rangle|^2,\qquad \xi^{-1}=\ln\frac{\lambda_1}{\lambda_2}.

At $h=0$ , $\lambda_{1,2}=e^{\beta J}\pm e^{-\beta J}$ , so $\xi^{-1}=\ln\coth(\beta J)$ , finite for every $T>0$ and diverging only as $T\to0$ . Why: a finite $\xi$ at all $T>0$ is the mathematical statement that correlations are always short-ranged in 1D — no order, no transition. *Verified: verify_aq-sm-correlation-functions-lattice-models.py computes $\lambda_{1,2}=e^{K}\pm e^{-K}$ , $\xi^{-1}=\ln\coth K$ , confirms $G(r)=(\tanh K)^r$ , and checks $\xi$ finite for all $T>0$ with $\xi\to\infty$ only as $K\to\infty$ .*

Go deeper · Derivation B · 1 minSusceptibility as the integral of correlation, and Fisher's relation

Derivation BSusceptibility as the integral of correlation, and Fisher's relation

We want to prove that the susceptibility is the spatial sum of the correlation function, and then use the critical form of $G$ to derive Fisher's scaling relation $\gamma=\nu(2-\eta)$ . The strategy: differentiate the magnetization with respect to a uniform field (which brings down a factor of the total spin, generating the correlation sum), then integrate the critical power-law correlation cut off at $\xi$ to extract the temperature dependence of $\chi$ . The subtlety is that the integral is dominated by the largest scale, $r\sim\xi$ , so the divergence of $\xi$ directly controls the divergence of $\chi$ .

Step 1. Add a uniform field conjugate to the order parameter, $\mathcal{H}=\mathcal{H}_0-h\sum_i s_i$ , and write the per-site magnetization $m=\langle s_j\rangle$ . Why: susceptibility is the response of $m$ to $h$ , so we need the field explicitly in the Boltzmann weight.

Step 2. Differentiate; each $\partial_h$ brings down $\beta\sum_i s_i$ ,

\chi=\frac{\partial m}{\partial h}=\beta\Big(\langle s_j\textstyle\sum_i s_i\rangle-\langle s_j\rangle\langle\sum_i s_i\rangle\Big).

Why: differentiating $e^{\beta h\sum s}$ produces a factor $\beta\sum_i s_i$ , generating exactly the connected correlation between site $j$ and every other site.

Step 3. Use translational invariance to write this as the correlation sum,

\chi=\beta\sum_i G(\mathbf{r}_i-\mathbf{r}_j)=\beta\sum_{\mathbf{r}}G(\mathbf{r}).

Why: the correlation depends only on separation, so the sum over $i$ is independent of the reference site $j$ — response is the integrated correlation.

Step 4. At criticality insert the power-law form $G(r)\sim r^{-(d-2+\eta)}$ and cut the integral off at $\xi$ ,

\chi\sim\int^{\xi}\mathrm{d}^d r\;r^{-(d-2+\eta)}\sim\xi^{2-\eta}.

Why: the $d$ -dimensional volume element $r^{d-1}\mathrm{d}r$ times $r^{-(d-2+\eta)}$ gives $r^{1-\eta}\mathrm{d}r$ , whose integral up to $\xi$ is $\xi^{2-\eta}$ (for $\eta<2$ ).

Step 5. Substitute $\xi\sim|t|^{-\nu}$ and compare with $\chi\sim|t|^{-\gamma}$ ,

\chi\sim|t|^{-\nu(2-\eta)}\ \Longrightarrow\ \gamma=\nu(2-\eta).

Why: this is Fisher's scaling relation, tying the susceptibility exponent to the correlation-length and anomalous-dimension exponents — a rigid consequence of the fluctuation–response link. *Verified: verify_aq-sm-correlation-functions-lattice-models.py derives $\chi=\beta\sum_r G(r)$ , checks the integral $\int^\xi r^{1-\eta}\mathrm{d}r\sim\xi^{2-\eta}$ , and verifies Fisher's relation $\gamma=\nu(2-\eta)$ for the exact 2D Ising values ( $\nu=1,\eta=1/4,\gamma=7/4$ ) and the mean-field values ( $\nu=1/2,\eta=0,\gamma=1$ ).*

Limiting cases

The decay form of $G(\mathbf{r})$ depends sharply on temperature and dimension.

Case	Result	Meaning
High $T$ ( $T\gg T_c$ )	$G(r)\sim e^{-r/\xi}$ , $\xi\approx\dfrac{a}{\ln(1/\tanh\beta J)}\to0$	Short-range, $\xi$ of order a lattice spacing; spins nearly independent.
Critical ( $T=T_c$ , $d>2$ )	$G(r)\sim r^{-(d-2+\eta)}$	Power-law decay; the single length $\xi$ has diverged, leaving scale invariance and $\eta$ .
Mean-field / Ornstein–Zernike ( $d>4$ )	$G(r)\sim\dfrac{e^{-r/\xi}}{r^{(d-1)/2}}$ , $\xi\sim\|t\|^{-1/2}$	Classical exponents $\nu=1/2$ , $\eta=0$ ; exact above the upper critical dimension $d_u=4$ .
Ordered phase ( $T<T_c$ )	$\langle s_0 s_{\mathbf{r}}\rangle\to m^2$ , $G_c(r)\sim e^{-r/\xi}$	Full function saturates at $m^2$ ; only the connected part decays.

Worked examples

Example 1 — Correlation length of the 1D Ising chain at a given temperature.

Step 1. At $h=0$ the transfer-matrix eigenvalues are $\lambda_{1,2}=e^{K}\pm e^{-K}$ with $K=\beta J$ , so $\lambda_2/\lambda_1=\tanh K$ and $\xi^{-1}=\ln\coth K=-\ln\tanh K$ .

Step 2. Take $K=\beta J=1$ (temperature $k_BT=J$ ). Then $\tanh(1)=0.7616$ .

Step 3. Correlation length: $\xi=-1/\ln(0.7616)=-1/(-0.2724)=3.67$ lattice spacings.

Step 4. Sanity: at $r=5$ , $G(5)=(\tanh 1)^5=0.7616^5=0.256$ ; at $r=10$ , $G(10)=0.7616^{10}=0.0658$ . Correlations reach a few sites, exactly as $\xi\approx3.7$ predicts, and never diverge at this finite temperature. The 1D Ising correlation length at $k_BT=J$ is $\xi=-1/\ln(\tanh 1)=3.67$ lattice spacings; correlations decay as $G(r)=0.7616^r$ .

Example 2 — Checking Fisher's relation with the exact 2D Ising exponents.

Step 1. Fisher's scaling relation is $\gamma=\nu(2-\eta)$ , derived from $\chi\sim\xi^{2-\eta}$ and $\xi\sim|t|^{-\nu}$ .

Step 2. Insert the exact 2D Ising values (Onsager): $\nu=1$ , $\eta=1/4$ .

Step 3. Compute: $\gamma=1\cdot(2-1/4)=1\cdot7/4=7/4=1.75$ .

Step 4. This is exactly the known 2D Ising susceptibility exponent $\gamma=7/4$ — the relation is satisfied. For comparison, the mean-field values $\nu=1/2$ , $\eta=0$ give $\gamma=\tfrac12\cdot2=1$ , the classical result, also consistent. Fisher's relation gives $\gamma=\nu(2-\eta)=1\cdot(2-1/4)=7/4$ for 2D Ising, matching Onsager, and $\gamma=1$ in mean field.

Origins · 1 mincorrelation length and how this was discovered

correlation length

the emergent length scale of a critical system · The correlation length ξ is the characteristic distance over which fluctuations of the order parameter remain correlated: G(r) ∼ e^(−r/ξ) away from criticality. It grows without bound as a continuous transition is approached, ξ ∼

History

1914

— Ornstein and Zernike introduce the correlation function and its Lorentzian structure factor to explain critical opalescence

1936

— Ising-chain correlations analyzed via the transfer matrix (Kramers and Wannier method)

1944

— Onsager's solution yields the exact 2D Ising correlation length exponent ν = 1

1959

— Fisher formulates the scaling relation γ = ν(2 − η) linking response to correlation

1966

— Kadanoff's block-spin picture explains why ξ organizes the approach to criticality

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 12–13 · Kardar, Statistical Physics of Fields, Ch. 2–4 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 10 · Landau & Lifshitz, Statistical Physics Part 1, §116, §146 · Reif, Fundamentals of Statistical and Thermal Physics, Ch. 7.

Provenance. All equations are standard. verify_aq-sm-correlation-functions-lattice-models.py reproduces the transfer-matrix correlation $G(r)=(\lambda_2/\lambda_1)^r$ and $\xi^{-1}=\ln(\lambda_1/\lambda_2)$ , with $\lambda_{1,2}=e^{\beta J}\pm e^{-\beta J}$ at $h=0$ giving $\xi^{-1}=\ln\coth(\beta J)$ (Pathria §13, Kardar Ch. 6); derives $\chi=\beta\sum_r G(r)$ from $\partial m/\partial h$ (Reif Ch. 7); confirms the Ornstein–Zernike Lorentzian $S(k)=(k_BT/c)/(k^2+\xi^{-2})$ from a Gaussian LG free energy with $\nu=1/2,\eta=0$ (Landau–Lifshitz §146); and checks $\gamma=\nu(2-\eta)$ with the exact 2D Ising values $\nu=1,\eta=1/4,\gamma=7/4$ (Onsager; Kardar). The general critical decay $G(r)\sim r^{-(d-2+\eta)}$ and $\xi\sim|t|^{-\nu}$ are standard; the 2D Ising $\nu=1$ result is recalled from the Onsager solution by name.

Builds on: The Ising Model: Definition and Hamiltonian · Transfer Matrix Formalism for Lattice Models · One-Dimensional Ising Model and Transfer Matrix Method · Order Parameter and Spontaneous Magnetization. Leads to: Mean-Field (Weiss) Theory of the Ising Model · High- and Low-Temperature Series Expansions · Kramers-Wannier Duality · Domain Walls and Interfacial Excitations.

What is the difference between the correlation function and the correlation length?

The correlation function $G(r)$ is a function of distance — it tells you, for every separation $r$ , how strongly two spins are correlated. The correlation length $\xi$ is a single number extracted from it: the characteristic scale over which $G$ decays, $G(r)\sim e^{-r/\xi}$ . Think of $G(r)$ as the full decay curve and $\xi$ as the distance at which the correlation has dropped by a factor $e$ . Away from criticality $\xi$ is a few lattice spacings; at a continuous transition $\xi\to\infty$ and the exponential decay is replaced by a slow power law, so a single $\xi$ no longer captures the whole curve.

Why does the 1D Ising chain never order, no matter how cold?

Because its correlation length, $\xi=-1/\ln(\tanh\beta J)$ , is finite for every finite temperature. Correlations always decay exponentially over a finite distance, so no matter how cold you go (short of $T=0$ ), a spin has no idea what a sufficiently distant spin is doing — there is no long-range order. Only as $T\to0$ does $\xi$ diverge (as $\xi\sim\tfrac12 e^{2\beta J}$ ), so order appears exactly at, and only at, absolute zero. This is the analytic face of the Peierls/domain-wall argument: in 1D a single kink has finite energy but grows in entropy as $\ln L$ , so kinks always proliferate at $T>0$ .

How can susceptibility — a bulk thermodynamic quantity — equal a sum over microscopic correlations?

It follows from a two-line calculation. The susceptibility is the response of the magnetization to a field, $\chi=\partial m/\partial h$ . Differentiating the Boltzmann weight $e^{\beta h\sum_i s_i}$ brings down a factor $\beta\sum_i s_i$ , so $\chi=\beta\sum_i(\langle s_0 s_i\rangle-\langle s_0\rangle\langle s_i\rangle)=\beta\sum_r G(r)$ . In words: poking one spin with a field creates a response that spreads out to a distance $\xi$ , and the total response — the susceptibility — is the sum of all those correlations. This is why $\chi$ diverges exactly when the correlation sum (i.e. $\xi$ ) does.

What does the anomalous dimension η actually mean physically?

At a critical point correlations decay not exponentially but as a pure power law, $G(r)\sim r^{-(d-2+\eta)}$ . The " $d-2$ " part is what you would get from a simple Gaussian (Ornstein–Zernike) theory — the Coulomb-like $r^{-(d-2)}$ of a massless field. The extra $\eta$ is the anomalous correction from fluctuations beyond Gaussian: it measures how much the true critical decay differs from the naïve field-theory expectation. In mean field $\eta=0$ ; in the exact 2D Ising model $\eta=1/4$ , a genuinely non-classical value that signals strong critical fluctuations. It appears in Fisher's relation $\gamma=\nu(2-\eta)$ , tying it to the susceptibility.

Why does a fluid turn milky near its critical point?

This is critical opalescence, and correlation functions explain it exactly. The intensity of light scattered at wavevector $k$ is proportional to the structure factor $S(k)$ , the Fourier transform of the density correlation. Near the critical point the correlation length $\xi$ diverges, so $S(k)\simeq S(0)/(1+k^2\xi^2)$ develops a huge, sharp forward peak — the fluid scatters light strongly at long wavelengths comparable to $\xi$ . When $\xi$ reaches the wavelength of visible light, the sample scatters so much that it turns cloudy. Historically this was the first direct evidence that correlations diverge at a critical point, which is why Ornstein and Zernike built the whole theory around it in 1914.

How is the correlation length connected to the renormalization group?

The correlation length is the RG's central character. A renormalization-group step coarse-grains the system, rescaling all lengths by a factor $b$ , so the correlation length transforms simply: $\xi\to\xi/b$ . A fixed point of the RG must therefore have either $\xi=0$ (trivial, high- or low- $T$ ) or $\xi=\infty$ — and $\xi=\infty$ is precisely the critical point $T_c$ . Linearizing the RG flow near that fixed point gives the exponent $\nu$ (how fast $\xi$ diverges), and the fixed-point value of the anomalous dimension gives $\eta$ . So the divergence of $\xi$ is not just a symptom of criticality — it is the organizing principle that lets the RG explain why utterly different microscopic models share the same critical exponents (universality).