All lessons

Module 6: Thermodynamics and statistical physics · Lesson 6.3 · about 55 min

The Ising model and Landau theory

Exact and approximate Ising results for CSIR NET: small clusters, the one-dimensional chain by transfer matrix with its correlation length, why 1D cannot order and 2D can (Peierls), Onsager and Kramers-Wannier duality, mean field against Bethe, Metropolis Monte Carlo, and Landau theory beyond the basics: first-order and tricritical points, Ginzburg-Landau and the Ginzburg criterion.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 6 · about 55 minutes. The basics are in the GATE lesson on phase transitions: Landau’s m2+m4 free energy, the mean-field Ising model with kBTc=zJ, and the table of critical exponents. CSIR NET Part C goes further and asks for exact results: partition functions of small clusters and chains, correlation lengths, why one dimension cannot order, the exact 2D critical temperature, and Landau theories with a first-order or tricritical point. That is this lesson. The partition-function machinery is in partition functions.

Order is a contest between energy, which wants all spins aligned, and entropy, which wants them random. In one dimension a single domain wall costs a fixed energy but gains entropy k_B ln N, so entropy always wins at T > 0. In two dimensions a wall's cost grows with its length, so order survives up to a finite T_c. Mean-field theory gets the existence right and the numbers wrong; exact solutions and Landau's symmetry argument get the rest.

1. The model and small clusters

Spins si=±1 on a lattice, nearest neighbours coupled:

H=−J∑⟨ij⟩sisj−h∑isi,Z=∑{s}e−βH,K=βJ

For a handful of spins, just list the states. Two spins: Z=2eK+2e−K=4cosh⁡K at h=0, so ⟨s1s2⟩=tanh⁡K and ⟨E⟩=−Jtanh⁡K. Three spins on a triangle: 2 aligned states with E=−3J and 6 with E=+J, so Z=2e3K+6e−K. With antiferromagnetic coupling (J<0) on the triangle, no state satisfies all three bonds: the ground state is 6-fold degenerate and keeps entropy kBln⁡6 at T=0. That is frustration, and on the full triangular lattice it leaves a residual entropy of 0.323kB per spin (Wannier).

2. The one-dimensional chain, exactly

Write Z for a ring of N spins as a trace of a product of 2 × 2 transfer matrices:

Z=TrTN,T=(eK+βhe−Ke−KeK−βh),λ±=eKcosh⁡βh±e2Ksinh2⁡βh+e−2K

So Z=λ+N+λ−N, and for large N only λ+ matters: f=−kBTln⁡λ+. At h=0, λ±=2cosh⁡K,2sinh⁡K, and

m(h)=sinh⁡βhsinh2⁡βh+e−4K,⟨s0sr⟩=tanhr⁡K=e−r/ξ,ξ=−1ln⁡tanh⁡K
  • m→0 as h→0 at every T>0: no spontaneous magnetisation. The zero-field susceptibility per spin is χ=βe2K, large but finite.
  • The correlation length is finite at every T>0: ξ=3.7 lattice spacings at K=1, 1.3 at K=0.5. At low temperature ξ≈12e2K grows exponentially but diverges only at T=0.
  • For a short ring the second eigenvalue still counts: the nearest-neighbour correlation is (t+tN−1)/(1+tN) with t=tanh⁡K.

Why 1D cannot order (Landau’s argument). Start from all spins up and flip a block, creating two walls. Each wall costs 2J and can sit at any of about N places, so the free energy change of one wall is ΔF=2J−kBTln⁡N, negative for large N at any T>0. Walls proliferate and destroy order. Their density is about e−2K, so the typical domain is about e2K long, matching ξ.

3. Two dimensions: Peierls, duality and Onsager

Peierls’ argument. In 2D, a domain of up spins inside a down sea is bounded by a closed wall of length L, costing 2JL. The number of such walls grows at most as 3L (a walk that cannot immediately backtrack), so the free energy is roughly L(2J−kBTln⁡3): walls are suppressed below kBT≈2J/ln⁡3=1.82J, and long-range order exists at low T. This is a rigorous lower bound on Tc.

Kramers-Wannier duality. The high-temperature expansion of Z (in powers of tanh⁡K, counting closed loops of bonds) equals the low-temperature expansion (in powers of e−2K∗, counting domain walls on the dual lattice) when

tanh⁡K∗=e−2K⟺sinh⁡2Ksinh⁡2K∗=1

The square lattice is its own dual. If there is one critical point, it is self-dual: sinh⁡2Kc=1, so

Kc=12ln⁡(1+2)=0.4407,kBTc=2Jln⁡(1+2)=2.269J

Onsager and Yang. Onsager solved the zero-field 2D model exactly in 1944; Yang derived the spontaneous magnetisation in 1952:

m=[1−sinh−4⁡(2βJ)]1/8  (T<Tc),C∼−ln⁡|T−Tc|

so β=1/8 and α=0 (logarithmic). At T=0.9Tc, m=0.896, much larger than mean field’s 0.525 at the same reduced temperature, because the exact curve is so steep (β=1/8). The other exponents are γ=7/4, ν=1, η=1/4, δ=15. No exact solution exists in 3D or in 2D with a field.

K = J/kBT
reduced coupling; large K is low temperature
Transfer matrix
Tss′ = eKss′ + βh(s + s′)/2; Z = Tr TN; the largest eigenvalue gives f, the ratio of the two gives ξ
ξ
correlation length: ⟨s0sr⟩ − m² ~ e−r/ξ; diverges at a critical point
Duality
exact map between the high- and low-temperature expansions of Z
Physical meaning
Whether a system orders is decided by how the energy cost of a domain wall grows with its size compared with its entropy. Dimension and symmetry decide it; the details of the lattice only shift Tc.

4. Approximations compared

Mean field replaces neighbours by their average, m=tanh⁡[β(zJm+h)], giving kBTc=zJ and β=12, γ=1, δ=3. It ignores fluctuations, so it overestimates Tc and predicts a transition even in 1D. The Bethe approximation treats one spin and its z neighbours exactly, with the rest as a field on the neighbours:

kBTcBethe=2Jln⁡zz−2
Lattice z Mean field Bethe Exact
chain 2 2J 0 (correct) 0
square 4 4J 2.885J 2.269J
simple cubic 6 6J 4.933J 4.511J (numerical)

Bethe is exact on a tree (Bethe lattice) and gets 1D right, but its exponents are still mean-field.

Monte Carlo. For anything not solvable, simulate. The Metropolis rule: pick a spin, compute the energy change ΔE=2Jsi∑jsj+2hsi of flipping it, flip if ΔE≤0, otherwise flip with probability e−βΔE. This obeys detailed balance, so the spins sample the Boltzmann distribution. On the square lattice in zero field, ΔE∈{−8J,−4J,0,4J,8J}, and only e−4K and e−8K need computing. Near Tc domains become huge and decorrelate slowly (critical slowing down, relaxation time ∼ξz with z≈2.17 for Metropolis); cluster algorithms (Wolff, Swendsen-Wang) flip whole domains and avoid it. Finite lattices round the transition; the Binder cumulant U=1−⟨m4⟩/3⟨m2⟩2 goes from 0 (disordered, Gaussian m) to 23 (ordered), and its curves for different sizes cross at Tc.

A 64 × 64 Ising lattice updated by the Metropolis rule. Move the temperature slider and watch |m| settle against the exact Onsager-Yang curve (dashed). Near kTc = 2.269 J domains of every size appear and the simulation slows down; well above it the lattice looks like noise.

5. Landau theory beyond the basics

Landau’s free energy is a power series in the order parameter allowed by symmetry. With m→−m symmetry and b>0, the transition is continuous (see the GATE lesson). Three extensions come up in NET.

A cubic term (no m→−m symmetry, for example the nematic-isotropic transition or a three-state Potts model): F=am2−cm3+bm4. A non-zero minimum appears before a reaches zero, and the transition is first order at a=c2/4b with jump m=c/2b.

Negative quartic term: F=αm2−βm4+γm6, α=a(T−T0).

Coexistence. The ordered minimum is level with m = 0 when F(m) = 0 and F′(m) = 0 together: α − βm² + γm⁴ = 0 and α − 2βm² + 3γm⁴ = 0.

Solve. Subtracting gives m² = β/2γ, and then α = β²/4γ.

Transition. First order at T₁ = T₀ + β²/(4aγ), where m jumps from 0 to √(β/2γ) and a latent heat T₁ΔS = aT₁β/(2γ) is released.

Spinodals. The m = 0 state stays metastable down to T₀ (supercooling limit), and the ordered state up to T₀ + β²/(3aγ) (superheating limit), so hysteresis spans that range.

When the quartic coefficient passes through zero, the line of continuous transitions turns first order at a tricritical point. There F=αm2+γm6 and m∝(Tc−T)1/4: β=14, γ=1, δ=5, α=12. Helium-3/helium-4 mixtures show one.

Ginzburg-Landau. Let m vary in space and charge for gradients:

F=∫ddr[12κ(∇m)2+am2+bm4−hm]

A small field h(r→) gives the Ornstein-Zernike correlation G(k)∝1/(k2+ξ−2) with ξ=κ/2a∝|T−Tc|−1/2, so ν=12 and η=0. Below Tc, ξ=κ/4|a|, a factor 2 shorter at the same |T−Tc|. The same functional, with complex ψ and a vector potential, is the Ginzburg-Landau theory of superconductors.

When does mean field fail? (Ginzburg criterion). The mean square fluctuation of m averaged over a correlation volume, kBTχ/ξd, compared with m2 itself, scales as |t|dν−γ−2β=|t|(d−4)/2 with mean-field exponents. For d>4 they vanish at the transition and mean field is exact; for d<4 they win close enough to Tc. d=4 is the upper critical dimension. In conventional superconductors the window where fluctuations matter is only about 10−14 of Tc wide, which is why BCS mean field works so well there.

“Mean-field theory is a rough guess that is wrong everywhere.” It is wrong near Tc in low dimensions, where fluctuations dominate. But it is exact above four dimensions, exact for infinite-range interactions, accurate far from Tc, and nearly exact for conventional superconductors and many structural transitions with a long interaction range. Its predictions of Tc are upper bounds, and its exponents are the starting point for the renormalisation group.

  • Small clusters: enumerate. Pair Z=4cosh⁡K; triangle 2e3K+6e−K; antiferromagnetic triangle is frustrated, ground-state entropy kBln⁡6.
  • 1D: Z=λ+N+λ−N, λ±=2cosh⁡K,2sinh⁡K at h=0; f=−kBTln⁡(2cosh⁡K); ⟨s0sr⟩=tanhr⁡K, ξ=−1/ln⁡tanh⁡K; no order at T>0 (one wall: 2J−kBTln⁡N).
  • 2D: Peierls proves order; duality sinh⁡2Ksinh⁡2K∗=1 gives kBTc=2J/ln⁡(1+2)=2.269J; m=[1−sinh−4⁡2K]1/8; β=1/8, γ=7/4, ν=1, η=1/4, δ=15, α=0 (log).
  • Tc: mean field zJ > Bethe 2J/ln⁡[z/(z−2)] > exact.
  • Metropolis: accept with min⁡(1,e−βΔE); critical slowing down; Binder cumulant 0 to 2/3, crossing at Tc.
  • Landau: cubic term or negative quartic gives first order (α=β2/4γ, m2=β/2γ); tricritical β=1/4; Ginzburg-Landau ξ∝|t|−1/2, Ornstein-Zernike; upper critical dimension 4.

Practise: exam questions on this lesson

NET Part C style, each with a full solution. More on the CSIR NET Physics practice page.

Next: the kinetic theory of transport, from mean free paths to viscosity, conductivity and the Boltzmann equation. Watch domains grow on a larger lattice in fibub’s Ising model simulator.

Check yourself

1For the zero-field Ising chain, how does the spin-spin correlation ⟨s₀s_r⟩ depend on distance r?
2On the square lattice, Kramers-Wannier duality relates couplings K and K* by sinh 2K sinh 2K* = 1. What follows if the transition is unique?
3A Landau free energy F = αm² − βm⁴ + γm⁶ with β, γ > 0 and α = a(T − T₀). What kind of transition occurs and where?

Sources

  • R. J. Baxter, Exactly Solved Models in Statistical Mechanics (1982), chapters 2, 3 and 7
  • N. Goldenfeld, Lectures on Phase Transitions and the Renormalization Group (1992), chapters 2 to 5
  • L. Onsager, "Crystal statistics. I. A two-dimensional model with an order-disorder transition", Physical Review 65, 117 (1944)
  • C. N. Yang, "The spontaneous magnetization of a two-dimensional Ising model", Physical Review 85, 808 (1952)
  • CSIR-UGC NET Physical Sciences syllabus, Thermodynamic and Statistical Physics unit, published by the National Testing Agency

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Not affiliated with CSIR, UGC or the National Testing Agency. Check the official CSIR NET website for the current pattern. All questions are original.