9.6 Correlation functions

Definition of correlation functions

Correlation functions quantify how the properties of one part of a system are statistically related to properties of another part. They're central to statistical mechanics because they connect microscopic particle arrangements to macroscopic observables like pressure, susceptibility, and scattering intensities. Nearly every measurable quantity in a many-body system can be expressed in terms of some correlation function.

Pair correlation function

The pair correlation function $g(r)$ measures how likely you are to find a particle at distance $r$ from a reference particle, compared to what you'd expect if particles were distributed completely at random.

• In a uniform system with number density $\rho$, the quantity $\rho \, g(r) \, 4\pi r^2 \, dr$ gives the average number of particles in a shell of thickness $dr$ at distance $r$ from a given particle.
• $g(r) = 1$ means the local density equals the bulk density (no correlation).
• $g(r) > 1$ means particles are more likely to be found at that distance than random chance would predict; $g(r) < 1$ means they're less likely.
• At large distances in a homogeneous fluid, $g(r) \to 1$, meaning correlations die out and particles "don't know about each other."

Higher-order correlation functions

Pair correlations capture a lot, but they can't tell you everything. Higher-order correlation functions describe relationships among three or more particles simultaneously.

• A three-point correlation function captures information about preferred bond angles (e.g., whether triplets of particles tend to form triangles of a particular shape).
• Four-point and higher functions encode increasingly detailed structural and dynamical information.
• These become very expensive to compute and harder to interpret, so in practice most analyses rely on pair correlations unless there's a specific reason to go higher.

Spatial vs. temporal correlations

• Spatial correlations describe how a quantity measured at position $\mathbf{r}_1$ relates to the same quantity at position $\mathbf{r}_2$ at the same time.
• Temporal correlations (autocorrelation functions) track how a quantity at one time relates to itself at a later time, capturing memory effects and relaxation.
• Space-time correlations combine both. The Van Hove function $G(\mathbf{r}, t)$ is a classic example: it describes the probability of finding a particle at position $\mathbf{r}$ at time $t$ given that a particle was at the origin at $t = 0$. This is essential for understanding diffusion, relaxation, and other dynamic processes.

Mathematical formulation

Two-point correlation function

The connected two-point correlation function (also called the covariance) for a density field $\rho(\mathbf{r})$ is:

$G(\mathbf{r}_1, \mathbf{r}_2) = \langle \rho(\mathbf{r}_1)\,\rho(\mathbf{r}_2) \rangle - \langle \rho(\mathbf{r}_1) \rangle \langle \rho(\mathbf{r}_2) \rangle$

The subtraction removes the trivial "background" so that $G$ isolates genuine statistical correlations. If the two points are uncorrelated, $G = 0$.

For a homogeneous, isotropic system, translational and rotational symmetry mean $G$ depends only on the scalar distance $r = |\mathbf{r}_1 - \mathbf{r}_2|$. The pair correlation function $g(r)$ is the normalized version of the full two-point density-density correlation.

N-point correlation functions

The N-point connected correlation function generalizes this:

$G_N(\mathbf{r}_1, \ldots, \mathbf{r}_N) = \langle \rho(\mathbf{r}_1) \cdots \rho(\mathbf{r}_N) \rangle_c$

where the subscript $c$ means "connected," indicating that all lower-order (disconnected) contributions have been subtracted out. This subtraction ensures $G_N$ captures only the genuinely new N-body correlations not already described by lower-order functions. Computing these for large $N$ is extremely demanding, both analytically and numerically.

Correlation length

The correlation length $\xi$ characterizes how far correlations extend in space. In many systems away from criticality, the connected correlation function decays exponentially:

$G(r) \sim e^{-r/\xi}$

• $\xi$ sets the length scale: at distances much larger than $\xi$, different parts of the system are effectively independent.
• Near a critical point, $\xi$ diverges. This is the physical origin of critical opalescence and other dramatic phenomena: fluctuations become correlated across the entire system.
• The divergence of $\xi$ is what makes mean-field theories break down near criticality, since they assume correlations are short-ranged.

Physical interpretation

Microscopic structure

Correlation functions are your window into how particles actually arrange themselves.

• Peaks in $g(r)$ correspond to preferred interparticle distances. The first peak typically sits near the particle diameter, reflecting nearest neighbors.
• Oscillations in $g(r)$ (common in liquids) indicate a shell-like structure: a first coordination shell, a second shell, and so on, with the oscillations damping out at larger $r$.
• A featureless $g(r) \approx 1$ everywhere signals a completely disordered, gas-like arrangement.

Fluctuations and correlations

Correlations and fluctuations are two sides of the same coin.

• The fluctuation-dissipation theorem connects equilibrium correlation functions to response functions (susceptibilities). For example, the magnetic susceptibility $\chi$ is proportional to the integral of the spin-spin correlation function.
• Larger spatial correlations generally mean larger thermodynamic fluctuations. Near a critical point, both the correlation length and the magnitude of fluctuations diverge together.

Order parameters

Correlation functions provide a rigorous way to distinguish ordered from disordered phases.

• Long-range order means the correlation function approaches a nonzero constant as $r \to \infty$. In a ferromagnet below $T_c$, the spin-spin correlation $\langle S(\mathbf{0}) \cdot S(\mathbf{r}) \rangle \to m^2$ where $m$ is the spontaneous magnetization.
• Short-range order means correlations decay to zero (exponentially or otherwise) at large distances.
• Quasi-long-range order (as in the 2D XY model below the Kosterlitz-Thouless transition) shows up as power-law decay of correlations, which is slower than exponential but still goes to zero.

Applications in statistical mechanics

Phase transitions

Correlation functions change character across phase boundaries, making them natural diagnostic tools.

• In the disordered (high-temperature) phase, correlations decay exponentially with a finite $\xi$.
• At the critical point, correlations decay as a power law (no characteristic length scale).
• In the ordered (low-temperature) phase, correlations persist to infinite distance.
• Critical exponents describe the singular behavior of correlation functions near the transition. Systems with the same spatial dimension and order-parameter symmetry share the same exponents, which is the basis of universality classes.

Critical phenomena

At a critical point, the correlation function takes the power-law form:

$G(r) \sim \frac{1}{r^{d-2+\eta}}$

where $d$ is the spatial dimension and $\eta$ is the anomalous dimension (a critical exponent that measures deviations from mean-field behavior).

• For the 3D Ising model, $\eta \approx 0.036$, which is small but nonzero.
• Scaling relations (like Fisher's relation $\gamma = (2 - \eta)\nu$) connect $\eta$ to other critical exponents such as $\nu$ (correlation length exponent) and $\gamma$ (susceptibility exponent).
• The renormalization group provides the theoretical framework for understanding why power-law correlations emerge at criticality and for computing critical exponents systematically.

Structural analysis

• The static structure factor $S(\mathbf{q})$ is related to the Fourier transform of $g(r)$:

$S(\mathbf{q}) = 1 + \rho \int [g(\mathbf{r}) - 1] \, e^{i\mathbf{q}\cdot\mathbf{r}} \, d\mathbf{r}$

• This is directly measurable in scattering experiments: the scattered intensity from X-rays or neutrons is proportional to $S(\mathbf{q})$.
• Peaks in $S(\mathbf{q})$ correspond to characteristic length scales in the system (e.g., lattice spacings in crystals, nearest-neighbor distances in liquids).
• Higher-order correlations provide information about bond angles and local molecular geometries that $g(r)$ alone cannot capture.

Experimental techniques

Scattering experiments

Scattering is the most direct experimental probe of correlation functions.

• X-ray scattering couples to electron density, so it measures electron density-density correlations. Bragg peaks in crystals correspond to long-range positional order.
• Neutron scattering couples to nuclear positions and magnetic moments, making it sensitive to both structural and magnetic correlations.
• Small-angle scattering (SAXS/SANS) probes correlations on larger length scales (nanometers to micrometers), useful for polymers, colloids, and biological macromolecules.
• Inelastic scattering measures the dynamic structure factor $S(\mathbf{q}, \omega)$, which is the space-time Fourier transform of the Van Hove function, giving access to dynamic correlations.

Microscopy methods

• Electron microscopy (TEM/SEM) can image atomic arrangements directly, from which real-space correlation functions can be computed.
• Scanning tunneling microscopy (STM) probes local electronic density of states, giving access to electronic correlations at surfaces.
• Atomic force microscopy (AFM) measures surface topography and force interactions at the nanoscale.
• Super-resolution optical microscopy extends correlation analysis to biological systems, where fluorescent labels track molecular positions.

Spectroscopic measurements

• NMR is sensitive to local magnetic environments and can probe spin-spin correlations and molecular dynamics.
• Raman and infrared spectroscopy measure vibrational correlations in molecules and solids.
• Dynamic light scattering (photon correlation spectroscopy) measures the time autocorrelation of scattered light intensity, yielding information about particle sizes and diffusion coefficients in colloidal suspensions.

Computational methods

Monte Carlo simulations

Monte Carlo methods generate representative configurations sampled from the equilibrium distribution, then compute correlation functions by averaging over those configurations.

1. Start from an initial configuration of particles (or spins, etc.).
2. Propose a random move (e.g., displace a particle, flip a spin).
3. Accept or reject the move using the Metropolis criterion: accept with probability $\min(1, e^{-\beta \Delta E})$ where $\Delta E$ is the energy change.
4. After equilibration, measure $g(r)$ or other correlation functions by averaging over many saved configurations.

Cluster algorithms (Wolff, Swendsen-Wang) are critical near phase transitions, where single-spin-flip methods suffer from critical slowing down because the correlation length is large.

Molecular dynamics

Molecular dynamics (MD) integrates Newton's equations of motion numerically, producing particle trajectories from which both static and dynamic correlation functions can be extracted.

1. Initialize positions and velocities for all particles.
2. Compute forces from the interparticle potential.
3. Update positions and velocities using an integrator (e.g., velocity Verlet).
4. Repeat, saving configurations at regular intervals.
5. Compute time-dependent correlations like the velocity autocorrelation function $C_v(t) = \langle \mathbf{v}(0) \cdot \mathbf{v}(t) \rangle$ directly from the trajectory.

Thermostats (Nosé-Hoover, Langevin) and barostats control temperature and pressure to simulate the desired ensemble (NVT, NPT, etc.).

Density functional theory

Density functional theory (DFT) is primarily an electronic structure method, but it connects to correlation functions through the electron density.

• The Kohn-Sham framework maps the interacting electron problem onto a set of non-interacting equations with an effective potential.
• The exchange-correlation functional encodes all many-body correlation effects. Its exact form is unknown, so approximations (LDA, GGA, hybrid functionals) are used.
• Time-dependent DFT (TD-DFT) extends the framework to compute dynamic response functions and excitation spectra, giving access to frequency-dependent correlations.

Correlation functions in specific systems

Ideal gases

The ideal gas is the simplest reference case.

• $g(r) = 1$ for all $r$ because there are no interactions, so there's no preferred spacing between particles.
• All higher-order correlation functions factorize into products of lower-order ones (the system is trivially uncorrelated).
• Velocity autocorrelations in an ideal gas are a delta function in time: $\langle \mathbf{v}(0) \cdot \mathbf{v}(t) \rangle = \frac{3k_BT}{m}\,\delta(t)$, reflecting the absence of any memory.

Liquids and dense fluids

Liquids sit between the extremes of ideal gases and crystals.

• $g(r)$ shows a strong first peak (nearest-neighbor shell), followed by damped oscillations that decay to 1 within a few molecular diameters. This is short-range order without long-range periodicity.
• The Van Hove function $G(r, t)$ separates into a "self" part (tracking a tagged particle's displacement over time, related to diffusion) and a "distinct" part (tracking how surrounding particles rearrange).
• Dynamic correlations reveal collective modes like sound waves at long wavelengths and diffusive relaxation at shorter scales.

Solids and crystals

• $g(r)$ in a crystal shows sharp peaks at lattice positions, reflecting long-range translational order. The peaks persist to arbitrarily large $r$.
• The structure and symmetry of the crystal (FCC, BCC, HCP, etc.) are encoded in the positions and relative heights of these peaks.
• Phonon correlations describe collective lattice vibrations. The displacement-displacement correlation function is directly related to the phonon spectrum.
• Defects (vacancies, dislocations) and thermal disorder broaden the peaks and introduce deviations from perfect crystalline correlations.

Connection to other concepts

Thermodynamic properties

Correlation functions provide a bridge from microscopic interactions to bulk thermodynamics.

• The pressure equation relates the equation of state to $g(r)$ and the pair potential $u(r)$:

$P = \rho k_B T - \frac{\rho^2}{6} \int r \, u'(r) \, g(r) \, 4\pi r^2 \, dr$

• The compressibility equation connects isothermal compressibility to the integral of $[g(r) - 1]$.
• Internal energy, entropy, and free energy can all be expressed as integrals involving correlation functions, though higher-order correlations are needed for exact results beyond the pair level.

Transport coefficients

The Green-Kubo relations express transport coefficients as time integrals of equilibrium correlation functions:

• Diffusion coefficient: $D = \frac{1}{3}\int_0^\infty \langle \mathbf{v}(0) \cdot \mathbf{v}(t) \rangle \, dt$
• Shear viscosity: $\eta = \frac{V}{k_BT}\int_0^\infty \langle \sigma_{xy}(0)\,\sigma_{xy}(t) \rangle \, dt$ where $\sigma_{xy}$ is the off-diagonal stress tensor component.
• Thermal conductivity: related to the energy current autocorrelation function in the same way.

These relations are powerful because they let you compute nonequilibrium transport properties from equilibrium simulations.

Susceptibilities

Linear response theory (Kubo formalism) directly links generalized susceptibilities to correlation functions.

• The magnetic susceptibility $\chi$ is proportional to the spatial integral of the spin-spin correlation function: $\chi \propto \int \langle S(\mathbf{0})\,S(\mathbf{r}) \rangle_c \, d\mathbf{r}$.
• The dielectric susceptibility is linked to dipole-dipole correlations.
• This is why susceptibilities diverge at critical points: the correlation length diverges, so the integral of the correlation function over all space diverges too.

Advanced topics

Dynamic correlation functions

Time-dependent correlation functions $C(t) = \langle A(0)\,A(t) \rangle$ describe how the system loses memory of its initial state.

• The spectral density (or power spectrum) is the Fourier transform of $C(t)$ and reveals characteristic frequencies of the system.
• Exponential decay of $C(t)$ corresponds to a Lorentzian spectral density, typical of simple relaxation processes.
• The fluctuation-dissipation theorem in its frequency-dependent form relates the imaginary part of the response function $\chi''(\omega)$ to the spectral density, connecting equilibrium fluctuations to energy absorption.

Quantum correlation functions

In quantum systems, observables become operators, and their ordering matters.

• The imaginary-time (Matsubara) Green's function $G(\tau) = -\langle T_\tau \hat{A}(\tau)\hat{B}(0) \rangle$ is defined with imaginary time $\tau = it$ and is periodic (bosons) or antiperiodic (fermions) with period $\beta = 1/k_BT$. This periodicity leads to discrete Matsubara frequencies in the Fourier representation.
• Retarded and advanced Green's functions describe causal response and are analytically continued from Matsubara functions.
• The path integral formulation provides an alternative route to quantum correlations, expressing them as functional integrals over field configurations.

Non-equilibrium correlations

When a system is driven out of equilibrium, the standard equilibrium framework must be extended.

• Fluctuation theorems (Jarzynski equality, Crooks relation) constrain the statistics of non-equilibrium work and entropy production, generalizing the fluctuation-dissipation theorem.
• In aging systems (e.g., glasses after a quench), two-time correlation functions $C(t, t_w)$ depend on both the observation time $t$ and the waiting time $t_w$ since the quench, breaking time-translation invariance.
• The Keldysh (closed-time-path) formalism provides a systematic diagrammatic framework for computing correlation functions in quantum systems out of equilibrium, where the Matsubara approach no longer applies.