6.5 Mean field theory

Mean field theory tackles one of the hardest problems in statistical mechanics: how do you analyze a system where every particle interacts with every other particle? The answer is to replace all those complicated individual interactions with a single average (or "mean") field that each particle feels. This reduces an intractable many-body problem to a solvable single-body problem, and it turns out to capture the essential physics of phase transitions surprisingly well.

That said, mean field theory has a well-known blind spot: it ignores fluctuations and correlations between particles. Near critical points, where these fluctuations diverge, the theory gives qualitatively right but quantitatively wrong predictions. Understanding both the power and the failure modes of mean field theory is central to studying critical phenomena.

Foundations of mean field theory

The core idea is straightforward: instead of tracking how each particle interacts with its specific neighbors, you assume every particle sees the same effective field produced by the average behavior of all the others. This converts a many-body problem into a single-body problem in an effective field.

Basic principles

• Each particle "feels" a mean field created by the average state of all other particles, rather than interacting with each neighbor individually.
• Fluctuations and correlations between individual particles are neglected. The theory treats deviations from the average as negligible.
• The mean field must be determined self-consistently: the average you assume must match the average you calculate from the resulting single-particle problem.

Historical development

• Pierre Weiss (1907) introduced the molecular field theory of ferromagnetism, the first true mean field theory. He proposed that each spin feels an effective field proportional to the average magnetization.
• Lev Landau (1937) generalized the approach by expanding the free energy in powers of an order parameter, creating a unified framework for continuous phase transitions.
• Pierre-Gilles de Gennes (1960s) extended mean field ideas to liquid crystals and polymers.
• Modern applications span superconductivity, quantum many-body physics, complex networks, and beyond.

Applications in physics

Mean field theory appears across many subfields:

• Ferromagnetism: Describes the paramagnetic-to-ferromagnetic transition in materials like iron and nickel.
• Superconductivity: BCS theory (Bardeen-Cooper-Schrieffer) is fundamentally a mean field theory for electron pairing.
• Liquid crystals: Models the isotropic-to-nematic transition.
• Fluids: Analyzes the liquid-gas critical point (van der Waals equation is a mean field theory).
• Ultracold gases: Describes Bose-Einstein condensation via the Gross-Pitaevskii equation.

Mean field approximation

Assumptions and limitations

Mean field theory rests on the assumption that correlations between particles are weak enough to ignore. This assumption works well when:

• The system has many neighbors (high coordination number $z$).
• Interactions are long-range (e.g., Coulomb or dipolar).
• The spatial dimension is high (above the upper critical dimension).

It breaks down when:

• The system is low-dimensional (1D or 2D), where fluctuations dominate.
• You're very close to the critical point, where the correlation length diverges.
• Interactions are short-range and the coordination number is small.

A common consequence of these failures: mean field theory overestimates the critical temperature and the stability of ordered phases.

Mathematical formulation

The procedure has three key steps:

1. Decompose the interaction: Write the Hamiltonian so that the interaction between particles $i$ and $j$ is expressed in terms of deviations from the mean. For a spin variable $s_i$, write $s_i = \langle s_i \rangle + \delta s_i$ and neglect terms quadratic in the fluctuations $\delta s_i$.

2. Construct the mean field Hamiltonian: The result is a single-particle Hamiltonian where each particle couples to the mean field:

$H_{MF} = \sum_i h_i(\langle s_j \rangle) + \text{const}$

1. Impose self-consistency: The thermal average computed from $H_{MF}$ must reproduce the assumed mean field:

$\langle s_i \rangle = \frac{\text{Tr}(s_i \, e^{-\beta H_{MF}})}{\text{Tr}(e^{-\beta H_{MF}})}$

This equation is solved iteratively or variationally until the input and output averages agree.

Validity criteria

• The Ginzburg criterion quantifies how close to $T_c$ you can get before fluctuations invalidate mean field theory. It defines a reduced temperature window within which mean field predictions fail.
• The upper critical dimension $d_u$ (typically $d_u = 4$ for short-range models) is the dimension above which mean field exponents become exact.
• For models with infinite-range interactions (like the Curie-Weiss model), mean field theory is exact in the thermodynamic limit.

Ising model and mean field theory

The Ising model is the standard testing ground for mean field theory. It's simple enough to have exact solutions in 1D and 2D, which makes it perfect for benchmarking the approximation.

Ising model basics

The model places discrete spins $s_i = \pm 1$ on a lattice. Each spin interacts with its nearest neighbors:

$H = -J\sum_{\langle i,j \rangle} s_i s_j - h\sum_i s_i$

Here $J$ is the coupling constant ($J > 0$ for ferromagnetic interactions), $h$ is an external magnetic field, and the first sum runs over nearest-neighbor pairs. The model exhibits a phase transition for dimensions $d \geq 2$. The 2D case was solved exactly by Onsager in 1944.

Mean field solution

To apply mean field theory to the Ising model:

1. Replace neighbor spins with their average: Write $s_j \approx m$ where $m = \langle s_i \rangle$ is the magnetization per spin.

2. Construct the effective single-spin Hamiltonian: Each spin sees an effective field $h_{\text{eff}} = h + Jzm$, where $z$ is the coordination number (number of nearest neighbors). The mean field Hamiltonian becomes:

$H_{MF} = -\frac{NJzm^2}{2} - (h + Jzm)\sum_i s_i$

1. Solve the self-consistency equation: The thermal average of a single spin in the effective field gives:

$m = \tanh\!\bigl(\beta(h + Jzm)\bigr)$

This transcendental equation must be solved graphically or numerically. For $h = 0$, the $m = 0$ solution always exists, but below a critical temperature a nonzero solution (spontaneous magnetization) appears.

Critical temperature prediction

Setting $h = 0$ and linearizing the self-consistency equation near $m = 0$, the condition for a nonzero solution to appear is $\beta_c J z = 1$, giving:

$T_c^{MF} = \frac{Jz}{k_B}$

For the 2D square lattice ($z = 4$), this predicts $T_c^{MF} = 4J/k_B$. The exact Onsager result is $T_c^{\text{exact}} \approx 2.269\, J/k_B$, nearly a factor of two lower. This overestimate is typical: mean field theory always predicts too much order because it ignores the fluctuations that tend to destroy it.

As dimensionality increases, the mean field prediction improves and becomes exact as $d \to \infty$.

Phase transitions in mean field theory

Order parameters

An order parameter is a quantity that is zero in the disordered phase and nonzero in the ordered phase. Choosing the right order parameter is the first step in any mean field analysis.

• Ferromagnets: magnetization $m$
• Liquid-gas transition: density difference $\rho_l - \rho_g$
• Ferroelectrics: electric polarization $P$

Near the critical point, mean field theory predicts the order parameter vanishes as a power law:

$m \sim (T_c - T)^{\beta}$

with $\beta = 1/2$. For first-order transitions, the order parameter jumps discontinuously, and simple mean field theory handles these less cleanly (though Landau theory can describe them by including higher-order terms in the free energy expansion).

Critical exponents

Critical exponents describe how thermodynamic quantities diverge or vanish near $T_c$. Mean field theory predicts a universal set:

Universality classes

In reality, systems with the same symmetry of the order parameter and the same spatial dimensionality share the same critical exponents. This grouping is called a universality class. Common examples include:

• Ising class: Scalar order parameter, $\mathbb{Z}_2$ symmetry (e.g., uniaxial magnets, liquid-gas)
• XY class: 2-component order parameter, $O(2)$ symmetry (e.g., superfluid helium)
• Heisenberg class: 3-component order parameter, $O(3)$ symmetry (e.g., isotropic magnets)

Mean field theory predicts the same exponents for all of these, which is one of its key failures. The distinction between universality classes only emerges when fluctuations are properly included.

Beyond mean field theory

Fluctuations and correlations

Near the critical point, the correlation length $\xi$ diverges:

$\xi \sim |T - T_c|^{-\nu}$

When $\xi$ becomes large, fluctuations on all length scales matter, and the mean field assumption of weak correlations breaks down. The Ginzburg criterion quantifies this: mean field theory fails when

$\frac{|T - T_c|}{T_c} < Gi$

where $Gi$ is the Ginzburg number, which depends on the system. For conventional superconductors, $Gi$ is extremely small ($\sim 10^{-8}$), so mean field theory works almost all the way to $T_c$. For high-$T_c$ superconductors and many magnets, $Gi$ is much larger, and fluctuation effects are prominent.

Observable consequences of critical fluctuations include critical opalescence in fluids (light scattering from density fluctuations) and critical scattering in neutron diffraction from magnets.

Renormalization group approach

The renormalization group (RG) is the systematic framework for going beyond mean field theory. The key ideas:

• Progressively integrate out short-wavelength fluctuations, producing an effective theory at longer length scales.
• Track how coupling constants change ("flow") under this coarse-graining. These are the RG flow equations.
• Fixed points of the flow determine the universality classes and yield the correct critical exponents.
• The $\varepsilon$-expansion (where $\varepsilon = 4 - d$) provides a perturbative way to calculate non-mean-field exponents near the upper critical dimension.

RG explains why universality exists and why mean field theory works above $d = 4$: the Gaussian fixed point (corresponding to mean field theory) is stable for $d > 4$ but unstable below it.

Corrections to mean field

Several methods systematically improve on mean field results:

• Gaussian approximation: Include quadratic fluctuations around the mean field saddle point (equivalent to Landau-Ginzburg theory with Gaussian integrals).
• $\varepsilon$-expansion: Perturbative expansion in $\varepsilon = 4 - d$ around the upper critical dimension.
• High-temperature series expansions: Systematic power series in $\beta J$ that can be extrapolated to the critical region.
• Cluster and variational methods: Treat small clusters of spins exactly and embed them in a mean field environment (e.g., Bethe-Peierls approximation).

Mean field theory in complex systems

Spin glasses

Spin glasses are disordered magnets with randomly frustrated interactions (some ferromagnetic, some antiferromagnetic). The Sherrington-Kirkpatrick (SK) model is the infinite-range mean field version. Its solution, found by Parisi, requires replica symmetry breaking, reflecting the fact that the free energy landscape has an enormous number of metastable states organized in an ultrametric hierarchy. These ideas have found applications far beyond physics, in optimization (e.g., the traveling salesman problem) and neural network theory.

Neural networks

The Hopfield model of associative memory maps directly onto a spin glass framework. Neurons are binary variables ($\pm 1$), synaptic connections play the role of couplings, and memory retrieval corresponds to the system relaxing to a stored pattern. Mean field analysis predicts a storage capacity of approximately $0.14N$ patterns for $N$ neurons, above which the network undergoes a phase transition to a "confused" state where retrieval fails.

Ecological models

Mean field approaches appear in theoretical ecology through spatially averaged Lotka-Volterra equations. By ignoring spatial correlations in population densities, these models predict population oscillations and extinction thresholds. The mean field approximation works reasonably well for well-mixed populations but can miss important spatial effects like pattern formation and local extinction-recolonization dynamics.

Numerical methods for mean field theory

Monte Carlo simulations

Monte Carlo methods don't assume mean field theory; instead, they sample the full configuration space stochastically. This makes them ideal for testing mean field predictions.

• The Metropolis algorithm proposes random spin flips and accepts them with a probability that satisfies detailed balance.
• Cluster algorithms (Wolff, Swendsen-Wang) flip correlated groups of spins at once, dramatically reducing critical slowing down near $T_c$.
• By comparing Monte Carlo results with mean field predictions, you can directly see where the approximation succeeds and where it fails.

Molecular dynamics approaches

Molecular dynamics integrates Newton's equations numerically for all particles (commonly using the Verlet algorithm). While computationally expensive, it provides dynamical information that Monte Carlo cannot. Mean field potentials can be incorporated to speed up simulations of large systems, and finite-size scaling analysis helps connect simulation results to the thermodynamic limit.

Self-consistent field calculations

This is the numerical implementation of the mean field self-consistency condition:

1. Start with an initial guess for the mean field (e.g., $m = 0.5$).
2. Compute the thermal averages using the current mean field.
3. Update the mean field based on the new averages.
4. Repeat until the input and output fields converge to within a chosen tolerance.

This iterative scheme is the backbone of Hartree-Fock calculations in quantum chemistry, self-consistent field theory in polymer physics, and density functional theory in electronic structure.

Mean field theory vs exact solutions

Strengths and weaknesses

Strengths:

• Analytically tractable and provides closed-form predictions.
• Captures the qualitative physics of phase transitions: symmetry breaking, order parameters, critical temperatures.
• Gives exact results above the upper critical dimension and for infinite-range models.

Weaknesses:

• Overestimates critical temperatures in finite dimensions.
• Predicts wrong critical exponents for $d < 4$.
• Completely fails in 1D (predicts a phase transition where none exists).
• Misses important physics like the Kosterlitz-Thouless transition.

Comparison with experimental data

Mean field theory often gives qualitatively correct phase diagrams, which is why it remains so widely used. For conventional superconductors, where the Ginzburg number is tiny, BCS mean field theory is quantitatively excellent. For critical exponents in magnets and fluids, the predictions are off by 10-30% in 3D, which is enough to matter for precision work but still useful for building intuition.

Limitations in low dimensions

• 1D Ising model: The exact solution shows no phase transition at any nonzero temperature. Mean field theory incorrectly predicts $T_c = 2J/k_B$ (for $z = 2$).
• 2D Ising model: Onsager's exact $T_c \approx 2.269\, J/k_B$ is much lower than the mean field $T_c = 4J/k_B$.
• Mermin-Wagner theorem: Proves that continuous symmetries cannot be spontaneously broken in 2D at finite temperature for short-range interactions. Mean field theory has no mechanism to enforce this.
• Kosterlitz-Thouless transition: The 2D XY model undergoes a topological transition driven by vortex unbinding, a phenomenon entirely outside the mean field framework.

Advanced topics in mean field theory

Dynamical mean field theory (DMFT)

DMFT maps a lattice problem onto a single impurity coupled to a self-consistent bath. Unlike static mean field theory, it retains the full local quantum dynamics (all local temporal fluctuations). This makes it powerful for strongly correlated electron systems like the Hubbard model, where it successfully describes the Mott metal-insulator transition. DMFT becomes exact in the limit of infinite coordination number and has been extended to cluster versions (cellular DMFT) that capture some spatial correlations.

Quantum mean field theory

The Hartree-Fock approximation is the quantum analog of classical mean field theory: each electron moves in the average potential created by all other electrons. For bosonic systems, the Gross-Pitaevskii equation provides a mean field description of Bose-Einstein condensates. Extensions like the Bogoliubov theory and the random phase approximation (RPA) systematically include quantum fluctuations on top of the mean field ground state.

Non-equilibrium mean field theory

Mean field ideas extend to systems driven out of equilibrium. Time-dependent mean field equations describe relaxation dynamics, and steady-state mean field solutions characterize driven systems (e.g., reaction-diffusion models, population dynamics). These approaches connect to large deviation theory and fluctuation theorems, which provide exact results for certain non-equilibrium quantities even when the full dynamics is intractable.