Mean-Field Theory

The Ising model on a lattice has every spin talking to every neighbor. On a 3D cubic lattice that's six neighbors per site. Do the math for N sites and you have O(N) interaction terms. The partition function is a sum over 2^N configurations. Exponential. Intractable. The universe is kind, so physicists invented a lie.

Mean-field theory is that lie. It is a useful lie. It is the kind of lie every physicist tells at least once.

Here's the trick. Pick a single spin, σ_i. Its neighbors are talking to it, but you don't care about their individual states. You replace each neighbor's σ_j with its average value m = ⟨σ_j⟩. Suddenly the problem collapses. The spin no longer interacts with six specific others. It interacts with an average field — a mean field — that all the neighbors collectively produce.

The effective field seen by spin i becomes H_eff = H_external + zJm, where z is the coordination number and m is the magnetization per site. The partition function factorizes. Each spin is independent. The system becomes a gas of non-interacting two-level systems sitting in a field. Solvable in one line.

But m is unknown. m is defined as the thermal average of a spin in that same field. So you get an equation:

m = tanh[(H + zJm) / (k_B T)]

This is the mean-field self-consistency equation. You solve it self-consistently — guess m, plug it in, compute a new m, repeat until convergence. The fixed point is the solution. Fixed-point equations in statistical physics always feel like cheating, and they are. The accuracy varies. The structure is universal.

Below a critical temperature T_c = zJ/k_B, the equation develops a nonzero solution. At T_c, the transition is continuous — m grows as (T_c − T)^{1/2}. The exponent β = 1/2 is the mean-field prediction. It's wrong for the 2D Ising model (the real value is 1/8), but it's in the right family. The mean-field exponents form a class — the mean-field universality class — and any system whose interactions are sufficiently long-range or sufficiently high-dimensional falls into it.

The reason is dimensional. In high enough dimensions, each spin has so many neighbors that the central limit theorem kicks in. The fluctuations of the local field scale as 1/√(number of neighbors). When that number goes to infinity, the field becomes sharp. The approximation becomes exact. Six dimensions is the magic number for the Ising model — above d = 4, mean-field theory is asymptotically correct.

What does mean-field theory get right? The qualitative picture of a phase transition. The existence of an ordered phase at low temperature and a disordered one at high temperature. The continuous nature of the transition. What it gets wrong? Critical exponents below the upper critical dimension. Fluctuations. Correlations. The world near T_c is messy, and mean-field theory smooths the mess away like a bad interior decorator.

Still, the self-consistency equation m = tanh[(H + zJm)/k_B T] is one of the most important equations you'll encounter in condensed matter. It appears in the Curie-Weiss theory of paramagnetism. It appears in the Bragg-Williams approximation for order-disorder transitions. It appears in the Hartree-Fock approximation for interacting electrons. It appears in neural network models. The structure repeats because the logic repeats: replace complicated interactions with an average, solve, demand consistency.

Mean-field theory is not a calculation. It's a way of thinking. You isolate a single degree of freedom, surround it with the average effect of everything else, and let self-consistency do the work. It is approximately right everywhere and exactly right in the limit that matters.