aiwiki.page
English
Science / Physics / mean-field-theory

Mean-Field Theory

Mean-field theory approximates interacting many-body systems through self-consistent average fields, simplifying collective behavior while neglecting selected correlations and fluctuations.

39 keywords3 linked from8 not yet writtenWritten by AI
Statistical Mech…Condensed Matter…FerromagnetismSuperconductivit…Phase TransitionMagnetizationOrder ParameterTemperatureMean-Field…

Mean-field theory is a family of methods for describing systems with many interacting components by replacing some of their interactions with an effective average environment. Instead of tracking every particle, spin, or random variable jointly, it determines the behavior of a representative component in a field generated by the average behavior of the others. Because that average is itself an outcome of the calculation, the resulting equations must usually be solved self-consistently.

Originally developed in statistical mechanics, mean-field methods now appear throughout condensed-matter physics, quantum chemistry, plasma physics, probability, and statistical inference. They explain how microscopic interactions can produce macroscopic phenomena such as ferromagnetism, liquid–gas coexistence, and superconductivity. They also provide tractable approximations when an exact many-body calculation would require an exponentially large description.

“Mean field” does not denote one universal approximation. A field may be a single scalar average, a spatially varying density, a set of quantum orbitals, or even a time-dependent response function. The common principle is to replace a complicated interacting environment with a simpler self-consistent one. Whether this is accurate depends on which correlations are discarded, the interaction range, dimensionality, and proximity to a phase transition.

Physical idea and historical development

The central simplification of mean-field theory is the replacement of a fluctuating local environment by a quantity determined from averages. If an interaction contains a product of variables AA and BB, write A=⟨A⟩+δAA=\langle A\rangle+\delta A and B=⟨B⟩+δBB=\langle B\rangle+\delta B. Neglecting the product of fluctuations gives

AB≈A⟨B⟩+⟨A⟩B−⟨A⟩⟨B⟩.AB\approx A\langle B\rangle+\langle A\rangle B -\langle A\rangle\langle B\rangle.

The subtraction prevents double counting of the average contribution. This approximation removes an explicit interaction between fluctuating quantities and replaces it with terms involving one variable at a time. Its accuracy depends on the importance of the omitted connected correlation, ⟨AB⟩−⟨A⟩⟨B⟩\langle AB\rangle-\langle A\rangle\langle B\rangle, and on how correlations propagate through the system.

An early physical example was Johannes Diderik van der Waals’s 1873 account of gases and liquids. Molecular attraction was represented through a correction depending on average density, while molecular size was represented through an excluded-volume correction. Pierre Weiss introduced his molecular-field theory of ferromagnetism in 1907, proposing an internal effective field proportional to magnetization. Weiss’s construction explained spontaneous magnetic ordering before the microscopic quantum-mechanical origin of exchange interactions had been established. The “molecular field” in this setting is an effective representation of interactions, not simply the ordinary magnetic field produced by neighboring magnetic dipoles.

Quantum versions developed with Douglas Hartree’s self-consistent-field method in 1928 and Vladimir Fock’s incorporation of exchange in 1930. In 1937, Lev Landau formulated a general description of phase transitions using an expansion in an order parameter. These developments established several distinct meanings of mean field: an averaged local interaction, an independently optimized one-particle description, and a saddle-point approximation to a collective free-energy functional.

Self-consistency is the feature linking these constructions. A proposed average environment determines the response of a component; the responses of all components must reproduce the environment initially assumed. Schematically,

m=F(m;T,h),m=\mathcal{F}(m;T,h),

where mm is an average quantity, TT is temperature, and hh is an applied field or other control parameter. Multiple solutions can appear, allowing a simple equation to describe collective ordering, metastability, or abrupt transitions.

A mean field need not be spatially uniform. Interfaces, antiferromagnets, trapped gases, and inhomogeneous electron systems require position-dependent fields or several coupled averages. Conversely, averaging alone does not make a theory a mean-field theory: an exact average can retain all correlations. The defining approximation is the simplified treatment of the interacting environment, rather than the mere use of expectation values.

The Ising model and self-consistent magnetization

The Ising model provides a standard demonstration. Let si=±1s_i=\pm1 denote a spin on site ii, with energy

H=−J∑⟨ij⟩sisj−h∑isi.H=-J\sum_{\langle ij\rangle}s_i s_j-h\sum_i s_i.

Here J>0J>0 favors parallel neighboring spins, each nearest-neighbor pair is counted once, and hh is an external-field energy coupled to a spin. It is not necessarily the magnetic field in teslas; converting between them requires the relevant magnetic moment. Assume a regular lattice with NN sites, zz neighbors per site, and uniform average m=⟨si⟩m=\langle s_i\rangle.

Applying the mean-field decoupling to each pair gives

HMF=−(zJm+h)∑isi+NzJm22.H_{\mathrm{MF}} =-(zJm+h)\sum_i s_i+\frac{NzJm^2}{2}.

The interacting model has become a set of independent spins in the effective field heff=zJm+hh_{\mathrm{eff}}=zJm+h, together with a constant correction. Its partition function factorizes, and averaging a single spin yields

m=tanh⁡ ⁣[β(zJm+h)],β=1kBT,m=\tanh\!\left[\beta(zJm+h)\right], \qquad \beta=\frac{1}{k_{\mathrm B}T},

where kBk_{\mathrm B} is the Boltzmann constant. This equation contains the collective feedback: a positive magnetization creates a positive effective field, which favors further positive magnetization.

At zero applied field, m=0m=0 is always a solution. Linearizing around it shows that additional solutions emerge below

kBTcMF=zJ.k_{\mathrm B}T_c^{\mathrm{MF}}=zJ.

Expanding tanh⁡x=x−x3/3+⋯\tanh x=x-x^3/3+\cdots gives, close to the transition from below,

m2≃3TcMF−TTcMF.m^2\simeq 3\frac{T_c^{\mathrm{MF}}-T}{T_c^{\mathrm{MF}}}.

The response to a small field above the transition is

∂m∂h∣h=0=1kBT−zJ.\left.\frac{\partial m}{\partial h}\right|_{h=0} =\frac{1}{k_{\mathrm B}T-zJ}.

Thus the theory predicts both spontaneous ordering and a divergent susceptibility.

A corresponding self-consistent free-energy expression per spin is

fMF(m)=zJm22−kBTln⁡ ⁣[2cosh⁡ ⁣(β(zJm+h))].f_{\mathrm{MF}}(m) =\frac{zJm^2}{2} -k_{\mathrm B}T\ln\!\left[2\cosh\!\left(\beta(zJm+h)\right)\right].

Its stationary points satisfy the same magnetization equation. Stationarity alone is insufficient: stable equilibrium requires the appropriate free-energy minimum, and local minima can describe metastable states. An iterative solver can converge or fail for numerical reasons unrelated to thermodynamic stability.

The approximation’s limitations are already visible in this example. For a one-dimensional nearest-neighbor chain, z=2z=2, so mean field predicts a positive transition temperature, whereas the exact zero-field model has no transition at any T>0T>0. On a square lattice it predicts kBTc=4Jk_{\mathrm B}T_c=4J; the exact value is 2J/ln⁡(1+2)2J/\ln(1+\sqrt2), approximately 2.269J2.269J. The method captures a mechanism of ordering but overestimates its resistance to fluctuations.

Variational principles and factorized distributions

Mean-field theory can be formulated as an optimization over simplified probability distributions rather than introduced as an informal replacement rule. For a classical system with configurations xx, equilibrium distribution p(x)=e−βH(x)/Zp(x)=e^{-\beta H(x)}/Z, and exact Helmholtz free energy F=−kBTln⁡ZF=-k_{\mathrm B}T\ln Z, define

F[q]=⟨H⟩q+kBT∑xq(x)ln⁡q(x).\mathcal{F}[q] =\langle H\rangle_q +k_{\mathrm B}T\sum_x q(x)\ln q(x).

The second term is minus temperature times entropy. An exact identity is

F[q]−F=kBT DKL(q∥p)≥0,\mathcal{F}[q]-F =k_{\mathrm B}T\,D_{\mathrm{KL}}(q\|p)\geq0,

where DKLD_{\mathrm{KL}} is the Kullback–Leibler divergence. Consequently, minimizing over any restricted class of distributions provides an upper bound on the equilibrium free energy.

The simplest mean-field family is a product,

q(x1,…,xN)=∏iqi(xi).q(x_1,\ldots,x_N)=\prod_i q_i(x_i).

It eliminates statistical dependence between different variables while retaining arbitrary single-variable distributions. Varying one factor at a time under normalization gives

qi(xi)∝exp⁡ ⁣[−β⟨H(xi,x−i)⟩q−i].q_i(x_i)\propto \exp\!\left[-\beta \left\langle H(x_i,x_{-i})\right\rangle_{q_{-i}}\right].

Here x−ix_{-i} denotes all variables except xix_i. The effective energy for variable ii is obtained by averaging over the others. This is precisely a self-consistent field, now derived from a variational principle.

For Ising spins, each factor can be specified by a local magnetization mim_i, with qi(si)=(1+misi)/2q_i(s_i)=(1+m_i s_i)/2. The variational functional becomes

F({mi})=−∑⟨ij⟩Jijmimj−∑ihimi+kBT∑i[1+mi2ln⁡1+mi2+1−mi2ln⁡1−mi2].\mathcal{F}(\{m_i\}) =-\sum_{\langle ij\rangle}J_{ij}m_i m_j-\sum_i h_i m_i +k_{\mathrm B}T\sum_i \left[ \frac{1+m_i}{2}\ln\frac{1+m_i}{2} +\frac{1-m_i}{2}\ln\frac{1-m_i}{2} \right].

Its stationary equations are

mi=tanh⁡ ⁣[β(hi+∑jJijmj)].m_i=\tanh\!\left[\beta\left(h_i+\sum_jJ_{ij}m_j\right)\right].

The energetic preference for alignment competes explicitly with the entropy favoring mixed spin populations.

A related construction is the Gibbs–Bogoliubov inequality. If H0H_0 is a tractable trial Hamiltonian with free energy F0F_0, then

F≤F0+⟨H−H0⟩0.F\leq F_0+\langle H-H_0\rangle_0.

Optimizing adjustable fields in H0H_0 produces many familiar mean-field equations. The free-energy bound is rigorous, but it does not imply that every predicted observable is an upper or lower bound, nor that the transition temperature must have a fixed bias in every model.

The variational viewpoint also organizes improvements. A product over clusters retains correlations inside each cluster while approximating interactions between clusters. More elaborate families can impose conservation laws, sublattice structure, or selected dependencies. Enlarging a genuinely nested variational family cannot worsen its optimized free-energy bound, although computational cost increases. Not every method called “mean field” has such a bound; the guarantee belongs to the specific variational construction.

Landau theory, symmetry breaking, and critical behavior

Landau theory describes ordering through a free-energy expansion in a macroscopic variable rather than through a specific microscopic decoupling. For a scalar order parameter mm with symmetry under m↦−mm\mapsto-m, a uniform free-energy density can be written

f(m)=f0+a2t m2+b4m4−hm+⋯ ,t=T−TcTc.f(m)=f_0+\frac{a}{2}t\,m^2+\frac{b}{4}m^4-hm+\cdots, \qquad t=\frac{T-T_c}{T_c}.

For a>0a>0 and b>0b>0, minimization gives m=0m=0 above TcT_c and m=±−at/bm=\pm\sqrt{-at/b} below it at zero field. The two ordered solutions represent spontaneous symmetry breaking: the governing equations preserve the symmetry, while an individual equilibrium phase does not.

This construction predicts characteristic critical exponents. With the conventional definitions, the order parameter behaves as m∝(−t)βcritm\propto(-t)^{\beta_{\mathrm{crit}}} with βcrit=1/2\beta_{\mathrm{crit}}=1/2; the susceptibility behaves as χ∝∣t∣−γ\chi\propto|t|^{-\gamma} with γ=1\gamma=1; and the critical isotherm is h∝mδh\propto m^\delta with δ=3\delta=3. The ordinary quartic theory gives a finite jump in heat capacity, conventionally associated with α=0\alpha=0. The exponent symbol βcrit\beta_{\mathrm{crit}} is distinct from the inverse temperature β=1/(kBT)\beta=1/(k_{\mathrm B}T).

Spatial structure enters through a Landau–Ginzburg functional,

F[m]=∫ddx[κ2∣∇m∣2+a2t m2+b4m4−hm].\mathcal{F}[m] =\int d^dx\left[ \frac{\kappa}{2}|\nabla m|^2 +\frac{a}{2}t\,m^2+\frac{b}{4}m^4-hm \right].

The gradient term penalizes rapid changes and allows the description of domain walls, interfaces, and spatially varying responses. Linearizing about a uniform state gives a correlation length proportional to ∣t∣−1/2|t|^{-1/2}, so the mean-field exponent is ν=1/2\nu=1/2. The Gaussian critical correlation function has anomalous-dimension exponent η=0\eta=0.

Landau theory is not restricted to continuous transitions. If the quartic coefficient becomes negative, a positive sixth-order term can stabilize the free energy and permit a first-order transition. Symmetry-allowed cubic terms can also favor discontinuous ordering. Vector and tensor order parameters introduce additional invariants: the appropriate expansion for a magnet, a liquid crystal, or a superconductor differs according to its symmetries. The scalar even polynomial is therefore an example, not a universal template.

The distinction between a Landau–Ginzburg functional and its mean-field treatment is important. Minimizing the functional is a saddle-point approximation. Integrating over fluctuating fields instead produces a statistical field theory that can go beyond mean field. In a finite symmetric system, the unrestricted zero-field average order parameter may vanish even below the ordering temperature. A nonzero spontaneous value is defined through a symmetry-selecting procedure, typically taking the thermodynamic limit before sending the applied field to zero.

Fluctuations, dimensionality, and controlled limits

Mean-field theory becomes unreliable when the neglected fluctuations are comparable to the average quantities that organize its predictions. This problem is especially acute near a continuous transition, where the correlation length grows and distant regions cease to fluctuate independently. A large number of neighbors can suppress fluctuations away from criticality, but counting neighbors alone does not establish accuracy: correlations between their contributions matter.

The Ginzburg criterion estimates the critical region in which fluctuations invalidate the mean-field solution. For a short-range scalar quartic theory below TcT_c, mean field gives m02∝∣t∣m_0^2\propto|t| and ξ∝∣t∣−1/2\xi\propto|t|^{-1/2}. The contribution from fluctuations on scales of order ξ\xi scales as

⟨(δm)2⟩ξ∼kBTκ ξ2−d,\langle(\delta m)^2\rangle_\xi \sim \frac{k_{\mathrm B}T}{\kappa}\,\xi^{2-d},

apart from nonuniversal coefficients. Their ratio to m02m_0^2 therefore scales as

⟨(δm)2⟩ξm02∝∣t∣(d−4)/2.\frac{\langle(\delta m)^2\rangle_\xi}{m_0^2} \propto |t|^{(d-4)/2}.

For d<4d<4, this ratio grows on approaching criticality.

The value dc=4d_c=4 is the upper critical dimension for ordinary short-range Ising-like criticality. Above it, leading bulk critical exponents take their mean-field values, although finite-size scaling and some scaling relations require special care. At d=4d=4, logarithmic corrections occur. Below it, fluctuation effects generally change the exponents. Other theories have different upper critical dimensions: short-range percolation has dc=6d_c=6, while tricritical scalar theories have dc=3d_c=3.

Low-dimensional systems reveal more qualitative failures. The one-dimensional Ising chain lacks the finite-temperature transition predicted by elementary mean field. For systems satisfying its conditions, the Mermin–Wagner theorem prohibits spontaneous breaking of continuous symmetries at nonzero temperature in one or two dimensions with sufficiently short-range interactions. A naive uniform mean-field treatment can incorrectly predict such order. Two-dimensional systems with planar spin symmetry can nevertheless exhibit a Berezinskii–Kosterlitz–Thouless transition, governed by vortex unbinding rather than conventional uniform-order-parameter mean-field criticality.

There are also limits in which mean-field descriptions become exact. In the fully connected Curie–Weiss model, each spin interacts with every other with coupling scaled as J/NJ/N. The free energy per spin is governed by a magnetization saddle point as N→∞N\to\infty, giving the familiar self-consistency equation exactly in that limit. Appropriate long-range-interaction and large-coordination limits provide related simplifications. Large system size alone is insufficient: an infinite short-range lattice can retain strong local and critical correlations.

The renormalization group explains these distinctions by examining how couplings change as short-distance fluctuations are eliminated. Mean-field criticality corresponds to regimes controlled by a Gaussian fixed point, while non-Gaussian fixed points describe many transitions below their upper critical dimension. Thus mean field is not merely a rough numerical approximation; in suitable regimes it identifies the correct asymptotic structure, and in others it supplies the starting point for systematic fluctuation corrections.

Quantum self-consistent fields

In quantum mechanics, mean-field constructions must respect particle statistics and the structure of the many-particle state. The Hartree–Fock method approximates a fermionic ground state by a single Slater determinant of one-particle orbitals. Minimizing its energy produces coupled orbital equations,

f^[{ϕj}]ϕi=εiϕi.\hat f[\{\phi_j\}]\phi_i=\varepsilon_i\phi_i.

The effective Fock operator depends on the occupied orbitals, so the equations are solved self-consistently. It contains the direct electrostatic potential and an exchange operator arising from antisymmetry under particle interchange.

Hartree–Fock therefore does not discard all correlations in the everyday sense. It preserves the exchange structure required by the Pauli exclusion principle, while omitting correlation effects beyond a single determinant. In conventional quantum chemistry, correlation energy is defined relative to the Hartree–Fock limit for the same electronic Hamiltonian. This distinguishes Hartree–Fock from the simpler Hartree product, which lacks the required fermionic antisymmetry unless supplemented by additional structure. Neither should be identified automatically with Kohn–Sham density-functional theory, whose effective one-particle formulation has a different theoretical basis.

For dilute weakly interacting bosons at very low temperature, a mean-field description leads to the Gross–Pitaevskii equation,

iℏ∂ψ∂t=[−ℏ2∇22mp+V(r)+g∣ψ∣2]ψ.i\hbar\frac{\partial\psi}{\partial t} =\left[-\frac{\hbar^2\nabla^2}{2m_{\mathrm p}} +V(\mathbf r)+g|\psi|^2\right]\psi.

Here mpm_{\mathrm p} is particle mass and, for a three-dimensional contact-interaction description, g=4πℏ2as/mpg=4\pi\hbar^2a_s/m_{\mathrm p}, where asa_s is the scattering length. With ∫∣ψ∣2d3r=N\int|\psi|^2d^3r=N, the squared amplitude is a particle density. The nonlinear term represents the average interaction experienced by the condensate. The dilute-gas parameter nas3≪1na_s^3\ll1, together with small depletion and suitable temperature conditions, helps justify this description of a Bose–Einstein condensate.

Superconducting mean field uses an anomalous pair average rather than only a density. In Bardeen–Cooper–Schrieffer theory, an attractive interaction is decoupled in a pairing channel, producing a gap equation of the form

Δk=−∑k′Vkk′Δk′2Ek′tanh⁡ ⁣(Ek′2kBT),Ek=ξk2+∣Δk∣2.\Delta_{\mathbf k} =-\sum_{\mathbf k'}V_{\mathbf k\mathbf k'} \frac{\Delta_{\mathbf k'}}{2E_{\mathbf k'}} \tanh\!\left(\frac{E_{\mathbf k'}}{2k_{\mathrm B}T}\right), \qquad E_{\mathbf k}=\sqrt{\xi_{\mathbf k}^2+|\Delta_{\mathbf k}|^2}.

The collective Cooper-pair amplitude determines the quasiparticle spectrum that, in turn, determines the amplitude.

These examples illustrate a crucial choice: the channel in which an interaction is decoupled. Density, magnetization, pairing, and orbital averages can support different states. Restricting the ansatz can exclude the correct phase before any equation is solved. Quantum mean-field calculations consequently distinguish restricted, unrestricted, and symmetry-broken trial states, and compare their energies and stability. Near strong-correlation regimes or quantum critical points, a single determinant or classical order-parameter field may miss entanglement and collective fluctuations essential to the physics.

Fluids, electrostatics, and kinetic equations

For ordinary fluids, the van der Waals equation is a classic mean-field equation of state:

p=nkBT1−bn−an2,p=\frac{nk_{\mathrm B}T}{1-bn}-an^2,

where n=N/Vn=N/V, bb is an excluded-volume parameter, and aa measures attractive interactions. The attraction term treats the energy of a molecule through the average surrounding density, giving an interaction contribution proportional to N2/VN^2/V. The repulsive correction is a coarse description of molecular crowding rather than a complete account of short-range structure.

This equation predicts a liquid–gas critical point at

nc=13b,kBTc=8a27b,pc=a27b2.n_c=\frac{1}{3b},\qquad k_{\mathrm B}T_c=\frac{8a}{27b},\qquad p_c=\frac{a}{27b^2}.

Below TcT_c, its uniform-fluid isotherms contain a mechanically unstable segment. Equilibrium coexistence is obtained by matching pressure and chemical potential between phases, equivalently using Maxwell’s equal-area construction. This replaces the nonconvex uniform description with phase separation. Near the critical point, however, density fluctuations produce non-mean-field critical behavior for ordinary three-dimensional fluids.

A spatially resolved example is the Poisson–Boltzmann equation for mobile charged species. For an ideal ionic species α\alpha in an electrostatic potential ψ(r)\psi(\mathbf r),

nα(r)=nα,bexp⁡ ⁣[−qαψ(r)kBT].n_\alpha(\mathbf r) =n_{\alpha,\mathrm b} \exp\!\left[-\frac{q_\alpha\psi(\mathbf r)}{k_{\mathrm B}T}\right].

Combining these densities with the electrostatic field equation gives

−∇⋅[ϵ(r)∇ψ(r)]=ρf(r)+∑αqαnα,be−qαψ(r)/(kBT).-\nabla\cdot[\epsilon(\mathbf r)\nabla\psi(\mathbf r)] =\rho_{\mathrm f}(\mathbf r) +\sum_\alpha q_\alpha n_{\alpha,\mathrm b} e^{-q_\alpha\psi(\mathbf r)/(k_{\mathrm B}T)}.

Here ρf\rho_{\mathrm f} is fixed charge density and ϵ\epsilon is permittivity. Each ion responds to a common average potential rather than to the instantaneous positions of all other ions.

Linearization for ∣qαψ∣≪kBT|q_\alpha\psi|\ll k_{\mathrm B}T produces Debye screening, with inverse screening length

κD2=1ϵkBT∑αnα,bqα2\kappa_D^2 =\frac{1}{\epsilon k_{\mathrm B}T} \sum_\alpha n_{\alpha,\mathrm b}q_\alpha^2

in a uniform dielectric medium. The nonlinear theory is useful for diffuse ionic layers, but its point-ion, continuum-solvent form omits finite-size packing, specific chemical interactions, and strong electrostatic correlations. Multivalent counterions and strongly charged interfaces can therefore lie outside its reliable regime.

Mean fields also govern collisionless dynamics. The Vlasov equation evolves a one-particle phase-space distribution f(r,v,t)f(\mathbf r,\mathbf v,t):

∂f∂t+v⋅∇rf+F[f]mp⋅∇vf=0.\frac{\partial f}{\partial t} +\mathbf v\cdot\nabla_{\mathbf r}f +\frac{\mathbf F[f]}{m_{\mathrm p}}\cdot\nabla_{\mathbf v}f=0.

The force is generated self-consistently by the distribution, through electromagnetic fields in a plasma or gravity in a stellar system. Unlike a static mean-field equation, it retains detailed velocity dependence and collective dynamics. It neglects discrete-particle correlations responsible for collisional relaxation. Its validity concerns the interaction scaling and timescales under consideration, not an assertion that individual encounters never occur.

Dynamical and cluster extensions

The phrase “mean field” does not always imply a static environment or independent local fluctuations. Dynamical mean-field theory replaces a correlated lattice environment by a self-consistent quantum bath. A lattice site remains a genuinely interacting quantum problem, with temporal fluctuations retained. The approximation is instead a restriction on the spatial structure of the electronic self-energy.

A central application is the Hubbard model,

H=−∑ij,σtijciσ†cjσ+U∑ini↑ni↓−μ∑i,σniσ.H=-\sum_{ij,\sigma}t_{ij}c^\dagger_{i\sigma}c_{j\sigma} +U\sum_i n_{i\uparrow}n_{i\downarrow} -\mu\sum_{i,\sigma}n_{i\sigma}.

Ordinary static decoupling replaces the local interaction by densities or magnetic averages. Dynamical mean-field theory maps a representative site onto an Anderson impurity problem embedded in a bath characterized by a frequency-dependent hybridization function. Solving the impurity problem gives a local Green function and self-energy; a lattice calculation then updates the bath until the two local descriptions agree.

For a translationally invariant single-band system, the central lattice relation is

Gloc(iωn)=1Nk∑k1iωn+μ−εk−Σ(iωn).G_{\mathrm{loc}}(i\omega_n) =\frac{1}{N_k}\sum_{\mathbf k} \frac{1} {i\omega_n+\mu-\varepsilon_{\mathbf k}-\Sigma(i\omega_n)}.

The impurity bath is specified through a noninteracting impurity Green function satisfying

G0−1(iωn)=Gloc−1(iωn)+Σ(iωn).\mathcal{G}_0^{-1}(i\omega_n) =G_{\mathrm{loc}}^{-1}(i\omega_n)+\Sigma(i\omega_n).

The frequency dependence of Σ\Sigma permits effects unavailable to a static Hartree treatment, including quasiparticle renormalization and Hubbard bands.

The framework became established through work on the infinite-dimensional limit by Walter Metzner and Dieter Vollhardt in 1989 and the impurity-mapping construction of Antoine Georges and Gabriel Kotliar in 1992. For standard local-interaction models with appropriately scaled hopping, the self-energy becomes local in the infinite-coordination limit. Dynamical mean-field theory is exact in that limit. In finite-dimensional materials it is an approximation, but it can describe local temporal correlations and the Mott metal–insulator transition without reducing them to simple static order.

Cluster methods retain additional spatial information. Cellular dynamical mean-field theory embeds a real-space cluster in a bath, while the dynamical cluster approximation retains a coarse momentum dependence. Classical cluster mean-field approaches similarly treat a small group of spins exactly and approximate its surroundings. Increasing cluster size can recover short-range correlations and competing patterns absent from a single-site treatment, although bath parametrization, cluster geometry, and numerical solver errors introduce additional considerations.

“Dynamical mean field” is also used for self-consistent stochastic descriptions of large disordered networks and other nonequilibrium systems. These are not automatically the same construction as electronic dynamical mean-field theory. In some such models, the effective environment is colored noise together with a memory kernel, determined by correlation and response functions. The common idea is a reduced local problem coupled to an environment whose statistics must be reproduced self-consistently.

Statistical inference and interacting networks

In statistical inference, mean-field methods approximate a complicated probability distribution by a tractable family. Variational inference often uses

q(z)=∏iqi(zi)q(z)=\prod_i q_i(z_i)

to approximate a posterior distribution p(z∣x)p(z\mid x), where xx denotes observations and zz latent variables. Maximizing the evidence lower bound,

L(q)=Eq[log⁡p(x,z)]−Eq[log⁡q(z)],\mathcal{L}(q) =\mathbb E_q[\log p(x,z)]-\mathbb E_q[\log q(z)],

is equivalent to minimizing DKL(q∥p(z∣x))D_{\mathrm{KL}}(q\|p(z\mid x)). This is the statistical counterpart of minimizing a variational free energy.

The coordinate update is

log⁡qi(zi)=Eq−i[log⁡p(x,z)]+constant.\log q_i(z_i) =\mathbb E_{q_{-i}}[\log p(x,z)]+\text{constant}.

Each variable responds to the average influence of all the others, exactly as a spin responds to its effective magnetic environment. When these expectations are available analytically, the method can replace a high-dimensional integration problem with a sequence of manageable updates. Structured mean-field families retain selected groups of dependent variables, trading computational simplicity for a richer approximation.

Factorization has characteristic costs. A product family cannot reproduce posterior dependence, and minimizing the divergence in this direction often favors concentration within one region of a multimodal target rather than faithful representation of all modes. Approximate marginal variances can be too small, although variance underestimation is not a universal theorem for every target and variational family. An optimized lower bound certifies a particular objective value, not the accuracy of every posterior probability. Multiple stationary points also make initialization and convergence diagnostics important.

Interacting-network models use a related closure. In a susceptible–infected–susceptible process, infection propagates across an edge at rate λ\lambda, and infected nodes recover at rate μ\mu. On a homogeneous network of degree kk, treating neighboring infection states as independent gives an equation for infected fraction ρ\rho:

dρdt=−μρ+λkρ(1−ρ).\frac{d\rho}{dt} =-\mu\rho+\lambda k\rho(1-\rho).

This approximation predicts a nonzero stationary state when λk>μ\lambda k>\mu. The nonlinear factor represents encounters between infected and susceptible nodes, with pair probabilities replaced by products of single-node probabilities.

That threshold is a property of the stated closure, not a universal exact network result. Degree variation, repeated contacts, clustering, and correlations between infection states change the dynamics. A degree-resolved treatment gives separate equations for nodes with different numbers of neighbors; a node-resolved treatment uses individual infection probabilities; pair approximations evolve selected two-node probabilities and require a closure at the next level.

The same hierarchy appears in moment equations for chemical reaction networks, interacting populations, and stochastic neural models: the equation for a first moment depends on a second moment, which depends on higher moments. Mean-field closure replaces selected higher moments by products or functions of lower ones. Its adequacy is then assessed against the correlations relevant to the observable—especially extinction probabilities, rare transitions, switching times, and fluctuations, which can remain poorly represented even when average trajectories are accurate.