PRINCIPLES OF STATISTICAL MECHANICS

Paul A. Pearce1

Mathematics Department, University of Melbourne,Parkville, Victoria 3052, Australia

Abstract

These lectures comprise an introductory course in statistical mechanics. The Gibbs formu-lation of the canonical ensemble is introduced and illustrated by application to simple modelsof magnets and fluids, specifically the ideal gas and the magnetic Ising spin chain. In addition,the classical mean field theories of fluids and ferromagnets, namely the van der Waals theoryof fluids and the Curie-Weiss theory of ferromagnets, are described and their critical behaviourelucidated. The lectures conclude with a general discussion of the principles of universalityand scaling which form the cornerstones of the modern theory of phase transitions and criticalphenomena.

1 Formulation of Statistical Mechanics

Statistical mechanics was the topic of the first A.N.U. Physics Summer School heldin 1988. On that occasion, Colin Thompson and I shared the pleasure of presentinglectures on the general principles of statistical mechanics [1, 2]. Those lectures, althoughunpublished, are available in the proceedings of that first A.N.U. Summer School. Theprinciples of statistical mechanics were in fact laid down at the turn of the century andso, not surprisingly, they have not changed since 1988. I will therefore be covering muchthe same ground in this series of lectures. Although this introduction is brief, a morecomprehensive account of the topics covered here can be found in Colin Thompson’sexcellent book [3] and a review article I wrote in 1983 [4]. Other very useful generalreferences are the books by Callen [5], Stanley [6], Huang [7], Baxter [8] and Yeomans[9]. The definitive reference for many topics in this area is of course “Phase Transitionsand Critical Phenomena” edited by Domb and Green and now Domb and Lebowitz [10].

1.1 Classical Mechanics and Phase Space

Thermodynamics describes the physical properties of bulk matter (solids, liquids andgases) in terms of a few variables such as absolute temperature T and pressure P . Ul-timately, of course, all of the macroscopic properties of matter (in equilibrium with itssurroundings) should be derivable from a knowledge of the fundamental interactions be-tween the constituent particles. This is the goal of statistical mechanics.

1Email: pap@mundoe.maths.mu.oz.au

1

The term statistical mechanics is a combination of mechanics and statistics. From amechanical viewpoint bulk matter, such as 22.4 litres of gas or 60 gram of iron at roomtemperature and pressure, typically consists of a system of

N ≈ NA = 6.0225× 1023 (Avogadro’s number) (1.1)

particles. In classical mechanics (we will not consider quantum mechanics in these lec-tures) the behaviour of a system of N particles with mass m is determined by writingdown a Hamiltonian describing the interactions between the particles and solving, usuallynumerically, the resulting equations of motion in the 6N -dimensional phase space

Γ = 〈{σ}〉 = 〈{(p1, q1,p2, q2, . . . ,pN , qN)}〉 (1.2)

spanned by the coordinates qi and their conjugate momenta pi in 3-dimensional space.The 6N -dimensional vector σ describes the microscopic state of the system. Explicitly,the Hamiltonian takes the form

H(σ) =N∑i=1

p2i

2m+ V (q1, q2, . . . , qN) (1.3)

= Kinetic + Potential Energy

and the equations of motion are

pi = −∂H∂qi

(= −∂V∂qi

= Force) (1.4)

qi =∂H

∂pi(=

pim

= velocity) (1.5)

Here derivatives with respect to vectors denote the appropriate gradients and the expres-sions in brackets give the Newtonian interpretation equivalent to F i = mqi.

Unfortunately, to solve the problem this way for a system of 6× 1023 particles wouldtake a millenia of CPU time on a Cray! Moreover, in practice, we never want to know themicroscopic details of every particle, rather, we want to know overall or average quantitiessuch as the average pressure exerted by a gas on its container. In other words we needto understand the statistics of the mechanical motion of a large number of particles.

1.2 Canonical Ensemble

The foundations of statistical mechanics were laid down by J.W. Gibbs in 1902. His workprovides an elegant mathematical prescription to calculate the thermodynamic quantitiesof a macroscopic system. The canonical ensemble describes a system of a fixed numberN of particles weakly coupled to and in thermal equilibrium with an infinitely large heatreservoir at absolute temperature T . Essentially, the system is to be regarded as isolatedbut maintained at a temperature T . The fundamental postulate (in general there is norigorous derivation) due to Gibbs is that the probability density ρ(σ) of points in phasespace Γ describing such an equilibrium system is given by

ρ(σ) =exp(−βH(σ))∫

Γexp(−βH(σ)) dΓ

, σ ∈ Γ (1.6)

2

where H(σ) is the Hamiltonian of the system (excluding interactions with the heat reser-voir), the integral is over all of the accessible phase space and

β =1

kT= inverse temperature (1.7)

where the absolute temperature T is measured in Kelvin (273K = 0◦C) and

k = 1.3805× 10−23 Joules/Kelvin = Boltzmann’s constant. (1.8)

Crudely speaking, the Boltzmann factor ρ(σ) gives the statistical probability of findingthe system in the state σ. Notice that the properties

ρ(σ) ≥ 0 and∫

Γρ(σ) dΓ = 1 (1.9)

are in accord with the interpretation of ρ(σ) as a probability. The normalization factor

ZN =∫

Γexp(−βH(σ)) dΓ (1.10)

is a fundamental quantity called the canonical partition function. Notice also that athigh temperatures (T →∞, β → 0)

ρ(σ) ∼ 1∫Γ dΓ

(1.11)

and all states are equally likely (random) while at low temperatures (T → 0, β → ∞)the low energy (ground) states are most probable.

1.3 Maxwell-Boltzmann Distribution

A simple heuristic argument can be given in favour of the Gibbs postulate in the case ofN particles distributed among a number of discrete energy levels Ej, j = 1, 2, 3, . . . withnj particles in each level as shown in Figure 1. The probability of a given distribution(partition) is

P =N !

n1!n2!n3! . . .(1.12)

Consider the thermodynamic limit so that N → ∞ and nj → ∞. Then by Stirling’sformula

N ! ∼(N

e

)N, etc. (1.13)

andlogP ∼ N logN −N −

∑j

(nj log nj − nj). (1.14)

To find the most probable configuration we must maximize logP subject to the twoconstraints ∑

j

nj = N,∑j

njEj = E = total energy. (1.15)

3

E1

E2

E3

E4

E5

E

n1

n2

n3

n4

n5

Figure 1. Distribution of N particles among a number of discrete energy levels.

We therefore cavalierly treat nj as continuous variables (properly we should use ρj =nj/N), introduce Lagrange multipliers α and β and maximize

N logN −N −∑j

(nj log nj − nj)− α∑j

nj − β∑j

njEj. (1.16)

Differentiating with respect to nj gives

log nj + α + βEj = 0 or nj = exp(−α− βEj). (1.17)

ThereforeN =

∑j

nj = e−α∑j

e−βEj = e−αZN (1.18)

and

ρj =njN

=e−βEj∑j

e−βEj= density of states (1.19)

gives the probability of finding the system in the energy state Ej. This distribution isthe Maxwell-Boltzmann distribution. The constant β is in fact identified as the inversetemperature.

1.4 Connection with Thermodynamics

Mathematically, the canonical partition function ZN can be regarded as a generatingfunction for the usual thermodynamic functions. Explicitly, the (Helmholtz) free energyΨ = Ψ(T, V ) is related to the canonical parition function by

Ψ = −kT logZN = U − TS (1.20)

where U = U(T, V ) is the internal energy, S = S(T, V ) is the entropy and V is thevolume. Other quantities are obtained by differentiation using the usual relations of

4

thermodynamics. The internal energy is

U = −T 2 ∂

∂T

(Ψ

T

)= Ψ− T ∂Ψ

∂T= Ψ + TS. (1.21)

Hence the entropy is

S = −∂Ψ

∂T. (1.22)

Similarly, the specific heat (at constant volume) is

CV =∂U

∂T= −T ∂

2Ψ

∂T 2(1.23)

and the pressure is

P = −∂Ψ

∂V. (1.24)

More generally, the experimental value of an observable A = A(σ) is obtained fromthe ensemble average or thermal expectation

〈A〉 =

∫ΓA(σ) exp(−βH(σ)) dΓ∫

Γexp(−βH(σ)) dΓ

. (1.25)

So, for example, the internal energy is indeed given by

U = 〈H〉 =

∫H(σ) exp(−βH(σ)) dΓ∫

exp(−βH(σ)) dΓ

= − ∂

∂βlog

∫exp(−βH(σ)) dΓ = − ∂

∂βlogZN (1.26)

= −(∂β

∂T

)−1∂

∂T

(− Ψ

kT

)= −T 2 ∂

∂T

(Ψ

T

).

There are in fact three different ensembles in general use — the microcanonical en-semble, the canonical ensemble and the grand canonical ensemble. In the microcanonicalensemble the system is completely isolated so that the number of particles N is fixed andthe total energy E is conserved. By contrast, as we have seen, in the canonical ensemblethe number of particles N is fixed but the average energy is determined by a temperatureT . Finally, in the grand canonical ensemble neither the number of particles N nor theenergy E is fixed, rather the average number of particles is controlled by the fugacity zand the average energy is controlled by the temperature T . The situation is summarizedin Table 1.

In each ensemble the details of the connection with thermodynamics is different buteach ensemble yields equivalent results in the thermodynamic limit when N and V aretaken to be infinitely large. Although there is in general no rigorous derivation of theGibbs ensembles and their connection with thermodynamics the fundamental postulatesare well confirmed in applications. Physicists adopt the attitude that the postulates

5

Table 1. Fixed quantities in the three standard ensembles.

Ensemble Fixed Quantities

Microcanonical N ECanonical N TGrand Canonical z T

are almost certainly correct and use them without question as the starting point forcalculations. At the same time, on a mathematical level, the whole edifice of statisticalmechanics is on shaky foundations! Needless to say, I will not dwell further on thefoundations of statistical mechanics. Instead, I will take the Gibbs prescription as givenand pragmatically concentrate on practical applications.

1.5 Ideal Gas

The ideal gas is a system of noninteracting point particles of mass m as shown in Figure 2.The Hamiltonian is

H =N∑i=1

p2i

2m(1.27)

and the canonical partition function is

ZN =1

N !

∫V· · ·

∫Vdq1 . . . dqN

∫ ∞−∞· · ·

∫ ∞−∞

dp1 . . . dpN exp

(− β

2m

N∑i=1

p2i

). (1.28)

The N ! appears here because the N particles are regarded as indistinguishable, thatis, configurations obtained by permuting the particles are considered to be identical(Boltzmann counting). This is not strictly correct. A proper treatment uses quantummechanics applied to a system of bosons or fermions.

Using the integral formula ∫ ∞−∞

exp(−λx2) dx =

√π

λ(1.29)

the partition function of the ideal gas is evaluated as

ZN =V N

N !

[∫ ∞−∞

exp

(− β

2mp2

)dp

]3N

=V N

N !

(2πm

β

)3N/2

(1.30)

so the pressure is given by

P = −∂Ψ

∂V=

∂

∂V(kT logZN) =

NkT

V. (1.31)

6

V

T

Figure 2. Ideal gas of noninteracting particles in a box of volume V . The temperature T is ameasure of the average internal energy

The ideal gas law is thusPV = NkT = nRT (1.32)

where n is the number of moles of gas (N = nNA) and

R = NAk = 8.315 Joules/Kelvin (1.33)

is the ideal gas constant.The internal energy is

U = − ∂

∂βlogZN =

3N

2β= 3N(1

2kT ). (1.34)

This illustrates the equipartition of energy between the 3N degrees of freedom. It alsoshows that the temperature T is indeed a measure of the average internal energy.

2 Lattice Spin Models

In many applications of statistical mechanics the particles are fixed or localized in space.In such cases the kinetic energy contribution to the Hamiltonian can be neglected. Withina magnetic crystal, for example, the atoms inherit an intrinsic angular momentum or spinfrom their valence electrons. These spins are restricted to certain discrete values and itis the interactions between these spins that are responsible for the magnetic propertiesof the material. A spin Hamiltonian for a system of N atoms (particles) is therefore ofthe form

H(σ) = V (σ1, σ2, . . . , σN) (2.1)

where σ = {σ1, σ2, . . . , σN} denotes the microscopic configuration and the spin σi is adiscrete variable describing the the state of the particle at the lattice site i. The canonicalpartition function of a lattice spin model is thus given by the configurational sum

ZN =∑σ

exp(−βH(σ)). (2.2)

7

2.1 Ising Paramagnet

In the presence of an external magnetic field h, the magnetic moments or spins of atomswithin magnetic materials tend to align themselves with the field. The first explanationof this phenomenon was due to Langevin [11] in 1905. For simplicity, let us assume thatthe spins are given by

σi ={

+1, if spin i is parallel to h−1, if spin i is antiparallel to h.

(2.3)

Such two-valued spins are called Ising spins after E. Ising [12] who first studied suchmodels in 1925. The Hamiltonian or energy function for the Ising paramagnet is

H = −hN∑i=1

σi, h ≥ 0. (2.4)

Clearly, the lowest energy (ground) state occurs when all the spins align with the externalfield, that is, σi = +1 for all i.

If we set B = βh, the canonical partition function is

ZN =∑σ

exp

(βh

N∑i=1

σi

)=

∑σ1=±1

∑σ2=±1

. . .∑

σN=±1

eBσ1eBσ2 . . . eBσN

= (2 coshB)N . (2.5)

Hence− βΨ = logZN = N log(2 coshB). (2.6)

The average magnetic moment or magnetization is given by

m =

⟨1

N

∑i

σi

⟩=

∑σ

(1

N

∑i

σi

)exp

(B

N∑i=1

σi

)∑σ

exp

(B

N∑i=1

σi

)

=1

N

∂

∂BlogZN = − 1

N

∂

∂B(βΨ) = tanhB. (2.7)

It follows that the magnetization of a paramagnet vanishes in the absence of an externalmagnetic field and that there is no permanent magnetization as in ferromagnetic materialssuch as iron and nickel.

2.2 Ising Ferromagnets

In ferromagnets there must be additional interactions between spins to explain the coop-erative alignment of spins in the absence of an external magnetic field. The first modelof a ferromagnet to entail such explicit interactions between spins was due to Ising andhis Ph.D. supervisor Lenz. The Hamiltonian of an Ising ferromagnet is

H = −∑〈i,j〉

Jijσiσj − h∑i

σi (2.8)

8

where the first sum is over all pairs 〈i, j〉 of lattice sites and Jij is the strength of in-teraction between the spins at sites i and j. Clearly, to energetically favour the mutualalignment of spins, as in a ferromagnet, we need Jij ≥ 0 for all i and j. Physically,the interactions between spins in a magnetic substance are short ranged. The simplestpossibility is that only nearest neighbour spins on a regular lattice interact so that

Jij ={J, i, j adjacent0, otherwise.

(2.9)

In general, it is impossible to evaluate the partition function even of a nearest-neighbour Ising model on a regular lattice consisting of N = 6×1023 sites. In practice, Nis effectively infinite and what we really need to evaluate is not the free energy Ψ, whichis extensive and grows with the size of the system, but rather the intensive quantity ψgiven by the free energy per site in the thermodynamic limit

− βψ = − limN→∞

1

NΨ = lim

N→∞

1

NlogZN . (2.10)

It can be shown quite generally that this limit exists, however, the resulting limitingfree energy need not in general be an analytic function of the thermodynamic variablessuch as h and T . Points where the limiting free energy is singular are called phasetransition points. In the special case of the nearest-neighbour Ising model on a periodic(one-dimensional) chain the limiting free energy can in fact be evaluated without muchdifficulty. This was first done by Ising in 1925. Here we will solve this model explicitlyusing transfer matrices.

2.3 Ising Spin Chain

The Hamiltonian of the Ising spin chain is

H = −JN∑i=1

σiσi+1 − hN∑i=1

σi, J > 0. (2.11)

The problem is to evaluate the partition function

ZN =∑σ

exp

(K

N∑i=1

σiσi+1 +BN∑i=1

σi

)(2.12)

where K = βJ and B = βh. Since the chain is periodic σN+1 = σ1. Let us define a 2× 2transfer matrix T with elements

〈σ|T |σ′〉 = exp[Kσσ′ + 1

2B(σ + σ′)

](2.13)

that is

T =

(〈1|T |1〉 〈1|T | − 1〉〈−1|T |1〉 〈−1|T | − 1〉

)=

(eK+B e−K

e−K eK−B

). (2.14)

We can then write

ZN =∑σ

N∏i=1

exp[Kσiσi+1 + 1

2B(σi + σi+1)

]=

∑σ1

· · ·∑σN

〈σ1|T |σ2〉〈σ2|T |σ3〉 . . . 〈σN−1|T |σN〉〈σN |T |σ1〉. (2.15)

9

These are matrix products so

ZN =∑σ1

〈σ1|TN |σ1〉 = TrTN = λN+ + λN− (2.16)

where λ+ ≥ λ− are the eigenvalues of the real symmetric matrix T . The characteristicpolynomial is

λ2 − (2eK coshB)λ+ 2 sinh 2K = 0 (2.17)

so

λ± = eK coshB ±√e2K sinh2 B + e−2K (2.18)

and for T > 0 we have λ+ > λ− > 0.It follows that

1

NlogZN =

1

Nlog(λN+ + λN− ) =

1

Nlog λN+

1 +

(λ−λ+

)N= log λ+ +

1

Nlog

1 +

(λ−λ+

)N→ log λ+ as N →∞. (2.19)

Hence the free energy per spin ψ in the thermodyamic limit is given by

− βψ = limN→∞

1

NlogZN = log λ+. (2.20)

The magnetization is

m = − ∂

∂B(βψ) =

∂

∂Blog λ+ =

sinhB√sinh2 B + e−4K

. (2.21)

Clearly, 0 ≤ m ≤ 1 and m→ 1 if either J →∞, h→∞ or T → 0.In zero field (h → 0±) the magnetization vanishes, so there is no residual or spon-

taneous magnetization! Moreover, for T > 0, the limiting free energy is an analyticfunction of h and T so the Ising spin chain does not undergo a phase transition. Asasserted by van Hove’s theorem [13], this is a general feature of one-dimensional modelswith finite-range interactions. In sharp contrast, Ising models on cubic lattices in twoor more dimensions do exhibit spontaneous magnetization! The zero-field Ising modelon the square lattice was in fact solved by L. Onsager [14] in 1944 and the spontaneousmagnetization was first calculated by C.N. Yang [15] in 1952.

Theorem 1 (van Hove) The limiting free energy ψ(h, T ) of the one-dimensional finite-range Ising model

H = −∑

1≤i<j≤NJ(j − i)σiσj − h

N∑i=1

σi (2.22)

withJ(k) = 0 for k > R (2.23)

is an analytic function of h and T for T > 0 and

limh→0±

m(h, T ) = 0 for T > 0. (2.24)

10

Sketch of Proof: For finite-range interactions we can always define a finite-dimensionaltransfer matrix T . Since the elements of T are Boltzmann weights, T is a positive matrix(has all positive entries). Hence, by the Frobenius theorem, the eigenvalues of T satisfy

λmax = λ1 > |λ2| ≥ |λ3| ≥ · · · |λn| (2.25)

that is, the largest eigenvalue is real, positive and nondegenerate. Hence

1

NlogZN =

1

Nlog(λN1 + · · ·+ λNn )

= log λ1 +1

Nlog

1 +

(λ2

λ1

)N+

(λ3

λ1

)N+ · · ·+

(λnλ1

)N (2.26)

→ log λmax as N →∞.So the free energy

− βψ(h, T ) = limN→∞

1

NlogZN = log λmax (2.27)

and the magnetization m(h, T ) are analytic because λmax > 0 is analytic. But themagnetization is an odd function of h, that is m(h, T ) = −m(−h, T ), so m(0, T ) = 0 =m(0±, T ) by continuity and there is no spontaneous magnetization .

The Ising model can also be regarded as a model of a lattice gas by transforming tooccupation numbers

ti = 12(1− σi) =

{1, if site i is occupied (σi = −1)0, if site i is unoccupied (σi = +1).

(2.28)

In this case the hard-core repulsion between atoms excludes multiple occupancy of a site.A more realistic model of a gas, however, should allow for a continuous distribution ofparticles.

3 Tonks and van der Waals Gases

3.1 Tonks Gas

Tonks gas [16] is a model of hard-core particles (spheres or rods) in one dimension asshown in Figure 3. Effectively, this is a model of one-dimensional billiard balls. The pairinteraction (hard-core) potential is

φ(r) = φhc(r) ={∞, 0 ≤ r < a

0, r ≥ a.(3.1)

Hence the Hamiltonian for N rods on the interval 0 ≤ x ≤ L is

H =∑

1≤i<j≤Nφhc(|xi − xj|). (3.2)

Note that H is a symmetric function of x1, x2, . . . , xN and we have omitted the kineticenergy term. The canonical partition function can therefore be written as

ZN =1

N !

∫ L

0· · ·

∫ L

0exp

−β∑i<j

φhc(|xi − xj|) dx1 . . . dxN

=∫· · ·

∫R

exp

−β∑i<j

φhc(|xi − xj|) dx1 . . . dxN (3.3)

11

0 L L+a

x1 x2 x3 xN+1

where R is the regionR : 0 ≤ x1 ≤ x2 ≤ . . . ≤ xN ≤ L. (3.4)

Figure 3. Tonks gas of hard-core particles distributed along a line of length L.

Now exp(−βφhc) = 0 or 1 so

ZN =∫· · ·

∫R′dx1 . . . dxN (3.5)

where

R′ : 0 ≤ x1 ≤ x2 − a, a ≤ x2 ≤ x3 − a, . . . , (i− 1)a ≤ xi ≤ xi+1 − a, . . .. . . , (N − 1)a ≤ xN ≤ xN+1 − a = L. (3.6)

If we now change variables to yi = xi − (i− 1)a we find that

0 ≤ yi ≤ xi+1 − a− (i− 1)a = xi+1 − ia = yi+1 (3.7)

and hence

ZN =∫ `

0dyN

∫ yN−1

0dyN−1 · · ·

∫ y3

0dy2

∫ y2

0dy1 =

`N

N !(3.8)

where ` = L − (N − 1)a is the effective volume. Again using Stirling’s formula N ! ∼(N/e)N , the limiting free energy is given by

−βψ = limN,L→∞L/N=v

1

NlogZN

= 1 + limN,L→∞

logL− (N − 1)a

N= 1 + log(v − a) (3.9)

where v > a is the volume per particle. The limit v → a is the close packing limit.The pressure is

P = −∂ψ∂v

=kT

v − a (3.10)

so the equation of state isP (v − a) = kT (3.11)

which is the ideal gas law with the volume per particle V/N replaced with the free volumeper particle v−a. Notice that there is no phase transition since the free energy is analyticfor v > a.

12

r

φ

−αL

L

3.2 Tonks–van der Waals Gas

In one dimension, Tonks gas is an improvement over the ideal gas since it takes intoaccount the finite size of particles and the hard-core exclusion between them. However,real particles also interact through attractive Lennard-Jones dispersion forces. A simpleway to model this is provided by the Tonks–van der Waals potential shown in Figure 4

φ(r) = φhc(r)−α

L, α > 0. (3.12)

This is unrealistic because the strength of the potential should not depend upon the size

Figure 4. Tonks–van der Waals potential.

L of the system. Nevertheless this assumption leads to a model that is both tractableand interesting. The Hamiltonian is

H =∑

1≤i<j≤N

[φhc(|xi − xj|)−

α

L

]= −αN(N − 1)

2L+HTonks. (3.13)

Hence

ZN = exp

[βαN(N − 1)

2L

]ZTonksN (3.14)

and

− βψ = limN,L→∞L/N=v

1

NlogZN =

βα

2v+ 1 + log(v − a). (3.15)

It follows that the pressure is

P = −∂ψ∂v

=kT

v − a −α

2v2= Phc −

α

2v2. (3.16)

The outcome is that the pressure is reduced due to the attractive interactions of theparticles by an amount proportional to the strength α of the interactions and also pro-portional to the square of the density 1

2ρ2 = 1/2v2 which gives the probability of two

particles interacting. The equation of state is thus modified to(P +

α

2v2

)(v − a) = kT (3.17)

which is the celebrated equation of state proposed, on phenomenological grounds, by vander Waals [17] in 1873.

13

3.3 Maxwell Construction

There is a well known problem with the van der Waals equation of state. If the isothermsfor P as a function of v are plotted it is found that there are wiggles at low temperatureswhere

∂v

∂P> 0. (3.18)

This asserts that the gas actually expands as you labour to compress it and violatesthermodynamic stability! The remedy for this situation was provided by Maxwell [18]who proposed the double tangent formula

ψ =Convex

Envelope

[ψhc(v)− α

2v

]. (3.19)

This construction is equivalent to placing horizontal segments in the isotherms accordingto an equal area rule as illustrated in Figure 5. Such flat regions are found experimentallyin isotherms at low temperatures throughout the gas-liquid coexistence region. TheMaxwell construction results in a free energy which is no longer analytic and thus leads toa phase transition. The van der Waals–Maxwell theory can in fact be obtained rigorouslyby taking a limit of infinitely weak long-range potentials after the thermodynamic limit[19, 20].

4 Curie-Weiss Ferromagnet

4.1 Weiss Phenomenological Theory

The classical theory of ferromagnetism was proposed by Weiss in 1907. This phenomeno-logical theory is based on the paramagnet

H = heff

N∑i=1

σi (4.1)

and assumes that, in a ferromagnet, there is an internal field in addition to the externalfield h. On the average the internal field, due to the cooperative alignment of the spins,is proportional to the magnetization m so that the local effective field seen by a spin is

heff = Jm+ h (4.2)

where the constant of proportionality J is called the mean-field parameter. Evaluatingthe magnetization then leads to the transcendental equation of state

m = tanh(βheff) = tanh(Km+B), K = βJ, B = βh. (4.3)

This self-consistency equation is used to determine m = m(h, T ). Since this equationcannot be solved analytically it is solved graphically, as shown in Figure 6, in the form

Jm+ h

kT= tanh−1 m. (4.4)

Given h and T , this equation can admit one, two or three solutions for the magnetization

14

P

ψ

A = B

vL v

v

vG

AB

vL vG

Figure 5. Schematic representation of the van der Waals wiggle in a low temperature isotherm ofP = −∂ψ

∂v plotted against v. The wiggle is removed by placing a flat segment into the isothermaccording to an equal area rule as illustrated. Also shown is the corresponding kink in the freeenergy isotherm with the equivalent Maxwell double tangent (convex envelope) construction.

m. If the slope of the straight line is less than the critical value

J

kTc= 1 (4.5)

there is just one solution. In general there can be more than one solution but, if weassume that m and h have the same sign as is physically reasonable, then for h > 0 theequation of state determines m uniquely. This assumption is analogous to the Maxwellconstruction for the van der Waals fluid and leads to the magnetic isotherms shown inFigure 7.

The spontaneous magnetization m0 is defined by

m0 = limh→0+

m(h, T ) (4.6)

and

m0(T ) =

{0, T ≥ Tc = J/kx, T < Tc = J/k

(4.7)

where x is the positive solution of

x = tanhJx

kT. (4.8)

15

y = (Jx+h)/kT

y

x

y = tanh-1 x

-1.5 -1 -0.5 0.5 1 1.5

-1

-0.75

-0.5

-0.25

0.25

0.5

0.75

1 m

h

Figure 6. Graphical solution of the self-consistency equation Jm+hkT = tanh−1m. At low tem-

peratures and for small magnetic fields h there are three solutions as shown. As h → 0 thepositive solution gives the spontaneous magnetization m0.

Figure 7. Magnetic isotherms for the Weiss theory. The three isotherms shown are respectivelyfor T < Tc, T = Tc and T > Tc. The dashed line shows the discarded solution.

16

TTc

h

m>0

m<0

Thus the Weiss theory correctly predicts spontaneous magnetization and a phase transi-tion in zero field from a magnetized phase (T < Tc, m > 0) to a paramagnetic (nonmag-netized) phase (T > Tc, m = 0) as the temperature is raised through the critical valueTc = J/k called the Curie point. The phase diagram is shown in Figure 8.

Figure 8. Phase diagram of the Ising ferromagnet. A first-order line extends along the lineh = 0 from zero temperature to the Curie point at T = Tc.

4.2 Equivalent Neighbour Ising Model

The results of the Curie-Weiss theory can be obtained using the canonical ensemble bystarting with the equivalent neighbour Hamiltonian

H = − JN

∑1≤i<j≤N

σiσj − hN∑i=1

σi, J > 0. (4.9)

Here the sites labelled 1, 2, . . . , N are all equivalent and no lattice structure is assumedor needed. As for the Tonks–van der Waals gas, the interactions are unphysical becausethey are independent of separation and depend on the size of the system N . (This factorof N is needed to ensure the existence of the thermodynamic limit.)

To calculate the partition function we begin by writing

H = 12J − J

2N

(N∑i=1

σi

)2

− hN∑i=1

σi (4.10)

so that

ZN = e−K/2∑σ

exp

K2N

(N∑i=1

σi

)2

+BN∑i=1

σi

(4.11)

where K = βJ and B = βh. Next we use the identity

exp(12αS2) =

√α

2π

∫ ∞−∞

dx exp(−12αx2 + αxS) (4.12)

17

with

α = NK and S =1

N

N∑i=1

σi (4.13)

to obtain

ZN =

√NK

2πe−K/2

∑σ

∫ ∞−∞

dx exp

[−1

2NKx2 + (Kx+B)

N∑i=1

σi

]

=

√NK

2πe−K/2

∫ ∞−∞

dx e−12NKx2 ∑

σ1=±1

· · ·∑

σN=±1

e(Kx+B)σ1 . . . e(Kx+B)σN

=

√NK

2πe−K/2

∫ ∞−∞

dx e−12NKx2

[2 cosh(Kx+B)]N (4.14)

=

√NK

2πe−K/2

∫ ∞−∞

dx exp{N[−1

2Kx2 + log 2 cosh(Kx+B)

]}The limiting free energy per spin ψ is given by

− βψ = limN→∞

1

NlogZN = lim

N→∞

1

Nlog IN (4.15)

whereIN =

∫ ∞−∞

exp[Nf(x)] dx (4.16)

andf(x) = −1

2Kx2 + log 2 cosh(Kx+B). (4.17)

This follows since

limN→∞

1

Nlog

√NK2π

e−K/2

= 0. (4.18)

For large N , IN is dominated by the maximum value of the integrand and ψ can beobtained by Laplace’s method giving

− βψ = maxx

f(x) = maxx

[−1

2Kx2 + log 2 cosh(Kx+B)

]. (4.19)

Differentiating with respect to x, we see that the equation determining the maximum is

x = tanh(Kx+B) (4.20)

which is precisely the Curie-Weiss equation of state where x is identified with the mag-netization

m =∂

∂B(−βψ) = tanh(Kx+B) = x. (4.21)

Furthermore, for the maximizing solution we see that m = x and h are always of thesame sign as was previously assumed. The equivalent neighbour Ising model thereforeundergoes a phase transition at the critical point h = 0, T = Tc = J/k. Again the Curie-Weiss theory can be obtained rigorously by taking a limit of infinitely weak long-rangepotentials [21, 20] or alternatively by taking the lattice dimensionality d→∞ after thethermodynamic limit [22].

18

5 Critical Exponents, Universality and Scaling

5.1 Critical Exponents

The behaviour of thermodynamic functions in the vicinity of a critical point is character-ized by critical exponents. These describe the power law behaviour asymptotically closeto the critical point. We write

f(x) ∼ xε as x→ 0+ (5.1)

whenever the limit

ε = limx→0+

log f(x)

log x(5.2)

exists. This limit defines the critical exponent ε of the function f(x) at the critical pointx = 0. Similarly, we define two-sided limits and one-sided limits for x→ 0−.

In statistical mechanics there has been a proliferation of critical exponents which nowexhaust the Greek alphabet! The standard critical exponents for magnetic and fluidsystems are defined in Table 2.

Table 2. The definition of some critical exponents for magnetic and fluid systems. Heret = (T − Tc)/Tc, C0 = zero-field specific heat, χ0 = zero-field susceptibility, KT = isothermalcompressibility and ρL,G = liquid, gas density.

Exponent Magnet Fluid

α C0 ∼ |t|−α, t→ 0 CV ∼ |t|−α, t→ 0β m0 ∼ |t|β, t→ 0− ρL − ρG ∼ |t|β, t→ 0−γ χ0 ∼ |t|−γ, t→ 0 KT ∼ |t|−γ, t→ 0δ h ∼ sgn(m)|m|δ, h→ 0, T = Tc P − Pc ∼ sgn(ρ− ρc)|ρ− ρc|δ,

|ρ− ρc| → 0, T = Tc

5.2 Mean-field Critical Exponents

In this section we obtain the critical exponents of the Curie-Weiss ferromagnet.

α = 0disc

The non-analytic zero-field free energy is

− ψ

kT=

log 2, T ≥ Tc

−Jm20

2kT+ log 2 cosh

(Jm0

kT

), T < Tc

(5.3)

withm0 = tanhKm0. (5.4)

19

Hence

U = −T 2 d

dT

(ψ

T

)=

{0, T ≥ Tc−1

2Jm2

0, T < Tc(5.5)

and

C0 =dU

dT=

0, T ≥ Tc

−12Jdm2

0

dT, T < Tc

(5.6)

This yields a jump discontinuity in C0 so α = 0disc.

β = 1/2

If we set h = 0 then m0 is small near the critical point T = Tc = J/k so we can Taylorexpand the equation of state

Jm0

kT=TcTm0 = tanh−1 m0 = m0 + 1

3m3

0 + . . . (5.7)

This gives

m20 ∼ 3

(TcT− 1

)= 3

TcT

(1− T

Tc

)(5.8)

and hence as T → Tc−

m0 ∼(

1− T

Tc

)βwith β = 1/2. (5.9)

γ = 1

The zero-field susceptibility is

χ0 = βdm

dB

∣∣∣∣∣B=0

. (5.10)

But differentiating m = tanh(Km+B) implicitly with respect to B gives

dm

dB=

1−m2

1−K(1−m2)and so χ0 =

β(1−m20)

1−K(1−m20). (5.11)

It follows that as T → Tc

χ0 ∼∣∣∣∣1− T

Tc

∣∣∣∣−γ with γ = 1. (5.12)

δ = 3

Finally, if we set T = Tc = J/k then m is small for small h so we can Taylor expandthe equation of state along the critical isotherm

J

kTcm+

h

kTc= m+

h

J= tanh−1 m ∼ m+ 1

3m3 + . . . (5.13)

20

Hence we conclude that as h→ 0

m ∼ h1/δ with δ = 3. (5.14)

In summary, the critical exponents of the Curie-Weiss ferromagnet are

α = 0disc, β = 1/2, γ = 1, δ = 3. (5.15)

These are the classical values for the critical exponents. A similar analysis shows thatthe van der Waals–Maxwell fluid has precisely the same classical values for the criticalexponents. Typical experimental values for these exponents are

α ≈ 0.1, β ≈ 0.33, γ ≈ 1.2, δ ≈ 4.2 (5.16)

for both fluid and magnetic systems in three dimensions. Clearly the classical criticalbehaviour and exponents are wrong!

5.3 Universality and Scaling

The critical exponents appear to be insensitive to the microscopic details of the system.This empirical fact is embodied in the following:

Universality Hypothesis

For short-range interactions, the critical exponents depend only on the spatial dimen-sion d and the symmetries of the Hamiltonian H.

Critical exponents for some much studied models in statistical mechanics are given inTable 3. The hard hexagon [23] and Potts models [24] are discussed further in RodneyBaxter’s lectures at this summer school. The fact that the mean-field theories of bothmagnets and fluids have the same classical values values is one example of universality.Another example is that the Ising model in two dimensions exhibits precisely the samecritical exponents for all the regular lattices, eg. the square, the triangular and thehexagonal lattices. Finally, the hard hexagon and 3-state Potts models in two dimensionsare seen to have the same exponents. This is a consequence of universality since bothmodels possess a Z3 symmetry.

Another important hypothesis in the modern theory of critical phenomena is thescaling hypothesis. For a simple magnetic system, which has just two relevant thermo-dynamic fields h and t, this takes the following form:

Scaling Hypothesis

There exist two exponents x and y such that, asymptotically close to the critical point,the free energy can be written as

ψ = ψanal + ψsing (5.17)

where ψanal is analytic and the singular part ψsing satisfies

ψsing(λxt, λyh) = λψsing(t, h) (5.18)

21

Table 3. Critical exponents of some much studied lattice spin models. These values are allknown exactly except for the critical exponents of the Ising model in three dimensions.

Model α β γ δ

Classical 0disc 1/2 1 3Ising (d = 2) 0log 1/8 7/4 15Ising (d = 3) 0.10 0.33 1.24 4.8Ising (d ≥ 4) 0 1/2 1 3Hard Hexagons (d = 2) 1/3 1/9 13/9 143-state Potts (d = 2) 1/3 1/9 13/9 144-state Potts (d = 2) 2/3 1/12 7/6 15

for all values of the scaling parameter λ, that is, ψsing is a generalized homogeneousfunction of t = (T − Tc)/Tc and h.

By differentiating the homogeneous relation satisfied by ψsing, it is possible to obtainthe exponents α, β, γ, δ in terms of x and y. Hence only two of these exponents areindependent. In particular, the exponents α, β, γ, δ are thus found to satisfy the scalingrelations

α + 2β + γ = 2, γ = β(δ − 1). (5.19)

In most cases, these relations are used to calculate the values of γ and δ in Table 3.However, these relations are clearly satisfied by the classical values. In fact, for theCurie-Weiss ferromagnet, it can be shown that

ψsing(t, h) = mins{−hs+ 1

2Jts2 + 1

12Js4}. (5.20)

This is a generalized homogeneous function with exponents x = 1/2 and y = 3/4.

6 Acknowledgement

It is a pleasure to acknowledge my indebtedness to Colin Thompson from whom I firstlearned much of the material presented here. I thank Brian Robson, Vladimir Bazhanovand Rodney Baxter for their hospitality in Canberra.

References

1. C. J. Thompson, “General Principles of Statistical Mechanics I”, First Physics Sum-mer School, A.N.U. unpublished 1988.

2. P. A. Pearce, “General Principles of Statistical Mechanics II”, First Physics SummerSchool, A.N.U. unpublished 1988.

22

3. C. J. Thompson, “Classical Equilibrium Statistical Mechanics”, Clarendon Press,Oxford, 1988.

4. P. A. Pearce, “Thermodynamic Models with Exact Solutions”, Science Progress(Oxford) 68 (1983) 189.

5. H. B. Callen, “Thermodynamics”, Wiley, New York, 1960.

6. H. E. Stanley, “Introduction to Phase Transitios and Critical Phenomena”, OxfordUniversity Press, Oxford, 1971.

7. K. Huang, “Statistical Mechanics”, 2nd edition, Wiley, New York, 1987.

8. R. J. Baxter, “Exactly Solved Models in Statistical Mechanics”, Academic Press,London, 1982.

9. J. M. Yeomans, “Statistical Mechanics of Phase Transitions”, Clarendon Press, Ox-ford, 1992.

10. C. Domb and M. S. Green/C. Domb and J. L. Lebowitz, Phase Transitions andCritical Phenomena”, Vols. 1–14, Academic Press, London, 1972–1994.

11. P. Langevin, J. de Phys. 4 (1905) 678.

12. E. Ising, Z. Phys. 31 (1925) 253.

13. L. van Hove, Physica 15 (1949) 951.

14. L. Onsager, Phys. Rev. 65 (1944) 117.

15. C. N. Yang, Phys. Rev. 85 (1952) 809.

16. L. Tonks, Phys. Rev. 50 (1936) 955.

17. J. D. van der Waals, Ph. D. Thesis, University of Leiden (1873).

18. J. C. Maxwell, Nature 11 (1874) 53.

19. J. L. Lebowitz and O. Penrose, J. Math. Phys. 6 (1966) 1282.

20. C. J. Thompson, Prog. Theor. Phys. 87 (1992) 535, and references therein.

21. C. J. Thompson and H. Silver, Commun. Math. Phys. 33 (1973) 53.

22. C. J. Thompson, Commun. Math. Phys. 36 (1974) 255; P. A. Pearce and C. J.Thompson, Commun. Math. Phys. 58 (1978) 131.

23. R. J. Baxter, J. Phys. A 13 (1980) L61.

24. R. B. Potts, Proc. Camb. Phil. Soc. 48 (1952) 106.

23