Mean-Field Theory

Every interesting statistical-mechanical model is hard for the same reason: each degree of freedom is coupled to many others, and those couplings fluctuate wildly with temperature. Mean-field theory cuts the knot with one audacious move — it lets each spin, atom, or molecule respond not to its jostling neighbours but to their thermal average, an effective "molecular field". The impossible many-body problem collapses to a single degree of freedom in an external field, with one twist: that field is itself built from the very response it induces, so it must be found self-consistently. That feedback loop is not a bug — it is exactly what manufactures a phase transition. This article owns the general mean-field framework and its central number, the mean-field critical temperature. For the Ising ferromagnet the recipe is explicit: write each spin as its mean plus a fluctuation, $S_i=m+\delta S_i$ , discard the product of two fluctuations, and the bilinear interaction linearizes. What remains is a self-consistency equation $m=\tanh(\beta(zJm+h))$ whose nontrivial solution switches on below $T_c=zJ/k_B$ , capturing spontaneous symmetry breaking, a continuous transition, and a diverging susceptibility. The same logic yields the van der Waals fluid, the BCS gap equation, and Weiss's 1907 theory of ferromagnetism — one approximation wearing many costumes. Mean-field theory is more than a heuristic: it is the exact leading order of controlled expansions (a saddle point after Hubbard–Stratonovich; the first term of a $1/z$ , $1/d$ , or $1/N$ series) and a rigorous variational upper bound on the free energy via the Bogoliubov–Gibbs inequality. Its failures are as instructive as its successes: by ignoring fluctuation correlations it overestimates ordering, inventing a finite $T_c$ even in one dimension where none exists, and it predicts critical exponents blind to dimension. Knowing precisely where it breaks — the Ginzburg criterion and the upper critical dimension $d_c=4$ — is the doorway to the renormalization group later in this track.

Key takeaways

Mean-field theory replaces the fluctuating interaction felt by each degree of freedom with its thermal average, turning a many-body problem into a single-body problem in a self-consistently determined effective field.
The sole approximation is dropping the product of two fluctuations, $\delta S_i\,\delta S_j\to0$ ; the exact bilinear $S_iS_j$ becomes linear, $m(S_i+S_j)-m^2$ .
For the Ising model the order parameter obeys $m=\tanh(\beta(zJm+h))$ , which spontaneously breaks the $S\to-S$ symmetry below the mean-field Curie temperature $k_BT_c=zJ$ ( $z$ = coordination number).
The mean-field free energy per spin $f(m)=\tfrac12 zJm^2-k_BT\ln[2\cosh(\beta(zJm+h))]$ has $m=0$ as its unique minimum above $T_c$ and a double well below — the transition is a pitchfork bifurcation.
It becomes exact for infinite-range couplings, in infinite dimensions, and above $d_c=4$ ; below that, fluctuations swamp it near $T_c$ (Ginzburg criterion) and its universal exponents are wrong.
Key idea: replace $\sum_{\langle ij\rangle}JS_iS_j$ by a single-site coupling to a self-consistent average field $h_{\text{eff}}=h+zJm$ ; the order parameter then solves a transcendental self-consistency equation that becomes exact as the coordination number, dimension, or component count diverges.

The core insight

Replace the interaction $\sum_{\langle ij\rangle}JS_iS_j$ by a single-site coupling to a self-consistent average field $h_{\text{eff}}=h+zJ\langle S\rangle$ ; the order parameter then solves a transcendental self-consistency equation that becomes exact as the coordination number, dimension, or component number diverges.

Where this lives today

Mean-field theory is the indispensable first pass on essentially every ordering problem in modern physics, because it is analytically tractable and its domain of validity is now understood. Dynamical mean-field theory (DMFT), exact in infinite dimensions, is the workhorse for strongly correlated electrons — Mott insulators, heavy-fermion metals, high- $T_c$ superconductors — capturing the local physics band theory misses. In soft matter it underlies Poisson–Boltzmann theory of electrolytes, Flory–Huggins polymer mixing, and self-consistent field theory of block copolymers; machine learning inherited the same self-consistency loop as mean-field variational inference and the Boltzmann machine. What makes the theory permanent is that its errors are quantitative and bounded. The Ginzburg criterion says exactly when to trust it, the upper critical dimension $d_c=4$ marks where its exponents become exact, and its corrections organize as the loop expansion of statistical field theory. When it fails — near a critical point in low dimensions — the failure is the entrance to the renormalization group, which repairs the exponents mean-field theory gets wrong. Every serious calculation of a phase transition begins with the mean-field answer and then asks how large the neglected fluctuations are.

The physical picture

Picture one spin buried in a crowd of neighbours. Tracking every neighbour's instantaneous orientation is hopeless, so mean-field theory replaces the crowd by its average: the spin feels a steady molecular field proportional to the bulk magnetization $m$ . But every spin obeys the same rule, and the bulk magnetization is built from all of them — so the field each spin feels depends on the ordering it helps create. That is the self-consistency loop, and it is the entire content of the theory. Above a critical temperature the only consistent answer is "no field, no order"; below it, the loop amplifies any infinitesimal seed of order into a finite magnetization. The approximation becomes exact when each spin has infinitely many neighbours, because then the law of large numbers makes the neighbour-sum genuinely non-fluctuating — its relative spread shrinks like $1/\sqrt{z}$ . This is the deep reason mean-field theory improves with coordination number $z$ , spatial dimension $d$ (where $z\sim 2d$ ), and component count $N$ : each dilutes the fluctuations it neglects. The systematic small parameter is $1/z$ (or $1/d$ , $1/N$ ), and the corrections are precisely the loop expansions of statistical field theory. The heart of the theory is the graphical solution of the self-consistency equation. Plot the straight line $y=m$ against the curve $y=\tanh(\beta zJ\,m)$ . At high temperature the $\tanh$ starts out shallow (slope $\beta zJ<1$ ) and the only intersection is the origin — no spontaneous order. As $T$ falls, the initial slope $\beta zJ$ steepens; the moment it exceeds $1$ (i.e. $k_BT<zJ$ ) the $\tanh$ pulls away from the line and two new symmetric intersections $\pm m_0$ appear. Those crossings are the spontaneous magnetization.

Figure

Graphical solution of m = tanh(βzJm): above T_c the tanh curve (slope βzJ ≪ 1) meets the line y=m only at the origin; below T_c the steeper tanh (slope βzJ > 1) crosses the line at three points, the outer two being the spontaneous magnetization ±m₀

The mean-field free energy makes the same story quantitative. Below shows $f(m)$ at three temperatures: a single well centred at $m=0$ above $T_c$ , a marginally flat bottom exactly at $T_c$ , and a symmetric double well below $T_c$ whose minima sit at $\pm m_0$ . As the system cools through $T_c$ the single minimum splits smoothly in two — a pitchfork bifurcation, and the geometric signature of a continuous (second-order) transition. Choosing one of the two degenerate wells is spontaneous symmetry breaking, the subject the sibling article aq-sm-mean-field-self-consistency develops in full.

Figure

The mean-field free energy f(m) as temperature crosses T_c: a single minimum at m=0 for T>T_c, a flat quartic bottom at T=T_c, and a symmetric double well with minima at ±m₀ for T<T_c — the pitchfork of a continuous transition

There is a useful one-sided guarantee. By the Bogoliubov inequality the mean-field free energy is a rigorous upper bound on the true free energy, so the theory can never underestimate the optimized free energy of its trial family — mean-field theory is a genuine variational principle, not merely a guess, and the "best" molecular field is the one that minimizes this bound.

The core equations

The starting point is the nearest-neighbour Ising ferromagnet, whose only difficulty is the bilinear coupling $S_iS_j$ tying every pair of neighbours. The field $h$ is in energy units so $\beta h$ is dimensionless.

\mathcal{H}=-J\sum_{\langle ij\rangle}S_iS_j-h\sum_i S_i,\qquad S_i=\pm1 \qquad (1)

where: $J>0$ — ferromagnetic nearest-neighbour exchange; $\langle ij\rangle$ — distinct nearest-neighbour bonds; $h$ — external field; $S_i=\pm1$ — Ising spin at site $i$ . Everything downstream tames the one hard term $-J\sum_{\langle ij\rangle}S_iS_j$ without solving the full lattice, which for $d\ge3$ has no closed form. The single mean-field approximation is to write each spin as its average plus a fluctuation and throw away the product of two fluctuations. This is the whole theory in one line; every result below follows from it.

S_iS_j=m(S_i+S_j)-m^2+\underbrace{\delta S_i\,\delta S_j}_{\to\,0},\qquad m=\langle S_i\rangle,\ \ \delta S_i=S_i-m \qquad (2)

where: $m$ — order parameter (magnetization per spin); $\delta S_i$ — fluctuation about the mean. The retained terms are linear in the spins; the discarded $\delta S_i\,\delta S_j$ is the two-point fluctuation correlation, and dropping it is exactly why mean-field theory misses correlations — and is exact when fluctuations vanish. Substituting (2) into (1) and counting bonds (each site has $z$ neighbours, each bond shared by two sites, so there are $Nz/2$ bonds) collapses the interacting Hamiltonian into $N$ independent single-site problems in a common effective field.

\mathcal{H}_{\text{MF}}=-(zJm+h)\sum_i S_i+\tfrac12 NzJm^2 \qquad (3)

where: $z$ — coordination number; $N$ — sites; $h_{\text{eff}}=zJm+h$ — the self-consistent molecular field each spin feels; the constant $\tfrac12 NzJm^2$ corrects the double-counted mean energy. The spins are now decoupled — the many-body problem is gone — at the price of an unknown $m$ still to determine. Each spin is now a two-state system in the field $h_{\text{eff}}$ , so its thermal average is $\tanh(\beta h_{\text{eff}})$ . Demanding that this average equal the $m$ we assumed closes the loop and gives the central self-consistency equation.

m=\tanh\!\big(\beta(zJm+h)\big),\qquad \beta=\frac1{k_BT} \qquad (4)

where: $\beta$ — inverse temperature; $\tanh$ — the exact single-spin average in a field. Equation (4) is transcendental: only $m=0$ at high $T$ , but extra solutions $\pm m_0$ appear once $T$ drops below $T_c$ . The number of solutions changes at $T_c$ — that is the phase transition. The critical temperature is fixed by when a nonzero solution first appears. At $h=0$ the right side of (4) has slope $\beta zJ$ at the origin; a nonzero crossing exists precisely when that slope exceeds $1$ .

\boxed{\,k_BT_c=zJ\,},\qquad \left.\frac{\mathrm d}{\mathrm dm}\tanh(\beta zJ m)\right|_{m=0}=\beta zJ=1 \qquad (5)

where: $T_c$ — mean-field Curie temperature. The result $k_BT_c=zJ$ is the number this article owns: order survives to higher temperature when spins have more neighbours ( $z$ ) or stronger exchange ( $J$ ) — both make the molecular field stiffer and the relative fluctuations smaller. Finally, the mean-field free energy per spin turns the self-consistency equation into a variational statement: its stationary points are the solutions of (4), and its curvature at the origin decides their stability.

f(m)=\tfrac12 zJm^2-k_BT\ln\!\big[2\cosh(\beta(zJm+h))\big] \qquad (6)

where: $f=F/N$ — Helmholtz free energy per site. Setting $\partial f/\partial m=0$ recovers (4); the second derivative at $m=0,h=0$ is $f''(0)=zJ(1-\beta zJ)$ , positive ( $m=0$ a minimum) above $T_c$ and negative ( $m=0$ a maximum, order sets in) below it. Expanding $f$ in small $m$ gives the Landau form $f\approx f_0+\tfrac12 a\,t\,m^2+\tfrac14 b\,m^4-hm$ with $t=(T-T_c)/T_c$ , the starting point of aq-sm-landau-theory-phase-transitions.

Derivation AFrom the Ising model to the self-consistency equation and the critical temperature

The goal of this derivation is to carry the nearest-neighbour Ising model all the way from its exact Hamiltonian to the self-consistency equation $m=\tanh(\beta(zJm+h))$ and the mean-field critical temperature $k_BT_c=zJ$ , making every approximation explicit. The strategy is to decouple the bilinear interaction with the single mean-field replacement of equation (2), count the bonds carefully so the constant energy is right, factorize the resulting single-site partition function, and then close the loop by demanding self-consistency. The one thing to watch is the bond counting: each spin sits in $z$ bonds, but each bond is shared by two spins, and getting the factor of $\tfrac12$ wrong corrupts both the free energy and $T_c$ .

Step 1. Decouple the interaction. Substitute $S_iS_j\approx m(S_i+S_j)-m^2$ into $-J\sum_{\langle ij\rangle}S_iS_j$ . Since each site appears in $z$ bonds, $\sum_{\langle ij\rangle}(S_i+S_j)=z\sum_i S_i$ ; since there are $Nz/2$ bonds, $\sum_{\langle ij\rangle}m^2=\tfrac{Nz}{2}m^2$ . The interaction becomes $-zJm\sum_i S_i+\tfrac12 NzJm^2$ . Why: the decoupling turns a sum over bonds into a sum over sites; the constant $\tfrac12 NzJm^2$ subtracts the mean energy that would otherwise be double counted when each spin is treated as if it alone carried the full molecular field.

Step 2. Assemble the single-site Hamiltonian. Adding the field term $-h\sum_i S_i$ gives $\mathcal{H}_{\text{MF}}=-(zJm+h)\sum_i S_i+\tfrac12 NzJm^2$ . The spins are now independent, coupled only through the common field $h_{\text{eff}}=zJm+h$ . Why: with no surviving $S_iS_j$ term the partition function factorizes — this is the entire payoff of the approximation, the reduction from $2^N$ coupled configurations to $N$ independent two-state problems.

Step 3. Factorize the partition function. Because the sites decouple, $Z_{\text{MF}}=e^{-\beta NzJm^2/2}\,[2\cosh(\beta h_{\text{eff}})]^N$ , the product of $N$ identical single-spin partition functions $Z_1=2\cosh(\beta h_{\text{eff}})$ times the constant Boltzmann factor. Why: a two-state spin in field $h_{\text{eff}}$ has $Z_1=e^{\beta h_{\text{eff}}}+e^{-\beta h_{\text{eff}}}=2\cosh(\beta h_{\text{eff}})$ ; independence lets the $N$ factors simply multiply.

Step 4. Impose self-consistency. The single-spin average is $\langle S\rangle=\dfrac{1}{\beta N}\dfrac{\partial\ln Z_{\text{MF}}}{\partial h}=\tanh(\beta h_{\text{eff}})=\tanh(\beta(zJm+h))$ . Setting $\langle S\rangle=m$ — the value we assumed at the outset — gives the transcendental self-consistency equation $m=\tanh(\beta(zJm+h))$ . Why: the assumed mean $m$ and the computed mean $\langle S\rangle$ must agree, or the "average field" was inconsistent with the ordering it produces; this closure is what makes the field self-generated rather than external.

Step 5. Locate the critical temperature. At $h=0$ , expand the right side for small $m$ : $\tanh(\beta zJm)=\beta zJ\,m-\tfrac13(\beta zJ\,m)^3+\cdots$ . A nonzero root branches off when the linear coefficient reaches $1$ , i.e. $\beta zJ=1$ , which is $k_BT_c=zJ$ . Below $T_c$ two symmetric roots $\pm m_0$ appear, with $m_0\simeq\sqrt{3}\,(1-T/T_c)^{1/2}$ from the cubic term. Why: the transition is where the tangent of the $\tanh$ at the origin first outruns the line $y=m$ ; the cubic term then pins the magnitude of the new roots, giving the mean-field exponent $\beta=1/2$ (developed in aq-sm-mean-field-critical-exponents). *Verified: the decoupling identity, the bond-counted Hamiltonian $\mathcal{H}_{\text{MF}}=-(zJm+h)\sum S_i+\tfrac12 NzJm^2$ , the self-consistency equation, and $k_BT_c=zJ$ from the tangent-slope condition are all checked symbolically in verify_mean-field-theory.py, reproducing Pathria & Beale §12.5 and Reif's Weiss molecular-field results.*

Go deeper · Derivation B · 2 minMean-field theory as a variational upper bound on the free energy

Derivation BMean-field theory as a variational upper bound on the free energy

The goal here is to show that mean-field theory is not an ad hoc averaging trick but a rigorous variational principle: the mean-field free energy is a genuine upper bound on the exact free energy, and the molecular field $h_{\text{eff}}=zJm+h$ is the optimal choice of a variational parameter. The route is the Gibbs–Bogoliubov inequality, which bounds the true free energy in terms of any trial Hamiltonian, followed by choosing a single-site trial Hamiltonian with a free field $\lambda$ and optimizing over it. The subtlety to watch is that the inequality is one-sided — it only ever bounds $F$ from above — which is exactly why mean-field theory systematically overestimates order rather than underestimating it.

Step 1. State the inequality. For any trial Hamiltonian $\mathcal{H}_0$ with free energy $F_0$ , convexity of the exponential (Jensen's inequality) gives $F\le F_0+\langle\mathcal{H}-\mathcal{H}_0\rangle_0$ , where $\langle\cdot\rangle_0$ is the average in the trial ensemble. Why: the exact $F$ is bounded above by the trial free energy plus the mean energy mismatch; equality holds only if $\mathcal{H}_0=\mathcal{H}$ , so any decoupled trial gives a strict, computable upper bound.

Step 2. Choose independent spins in a trial field. Take $\mathcal{H}_0=-\lambda\sum_i S_i$ , a set of independent spins in a variational field $\lambda$ . Then each spin has $\langle S_i\rangle_0=\tanh(\beta\lambda)\equiv m$ and the trial free energy is $F_0=-Nk_BT\ln[2\cosh(\beta\lambda)]$ . Why: a product (independent-site) trial state is the simplest one whose averages we can compute exactly, and its single parameter $\lambda$ plays the role of the yet-unknown molecular field.

Step 3. Evaluate the bound. With $\langle\mathcal{H}\rangle_0=-\tfrac12 NzJm^2-Nhm$ (using $\langle S_iS_j\rangle_0=m^2$ for independent spins) and $\langle\mathcal{H}_0\rangle_0=-N\lambda m$ , the bound becomes $F_{\text{var}}(\lambda)=-Nk_BT\ln[2\cosh(\beta\lambda)]-\tfrac12 NzJm^2+N(\lambda-h)m$ , with $m=\tanh(\beta\lambda)$ . Why: substituting the trial averages turns the abstract inequality into an explicit function of the single variational parameter $\lambda$ , ready to minimize.

Step 4. Optimize. Setting $\partial F_{\text{var}}/\partial\lambda=0$ yields $\lambda=zJm+h$ — the optimal trial field is exactly the molecular field. Substituting back reproduces the mean-field free energy per spin $f(m)$ of equation (6) and, through $m=\tanh(\beta\lambda)$ , the self-consistency equation. Why: the best independent-spin approximation to the true system is the one whose effective field equals the self-consistent molecular field; this identifies mean-field theory as the variational optimum over product states and guarantees $F_{\text{MF}}\ge F_{\text{exact}}$ . *Verified: the stationarity condition $\partial F_{\text{var}}/\partial\lambda=0\Rightarrow\lambda=zJm+h$ and the recovery of $f(m)$ and the self-consistency equation are checked in verify_mean-field-theory.py; the upper-bound property follows from Jensen's inequality (Sethna ch. 8, Pathria §11.4).*

Limiting cases

Mean-field theory is the controlled leading order of several limits, each of which sends the relative size of the neglected fluctuations to zero.

Case	Result	Meaning
Infinite-range / Curie–Weiss ( $J_{ij}=J_0/N$ for all $i,j$ )	MFT exact for all $T$	Every spin couples equally to all others; the molecular field is genuinely non-fluctuating, so dropping $\delta S_i\,\delta S_j$ costs nothing.
Large coordination / high dimension ( $z\sim2d\to\infty$ )	corrections $\sim1/z$ ; exact in the limit	The relative fluctuation of the neighbour sum $\sim1/\sqrt{z}\to0$ ; mean-field is the leading term of a $1/z$ expansion.
Many components ( $O(N)$ model, $N\to\infty$ )	saddle point exact; $1/N$ corrections	The large- $N$ (spherical) limit makes the saddle point exact, delivering mean-field exponents with controlled corrections.
Above the upper critical dimension ( $d\ge d_c=4$ )	mean-field exponents correct	For short-range models the Ginzburg criterion shows fluctuations are irrelevant at criticality above four dimensions.
Low dimension near $T_c$ ( $d<4,\ t\to0$ )	mean-field fails	Fluctuations dominate (Ginzburg criterion violated); exponents change and the renormalization group is required.

Common trapMean-field theory can be trusted right at the critical point

It usually cannot. Because it discards the fluctuation correlation $\langle\delta S_i\,\delta S_j\rangle$ , mean-field theory overestimates ordering and gives the wrong critical behaviour in low dimensions — it even predicts a spurious transition where none exists. Naive mean-field theory hands the one-dimensional Ising chain a finite $k_BT_c=2J$ (using $z=2$ ), yet the exact answer is $T_c=0$ : a single domain wall costs finite energy but gains extensive entropy, so any $T>0$ destroys long-range order in 1D. Mean-field theory is reliable away from $T_c$ , for large coordination number, or above $d_c=4$ — never trust its exponents near criticality without checking the Ginzburg criterion.

Worked examples

1 · The Curie temperature of a mean-field cubic ferromagnet (confidence builder)

We compute $T_c$ for a simple-cubic Ising ferromagnet and then find how strongly ordered it is well below $T_c$ , using only $k_BT_c=zJ$ and the self-consistency equation.

Step 1. A simple-cubic lattice has coordination number $z=6$ (six nearest neighbours). Take the exchange strength $J/k_B=1$ K. Then the mean-field Curie temperature is $k_BT_c=zJ$ , i.e. $T_c=zJ/k_B=6\times1=\mathbf{6\ K}$ .

Step 2. At $T=3$ K $=T_c/2$ solve $m=\tanh(zJm/k_BT)=\tanh(6m/3)=\tanh(2m)$ by iteration. Starting from $m=0.9$ the map converges to $m=\tanh(2m)$ at $m=\mathbf{0.9575}$ — the lattice is almost fully magnetized at half the critical temperature.

Step 3. At $T=5$ K (just below $T_c$ ) solve $m=\tanh(6m/5)=\tanh(1.2\,m)$ ; iteration gives $m=\mathbf{0.6586}$ . The magnetization has dropped substantially as $T$ climbs toward $T_c$ , and the near-critical estimate $m_0\simeq\sqrt{3(1-T/T_c)}=\sqrt{3(1-5/6)}=0.707$ brackets it correctly (the leading square-root law slightly overshoots because $T=5$ K is not yet very close to $T_c$ ).

Step 4. At $T=7$ K $>T_c$ the slope $zJ/k_BT=6/7<1$ , so the only solution of $m=\tanh(6m/7)$ is $m=\mathbf{0}$ — the ferromagnet has melted into a paramagnet. The single number $k_BT_c=zJ$ separates an ordered from a disordered phase, and the self-consistency equation traces the magnetization curve $m(T)$ across the whole ordered side.

2 · Curie–Weiss susceptibility and its divergence (exam level)

We derive the paramagnetic susceptibility above $T_c$ and use it to extract $T_c$ from a hypothetical measurement — the standard experimental route.

Step 1. Above $T_c$ , $m$ and $h$ are small, so linearize equation (4): $m\approx\beta(zJm+h)$ , giving $m(1-\beta zJ)=\beta h$ . Hence $\chi=\left.\partial m/\partial h\right|_{h=0}=\dfrac{\beta}{1-\beta zJ}$ .

Step 2. Multiply top and bottom by $k_BT$ : $\chi=\dfrac{1}{k_BT-zJ}=\dfrac{1}{k_B(T-T_c)}$ , the Curie–Weiss law with a divergence at $T_c$ and critical exponent $\gamma=1$ (developed further in aq-sm-weiss-molecular-field and aq-sm-mean-field-critical-exponents).

Step 3. Suppose a material shows $1/\chi$ extrapolating linearly to zero at $T=6$ K when plotted against $T$ . Reading off the intercept gives $T_c=\mathbf{6\ K}$ , and the slope of $1/\chi=k_B(T-T_c)$ fixes the Curie constant. Plotting $1/\chi$ versus $T$ and reading its zero crossing is exactly how experimentalists measure the mean-field $T_c$ — the straight-line Curie–Weiss behaviour holds well away from $T_c$ even in real magnets, where only the near-critical curvature betrays the failure of mean-field exponents.

Step 4. As a consistency check, at $T=12$ K $=2T_c$ the susceptibility is $\chi=1/[k_B(12-6)]=1/(6k_B)$ , half the value at $T=9$ K where $\chi=1/(3k_B)$ . The susceptibility grows without bound as $T\to T_c^+$ , the paramagnetic fingerprint of an impending transition.

Origins · 1 minmean-field theory and how this was discovered

mean-field theory

the recurring approximation of statistical physics (van der Waals 1873; Weiss 1907; Landau 1937) · replace fluctuating neighbours by a self-consistent average field

Mean-field theory is less a single discovery than a idea rediscovered in every corner of physics. Its first incarnation was van der Waals's 1873 doctoral thesis, which modelled a fluid by giving each molecule an averaged background attraction and an excluded volume — the first mean-field equation of state. Pierre Weiss made the magnetic version explicit in 1907 with his "molecular field", positing an internal field proportional to the magnetization that explained the Curie point and spontaneous magnetization decades before quantum exchange was understood. William Lawrence Bragg and Evan James Williams cast the same approximation in the language of order–disorder in alloys in 1934–35, and Lev Landau's 1937 order-parameter theory abstracted it into a universal free-energy expansion, exposing why magnets, fluids, and alloys share identical mean-field critical behaviour. The unifying modern view — mean-field theory as the saddle point of an exactly rewritten partition function, controlled by $1/z$ , $1/d$ , or $1/N$ — came with the Hubbard–Stratonovich transformation and large- $N$ methods. Onsager's 1944 exact solution of the two-dimensional Ising model was the turning point: it showed the mean-field exponents were simply wrong in low dimensions, a discrepancy that drove Wilson's renormalization group and the modern theory of critical phenomena. Mean-field theory endures because it is almost always the right first move — exact where fluctuations are weak, and a precise map of its own failure where they are not.

History

1873

— van der Waals's thesis introduces the first mean-field equation of state for fluids.

1895

— Pierre Curie establishes the Curie law and the concept of a paramagnet-to-ferromagnet transition temperature.

1907

— Weiss proposes the molecular (mean) field, explaining spontaneous magnetization and the Curie–Weiss law.

1934

— Bragg and Williams recast mean-field theory as order–disorder in alloys.

1937

— Landau formulates the order-parameter free-energy theory unifying all mean-field transitions.

1944

— Onsager's exact 2D Ising solution reveals mean-field critical exponents are wrong in low dimensions.

1972

— Wilson and Fisher's renormalization group explains why, via the upper critical dimension d_c=4.

Going further

Texts. Pathria & Beale, Statistical Mechanics, 3rd ed., ch. 11–13 (phase transitions, mean-field and Landau theory). Kardar, Statistical Physics of Fields, ch. 2–3 (saddle point, Landau–Ginzburg, Ginzburg criterion). Reif, Fundamentals of Statistical and Thermal Physics, ch. on the Weiss molecular field. Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, ch. 8–12. Goldenfeld, Lectures on Phase Transitions and the Renormalization Group, ch. 4–5. Chaikin & Lubensky, Principles of Condensed Matter Physics, ch. 4.

Provenance. Base article: AllQFT track `sm`, "Mean-Field Theory" (STAT.INTER.MFT.1254); source Pathria & Beale, Statistical Mechanics, 3rd ed., ch. 11–13. All symbolic and numeric claims re-verified here in verify_mean-field-theory.py (the decoupling identity, the bond-counted Hamiltonian, the self-consistency equation, $k_BT_c=zJ$ from the tangent-slope bifurcation, the free-energy curvature $f''(0)=zJ(1-\beta zJ)$ , and the numeric $m(T)$ at $z=6$ , $J/k_B=1$ K giving $T_c=6$ K, $m(3\,\text{K})=0.9575$ , $m(5\,\text{K})=0.6586$ ); no base physics error found. The critical exponents are derived in the sibling article aq-sm-mean-field-critical-exponents, which this article owns the framework for and hands forward by name.

Prerequisites. Partition Function and Free Energy · Ising Model · Configuration Integral and the Interacting Partition Function · Thermodynamic Potentials and the Fundamental Relation.

Leads to. Van der Waals Equation of State · Maxwell Construction and Coexistence · Landau Theory of Phase Transitions · Ginzburg Criterion · Renormalization Group.

If mean-field theory drops fluctuations entirely, why does it get anything right at all?

Because it keeps the average effect of the interaction exactly, and for many purposes that is the dominant part. Away from a critical point, correlations between fluctuations are short-ranged and small, so replacing a neighbour by its mean is an excellent approximation — the Curie–Weiss susceptibility and the location of the transition are qualitatively (and, at high coordination, quantitatively) correct. What mean-field theory misses is the collective behaviour of fluctuations, which only becomes dominant in a narrow window around $T_c$ in low dimensions. So it gets the big picture — there is a transition, an order parameter, a susceptibility that diverges — right, and only the fine critical details wrong.

What exactly is the coordination number z, and how do I find it for my lattice?

The coordination number $z$ is the number of nearest neighbours each site has. It counts the bonds meeting at a site, which is what determines how many spins contribute to one site's molecular field. For common lattices: a 1D chain has $z=2$ ; a 2D square lattice $z=4$ ; a 2D triangular lattice $z=6$ ; a 3D simple cubic lattice $z=6$ ; body-centred cubic $z=8$ ; face-centred cubic $z=12$ . In general a $d$ -dimensional hypercubic lattice has $z=2d$ . Since $k_BT_c=zJ$ in mean-field theory, higher-coordination lattices are predicted to order at higher temperatures — and because mean-field theory improves as $z$ grows, the prediction is most reliable for the densely coordinated (fcc, bcc, high- $d$ ) cases.

Is m the same thing as the magnetization I measure in the lab?

Essentially yes, up to units. Here $m=\langle S_i\rangle$ is the average spin per site, a dimensionless number between $-1$ and $+1$ (the reduced magnetization). The laboratory magnetization is $M=n\mu\,m$ , where $n$ is the number density of spins and $\mu$ is the magnetic moment per spin, so $m$ is $M$ normalized to its saturation value $M_{\text{sat}}=n\mu$ . Calling $m$ the "order parameter" emphasizes its role: it is zero in the disordered (paramagnetic) phase and nonzero in the ordered (ferromagnetic) phase, and its onset marks the transition. The self-consistency equation is cleanest in these reduced units, which is why we work with $m$ rather than $M$ throughout.

The 1D Ising model has T_c = 0, but mean-field theory gives a finite T_c. Isn't that a fatal flaw?

It is a genuine failure, but a diagnostic one rather than a fatal one. Mean-field theory replaces a fluctuating neighbour by a static average, which in one dimension is disastrous: a single domain wall (a spot where the chain flips) costs only finite energy $2J$ but can sit at any of $\sim N$ positions, so its entropy $\sim k_B\ln N$ always wins at any $T>0$ , destroying long-range order. Mean-field theory cannot see this because it has thrown away the fluctuation correlations that domain walls live in. The lesson is that mean-field theory is least trustworthy in low dimensions and near $T_c$ , and the Ginzburg criterion makes this precise — so the "flaw" is exactly the signal that tells you when to reach for the renormalization group instead.

Why is the mean-field free energy an upper bound and not a lower bound on the true free energy?

Because of the direction of the Gibbs–Bogoliubov (Jensen) inequality, $F\le F_0+\langle\mathcal{H}-\mathcal{H}_0\rangle_0$ . It says that approximating the true Boltzmann weight by a simpler trial one can only raise the estimated free energy, never lower it — the true system always finds a lower free energy than any restricted (here, independent-spin) trial family can. This one-sidedness is physically meaningful: it is the mathematical statement that mean-field theory overestimates ordering (a lower free energy for the ordered state than the truth would give), which is why it invents transitions that do not exist. Optimizing the trial field $\lambda$ minimizes the bound, giving the best mean-field free energy, but it is still an upper bound.

How does this general framework relate to the van der Waals equation and BCS superconductivity?

They are the same self-consistency loop with different order parameters. For a fluid, treating the attractive tail of the interparticle potential as a weak long-range background gives each molecule an average attraction $\propto$ density and an excluded volume, producing the van der Waals equation — fluid mean-field theory, with density as the order parameter (see aq-sm-van-der-waals-equation). For a superconductor, BCS theory decouples the four-fermion pairing interaction about the average $\Delta=\langle c_{-k\downarrow}c_{k\uparrow}\rangle$ , yielding a self-consistent gap equation structurally identical to $m=\tanh(\cdots)$ , with the gap $\Delta$ as the order parameter. In each case the recipe is: identify the order parameter, replace the interaction by a coupling to its average, and solve the resulting single-body problem self-consistently.

Where do the mean-field critical exponents β=1/2, γ=1, δ=3, α=0 come from?

They all follow from expanding the self-consistency equation near $T_c$ , and are derived in full in the sibling article aq-sm-mean-field-critical-exponents. Briefly: the cubic term in $\tanh$ gives $m\propto(T_c-T)^{1/2}$ , so $\beta=1/2$ ; linearizing above $T_c$ gives Curie–Weiss $\chi\propto(T-T_c)^{-1}$ , so $\gamma=1$ ; the critical isotherm gives $m\propto h^{1/3}$ , so $\delta=3$ ; and the specific heat has a finite jump, so $\alpha=0$ . The striking feature is that these numbers depend on nothing — not the dimension, lattice, or interaction range — which is precisely the fingerprint (and the flaw) of mean-field theory. This article owns the framework; the exponent article owns the derivation.